]> Git Repo - linux.git/blob - mm/hugetlb.c
Merge tag 'ceph-for-6.9-rc1' of https://github.com/ceph/ceph-client
[linux.git] / mm / hugetlb.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Generic hugetlb support.
4  * (C) Nadia Yvette Chambers, April 2004
5  */
6 #include <linux/list.h>
7 #include <linux/init.h>
8 #include <linux/mm.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/memblock.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/sched/mm.h>
23 #include <linux/mmdebug.h>
24 #include <linux/sched/signal.h>
25 #include <linux/rmap.h>
26 #include <linux/string_helpers.h>
27 #include <linux/swap.h>
28 #include <linux/swapops.h>
29 #include <linux/jhash.h>
30 #include <linux/numa.h>
31 #include <linux/llist.h>
32 #include <linux/cma.h>
33 #include <linux/migrate.h>
34 #include <linux/nospec.h>
35 #include <linux/delayacct.h>
36 #include <linux/memory.h>
37 #include <linux/mm_inline.h>
38 #include <linux/padata.h>
39
40 #include <asm/page.h>
41 #include <asm/pgalloc.h>
42 #include <asm/tlb.h>
43
44 #include <linux/io.h>
45 #include <linux/hugetlb.h>
46 #include <linux/hugetlb_cgroup.h>
47 #include <linux/node.h>
48 #include <linux/page_owner.h>
49 #include "internal.h"
50 #include "hugetlb_vmemmap.h"
51
52 int hugetlb_max_hstate __read_mostly;
53 unsigned int default_hstate_idx;
54 struct hstate hstates[HUGE_MAX_HSTATE];
55
56 #ifdef CONFIG_CMA
57 static struct cma *hugetlb_cma[MAX_NUMNODES];
58 static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata;
59 static bool hugetlb_cma_folio(struct folio *folio, unsigned int order)
60 {
61         return cma_pages_valid(hugetlb_cma[folio_nid(folio)], &folio->page,
62                                 1 << order);
63 }
64 #else
65 static bool hugetlb_cma_folio(struct folio *folio, unsigned int order)
66 {
67         return false;
68 }
69 #endif
70 static unsigned long hugetlb_cma_size __initdata;
71
72 __initdata struct list_head huge_boot_pages[MAX_NUMNODES];
73
74 /* for command line parsing */
75 static struct hstate * __initdata parsed_hstate;
76 static unsigned long __initdata default_hstate_max_huge_pages;
77 static bool __initdata parsed_valid_hugepagesz = true;
78 static bool __initdata parsed_default_hugepagesz;
79 static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata;
80
81 /*
82  * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
83  * free_huge_pages, and surplus_huge_pages.
84  */
85 DEFINE_SPINLOCK(hugetlb_lock);
86
87 /*
88  * Serializes faults on the same logical page.  This is used to
89  * prevent spurious OOMs when the hugepage pool is fully utilized.
90  */
91 static int num_fault_mutexes;
92 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
93
94 /* Forward declaration */
95 static int hugetlb_acct_memory(struct hstate *h, long delta);
96 static void hugetlb_vma_lock_free(struct vm_area_struct *vma);
97 static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma);
98 static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma);
99 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
100                 unsigned long start, unsigned long end);
101 static struct resv_map *vma_resv_map(struct vm_area_struct *vma);
102
103 static inline bool subpool_is_free(struct hugepage_subpool *spool)
104 {
105         if (spool->count)
106                 return false;
107         if (spool->max_hpages != -1)
108                 return spool->used_hpages == 0;
109         if (spool->min_hpages != -1)
110                 return spool->rsv_hpages == spool->min_hpages;
111
112         return true;
113 }
114
115 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool,
116                                                 unsigned long irq_flags)
117 {
118         spin_unlock_irqrestore(&spool->lock, irq_flags);
119
120         /* If no pages are used, and no other handles to the subpool
121          * remain, give up any reservations based on minimum size and
122          * free the subpool */
123         if (subpool_is_free(spool)) {
124                 if (spool->min_hpages != -1)
125                         hugetlb_acct_memory(spool->hstate,
126                                                 -spool->min_hpages);
127                 kfree(spool);
128         }
129 }
130
131 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
132                                                 long min_hpages)
133 {
134         struct hugepage_subpool *spool;
135
136         spool = kzalloc(sizeof(*spool), GFP_KERNEL);
137         if (!spool)
138                 return NULL;
139
140         spin_lock_init(&spool->lock);
141         spool->count = 1;
142         spool->max_hpages = max_hpages;
143         spool->hstate = h;
144         spool->min_hpages = min_hpages;
145
146         if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
147                 kfree(spool);
148                 return NULL;
149         }
150         spool->rsv_hpages = min_hpages;
151
152         return spool;
153 }
154
155 void hugepage_put_subpool(struct hugepage_subpool *spool)
156 {
157         unsigned long flags;
158
159         spin_lock_irqsave(&spool->lock, flags);
160         BUG_ON(!spool->count);
161         spool->count--;
162         unlock_or_release_subpool(spool, flags);
163 }
164
165 /*
166  * Subpool accounting for allocating and reserving pages.
167  * Return -ENOMEM if there are not enough resources to satisfy the
168  * request.  Otherwise, return the number of pages by which the
169  * global pools must be adjusted (upward).  The returned value may
170  * only be different than the passed value (delta) in the case where
171  * a subpool minimum size must be maintained.
172  */
173 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
174                                       long delta)
175 {
176         long ret = delta;
177
178         if (!spool)
179                 return ret;
180
181         spin_lock_irq(&spool->lock);
182
183         if (spool->max_hpages != -1) {          /* maximum size accounting */
184                 if ((spool->used_hpages + delta) <= spool->max_hpages)
185                         spool->used_hpages += delta;
186                 else {
187                         ret = -ENOMEM;
188                         goto unlock_ret;
189                 }
190         }
191
192         /* minimum size accounting */
193         if (spool->min_hpages != -1 && spool->rsv_hpages) {
194                 if (delta > spool->rsv_hpages) {
195                         /*
196                          * Asking for more reserves than those already taken on
197                          * behalf of subpool.  Return difference.
198                          */
199                         ret = delta - spool->rsv_hpages;
200                         spool->rsv_hpages = 0;
201                 } else {
202                         ret = 0;        /* reserves already accounted for */
203                         spool->rsv_hpages -= delta;
204                 }
205         }
206
207 unlock_ret:
208         spin_unlock_irq(&spool->lock);
209         return ret;
210 }
211
212 /*
213  * Subpool accounting for freeing and unreserving pages.
214  * Return the number of global page reservations that must be dropped.
215  * The return value may only be different than the passed value (delta)
216  * in the case where a subpool minimum size must be maintained.
217  */
218 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
219                                        long delta)
220 {
221         long ret = delta;
222         unsigned long flags;
223
224         if (!spool)
225                 return delta;
226
227         spin_lock_irqsave(&spool->lock, flags);
228
229         if (spool->max_hpages != -1)            /* maximum size accounting */
230                 spool->used_hpages -= delta;
231
232          /* minimum size accounting */
233         if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
234                 if (spool->rsv_hpages + delta <= spool->min_hpages)
235                         ret = 0;
236                 else
237                         ret = spool->rsv_hpages + delta - spool->min_hpages;
238
239                 spool->rsv_hpages += delta;
240                 if (spool->rsv_hpages > spool->min_hpages)
241                         spool->rsv_hpages = spool->min_hpages;
242         }
243
244         /*
245          * If hugetlbfs_put_super couldn't free spool due to an outstanding
246          * quota reference, free it now.
247          */
248         unlock_or_release_subpool(spool, flags);
249
250         return ret;
251 }
252
253 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
254 {
255         return HUGETLBFS_SB(inode->i_sb)->spool;
256 }
257
258 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
259 {
260         return subpool_inode(file_inode(vma->vm_file));
261 }
262
263 /*
264  * hugetlb vma_lock helper routines
265  */
266 void hugetlb_vma_lock_read(struct vm_area_struct *vma)
267 {
268         if (__vma_shareable_lock(vma)) {
269                 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
270
271                 down_read(&vma_lock->rw_sema);
272         } else if (__vma_private_lock(vma)) {
273                 struct resv_map *resv_map = vma_resv_map(vma);
274
275                 down_read(&resv_map->rw_sema);
276         }
277 }
278
279 void hugetlb_vma_unlock_read(struct vm_area_struct *vma)
280 {
281         if (__vma_shareable_lock(vma)) {
282                 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
283
284                 up_read(&vma_lock->rw_sema);
285         } else if (__vma_private_lock(vma)) {
286                 struct resv_map *resv_map = vma_resv_map(vma);
287
288                 up_read(&resv_map->rw_sema);
289         }
290 }
291
292 void hugetlb_vma_lock_write(struct vm_area_struct *vma)
293 {
294         if (__vma_shareable_lock(vma)) {
295                 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
296
297                 down_write(&vma_lock->rw_sema);
298         } else if (__vma_private_lock(vma)) {
299                 struct resv_map *resv_map = vma_resv_map(vma);
300
301                 down_write(&resv_map->rw_sema);
302         }
303 }
304
305 void hugetlb_vma_unlock_write(struct vm_area_struct *vma)
306 {
307         if (__vma_shareable_lock(vma)) {
308                 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
309
310                 up_write(&vma_lock->rw_sema);
311         } else if (__vma_private_lock(vma)) {
312                 struct resv_map *resv_map = vma_resv_map(vma);
313
314                 up_write(&resv_map->rw_sema);
315         }
316 }
317
318 int hugetlb_vma_trylock_write(struct vm_area_struct *vma)
319 {
320
321         if (__vma_shareable_lock(vma)) {
322                 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
323
324                 return down_write_trylock(&vma_lock->rw_sema);
325         } else if (__vma_private_lock(vma)) {
326                 struct resv_map *resv_map = vma_resv_map(vma);
327
328                 return down_write_trylock(&resv_map->rw_sema);
329         }
330
331         return 1;
332 }
333
334 void hugetlb_vma_assert_locked(struct vm_area_struct *vma)
335 {
336         if (__vma_shareable_lock(vma)) {
337                 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
338
339                 lockdep_assert_held(&vma_lock->rw_sema);
340         } else if (__vma_private_lock(vma)) {
341                 struct resv_map *resv_map = vma_resv_map(vma);
342
343                 lockdep_assert_held(&resv_map->rw_sema);
344         }
345 }
346
347 void hugetlb_vma_lock_release(struct kref *kref)
348 {
349         struct hugetlb_vma_lock *vma_lock = container_of(kref,
350                         struct hugetlb_vma_lock, refs);
351
352         kfree(vma_lock);
353 }
354
355 static void __hugetlb_vma_unlock_write_put(struct hugetlb_vma_lock *vma_lock)
356 {
357         struct vm_area_struct *vma = vma_lock->vma;
358
359         /*
360          * vma_lock structure may or not be released as a result of put,
361          * it certainly will no longer be attached to vma so clear pointer.
362          * Semaphore synchronizes access to vma_lock->vma field.
363          */
364         vma_lock->vma = NULL;
365         vma->vm_private_data = NULL;
366         up_write(&vma_lock->rw_sema);
367         kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
368 }
369
370 static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma)
371 {
372         if (__vma_shareable_lock(vma)) {
373                 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
374
375                 __hugetlb_vma_unlock_write_put(vma_lock);
376         } else if (__vma_private_lock(vma)) {
377                 struct resv_map *resv_map = vma_resv_map(vma);
378
379                 /* no free for anon vmas, but still need to unlock */
380                 up_write(&resv_map->rw_sema);
381         }
382 }
383
384 static void hugetlb_vma_lock_free(struct vm_area_struct *vma)
385 {
386         /*
387          * Only present in sharable vmas.
388          */
389         if (!vma || !__vma_shareable_lock(vma))
390                 return;
391
392         if (vma->vm_private_data) {
393                 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
394
395                 down_write(&vma_lock->rw_sema);
396                 __hugetlb_vma_unlock_write_put(vma_lock);
397         }
398 }
399
400 static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma)
401 {
402         struct hugetlb_vma_lock *vma_lock;
403
404         /* Only establish in (flags) sharable vmas */
405         if (!vma || !(vma->vm_flags & VM_MAYSHARE))
406                 return;
407
408         /* Should never get here with non-NULL vm_private_data */
409         if (vma->vm_private_data)
410                 return;
411
412         vma_lock = kmalloc(sizeof(*vma_lock), GFP_KERNEL);
413         if (!vma_lock) {
414                 /*
415                  * If we can not allocate structure, then vma can not
416                  * participate in pmd sharing.  This is only a possible
417                  * performance enhancement and memory saving issue.
418                  * However, the lock is also used to synchronize page
419                  * faults with truncation.  If the lock is not present,
420                  * unlikely races could leave pages in a file past i_size
421                  * until the file is removed.  Warn in the unlikely case of
422                  * allocation failure.
423                  */
424                 pr_warn_once("HugeTLB: unable to allocate vma specific lock\n");
425                 return;
426         }
427
428         kref_init(&vma_lock->refs);
429         init_rwsem(&vma_lock->rw_sema);
430         vma_lock->vma = vma;
431         vma->vm_private_data = vma_lock;
432 }
433
434 /* Helper that removes a struct file_region from the resv_map cache and returns
435  * it for use.
436  */
437 static struct file_region *
438 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
439 {
440         struct file_region *nrg;
441
442         VM_BUG_ON(resv->region_cache_count <= 0);
443
444         resv->region_cache_count--;
445         nrg = list_first_entry(&resv->region_cache, struct file_region, link);
446         list_del(&nrg->link);
447
448         nrg->from = from;
449         nrg->to = to;
450
451         return nrg;
452 }
453
454 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
455                                               struct file_region *rg)
456 {
457 #ifdef CONFIG_CGROUP_HUGETLB
458         nrg->reservation_counter = rg->reservation_counter;
459         nrg->css = rg->css;
460         if (rg->css)
461                 css_get(rg->css);
462 #endif
463 }
464
465 /* Helper that records hugetlb_cgroup uncharge info. */
466 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
467                                                 struct hstate *h,
468                                                 struct resv_map *resv,
469                                                 struct file_region *nrg)
470 {
471 #ifdef CONFIG_CGROUP_HUGETLB
472         if (h_cg) {
473                 nrg->reservation_counter =
474                         &h_cg->rsvd_hugepage[hstate_index(h)];
475                 nrg->css = &h_cg->css;
476                 /*
477                  * The caller will hold exactly one h_cg->css reference for the
478                  * whole contiguous reservation region. But this area might be
479                  * scattered when there are already some file_regions reside in
480                  * it. As a result, many file_regions may share only one css
481                  * reference. In order to ensure that one file_region must hold
482                  * exactly one h_cg->css reference, we should do css_get for
483                  * each file_region and leave the reference held by caller
484                  * untouched.
485                  */
486                 css_get(&h_cg->css);
487                 if (!resv->pages_per_hpage)
488                         resv->pages_per_hpage = pages_per_huge_page(h);
489                 /* pages_per_hpage should be the same for all entries in
490                  * a resv_map.
491                  */
492                 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
493         } else {
494                 nrg->reservation_counter = NULL;
495                 nrg->css = NULL;
496         }
497 #endif
498 }
499
500 static void put_uncharge_info(struct file_region *rg)
501 {
502 #ifdef CONFIG_CGROUP_HUGETLB
503         if (rg->css)
504                 css_put(rg->css);
505 #endif
506 }
507
508 static bool has_same_uncharge_info(struct file_region *rg,
509                                    struct file_region *org)
510 {
511 #ifdef CONFIG_CGROUP_HUGETLB
512         return rg->reservation_counter == org->reservation_counter &&
513                rg->css == org->css;
514
515 #else
516         return true;
517 #endif
518 }
519
520 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
521 {
522         struct file_region *nrg, *prg;
523
524         prg = list_prev_entry(rg, link);
525         if (&prg->link != &resv->regions && prg->to == rg->from &&
526             has_same_uncharge_info(prg, rg)) {
527                 prg->to = rg->to;
528
529                 list_del(&rg->link);
530                 put_uncharge_info(rg);
531                 kfree(rg);
532
533                 rg = prg;
534         }
535
536         nrg = list_next_entry(rg, link);
537         if (&nrg->link != &resv->regions && nrg->from == rg->to &&
538             has_same_uncharge_info(nrg, rg)) {
539                 nrg->from = rg->from;
540
541                 list_del(&rg->link);
542                 put_uncharge_info(rg);
543                 kfree(rg);
544         }
545 }
546
547 static inline long
548 hugetlb_resv_map_add(struct resv_map *map, struct list_head *rg, long from,
549                      long to, struct hstate *h, struct hugetlb_cgroup *cg,
550                      long *regions_needed)
551 {
552         struct file_region *nrg;
553
554         if (!regions_needed) {
555                 nrg = get_file_region_entry_from_cache(map, from, to);
556                 record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);
557                 list_add(&nrg->link, rg);
558                 coalesce_file_region(map, nrg);
559         } else
560                 *regions_needed += 1;
561
562         return to - from;
563 }
564
565 /*
566  * Must be called with resv->lock held.
567  *
568  * Calling this with regions_needed != NULL will count the number of pages
569  * to be added but will not modify the linked list. And regions_needed will
570  * indicate the number of file_regions needed in the cache to carry out to add
571  * the regions for this range.
572  */
573 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
574                                      struct hugetlb_cgroup *h_cg,
575                                      struct hstate *h, long *regions_needed)
576 {
577         long add = 0;
578         struct list_head *head = &resv->regions;
579         long last_accounted_offset = f;
580         struct file_region *iter, *trg = NULL;
581         struct list_head *rg = NULL;
582
583         if (regions_needed)
584                 *regions_needed = 0;
585
586         /* In this loop, we essentially handle an entry for the range
587          * [last_accounted_offset, iter->from), at every iteration, with some
588          * bounds checking.
589          */
590         list_for_each_entry_safe(iter, trg, head, link) {
591                 /* Skip irrelevant regions that start before our range. */
592                 if (iter->from < f) {
593                         /* If this region ends after the last accounted offset,
594                          * then we need to update last_accounted_offset.
595                          */
596                         if (iter->to > last_accounted_offset)
597                                 last_accounted_offset = iter->to;
598                         continue;
599                 }
600
601                 /* When we find a region that starts beyond our range, we've
602                  * finished.
603                  */
604                 if (iter->from >= t) {
605                         rg = iter->link.prev;
606                         break;
607                 }
608
609                 /* Add an entry for last_accounted_offset -> iter->from, and
610                  * update last_accounted_offset.
611                  */
612                 if (iter->from > last_accounted_offset)
613                         add += hugetlb_resv_map_add(resv, iter->link.prev,
614                                                     last_accounted_offset,
615                                                     iter->from, h, h_cg,
616                                                     regions_needed);
617
618                 last_accounted_offset = iter->to;
619         }
620
621         /* Handle the case where our range extends beyond
622          * last_accounted_offset.
623          */
624         if (!rg)
625                 rg = head->prev;
626         if (last_accounted_offset < t)
627                 add += hugetlb_resv_map_add(resv, rg, last_accounted_offset,
628                                             t, h, h_cg, regions_needed);
629
630         return add;
631 }
632
633 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
634  */
635 static int allocate_file_region_entries(struct resv_map *resv,
636                                         int regions_needed)
637         __must_hold(&resv->lock)
638 {
639         LIST_HEAD(allocated_regions);
640         int to_allocate = 0, i = 0;
641         struct file_region *trg = NULL, *rg = NULL;
642
643         VM_BUG_ON(regions_needed < 0);
644
645         /*
646          * Check for sufficient descriptors in the cache to accommodate
647          * the number of in progress add operations plus regions_needed.
648          *
649          * This is a while loop because when we drop the lock, some other call
650          * to region_add or region_del may have consumed some region_entries,
651          * so we keep looping here until we finally have enough entries for
652          * (adds_in_progress + regions_needed).
653          */
654         while (resv->region_cache_count <
655                (resv->adds_in_progress + regions_needed)) {
656                 to_allocate = resv->adds_in_progress + regions_needed -
657                               resv->region_cache_count;
658
659                 /* At this point, we should have enough entries in the cache
660                  * for all the existing adds_in_progress. We should only be
661                  * needing to allocate for regions_needed.
662                  */
663                 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
664
665                 spin_unlock(&resv->lock);
666                 for (i = 0; i < to_allocate; i++) {
667                         trg = kmalloc(sizeof(*trg), GFP_KERNEL);
668                         if (!trg)
669                                 goto out_of_memory;
670                         list_add(&trg->link, &allocated_regions);
671                 }
672
673                 spin_lock(&resv->lock);
674
675                 list_splice(&allocated_regions, &resv->region_cache);
676                 resv->region_cache_count += to_allocate;
677         }
678
679         return 0;
680
681 out_of_memory:
682         list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
683                 list_del(&rg->link);
684                 kfree(rg);
685         }
686         return -ENOMEM;
687 }
688
689 /*
690  * Add the huge page range represented by [f, t) to the reserve
691  * map.  Regions will be taken from the cache to fill in this range.
692  * Sufficient regions should exist in the cache due to the previous
693  * call to region_chg with the same range, but in some cases the cache will not
694  * have sufficient entries due to races with other code doing region_add or
695  * region_del.  The extra needed entries will be allocated.
696  *
697  * regions_needed is the out value provided by a previous call to region_chg.
698  *
699  * Return the number of new huge pages added to the map.  This number is greater
700  * than or equal to zero.  If file_region entries needed to be allocated for
701  * this operation and we were not able to allocate, it returns -ENOMEM.
702  * region_add of regions of length 1 never allocate file_regions and cannot
703  * fail; region_chg will always allocate at least 1 entry and a region_add for
704  * 1 page will only require at most 1 entry.
705  */
706 static long region_add(struct resv_map *resv, long f, long t,
707                        long in_regions_needed, struct hstate *h,
708                        struct hugetlb_cgroup *h_cg)
709 {
710         long add = 0, actual_regions_needed = 0;
711
712         spin_lock(&resv->lock);
713 retry:
714
715         /* Count how many regions are actually needed to execute this add. */
716         add_reservation_in_range(resv, f, t, NULL, NULL,
717                                  &actual_regions_needed);
718
719         /*
720          * Check for sufficient descriptors in the cache to accommodate
721          * this add operation. Note that actual_regions_needed may be greater
722          * than in_regions_needed, as the resv_map may have been modified since
723          * the region_chg call. In this case, we need to make sure that we
724          * allocate extra entries, such that we have enough for all the
725          * existing adds_in_progress, plus the excess needed for this
726          * operation.
727          */
728         if (actual_regions_needed > in_regions_needed &&
729             resv->region_cache_count <
730                     resv->adds_in_progress +
731                             (actual_regions_needed - in_regions_needed)) {
732                 /* region_add operation of range 1 should never need to
733                  * allocate file_region entries.
734                  */
735                 VM_BUG_ON(t - f <= 1);
736
737                 if (allocate_file_region_entries(
738                             resv, actual_regions_needed - in_regions_needed)) {
739                         return -ENOMEM;
740                 }
741
742                 goto retry;
743         }
744
745         add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
746
747         resv->adds_in_progress -= in_regions_needed;
748
749         spin_unlock(&resv->lock);
750         return add;
751 }
752
753 /*
754  * Examine the existing reserve map and determine how many
755  * huge pages in the specified range [f, t) are NOT currently
756  * represented.  This routine is called before a subsequent
757  * call to region_add that will actually modify the reserve
758  * map to add the specified range [f, t).  region_chg does
759  * not change the number of huge pages represented by the
760  * map.  A number of new file_region structures is added to the cache as a
761  * placeholder, for the subsequent region_add call to use. At least 1
762  * file_region structure is added.
763  *
764  * out_regions_needed is the number of regions added to the
765  * resv->adds_in_progress.  This value needs to be provided to a follow up call
766  * to region_add or region_abort for proper accounting.
767  *
768  * Returns the number of huge pages that need to be added to the existing
769  * reservation map for the range [f, t).  This number is greater or equal to
770  * zero.  -ENOMEM is returned if a new file_region structure or cache entry
771  * is needed and can not be allocated.
772  */
773 static long region_chg(struct resv_map *resv, long f, long t,
774                        long *out_regions_needed)
775 {
776         long chg = 0;
777
778         spin_lock(&resv->lock);
779
780         /* Count how many hugepages in this range are NOT represented. */
781         chg = add_reservation_in_range(resv, f, t, NULL, NULL,
782                                        out_regions_needed);
783
784         if (*out_regions_needed == 0)
785                 *out_regions_needed = 1;
786
787         if (allocate_file_region_entries(resv, *out_regions_needed))
788                 return -ENOMEM;
789
790         resv->adds_in_progress += *out_regions_needed;
791
792         spin_unlock(&resv->lock);
793         return chg;
794 }
795
796 /*
797  * Abort the in progress add operation.  The adds_in_progress field
798  * of the resv_map keeps track of the operations in progress between
799  * calls to region_chg and region_add.  Operations are sometimes
800  * aborted after the call to region_chg.  In such cases, region_abort
801  * is called to decrement the adds_in_progress counter. regions_needed
802  * is the value returned by the region_chg call, it is used to decrement
803  * the adds_in_progress counter.
804  *
805  * NOTE: The range arguments [f, t) are not needed or used in this
806  * routine.  They are kept to make reading the calling code easier as
807  * arguments will match the associated region_chg call.
808  */
809 static void region_abort(struct resv_map *resv, long f, long t,
810                          long regions_needed)
811 {
812         spin_lock(&resv->lock);
813         VM_BUG_ON(!resv->region_cache_count);
814         resv->adds_in_progress -= regions_needed;
815         spin_unlock(&resv->lock);
816 }
817
818 /*
819  * Delete the specified range [f, t) from the reserve map.  If the
820  * t parameter is LONG_MAX, this indicates that ALL regions after f
821  * should be deleted.  Locate the regions which intersect [f, t)
822  * and either trim, delete or split the existing regions.
823  *
824  * Returns the number of huge pages deleted from the reserve map.
825  * In the normal case, the return value is zero or more.  In the
826  * case where a region must be split, a new region descriptor must
827  * be allocated.  If the allocation fails, -ENOMEM will be returned.
828  * NOTE: If the parameter t == LONG_MAX, then we will never split
829  * a region and possibly return -ENOMEM.  Callers specifying
830  * t == LONG_MAX do not need to check for -ENOMEM error.
831  */
832 static long region_del(struct resv_map *resv, long f, long t)
833 {
834         struct list_head *head = &resv->regions;
835         struct file_region *rg, *trg;
836         struct file_region *nrg = NULL;
837         long del = 0;
838
839 retry:
840         spin_lock(&resv->lock);
841         list_for_each_entry_safe(rg, trg, head, link) {
842                 /*
843                  * Skip regions before the range to be deleted.  file_region
844                  * ranges are normally of the form [from, to).  However, there
845                  * may be a "placeholder" entry in the map which is of the form
846                  * (from, to) with from == to.  Check for placeholder entries
847                  * at the beginning of the range to be deleted.
848                  */
849                 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
850                         continue;
851
852                 if (rg->from >= t)
853                         break;
854
855                 if (f > rg->from && t < rg->to) { /* Must split region */
856                         /*
857                          * Check for an entry in the cache before dropping
858                          * lock and attempting allocation.
859                          */
860                         if (!nrg &&
861                             resv->region_cache_count > resv->adds_in_progress) {
862                                 nrg = list_first_entry(&resv->region_cache,
863                                                         struct file_region,
864                                                         link);
865                                 list_del(&nrg->link);
866                                 resv->region_cache_count--;
867                         }
868
869                         if (!nrg) {
870                                 spin_unlock(&resv->lock);
871                                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
872                                 if (!nrg)
873                                         return -ENOMEM;
874                                 goto retry;
875                         }
876
877                         del += t - f;
878                         hugetlb_cgroup_uncharge_file_region(
879                                 resv, rg, t - f, false);
880
881                         /* New entry for end of split region */
882                         nrg->from = t;
883                         nrg->to = rg->to;
884
885                         copy_hugetlb_cgroup_uncharge_info(nrg, rg);
886
887                         INIT_LIST_HEAD(&nrg->link);
888
889                         /* Original entry is trimmed */
890                         rg->to = f;
891
892                         list_add(&nrg->link, &rg->link);
893                         nrg = NULL;
894                         break;
895                 }
896
897                 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
898                         del += rg->to - rg->from;
899                         hugetlb_cgroup_uncharge_file_region(resv, rg,
900                                                             rg->to - rg->from, true);
901                         list_del(&rg->link);
902                         kfree(rg);
903                         continue;
904                 }
905
906                 if (f <= rg->from) {    /* Trim beginning of region */
907                         hugetlb_cgroup_uncharge_file_region(resv, rg,
908                                                             t - rg->from, false);
909
910                         del += t - rg->from;
911                         rg->from = t;
912                 } else {                /* Trim end of region */
913                         hugetlb_cgroup_uncharge_file_region(resv, rg,
914                                                             rg->to - f, false);
915
916                         del += rg->to - f;
917                         rg->to = f;
918                 }
919         }
920
921         spin_unlock(&resv->lock);
922         kfree(nrg);
923         return del;
924 }
925
926 /*
927  * A rare out of memory error was encountered which prevented removal of
928  * the reserve map region for a page.  The huge page itself was free'ed
929  * and removed from the page cache.  This routine will adjust the subpool
930  * usage count, and the global reserve count if needed.  By incrementing
931  * these counts, the reserve map entry which could not be deleted will
932  * appear as a "reserved" entry instead of simply dangling with incorrect
933  * counts.
934  */
935 void hugetlb_fix_reserve_counts(struct inode *inode)
936 {
937         struct hugepage_subpool *spool = subpool_inode(inode);
938         long rsv_adjust;
939         bool reserved = false;
940
941         rsv_adjust = hugepage_subpool_get_pages(spool, 1);
942         if (rsv_adjust > 0) {
943                 struct hstate *h = hstate_inode(inode);
944
945                 if (!hugetlb_acct_memory(h, 1))
946                         reserved = true;
947         } else if (!rsv_adjust) {
948                 reserved = true;
949         }
950
951         if (!reserved)
952                 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
953 }
954
955 /*
956  * Count and return the number of huge pages in the reserve map
957  * that intersect with the range [f, t).
958  */
959 static long region_count(struct resv_map *resv, long f, long t)
960 {
961         struct list_head *head = &resv->regions;
962         struct file_region *rg;
963         long chg = 0;
964
965         spin_lock(&resv->lock);
966         /* Locate each segment we overlap with, and count that overlap. */
967         list_for_each_entry(rg, head, link) {
968                 long seg_from;
969                 long seg_to;
970
971                 if (rg->to <= f)
972                         continue;
973                 if (rg->from >= t)
974                         break;
975
976                 seg_from = max(rg->from, f);
977                 seg_to = min(rg->to, t);
978
979                 chg += seg_to - seg_from;
980         }
981         spin_unlock(&resv->lock);
982
983         return chg;
984 }
985
986 /*
987  * Convert the address within this vma to the page offset within
988  * the mapping, huge page units here.
989  */
990 static pgoff_t vma_hugecache_offset(struct hstate *h,
991                         struct vm_area_struct *vma, unsigned long address)
992 {
993         return ((address - vma->vm_start) >> huge_page_shift(h)) +
994                         (vma->vm_pgoff >> huge_page_order(h));
995 }
996
997 /**
998  * vma_kernel_pagesize - Page size granularity for this VMA.
999  * @vma: The user mapping.
1000  *
1001  * Folios in this VMA will be aligned to, and at least the size of the
1002  * number of bytes returned by this function.
1003  *
1004  * Return: The default size of the folios allocated when backing a VMA.
1005  */
1006 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
1007 {
1008         if (vma->vm_ops && vma->vm_ops->pagesize)
1009                 return vma->vm_ops->pagesize(vma);
1010         return PAGE_SIZE;
1011 }
1012 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
1013
1014 /*
1015  * Return the page size being used by the MMU to back a VMA. In the majority
1016  * of cases, the page size used by the kernel matches the MMU size. On
1017  * architectures where it differs, an architecture-specific 'strong'
1018  * version of this symbol is required.
1019  */
1020 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
1021 {
1022         return vma_kernel_pagesize(vma);
1023 }
1024
1025 /*
1026  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
1027  * bits of the reservation map pointer, which are always clear due to
1028  * alignment.
1029  */
1030 #define HPAGE_RESV_OWNER    (1UL << 0)
1031 #define HPAGE_RESV_UNMAPPED (1UL << 1)
1032 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
1033
1034 /*
1035  * These helpers are used to track how many pages are reserved for
1036  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
1037  * is guaranteed to have their future faults succeed.
1038  *
1039  * With the exception of hugetlb_dup_vma_private() which is called at fork(),
1040  * the reserve counters are updated with the hugetlb_lock held. It is safe
1041  * to reset the VMA at fork() time as it is not in use yet and there is no
1042  * chance of the global counters getting corrupted as a result of the values.
1043  *
1044  * The private mapping reservation is represented in a subtly different
1045  * manner to a shared mapping.  A shared mapping has a region map associated
1046  * with the underlying file, this region map represents the backing file
1047  * pages which have ever had a reservation assigned which this persists even
1048  * after the page is instantiated.  A private mapping has a region map
1049  * associated with the original mmap which is attached to all VMAs which
1050  * reference it, this region map represents those offsets which have consumed
1051  * reservation ie. where pages have been instantiated.
1052  */
1053 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
1054 {
1055         return (unsigned long)vma->vm_private_data;
1056 }
1057
1058 static void set_vma_private_data(struct vm_area_struct *vma,
1059                                                         unsigned long value)
1060 {
1061         vma->vm_private_data = (void *)value;
1062 }
1063
1064 static void
1065 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
1066                                           struct hugetlb_cgroup *h_cg,
1067                                           struct hstate *h)
1068 {
1069 #ifdef CONFIG_CGROUP_HUGETLB
1070         if (!h_cg || !h) {
1071                 resv_map->reservation_counter = NULL;
1072                 resv_map->pages_per_hpage = 0;
1073                 resv_map->css = NULL;
1074         } else {
1075                 resv_map->reservation_counter =
1076                         &h_cg->rsvd_hugepage[hstate_index(h)];
1077                 resv_map->pages_per_hpage = pages_per_huge_page(h);
1078                 resv_map->css = &h_cg->css;
1079         }
1080 #endif
1081 }
1082
1083 struct resv_map *resv_map_alloc(void)
1084 {
1085         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
1086         struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
1087
1088         if (!resv_map || !rg) {
1089                 kfree(resv_map);
1090                 kfree(rg);
1091                 return NULL;
1092         }
1093
1094         kref_init(&resv_map->refs);
1095         spin_lock_init(&resv_map->lock);
1096         INIT_LIST_HEAD(&resv_map->regions);
1097         init_rwsem(&resv_map->rw_sema);
1098
1099         resv_map->adds_in_progress = 0;
1100         /*
1101          * Initialize these to 0. On shared mappings, 0's here indicate these
1102          * fields don't do cgroup accounting. On private mappings, these will be
1103          * re-initialized to the proper values, to indicate that hugetlb cgroup
1104          * reservations are to be un-charged from here.
