1 // SPDX-License-Identifier: GPL-2.0
3 * Page table allocation functions
5 * Copyright IBM Corp. 2016
9 #include <linux/sysctl.h>
10 #include <linux/slab.h>
12 #include <asm/mmu_context.h>
13 #include <asm/pgalloc.h>
16 #include <asm/tlbflush.h>
20 static int page_table_allocate_pgste_min = 0;
21 static int page_table_allocate_pgste_max = 1;
22 int page_table_allocate_pgste = 0;
23 EXPORT_SYMBOL(page_table_allocate_pgste);
25 static struct ctl_table page_table_sysctl[] = {
27 .procname = "allocate_pgste",
28 .data = &page_table_allocate_pgste,
29 .maxlen = sizeof(int),
30 .mode = S_IRUGO | S_IWUSR,
31 .proc_handler = proc_dointvec_minmax,
32 .extra1 = &page_table_allocate_pgste_min,
33 .extra2 = &page_table_allocate_pgste_max,
38 static struct ctl_table page_table_sysctl_dir[] = {
43 .child = page_table_sysctl,
48 static int __init page_table_register_sysctl(void)
50 return register_sysctl_table(page_table_sysctl_dir) ? 0 : -ENOMEM;
52 __initcall(page_table_register_sysctl);
54 #endif /* CONFIG_PGSTE */
56 unsigned long *crst_table_alloc(struct mm_struct *mm)
58 struct page *page = alloc_pages(GFP_KERNEL, 2);
62 arch_set_page_dat(page, 2);
63 return (unsigned long *) page_to_phys(page);
66 void crst_table_free(struct mm_struct *mm, unsigned long *table)
68 free_pages((unsigned long) table, 2);
71 static void __crst_table_upgrade(void *arg)
73 struct mm_struct *mm = arg;
75 if (current->active_mm == mm)
80 int crst_table_upgrade(struct mm_struct *mm, unsigned long end)
82 unsigned long *table, *pgd;
85 /* upgrade should only happen from 3 to 4, 3 to 5, or 4 to 5 levels */
86 VM_BUG_ON(mm->context.asce_limit < _REGION2_SIZE);
89 while (mm->context.asce_limit < end) {
90 table = crst_table_alloc(mm);
95 spin_lock_bh(&mm->page_table_lock);
96 pgd = (unsigned long *) mm->pgd;
97 if (mm->context.asce_limit == _REGION2_SIZE) {
98 crst_table_init(table, _REGION2_ENTRY_EMPTY);
99 p4d_populate(mm, (p4d_t *) table, (pud_t *) pgd);
100 mm->pgd = (pgd_t *) table;
101 mm->context.asce_limit = _REGION1_SIZE;
102 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
103 _ASCE_USER_BITS | _ASCE_TYPE_REGION2;
106 crst_table_init(table, _REGION1_ENTRY_EMPTY);
107 pgd_populate(mm, (pgd_t *) table, (p4d_t *) pgd);
108 mm->pgd = (pgd_t *) table;
109 mm->context.asce_limit = -PAGE_SIZE;
110 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
111 _ASCE_USER_BITS | _ASCE_TYPE_REGION1;
114 spin_unlock_bh(&mm->page_table_lock);
117 on_each_cpu(__crst_table_upgrade, mm, 0);
121 void crst_table_downgrade(struct mm_struct *mm)
125 /* downgrade should only happen from 3 to 2 levels (compat only) */
126 VM_BUG_ON(mm->context.asce_limit != _REGION2_SIZE);
128 if (current->active_mm == mm) {
135 mm->pgd = (pgd_t *) (pgd_val(*pgd) & _REGION_ENTRY_ORIGIN);
136 mm->context.asce_limit = _REGION3_SIZE;
137 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
138 _ASCE_USER_BITS | _ASCE_TYPE_SEGMENT;
139 crst_table_free(mm, (unsigned long *) pgd);
141 if (current->active_mm == mm)
145 static inline unsigned int atomic_xor_bits(atomic_t *v, unsigned int bits)
147 unsigned int old, new;
150 old = atomic_read(v);
152 } while (atomic_cmpxchg(v, old, new) != old);
158 struct page *page_table_alloc_pgste(struct mm_struct *mm)
163 page = alloc_page(GFP_KERNEL);
165 table = (u64 *)page_to_phys(page);
166 memset64(table, _PAGE_INVALID, PTRS_PER_PTE);
167 memset64(table + PTRS_PER_PTE, 0, PTRS_PER_PTE);
172 void page_table_free_pgste(struct page *page)
177 #endif /* CONFIG_PGSTE */
180 * page table entry allocation/free routines.
