Loading...
Note: File does not exist in v4.17.
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * DAMON Primitives for Virtual Address Spaces
4 *
5 * Author: SeongJae Park <sj@kernel.org>
6 */
7
8#define pr_fmt(fmt) "damon-va: " fmt
9
10#include <linux/highmem.h>
11#include <linux/hugetlb.h>
12#include <linux/mman.h>
13#include <linux/mmu_notifier.h>
14#include <linux/page_idle.h>
15#include <linux/pagewalk.h>
16#include <linux/sched/mm.h>
17
18#include "ops-common.h"
19
20#ifdef CONFIG_DAMON_VADDR_KUNIT_TEST
21#undef DAMON_MIN_REGION
22#define DAMON_MIN_REGION 1
23#endif
24
25/*
26 * 't->pid' should be the pointer to the relevant 'struct pid' having reference
27 * count. Caller must put the returned task, unless it is NULL.
28 */
29static inline struct task_struct *damon_get_task_struct(struct damon_target *t)
30{
31 return get_pid_task(t->pid, PIDTYPE_PID);
32}
33
34/*
35 * Get the mm_struct of the given target
36 *
37 * Caller _must_ put the mm_struct after use, unless it is NULL.
38 *
39 * Returns the mm_struct of the target on success, NULL on failure
40 */
41static struct mm_struct *damon_get_mm(struct damon_target *t)
42{
43 struct task_struct *task;
44 struct mm_struct *mm;
45
46 task = damon_get_task_struct(t);
47 if (!task)
48 return NULL;
49
50 mm = get_task_mm(task);
51 put_task_struct(task);
52 return mm;
53}
54
55/*
56 * Functions for the initial monitoring target regions construction
57 */
58
59/*
60 * Size-evenly split a region into 'nr_pieces' small regions
61 *
62 * Returns 0 on success, or negative error code otherwise.
63 */
64static int damon_va_evenly_split_region(struct damon_target *t,
65 struct damon_region *r, unsigned int nr_pieces)
66{
67 unsigned long sz_orig, sz_piece, orig_end;
68 struct damon_region *n = NULL, *next;
69 unsigned long start;
70 unsigned int i;
71
72 if (!r || !nr_pieces)
73 return -EINVAL;
74
75 if (nr_pieces == 1)
76 return 0;
77
78 orig_end = r->ar.end;
79 sz_orig = damon_sz_region(r);
80 sz_piece = ALIGN_DOWN(sz_orig / nr_pieces, DAMON_MIN_REGION);
81
82 if (!sz_piece)
83 return -EINVAL;
84
85 r->ar.end = r->ar.start + sz_piece;
86 next = damon_next_region(r);
87 for (start = r->ar.end, i = 1; i < nr_pieces; start += sz_piece, i++) {
88 n = damon_new_region(start, start + sz_piece);
89 if (!n)
90 return -ENOMEM;
91 damon_insert_region(n, r, next, t);
92 r = n;
93 }
94 /* complement last region for possible rounding error */
95 if (n)
96 n->ar.end = orig_end;
97
98 return 0;
99}
100
101static unsigned long sz_range(struct damon_addr_range *r)
102{
103 return r->end - r->start;
104}
105
106/*
107 * Find three regions separated by two biggest unmapped regions
108 *
109 * vma the head vma of the target address space
110 * regions an array of three address ranges that results will be saved
111 *
112 * This function receives an address space and finds three regions in it which
113 * separated by the two biggest unmapped regions in the space. Please refer to
114 * below comments of '__damon_va_init_regions()' function to know why this is
115 * necessary.
116 *
117 * Returns 0 if success, or negative error code otherwise.
118 */
119static int __damon_va_three_regions(struct mm_struct *mm,
120 struct damon_addr_range regions[3])
121{
122 struct damon_addr_range first_gap = {0}, second_gap = {0};
123 VMA_ITERATOR(vmi, mm, 0);
124 struct vm_area_struct *vma, *prev = NULL;
125 unsigned long start;
126
127 /*
128 * Find the two biggest gaps so that first_gap > second_gap > others.
