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 if (current->active_mm == mm)
78 int crst_table_upgrade(struct mm_struct *mm, unsigned long end)
80 unsigned long *pgd = NULL, *p4d = NULL, *__pgd;
81 unsigned long asce_limit = mm->context.asce_limit;
83 /* upgrade should only happen from 3 to 4, 3 to 5, or 4 to 5 levels */
84 VM_BUG_ON(asce_limit < _REGION2_SIZE);
86 if (end <= asce_limit)
89 if (asce_limit == _REGION2_SIZE) {
90 p4d = crst_table_alloc(mm);
93 crst_table_init(p4d, _REGION2_ENTRY_EMPTY);
95 if (end > _REGION1_SIZE) {
96 pgd = crst_table_alloc(mm);
99 crst_table_init(pgd, _REGION1_ENTRY_EMPTY);
102 spin_lock_bh(&mm->page_table_lock);
105 * This routine gets called with mmap_sem lock held and there is
106 * no reason to optimize for the case of otherwise. However, if
107 * that would ever change, the below check will let us know.
109 VM_BUG_ON(asce_limit != mm->context.asce_limit);
112 __pgd = (unsigned long *) mm->pgd;
113 p4d_populate(mm, (p4d_t *) p4d, (pud_t *) __pgd);
114 mm->pgd = (pgd_t *) p4d;
115 mm->context.asce_limit = _REGION1_SIZE;
116 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
117 _ASCE_USER_BITS | _ASCE_TYPE_REGION2;
121 __pgd = (unsigned long *) mm->pgd;
122 pgd_populate(mm, (pgd_t *) pgd, (p4d_t *) __pgd);
123 mm->pgd = (pgd_t *) pgd;
124 mm->context.asce_limit = TASK_SIZE_MAX;
125 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
126 _ASCE_USER_BITS | _ASCE_TYPE_REGION1;
129 spin_unlock_bh(&mm->page_table_lock);
131 on_each_cpu(__crst_table_upgrade, mm, 0);
136 crst_table_free(mm, p4d);
141 static inline unsigned int atomic_xor_bits(atomic_t *v, unsigned int bits)
143 unsigned int old, new;
146 old = atomic_read(v);
148 } while (atomic_cmpxchg(v, old, new) != old);
154 struct page *page_table_alloc_pgste(struct mm_struct *mm)
159 page = alloc_page(GFP_KERNEL);
161 table = (u64 *)page_to_phys(page);
162 memset64(table, _PAGE_INVALID, PTRS_PER_PTE);
163 memset64(table + PTRS_PER_PTE, 0, PTRS_PER_PTE);
168 void page_table_free_pgste(struct page *page)
173 #endif /* CONFIG_PGSTE */
176 * page table entry allocation/free routines.