1105          */
1106         resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
1107
1108         INIT_LIST_HEAD(&resv_map->region_cache);
1109         list_add(&rg->link, &resv_map->region_cache);
1110         resv_map->region_cache_count = 1;
1111
1112         return resv_map;
1113 }
1114
1115 void resv_map_release(struct kref *ref)
1116 {
1117         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
1118         struct list_head *head = &resv_map->region_cache;
1119         struct file_region *rg, *trg;
1120
1121         /* Clear out any active regions before we release the map. */
1122         region_del(resv_map, 0, LONG_MAX);
1123
1124         /* ... and any entries left in the cache */
1125         list_for_each_entry_safe(rg, trg, head, link) {
1126                 list_del(&rg->link);
1127                 kfree(rg);
1128         }
1129
1130         VM_BUG_ON(resv_map->adds_in_progress);
1131
1132         kfree(resv_map);
1133 }
1134
1135 static inline struct resv_map *inode_resv_map(struct inode *inode)
1136 {
1137         /*
1138          * At inode evict time, i_mapping may not point to the original
1139          * address space within the inode.  This original address space
1140          * contains the pointer to the resv_map.  So, always use the
1141          * address space embedded within the inode.
1142          * The VERY common case is inode->mapping == &inode->i_data but,
1143          * this may not be true for device special inodes.
1144          */
1145         return (struct resv_map *)(&inode->i_data)->i_private_data;
1146 }
1147
1148 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
1149 {
1150         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1151         if (vma->vm_flags & VM_MAYSHARE) {
1152                 struct address_space *mapping = vma->vm_file->f_mapping;
1153                 struct inode *inode = mapping->host;
1154
1155                 return inode_resv_map(inode);
1156
1157         } else {
1158                 return (struct resv_map *)(get_vma_private_data(vma) &
1159                                                         ~HPAGE_RESV_MASK);
1160         }
1161 }
1162
1163 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
1164 {
1165         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1166         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
1167
1168         set_vma_private_data(vma, (unsigned long)map);
1169 }
1170
1171 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
1172 {
1173         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1174         VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
1175
1176         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
1177 }
1178
1179 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
1180 {
1181         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1182
1183         return (get_vma_private_data(vma) & flag) != 0;
1184 }
1185
1186 bool __vma_private_lock(struct vm_area_struct *vma)
1187 {
1188         return !(vma->vm_flags & VM_MAYSHARE) &&
1189                 get_vma_private_data(vma) & ~HPAGE_RESV_MASK &&
1190                 is_vma_resv_set(vma, HPAGE_RESV_OWNER);
1191 }
1192
1193 void hugetlb_dup_vma_private(struct vm_area_struct *vma)
1194 {
1195         VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1196         /*
1197          * Clear vm_private_data
1198          * - For shared mappings this is a per-vma semaphore that may be
1199          *   allocated in a subsequent call to hugetlb_vm_op_open.
1200          *   Before clearing, make sure pointer is not associated with vma
1201          *   as this will leak the structure.  This is the case when called
1202          *   via clear_vma_resv_huge_pages() and hugetlb_vm_op_open has already
1203          *   been called to allocate a new structure.
1204          * - For MAP_PRIVATE mappings, this is the reserve map which does
1205          *   not apply to children.  Faults generated by the children are
1206          *   not guaranteed to succeed, even if read-only.
1207          */
1208         if (vma->vm_flags & VM_MAYSHARE) {
1209                 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
1210
1211                 if (vma_lock && vma_lock->vma != vma)
1212                         vma->vm_private_data = NULL;
1213         } else
1214                 vma->vm_private_data = NULL;
1215 }
1216
1217 /*
1218  * Reset and decrement one ref on hugepage private reservation.
1219  * Called with mm->mmap_lock writer semaphore held.
1220  * This function should be only used by move_vma() and operate on
1221  * same sized vma. It should never come here with last ref on the
1222  * reservation.
1223  */
1224 void clear_vma_resv_huge_pages(struct vm_area_struct *vma)
1225 {
1226         /*
1227          * Clear the old hugetlb private page reservation.
1228          * It has already been transferred to new_vma.
1229          *
1230          * During a mremap() operation of a hugetlb vma we call move_vma()
1231          * which copies vma into new_vma and unmaps vma. After the copy
1232          * operation both new_vma and vma share a reference to the resv_map
1233          * struct, and at that point vma is about to be unmapped. We don't
1234          * want to return the reservation to the pool at unmap of vma because
1235          * the reservation still lives on in new_vma, so simply decrement the
1236          * ref here and remove the resv_map reference from this vma.
1237          */
1238         struct resv_map *reservations = vma_resv_map(vma);
1239
1240         if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1241                 resv_map_put_hugetlb_cgroup_uncharge_info(reservations);
1242                 kref_put(&reservations->refs, resv_map_release);
1243         }
1244
1245         hugetlb_dup_vma_private(vma);
1246 }
1247
1248 /* Returns true if the VMA has associated reserve pages */
1249 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
1250 {
1251         if (vma->vm_flags & VM_NORESERVE) {
1252                 /*
1253                  * This address is already reserved by other process(chg == 0),
1254                  * so, we should decrement reserved count. Without decrementing,
1255                  * reserve count remains after releasing inode, because this
1256                  * allocated page will go into page cache and is regarded as
1257                  * coming from reserved pool in releasing step.  Currently, we
1258                  * don't have any other solution to deal with this situation
1259                  * properly, so add work-around here.
1260                  */
1261                 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1262                         return true;
1263                 else
1264                         return false;
1265         }
1266
1267         /* Shared mappings always use reserves */
1268         if (vma->vm_flags & VM_MAYSHARE) {
1269                 /*
1270                  * We know VM_NORESERVE is not set.  Therefore, there SHOULD
1271                  * be a region map for all pages.  The only situation where
1272                  * there is no region map is if a hole was punched via
1273                  * fallocate.  In this case, there really are no reserves to
1274                  * use.  This situation is indicated if chg != 0.
1275                  */
1276                 if (chg)
1277                         return false;
1278                 else
1279                         return true;
1280         }
1281
1282         /*
1283          * Only the process that called mmap() has reserves for
1284          * private mappings.
1285          */
1286         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1287                 /*
1288                  * Like the shared case above, a hole punch or truncate
1289                  * could have been performed on the private mapping.
1290                  * Examine the value of chg to determine if reserves
1291                  * actually exist or were previously consumed.
1292                  * Very Subtle - The value of chg comes from a previous
1293                  * call to vma_needs_reserves().  The reserve map for
1294                  * private mappings has different (opposite) semantics
1295                  * than that of shared mappings.  vma_needs_reserves()
1296                  * has already taken this difference in semantics into
1297                  * account.  Therefore, the meaning of chg is the same
1298                  * as in the shared case above.  Code could easily be
1299                  * combined, but keeping it separate draws attention to
1300                  * subtle differences.
1301                  */
1302                 if (chg)
1303                         return false;
1304                 else
1305                         return true;
1306         }
1307
1308         return false;
1309 }
1310
1311 static void enqueue_hugetlb_folio(struct hstate *h, struct folio *folio)
1312 {
1313         int nid = folio_nid(folio);
1314
1315         lockdep_assert_held(&hugetlb_lock);
1316         VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
1317
1318         list_move(&folio->lru, &h->hugepage_freelists[nid]);
1319         h->free_huge_pages++;
1320         h->free_huge_pages_node[nid]++;
1321         folio_set_hugetlb_freed(folio);
1322 }
1323
1324 static struct folio *dequeue_hugetlb_folio_node_exact(struct hstate *h,
1325                                                                 int nid)
1326 {
1327         struct folio *folio;
1328         bool pin = !!(current->flags & PF_MEMALLOC_PIN);
1329
1330         lockdep_assert_held(&hugetlb_lock);
1331         list_for_each_entry(folio, &h->hugepage_freelists[nid], lru) {
1332                 if (pin && !folio_is_longterm_pinnable(folio))
1333                         continue;
1334
1335                 if (folio_test_hwpoison(folio))
1336                         continue;
1337
1338                 list_move(&folio->lru, &h->hugepage_activelist);
1339                 folio_ref_unfreeze(folio, 1);
1340                 folio_clear_hugetlb_freed(folio);
1341                 h->free_huge_pages--;
1342                 h->free_huge_pages_node[nid]--;
1343                 return folio;
1344         }
1345
1346         return NULL;
1347 }
1348
1349 static struct folio *dequeue_hugetlb_folio_nodemask(struct hstate *h, gfp_t gfp_mask,
1350                                                         int nid, nodemask_t *nmask)
1351 {
1352         unsigned int cpuset_mems_cookie;
1353         struct zonelist *zonelist;
1354         struct zone *zone;
1355         struct zoneref *z;
1356         int node = NUMA_NO_NODE;
1357
1358         zonelist = node_zonelist(nid, gfp_mask);
1359
1360 retry_cpuset:
1361         cpuset_mems_cookie = read_mems_allowed_begin();
1362         for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1363                 struct folio *folio;
1364
1365                 if (!cpuset_zone_allowed(zone, gfp_mask))
1366                         continue;
1367                 /*
1368                  * no need to ask again on the same node. Pool is node rather than
1369                  * zone aware
1370                  */
1371                 if (zone_to_nid(zone) == node)
1372                         continue;
1373                 node = zone_to_nid(zone);
1374
1375                 folio = dequeue_hugetlb_folio_node_exact(h, node);
1376                 if (folio)
1377                         return folio;
1378         }
1379         if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1380                 goto retry_cpuset;
1381
1382         return NULL;
1383 }
1384
1385 static unsigned long available_huge_pages(struct hstate *h)
1386 {
1387         return h->free_huge_pages - h->resv_huge_pages;
1388 }
1389
1390 static struct folio *dequeue_hugetlb_folio_vma(struct hstate *h,
1391                                 struct vm_area_struct *vma,
1392                                 unsigned long address, int avoid_reserve,
1393                                 long chg)
1394 {
1395         struct folio *folio = NULL;
1396         struct mempolicy *mpol;
1397         gfp_t gfp_mask;
1398         nodemask_t *nodemask;
1399         int nid;
1400
1401         /*
1402          * A child process with MAP_PRIVATE mappings created by their parent
1403          * have no page reserves. This check ensures that reservations are
1404          * not "stolen". The child may still get SIGKILLed
1405          */
1406         if (!vma_has_reserves(vma, chg) && !available_huge_pages(h))
1407                 goto err;
1408
1409         /* If reserves cannot be used, ensure enough pages are in the pool */
1410         if (avoid_reserve && !available_huge_pages(h))
1411                 goto err;
1412
1413         gfp_mask = htlb_alloc_mask(h);
1414         nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1415
1416         if (mpol_is_preferred_many(mpol)) {
1417                 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
1418                                                         nid, nodemask);
1419
1420                 /* Fallback to all nodes if page==NULL */
1421                 nodemask = NULL;
1422         }
1423
1424         if (!folio)
1425                 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
1426                                                         nid, nodemask);
1427
1428         if (folio && !avoid_reserve && vma_has_reserves(vma, chg)) {
1429                 folio_set_hugetlb_restore_reserve(folio);
1430                 h->resv_huge_pages--;
1431         }
1432
1433         mpol_cond_put(mpol);
1434         return folio;
1435
1436 err:
1437         return NULL;
1438 }
1439
1440 /*
1441  * common helper functions for hstate_next_node_to_{alloc|free}.
1442  * We may have allocated or freed a huge page based on a different
1443  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1444  * be outside of *nodes_allowed.  Ensure that we use an allowed
1445  * node for alloc or free.
1446  */
1447 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1448 {
1449         nid = next_node_in(nid, *nodes_allowed);
1450         VM_BUG_ON(nid >= MAX_NUMNODES);
1451
1452         return nid;
1453 }
1454
1455 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1456 {
1457         if (!node_isset(nid, *nodes_allowed))
1458                 nid = next_node_allowed(nid, nodes_allowed);
1459         return nid;
1460 }
1461
1462 /*
1463  * returns the previously saved node ["this node"] from which to
1464  * allocate a persistent huge page for the pool and advance the
1465  * next node from which to allocate, handling wrap at end of node
1466  * mask.
1467  */
1468 static int hstate_next_node_to_alloc(int *next_node,
1469                                         nodemask_t *nodes_allowed)
1470 {
1471         int nid;
1472
1473         VM_BUG_ON(!nodes_allowed);
1474
1475         nid = get_valid_node_allowed(*next_node, nodes_allowed);
1476         *next_node = next_node_allowed(nid, nodes_allowed);
1477
1478         return nid;
1479 }
1480
1481 /*
1482  * helper for remove_pool_hugetlb_folio() - return the previously saved
1483  * node ["this node"] from which to free a huge page.  Advance the
1484  * next node id whether or not we find a free huge page to free so
1485  * that the next attempt to free addresses the next node.
1486  */
1487 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1488 {
1489         int nid;
1490
1491         VM_BUG_ON(!nodes_allowed);
1492
1493         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1494         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1495
1496         return nid;
1497 }
1498
1499 #define for_each_node_mask_to_alloc(next_node, nr_nodes, node, mask)            \
1500         for (nr_nodes = nodes_weight(*mask);                            \
1501                 nr_nodes > 0 &&                                         \
1502                 ((node = hstate_next_node_to_alloc(next_node, mask)) || 1);     \
1503                 nr_nodes--)
1504
1505 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask)            \
1506         for (nr_nodes = nodes_weight(*mask);                            \
1507                 nr_nodes > 0 &&                                         \
1508                 ((node = hstate_next_node_to_free(hs, mask)) || 1);     \
1509                 nr_nodes--)
1510
1511 /* used to demote non-gigantic_huge pages as well */
1512 static void __destroy_compound_gigantic_folio(struct folio *folio,
1513                                         unsigned int order, bool demote)
1514 {
1515         int i;
1516         int nr_pages = 1 << order;
1517         struct page *p;
1518
1519         atomic_set(&folio->_entire_mapcount, 0);
1520         atomic_set(&folio->_nr_pages_mapped, 0);
1521         atomic_set(&folio->_pincount, 0);
1522
1523         for (i = 1; i < nr_pages; i++) {
1524                 p = folio_page(folio, i);
1525                 p->flags &= ~PAGE_FLAGS_CHECK_AT_FREE;
1526                 p->mapping = NULL;
1527                 clear_compound_head(p);
1528                 if (!demote)
1529                         set_page_refcounted(p);
1530         }
1531
1532         __folio_clear_head(folio);
1533 }
1534
1535 static void destroy_compound_hugetlb_folio_for_demote(struct folio *folio,
1536                                         unsigned int order)
1537 {
1538         __destroy_compound_gigantic_folio(folio, order, true);
1539 }
1540
1541 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1542 static void destroy_compound_gigantic_folio(struct folio *folio,
1543                                         unsigned int order)
1544 {
1545         __destroy_compound_gigantic_folio(folio, order, false);
1546 }
1547
1548 static void free_gigantic_folio(struct folio *folio, unsigned int order)
1549 {
1550         /*
1551          * If the page isn't allocated using the cma allocator,
1552          * cma_release() returns false.
1553          */
1554 #ifdef CONFIG_CMA
1555         int nid = folio_nid(folio);
1556
1557         if (cma_release(hugetlb_cma[nid], &folio->page, 1 << order))
1558                 return;
1559 #endif
1560
1561         free_contig_range(folio_pfn(folio), 1 << order);
1562 }
1563
1564 #ifdef CONFIG_CONTIG_ALLOC
1565 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1566                 int nid, nodemask_t *nodemask)
1567 {
1568         struct page *page;
1569         unsigned long nr_pages = pages_per_huge_page(h);
1570         if (nid == NUMA_NO_NODE)
1571                 nid = numa_mem_id();
1572
1573 #ifdef CONFIG_CMA
1574         {
1575                 int node;
1576
1577                 if (hugetlb_cma[nid]) {
1578                         page = cma_alloc(hugetlb_cma[nid], nr_pages,
1579                                         huge_page_order(h), true);
1580                         if (page)
1581                                 return page_folio(page);
1582                 }
1583
1584                 if (!(gfp_mask & __GFP_THISNODE)) {
1585                         for_each_node_mask(node, *nodemask) {
1586                                 if (node == nid || !hugetlb_cma[node])
1587                                         continue;
1588
1589                                 page = cma_alloc(hugetlb_cma[node], nr_pages,
1590                                                 huge_page_order(h), true);
1591                                 if (page)
1592                                         return page_folio(page);
1593                         }
1594                 }
1595         }
1596 #endif
1597
1598         page = alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1599         return page ? page_folio(page) : NULL;
1600 }
1601
1602 #else /* !CONFIG_CONTIG_ALLOC */
1603 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1604                                         int nid, nodemask_t *nodemask)
1605 {
1606         return NULL;
1607 }
1608 #endif /* CONFIG_CONTIG_ALLOC */
1609
1610 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1611 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1612                                         int nid, nodemask_t *nodemask)
1613 {
1614         return NULL;
1615 }
1616 static inline void free_gigantic_folio(struct folio *folio,
1617                                                 unsigned int order) { }
1618 static inline void destroy_compound_gigantic_folio(struct folio *folio,
1619                                                 unsigned int order) { }
1620 #endif
1621
1622 static inline void __clear_hugetlb_destructor(struct hstate *h,
1623                                                 struct folio *folio)
1624 {
1625         lockdep_assert_held(&hugetlb_lock);
1626
1627         folio_clear_hugetlb(folio);
1628 }
1629
1630 /*
1631  * Remove hugetlb folio from lists.
1632  * If vmemmap exists for the folio, update dtor so that the folio appears
1633  * as just a compound page.  Otherwise, wait until after allocating vmemmap
1634  * to update dtor.
1635  *
1636  * A reference is held on the folio, except in the case of demote.
1637  *
1638  * Must be called with hugetlb lock held.
1639  */
1640 static void __remove_hugetlb_folio(struct hstate *h, struct folio *folio,
1641                                                         bool adjust_surplus,
1642                                                         bool demote)
1643 {
1644         int nid = folio_nid(folio);
1645
1646         VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio(folio), folio);
1647         VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio_rsvd(folio), folio);
1648
1649         lockdep_assert_held(&hugetlb_lock);
1650         if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1651                 return;
1652
1653         list_del(&folio->lru);
1654
1655         if (folio_test_hugetlb_freed(folio)) {
1656                 h->free_huge_pages--;
1657                 h->free_huge_pages_node[nid]--;
1658         }
1659         if (adjust_surplus) {
1660                 h->surplus_huge_pages--;
1661                 h->surplus_huge_pages_node[nid]--;
1662         }
1663
1664         /*
1665          * We can only clear the hugetlb destructor after allocating vmemmap
1666          * pages.  Otherwise, someone (memory error handling) may try to write
1667          * to tail struct pages.
1668          */
1669         if (!folio_test_hugetlb_vmemmap_optimized(folio))
1670                 __clear_hugetlb_destructor(h, folio);
1671
1672          /*
1673           * In the case of demote we do not ref count the page as it will soon
1674           * be turned into a page of smaller size.
1675          */
1676         if (!demote)
1677                 folio_ref_unfreeze(folio, 1);
1678
1679         h->nr_huge_pages--;
1680         h->nr_huge_pages_node[nid]--;
1681 }
1682
1683 static void remove_hugetlb_folio(struct hstate *h, struct folio *folio,
1684                                                         bool adjust_surplus)
1685 {
1686         __remove_hugetlb_folio(h, folio, adjust_surplus, false);
1687 }
1688
1689 static void remove_hugetlb_folio_for_demote(struct hstate *h, struct folio *folio,
1690                                                         bool adjust_surplus)
1691 {
1692         __remove_hugetlb_folio(h, folio, adjust_surplus, true);
1693 }
1694
1695 static void add_hugetlb_folio(struct hstate *h, struct folio *folio,
1696                              bool adjust_surplus)
1697 {
1698         int zeroed;
1699         int nid = folio_nid(folio);
1700
1701         VM_BUG_ON_FOLIO(!folio_test_hugetlb_vmemmap_optimized(folio), folio);
1702
1703         lockdep_assert_held(&hugetlb_lock);
1704
1705         INIT_LIST_HEAD(&folio->lru);
1706         h->nr_huge_pages++;
1707         h->nr_huge_pages_node[nid]++;
1708
1709         if (adjust_surplus) {
1710                 h->surplus_huge_pages++;
1711                 h->surplus_huge_pages_node[nid]++;
1712         }
1713
1714         folio_set_hugetlb(folio);
1715         folio_change_private(folio, NULL);
1716         /*
1717          * We have to set hugetlb_vmemmap_optimized again as above
1718          * folio_change_private(folio, NULL) cleared it.
1719          */
1720         folio_set_hugetlb_vmemmap_optimized(folio);
1721
1722         /*
1723          * This folio is about to be managed by the hugetlb allocator and
1724          * should have no users.  Drop our reference, and check for others
1725          * just in case.
1726          */
1727         zeroed = folio_put_testzero(folio);
1728         if (unlikely(!zeroed))
1729                 /*
1730                  * It is VERY unlikely soneone else has taken a ref
1731                  * on the folio.  In this case, we simply return as
1732                  * free_huge_folio() will be called when this other ref
1733                  * is dropped.
1734                  */
1735                 return;
1736
1737         arch_clear_hugepage_flags(&folio->page);
1738         enqueue_hugetlb_folio(h, folio);
1739 }
1740
1741 static void __update_and_free_hugetlb_folio(struct hstate *h,
1742                                                 struct folio *folio)
1743 {
1744         bool clear_dtor = folio_test_hugetlb_vmemmap_optimized(folio);
1745
1746         if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1747                 return;
1748
1749         /*
1750          * If we don't know which subpages are hwpoisoned, we can't free
1751          * the hugepage, so it's leaked intentionally.
1752          */
1753         if (folio_test_hugetlb_raw_hwp_unreliable(folio))
1754                 return;
1755
1756         /*
1757          * If folio is not vmemmap optimized (!clear_dtor), then the folio
1758          * is no longer identified as a hugetlb page.  hugetlb_vmemmap_restore_folio
1759          * can only be passed hugetlb pages and will BUG otherwise.
1760          */
1761         if (clear_dtor && hugetlb_vmemmap_restore_folio(h, folio)) {
1762                 spin_lock_irq(&hugetlb_lock);
1763                 /*
1764                  * If we cannot allocate vmemmap pages, just refuse to free the
1765                  * page and put the page back on the hugetlb free list and treat
1766                  * as a surplus page.
1767                  */
1768                 add_hugetlb_folio(h, folio, true);
1769                 spin_unlock_irq(&hugetlb_lock);
1770                 return;
1771         }
1772
1773         /*
1774          * Move PageHWPoison flag from head page to the raw error pages,
1775          * which makes any healthy subpages reusable.
1776          */
1777         if (unlikely(folio_test_hwpoison(folio)))
1778                 folio_clear_hugetlb_hwpoison(folio);
1779
1780         /*
1781          * If vmemmap pages were allocated above, then we need to clear the
1782          * hugetlb destructor under the hugetlb lock.
1783          */
1784         if (clear_dtor) {
1785                 spin_lock_irq(&hugetlb_lock);
1786                 __clear_hugetlb_destructor(h, folio);
1787                 spin_unlock_irq(&hugetlb_lock);
1788         }
1789
1790         /*
1791          * Non-gigantic pages demoted from CMA allocated gigantic pages
1792          * need to be given back to CMA in free_gigantic_folio.
1793          */
1794         if (hstate_is_gigantic(h) ||
1795             hugetlb_cma_folio(folio, huge_page_order(h))) {
1796                 destroy_compound_gigantic_folio(folio, huge_page_order(h));
1797                 free_gigantic_folio(folio, huge_page_order(h));
1798         } else {
1799                 __free_pages(&folio->page, huge_page_order(h));
1800         }
1801 }
1802
1803 /*
1804  * As update_and_free_hugetlb_folio() can be called under any context, so we cannot
1805  * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1806  * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1807  * the vmemmap pages.
1808  *
1809  * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1810  * freed and frees them one-by-one. As the page->mapping pointer is going
1811  * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1812  * structure of a lockless linked list of huge pages to be freed.
1813  */
1814 static LLIST_HEAD(hpage_freelist);
1815
1816 static void free_hpage_workfn(struct work_struct *work)
1817 {
1818         struct llist_node *node;
1819
1820         node = llist_del_all(&hpage_freelist);
1821
1822         while (node) {
1823                 struct folio *folio;
1824                 struct hstate *h;
1825
1826                 folio = container_of((struct address_space **)node,
1827                                      struct folio, mapping);
1828                 node = node->next;
1829                 folio->mapping = NULL;
1830                 /*
1831                  * The VM_BUG_ON_FOLIO(!folio_test_hugetlb(folio), folio) in
1832                  * folio_hstate() is going to trigger because a previous call to
1833                  * remove_hugetlb_folio() will clear the hugetlb bit, so do
1834                  * not use folio_hstate() directly.
1835                  */
1836                 h = size_to_hstate(folio_size(folio));
1837
1838                 __update_and_free_hugetlb_folio(h, folio);
1839
1840                 cond_resched();
1841         }
1842 }
1843 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1844
1845 static inline void flush_free_hpage_work(struct hstate *h)
1846 {
1847         if (hugetlb_vmemmap_optimizable(h))
1848                 flush_work(&free_hpage_work);
1849 }
1850
1851 static void update_and_free_hugetlb_folio(struct hstate *h, struct folio *folio,
1852                                  bool atomic)
1853 {
1854         if (!folio_test_hugetlb_vmemmap_optimized(folio) || !atomic) {
1855                 __update_and_free_hugetlb_folio(h, folio);
1856                 return;
1857         }
1858
1859         /*
1860          * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1861          *
1862          * Only call schedule_work() if hpage_freelist is previously
1863          * empty. Otherwise, schedule_work() had been called but the workfn
1864          * hasn't retrieved the list yet.
1865          */
1866         if (llist_add((struct llist_node *)&folio->mapping, &hpage_freelist))
1867                 schedule_work(&free_hpage_work);
1868 }
1869
1870 static void bulk_vmemmap_restore_error(struct hstate *h,
1871                                         struct list_head *folio_list,
1872                                         struct list_head *non_hvo_folios)
1873 {
1874         struct folio *folio, *t_folio;
1875
1876         if (!list_empty(non_hvo_folios)) {
1877                 /*
1878                  * Free any restored hugetlb pages so that restore of the
1879                  * entire list can be retried.
1880                  * The idea is that in the common case of ENOMEM errors freeing
1881                  * hugetlb pages with vmemmap we will free up memory so that we
1882                  * can allocate vmemmap for more hugetlb pages.
1883                  */
1884                 list_for_each_entry_safe(folio, t_folio, non_hvo_folios, lru) {
1885                         list_del(&folio->lru);
1886                         spin_lock_irq(&hugetlb_lock);
1887                         __clear_hugetlb_destructor(h, folio);
1888                         spin_unlock_irq(&hugetlb_lock);
1889                         update_and_free_hugetlb_folio(h, folio, false);
1890                         cond_resched();
1891                 }
1892         } else {
1893                 /*
1894                  * In the case where there are no folios which can be
1895                  * immediately freed, we loop through the list trying to restore
1896                  * vmemmap individually in the hope that someone elsewhere may
1897                  * have done something to cause success (such as freeing some
1898                  * memory).  If unable to restore a hugetlb page, the hugetlb
1899                  * page is made a surplus page and removed from the list.
1900                  * If are able to restore vmemmap and free one hugetlb page, we
1901                  * quit processing the list to retry the bulk operation.
1902                  */
1903                 list_for_each_entry_safe(folio, t_folio, folio_list, lru)
1904                         if (hugetlb_vmemmap_restore_folio(h, folio)) {
1905                                 list_del(&folio->lru);
1906                                 spin_lock_irq(&hugetlb_lock);
1907                                 add_hugetlb_folio(h, folio, true);
1908                                 spin_unlock_irq(&hugetlb_lock);
1909                         } else {
1910                                 list_del(&folio->lru);
1911                                 spin_lock_irq(&hugetlb_lock);
1912                                 __clear_hugetlb_destructor(h, folio);
1913                                 spin_unlock_irq(&hugetlb_lock);
1914                                 update_and_free_hugetlb_folio(h, folio, false);
1915                                 cond_resched();
1916                                 break;
1917                         }
1918         }
1919 }
1920
1921 static void update_and_free_pages_bulk(struct hstate *h,
1922                                                 struct list_head *folio_list)
1923 {
1924         long ret;
1925         struct folio *folio, *t_folio;
1926         LIST_HEAD(non_hvo_folios);
1927
1928         /*
1929          * First allocate required vmemmmap (if necessary) for all folios.
1930          * Carefully handle errors and free up any available hugetlb pages
1931          * in an effort to make forward progress.
1932          */
1933 retry:
1934         ret = hugetlb_vmemmap_restore_folios(h, folio_list, &non_hvo_folios);
1935         if (ret < 0) {
1936                 bulk_vmemmap_restore_error(h, folio_list, &non_hvo_folios);
1937                 goto retry;
1938         }
1939
1940         /*
1941          * At this point, list should be empty, ret should be >= 0 and there
1942          * should only be pages on the non_hvo_folios list.
1943          * Do note that the non_hvo_folios list could be empty.
1944          * Without HVO enabled, ret will be 0 and there is no need to call
1945          * __clear_hugetlb_destructor as this was done previously.
1946          */
1947         VM_WARN_ON(!list_empty(folio_list));
1948         VM_WARN_ON(ret < 0);
1949         if (!list_empty(&non_hvo_folios) && ret) {
1950                 spin_lock_irq(&hugetlb_lock);
1951                 list_for_each_entry(folio, &non_hvo_folios, lru)
1952                         __clear_hugetlb_destructor(h, folio);
1953                 spin_unlock_irq(&hugetlb_lock);
1954         }
1955
1956         list_for_each_entry_safe(folio, t_folio, &non_hvo_folios, lru) {
1957                 update_and_free_hugetlb_folio(h, folio, false);
1958                 cond_resched();
1959         }
1960 }
1961
1962 struct hstate *size_to_hstate(unsigned long size)
1963 {
1964         struct hstate *h;
1965
1966         for_each_hstate(h) {
1967                 if (huge_page_size(h) == size)
1968                         return h;
1969         }
1970         return NULL;
1971 }
1972
1973 void free_huge_folio(struct folio *folio)
1974 {
1975         /*
1976          * Can't pass hstate in here because it is called from the
1977          * compound page destructor.
1978          */
1979         struct hstate *h = folio_hstate(folio);
1980         int nid = folio_nid(folio);
1981         struct hugepage_subpool *spool = hugetlb_folio_subpool(folio);
1982         bool restore_reserve;
1983         unsigned long flags;
1984
1985         VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
1986         VM_BUG_ON_FOLIO(folio_mapcount(folio), folio);
1987
1988         hugetlb_set_folio_subpool(folio, NULL);
1989         if (folio_test_anon(folio))
1990                 __ClearPageAnonExclusive(&folio->page);
1991         folio->mapping = NULL;
1992         restore_reserve = folio_test_hugetlb_restore_reserve(folio);
1993         folio_clear_hugetlb_restore_reserve(folio);
1994
1995         /*
1996          * If HPageRestoreReserve was set on page, page allocation consumed a
1997          * reservation.  If the page was associated with a subpool, there
1998          * would have been a page reserved in the subpool before allocation
1999          * via hugepage_subpool_get_pages().  Since we are 'restoring' the
2000          * reservation, do not call hugepage_subpool_put_pages() as this will
2001          * remove the reserved page from the subpool.
2002          */
2003         if (!restore_reserve) {
2004                 /*
2005                  * A return code of zero implies that the subpool will be
2006                  * under its minimum size if the reservation is not restored
2007                  * after page is free.  Therefore, force restore_reserve
2008                  * operation.
2009                  */
2010                 if (hugepage_subpool_put_pages(spool, 1) == 0)
2011                         restore_reserve = true;
2012         }
2013
2014         spin_lock_irqsave(&hugetlb_lock, flags);
2015         folio_clear_hugetlb_migratable(folio);
2016         hugetlb_cgroup_uncharge_folio(hstate_index(h),
2017                                      pages_per_huge_page(h), folio);
2018         hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),
2019                                           pages_per_huge_page(h), folio);
2020         mem_cgroup_uncharge(folio);
2021         if (restore_reserve)
2022                 h->resv_huge_pages++;
2023
2024         if (folio_test_hugetlb_temporary(folio)) {
2025                 remove_hugetlb_folio(h, folio, false);
2026                 spin_unlock_irqrestore(&hugetlb_lock, flags);
2027                 update_and_free_hugetlb_folio(h, folio, true);
2028         } else if (h->surplus_huge_pages_node[nid]) {
2029                 /* remove the page from active list */
2030                 remove_hugetlb_folio(h, folio, true);
2031                 spin_unlock_irqrestore(&hugetlb_lock, flags);
2032                 update_and_free_hugetlb_folio(h, folio, true);
2033         } else {
2034                 arch_clear_hugepage_flags(&folio->page);
2035                 enqueue_hugetlb_folio(h, folio);
2036                 spin_unlock_irqrestore(&hugetlb_lock, flags);
2037         }
2038 }
2039
2040 /*
2041  * Must be called with the hugetlb lock held
2042  */
2043 static void __prep_account_new_huge_page(struct hstate *h, int nid)
2044 {
2045         lockdep_assert_held(&hugetlb_lock);
2046         h->nr_huge_pages++;
2047         h->nr_huge_pages_node[nid]++;
2048 }
2049
2050 static void init_new_hugetlb_folio(struct hstate *h, struct folio *folio)
2051 {
2052         folio_set_hugetlb(folio);
2053         INIT_LIST_HEAD(&folio->lru);
2054         hugetlb_set_folio_subpool(folio, NULL);
2055         set_hugetlb_cgroup(folio, NULL);
2056         set_hugetlb_cgroup_rsvd(folio, NULL);
2057 }
2058
2059 static void __prep_new_hugetlb_folio(struct hstate *h, struct folio *folio)
2060 {
2061         init_new_hugetlb_folio(h, folio);
2062         hugetlb_vmemmap_optimize_folio(h, folio);
2063 }
2064
2065 static void prep_new_hugetlb_folio(struct hstate *h, struct folio *folio, int nid)
2066 {
2067         __prep_new_hugetlb_folio(h, folio);
2068         spin_lock_irq(&hugetlb_lock);
2069         __prep_account_new_huge_page(h, nid);
2070         spin_unlock_irq(&hugetlb_lock);
2071 }
2072
2073 static bool __prep_compound_gigantic_folio(struct folio *folio,
2074                                         unsigned int order, bool demote)
2075 {
2076         int i, j;
2077         int nr_pages = 1 << order;
2078         struct page *p;
2079
2080         __folio_clear_reserved(folio);
2081         for (i = 0; i < nr_pages; i++) {
2082                 p = folio_page(folio, i);
2083
2084                 /*
2085                  * For gigantic hugepages allocated through bootmem at
2086                  * boot, it's safer to be consistent with the not-gigantic
2087                  * hugepages and clear the PG_reserved bit from all tail pages
2088                  * too.  Otherwise drivers using get_user_pages() to access tail
2089                  * pages may get the reference counting wrong if they see
2090                  * PG_reserved set on a tail page (despite the head page not
2091                  * having PG_reserved set).  Enforcing this consistency between
2092                  * head and tail pages allows drivers to optimize away a check
2093                  * on the head page when they need know if put_page() is needed
2094                  * after get_user_pages().
