forked from ~ljy/RK356X_SDK_RELEASE
blame | history | patch | commit | commitdiff
hc
2024-10-09 244b2c5ca8b14627e4a17755e5922221e121c771 
 kernel/mm/workingset.c
....@@ -6,6 +6,7 @@
66  */
77 
88 #include <linux/memcontrol.h>
9
+#include <linux/mm_inline.h>
910 #include <linux/writeback.h>
1011 #include <linux/shmem_fs.h>
1112 #include <linux/pagemap.h>
....@@ -156,24 +157,24 @@
156157  *
157158  *		Implementation
158159  *
159
- * For each node's file LRU lists, a counter for inactive evictions
160
- * and activations is maintained (node->inactive_age).
160
+ * For each node's LRU lists, a counter for inactive evictions and
161
+ * activations is maintained (node->nonresident_age).
161162  *
162163  * On eviction, a snapshot of this counter (along with some bits to
163
- * identify the node) is stored in the now empty page cache radix tree
164
+ * identify the node) is stored in the now empty page cache
164165  * slot of the evicted page.  This is called a shadow entry.
165166  *
166167  * On cache misses for which there are shadow entries, an eligible
167168  * refault distance will immediately activate the refaulting page.
168169  */
169170 
170
-#define EVICTION_SHIFT	(RADIX_TREE_EXCEPTIONAL_ENTRY + \
171
+#define EVICTION_SHIFT	((BITS_PER_LONG - BITS_PER_XA_VALUE) +	\
171172 			 1 + NODES_SHIFT + MEM_CGROUP_ID_SHIFT)
172173 #define EVICTION_MASK	(~0UL >> EVICTION_SHIFT)
173174 
174175 /*
175176  * Eviction timestamps need to be able to cover the full range of
176
- * actionable refaults. However, bits are tight in the radix tree
177
+ * actionable refaults. However, bits are tight in the xarray
177178  * entry, and after storing the identifier for the lruvec there might
178179  * not be enough left to represent every single actionable refault. In
179180  * that case, we have to sacrifice granularity for distance, and group
....@@ -185,22 +186,21 @@
185186 			 bool workingset)
186187 {
187188 	eviction >>= bucket_order;
189
+	eviction &= EVICTION_MASK;
188190 	eviction = (eviction << MEM_CGROUP_ID_SHIFT) | memcgid;
189191 	eviction = (eviction << NODES_SHIFT) | pgdat->node_id;
190192 	eviction = (eviction << 1) | workingset;
191
-	eviction = (eviction << RADIX_TREE_EXCEPTIONAL_SHIFT);
192193 
193
-	return (void *)(eviction | RADIX_TREE_EXCEPTIONAL_ENTRY);
194
+	return xa_mk_value(eviction);
194195 }
195196 
196197 static void unpack_shadow(void *shadow, int *memcgidp, pg_data_t **pgdat,
197198 			  unsigned long *evictionp, bool *workingsetp)
198199 {
199
-	unsigned long entry = (unsigned long)shadow;
200
+	unsigned long entry = xa_to_value(shadow);
200201 	int memcgid, nid;
201202 	bool workingset;
202203 
203
-	entry >>= RADIX_TREE_EXCEPTIONAL_SHIFT;
204204 	workingset = entry & 1;
205205 	entry >>= 1;
206206 	nid = entry & ((1UL << NODES_SHIFT) - 1);
....@@ -215,28 +215,58 @@
215215 }
216216 
217217 /**
218
+ * workingset_age_nonresident - age non-resident entries as LRU ages
219
+ * @lruvec: the lruvec that was aged
220
+ * @nr_pages: the number of pages to count
221
+ *
222
+ * As in-memory pages are aged, non-resident pages need to be aged as
223
+ * well, in order for the refault distances later on to be comparable
224
+ * to the in-memory dimensions. This function allows reclaim and LRU
225
+ * operations to drive the non-resident aging along in parallel.
226
+ */
227
+void workingset_age_nonresident(struct lruvec *lruvec, unsigned long nr_pages)
228
+{
229
+	/*
230
+	 * Reclaiming a cgroup means reclaiming all its children in a
231
+	 * round-robin fashion. That means that each cgroup has an LRU
232
+	 * order that is composed of the LRU orders of its child
233
+	 * cgroups; and every page has an LRU position not just in the
234
+	 * cgroup that owns it, but in all of that group's ancestors.
