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 int page_table_allocate_pgste = 0;
21 EXPORT_SYMBOL(page_table_allocate_pgste);
23 static struct ctl_table page_table_sysctl[] = {
25 .procname = "allocate_pgste",
26 .data = &page_table_allocate_pgste,
27 .maxlen = sizeof(int),
28 .mode = S_IRUGO | S_IWUSR,
29 .proc_handler = proc_dointvec_minmax,
30 .extra1 = SYSCTL_ZERO,
36 static struct ctl_table page_table_sysctl_dir[] = {
41 .child = page_table_sysctl,
46 static int __init page_table_register_sysctl(void)
48 return register_sysctl_table(page_table_sysctl_dir) ? 0 : -ENOMEM;
50 __initcall(page_table_register_sysctl);
52 #endif /* CONFIG_PGSTE */
54 unsigned long *crst_table_alloc(struct mm_struct *mm)
56 struct page *page = alloc_pages(GFP_KERNEL, 2);
60 arch_set_page_dat(page, 2);
61 return (unsigned long *) page_to_phys(page);
64 void crst_table_free(struct mm_struct *mm, unsigned long *table)
66 free_pages((unsigned long) table, 2);
69 static void __crst_table_upgrade(void *arg)
71 struct mm_struct *mm = arg;
73 /* we must change all active ASCEs to avoid the creation of new TLBs */
74 if (current->active_mm == mm) {
75 S390_lowcore.user_asce = mm->context.asce;
76 if (current->thread.mm_segment == USER_DS) {
77 __ctl_load(S390_lowcore.user_asce, 1, 1);
78 /* Mark user-ASCE present in CR1 */
79 clear_cpu_flag(CIF_ASCE_PRIMARY);
81 if (current->thread.mm_segment == USER_DS_SACF) {
82 __ctl_load(S390_lowcore.user_asce, 7, 7);
83 /* enable_sacf_uaccess does all or nothing */
84 WARN_ON(!test_cpu_flag(CIF_ASCE_SECONDARY));
90 int crst_table_upgrade(struct mm_struct *mm, unsigned long end)
92 unsigned long *pgd = NULL, *p4d = NULL, *__pgd;
93 unsigned long asce_limit = mm->context.asce_limit;
95 /* upgrade should only happen from 3 to 4, 3 to 5, or 4 to 5 levels */
96 VM_BUG_ON(asce_limit < _REGION2_SIZE);
98 if (end <= asce_limit)
101 if (asce_limit == _REGION2_SIZE) {
102 p4d = crst_table_alloc(mm);
105 crst_table_init(p4d, _REGION2_ENTRY_EMPTY);
107 if (end > _REGION1_SIZE) {
108 pgd = crst_table_alloc(mm);
111 crst_table_init(pgd, _REGION1_ENTRY_EMPTY);
114 spin_lock_bh(&mm->page_table_lock);
117 * This routine gets called with mmap_lock lock held and there is
118 * no reason to optimize for the case of otherwise. However, if
119 * that would ever change, the below check will let us know.
121 VM_BUG_ON(asce_limit != mm->context.asce_limit);
124 __pgd = (unsigned long *) mm->pgd;
125 p4d_populate(mm, (p4d_t *) p4d, (pud_t *) __pgd);
126 mm->pgd = (pgd_t *) p4d;
127 mm->context.asce_limit = _REGION1_SIZE;
128 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
129 _ASCE_USER_BITS | _ASCE_TYPE_REGION2;
133 __pgd = (unsigned long *) mm->pgd;
134 pgd_populate(mm, (pgd_t *) pgd, (p4d_t *) __pgd);
135 mm->pgd = (pgd_t *) pgd;
136 mm->context.asce_limit = TASK_SIZE_MAX;
137 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
138 _ASCE_USER_BITS | _ASCE_TYPE_REGION1;
141 spin_unlock_bh(&mm->page_table_lock);
143 on_each_cpu(__crst_table_upgrade, mm, 0);
148 crst_table_free(mm, p4d);
153 static inline unsigned int atomic_xor_bits(atomic_t *v, unsigned int bits)
155 unsigned int old, new;
158 old = atomic_read(v);
160 } while (atomic_cmpxchg(v, old, new) != old);
166 struct page *page_table_alloc_pgste(struct mm_struct *mm)
171 page = alloc_page(GFP_KERNEL);
173 table = (u64 *)page_to_phys(page);
174 memset64(table, _PAGE_INVALID, PTRS_PER_PTE);
175 memset64(table + PTRS_PER_PTE, 0, PTRS_PER_PTE);
180 void page_table_free_pgste(struct page *page)
185 #endif /* CONFIG_PGSTE */
188 * page table entry allocation/free routines.
