~hc/RK356X_SDK_RELEASE.git

..	..	@@ -50,6 +50,24 @@
50	50	#define pageblock_start_pfn(pfn) block_start_pfn(pfn, pageblock_order)
51	51	#define pageblock_end_pfn(pfn) block_end_pfn(pfn, pageblock_order)
52	52
	53	+/*
	54	+ * Fragmentation score check interval for proactive compaction purposes.
	55	+ */
	56	+static const unsigned int HPAGE_FRAG_CHECK_INTERVAL_MSEC = 500;
	57	+
	58	+/*
	59	+ * Page order with-respect-to which proactive compaction
	60	+ * calculates external fragmentation, which is used as
	61	+ * the "fragmentation score" of a node/zone.
	62	+ */
	63	+#if defined CONFIG_TRANSPARENT_HUGEPAGE
	64	+#define COMPACTION_HPAGE_ORDER HPAGE_PMD_ORDER
	65	+#elif defined CONFIG_HUGETLBFS
	66	+#define COMPACTION_HPAGE_ORDER HUGETLB_PAGE_ORDER
	67	+#else
	68	+#define COMPACTION_HPAGE_ORDER (PMD_SHIFT - PAGE_SHIFT)
	69	+#endif
	70	+
53	71	static unsigned long release_freepages(struct list_head *freelist)
54	72	{
55	73	struct page page, next;
..	..	@@ -66,7 +84,7 @@
66	84	return high_pfn;
67	85	}
68	86
69		-static void map_pages(struct list_head *list)
	87	+static void split_map_pages(struct list_head *list)
70	88	{
71	89	unsigned int i, order, nr_pages;
72	90	struct page page, next;
..	..	@@ -136,7 +154,7 @@
136	154
137	155	/*
138	156	* Compaction is deferred when compaction fails to result in a page
139		- * allocation success. 1 << compact_defer_limit compactions are skipped up
	157	+ * allocation success. 1 << compact_defer_shift, compactions are skipped up
140	158	* to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
141	159	*/
142	160	void defer_compaction(struct zone *zone, int order)
..	..	@@ -162,11 +180,10 @@
162	180	return false;
163	181
164	182	/* Avoid possible overflow */
165		- if (++zone->compact_considered > defer_limit)
	183	+ if (++zone->compact_considered >= defer_limit) {
166	184	zone->compact_considered = defer_limit;
167		-
168		- if (zone->compact_considered >= defer_limit)
169	185	return false;
	186	+ }
170	187
171	188	trace_mm_compaction_deferred(zone, order);
172	189
..	..	@@ -237,6 +254,78 @@
237	254	return false;
238	255	}
239	256
	257	+static bool
	258	+__reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source,
	259	+ bool check_target)
	260	+{
	261	+ struct page *page = pfn_to_online_page(pfn);
	262	+ struct page *block_page;
	263	+ struct page *end_page;
	264	+ unsigned long block_pfn;
	265	+
	266	+ if (!page)
	267	+ return false;
	268	+ if (zone != page_zone(page))
	269	+ return false;
	270	+ if (pageblock_skip_persistent(page))
	271	+ return false;
	272	+
	273	+ /*
	274	+ * If skip is already cleared do no further checking once the
	275	+ * restart points have been set.
	276	+ */
	277	+ if (check_source && check_target && !get_pageblock_skip(page))
	278	+ return true;
	279	+
	280	+ /*
	281	+ * If clearing skip for the target scanner, do not select a
	282	+ * non-movable pageblock as the starting point.
	283	+ */
	284	+ if (!check_source && check_target &&
	285	+ get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
	286	+ return false;
	287	+
	288	+ /* Ensure the start of the pageblock or zone is online and valid */
	289	+ block_pfn = pageblock_start_pfn(pfn);
	290	+ block_pfn = max(block_pfn, zone->zone_start_pfn);
	291	+ block_page = pfn_to_online_page(block_pfn);
	292	+ if (block_page) {
	293	+ page = block_page;
	294	+ pfn = block_pfn;
	295	+ }
	296	+
	297	+ /* Ensure the end of the pageblock or zone is online and valid */
	298	+ block_pfn = pageblock_end_pfn(pfn) - 1;
	299	+ block_pfn = min(block_pfn, zone_end_pfn(zone) - 1);
	300	+ end_page = pfn_to_online_page(block_pfn);
	301	+ if (!end_page)
	302	+ return false;
	303	+
	304	+ /*
	305	+ * Only clear the hint if a sample indicates there is either a
	306	+ * free page or an LRU page in the block. One or other condition
	307	+ * is necessary for the block to be a migration source/target.
	308	+ */
	309	+ do {
	310	+ if (pfn_valid_within(pfn)) {
	311	+ if (check_source && PageLRU(page)) {
	312	+ clear_pageblock_skip(page);
	313	+ return true;
	314	+ }
	315	+
	316	+ if (check_target && PageBuddy(page)) {
	317	+ clear_pageblock_skip(page);
	318	+ return true;
	319	+ }
	320	+ }
	321	+
	322	+ page += (1 << PAGE_ALLOC_COSTLY_ORDER);
	323	+ pfn += (1 << PAGE_ALLOC_COSTLY_ORDER);
	324	+ } while (page <= end_page);
	325	+
	326	+ return false;
	327	+}
	328	+
240	329	/*
241	330	* This function is called to clear all cached information on pageblocks that
242	331	* should be skipped for page isolation when the migrate and free page scanner
..	..	@@ -244,30 +333,54 @@
244	333	*/
245	334	static void __reset_isolation_suitable(struct zone *zone)
246	335	{
247		- unsigned long start_pfn = zone->zone_start_pfn;
248		- unsigned long end_pfn = zone_end_pfn(zone);
249		- unsigned long pfn;
	336	+ unsigned long migrate_pfn = zone->zone_start_pfn;
	337	+ unsigned long free_pfn = zone_end_pfn(zone) - 1;
	338	+ unsigned long reset_migrate = free_pfn;
	339	+ unsigned long reset_free = migrate_pfn;
	340	+ bool source_set = false;
	341	+ bool free_set = false;
	342	+
	343	+ if (!zone->compact_blockskip_flush)
	344	+ return;
250	345
251	346	zone->compact_blockskip_flush = false;
252	347
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		-
	348	+ /*
	349	+ * Walk the zone and update pageblock skip information. Source looks
	350	+ * for PageLRU while target looks for PageBuddy. When the scanner
	351	+ * is found, both PageBuddy and PageLRU are checked as the pageblock
	352	+ * is suitable as both source and target.
	353	+ */
	354	+ for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages,
	355	+ free_pfn -= pageblock_nr_pages) {
257	356	cond_resched();
258	357
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;
	358	+ /* Update the migrate PFN */
	359	+ if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) &&
	360	+ migrate_pfn < reset_migrate) {
	361	+ source_set = true;
	362	+ reset_migrate = migrate_pfn;
	363	+ zone->compact_init_migrate_pfn = reset_migrate;
	364	+ zone->compact_cached_migrate_pfn[0] = reset_migrate;
	365	+ zone->compact_cached_migrate_pfn[1] = reset_migrate;
	366	+ }
266	367
267		- clear_pageblock_skip(page);
	368	+ /* Update the free PFN */
	369	+ if (__reset_isolation_pfn(zone, free_pfn, free_set, true) &&
	370	+ free_pfn > reset_free) {
	371	+ free_set = true;
	372	+ reset_free = free_pfn;
	373	+ zone->compact_init_free_pfn = reset_free;
	374	+ zone->compact_cached_free_pfn = reset_free;
	375	+ }
268	376	}
269	377
270		- reset_cached_positions(zone);
	378	+ /* Leave no distance if no suitable block was reset */
	379	+ if (reset_migrate >= reset_free) {
	380	+ zone->compact_cached_migrate_pfn[0] = migrate_pfn;
	381	+ zone->compact_cached_migrate_pfn[1] = migrate_pfn;
	382	+ zone->compact_cached_free_pfn = free_pfn;
	383	+ }
271	384	}
272	385
273	386	void reset_isolation_suitable(pg_data_t *pgdat)
..	..	@@ -286,15 +399,53 @@
286	399	}
287	400
288	401	/*
	402	+ * Sets the pageblock skip bit if it was clear. Note that this is a hint as
	403	+ * locks are not required for read/writers. Returns true if it was already set.
	404	+ */
	405	+static bool test_and_set_skip(struct compact_control cc, struct page page,
	406	+ unsigned long pfn)
	407	+{
	408	+ bool skip;
	409	+
	410	+ /* Do no update if skip hint is being ignored */
	411	+ if (cc->ignore_skip_hint)
	412	+ return false;
	413	+
	414	+ if (!IS_ALIGNED(pfn, pageblock_nr_pages))
	415	+ return false;
	416	+
	417	+ skip = get_pageblock_skip(page);
	418	+ if (!skip && !cc->no_set_skip_hint)
	419	+ set_pageblock_skip(page);
	420	+
	421	+ return skip;
	422	+}
	423	+
	424	+static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
	425	+{
	426	+ struct zone *zone = cc->zone;
	427	+
	428	+ pfn = pageblock_end_pfn(pfn);
	429	+
	430	+ /* Set for isolation rather than compaction */
	431	+ if (cc->no_set_skip_hint)
	432	+ return;
	433	+
	434	+ if (pfn > zone->compact_cached_migrate_pfn[0])
	435	+ zone->compact_cached_migrate_pfn[0] = pfn;
	436	+ if (cc->mode != MIGRATE_ASYNC &&
	437	+ pfn > zone->compact_cached_migrate_pfn[1])
	438	+ zone->compact_cached_migrate_pfn[1] = pfn;
	439	+}
	440	+
	441	+/*
289	442	* If no pages were isolated then mark this pageblock to be skipped in the
290	443	* future. The information is later cleared by __reset_isolation_suitable().
291	444	*/
292	445	static void update_pageblock_skip(struct compact_control *cc,
293		- struct page *page, unsigned long nr_isolated,
294		- bool migrate_scanner)
	446	+ struct page *page, unsigned long pfn)
295	447	{
296	448	struct zone *zone = cc->zone;
297		- unsigned long pfn;
298	449
299	450	if (cc->no_set_skip_hint)
300	451	return;
..	..	@@ -302,24 +453,11 @@
302	453	if (!page)
303	454	return;
304	455
305		- if (nr_isolated)
306		- return;
307		-
308	456	set_pageblock_skip(page);
309	457
310		- pfn = page_to_pfn(page);
311		-
312	458	/* 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		- }
	459	+ if (pfn < zone->compact_cached_free_pfn)
	460	+ zone->compact_cached_free_pfn = pfn;
323	461	}
324	462	#else
325	463	static inline bool isolation_suitable(struct compact_control *cc,
..	..	@@ -334,32 +472,43 @@
334	472	}
335	473
336	474	static inline void update_pageblock_skip(struct compact_control *cc,
337		- struct page *page, unsigned long nr_isolated,
338		- bool migrate_scanner)
	475	+ struct page *page, unsigned long pfn)
339	476	{
	477	+}
	478	+
	479	+static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
	480	+{
	481	+}
	482	+
	483	+static bool test_and_set_skip(struct compact_control cc, struct page page,
	484	+ unsigned long pfn)
	485	+{
	486	+ return false;
340	487	}
341	488	#endif /* CONFIG_COMPACTION */
342	489
343	490	/*
344	491	* 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.