2095                  */
2096                 if (i != 0)     /* head page cleared above */
2097                         __ClearPageReserved(p);
2098                 /*
2099                  * Subtle and very unlikely
2100                  *
2101                  * Gigantic 'page allocators' such as memblock or cma will
2102                  * return a set of pages with each page ref counted.  We need
2103                  * to turn this set of pages into a compound page with tail
2104                  * page ref counts set to zero.  Code such as speculative page
2105                  * cache adding could take a ref on a 'to be' tail page.
2106                  * We need to respect any increased ref count, and only set
2107                  * the ref count to zero if count is currently 1.  If count
2108                  * is not 1, we return an error.  An error return indicates
2109                  * the set of pages can not be converted to a gigantic page.
2110                  * The caller who allocated the pages should then discard the
2111                  * pages using the appropriate free interface.
2112                  *
2113                  * In the case of demote, the ref count will be zero.
2114                  */
2115                 if (!demote) {
2116                         if (!page_ref_freeze(p, 1)) {
2117                                 pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
2118                                 goto out_error;
2119                         }
2120                 } else {
2121                         VM_BUG_ON_PAGE(page_count(p), p);
2122                 }
2123                 if (i != 0)
2124                         set_compound_head(p, &folio->page);
2125         }
2126         __folio_set_head(folio);
2127         /* we rely on prep_new_hugetlb_folio to set the destructor */
2128         folio_set_order(folio, order);
2129         atomic_set(&folio->_entire_mapcount, -1);
2130         atomic_set(&folio->_nr_pages_mapped, 0);
2131         atomic_set(&folio->_pincount, 0);
2132         return true;
2133
2134 out_error:
2135         /* undo page modifications made above */
2136         for (j = 0; j < i; j++) {
2137                 p = folio_page(folio, j);
2138                 if (j != 0)
2139                         clear_compound_head(p);
2140                 set_page_refcounted(p);
2141         }
2142         /* need to clear PG_reserved on remaining tail pages  */
2143         for (; j < nr_pages; j++) {
2144                 p = folio_page(folio, j);
2145                 __ClearPageReserved(p);
2146         }
2147         return false;
2148 }
2149
2150 static bool prep_compound_gigantic_folio(struct folio *folio,
2151                                                         unsigned int order)
2152 {
2153         return __prep_compound_gigantic_folio(folio, order, false);
2154 }
2155
2156 static bool prep_compound_gigantic_folio_for_demote(struct folio *folio,
2157                                                         unsigned int order)
2158 {
2159         return __prep_compound_gigantic_folio(folio, order, true);
2160 }
2161
2162 /*
2163  * PageHuge() only returns true for hugetlbfs pages, but not for normal or
2164  * transparent huge pages.  See the PageTransHuge() documentation for more
2165  * details.
2166  */
2167 int PageHuge(const struct page *page)
2168 {
2169         const struct folio *folio;
2170
2171         if (!PageCompound(page))
2172                 return 0;
2173         folio = page_folio(page);
2174         return folio_test_hugetlb(folio);
2175 }
2176 EXPORT_SYMBOL_GPL(PageHuge);
2177
2178 /*
2179  * Find and lock address space (mapping) in write mode.
2180  *
2181  * Upon entry, the page is locked which means that page_mapping() is
2182  * stable.  Due to locking order, we can only trylock_write.  If we can
2183  * not get the lock, simply return NULL to caller.
2184  */
2185 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
2186 {
2187         struct address_space *mapping = page_mapping(hpage);
2188
2189         if (!mapping)
2190                 return mapping;
2191
2192         if (i_mmap_trylock_write(mapping))
2193                 return mapping;
2194
2195         return NULL;
2196 }
2197
2198 static struct folio *alloc_buddy_hugetlb_folio(struct hstate *h,
2199                 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2200                 nodemask_t *node_alloc_noretry)
2201 {
2202         int order = huge_page_order(h);
2203         struct page *page;
2204         bool alloc_try_hard = true;
2205         bool retry = true;
2206
2207         /*
2208          * By default we always try hard to allocate the page with
2209          * __GFP_RETRY_MAYFAIL flag.  However, if we are allocating pages in
2210          * a loop (to adjust global huge page counts) and previous allocation
2211          * failed, do not continue to try hard on the same node.  Use the
2212          * node_alloc_noretry bitmap to manage this state information.
2213          */
2214         if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
2215                 alloc_try_hard = false;
2216         gfp_mask |= __GFP_COMP|__GFP_NOWARN;
2217         if (alloc_try_hard)
2218                 gfp_mask |= __GFP_RETRY_MAYFAIL;
2219         if (nid == NUMA_NO_NODE)
2220                 nid = numa_mem_id();
2221 retry:
2222         page = __alloc_pages(gfp_mask, order, nid, nmask);
2223
2224         /* Freeze head page */
2225         if (page && !page_ref_freeze(page, 1)) {
2226                 __free_pages(page, order);
2227                 if (retry) {    /* retry once */
2228                         retry = false;
2229                         goto retry;
2230                 }
2231                 /* WOW!  twice in a row. */
2232                 pr_warn("HugeTLB head page unexpected inflated ref count\n");
2233                 page = NULL;
2234         }
2235
2236         /*
2237          * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
2238          * indicates an overall state change.  Clear bit so that we resume
2239          * normal 'try hard' allocations.
2240          */
2241         if (node_alloc_noretry && page && !alloc_try_hard)
2242                 node_clear(nid, *node_alloc_noretry);
2243
2244         /*
2245          * If we tried hard to get a page but failed, set bit so that
2246          * subsequent attempts will not try as hard until there is an
2247          * overall state change.
2248          */
2249         if (node_alloc_noretry && !page && alloc_try_hard)
2250                 node_set(nid, *node_alloc_noretry);
2251
2252         if (!page) {
2253                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
2254                 return NULL;
2255         }
2256
2257         __count_vm_event(HTLB_BUDDY_PGALLOC);
2258         return page_folio(page);
2259 }
2260
2261 static struct folio *__alloc_fresh_hugetlb_folio(struct hstate *h,
2262                                 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2263                                 nodemask_t *node_alloc_noretry)
2264 {
2265         struct folio *folio;
2266         bool retry = false;
2267
2268 retry:
2269         if (hstate_is_gigantic(h))
2270                 folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask);
2271         else
2272                 folio = alloc_buddy_hugetlb_folio(h, gfp_mask,
2273                                 nid, nmask, node_alloc_noretry);
2274         if (!folio)
2275                 return NULL;
2276
2277         if (hstate_is_gigantic(h)) {
2278                 if (!prep_compound_gigantic_folio(folio, huge_page_order(h))) {
2279                         /*
2280                          * Rare failure to convert pages to compound page.
2281                          * Free pages and try again - ONCE!
2282                          */
2283                         free_gigantic_folio(folio, huge_page_order(h));
2284                         if (!retry) {
2285                                 retry = true;
2286                                 goto retry;
2287                         }
2288                         return NULL;
2289                 }
2290         }
2291
2292         return folio;
2293 }
2294
2295 static struct folio *only_alloc_fresh_hugetlb_folio(struct hstate *h,
2296                 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2297                 nodemask_t *node_alloc_noretry)
2298 {
2299         struct folio *folio;
2300
2301         folio = __alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask,
2302                                                 node_alloc_noretry);
2303         if (folio)
2304                 init_new_hugetlb_folio(h, folio);
2305         return folio;
2306 }
2307
2308 /*
2309  * Common helper to allocate a fresh hugetlb page. All specific allocators
2310  * should use this function to get new hugetlb pages
2311  *
2312  * Note that returned page is 'frozen':  ref count of head page and all tail
2313  * pages is zero.
2314  */
2315 static struct folio *alloc_fresh_hugetlb_folio(struct hstate *h,
2316                 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2317                 nodemask_t *node_alloc_noretry)
2318 {
2319         struct folio *folio;
2320
2321         folio = __alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask,
2322                                                 node_alloc_noretry);
2323         if (!folio)
2324                 return NULL;
2325
2326         prep_new_hugetlb_folio(h, folio, folio_nid(folio));
2327         return folio;
2328 }
2329
2330 static void prep_and_add_allocated_folios(struct hstate *h,
2331                                         struct list_head *folio_list)
2332 {
2333         unsigned long flags;
2334         struct folio *folio, *tmp_f;
2335
2336         /* Send list for bulk vmemmap optimization processing */
2337         hugetlb_vmemmap_optimize_folios(h, folio_list);
2338
2339         /* Add all new pool pages to free lists in one lock cycle */
2340         spin_lock_irqsave(&hugetlb_lock, flags);
2341         list_for_each_entry_safe(folio, tmp_f, folio_list, lru) {
2342                 __prep_account_new_huge_page(h, folio_nid(folio));
2343                 enqueue_hugetlb_folio(h, folio);
2344         }
2345         spin_unlock_irqrestore(&hugetlb_lock, flags);
2346 }
2347
2348 /*
2349  * Allocates a fresh hugetlb page in a node interleaved manner.  The page
2350  * will later be added to the appropriate hugetlb pool.
2351  */
2352 static struct folio *alloc_pool_huge_folio(struct hstate *h,
2353                                         nodemask_t *nodes_allowed,
2354                                         nodemask_t *node_alloc_noretry,
2355                                         int *next_node)
2356 {
2357         gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2358         int nr_nodes, node;
2359
2360         for_each_node_mask_to_alloc(next_node, nr_nodes, node, nodes_allowed) {
2361                 struct folio *folio;
2362
2363                 folio = only_alloc_fresh_hugetlb_folio(h, gfp_mask, node,
2364                                         nodes_allowed, node_alloc_noretry);
2365                 if (folio)
2366                         return folio;
2367         }
2368
2369         return NULL;
2370 }
2371
2372 /*
2373  * Remove huge page from pool from next node to free.  Attempt to keep
2374  * persistent huge pages more or less balanced over allowed nodes.
2375  * This routine only 'removes' the hugetlb page.  The caller must make
2376  * an additional call to free the page to low level allocators.
2377  * Called with hugetlb_lock locked.
2378  */
2379 static struct folio *remove_pool_hugetlb_folio(struct hstate *h,
2380                 nodemask_t *nodes_allowed, bool acct_surplus)
2381 {
2382         int nr_nodes, node;
2383         struct folio *folio = NULL;
2384
2385         lockdep_assert_held(&hugetlb_lock);
2386         for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2387                 /*
2388                  * If we're returning unused surplus pages, only examine
2389                  * nodes with surplus pages.
2390                  */
2391                 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
2392                     !list_empty(&h->hugepage_freelists[node])) {
2393                         folio = list_entry(h->hugepage_freelists[node].next,
2394                                           struct folio, lru);
2395                         remove_hugetlb_folio(h, folio, acct_surplus);
2396                         break;
2397                 }
2398         }
2399
2400         return folio;
2401 }
2402
2403 /*
2404  * Dissolve a given free hugepage into free buddy pages. This function does
2405  * nothing for in-use hugepages and non-hugepages.
2406  * This function returns values like below:
2407  *
2408  *  -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
2409  *           when the system is under memory pressure and the feature of
2410  *           freeing unused vmemmap pages associated with each hugetlb page
2411  *           is enabled.
2412  *  -EBUSY:  failed to dissolved free hugepages or the hugepage is in-use
2413  *           (allocated or reserved.)
2414  *       0:  successfully dissolved free hugepages or the page is not a
2415  *           hugepage (considered as already dissolved)
2416  */
2417 int dissolve_free_huge_page(struct page *page)
2418 {
2419         int rc = -EBUSY;
2420         struct folio *folio = page_folio(page);
2421
2422 retry:
2423         /* Not to disrupt normal path by vainly holding hugetlb_lock */
2424         if (!folio_test_hugetlb(folio))
2425                 return 0;
2426
2427         spin_lock_irq(&hugetlb_lock);
2428         if (!folio_test_hugetlb(folio)) {
2429                 rc = 0;
2430                 goto out;
2431         }
2432
2433         if (!folio_ref_count(folio)) {
2434                 struct hstate *h = folio_hstate(folio);
2435                 if (!available_huge_pages(h))
2436                         goto out;
2437
2438                 /*
2439                  * We should make sure that the page is already on the free list
2440                  * when it is dissolved.
2441                  */
2442                 if (unlikely(!folio_test_hugetlb_freed(folio))) {
2443                         spin_unlock_irq(&hugetlb_lock);
2444                         cond_resched();
2445
2446                         /*
2447                          * Theoretically, we should return -EBUSY when we
2448                          * encounter this race. In fact, we have a chance
2449                          * to successfully dissolve the page if we do a
2450                          * retry. Because the race window is quite small.
2451                          * If we seize this opportunity, it is an optimization
2452                          * for increasing the success rate of dissolving page.
2453                          */
2454                         goto retry;
2455                 }
2456
2457                 remove_hugetlb_folio(h, folio, false);
2458                 h->max_huge_pages--;
2459                 spin_unlock_irq(&hugetlb_lock);
2460
2461                 /*
2462                  * Normally update_and_free_hugtlb_folio will allocate required vmemmmap
2463                  * before freeing the page.  update_and_free_hugtlb_folio will fail to
2464                  * free the page if it can not allocate required vmemmap.  We
2465                  * need to adjust max_huge_pages if the page is not freed.
2466                  * Attempt to allocate vmemmmap here so that we can take
2467                  * appropriate action on failure.
2468                  *
2469                  * The folio_test_hugetlb check here is because
2470                  * remove_hugetlb_folio will clear hugetlb folio flag for
2471                  * non-vmemmap optimized hugetlb folios.
2472                  */
2473                 if (folio_test_hugetlb(folio)) {
2474                         rc = hugetlb_vmemmap_restore_folio(h, folio);
2475                         if (rc) {
2476                                 spin_lock_irq(&hugetlb_lock);
2477                                 add_hugetlb_folio(h, folio, false);
2478                                 h->max_huge_pages++;
2479                                 goto out;
2480                         }
2481                 } else
2482                         rc = 0;
2483
2484                 update_and_free_hugetlb_folio(h, folio, false);
2485                 return rc;
2486         }
2487 out:
2488         spin_unlock_irq(&hugetlb_lock);
2489         return rc;
2490 }
2491
2492 /*
2493  * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
2494  * make specified memory blocks removable from the system.
2495  * Note that this will dissolve a free gigantic hugepage completely, if any
2496  * part of it lies within the given range.
2497  * Also note that if dissolve_free_huge_page() returns with an error, all
2498  * free hugepages that were dissolved before that error are lost.
2499  */
2500 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
2501 {
2502         unsigned long pfn;
2503         struct page *page;
2504         int rc = 0;
2505         unsigned int order;
2506         struct hstate *h;
2507
2508         if (!hugepages_supported())
2509                 return rc;
2510
2511         order = huge_page_order(&default_hstate);
2512         for_each_hstate(h)
2513                 order = min(order, huge_page_order(h));
2514
2515         for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) {
2516                 page = pfn_to_page(pfn);
2517                 rc = dissolve_free_huge_page(page);
2518                 if (rc)
2519                         break;
2520         }
2521
2522         return rc;
2523 }
2524
2525 /*
2526  * Allocates a fresh surplus page from the page allocator.
2527  */
2528 static struct folio *alloc_surplus_hugetlb_folio(struct hstate *h,
2529                                 gfp_t gfp_mask, int nid, nodemask_t *nmask)
2530 {
2531         struct folio *folio = NULL;
2532
2533         if (hstate_is_gigantic(h))
2534                 return NULL;
2535
2536         spin_lock_irq(&hugetlb_lock);
2537         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
2538                 goto out_unlock;
2539         spin_unlock_irq(&hugetlb_lock);
2540
2541         folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);
2542         if (!folio)
2543                 return NULL;
2544
2545         spin_lock_irq(&hugetlb_lock);
2546         /*
2547          * We could have raced with the pool size change.
2548          * Double check that and simply deallocate the new page
2549          * if we would end up overcommiting the surpluses. Abuse
2550          * temporary page to workaround the nasty free_huge_folio
2551          * codeflow
2552          */
2553         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
2554                 folio_set_hugetlb_temporary(folio);
2555                 spin_unlock_irq(&hugetlb_lock);
2556                 free_huge_folio(folio);
2557                 return NULL;
2558         }
2559
2560         h->surplus_huge_pages++;
2561         h->surplus_huge_pages_node[folio_nid(folio)]++;
2562
2563 out_unlock:
2564         spin_unlock_irq(&hugetlb_lock);
2565
2566         return folio;
2567 }
2568
2569 static struct folio *alloc_migrate_hugetlb_folio(struct hstate *h, gfp_t gfp_mask,
2570                                      int nid, nodemask_t *nmask)
2571 {
2572         struct folio *folio;
2573
2574         if (hstate_is_gigantic(h))
2575                 return NULL;
2576
2577         folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);
2578         if (!folio)
2579                 return NULL;
2580
2581         /* fresh huge pages are frozen */
2582         folio_ref_unfreeze(folio, 1);
2583         /*
2584          * We do not account these pages as surplus because they are only
2585          * temporary and will be released properly on the last reference
2586          */
2587         folio_set_hugetlb_temporary(folio);
2588
2589         return folio;
2590 }
2591
2592 /*
2593  * Use the VMA's mpolicy to allocate a huge page from the buddy.
2594  */
2595 static
2596 struct folio *alloc_buddy_hugetlb_folio_with_mpol(struct hstate *h,
2597                 struct vm_area_struct *vma, unsigned long addr)
2598 {
2599         struct folio *folio = NULL;
2600         struct mempolicy *mpol;
2601         gfp_t gfp_mask = htlb_alloc_mask(h);
2602         int nid;
2603         nodemask_t *nodemask;
2604
2605         nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
2606         if (mpol_is_preferred_many(mpol)) {
2607                 gfp_t gfp = gfp_mask | __GFP_NOWARN;
2608
2609                 gfp &=  ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2610                 folio = alloc_surplus_hugetlb_folio(h, gfp, nid, nodemask);
2611
2612                 /* Fallback to all nodes if page==NULL */
2613                 nodemask = NULL;
2614         }
2615
2616         if (!folio)
2617                 folio = alloc_surplus_hugetlb_folio(h, gfp_mask, nid, nodemask);
2618         mpol_cond_put(mpol);
2619         return folio;
2620 }
2621
2622 /* folio migration callback function */
2623 struct folio *alloc_hugetlb_folio_nodemask(struct hstate *h, int preferred_nid,
2624                 nodemask_t *nmask, gfp_t gfp_mask)
2625 {
2626         spin_lock_irq(&hugetlb_lock);
2627         if (available_huge_pages(h)) {
2628                 struct folio *folio;
2629
2630                 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
2631                                                 preferred_nid, nmask);
2632                 if (folio) {
2633                         spin_unlock_irq(&hugetlb_lock);
2634                         return folio;
2635                 }
2636         }
2637         spin_unlock_irq(&hugetlb_lock);
2638
2639         return alloc_migrate_hugetlb_folio(h, gfp_mask, preferred_nid, nmask);
2640 }
2641
2642 /*
2643  * Increase the hugetlb pool such that it can accommodate a reservation
2644  * of size 'delta'.
2645  */
2646 static int gather_surplus_pages(struct hstate *h, long delta)
2647         __must_hold(&hugetlb_lock)
2648 {
2649         LIST_HEAD(surplus_list);
2650         struct folio *folio, *tmp;
2651         int ret;
2652         long i;
2653         long needed, allocated;
2654         bool alloc_ok = true;
2655
2656         lockdep_assert_held(&hugetlb_lock);
2657         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2658         if (needed <= 0) {
2659                 h->resv_huge_pages += delta;
2660                 return 0;
2661         }
2662
2663         allocated = 0;
2664
2665         ret = -ENOMEM;
2666 retry:
2667         spin_unlock_irq(&hugetlb_lock);
2668         for (i = 0; i < needed; i++) {
2669                 folio = alloc_surplus_hugetlb_folio(h, htlb_alloc_mask(h),
2670                                 NUMA_NO_NODE, NULL);
2671                 if (!folio) {
2672                         alloc_ok = false;
2673                         break;
2674                 }
2675                 list_add(&folio->lru, &surplus_list);
2676                 cond_resched();
2677         }
2678         allocated += i;
2679
2680         /*
2681          * After retaking hugetlb_lock, we need to recalculate 'needed'
2682          * because either resv_huge_pages or free_huge_pages may have changed.
2683          */
2684         spin_lock_irq(&hugetlb_lock);
2685         needed = (h->resv_huge_pages + delta) -
2686                         (h->free_huge_pages + allocated);
2687         if (needed > 0) {
2688                 if (alloc_ok)
2689                         goto retry;
2690                 /*
2691                  * We were not able to allocate enough pages to
2692                  * satisfy the entire reservation so we free what
2693                  * we've allocated so far.
2694                  */
2695                 goto free;
2696         }
2697         /*
2698          * The surplus_list now contains _at_least_ the number of extra pages
2699          * needed to accommodate the reservation.  Add the appropriate number
2700          * of pages to the hugetlb pool and free the extras back to the buddy
2701          * allocator.  Commit the entire reservation here to prevent another
2702          * process from stealing the pages as they are added to the pool but
2703          * before they are reserved.
2704          */
2705         needed += allocated;
2706         h->resv_huge_pages += delta;
2707         ret = 0;
2708
2709         /* Free the needed pages to the hugetlb pool */
2710         list_for_each_entry_safe(folio, tmp, &surplus_list, lru) {
2711                 if ((--needed) < 0)
2712                         break;
2713                 /* Add the page to the hugetlb allocator */
2714                 enqueue_hugetlb_folio(h, folio);
2715         }
2716 free:
2717         spin_unlock_irq(&hugetlb_lock);
2718
2719         /*
2720          * Free unnecessary surplus pages to the buddy allocator.
2721          * Pages have no ref count, call free_huge_folio directly.
2722          */
2723         list_for_each_entry_safe(folio, tmp, &surplus_list, lru)
2724                 free_huge_folio(folio);
2725         spin_lock_irq(&hugetlb_lock);
2726
2727         return ret;
2728 }
2729
2730 /*
2731  * This routine has two main purposes:
2732  * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2733  *    in unused_resv_pages.  This corresponds to the prior adjustments made
2734  *    to the associated reservation map.
2735  * 2) Free any unused surplus pages that may have been allocated to satisfy
2736  *    the reservation.  As many as unused_resv_pages may be freed.
2737  */
2738 static void return_unused_surplus_pages(struct hstate *h,
2739                                         unsigned long unused_resv_pages)
2740 {
2741         unsigned long nr_pages;
2742         LIST_HEAD(page_list);
2743
2744         lockdep_assert_held(&hugetlb_lock);
2745         /* Uncommit the reservation */
2746         h->resv_huge_pages -= unused_resv_pages;
2747
2748         if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2749                 goto out;
2750
2751         /*
2752          * Part (or even all) of the reservation could have been backed
2753          * by pre-allocated pages. Only free surplus pages.
2754          */
2755         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2756
2757         /*
2758          * We want to release as many surplus pages as possible, spread
2759          * evenly across all nodes with memory. Iterate across these nodes
2760          * until we can no longer free unreserved surplus pages. This occurs
2761          * when the nodes with surplus pages have no free pages.
2762          * remove_pool_hugetlb_folio() will balance the freed pages across the
2763          * on-line nodes with memory and will handle the hstate accounting.
2764          */
2765         while (nr_pages--) {
2766                 struct folio *folio;
2767
2768                 folio = remove_pool_hugetlb_folio(h, &node_states[N_MEMORY], 1);
2769                 if (!folio)
2770                         goto out;
2771
2772                 list_add(&folio->lru, &page_list);
2773         }
2774
2775 out:
2776         spin_unlock_irq(&hugetlb_lock);
2777         update_and_free_pages_bulk(h, &page_list);
2778         spin_lock_irq(&hugetlb_lock);
2779 }
2780
2781
2782 /*
2783  * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2784  * are used by the huge page allocation routines to manage reservations.
2785  *
2786  * vma_needs_reservation is called to determine if the huge page at addr
2787  * within the vma has an associated reservation.  If a reservation is
2788  * needed, the value 1 is returned.  The caller is then responsible for
2789  * managing the global reservation and subpool usage counts.  After
2790  * the huge page has been allocated, vma_commit_reservation is called
2791  * to add the page to the reservation map.  If the page allocation fails,
2792  * the reservation must be ended instead of committed.  vma_end_reservation
2793  * is called in such cases.
2794  *
2795  * In the normal case, vma_commit_reservation returns the same value
2796  * as the preceding vma_needs_reservation call.  The only time this
2797  * is not the case is if a reserve map was changed between calls.  It
2798  * is the responsibility of the caller to notice the difference and
2799  * take appropriate action.
2800  *
2801  * vma_add_reservation is used in error paths where a reservation must
2802  * be restored when a newly allocated huge page must be freed.  It is
2803  * to be called after calling vma_needs_reservation to determine if a
2804  * reservation exists.
2805  *
2806  * vma_del_reservation is used in error paths where an entry in the reserve
2807  * map was created during huge page allocation and must be removed.  It is to
2808  * be called after calling vma_needs_reservation to determine if a reservation
2809  * exists.
2810  */
2811 enum vma_resv_mode {
2812         VMA_NEEDS_RESV,
2813         VMA_COMMIT_RESV,
2814         VMA_END_RESV,
2815         VMA_ADD_RESV,
2816         VMA_DEL_RESV,
2817 };
2818 static long __vma_reservation_common(struct hstate *h,
2819                                 struct vm_area_struct *vma, unsigned long addr,
2820                                 enum vma_resv_mode mode)
2821 {
2822         struct resv_map *resv;
2823         pgoff_t idx;
2824         long ret;
2825         long dummy_out_regions_needed;
2826
2827         resv = vma_resv_map(vma);
2828         if (!resv)
2829                 return 1;
2830
2831         idx = vma_hugecache_offset(h, vma, addr);
2832         switch (mode) {
2833         case VMA_NEEDS_RESV:
2834                 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2835                 /* We assume that vma_reservation_* routines always operate on
2836                  * 1 page, and that adding to resv map a 1 page entry can only
2837                  * ever require 1 region.
2838                  */
2839                 VM_BUG_ON(dummy_out_regions_needed != 1);
2840                 break;
2841         case VMA_COMMIT_RESV:
2842                 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2843                 /* region_add calls of range 1 should never fail. */
2844                 VM_BUG_ON(ret < 0);
2845                 break;
2846         case VMA_END_RESV:
2847                 region_abort(resv, idx, idx + 1, 1);
2848                 ret = 0;
2849                 break;
2850         case VMA_ADD_RESV:
2851                 if (vma->vm_flags & VM_MAYSHARE) {
2852                         ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2853                         /* region_add calls of range 1 should never fail. */
2854                         VM_BUG_ON(ret < 0);
2855                 } else {
2856                         region_abort(resv, idx, idx + 1, 1);
2857                         ret = region_del(resv, idx, idx + 1);
2858                 }
2859                 break;
2860         case VMA_DEL_RESV:
2861                 if (vma->vm_flags & VM_MAYSHARE) {
2862                         region_abort(resv, idx, idx + 1, 1);
2863                         ret = region_del(resv, idx, idx + 1);
2864                 } else {
2865                         ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2866                         /* region_add calls of range 1 should never fail. */
2867                         VM_BUG_ON(ret < 0);
2868                 }
2869                 break;
2870         default:
2871                 BUG();
2872         }
2873
2874         if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
2875                 return ret;
2876         /*
2877          * We know private mapping must have HPAGE_RESV_OWNER set.
2878          *
2879          * In most cases, reserves always exist for private mappings.
2880          * However, a file associated with mapping could have been
2881          * hole punched or truncated after reserves were consumed.
2882          * As subsequent fault on such a range will not use reserves.
2883          * Subtle - The reserve map for private mappings has the
2884          * opposite meaning than that of shared mappings.  If NO
2885          * entry is in the reserve map, it means a reservation exists.
2886          * If an entry exists in the reserve map, it means the
2887          * reservation has already been consumed.  As a result, the
2888          * return value of this routine is the opposite of the
2889          * value returned from reserve map manipulation routines above.
2890          */
2891         if (ret > 0)
2892                 return 0;
2893         if (ret == 0)
2894                 return 1;
2895         return ret;
2896 }
2897
2898 static long vma_needs_reservation(struct hstate *h,
2899                         struct vm_area_struct *vma, unsigned long addr)
2900 {
2901         return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2902 }
2903
2904 static long vma_commit_reservation(struct hstate *h,
2905                         struct vm_area_struct *vma, unsigned long addr)
2906 {
2907         return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2908 }
2909
2910 static void vma_end_reservation(struct hstate *h,
2911                         struct vm_area_struct *vma, unsigned long addr)
2912 {
2913         (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2914 }
2915
2916 static long vma_add_reservation(struct hstate *h,
2917                         struct vm_area_struct *vma, unsigned long addr)
2918 {
2919         return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2920 }
2921
2922 static long vma_del_reservation(struct hstate *h,
2923                         struct vm_area_struct *vma, unsigned long addr)
2924 {
2925         return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2926 }
2927
2928 /*
2929  * This routine is called to restore reservation information on error paths.
2930  * It should ONLY be called for folios allocated via alloc_hugetlb_folio(),
2931  * and the hugetlb mutex should remain held when calling this routine.
2932  *
2933  * It handles two specific cases:
2934  * 1) A reservation was in place and the folio consumed the reservation.
2935  *    hugetlb_restore_reserve is set in the folio.
2936  * 2) No reservation was in place for the page, so hugetlb_restore_reserve is
2937  *    not set.  However, alloc_hugetlb_folio always updates the reserve map.
2938  *
2939  * In case 1, free_huge_folio later in the error path will increment the
2940  * global reserve count.  But, free_huge_folio does not have enough context
2941  * to adjust the reservation map.  This case deals primarily with private
2942  * mappings.  Adjust the reserve map here to be consistent with global
2943  * reserve count adjustments to be made by free_huge_folio.  Make sure the
2944  * reserve map indicates there is a reservation present.
2945  *
2946  * In case 2, simply undo reserve map modifications done by alloc_hugetlb_folio.
2947  */
2948 void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2949                         unsigned long address, struct folio *folio)
2950 {
2951         long rc = vma_needs_reservation(h, vma, address);
2952
2953         if (folio_test_hugetlb_restore_reserve(folio)) {
2954                 if (unlikely(rc < 0))
2955                         /*
2956                          * Rare out of memory condition in reserve map
2957                          * manipulation.  Clear hugetlb_restore_reserve so
2958                          * that global reserve count will not be incremented
2959                          * by free_huge_folio.  This will make it appear
2960                          * as though the reservation for this folio was
2961                          * consumed.  This may prevent the task from
2962                          * faulting in the folio at a later time.  This
2963                          * is better than inconsistent global huge page
2964                          * accounting of reserve counts.
2965                          */
2966                         folio_clear_hugetlb_restore_reserve(folio);
2967                 else if (rc)
2968                         (void)vma_add_reservation(h, vma, address);
2969                 else
2970                         vma_end_reservation(h, vma, address);
2971         } else {
2972                 if (!rc) {
2973                         /*
2974                          * This indicates there is an entry in the reserve map
2975                          * not added by alloc_hugetlb_folio.  We know it was added
2976                          * before the alloc_hugetlb_folio call, otherwise
2977                          * hugetlb_restore_reserve would be set on the folio.
2978                          * Remove the entry so that a subsequent allocation
2979                          * does not consume a reservation.
2980                          */
2981                         rc = vma_del_reservation(h, vma, address);
2982                         if (rc < 0)
2983                                 /*
2984                                  * VERY rare out of memory condition.  Since
2985                                  * we can not delete the entry, set
2986                                  * hugetlb_restore_reserve so that the reserve
2987                                  * count will be incremented when the folio
2988                                  * is freed.  This reserve will be consumed
2989                                  * on a subsequent allocation.
2990                                  */
2991                                 folio_set_hugetlb_restore_reserve(folio);
2992                 } else if (rc < 0) {
2993                         /*
2994                          * Rare out of memory condition from
2995                          * vma_needs_reservation call.  Memory allocation is
2996                          * only attempted if a new entry is needed.  Therefore,
2997                          * this implies there is not an entry in the
2998                          * reserve map.
2999                          *
3000                          * For shared mappings, no entry in the map indicates
3001                          * no reservation.  We are done.
3002                          */
3003                         if (!(vma->vm_flags & VM_MAYSHARE))
3004                                 /*
3005                                  * For private mappings, no entry indicates
3006                                  * a reservation is present.  Since we can
3007                                  * not add an entry, set hugetlb_restore_reserve
3008                                  * on the folio so reserve count will be
3009                                  * incremented when freed.  This reserve will
3010                                  * be consumed on a subsequent allocation.
3011                                  */
3012                                 folio_set_hugetlb_restore_reserve(folio);
3013                 } else
3014                         /*
3015                          * No reservation present, do nothing
3016                          */
3017                          vma_end_reservation(h, vma, address);
3018         }
3019 }
3020
3021 /*
3022  * alloc_and_dissolve_hugetlb_folio - Allocate a new folio and dissolve
3023  * the old one
3024  * @h: struct hstate old page belongs to
3025  * @old_folio: Old folio to dissolve
3026  * @list: List to isolate the page in case we need to
3027  * Returns 0 on success, otherwise negated error.