129 * If this is too slow, it can be optimised to examine the maple
130 * tree gaps.
131 */
132 rcu_read_lock();
133 for_each_vma(vmi, vma) {
134 unsigned long gap;
135
136 if (!prev) {
137 start = vma->vm_start;
138 goto next;
139 }
140 gap = vma->vm_start - prev->vm_end;
141
142 if (gap > sz_range(&first_gap)) {
143 second_gap = first_gap;
144 first_gap.start = prev->vm_end;
145 first_gap.end = vma->vm_start;
146 } else if (gap > sz_range(&second_gap)) {
147 second_gap.start = prev->vm_end;
148 second_gap.end = vma->vm_start;
149 }
150next:
151 prev = vma;
152 }
153 rcu_read_unlock();
154
155 if (!sz_range(&second_gap) || !sz_range(&first_gap))
156 return -EINVAL;
157
158 /* Sort the two biggest gaps by address */
159 if (first_gap.start > second_gap.start)
160 swap(first_gap, second_gap);
161
162 /* Store the result */
163 regions[0].start = ALIGN(start, DAMON_MIN_REGION);
164 regions[0].end = ALIGN(first_gap.start, DAMON_MIN_REGION);
165 regions[1].start = ALIGN(first_gap.end, DAMON_MIN_REGION);
166 regions[1].end = ALIGN(second_gap.start, DAMON_MIN_REGION);
167 regions[2].start = ALIGN(second_gap.end, DAMON_MIN_REGION);
168 regions[2].end = ALIGN(prev->vm_end, DAMON_MIN_REGION);
169
170 return 0;
171}
172
173/*
174 * Get the three regions in the given target (task)
175 *
176 * Returns 0 on success, negative error code otherwise.
177 */
178static int damon_va_three_regions(struct damon_target *t,
179 struct damon_addr_range regions[3])
180{
181 struct mm_struct *mm;
182 int rc;
183
184 mm = damon_get_mm(t);
185 if (!mm)
186 return -EINVAL;
187
188 mmap_read_lock(mm);
189 rc = __damon_va_three_regions(mm, regions);
190 mmap_read_unlock(mm);
191
192 mmput(mm);
193 return rc;
194}
195
196/*
197 * Initialize the monitoring target regions for the given target (task)
198 *
199 * t the given target
200 *
201 * Because only a number of small portions of the entire address space
202 * is actually mapped to the memory and accessed, monitoring the unmapped
203 * regions is wasteful. That said, because we can deal with small noises,
204 * tracking every mapping is not strictly required but could even incur a high
205 * overhead if the mapping frequently changes or the number of mappings is
206 * high. The adaptive regions adjustment mechanism will further help to deal
207 * with the noise by simply identifying the unmapped areas as a region that
208 * has no access. Moreover, applying the real mappings that would have many
209 * unmapped areas inside will make the adaptive mechanism quite complex. That
210 * said, too huge unmapped areas inside the monitoring target should be removed
211 * to not take the time for the adaptive mechanism.
212 *
213 * For the reason, we convert the complex mappings to three distinct regions
214 * that cover every mapped area of the address space. Also the two gaps
215 * between the three regions are the two biggest unmapped areas in the given
216 * address space. In detail, this function first identifies the start and the
217 * end of the mappings and the two biggest unmapped areas of the address space.
218 * Then, it constructs the three regions as below:
219 *
220 * [mappings[0]->start, big_two_unmapped_areas[0]->start)
221 * [big_two_unmapped_areas[0]->end, big_two_unmapped_areas[1]->start)
222 * [big_two_unmapped_areas[1]->end, mappings[nr_mappings - 1]->end)
223 *
224 * As usual memory map of processes is as below, the gap between the heap and
225 * the uppermost mmap()-ed region, and the gap between the lowermost mmap()-ed
226 * region and the stack will be two biggest unmapped regions. Because these
227 * gaps are exceptionally huge areas in usual address space, excluding these
228 * two biggest unmapped regions will be sufficient to make a trade-off.