235
+	 *
236
+	 * So when the physical inactive list of a leaf cgroup ages,
237
+	 * the virtual inactive lists of all its parents, including
238
+	 * the root cgroup's, age as well.
239
+	 */
240
+	do {
241
+		atomic_long_add(nr_pages, &lruvec->nonresident_age);
242
+	} while ((lruvec = parent_lruvec(lruvec)));
243
+}
244
+
245
+/**
218246  * workingset_eviction - note the eviction of a page from memory
219
- * @mapping: address space the page was backing
247
+ * @target_memcg: the cgroup that is causing the reclaim
220248  * @page: the page being evicted
221249  *
222
- * Returns a shadow entry to be stored in @mapping->i_pages in place
250
+ * Returns a shadow entry to be stored in @page->mapping->i_pages in place
223251  * of the evicted @page so that a later refault can be detected.
224252  */
225
-void *workingset_eviction(struct address_space *mapping, struct page *page)
253
+void *workingset_eviction(struct page *page, struct mem_cgroup *target_memcg)
226254 {
227255 	struct pglist_data *pgdat = page_pgdat(page);
228
-	struct mem_cgroup *memcg = page_memcg(page);
229
-	int memcgid = mem_cgroup_id(memcg);
230256 	unsigned long eviction;
231257 	struct lruvec *lruvec;
258
+	int memcgid;
232259 
233260 	/* Page is fully exclusive and pins page->mem_cgroup */
234261 	VM_BUG_ON_PAGE(PageLRU(page), page);
235262 	VM_BUG_ON_PAGE(page_count(page), page);
236263 	VM_BUG_ON_PAGE(!PageLocked(page), page);
237264 
238
-	lruvec = mem_cgroup_lruvec(pgdat, memcg);
239
-	eviction = atomic_long_inc_return(&lruvec->inactive_age);
265
+	lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
266
+	workingset_age_nonresident(lruvec, thp_nr_pages(page));
267
+	/* XXX: target_memcg can be NULL, go through lruvec */
268
+	memcgid = mem_cgroup_id(lruvec_memcg(lruvec));
269
+	eviction = atomic_long_read(&lruvec->nonresident_age);
240270 	return pack_shadow(memcgid, pgdat, eviction, PageWorkingset(page));
241271 }
242272 
....@@ -246,13 +276,17 @@
246276  * @shadow: shadow entry of the evicted page
247277  *
248278  * Calculates and evaluates the refault distance of the previously
249
- * evicted page in the context of the node it was allocated in.
279
+ * evicted page in the context of the node and the memcg whose memory
280
+ * pressure caused the eviction.
250281  */
251282 void workingset_refault(struct page *page, void *shadow)
252283 {
284
+	bool file = page_is_file_lru(page);
285
+	struct mem_cgroup *eviction_memcg;
286
+	struct lruvec *eviction_lruvec;
253287 	unsigned long refault_distance;
288
+	unsigned long workingset_size;
254289 	struct pglist_data *pgdat;
255
-	unsigned long active_file;
256290 	struct mem_cgroup *memcg;
257291 	unsigned long eviction;
258292 	struct lruvec *lruvec;
....@@ -279,49 +313,78 @@
279313 	 * would be better if the root_mem_cgroup existed in all
280314 	 * configurations instead.
281315 	 */
282
-	memcg = mem_cgroup_from_id(memcgid);
283
-	if (!mem_cgroup_disabled() && !memcg)
316
+	eviction_memcg = mem_cgroup_from_id(memcgid);
317
+	if (!mem_cgroup_disabled() && !eviction_memcg)
284318 		goto out;
285
-	lruvec = mem_cgroup_lruvec(pgdat, memcg);
286
-	refault = atomic_long_read(&lruvec->inactive_age);
287
-	active_file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES);
319
+	eviction_lruvec = mem_cgroup_lruvec(eviction_memcg, pgdat);
320
+	refault = atomic_long_read(&eviction_lruvec->nonresident_age);
288321 
289322 	/*
290323 	 * Calculate the refault distance
291324 	 *
292325 	 * The unsigned subtraction here gives an accurate distance
293
-	 * across inactive_age overflows in most cases. There is a
326
+	 * across nonresident_age overflows in most cases. There is a
294327 	 * special case: usually, shadow entries have a short lifetime
295328 	 * and are either refaulted or reclaimed along with the inode
296329 	 * before they get too old.  But it is not impossible for the
297
-	 * inactive_age to lap a shadow entry in the field, which can
298
-	 * then result in a false small refault distance, leading to a
299
-	 * false activation should this old entry actually refault
300
-	 * again.  However, earlier kernels used to deactivate
330
+	 * nonresident_age to lap a shadow entry in the field, which
331
+	 * can then result in a false small refault distance, leading
332
+	 * to a false activation should this old entry actually
333
+	 * refault again.  However, earlier kernels used to deactivate
301334 	 * unconditionally with *every* reclaim invocation for the
302335 	 * longest time, so the occasional inappropriate activation
303336 	 * leading to pressure on the active list is not a problem.