	492	+ * very heavily contended. For async compaction, trylock and record if the
	493	+ * lock is contended. The lock will still be acquired but compaction will
	494	+ * abort when the current block is finished regardless of success rate.
	495	+ * Sync compaction acquires the lock.
347	496	*
348		- * Returns true if the lock is held
349		- * Returns false if the lock is not held and compaction should abort
	497	+ * Always returns true which makes it easier to track lock state in callers.
350	498	*/
351		-static bool compact_trylock_irqsave(spinlock_t lock, unsigned long flags,
	499	+static bool compact_lock_irqsave(spinlock_t lock, unsigned long flags,
352	500	struct compact_control *cc)
	501	+ __acquires(lock)
353	502	{
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);
	503	+ /* Track if the lock is contended in async mode */
	504	+ if (cc->mode == MIGRATE_ASYNC && !cc->contended) {
	505	+ if (spin_trylock_irqsave(lock, *flags))
	506	+ return true;
	507	+
	508	+ cc->contended = true;
361	509	}
362	510
	511	+ spin_lock_irqsave(lock, *flags);
363	512	return true;
364	513	}
365	514
..	..	@@ -391,37 +540,7 @@
391	540	return true;
392	541	}
393	542
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		- }
	543	+ cond_resched();
425	544
426	545	return false;
427	546	}
..	..	@@ -435,19 +554,24 @@
435	554	unsigned long *start_pfn,
436	555	unsigned long end_pfn,
437	556	struct list_head *freelist,
	557	+ unsigned int stride,
438	558	bool strict)
439	559	{
440	560	int nr_scanned = 0, total_isolated = 0;
441		- struct page cursor, valid_page = NULL;
	561	+ struct page *cursor;
442	562	unsigned long flags = 0;
443	563	bool locked = false;
444	564	unsigned long blockpfn = *start_pfn;
445	565	unsigned int order;
446	566
	567	+ /* Strict mode is for isolation, speed is secondary */
	568	+ if (strict)
	569	+ stride = 1;
	570	+
447	571	cursor = pfn_to_page(blockpfn);
448	572
449	573	/* Isolate free pages. */
450		- for (; blockpfn < end_pfn; blockpfn++, cursor++) {
	574	+ for (; blockpfn < end_pfn; blockpfn += stride, cursor += stride) {
451	575	int isolated;
452	576	struct page *page = cursor;
453	577
..	..	@@ -464,9 +588,6 @@
464	588	nr_scanned++;
465	589	if (!pfn_valid_within(blockpfn))
466	590	goto isolate_fail;
467		-
468		- if (!valid_page)
469		- valid_page = page;
470	591
471	592	/*
472	593	* For compound pages such as THP and hugetlbfs, we can save
..	..	@@ -495,18 +616,8 @@
495	616	* recheck as well.
496	617	*/
497	618	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,
	619	+ locked = compact_lock_irqsave(&cc->zone->lock,
507	620	&flags, cc);
508		- if (!locked)
509		- break;
510	621
511	622	/* Recheck this is a buddy page under lock */
512	623	if (!PageBuddy(page))
..	..	@@ -514,7 +625,7 @@
514	625	}
515	626
516	627	/* Found a free page, will break it into order-0 pages */
517		- order = page_order(page);
	628	+ order = buddy_order(page);
518	629	isolated = __isolate_free_page(page, order);
519	630	if (!isolated)
520	631	break;
..	..	@@ -564,10 +675,6 @@
564	675	*/
565	676	if (strict && blockpfn < end_pfn)
566	677	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	678
572	679	cc->total_free_scanned += nr_scanned;
573	680	if (total_isolated)
..	..	@@ -626,7 +733,7 @@
626	733	break;
627	734
628	735	isolated = isolate_freepages_block(cc, &isolate_start_pfn,
629		- block_end_pfn, &freelist, true);
	736	+ block_end_pfn, &freelist, 0, true);
630	737
631	738	/*
632	739	* In strict mode, isolate_freepages_block() returns 0 if
..	..	@@ -644,7 +751,7 @@
644	751	}
645	752
646	753	/* __isolate_free_page() does not map the pages */
647		- map_pages(&freelist);
	754	+ split_map_pages(&freelist);
648	755
649	756	if (pfn < end_pfn) {
650	757	/* Loop terminated early, cleanup. */
..	..	@@ -656,17 +763,42 @@
656	763	return pfn;
657	764	}
658	765
	766	+#ifdef CONFIG_COMPACTION
	767	+unsigned long isolate_and_split_free_page(struct page *page,
	768	+ struct list_head *list)
	769	+{
	770	+ unsigned long isolated;
	771	+ unsigned int order;
	772	+
	773	+ if (!PageBuddy(page))
	774	+ return 0;
	775	+
	776	+ order = buddy_order(page);
	777	+ isolated = __isolate_free_page(page, order);
	778	+ if (!isolated)
	779	+ return 0;
	780	+
	781	+ set_page_private(page, order);
	782	+ list_add(&page->lru, list);
	783	+
	784	+ split_map_pages(list);
	785	+
	786	+ return isolated;
	787	+}
	788	+EXPORT_SYMBOL_GPL(isolate_and_split_free_page);
	789	+#endif
	790	+
659	791	/* Similar to reclaim, but different enough that they don't share logic */
660		-static bool too_many_isolated(struct zone *zone)
	792	+static bool too_many_isolated(pg_data_t *pgdat)
661	793	{
662	794	unsigned long active, inactive, isolated;
663	795
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);
	796	+ inactive = node_page_state(pgdat, NR_INACTIVE_FILE) +
	797	+ node_page_state(pgdat, NR_INACTIVE_ANON);
	798	+ active = node_page_state(pgdat, NR_ACTIVE_FILE) +
	799	+ node_page_state(pgdat, NR_ACTIVE_ANON);
	800	+ isolated = node_page_state(pgdat, NR_ISOLATED_FILE) +
	801	+ node_page_state(pgdat, NR_ISOLATED_ANON);
670	802
671	803	return isolated > (inactive + active) / 2;
672	804	}
..	..	@@ -693,7 +825,7 @@
693	825	isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
694	826	unsigned long end_pfn, isolate_mode_t isolate_mode)
695	827	{
696		- struct zone *zone = cc->zone;
	828	+ pg_data_t *pgdat = cc->zone->zone_pgdat;
697	829	unsigned long nr_scanned = 0, nr_isolated = 0;
698	830	struct lruvec *lruvec;
699	831	unsigned long flags = 0;
..	..	@@ -702,13 +834,18 @@
702	834	unsigned long start_pfn = low_pfn;
703	835	bool skip_on_failure = false;
704	836	unsigned long next_skip_pfn = 0;
	837	+ bool skip_updated = false;
705	838
706	839	/*
707	840	* Ensure that there are not too many pages isolated from the LRU
708	841	* list by either parallel reclaimers or compaction. If there are,
709	842	* delay for some time until fewer pages are isolated
710	843	*/
711		- while (unlikely(too_many_isolated(zone))) {
	844	+ while (unlikely(too_many_isolated(pgdat))) {
	845	+ /* stop isolation if there are still pages not migrated */
	846	+ if (cc->nr_migratepages)
	847	+ return 0;
	848	+
712	849	/* async migration should just abort */
713	850	if (cc->mode == MIGRATE_ASYNC)
714	851	return 0;
..	..	@@ -719,8 +856,7 @@
719	856	return 0;
720	857	}
721	858
722		- if (compact_should_abort(cc))
723		- return 0;
	859	+ cond_resched();
724	860
725	861	if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
726	862	skip_on_failure = true;
..	..	@@ -754,13 +890,15 @@
754	890
755	891	/*
756	892	* Periodically drop the lock (if held) regardless of its
757		- * contention, to give chance to IRQs. Abort async compaction
758		- * if contended.
	893	+ * contention, to give chance to IRQs. Abort completely if
	894	+ * a fatal signal is pending.
759	895	*/
760	896	if (!(low_pfn % SWAP_CLUSTER_MAX)
761		- && compact_unlock_should_abort(zone_lru_lock(zone), flags,
762		- &locked, cc))
763		- break;
	897	+ && compact_unlock_should_abort(&pgdat->lru_lock,
	898	+ flags, &locked, cc)) {
	899	+ low_pfn = 0;
	900	+ goto fatal_pending;
	901	+ }
764	902
765	903	if (!pfn_valid_within(low_pfn))
766	904	goto isolate_fail;
..	..	@@ -768,8 +906,19 @@
768	906
769	907	page = pfn_to_page(low_pfn);
770	908
771		- if (!valid_page)
	909	+ /*
	910	+ * Check if the pageblock has already been marked skipped.
	911	+ * Only the aligned PFN is checked as the caller isolates
	912	+ * COMPACT_CLUSTER_MAX at a time so the second call must
	913	+ * not falsely conclude that the block should be skipped.
	914	+ */
	915	+ if (!valid_page && IS_ALIGNED(low_pfn, pageblock_nr_pages)) {
	916	+ if (!cc->ignore_skip_hint && get_pageblock_skip(page)) {
	917	+ low_pfn = end_pfn;
	918	+ goto isolate_abort;
	919	+ }
772	920	valid_page = page;
	921	+ }
773	922
774	923	/*
775	924	* Skip if free. We read page order here without zone lock
..	..	@@ -778,7 +927,7 @@
778	927	* potential isolation targets.
779	928	*/
780	929	if (PageBuddy(page)) {
781		- unsigned long freepage_order = page_order_unsafe(page);
	930	+ unsigned long freepage_order = buddy_order_unsafe(page);
782	931
783	932	/*
784	933	* Without lock, we cannot be sure that what we got is
..	..	@@ -792,12 +941,13 @@
792	941
793	942	/*
794	943	* 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.
	944	+ * hugetlbfs are not to be compacted unless we are attempting
	945	+ * an allocation much larger than the huge page size (eg CMA).
	946	+ * We can potentially save a lot of iterations if we skip them
	947	+ * at once. The check is racy, but we can consider only valid
	948	+ * values and the only danger is skipping too much.
799	949	*/
800		- if (PageCompound(page)) {
	950	+ if (PageCompound(page) && !cc->alloc_contig) {
801	951	const unsigned int order = compound_order(page);
802	952
803	953	if (likely(order < MAX_ORDER))
..	..	@@ -818,7 +968,7 @@
818	968	if (unlikely(__PageMovable(page)) &&
819	969	!PageIsolated(page)) {
820	970	if (locked) {
821		- spin_unlock_irqrestore(zone_lru_lock(zone),
	971	+ spin_unlock_irqrestore(&pgdat->lru_lock,
822	972	flags);
823	973	locked = false;
824	974	}
..	..	@@ -848,10 +998,15 @@
848	998
849	999	/* If we already hold the lock, we can skip some rechecking */
850	1000	if (!locked) {
851		- locked = compact_trylock_irqsave(zone_lru_lock(zone),
	1001	+ locked = compact_lock_irqsave(&pgdat->lru_lock,
852	1002	&flags, cc);
853		- if (!locked)
854		- break;
	1003	+
	1004	+ /* Try get exclusive access under lock */
	1005	+ if (!skip_updated) {
	1006	+ skip_updated = true;
	1007	+ if (test_and_set_skip(cc, page, low_pfn))
	1008	+ goto isolate_abort;
	1009	+ }
855	1010
856	1011	/* Recheck PageLRU and PageCompound under lock */
857	1012	if (!PageLRU(page))
..	..	@@ -862,41 +1017,41 @@
862	1017	* and it's on LRU. It can only be a THP so the order
863	1018	* is safe to read and it's 0 for tail pages.
