~hc/RK356X_SDK_RELEASE.git

..	..	@@ -45,10 +45,33 @@
45	45	#define CREATE_TRACE_POINTS
46	46	#include <trace/events/compaction.h>
47	47
	48	+#undef CREATE_TRACE_POINTS
	49	+#ifndef __GENKSYMS__
	50	+#include <trace/hooks/mm.h>
	51	+#endif
	52	+
48	53	#define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order))
49	54	#define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order))
50	55	#define pageblock_start_pfn(pfn) block_start_pfn(pfn, pageblock_order)
51	56	#define pageblock_end_pfn(pfn) block_end_pfn(pfn, pageblock_order)
	57	+
	58	+/*
	59	+ * Fragmentation score check interval for proactive compaction purposes.
	60	+ */
	61	+static const unsigned int HPAGE_FRAG_CHECK_INTERVAL_MSEC = 500;
	62	+
	63	+/*
	64	+ * Page order with-respect-to which proactive compaction
	65	+ * calculates external fragmentation, which is used as
	66	+ * the "fragmentation score" of a node/zone.
	67	+ */
	68	+#if defined CONFIG_TRANSPARENT_HUGEPAGE
	69	+#define COMPACTION_HPAGE_ORDER HPAGE_PMD_ORDER
	70	+#elif defined CONFIG_HUGETLBFS
	71	+#define COMPACTION_HPAGE_ORDER HUGETLB_PAGE_ORDER
	72	+#else
	73	+#define COMPACTION_HPAGE_ORDER (PMD_SHIFT - PAGE_SHIFT)
	74	+#endif
52	75
53	76	static unsigned long release_freepages(struct list_head *freelist)
54	77	{
..	..	@@ -66,7 +89,7 @@
66	89	return high_pfn;
67	90	}
68	91
69		-static void map_pages(struct list_head *list)
	92	+static void split_map_pages(struct list_head *list)
70	93	{
71	94	unsigned int i, order, nr_pages;
72	95	struct page page, next;
..	..	@@ -136,7 +159,7 @@
136	159
137	160	/*
138	161	* Compaction is deferred when compaction fails to result in a page
139		- * allocation success. 1 << compact_defer_limit compactions are skipped up
	162	+ * allocation success. 1 << compact_defer_shift, compactions are skipped up
140	163	* to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
141	164	*/
142	165	void defer_compaction(struct zone *zone, int order)
..	..	@@ -162,11 +185,10 @@
162	185	return false;
163	186
164	187	/* Avoid possible overflow */
165		- if (++zone->compact_considered > defer_limit)
	188	+ if (++zone->compact_considered >= defer_limit) {
166	189	zone->compact_considered = defer_limit;
167		-
168		- if (zone->compact_considered >= defer_limit)
169	190	return false;
	191	+ }
170	192
171	193	trace_mm_compaction_deferred(zone, order);
172	194
..	..	@@ -237,6 +259,78 @@
237	259	return false;
238	260	}
239	261
	262	+static bool
	263	+__reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source,
	264	+ bool check_target)
	265	+{
	266	+ struct page *page = pfn_to_online_page(pfn);
	267	+ struct page *block_page;
	268	+ struct page *end_page;
	269	+ unsigned long block_pfn;
	270	+
	271	+ if (!page)
	272	+ return false;
	273	+ if (zone != page_zone(page))
	274	+ return false;
	275	+ if (pageblock_skip_persistent(page))
	276	+ return false;
	277	+
	278	+ /*
	279	+ * If skip is already cleared do no further checking once the
	280	+ * restart points have been set.
	281	+ */
	282	+ if (check_source && check_target && !get_pageblock_skip(page))
	283	+ return true;
	284	+
	285	+ /*
	286	+ * If clearing skip for the target scanner, do not select a
	287	+ * non-movable pageblock as the starting point.
	288	+ */
	289	+ if (!check_source && check_target &&
	290	+ get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
	291	+ return false;
	292	+
	293	+ /* Ensure the start of the pageblock or zone is online and valid */
	294	+ block_pfn = pageblock_start_pfn(pfn);
	295	+ block_pfn = max(block_pfn, zone->zone_start_pfn);
	296	+ block_page = pfn_to_online_page(block_pfn);
	297	+ if (block_page) {
	298	+ page = block_page;
	299	+ pfn = block_pfn;
	300	+ }
	301	+
	302	+ /* Ensure the end of the pageblock or zone is online and valid */
	303	+ block_pfn = pageblock_end_pfn(pfn) - 1;
	304	+ block_pfn = min(block_pfn, zone_end_pfn(zone) - 1);
	305	+ end_page = pfn_to_online_page(block_pfn);
	306	+ if (!end_page)
	307	+ return false;
	308	+
	309	+ /*
	310	+ * Only clear the hint if a sample indicates there is either a
	311	+ * free page or an LRU page in the block. One or other condition
	312	+ * is necessary for the block to be a migration source/target.
	313	+ */
	314	+ do {
	315	+ if (pfn_valid_within(pfn)) {
	316	+ if (check_source && PageLRU(page)) {
	317	+ clear_pageblock_skip(page);
	318	+ return true;
	319	+ }
	320	+
	321	+ if (check_target && PageBuddy(page)) {
	322	+ clear_pageblock_skip(page);
	323	+ return true;
	324	+ }
	325	+ }
	326	+
	327	+ page += (1 << PAGE_ALLOC_COSTLY_ORDER);
	328	+ pfn += (1 << PAGE_ALLOC_COSTLY_ORDER);
	329	+ } while (page <= end_page);
	330	+
	331	+ return false;
	332	+}
	333	+
240	334	/*
241	335	* This function is called to clear all cached information on pageblocks that
242	336	* should be skipped for page isolation when the migrate and free page scanner
..	..	@@ -244,30 +338,54 @@
244	338	*/
245	339	static void __reset_isolation_suitable(struct zone *zone)
246	340	{
247		- unsigned long start_pfn = zone->zone_start_pfn;
248		- unsigned long end_pfn = zone_end_pfn(zone);
249		- unsigned long pfn;
	341	+ unsigned long migrate_pfn = zone->zone_start_pfn;
	342	+ unsigned long free_pfn = zone_end_pfn(zone) - 1;
	343	+ unsigned long reset_migrate = free_pfn;
	344	+ unsigned long reset_free = migrate_pfn;
	345	+ bool source_set = false;
	346	+ bool free_set = false;
	347	+
	348	+ if (!zone->compact_blockskip_flush)
	349	+ return;
250	350
251	351	zone->compact_blockskip_flush = false;
252	352
253		- /* Walk the zone and mark every pageblock as suitable for isolation */
254		- for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
255		- struct page *page;
256		-
	353	+ /*
	354	+ * Walk the zone and update pageblock skip information. Source looks
	355	+ * for PageLRU while target looks for PageBuddy. When the scanner
	356	+ * is found, both PageBuddy and PageLRU are checked as the pageblock
	357	+ * is suitable as both source and target.
	358	+ */
	359	+ for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages,
	360	+ free_pfn -= pageblock_nr_pages) {
257	361	cond_resched();
258	362
259		- page = pfn_to_online_page(pfn);
260		- if (!page)
261		- continue;
262		- if (zone != page_zone(page))
263		- continue;
264		- if (pageblock_skip_persistent(page))
265		- continue;
	363	+ /* Update the migrate PFN */
	364	+ if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) &&
	365	+ migrate_pfn < reset_migrate) {
	366	+ source_set = true;
	367	+ reset_migrate = migrate_pfn;
	368	+ zone->compact_init_migrate_pfn = reset_migrate;
	369	+ zone->compact_cached_migrate_pfn[0] = reset_migrate;
	370	+ zone->compact_cached_migrate_pfn[1] = reset_migrate;
	371	+ }
266	372
267		- clear_pageblock_skip(page);
	373	+ /* Update the free PFN */
	374	+ if (__reset_isolation_pfn(zone, free_pfn, free_set, true) &&
	375	+ free_pfn > reset_free) {
	376	+ free_set = true;
	377	+ reset_free = free_pfn;
	378	+ zone->compact_init_free_pfn = reset_free;
	379	+ zone->compact_cached_free_pfn = reset_free;
	380	+ }
268	381	}
269	382
270		- reset_cached_positions(zone);
	383	+ /* Leave no distance if no suitable block was reset */
	384	+ if (reset_migrate >= reset_free) {
	385	+ zone->compact_cached_migrate_pfn[0] = migrate_pfn;
	386	+ zone->compact_cached_migrate_pfn[1] = migrate_pfn;
	387	+ zone->compact_cached_free_pfn = free_pfn;
	388	+ }
271	389	}
272	390
273	391	void reset_isolation_suitable(pg_data_t *pgdat)
..	..	@@ -286,15 +404,53 @@
286	404	}
287	405
288	406	/*
	407	+ * Sets the pageblock skip bit if it was clear. Note that this is a hint as
	408	+ * locks are not required for read/writers. Returns true if it was already set.
	409	+ */
	410	+static bool test_and_set_skip(struct compact_control cc, struct page page,
	411	+ unsigned long pfn)
	412	+{
	413	+ bool skip;
	414	+
	415	+ /* Do no update if skip hint is being ignored */
	416	+ if (cc->ignore_skip_hint)
	417	+ return false;
	418	+
	419	+ if (!IS_ALIGNED(pfn, pageblock_nr_pages))
	420	+ return false;
	421	+
	422	+ skip = get_pageblock_skip(page);
	423	+ if (!skip && !cc->no_set_skip_hint)
	424	+ set_pageblock_skip(page);
	425	+
	426	+ return skip;
	427	+}
	428	+
	429	+static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
	430	+{
	431	+ struct zone *zone = cc->zone;
	432	+
	433	+ pfn = pageblock_end_pfn(pfn);
	434	+
	435	+ /* Set for isolation rather than compaction */
	436	+ if (cc->no_set_skip_hint)
	437	+ return;
	438	+
	439	+ if (pfn > zone->compact_cached_migrate_pfn[0])
	440	+ zone->compact_cached_migrate_pfn[0] = pfn;
	441	+ if (cc->mode != MIGRATE_ASYNC &&
	442	+ pfn > zone->compact_cached_migrate_pfn[1])
	443	+ zone->compact_cached_migrate_pfn[1] = pfn;
	444	+}
	445	+
	446	+/*
289	447	* If no pages were isolated then mark this pageblock to be skipped in the
290	448	* future. The information is later cleared by __reset_isolation_suitable().
291	449	*/
292	450	static void update_pageblock_skip(struct compact_control *cc,
293		- struct page *page, unsigned long nr_isolated,
294		- bool migrate_scanner)
	451	+ struct page *page, unsigned long pfn)
295	452	{
296	453	struct zone *zone = cc->zone;
297		- unsigned long pfn;
298	454
299	455	if (cc->no_set_skip_hint)
300	456	return;
..	..	@@ -302,24 +458,11 @@
302	458	if (!page)
303	459	return;
304	460
305		- if (nr_isolated)
306		- return;
307		-
308	461	set_pageblock_skip(page);
309	462
310		- pfn = page_to_pfn(page);
311		-
312	463	/* Update where async and sync compaction should restart */
313		- if (migrate_scanner) {
314		- if (pfn > zone->compact_cached_migrate_pfn[0])
315		- zone->compact_cached_migrate_pfn[0] = pfn;
316		- if (cc->mode != MIGRATE_ASYNC &&
317		- pfn > zone->compact_cached_migrate_pfn[1])
318		- zone->compact_cached_migrate_pfn[1] = pfn;
319		- } else {
320		- if (pfn < zone->compact_cached_free_pfn)
321		- zone->compact_cached_free_pfn = pfn;
322		- }
	464	+ if (pfn < zone->compact_cached_free_pfn)
	465	+ zone->compact_cached_free_pfn = pfn;
323	466	}
324	467	#else
325	468	static inline bool isolation_suitable(struct compact_control *cc,
..	..	@@ -334,32 +477,43 @@
334	477	}
335	478
336	479	static inline void update_pageblock_skip(struct compact_control *cc,
337		- struct page *page, unsigned long nr_isolated,
338		- bool migrate_scanner)
	480	+ struct page *page, unsigned long pfn)
339	481	{
	482	+}
	483	+
	484	+static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
	485	+{
	486	+}
	487	+
	488	+static bool test_and_set_skip(struct compact_control cc, struct page page,
	489	+ unsigned long pfn)
	490	+{
	491	+ return false;
340	492	}
341	493	#endif /* CONFIG_COMPACTION */
342	494
343	495	/*
344	496	* Compaction requires the taking of some coarse locks that are potentially
345		- * very heavily contended. For async compaction, back out if the lock cannot
346		- * be taken immediately. For sync compaction, spin on the lock if needed.