3028  */
3029 static int alloc_and_dissolve_hugetlb_folio(struct hstate *h,
3030                         struct folio *old_folio, struct list_head *list)
3031 {
3032         gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
3033         int nid = folio_nid(old_folio);
3034         struct folio *new_folio = NULL;
3035         int ret = 0;
3036
3037 retry:
3038         spin_lock_irq(&hugetlb_lock);
3039         if (!folio_test_hugetlb(old_folio)) {
3040                 /*
3041                  * Freed from under us. Drop new_folio too.
3042                  */
3043                 goto free_new;
3044         } else if (folio_ref_count(old_folio)) {
3045                 bool isolated;
3046
3047                 /*
3048                  * Someone has grabbed the folio, try to isolate it here.
3049                  * Fail with -EBUSY if not possible.
3050                  */
3051                 spin_unlock_irq(&hugetlb_lock);
3052                 isolated = isolate_hugetlb(old_folio, list);
3053                 ret = isolated ? 0 : -EBUSY;
3054                 spin_lock_irq(&hugetlb_lock);
3055                 goto free_new;
3056         } else if (!folio_test_hugetlb_freed(old_folio)) {
3057                 /*
3058                  * Folio's refcount is 0 but it has not been enqueued in the
3059                  * freelist yet. Race window is small, so we can succeed here if
3060                  * we retry.
3061                  */
3062                 spin_unlock_irq(&hugetlb_lock);
3063                 cond_resched();
3064                 goto retry;
3065         } else {
3066                 if (!new_folio) {
3067                         spin_unlock_irq(&hugetlb_lock);
3068                         new_folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid,
3069                                                               NULL, NULL);
3070                         if (!new_folio)
3071                                 return -ENOMEM;
3072                         __prep_new_hugetlb_folio(h, new_folio);
3073                         goto retry;
3074                 }
3075
3076                 /*
3077                  * Ok, old_folio is still a genuine free hugepage. Remove it from
3078                  * the freelist and decrease the counters. These will be
3079                  * incremented again when calling __prep_account_new_huge_page()
3080                  * and enqueue_hugetlb_folio() for new_folio. The counters will
3081                  * remain stable since this happens under the lock.
3082                  */
3083                 remove_hugetlb_folio(h, old_folio, false);
3084
3085                 /*
3086                  * Ref count on new_folio is already zero as it was dropped
3087                  * earlier.  It can be directly added to the pool free list.
3088                  */
3089                 __prep_account_new_huge_page(h, nid);
3090                 enqueue_hugetlb_folio(h, new_folio);
3091
3092                 /*
3093                  * Folio has been replaced, we can safely free the old one.
3094                  */
3095                 spin_unlock_irq(&hugetlb_lock);
3096                 update_and_free_hugetlb_folio(h, old_folio, false);
3097         }
3098
3099         return ret;
3100
3101 free_new:
3102         spin_unlock_irq(&hugetlb_lock);
3103         if (new_folio) {
3104                 /* Folio has a zero ref count, but needs a ref to be freed */
3105                 folio_ref_unfreeze(new_folio, 1);
3106                 update_and_free_hugetlb_folio(h, new_folio, false);
3107         }
3108
3109         return ret;
3110 }
3111
3112 int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
3113 {
3114         struct hstate *h;
3115         struct folio *folio = page_folio(page);
3116         int ret = -EBUSY;
3117
3118         /*
3119          * The page might have been dissolved from under our feet, so make sure
3120          * to carefully check the state under the lock.
3121          * Return success when racing as if we dissolved the page ourselves.
3122          */
3123         spin_lock_irq(&hugetlb_lock);
3124         if (folio_test_hugetlb(folio)) {
3125                 h = folio_hstate(folio);
3126         } else {
3127                 spin_unlock_irq(&hugetlb_lock);
3128                 return 0;
3129         }
3130         spin_unlock_irq(&hugetlb_lock);
3131
3132         /*
3133          * Fence off gigantic pages as there is a cyclic dependency between
3134          * alloc_contig_range and them. Return -ENOMEM as this has the effect
3135          * of bailing out right away without further retrying.
3136          */
3137         if (hstate_is_gigantic(h))
3138                 return -ENOMEM;
3139
3140         if (folio_ref_count(folio) && isolate_hugetlb(folio, list))
3141                 ret = 0;
3142         else if (!folio_ref_count(folio))
3143                 ret = alloc_and_dissolve_hugetlb_folio(h, folio, list);
3144
3145         return ret;
3146 }
3147
3148 struct folio *alloc_hugetlb_folio(struct vm_area_struct *vma,
3149                                     unsigned long addr, int avoid_reserve)
3150 {
3151         struct hugepage_subpool *spool = subpool_vma(vma);
3152         struct hstate *h = hstate_vma(vma);
3153         struct folio *folio;
3154         long map_chg, map_commit, nr_pages = pages_per_huge_page(h);
3155         long gbl_chg;
3156         int memcg_charge_ret, ret, idx;
3157         struct hugetlb_cgroup *h_cg = NULL;
3158         struct mem_cgroup *memcg;
3159         bool deferred_reserve;
3160         gfp_t gfp = htlb_alloc_mask(h) | __GFP_RETRY_MAYFAIL;
3161
3162         memcg = get_mem_cgroup_from_current();
3163         memcg_charge_ret = mem_cgroup_hugetlb_try_charge(memcg, gfp, nr_pages);
3164         if (memcg_charge_ret == -ENOMEM) {
3165                 mem_cgroup_put(memcg);
3166                 return ERR_PTR(-ENOMEM);
3167         }
3168
3169         idx = hstate_index(h);
3170         /*
3171          * Examine the region/reserve map to determine if the process
3172          * has a reservation for the page to be allocated.  A return
3173          * code of zero indicates a reservation exists (no change).
3174          */
3175         map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
3176         if (map_chg < 0) {
3177                 if (!memcg_charge_ret)
3178                         mem_cgroup_cancel_charge(memcg, nr_pages);
3179                 mem_cgroup_put(memcg);
3180                 return ERR_PTR(-ENOMEM);
3181         }
3182
3183         /*
3184          * Processes that did not create the mapping will have no
3185          * reserves as indicated by the region/reserve map. Check
3186          * that the allocation will not exceed the subpool limit.
3187          * Allocations for MAP_NORESERVE mappings also need to be
3188          * checked against any subpool limit.
3189          */
3190         if (map_chg || avoid_reserve) {
3191                 gbl_chg = hugepage_subpool_get_pages(spool, 1);
3192                 if (gbl_chg < 0)
3193                         goto out_end_reservation;
3194
3195                 /*
3196                  * Even though there was no reservation in the region/reserve
3197                  * map, there could be reservations associated with the
3198                  * subpool that can be used.  This would be indicated if the
3199                  * return value of hugepage_subpool_get_pages() is zero.
3200                  * However, if avoid_reserve is specified we still avoid even
3201                  * the subpool reservations.
3202                  */
3203                 if (avoid_reserve)
3204                         gbl_chg = 1;
3205         }
3206
3207         /* If this allocation is not consuming a reservation, charge it now.
3208          */
3209         deferred_reserve = map_chg || avoid_reserve;
3210         if (deferred_reserve) {
3211                 ret = hugetlb_cgroup_charge_cgroup_rsvd(
3212                         idx, pages_per_huge_page(h), &h_cg);
3213                 if (ret)
3214                         goto out_subpool_put;
3215         }
3216
3217         ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
3218         if (ret)
3219                 goto out_uncharge_cgroup_reservation;
3220
3221         spin_lock_irq(&hugetlb_lock);
3222         /*
3223          * glb_chg is passed to indicate whether or not a page must be taken
3224          * from the global free pool (global change).  gbl_chg == 0 indicates
3225          * a reservation exists for the allocation.
3226          */
3227         folio = dequeue_hugetlb_folio_vma(h, vma, addr, avoid_reserve, gbl_chg);
3228         if (!folio) {
3229                 spin_unlock_irq(&hugetlb_lock);
3230                 folio = alloc_buddy_hugetlb_folio_with_mpol(h, vma, addr);
3231                 if (!folio)
3232                         goto out_uncharge_cgroup;
3233                 spin_lock_irq(&hugetlb_lock);
3234                 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
3235                         folio_set_hugetlb_restore_reserve(folio);
3236                         h->resv_huge_pages--;
3237                 }
3238                 list_add(&folio->lru, &h->hugepage_activelist);
3239                 folio_ref_unfreeze(folio, 1);
3240                 /* Fall through */
3241         }
3242
3243         hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, folio);
3244         /* If allocation is not consuming a reservation, also store the
3245          * hugetlb_cgroup pointer on the page.
3246          */
3247         if (deferred_reserve) {
3248                 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
3249                                                   h_cg, folio);
3250         }
3251
3252         spin_unlock_irq(&hugetlb_lock);
3253
3254         hugetlb_set_folio_subpool(folio, spool);
3255
3256         map_commit = vma_commit_reservation(h, vma, addr);
3257         if (unlikely(map_chg > map_commit)) {
3258                 /*
3259                  * The page was added to the reservation map between
3260                  * vma_needs_reservation and vma_commit_reservation.
3261                  * This indicates a race with hugetlb_reserve_pages.
3262                  * Adjust for the subpool count incremented above AND
3263                  * in hugetlb_reserve_pages for the same page.  Also,
3264                  * the reservation count added in hugetlb_reserve_pages
3265                  * no longer applies.
3266                  */
3267                 long rsv_adjust;
3268
3269                 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
3270                 hugetlb_acct_memory(h, -rsv_adjust);
3271                 if (deferred_reserve)
3272                         hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),
3273                                         pages_per_huge_page(h), folio);
3274         }
3275
3276         if (!memcg_charge_ret)
3277                 mem_cgroup_commit_charge(folio, memcg);
3278         mem_cgroup_put(memcg);
3279
3280         return folio;
3281
3282 out_uncharge_cgroup:
3283         hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
3284 out_uncharge_cgroup_reservation:
3285         if (deferred_reserve)
3286                 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
3287                                                     h_cg);
3288 out_subpool_put:
3289         if (map_chg || avoid_reserve)
3290                 hugepage_subpool_put_pages(spool, 1);
3291 out_end_reservation:
3292         vma_end_reservation(h, vma, addr);
3293         if (!memcg_charge_ret)
3294                 mem_cgroup_cancel_charge(memcg, nr_pages);
3295         mem_cgroup_put(memcg);
3296         return ERR_PTR(-ENOSPC);
3297 }
3298
3299 int alloc_bootmem_huge_page(struct hstate *h, int nid)
3300         __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
3301 int __alloc_bootmem_huge_page(struct hstate *h, int nid)
3302 {
3303         struct huge_bootmem_page *m = NULL; /* initialize for clang */
3304         int nr_nodes, node = nid;
3305
3306         /* do node specific alloc */
3307         if (nid != NUMA_NO_NODE) {
3308                 m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h),
3309                                 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
3310                 if (!m)
3311                         return 0;
3312                 goto found;
3313         }
3314         /* allocate from next node when distributing huge pages */
3315         for_each_node_mask_to_alloc(&h->next_nid_to_alloc, nr_nodes, node, &node_states[N_MEMORY]) {
3316                 m = memblock_alloc_try_nid_raw(
3317                                 huge_page_size(h), huge_page_size(h),
3318                                 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
3319                 /*
3320                  * Use the beginning of the huge page to store the
3321                  * huge_bootmem_page struct (until gather_bootmem
3322                  * puts them into the mem_map).
3323                  */
3324                 if (!m)
3325                         return 0;
3326                 goto found;
3327         }
3328
3329 found:
3330
3331         /*
3332          * Only initialize the head struct page in memmap_init_reserved_pages,
3333          * rest of the struct pages will be initialized by the HugeTLB
3334          * subsystem itself.
3335          * The head struct page is used to get folio information by the HugeTLB
3336          * subsystem like zone id and node id.
3337          */
3338         memblock_reserved_mark_noinit(virt_to_phys((void *)m + PAGE_SIZE),
3339                 huge_page_size(h) - PAGE_SIZE);
3340         /* Put them into a private list first because mem_map is not up yet */
3341         INIT_LIST_HEAD(&m->list);
3342         list_add(&m->list, &huge_boot_pages[node]);
3343         m->hstate = h;
3344         return 1;
3345 }
3346
3347 /* Initialize [start_page:end_page_number] tail struct pages of a hugepage */
3348 static void __init hugetlb_folio_init_tail_vmemmap(struct folio *folio,
3349                                         unsigned long start_page_number,
3350                                         unsigned long end_page_number)
3351 {
3352         enum zone_type zone = zone_idx(folio_zone(folio));
3353         int nid = folio_nid(folio);
3354         unsigned long head_pfn = folio_pfn(folio);
3355         unsigned long pfn, end_pfn = head_pfn + end_page_number;
3356         int ret;
3357
3358         for (pfn = head_pfn + start_page_number; pfn < end_pfn; pfn++) {
3359                 struct page *page = pfn_to_page(pfn);
3360
3361                 __init_single_page(page, pfn, zone, nid);
3362                 prep_compound_tail((struct page *)folio, pfn - head_pfn);
3363                 ret = page_ref_freeze(page, 1);
3364                 VM_BUG_ON(!ret);
3365         }
3366 }
3367
3368 static void __init hugetlb_folio_init_vmemmap(struct folio *folio,
3369                                               struct hstate *h,
3370                                               unsigned long nr_pages)
3371 {
3372         int ret;
3373
3374         /* Prepare folio head */
3375         __folio_clear_reserved(folio);
3376         __folio_set_head(folio);
3377         ret = folio_ref_freeze(folio, 1);
3378         VM_BUG_ON(!ret);
3379         /* Initialize the necessary tail struct pages */
3380         hugetlb_folio_init_tail_vmemmap(folio, 1, nr_pages);
3381         prep_compound_head((struct page *)folio, huge_page_order(h));
3382 }
3383
3384 static void __init prep_and_add_bootmem_folios(struct hstate *h,
3385                                         struct list_head *folio_list)
3386 {
3387         unsigned long flags;
3388         struct folio *folio, *tmp_f;
3389
3390         /* Send list for bulk vmemmap optimization processing */
3391         hugetlb_vmemmap_optimize_folios(h, folio_list);
3392
3393         list_for_each_entry_safe(folio, tmp_f, folio_list, lru) {
3394                 if (!folio_test_hugetlb_vmemmap_optimized(folio)) {
3395                         /*
3396                          * If HVO fails, initialize all tail struct pages
3397                          * We do not worry about potential long lock hold
3398                          * time as this is early in boot and there should
3399                          * be no contention.
3400                          */
3401                         hugetlb_folio_init_tail_vmemmap(folio,
3402                                         HUGETLB_VMEMMAP_RESERVE_PAGES,
3403                                         pages_per_huge_page(h));
3404                 }
3405                 /* Subdivide locks to achieve better parallel performance */
3406                 spin_lock_irqsave(&hugetlb_lock, flags);
3407                 __prep_account_new_huge_page(h, folio_nid(folio));
3408                 enqueue_hugetlb_folio(h, folio);
3409                 spin_unlock_irqrestore(&hugetlb_lock, flags);
3410         }
3411 }
3412
3413 /*
3414  * Put bootmem huge pages into the standard lists after mem_map is up.
3415  * Note: This only applies to gigantic (order > MAX_PAGE_ORDER) pages.
3416  */
3417 static void __init gather_bootmem_prealloc_node(unsigned long nid)
3418 {
3419         LIST_HEAD(folio_list);
3420         struct huge_bootmem_page *m;
3421         struct hstate *h = NULL, *prev_h = NULL;
3422
3423         list_for_each_entry(m, &huge_boot_pages[nid], list) {
3424                 struct page *page = virt_to_page(m);
3425                 struct folio *folio = (void *)page;
3426
3427                 h = m->hstate;
3428                 /*
3429                  * It is possible to have multiple huge page sizes (hstates)
3430                  * in this list.  If so, process each size separately.
3431                  */
3432                 if (h != prev_h && prev_h != NULL)
3433                         prep_and_add_bootmem_folios(prev_h, &folio_list);
3434                 prev_h = h;
3435
3436                 VM_BUG_ON(!hstate_is_gigantic(h));
3437                 WARN_ON(folio_ref_count(folio) != 1);
3438
3439                 hugetlb_folio_init_vmemmap(folio, h,
3440                                            HUGETLB_VMEMMAP_RESERVE_PAGES);
3441                 init_new_hugetlb_folio(h, folio);
3442                 list_add(&folio->lru, &folio_list);
3443
3444                 /*
3445                  * We need to restore the 'stolen' pages to totalram_pages
3446                  * in order to fix confusing memory reports from free(1) and
3447                  * other side-effects, like CommitLimit going negative.
3448                  */
3449                 adjust_managed_page_count(page, pages_per_huge_page(h));
3450                 cond_resched();
3451         }
3452
3453         prep_and_add_bootmem_folios(h, &folio_list);
3454 }
3455
3456 static void __init gather_bootmem_prealloc_parallel(unsigned long start,
3457                                                     unsigned long end, void *arg)
3458 {
3459         int nid;
3460
3461         for (nid = start; nid < end; nid++)
3462                 gather_bootmem_prealloc_node(nid);
3463 }
3464
3465 static void __init gather_bootmem_prealloc(void)
3466 {
3467         struct padata_mt_job job = {
3468                 .thread_fn      = gather_bootmem_prealloc_parallel,
3469                 .fn_arg         = NULL,
3470                 .start          = 0,
3471                 .size           = num_node_state(N_MEMORY),
3472                 .align          = 1,
3473                 .min_chunk      = 1,
3474                 .max_threads    = num_node_state(N_MEMORY),
3475                 .numa_aware     = true,
3476         };
3477
3478         padata_do_multithreaded(&job);
3479 }
3480
3481 static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid)
3482 {
3483         unsigned long i;
3484         char buf[32];
3485
3486         for (i = 0; i < h->max_huge_pages_node[nid]; ++i) {
3487                 if (hstate_is_gigantic(h)) {
3488                         if (!alloc_bootmem_huge_page(h, nid))
3489                                 break;
3490                 } else {
3491                         struct folio *folio;
3492                         gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
3493
3494                         folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid,
3495                                         &node_states[N_MEMORY], NULL);
3496                         if (!folio)
3497                                 break;
3498                         free_huge_folio(folio); /* free it into the hugepage allocator */
3499                 }
3500                 cond_resched();
3501         }
3502         if (i == h->max_huge_pages_node[nid])
3503                 return;
3504
3505         string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3506         pr_warn("HugeTLB: allocating %u of page size %s failed node%d.  Only allocated %lu hugepages.\n",
3507                 h->max_huge_pages_node[nid], buf, nid, i);
3508         h->max_huge_pages -= (h->max_huge_pages_node[nid] - i);
3509         h->max_huge_pages_node[nid] = i;
3510 }
3511
3512 static bool __init hugetlb_hstate_alloc_pages_specific_nodes(struct hstate *h)
3513 {
3514         int i;
3515         bool node_specific_alloc = false;
3516
3517         for_each_online_node(i) {
3518                 if (h->max_huge_pages_node[i] > 0) {
3519                         hugetlb_hstate_alloc_pages_onenode(h, i);
3520                         node_specific_alloc = true;
3521                 }
3522         }
3523
3524         return node_specific_alloc;
3525 }
3526
3527 static void __init hugetlb_hstate_alloc_pages_errcheck(unsigned long allocated, struct hstate *h)
3528 {
3529         if (allocated < h->max_huge_pages) {
3530                 char buf[32];
3531
3532                 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3533                 pr_warn("HugeTLB: allocating %lu of page size %s failed.  Only allocated %lu hugepages.\n",
3534                         h->max_huge_pages, buf, allocated);
3535                 h->max_huge_pages = allocated;
3536         }
3537 }
3538
3539 static void __init hugetlb_pages_alloc_boot_node(unsigned long start, unsigned long end, void *arg)
3540 {
3541         struct hstate *h = (struct hstate *)arg;
3542         int i, num = end - start;
3543         nodemask_t node_alloc_noretry;
3544         LIST_HEAD(folio_list);
3545         int next_node = first_online_node;
3546
3547         /* Bit mask controlling how hard we retry per-node allocations.*/
3548         nodes_clear(node_alloc_noretry);
3549
3550         for (i = 0; i < num; ++i) {
3551                 struct folio *folio = alloc_pool_huge_folio(h, &node_states[N_MEMORY],
3552                                                 &node_alloc_noretry, &next_node);
3553                 if (!folio)
3554                         break;
3555
3556                 list_move(&folio->lru, &folio_list);
3557                 cond_resched();
3558         }
3559
3560         prep_and_add_allocated_folios(h, &folio_list);
3561 }
3562
3563 static unsigned long __init hugetlb_gigantic_pages_alloc_boot(struct hstate *h)
3564 {
3565         unsigned long i;
3566
3567         for (i = 0; i < h->max_huge_pages; ++i) {
3568                 if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE))
3569                         break;
3570                 cond_resched();
3571         }
3572
3573         return i;
3574 }
3575
3576 static unsigned long __init hugetlb_pages_alloc_boot(struct hstate *h)
3577 {
3578         struct padata_mt_job job = {
3579                 .fn_arg         = h,
3580                 .align          = 1,
3581                 .numa_aware     = true
3582         };
3583
3584         job.thread_fn   = hugetlb_pages_alloc_boot_node;
3585         job.start       = 0;
3586         job.size        = h->max_huge_pages;
3587
3588         /*
3589          * job.max_threads is twice the num_node_state(N_MEMORY),
3590          *
3591          * Tests below indicate that a multiplier of 2 significantly improves
3592          * performance, and although larger values also provide improvements,
3593          * the gains are marginal.
3594          *
3595          * Therefore, choosing 2 as the multiplier strikes a good balance between
3596          * enhancing parallel processing capabilities and maintaining efficient
3597          * resource management.
3598          *
3599          * +------------+-------+-------+-------+-------+-------+
3600          * | multiplier |   1   |   2   |   3   |   4   |   5   |
3601          * +------------+-------+-------+-------+-------+-------+
3602          * | 256G 2node | 358ms | 215ms | 157ms | 134ms | 126ms |
3603          * | 2T   4node | 979ms | 679ms | 543ms | 489ms | 481ms |
3604          * | 50G  2node | 71ms  | 44ms  | 37ms  | 30ms  | 31ms  |
3605          * +------------+-------+-------+-------+-------+-------+
3606          */
3607         job.max_threads = num_node_state(N_MEMORY) * 2;
3608         job.min_chunk   = h->max_huge_pages / num_node_state(N_MEMORY) / 2;
3609         padata_do_multithreaded(&job);
3610
3611         return h->nr_huge_pages;
3612 }
3613
3614 /*
3615  * NOTE: this routine is called in different contexts for gigantic and
3616  * non-gigantic pages.
3617  * - For gigantic pages, this is called early in the boot process and
3618  *   pages are allocated from memblock allocated or something similar.
3619  *   Gigantic pages are actually added to pools later with the routine
3620  *   gather_bootmem_prealloc.
3621  * - For non-gigantic pages, this is called later in the boot process after
3622  *   all of mm is up and functional.  Pages are allocated from buddy and
3623  *   then added to hugetlb pools.
3624  */
3625 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
3626 {
3627         unsigned long allocated;
3628         static bool initialized __initdata;
3629
3630         /* skip gigantic hugepages allocation if hugetlb_cma enabled */
3631         if (hstate_is_gigantic(h) && hugetlb_cma_size) {
3632                 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
3633                 return;
3634         }
3635
3636         /* hugetlb_hstate_alloc_pages will be called many times, initialize huge_boot_pages once */
3637         if (!initialized) {
3638                 int i = 0;
3639
3640                 for (i = 0; i < MAX_NUMNODES; i++)
3641                         INIT_LIST_HEAD(&huge_boot_pages[i]);
3642                 initialized = true;
3643         }
3644
3645         /* do node specific alloc */
3646         if (hugetlb_hstate_alloc_pages_specific_nodes(h))
3647                 return;
3648
3649         /* below will do all node balanced alloc */
3650         if (hstate_is_gigantic(h))
3651                 allocated = hugetlb_gigantic_pages_alloc_boot(h);
3652         else
3653                 allocated = hugetlb_pages_alloc_boot(h);
3654
3655         hugetlb_hstate_alloc_pages_errcheck(allocated, h);
3656 }
3657
3658 static void __init hugetlb_init_hstates(void)
3659 {
3660         struct hstate *h, *h2;
3661
3662         for_each_hstate(h) {
3663                 /* oversize hugepages were init'ed in early boot */
3664                 if (!hstate_is_gigantic(h))
3665                         hugetlb_hstate_alloc_pages(h);
3666
3667                 /*
3668                  * Set demote order for each hstate.  Note that
3669                  * h->demote_order is initially 0.
3670                  * - We can not demote gigantic pages if runtime freeing
3671                  *   is not supported, so skip this.
3672                  * - If CMA allocation is possible, we can not demote
3673                  *   HUGETLB_PAGE_ORDER or smaller size pages.
3674                  */
3675                 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3676                         continue;
3677                 if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER)
3678                         continue;
3679                 for_each_hstate(h2) {
3680                         if (h2 == h)
3681                                 continue;
3682                         if (h2->order < h->order &&
3683                             h2->order > h->demote_order)
3684                                 h->demote_order = h2->order;
3685                 }
3686         }
3687 }
3688
3689 static void __init report_hugepages(void)
3690 {
3691         struct hstate *h;
3692
3693         for_each_hstate(h) {
3694                 char buf[32];
3695
3696                 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3697                 pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n",
3698                         buf, h->free_huge_pages);
3699                 pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n",
3700                         hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf);
3701         }
3702 }
3703
3704 #ifdef CONFIG_HIGHMEM
3705 static void try_to_free_low(struct hstate *h, unsigned long count,
3706                                                 nodemask_t *nodes_allowed)
3707 {
3708         int i;
3709         LIST_HEAD(page_list);
3710
3711         lockdep_assert_held(&hugetlb_lock);
3712         if (hstate_is_gigantic(h))
3713                 return;
3714
3715         /*
3716          * Collect pages to be freed on a list, and free after dropping lock
3717          */
3718         for_each_node_mask(i, *nodes_allowed) {
3719                 struct folio *folio, *next;
3720                 struct list_head *freel = &h->hugepage_freelists[i];
3721                 list_for_each_entry_safe(folio, next, freel, lru) {
3722                         if (count >= h->nr_huge_pages)
3723                                 goto out;
3724                         if (folio_test_highmem(folio))
3725                                 continue;
3726                         remove_hugetlb_folio(h, folio, false);
3727                         list_add(&folio->lru, &page_list);
3728                 }
3729         }
3730
3731 out:
3732         spin_unlock_irq(&hugetlb_lock);
3733         update_and_free_pages_bulk(h, &page_list);
3734         spin_lock_irq(&hugetlb_lock);
3735 }
3736 #else
3737 static inline void try_to_free_low(struct hstate *h, unsigned long count,
3738                                                 nodemask_t *nodes_allowed)
3739 {
3740 }
3741 #endif
3742
3743 /*
3744  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
3745  * balanced by operating on them in a round-robin fashion.
3746  * Returns 1 if an adjustment was made.
3747  */
3748 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
3749                                 int delta)
3750 {
3751         int nr_nodes, node;
3752
3753         lockdep_assert_held(&hugetlb_lock);
3754         VM_BUG_ON(delta != -1 && delta != 1);
3755
3756         if (delta < 0) {
3757                 for_each_node_mask_to_alloc(&h->next_nid_to_alloc, nr_nodes, node, nodes_allowed) {
3758                         if (h->surplus_huge_pages_node[node])
3759                                 goto found;
3760                 }
3761         } else {
3762                 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3763                         if (h->surplus_huge_pages_node[node] <
3764                                         h->nr_huge_pages_node[node])
3765                                 goto found;
3766                 }
3767         }
3768         return 0;
3769
3770 found:
3771         h->surplus_huge_pages += delta;
3772         h->surplus_huge_pages_node[node] += delta;
3773         return 1;
3774 }
3775
3776 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
3777 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
3778                               nodemask_t *nodes_allowed)
3779 {
3780         unsigned long min_count;
3781         unsigned long allocated;
3782         struct folio *folio;
3783         LIST_HEAD(page_list);
3784         NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
3785
3786         /*
3787          * Bit mask controlling how hard we retry per-node allocations.
3788          * If we can not allocate the bit mask, do not attempt to allocate
3789          * the requested huge pages.
3790          */
3791         if (node_alloc_noretry)
3792                 nodes_clear(*node_alloc_noretry);
3793         else
3794                 return -ENOMEM;
3795
3796         /*
3797          * resize_lock mutex prevents concurrent adjustments to number of
3798          * pages in hstate via the proc/sysfs interfaces.
3799          */
3800         mutex_lock(&h->resize_lock);
3801         flush_free_hpage_work(h);
3802         spin_lock_irq(&hugetlb_lock);
3803
3804         /*
3805          * Check for a node specific request.
3806          * Changing node specific huge page count may require a corresponding
3807          * change to the global count.  In any case, the passed node mask
3808          * (nodes_allowed) will restrict alloc/free to the specified node.
3809          */
3810         if (nid != NUMA_NO_NODE) {
3811                 unsigned long old_count = count;
3812
3813                 count += persistent_huge_pages(h) -
3814                          (h->nr_huge_pages_node[nid] -
3815                           h->surplus_huge_pages_node[nid]);
3816                 /*
3817                  * User may have specified a large count value which caused the
3818                  * above calculation to overflow.  In this case, they wanted
3819                  * to allocate as many huge pages as possible.  Set count to
3820                  * largest possible value to align with their intention.
3821                  */
3822                 if (count < old_count)
3823                         count = ULONG_MAX;
3824         }
3825
3826         /*
3827          * Gigantic pages runtime allocation depend on the capability for large
3828          * page range allocation.
3829          * If the system does not provide this feature, return an error when
3830          * the user tries to allocate gigantic pages but let the user free the
3831          * boottime allocated gigantic pages.
3832          */
3833         if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3834                 if (count > persistent_huge_pages(h)) {
3835                         spin_unlock_irq(&hugetlb_lock);
3836                         mutex_unlock(&h->resize_lock);
3837                         NODEMASK_FREE(node_alloc_noretry);
3838                         return -EINVAL;
3839                 }
3840                 /* Fall through to decrease pool */
3841         }
3842
3843         /*
3844          * Increase the pool size
3845          * First take pages out of surplus state.  Then make up the
3846          * remaining difference by allocating fresh huge pages.
3847          *
3848          * We might race with alloc_surplus_hugetlb_folio() here and be unable
3849          * to convert a surplus huge page to a normal huge page. That is
3850          * not critical, though, it just means the overall size of the
3851          * pool might be one hugepage larger than it needs to be, but
3852          * within all the constraints specified by the sysctls.
3853          */
3854         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3855                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
3856                         break;
3857         }
3858
3859         allocated = 0;
3860         while (count > (persistent_huge_pages(h) + allocated)) {
3861                 /*
3862                  * If this allocation races such that we no longer need the
3863                  * page, free_huge_folio will handle it by freeing the page
3864                  * and reducing the surplus.
3865                  */
3866                 spin_unlock_irq(&hugetlb_lock);
3867
3868                 /* yield cpu to avoid soft lockup */
3869                 cond_resched();
3870
3871                 folio = alloc_pool_huge_folio(h, nodes_allowed,
3872                                                 node_alloc_noretry,
3873                                                 &h->next_nid_to_alloc);
3874                 if (!folio) {
3875                         prep_and_add_allocated_folios(h, &page_list);
3876                         spin_lock_irq(&hugetlb_lock);
3877                         goto out;
3878                 }
3879
3880                 list_add(&folio->lru, &page_list);
3881                 allocated++;
3882
3883                 /* Bail for signals. Probably ctrl-c from user */
3884                 if (signal_pending(current)) {
3885                         prep_and_add_allocated_folios(h, &page_list);
3886                         spin_lock_irq(&hugetlb_lock);
3887                         goto out;
3888                 }
3889
3890                 spin_lock_irq(&hugetlb_lock);
3891         }
3892
3893         /* Add allocated pages to the pool */
3894         if (!list_empty(&page_list)) {
3895                 spin_unlock_irq(&hugetlb_lock);
3896                 prep_and_add_allocated_folios(h, &page_list);
3897                 spin_lock_irq(&hugetlb_lock);
3898         }
3899
3900         /*
3901          * Decrease the pool size
3902          * First return free pages to the buddy allocator (being careful
3903          * to keep enough around to satisfy reservations).  Then place
3904          * pages into surplus state as needed so the pool will shrink
3905          * to the desired size as pages become free.
3906          *
3907          * By placing pages into the surplus state independent of the
3908          * overcommit value, we are allowing the surplus pool size to
3909          * exceed overcommit. There are few sane options here. Since
3910          * alloc_surplus_hugetlb_folio() is checking the global counter,
3911          * though, we'll note that we're not allowed to exceed surplus
3912          * and won't grow the pool anywhere else. Not until one of the
3913          * sysctls are changed, or the surplus pages go out of use.
3914          */
3915         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3916         min_count = max(count, min_count);
3917         try_to_free_low(h, min_count, nodes_allowed);
3918
3919         /*
3920          * Collect pages to be removed on list without dropping lock
3921          */
3922         while (min_count < persistent_huge_pages(h)) {
3923                 folio = remove_pool_hugetlb_folio(h, nodes_allowed, 0);
3924                 if (!folio)
3925                         break;
3926
3927                 list_add(&folio->lru, &page_list);
3928         }
3929         /* free the pages after dropping lock */
3930         spin_unlock_irq(&hugetlb_lock);
3931         update_and_free_pages_bulk(h, &page_list);
3932         flush_free_hpage_work(h);
3933         spin_lock_irq(&hugetlb_lock);
3934
3935         while (count < persistent_huge_pages(h)) {
3936                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
3937                         break;
3938         }
3939 out:
3940         h->max_huge_pages = persistent_huge_pages(h);
3941         spin_unlock_irq(&hugetlb_lock);
3942         mutex_unlock(&h->resize_lock);
3943
3944         NODEMASK_FREE(node_alloc_noretry);
3945
3946         return 0;
3947 }
3948
3949 static int demote_free_hugetlb_folio(struct hstate *h, struct folio *folio)
3950 {
3951         int i, nid = folio_nid(folio);
3952         struct hstate *target_hstate;
3953         struct page *subpage;
3954         struct folio *inner_folio;
3955         int rc = 0;
3956
3957         target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order);
3958
3959         remove_hugetlb_folio_for_demote(h, folio, false);
3960         spin_unlock_irq(&hugetlb_lock);
3961
3962         /*
3963          * If vmemmap already existed for folio, the remove routine above would
3964          * have cleared the hugetlb folio flag.  Hence the folio is technically
3965          * no longer a hugetlb folio.  hugetlb_vmemmap_restore_folio can only be
3966          * passed hugetlb folios and will BUG otherwise.