178 unsigned long *page_table_alloc(struct mm_struct *mm)
180 unsigned long *table;
182 unsigned int mask, bit;
184 /* Try to get a fragment of a 4K page as a 2K page table */
185 if (!mm_alloc_pgste(mm)) {
187 spin_lock_bh(&mm->context.lock);
188 if (!list_empty(&mm->context.pgtable_list)) {
189 page = list_first_entry(&mm->context.pgtable_list,
191 mask = atomic_read(&page->_refcount) >> 24;
192 mask = (mask | (mask >> 4)) & 3;
194 table = (unsigned long *) page_to_phys(page);
195 bit = mask & 1; /* =1 -> second 2K */
197 table += PTRS_PER_PTE;
198 atomic_xor_bits(&page->_refcount,
200 list_del(&page->lru);
203 spin_unlock_bh(&mm->context.lock);
207 /* Allocate a fresh page */
208 page = alloc_page(GFP_KERNEL);
211 if (!pgtable_pte_page_ctor(page)) {
215 arch_set_page_dat(page, 0);
216 /* Initialize page table */
217 table = (unsigned long *) page_to_phys(page);
218 if (mm_alloc_pgste(mm)) {
219 /* Return 4K page table with PGSTEs */
220 atomic_xor_bits(&page->_refcount, 3 << 24);
221 memset64((u64 *)table, _PAGE_INVALID, PTRS_PER_PTE);
222 memset64((u64 *)table + PTRS_PER_PTE, 0, PTRS_PER_PTE);
224 /* Return the first 2K fragment of the page */
225 atomic_xor_bits(&page->_refcount, 1 << 24);
226 memset64((u64 *)table, _PAGE_INVALID, 2 * PTRS_PER_PTE);
227 spin_lock_bh(&mm->context.lock);
228 list_add(&page->lru, &mm->context.pgtable_list);
229 spin_unlock_bh(&mm->context.lock);
234 void page_table_free(struct mm_struct *mm, unsigned long *table)
237 unsigned int bit, mask;
239 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
240 if (!mm_alloc_pgste(mm)) {
241 /* Free 2K page table fragment of a 4K page */
242 bit = (__pa(table) & ~PAGE_MASK)/(PTRS_PER_PTE*sizeof(pte_t));
243 spin_lock_bh(&mm->context.lock);
244 mask = atomic_xor_bits(&page->_refcount, 1U << (bit + 24));
247 list_add(&page->lru, &mm->context.pgtable_list);
249 list_del(&page->lru);
250 spin_unlock_bh(&mm->context.lock);
254 atomic_xor_bits(&page->_refcount, 3U << 24);
257 pgtable_pte_page_dtor(page);
261 void page_table_free_rcu(struct mmu_gather *tlb, unsigned long *table,
262 unsigned long vmaddr)
264 struct mm_struct *mm;
266 unsigned int bit, mask;
269 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
270 if (mm_alloc_pgste(mm)) {
271 gmap_unlink(mm, table, vmaddr);
272 table = (unsigned long *) (__pa(table) | 3);
273 tlb_remove_table(tlb, table);
276 bit = (__pa(table) & ~PAGE_MASK) / (PTRS_PER_PTE*sizeof(pte_t));
277 spin_lock_bh(&mm->context.lock);
278 mask = atomic_xor_bits(&page->_refcount, 0x11U << (bit + 24));
281 list_add_tail(&page->lru, &mm->context.pgtable_list);
283 list_del(&page->lru);
284 spin_unlock_bh(&mm->context.lock);
285 table = (unsigned long *) (__pa(table) | (1U << bit));
286 tlb_remove_table(tlb, table);
289 void __tlb_remove_table(void *_table)
291 unsigned int mask = (unsigned long) _table & 3;
292 void *table = (void *)((unsigned long) _table ^ mask);
293 struct page *page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
296 case 0: /* pmd, pud, or p4d */
297 free_pages((unsigned long) table, 2);
299 case 1: /* lower 2K of a 4K page table */
300 case 2: /* higher 2K of a 4K page table */
301 mask = atomic_xor_bits(&page->_refcount, mask << (4 + 24));
306 case 3: /* 4K page table with pgstes */
308 atomic_xor_bits(&page->_refcount, 3 << 24);
309 pgtable_pte_page_dtor(page);
316 * Base infrastructure required to generate basic asces, region, segment,
317 * and page tables that do not make use of enhanced features like EDAT1.
