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/pgalloc.h>
14#include <asm/gmap.h>
15#include <asm/tlb.h>
16#include <asm/tlbflush.h>
17
18#ifdef CONFIG_PGSTE
19
20static int page_table_allocate_pgste_min = 0;
21static int page_table_allocate_pgste_max = 1;
22int page_table_allocate_pgste = 0;
23EXPORT_SYMBOL(page_table_allocate_pgste);
24
25static struct ctl_table page_table_sysctl[] = {
26 {
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,
32 .extra1 = &page_table_allocate_pgste_min,
33 .extra2 = &page_table_allocate_pgste_max,
34 },
35 { }
36};
37
38static struct ctl_table page_table_sysctl_dir[] = {
39 {
40 .procname = "vm",
41 .maxlen = 0,
42 .mode = 0555,
43 .child = page_table_sysctl,
44 },
45 { }
46};
47
48static int __init page_table_register_sysctl(void)
49{
50 return register_sysctl_table(page_table_sysctl_dir) ? 0 : -ENOMEM;
51}
52__initcall(page_table_register_sysctl);
53
54#endif /* CONFIG_PGSTE */
55
56unsigned long *crst_table_alloc(struct mm_struct *mm)
57{
58 struct page *page = alloc_pages(GFP_KERNEL, 2);
59
60 if (!page)
61 return NULL;
62 arch_set_page_dat(page, 2);
63 return (unsigned long *) page_to_phys(page);
64}
65
66void crst_table_free(struct mm_struct *mm, unsigned long *table)
67{
68 free_pages((unsigned long) table, 2);
69}
70
71static void __crst_table_upgrade(void *arg)
72{
73 struct mm_struct *mm = arg;
74
75 if (current->active_mm == mm)
76 set_user_asce(mm);
77 __tlb_flush_local();
78}
79
80int crst_table_upgrade(struct mm_struct *mm, unsigned long end)
81{
82 unsigned long *table, *pgd;
83 int rc, notify;
84
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);
87 rc = 0;
88 notify = 0;
89 while (mm->context.asce_limit < end) {
90 table = crst_table_alloc(mm);
91 if (!table) {
92 rc = -ENOMEM;
93 break;
94 }
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;
104 } else {
105 crst_table_init(table, _REGION1_ENTRY_EMPTY);
106 pgd_populate(mm, (pgd_t *) table, (p4d_t *) pgd);
107 mm->pgd = (pgd_t *) table;
108 mm->context.asce_limit = -PAGE_SIZE;
109 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
110 _ASCE_USER_BITS | _ASCE_TYPE_REGION1;
111 }
112 notify = 1;
113 spin_unlock_bh(&mm->page_table_lock);
114 }
115 if (notify)
116 on_each_cpu(__crst_table_upgrade, mm, 0);
117 return rc;
118}
119
120void crst_table_downgrade(struct mm_struct *mm)
121{
122 pgd_t *pgd;
123
124 /* downgrade should only happen from 3 to 2 levels (compat only) */
125 VM_BUG_ON(mm->context.asce_limit != _REGION2_SIZE);
126
127 if (current->active_mm == mm) {
128 clear_user_asce();
129 __tlb_flush_mm(mm);
130 }
131
132 pgd = mm->pgd;
133 mm->pgd = (pgd_t *) (pgd_val(*pgd) & _REGION_ENTRY_ORIGIN);
134 mm->context.asce_limit = _REGION3_SIZE;
135 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
136 _ASCE_USER_BITS | _ASCE_TYPE_SEGMENT;
137 crst_table_free(mm, (unsigned long *) pgd);
138
139 if (current->active_mm == mm)
140 set_user_asce(mm);
141}
142
143static inline unsigned int atomic_xor_bits(atomic_t *v, unsigned int bits)
144{
145 unsigned int old, new;
146
147 do {
148 old = atomic_read(v);
149 new = old ^ bits;
150 } while (atomic_cmpxchg(v, old, new) != old);
151 return new;
152}
153
154#ifdef CONFIG_PGSTE
155
156struct page *page_table_alloc_pgste(struct mm_struct *mm)
157{
158 struct page *page;
159 u64 *table;
160
161 page = alloc_page(GFP_KERNEL);
162 if (page) {
163 table = (u64 *)page_to_phys(page);
164 memset64(table, _PAGE_INVALID, PTRS_PER_PTE);
165 memset64(table + PTRS_PER_PTE, 0, PTRS_PER_PTE);
166 }
167 return page;
168}
169
170void page_table_free_pgste(struct page *page)
171{
172 __free_page(page);
173}
174
175#endif /* CONFIG_PGSTE */
176
177/*
178 * page table entry allocation/free routines.