	497	+ * very heavily contended. For async compaction, trylock and record if the
	498	+ * lock is contended. The lock will still be acquired but compaction will
	499	+ * abort when the current block is finished regardless of success rate.
	500	+ * Sync compaction acquires the lock.
347	501	*
348		- * Returns true if the lock is held
349		- * Returns false if the lock is not held and compaction should abort
	502	+ * Always returns true which makes it easier to track lock state in callers.
350	503	*/
351		-static bool compact_trylock_irqsave(spinlock_t lock, unsigned long flags,
	504	+static bool compact_lock_irqsave(spinlock_t lock, unsigned long flags,
352	505	struct compact_control *cc)
	506	+ __acquires(lock)
353	507	{
354		- if (cc->mode == MIGRATE_ASYNC) {
355		- if (!spin_trylock_irqsave(lock, *flags)) {
356		- cc->contended = true;
357		- return false;
358		- }
359		- } else {
360		- spin_lock_irqsave(lock, *flags);
	508	+ /* Track if the lock is contended in async mode */
	509	+ if (cc->mode == MIGRATE_ASYNC && !cc->contended) {
	510	+ if (spin_trylock_irqsave(lock, *flags))
	511	+ return true;
	512	+
	513	+ cc->contended = true;
361	514	}
362	515
	516	+ spin_lock_irqsave(lock, *flags);
363	517	return true;
364	518	}
365	519
..	..	@@ -391,37 +545,7 @@
391	545	return true;
392	546	}
393	547
394		- if (need_resched()) {
395		- if (cc->mode == MIGRATE_ASYNC) {
396		- cc->contended = true;
397		- return true;
398		- }
399		- cond_resched();
400		- }
401		-
402		- return false;
403		-}
404		-
405		-/*
406		- * Aside from avoiding lock contention, compaction also periodically checks
407		- * need_resched() and either schedules in sync compaction or aborts async
408		- * compaction. This is similar to what compact_unlock_should_abort() does, but
409		- * is used where no lock is concerned.
410		- *
411		- * Returns false when no scheduling was needed, or sync compaction scheduled.
412		- * Returns true when async compaction should abort.
413		- */
414		-static inline bool compact_should_abort(struct compact_control *cc)
415		-{
416		- /* async compaction aborts if contended */
417		- if (need_resched()) {
418		- if (cc->mode == MIGRATE_ASYNC) {
419		- cc->contended = true;
420		- return true;
421		- }
422		-
423		- cond_resched();
424		- }
	548	+ cond_resched();
425	549
426	550	return false;
427	551	}
..	..	@@ -435,19 +559,24 @@
435	559	unsigned long *start_pfn,
436	560	unsigned long end_pfn,
437	561	struct list_head *freelist,
	562	+ unsigned int stride,
438	563	bool strict)
439	564	{
440	565	int nr_scanned = 0, total_isolated = 0;
441		- struct page cursor, valid_page = NULL;
	566	+ struct page *cursor;
442	567	unsigned long flags = 0;
443	568	bool locked = false;
444	569	unsigned long blockpfn = *start_pfn;
445	570	unsigned int order;
446	571
	572	+ /* Strict mode is for isolation, speed is secondary */
	573	+ if (strict)
	574	+ stride = 1;
	575	+
447	576	cursor = pfn_to_page(blockpfn);
448	577
449	578	/* Isolate free pages. */
450		- for (; blockpfn < end_pfn; blockpfn++, cursor++) {
	579	+ for (; blockpfn < end_pfn; blockpfn += stride, cursor += stride) {
451	580	int isolated;
452	581	struct page *page = cursor;
453	582
..	..	@@ -464,9 +593,6 @@
464	593	nr_scanned++;
465	594	if (!pfn_valid_within(blockpfn))
466	595	goto isolate_fail;
467		-
468		- if (!valid_page)
469		- valid_page = page;
470	596
471	597	/*
472	598	* For compound pages such as THP and hugetlbfs, we can save
..	..	@@ -495,18 +621,8 @@
495	621	* recheck as well.
496	622	*/
497	623	if (!locked) {
498		- /*
499		- * The zone lock must be held to isolate freepages.
500		- * Unfortunately this is a very coarse lock and can be
501		- * heavily contended if there are parallel allocations
502		- * or parallel compactions. For async compaction do not
503		- * spin on the lock and we acquire the lock as late as
504		- * possible.
505		- */
506		- locked = compact_trylock_irqsave(&cc->zone->lock,
	624	+ locked = compact_lock_irqsave(&cc->zone->lock,
507	625	&flags, cc);
508		- if (!locked)
509		- break;
510	626
511	627	/* Recheck this is a buddy page under lock */
512	628	if (!PageBuddy(page))
..	..	@@ -514,7 +630,7 @@
514	630	}
515	631
516	632	/* Found a free page, will break it into order-0 pages */
517		- order = page_order(page);
	633	+ order = buddy_order(page);
518	634	isolated = __isolate_free_page(page, order);
519	635	if (!isolated)
520	636	break;
..	..	@@ -564,10 +680,6 @@
564	680	*/
565	681	if (strict && blockpfn < end_pfn)
566	682	total_isolated = 0;
567		-
568		- /* Update the pageblock-skip if the whole pageblock was scanned */
569		- if (blockpfn == end_pfn)
570		- update_pageblock_skip(cc, valid_page, total_isolated, false);
571	683
572	684	cc->total_free_scanned += nr_scanned;
573	685	if (total_isolated)
..	..	@@ -626,7 +738,7 @@
626	738	break;
627	739
628	740	isolated = isolate_freepages_block(cc, &isolate_start_pfn,
629		- block_end_pfn, &freelist, true);
	741	+ block_end_pfn, &freelist, 0, true);
630	742
631	743	/*
632	744	* In strict mode, isolate_freepages_block() returns 0 if
..	..	@@ -644,7 +756,7 @@
644	756	}
645	757
646	758	/* __isolate_free_page() does not map the pages */
647		- map_pages(&freelist);
	759	+ split_map_pages(&freelist);
648	760
649	761	if (pfn < end_pfn) {
650	762	/* Loop terminated early, cleanup. */
..	..	@@ -656,17 +768,42 @@
656	768	return pfn;
657	769	}
658	770
	771	+#ifdef CONFIG_COMPACTION
	772	+unsigned long isolate_and_split_free_page(struct page *page,
	773	+ struct list_head *list)
	774	+{
	775	+ unsigned long isolated;
	776	+ unsigned int order;
	777	+
	778	+ if (!PageBuddy(page))
	779	+ return 0;
	780	+
	781	+ order = buddy_order(page);
	782	+ isolated = __isolate_free_page(page, order);
	783	+ if (!isolated)
	784	+ return 0;
	785	+
	786	+ set_page_private(page, order);
	787	+ list_add(&page->lru, list);
	788	+
	789	+ split_map_pages(list);
	790	+
	791	+ return isolated;
	792	+}
	793	+EXPORT_SYMBOL_GPL(isolate_and_split_free_page);
	794	+#endif
	795	+
659	796	/* Similar to reclaim, but different enough that they don't share logic */
660		-static bool too_many_isolated(struct zone *zone)
	797	+static bool too_many_isolated(pg_data_t *pgdat)
661	798	{
662	799	unsigned long active, inactive, isolated;
663	800
664		- inactive = node_page_state(zone->zone_pgdat, NR_INACTIVE_FILE) +
665		- node_page_state(zone->zone_pgdat, NR_INACTIVE_ANON);
666		- active = node_page_state(zone->zone_pgdat, NR_ACTIVE_FILE) +
667		- node_page_state(zone->zone_pgdat, NR_ACTIVE_ANON);
668		- isolated = node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE) +
669		- node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON);
	801	+ inactive = node_page_state(pgdat, NR_INACTIVE_FILE) +
	802	+ node_page_state(pgdat, NR_INACTIVE_ANON);
	803	+ active = node_page_state(pgdat, NR_ACTIVE_FILE) +
	804	+ node_page_state(pgdat, NR_ACTIVE_ANON);
	805	+ isolated = node_page_state(pgdat, NR_ISOLATED_FILE) +
	806	+ node_page_state(pgdat, NR_ISOLATED_ANON);
670	807
671	808	return isolated > (inactive + active) / 2;
672	809	}
..	..	@@ -693,7 +830,7 @@
693	830	isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
694	831	unsigned long end_pfn, isolate_mode_t isolate_mode)
695	832	{
696		- struct zone *zone = cc->zone;
	833	+ pg_data_t *pgdat = cc->zone->zone_pgdat;
697	834	unsigned long nr_scanned = 0, nr_isolated = 0;
698	835	struct lruvec *lruvec;
699	836	unsigned long flags = 0;
..	..	@@ -702,13 +839,18 @@
702	839	unsigned long start_pfn = low_pfn;
703	840	bool skip_on_failure = false;
704	841	unsigned long next_skip_pfn = 0;
	842	+ bool skip_updated = false;
705	843
706	844	/*
707	845	* Ensure that there are not too many pages isolated from the LRU
708	846	* list by either parallel reclaimers or compaction. If there are,
709	847	* delay for some time until fewer pages are isolated
710	848	*/
711		- while (unlikely(too_many_isolated(zone))) {
	849	+ while (unlikely(too_many_isolated(pgdat))) {
	850	+ /* stop isolation if there are still pages not migrated */
	851	+ if (cc->nr_migratepages)
	852	+ return 0;
	853	+
712	854	/* async migration should just abort */
713	855	if (cc->mode == MIGRATE_ASYNC)
714	856	return 0;
..	..	@@ -719,8 +861,7 @@
719	861	return 0;
720	862	}
721	863
722		- if (compact_should_abort(cc))
723		- return 0;
	864	+ cond_resched();
724	865
725	866	if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
726	867	skip_on_failure = true;
..	..	@@ -754,13 +895,15 @@
754	895
755	896	/*
756	897	* Periodically drop the lock (if held) regardless of its
757		- * contention, to give chance to IRQs. Abort async compaction
758		- * if contended.
	898	+ * contention, to give chance to IRQs. Abort completely if
	899	+ * a fatal signal is pending.
759	900	*/
760	901	if (!(low_pfn % SWAP_CLUSTER_MAX)
761		- && compact_unlock_should_abort(zone_lru_lock(zone), flags,
762		- &locked, cc))
763		- break;
	902	+ && compact_unlock_should_abort(&pgdat->lru_lock,
	903	+ flags, &locked, cc)) {
	904	+ low_pfn = 0;
	905	+ goto fatal_pending;
	906	+ }
764	907
765	908	if (!pfn_valid_within(low_pfn))
766	909	goto isolate_fail;
..	..	@@ -768,8 +911,19 @@
768	911
769	912	page = pfn_to_page(low_pfn);
770	913
771		- if (!valid_page)
	914	+ /*
	915	+ * Check if the pageblock has already been marked skipped.
	916	+ * Only the aligned PFN is checked as the caller isolates
	917	+ * COMPACT_CLUSTER_MAX at a time so the second call must
	918	+ * not falsely conclude that the block should be skipped.
	919	+ */
	920	+ if (!valid_page && IS_ALIGNED(low_pfn, pageblock_nr_pages)) {
	921	+ if (!cc->ignore_skip_hint && get_pageblock_skip(page)) {
	922	+ low_pfn = end_pfn;
	923	+ goto isolate_abort;
	924	+ }
772	925	valid_page = page;
	926	+ }
773	927
774	928	/*
775	929	* Skip if free. We read page order here without zone lock
..	..	@@ -778,7 +932,7 @@
778	932	* potential isolation targets.