864	1019	*/
865		- if (unlikely(PageCompound(page))) {
866		- low_pfn += (1UL << compound_order(page)) - 1;
	1020	+ if (unlikely(PageCompound(page) && !cc->alloc_contig)) {
	1021	+ low_pfn += compound_nr(page) - 1;
867	1022	goto isolate_fail;
868	1023	}
869	1024	}
870	1025
871		- lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
	1026	+ lruvec = mem_cgroup_page_lruvec(page, pgdat);
872	1027
873	1028	/* Try isolate the page */
874	1029	if (__isolate_lru_page(page, isolate_mode) != 0)
875	1030	goto isolate_fail;
876	1031
877		- VM_BUG_ON_PAGE(PageCompound(page), page);
	1032	+ /* The whole page is taken off the LRU; skip the tail pages. */
	1033	+ if (PageCompound(page))
	1034	+ low_pfn += compound_nr(page) - 1;
878	1035
879	1036	/* Successfully isolated */
880	1037	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));
	1038	+ mod_node_page_state(page_pgdat(page),
	1039	+ NR_ISOLATED_ANON + page_is_file_lru(page),
	1040	+ thp_nr_pages(page));
883	1041
884	1042	isolate_success:
885	1043	list_add(&page->lru, &cc->migratepages);
886		- cc->nr_migratepages++;
887		- nr_isolated++;
	1044	+ cc->nr_migratepages += compound_nr(page);
	1045	+ nr_isolated += compound_nr(page);
888	1046
889	1047	/*
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.
	1048	+ * Avoid isolating too much unless this block is being
	1049	+ * rescanned (e.g. dirty/writeback pages, parallel allocation)
	1050	+ * or a lock is contended. For contention, isolate quickly to
	1051	+ * potentially remove one source of contention.
894	1052	*/
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) {
	1053	+ if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX &&
	1054	+ !cc->rescan && !cc->contended) {
900	1055	++low_pfn;
901	1056	break;
902	1057	}
..	..	@@ -913,12 +1068,11 @@
913	1068	*/
914	1069	if (nr_isolated) {
915	1070	if (locked) {
916		- spin_unlock_irqrestore(zone_lru_lock(zone), flags);
	1071	+ spin_unlock_irqrestore(&pgdat->lru_lock, flags);
917	1072	locked = false;
918	1073	}
919	1074	putback_movable_pages(&cc->migratepages);
920	1075	cc->nr_migratepages = 0;
921		- cc->last_migrated_pfn = 0;
922	1076	nr_isolated = 0;
923	1077	}
924	1078
..	..	@@ -939,19 +1093,28 @@
939	1093	if (unlikely(low_pfn > end_pfn))
940	1094	low_pfn = end_pfn;
941	1095
	1096	+isolate_abort:
942	1097	if (locked)
943		- spin_unlock_irqrestore(zone_lru_lock(zone), flags);
	1098	+ spin_unlock_irqrestore(&pgdat->lru_lock, flags);
944	1099
945	1100	/*
946		- * Update the pageblock-skip information and cached scanner pfn,
947		- * if the whole pageblock was scanned without isolating any page.
	1101	+ * Updated the cached scanner pfn once the pageblock has been scanned
	1102	+ * Pages will either be migrated in which case there is no point
	1103	+ * scanning in the near future or migration failed in which case the
	1104	+ * failure reason may persist. The block is marked for skipping if
	1105	+ * there were no pages isolated in the block or if the block is
	1106	+ * rescanned twice in a row.
948	1107	*/
949		- if (low_pfn == end_pfn)
950		- update_pageblock_skip(cc, valid_page, nr_isolated, true);
	1108	+ if (low_pfn == end_pfn && (!nr_isolated \|\| cc->rescan)) {
	1109	+ if (valid_page && !skip_updated)
	1110	+ set_pageblock_skip(valid_page);
	1111	+ update_cached_migrate(cc, low_pfn);
	1112	+ }
951	1113
952	1114	trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
953	1115	nr_scanned, nr_isolated);
954	1116
	1117	+fatal_pending:
955	1118	cc->total_migrate_scanned += nr_scanned;
956	1119	if (nr_isolated)
957	1120	count_compact_events(COMPACTISOLATED, nr_isolated);
..	..	@@ -998,7 +1161,7 @@
998	1161	if (!pfn)
999	1162	break;
1000	1163
1001		- if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
	1164	+ if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX)
1002	1165	break;
1003	1166	}
1004	1167
..	..	@@ -1012,6 +1175,9 @@
1012	1175	struct page *page)
1013	1176	{
1014	1177	int block_mt;
	1178	+
	1179	+ if (pageblock_skip_persistent(page))
	1180	+ return false;
1015	1181
1016	1182	if ((cc->mode != MIGRATE_ASYNC) \|\| !cc->direct_compaction)
1017	1183	return true;
..	..	@@ -1035,7 +1201,7 @@
1035	1201	* the only small danger is that we skip a potentially suitable
1036	1202	* pageblock, so it's not worth to check order for valid range.
1037	1203	*/
1038		- if (page_order_unsafe(page) >= pageblock_order)
	1204	+ if (buddy_order_unsafe(page) >= pageblock_order)
1039	1205	return false;
1040	1206	}
1041	1207
..	..	@@ -1050,6 +1216,14 @@
1050	1216	return false;
1051	1217	}
1052	1218
	1219	+static inline unsigned int
	1220	+freelist_scan_limit(struct compact_control *cc)
	1221	+{
	1222	+ unsigned short shift = BITS_PER_LONG - 1;
	1223	+
	1224	+ return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1;
	1225	+}
	1226	+
1053	1227	/*
1054	1228	* Test whether the free scanner has reached the same or lower pageblock than
1055	1229	* the migration scanner, and compaction should thus terminate.
..	..	@@ -1058,6 +1232,257 @@
1058	1232	{
1059	1233	return (cc->free_pfn >> pageblock_order)
1060	1234	<= (cc->migrate_pfn >> pageblock_order);
	1235	+}
	1236	+
	1237	+/*
	1238	+ * Used when scanning for a suitable migration target which scans freelists
	1239	+ * in reverse. Reorders the list such as the unscanned pages are scanned
	1240	+ * first on the next iteration of the free scanner
	1241	+ */
	1242	+static void
	1243	+move_freelist_head(struct list_head freelist, struct page freepage)
	1244	+{
	1245	+ LIST_HEAD(sublist);
	1246	+
	1247	+ if (!list_is_last(freelist, &freepage->lru)) {
	1248	+ list_cut_before(&sublist, freelist, &freepage->lru);
	1249	+ if (!list_empty(&sublist))
	1250	+ list_splice_tail(&sublist, freelist);
	1251	+ }
	1252	+}
	1253	+
	1254	+/*
	1255	+ * Similar to move_freelist_head except used by the migration scanner
	1256	+ * when scanning forward. It's possible for these list operations to
	1257	+ * move against each other if they search the free list exactly in
	1258	+ * lockstep.
	1259	+ */
	1260	+static void
	1261	+move_freelist_tail(struct list_head freelist, struct page freepage)
	1262	+{
	1263	+ LIST_HEAD(sublist);
	1264	+
	1265	+ if (!list_is_first(freelist, &freepage->lru)) {
	1266	+ list_cut_position(&sublist, freelist, &freepage->lru);
	1267	+ if (!list_empty(&sublist))
	1268	+ list_splice_tail(&sublist, freelist);
	1269	+ }
	1270	+}
	1271	+
	1272	+static void
	1273	+fast_isolate_around(struct compact_control *cc, unsigned long pfn, unsigned long nr_isolated)
	1274	+{
	1275	+ unsigned long start_pfn, end_pfn;
	1276	+ struct page *page;
	1277	+
	1278	+ /* Do not search around if there are enough pages already */
	1279	+ if (cc->nr_freepages >= cc->nr_migratepages)
	1280	+ return;
	1281	+
	1282	+ /* Minimise scanning during async compaction */
	1283	+ if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC)
	1284	+ return;
	1285	+
	1286	+ /* Pageblock boundaries */
	1287	+ start_pfn = max(pageblock_start_pfn(pfn), cc->zone->zone_start_pfn);
	1288	+ end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone));
	1289	+
	1290	+ page = pageblock_pfn_to_page(start_pfn, end_pfn, cc->zone);
	1291	+ if (!page)
	1292	+ return;
	1293	+
	1294	+ /* Scan before */
	1295	+ if (start_pfn != pfn) {
	1296	+ isolate_freepages_block(cc, &start_pfn, pfn, &cc->freepages, 1, false);
	1297	+ if (cc->nr_freepages >= cc->nr_migratepages)
	1298	+ return;
	1299	+ }
	1300	+
	1301	+ /* Scan after */
	1302	+ start_pfn = pfn + nr_isolated;
	1303	+ if (start_pfn < end_pfn)
	1304	+ isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, 1, false);
	1305	+
	1306	+ /* Skip this pageblock in the future as it's full or nearly full */
	1307	+ if (cc->nr_freepages < cc->nr_migratepages)
	1308	+ set_pageblock_skip(page);
	1309	+}
	1310	+
	1311	+/* Search orders in round-robin fashion */
	1312	+static int next_search_order(struct compact_control *cc, int order)
	1313	+{
	1314	+ order--;
	1315	+ if (order < 0)
	1316	+ order = cc->order - 1;
	1317	+
	1318	+ /* Search wrapped around? */
	1319	+ if (order == cc->search_order) {
	1320	+ cc->search_order--;
	1321	+ if (cc->search_order < 0)
	1322	+ cc->search_order = cc->order - 1;
	1323	+ return -1;
	1324	+ }
	1325	+
	1326	+ return order;
	1327	+}
	1328	+
	1329	+static unsigned long
	1330	+fast_isolate_freepages(struct compact_control *cc)
	1331	+{
	1332	+ unsigned int limit = min(1U, freelist_scan_limit(cc) >> 1);
	1333	+ unsigned int nr_scanned = 0;
	1334	+ unsigned long low_pfn, min_pfn, highest = 0;
	1335	+ unsigned long nr_isolated = 0;
	1336	+ unsigned long distance;
	1337	+ struct page *page = NULL;
	1338	+ bool scan_start = false;
	1339	+ int order;
	1340	+
	1341	+ /* Full compaction passes in a negative order */
	1342	+ if (cc->order <= 0)
	1343	+ return cc->free_pfn;
	1344	+
	1345	+ /*
	1346	+ * If starting the scan, use a deeper search and use the highest
	1347	+ * PFN found if a suitable one is not found.
	1348	+ */
	1349	+ if (cc->free_pfn >= cc->zone->compact_init_free_pfn) {
	1350	+ limit = pageblock_nr_pages >> 1;
	1351	+ scan_start = true;
	1352	+ }
	1353	+
	1354	+ /*
	1355	+ * Preferred point is in the top quarter of the scan space but take
	1356	+ * a pfn from the top half if the search is problematic.
