1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * DAMON Primitives for The Physical Address Space
4 *
5 * Author: SeongJae Park <sj@kernel.org>
6 */
7
8 #define pr_fmt(fmt) "damon-pa: " fmt
9
10 #include <linux/mmu_notifier.h>
11 #include <linux/page_idle.h>
12 #include <linux/pagemap.h>
13 #include <linux/rmap.h>
14 #include <linux/swap.h>
15 #include <linux/memory-tiers.h>
16 #include <linux/migrate.h>
17 #include <linux/mm_inline.h>
18
19 #include "../internal.h"
20 #include "ops-common.h"
21
damon_folio_mkold_one(struct folio * folio,struct vm_area_struct * vma,unsigned long addr,void * arg)22 static bool damon_folio_mkold_one(struct folio *folio,
23 struct vm_area_struct *vma, unsigned long addr, void *arg)
24 {
25 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, addr, 0);
26
27 while (page_vma_mapped_walk(&pvmw)) {
28 addr = pvmw.address;
29 if (pvmw.pte)
30 damon_ptep_mkold(pvmw.pte, vma, addr);
31 else
32 damon_pmdp_mkold(pvmw.pmd, vma, addr);
33 }
34 return true;
35 }
36
damon_folio_mkold(struct folio * folio)37 static void damon_folio_mkold(struct folio *folio)
38 {
39 struct rmap_walk_control rwc = {
40 .rmap_one = damon_folio_mkold_one,
41 .anon_lock = folio_lock_anon_vma_read,
42 };
43 bool need_lock;
44
45 if (!folio_mapped(folio) || !folio_raw_mapping(folio)) {
46 folio_set_idle(folio);
47 return;
48 }
49
50 need_lock = !folio_test_anon(folio) || folio_test_ksm(folio);
51 if (need_lock && !folio_trylock(folio))
52 return;
53
54 rmap_walk(folio, &rwc);
55
56 if (need_lock)
57 folio_unlock(folio);
58
59 }
60
damon_pa_mkold(unsigned long paddr)61 static void damon_pa_mkold(unsigned long paddr)
62 {
63 struct folio *folio = damon_get_folio(PHYS_PFN(paddr));
64
65 if (!folio)
66 return;
67
68 damon_folio_mkold(folio);
69 folio_put(folio);
70 }
71
__damon_pa_prepare_access_check(struct damon_region * r)72 static void __damon_pa_prepare_access_check(struct damon_region *r)
73 {
74 r->sampling_addr = damon_rand(r->ar.start, r->ar.end);
75
76 damon_pa_mkold(r->sampling_addr);
77 }
78
damon_pa_prepare_access_checks(struct damon_ctx * ctx)79 static void damon_pa_prepare_access_checks(struct damon_ctx *ctx)
80 {
81 struct damon_target *t;
82 struct damon_region *r;
83
84 damon_for_each_target(t, ctx) {
85 damon_for_each_region(r, t)
86 __damon_pa_prepare_access_check(r);
87 }
88 }
89
damon_folio_young_one(struct folio * folio,struct vm_area_struct * vma,unsigned long addr,void * arg)90 static bool damon_folio_young_one(struct folio *folio,
91 struct vm_area_struct *vma, unsigned long addr, void *arg)
92 {
93 bool *accessed = arg;
94 DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, addr, 0);
95
96 *accessed = false;
97 while (page_vma_mapped_walk(&pvmw)) {
98 addr = pvmw.address;
99 if (pvmw.pte) {
100 *accessed = pte_young(ptep_get(pvmw.pte)) ||
101 !folio_test_idle(folio) ||
102 mmu_notifier_test_young(vma->vm_mm, addr);
103 } else {
104 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
105 *accessed = pmd_young(pmdp_get(pvmw.pmd)) ||
106 !folio_test_idle(folio) ||
107 mmu_notifier_test_young(vma->vm_mm, addr);
108 #else
109 WARN_ON_ONCE(1);
110 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
111 }
112 if (*accessed) {
113 page_vma_mapped_walk_done(&pvmw);
