1 // SPDX-License-Identifier: GPL-2.0
3 * Page table allocation functions
5 * Copyright IBM Corp. 2016
6 * Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
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 *table, *pgd;
95 /* upgrade should only happen from 3 to 4, 3 to 5, or 4 to 5 levels */
96 VM_BUG_ON(mm->context.asce_limit < _REGION2_SIZE);
99 while (mm->context.asce_limit < end) {
100 table = crst_table_alloc(mm);
105 spin_lock_bh(&mm->page_table_lock);
106 pgd = (unsigned long *) mm->pgd;
107 if (mm->context.asce_limit == _REGION2_SIZE) {
108 crst_table_init(table, _REGION2_ENTRY_EMPTY);
109 p4d_populate(mm, (p4d_t *) table, (pud_t *) pgd);
110 mm->pgd = (pgd_t *) table;
111 mm->context.asce_limit = _REGION1_SIZE;
112 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
113 _ASCE_USER_BITS | _ASCE_TYPE_REGION2;
116 crst_table_init(table, _REGION1_ENTRY_EMPTY);
117 pgd_populate(mm, (pgd_t *) table, (p4d_t *) pgd);
118 mm->pgd = (pgd_t *) table;
119 mm->context.asce_limit = -PAGE_SIZE;
120 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
121 _ASCE_USER_BITS | _ASCE_TYPE_REGION1;
124 spin_unlock_bh(&mm->page_table_lock);
127 on_each_cpu(__crst_table_upgrade, mm, 0);
131 void crst_table_downgrade(struct mm_struct *mm)
135 /* downgrade should only happen from 3 to 2 levels (compat only) */
136 VM_BUG_ON(mm->context.asce_limit != _REGION2_SIZE);
138 if (current->active_mm == mm) {
145 mm->pgd = (pgd_t *) (pgd_val(*pgd) & _REGION_ENTRY_ORIGIN);
146 mm->context.asce_limit = _REGION3_SIZE;
147 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
148 _ASCE_USER_BITS | _ASCE_TYPE_SEGMENT;
149 crst_table_free(mm, (unsigned long *) pgd);
151 if (current->active_mm == mm)
155 static inline unsigned int atomic_xor_bits(atomic_t *v, unsigned int bits)
157 unsigned int old, new;
160 old = atomic_read(v);
162 } while (atomic_cmpxchg(v, old, new) != old);
168 struct page *page_table_alloc_pgste(struct mm_struct *mm)
173 page = alloc_page(GFP_KERNEL);
175 table = (u64 *)page_to_phys(page);
176 memset64(table, _PAGE_INVALID, PTRS_PER_PTE);
177 memset64(table + PTRS_PER_PTE, 0, PTRS_PER_PTE);
182 void page_table_free_pgste(struct page *page)
187 #endif /* CONFIG_PGSTE */
190 * page table entry allocation/free routines.