	1357	+ */
	1358	+ distance = (cc->free_pfn - cc->migrate_pfn);
	1359	+ low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2));
	1360	+ min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1));
	1361	+
	1362	+ if (WARN_ON_ONCE(min_pfn > low_pfn))
	1363	+ low_pfn = min_pfn;
	1364	+
	1365	+ /*
	1366	+ * Search starts from the last successful isolation order or the next
	1367	+ * order to search after a previous failure
	1368	+ */
	1369	+ cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order);
	1370	+
	1371	+ for (order = cc->search_order;
	1372	+ !page && order >= 0;
	1373	+ order = next_search_order(cc, order)) {
	1374	+ struct free_area *area = &cc->zone->free_area[order];
	1375	+ struct list_head *freelist;
	1376	+ struct page *freepage;
	1377	+ unsigned long flags;
	1378	+ unsigned int order_scanned = 0;
	1379	+ unsigned long high_pfn = 0;
	1380	+
	1381	+ if (!area->nr_free)
	1382	+ continue;
	1383	+
	1384	+ spin_lock_irqsave(&cc->zone->lock, flags);
	1385	+ freelist = &area->free_list[MIGRATE_MOVABLE];
	1386	+ list_for_each_entry_reverse(freepage, freelist, lru) {
	1387	+ unsigned long pfn;
	1388	+
	1389	+ order_scanned++;
	1390	+ nr_scanned++;
	1391	+ pfn = page_to_pfn(freepage);
	1392	+
	1393	+ if (pfn >= highest)
	1394	+ highest = max(pageblock_start_pfn(pfn),
	1395	+ cc->zone->zone_start_pfn);
	1396	+
	1397	+ if (pfn >= low_pfn) {
	1398	+ cc->fast_search_fail = 0;
	1399	+ cc->search_order = order;
	1400	+ page = freepage;
	1401	+ break;
	1402	+ }
	1403	+
	1404	+ if (pfn >= min_pfn && pfn > high_pfn) {
	1405	+ high_pfn = pfn;
	1406	+
	1407	+ /* Shorten the scan if a candidate is found */
	1408	+ limit >>= 1;
	1409	+ }
	1410	+
	1411	+ if (order_scanned >= limit)
	1412	+ break;
	1413	+ }
	1414	+
	1415	+ /* Use a minimum pfn if a preferred one was not found */
	1416	+ if (!page && high_pfn) {
	1417	+ page = pfn_to_page(high_pfn);
	1418	+
	1419	+ /* Update freepage for the list reorder below */
	1420	+ freepage = page;
	1421	+ }
	1422	+
	1423	+ /* Reorder to so a future search skips recent pages */
	1424	+ move_freelist_head(freelist, freepage);
	1425	+
	1426	+ /* Isolate the page if available */
	1427	+ if (page) {
	1428	+ if (__isolate_free_page(page, order)) {
	1429	+ set_page_private(page, order);
	1430	+ nr_isolated = 1 << order;
	1431	+ cc->nr_freepages += nr_isolated;
	1432	+ list_add_tail(&page->lru, &cc->freepages);
	1433	+ count_compact_events(COMPACTISOLATED, nr_isolated);
	1434	+ } else {
	1435	+ /* If isolation fails, abort the search */
	1436	+ order = cc->search_order + 1;
	1437	+ page = NULL;
	1438	+ }
	1439	+ }
	1440	+
	1441	+ spin_unlock_irqrestore(&cc->zone->lock, flags);
	1442	+
	1443	+ /*
	1444	+ * Smaller scan on next order so the total scan ig related
	1445	+ * to freelist_scan_limit.
	1446	+ */
	1447	+ if (order_scanned >= limit)
	1448	+ limit = min(1U, limit >> 1);
	1449	+ }
	1450	+
	1451	+ if (!page) {
	1452	+ cc->fast_search_fail++;
	1453	+ if (scan_start) {
	1454	+ /*
	1455	+ * Use the highest PFN found above min. If one was
	1456	+ * not found, be pessimistic for direct compaction
	1457	+ * and use the min mark.
	1458	+ */
	1459	+ if (highest) {
	1460	+ page = pfn_to_page(highest);
	1461	+ cc->free_pfn = highest;
	1462	+ } else {
	1463	+ if (cc->direct_compaction && pfn_valid(min_pfn)) {
	1464	+ page = pageblock_pfn_to_page(min_pfn,
	1465	+ min(pageblock_end_pfn(min_pfn),
	1466	+ zone_end_pfn(cc->zone)),
	1467	+ cc->zone);
	1468	+ cc->free_pfn = min_pfn;
	1469	+ }
	1470	+ }
	1471	+ }
	1472	+ }
	1473	+
	1474	+ if (highest && highest >= cc->zone->compact_cached_free_pfn) {
	1475	+ highest -= pageblock_nr_pages;
	1476	+ cc->zone->compact_cached_free_pfn = highest;
	1477	+ }
	1478	+
	1479	+ cc->total_free_scanned += nr_scanned;
	1480	+ if (!page)
	1481	+ return cc->free_pfn;
	1482	+
	1483	+ low_pfn = page_to_pfn(page);
	1484	+ fast_isolate_around(cc, low_pfn, nr_isolated);
	1485	+ return low_pfn;
1061	1486	}
1062	1487
1063	1488	/*
..	..	@@ -1073,6 +1498,12 @@
1073	1498	unsigned long block_end_pfn; /* end of current pageblock */
1074	1499	unsigned long low_pfn; /* lowest pfn scanner is able to scan */
1075	1500	struct list_head *freelist = &cc->freepages;
	1501	+ unsigned int stride;
	1502	+
	1503	+ /* Try a small search of the free lists for a candidate */
	1504	+ isolate_start_pfn = fast_isolate_freepages(cc);
	1505	+ if (cc->nr_freepages)
	1506	+ goto splitmap;
1076	1507
1077	1508	/*
1078	1509	* Initialise the free scanner. The starting point is where we last
..	..	@@ -1081,15 +1512,16 @@
1081	1512	* this pfn aligned down to the pageblock boundary, because we do
1082	1513	* block_start_pfn -= pageblock_nr_pages in the for loop.
1083	1514	* For ending point, take care when isolating in last pageblock of a
1084		- * a zone which ends in the middle of a pageblock.
	1515	+ * zone which ends in the middle of a pageblock.
1085	1516	* The low boundary is the end of the pageblock the migration scanner
1086	1517	* is using.
1087	1518	*/
1088	1519	isolate_start_pfn = cc->free_pfn;
1089		- block_start_pfn = pageblock_start_pfn(cc->free_pfn);
	1520	+ block_start_pfn = pageblock_start_pfn(isolate_start_pfn);
1090	1521	block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1091	1522	zone_end_pfn(zone));
1092	1523	low_pfn = pageblock_end_pfn(cc->migrate_pfn);
	1524	+ stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1;
1093	1525
1094	1526	/*
1095	1527	* Isolate free pages until enough are available to migrate the
..	..	@@ -1100,14 +1532,14 @@
1100	1532	block_end_pfn = block_start_pfn,
1101	1533	block_start_pfn -= pageblock_nr_pages,
1102	1534	isolate_start_pfn = block_start_pfn) {
	1535	+ unsigned long nr_isolated;
	1536	+
1103	1537	/*
1104	1538	* 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.
	1539	+ * suitable migration targets, so periodically check resched.
1107	1540	*/
1108		- if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1109		- && compact_should_abort(cc))
1110		- break;
	1541	+ if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)))
	1542	+ cond_resched();
1111	1543
1112	1544	page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1113	1545	zone);
..	..	@@ -1123,15 +1555,15 @@
1123	1555	continue;
1124	1556
1125	1557	/* Found a block suitable for isolating free pages from. */
1126		- isolate_freepages_block(cc, &isolate_start_pfn, block_end_pfn,
1127		- freelist, false);
	1558	+ nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn,
	1559	+ block_end_pfn, freelist, stride, false);
1128	1560
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) {
	1561	+ /* Update the skip hint if the full pageblock was scanned */
	1562	+ if (isolate_start_pfn == block_end_pfn)
	1563	+ update_pageblock_skip(cc, page, block_start_pfn);
	1564	+
	1565	+ /* Are enough freepages isolated? */
	1566	+ if (cc->nr_freepages >= cc->nr_migratepages) {
1135	1567	if (isolate_start_pfn >= block_end_pfn) {
1136	1568	/*
1137	1569	* Restart at previous pageblock if more
..	..	@@ -1148,10 +1580,14 @@
1148	1580	*/
1149	1581	break;
1150	1582	}
1151		- }
1152	1583
1153		- /* __isolate_free_page() does not map the pages */
1154		- map_pages(freelist);
	1584	+ /* Adjust stride depending on isolation */
	1585	+ if (nr_isolated) {
	1586	+ stride = 1;
	1587	+ continue;
	1588	+ }
	1589	+ stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1);
	1590	+ }
1155	1591
1156	1592	/*
1157	1593	* Record where the free scanner will restart next time. Either we
..	..	@@ -1160,6 +1596,10 @@
1160	1596	* and the loop terminated due to isolate_start_pfn < low_pfn
1161	1597	*/
1162	1598	cc->free_pfn = isolate_start_pfn;
	1599	+
	1600	+splitmap:
	1601	+ /* __isolate_free_page() does not map the pages */
	1602	+ split_map_pages(freelist);
1163	1603	}
1164	1604
1165	1605	/*
..	..	@@ -1172,13 +1612,8 @@
1172	1612	struct compact_control cc = (struct compact_control )data;
1173	1613	struct page *freepage;
1174	1614
1175		- /*
1176		- * Isolate free pages if necessary, and if we are not aborting due to
1177		- * contention.
1178		- */
1179	1615	if (list_empty(&cc->freepages)) {
1180		- if (!cc->contended)
1181		- isolate_freepages(cc);
	1616	+ isolate_freepages(cc);
1182	1617
1183	1618	if (list_empty(&cc->freepages))
1184	1619	return NULL;
..	..	@@ -1215,15 +1650,158 @@
1215	1650	* Allow userspace to control policy on scanning the unevictable LRU for
1216	1651	* compactable pages.
1217	1652	*/
	1653	+#ifdef CONFIG_PREEMPT_RT
	1654	+int sysctl_compact_unevictable_allowed __read_mostly = 0;
	1655	+#else
1218	1656	int sysctl_compact_unevictable_allowed __read_mostly = 1;
	1657	+#endif
	1658	+
	1659	+static inline void
	1660	+update_fast_start_pfn(struct compact_control *cc, unsigned long pfn)
	1661	+{
	1662	+ if (cc->fast_start_pfn == ULONG_MAX)
	1663	+ return;
	1664	+
	1665	+ if (!cc->fast_start_pfn)
	1666	+ cc->fast_start_pfn = pfn;
	1667	+
	1668	+ cc->fast_start_pfn = min(cc->fast_start_pfn, pfn);
	1669	+}
	1670	+
	1671	+static inline unsigned long
	1672	+reinit_migrate_pfn(struct compact_control *cc)
	1673	+{
	1674	+ if (!cc->fast_start_pfn \|\| cc->fast_start_pfn == ULONG_MAX)
	1675	+ return cc->migrate_pfn;
	1676	+
	1677	+ cc->migrate_pfn = cc->fast_start_pfn;
	1678	+ cc->fast_start_pfn = ULONG_MAX;
	1679	+
	1680	+ return cc->migrate_pfn;
	1681	+}
	1682	+
	1683	+/*
	1684	+ * Briefly search the free lists for a migration source that already has
	1685	+ * some free pages to reduce the number of pages that need migration
	1686	+ * before a pageblock is free.
