xref: /openbmc/linux/mm/compaction.c (revision 748446bb6b5a9390b546af38ec899c868a9dbcf0)
1 /*
2  * linux/mm/compaction.c
3  *
4  * Memory compaction for the reduction of external fragmentation. Note that
5  * this heavily depends upon page migration to do all the real heavy
6  * lifting
7  *
8  * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie>
9  */
10 #include <linux/swap.h>
11 #include <linux/migrate.h>
12 #include <linux/compaction.h>
13 #include <linux/mm_inline.h>
14 #include <linux/backing-dev.h>
15 #include "internal.h"
16 
17 /*
18  * compact_control is used to track pages being migrated and the free pages
19  * they are being migrated to during memory compaction. The free_pfn starts
20  * at the end of a zone and migrate_pfn begins at the start. Movable pages
21  * are moved to the end of a zone during a compaction run and the run
22  * completes when free_pfn <= migrate_pfn
23  */
24 struct compact_control {
25 	struct list_head freepages;	/* List of free pages to migrate to */
26 	struct list_head migratepages;	/* List of pages being migrated */
27 	unsigned long nr_freepages;	/* Number of isolated free pages */
28 	unsigned long nr_migratepages;	/* Number of pages to migrate */
29 	unsigned long free_pfn;		/* isolate_freepages search base */
30 	unsigned long migrate_pfn;	/* isolate_migratepages search base */
31 
32 	/* Account for isolated anon and file pages */
33 	unsigned long nr_anon;
34 	unsigned long nr_file;
35 
36 	struct zone *zone;
37 };
38 
39 static unsigned long release_freepages(struct list_head *freelist)
40 {
41 	struct page *page, *next;
42 	unsigned long count = 0;
43 
44 	list_for_each_entry_safe(page, next, freelist, lru) {
45 		list_del(&page->lru);
46 		__free_page(page);
47 		count++;
48 	}
49 
50 	return count;
51 }
52 
53 /* Isolate free pages onto a private freelist. Must hold zone->lock */
54 static unsigned long isolate_freepages_block(struct zone *zone,
55 				unsigned long blockpfn,
56 				struct list_head *freelist)
57 {
58 	unsigned long zone_end_pfn, end_pfn;
59 	int total_isolated = 0;
60 	struct page *cursor;
61 
62 	/* Get the last PFN we should scan for free pages at */
63 	zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
64 	end_pfn = min(blockpfn + pageblock_nr_pages, zone_end_pfn);
65 
66 	/* Find the first usable PFN in the block to initialse page cursor */
67 	for (; blockpfn < end_pfn; blockpfn++) {
68 		if (pfn_valid_within(blockpfn))
69 			break;
70 	}
71 	cursor = pfn_to_page(blockpfn);
72 
73 	/* Isolate free pages. This assumes the block is valid */
74 	for (; blockpfn < end_pfn; blockpfn++, cursor++) {
75 		int isolated, i;
76 		struct page *page = cursor;
77 
78 		if (!pfn_valid_within(blockpfn))
79 			continue;
80 
81 		if (!PageBuddy(page))
82 			continue;
83 
84 		/* Found a free page, break it into order-0 pages */
85 		isolated = split_free_page(page);
86 		total_isolated += isolated;
87 		for (i = 0; i < isolated; i++) {
88 			list_add(&page->lru, freelist);
89 			page++;
90 		}
91 
92 		/* If a page was split, advance to the end of it */
93 		if (isolated) {
94 			blockpfn += isolated - 1;
95 			cursor += isolated - 1;
96 		}
97 	}
98 
99 	return total_isolated;
100 }
101 
102 /* Returns true if the page is within a block suitable for migration to */
103 static bool suitable_migration_target(struct page *page)
104 {
105 
106 	int migratetype = get_pageblock_migratetype(page);
107 
108 	/* Don't interfere with memory hot-remove or the min_free_kbytes blocks */
109 	if (migratetype == MIGRATE_ISOLATE || migratetype == MIGRATE_RESERVE)
110 		return false;
111 
112 	/* If the page is a large free page, then allow migration */
113 	if (PageBuddy(page) && page_order(page) >= pageblock_order)
114 		return true;
115 
116 	/* If the block is MIGRATE_MOVABLE, allow migration */
117 	if (migratetype == MIGRATE_MOVABLE)
118 		return true;
119 
120 	/* Otherwise skip the block */
121 	return false;
122 }
123 
124 /*
125  * Based on information in the current compact_control, find blocks
126  * suitable for isolating free pages from and then isolate them.
