Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Page table allocation functions
4 *
5 * Copyright IBM Corp. 2016
6 * Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
7 */
8
9#include <linux/sysctl.h>
10#include <linux/slab.h>
11#include <linux/mm.h>
12#include <asm/mmu_context.h>
13#include <asm/pgalloc.h>
14#include <asm/gmap.h>
15#include <asm/tlb.h>
16#include <asm/tlbflush.h>
17
18#ifdef CONFIG_PGSTE
19
20int page_table_allocate_pgste = 0;
21EXPORT_SYMBOL(page_table_allocate_pgste);
22
23static struct ctl_table page_table_sysctl[] = {
24 {
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,
31 .extra2 = SYSCTL_ONE,
32 },
33 { }
34};
35
36static struct ctl_table page_table_sysctl_dir[] = {
37 {
38 .procname = "vm",
39 .maxlen = 0,
40 .mode = 0555,
41 .child = page_table_sysctl,
42 },
43 { }
44};
45
46static int __init page_table_register_sysctl(void)
47{
48 return register_sysctl_table(page_table_sysctl_dir) ? 0 : -ENOMEM;
49}
50__initcall(page_table_register_sysctl);
51
52#endif /* CONFIG_PGSTE */
53
54unsigned long *crst_table_alloc(struct mm_struct *mm)
55{
56 struct page *page = alloc_pages(GFP_KERNEL, CRST_ALLOC_ORDER);
57
58 if (!page)
59 return NULL;
60 arch_set_page_dat(page, CRST_ALLOC_ORDER);
61 return (unsigned long *) page_to_virt(page);
62}
63
64void crst_table_free(struct mm_struct *mm, unsigned long *table)
65{
66 free_pages((unsigned long)table, CRST_ALLOC_ORDER);
67}
68
69static void __crst_table_upgrade(void *arg)
70{
71 struct mm_struct *mm = arg;
72
73 /* 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 __ctl_load(S390_lowcore.user_asce, 7, 7);
77 }
78 __tlb_flush_local();
79}
80
81int crst_table_upgrade(struct mm_struct *mm, unsigned long end)
82{
83 unsigned long *pgd = NULL, *p4d = NULL, *__pgd;
84 unsigned long asce_limit = mm->context.asce_limit;
85
86 /* upgrade should only happen from 3 to 4, 3 to 5, or 4 to 5 levels */
87 VM_BUG_ON(asce_limit < _REGION2_SIZE);
88
89 if (end <= asce_limit)
90 return 0;
91
92 if (asce_limit == _REGION2_SIZE) {
93 p4d = crst_table_alloc(mm);
94 if (unlikely(!p4d))
95 goto err_p4d;
96 crst_table_init(p4d, _REGION2_ENTRY_EMPTY);
97 }
98 if (end > _REGION1_SIZE) {
99 pgd = crst_table_alloc(mm);
100 if (unlikely(!pgd))
101 goto err_pgd;
102 crst_table_init(pgd, _REGION1_ENTRY_EMPTY);
103 }
104
105 spin_lock_bh(&mm->page_table_lock);
106
107 /*
108 * This routine gets called with mmap_lock lock held and there is
109 * no reason to optimize for the case of otherwise. However, if
110 * that would ever change, the below check will let us know.
111 */
112 VM_BUG_ON(asce_limit != mm->context.asce_limit);
113
114 if (p4d) {
115 __pgd = (unsigned long *) mm->pgd;
116 p4d_populate(mm, (p4d_t *) p4d, (pud_t *) __pgd);
117 mm->pgd = (pgd_t *) p4d;
118 mm->context.asce_limit = _REGION1_SIZE;
119 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
120 _ASCE_USER_BITS | _ASCE_TYPE_REGION2;
121 mm_inc_nr_puds(mm);
122 }
123 if (pgd) {
124 __pgd = (unsigned long *) mm->pgd;
125 pgd_populate(mm, (pgd_t *) pgd, (p4d_t *) __pgd);
126 mm->pgd = (pgd_t *) pgd;
127 mm->context.asce_limit = TASK_SIZE_MAX;
128 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
129 _ASCE_USER_BITS | _ASCE_TYPE_REGION1;
130 }
131
132 spin_unlock_bh(&mm->page_table_lock);
133
134 on_each_cpu(__crst_table_upgrade, mm, 0);
135
136 return 0;
137
138err_pgd:
139 crst_table_free(mm, p4d);
140err_p4d:
141 return -ENOMEM;
142}
143
144static inline unsigned int atomic_xor_bits(atomic_t *v, unsigned int bits)
145{
146 unsigned int old, new;
147
148 do {
149 old = atomic_read(v);
150 new = old ^ bits;
151 } while (atomic_cmpxchg(v, old, new) != old);
152 return new;
153}
154
155#ifdef CONFIG_PGSTE
156
157struct page *page_table_alloc_pgste(struct mm_struct *mm)
158{
159 struct page *page;
160 u64 *table;
161
162 page = alloc_page(GFP_KERNEL);
163 if (page) {
164 table = (u64 *)page_to_virt(page);
165 memset64(table, _PAGE_INVALID, PTRS_PER_PTE);
166 memset64(table + PTRS_PER_PTE, 0, PTRS_PER_PTE);
167 }
168 return page;
169}
170
171void page_table_free_pgste(struct page *page)
172{
173 __free_page(page);
174}
175
176#endif /* CONFIG_PGSTE */
177
178/*
179 * A 2KB-pgtable is either upper or lower half of a normal page.
180 * The second half of the page may be unused or used as another
181 * 2KB-pgtable.