229 *
230 * <heap>
231 * <BIG UNMAPPED REGION 1>
232 * <uppermost mmap()-ed region>
233 * (other mmap()-ed regions and small unmapped regions)
234 * <lowermost mmap()-ed region>
235 * <BIG UNMAPPED REGION 2>
236 * <stack>
237 */
238static void __damon_va_init_regions(struct damon_ctx *ctx,
239 struct damon_target *t)
240{
241 struct damon_target *ti;
242 struct damon_region *r;
243 struct damon_addr_range regions[3];
244 unsigned long sz = 0, nr_pieces;
245 int i, tidx = 0;
246
247 if (damon_va_three_regions(t, regions)) {
248 damon_for_each_target(ti, ctx) {
249 if (ti == t)
250 break;
251 tidx++;
252 }
253 pr_debug("Failed to get three regions of %dth target\n", tidx);
254 return;
255 }
256
257 for (i = 0; i < 3; i++)
258 sz += regions[i].end - regions[i].start;
259 if (ctx->attrs.min_nr_regions)
260 sz /= ctx->attrs.min_nr_regions;
261 if (sz < DAMON_MIN_REGION)
262 sz = DAMON_MIN_REGION;
263
264 /* Set the initial three regions of the target */
265 for (i = 0; i < 3; i++) {
266 r = damon_new_region(regions[i].start, regions[i].end);
267 if (!r) {
268 pr_err("%d'th init region creation failed\n", i);
269 return;
270 }
271 damon_add_region(r, t);
272
273 nr_pieces = (regions[i].end - regions[i].start) / sz;
274 damon_va_evenly_split_region(t, r, nr_pieces);
275 }
276}
277
278/* Initialize '->regions_list' of every target (task) */
279static void damon_va_init(struct damon_ctx *ctx)
280{
281 struct damon_target *t;
282
283 damon_for_each_target(t, ctx) {
284 /* the user may set the target regions as they want */
285 if (!damon_nr_regions(t))
286 __damon_va_init_regions(ctx, t);
287 }
288}
289
290/*
291 * Update regions for current memory mappings
292 */
293static void damon_va_update(struct damon_ctx *ctx)
294{
295 struct damon_addr_range three_regions[3];
296 struct damon_target *t;
297
298 damon_for_each_target(t, ctx) {
299 if (damon_va_three_regions(t, three_regions))
300 continue;
301 damon_set_regions(t, three_regions, 3);
302 }
303}
304
305static int damon_mkold_pmd_entry(pmd_t *pmd, unsigned long addr,
306 unsigned long next, struct mm_walk *walk)
307{
308 pte_t *pte;
309 pmd_t pmde;
310 spinlock_t *ptl;
311
312 if (pmd_trans_huge(pmdp_get(pmd))) {
313 ptl = pmd_lock(walk->mm, pmd);
314 pmde = pmdp_get(pmd);
315
316 if (!pmd_present(pmde)) {
317 spin_unlock(ptl);
318 return 0;
319 }
320
321 if (pmd_trans_huge(pmde)) {
322 damon_pmdp_mkold(pmd, walk->vma, addr);
323 spin_unlock(ptl);
324 return 0;
325 }
326 spin_unlock(ptl);
327 }
328
329 pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
330 if (!pte) {
331 walk->action = ACTION_AGAIN;
332 return 0;
333 }
334 if (!pte_present(ptep_get(pte)))
335 goto out;
336 damon_ptep_mkold(pte, walk->vma, addr);
337out:
338 pte_unmap_unlock(pte, ptl);
339 return 0;
340}
341
342#ifdef CONFIG_HUGETLB_PAGE
343static void damon_hugetlb_mkold(pte_t *pte, struct mm_struct *mm,
344 struct vm_area_struct *vma, unsigned long addr)
345{
346 bool referenced = false;
347 pte_t entry = huge_ptep_get(mm, addr, pte);
348 struct folio *folio = pfn_folio(pte_pfn(entry));
349 unsigned long psize = huge_page_size(hstate_vma(vma));
350
351 folio_get(folio);
352
353 if (pte_young(entry)) {
354 referenced = true;
