1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
4  *
5  *  Swap reorganised 29.12.95, Stephen Tweedie.
6  *  kswapd added: 7.1.96  sct
7  *  Removed kswapd_ctl limits, and swap out as many pages as needed
8  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
9  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
10  *  Multiqueue VM started 5.8.00, Rik van Riel.
11  */
12 
13 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
14 
15 #include <linux/mm.h>
16 #include <linux/sched/mm.h>
17 #include <linux/module.h>
18 #include <linux/gfp.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/swap.h>
21 #include <linux/pagemap.h>
22 #include <linux/init.h>
23 #include <linux/highmem.h>
24 #include <linux/vmpressure.h>
25 #include <linux/vmstat.h>
26 #include <linux/file.h>
27 #include <linux/writeback.h>
28 #include <linux/blkdev.h>
29 #include <linux/buffer_head.h>	/* for buffer_heads_over_limit */
30 #include <linux/mm_inline.h>
31 #include <linux/backing-dev.h>
32 #include <linux/rmap.h>
33 #include <linux/topology.h>
34 #include <linux/cpu.h>
35 #include <linux/cpuset.h>
36 #include <linux/compaction.h>
37 #include <linux/notifier.h>
38 #include <linux/rwsem.h>
39 #include <linux/delay.h>
40 #include <linux/kthread.h>
41 #include <linux/freezer.h>
42 #include <linux/memcontrol.h>
43 #include <linux/migrate.h>
44 #include <linux/delayacct.h>
45 #include <linux/sysctl.h>
46 #include <linux/memory-tiers.h>
47 #include <linux/oom.h>
48 #include <linux/pagevec.h>
49 #include <linux/prefetch.h>
50 #include <linux/printk.h>
51 #include <linux/dax.h>
52 #include <linux/psi.h>
53 #include <linux/pagewalk.h>
54 #include <linux/shmem_fs.h>
55 #include <linux/ctype.h>
56 #include <linux/debugfs.h>
57 
58 #include <asm/tlbflush.h>
59 #include <asm/div64.h>
60 
61 #include <linux/swapops.h>
62 #include <linux/balloon_compaction.h>
63 #include <linux/sched/sysctl.h>
64 
65 #include "internal.h"
66 #include "swap.h"
67 
68 #define CREATE_TRACE_POINTS
69 #include <trace/events/vmscan.h>
70 
71 struct scan_control {
72 	/* How many pages shrink_list() should reclaim */
73 	unsigned long nr_to_reclaim;
74 
75 	/*
76 	 * Nodemask of nodes allowed by the caller. If NULL, all nodes
77 	 * are scanned.
78 	 */
79 	nodemask_t	*nodemask;
80 
81 	/*
82 	 * The memory cgroup that hit its limit and as a result is the
83 	 * primary target of this reclaim invocation.
84 	 */
85 	struct mem_cgroup *target_mem_cgroup;
86 
87 	/*
88 	 * Scan pressure balancing between anon and file LRUs
89 	 */
90 	unsigned long	anon_cost;
91 	unsigned long	file_cost;
92 
93 	/* Can active folios be deactivated as part of reclaim? */
94 #define DEACTIVATE_ANON 1
95 #define DEACTIVATE_FILE 2
96 	unsigned int may_deactivate:2;
97 	unsigned int force_deactivate:1;
98 	unsigned int skipped_deactivate:1;
99 
100 	/* Writepage batching in laptop mode; RECLAIM_WRITE */
101 	unsigned int may_writepage:1;
102 
103 	/* Can mapped folios be reclaimed? */
104 	unsigned int may_unmap:1;
105 
106 	/* Can folios be swapped as part of reclaim? */
107 	unsigned int may_swap:1;
108 
109 	/* Proactive reclaim invoked by userspace through memory.reclaim */
110 	unsigned int proactive:1;
111 
112 	/*
113 	 * Cgroup memory below memory.low is protected as long as we
114 	 * don't threaten to OOM. If any cgroup is reclaimed at
115 	 * reduced force or passed over entirely due to its memory.low
116 	 * setting (memcg_low_skipped), and nothing is reclaimed as a
117 	 * result, then go back for one more cycle that reclaims the protected
118 	 * memory (memcg_low_reclaim) to avert OOM.
119 	 */
120 	unsigned int memcg_low_reclaim:1;
121 	unsigned int memcg_low_skipped:1;
122 
123 	unsigned int hibernation_mode:1;
124 
125 	/* One of the zones is ready for compaction */
126 	unsigned int compaction_ready:1;
127 
128 	/* There is easily reclaimable cold cache in the current node */
129 	unsigned int cache_trim_mode:1;
130 
131 	/* The file folios on the current node are dangerously low */
132 	unsigned int file_is_tiny:1;
133 
134 	/* Always discard instead of demoting to lower tier memory */
135 	unsigned int no_demotion:1;
136 
137 #ifdef CONFIG_LRU_GEN
138 	/* help kswapd make better choices among multiple memcgs */
139 	unsigned int memcgs_need_aging:1;
140 	unsigned long last_reclaimed;
141 #endif
142 
143 	/* Allocation order */
144 	s8 order;
145 
146 	/* Scan (total_size >> priority) pages at once */
147 	s8 priority;
148 
149 	/* The highest zone to isolate folios for reclaim from */
150 	s8 reclaim_idx;
151 
152 	/* This context's GFP mask */
153 	gfp_t gfp_mask;
154 
155 	/* Incremented by the number of inactive pages that were scanned */
156 	unsigned long nr_scanned;
157 
158 	/* Number of pages freed so far during a call to shrink_zones() */
159 	unsigned long nr_reclaimed;
160 
161 	struct {
162 		unsigned int dirty;
163 		unsigned int unqueued_dirty;
164 		unsigned int congested;
165 		unsigned int writeback;
166 		unsigned int immediate;
167 		unsigned int file_taken;
168 		unsigned int taken;
169 	} nr;
170 
171 	/* for recording the reclaimed slab by now */
172 	struct reclaim_state reclaim_state;
173 };
174 
175 #ifdef ARCH_HAS_PREFETCHW
176 #define prefetchw_prev_lru_folio(_folio, _base, _field)			\
177 	do {								\
178 		if ((_folio)->lru.prev != _base) {			\
179 			struct folio *prev;				\
180 									\
181 			prev = lru_to_folio(&(_folio->lru));		\
182 			prefetchw(&prev->_field);			\
183 		}							\
184 	} while (0)
185 #else
186 #define prefetchw_prev_lru_folio(_folio, _base, _field) do { } while (0)
187 #endif
188 
189 /*
190  * From 0 .. 200.  Higher means more swappy.
191  */
192 int vm_swappiness = 60;
193 
set_task_reclaim_state(struct task_struct * task,struct reclaim_state * rs)194 static void set_task_reclaim_state(struct task_struct *task,
195 				   struct reclaim_state *rs)
196 {
197 	/* Check for an overwrite */
198 	WARN_ON_ONCE(rs && task->reclaim_state);
199 
200 	/* Check for the nulling of an already-nulled member */
201 	WARN_ON_ONCE(!rs && !task->reclaim_state);
202 
203 	task->reclaim_state = rs;
204 }
205 
206 LIST_HEAD(shrinker_list);
207 DECLARE_RWSEM(shrinker_rwsem);
208 
209 #ifdef CONFIG_MEMCG
210 static int shrinker_nr_max;
211 
212 /* The shrinker_info is expanded in a batch of BITS_PER_LONG */
shrinker_map_size(int nr_items)213 static inline int shrinker_map_size(int nr_items)
214 {
215 	return (DIV_ROUND_UP(nr_items, BITS_PER_LONG) * sizeof(unsigned long));
216 }
217 
shrinker_defer_size(int nr_items)218 static inline int shrinker_defer_size(int nr_items)
219 {
220 	return (round_up(nr_items, BITS_PER_LONG) * sizeof(atomic_long_t));
221 }
222 
shrinker_info_protected(struct mem_cgroup * memcg,int nid)223 static struct shrinker_info *shrinker_info_protected(struct mem_cgroup *memcg,
224 						     int nid)
225 {
226 	return rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_info,
227 					 lockdep_is_held(&shrinker_rwsem));
228 }
229 
expand_one_shrinker_info(struct mem_cgroup * memcg,int map_size,int defer_size,int old_map_size,int old_defer_size)230 static int expand_one_shrinker_info(struct mem_cgroup *memcg,
231 				    int map_size, int defer_size,
232 				    int old_map_size, int old_defer_size)
233 {
234 	struct shrinker_info *new, *old;
235 	struct mem_cgroup_per_node *pn;
236 	int nid;
237 	int size = map_size + defer_size;
238 
239 	for_each_node(nid) {
240 		pn = memcg->nodeinfo[nid];
241 		old = shrinker_info_protected(memcg, nid);
242 		/* Not yet online memcg */
243 		if (!old)
244 			return 0;
245 
246 		new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
247 		if (!new)
248 			return -ENOMEM;
249 
250 		new->nr_deferred = (atomic_long_t *)(new + 1);
251 		new->map = (void *)new->nr_deferred + defer_size;
252 
253 		/* map: set all old bits, clear all new bits */
254 		memset(new->map, (int)0xff, old_map_size);
255 		memset((void *)new->map + old_map_size, 0, map_size - old_map_size);
256 		/* nr_deferred: copy old values, clear all new values */
257 		memcpy(new->nr_deferred, old->nr_deferred, old_defer_size);
258 		memset((void *)new->nr_deferred + old_defer_size, 0,
259 		       defer_size - old_defer_size);
260 
261 		rcu_assign_pointer(pn->shrinker_info, new);
262 		kvfree_rcu(old, rcu);
263 	}
264 
265 	return 0;
266 }
267 
free_shrinker_info(struct mem_cgroup * memcg)268 void free_shrinker_info(struct mem_cgroup *memcg)
269 {
270 	struct mem_cgroup_per_node *pn;
271 	struct shrinker_info *info;
272 	int nid;
273 
274 	for_each_node(nid) {
275 		pn = memcg->nodeinfo[nid];
276 		info = rcu_dereference_protected(pn->shrinker_info, true);
277 		kvfree(info);
278 		rcu_assign_pointer(pn->shrinker_info, NULL);
279 	}
280 }
281 
alloc_shrinker_info(struct mem_cgroup * memcg)282 int alloc_shrinker_info(struct mem_cgroup *memcg)
283 {
284 	struct shrinker_info *info;
285 	int nid, size, ret = 0;
286 	int map_size, defer_size = 0;
287 
288 	down_write(&shrinker_rwsem);
289 	map_size = shrinker_map_size(shrinker_nr_max);
290 	defer_size = shrinker_defer_size(shrinker_nr_max);
291 	size = map_size + defer_size;
292 	for_each_node(nid) {
293 		info = kvzalloc_node(sizeof(*info) + size, GFP_KERNEL, nid);
294 		if (!info) {
295 			free_shrinker_info(memcg);
296 			ret = -ENOMEM;
297 			break;
298 		}
299 		info->nr_deferred = (atomic_long_t *)(info + 1);
300 		info->map = (void *)info->nr_deferred + defer_size;
301 		rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_info, info);
302 	}
303 	up_write(&shrinker_rwsem);
304 
305 	return ret;
306 }
307 
need_expand(int nr_max)308 static inline bool need_expand(int nr_max)
309 {
310 	return round_up(nr_max, BITS_PER_LONG) >
311 	       round_up(shrinker_nr_max, BITS_PER_LONG);
312 }
313 
expand_shrinker_info(int new_id)314 static int expand_shrinker_info(int new_id)
315 {
316 	int ret = 0;
317 	int new_nr_max = new_id + 1;
318 	int map_size, defer_size = 0;
319 	int old_map_size, old_defer_size = 0;
320 	struct mem_cgroup *memcg;
321 
322 	if (!need_expand(new_nr_max))
323 		goto out;
324 
325 	if (!root_mem_cgroup)
326 		goto out;
327 
328 	lockdep_assert_held(&shrinker_rwsem);
329 
330 	map_size = shrinker_map_size(new_nr_max);
331 	defer_size = shrinker_defer_size(new_nr_max);
332 	old_map_size = shrinker_map_size(shrinker_nr_max);
333 	old_defer_size = shrinker_defer_size(shrinker_nr_max);
334 
335 	memcg = mem_cgroup_iter(NULL, NULL, NULL);
336 	do {
337 		ret = expand_one_shrinker_info(memcg, map_size, defer_size,
338 					       old_map_size, old_defer_size);
339 		if (ret) {
340 			mem_cgroup_iter_break(NULL, memcg);
341 			goto out;
342 		}
343 	} while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
344 out:
345 	if (!ret)
346 		shrinker_nr_max = new_nr_max;
347 
348 	return ret;
349 }
350 
set_shrinker_bit(struct mem_cgroup * memcg,int nid,int shrinker_id)351 void set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
352 {
353 	if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
354 		struct shrinker_info *info;
355 
356 		rcu_read_lock();
357 		info = rcu_dereference(memcg->nodeinfo[nid]->shrinker_info);
358 		/* Pairs with smp mb in shrink_slab() */
359 		smp_mb__before_atomic();
360 		set_bit(shrinker_id, info->map);
361 		rcu_read_unlock();
362 	}
363 }
364 
365 static DEFINE_IDR(shrinker_idr);
366 
prealloc_memcg_shrinker(struct shrinker * shrinker)367 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
368 {
369 	int id, ret = -ENOMEM;
370 
371 	if (mem_cgroup_disabled())
372 		return -ENOSYS;
373 
374 	down_write(&shrinker_rwsem);
375 	/* This may call shrinker, so it must use down_read_trylock() */
376 	id = idr_alloc(&shrinker_idr, shrinker, 0, 0, GFP_KERNEL);
377 	if (id < 0)
378 		goto unlock;
379 
380 	if (id >= shrinker_nr_max) {
381 		if (expand_shrinker_info(id)) {
382 			idr_remove(&shrinker_idr, id);
383 			goto unlock;
384 		}
385 	}
386 	shrinker->id = id;
387 	ret = 0;
388 unlock:
389 	up_write(&shrinker_rwsem);
390 	return ret;
391 }
392 
unregister_memcg_shrinker(struct shrinker * shrinker)393 static void unregister_memcg_shrinker(struct shrinker *shrinker)
394 {
395 	int id = shrinker->id;
396 
397 	BUG_ON(id < 0);
398 
399 	lockdep_assert_held(&shrinker_rwsem);
400 
401 	idr_remove(&shrinker_idr, id);
402 }
403 
xchg_nr_deferred_memcg(int nid,struct shrinker * shrinker,struct mem_cgroup * memcg)404 static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
405 				   struct mem_cgroup *memcg)
406 {
407 	struct shrinker_info *info;
408 
409 	info = shrinker_info_protected(memcg, nid);
410 	return atomic_long_xchg(&info->nr_deferred[shrinker->id], 0);
411 }
412 
add_nr_deferred_memcg(long nr,int nid,struct shrinker * shrinker,struct mem_cgroup * memcg)413 static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker,
414 				  struct mem_cgroup *memcg)
415 {
416 	struct shrinker_info *info;
417 
418 	info = shrinker_info_protected(memcg, nid);
419 	return atomic_long_add_return(nr, &info->nr_deferred[shrinker->id]);
420 }
421 
reparent_shrinker_deferred(struct mem_cgroup * memcg)422 void reparent_shrinker_deferred(struct mem_cgroup *memcg)
423 {
424 	int i, nid;
425 	long nr;
426 	struct mem_cgroup *parent;
427 	struct shrinker_info *child_info, *parent_info;
428 
429 	parent = parent_mem_cgroup(memcg);
430 	if (!parent)
431 		parent = root_mem_cgroup;
432 
433 	/* Prevent from concurrent shrinker_info expand */
434 	down_read(&shrinker_rwsem);
435 	for_each_node(nid) {
436 		child_info = shrinker_info_protected(memcg, nid);
437 		parent_info = shrinker_info_protected(parent, nid);
438 		for (i = 0; i < shrinker_nr_max; i++) {
439 			nr = atomic_long_read(&child_info->nr_deferred[i]);
440 			atomic_long_add(nr, &parent_info->nr_deferred[i]);
441 		}
442 	}
443 	up_read(&shrinker_rwsem);
444 }
445 
cgroup_reclaim(struct scan_control * sc)446 static bool cgroup_reclaim(struct scan_control *sc)
447 {
448 	return sc->target_mem_cgroup;
449 }
450 
451 /**
452  * writeback_throttling_sane - is the usual dirty throttling mechanism available?
453  * @sc: scan_control in question
454  *
455  * The normal page dirty throttling mechanism in balance_dirty_pages() is
456  * completely broken with the legacy memcg and direct stalling in
457  * shrink_folio_list() is used for throttling instead, which lacks all the
458  * niceties such as fairness, adaptive pausing, bandwidth proportional
459  * allocation and configurability.
460  *
461  * This function tests whether the vmscan currently in progress can assume
462  * that the normal dirty throttling mechanism is operational.
463  */
writeback_throttling_sane(struct scan_control * sc)464 static bool writeback_throttling_sane(struct scan_control *sc)
465 {
466 	if (!cgroup_reclaim(sc))
467 		return true;
468 #ifdef CONFIG_CGROUP_WRITEBACK
469 	if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
470 		return true;
471 #endif
472 	return false;
473 }
474 #else
prealloc_memcg_shrinker(struct shrinker * shrinker)475 static int prealloc_memcg_shrinker(struct shrinker *shrinker)
476 {
477 	return -ENOSYS;
478 }
479 
unregister_memcg_shrinker(struct shrinker * shrinker)480 static void unregister_memcg_shrinker(struct shrinker *shrinker)
481 {
482 }
483 
xchg_nr_deferred_memcg(int nid,struct shrinker * shrinker,struct mem_cgroup * memcg)484 static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker,
485 				   struct mem_cgroup *memcg)
486 {
487 	return 0;
488 }
489 
add_nr_deferred_memcg(long nr,int nid,struct shrinker * shrinker,struct mem_cgroup * memcg)490 static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker,
491 				  struct mem_cgroup *memcg)
492 {
493 	return 0;
494 }
495 
cgroup_reclaim(struct scan_control * sc)496 static bool cgroup_reclaim(struct scan_control *sc)
497 {
498 	return false;
499 }
500 
writeback_throttling_sane(struct scan_control * sc)501 static bool writeback_throttling_sane(struct scan_control *sc)
502 {
503 	return true;
504 }
505 #endif
506 
xchg_nr_deferred(struct shrinker * shrinker,struct shrink_control * sc)507 static long xchg_nr_deferred(struct shrinker *shrinker,
508 			     struct shrink_control *sc)
509 {
510 	int nid = sc->nid;
511 
512 	if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
513 		nid = 0;
514 
515 	if (sc->memcg &&
516 	    (shrinker->flags & SHRINKER_MEMCG_AWARE))
517 		return xchg_nr_deferred_memcg(nid, shrinker,
518 					      sc->memcg);
519 
520 	return atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
521 }
522 
523 
add_nr_deferred(long nr,struct shrinker * shrinker,struct shrink_control * sc)524 static long add_nr_deferred(long nr, struct shrinker *shrinker,
525 			    struct shrink_control *sc)
526 {
527 	int nid = sc->nid;
528 
529 	if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
530 		nid = 0;
531 
532 	if (sc->memcg &&
533 	    (shrinker->flags & SHRINKER_MEMCG_AWARE))
534 		return add_nr_deferred_memcg(nr, nid, shrinker,
535 					     sc->memcg);
536 
537 	return atomic_long_add_return(nr, &shrinker->nr_deferred[nid]);
538 }
539 
can_demote(int nid,struct scan_control * sc)540 static bool can_demote(int nid, struct scan_control *sc)
541 {
542 	if (!numa_demotion_enabled)
543 		return false;
544 	if (sc && sc->no_demotion)
545 		return false;
546 	if (next_demotion_node(nid) == NUMA_NO_NODE)
547 		return false;
548 
549 	return true;
550 }
551 
can_reclaim_anon_pages(struct mem_cgroup * memcg,int nid,struct scan_control * sc)552 static inline bool can_reclaim_anon_pages(struct mem_cgroup *memcg,
553 					  int nid,
554 					  struct scan_control *sc)
555 {
556 	if (memcg == NULL) {
557 		/*
558 		 * For non-memcg reclaim, is there
559 		 * space in any swap device?
560 		 */
561 		if (get_nr_swap_pages() > 0)
562 			return true;
563 	} else {
564 		/* Is the memcg below its swap limit? */
565 		if (mem_cgroup_get_nr_swap_pages(memcg) > 0)
566 			return true;
567 	}
568 
569 	/*
570 	 * The page can not be swapped.
571 	 *
572 	 * Can it be reclaimed from this node via demotion?
573 	 */
574 	return can_demote(nid, sc);
575 }
576 
577 /*
578  * This misses isolated folios which are not accounted for to save counters.
579  * As the data only determines if reclaim or compaction continues, it is
580  * not expected that isolated folios will be a dominating factor.
581  */
zone_reclaimable_pages(struct zone * zone)582 unsigned long zone_reclaimable_pages(struct zone *zone)
583 {
584 	unsigned long nr;
585 
586 	nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
587 		zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
588 	if (can_reclaim_anon_pages(NULL, zone_to_nid(zone), NULL))
589 		nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
590 			zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
591 
592 	return nr;
593 }
594 
595 /**
596  * lruvec_lru_size -  Returns the number of pages on the given LRU list.
597  * @lruvec: lru vector
598  * @lru: lru to use
599  * @zone_idx: zones to consider (use MAX_NR_ZONES - 1 for the whole LRU list)
600  */
lruvec_lru_size(struct lruvec * lruvec,enum lru_list lru,int zone_idx)601 static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru,
602 				     int zone_idx)
603 {
604 	unsigned long size = 0;
605 	int zid;
606 
607 	for (zid = 0; zid <= zone_idx; zid++) {
608 		struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
609 
610 		if (!managed_zone(zone))
611 			continue;
612 
613 		if (!mem_cgroup_disabled())
614 			size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
615 		else
616 			size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru);
617 	}
618 	return size;
619 }
620 
621 /*
622  * Add a shrinker callback to be called from the vm.
623  */
__prealloc_shrinker(struct shrinker * shrinker)624 static int __prealloc_shrinker(struct shrinker *shrinker)
625 {
626 	unsigned int size;
627 	int err;
628 
629 	if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
630 		err = prealloc_memcg_shrinker(shrinker);
631 		if (err != -ENOSYS)
632 			return err;
633 
634 		shrinker->flags &= ~SHRINKER_MEMCG_AWARE;
635 	}
636 
637 	size = sizeof(*shrinker->nr_deferred);
638 	if (shrinker->flags & SHRINKER_NUMA_AWARE)
639 		size *= nr_node_ids;
640 
641 	shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
642 	if (!shrinker->nr_deferred)
643 		return -ENOMEM;
644 
645 	return 0;
646 }
647 
648 #ifdef CONFIG_SHRINKER_DEBUG
prealloc_shrinker(struct shrinker * shrinker,const char * fmt,...)649 int prealloc_shrinker(struct shrinker *shrinker, const char *fmt, ...)
650 {
651 	va_list ap;
652 	int err;
653 
654 	va_start(ap, fmt);
655 	shrinker->name = kvasprintf_const(GFP_KERNEL, fmt, ap);
656 	va_end(ap);
657 	if (!shrinker->name)
658 		return -ENOMEM;
659 
660 	err = __prealloc_shrinker(shrinker);
661 	if (err) {
662 		kfree_const(shrinker->name);
663 		shrinker->name = NULL;
664 	}
665 
666 	return err;
667 }
668 #else
prealloc_shrinker(struct shrinker * shrinker,const char * fmt,...)669 int prealloc_shrinker(struct shrinker *shrinker, const char *fmt, ...)
670 {
671 	return __prealloc_shrinker(shrinker);
672 }
673 #endif
674 
free_prealloced_shrinker(struct shrinker * shrinker)675 void free_prealloced_shrinker(struct shrinker *shrinker)
676 {
677 #ifdef CONFIG_SHRINKER_DEBUG
678 	kfree_const(shrinker->name);
679 	shrinker->name = NULL;
680 #endif
681 	if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
682 		down_write(&shrinker_rwsem);
683 		unregister_memcg_shrinker(shrinker);
684 		up_write(&shrinker_rwsem);
685 		return;
686 	}
687 
688 	kfree(shrinker->nr_deferred);
689 	shrinker->nr_deferred = NULL;
690 }
691 
register_shrinker_prepared(struct shrinker * shrinker)692 void register_shrinker_prepared(struct shrinker *shrinker)
693 {
694 	down_write(&shrinker_rwsem);
695 	list_add_tail(&shrinker->list, &shrinker_list);
696 	shrinker->flags |= SHRINKER_REGISTERED;
697 	shrinker_debugfs_add(shrinker);
698 	up_write(&shrinker_rwsem);
699 }
700 
__register_shrinker(struct shrinker * shrinker)701 static int __register_shrinker(struct shrinker *shrinker)
702 {
703 	int err = __prealloc_shrinker(shrinker);
704 
705 	if (err)
706 		return err;
707 	register_shrinker_prepared(shrinker);
708 	return 0;
709 }
710 
711 #ifdef CONFIG_SHRINKER_DEBUG
register_shrinker(struct shrinker * shrinker,const char * fmt,...)712 int register_shrinker(struct shrinker *shrinker, const char *fmt, ...)
713 {
714 	va_list ap;
715 	int err;
716 
717 	va_start(ap, fmt);
718 	shrinker->name = kvasprintf_const(GFP_KERNEL, fmt, ap);
719 	va_end(ap);
720 	if (!shrinker->name)
721 		return -ENOMEM;
722 
723 	err = __register_shrinker(shrinker);
724 	if (err) {
725 		kfree_const(shrinker->name);
726 		shrinker->name = NULL;
727 	}
728 	return err;
729 }
730 #else
register_shrinker(struct shrinker * shrinker,const char * fmt,...)731 int register_shrinker(struct shrinker *shrinker, const char *fmt, ...)
732 {
733 	return __register_shrinker(shrinker);
734 }
735 #endif
736 EXPORT_SYMBOL(register_shrinker);
737 
738 /*
739  * Remove one
740  */
unregister_shrinker(struct shrinker * shrinker)741 void unregister_shrinker(struct shrinker *shrinker)
742 {
743 	if (!(shrinker->flags & SHRINKER_REGISTERED))
744 		return;
745 
746 	down_write(&shrinker_rwsem);
747 	list_del(&shrinker->list);
748 	shrinker->flags &= ~SHRINKER_REGISTERED;
749 	if (shrinker->flags & SHRINKER_MEMCG_AWARE)
750 		unregister_memcg_shrinker(shrinker);
751 	shrinker_debugfs_remove(shrinker);
752 	up_write(&shrinker_rwsem);
753 
754 	kfree(shrinker->nr_deferred);
755 	shrinker->nr_deferred = NULL;
756 }
757 EXPORT_SYMBOL(unregister_shrinker);
758 
759 /**
760  * synchronize_shrinkers - Wait for all running shrinkers to complete.
761  *
762  * This is equivalent to calling unregister_shrink() and register_shrinker(),
763  * but atomically and with less overhead. This is useful to guarantee that all
764  * shrinker invocations have seen an update, before freeing memory, similar to
765  * rcu.
766  */
synchronize_shrinkers(void)767 void synchronize_shrinkers(void)
768 {
769 	down_write(&shrinker_rwsem);
770 	up_write(&shrinker_rwsem);
771 }
772 EXPORT_SYMBOL(synchronize_shrinkers);
773 
774 #define SHRINK_BATCH 128
775 
do_shrink_slab(struct shrink_control * shrinkctl,struct shrinker * shrinker,int priority)776 static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
777 				    struct shrinker *shrinker, int priority)
778 {
779 	unsigned long freed = 0;
780 	unsigned long long delta;
781 	long total_scan;
782 	long freeable;
783 	long nr;
784 	long new_nr;
785 	long batch_size = shrinker->batch ? shrinker->batch
786 					  : SHRINK_BATCH;
787 	long scanned = 0, next_deferred;
788 
789 	freeable = shrinker->count_objects(shrinker, shrinkctl);
790 	if (freeable == 0 || freeable == SHRINK_EMPTY)
791 		return freeable;
792 
793 	/*
794 	 * copy the current shrinker scan count into a local variable
795 	 * and zero it so that other concurrent shrinker invocations
796 	 * don't also do this scanning work.
797 	 */
798 	nr = xchg_nr_deferred(shrinker, shrinkctl);
799 
800 	if (shrinker->seeks) {
801 		delta = freeable >> priority;
802 		delta *= 4;
803 		do_div(delta, shrinker->seeks);
804 	} else {
805 		/*
806 		 * These objects don't require any IO to create. Trim
807 		 * them aggressively under memory pressure to keep
808 		 * them from causing refetches in the IO caches.
809 		 */
810 		delta = freeable / 2;
811 	}
812 
813 	total_scan = nr >> priority;
814 	total_scan += delta;
815 	total_scan = min(total_scan, (2 * freeable));
816 
817 	trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
818 				   freeable, delta, total_scan, priority);
819 
820 	/*
821 	 * Normally, we should not scan less than batch_size objects in one
822 	 * pass to avoid too frequent shrinker calls, but if the slab has less
823 	 * than batch_size objects in total and we are really tight on memory,
824 	 * we will try to reclaim all available objects, otherwise we can end
825 	 * up failing allocations although there are plenty of reclaimable
826 	 * objects spread over several slabs with usage less than the
827 	 * batch_size.
828 	 *
829 	 * We detect the "tight on memory" situations by looking at the total
830 	 * number of objects we want to scan (total_scan). If it is greater
831 	 * than the total number of objects on slab (freeable), we must be
832 	 * scanning at high prio and therefore should try to reclaim as much as
833 	 * possible.
834 	 */
835 	while (total_scan >= batch_size ||
836 	       total_scan >= freeable) {
837 		unsigned long ret;
838 		unsigned long nr_to_scan = min(batch_size, total_scan);
839 
840 		shrinkctl->nr_to_scan = nr_to_scan;
841 		shrinkctl->nr_scanned = nr_to_scan;
842 		ret = shrinker->scan_objects(shrinker, shrinkctl);
843 		if (ret == SHRINK_STOP)
844 			break;
845 		freed += ret;
846 
847 		count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
848 		total_scan -= shrinkctl->nr_scanned;
849 		scanned += shrinkctl->nr_scanned;
850 
851 		cond_resched();
852 	}
853 
854 	/*
855 	 * The deferred work is increased by any new work (delta) that wasn't
856 	 * done, decreased by old deferred work that was done now.
857 	 *
858 	 * And it is capped to two times of the freeable items.
859 	 */
860 	next_deferred = max_t(long, (nr + delta - scanned), 0);
861 	next_deferred = min(next_deferred, (2 * freeable));
862 
863 	/*
864 	 * move the unused scan count back into the shrinker in a
865 	 * manner that handles concurrent updates.
866 	 */
867 	new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl);
868 
869 	trace_mm_shrink_slab_end(shrinker, shrinkctl->nid, freed, nr, new_nr, total_scan);
870 	return freed;
871 }
872 
873 #ifdef CONFIG_MEMCG
shrink_slab_memcg(gfp_t gfp_mask,int nid,struct mem_cgroup * memcg,int priority)874 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
875 			struct mem_cgroup *memcg, int priority)
876 {
877 	struct shrinker_info *info;
878 	unsigned long ret, freed = 0;
879 	int i;
880 
881 	if (!mem_cgroup_online(memcg))
882 		return 0;
883 
884 	if (!down_read_trylock(&shrinker_rwsem))
885 		return 0;
886 
887 	info = shrinker_info_protected(memcg, nid);
888 	if (unlikely(!info))
889 		goto unlock;
890 
891 	for_each_set_bit(i, info->map, shrinker_nr_max) {
892 		struct shrink_control sc = {
893 			.gfp_mask = gfp_mask,
894 			.nid = nid,
895 			.memcg = memcg,
896 		};
897 		struct shrinker *shrinker;
898 
899 		shrinker = idr_find(&shrinker_idr, i);
900 		if (unlikely(!shrinker || !(shrinker->flags & SHRINKER_REGISTERED))) {
901 			if (!shrinker)
902 				clear_bit(i, info->map);
903 			continue;
904 		}
905 
906 		/* Call non-slab shrinkers even though kmem is disabled */
907 		if (!memcg_kmem_enabled() &&
908 		    !(shrinker->flags & SHRINKER_NONSLAB))
909 			continue;
910 
911 		ret = do_shrink_slab(&sc, shrinker, priority);
912 		if (ret == SHRINK_EMPTY) {
913 			clear_bit(i, info->map);
914 			/*
915 			 * After the shrinker reported that it had no objects to
916 			 * free, but before we cleared the corresponding bit in
917 			 * the memcg shrinker map, a new object might have been
918 			 * added. To make sure, we have the bit set in this
919 			 * case, we invoke the shrinker one more time and reset
920 			 * the bit if it reports that it is not empty anymore.
921 			 * The memory barrier here pairs with the barrier in
922 			 * set_shrinker_bit():
923 			 *
924 			 * list_lru_add()     shrink_slab_memcg()
925 			 *   list_add_tail()    clear_bit()
926 			 *   <MB>               <MB>
927 			 *   set_bit()          do_shrink_slab()
928 			 */
929 			smp_mb__after_atomic();
930 			ret = do_shrink_slab(&sc, shrinker, priority);
931 			if (ret == SHRINK_EMPTY)
932 				ret = 0;
933 			else
934 				set_shrinker_bit(memcg, nid, i);
935 		}
936 		freed += ret;
937 
938 		if (rwsem_is_contended(&shrinker_rwsem)) {
939 			freed = freed ? : 1;
940 			break;
941 		}
942 	}
943 unlock:
944 	up_read(&shrinker_rwsem);
945 	return freed;
946 }
947 #else /* CONFIG_MEMCG */
shrink_slab_memcg(gfp_t gfp_mask,int nid,struct mem_cgroup * memcg,int priority)948 static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
949 			struct mem_cgroup *memcg, int priority)
950 {
951 	return 0;
952 }
953 #endif /* CONFIG_MEMCG */
954 
955 /**
956  * shrink_slab - shrink slab caches
957  * @gfp_mask: allocation context
958  * @nid: node whose slab caches to target
959  * @memcg: memory cgroup whose slab caches to target
960  * @priority: the reclaim priority
961  *
962  * Call the shrink functions to age shrinkable caches.
963  *
964  * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
965  * unaware shrinkers will receive a node id of 0 instead.
966  *
967  * @memcg specifies the memory cgroup to target. Unaware shrinkers
968  * are called only if it is the root cgroup.
969  *
970  * @priority is sc->priority, we take the number of objects and >> by priority
971  * in order to get the scan target.
972  *
973  * Returns the number of reclaimed slab objects.
974  */
shrink_slab(gfp_t gfp_mask,int nid,struct mem_cgroup * memcg,int priority)975 static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
976 				 struct mem_cgroup *memcg,
977 				 int priority)
978 {
979 	unsigned long ret, freed = 0;
980 	struct shrinker *shrinker;
981 
982 	/*
983 	 * The root memcg might be allocated even though memcg is disabled
984 	 * via "cgroup_disable=memory" boot parameter.  This could make
985 	 * mem_cgroup_is_root() return false, then just run memcg slab
986 	 * shrink, but skip global shrink.  This may result in premature
987 	 * oom.
988 	 */
989 	if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
990 		return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
991 
992 	if (!down_read_trylock(&shrinker_rwsem))
993 		goto out;
994 
995 	list_for_each_entry(shrinker, &shrinker_list, list) {
996 		struct shrink_control sc = {
997 			.gfp_mask = gfp_mask,
998 			.nid = nid,
999 			.memcg = memcg,
1000 		};
1001 
1002 		ret = do_shrink_slab(&sc, shrinker, priority);
1003 		if (ret == SHRINK_EMPTY)
1004 			ret = 0;
1005 		freed += ret;
1006 		/*
1007 		 * Bail out if someone want to register a new shrinker to
1008 		 * prevent the registration from being stalled for long periods
1009 		 * by parallel ongoing shrinking.
1010 		 */
1011 		if (rwsem_is_contended(&shrinker_rwsem)) {
1012 			freed = freed ? : 1;
1013 			break;
1014 		}
1015 	}
1016 
1017 	up_read(&shrinker_rwsem);
1018 out:
1019 	cond_resched();
1020 	return freed;
1021 }
1022 
drop_slab_node(int nid)1023 static void drop_slab_node(int nid)
1024 {
1025 	unsigned long freed;
1026 	int shift = 0;
1027 
1028 	do {
1029 		struct mem_cgroup *memcg = NULL;
1030 
1031 		if (fatal_signal_pending(current))
1032 			return;
1033 
1034 		freed = 0;
1035 		memcg = mem_cgroup_iter(NULL, NULL, NULL);
1036 		do {
1037 			freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
1038 		} while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
1039 	} while ((freed >> shift++) > 1);
1040 }
1041 
drop_slab(void)1042 void drop_slab(void)
1043 {
1044 	int nid;
1045 
1046 	for_each_online_node(nid)
1047 		drop_slab_node(nid);
1048 }
1049 
is_page_cache_freeable(struct folio * folio)1050 static inline int is_page_cache_freeable(struct folio *folio)
1051 {
1052 	/*
1053 	 * A freeable page cache folio is referenced only by the caller
1054 	 * that isolated the folio, the page cache and optional filesystem
1055 	 * private data at folio->private.
1056 	 */
1057 	return folio_ref_count(folio) - folio_test_private(folio) ==
1058 		1 + folio_nr_pages(folio);
1059 }
1060 
1061 /*
1062  * We detected a synchronous write error writing a folio out.  Probably
1063  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
1064  * fsync(), msync() or close().
1065  *
1066  * The tricky part is that after writepage we cannot touch the mapping: nothing
1067  * prevents it from being freed up.  But we have a ref on the folio and once
1068  * that folio is locked, the mapping is pinned.
1069  *
1070  * We're allowed to run sleeping folio_lock() here because we know the caller has
1071  * __GFP_FS.