114 break;
115 }
116 }
117
118 /* If accessed, stop walking */
119 return *accessed == false;
120 }
121
damon_folio_young(struct folio * folio)122 static bool damon_folio_young(struct folio *folio)
123 {
124 bool accessed = false;
125 struct rmap_walk_control rwc = {
126 .arg = &accessed,
127 .rmap_one = damon_folio_young_one,
128 .anon_lock = folio_lock_anon_vma_read,
129 };
130 bool need_lock;
131
132 if (!folio_mapped(folio) || !folio_raw_mapping(folio)) {
133 if (folio_test_idle(folio))
134 return false;
135 else
136 return true;
137 }
138
139 need_lock = !folio_test_anon(folio) || folio_test_ksm(folio);
140 if (need_lock && !folio_trylock(folio))
141 return false;
142
143 rmap_walk(folio, &rwc);
144
145 if (need_lock)
146 folio_unlock(folio);
147
148 return accessed;
149 }
150
damon_pa_young(unsigned long paddr,unsigned long * folio_sz)151 static bool damon_pa_young(unsigned long paddr, unsigned long *folio_sz)
152 {
153 struct folio *folio = damon_get_folio(PHYS_PFN(paddr));
154 bool accessed;
155
156 if (!folio)
157 return false;
158
159 accessed = damon_folio_young(folio);
160 *folio_sz = folio_size(folio);
161 folio_put(folio);
162 return accessed;
163 }
164
__damon_pa_check_access(struct damon_region * r,struct damon_attrs * attrs)165 static void __damon_pa_check_access(struct damon_region *r,
166 struct damon_attrs *attrs)
167 {
168 static unsigned long last_addr;
169 static unsigned long last_folio_sz = PAGE_SIZE;
170 static bool last_accessed;
171
172 /* If the region is in the last checked page, reuse the result */
173 if (ALIGN_DOWN(last_addr, last_folio_sz) ==
174 ALIGN_DOWN(r->sampling_addr, last_folio_sz)) {
175 damon_update_region_access_rate(r, last_accessed, attrs);
176 return;
177 }
178
179 last_accessed = damon_pa_young(r->sampling_addr, &last_folio_sz);
180 damon_update_region_access_rate(r, last_accessed, attrs);
181
182 last_addr = r->sampling_addr;
183 }
184
damon_pa_check_accesses(struct damon_ctx * ctx)185 static unsigned int damon_pa_check_accesses(struct damon_ctx *ctx)
186 {
187 struct damon_target *t;
188 struct damon_region *r;
189 unsigned int max_nr_accesses = 0;
190
191 damon_for_each_target(t, ctx) {
192 damon_for_each_region(r, t) {
193 __damon_pa_check_access(r, &ctx->attrs);
194 max_nr_accesses = max(r->nr_accesses, max_nr_accesses);
195 }
196 }
197
198 return max_nr_accesses;
199 }
200
damos_pa_filter_match(struct damos_filter * filter,struct folio * folio)201 static bool damos_pa_filter_match(struct damos_filter *filter,
202 struct folio *folio)
203 {
204 bool matched = false;
205 struct mem_cgroup *memcg;
206
207 switch (filter->type) {
208 case DAMOS_FILTER_TYPE_ANON:
209 matched = folio_test_anon(folio);
210 break;
211 case DAMOS_FILTER_TYPE_MEMCG:
212 rcu_read_lock();
213 memcg = folio_memcg_check(folio);
214 if (!memcg)
215 matched = false;
216 else
217 matched = filter->memcg_id == mem_cgroup_id(memcg);
218 rcu_read_unlock();
219 break;
220 case DAMOS_FILTER_TYPE_YOUNG:
221 matched = damon_folio_young(folio);
222 if (matched)
223 damon_folio_mkold(folio);
224 break;
225 default:
226 break;
227 }
228
229 return matched == filter->matching;
230 }
231
232 /*
233 * damos_pa_filter_out - Return true if the page should be filtered out.
