Diff markup
1 /* 1 /*
2 * linux/mm/vmscan.c 2 * linux/mm/vmscan.c
3 * 3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linu 4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 * 5 *
6 * Swap reorganised 29.12.95, Stephen Tweedie 6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct 7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as 8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high 9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sar 10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel 11 * Multiqueue VM started 5.8.00, Rik van Riel.
12 */ 12 */
13 13
14 #include <linux/mm.h> 14 #include <linux/mm.h>
15 #include <linux/module.h> 15 #include <linux/module.h>
16 #include <linux/slab.h> 16 #include <linux/slab.h>
17 #include <linux/kernel_stat.h> 17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h> 18 #include <linux/swap.h>
19 #include <linux/pagemap.h> 19 #include <linux/pagemap.h>
20 #include <linux/init.h> 20 #include <linux/init.h>
21 #include <linux/highmem.h> 21 #include <linux/highmem.h>
22 #include <linux/vmstat.h> 22 #include <linux/vmstat.h>
23 #include <linux/file.h> 23 #include <linux/file.h>
24 #include <linux/writeback.h> 24 #include <linux/writeback.h>
25 #include <linux/blkdev.h> 25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h> /* for try_to_ 26 #include <linux/buffer_head.h> /* for try_to_release_page(),
27 buffer 27 buffer_heads_over_limit */
28 #include <linux/mm_inline.h> 28 #include <linux/mm_inline.h>
29 #include <linux/pagevec.h> 29 #include <linux/pagevec.h>
30 #include <linux/backing-dev.h> 30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h> 31 #include <linux/rmap.h>
32 #include <linux/topology.h> 32 #include <linux/topology.h>
33 #include <linux/cpu.h> 33 #include <linux/cpu.h>
34 #include <linux/cpuset.h> 34 #include <linux/cpuset.h>
35 #include <linux/notifier.h> 35 #include <linux/notifier.h>
36 #include <linux/rwsem.h> 36 #include <linux/rwsem.h>
37 #include <linux/delay.h> 37 #include <linux/delay.h>
38 #include <linux/kthread.h> 38 #include <linux/kthread.h>
39 #include <linux/freezer.h> 39 #include <linux/freezer.h>
40 #include <linux/memcontrol.h> 40 #include <linux/memcontrol.h>
>> 41 #include <linux/delayacct.h>
>> 42 #include <linux/sysctl.h>
41 43
42 #include <asm/tlbflush.h> 44 #include <asm/tlbflush.h>
43 #include <asm/div64.h> 45 #include <asm/div64.h>
44 46
45 #include <linux/swapops.h> 47 #include <linux/swapops.h>
46 48
47 #include "internal.h" 49 #include "internal.h"
48 50
49 struct scan_control { 51 struct scan_control {
50 /* Incremented by the number of inacti 52 /* Incremented by the number of inactive pages that were scanned */
51 unsigned long nr_scanned; 53 unsigned long nr_scanned;
52 54
>> 55 /* Number of pages freed so far during a call to shrink_zones() */
>> 56 unsigned long nr_reclaimed;
>> 57
53 /* This context's GFP mask */ 58 /* This context's GFP mask */
54 gfp_t gfp_mask; 59 gfp_t gfp_mask;
55 60
56 int may_writepage; 61 int may_writepage;
57 62
>> 63 /* Can mapped pages be reclaimed? */
>> 64 int may_unmap;
>> 65
58 /* Can pages be swapped as part of rec 66 /* Can pages be swapped as part of reclaim? */
59 int may_swap; 67 int may_swap;
60 68
61 /* This context's SWAP_CLUSTER_MAX. If 69 /* This context's SWAP_CLUSTER_MAX. If freeing memory for
62 * suspend, we effectively ignore SWAP 70 * suspend, we effectively ignore SWAP_CLUSTER_MAX.
63 * In this context, it doesn't matter 71 * In this context, it doesn't matter that we scan the
64 * whole list at once. */ 72 * whole list at once. */
65 int swap_cluster_max; 73 int swap_cluster_max;
66 74
67 int swappiness; 75 int swappiness;
68 76
69 int all_unreclaimable; 77 int all_unreclaimable;
70 78
71 int order; 79 int order;
72 80
73 /* Which cgroup do we reclaim from */ 81 /* Which cgroup do we reclaim from */
74 struct mem_cgroup *mem_cgroup; 82 struct mem_cgroup *mem_cgroup;
75 83
>> 84 /*
>> 85 * Nodemask of nodes allowed by the caller. If NULL, all nodes
>> 86 * are scanned.
>> 87 */
>> 88 nodemask_t *nodemask;
>> 89
76 /* Pluggable isolate pages callback */ 90 /* Pluggable isolate pages callback */
77 unsigned long (*isolate_pages)(unsigne 91 unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst,
78 unsigned long *scanned 92 unsigned long *scanned, int order, int mode,
79 struct zone *z, struct 93 struct zone *z, struct mem_cgroup *mem_cont,
80 int active); !! 94 int active, int file);
81 }; 95 };
82 96
83 #define lru_to_page(_head) (list_entry((_head) 97 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
84 98
85 #ifdef ARCH_HAS_PREFETCH 99 #ifdef ARCH_HAS_PREFETCH
86 #define prefetch_prev_lru_page(_page, _base, _ 100 #define prefetch_prev_lru_page(_page, _base, _field) \
87 do { 101 do { \
88 if ((_page)->lru.prev != _base 102 if ((_page)->lru.prev != _base) { \
89 struct page *prev; 103 struct page *prev; \
90 104 \
91 prev = lru_to_page(&(_ 105 prev = lru_to_page(&(_page->lru)); \
92 prefetch(&prev->_field 106 prefetch(&prev->_field); \
93 } 107 } \
94 } while (0) 108 } while (0)
95 #else 109 #else
96 #define prefetch_prev_lru_page(_page, _base, _ 110 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
97 #endif 111 #endif
98 112
99 #ifdef ARCH_HAS_PREFETCHW 113 #ifdef ARCH_HAS_PREFETCHW
100 #define prefetchw_prev_lru_page(_page, _base, 114 #define prefetchw_prev_lru_page(_page, _base, _field) \
101 do { 115 do { \
102 if ((_page)->lru.prev != _base 116 if ((_page)->lru.prev != _base) { \
103 struct page *prev; 117 struct page *prev; \
104 118 \
105 prev = lru_to_page(&(_ 119 prev = lru_to_page(&(_page->lru)); \
106 prefetchw(&prev->_fiel 120 prefetchw(&prev->_field); \
107 } 121 } \
108 } while (0) 122 } while (0)
109 #else 123 #else
110 #define prefetchw_prev_lru_page(_page, _base, 124 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
111 #endif 125 #endif
112 126
113 /* 127 /*
114 * From 0 .. 100. Higher means more swappy. 128 * From 0 .. 100. Higher means more swappy.
115 */ 129 */
116 int vm_swappiness = 60; 130 int vm_swappiness = 60;
117 long vm_total_pages; /* The total number of 131 long vm_total_pages; /* The total number of pages which the VM controls */
118 132
119 static LIST_HEAD(shrinker_list); 133 static LIST_HEAD(shrinker_list);
120 static DECLARE_RWSEM(shrinker_rwsem); 134 static DECLARE_RWSEM(shrinker_rwsem);
121 135
122 #ifdef CONFIG_CGROUP_MEM_RES_CTLR 136 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
123 #define scan_global_lru(sc) (!(sc)->mem_cg !! 137 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
124 #else 138 #else
125 #define scan_global_lru(sc) (1) !! 139 #define scanning_global_lru(sc) (1)
126 #endif 140 #endif
127 141
>> 142 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
>> 143 struct scan_control *sc)
>> 144 {
>> 145 if (!scanning_global_lru(sc))
>> 146 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
>> 147
>> 148 return &zone->reclaim_stat;
>> 149 }
>> 150
>> 151 static unsigned long zone_nr_pages(struct zone *zone, struct scan_control *sc,
>> 152 enum lru_list lru)
>> 153 {
>> 154 if (!scanning_global_lru(sc))
>> 155 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
>> 156
>> 157 return zone_page_state(zone, NR_LRU_BASE + lru);
>> 158 }
>> 159
>> 160
128 /* 161 /*
129 * Add a shrinker callback to be called from t 162 * Add a shrinker callback to be called from the vm
130 */ 163 */
131 void register_shrinker(struct shrinker *shrink 164 void register_shrinker(struct shrinker *shrinker)
132 { 165 {
133 shrinker->nr = 0; 166 shrinker->nr = 0;
134 down_write(&shrinker_rwsem); 167 down_write(&shrinker_rwsem);
135 list_add_tail(&shrinker->list, &shrink 168 list_add_tail(&shrinker->list, &shrinker_list);
136 up_write(&shrinker_rwsem); 169 up_write(&shrinker_rwsem);
137 } 170 }
138 EXPORT_SYMBOL(register_shrinker); 171 EXPORT_SYMBOL(register_shrinker);
139 172
140 /* 173 /*
141 * Remove one 174 * Remove one
142 */ 175 */
143 void unregister_shrinker(struct shrinker *shri 176 void unregister_shrinker(struct shrinker *shrinker)
144 { 177 {
145 down_write(&shrinker_rwsem); 178 down_write(&shrinker_rwsem);
146 list_del(&shrinker->list); 179 list_del(&shrinker->list);
147 up_write(&shrinker_rwsem); 180 up_write(&shrinker_rwsem);
148 } 181 }
149 EXPORT_SYMBOL(unregister_shrinker); 182 EXPORT_SYMBOL(unregister_shrinker);
150 183
151 #define SHRINK_BATCH 128 184 #define SHRINK_BATCH 128
152 /* 185 /*
153 * Call the shrink functions to age shrinkable 186 * Call the shrink functions to age shrinkable caches
154 * 187 *
155 * Here we assume it costs one seek to replace 188 * Here we assume it costs one seek to replace a lru page and that it also
156 * takes a seek to recreate a cache object. W 189 * takes a seek to recreate a cache object. With this in mind we age equal
157 * percentages of the lru and ageable caches. 190 * percentages of the lru and ageable caches. This should balance the seeks
158 * generated by these structures. 191 * generated by these structures.
159 * 192 *
160 * If the vm encountered mapped pages on the L 193 * If the vm encountered mapped pages on the LRU it increase the pressure on
161 * slab to avoid swapping. 194 * slab to avoid swapping.
162 * 195 *
163 * We do weird things to avoid (scanned*seeks* 196 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
164 * 197 *
165 * `lru_pages' represents the number of on-LRU 198 * `lru_pages' represents the number of on-LRU pages in all the zones which
166 * are eligible for the caller's allocation at 199 * are eligible for the caller's allocation attempt. It is used for balancing
167 * slab reclaim versus page reclaim. 200 * slab reclaim versus page reclaim.
168 * 201 *
169 * Returns the number of slab objects which we 202 * Returns the number of slab objects which we shrunk.
170 */ 203 */
171 unsigned long shrink_slab(unsigned long scanne 204 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
172 unsigned long lru_page 205 unsigned long lru_pages)
173 { 206 {
174 struct shrinker *shrinker; 207 struct shrinker *shrinker;
175 unsigned long ret = 0; 208 unsigned long ret = 0;
176 209
177 if (scanned == 0) 210 if (scanned == 0)
178 scanned = SWAP_CLUSTER_MAX; 211 scanned = SWAP_CLUSTER_MAX;
179 212
180 if (!down_read_trylock(&shrinker_rwsem 213 if (!down_read_trylock(&shrinker_rwsem))
181 return 1; /* Assume we'l 214 return 1; /* Assume we'll be able to shrink next time */
182 215
183 list_for_each_entry(shrinker, &shrinke 216 list_for_each_entry(shrinker, &shrinker_list, list) {
184 unsigned long long delta; 217 unsigned long long delta;
185 unsigned long total_scan; 218 unsigned long total_scan;
186 unsigned long max_pass = (*shr 219 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
187 220
188 delta = (4 * scanned) / shrink 221 delta = (4 * scanned) / shrinker->seeks;
189 delta *= max_pass; 222 delta *= max_pass;
190 do_div(delta, lru_pages + 1); 223 do_div(delta, lru_pages + 1);
191 shrinker->nr += delta; 224 shrinker->nr += delta;
192 if (shrinker->nr < 0) { 225 if (shrinker->nr < 0) {
193 printk(KERN_ERR "%s: n !! 226 printk(KERN_ERR "shrink_slab: %pF negative objects to "
194 __FUNC !! 227 "delete nr=%ld\n",
>> 228 shrinker->shrink, shrinker->nr);
195 shrinker->nr = max_pas 229 shrinker->nr = max_pass;
196 } 230 }
197 231
198 /* 232 /*
199 * Avoid risking looping forev 233 * Avoid risking looping forever due to too large nr value:
200 * never try to free more than 234 * never try to free more than twice the estimate number of
201 * freeable entries. 235 * freeable entries.
202 */ 236 */
203 if (shrinker->nr > max_pass * 237 if (shrinker->nr > max_pass * 2)
204 shrinker->nr = max_pas 238 shrinker->nr = max_pass * 2;
205 239
206 total_scan = shrinker->nr; 240 total_scan = shrinker->nr;
207 shrinker->nr = 0; 241 shrinker->nr = 0;
208 242
209 while (total_scan >= SHRINK_BA 243 while (total_scan >= SHRINK_BATCH) {
210 long this_scan = SHRIN 244 long this_scan = SHRINK_BATCH;
211 int shrink_ret; 245 int shrink_ret;
212 int nr_before; 246 int nr_before;
213 247
214 nr_before = (*shrinker 248 nr_before = (*shrinker->shrink)(0, gfp_mask);
215 shrink_ret = (*shrinke 249 shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
216 if (shrink_ret == -1) 250 if (shrink_ret == -1)
217 break; 251 break;
218 if (shrink_ret < nr_be 252 if (shrink_ret < nr_before)
219 ret += nr_befo 253 ret += nr_before - shrink_ret;
220 count_vm_events(SLABS_ 254 count_vm_events(SLABS_SCANNED, this_scan);
221 total_scan -= this_sca 255 total_scan -= this_scan;
222 256
223 cond_resched(); 257 cond_resched();
224 } 258 }
225 259
226 shrinker->nr += total_scan; 260 shrinker->nr += total_scan;
227 } 261 }
228 up_read(&shrinker_rwsem); 262 up_read(&shrinker_rwsem);
229 return ret; 263 return ret;
230 } 264 }
231 265
232 /* Called without lock on whether page is mapp 266 /* Called without lock on whether page is mapped, so answer is unstable */
233 static inline int page_mapping_inuse(struct pa 267 static inline int page_mapping_inuse(struct page *page)
234 { 268 {
235 struct address_space *mapping; 269 struct address_space *mapping;
236 270
237 /* Page is in somebody's page tables. 271 /* Page is in somebody's page tables. */
238 if (page_mapped(page)) 272 if (page_mapped(page))
239 return 1; 273 return 1;
240 274
241 /* Be more reluctant to reclaim swapca 275 /* Be more reluctant to reclaim swapcache than pagecache */
242 if (PageSwapCache(page)) 276 if (PageSwapCache(page))
243 return 1; 277 return 1;
244 278
245 mapping = page_mapping(page); 279 mapping = page_mapping(page);
246 if (!mapping) 280 if (!mapping)
247 return 0; 281 return 0;
248 282
249 /* File is mmap'd by somebody? */ 283 /* File is mmap'd by somebody? */
250 return mapping_mapped(mapping); 284 return mapping_mapped(mapping);
251 } 285 }
252 286
253 static inline int is_page_cache_freeable(struc 287 static inline int is_page_cache_freeable(struct page *page)
254 { 288 {
255 return page_count(page) - !!PagePrivat !! 289 return page_count(page) - !!page_has_private(page) == 2;
256 } 290 }
257 291
258 static int may_write_to_queue(struct backing_d 292 static int may_write_to_queue(struct backing_dev_info *bdi)
259 { 293 {
260 if (current->flags & PF_SWAPWRITE) 294 if (current->flags & PF_SWAPWRITE)
261 return 1; 295 return 1;
262 if (!bdi_write_congested(bdi)) 296 if (!bdi_write_congested(bdi))
263 return 1; 297 return 1;
264 if (bdi == current->backing_dev_info) 298 if (bdi == current->backing_dev_info)
265 return 1; 299 return 1;
266 return 0; 300 return 0;
267 } 301 }
268 302
269 /* 303 /*
270 * We detected a synchronous write error writi 304 * We detected a synchronous write error writing a page out. Probably
271 * -ENOSPC. We need to propagate that into th 305 * -ENOSPC. We need to propagate that into the address_space for a subsequent
272 * fsync(), msync() or close(). 306 * fsync(), msync() or close().
273 * 307 *
274 * The tricky part is that after writepage we 308 * The tricky part is that after writepage we cannot touch the mapping: nothing
275 * prevents it from being freed up. But we ha 309 * prevents it from being freed up. But we have a ref on the page and once
276 * that page is locked, the mapping is pinned. 310 * that page is locked, the mapping is pinned.
277 * 311 *
278 * We're allowed to run sleeping lock_page() h 312 * We're allowed to run sleeping lock_page() here because we know the caller has
279 * __GFP_FS. 313 * __GFP_FS.
280 */ 314 */
281 static void handle_write_error(struct address_ 315 static void handle_write_error(struct address_space *mapping,
282 struct page *p 316 struct page *page, int error)
283 { 317 {
284 lock_page(page); 318 lock_page(page);
285 if (page_mapping(page) == mapping) 319 if (page_mapping(page) == mapping)
286 mapping_set_error(mapping, err 320 mapping_set_error(mapping, error);
287 unlock_page(page); 321 unlock_page(page);
288 } 322 }
289 323
290 /* Request for sync pageout. */ 324 /* Request for sync pageout. */
291 enum pageout_io { 325 enum pageout_io {
292 PAGEOUT_IO_ASYNC, 326 PAGEOUT_IO_ASYNC,
293 PAGEOUT_IO_SYNC, 327 PAGEOUT_IO_SYNC,
294 }; 328 };
295 329
296 /* possible outcome of pageout() */ 330 /* possible outcome of pageout() */
297 typedef enum { 331 typedef enum {
298 /* failed to write page out, page is l 332 /* failed to write page out, page is locked */
299 PAGE_KEEP, 333 PAGE_KEEP,
300 /* move page to the active list, page 334 /* move page to the active list, page is locked */
301 PAGE_ACTIVATE, 335 PAGE_ACTIVATE,
302 /* page has been sent to the disk succ 336 /* page has been sent to the disk successfully, page is unlocked */
303 PAGE_SUCCESS, 337 PAGE_SUCCESS,
304 /* page is clean and locked */ 338 /* page is clean and locked */
305 PAGE_CLEAN, 339 PAGE_CLEAN,
306 } pageout_t; 340 } pageout_t;
307 341
308 /* 342 /*
309 * pageout is called by shrink_page_list() for 343 * pageout is called by shrink_page_list() for each dirty page.
310 * Calls ->writepage(). 344 * Calls ->writepage().
311 */ 345 */
312 static pageout_t pageout(struct page *page, st 346 static pageout_t pageout(struct page *page, struct address_space *mapping,
313 347 enum pageout_io sync_writeback)
314 { 348 {
315 /* 349 /*
316 * If the page is dirty, only perform 350 * If the page is dirty, only perform writeback if that write
317 * will be non-blocking. To prevent t 351 * will be non-blocking. To prevent this allocation from being
318 * stalled by pagecache activity. But 352 * stalled by pagecache activity. But note that there may be
319 * stalls if we need to run get_block( 353 * stalls if we need to run get_block(). We could test
320 * PagePrivate for that. 354 * PagePrivate for that.