182 *
183 * Whenever possible the parent page for a new 2KB-pgtable is picked
184 * from the list of partially allocated pages mm_context_t::pgtable_list.
185 * In case the list is empty a new parent page is allocated and added to
186 * the list.
187 *
188 * When a parent page gets fully allocated it contains 2KB-pgtables in both
189 * upper and lower halves and is removed from mm_context_t::pgtable_list.
190 *
191 * When 2KB-pgtable is freed from to fully allocated parent page that
192 * page turns partially allocated and added to mm_context_t::pgtable_list.
193 *
194 * If 2KB-pgtable is freed from the partially allocated parent page that
195 * page turns unused and gets removed from mm_context_t::pgtable_list.
196 * Furthermore, the unused parent page is released.
197 *
198 * As follows from the above, no unallocated or fully allocated parent
199 * pages are contained in mm_context_t::pgtable_list.
200 *
201 * The upper byte (bits 24-31) of the parent page _refcount is used
202 * for tracking contained 2KB-pgtables and has the following format:
203 *
204 * PP AA
205 * 01234567 upper byte (bits 24-31) of struct page::_refcount
206 * || ||
207 * || |+--- upper 2KB-pgtable is allocated
208 * || +---- lower 2KB-pgtable is allocated
209 * |+------- upper 2KB-pgtable is pending for removal
210 * +-------- lower 2KB-pgtable is pending for removal
211 *
212 * (See commit 620b4e903179 ("s390: use _refcount for pgtables") on why
213 * using _refcount is possible).
214 *
215 * When 2KB-pgtable is allocated the corresponding AA bit is set to 1.
216 * The parent page is either:
217 * - added to mm_context_t::pgtable_list in case the second half of the
218 * parent page is still unallocated;
219 * - removed from mm_context_t::pgtable_list in case both hales of the
220 * parent page are allocated;
221 * These operations are protected with mm_context_t::lock.
222 *
223 * When 2KB-pgtable is deallocated the corresponding AA bit is set to 0
224 * and the corresponding PP bit is set to 1 in a single atomic operation.
225 * Thus, PP and AA bits corresponding to the same 2KB-pgtable are mutually
226 * exclusive and may never be both set to 1!
227 * The parent page is either:
228 * - added to mm_context_t::pgtable_list in case the second half of the
229 * parent page is still allocated;
230 * - removed from mm_context_t::pgtable_list in case the second half of
231 * the parent page is unallocated;
232 * These operations are protected with mm_context_t::lock.
233 *
234 * It is important to understand that mm_context_t::lock only protects
235 * mm_context_t::pgtable_list and AA bits, but not the parent page itself
236 * and PP bits.
237 *
238 * Releasing the parent page happens whenever the PP bit turns from 1 to 0,
239 * while both AA bits and the second PP bit are already unset. Then the
240 * parent page does not contain any 2KB-pgtable fragment anymore, and it has
241 * also been removed from mm_context_t::pgtable_list. It is safe to release
242 * the page therefore.
243 *
244 * PGSTE memory spaces use full 4KB-pgtables and do not need most of the
245 * logic described above. Both AA bits are set to 1 to denote a 4KB-pgtable
246 * while the PP bits are never used, nor such a page is added to or removed
247 * from mm_context_t::pgtable_list.
248 */
249unsigned long *page_table_alloc(struct mm_struct *mm)
250{
251 unsigned long *table;
252 struct page *page;
253 unsigned int mask, bit;
254
255 /* Try to get a fragment of a 4K page as a 2K page table */
256 if (!mm_alloc_pgste(mm)) {
257 table = NULL;
258 spin_lock_bh(&mm->context.lock);
259 if (!list_empty(&mm->context.pgtable_list)) {
260 page = list_first_entry(&mm->context.pgtable_list,
261 struct page, lru);
262 mask = atomic_read(&page->_refcount) >> 24;
263 /*
264 * The pending removal bits must also be checked.
265 * Failure to do so might lead to an impossible
266 * value of (i.e 0x13 or 0x23) written to _refcount.
267 * Such values violate the assumption that pending and
268 * allocation bits are mutually exclusive, and the rest
269 * of the code unrails as result. That could lead to
270 * a whole bunch of races and corruptions.