779	933	*/
780	934	if (PageBuddy(page)) {
781		- unsigned long freepage_order = page_order_unsafe(page);
	935	+ unsigned long freepage_order = buddy_order_unsafe(page);
782	936
783	937	/*
784	938	* Without lock, we cannot be sure that what we got is
..	..	@@ -792,12 +946,13 @@
792	946
793	947	/*
794	948	* Regardless of being on LRU, compound pages such as THP and
795		- * hugetlbfs are not to be compacted. We can potentially save
796		- * a lot of iterations if we skip them at once. The check is
797		- * racy, but we can consider only valid values and the only
798		- * danger is skipping too much.
	949	+ * hugetlbfs are not to be compacted unless we are attempting
	950	+ * an allocation much larger than the huge page size (eg CMA).
	951	+ * We can potentially save a lot of iterations if we skip them
	952	+ * at once. The check is racy, but we can consider only valid
	953	+ * values and the only danger is skipping too much.
799	954	*/
800		- if (PageCompound(page)) {
	955	+ if (PageCompound(page) && !cc->alloc_contig) {
801	956	const unsigned int order = compound_order(page);
802	957
803	958	if (likely(order < MAX_ORDER))
..	..	@@ -818,7 +973,7 @@
818	973	if (unlikely(__PageMovable(page)) &&
819	974	!PageIsolated(page)) {
820	975	if (locked) {
821		- spin_unlock_irqrestore(zone_lru_lock(zone),
	976	+ spin_unlock_irqrestore(&pgdat->lru_lock,
822	977	flags);
823	978	locked = false;
824	979	}
..	..	@@ -848,10 +1003,15 @@
848	1003
849	1004	/* If we already hold the lock, we can skip some rechecking */
850	1005	if (!locked) {
851		- locked = compact_trylock_irqsave(zone_lru_lock(zone),
	1006	+ locked = compact_lock_irqsave(&pgdat->lru_lock,
852	1007	&flags, cc);
853		- if (!locked)
854		- break;
	1008	+
	1009	+ /* Try get exclusive access under lock */
	1010	+ if (!skip_updated) {
	1011	+ skip_updated = true;
	1012	+ if (test_and_set_skip(cc, page, low_pfn))
	1013	+ goto isolate_abort;
	1014	+ }
855	1015
856	1016	/* Recheck PageLRU and PageCompound under lock */
857	1017	if (!PageLRU(page))
..	..	@@ -862,41 +1022,41 @@
862	1022	* and it's on LRU. It can only be a THP so the order
863	1023	* is safe to read and it's 0 for tail pages.
864	1024	*/
865		- if (unlikely(PageCompound(page))) {
866		- low_pfn += (1UL << compound_order(page)) - 1;
	1025	+ if (unlikely(PageCompound(page) && !cc->alloc_contig)) {
	1026	+ low_pfn += compound_nr(page) - 1;
867	1027	goto isolate_fail;
868	1028	}
869	1029	}
870	1030
871		- lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
	1031	+ lruvec = mem_cgroup_page_lruvec(page, pgdat);
872	1032
873	1033	/* Try isolate the page */
874	1034	if (__isolate_lru_page(page, isolate_mode) != 0)
875	1035	goto isolate_fail;
876	1036
877		- VM_BUG_ON_PAGE(PageCompound(page), page);
	1037	+ /* The whole page is taken off the LRU; skip the tail pages. */
	1038	+ if (PageCompound(page))
	1039	+ low_pfn += compound_nr(page) - 1;
878	1040
879	1041	/* Successfully isolated */
880	1042	del_page_from_lru_list(page, lruvec, page_lru(page));
881		- inc_node_page_state(page,
882		- NR_ISOLATED_ANON + page_is_file_cache(page));
	1043	+ mod_node_page_state(page_pgdat(page),
	1044	+ NR_ISOLATED_ANON + page_is_file_lru(page),
	1045	+ thp_nr_pages(page));
883	1046
884	1047	isolate_success:
885	1048	list_add(&page->lru, &cc->migratepages);
886		- cc->nr_migratepages++;
887		- nr_isolated++;
	1049	+ cc->nr_migratepages += compound_nr(page);
	1050	+ nr_isolated += compound_nr(page);
888	1051
889	1052	/*
890		- * Record where we could have freed pages by migration and not
891		- * yet flushed them to buddy allocator.
892		- * - this is the lowest page that was isolated and likely be
893		- * then freed by migration.
	1053	+ * Avoid isolating too much unless this block is being
	1054	+ * rescanned (e.g. dirty/writeback pages, parallel allocation)
	1055	+ * or a lock is contended. For contention, isolate quickly to
	1056	+ * potentially remove one source of contention.
894	1057	*/
895		- if (!cc->last_migrated_pfn)
896		- cc->last_migrated_pfn = low_pfn;
897		-
898		- /* Avoid isolating too much */
899		- if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
	1058	+ if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX &&
	1059	+ !cc->rescan && !cc->contended) {
900	1060	++low_pfn;
901	1061	break;
902	1062	}
..	..	@@ -913,12 +1073,11 @@
913	1073	*/
914	1074	if (nr_isolated) {
915	1075	if (locked) {
916		- spin_unlock_irqrestore(zone_lru_lock(zone), flags);
	1076	+ spin_unlock_irqrestore(&pgdat->lru_lock, flags);
917	1077	locked = false;
918	1078	}
919	1079	putback_movable_pages(&cc->migratepages);
920	1080	cc->nr_migratepages = 0;
921		- cc->last_migrated_pfn = 0;
922	1081	nr_isolated = 0;
923	1082	}
924	1083
..	..	@@ -939,19 +1098,28 @@
939	1098	if (unlikely(low_pfn > end_pfn))
940	1099	low_pfn = end_pfn;
941	1100
	1101	+isolate_abort:
942	1102	if (locked)
943		- spin_unlock_irqrestore(zone_lru_lock(zone), flags);
	1103	+ spin_unlock_irqrestore(&pgdat->lru_lock, flags);
944	1104
945	1105	/*
946		- * Update the pageblock-skip information and cached scanner pfn,
947		- * if the whole pageblock was scanned without isolating any page.
	1106	+ * Updated the cached scanner pfn once the pageblock has been scanned
	1107	+ * Pages will either be migrated in which case there is no point
	1108	+ * scanning in the near future or migration failed in which case the
	1109	+ * failure reason may persist. The block is marked for skipping if
	1110	+ * there were no pages isolated in the block or if the block is
	1111	+ * rescanned twice in a row.
948	1112	*/
949		- if (low_pfn == end_pfn)
950		- update_pageblock_skip(cc, valid_page, nr_isolated, true);
	1113	+ if (low_pfn == end_pfn && (!nr_isolated \|\| cc->rescan)) {
	1114	+ if (valid_page && !skip_updated)
	1115	+ set_pageblock_skip(valid_page);
	1116	+ update_cached_migrate(cc, low_pfn);
	1117	+ }
951	1118
952	1119	trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
953	1120	nr_scanned, nr_isolated);
954	1121
	1122	+fatal_pending:
955	1123	cc->total_migrate_scanned += nr_scanned;
956	1124	if (nr_isolated)
957	1125	count_compact_events(COMPACTISOLATED, nr_isolated);
..	..	@@ -998,7 +1166,7 @@
998	1166	if (!pfn)
999	1167	break;
1000	1168
1001		- if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
	1169	+ if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX)
1002	1170	break;
1003	1171	}
1004	1172
..	..	@@ -1012,6 +1180,9 @@
1012	1180	struct page *page)
1013	1181	{
1014	1182	int block_mt;
	1183	+
	1184	+ if (pageblock_skip_persistent(page))
	1185	+ return false;
1015	1186
1016	1187	if ((cc->mode != MIGRATE_ASYNC) \|\| !cc->direct_compaction)
1017	1188	return true;
..	..	@@ -1035,7 +1206,7 @@
1035	1206	* the only small danger is that we skip a potentially suitable
1036	1207	* pageblock, so it's not worth to check order for valid range.
1037	1208	*/
1038		- if (page_order_unsafe(page) >= pageblock_order)
	1209	+ if (buddy_order_unsafe(page) >= pageblock_order)
1039	1210	return false;
1040	1211	}
1041	1212
..	..	@@ -1050,6 +1221,14 @@
1050	1221	return false;
1051	1222	}
1052	1223
	1224	+static inline unsigned int
	1225	+freelist_scan_limit(struct compact_control *cc)
	1226	+{
	1227	+ unsigned short shift = BITS_PER_LONG - 1;
	1228	+
	1229	+ return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1;
	1230	+}
	1231	+
1053	1232	/*
1054	1233	* Test whether the free scanner has reached the same or lower pageblock than
1055	1234	* the migration scanner, and compaction should thus terminate.
..	..	@@ -1058,6 +1237,249 @@
1058	1237	{
1059	1238	return (cc->free_pfn >> pageblock_order)
1060	1239	<= (cc->migrate_pfn >> pageblock_order);
	1240	+}
	1241	+
	1242	+/*
	1243	+ * Used when scanning for a suitable migration target which scans freelists
	1244	+ * in reverse. Reorders the list such as the unscanned pages are scanned
	1245	+ * first on the next iteration of the free scanner
	1246	+ */
	1247	+static void
	1248	+move_freelist_head(struct list_head freelist, struct page freepage)
	1249	+{
	1250	+ LIST_HEAD(sublist);
	1251	+
	1252	+ if (!list_is_last(freelist, &freepage->lru)) {
	1253	+ list_cut_before(&sublist, freelist, &freepage->lru);
	1254	+ if (!list_empty(&sublist))
	1255	+ list_splice_tail(&sublist, freelist);
	1256	+ }
	1257	+}
	1258	+
	1259	+/*
	1260	+ * Similar to move_freelist_head except used by the migration scanner
	1261	+ * when scanning forward. It's possible for these list operations to
	1262	+ * move against each other if they search the free list exactly in
	1263	+ * lockstep.
	1264	+ */
	1265	+static void
	1266	+move_freelist_tail(struct list_head freelist, struct page freepage)
	1267	+{
	1268	+ LIST_HEAD(sublist);
	1269	+
	1270	+ if (!list_is_first(freelist, &freepage->lru)) {
	1271	+ list_cut_position(&sublist, freelist, &freepage->lru);
	1272	+ if (!list_empty(&sublist))
	1273	+ list_splice_tail(&sublist, freelist);
	1274	+ }
	1275	+}
	1276	+
	1277	+static void
	1278	+fast_isolate_around(struct compact_control *cc, unsigned long pfn)
	1279	+{
	1280	+ unsigned long start_pfn, end_pfn;
	1281	+ struct page *page;
	1282	+
	1283	+ /* Do not search around if there are enough pages already */
	1284	+ if (cc->nr_freepages >= cc->nr_migratepages)
	1285	+ return;
	1286	+
	1287	+ /* Minimise scanning during async compaction */
	1288	+ if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC)
	1289	+ return;
	1290	+
	1291	+ /* Pageblock boundaries */
	1292	+ start_pfn = max(pageblock_start_pfn(pfn), cc->zone->zone_start_pfn);
	1293	+ end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone));
	1294	+
	1295	+ page = pageblock_pfn_to_page(start_pfn, end_pfn, cc->zone);
	1296	+ if (!page)
	1297	+ return;
	1298	+
	1299	+ isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, 1, false);
	1300	+
	1301	+ /* Skip this pageblock in the future as it's full or nearly full */
	1302	+ if (cc->nr_freepages < cc->nr_migratepages)
	1303	+ set_pageblock_skip(page);
	1304	+
	1305	+ return;
	1306	+}
	1307	+
	1308	+/* Search orders in round-robin fashion */
	1309	+static int next_search_order(struct compact_control *cc, int order)
	1310	+{
	1311	+ order--;
	1312	+ if (order < 0)
	1313	+ order = cc->order - 1;
	1314	+
	1315	+ /* Search wrapped around? */
	1316	+ if (order == cc->search_order) {
	1317	+ cc->search_order--;
	1318	+ if (cc->search_order < 0)
	1319	+ cc->search_order = cc->order - 1;
	1320	+ return -1;
	1321	+ }
	1322	+
	1323	+ return order;
	1324	+}
	1325	+
	1326	+static unsigned long
	1327	+fast_isolate_freepages(struct compact_control *cc)
	1328	+{
	1329	+ unsigned int limit = min(1U, freelist_scan_limit(cc) >> 1);
	1330	+ unsigned int nr_scanned = 0;
	1331	+ unsigned long low_pfn, min_pfn, highest = 0;
	1332	+ unsigned long nr_isolated = 0;
	1333	+ unsigned long distance;
	1334	+ struct page *page = NULL;
	1335	+ bool scan_start = false;
	1336	+ int order;
	1337	+
	1338	+ /* Full compaction passes in a negative order */
	1339	+ if (cc->order <= 0)
	1340	+ return cc->free_pfn;
	1341	+
	1342	+ /*
	1343	+ * If starting the scan, use a deeper search and use the highest
	1344	+ * PFN found if a suitable one is not found.