321 * 355 *
322 * If this process is currently in gen 356 * If this process is currently in generic_file_write() against
323 * this page's queue, we can perform w 357 * this page's queue, we can perform writeback even if that
324 * will block. 358 * will block.
325 * 359 *
326 * If the page is swapcache, write it 360 * If the page is swapcache, write it back even if that would
327 * block, for some throttling. This ha 361 * block, for some throttling. This happens by accident, because
328 * swap_backing_dev_info is bust: it d 362 * swap_backing_dev_info is bust: it doesn't reflect the
329 * congestion state of the swapdevs. 363 * congestion state of the swapdevs. Easy to fix, if needed.
330 * See swapfile.c:page_queue_congested 364 * See swapfile.c:page_queue_congested().
331 */ 365 */
332 if (!is_page_cache_freeable(page)) 366 if (!is_page_cache_freeable(page))
333 return PAGE_KEEP; 367 return PAGE_KEEP;
334 if (!mapping) { 368 if (!mapping) {
335 /* 369 /*
336 * Some data journaling orphan 370 * Some data journaling orphaned pages can have
337 * page->mapping == NULL while 371 * page->mapping == NULL while being dirty with clean buffers.
338 */ 372 */
339 if (PagePrivate(page)) { !! 373 if (page_has_private(page)) {
340 if (try_to_free_buffer 374 if (try_to_free_buffers(page)) {
341 ClearPageDirty 375 ClearPageDirty(page);
342 printk("%s: or !! 376 printk("%s: orphaned page\n", __func__);
343 return PAGE_CL 377 return PAGE_CLEAN;
344 } 378 }
345 } 379 }
346 return PAGE_KEEP; 380 return PAGE_KEEP;
347 } 381 }
348 if (mapping->a_ops->writepage == NULL) 382 if (mapping->a_ops->writepage == NULL)
349 return PAGE_ACTIVATE; 383 return PAGE_ACTIVATE;
350 if (!may_write_to_queue(mapping->backi 384 if (!may_write_to_queue(mapping->backing_dev_info))
351 return PAGE_KEEP; 385 return PAGE_KEEP;
352 386
353 if (clear_page_dirty_for_io(page)) { 387 if (clear_page_dirty_for_io(page)) {
354 int res; 388 int res;
355 struct writeback_control wbc = 389 struct writeback_control wbc = {
356 .sync_mode = WB_SYNC_N 390 .sync_mode = WB_SYNC_NONE,
357 .nr_to_write = SWAP_CL 391 .nr_to_write = SWAP_CLUSTER_MAX,
358 .range_start = 0, 392 .range_start = 0,
359 .range_end = LLONG_MAX 393 .range_end = LLONG_MAX,
360 .nonblocking = 1, 394 .nonblocking = 1,
361 .for_reclaim = 1, 395 .for_reclaim = 1,
362 }; 396 };
363 397
364 SetPageReclaim(page); 398 SetPageReclaim(page);
365 res = mapping->a_ops->writepag 399 res = mapping->a_ops->writepage(page, &wbc);
366 if (res < 0) 400 if (res < 0)
367 handle_write_error(map 401 handle_write_error(mapping, page, res);
368 if (res == AOP_WRITEPAGE_ACTIV 402 if (res == AOP_WRITEPAGE_ACTIVATE) {
369 ClearPageReclaim(page) 403 ClearPageReclaim(page);
370 return PAGE_ACTIVATE; 404 return PAGE_ACTIVATE;
371 } 405 }
372 406
373 /* 407 /*
374 * Wait on writeback if reques 408 * Wait on writeback if requested to. This happens when
375 * direct reclaiming a large c 409 * direct reclaiming a large contiguous area and the
376 * first attempt to free a ran 410 * first attempt to free a range of pages fails.
377 */ 411 */
378 if (PageWriteback(page) && syn 412 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
379 wait_on_page_writeback 413 wait_on_page_writeback(page);
380 414
381 if (!PageWriteback(page)) { 415 if (!PageWriteback(page)) {
382 /* synchronous write o 416 /* synchronous write or broken a_ops? */
383 ClearPageReclaim(page) 417 ClearPageReclaim(page);
384 } 418 }
385 inc_zone_page_state(page, NR_V 419 inc_zone_page_state(page, NR_VMSCAN_WRITE);
386 return PAGE_SUCCESS; 420 return PAGE_SUCCESS;
387 } 421 }
388 422
389 return PAGE_CLEAN; 423 return PAGE_CLEAN;
390 } 424 }
391 425
392 /* 426 /*
393 * Attempt to detach a locked page from its -> !! 427 * Same as remove_mapping, but if the page is removed from the mapping, it
394 * someone else has a ref on the page, abort a !! 428 * gets returned with a refcount of 0.
395 * successfully detached, return 1. Assumes t <<
396 * this page. <<
397 */ 429 */
398 int remove_mapping(struct address_space *mappi !! 430 static int __remove_mapping(struct address_space *mapping, struct page *page)
399 { 431 {
400 BUG_ON(!PageLocked(page)); 432 BUG_ON(!PageLocked(page));
401 BUG_ON(mapping != page_mapping(page)); 433 BUG_ON(mapping != page_mapping(page));
402 434
403 write_lock_irq(&mapping->tree_lock); !! 435 spin_lock_irq(&mapping->tree_lock);
404 /* 436 /*
405 * The non racy check for a busy page. 437 * The non racy check for a busy page.
406 * 438 *
407 * Must be careful with the order of t 439 * Must be careful with the order of the tests. When someone has
408 * a ref to the page, it may be possib 440 * a ref to the page, it may be possible that they dirty it then
409 * drop the reference. So if PageDirty 441 * drop the reference. So if PageDirty is tested before page_count
410 * here, then the following race may o 442 * here, then the following race may occur:
411 * 443 *
412 * get_user_pages(&page); 444 * get_user_pages(&page);
413 * [user mapping goes away] 445 * [user mapping goes away]
414 * write_to(page); 446 * write_to(page);
415 * !PageD 447 * !PageDirty(page) [good]
416 * SetPageDirty(page); 448 * SetPageDirty(page);
417 * put_page(page); 449 * put_page(page);
418 * !page_ 450 * !page_count(page) [good, discard it]
419 * 451 *
420 * [oops, our write_to data is lost] 452 * [oops, our write_to data is lost]
421 * 453 *
422 * Reversing the order of the tests en 454 * Reversing the order of the tests ensures such a situation cannot
423 * escape unnoticed. The smp_rmb is ne 455 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
424 * load is not satisfied before that o 456 * load is not satisfied before that of page->_count.
425 * 457 *
426 * Note that if SetPageDirty is always 458 * Note that if SetPageDirty is always performed via set_page_dirty,
427 * and thus under tree_lock, then this 459 * and thus under tree_lock, then this ordering is not required.
428 */ 460 */
429 if (unlikely(page_count(page) != 2)) !! 461 if (!page_freeze_refs(page, 2))
430 goto cannot_free; 462 goto cannot_free;
431 smp_rmb(); !! 463 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
432 if (unlikely(PageDirty(page))) !! 464 if (unlikely(PageDirty(page))) {
>> 465 page_unfreeze_refs(page, 2);
433 goto cannot_free; 466 goto cannot_free;
>> 467 }
434 468
435 if (PageSwapCache(page)) { 469 if (PageSwapCache(page)) {
436 swp_entry_t swap = { .val = pa 470 swp_entry_t swap = { .val = page_private(page) };
437 __delete_from_swap_cache(page) 471 __delete_from_swap_cache(page);
438 write_unlock_irq(&mapping->tre !! 472 spin_unlock_irq(&mapping->tree_lock);
439 swap_free(swap); !! 473 swapcache_free(swap, page);
440 __put_page(page); /* The !! 474 } else {
441 return 1; !! 475 __remove_from_page_cache(page);
>> 476 spin_unlock_irq(&mapping->tree_lock);
>> 477 mem_cgroup_uncharge_cache_page(page);
442 } 478 }
443 479
444 __remove_from_page_cache(page); <<
445 write_unlock_irq(&mapping->tree_lock); <<
446 __put_page(page); <<
447 return 1; 480 return 1;
448 481
449 cannot_free: 482 cannot_free:
450 write_unlock_irq(&mapping->tree_lock); !! 483 spin_unlock_irq(&mapping->tree_lock);
>> 484 return 0;
>> 485 }
>> 486
>> 487 /*
>> 488 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
>> 489 * someone else has a ref on the page, abort and return 0. If it was
>> 490 * successfully detached, return 1. Assumes the caller has a single ref on
>> 491 * this page.
>> 492 */
>> 493 int remove_mapping(struct address_space *mapping, struct page *page)
>> 494 {
>> 495 if (__remove_mapping(mapping, page)) {
>> 496 /*
>> 497 * Unfreezing the refcount with 1 rather than 2 effectively
>> 498 * drops the pagecache ref for us without requiring another
>> 499 * atomic operation.
>> 500 */
>> 501 page_unfreeze_refs(page, 1);
>> 502 return 1;
>> 503 }
451 return 0; 504 return 0;
452 } 505 }
453 506
>> 507 /**
>> 508 * putback_lru_page - put previously isolated page onto appropriate LRU list
>> 509 * @page: page to be put back to appropriate lru list
>> 510 *
>> 511 * Add previously isolated @page to appropriate LRU list.
>> 512 * Page may still be unevictable for other reasons.
>> 513 *
>> 514 * lru_lock must not be held, interrupts must be enabled.
>> 515 */
>> 516 void putback_lru_page(struct page *page)
>> 517 {
>> 518 int lru;
>> 519 int active = !!TestClearPageActive(page);
>> 520 int was_unevictable = PageUnevictable(page);
>> 521
>> 522 VM_BUG_ON(PageLRU(page));
>> 523
>> 524 redo:
>> 525 ClearPageUnevictable(page);
>> 526
>> 527 if (page_evictable(page, NULL)) {
>> 528 /*
>> 529 * For evictable pages, we can use the cache.
>> 530 * In event of a race, worst case is we end up with an
>> 531 * unevictable page on [in]active list.
>> 532 * We know how to handle that.
>> 533 */
>> 534 lru = active + page_is_file_cache(page);
>> 535 lru_cache_add_lru(page, lru);
>> 536 } else {
>> 537 /*
>> 538 * Put unevictable pages directly on zone's unevictable
>> 539 * list.
>> 540 */
>> 541 lru = LRU_UNEVICTABLE;
>> 542 add_page_to_unevictable_list(page);
>> 543 }
>> 544
>> 545 /*
>> 546 * page's status can change while we move it among lru. If an evictable
>> 547 * page is on unevictable list, it never be freed. To avoid that,
>> 548 * check after we added it to the list, again.
>> 549 */
>> 550 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
>> 551 if (!isolate_lru_page(page)) {
>> 552 put_page(page);
>> 553 goto redo;
>> 554 }
>> 555 /* This means someone else dropped this page from LRU
>> 556 * So, it will be freed or putback to LRU again. There is
>> 557 * nothing to do here.
>> 558 */
>> 559 }
>> 560
>> 561 if (was_unevictable && lru != LRU_UNEVICTABLE)
>> 562 count_vm_event(UNEVICTABLE_PGRESCUED);
>> 563 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
>> 564 count_vm_event(UNEVICTABLE_PGCULLED);
>> 565
>> 566 put_page(page); /* drop ref from isolate */
>> 567 }
>> 568
454 /* 569 /*
455 * shrink_page_list() returns the number of re 570 * shrink_page_list() returns the number of reclaimed pages
456 */ 571 */
457 static unsigned long shrink_page_list(struct l 572 static unsigned long shrink_page_list(struct list_head *page_list,
458 struct 573 struct scan_control *sc,
459 enum p 574 enum pageout_io sync_writeback)
460 { 575 {
461 LIST_HEAD(ret_pages); 576 LIST_HEAD(ret_pages);
462 struct pagevec freed_pvec; 577 struct pagevec freed_pvec;
463 int pgactivate = 0; 578 int pgactivate = 0;
464 unsigned long nr_reclaimed = 0; 579 unsigned long nr_reclaimed = 0;
>> 580 unsigned long vm_flags;
465 581
466 cond_resched(); 582 cond_resched();
467 583
468 pagevec_init(&freed_pvec, 1); 584 pagevec_init(&freed_pvec, 1);
469 while (!list_empty(page_list)) { 585 while (!list_empty(page_list)) {
470 struct address_space *mapping; 586 struct address_space *mapping;
471 struct page *page; 587 struct page *page;
472 int may_enter_fs; 588 int may_enter_fs;
473 int referenced; 589 int referenced;
474 590
475 cond_resched(); 591 cond_resched();
476 592
477 page = lru_to_page(page_list); 593 page = lru_to_page(page_list);
478 list_del(&page->lru); 594 list_del(&page->lru);
479 595
480 if (TestSetPageLocked(page)) !! 596 if (!trylock_page(page))
481 goto keep; 597 goto keep;
482 598
483 VM_BUG_ON(PageActive(page)); 599 VM_BUG_ON(PageActive(page));
484 600
485 sc->nr_scanned++; 601 sc->nr_scanned++;
486 602
487 if (!sc->may_swap && page_mapp !! 603 if (unlikely(!page_evictable(page, NULL)))
>> 604 goto cull_mlocked;
>> 605
>> 606 if (!sc->may_unmap && page_mapped(page))
488 goto keep_locked; 607 goto keep_locked;
489 608
490 /* Double the slab pressure fo 609 /* Double the slab pressure for mapped and swapcache pages */
491 if (page_mapped(page) || PageS 610 if (page_mapped(page) || PageSwapCache(page))
492 sc->nr_scanned++; 611 sc->nr_scanned++;
493 612
494 may_enter_fs = (sc->gfp_mask & 613 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
495 (PageSwapCache(page) & 614 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
496 615
497 if (PageWriteback(page)) { 616 if (PageWriteback(page)) {
498 /* 617 /*
499 * Synchronous reclaim 618 * Synchronous reclaim is performed in two passes,
500 * first an asynchrono 619 * first an asynchronous pass over the list to
501 * start parallel writ 620 * start parallel writeback, and a second synchronous
502 * pass to wait for th 621 * pass to wait for the IO to complete. Wait here
503 * for any page for wh 622 * for any page for which writeback has already
504 * started. 623 * started.
505 */ 624 */
506 if (sync_writeback == 625 if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
507 wait_on_page_w 626 wait_on_page_writeback(page);
508 else 627 else
509 goto keep_lock 628 goto keep_locked;
510 } 629 }
511 630
512 referenced = page_referenced(p !! 631 referenced = page_referenced(page, 1,
513 /* In active use or really unf !! 632 sc->mem_cgroup, &vm_flags);
>> 633 /*
>> 634 * In active use or really unfreeable? Activate it.
>> 635 * If page which have PG_mlocked lost isoltation race,
>> 636 * try_to_unmap moves it to unevictable list
>> 637 */
514 if (sc->order <= PAGE_ALLOC_CO 638 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
515 refere !! 639 referenced && page_mapping_inuse(page)
>> 640 && !(vm_flags & VM_LOCKED))
516 goto activate_locked; 641 goto activate_locked;
517 642
518 #ifdef CONFIG_SWAP <<
519 /* 643 /*
520 * Anonymous process memory ha 644 * Anonymous process memory has backing store?
521 * Try to allocate it some swa 645 * Try to allocate it some swap space here.
522 */ 646 */
523 if (PageAnon(page) && !PageSwa !! 647 if (PageAnon(page) && !PageSwapCache(page)) {
524 if (!add_to_swap(page, !! 648 if (!(sc->gfp_mask & __GFP_IO))
>> 649 goto keep_locked;
>> 650 if (!add_to_swap(page))
525 goto activate_ 651 goto activate_locked;
526 #endif /* CONFIG_SWAP */ !! 652 may_enter_fs = 1;
>> 653 }
527 654
528 mapping = page_mapping(page); 655 mapping = page_mapping(page);
529 656
530 /* 657 /*
531 * The page is mapped into the 658 * The page is mapped into the page tables of one or more
532 * processes. Try to unmap it 659 * processes. Try to unmap it here.
533 */ 660 */
534 if (page_mapped(page) && mappi 661 if (page_mapped(page) && mapping) {
535 switch (try_to_unmap(p 662 switch (try_to_unmap(page, 0)) {
536 case SWAP_FAIL: 663 case SWAP_FAIL:
537 goto activate_ 664 goto activate_locked;
538 case SWAP_AGAIN: 665 case SWAP_AGAIN:
539 goto keep_lock 666 goto keep_locked;
>> 667 case SWAP_MLOCK:
>> 668 goto cull_mlocked;
540 case SWAP_SUCCESS: 669 case SWAP_SUCCESS:
541 ; /* try to fr 670 ; /* try to free the page below */
542 } 671 }
543 } 672 }
544 673
545 if (PageDirty(page)) { 674 if (PageDirty(page)) {
546 if (sc->order <= PAGE_ 675 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
547 goto keep_lock 676 goto keep_locked;
548 if (!may_enter_fs) 677 if (!may_enter_fs)
549 goto keep_lock 678 goto keep_locked;
550 if (!sc->may_writepage 679 if (!sc->may_writepage)
551 goto keep_lock 680 goto keep_locked;
552 681
553 /* Page is dirty, try 682 /* Page is dirty, try to write it out here */
554 switch (pageout(page, 683 switch (pageout(page, mapping, sync_writeback)) {
555 case PAGE_KEEP: 684 case PAGE_KEEP:
556 goto keep_lock 685 goto keep_locked;
557 case PAGE_ACTIVATE: 686 case PAGE_ACTIVATE:
558 goto activate_ 687 goto activate_locked;
559 case PAGE_SUCCESS: 688 case PAGE_SUCCESS:
560 if (PageWriteb 689 if (PageWriteback(page) || PageDirty(page))
561 goto k 690 goto keep;
562 /* 691 /*
563 * A synchrono 692 * A synchronous write - probably a ramdisk. Go
564 * ahead and t 693 * ahead and try to reclaim the page.
565 */ 694 */
566 if (TestSetPag !! 695 if (!trylock_page(page))
567 goto k 696 goto keep;
568 if (PageDirty( 697 if (PageDirty(page) || PageWriteback(page))
569 goto k 698 goto keep_locked;
570 mapping = page 699 mapping = page_mapping(page);
571 case PAGE_CLEAN: 700 case PAGE_CLEAN:
572 ; /* try to fr 701 ; /* try to free the page below */
573 } 702 }
574 } 703 }
575 704
576 /* 705 /*
577 * If the page has buffers, tr 706 * If the page has buffers, try to free the buffer mappings
578 * associated with this page. 707 * associated with this page. If we succeed we try to free
579 * the page as well. 708 * the page as well.
580 * 709 *
581 * We do this even if the page 710 * We do this even if the page is PageDirty().
582 * try_to_release_page() does 711 * try_to_release_page() does not perform I/O, but it is
583 * possible for a page to have 712 * possible for a page to have PageDirty set, but it is actually
584 * clean (all its buffers are 713 * clean (all its buffers are clean). This happens if the
585 * buffers were written out di 714 * buffers were written out directly, with submit_bh(). ext3
586 * will do this, as well as th !! 715 * will do this, as well as the blockdev mapping.
587 * try_to_release_page() will 716 * try_to_release_page() will discover that cleanness and will
588 * drop the buffers and mark t 717 * drop the buffers and mark the page clean - it can be freed.
589 * 718 *
590 * Rarely, pages can have buff 719 * Rarely, pages can have buffers and no ->mapping. These are
591 * the pages which were not su 720 * the pages which were not successfully invalidated in
592 * truncate_complete_page(). 721 * truncate_complete_page(). We try to drop those buffers here
593 * and if that worked, and the 722 * and if that worked, and the page is no longer mapped into
594 * process address space (page 723 * process address space (page_count == 1) it can be freed.