179 */
180unsigned long *page_table_alloc(struct mm_struct *mm)
181{
182 unsigned long *table;
183 struct page *page;
184 unsigned int mask, bit;
185
186 /* Try to get a fragment of a 4K page as a 2K page table */
187 if (!mm_alloc_pgste(mm)) {
188 table = NULL;
189 spin_lock_bh(&mm->context.lock);
190 if (!list_empty(&mm->context.pgtable_list)) {
191 page = list_first_entry(&mm->context.pgtable_list,
192 struct page, lru);
193 mask = atomic_read(&page->_mapcount);
194 mask = (mask | (mask >> 4)) & 3;
195 if (mask != 3) {
196 table = (unsigned long *) page_to_phys(page);
197 bit = mask & 1; /* =1 -> second 2K */
198 if (bit)
199 table += PTRS_PER_PTE;
200 atomic_xor_bits(&page->_mapcount, 1U << bit);
201 list_del(&page->lru);
202 }
203 }
204 spin_unlock_bh(&mm->context.lock);
205 if (table)
206 return table;
207 }
208 /* Allocate a fresh page */
209 page = alloc_page(GFP_KERNEL);
210 if (!page)
211 return NULL;
212 if (!pgtable_page_ctor(page)) {
213 __free_page(page);
214 return NULL;
215 }
216 arch_set_page_dat(page, 0);
217 /* Initialize page table */
218 table = (unsigned long *) page_to_phys(page);
219 if (mm_alloc_pgste(mm)) {
220 /* Return 4K page table with PGSTEs */
221 atomic_set(&page->_mapcount, 3);
222 memset64((u64 *)table, _PAGE_INVALID, PTRS_PER_PTE);
223 memset64((u64 *)table + PTRS_PER_PTE, 0, PTRS_PER_PTE);
224 } else {
225 /* Return the first 2K fragment of the page */
226 atomic_set(&page->_mapcount, 1);
227 memset64((u64 *)table, _PAGE_INVALID, 2 * PTRS_PER_PTE);
228 spin_lock_bh(&mm->context.lock);
229 list_add(&page->lru, &mm->context.pgtable_list);
230 spin_unlock_bh(&mm->context.lock);
231 }
232 return table;
233}
234
235void page_table_free(struct mm_struct *mm, unsigned long *table)
236{
237 struct page *page;
238 unsigned int bit, mask;
239
240 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
241 if (!mm_alloc_pgste(mm)) {
242 /* Free 2K page table fragment of a 4K page */
243 bit = (__pa(table) & ~PAGE_MASK)/(PTRS_PER_PTE*sizeof(pte_t));
244 spin_lock_bh(&mm->context.lock);
245 mask = atomic_xor_bits(&page->_mapcount, 1U << bit);
246 if (mask & 3)
247 list_add(&page->lru, &mm->context.pgtable_list);
248 else
249 list_del(&page->lru);
250 spin_unlock_bh(&mm->context.lock);
251 if (mask != 0)
252 return;
253 }
254
255 pgtable_page_dtor(page);
256 atomic_set(&page->_mapcount, -1);
257 __free_page(page);
258}
259
260void page_table_free_rcu(struct mmu_gather *tlb, unsigned long *table,
261 unsigned long vmaddr)
262{
263 struct mm_struct *mm;
264 struct page *page;
265 unsigned int bit, mask;
266
267 mm = tlb->mm;
268 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
269 if (mm_alloc_pgste(mm)) {
270 gmap_unlink(mm, table, vmaddr);
271 table = (unsigned long *) (__pa(table) | 3);
272 tlb_remove_table(tlb, table);
273 return;
274 }
275 bit = (__pa(table) & ~PAGE_MASK) / (PTRS_PER_PTE*sizeof(pte_t));
276 spin_lock_bh(&mm->context.lock);
277 mask = atomic_xor_bits(&page->_mapcount, 0x11U << bit);
278 if (mask & 3)
279 list_add_tail(&page->lru, &mm->context.pgtable_list);
280 else
281 list_del(&page->lru);
282 spin_unlock_bh(&mm->context.lock);
283 table = (unsigned long *) (__pa(table) | (1U << bit));
284 tlb_remove_table(tlb, table);
285}
286
287static void __tlb_remove_table(void *_table)
288{
289 unsigned int mask = (unsigned long) _table & 3;
290 void *table = (void *)((unsigned long) _table ^ mask);
291 struct page *page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
292
293 switch (mask) {
294 case 0: /* pmd, pud, or p4d */
295 free_pages((unsigned long) table, 2);
296 break;
297 case 1: /* lower 2K of a 4K page table */
298 case 2: /* higher 2K of a 4K page table */
299 if (atomic_xor_bits(&page->_mapcount, mask << 4) != 0)
300 break;
301 /* fallthrough */
302 case 3: /* 4K page table with pgstes */
303 pgtable_page_dtor(page);
304 atomic_set(&page->_mapcount, -1);
305 __free_page(page);
306 break;
307 }
308}
309
310static void tlb_remove_table_smp_sync(void *arg)
311{
312 /* Simply deliver the interrupt */
313}
314
315static void tlb_remove_table_one(void *table)
316{
317 /*
318 * This isn't an RCU grace period and hence the page-tables cannot be
319 * assumed to be actually RCU-freed.