	1345	+ */
	1346	+ if (cc->free_pfn >= cc->zone->compact_init_free_pfn) {
	1347	+ limit = pageblock_nr_pages >> 1;
	1348	+ scan_start = true;
	1349	+ }
	1350	+
	1351	+ /*
	1352	+ * Preferred point is in the top quarter of the scan space but take
	1353	+ * a pfn from the top half if the search is problematic.
	1354	+ */
	1355	+ distance = (cc->free_pfn - cc->migrate_pfn);
	1356	+ low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2));
	1357	+ min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1));
	1358	+
	1359	+ if (WARN_ON_ONCE(min_pfn > low_pfn))
	1360	+ low_pfn = min_pfn;
	1361	+
	1362	+ /*
	1363	+ * Search starts from the last successful isolation order or the next
	1364	+ * order to search after a previous failure
	1365	+ */
	1366	+ cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order);
	1367	+
	1368	+ for (order = cc->search_order;
	1369	+ !page && order >= 0;
	1370	+ order = next_search_order(cc, order)) {
	1371	+ struct free_area *area = &cc->zone->free_area[order];
	1372	+ struct list_head *freelist;
	1373	+ struct page *freepage;
	1374	+ unsigned long flags;
	1375	+ unsigned int order_scanned = 0;
	1376	+ unsigned long high_pfn = 0;
	1377	+
	1378	+ if (!area->nr_free)
	1379	+ continue;
	1380	+
	1381	+ spin_lock_irqsave(&cc->zone->lock, flags);
	1382	+ freelist = &area->free_list[MIGRATE_MOVABLE];
	1383	+ list_for_each_entry_reverse(freepage, freelist, lru) {
	1384	+ unsigned long pfn;
	1385	+
	1386	+ order_scanned++;
	1387	+ nr_scanned++;
	1388	+ pfn = page_to_pfn(freepage);
	1389	+
	1390	+ if (pfn >= highest)
	1391	+ highest = max(pageblock_start_pfn(pfn),
	1392	+ cc->zone->zone_start_pfn);
	1393	+
	1394	+ if (pfn >= low_pfn) {
	1395	+ cc->fast_search_fail = 0;
	1396	+ cc->search_order = order;
	1397	+ page = freepage;
	1398	+ break;
	1399	+ }
	1400	+
	1401	+ if (pfn >= min_pfn && pfn > high_pfn) {
	1402	+ high_pfn = pfn;
	1403	+
	1404	+ /* Shorten the scan if a candidate is found */
	1405	+ limit >>= 1;
	1406	+ }
	1407	+
	1408	+ if (order_scanned >= limit)
	1409	+ break;
	1410	+ }
	1411	+
	1412	+ /* Use a minimum pfn if a preferred one was not found */
	1413	+ if (!page && high_pfn) {
	1414	+ page = pfn_to_page(high_pfn);
	1415	+
	1416	+ /* Update freepage for the list reorder below */
	1417	+ freepage = page;
	1418	+ }
	1419	+
	1420	+ /* Reorder to so a future search skips recent pages */
	1421	+ move_freelist_head(freelist, freepage);
	1422	+
	1423	+ /* Isolate the page if available */
	1424	+ if (page) {
	1425	+ if (__isolate_free_page(page, order)) {
	1426	+ set_page_private(page, order);
	1427	+ nr_isolated = 1 << order;
	1428	+ cc->nr_freepages += nr_isolated;
	1429	+ list_add_tail(&page->lru, &cc->freepages);
	1430	+ count_compact_events(COMPACTISOLATED, nr_isolated);
	1431	+ } else {
	1432	+ /* If isolation fails, abort the search */
	1433	+ order = cc->search_order + 1;
	1434	+ page = NULL;
	1435	+ }
	1436	+ }
	1437	+
	1438	+ spin_unlock_irqrestore(&cc->zone->lock, flags);
	1439	+
	1440	+ /*
	1441	+ * Smaller scan on next order so the total scan ig related
	1442	+ * to freelist_scan_limit.
	1443	+ */
	1444	+ if (order_scanned >= limit)
	1445	+ limit = min(1U, limit >> 1);
	1446	+ }
	1447	+
	1448	+ if (!page) {
	1449	+ cc->fast_search_fail++;
	1450	+ if (scan_start) {
	1451	+ /*
	1452	+ * Use the highest PFN found above min. If one was
	1453	+ * not found, be pessimistic for direct compaction
	1454	+ * and use the min mark.
	1455	+ */
	1456	+ if (highest) {
	1457	+ page = pfn_to_page(highest);
	1458	+ cc->free_pfn = highest;
	1459	+ } else {
	1460	+ if (cc->direct_compaction && pfn_valid(min_pfn)) {
	1461	+ page = pageblock_pfn_to_page(min_pfn,
	1462	+ min(pageblock_end_pfn(min_pfn),
	1463	+ zone_end_pfn(cc->zone)),
	1464	+ cc->zone);
	1465	+ cc->free_pfn = min_pfn;
	1466	+ }
	1467	+ }
	1468	+ }
	1469	+ }
	1470	+
	1471	+ if (highest && highest >= cc->zone->compact_cached_free_pfn) {
	1472	+ highest -= pageblock_nr_pages;
	1473	+ cc->zone->compact_cached_free_pfn = highest;
	1474	+ }
	1475	+
	1476	+ cc->total_free_scanned += nr_scanned;
	1477	+ if (!page)
	1478	+ return cc->free_pfn;
	1479	+
	1480	+ low_pfn = page_to_pfn(page);
	1481	+ fast_isolate_around(cc, low_pfn);
	1482	+ return low_pfn;
1061	1483	}
1062	1484
1063	1485	/*
..	..	@@ -1073,6 +1495,12 @@
1073	1495	unsigned long block_end_pfn; /* end of current pageblock */
1074	1496	unsigned long low_pfn; /* lowest pfn scanner is able to scan */
1075	1497	struct list_head *freelist = &cc->freepages;
	1498	+ unsigned int stride;
	1499	+
	1500	+ /* Try a small search of the free lists for a candidate */
	1501	+ isolate_start_pfn = fast_isolate_freepages(cc);
	1502	+ if (cc->nr_freepages)
	1503	+ goto splitmap;
1076	1504
1077	1505	/*
1078	1506	* Initialise the free scanner. The starting point is where we last
..	..	@@ -1081,15 +1509,16 @@
1081	1509	* this pfn aligned down to the pageblock boundary, because we do
1082	1510	* block_start_pfn -= pageblock_nr_pages in the for loop.
1083	1511	* For ending point, take care when isolating in last pageblock of a
1084		- * a zone which ends in the middle of a pageblock.
	1512	+ * zone which ends in the middle of a pageblock.
1085	1513	* The low boundary is the end of the pageblock the migration scanner
1086	1514	* is using.
1087	1515	*/
1088	1516	isolate_start_pfn = cc->free_pfn;
1089		- block_start_pfn = pageblock_start_pfn(cc->free_pfn);
	1517	+ block_start_pfn = pageblock_start_pfn(isolate_start_pfn);
1090	1518	block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1091	1519	zone_end_pfn(zone));
1092	1520	low_pfn = pageblock_end_pfn(cc->migrate_pfn);
	1521	+ stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1;
1093	1522
1094	1523	/*
1095	1524	* Isolate free pages until enough are available to migrate the
..	..	@@ -1100,14 +1529,14 @@
1100	1529	block_end_pfn = block_start_pfn,
1101	1530	block_start_pfn -= pageblock_nr_pages,
1102	1531	isolate_start_pfn = block_start_pfn) {
	1532	+ unsigned long nr_isolated;
	1533	+
1103	1534	/*
1104	1535	* This can iterate a massively long zone without finding any
1105		- * suitable migration targets, so periodically check if we need
1106		- * to schedule, or even abort async compaction.
	1536	+ * suitable migration targets, so periodically check resched.
1107	1537	*/
1108		- if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1109		- && compact_should_abort(cc))
1110		- break;
	1538	+ if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)))
	1539	+ cond_resched();
1111	1540
1112	1541	page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1113	1542	zone);
..	..	@@ -1123,15 +1552,15 @@
1123	1552	continue;
1124	1553
1125	1554	/* Found a block suitable for isolating free pages from. */
1126		- isolate_freepages_block(cc, &isolate_start_pfn, block_end_pfn,
1127		- freelist, false);
	1555	+ nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn,
	1556	+ block_end_pfn, freelist, stride, false);
1128	1557
1129		- /*
1130		- * If we isolated enough freepages, or aborted due to lock
1131		- * contention, terminate.
1132		- */
1133		- if ((cc->nr_freepages >= cc->nr_migratepages)
1134		- \|\| cc->contended) {
	1558	+ /* Update the skip hint if the full pageblock was scanned */
	1559	+ if (isolate_start_pfn == block_end_pfn)
	1560	+ update_pageblock_skip(cc, page, block_start_pfn);
	1561	+
	1562	+ /* Are enough freepages isolated? */
	1563	+ if (cc->nr_freepages >= cc->nr_migratepages) {
1135	1564	if (isolate_start_pfn >= block_end_pfn) {
1136	1565	/*
1137	1566	* Restart at previous pageblock if more
..	..	@@ -1148,10 +1577,14 @@
1148	1577	*/
1149	1578	break;
1150	1579	}
1151		- }
1152	1580
1153		- /* __isolate_free_page() does not map the pages */
1154		- map_pages(freelist);
	1581	+ /* Adjust stride depending on isolation */
	1582	+ if (nr_isolated) {
	1583	+ stride = 1;
	1584	+ continue;
	1585	+ }
	1586	+ stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1);
	1587	+ }
1155	1588
1156	1589	/*
1157	1590	* Record where the free scanner will restart next time. Either we
..	..	@@ -1160,6 +1593,10 @@
1160	1593	* and the loop terminated due to isolate_start_pfn < low_pfn
1161	1594	*/
1162	1595	cc->free_pfn = isolate_start_pfn;
	1596	+
	1597	+splitmap:
	1598	+ /* __isolate_free_page() does not map the pages */
	1599	+ split_map_pages(freelist);
1163	1600	}
1164	1601
1165	1602	/*
..	..	@@ -1172,13 +1609,8 @@
1172	1609	struct compact_control cc = (struct compact_control )data;
1173	1610	struct page *freepage;
1174	1611
1175		- /*
1176		- * Isolate free pages if necessary, and if we are not aborting due to
1177		- * contention.
1178		- */
1179	1612	if (list_empty(&cc->freepages)) {
1180		- if (!cc->contended)
1181		- isolate_freepages(cc);
	1613	+ isolate_freepages(cc);
1182	1614
1183	1615	if (list_empty(&cc->freepages))
1184	1616	return NULL;
..	..	@@ -1215,15 +1647,158 @@
1215	1647	* Allow userspace to control policy on scanning the unevictable LRU for
1216	1648	* compactable pages.
