aboutsummaryrefslogtreecommitdiffstats
path: root/mm/memcontrol.c
diff options
context:
space:
mode:
Diffstat (limited to 'mm/memcontrol.c')
-rw-r--r--mm/memcontrol.c1386
1 files changed, 1163 insertions, 223 deletions
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index d813823ab08f..7973b5221fb8 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -6,6 +6,10 @@
6 * Copyright 2007 OpenVZ SWsoft Inc 6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org> 7 * Author: Pavel Emelianov <xemul@openvz.org>
8 * 8 *
9 * Memory thresholds
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
12 *
9 * This program is free software; you can redistribute it and/or modify 13 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of the GNU General Public License as published by 14 * it under the terms of the GNU General Public License as published by
11 * the Free Software Foundation; either version 2 of the License, or 15 * the Free Software Foundation; either version 2 of the License, or
@@ -21,6 +25,7 @@
21#include <linux/memcontrol.h> 25#include <linux/memcontrol.h>
22#include <linux/cgroup.h> 26#include <linux/cgroup.h>
23#include <linux/mm.h> 27#include <linux/mm.h>
28#include <linux/hugetlb.h>
24#include <linux/pagemap.h> 29#include <linux/pagemap.h>
25#include <linux/smp.h> 30#include <linux/smp.h>
26#include <linux/page-flags.h> 31#include <linux/page-flags.h>
@@ -32,7 +37,10 @@
32#include <linux/rbtree.h> 37#include <linux/rbtree.h>
33#include <linux/slab.h> 38#include <linux/slab.h>
34#include <linux/swap.h> 39#include <linux/swap.h>
40#include <linux/swapops.h>
35#include <linux/spinlock.h> 41#include <linux/spinlock.h>
42#include <linux/eventfd.h>
43#include <linux/sort.h>
36#include <linux/fs.h> 44#include <linux/fs.h>
37#include <linux/seq_file.h> 45#include <linux/seq_file.h>
38#include <linux/vmalloc.h> 46#include <linux/vmalloc.h>
@@ -55,7 +63,15 @@ static int really_do_swap_account __initdata = 1; /* for remember boot option*/
55#define do_swap_account (0) 63#define do_swap_account (0)
56#endif 64#endif
57 65
58#define SOFTLIMIT_EVENTS_THRESH (1000) 66/*
67 * Per memcg event counter is incremented at every pagein/pageout. This counter
68 * is used for trigger some periodic events. This is straightforward and better
69 * than using jiffies etc. to handle periodic memcg event.
70 *
71 * These values will be used as !((event) & ((1 <<(thresh)) - 1))
72 */
73#define THRESHOLDS_EVENTS_THRESH (7) /* once in 128 */
74#define SOFTLIMIT_EVENTS_THRESH (10) /* once in 1024 */
59 75
60/* 76/*
61 * Statistics for memory cgroup. 77 * Statistics for memory cgroup.
@@ -69,62 +85,16 @@ enum mem_cgroup_stat_index {
69 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */ 85 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
70 MEM_CGROUP_STAT_PGPGIN_COUNT, /* # of pages paged in */ 86 MEM_CGROUP_STAT_PGPGIN_COUNT, /* # of pages paged in */
71 MEM_CGROUP_STAT_PGPGOUT_COUNT, /* # of pages paged out */ 87 MEM_CGROUP_STAT_PGPGOUT_COUNT, /* # of pages paged out */
72 MEM_CGROUP_STAT_EVENTS, /* sum of pagein + pageout for internal use */
73 MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */ 88 MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
89 MEM_CGROUP_EVENTS, /* incremented at every pagein/pageout */
74 90
75 MEM_CGROUP_STAT_NSTATS, 91 MEM_CGROUP_STAT_NSTATS,
76}; 92};
77 93
78struct mem_cgroup_stat_cpu { 94struct mem_cgroup_stat_cpu {
79 s64 count[MEM_CGROUP_STAT_NSTATS]; 95 s64 count[MEM_CGROUP_STAT_NSTATS];
80} ____cacheline_aligned_in_smp;
81
82struct mem_cgroup_stat {
83 struct mem_cgroup_stat_cpu cpustat[0];
84}; 96};
85 97
86static inline void
87__mem_cgroup_stat_reset_safe(struct mem_cgroup_stat_cpu *stat,
88 enum mem_cgroup_stat_index idx)
89{
90 stat->count[idx] = 0;
91}
92
93static inline s64
94__mem_cgroup_stat_read_local(struct mem_cgroup_stat_cpu *stat,
95 enum mem_cgroup_stat_index idx)
96{
97 return stat->count[idx];
98}
99
100/*
101 * For accounting under irq disable, no need for increment preempt count.
102 */
103static inline void __mem_cgroup_stat_add_safe(struct mem_cgroup_stat_cpu *stat,
104 enum mem_cgroup_stat_index idx, int val)
105{
106 stat->count[idx] += val;
107}
108
109static s64 mem_cgroup_read_stat(struct mem_cgroup_stat *stat,
110 enum mem_cgroup_stat_index idx)
111{
112 int cpu;
113 s64 ret = 0;
114 for_each_possible_cpu(cpu)
115 ret += stat->cpustat[cpu].count[idx];
116 return ret;
117}
118
119static s64 mem_cgroup_local_usage(struct mem_cgroup_stat *stat)
120{
121 s64 ret;
122
123 ret = mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_CACHE);
124 ret += mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_RSS);
125 return ret;
126}
127
128/* 98/*
129 * per-zone information in memory controller. 99 * per-zone information in memory controller.
130 */ 100 */
@@ -174,6 +144,22 @@ struct mem_cgroup_tree {
174 144
175static struct mem_cgroup_tree soft_limit_tree __read_mostly; 145static struct mem_cgroup_tree soft_limit_tree __read_mostly;
176 146
147struct mem_cgroup_threshold {
148 struct eventfd_ctx *eventfd;
149 u64 threshold;
150};
151
152struct mem_cgroup_threshold_ary {
153 /* An array index points to threshold just below usage. */
154 atomic_t current_threshold;
155 /* Size of entries[] */
156 unsigned int size;
157 /* Array of thresholds */
158 struct mem_cgroup_threshold entries[0];
159};
160
161static void mem_cgroup_threshold(struct mem_cgroup *mem);
162
177/* 163/*
178 * The memory controller data structure. The memory controller controls both 164 * The memory controller data structure. The memory controller controls both
179 * page cache and RSS per cgroup. We would eventually like to provide 165 * page cache and RSS per cgroup. We would eventually like to provide
@@ -217,7 +203,7 @@ struct mem_cgroup {
217 * Should the accounting and control be hierarchical, per subtree? 203 * Should the accounting and control be hierarchical, per subtree?
218 */ 204 */
219 bool use_hierarchy; 205 bool use_hierarchy;
220 unsigned long last_oom_jiffies; 206 atomic_t oom_lock;
221 atomic_t refcnt; 207 atomic_t refcnt;
222 208
223 unsigned int swappiness; 209 unsigned int swappiness;
@@ -225,10 +211,48 @@ struct mem_cgroup {
225 /* set when res.limit == memsw.limit */ 211 /* set when res.limit == memsw.limit */
226 bool memsw_is_minimum; 212 bool memsw_is_minimum;
227 213
214 /* protect arrays of thresholds */
215 struct mutex thresholds_lock;
216
217 /* thresholds for memory usage. RCU-protected */
218 struct mem_cgroup_threshold_ary *thresholds;
219
220 /* thresholds for mem+swap usage. RCU-protected */
221 struct mem_cgroup_threshold_ary *memsw_thresholds;
222
228 /* 223 /*
229 * statistics. This must be placed at the end of memcg. 224 * Should we move charges of a task when a task is moved into this
225 * mem_cgroup ? And what type of charges should we move ?
