Loading...
Note: File does not exist in v5.9.
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * DAMON Primitives for Virtual Address Spaces
4 *
5 * Author: SeongJae Park <sjpark@amazon.de>
6 */
7
8#define pr_fmt(fmt) "damon-va: " fmt
9
10#include <asm-generic/mman-common.h>
11#include <linux/highmem.h>
12#include <linux/hugetlb.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
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
98static 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 */
116static 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 for_each_vma(vmi, vma) {
130 unsigned long gap;
131
132 if (!prev) {
133 start = vma->vm_start;
134 goto next;
135 }
136 gap = vma->vm_start - prev->vm_end;
137
138 if (gap > sz_range(&first_gap)) {
139 second_gap = first_gap;
140 first_gap.start = prev->vm_end;
141 first_gap.end = vma->vm_start;
142 } else if (gap > sz_range(&second_gap)) {
143 second_gap.start = prev->vm_end;
144 second_gap.end = vma->vm_start;
145 }
146next:
147 prev = vma;
148 }
149
150 if (!sz_range(&second_gap) || !sz_range(&first_gap))
151 return -EINVAL;
152
153 /* Sort the two biggest gaps by address */
154 if (first_gap.start > second_gap.start)
155 swap(first_gap, second_gap);
156
157 /* Store the result */
158 regions[0].start = ALIGN(start, DAMON_MIN_REGION);
159 regions[0].end = ALIGN(first_gap.start, DAMON_MIN_REGION);
160 regions[1].start = ALIGN(first_gap.end, DAMON_MIN_REGION);
161 regions[1].end = ALIGN(second_gap.start, DAMON_MIN_REGION);
162 regions[2].start = ALIGN(second_gap.end, DAMON_MIN_REGION);
163 regions[2].end = ALIGN(prev->vm_end, DAMON_MIN_REGION);
164
165 return 0;
166}
167
168/*
169 * Get the three regions in the given target (task)
170 *
171 * Returns 0 on success, negative error code otherwise.
172 */
173static int damon_va_three_regions(struct damon_target *t,
174 struct damon_addr_range regions[3])
175{
176 struct mm_struct *mm;
177 int rc;
178
179 mm = damon_get_mm(t);
180 if (!mm)
181 return -EINVAL;
182
183 mmap_read_lock(mm);
184 rc = __damon_va_three_regions(mm, regions);
185 mmap_read_unlock(mm);
186
187 mmput(mm);
188 return rc;
189}
190
191/*
192 * Initialize the monitoring target regions for the given target (task)
193 *
194 * t the given target
195 *
196 * Because only a number of small portions of the entire address space
197 * is actually mapped to the memory and accessed, monitoring the unmapped
198 * regions is wasteful. That said, because we can deal with small noises,
199 * tracking every mapping is not strictly required but could even incur a high
200 * overhead if the mapping frequently changes or the number of mappings is
201 * high. The adaptive regions adjustment mechanism will further help to deal
202 * with the noise by simply identifying the unmapped areas as a region that
203 * has no access. Moreover, applying the real mappings that would have many
204 * unmapped areas inside will make the adaptive mechanism quite complex. That
205 * said, too huge unmapped areas inside the monitoring target should be removed
206 * to not take the time for the adaptive mechanism.
207 *
208 * For the reason, we convert the complex mappings to three distinct regions
209 * that cover every mapped area of the address space. Also the two gaps
210 * between the three regions are the two biggest unmapped areas in the given
211 * address space. In detail, this function first identifies the start and the
212 * end of the mappings and the two biggest unmapped areas of the address space.
213 * Then, it constructs the three regions as below:
214 *
215 * [mappings[0]->start, big_two_unmapped_areas[0]->start)
216 * [big_two_unmapped_areas[0]->end, big_two_unmapped_areas[1]->start)
217 * [big_two_unmapped_areas[1]->end, mappings[nr_mappings - 1]->end)
218 *
219 * As usual memory map of processes is as below, the gap between the heap and
220 * the uppermost mmap()-ed region, and the gap between the lowermost mmap()-ed
221 * region and the stack will be two biggest unmapped regions. Because these
222 * gaps are exceptionally huge areas in usual address space, excluding these
223 * two biggest unmapped regions will be sufficient to make a trade-off.