1217	1649	*/
	1650	+#ifdef CONFIG_PREEMPT_RT
	1651	+int sysctl_compact_unevictable_allowed __read_mostly = 0;
	1652	+#else
1218	1653	int sysctl_compact_unevictable_allowed __read_mostly = 1;
	1654	+#endif
	1655	+
	1656	+static inline void
	1657	+update_fast_start_pfn(struct compact_control *cc, unsigned long pfn)
	1658	+{
	1659	+ if (cc->fast_start_pfn == ULONG_MAX)
	1660	+ return;
	1661	+
	1662	+ if (!cc->fast_start_pfn)
	1663	+ cc->fast_start_pfn = pfn;
	1664	+
	1665	+ cc->fast_start_pfn = min(cc->fast_start_pfn, pfn);
	1666	+}
	1667	+
	1668	+static inline unsigned long
	1669	+reinit_migrate_pfn(struct compact_control *cc)
	1670	+{
	1671	+ if (!cc->fast_start_pfn \|\| cc->fast_start_pfn == ULONG_MAX)
	1672	+ return cc->migrate_pfn;
	1673	+
	1674	+ cc->migrate_pfn = cc->fast_start_pfn;
	1675	+ cc->fast_start_pfn = ULONG_MAX;
	1676	+
	1677	+ return cc->migrate_pfn;
	1678	+}
	1679	+
	1680	+/*
	1681	+ * Briefly search the free lists for a migration source that already has
	1682	+ * some free pages to reduce the number of pages that need migration
	1683	+ * before a pageblock is free.
	1684	+ */
	1685	+static unsigned long fast_find_migrateblock(struct compact_control *cc)
	1686	+{
	1687	+ unsigned int limit = freelist_scan_limit(cc);
	1688	+ unsigned int nr_scanned = 0;
	1689	+ unsigned long distance;
	1690	+ unsigned long pfn = cc->migrate_pfn;
	1691	+ unsigned long high_pfn;
	1692	+ int order;
	1693	+ bool found_block = false;
	1694	+
	1695	+ /* Skip hints are relied on to avoid repeats on the fast search */
	1696	+ if (cc->ignore_skip_hint)
	1697	+ return pfn;
	1698	+
	1699	+ /*
	1700	+ * If the migrate_pfn is not at the start of a zone or the start
	1701	+ * of a pageblock then assume this is a continuation of a previous
	1702	+ * scan restarted due to COMPACT_CLUSTER_MAX.
	1703	+ */
	1704	+ if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn))
	1705	+ return pfn;
	1706	+
	1707	+ /*
	1708	+ * For smaller orders, just linearly scan as the number of pages
	1709	+ * to migrate should be relatively small and does not necessarily
	1710	+ * justify freeing up a large block for a small allocation.
	1711	+ */
	1712	+ if (cc->order <= PAGE_ALLOC_COSTLY_ORDER)
	1713	+ return pfn;
	1714	+
	1715	+ /*
	1716	+ * Only allow kcompactd and direct requests for movable pages to
	1717	+ * quickly clear out a MOVABLE pageblock for allocation. This
	1718	+ * reduces the risk that a large movable pageblock is freed for
	1719	+ * an unmovable/reclaimable small allocation.
	1720	+ */
	1721	+ if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE)
	1722	+ return pfn;
	1723	+
	1724	+ /*
	1725	+ * When starting the migration scanner, pick any pageblock within the
	1726	+ * first half of the search space. Otherwise try and pick a pageblock
	1727	+ * within the first eighth to reduce the chances that a migration
	1728	+ * target later becomes a source.
	1729	+ */
	1730	+ distance = (cc->free_pfn - cc->migrate_pfn) >> 1;
	1731	+ if (cc->migrate_pfn != cc->zone->zone_start_pfn)
	1732	+ distance >>= 2;
	1733	+ high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance);
	1734	+
	1735	+ for (order = cc->order - 1;
	1736	+ order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit;
	1737	+ order--) {
	1738	+ struct free_area *area = &cc->zone->free_area[order];
	1739	+ struct list_head *freelist;
	1740	+ unsigned long flags;
	1741	+ struct page *freepage;
	1742	+
	1743	+ if (!area->nr_free)
	1744	+ continue;
	1745	+
	1746	+ spin_lock_irqsave(&cc->zone->lock, flags);
	1747	+ freelist = &area->free_list[MIGRATE_MOVABLE];
	1748	+ list_for_each_entry(freepage, freelist, lru) {
	1749	+ unsigned long free_pfn;
	1750	+
	1751	+ if (nr_scanned++ >= limit) {
	1752	+ move_freelist_tail(freelist, freepage);
	1753	+ break;
	1754	+ }
	1755	+
	1756	+ free_pfn = page_to_pfn(freepage);
	1757	+ if (free_pfn < high_pfn) {
	1758	+ /*
	1759	+ * Avoid if skipped recently. Ideally it would
	1760	+ * move to the tail but even safe iteration of
	1761	+ * the list assumes an entry is deleted, not
	1762	+ * reordered.
	1763	+ */
	1764	+ if (get_pageblock_skip(freepage))
	1765	+ continue;
	1766	+
	1767	+ /* Reorder to so a future search skips recent pages */
	1768	+ move_freelist_tail(freelist, freepage);
	1769	+
	1770	+ update_fast_start_pfn(cc, free_pfn);
	1771	+ pfn = pageblock_start_pfn(free_pfn);
	1772	+ if (pfn < cc->zone->zone_start_pfn)
	1773	+ pfn = cc->zone->zone_start_pfn;
	1774	+ cc->fast_search_fail = 0;
	1775	+ found_block = true;
	1776	+ set_pageblock_skip(freepage);
	1777	+ break;
	1778	+ }
	1779	+ }
	1780	+ spin_unlock_irqrestore(&cc->zone->lock, flags);
	1781	+ }
	1782	+
	1783	+ cc->total_migrate_scanned += nr_scanned;
	1784	+
	1785	+ /*
	1786	+ * If fast scanning failed then use a cached entry for a page block
	1787	+ * that had free pages as the basis for starting a linear scan.
	1788	+ */
	1789	+ if (!found_block) {
	1790	+ cc->fast_search_fail++;
	1791	+ pfn = reinit_migrate_pfn(cc);
	1792	+ }
	1793	+ return pfn;
	1794	+}
1219	1795
1220	1796	/*
1221	1797	* Isolate all pages that can be migrated from the first suitable block,
1222	1798	* starting at the block pointed to by the migrate scanner pfn within
1223	1799	* compact_control.
1224	1800	*/
1225		-static isolate_migrate_t isolate_migratepages(struct zone *zone,
1226		- struct compact_control *cc)
	1801	+static isolate_migrate_t isolate_migratepages(struct compact_control *cc)
1227	1802	{
1228	1803	unsigned long block_start_pfn;
1229	1804	unsigned long block_end_pfn;
..	..	@@ -1232,15 +1807,24 @@
1232	1807	const isolate_mode_t isolate_mode =
1233	1808	(sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) \|
1234	1809	(cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
	1810	+ bool fast_find_block;
1235	1811
1236	1812	/*
1237	1813	* Start at where we last stopped, or beginning of the zone as
1238		- * initialized by compact_zone()
	1814	+ * initialized by compact_zone(). The first failure will use
	1815	+ * the lowest PFN as the starting point for linear scanning.
1239	1816	*/
1240		- low_pfn = cc->migrate_pfn;
	1817	+ low_pfn = fast_find_migrateblock(cc);
1241	1818	block_start_pfn = pageblock_start_pfn(low_pfn);
1242		- if (block_start_pfn < zone->zone_start_pfn)
1243		- block_start_pfn = zone->zone_start_pfn;
	1819	+ if (block_start_pfn < cc->zone->zone_start_pfn)
	1820	+ block_start_pfn = cc->zone->zone_start_pfn;
	1821	+
	1822	+ /*
	1823	+ * fast_find_migrateblock marks a pageblock skipped so to avoid
	1824	+ * the isolation_suitable check below, check whether the fast
	1825	+ * search was successful.
	1826	+ */
	1827	+ fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail;
1244	1828
1245	1829	/* Only scan within a pageblock boundary */
1246	1830	block_end_pfn = pageblock_end_pfn(low_pfn);
..	..	@@ -1250,6 +1834,7 @@
1250	1834	* Do not cross the free scanner.
1251	1835	*/
1252	1836	for (; block_end_pfn <= cc->free_pfn;
	1837	+ fast_find_block = false,
1253	1838	low_pfn = block_end_pfn,
1254	1839	block_start_pfn = block_end_pfn,
1255	1840	block_end_pfn += pageblock_nr_pages) {
..	..	@@ -1257,34 +1842,45 @@
1257	1842	/*
1258	1843	* This can potentially iterate a massively long zone with
1259	1844	* many pageblocks unsuitable, so periodically check if we
1260		- * need to schedule, or even abort async compaction.
	1845	+ * need to schedule.
1261	1846	*/
1262		- if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1263		- && compact_should_abort(cc))
1264		- break;
	1847	+ if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)))
	1848	+ cond_resched();
1265	1849
1266		- page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1267		- zone);
	1850	+ page = pageblock_pfn_to_page(block_start_pfn,
	1851	+ block_end_pfn, cc->zone);
1268	1852	if (!page)
1269	1853	continue;
1270	1854
1271		- /* If isolation recently failed, do not retry */
1272		- if (!isolation_suitable(cc, page))
	1855	+ /*
	1856	+ * If isolation recently failed, do not retry. Only check the
	1857	+ * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
	1858	+ * to be visited multiple times. Assume skip was checked
	1859	+ * before making it "skip" so other compaction instances do
	1860	+ * not scan the same block.
	1861	+ */
	1862	+ if (IS_ALIGNED(low_pfn, pageblock_nr_pages) &&
	1863	+ !fast_find_block && !isolation_suitable(cc, page))
1273	1864	continue;
1274	1865
1275	1866	/*
1276		- * For async compaction, also only scan in MOVABLE blocks.
1277		- * Async compaction is optimistic to see if the minimum amount
1278		- * of work satisfies the allocation.
	1867	+ * For async compaction, also only scan in MOVABLE blocks
	1868	+ * without huge pages. Async compaction is optimistic to see
	1869	+ * if the minimum amount of work satisfies the allocation.
	1870	+ * The cached PFN is updated as it's possible that all
	1871	+ * remaining blocks between source and target are unsuitable
	1872	+ * and the compaction scanners fail to meet.
1279	1873	*/
1280		- if (!suitable_migration_source(cc, page))
	1874	+ if (!suitable_migration_source(cc, page)) {
	1875	+ update_cached_migrate(cc, block_end_pfn);
1281	1876	continue;
	1877	+ }
1282	1878
1283	1879	/* Perform the isolation */
1284	1880	low_pfn = isolate_migratepages_block(cc, low_pfn,
1285	1881	block_end_pfn, isolate_mode);
1286	1882
1287		- if (!low_pfn \|\| cc->contended)
	1883	+ if (!low_pfn)
1288	1884	return ISOLATE_ABORT;
1289	1885
1290	1886	/*
..	..	@@ -1310,19 +1906,95 @@
1310	1906	return order == -1;
1311	1907	}
1312	1908
1313		-static enum compact_result __compact_finished(struct zone *zone,
1314		- struct compact_control *cc)
	1909	+static bool kswapd_is_running(pg_data_t *pgdat)
	1910	+{
	1911	+ return pgdat->kswapd && (pgdat->kswapd->state == TASK_RUNNING);
	1912	+}
	1913	+
	1914	+/*
	1915	+ * A zone's fragmentation score is the external fragmentation wrt to the
	1916	+ * COMPACTION_HPAGE_ORDER. It returns a value in the range [0, 100].
	1917	+ */
	1918	+static unsigned int fragmentation_score_zone(struct zone *zone)
	1919	+{
	1920	+ return extfrag_for_order(zone, COMPACTION_HPAGE_ORDER);
	1921	+}
	1922	+
	1923	+/*
	1924	+ * A weighted zone's fragmentation score is the external fragmentation
	1925	+ * wrt to the COMPACTION_HPAGE_ORDER scaled by the zone's size. It
	1926	+ * returns a value in the range [0, 100].
	1927	+ *
	1928	+ * The scaling factor ensures that proactive compaction focuses on larger
	1929	+ * zones like ZONE_NORMAL, rather than smaller, specialized zones like
	1930	+ * ZONE_DMA32. For smaller zones, the score value remains close to zero,
	1931	+ * and thus never exceeds the high threshold for proactive compaction.