127  */
128 static void isolate_freepages(struct zone *zone,
129 				struct compact_control *cc)
130 {
131 	struct page *page;
132 	unsigned long high_pfn, low_pfn, pfn;
133 	unsigned long flags;
134 	int nr_freepages = cc->nr_freepages;
135 	struct list_head *freelist = &cc->freepages;
136 
137 	pfn = cc->free_pfn;
138 	low_pfn = cc->migrate_pfn + pageblock_nr_pages;
139 	high_pfn = low_pfn;
140 
141 	/*
142 	 * Isolate free pages until enough are available to migrate the
143 	 * pages on cc->migratepages. We stop searching if the migrate
144 	 * and free page scanners meet or enough free pages are isolated.
145 	 */
146 	spin_lock_irqsave(&zone->lock, flags);
147 	for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages;
148 					pfn -= pageblock_nr_pages) {
149 		unsigned long isolated;
150 
151 		if (!pfn_valid(pfn))
152 			continue;
153 
154 		/*
155 		 * Check for overlapping nodes/zones. It's possible on some
156 		 * configurations to have a setup like
157 		 * node0 node1 node0
158 		 * i.e. it's possible that all pages within a zones range of
159 		 * pages do not belong to a single zone.
160 		 */
161 		page = pfn_to_page(pfn);
162 		if (page_zone(page) != zone)
163 			continue;
164 
165 		/* Check the block is suitable for migration */
166 		if (!suitable_migration_target(page))
167 			continue;
168 
169 		/* Found a block suitable for isolating free pages from */
170 		isolated = isolate_freepages_block(zone, pfn, freelist);
171 		nr_freepages += isolated;
172 
173 		/*
174 		 * Record the highest PFN we isolated pages from. When next
175 		 * looking for free pages, the search will restart here as
176 		 * page migration may have returned some pages to the allocator
177 		 */
178 		if (isolated)
179 			high_pfn = max(high_pfn, pfn);
180 	}
181 	spin_unlock_irqrestore(&zone->lock, flags);
182 
183 	/* split_free_page does not map the pages */
184 	list_for_each_entry(page, freelist, lru) {
185 		arch_alloc_page(page, 0);
186 		kernel_map_pages(page, 1, 1);
187 	}
188 
189 	cc->free_pfn = high_pfn;
190 	cc->nr_freepages = nr_freepages;
191 }
192 
193 /* Update the number of anon and file isolated pages in the zone */
194 static void acct_isolated(struct zone *zone, struct compact_control *cc)
195 {
196 	struct page *page;
197 	unsigned int count[NR_LRU_LISTS] = { 0, };
198 
199 	list_for_each_entry(page, &cc->migratepages, lru) {
200 		int lru = page_lru_base_type(page);
201 		count[lru]++;
202 	}
203 
204 	cc->nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
205 	cc->nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
206 	__mod_zone_page_state(zone, NR_ISOLATED_ANON, cc->nr_anon);
207 	__mod_zone_page_state(zone, NR_ISOLATED_FILE, cc->nr_file);
208 }
209 
210 /* Similar to reclaim, but different enough that they don't share logic */
211 static bool too_many_isolated(struct zone *zone)
212 {
213 
214 	unsigned long inactive, isolated;
215 
216 	inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
217 					zone_page_state(zone, NR_INACTIVE_ANON);
218 	isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
219 					zone_page_state(zone, NR_ISOLATED_ANON);
220 
221 	return isolated > inactive;
222 }
223 
224 /*
225  * Isolate all pages that can be migrated from the block pointed to by
226  * the migrate scanner within compact_control.