224 *
225 * <heap>
226 * <BIG UNMAPPED REGION 1>
227 * <uppermost mmap()-ed region>
228 * (other mmap()-ed regions and small unmapped regions)
229 * <lowermost mmap()-ed region>
230 * <BIG UNMAPPED REGION 2>
231 * <stack>
232 */
233static void __damon_va_init_regions(struct damon_ctx *ctx,
234 struct damon_target *t)
235{
236 struct damon_target *ti;
237 struct damon_region *r;
238 struct damon_addr_range regions[3];
239 unsigned long sz = 0, nr_pieces;
240 int i, tidx = 0;
241
242 if (damon_va_three_regions(t, regions)) {
243 damon_for_each_target(ti, ctx) {
244 if (ti == t)
245 break;
246 tidx++;
247 }
248 pr_debug("Failed to get three regions of %dth target\n", tidx);
249 return;
250 }
251
252 for (i = 0; i < 3; i++)
253 sz += regions[i].end - regions[i].start;
254 if (ctx->attrs.min_nr_regions)
255 sz /= ctx->attrs.min_nr_regions;
256 if (sz < DAMON_MIN_REGION)
257 sz = DAMON_MIN_REGION;
258
259 /* Set the initial three regions of the target */
260 for (i = 0; i < 3; i++) {
261 r = damon_new_region(regions[i].start, regions[i].end);
262 if (!r) {
263 pr_err("%d'th init region creation failed\n", i);
264 return;
265 }
266 damon_add_region(r, t);
267
268 nr_pieces = (regions[i].end - regions[i].start) / sz;
269 damon_va_evenly_split_region(t, r, nr_pieces);
270 }
271}
272
273/* Initialize '->regions_list' of every target (task) */
274static void damon_va_init(struct damon_ctx *ctx)
275{
276 struct damon_target *t;
277
278 damon_for_each_target(t, ctx) {
279 /* the user may set the target regions as they want */
280 if (!damon_nr_regions(t))
281 __damon_va_init_regions(ctx, t);
282 }
283}
284
285/*
286 * Update regions for current memory mappings
287 */
288static void damon_va_update(struct damon_ctx *ctx)
289{
290 struct damon_addr_range three_regions[3];
291 struct damon_target *t;
292
293 damon_for_each_target(t, ctx) {
294 if (damon_va_three_regions(t, three_regions))
295 continue;
296 damon_set_regions(t, three_regions, 3);
297 }
298}
299
300static int damon_mkold_pmd_entry(pmd_t *pmd, unsigned long addr,
301 unsigned long next, struct mm_walk *walk)
302{
303 pte_t *pte;
304 spinlock_t *ptl;
305
306 if (pmd_trans_huge(*pmd)) {
307 ptl = pmd_lock(walk->mm, pmd);
308 if (!pmd_present(*pmd)) {
309 spin_unlock(ptl);
310 return 0;
311 }
312
313 if (pmd_trans_huge(*pmd)) {
314 damon_pmdp_mkold(pmd, walk->mm, addr);
315 spin_unlock(ptl);
316 return 0;
317 }
318 spin_unlock(ptl);
319 }
320
321 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
322 return 0;
323 pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
324 if (!pte_present(*pte))
325 goto out;
326 damon_ptep_mkold(pte, walk->mm, addr);
327out:
328 pte_unmap_unlock(pte, ptl);
329 return 0;
330}
331
332#ifdef CONFIG_HUGETLB_PAGE
333static void damon_hugetlb_mkold(pte_t *pte, struct mm_struct *mm,
334 struct vm_area_struct *vma, unsigned long addr)
335{
336 bool referenced = false;
337 pte_t entry = huge_ptep_get(pte);
338 struct page *page = pte_page(entry);
339
340 get_page(page);
341
342 if (pte_young(entry)) {
343 referenced = true;