230 */ 226 */
231 struct mem_cgroup_stat stat; 227 unsigned long move_charge_at_immigrate;
228
229 /*
230 * percpu counter.
231 */
232 struct mem_cgroup_stat_cpu *stat;
233};
234
235/* Stuffs for move charges at task migration. */
236/*
237 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
238 * left-shifted bitmap of these types.
239 */
240enum move_type {
241 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
242 NR_MOVE_TYPE,
243};
244
245/* "mc" and its members are protected by cgroup_mutex */
246static struct move_charge_struct {
247 struct mem_cgroup *from;
248 struct mem_cgroup *to;
249 unsigned long precharge;
250 unsigned long moved_charge;
251 unsigned long moved_swap;
252 struct task_struct *moving_task; /* a task moving charges */
253 wait_queue_head_t waitq; /* a waitq for other context */
254} mc = {
255 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
232}; 256};
233 257
234/* 258/*
@@ -371,23 +395,6 @@ mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
371 spin_unlock(&mctz->lock); 395 spin_unlock(&mctz->lock);
372} 396}
373 397
374static bool mem_cgroup_soft_limit_check(struct mem_cgroup *mem)
375{
376 bool ret = false;
377 int cpu;
378 s64 val;
379 struct mem_cgroup_stat_cpu *cpustat;
380
381 cpu = get_cpu();
382 cpustat = &mem->stat.cpustat[cpu];
383 val = __mem_cgroup_stat_read_local(cpustat, MEM_CGROUP_STAT_EVENTS);
384 if (unlikely(val > SOFTLIMIT_EVENTS_THRESH)) {
385 __mem_cgroup_stat_reset_safe(cpustat, MEM_CGROUP_STAT_EVENTS);
386 ret = true;
387 }
388 put_cpu();
389 return ret;
390}
391 398
392static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page) 399static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
393{ 400{
@@ -481,17 +488,31 @@ mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
481 return mz; 488 return mz;
482} 489}
483 490
491static s64 mem_cgroup_read_stat(struct mem_cgroup *mem,
492 enum mem_cgroup_stat_index idx)
493{
494 int cpu;
495 s64 val = 0;
496
497 for_each_possible_cpu(cpu)
498 val += per_cpu(mem->stat->count[idx], cpu);
499 return val;
500}
501
502static s64 mem_cgroup_local_usage(struct mem_cgroup *mem)
503{
504 s64 ret;
505
506 ret = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
507 ret += mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
508 return ret;
509}
510
484static void mem_cgroup_swap_statistics(struct mem_cgroup *mem, 511static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
485 bool charge) 512 bool charge)
486{ 513{
487 int val = (charge) ? 1 : -1; 514 int val = (charge) ? 1 : -1;
488 struct mem_cgroup_stat *stat = &mem->stat; 515 this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
489 struct mem_cgroup_stat_cpu *cpustat;
490 int cpu = get_cpu();
491
492 cpustat = &stat->cpustat[cpu];
493 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_SWAPOUT, val);
494 put_cpu();
495} 516}
496 517
497static void mem_cgroup_charge_statistics(struct mem_cgroup *mem, 518static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
@@ -499,24 +520,21 @@ static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
499 bool charge) 520 bool charge)
500{ 521{
501 int val = (charge) ? 1 : -1; 522 int val = (charge) ? 1 : -1;
502 struct mem_cgroup_stat *stat = &mem->stat;
503 struct mem_cgroup_stat_cpu *cpustat;
504 int cpu = get_cpu();
505 523
506 cpustat = &stat->cpustat[cpu]; 524 preempt_disable();
525
507 if (PageCgroupCache(pc)) 526 if (PageCgroupCache(pc))
508 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_CACHE, val); 527 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], val);
509 else 528 else
510 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_RSS, val); 529 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], val);
511 530
512 if (charge) 531 if (charge)
513 __mem_cgroup_stat_add_safe(cpustat, 532 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGIN_COUNT]);
514 MEM_CGROUP_STAT_PGPGIN_COUNT, 1);
515 else 533 else
516 __mem_cgroup_stat_add_safe(cpustat, 534 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGOUT_COUNT]);
517 MEM_CGROUP_STAT_PGPGOUT_COUNT, 1); 535 __this_cpu_inc(mem->stat->count[MEM_CGROUP_EVENTS]);
518 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_EVENTS, 1); 536
519 put_cpu(); 537 preempt_enable();
520} 538}
521 539
522static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem, 540static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
@@ -534,6 +552,29 @@ static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
534 return total; 552 return total;
535} 553}
536 554
555static bool __memcg_event_check(struct mem_cgroup *mem, int event_mask_shift)
556{
557 s64 val;
558
559 val = this_cpu_read(mem->stat->count[MEM_CGROUP_EVENTS]);
560
561 return !(val & ((1 << event_mask_shift) - 1));
562}
563
564/*
565 * Check events in order.
566 *
567 */
568static void memcg_check_events(struct mem_cgroup *mem, struct page *page)
569{
570 /* threshold event is triggered in finer grain than soft limit */
571 if (unlikely(__memcg_event_check(mem, THRESHOLDS_EVENTS_THRESH))) {
572 mem_cgroup_threshold(mem);
573 if (unlikely(__memcg_event_check(mem, SOFTLIMIT_EVENTS_THRESH)))
574 mem_cgroup_update_tree(mem, page);
575 }
576}
577
537static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont) 578static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
538{ 579{
539 return container_of(cgroup_subsys_state(cont, 580 return container_of(cgroup_subsys_state(cont,
@@ -1000,7 +1041,7 @@ static int mem_cgroup_count_children_cb(struct mem_cgroup *mem, void *data)
1000} 1041}
1001 1042
1002/** 1043/**
1003 * mem_cgroup_print_mem_info: Called from OOM with tasklist_lock held in read mode. 1044 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1004 * @memcg: The memory cgroup that went over limit 1045 * @memcg: The memory cgroup that went over limit
1005 * @p: Task that is going to be killed 1046 * @p: Task that is going to be killed
1006 * 1047 *
@@ -1174,7 +1215,7 @@ static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1174 } 1215 }
1175 } 1216 }
1176 } 1217 }
1177 if (!mem_cgroup_local_usage(&victim->stat)) { 1218 if (!mem_cgroup_local_usage(victim)) {
1178 /* this cgroup's local usage == 0 */ 1219 /* this cgroup's local usage == 0 */
1179 css_put(&victim->css); 1220 css_put(&victim->css);
1180 continue; 1221 continue;
@@ -1205,32 +1246,102 @@ static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1205 return total; 1246 return total;
1206} 1247}
1207 1248
1208bool mem_cgroup_oom_called(struct task_struct *task) 1249static int mem_cgroup_oom_lock_cb(struct mem_cgroup *mem, void *data)
1209{ 1250{
1210 bool ret = false; 1251 int *val = (int *)data;
1211 struct mem_cgroup *mem; 1252 int x;
1212 struct mm_struct *mm; 1253 /*
1254 * Logically, we can stop scanning immediately when we find
1255 * a memcg is already locked. But condidering unlock ops and
1256 * creation/removal of memcg, scan-all is simple operation.