595 * Otherwise, leave the page o 724 * Otherwise, leave the page on the LRU so it is swappable.
596 */ 725 */
597 if (PagePrivate(page)) { !! 726 if (page_has_private(page)) {
598 if (!try_to_release_pa 727 if (!try_to_release_page(page, sc->gfp_mask))
599 goto activate_ 728 goto activate_locked;
600 if (!mapping && page_c !! 729 if (!mapping && page_count(page) == 1) {
601 goto free_it; !! 730 unlock_page(page);
>> 731 if (put_page_testzero(page))
>> 732 goto free_it;
>> 733 else {
>> 734 /*
>> 735 * rare race with speculative reference.
>> 736 * the speculative reference will free
>> 737 * this page shortly, so we may
>> 738 * increment nr_reclaimed here (and
>> 739 * leave it off the LRU).
>> 740 */
>> 741 nr_reclaimed++;
>> 742 continue;
>> 743 }
>> 744 }
602 } 745 }
603 746
604 if (!mapping || !remove_mappin !! 747 if (!mapping || !__remove_mapping(mapping, page))
605 goto keep_locked; 748 goto keep_locked;
606 749
>> 750 /*
>> 751 * At this point, we have no other references and there is
>> 752 * no way to pick any more up (removed from LRU, removed
>> 753 * from pagecache). Can use non-atomic bitops now (and
>> 754 * we obviously don't have to worry about waking up a process
>> 755 * waiting on the page lock, because there are no references.
>> 756 */
>> 757 __clear_page_locked(page);
607 free_it: 758 free_it:
608 unlock_page(page); <<
609 nr_reclaimed++; 759 nr_reclaimed++;
610 if (!pagevec_add(&freed_pvec, !! 760 if (!pagevec_add(&freed_pvec, page)) {
611 __pagevec_release_nonl !! 761 __pagevec_free(&freed_pvec);
>> 762 pagevec_reinit(&freed_pvec);
>> 763 }
>> 764 continue;
>> 765
>> 766 cull_mlocked:
>> 767 if (PageSwapCache(page))
>> 768 try_to_free_swap(page);
>> 769 unlock_page(page);
>> 770 putback_lru_page(page);
612 continue; 771 continue;
613 772
614 activate_locked: 773 activate_locked:
>> 774 /* Not a candidate for swapping, so reclaim swap space. */
>> 775 if (PageSwapCache(page) && vm_swap_full())
>> 776 try_to_free_swap(page);
>> 777 VM_BUG_ON(PageActive(page));
615 SetPageActive(page); 778 SetPageActive(page);
616 pgactivate++; 779 pgactivate++;
617 keep_locked: 780 keep_locked:
618 unlock_page(page); 781 unlock_page(page);
619 keep: 782 keep:
620 list_add(&page->lru, &ret_page 783 list_add(&page->lru, &ret_pages);
621 VM_BUG_ON(PageLRU(page)); !! 784 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
622 } 785 }
623 list_splice(&ret_pages, page_list); 786 list_splice(&ret_pages, page_list);
624 if (pagevec_count(&freed_pvec)) 787 if (pagevec_count(&freed_pvec))
625 __pagevec_release_nonlru(&free !! 788 __pagevec_free(&freed_pvec);
626 count_vm_events(PGACTIVATE, pgactivate 789 count_vm_events(PGACTIVATE, pgactivate);
627 return nr_reclaimed; 790 return nr_reclaimed;
628 } 791 }
629 792
630 /* LRU Isolation modes. */ 793 /* LRU Isolation modes. */
631 #define ISOLATE_INACTIVE 0 /* Isolate ina 794 #define ISOLATE_INACTIVE 0 /* Isolate inactive pages. */
632 #define ISOLATE_ACTIVE 1 /* Isolate act 795 #define ISOLATE_ACTIVE 1 /* Isolate active pages. */
633 #define ISOLATE_BOTH 2 /* Isolate bot 796 #define ISOLATE_BOTH 2 /* Isolate both active and inactive pages. */
634 797
635 /* 798 /*
636 * Attempt to remove the specified page from i 799 * Attempt to remove the specified page from its LRU. Only take this page
637 * if it is of the appropriate PageActive stat 800 * if it is of the appropriate PageActive status. Pages which are being
638 * freed elsewhere are also ignored. 801 * freed elsewhere are also ignored.
639 * 802 *
640 * page: page to consider 803 * page: page to consider
641 * mode: one of the LRU isolation modes 804 * mode: one of the LRU isolation modes defined above
642 * 805 *
643 * returns 0 on success, -ve errno on failure. 806 * returns 0 on success, -ve errno on failure.
644 */ 807 */
645 int __isolate_lru_page(struct page *page, int !! 808 int __isolate_lru_page(struct page *page, int mode, int file)
646 { 809 {
647 int ret = -EINVAL; 810 int ret = -EINVAL;
648 811
649 /* Only take pages on the LRU. */ 812 /* Only take pages on the LRU. */
650 if (!PageLRU(page)) 813 if (!PageLRU(page))
651 return ret; 814 return ret;
652 815
653 /* 816 /*
654 * When checking the active state, we 817 * When checking the active state, we need to be sure we are
655 * dealing with comparible boolean val 818 * dealing with comparible boolean values. Take the logical not
656 * of each. 819 * of each.
657 */ 820 */
658 if (mode != ISOLATE_BOTH && (!PageActi 821 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
659 return ret; 822 return ret;
660 823
>> 824 if (mode != ISOLATE_BOTH && (!page_is_file_cache(page) != !file))
>> 825 return ret;
>> 826
>> 827 /*
>> 828 * When this function is being called for lumpy reclaim, we
>> 829 * initially look into all LRU pages, active, inactive and
>> 830 * unevictable; only give shrink_page_list evictable pages.
>> 831 */
>> 832 if (PageUnevictable(page))
>> 833 return ret;
>> 834
661 ret = -EBUSY; 835 ret = -EBUSY;
>> 836
662 if (likely(get_page_unless_zero(page)) 837 if (likely(get_page_unless_zero(page))) {
663 /* 838 /*
664 * Be careful not to clear Pag 839 * Be careful not to clear PageLRU until after we're
665 * sure the page is not being 840 * sure the page is not being freed elsewhere -- the
666 * page release code relies on 841 * page release code relies on it.
667 */ 842 */
668 ClearPageLRU(page); 843 ClearPageLRU(page);
669 ret = 0; 844 ret = 0;
670 } 845 }
671 846
672 return ret; 847 return ret;
673 } 848 }
674 849
675 /* 850 /*
676 * zone->lru_lock is heavily contended. Some 851 * zone->lru_lock is heavily contended. Some of the functions that
677 * shrink the lists perform better by taking o 852 * shrink the lists perform better by taking out a batch of pages
678 * and working on them outside the LRU lock. 853 * and working on them outside the LRU lock.
679 * 854 *
680 * For pagecache intensive workloads, this fun 855 * For pagecache intensive workloads, this function is the hottest
681 * spot in the kernel (apart from copy_*_user 856 * spot in the kernel (apart from copy_*_user functions).
682 * 857 *
683 * Appropriate locks must be held before calli 858 * Appropriate locks must be held before calling this function.
684 * 859 *
685 * @nr_to_scan: The number of pages to look th 860 * @nr_to_scan: The number of pages to look through on the list.
686 * @src: The LRU list to pull pages off 861 * @src: The LRU list to pull pages off.
687 * @dst: The temp list to put pages on 862 * @dst: The temp list to put pages on to.
688 * @scanned: The number of pages that were 863 * @scanned: The number of pages that were scanned.
689 * @order: The caller's attempted allocat 864 * @order: The caller's attempted allocation order
690 * @mode: One of the LRU isolation modes 865 * @mode: One of the LRU isolation modes
>> 866 * @file: True [1] if isolating file [!anon] pages
691 * 867 *
692 * returns how many pages were moved onto *@ds 868 * returns how many pages were moved onto *@dst.
693 */ 869 */
694 static unsigned long isolate_lru_pages(unsigne 870 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
695 struct list_head *src, struct 871 struct list_head *src, struct list_head *dst,
696 unsigned long *scanned, int or !! 872 unsigned long *scanned, int order, int mode, int file)
697 { 873 {
698 unsigned long nr_taken = 0; 874 unsigned long nr_taken = 0;
699 unsigned long scan; 875 unsigned long scan;
700 876
701 for (scan = 0; scan < nr_to_scan && !l 877 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
702 struct page *page; 878 struct page *page;
703 unsigned long pfn; 879 unsigned long pfn;
704 unsigned long end_pfn; 880 unsigned long end_pfn;
705 unsigned long page_pfn; 881 unsigned long page_pfn;
706 int zone_id; 882 int zone_id;
707 883
708 page = lru_to_page(src); 884 page = lru_to_page(src);
709 prefetchw_prev_lru_page(page, 885 prefetchw_prev_lru_page(page, src, flags);
710 886
711 VM_BUG_ON(!PageLRU(page)); 887 VM_BUG_ON(!PageLRU(page));
712 888
713 switch (__isolate_lru_page(pag !! 889 switch (__isolate_lru_page(page, mode, file)) {
714 case 0: 890 case 0:
715 list_move(&page->lru, 891 list_move(&page->lru, dst);
>> 892 mem_cgroup_del_lru(page);
716 nr_taken++; 893 nr_taken++;
717 break; 894 break;
718 895
719 case -EBUSY: 896 case -EBUSY:
720 /* else it is being fr 897 /* else it is being freed elsewhere */
721 list_move(&page->lru, 898 list_move(&page->lru, src);
>> 899 mem_cgroup_rotate_lru_list(page, page_lru(page));
722 continue; 900 continue;
723 901
724 default: 902 default:
725 BUG(); 903 BUG();
726 } 904 }
727 905
728 if (!order) 906 if (!order)
729 continue; 907 continue;
730 908
731 /* 909 /*
732 * Attempt to take all pages i 910 * Attempt to take all pages in the order aligned region
733 * surrounding the tag page. 911 * surrounding the tag page. Only take those pages of
734 * the same active state as th 912 * the same active state as that tag page. We may safely
735 * round the target page pfn d 913 * round the target page pfn down to the requested order
736 * as the mem_map is guarentee 914 * as the mem_map is guarenteed valid out to MAX_ORDER,
737 * where that page is in a dif 915 * where that page is in a different zone we will detect
738 * it from its zone id and abo 916 * it from its zone id and abort this block scan.
739 */ 917 */
740 zone_id = page_zone_id(page); 918 zone_id = page_zone_id(page);
741 page_pfn = page_to_pfn(page); 919 page_pfn = page_to_pfn(page);
742 pfn = page_pfn & ~((1 << order 920 pfn = page_pfn & ~((1 << order) - 1);
743 end_pfn = pfn + (1 << order); 921 end_pfn = pfn + (1 << order);
744 for (; pfn < end_pfn; pfn++) { 922 for (; pfn < end_pfn; pfn++) {
745 struct page *cursor_pa 923 struct page *cursor_page;
746 924
747 /* The target page is 925 /* The target page is in the block, ignore it. */
748 if (unlikely(pfn == pa 926 if (unlikely(pfn == page_pfn))
749 continue; 927 continue;
750 928
751 /* Avoid holes within 929 /* Avoid holes within the zone. */
752 if (unlikely(!pfn_vali 930 if (unlikely(!pfn_valid_within(pfn)))
753 break; 931 break;
754 932
755 cursor_page = pfn_to_p 933 cursor_page = pfn_to_page(pfn);
>> 934
756 /* Check that we have 935 /* Check that we have not crossed a zone boundary. */
757 if (unlikely(page_zone 936 if (unlikely(page_zone_id(cursor_page) != zone_id))
758 continue; 937 continue;
759 switch (__isolate_lru_ !! 938 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
760 case 0: <<
761 list_move(&cur 939 list_move(&cursor_page->lru, dst);
>> 940 mem_cgroup_del_lru(cursor_page);
762 nr_taken++; 941 nr_taken++;
763 scan++; 942 scan++;
764 break; <<
765 <<
766 case -EBUSY: <<
767 /* else it is <<
768 list_move(&cur <<
769 default: <<
770 break; <<
771 } 943 }
772 } 944 }
773 } 945 }
774 946
775 *scanned = scan; 947 *scanned = scan;
776 return nr_taken; 948 return nr_taken;
777 } 949 }
778 950
779 static unsigned long isolate_pages_global(unsi 951 static unsigned long isolate_pages_global(unsigned long nr,
780 struct 952 struct list_head *dst,
781 unsign 953 unsigned long *scanned, int order,
782 int mo 954 int mode, struct zone *z,
783 struct 955 struct mem_cgroup *mem_cont,
784 int ac !! 956 int active, int file)
785 { 957 {
>> 958 int lru = LRU_BASE;
786 if (active) 959 if (active)
787 return isolate_lru_pages(nr, & !! 960 lru += LRU_ACTIVE;
788 !! 961 if (file)
789 else !! 962 lru += LRU_FILE;
790 return isolate_lru_pages(nr, & !! 963 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
791 !! 964 mode, !!file);
792 } 965 }
793 966
794 /* 967 /*
795 * clear_active_flags() is a helper for shrink 968 * clear_active_flags() is a helper for shrink_active_list(), clearing
796 * any active bits from the pages in the list. 969 * any active bits from the pages in the list.
797 */ 970 */
798 static unsigned long clear_active_flags(struct !! 971 static unsigned long clear_active_flags(struct list_head *page_list,
>> 972 unsigned int *count)
799 { 973 {
800 int nr_active = 0; 974 int nr_active = 0;
>> 975 int lru;
801 struct page *page; 976 struct page *page;
802 977
803 list_for_each_entry(page, page_list, l !! 978 list_for_each_entry(page, page_list, lru) {
>> 979 lru = page_is_file_cache(page);
804 if (PageActive(page)) { 980 if (PageActive(page)) {
>> 981 lru += LRU_ACTIVE;
805 ClearPageActive(page); 982 ClearPageActive(page);
806 nr_active++; 983 nr_active++;
807 } 984 }
>> 985 count[lru]++;
>> 986 }
808 987
809 return nr_active; 988 return nr_active;
810 } 989 }
811 990
>> 991 /**
>> 992 * isolate_lru_page - tries to isolate a page from its LRU list
>> 993 * @page: page to isolate from its LRU list
>> 994 *
>> 995 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
>> 996 * vmstat statistic corresponding to whatever LRU list the page was on.
>> 997 *
>> 998 * Returns 0 if the page was removed from an LRU list.
>> 999 * Returns -EBUSY if the page was not on an LRU list.
>> 1000 *
>> 1001 * The returned page will have PageLRU() cleared. If it was found on
>> 1002 * the active list, it will have PageActive set. If it was found on
>> 1003 * the unevictable list, it will have the PageUnevictable bit set. That flag
>> 1004 * may need to be cleared by the caller before letting the page go.
>> 1005 *
>> 1006 * The vmstat statistic corresponding to the list on which the page was
>> 1007 * found will be decremented.
>> 1008 *
>> 1009 * Restrictions:
>> 1010 * (1) Must be called with an elevated refcount on the page. This is a
>> 1011 * fundamentnal difference from isolate_lru_pages (which is called
>> 1012 * without a stable reference).
>> 1013 * (2) the lru_lock must not be held.
>> 1014 * (3) interrupts must be enabled.
>> 1015 */
>> 1016 int isolate_lru_page(struct page *page)
>> 1017 {
>> 1018 int ret = -EBUSY;
>> 1019
>> 1020 if (PageLRU(page)) {
>> 1021 struct zone *zone = page_zone(page);
>> 1022
>> 1023 spin_lock_irq(&zone->lru_lock);
>> 1024 if (PageLRU(page) && get_page_unless_zero(page)) {
>> 1025 int lru = page_lru(page);
>> 1026 ret = 0;
>> 1027 ClearPageLRU(page);
>> 1028
>> 1029 del_page_from_lru_list(zone, page, lru);
>> 1030 }
>> 1031 spin_unlock_irq(&zone->lru_lock);
>> 1032 }
>> 1033 return ret;
>> 1034 }
>> 1035
812 /* 1036 /*
813 * shrink_inactive_list() is a helper for shri 1037 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
814 * of reclaimed pages 1038 * of reclaimed pages
815 */ 1039 */
816 static unsigned long shrink_inactive_list(unsi 1040 static unsigned long shrink_inactive_list(unsigned long max_scan,
817 struct zone *z !! 1041 struct zone *zone, struct scan_control *sc,
>> 1042 int priority, int file)
818 { 1043 {
819 LIST_HEAD(page_list); 1044 LIST_HEAD(page_list);
820 struct pagevec pvec; 1045 struct pagevec pvec;
821 unsigned long nr_scanned = 0; 1046 unsigned long nr_scanned = 0;
822 unsigned long nr_reclaimed = 0; 1047 unsigned long nr_reclaimed = 0;
>> 1048 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
>> 1049 int lumpy_reclaim = 0;
>> 1050
>> 1051 /*
>> 1052 * If we need a large contiguous chunk of memory, or have
>> 1053 * trouble getting a small set of contiguous pages, we
>> 1054 * will reclaim both active and inactive pages.
>> 1055 *
>> 1056 * We use the same threshold as pageout congestion_wait below.
>> 1057 */
>> 1058 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
>> 1059 lumpy_reclaim = 1;
>> 1060 else if (sc->order && priority < DEF_PRIORITY - 2)
>> 1061 lumpy_reclaim = 1;
823 1062
824 pagevec_init(&pvec, 1); 1063 pagevec_init(&pvec, 1);
825 1064
826 lru_add_drain(); 1065 lru_add_drain();
827 spin_lock_irq(&zone->lru_lock); 1066 spin_lock_irq(&zone->lru_lock);
828 do { 1067 do {
829 struct page *page; 1068 struct page *page;
830 unsigned long nr_taken; 1069 unsigned long nr_taken;
831 unsigned long nr_scan; 1070 unsigned long nr_scan;
832 unsigned long nr_freed; 1071 unsigned long nr_freed;
833 unsigned long nr_active; 1072 unsigned long nr_active;
>> 1073 unsigned int count[NR_LRU_LISTS] = { 0, };
>> 1074 int mode = lumpy_reclaim ? ISOLATE_BOTH : ISOLATE_INACTIVE;
834 1075
835 nr_taken = sc->isolate_pages(s 1076 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
836 &page_list, &nr_s !! 1077 &page_list, &nr_scan, sc->order, mode,
837 (sc->order > PAGE !! 1078 zone, sc->mem_cgroup, 0, file);
838 I !! 1079 nr_active = clear_active_flags(&page_list, count);
839 zone, sc->mem_ <<
840 nr_active = clear_active_flags <<
841 __count_vm_events(PGDEACTIVATE 1080 __count_vm_events(PGDEACTIVATE, nr_active);
842 1081
843 __mod_zone_page_state(zone, NR !! 1082 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
844 __mod_zone_page_state(zone, NR !! 1083 -count[LRU_ACTIVE_FILE]);
845 !! 1084 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
846 if (scan_global_lru(sc)) !! 1085 -count[LRU_INACTIVE_FILE]);
>> 1086 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
>> 1087 -count[LRU_ACTIVE_ANON]);
>> 1088 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
>> 1089 -count[LRU_INACTIVE_ANON]);
>> 1090
>> 1091 if (scanning_global_lru(sc))
847 zone->pages_scanned += 1092 zone->pages_scanned += nr_scan;
>> 1093
>> 1094 reclaim_stat->recent_scanned[0] += count[LRU_INACTIVE_ANON];
>> 1095 reclaim_stat->recent_scanned[0] += count[LRU_ACTIVE_ANON];
>> 1096 reclaim_stat->recent_scanned[1] += count[LRU_INACTIVE_FILE];
>> 1097 reclaim_stat->recent_scanned[1] += count[LRU_ACTIVE_FILE];
>> 1098
848 spin_unlock_irq(&zone->lru_loc 1099 spin_unlock_irq(&zone->lru_lock);
849 1100
850 nr_scanned += nr_scan; 1101 nr_scanned += nr_scan;
851 nr_freed = shrink_page_list(&p 1102 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
852 1103
853 /* 1104 /*
854 * If we are direct reclaiming 1105 * If we are direct reclaiming for contiguous pages and we do
855 * not reclaim everything in t 1106 * not reclaim everything in the list, try again and wait
856 * for IO to complete. This wi 1107 * for IO to complete. This will stall high-order allocations
857 * but that should be acceptab 1108 * but that should be acceptable to the caller
858 */ 1109 */
859 if (nr_freed < nr_taken && !cu 1110 if (nr_freed < nr_taken && !current_is_kswapd() &&
860 sc->or !! 1111 lumpy_reclaim) {
861 congestion_wait(WRITE, !! 1112 congestion_wait(BLK_RW_ASYNC, HZ/10);
862 1113
863 /* 1114 /*
864 * The attempt at page 1115 * The attempt at page out may have made some
865 * of the pages active 1116 * of the pages active, mark them inactive again.