	1932	+ */
	1933	+static unsigned int fragmentation_score_zone_weighted(struct zone *zone)
	1934	+{
	1935	+ unsigned long score;
	1936	+
	1937	+ score = zone->present_pages * fragmentation_score_zone(zone);
	1938	+ return div64_ul(score, zone->zone_pgdat->node_present_pages + 1);
	1939	+}
	1940	+
	1941	+/*
	1942	+ * The per-node proactive (background) compaction process is started by its
	1943	+ * corresponding kcompactd thread when the node's fragmentation score
	1944	+ * exceeds the high threshold. The compaction process remains active till
	1945	+ * the node's score falls below the low threshold, or one of the back-off
	1946	+ * conditions is met.
	1947	+ */
	1948	+static unsigned int fragmentation_score_node(pg_data_t *pgdat)
	1949	+{
	1950	+ unsigned int score = 0;
	1951	+ int zoneid;
	1952	+
	1953	+ for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
	1954	+ struct zone *zone;
	1955	+
	1956	+ zone = &pgdat->node_zones[zoneid];
	1957	+ score += fragmentation_score_zone_weighted(zone);
	1958	+ }
	1959	+
	1960	+ return score;
	1961	+}
	1962	+
	1963	+static unsigned int fragmentation_score_wmark(pg_data_t *pgdat, bool low)
	1964	+{
	1965	+ unsigned int wmark_low;
	1966	+
	1967	+ /*
	1968	+ * Cap the low watermak to avoid excessive compaction
	1969	+ * activity in case a user sets the proactivess tunable
	1970	+ * close to 100 (maximum).
	1971	+ */
	1972	+ wmark_low = max(100U - sysctl_compaction_proactiveness, 5U);
	1973	+ return low ? wmark_low : min(wmark_low + 10, 100U);
	1974	+}
	1975	+
	1976	+static bool should_proactive_compact_node(pg_data_t *pgdat)
	1977	+{
	1978	+ int wmark_high;
	1979	+
	1980	+ if (!sysctl_compaction_proactiveness \|\| kswapd_is_running(pgdat))
	1981	+ return false;
	1982	+
	1983	+ wmark_high = fragmentation_score_wmark(pgdat, false);
	1984	+ return fragmentation_score_node(pgdat) > wmark_high;
	1985	+}
	1986	+
	1987	+static enum compact_result __compact_finished(struct compact_control *cc)
1315	1988	{
1316	1989	unsigned int order;
1317	1990	const int migratetype = cc->migratetype;
1318		-
1319		- if (cc->contended \|\| fatal_signal_pending(current))
1320		- return COMPACT_CONTENDED;
	1991	+ int ret;
	1992	+ bool abort_compact = false;
1321	1993
1322	1994	/* Compaction run completes if the migrate and free scanner meet */
1323	1995	if (compact_scanners_met(cc)) {
1324	1996	/* Let the next compaction start anew. */
1325		- reset_cached_positions(zone);
	1997	+ reset_cached_positions(cc->zone);
1326	1998
1327	1999	/*
1328	2000	* Mark that the PG_migrate_skip information should be cleared
..	..	@@ -1331,7 +2003,7 @@
1331	2003	* based on an allocation request.
1332	2004	*/
1333	2005	if (cc->direct_compaction)
1334		- zone->compact_blockskip_flush = true;
	2006	+ cc->zone->compact_blockskip_flush = true;
1335	2007
1336	2008	if (cc->whole_zone)
1337	2009	return COMPACT_COMPLETE;
..	..	@@ -1339,33 +2011,51 @@
1339	2011	return COMPACT_PARTIAL_SKIPPED;
1340	2012	}
1341	2013
	2014	+ if (cc->proactive_compaction) {
	2015	+ int score, wmark_low;
	2016	+ pg_data_t *pgdat;
	2017	+
	2018	+ pgdat = cc->zone->zone_pgdat;
	2019	+ if (kswapd_is_running(pgdat))
	2020	+ return COMPACT_PARTIAL_SKIPPED;
	2021	+
	2022	+ score = fragmentation_score_zone(cc->zone);
	2023	+ wmark_low = fragmentation_score_wmark(pgdat, true);
	2024	+
	2025	+ if (score > wmark_low)
	2026	+ ret = COMPACT_CONTINUE;
	2027	+ else
	2028	+ ret = COMPACT_SUCCESS;
	2029	+
	2030	+ goto out;
	2031	+ }
	2032	+
1342	2033	if (is_via_compact_memory(cc->order))
1343	2034	return COMPACT_CONTINUE;
1344	2035
1345		- if (cc->finishing_block) {
1346		- /*
1347		- * We have finished the pageblock, but better check again that
1348		- * we really succeeded.
1349		- */
1350		- if (IS_ALIGNED(cc->migrate_pfn, pageblock_nr_pages))
1351		- cc->finishing_block = false;
1352		- else
1353		- return COMPACT_CONTINUE;
1354		- }
	2036	+ /*
	2037	+ * Always finish scanning a pageblock to reduce the possibility of
	2038	+ * fallbacks in the future. This is particularly important when
	2039	+ * migration source is unmovable/reclaimable but it's not worth
	2040	+ * special casing.
	2041	+ */
	2042	+ if (!IS_ALIGNED(cc->migrate_pfn, pageblock_nr_pages))
	2043	+ return COMPACT_CONTINUE;
1355	2044
1356	2045	/* Direct compactor: Is a suitable page free? */
	2046	+ ret = COMPACT_NO_SUITABLE_PAGE;
1357	2047	for (order = cc->order; order < MAX_ORDER; order++) {
1358		- struct free_area *area = &zone->free_area[order];
	2048	+ struct free_area *area = &cc->zone->free_area[order];
1359	2049	bool can_steal;
1360	2050
1361	2051	/* Job done if page is free of the right migratetype */
1362		- if (!list_empty(&area->free_list[migratetype]))
	2052	+ if (!free_area_empty(area, migratetype))
1363	2053	return COMPACT_SUCCESS;
1364	2054
1365	2055	#ifdef CONFIG_CMA
1366	2056	/* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1367	2057	if (migratetype == MIGRATE_MOVABLE &&
1368		- !list_empty(&area->free_list[MIGRATE_CMA]))
	2058	+ !free_area_empty(area, MIGRATE_CMA))
1369	2059	return COMPACT_SUCCESS;
1370	2060	#endif
1371	2061	/*
..	..	@@ -1393,21 +2083,25 @@
1393	2083	return COMPACT_SUCCESS;
1394	2084	}
1395	2085
1396		- cc->finishing_block = true;
1397		- return COMPACT_CONTINUE;
	2086	+ ret = COMPACT_CONTINUE;
	2087	+ break;
1398	2088	}
1399	2089	}
1400	2090
1401		- return COMPACT_NO_SUITABLE_PAGE;
	2091	+out:
	2092	+ trace_android_vh_compact_finished(&abort_compact);
	2093	+ if (cc->contended \|\| fatal_signal_pending(current) \|\| abort_compact)
	2094	+ ret = COMPACT_CONTENDED;
	2095	+
	2096	+ return ret;
1402	2097	}
1403	2098
1404		-static enum compact_result compact_finished(struct zone *zone,
1405		- struct compact_control *cc)
	2099	+static enum compact_result compact_finished(struct compact_control *cc)
1406	2100	{
1407	2101	int ret;
1408	2102
1409		- ret = __compact_finished(zone, cc);
1410		- trace_mm_compaction_finished(zone, cc->order, ret);
	2103	+ ret = __compact_finished(cc);
	2104	+ trace_mm_compaction_finished(cc->zone, cc->order, ret);
1411	2105	if (ret == COMPACT_NO_SUITABLE_PAGE)
1412	2106	ret = COMPACT_CONTINUE;
1413	2107
..	..	@@ -1423,7 +2117,7 @@
1423	2117	*/
1424	2118	static enum compact_result __compaction_suitable(struct zone *zone, int order,
1425	2119	unsigned int alloc_flags,
1426		- int classzone_idx,
	2120	+ int highest_zoneidx,
1427	2121	unsigned long wmark_target)
1428	2122	{
1429	2123	unsigned long watermark;
..	..	@@ -1431,12 +2125,12 @@
1431	2125	if (is_via_compact_memory(order))
1432	2126	return COMPACT_CONTINUE;
1433	2127
1434		- watermark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
	2128	+ watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
1435	2129	/*
1436	2130	* If watermarks for high-order allocation are already met, there
1437	2131	* should be no need for compaction at all.
1438	2132	*/
1439		- if (zone_watermark_ok(zone, order, watermark, classzone_idx,
	2133	+ if (zone_watermark_ok(zone, order, watermark, highest_zoneidx,
1440	2134	alloc_flags))
1441	2135	return COMPACT_SUCCESS;
1442	2136
..	..	@@ -1446,9 +2140,9 @@
1446	2140	* watermark and alloc_flags have to match, or be more pessimistic than
1447	2141	* the check in __isolate_free_page(). We don't use the direct
1448	2142	* compactor's alloc_flags, as they are not relevant for freepage
1449		- * isolation. We however do use the direct compactor's classzone_idx to
1450		- * skip over zones where lowmem reserves would prevent allocation even
1451		- * if compaction succeeds.
	2143	+ * isolation. We however do use the direct compactor's highest_zoneidx
	2144	+ * to skip over zones where lowmem reserves would prevent allocation
	2145	+ * even if compaction succeeds.
1452	2146	* For costly orders, we require low watermark instead of min for
1453	2147	* compaction to proceed to increase its chances.
1454	2148	* ALLOC_CMA is used, as pages in CMA pageblocks are considered
..	..	@@ -1457,7 +2151,7 @@
1457	2151	watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
1458	2152	low_wmark_pages(zone) : min_wmark_pages(zone);
1459	2153	watermark += compact_gap(order);
1460		- if (!__zone_watermark_ok(zone, 0, watermark, classzone_idx,
	2154	+ if (!__zone_watermark_ok(zone, 0, watermark, highest_zoneidx,
1461	2155	ALLOC_CMA, wmark_target))
1462	2156	return COMPACT_SKIPPED;
1463	2157
..	..	@@ -1466,12 +2160,12 @@
1466	2160
1467	2161	enum compact_result compaction_suitable(struct zone *zone, int order,
1468	2162	unsigned int alloc_flags,
1469		- int classzone_idx)
	2163	+ int highest_zoneidx)
1470	2164	{
1471	2165	enum compact_result ret;
1472	2166	int fragindex;
1473	2167
1474		- ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx,
	2168	+ ret = __compaction_suitable(zone, order, alloc_flags, highest_zoneidx,
1475	2169	zone_page_state(zone, NR_FREE_PAGES));
1476	2170	/*
1477	2171	* fragmentation index determines if allocation failures are due to
..	..	@@ -1512,8 +2206,8 @@
1512	2206	* Make sure at least one zone would pass __compaction_suitable if we continue
1513	2207	* retrying the reclaim.