190 unsigned long *page_table_alloc(struct mm_struct *mm)
192 unsigned long *table;
194 unsigned int mask, bit;
196 /* Try to get a fragment of a 4K page as a 2K page table */
197 if (!mm_alloc_pgste(mm)) {
199 spin_lock_bh(&mm->context.lock);
200 if (!list_empty(&mm->context.pgtable_list)) {
201 page = list_first_entry(&mm->context.pgtable_list,
203 mask = atomic_read(&page->_refcount) >> 24;
204 mask = (mask | (mask >> 4)) & 3;
206 table = (unsigned long *) page_to_phys(page);
207 bit = mask & 1; /* =1 -> second 2K */
209 table += PTRS_PER_PTE;
210 atomic_xor_bits(&page->_refcount,
212 list_del(&page->lru);
215 spin_unlock_bh(&mm->context.lock);
219 /* Allocate a fresh page */
220 page = alloc_page(GFP_KERNEL);
223 if (!pgtable_pte_page_ctor(page)) {
227 arch_set_page_dat(page, 0);
228 /* Initialize page table */
229 table = (unsigned long *) page_to_phys(page);
230 if (mm_alloc_pgste(mm)) {
231 /* Return 4K page table with PGSTEs */
232 atomic_xor_bits(&page->_refcount, 3 << 24);
233 memset64((u64 *)table, _PAGE_INVALID, PTRS_PER_PTE);
234 memset64((u64 *)table + PTRS_PER_PTE, 0, PTRS_PER_PTE);
236 /* Return the first 2K fragment of the page */
237 atomic_xor_bits(&page->_refcount, 1 << 24);
238 memset64((u64 *)table, _PAGE_INVALID, 2 * PTRS_PER_PTE);
239 spin_lock_bh(&mm->context.lock);
240 list_add(&page->lru, &mm->context.pgtable_list);
241 spin_unlock_bh(&mm->context.lock);
246 void page_table_free(struct mm_struct *mm, unsigned long *table)
249 unsigned int bit, mask;
251 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
252 if (!mm_alloc_pgste(mm)) {
253 /* Free 2K page table fragment of a 4K page */
254 bit = (__pa(table) & ~PAGE_MASK)/(PTRS_PER_PTE*sizeof(pte_t));
255 spin_lock_bh(&mm->context.lock);
256 mask = atomic_xor_bits(&page->_refcount, 1U << (bit + 24));
259 list_add(&page->lru, &mm->context.pgtable_list);
261 list_del(&page->lru);
262 spin_unlock_bh(&mm->context.lock);
266 atomic_xor_bits(&page->_refcount, 3U << 24);
269 pgtable_pte_page_dtor(page);
273 void page_table_free_rcu(struct mmu_gather *tlb, unsigned long *table,
274 unsigned long vmaddr)
276 struct mm_struct *mm;
278 unsigned int bit, mask;
281 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
282 if (mm_alloc_pgste(mm)) {
283 gmap_unlink(mm, table, vmaddr);
284 table = (unsigned long *) (__pa(table) | 3);
285 tlb_remove_table(tlb, table);
288 bit = (__pa(table) & ~PAGE_MASK) / (PTRS_PER_PTE*sizeof(pte_t));
289 spin_lock_bh(&mm->context.lock);
290 mask = atomic_xor_bits(&page->_refcount, 0x11U << (bit + 24));
293 list_add_tail(&page->lru, &mm->context.pgtable_list);
295 list_del(&page->lru);
296 spin_unlock_bh(&mm->context.lock);
297 table = (unsigned long *) (__pa(table) | (1U << bit));
298 tlb_remove_table(tlb, table);
301 void __tlb_remove_table(void *_table)
303 unsigned int mask = (unsigned long) _table & 3;
304 void *table = (void *)((unsigned long) _table ^ mask);
305 struct page *page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
308 case 0: /* pmd, pud, or p4d */
309 free_pages((unsigned long) table, 2);
311 case 1: /* lower 2K of a 4K page table */
312 case 2: /* higher 2K of a 4K page table */
313 mask = atomic_xor_bits(&page->_refcount, mask << (4 + 24));
318 case 3: /* 4K page table with pgstes */
320 atomic_xor_bits(&page->_refcount, 3 << 24);
321 pgtable_pte_page_dtor(page);
328 * Base infrastructure required to generate basic asces, region, segment,
329 * and page tables that do not make use of enhanced features like EDAT1.