344 entry = pte_mkold(entry);
345 set_huge_pte_at(mm, addr, pte, entry);
346 }
347
348#ifdef CONFIG_MMU_NOTIFIER
349 if (mmu_notifier_clear_young(mm, addr,
350 addr + huge_page_size(hstate_vma(vma))))
351 referenced = true;
352#endif /* CONFIG_MMU_NOTIFIER */
353
354 if (referenced)
355 set_page_young(page);
356
357 set_page_idle(page);
358 put_page(page);
359}
360
361static int damon_mkold_hugetlb_entry(pte_t *pte, unsigned long hmask,
362 unsigned long addr, unsigned long end,
363 struct mm_walk *walk)
364{
365 struct hstate *h = hstate_vma(walk->vma);
366 spinlock_t *ptl;
367 pte_t entry;
368
369 ptl = huge_pte_lock(h, walk->mm, pte);
370 entry = huge_ptep_get(pte);
371 if (!pte_present(entry))
372 goto out;
373
374 damon_hugetlb_mkold(pte, walk->mm, walk->vma, addr);
375
376out:
377 spin_unlock(ptl);
378 return 0;
379}
380#else
381#define damon_mkold_hugetlb_entry NULL
382#endif /* CONFIG_HUGETLB_PAGE */
383
384static const struct mm_walk_ops damon_mkold_ops = {
385 .pmd_entry = damon_mkold_pmd_entry,
386 .hugetlb_entry = damon_mkold_hugetlb_entry,
387};
388
389static void damon_va_mkold(struct mm_struct *mm, unsigned long addr)
390{
391 mmap_read_lock(mm);
392 walk_page_range(mm, addr, addr + 1, &damon_mkold_ops, NULL);
393 mmap_read_unlock(mm);
394}
395
396/*
397 * Functions for the access checking of the regions
398 */
399
400static void __damon_va_prepare_access_check(struct mm_struct *mm,
401 struct damon_region *r)
402{
403 r->sampling_addr = damon_rand(r->ar.start, r->ar.end);
404
405 damon_va_mkold(mm, r->sampling_addr);
406}
407
408static void damon_va_prepare_access_checks(struct damon_ctx *ctx)
409{
410 struct damon_target *t;
411 struct mm_struct *mm;
412 struct damon_region *r;
413
414 damon_for_each_target(t, ctx) {
415 mm = damon_get_mm(t);
416 if (!mm)
417 continue;
418 damon_for_each_region(r, t)
419 __damon_va_prepare_access_check(mm, r);
420 mmput(mm);
421 }
422}
423
424struct damon_young_walk_private {
425 unsigned long *page_sz;
426 bool young;
427};
428
429static int damon_young_pmd_entry(pmd_t *pmd, unsigned long addr,
430 unsigned long next, struct mm_walk *walk)
431{
432 pte_t *pte;
433 spinlock_t *ptl;
434 struct page *page;
435 struct damon_young_walk_private *priv = walk->private;
436
437#ifdef CONFIG_TRANSPARENT_HUGEPAGE
438 if (pmd_trans_huge(*pmd)) {
439 ptl = pmd_lock(walk->mm, pmd);
440 if (!pmd_present(*pmd)) {
441 spin_unlock(ptl);
442 return 0;
443 }
444
445 if (!pmd_trans_huge(*pmd)) {
446 spin_unlock(ptl);
447 goto regular_page;
448 }
449 page = damon_get_page(pmd_pfn(*pmd));
450 if (!page)
451 goto huge_out;
452 if (pmd_young(*pmd) || !page_is_idle(page) ||
453 mmu_notifier_test_young(walk->mm,
454 addr)) {
455 *priv->page_sz = HPAGE_PMD_SIZE;
456 priv->young = true;
457 }
458 put_page(page);
459huge_out:
460 spin_unlock(ptl);
461 return 0;
462 }
463
464regular_page:
465#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
466
467 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
468 return -EINVAL;
469 pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
