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
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
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
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
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
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
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
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
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
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
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
201 static bool __damos_pa_filter_out(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 */
235 static bool damos_pa_filter_out(struct damos *scheme, struct folio *folio)
236 {
237 struct damos_filter *filter;
238
239 damos_for_each_filter(filter, scheme) {
240 if (__damos_pa_filter_out(filter, folio))
241 return true;
242 }
243 return false;
244 }
245
246 static unsigned long damon_pa_pageout(struct damon_region *r, struct damos *s)
247 {
248 unsigned long addr, applied;
249 LIST_HEAD(folio_list);
250 bool install_young_filter = true;
251 struct damos_filter *filter;
252
253 /* check access in page level again by default */
254 damos_for_each_filter(filter, s) {
255 if (filter->type == DAMOS_FILTER_TYPE_YOUNG) {
256 install_young_filter = false;
257 break;
258 }
259 }
260 if (install_young_filter) {
261 filter = damos_new_filter(DAMOS_FILTER_TYPE_YOUNG, true);
262 if (!filter)
263 return 0;
264 damos_add_filter(s, filter);
265 }
266
267 for (addr = r->ar.start; addr < r->ar.end; addr += PAGE_SIZE) {
268 struct folio *folio = damon_get_folio(PHYS_PFN(addr));
269
270 if (!folio)
271 continue;
272
273 if (damos_pa_filter_out(s, folio))
274 goto put_folio;
275
276 folio_clear_referenced(folio);
277 folio_test_clear_young(folio);
278 if (!folio_isolate_lru(folio))
279 goto put_folio;
280 if (folio_test_unevictable(folio))
281 folio_putback_lru(folio);
282 else
283 list_add(&folio->lru, &folio_list);
284 put_folio:
285 folio_put(folio);
286 }
287 if (install_young_filter)
288 damos_destroy_filter(filter);
289 applied = reclaim_pages(&folio_list);
290 cond_resched();
291 return applied * PAGE_SIZE;
292 }
293
294 static inline unsigned long damon_pa_mark_accessed_or_deactivate(
295 struct damon_region *r, struct damos *s, bool mark_accessed)
296 {
297 unsigned long addr, applied = 0;
298
299 for (addr = r->ar.start; addr < r->ar.end; addr += PAGE_SIZE) {
300 struct folio *folio = damon_get_folio(PHYS_PFN(addr));
301
302 if (!folio)
303 continue;
304
305 if (damos_pa_filter_out(s, folio))
306 goto put_folio;
307
308 if (mark_accessed)
309 folio_mark_accessed(folio);
310 else
311 folio_deactivate(folio);
312 applied += folio_nr_pages(folio);
313 put_folio:
314 folio_put(folio);
315 }
316 return applied * PAGE_SIZE;
317 }
318
319 static unsigned long damon_pa_mark_accessed(struct damon_region *r,
320 struct damos *s)
321 {
322 return damon_pa_mark_accessed_or_deactivate(r, s, true);
323 }
324
325 static unsigned long damon_pa_deactivate_pages(struct damon_region *r,
326 struct damos *s)
327 {
328 return damon_pa_mark_accessed_or_deactivate(r, s, false);
329 }
330
331 static unsigned int __damon_pa_migrate_folio_list(
332 struct list_head *migrate_folios, struct pglist_data *pgdat,
333 int target_nid)
334 {
335 unsigned int nr_succeeded = 0;
336 nodemask_t allowed_mask = NODE_MASK_NONE;
337 struct migration_target_control mtc = {
338 /*
339 * Allocate from 'node', or fail quickly and quietly.
340 * When this happens, 'page' will likely just be discarded
341 * instead of migrated.