355 entry = pte_mkold(entry);
356 set_huge_pte_at(mm, addr, pte, entry, psize);
357 }
358
359 if (mmu_notifier_clear_young(mm, addr,
360 addr + huge_page_size(hstate_vma(vma))))
361 referenced = true;
362
363 if (referenced)
364 folio_set_young(folio);
365
366 folio_set_idle(folio);
367 folio_put(folio);
368}
369
370static int damon_mkold_hugetlb_entry(pte_t *pte, unsigned long hmask,
371 unsigned long addr, unsigned long end,
372 struct mm_walk *walk)
373{
374 struct hstate *h = hstate_vma(walk->vma);
375 spinlock_t *ptl;
376 pte_t entry;
377
378 ptl = huge_pte_lock(h, walk->mm, pte);
379 entry = huge_ptep_get(walk->mm, addr, pte);
380 if (!pte_present(entry))
381 goto out;
382
383 damon_hugetlb_mkold(pte, walk->mm, walk->vma, addr);
384
385out:
386 spin_unlock(ptl);
387 return 0;
388}
389#else
390#define damon_mkold_hugetlb_entry NULL
391#endif /* CONFIG_HUGETLB_PAGE */
392
393static const struct mm_walk_ops damon_mkold_ops = {
394 .pmd_entry = damon_mkold_pmd_entry,
395 .hugetlb_entry = damon_mkold_hugetlb_entry,
396 .walk_lock = PGWALK_RDLOCK,
397};
398
399static void damon_va_mkold(struct mm_struct *mm, unsigned long addr)
400{
401 mmap_read_lock(mm);
402 walk_page_range(mm, addr, addr + 1, &damon_mkold_ops, NULL);
403 mmap_read_unlock(mm);
404}
405
406/*
407 * Functions for the access checking of the regions
408 */
409
410static void __damon_va_prepare_access_check(struct mm_struct *mm,
411 struct damon_region *r)
412{
413 r->sampling_addr = damon_rand(r->ar.start, r->ar.end);
414
415 damon_va_mkold(mm, r->sampling_addr);
416}
417
418static void damon_va_prepare_access_checks(struct damon_ctx *ctx)
419{
420 struct damon_target *t;
421 struct mm_struct *mm;
422 struct damon_region *r;
423
424 damon_for_each_target(t, ctx) {
425 mm = damon_get_mm(t);
426 if (!mm)
427 continue;
428 damon_for_each_region(r, t)
429 __damon_va_prepare_access_check(mm, r);
430 mmput(mm);
431 }
432}
433
434struct damon_young_walk_private {
435 /* size of the folio for the access checked virtual memory address */
436 unsigned long *folio_sz;
437 bool young;
438};
439
440static int damon_young_pmd_entry(pmd_t *pmd, unsigned long addr,
441 unsigned long next, struct mm_walk *walk)
442{
443 pte_t *pte;
444 pte_t ptent;
445 spinlock_t *ptl;
446 struct folio *folio;
447 struct damon_young_walk_private *priv = walk->private;
448
449#ifdef CONFIG_TRANSPARENT_HUGEPAGE
450 if (pmd_trans_huge(pmdp_get(pmd))) {
451 pmd_t pmde;
452
453 ptl = pmd_lock(walk->mm, pmd);
454 pmde = pmdp_get(pmd);
455
456 if (!pmd_present(pmde)) {
457 spin_unlock(ptl);
458 return 0;
459 }
460
461 if (!pmd_trans_huge(pmde)) {
462 spin_unlock(ptl);
463 goto regular_page;
464 }
465 folio = damon_get_folio(pmd_pfn(pmde));
466 if (!folio)
467 goto huge_out;
468 if (pmd_young(pmde) || !folio_test_idle(folio) ||
469 mmu_notifier_test_young(walk->mm,
470 addr))
471 priv->young = true;
472 *priv->folio_sz = HPAGE_PMD_SIZE;
473 folio_put(folio);
474huge_out:
475 spin_unlock(ptl);
476 return 0;
477 }
478
479regular_page:
480#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
481
482 pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
483 if (!pte) {
484 walk->action = ACTION_AGAIN;