234 */
damos_pa_filter_out(struct damos * scheme,struct folio * folio)235 static bool damos_pa_filter_out(struct damos *scheme, struct folio *folio)
236 {
237 struct damos_filter *filter;
238
239 if (scheme->core_filters_allowed)
240 return false;
241
242 damos_for_each_filter(filter, scheme) {
243 if (damos_pa_filter_match(filter, folio))
244 return !filter->allow;
245 }
246 return false;
247 }
248
damon_pa_pageout(struct damon_region * r,struct damos * s,unsigned long * sz_filter_passed)249 static unsigned long damon_pa_pageout(struct damon_region *r, struct damos *s,
250 unsigned long *sz_filter_passed)
251 {
252 unsigned long addr, applied;
253 LIST_HEAD(folio_list);
254 bool install_young_filter = true;
255 struct damos_filter *filter;
256
257 /* check access in page level again by default */
258 damos_for_each_filter(filter, s) {
259 if (filter->type == DAMOS_FILTER_TYPE_YOUNG) {
260 install_young_filter = false;
261 break;
262 }
263 }
264 if (install_young_filter) {
265 filter = damos_new_filter(
266 DAMOS_FILTER_TYPE_YOUNG, true, false);
267 if (!filter)
268 return 0;
269 damos_add_filter(s, filter);
270 }
271
272 for (addr = r->ar.start; addr < r->ar.end; addr += PAGE_SIZE) {
273 struct folio *folio = damon_get_folio(PHYS_PFN(addr));
274
275 if (!folio)
276 continue;
277
278 if (damos_pa_filter_out(s, folio))
279 goto put_folio;
280 else
281 *sz_filter_passed += folio_size(folio);
282
283 folio_clear_referenced(folio);
284 folio_test_clear_young(folio);
285 if (!folio_isolate_lru(folio))
286 goto put_folio;
287 if (folio_test_unevictable(folio))
288 folio_putback_lru(folio);
289 else
290 list_add(&folio->lru, &folio_list);
291 put_folio:
292 folio_put(folio);
293 }
294 if (install_young_filter)
295 damos_destroy_filter(filter);
296 applied = reclaim_pages(&folio_list);
297 cond_resched();
298 return applied * PAGE_SIZE;
299 }
300
damon_pa_mark_accessed_or_deactivate(struct damon_region * r,struct damos * s,bool mark_accessed,unsigned long * sz_filter_passed)301 static inline unsigned long damon_pa_mark_accessed_or_deactivate(
302 struct damon_region *r, struct damos *s, bool mark_accessed,
303 unsigned long *sz_filter_passed)
304 {
305 unsigned long addr, applied = 0;
306
307 for (addr = r->ar.start; addr < r->ar.end; addr += PAGE_SIZE) {
308 struct folio *folio = damon_get_folio(PHYS_PFN(addr));
309
310 if (!folio)
311 continue;
312
313 if (damos_pa_filter_out(s, folio))
314 goto put_folio;
315 else
316 *sz_filter_passed += folio_size(folio);
317
318 if (mark_accessed)
319 folio_mark_accessed(folio);
320 else
321 folio_deactivate(folio);
322 applied += folio_nr_pages(folio);
323 put_folio:
324 folio_put(folio);
325 }
326 return applied * PAGE_SIZE;
327 }
328
damon_pa_mark_accessed(struct damon_region * r,struct damos * s,unsigned long * sz_filter_passed)329 static unsigned long damon_pa_mark_accessed(struct damon_region *r,
330 struct damos *s, unsigned long *sz_filter_passed)
331 {
332 return damon_pa_mark_accessed_or_deactivate(r, s, true,
333 sz_filter_passed);
334 }
335
damon_pa_deactivate_pages(struct damon_region * r,struct damos * s,unsigned long * sz_filter_passed)336 static unsigned long damon_pa_deactivate_pages(struct damon_region *r,
337 struct damos *s, unsigned long *sz_filter_passed)
338 {
339 return damon_pa_mark_accessed_or_deactivate(r, s, false,
340 sz_filter_passed);
341 }
342
__damon_pa_migrate_folio_list(struct list_head * migrate_folios,struct pglist_data * pgdat,int target_nid)343 static unsigned int __damon_pa_migrate_folio_list(
344 struct list_head *migrate_folios, struct pglist_data *pgdat,
345 int target_nid)
346 {
347 unsigned int nr_succeeded = 0;
348 nodemask_t allowed_mask = NODE_MASK_NONE;
349 struct migration_target_control mtc = {
350 /*
351 * Allocate from 'node', or fail quickly and quietly.
352 * When this happens, 'page' will likely just be discarded
353 * instead of migrated.