320 static struct kmem_cache *base_pgt_cache;
322 static unsigned long base_pgt_alloc(void)
326 table = kmem_cache_alloc(base_pgt_cache, GFP_KERNEL);
328 memset64(table, _PAGE_INVALID, PTRS_PER_PTE);
329 return (unsigned long) table;
332 static void base_pgt_free(unsigned long table)
334 kmem_cache_free(base_pgt_cache, (void *) table);
337 static unsigned long base_crst_alloc(unsigned long val)
341 table = __get_free_pages(GFP_KERNEL, CRST_ALLOC_ORDER);
343 crst_table_init((unsigned long *)table, val);
347 static void base_crst_free(unsigned long table)
349 free_pages(table, CRST_ALLOC_ORDER);
352 #define BASE_ADDR_END_FUNC(NAME, SIZE) \
353 static inline unsigned long base_##NAME##_addr_end(unsigned long addr, \
356 unsigned long next = (addr + (SIZE)) & ~((SIZE) - 1); \
358 return (next - 1) < (end - 1) ? next : end; \
361 BASE_ADDR_END_FUNC(page, _PAGE_SIZE)
362 BASE_ADDR_END_FUNC(segment, _SEGMENT_SIZE)
363 BASE_ADDR_END_FUNC(region3, _REGION3_SIZE)
364 BASE_ADDR_END_FUNC(region2, _REGION2_SIZE)
365 BASE_ADDR_END_FUNC(region1, _REGION1_SIZE)
367 static inline unsigned long base_lra(unsigned long address)
373 : "=d" (real) : "a" (address) : "cc");
377 static int base_page_walk(unsigned long origin, unsigned long addr,
378 unsigned long end, int alloc)
380 unsigned long *pte, next;
384 pte = (unsigned long *) origin;
385 pte += (addr & _PAGE_INDEX) >> _PAGE_SHIFT;
387 next = base_page_addr_end(addr, end);
388 *pte = base_lra(addr);
389 } while (pte++, addr = next, addr < end);
393 static int base_segment_walk(unsigned long origin, unsigned long addr,
394 unsigned long end, int alloc)
396 unsigned long *ste, next, table;
399 ste = (unsigned long *) origin;
400 ste += (addr & _SEGMENT_INDEX) >> _SEGMENT_SHIFT;
402 next = base_segment_addr_end(addr, end);
403 if (*ste & _SEGMENT_ENTRY_INVALID) {
406 table = base_pgt_alloc();
409 *ste = table | _SEGMENT_ENTRY;
411 table = *ste & _SEGMENT_ENTRY_ORIGIN;
412 rc = base_page_walk(table, addr, next, alloc);
416 base_pgt_free(table);
418 } while (ste++, addr = next, addr < end);
422 static int base_region3_walk(unsigned long origin, unsigned long addr,
423 unsigned long end, int alloc)
425 unsigned long *rtte, next, table;
428 rtte = (unsigned long *) origin;
429 rtte += (addr & _REGION3_INDEX) >> _REGION3_SHIFT;
431 next = base_region3_addr_end(addr, end);
432 if (*rtte & _REGION_ENTRY_INVALID) {
435 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY);
438 *rtte = table | _REGION3_ENTRY;
440 table = *rtte & _REGION_ENTRY_ORIGIN;
441 rc = base_segment_walk(table, addr, next, alloc);
445 base_crst_free(table);
446 } while (rtte++, addr = next, addr < end);
450 static int base_region2_walk(unsigned long origin, unsigned long addr,
451 unsigned long end, int alloc)
453 unsigned long *rste, next, table;
456 rste = (unsigned long *) origin;
457 rste += (addr & _REGION2_INDEX) >> _REGION2_SHIFT;
459 next = base_region2_addr_end(addr, end);
460 if (*rste & _REGION_ENTRY_INVALID) {
463 table = base_crst_alloc(_REGION3_ENTRY_EMPTY);
466 *rste = table | _REGION2_ENTRY;
468 table = *rste & _REGION_ENTRY_ORIGIN;
469 rc = base_region3_walk(table, addr, next, alloc);
473 base_crst_free(table);
474 } while (rste++, addr = next, addr < end);
478 static int base_region1_walk(unsigned long origin, unsigned long addr,
479 unsigned long end, int alloc)
481 unsigned long *rfte, next, table;
484 rfte = (unsigned long *) origin;
485 rfte += (addr & _REGION1_INDEX) >> _REGION1_SHIFT;
487 next = base_region1_addr_end(addr, end);
488 if (*rfte & _REGION_ENTRY_INVALID) {
491 table = base_crst_alloc(_REGION2_ENTRY_EMPTY);
494 *rfte = table | _REGION1_ENTRY;
496 table = *rfte & _REGION_ENTRY_ORIGIN;
497 rc = base_region2_walk(table, addr, next, alloc);
501 base_crst_free(table);
502 } while (rfte++, addr = next, addr < end);
507 * base_asce_free - free asce and tables returned from base_asce_alloc()
508 * @asce: asce to be freed
510 * Frees all region, segment, and page tables that were allocated with a
511 * corresponding base_asce_alloc() call.