271 */
272 mask = (mask | (mask >> 4)) & 0x03U;
273 if (mask != 0x03U) {
274 table = (unsigned long *) page_to_virt(page);
275 bit = mask & 1; /* =1 -> second 2K */
276 if (bit)
277 table += PTRS_PER_PTE;
278 atomic_xor_bits(&page->_refcount,
279 0x01U << (bit + 24));
280 list_del(&page->lru);
281 }
282 }
283 spin_unlock_bh(&mm->context.lock);
284 if (table)
285 return table;
286 }
287 /* Allocate a fresh page */
288 page = alloc_page(GFP_KERNEL);
289 if (!page)
290 return NULL;
291 if (!pgtable_pte_page_ctor(page)) {
292 __free_page(page);
293 return NULL;
294 }
295 arch_set_page_dat(page, 0);
296 /* Initialize page table */
297 table = (unsigned long *) page_to_virt(page);
298 if (mm_alloc_pgste(mm)) {
299 /* Return 4K page table with PGSTEs */
300 atomic_xor_bits(&page->_refcount, 0x03U << 24);
301 memset64((u64 *)table, _PAGE_INVALID, PTRS_PER_PTE);
302 memset64((u64 *)table + PTRS_PER_PTE, 0, PTRS_PER_PTE);
303 } else {
304 /* Return the first 2K fragment of the page */
305 atomic_xor_bits(&page->_refcount, 0x01U << 24);
306 memset64((u64 *)table, _PAGE_INVALID, 2 * PTRS_PER_PTE);
307 spin_lock_bh(&mm->context.lock);
308 list_add(&page->lru, &mm->context.pgtable_list);
309 spin_unlock_bh(&mm->context.lock);
310 }
311 return table;
312}
313
314static void page_table_release_check(struct page *page, void *table,
315 unsigned int half, unsigned int mask)
316{
317 char msg[128];
318
319 if (!IS_ENABLED(CONFIG_DEBUG_VM) || !mask)
320 return;
321 snprintf(msg, sizeof(msg),
322 "Invalid pgtable %p release half 0x%02x mask 0x%02x",
323 table, half, mask);
324 dump_page(page, msg);
325}
326
327void page_table_free(struct mm_struct *mm, unsigned long *table)
328{
329 unsigned int mask, bit, half;
330 struct page *page;
331
332 page = virt_to_page(table);
333 if (!mm_alloc_pgste(mm)) {
334 /* Free 2K page table fragment of a 4K page */
335 bit = ((unsigned long) table & ~PAGE_MASK)/(PTRS_PER_PTE*sizeof(pte_t));
336 spin_lock_bh(&mm->context.lock);
337 /*
338 * Mark the page for delayed release. The actual release
339 * will happen outside of the critical section from this
340 * function or from __tlb_remove_table()
341 */
342 mask = atomic_xor_bits(&page->_refcount, 0x11U << (bit + 24));
343 mask >>= 24;
344 if (mask & 0x03U)
345 list_add(&page->lru, &mm->context.pgtable_list);
346 else
347 list_del(&page->lru);
348 spin_unlock_bh(&mm->context.lock);
349 mask = atomic_xor_bits(&page->_refcount, 0x10U << (bit + 24));
350 mask >>= 24;
351 if (mask != 0x00U)
352 return;
353 half = 0x01U << bit;
354 } else {
355 half = 0x03U;
356 mask = atomic_xor_bits(&page->_refcount, 0x03U << 24);
357 mask >>= 24;
358 }
359
360 page_table_release_check(page, table, half, mask);
361 pgtable_pte_page_dtor(page);
362 __free_page(page);
363}
364
365void page_table_free_rcu(struct mmu_gather *tlb, unsigned long *table,
366 unsigned long vmaddr)
367{
368 struct mm_struct *mm;
369 struct page *page;
370 unsigned int bit, mask;
371
372 mm = tlb->mm;
373 page = virt_to_page(table);
374 if (mm_alloc_pgste(mm)) {
375 gmap_unlink(mm, table, vmaddr);
376 table = (unsigned long *) ((unsigned long)table | 0x03U);
377 tlb_remove_table(tlb, table);
378 return;
379 }
380 bit = ((unsigned long) table & ~PAGE_MASK) / (PTRS_PER_PTE*sizeof(pte_t));
381 spin_lock_bh(&mm->context.lock);
382 /*
383 * Mark the page for delayed release. The actual release will happen
384 * outside of the critical section from __tlb_remove_table() or from
385 * page_table_free()
386 */
387 mask = atomic_xor_bits(&page->_refcount, 0x11U << (bit + 24));
388 mask >>= 24;
389 if (mask & 0x03U)
390 list_add_tail(&page->lru, &mm->context.pgtable_list);
391 else
392 list_del(&page->lru);
393 spin_unlock_bh(&mm->context.lock);
394 table = (unsigned long *) ((unsigned long) table | (0x01U << bit));
395 tlb_remove_table(tlb, table);
396}
397
398void __tlb_remove_table(void *_table)
399{
400 unsigned int mask = (unsigned long) _table & 0x03U, half = mask;
401 void *table = (void *)((unsigned long) _table ^ mask);
402 struct page *page = virt_to_page(table);
403
404 switch (half) {
405 case 0x00U: /* pmd, pud, or p4d */
406 free_pages((unsigned long)table, CRST_ALLOC_ORDER);
407 return;
408 case 0x01U: /* lower 2K of a 4K page table */
409 case 0x02U: /* higher 2K of a 4K page table */
410 mask = atomic_xor_bits(&page->_refcount, mask << (4 + 24));
411 mask >>= 24;
412 if (mask != 0x00U)
413 return;
414 break;
415 case 0x03U: /* 4K page table with pgstes */
416 mask = atomic_xor_bits(&page->_refcount, 0x03U << 24);
417 mask >>= 24;
418 break;
419 }
420
421 page_table_release_check(page, table, half, mask);
422 pgtable_pte_page_dtor(page);
423 __free_page(page);
424}
425
426/*
427 * Base infrastructure required to generate basic asces, region, segment,
428 * and page tables that do not make use of enhanced features like EDAT1.