1257 */
1258 x = atomic_inc_return(&mem->oom_lock);
1259 *val = max(x, *val);
1260 return 0;
1261}
1262/*
1263 * Check OOM-Killer is already running under our hierarchy.
1264 * If someone is running, return false.
1265 */
1266static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1267{
1268 int lock_count = 0;
1213 1269
1214 rcu_read_lock(); 1270 mem_cgroup_walk_tree(mem, &lock_count, mem_cgroup_oom_lock_cb);
1215 mm = task->mm; 1271
1216 if (!mm) 1272 if (lock_count == 1)
1217 mm = &init_mm; 1273 return true;
1218 mem = mem_cgroup_from_task(rcu_dereference(mm->owner)); 1274 return false;
1219 if (mem && time_before(jiffies, mem->last_oom_jiffies + HZ/10))
1220 ret = true;
1221 rcu_read_unlock();
1222 return ret;
1223} 1275}
1224 1276
1225static int record_last_oom_cb(struct mem_cgroup *mem, void *data) 1277static int mem_cgroup_oom_unlock_cb(struct mem_cgroup *mem, void *data)
1226{ 1278{
1227 mem->last_oom_jiffies = jiffies; 1279 /*
1280 * When a new child is created while the hierarchy is under oom,
1281 * mem_cgroup_oom_lock() may not be called. We have to use
1282 * atomic_add_unless() here.
1283 */
1284 atomic_add_unless(&mem->oom_lock, -1, 0);
1228 return 0; 1285 return 0;
1229} 1286}
1230 1287
1231static void record_last_oom(struct mem_cgroup *mem) 1288static void mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1232{ 1289{
1233 mem_cgroup_walk_tree(mem, NULL, record_last_oom_cb); 1290 mem_cgroup_walk_tree(mem, NULL, mem_cgroup_oom_unlock_cb);
1291}
1292
1293static DEFINE_MUTEX(memcg_oom_mutex);
1294static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1295
1296/*
1297 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1298 */
1299bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1300{
1301 DEFINE_WAIT(wait);
1302 bool locked;
1303
1304 /* At first, try to OOM lock hierarchy under mem.*/
1305 mutex_lock(&memcg_oom_mutex);
1306 locked = mem_cgroup_oom_lock(mem);
1307 /*
1308 * Even if signal_pending(), we can't quit charge() loop without
1309 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1310 * under OOM is always welcomed, use TASK_KILLABLE here.
1311 */
1312 if (!locked)
1313 prepare_to_wait(&memcg_oom_waitq, &wait, TASK_KILLABLE);
1314 mutex_unlock(&memcg_oom_mutex);
1315
1316 if (locked)
1317 mem_cgroup_out_of_memory(mem, mask);
1318 else {
1319 schedule();
1320 finish_wait(&memcg_oom_waitq, &wait);
1321 }
1322 mutex_lock(&memcg_oom_mutex);
1323 mem_cgroup_oom_unlock(mem);
1324 /*
1325 * Here, we use global waitq .....more fine grained waitq ?
1326 * Assume following hierarchy.
1327 * A/
1328 * 01
1329 * 02
1330 * assume OOM happens both in A and 01 at the same time. Tthey are
1331 * mutually exclusive by lock. (kill in 01 helps A.)
1332 * When we use per memcg waitq, we have to wake up waiters on A and 02
1333 * in addtion to waiters on 01. We use global waitq for avoiding mess.
1334 * It will not be a big problem.
1335 * (And a task may be moved to other groups while it's waiting for OOM.)
1336 */
1337 wake_up_all(&memcg_oom_waitq);
1338 mutex_unlock(&memcg_oom_mutex);
1339
1340 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1341 return false;
1342 /* Give chance to dying process */
1343 schedule_timeout(1);
1344 return true;
1234} 1345}
1235 1346
1236/* 1347/*
@@ -1240,9 +1351,6 @@ static void record_last_oom(struct mem_cgroup *mem)
1240void mem_cgroup_update_file_mapped(struct page *page, int val) 1351void mem_cgroup_update_file_mapped(struct page *page, int val)
1241{ 1352{
1242 struct mem_cgroup *mem; 1353 struct mem_cgroup *mem;
1243 struct mem_cgroup_stat *stat;
1244 struct mem_cgroup_stat_cpu *cpustat;
1245 int cpu;
1246 struct page_cgroup *pc; 1354 struct page_cgroup *pc;
1247 1355
1248 pc = lookup_page_cgroup(page); 1356 pc = lookup_page_cgroup(page);
@@ -1258,13 +1366,10 @@ void mem_cgroup_update_file_mapped(struct page *page, int val)
1258 goto done; 1366 goto done;
1259 1367
1260 /* 1368 /*
1261 * Preemption is already disabled, we don't need get_cpu() 1369 * Preemption is already disabled. We can use __this_cpu_xxx
1262 */ 1370 */
1263 cpu = smp_processor_id(); 1371 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_FILE_MAPPED], val);
1264 stat = &mem->stat;
1265 cpustat = &stat->cpustat[cpu];
1266 1372
1267 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_FILE_MAPPED, val);
1268done: 1373done:
1269 unlock_page_cgroup(pc); 1374 unlock_page_cgroup(pc);
1270} 1375}
@@ -1401,19 +1506,21 @@ static int __cpuinit memcg_stock_cpu_callback(struct notifier_block *nb,
1401 * oom-killer can be invoked. 1506 * oom-killer can be invoked.
1402 */ 1507 */
1403static int __mem_cgroup_try_charge(struct mm_struct *mm, 1508static int __mem_cgroup_try_charge(struct mm_struct *mm,
1404 gfp_t gfp_mask, struct mem_cgroup **memcg, 1509 gfp_t gfp_mask, struct mem_cgroup **memcg, bool oom)
1405 bool oom, struct page *page)
1406{ 1510{
1407 struct mem_cgroup *mem, *mem_over_limit; 1511 struct mem_cgroup *mem, *mem_over_limit;
1408 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES; 1512 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1409 struct res_counter *fail_res; 1513 struct res_counter *fail_res;
1410 int csize = CHARGE_SIZE; 1514 int csize = CHARGE_SIZE;
1411 1515
1412 if (unlikely(test_thread_flag(TIF_MEMDIE))) { 1516 /*
1413 /* Don't account this! */ 1517 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
1414 *memcg = NULL; 1518 * in system level. So, allow to go ahead dying process in addition to
1415 return 0; 1519 * MEMDIE process.
1416 } 1520 */
1521 if (unlikely(test_thread_flag(TIF_MEMDIE)
1522 || fatal_signal_pending(current)))
1523 goto bypass;
1417 1524
1418 /* 1525 /*
1419 * We always charge the cgroup the mm_struct belongs to. 1526 * We always charge the cgroup the mm_struct belongs to.