192 unsigned long *page_table_alloc(struct mm_struct *mm)
194 unsigned long *table;
196 unsigned int mask, bit;
198 /* Try to get a fragment of a 4K page as a 2K page table */
199 if (!mm_alloc_pgste(mm)) {
201 spin_lock_bh(&mm->context.lock);
202 if (!list_empty(&mm->context.pgtable_list)) {
203 page = list_first_entry(&mm->context.pgtable_list,
205 mask = atomic_read(&page->_refcount) >> 24;
206 mask = (mask | (mask >> 4)) & 3;
208 table = (unsigned long *) page_to_phys(page);
209 bit = mask & 1; /* =1 -> second 2K */
211 table += PTRS_PER_PTE;
212 atomic_xor_bits(&page->_refcount,
214 list_del(&page->lru);
217 spin_unlock_bh(&mm->context.lock);
221 /* Allocate a fresh page */
222 page = alloc_page(GFP_KERNEL);
225 if (!pgtable_pte_page_ctor(page)) {
229 arch_set_page_dat(page, 0);
230 /* Initialize page table */
231 table = (unsigned long *) page_to_phys(page);
232 if (mm_alloc_pgste(mm)) {
233 /* Return 4K page table with PGSTEs */
234 atomic_xor_bits(&page->_refcount, 3 << 24);
235 memset64((u64 *)table, _PAGE_INVALID, PTRS_PER_PTE);
236 memset64((u64 *)table + PTRS_PER_PTE, 0, PTRS_PER_PTE);
238 /* Return the first 2K fragment of the page */
239 atomic_xor_bits(&page->_refcount, 1 << 24);
240 memset64((u64 *)table, _PAGE_INVALID, 2 * PTRS_PER_PTE);
241 spin_lock_bh(&mm->context.lock);
242 list_add(&page->lru, &mm->context.pgtable_list);
243 spin_unlock_bh(&mm->context.lock);
248 void page_table_free(struct mm_struct *mm, unsigned long *table)
251 unsigned int bit, mask;
253 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
254 if (!mm_alloc_pgste(mm)) {
255 /* Free 2K page table fragment of a 4K page */
256 bit = (__pa(table) & ~PAGE_MASK)/(PTRS_PER_PTE*sizeof(pte_t));
257 spin_lock_bh(&mm->context.lock);
258 mask = atomic_xor_bits(&page->_refcount, 0x11U << (bit + 24));
261 list_add(&page->lru, &mm->context.pgtable_list);
263 list_del(&page->lru);
264 spin_unlock_bh(&mm->context.lock);
265 mask = atomic_xor_bits(&page->_refcount, 0x10U << (bit + 24));
270 atomic_xor_bits(&page->_refcount, 3U << 24);
273 pgtable_pte_page_dtor(page);
277 void page_table_free_rcu(struct mmu_gather *tlb, unsigned long *table,
278 unsigned long vmaddr)
280 struct mm_struct *mm;
282 unsigned int bit, mask;
285 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
286 if (mm_alloc_pgste(mm)) {
287 gmap_unlink(mm, table, vmaddr);
288 table = (unsigned long *) (__pa(table) | 3);
289 tlb_remove_table(tlb, table);
292 bit = (__pa(table) & ~PAGE_MASK) / (PTRS_PER_PTE*sizeof(pte_t));
293 spin_lock_bh(&mm->context.lock);
294 mask = atomic_xor_bits(&page->_refcount, 0x11U << (bit + 24));
297 list_add_tail(&page->lru, &mm->context.pgtable_list);
299 list_del(&page->lru);
300 spin_unlock_bh(&mm->context.lock);
301 table = (unsigned long *) (__pa(table) | (1U << bit));
302 tlb_remove_table(tlb, table);
305 void __tlb_remove_table(void *_table)
307 unsigned int mask = (unsigned long) _table & 3;
308 void *table = (void *)((unsigned long) _table ^ mask);
309 struct page *page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
312 case 0: /* pmd, pud, or p4d */
313 free_pages((unsigned long) table, 2);
315 case 1: /* lower 2K of a 4K page table */
316 case 2: /* higher 2K of a 4K page table */
317 mask = atomic_xor_bits(&page->_refcount, mask << (4 + 24));
322 case 3: /* 4K page table with pgstes */
324 atomic_xor_bits(&page->_refcount, 3 << 24);
325 pgtable_pte_page_dtor(page);
332 * Base infrastructure required to generate basic asces, region, segment,
333 * and page tables that do not make use of enhanced features like EDAT1.