332 static struct kmem_cache *base_pgt_cache;
334 static unsigned long base_pgt_alloc(void)
338 table = kmem_cache_alloc(base_pgt_cache, GFP_KERNEL);
340 memset64(table, _PAGE_INVALID, PTRS_PER_PTE);
341 return (unsigned long) table;
344 static void base_pgt_free(unsigned long table)
346 kmem_cache_free(base_pgt_cache, (void *) table);
349 static unsigned long base_crst_alloc(unsigned long val)
353 table = __get_free_pages(GFP_KERNEL, CRST_ALLOC_ORDER);
355 crst_table_init((unsigned long *)table, val);
359 static void base_crst_free(unsigned long table)
361 free_pages(table, CRST_ALLOC_ORDER);
364 #define BASE_ADDR_END_FUNC(NAME, SIZE) \
365 static inline unsigned long base_##NAME##_addr_end(unsigned long addr, \
368 unsigned long next = (addr + (SIZE)) & ~((SIZE) - 1); \
370 return (next - 1) < (end - 1) ? next : end; \
373 BASE_ADDR_END_FUNC(page, _PAGE_SIZE)
374 BASE_ADDR_END_FUNC(segment, _SEGMENT_SIZE)
375 BASE_ADDR_END_FUNC(region3, _REGION3_SIZE)
376 BASE_ADDR_END_FUNC(region2, _REGION2_SIZE)
377 BASE_ADDR_END_FUNC(region1, _REGION1_SIZE)
379 static inline unsigned long base_lra(unsigned long address)
385 : "=d" (real) : "a" (address) : "cc");
389 static int base_page_walk(unsigned long origin, unsigned long addr,
390 unsigned long end, int alloc)
392 unsigned long *pte, next;
396 pte = (unsigned long *) origin;
397 pte += (addr & _PAGE_INDEX) >> _PAGE_SHIFT;
399 next = base_page_addr_end(addr, end);
400 *pte = base_lra(addr);
401 } while (pte++, addr = next, addr < end);
405 static int base_segment_walk(unsigned long origin, unsigned long addr,
406 unsigned long end, int alloc)
408 unsigned long *ste, next, table;
411 ste = (unsigned long *) origin;
412 ste += (addr & _SEGMENT_INDEX) >> _SEGMENT_SHIFT;
414 next = base_segment_addr_end(addr, end);
415 if (*ste & _SEGMENT_ENTRY_INVALID) {
418 table = base_pgt_alloc();
421 *ste = table | _SEGMENT_ENTRY;
423 table = *ste & _SEGMENT_ENTRY_ORIGIN;
424 rc = base_page_walk(table, addr, next, alloc);
428 base_pgt_free(table);
430 } while (ste++, addr = next, addr < end);
434 static int base_region3_walk(unsigned long origin, unsigned long addr,
435 unsigned long end, int alloc)
437 unsigned long *rtte, next, table;
440 rtte = (unsigned long *) origin;
441 rtte += (addr & _REGION3_INDEX) >> _REGION3_SHIFT;
443 next = base_region3_addr_end(addr, end);
444 if (*rtte & _REGION_ENTRY_INVALID) {
447 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY);
450 *rtte = table | _REGION3_ENTRY;
452 table = *rtte & _REGION_ENTRY_ORIGIN;
453 rc = base_segment_walk(table, addr, next, alloc);
457 base_crst_free(table);
458 } while (rtte++, addr = next, addr < end);
462 static int base_region2_walk(unsigned long origin, unsigned long addr,
463 unsigned long end, int alloc)
465 unsigned long *rste, next, table;
468 rste = (unsigned long *) origin;
469 rste += (addr & _REGION2_INDEX) >> _REGION2_SHIFT;
471 next = base_region2_addr_end(addr, end);
472 if (*rste & _REGION_ENTRY_INVALID) {
475 table = base_crst_alloc(_REGION3_ENTRY_EMPTY);
478 *rste = table | _REGION2_ENTRY;
480 table = *rste & _REGION_ENTRY_ORIGIN;
481 rc = base_region3_walk(table, addr, next, alloc);
485 base_crst_free(table);
486 } while (rste++, addr = next, addr < end);
490 static int base_region1_walk(unsigned long origin, unsigned long addr,
491 unsigned long end, int alloc)
493 unsigned long *rfte, next, table;
496 rfte = (unsigned long *) origin;
497 rfte += (addr & _REGION1_INDEX) >> _REGION1_SHIFT;
499 next = base_region1_addr_end(addr, end);
500 if (*rfte & _REGION_ENTRY_INVALID) {
503 table = base_crst_alloc(_REGION2_ENTRY_EMPTY);
506 *rfte = table | _REGION1_ENTRY;
508 table = *rfte & _REGION_ENTRY_ORIGIN;
509 rc = base_region2_walk(table, addr, next, alloc);
513 base_crst_free(table);
514 } while (rfte++, addr = next, addr < end);
519 * base_asce_free - free asce and tables returned from base_asce_alloc()
520 * @asce: asce to be freed
522 * Frees all region, segment, and page tables that were allocated with a
523 * corresponding base_asce_alloc() call.