1514	2208	*/
1515		- for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1516		- ac->nodemask) {
	2209	+ for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
	2210	+ ac->highest_zoneidx, ac->nodemask) {
1517	2211	unsigned long available;
1518	2212	enum compact_result compact_result;
1519	2213
..	..	@@ -1526,7 +2220,7 @@
1526	2220	available = zone_reclaimable_pages(zone) / order;
1527	2221	available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
1528	2222	compact_result = __compaction_suitable(zone, order, alloc_flags,
1529		- ac_classzone_idx(ac), available);
	2223	+ ac->highest_zoneidx, available);
1530	2224	if (compact_result != COMPACT_SKIPPED)
1531	2225	return true;
1532	2226	}
..	..	@@ -1534,12 +2228,15 @@
1534	2228	return false;
1535	2229	}
1536	2230
1537		-static enum compact_result compact_zone(struct zone zone, struct compact_control cc)
	2231	+static enum compact_result
	2232	+compact_zone(struct compact_control cc, struct capture_control capc)
1538	2233	{
1539	2234	enum compact_result ret;
1540		- unsigned long start_pfn = zone->zone_start_pfn;
1541		- unsigned long end_pfn = zone_end_pfn(zone);
	2235	+ unsigned long start_pfn = cc->zone->zone_start_pfn;
	2236	+ unsigned long end_pfn = zone_end_pfn(cc->zone);
	2237	+ unsigned long last_migrated_pfn;
1542	2238	const bool sync = cc->mode != MIGRATE_ASYNC;
	2239	+ bool update_cached;
1543	2240
1544	2241	/*
1545	2242	* These counters track activities during zone compaction. Initialize
..	..	@@ -1552,9 +2249,9 @@
1552	2249	INIT_LIST_HEAD(&cc->freepages);
1553	2250	INIT_LIST_HEAD(&cc->migratepages);
1554	2251
1555		- cc->migratetype = gfpflags_to_migratetype(cc->gfp_mask);
1556		- ret = compaction_suitable(zone, cc->order, cc->alloc_flags,
1557		- cc->classzone_idx);
	2252	+ cc->migratetype = gfp_migratetype(cc->gfp_mask);
	2253	+ ret = compaction_suitable(cc->zone, cc->order, cc->alloc_flags,
	2254	+ cc->highest_zoneidx);
1558	2255	/* Compaction is likely to fail */
1559	2256	if (ret == COMPACT_SUCCESS \|\| ret == COMPACT_SKIPPED)
1560	2257	return ret;
..	..	@@ -1566,8 +2263,8 @@
1566	2263	* Clear pageblock skip if there were failures recently and compaction
1567	2264	* is about to be retried after being deferred.
1568	2265	*/
1569		- if (compaction_restarting(zone, cc->order))
1570		- __reset_isolation_suitable(zone);
	2266	+ if (compaction_restarting(cc->zone, cc->order))
	2267	+ __reset_isolation_suitable(cc->zone);
1571	2268
1572	2269	/*
1573	2270	* Setup to move all movable pages to the end of the zone. Used cached
..	..	@@ -1575,43 +2272,76 @@
1575	2272	* want to compact the whole zone), but check that it is initialised
1576	2273	* by ensuring the values are within zone boundaries.
1577	2274	*/
	2275	+ cc->fast_start_pfn = 0;
1578	2276	if (cc->whole_zone) {
1579	2277	cc->migrate_pfn = start_pfn;
1580	2278	cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
1581	2279	} else {
1582		- cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
1583		- cc->free_pfn = zone->compact_cached_free_pfn;
	2280	+ cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync];
	2281	+ cc->free_pfn = cc->zone->compact_cached_free_pfn;
1584	2282	if (cc->free_pfn < start_pfn \|\| cc->free_pfn >= end_pfn) {
1585	2283	cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
1586		- zone->compact_cached_free_pfn = cc->free_pfn;
	2284	+ cc->zone->compact_cached_free_pfn = cc->free_pfn;
1587	2285	}
1588	2286	if (cc->migrate_pfn < start_pfn \|\| cc->migrate_pfn >= end_pfn) {
1589	2287	cc->migrate_pfn = start_pfn;
1590		- zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1591		- zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
	2288	+ cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
	2289	+ cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1592	2290	}
1593	2291
1594		- if (cc->migrate_pfn == start_pfn)
	2292	+ if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn)
1595	2293	cc->whole_zone = true;
1596	2294	}
1597	2295
1598		- cc->last_migrated_pfn = 0;
	2296	+ last_migrated_pfn = 0;
	2297	+
	2298	+ /*
	2299	+ * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
	2300	+ * the basis that some migrations will fail in ASYNC mode. However,
	2301	+ * if the cached PFNs match and pageblocks are skipped due to having
	2302	+ * no isolation candidates, then the sync state does not matter.
	2303	+ * Until a pageblock with isolation candidates is found, keep the
	2304	+ * cached PFNs in sync to avoid revisiting the same blocks.
	2305	+ */
	2306	+ update_cached = !sync &&
	2307	+ cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1];
1599	2308
1600	2309	trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
1601	2310	cc->free_pfn, end_pfn, sync);
1602	2311
1603		- migrate_prep_local();
	2312	+ /* lru_add_drain_all could be expensive with involving other CPUs */
	2313	+ lru_add_drain();
1604	2314
1605		- while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
	2315	+ while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) {
1606	2316	int err;
	2317	+ unsigned long start_pfn = cc->migrate_pfn;
1607	2318
1608		- switch (isolate_migratepages(zone, cc)) {
	2319	+ /*
	2320	+ * Avoid multiple rescans which can happen if a page cannot be
	2321	+ * isolated (dirty/writeback in async mode) or if the migrated
	2322	+ * pages are being allocated before the pageblock is cleared.
	2323	+ * The first rescan will capture the entire pageblock for
	2324	+ * migration. If it fails, it'll be marked skip and scanning
	2325	+ * will proceed as normal.
	2326	+ */
	2327	+ cc->rescan = false;
	2328	+ if (pageblock_start_pfn(last_migrated_pfn) ==
	2329	+ pageblock_start_pfn(start_pfn)) {
	2330	+ cc->rescan = true;
	2331	+ }
	2332	+
	2333	+ switch (isolate_migratepages(cc)) {
1609	2334	case ISOLATE_ABORT:
1610	2335	ret = COMPACT_CONTENDED;
1611	2336	putback_movable_pages(&cc->migratepages);
1612	2337	cc->nr_migratepages = 0;
1613	2338	goto out;
1614	2339	case ISOLATE_NONE:
	2340	+ if (update_cached) {
	2341	+ cc->zone->compact_cached_migrate_pfn[1] =
	2342	+ cc->zone->compact_cached_migrate_pfn[0];
	2343	+ }
	2344	+
1615	2345	/*
1616	2346	* We haven't isolated and migrated anything, but
1617	2347	* there might still be unflushed migrations from
..	..	@@ -1619,6 +2349,8 @@
1619	2349	*/
1620	2350	goto check_drain;
1621	2351	case ISOLATE_SUCCESS:
	2352	+ update_cached = false;
	2353	+ last_migrated_pfn = start_pfn;
1622	2354	;
1623	2355	}
1624	2356
..	..	@@ -1650,8 +2382,7 @@
1650	2382	cc->migrate_pfn = block_end_pfn(
1651	2383	cc->migrate_pfn - 1, cc->order);
1652	2384	/* Draining pcplists is useless in this case */
1653		- cc->last_migrated_pfn = 0;
1654		-
	2385	+ last_migrated_pfn = 0;
1655	2386	}
1656	2387	}
1657	2388
..	..	@@ -1663,23 +2394,22 @@
1663	2394	* compact_finished() can detect immediately if allocation
1664	2395	* would succeed.
1665	2396	*/
1666		- if (cc->order > 0 && cc->last_migrated_pfn) {
1667		- int cpu;
	2397	+ if (cc->order > 0 && last_migrated_pfn) {
1668	2398	unsigned long current_block_start =
1669	2399	block_start_pfn(cc->migrate_pfn, cc->order);
1670	2400
1671		- if (cc->last_migrated_pfn < current_block_start) {
1672		- cpu = get_cpu_light();
1673		- local_lock_irq(swapvec_lock);
1674		- lru_add_drain_cpu(cpu);
1675		- local_unlock_irq(swapvec_lock);
1676		- drain_local_pages(zone);
1677		- put_cpu_light();
	2401	+ if (last_migrated_pfn < current_block_start) {
	2402	+ lru_add_drain_cpu_zone(cc->zone);
1678	2403	/* No more flushing until we migrate again */
1679		- cc->last_migrated_pfn = 0;
	2404	+ last_migrated_pfn = 0;
1680	2405	}
1681	2406	}
1682	2407
	2408	+ /* Stop if a page has been captured */
	2409	+ if (capc && capc->page) {
	2410	+ ret = COMPACT_SUCCESS;
	2411	+ break;
	2412	+ }
1683	2413	}
1684	2414
1685	2415	out:
..	..	@@ -1698,8 +2428,8 @@
1698	2428	* Only go back, not forward. The cached pfn might have been
1699	2429	* already reset to zone end in compact_finished()
1700	2430	*/
1701		- if (free_pfn > zone->compact_cached_free_pfn)
1702		- zone->compact_cached_free_pfn = free_pfn;
	2431	+ if (free_pfn > cc->zone->compact_cached_free_pfn)
	2432	+ cc->zone->compact_cached_free_pfn = free_pfn;
1703	2433	}
1704	2434
1705	2435	count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
..	..	@@ -1713,27 +2443,49 @@
1713	2443
1714	2444	static enum compact_result compact_zone_order(struct zone *zone, int order,
1715	2445	gfp_t gfp_mask, enum compact_priority prio,
1716		- unsigned int alloc_flags, int classzone_idx)
	2446	+ unsigned int alloc_flags, int highest_zoneidx,
	2447	+ struct page **capture)
1717	2448	{
1718	2449	enum compact_result ret;
1719	2450	struct compact_control cc = {
1720	2451	.order = order,
	2452	+ .search_order = order,
1721	2453	.gfp_mask = gfp_mask,
1722	2454	.zone = zone,
1723	2455	.mode = (prio == COMPACT_PRIO_ASYNC) ?
1724	2456	MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
1725	2457	.alloc_flags = alloc_flags,
1726		- .classzone_idx = classzone_idx,
	2458	+ .highest_zoneidx = highest_zoneidx,
1727	2459	.direct_compaction = true,
1728	2460	.whole_zone = (prio == MIN_COMPACT_PRIORITY),
1729	2461	.ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
1730	2462	.ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
1731	2463	};
	2464	+ struct capture_control capc = {
	2465	+ .cc = &cc,
	2466	+ .page = NULL,
	2467	+ };
1732	2468
1733		- ret = compact_zone(zone, &cc);
	2469	+ /*
	2470	+ * Make sure the structs are really initialized before we expose the
	2471	+ * capture control, in case we are interrupted and the interrupt handler
	2472	+ * frees a page.
	2473	+ */
	2474	+ barrier();
	2475	+ WRITE_ONCE(current->capture_control, &capc);
	2476	+
	2477	+ ret = compact_zone(&cc, &capc);
1734	2478
1735	2479	VM_BUG_ON(!list_empty(&cc.freepages));
1736	2480	VM_BUG_ON(!list_empty(&cc.migratepages));
	2481	+
	2482	+ /*
	2483	+ * Make sure we hide capture control first before we read the captured
	2484	+ * page pointer, otherwise an interrupt could free and capture a page
	2485	+ * and we would leak it.
	2486	+ */
	2487	+ WRITE_ONCE(current->capture_control, NULL);
	2488	+ *capture = READ_ONCE(capc.page);
1737	2489
1738	2490	return ret;
1739	2491	}
..	..	@@ -1747,12 +2499,13 @@
1747	2499	* @alloc_flags: The allocation flags of the current allocation
1748	2500	* @ac: The context of current allocation
1749	2501	* @prio: Determines how hard direct compaction should try to succeed
	2502	+ * @capture: Pointer to free page created by compaction will be stored here
1750	2503	*
1751	2504	* This is the main entry point for direct page compaction.
1752	2505	*/
1753	2506	enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
1754	2507	unsigned int alloc_flags, const struct alloc_context *ac,
1755		- enum compact_priority prio)
	2508	+ enum compact_priority prio, struct page **capture)
1756	2509	{
1757	2510	int may_perform_io = gfp_mask & __GFP_IO;
1758	2511	struct zoneref *z;
..	..	@@ -1769,8 +2522,8 @@
1769	2522	trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
1770	2523
1771	2524	/* Compact each zone in the list */
1772		- for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1773		- ac->nodemask) {
	2525	+ for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
	2526	+ ac->highest_zoneidx, ac->nodemask) {
1774	2527	enum compact_result status;
1775	2528
1776	2529	if (prio > MIN_COMPACT_PRIORITY
..	..	@@ -1780,7 +2533,7 @@
1780	2533	}
1781	2534
1782	2535	status = compact_zone_order(zone, order, gfp_mask, prio,
1783		- alloc_flags, ac_classzone_idx(ac));
	2536	+ alloc_flags, ac->highest_zoneidx, capture);
1784	2537	rc = max(status, rc);
1785	2538
1786	2539	/* The allocation should succeed, stop compacting */
..	..	@@ -1818,6 +2571,41 @@
1818	2571	return rc;
1819	2572	}
1820	2573
	2574	+/*
	2575	+ * Compact all zones within a node till each zone's fragmentation score
	2576	+ * reaches within proactive compaction thresholds (as determined by the
	2577	+ * proactiveness tunable).