485 return 0;
486 }
487 ptent = ptep_get(pte);
488 if (!pte_present(ptent))
489 goto out;
490 folio = damon_get_folio(pte_pfn(ptent));
491 if (!folio)
492 goto out;
493 if (pte_young(ptent) || !folio_test_idle(folio) ||
494 mmu_notifier_test_young(walk->mm, addr))
495 priv->young = true;
496 *priv->folio_sz = folio_size(folio);
497 folio_put(folio);
498out:
499 pte_unmap_unlock(pte, ptl);
500 return 0;
501}
502
503#ifdef CONFIG_HUGETLB_PAGE
504static int damon_young_hugetlb_entry(pte_t *pte, unsigned long hmask,
505 unsigned long addr, unsigned long end,
506 struct mm_walk *walk)
507{
508 struct damon_young_walk_private *priv = walk->private;
509 struct hstate *h = hstate_vma(walk->vma);
510 struct folio *folio;
511 spinlock_t *ptl;
512 pte_t entry;
513
514 ptl = huge_pte_lock(h, walk->mm, pte);
515 entry = huge_ptep_get(walk->mm, addr, pte);
516 if (!pte_present(entry))
517 goto out;
518
519 folio = pfn_folio(pte_pfn(entry));
520 folio_get(folio);
521
522 if (pte_young(entry) || !folio_test_idle(folio) ||
523 mmu_notifier_test_young(walk->mm, addr))
524 priv->young = true;
525 *priv->folio_sz = huge_page_size(h);
526
527 folio_put(folio);
528
529out:
530 spin_unlock(ptl);
531 return 0;
532}
533#else
534#define damon_young_hugetlb_entry NULL
535#endif /* CONFIG_HUGETLB_PAGE */
536
537static const struct mm_walk_ops damon_young_ops = {
538 .pmd_entry = damon_young_pmd_entry,
539 .hugetlb_entry = damon_young_hugetlb_entry,
540 .walk_lock = PGWALK_RDLOCK,
541};
542
543static bool damon_va_young(struct mm_struct *mm, unsigned long addr,
544 unsigned long *folio_sz)
545{
546 struct damon_young_walk_private arg = {
547 .folio_sz = folio_sz,
548 .young = false,
549 };
550
551 mmap_read_lock(mm);
552 walk_page_range(mm, addr, addr + 1, &damon_young_ops, &arg);
553 mmap_read_unlock(mm);
554 return arg.young;
555}
556
557/*
558 * Check whether the region was accessed after the last preparation
559 *
560 * mm 'mm_struct' for the given virtual address space
561 * r the region to be checked
562 */
563static void __damon_va_check_access(struct mm_struct *mm,
564 struct damon_region *r, bool same_target,
565 struct damon_attrs *attrs)
566{
567 static unsigned long last_addr;
568 static unsigned long last_folio_sz = PAGE_SIZE;
569 static bool last_accessed;
570
571 if (!mm) {
572 damon_update_region_access_rate(r, false, attrs);
573 return;
574 }
575
576 /* If the region is in the last checked page, reuse the result */
577 if (same_target && (ALIGN_DOWN(last_addr, last_folio_sz) ==
578 ALIGN_DOWN(r->sampling_addr, last_folio_sz))) {
579 damon_update_region_access_rate(r, last_accessed, attrs);
580 return;
581 }
582
583 last_accessed = damon_va_young(mm, r->sampling_addr, &last_folio_sz);
584 damon_update_region_access_rate(r, last_accessed, attrs);
585
586 last_addr = r->sampling_addr;
587}
588
589static unsigned int damon_va_check_accesses(struct damon_ctx *ctx)
590{
591 struct damon_target *t;
592 struct mm_struct *mm;
593 struct damon_region *r;
594 unsigned int max_nr_accesses = 0;
595 bool same_target;
596
597 damon_for_each_target(t, ctx) {
598 mm = damon_get_mm(t);
599 same_target = false;
600 damon_for_each_region(r, t) {
601 __damon_va_check_access(mm, r, same_target,