@@ -1440,7 +1547,7 @@ static int __mem_cgroup_try_charge(struct mm_struct *mm,
1440 unsigned long flags = 0; 1547 unsigned long flags = 0;
1441 1548
1442 if (consume_stock(mem)) 1549 if (consume_stock(mem))
1443 goto charged; 1550 goto done;
1444 1551
1445 ret = res_counter_charge(&mem->res, csize, &fail_res); 1552 ret = res_counter_charge(&mem->res, csize, &fail_res);
1446 if (likely(!ret)) { 1553 if (likely(!ret)) {
@@ -1483,28 +1590,70 @@ static int __mem_cgroup_try_charge(struct mm_struct *mm,
1483 if (mem_cgroup_check_under_limit(mem_over_limit)) 1590 if (mem_cgroup_check_under_limit(mem_over_limit))
1484 continue; 1591 continue;
1485 1592
1593 /* try to avoid oom while someone is moving charge */
1594 if (mc.moving_task && current != mc.moving_task) {
1595 struct mem_cgroup *from, *to;
1596 bool do_continue = false;
1597 /*
1598 * There is a small race that "from" or "to" can be
1599 * freed by rmdir, so we use css_tryget().
1600 */
1601 rcu_read_lock();
1602 from = mc.from;
1603 to = mc.to;
1604 if (from && css_tryget(&from->css)) {
1605 if (mem_over_limit->use_hierarchy)
1606 do_continue = css_is_ancestor(
1607 &from->css,
1608 &mem_over_limit->css);
1609 else
1610 do_continue = (from == mem_over_limit);
1611 css_put(&from->css);
1612 }
1613 if (!do_continue && to && css_tryget(&to->css)) {
1614 if (mem_over_limit->use_hierarchy)
1615 do_continue = css_is_ancestor(
1616 &to->css,
1617 &mem_over_limit->css);
1618 else
1619 do_continue = (to == mem_over_limit);
1620 css_put(&to->css);
1621 }
1622 rcu_read_unlock();
1623 if (do_continue) {
1624 DEFINE_WAIT(wait);
1625 prepare_to_wait(&mc.waitq, &wait,
1626 TASK_INTERRUPTIBLE);
1627 /* moving charge context might have finished. */
1628 if (mc.moving_task)
1629 schedule();
1630 finish_wait(&mc.waitq, &wait);
1631 continue;
1632 }
1633 }
1634
1486 if (!nr_retries--) { 1635 if (!nr_retries--) {
1487 if (oom) { 1636 if (!oom)
1488 mem_cgroup_out_of_memory(mem_over_limit, gfp_mask); 1637 goto nomem;
1489 record_last_oom(mem_over_limit); 1638 if (mem_cgroup_handle_oom(mem_over_limit, gfp_mask)) {
1639 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1640 continue;
1490 } 1641 }
1491 goto nomem; 1642 /* When we reach here, current task is dying .*/
1643 css_put(&mem->css);
1644 goto bypass;
1492 } 1645 }
1493 } 1646 }
1494 if (csize > PAGE_SIZE) 1647 if (csize > PAGE_SIZE)
1495 refill_stock(mem, csize - PAGE_SIZE); 1648 refill_stock(mem, csize - PAGE_SIZE);
1496charged:
1497 /*
1498 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
1499 * if they exceeds softlimit.
1500 */
1501 if (mem_cgroup_soft_limit_check(mem))
1502 mem_cgroup_update_tree(mem, page);
1503done: 1649done:
1504 return 0; 1650 return 0;
1505nomem: 1651nomem:
1506 css_put(&mem->css); 1652 css_put(&mem->css);
1507 return -ENOMEM; 1653 return -ENOMEM;
1654bypass:
1655 *memcg = NULL;
1656 return 0;
1508} 1657}
1509 1658
1510/* 1659/*
@@ -1512,14 +1661,23 @@ nomem:
1512 * This function is for that and do uncharge, put css's refcnt. 1661 * This function is for that and do uncharge, put css's refcnt.
1513 * gotten by try_charge(). 1662 * gotten by try_charge().
1514 */ 1663 */
1515static void mem_cgroup_cancel_charge(struct mem_cgroup *mem) 1664static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
1665 unsigned long count)
1516{ 1666{
1517 if (!mem_cgroup_is_root(mem)) { 1667 if (!mem_cgroup_is_root(mem)) {
1518 res_counter_uncharge(&mem->res, PAGE_SIZE); 1668 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
1519 if (do_swap_account) 1669 if (do_swap_account)
1520 res_counter_uncharge(&mem->memsw, PAGE_SIZE); 1670 res_counter_uncharge(&mem->memsw, PAGE_SIZE * count);
1671 VM_BUG_ON(test_bit(CSS_ROOT, &mem->css.flags));
1672 WARN_ON_ONCE(count > INT_MAX);
1673 __css_put(&mem->css, (int)count);
1521 } 1674 }
1522 css_put(&mem->css); 1675 /* we don't need css_put for root */
1676}
1677
1678static void mem_cgroup_cancel_charge(struct mem_cgroup *mem)
1679{
1680 __mem_cgroup_cancel_charge(mem, 1);
1523} 1681}
1524 1682
1525/* 1683/*
@@ -1615,6 +1773,12 @@ static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
1615 mem_cgroup_charge_statistics(mem, pc, true); 1773 mem_cgroup_charge_statistics(mem, pc, true);
1616 1774
1617 unlock_page_cgroup(pc); 1775 unlock_page_cgroup(pc);
1776 /*
1777 * "charge_statistics" updated event counter. Then, check it.
1778 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
1779 * if they exceeds softlimit.
1780 */
1781 memcg_check_events(mem, pc->page);
1618} 1782}
1619 1783
1620/** 1784/**
@@ -1622,22 +1786,22 @@ static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
1622 * @pc: page_cgroup of the page. 1786 * @pc: page_cgroup of the page.
1623 * @from: mem_cgroup which the page is moved from. 1787 * @from: mem_cgroup which the page is moved from.
1624 * @to: mem_cgroup which the page is moved to. @from != @to. 1788 * @to: mem_cgroup which the page is moved to. @from != @to.
1789 * @uncharge: whether we should call uncharge and css_put against @from.
1625 * 1790 *
1626 * The caller must confirm following. 1791 * The caller must confirm following.
1627 * - page is not on LRU (isolate_page() is useful.) 1792 * - page is not on LRU (isolate_page() is useful.)
1628 * - the pc is locked, used, and ->mem_cgroup points to @from. 1793 * - the pc is locked, used, and ->mem_cgroup points to @from.
1629 * 1794 *
1630 * This function does "uncharge" from old cgroup but doesn't do "charge" to 1795 * This function doesn't do "charge" nor css_get to new cgroup. It should be
1631 * new cgroup. It should be done by a caller. 1796 * done by a caller(__mem_cgroup_try_charge would be usefull). If @uncharge is
1797 * true, this function does "uncharge" from old cgroup, but it doesn't if
1798 * @uncharge is false, so a caller should do "uncharge".