320 *
321 * It is however sufficient for software page-table walkers that rely
322 * on IRQ disabling. See the comment near struct mmu_table_batch.
323 */
324 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
325 __tlb_remove_table(table);
326}
327
328static void tlb_remove_table_rcu(struct rcu_head *head)
329{
330 struct mmu_table_batch *batch;
331 int i;
332
333 batch = container_of(head, struct mmu_table_batch, rcu);
334
335 for (i = 0; i < batch->nr; i++)
336 __tlb_remove_table(batch->tables[i]);
337
338 free_page((unsigned long)batch);
339}
340
341void tlb_table_flush(struct mmu_gather *tlb)
342{
343 struct mmu_table_batch **batch = &tlb->batch;
344
345 if (*batch) {
346 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
347 *batch = NULL;
348 }
349}
350
351void tlb_remove_table(struct mmu_gather *tlb, void *table)
352{
353 struct mmu_table_batch **batch = &tlb->batch;
354
355 tlb->mm->context.flush_mm = 1;
356 if (*batch == NULL) {
357 *batch = (struct mmu_table_batch *)
358 __get_free_page(GFP_NOWAIT | __GFP_NOWARN);
359 if (*batch == NULL) {
360 __tlb_flush_mm_lazy(tlb->mm);
361 tlb_remove_table_one(table);
362 return;
363 }
364 (*batch)->nr = 0;
365 }
366 (*batch)->tables[(*batch)->nr++] = table;
367 if ((*batch)->nr == MAX_TABLE_BATCH)
368 tlb_flush_mmu(tlb);
369}
370
371/*
372 * Base infrastructure required to generate basic asces, region, segment,
373 * and page tables that do not make use of enhanced features like EDAT1.
374 */
375
376static struct kmem_cache *base_pgt_cache;
377
378static unsigned long base_pgt_alloc(void)
379{
380 u64 *table;
381
382 table = kmem_cache_alloc(base_pgt_cache, GFP_KERNEL);
383 if (table)
384 memset64(table, _PAGE_INVALID, PTRS_PER_PTE);
385 return (unsigned long) table;
386}
387
388static void base_pgt_free(unsigned long table)
389{
390 kmem_cache_free(base_pgt_cache, (void *) table);
391}
392
393static unsigned long base_crst_alloc(unsigned long val)
394{
395 unsigned long table;
396
397 table = __get_free_pages(GFP_KERNEL, CRST_ALLOC_ORDER);
398 if (table)
399 crst_table_init((unsigned long *)table, val);
400 return table;
401}
402
403static void base_crst_free(unsigned long table)
404{
405 free_pages(table, CRST_ALLOC_ORDER);
406}
407
408#define BASE_ADDR_END_FUNC(NAME, SIZE) \
409static inline unsigned long base_##NAME##_addr_end(unsigned long addr, \
410 unsigned long end) \
411{ \
412 unsigned long next = (addr + (SIZE)) & ~((SIZE) - 1); \
413 \
414 return (next - 1) < (end - 1) ? next : end; \
415}
416
417BASE_ADDR_END_FUNC(page, _PAGE_SIZE)
418BASE_ADDR_END_FUNC(segment, _SEGMENT_SIZE)
419BASE_ADDR_END_FUNC(region3, _REGION3_SIZE)
420BASE_ADDR_END_FUNC(region2, _REGION2_SIZE)
421BASE_ADDR_END_FUNC(region1, _REGION1_SIZE)
422
423static inline unsigned long base_lra(unsigned long address)
424{
425 unsigned long real;
426
427 asm volatile(
428 " lra %0,0(%1)\n"
429 : "=d" (real) : "a" (address) : "cc");
430 return real;
431}
432
433static int base_page_walk(unsigned long origin, unsigned long addr,
434 unsigned long end, int alloc)
435{
436 unsigned long *pte, next;
437
438 if (!alloc)
439 return 0;
440 pte = (unsigned long *) origin;
441 pte += (addr & _PAGE_INDEX) >> _PAGE_SHIFT;
442 do {
443 next = base_page_addr_end(addr, end);
444 *pte = base_lra(addr);