	2578	+ *
	2579	+ * It is possible that the function returns before reaching score targets
	2580	+ * due to various back-off conditions, such as, contention on per-node or
	2581	+ * per-zone locks.
	2582	+ */
	2583	+static void proactive_compact_node(pg_data_t *pgdat)
	2584	+{
	2585	+ int zoneid;
	2586	+ struct zone *zone;
	2587	+ struct compact_control cc = {
	2588	+ .order = -1,
	2589	+ .mode = MIGRATE_SYNC_LIGHT,
	2590	+ .ignore_skip_hint = true,
	2591	+ .whole_zone = true,
	2592	+ .gfp_mask = GFP_KERNEL,
	2593	+ .proactive_compaction = true,
	2594	+ };
	2595	+
	2596	+ for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
	2597	+ zone = &pgdat->node_zones[zoneid];
	2598	+ if (!populated_zone(zone))
	2599	+ continue;
	2600	+
	2601	+ cc.zone = zone;
	2602	+
	2603	+ compact_zone(&cc, NULL);
	2604	+
	2605	+ VM_BUG_ON(!list_empty(&cc.freepages));
	2606	+ VM_BUG_ON(!list_empty(&cc.migratepages));
	2607	+ }
	2608	+}
1821	2609
1822	2610	/* Compact all zones within a node */
1823	2611	static void compact_node(int nid)
..	..	@@ -1842,7 +2630,7 @@
1842	2630
1843	2631	cc.zone = zone;
1844	2632
1845		- compact_zone(zone, &cc);
	2633	+ compact_zone(&cc, NULL);
1846	2634
1847	2635	VM_BUG_ON(!list_empty(&cc.freepages));
1848	2636	VM_BUG_ON(!list_empty(&cc.migratepages));
..	..	@@ -1865,22 +2653,45 @@
1865	2653	int sysctl_compact_memory;
1866	2654
1867	2655	/*
1868		- * This is the entry point for compacting all nodes via
1869		- * /proc/sys/vm/compact_memory
	2656	+ * Tunable for proactive compaction. It determines how
	2657	+ * aggressively the kernel should compact memory in the
	2658	+ * background. It takes values in the range [0, 100].
1870	2659	*/
1871		-int sysctl_compaction_handler(struct ctl_table *table, int write,
1872		- void __user buffer, size_t length, loff_t *ppos)
	2660	+unsigned int __read_mostly sysctl_compaction_proactiveness = 20;
	2661	+
	2662	+int compaction_proactiveness_sysctl_handler(struct ctl_table *table, int write,
	2663	+ void buffer, size_t length, loff_t *ppos)
1873	2664	{
1874		- if (write)
1875		- compact_nodes();
	2665	+ int rc, nid;
	2666	+
	2667	+ rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
	2668	+ if (rc)
	2669	+ return rc;
	2670	+
	2671	+ if (write && sysctl_compaction_proactiveness) {
	2672	+ for_each_online_node(nid) {
	2673	+ pg_data_t *pgdat = NODE_DATA(nid);
	2674	+
	2675	+ if (pgdat->proactive_compact_trigger)
	2676	+ continue;
	2677	+
	2678	+ pgdat->proactive_compact_trigger = true;
	2679	+ wake_up_interruptible(&pgdat->kcompactd_wait);
	2680	+ }
	2681	+ }
1876	2682
1877	2683	return 0;
1878	2684	}
1879	2685
1880		-int sysctl_extfrag_handler(struct ctl_table *table, int write,
1881		- void __user buffer, size_t length, loff_t *ppos)
	2686	+/*
	2687	+ * This is the entry point for compacting all nodes via
	2688	+ * /proc/sys/vm/compact_memory
	2689	+ */
	2690	+int sysctl_compaction_handler(struct ctl_table *table, int write,
	2691	+ void buffer, size_t length, loff_t *ppos)
1882	2692	{
1883		- proc_dointvec_minmax(table, write, buffer, length, ppos);
	2693	+ if (write)
	2694	+ compact_nodes();
1884	2695
1885	2696	return 0;
1886	2697	}
..	..	@@ -1916,23 +2727,24 @@
1916	2727
1917	2728	static inline bool kcompactd_work_requested(pg_data_t *pgdat)
1918	2729	{
1919		- return pgdat->kcompactd_max_order > 0 \|\| kthread_should_stop();
	2730	+ return pgdat->kcompactd_max_order > 0 \|\| kthread_should_stop() \|\|
	2731	+ pgdat->proactive_compact_trigger;
1920	2732	}
1921	2733
1922	2734	static bool kcompactd_node_suitable(pg_data_t *pgdat)
1923	2735	{
1924	2736	int zoneid;
1925	2737	struct zone *zone;
1926		- enum zone_type classzone_idx = pgdat->kcompactd_classzone_idx;
	2738	+ enum zone_type highest_zoneidx = pgdat->kcompactd_highest_zoneidx;
1927	2739
1928		- for (zoneid = 0; zoneid <= classzone_idx; zoneid++) {
	2740	+ for (zoneid = 0; zoneid <= highest_zoneidx; zoneid++) {
1929	2741	zone = &pgdat->node_zones[zoneid];
1930	2742
1931	2743	if (!populated_zone(zone))
1932	2744	continue;
1933	2745
1934	2746	if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0,
1935		- classzone_idx) == COMPACT_CONTINUE)
	2747	+ highest_zoneidx) == COMPACT_CONTINUE)
1936	2748	return true;
1937	2749	}
1938	2750
..	..	@@ -1949,16 +2761,17 @@
1949	2761	struct zone *zone;
1950	2762	struct compact_control cc = {
1951	2763	.order = pgdat->kcompactd_max_order,
1952		- .classzone_idx = pgdat->kcompactd_classzone_idx,
	2764	+ .search_order = pgdat->kcompactd_max_order,
	2765	+ .highest_zoneidx = pgdat->kcompactd_highest_zoneidx,
1953	2766	.mode = MIGRATE_SYNC_LIGHT,
1954	2767	.ignore_skip_hint = false,
1955	2768	.gfp_mask = GFP_KERNEL,
1956	2769	};
1957	2770	trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
1958		- cc.classzone_idx);
	2771	+ cc.highest_zoneidx);
1959	2772	count_compact_event(KCOMPACTD_WAKE);
1960	2773
1961		- for (zoneid = 0; zoneid <= cc.classzone_idx; zoneid++) {
	2774	+ for (zoneid = 0; zoneid <= cc.highest_zoneidx; zoneid++) {
1962	2775	int status;
1963	2776
1964	2777	zone = &pgdat->node_zones[zoneid];
..	..	@@ -1976,7 +2789,7 @@
1976	2789	return;
1977	2790
1978	2791	cc.zone = zone;
1979		- status = compact_zone(zone, &cc);
	2792	+ status = compact_zone(&cc, NULL);
1980	2793
1981	2794	if (status == COMPACT_SUCCESS) {
1982	2795	compaction_defer_reset(zone, cc.order, false);
..	..	@@ -2007,16 +2820,16 @@
2007	2820
2008	2821	/*
2009	2822	* Regardless of success, we are done until woken up next. But remember
2010		- * the requested order/classzone_idx in case it was higher/tighter than
2011		- * our current ones
	2823	+ * the requested order/highest_zoneidx in case it was higher/tighter
	2824	+ * than our current ones
2012	2825	*/
2013	2826	if (pgdat->kcompactd_max_order <= cc.order)
2014	2827	pgdat->kcompactd_max_order = 0;
2015		- if (pgdat->kcompactd_classzone_idx >= cc.classzone_idx)
2016		- pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
	2828	+ if (pgdat->kcompactd_highest_zoneidx >= cc.highest_zoneidx)
	2829	+ pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
2017	2830	}
2018	2831
2019		-void wakeup_kcompactd(pg_data_t *pgdat, int order, int classzone_idx)
	2832	+void wakeup_kcompactd(pg_data_t *pgdat, int order, int highest_zoneidx)
2020	2833	{
2021	2834	if (!order)
2022	2835	return;
..	..	@@ -2024,8 +2837,8 @@
2024	2837	if (pgdat->kcompactd_max_order < order)
2025	2838	pgdat->kcompactd_max_order = order;
2026	2839
2027		- if (pgdat->kcompactd_classzone_idx > classzone_idx)
2028		- pgdat->kcompactd_classzone_idx = classzone_idx;
	2840	+ if (pgdat->kcompactd_highest_zoneidx > highest_zoneidx)
	2841	+ pgdat->kcompactd_highest_zoneidx = highest_zoneidx;
2029	2842
2030	2843	/*
2031	2844	* Pairs with implicit barrier in wait_event_freezable()
..	..	@@ -2038,7 +2851,7 @@
2038	2851	return;
2039	2852
2040	2853	trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
2041		- classzone_idx);
	2854	+ highest_zoneidx);
2042	2855	wake_up_interruptible(&pgdat->kcompactd_wait);
2043	2856	}
2044	2857
..	..	@@ -2050,6 +2863,7 @@
2050	2863	{
2051	2864	pg_data_t pgdat = (pg_data_t)p;
2052	2865	struct task_struct *tsk = current;
	2866	+ unsigned int proactive_defer = 0;
2053	2867
2054	2868	const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2055	2869
..	..	@@ -2059,18 +2873,56 @@
2059	2873	set_freezable();
2060	2874
2061	2875	pgdat->kcompactd_max_order = 0;
2062		- pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
	2876	+ pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
2063	2877
2064	2878	while (!kthread_should_stop()) {
2065	2879	unsigned long pflags;
	2880	+ long timeout;
2066	2881
	2882	+ timeout = sysctl_compaction_proactiveness ?
	2883	+ msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC) :
	2884	+ MAX_SCHEDULE_TIMEOUT;
2067	2885	trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
2068		- wait_event_freezable(pgdat->kcompactd_wait,
2069		- kcompactd_work_requested(pgdat));
	2886	+ if (wait_event_freezable_timeout(pgdat->kcompactd_wait,
	2887	+ kcompactd_work_requested(pgdat), timeout) &&
	2888	+ !pgdat->proactive_compact_trigger) {
2070	2889
2071		- psi_memstall_enter(&pflags);
2072		- kcompactd_do_work(pgdat);
2073		- psi_memstall_leave(&pflags);
	2890	+ psi_memstall_enter(&pflags);
	2891	+ kcompactd_do_work(pgdat);
	2892	+ psi_memstall_leave(&pflags);
	2893	+ continue;
	2894	+ }
	2895	+
	2896	+ /* kcompactd wait timeout */
	2897	+ if (should_proactive_compact_node(pgdat)) {
	2898	+ unsigned int prev_score, score;
	2899	+
	2900	+ /*
	2901	+ * On wakeup of proactive compaction by sysctl
	2902	+ * write, ignore the accumulated defer score.
	2903	+ * Anyway, if the proactive compaction didn't
	2904	+ * make any progress for the new value, it will
	2905	+ * be further deferred by 2^COMPACT_MAX_DEFER_SHIFT
	2906	+ * times.
	2907	+ */
	2908	+ if (proactive_defer &&
	2909	+ !pgdat->proactive_compact_trigger) {
	2910	+ proactive_defer--;
	2911	+ continue;
	2912	+ }
	2913	+
	2914	+ prev_score = fragmentation_score_node(pgdat);
	2915	+ proactive_compact_node(pgdat);
	2916	+ score = fragmentation_score_node(pgdat);
	2917	+ /*
	2918	+ * Defer proactive compaction if the fragmentation
	2919	+ * score did not go down i.e. no progress made.
	2920	+ */
	2921	+ proactive_defer = score < prev_score ?
	2922	+ 0 : 1 << COMPACT_MAX_DEFER_SHIFT;
	2923	+ }
	2924	+ if (pgdat->proactive_compact_trigger)
	2925	+ pgdat->proactive_compact_trigger = false;
2074	2926	}
2075	2927
2076	2928	return 0;