336 static struct kmem_cache *base_pgt_cache;
338 static unsigned long base_pgt_alloc(void)
342 table = kmem_cache_alloc(base_pgt_cache, GFP_KERNEL);
344 memset64(table, _PAGE_INVALID, PTRS_PER_PTE);
345 return (unsigned long) table;
348 static void base_pgt_free(unsigned long table)
350 kmem_cache_free(base_pgt_cache, (void *) table);
353 static unsigned long base_crst_alloc(unsigned long val)
357 table = __get_free_pages(GFP_KERNEL, CRST_ALLOC_ORDER);
359 crst_table_init((unsigned long *)table, val);
363 static void base_crst_free(unsigned long table)
365 free_pages(table, CRST_ALLOC_ORDER);
368 #define BASE_ADDR_END_FUNC(NAME, SIZE) \
369 static inline unsigned long base_##NAME##_addr_end(unsigned long addr, \
372 unsigned long next = (addr + (SIZE)) & ~((SIZE) - 1); \
374 return (next - 1) < (end - 1) ? next : end; \
377 BASE_ADDR_END_FUNC(page, _PAGE_SIZE)
378 BASE_ADDR_END_FUNC(segment, _SEGMENT_SIZE)
379 BASE_ADDR_END_FUNC(region3, _REGION3_SIZE)
380 BASE_ADDR_END_FUNC(region2, _REGION2_SIZE)
381 BASE_ADDR_END_FUNC(region1, _REGION1_SIZE)
383 static inline unsigned long base_lra(unsigned long address)
389 : "=d" (real) : "a" (address) : "cc");
393 static int base_page_walk(unsigned long origin, unsigned long addr,
394 unsigned long end, int alloc)
396 unsigned long *pte, next;
400 pte = (unsigned long *) origin;
401 pte += (addr & _PAGE_INDEX) >> _PAGE_SHIFT;
403 next = base_page_addr_end(addr, end);
404 *pte = base_lra(addr);
405 } while (pte++, addr = next, addr < end);
409 static int base_segment_walk(unsigned long origin, unsigned long addr,
410 unsigned long end, int alloc)
412 unsigned long *ste, next, table;
415 ste = (unsigned long *) origin;
416 ste += (addr & _SEGMENT_INDEX) >> _SEGMENT_SHIFT;
418 next = base_segment_addr_end(addr, end);
419 if (*ste & _SEGMENT_ENTRY_INVALID) {
422 table = base_pgt_alloc();
425 *ste = table | _SEGMENT_ENTRY;
427 table = *ste & _SEGMENT_ENTRY_ORIGIN;
428 rc = base_page_walk(table, addr, next, alloc);
432 base_pgt_free(table);
434 } while (ste++, addr = next, addr < end);
438 static int base_region3_walk(unsigned long origin, unsigned long addr,
439 unsigned long end, int alloc)
441 unsigned long *rtte, next, table;
444 rtte = (unsigned long *) origin;
445 rtte += (addr & _REGION3_INDEX) >> _REGION3_SHIFT;
447 next = base_region3_addr_end(addr, end);
448 if (*rtte & _REGION_ENTRY_INVALID) {
451 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY);
454 *rtte = table | _REGION3_ENTRY;
456 table = *rtte & _REGION_ENTRY_ORIGIN;
457 rc = base_segment_walk(table, addr, next, alloc);
461 base_crst_free(table);
462 } while (rtte++, addr = next, addr < end);
466 static int base_region2_walk(unsigned long origin, unsigned long addr,
467 unsigned long end, int alloc)
469 unsigned long *rste, next, table;
472 rste = (unsigned long *) origin;
473 rste += (addr & _REGION2_INDEX) >> _REGION2_SHIFT;
475 next = base_region2_addr_end(addr, end);
476 if (*rste & _REGION_ENTRY_INVALID) {
479 table = base_crst_alloc(_REGION3_ENTRY_EMPTY);
482 *rste = table | _REGION2_ENTRY;
484 table = *rste & _REGION_ENTRY_ORIGIN;
485 rc = base_region3_walk(table, addr, next, alloc);
489 base_crst_free(table);
490 } while (rste++, addr = next, addr < end);
494 static int base_region1_walk(unsigned long origin, unsigned long addr,
495 unsigned long end, int alloc)
497 unsigned long *rfte, next, table;
500 rfte = (unsigned long *) origin;
501 rfte += (addr & _REGION1_INDEX) >> _REGION1_SHIFT;
503 next = base_region1_addr_end(addr, end);
504 if (*rfte & _REGION_ENTRY_INVALID) {
507 table = base_crst_alloc(_REGION2_ENTRY_EMPTY);
510 *rfte = table | _REGION1_ENTRY;
512 table = *rfte & _REGION_ENTRY_ORIGIN;
513 rc = base_region2_walk(table, addr, next, alloc);
517 base_crst_free(table);
518 } while (rfte++, addr = next, addr < end);
523 * base_asce_free - free asce and tables returned from base_asce_alloc()
524 * @asce: asce to be freed
526 * Frees all region, segment, and page tables that were allocated with a
527 * corresponding base_asce_alloc() call.