304337 	 */
305338 	refault_distance = (refault - eviction) & EVICTION_MASK;
306339 
307
-	inc_lruvec_state(lruvec, WORKINGSET_REFAULT);
340
+	/*
341
+	 * The activation decision for this page is made at the level
342
+	 * where the eviction occurred, as that is where the LRU order
343
+	 * during page reclaim is being determined.
344
+	 *
345
+	 * However, the cgroup that will own the page is the one that
346
+	 * is actually experiencing the refault event.
347
+	 */
348
+	memcg = page_memcg(page);
349
+	lruvec = mem_cgroup_lruvec(memcg, pgdat);
350
+
351
+	inc_lruvec_state(lruvec, WORKINGSET_REFAULT_BASE + file);
308352 
309353 	/*
310354 	 * Compare the distance to the existing workingset size. We
311
-	 * don't act on pages that couldn't stay resident even if all
312
-	 * the memory was available to the page cache.
355
+	 * don't activate pages that couldn't stay resident even if
356
+	 * all the memory was available to the workingset. Whether
357
+	 * workingset competition needs to consider anon or not depends
358
+	 * on having swap.
313359 	 */
314
-	if (refault_distance > active_file)
360
+	workingset_size = lruvec_page_state(eviction_lruvec, NR_ACTIVE_FILE);
361
+	if (!file) {
362
+		workingset_size += lruvec_page_state(eviction_lruvec,
363
+						     NR_INACTIVE_FILE);
364
+	}
365
+	if (mem_cgroup_get_nr_swap_pages(memcg) > 0) {
366
+		workingset_size += lruvec_page_state(eviction_lruvec,
367
+						     NR_ACTIVE_ANON);
368
+		if (file) {
369
+			workingset_size += lruvec_page_state(eviction_lruvec,
370
+						     NR_INACTIVE_ANON);
371
+		}
372
+	}
373
+	if (refault_distance > workingset_size)
315374 		goto out;
316375 
317376 	SetPageActive(page);
318
-	atomic_long_inc(&lruvec->inactive_age);
319
-	inc_lruvec_state(lruvec, WORKINGSET_ACTIVATE);
377
+	workingset_age_nonresident(lruvec, thp_nr_pages(page));
378
+	inc_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + file);
320379 
321380 	/* Page was active prior to eviction */
322381 	if (workingset) {
323382 		SetPageWorkingset(page);
324
-		inc_lruvec_state(lruvec, WORKINGSET_RESTORE);
383
+		/* XXX: Move to lru_cache_add() when it supports new vs putback */
384
+		spin_lock_irq(&page_pgdat(page)->lru_lock);
385
+		lru_note_cost_page(page);
386
+		spin_unlock_irq(&page_pgdat(page)->lru_lock);
387
+		inc_lruvec_state(lruvec, WORKINGSET_RESTORE_BASE + file);
325388 	}
326389 out:
327390 	rcu_read_unlock();
....@@ -347,8 +410,8 @@
347410 	memcg = page_memcg_rcu(page);
348411 	if (!mem_cgroup_disabled() && !memcg)
349412 		goto out;
350
-	lruvec = mem_cgroup_lruvec(page_pgdat(page), memcg);
351
-	atomic_long_inc(&lruvec->inactive_age);
413
+	lruvec = mem_cgroup_page_lruvec(page, page_pgdat(page));
414
+	workingset_age_nonresident(lruvec, thp_nr_pages(page));
352415 out:
353416 	rcu_read_unlock();
354417 }
....@@ -367,7 +430,7 @@
367430 
368431 static struct list_lru shadow_nodes;
369432 
370
-void workingset_update_node(struct radix_tree_node *node)
433
+void workingset_update_node(struct xa_node *node)
371434 {
372435 	/*
373436 	 * Track non-empty nodes that contain only shadow entries;
....@@ -377,12 +440,18 @@
377440 	 * already where they should be. The list_empty() test is safe
378441 	 * as node->private_list is protected by the i_pages lock.