866 */ 1117 */
867 nr_active = clear_acti !! 1118 nr_active = clear_active_flags(&page_list, count);
868 count_vm_events(PGDEAC 1119 count_vm_events(PGDEACTIVATE, nr_active);
869 1120
870 nr_freed += shrink_pag 1121 nr_freed += shrink_page_list(&page_list, sc,
871 1122 PAGEOUT_IO_SYNC);
872 } 1123 }
873 1124
874 nr_reclaimed += nr_freed; 1125 nr_reclaimed += nr_freed;
875 local_irq_disable(); 1126 local_irq_disable();
876 if (current_is_kswapd()) { 1127 if (current_is_kswapd()) {
877 __count_zone_vm_events 1128 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
878 __count_vm_events(KSWA 1129 __count_vm_events(KSWAPD_STEAL, nr_freed);
879 } else if (scan_global_lru(sc) !! 1130 } else if (scanning_global_lru(sc))
880 __count_zone_vm_events 1131 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
881 1132
882 __count_zone_vm_events(PGSTEAL 1133 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
883 1134
884 if (nr_taken == 0) 1135 if (nr_taken == 0)
885 goto done; 1136 goto done;
886 1137
887 spin_lock(&zone->lru_lock); 1138 spin_lock(&zone->lru_lock);
888 /* 1139 /*
889 * Put back any unfreeable pag 1140 * Put back any unfreeable pages.
890 */ 1141 */
891 while (!list_empty(&page_list) 1142 while (!list_empty(&page_list)) {
>> 1143 int lru;
892 page = lru_to_page(&pa 1144 page = lru_to_page(&page_list);
893 VM_BUG_ON(PageLRU(page 1145 VM_BUG_ON(PageLRU(page));
894 SetPageLRU(page); <<
895 list_del(&page->lru); 1146 list_del(&page->lru);
896 if (PageActive(page)) !! 1147 if (unlikely(!page_evictable(page, NULL))) {
897 add_page_to_ac !! 1148 spin_unlock_irq(&zone->lru_lock);
898 else !! 1149 putback_lru_page(page);
899 add_page_to_in !! 1150 spin_lock_irq(&zone->lru_lock);
>> 1151 continue;
>> 1152 }
>> 1153 SetPageLRU(page);
>> 1154 lru = page_lru(page);
>> 1155 add_page_to_lru_list(zone, page, lru);
>> 1156 if (PageActive(page)) {
>> 1157 int file = !!page_is_file_cache(page);
>> 1158 reclaim_stat->recent_rotated[file]++;
>> 1159 }
900 if (!pagevec_add(&pvec 1160 if (!pagevec_add(&pvec, page)) {
901 spin_unlock_ir 1161 spin_unlock_irq(&zone->lru_lock);
902 __pagevec_rele 1162 __pagevec_release(&pvec);
903 spin_lock_irq( 1163 spin_lock_irq(&zone->lru_lock);
904 } 1164 }
905 } 1165 }
906 } while (nr_scanned < max_scan); 1166 } while (nr_scanned < max_scan);
907 spin_unlock(&zone->lru_lock); 1167 spin_unlock(&zone->lru_lock);
908 done: 1168 done:
909 local_irq_enable(); 1169 local_irq_enable();
910 pagevec_release(&pvec); 1170 pagevec_release(&pvec);
911 return nr_reclaimed; 1171 return nr_reclaimed;
912 } 1172 }
913 1173
914 /* 1174 /*
915 * We are about to scan this zone at a certain 1175 * We are about to scan this zone at a certain priority level. If that priority
916 * level is smaller (ie: more urgent) than the 1176 * level is smaller (ie: more urgent) than the previous priority, then note
917 * that priority level within the zone. This 1177 * that priority level within the zone. This is done so that when the next
918 * process comes in to scan this zone, it will 1178 * process comes in to scan this zone, it will immediately start out at this
919 * priority level rather than having to build 1179 * priority level rather than having to build up its own scanning priority.
920 * Here, this priority affects only the reclai 1180 * Here, this priority affects only the reclaim-mapped threshold.
921 */ 1181 */
922 static inline void note_zone_scanning_priority 1182 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
923 { 1183 {
924 if (priority < zone->prev_priority) 1184 if (priority < zone->prev_priority)
925 zone->prev_priority = priority 1185 zone->prev_priority = priority;
926 } 1186 }
927 1187
928 static inline int zone_is_near_oom(struct zone <<
929 { <<
930 return zone->pages_scanned >= (zone_pa <<
931 + zone_page_st <<
932 } <<
933 <<
934 /* <<
935 * Determine we should try to reclaim mapped p <<
936 * This is called only when sc->mem_cgroup is <<
937 */ <<
938 static int calc_reclaim_mapped(struct scan_con <<
939 int priority) <<
940 { <<
941 long mapped_ratio; <<
942 long distress; <<
943 long swap_tendency; <<
944 long imbalance; <<
945 int reclaim_mapped = 0; <<
946 int prev_priority; <<
947 <<
948 if (scan_global_lru(sc) && zone_is_nea <<
949 return 1; <<
950 /* <<
951 * `distress' is a measure of how much <<
952 * reclaiming pages. 0 -> no problems <<
953 */ <<
954 if (scan_global_lru(sc)) <<
955 prev_priority = zone->prev_pri <<
956 else <<
957 prev_priority = mem_cgroup_get <<
958 <<
959 distress = 100 >> min(prev_priority, p <<
960 <<
961 /* <<
962 * The point of this algorithm is to d <<
963 * reclaiming mapped memory instead of <<
964 * how much memory <<
965 * is mapped. <<
966 */ <<
967 if (scan_global_lru(sc)) <<
968 mapped_ratio = ((global_page_s <<
969 global_page_st <<
970 vm_tot <<
971 else <<
972 mapped_ratio = mem_cgroup_calc <<
973 <<
974 /* <<
975 * Now decide how much we really want <<
976 * mapped ratio is downgraded - just b <<
977 * mapped memory doesn't necessarily m <<
978 * isn't succeeding. <<
979 * <<
980 * The distress ratio is important - w <<
981 * going oom. <<
982 * <<
983 * A 100% value of vm_swappiness overr <<
984 * altogether. <<
985 */ <<
986 swap_tendency = mapped_ratio / 2 + dis <<
987 <<
988 /* <<
989 * If there's huge imbalance between a <<
990 * (think active 100 times larger than <<
991 * become more permissive, or the syst <<
992 * cpu before it start swapping during <<
993 * Distress is about avoiding early-oo <<
994 * making swappiness graceful despite <<
995 * values. <<
996 * <<
997 * Avoid div by zero with nr_inactive+ <<
998 * value is vm_total_pages. <<
999 */ <<
1000 if (scan_global_lru(sc)) { <<
1001 imbalance = zone_page_state( <<
1002 imbalance /= zone_page_state( <<
1003 } else <<
1004 imbalance = mem_cgroup_reclai <<
1005 <<
1006 /* <<
1007 * Reduce the effect of imbalance if <<
1008 * this means for a swappiness very l <<
1009 * must be much higher than 100 for t <<
1010 * the difference. <<
1011 * <<
1012 * Max temporary value is vm_total_pa <<
1013 */ <<
1014 imbalance *= (vm_swappiness + 1); <<
1015 imbalance /= 100; <<
1016 <<
1017 /* <<
1018 * If not much of the ram is mapped, <<
1019 * less relevant, it's high priority <<
1020 * list with mapped pages only in pre <<
1021 * mapped pages. <<
1022 * <<
1023 * Max temporary value is vm_total_pa <<
1024 */ <<
1025 imbalance *= mapped_ratio; <<
1026 imbalance /= 100; <<
1027 <<
1028 /* apply imbalance feedback to swap_t <<
1029 swap_tendency += imbalance; <<
1030 <<
1031 /* <<
1032 * Now use this metric to decide whet <<
1033 * memory onto the inactive list. <<
1034 */ <<
1035 if (swap_tendency >= 100) <<
1036 reclaim_mapped = 1; <<
1037 <<
1038 return reclaim_mapped; <<
1039 } <<
1040 <<
1041 /* 1188 /*
1042 * This moves pages from the active list to t 1189 * This moves pages from the active list to the inactive list.
1043 * 1190 *
1044 * We move them the other way if the page is 1191 * We move them the other way if the page is referenced by one or more
1045 * processes, from rmap. 1192 * processes, from rmap.
1046 * 1193 *
1047 * If the pages are mostly unmapped, the proc 1194 * If the pages are mostly unmapped, the processing is fast and it is
1048 * appropriate to hold zone->lru_lock across 1195 * appropriate to hold zone->lru_lock across the whole operation. But if
1049 * the pages are mapped, the processing is sl 1196 * the pages are mapped, the processing is slow (page_referenced()) so we
1050 * should drop zone->lru_lock around each pag 1197 * should drop zone->lru_lock around each page. It's impossible to balance
1051 * this, so instead we remove the pages from 1198 * this, so instead we remove the pages from the LRU while processing them.
1052 * It is safe to rely on PG_active against th 1199 * It is safe to rely on PG_active against the non-LRU pages in here because
1053 * nobody will play with that bit on a non-LR 1200 * nobody will play with that bit on a non-LRU page.
1054 * 1201 *
1055 * The downside is that we have to touch page 1202 * The downside is that we have to touch page->_count against each page.
1056 * But we had to alter page->flags anyway. 1203 * But we had to alter page->flags anyway.
1057 */ 1204 */
1058 1205
>> 1206 static void move_active_pages_to_lru(struct zone *zone,
>> 1207 struct list_head *list,
>> 1208 enum lru_list lru)
>> 1209 {
>> 1210 unsigned long pgmoved = 0;
>> 1211 struct pagevec pvec;
>> 1212 struct page *page;
>> 1213
>> 1214 pagevec_init(&pvec, 1);
>> 1215
>> 1216 while (!list_empty(list)) {
>> 1217 page = lru_to_page(list);
>> 1218 prefetchw_prev_lru_page(page, list, flags);
>> 1219
>> 1220 VM_BUG_ON(PageLRU(page));
>> 1221 SetPageLRU(page);
>> 1222
>> 1223 VM_BUG_ON(!PageActive(page));
>> 1224 if (!is_active_lru(lru))
>> 1225 ClearPageActive(page); /* we are de-activating */
>> 1226
>> 1227 list_move(&page->lru, &zone->lru[lru].list);
>> 1228 mem_cgroup_add_lru_list(page, lru);
>> 1229 pgmoved++;
>> 1230
>> 1231 if (!pagevec_add(&pvec, page) || list_empty(list)) {
>> 1232 spin_unlock_irq(&zone->lru_lock);
>> 1233 if (buffer_heads_over_limit)
>> 1234 pagevec_strip(&pvec);
>> 1235 __pagevec_release(&pvec);
>> 1236 spin_lock_irq(&zone->lru_lock);
>> 1237 }
>> 1238 }
>> 1239 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
>> 1240 if (!is_active_lru(lru))
>> 1241 __count_vm_events(PGDEACTIVATE, pgmoved);
>> 1242 }
1059 1243
1060 static void shrink_active_list(unsigned long 1244 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1061 struct scan_c !! 1245 struct scan_control *sc, int priority, int file)
1062 { 1246 {
1063 unsigned long pgmoved; 1247 unsigned long pgmoved;
1064 int pgdeactivate = 0; <<
1065 unsigned long pgscanned; 1248 unsigned long pgscanned;
>> 1249 unsigned long vm_flags;
1066 LIST_HEAD(l_hold); /* The pages 1250 LIST_HEAD(l_hold); /* The pages which were snipped off */
1067 LIST_HEAD(l_inactive); /* Pages to g !! 1251 LIST_HEAD(l_active);
1068 LIST_HEAD(l_active); /* Pages to g !! 1252 LIST_HEAD(l_inactive);
1069 struct page *page; 1253 struct page *page;
1070 struct pagevec pvec; !! 1254 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1071 int reclaim_mapped = 0; <<
1072 <<
1073 if (sc->may_swap) <<
1074 reclaim_mapped = calc_reclaim <<
1075 1255
1076 lru_add_drain(); 1256 lru_add_drain();
1077 spin_lock_irq(&zone->lru_lock); 1257 spin_lock_irq(&zone->lru_lock);
1078 pgmoved = sc->isolate_pages(nr_pages, 1258 pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1079 ISOLA 1259 ISOLATE_ACTIVE, zone,
1080 sc->m !! 1260 sc->mem_cgroup, 1, file);
1081 /* 1261 /*
1082 * zone->pages_scanned is used for de 1262 * zone->pages_scanned is used for detect zone's oom
1083 * mem_cgroup remembers nr_scan by it 1263 * mem_cgroup remembers nr_scan by itself.
1084 */ 1264 */
1085 if (scan_global_lru(sc)) !! 1265 if (scanning_global_lru(sc)) {
1086 zone->pages_scanned += pgscan 1266 zone->pages_scanned += pgscanned;
>> 1267 }
>> 1268 reclaim_stat->recent_scanned[!!file] += pgmoved;
1087 1269
1088 __mod_zone_page_state(zone, NR_ACTIVE !! 1270 __count_zone_vm_events(PGREFILL, zone, pgscanned);
>> 1271 if (file)
>> 1272 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -pgmoved);
>> 1273 else
>> 1274 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -pgmoved);
1089 spin_unlock_irq(&zone->lru_lock); 1275 spin_unlock_irq(&zone->lru_lock);
1090 1276
>> 1277 pgmoved = 0; /* count referenced (mapping) mapped pages */
1091 while (!list_empty(&l_hold)) { 1278 while (!list_empty(&l_hold)) {
1092 cond_resched(); 1279 cond_resched();
1093 page = lru_to_page(&l_hold); 1280 page = lru_to_page(&l_hold);
1094 list_del(&page->lru); 1281 list_del(&page->lru);
1095 if (page_mapped(page)) { !! 1282
1096 if (!reclaim_mapped | !! 1283 if (unlikely(!page_evictable(page, NULL))) {
1097 (total_swap_pages !! 1284 putback_lru_page(page);
1098 page_referenced(p !! 1285 continue;
>> 1286 }
>> 1287
>> 1288 /* page_referenced clears PageReferenced */
>> 1289 if (page_mapping_inuse(page) &&
>> 1290 page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
>> 1291 pgmoved++;
>> 1292 /*
>> 1293 * Identify referenced, file-backed active pages and
>> 1294 * give them one more trip around the active list. So
>> 1295 * that executable code get better chances to stay in
>> 1296 * memory under moderate memory pressure. Anon pages
>> 1297 * are not likely to be evicted by use-once streaming
>> 1298 * IO, plus JVM can create lots of anon VM_EXEC pages,
>> 1299 * so we ignore them here.
>> 1300 */
>> 1301 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1099 list_add(&pag 1302 list_add(&page->lru, &l_active);
1100 continue; 1303 continue;
1101 } 1304 }
1102 } 1305 }
>> 1306
1103 list_add(&page->lru, &l_inact 1307 list_add(&page->lru, &l_inactive);
1104 } 1308 }
1105 1309
1106 pagevec_init(&pvec, 1); !! 1310 /*
1107 pgmoved = 0; !! 1311 * Move pages back to the lru list.
>> 1312 */
1108 spin_lock_irq(&zone->lru_lock); 1313 spin_lock_irq(&zone->lru_lock);
1109 while (!list_empty(&l_inactive)) { !! 1314 /*
1110 page = lru_to_page(&l_inactiv !! 1315 * Count referenced pages from currently used mappings as rotated,
1111 prefetchw_prev_lru_page(page, !! 1316 * even though only some of them are actually re-activated. This
1112 VM_BUG_ON(PageLRU(page)); !! 1317 * helps balance scan pressure between file and anonymous pages in
1113 SetPageLRU(page); !! 1318 * get_scan_ratio.
1114 VM_BUG_ON(!PageActive(page)); !! 1319 */
1115 ClearPageActive(page); !! 1320 reclaim_stat->recent_rotated[!!file] += pgmoved;
>> 1321
>> 1322 move_active_pages_to_lru(zone, &l_active,
>> 1323 LRU_ACTIVE + file * LRU_FILE);
>> 1324 move_active_pages_to_lru(zone, &l_inactive,
>> 1325 LRU_BASE + file * LRU_FILE);
1116 1326
1117 list_move(&page->lru, &zone-> !! 1327 spin_unlock_irq(&zone->lru_lock);
1118 mem_cgroup_move_lists(page, f !! 1328 }
1119 pgmoved++; !! 1329
1120 if (!pagevec_add(&pvec, page) !! 1330 static int inactive_anon_is_low_global(struct zone *zone)
1121 __mod_zone_page_state !! 1331 {
1122 spin_unlock_irq(&zone !! 1332 unsigned long active, inactive;
1123 pgdeactivate += pgmov !! 1333
1124 pgmoved = 0; !! 1334 active = zone_page_state(zone, NR_ACTIVE_ANON);
1125 if (buffer_heads_over !! 1335 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1126 pagevec_strip !! 1336
1127 __pagevec_release(&pv !! 1337 if (inactive * zone->inactive_ratio < active)
1128 spin_lock_irq(&zone-> !! 1338 return 1;
>> 1339
>> 1340 return 0;
>> 1341 }
>> 1342
>> 1343 /**
>> 1344 * inactive_anon_is_low - check if anonymous pages need to be deactivated
>> 1345 * @zone: zone to check
>> 1346 * @sc: scan control of this context
>> 1347 *
>> 1348 * Returns true if the zone does not have enough inactive anon pages,
>> 1349 * meaning some active anon pages need to be deactivated.
>> 1350 */
>> 1351 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
>> 1352 {
>> 1353 int low;
>> 1354
>> 1355 if (scanning_global_lru(sc))
>> 1356 low = inactive_anon_is_low_global(zone);
>> 1357 else
>> 1358 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
>> 1359 return low;
>> 1360 }
>> 1361
>> 1362 static int inactive_file_is_low_global(struct zone *zone)
>> 1363 {
>> 1364 unsigned long active, inactive;
>> 1365
>> 1366 active = zone_page_state(zone, NR_ACTIVE_FILE);
>> 1367 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
>> 1368
>> 1369 return (active > inactive);
>> 1370 }
>> 1371
>> 1372 /**
>> 1373 * inactive_file_is_low - check if file pages need to be deactivated
>> 1374 * @zone: zone to check
>> 1375 * @sc: scan control of this context
>> 1376 *
>> 1377 * When the system is doing streaming IO, memory pressure here
>> 1378 * ensures that active file pages get deactivated, until more
>> 1379 * than half of the file pages are on the inactive list.