429 */
430
431static struct kmem_cache *base_pgt_cache;
432
433static unsigned long *base_pgt_alloc(void)
434{
435 unsigned long *table;
436
437 table = kmem_cache_alloc(base_pgt_cache, GFP_KERNEL);
438 if (table)
439 memset64((u64 *)table, _PAGE_INVALID, PTRS_PER_PTE);
440 return table;
441}
442
443static void base_pgt_free(unsigned long *table)
444{
445 kmem_cache_free(base_pgt_cache, table);
446}
447
448static unsigned long *base_crst_alloc(unsigned long val)
449{
450 unsigned long *table;
451
452 table = (unsigned long *)__get_free_pages(GFP_KERNEL, CRST_ALLOC_ORDER);
453 if (table)
454 crst_table_init(table, val);
455 return table;
456}
457
458static void base_crst_free(unsigned long *table)
459{
460 free_pages((unsigned long)table, CRST_ALLOC_ORDER);
461}
462
463#define BASE_ADDR_END_FUNC(NAME, SIZE) \
464static inline unsigned long base_##NAME##_addr_end(unsigned long addr, \
465 unsigned long end) \
466{ \
467 unsigned long next = (addr + (SIZE)) & ~((SIZE) - 1); \
468 \
469 return (next - 1) < (end - 1) ? next : end; \
470}
471
472BASE_ADDR_END_FUNC(page, _PAGE_SIZE)
473BASE_ADDR_END_FUNC(segment, _SEGMENT_SIZE)
474BASE_ADDR_END_FUNC(region3, _REGION3_SIZE)
475BASE_ADDR_END_FUNC(region2, _REGION2_SIZE)
476BASE_ADDR_END_FUNC(region1, _REGION1_SIZE)
477
478static inline unsigned long base_lra(unsigned long address)
479{
480 unsigned long real;
481
482 asm volatile(
483 " lra %0,0(%1)\n"
484 : "=d" (real) : "a" (address) : "cc");
485 return real;
486}
487
488static int base_page_walk(unsigned long *origin, unsigned long addr,
489 unsigned long end, int alloc)
490{
491 unsigned long *pte, next;
492
493 if (!alloc)
494 return 0;
495 pte = origin;
496 pte += (addr & _PAGE_INDEX) >> _PAGE_SHIFT;
497 do {
498 next = base_page_addr_end(addr, end);
499 *pte = base_lra(addr);
500 } while (pte++, addr = next, addr < end);
501 return 0;
502}
503
504static int base_segment_walk(unsigned long *origin, unsigned long addr,
505 unsigned long end, int alloc)
506{
507 unsigned long *ste, next, *table;
508 int rc;
509
510 ste = origin;
511 ste += (addr & _SEGMENT_INDEX) >> _SEGMENT_SHIFT;
512 do {
513 next = base_segment_addr_end(addr, end);
514 if (*ste & _SEGMENT_ENTRY_INVALID) {
515 if (!alloc)
516 continue;
517 table = base_pgt_alloc();
518 if (!table)
519 return -ENOMEM;
520 *ste = __pa(table) | _SEGMENT_ENTRY;
521 }
522 table = __va(*ste & _SEGMENT_ENTRY_ORIGIN);
523 rc = base_page_walk(table, addr, next, alloc);
524 if (rc)
525 return rc;
526 if (!alloc)
527 base_pgt_free(table);
528 cond_resched();
529 } while (ste++, addr = next, addr < end);
530 return 0;
531}
532
533static int base_region3_walk(unsigned long *origin, unsigned long addr,
534 unsigned long end, int alloc)
535{
536 unsigned long *rtte, next, *table;
537 int rc;
538
539 rtte = origin;
540 rtte += (addr & _REGION3_INDEX) >> _REGION3_SHIFT;
541 do {
542 next = base_region3_addr_end(addr, end);
543 if (*rtte & _REGION_ENTRY_INVALID) {
544 if (!alloc)
545 continue;
546 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY);
547 if (!table)
548 return -ENOMEM;
549 *rtte = __pa(table) | _REGION3_ENTRY;
550 }
551 table = __va(*rtte & _REGION_ENTRY_ORIGIN);
552 rc = base_segment_walk(table, addr, next, alloc);
553 if (rc)
554 return rc;
555 if (!alloc)
556 base_crst_free(table);
557 } while (rtte++, addr = next, addr < end);
558 return 0;
559}
560
561static int base_region2_walk(unsigned long *origin, unsigned long addr,
562 unsigned long end, int alloc)
563{
564 unsigned long *rste, next, *table;
565 int rc;
566
567 rste = origin;
568 rste += (addr & _REGION2_INDEX) >> _REGION2_SHIFT;
569 do {
570 next = base_region2_addr_end(addr, end);
571 if (*rste & _REGION_ENTRY_INVALID) {
572 if (!alloc)
573 continue;
574 table = base_crst_alloc(_REGION3_ENTRY_EMPTY);
575 if (!table)
576 return -ENOMEM;
577 *rste = __pa(table) | _REGION2_ENTRY;
578 }
579 table = __va(*rste & _REGION_ENTRY_ORIGIN);
580 rc = base_region3_walk(table, addr, next, alloc);
581 if (rc)
582 return rc;
583 if (!alloc)
584 base_crst_free(table);
585 } while (rste++, addr = next, addr < end);
586 return 0;
587}
588
589static int base_region1_walk(unsigned long *origin, unsigned long addr,
590 unsigned long end, int alloc)
591{
592 unsigned long *rfte, next, *table;
593 int rc;
594
595 rfte = origin;
596 rfte += (addr & _REGION1_INDEX) >> _REGION1_SHIFT;
597 do {
598 next = base_region1_addr_end(addr, end);
599 if (*rfte & _REGION_ENTRY_INVALID) {
600 if (!alloc)
601 continue;
602 table = base_crst_alloc(_REGION2_ENTRY_EMPTY);
603 if (!table)
604 return -ENOMEM;
605 *rfte = __pa(table) | _REGION1_ENTRY;
606 }
607 table = __va(*rfte & _REGION_ENTRY_ORIGIN);
608 rc = base_region2_walk(table, addr, next, alloc);
609 if (rc)
610 return rc;
611 if (!alloc)
612 base_crst_free(table);
613 } while (rfte++, addr = next, addr < end);
614 return 0;
615}
616
617/**
618 * base_asce_free - free asce and tables returned from base_asce_alloc()
619 * @asce: asce to be freed
620 *
621 * Frees all region, segment, and page tables that were allocated with a
622 * corresponding base_asce_alloc() call.
