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