3967          */
3968         if (folio_test_hugetlb(folio)) {
3969                 rc = hugetlb_vmemmap_restore_folio(h, folio);
3970                 if (rc) {
3971                         /* Allocation of vmemmmap failed, we can not demote folio */
3972                         spin_lock_irq(&hugetlb_lock);
3973                         folio_ref_unfreeze(folio, 1);
3974                         add_hugetlb_folio(h, folio, false);
3975                         return rc;
3976                 }
3977         }
3978
3979         /*
3980          * Use destroy_compound_hugetlb_folio_for_demote for all huge page
3981          * sizes as it will not ref count folios.
3982          */
3983         destroy_compound_hugetlb_folio_for_demote(folio, huge_page_order(h));
3984
3985         /*
3986          * Taking target hstate mutex synchronizes with set_max_huge_pages.
3987          * Without the mutex, pages added to target hstate could be marked
3988          * as surplus.
3989          *
3990          * Note that we already hold h->resize_lock.  To prevent deadlock,
3991          * use the convention of always taking larger size hstate mutex first.
3992          */
3993         mutex_lock(&target_hstate->resize_lock);
3994         for (i = 0; i < pages_per_huge_page(h);
3995                                 i += pages_per_huge_page(target_hstate)) {
3996                 subpage = folio_page(folio, i);
3997                 inner_folio = page_folio(subpage);
3998                 if (hstate_is_gigantic(target_hstate))
3999                         prep_compound_gigantic_folio_for_demote(inner_folio,
4000                                                         target_hstate->order);
4001                 else
4002                         prep_compound_page(subpage, target_hstate->order);
4003                 folio_change_private(inner_folio, NULL);
4004                 prep_new_hugetlb_folio(target_hstate, inner_folio, nid);
4005                 free_huge_folio(inner_folio);
4006         }
4007         mutex_unlock(&target_hstate->resize_lock);
4008
4009         spin_lock_irq(&hugetlb_lock);
4010
4011         /*
4012          * Not absolutely necessary, but for consistency update max_huge_pages
4013          * based on pool changes for the demoted page.
4014          */
4015         h->max_huge_pages--;
4016         target_hstate->max_huge_pages +=
4017                 pages_per_huge_page(h) / pages_per_huge_page(target_hstate);
4018
4019         return rc;
4020 }
4021
4022 static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
4023         __must_hold(&hugetlb_lock)
4024 {
4025         int nr_nodes, node;
4026         struct folio *folio;
4027
4028         lockdep_assert_held(&hugetlb_lock);
4029
4030         /* We should never get here if no demote order */
4031         if (!h->demote_order) {
4032                 pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n");
4033                 return -EINVAL;         /* internal error */
4034         }
4035
4036         for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
4037                 list_for_each_entry(folio, &h->hugepage_freelists[node], lru) {
4038                         if (folio_test_hwpoison(folio))
4039                                 continue;
4040                         return demote_free_hugetlb_folio(h, folio);
4041                 }
4042         }
4043
4044         /*
4045          * Only way to get here is if all pages on free lists are poisoned.
4046          * Return -EBUSY so that caller will not retry.
4047          */
4048         return -EBUSY;
4049 }
4050
4051 #define HSTATE_ATTR_RO(_name) \
4052         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
4053
4054 #define HSTATE_ATTR_WO(_name) \
4055         static struct kobj_attribute _name##_attr = __ATTR_WO(_name)
4056
4057 #define HSTATE_ATTR(_name) \
4058         static struct kobj_attribute _name##_attr = __ATTR_RW(_name)
4059
4060 static struct kobject *hugepages_kobj;
4061 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
4062
4063 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
4064
4065 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
4066 {
4067         int i;
4068
4069         for (i = 0; i < HUGE_MAX_HSTATE; i++)
4070                 if (hstate_kobjs[i] == kobj) {
4071                         if (nidp)
4072                                 *nidp = NUMA_NO_NODE;
4073                         return &hstates[i];
4074                 }
4075
4076         return kobj_to_node_hstate(kobj, nidp);
4077 }
4078
4079 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
4080                                         struct kobj_attribute *attr, char *buf)
4081 {
4082         struct hstate *h;
4083         unsigned long nr_huge_pages;
4084         int nid;
4085
4086         h = kobj_to_hstate(kobj, &nid);
4087         if (nid == NUMA_NO_NODE)
4088                 nr_huge_pages = h->nr_huge_pages;
4089         else
4090                 nr_huge_pages = h->nr_huge_pages_node[nid];
4091
4092         return sysfs_emit(buf, "%lu\n", nr_huge_pages);
4093 }
4094
4095 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
4096                                            struct hstate *h, int nid,
4097                                            unsigned long count, size_t len)
4098 {
4099         int err;
4100         nodemask_t nodes_allowed, *n_mask;
4101
4102         if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
4103                 return -EINVAL;
4104
4105         if (nid == NUMA_NO_NODE) {
4106                 /*
4107                  * global hstate attribute
4108                  */
4109                 if (!(obey_mempolicy &&
4110                                 init_nodemask_of_mempolicy(&nodes_allowed)))
4111                         n_mask = &node_states[N_MEMORY];
4112                 else
4113                         n_mask = &nodes_allowed;
4114         } else {
4115                 /*
4116                  * Node specific request.  count adjustment happens in
4117                  * set_max_huge_pages() after acquiring hugetlb_lock.
4118                  */
4119                 init_nodemask_of_node(&nodes_allowed, nid);
4120                 n_mask = &nodes_allowed;
4121         }
4122
4123         err = set_max_huge_pages(h, count, nid, n_mask);
4124
4125         return err ? err : len;
4126 }
4127
4128 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
4129                                          struct kobject *kobj, const char *buf,
4130                                          size_t len)
4131 {
4132         struct hstate *h;
4133         unsigned long count;
4134         int nid;
4135         int err;
4136
4137         err = kstrtoul(buf, 10, &count);
4138         if (err)
4139                 return err;
4140
4141         h = kobj_to_hstate(kobj, &nid);
4142         return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
4143 }
4144
4145 static ssize_t nr_hugepages_show(struct kobject *kobj,
4146                                        struct kobj_attribute *attr, char *buf)
4147 {
4148         return nr_hugepages_show_common(kobj, attr, buf);
4149 }
4150
4151 static ssize_t nr_hugepages_store(struct kobject *kobj,
4152                struct kobj_attribute *attr, const char *buf, size_t len)
4153 {
4154         return nr_hugepages_store_common(false, kobj, buf, len);
4155 }
4156 HSTATE_ATTR(nr_hugepages);
4157
4158 #ifdef CONFIG_NUMA
4159
4160 /*
4161  * hstate attribute for optionally mempolicy-based constraint on persistent
4162  * huge page alloc/free.
4163  */
4164 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
4165                                            struct kobj_attribute *attr,
4166                                            char *buf)
4167 {
4168         return nr_hugepages_show_common(kobj, attr, buf);
4169 }
4170
4171 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
4172                struct kobj_attribute *attr, const char *buf, size_t len)
4173 {
4174         return nr_hugepages_store_common(true, kobj, buf, len);
4175 }
4176 HSTATE_ATTR(nr_hugepages_mempolicy);
4177 #endif
4178
4179
4180 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
4181                                         struct kobj_attribute *attr, char *buf)
4182 {
4183         struct hstate *h = kobj_to_hstate(kobj, NULL);
4184         return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
4185 }
4186
4187 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
4188                 struct kobj_attribute *attr, const char *buf, size_t count)
4189 {
4190         int err;
4191         unsigned long input;
4192         struct hstate *h = kobj_to_hstate(kobj, NULL);
4193
4194         if (hstate_is_gigantic(h))
4195                 return -EINVAL;
4196
4197         err = kstrtoul(buf, 10, &input);
4198         if (err)
4199                 return err;
4200
4201         spin_lock_irq(&hugetlb_lock);
4202         h->nr_overcommit_huge_pages = input;
4203         spin_unlock_irq(&hugetlb_lock);
4204
4205         return count;
4206 }
4207 HSTATE_ATTR(nr_overcommit_hugepages);
4208
4209 static ssize_t free_hugepages_show(struct kobject *kobj,
4210                                         struct kobj_attribute *attr, char *buf)
4211 {
4212         struct hstate *h;
4213         unsigned long free_huge_pages;
4214         int nid;
4215
4216         h = kobj_to_hstate(kobj, &nid);
4217         if (nid == NUMA_NO_NODE)
4218                 free_huge_pages = h->free_huge_pages;
4219         else
4220                 free_huge_pages = h->free_huge_pages_node[nid];
4221
4222         return sysfs_emit(buf, "%lu\n", free_huge_pages);
4223 }
4224 HSTATE_ATTR_RO(free_hugepages);
4225
4226 static ssize_t resv_hugepages_show(struct kobject *kobj,
4227                                         struct kobj_attribute *attr, char *buf)
4228 {
4229         struct hstate *h = kobj_to_hstate(kobj, NULL);
4230         return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
4231 }
4232 HSTATE_ATTR_RO(resv_hugepages);
4233
4234 static ssize_t surplus_hugepages_show(struct kobject *kobj,
4235                                         struct kobj_attribute *attr, char *buf)
4236 {
4237         struct hstate *h;
4238         unsigned long surplus_huge_pages;
4239         int nid;
4240
4241         h = kobj_to_hstate(kobj, &nid);
4242         if (nid == NUMA_NO_NODE)
4243                 surplus_huge_pages = h->surplus_huge_pages;
4244         else
4245                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
4246
4247         return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
4248 }
4249 HSTATE_ATTR_RO(surplus_hugepages);
4250
4251 static ssize_t demote_store(struct kobject *kobj,
4252                struct kobj_attribute *attr, const char *buf, size_t len)
4253 {
4254         unsigned long nr_demote;
4255         unsigned long nr_available;
4256         nodemask_t nodes_allowed, *n_mask;
4257         struct hstate *h;
4258         int err;
4259         int nid;
4260
4261         err = kstrtoul(buf, 10, &nr_demote);
4262         if (err)
4263                 return err;
4264         h = kobj_to_hstate(kobj, &nid);
4265
4266         if (nid != NUMA_NO_NODE) {
4267                 init_nodemask_of_node(&nodes_allowed, nid);
4268                 n_mask = &nodes_allowed;
4269         } else {
4270                 n_mask = &node_states[N_MEMORY];
4271         }
4272
4273         /* Synchronize with other sysfs operations modifying huge pages */
4274         mutex_lock(&h->resize_lock);
4275         spin_lock_irq(&hugetlb_lock);
4276
4277         while (nr_demote) {
4278                 /*
4279                  * Check for available pages to demote each time thorough the
4280                  * loop as demote_pool_huge_page will drop hugetlb_lock.
4281                  */
4282                 if (nid != NUMA_NO_NODE)
4283                         nr_available = h->free_huge_pages_node[nid];
4284                 else
4285                         nr_available = h->free_huge_pages;
4286                 nr_available -= h->resv_huge_pages;
4287                 if (!nr_available)
4288                         break;
4289
4290                 err = demote_pool_huge_page(h, n_mask);
4291                 if (err)
4292                         break;
4293
4294                 nr_demote--;
4295         }
4296
4297         spin_unlock_irq(&hugetlb_lock);
4298         mutex_unlock(&h->resize_lock);
4299
4300         if (err)
4301                 return err;
4302         return len;
4303 }
4304 HSTATE_ATTR_WO(demote);
4305
4306 static ssize_t demote_size_show(struct kobject *kobj,
4307                                         struct kobj_attribute *attr, char *buf)
4308 {
4309         struct hstate *h = kobj_to_hstate(kobj, NULL);
4310         unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K;
4311
4312         return sysfs_emit(buf, "%lukB\n", demote_size);
4313 }
4314
4315 static ssize_t demote_size_store(struct kobject *kobj,
4316                                         struct kobj_attribute *attr,
4317                                         const char *buf, size_t count)
4318 {
4319         struct hstate *h, *demote_hstate;
4320         unsigned long demote_size;
4321         unsigned int demote_order;
4322
4323         demote_size = (unsigned long)memparse(buf, NULL);
4324
4325         demote_hstate = size_to_hstate(demote_size);
4326         if (!demote_hstate)
4327                 return -EINVAL;
4328         demote_order = demote_hstate->order;
4329         if (demote_order < HUGETLB_PAGE_ORDER)
4330                 return -EINVAL;
4331
4332         /* demote order must be smaller than hstate order */
4333         h = kobj_to_hstate(kobj, NULL);
4334         if (demote_order >= h->order)
4335                 return -EINVAL;
4336
4337         /* resize_lock synchronizes access to demote size and writes */
4338         mutex_lock(&h->resize_lock);
4339         h->demote_order = demote_order;
4340         mutex_unlock(&h->resize_lock);
4341
4342         return count;
4343 }
4344 HSTATE_ATTR(demote_size);
4345
4346 static struct attribute *hstate_attrs[] = {
4347         &nr_hugepages_attr.attr,
4348         &nr_overcommit_hugepages_attr.attr,
4349         &free_hugepages_attr.attr,
4350         &resv_hugepages_attr.attr,
4351         &surplus_hugepages_attr.attr,
4352 #ifdef CONFIG_NUMA
4353         &nr_hugepages_mempolicy_attr.attr,
4354 #endif
4355         NULL,
4356 };
4357
4358 static const struct attribute_group hstate_attr_group = {
4359         .attrs = hstate_attrs,
4360 };
4361
4362 static struct attribute *hstate_demote_attrs[] = {
4363         &demote_size_attr.attr,
4364         &demote_attr.attr,
4365         NULL,
4366 };
4367
4368 static const struct attribute_group hstate_demote_attr_group = {
4369         .attrs = hstate_demote_attrs,
4370 };
4371
4372 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
4373                                     struct kobject **hstate_kobjs,
4374                                     const struct attribute_group *hstate_attr_group)
4375 {
4376         int retval;
4377         int hi = hstate_index(h);
4378
4379         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
4380         if (!hstate_kobjs[hi])
4381                 return -ENOMEM;
4382
4383         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
4384         if (retval) {
4385                 kobject_put(hstate_kobjs[hi]);
4386                 hstate_kobjs[hi] = NULL;
4387                 return retval;
4388         }
4389
4390         if (h->demote_order) {
4391                 retval = sysfs_create_group(hstate_kobjs[hi],
4392                                             &hstate_demote_attr_group);
4393                 if (retval) {
4394                         pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name);
4395                         sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group);
4396                         kobject_put(hstate_kobjs[hi]);
4397                         hstate_kobjs[hi] = NULL;
4398                         return retval;
4399                 }
4400         }
4401
4402         return 0;
4403 }
4404
4405 #ifdef CONFIG_NUMA
4406 static bool hugetlb_sysfs_initialized __ro_after_init;
4407
4408 /*
4409  * node_hstate/s - associate per node hstate attributes, via their kobjects,
4410  * with node devices in node_devices[] using a parallel array.  The array
4411  * index of a node device or _hstate == node id.
4412  * This is here to avoid any static dependency of the node device driver, in
4413  * the base kernel, on the hugetlb module.
4414  */
4415 struct node_hstate {
4416         struct kobject          *hugepages_kobj;
4417         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
4418 };
4419 static struct node_hstate node_hstates[MAX_NUMNODES];
4420
4421 /*
4422  * A subset of global hstate attributes for node devices
4423  */
4424 static struct attribute *per_node_hstate_attrs[] = {
4425         &nr_hugepages_attr.attr,
4426         &free_hugepages_attr.attr,
4427         &surplus_hugepages_attr.attr,
4428         NULL,
4429 };
4430
4431 static const struct attribute_group per_node_hstate_attr_group = {
4432         .attrs = per_node_hstate_attrs,
4433 };
4434
4435 /*
4436  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
4437  * Returns node id via non-NULL nidp.
4438  */
4439 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4440 {
4441         int nid;
4442
4443         for (nid = 0; nid < nr_node_ids; nid++) {
4444                 struct node_hstate *nhs = &node_hstates[nid];
4445                 int i;
4446                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
4447                         if (nhs->hstate_kobjs[i] == kobj) {
4448                                 if (nidp)
4449                                         *nidp = nid;
4450                                 return &hstates[i];
4451                         }
4452         }
4453
4454         BUG();
4455         return NULL;
4456 }
4457
4458 /*
4459  * Unregister hstate attributes from a single node device.
4460  * No-op if no hstate attributes attached.
4461  */
4462 void hugetlb_unregister_node(struct node *node)
4463 {
4464         struct hstate *h;
4465         struct node_hstate *nhs = &node_hstates[node->dev.id];
4466
4467         if (!nhs->hugepages_kobj)
4468                 return;         /* no hstate attributes */
4469
4470         for_each_hstate(h) {
4471                 int idx = hstate_index(h);
4472                 struct kobject *hstate_kobj = nhs->hstate_kobjs[idx];
4473
4474                 if (!hstate_kobj)
4475                         continue;
4476                 if (h->demote_order)
4477                         sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group);
4478                 sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group);
4479                 kobject_put(hstate_kobj);
4480                 nhs->hstate_kobjs[idx] = NULL;
4481         }
4482
4483         kobject_put(nhs->hugepages_kobj);
4484         nhs->hugepages_kobj = NULL;
4485 }
4486
4487
4488 /*
4489  * Register hstate attributes for a single node device.
4490  * No-op if attributes already registered.
4491  */
4492 void hugetlb_register_node(struct node *node)
4493 {
4494         struct hstate *h;
4495         struct node_hstate *nhs = &node_hstates[node->dev.id];
4496         int err;
4497
4498         if (!hugetlb_sysfs_initialized)
4499                 return;
4500
4501         if (nhs->hugepages_kobj)
4502                 return;         /* already allocated */
4503
4504         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
4505                                                         &node->dev.kobj);
4506         if (!nhs->hugepages_kobj)
4507                 return;
4508
4509         for_each_hstate(h) {
4510                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
4511                                                 nhs->hstate_kobjs,
4512                                                 &per_node_hstate_attr_group);
4513                 if (err) {
4514                         pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
4515                                 h->name, node->dev.id);
4516                         hugetlb_unregister_node(node);
4517                         break;
4518                 }
4519         }
4520 }
4521
4522 /*
4523  * hugetlb init time:  register hstate attributes for all registered node
4524  * devices of nodes that have memory.  All on-line nodes should have
4525  * registered their associated device by this time.
4526  */
4527 static void __init hugetlb_register_all_nodes(void)
4528 {
4529         int nid;
4530
4531         for_each_online_node(nid)
4532                 hugetlb_register_node(node_devices[nid]);
4533 }
4534 #else   /* !CONFIG_NUMA */
4535
4536 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4537 {
4538         BUG();
4539         if (nidp)
4540                 *nidp = -1;
4541         return NULL;
4542 }
4543
4544 static void hugetlb_register_all_nodes(void) { }
4545
4546 #endif
4547
4548 #ifdef CONFIG_CMA
4549 static void __init hugetlb_cma_check(void);
4550 #else
4551 static inline __init void hugetlb_cma_check(void)
4552 {
4553 }
4554 #endif
4555
4556 static void __init hugetlb_sysfs_init(void)
4557 {
4558         struct hstate *h;
4559         int err;
4560
4561         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
4562         if (!hugepages_kobj)
4563                 return;
4564
4565         for_each_hstate(h) {
4566                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
4567                                          hstate_kobjs, &hstate_attr_group);
4568                 if (err)
4569                         pr_err("HugeTLB: Unable to add hstate %s", h->name);
4570         }
4571
4572 #ifdef CONFIG_NUMA
4573         hugetlb_sysfs_initialized = true;
4574 #endif
4575         hugetlb_register_all_nodes();
4576 }
4577
4578 #ifdef CONFIG_SYSCTL
4579 static void hugetlb_sysctl_init(void);
4580 #else
4581 static inline void hugetlb_sysctl_init(void) { }
4582 #endif
4583
4584 static int __init hugetlb_init(void)
4585 {
4586         int i;
4587
4588         BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
4589                         __NR_HPAGEFLAGS);
4590
4591         if (!hugepages_supported()) {
4592                 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
4593                         pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
4594                 return 0;
4595         }
4596
4597         /*
4598          * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists.  Some
4599          * architectures depend on setup being done here.
4600          */
4601         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
4602         if (!parsed_default_hugepagesz) {
4603                 /*
4604                  * If we did not parse a default huge page size, set
4605                  * default_hstate_idx to HPAGE_SIZE hstate. And, if the
4606                  * number of huge pages for this default size was implicitly
4607                  * specified, set that here as well.
4608                  * Note that the implicit setting will overwrite an explicit
4609                  * setting.  A warning will be printed in this case.
4610                  */
4611                 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
4612                 if (default_hstate_max_huge_pages) {
4613                         if (default_hstate.max_huge_pages) {
4614                                 char buf[32];
4615
4616                                 string_get_size(huge_page_size(&default_hstate),
4617                                         1, STRING_UNITS_2, buf, 32);
4618                                 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
4619                                         default_hstate.max_huge_pages, buf);
4620                                 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
4621                                         default_hstate_max_huge_pages);
4622                         }
4623                         default_hstate.max_huge_pages =
4624                                 default_hstate_max_huge_pages;
4625
4626                         for_each_online_node(i)
4627                                 default_hstate.max_huge_pages_node[i] =
4628                                         default_hugepages_in_node[i];
4629                 }
4630         }
4631
4632         hugetlb_cma_check();
4633         hugetlb_init_hstates();
4634         gather_bootmem_prealloc();
4635         report_hugepages();
4636
4637         hugetlb_sysfs_init();
4638         hugetlb_cgroup_file_init();
4639         hugetlb_sysctl_init();
4640
4641 #ifdef CONFIG_SMP
4642         num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
4643 #else
4644         num_fault_mutexes = 1;
4645 #endif
4646         hugetlb_fault_mutex_table =
4647                 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
4648                               GFP_KERNEL);
4649         BUG_ON(!hugetlb_fault_mutex_table);
4650
4651         for (i = 0; i < num_fault_mutexes; i++)
4652                 mutex_init(&hugetlb_fault_mutex_table[i]);
4653         return 0;
4654 }
4655 subsys_initcall(hugetlb_init);
4656
4657 /* Overwritten by architectures with more huge page sizes */
4658 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
4659 {
4660         return size == HPAGE_SIZE;
4661 }
4662
4663 void __init hugetlb_add_hstate(unsigned int order)
4664 {
4665         struct hstate *h;
4666         unsigned long i;
4667
4668         if (size_to_hstate(PAGE_SIZE << order)) {
4669                 return;
4670         }
4671         BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
4672         BUG_ON(order < order_base_2(__NR_USED_SUBPAGE));
4673         h = &hstates[hugetlb_max_hstate++];
4674         mutex_init(&h->resize_lock);
4675         h->order = order;
4676         h->mask = ~(huge_page_size(h) - 1);
4677         for (i = 0; i < MAX_NUMNODES; ++i)
4678                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
4679         INIT_LIST_HEAD(&h->hugepage_activelist);
4680         h->next_nid_to_alloc = first_memory_node;
4681         h->next_nid_to_free = first_memory_node;
4682         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
4683                                         huge_page_size(h)/SZ_1K);
4684
4685         parsed_hstate = h;
4686 }
4687
4688 bool __init __weak hugetlb_node_alloc_supported(void)
4689 {
4690         return true;
4691 }
4692
4693 static void __init hugepages_clear_pages_in_node(void)
4694 {
4695         if (!hugetlb_max_hstate) {
4696                 default_hstate_max_huge_pages = 0;
4697                 memset(default_hugepages_in_node, 0,
4698                         sizeof(default_hugepages_in_node));
4699         } else {
4700                 parsed_hstate->max_huge_pages = 0;
4701                 memset(parsed_hstate->max_huge_pages_node, 0,
4702                         sizeof(parsed_hstate->max_huge_pages_node));
4703         }
4704 }
4705
4706 /*
4707  * hugepages command line processing
4708  * hugepages normally follows a valid hugepagsz or default_hugepagsz
4709  * specification.  If not, ignore the hugepages value.  hugepages can also
4710  * be the first huge page command line  option in which case it implicitly
4711  * specifies the number of huge pages for the default size.
4712  */
4713 static int __init hugepages_setup(char *s)
4714 {
4715         unsigned long *mhp;
4716         static unsigned long *last_mhp;
4717         int node = NUMA_NO_NODE;
4718         int count;
4719         unsigned long tmp;
4720         char *p = s;
4721
4722         if (!parsed_valid_hugepagesz) {
4723                 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
4724                 parsed_valid_hugepagesz = true;
4725                 return 1;
4726         }
4727
4728         /*
4729          * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
4730          * yet, so this hugepages= parameter goes to the "default hstate".
4731          * Otherwise, it goes with the previously parsed hugepagesz or
4732          * default_hugepagesz.
4733          */
4734         else if (!hugetlb_max_hstate)
4735                 mhp = &default_hstate_max_huge_pages;
4736         else
4737                 mhp = &parsed_hstate->max_huge_pages;
4738
4739         if (mhp == last_mhp) {
4740                 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
4741                 return 1;
4742         }
4743
4744         while (*p) {
4745                 count = 0;
4746                 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4747                         goto invalid;
4748                 /* Parameter is node format */
4749                 if (p[count] == ':') {
4750                         if (!hugetlb_node_alloc_supported()) {
4751                                 pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n");
4752                                 return 1;
4753                         }
4754                         if (tmp >= MAX_NUMNODES || !node_online(tmp))
4755                                 goto invalid;
4756                         node = array_index_nospec(tmp, MAX_NUMNODES);
4757                         p += count + 1;
4758                         /* Parse hugepages */
4759                         if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4760                                 goto invalid;
4761                         if (!hugetlb_max_hstate)
4762                                 default_hugepages_in_node[node] = tmp;
4763                         else
4764                                 parsed_hstate->max_huge_pages_node[node] = tmp;
4765                         *mhp += tmp;
4766                         /* Go to parse next node*/
4767                         if (p[count] == ',')
4768                                 p += count + 1;
4769                         else
4770                                 break;
4771                 } else {
4772                         if (p != s)
4773                                 goto invalid;
4774                         *mhp = tmp;
4775                         break;
4776                 }
4777         }
4778
4779         /*
4780          * Global state is always initialized later in hugetlb_init.
4781          * But we need to allocate gigantic hstates here early to still
4782          * use the bootmem allocator.
4783          */
4784         if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
4785                 hugetlb_hstate_alloc_pages(parsed_hstate);
4786
4787         last_mhp = mhp;
4788
4789         return 1;
4790
4791 invalid:
4792         pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p);
4793         hugepages_clear_pages_in_node();
4794         return 1;
4795 }
4796 __setup("hugepages=", hugepages_setup);
4797
4798 /*
4799  * hugepagesz command line processing
4800  * A specific huge page size can only be specified once with hugepagesz.
4801  * hugepagesz is followed by hugepages on the command line.  The global
4802  * variable 'parsed_valid_hugepagesz' is used to determine if prior
4803  * hugepagesz argument was valid.
4804  */
4805 static int __init hugepagesz_setup(char *s)
4806 {
4807         unsigned long size;
4808         struct hstate *h;
4809
4810         parsed_valid_hugepagesz = false;
4811         size = (unsigned long)memparse(s, NULL);
4812
4813         if (!arch_hugetlb_valid_size(size)) {
4814                 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
4815                 return 1;
4816         }
4817
4818         h = size_to_hstate(size);
4819         if (h) {
4820                 /*
4821                  * hstate for this size already exists.  This is normally
4822                  * an error, but is allowed if the existing hstate is the
4823                  * default hstate.  More specifically, it is only allowed if
4824                  * the number of huge pages for the default hstate was not
4825                  * previously specified.
4826                  */
4827                 if (!parsed_default_hugepagesz ||  h != &default_hstate ||
4828                     default_hstate.max_huge_pages) {
4829                         pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
4830                         return 1;
4831                 }
4832
4833                 /*
4834                  * No need to call hugetlb_add_hstate() as hstate already
4835                  * exists.  But, do set parsed_hstate so that a following
4836                  * hugepages= parameter will be applied to this hstate.
4837                  */
4838                 parsed_hstate = h;
4839                 parsed_valid_hugepagesz = true;
4840                 return 1;
4841         }
4842
4843         hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4844         parsed_valid_hugepagesz = true;
4845         return 1;
4846 }
4847 __setup("hugepagesz=", hugepagesz_setup);
4848
4849 /*
4850  * default_hugepagesz command line input
4851  * Only one instance of default_hugepagesz allowed on command line.
4852  */
4853 static int __init default_hugepagesz_setup(char *s)
4854 {
4855         unsigned long size;
4856         int i;
4857
4858         parsed_valid_hugepagesz = false;
4859         if (parsed_default_hugepagesz) {
4860                 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
4861                 return 1;
4862         }
4863
4864         size = (unsigned long)memparse(s, NULL);
4865
4866         if (!arch_hugetlb_valid_size(size)) {
4867                 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
4868                 return 1;
4869         }
4870
4871         hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4872         parsed_valid_hugepagesz = true;
4873         parsed_default_hugepagesz = true;
4874         default_hstate_idx = hstate_index(size_to_hstate(size));
4875
4876         /*
4877          * The number of default huge pages (for this size) could have been
4878          * specified as the first hugetlb parameter: hugepages=X.  If so,
4879          * then default_hstate_max_huge_pages is set.  If the default huge
4880          * page size is gigantic (> MAX_PAGE_ORDER), then the pages must be
4881          * allocated here from bootmem allocator.
4882          */
4883         if (default_hstate_max_huge_pages) {
4884                 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
4885                 for_each_online_node(i)
4886                         default_hstate.max_huge_pages_node[i] =
4887                                 default_hugepages_in_node[i];
4888                 if (hstate_is_gigantic(&default_hstate))
4889                         hugetlb_hstate_alloc_pages(&default_hstate);
4890                 default_hstate_max_huge_pages = 0;
4891         }
4892
4893         return 1;
4894 }
4895 __setup("default_hugepagesz=", default_hugepagesz_setup);
4896
4897 static nodemask_t *policy_mbind_nodemask(gfp_t gfp)
4898 {
4899 #ifdef CONFIG_NUMA
4900         struct mempolicy *mpol = get_task_policy(current);
4901
4902         /*
4903          * Only enforce MPOL_BIND policy which overlaps with cpuset policy
4904          * (from policy_nodemask) specifically for hugetlb case
4905          */
4906         if (mpol->mode == MPOL_BIND &&
4907                 (apply_policy_zone(mpol, gfp_zone(gfp)) &&
4908                  cpuset_nodemask_valid_mems_allowed(&mpol->nodes)))
4909                 return &mpol->nodes;
4910 #endif
4911         return NULL;
4912 }
4913
4914 static unsigned int allowed_mems_nr(struct hstate *h)
4915 {
4916         int node;
4917         unsigned int nr = 0;
4918         nodemask_t *mbind_nodemask;
4919         unsigned int *array = h->free_huge_pages_node;
4920         gfp_t gfp_mask = htlb_alloc_mask(h);
4921
4922         mbind_nodemask = policy_mbind_nodemask(gfp_mask);
4923         for_each_node_mask(node, cpuset_current_mems_allowed) {
4924                 if (!mbind_nodemask || node_isset(node, *mbind_nodemask))
4925                         nr += array[node];
4926         }
4927
4928         return nr;
4929 }
4930
4931 #ifdef CONFIG_SYSCTL
4932 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
4933                                           void *buffer, size_t *length,
4934                                           loff_t *ppos, unsigned long *out)
4935 {
4936         struct ctl_table dup_table;
4937
4938         /*
4939          * In order to avoid races with __do_proc_doulongvec_minmax(), we
4940          * can duplicate the @table and alter the duplicate of it.