182 unsigned long *page_table_alloc(struct mm_struct *mm)
184 unsigned long *table;
186 unsigned int mask, bit;
188 /* Try to get a fragment of a 4K page as a 2K page table */
189 if (!mm_alloc_pgste(mm)) {
191 spin_lock_bh(&mm->context.lock);
192 if (!list_empty(&mm->context.pgtable_list)) {
193 page = list_first_entry(&mm->context.pgtable_list,
195 mask = atomic_read(&page->_refcount) >> 24;
196 mask = (mask | (mask >> 4)) & 3;
198 table = (unsigned long *) page_to_phys(page);
199 bit = mask & 1; /* =1 -> second 2K */
201 table += PTRS_PER_PTE;
202 atomic_xor_bits(&page->_refcount,
204 list_del(&page->lru);
207 spin_unlock_bh(&mm->context.lock);
211 /* Allocate a fresh page */
212 page = alloc_page(GFP_KERNEL);
215 if (!pgtable_page_ctor(page)) {
219 arch_set_page_dat(page, 0);
220 /* Initialize page table */
221 table = (unsigned long *) page_to_phys(page);
222 if (mm_alloc_pgste(mm)) {
223 /* Return 4K page table with PGSTEs */
224 atomic_xor_bits(&page->_refcount, 3 << 24);
225 memset64((u64 *)table, _PAGE_INVALID, PTRS_PER_PTE);
226 memset64((u64 *)table + PTRS_PER_PTE, 0, PTRS_PER_PTE);
228 /* Return the first 2K fragment of the page */
229 atomic_xor_bits(&page->_refcount, 1 << 24);
230 memset64((u64 *)table, _PAGE_INVALID, 2 * PTRS_PER_PTE);
231 spin_lock_bh(&mm->context.lock);
232 list_add(&page->lru, &mm->context.pgtable_list);
233 spin_unlock_bh(&mm->context.lock);
238 void page_table_free(struct mm_struct *mm, unsigned long *table)
241 unsigned int bit, mask;
243 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
244 if (!mm_alloc_pgste(mm)) {
245 /* Free 2K page table fragment of a 4K page */
246 bit = (__pa(table) & ~PAGE_MASK)/(PTRS_PER_PTE*sizeof(pte_t));
247 spin_lock_bh(&mm->context.lock);
248 mask = atomic_xor_bits(&page->_refcount, 1U << (bit + 24));
251 list_add(&page->lru, &mm->context.pgtable_list);
253 list_del(&page->lru);
254 spin_unlock_bh(&mm->context.lock);
258 atomic_xor_bits(&page->_refcount, 3U << 24);
261 pgtable_page_dtor(page);
265 void page_table_free_rcu(struct mmu_gather *tlb, unsigned long *table,
266 unsigned long vmaddr)
268 struct mm_struct *mm;
270 unsigned int bit, mask;
273 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
274 if (mm_alloc_pgste(mm)) {
275 gmap_unlink(mm, table, vmaddr);
276 table = (unsigned long *) (__pa(table) | 3);
277 tlb_remove_table(tlb, table);
280 bit = (__pa(table) & ~PAGE_MASK) / (PTRS_PER_PTE*sizeof(pte_t));
281 spin_lock_bh(&mm->context.lock);
282 mask = atomic_xor_bits(&page->_refcount, 0x11U << (bit + 24));
285 list_add_tail(&page->lru, &mm->context.pgtable_list);
287 list_del(&page->lru);
288 spin_unlock_bh(&mm->context.lock);
289 table = (unsigned long *) (__pa(table) | (1U << bit));
290 tlb_remove_table(tlb, table);
293 static void __tlb_remove_table(void *_table)
295 unsigned int mask = (unsigned long) _table & 3;
296 void *table = (void *)((unsigned long) _table ^ mask);
297 struct page *page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
300 case 0: /* pmd, pud, or p4d */
301 free_pages((unsigned long) table, 2);
303 case 1: /* lower 2K of a 4K page table */
304 case 2: /* higher 2K of a 4K page table */
305 mask = atomic_xor_bits(&page->_refcount, mask << (4 + 24));
310 case 3: /* 4K page table with pgstes */
312 atomic_xor_bits(&page->_refcount, 3 << 24);
313 pgtable_page_dtor(page);
319 static void tlb_remove_table_smp_sync(void *arg)
321 /* Simply deliver the interrupt */
324 static void tlb_remove_table_one(void *table)
327 * This isn't an RCU grace period and hence the page-tables cannot be
328 * assumed to be actually RCU-freed.
330 * It is however sufficient for software page-table walkers that rely
331 * on IRQ disabling. See the comment near struct mmu_table_batch.