445 } while (pte++, addr = next, addr < end);
446 return 0;
447}
448
449static int base_segment_walk(unsigned long origin, unsigned long addr,
450 unsigned long end, int alloc)
451{
452 unsigned long *ste, next, table;
453 int rc;
454
455 ste = (unsigned long *) origin;
456 ste += (addr & _SEGMENT_INDEX) >> _SEGMENT_SHIFT;
457 do {
458 next = base_segment_addr_end(addr, end);
459 if (*ste & _SEGMENT_ENTRY_INVALID) {
460 if (!alloc)
461 continue;
462 table = base_pgt_alloc();
463 if (!table)
464 return -ENOMEM;
465 *ste = table | _SEGMENT_ENTRY;
466 }
467 table = *ste & _SEGMENT_ENTRY_ORIGIN;
468 rc = base_page_walk(table, addr, next, alloc);
469 if (rc)
470 return rc;
471 if (!alloc)
472 base_pgt_free(table);
473 cond_resched();
474 } while (ste++, addr = next, addr < end);
475 return 0;
476}
477
478static int base_region3_walk(unsigned long origin, unsigned long addr,
479 unsigned long end, int alloc)
480{
481 unsigned long *rtte, next, table;
482 int rc;
483
484 rtte = (unsigned long *) origin;
485 rtte += (addr & _REGION3_INDEX) >> _REGION3_SHIFT;
486 do {
487 next = base_region3_addr_end(addr, end);
488 if (*rtte & _REGION_ENTRY_INVALID) {
489 if (!alloc)
490 continue;
491 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY);
492 if (!table)
493 return -ENOMEM;
494 *rtte = table | _REGION3_ENTRY;
495 }
496 table = *rtte & _REGION_ENTRY_ORIGIN;
497 rc = base_segment_walk(table, addr, next, alloc);
498 if (rc)
499 return rc;
500 if (!alloc)
501 base_crst_free(table);
502 } while (rtte++, addr = next, addr < end);
503 return 0;
504}
505
506static int base_region2_walk(unsigned long origin, unsigned long addr,
507 unsigned long end, int alloc)
508{
509 unsigned long *rste, next, table;
510 int rc;
511
512 rste = (unsigned long *) origin;
513 rste += (addr & _REGION2_INDEX) >> _REGION2_SHIFT;
514 do {
515 next = base_region2_addr_end(addr, end);
516 if (*rste & _REGION_ENTRY_INVALID) {
517 if (!alloc)
518 continue;
519 table = base_crst_alloc(_REGION3_ENTRY_EMPTY);
520 if (!table)
521 return -ENOMEM;
522 *rste = table | _REGION2_ENTRY;
523 }
524 table = *rste & _REGION_ENTRY_ORIGIN;
525 rc = base_region3_walk(table, addr, next, alloc);
526 if (rc)
527 return rc;
528 if (!alloc)
529 base_crst_free(table);
530 } while (rste++, addr = next, addr < end);
531 return 0;
532}
533
534static int base_region1_walk(unsigned long origin, unsigned long addr,
535 unsigned long end, int alloc)
536{
537 unsigned long *rfte, next, table;
538 int rc;
539
540 rfte = (unsigned long *) origin;
541 rfte += (addr & _REGION1_INDEX) >> _REGION1_SHIFT;
542 do {
543 next = base_region1_addr_end(addr, end);
544 if (*rfte & _REGION_ENTRY_INVALID) {
545 if (!alloc)
546 continue;
547 table = base_crst_alloc(_REGION2_ENTRY_EMPTY);
548 if (!table)
549 return -ENOMEM;
550 *rfte = table | _REGION1_ENTRY;
551 }
552 table = *rfte & _REGION_ENTRY_ORIGIN;
553 rc = base_region2_walk(table, addr, next, alloc);
554 if (rc)
555 return rc;
556 if (!alloc)
557 base_crst_free(table);
558 } while (rfte++, addr = next, addr < end);
559 return 0;
560}
561
562/**
563 * base_asce_free - free asce and tables returned from base_asce_alloc()
564 * @asce: asce to be freed
565 *
566 * Frees all region, segment, and page tables that were allocated with a
567 * corresponding base_asce_alloc() call.