>> 1380 *
>> 1381 * Once we get to that situation, protect the system's working
>> 1382 * set from being evicted by disabling active file page aging.
>> 1383 *
>> 1384 * This uses a different ratio than the anonymous pages, because
>> 1385 * the page cache uses a use-once replacement algorithm.
>> 1386 */
>> 1387 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
>> 1388 {
>> 1389 int low;
>> 1390
>> 1391 if (scanning_global_lru(sc))
>> 1392 low = inactive_file_is_low_global(zone);
>> 1393 else
>> 1394 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
>> 1395 return low;
>> 1396 }
>> 1397
>> 1398 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
>> 1399 struct zone *zone, struct scan_control *sc, int priority)
>> 1400 {
>> 1401 int file = is_file_lru(lru);
>> 1402
>> 1403 if (lru == LRU_ACTIVE_FILE && inactive_file_is_low(zone, sc)) {
>> 1404 shrink_active_list(nr_to_scan, zone, sc, priority, file);
>> 1405 return 0;
>> 1406 }
>> 1407
>> 1408 if (lru == LRU_ACTIVE_ANON && inactive_anon_is_low(zone, sc)) {
>> 1409 shrink_active_list(nr_to_scan, zone, sc, priority, file);
>> 1410 return 0;
>> 1411 }
>> 1412 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
>> 1413 }
>> 1414
>> 1415 /*
>> 1416 * Determine how aggressively the anon and file LRU lists should be
>> 1417 * scanned. The relative value of each set of LRU lists is determined
>> 1418 * by looking at the fraction of the pages scanned we did rotate back
>> 1419 * onto the active list instead of evict.
>> 1420 *
>> 1421 * percent[0] specifies how much pressure to put on ram/swap backed
>> 1422 * memory, while percent[1] determines pressure on the file LRUs.
>> 1423 */
>> 1424 static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
>> 1425 unsigned long *percent)
>> 1426 {
>> 1427 unsigned long anon, file, free;
>> 1428 unsigned long anon_prio, file_prio;
>> 1429 unsigned long ap, fp;
>> 1430 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
>> 1431
>> 1432 anon = zone_nr_pages(zone, sc, LRU_ACTIVE_ANON) +
>> 1433 zone_nr_pages(zone, sc, LRU_INACTIVE_ANON);
>> 1434 file = zone_nr_pages(zone, sc, LRU_ACTIVE_FILE) +
>> 1435 zone_nr_pages(zone, sc, LRU_INACTIVE_FILE);
>> 1436
>> 1437 if (scanning_global_lru(sc)) {
>> 1438 free = zone_page_state(zone, NR_FREE_PAGES);
>> 1439 /* If we have very few page cache pages,
>> 1440 force-scan anon pages. */
>> 1441 if (unlikely(file + free <= high_wmark_pages(zone))) {
>> 1442 percent[0] = 100;
>> 1443 percent[1] = 0;
>> 1444 return;
1129 } 1445 }
1130 } 1446 }
1131 __mod_zone_page_state(zone, NR_INACTI !! 1447
1132 pgdeactivate += pgmoved; !! 1448 /*
1133 if (buffer_heads_over_limit) { !! 1449 * OK, so we have swap space and a fair amount of page cache
>> 1450 * pages. We use the recently rotated / recently scanned
>> 1451 * ratios to determine how valuable each cache is.
>> 1452 *
>> 1453 * Because workloads change over time (and to avoid overflow)
>> 1454 * we keep these statistics as a floating average, which ends
>> 1455 * up weighing recent references more than old ones.
>> 1456 *
>> 1457 * anon in [0], file in [1]
>> 1458 */
>> 1459 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
>> 1460 spin_lock_irq(&zone->lru_lock);
>> 1461 reclaim_stat->recent_scanned[0] /= 2;
>> 1462 reclaim_stat->recent_rotated[0] /= 2;
1134 spin_unlock_irq(&zone->lru_lo 1463 spin_unlock_irq(&zone->lru_lock);
1135 pagevec_strip(&pvec); !! 1464 }
>> 1465
>> 1466 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1136 spin_lock_irq(&zone->lru_lock 1467 spin_lock_irq(&zone->lru_lock);
>> 1468 reclaim_stat->recent_scanned[1] /= 2;
>> 1469 reclaim_stat->recent_rotated[1] /= 2;
>> 1470 spin_unlock_irq(&zone->lru_lock);
1137 } 1471 }
1138 1472
1139 pgmoved = 0; !! 1473 /*
1140 while (!list_empty(&l_active)) { !! 1474 * With swappiness at 100, anonymous and file have the same priority.
1141 page = lru_to_page(&l_active) !! 1475 * This scanning priority is essentially the inverse of IO cost.
1142 prefetchw_prev_lru_page(page, !! 1476 */
1143 VM_BUG_ON(PageLRU(page)); !! 1477 anon_prio = sc->swappiness;
1144 SetPageLRU(page); !! 1478 file_prio = 200 - sc->swappiness;
1145 VM_BUG_ON(!PageActive(page)); <<
1146 1479
1147 list_move(&page->lru, &zone-> !! 1480 /*
1148 mem_cgroup_move_lists(page, t !! 1481 * The amount of pressure on anon vs file pages is inversely
1149 pgmoved++; !! 1482 * proportional to the fraction of recently scanned pages on
1150 if (!pagevec_add(&pvec, page) !! 1483 * each list that were recently referenced and in active use.
1151 __mod_zone_page_state !! 1484 */
1152 pgmoved = 0; !! 1485 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1153 spin_unlock_irq(&zone !! 1486 ap /= reclaim_stat->recent_rotated[0] + 1;
1154 __pagevec_release(&pv <<
1155 spin_lock_irq(&zone-> <<
1156 } <<
1157 } <<
1158 __mod_zone_page_state(zone, NR_ACTIVE <<
1159 1487
1160 __count_zone_vm_events(PGREFILL, zone !! 1488 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1161 __count_vm_events(PGDEACTIVATE, pgdea !! 1489 fp /= reclaim_stat->recent_rotated[1] + 1;
1162 spin_unlock_irq(&zone->lru_lock); <<
1163 1490
1164 pagevec_release(&pvec); !! 1491 /* Normalize to percentages */
>> 1492 percent[0] = 100 * ap / (ap + fp + 1);
>> 1493 percent[1] = 100 - percent[0];
>> 1494 }
>> 1495
>> 1496 /*
>> 1497 * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
>> 1498 * until we collected @swap_cluster_max pages to scan.
>> 1499 */
>> 1500 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
>> 1501 unsigned long *nr_saved_scan,
>> 1502 unsigned long swap_cluster_max)
>> 1503 {
>> 1504 unsigned long nr;
>> 1505
>> 1506 *nr_saved_scan += nr_to_scan;
>> 1507 nr = *nr_saved_scan;
>> 1508
>> 1509 if (nr >= swap_cluster_max)
>> 1510 *nr_saved_scan = 0;
>> 1511 else
>> 1512 nr = 0;
>> 1513
>> 1514 return nr;
1165 } 1515 }
1166 1516
1167 /* 1517 /*
1168 * This is a basic per-zone page freer. Used 1518 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1169 */ 1519 */
1170 static unsigned long shrink_zone(int priority !! 1520 static void shrink_zone(int priority, struct zone *zone,
1171 struct scan_c 1521 struct scan_control *sc)
1172 { 1522 {
1173 unsigned long nr_active; !! 1523 unsigned long nr[NR_LRU_LISTS];
1174 unsigned long nr_inactive; <<
1175 unsigned long nr_to_scan; 1524 unsigned long nr_to_scan;
1176 unsigned long nr_reclaimed = 0; !! 1525 unsigned long percent[2]; /* anon @ 0; file @ 1 */
>> 1526 enum lru_list l;
>> 1527 unsigned long nr_reclaimed = sc->nr_reclaimed;
>> 1528 unsigned long swap_cluster_max = sc->swap_cluster_max;
>> 1529 int noswap = 0;
>> 1530
>> 1531 /* If we have no swap space, do not bother scanning anon pages. */
>> 1532 if (!sc->may_swap || (nr_swap_pages <= 0)) {
>> 1533 noswap = 1;
>> 1534 percent[0] = 0;
>> 1535 percent[1] = 100;
>> 1536 } else
>> 1537 get_scan_ratio(zone, sc, percent);
1177 1538
1178 if (scan_global_lru(sc)) { !! 1539 for_each_evictable_lru(l) {
1179 /* !! 1540 int file = is_file_lru(l);
1180 * Add one to nr_to_scan just !! 1541 unsigned long scan;
1181 * will slowly sift through t !! 1542
1182 */ !! 1543 scan = zone_nr_pages(zone, sc, l);
1183 zone->nr_scan_active += !! 1544 if (priority || noswap) {
1184 (zone_page_state(zone !! 1545 scan >>= priority;
1185 nr_active = zone->nr_scan_act !! 1546 scan = (scan * percent[file]) / 100;
1186 zone->nr_scan_inactive += !! 1547 }
1187 (zone_page_state(zone !! 1548 if (scanning_global_lru(sc))
1188 nr_inactive = zone->nr_scan_i !! 1549 nr[l] = nr_scan_try_batch(scan,
1189 if (nr_inactive >= sc->swap_c !! 1550 &zone->lru[l].nr_saved_scan,
1190 zone->nr_scan_inactiv !! 1551 swap_cluster_max);
1191 else 1552 else
1192 nr_inactive = 0; !! 1553 nr[l] = scan;
>> 1554 }
1193 1555
1194 if (nr_active >= sc->swap_clu !! 1556 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1195 zone->nr_scan_active !! 1557 nr[LRU_INACTIVE_FILE]) {
1196 else !! 1558 for_each_evictable_lru(l) {
1197 nr_active = 0; !! 1559 if (nr[l]) {
1198 } else { !! 1560 nr_to_scan = min(nr[l], swap_cluster_max);
>> 1561 nr[l] -= nr_to_scan;
>> 1562
>> 1563 nr_reclaimed += shrink_list(l, nr_to_scan,
>> 1564 zone, sc, priority);
>> 1565 }
>> 1566 }
1199 /* 1567 /*
1200 * This reclaim occurs not be !! 1568 * On large memory systems, scan >> priority can become
1201 * because memory controller !! 1569 * really large. This is fine for the starting priority;
1202 * Then, don't modify zone re !! 1570 * we want to put equal scanning pressure on each zone.
>> 1571 * However, if the VM has a harder time of freeing pages,
>> 1572 * with multiple processes reclaiming pages, the total
>> 1573 * freeing target can get unreasonably large.
1203 */ 1574 */
1204 nr_active = mem_cgroup_calc_r !! 1575 if (nr_reclaimed > swap_cluster_max &&
1205 zone, !! 1576 priority < DEF_PRIORITY && !current_is_kswapd())
1206 !! 1577 break;
1207 nr_inactive = mem_cgroup_calc <<
1208 zone, <<
1209 } 1578 }
1210 1579
>> 1580 sc->nr_reclaimed = nr_reclaimed;
1211 1581
1212 while (nr_active || nr_inactive) { !! 1582 /*
1213 if (nr_active) { !! 1583 * Even if we did not try to evict anon pages at all, we want to
1214 nr_to_scan = min(nr_a !! 1584 * rebalance the anon lru active/inactive ratio.
1215 (unsi !! 1585 */
1216 nr_active -= nr_to_sc !! 1586 if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0)
1217 shrink_active_list(nr !! 1587 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1218 } <<
1219 <<
1220 if (nr_inactive) { <<
1221 nr_to_scan = min(nr_i <<
1222 (unsi <<
1223 nr_inactive -= nr_to_ <<
1224 nr_reclaimed += shrin <<
1225 <<
1226 } <<
1227 } <<
1228 1588
1229 throttle_vm_writeout(sc->gfp_mask); 1589 throttle_vm_writeout(sc->gfp_mask);
1230 return nr_reclaimed; <<
1231 } 1590 }
1232 1591
1233 /* 1592 /*
1234 * This is the direct reclaim path, for page- 1593 * This is the direct reclaim path, for page-allocating processes. We only
1235 * try to reclaim pages from zones which will 1594 * try to reclaim pages from zones which will satisfy the caller's allocation
1236 * request. 1595 * request.
1237 * 1596 *
1238 * We reclaim from a zone even if that zone i !! 1597 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
>> 1598 * Because:
1239 * a) The caller may be trying to free *extra 1599 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1240 * allocation or 1600 * allocation or
1241 * b) The zones may be over pages_high but th !! 1601 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1242 * satisfy the `incremental min' zone defe !! 1602 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1243 * !! 1603 * zone defense algorithm.
1244 * Returns the number of reclaimed pages. <<
1245 * 1604 *
1246 * If a zone is deemed to be full of pinned p 1605 * If a zone is deemed to be full of pinned pages then just give it a light
1247 * scan then give up on it. 1606 * scan then give up on it.
1248 */ 1607 */
1249 static unsigned long shrink_zones(int priorit !! 1608 static void shrink_zones(int priority, struct zonelist *zonelist,
1250 struc 1609 struct scan_control *sc)
1251 { 1610 {
1252 unsigned long nr_reclaimed = 0; !! 1611 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1253 int i; !! 1612 struct zoneref *z;
1254 !! 1613 struct zone *zone;
1255 1614
1256 sc->all_unreclaimable = 1; 1615 sc->all_unreclaimable = 1;
1257 for (i = 0; zones[i] != NULL; i++) { !! 1616 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1258 struct zone *zone = zones[i]; !! 1617 sc->nodemask) {
1259 <<
1260 if (!populated_zone(zone)) 1618 if (!populated_zone(zone))
1261 continue; 1619 continue;
1262 /* 1620 /*
1263 * Take care memory controlle 1621 * Take care memory controller reclaiming has small influence
1264 * to global LRU. 1622 * to global LRU.
1265 */ 1623 */
1266 if (scan_global_lru(sc)) { !! 1624 if (scanning_global_lru(sc)) {
1267 if (!cpuset_zone_allo 1625 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1268 continue; 1626 continue;
1269 note_zone_scanning_pr 1627 note_zone_scanning_priority(zone, priority);
1270 1628
1271 if (zone_is_all_unrec 1629 if (zone_is_all_unreclaimable(zone) &&
1272 1630 priority != DEF_PRIORITY)
1273 continue; 1631 continue; /* Let kswapd poll it */
1274 sc->all_unreclaimable 1632 sc->all_unreclaimable = 0;
1275 } else { 1633 } else {
1276 /* 1634 /*
1277 * Ignore cpuset limi 1635 * Ignore cpuset limitation here. We just want to reduce
1278 * # of used pages by 1636 * # of used pages by us regardless of memory shortage.
1279 */ 1637 */
1280 sc->all_unreclaimable 1638 sc->all_unreclaimable = 0;
1281 mem_cgroup_note_recla 1639 mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1282 1640 priority);
1283 } 1641 }
1284 1642
1285 nr_reclaimed += shrink_zone(p !! 1643 shrink_zone(priority, zone, sc);
1286 } 1644 }
1287 <<
1288 return nr_reclaimed; <<
1289 } 1645 }
1290 !! 1646
1291 /* 1647 /*
1292 * This is the main entry point to direct pag 1648 * This is the main entry point to direct page reclaim.
1293 * 1649 *
1294 * If a full scan of the inactive list fails 1650 * If a full scan of the inactive list fails to free enough memory then we
1295 * are "out of memory" and something needs to 1651 * are "out of memory" and something needs to be killed.
1296 * 1652 *
1297 * If the caller is !__GFP_FS then the probab 1653 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1298 * high - the zone may be full of dirty or un 1654 * high - the zone may be full of dirty or under-writeback pages, which this
1299 * caller can't do much about. We kick pdflu 1655 * caller can't do much about. We kick pdflush and take explicit naps in the
1300 * hope that some of these pages can be writt 1656 * hope that some of these pages can be written. But if the allocating task
1301 * holds filesystem locks which prevent write 1657 * holds filesystem locks which prevent writeout this might not work, and the
1302 * allocation attempt will fail. 1658 * allocation attempt will fail.
>> 1659 *
>> 1660 * returns: 0, if no pages reclaimed
>> 1661 * else, the number of pages reclaimed
1303 */ 1662 */
1304 static unsigned long do_try_to_free_pages(str !! 1663 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1305 str !! 1664 struct scan_control *sc)
1306 { 1665 {
1307 int priority; 1666 int priority;
1308 int ret = 0; !! 1667 unsigned long ret = 0;
1309 unsigned long total_scanned = 0; 1668 unsigned long total_scanned = 0;
1310 unsigned long nr_reclaimed = 0; <<
1311 struct reclaim_state *reclaim_state = 1669 struct reclaim_state *reclaim_state = current->reclaim_state;
1312 unsigned long lru_pages = 0; 1670 unsigned long lru_pages = 0;
1313 int i; !! 1671 struct zoneref *z;
>> 1672 struct zone *zone;
>> 1673 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
1314 1674
1315 if (scan_global_lru(sc)) !! 1675 delayacct_freepages_start();
>> 1676
>> 1677 if (scanning_global_lru(sc))
1316 count_vm_event(ALLOCSTALL); 1678 count_vm_event(ALLOCSTALL);
1317 /* 1679 /*
1318 * mem_cgroup will not do shrink_slab 1680 * mem_cgroup will not do shrink_slab.
1319 */ 1681 */
1320 if (scan_global_lru(sc)) { !! 1682 if (scanning_global_lru(sc)) {
1321 for (i = 0; zones[i] != NULL; !! 1683 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1322 struct zone *zone = z <<
1323 1684
1324 if (!cpuset_zone_allo 1685 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1325 continue; 1686 continue;
1326 1687
1327 lru_pages += zone_pag !! 1688 lru_pages += zone_lru_pages(zone);
1328 + zon <<
1329 } 1689 }
1330 } 1690 }
1331 1691
1332 for (priority = DEF_PRIORITY; priorit 1692 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1333 sc->nr_scanned = 0; 1693 sc->nr_scanned = 0;
1334 if (!priority) 1694 if (!priority)
1335 disable_swap_token(); 1695 disable_swap_token();
1336 nr_reclaimed += shrink_zones( !! 1696 shrink_zones(priority, zonelist, sc);
1337 /* 1697 /*
1338 * Don't shrink slabs when re 1698 * Don't shrink slabs when reclaiming memory from
1339 * over limit cgroups 1699 * over limit cgroups
1340 */ 1700 */
1341 if (scan_global_lru(sc)) { !! 1701 if (scanning_global_lru(sc)) {
1342 shrink_slab(sc->nr_sc !! 1702 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1343 if (reclaim_state) { 1703 if (reclaim_state) {
1344 nr_reclaimed !! 1704 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1345 reclaim_state 1705 reclaim_state->reclaimed_slab = 0;
1346 } 1706 }
1347 } 1707 }
1348 total_scanned += sc->nr_scann 1708 total_scanned += sc->nr_scanned;
1349 if (nr_reclaimed >= sc->swap_ !! 1709 if (sc->nr_reclaimed >= sc->swap_cluster_max) {
1350 ret = 1; !! 1710 ret = sc->nr_reclaimed;
1351 goto out; 1711 goto out;
1352 } 1712 }
1353 1713
1354 /* 1714 /*
1355 * Try to write back as many 1715 * Try to write back as many pages as we just scanned. This
1356 * tends to cause slow stream 1716 * tends to cause slow streaming writers to write data to the
1357 * disk smoothly, at the dirt 1717 * disk smoothly, at the dirtying rate, which is nice. But
1358 * that's undesirable in lapt 1718 * that's undesirable in laptop mode, where we *want* lumpy
1359 * writeout. So in laptop mo 1719 * writeout. So in laptop mode, write out the whole world.