470 if (!pte_present(*pte))
471 goto out;
472 page = damon_get_page(pte_pfn(*pte));
473 if (!page)
474 goto out;
475 if (pte_young(*pte) || !page_is_idle(page) ||
476 mmu_notifier_test_young(walk->mm, addr)) {
477 *priv->page_sz = PAGE_SIZE;
478 priv->young = true;
479 }
480 put_page(page);
481out:
482 pte_unmap_unlock(pte, ptl);
483 return 0;
484}
485
486#ifdef CONFIG_HUGETLB_PAGE
487static int damon_young_hugetlb_entry(pte_t *pte, unsigned long hmask,
488 unsigned long addr, unsigned long end,
489 struct mm_walk *walk)
490{
491 struct damon_young_walk_private *priv = walk->private;
492 struct hstate *h = hstate_vma(walk->vma);
493 struct page *page;
494 spinlock_t *ptl;
495 pte_t entry;
496
497 ptl = huge_pte_lock(h, walk->mm, pte);
498 entry = huge_ptep_get(pte);
499 if (!pte_present(entry))
500 goto out;
501
502 page = pte_page(entry);
503 get_page(page);
504
505 if (pte_young(entry) || !page_is_idle(page) ||
506 mmu_notifier_test_young(walk->mm, addr)) {
507 *priv->page_sz = huge_page_size(h);
508 priv->young = true;
509 }
510
511 put_page(page);
512
513out:
514 spin_unlock(ptl);
515 return 0;
516}
517#else
518#define damon_young_hugetlb_entry NULL
519#endif /* CONFIG_HUGETLB_PAGE */
520
521static const struct mm_walk_ops damon_young_ops = {
522 .pmd_entry = damon_young_pmd_entry,
523 .hugetlb_entry = damon_young_hugetlb_entry,
524};
525
526static bool damon_va_young(struct mm_struct *mm, unsigned long addr,
527 unsigned long *page_sz)
528{
529 struct damon_young_walk_private arg = {
530 .page_sz = page_sz,
531 .young = false,
532 };
533
534 mmap_read_lock(mm);
535 walk_page_range(mm, addr, addr + 1, &damon_young_ops, &arg);
536 mmap_read_unlock(mm);
537 return arg.young;
538}
539
540/*
541 * Check whether the region was accessed after the last preparation
542 *
543 * mm 'mm_struct' for the given virtual address space
544 * r the region to be checked
545 */
546static void __damon_va_check_access(struct mm_struct *mm,
547 struct damon_region *r, bool same_target)
548{
549 static unsigned long last_addr;
550 static unsigned long last_page_sz = PAGE_SIZE;
551 static bool last_accessed;
552
553 /* If the region is in the last checked page, reuse the result */
554 if (same_target && (ALIGN_DOWN(last_addr, last_page_sz) ==
555 ALIGN_DOWN(r->sampling_addr, last_page_sz))) {
556 if (last_accessed)
557 r->nr_accesses++;
558 return;
559 }
560
561 last_accessed = damon_va_young(mm, r->sampling_addr, &last_page_sz);
562 if (last_accessed)
563 r->nr_accesses++;
564
565 last_addr = r->sampling_addr;
566}
567
568static unsigned int damon_va_check_accesses(struct damon_ctx *ctx)
569{
570 struct damon_target *t;
571 struct mm_struct *mm;
572 struct damon_region *r;
573 unsigned int max_nr_accesses = 0;
574 bool same_target;
575
576 damon_for_each_target(t, ctx) {
577 mm = damon_get_mm(t);
578 if (!mm)
579 continue;
580 same_target = false;
581 damon_for_each_region(r, t) {
582 __damon_va_check_access(mm, r, same_target);
583 max_nr_accesses = max(r->nr_accesses, max_nr_accesses);
584 same_target = true;