1072  */
handle_write_error(struct address_space * mapping,struct folio * folio,int error)1073 static void handle_write_error(struct address_space *mapping,
1074 				struct folio *folio, int error)
1075 {
1076 	folio_lock(folio);
1077 	if (folio_mapping(folio) == mapping)
1078 		mapping_set_error(mapping, error);
1079 	folio_unlock(folio);
1080 }
1081 
skip_throttle_noprogress(pg_data_t * pgdat)1082 static bool skip_throttle_noprogress(pg_data_t *pgdat)
1083 {
1084 	int reclaimable = 0, write_pending = 0;
1085 	int i;
1086 
1087 	/*
1088 	 * If kswapd is disabled, reschedule if necessary but do not
1089 	 * throttle as the system is likely near OOM.
1090 	 */
1091 	if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
1092 		return true;
1093 
1094 	/*
1095 	 * If there are a lot of dirty/writeback folios then do not
1096 	 * throttle as throttling will occur when the folios cycle
1097 	 * towards the end of the LRU if still under writeback.
1098 	 */
1099 	for (i = 0; i < MAX_NR_ZONES; i++) {
1100 		struct zone *zone = pgdat->node_zones + i;
1101 
1102 		if (!managed_zone(zone))
1103 			continue;
1104 
1105 		reclaimable += zone_reclaimable_pages(zone);
1106 		write_pending += zone_page_state_snapshot(zone,
1107 						  NR_ZONE_WRITE_PENDING);
1108 	}
1109 	if (2 * write_pending <= reclaimable)
1110 		return true;
1111 
1112 	return false;
1113 }
1114 
reclaim_throttle(pg_data_t * pgdat,enum vmscan_throttle_state reason)1115 void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason)
1116 {
1117 	wait_queue_head_t *wqh = &pgdat->reclaim_wait[reason];
1118 	long timeout, ret;
1119 	DEFINE_WAIT(wait);
1120 
1121 	/*
1122 	 * Do not throttle IO workers, kthreads other than kswapd or
1123 	 * workqueues. They may be required for reclaim to make
1124 	 * forward progress (e.g. journalling workqueues or kthreads).
1125 	 */
1126 	if (!current_is_kswapd() &&
1127 	    current->flags & (PF_IO_WORKER|PF_KTHREAD)) {
1128 		cond_resched();
1129 		return;
1130 	}
1131 
1132 	/*
1133 	 * These figures are pulled out of thin air.
1134 	 * VMSCAN_THROTTLE_ISOLATED is a transient condition based on too many
1135 	 * parallel reclaimers which is a short-lived event so the timeout is
1136 	 * short. Failing to make progress or waiting on writeback are
1137 	 * potentially long-lived events so use a longer timeout. This is shaky
1138 	 * logic as a failure to make progress could be due to anything from
1139 	 * writeback to a slow device to excessive referenced folios at the tail
1140 	 * of the inactive LRU.
1141 	 */
1142 	switch(reason) {
1143 	case VMSCAN_THROTTLE_WRITEBACK:
1144 		timeout = HZ/10;
1145 
1146 		if (atomic_inc_return(&pgdat->nr_writeback_throttled) == 1) {
1147 			WRITE_ONCE(pgdat->nr_reclaim_start,
1148 				node_page_state(pgdat, NR_THROTTLED_WRITTEN));
1149 		}
1150 
1151 		break;
1152 	case VMSCAN_THROTTLE_CONGESTED:
1153 		fallthrough;
1154 	case VMSCAN_THROTTLE_NOPROGRESS:
1155 		if (skip_throttle_noprogress(pgdat)) {
1156 			cond_resched();
1157 			return;
1158 		}
1159 
1160 		timeout = 1;
1161 
1162 		break;
1163 	case VMSCAN_THROTTLE_ISOLATED:
1164 		timeout = HZ/50;
1165 		break;
1166 	default:
1167 		WARN_ON_ONCE(1);
1168 		timeout = HZ;
1169 		break;
1170 	}
1171 
1172 	prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1173 	ret = schedule_timeout(timeout);
1174 	finish_wait(wqh, &wait);
1175 
1176 	if (reason == VMSCAN_THROTTLE_WRITEBACK)
1177 		atomic_dec(&pgdat->nr_writeback_throttled);
1178 
1179 	trace_mm_vmscan_throttled(pgdat->node_id, jiffies_to_usecs(timeout),
1180 				jiffies_to_usecs(timeout - ret),
1181 				reason);
1182 }
1183 
1184 /*
1185  * Account for folios written if tasks are throttled waiting on dirty
1186  * folios to clean. If enough folios have been cleaned since throttling
1187  * started then wakeup the throttled tasks.
1188  */
__acct_reclaim_writeback(pg_data_t * pgdat,struct folio * folio,int nr_throttled)1189 void __acct_reclaim_writeback(pg_data_t *pgdat, struct folio *folio,
1190 							int nr_throttled)
1191 {
1192 	unsigned long nr_written;
1193 
1194 	node_stat_add_folio(folio, NR_THROTTLED_WRITTEN);
1195 
1196 	/*
1197 	 * This is an inaccurate read as the per-cpu deltas may not
1198 	 * be synchronised. However, given that the system is
1199 	 * writeback throttled, it is not worth taking the penalty
1200 	 * of getting an accurate count. At worst, the throttle
1201 	 * timeout guarantees forward progress.
1202 	 */
1203 	nr_written = node_page_state(pgdat, NR_THROTTLED_WRITTEN) -
1204 		READ_ONCE(pgdat->nr_reclaim_start);
1205 
1206 	if (nr_written > SWAP_CLUSTER_MAX * nr_throttled)
1207 		wake_up(&pgdat->reclaim_wait[VMSCAN_THROTTLE_WRITEBACK]);
1208 }
1209 
1210 /* possible outcome of pageout() */
1211 typedef enum {
1212 	/* failed to write folio out, folio is locked */
1213 	PAGE_KEEP,
1214 	/* move folio to the active list, folio is locked */
1215 	PAGE_ACTIVATE,
1216 	/* folio has been sent to the disk successfully, folio is unlocked */
1217 	PAGE_SUCCESS,
1218 	/* folio is clean and locked */
1219 	PAGE_CLEAN,
1220 } pageout_t;
1221 
1222 /*
1223  * pageout is called by shrink_folio_list() for each dirty folio.
1224  * Calls ->writepage().
1225  */
pageout(struct folio * folio,struct address_space * mapping,struct swap_iocb ** plug)1226 static pageout_t pageout(struct folio *folio, struct address_space *mapping,
1227 			 struct swap_iocb **plug)
1228 {
1229 	/*
1230 	 * If the folio is dirty, only perform writeback if that write
1231 	 * will be non-blocking.  To prevent this allocation from being
1232 	 * stalled by pagecache activity.  But note that there may be
1233 	 * stalls if we need to run get_block().  We could test
1234 	 * PagePrivate for that.
1235 	 *
1236 	 * If this process is currently in __generic_file_write_iter() against
1237 	 * this folio's queue, we can perform writeback even if that
1238 	 * will block.
1239 	 *
1240 	 * If the folio is swapcache, write it back even if that would
1241 	 * block, for some throttling. This happens by accident, because
1242 	 * swap_backing_dev_info is bust: it doesn't reflect the
1243 	 * congestion state of the swapdevs.  Easy to fix, if needed.
1244 	 */
1245 	if (!is_page_cache_freeable(folio))
1246 		return PAGE_KEEP;
1247 	if (!mapping) {
1248 		/*
1249 		 * Some data journaling orphaned folios can have
1250 		 * folio->mapping == NULL while being dirty with clean buffers.
1251 		 */
1252 		if (folio_test_private(folio)) {
1253 			if (try_to_free_buffers(folio)) {
1254 				folio_clear_dirty(folio);
1255 				pr_info("%s: orphaned folio\n", __func__);
1256 				return PAGE_CLEAN;
1257 			}
1258 		}
1259 		return PAGE_KEEP;
1260 	}
1261 	if (mapping->a_ops->writepage == NULL)
1262 		return PAGE_ACTIVATE;
1263 
1264 	if (folio_clear_dirty_for_io(folio)) {
1265 		int res;
1266 		struct writeback_control wbc = {
1267 			.sync_mode = WB_SYNC_NONE,
1268 			.nr_to_write = SWAP_CLUSTER_MAX,
1269 			.range_start = 0,
1270 			.range_end = LLONG_MAX,
1271 			.for_reclaim = 1,
1272 			.swap_plug = plug,
1273 		};
1274 
1275 		folio_set_reclaim(folio);
1276 		res = mapping->a_ops->writepage(&folio->page, &wbc);
1277 		if (res < 0)
1278 			handle_write_error(mapping, folio, res);
1279 		if (res == AOP_WRITEPAGE_ACTIVATE) {
1280 			folio_clear_reclaim(folio);
1281 			return PAGE_ACTIVATE;
1282 		}
1283 
1284 		if (!folio_test_writeback(folio)) {
1285 			/* synchronous write or broken a_ops? */
1286 			folio_clear_reclaim(folio);
1287 		}
1288 		trace_mm_vmscan_write_folio(folio);
1289 		node_stat_add_folio(folio, NR_VMSCAN_WRITE);
1290 		return PAGE_SUCCESS;
1291 	}
1292 
1293 	return PAGE_CLEAN;
1294 }
1295 
1296 /*
1297  * Same as remove_mapping, but if the folio is removed from the mapping, it
1298  * gets returned with a refcount of 0.
1299  */
__remove_mapping(struct address_space * mapping,struct folio * folio,bool reclaimed,struct mem_cgroup * target_memcg)1300 static int __remove_mapping(struct address_space *mapping, struct folio *folio,
1301 			    bool reclaimed, struct mem_cgroup *target_memcg)
1302 {
1303 	int refcount;
1304 	void *shadow = NULL;
1305 
1306 	BUG_ON(!folio_test_locked(folio));
1307 	BUG_ON(mapping != folio_mapping(folio));
1308 
1309 	if (!folio_test_swapcache(folio))
1310 		spin_lock(&mapping->host->i_lock);
1311 	xa_lock_irq(&mapping->i_pages);
1312 	/*
1313 	 * The non racy check for a busy folio.
1314 	 *
1315 	 * Must be careful with the order of the tests. When someone has
1316 	 * a ref to the folio, it may be possible that they dirty it then
1317 	 * drop the reference. So if the dirty flag is tested before the
1318 	 * refcount here, then the following race may occur:
1319 	 *
1320 	 * get_user_pages(&page);
1321 	 * [user mapping goes away]
1322 	 * write_to(page);
1323 	 *				!folio_test_dirty(folio)    [good]
1324 	 * folio_set_dirty(folio);
1325 	 * folio_put(folio);
1326 	 *				!refcount(folio)   [good, discard it]
1327 	 *
1328 	 * [oops, our write_to data is lost]
1329 	 *
1330 	 * Reversing the order of the tests ensures such a situation cannot
1331 	 * escape unnoticed. The smp_rmb is needed to ensure the folio->flags
1332 	 * load is not satisfied before that of folio->_refcount.
1333 	 *
1334 	 * Note that if the dirty flag is always set via folio_mark_dirty,
1335 	 * and thus under the i_pages lock, then this ordering is not required.
1336 	 */
1337 	refcount = 1 + folio_nr_pages(folio);
1338 	if (!folio_ref_freeze(folio, refcount))
1339 		goto cannot_free;
1340 	/* note: atomic_cmpxchg in folio_ref_freeze provides the smp_rmb */
1341 	if (unlikely(folio_test_dirty(folio))) {
1342 		folio_ref_unfreeze(folio, refcount);
1343 		goto cannot_free;
1344 	}
1345 
1346 	if (folio_test_swapcache(folio)) {
1347 		swp_entry_t swap = folio_swap_entry(folio);
1348 
1349 		/* get a shadow entry before mem_cgroup_swapout() clears folio_memcg() */
1350 		if (reclaimed && !mapping_exiting(mapping))
1351 			shadow = workingset_eviction(folio, target_memcg);
1352 		mem_cgroup_swapout(folio, swap);
1353 		__delete_from_swap_cache(folio, swap, shadow);
1354 		xa_unlock_irq(&mapping->i_pages);
1355 		put_swap_folio(folio, swap);
1356 	} else {
1357 		void (*free_folio)(struct folio *);
1358 
1359 		free_folio = mapping->a_ops->free_folio;
1360 		/*
1361 		 * Remember a shadow entry for reclaimed file cache in
1362 		 * order to detect refaults, thus thrashing, later on.
1363 		 *
1364 		 * But don't store shadows in an address space that is
1365 		 * already exiting.  This is not just an optimization,
1366 		 * inode reclaim needs to empty out the radix tree or
1367 		 * the nodes are lost.  Don't plant shadows behind its
1368 		 * back.
1369 		 *
1370 		 * We also don't store shadows for DAX mappings because the
1371 		 * only page cache folios found in these are zero pages
1372 		 * covering holes, and because we don't want to mix DAX
1373 		 * exceptional entries and shadow exceptional entries in the
1374 		 * same address_space.
1375 		 */
1376 		if (reclaimed && folio_is_file_lru(folio) &&
1377 		    !mapping_exiting(mapping) && !dax_mapping(mapping))
1378 			shadow = workingset_eviction(folio, target_memcg);
1379 		__filemap_remove_folio(folio, shadow);
1380 		xa_unlock_irq(&mapping->i_pages);
1381 		if (mapping_shrinkable(mapping))
1382 			inode_add_lru(mapping->host);
1383 		spin_unlock(&mapping->host->i_lock);
1384 
1385 		if (free_folio)
1386 			free_folio(folio);
1387 	}
1388 
1389 	return 1;
1390 
1391 cannot_free:
1392 	xa_unlock_irq(&mapping->i_pages);
1393 	if (!folio_test_swapcache(folio))
1394 		spin_unlock(&mapping->host->i_lock);
1395 	return 0;
1396 }
1397 
1398 /**
1399  * remove_mapping() - Attempt to remove a folio from its mapping.
1400  * @mapping: The address space.
1401  * @folio: The folio to remove.
1402  *
1403  * If the folio is dirty, under writeback or if someone else has a ref
1404  * on it, removal will fail.
1405  * Return: The number of pages removed from the mapping.  0 if the folio
1406  * could not be removed.
1407  * Context: The caller should have a single refcount on the folio and
1408  * hold its lock.
1409  */
remove_mapping(struct address_space * mapping,struct folio * folio)1410 long remove_mapping(struct address_space *mapping, struct folio *folio)
1411 {
1412 	if (__remove_mapping(mapping, folio, false, NULL)) {
1413 		/*
1414 		 * Unfreezing the refcount with 1 effectively
1415 		 * drops the pagecache ref for us without requiring another
1416 		 * atomic operation.
1417 		 */
1418 		folio_ref_unfreeze(folio, 1);
1419 		return folio_nr_pages(folio);
1420 	}
1421 	return 0;
1422 }
1423 
1424 /**
1425  * folio_putback_lru - Put previously isolated folio onto appropriate LRU list.
1426  * @folio: Folio to be returned to an LRU list.
1427  *
1428  * Add previously isolated @folio to appropriate LRU list.
1429  * The folio may still be unevictable for other reasons.
1430  *
1431  * Context: lru_lock must not be held, interrupts must be enabled.
1432  */
folio_putback_lru(struct folio * folio)1433 void folio_putback_lru(struct folio *folio)
1434 {
1435 	folio_add_lru(folio);
1436 	folio_put(folio);		/* drop ref from isolate */
1437 }
1438 
1439 enum folio_references {
1440 	FOLIOREF_RECLAIM,
1441 	FOLIOREF_RECLAIM_CLEAN,
1442 	FOLIOREF_KEEP,
1443 	FOLIOREF_ACTIVATE,
1444 };
1445 
folio_check_references(struct folio * folio,struct scan_control * sc)1446 static enum folio_references folio_check_references(struct folio *folio,
1447 						  struct scan_control *sc)
1448 {
1449 	int referenced_ptes, referenced_folio;
1450 	unsigned long vm_flags;
1451 
1452 	referenced_ptes = folio_referenced(folio, 1, sc->target_mem_cgroup,
1453 					   &vm_flags);
1454 	referenced_folio = folio_test_clear_referenced(folio);
1455 
1456 	/*
1457 	 * The supposedly reclaimable folio was found to be in a VM_LOCKED vma.
1458 	 * Let the folio, now marked Mlocked, be moved to the unevictable list.
1459 	 */
1460 	if (vm_flags & VM_LOCKED)
1461 		return FOLIOREF_ACTIVATE;
1462 
1463 	/* rmap lock contention: rotate */
1464 	if (referenced_ptes == -1)
1465 		return FOLIOREF_KEEP;
1466 
1467 	if (referenced_ptes) {
1468 		/*
1469 		 * All mapped folios start out with page table
1470 		 * references from the instantiating fault, so we need
1471 		 * to look twice if a mapped file/anon folio is used more
1472 		 * than once.
1473 		 *
1474 		 * Mark it and spare it for another trip around the
1475 		 * inactive list.  Another page table reference will
1476 		 * lead to its activation.
1477 		 *
1478 		 * Note: the mark is set for activated folios as well
1479 		 * so that recently deactivated but used folios are
1480 		 * quickly recovered.
1481 		 */
1482 		folio_set_referenced(folio);
1483 
1484 		if (referenced_folio || referenced_ptes > 1)
1485 			return FOLIOREF_ACTIVATE;
1486 
1487 		/*
1488 		 * Activate file-backed executable folios after first usage.
1489 		 */
1490 		if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio))
1491 			return FOLIOREF_ACTIVATE;
1492 
1493 		return FOLIOREF_KEEP;
1494 	}
1495 
1496 	/* Reclaim if clean, defer dirty folios to writeback */
1497 	if (referenced_folio && folio_is_file_lru(folio))
1498 		return FOLIOREF_RECLAIM_CLEAN;
1499 
1500 	return FOLIOREF_RECLAIM;
1501 }
1502 
1503 /* Check if a folio is dirty or under writeback */
folio_check_dirty_writeback(struct folio * folio,bool * dirty,bool * writeback)1504 static void folio_check_dirty_writeback(struct folio *folio,
1505 				       bool *dirty, bool *writeback)
1506 {
1507 	struct address_space *mapping;
1508 
1509 	/*
1510 	 * Anonymous folios are not handled by flushers and must be written
1511 	 * from reclaim context. Do not stall reclaim based on them.
1512 	 * MADV_FREE anonymous folios are put into inactive file list too.
1513 	 * They could be mistakenly treated as file lru. So further anon
1514 	 * test is needed.
1515 	 */
1516 	if (!folio_is_file_lru(folio) ||
1517 	    (folio_test_anon(folio) && !folio_test_swapbacked(folio))) {
1518 		*dirty = false;
1519 		*writeback = false;
1520 		return;
1521 	}
1522 
1523 	/* By default assume that the folio flags are accurate */
1524 	*dirty = folio_test_dirty(folio);
1525 	*writeback = folio_test_writeback(folio);
1526 
1527 	/* Verify dirty/writeback state if the filesystem supports it */
1528 	if (!folio_test_private(folio))
1529 		return;
1530 
1531 	mapping = folio_mapping(folio);
1532 	if (mapping && mapping->a_ops->is_dirty_writeback)
1533 		mapping->a_ops->is_dirty_writeback(folio, dirty, writeback);
1534 }
1535 
alloc_demote_page(struct page * page,unsigned long private)1536 static struct page *alloc_demote_page(struct page *page, unsigned long private)
1537 {
1538 	struct page *target_page;
1539 	nodemask_t *allowed_mask;
1540 	struct migration_target_control *mtc;
1541 
1542 	mtc = (struct migration_target_control *)private;
1543 
1544 	allowed_mask = mtc->nmask;
1545 	/*
1546 	 * make sure we allocate from the target node first also trying to
1547 	 * demote or reclaim pages from the target node via kswapd if we are
1548 	 * low on free memory on target node. If we don't do this and if
1549 	 * we have free memory on the slower(lower) memtier, we would start
1550 	 * allocating pages from slower(lower) memory tiers without even forcing
1551 	 * a demotion of cold pages from the target memtier. This can result
1552 	 * in the kernel placing hot pages in slower(lower) memory tiers.
1553 	 */
1554 	mtc->nmask = NULL;
1555 	mtc->gfp_mask |= __GFP_THISNODE;
1556 	target_page = alloc_migration_target(page, (unsigned long)mtc);
1557 	if (target_page)
1558 		return target_page;
1559 
1560 	mtc->gfp_mask &= ~__GFP_THISNODE;
1561 	mtc->nmask = allowed_mask;
1562 
1563 	return alloc_migration_target(page, (unsigned long)mtc);
1564 }
1565 
1566 /*
1567  * Take folios on @demote_folios and attempt to demote them to another node.
1568  * Folios which are not demoted are left on @demote_folios.
1569  */
demote_folio_list(struct list_head * demote_folios,struct pglist_data * pgdat)1570 static unsigned int demote_folio_list(struct list_head *demote_folios,
1571 				     struct pglist_data *pgdat)
1572 {
1573 	int target_nid = next_demotion_node(pgdat->node_id);
1574 	unsigned int nr_succeeded;
1575 	nodemask_t allowed_mask;
1576 
1577 	struct migration_target_control mtc = {
1578 		/*
1579 		 * Allocate from 'node', or fail quickly and quietly.
1580 		 * When this happens, 'page' will likely just be discarded
1581 		 * instead of migrated.
1582 		 */
1583 		.gfp_mask = (GFP_HIGHUSER_MOVABLE & ~__GFP_RECLAIM) | __GFP_NOWARN |
1584 			__GFP_NOMEMALLOC | GFP_NOWAIT,
1585 		.nid = target_nid,
1586 		.nmask = &allowed_mask
1587 	};
1588 
1589 	if (list_empty(demote_folios))
1590 		return 0;
1591 
1592 	if (target_nid == NUMA_NO_NODE)
1593 		return 0;
1594 
1595 	node_get_allowed_targets(pgdat, &allowed_mask);
1596 
1597 	/* Demotion ignores all cpuset and mempolicy settings */
1598 	migrate_pages(demote_folios, alloc_demote_page, NULL,
1599 		      (unsigned long)&mtc, MIGRATE_ASYNC, MR_DEMOTION,
1600 		      &nr_succeeded);
1601 
1602 	if (current_is_kswapd())
1603 		__count_vm_events(PGDEMOTE_KSWAPD, nr_succeeded);
1604 	else
1605 		__count_vm_events(PGDEMOTE_DIRECT, nr_succeeded);
1606 
1607 	return nr_succeeded;
1608 }
1609 
may_enter_fs(struct folio * folio,gfp_t gfp_mask)1610 static bool may_enter_fs(struct folio *folio, gfp_t gfp_mask)
1611 {
1612 	if (gfp_mask & __GFP_FS)
1613 		return true;
1614 	if (!folio_test_swapcache(folio) || !(gfp_mask & __GFP_IO))
1615 		return false;
1616 	/*
1617 	 * We can "enter_fs" for swap-cache with only __GFP_IO
1618 	 * providing this isn't SWP_FS_OPS.
1619 	 * ->flags can be updated non-atomicially (scan_swap_map_slots),
1620 	 * but that will never affect SWP_FS_OPS, so the data_race
1621 	 * is safe.
1622 	 */
1623 	return !data_race(folio_swap_flags(folio) & SWP_FS_OPS);
1624 }
1625 
1626 /*
1627  * shrink_folio_list() returns the number of reclaimed pages
1628  */
shrink_folio_list(struct list_head * folio_list,struct pglist_data * pgdat,struct scan_control * sc,struct reclaim_stat * stat,bool ignore_references)1629 static unsigned int shrink_folio_list(struct list_head *folio_list,
1630 		struct pglist_data *pgdat, struct scan_control *sc,
1631 		struct reclaim_stat *stat, bool ignore_references)
1632 {
1633 	LIST_HEAD(ret_folios);
1634 	LIST_HEAD(free_folios);
1635 	LIST_HEAD(demote_folios);
1636 	unsigned int nr_reclaimed = 0;
1637 	unsigned int pgactivate = 0;
1638 	bool do_demote_pass;
1639 	struct swap_iocb *plug = NULL;
1640 
1641 	memset(stat, 0, sizeof(*stat));
1642 	cond_resched();
1643 	do_demote_pass = can_demote(pgdat->node_id, sc);
1644 
1645 retry:
1646 	while (!list_empty(folio_list)) {
1647 		struct address_space *mapping;
1648 		struct folio *folio;
1649 		enum folio_references references = FOLIOREF_RECLAIM;
1650 		bool dirty, writeback;
1651 		unsigned int nr_pages;
1652 
1653 		cond_resched();
1654 
1655 		folio = lru_to_folio(folio_list);
1656 		list_del(&folio->lru);
1657 
1658 		if (!folio_trylock(folio))
1659 			goto keep;
1660 
1661 		VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
1662 
1663 		nr_pages = folio_nr_pages(folio);
1664 
1665 		/* Account the number of base pages */
1666 		sc->nr_scanned += nr_pages;
1667 
1668 		if (unlikely(!folio_evictable(folio)))
1669 			goto activate_locked;
1670 
1671 		if (!sc->may_unmap && folio_mapped(folio))
1672 			goto keep_locked;
1673 
1674 		/* folio_update_gen() tried to promote this page? */
1675 		if (lru_gen_enabled() && !ignore_references &&
1676 		    folio_mapped(folio) && folio_test_referenced(folio))
1677 			goto keep_locked;
1678 
1679 		/*
1680 		 * The number of dirty pages determines if a node is marked
1681 		 * reclaim_congested. kswapd will stall and start writing
1682 		 * folios if the tail of the LRU is all dirty unqueued folios.
1683 		 */
1684 		folio_check_dirty_writeback(folio, &dirty, &writeback);
1685 		if (dirty || writeback)
1686 			stat->nr_dirty += nr_pages;
1687 
1688 		if (dirty && !writeback)
1689 			stat->nr_unqueued_dirty += nr_pages;
1690 
1691 		/*
1692 		 * Treat this folio as congested if folios are cycling
1693 		 * through the LRU so quickly that the folios marked
1694 		 * for immediate reclaim are making it to the end of
1695 		 * the LRU a second time.
1696 		 */
1697 		if (writeback && folio_test_reclaim(folio))
1698 			stat->nr_congested += nr_pages;
1699 
1700 		/*
1701 		 * If a folio at the tail of the LRU is under writeback, there
1702 		 * are three cases to consider.
1703 		 *
1704 		 * 1) If reclaim is encountering an excessive number
1705 		 *    of folios under writeback and this folio has both
1706 		 *    the writeback and reclaim flags set, then it
1707 		 *    indicates that folios are being queued for I/O but
1708 		 *    are being recycled through the LRU before the I/O
1709 		 *    can complete. Waiting on the folio itself risks an
1710 		 *    indefinite stall if it is impossible to writeback
1711 		 *    the folio due to I/O error or disconnected storage
1712 		 *    so instead note that the LRU is being scanned too
1713 		 *    quickly and the caller can stall after the folio
1714 		 *    list has been processed.
1715 		 *
1716 		 * 2) Global or new memcg reclaim encounters a folio that is
1717 		 *    not marked for immediate reclaim, or the caller does not
1718 		 *    have __GFP_FS (or __GFP_IO if it's simply going to swap,
1719 		 *    not to fs). In this case mark the folio for immediate
1720 		 *    reclaim and continue scanning.
1721 		 *
1722 		 *    Require may_enter_fs() because we would wait on fs, which
1723 		 *    may not have submitted I/O yet. And the loop driver might
1724 		 *    enter reclaim, and deadlock if it waits on a folio for
1725 		 *    which it is needed to do the write (loop masks off
1726 		 *    __GFP_IO|__GFP_FS for this reason); but more thought
1727 		 *    would probably show more reasons.
1728 		 *
1729 		 * 3) Legacy memcg encounters a folio that already has the
1730 		 *    reclaim flag set. memcg does not have any dirty folio
1731 		 *    throttling so we could easily OOM just because too many
1732 		 *    folios are in writeback and there is nothing else to
1733 		 *    reclaim. Wait for the writeback to complete.
1734 		 *
1735 		 * In cases 1) and 2) we activate the folios to get them out of
1736 		 * the way while we continue scanning for clean folios on the
1737 		 * inactive list and refilling from the active list. The
1738 		 * observation here is that waiting for disk writes is more
1739 		 * expensive than potentially causing reloads down the line.
1740 		 * Since they're marked for immediate reclaim, they won't put
1741 		 * memory pressure on the cache working set any longer than it
1742 		 * takes to write them to disk.
1743 		 */
1744 		if (folio_test_writeback(folio)) {
1745 			/* Case 1 above */
1746 			if (current_is_kswapd() &&
1747 			    folio_test_reclaim(folio) &&
1748 			    test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1749 				stat->nr_immediate += nr_pages;
1750 				goto activate_locked;
1751 
1752 			/* Case 2 above */
1753 			} else if (writeback_throttling_sane(sc) ||
1754 			    !folio_test_reclaim(folio) ||
1755 			    !may_enter_fs(folio, sc->gfp_mask)) {
1756 				/*
1757 				 * This is slightly racy -
1758 				 * folio_end_writeback() might have
1759 				 * just cleared the reclaim flag, then
1760 				 * setting the reclaim flag here ends up
1761 				 * interpreted as the readahead flag - but
1762 				 * that does not matter enough to care.
1763 				 * What we do want is for this folio to
1764 				 * have the reclaim flag set next time
1765 				 * memcg reclaim reaches the tests above,
1766 				 * so it will then wait for writeback to
1767 				 * avoid OOM; and it's also appropriate
1768 				 * in global reclaim.
1769 				 */
1770 				folio_set_reclaim(folio);
1771 				stat->nr_writeback += nr_pages;
1772 				goto activate_locked;
1773 
1774 			/* Case 3 above */
1775 			} else {
1776 				folio_unlock(folio);
1777 				folio_wait_writeback(folio);
1778 				/* then go back and try same folio again */
1779 				list_add_tail(&folio->lru, folio_list);
1780 				continue;
1781 			}
1782 		}
1783 
1784 		if (!ignore_references)
1785 			references = folio_check_references(folio, sc);
1786 
1787 		switch (references) {
1788 		case FOLIOREF_ACTIVATE:
1789 			goto activate_locked;
1790 		case FOLIOREF_KEEP:
1791 			stat->nr_ref_keep += nr_pages;
1792 			goto keep_locked;
1793 		case FOLIOREF_RECLAIM:
1794 		case FOLIOREF_RECLAIM_CLEAN:
1795 			; /* try to reclaim the folio below */
1796 		}
1797 
1798 		/*
1799 		 * Before reclaiming the folio, try to relocate
1800 		 * its contents to another node.
1801 		 */
1802 		if (do_demote_pass &&
1803 		    (thp_migration_supported() || !folio_test_large(folio))) {
1804 			list_add(&folio->lru, &demote_folios);
1805 			folio_unlock(folio);
1806 			continue;
1807 		}
1808 
1809 		/*
1810 		 * Anonymous process memory has backing store?
1811 		 * Try to allocate it some swap space here.
1812 		 * Lazyfree folio could be freed directly
1813 		 */
1814 		if (folio_test_anon(folio) && folio_test_swapbacked(folio)) {
1815 			if (!folio_test_swapcache(folio)) {
1816 				if (!(sc->gfp_mask & __GFP_IO))
1817 					goto keep_locked;
1818 				if (folio_maybe_dma_pinned(folio))
1819 					goto keep_locked;
1820 				if (folio_test_large(folio)) {
1821 					/* cannot split folio, skip it */
1822 					if (!can_split_folio(folio, NULL))
1823 						goto activate_locked;
1824 					/*
1825 					 * Split folios without a PMD map right
1826 					 * away. Chances are some or all of the
1827 					 * tail pages can be freed without IO.
1828 					 */
1829 					if (!folio_entire_mapcount(folio) &&
1830 					    split_folio_to_list(folio,
1831 								folio_list))
1832 						goto activate_locked;
1833 				}
1834 				if (!add_to_swap(folio)) {
1835 					if (!folio_test_large(folio))
1836 						goto activate_locked_split;
1837 					/* Fallback to swap normal pages */
1838 					if (split_folio_to_list(folio,
1839 								folio_list))
1840 						goto activate_locked;
1841 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1842 					count_vm_event(THP_SWPOUT_FALLBACK);
1843 #endif
1844 					if (!add_to_swap(folio))
1845 						goto activate_locked_split;
1846 				}
1847 			}
1848 		} else if (folio_test_swapbacked(folio) &&
1849 			   folio_test_large(folio)) {
1850 			/* Split shmem folio */
1851 			if (split_folio_to_list(folio, folio_list))
1852 				goto keep_locked;
1853 		}
1854 
1855 		/*
1856 		 * If the folio was split above, the tail pages will make
1857 		 * their own pass through this function and be accounted
1858 		 * then.
1859 		 */
1860 		if ((nr_pages > 1) && !folio_test_large(folio)) {
1861 			sc->nr_scanned -= (nr_pages - 1);
1862 			nr_pages = 1;
1863 		}
1864 
1865 		/*
1866 		 * The folio is mapped into the page tables of one or more
1867 		 * processes. Try to unmap it here.
1868 		 */
1869 		if (folio_mapped(folio)) {
1870 			enum ttu_flags flags = TTU_BATCH_FLUSH;
1871 			bool was_swapbacked = folio_test_swapbacked(folio);
1872 
1873 			if (folio_test_pmd_mappable(folio))
1874 				flags |= TTU_SPLIT_HUGE_PMD;
1875 
1876 			try_to_unmap(folio, flags);
1877 			if (folio_mapped(folio)) {
1878 				stat->nr_unmap_fail += nr_pages;
1879 				if (!was_swapbacked &&
1880 				    folio_test_swapbacked(folio))
1881 					stat->nr_lazyfree_fail += nr_pages;
1882 				goto activate_locked;
1883 			}
1884 		}
1885 
1886 		mapping = folio_mapping(folio);
1887 		if (folio_test_dirty(folio)) {
1888 			/*
1889 			 * Only kswapd can writeback filesystem folios
1890 			 * to avoid risk of stack overflow. But avoid
1891 			 * injecting inefficient single-folio I/O into
1892 			 * flusher writeback as much as possible: only
1893 			 * write folios when we've encountered many
1894 			 * dirty folios, and when we've already scanned
1895 			 * the rest of the LRU for clean folios and see
1896 			 * the same dirty folios again (with the reclaim
1897 			 * flag set).
1898 			 */
1899 			if (folio_is_file_lru(folio) &&
1900 			    (!current_is_kswapd() ||
1901 			     !folio_test_reclaim(folio) ||
1902 			     !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1903 				/*
1904 				 * Immediately reclaim when written back.
1905 				 * Similar in principle to deactivate_page()
1906 				 * except we already have the folio isolated
1907 				 * and know it's dirty
1908 				 */
1909 				node_stat_mod_folio(folio, NR_VMSCAN_IMMEDIATE,
1910 						nr_pages);
1911 				folio_set_reclaim(folio);
1912 
1913 				goto activate_locked;
1914 			}
1915 
1916 			if (references == FOLIOREF_RECLAIM_CLEAN)
1917 				goto keep_locked;
1918 			if (!may_enter_fs(folio, sc->gfp_mask))
1919 				goto keep_locked;
1920 			if (!sc->may_writepage)
1921 				goto keep_locked;
1922 
1923 			/*
1924 			 * Folio is dirty. Flush the TLB if a writable entry
1925 			 * potentially exists to avoid CPU writes after I/O
1926 			 * starts and then write it out here.
1927 			 */
1928 			try_to_unmap_flush_dirty();
1929 			switch (pageout(folio, mapping, &plug)) {
1930 			case PAGE_KEEP:
1931 				goto keep_locked;
1932 			case PAGE_ACTIVATE:
1933 				goto activate_locked;
1934 			case PAGE_SUCCESS:
1935 				stat->nr_pageout += nr_pages;
1936 
1937 				if (folio_test_writeback(folio))
1938 					goto keep;
1939 				if (folio_test_dirty(folio))
1940 					goto keep;
1941 
1942 				/*
1943 				 * A synchronous write - probably a ramdisk.  Go
1944 				 * ahead and try to reclaim the folio.
1945 				 */
1946 				if (!folio_trylock(folio))
1947 					goto keep;
1948 				if (folio_test_dirty(folio) ||
1949 				    folio_test_writeback(folio))
1950 					goto keep_locked;
1951 				mapping = folio_mapping(folio);
1952 				fallthrough;
1953 			case PAGE_CLEAN:
1954 				; /* try to free the folio below */
1955 			}
1956 		}
1957 
1958 		/*
1959 		 * If the folio has buffers, try to free the buffer
1960 		 * mappings associated with this folio. If we succeed
1961 		 * we try to free the folio as well.
1962 		 *
1963 		 * We do this even if the folio is dirty.
1964 		 * filemap_release_folio() does not perform I/O, but it
1965 		 * is possible for a folio to have the dirty flag set,
1966 		 * but it is actually clean (all its buffers are clean).
1967 		 * This happens if the buffers were written out directly,
1968 		 * with submit_bh(). ext3 will do this, as well as
1969 		 * the blockdev mapping.  filemap_release_folio() will
1970 		 * discover that cleanness and will drop the buffers
1971 		 * and mark the folio clean - it can be freed.
1972 		 *
1973 		 * Rarely, folios can have buffers and no ->mapping.
1974 		 * These are the folios which were not successfully
1975 		 * invalidated in truncate_cleanup_folio().  We try to
1976 		 * drop those buffers here and if that worked, and the
1977 		 * folio is no longer mapped into process address space
1978 		 * (refcount == 1) it can be freed.  Otherwise, leave
1979 		 * the folio on the LRU so it is swappable.
1980 		 */
1981 		if (folio_has_private(folio)) {
1982 			if (!filemap_release_folio(folio, sc->gfp_mask))
1983 				goto activate_locked;
1984 			if (!mapping && folio_ref_count(folio) == 1) {
1985 				folio_unlock(folio);
1986 				if (folio_put_testzero(folio))
1987 					goto free_it;
1988 				else {
1989 					/*
1990 					 * rare race with speculative reference.