1360 */ 1720 */
1361 if (total_scanned > sc->swap_ 1721 if (total_scanned > sc->swap_cluster_max +
1362 sc->s 1722 sc->swap_cluster_max / 2) {
1363 wakeup_pdflush(laptop 1723 wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1364 sc->may_writepage = 1 1724 sc->may_writepage = 1;
1365 } 1725 }
1366 1726
1367 /* Take a nap, wait for some 1727 /* Take a nap, wait for some writeback to complete */
1368 if (sc->nr_scanned && priorit 1728 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
1369 congestion_wait(WRITE !! 1729 congestion_wait(BLK_RW_ASYNC, HZ/10);
1370 } 1730 }
1371 /* top priority shrink_caches still h !! 1731 /* top priority shrink_zones still had more to do? don't OOM, then */
1372 if (!sc->all_unreclaimable && scan_gl !! 1732 if (!sc->all_unreclaimable && scanning_global_lru(sc))
1373 ret = 1; !! 1733 ret = sc->nr_reclaimed;
1374 out: 1734 out:
1375 /* 1735 /*
1376 * Now that we've scanned all the zon 1736 * Now that we've scanned all the zones at this priority level, note
1377 * that level within the zone so that 1737 * that level within the zone so that the next thread which performs
1378 * scanning of this zone will immedia 1738 * scanning of this zone will immediately start out at this priority
1379 * level. This affects only the deci 1739 * level. This affects only the decision whether or not to bring
1380 * mapped pages onto the inactive lis 1740 * mapped pages onto the inactive list.
1381 */ 1741 */
1382 if (priority < 0) 1742 if (priority < 0)
1383 priority = 0; 1743 priority = 0;
1384 1744
1385 if (scan_global_lru(sc)) { !! 1745 if (scanning_global_lru(sc)) {
1386 for (i = 0; zones[i] != NULL; !! 1746 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1387 struct zone *zone = z <<
1388 1747
1389 if (!cpuset_zone_allo 1748 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1390 continue; 1749 continue;
1391 1750
1392 zone->prev_priority = 1751 zone->prev_priority = priority;
1393 } 1752 }
1394 } else 1753 } else
1395 mem_cgroup_record_reclaim_pri 1754 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1396 1755
>> 1756 delayacct_freepages_end();
>> 1757
1397 return ret; 1758 return ret;
1398 } 1759 }
1399 1760
1400 unsigned long try_to_free_pages(struct zone * !! 1761 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
>> 1762 gfp_t gfp_mask, nodemask_t *nodemask)
1401 { 1763 {
1402 struct scan_control sc = { 1764 struct scan_control sc = {
1403 .gfp_mask = gfp_mask, 1765 .gfp_mask = gfp_mask,
1404 .may_writepage = !laptop_mode 1766 .may_writepage = !laptop_mode,
1405 .swap_cluster_max = SWAP_CLUS 1767 .swap_cluster_max = SWAP_CLUSTER_MAX,
>> 1768 .may_unmap = 1,
1406 .may_swap = 1, 1769 .may_swap = 1,
1407 .swappiness = vm_swappiness, 1770 .swappiness = vm_swappiness,
1408 .order = order, 1771 .order = order,
1409 .mem_cgroup = NULL, 1772 .mem_cgroup = NULL,
1410 .isolate_pages = isolate_page 1773 .isolate_pages = isolate_pages_global,
>> 1774 .nodemask = nodemask,
1411 }; 1775 };
1412 1776
1413 return do_try_to_free_pages(zones, gf !! 1777 return do_try_to_free_pages(zonelist, &sc);
1414 } 1778 }
1415 1779
1416 #ifdef CONFIG_CGROUP_MEM_RES_CTLR 1780 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
1417 1781
1418 unsigned long try_to_free_mem_cgroup_pages(st 1782 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1419 !! 1783 gfp_t gfp_mask,
>> 1784 bool noswap,
>> 1785 unsigned int swappiness)
1420 { 1786 {
1421 struct scan_control sc = { 1787 struct scan_control sc = {
1422 .gfp_mask = gfp_mask, <<
1423 .may_writepage = !laptop_mode 1788 .may_writepage = !laptop_mode,
1424 .may_swap = 1, !! 1789 .may_unmap = 1,
>> 1790 .may_swap = !noswap,
1425 .swap_cluster_max = SWAP_CLUS 1791 .swap_cluster_max = SWAP_CLUSTER_MAX,
1426 .swappiness = vm_swappiness, !! 1792 .swappiness = swappiness,
1427 .order = 0, 1793 .order = 0,
1428 .mem_cgroup = mem_cont, 1794 .mem_cgroup = mem_cont,
1429 .isolate_pages = mem_cgroup_i 1795 .isolate_pages = mem_cgroup_isolate_pages,
>> 1796 .nodemask = NULL, /* we don't care the placement */
1430 }; 1797 };
1431 struct zone **zones; !! 1798 struct zonelist *zonelist;
1432 int target_zone = gfp_zone(GFP_HIGHUS <<
1433 1799
1434 zones = NODE_DATA(numa_node_id())->no !! 1800 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1435 if (do_try_to_free_pages(zones, sc.gf !! 1801 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1436 return 1; !! 1802 zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1437 return 0; !! 1803 return do_try_to_free_pages(zonelist, &sc);
1438 } 1804 }
1439 #endif 1805 #endif
1440 1806
1441 /* 1807 /*
1442 * For kswapd, balance_pgdat() will work acro 1808 * For kswapd, balance_pgdat() will work across all this node's zones until
1443 * they are all at pages_high. !! 1809 * they are all at high_wmark_pages(zone).
1444 * 1810 *
1445 * Returns the number of pages which were act 1811 * Returns the number of pages which were actually freed.
1446 * 1812 *
1447 * There is special handling here for zones w 1813 * There is special handling here for zones which are full of pinned pages.
1448 * This can happen if the pages are all mlock 1814 * This can happen if the pages are all mlocked, or if they are all used by
1449 * device drivers (say, ZONE_DMA). Or if the 1815 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1450 * What we do is to detect the case where all 1816 * What we do is to detect the case where all pages in the zone have been
1451 * scanned twice and there has been zero succ 1817 * scanned twice and there has been zero successful reclaim. Mark the zone as
1452 * dead and from now on, only perform a short 1818 * dead and from now on, only perform a short scan. Basically we're polling
1453 * the zone for when the problem goes away. 1819 * the zone for when the problem goes away.
1454 * 1820 *
1455 * kswapd scans the zones in the highmem->nor 1821 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1456 * zones which have free_pages > pages_high, !! 1822 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
1457 * free_pages <= pages_high, we scan that zon !! 1823 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
1458 * of the number of free pages in the lower z !! 1824 * lower zones regardless of the number of free pages in the lower zones. This
1459 * the page allocator fallback scheme to ensu !! 1825 * interoperates with the page allocator fallback scheme to ensure that aging
1460 * across the zones. !! 1826 * of pages is balanced across the zones.
1461 */ 1827 */
1462 static unsigned long balance_pgdat(pg_data_t 1828 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1463 { 1829 {
1464 int all_zones_ok; 1830 int all_zones_ok;
1465 int priority; 1831 int priority;
1466 int i; 1832 int i;
1467 unsigned long total_scanned; 1833 unsigned long total_scanned;
1468 unsigned long nr_reclaimed; <<
1469 struct reclaim_state *reclaim_state = 1834 struct reclaim_state *reclaim_state = current->reclaim_state;
1470 struct scan_control sc = { 1835 struct scan_control sc = {
1471 .gfp_mask = GFP_KERNEL, 1836 .gfp_mask = GFP_KERNEL,
>> 1837 .may_unmap = 1,
1472 .may_swap = 1, 1838 .may_swap = 1,
1473 .swap_cluster_max = SWAP_CLUS 1839 .swap_cluster_max = SWAP_CLUSTER_MAX,
1474 .swappiness = vm_swappiness, 1840 .swappiness = vm_swappiness,
1475 .order = order, 1841 .order = order,
1476 .mem_cgroup = NULL, 1842 .mem_cgroup = NULL,
1477 .isolate_pages = isolate_page 1843 .isolate_pages = isolate_pages_global,
1478 }; 1844 };
1479 /* 1845 /*
1480 * temp_priority is used to remember 1846 * temp_priority is used to remember the scanning priority at which
1481 * this zone was successfully refille !! 1847 * this zone was successfully refilled to
>> 1848 * free_pages == high_wmark_pages(zone).
1482 */ 1849 */
1483 int temp_priority[MAX_NR_ZONES]; 1850 int temp_priority[MAX_NR_ZONES];
1484 1851
1485 loop_again: 1852 loop_again:
1486 total_scanned = 0; 1853 total_scanned = 0;
1487 nr_reclaimed = 0; !! 1854 sc.nr_reclaimed = 0;
1488 sc.may_writepage = !laptop_mode; 1855 sc.may_writepage = !laptop_mode;
1489 count_vm_event(PAGEOUTRUN); 1856 count_vm_event(PAGEOUTRUN);
1490 1857
1491 for (i = 0; i < pgdat->nr_zones; i++) 1858 for (i = 0; i < pgdat->nr_zones; i++)
1492 temp_priority[i] = DEF_PRIORI 1859 temp_priority[i] = DEF_PRIORITY;
1493 1860
1494 for (priority = DEF_PRIORITY; priorit 1861 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1495 int end_zone = 0; /* In 1862 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
1496 unsigned long lru_pages = 0; 1863 unsigned long lru_pages = 0;
1497 1864
1498 /* The swap token gets in the 1865 /* The swap token gets in the way of swapout... */
1499 if (!priority) 1866 if (!priority)
1500 disable_swap_token(); 1867 disable_swap_token();
1501 1868
1502 all_zones_ok = 1; 1869 all_zones_ok = 1;
1503 1870
1504 /* 1871 /*
1505 * Scan in the highmem->dma d 1872 * Scan in the highmem->dma direction for the highest
1506 * zone which needs scanning 1873 * zone which needs scanning
1507 */ 1874 */
1508 for (i = pgdat->nr_zones - 1; 1875 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1509 struct zone *zone = p 1876 struct zone *zone = pgdat->node_zones + i;
1510 1877
1511 if (!populated_zone(z 1878 if (!populated_zone(zone))
1512 continue; 1879 continue;
1513 1880
1514 if (zone_is_all_unrec 1881 if (zone_is_all_unreclaimable(zone) &&
1515 priority != DEF_P 1882 priority != DEF_PRIORITY)
1516 continue; 1883 continue;
1517 1884
1518 if (!zone_watermark_o !! 1885 /*
1519 !! 1886 * Do some background aging of the anon list, to give
>> 1887 * pages a chance to be referenced before reclaiming.
>> 1888 */
>> 1889 if (inactive_anon_is_low(zone, &sc))
>> 1890 shrink_active_list(SWAP_CLUSTER_MAX, zone,
>> 1891 &sc, priority, 0);
>> 1892
>> 1893 if (!zone_watermark_ok(zone, order,
>> 1894 high_wmark_pages(zone), 0, 0)) {
1520 end_zone = i; 1895 end_zone = i;
1521 break; 1896 break;
1522 } 1897 }
1523 } 1898 }
1524 if (i < 0) 1899 if (i < 0)
1525 goto out; 1900 goto out;
1526 1901
1527 for (i = 0; i <= end_zone; i+ 1902 for (i = 0; i <= end_zone; i++) {
1528 struct zone *zone = p 1903 struct zone *zone = pgdat->node_zones + i;
1529 1904
1530 lru_pages += zone_pag !! 1905 lru_pages += zone_lru_pages(zone);
1531 + zon <<
1532 } 1906 }
1533 1907
1534 /* 1908 /*
1535 * Now scan the zone in the d 1909 * Now scan the zone in the dma->highmem direction, stopping
1536 * at the last zone which nee 1910 * at the last zone which needs scanning.
1537 * 1911 *
1538 * We do this because the pag 1912 * We do this because the page allocator works in the opposite
1539 * direction. This prevents 1913 * direction. This prevents the page allocator from allocating
1540 * pages behind kswapd's dire 1914 * pages behind kswapd's direction of progress, which would
1541 * cause too much scanning of 1915 * cause too much scanning of the lower zones.
1542 */ 1916 */
1543 for (i = 0; i <= end_zone; i+ 1917 for (i = 0; i <= end_zone; i++) {
1544 struct zone *zone = p 1918 struct zone *zone = pgdat->node_zones + i;
1545 int nr_slab; 1919 int nr_slab;
1546 1920
1547 if (!populated_zone(z 1921 if (!populated_zone(zone))
1548 continue; 1922 continue;
1549 1923
1550 if (zone_is_all_unrec 1924 if (zone_is_all_unreclaimable(zone) &&
1551 prior 1925 priority != DEF_PRIORITY)
1552 continue; 1926 continue;
1553 1927
1554 if (!zone_watermark_o !! 1928 if (!zone_watermark_ok(zone, order,
1555 !! 1929 high_wmark_pages(zone), end_zone, 0))
1556 all_zones_ok 1930 all_zones_ok = 0;
1557 temp_priority[i] = pr 1931 temp_priority[i] = priority;
1558 sc.nr_scanned = 0; 1932 sc.nr_scanned = 0;
1559 note_zone_scanning_pr 1933 note_zone_scanning_priority(zone, priority);
1560 /* 1934 /*
1561 * We put equal press 1935 * We put equal pressure on every zone, unless one
1562 * zone has way too m 1936 * zone has way too many pages free already.
1563 */ 1937 */
1564 if (!zone_watermark_o !! 1938 if (!zone_watermark_ok(zone, order,
1565 !! 1939 8*high_wmark_pages(zone), end_zone, 0))
1566 nr_reclaimed !! 1940 shrink_zone(priority, zone, &sc);
1567 reclaim_state->reclai 1941 reclaim_state->reclaimed_slab = 0;
1568 nr_slab = shrink_slab 1942 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1569 1943 lru_pages);
1570 nr_reclaimed += recla !! 1944 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
1571 total_scanned += sc.n 1945 total_scanned += sc.nr_scanned;
1572 if (zone_is_all_unrec 1946 if (zone_is_all_unreclaimable(zone))
1573 continue; 1947 continue;
1574 if (nr_slab == 0 && z 1948 if (nr_slab == 0 && zone->pages_scanned >=
1575 (zone_page_st !! 1949 (zone_lru_pages(zone) * 6))
1576 + zone_page_s <<
1577 zone_ 1950 zone_set_flag(zone,
1578 1951 ZONE_ALL_UNRECLAIMABLE);
1579 /* 1952 /*
1580 * If we've done a de 1953 * If we've done a decent amount of scanning and
1581 * the reclaim ratio 1954 * the reclaim ratio is low, start doing writepage
1582 * even in laptop mod 1955 * even in laptop mode
1583 */ 1956 */
1584 if (total_scanned > S 1957 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1585 total_scanned > n !! 1958 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
1586 sc.may_writep 1959 sc.may_writepage = 1;
1587 } 1960 }
1588 if (all_zones_ok) 1961 if (all_zones_ok)
1589 break; /* ks 1962 break; /* kswapd: all done */
1590 /* 1963 /*
1591 * OK, kswapd is getting into 1964 * OK, kswapd is getting into trouble. Take a nap, then take
1592 * another pass across the zo 1965 * another pass across the zones.
1593 */ 1966 */
1594 if (total_scanned && priority 1967 if (total_scanned && priority < DEF_PRIORITY - 2)
1595 congestion_wait(WRITE !! 1968 congestion_wait(BLK_RW_ASYNC, HZ/10);
1596 1969
1597 /* 1970 /*
1598 * We do this so kswapd doesn 1971 * We do this so kswapd doesn't build up large priorities for
1599 * example when it is freeing 1972 * example when it is freeing in parallel with allocators. It
1600 * matches the direct reclaim 1973 * matches the direct reclaim path behaviour in terms of impact
1601 * on zone->*_priority. 1974 * on zone->*_priority.
1602 */ 1975 */
1603 if (nr_reclaimed >= SWAP_CLUS !! 1976 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
1604 break; 1977 break;
1605 } 1978 }
1606 out: 1979 out:
1607 /* 1980 /*
1608 * Note within each zone the priority 1981 * Note within each zone the priority level at which this zone was
1609 * brought into a happy state. So th 1982 * brought into a happy state. So that the next thread which scans this
1610 * zone will start out at that priori 1983 * zone will start out at that priority level.
1611 */ 1984 */
1612 for (i = 0; i < pgdat->nr_zones; i++) 1985 for (i = 0; i < pgdat->nr_zones; i++) {
1613 struct zone *zone = pgdat->no 1986 struct zone *zone = pgdat->node_zones + i;
1614 1987
1615 zone->prev_priority = temp_pr 1988 zone->prev_priority = temp_priority[i];
1616 } 1989 }
1617 if (!all_zones_ok) { 1990 if (!all_zones_ok) {
1618 cond_resched(); 1991 cond_resched();
1619 1992
1620 try_to_freeze(); 1993 try_to_freeze();
1621 1994
>> 1995 /*
>> 1996 * Fragmentation may mean that the system cannot be
>> 1997 * rebalanced for high-order allocations in all zones.
>> 1998 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
>> 1999 * it means the zones have been fully scanned and are still
>> 2000 * not balanced. For high-order allocations, there is
>> 2001 * little point trying all over again as kswapd may
>> 2002 * infinite loop.
>> 2003 *
>> 2004 * Instead, recheck all watermarks at order-0 as they
>> 2005 * are the most important. If watermarks are ok, kswapd will go
>> 2006 * back to sleep. High-order users can still perform direct
>> 2007 * reclaim if they wish.
>> 2008 */
>> 2009 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
>> 2010 order = sc.order = 0;
>> 2011
1622 goto loop_again; 2012 goto loop_again;
1623 } 2013 }
1624 2014
1625 return nr_reclaimed; !! 2015 return sc.nr_reclaimed;
1626 } 2016 }
1627 2017
1628 /* 2018 /*
1629 * The background pageout daemon, started as 2019 * The background pageout daemon, started as a kernel thread
1630 * from the init process. !! 2020 * from the init process.
1631 * 2021 *
1632 * This basically trickles out pages so that 2022 * This basically trickles out pages so that we have _some_
1633 * free memory available even if there is no 2023 * free memory available even if there is no other activity
1634 * that frees anything up. This is needed for 2024 * that frees anything up. This is needed for things like routing
1635 * etc, where we otherwise might have all act 2025 * etc, where we otherwise might have all activity going on in
1636 * asynchronous contexts that cannot page thi 2026 * asynchronous contexts that cannot page things out.
1637 * 2027 *
1638 * If there are applications that are active 2028 * If there are applications that are active memory-allocators
1639 * (most normal use), this basically shouldn' 2029 * (most normal use), this basically shouldn't matter.