623 */
624void base_asce_free(unsigned long asce)
625{
626 unsigned long *table = __va(asce & _ASCE_ORIGIN);
627
628 if (!asce)
629 return;
630 switch (asce & _ASCE_TYPE_MASK) {
631 case _ASCE_TYPE_SEGMENT:
632 base_segment_walk(table, 0, _REGION3_SIZE, 0);
633 break;
634 case _ASCE_TYPE_REGION3:
635 base_region3_walk(table, 0, _REGION2_SIZE, 0);
636 break;
637 case _ASCE_TYPE_REGION2:
638 base_region2_walk(table, 0, _REGION1_SIZE, 0);
639 break;
640 case _ASCE_TYPE_REGION1:
641 base_region1_walk(table, 0, TASK_SIZE_MAX, 0);
642 break;
643 }
644 base_crst_free(table);
645}
646
647static int base_pgt_cache_init(void)
648{
649 static DEFINE_MUTEX(base_pgt_cache_mutex);
650 unsigned long sz = _PAGE_TABLE_SIZE;
651
652 if (base_pgt_cache)
653 return 0;
654 mutex_lock(&base_pgt_cache_mutex);
655 if (!base_pgt_cache)
656 base_pgt_cache = kmem_cache_create("base_pgt", sz, sz, 0, NULL);
657 mutex_unlock(&base_pgt_cache_mutex);
658 return base_pgt_cache ? 0 : -ENOMEM;
659}
660
661/**
662 * base_asce_alloc - create kernel mapping without enhanced DAT features
663 * @addr: virtual start address of kernel mapping
664 * @num_pages: number of consecutive pages
665 *
666 * Generate an asce, including all required region, segment and page tables,
667 * that can be used to access the virtual kernel mapping. The difference is
668 * that the returned asce does not make use of any enhanced DAT features like
669 * e.g. large pages. This is required for some I/O functions that pass an
670 * asce, like e.g. some service call requests.
671 *
672 * Note: the returned asce may NEVER be attached to any cpu. It may only be
673 * used for I/O requests. tlb entries that might result because the
674 * asce was attached to a cpu won't be cleared.
675 */
676unsigned long base_asce_alloc(unsigned long addr, unsigned long num_pages)
677{
678 unsigned long asce, *table, end;
679 int rc;
680
681 if (base_pgt_cache_init())
682 return 0;
683 end = addr + num_pages * PAGE_SIZE;
684 if (end <= _REGION3_SIZE) {
685 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY);
686 if (!table)
687 return 0;
688 rc = base_segment_walk(table, addr, end, 1);
689 asce = __pa(table) | _ASCE_TYPE_SEGMENT | _ASCE_TABLE_LENGTH;
690 } else if (end <= _REGION2_SIZE) {
691 table = base_crst_alloc(_REGION3_ENTRY_EMPTY);
692 if (!table)
693 return 0;
694 rc = base_region3_walk(table, addr, end, 1);
695 asce = __pa(table) | _ASCE_TYPE_REGION3 | _ASCE_TABLE_LENGTH;
696 } else if (end <= _REGION1_SIZE) {
697 table = base_crst_alloc(_REGION2_ENTRY_EMPTY);
698 if (!table)
699 return 0;
700 rc = base_region2_walk(table, addr, end, 1);
701 asce = __pa(table) | _ASCE_TYPE_REGION2 | _ASCE_TABLE_LENGTH;
702 } else {
703 table = base_crst_alloc(_REGION1_ENTRY_EMPTY);
704 if (!table)
705 return 0;
706 rc = base_region1_walk(table, addr, end, 1);
707 asce = __pa(table) | _ASCE_TYPE_REGION1 | _ASCE_TABLE_LENGTH;
708 }
709 if (rc) {
710 base_asce_free(asce);
711 asce = 0;
712 }
713 return asce;
714}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Page table allocation functions
4 *
5 * Copyright IBM Corp. 2016
6 * Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
7 */
8
9#include <linux/sysctl.h>
10#include <linux/slab.h>
11#include <linux/mm.h>
12#include <asm/mmu_context.h>
13#include <asm/page-states.h>
14#include <asm/pgalloc.h>
15#include <asm/gmap.h>
16#include <asm/tlb.h>
17#include <asm/tlbflush.h>
18
19#ifdef CONFIG_PGSTE
20
21int page_table_allocate_pgste = 0;
22EXPORT_SYMBOL(page_table_allocate_pgste);
23
24static struct ctl_table page_table_sysctl[] = {
25 {
26 .procname = "allocate_pgste",
27 .data = &page_table_allocate_pgste,
28 .maxlen = sizeof(int),
29 .mode = S_IRUGO | S_IWUSR,
30 .proc_handler = proc_dointvec_minmax,
31 .extra1 = SYSCTL_ZERO,
32 .extra2 = SYSCTL_ONE,
33 },
34};
35
36static int __init page_table_register_sysctl(void)
37{
38 return register_sysctl("vm", page_table_sysctl) ? 0 : -ENOMEM;
39}
40__initcall(page_table_register_sysctl);
41
42#endif /* CONFIG_PGSTE */
43
44unsigned long *crst_table_alloc(struct mm_struct *mm)
45{
46 struct ptdesc *ptdesc = pagetable_alloc(GFP_KERNEL, CRST_ALLOC_ORDER);
47 unsigned long *table;
48
49 if (!ptdesc)
50 return NULL;
51 table = ptdesc_to_virt(ptdesc);
52 __arch_set_page_dat(table, 1UL << CRST_ALLOC_ORDER);
53 return table;
54}
55
56void crst_table_free(struct mm_struct *mm, unsigned long *table)
57{
58 pagetable_free(virt_to_ptdesc(table));
59}
60
61static void __crst_table_upgrade(void *arg)
62{
63 struct mm_struct *mm = arg;
64
65 /* change all active ASCEs to avoid the creation of new TLBs */
66 if (current->active_mm == mm) {
67 S390_lowcore.user_asce.val = mm->context.asce;
68 local_ctl_load(7, &S390_lowcore.user_asce);
69 }
70 __tlb_flush_local();
71}
72
73int crst_table_upgrade(struct mm_struct *mm, unsigned long end)
74{
75 unsigned long *pgd = NULL, *p4d = NULL, *__pgd;
76 unsigned long asce_limit = mm->context.asce_limit;
77
78 /* upgrade should only happen from 3 to 4, 3 to 5, or 4 to 5 levels */
79 VM_BUG_ON(asce_limit < _REGION2_SIZE);
80
81 if (end <= asce_limit)
82 return 0;
83
84 if (asce_limit == _REGION2_SIZE) {
85 p4d = crst_table_alloc(mm);
86 if (unlikely(!p4d))
87 goto err_p4d;
88 crst_table_init(p4d, _REGION2_ENTRY_EMPTY);
89 }
90 if (end > _REGION1_SIZE) {
91 pgd = crst_table_alloc(mm);
92 if (unlikely(!pgd))
93 goto err_pgd;
94 crst_table_init(pgd, _REGION1_ENTRY_EMPTY);
95 }
96
97 spin_lock_bh(&mm->page_table_lock);
98
99 /*
100 * This routine gets called with mmap_lock lock held and there is
101 * no reason to optimize for the case of otherwise. However, if
102 * that would ever change, the below check will let us know.