	1687	+ */
	1688	+static unsigned long fast_find_migrateblock(struct compact_control *cc)
	1689	+{
	1690	+ unsigned int limit = freelist_scan_limit(cc);
	1691	+ unsigned int nr_scanned = 0;
	1692	+ unsigned long distance;
	1693	+ unsigned long pfn = cc->migrate_pfn;
	1694	+ unsigned long high_pfn;
	1695	+ int order;
	1696	+ bool found_block = false;
	1697	+
	1698	+ /* Skip hints are relied on to avoid repeats on the fast search */
	1699	+ if (cc->ignore_skip_hint)
	1700	+ return pfn;
	1701	+
	1702	+ /*
	1703	+ * If the migrate_pfn is not at the start of a zone or the start
	1704	+ * of a pageblock then assume this is a continuation of a previous
	1705	+ * scan restarted due to COMPACT_CLUSTER_MAX.
	1706	+ */
	1707	+ if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn))
	1708	+ return pfn;
	1709	+
	1710	+ /*
	1711	+ * For smaller orders, just linearly scan as the number of pages
	1712	+ * to migrate should be relatively small and does not necessarily
	1713	+ * justify freeing up a large block for a small allocation.
	1714	+ */
	1715	+ if (cc->order <= PAGE_ALLOC_COSTLY_ORDER)
	1716	+ return pfn;
	1717	+
	1718	+ /*
	1719	+ * Only allow kcompactd and direct requests for movable pages to
	1720	+ * quickly clear out a MOVABLE pageblock for allocation. This
	1721	+ * reduces the risk that a large movable pageblock is freed for
	1722	+ * an unmovable/reclaimable small allocation.
	1723	+ */
	1724	+ if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE)
	1725	+ return pfn;
	1726	+
	1727	+ /*
	1728	+ * When starting the migration scanner, pick any pageblock within the
	1729	+ * first half of the search space. Otherwise try and pick a pageblock
	1730	+ * within the first eighth to reduce the chances that a migration
	1731	+ * target later becomes a source.
	1732	+ */
	1733	+ distance = (cc->free_pfn - cc->migrate_pfn) >> 1;
	1734	+ if (cc->migrate_pfn != cc->zone->zone_start_pfn)
	1735	+ distance >>= 2;
	1736	+ high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance);
	1737	+
	1738	+ for (order = cc->order - 1;
	1739	+ order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit;
	1740	+ order--) {
	1741	+ struct free_area *area = &cc->zone->free_area[order];
	1742	+ struct list_head *freelist;
	1743	+ unsigned long flags;
	1744	+ struct page *freepage;
	1745	+
	1746	+ if (!area->nr_free)
	1747	+ continue;
	1748	+
	1749	+ spin_lock_irqsave(&cc->zone->lock, flags);
	1750	+ freelist = &area->free_list[MIGRATE_MOVABLE];
	1751	+ list_for_each_entry(freepage, freelist, lru) {
	1752	+ unsigned long free_pfn;
	1753	+
	1754	+ if (nr_scanned++ >= limit) {
	1755	+ move_freelist_tail(freelist, freepage);
	1756	+ break;
	1757	+ }
	1758	+
	1759	+ free_pfn = page_to_pfn(freepage);
	1760	+ if (free_pfn < high_pfn) {
	1761	+ /*
	1762	+ * Avoid if skipped recently. Ideally it would
	1763	+ * move to the tail but even safe iteration of
	1764	+ * the list assumes an entry is deleted, not
	1765	+ * reordered.
	1766	+ */
	1767	+ if (get_pageblock_skip(freepage))
	1768	+ continue;
	1769	+
	1770	+ /* Reorder to so a future search skips recent pages */
	1771	+ move_freelist_tail(freelist, freepage);
	1772	+
	1773	+ update_fast_start_pfn(cc, free_pfn);
	1774	+ pfn = pageblock_start_pfn(free_pfn);
	1775	+ if (pfn < cc->zone->zone_start_pfn)
	1776	+ pfn = cc->zone->zone_start_pfn;
	1777	+ cc->fast_search_fail = 0;
	1778	+ found_block = true;
	1779	+ set_pageblock_skip(freepage);
	1780	+ break;
	1781	+ }
	1782	+ }
	1783	+ spin_unlock_irqrestore(&cc->zone->lock, flags);
	1784	+ }
	1785	+
	1786	+ cc->total_migrate_scanned += nr_scanned;
	1787	+
	1788	+ /*
	1789	+ * If fast scanning failed then use a cached entry for a page block
	1790	+ * that had free pages as the basis for starting a linear scan.
	1791	+ */
	1792	+ if (!found_block) {
	1793	+ cc->fast_search_fail++;
	1794	+ pfn = reinit_migrate_pfn(cc);
	1795	+ }
	1796	+ return pfn;
	1797	+}
1219	1798
1220	1799	/*
1221	1800	* Isolate all pages that can be migrated from the first suitable block,
1222	1801	* starting at the block pointed to by the migrate scanner pfn within
1223	1802	* compact_control.
1224	1803	*/
1225		-static isolate_migrate_t isolate_migratepages(struct zone *zone,
1226		- struct compact_control *cc)
	1804	+static isolate_migrate_t isolate_migratepages(struct compact_control *cc)
1227	1805	{
1228	1806	unsigned long block_start_pfn;
1229	1807	unsigned long block_end_pfn;
..	..	@@ -1232,15 +1810,24 @@
1232	1810	const isolate_mode_t isolate_mode =
1233	1811	(sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) \|
1234	1812	(cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
	1813	+ bool fast_find_block;
1235	1814
1236	1815	/*
1237	1816	* Start at where we last stopped, or beginning of the zone as
1238		- * initialized by compact_zone()
	1817	+ * initialized by compact_zone(). The first failure will use
	1818	+ * the lowest PFN as the starting point for linear scanning.
1239	1819	*/
1240		- low_pfn = cc->migrate_pfn;
	1820	+ low_pfn = fast_find_migrateblock(cc);
1241	1821	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;
	1822	+ if (block_start_pfn < cc->zone->zone_start_pfn)
	1823	+ block_start_pfn = cc->zone->zone_start_pfn;
	1824	+
	1825	+ /*
	1826	+ * fast_find_migrateblock marks a pageblock skipped so to avoid
	1827	+ * the isolation_suitable check below, check whether the fast
	1828	+ * search was successful.
	1829	+ */
	1830	+ fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail;
1244	1831
1245	1832	/* Only scan within a pageblock boundary */
1246	1833	block_end_pfn = pageblock_end_pfn(low_pfn);
..	..	@@ -1250,6 +1837,7 @@
1250	1837	* Do not cross the free scanner.
1251	1838	*/
1252	1839	for (; block_end_pfn <= cc->free_pfn;
	1840	+ fast_find_block = false,
1253	1841	low_pfn = block_end_pfn,
1254	1842	block_start_pfn = block_end_pfn,
1255	1843	block_end_pfn += pageblock_nr_pages) {
..	..	@@ -1257,34 +1845,45 @@
1257	1845	/*
1258	1846	* This can potentially iterate a massively long zone with
1259	1847	* many pageblocks unsuitable, so periodically check if we
1260		- * need to schedule, or even abort async compaction.
	1848	+ * need to schedule.
1261	1849	*/
1262		- if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1263		- && compact_should_abort(cc))
1264		- break;
	1850	+ if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)))
	1851	+ cond_resched();
1265	1852
1266		- page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1267		- zone);
	1853	+ page = pageblock_pfn_to_page(block_start_pfn,
	1854	+ block_end_pfn, cc->zone);
1268	1855	if (!page)
1269	1856	continue;
1270	1857
1271		- /* If isolation recently failed, do not retry */
1272		- if (!isolation_suitable(cc, page))
	1858	+ /*
	1859	+ * If isolation recently failed, do not retry. Only check the
	1860	+ * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
	1861	+ * to be visited multiple times. Assume skip was checked
	1862	+ * before making it "skip" so other compaction instances do
	1863	+ * not scan the same block.
	1864	+ */
	1865	+ if (IS_ALIGNED(low_pfn, pageblock_nr_pages) &&
	1866	+ !fast_find_block && !isolation_suitable(cc, page))
1273	1867	continue;
1274	1868
1275	1869	/*
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.
	1870	+ * For async compaction, also only scan in MOVABLE blocks
	1871	+ * without huge pages. Async compaction is optimistic to see
	1872	+ * if the minimum amount of work satisfies the allocation.
	1873	+ * The cached PFN is updated as it's possible that all
	1874	+ * remaining blocks between source and target are unsuitable
	1875	+ * and the compaction scanners fail to meet.
1279	1876	*/
1280		- if (!suitable_migration_source(cc, page))
	1877	+ if (!suitable_migration_source(cc, page)) {
	1878	+ update_cached_migrate(cc, block_end_pfn);
1281	1879	continue;
	1880	+ }
1282	1881
1283	1882	/* Perform the isolation */
1284	1883	low_pfn = isolate_migratepages_block(cc, low_pfn,
1285	1884	block_end_pfn, isolate_mode);
1286	1885
1287		- if (!low_pfn \|\| cc->contended)
	1886	+ if (!low_pfn)
1288	1887	return ISOLATE_ABORT;
1289	1888
1290	1889	/*
..	..	@@ -1310,19 +1909,94 @@
1310	1909	return order == -1;
1311	1910	}
1312	1911
1313		-static enum compact_result __compact_finished(struct zone *zone,
1314		- struct compact_control *cc)
	1912	+static bool kswapd_is_running(pg_data_t *pgdat)
	1913	+{
	1914	+ return pgdat->kswapd && (pgdat->kswapd->state == TASK_RUNNING);
	1915	+}
	1916	+
	1917	+/*
	1918	+ * A zone's fragmentation score is the external fragmentation wrt to the
	1919	+ * COMPACTION_HPAGE_ORDER. It returns a value in the range [0, 100].
	1920	+ */
	1921	+static unsigned int fragmentation_score_zone(struct zone *zone)
	1922	+{
	1923	+ return extfrag_for_order(zone, COMPACTION_HPAGE_ORDER);
	1924	+}
	1925	+
	1926	+/*
	1927	+ * A weighted zone's fragmentation score is the external fragmentation
	1928	+ * wrt to the COMPACTION_HPAGE_ORDER scaled by the zone's size. It
	1929	+ * returns a value in the range [0, 100].
	1930	+ *
	1931	+ * The scaling factor ensures that proactive compaction focuses on larger
	1932	+ * zones like ZONE_NORMAL, rather than smaller, specialized zones like
	1933	+ * ZONE_DMA32. For smaller zones, the score value remains close to zero,
	1934	+ * and thus never exceeds the high threshold for proactive compaction.
	1935	+ */
	1936	+static unsigned int fragmentation_score_zone_weighted(struct zone *zone)
	1937	+{
	1938	+ unsigned long score;
	1939	+
	1940	+ score = zone->present_pages * fragmentation_score_zone(zone);
	1941	+ return div64_ul(score, zone->zone_pgdat->node_present_pages + 1);
	1942	+}
	1943	+
	1944	+/*
	1945	+ * The per-node proactive (background) compaction process is started by its
	1946	+ * corresponding kcompactd thread when the node's fragmentation score
	1947	+ * exceeds the high threshold. The compaction process remains active till
	1948	+ * the node's score falls below the low threshold, or one of the back-off
	1949	+ * conditions is met.
	1950	+ */
	1951	+static unsigned int fragmentation_score_node(pg_data_t *pgdat)
	1952	+{
	1953	+ unsigned int score = 0;
	1954	+ int zoneid;
	1955	+
	1956	+ for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
	1957	+ struct zone *zone;
	1958	+
	1959	+ zone = &pgdat->node_zones[zoneid];
	1960	+ score += fragmentation_score_zone_weighted(zone);
	1961	+ }
	1962	+
	1963	+ return score;
	1964	+}
	1965	+
	1966	+static unsigned int fragmentation_score_wmark(pg_data_t *pgdat, bool low)
	1967	+{
	1968	+ unsigned int wmark_low;
	1969	+
	1970	+ /*
	1971	+ * Cap the low watermak to avoid excessive compaction
	1972	+ * activity in case a user sets the proactivess tunable
	1973	+ * close to 100 (maximum).