354 */
355 .gfp_mask = (GFP_HIGHUSER_MOVABLE & ~__GFP_RECLAIM) |
356 __GFP_NOWARN | __GFP_NOMEMALLOC | GFP_NOWAIT,
357 .nid = target_nid,
358 .nmask = &allowed_mask
359 };
360
361 if (pgdat->node_id == target_nid || target_nid == NUMA_NO_NODE)
362 return 0;
363
364 if (list_empty(migrate_folios))
365 return 0;
366
367 /* Migration ignores all cpuset and mempolicy settings */
368 migrate_pages(migrate_folios, alloc_migrate_folio, NULL,
369 (unsigned long)&mtc, MIGRATE_ASYNC, MR_DAMON,
370 &nr_succeeded);
371
372 return nr_succeeded;
373 }
374
damon_pa_migrate_folio_list(struct list_head * folio_list,struct pglist_data * pgdat,int target_nid)375 static unsigned int damon_pa_migrate_folio_list(struct list_head *folio_list,
376 struct pglist_data *pgdat,
377 int target_nid)
378 {
379 unsigned int nr_migrated = 0;
380 struct folio *folio;
381 LIST_HEAD(ret_folios);
382 LIST_HEAD(migrate_folios);
383
384 while (!list_empty(folio_list)) {
385 struct folio *folio;
386
387 cond_resched();
388
389 folio = lru_to_folio(folio_list);
390 list_del(&folio->lru);
391
392 if (!folio_trylock(folio))
393 goto keep;
394
395 /* Relocate its contents to another node. */
396 list_add(&folio->lru, &migrate_folios);
397 folio_unlock(folio);
398 continue;
399 keep:
400 list_add(&folio->lru, &ret_folios);
401 }
402 /* 'folio_list' is always empty here */
403
404 /* Migrate folios selected for migration */
405 nr_migrated += __damon_pa_migrate_folio_list(
406 &migrate_folios, pgdat, target_nid);
407 /*
408 * Folios that could not be migrated are still in @migrate_folios. Add
409 * those back on @folio_list
410 */
411 if (!list_empty(&migrate_folios))
412 list_splice_init(&migrate_folios, folio_list);
413
414 try_to_unmap_flush();
415
416 list_splice(&ret_folios, folio_list);
417
418 while (!list_empty(folio_list)) {
419 folio = lru_to_folio(folio_list);
420 list_del(&folio->lru);
421 folio_putback_lru(folio);
422 }
423
424 return nr_migrated;
425 }
426
damon_pa_migrate_pages(struct list_head * folio_list,int target_nid)427 static unsigned long damon_pa_migrate_pages(struct list_head *folio_list,
428 int target_nid)
429 {
430 int nid;
431 unsigned long nr_migrated = 0;
432 LIST_HEAD(node_folio_list);
433 unsigned int noreclaim_flag;
434
435 if (list_empty(folio_list))
436 return nr_migrated;
437
438 noreclaim_flag = memalloc_noreclaim_save();
439
440 nid = folio_nid(lru_to_folio(folio_list));
441 do {
442 struct folio *folio = lru_to_folio(folio_list);
443
444 if (nid == folio_nid(folio)) {
445 list_move(&folio->lru, &node_folio_list);
446 continue;
447 }
448
449 nr_migrated += damon_pa_migrate_folio_list(&node_folio_list,
450 NODE_DATA(nid),
451 target_nid);
452 nid = folio_nid(lru_to_folio(folio_list));
453 } while (!list_empty(folio_list));
454
455 nr_migrated += damon_pa_migrate_folio_list(&node_folio_list,
456 NODE_DATA(nid),
457 target_nid);
458
459 memalloc_noreclaim_restore(noreclaim_flag);
460
461 return nr_migrated;
462 }
463
damon_pa_migrate(struct damon_region * r,struct damos * s,unsigned long * sz_filter_passed)464 static unsigned long damon_pa_migrate(struct damon_region *r, struct damos *s,
465 unsigned long *sz_filter_passed)
466 {
467 unsigned long addr, applied;
468 LIST_HEAD(folio_list);
469
470 for (addr = r->ar.start; addr < r->ar.end; addr += PAGE_SIZE) {
471 struct folio *folio = damon_get_folio(PHYS_PFN(addr));
472
473 if (!folio)
474 continue;
475
476 if (damos_pa_filter_out(s, folio))
477 goto put_folio;