1632 */ 1799 */
1633 1800
1634static void __mem_cgroup_move_account(struct page_cgroup *pc, 1801static void __mem_cgroup_move_account(struct page_cgroup *pc,
1635 struct mem_cgroup *from, struct mem_cgroup *to) 1802 struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge)
1636{ 1803{
1637 struct page *page; 1804 struct page *page;
1638 int cpu;
1639 struct mem_cgroup_stat *stat;
1640 struct mem_cgroup_stat_cpu *cpustat;
1641 1805
1642 VM_BUG_ON(from == to); 1806 VM_BUG_ON(from == to);
1643 VM_BUG_ON(PageLRU(pc->page)); 1807 VM_BUG_ON(PageLRU(pc->page));
@@ -1645,38 +1809,28 @@ static void __mem_cgroup_move_account(struct page_cgroup *pc,
1645 VM_BUG_ON(!PageCgroupUsed(pc)); 1809 VM_BUG_ON(!PageCgroupUsed(pc));
1646 VM_BUG_ON(pc->mem_cgroup != from); 1810 VM_BUG_ON(pc->mem_cgroup != from);
1647 1811
1648 if (!mem_cgroup_is_root(from))
1649 res_counter_uncharge(&from->res, PAGE_SIZE);
1650 mem_cgroup_charge_statistics(from, pc, false);
1651
1652 page = pc->page; 1812 page = pc->page;
1653 if (page_mapped(page) && !PageAnon(page)) { 1813 if (page_mapped(page) && !PageAnon(page)) {
1654 cpu = smp_processor_id(); 1814 /* Update mapped_file data for mem_cgroup */
1655 /* Update mapped_file data for mem_cgroup "from" */ 1815 preempt_disable();
1656 stat = &from->stat; 1816 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
1657 cpustat = &stat->cpustat[cpu]; 1817 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
1658 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_FILE_MAPPED, 1818 preempt_enable();
1659 -1);
1660
1661 /* Update mapped_file data for mem_cgroup "to" */
1662 stat = &to->stat;
1663 cpustat = &stat->cpustat[cpu];
1664 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_FILE_MAPPED,
1665 1);
1666 } 1819 }
1820 mem_cgroup_charge_statistics(from, pc, false);
1821 if (uncharge)
1822 /* This is not "cancel", but cancel_charge does all we need. */
1823 mem_cgroup_cancel_charge(from);
1667 1824
1668 if (do_swap_account && !mem_cgroup_is_root(from)) 1825 /* caller should have done css_get */
1669 res_counter_uncharge(&from->memsw, PAGE_SIZE);
1670 css_put(&from->css);
1671
1672 css_get(&to->css);
1673 pc->mem_cgroup = to; 1826 pc->mem_cgroup = to;
1674 mem_cgroup_charge_statistics(to, pc, true); 1827 mem_cgroup_charge_statistics(to, pc, true);
1675 /* 1828 /*
1676 * We charges against "to" which may not have any tasks. Then, "to" 1829 * We charges against "to" which may not have any tasks. Then, "to"
1677 * can be under rmdir(). But in current implementation, caller of 1830 * can be under rmdir(). But in current implementation, caller of
1678 * this function is just force_empty() and it's garanteed that 1831 * this function is just force_empty() and move charge, so it's
1679 * "to" is never removed. So, we don't check rmdir status here. 1832 * garanteed that "to" is never removed. So, we don't check rmdir
1833 * status here.
1680 */ 1834 */
1681} 1835}
1682 1836
@@ -1685,15 +1839,20 @@ static void __mem_cgroup_move_account(struct page_cgroup *pc,
1685 * __mem_cgroup_move_account() 1839 * __mem_cgroup_move_account()
1686 */ 1840 */
1687static int mem_cgroup_move_account(struct page_cgroup *pc, 1841static int mem_cgroup_move_account(struct page_cgroup *pc,
1688 struct mem_cgroup *from, struct mem_cgroup *to) 1842 struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge)
1689{ 1843{
1690 int ret = -EINVAL; 1844 int ret = -EINVAL;
1691 lock_page_cgroup(pc); 1845 lock_page_cgroup(pc);
1692 if (PageCgroupUsed(pc) && pc->mem_cgroup == from) { 1846 if (PageCgroupUsed(pc) && pc->mem_cgroup == from) {
1693 __mem_cgroup_move_account(pc, from, to); 1847 __mem_cgroup_move_account(pc, from, to, uncharge);
1694 ret = 0; 1848 ret = 0;
1695 } 1849 }
1696 unlock_page_cgroup(pc); 1850 unlock_page_cgroup(pc);
1851 /*
1852 * check events
1853 */
1854 memcg_check_events(to, pc->page);
1855 memcg_check_events(from, pc->page);
1697 return ret; 1856 return ret;
1698} 1857}
1699 1858
@@ -1722,15 +1881,13 @@ static int mem_cgroup_move_parent(struct page_cgroup *pc,
1722 goto put; 1881 goto put;
1723 1882
1724 parent = mem_cgroup_from_cont(pcg); 1883 parent = mem_cgroup_from_cont(pcg);
1725 ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false, page); 1884 ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false);
1726 if (ret || !parent) 1885 if (ret || !parent)
1727 goto put_back; 1886 goto put_back;
1728 1887
1729 ret = mem_cgroup_move_account(pc, child, parent); 1888 ret = mem_cgroup_move_account(pc, child, parent, true);
1730 if (!ret) 1889 if (ret)
1731 css_put(&parent->css); /* drop extra refcnt by try_charge() */ 1890 mem_cgroup_cancel_charge(parent);
1732 else
1733 mem_cgroup_cancel_charge(parent); /* does css_put */
1734put_back: 1891put_back:
1735 putback_lru_page(page); 1892 putback_lru_page(page);
1736put: 1893put:
@@ -1760,7 +1917,7 @@ static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
1760 prefetchw(pc); 1917 prefetchw(pc);
1761 1918
1762 mem = memcg; 1919 mem = memcg;
1763 ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true, page); 1920 ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true);
1764 if (ret || !mem) 1921 if (ret || !mem)
1765 return ret; 1922 return ret;
1766 1923
@@ -1880,14 +2037,14 @@ int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
1880 if (!mem) 2037 if (!mem)
1881 goto charge_cur_mm; 2038 goto charge_cur_mm;
1882 *ptr = mem; 2039 *ptr = mem;
1883 ret = __mem_cgroup_try_charge(NULL, mask, ptr, true, page); 2040 ret = __mem_cgroup_try_charge(NULL, mask, ptr, true);
1884 /* drop extra refcnt from tryget */ 2041 /* drop extra refcnt from tryget */
1885 css_put(&mem->css); 2042 css_put(&mem->css);
1886 return ret; 2043 return ret;
1887charge_cur_mm: 2044charge_cur_mm:
1888 if (unlikely(!mm)) 2045 if (unlikely(!mm))
1889 mm = &init_mm; 2046 mm = &init_mm;
1890 return __mem_cgroup_try_charge(mm, mask, ptr, true, page); 2047 return __mem_cgroup_try_charge(mm, mask, ptr, true);
1891} 2048}
1892 2049
1893static void 2050static void
@@ -2064,8 +2221,7 @@ __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2064 mz = page_cgroup_zoneinfo(pc); 2221 mz = page_cgroup_zoneinfo(pc);
2065 unlock_page_cgroup(pc); 2222 unlock_page_cgroup(pc);
2066 2223
2067 if (mem_cgroup_soft_limit_check(mem)) 2224 memcg_check_events(mem, page);
2068 mem_cgroup_update_tree(mem, page);
2069 /* at swapout, this memcg will be accessed to record to swap */ 2225 /* at swapout, this memcg will be accessed to record to swap */
2070 if (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT) 2226 if (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2071 css_put(&mem->css); 2227 css_put(&mem->css);
@@ -2192,6 +2348,64 @@ void mem_cgroup_uncharge_swap(swp_entry_t ent)
2192 } 2348 }
2193 rcu_read_unlock(); 2349 rcu_read_unlock();
2194} 2350}
2351
2352/**
2353 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2354 * @entry: swap entry to be moved
2355 * @from: mem_cgroup which the entry is moved from
2356 * @to: mem_cgroup which the entry is moved to
2357 * @need_fixup: whether we should fixup res_counters and refcounts.