	1974	+ */
	1975	+ wmark_low = max(100U - sysctl_compaction_proactiveness, 5U);
	1976	+ return low ? wmark_low : min(wmark_low + 10, 100U);
	1977	+}
	1978	+
	1979	+static bool should_proactive_compact_node(pg_data_t *pgdat)
	1980	+{
	1981	+ int wmark_high;
	1982	+
	1983	+ if (!sysctl_compaction_proactiveness \|\| kswapd_is_running(pgdat))
	1984	+ return false;
	1985	+
	1986	+ wmark_high = fragmentation_score_wmark(pgdat, false);
	1987	+ return fragmentation_score_node(pgdat) > wmark_high;
	1988	+}
	1989	+
	1990	+static enum compact_result __compact_finished(struct compact_control *cc)
1315	1991	{
1316	1992	unsigned int order;
1317	1993	const int migratetype = cc->migratetype;
1318		-
1319		- if (cc->contended \|\| fatal_signal_pending(current))
1320		- return COMPACT_CONTENDED;
	1994	+ int ret;
1321	1995
1322	1996	/* Compaction run completes if the migrate and free scanner meet */
1323	1997	if (compact_scanners_met(cc)) {
1324	1998	/* Let the next compaction start anew. */
1325		- reset_cached_positions(zone);
	1999	+ reset_cached_positions(cc->zone);
1326	2000
1327	2001	/*
1328	2002	* Mark that the PG_migrate_skip information should be cleared
..	..	@@ -1331,7 +2005,7 @@
1331	2005	* based on an allocation request.
1332	2006	*/
1333	2007	if (cc->direct_compaction)
1334		- zone->compact_blockskip_flush = true;
	2008	+ cc->zone->compact_blockskip_flush = true;
1335	2009
1336	2010	if (cc->whole_zone)
1337	2011	return COMPACT_COMPLETE;
..	..	@@ -1339,33 +2013,51 @@
1339	2013	return COMPACT_PARTIAL_SKIPPED;
1340	2014	}
1341	2015
	2016	+ if (cc->proactive_compaction) {
	2017	+ int score, wmark_low;
	2018	+ pg_data_t *pgdat;
	2019	+
	2020	+ pgdat = cc->zone->zone_pgdat;
	2021	+ if (kswapd_is_running(pgdat))
	2022	+ return COMPACT_PARTIAL_SKIPPED;
	2023	+
	2024	+ score = fragmentation_score_zone(cc->zone);
	2025	+ wmark_low = fragmentation_score_wmark(pgdat, true);
	2026	+
	2027	+ if (score > wmark_low)
	2028	+ ret = COMPACT_CONTINUE;
	2029	+ else
	2030	+ ret = COMPACT_SUCCESS;
	2031	+
	2032	+ goto out;
	2033	+ }
	2034	+
1342	2035	if (is_via_compact_memory(cc->order))
1343	2036	return COMPACT_CONTINUE;
1344	2037
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		- }
	2038	+ /*
	2039	+ * Always finish scanning a pageblock to reduce the possibility of
	2040	+ * fallbacks in the future. This is particularly important when
	2041	+ * migration source is unmovable/reclaimable but it's not worth
	2042	+ * special casing.
	2043	+ */
	2044	+ if (!IS_ALIGNED(cc->migrate_pfn, pageblock_nr_pages))
	2045	+ return COMPACT_CONTINUE;
1355	2046
1356	2047	/* Direct compactor: Is a suitable page free? */
	2048	+ ret = COMPACT_NO_SUITABLE_PAGE;
1357	2049	for (order = cc->order; order < MAX_ORDER; order++) {
1358		- struct free_area *area = &zone->free_area[order];
	2050	+ struct free_area *area = &cc->zone->free_area[order];
1359	2051	bool can_steal;
1360	2052
1361	2053	/* Job done if page is free of the right migratetype */
1362		- if (!list_empty(&area->free_list[migratetype]))
	2054	+ if (!free_area_empty(area, migratetype))
1363	2055	return COMPACT_SUCCESS;
1364	2056
1365	2057	#ifdef CONFIG_CMA
1366	2058	/* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1367	2059	if (migratetype == MIGRATE_MOVABLE &&
1368		- !list_empty(&area->free_list[MIGRATE_CMA]))
	2060	+ !free_area_empty(area, MIGRATE_CMA))
1369	2061	return COMPACT_SUCCESS;
1370	2062	#endif
1371	2063	/*
..	..	@@ -1393,21 +2085,24 @@
1393	2085	return COMPACT_SUCCESS;
1394	2086	}
1395	2087
1396		- cc->finishing_block = true;
1397		- return COMPACT_CONTINUE;
	2088	+ ret = COMPACT_CONTINUE;
	2089	+ break;
1398	2090	}
1399	2091	}
1400	2092
1401		- return COMPACT_NO_SUITABLE_PAGE;
	2093	+out:
	2094	+ if (cc->contended \|\| fatal_signal_pending(current))
	2095	+ ret = COMPACT_CONTENDED;
	2096	+
	2097	+ return ret;
1402	2098	}
1403	2099
1404		-static enum compact_result compact_finished(struct zone *zone,
1405		- struct compact_control *cc)
	2100	+static enum compact_result compact_finished(struct compact_control *cc)
1406	2101	{
1407	2102	int ret;
1408	2103
1409		- ret = __compact_finished(zone, cc);
1410		- trace_mm_compaction_finished(zone, cc->order, ret);
	2104	+ ret = __compact_finished(cc);
	2105	+ trace_mm_compaction_finished(cc->zone, cc->order, ret);
1411	2106	if (ret == COMPACT_NO_SUITABLE_PAGE)
1412	2107	ret = COMPACT_CONTINUE;
1413	2108
..	..	@@ -1423,7 +2118,7 @@
1423	2118	*/
1424	2119	static enum compact_result __compaction_suitable(struct zone *zone, int order,
1425	2120	unsigned int alloc_flags,
1426		- int classzone_idx,
	2121	+ int highest_zoneidx,
1427	2122	unsigned long wmark_target)
1428	2123	{
1429	2124	unsigned long watermark;
..	..	@@ -1431,12 +2126,12 @@
1431	2126	if (is_via_compact_memory(order))
1432	2127	return COMPACT_CONTINUE;
1433	2128
1434		- watermark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
	2129	+ watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
1435	2130	/*
1436	2131	* If watermarks for high-order allocation are already met, there
1437	2132	* should be no need for compaction at all.
1438	2133	*/
1439		- if (zone_watermark_ok(zone, order, watermark, classzone_idx,
	2134	+ if (zone_watermark_ok(zone, order, watermark, highest_zoneidx,
1440	2135	alloc_flags))
1441	2136	return COMPACT_SUCCESS;
1442	2137
..	..	@@ -1446,9 +2141,9 @@
1446	2141	* watermark and alloc_flags have to match, or be more pessimistic than
1447	2142	* the check in __isolate_free_page(). We don't use the direct
1448	2143	* 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.
	2144	+ * isolation. We however do use the direct compactor's highest_zoneidx
	2145	+ * to skip over zones where lowmem reserves would prevent allocation
	2146	+ * even if compaction succeeds.
1452	2147	* For costly orders, we require low watermark instead of min for
1453	2148	* compaction to proceed to increase its chances.
1454	2149	* ALLOC_CMA is used, as pages in CMA pageblocks are considered
..	..	@@ -1457,7 +2152,7 @@
1457	2152	watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
1458	2153	low_wmark_pages(zone) : min_wmark_pages(zone);
1459	2154	watermark += compact_gap(order);
1460		- if (!__zone_watermark_ok(zone, 0, watermark, classzone_idx,
	2155	+ if (!__zone_watermark_ok(zone, 0, watermark, highest_zoneidx,
1461	2156	ALLOC_CMA, wmark_target))
1462	2157	return COMPACT_SKIPPED;
1463	2158
..	..	@@ -1466,12 +2161,12 @@
1466	2161
1467	2162	enum compact_result compaction_suitable(struct zone *zone, int order,
1468	2163	unsigned int alloc_flags,
1469		- int classzone_idx)
	2164	+ int highest_zoneidx)
1470	2165	{
1471	2166	enum compact_result ret;
1472	2167	int fragindex;
1473	2168
1474		- ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx,
	2169	+ ret = __compaction_suitable(zone, order, alloc_flags, highest_zoneidx,
1475	2170	zone_page_state(zone, NR_FREE_PAGES));
1476	2171	/*
1477	2172	* fragmentation index determines if allocation failures are due to
..	..	@@ -1512,8 +2207,8 @@
1512	2207	* Make sure at least one zone would pass __compaction_suitable if we continue
1513	2208	* retrying the reclaim.
1514	2209	*/
1515		- for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1516		- ac->nodemask) {
	2210	+ for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
	2211	+ ac->highest_zoneidx, ac->nodemask) {
1517	2212	unsigned long available;
1518	2213	enum compact_result compact_result;
1519	2214
..	..	@@ -1526,7 +2221,7 @@
1526	2221	available = zone_reclaimable_pages(zone) / order;
1527	2222	available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
1528	2223	compact_result = __compaction_suitable(zone, order, alloc_flags,
1529		- ac_classzone_idx(ac), available);
	2224	+ ac->highest_zoneidx, available);
1530	2225	if (compact_result != COMPACT_SKIPPED)
1531	2226	return true;
1532	2227	}
..	..	@@ -1534,12 +2229,15 @@
1534	2229	return false;
1535	2230	}
1536	2231
1537		-static enum compact_result compact_zone(struct zone zone, struct compact_control cc)
	2232	+static enum compact_result
	2233	+compact_zone(struct compact_control cc, struct capture_control capc)
1538	2234	{
1539	2235	enum compact_result ret;
1540		- unsigned long start_pfn = zone->zone_start_pfn;
1541		- unsigned long end_pfn = zone_end_pfn(zone);
	2236	+ unsigned long start_pfn = cc->zone->zone_start_pfn;
	2237	+ unsigned long end_pfn = zone_end_pfn(cc->zone);
	2238	+ unsigned long last_migrated_pfn;
1542	2239	const bool sync = cc->mode != MIGRATE_ASYNC;
	2240	+ bool update_cached;
1543	2241
1544	2242	/*
1545	2243	* These counters track activities during zone compaction. Initialize
..	..	@@ -1552,9 +2250,9 @@
1552	2250	INIT_LIST_HEAD(&cc->freepages);
1553	2251	INIT_LIST_HEAD(&cc->migratepages);
1554	2252
1555		- cc->migratetype = gfpflags_to_migratetype(cc->gfp_mask);
1556		- ret = compaction_suitable(zone, cc->order, cc->alloc_flags,
1557		- cc->classzone_idx);
	2253	+ cc->migratetype = gfp_migratetype(cc->gfp_mask);
	2254	+ ret = compaction_suitable(cc->zone, cc->order, cc->alloc_flags,
	2255	+ cc->highest_zoneidx);
1558	2256	/* Compaction is likely to fail */
1559	2257	if (ret == COMPACT_SUCCESS \|\| ret == COMPACT_SKIPPED)
1560	2258	return ret;
..	..	@@ -1566,8 +2264,8 @@
1566	2264	* Clear pageblock skip if there were failures recently and compaction
1567	2265	* is about to be retried after being deferred.