4941          */
4942         dup_table = *table;
4943         dup_table.data = out;
4944
4945         return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
4946 }
4947
4948 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
4949                          struct ctl_table *table, int write,
4950                          void *buffer, size_t *length, loff_t *ppos)
4951 {
4952         struct hstate *h = &default_hstate;
4953         unsigned long tmp = h->max_huge_pages;
4954         int ret;
4955
4956         if (!hugepages_supported())
4957                 return -EOPNOTSUPP;
4958
4959         ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4960                                              &tmp);
4961         if (ret)
4962                 goto out;
4963
4964         if (write)
4965                 ret = __nr_hugepages_store_common(obey_mempolicy, h,
4966                                                   NUMA_NO_NODE, tmp, *length);
4967 out:
4968         return ret;
4969 }
4970
4971 static int hugetlb_sysctl_handler(struct ctl_table *table, int write,
4972                           void *buffer, size_t *length, loff_t *ppos)
4973 {
4974
4975         return hugetlb_sysctl_handler_common(false, table, write,
4976                                                         buffer, length, ppos);
4977 }
4978
4979 #ifdef CONFIG_NUMA
4980 static int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
4981                           void *buffer, size_t *length, loff_t *ppos)
4982 {
4983         return hugetlb_sysctl_handler_common(true, table, write,
4984                                                         buffer, length, ppos);
4985 }
4986 #endif /* CONFIG_NUMA */
4987
4988 static int hugetlb_overcommit_handler(struct ctl_table *table, int write,
4989                 void *buffer, size_t *length, loff_t *ppos)
4990 {
4991         struct hstate *h = &default_hstate;
4992         unsigned long tmp;
4993         int ret;
4994
4995         if (!hugepages_supported())
4996                 return -EOPNOTSUPP;
4997
4998         tmp = h->nr_overcommit_huge_pages;
4999
5000         if (write && hstate_is_gigantic(h))
5001                 return -EINVAL;
5002
5003         ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
5004                                              &tmp);
5005         if (ret)
5006                 goto out;
5007
5008         if (write) {
5009                 spin_lock_irq(&hugetlb_lock);
5010                 h->nr_overcommit_huge_pages = tmp;
5011                 spin_unlock_irq(&hugetlb_lock);
5012         }
5013 out:
5014         return ret;
5015 }
5016
5017 static struct ctl_table hugetlb_table[] = {
5018         {
5019                 .procname       = "nr_hugepages",
5020                 .data           = NULL,
5021                 .maxlen         = sizeof(unsigned long),
5022                 .mode           = 0644,
5023                 .proc_handler   = hugetlb_sysctl_handler,
5024         },
5025 #ifdef CONFIG_NUMA
5026         {
5027                 .procname       = "nr_hugepages_mempolicy",
5028                 .data           = NULL,
5029                 .maxlen         = sizeof(unsigned long),
5030                 .mode           = 0644,
5031                 .proc_handler   = &hugetlb_mempolicy_sysctl_handler,
5032         },
5033 #endif
5034         {
5035                 .procname       = "hugetlb_shm_group",
5036                 .data           = &sysctl_hugetlb_shm_group,
5037                 .maxlen         = sizeof(gid_t),
5038                 .mode           = 0644,
5039                 .proc_handler   = proc_dointvec,
5040         },
5041         {
5042                 .procname       = "nr_overcommit_hugepages",
5043                 .data           = NULL,
5044                 .maxlen         = sizeof(unsigned long),
5045                 .mode           = 0644,
5046                 .proc_handler   = hugetlb_overcommit_handler,
5047         },
5048         { }
5049 };
5050
5051 static void hugetlb_sysctl_init(void)
5052 {
5053         register_sysctl_init("vm", hugetlb_table);
5054 }
5055 #endif /* CONFIG_SYSCTL */
5056
5057 void hugetlb_report_meminfo(struct seq_file *m)
5058 {
5059         struct hstate *h;
5060         unsigned long total = 0;
5061
5062         if (!hugepages_supported())
5063                 return;
5064
5065         for_each_hstate(h) {
5066                 unsigned long count = h->nr_huge_pages;
5067
5068                 total += huge_page_size(h) * count;
5069
5070                 if (h == &default_hstate)
5071                         seq_printf(m,
5072                                    "HugePages_Total:   %5lu\n"
5073                                    "HugePages_Free:    %5lu\n"
5074                                    "HugePages_Rsvd:    %5lu\n"
5075                                    "HugePages_Surp:    %5lu\n"
5076                                    "Hugepagesize:   %8lu kB\n",
5077                                    count,
5078                                    h->free_huge_pages,
5079                                    h->resv_huge_pages,
5080                                    h->surplus_huge_pages,
5081                                    huge_page_size(h) / SZ_1K);
5082         }
5083
5084         seq_printf(m, "Hugetlb:        %8lu kB\n", total / SZ_1K);
5085 }
5086
5087 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
5088 {
5089         struct hstate *h = &default_hstate;
5090
5091         if (!hugepages_supported())
5092                 return 0;
5093
5094         return sysfs_emit_at(buf, len,
5095                              "Node %d HugePages_Total: %5u\n"
5096                              "Node %d HugePages_Free:  %5u\n"
5097                              "Node %d HugePages_Surp:  %5u\n",
5098                              nid, h->nr_huge_pages_node[nid],
5099                              nid, h->free_huge_pages_node[nid],
5100                              nid, h->surplus_huge_pages_node[nid]);
5101 }
5102
5103 void hugetlb_show_meminfo_node(int nid)
5104 {
5105         struct hstate *h;
5106
5107         if (!hugepages_supported())
5108                 return;
5109
5110         for_each_hstate(h)
5111                 printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
5112                         nid,
5113                         h->nr_huge_pages_node[nid],
5114                         h->free_huge_pages_node[nid],
5115                         h->surplus_huge_pages_node[nid],
5116                         huge_page_size(h) / SZ_1K);
5117 }
5118
5119 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
5120 {
5121         seq_printf(m, "HugetlbPages:\t%8lu kB\n",
5122                    K(atomic_long_read(&mm->hugetlb_usage)));
5123 }
5124
5125 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
5126 unsigned long hugetlb_total_pages(void)
5127 {
5128         struct hstate *h;
5129         unsigned long nr_total_pages = 0;
5130
5131         for_each_hstate(h)
5132                 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
5133         return nr_total_pages;
5134 }
5135
5136 static int hugetlb_acct_memory(struct hstate *h, long delta)
5137 {
5138         int ret = -ENOMEM;
5139
5140         if (!delta)
5141                 return 0;
5142
5143         spin_lock_irq(&hugetlb_lock);
5144         /*
5145          * When cpuset is configured, it breaks the strict hugetlb page
5146          * reservation as the accounting is done on a global variable. Such
5147          * reservation is completely rubbish in the presence of cpuset because
5148          * the reservation is not checked against page availability for the
5149          * current cpuset. Application can still potentially OOM'ed by kernel
5150          * with lack of free htlb page in cpuset that the task is in.
5151          * Attempt to enforce strict accounting with cpuset is almost
5152          * impossible (or too ugly) because cpuset is too fluid that
5153          * task or memory node can be dynamically moved between cpusets.
5154          *
5155          * The change of semantics for shared hugetlb mapping with cpuset is
5156          * undesirable. However, in order to preserve some of the semantics,
5157          * we fall back to check against current free page availability as
5158          * a best attempt and hopefully to minimize the impact of changing
5159          * semantics that cpuset has.
5160          *
5161          * Apart from cpuset, we also have memory policy mechanism that
5162          * also determines from which node the kernel will allocate memory
5163          * in a NUMA system. So similar to cpuset, we also should consider
5164          * the memory policy of the current task. Similar to the description
5165          * above.
5166          */
5167         if (delta > 0) {
5168                 if (gather_surplus_pages(h, delta) < 0)
5169                         goto out;
5170
5171                 if (delta > allowed_mems_nr(h)) {
5172                         return_unused_surplus_pages(h, delta);
5173                         goto out;
5174                 }
5175         }
5176
5177         ret = 0;
5178         if (delta < 0)
5179                 return_unused_surplus_pages(h, (unsigned long) -delta);
5180
5181 out:
5182         spin_unlock_irq(&hugetlb_lock);
5183         return ret;
5184 }
5185
5186 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
5187 {
5188         struct resv_map *resv = vma_resv_map(vma);
5189
5190         /*
5191          * HPAGE_RESV_OWNER indicates a private mapping.
5192          * This new VMA should share its siblings reservation map if present.
5193          * The VMA will only ever have a valid reservation map pointer where
5194          * it is being copied for another still existing VMA.  As that VMA
5195          * has a reference to the reservation map it cannot disappear until
5196          * after this open call completes.  It is therefore safe to take a
5197          * new reference here without additional locking.
5198          */
5199         if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
5200                 resv_map_dup_hugetlb_cgroup_uncharge_info(resv);
5201                 kref_get(&resv->refs);
5202         }
5203
5204         /*
5205          * vma_lock structure for sharable mappings is vma specific.
5206          * Clear old pointer (if copied via vm_area_dup) and allocate
5207          * new structure.  Before clearing, make sure vma_lock is not
5208          * for this vma.
5209          */
5210         if (vma->vm_flags & VM_MAYSHARE) {
5211                 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
5212
5213                 if (vma_lock) {
5214                         if (vma_lock->vma != vma) {
5215                                 vma->vm_private_data = NULL;
5216                                 hugetlb_vma_lock_alloc(vma);
5217                         } else
5218                                 pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__);
5219                 } else
5220                         hugetlb_vma_lock_alloc(vma);
5221         }
5222 }
5223
5224 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
5225 {
5226         struct hstate *h = hstate_vma(vma);
5227         struct resv_map *resv;
5228         struct hugepage_subpool *spool = subpool_vma(vma);
5229         unsigned long reserve, start, end;
5230         long gbl_reserve;
5231
5232         hugetlb_vma_lock_free(vma);
5233
5234         resv = vma_resv_map(vma);
5235         if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5236                 return;
5237
5238         start = vma_hugecache_offset(h, vma, vma->vm_start);
5239         end = vma_hugecache_offset(h, vma, vma->vm_end);
5240
5241         reserve = (end - start) - region_count(resv, start, end);
5242         hugetlb_cgroup_uncharge_counter(resv, start, end);
5243         if (reserve) {
5244                 /*
5245                  * Decrement reserve counts.  The global reserve count may be
5246                  * adjusted if the subpool has a minimum size.
5247                  */
5248                 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
5249                 hugetlb_acct_memory(h, -gbl_reserve);
5250         }
5251
5252         kref_put(&resv->refs, resv_map_release);
5253 }
5254
5255 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
5256 {
5257         if (addr & ~(huge_page_mask(hstate_vma(vma))))
5258                 return -EINVAL;
5259
5260         /*
5261          * PMD sharing is only possible for PUD_SIZE-aligned address ranges
5262          * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this
5263          * split, unshare PMDs in the PUD_SIZE interval surrounding addr now.
5264          */
5265         if (addr & ~PUD_MASK) {
5266                 /*
5267                  * hugetlb_vm_op_split is called right before we attempt to
5268                  * split the VMA. We will need to unshare PMDs in the old and
5269                  * new VMAs, so let's unshare before we split.
5270                  */
5271                 unsigned long floor = addr & PUD_MASK;
5272                 unsigned long ceil = floor + PUD_SIZE;
5273
5274                 if (floor >= vma->vm_start && ceil <= vma->vm_end)
5275                         hugetlb_unshare_pmds(vma, floor, ceil);
5276         }
5277
5278         return 0;
5279 }
5280
5281 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
5282 {
5283         return huge_page_size(hstate_vma(vma));
5284 }
5285
5286 /*
5287  * We cannot handle pagefaults against hugetlb pages at all.  They cause
5288  * handle_mm_fault() to try to instantiate regular-sized pages in the
5289  * hugepage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
5290  * this far.
5291  */
5292 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
5293 {
5294         BUG();
5295         return 0;
5296 }
5297
5298 /*
5299  * When a new function is introduced to vm_operations_struct and added
5300  * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
5301  * This is because under System V memory model, mappings created via
5302  * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
5303  * their original vm_ops are overwritten with shm_vm_ops.
5304  */
5305 const struct vm_operations_struct hugetlb_vm_ops = {
5306         .fault = hugetlb_vm_op_fault,
5307         .open = hugetlb_vm_op_open,
5308         .close = hugetlb_vm_op_close,
5309         .may_split = hugetlb_vm_op_split,
5310         .pagesize = hugetlb_vm_op_pagesize,
5311 };
5312
5313 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
5314                                 int writable)
5315 {
5316         pte_t entry;
5317         unsigned int shift = huge_page_shift(hstate_vma(vma));
5318
5319         if (writable) {
5320                 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
5321                                          vma->vm_page_prot)));
5322         } else {
5323                 entry = huge_pte_wrprotect(mk_huge_pte(page,
5324                                            vma->vm_page_prot));
5325         }
5326         entry = pte_mkyoung(entry);
5327         entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
5328
5329         return entry;
5330 }
5331
5332 static void set_huge_ptep_writable(struct vm_area_struct *vma,
5333                                    unsigned long address, pte_t *ptep)
5334 {
5335         pte_t entry;
5336
5337         entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
5338         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
5339                 update_mmu_cache(vma, address, ptep);
5340 }
5341
5342 bool is_hugetlb_entry_migration(pte_t pte)
5343 {
5344         swp_entry_t swp;
5345
5346         if (huge_pte_none(pte) || pte_present(pte))
5347                 return false;
5348         swp = pte_to_swp_entry(pte);
5349         if (is_migration_entry(swp))
5350                 return true;
5351         else
5352                 return false;
5353 }
5354
5355 bool is_hugetlb_entry_hwpoisoned(pte_t pte)
5356 {
5357         swp_entry_t swp;
5358
5359         if (huge_pte_none(pte) || pte_present(pte))
5360                 return false;
5361         swp = pte_to_swp_entry(pte);
5362         if (is_hwpoison_entry(swp))
5363                 return true;
5364         else
5365                 return false;
5366 }
5367
5368 static void
5369 hugetlb_install_folio(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
5370                       struct folio *new_folio, pte_t old, unsigned long sz)
5371 {
5372         pte_t newpte = make_huge_pte(vma, &new_folio->page, 1);
5373
5374         __folio_mark_uptodate(new_folio);
5375         hugetlb_add_new_anon_rmap(new_folio, vma, addr);
5376         if (userfaultfd_wp(vma) && huge_pte_uffd_wp(old))
5377                 newpte = huge_pte_mkuffd_wp(newpte);
5378         set_huge_pte_at(vma->vm_mm, addr, ptep, newpte, sz);
5379         hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
5380         folio_set_hugetlb_migratable(new_folio);
5381 }
5382
5383 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
5384                             struct vm_area_struct *dst_vma,
5385                             struct vm_area_struct *src_vma)
5386 {
5387         pte_t *src_pte, *dst_pte, entry;
5388         struct folio *pte_folio;
5389         unsigned long addr;
5390         bool cow = is_cow_mapping(src_vma->vm_flags);
5391         struct hstate *h = hstate_vma(src_vma);
5392         unsigned long sz = huge_page_size(h);
5393         unsigned long npages = pages_per_huge_page(h);
5394         struct mmu_notifier_range range;
5395         unsigned long last_addr_mask;
5396         int ret = 0;
5397
5398         if (cow) {
5399                 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src,
5400                                         src_vma->vm_start,
5401                                         src_vma->vm_end);
5402                 mmu_notifier_invalidate_range_start(&range);
5403                 vma_assert_write_locked(src_vma);
5404                 raw_write_seqcount_begin(&src->write_protect_seq);
5405         } else {
5406                 /*
5407                  * For shared mappings the vma lock must be held before
5408                  * calling hugetlb_walk() in the src vma. Otherwise, the
5409                  * returned ptep could go away if part of a shared pmd and
5410                  * another thread calls huge_pmd_unshare.
5411                  */
5412                 hugetlb_vma_lock_read(src_vma);
5413         }
5414
5415         last_addr_mask = hugetlb_mask_last_page(h);
5416         for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) {
5417                 spinlock_t *src_ptl, *dst_ptl;
5418                 src_pte = hugetlb_walk(src_vma, addr, sz);
5419                 if (!src_pte) {
5420                         addr |= last_addr_mask;
5421                         continue;
5422                 }
5423                 dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz);
5424                 if (!dst_pte) {
5425                         ret = -ENOMEM;
5426                         break;
5427                 }
5428
5429                 /*
5430                  * If the pagetables are shared don't copy or take references.
5431                  *
5432                  * dst_pte == src_pte is the common case of src/dest sharing.
5433                  * However, src could have 'unshared' and dst shares with
5434                  * another vma. So page_count of ptep page is checked instead
5435                  * to reliably determine whether pte is shared.
5436                  */
5437                 if (page_count(virt_to_page(dst_pte)) > 1) {
5438                         addr |= last_addr_mask;
5439                         continue;
5440                 }
5441
5442                 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5443                 src_ptl = huge_pte_lockptr(h, src, src_pte);
5444                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5445                 entry = huge_ptep_get(src_pte);
5446 again:
5447                 if (huge_pte_none(entry)) {
5448                         /*
5449                          * Skip if src entry none.
5450                          */
5451                         ;
5452                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) {
5453                         if (!userfaultfd_wp(dst_vma))
5454                                 entry = huge_pte_clear_uffd_wp(entry);
5455                         set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5456                 } else if (unlikely(is_hugetlb_entry_migration(entry))) {
5457                         swp_entry_t swp_entry = pte_to_swp_entry(entry);
5458                         bool uffd_wp = pte_swp_uffd_wp(entry);
5459
5460                         if (!is_readable_migration_entry(swp_entry) && cow) {
5461                                 /*
5462                                  * COW mappings require pages in both
5463                                  * parent and child to be set to read.
5464                                  */
5465                                 swp_entry = make_readable_migration_entry(
5466                                                         swp_offset(swp_entry));
5467                                 entry = swp_entry_to_pte(swp_entry);
5468                                 if (userfaultfd_wp(src_vma) && uffd_wp)
5469                                         entry = pte_swp_mkuffd_wp(entry);
5470                                 set_huge_pte_at(src, addr, src_pte, entry, sz);
5471                         }
5472                         if (!userfaultfd_wp(dst_vma))
5473                                 entry = huge_pte_clear_uffd_wp(entry);
5474                         set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5475                 } else if (unlikely(is_pte_marker(entry))) {
5476                         pte_marker marker = copy_pte_marker(
5477                                 pte_to_swp_entry(entry), dst_vma);
5478
5479                         if (marker)
5480                                 set_huge_pte_at(dst, addr, dst_pte,
5481                                                 make_pte_marker(marker), sz);
5482                 } else {
5483                         entry = huge_ptep_get(src_pte);
5484                         pte_folio = page_folio(pte_page(entry));
5485                         folio_get(pte_folio);
5486
5487                         /*
5488                          * Failing to duplicate the anon rmap is a rare case
5489                          * where we see pinned hugetlb pages while they're
5490                          * prone to COW. We need to do the COW earlier during
5491                          * fork.
5492                          *
5493                          * When pre-allocating the page or copying data, we
5494                          * need to be without the pgtable locks since we could
5495                          * sleep during the process.
5496                          */
5497                         if (!folio_test_anon(pte_folio)) {
5498                                 hugetlb_add_file_rmap(pte_folio);
5499                         } else if (hugetlb_try_dup_anon_rmap(pte_folio, src_vma)) {
5500                                 pte_t src_pte_old = entry;
5501                                 struct folio *new_folio;
5502
5503                                 spin_unlock(src_ptl);
5504                                 spin_unlock(dst_ptl);
5505                                 /* Do not use reserve as it's private owned */
5506                                 new_folio = alloc_hugetlb_folio(dst_vma, addr, 1);
5507                                 if (IS_ERR(new_folio)) {
5508                                         folio_put(pte_folio);
5509                                         ret = PTR_ERR(new_folio);
5510                                         break;
5511                                 }
5512                                 ret = copy_user_large_folio(new_folio,
5513                                                             pte_folio,
5514                                                             addr, dst_vma);
5515                                 folio_put(pte_folio);
5516                                 if (ret) {
5517                                         folio_put(new_folio);
5518                                         break;
5519                                 }
5520
5521                                 /* Install the new hugetlb folio if src pte stable */
5522                                 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5523                                 src_ptl = huge_pte_lockptr(h, src, src_pte);
5524                                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5525                                 entry = huge_ptep_get(src_pte);
5526                                 if (!pte_same(src_pte_old, entry)) {
5527                                         restore_reserve_on_error(h, dst_vma, addr,
5528                                                                 new_folio);
5529                                         folio_put(new_folio);
5530                                         /* huge_ptep of dst_pte won't change as in child */
5531                                         goto again;
5532                                 }
5533                                 hugetlb_install_folio(dst_vma, dst_pte, addr,
5534                                                       new_folio, src_pte_old, sz);
5535                                 spin_unlock(src_ptl);
5536                                 spin_unlock(dst_ptl);
5537                                 continue;
5538                         }
5539
5540                         if (cow) {
5541                                 /*
5542                                  * No need to notify as we are downgrading page
5543                                  * table protection not changing it to point
5544                                  * to a new page.
5545                                  *
5546                                  * See Documentation/mm/mmu_notifier.rst
5547                                  */
5548                                 huge_ptep_set_wrprotect(src, addr, src_pte);
5549                                 entry = huge_pte_wrprotect(entry);
5550                         }
5551
5552                         if (!userfaultfd_wp(dst_vma))
5553                                 entry = huge_pte_clear_uffd_wp(entry);
5554
5555                         set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5556                         hugetlb_count_add(npages, dst);
5557                 }
5558                 spin_unlock(src_ptl);
5559                 spin_unlock(dst_ptl);
5560         }
5561
5562         if (cow) {
5563                 raw_write_seqcount_end(&src->write_protect_seq);
5564                 mmu_notifier_invalidate_range_end(&range);
5565         } else {
5566                 hugetlb_vma_unlock_read(src_vma);
5567         }
5568
5569         return ret;
5570 }
5571
5572 static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr,
5573                           unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte,
5574                           unsigned long sz)
5575 {
5576         struct hstate *h = hstate_vma(vma);
5577         struct mm_struct *mm = vma->vm_mm;
5578         spinlock_t *src_ptl, *dst_ptl;
5579         pte_t pte;
5580
5581         dst_ptl = huge_pte_lock(h, mm, dst_pte);
5582         src_ptl = huge_pte_lockptr(h, mm, src_pte);
5583
5584         /*
5585          * We don't have to worry about the ordering of src and dst ptlocks
5586          * because exclusive mmap_lock (or the i_mmap_lock) prevents deadlock.
5587          */
5588         if (src_ptl != dst_ptl)
5589                 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5590
5591         pte = huge_ptep_get_and_clear(mm, old_addr, src_pte);
5592         set_huge_pte_at(mm, new_addr, dst_pte, pte, sz);
5593
5594         if (src_ptl != dst_ptl)
5595                 spin_unlock(src_ptl);
5596         spin_unlock(dst_ptl);
5597 }
5598
5599 int move_hugetlb_page_tables(struct vm_area_struct *vma,
5600                              struct vm_area_struct *new_vma,
5601                              unsigned long old_addr, unsigned long new_addr,
5602                              unsigned long len)
5603 {
5604         struct hstate *h = hstate_vma(vma);
5605         struct address_space *mapping = vma->vm_file->f_mapping;
5606         unsigned long sz = huge_page_size(h);
5607         struct mm_struct *mm = vma->vm_mm;
5608         unsigned long old_end = old_addr + len;
5609         unsigned long last_addr_mask;
5610         pte_t *src_pte, *dst_pte;
5611         struct mmu_notifier_range range;
5612         bool shared_pmd = false;
5613
5614         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, old_addr,
5615                                 old_end);
5616         adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5617         /*
5618          * In case of shared PMDs, we should cover the maximum possible
5619          * range.
5620          */
5621         flush_cache_range(vma, range.start, range.end);
5622
5623         mmu_notifier_invalidate_range_start(&range);
5624         last_addr_mask = hugetlb_mask_last_page(h);
5625         /* Prevent race with file truncation */
5626         hugetlb_vma_lock_write(vma);
5627         i_mmap_lock_write(mapping);
5628         for (; old_addr < old_end; old_addr += sz, new_addr += sz) {
5629                 src_pte = hugetlb_walk(vma, old_addr, sz);
5630                 if (!src_pte) {
5631                         old_addr |= last_addr_mask;
5632                         new_addr |= last_addr_mask;
5633                         continue;
5634                 }
5635                 if (huge_pte_none(huge_ptep_get(src_pte)))
5636                         continue;
5637
5638                 if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) {
5639                         shared_pmd = true;
5640                         old_addr |= last_addr_mask;
5641                         new_addr |= last_addr_mask;
5642                         continue;
5643                 }
5644
5645                 dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz);
5646                 if (!dst_pte)
5647                         break;
5648
5649                 move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte, sz);
5650         }
5651
5652         if (shared_pmd)
5653                 flush_hugetlb_tlb_range(vma, range.start, range.end);
5654         else
5655                 flush_hugetlb_tlb_range(vma, old_end - len, old_end);
5656         mmu_notifier_invalidate_range_end(&range);
5657         i_mmap_unlock_write(mapping);
5658         hugetlb_vma_unlock_write(vma);
5659
5660         return len + old_addr - old_end;
5661 }
5662
5663 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
5664                             unsigned long start, unsigned long end,
5665                             struct page *ref_page, zap_flags_t zap_flags)
5666 {
5667         struct mm_struct *mm = vma->vm_mm;
5668         unsigned long address;
5669         pte_t *ptep;
5670         pte_t pte;
5671         spinlock_t *ptl;
5672         struct page *page;
5673         struct hstate *h = hstate_vma(vma);
5674         unsigned long sz = huge_page_size(h);
5675         bool adjust_reservation = false;
5676         unsigned long last_addr_mask;
5677         bool force_flush = false;
5678
5679         WARN_ON(!is_vm_hugetlb_page(vma));
5680         BUG_ON(start & ~huge_page_mask(h));
5681         BUG_ON(end & ~huge_page_mask(h));
5682
5683         /*
5684          * This is a hugetlb vma, all the pte entries should point
5685          * to huge page.
5686          */
5687         tlb_change_page_size(tlb, sz);
5688         tlb_start_vma(tlb, vma);
5689
5690         last_addr_mask = hugetlb_mask_last_page(h);
5691         address = start;
5692         for (; address < end; address += sz) {
5693                 ptep = hugetlb_walk(vma, address, sz);
5694                 if (!ptep) {
5695                         address |= last_addr_mask;
5696                         continue;
5697                 }
5698
5699                 ptl = huge_pte_lock(h, mm, ptep);
5700                 if (huge_pmd_unshare(mm, vma, address, ptep)) {
5701                         spin_unlock(ptl);
5702                         tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE);
5703                         force_flush = true;
5704                         address |= last_addr_mask;
5705                         continue;
5706                 }
5707
5708                 pte = huge_ptep_get(ptep);
5709                 if (huge_pte_none(pte)) {
5710                         spin_unlock(ptl);
5711                         continue;
5712                 }
5713
5714                 /*
5715                  * Migrating hugepage or HWPoisoned hugepage is already
5716                  * unmapped and its refcount is dropped, so just clear pte here.
5717                  */
5718                 if (unlikely(!pte_present(pte))) {
5719                         /*
5720                          * If the pte was wr-protected by uffd-wp in any of the
5721                          * swap forms, meanwhile the caller does not want to
5722                          * drop the uffd-wp bit in this zap, then replace the
5723                          * pte with a marker.
5724                          */
5725                         if (pte_swp_uffd_wp_any(pte) &&
5726                             !(zap_flags & ZAP_FLAG_DROP_MARKER))
5727                                 set_huge_pte_at(mm, address, ptep,
5728                                                 make_pte_marker(PTE_MARKER_UFFD_WP),
5729                                                 sz);
5730                         else
5731                                 huge_pte_clear(mm, address, ptep, sz);
5732                         spin_unlock(ptl);
5733                         continue;
5734                 }
5735
5736                 page = pte_page(pte);
5737                 /*
5738                  * If a reference page is supplied, it is because a specific
5739                  * page is being unmapped, not a range. Ensure the page we
5740                  * are about to unmap is the actual page of interest.
5741                  */
5742                 if (ref_page) {
5743                         if (page != ref_page) {
5744                                 spin_unlock(ptl);
5745                                 continue;
5746                         }
5747                         /*
5748                          * Mark the VMA as having unmapped its page so that
5749                          * future faults in this VMA will fail rather than
5750                          * looking like data was lost
5751                          */
5752                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
5753                 }
5754
5755                 pte = huge_ptep_get_and_clear(mm, address, ptep);
5756                 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
5757                 if (huge_pte_dirty(pte))
5758                         set_page_dirty(page);
5759                 /* Leave a uffd-wp pte marker if needed */
5760                 if (huge_pte_uffd_wp(pte) &&
5761                     !(zap_flags & ZAP_FLAG_DROP_MARKER))
5762                         set_huge_pte_at(mm, address, ptep,
5763                                         make_pte_marker(PTE_MARKER_UFFD_WP),
5764                                         sz);
5765                 hugetlb_count_sub(pages_per_huge_page(h), mm);
5766                 hugetlb_remove_rmap(page_folio(page));
5767
5768                 /*
5769                  * Restore the reservation for anonymous page, otherwise the
5770                  * backing page could be stolen by someone.
5771                  * If there we are freeing a surplus, do not set the restore
5772                  * reservation bit.
5773                  */
5774                 if (!h->surplus_huge_pages && __vma_private_lock(vma) &&
5775                     folio_test_anon(page_folio(page))) {
5776                         folio_set_hugetlb_restore_reserve(page_folio(page));
5777                         /* Reservation to be adjusted after the spin lock */
5778                         adjust_reservation = true;
5779                 }
5780
5781                 spin_unlock(ptl);
5782
5783                 /*
5784                  * Adjust the reservation for the region that will have the
5785                  * reserve restored. Keep in mind that vma_needs_reservation() changes
5786                  * resv->adds_in_progress if it succeeds. If this is not done,
5787                  * do_exit() will not see it, and will keep the reservation
5788                  * forever.
5789                  */
5790                 if (adjust_reservation && vma_needs_reservation(h, vma, address))
5791                         vma_add_reservation(h, vma, address);
5792
5793                 tlb_remove_page_size(tlb, page, huge_page_size(h));
5794                 /*
5795                  * Bail out after unmapping reference page if supplied
5796                  */
5797                 if (ref_page)
5798                         break;
5799         }
5800         tlb_end_vma(tlb, vma);
5801
5802         /*
5803          * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We
5804          * could defer the flush until now, since by holding i_mmap_rwsem we
5805          * guaranteed that the last refernece would not be dropped. But we must
5806          * do the flushing before we return, as otherwise i_mmap_rwsem will be
5807          * dropped and the last reference to the shared PMDs page might be
5808          * dropped as well.
5809          *
5810          * In theory we could defer the freeing of the PMD pages as well, but
5811          * huge_pmd_unshare() relies on the exact page_count for the PMD page to
5812          * detect sharing, so we cannot defer the release of the page either.
5813          * Instead, do flush now.
5814          */
5815         if (force_flush)
5816                 tlb_flush_mmu_tlbonly(tlb);
5817 }
5818
5819 void __hugetlb_zap_begin(struct vm_area_struct *vma,
5820                          unsigned long *start, unsigned long *end)
5821 {
5822         if (!vma->vm_file)      /* hugetlbfs_file_mmap error */
5823                 return;
5824
5825         adjust_range_if_pmd_sharing_possible(vma, start, end);
5826         hugetlb_vma_lock_write(vma);
5827         if (vma->vm_file)
5828                 i_mmap_lock_write(vma->vm_file->f_mapping);
5829 }
5830
5831 void __hugetlb_zap_end(struct vm_area_struct *vma,
5832                        struct zap_details *details)
5833 {
5834         zap_flags_t zap_flags = details ? details->zap_flags : 0;
5835
5836         if (!vma->vm_file)      /* hugetlbfs_file_mmap error */
5837                 return;
5838
5839         if (zap_flags & ZAP_FLAG_UNMAP) {       /* final unmap */
5840                 /*
5841                  * Unlock and free the vma lock before releasing i_mmap_rwsem.
5842                  * When the vma_lock is freed, this makes the vma ineligible
5843                  * for pmd sharing.  And, i_mmap_rwsem is required to set up
5844                  * pmd sharing.  This is important as page tables for this
5845                  * unmapped range will be asynchrously deleted.  If the page
5846                  * tables are shared, there will be issues when accessed by
5847                  * someone else.
5848                  */
5849                 __hugetlb_vma_unlock_write_free(vma);
5850         } else {
5851                 hugetlb_vma_unlock_write(vma);
5852         }
5853
5854         if (vma->vm_file)
5855                 i_mmap_unlock_write(vma->vm_file->f_mapping);
5856 }
5857
5858 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
5859                           unsigned long end, struct page *ref_page,
5860                           zap_flags_t zap_flags)
5861 {
5862         struct mmu_notifier_range range;
5863         struct mmu_gather tlb;
5864
5865         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
5866                                 start, end);
5867         adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5868         mmu_notifier_invalidate_range_start(&range);
5869         tlb_gather_mmu(&tlb, vma->vm_mm);
5870
5871         __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags);
5872
5873         mmu_notifier_invalidate_range_end(&range);
5874         tlb_finish_mmu(&tlb);
5875 }
5876
5877 /*
5878  * This is called when the original mapper is failing to COW a MAP_PRIVATE
5879  * mapping it owns the reserve page for. The intention is to unmap the page
5880  * from other VMAs and let the children be SIGKILLed if they are faulting the
5881  * same region.
5882  */
5883 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
5884                               struct page *page, unsigned long address)
5885 {
5886         struct hstate *h = hstate_vma(vma);
5887         struct vm_area_struct *iter_vma;
5888         struct address_space *mapping;
5889         pgoff_t pgoff;
5890
5891         /*
5892          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
5893          * from page cache lookup which is in HPAGE_SIZE units.
5894          */
5895         address = address & huge_page_mask(h);
5896         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
5897                         vma->vm_pgoff;
5898         mapping = vma->vm_file->f_mapping;
5899
5900         /*
5901          * Take the mapping lock for the duration of the table walk. As
5902          * this mapping should be shared between all the VMAs,
5903          * __unmap_hugepage_range() is called as the lock is already held
5904          */
5905         i_mmap_lock_write(mapping);
5906         vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
5907                 /* Do not unmap the current VMA */
5908                 if (iter_vma == vma)
5909                         continue;
5910
5911                 /*
5912                  * Shared VMAs have their own reserves and do not affect
5913                  * MAP_PRIVATE accounting but it is possible that a shared
5914                  * VMA is using the same page so check and skip such VMAs.
5915                  */
5916                 if (iter_vma->vm_flags & VM_MAYSHARE)
5917                         continue;
5918
5919                 /*
5920                  * Unmap the page from other VMAs without their own reserves.
5921                  * They get marked to be SIGKILLed if they fault in these
5922                  * areas. This is because a future no-page fault on this VMA
5923                  * could insert a zeroed page instead of the data existing
5924                  * from the time of fork. This would look like data corruption
5925                  */
5926                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
5927                         unmap_hugepage_range(iter_vma, address,
5928                                              address + huge_page_size(h), page, 0);
5929         }
5930         i_mmap_unlock_write(mapping);
5931 }
5932
5933 /*
5934  * hugetlb_wp() should be called with page lock of the original hugepage held.
5935  * Called with hugetlb_fault_mutex_table held and pte_page locked so we
5936  * cannot race with other handlers or page migration.
5937  * Keep the pte_same checks anyway to make transition from the mutex easier.
5938  */
5939 static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma,
5940                        unsigned long address, pte_t *ptep, unsigned int flags,
5941                        struct folio *pagecache_folio, spinlock_t *ptl,
5942                        struct vm_fault *vmf)
5943 {
5944         const bool unshare = flags & FAULT_FLAG_UNSHARE;
5945         pte_t pte = huge_ptep_get(ptep);
5946         struct hstate *h = hstate_vma(vma);
5947         struct folio *old_folio;
5948         struct folio *new_folio;
5949         int outside_reserve = 0;
5950         vm_fault_t ret = 0;
5951         unsigned long haddr = address & huge_page_mask(h);
5952         struct mmu_notifier_range range;
5953
5954         /*
5955          * Never handle CoW for uffd-wp protected pages.  It should be only
5956          * handled when the uffd-wp protection is removed.
5957          *
5958          * Note that only the CoW optimization path (in hugetlb_no_page())
5959          * can trigger this, because hugetlb_fault() will always resolve
5960          * uffd-wp bit first.
5961          */
5962         if (!unshare && huge_pte_uffd_wp(pte))
5963                 return 0;
5964
5965         /*
5966          * hugetlb does not support FOLL_FORCE-style write faults that keep the
5967          * PTE mapped R/O such as maybe_mkwrite() would do.