1991 					 * the speculative reference will free
1992 					 * this folio shortly, so we may
1993 					 * increment nr_reclaimed here (and
1994 					 * leave it off the LRU).
1995 					 */
1996 					nr_reclaimed += nr_pages;
1997 					continue;
1998 				}
1999 			}
2000 		}
2001 
2002 		if (folio_test_anon(folio) && !folio_test_swapbacked(folio)) {
2003 			/* follow __remove_mapping for reference */
2004 			if (!folio_ref_freeze(folio, 1))
2005 				goto keep_locked;
2006 			/*
2007 			 * The folio has only one reference left, which is
2008 			 * from the isolation. After the caller puts the
2009 			 * folio back on the lru and drops the reference, the
2010 			 * folio will be freed anyway. It doesn't matter
2011 			 * which lru it goes on. So we don't bother checking
2012 			 * the dirty flag here.
2013 			 */
2014 			count_vm_events(PGLAZYFREED, nr_pages);
2015 			count_memcg_folio_events(folio, PGLAZYFREED, nr_pages);
2016 		} else if (!mapping || !__remove_mapping(mapping, folio, true,
2017 							 sc->target_mem_cgroup))
2018 			goto keep_locked;
2019 
2020 		folio_unlock(folio);
2021 free_it:
2022 		/*
2023 		 * Folio may get swapped out as a whole, need to account
2024 		 * all pages in it.
2025 		 */
2026 		nr_reclaimed += nr_pages;
2027 
2028 		/*
2029 		 * Is there need to periodically free_folio_list? It would
2030 		 * appear not as the counts should be low
2031 		 */
2032 		if (unlikely(folio_test_large(folio)))
2033 			destroy_large_folio(folio);
2034 		else
2035 			list_add(&folio->lru, &free_folios);
2036 		continue;
2037 
2038 activate_locked_split:
2039 		/*
2040 		 * The tail pages that are failed to add into swap cache
2041 		 * reach here.  Fixup nr_scanned and nr_pages.
2042 		 */
2043 		if (nr_pages > 1) {
2044 			sc->nr_scanned -= (nr_pages - 1);
2045 			nr_pages = 1;
2046 		}
2047 activate_locked:
2048 		/* Not a candidate for swapping, so reclaim swap space. */
2049 		if (folio_test_swapcache(folio) &&
2050 		    (mem_cgroup_swap_full(folio) || folio_test_mlocked(folio)))
2051 			folio_free_swap(folio);
2052 		VM_BUG_ON_FOLIO(folio_test_active(folio), folio);
2053 		if (!folio_test_mlocked(folio)) {
2054 			int type = folio_is_file_lru(folio);
2055 			folio_set_active(folio);
2056 			stat->nr_activate[type] += nr_pages;
2057 			count_memcg_folio_events(folio, PGACTIVATE, nr_pages);
2058 		}
2059 keep_locked:
2060 		folio_unlock(folio);
2061 keep:
2062 		list_add(&folio->lru, &ret_folios);
2063 		VM_BUG_ON_FOLIO(folio_test_lru(folio) ||
2064 				folio_test_unevictable(folio), folio);
2065 	}
2066 	/* 'folio_list' is always empty here */
2067 
2068 	/* Migrate folios selected for demotion */
2069 	nr_reclaimed += demote_folio_list(&demote_folios, pgdat);
2070 	/* Folios that could not be demoted are still in @demote_folios */
2071 	if (!list_empty(&demote_folios)) {
2072 		/* Folios which weren't demoted go back on @folio_list for retry: */
2073 		list_splice_init(&demote_folios, folio_list);
2074 		do_demote_pass = false;
2075 		goto retry;
2076 	}
2077 
2078 	pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
2079 
2080 	mem_cgroup_uncharge_list(&free_folios);
2081 	try_to_unmap_flush();
2082 	free_unref_page_list(&free_folios);
2083 
2084 	list_splice(&ret_folios, folio_list);
2085 	count_vm_events(PGACTIVATE, pgactivate);
2086 
2087 	if (plug)
2088 		swap_write_unplug(plug);
2089 	return nr_reclaimed;
2090 }
2091 
reclaim_clean_pages_from_list(struct zone * zone,struct list_head * folio_list)2092 unsigned int reclaim_clean_pages_from_list(struct zone *zone,
2093 					   struct list_head *folio_list)
2094 {
2095 	struct scan_control sc = {
2096 		.gfp_mask = GFP_KERNEL,
2097 		.may_unmap = 1,
2098 	};
2099 	struct reclaim_stat stat;
2100 	unsigned int nr_reclaimed;
2101 	struct folio *folio, *next;
2102 	LIST_HEAD(clean_folios);
2103 	unsigned int noreclaim_flag;
2104 
2105 	list_for_each_entry_safe(folio, next, folio_list, lru) {
2106 		if (!folio_test_hugetlb(folio) && folio_is_file_lru(folio) &&
2107 		    !folio_test_dirty(folio) && !__folio_test_movable(folio) &&
2108 		    !folio_test_unevictable(folio)) {
2109 			folio_clear_active(folio);
2110 			list_move(&folio->lru, &clean_folios);
2111 		}
2112 	}
2113 
2114 	/*
2115 	 * We should be safe here since we are only dealing with file pages and
2116 	 * we are not kswapd and therefore cannot write dirty file pages. But
2117 	 * call memalloc_noreclaim_save() anyway, just in case these conditions
2118 	 * change in the future.
2119 	 */
2120 	noreclaim_flag = memalloc_noreclaim_save();
2121 	nr_reclaimed = shrink_folio_list(&clean_folios, zone->zone_pgdat, &sc,
2122 					&stat, true);
2123 	memalloc_noreclaim_restore(noreclaim_flag);
2124 
2125 	list_splice(&clean_folios, folio_list);
2126 	mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
2127 			    -(long)nr_reclaimed);
2128 	/*
2129 	 * Since lazyfree pages are isolated from file LRU from the beginning,
2130 	 * they will rotate back to anonymous LRU in the end if it failed to
2131 	 * discard so isolated count will be mismatched.
2132 	 * Compensate the isolated count for both LRU lists.
2133 	 */
2134 	mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON,
2135 			    stat.nr_lazyfree_fail);
2136 	mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE,
2137 			    -(long)stat.nr_lazyfree_fail);
2138 	return nr_reclaimed;
2139 }
2140 
2141 /*
2142  * Update LRU sizes after isolating pages. The LRU size updates must
2143  * be complete before mem_cgroup_update_lru_size due to a sanity check.
2144  */
update_lru_sizes(struct lruvec * lruvec,enum lru_list lru,unsigned long * nr_zone_taken)2145 static __always_inline void update_lru_sizes(struct lruvec *lruvec,
2146 			enum lru_list lru, unsigned long *nr_zone_taken)
2147 {
2148 	int zid;
2149 
2150 	for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2151 		if (!nr_zone_taken[zid])
2152 			continue;
2153 
2154 		update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
2155 	}
2156 
2157 }
2158 
2159 /*
2160  * Isolating page from the lruvec to fill in @dst list by nr_to_scan times.
2161  *
2162  * lruvec->lru_lock is heavily contended.  Some of the functions that
2163  * shrink the lists perform better by taking out a batch of pages
2164  * and working on them outside the LRU lock.
2165  *
2166  * For pagecache intensive workloads, this function is the hottest
2167  * spot in the kernel (apart from copy_*_user functions).
2168  *
2169  * Lru_lock must be held before calling this function.
2170  *
2171  * @nr_to_scan:	The number of eligible pages to look through on the list.
2172  * @lruvec:	The LRU vector to pull pages from.
2173  * @dst:	The temp list to put pages on to.
2174  * @nr_scanned:	The number of pages that were scanned.
2175  * @sc:		The scan_control struct for this reclaim session
2176  * @lru:	LRU list id for isolating
2177  *
2178  * returns how many pages were moved onto *@dst.
2179  */
isolate_lru_folios(unsigned long nr_to_scan,struct lruvec * lruvec,struct list_head * dst,unsigned long * nr_scanned,struct scan_control * sc,enum lru_list lru)2180 static unsigned long isolate_lru_folios(unsigned long nr_to_scan,
2181 		struct lruvec *lruvec, struct list_head *dst,
2182 		unsigned long *nr_scanned, struct scan_control *sc,
2183 		enum lru_list lru)
2184 {
2185 	struct list_head *src = &lruvec->lists[lru];
2186 	unsigned long nr_taken = 0;
2187 	unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
2188 	unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
2189 	unsigned long skipped = 0;
2190 	unsigned long scan, total_scan, nr_pages;
2191 	LIST_HEAD(folios_skipped);
2192 
2193 	total_scan = 0;
2194 	scan = 0;
2195 	while (scan < nr_to_scan && !list_empty(src)) {
2196 		struct list_head *move_to = src;
2197 		struct folio *folio;
2198 
2199 		folio = lru_to_folio(src);
2200 		prefetchw_prev_lru_folio(folio, src, flags);
2201 
2202 		nr_pages = folio_nr_pages(folio);
2203 		total_scan += nr_pages;
2204 
2205 		if (folio_zonenum(folio) > sc->reclaim_idx) {
2206 			nr_skipped[folio_zonenum(folio)] += nr_pages;
2207 			move_to = &folios_skipped;
2208 			goto move;
2209 		}
2210 
2211 		/*
2212 		 * Do not count skipped folios because that makes the function
2213 		 * return with no isolated folios if the LRU mostly contains
2214 		 * ineligible folios.  This causes the VM to not reclaim any
2215 		 * folios, triggering a premature OOM.
2216 		 * Account all pages in a folio.
2217 		 */
2218 		scan += nr_pages;
2219 
2220 		if (!folio_test_lru(folio))
2221 			goto move;
2222 		if (!sc->may_unmap && folio_mapped(folio))
2223 			goto move;
2224 
2225 		/*
2226 		 * Be careful not to clear the lru flag until after we're
2227 		 * sure the folio is not being freed elsewhere -- the
2228 		 * folio release code relies on it.
2229 		 */
2230 		if (unlikely(!folio_try_get(folio)))
2231 			goto move;
2232 
2233 		if (!folio_test_clear_lru(folio)) {
2234 			/* Another thread is already isolating this folio */
2235 			folio_put(folio);
2236 			goto move;
2237 		}
2238 
2239 		nr_taken += nr_pages;
2240 		nr_zone_taken[folio_zonenum(folio)] += nr_pages;
2241 		move_to = dst;
2242 move:
2243 		list_move(&folio->lru, move_to);
2244 	}
2245 
2246 	/*
2247 	 * Splice any skipped folios to the start of the LRU list. Note that
2248 	 * this disrupts the LRU order when reclaiming for lower zones but
2249 	 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
2250 	 * scanning would soon rescan the same folios to skip and waste lots
2251 	 * of cpu cycles.
2252 	 */
2253 	if (!list_empty(&folios_skipped)) {
2254 		int zid;
2255 
2256 		list_splice(&folios_skipped, src);
2257 		for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2258 			if (!nr_skipped[zid])
2259 				continue;
2260 
2261 			__count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
2262 			skipped += nr_skipped[zid];
2263 		}
2264 	}
2265 	*nr_scanned = total_scan;
2266 	trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
2267 				    total_scan, skipped, nr_taken,
2268 				    sc->may_unmap ? 0 : ISOLATE_UNMAPPED, lru);
2269 	update_lru_sizes(lruvec, lru, nr_zone_taken);
2270 	return nr_taken;
2271 }
2272 
2273 /**
2274  * folio_isolate_lru() - Try to isolate a folio from its LRU list.
2275  * @folio: Folio to isolate from its LRU list.
2276  *
2277  * Isolate a @folio from an LRU list and adjust the vmstat statistic
2278  * corresponding to whatever LRU list the folio was on.
2279  *
2280  * The folio will have its LRU flag cleared.  If it was found on the
2281  * active list, it will have the Active flag set.  If it was found on the
2282  * unevictable list, it will have the Unevictable flag set.  These flags
2283  * may need to be cleared by the caller before letting the page go.
2284  *
2285  * Context:
2286  *
2287  * (1) Must be called with an elevated refcount on the folio. This is a
2288  *     fundamental difference from isolate_lru_folios() (which is called
2289  *     without a stable reference).
2290  * (2) The lru_lock must not be held.
2291  * (3) Interrupts must be enabled.
2292  *
2293  * Return: 0 if the folio was removed from an LRU list.
2294  * -EBUSY if the folio was not on an LRU list.
2295  */
folio_isolate_lru(struct folio * folio)2296 int folio_isolate_lru(struct folio *folio)
2297 {
2298 	int ret = -EBUSY;
2299 
2300 	VM_BUG_ON_FOLIO(!folio_ref_count(folio), folio);
2301 
2302 	if (folio_test_clear_lru(folio)) {
2303 		struct lruvec *lruvec;
2304 
2305 		folio_get(folio);
2306 		lruvec = folio_lruvec_lock_irq(folio);
2307 		lruvec_del_folio(lruvec, folio);
2308 		unlock_page_lruvec_irq(lruvec);
2309 		ret = 0;
2310 	}
2311 
2312 	return ret;
2313 }
2314 
2315 /*
2316  * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
2317  * then get rescheduled. When there are massive number of tasks doing page
2318  * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
2319  * the LRU list will go small and be scanned faster than necessary, leading to
2320  * unnecessary swapping, thrashing and OOM.
2321  */
too_many_isolated(struct pglist_data * pgdat,int file,struct scan_control * sc)2322 static int too_many_isolated(struct pglist_data *pgdat, int file,
2323 		struct scan_control *sc)
2324 {
2325 	unsigned long inactive, isolated;
2326 	bool too_many;
2327 
2328 	if (current_is_kswapd())
2329 		return 0;
2330 
2331 	if (!writeback_throttling_sane(sc))
2332 		return 0;
2333 
2334 	if (file) {
2335 		inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
2336 		isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
2337 	} else {
2338 		inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
2339 		isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
2340 	}
2341 
2342 	/*
2343 	 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
2344 	 * won't get blocked by normal direct-reclaimers, forming a circular
2345 	 * deadlock.
2346 	 */
2347 	if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
2348 		inactive >>= 3;
2349 
2350 	too_many = isolated > inactive;
2351 
2352 	/* Wake up tasks throttled due to too_many_isolated. */
2353 	if (!too_many)
2354 		wake_throttle_isolated(pgdat);
2355 
2356 	return too_many;
2357 }
2358 
2359 /*
2360  * move_folios_to_lru() moves folios from private @list to appropriate LRU list.
2361  * On return, @list is reused as a list of folios to be freed by the caller.
2362  *
2363  * Returns the number of pages moved to the given lruvec.
2364  */
move_folios_to_lru(struct lruvec * lruvec,struct list_head * list)2365 static unsigned int move_folios_to_lru(struct lruvec *lruvec,
2366 		struct list_head *list)
2367 {
2368 	int nr_pages, nr_moved = 0;
2369 	LIST_HEAD(folios_to_free);
2370 
2371 	while (!list_empty(list)) {
2372 		struct folio *folio = lru_to_folio(list);
2373 
2374 		VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
2375 		list_del(&folio->lru);
2376 		if (unlikely(!folio_evictable(folio))) {
2377 			spin_unlock_irq(&lruvec->lru_lock);
2378 			folio_putback_lru(folio);
2379 			spin_lock_irq(&lruvec->lru_lock);
2380 			continue;
2381 		}
2382 
2383 		/*
2384 		 * The folio_set_lru needs to be kept here for list integrity.
2385 		 * Otherwise:
2386 		 *   #0 move_folios_to_lru             #1 release_pages
2387 		 *   if (!folio_put_testzero())
2388 		 *				      if (folio_put_testzero())
2389 		 *				        !lru //skip lru_lock
2390 		 *     folio_set_lru()
2391 		 *     list_add(&folio->lru,)
2392 		 *                                        list_add(&folio->lru,)
2393 		 */
2394 		folio_set_lru(folio);
2395 
2396 		if (unlikely(folio_put_testzero(folio))) {
2397 			__folio_clear_lru_flags(folio);
2398 
2399 			if (unlikely(folio_test_large(folio))) {
2400 				spin_unlock_irq(&lruvec->lru_lock);
2401 				destroy_large_folio(folio);
2402 				spin_lock_irq(&lruvec->lru_lock);
2403 			} else
2404 				list_add(&folio->lru, &folios_to_free);
2405 
2406 			continue;
2407 		}
2408 
2409 		/*
2410 		 * All pages were isolated from the same lruvec (and isolation
2411 		 * inhibits memcg migration).
2412 		 */
2413 		VM_BUG_ON_FOLIO(!folio_matches_lruvec(folio, lruvec), folio);
2414 		lruvec_add_folio(lruvec, folio);
2415 		nr_pages = folio_nr_pages(folio);
2416 		nr_moved += nr_pages;
2417 		if (folio_test_active(folio))
2418 			workingset_age_nonresident(lruvec, nr_pages);
2419 	}
2420 
2421 	/*
2422 	 * To save our caller's stack, now use input list for pages to free.
2423 	 */
2424 	list_splice(&folios_to_free, list);
2425 
2426 	return nr_moved;
2427 }
2428 
2429 /*
2430  * If a kernel thread (such as nfsd for loop-back mounts) services a backing
2431  * device by writing to the page cache it sets PF_LOCAL_THROTTLE. In this case
2432  * we should not throttle.  Otherwise it is safe to do so.
2433  */
current_may_throttle(void)2434 static int current_may_throttle(void)
2435 {
2436 	return !(current->flags & PF_LOCAL_THROTTLE);
2437 }
2438 
2439 /*
2440  * shrink_inactive_list() is a helper for shrink_node().  It returns the number
2441  * of reclaimed pages
2442  */
shrink_inactive_list(unsigned long nr_to_scan,struct lruvec * lruvec,struct scan_control * sc,enum lru_list lru)2443 static unsigned long shrink_inactive_list(unsigned long nr_to_scan,
2444 		struct lruvec *lruvec, struct scan_control *sc,
2445 		enum lru_list lru)
2446 {
2447 	LIST_HEAD(folio_list);
2448 	unsigned long nr_scanned;
2449 	unsigned int nr_reclaimed = 0;
2450 	unsigned long nr_taken;
2451 	struct reclaim_stat stat;
2452 	bool file = is_file_lru(lru);
2453 	enum vm_event_item item;
2454 	struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2455 	bool stalled = false;
2456 
2457 	while (unlikely(too_many_isolated(pgdat, file, sc))) {
2458 		if (stalled)
2459 			return 0;
2460 
2461 		/* wait a bit for the reclaimer. */
2462 		stalled = true;
2463 		reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED);
2464 
2465 		/* We are about to die and free our memory. Return now. */
2466 		if (fatal_signal_pending(current))
2467 			return SWAP_CLUSTER_MAX;
2468 	}
2469 
2470 	lru_add_drain();
2471 
2472 	spin_lock_irq(&lruvec->lru_lock);
2473 
2474 	nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &folio_list,
2475 				     &nr_scanned, sc, lru);
2476 
2477 	__mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2478 	item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
2479 	if (!cgroup_reclaim(sc))
2480 		__count_vm_events(item, nr_scanned);
2481 	__count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
2482 	__count_vm_events(PGSCAN_ANON + file, nr_scanned);
2483 
2484 	spin_unlock_irq(&lruvec->lru_lock);
2485 
2486 	if (nr_taken == 0)
2487 		return 0;
2488 
2489 	nr_reclaimed = shrink_folio_list(&folio_list, pgdat, sc, &stat, false);
2490 
2491 	spin_lock_irq(&lruvec->lru_lock);
2492 	move_folios_to_lru(lruvec, &folio_list);
2493 
2494 	__mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2495 	item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
2496 	if (!cgroup_reclaim(sc))
2497 		__count_vm_events(item, nr_reclaimed);
2498 	__count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
2499 	__count_vm_events(PGSTEAL_ANON + file, nr_reclaimed);
2500 	spin_unlock_irq(&lruvec->lru_lock);
2501 
2502 	lru_note_cost(lruvec, file, stat.nr_pageout);
2503 	mem_cgroup_uncharge_list(&folio_list);
2504 	free_unref_page_list(&folio_list);
2505 
2506 	/*
2507 	 * If dirty folios are scanned that are not queued for IO, it
2508 	 * implies that flushers are not doing their job. This can
2509 	 * happen when memory pressure pushes dirty folios to the end of
2510 	 * the LRU before the dirty limits are breached and the dirty
2511 	 * data has expired. It can also happen when the proportion of
2512 	 * dirty folios grows not through writes but through memory
2513 	 * pressure reclaiming all the clean cache. And in some cases,
2514 	 * the flushers simply cannot keep up with the allocation
2515 	 * rate. Nudge the flusher threads in case they are asleep.
2516 	 */
2517 	if (stat.nr_unqueued_dirty == nr_taken) {
2518 		wakeup_flusher_threads(WB_REASON_VMSCAN);
2519 		/*
2520 		 * For cgroupv1 dirty throttling is achieved by waking up
2521 		 * the kernel flusher here and later waiting on folios
2522 		 * which are in writeback to finish (see shrink_folio_list()).
2523 		 *
2524 		 * Flusher may not be able to issue writeback quickly
2525 		 * enough for cgroupv1 writeback throttling to work
2526 		 * on a large system.
2527 		 */
2528 		if (!writeback_throttling_sane(sc))
2529 			reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
2530 	}
2531 
2532 	sc->nr.dirty += stat.nr_dirty;
2533 	sc->nr.congested += stat.nr_congested;
2534 	sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2535 	sc->nr.writeback += stat.nr_writeback;
2536 	sc->nr.immediate += stat.nr_immediate;
2537 	sc->nr.taken += nr_taken;
2538 	if (file)
2539 		sc->nr.file_taken += nr_taken;
2540 
2541 	trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2542 			nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2543 	return nr_reclaimed;
2544 }
2545 
2546 /*
2547  * shrink_active_list() moves folios from the active LRU to the inactive LRU.
2548  *
2549  * We move them the other way if the folio is referenced by one or more
2550  * processes.
2551  *
2552  * If the folios are mostly unmapped, the processing is fast and it is
2553  * appropriate to hold lru_lock across the whole operation.  But if
2554  * the folios are mapped, the processing is slow (folio_referenced()), so
2555  * we should drop lru_lock around each folio.  It's impossible to balance
2556  * this, so instead we remove the folios from the LRU while processing them.
2557  * It is safe to rely on the active flag against the non-LRU folios in here
2558  * because nobody will play with that bit on a non-LRU folio.
2559  *
2560  * The downside is that we have to touch folio->_refcount against each folio.
2561  * But we had to alter folio->flags anyway.
2562  */
shrink_active_list(unsigned long nr_to_scan,struct lruvec * lruvec,struct scan_control * sc,enum lru_list lru)2563 static void shrink_active_list(unsigned long nr_to_scan,
2564 			       struct lruvec *lruvec,
2565 			       struct scan_control *sc,
2566 			       enum lru_list lru)
2567 {
2568 	unsigned long nr_taken;
2569 	unsigned long nr_scanned;
2570 	unsigned long vm_flags;
2571 	LIST_HEAD(l_hold);	/* The folios which were snipped off */
2572 	LIST_HEAD(l_active);
2573 	LIST_HEAD(l_inactive);
2574 	unsigned nr_deactivate, nr_activate;
2575 	unsigned nr_rotated = 0;
2576 	int file = is_file_lru(lru);
2577 	struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2578 
2579 	lru_add_drain();
2580 
2581 	spin_lock_irq(&lruvec->lru_lock);
2582 
2583 	nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &l_hold,
2584 				     &nr_scanned, sc, lru);
2585 
2586 	__mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2587 
2588 	if (!cgroup_reclaim(sc))
2589 		__count_vm_events(PGREFILL, nr_scanned);
2590 	__count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2591 
2592 	spin_unlock_irq(&lruvec->lru_lock);
2593 
2594 	while (!list_empty(&l_hold)) {
2595 		struct folio *folio;
2596 
2597 		cond_resched();
2598 		folio = lru_to_folio(&l_hold);
2599 		list_del(&folio->lru);
2600 
2601 		if (unlikely(!folio_evictable(folio))) {
2602 			folio_putback_lru(folio);
2603 			continue;
2604 		}
2605 
2606 		if (unlikely(buffer_heads_over_limit)) {
2607 			if (folio_test_private(folio) && folio_trylock(folio)) {
2608 				if (folio_test_private(folio))
2609 					filemap_release_folio(folio, 0);
2610 				folio_unlock(folio);
2611 			}
2612 		}
2613 
2614 		/* Referenced or rmap lock contention: rotate */
2615 		if (folio_referenced(folio, 0, sc->target_mem_cgroup,
2616 				     &vm_flags) != 0) {
2617 			/*
2618 			 * Identify referenced, file-backed active folios and
2619 			 * give them one more trip around the active list. So
2620 			 * that executable code get better chances to stay in
2621 			 * memory under moderate memory pressure.  Anon folios
2622 			 * are not likely to be evicted by use-once streaming
2623 			 * IO, plus JVM can create lots of anon VM_EXEC folios,
2624 			 * so we ignore them here.
2625 			 */
2626 			if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio)) {
2627 				nr_rotated += folio_nr_pages(folio);
2628 				list_add(&folio->lru, &l_active);
2629 				continue;
2630 			}
2631 		}
2632 
2633 		folio_clear_active(folio);	/* we are de-activating */
2634 		folio_set_workingset(folio);
2635 		list_add(&folio->lru, &l_inactive);
2636 	}
2637 
2638 	/*
2639 	 * Move folios back to the lru list.
2640 	 */
2641 	spin_lock_irq(&lruvec->lru_lock);
2642 
2643 	nr_activate = move_folios_to_lru(lruvec, &l_active);
2644 	nr_deactivate = move_folios_to_lru(lruvec, &l_inactive);
2645 	/* Keep all free folios in l_active list */
2646 	list_splice(&l_inactive, &l_active);
2647 
2648 	__count_vm_events(PGDEACTIVATE, nr_deactivate);
2649 	__count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2650 
2651 	__mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2652 	spin_unlock_irq(&lruvec->lru_lock);
2653 
2654 	mem_cgroup_uncharge_list(&l_active);
2655 	free_unref_page_list(&l_active);
2656 	trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2657 			nr_deactivate, nr_rotated, sc->priority, file);
2658 }
2659 
reclaim_folio_list(struct list_head * folio_list,struct pglist_data * pgdat)2660 static unsigned int reclaim_folio_list(struct list_head *folio_list,
2661 				      struct pglist_data *pgdat)
2662 {
2663 	struct reclaim_stat dummy_stat;
2664 	unsigned int nr_reclaimed;
2665 	struct folio *folio;
2666 	struct scan_control sc = {
2667 		.gfp_mask = GFP_KERNEL,
2668 		.may_writepage = 1,
2669 		.may_unmap = 1,
2670 		.may_swap = 1,
2671 		.no_demotion = 1,
2672 	};
2673 
2674 	nr_reclaimed = shrink_folio_list(folio_list, pgdat, &sc, &dummy_stat, false);
2675 	while (!list_empty(folio_list)) {
2676 		folio = lru_to_folio(folio_list);
2677 		list_del(&folio->lru);
2678 		folio_putback_lru(folio);
2679 	}
2680 
2681 	return nr_reclaimed;
2682 }
2683 
reclaim_pages(struct list_head * folio_list)2684 unsigned long reclaim_pages(struct list_head *folio_list)
2685 {
2686 	int nid;
2687 	unsigned int nr_reclaimed = 0;
2688 	LIST_HEAD(node_folio_list);
2689 	unsigned int noreclaim_flag;
2690 
2691 	if (list_empty(folio_list))
2692 		return nr_reclaimed;
2693 
2694 	noreclaim_flag = memalloc_noreclaim_save();
2695 
2696 	nid = folio_nid(lru_to_folio(folio_list));
2697 	do {
2698 		struct folio *folio = lru_to_folio(folio_list);
2699 
2700 		if (nid == folio_nid(folio)) {
2701 			folio_clear_active(folio);
2702 			list_move(&folio->lru, &node_folio_list);
2703 			continue;
2704 		}
2705 
2706 		nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid));
2707 		nid = folio_nid(lru_to_folio(folio_list));
2708 	} while (!list_empty(folio_list));
2709 
2710 	nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid));
2711 
2712 	memalloc_noreclaim_restore(noreclaim_flag);
2713 
2714 	return nr_reclaimed;
2715 }
2716 
shrink_list(enum lru_list lru,unsigned long nr_to_scan,struct lruvec * lruvec,struct scan_control * sc)2717 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2718 				 struct lruvec *lruvec, struct scan_control *sc)
2719 {
2720 	if (is_active_lru(lru)) {
2721 		if (sc->may_deactivate & (1 << is_file_lru(lru)))
2722 			shrink_active_list(nr_to_scan, lruvec, sc, lru);
2723 		else
2724 			sc->skipped_deactivate = 1;
2725 		return 0;
2726 	}
2727 
2728 	return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2729 }
2730 
2731 /*
2732  * The inactive anon list should be small enough that the VM never has
2733  * to do too much work.
2734  *
2735  * The inactive file list should be small enough to leave most memory
2736  * to the established workingset on the scan-resistant active list,
2737  * but large enough to avoid thrashing the aggregate readahead window.
2738  *
2739  * Both inactive lists should also be large enough that each inactive
2740  * folio has a chance to be referenced again before it is reclaimed.
2741  *
2742  * If that fails and refaulting is observed, the inactive list grows.
2743  *
2744  * The inactive_ratio is the target ratio of ACTIVE to INACTIVE folios
2745  * on this LRU, maintained by the pageout code. An inactive_ratio
2746  * of 3 means 3:1 or 25% of the folios are kept on the inactive list.
2747  *
2748  * total     target    max
2749  * memory    ratio     inactive
2750  * -------------------------------------
2751  *   10MB       1         5MB
2752  *  100MB       1        50MB
2753  *    1GB       3       250MB
2754  *   10GB      10       0.9GB
2755  *  100GB      31         3GB
2756  *    1TB     101        10GB
2757  *   10TB     320        32GB
2758  */
inactive_is_low(struct lruvec * lruvec,enum lru_list inactive_lru)2759 static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru)
2760 {
2761 	enum lru_list active_lru = inactive_lru + LRU_ACTIVE;
2762 	unsigned long inactive, active;
2763 	unsigned long inactive_ratio;
2764 	unsigned long gb;
2765 
2766 	inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru);
2767 	active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru);
2768 
2769 	gb = (inactive + active) >> (30 - PAGE_SHIFT);
2770 	if (gb)
2771 		inactive_ratio = int_sqrt(10 * gb);
2772 	else
2773 		inactive_ratio = 1;
2774 
2775 	return inactive * inactive_ratio < active;
2776 }
2777 
2778 enum scan_balance {
2779 	SCAN_EQUAL,
2780 	SCAN_FRACT,
2781 	SCAN_ANON,
2782 	SCAN_FILE,
2783 };
2784 
prepare_scan_count(pg_data_t * pgdat,struct scan_control * sc)2785 static void prepare_scan_count(pg_data_t *pgdat, struct scan_control *sc)
2786 {
2787 	unsigned long file;
2788 	struct lruvec *target_lruvec;
2789 
2790 	if (lru_gen_enabled())
2791 		return;
2792 
2793 	target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
2794 
2795 	/*
2796 	 * Flush the memory cgroup stats, so that we read accurate per-memcg
2797 	 * lruvec stats for heuristics.
2798 	 */
2799 	mem_cgroup_flush_stats();
2800 
2801 	/*
2802 	 * Determine the scan balance between anon and file LRUs.
2803 	 */
2804 	spin_lock_irq(&target_lruvec->lru_lock);
2805 	sc->anon_cost = target_lruvec->anon_cost;
2806 	sc->file_cost = target_lruvec->file_cost;
2807 	spin_unlock_irq(&target_lruvec->lru_lock);
2808 
2809 	/*
2810 	 * Target desirable inactive:active list ratios for the anon
2811 	 * and file LRU lists.
2812 	 */
2813 	if (!sc->force_deactivate) {
2814 		unsigned long refaults;
2815 
2816 		/*
2817 		 * When refaults are being observed, it means a new
2818 		 * workingset is being established. Deactivate to get
2819 		 * rid of any stale active pages quickly.
2820 		 */
2821 		refaults = lruvec_page_state(target_lruvec,
2822 				WORKINGSET_ACTIVATE_ANON);
2823 		if (refaults != target_lruvec->refaults[WORKINGSET_ANON] ||
2824 			inactive_is_low(target_lruvec, LRU_INACTIVE_ANON))
2825 			sc->may_deactivate |= DEACTIVATE_ANON;
2826 		else
2827 			sc->may_deactivate &= ~DEACTIVATE_ANON;
2828 
2829 		refaults = lruvec_page_state(target_lruvec,
2830 				WORKINGSET_ACTIVATE_FILE);
2831 		if (refaults != target_lruvec->refaults[WORKINGSET_FILE] ||
2832 		    inactive_is_low(target_lruvec, LRU_INACTIVE_FILE))
2833 			sc->may_deactivate |= DEACTIVATE_FILE;
2834 		else
2835 			sc->may_deactivate &= ~DEACTIVATE_FILE;
2836 	} else
2837 		sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE;
2838 
2839 	/*
2840 	 * If we have plenty of inactive file pages that aren't
2841 	 * thrashing, try to reclaim those first before touching
2842 	 * anonymous pages.
2843 	 */
2844 	file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
2845 	if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
2846 		sc->cache_trim_mode = 1;
2847 	else
2848 		sc->cache_trim_mode = 0;
2849 
2850 	/*
2851 	 * Prevent the reclaimer from falling into the cache trap: as
2852 	 * cache pages start out inactive, every cache fault will tip
2853 	 * the scan balance towards the file LRU.  And as the file LRU
2854 	 * shrinks, so does the window for rotation from references.
2855 	 * This means we have a runaway feedback loop where a tiny
2856 	 * thrashing file LRU becomes infinitely more attractive than
2857 	 * anon pages.  Try to detect this based on file LRU size.
2858 	 */
2859 	if (!cgroup_reclaim(sc)) {
2860 		unsigned long total_high_wmark = 0;
2861 		unsigned long free, anon;
2862 		int z;
2863 
2864 		free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2865 		file = node_page_state(pgdat, NR_ACTIVE_FILE) +
2866 			   node_page_state(pgdat, NR_INACTIVE_FILE);
2867 
2868 		for (z = 0; z < MAX_NR_ZONES; z++) {
2869 			struct zone *zone = &pgdat->node_zones[z];
2870 
2871 			if (!managed_zone(zone))
2872 				continue;
2873 
2874 			total_high_wmark += high_wmark_pages(zone);
2875 		}
2876 
2877 		/*
2878 		 * Consider anon: if that's low too, this isn't a
2879 		 * runaway file reclaim problem, but rather just
2880 		 * extreme pressure. Reclaim as per usual then.
2881 		 */
2882 		anon = node_page_state(pgdat, NR_INACTIVE_ANON);
2883 
2884 		sc->file_is_tiny =
2885 			file + free <= total_high_wmark &&
2886 			!(sc->may_deactivate & DEACTIVATE_ANON) &&
2887 			anon >> sc->priority;
2888 	}
2889 }
2890 
2891 /*
2892  * Determine how aggressively the anon and file LRU lists should be
2893  * scanned.
2894  *
2895  * nr[0] = anon inactive folios to scan; nr[1] = anon active folios to scan
2896  * nr[2] = file inactive folios to scan; nr[3] = file active folios to scan
2897  */
get_scan_count(struct lruvec * lruvec,struct scan_control * sc,unsigned long * nr)2898 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
2899 			   unsigned long *nr)
2900 {
2901 	struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2902 	struct mem_cgroup *memcg = lruvec_memcg(lruvec);
2903 	unsigned long anon_cost, file_cost, total_cost;
2904 	int swappiness = mem_cgroup_swappiness(memcg);
2905 	u64 fraction[ANON_AND_FILE];
2906 	u64 denominator = 0;	/* gcc */
2907 	enum scan_balance scan_balance;
2908 	unsigned long ap, fp;
2909 	enum lru_list lru;
2910 
2911 	/* If we have no swap space, do not bother scanning anon folios. */
2912 	if (!sc->may_swap || !can_reclaim_anon_pages(memcg, pgdat->node_id, sc)) {
2913 		scan_balance = SCAN_FILE;
2914 		goto out;
2915 	}
2916 
2917 	/*
2918 	 * Global reclaim will swap to prevent OOM even with no
2919 	 * swappiness, but memcg users want to use this knob to
2920 	 * disable swapping for individual groups completely when
2921 	 * using the memory controller's swap limit feature would be
2922 	 * too expensive.
2923 	 */
2924 	if (cgroup_reclaim(sc) && !swappiness) {
2925 		scan_balance = SCAN_FILE;
2926 		goto out;
2927 	}
2928 
2929 	/*
2930 	 * Do not apply any pressure balancing cleverness when the
2931 	 * system is close to OOM, scan both anon and file equally
2932 	 * (unless the swappiness setting disagrees with swapping).
2933 	 */
2934 	if (!sc->priority && swappiness) {
2935 		scan_balance = SCAN_EQUAL;
2936 		goto out;
2937 	}
2938 
2939 	/*
2940 	 * If the system is almost out of file pages, force-scan anon.
2941 	 */
2942 	if (sc->file_is_tiny) {
2943 		scan_balance = SCAN_ANON;
2944 		goto out;
2945 	}
2946 
2947 	/*
2948 	 * If there is enough inactive page cache, we do not reclaim
2949 	 * anything from the anonymous working right now.
2950 	 */
2951 	if (sc->cache_trim_mode) {
2952 		scan_balance = SCAN_FILE;
2953 		goto out;
2954 	}
2955 
2956 	scan_balance = SCAN_FRACT;
2957 	/*
2958 	 * Calculate the pressure balance between anon and file pages.
2959 	 *
2960 	 * The amount of pressure we put on each LRU is inversely
2961 	 * proportional to the cost of reclaiming each list, as
2962 	 * determined by the share of pages that are refaulting, times
2963 	 * the relative IO cost of bringing back a swapped out
2964 	 * anonymous page vs reloading a filesystem page (swappiness).