379442 	 */
380
-	if (node->count && node->count == node->exceptional) {
381
-		if (list_empty(&node->private_list))
443
+	VM_WARN_ON_ONCE(!irqs_disabled());  /* For __inc_lruvec_page_state */
444
+
445
+	if (node->count && node->count == node->nr_values) {
446
+		if (list_empty(&node->private_list)) {
382447 			list_lru_add(&shadow_nodes, &node->private_list);
448
+			__inc_lruvec_slab_state(node, WORKINGSET_NODES);
449
+		}
383450 	} else {
384
-		if (!list_empty(&node->private_list))
451
+		if (!list_empty(&node->private_list)) {
385452 			list_lru_del(&shadow_nodes, &node->private_list);
453
+			__dec_lruvec_slab_state(node, WORKINGSET_NODES);
454
+		}
386455 	}
387456 }
388457 
....@@ -391,12 +460,12 @@
391460 {
392461 	unsigned long max_nodes;
393462 	unsigned long nodes;
394
-	unsigned long cache;
463
+	unsigned long pages;
395464 
396465 	nodes = list_lru_shrink_count(&shadow_nodes, sc);
397466 
398467 	/*
399
-	 * Approximate a reasonable limit for the radix tree nodes
468
+	 * Approximate a reasonable limit for the nodes
400469 	 * containing shadow entries. We don't need to keep more
401470 	 * shadow entries than possible pages on the active list,
402471 	 * since refault distances bigger than that are dismissed.
....@@ -411,20 +480,30 @@
411480 	 * worst-case density of 1/8th. Below that, not all eligible
412481 	 * refaults can be detected anymore.
413482 	 *
414
-	 * On 64-bit with 7 radix_tree_nodes per page and 64 slots
483
+	 * On 64-bit with 7 xa_nodes per page and 64 slots
415484 	 * each, this will reclaim shadow entries when they consume
416485 	 * ~1.8% of available memory:
417486 	 *
418
-	 * PAGE_SIZE / radix_tree_nodes / node_entries * 8 / PAGE_SIZE
487
+	 * PAGE_SIZE / xa_nodes / node_entries * 8 / PAGE_SIZE
419488 	 */
489
+#ifdef CONFIG_MEMCG
420490 	if (sc->memcg) {
421
-		cache = mem_cgroup_node_nr_lru_pages(sc->memcg, sc->nid,
422
-						     LRU_ALL_FILE);
423
-	} else {
424
-		cache = node_page_state(NODE_DATA(sc->nid), NR_ACTIVE_FILE) +
425
-			node_page_state(NODE_DATA(sc->nid), NR_INACTIVE_FILE);
426
-	}
427
-	max_nodes = cache >> (RADIX_TREE_MAP_SHIFT - 3);
491
+		struct lruvec *lruvec;
492
+		int i;
493
+
494
+		lruvec = mem_cgroup_lruvec(sc->memcg, NODE_DATA(sc->nid));
495
+		for (pages = 0, i = 0; i < NR_LRU_LISTS; i++)
496
+			pages += lruvec_page_state_local(lruvec,
497
+							 NR_LRU_BASE + i);
498
+		pages += lruvec_page_state_local(
499
+			lruvec, NR_SLAB_RECLAIMABLE_B) >> PAGE_SHIFT;
500
+		pages += lruvec_page_state_local(
501
+			lruvec, NR_SLAB_UNRECLAIMABLE_B) >> PAGE_SHIFT;
502
+	} else
503
+#endif
504
+		pages = node_present_pages(sc->nid);
505
+
506
+	max_nodes = pages >> (XA_CHUNK_SHIFT - 3);
428507 
429508 	if (!nodes)
430509 		return SHRINK_EMPTY;
....@@ -437,27 +516,25 @@
437516 static enum lru_status shadow_lru_isolate(struct list_head *item,
438517 					  struct list_lru_one *lru,
439518 					  spinlock_t *lru_lock,
440
-					  void *arg)
519
+					  void *arg) __must_hold(lru_lock)
441520 {
521
+	struct xa_node *node = container_of(item, struct xa_node, private_list);
442522 	struct address_space *mapping;
443
-	struct radix_tree_node *node;
444
-	unsigned int i;
445523 	int ret;
446524 
447525 	/*
448
-	 * Page cache insertions and deletions synchroneously maintain
526
+	 * Page cache insertions and deletions synchronously maintain
449527 	 * the shadow node LRU under the i_pages lock and the
450528 	 * lru_lock.  Because the page cache tree is emptied before
451529 	 * the inode can be destroyed, holding the lru_lock pins any
452
-	 * address_space that has radix tree nodes on the LRU.