103 */
104 VM_BUG_ON(asce_limit != mm->context.asce_limit);
105
106 if (p4d) {
107 __pgd = (unsigned long *) mm->pgd;
108 p4d_populate(mm, (p4d_t *) p4d, (pud_t *) __pgd);
109 mm->pgd = (pgd_t *) p4d;
110 mm->context.asce_limit = _REGION1_SIZE;
111 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
112 _ASCE_USER_BITS | _ASCE_TYPE_REGION2;
113 mm_inc_nr_puds(mm);
114 }
115 if (pgd) {
116 __pgd = (unsigned long *) mm->pgd;
117 pgd_populate(mm, (pgd_t *) pgd, (p4d_t *) __pgd);
118 mm->pgd = (pgd_t *) pgd;
119 mm->context.asce_limit = TASK_SIZE_MAX;
120 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
121 _ASCE_USER_BITS | _ASCE_TYPE_REGION1;
122 }
123
124 spin_unlock_bh(&mm->page_table_lock);
125
126 on_each_cpu(__crst_table_upgrade, mm, 0);
127
128 return 0;
129
130err_pgd:
131 crst_table_free(mm, p4d);
132err_p4d:
133 return -ENOMEM;
134}
135
136#ifdef CONFIG_PGSTE
137
138struct ptdesc *page_table_alloc_pgste(struct mm_struct *mm)
139{
140 struct ptdesc *ptdesc;
141 u64 *table;
142
143 ptdesc = pagetable_alloc(GFP_KERNEL, 0);
144 if (ptdesc) {
145 table = (u64 *)ptdesc_to_virt(ptdesc);
146 __arch_set_page_dat(table, 1);
147 memset64(table, _PAGE_INVALID, PTRS_PER_PTE);
148 memset64(table + PTRS_PER_PTE, 0, PTRS_PER_PTE);
149 }
150 return ptdesc;
151}
152
153void page_table_free_pgste(struct ptdesc *ptdesc)
154{
155 pagetable_free(ptdesc);
156}
157
158#endif /* CONFIG_PGSTE */
159
160unsigned long *page_table_alloc(struct mm_struct *mm)
161{
162 struct ptdesc *ptdesc;
163 unsigned long *table;
164
165 ptdesc = pagetable_alloc(GFP_KERNEL, 0);
166 if (!ptdesc)
167 return NULL;
168 if (!pagetable_pte_ctor(ptdesc)) {
169 pagetable_free(ptdesc);
170 return NULL;
171 }
172 table = ptdesc_to_virt(ptdesc);
173 __arch_set_page_dat(table, 1);
174 /* pt_list is used by gmap only */
175 INIT_LIST_HEAD(&ptdesc->pt_list);
176 memset64((u64 *)table, _PAGE_INVALID, PTRS_PER_PTE);
177 memset64((u64 *)table + PTRS_PER_PTE, 0, PTRS_PER_PTE);
178 return table;
179}
180
181static void pagetable_pte_dtor_free(struct ptdesc *ptdesc)
182{
183 pagetable_pte_dtor(ptdesc);
184 pagetable_free(ptdesc);
185}
186
187void page_table_free(struct mm_struct *mm, unsigned long *table)
188{
189 struct ptdesc *ptdesc = virt_to_ptdesc(table);
190
191 pagetable_pte_dtor_free(ptdesc);
192}
193
194void __tlb_remove_table(void *table)
195{
196 struct ptdesc *ptdesc = virt_to_ptdesc(table);
197 struct page *page = ptdesc_page(ptdesc);
198
199 if (compound_order(page) == CRST_ALLOC_ORDER) {
200 /* pmd, pud, or p4d */
201 pagetable_free(ptdesc);
202 return;
203 }
204 pagetable_pte_dtor_free(ptdesc);
205}
206
207#ifdef CONFIG_TRANSPARENT_HUGEPAGE
208static void pte_free_now(struct rcu_head *head)
209{
210 struct ptdesc *ptdesc = container_of(head, struct ptdesc, pt_rcu_head);
211
212 pagetable_pte_dtor_free(ptdesc);
213}
214
215void pte_free_defer(struct mm_struct *mm, pgtable_t pgtable)
216{
217 struct ptdesc *ptdesc = virt_to_ptdesc(pgtable);
218
219 call_rcu(&ptdesc->pt_rcu_head, pte_free_now);
220 /*
221 * THPs are not allowed for KVM guests. Warn if pgste ever reaches here.