1640 */ 2030 */
1641 static int kswapd(void *p) 2031 static int kswapd(void *p)
1642 { 2032 {
1643 unsigned long order; 2033 unsigned long order;
1644 pg_data_t *pgdat = (pg_data_t*)p; 2034 pg_data_t *pgdat = (pg_data_t*)p;
1645 struct task_struct *tsk = current; 2035 struct task_struct *tsk = current;
1646 DEFINE_WAIT(wait); 2036 DEFINE_WAIT(wait);
1647 struct reclaim_state reclaim_state = 2037 struct reclaim_state reclaim_state = {
1648 .reclaimed_slab = 0, 2038 .reclaimed_slab = 0,
1649 }; 2039 };
1650 cpumask_t cpumask; !! 2040 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
>> 2041
>> 2042 lockdep_set_current_reclaim_state(GFP_KERNEL);
1651 2043
1652 cpumask = node_to_cpumask(pgdat->node !! 2044 if (!cpumask_empty(cpumask))
1653 if (!cpus_empty(cpumask)) !! 2045 set_cpus_allowed_ptr(tsk, cpumask);
1654 set_cpus_allowed(tsk, cpumask <<
1655 current->reclaim_state = &reclaim_sta 2046 current->reclaim_state = &reclaim_state;
1656 2047
1657 /* 2048 /*
1658 * Tell the memory management that we 2049 * Tell the memory management that we're a "memory allocator",
1659 * and that if we need more memory we 2050 * and that if we need more memory we should get access to it
1660 * regardless (see "__alloc_pages()") 2051 * regardless (see "__alloc_pages()"). "kswapd" should
1661 * never get caught in the normal pag 2052 * never get caught in the normal page freeing logic.
1662 * 2053 *
1663 * (Kswapd normally doesn't need memo 2054 * (Kswapd normally doesn't need memory anyway, but sometimes
1664 * you need a small amount of memory 2055 * you need a small amount of memory in order to be able to
1665 * page out something else, and this 2056 * page out something else, and this flag essentially protects
1666 * us from recursively trying to free 2057 * us from recursively trying to free more memory as we're
1667 * trying to free the first piece of 2058 * trying to free the first piece of memory in the first place).
1668 */ 2059 */
1669 tsk->flags |= PF_MEMALLOC | PF_SWAPWR 2060 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1670 set_freezable(); 2061 set_freezable();
1671 2062
1672 order = 0; 2063 order = 0;
1673 for ( ; ; ) { 2064 for ( ; ; ) {
1674 unsigned long new_order; 2065 unsigned long new_order;
1675 2066
1676 prepare_to_wait(&pgdat->kswap 2067 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1677 new_order = pgdat->kswapd_max 2068 new_order = pgdat->kswapd_max_order;
1678 pgdat->kswapd_max_order = 0; 2069 pgdat->kswapd_max_order = 0;
1679 if (order < new_order) { 2070 if (order < new_order) {
1680 /* 2071 /*
1681 * Don't sleep if som 2072 * Don't sleep if someone wants a larger 'order'
1682 * allocation 2073 * allocation
1683 */ 2074 */
1684 order = new_order; 2075 order = new_order;
1685 } else { 2076 } else {
1686 if (!freezing(current 2077 if (!freezing(current))
1687 schedule(); 2078 schedule();
1688 2079
1689 order = pgdat->kswapd 2080 order = pgdat->kswapd_max_order;
1690 } 2081 }
1691 finish_wait(&pgdat->kswapd_wa 2082 finish_wait(&pgdat->kswapd_wait, &wait);
1692 2083
1693 if (!try_to_freeze()) { 2084 if (!try_to_freeze()) {
1694 /* We can speed up th 2085 /* We can speed up thawing tasks if we don't call
1695 * balance_pgdat afte 2086 * balance_pgdat after returning from the refrigerator
1696 */ 2087 */
1697 balance_pgdat(pgdat, 2088 balance_pgdat(pgdat, order);
1698 } 2089 }
1699 } 2090 }
1700 return 0; 2091 return 0;
1701 } 2092 }
1702 2093
1703 /* 2094 /*
1704 * A zone is low on free memory, so wake its 2095 * A zone is low on free memory, so wake its kswapd task to service it.
1705 */ 2096 */
1706 void wakeup_kswapd(struct zone *zone, int ord 2097 void wakeup_kswapd(struct zone *zone, int order)
1707 { 2098 {
1708 pg_data_t *pgdat; 2099 pg_data_t *pgdat;
1709 2100
1710 if (!populated_zone(zone)) 2101 if (!populated_zone(zone))
1711 return; 2102 return;
1712 2103
1713 pgdat = zone->zone_pgdat; 2104 pgdat = zone->zone_pgdat;
1714 if (zone_watermark_ok(zone, order, zo !! 2105 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
1715 return; 2106 return;
1716 if (pgdat->kswapd_max_order < order) 2107 if (pgdat->kswapd_max_order < order)
1717 pgdat->kswapd_max_order = ord 2108 pgdat->kswapd_max_order = order;
1718 if (!cpuset_zone_allowed_hardwall(zon 2109 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1719 return; 2110 return;
1720 if (!waitqueue_active(&pgdat->kswapd_ 2111 if (!waitqueue_active(&pgdat->kswapd_wait))
1721 return; 2112 return;
1722 wake_up_interruptible(&pgdat->kswapd_ 2113 wake_up_interruptible(&pgdat->kswapd_wait);
1723 } 2114 }
1724 2115
1725 #ifdef CONFIG_PM !! 2116 unsigned long global_lru_pages(void)
>> 2117 {
>> 2118 return global_page_state(NR_ACTIVE_ANON)
>> 2119 + global_page_state(NR_ACTIVE_FILE)
>> 2120 + global_page_state(NR_INACTIVE_ANON)
>> 2121 + global_page_state(NR_INACTIVE_FILE);
>> 2122 }
>> 2123
>> 2124 #ifdef CONFIG_HIBERNATION
1726 /* 2125 /*
1727 * Helper function for shrink_all_memory(). 2126 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages
1728 * from LRU lists system-wide, for given pass !! 2127 * from LRU lists system-wide, for given pass and priority.
1729 * number of reclaimed pages <<
1730 * 2128 *
1731 * For pass > 3 we also try to shrink the LRU 2129 * For pass > 3 we also try to shrink the LRU lists that contain a few pages
1732 */ 2130 */
1733 static unsigned long shrink_all_zones(unsigne !! 2131 static void shrink_all_zones(unsigned long nr_pages, int prio,
1734 int pas 2132 int pass, struct scan_control *sc)
1735 { 2133 {
1736 struct zone *zone; 2134 struct zone *zone;
1737 unsigned long nr_to_scan, ret = 0; !! 2135 unsigned long nr_reclaimed = 0;
1738 2136
1739 for_each_zone(zone) { !! 2137 for_each_populated_zone(zone) {
1740 !! 2138 enum lru_list l;
1741 if (!populated_zone(zone)) <<
1742 continue; <<
1743 2139
1744 if (zone_is_all_unreclaimable 2140 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
1745 continue; 2141 continue;
1746 2142
1747 /* For pass = 0 we don't shri !! 2143 for_each_evictable_lru(l) {
1748 if (pass > 0) { !! 2144 enum zone_stat_item ls = NR_LRU_BASE + l;
1749 zone->nr_scan_active !! 2145 unsigned long lru_pages = zone_page_state(zone, ls);
1750 (zone_page_st !! 2146
1751 if (zone->nr_scan_act !! 2147 /* For pass = 0, we don't shrink the active list */
1752 zone->nr_scan !! 2148 if (pass == 0 && (l == LRU_ACTIVE_ANON ||
1753 nr_to_scan = !! 2149 l == LRU_ACTIVE_FILE))
1754 zone_ !! 2150 continue;
1755 shrink_active <<
1756 } <<
1757 } <<
1758 2151
1759 zone->nr_scan_inactive += !! 2152 zone->lru[l].nr_saved_scan += (lru_pages >> prio) + 1;
1760 (zone_page_state(zone !! 2153 if (zone->lru[l].nr_saved_scan >= nr_pages || pass > 3) {
1761 if (zone->nr_scan_inactive >= !! 2154 unsigned long nr_to_scan;
1762 zone->nr_scan_inactiv !! 2155
1763 nr_to_scan = min(nr_p !! 2156 zone->lru[l].nr_saved_scan = 0;
1764 zone_page_sta !! 2157 nr_to_scan = min(nr_pages, lru_pages);
1765 ret += shrink_inactiv !! 2158 nr_reclaimed += shrink_list(l, nr_to_scan, zone,
1766 if (ret >= nr_pages) !! 2159 sc, prio);
1767 return ret; !! 2160 if (nr_reclaimed >= nr_pages) {
>> 2161 sc->nr_reclaimed += nr_reclaimed;
>> 2162 return;
>> 2163 }
>> 2164 }
1768 } 2165 }
1769 } 2166 }
1770 !! 2167 sc->nr_reclaimed += nr_reclaimed;
1771 return ret; <<
1772 } <<
1773 <<
1774 static unsigned long count_lru_pages(void) <<
1775 { <<
1776 return global_page_state(NR_ACTIVE) + <<
1777 } 2168 }
1778 2169
1779 /* 2170 /*
1780 * Try to free `nr_pages' of memory, system-w 2171 * Try to free `nr_pages' of memory, system-wide, and return the number of
1781 * freed pages. 2172 * freed pages.
1782 * 2173 *
1783 * Rather than trying to age LRUs the aim is 2174 * Rather than trying to age LRUs the aim is to preserve the overall
1784 * LRU order by reclaiming preferentially 2175 * LRU order by reclaiming preferentially
1785 * inactive > active > active referenced > ac 2176 * inactive > active > active referenced > active mapped
1786 */ 2177 */
1787 unsigned long shrink_all_memory(unsigned long 2178 unsigned long shrink_all_memory(unsigned long nr_pages)
1788 { 2179 {
1789 unsigned long lru_pages, nr_slab; 2180 unsigned long lru_pages, nr_slab;
1790 unsigned long ret = 0; <<
1791 int pass; 2181 int pass;
1792 struct reclaim_state reclaim_state; 2182 struct reclaim_state reclaim_state;
1793 struct scan_control sc = { 2183 struct scan_control sc = {
1794 .gfp_mask = GFP_KERNEL, 2184 .gfp_mask = GFP_KERNEL,
1795 .may_swap = 0, !! 2185 .may_unmap = 0,
1796 .swap_cluster_max = nr_pages, <<
1797 .may_writepage = 1, 2186 .may_writepage = 1,
1798 .swappiness = vm_swappiness, <<
1799 .isolate_pages = isolate_page 2187 .isolate_pages = isolate_pages_global,
>> 2188 .nr_reclaimed = 0,
1800 }; 2189 };
1801 2190
1802 current->reclaim_state = &reclaim_sta 2191 current->reclaim_state = &reclaim_state;
1803 2192
1804 lru_pages = count_lru_pages(); !! 2193 lru_pages = global_lru_pages();
1805 nr_slab = global_page_state(NR_SLAB_R 2194 nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
1806 /* If slab caches are huge, it's bett 2195 /* If slab caches are huge, it's better to hit them first */
1807 while (nr_slab >= lru_pages) { 2196 while (nr_slab >= lru_pages) {
1808 reclaim_state.reclaimed_slab 2197 reclaim_state.reclaimed_slab = 0;
1809 shrink_slab(nr_pages, sc.gfp_ 2198 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
1810 if (!reclaim_state.reclaimed_ 2199 if (!reclaim_state.reclaimed_slab)
1811 break; 2200 break;
1812 2201
1813 ret += reclaim_state.reclaime !! 2202 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
1814 if (ret >= nr_pages) !! 2203 if (sc.nr_reclaimed >= nr_pages)
1815 goto out; 2204 goto out;
1816 2205
1817 nr_slab -= reclaim_state.recl 2206 nr_slab -= reclaim_state.reclaimed_slab;
1818 } 2207 }
1819 2208
1820 /* 2209 /*
1821 * We try to shrink LRUs in 5 passes: 2210 * We try to shrink LRUs in 5 passes:
1822 * 0 = Reclaim from inactive_list onl 2211 * 0 = Reclaim from inactive_list only
1823 * 1 = Reclaim from active list but d 2212 * 1 = Reclaim from active list but don't reclaim mapped
1824 * 2 = 2nd pass of type 1 2213 * 2 = 2nd pass of type 1
1825 * 3 = Reclaim mapped (normal reclaim 2214 * 3 = Reclaim mapped (normal reclaim)
1826 * 4 = 2nd pass of type 3 2215 * 4 = 2nd pass of type 3
1827 */ 2216 */
1828 for (pass = 0; pass < 5; pass++) { 2217 for (pass = 0; pass < 5; pass++) {
1829 int prio; 2218 int prio;
1830 2219
1831 /* Force reclaiming mapped pa 2220 /* Force reclaiming mapped pages in the passes #3 and #4 */
1832 if (pass > 2) { !! 2221 if (pass > 2)
1833 sc.may_swap = 1; !! 2222 sc.may_unmap = 1;
1834 sc.swappiness = 100; <<
1835 } <<
1836 2223
1837 for (prio = DEF_PRIORITY; pri 2224 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
1838 unsigned long nr_to_s !! 2225 unsigned long nr_to_scan = nr_pages - sc.nr_reclaimed;
1839 2226
1840 sc.nr_scanned = 0; 2227 sc.nr_scanned = 0;
1841 ret += shrink_all_zon !! 2228 sc.swap_cluster_max = nr_to_scan;
1842 if (ret >= nr_pages) !! 2229 shrink_all_zones(nr_to_scan, prio, pass, &sc);
>> 2230 if (sc.nr_reclaimed >= nr_pages)
1843 goto out; 2231 goto out;
1844 2232
1845 reclaim_state.reclaim 2233 reclaim_state.reclaimed_slab = 0;
1846 shrink_slab(sc.nr_sca 2234 shrink_slab(sc.nr_scanned, sc.gfp_mask,
1847 count !! 2235 global_lru_pages());
1848 ret += reclaim_state. !! 2236 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
1849 if (ret >= nr_pages) !! 2237 if (sc.nr_reclaimed >= nr_pages)
1850 goto out; 2238 goto out;
1851 2239
1852 if (sc.nr_scanned && 2240 if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
1853 congestion_wa !! 2241 congestion_wait(BLK_RW_ASYNC, HZ / 10);
1854 } 2242 }
1855 } 2243 }
1856 2244
1857 /* 2245 /*
1858 * If ret = 0, we could not shrink LR !! 2246 * If sc.nr_reclaimed = 0, we could not shrink LRUs, but there may be
1859 * in slab caches !! 2247 * something in slab caches
1860 */ 2248 */
1861 if (!ret) { !! 2249 if (!sc.nr_reclaimed) {
1862 do { 2250 do {
1863 reclaim_state.reclaim 2251 reclaim_state.reclaimed_slab = 0;
1864 shrink_slab(nr_pages, !! 2252 shrink_slab(nr_pages, sc.gfp_mask, global_lru_pages());
1865 ret += reclaim_state. !! 2253 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
1866 } while (ret < nr_pages && re !! 2254 } while (sc.nr_reclaimed < nr_pages &&
>> 2255 reclaim_state.reclaimed_slab > 0);
1867 } 2256 }
1868 2257
>> 2258
1869 out: 2259 out:
1870 current->reclaim_state = NULL; 2260 current->reclaim_state = NULL;
1871 2261
1872 return ret; !! 2262 return sc.nr_reclaimed;
1873 } 2263 }
1874 #endif !! 2264 #endif /* CONFIG_HIBERNATION */
1875 2265
1876 /* It's optimal to keep kswapds on the same C 2266 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1877 not required for correctness. So if the l 2267 not required for correctness. So if the last cpu in a node goes
1878 away, we get changed to run anywhere: as t 2268 away, we get changed to run anywhere: as the first one comes back,
1879 restore their cpu bindings. */ 2269 restore their cpu bindings. */
1880 static int __devinit cpu_callback(struct noti 2270 static int __devinit cpu_callback(struct notifier_block *nfb,
1881 unsigned lo 2271 unsigned long action, void *hcpu)
1882 { 2272 {
1883 pg_data_t *pgdat; <<
1884 cpumask_t mask; <<
1885 int nid; 2273 int nid;
1886 2274
1887 if (action == CPU_ONLINE || action == 2275 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
1888 for_each_node_state(nid, N_HI 2276 for_each_node_state(nid, N_HIGH_MEMORY) {
1889 pgdat = NODE_DATA(nid !! 2277 pg_data_t *pgdat = NODE_DATA(nid);
1890 mask = node_to_cpumas !! 2278 const struct cpumask *mask;
1891 if (any_online_cpu(ma !! 2279
>> 2280 mask = cpumask_of_node(pgdat->node_id);
>> 2281
>> 2282 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
1892 /* One of our 2283 /* One of our CPUs online: restore mask */
1893 set_cpus_allo !! 2284 set_cpus_allowed_ptr(pgdat->kswapd, mask);
1894 } 2285 }
1895 } 2286 }
1896 return NOTIFY_OK; 2287 return NOTIFY_OK;
1897 } 2288 }
1898 2289
1899 /* 2290 /*
1900 * This kswapd start function will be called 2291 * This kswapd start function will be called by init and node-hot-add.
1901 * On node-hot-add, kswapd will moved to prop 2292 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
1902 */ 2293 */
1903 int kswapd_run(int nid) 2294 int kswapd_run(int nid)
1904 { 2295 {
1905 pg_data_t *pgdat = NODE_DATA(nid); 2296 pg_data_t *pgdat = NODE_DATA(nid);
1906 int ret = 0; 2297 int ret = 0;
1907 2298
1908 if (pgdat->kswapd) 2299 if (pgdat->kswapd)
1909 return 0; 2300 return 0;
1910 2301
1911 pgdat->kswapd = kthread_run(kswapd, p 2302 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
1912 if (IS_ERR(pgdat->kswapd)) { 2303 if (IS_ERR(pgdat->kswapd)) {
1913 /* failure at boot is fatal * 2304 /* failure at boot is fatal */
1914 BUG_ON(system_state == SYSTEM 2305 BUG_ON(system_state == SYSTEM_BOOTING);
1915 printk("Failed to start kswap 2306 printk("Failed to start kswapd on node %d\n",nid);
1916 ret = -1; 2307 ret = -1;
1917 } 2308 }
1918 return ret; 2309 return ret;
1919 } 2310 }
1920 2311
1921 static int __init kswapd_init(void) 2312 static int __init kswapd_init(void)
1922 { 2313 {
1923 int nid; 2314 int nid;
1924 2315
1925 swap_setup(); 2316 swap_setup();
1926 for_each_node_state(nid, N_HIGH_MEMOR 2317 for_each_node_state(nid, N_HIGH_MEMORY)
1927 kswapd_run(nid); 2318 kswapd_run(nid);
1928 hotcpu_notifier(cpu_callback, 0); 2319 hotcpu_notifier(cpu_callback, 0);
1929 return 0; 2320 return 0;
1930 } 2321 }
1931 2322
1932 module_init(kswapd_init) 2323 module_init(kswapd_init)
1933 2324
1934 #ifdef CONFIG_NUMA 2325 #ifdef CONFIG_NUMA
1935 /* 2326 /*
1936 * Zone reclaim mode 2327 * Zone reclaim mode
1937 * 2328 *
1938 * If non-zero call zone_reclaim when the num 2329 * If non-zero call zone_reclaim when the number of free pages falls below
1939 * the watermarks. 2330 * the watermarks.
1940 */ 2331 */
1941 int zone_reclaim_mode __read_mostly; 2332 int zone_reclaim_mode __read_mostly;
1942 2333
1943 #define RECLAIM_OFF 0 2334 #define RECLAIM_OFF 0
1944 #define RECLAIM_ZONE (1<<0) /* Run shrink !! 2335 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
1945 #define RECLAIM_WRITE (1<<1) /* Writeout p 2336 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
1946 #define RECLAIM_SWAP (1<<2) /* Swap pages 2337 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
1947 2338
1948 /* 2339 /*
1949 * Priority for ZONE_RECLAIM. This determines 2340 * Priority for ZONE_RECLAIM. This determines the fraction of pages
1950 * of a node considered for each zone_reclaim 2341 * of a node considered for each zone_reclaim. 4 scans 1/16th of
1951 * a zone. 2342 * a zone.