525 void base_asce_free(unsigned long asce)
527 unsigned long table = asce & _ASCE_ORIGIN;
531 switch (asce & _ASCE_TYPE_MASK) {
532 case _ASCE_TYPE_SEGMENT:
533 base_segment_walk(table, 0, _REGION3_SIZE, 0);
535 case _ASCE_TYPE_REGION3:
536 base_region3_walk(table, 0, _REGION2_SIZE, 0);
538 case _ASCE_TYPE_REGION2:
539 base_region2_walk(table, 0, _REGION1_SIZE, 0);
541 case _ASCE_TYPE_REGION1:
542 base_region1_walk(table, 0, TASK_SIZE_MAX, 0);
545 base_crst_free(table);
548 static int base_pgt_cache_init(void)
550 static DEFINE_MUTEX(base_pgt_cache_mutex);
551 unsigned long sz = _PAGE_TABLE_SIZE;
555 mutex_lock(&base_pgt_cache_mutex);
557 base_pgt_cache = kmem_cache_create("base_pgt", sz, sz, 0, NULL);
558 mutex_unlock(&base_pgt_cache_mutex);
559 return base_pgt_cache ? 0 : -ENOMEM;
563 * base_asce_alloc - create kernel mapping without enhanced DAT features
564 * @addr: virtual start address of kernel mapping
565 * @num_pages: number of consecutive pages
567 * Generate an asce, including all required region, segment and page tables,
568 * that can be used to access the virtual kernel mapping. The difference is
569 * that the returned asce does not make use of any enhanced DAT features like
570 * e.g. large pages. This is required for some I/O functions that pass an
571 * asce, like e.g. some service call requests.
573 * Note: the returned asce may NEVER be attached to any cpu. It may only be
574 * used for I/O requests. tlb entries that might result because the
575 * asce was attached to a cpu won't be cleared.
577 unsigned long base_asce_alloc(unsigned long addr, unsigned long num_pages)
579 unsigned long asce, table, end;
582 if (base_pgt_cache_init())
584 end = addr + num_pages * PAGE_SIZE;
585 if (end <= _REGION3_SIZE) {
586 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY);
589 rc = base_segment_walk(table, addr, end, 1);
590 asce = table | _ASCE_TYPE_SEGMENT | _ASCE_TABLE_LENGTH;
591 } else if (end <= _REGION2_SIZE) {
592 table = base_crst_alloc(_REGION3_ENTRY_EMPTY);
595 rc = base_region3_walk(table, addr, end, 1);
596 asce = table | _ASCE_TYPE_REGION3 | _ASCE_TABLE_LENGTH;
597 } else if (end <= _REGION1_SIZE) {
598 table = base_crst_alloc(_REGION2_ENTRY_EMPTY);
601 rc = base_region2_walk(table, addr, end, 1);
602 asce = table | _ASCE_TYPE_REGION2 | _ASCE_TABLE_LENGTH;
604 table = base_crst_alloc(_REGION1_ENTRY_EMPTY);
607 rc = base_region1_walk(table, addr, end, 1);
608 asce = table | _ASCE_TYPE_REGION1 | _ASCE_TABLE_LENGTH;
611 base_asce_free(asce);