222 * Turn to the generic pte_free_defer() version once gmap is removed.
223 */
224 WARN_ON_ONCE(mm_has_pgste(mm));
225}
226#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
227
228/*
229 * Base infrastructure required to generate basic asces, region, segment,
230 * and page tables that do not make use of enhanced features like EDAT1.
231 */
232
233static struct kmem_cache *base_pgt_cache;
234
235static unsigned long *base_pgt_alloc(void)
236{
237 unsigned long *table;
238
239 table = kmem_cache_alloc(base_pgt_cache, GFP_KERNEL);
240 if (table)
241 memset64((u64 *)table, _PAGE_INVALID, PTRS_PER_PTE);
242 return table;
243}
244
245static void base_pgt_free(unsigned long *table)
246{
247 kmem_cache_free(base_pgt_cache, table);
248}
249
250static unsigned long *base_crst_alloc(unsigned long val)
251{
252 unsigned long *table;
253 struct ptdesc *ptdesc;
254
255 ptdesc = pagetable_alloc(GFP_KERNEL, CRST_ALLOC_ORDER);
256 if (!ptdesc)
257 return NULL;
258 table = ptdesc_address(ptdesc);
259 crst_table_init(table, val);
260 return table;
261}
262
263static void base_crst_free(unsigned long *table)
264{
265 pagetable_free(virt_to_ptdesc(table));
266}
267
268#define BASE_ADDR_END_FUNC(NAME, SIZE) \
269static inline unsigned long base_##NAME##_addr_end(unsigned long addr, \
270 unsigned long end) \
271{ \
272 unsigned long next = (addr + (SIZE)) & ~((SIZE) - 1); \
273 \
274 return (next - 1) < (end - 1) ? next : end; \
275}
276
277BASE_ADDR_END_FUNC(page, _PAGE_SIZE)
278BASE_ADDR_END_FUNC(segment, _SEGMENT_SIZE)
279BASE_ADDR_END_FUNC(region3, _REGION3_SIZE)
280BASE_ADDR_END_FUNC(region2, _REGION2_SIZE)
281BASE_ADDR_END_FUNC(region1, _REGION1_SIZE)
282
283static inline unsigned long base_lra(unsigned long address)
284{
285 unsigned long real;
286
287 asm volatile(
288 " lra %0,0(%1)\n"
289 : "=d" (real) : "a" (address) : "cc");
290 return real;
291}
292
293static int base_page_walk(unsigned long *origin, unsigned long addr,
294 unsigned long end, int alloc)
295{
296 unsigned long *pte, next;
297
298 if (!alloc)
299 return 0;
300 pte = origin;
301 pte += (addr & _PAGE_INDEX) >> _PAGE_SHIFT;
302 do {
303 next = base_page_addr_end(addr, end);
304 *pte = base_lra(addr);
305 } while (pte++, addr = next, addr < end);
306 return 0;
307}
308
309static int base_segment_walk(unsigned long *origin, unsigned long addr,
310 unsigned long end, int alloc)
311{
312 unsigned long *ste, next, *table;
313 int rc;
314
315 ste = origin;
316 ste += (addr & _SEGMENT_INDEX) >> _SEGMENT_SHIFT;
317 do {
318 next = base_segment_addr_end(addr, end);
319 if (*ste & _SEGMENT_ENTRY_INVALID) {
320 if (!alloc)
321 continue;
322 table = base_pgt_alloc();
323 if (!table)
324 return -ENOMEM;
325 *ste = __pa(table) | _SEGMENT_ENTRY;
326 }
327 table = __va(*ste & _SEGMENT_ENTRY_ORIGIN);
328 rc = base_page_walk(table, addr, next, alloc);
329 if (rc)
330 return rc;
331 if (!alloc)
332 base_pgt_free(table);
333 cond_resched();
334 } while (ste++, addr = next, addr < end);
335 return 0;
336}
337
338static int base_region3_walk(unsigned long *origin, unsigned long addr,
339 unsigned long end, int alloc)
340{
341 unsigned long *rtte, next, *table;
342 int rc;
343
344 rtte = origin;
345 rtte += (addr & _REGION3_INDEX) >> _REGION3_SHIFT;
346 do {
347 next = base_region3_addr_end(addr, end);
348 if (*rtte & _REGION_ENTRY_INVALID) {
349 if (!alloc)
350 continue;
351 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY);
352 if (!table)
353 return -ENOMEM;
354 *rtte = __pa(table) | _REGION3_ENTRY;
355 }
356 table = __va(*rtte & _REGION_ENTRY_ORIGIN);
357 rc = base_segment_walk(table, addr, next, alloc);
358 if (rc)
359 return rc;
360 if (!alloc)
361 base_crst_free(table);
362 } while (rtte++, addr = next, addr < end);
363 return 0;
364}
365
366static int base_region2_walk(unsigned long *origin, unsigned long addr,
367 unsigned long end, int alloc)
368{
369 unsigned long *rste, next, *table;
370 int rc;
371
372 rste = origin;
373 rste += (addr & _REGION2_INDEX) >> _REGION2_SHIFT;
374 do {
375 next = base_region2_addr_end(addr, end);
376 if (*rste & _REGION_ENTRY_INVALID) {
377 if (!alloc)
378 continue;
379 table = base_crst_alloc(_REGION3_ENTRY_EMPTY);
380 if (!table)
381 return -ENOMEM;