2358 *
2359 * It succeeds only when the swap_cgroup's record for this entry is the same
2360 * as the mem_cgroup's id of @from.
2361 *
2362 * Returns 0 on success, -EINVAL on failure.
2363 *
2364 * The caller must have charged to @to, IOW, called res_counter_charge() about
2365 * both res and memsw, and called css_get().
2366 */
2367static int mem_cgroup_move_swap_account(swp_entry_t entry,
2368 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2369{
2370 unsigned short old_id, new_id;
2371
2372 old_id = css_id(&from->css);
2373 new_id = css_id(&to->css);
2374
2375 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2376 mem_cgroup_swap_statistics(from, false);
2377 mem_cgroup_swap_statistics(to, true);
2378 /*
2379 * This function is only called from task migration context now.
2380 * It postpones res_counter and refcount handling till the end
2381 * of task migration(mem_cgroup_clear_mc()) for performance
2382 * improvement. But we cannot postpone mem_cgroup_get(to)
2383 * because if the process that has been moved to @to does
2384 * swap-in, the refcount of @to might be decreased to 0.
2385 */
2386 mem_cgroup_get(to);
2387 if (need_fixup) {
2388 if (!mem_cgroup_is_root(from))
2389 res_counter_uncharge(&from->memsw, PAGE_SIZE);
2390 mem_cgroup_put(from);
2391 /*
2392 * we charged both to->res and to->memsw, so we should
2393 * uncharge to->res.
2394 */
2395 if (!mem_cgroup_is_root(to))
2396 res_counter_uncharge(&to->res, PAGE_SIZE);
2397 css_put(&to->css);
2398 }
2399 return 0;
2400 }
2401 return -EINVAL;
2402}
2403#else
2404static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2405 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2406{
2407 return -EINVAL;
2408}
2195#endif 2409#endif
2196 2410
2197/* 2411/*
@@ -2216,8 +2430,7 @@ int mem_cgroup_prepare_migration(struct page *page, struct mem_cgroup **ptr)
2216 unlock_page_cgroup(pc); 2430 unlock_page_cgroup(pc);
2217 2431
2218 if (mem) { 2432 if (mem) {
2219 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false, 2433 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false);
2220 page);
2221 css_put(&mem->css); 2434 css_put(&mem->css);
2222 } 2435 }
2223 *ptr = mem; 2436 *ptr = mem;
@@ -2704,7 +2917,7 @@ static int
2704mem_cgroup_get_idx_stat(struct mem_cgroup *mem, void *data) 2917mem_cgroup_get_idx_stat(struct mem_cgroup *mem, void *data)
2705{ 2918{
2706 struct mem_cgroup_idx_data *d = data; 2919 struct mem_cgroup_idx_data *d = data;
2707 d->val += mem_cgroup_read_stat(&mem->stat, d->idx); 2920 d->val += mem_cgroup_read_stat(mem, d->idx);
2708 return 0; 2921 return 0;
2709} 2922}
2710 2923
@@ -2719,40 +2932,50 @@ mem_cgroup_get_recursive_idx_stat(struct mem_cgroup *mem,
2719 *val = d.val; 2932 *val = d.val;
2720} 2933}
2721 2934
2935static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
2936{
2937 u64 idx_val, val;
2938
2939 if (!mem_cgroup_is_root(mem)) {
2940 if (!swap)
2941 return res_counter_read_u64(&mem->res, RES_USAGE);
2942 else
2943 return res_counter_read_u64(&mem->memsw, RES_USAGE);
2944 }
2945
2946 mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_CACHE, &idx_val);
2947 val = idx_val;
2948 mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_RSS, &idx_val);
2949 val += idx_val;
2950
2951 if (swap) {
2952 mem_cgroup_get_recursive_idx_stat(mem,
2953 MEM_CGROUP_STAT_SWAPOUT, &idx_val);
2954 val += idx_val;
2955 }
2956
2957 return val << PAGE_SHIFT;
2958}
2959
2722static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft) 2960static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
2723{ 2961{
2724 struct mem_cgroup *mem = mem_cgroup_from_cont(cont); 2962 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2725 u64 idx_val, val; 2963 u64 val;
2726 int type, name; 2964 int type, name;
2727 2965
2728 type = MEMFILE_TYPE(cft->private); 2966 type = MEMFILE_TYPE(cft->private);
2729 name = MEMFILE_ATTR(cft->private); 2967 name = MEMFILE_ATTR(cft->private);
2730 switch (type) { 2968 switch (type) {
2731 case _MEM: 2969 case _MEM:
2732 if (name == RES_USAGE && mem_cgroup_is_root(mem)) { 2970 if (name == RES_USAGE)
2733 mem_cgroup_get_recursive_idx_stat(mem, 2971 val = mem_cgroup_usage(mem, false);
2734 MEM_CGROUP_STAT_CACHE, &idx_val); 2972 else
2735 val = idx_val;
2736 mem_cgroup_get_recursive_idx_stat(mem,
2737 MEM_CGROUP_STAT_RSS, &idx_val);
2738 val += idx_val;
2739 val <<= PAGE_SHIFT;
2740 } else
2741 val = res_counter_read_u64(&mem->res, name); 2973 val = res_counter_read_u64(&mem->res, name);
2742 break; 2974 break;
2743 case _MEMSWAP: 2975 case _MEMSWAP:
2744 if (name == RES_USAGE && mem_cgroup_is_root(mem)) { 2976 if (name == RES_USAGE)
2745 mem_cgroup_get_recursive_idx_stat(mem, 2977 val = mem_cgroup_usage(mem, true);
2746 MEM_CGROUP_STAT_CACHE, &idx_val); 2978 else
2747 val = idx_val;
2748 mem_cgroup_get_recursive_idx_stat(mem,
2749 MEM_CGROUP_STAT_RSS, &idx_val);
2750 val += idx_val;
2751 mem_cgroup_get_recursive_idx_stat(mem,
2752 MEM_CGROUP_STAT_SWAPOUT, &idx_val);
2753 val += idx_val;
2754 val <<= PAGE_SHIFT;
2755 } else
2756 val = res_counter_read_u64(&mem->memsw, name); 2979 val = res_counter_read_u64(&mem->memsw, name);
2757 break; 2980 break;
2758 default: 2981 default:
@@ -2865,6 +3088,39 @@ static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
2865 return 0; 3088 return 0;
2866} 3089}
2867 3090
3091static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3092 struct cftype *cft)
3093{
3094 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3095}
3096
3097#ifdef CONFIG_MMU
3098static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3099 struct cftype *cft, u64 val)
3100{
3101 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3102
3103 if (val >= (1 << NR_MOVE_TYPE))
3104 return -EINVAL;
3105 /*
3106 * We check this value several times in both in can_attach() and
3107 * attach(), so we need cgroup lock to prevent this value from being
3108 * inconsistent.