333 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
334 __tlb_remove_table(table);
337 static void tlb_remove_table_rcu(struct rcu_head *head)
339 struct mmu_table_batch *batch;
342 batch = container_of(head, struct mmu_table_batch, rcu);
344 for (i = 0; i < batch->nr; i++)
345 __tlb_remove_table(batch->tables[i]);
347 free_page((unsigned long)batch);
350 void tlb_table_flush(struct mmu_gather *tlb)
352 struct mmu_table_batch **batch = &tlb->batch;
355 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
360 void tlb_remove_table(struct mmu_gather *tlb, void *table)
362 struct mmu_table_batch **batch = &tlb->batch;
364 tlb->mm->context.flush_mm = 1;
365 if (*batch == NULL) {
366 *batch = (struct mmu_table_batch *)
367 __get_free_page(GFP_NOWAIT | __GFP_NOWARN);
368 if (*batch == NULL) {
369 __tlb_flush_mm_lazy(tlb->mm);
370 tlb_remove_table_one(table);
375 (*batch)->tables[(*batch)->nr++] = table;
376 if ((*batch)->nr == MAX_TABLE_BATCH)
381 * Base infrastructure required to generate basic asces, region, segment,
382 * and page tables that do not make use of enhanced features like EDAT1.
385 static struct kmem_cache *base_pgt_cache;
387 static unsigned long base_pgt_alloc(void)
391 table = kmem_cache_alloc(base_pgt_cache, GFP_KERNEL);
393 memset64(table, _PAGE_INVALID, PTRS_PER_PTE);
394 return (unsigned long) table;
397 static void base_pgt_free(unsigned long table)
399 kmem_cache_free(base_pgt_cache, (void *) table);
402 static unsigned long base_crst_alloc(unsigned long val)
406 table = __get_free_pages(GFP_KERNEL, CRST_ALLOC_ORDER);
408 crst_table_init((unsigned long *)table, val);
412 static void base_crst_free(unsigned long table)
414 free_pages(table, CRST_ALLOC_ORDER);
417 #define BASE_ADDR_END_FUNC(NAME, SIZE) \
418 static inline unsigned long base_##NAME##_addr_end(unsigned long addr, \
421 unsigned long next = (addr + (SIZE)) & ~((SIZE) - 1); \
423 return (next - 1) < (end - 1) ? next : end; \
426 BASE_ADDR_END_FUNC(page, _PAGE_SIZE)
427 BASE_ADDR_END_FUNC(segment, _SEGMENT_SIZE)
428 BASE_ADDR_END_FUNC(region3, _REGION3_SIZE)
429 BASE_ADDR_END_FUNC(region2, _REGION2_SIZE)
430 BASE_ADDR_END_FUNC(region1, _REGION1_SIZE)
432 static inline unsigned long base_lra(unsigned long address)
438 : "=d" (real) : "a" (address) : "cc");
442 static int base_page_walk(unsigned long origin, unsigned long addr,
443 unsigned long end, int alloc)
445 unsigned long *pte, next;
449 pte = (unsigned long *) origin;
450 pte += (addr & _PAGE_INDEX) >> _PAGE_SHIFT;
452 next = base_page_addr_end(addr, end);
453 *pte = base_lra(addr);
454 } while (pte++, addr = next, addr < end);
458 static int base_segment_walk(unsigned long origin, unsigned long addr,
459 unsigned long end, int alloc)
461 unsigned long *ste, next, table;
464 ste = (unsigned long *) origin;
465 ste += (addr & _SEGMENT_INDEX) >> _SEGMENT_SHIFT;
467 next = base_segment_addr_end(addr, end);
468 if (*ste & _SEGMENT_ENTRY_INVALID) {
471 table = base_pgt_alloc();
474 *ste = table | _SEGMENT_ENTRY;
476 table = *ste & _SEGMENT_ENTRY_ORIGIN;
477 rc = base_page_walk(table, addr, next, alloc);
481 base_pgt_free(table);
483 } while (ste++, addr = next, addr < end);
487 static int base_region3_walk(unsigned long origin, unsigned long addr,
488 unsigned long end, int alloc)
490 unsigned long *rtte, next, table;
493 rtte = (unsigned long *) origin;
494 rtte += (addr & _REGION3_INDEX) >> _REGION3_SHIFT;
496 next = base_region3_addr_end(addr, end);
497 if (*rtte & _REGION_ENTRY_INVALID) {
500 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY);
503 *rtte = table | _REGION3_ENTRY;
505 table = *rtte & _REGION_ENTRY_ORIGIN;
506 rc = base_segment_walk(table, addr, next, alloc);
510 base_crst_free(table);
511 } while (rtte++, addr = next, addr < end);
515 static int base_region2_walk(unsigned long origin, unsigned long addr,