382 *rste = __pa(table) | _REGION2_ENTRY;
383 }
384 table = __va(*rste & _REGION_ENTRY_ORIGIN);
385 rc = base_region3_walk(table, addr, next, alloc);
386 if (rc)
387 return rc;
388 if (!alloc)
389 base_crst_free(table);
390 } while (rste++, addr = next, addr < end);
391 return 0;
392}
393
394static int base_region1_walk(unsigned long *origin, unsigned long addr,
395 unsigned long end, int alloc)
396{
397 unsigned long *rfte, next, *table;
398 int rc;
399
400 rfte = origin;
401 rfte += (addr & _REGION1_INDEX) >> _REGION1_SHIFT;
402 do {
403 next = base_region1_addr_end(addr, end);
404 if (*rfte & _REGION_ENTRY_INVALID) {
405 if (!alloc)
406 continue;
407 table = base_crst_alloc(_REGION2_ENTRY_EMPTY);
408 if (!table)
409 return -ENOMEM;
410 *rfte = __pa(table) | _REGION1_ENTRY;
411 }
412 table = __va(*rfte & _REGION_ENTRY_ORIGIN);
413 rc = base_region2_walk(table, addr, next, alloc);
414 if (rc)
415 return rc;
416 if (!alloc)
417 base_crst_free(table);
418 } while (rfte++, addr = next, addr < end);
419 return 0;
420}
421
422/**
423 * base_asce_free - free asce and tables returned from base_asce_alloc()
424 * @asce: asce to be freed
425 *
426 * Frees all region, segment, and page tables that were allocated with a
427 * corresponding base_asce_alloc() call.
428 */
429void base_asce_free(unsigned long asce)
430{
431 unsigned long *table = __va(asce & _ASCE_ORIGIN);
432
433 if (!asce)
434 return;
435 switch (asce & _ASCE_TYPE_MASK) {
436 case _ASCE_TYPE_SEGMENT:
437 base_segment_walk(table, 0, _REGION3_SIZE, 0);
438 break;
439 case _ASCE_TYPE_REGION3:
440 base_region3_walk(table, 0, _REGION2_SIZE, 0);
441 break;
442 case _ASCE_TYPE_REGION2:
443 base_region2_walk(table, 0, _REGION1_SIZE, 0);
444 break;
445 case _ASCE_TYPE_REGION1:
446 base_region1_walk(table, 0, TASK_SIZE_MAX, 0);
447 break;
448 }
449 base_crst_free(table);
450}
451
452static int base_pgt_cache_init(void)
453{
454 static DEFINE_MUTEX(base_pgt_cache_mutex);
455 unsigned long sz = _PAGE_TABLE_SIZE;
456
457 if (base_pgt_cache)
458 return 0;
459 mutex_lock(&base_pgt_cache_mutex);
460 if (!base_pgt_cache)
461 base_pgt_cache = kmem_cache_create("base_pgt", sz, sz, 0, NULL);
462 mutex_unlock(&base_pgt_cache_mutex);
463 return base_pgt_cache ? 0 : -ENOMEM;
464}
465
466/**
467 * base_asce_alloc - create kernel mapping without enhanced DAT features
468 * @addr: virtual start address of kernel mapping
469 * @num_pages: number of consecutive pages
470 *
471 * Generate an asce, including all required region, segment and page tables,
472 * that can be used to access the virtual kernel mapping. The difference is
473 * that the returned asce does not make use of any enhanced DAT features like
474 * e.g. large pages. This is required for some I/O functions that pass an
475 * asce, like e.g. some service call requests.
476 *
477 * Note: the returned asce may NEVER be attached to any cpu. It may only be
478 * used for I/O requests. tlb entries that might result because the
479 * asce was attached to a cpu won't be cleared.
480 */
481unsigned long base_asce_alloc(unsigned long addr, unsigned long num_pages)
482{
483 unsigned long asce, *table, end;
484 int rc;
485
486 if (base_pgt_cache_init())
487 return 0;
488 end = addr + num_pages * PAGE_SIZE;
489 if (end <= _REGION3_SIZE) {
490 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY);
491 if (!table)
492 return 0;
493 rc = base_segment_walk(table, addr, end, 1);
494 asce = __pa(table) | _ASCE_TYPE_SEGMENT | _ASCE_TABLE_LENGTH;
495 } else if (end <= _REGION2_SIZE) {
496 table = base_crst_alloc(_REGION3_ENTRY_EMPTY);
497 if (!table)
498 return 0;
499 rc = base_region3_walk(table, addr, end, 1);
500 asce = __pa(table) | _ASCE_TYPE_REGION3 | _ASCE_TABLE_LENGTH;
501 } else if (end <= _REGION1_SIZE) {
502 table = base_crst_alloc(_REGION2_ENTRY_EMPTY);
503 if (!table)
504 return 0;
505 rc = base_region2_walk(table, addr, end, 1);
506 asce = __pa(table) | _ASCE_TYPE_REGION2 | _ASCE_TABLE_LENGTH;
507 } else {
508 table = base_crst_alloc(_REGION1_ENTRY_EMPTY);
509 if (!table)
510 return 0;
511 rc = base_region1_walk(table, addr, end, 1);
512 asce = __pa(table) | _ASCE_TYPE_REGION1 | _ASCE_TABLE_LENGTH;
513 }
514 if (rc) {
515 base_asce_free(asce);
516 asce = 0;
517 }
518 return asce;
519}