3109 */
3110 cgroup_lock();
3111 mem->move_charge_at_immigrate = val;
3112 cgroup_unlock();
3113
3114 return 0;
3115}
3116#else
3117static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3118 struct cftype *cft, u64 val)
3119{
3120 return -ENOSYS;
3121}
3122#endif
3123
2868 3124
2869/* For read statistics */ 3125/* For read statistics */
2870enum { 3126enum {
@@ -2910,18 +3166,18 @@ static int mem_cgroup_get_local_stat(struct mem_cgroup *mem, void *data)
2910 s64 val; 3166 s64 val;
2911 3167
2912 /* per cpu stat */ 3168 /* per cpu stat */
2913 val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_CACHE); 3169 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
2914 s->stat[MCS_CACHE] += val * PAGE_SIZE; 3170 s->stat[MCS_CACHE] += val * PAGE_SIZE;
2915 val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_RSS); 3171 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
2916 s->stat[MCS_RSS] += val * PAGE_SIZE; 3172 s->stat[MCS_RSS] += val * PAGE_SIZE;
2917 val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_FILE_MAPPED); 3173 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
2918 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE; 3174 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
2919 val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGIN_COUNT); 3175 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGIN_COUNT);
2920 s->stat[MCS_PGPGIN] += val; 3176 s->stat[MCS_PGPGIN] += val;
2921 val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGOUT_COUNT); 3177 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGOUT_COUNT);
2922 s->stat[MCS_PGPGOUT] += val; 3178 s->stat[MCS_PGPGOUT] += val;
2923 if (do_swap_account) { 3179 if (do_swap_account) {
2924 val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_SWAPOUT); 3180 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
2925 s->stat[MCS_SWAP] += val * PAGE_SIZE; 3181 s->stat[MCS_SWAP] += val * PAGE_SIZE;
2926 } 3182 }
2927 3183
@@ -3049,12 +3305,249 @@ static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
3049 return 0; 3305 return 0;
3050} 3306}
3051 3307
3308static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3309{
3310 struct mem_cgroup_threshold_ary *t;
3311 u64 usage;
3312 int i;
3313
3314 rcu_read_lock();
3315 if (!swap)
3316 t = rcu_dereference(memcg->thresholds);
3317 else
3318 t = rcu_dereference(memcg->memsw_thresholds);
3319
3320 if (!t)
3321 goto unlock;
3322
3323 usage = mem_cgroup_usage(memcg, swap);
3324
3325 /*
3326 * current_threshold points to threshold just below usage.
3327 * If it's not true, a threshold was crossed after last
3328 * call of __mem_cgroup_threshold().
3329 */
3330 i = atomic_read(&t->current_threshold);
3331
3332 /*
3333 * Iterate backward over array of thresholds starting from
3334 * current_threshold and check if a threshold is crossed.
3335 * If none of thresholds below usage is crossed, we read
3336 * only one element of the array here.
3337 */
3338 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3339 eventfd_signal(t->entries[i].eventfd, 1);
3340
3341 /* i = current_threshold + 1 */
3342 i++;
3343
3344 /*
3345 * Iterate forward over array of thresholds starting from
3346 * current_threshold+1 and check if a threshold is crossed.
3347 * If none of thresholds above usage is crossed, we read
3348 * only one element of the array here.
3349 */
3350 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3351 eventfd_signal(t->entries[i].eventfd, 1);
3352
3353 /* Update current_threshold */
3354 atomic_set(&t->current_threshold, i - 1);
3355unlock:
3356 rcu_read_unlock();
3357}
3358
3359static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3360{
3361 __mem_cgroup_threshold(memcg, false);
3362 if (do_swap_account)
3363 __mem_cgroup_threshold(memcg, true);
3364}
3365
3366static int compare_thresholds(const void *a, const void *b)
3367{
3368 const struct mem_cgroup_threshold *_a = a;
3369 const struct mem_cgroup_threshold *_b = b;
3370
3371 return _a->threshold - _b->threshold;
3372}
3373
3374static int mem_cgroup_register_event(struct cgroup *cgrp, struct cftype *cft,
3375 struct eventfd_ctx *eventfd, const char *args)
3376{
3377 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3378 struct mem_cgroup_threshold_ary *thresholds, *thresholds_new;
3379 int type = MEMFILE_TYPE(cft->private);
3380 u64 threshold, usage;
3381 int size;
3382 int i, ret;
3383
3384 ret = res_counter_memparse_write_strategy(args, &threshold);
3385 if (ret)
3386 return ret;
3387
3388 mutex_lock(&memcg->thresholds_lock);
3389 if (type == _MEM)
3390 thresholds = memcg->thresholds;
3391 else if (type == _MEMSWAP)
3392 thresholds = memcg->memsw_thresholds;
3393 else
3394 BUG();
3395
3396 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3397
3398 /* Check if a threshold crossed before adding a new one */
3399 if (thresholds)
3400 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3401
3402 if (thresholds)
3403 size = thresholds->size + 1;
3404 else
3405 size = 1;
3406
3407 /* Allocate memory for new array of thresholds */
3408 thresholds_new = kmalloc(sizeof(*thresholds_new) +
3409 size * sizeof(struct mem_cgroup_threshold),
3410 GFP_KERNEL);
3411 if (!thresholds_new) {
3412 ret = -ENOMEM;
3413 goto unlock;
3414 }
3415 thresholds_new->size = size;
3416
3417 /* Copy thresholds (if any) to new array */
3418 if (thresholds)
3419 memcpy(thresholds_new->entries, thresholds->entries,
3420 thresholds->size *
3421 sizeof(struct mem_cgroup_threshold));
3422 /* Add new threshold */
3423 thresholds_new->entries[size - 1].eventfd = eventfd;
3424 thresholds_new->entries[size - 1].threshold = threshold;
3425
3426 /* Sort thresholds. Registering of new threshold isn't time-critical */
3427 sort(thresholds_new->entries, size,
3428 sizeof(struct mem_cgroup_threshold),
3429 compare_thresholds, NULL);
3430
3431 /* Find current threshold */
3432 atomic_set(&thresholds_new->current_threshold, -1);
3433 for (i = 0; i < size; i++) {
3434 if (thresholds_new->entries[i].threshold < usage) {
3435 /*
3436 * thresholds_new->current_threshold will not be used
3437 * until rcu_assign_pointer(), so it's safe to increment
3438 * it here.