513 void base_asce_free(unsigned long asce)
515 unsigned long table = asce & _ASCE_ORIGIN;
519 switch (asce & _ASCE_TYPE_MASK) {
520 case _ASCE_TYPE_SEGMENT:
521 base_segment_walk(table, 0, _REGION3_SIZE, 0);
523 case _ASCE_TYPE_REGION3:
524 base_region3_walk(table, 0, _REGION2_SIZE, 0);
526 case _ASCE_TYPE_REGION2:
527 base_region2_walk(table, 0, _REGION1_SIZE, 0);
529 case _ASCE_TYPE_REGION1:
530 base_region1_walk(table, 0, TASK_SIZE_MAX, 0);
533 base_crst_free(table);
536 static int base_pgt_cache_init(void)
538 static DEFINE_MUTEX(base_pgt_cache_mutex);
539 unsigned long sz = _PAGE_TABLE_SIZE;
543 mutex_lock(&base_pgt_cache_mutex);
545 base_pgt_cache = kmem_cache_create("base_pgt", sz, sz, 0, NULL);
546 mutex_unlock(&base_pgt_cache_mutex);
547 return base_pgt_cache ? 0 : -ENOMEM;
551 * base_asce_alloc - create kernel mapping without enhanced DAT features
552 * @addr: virtual start address of kernel mapping
553 * @num_pages: number of consecutive pages
555 * Generate an asce, including all required region, segment and page tables,
556 * that can be used to access the virtual kernel mapping. The difference is
557 * that the returned asce does not make use of any enhanced DAT features like
558 * e.g. large pages. This is required for some I/O functions that pass an
559 * asce, like e.g. some service call requests.
561 * Note: the returned asce may NEVER be attached to any cpu. It may only be
562 * used for I/O requests. tlb entries that might result because the
563 * asce was attached to a cpu won't be cleared.
565 unsigned long base_asce_alloc(unsigned long addr, unsigned long num_pages)
567 unsigned long asce, table, end;
570 if (base_pgt_cache_init())
572 end = addr + num_pages * PAGE_SIZE;
573 if (end <= _REGION3_SIZE) {
574 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY);
577 rc = base_segment_walk(table, addr, end, 1);
578 asce = table | _ASCE_TYPE_SEGMENT | _ASCE_TABLE_LENGTH;
579 } else if (end <= _REGION2_SIZE) {
580 table = base_crst_alloc(_REGION3_ENTRY_EMPTY);
583 rc = base_region3_walk(table, addr, end, 1);
584 asce = table | _ASCE_TYPE_REGION3 | _ASCE_TABLE_LENGTH;
585 } else if (end <= _REGION1_SIZE) {
586 table = base_crst_alloc(_REGION2_ENTRY_EMPTY);
589 rc = base_region2_walk(table, addr, end, 1);
590 asce = table | _ASCE_TYPE_REGION2 | _ASCE_TABLE_LENGTH;
592 table = base_crst_alloc(_REGION1_ENTRY_EMPTY);
595 rc = base_region1_walk(table, addr, end, 1);
596 asce = table | _ASCE_TYPE_REGION1 | _ASCE_TABLE_LENGTH;
599 base_asce_free(asce);