478 else
479 *sz_filter_passed += folio_size(folio);
480
481 if (!folio_isolate_lru(folio))
482 goto put_folio;
483 list_add(&folio->lru, &folio_list);
484 put_folio:
485 folio_put(folio);
486 }
487 applied = damon_pa_migrate_pages(&folio_list, s->target_nid);
488 cond_resched();
489 return applied * PAGE_SIZE;
490 }
491
damon_pa_scheme_has_filter(struct damos * s)492 static bool damon_pa_scheme_has_filter(struct damos *s)
493 {
494 struct damos_filter *f;
495
496 damos_for_each_filter(f, s)
497 return true;
498 return false;
499 }
500
damon_pa_stat(struct damon_region * r,struct damos * s,unsigned long * sz_filter_passed)501 static unsigned long damon_pa_stat(struct damon_region *r, struct damos *s,
502 unsigned long *sz_filter_passed)
503 {
504 unsigned long addr;
505 LIST_HEAD(folio_list);
506
507 if (!damon_pa_scheme_has_filter(s))
508 return 0;
509
510 addr = r->ar.start;
511 while (addr < r->ar.end) {
512 struct folio *folio = damon_get_folio(PHYS_PFN(addr));
513
514 if (!folio) {
515 addr += PAGE_SIZE;
516 continue;
517 }
518
519 if (!damos_pa_filter_out(s, folio))
520 *sz_filter_passed += folio_size(folio);
521 addr += folio_size(folio);
522 folio_put(folio);
523 }
524 return 0;
525 }
526
damon_pa_apply_scheme(struct damon_ctx * ctx,struct damon_target * t,struct damon_region * r,struct damos * scheme,unsigned long * sz_filter_passed)527 static unsigned long damon_pa_apply_scheme(struct damon_ctx *ctx,
528 struct damon_target *t, struct damon_region *r,
529 struct damos *scheme, unsigned long *sz_filter_passed)
530 {
531 switch (scheme->action) {
532 case DAMOS_PAGEOUT:
533 return damon_pa_pageout(r, scheme, sz_filter_passed);
534 case DAMOS_LRU_PRIO:
535 return damon_pa_mark_accessed(r, scheme, sz_filter_passed);
536 case DAMOS_LRU_DEPRIO:
537 return damon_pa_deactivate_pages(r, scheme, sz_filter_passed);
538 case DAMOS_MIGRATE_HOT:
539 case DAMOS_MIGRATE_COLD:
540 return damon_pa_migrate(r, scheme, sz_filter_passed);
541 case DAMOS_STAT:
542 return damon_pa_stat(r, scheme, sz_filter_passed);
543 default:
544 /* DAMOS actions that not yet supported by 'paddr'. */
545 break;
546 }
547 return 0;
548 }
549
damon_pa_scheme_score(struct damon_ctx * context,struct damon_target * t,struct damon_region * r,struct damos * scheme)550 static int damon_pa_scheme_score(struct damon_ctx *context,
551 struct damon_target *t, struct damon_region *r,
552 struct damos *scheme)
553 {
554 switch (scheme->action) {
555 case DAMOS_PAGEOUT:
556 return damon_cold_score(context, r, scheme);
557 case DAMOS_LRU_PRIO:
558 return damon_hot_score(context, r, scheme);
559 case DAMOS_LRU_DEPRIO:
560 return damon_cold_score(context, r, scheme);
561 case DAMOS_MIGRATE_HOT:
562 return damon_hot_score(context, r, scheme);
563 case DAMOS_MIGRATE_COLD:
564 return damon_cold_score(context, r, scheme);
565 default:
566 break;
567 }
568
569 return DAMOS_MAX_SCORE;
570 }
571
damon_pa_initcall(void)572 static int __init damon_pa_initcall(void)
573 {
574 struct damon_operations ops = {
575 .id = DAMON_OPS_PADDR,
576 .init = NULL,
577 .update = NULL,
578 .prepare_access_checks = damon_pa_prepare_access_checks,
579 .check_accesses = damon_pa_check_accesses,
580 .reset_aggregated = NULL,
581 .target_valid = NULL,
582 .cleanup = NULL,
583 .apply_scheme = damon_pa_apply_scheme,
584 .get_scheme_score = damon_pa_scheme_score,
585 };
586
587 return damon_register_ops(&ops);
588 };
589
590 subsys_initcall(damon_pa_initcall);
591