2965 	 *
2966 	 * Although we limit that influence to ensure no list gets
2967 	 * left behind completely: at least a third of the pressure is
2968 	 * applied, before swappiness.
2969 	 *
2970 	 * With swappiness at 100, anon and file have equal IO cost.
2971 	 */
2972 	total_cost = sc->anon_cost + sc->file_cost;
2973 	anon_cost = total_cost + sc->anon_cost;
2974 	file_cost = total_cost + sc->file_cost;
2975 	total_cost = anon_cost + file_cost;
2976 
2977 	ap = swappiness * (total_cost + 1);
2978 	ap /= anon_cost + 1;
2979 
2980 	fp = (200 - swappiness) * (total_cost + 1);
2981 	fp /= file_cost + 1;
2982 
2983 	fraction[0] = ap;
2984 	fraction[1] = fp;
2985 	denominator = ap + fp;
2986 out:
2987 	for_each_evictable_lru(lru) {
2988 		int file = is_file_lru(lru);
2989 		unsigned long lruvec_size;
2990 		unsigned long low, min;
2991 		unsigned long scan;
2992 
2993 		lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2994 		mem_cgroup_protection(sc->target_mem_cgroup, memcg,
2995 				      &min, &low);
2996 
2997 		if (min || low) {
2998 			/*
2999 			 * Scale a cgroup's reclaim pressure by proportioning
3000 			 * its current usage to its memory.low or memory.min
3001 			 * setting.
3002 			 *
3003 			 * This is important, as otherwise scanning aggression
3004 			 * becomes extremely binary -- from nothing as we
3005 			 * approach the memory protection threshold, to totally
3006 			 * nominal as we exceed it.  This results in requiring
3007 			 * setting extremely liberal protection thresholds. It
3008 			 * also means we simply get no protection at all if we
3009 			 * set it too low, which is not ideal.
3010 			 *
3011 			 * If there is any protection in place, we reduce scan
3012 			 * pressure by how much of the total memory used is
3013 			 * within protection thresholds.
3014 			 *
3015 			 * There is one special case: in the first reclaim pass,
3016 			 * we skip over all groups that are within their low
3017 			 * protection. If that fails to reclaim enough pages to
3018 			 * satisfy the reclaim goal, we come back and override
3019 			 * the best-effort low protection. However, we still
3020 			 * ideally want to honor how well-behaved groups are in
3021 			 * that case instead of simply punishing them all
3022 			 * equally. As such, we reclaim them based on how much
3023 			 * memory they are using, reducing the scan pressure
3024 			 * again by how much of the total memory used is under
3025 			 * hard protection.
3026 			 */
3027 			unsigned long cgroup_size = mem_cgroup_size(memcg);
3028 			unsigned long protection;
3029 
3030 			/* memory.low scaling, make sure we retry before OOM */
3031 			if (!sc->memcg_low_reclaim && low > min) {
3032 				protection = low;
3033 				sc->memcg_low_skipped = 1;
3034 			} else {
3035 				protection = min;
3036 			}
3037 
3038 			/* Avoid TOCTOU with earlier protection check */
3039 			cgroup_size = max(cgroup_size, protection);
3040 
3041 			scan = lruvec_size - lruvec_size * protection /
3042 				(cgroup_size + 1);
3043 
3044 			/*
3045 			 * Minimally target SWAP_CLUSTER_MAX pages to keep
3046 			 * reclaim moving forwards, avoiding decrementing
3047 			 * sc->priority further than desirable.
3048 			 */
3049 			scan = max(scan, SWAP_CLUSTER_MAX);
3050 		} else {
3051 			scan = lruvec_size;
3052 		}
3053 
3054 		scan >>= sc->priority;
3055 
3056 		/*
3057 		 * If the cgroup's already been deleted, make sure to
3058 		 * scrape out the remaining cache.
3059 		 */
3060 		if (!scan && !mem_cgroup_online(memcg))
3061 			scan = min(lruvec_size, SWAP_CLUSTER_MAX);
3062 
3063 		switch (scan_balance) {
3064 		case SCAN_EQUAL:
3065 			/* Scan lists relative to size */
3066 			break;
3067 		case SCAN_FRACT:
3068 			/*
3069 			 * Scan types proportional to swappiness and
3070 			 * their relative recent reclaim efficiency.
3071 			 * Make sure we don't miss the last page on
3072 			 * the offlined memory cgroups because of a
3073 			 * round-off error.
3074 			 */
3075 			scan = mem_cgroup_online(memcg) ?
3076 			       div64_u64(scan * fraction[file], denominator) :
3077 			       DIV64_U64_ROUND_UP(scan * fraction[file],
3078 						  denominator);
3079 			break;
3080 		case SCAN_FILE:
3081 		case SCAN_ANON:
3082 			/* Scan one type exclusively */
3083 			if ((scan_balance == SCAN_FILE) != file)
3084 				scan = 0;
3085 			break;
3086 		default:
3087 			/* Look ma, no brain */
3088 			BUG();
3089 		}
3090 
3091 		nr[lru] = scan;
3092 	}
3093 }
3094 
3095 /*
3096  * Anonymous LRU management is a waste if there is
3097  * ultimately no way to reclaim the memory.
3098  */
can_age_anon_pages(struct pglist_data * pgdat,struct scan_control * sc)3099 static bool can_age_anon_pages(struct pglist_data *pgdat,
3100 			       struct scan_control *sc)
3101 {
3102 	/* Aging the anon LRU is valuable if swap is present: */
3103 	if (total_swap_pages > 0)
3104 		return true;
3105 
3106 	/* Also valuable if anon pages can be demoted: */
3107 	return can_demote(pgdat->node_id, sc);
3108 }
3109 
3110 #ifdef CONFIG_LRU_GEN
3111 
3112 #ifdef CONFIG_LRU_GEN_ENABLED
3113 DEFINE_STATIC_KEY_ARRAY_TRUE(lru_gen_caps, NR_LRU_GEN_CAPS);
3114 #define get_cap(cap)	static_branch_likely(&lru_gen_caps[cap])
3115 #else
3116 DEFINE_STATIC_KEY_ARRAY_FALSE(lru_gen_caps, NR_LRU_GEN_CAPS);
3117 #define get_cap(cap)	static_branch_unlikely(&lru_gen_caps[cap])
3118 #endif
3119 
3120 /******************************************************************************
3121  *                          shorthand helpers
3122  ******************************************************************************/
3123 
3124 #define LRU_REFS_FLAGS	(BIT(PG_referenced) | BIT(PG_workingset))
3125 
3126 #define DEFINE_MAX_SEQ(lruvec)						\
3127 	unsigned long max_seq = READ_ONCE((lruvec)->lrugen.max_seq)
3128 
3129 #define DEFINE_MIN_SEQ(lruvec)						\
3130 	unsigned long min_seq[ANON_AND_FILE] = {			\
3131 		READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_ANON]),	\
3132 		READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_FILE]),	\
3133 	}
3134 
3135 #define for_each_gen_type_zone(gen, type, zone)				\
3136 	for ((gen) = 0; (gen) < MAX_NR_GENS; (gen)++)			\
3137 		for ((type) = 0; (type) < ANON_AND_FILE; (type)++)	\
3138 			for ((zone) = 0; (zone) < MAX_NR_ZONES; (zone)++)
3139 
get_lruvec(struct mem_cgroup * memcg,int nid)3140 static struct lruvec *get_lruvec(struct mem_cgroup *memcg, int nid)
3141 {
3142 	struct pglist_data *pgdat = NODE_DATA(nid);
3143 
3144 #ifdef CONFIG_MEMCG
3145 	if (memcg) {
3146 		struct lruvec *lruvec = &memcg->nodeinfo[nid]->lruvec;
3147 
3148 		/* for hotadd_new_pgdat() */
3149 		if (!lruvec->pgdat)
3150 			lruvec->pgdat = pgdat;
3151 
3152 		return lruvec;
3153 	}
3154 #endif
3155 	VM_WARN_ON_ONCE(!mem_cgroup_disabled());
3156 
3157 	return pgdat ? &pgdat->__lruvec : NULL;
3158 }
3159 
get_swappiness(struct lruvec * lruvec,struct scan_control * sc)3160 static int get_swappiness(struct lruvec *lruvec, struct scan_control *sc)
3161 {
3162 	struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3163 	struct pglist_data *pgdat = lruvec_pgdat(lruvec);
3164 
3165 	if (!can_demote(pgdat->node_id, sc) &&
3166 	    mem_cgroup_get_nr_swap_pages(memcg) < MIN_LRU_BATCH)
3167 		return 0;
3168 
3169 	return mem_cgroup_swappiness(memcg);
3170 }
3171 
get_nr_gens(struct lruvec * lruvec,int type)3172 static int get_nr_gens(struct lruvec *lruvec, int type)
3173 {
3174 	return lruvec->lrugen.max_seq - lruvec->lrugen.min_seq[type] + 1;
3175 }
3176 
seq_is_valid(struct lruvec * lruvec)3177 static bool __maybe_unused seq_is_valid(struct lruvec *lruvec)
3178 {
3179 	/* see the comment on lru_gen_struct */
3180 	return get_nr_gens(lruvec, LRU_GEN_FILE) >= MIN_NR_GENS &&
3181 	       get_nr_gens(lruvec, LRU_GEN_FILE) <= get_nr_gens(lruvec, LRU_GEN_ANON) &&
3182 	       get_nr_gens(lruvec, LRU_GEN_ANON) <= MAX_NR_GENS;
3183 }
3184 
3185 /******************************************************************************
3186  *                          mm_struct list
3187  ******************************************************************************/
3188 
get_mm_list(struct mem_cgroup * memcg)3189 static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg)
3190 {
3191 	static struct lru_gen_mm_list mm_list = {
3192 		.fifo = LIST_HEAD_INIT(mm_list.fifo),
3193 		.lock = __SPIN_LOCK_UNLOCKED(mm_list.lock),
3194 	};
3195 
3196 #ifdef CONFIG_MEMCG
3197 	if (memcg)
3198 		return &memcg->mm_list;
3199 #endif
3200 	VM_WARN_ON_ONCE(!mem_cgroup_disabled());
3201 
3202 	return &mm_list;
3203 }
3204 
lru_gen_add_mm(struct mm_struct * mm)3205 void lru_gen_add_mm(struct mm_struct *mm)
3206 {
3207 	int nid;
3208 	struct mem_cgroup *memcg = get_mem_cgroup_from_mm(mm);
3209 	struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
3210 
3211 	VM_WARN_ON_ONCE(!list_empty(&mm->lru_gen.list));
3212 #ifdef CONFIG_MEMCG
3213 	VM_WARN_ON_ONCE(mm->lru_gen.memcg);
3214 	mm->lru_gen.memcg = memcg;
3215 #endif
3216 	spin_lock(&mm_list->lock);
3217 
3218 	for_each_node_state(nid, N_MEMORY) {
3219 		struct lruvec *lruvec = get_lruvec(memcg, nid);
3220 
3221 		if (!lruvec)
3222 			continue;
3223 
3224 		/* the first addition since the last iteration */
3225 		if (lruvec->mm_state.tail == &mm_list->fifo)
3226 			lruvec->mm_state.tail = &mm->lru_gen.list;
3227 	}
3228 
3229 	list_add_tail(&mm->lru_gen.list, &mm_list->fifo);
3230 
3231 	spin_unlock(&mm_list->lock);
3232 }
3233 
lru_gen_del_mm(struct mm_struct * mm)3234 void lru_gen_del_mm(struct mm_struct *mm)
3235 {
3236 	int nid;
3237 	struct lru_gen_mm_list *mm_list;
3238 	struct mem_cgroup *memcg = NULL;
3239 
3240 	if (list_empty(&mm->lru_gen.list))
3241 		return;
3242 
3243 #ifdef CONFIG_MEMCG
3244 	memcg = mm->lru_gen.memcg;
3245 #endif
3246 	mm_list = get_mm_list(memcg);
3247 
3248 	spin_lock(&mm_list->lock);
3249 
3250 	for_each_node(nid) {
3251 		struct lruvec *lruvec = get_lruvec(memcg, nid);
3252 
3253 		if (!lruvec)
3254 			continue;
3255 
3256 		/* where the last iteration ended (exclusive) */
3257 		if (lruvec->mm_state.tail == &mm->lru_gen.list)
3258 			lruvec->mm_state.tail = lruvec->mm_state.tail->next;
3259 
3260 		/* where the current iteration continues (inclusive) */
3261 		if (lruvec->mm_state.head != &mm->lru_gen.list)
3262 			continue;
3263 
3264 		lruvec->mm_state.head = lruvec->mm_state.head->next;
3265 		/* the deletion ends the current iteration */
3266 		if (lruvec->mm_state.head == &mm_list->fifo)
3267 			WRITE_ONCE(lruvec->mm_state.seq, lruvec->mm_state.seq + 1);
3268 	}
3269 
3270 	list_del_init(&mm->lru_gen.list);
3271 
3272 	spin_unlock(&mm_list->lock);
3273 
3274 #ifdef CONFIG_MEMCG
3275 	mem_cgroup_put(mm->lru_gen.memcg);
3276 	mm->lru_gen.memcg = NULL;
3277 #endif
3278 }
3279 
3280 #ifdef CONFIG_MEMCG
lru_gen_migrate_mm(struct mm_struct * mm)3281 void lru_gen_migrate_mm(struct mm_struct *mm)
3282 {
3283 	struct mem_cgroup *memcg;
3284 	struct task_struct *task = rcu_dereference_protected(mm->owner, true);
3285 
3286 	VM_WARN_ON_ONCE(task->mm != mm);
3287 	lockdep_assert_held(&task->alloc_lock);
3288 
3289 	/* for mm_update_next_owner() */
3290 	if (mem_cgroup_disabled())
3291 		return;
3292 
3293 	rcu_read_lock();
3294 	memcg = mem_cgroup_from_task(task);
3295 	rcu_read_unlock();
3296 	if (memcg == mm->lru_gen.memcg)
3297 		return;
3298 
3299 	VM_WARN_ON_ONCE(!mm->lru_gen.memcg);
3300 	VM_WARN_ON_ONCE(list_empty(&mm->lru_gen.list));
3301 
3302 	lru_gen_del_mm(mm);
3303 	lru_gen_add_mm(mm);
3304 }
3305 #endif
3306 
3307 /*
3308  * Bloom filters with m=1<<15, k=2 and the false positive rates of ~1/5 when
3309  * n=10,000 and ~1/2 when n=20,000, where, conventionally, m is the number of
3310  * bits in a bitmap, k is the number of hash functions and n is the number of
3311  * inserted items.
3312  *
3313  * Page table walkers use one of the two filters to reduce their search space.
3314  * To get rid of non-leaf entries that no longer have enough leaf entries, the
3315  * aging uses the double-buffering technique to flip to the other filter each
3316  * time it produces a new generation. For non-leaf entries that have enough
3317  * leaf entries, the aging carries them over to the next generation in
3318  * walk_pmd_range(); the eviction also report them when walking the rmap
3319  * in lru_gen_look_around().
3320  *
3321  * For future optimizations:
3322  * 1. It's not necessary to keep both filters all the time. The spare one can be
3323  *    freed after the RCU grace period and reallocated if needed again.
3324  * 2. And when reallocating, it's worth scaling its size according to the number
3325  *    of inserted entries in the other filter, to reduce the memory overhead on
3326  *    small systems and false positives on large systems.
3327  * 3. Jenkins' hash function is an alternative to Knuth's.
3328  */
3329 #define BLOOM_FILTER_SHIFT	15
3330 
filter_gen_from_seq(unsigned long seq)3331 static inline int filter_gen_from_seq(unsigned long seq)
3332 {
3333 	return seq % NR_BLOOM_FILTERS;
3334 }
3335 
get_item_key(void * item,int * key)3336 static void get_item_key(void *item, int *key)
3337 {
3338 	u32 hash = hash_ptr(item, BLOOM_FILTER_SHIFT * 2);
3339 
3340 	BUILD_BUG_ON(BLOOM_FILTER_SHIFT * 2 > BITS_PER_TYPE(u32));
3341 
3342 	key[0] = hash & (BIT(BLOOM_FILTER_SHIFT) - 1);
3343 	key[1] = hash >> BLOOM_FILTER_SHIFT;
3344 }
3345 
reset_bloom_filter(struct lruvec * lruvec,unsigned long seq)3346 static void reset_bloom_filter(struct lruvec *lruvec, unsigned long seq)
3347 {
3348 	unsigned long *filter;
3349 	int gen = filter_gen_from_seq(seq);
3350 
3351 	filter = lruvec->mm_state.filters[gen];
3352 	if (filter) {
3353 		bitmap_clear(filter, 0, BIT(BLOOM_FILTER_SHIFT));
3354 		return;
3355 	}
3356 
3357 	filter = bitmap_zalloc(BIT(BLOOM_FILTER_SHIFT),
3358 			       __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
3359 	WRITE_ONCE(lruvec->mm_state.filters[gen], filter);
3360 }
3361 
update_bloom_filter(struct lruvec * lruvec,unsigned long seq,void * item)3362 static void update_bloom_filter(struct lruvec *lruvec, unsigned long seq, void *item)
3363 {
3364 	int key[2];
3365 	unsigned long *filter;
3366 	int gen = filter_gen_from_seq(seq);
3367 
3368 	filter = READ_ONCE(lruvec->mm_state.filters[gen]);
3369 	if (!filter)
3370 		return;
3371 
3372 	get_item_key(item, key);
3373 
3374 	if (!test_bit(key[0], filter))
3375 		set_bit(key[0], filter);
3376 	if (!test_bit(key[1], filter))
3377 		set_bit(key[1], filter);
3378 }
3379 
test_bloom_filter(struct lruvec * lruvec,unsigned long seq,void * item)3380 static bool test_bloom_filter(struct lruvec *lruvec, unsigned long seq, void *item)
3381 {
3382 	int key[2];
3383 	unsigned long *filter;
3384 	int gen = filter_gen_from_seq(seq);
3385 
3386 	filter = READ_ONCE(lruvec->mm_state.filters[gen]);
3387 	if (!filter)
3388 		return true;
3389 
3390 	get_item_key(item, key);
3391 
3392 	return test_bit(key[0], filter) && test_bit(key[1], filter);
3393 }
3394 
reset_mm_stats(struct lruvec * lruvec,struct lru_gen_mm_walk * walk,bool last)3395 static void reset_mm_stats(struct lruvec *lruvec, struct lru_gen_mm_walk *walk, bool last)
3396 {
3397 	int i;
3398 	int hist;
3399 
3400 	lockdep_assert_held(&get_mm_list(lruvec_memcg(lruvec))->lock);
3401 
3402 	if (walk) {
3403 		hist = lru_hist_from_seq(walk->max_seq);
3404 
3405 		for (i = 0; i < NR_MM_STATS; i++) {
3406 			WRITE_ONCE(lruvec->mm_state.stats[hist][i],
3407 				   lruvec->mm_state.stats[hist][i] + walk->mm_stats[i]);
3408 			walk->mm_stats[i] = 0;
3409 		}
3410 	}
3411 
3412 	if (NR_HIST_GENS > 1 && last) {
3413 		hist = lru_hist_from_seq(lruvec->mm_state.seq + 1);
3414 
3415 		for (i = 0; i < NR_MM_STATS; i++)
3416 			WRITE_ONCE(lruvec->mm_state.stats[hist][i], 0);
3417 	}
3418 }
3419 
should_skip_mm(struct mm_struct * mm,struct lru_gen_mm_walk * walk)3420 static bool should_skip_mm(struct mm_struct *mm, struct lru_gen_mm_walk *walk)
3421 {
3422 	int type;
3423 	unsigned long size = 0;
3424 	struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3425 	int key = pgdat->node_id % BITS_PER_TYPE(mm->lru_gen.bitmap);
3426 
3427 	if (!walk->force_scan && !test_bit(key, &mm->lru_gen.bitmap))
3428 		return true;
3429 
3430 	clear_bit(key, &mm->lru_gen.bitmap);
3431 
3432 	for (type = !walk->can_swap; type < ANON_AND_FILE; type++) {
3433 		size += type ? get_mm_counter(mm, MM_FILEPAGES) :
3434 			       get_mm_counter(mm, MM_ANONPAGES) +
3435 			       get_mm_counter(mm, MM_SHMEMPAGES);
3436 	}
3437 
3438 	if (size < MIN_LRU_BATCH)
3439 		return true;
3440 
3441 	return !mmget_not_zero(mm);
3442 }
3443 
iterate_mm_list(struct lruvec * lruvec,struct lru_gen_mm_walk * walk,struct mm_struct ** iter)3444 static bool iterate_mm_list(struct lruvec *lruvec, struct lru_gen_mm_walk *walk,
3445 			    struct mm_struct **iter)
3446 {
3447 	bool first = false;
3448 	bool last = true;
3449 	struct mm_struct *mm = NULL;
3450 	struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3451 	struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
3452 	struct lru_gen_mm_state *mm_state = &lruvec->mm_state;
3453 
3454 	/*
3455 	 * There are four interesting cases for this page table walker:
3456 	 * 1. It tries to start a new iteration of mm_list with a stale max_seq;
3457 	 *    there is nothing left to do.
3458 	 * 2. It's the first of the current generation, and it needs to reset
3459 	 *    the Bloom filter for the next generation.
3460 	 * 3. It reaches the end of mm_list, and it needs to increment
3461 	 *    mm_state->seq; the iteration is done.
3462 	 * 4. It's the last of the current generation, and it needs to reset the
3463 	 *    mm stats counters for the next generation.
3464 	 */
3465 	spin_lock(&mm_list->lock);
3466 
3467 	VM_WARN_ON_ONCE(mm_state->seq + 1 < walk->max_seq);
3468 	VM_WARN_ON_ONCE(*iter && mm_state->seq > walk->max_seq);
3469 	VM_WARN_ON_ONCE(*iter && !mm_state->nr_walkers);
3470 
3471 	if (walk->max_seq <= mm_state->seq) {
3472 		if (!*iter)
3473 			last = false;
3474 		goto done;
3475 	}
3476 
3477 	if (!mm_state->nr_walkers) {
3478 		VM_WARN_ON_ONCE(mm_state->head && mm_state->head != &mm_list->fifo);
3479 
3480 		mm_state->head = mm_list->fifo.next;
3481 		first = true;
3482 	}
3483 
3484 	while (!mm && mm_state->head != &mm_list->fifo) {
3485 		mm = list_entry(mm_state->head, struct mm_struct, lru_gen.list);
3486 
3487 		mm_state->head = mm_state->head->next;
3488 
3489 		/* force scan for those added after the last iteration */
3490 		if (!mm_state->tail || mm_state->tail == &mm->lru_gen.list) {
3491 			mm_state->tail = mm_state->head;
3492 			walk->force_scan = true;
3493 		}
3494 
3495 		if (should_skip_mm(mm, walk))
3496 			mm = NULL;
3497 	}
3498 
3499 	if (mm_state->head == &mm_list->fifo)
3500 		WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
3501 done:
3502 	if (*iter && !mm)
3503 		mm_state->nr_walkers--;
3504 	if (!*iter && mm)
3505 		mm_state->nr_walkers++;
3506 
3507 	if (mm_state->nr_walkers)
3508 		last = false;
3509 
3510 	if (*iter || last)
3511 		reset_mm_stats(lruvec, walk, last);
3512 
3513 	spin_unlock(&mm_list->lock);
3514 
3515 	if (mm && first)
3516 		reset_bloom_filter(lruvec, walk->max_seq + 1);
3517 
3518 	if (*iter)
3519 		mmput_async(*iter);
3520 
3521 	*iter = mm;
3522 
3523 	return last;
3524 }
3525 
iterate_mm_list_nowalk(struct lruvec * lruvec,unsigned long max_seq)3526 static bool iterate_mm_list_nowalk(struct lruvec *lruvec, unsigned long max_seq)
3527 {
3528 	bool success = false;
3529 	struct mem_cgroup *memcg = lruvec_memcg(lruvec);
3530 	struct lru_gen_mm_list *mm_list = get_mm_list(memcg);
3531 	struct lru_gen_mm_state *mm_state = &lruvec->mm_state;
3532 
3533 	spin_lock(&mm_list->lock);
3534 
3535 	VM_WARN_ON_ONCE(mm_state->seq + 1 < max_seq);
3536 
3537 	if (max_seq > mm_state->seq && !mm_state->nr_walkers) {
3538 		VM_WARN_ON_ONCE(mm_state->head && mm_state->head != &mm_list->fifo);
3539 
3540 		WRITE_ONCE(mm_state->seq, mm_state->seq + 1);
3541 		reset_mm_stats(lruvec, NULL, true);
3542 		success = true;
3543 	}
3544 
3545 	spin_unlock(&mm_list->lock);
3546 
3547 	return success;
3548 }
3549 
3550 /******************************************************************************
3551  *                          refault feedback loop
3552  ******************************************************************************/
3553 
3554 /*
3555  * A feedback loop based on Proportional-Integral-Derivative (PID) controller.
3556  *
3557  * The P term is refaulted/(evicted+protected) from a tier in the generation
3558  * currently being evicted; the I term is the exponential moving average of the
3559  * P term over the generations previously evicted, using the smoothing factor
3560  * 1/2; the D term isn't supported.
3561  *
3562  * The setpoint (SP) is always the first tier of one type; the process variable
3563  * (PV) is either any tier of the other type or any other tier of the same
3564  * type.
3565  *
3566  * The error is the difference between the SP and the PV; the correction is to
3567  * turn off protection when SP>PV or turn on protection when SP<PV.
3568  *
3569  * For future optimizations:
3570  * 1. The D term may discount the other two terms over time so that long-lived
3571  *    generations can resist stale information.
3572  */
3573 struct ctrl_pos {
3574 	unsigned long refaulted;
3575 	unsigned long total;
3576 	int gain;
3577 };
3578 
read_ctrl_pos(struct lruvec * lruvec,int type,int tier,int gain,struct ctrl_pos * pos)3579 static void read_ctrl_pos(struct lruvec *lruvec, int type, int tier, int gain,
3580 			  struct ctrl_pos *pos)
3581 {
3582 	struct lru_gen_struct *lrugen = &lruvec->lrugen;
3583 	int hist = lru_hist_from_seq(lrugen->min_seq[type]);
3584 
3585 	pos->refaulted = lrugen->avg_refaulted[type][tier] +
3586 			 atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3587 	pos->total = lrugen->avg_total[type][tier] +
3588 		     atomic_long_read(&lrugen->evicted[hist][type][tier]);
3589 	if (tier)
3590 		pos->total += lrugen->protected[hist][type][tier - 1];
3591 	pos->gain = gain;
3592 }
3593 
reset_ctrl_pos(struct lruvec * lruvec,int type,bool carryover)3594 static void reset_ctrl_pos(struct lruvec *lruvec, int type, bool carryover)
3595 {
3596 	int hist, tier;
3597 	struct lru_gen_struct *lrugen = &lruvec->lrugen;
3598 	bool clear = carryover ? NR_HIST_GENS == 1 : NR_HIST_GENS > 1;
3599 	unsigned long seq = carryover ? lrugen->min_seq[type] : lrugen->max_seq + 1;
3600 
3601 	lockdep_assert_held(&lruvec->lru_lock);
3602 
3603 	if (!carryover && !clear)
3604 		return;
3605 
3606 	hist = lru_hist_from_seq(seq);
3607 
3608 	for (tier = 0; tier < MAX_NR_TIERS; tier++) {
3609 		if (carryover) {
3610 			unsigned long sum;
3611 
3612 			sum = lrugen->avg_refaulted[type][tier] +
3613 			      atomic_long_read(&lrugen->refaulted[hist][type][tier]);
3614 			WRITE_ONCE(lrugen->avg_refaulted[type][tier], sum / 2);
3615 
3616 			sum = lrugen->avg_total[type][tier] +
3617 			      atomic_long_read(&lrugen->evicted[hist][type][tier]);
3618 			if (tier)
3619 				sum += lrugen->protected[hist][type][tier - 1];
3620 			WRITE_ONCE(lrugen->avg_total[type][tier], sum / 2);
3621 		}
3622 
3623 		if (clear) {
3624 			atomic_long_set(&lrugen->refaulted[hist][type][tier], 0);
3625 			atomic_long_set(&lrugen->evicted[hist][type][tier], 0);
3626 			if (tier)
3627 				WRITE_ONCE(lrugen->protected[hist][type][tier - 1], 0);
3628 		}
3629 	}
3630 }
3631 
positive_ctrl_err(struct ctrl_pos * sp,struct ctrl_pos * pv)3632 static bool positive_ctrl_err(struct ctrl_pos *sp, struct ctrl_pos *pv)
3633 {
3634 	/*
3635 	 * Return true if the PV has a limited number of refaults or a lower
3636 	 * refaulted/total than the SP.
3637 	 */
3638 	return pv->refaulted < MIN_LRU_BATCH ||
3639 	       pv->refaulted * (sp->total + MIN_LRU_BATCH) * sp->gain <=
3640 	       (sp->refaulted + 1) * pv->total * pv->gain;
3641 }
3642 
3643 /******************************************************************************
3644  *                          the aging
3645  ******************************************************************************/
3646 
3647 /* promote pages accessed through page tables */
folio_update_gen(struct folio * folio,int gen)3648 static int folio_update_gen(struct folio *folio, int gen)
3649 {
3650 	unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
3651 
3652 	VM_WARN_ON_ONCE(gen >= MAX_NR_GENS);
3653 	VM_WARN_ON_ONCE(!rcu_read_lock_held());
3654 
3655 	do {
3656 		/* lru_gen_del_folio() has isolated this page? */
3657 		if (!(old_flags & LRU_GEN_MASK)) {
3658 			/* for shrink_folio_list() */
3659 			new_flags = old_flags | BIT(PG_referenced);
3660 			continue;
3661 		}
3662 
3663 		new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS);
3664 		new_flags |= (gen + 1UL) << LRU_GEN_PGOFF;
3665 	} while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
3666 
3667 	return ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3668 }
3669 
3670 /* protect pages accessed multiple times through file descriptors */
folio_inc_gen(struct lruvec * lruvec,struct folio * folio,bool reclaiming)3671 static int folio_inc_gen(struct lruvec *lruvec, struct folio *folio, bool reclaiming)
3672 {
3673 	int type = folio_is_file_lru(folio);
3674 	struct lru_gen_struct *lrugen = &lruvec->lrugen;
3675 	int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
3676 	unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
3677 
3678 	VM_WARN_ON_ONCE_FOLIO(!(old_flags & LRU_GEN_MASK), folio);
3679 
3680 	do {
3681 		new_gen = ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1;
3682 		/* folio_update_gen() has promoted this page? */
3683 		if (new_gen >= 0 && new_gen != old_gen)
3684 			return new_gen;
3685 
3686 		new_gen = (old_gen + 1) % MAX_NR_GENS;
3687 
3688 		new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS);
3689 		new_flags |= (new_gen + 1UL) << LRU_GEN_PGOFF;
3690 		/* for folio_end_writeback() */
3691 		if (reclaiming)
3692 			new_flags |= BIT(PG_reclaim);
3693 	} while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
3694 
3695 	lru_gen_update_size(lruvec, folio, old_gen, new_gen);
3696 
3697 	return new_gen;
3698 }
3699 
update_batch_size(struct lru_gen_mm_walk * walk,struct folio * folio,int old_gen,int new_gen)3700 static void update_batch_size(struct lru_gen_mm_walk *walk, struct folio *folio,
3701 			      int old_gen, int new_gen)
3702 {
3703 	int type = folio_is_file_lru(folio);
3704 	int zone = folio_zonenum(folio);
3705 	int delta = folio_nr_pages(folio);
3706 
3707 	VM_WARN_ON_ONCE(old_gen >= MAX_NR_GENS);
3708 	VM_WARN_ON_ONCE(new_gen >= MAX_NR_GENS);
3709 
3710 	walk->batched++;
3711 
3712 	walk->nr_pages[old_gen][type][zone] -= delta;
3713 	walk->nr_pages[new_gen][type][zone] += delta;
3714 }
3715 
reset_batch_size(struct lruvec * lruvec,struct lru_gen_mm_walk * walk)3716 static void reset_batch_size(struct lruvec *lruvec, struct lru_gen_mm_walk *walk)
3717 {
3718 	int gen, type, zone;
3719 	struct lru_gen_struct *lrugen = &lruvec->lrugen;
3720 
3721 	walk->batched = 0;
3722 
3723 	for_each_gen_type_zone(gen, type, zone) {
3724 		enum lru_list lru = type * LRU_INACTIVE_FILE;
3725 		int delta = walk->nr_pages[gen][type][zone];
3726 
3727 		if (!delta)
3728 			continue;
3729 
3730 		walk->nr_pages[gen][type][zone] = 0;
3731 		WRITE_ONCE(lrugen->nr_pages[gen][type][zone],
3732 			   lrugen->nr_pages[gen][type][zone] + delta);
3733 
3734 		if (lru_gen_is_active(lruvec, gen))
3735 			lru += LRU_ACTIVE;
3736 		__update_lru_size(lruvec, lru, zone, delta);
3737 	}
3738 }
3739 
should_skip_vma(unsigned long start,unsigned long end,struct mm_walk * args)3740 static int should_skip_vma(unsigned long start, unsigned long end, struct mm_walk *args)
3741 {
3742 	struct address_space *mapping;
3743 	struct vm_area_struct *vma = args->vma;
3744 	struct lru_gen_mm_walk *walk = args->private;
3745 
3746 	if (!vma_is_accessible(vma))
3747 		return true;
3748 
3749 	if (is_vm_hugetlb_page(vma))
3750 		return true;
3751 
3752 	if (vma->vm_flags & (VM_LOCKED | VM_SPECIAL | VM_SEQ_READ | VM_RAND_READ))
3753 		return true;
3754 
3755 	if (vma == get_gate_vma(vma->vm_mm))
3756 		return true;
3757 
3758 	if (vma_is_anonymous(vma))
3759 		return !walk->can_swap;
3760 
3761 	if (WARN_ON_ONCE(!vma->vm_file || !vma->vm_file->f_mapping))
3762 		return true;
3763 
3764 	mapping = vma->vm_file->f_mapping;
3765 	if (mapping_unevictable(mapping))
3766 		return true;
3767 
3768 	if (shmem_mapping(mapping))
3769 		return !walk->can_swap;
3770 
3771 	/* to exclude special mappings like dax, etc. */
3772 	return !mapping->a_ops->read_folio;
3773 }
3774 
3775 /*
3776  * Some userspace memory allocators map many single-page VMAs. Instead of
3777  * returning back to the PGD table for each of such VMAs, finish an entire PMD
3778  * table to reduce zigzags and improve cache performance.