1568	2266	*/
1569		- if (compaction_restarting(zone, cc->order))
1570		- __reset_isolation_suitable(zone);
	2267	+ if (compaction_restarting(cc->zone, cc->order))
	2268	+ __reset_isolation_suitable(cc->zone);
1571	2269
1572	2270	/*
1573	2271	* Setup to move all movable pages to the end of the zone. Used cached
..	..	@@ -1575,43 +2273,76 @@
1575	2273	* want to compact the whole zone), but check that it is initialised
1576	2274	* by ensuring the values are within zone boundaries.
1577	2275	*/
	2276	+ cc->fast_start_pfn = 0;
1578	2277	if (cc->whole_zone) {
1579	2278	cc->migrate_pfn = start_pfn;
1580	2279	cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
1581	2280	} else {
1582		- cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
1583		- cc->free_pfn = zone->compact_cached_free_pfn;
	2281	+ cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync];
	2282	+ cc->free_pfn = cc->zone->compact_cached_free_pfn;
1584	2283	if (cc->free_pfn < start_pfn \|\| cc->free_pfn >= end_pfn) {
1585	2284	cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
1586		- zone->compact_cached_free_pfn = cc->free_pfn;
	2285	+ cc->zone->compact_cached_free_pfn = cc->free_pfn;
1587	2286	}
1588	2287	if (cc->migrate_pfn < start_pfn \|\| cc->migrate_pfn >= end_pfn) {
1589	2288	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;
	2289	+ cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
	2290	+ cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1592	2291	}
1593	2292
1594		- if (cc->migrate_pfn == start_pfn)
	2293	+ if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn)
1595	2294	cc->whole_zone = true;
1596	2295	}
1597	2296
1598		- cc->last_migrated_pfn = 0;
	2297	+ last_migrated_pfn = 0;
	2298	+
	2299	+ /*
	2300	+ * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
	2301	+ * the basis that some migrations will fail in ASYNC mode. However,
	2302	+ * if the cached PFNs match and pageblocks are skipped due to having
	2303	+ * no isolation candidates, then the sync state does not matter.
	2304	+ * Until a pageblock with isolation candidates is found, keep the
	2305	+ * cached PFNs in sync to avoid revisiting the same blocks.
	2306	+ */
	2307	+ update_cached = !sync &&
	2308	+ cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1];
1599	2309
1600	2310	trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
1601	2311	cc->free_pfn, end_pfn, sync);
1602	2312
1603		- migrate_prep_local();
	2313	+ /* lru_add_drain_all could be expensive with involving other CPUs */
	2314	+ lru_add_drain();
1604	2315
1605		- while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
	2316	+ while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) {
1606	2317	int err;
	2318	+ unsigned long start_pfn = cc->migrate_pfn;
1607	2319
1608		- switch (isolate_migratepages(zone, cc)) {
	2320	+ /*
	2321	+ * Avoid multiple rescans which can happen if a page cannot be
	2322	+ * isolated (dirty/writeback in async mode) or if the migrated
	2323	+ * pages are being allocated before the pageblock is cleared.
	2324	+ * The first rescan will capture the entire pageblock for
	2325	+ * migration. If it fails, it'll be marked skip and scanning
	2326	+ * will proceed as normal.
	2327	+ */
	2328	+ cc->rescan = false;
	2329	+ if (pageblock_start_pfn(last_migrated_pfn) ==
	2330	+ pageblock_start_pfn(start_pfn)) {
	2331	+ cc->rescan = true;
	2332	+ }
	2333	+
	2334	+ switch (isolate_migratepages(cc)) {
1609	2335	case ISOLATE_ABORT:
1610	2336	ret = COMPACT_CONTENDED;
1611	2337	putback_movable_pages(&cc->migratepages);
1612	2338	cc->nr_migratepages = 0;
1613	2339	goto out;
1614	2340	case ISOLATE_NONE:
	2341	+ if (update_cached) {
	2342	+ cc->zone->compact_cached_migrate_pfn[1] =
	2343	+ cc->zone->compact_cached_migrate_pfn[0];
	2344	+ }
	2345	+
1615	2346	/*
1616	2347	* We haven't isolated and migrated anything, but
1617	2348	* there might still be unflushed migrations from
..	..	@@ -1619,6 +2350,8 @@
1619	2350	*/
1620	2351	goto check_drain;
1621	2352	case ISOLATE_SUCCESS:
	2353	+ update_cached = false;
	2354	+ last_migrated_pfn = start_pfn;
1622	2355	;
1623	2356	}
1624	2357
..	..	@@ -1650,8 +2383,7 @@
1650	2383	cc->migrate_pfn = block_end_pfn(
1651	2384	cc->migrate_pfn - 1, cc->order);
1652	2385	/* Draining pcplists is useless in this case */
1653		- cc->last_migrated_pfn = 0;
1654		-
	2386	+ last_migrated_pfn = 0;
1655	2387	}
1656	2388	}
1657	2389
..	..	@@ -1663,21 +2395,22 @@
1663	2395	* compact_finished() can detect immediately if allocation
1664	2396	* would succeed.
1665	2397	*/
1666		- if (cc->order > 0 && cc->last_migrated_pfn) {
1667		- int cpu;
	2398	+ if (cc->order > 0 && last_migrated_pfn) {
1668	2399	unsigned long current_block_start =
1669	2400	block_start_pfn(cc->migrate_pfn, cc->order);
1670	2401
1671		- if (cc->last_migrated_pfn < current_block_start) {
1672		- cpu = get_cpu();
1673		- lru_add_drain_cpu(cpu);
1674		- drain_local_pages(zone);
1675		- put_cpu();
	2402	+ if (last_migrated_pfn < current_block_start) {
	2403	+ lru_add_drain_cpu_zone(cc->zone);
1676	2404	/* No more flushing until we migrate again */
1677		- cc->last_migrated_pfn = 0;
	2405	+ last_migrated_pfn = 0;
1678	2406	}
1679	2407	}
1680	2408
	2409	+ /* Stop if a page has been captured */
	2410	+ if (capc && capc->page) {
	2411	+ ret = COMPACT_SUCCESS;
	2412	+ break;
	2413	+ }
1681	2414	}
1682	2415
1683	2416	out:
..	..	@@ -1696,8 +2429,8 @@
1696	2429	* Only go back, not forward. The cached pfn might have been
1697	2430	* already reset to zone end in compact_finished()
1698	2431	*/
1699		- if (free_pfn > zone->compact_cached_free_pfn)
1700		- zone->compact_cached_free_pfn = free_pfn;
	2432	+ if (free_pfn > cc->zone->compact_cached_free_pfn)
	2433	+ cc->zone->compact_cached_free_pfn = free_pfn;
1701	2434	}
1702	2435
1703	2436	count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
..	..	@@ -1711,27 +2444,49 @@
1711	2444
1712	2445	static enum compact_result compact_zone_order(struct zone *zone, int order,
1713	2446	gfp_t gfp_mask, enum compact_priority prio,
1714		- unsigned int alloc_flags, int classzone_idx)
	2447	+ unsigned int alloc_flags, int highest_zoneidx,
	2448	+ struct page **capture)
1715	2449	{
1716	2450	enum compact_result ret;
1717	2451	struct compact_control cc = {
1718	2452	.order = order,
	2453	+ .search_order = order,
1719	2454	.gfp_mask = gfp_mask,
1720	2455	.zone = zone,
1721	2456	.mode = (prio == COMPACT_PRIO_ASYNC) ?
1722	2457	MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
1723	2458	.alloc_flags = alloc_flags,
1724		- .classzone_idx = classzone_idx,
	2459	+ .highest_zoneidx = highest_zoneidx,
1725	2460	.direct_compaction = true,
1726	2461	.whole_zone = (prio == MIN_COMPACT_PRIORITY),
1727	2462	.ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
1728	2463	.ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
1729	2464	};
	2465	+ struct capture_control capc = {
	2466	+ .cc = &cc,
	2467	+ .page = NULL,
	2468	+ };
1730	2469
1731		- ret = compact_zone(zone, &cc);
	2470	+ /*
	2471	+ * Make sure the structs are really initialized before we expose the
	2472	+ * capture control, in case we are interrupted and the interrupt handler
	2473	+ * frees a page.
	2474	+ */
	2475	+ barrier();
	2476	+ WRITE_ONCE(current->capture_control, &capc);
	2477	+
	2478	+ ret = compact_zone(&cc, &capc);
1732	2479
1733	2480	VM_BUG_ON(!list_empty(&cc.freepages));
1734	2481	VM_BUG_ON(!list_empty(&cc.migratepages));
	2482	+
	2483	+ /*
	2484	+ * Make sure we hide capture control first before we read the captured
	2485	+ * page pointer, otherwise an interrupt could free and capture a page
	2486	+ * and we would leak it.
	2487	+ */
	2488	+ WRITE_ONCE(current->capture_control, NULL);
	2489	+ *capture = READ_ONCE(capc.page);
1735	2490
1736	2491	return ret;
1737	2492	}
..	..	@@ -1745,12 +2500,13 @@
1745	2500	* @alloc_flags: The allocation flags of the current allocation
1746	2501	* @ac: The context of current allocation
1747	2502	* @prio: Determines how hard direct compaction should try to succeed
	2503	+ * @capture: Pointer to free page created by compaction will be stored here
1748	2504	*
1749	2505	* This is the main entry point for direct page compaction.
1750	2506	*/
1751	2507	enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
1752	2508	unsigned int alloc_flags, const struct alloc_context *ac,
1753		- enum compact_priority prio)
	2509	+ enum compact_priority prio, struct page **capture)
1754	2510	{
1755	2511	int may_perform_io = gfp_mask & __GFP_IO;
1756	2512	struct zoneref *z;
..	..	@@ -1767,8 +2523,8 @@
1767	2523	trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
1768	2524
1769	2525	/* Compact each zone in the list */
1770		- for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1771		- ac->nodemask) {
	2526	+ for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
	2527	+ ac->highest_zoneidx, ac->nodemask) {
1772	2528	enum compact_result status;
1773	2529
1774	2530	if (prio > MIN_COMPACT_PRIORITY
..	..	@@ -1778,7 +2534,7 @@
1778	2534	}
1779	2535
1780	2536	status = compact_zone_order(zone, order, gfp_mask, prio,
1781		- alloc_flags, ac_classzone_idx(ac));
	2537	+ alloc_flags, ac->highest_zoneidx, capture);
1782	2538	rc = max(status, rc);
1783	2539
1784	2540	/* The allocation should succeed, stop compacting */
..	..	@@ -1816,6 +2572,41 @@
1816	2572	return rc;
1817	2573	}
1818	2574
	2575	+/*
	2576	+ * Compact all zones within a node till each zone's fragmentation score
	2577	+ * reaches within proactive compaction thresholds (as determined by the
	2578	+ * proactiveness tunable).
	2579	+ *
	2580	+ * It is possible that the function returns before reaching score targets
	2581	+ * due to various back-off conditions, such as, contention on per-node or
	2582	+ * per-zone locks.