5968          */
5969         if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE)))
5970                 return VM_FAULT_SIGSEGV;
5971
5972         /* Let's take out MAP_SHARED mappings first. */
5973         if (vma->vm_flags & VM_MAYSHARE) {
5974                 set_huge_ptep_writable(vma, haddr, ptep);
5975                 return 0;
5976         }
5977
5978         old_folio = page_folio(pte_page(pte));
5979
5980         delayacct_wpcopy_start();
5981
5982 retry_avoidcopy:
5983         /*
5984          * If no-one else is actually using this page, we're the exclusive
5985          * owner and can reuse this page.
5986          */
5987         if (folio_mapcount(old_folio) == 1 && folio_test_anon(old_folio)) {
5988                 if (!PageAnonExclusive(&old_folio->page)) {
5989                         folio_move_anon_rmap(old_folio, vma);
5990                         SetPageAnonExclusive(&old_folio->page);
5991                 }
5992                 if (likely(!unshare))
5993                         set_huge_ptep_writable(vma, haddr, ptep);
5994
5995                 delayacct_wpcopy_end();
5996                 return 0;
5997         }
5998         VM_BUG_ON_PAGE(folio_test_anon(old_folio) &&
5999                        PageAnonExclusive(&old_folio->page), &old_folio->page);
6000
6001         /*
6002          * If the process that created a MAP_PRIVATE mapping is about to
6003          * perform a COW due to a shared page count, attempt to satisfy
6004          * the allocation without using the existing reserves. The pagecache
6005          * page is used to determine if the reserve at this address was
6006          * consumed or not. If reserves were used, a partial faulted mapping
6007          * at the time of fork() could consume its reserves on COW instead
6008          * of the full address range.
6009          */
6010         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
6011                         old_folio != pagecache_folio)
6012                 outside_reserve = 1;
6013
6014         folio_get(old_folio);
6015
6016         /*
6017          * Drop page table lock as buddy allocator may be called. It will
6018          * be acquired again before returning to the caller, as expected.
6019          */
6020         spin_unlock(ptl);
6021         new_folio = alloc_hugetlb_folio(vma, haddr, outside_reserve);
6022
6023         if (IS_ERR(new_folio)) {
6024                 /*
6025                  * If a process owning a MAP_PRIVATE mapping fails to COW,
6026                  * it is due to references held by a child and an insufficient
6027                  * huge page pool. To guarantee the original mappers
6028                  * reliability, unmap the page from child processes. The child
6029                  * may get SIGKILLed if it later faults.
6030                  */
6031                 if (outside_reserve) {
6032                         struct address_space *mapping = vma->vm_file->f_mapping;
6033                         pgoff_t idx;
6034                         u32 hash;
6035
6036                         folio_put(old_folio);
6037                         /*
6038                          * Drop hugetlb_fault_mutex and vma_lock before
6039                          * unmapping.  unmapping needs to hold vma_lock
6040                          * in write mode.  Dropping vma_lock in read mode
6041                          * here is OK as COW mappings do not interact with
6042                          * PMD sharing.
6043                          *
6044                          * Reacquire both after unmap operation.
6045                          */
6046                         idx = vma_hugecache_offset(h, vma, haddr);
6047                         hash = hugetlb_fault_mutex_hash(mapping, idx);
6048                         hugetlb_vma_unlock_read(vma);
6049                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6050
6051                         unmap_ref_private(mm, vma, &old_folio->page, haddr);
6052
6053                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
6054                         hugetlb_vma_lock_read(vma);
6055                         spin_lock(ptl);
6056                         ptep = hugetlb_walk(vma, haddr, huge_page_size(h));
6057                         if (likely(ptep &&
6058                                    pte_same(huge_ptep_get(ptep), pte)))
6059                                 goto retry_avoidcopy;
6060                         /*
6061                          * race occurs while re-acquiring page table
6062                          * lock, and our job is done.
6063                          */
6064                         delayacct_wpcopy_end();
6065                         return 0;
6066                 }
6067
6068                 ret = vmf_error(PTR_ERR(new_folio));
6069                 goto out_release_old;
6070         }
6071
6072         /*
6073          * When the original hugepage is shared one, it does not have
6074          * anon_vma prepared.
6075          */
6076         ret = vmf_anon_prepare(vmf);
6077         if (unlikely(ret))
6078                 goto out_release_all;
6079
6080         if (copy_user_large_folio(new_folio, old_folio, address, vma)) {
6081                 ret = VM_FAULT_HWPOISON_LARGE;
6082                 goto out_release_all;
6083         }
6084         __folio_mark_uptodate(new_folio);
6085
6086         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, haddr,
6087                                 haddr + huge_page_size(h));
6088         mmu_notifier_invalidate_range_start(&range);
6089
6090         /*
6091          * Retake the page table lock to check for racing updates
6092          * before the page tables are altered
6093          */
6094         spin_lock(ptl);
6095         ptep = hugetlb_walk(vma, haddr, huge_page_size(h));
6096         if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
6097                 pte_t newpte = make_huge_pte(vma, &new_folio->page, !unshare);
6098
6099                 /* Break COW or unshare */
6100                 huge_ptep_clear_flush(vma, haddr, ptep);
6101                 hugetlb_remove_rmap(old_folio);
6102                 hugetlb_add_new_anon_rmap(new_folio, vma, haddr);
6103                 if (huge_pte_uffd_wp(pte))
6104                         newpte = huge_pte_mkuffd_wp(newpte);
6105                 set_huge_pte_at(mm, haddr, ptep, newpte, huge_page_size(h));
6106                 folio_set_hugetlb_migratable(new_folio);
6107                 /* Make the old page be freed below */
6108                 new_folio = old_folio;
6109         }
6110         spin_unlock(ptl);
6111         mmu_notifier_invalidate_range_end(&range);
6112 out_release_all:
6113         /*
6114          * No restore in case of successful pagetable update (Break COW or
6115          * unshare)
6116          */
6117         if (new_folio != old_folio)
6118                 restore_reserve_on_error(h, vma, haddr, new_folio);
6119         folio_put(new_folio);
6120 out_release_old:
6121         folio_put(old_folio);
6122
6123         spin_lock(ptl); /* Caller expects lock to be held */
6124
6125         delayacct_wpcopy_end();
6126         return ret;
6127 }
6128
6129 /*
6130  * Return whether there is a pagecache page to back given address within VMA.
6131  */
6132 static bool hugetlbfs_pagecache_present(struct hstate *h,
6133                         struct vm_area_struct *vma, unsigned long address)
6134 {
6135         struct address_space *mapping = vma->vm_file->f_mapping;
6136         pgoff_t idx = linear_page_index(vma, address);
6137         struct folio *folio;
6138
6139         folio = filemap_get_folio(mapping, idx);
6140         if (IS_ERR(folio))
6141                 return false;
6142         folio_put(folio);
6143         return true;
6144 }
6145
6146 int hugetlb_add_to_page_cache(struct folio *folio, struct address_space *mapping,
6147                            pgoff_t idx)
6148 {
6149         struct inode *inode = mapping->host;
6150         struct hstate *h = hstate_inode(inode);
6151         int err;
6152
6153         idx <<= huge_page_order(h);
6154         __folio_set_locked(folio);
6155         err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL);
6156
6157         if (unlikely(err)) {
6158                 __folio_clear_locked(folio);
6159                 return err;
6160         }
6161         folio_clear_hugetlb_restore_reserve(folio);
6162
6163         /*
6164          * mark folio dirty so that it will not be removed from cache/file
6165          * by non-hugetlbfs specific code paths.
6166          */
6167         folio_mark_dirty(folio);
6168
6169         spin_lock(&inode->i_lock);
6170         inode->i_blocks += blocks_per_huge_page(h);
6171         spin_unlock(&inode->i_lock);
6172         return 0;
6173 }
6174
6175 static inline vm_fault_t hugetlb_handle_userfault(struct vm_fault *vmf,
6176                                                   struct address_space *mapping,
6177                                                   unsigned long reason)
6178 {
6179         u32 hash;
6180
6181         /*
6182          * vma_lock and hugetlb_fault_mutex must be dropped before handling
6183          * userfault. Also mmap_lock could be dropped due to handling
6184          * userfault, any vma operation should be careful from here.
6185          */
6186         hugetlb_vma_unlock_read(vmf->vma);
6187         hash = hugetlb_fault_mutex_hash(mapping, vmf->pgoff);
6188         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6189         return handle_userfault(vmf, reason);
6190 }
6191
6192 /*
6193  * Recheck pte with pgtable lock.  Returns true if pte didn't change, or
6194  * false if pte changed or is changing.
6195  */
6196 static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm,
6197                                pte_t *ptep, pte_t old_pte)
6198 {
6199         spinlock_t *ptl;
6200         bool same;
6201
6202         ptl = huge_pte_lock(h, mm, ptep);
6203         same = pte_same(huge_ptep_get(ptep), old_pte);
6204         spin_unlock(ptl);
6205
6206         return same;
6207 }
6208
6209 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
6210                         struct vm_area_struct *vma,
6211                         struct address_space *mapping, pgoff_t idx,
6212                         unsigned long address, pte_t *ptep,
6213                         pte_t old_pte, unsigned int flags,
6214                         struct vm_fault *vmf)
6215 {
6216         struct hstate *h = hstate_vma(vma);
6217         vm_fault_t ret = VM_FAULT_SIGBUS;
6218         int anon_rmap = 0;
6219         unsigned long size;
6220         struct folio *folio;
6221         pte_t new_pte;
6222         spinlock_t *ptl;
6223         unsigned long haddr = address & huge_page_mask(h);
6224         bool new_folio, new_pagecache_folio = false;
6225         u32 hash = hugetlb_fault_mutex_hash(mapping, idx);
6226
6227         /*
6228          * Currently, we are forced to kill the process in the event the
6229          * original mapper has unmapped pages from the child due to a failed
6230          * COW/unsharing. Warn that such a situation has occurred as it may not
6231          * be obvious.
6232          */
6233         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
6234                 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
6235                            current->pid);
6236                 goto out;
6237         }
6238
6239         /*
6240          * Use page lock to guard against racing truncation
6241          * before we get page_table_lock.
6242          */
6243         new_folio = false;
6244         folio = filemap_lock_hugetlb_folio(h, mapping, idx);
6245         if (IS_ERR(folio)) {
6246                 size = i_size_read(mapping->host) >> huge_page_shift(h);
6247                 if (idx >= size)
6248                         goto out;
6249                 /* Check for page in userfault range */
6250                 if (userfaultfd_missing(vma)) {
6251                         /*
6252                          * Since hugetlb_no_page() was examining pte
6253                          * without pgtable lock, we need to re-test under
6254                          * lock because the pte may not be stable and could
6255                          * have changed from under us.  Try to detect
6256                          * either changed or during-changing ptes and retry
6257                          * properly when needed.
6258                          *
6259                          * Note that userfaultfd is actually fine with
6260                          * false positives (e.g. caused by pte changed),
6261                          * but not wrong logical events (e.g. caused by
6262                          * reading a pte during changing).  The latter can
6263                          * confuse the userspace, so the strictness is very
6264                          * much preferred.  E.g., MISSING event should
6265                          * never happen on the page after UFFDIO_COPY has
6266                          * correctly installed the page and returned.
6267                          */
6268                         if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
6269                                 ret = 0;
6270                                 goto out;
6271                         }
6272
6273                         return hugetlb_handle_userfault(vmf, mapping,
6274                                                         VM_UFFD_MISSING);
6275                 }
6276
6277                 folio = alloc_hugetlb_folio(vma, haddr, 0);
6278                 if (IS_ERR(folio)) {
6279                         /*
6280                          * Returning error will result in faulting task being
6281                          * sent SIGBUS.  The hugetlb fault mutex prevents two
6282                          * tasks from racing to fault in the same page which
6283                          * could result in false unable to allocate errors.
6284                          * Page migration does not take the fault mutex, but
6285                          * does a clear then write of pte's under page table
6286                          * lock.  Page fault code could race with migration,
6287                          * notice the clear pte and try to allocate a page
6288                          * here.  Before returning error, get ptl and make
6289                          * sure there really is no pte entry.
6290                          */
6291                         if (hugetlb_pte_stable(h, mm, ptep, old_pte))
6292                                 ret = vmf_error(PTR_ERR(folio));
6293                         else
6294                                 ret = 0;
6295                         goto out;
6296                 }
6297                 clear_huge_page(&folio->page, address, pages_per_huge_page(h));
6298                 __folio_mark_uptodate(folio);
6299                 new_folio = true;
6300
6301                 if (vma->vm_flags & VM_MAYSHARE) {
6302                         int err = hugetlb_add_to_page_cache(folio, mapping, idx);
6303                         if (err) {
6304                                 /*
6305                                  * err can't be -EEXIST which implies someone
6306                                  * else consumed the reservation since hugetlb
6307                                  * fault mutex is held when add a hugetlb page
6308                                  * to the page cache. So it's safe to call
6309                                  * restore_reserve_on_error() here.
6310                                  */
6311                                 restore_reserve_on_error(h, vma, haddr, folio);
6312                                 folio_put(folio);
6313                                 goto out;
6314                         }
6315                         new_pagecache_folio = true;
6316                 } else {
6317                         folio_lock(folio);
6318
6319                         ret = vmf_anon_prepare(vmf);
6320                         if (unlikely(ret))
6321                                 goto backout_unlocked;
6322                         anon_rmap = 1;
6323                 }
6324         } else {
6325                 /*
6326                  * If memory error occurs between mmap() and fault, some process
6327                  * don't have hwpoisoned swap entry for errored virtual address.
6328                  * So we need to block hugepage fault by PG_hwpoison bit check.
6329                  */
6330                 if (unlikely(folio_test_hwpoison(folio))) {
6331                         ret = VM_FAULT_HWPOISON_LARGE |
6332                                 VM_FAULT_SET_HINDEX(hstate_index(h));
6333                         goto backout_unlocked;
6334                 }
6335
6336                 /* Check for page in userfault range. */
6337                 if (userfaultfd_minor(vma)) {
6338                         folio_unlock(folio);
6339                         folio_put(folio);
6340                         /* See comment in userfaultfd_missing() block above */
6341                         if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
6342                                 ret = 0;
6343                                 goto out;
6344                         }
6345                         return hugetlb_handle_userfault(vmf, mapping,
6346                                                         VM_UFFD_MINOR);
6347                 }
6348         }
6349
6350         /*
6351          * If we are going to COW a private mapping later, we examine the
6352          * pending reservations for this page now. This will ensure that
6353          * any allocations necessary to record that reservation occur outside
6354          * the spinlock.
6355          */
6356         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
6357                 if (vma_needs_reservation(h, vma, haddr) < 0) {
6358                         ret = VM_FAULT_OOM;
6359                         goto backout_unlocked;
6360                 }
6361                 /* Just decrements count, does not deallocate */
6362                 vma_end_reservation(h, vma, haddr);
6363         }
6364
6365         ptl = huge_pte_lock(h, mm, ptep);
6366         ret = 0;
6367         /* If pte changed from under us, retry */
6368         if (!pte_same(huge_ptep_get(ptep), old_pte))
6369                 goto backout;
6370
6371         if (anon_rmap)
6372                 hugetlb_add_new_anon_rmap(folio, vma, haddr);
6373         else
6374                 hugetlb_add_file_rmap(folio);
6375         new_pte = make_huge_pte(vma, &folio->page, ((vma->vm_flags & VM_WRITE)
6376                                 && (vma->vm_flags & VM_SHARED)));
6377         /*
6378          * If this pte was previously wr-protected, keep it wr-protected even
6379          * if populated.
6380          */
6381         if (unlikely(pte_marker_uffd_wp(old_pte)))
6382                 new_pte = huge_pte_mkuffd_wp(new_pte);
6383         set_huge_pte_at(mm, haddr, ptep, new_pte, huge_page_size(h));
6384
6385         hugetlb_count_add(pages_per_huge_page(h), mm);
6386         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
6387                 /* Optimization, do the COW without a second fault */
6388                 ret = hugetlb_wp(mm, vma, address, ptep, flags, folio, ptl, vmf);
6389         }
6390
6391         spin_unlock(ptl);
6392
6393         /*
6394          * Only set hugetlb_migratable in newly allocated pages.  Existing pages
6395          * found in the pagecache may not have hugetlb_migratable if they have
6396          * been isolated for migration.
6397          */
6398         if (new_folio)
6399                 folio_set_hugetlb_migratable(folio);
6400
6401         folio_unlock(folio);
6402 out:
6403         hugetlb_vma_unlock_read(vma);
6404         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6405         return ret;
6406
6407 backout:
6408         spin_unlock(ptl);
6409 backout_unlocked:
6410         if (new_folio && !new_pagecache_folio)
6411                 restore_reserve_on_error(h, vma, haddr, folio);
6412
6413         folio_unlock(folio);
6414         folio_put(folio);
6415         goto out;
6416 }
6417
6418 #ifdef CONFIG_SMP
6419 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
6420 {
6421         unsigned long key[2];
6422         u32 hash;
6423
6424         key[0] = (unsigned long) mapping;
6425         key[1] = idx;
6426
6427         hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
6428
6429         return hash & (num_fault_mutexes - 1);
6430 }
6431 #else
6432 /*
6433  * For uniprocessor systems we always use a single mutex, so just
6434  * return 0 and avoid the hashing overhead.
6435  */
6436 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
6437 {
6438         return 0;
6439 }
6440 #endif
6441
6442 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
6443                         unsigned long address, unsigned int flags)
6444 {
6445         pte_t *ptep, entry;
6446         spinlock_t *ptl;
6447         vm_fault_t ret;
6448         u32 hash;
6449         struct folio *folio = NULL;
6450         struct folio *pagecache_folio = NULL;
6451         struct hstate *h = hstate_vma(vma);
6452         struct address_space *mapping;
6453         int need_wait_lock = 0;
6454         unsigned long haddr = address & huge_page_mask(h);
6455         struct vm_fault vmf = {
6456                 .vma = vma,
6457                 .address = haddr,
6458                 .real_address = address,
6459                 .flags = flags,
6460                 .pgoff = vma_hugecache_offset(h, vma, haddr),
6461                 /* TODO: Track hugetlb faults using vm_fault */
6462
6463                 /*
6464                  * Some fields may not be initialized, be careful as it may
6465                  * be hard to debug if called functions make assumptions
6466                  */
6467         };
6468
6469         /*
6470          * Serialize hugepage allocation and instantiation, so that we don't
6471          * get spurious allocation failures if two CPUs race to instantiate
6472          * the same page in the page cache.
6473          */
6474         mapping = vma->vm_file->f_mapping;
6475         hash = hugetlb_fault_mutex_hash(mapping, vmf.pgoff);
6476         mutex_lock(&hugetlb_fault_mutex_table[hash]);
6477
6478         /*
6479          * Acquire vma lock before calling huge_pte_alloc and hold
6480          * until finished with ptep.  This prevents huge_pmd_unshare from
6481          * being called elsewhere and making the ptep no longer valid.
6482          */
6483         hugetlb_vma_lock_read(vma);
6484         ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
6485         if (!ptep) {
6486                 hugetlb_vma_unlock_read(vma);
6487                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6488                 return VM_FAULT_OOM;
6489         }
6490
6491         entry = huge_ptep_get(ptep);
6492         if (huge_pte_none_mostly(entry)) {
6493                 if (is_pte_marker(entry)) {
6494                         pte_marker marker =
6495                                 pte_marker_get(pte_to_swp_entry(entry));
6496
6497                         if (marker & PTE_MARKER_POISONED) {
6498                                 ret = VM_FAULT_HWPOISON_LARGE;
6499                                 goto out_mutex;
6500                         }
6501                 }
6502
6503                 /*
6504                  * Other PTE markers should be handled the same way as none PTE.
6505                  *
6506                  * hugetlb_no_page will drop vma lock and hugetlb fault
6507                  * mutex internally, which make us return immediately.
6508                  */
6509                 return hugetlb_no_page(mm, vma, mapping, vmf.pgoff, address,
6510                                         ptep, entry, flags, &vmf);
6511         }
6512
6513         ret = 0;
6514
6515         /*
6516          * entry could be a migration/hwpoison entry at this point, so this
6517          * check prevents the kernel from going below assuming that we have
6518          * an active hugepage in pagecache. This goto expects the 2nd page
6519          * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
6520          * properly handle it.
6521          */
6522         if (!pte_present(entry)) {
6523                 if (unlikely(is_hugetlb_entry_migration(entry))) {
6524                         /*
6525                          * Release the hugetlb fault lock now, but retain
6526                          * the vma lock, because it is needed to guard the
6527                          * huge_pte_lockptr() later in
6528                          * migration_entry_wait_huge(). The vma lock will
6529                          * be released there.
6530                          */
6531                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6532                         migration_entry_wait_huge(vma, ptep);
6533                         return 0;
6534                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
6535                         ret = VM_FAULT_HWPOISON_LARGE |
6536                             VM_FAULT_SET_HINDEX(hstate_index(h));
6537                 goto out_mutex;
6538         }
6539
6540         /*
6541          * If we are going to COW/unshare the mapping later, we examine the
6542          * pending reservations for this page now. This will ensure that any
6543          * allocations necessary to record that reservation occur outside the
6544          * spinlock. Also lookup the pagecache page now as it is used to
6545          * determine if a reservation has been consumed.
6546          */
6547         if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
6548             !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) {
6549                 if (vma_needs_reservation(h, vma, haddr) < 0) {
6550                         ret = VM_FAULT_OOM;
6551                         goto out_mutex;
6552                 }
6553                 /* Just decrements count, does not deallocate */
6554                 vma_end_reservation(h, vma, haddr);
6555
6556                 pagecache_folio = filemap_lock_hugetlb_folio(h, mapping,
6557                                                              vmf.pgoff);
6558                 if (IS_ERR(pagecache_folio))
6559                         pagecache_folio = NULL;
6560         }
6561
6562         ptl = huge_pte_lock(h, mm, ptep);
6563
6564         /* Check for a racing update before calling hugetlb_wp() */
6565         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
6566                 goto out_ptl;
6567
6568         /* Handle userfault-wp first, before trying to lock more pages */
6569         if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) &&
6570             (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
6571                 if (!userfaultfd_wp_async(vma)) {
6572                         spin_unlock(ptl);
6573                         if (pagecache_folio) {
6574                                 folio_unlock(pagecache_folio);
6575                                 folio_put(pagecache_folio);
6576                         }
6577                         hugetlb_vma_unlock_read(vma);
6578                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6579                         return handle_userfault(&vmf, VM_UFFD_WP);
6580                 }
6581
6582                 entry = huge_pte_clear_uffd_wp(entry);
6583                 set_huge_pte_at(mm, haddr, ptep, entry,
6584                                 huge_page_size(hstate_vma(vma)));
6585                 /* Fallthrough to CoW */
6586         }
6587
6588         /*
6589          * hugetlb_wp() requires page locks of pte_page(entry) and
6590          * pagecache_folio, so here we need take the former one
6591          * when folio != pagecache_folio or !pagecache_folio.
6592          */
6593         folio = page_folio(pte_page(entry));
6594         if (folio != pagecache_folio)
6595                 if (!folio_trylock(folio)) {
6596                         need_wait_lock = 1;
6597                         goto out_ptl;
6598                 }
6599
6600         folio_get(folio);
6601
6602         if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
6603                 if (!huge_pte_write(entry)) {
6604                         ret = hugetlb_wp(mm, vma, address, ptep, flags,
6605                                          pagecache_folio, ptl, &vmf);
6606                         goto out_put_page;
6607                 } else if (likely(flags & FAULT_FLAG_WRITE)) {
6608                         entry = huge_pte_mkdirty(entry);
6609                 }
6610         }
6611         entry = pte_mkyoung(entry);
6612         if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
6613                                                 flags & FAULT_FLAG_WRITE))
6614                 update_mmu_cache(vma, haddr, ptep);
6615 out_put_page:
6616         if (folio != pagecache_folio)
6617                 folio_unlock(folio);
6618         folio_put(folio);
6619 out_ptl:
6620         spin_unlock(ptl);
6621
6622         if (pagecache_folio) {
6623                 folio_unlock(pagecache_folio);
6624                 folio_put(pagecache_folio);
6625         }
6626 out_mutex:
6627         hugetlb_vma_unlock_read(vma);
6628         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6629         /*
6630          * Generally it's safe to hold refcount during waiting page lock. But
6631          * here we just wait to defer the next page fault to avoid busy loop and
6632          * the page is not used after unlocked before returning from the current
6633          * page fault. So we are safe from accessing freed page, even if we wait
6634          * here without taking refcount.
6635          */
6636         if (need_wait_lock)
6637                 folio_wait_locked(folio);
6638         return ret;
6639 }
6640
6641 #ifdef CONFIG_USERFAULTFD
6642 /*
6643  * Can probably be eliminated, but still used by hugetlb_mfill_atomic_pte().
6644  */
6645 static struct folio *alloc_hugetlb_folio_vma(struct hstate *h,
6646                 struct vm_area_struct *vma, unsigned long address)
6647 {
6648         struct mempolicy *mpol;
6649         nodemask_t *nodemask;
6650         struct folio *folio;
6651         gfp_t gfp_mask;
6652         int node;
6653
6654         gfp_mask = htlb_alloc_mask(h);
6655         node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
6656         folio = alloc_hugetlb_folio_nodemask(h, node, nodemask, gfp_mask);
6657         mpol_cond_put(mpol);
6658
6659         return folio;
6660 }
6661
6662 /*
6663  * Used by userfaultfd UFFDIO_* ioctls. Based on userfaultfd's mfill_atomic_pte
6664  * with modifications for hugetlb pages.
6665  */
6666 int hugetlb_mfill_atomic_pte(pte_t *dst_pte,
6667                              struct vm_area_struct *dst_vma,
6668                              unsigned long dst_addr,
6669                              unsigned long src_addr,
6670                              uffd_flags_t flags,
6671                              struct folio **foliop)
6672 {
6673         struct mm_struct *dst_mm = dst_vma->vm_mm;
6674         bool is_continue = uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE);
6675         bool wp_enabled = (flags & MFILL_ATOMIC_WP);
6676         struct hstate *h = hstate_vma(dst_vma);
6677         struct address_space *mapping = dst_vma->vm_file->f_mapping;
6678         pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
6679         unsigned long size;
6680         int vm_shared = dst_vma->vm_flags & VM_SHARED;
6681         pte_t _dst_pte;
6682         spinlock_t *ptl;
6683         int ret = -ENOMEM;
6684         struct folio *folio;
6685         int writable;
6686         bool folio_in_pagecache = false;
6687
6688         if (uffd_flags_mode_is(flags, MFILL_ATOMIC_POISON)) {
6689                 ptl = huge_pte_lock(h, dst_mm, dst_pte);
6690
6691                 /* Don't overwrite any existing PTEs (even markers) */
6692                 if (!huge_pte_none(huge_ptep_get(dst_pte))) {
6693                         spin_unlock(ptl);
6694                         return -EEXIST;
6695                 }
6696
6697                 _dst_pte = make_pte_marker(PTE_MARKER_POISONED);
6698                 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte,
6699                                 huge_page_size(h));
6700
6701                 /* No need to invalidate - it was non-present before */
6702                 update_mmu_cache(dst_vma, dst_addr, dst_pte);
6703
6704                 spin_unlock(ptl);
6705                 return 0;
6706         }
6707
6708         if (is_continue) {
6709                 ret = -EFAULT;
6710                 folio = filemap_lock_hugetlb_folio(h, mapping, idx);
6711                 if (IS_ERR(folio))
6712                         goto out;
6713                 folio_in_pagecache = true;
6714         } else if (!*foliop) {
6715                 /* If a folio already exists, then it's UFFDIO_COPY for
6716                  * a non-missing case. Return -EEXIST.
6717                  */
6718                 if (vm_shared &&
6719                     hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6720                         ret = -EEXIST;
6721                         goto out;
6722                 }
6723
6724                 folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0);
6725                 if (IS_ERR(folio)) {
6726                         ret = -ENOMEM;
6727                         goto out;
6728                 }
6729
6730                 ret = copy_folio_from_user(folio, (const void __user *) src_addr,
6731                                            false);
6732
6733                 /* fallback to copy_from_user outside mmap_lock */
6734                 if (unlikely(ret)) {
6735                         ret = -ENOENT;
6736                         /* Free the allocated folio which may have
6737                          * consumed a reservation.
6738                          */
6739                         restore_reserve_on_error(h, dst_vma, dst_addr, folio);
6740                         folio_put(folio);
6741
6742                         /* Allocate a temporary folio to hold the copied
6743                          * contents.
6744                          */
6745                         folio = alloc_hugetlb_folio_vma(h, dst_vma, dst_addr);
6746                         if (!folio) {
6747                                 ret = -ENOMEM;
6748                                 goto out;
6749                         }
6750                         *foliop = folio;
6751                         /* Set the outparam foliop and return to the caller to
6752                          * copy the contents outside the lock. Don't free the
6753                          * folio.
6754                          */
6755                         goto out;
6756                 }
6757         } else {
6758                 if (vm_shared &&
6759                     hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6760                         folio_put(*foliop);
6761                         ret = -EEXIST;
6762                         *foliop = NULL;
6763                         goto out;
6764                 }
6765
6766                 folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0);
6767                 if (IS_ERR(folio)) {
6768                         folio_put(*foliop);
6769                         ret = -ENOMEM;
6770                         *foliop = NULL;
6771                         goto out;
6772                 }
6773                 ret = copy_user_large_folio(folio, *foliop, dst_addr, dst_vma);
6774                 folio_put(*foliop);
6775                 *foliop = NULL;
6776                 if (ret) {
6777                         folio_put(folio);
6778                         goto out;
6779                 }
6780         }
6781
6782         /*
6783          * If we just allocated a new page, we need a memory barrier to ensure
6784          * that preceding stores to the page become visible before the
6785          * set_pte_at() write. The memory barrier inside __folio_mark_uptodate
6786          * is what we need.
6787          *
6788          * In the case where we have not allocated a new page (is_continue),
6789          * the page must already be uptodate. UFFDIO_CONTINUE already includes
6790          * an earlier smp_wmb() to ensure that prior stores will be visible
6791          * before the set_pte_at() write.
6792          */
6793         if (!is_continue)
6794                 __folio_mark_uptodate(folio);
6795         else
6796                 WARN_ON_ONCE(!folio_test_uptodate(folio));
6797
6798         /* Add shared, newly allocated pages to the page cache. */
6799         if (vm_shared && !is_continue) {
6800                 size = i_size_read(mapping->host) >> huge_page_shift(h);
6801                 ret = -EFAULT;
6802                 if (idx >= size)
6803                         goto out_release_nounlock;
6804
6805                 /*
6806                  * Serialization between remove_inode_hugepages() and
6807                  * hugetlb_add_to_page_cache() below happens through the
6808                  * hugetlb_fault_mutex_table that here must be hold by
6809                  * the caller.
6810                  */
6811                 ret = hugetlb_add_to_page_cache(folio, mapping, idx);
6812                 if (ret)
6813                         goto out_release_nounlock;
6814                 folio_in_pagecache = true;
6815         }
6816
6817         ptl = huge_pte_lock(h, dst_mm, dst_pte);
6818
6819         ret = -EIO;
6820         if (folio_test_hwpoison(folio))
6821                 goto out_release_unlock;
6822
6823         /*
6824          * We allow to overwrite a pte marker: consider when both MISSING|WP
6825          * registered, we firstly wr-protect a none pte which has no page cache
6826          * page backing it, then access the page.
6827          */
6828         ret = -EEXIST;
6829         if (!huge_pte_none_mostly(huge_ptep_get(dst_pte)))
6830                 goto out_release_unlock;
6831
6832         if (folio_in_pagecache)
6833                 hugetlb_add_file_rmap(folio);
6834         else
6835                 hugetlb_add_new_anon_rmap(folio, dst_vma, dst_addr);
6836
6837         /*
6838          * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY
6839          * with wp flag set, don't set pte write bit.
6840          */
6841         if (wp_enabled || (is_continue && !vm_shared))
6842                 writable = 0;
6843         else
6844                 writable = dst_vma->vm_flags & VM_WRITE;
6845
6846         _dst_pte = make_huge_pte(dst_vma, &folio->page, writable);
6847         /*
6848          * Always mark UFFDIO_COPY page dirty; note that this may not be
6849          * extremely important for hugetlbfs for now since swapping is not
6850          * supported, but we should still be clear in that this page cannot be
6851          * thrown away at will, even if write bit not set.
6852          */
6853         _dst_pte = huge_pte_mkdirty(_dst_pte);
6854         _dst_pte = pte_mkyoung(_dst_pte);
6855
6856         if (wp_enabled)
6857                 _dst_pte = huge_pte_mkuffd_wp(_dst_pte);
6858
6859         set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte, huge_page_size(h));
6860
6861         hugetlb_count_add(pages_per_huge_page(h), dst_mm);
6862
6863         /* No need to invalidate - it was non-present before */
6864         update_mmu_cache(dst_vma, dst_addr, dst_pte);
6865
6866         spin_unlock(ptl);
6867         if (!is_continue)
6868                 folio_set_hugetlb_migratable(folio);
6869         if (vm_shared || is_continue)
6870                 folio_unlock(folio);
6871         ret = 0;
6872 out:
6873         return ret;
6874 out_release_unlock:
6875         spin_unlock(ptl);
6876         if (vm_shared || is_continue)
6877                 folio_unlock(folio);
6878 out_release_nounlock:
6879         if (!folio_in_pagecache)
6880                 restore_reserve_on_error(h, dst_vma, dst_addr, folio);
6881         folio_put(folio);
6882         goto out;
6883 }
6884 #endif /* CONFIG_USERFAULTFD */
6885
6886 struct page *hugetlb_follow_page_mask(struct vm_area_struct *vma,
6887                                       unsigned long address, unsigned int flags,
6888                                       unsigned int *page_mask)
6889 {
6890         struct hstate *h = hstate_vma(vma);
6891         struct mm_struct *mm = vma->vm_mm;
6892         unsigned long haddr = address & huge_page_mask(h);
6893         struct page *page = NULL;
6894         spinlock_t *ptl;
6895         pte_t *pte, entry;
6896         int ret;
6897
6898         hugetlb_vma_lock_read(vma);
6899         pte = hugetlb_walk(vma, haddr, huge_page_size(h));
6900         if (!pte)
6901                 goto out_unlock;
6902
6903         ptl = huge_pte_lock(h, mm, pte);
6904         entry = huge_ptep_get(pte);
6905         if (pte_present(entry)) {
6906                 page = pte_page(entry);
6907
6908                 if (!huge_pte_write(entry)) {
6909                         if (flags & FOLL_WRITE) {
6910                                 page = NULL;
6911                                 goto out;
6912                         }
6913
6914                         if (gup_must_unshare(vma, flags, page)) {
6915                                 /* Tell the caller to do unsharing */
6916                                 page = ERR_PTR(-EMLINK);
6917                                 goto out;
6918                         }
6919                 }
6920
6921                 page = nth_page(page, ((address & ~huge_page_mask(h)) >> PAGE_SHIFT));
6922
6923                 /*
6924                  * Note that page may be a sub-page, and with vmemmap
6925                  * optimizations the page struct may be read only.