3779  */
get_next_vma(unsigned long mask,unsigned long size,struct mm_walk * args,unsigned long * vm_start,unsigned long * vm_end)3780 static bool get_next_vma(unsigned long mask, unsigned long size, struct mm_walk *args,
3781 			 unsigned long *vm_start, unsigned long *vm_end)
3782 {
3783 	unsigned long start = round_up(*vm_end, size);
3784 	unsigned long end = (start | ~mask) + 1;
3785 	VMA_ITERATOR(vmi, args->mm, start);
3786 
3787 	VM_WARN_ON_ONCE(mask & size);
3788 	VM_WARN_ON_ONCE((start & mask) != (*vm_start & mask));
3789 
3790 	for_each_vma(vmi, args->vma) {
3791 		if (end && end <= args->vma->vm_start)
3792 			return false;
3793 
3794 		if (should_skip_vma(args->vma->vm_start, args->vma->vm_end, args))
3795 			continue;
3796 
3797 		*vm_start = max(start, args->vma->vm_start);
3798 		*vm_end = min(end - 1, args->vma->vm_end - 1) + 1;
3799 
3800 		return true;
3801 	}
3802 
3803 	return false;
3804 }
3805 
get_pte_pfn(pte_t pte,struct vm_area_struct * vma,unsigned long addr)3806 static unsigned long get_pte_pfn(pte_t pte, struct vm_area_struct *vma, unsigned long addr)
3807 {
3808 	unsigned long pfn = pte_pfn(pte);
3809 
3810 	VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3811 
3812 	if (!pte_present(pte) || is_zero_pfn(pfn))
3813 		return -1;
3814 
3815 	if (WARN_ON_ONCE(pte_devmap(pte) || pte_special(pte)))
3816 		return -1;
3817 
3818 	if (WARN_ON_ONCE(!pfn_valid(pfn)))
3819 		return -1;
3820 
3821 	return pfn;
3822 }
3823 
3824 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG)
get_pmd_pfn(pmd_t pmd,struct vm_area_struct * vma,unsigned long addr)3825 static unsigned long get_pmd_pfn(pmd_t pmd, struct vm_area_struct *vma, unsigned long addr)
3826 {
3827 	unsigned long pfn = pmd_pfn(pmd);
3828 
3829 	VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end);
3830 
3831 	if (!pmd_present(pmd) || is_huge_zero_pmd(pmd))
3832 		return -1;
3833 
3834 	if (WARN_ON_ONCE(pmd_devmap(pmd)))
3835 		return -1;
3836 
3837 	if (WARN_ON_ONCE(!pfn_valid(pfn)))
3838 		return -1;
3839 
3840 	return pfn;
3841 }
3842 #endif
3843 
get_pfn_folio(unsigned long pfn,struct mem_cgroup * memcg,struct pglist_data * pgdat,bool can_swap)3844 static struct folio *get_pfn_folio(unsigned long pfn, struct mem_cgroup *memcg,
3845 				   struct pglist_data *pgdat, bool can_swap)
3846 {
3847 	struct folio *folio;
3848 
3849 	/* try to avoid unnecessary memory loads */
3850 	if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
3851 		return NULL;
3852 
3853 	folio = pfn_folio(pfn);
3854 	if (folio_nid(folio) != pgdat->node_id)
3855 		return NULL;
3856 
3857 	if (folio_memcg_rcu(folio) != memcg)
3858 		return NULL;
3859 
3860 	/* file VMAs can contain anon pages from COW */
3861 	if (!folio_is_file_lru(folio) && !can_swap)
3862 		return NULL;
3863 
3864 	return folio;
3865 }
3866 
suitable_to_scan(int total,int young)3867 static bool suitable_to_scan(int total, int young)
3868 {
3869 	int n = clamp_t(int, cache_line_size() / sizeof(pte_t), 2, 8);
3870 
3871 	/* suitable if the average number of young PTEs per cacheline is >=1 */
3872 	return young * n >= total;
3873 }
3874 
walk_pte_range(pmd_t * pmd,unsigned long start,unsigned long end,struct mm_walk * args)3875 static bool walk_pte_range(pmd_t *pmd, unsigned long start, unsigned long end,
3876 			   struct mm_walk *args)
3877 {
3878 	int i;
3879 	pte_t *pte;
3880 	spinlock_t *ptl;
3881 	unsigned long addr;
3882 	int total = 0;
3883 	int young = 0;
3884 	struct lru_gen_mm_walk *walk = args->private;
3885 	struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
3886 	struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3887 	int old_gen, new_gen = lru_gen_from_seq(walk->max_seq);
3888 
3889 	VM_WARN_ON_ONCE(pmd_leaf(*pmd));
3890 
3891 	ptl = pte_lockptr(args->mm, pmd);
3892 	if (!spin_trylock(ptl))
3893 		return false;
3894 
3895 	arch_enter_lazy_mmu_mode();
3896 
3897 	pte = pte_offset_map(pmd, start & PMD_MASK);
3898 restart:
3899 	for (i = pte_index(start), addr = start; addr != end; i++, addr += PAGE_SIZE) {
3900 		unsigned long pfn;
3901 		struct folio *folio;
3902 
3903 		total++;
3904 		walk->mm_stats[MM_LEAF_TOTAL]++;
3905 
3906 		pfn = get_pte_pfn(pte[i], args->vma, addr);
3907 		if (pfn == -1)
3908 			continue;
3909 
3910 		if (!pte_young(pte[i])) {
3911 			walk->mm_stats[MM_LEAF_OLD]++;
3912 			continue;
3913 		}
3914 
3915 		folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap);
3916 		if (!folio)
3917 			continue;
3918 
3919 		if (!ptep_test_and_clear_young(args->vma, addr, pte + i))
3920 			VM_WARN_ON_ONCE(true);
3921 
3922 		young++;
3923 		walk->mm_stats[MM_LEAF_YOUNG]++;
3924 
3925 		if (pte_dirty(pte[i]) && !folio_test_dirty(folio) &&
3926 		    !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
3927 		      !folio_test_swapcache(folio)))
3928 			folio_mark_dirty(folio);
3929 
3930 		old_gen = folio_update_gen(folio, new_gen);
3931 		if (old_gen >= 0 && old_gen != new_gen)
3932 			update_batch_size(walk, folio, old_gen, new_gen);
3933 	}
3934 
3935 	if (i < PTRS_PER_PTE && get_next_vma(PMD_MASK, PAGE_SIZE, args, &start, &end))
3936 		goto restart;
3937 
3938 	pte_unmap(pte);
3939 
3940 	arch_leave_lazy_mmu_mode();
3941 	spin_unlock(ptl);
3942 
3943 	return suitable_to_scan(total, young);
3944 }
3945 
3946 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG)
walk_pmd_range_locked(pud_t * pud,unsigned long next,struct vm_area_struct * vma,struct mm_walk * args,unsigned long * bitmap,unsigned long * start)3947 static void walk_pmd_range_locked(pud_t *pud, unsigned long next, struct vm_area_struct *vma,
3948 				  struct mm_walk *args, unsigned long *bitmap, unsigned long *start)
3949 {
3950 	int i;
3951 	pmd_t *pmd;
3952 	spinlock_t *ptl;
3953 	struct lru_gen_mm_walk *walk = args->private;
3954 	struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec);
3955 	struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
3956 	int old_gen, new_gen = lru_gen_from_seq(walk->max_seq);
3957 
3958 	VM_WARN_ON_ONCE(pud_leaf(*pud));
3959 
3960 	/* try to batch at most 1+MIN_LRU_BATCH+1 entries */
3961 	if (*start == -1) {
3962 		*start = next;
3963 		return;
3964 	}
3965 
3966 	i = next == -1 ? 0 : pmd_index(next) - pmd_index(*start);
3967 	if (i && i <= MIN_LRU_BATCH) {
3968 		__set_bit(i - 1, bitmap);
3969 		return;
3970 	}
3971 
3972 	pmd = pmd_offset(pud, *start);
3973 
3974 	ptl = pmd_lockptr(args->mm, pmd);
3975 	if (!spin_trylock(ptl))
3976 		goto done;
3977 
3978 	arch_enter_lazy_mmu_mode();
3979 
3980 	do {
3981 		unsigned long pfn;
3982 		struct folio *folio;
3983 		unsigned long addr = i ? (*start & PMD_MASK) + i * PMD_SIZE : *start;
3984 
3985 		pfn = get_pmd_pfn(pmd[i], vma, addr);
3986 		if (pfn == -1)
3987 			goto next;
3988 
3989 		if (!pmd_trans_huge(pmd[i])) {
3990 			if (arch_has_hw_nonleaf_pmd_young() &&
3991 			    get_cap(LRU_GEN_NONLEAF_YOUNG))
3992 				pmdp_test_and_clear_young(vma, addr, pmd + i);
3993 			goto next;
3994 		}
3995 
3996 		folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap);
3997 		if (!folio)
3998 			goto next;
3999 
4000 		if (!pmdp_test_and_clear_young(vma, addr, pmd + i))
4001 			goto next;
4002 
4003 		walk->mm_stats[MM_LEAF_YOUNG]++;
4004 
4005 		if (pmd_dirty(pmd[i]) && !folio_test_dirty(folio) &&
4006 		    !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
4007 		      !folio_test_swapcache(folio)))
4008 			folio_mark_dirty(folio);
4009 
4010 		old_gen = folio_update_gen(folio, new_gen);
4011 		if (old_gen >= 0 && old_gen != new_gen)
4012 			update_batch_size(walk, folio, old_gen, new_gen);
4013 next:
4014 		i = i > MIN_LRU_BATCH ? 0 : find_next_bit(bitmap, MIN_LRU_BATCH, i) + 1;
4015 	} while (i <= MIN_LRU_BATCH);
4016 
4017 	arch_leave_lazy_mmu_mode();
4018 	spin_unlock(ptl);
4019 done:
4020 	*start = -1;
4021 	bitmap_zero(bitmap, MIN_LRU_BATCH);
4022 }
4023 #else
walk_pmd_range_locked(pud_t * pud,unsigned long next,struct vm_area_struct * vma,struct mm_walk * args,unsigned long * bitmap,unsigned long * start)4024 static void walk_pmd_range_locked(pud_t *pud, unsigned long next, struct vm_area_struct *vma,
4025 				  struct mm_walk *args, unsigned long *bitmap, unsigned long *start)
4026 {
4027 }
4028 #endif
4029 
walk_pmd_range(pud_t * pud,unsigned long start,unsigned long end,struct mm_walk * args)4030 static void walk_pmd_range(pud_t *pud, unsigned long start, unsigned long end,
4031 			   struct mm_walk *args)
4032 {
4033 	int i;
4034 	pmd_t *pmd;
4035 	unsigned long next;
4036 	unsigned long addr;
4037 	struct vm_area_struct *vma;
4038 	unsigned long pos = -1;
4039 	struct lru_gen_mm_walk *walk = args->private;
4040 	unsigned long bitmap[BITS_TO_LONGS(MIN_LRU_BATCH)] = {};
4041 
4042 	VM_WARN_ON_ONCE(pud_leaf(*pud));
4043 
4044 	/*
4045 	 * Finish an entire PMD in two passes: the first only reaches to PTE
4046 	 * tables to avoid taking the PMD lock; the second, if necessary, takes
4047 	 * the PMD lock to clear the accessed bit in PMD entries.
4048 	 */
4049 	pmd = pmd_offset(pud, start & PUD_MASK);
4050 restart:
4051 	/* walk_pte_range() may call get_next_vma() */
4052 	vma = args->vma;
4053 	for (i = pmd_index(start), addr = start; addr != end; i++, addr = next) {
4054 		pmd_t val = pmd_read_atomic(pmd + i);
4055 
4056 		/* for pmd_read_atomic() */
4057 		barrier();
4058 
4059 		next = pmd_addr_end(addr, end);
4060 
4061 		if (!pmd_present(val) || is_huge_zero_pmd(val)) {
4062 			walk->mm_stats[MM_LEAF_TOTAL]++;
4063 			continue;
4064 		}
4065 
4066 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4067 		if (pmd_trans_huge(val)) {
4068 			unsigned long pfn = pmd_pfn(val);
4069 			struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
4070 
4071 			walk->mm_stats[MM_LEAF_TOTAL]++;
4072 
4073 			if (!pmd_young(val)) {
4074 				walk->mm_stats[MM_LEAF_OLD]++;
4075 				continue;
4076 			}
4077 
4078 			/* try to avoid unnecessary memory loads */
4079 			if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat))
4080 				continue;
4081 
4082 			walk_pmd_range_locked(pud, addr, vma, args, bitmap, &pos);
4083 			continue;
4084 		}
4085 #endif
4086 		walk->mm_stats[MM_NONLEAF_TOTAL]++;
4087 
4088 		if (arch_has_hw_nonleaf_pmd_young() &&
4089 		    get_cap(LRU_GEN_NONLEAF_YOUNG)) {
4090 			if (!pmd_young(val))
4091 				continue;
4092 
4093 			walk_pmd_range_locked(pud, addr, vma, args, bitmap, &pos);
4094 		}
4095 
4096 		if (!walk->force_scan && !test_bloom_filter(walk->lruvec, walk->max_seq, pmd + i))
4097 			continue;
4098 
4099 		walk->mm_stats[MM_NONLEAF_FOUND]++;
4100 
4101 		if (!walk_pte_range(&val, addr, next, args))
4102 			continue;
4103 
4104 		walk->mm_stats[MM_NONLEAF_ADDED]++;
4105 
4106 		/* carry over to the next generation */
4107 		update_bloom_filter(walk->lruvec, walk->max_seq + 1, pmd + i);
4108 	}
4109 
4110 	walk_pmd_range_locked(pud, -1, vma, args, bitmap, &pos);
4111 
4112 	if (i < PTRS_PER_PMD && get_next_vma(PUD_MASK, PMD_SIZE, args, &start, &end))
4113 		goto restart;
4114 }
4115 
walk_pud_range(p4d_t * p4d,unsigned long start,unsigned long end,struct mm_walk * args)4116 static int walk_pud_range(p4d_t *p4d, unsigned long start, unsigned long end,
4117 			  struct mm_walk *args)
4118 {
4119 	int i;
4120 	pud_t *pud;
4121 	unsigned long addr;
4122 	unsigned long next;
4123 	struct lru_gen_mm_walk *walk = args->private;
4124 
4125 	VM_WARN_ON_ONCE(p4d_leaf(*p4d));
4126 
4127 	pud = pud_offset(p4d, start & P4D_MASK);
4128 restart:
4129 	for (i = pud_index(start), addr = start; addr != end; i++, addr = next) {
4130 		pud_t val = READ_ONCE(pud[i]);
4131 
4132 		next = pud_addr_end(addr, end);
4133 
4134 		if (!pud_present(val) || WARN_ON_ONCE(pud_leaf(val)))
4135 			continue;
4136 
4137 		walk_pmd_range(&val, addr, next, args);
4138 
4139 		/* a racy check to curtail the waiting time */
4140 		if (wq_has_sleeper(&walk->lruvec->mm_state.wait))
4141 			return 1;
4142 
4143 		if (need_resched() || walk->batched >= MAX_LRU_BATCH) {
4144 			end = (addr | ~PUD_MASK) + 1;
4145 			goto done;
4146 		}
4147 	}
4148 
4149 	if (i < PTRS_PER_PUD && get_next_vma(P4D_MASK, PUD_SIZE, args, &start, &end))
4150 		goto restart;
4151 
4152 	end = round_up(end, P4D_SIZE);
4153 done:
4154 	if (!end || !args->vma)
4155 		return 1;
4156 
4157 	walk->next_addr = max(end, args->vma->vm_start);
4158 
4159 	return -EAGAIN;
4160 }
4161 
walk_mm(struct lruvec * lruvec,struct mm_struct * mm,struct lru_gen_mm_walk * walk)4162 static void walk_mm(struct lruvec *lruvec, struct mm_struct *mm, struct lru_gen_mm_walk *walk)
4163 {
4164 	static const struct mm_walk_ops mm_walk_ops = {
4165 		.test_walk = should_skip_vma,
4166 		.p4d_entry = walk_pud_range,
4167 	};
4168 
4169 	int err;
4170 	struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4171 
4172 	walk->next_addr = FIRST_USER_ADDRESS;
4173 
4174 	do {
4175 		err = -EBUSY;
4176 
4177 		/* folio_update_gen() requires stable folio_memcg() */
4178 		if (!mem_cgroup_trylock_pages(memcg))
4179 			break;
4180 
4181 		/* the caller might be holding the lock for write */
4182 		if (mmap_read_trylock(mm)) {
4183 			err = walk_page_range(mm, walk->next_addr, ULONG_MAX, &mm_walk_ops, walk);
4184 
4185 			mmap_read_unlock(mm);
4186 		}
4187 
4188 		mem_cgroup_unlock_pages();
4189 
4190 		if (walk->batched) {
4191 			spin_lock_irq(&lruvec->lru_lock);
4192 			reset_batch_size(lruvec, walk);
4193 			spin_unlock_irq(&lruvec->lru_lock);
4194 		}
4195 
4196 		cond_resched();
4197 	} while (err == -EAGAIN);
4198 }
4199 
set_mm_walk(struct pglist_data * pgdat)4200 static struct lru_gen_mm_walk *set_mm_walk(struct pglist_data *pgdat)
4201 {
4202 	struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
4203 
4204 	if (pgdat && current_is_kswapd()) {
4205 		VM_WARN_ON_ONCE(walk);
4206 
4207 		walk = &pgdat->mm_walk;
4208 	} else if (!pgdat && !walk) {
4209 		VM_WARN_ON_ONCE(current_is_kswapd());
4210 
4211 		walk = kzalloc(sizeof(*walk), __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN);
4212 	}
4213 
4214 	current->reclaim_state->mm_walk = walk;
4215 
4216 	return walk;
4217 }
4218 
clear_mm_walk(void)4219 static void clear_mm_walk(void)
4220 {
4221 	struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk;
4222 
4223 	VM_WARN_ON_ONCE(walk && memchr_inv(walk->nr_pages, 0, sizeof(walk->nr_pages)));
4224 	VM_WARN_ON_ONCE(walk && memchr_inv(walk->mm_stats, 0, sizeof(walk->mm_stats)));
4225 
4226 	current->reclaim_state->mm_walk = NULL;
4227 
4228 	if (!current_is_kswapd())
4229 		kfree(walk);
4230 }
4231 
inc_min_seq(struct lruvec * lruvec,int type,bool can_swap)4232 static bool inc_min_seq(struct lruvec *lruvec, int type, bool can_swap)
4233 {
4234 	int zone;
4235 	int remaining = MAX_LRU_BATCH;
4236 	struct lru_gen_struct *lrugen = &lruvec->lrugen;
4237 	int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]);
4238 
4239 	if (type == LRU_GEN_ANON && !can_swap)
4240 		goto done;
4241 
4242 	/* prevent cold/hot inversion if force_scan is true */
4243 	for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4244 		struct list_head *head = &lrugen->lists[old_gen][type][zone];
4245 
4246 		while (!list_empty(head)) {
4247 			struct folio *folio = lru_to_folio(head);
4248 
4249 			VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
4250 			VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
4251 			VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
4252 			VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
4253 
4254 			new_gen = folio_inc_gen(lruvec, folio, false);
4255 			list_move_tail(&folio->lru, &lrugen->lists[new_gen][type][zone]);
4256 
4257 			if (!--remaining)
4258 				return false;
4259 		}
4260 	}
4261 done:
4262 	reset_ctrl_pos(lruvec, type, true);
4263 	WRITE_ONCE(lrugen->min_seq[type], lrugen->min_seq[type] + 1);
4264 
4265 	return true;
4266 }
4267 
try_to_inc_min_seq(struct lruvec * lruvec,bool can_swap)4268 static bool try_to_inc_min_seq(struct lruvec *lruvec, bool can_swap)
4269 {
4270 	int gen, type, zone;
4271 	bool success = false;
4272 	struct lru_gen_struct *lrugen = &lruvec->lrugen;
4273 	DEFINE_MIN_SEQ(lruvec);
4274 
4275 	VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
4276 
4277 	/* find the oldest populated generation */
4278 	for (type = !can_swap; type < ANON_AND_FILE; type++) {
4279 		while (min_seq[type] + MIN_NR_GENS <= lrugen->max_seq) {
4280 			gen = lru_gen_from_seq(min_seq[type]);
4281 
4282 			for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4283 				if (!list_empty(&lrugen->lists[gen][type][zone]))
4284 					goto next;
4285 			}
4286 
4287 			min_seq[type]++;
4288 		}
4289 next:
4290 		;
4291 	}
4292 
4293 	/* see the comment on lru_gen_struct */
4294 	if (can_swap) {
4295 		min_seq[LRU_GEN_ANON] = min(min_seq[LRU_GEN_ANON], min_seq[LRU_GEN_FILE]);
4296 		min_seq[LRU_GEN_FILE] = max(min_seq[LRU_GEN_ANON], lrugen->min_seq[LRU_GEN_FILE]);
4297 	}
4298 
4299 	for (type = !can_swap; type < ANON_AND_FILE; type++) {
4300 		if (min_seq[type] == lrugen->min_seq[type])
4301 			continue;
4302 
4303 		reset_ctrl_pos(lruvec, type, true);
4304 		WRITE_ONCE(lrugen->min_seq[type], min_seq[type]);
4305 		success = true;
4306 	}
4307 
4308 	return success;
4309 }
4310 
inc_max_seq(struct lruvec * lruvec,bool can_swap,bool force_scan)4311 static void inc_max_seq(struct lruvec *lruvec, bool can_swap, bool force_scan)
4312 {
4313 	int prev, next;
4314 	int type, zone;
4315 	struct lru_gen_struct *lrugen = &lruvec->lrugen;
4316 
4317 	spin_lock_irq(&lruvec->lru_lock);
4318 
4319 	VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
4320 
4321 	for (type = ANON_AND_FILE - 1; type >= 0; type--) {
4322 		if (get_nr_gens(lruvec, type) != MAX_NR_GENS)
4323 			continue;
4324 
4325 		VM_WARN_ON_ONCE(!force_scan && (type == LRU_GEN_FILE || can_swap));
4326 
4327 		while (!inc_min_seq(lruvec, type, can_swap)) {
4328 			spin_unlock_irq(&lruvec->lru_lock);
4329 			cond_resched();
4330 			spin_lock_irq(&lruvec->lru_lock);
4331 		}
4332 	}
4333 
4334 	/*
4335 	 * Update the active/inactive LRU sizes for compatibility. Both sides of
4336 	 * the current max_seq need to be covered, since max_seq+1 can overlap
4337 	 * with min_seq[LRU_GEN_ANON] if swapping is constrained. And if they do
4338 	 * overlap, cold/hot inversion happens.
4339 	 */
4340 	prev = lru_gen_from_seq(lrugen->max_seq - 1);
4341 	next = lru_gen_from_seq(lrugen->max_seq + 1);
4342 
4343 	for (type = 0; type < ANON_AND_FILE; type++) {
4344 		for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4345 			enum lru_list lru = type * LRU_INACTIVE_FILE;
4346 			long delta = lrugen->nr_pages[prev][type][zone] -
4347 				     lrugen->nr_pages[next][type][zone];
4348 
4349 			if (!delta)
4350 				continue;
4351 
4352 			__update_lru_size(lruvec, lru, zone, delta);
4353 			__update_lru_size(lruvec, lru + LRU_ACTIVE, zone, -delta);
4354 		}
4355 	}
4356 
4357 	for (type = 0; type < ANON_AND_FILE; type++)
4358 		reset_ctrl_pos(lruvec, type, false);
4359 
4360 	WRITE_ONCE(lrugen->timestamps[next], jiffies);
4361 	/* make sure preceding modifications appear */
4362 	smp_store_release(&lrugen->max_seq, lrugen->max_seq + 1);
4363 
4364 	spin_unlock_irq(&lruvec->lru_lock);
4365 }
4366 
try_to_inc_max_seq(struct lruvec * lruvec,unsigned long max_seq,struct scan_control * sc,bool can_swap,bool force_scan)4367 static bool try_to_inc_max_seq(struct lruvec *lruvec, unsigned long max_seq,
4368 			       struct scan_control *sc, bool can_swap, bool force_scan)
4369 {
4370 	bool success;
4371 	struct lru_gen_mm_walk *walk;
4372 	struct mm_struct *mm = NULL;
4373 	struct lru_gen_struct *lrugen = &lruvec->lrugen;
4374 
4375 	VM_WARN_ON_ONCE(max_seq > READ_ONCE(lrugen->max_seq));
4376 
4377 	/* see the comment in iterate_mm_list() */
4378 	if (max_seq <= READ_ONCE(lruvec->mm_state.seq)) {
4379 		success = false;
4380 		goto done;
4381 	}
4382 
4383 	/*
4384 	 * If the hardware doesn't automatically set the accessed bit, fallback
4385 	 * to lru_gen_look_around(), which only clears the accessed bit in a
4386 	 * handful of PTEs. Spreading the work out over a period of time usually
4387 	 * is less efficient, but it avoids bursty page faults.
4388 	 */
4389 	if (!force_scan && !(arch_has_hw_pte_young() && get_cap(LRU_GEN_MM_WALK))) {
4390 		success = iterate_mm_list_nowalk(lruvec, max_seq);
4391 		goto done;
4392 	}
4393 
4394 	walk = set_mm_walk(NULL);
4395 	if (!walk) {
4396 		success = iterate_mm_list_nowalk(lruvec, max_seq);
4397 		goto done;
4398 	}
4399 
4400 	walk->lruvec = lruvec;
4401 	walk->max_seq = max_seq;
4402 	walk->can_swap = can_swap;
4403 	walk->force_scan = force_scan;
4404 
4405 	do {
4406 		success = iterate_mm_list(lruvec, walk, &mm);
4407 		if (mm)
4408 			walk_mm(lruvec, mm, walk);
4409 
4410 		cond_resched();
4411 	} while (mm);
4412 done:
4413 	if (!success) {
4414 		if (sc->priority <= DEF_PRIORITY - 2)
4415 			wait_event_killable(lruvec->mm_state.wait,
4416 					    max_seq < READ_ONCE(lrugen->max_seq));
4417 
4418 		return max_seq < READ_ONCE(lrugen->max_seq);
4419 	}
4420 
4421 	VM_WARN_ON_ONCE(max_seq != READ_ONCE(lrugen->max_seq));
4422 
4423 	inc_max_seq(lruvec, can_swap, force_scan);
4424 	/* either this sees any waiters or they will see updated max_seq */
4425 	if (wq_has_sleeper(&lruvec->mm_state.wait))
4426 		wake_up_all(&lruvec->mm_state.wait);
4427 
4428 	return true;
4429 }
4430 
should_run_aging(struct lruvec * lruvec,unsigned long max_seq,unsigned long * min_seq,struct scan_control * sc,bool can_swap,unsigned long * nr_to_scan)4431 static bool should_run_aging(struct lruvec *lruvec, unsigned long max_seq, unsigned long *min_seq,
4432 			     struct scan_control *sc, bool can_swap, unsigned long *nr_to_scan)
4433 {
4434 	int gen, type, zone;
4435 	unsigned long old = 0;
4436 	unsigned long young = 0;
4437 	unsigned long total = 0;
4438 	struct lru_gen_struct *lrugen = &lruvec->lrugen;
4439 	struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4440 
4441 	for (type = !can_swap; type < ANON_AND_FILE; type++) {
4442 		unsigned long seq;
4443 
4444 		for (seq = min_seq[type]; seq <= max_seq; seq++) {
4445 			unsigned long size = 0;
4446 
4447 			gen = lru_gen_from_seq(seq);
4448 
4449 			for (zone = 0; zone < MAX_NR_ZONES; zone++)
4450 				size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
4451 
4452 			total += size;
4453 			if (seq == max_seq)
4454 				young += size;
4455 			else if (seq + MIN_NR_GENS == max_seq)
4456 				old += size;
4457 		}
4458 	}
4459 
4460 	/* try to scrape all its memory if this memcg was deleted */
4461 	*nr_to_scan = mem_cgroup_online(memcg) ? (total >> sc->priority) : total;
4462 
4463 	/*
4464 	 * The aging tries to be lazy to reduce the overhead, while the eviction
4465 	 * stalls when the number of generations reaches MIN_NR_GENS. Hence, the
4466 	 * ideal number of generations is MIN_NR_GENS+1.
4467 	 */
4468 	if (min_seq[!can_swap] + MIN_NR_GENS > max_seq)
4469 		return true;
4470 	if (min_seq[!can_swap] + MIN_NR_GENS < max_seq)
4471 		return false;
4472 
4473 	/*
4474 	 * It's also ideal to spread pages out evenly, i.e., 1/(MIN_NR_GENS+1)
4475 	 * of the total number of pages for each generation. A reasonable range
4476 	 * for this average portion is [1/MIN_NR_GENS, 1/(MIN_NR_GENS+2)]. The
4477 	 * aging cares about the upper bound of hot pages, while the eviction
4478 	 * cares about the lower bound of cold pages.
4479 	 */
4480 	if (young * MIN_NR_GENS > total)
4481 		return true;
4482 	if (old * (MIN_NR_GENS + 2) < total)
4483 		return true;
4484 
4485 	return false;
4486 }
4487 
age_lruvec(struct lruvec * lruvec,struct scan_control * sc,unsigned long min_ttl)4488 static bool age_lruvec(struct lruvec *lruvec, struct scan_control *sc, unsigned long min_ttl)
4489 {
4490 	bool need_aging;
4491 	unsigned long nr_to_scan;
4492 	int swappiness = get_swappiness(lruvec, sc);
4493 	struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4494 	DEFINE_MAX_SEQ(lruvec);
4495 	DEFINE_MIN_SEQ(lruvec);
4496 
4497 	VM_WARN_ON_ONCE(sc->memcg_low_reclaim);
4498 
4499 	mem_cgroup_calculate_protection(NULL, memcg);
4500 
4501 	if (mem_cgroup_below_min(memcg))
4502 		return false;
4503 
4504 	need_aging = should_run_aging(lruvec, max_seq, min_seq, sc, swappiness, &nr_to_scan);
4505 
4506 	if (min_ttl) {
4507 		int gen = lru_gen_from_seq(min_seq[LRU_GEN_FILE]);
4508 		unsigned long birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
4509 
4510 		if (time_is_after_jiffies(birth + min_ttl))
4511 			return false;
4512 
4513 		/* the size is likely too small to be helpful */
4514 		if (!nr_to_scan && sc->priority != DEF_PRIORITY)
4515 			return false;
4516 	}
4517 
4518 	if (need_aging)
4519 		try_to_inc_max_seq(lruvec, max_seq, sc, swappiness, false);
4520 
4521 	return true;
4522 }
4523 
4524 /* to protect the working set of the last N jiffies */
4525 static unsigned long lru_gen_min_ttl __read_mostly;
4526 
lru_gen_age_node(struct pglist_data * pgdat,struct scan_control * sc)4527 static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
4528 {
4529 	struct mem_cgroup *memcg;
4530 	bool success = false;
4531 	unsigned long min_ttl = READ_ONCE(lru_gen_min_ttl);
4532 
4533 	VM_WARN_ON_ONCE(!current_is_kswapd());
4534 
4535 	sc->last_reclaimed = sc->nr_reclaimed;
4536 
4537 	/*
4538 	 * To reduce the chance of going into the aging path, which can be
4539 	 * costly, optimistically skip it if the flag below was cleared in the
4540 	 * eviction path. This improves the overall performance when multiple
4541 	 * memcgs are available.
4542 	 */
4543 	if (!sc->memcgs_need_aging) {
4544 		sc->memcgs_need_aging = true;
4545 		return;
4546 	}
4547 
4548 	set_mm_walk(pgdat);
4549 
4550 	memcg = mem_cgroup_iter(NULL, NULL, NULL);
4551 	do {
4552 		struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
4553 
4554 		if (age_lruvec(lruvec, sc, min_ttl))
4555 			success = true;
4556 
4557 		cond_resched();
4558 	} while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
4559 
4560 	clear_mm_walk();
4561 
4562 	/* check the order to exclude compaction-induced reclaim */
4563 	if (success || !min_ttl || sc->order)
4564 		return;
4565 
4566 	/*
4567 	 * The main goal is to OOM kill if every generation from all memcgs is
4568 	 * younger than min_ttl. However, another possibility is all memcgs are
4569 	 * either below min or empty.
4570 	 */
4571 	if (mutex_trylock(&oom_lock)) {
4572 		struct oom_control oc = {
4573 			.gfp_mask = sc->gfp_mask,
4574 		};
4575 
4576 		out_of_memory(&oc);
4577 
4578 		mutex_unlock(&oom_lock);
4579 	}
4580 }
4581 
4582 /*
4583  * This function exploits spatial locality when shrink_folio_list() walks the
4584  * rmap. It scans the adjacent PTEs of a young PTE and promotes hot pages. If
4585  * the scan was done cacheline efficiently, it adds the PMD entry pointing to
4586  * the PTE table to the Bloom filter. This forms a feedback loop between the
4587  * eviction and the aging.
4588  */
lru_gen_look_around(struct page_vma_mapped_walk * pvmw)4589 void lru_gen_look_around(struct page_vma_mapped_walk *pvmw)
4590 {
4591 	int i;
4592 	pte_t *pte;
4593 	unsigned long start;
4594 	unsigned long end;
4595 	unsigned long addr;
4596 	struct lru_gen_mm_walk *walk;
4597 	int young = 0;
4598 	unsigned long bitmap[BITS_TO_LONGS(MIN_LRU_BATCH)] = {};
4599 	struct folio *folio = pfn_folio(pvmw->pfn);
4600 	struct mem_cgroup *memcg = folio_memcg(folio);
4601 	struct pglist_data *pgdat = folio_pgdat(folio);
4602 	struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
4603 	DEFINE_MAX_SEQ(lruvec);
4604 	int old_gen, new_gen = lru_gen_from_seq(max_seq);
4605 
4606 	lockdep_assert_held(pvmw->ptl);
4607 	VM_WARN_ON_ONCE_FOLIO(folio_test_lru(folio), folio);
4608 
4609 	if (spin_is_contended(pvmw->ptl))
4610 		return;
4611 
4612 	/* avoid taking the LRU lock under the PTL when possible */
4613 	walk = current->reclaim_state ? current->reclaim_state->mm_walk : NULL;
4614 
4615 	start = max(pvmw->address & PMD_MASK, pvmw->vma->vm_start);
4616 	end = min(pvmw->address | ~PMD_MASK, pvmw->vma->vm_end - 1) + 1;
4617 
4618 	if (end - start > MIN_LRU_BATCH * PAGE_SIZE) {
4619 		if (pvmw->address - start < MIN_LRU_BATCH * PAGE_SIZE / 2)
4620 			end = start + MIN_LRU_BATCH * PAGE_SIZE;
4621 		else if (end - pvmw->address < MIN_LRU_BATCH * PAGE_SIZE / 2)
4622 			start = end - MIN_LRU_BATCH * PAGE_SIZE;
4623 		else {
4624 			start = pvmw->address - MIN_LRU_BATCH * PAGE_SIZE / 2;
4625 			end = pvmw->address + MIN_LRU_BATCH * PAGE_SIZE / 2;
4626 		}
4627 	}
4628 
4629 	pte = pvmw->pte - (pvmw->address - start) / PAGE_SIZE;
4630 
4631 	rcu_read_lock();
4632 	arch_enter_lazy_mmu_mode();
4633 
4634 	for (i = 0, addr = start; addr != end; i++, addr += PAGE_SIZE) {
4635 		unsigned long pfn;
4636 
4637 		pfn = get_pte_pfn(pte[i], pvmw->vma, addr);
4638 		if (pfn == -1)
4639 			continue;
4640 
4641 		if (!pte_young(pte[i]))
4642 			continue;
4643 
4644 		folio = get_pfn_folio(pfn, memcg, pgdat, !walk || walk->can_swap);
4645 		if (!folio)
4646 			continue;
4647 
4648 		if (!ptep_test_and_clear_young(pvmw->vma, addr, pte + i))
4649 			VM_WARN_ON_ONCE(true);
4650 
4651 		young++;
4652 
4653 		if (pte_dirty(pte[i]) && !folio_test_dirty(folio) &&
4654 		    !(folio_test_anon(folio) && folio_test_swapbacked(folio) &&
4655 		      !folio_test_swapcache(folio)))
4656 			folio_mark_dirty(folio);
4657 
4658 		old_gen = folio_lru_gen(folio);
4659 		if (old_gen < 0)
4660 			folio_set_referenced(folio);
4661 		else if (old_gen != new_gen)
4662 			__set_bit(i, bitmap);
4663 	}
4664 
4665 	arch_leave_lazy_mmu_mode();
4666 	rcu_read_unlock();
4667 
4668 	/* feedback from rmap walkers to page table walkers */
4669 	if (suitable_to_scan(i, young))
4670 		update_bloom_filter(lruvec, max_seq, pvmw->pmd);
4671 
4672 	if (!walk && bitmap_weight(bitmap, MIN_LRU_BATCH) < PAGEVEC_SIZE) {
4673 		for_each_set_bit(i, bitmap, MIN_LRU_BATCH) {
4674 			folio = pfn_folio(pte_pfn(pte[i]));
4675 			folio_activate(folio);
4676 		}
4677 		return;
4678 	}
4679 
4680 	/* folio_update_gen() requires stable folio_memcg() */
4681 	if (!mem_cgroup_trylock_pages(memcg))
4682 		return;
4683 
4684 	if (!walk) {
4685 		spin_lock_irq(&lruvec->lru_lock);
4686 		new_gen = lru_gen_from_seq(lruvec->lrugen.max_seq);
4687 	}
4688 
4689 	for_each_set_bit(i, bitmap, MIN_LRU_BATCH) {
4690 		folio = pfn_folio(pte_pfn(pte[i]));
4691 		if (folio_memcg_rcu(folio) != memcg)
4692 			continue;
4693 
4694 		old_gen = folio_update_gen(folio, new_gen);
4695 		if (old_gen < 0 || old_gen == new_gen)
4696 			continue;
4697 
4698 		if (walk)
4699 			update_batch_size(walk, folio, old_gen, new_gen);
4700 		else
4701 			lru_gen_update_size(lruvec, folio, old_gen, new_gen);
4702 	}
4703 
4704 	if (!walk)
4705 		spin_unlock_irq(&lruvec->lru_lock);
4706 
4707 	mem_cgroup_unlock_pages();
4708 }
4709 
4710 /******************************************************************************
4711  *                          the eviction
4712  ******************************************************************************/
4713 
sort_folio(struct lruvec * lruvec,struct folio * folio,int tier_idx)4714 static bool sort_folio(struct lruvec *lruvec, struct folio *folio, int tier_idx)
4715 {
4716 	bool success;
4717 	int gen = folio_lru_gen(folio);
4718 	int type = folio_is_file_lru(folio);
4719 	int zone = folio_zonenum(folio);
4720 	int delta = folio_nr_pages(folio);
4721 	int refs = folio_lru_refs(folio);
4722 	int tier = lru_tier_from_refs(refs);
4723 	struct lru_gen_struct *lrugen = &lruvec->lrugen;
4724 
4725 	VM_WARN_ON_ONCE_FOLIO(gen >= MAX_NR_GENS, folio);
4726 
4727 	/* unevictable */
4728 	if (!folio_evictable(folio)) {
4729 		success = lru_gen_del_folio(lruvec, folio, true);
4730 		VM_WARN_ON_ONCE_FOLIO(!success, folio);
4731 		folio_set_unevictable(folio);
4732 		lruvec_add_folio(lruvec, folio);
4733 		__count_vm_events(UNEVICTABLE_PGCULLED, delta);
4734 		return true;
4735 	}
4736 
4737 	/* dirty lazyfree */
4738 	if (type == LRU_GEN_FILE && folio_test_anon(folio) && folio_test_dirty(folio)) {
4739 		success = lru_gen_del_folio(lruvec, folio, true);
4740 		VM_WARN_ON_ONCE_FOLIO(!success, folio);
4741 		folio_set_swapbacked(folio);
4742 		lruvec_add_folio_tail(lruvec, folio);
4743 		return true;
4744 	}
4745 
4746 	/* promoted */
4747 	if (gen != lru_gen_from_seq(lrugen->min_seq[type])) {
4748 		list_move(&folio->lru, &lrugen->lists[gen][type][zone]);
4749 		return true;
4750 	}
4751 
4752 	/* protected */
4753 	if (tier > tier_idx) {
4754 		int hist = lru_hist_from_seq(lrugen->min_seq[type]);
4755 
4756 		gen = folio_inc_gen(lruvec, folio, false);
4757 		list_move_tail(&folio->lru, &lrugen->lists[gen][type][zone]);
4758 
4759 		WRITE_ONCE(lrugen->protected[hist][type][tier - 1],
4760 			   lrugen->protected[hist][type][tier - 1] + delta);
4761 		__mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + type, delta);
4762 		return true;
4763 	}
4764 
4765 	/* waiting for writeback */
4766 	if (folio_test_locked(folio) || folio_test_writeback(folio) ||
4767 	    (type == LRU_GEN_FILE && folio_test_dirty(folio))) {
4768 		gen = folio_inc_gen(lruvec, folio, true);
4769 		list_move(&folio->lru, &lrugen->lists[gen][type][zone]);
4770 		return true;
4771 	}
4772 
4773 	return false;
4774 }
4775 
isolate_folio(struct lruvec * lruvec,struct folio * folio,struct scan_control * sc)4776 static bool isolate_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc)
4777 {
4778 	bool success;
4779 
4780 	/* unmapping inhibited */
4781 	if (!sc->may_unmap && folio_mapped(folio))
4782 		return false;
4783 
4784 	/* swapping inhibited */
4785 	if (!(sc->may_writepage && (sc->gfp_mask & __GFP_IO)) &&
4786 	    (folio_test_dirty(folio) ||
4787 	     (folio_test_anon(folio) && !folio_test_swapcache(folio))))
4788 		return false;
4789 
4790 	/* raced with release_pages() */
4791 	if (!folio_try_get(folio))
4792 		return false;
4793 
4794 	/* raced with another isolation */
4795 	if (!folio_test_clear_lru(folio)) {
4796 		folio_put(folio);
4797 		return false;
4798 	}
4799 
4800 	/* see the comment on MAX_NR_TIERS */
4801 	if (!folio_test_referenced(folio))
4802 		set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS, 0);
4803 
4804 	/* for shrink_folio_list() */
4805 	folio_clear_reclaim(folio);
4806 	folio_clear_referenced(folio);
4807 
4808 	success = lru_gen_del_folio(lruvec, folio, true);
4809 	VM_WARN_ON_ONCE_FOLIO(!success, folio);
4810 
4811 	return true;
4812 }
4813 
scan_folios(struct lruvec * lruvec,struct scan_control * sc,int type,int tier,struct list_head * list)4814 static int scan_folios(struct lruvec *lruvec, struct scan_control *sc,
4815 		       int type, int tier, struct list_head *list)
4816 {
4817 	int gen, zone;
4818 	enum vm_event_item item;
4819 	int sorted = 0;
4820 	int scanned = 0;
4821 	int isolated = 0;
4822 	int remaining = MAX_LRU_BATCH;
4823 	struct lru_gen_struct *lrugen = &lruvec->lrugen;
4824 	struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4825 
4826 	VM_WARN_ON_ONCE(!list_empty(list));
4827 
4828 	if (get_nr_gens(lruvec, type) == MIN_NR_GENS)
4829 		return 0;
4830 
4831 	gen = lru_gen_from_seq(lrugen->min_seq[type]);
4832 
4833 	for (zone = sc->reclaim_idx; zone >= 0; zone--) {
4834 		LIST_HEAD(moved);
4835 		int skipped = 0;
4836 		struct list_head *head = &lrugen->lists[gen][type][zone];
4837 
4838 		while (!list_empty(head)) {
4839 			struct folio *folio = lru_to_folio(head);
4840 			int delta = folio_nr_pages(folio);
4841 
4842 			VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
4843 			VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
4844 			VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
4845 			VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
4846 
4847 			scanned += delta;
4848 
4849 			if (sort_folio(lruvec, folio, tier))
4850 				sorted += delta;
4851 			else if (isolate_folio(lruvec, folio, sc)) {
4852 				list_add(&folio->lru, list);
4853 				isolated += delta;
4854 			} else {
4855 				list_move(&folio->lru, &moved);
4856 				skipped += delta;
4857 			}
4858 
4859 			if (!--remaining || max(isolated, skipped) >= MIN_LRU_BATCH)
4860 				break;
4861 		}
4862 
4863 		if (skipped) {
4864 			list_splice(&moved, head);
4865 			__count_zid_vm_events(PGSCAN_SKIP, zone, skipped);
4866 		}
4867 
4868 		if (!remaining || isolated >= MIN_LRU_BATCH)
4869 			break;
4870 	}
4871 
4872 	item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
4873 	if (!cgroup_reclaim(sc)) {
4874 		__count_vm_events(item, isolated);
4875 		__count_vm_events(PGREFILL, sorted);
4876 	}
4877 	__count_memcg_events(memcg, item, isolated);
4878 	__count_memcg_events(memcg, PGREFILL, sorted);
4879 	__count_vm_events(PGSCAN_ANON + type, isolated);
4880 
4881 	/*
4882 	 * There might not be eligible pages due to reclaim_idx, may_unmap and
4883 	 * may_writepage. Check the remaining to prevent livelock if it's not
4884 	 * making progress.