585 }
586 mmput(mm);
587 }
588
589 return max_nr_accesses;
590}
591
592/*
593 * Functions for the target validity check and cleanup
594 */
595
596static bool damon_va_target_valid(struct damon_target *t)
597{
598 struct task_struct *task;
599
600 task = damon_get_task_struct(t);
601 if (task) {
602 put_task_struct(task);
603 return true;
604 }
605
606 return false;
607}
608
609#ifndef CONFIG_ADVISE_SYSCALLS
610static unsigned long damos_madvise(struct damon_target *target,
611 struct damon_region *r, int behavior)
612{
613 return 0;
614}
615#else
616static unsigned long damos_madvise(struct damon_target *target,
617 struct damon_region *r, int behavior)
618{
619 struct mm_struct *mm;
620 unsigned long start = PAGE_ALIGN(r->ar.start);
621 unsigned long len = PAGE_ALIGN(damon_sz_region(r));
622 unsigned long applied;
623
624 mm = damon_get_mm(target);
625 if (!mm)
626 return 0;
627
628 applied = do_madvise(mm, start, len, behavior) ? 0 : len;
629 mmput(mm);
630
631 return applied;
632}
633#endif /* CONFIG_ADVISE_SYSCALLS */
634
635static unsigned long damon_va_apply_scheme(struct damon_ctx *ctx,
636 struct damon_target *t, struct damon_region *r,
637 struct damos *scheme)
638{
639 int madv_action;
640
641 switch (scheme->action) {
642 case DAMOS_WILLNEED:
643 madv_action = MADV_WILLNEED;
644 break;
645 case DAMOS_COLD:
646 madv_action = MADV_COLD;
647 break;
648 case DAMOS_PAGEOUT:
649 madv_action = MADV_PAGEOUT;
650 break;
651 case DAMOS_HUGEPAGE:
652 madv_action = MADV_HUGEPAGE;
653 break;
654 case DAMOS_NOHUGEPAGE:
655 madv_action = MADV_NOHUGEPAGE;
656 break;
657 case DAMOS_STAT:
658 return 0;
659 default:
660 /*
661 * DAMOS actions that are not yet supported by 'vaddr'.
662 */
663 return 0;
664 }
665
666 return damos_madvise(t, r, madv_action);
667}
668
669static int damon_va_scheme_score(struct damon_ctx *context,
670 struct damon_target *t, struct damon_region *r,
671 struct damos *scheme)
672{
673
674 switch (scheme->action) {
675 case DAMOS_PAGEOUT:
676 return damon_cold_score(context, r, scheme);
677 default:
678 break;
679 }
680
681 return DAMOS_MAX_SCORE;
682}
683
684static int __init damon_va_initcall(void)
685{
686 struct damon_operations ops = {
687 .id = DAMON_OPS_VADDR,
688 .init = damon_va_init,
689 .update = damon_va_update,
690 .prepare_access_checks = damon_va_prepare_access_checks,
691 .check_accesses = damon_va_check_accesses,
692 .reset_aggregated = NULL,
693 .target_valid = damon_va_target_valid,
694 .cleanup = NULL,
695 .apply_scheme = damon_va_apply_scheme,
696 .get_scheme_score = damon_va_scheme_score,
697 };
698 /* ops for fixed virtual address ranges */
699 struct damon_operations ops_fvaddr = ops;
700 int err;
701
702 /* Don't set the monitoring target regions for the entire mapping */
703 ops_fvaddr.id = DAMON_OPS_FVADDR;
704 ops_fvaddr.init = NULL;
705 ops_fvaddr.update = NULL;
706
707 err = damon_register_ops(&ops);
708 if (err)
709 return err;
710 return damon_register_ops(&ops_fvaddr);
711};
712
713subsys_initcall(damon_va_initcall);
714
715#include "vaddr-test.h"