6926                  * try_grab_page() will increase the ref count on the
6927                  * head page, so this will be OK.
6928                  *
6929                  * try_grab_page() should always be able to get the page here,
6930                  * because we hold the ptl lock and have verified pte_present().
6931                  */
6932                 ret = try_grab_page(page, flags);
6933
6934                 if (WARN_ON_ONCE(ret)) {
6935                         page = ERR_PTR(ret);
6936                         goto out;
6937                 }
6938
6939                 *page_mask = (1U << huge_page_order(h)) - 1;
6940         }
6941 out:
6942         spin_unlock(ptl);
6943 out_unlock:
6944         hugetlb_vma_unlock_read(vma);
6945
6946         /*
6947          * Fixup retval for dump requests: if pagecache doesn't exist,
6948          * don't try to allocate a new page but just skip it.
6949          */
6950         if (!page && (flags & FOLL_DUMP) &&
6951             !hugetlbfs_pagecache_present(h, vma, address))
6952                 page = ERR_PTR(-EFAULT);
6953
6954         return page;
6955 }
6956
6957 long hugetlb_change_protection(struct vm_area_struct *vma,
6958                 unsigned long address, unsigned long end,
6959                 pgprot_t newprot, unsigned long cp_flags)
6960 {
6961         struct mm_struct *mm = vma->vm_mm;
6962         unsigned long start = address;
6963         pte_t *ptep;
6964         pte_t pte;
6965         struct hstate *h = hstate_vma(vma);
6966         long pages = 0, psize = huge_page_size(h);
6967         bool shared_pmd = false;
6968         struct mmu_notifier_range range;
6969         unsigned long last_addr_mask;
6970         bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
6971         bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
6972
6973         /*
6974          * In the case of shared PMDs, the area to flush could be beyond
6975          * start/end.  Set range.start/range.end to cover the maximum possible
6976          * range if PMD sharing is possible.
6977          */
6978         mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
6979                                 0, mm, start, end);
6980         adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
6981
6982         BUG_ON(address >= end);
6983         flush_cache_range(vma, range.start, range.end);
6984
6985         mmu_notifier_invalidate_range_start(&range);
6986         hugetlb_vma_lock_write(vma);
6987         i_mmap_lock_write(vma->vm_file->f_mapping);
6988         last_addr_mask = hugetlb_mask_last_page(h);
6989         for (; address < end; address += psize) {
6990                 spinlock_t *ptl;
6991                 ptep = hugetlb_walk(vma, address, psize);
6992                 if (!ptep) {
6993                         if (!uffd_wp) {
6994                                 address |= last_addr_mask;
6995                                 continue;
6996                         }
6997                         /*
6998                          * Userfaultfd wr-protect requires pgtable
6999                          * pre-allocations to install pte markers.
7000                          */
7001                         ptep = huge_pte_alloc(mm, vma, address, psize);
7002                         if (!ptep) {
7003                                 pages = -ENOMEM;
7004                                 break;
7005                         }
7006                 }
7007                 ptl = huge_pte_lock(h, mm, ptep);
7008                 if (huge_pmd_unshare(mm, vma, address, ptep)) {
7009                         /*
7010                          * When uffd-wp is enabled on the vma, unshare
7011                          * shouldn't happen at all.  Warn about it if it
7012                          * happened due to some reason.
7013                          */
7014                         WARN_ON_ONCE(uffd_wp || uffd_wp_resolve);
7015                         pages++;
7016                         spin_unlock(ptl);
7017                         shared_pmd = true;
7018                         address |= last_addr_mask;
7019                         continue;
7020                 }
7021                 pte = huge_ptep_get(ptep);
7022                 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
7023                         /* Nothing to do. */
7024                 } else if (unlikely(is_hugetlb_entry_migration(pte))) {
7025                         swp_entry_t entry = pte_to_swp_entry(pte);
7026                         struct page *page = pfn_swap_entry_to_page(entry);
7027                         pte_t newpte = pte;
7028
7029                         if (is_writable_migration_entry(entry)) {
7030                                 if (PageAnon(page))
7031                                         entry = make_readable_exclusive_migration_entry(
7032                                                                 swp_offset(entry));
7033                                 else
7034                                         entry = make_readable_migration_entry(
7035                                                                 swp_offset(entry));
7036                                 newpte = swp_entry_to_pte(entry);
7037                                 pages++;
7038                         }
7039
7040                         if (uffd_wp)
7041                                 newpte = pte_swp_mkuffd_wp(newpte);
7042                         else if (uffd_wp_resolve)
7043                                 newpte = pte_swp_clear_uffd_wp(newpte);
7044                         if (!pte_same(pte, newpte))
7045                                 set_huge_pte_at(mm, address, ptep, newpte, psize);
7046                 } else if (unlikely(is_pte_marker(pte))) {
7047                         /* No other markers apply for now. */
7048                         WARN_ON_ONCE(!pte_marker_uffd_wp(pte));
7049                         if (uffd_wp_resolve)
7050                                 /* Safe to modify directly (non-present->none). */
7051                                 huge_pte_clear(mm, address, ptep, psize);
7052                 } else if (!huge_pte_none(pte)) {
7053                         pte_t old_pte;
7054                         unsigned int shift = huge_page_shift(hstate_vma(vma));
7055
7056                         old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
7057                         pte = huge_pte_modify(old_pte, newprot);
7058                         pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
7059                         if (uffd_wp)
7060                                 pte = huge_pte_mkuffd_wp(pte);
7061                         else if (uffd_wp_resolve)
7062                                 pte = huge_pte_clear_uffd_wp(pte);
7063                         huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
7064                         pages++;
7065                 } else {
7066                         /* None pte */
7067                         if (unlikely(uffd_wp))
7068                                 /* Safe to modify directly (none->non-present). */
7069                                 set_huge_pte_at(mm, address, ptep,
7070                                                 make_pte_marker(PTE_MARKER_UFFD_WP),
7071                                                 psize);
7072                 }
7073                 spin_unlock(ptl);
7074         }
7075         /*
7076          * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
7077          * may have cleared our pud entry and done put_page on the page table:
7078          * once we release i_mmap_rwsem, another task can do the final put_page
7079          * and that page table be reused and filled with junk.  If we actually
7080          * did unshare a page of pmds, flush the range corresponding to the pud.
7081          */
7082         if (shared_pmd)
7083                 flush_hugetlb_tlb_range(vma, range.start, range.end);
7084         else
7085                 flush_hugetlb_tlb_range(vma, start, end);
7086         /*
7087          * No need to call mmu_notifier_arch_invalidate_secondary_tlbs() we are
7088          * downgrading page table protection not changing it to point to a new
7089          * page.
7090          *
7091          * See Documentation/mm/mmu_notifier.rst
7092          */
7093         i_mmap_unlock_write(vma->vm_file->f_mapping);
7094         hugetlb_vma_unlock_write(vma);
7095         mmu_notifier_invalidate_range_end(&range);
7096
7097         return pages > 0 ? (pages << h->order) : pages;
7098 }
7099
7100 /* Return true if reservation was successful, false otherwise.  */
7101 bool hugetlb_reserve_pages(struct inode *inode,
7102                                         long from, long to,
7103                                         struct vm_area_struct *vma,
7104                                         vm_flags_t vm_flags)
7105 {
7106         long chg = -1, add = -1;
7107         struct hstate *h = hstate_inode(inode);
7108         struct hugepage_subpool *spool = subpool_inode(inode);
7109         struct resv_map *resv_map;
7110         struct hugetlb_cgroup *h_cg = NULL;
7111         long gbl_reserve, regions_needed = 0;
7112
7113         /* This should never happen */
7114         if (from > to) {
7115                 VM_WARN(1, "%s called with a negative range\n", __func__);
7116                 return false;
7117         }
7118
7119         /*
7120          * vma specific semaphore used for pmd sharing and fault/truncation
7121          * synchronization
7122          */
7123         hugetlb_vma_lock_alloc(vma);
7124
7125         /*
7126          * Only apply hugepage reservation if asked. At fault time, an
7127          * attempt will be made for VM_NORESERVE to allocate a page
7128          * without using reserves
7129          */
7130         if (vm_flags & VM_NORESERVE)
7131                 return true;
7132
7133         /*
7134          * Shared mappings base their reservation on the number of pages that
7135          * are already allocated on behalf of the file. Private mappings need
7136          * to reserve the full area even if read-only as mprotect() may be
7137          * called to make the mapping read-write. Assume !vma is a shm mapping
7138          */
7139         if (!vma || vma->vm_flags & VM_MAYSHARE) {
7140                 /*
7141                  * resv_map can not be NULL as hugetlb_reserve_pages is only
7142                  * called for inodes for which resv_maps were created (see
7143                  * hugetlbfs_get_inode).
7144                  */
7145                 resv_map = inode_resv_map(inode);
7146
7147                 chg = region_chg(resv_map, from, to, &regions_needed);
7148         } else {
7149                 /* Private mapping. */
7150                 resv_map = resv_map_alloc();
7151                 if (!resv_map)
7152                         goto out_err;
7153
7154                 chg = to - from;
7155
7156                 set_vma_resv_map(vma, resv_map);
7157                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
7158         }
7159
7160         if (chg < 0)
7161                 goto out_err;
7162
7163         if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
7164                                 chg * pages_per_huge_page(h), &h_cg) < 0)
7165                 goto out_err;
7166
7167         if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
7168                 /* For private mappings, the hugetlb_cgroup uncharge info hangs
7169                  * of the resv_map.
7170                  */
7171                 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
7172         }
7173
7174         /*
7175          * There must be enough pages in the subpool for the mapping. If
7176          * the subpool has a minimum size, there may be some global
7177          * reservations already in place (gbl_reserve).
7178          */
7179         gbl_reserve = hugepage_subpool_get_pages(spool, chg);
7180         if (gbl_reserve < 0)
7181                 goto out_uncharge_cgroup;
7182
7183         /*
7184          * Check enough hugepages are available for the reservation.
7185          * Hand the pages back to the subpool if there are not
7186          */
7187         if (hugetlb_acct_memory(h, gbl_reserve) < 0)
7188                 goto out_put_pages;
7189
7190         /*
7191          * Account for the reservations made. Shared mappings record regions
7192          * that have reservations as they are shared by multiple VMAs.
7193          * When the last VMA disappears, the region map says how much
7194          * the reservation was and the page cache tells how much of
7195          * the reservation was consumed. Private mappings are per-VMA and
7196          * only the consumed reservations are tracked. When the VMA
7197          * disappears, the original reservation is the VMA size and the
7198          * consumed reservations are stored in the map. Hence, nothing
7199          * else has to be done for private mappings here
7200          */
7201         if (!vma || vma->vm_flags & VM_MAYSHARE) {
7202                 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
7203
7204                 if (unlikely(add < 0)) {
7205                         hugetlb_acct_memory(h, -gbl_reserve);
7206                         goto out_put_pages;
7207                 } else if (unlikely(chg > add)) {
7208                         /*
7209                          * pages in this range were added to the reserve
7210                          * map between region_chg and region_add.  This
7211                          * indicates a race with alloc_hugetlb_folio.  Adjust
7212                          * the subpool and reserve counts modified above
7213                          * based on the difference.
7214                          */
7215                         long rsv_adjust;
7216
7217                         /*
7218                          * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
7219                          * reference to h_cg->css. See comment below for detail.
7220                          */
7221                         hugetlb_cgroup_uncharge_cgroup_rsvd(
7222                                 hstate_index(h),
7223                                 (chg - add) * pages_per_huge_page(h), h_cg);
7224
7225                         rsv_adjust = hugepage_subpool_put_pages(spool,
7226                                                                 chg - add);
7227                         hugetlb_acct_memory(h, -rsv_adjust);
7228                 } else if (h_cg) {
7229                         /*
7230                          * The file_regions will hold their own reference to
7231                          * h_cg->css. So we should release the reference held
7232                          * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
7233                          * done.
7234                          */
7235                         hugetlb_cgroup_put_rsvd_cgroup(h_cg);
7236                 }
7237         }
7238         return true;
7239
7240 out_put_pages:
7241         /* put back original number of pages, chg */
7242         (void)hugepage_subpool_put_pages(spool, chg);
7243 out_uncharge_cgroup:
7244         hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
7245                                             chg * pages_per_huge_page(h), h_cg);
7246 out_err:
7247         hugetlb_vma_lock_free(vma);
7248         if (!vma || vma->vm_flags & VM_MAYSHARE)
7249                 /* Only call region_abort if the region_chg succeeded but the
7250                  * region_add failed or didn't run.
7251                  */
7252                 if (chg >= 0 && add < 0)
7253                         region_abort(resv_map, from, to, regions_needed);
7254         if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
7255                 kref_put(&resv_map->refs, resv_map_release);
7256                 set_vma_resv_map(vma, NULL);
7257         }
7258         return false;
7259 }
7260
7261 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
7262                                                                 long freed)
7263 {
7264         struct hstate *h = hstate_inode(inode);
7265         struct resv_map *resv_map = inode_resv_map(inode);
7266         long chg = 0;
7267         struct hugepage_subpool *spool = subpool_inode(inode);
7268         long gbl_reserve;
7269
7270         /*
7271          * Since this routine can be called in the evict inode path for all
7272          * hugetlbfs inodes, resv_map could be NULL.
7273          */
7274         if (resv_map) {
7275                 chg = region_del(resv_map, start, end);
7276                 /*
7277                  * region_del() can fail in the rare case where a region
7278                  * must be split and another region descriptor can not be
7279                  * allocated.  If end == LONG_MAX, it will not fail.
7280                  */
7281                 if (chg < 0)
7282                         return chg;
7283         }
7284
7285         spin_lock(&inode->i_lock);
7286         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
7287         spin_unlock(&inode->i_lock);
7288
7289         /*
7290          * If the subpool has a minimum size, the number of global
7291          * reservations to be released may be adjusted.
7292          *
7293          * Note that !resv_map implies freed == 0. So (chg - freed)
7294          * won't go negative.
7295          */
7296         gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
7297         hugetlb_acct_memory(h, -gbl_reserve);
7298
7299         return 0;
7300 }
7301
7302 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
7303 static unsigned long page_table_shareable(struct vm_area_struct *svma,
7304                                 struct vm_area_struct *vma,
7305                                 unsigned long addr, pgoff_t idx)
7306 {
7307         unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
7308                                 svma->vm_start;
7309         unsigned long sbase = saddr & PUD_MASK;
7310         unsigned long s_end = sbase + PUD_SIZE;
7311
7312         /* Allow segments to share if only one is marked locked */
7313         unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED_MASK;
7314         unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED_MASK;
7315
7316         /*
7317          * match the virtual addresses, permission and the alignment of the
7318          * page table page.
7319          *
7320          * Also, vma_lock (vm_private_data) is required for sharing.
7321          */
7322         if (pmd_index(addr) != pmd_index(saddr) ||
7323             vm_flags != svm_flags ||
7324             !range_in_vma(svma, sbase, s_end) ||
7325             !svma->vm_private_data)
7326                 return 0;
7327
7328         return saddr;
7329 }
7330
7331 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
7332 {
7333         unsigned long start = addr & PUD_MASK;
7334         unsigned long end = start + PUD_SIZE;
7335
7336 #ifdef CONFIG_USERFAULTFD
7337         if (uffd_disable_huge_pmd_share(vma))
7338                 return false;
7339 #endif
7340         /*
7341          * check on proper vm_flags and page table alignment
7342          */
7343         if (!(vma->vm_flags & VM_MAYSHARE))
7344                 return false;
7345         if (!vma->vm_private_data)      /* vma lock required for sharing */
7346                 return false;
7347         if (!range_in_vma(vma, start, end))
7348                 return false;
7349         return true;
7350 }
7351
7352 /*
7353  * Determine if start,end range within vma could be mapped by shared pmd.
7354  * If yes, adjust start and end to cover range associated with possible
7355  * shared pmd mappings.
7356  */
7357 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
7358                                 unsigned long *start, unsigned long *end)
7359 {
7360         unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
7361                 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
7362
7363         /*
7364          * vma needs to span at least one aligned PUD size, and the range
7365          * must be at least partially within in.
7366          */
7367         if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
7368                 (*end <= v_start) || (*start >= v_end))
7369                 return;
7370
7371         /* Extend the range to be PUD aligned for a worst case scenario */
7372         if (*start > v_start)
7373                 *start = ALIGN_DOWN(*start, PUD_SIZE);
7374
7375         if (*end < v_end)
7376                 *end = ALIGN(*end, PUD_SIZE);
7377 }
7378
7379 /*
7380  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
7381  * and returns the corresponding pte. While this is not necessary for the
7382  * !shared pmd case because we can allocate the pmd later as well, it makes the
7383  * code much cleaner. pmd allocation is essential for the shared case because
7384  * pud has to be populated inside the same i_mmap_rwsem section - otherwise
7385  * racing tasks could either miss the sharing (see huge_pte_offset) or select a
7386  * bad pmd for sharing.
7387  */
7388 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
7389                       unsigned long addr, pud_t *pud)
7390 {
7391         struct address_space *mapping = vma->vm_file->f_mapping;
7392         pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
7393                         vma->vm_pgoff;
7394         struct vm_area_struct *svma;
7395         unsigned long saddr;
7396         pte_t *spte = NULL;
7397         pte_t *pte;
7398
7399         i_mmap_lock_read(mapping);
7400         vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
7401                 if (svma == vma)
7402                         continue;
7403
7404                 saddr = page_table_shareable(svma, vma, addr, idx);
7405                 if (saddr) {
7406                         spte = hugetlb_walk(svma, saddr,
7407                                             vma_mmu_pagesize(svma));
7408                         if (spte) {
7409                                 get_page(virt_to_page(spte));
7410                                 break;
7411                         }
7412                 }
7413         }
7414
7415         if (!spte)
7416                 goto out;
7417
7418         spin_lock(&mm->page_table_lock);
7419         if (pud_none(*pud)) {
7420                 pud_populate(mm, pud,
7421                                 (pmd_t *)((unsigned long)spte & PAGE_MASK));
7422                 mm_inc_nr_pmds(mm);
7423         } else {
7424                 put_page(virt_to_page(spte));
7425         }
7426         spin_unlock(&mm->page_table_lock);
7427 out:
7428         pte = (pte_t *)pmd_alloc(mm, pud, addr);
7429         i_mmap_unlock_read(mapping);
7430         return pte;
7431 }
7432
7433 /*
7434  * unmap huge page backed by shared pte.
7435  *
7436  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
7437  * indicated by page_count > 1, unmap is achieved by clearing pud and
7438  * decrementing the ref count. If count == 1, the pte page is not shared.
7439  *
7440  * Called with page table lock held.
7441  *
7442  * returns: 1 successfully unmapped a shared pte page
7443  *          0 the underlying pte page is not shared, or it is the last user
7444  */
7445 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7446                                         unsigned long addr, pte_t *ptep)
7447 {
7448         pgd_t *pgd = pgd_offset(mm, addr);
7449         p4d_t *p4d = p4d_offset(pgd, addr);
7450         pud_t *pud = pud_offset(p4d, addr);
7451
7452         i_mmap_assert_write_locked(vma->vm_file->f_mapping);
7453         hugetlb_vma_assert_locked(vma);
7454         BUG_ON(page_count(virt_to_page(ptep)) == 0);
7455         if (page_count(virt_to_page(ptep)) == 1)
7456                 return 0;
7457
7458         pud_clear(pud);
7459         put_page(virt_to_page(ptep));
7460         mm_dec_nr_pmds(mm);
7461         return 1;
7462 }
7463
7464 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7465
7466 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
7467                       unsigned long addr, pud_t *pud)
7468 {
7469         return NULL;
7470 }
7471
7472 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7473                                 unsigned long addr, pte_t *ptep)
7474 {
7475         return 0;
7476 }
7477
7478 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
7479                                 unsigned long *start, unsigned long *end)
7480 {
7481 }
7482
7483 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
7484 {
7485         return false;
7486 }
7487 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7488
7489 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
7490 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
7491                         unsigned long addr, unsigned long sz)
7492 {
7493         pgd_t *pgd;
7494         p4d_t *p4d;
7495         pud_t *pud;
7496         pte_t *pte = NULL;
7497
7498         pgd = pgd_offset(mm, addr);
7499         p4d = p4d_alloc(mm, pgd, addr);
7500         if (!p4d)
7501                 return NULL;
7502         pud = pud_alloc(mm, p4d, addr);
7503         if (pud) {
7504                 if (sz == PUD_SIZE) {
7505                         pte = (pte_t *)pud;
7506                 } else {
7507                         BUG_ON(sz != PMD_SIZE);
7508                         if (want_pmd_share(vma, addr) && pud_none(*pud))
7509                                 pte = huge_pmd_share(mm, vma, addr, pud);
7510                         else
7511                                 pte = (pte_t *)pmd_alloc(mm, pud, addr);
7512                 }
7513         }
7514
7515         if (pte) {
7516                 pte_t pteval = ptep_get_lockless(pte);
7517
7518                 BUG_ON(pte_present(pteval) && !pte_huge(pteval));
7519         }
7520
7521         return pte;
7522 }
7523
7524 /*
7525  * huge_pte_offset() - Walk the page table to resolve the hugepage
7526  * entry at address @addr
7527  *
7528  * Return: Pointer to page table entry (PUD or PMD) for
7529  * address @addr, or NULL if a !p*d_present() entry is encountered and the
7530  * size @sz doesn't match the hugepage size at this level of the page
7531  * table.
7532  */
7533 pte_t *huge_pte_offset(struct mm_struct *mm,
7534                        unsigned long addr, unsigned long sz)
7535 {
7536         pgd_t *pgd;
7537         p4d_t *p4d;
7538         pud_t *pud;
7539         pmd_t *pmd;
7540
7541         pgd = pgd_offset(mm, addr);
7542         if (!pgd_present(*pgd))
7543                 return NULL;
7544         p4d = p4d_offset(pgd, addr);
7545         if (!p4d_present(*p4d))
7546                 return NULL;
7547
7548         pud = pud_offset(p4d, addr);
7549         if (sz == PUD_SIZE)
7550                 /* must be pud huge, non-present or none */
7551                 return (pte_t *)pud;
7552         if (!pud_present(*pud))
7553                 return NULL;
7554         /* must have a valid entry and size to go further */
7555
7556         pmd = pmd_offset(pud, addr);
7557         /* must be pmd huge, non-present or none */
7558         return (pte_t *)pmd;
7559 }
7560
7561 /*
7562  * Return a mask that can be used to update an address to the last huge
7563  * page in a page table page mapping size.  Used to skip non-present
7564  * page table entries when linearly scanning address ranges.  Architectures
7565  * with unique huge page to page table relationships can define their own
7566  * version of this routine.
7567  */
7568 unsigned long hugetlb_mask_last_page(struct hstate *h)
7569 {
7570         unsigned long hp_size = huge_page_size(h);
7571
7572         if (hp_size == PUD_SIZE)
7573                 return P4D_SIZE - PUD_SIZE;
7574         else if (hp_size == PMD_SIZE)
7575                 return PUD_SIZE - PMD_SIZE;
7576         else
7577                 return 0UL;
7578 }
7579
7580 #else
7581
7582 /* See description above.  Architectures can provide their own version. */
7583 __weak unsigned long hugetlb_mask_last_page(struct hstate *h)
7584 {
7585 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
7586         if (huge_page_size(h) == PMD_SIZE)
7587                 return PUD_SIZE - PMD_SIZE;
7588 #endif
7589         return 0UL;
7590 }
7591
7592 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
7593
7594 /*
7595  * These functions are overwritable if your architecture needs its own
7596  * behavior.
7597  */
7598 bool isolate_hugetlb(struct folio *folio, struct list_head *list)
7599 {
7600         bool ret = true;
7601
7602         spin_lock_irq(&hugetlb_lock);
7603         if (!folio_test_hugetlb(folio) ||
7604             !folio_test_hugetlb_migratable(folio) ||
7605             !folio_try_get(folio)) {
7606                 ret = false;
7607                 goto unlock;
7608         }
7609         folio_clear_hugetlb_migratable(folio);
7610         list_move_tail(&folio->lru, list);
7611 unlock:
7612         spin_unlock_irq(&hugetlb_lock);
7613         return ret;
7614 }
7615
7616 int get_hwpoison_hugetlb_folio(struct folio *folio, bool *hugetlb, bool unpoison)
7617 {
7618         int ret = 0;
7619
7620         *hugetlb = false;
7621         spin_lock_irq(&hugetlb_lock);
7622         if (folio_test_hugetlb(folio)) {
7623                 *hugetlb = true;
7624                 if (folio_test_hugetlb_freed(folio))
7625                         ret = 0;
7626                 else if (folio_test_hugetlb_migratable(folio) || unpoison)
7627                         ret = folio_try_get(folio);
7628                 else
7629                         ret = -EBUSY;
7630         }
7631         spin_unlock_irq(&hugetlb_lock);
7632         return ret;
7633 }
7634
7635 int get_huge_page_for_hwpoison(unsigned long pfn, int flags,
7636                                 bool *migratable_cleared)
7637 {
7638         int ret;
7639
7640         spin_lock_irq(&hugetlb_lock);
7641         ret = __get_huge_page_for_hwpoison(pfn, flags, migratable_cleared);
7642         spin_unlock_irq(&hugetlb_lock);
7643         return ret;
7644 }
7645
7646 void folio_putback_active_hugetlb(struct folio *folio)
7647 {
7648         spin_lock_irq(&hugetlb_lock);
7649         folio_set_hugetlb_migratable(folio);
7650         list_move_tail(&folio->lru, &(folio_hstate(folio))->hugepage_activelist);
7651         spin_unlock_irq(&hugetlb_lock);
7652         folio_put(folio);
7653 }
7654
7655 void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int reason)
7656 {
7657         struct hstate *h = folio_hstate(old_folio);
7658
7659         hugetlb_cgroup_migrate(old_folio, new_folio);
7660         set_page_owner_migrate_reason(&new_folio->page, reason);
7661
7662         /*
7663          * transfer temporary state of the new hugetlb folio. This is
7664          * reverse to other transitions because the newpage is going to
7665          * be final while the old one will be freed so it takes over
7666          * the temporary status.
7667          *
7668          * Also note that we have to transfer the per-node surplus state
7669          * here as well otherwise the global surplus count will not match
7670          * the per-node's.
7671          */
7672         if (folio_test_hugetlb_temporary(new_folio)) {
7673                 int old_nid = folio_nid(old_folio);
7674                 int new_nid = folio_nid(new_folio);
7675
7676                 folio_set_hugetlb_temporary(old_folio);
7677                 folio_clear_hugetlb_temporary(new_folio);
7678
7679
7680                 /*
7681                  * There is no need to transfer the per-node surplus state
7682                  * when we do not cross the node.
7683                  */
7684                 if (new_nid == old_nid)
7685                         return;
7686                 spin_lock_irq(&hugetlb_lock);
7687                 if (h->surplus_huge_pages_node[old_nid]) {
7688                         h->surplus_huge_pages_node[old_nid]--;
7689                         h->surplus_huge_pages_node[new_nid]++;
7690                 }
7691                 spin_unlock_irq(&hugetlb_lock);
7692         }
7693 }
7694
7695 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
7696                                    unsigned long start,
7697                                    unsigned long end)
7698 {
7699         struct hstate *h = hstate_vma(vma);
7700         unsigned long sz = huge_page_size(h);
7701         struct mm_struct *mm = vma->vm_mm;
7702         struct mmu_notifier_range range;
7703         unsigned long address;
7704         spinlock_t *ptl;
7705         pte_t *ptep;
7706
7707         if (!(vma->vm_flags & VM_MAYSHARE))
7708                 return;
7709
7710         if (start >= end)
7711                 return;
7712
7713         flush_cache_range(vma, start, end);
7714         /*
7715          * No need to call adjust_range_if_pmd_sharing_possible(), because
7716          * we have already done the PUD_SIZE alignment.
7717          */
7718         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
7719                                 start, end);
7720         mmu_notifier_invalidate_range_start(&range);
7721         hugetlb_vma_lock_write(vma);
7722         i_mmap_lock_write(vma->vm_file->f_mapping);
7723         for (address = start; address < end; address += PUD_SIZE) {
7724                 ptep = hugetlb_walk(vma, address, sz);
7725                 if (!ptep)
7726                         continue;
7727                 ptl = huge_pte_lock(h, mm, ptep);
7728                 huge_pmd_unshare(mm, vma, address, ptep);
7729                 spin_unlock(ptl);
7730         }
7731         flush_hugetlb_tlb_range(vma, start, end);
7732         i_mmap_unlock_write(vma->vm_file->f_mapping);
7733         hugetlb_vma_unlock_write(vma);
7734         /*
7735          * No need to call mmu_notifier_arch_invalidate_secondary_tlbs(), see
7736          * Documentation/mm/mmu_notifier.rst.
7737          */
7738         mmu_notifier_invalidate_range_end(&range);
7739 }
7740
7741 /*
7742  * This function will unconditionally remove all the shared pmd pgtable entries
7743  * within the specific vma for a hugetlbfs memory range.
7744  */
7745 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
7746 {
7747         hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE),
7748                         ALIGN_DOWN(vma->vm_end, PUD_SIZE));
7749 }
7750
7751 #ifdef CONFIG_CMA
7752 static bool cma_reserve_called __initdata;
7753
7754 static int __init cmdline_parse_hugetlb_cma(char *p)
7755 {
7756         int nid, count = 0;
7757         unsigned long tmp;
7758         char *s = p;
7759
7760         while (*s) {
7761                 if (sscanf(s, "%lu%n", &tmp, &count) != 1)
7762                         break;
7763
7764                 if (s[count] == ':') {
7765                         if (tmp >= MAX_NUMNODES)
7766                                 break;
7767                         nid = array_index_nospec(tmp, MAX_NUMNODES);
7768
7769                         s += count + 1;
7770                         tmp = memparse(s, &s);
7771                         hugetlb_cma_size_in_node[nid] = tmp;
7772                         hugetlb_cma_size += tmp;
7773
7774                         /*
7775                          * Skip the separator if have one, otherwise
7776                          * break the parsing.
7777                          */
7778                         if (*s == ',')
7779                                 s++;
7780                         else
7781                                 break;
7782                 } else {
7783                         hugetlb_cma_size = memparse(p, &p);
7784                         break;
7785                 }
7786         }
7787
7788         return 0;
7789 }
7790
7791 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
7792
7793 void __init hugetlb_cma_reserve(int order)
7794 {
7795         unsigned long size, reserved, per_node;
7796         bool node_specific_cma_alloc = false;
7797         int nid;
7798
7799         /*
7800          * HugeTLB CMA reservation is required for gigantic
7801          * huge pages which could not be allocated via the
7802          * page allocator. Just warn if there is any change
7803          * breaking this assumption.
7804          */
7805         VM_WARN_ON(order <= MAX_PAGE_ORDER);
7806         cma_reserve_called = true;
7807
7808         if (!hugetlb_cma_size)
7809                 return;
7810
7811         for (nid = 0; nid < MAX_NUMNODES; nid++) {
7812                 if (hugetlb_cma_size_in_node[nid] == 0)
7813                         continue;
7814
7815                 if (!node_online(nid)) {
7816                         pr_warn("hugetlb_cma: invalid node %d specified\n", nid);
7817                         hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7818                         hugetlb_cma_size_in_node[nid] = 0;
7819                         continue;
7820                 }
7821
7822                 if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) {
7823                         pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n",
7824                                 nid, (PAGE_SIZE << order) / SZ_1M);
7825                         hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7826                         hugetlb_cma_size_in_node[nid] = 0;
7827                 } else {
7828                         node_specific_cma_alloc = true;
7829                 }
7830         }
7831
7832         /* Validate the CMA size again in case some invalid nodes specified. */
7833         if (!hugetlb_cma_size)
7834                 return;
7835
7836         if (hugetlb_cma_size < (PAGE_SIZE << order)) {
7837                 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
7838                         (PAGE_SIZE << order) / SZ_1M);
7839                 hugetlb_cma_size = 0;
7840                 return;
7841         }
7842
7843         if (!node_specific_cma_alloc) {
7844                 /*
7845                  * If 3 GB area is requested on a machine with 4 numa nodes,
7846                  * let's allocate 1 GB on first three nodes and ignore the last one.
7847                  */
7848                 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
7849                 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
7850                         hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
7851         }
7852
7853         reserved = 0;
7854         for_each_online_node(nid) {
7855                 int res;
7856                 char name[CMA_MAX_NAME];
7857
7858                 if (node_specific_cma_alloc) {
7859                         if (hugetlb_cma_size_in_node[nid] == 0)
7860                                 continue;
7861
7862                         size = hugetlb_cma_size_in_node[nid];
7863                 } else {
7864                         size = min(per_node, hugetlb_cma_size - reserved);
7865                 }
7866
7867                 size = round_up(size, PAGE_SIZE << order);
7868
7869                 snprintf(name, sizeof(name), "hugetlb%d", nid);
7870                 /*
7871                  * Note that 'order per bit' is based on smallest size that
7872                  * may be returned to CMA allocator in the case of
7873                  * huge page demotion.
7874                  */
7875                 res = cma_declare_contiguous_nid(0, size, 0,
7876                                                 PAGE_SIZE << HUGETLB_PAGE_ORDER,
7877                                                  0, false, name,
7878                                                  &hugetlb_cma[nid], nid);
7879                 if (res) {
7880                         pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
7881                                 res, nid);
7882                         continue;
7883                 }
7884
7885                 reserved += size;
7886                 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
7887                         size / SZ_1M, nid);
7888
7889                 if (reserved >= hugetlb_cma_size)
7890                         break;
7891         }
7892
7893         if (!reserved)
7894                 /*
7895                  * hugetlb_cma_size is used to determine if allocations from
7896                  * cma are possible.  Set to zero if no cma regions are set up.
7897                  */
7898                 hugetlb_cma_size = 0;
7899 }
7900
7901 static void __init hugetlb_cma_check(void)
7902 {
7903         if (!hugetlb_cma_size || cma_reserve_called)
7904                 return;
7905
7906         pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
7907 }
7908
7909 #endif /* CONFIG_CMA */
This page took 0.502041 seconds and 4 git commands to generate.