4885 	 */
4886 	return isolated || !remaining ? scanned : 0;
4887 }
4888 
get_tier_idx(struct lruvec * lruvec,int type)4889 static int get_tier_idx(struct lruvec *lruvec, int type)
4890 {
4891 	int tier;
4892 	struct ctrl_pos sp, pv;
4893 
4894 	/*
4895 	 * To leave a margin for fluctuations, use a larger gain factor (1:2).
4896 	 * This value is chosen because any other tier would have at least twice
4897 	 * as many refaults as the first tier.
4898 	 */
4899 	read_ctrl_pos(lruvec, type, 0, 1, &sp);
4900 	for (tier = 1; tier < MAX_NR_TIERS; tier++) {
4901 		read_ctrl_pos(lruvec, type, tier, 2, &pv);
4902 		if (!positive_ctrl_err(&sp, &pv))
4903 			break;
4904 	}
4905 
4906 	return tier - 1;
4907 }
4908 
get_type_to_scan(struct lruvec * lruvec,int swappiness,int * tier_idx)4909 static int get_type_to_scan(struct lruvec *lruvec, int swappiness, int *tier_idx)
4910 {
4911 	int type, tier;
4912 	struct ctrl_pos sp, pv;
4913 	int gain[ANON_AND_FILE] = { swappiness, 200 - swappiness };
4914 
4915 	/*
4916 	 * Compare the first tier of anon with that of file to determine which
4917 	 * type to scan. Also need to compare other tiers of the selected type
4918 	 * with the first tier of the other type to determine the last tier (of
4919 	 * the selected type) to evict.
4920 	 */
4921 	read_ctrl_pos(lruvec, LRU_GEN_ANON, 0, gain[LRU_GEN_ANON], &sp);
4922 	read_ctrl_pos(lruvec, LRU_GEN_FILE, 0, gain[LRU_GEN_FILE], &pv);
4923 	type = positive_ctrl_err(&sp, &pv);
4924 
4925 	read_ctrl_pos(lruvec, !type, 0, gain[!type], &sp);
4926 	for (tier = 1; tier < MAX_NR_TIERS; tier++) {
4927 		read_ctrl_pos(lruvec, type, tier, gain[type], &pv);
4928 		if (!positive_ctrl_err(&sp, &pv))
4929 			break;
4930 	}
4931 
4932 	*tier_idx = tier - 1;
4933 
4934 	return type;
4935 }
4936 
isolate_folios(struct lruvec * lruvec,struct scan_control * sc,int swappiness,int * type_scanned,struct list_head * list)4937 static int isolate_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness,
4938 			  int *type_scanned, struct list_head *list)
4939 {
4940 	int i;
4941 	int type;
4942 	int scanned;
4943 	int tier = -1;
4944 	DEFINE_MIN_SEQ(lruvec);
4945 
4946 	/*
4947 	 * Try to make the obvious choice first. When anon and file are both
4948 	 * available from the same generation, interpret swappiness 1 as file
4949 	 * first and 200 as anon first.
4950 	 */
4951 	if (!swappiness)
4952 		type = LRU_GEN_FILE;
4953 	else if (min_seq[LRU_GEN_ANON] < min_seq[LRU_GEN_FILE])
4954 		type = LRU_GEN_ANON;
4955 	else if (swappiness == 1)
4956 		type = LRU_GEN_FILE;
4957 	else if (swappiness == 200)
4958 		type = LRU_GEN_ANON;
4959 	else
4960 		type = get_type_to_scan(lruvec, swappiness, &tier);
4961 
4962 	for (i = !swappiness; i < ANON_AND_FILE; i++) {
4963 		if (tier < 0)
4964 			tier = get_tier_idx(lruvec, type);
4965 
4966 		scanned = scan_folios(lruvec, sc, type, tier, list);
4967 		if (scanned)
4968 			break;
4969 
4970 		type = !type;
4971 		tier = -1;
4972 	}
4973 
4974 	*type_scanned = type;
4975 
4976 	return scanned;
4977 }
4978 
evict_folios(struct lruvec * lruvec,struct scan_control * sc,int swappiness,bool * need_swapping)4979 static int evict_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness,
4980 			bool *need_swapping)
4981 {
4982 	int type;
4983 	int scanned;
4984 	int reclaimed;
4985 	LIST_HEAD(list);
4986 	LIST_HEAD(clean);
4987 	struct folio *folio;
4988 	struct folio *next;
4989 	enum vm_event_item item;
4990 	struct reclaim_stat stat;
4991 	struct lru_gen_mm_walk *walk;
4992 	bool skip_retry = false;
4993 	struct mem_cgroup *memcg = lruvec_memcg(lruvec);
4994 	struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4995 
4996 	spin_lock_irq(&lruvec->lru_lock);
4997 
4998 	scanned = isolate_folios(lruvec, sc, swappiness, &type, &list);
4999 
5000 	scanned += try_to_inc_min_seq(lruvec, swappiness);
5001 
5002 	if (get_nr_gens(lruvec, !swappiness) == MIN_NR_GENS)
5003 		scanned = 0;
5004 
5005 	spin_unlock_irq(&lruvec->lru_lock);
5006 
5007 	if (list_empty(&list))
5008 		return scanned;
5009 retry:
5010 	reclaimed = shrink_folio_list(&list, pgdat, sc, &stat, false);
5011 	sc->nr_reclaimed += reclaimed;
5012 
5013 	list_for_each_entry_safe_reverse(folio, next, &list, lru) {
5014 		if (!folio_evictable(folio)) {
5015 			list_del(&folio->lru);
5016 			folio_putback_lru(folio);
5017 			continue;
5018 		}
5019 
5020 		if (folio_test_reclaim(folio) &&
5021 		    (folio_test_dirty(folio) || folio_test_writeback(folio))) {
5022 			/* restore LRU_REFS_FLAGS cleared by isolate_folio() */
5023 			if (folio_test_workingset(folio))
5024 				folio_set_referenced(folio);
5025 			continue;
5026 		}
5027 
5028 		if (skip_retry || folio_test_active(folio) || folio_test_referenced(folio) ||
5029 		    folio_mapped(folio) || folio_test_locked(folio) ||
5030 		    folio_test_dirty(folio) || folio_test_writeback(folio)) {
5031 			/* don't add rejected folios to the oldest generation */
5032 			set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS,
5033 				      BIT(PG_active));
5034 			continue;
5035 		}
5036 
5037 		/* retry folios that may have missed folio_rotate_reclaimable() */
5038 		list_move(&folio->lru, &clean);
5039 		sc->nr_scanned -= folio_nr_pages(folio);
5040 	}
5041 
5042 	spin_lock_irq(&lruvec->lru_lock);
5043 
5044 	move_folios_to_lru(lruvec, &list);
5045 
5046 	walk = current->reclaim_state->mm_walk;
5047 	if (walk && walk->batched)
5048 		reset_batch_size(lruvec, walk);
5049 
5050 	item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
5051 	if (!cgroup_reclaim(sc))
5052 		__count_vm_events(item, reclaimed);
5053 	__count_memcg_events(memcg, item, reclaimed);
5054 	__count_vm_events(PGSTEAL_ANON + type, reclaimed);
5055 
5056 	spin_unlock_irq(&lruvec->lru_lock);
5057 
5058 	mem_cgroup_uncharge_list(&list);
5059 	free_unref_page_list(&list);
5060 
5061 	INIT_LIST_HEAD(&list);
5062 	list_splice_init(&clean, &list);
5063 
5064 	if (!list_empty(&list)) {
5065 		skip_retry = true;
5066 		goto retry;
5067 	}
5068 
5069 	if (need_swapping && type == LRU_GEN_ANON)
5070 		*need_swapping = true;
5071 
5072 	return scanned;
5073 }
5074 
5075 /*
5076  * For future optimizations:
5077  * 1. Defer try_to_inc_max_seq() to workqueues to reduce latency for memcg
5078  *    reclaim.
5079  */
get_nr_to_scan(struct lruvec * lruvec,struct scan_control * sc,bool can_swap,bool * need_aging)5080 static unsigned long get_nr_to_scan(struct lruvec *lruvec, struct scan_control *sc,
5081 				    bool can_swap, bool *need_aging)
5082 {
5083 	unsigned long nr_to_scan;
5084 	struct mem_cgroup *memcg = lruvec_memcg(lruvec);
5085 	DEFINE_MAX_SEQ(lruvec);
5086 	DEFINE_MIN_SEQ(lruvec);
5087 
5088 	if (mem_cgroup_below_min(memcg) ||
5089 	    (mem_cgroup_below_low(memcg) && !sc->memcg_low_reclaim))
5090 		return 0;
5091 
5092 	*need_aging = should_run_aging(lruvec, max_seq, min_seq, sc, can_swap, &nr_to_scan);
5093 	if (!*need_aging)
5094 		return nr_to_scan;
5095 
5096 	/* skip the aging path at the default priority */
5097 	if (sc->priority == DEF_PRIORITY)
5098 		goto done;
5099 
5100 	/* leave the work to lru_gen_age_node() */
5101 	if (current_is_kswapd())
5102 		return 0;
5103 
5104 	if (try_to_inc_max_seq(lruvec, max_seq, sc, can_swap, false))
5105 		return nr_to_scan;
5106 done:
5107 	return min_seq[!can_swap] + MIN_NR_GENS <= max_seq ? nr_to_scan : 0;
5108 }
5109 
should_abort_scan(struct lruvec * lruvec,unsigned long seq,struct scan_control * sc,bool need_swapping)5110 static bool should_abort_scan(struct lruvec *lruvec, unsigned long seq,
5111 			      struct scan_control *sc, bool need_swapping)
5112 {
5113 	int i;
5114 	DEFINE_MAX_SEQ(lruvec);
5115 
5116 	if (!current_is_kswapd()) {
5117 		/* age each memcg at most once to ensure fairness */
5118 		if (max_seq - seq > 1)
5119 			return true;
5120 
5121 		/* over-swapping can increase allocation latency */
5122 		if (sc->nr_reclaimed >= sc->nr_to_reclaim && need_swapping)
5123 			return true;
5124 
5125 		/* give this thread a chance to exit and free its memory */
5126 		if (fatal_signal_pending(current)) {
5127 			sc->nr_reclaimed += MIN_LRU_BATCH;
5128 			return true;
5129 		}
5130 
5131 		if (cgroup_reclaim(sc))
5132 			return false;
5133 	} else if (sc->nr_reclaimed - sc->last_reclaimed < sc->nr_to_reclaim)
5134 		return false;
5135 
5136 	/* keep scanning at low priorities to ensure fairness */
5137 	if (sc->priority > DEF_PRIORITY - 2)
5138 		return false;
5139 
5140 	/*
5141 	 * A minimum amount of work was done under global memory pressure. For
5142 	 * kswapd, it may be overshooting. For direct reclaim, the allocation
5143 	 * may succeed if all suitable zones are somewhat safe. In either case,
5144 	 * it's better to stop now, and restart later if necessary.
5145 	 */
5146 	for (i = 0; i <= sc->reclaim_idx; i++) {
5147 		unsigned long wmark;
5148 		struct zone *zone = lruvec_pgdat(lruvec)->node_zones + i;
5149 
5150 		if (!managed_zone(zone))
5151 			continue;
5152 
5153 		wmark = current_is_kswapd() ? high_wmark_pages(zone) : low_wmark_pages(zone);
5154 		if (wmark > zone_page_state(zone, NR_FREE_PAGES))
5155 			return false;
5156 	}
5157 
5158 	sc->nr_reclaimed += MIN_LRU_BATCH;
5159 
5160 	return true;
5161 }
5162 
lru_gen_shrink_lruvec(struct lruvec * lruvec,struct scan_control * sc)5163 static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5164 {
5165 	struct blk_plug plug;
5166 	bool need_aging = false;
5167 	bool need_swapping = false;
5168 	unsigned long scanned = 0;
5169 	unsigned long reclaimed = sc->nr_reclaimed;
5170 	DEFINE_MAX_SEQ(lruvec);
5171 
5172 	lru_add_drain();
5173 
5174 	blk_start_plug(&plug);
5175 
5176 	set_mm_walk(lruvec_pgdat(lruvec));
5177 
5178 	while (true) {
5179 		int delta;
5180 		int swappiness;
5181 		unsigned long nr_to_scan;
5182 
5183 		if (sc->may_swap)
5184 			swappiness = get_swappiness(lruvec, sc);
5185 		else if (!cgroup_reclaim(sc) && get_swappiness(lruvec, sc))
5186 			swappiness = 1;
5187 		else
5188 			swappiness = 0;
5189 
5190 		nr_to_scan = get_nr_to_scan(lruvec, sc, swappiness, &need_aging);
5191 		if (!nr_to_scan)
5192 			goto done;
5193 
5194 		delta = evict_folios(lruvec, sc, swappiness, &need_swapping);
5195 		if (!delta)
5196 			goto done;
5197 
5198 		scanned += delta;
5199 		if (scanned >= nr_to_scan)
5200 			break;
5201 
5202 		if (should_abort_scan(lruvec, max_seq, sc, need_swapping))
5203 			break;
5204 
5205 		cond_resched();
5206 	}
5207 
5208 	/* see the comment in lru_gen_age_node() */
5209 	if (sc->nr_reclaimed - reclaimed >= MIN_LRU_BATCH && !need_aging)
5210 		sc->memcgs_need_aging = false;
5211 done:
5212 	clear_mm_walk();
5213 
5214 	blk_finish_plug(&plug);
5215 }
5216 
5217 /******************************************************************************
5218  *                          state change
5219  ******************************************************************************/
5220 
state_is_valid(struct lruvec * lruvec)5221 static bool __maybe_unused state_is_valid(struct lruvec *lruvec)
5222 {
5223 	struct lru_gen_struct *lrugen = &lruvec->lrugen;
5224 
5225 	if (lrugen->enabled) {
5226 		enum lru_list lru;
5227 
5228 		for_each_evictable_lru(lru) {
5229 			if (!list_empty(&lruvec->lists[lru]))
5230 				return false;
5231 		}
5232 	} else {
5233 		int gen, type, zone;
5234 
5235 		for_each_gen_type_zone(gen, type, zone) {
5236 			if (!list_empty(&lrugen->lists[gen][type][zone]))
5237 				return false;
5238 		}
5239 	}
5240 
5241 	return true;
5242 }
5243 
fill_evictable(struct lruvec * lruvec)5244 static bool fill_evictable(struct lruvec *lruvec)
5245 {
5246 	enum lru_list lru;
5247 	int remaining = MAX_LRU_BATCH;
5248 
5249 	for_each_evictable_lru(lru) {
5250 		int type = is_file_lru(lru);
5251 		bool active = is_active_lru(lru);
5252 		struct list_head *head = &lruvec->lists[lru];
5253 
5254 		while (!list_empty(head)) {
5255 			bool success;
5256 			struct folio *folio = lru_to_folio(head);
5257 
5258 			VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
5259 			VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio) != active, folio);
5260 			VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
5261 			VM_WARN_ON_ONCE_FOLIO(folio_lru_gen(folio) != -1, folio);
5262 
5263 			lruvec_del_folio(lruvec, folio);
5264 			success = lru_gen_add_folio(lruvec, folio, false);
5265 			VM_WARN_ON_ONCE(!success);
5266 
5267 			if (!--remaining)
5268 				return false;
5269 		}
5270 	}
5271 
5272 	return true;
5273 }
5274 
drain_evictable(struct lruvec * lruvec)5275 static bool drain_evictable(struct lruvec *lruvec)
5276 {
5277 	int gen, type, zone;
5278 	int remaining = MAX_LRU_BATCH;
5279 
5280 	for_each_gen_type_zone(gen, type, zone) {
5281 		struct list_head *head = &lruvec->lrugen.lists[gen][type][zone];
5282 
5283 		while (!list_empty(head)) {
5284 			bool success;
5285 			struct folio *folio = lru_to_folio(head);
5286 
5287 			VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio);
5288 			VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio);
5289 			VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio);
5290 			VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio);
5291 
5292 			success = lru_gen_del_folio(lruvec, folio, false);
5293 			VM_WARN_ON_ONCE(!success);
5294 			lruvec_add_folio(lruvec, folio);
5295 
5296 			if (!--remaining)
5297 				return false;
5298 		}
5299 	}
5300 
5301 	return true;
5302 }
5303 
lru_gen_change_state(bool enabled)5304 static void lru_gen_change_state(bool enabled)
5305 {
5306 	static DEFINE_MUTEX(state_mutex);
5307 
5308 	struct mem_cgroup *memcg;
5309 
5310 	cgroup_lock();
5311 	cpus_read_lock();
5312 	get_online_mems();
5313 	mutex_lock(&state_mutex);
5314 
5315 	if (enabled == lru_gen_enabled())
5316 		goto unlock;
5317 
5318 	if (enabled)
5319 		static_branch_enable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5320 	else
5321 		static_branch_disable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]);
5322 
5323 	memcg = mem_cgroup_iter(NULL, NULL, NULL);
5324 	do {
5325 		int nid;
5326 
5327 		for_each_node(nid) {
5328 			struct lruvec *lruvec = get_lruvec(memcg, nid);
5329 
5330 			if (!lruvec)
5331 				continue;
5332 
5333 			spin_lock_irq(&lruvec->lru_lock);
5334 
5335 			VM_WARN_ON_ONCE(!seq_is_valid(lruvec));
5336 			VM_WARN_ON_ONCE(!state_is_valid(lruvec));
5337 
5338 			lruvec->lrugen.enabled = enabled;
5339 
5340 			while (!(enabled ? fill_evictable(lruvec) : drain_evictable(lruvec))) {
5341 				spin_unlock_irq(&lruvec->lru_lock);
5342 				cond_resched();
5343 				spin_lock_irq(&lruvec->lru_lock);
5344 			}
5345 
5346 			spin_unlock_irq(&lruvec->lru_lock);
5347 		}
5348 
5349 		cond_resched();
5350 	} while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5351 unlock:
5352 	mutex_unlock(&state_mutex);
5353 	put_online_mems();
5354 	cpus_read_unlock();
5355 	cgroup_unlock();
5356 }
5357 
5358 /******************************************************************************
5359  *                          sysfs interface
5360  ******************************************************************************/
5361 
show_min_ttl(struct kobject * kobj,struct kobj_attribute * attr,char * buf)5362 static ssize_t show_min_ttl(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5363 {
5364 	return sprintf(buf, "%u\n", jiffies_to_msecs(READ_ONCE(lru_gen_min_ttl)));
5365 }
5366 
5367 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
store_min_ttl(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t len)5368 static ssize_t store_min_ttl(struct kobject *kobj, struct kobj_attribute *attr,
5369 			     const char *buf, size_t len)
5370 {
5371 	unsigned int msecs;
5372 
5373 	if (kstrtouint(buf, 0, &msecs))
5374 		return -EINVAL;
5375 
5376 	WRITE_ONCE(lru_gen_min_ttl, msecs_to_jiffies(msecs));
5377 
5378 	return len;
5379 }
5380 
5381 static struct kobj_attribute lru_gen_min_ttl_attr = __ATTR(
5382 	min_ttl_ms, 0644, show_min_ttl, store_min_ttl
5383 );
5384 
show_enabled(struct kobject * kobj,struct kobj_attribute * attr,char * buf)5385 static ssize_t show_enabled(struct kobject *kobj, struct kobj_attribute *attr, char *buf)
5386 {
5387 	unsigned int caps = 0;
5388 
5389 	if (get_cap(LRU_GEN_CORE))
5390 		caps |= BIT(LRU_GEN_CORE);
5391 
5392 	if (arch_has_hw_pte_young() && get_cap(LRU_GEN_MM_WALK))
5393 		caps |= BIT(LRU_GEN_MM_WALK);
5394 
5395 	if (arch_has_hw_nonleaf_pmd_young() && get_cap(LRU_GEN_NONLEAF_YOUNG))
5396 		caps |= BIT(LRU_GEN_NONLEAF_YOUNG);
5397 
5398 	return snprintf(buf, PAGE_SIZE, "0x%04x\n", caps);
5399 }
5400 
5401 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
store_enabled(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t len)5402 static ssize_t store_enabled(struct kobject *kobj, struct kobj_attribute *attr,
5403 			     const char *buf, size_t len)
5404 {
5405 	int i;
5406 	unsigned int caps;
5407 
5408 	if (tolower(*buf) == 'n')
5409 		caps = 0;
5410 	else if (tolower(*buf) == 'y')
5411 		caps = -1;
5412 	else if (kstrtouint(buf, 0, &caps))
5413 		return -EINVAL;
5414 
5415 	for (i = 0; i < NR_LRU_GEN_CAPS; i++) {
5416 		bool enabled = caps & BIT(i);
5417 
5418 		if (i == LRU_GEN_CORE)
5419 			lru_gen_change_state(enabled);
5420 		else if (enabled)
5421 			static_branch_enable(&lru_gen_caps[i]);
5422 		else
5423 			static_branch_disable(&lru_gen_caps[i]);
5424 	}
5425 
5426 	return len;
5427 }
5428 
5429 static struct kobj_attribute lru_gen_enabled_attr = __ATTR(
5430 	enabled, 0644, show_enabled, store_enabled
5431 );
5432 
5433 static struct attribute *lru_gen_attrs[] = {
5434 	&lru_gen_min_ttl_attr.attr,
5435 	&lru_gen_enabled_attr.attr,
5436 	NULL
5437 };
5438 
5439 static struct attribute_group lru_gen_attr_group = {
5440 	.name = "lru_gen",
5441 	.attrs = lru_gen_attrs,
5442 };
5443 
5444 /******************************************************************************
5445  *                          debugfs interface
5446  ******************************************************************************/
5447 
lru_gen_seq_start(struct seq_file * m,loff_t * pos)5448 static void *lru_gen_seq_start(struct seq_file *m, loff_t *pos)
5449 {
5450 	struct mem_cgroup *memcg;
5451 	loff_t nr_to_skip = *pos;
5452 
5453 	m->private = kvmalloc(PATH_MAX, GFP_KERNEL);
5454 	if (!m->private)
5455 		return ERR_PTR(-ENOMEM);
5456 
5457 	memcg = mem_cgroup_iter(NULL, NULL, NULL);
5458 	do {
5459 		int nid;
5460 
5461 		for_each_node_state(nid, N_MEMORY) {
5462 			if (!nr_to_skip--)
5463 				return get_lruvec(memcg, nid);
5464 		}
5465 	} while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)));
5466 
5467 	return NULL;
5468 }
5469 
lru_gen_seq_stop(struct seq_file * m,void * v)5470 static void lru_gen_seq_stop(struct seq_file *m, void *v)
5471 {
5472 	if (!IS_ERR_OR_NULL(v))
5473 		mem_cgroup_iter_break(NULL, lruvec_memcg(v));
5474 
5475 	kvfree(m->private);
5476 	m->private = NULL;
5477 }
5478 
lru_gen_seq_next(struct seq_file * m,void * v,loff_t * pos)5479 static void *lru_gen_seq_next(struct seq_file *m, void *v, loff_t *pos)
5480 {
5481 	int nid = lruvec_pgdat(v)->node_id;
5482 	struct mem_cgroup *memcg = lruvec_memcg(v);
5483 
5484 	++*pos;
5485 
5486 	nid = next_memory_node(nid);
5487 	if (nid == MAX_NUMNODES) {
5488 		memcg = mem_cgroup_iter(NULL, memcg, NULL);
5489 		if (!memcg)
5490 			return NULL;
5491 
5492 		nid = first_memory_node;
5493 	}
5494 
5495 	return get_lruvec(memcg, nid);
5496 }
5497 
lru_gen_seq_show_full(struct seq_file * m,struct lruvec * lruvec,unsigned long max_seq,unsigned long * min_seq,unsigned long seq)5498 static void lru_gen_seq_show_full(struct seq_file *m, struct lruvec *lruvec,
5499 				  unsigned long max_seq, unsigned long *min_seq,
5500 				  unsigned long seq)
5501 {
5502 	int i;
5503 	int type, tier;
5504 	int hist = lru_hist_from_seq(seq);
5505 	struct lru_gen_struct *lrugen = &lruvec->lrugen;
5506 
5507 	for (tier = 0; tier < MAX_NR_TIERS; tier++) {
5508 		seq_printf(m, "            %10d", tier);
5509 		for (type = 0; type < ANON_AND_FILE; type++) {
5510 			const char *s = "   ";
5511 			unsigned long n[3] = {};
5512 
5513 			if (seq == max_seq) {
5514 				s = "RT ";
5515 				n[0] = READ_ONCE(lrugen->avg_refaulted[type][tier]);
5516 				n[1] = READ_ONCE(lrugen->avg_total[type][tier]);
5517 			} else if (seq == min_seq[type] || NR_HIST_GENS > 1) {
5518 				s = "rep";
5519 				n[0] = atomic_long_read(&lrugen->refaulted[hist][type][tier]);
5520 				n[1] = atomic_long_read(&lrugen->evicted[hist][type][tier]);
5521 				if (tier)
5522 					n[2] = READ_ONCE(lrugen->protected[hist][type][tier - 1]);
5523 			}
5524 
5525 			for (i = 0; i < 3; i++)
5526 				seq_printf(m, " %10lu%c", n[i], s[i]);
5527 		}
5528 		seq_putc(m, '\n');
5529 	}
5530 
5531 	seq_puts(m, "                      ");
5532 	for (i = 0; i < NR_MM_STATS; i++) {
5533 		const char *s = "      ";
5534 		unsigned long n = 0;
5535 
5536 		if (seq == max_seq && NR_HIST_GENS == 1) {
5537 			s = "LOYNFA";
5538 			n = READ_ONCE(lruvec->mm_state.stats[hist][i]);
5539 		} else if (seq != max_seq && NR_HIST_GENS > 1) {
5540 			s = "loynfa";
5541 			n = READ_ONCE(lruvec->mm_state.stats[hist][i]);
5542 		}
5543 
5544 		seq_printf(m, " %10lu%c", n, s[i]);
5545 	}
5546 	seq_putc(m, '\n');
5547 }
5548 
5549 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
lru_gen_seq_show(struct seq_file * m,void * v)5550 static int lru_gen_seq_show(struct seq_file *m, void *v)
5551 {
5552 	unsigned long seq;
5553 	bool full = !debugfs_real_fops(m->file)->write;
5554 	struct lruvec *lruvec = v;
5555 	struct lru_gen_struct *lrugen = &lruvec->lrugen;
5556 	int nid = lruvec_pgdat(lruvec)->node_id;
5557 	struct mem_cgroup *memcg = lruvec_memcg(lruvec);
5558 	DEFINE_MAX_SEQ(lruvec);
5559 	DEFINE_MIN_SEQ(lruvec);
5560 
5561 	if (nid == first_memory_node) {
5562 		const char *path = memcg ? m->private : "";
5563 
5564 #ifdef CONFIG_MEMCG
5565 		if (memcg)
5566 			cgroup_path(memcg->css.cgroup, m->private, PATH_MAX);
5567 #endif
5568 		seq_printf(m, "memcg %5hu %s\n", mem_cgroup_id(memcg), path);
5569 	}
5570 
5571 	seq_printf(m, " node %5d\n", nid);
5572 
5573 	if (!full)
5574 		seq = min_seq[LRU_GEN_ANON];
5575 	else if (max_seq >= MAX_NR_GENS)
5576 		seq = max_seq - MAX_NR_GENS + 1;
5577 	else
5578 		seq = 0;
5579 
5580 	for (; seq <= max_seq; seq++) {
5581 		int type, zone;
5582 		int gen = lru_gen_from_seq(seq);
5583 		unsigned long birth = READ_ONCE(lruvec->lrugen.timestamps[gen]);
5584 
5585 		seq_printf(m, " %10lu %10u", seq, jiffies_to_msecs(jiffies - birth));
5586 
5587 		for (type = 0; type < ANON_AND_FILE; type++) {
5588 			unsigned long size = 0;
5589 			char mark = full && seq < min_seq[type] ? 'x' : ' ';
5590 
5591 			for (zone = 0; zone < MAX_NR_ZONES; zone++)
5592 				size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L);
5593 
5594 			seq_printf(m, " %10lu%c", size, mark);
5595 		}
5596 
5597 		seq_putc(m, '\n');
5598 
5599 		if (full)
5600 			lru_gen_seq_show_full(m, lruvec, max_seq, min_seq, seq);
5601 	}
5602 
5603 	return 0;
5604 }
5605 
5606 static const struct seq_operations lru_gen_seq_ops = {
5607 	.start = lru_gen_seq_start,
5608 	.stop = lru_gen_seq_stop,
5609 	.next = lru_gen_seq_next,
5610 	.show = lru_gen_seq_show,
5611 };
5612 
run_aging(struct lruvec * lruvec,unsigned long seq,struct scan_control * sc,bool can_swap,bool force_scan)5613 static int run_aging(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc,
5614 		     bool can_swap, bool force_scan)
5615 {
5616 	DEFINE_MAX_SEQ(lruvec);
5617 	DEFINE_MIN_SEQ(lruvec);
5618 
5619 	if (seq < max_seq)
5620 		return 0;
5621 
5622 	if (seq > max_seq)
5623 		return -EINVAL;
5624 
5625 	if (!force_scan && min_seq[!can_swap] + MAX_NR_GENS - 1 <= max_seq)
5626 		return -ERANGE;
5627 
5628 	try_to_inc_max_seq(lruvec, max_seq, sc, can_swap, force_scan);
5629 
5630 	return 0;
5631 }
5632 
run_eviction(struct lruvec * lruvec,unsigned long seq,struct scan_control * sc,int swappiness,unsigned long nr_to_reclaim)5633 static int run_eviction(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc,
5634 			int swappiness, unsigned long nr_to_reclaim)
5635 {
5636 	DEFINE_MAX_SEQ(lruvec);
5637 
5638 	if (seq + MIN_NR_GENS > max_seq)
5639 		return -EINVAL;
5640 
5641 	sc->nr_reclaimed = 0;
5642 
5643 	while (!signal_pending(current)) {
5644 		DEFINE_MIN_SEQ(lruvec);
5645 
5646 		if (seq < min_seq[!swappiness])
5647 			return 0;
5648 
5649 		if (sc->nr_reclaimed >= nr_to_reclaim)
5650 			return 0;
5651 
5652 		if (!evict_folios(lruvec, sc, swappiness, NULL))
5653 			return 0;
5654 
5655 		cond_resched();
5656 	}
5657 
5658 	return -EINTR;
5659 }
5660 
run_cmd(char cmd,int memcg_id,int nid,unsigned long seq,struct scan_control * sc,int swappiness,unsigned long opt)5661 static int run_cmd(char cmd, int memcg_id, int nid, unsigned long seq,
5662 		   struct scan_control *sc, int swappiness, unsigned long opt)
5663 {
5664 	struct lruvec *lruvec;
5665 	int err = -EINVAL;
5666 	struct mem_cgroup *memcg = NULL;
5667 
5668 	if (nid < 0 || nid >= MAX_NUMNODES || !node_state(nid, N_MEMORY))
5669 		return -EINVAL;
5670 
5671 	if (!mem_cgroup_disabled()) {
5672 		rcu_read_lock();
5673 		memcg = mem_cgroup_from_id(memcg_id);
5674 #ifdef CONFIG_MEMCG
5675 		if (memcg && !css_tryget(&memcg->css))
5676 			memcg = NULL;
5677 #endif
5678 		rcu_read_unlock();
5679 
5680 		if (!memcg)
5681 			return -EINVAL;
5682 	}
5683 
5684 	if (memcg_id != mem_cgroup_id(memcg))
5685 		goto done;
5686 
5687 	lruvec = get_lruvec(memcg, nid);
5688 
5689 	if (swappiness < 0)
5690 		swappiness = get_swappiness(lruvec, sc);
5691 	else if (swappiness > 200)
5692 		goto done;
5693 
5694 	switch (cmd) {
5695 	case '+':
5696 		err = run_aging(lruvec, seq, sc, swappiness, opt);
5697 		break;
5698 	case '-':
5699 		err = run_eviction(lruvec, seq, sc, swappiness, opt);
5700 		break;
5701 	}
5702 done:
5703 	mem_cgroup_put(memcg);
5704 
5705 	return err;
5706 }
5707 
5708 /* see Documentation/admin-guide/mm/multigen_lru.rst for details */
lru_gen_seq_write(struct file * file,const char __user * src,size_t len,loff_t * pos)5709 static ssize_t lru_gen_seq_write(struct file *file, const char __user *src,
5710 				 size_t len, loff_t *pos)
5711 {
5712 	void *buf;
5713 	char *cur, *next;
5714 	unsigned int flags;
5715 	struct blk_plug plug;
5716 	int err = -EINVAL;
5717 	struct scan_control sc = {
5718 		.may_writepage = true,
5719 		.may_unmap = true,
5720 		.may_swap = true,
5721 		.reclaim_idx = MAX_NR_ZONES - 1,
5722 		.gfp_mask = GFP_KERNEL,
5723 	};
5724 
5725 	buf = kvmalloc(len + 1, GFP_KERNEL);
5726 	if (!buf)
5727 		return -ENOMEM;
5728 
5729 	if (copy_from_user(buf, src, len)) {
5730 		kvfree(buf);
5731 		return -EFAULT;
5732 	}
5733 
5734 	set_task_reclaim_state(current, &sc.reclaim_state);
5735 	flags = memalloc_noreclaim_save();
5736 	blk_start_plug(&plug);
5737 	if (!set_mm_walk(NULL)) {
5738 		err = -ENOMEM;
5739 		goto done;
5740 	}
5741 
5742 	next = buf;
5743 	next[len] = '\0';
5744 
5745 	while ((cur = strsep(&next, ",;\n"))) {
5746 		int n;
5747 		int end;
5748 		char cmd;
5749 		unsigned int memcg_id;
5750 		unsigned int nid;
5751 		unsigned long seq;
5752 		unsigned int swappiness = -1;
5753 		unsigned long opt = -1;
5754 
5755 		cur = skip_spaces(cur);
5756 		if (!*cur)
5757 			continue;
5758 
5759 		n = sscanf(cur, "%c %u %u %lu %n %u %n %lu %n", &cmd, &memcg_id, &nid,
5760 			   &seq, &end, &swappiness, &end, &opt, &end);
5761 		if (n < 4 || cur[end]) {
5762 			err = -EINVAL;
5763 			break;
5764 		}
5765 
5766 		err = run_cmd(cmd, memcg_id, nid, seq, &sc, swappiness, opt);
5767 		if (err)
5768 			break;
5769 	}
5770 done:
5771 	clear_mm_walk();
5772 	blk_finish_plug(&plug);
5773 	memalloc_noreclaim_restore(flags);
5774 	set_task_reclaim_state(current, NULL);
5775 
5776 	kvfree(buf);
5777 
5778 	return err ? : len;
5779 }
5780 
lru_gen_seq_open(struct inode * inode,struct file * file)5781 static int lru_gen_seq_open(struct inode *inode, struct file *file)
5782 {
5783 	return seq_open(file, &lru_gen_seq_ops);
5784 }
5785 
5786 static const struct file_operations lru_gen_rw_fops = {
5787 	.open = lru_gen_seq_open,
5788 	.read = seq_read,
5789 	.write = lru_gen_seq_write,
5790 	.llseek = seq_lseek,
5791 	.release = seq_release,
5792 };
5793 
5794 static const struct file_operations lru_gen_ro_fops = {
5795 	.open = lru_gen_seq_open,
5796 	.read = seq_read,
5797 	.llseek = seq_lseek,
5798 	.release = seq_release,
5799 };
5800 
5801 /******************************************************************************
5802  *                          initialization
5803  ******************************************************************************/
5804 
lru_gen_init_lruvec(struct lruvec * lruvec)5805 void lru_gen_init_lruvec(struct lruvec *lruvec)
5806 {
5807 	int i;
5808 	int gen, type, zone;
5809 	struct lru_gen_struct *lrugen = &lruvec->lrugen;
5810 
5811 	lrugen->max_seq = MIN_NR_GENS + 1;
5812 	lrugen->enabled = lru_gen_enabled();
5813 
5814 	for (i = 0; i <= MIN_NR_GENS + 1; i++)
5815 		lrugen->timestamps[i] = jiffies;
5816 
5817 	for_each_gen_type_zone(gen, type, zone)
5818 		INIT_LIST_HEAD(&lrugen->lists[gen][type][zone]);
5819 
5820 	lruvec->mm_state.seq = MIN_NR_GENS;
5821 	init_waitqueue_head(&lruvec->mm_state.wait);
5822 }
5823 
5824 #ifdef CONFIG_MEMCG
lru_gen_init_memcg(struct mem_cgroup * memcg)5825 void lru_gen_init_memcg(struct mem_cgroup *memcg)
5826 {
5827 	INIT_LIST_HEAD(&memcg->mm_list.fifo);
5828 	spin_lock_init(&memcg->mm_list.lock);
5829 }
5830 
lru_gen_exit_memcg(struct mem_cgroup * memcg)5831 void lru_gen_exit_memcg(struct mem_cgroup *memcg)
5832 {
5833 	int i;
5834 	int nid;
5835 
5836 	for_each_node(nid) {
5837 		struct lruvec *lruvec = get_lruvec(memcg, nid);
5838 
5839 		VM_WARN_ON_ONCE(memchr_inv(lruvec->lrugen.nr_pages, 0,
5840 					   sizeof(lruvec->lrugen.nr_pages)));
5841 
5842 		for (i = 0; i < NR_BLOOM_FILTERS; i++) {
5843 			bitmap_free(lruvec->mm_state.filters[i]);
5844 			lruvec->mm_state.filters[i] = NULL;
5845 		}
5846 	}
5847 }
5848 #endif
5849 
init_lru_gen(void)5850 static int __init init_lru_gen(void)
5851 {
5852 	BUILD_BUG_ON(MIN_NR_GENS + 1 >= MAX_NR_GENS);
5853 	BUILD_BUG_ON(BIT(LRU_GEN_WIDTH) <= MAX_NR_GENS);
5854 
5855 	if (sysfs_create_group(mm_kobj, &lru_gen_attr_group))
5856 		pr_err("lru_gen: failed to create sysfs group\n");
5857 
5858 	debugfs_create_file("lru_gen", 0644, NULL, NULL, &lru_gen_rw_fops);
5859 	debugfs_create_file("lru_gen_full", 0444, NULL, NULL, &lru_gen_ro_fops);
5860 
5861 	return 0;
5862 };
5863 late_initcall(init_lru_gen);
5864 
5865 #else /* !CONFIG_LRU_GEN */
5866 
lru_gen_age_node(struct pglist_data * pgdat,struct scan_control * sc)5867 static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
5868 {
5869 }
5870 
lru_gen_shrink_lruvec(struct lruvec * lruvec,struct scan_control * sc)5871 static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5872 {
5873 }
5874 
5875 #endif /* CONFIG_LRU_GEN */
5876 
shrink_lruvec(struct lruvec * lruvec,struct scan_control * sc)5877 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
5878 {
5879 	unsigned long nr[NR_LRU_LISTS];
5880 	unsigned long targets[NR_LRU_LISTS];
5881 	unsigned long nr_to_scan;
5882 	enum lru_list lru;
5883 	unsigned long nr_reclaimed = 0;
5884 	unsigned long nr_to_reclaim = sc->nr_to_reclaim;
5885 	bool proportional_reclaim;
5886 	struct blk_plug plug;
5887 
5888 	if (lru_gen_enabled()) {
5889 		lru_gen_shrink_lruvec(lruvec, sc);
5890 		return;
5891 	}
5892 
5893 	get_scan_count(lruvec, sc, nr);
5894 
5895 	/* Record the original scan target for proportional adjustments later */
5896 	memcpy(targets, nr, sizeof(nr));
5897 
5898 	/*
5899 	 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
5900 	 * event that can occur when there is little memory pressure e.g.