602 &ctx->attrs);
603 max_nr_accesses = max(r->nr_accesses, max_nr_accesses);
604 same_target = true;
605 }
606 if (mm)
607 mmput(mm);
608 }
609
610 return max_nr_accesses;
611}
612
613/*
614 * Functions for the target validity check and cleanup
615 */
616
617static bool damon_va_target_valid(struct damon_target *t)
618{
619 struct task_struct *task;
620
621 task = damon_get_task_struct(t);
622 if (task) {
623 put_task_struct(task);
624 return true;
625 }
626
627 return false;
628}
629
630#ifndef CONFIG_ADVISE_SYSCALLS
631static unsigned long damos_madvise(struct damon_target *target,
632 struct damon_region *r, int behavior)
633{
634 return 0;
635}
636#else
637static unsigned long damos_madvise(struct damon_target *target,
638 struct damon_region *r, int behavior)
639{
640 struct mm_struct *mm;
641 unsigned long start = PAGE_ALIGN(r->ar.start);
642 unsigned long len = PAGE_ALIGN(damon_sz_region(r));
643 unsigned long applied;
644
645 mm = damon_get_mm(target);
646 if (!mm)
647 return 0;
648
649 applied = do_madvise(mm, start, len, behavior) ? 0 : len;
650 mmput(mm);
651
652 return applied;
653}
654#endif /* CONFIG_ADVISE_SYSCALLS */
655
656static unsigned long damon_va_apply_scheme(struct damon_ctx *ctx,
657 struct damon_target *t, struct damon_region *r,
658 struct damos *scheme)
659{
660 int madv_action;
661
662 switch (scheme->action) {
663 case DAMOS_WILLNEED:
664 madv_action = MADV_WILLNEED;
665 break;
666 case DAMOS_COLD:
667 madv_action = MADV_COLD;
668 break;
669 case DAMOS_PAGEOUT:
670 madv_action = MADV_PAGEOUT;
671 break;
672 case DAMOS_HUGEPAGE:
673 madv_action = MADV_HUGEPAGE;
674 break;
675 case DAMOS_NOHUGEPAGE:
676 madv_action = MADV_NOHUGEPAGE;
677 break;
678 case DAMOS_STAT:
679 return 0;
680 default:
681 /*
682 * DAMOS actions that are not yet supported by 'vaddr'.
683 */
684 return 0;
685 }
686
687 return damos_madvise(t, r, madv_action);
688}
689
690static int damon_va_scheme_score(struct damon_ctx *context,
691 struct damon_target *t, struct damon_region *r,
692 struct damos *scheme)
693{
694
695 switch (scheme->action) {
696 case DAMOS_PAGEOUT:
697 return damon_cold_score(context, r, scheme);
698 default:
699 break;
700 }
701
702 return DAMOS_MAX_SCORE;
703}
704
705static int __init damon_va_initcall(void)
706{
707 struct damon_operations ops = {
708 .id = DAMON_OPS_VADDR,
709 .init = damon_va_init,
710 .update = damon_va_update,
711 .prepare_access_checks = damon_va_prepare_access_checks,
712 .check_accesses = damon_va_check_accesses,
713 .reset_aggregated = NULL,
714 .target_valid = damon_va_target_valid,
715 .cleanup = NULL,
716 .apply_scheme = damon_va_apply_scheme,
717 .get_scheme_score = damon_va_scheme_score,
718 };
719 /* ops for fixed virtual address ranges */
720 struct damon_operations ops_fvaddr = ops;
721 int err;
722
723 /* Don't set the monitoring target regions for the entire mapping */
724 ops_fvaddr.id = DAMON_OPS_FVADDR;
725 ops_fvaddr.init = NULL;
726 ops_fvaddr.update = NULL;
727
728 err = damon_register_ops(&ops);
729 if (err)
730 return err;
731 return damon_register_ops(&ops_fvaddr);
732};
733
734subsys_initcall(damon_va_initcall);
735
736#include "tests/vaddr-kunit.h"