568 */
569void base_asce_free(unsigned long asce)
570{
571 unsigned long table = asce & _ASCE_ORIGIN;
572
573 if (!asce)
574 return;
575 switch (asce & _ASCE_TYPE_MASK) {
576 case _ASCE_TYPE_SEGMENT:
577 base_segment_walk(table, 0, _REGION3_SIZE, 0);
578 break;
579 case _ASCE_TYPE_REGION3:
580 base_region3_walk(table, 0, _REGION2_SIZE, 0);
581 break;
582 case _ASCE_TYPE_REGION2:
583 base_region2_walk(table, 0, _REGION1_SIZE, 0);
584 break;
585 case _ASCE_TYPE_REGION1:
586 base_region1_walk(table, 0, -_PAGE_SIZE, 0);
587 break;
588 }
589 base_crst_free(table);
590}
591
592static int base_pgt_cache_init(void)
593{
594 static DEFINE_MUTEX(base_pgt_cache_mutex);
595 unsigned long sz = _PAGE_TABLE_SIZE;
596
597 if (base_pgt_cache)
598 return 0;
599 mutex_lock(&base_pgt_cache_mutex);
600 if (!base_pgt_cache)
601 base_pgt_cache = kmem_cache_create("base_pgt", sz, sz, 0, NULL);
602 mutex_unlock(&base_pgt_cache_mutex);
603 return base_pgt_cache ? 0 : -ENOMEM;
604}
605
606/**
607 * base_asce_alloc - create kernel mapping without enhanced DAT features
608 * @addr: virtual start address of kernel mapping
609 * @num_pages: number of consecutive pages
610 *
611 * Generate an asce, including all required region, segment and page tables,
612 * that can be used to access the virtual kernel mapping. The difference is
613 * that the returned asce does not make use of any enhanced DAT features like
614 * e.g. large pages. This is required for some I/O functions that pass an
615 * asce, like e.g. some service call requests.
616 *
617 * Note: the returned asce may NEVER be attached to any cpu. It may only be
618 * used for I/O requests. tlb entries that might result because the
619 * asce was attached to a cpu won't be cleared.
620 */
621unsigned long base_asce_alloc(unsigned long addr, unsigned long num_pages)
622{
623 unsigned long asce, table, end;
624 int rc;
625
626 if (base_pgt_cache_init())
627 return 0;
628 end = addr + num_pages * PAGE_SIZE;
629 if (end <= _REGION3_SIZE) {
630 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY);
631 if (!table)
632 return 0;
633 rc = base_segment_walk(table, addr, end, 1);
634 asce = table | _ASCE_TYPE_SEGMENT | _ASCE_TABLE_LENGTH;
635 } else if (end <= _REGION2_SIZE) {
636 table = base_crst_alloc(_REGION3_ENTRY_EMPTY);
637 if (!table)
638 return 0;
639 rc = base_region3_walk(table, addr, end, 1);
640 asce = table | _ASCE_TYPE_REGION3 | _ASCE_TABLE_LENGTH;
641 } else if (end <= _REGION1_SIZE) {
642 table = base_crst_alloc(_REGION2_ENTRY_EMPTY);
643 if (!table)
644 return 0;
645 rc = base_region2_walk(table, addr, end, 1);
646 asce = table | _ASCE_TYPE_REGION2 | _ASCE_TABLE_LENGTH;
647 } else {
648 table = base_crst_alloc(_REGION1_ENTRY_EMPTY);
649 if (!table)
650 return 0;
651 rc = base_region1_walk(table, addr, end, 1);
652 asce = table | _ASCE_TYPE_REGION1 | _ASCE_TABLE_LENGTH;
653 }
654 if (rc) {
655 base_asce_free(asce);
656 asce = 0;
657 }
658 return asce;
659}