5901 	 * multiple streaming readers/writers. Hence, we do not abort scanning
5902 	 * when the requested number of pages are reclaimed when scanning at
5903 	 * DEF_PRIORITY on the assumption that the fact we are direct
5904 	 * reclaiming implies that kswapd is not keeping up and it is best to
5905 	 * do a batch of work at once. For memcg reclaim one check is made to
5906 	 * abort proportional reclaim if either the file or anon lru has already
5907 	 * dropped to zero at the first pass.
5908 	 */
5909 	proportional_reclaim = (!cgroup_reclaim(sc) && !current_is_kswapd() &&
5910 				sc->priority == DEF_PRIORITY);
5911 
5912 	blk_start_plug(&plug);
5913 	while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
5914 					nr[LRU_INACTIVE_FILE]) {
5915 		unsigned long nr_anon, nr_file, percentage;
5916 		unsigned long nr_scanned;
5917 
5918 		for_each_evictable_lru(lru) {
5919 			if (nr[lru]) {
5920 				nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
5921 				nr[lru] -= nr_to_scan;
5922 
5923 				nr_reclaimed += shrink_list(lru, nr_to_scan,
5924 							    lruvec, sc);
5925 			}
5926 		}
5927 
5928 		cond_resched();
5929 
5930 		if (nr_reclaimed < nr_to_reclaim || proportional_reclaim)
5931 			continue;
5932 
5933 		/*
5934 		 * For kswapd and memcg, reclaim at least the number of pages
5935 		 * requested. Ensure that the anon and file LRUs are scanned
5936 		 * proportionally what was requested by get_scan_count(). We
5937 		 * stop reclaiming one LRU and reduce the amount scanning
5938 		 * proportional to the original scan target.
5939 		 */
5940 		nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
5941 		nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
5942 
5943 		/*
5944 		 * It's just vindictive to attack the larger once the smaller
5945 		 * has gone to zero.  And given the way we stop scanning the
5946 		 * smaller below, this makes sure that we only make one nudge
5947 		 * towards proportionality once we've got nr_to_reclaim.
5948 		 */
5949 		if (!nr_file || !nr_anon)
5950 			break;
5951 
5952 		if (nr_file > nr_anon) {
5953 			unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
5954 						targets[LRU_ACTIVE_ANON] + 1;
5955 			lru = LRU_BASE;
5956 			percentage = nr_anon * 100 / scan_target;
5957 		} else {
5958 			unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
5959 						targets[LRU_ACTIVE_FILE] + 1;
5960 			lru = LRU_FILE;
5961 			percentage = nr_file * 100 / scan_target;
5962 		}
5963 
5964 		/* Stop scanning the smaller of the LRU */
5965 		nr[lru] = 0;
5966 		nr[lru + LRU_ACTIVE] = 0;
5967 
5968 		/*
5969 		 * Recalculate the other LRU scan count based on its original
5970 		 * scan target and the percentage scanning already complete
5971 		 */
5972 		lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
5973 		nr_scanned = targets[lru] - nr[lru];
5974 		nr[lru] = targets[lru] * (100 - percentage) / 100;
5975 		nr[lru] -= min(nr[lru], nr_scanned);
5976 
5977 		lru += LRU_ACTIVE;
5978 		nr_scanned = targets[lru] - nr[lru];
5979 		nr[lru] = targets[lru] * (100 - percentage) / 100;
5980 		nr[lru] -= min(nr[lru], nr_scanned);
5981 	}
5982 	blk_finish_plug(&plug);
5983 	sc->nr_reclaimed += nr_reclaimed;
5984 
5985 	/*
5986 	 * Even if we did not try to evict anon pages at all, we want to
5987 	 * rebalance the anon lru active/inactive ratio.
5988 	 */
5989 	if (can_age_anon_pages(lruvec_pgdat(lruvec), sc) &&
5990 	    inactive_is_low(lruvec, LRU_INACTIVE_ANON))
5991 		shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
5992 				   sc, LRU_ACTIVE_ANON);
5993 }
5994 
5995 /* Use reclaim/compaction for costly allocs or under memory pressure */
in_reclaim_compaction(struct scan_control * sc)5996 static bool in_reclaim_compaction(struct scan_control *sc)
5997 {
5998 	if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
5999 			(sc->order > PAGE_ALLOC_COSTLY_ORDER ||
6000 			 sc->priority < DEF_PRIORITY - 2))
6001 		return true;
6002 
6003 	return false;
6004 }
6005 
6006 /*
6007  * Reclaim/compaction is used for high-order allocation requests. It reclaims
6008  * order-0 pages before compacting the zone. should_continue_reclaim() returns
6009  * true if more pages should be reclaimed such that when the page allocator
6010  * calls try_to_compact_pages() that it will have enough free pages to succeed.
6011  * It will give up earlier than that if there is difficulty reclaiming pages.
6012  */
should_continue_reclaim(struct pglist_data * pgdat,unsigned long nr_reclaimed,struct scan_control * sc)6013 static inline bool should_continue_reclaim(struct pglist_data *pgdat,
6014 					unsigned long nr_reclaimed,
6015 					struct scan_control *sc)
6016 {
6017 	unsigned long pages_for_compaction;
6018 	unsigned long inactive_lru_pages;
6019 	int z;
6020 
6021 	/* If not in reclaim/compaction mode, stop */
6022 	if (!in_reclaim_compaction(sc))
6023 		return false;
6024 
6025 	/*
6026 	 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
6027 	 * number of pages that were scanned. This will return to the caller
6028 	 * with the risk reclaim/compaction and the resulting allocation attempt
6029 	 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
6030 	 * allocations through requiring that the full LRU list has been scanned
6031 	 * first, by assuming that zero delta of sc->nr_scanned means full LRU
6032 	 * scan, but that approximation was wrong, and there were corner cases
6033 	 * where always a non-zero amount of pages were scanned.
6034 	 */
6035 	if (!nr_reclaimed)
6036 		return false;
6037 
6038 	/* If compaction would go ahead or the allocation would succeed, stop */
6039 	for (z = 0; z <= sc->reclaim_idx; z++) {
6040 		struct zone *zone = &pgdat->node_zones[z];
6041 		if (!managed_zone(zone))
6042 			continue;
6043 
6044 		switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
6045 		case COMPACT_SUCCESS:
6046 		case COMPACT_CONTINUE:
6047 			return false;
6048 		default:
6049 			/* check next zone */
6050 			;
6051 		}
6052 	}
6053 
6054 	/*
6055 	 * If we have not reclaimed enough pages for compaction and the
6056 	 * inactive lists are large enough, continue reclaiming
6057 	 */
6058 	pages_for_compaction = compact_gap(sc->order);
6059 	inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
6060 	if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc))
6061 		inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
6062 
6063 	return inactive_lru_pages > pages_for_compaction;
6064 }
6065 
shrink_node_memcgs(pg_data_t * pgdat,struct scan_control * sc)6066 static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
6067 {
6068 	struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
6069 	struct mem_cgroup *memcg;
6070 
6071 	memcg = mem_cgroup_iter(target_memcg, NULL, NULL);
6072 	do {
6073 		struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
6074 		unsigned long reclaimed;
6075 		unsigned long scanned;
6076 
6077 		/*
6078 		 * This loop can become CPU-bound when target memcgs
6079 		 * aren't eligible for reclaim - either because they
6080 		 * don't have any reclaimable pages, or because their
6081 		 * memory is explicitly protected. Avoid soft lockups.
6082 		 */
6083 		cond_resched();
6084 
6085 		mem_cgroup_calculate_protection(target_memcg, memcg);
6086 
6087 		if (mem_cgroup_below_min(memcg)) {
6088 			/*
6089 			 * Hard protection.
6090 			 * If there is no reclaimable memory, OOM.
6091 			 */
6092 			continue;
6093 		} else if (mem_cgroup_below_low(memcg)) {
6094 			/*
6095 			 * Soft protection.
6096 			 * Respect the protection only as long as
6097 			 * there is an unprotected supply
6098 			 * of reclaimable memory from other cgroups.
6099 			 */
6100 			if (!sc->memcg_low_reclaim) {
6101 				sc->memcg_low_skipped = 1;
6102 				continue;
6103 			}
6104 			memcg_memory_event(memcg, MEMCG_LOW);
6105 		}
6106 
6107 		reclaimed = sc->nr_reclaimed;
6108 		scanned = sc->nr_scanned;
6109 
6110 		shrink_lruvec(lruvec, sc);
6111 
6112 		shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
6113 			    sc->priority);
6114 
6115 		/* Record the group's reclaim efficiency */
6116 		if (!sc->proactive)
6117 			vmpressure(sc->gfp_mask, memcg, false,
6118 				   sc->nr_scanned - scanned,
6119 				   sc->nr_reclaimed - reclaimed);
6120 
6121 	} while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL)));
6122 }
6123 
shrink_node(pg_data_t * pgdat,struct scan_control * sc)6124 static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
6125 {
6126 	struct reclaim_state *reclaim_state = current->reclaim_state;
6127 	unsigned long nr_reclaimed, nr_scanned;
6128 	struct lruvec *target_lruvec;
6129 	bool reclaimable = false;
6130 
6131 	target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
6132 
6133 again:
6134 	memset(&sc->nr, 0, sizeof(sc->nr));
6135 
6136 	nr_reclaimed = sc->nr_reclaimed;
6137 	nr_scanned = sc->nr_scanned;
6138 
6139 	prepare_scan_count(pgdat, sc);
6140 
6141 	shrink_node_memcgs(pgdat, sc);
6142 
6143 	if (reclaim_state) {
6144 		sc->nr_reclaimed += reclaim_state->reclaimed_slab;
6145 		reclaim_state->reclaimed_slab = 0;
6146 	}
6147 
6148 	/* Record the subtree's reclaim efficiency */
6149 	if (!sc->proactive)
6150 		vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
6151 			   sc->nr_scanned - nr_scanned,
6152 			   sc->nr_reclaimed - nr_reclaimed);
6153 
6154 	if (sc->nr_reclaimed - nr_reclaimed)
6155 		reclaimable = true;
6156 
6157 	if (current_is_kswapd()) {
6158 		/*
6159 		 * If reclaim is isolating dirty pages under writeback,
6160 		 * it implies that the long-lived page allocation rate
6161 		 * is exceeding the page laundering rate. Either the
6162 		 * global limits are not being effective at throttling
6163 		 * processes due to the page distribution throughout
6164 		 * zones or there is heavy usage of a slow backing
6165 		 * device. The only option is to throttle from reclaim
6166 		 * context which is not ideal as there is no guarantee
6167 		 * the dirtying process is throttled in the same way
6168 		 * balance_dirty_pages() manages.
6169 		 *
6170 		 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
6171 		 * count the number of pages under pages flagged for
6172 		 * immediate reclaim and stall if any are encountered
6173 		 * in the nr_immediate check below.
6174 		 */
6175 		if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
6176 			set_bit(PGDAT_WRITEBACK, &pgdat->flags);
6177 
6178 		/* Allow kswapd to start writing pages during reclaim.*/
6179 		if (sc->nr.unqueued_dirty == sc->nr.file_taken)
6180 			set_bit(PGDAT_DIRTY, &pgdat->flags);
6181 
6182 		/*
6183 		 * If kswapd scans pages marked for immediate
6184 		 * reclaim and under writeback (nr_immediate), it
6185 		 * implies that pages are cycling through the LRU
6186 		 * faster than they are written so forcibly stall
6187 		 * until some pages complete writeback.
6188 		 */
6189 		if (sc->nr.immediate)
6190 			reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK);
6191 	}
6192 
6193 	/*
6194 	 * Tag a node/memcg as congested if all the dirty pages were marked
6195 	 * for writeback and immediate reclaim (counted in nr.congested).
6196 	 *
6197 	 * Legacy memcg will stall in page writeback so avoid forcibly
6198 	 * stalling in reclaim_throttle().
6199 	 */
6200 	if ((current_is_kswapd() ||
6201 	     (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) &&
6202 	    sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
6203 		set_bit(LRUVEC_CONGESTED, &target_lruvec->flags);
6204 
6205 	/*
6206 	 * Stall direct reclaim for IO completions if the lruvec is
6207 	 * node is congested. Allow kswapd to continue until it
6208 	 * starts encountering unqueued dirty pages or cycling through
6209 	 * the LRU too quickly.
6210 	 */
6211 	if (!current_is_kswapd() && current_may_throttle() &&
6212 	    !sc->hibernation_mode &&
6213 	    test_bit(LRUVEC_CONGESTED, &target_lruvec->flags))
6214 		reclaim_throttle(pgdat, VMSCAN_THROTTLE_CONGESTED);
6215 
6216 	if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
6217 				    sc))
6218 		goto again;
6219 
6220 	/*
6221 	 * Kswapd gives up on balancing particular nodes after too
6222 	 * many failures to reclaim anything from them and goes to
6223 	 * sleep. On reclaim progress, reset the failure counter. A
6224 	 * successful direct reclaim run will revive a dormant kswapd.
6225 	 */
6226 	if (reclaimable)
6227 		pgdat->kswapd_failures = 0;
6228 }
6229 
6230 /*
6231  * Returns true if compaction should go ahead for a costly-order request, or
6232  * the allocation would already succeed without compaction. Return false if we
6233  * should reclaim first.
6234  */
compaction_ready(struct zone * zone,struct scan_control * sc)6235 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
6236 {
6237 	unsigned long watermark;
6238 	enum compact_result suitable;
6239 
6240 	suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
6241 	if (suitable == COMPACT_SUCCESS)
6242 		/* Allocation should succeed already. Don't reclaim. */
6243 		return true;
6244 	if (suitable == COMPACT_SKIPPED)
6245 		/* Compaction cannot yet proceed. Do reclaim. */
6246 		return false;
6247 
6248 	/*
6249 	 * Compaction is already possible, but it takes time to run and there
6250 	 * are potentially other callers using the pages just freed. So proceed
6251 	 * with reclaim to make a buffer of free pages available to give
6252 	 * compaction a reasonable chance of completing and allocating the page.
6253 	 * Note that we won't actually reclaim the whole buffer in one attempt
6254 	 * as the target watermark in should_continue_reclaim() is lower. But if
6255 	 * we are already above the high+gap watermark, don't reclaim at all.
6256 	 */
6257 	watermark = high_wmark_pages(zone) + compact_gap(sc->order);
6258 
6259 	return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
6260 }
6261 
consider_reclaim_throttle(pg_data_t * pgdat,struct scan_control * sc)6262 static void consider_reclaim_throttle(pg_data_t *pgdat, struct scan_control *sc)
6263 {
6264 	/*
6265 	 * If reclaim is making progress greater than 12% efficiency then
6266 	 * wake all the NOPROGRESS throttled tasks.
6267 	 */
6268 	if (sc->nr_reclaimed > (sc->nr_scanned >> 3)) {
6269 		wait_queue_head_t *wqh;
6270 
6271 		wqh = &pgdat->reclaim_wait[VMSCAN_THROTTLE_NOPROGRESS];
6272 		if (waitqueue_active(wqh))
6273 			wake_up(wqh);
6274 
6275 		return;
6276 	}
6277 
6278 	/*
6279 	 * Do not throttle kswapd or cgroup reclaim on NOPROGRESS as it will
6280 	 * throttle on VMSCAN_THROTTLE_WRITEBACK if there are too many pages
6281 	 * under writeback and marked for immediate reclaim at the tail of the
6282 	 * LRU.
6283 	 */
6284 	if (current_is_kswapd() || cgroup_reclaim(sc))
6285 		return;
6286 
6287 	/* Throttle if making no progress at high prioities. */
6288 	if (sc->priority == 1 && !sc->nr_reclaimed)
6289 		reclaim_throttle(pgdat, VMSCAN_THROTTLE_NOPROGRESS);
6290 }
6291 
6292 /*
6293  * This is the direct reclaim path, for page-allocating processes.  We only
6294  * try to reclaim pages from zones which will satisfy the caller's allocation
6295  * request.
6296  *
6297  * If a zone is deemed to be full of pinned pages then just give it a light
6298  * scan then give up on it.
6299  */
shrink_zones(struct zonelist * zonelist,struct scan_control * sc)6300 static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
6301 {
6302 	struct zoneref *z;
6303 	struct zone *zone;
6304 	unsigned long nr_soft_reclaimed;
6305 	unsigned long nr_soft_scanned;
6306 	gfp_t orig_mask;
6307 	pg_data_t *last_pgdat = NULL;
6308 	pg_data_t *first_pgdat = NULL;
6309 
6310 	/*
6311 	 * If the number of buffer_heads in the machine exceeds the maximum
6312 	 * allowed level, force direct reclaim to scan the highmem zone as
6313 	 * highmem pages could be pinning lowmem pages storing buffer_heads
6314 	 */
6315 	orig_mask = sc->gfp_mask;
6316 	if (buffer_heads_over_limit) {
6317 		sc->gfp_mask |= __GFP_HIGHMEM;
6318 		sc->reclaim_idx = gfp_zone(sc->gfp_mask);
6319 	}
6320 
6321 	for_each_zone_zonelist_nodemask(zone, z, zonelist,
6322 					sc->reclaim_idx, sc->nodemask) {
6323 		/*
6324 		 * Take care memory controller reclaiming has small influence
6325 		 * to global LRU.
6326 		 */
6327 		if (!cgroup_reclaim(sc)) {
6328 			if (!cpuset_zone_allowed(zone,
6329 						 GFP_KERNEL | __GFP_HARDWALL))
6330 				continue;
6331 
6332 			/*
6333 			 * If we already have plenty of memory free for
6334 			 * compaction in this zone, don't free any more.
6335 			 * Even though compaction is invoked for any
6336 			 * non-zero order, only frequent costly order
6337 			 * reclamation is disruptive enough to become a
6338 			 * noticeable problem, like transparent huge
6339 			 * page allocations.
6340 			 */
6341 			if (IS_ENABLED(CONFIG_COMPACTION) &&
6342 			    sc->order > PAGE_ALLOC_COSTLY_ORDER &&
6343 			    compaction_ready(zone, sc)) {
6344 				sc->compaction_ready = true;
6345 				continue;
6346 			}
6347 
6348 			/*
6349 			 * Shrink each node in the zonelist once. If the
6350 			 * zonelist is ordered by zone (not the default) then a
6351 			 * node may be shrunk multiple times but in that case
6352 			 * the user prefers lower zones being preserved.
6353 			 */
6354 			if (zone->zone_pgdat == last_pgdat)
6355 				continue;
6356 
6357 			/*
6358 			 * This steals pages from memory cgroups over softlimit
6359 			 * and returns the number of reclaimed pages and
6360 			 * scanned pages. This works for global memory pressure
6361 			 * and balancing, not for a memcg's limit.
6362 			 */
6363 			nr_soft_scanned = 0;
6364 			nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
6365 						sc->order, sc->gfp_mask,
6366 						&nr_soft_scanned);
6367 			sc->nr_reclaimed += nr_soft_reclaimed;
6368 			sc->nr_scanned += nr_soft_scanned;
6369 			/* need some check for avoid more shrink_zone() */
6370 		}
6371 
6372 		if (!first_pgdat)
6373 			first_pgdat = zone->zone_pgdat;
6374 
6375 		/* See comment about same check for global reclaim above */
6376 		if (zone->zone_pgdat == last_pgdat)
6377 			continue;
6378 		last_pgdat = zone->zone_pgdat;
6379 		shrink_node(zone->zone_pgdat, sc);
6380 	}
6381 
6382 	if (first_pgdat)
6383 		consider_reclaim_throttle(first_pgdat, sc);
6384 
6385 	/*
6386 	 * Restore to original mask to avoid the impact on the caller if we
6387 	 * promoted it to __GFP_HIGHMEM.
6388 	 */
6389 	sc->gfp_mask = orig_mask;
6390 }
6391 
snapshot_refaults(struct mem_cgroup * target_memcg,pg_data_t * pgdat)6392 static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat)
6393 {
6394 	struct lruvec *target_lruvec;
6395 	unsigned long refaults;
6396 
6397 	if (lru_gen_enabled())
6398 		return;
6399 
6400 	target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
6401 	refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON);
6402 	target_lruvec->refaults[WORKINGSET_ANON] = refaults;
6403 	refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE);
6404 	target_lruvec->refaults[WORKINGSET_FILE] = refaults;
6405 }
6406 
6407 /*
6408  * This is the main entry point to direct page reclaim.
6409  *
6410  * If a full scan of the inactive list fails to free enough memory then we
6411  * are "out of memory" and something needs to be killed.
6412  *
6413  * If the caller is !__GFP_FS then the probability of a failure is reasonably
6414  * high - the zone may be full of dirty or under-writeback pages, which this
6415  * caller can't do much about.  We kick the writeback threads and take explicit
6416  * naps in the hope that some of these pages can be written.  But if the
6417  * allocating task holds filesystem locks which prevent writeout this might not
6418  * work, and the allocation attempt will fail.
6419  *
6420  * returns:	0, if no pages reclaimed
6421  * 		else, the number of pages reclaimed
6422  */
do_try_to_free_pages(struct zonelist * zonelist,struct scan_control * sc)6423 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
6424 					  struct scan_control *sc)
6425 {
6426 	int initial_priority = sc->priority;
6427 	pg_data_t *last_pgdat;
6428 	struct zoneref *z;
6429 	struct zone *zone;
6430 retry:
6431 	delayacct_freepages_start();
6432 
6433 	if (!cgroup_reclaim(sc))
6434 		__count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
6435 
6436 	do {
6437 		if (!sc->proactive)
6438 			vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
6439 					sc->priority);
6440 		sc->nr_scanned = 0;
6441 		shrink_zones(zonelist, sc);
6442 
6443 		if (sc->nr_reclaimed >= sc->nr_to_reclaim)
6444 			break;
6445 
6446 		if (sc->compaction_ready)
6447 			break;
6448 
6449 		/*
6450 		 * If we're getting trouble reclaiming, start doing
6451 		 * writepage even in laptop mode.
6452 		 */
6453 		if (sc->priority < DEF_PRIORITY - 2)
6454 			sc->may_writepage = 1;
6455 	} while (--sc->priority >= 0);
6456 
6457 	last_pgdat = NULL;
6458 	for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
6459 					sc->nodemask) {
6460 		if (zone->zone_pgdat == last_pgdat)
6461 			continue;
6462 		last_pgdat = zone->zone_pgdat;
6463 
6464 		snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
6465 
6466 		if (cgroup_reclaim(sc)) {
6467 			struct lruvec *lruvec;
6468 
6469 			lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup,
6470 						   zone->zone_pgdat);
6471 			clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
6472 		}
6473 	}
6474 
6475 	delayacct_freepages_end();
6476 
6477 	if (sc->nr_reclaimed)
6478 		return sc->nr_reclaimed;
6479 
6480 	/* Aborted reclaim to try compaction? don't OOM, then */
6481 	if (sc->compaction_ready)
6482 		return 1;
6483 
6484 	/*
6485 	 * We make inactive:active ratio decisions based on the node's
6486 	 * composition of memory, but a restrictive reclaim_idx or a
6487 	 * memory.low cgroup setting can exempt large amounts of
6488 	 * memory from reclaim. Neither of which are very common, so
6489 	 * instead of doing costly eligibility calculations of the
6490 	 * entire cgroup subtree up front, we assume the estimates are
6491 	 * good, and retry with forcible deactivation if that fails.
6492 	 */
6493 	if (sc->skipped_deactivate) {
6494 		sc->priority = initial_priority;
6495 		sc->force_deactivate = 1;
6496 		sc->skipped_deactivate = 0;
6497 		goto retry;
6498 	}
6499 
6500 	/* Untapped cgroup reserves?  Don't OOM, retry. */
6501 	if (sc->memcg_low_skipped) {
6502 		sc->priority = initial_priority;
6503 		sc->force_deactivate = 0;
6504 		sc->memcg_low_reclaim = 1;
6505 		sc->memcg_low_skipped = 0;
6506 		goto retry;
6507 	}
6508 
6509 	return 0;
6510 }
6511 
allow_direct_reclaim(pg_data_t * pgdat)6512 static bool allow_direct_reclaim(pg_data_t *pgdat)
6513 {
6514 	struct zone *zone;
6515 	unsigned long pfmemalloc_reserve = 0;
6516 	unsigned long free_pages = 0;
6517 	int i;
6518 	bool wmark_ok;
6519 
6520 	if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6521 		return true;
6522 
6523 	for (i = 0; i <= ZONE_NORMAL; i++) {
6524 		zone = &pgdat->node_zones[i];
6525 		if (!managed_zone(zone))
6526 			continue;
6527 
6528 		if (!zone_reclaimable_pages(zone))
6529 			continue;
6530 
6531 		pfmemalloc_reserve += min_wmark_pages(zone);
6532 		free_pages += zone_page_state(zone, NR_FREE_PAGES);
6533 	}
6534 
6535 	/* If there are no reserves (unexpected config) then do not throttle */
6536 	if (!pfmemalloc_reserve)
6537 		return true;
6538 
6539 	wmark_ok = free_pages > pfmemalloc_reserve / 2;
6540 
6541 	/* kswapd must be awake if processes are being throttled */
6542 	if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
6543 		if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL)
6544 			WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL);
6545 
6546 		wake_up_interruptible(&pgdat->kswapd_wait);
6547 	}
6548 
6549 	return wmark_ok;
6550 }
6551 
6552 /*
6553  * Throttle direct reclaimers if backing storage is backed by the network
6554  * and the PFMEMALLOC reserve for the preferred node is getting dangerously
6555  * depleted. kswapd will continue to make progress and wake the processes
6556  * when the low watermark is reached.
6557  *
6558  * Returns true if a fatal signal was delivered during throttling. If this
6559  * happens, the page allocator should not consider triggering the OOM killer.
6560  */
throttle_direct_reclaim(gfp_t gfp_mask,struct zonelist * zonelist,nodemask_t * nodemask)6561 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
6562 					nodemask_t *nodemask)
6563 {
6564 	struct zoneref *z;
6565 	struct zone *zone;
6566 	pg_data_t *pgdat = NULL;
6567 
6568 	/*
6569 	 * Kernel threads should not be throttled as they may be indirectly
6570 	 * responsible for cleaning pages necessary for reclaim to make forward
6571 	 * progress. kjournald for example may enter direct reclaim while
6572 	 * committing a transaction where throttling it could forcing other
6573 	 * processes to block on log_wait_commit().
6574 	 */
6575 	if (current->flags & PF_KTHREAD)
6576 		goto out;
6577 
6578 	/*
6579 	 * If a fatal signal is pending, this process should not throttle.
6580 	 * It should return quickly so it can exit and free its memory
6581 	 */
6582 	if (fatal_signal_pending(current))
6583 		goto out;
6584 
6585 	/*
6586 	 * Check if the pfmemalloc reserves are ok by finding the first node
6587 	 * with a usable ZONE_NORMAL or lower zone. The expectation is that
6588 	 * GFP_KERNEL will be required for allocating network buffers when
6589 	 * swapping over the network so ZONE_HIGHMEM is unusable.
6590 	 *
6591 	 * Throttling is based on the first usable node and throttled processes
6592 	 * wait on a queue until kswapd makes progress and wakes them. There
6593 	 * is an affinity then between processes waking up and where reclaim
6594 	 * progress has been made assuming the process wakes on the same node.
6595 	 * More importantly, processes running on remote nodes will not compete
6596 	 * for remote pfmemalloc reserves and processes on different nodes
6597 	 * should make reasonable progress.
6598 	 */
6599 	for_each_zone_zonelist_nodemask(zone, z, zonelist,
6600 					gfp_zone(gfp_mask), nodemask) {
6601 		if (zone_idx(zone) > ZONE_NORMAL)
6602 			continue;
6603 
6604 		/* Throttle based on the first usable node */
6605 		pgdat = zone->zone_pgdat;
6606 		if (allow_direct_reclaim(pgdat))
6607 			goto out;
6608 		break;
6609 	}
6610 
6611 	/* If no zone was usable by the allocation flags then do not throttle */
6612 	if (!pgdat)
6613 		goto out;
6614 
6615 	/* Account for the throttling */
6616 	count_vm_event(PGSCAN_DIRECT_THROTTLE);
6617 
6618 	/*
6619 	 * If the caller cannot enter the filesystem, it's possible that it
6620 	 * is due to the caller holding an FS lock or performing a journal
6621 	 * transaction in the case of a filesystem like ext[3|4]. In this case,
6622 	 * it is not safe to block on pfmemalloc_wait as kswapd could be
6623 	 * blocked waiting on the same lock. Instead, throttle for up to a
6624 	 * second before continuing.
6625 	 */
6626 	if (!(gfp_mask & __GFP_FS))
6627 		wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
6628 			allow_direct_reclaim(pgdat), HZ);
6629 	else
6630 		/* Throttle until kswapd wakes the process */
6631 		wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
6632 			allow_direct_reclaim(pgdat));
6633 
6634 	if (fatal_signal_pending(current))
6635 		return true;
6636 
6637 out:
6638 	return false;
6639 }
6640 
try_to_free_pages(struct zonelist * zonelist,int order,gfp_t gfp_mask,nodemask_t * nodemask)6641 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
6642 				gfp_t gfp_mask, nodemask_t *nodemask)
6643 {
6644 	unsigned long nr_reclaimed;
6645 	struct scan_control sc = {
6646 		.nr_to_reclaim = SWAP_CLUSTER_MAX,
6647 		.gfp_mask = current_gfp_context(gfp_mask),
6648 		.reclaim_idx = gfp_zone(gfp_mask),
6649 		.order = order,
6650 		.nodemask = nodemask,
6651 		.priority = DEF_PRIORITY,
6652 		.may_writepage = !laptop_mode,
6653 		.may_unmap = 1,
6654 		.may_swap = 1,
6655 	};
6656 
6657 	/*
6658 	 * scan_control uses s8 fields for order, priority, and reclaim_idx.
6659 	 * Confirm they are large enough for max values.
6660 	 */
6661 	BUILD_BUG_ON(MAX_ORDER > S8_MAX);
6662 	BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
6663 	BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
6664 
6665 	/*
6666 	 * Do not enter reclaim if fatal signal was delivered while throttled.
6667 	 * 1 is returned so that the page allocator does not OOM kill at this
6668 	 * point.
6669 	 */
6670 	if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
6671 		return 1;
6672 
6673 	set_task_reclaim_state(current, &sc.reclaim_state);
6674 	trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
6675 
6676 	nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
6677 
6678 	trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
6679 	set_task_reclaim_state(current, NULL);
6680 
6681 	return nr_reclaimed;
6682 }
6683 
6684 #ifdef CONFIG_MEMCG
6685 
6686 /* Only used by soft limit reclaim. Do not reuse for anything else. */
mem_cgroup_shrink_node(struct mem_cgroup * memcg,gfp_t gfp_mask,bool noswap,pg_data_t * pgdat,unsigned long * nr_scanned)6687 unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
6688 						gfp_t gfp_mask, bool noswap,
6689 						pg_data_t *pgdat,
6690 						unsigned long *nr_scanned)
6691 {
6692 	struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat);
6693 	struct scan_control sc = {
6694 		.nr_to_reclaim = SWAP_CLUSTER_MAX,
6695 		.target_mem_cgroup = memcg,
6696 		.may_writepage = !laptop_mode,
6697 		.may_unmap = 1,
6698 		.reclaim_idx = MAX_NR_ZONES - 1,
6699 		.may_swap = !noswap,
6700 	};
6701 
6702 	WARN_ON_ONCE(!current->reclaim_state);
6703 
6704 	sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
6705 			(GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
6706 
6707 	trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
6708 						      sc.gfp_mask);
6709 
6710 	/*
6711 	 * NOTE: Although we can get the priority field, using it
6712 	 * here is not a good idea, since it limits the pages we can scan.
6713 	 * if we don't reclaim here, the shrink_node from balance_pgdat
6714 	 * will pick up pages from other mem cgroup's as well. We hack
6715 	 * the priority and make it zero.
6716 	 */
6717 	shrink_lruvec(lruvec, &sc);
6718 
6719 	trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
6720 
6721 	*nr_scanned = sc.nr_scanned;
6722 
6723 	return sc.nr_reclaimed;
6724 }
6725 
try_to_free_mem_cgroup_pages(struct mem_cgroup * memcg,unsigned long nr_pages,gfp_t gfp_mask,unsigned int reclaim_options)6726 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
6727 					   unsigned long nr_pages,
6728 					   gfp_t gfp_mask,
6729 					   unsigned int reclaim_options)
6730 {
6731 	unsigned long nr_reclaimed;
6732 	unsigned int noreclaim_flag;
6733 	struct scan_control sc = {
6734 		.nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
6735 		.gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
6736 				(GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
6737 		.reclaim_idx = MAX_NR_ZONES - 1,
6738 		.target_mem_cgroup = memcg,
6739 		.priority = DEF_PRIORITY,
6740 		.may_writepage = !laptop_mode,
6741 		.may_unmap = 1,
6742 		.may_swap = !!(reclaim_options & MEMCG_RECLAIM_MAY_SWAP),
6743 		.proactive = !!(reclaim_options & MEMCG_RECLAIM_PROACTIVE),
6744 	};
6745 	/*
6746 	 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
6747 	 * equal pressure on all the nodes. This is based on the assumption that
6748 	 * the reclaim does not bail out early.