342 */
343 .gfp_mask = (GFP_HIGHUSER_MOVABLE & ~__GFP_RECLAIM) |
344 __GFP_NOWARN | __GFP_NOMEMALLOC | GFP_NOWAIT,
345 .nid = target_nid,
346 .nmask = &allowed_mask
347 };
348
349 if (pgdat->node_id == target_nid || target_nid == NUMA_NO_NODE)
350 return 0;
351
352 if (list_empty(migrate_folios))
353 return 0;
354
355 /* Migration ignores all cpuset and mempolicy settings */
356 migrate_pages(migrate_folios, alloc_migrate_folio, NULL,
357 (unsigned long)&mtc, MIGRATE_ASYNC, MR_DAMON,
358 &nr_succeeded);
359
360 return nr_succeeded;
361 }
362
363 static unsigned int damon_pa_migrate_folio_list(struct list_head *folio_list,
364 struct pglist_data *pgdat,
365 int target_nid)
366 {
367 unsigned int nr_migrated = 0;
368 struct folio *folio;
369 LIST_HEAD(ret_folios);
370 LIST_HEAD(migrate_folios);
371
372 while (!list_empty(folio_list)) {
373 struct folio *folio;
374
375 cond_resched();
376
377 folio = lru_to_folio(folio_list);
378 list_del(&folio->lru);
379
380 if (!folio_trylock(folio))
381 goto keep;
382
383 /* Relocate its contents to another node. */
384 list_add(&folio->lru, &migrate_folios);
385 folio_unlock(folio);
386 continue;
387 keep:
388 list_add(&folio->lru, &ret_folios);
389 }
390 /* 'folio_list' is always empty here */
391
392 /* Migrate folios selected for migration */
393 nr_migrated += __damon_pa_migrate_folio_list(
394 &migrate_folios, pgdat, target_nid);
395 /*
396 * Folios that could not be migrated are still in @migrate_folios. Add
397 * those back on @folio_list
398 */
399 if (!list_empty(&migrate_folios))
400 list_splice_init(&migrate_folios, folio_list);
401
402 try_to_unmap_flush();
403
404 list_splice(&ret_folios, folio_list);
405
406 while (!list_empty(folio_list)) {
407 folio = lru_to_folio(folio_list);
408 list_del(&folio->lru);
409 folio_putback_lru(folio);
410 }
411
412 return nr_migrated;
413 }
414
415 static unsigned long damon_pa_migrate_pages(struct list_head *folio_list,
416 int target_nid)
417 {
418 int nid;
419 unsigned long nr_migrated = 0;
420 LIST_HEAD(node_folio_list);
421 unsigned int noreclaim_flag;
422
423 if (list_empty(folio_list))
424 return nr_migrated;
425
426 noreclaim_flag = memalloc_noreclaim_save();
427
428 nid = folio_nid(lru_to_folio(folio_list));
429 do {
430 struct folio *folio = lru_to_folio(folio_list);
431
432 if (nid == folio_nid(folio)) {
433 list_move(&folio->lru, &node_folio_list);
434 continue;
435 }
436
437 nr_migrated += damon_pa_migrate_folio_list(&node_folio_list,
438 NODE_DATA(nid),
439 target_nid);
440 nid = folio_nid(lru_to_folio(folio_list));
441 } while (!list_empty(folio_list));
442
443 nr_migrated += damon_pa_migrate_folio_list(&node_folio_list,
444 NODE_DATA(nid),
445 target_nid);
446
447 memalloc_noreclaim_restore(noreclaim_flag);
448
449 return nr_migrated;
450 }
451
452 static unsigned long damon_pa_migrate(struct damon_region *r, struct damos *s)
453 {
454 unsigned long addr, applied;
455 LIST_HEAD(folio_list);
456
457 for (addr = r->ar.start; addr < r->ar.end; addr += PAGE_SIZE) {
458 struct folio *folio = damon_get_folio(PHYS_PFN(addr));
459
460 if (!folio)
461 continue;
462
463 if (damos_pa_filter_out(s, folio))
464 goto put_folio;
465
466 if (!folio_isolate_lru(folio))
467 goto put_folio;
468 list_add(&folio->lru, &folio_list);
469 put_folio:
470 folio_put(folio);
471 }
472 applied = damon_pa_migrate_pages(&folio_list, s->target_nid);
473 cond_resched();
474 return applied * PAGE_SIZE;
475 }
476
477
478 static unsigned long damon_pa_apply_scheme(struct damon_ctx *ctx,
479 struct damon_target *t, struct damon_region *r,
480 struct damos *scheme)
481 {
482 switch (scheme->action) {
483 case DAMOS_PAGEOUT:
484 return damon_pa_pageout(r, scheme);
485 case DAMOS_LRU_PRIO:
486 return damon_pa_mark_accessed(r, scheme);
487 case DAMOS_LRU_DEPRIO:
488 return damon_pa_deactivate_pages(r, scheme);
489 case DAMOS_MIGRATE_HOT:
490 case DAMOS_MIGRATE_COLD:
491 return damon_pa_migrate(r, scheme);
492 case DAMOS_STAT:
493 break;
494 default:
495 /* DAMOS actions that not yet supported by 'paddr'. */
496 break;
497 }
498 return 0;
499 }
500
501 static int damon_pa_scheme_score(struct damon_ctx *context,
502 struct damon_target *t, struct damon_region *r,
503 struct damos *scheme)
504 {
505 switch (scheme->action) {
506 case DAMOS_PAGEOUT:
507 return damon_cold_score(context, r, scheme);
508 case DAMOS_LRU_PRIO:
509 return damon_hot_score(context, r, scheme);
510 case DAMOS_LRU_DEPRIO:
511 return damon_cold_score(context, r, scheme);
512 case DAMOS_MIGRATE_HOT:
513 return damon_hot_score(context, r, scheme);
514 case DAMOS_MIGRATE_COLD:
515 return damon_cold_score(context, r, scheme);
516 default:
517 break;
518 }
519
520 return DAMOS_MAX_SCORE;
521 }
522
523 static int __init damon_pa_initcall(void)
524 {
525 struct damon_operations ops = {
526 .id = DAMON_OPS_PADDR,
527 .init = NULL,
528 .update = NULL,
529 .prepare_access_checks = damon_pa_prepare_access_checks,
530 .check_accesses = damon_pa_check_accesses,
531 .reset_aggregated = NULL,
532 .target_valid = NULL,
533 .cleanup = NULL,
534 .apply_scheme = damon_pa_apply_scheme,
535 .get_scheme_score = damon_pa_scheme_score,
536 };
537
538 return damon_register_ops(&ops);
539 };
540
541 subsys_initcall(damon_pa_initcall);
542
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.