516 unsigned long end, int alloc)
518 unsigned long *rste, next, table;
521 rste = (unsigned long *) origin;
522 rste += (addr & _REGION2_INDEX) >> _REGION2_SHIFT;
524 next = base_region2_addr_end(addr, end);
525 if (*rste & _REGION_ENTRY_INVALID) {
528 table = base_crst_alloc(_REGION3_ENTRY_EMPTY);
531 *rste = table | _REGION2_ENTRY;
533 table = *rste & _REGION_ENTRY_ORIGIN;
534 rc = base_region3_walk(table, addr, next, alloc);
538 base_crst_free(table);
539 } while (rste++, addr = next, addr < end);
543 static int base_region1_walk(unsigned long origin, unsigned long addr,
544 unsigned long end, int alloc)
546 unsigned long *rfte, next, table;
549 rfte = (unsigned long *) origin;
550 rfte += (addr & _REGION1_INDEX) >> _REGION1_SHIFT;
552 next = base_region1_addr_end(addr, end);
553 if (*rfte & _REGION_ENTRY_INVALID) {
556 table = base_crst_alloc(_REGION2_ENTRY_EMPTY);
559 *rfte = table | _REGION1_ENTRY;
561 table = *rfte & _REGION_ENTRY_ORIGIN;
562 rc = base_region2_walk(table, addr, next, alloc);
566 base_crst_free(table);
567 } while (rfte++, addr = next, addr < end);
572 * base_asce_free - free asce and tables returned from base_asce_alloc()
573 * @asce: asce to be freed
575 * Frees all region, segment, and page tables that were allocated with a
576 * corresponding base_asce_alloc() call.
578 void base_asce_free(unsigned long asce)
580 unsigned long table = asce & _ASCE_ORIGIN;
584 switch (asce & _ASCE_TYPE_MASK) {
585 case _ASCE_TYPE_SEGMENT:
586 base_segment_walk(table, 0, _REGION3_SIZE, 0);
588 case _ASCE_TYPE_REGION3:
589 base_region3_walk(table, 0, _REGION2_SIZE, 0);
591 case _ASCE_TYPE_REGION2:
592 base_region2_walk(table, 0, _REGION1_SIZE, 0);
594 case _ASCE_TYPE_REGION1:
595 base_region1_walk(table, 0, -_PAGE_SIZE, 0);
598 base_crst_free(table);
601 static int base_pgt_cache_init(void)
603 static DEFINE_MUTEX(base_pgt_cache_mutex);
604 unsigned long sz = _PAGE_TABLE_SIZE;
608 mutex_lock(&base_pgt_cache_mutex);
610 base_pgt_cache = kmem_cache_create("base_pgt", sz, sz, 0, NULL);
611 mutex_unlock(&base_pgt_cache_mutex);
612 return base_pgt_cache ? 0 : -ENOMEM;
616 * base_asce_alloc - create kernel mapping without enhanced DAT features
617 * @addr: virtual start address of kernel mapping
618 * @num_pages: number of consecutive pages
620 * Generate an asce, including all required region, segment and page tables,
621 * that can be used to access the virtual kernel mapping. The difference is
622 * that the returned asce does not make use of any enhanced DAT features like
623 * e.g. large pages. This is required for some I/O functions that pass an
624 * asce, like e.g. some service call requests.
626 * Note: the returned asce may NEVER be attached to any cpu. It may only be
627 * used for I/O requests. tlb entries that might result because the
628 * asce was attached to a cpu won't be cleared.
630 unsigned long base_asce_alloc(unsigned long addr, unsigned long num_pages)
632 unsigned long asce, table, end;
635 if (base_pgt_cache_init())
637 end = addr + num_pages * PAGE_SIZE;
638 if (end <= _REGION3_SIZE) {
639 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY);
642 rc = base_segment_walk(table, addr, end, 1);
643 asce = table | _ASCE_TYPE_SEGMENT | _ASCE_TABLE_LENGTH;
644 } else if (end <= _REGION2_SIZE) {
645 table = base_crst_alloc(_REGION3_ENTRY_EMPTY);
648 rc = base_region3_walk(table, addr, end, 1);
649 asce = table | _ASCE_TYPE_REGION3 | _ASCE_TABLE_LENGTH;
650 } else if (end <= _REGION1_SIZE) {
651 table = base_crst_alloc(_REGION2_ENTRY_EMPTY);
654 rc = base_region2_walk(table, addr, end, 1);
655 asce = table | _ASCE_TYPE_REGION2 | _ASCE_TABLE_LENGTH;
657 table = base_crst_alloc(_REGION1_ENTRY_EMPTY);
660 rc = base_region1_walk(table, addr, end, 1);
661 asce = table | _ASCE_TYPE_REGION1 | _ASCE_TABLE_LENGTH;
664 base_asce_free(asce);