6749 	 */
6750 	struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
6751 
6752 	set_task_reclaim_state(current, &sc.reclaim_state);
6753 	trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
6754 	noreclaim_flag = memalloc_noreclaim_save();
6755 
6756 	nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
6757 
6758 	memalloc_noreclaim_restore(noreclaim_flag);
6759 	trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
6760 	set_task_reclaim_state(current, NULL);
6761 
6762 	return nr_reclaimed;
6763 }
6764 #endif
6765 
kswapd_age_node(struct pglist_data * pgdat,struct scan_control * sc)6766 static void kswapd_age_node(struct pglist_data *pgdat, struct scan_control *sc)
6767 {
6768 	struct mem_cgroup *memcg;
6769 	struct lruvec *lruvec;
6770 
6771 	if (lru_gen_enabled()) {
6772 		lru_gen_age_node(pgdat, sc);
6773 		return;
6774 	}
6775 
6776 	if (!can_age_anon_pages(pgdat, sc))
6777 		return;
6778 
6779 	lruvec = mem_cgroup_lruvec(NULL, pgdat);
6780 	if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON))
6781 		return;
6782 
6783 	memcg = mem_cgroup_iter(NULL, NULL, NULL);
6784 	do {
6785 		lruvec = mem_cgroup_lruvec(memcg, pgdat);
6786 		shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
6787 				   sc, LRU_ACTIVE_ANON);
6788 		memcg = mem_cgroup_iter(NULL, memcg, NULL);
6789 	} while (memcg);
6790 }
6791 
pgdat_watermark_boosted(pg_data_t * pgdat,int highest_zoneidx)6792 static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx)
6793 {
6794 	int i;
6795 	struct zone *zone;
6796 
6797 	/*
6798 	 * Check for watermark boosts top-down as the higher zones
6799 	 * are more likely to be boosted. Both watermarks and boosts
6800 	 * should not be checked at the same time as reclaim would
6801 	 * start prematurely when there is no boosting and a lower
6802 	 * zone is balanced.
6803 	 */
6804 	for (i = highest_zoneidx; i >= 0; i--) {
6805 		zone = pgdat->node_zones + i;
6806 		if (!managed_zone(zone))
6807 			continue;
6808 
6809 		if (zone->watermark_boost)
6810 			return true;
6811 	}
6812 
6813 	return false;
6814 }
6815 
6816 /*
6817  * Returns true if there is an eligible zone balanced for the request order
6818  * and highest_zoneidx
6819  */
pgdat_balanced(pg_data_t * pgdat,int order,int highest_zoneidx)6820 static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
6821 {
6822 	int i;
6823 	unsigned long mark = -1;
6824 	struct zone *zone;
6825 
6826 	/*
6827 	 * Check watermarks bottom-up as lower zones are more likely to
6828 	 * meet watermarks.
6829 	 */
6830 	for (i = 0; i <= highest_zoneidx; i++) {
6831 		zone = pgdat->node_zones + i;
6832 
6833 		if (!managed_zone(zone))
6834 			continue;
6835 
6836 		if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)
6837 			mark = wmark_pages(zone, WMARK_PROMO);
6838 		else
6839 			mark = high_wmark_pages(zone);
6840 		if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx))
6841 			return true;
6842 	}
6843 
6844 	/*
6845 	 * If a node has no managed zone within highest_zoneidx, it does not
6846 	 * need balancing by definition. This can happen if a zone-restricted
6847 	 * allocation tries to wake a remote kswapd.
6848 	 */
6849 	if (mark == -1)
6850 		return true;
6851 
6852 	return false;
6853 }
6854 
6855 /* Clear pgdat state for congested, dirty or under writeback. */
clear_pgdat_congested(pg_data_t * pgdat)6856 static void clear_pgdat_congested(pg_data_t *pgdat)
6857 {
6858 	struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat);
6859 
6860 	clear_bit(LRUVEC_CONGESTED, &lruvec->flags);
6861 	clear_bit(PGDAT_DIRTY, &pgdat->flags);
6862 	clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
6863 }
6864 
6865 /*
6866  * Prepare kswapd for sleeping. This verifies that there are no processes
6867  * waiting in throttle_direct_reclaim() and that watermarks have been met.
6868  *
6869  * Returns true if kswapd is ready to sleep
6870  */
prepare_kswapd_sleep(pg_data_t * pgdat,int order,int highest_zoneidx)6871 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order,
6872 				int highest_zoneidx)
6873 {
6874 	/*
6875 	 * The throttled processes are normally woken up in balance_pgdat() as
6876 	 * soon as allow_direct_reclaim() is true. But there is a potential
6877 	 * race between when kswapd checks the watermarks and a process gets
6878 	 * throttled. There is also a potential race if processes get
6879 	 * throttled, kswapd wakes, a large process exits thereby balancing the
6880 	 * zones, which causes kswapd to exit balance_pgdat() before reaching
6881 	 * the wake up checks. If kswapd is going to sleep, no process should
6882 	 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
6883 	 * the wake up is premature, processes will wake kswapd and get
6884 	 * throttled again. The difference from wake ups in balance_pgdat() is
6885 	 * that here we are under prepare_to_wait().
6886 	 */
6887 	if (waitqueue_active(&pgdat->pfmemalloc_wait))
6888 		wake_up_all(&pgdat->pfmemalloc_wait);
6889 
6890 	/* Hopeless node, leave it to direct reclaim */
6891 	if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
6892 		return true;
6893 
6894 	if (pgdat_balanced(pgdat, order, highest_zoneidx)) {
6895 		clear_pgdat_congested(pgdat);
6896 		return true;
6897 	}
6898 
6899 	return false;
6900 }
6901 
6902 /*
6903  * kswapd shrinks a node of pages that are at or below the highest usable
6904  * zone that is currently unbalanced.
6905  *
6906  * Returns true if kswapd scanned at least the requested number of pages to
6907  * reclaim or if the lack of progress was due to pages under writeback.
6908  * This is used to determine if the scanning priority needs to be raised.
6909  */
kswapd_shrink_node(pg_data_t * pgdat,struct scan_control * sc)6910 static bool kswapd_shrink_node(pg_data_t *pgdat,
6911 			       struct scan_control *sc)
6912 {
6913 	struct zone *zone;
6914 	int z;
6915 
6916 	/* Reclaim a number of pages proportional to the number of zones */
6917 	sc->nr_to_reclaim = 0;
6918 	for (z = 0; z <= sc->reclaim_idx; z++) {
6919 		zone = pgdat->node_zones + z;
6920 		if (!managed_zone(zone))
6921 			continue;
6922 
6923 		sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
6924 	}
6925 
6926 	/*
6927 	 * Historically care was taken to put equal pressure on all zones but
6928 	 * now pressure is applied based on node LRU order.
6929 	 */
6930 	shrink_node(pgdat, sc);
6931 
6932 	/*
6933 	 * Fragmentation may mean that the system cannot be rebalanced for
6934 	 * high-order allocations. If twice the allocation size has been
6935 	 * reclaimed then recheck watermarks only at order-0 to prevent
6936 	 * excessive reclaim. Assume that a process requested a high-order
6937 	 * can direct reclaim/compact.
6938 	 */
6939 	if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
6940 		sc->order = 0;
6941 
6942 	return sc->nr_scanned >= sc->nr_to_reclaim;
6943 }
6944 
6945 /* Page allocator PCP high watermark is lowered if reclaim is active. */
6946 static inline void
update_reclaim_active(pg_data_t * pgdat,int highest_zoneidx,bool active)6947 update_reclaim_active(pg_data_t *pgdat, int highest_zoneidx, bool active)
6948 {
6949 	int i;
6950 	struct zone *zone;
6951 
6952 	for (i = 0; i <= highest_zoneidx; i++) {
6953 		zone = pgdat->node_zones + i;
6954 
6955 		if (!managed_zone(zone))
6956 			continue;
6957 
6958 		if (active)
6959 			set_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
6960 		else
6961 			clear_bit(ZONE_RECLAIM_ACTIVE, &zone->flags);
6962 	}
6963 }
6964 
6965 static inline void
set_reclaim_active(pg_data_t * pgdat,int highest_zoneidx)6966 set_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
6967 {
6968 	update_reclaim_active(pgdat, highest_zoneidx, true);
6969 }
6970 
6971 static inline void
clear_reclaim_active(pg_data_t * pgdat,int highest_zoneidx)6972 clear_reclaim_active(pg_data_t *pgdat, int highest_zoneidx)
6973 {
6974 	update_reclaim_active(pgdat, highest_zoneidx, false);
6975 }
6976 
6977 /*
6978  * For kswapd, balance_pgdat() will reclaim pages across a node from zones
6979  * that are eligible for use by the caller until at least one zone is
6980  * balanced.
6981  *
6982  * Returns the order kswapd finished reclaiming at.
6983  *
6984  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
6985  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
6986  * found to have free_pages <= high_wmark_pages(zone), any page in that zone
6987  * or lower is eligible for reclaim until at least one usable zone is
6988  * balanced.
6989  */
balance_pgdat(pg_data_t * pgdat,int order,int highest_zoneidx)6990 static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
6991 {
6992 	int i;
6993 	unsigned long nr_soft_reclaimed;
6994 	unsigned long nr_soft_scanned;
6995 	unsigned long pflags;
6996 	unsigned long nr_boost_reclaim;
6997 	unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
6998 	bool boosted;
6999 	struct zone *zone;
7000 	struct scan_control sc = {
7001 		.gfp_mask = GFP_KERNEL,
7002 		.order = order,
7003 		.may_unmap = 1,
7004 	};
7005 
7006 	set_task_reclaim_state(current, &sc.reclaim_state);
7007 	psi_memstall_enter(&pflags);
7008 	__fs_reclaim_acquire(_THIS_IP_);
7009 
7010 	count_vm_event(PAGEOUTRUN);
7011 
7012 	/*
7013 	 * Account for the reclaim boost. Note that the zone boost is left in
7014 	 * place so that parallel allocations that are near the watermark will
7015 	 * stall or direct reclaim until kswapd is finished.
7016 	 */
7017 	nr_boost_reclaim = 0;
7018 	for (i = 0; i <= highest_zoneidx; i++) {
7019 		zone = pgdat->node_zones + i;
7020 		if (!managed_zone(zone))
7021 			continue;
7022 
7023 		nr_boost_reclaim += zone->watermark_boost;
7024 		zone_boosts[i] = zone->watermark_boost;
7025 	}
7026 	boosted = nr_boost_reclaim;
7027 
7028 restart:
7029 	set_reclaim_active(pgdat, highest_zoneidx);
7030 	sc.priority = DEF_PRIORITY;
7031 	do {
7032 		unsigned long nr_reclaimed = sc.nr_reclaimed;
7033 		bool raise_priority = true;
7034 		bool balanced;
7035 		bool ret;
7036 
7037 		sc.reclaim_idx = highest_zoneidx;
7038 
7039 		/*
7040 		 * If the number of buffer_heads exceeds the maximum allowed
7041 		 * then consider reclaiming from all zones. This has a dual
7042 		 * purpose -- on 64-bit systems it is expected that
7043 		 * buffer_heads are stripped during active rotation. On 32-bit
7044 		 * systems, highmem pages can pin lowmem memory and shrinking
7045 		 * buffers can relieve lowmem pressure. Reclaim may still not
7046 		 * go ahead if all eligible zones for the original allocation
7047 		 * request are balanced to avoid excessive reclaim from kswapd.
7048 		 */
7049 		if (buffer_heads_over_limit) {
7050 			for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
7051 				zone = pgdat->node_zones + i;
7052 				if (!managed_zone(zone))
7053 					continue;
7054 
7055 				sc.reclaim_idx = i;
7056 				break;
7057 			}
7058 		}
7059 
7060 		/*
7061 		 * If the pgdat is imbalanced then ignore boosting and preserve
7062 		 * the watermarks for a later time and restart. Note that the
7063 		 * zone watermarks will be still reset at the end of balancing
7064 		 * on the grounds that the normal reclaim should be enough to
7065 		 * re-evaluate if boosting is required when kswapd next wakes.
7066 		 */
7067 		balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx);
7068 		if (!balanced && nr_boost_reclaim) {
7069 			nr_boost_reclaim = 0;
7070 			goto restart;
7071 		}
7072 
7073 		/*
7074 		 * If boosting is not active then only reclaim if there are no
7075 		 * eligible zones. Note that sc.reclaim_idx is not used as
7076 		 * buffer_heads_over_limit may have adjusted it.
7077 		 */
7078 		if (!nr_boost_reclaim && balanced)
7079 			goto out;
7080 
7081 		/* Limit the priority of boosting to avoid reclaim writeback */
7082 		if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
7083 			raise_priority = false;
7084 
7085 		/*
7086 		 * Do not writeback or swap pages for boosted reclaim. The
7087 		 * intent is to relieve pressure not issue sub-optimal IO
7088 		 * from reclaim context. If no pages are reclaimed, the
7089 		 * reclaim will be aborted.
7090 		 */
7091 		sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
7092 		sc.may_swap = !nr_boost_reclaim;
7093 
7094 		/*
7095 		 * Do some background aging, to give pages a chance to be
7096 		 * referenced before reclaiming. All pages are rotated
7097 		 * regardless of classzone as this is about consistent aging.
7098 		 */
7099 		kswapd_age_node(pgdat, &sc);
7100 
7101 		/*
7102 		 * If we're getting trouble reclaiming, start doing writepage
7103 		 * even in laptop mode.
7104 		 */
7105 		if (sc.priority < DEF_PRIORITY - 2)
7106 			sc.may_writepage = 1;
7107 
7108 		/* Call soft limit reclaim before calling shrink_node. */
7109 		sc.nr_scanned = 0;
7110 		nr_soft_scanned = 0;
7111 		nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
7112 						sc.gfp_mask, &nr_soft_scanned);
7113 		sc.nr_reclaimed += nr_soft_reclaimed;
7114 
7115 		/*
7116 		 * There should be no need to raise the scanning priority if
7117 		 * enough pages are already being scanned that that high
7118 		 * watermark would be met at 100% efficiency.
7119 		 */
7120 		if (kswapd_shrink_node(pgdat, &sc))
7121 			raise_priority = false;
7122 
7123 		/*
7124 		 * If the low watermark is met there is no need for processes
7125 		 * to be throttled on pfmemalloc_wait as they should not be
7126 		 * able to safely make forward progress. Wake them
7127 		 */
7128 		if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
7129 				allow_direct_reclaim(pgdat))
7130 			wake_up_all(&pgdat->pfmemalloc_wait);
7131 
7132 		/* Check if kswapd should be suspending */
7133 		__fs_reclaim_release(_THIS_IP_);
7134 		ret = try_to_freeze();
7135 		__fs_reclaim_acquire(_THIS_IP_);
7136 		if (ret || kthread_should_stop())
7137 			break;
7138 
7139 		/*
7140 		 * Raise priority if scanning rate is too low or there was no
7141 		 * progress in reclaiming pages
7142 		 */
7143 		nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
7144 		nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
7145 
7146 		/*
7147 		 * If reclaim made no progress for a boost, stop reclaim as
7148 		 * IO cannot be queued and it could be an infinite loop in
7149 		 * extreme circumstances.
7150 		 */
7151 		if (nr_boost_reclaim && !nr_reclaimed)
7152 			break;
7153 
7154 		if (raise_priority || !nr_reclaimed)
7155 			sc.priority--;
7156 	} while (sc.priority >= 1);
7157 
7158 	if (!sc.nr_reclaimed)
7159 		pgdat->kswapd_failures++;
7160 
7161 out:
7162 	clear_reclaim_active(pgdat, highest_zoneidx);
7163 
7164 	/* If reclaim was boosted, account for the reclaim done in this pass */
7165 	if (boosted) {
7166 		unsigned long flags;
7167 
7168 		for (i = 0; i <= highest_zoneidx; i++) {
7169 			if (!zone_boosts[i])
7170 				continue;
7171 
7172 			/* Increments are under the zone lock */
7173 			zone = pgdat->node_zones + i;
7174 			spin_lock_irqsave(&zone->lock, flags);
7175 			zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
7176 			spin_unlock_irqrestore(&zone->lock, flags);
7177 		}
7178 
7179 		/*
7180 		 * As there is now likely space, wakeup kcompact to defragment
7181 		 * pageblocks.
7182 		 */
7183 		wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx);
7184 	}
7185 
7186 	snapshot_refaults(NULL, pgdat);
7187 	__fs_reclaim_release(_THIS_IP_);
7188 	psi_memstall_leave(&pflags);
7189 	set_task_reclaim_state(current, NULL);
7190 
7191 	/*
7192 	 * Return the order kswapd stopped reclaiming at as
7193 	 * prepare_kswapd_sleep() takes it into account. If another caller
7194 	 * entered the allocator slow path while kswapd was awake, order will
7195 	 * remain at the higher level.
7196 	 */
7197 	return sc.order;
7198 }
7199 
7200 /*
7201  * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to
7202  * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is
7203  * not a valid index then either kswapd runs for first time or kswapd couldn't
7204  * sleep after previous reclaim attempt (node is still unbalanced). In that
7205  * case return the zone index of the previous kswapd reclaim cycle.
7206  */
kswapd_highest_zoneidx(pg_data_t * pgdat,enum zone_type prev_highest_zoneidx)7207 static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat,
7208 					   enum zone_type prev_highest_zoneidx)
7209 {
7210 	enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
7211 
7212 	return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx;
7213 }
7214 
kswapd_try_to_sleep(pg_data_t * pgdat,int alloc_order,int reclaim_order,unsigned int highest_zoneidx)7215 static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
7216 				unsigned int highest_zoneidx)
7217 {
7218 	long remaining = 0;
7219 	DEFINE_WAIT(wait);
7220 
7221 	if (freezing(current) || kthread_should_stop())
7222 		return;
7223 
7224 	prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
7225 
7226 	/*
7227 	 * Try to sleep for a short interval. Note that kcompactd will only be
7228 	 * woken if it is possible to sleep for a short interval. This is
7229 	 * deliberate on the assumption that if reclaim cannot keep an
7230 	 * eligible zone balanced that it's also unlikely that compaction will
7231 	 * succeed.
7232 	 */
7233 	if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
7234 		/*
7235 		 * Compaction records what page blocks it recently failed to
7236 		 * isolate pages from and skips them in the future scanning.
7237 		 * When kswapd is going to sleep, it is reasonable to assume
7238 		 * that pages and compaction may succeed so reset the cache.
7239 		 */
7240 		reset_isolation_suitable(pgdat);
7241 
7242 		/*
7243 		 * We have freed the memory, now we should compact it to make
7244 		 * allocation of the requested order possible.
7245 		 */
7246 		wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx);
7247 
7248 		remaining = schedule_timeout(HZ/10);
7249 
7250 		/*
7251 		 * If woken prematurely then reset kswapd_highest_zoneidx and
7252 		 * order. The values will either be from a wakeup request or
7253 		 * the previous request that slept prematurely.
7254 		 */
7255 		if (remaining) {
7256 			WRITE_ONCE(pgdat->kswapd_highest_zoneidx,
7257 					kswapd_highest_zoneidx(pgdat,
7258 							highest_zoneidx));
7259 
7260 			if (READ_ONCE(pgdat->kswapd_order) < reclaim_order)
7261 				WRITE_ONCE(pgdat->kswapd_order, reclaim_order);
7262 		}
7263 
7264 		finish_wait(&pgdat->kswapd_wait, &wait);
7265 		prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
7266 	}
7267 
7268 	/*
7269 	 * After a short sleep, check if it was a premature sleep. If not, then
7270 	 * go fully to sleep until explicitly woken up.
7271 	 */
7272 	if (!remaining &&
7273 	    prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) {
7274 		trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
7275 
7276 		/*
7277 		 * vmstat counters are not perfectly accurate and the estimated
7278 		 * value for counters such as NR_FREE_PAGES can deviate from the
7279 		 * true value by nr_online_cpus * threshold. To avoid the zone
7280 		 * watermarks being breached while under pressure, we reduce the
7281 		 * per-cpu vmstat threshold while kswapd is awake and restore
7282 		 * them before going back to sleep.
7283 		 */
7284 		set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
7285 
7286 		if (!kthread_should_stop())
7287 			schedule();
7288 
7289 		set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
7290 	} else {
7291 		if (remaining)
7292 			count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
7293 		else
7294 			count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
7295 	}
7296 	finish_wait(&pgdat->kswapd_wait, &wait);
7297 }
7298 
7299 /*
7300  * The background pageout daemon, started as a kernel thread
7301  * from the init process.
7302  *
7303  * This basically trickles out pages so that we have _some_
7304  * free memory available even if there is no other activity
7305  * that frees anything up. This is needed for things like routing
7306  * etc, where we otherwise might have all activity going on in
7307  * asynchronous contexts that cannot page things out.
7308  *
7309  * If there are applications that are active memory-allocators
7310  * (most normal use), this basically shouldn't matter.
7311  */
kswapd(void * p)7312 static int kswapd(void *p)
7313 {
7314 	unsigned int alloc_order, reclaim_order;
7315 	unsigned int highest_zoneidx = MAX_NR_ZONES - 1;
7316 	pg_data_t *pgdat = (pg_data_t *)p;
7317 	struct task_struct *tsk = current;
7318 	const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
7319 
7320 	if (!cpumask_empty(cpumask))
7321 		set_cpus_allowed_ptr(tsk, cpumask);
7322 
7323 	/*
7324 	 * Tell the memory management that we're a "memory allocator",
7325 	 * and that if we need more memory we should get access to it
7326 	 * regardless (see "__alloc_pages()"). "kswapd" should
7327 	 * never get caught in the normal page freeing logic.
7328 	 *
7329 	 * (Kswapd normally doesn't need memory anyway, but sometimes
7330 	 * you need a small amount of memory in order to be able to
7331 	 * page out something else, and this flag essentially protects
7332 	 * us from recursively trying to free more memory as we're
7333 	 * trying to free the first piece of memory in the first place).
7334 	 */
7335 	tsk->flags |= PF_MEMALLOC | PF_KSWAPD;
7336 	set_freezable();
7337 
7338 	WRITE_ONCE(pgdat->kswapd_order, 0);
7339 	WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7340 	atomic_set(&pgdat->nr_writeback_throttled, 0);
7341 	for ( ; ; ) {
7342 		bool ret;
7343 
7344 		alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order);
7345 		highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7346 							highest_zoneidx);
7347 
7348 kswapd_try_sleep:
7349 		kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
7350 					highest_zoneidx);
7351 
7352 		/* Read the new order and highest_zoneidx */
7353 		alloc_order = READ_ONCE(pgdat->kswapd_order);
7354 		highest_zoneidx = kswapd_highest_zoneidx(pgdat,
7355 							highest_zoneidx);
7356 		WRITE_ONCE(pgdat->kswapd_order, 0);
7357 		WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES);
7358 
7359 		ret = try_to_freeze();
7360 		if (kthread_should_stop())
7361 			break;
7362 
7363 		/*
7364 		 * We can speed up thawing tasks if we don't call balance_pgdat
7365 		 * after returning from the refrigerator
7366 		 */
7367 		if (ret)
7368 			continue;
7369 
7370 		/*
7371 		 * Reclaim begins at the requested order but if a high-order
7372 		 * reclaim fails then kswapd falls back to reclaiming for
7373 		 * order-0. If that happens, kswapd will consider sleeping
7374 		 * for the order it finished reclaiming at (reclaim_order)
7375 		 * but kcompactd is woken to compact for the original
7376 		 * request (alloc_order).
7377 		 */
7378 		trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx,
7379 						alloc_order);
7380 		reclaim_order = balance_pgdat(pgdat, alloc_order,
7381 						highest_zoneidx);
7382 		if (reclaim_order < alloc_order)
7383 			goto kswapd_try_sleep;
7384 	}
7385 
7386 	tsk->flags &= ~(PF_MEMALLOC | PF_KSWAPD);
7387 
7388 	return 0;
7389 }
7390 
7391 /*
7392  * A zone is low on free memory or too fragmented for high-order memory.  If
7393  * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
7394  * pgdat.  It will wake up kcompactd after reclaiming memory.  If kswapd reclaim
7395  * has failed or is not needed, still wake up kcompactd if only compaction is
7396  * needed.
7397  */
wakeup_kswapd(struct zone * zone,gfp_t gfp_flags,int order,enum zone_type highest_zoneidx)7398 void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
7399 		   enum zone_type highest_zoneidx)
7400 {
7401 	pg_data_t *pgdat;
7402 	enum zone_type curr_idx;
7403 
7404 	if (!managed_zone(zone))
7405 		return;
7406 
7407 	if (!cpuset_zone_allowed(zone, gfp_flags))
7408 		return;
7409 
7410 	pgdat = zone->zone_pgdat;
7411 	curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx);
7412 
7413 	if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx)
7414 		WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx);
7415 
7416 	if (READ_ONCE(pgdat->kswapd_order) < order)
7417 		WRITE_ONCE(pgdat->kswapd_order, order);
7418 
7419 	if (!waitqueue_active(&pgdat->kswapd_wait))
7420 		return;
7421 
7422 	/* Hopeless node, leave it to direct reclaim if possible */
7423 	if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
7424 	    (pgdat_balanced(pgdat, order, highest_zoneidx) &&
7425 	     !pgdat_watermark_boosted(pgdat, highest_zoneidx))) {
7426 		/*
7427 		 * There may be plenty of free memory available, but it's too
7428 		 * fragmented for high-order allocations.  Wake up kcompactd
7429 		 * and rely on compaction_suitable() to determine if it's
7430 		 * needed.  If it fails, it will defer subsequent attempts to
7431 		 * ratelimit its work.
7432 		 */
7433 		if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
7434 			wakeup_kcompactd(pgdat, order, highest_zoneidx);
7435 		return;
7436 	}
7437 
7438 	trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order,
7439 				      gfp_flags);
7440 	wake_up_interruptible(&pgdat->kswapd_wait);
7441 }
7442 
7443 #ifdef CONFIG_HIBERNATION
7444 /*
7445  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
7446  * freed pages.
7447  *
7448  * Rather than trying to age LRUs the aim is to preserve the overall
7449  * LRU order by reclaiming preferentially
7450  * inactive > active > active referenced > active mapped
7451  */
shrink_all_memory(unsigned long nr_to_reclaim)7452 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
7453 {
7454 	struct scan_control sc = {
7455 		.nr_to_reclaim = nr_to_reclaim,
7456 		.gfp_mask = GFP_HIGHUSER_MOVABLE,
7457 		.reclaim_idx = MAX_NR_ZONES - 1,
7458 		.priority = DEF_PRIORITY,
7459 		.may_writepage = 1,
7460 		.may_unmap = 1,
7461 		.may_swap = 1,
7462 		.hibernation_mode = 1,
7463 	};
7464 	struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
7465 	unsigned long nr_reclaimed;
7466 	unsigned int noreclaim_flag;
7467 
7468 	fs_reclaim_acquire(sc.gfp_mask);
7469 	noreclaim_flag = memalloc_noreclaim_save();
7470 	set_task_reclaim_state(current, &sc.reclaim_state);
7471 
7472 	nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
7473 
7474 	set_task_reclaim_state(current, NULL);
7475 	memalloc_noreclaim_restore(noreclaim_flag);
7476 	fs_reclaim_release(sc.gfp_mask);
7477 
7478 	return nr_reclaimed;
7479 }
7480 #endif /* CONFIG_HIBERNATION */
7481 
7482 /*
7483  * This kswapd start function will be called by init and node-hot-add.
7484  */
kswapd_run(int nid)7485 void kswapd_run(int nid)
7486 {
7487 	pg_data_t *pgdat = NODE_DATA(nid);
7488 
7489 	pgdat_kswapd_lock(pgdat);
7490 	if (!pgdat->kswapd) {
7491 		pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
7492 		if (IS_ERR(pgdat->kswapd)) {
7493 			/* failure at boot is fatal */
7494 			BUG_ON(system_state < SYSTEM_RUNNING);
7495 			pr_err("Failed to start kswapd on node %d\n", nid);
7496 			pgdat->kswapd = NULL;
7497 		}
7498 	}
7499 	pgdat_kswapd_unlock(pgdat);
7500 }
7501 
7502 /*
7503  * Called by memory hotplug when all memory in a node is offlined.  Caller must
7504  * be holding mem_hotplug_begin/done().
7505  */
kswapd_stop(int nid)7506 void kswapd_stop(int nid)
7507 {
7508 	pg_data_t *pgdat = NODE_DATA(nid);
7509 	struct task_struct *kswapd;
7510 
7511 	pgdat_kswapd_lock(pgdat);
7512 	kswapd = pgdat->kswapd;
7513 	if (kswapd) {
7514 		kthread_stop(kswapd);
7515 		pgdat->kswapd = NULL;
7516 	}
7517 	pgdat_kswapd_unlock(pgdat);
7518 }
7519 
kswapd_init(void)7520 static int __init kswapd_init(void)
7521 {
7522 	int nid;
7523 
7524 	swap_setup();
7525 	for_each_node_state(nid, N_MEMORY)
7526  		kswapd_run(nid);
7527 	return 0;
7528 }
7529 
7530 module_init(kswapd_init)
7531 
7532 #ifdef CONFIG_NUMA
7533 /*
7534  * Node reclaim mode
7535  *
7536  * If non-zero call node_reclaim when the number of free pages falls below
7537  * the watermarks.
7538  */
7539 int node_reclaim_mode __read_mostly;
7540 
7541 /*
7542  * Priority for NODE_RECLAIM. This determines the fraction of pages
7543  * of a node considered for each zone_reclaim. 4 scans 1/16th of
7544  * a zone.
7545  */
7546 #define NODE_RECLAIM_PRIORITY 4
7547 
7548 /*
7549  * Percentage of pages in a zone that must be unmapped for node_reclaim to
7550  * occur.
7551  */
7552 int sysctl_min_unmapped_ratio = 1;
7553 
7554 /*
7555  * If the number of slab pages in a zone grows beyond this percentage then
7556  * slab reclaim needs to occur.
7557  */
7558 int sysctl_min_slab_ratio = 5;
7559 
node_unmapped_file_pages(struct pglist_data * pgdat)7560 static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
7561 {
7562 	unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
7563 	unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
7564 		node_page_state(pgdat, NR_ACTIVE_FILE);
7565 
7566 	/*
7567 	 * It's possible for there to be more file mapped pages than
7568 	 * accounted for by the pages on the file LRU lists because
7569 	 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
7570 	 */
7571 	return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
7572 }
7573 
7574 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
node_pagecache_reclaimable(struct pglist_data * pgdat)7575 static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
7576 {
7577 	unsigned long nr_pagecache_reclaimable;
7578 	unsigned long delta = 0;
7579 
7580 	/*
7581 	 * If RECLAIM_UNMAP is set, then all file pages are considered
7582 	 * potentially reclaimable. Otherwise, we have to worry about
7583 	 * pages like swapcache and node_unmapped_file_pages() provides
7584 	 * a better estimate
7585 	 */
7586 	if (node_reclaim_mode & RECLAIM_UNMAP)
7587 		nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
7588 	else
7589 		nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
7590 
7591 	/* If we can't clean pages, remove dirty pages from consideration */
7592 	if (!(node_reclaim_mode & RECLAIM_WRITE))
7593 		delta += node_page_state(pgdat, NR_FILE_DIRTY);
7594 
7595 	/* Watch for any possible underflows due to delta */
7596 	if (unlikely(delta > nr_pagecache_reclaimable))
7597 		delta = nr_pagecache_reclaimable;
7598 
7599 	return nr_pagecache_reclaimable - delta;
7600 }
7601 
7602 /*
7603  * Try to free up some pages from this node through reclaim.
7604  */
__node_reclaim(struct pglist_data * pgdat,gfp_t gfp_mask,unsigned int order)7605 static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
7606 {
7607 	/* Minimum pages needed in order to stay on node */
7608 	const unsigned long nr_pages = 1 << order;
7609 	struct task_struct *p = current;
7610 	unsigned int noreclaim_flag;
7611 	struct scan_control sc = {
7612 		.nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
7613 		.gfp_mask = current_gfp_context(gfp_mask),
7614 		.order = order,
7615 		.priority = NODE_RECLAIM_PRIORITY,
7616 		.may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
7617 		.may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
7618 		.may_swap = 1,
7619 		.reclaim_idx = gfp_zone(gfp_mask),
7620 	};
7621 	unsigned long pflags;
7622 
7623 	trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
7624 					   sc.gfp_mask);
7625 
7626 	cond_resched();
7627 	psi_memstall_enter(&pflags);
7628 	fs_reclaim_acquire(sc.gfp_mask);
7629 	/*
7630 	 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
7631 	 */
7632 	noreclaim_flag = memalloc_noreclaim_save();
7633 	set_task_reclaim_state(p, &sc.reclaim_state);
7634 
7635 	if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages ||
7636 	    node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) > pgdat->min_slab_pages) {
7637 		/*
7638 		 * Free memory by calling shrink node with increasing
7639 		 * priorities until we have enough memory freed.
7640 		 */
7641 		do {
7642 			shrink_node(pgdat, &sc);
7643 		} while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
7644 	}
7645 
7646 	set_task_reclaim_state(p, NULL);
7647 	memalloc_noreclaim_restore(noreclaim_flag);
7648 	fs_reclaim_release(sc.gfp_mask);
7649 	psi_memstall_leave(&pflags);
7650 
7651 	trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
7652 
7653 	return sc.nr_reclaimed >= nr_pages;
7654 }
7655 
node_reclaim(struct pglist_data * pgdat,gfp_t gfp_mask,unsigned int order)7656 int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
7657 {
7658 	int ret;
7659 
7660 	/*
7661 	 * Node reclaim reclaims unmapped file backed pages and
7662 	 * slab pages if we are over the defined limits.
7663 	 *
7664 	 * A small portion of unmapped file backed pages is needed for
7665 	 * file I/O otherwise pages read by file I/O will be immediately
7666 	 * thrown out if the node is overallocated. So we do not reclaim
7667 	 * if less than a specified percentage of the node is used by
7668 	 * unmapped file backed pages.
7669 	 */
7670 	if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
7671 	    node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <=
7672 	    pgdat->min_slab_pages)
7673 		return NODE_RECLAIM_FULL;
7674 
7675 	/*
7676 	 * Do not scan if the allocation should not be delayed.
7677 	 */
7678 	if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
7679 		return NODE_RECLAIM_NOSCAN;
7680 
7681 	/*
7682 	 * Only run node reclaim on the local node or on nodes that do not
7683 	 * have associated processors. This will favor the local processor
7684 	 * over remote processors and spread off node memory allocations
7685 	 * as wide as possible.
7686 	 */
7687 	if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
7688 		return NODE_RECLAIM_NOSCAN;
7689 
7690 	if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
7691 		return NODE_RECLAIM_NOSCAN;
7692 
7693 	ret = __node_reclaim(pgdat, gfp_mask, order);
7694 	clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
7695 
7696 	if (!ret)
7697 		count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
7698 
7699 	return ret;
7700 }
7701 #endif
7702 
check_move_unevictable_pages(struct pagevec * pvec)7703 void check_move_unevictable_pages(struct pagevec *pvec)
7704 {
7705 	struct folio_batch fbatch;
7706 	unsigned i;
7707 
7708 	folio_batch_init(&fbatch);
7709 	for (i = 0; i < pvec->nr; i++) {
7710 		struct page *page = pvec->pages[i];
7711 
7712 		if (PageTransTail(page))
7713 			continue;
7714 		folio_batch_add(&fbatch, page_folio(page));
7715 	}
7716 	check_move_unevictable_folios(&fbatch);
7717 }
7718 EXPORT_SYMBOL_GPL(check_move_unevictable_pages);
7719 
7720 /**
7721  * check_move_unevictable_folios - Move evictable folios to appropriate zone
7722  * lru list
7723  * @fbatch: Batch of lru folios to check.
7724  *
7725  * Checks folios for evictability, if an evictable folio is in the unevictable
7726  * lru list, moves it to the appropriate evictable lru list. This function
7727  * should be only used for lru folios.
7728  */
check_move_unevictable_folios(struct folio_batch * fbatch)7729 void check_move_unevictable_folios(struct folio_batch *fbatch)
7730 {
7731 	struct lruvec *lruvec = NULL;
7732 	int pgscanned = 0;
7733 	int pgrescued = 0;
7734 	int i;
7735 
7736 	for (i = 0; i < fbatch->nr; i++) {
7737 		struct folio *folio = fbatch->folios[i];
7738 		int nr_pages = folio_nr_pages(folio);
7739 
7740 		pgscanned += nr_pages;
7741 
7742 		/* block memcg migration while the folio moves between lrus */
7743 		if (!folio_test_clear_lru(folio))
7744 			continue;
7745 
7746 		lruvec = folio_lruvec_relock_irq(folio, lruvec);
7747 		if (folio_evictable(folio) && folio_test_unevictable(folio)) {
7748 			lruvec_del_folio(lruvec, folio);
7749 			folio_clear_unevictable(folio);
7750 			lruvec_add_folio(lruvec, folio);
7751 			pgrescued += nr_pages;
7752 		}
7753 		folio_set_lru(folio);
7754 	}
7755 
7756 	if (lruvec) {
7757 		__count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
7758 		__count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
7759 		unlock_page_lruvec_irq(lruvec);
7760 	} else if (pgscanned) {
7761 		count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
7762 	}
7763 }
7764 EXPORT_SYMBOL_GPL(check_move_unevictable_folios);
7765