1952 */ 2343 */
1953 #define ZONE_RECLAIM_PRIORITY 4 2344 #define ZONE_RECLAIM_PRIORITY 4
1954 2345
1955 /* 2346 /*
1956 * Percentage of pages in a zone that must be 2347 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
1957 * occur. 2348 * occur.
1958 */ 2349 */
1959 int sysctl_min_unmapped_ratio = 1; 2350 int sysctl_min_unmapped_ratio = 1;
1960 2351
1961 /* 2352 /*
1962 * If the number of slab pages in a zone grow 2353 * If the number of slab pages in a zone grows beyond this percentage then
1963 * slab reclaim needs to occur. 2354 * slab reclaim needs to occur.
1964 */ 2355 */
1965 int sysctl_min_slab_ratio = 5; 2356 int sysctl_min_slab_ratio = 5;
1966 2357
>> 2358 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
>> 2359 {
>> 2360 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
>> 2361 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
>> 2362 zone_page_state(zone, NR_ACTIVE_FILE);
>> 2363
>> 2364 /*
>> 2365 * It's possible for there to be more file mapped pages than
>> 2366 * accounted for by the pages on the file LRU lists because
>> 2367 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
>> 2368 */
>> 2369 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
>> 2370 }
>> 2371
>> 2372 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
>> 2373 static long zone_pagecache_reclaimable(struct zone *zone)
>> 2374 {
>> 2375 long nr_pagecache_reclaimable;
>> 2376 long delta = 0;
>> 2377
>> 2378 /*
>> 2379 * If RECLAIM_SWAP is set, then all file pages are considered
>> 2380 * potentially reclaimable. Otherwise, we have to worry about
>> 2381 * pages like swapcache and zone_unmapped_file_pages() provides
>> 2382 * a better estimate
>> 2383 */
>> 2384 if (zone_reclaim_mode & RECLAIM_SWAP)
>> 2385 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
>> 2386 else
>> 2387 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
>> 2388
>> 2389 /* If we can't clean pages, remove dirty pages from consideration */
>> 2390 if (!(zone_reclaim_mode & RECLAIM_WRITE))
>> 2391 delta += zone_page_state(zone, NR_FILE_DIRTY);
>> 2392
>> 2393 /* Watch for any possible underflows due to delta */
>> 2394 if (unlikely(delta > nr_pagecache_reclaimable))
>> 2395 delta = nr_pagecache_reclaimable;
>> 2396
>> 2397 return nr_pagecache_reclaimable - delta;
>> 2398 }
>> 2399
1967 /* 2400 /*
1968 * Try to free up some pages from this zone t 2401 * Try to free up some pages from this zone through reclaim.
1969 */ 2402 */
1970 static int __zone_reclaim(struct zone *zone, 2403 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1971 { 2404 {
1972 /* Minimum pages needed in order to s 2405 /* Minimum pages needed in order to stay on node */
1973 const unsigned long nr_pages = 1 << o 2406 const unsigned long nr_pages = 1 << order;
1974 struct task_struct *p = current; 2407 struct task_struct *p = current;
1975 struct reclaim_state reclaim_state; 2408 struct reclaim_state reclaim_state;
1976 int priority; 2409 int priority;
1977 unsigned long nr_reclaimed = 0; <<
1978 struct scan_control sc = { 2410 struct scan_control sc = {
1979 .may_writepage = !!(zone_recl 2411 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
1980 .may_swap = !!(zone_reclaim_m !! 2412 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
>> 2413 .may_swap = 1,
1981 .swap_cluster_max = max_t(uns 2414 .swap_cluster_max = max_t(unsigned long, nr_pages,
1982 SWAP_ 2415 SWAP_CLUSTER_MAX),
1983 .gfp_mask = gfp_mask, 2416 .gfp_mask = gfp_mask,
1984 .swappiness = vm_swappiness, 2417 .swappiness = vm_swappiness,
>> 2418 .order = order,
1985 .isolate_pages = isolate_page 2419 .isolate_pages = isolate_pages_global,
1986 }; 2420 };
1987 unsigned long slab_reclaimable; 2421 unsigned long slab_reclaimable;
1988 2422
1989 disable_swap_token(); 2423 disable_swap_token();
1990 cond_resched(); 2424 cond_resched();
1991 /* 2425 /*
1992 * We need to be able to allocate fro 2426 * We need to be able to allocate from the reserves for RECLAIM_SWAP
1993 * and we also need to be able to wri 2427 * and we also need to be able to write out pages for RECLAIM_WRITE
1994 * and RECLAIM_SWAP. 2428 * and RECLAIM_SWAP.
1995 */ 2429 */
1996 p->flags |= PF_MEMALLOC | PF_SWAPWRIT 2430 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
1997 reclaim_state.reclaimed_slab = 0; 2431 reclaim_state.reclaimed_slab = 0;
1998 p->reclaim_state = &reclaim_state; 2432 p->reclaim_state = &reclaim_state;
1999 2433
2000 if (zone_page_state(zone, NR_FILE_PAG !! 2434 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2001 zone_page_state(zone, NR_FILE <<
2002 zone->min_unmapped_pages) { <<
2003 /* 2435 /*
2004 * Free memory by calling shr 2436 * Free memory by calling shrink zone with increasing
2005 * priorities until we have e 2437 * priorities until we have enough memory freed.
2006 */ 2438 */
2007 priority = ZONE_RECLAIM_PRIOR 2439 priority = ZONE_RECLAIM_PRIORITY;
2008 do { 2440 do {
2009 note_zone_scanning_pr 2441 note_zone_scanning_priority(zone, priority);
2010 nr_reclaimed += shrin !! 2442 shrink_zone(priority, zone, &sc);
2011 priority--; 2443 priority--;
2012 } while (priority >= 0 && nr_ !! 2444 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2013 } 2445 }
2014 2446
2015 slab_reclaimable = zone_page_state(zo 2447 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2016 if (slab_reclaimable > zone->min_slab 2448 if (slab_reclaimable > zone->min_slab_pages) {
2017 /* 2449 /*
2018 * shrink_slab() does not cur 2450 * shrink_slab() does not currently allow us to determine how
2019 * many pages were freed in t 2451 * many pages were freed in this zone. So we take the current
2020 * number of slab pages and s 2452 * number of slab pages and shake the slab until it is reduced
2021 * by the same nr_pages that 2453 * by the same nr_pages that we used for reclaiming unmapped
2022 * pages. 2454 * pages.
2023 * 2455 *
2024 * Note that shrink_slab will 2456 * Note that shrink_slab will free memory on all zones and may
2025 * take a long time. 2457 * take a long time.
2026 */ 2458 */
2027 while (shrink_slab(sc.nr_scan 2459 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
2028 zone_page_state(zone, 2460 zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2029 slab_reclaima 2461 slab_reclaimable - nr_pages)
2030 ; 2462 ;
2031 2463
2032 /* 2464 /*
2033 * Update nr_reclaimed by the 2465 * Update nr_reclaimed by the number of slab pages we
2034 * reclaimed from this zone. 2466 * reclaimed from this zone.
2035 */ 2467 */
2036 nr_reclaimed += slab_reclaima !! 2468 sc.nr_reclaimed += slab_reclaimable -
2037 zone_page_state(zone, 2469 zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2038 } 2470 }
2039 2471
2040 p->reclaim_state = NULL; 2472 p->reclaim_state = NULL;
2041 current->flags &= ~(PF_MEMALLOC | PF_ 2473 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2042 return nr_reclaimed >= nr_pages; !! 2474 return sc.nr_reclaimed >= nr_pages;
2043 } 2475 }
2044 2476
2045 int zone_reclaim(struct zone *zone, gfp_t gfp 2477 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2046 { 2478 {
2047 int node_id; 2479 int node_id;
2048 int ret; 2480 int ret;
2049 2481
2050 /* 2482 /*
2051 * Zone reclaim reclaims unmapped fil 2483 * Zone reclaim reclaims unmapped file backed pages and
2052 * slab pages if we are over the defi 2484 * slab pages if we are over the defined limits.
2053 * 2485 *
2054 * A small portion of unmapped file b 2486 * A small portion of unmapped file backed pages is needed for
2055 * file I/O otherwise pages read by f 2487 * file I/O otherwise pages read by file I/O will be immediately
2056 * thrown out if the zone is overallo 2488 * thrown out if the zone is overallocated. So we do not reclaim
2057 * if less than a specified percentag 2489 * if less than a specified percentage of the zone is used by
2058 * unmapped file backed pages. 2490 * unmapped file backed pages.
2059 */ 2491 */
2060 if (zone_page_state(zone, NR_FILE_PAG !! 2492 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2061 zone_page_state(zone, NR_FILE_MAP !! 2493 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
2062 && zone_page_state(zone, NR_SLAB_ !! 2494 return ZONE_RECLAIM_FULL;
2063 <= zone->min_slab_pag <<
2064 return 0; <<
2065 2495
2066 if (zone_is_all_unreclaimable(zone)) 2496 if (zone_is_all_unreclaimable(zone))
2067 return 0; !! 2497 return ZONE_RECLAIM_FULL;
2068 2498
2069 /* 2499 /*
2070 * Do not scan if the allocation shou 2500 * Do not scan if the allocation should not be delayed.
2071 */ 2501 */
2072 if (!(gfp_mask & __GFP_WAIT) || (curr 2502 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2073 return 0; !! 2503 return ZONE_RECLAIM_NOSCAN;
2074 2504
2075 /* 2505 /*
2076 * Only run zone reclaim on the local 2506 * Only run zone reclaim on the local zone or on zones that do not
2077 * have associated processors. This w 2507 * have associated processors. This will favor the local processor
2078 * over remote processors and spread 2508 * over remote processors and spread off node memory allocations
2079 * as wide as possible. 2509 * as wide as possible.
2080 */ 2510 */
2081 node_id = zone_to_nid(zone); 2511 node_id = zone_to_nid(zone);
2082 if (node_state(node_id, N_CPU) && nod 2512 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2083 return 0; !! 2513 return ZONE_RECLAIM_NOSCAN;
2084 2514
2085 if (zone_test_and_set_flag(zone, ZONE 2515 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2086 return 0; !! 2516 return ZONE_RECLAIM_NOSCAN;
>> 2517
2087 ret = __zone_reclaim(zone, gfp_mask, 2518 ret = __zone_reclaim(zone, gfp_mask, order);
2088 zone_clear_flag(zone, ZONE_RECLAIM_LO 2519 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2089 2520
>> 2521 if (!ret)
>> 2522 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
>> 2523
2090 return ret; 2524 return ret;
2091 } 2525 }
2092 #endif 2526 #endif
>> 2527
>> 2528 /*
>> 2529 * page_evictable - test whether a page is evictable
>> 2530 * @page: the page to test
>> 2531 * @vma: the VMA in which the page is or will be mapped, may be NULL
>> 2532 *
>> 2533 * Test whether page is evictable--i.e., should be placed on active/inactive
>> 2534 * lists vs unevictable list. The vma argument is !NULL when called from the
>> 2535 * fault path to determine how to instantate a new page.
>> 2536 *
>> 2537 * Reasons page might not be evictable:
>> 2538 * (1) page's mapping marked unevictable
>> 2539 * (2) page is part of an mlocked VMA
>> 2540 *
>> 2541 */
>> 2542 int page_evictable(struct page *page, struct vm_area_struct *vma)
>> 2543 {
>> 2544
>> 2545 if (mapping_unevictable(page_mapping(page)))
>> 2546 return 0;
>> 2547
>> 2548 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
>> 2549 return 0;
>> 2550
>> 2551 return 1;
>> 2552 }
>> 2553
>> 2554 /**
>> 2555 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
>> 2556 * @page: page to check evictability and move to appropriate lru list
>> 2557 * @zone: zone page is in
>> 2558 *
>> 2559 * Checks a page for evictability and moves the page to the appropriate
>> 2560 * zone lru list.
>> 2561 *
>> 2562 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
>> 2563 * have PageUnevictable set.
>> 2564 */
>> 2565 static void check_move_unevictable_page(struct page *page, struct zone *zone)
>> 2566 {
>> 2567 VM_BUG_ON(PageActive(page));
>> 2568
>> 2569 retry:
>> 2570 ClearPageUnevictable(page);
>> 2571 if (page_evictable(page, NULL)) {
>> 2572 enum lru_list l = LRU_INACTIVE_ANON + page_is_file_cache(page);
>> 2573
>> 2574 __dec_zone_state(zone, NR_UNEVICTABLE);
>> 2575 list_move(&page->lru, &zone->lru[l].list);
>> 2576 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
>> 2577 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
>> 2578 __count_vm_event(UNEVICTABLE_PGRESCUED);
>> 2579 } else {
>> 2580 /*
>> 2581 * rotate unevictable list
>> 2582 */
>> 2583 SetPageUnevictable(page);
>> 2584 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
>> 2585 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
>> 2586 if (page_evictable(page, NULL))
>> 2587 goto retry;
>> 2588 }
>> 2589 }
>> 2590
>> 2591 /**
>> 2592 * scan_mapping_unevictable_pages - scan an address space for evictable pages
>> 2593 * @mapping: struct address_space to scan for evictable pages
>> 2594 *
>> 2595 * Scan all pages in mapping. Check unevictable pages for
>> 2596 * evictability and move them to the appropriate zone lru list.
>> 2597 */
>> 2598 void scan_mapping_unevictable_pages(struct address_space *mapping)
>> 2599 {
>> 2600 pgoff_t next = 0;
>> 2601 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
>> 2602 PAGE_CACHE_SHIFT;
>> 2603 struct zone *zone;
>> 2604 struct pagevec pvec;
>> 2605
>> 2606 if (mapping->nrpages == 0)
>> 2607 return;
>> 2608
>> 2609 pagevec_init(&pvec, 0);
>> 2610 while (next < end &&
>> 2611 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
>> 2612 int i;
>> 2613 int pg_scanned = 0;
>> 2614
>> 2615 zone = NULL;
>> 2616
>> 2617 for (i = 0; i < pagevec_count(&pvec); i++) {
>> 2618 struct page *page = pvec.pages[i];
>> 2619 pgoff_t page_index = page->index;
>> 2620 struct zone *pagezone = page_zone(page);
>> 2621
>> 2622 pg_scanned++;
>> 2623 if (page_index > next)
>> 2624 next = page_index;
>> 2625 next++;
>> 2626
>> 2627 if (pagezone != zone) {
>> 2628 if (zone)
>> 2629 spin_unlock_irq(&zone->lru_lock);
>> 2630 zone = pagezone;
>> 2631 spin_lock_irq(&zone->lru_lock);
>> 2632 }
>> 2633
>> 2634 if (PageLRU(page) && PageUnevictable(page))
>> 2635 check_move_unevictable_page(page, zone);
>> 2636 }
>> 2637 if (zone)
>> 2638 spin_unlock_irq(&zone->lru_lock);
>> 2639 pagevec_release(&pvec);
>> 2640
>> 2641 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
>> 2642 }
>> 2643
>> 2644 }
>> 2645
>> 2646 /**
>> 2647 * scan_zone_unevictable_pages - check unevictable list for evictable pages
>> 2648 * @zone - zone of which to scan the unevictable list
>> 2649 *
>> 2650 * Scan @zone's unevictable LRU lists to check for pages that have become
>> 2651 * evictable. Move those that have to @zone's inactive list where they
>> 2652 * become candidates for reclaim, unless shrink_inactive_zone() decides
>> 2653 * to reactivate them. Pages that are still unevictable are rotated
>> 2654 * back onto @zone's unevictable list.
>> 2655 */
>> 2656 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
>> 2657 static void scan_zone_unevictable_pages(struct zone *zone)
>> 2658 {
>> 2659 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
>> 2660 unsigned long scan;
>> 2661 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
>> 2662
>> 2663 while (nr_to_scan > 0) {
>> 2664 unsigned long batch_size = min(nr_to_scan,
>> 2665 SCAN_UNEVICTABLE_BATCH_SIZE);
>> 2666
>> 2667 spin_lock_irq(&zone->lru_lock);
>> 2668 for (scan = 0; scan < batch_size; scan++) {
>> 2669 struct page *page = lru_to_page(l_unevictable);
>> 2670
>> 2671 if (!trylock_page(page))
>> 2672 continue;
>> 2673
>> 2674 prefetchw_prev_lru_page(page, l_unevictable, flags);
>> 2675
>> 2676 if (likely(PageLRU(page) && PageUnevictable(page)))
>> 2677 check_move_unevictable_page(page, zone);
>> 2678
>> 2679 unlock_page(page);
>> 2680 }
>> 2681 spin_unlock_irq(&zone->lru_lock);
>> 2682
>> 2683 nr_to_scan -= batch_size;
>> 2684 }
>> 2685 }
>> 2686
>> 2687
>> 2688 /**
>> 2689 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
>> 2690 *
>> 2691 * A really big hammer: scan all zones' unevictable LRU lists to check for
>> 2692 * pages that have become evictable. Move those back to the zones'
>> 2693 * inactive list where they become candidates for reclaim.
>> 2694 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
>> 2695 * and we add swap to the system. As such, it runs in the context of a task
>> 2696 * that has possibly/probably made some previously unevictable pages
>> 2697 * evictable.
>> 2698 */
>> 2699 static void scan_all_zones_unevictable_pages(void)
>> 2700 {
>> 2701 struct zone *zone;
>> 2702
>> 2703 for_each_zone(zone) {
>> 2704 scan_zone_unevictable_pages(zone);
>> 2705 }
>> 2706 }
>> 2707
>> 2708 /*
>> 2709 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
>> 2710 * all nodes' unevictable lists for evictable pages
>> 2711 */
>> 2712 unsigned long scan_unevictable_pages;
>> 2713
>> 2714 int scan_unevictable_handler(struct ctl_table *table, int write,
>> 2715 struct file *file, void __user *buffer,
>> 2716 size_t *length, loff_t *ppos)
>> 2717 {
>> 2718 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
>> 2719
>> 2720 if (write && *(unsigned long *)table->data)
>> 2721 scan_all_zones_unevictable_pages();
>> 2722
>> 2723 scan_unevictable_pages = 0;
>> 2724 return 0;
>> 2725 }
>> 2726
>> 2727 /*
>> 2728 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
>> 2729 * a specified node's per zone unevictable lists for evictable pages.
>> 2730 */
>> 2731
>> 2732 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
>> 2733 struct sysdev_attribute *attr,
>> 2734 char *buf)
>> 2735 {
>> 2736 return sprintf(buf, "\n"); /* always zero; should fit... */
>> 2737 }
>> 2738
>> 2739 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
>> 2740 struct sysdev_attribute *attr,
>> 2741 const char *buf, size_t count)
>> 2742 {
>> 2743 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
>> 2744 struct zone *zone;
>> 2745 unsigned long res;
>> 2746 unsigned long req = strict_strtoul(buf, 10, &res);
>> 2747
>> 2748 if (!req)
>> 2749 return 1; /* zero is no-op */
>> 2750
>> 2751 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
>> 2752 if (!populated_zone(zone))
>> 2753 continue;
>> 2754 scan_zone_unevictable_pages(zone);
>> 2755 }
>> 2756 return 1;
>> 2757 }
>> 2758
>> 2759
>> 2760 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
>> 2761 read_scan_unevictable_node,
>> 2762 write_scan_unevictable_node);
>> 2763
>> 2764 int scan_unevictable_register_node(struct node *node)
>> 2765 {
>> 2766 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
>> 2767 }
>> 2768
>> 2769 void scan_unevictable_unregister_node(struct node *node)
>> 2770 {
>> 2771 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
>> 2772 }
>> 2773
2093 2774
|
This page was automatically generated by the
LXR engine.
|