529 void base_asce_free(unsigned long asce)
531 unsigned long table = asce & _ASCE_ORIGIN;
535 switch (asce & _ASCE_TYPE_MASK) {
536 case _ASCE_TYPE_SEGMENT:
537 base_segment_walk(table, 0, _REGION3_SIZE, 0);
539 case _ASCE_TYPE_REGION3:
540 base_region3_walk(table, 0, _REGION2_SIZE, 0);
542 case _ASCE_TYPE_REGION2:
543 base_region2_walk(table, 0, _REGION1_SIZE, 0);
545 case _ASCE_TYPE_REGION1:
546 base_region1_walk(table, 0, -_PAGE_SIZE, 0);
549 base_crst_free(table);
552 static int base_pgt_cache_init(void)
554 static DEFINE_MUTEX(base_pgt_cache_mutex);
555 unsigned long sz = _PAGE_TABLE_SIZE;
559 mutex_lock(&base_pgt_cache_mutex);
561 base_pgt_cache = kmem_cache_create("base_pgt", sz, sz, 0, NULL);
562 mutex_unlock(&base_pgt_cache_mutex);
563 return base_pgt_cache ? 0 : -ENOMEM;
567 * base_asce_alloc - create kernel mapping without enhanced DAT features
568 * @addr: virtual start address of kernel mapping
569 * @num_pages: number of consecutive pages
571 * Generate an asce, including all required region, segment and page tables,
572 * that can be used to access the virtual kernel mapping. The difference is
573 * that the returned asce does not make use of any enhanced DAT features like
574 * e.g. large pages. This is required for some I/O functions that pass an
575 * asce, like e.g. some service call requests.
577 * Note: the returned asce may NEVER be attached to any cpu. It may only be
578 * used for I/O requests. tlb entries that might result because the
579 * asce was attached to a cpu won't be cleared.
581 unsigned long base_asce_alloc(unsigned long addr, unsigned long num_pages)
583 unsigned long asce, table, end;
586 if (base_pgt_cache_init())
588 end = addr + num_pages * PAGE_SIZE;
589 if (end <= _REGION3_SIZE) {
590 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY);
593 rc = base_segment_walk(table, addr, end, 1);
594 asce = table | _ASCE_TYPE_SEGMENT | _ASCE_TABLE_LENGTH;
595 } else if (end <= _REGION2_SIZE) {
596 table = base_crst_alloc(_REGION3_ENTRY_EMPTY);
599 rc = base_region3_walk(table, addr, end, 1);
600 asce = table | _ASCE_TYPE_REGION3 | _ASCE_TABLE_LENGTH;
601 } else if (end <= _REGION1_SIZE) {
602 table = base_crst_alloc(_REGION2_ENTRY_EMPTY);
605 rc = base_region2_walk(table, addr, end, 1);
606 asce = table | _ASCE_TYPE_REGION2 | _ASCE_TABLE_LENGTH;
608 table = base_crst_alloc(_REGION1_ENTRY_EMPTY);
611 rc = base_region1_walk(table, addr, end, 1);
612 asce = table | _ASCE_TYPE_REGION1 | _ASCE_TABLE_LENGTH;
615 base_asce_free(asce);