530
+	 * address_space that has nodes on the LRU.
453531 	 *
454532 	 * We can then safely transition to the i_pages lock to
455533 	 * pin only the address_space of the particular node we want
456534 	 * to reclaim, take the node off-LRU, and drop the lru_lock.
457535 	 */
458536 
459
-	node = container_of(item, struct radix_tree_node, private_list);
460
-	mapping = container_of(node->root, struct address_space, i_pages);
537
+	mapping = container_of(node->array, struct address_space, i_pages);
461538 
462539 	/* Coming from the list, invert the lock order */
463540 	if (!xa_trylock(&mapping->i_pages)) {
....@@ -467,6 +544,8 @@
467544 	}
468545 
469546 	list_lru_isolate(lru, item);
547
+	__dec_lruvec_slab_state(node, WORKINGSET_NODES);
548
+
470549 	spin_unlock(lru_lock);
471550 
472551 	/*
....@@ -474,29 +553,13 @@
474553 	 * no pages, so we expect to be able to remove them all and
475554 	 * delete and free the empty node afterwards.
476555 	 */
477
-	if (WARN_ON_ONCE(!node->exceptional))
556
+	if (WARN_ON_ONCE(!node->nr_values))
478557 		goto out_invalid;
479
-	if (WARN_ON_ONCE(node->count != node->exceptional))
558
+	if (WARN_ON_ONCE(node->count != node->nr_values))
480559 		goto out_invalid;
481
-	for (i = 0; i < RADIX_TREE_MAP_SIZE; i++) {
482
-		if (node->slots[i]) {
483
-			if (WARN_ON_ONCE(!radix_tree_exceptional_entry(node->slots[i])))
484
-				goto out_invalid;
485
-			if (WARN_ON_ONCE(!node->exceptional))
486
-				goto out_invalid;
487
-			if (WARN_ON_ONCE(!mapping->nrexceptional))
488
-				goto out_invalid;
489
-			node->slots[i] = NULL;
490
-			node->exceptional--;
491
-			node->count--;
492
-			mapping->nrexceptional--;
493
-		}
494
-	}
495
-	if (WARN_ON_ONCE(node->exceptional))
496
-		goto out_invalid;
497
-	inc_lruvec_page_state(virt_to_page(node), WORKINGSET_NODERECLAIM);
498
-	__radix_tree_delete_node(&mapping->i_pages, node,
499
-				 workingset_lookup_update(mapping));
560
+	mapping->nrexceptional -= node->nr_values;
561
+	xa_delete_node(node, workingset_update_node);
562
+	__inc_lruvec_slab_state(node, WORKINGSET_NODERECLAIM);
500563 
501564 out_invalid:
502565 	xa_unlock_irq(&mapping->i_pages);
....@@ -518,7 +581,7 @@
518581 static struct shrinker workingset_shadow_shrinker = {
519582 	.count_objects = count_shadow_nodes,
520583 	.scan_objects = scan_shadow_nodes,
521
-	.seeks = DEFAULT_SEEKS,
584
+	.seeks = 0, /* ->count reports only fully expendable nodes */
522585 	.flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE,
523586 };
524587 
....@@ -543,7 +606,7 @@
543606 	 * double the initial memory by using totalram_pages as-is.
544607 	 */
545608 	timestamp_bits = BITS_PER_LONG - EVICTION_SHIFT;
546
-	max_order = fls_long(totalram_pages - 1);
609
+	max_order = fls_long(totalram_pages() - 1);
547610 	if (max_order > timestamp_bits)
548611 		bucket_order = max_order - timestamp_bits;
549612 	pr_info("workingset: timestamp_bits=%d max_order=%d bucket_order=%u\n",