227  */
228 static unsigned long isolate_migratepages(struct zone *zone,
229 					struct compact_control *cc)
230 {
231 	unsigned long low_pfn, end_pfn;
232 	struct list_head *migratelist = &cc->migratepages;
233 
234 	/* Do not scan outside zone boundaries */
235 	low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
236 
237 	/* Only scan within a pageblock boundary */
238 	end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages);
239 
240 	/* Do not cross the free scanner or scan within a memory hole */
241 	if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
242 		cc->migrate_pfn = end_pfn;
243 		return 0;
244 	}
245 
246 	/*
247 	 * Ensure that there are not too many pages isolated from the LRU
248 	 * list by either parallel reclaimers or compaction. If there are,
249 	 * delay for some time until fewer pages are isolated
250 	 */
251 	while (unlikely(too_many_isolated(zone))) {
252 		congestion_wait(BLK_RW_ASYNC, HZ/10);
253 
254 		if (fatal_signal_pending(current))
255 			return 0;
256 	}
257 
258 	/* Time to isolate some pages for migration */
259 	spin_lock_irq(&zone->lru_lock);
260 	for (; low_pfn < end_pfn; low_pfn++) {
261 		struct page *page;
262 		if (!pfn_valid_within(low_pfn))
263 			continue;
264 
265 		/* Get the page and skip if free */
266 		page = pfn_to_page(low_pfn);
267 		if (PageBuddy(page))
268 			continue;
269 
270 		/* Try isolate the page */
271 		if (__isolate_lru_page(page, ISOLATE_BOTH, 0) != 0)
272 			continue;
273 
274 		/* Successfully isolated */
275 		del_page_from_lru_list(zone, page, page_lru(page));
276 		list_add(&page->lru, migratelist);
277 		mem_cgroup_del_lru(page);
278 		cc->nr_migratepages++;
279 
280 		/* Avoid isolating too much */
281 		if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
282 			break;
283 	}
284 
285 	acct_isolated(zone, cc);
286 
287 	spin_unlock_irq(&zone->lru_lock);
288 	cc->migrate_pfn = low_pfn;
289 
290 	return cc->nr_migratepages;
291 }
292 
293 /*
294  * This is a migrate-callback that "allocates" freepages by taking pages
295  * from the isolated freelists in the block we are migrating to.
296  */
297 static struct page *compaction_alloc(struct page *migratepage,
298 					unsigned long data,
299 					int **result)
300 {
301 	struct compact_control *cc = (struct compact_control *)data;
302 	struct page *freepage;
303 
304 	/* Isolate free pages if necessary */
305 	if (list_empty(&cc->freepages)) {
306 		isolate_freepages(cc->zone, cc);
307 
308 		if (list_empty(&cc->freepages))
309 			return NULL;
310 	}
311 
312 	freepage = list_entry(cc->freepages.next, struct page, lru);
313 	list_del(&freepage->lru);
314 	cc->nr_freepages--;
315 
316 	return freepage;
317 }
318 
319 /*
320  * We cannot control nr_migratepages and nr_freepages fully when migration is
321  * running as migrate_pages() has no knowledge of compact_control. When
322  * migration is complete, we count the number of pages on the lists by hand.
323  */
324 static void update_nr_listpages(struct compact_control *cc)
325 {
326 	int nr_migratepages = 0;
327 	int nr_freepages = 0;
328 	struct page *page;
329 
330 	list_for_each_entry(page, &cc->migratepages, lru)
331 		nr_migratepages++;
332 	list_for_each_entry(page, &cc->freepages, lru)
333 		nr_freepages++;
334 
335 	cc->nr_migratepages = nr_migratepages;
336 	cc->nr_freepages = nr_freepages;
337 }
338 
339 static int compact_finished(struct zone *zone,
340 						struct compact_control *cc)
341 {
342 	if (fatal_signal_pending(current))
343 		return COMPACT_PARTIAL;
344 
345 	/* Compaction run completes if the migrate and free scanner meet */
346 	if (cc->free_pfn <= cc->migrate_pfn)
347 		return COMPACT_COMPLETE;
348 
349 	return COMPACT_CONTINUE;
350 }
351 
352 static int compact_zone(struct zone *zone, struct compact_control *cc)
353 {
354 	int ret;
355 
356 	/* Setup to move all movable pages to the end of the zone */
357 	cc->migrate_pfn = zone->zone_start_pfn;
358 	cc->free_pfn = cc->migrate_pfn + zone->spanned_pages;
359 	cc->free_pfn &= ~(pageblock_nr_pages-1);
360 
361 	migrate_prep_local();
362 
363 	while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
364 		unsigned long nr_migrate, nr_remaining;
365 
366 		if (!isolate_migratepages(zone, cc))
367 			continue;
368 
369 		nr_migrate = cc->nr_migratepages;
370 		migrate_pages(&cc->migratepages, compaction_alloc,
371 						(unsigned long)cc, 0);
372 		update_nr_listpages(cc);
373 		nr_remaining = cc->nr_migratepages;
374 
375 		count_vm_event(COMPACTBLOCKS);
376 		count_vm_events(COMPACTPAGES, nr_migrate - nr_remaining);
377 		if (nr_remaining)
378 			count_vm_events(COMPACTPAGEFAILED, nr_remaining);
379 
380 		/* Release LRU pages not migrated */
381 		if (!list_empty(&cc->migratepages)) {
382 			putback_lru_pages(&cc->migratepages);
383 			cc->nr_migratepages = 0;
384 		}
385 
386 	}
387 
388 	/* Release free pages and check accounting */
389 	cc->nr_freepages -= release_freepages(&cc->freepages);
390 	VM_BUG_ON(cc->nr_freepages != 0);
391 
392 	return ret;
393 }
394