	2583	+ */
	2584	+static void proactive_compact_node(pg_data_t *pgdat)
	2585	+{
	2586	+ int zoneid;
	2587	+ struct zone *zone;
	2588	+ struct compact_control cc = {
	2589	+ .order = -1,
	2590	+ .mode = MIGRATE_SYNC_LIGHT,
	2591	+ .ignore_skip_hint = true,
	2592	+ .whole_zone = true,
	2593	+ .gfp_mask = GFP_KERNEL,
	2594	+ .proactive_compaction = true,
	2595	+ };
	2596	+
	2597	+ for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
	2598	+ zone = &pgdat->node_zones[zoneid];
	2599	+ if (!populated_zone(zone))
	2600	+ continue;
	2601	+
	2602	+ cc.zone = zone;
	2603	+
	2604	+ compact_zone(&cc, NULL);
	2605	+
	2606	+ VM_BUG_ON(!list_empty(&cc.freepages));
	2607	+ VM_BUG_ON(!list_empty(&cc.migratepages));
	2608	+ }
	2609	+}
1819	2610
1820	2611	/* Compact all zones within a node */
1821	2612	static void compact_node(int nid)
..	..	@@ -1840,7 +2631,7 @@
1840	2631
1841	2632	cc.zone = zone;
1842	2633
1843		- compact_zone(zone, &cc);
	2634	+ compact_zone(&cc, NULL);
1844	2635
1845	2636	VM_BUG_ON(!list_empty(&cc.freepages));
1846	2637	VM_BUG_ON(!list_empty(&cc.migratepages));
..	..	@@ -1863,22 +2654,45 @@
1863	2654	int sysctl_compact_memory;
1864	2655
1865	2656	/*
1866		- * This is the entry point for compacting all nodes via
1867		- * /proc/sys/vm/compact_memory
	2657	+ * Tunable for proactive compaction. It determines how
	2658	+ * aggressively the kernel should compact memory in the
	2659	+ * background. It takes values in the range [0, 100].
1868	2660	*/
1869		-int sysctl_compaction_handler(struct ctl_table *table, int write,
1870		- void __user buffer, size_t length, loff_t *ppos)
	2661	+unsigned int __read_mostly sysctl_compaction_proactiveness = 20;
	2662	+
	2663	+int compaction_proactiveness_sysctl_handler(struct ctl_table *table, int write,
	2664	+ void buffer, size_t length, loff_t *ppos)
1871	2665	{
1872		- if (write)
1873		- compact_nodes();
	2666	+ int rc, nid;
	2667	+
	2668	+ rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
	2669	+ if (rc)
	2670	+ return rc;
	2671	+
	2672	+ if (write && sysctl_compaction_proactiveness) {
	2673	+ for_each_online_node(nid) {
	2674	+ pg_data_t *pgdat = NODE_DATA(nid);
	2675	+
	2676	+ if (pgdat->proactive_compact_trigger)
	2677	+ continue;
	2678	+
	2679	+ pgdat->proactive_compact_trigger = true;
	2680	+ wake_up_interruptible(&pgdat->kcompactd_wait);
	2681	+ }
	2682	+ }
1874	2683
1875	2684	return 0;
1876	2685	}
1877	2686
1878		-int sysctl_extfrag_handler(struct ctl_table *table, int write,
1879		- void __user buffer, size_t length, loff_t *ppos)
	2687	+/*
	2688	+ * This is the entry point for compacting all nodes via
	2689	+ * /proc/sys/vm/compact_memory
	2690	+ */
	2691	+int sysctl_compaction_handler(struct ctl_table *table, int write,
	2692	+ void buffer, size_t length, loff_t *ppos)
1880	2693	{
1881		- proc_dointvec_minmax(table, write, buffer, length, ppos);
	2694	+ if (write)
	2695	+ compact_nodes();
1882	2696
1883	2697	return 0;
1884	2698	}
..	..	@@ -1914,23 +2728,24 @@
1914	2728
1915	2729	static inline bool kcompactd_work_requested(pg_data_t *pgdat)
1916	2730	{
1917		- return pgdat->kcompactd_max_order > 0 \|\| kthread_should_stop();
	2731	+ return pgdat->kcompactd_max_order > 0 \|\| kthread_should_stop() \|\|
	2732	+ pgdat->proactive_compact_trigger;
1918	2733	}
1919	2734
1920	2735	static bool kcompactd_node_suitable(pg_data_t *pgdat)
1921	2736	{
1922	2737	int zoneid;
1923	2738	struct zone *zone;
1924		- enum zone_type classzone_idx = pgdat->kcompactd_classzone_idx;
	2739	+ enum zone_type highest_zoneidx = pgdat->kcompactd_highest_zoneidx;
1925	2740
1926		- for (zoneid = 0; zoneid <= classzone_idx; zoneid++) {
	2741	+ for (zoneid = 0; zoneid <= highest_zoneidx; zoneid++) {
1927	2742	zone = &pgdat->node_zones[zoneid];
1928	2743
1929	2744	if (!populated_zone(zone))
1930	2745	continue;
1931	2746
1932	2747	if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0,
1933		- classzone_idx) == COMPACT_CONTINUE)
	2748	+ highest_zoneidx) == COMPACT_CONTINUE)
1934	2749	return true;
1935	2750	}
1936	2751
..	..	@@ -1947,16 +2762,17 @@
1947	2762	struct zone *zone;
1948	2763	struct compact_control cc = {
1949	2764	.order = pgdat->kcompactd_max_order,
1950		- .classzone_idx = pgdat->kcompactd_classzone_idx,
	2765	+ .search_order = pgdat->kcompactd_max_order,
	2766	+ .highest_zoneidx = pgdat->kcompactd_highest_zoneidx,
1951	2767	.mode = MIGRATE_SYNC_LIGHT,
1952	2768	.ignore_skip_hint = false,
1953	2769	.gfp_mask = GFP_KERNEL,
1954	2770	};
1955	2771	trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
1956		- cc.classzone_idx);
	2772	+ cc.highest_zoneidx);
1957	2773	count_compact_event(KCOMPACTD_WAKE);
1958	2774
1959		- for (zoneid = 0; zoneid <= cc.classzone_idx; zoneid++) {
	2775	+ for (zoneid = 0; zoneid <= cc.highest_zoneidx; zoneid++) {
1960	2776	int status;
1961	2777
1962	2778	zone = &pgdat->node_zones[zoneid];
..	..	@@ -1974,7 +2790,7 @@
1974	2790	return;
1975	2791
1976	2792	cc.zone = zone;
1977		- status = compact_zone(zone, &cc);
	2793	+ status = compact_zone(&cc, NULL);
1978	2794
1979	2795	if (status == COMPACT_SUCCESS) {
1980	2796	compaction_defer_reset(zone, cc.order, false);
..	..	@@ -2005,16 +2821,16 @@
2005	2821
2006	2822	/*
2007	2823	* Regardless of success, we are done until woken up next. But remember
2008		- * the requested order/classzone_idx in case it was higher/tighter than
2009		- * our current ones
	2824	+ * the requested order/highest_zoneidx in case it was higher/tighter
	2825	+ * than our current ones
2010	2826	*/
2011	2827	if (pgdat->kcompactd_max_order <= cc.order)
2012	2828	pgdat->kcompactd_max_order = 0;
2013		- if (pgdat->kcompactd_classzone_idx >= cc.classzone_idx)
2014		- pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
	2829	+ if (pgdat->kcompactd_highest_zoneidx >= cc.highest_zoneidx)
	2830	+ pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
2015	2831	}
2016	2832
2017		-void wakeup_kcompactd(pg_data_t *pgdat, int order, int classzone_idx)
	2833	+void wakeup_kcompactd(pg_data_t *pgdat, int order, int highest_zoneidx)
2018	2834	{
2019	2835	if (!order)
2020	2836	return;
..	..	@@ -2022,8 +2838,8 @@
2022	2838	if (pgdat->kcompactd_max_order < order)
2023	2839	pgdat->kcompactd_max_order = order;
2024	2840
2025		- if (pgdat->kcompactd_classzone_idx > classzone_idx)
2026		- pgdat->kcompactd_classzone_idx = classzone_idx;
	2841	+ if (pgdat->kcompactd_highest_zoneidx > highest_zoneidx)
	2842	+ pgdat->kcompactd_highest_zoneidx = highest_zoneidx;
2027	2843
2028	2844	/*
2029	2845	* Pairs with implicit barrier in wait_event_freezable()
..	..	@@ -2036,7 +2852,7 @@
2036	2852	return;
2037	2853
2038	2854	trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
2039		- classzone_idx);
	2855	+ highest_zoneidx);
2040	2856	wake_up_interruptible(&pgdat->kcompactd_wait);
2041	2857	}
2042	2858
..	..	@@ -2048,6 +2864,7 @@
2048	2864	{
2049	2865	pg_data_t pgdat = (pg_data_t)p;
2050	2866	struct task_struct *tsk = current;
	2867	+ unsigned int proactive_defer = 0;
2051	2868
2052	2869	const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2053	2870
..	..	@@ -2057,18 +2874,56 @@
2057	2874	set_freezable();
2058	2875
2059	2876	pgdat->kcompactd_max_order = 0;
2060		- pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
	2877	+ pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
2061	2878
2062	2879	while (!kthread_should_stop()) {
2063	2880	unsigned long pflags;
	2881	+ long timeout;
2064	2882
	2883	+ timeout = sysctl_compaction_proactiveness ?
	2884	+ msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC) :
	2885	+ MAX_SCHEDULE_TIMEOUT;
2065	2886	trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
2066		- wait_event_freezable(pgdat->kcompactd_wait,
2067		- kcompactd_work_requested(pgdat));
	2887	+ if (wait_event_freezable_timeout(pgdat->kcompactd_wait,
	2888	+ kcompactd_work_requested(pgdat), timeout) &&
	2889	+ !pgdat->proactive_compact_trigger) {
2068	2890
2069		- psi_memstall_enter(&pflags);
2070		- kcompactd_do_work(pgdat);
2071		- psi_memstall_leave(&pflags);
	2891	+ psi_memstall_enter(&pflags);
	2892	+ kcompactd_do_work(pgdat);
	2893	+ psi_memstall_leave(&pflags);
	2894	+ continue;
	2895	+ }
	2896	+
	2897	+ /* kcompactd wait timeout */
	2898	+ if (should_proactive_compact_node(pgdat)) {
	2899	+ unsigned int prev_score, score;
	2900	+
	2901	+ /*
	2902	+ * On wakeup of proactive compaction by sysctl
	2903	+ * write, ignore the accumulated defer score.
	2904	+ * Anyway, if the proactive compaction didn't
	2905	+ * make any progress for the new value, it will
	2906	+ * be further deferred by 2^COMPACT_MAX_DEFER_SHIFT
	2907	+ * times.
	2908	+ */
	2909	+ if (proactive_defer &&
	2910	+ !pgdat->proactive_compact_trigger) {
	2911	+ proactive_defer--;
	2912	+ continue;
	2913	+ }
	2914	+
	2915	+ prev_score = fragmentation_score_node(pgdat);
	2916	+ proactive_compact_node(pgdat);
	2917	+ score = fragmentation_score_node(pgdat);
	2918	+ /*
	2919	+ * Defer proactive compaction if the fragmentation
	2920	+ * score did not go down i.e. no progress made.
	2921	+ */
	2922	+ proactive_defer = score < prev_score ?
	2923	+ 0 : 1 << COMPACT_MAX_DEFER_SHIFT;
	2924	+ }
	2925	+ if (pgdat->proactive_compact_trigger)
	2926	+ pgdat->proactive_compact_trigger = false;
2072	2927	}
2073	2928
2074	2929	return 0;