3439 */
3440 atomic_inc(&thresholds_new->current_threshold);
3441 }
3442 }
3443
3444 if (type == _MEM)
3445 rcu_assign_pointer(memcg->thresholds, thresholds_new);
3446 else
3447 rcu_assign_pointer(memcg->memsw_thresholds, thresholds_new);
3448
3449 /* To be sure that nobody uses thresholds before freeing it */
3450 synchronize_rcu();
3451
3452 kfree(thresholds);
3453unlock:
3454 mutex_unlock(&memcg->thresholds_lock);
3455
3456 return ret;
3457}
3458
3459static int mem_cgroup_unregister_event(struct cgroup *cgrp, struct cftype *cft,
3460 struct eventfd_ctx *eventfd)
3461{
3462 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3463 struct mem_cgroup_threshold_ary *thresholds, *thresholds_new;
3464 int type = MEMFILE_TYPE(cft->private);
3465 u64 usage;
3466 int size = 0;
3467 int i, j, ret;
3468
3469 mutex_lock(&memcg->thresholds_lock);
3470 if (type == _MEM)
3471 thresholds = memcg->thresholds;
3472 else if (type == _MEMSWAP)
3473 thresholds = memcg->memsw_thresholds;
3474 else
3475 BUG();
3476
3477 /*
3478 * Something went wrong if we trying to unregister a threshold
3479 * if we don't have thresholds
3480 */
3481 BUG_ON(!thresholds);
3482
3483 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3484
3485 /* Check if a threshold crossed before removing */
3486 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3487
3488 /* Calculate new number of threshold */
3489 for (i = 0; i < thresholds->size; i++) {
3490 if (thresholds->entries[i].eventfd != eventfd)
3491 size++;
3492 }
3493
3494 /* Set thresholds array to NULL if we don't have thresholds */
3495 if (!size) {
3496 thresholds_new = NULL;
3497 goto assign;
3498 }
3499
3500 /* Allocate memory for new array of thresholds */
3501 thresholds_new = kmalloc(sizeof(*thresholds_new) +
3502 size * sizeof(struct mem_cgroup_threshold),
3503 GFP_KERNEL);
3504 if (!thresholds_new) {
3505 ret = -ENOMEM;
3506 goto unlock;
3507 }
3508 thresholds_new->size = size;
3509
3510 /* Copy thresholds and find current threshold */
3511 atomic_set(&thresholds_new->current_threshold, -1);
3512 for (i = 0, j = 0; i < thresholds->size; i++) {
3513 if (thresholds->entries[i].eventfd == eventfd)
3514 continue;
3515
3516 thresholds_new->entries[j] = thresholds->entries[i];
3517 if (thresholds_new->entries[j].threshold < usage) {
3518 /*
3519 * thresholds_new->current_threshold will not be used
3520 * until rcu_assign_pointer(), so it's safe to increment
3521 * it here.
3522 */
3523 atomic_inc(&thresholds_new->current_threshold);
3524 }
3525 j++;
3526 }
3527
3528assign:
3529 if (type == _MEM)
3530 rcu_assign_pointer(memcg->thresholds, thresholds_new);
3531 else
3532 rcu_assign_pointer(memcg->memsw_thresholds, thresholds_new);
3533
3534 /* To be sure that nobody uses thresholds before freeing it */
3535 synchronize_rcu();
3536
3537 kfree(thresholds);
3538unlock:
3539 mutex_unlock(&memcg->thresholds_lock);
3540
3541 return ret;
3542}
3052 3543
3053static struct cftype mem_cgroup_files[] = { 3544static struct cftype mem_cgroup_files[] = {
3054 { 3545 {
3055 .name = "usage_in_bytes", 3546 .name = "usage_in_bytes",
3056 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE), 3547 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
3057 .read_u64 = mem_cgroup_read, 3548 .read_u64 = mem_cgroup_read,
3549 .register_event = mem_cgroup_register_event,
3550 .unregister_event = mem_cgroup_unregister_event,
3058 }, 3551 },
3059 { 3552 {
3060 .name = "max_usage_in_bytes", 3553 .name = "max_usage_in_bytes",
@@ -3098,6 +3591,11 @@ static struct cftype mem_cgroup_files[] = {
3098 .read_u64 = mem_cgroup_swappiness_read, 3591 .read_u64 = mem_cgroup_swappiness_read,
3099 .write_u64 = mem_cgroup_swappiness_write, 3592 .write_u64 = mem_cgroup_swappiness_write,
3100 }, 3593 },
3594 {
3595 .name = "move_charge_at_immigrate",
3596 .read_u64 = mem_cgroup_move_charge_read,
3597 .write_u64 = mem_cgroup_move_charge_write,
3598 },
3101}; 3599};
3102 3600
3103#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 3601#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
@@ -3106,6 +3604,8 @@ static struct cftype memsw_cgroup_files[] = {
3106 .name = "memsw.usage_in_bytes", 3604 .name = "memsw.usage_in_bytes",
3107 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE), 3605 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
3108 .read_u64 = mem_cgroup_read, 3606 .read_u64 = mem_cgroup_read,
3607 .register_event = mem_cgroup_register_event,
3608 .unregister_event = mem_cgroup_unregister_event,
3109 }, 3609 },
3110 { 3610 {
3111 .name = "memsw.max_usage_in_bytes", 3611 .name = "memsw.max_usage_in_bytes",
@@ -3180,17 +3680,12 @@ static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
3180 kfree(mem->info.nodeinfo[node]); 3680 kfree(mem->info.nodeinfo[node]);
3181} 3681}
3182 3682
3183static int mem_cgroup_size(void)
3184{
3185 int cpustat_size = nr_cpu_ids * sizeof(struct mem_cgroup_stat_cpu);
3186 return sizeof(struct mem_cgroup) + cpustat_size;
3187}
3188
3189static struct mem_cgroup *mem_cgroup_alloc(void) 3683static struct mem_cgroup *mem_cgroup_alloc(void)
3190{ 3684{
3191 struct mem_cgroup *mem; 3685 struct mem_cgroup *mem;
3192 int size = mem_cgroup_size(); 3686 int size = sizeof(struct mem_cgroup);
3193 3687
3688 /* Can be very big if MAX_NUMNODES is very big */
3194 if (size < PAGE_SIZE) 3689 if (size < PAGE_SIZE)
3195 mem = kmalloc(size, GFP_KERNEL); 3690 mem = kmalloc(size, GFP_KERNEL);
3196 else 3691 else
@@ -3198,6 +3693,14 @@ static struct mem_cgroup *mem_cgroup_alloc(void)
3198 3693
3199 if (mem) 3694 if (mem)
3200 memset(mem, 0, size); 3695 memset(mem, 0, size);
3696 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
3697 if (!mem->stat) {
3698 if (size < PAGE_SIZE)
3699 kfree(mem);
3700 else
3701 vfree(mem);
3702 mem = NULL;
3703 }
3201 return mem; 3704 return mem;
3202} 3705}
3203 3706
@@ -3222,7 +3725,8 @@ static void __mem_cgroup_free(struct mem_cgroup *mem)
3222 for_each_node_state(node, N_POSSIBLE) 3725 for_each_node_state(node, N_POSSIBLE)
3223 free_mem_cgroup_per_zone_info(mem, node); 3726 free_mem_cgroup_per_zone_info(mem, node);
3224 3727
3225 if (mem_cgroup_size() < PAGE_SIZE) 3728 free_percpu(mem->stat);
3729 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
3226 kfree(mem); 3730 kfree(mem);
3227 else 3731 else
3228 vfree(mem); 3732 vfree(mem);
@@ -3233,9 +3737,9 @@ static void mem_cgroup_get(struct mem_cgroup *mem)
3233 atomic_inc(&mem->refcnt); 3737 atomic_inc(&mem->refcnt);
3234} 3738}
3235 3739
3236static void mem_cgroup_put(struct mem_cgroup *mem) 3740static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
3237{ 3741{
3238 if (atomic_dec_and_test(&mem->refcnt)) { 3742 if (atomic_sub_and_test(count, &mem->refcnt)) {
3239 struct mem_cgroup *parent = parent_mem_cgroup(mem); 3743 struct mem_cgroup *parent = parent_mem_cgroup(mem);
3240 __mem_cgroup_free(mem); 3744 __mem_cgroup_free(mem);
3241 if (parent) 3745 if (parent)
@@ -3243,6 +3747,11 @@ static void mem_cgroup_put(struct mem_cgroup *mem)
3243 } 3747 }
3244} 3748}
3245 3749
3750static void mem_cgroup_put(struct mem_cgroup *mem)
3751{
3752 __mem_cgroup_put(mem, 1);
3753}
3754
3246/* 3755/*
3247 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled. 3756 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
3248 */ 3757 */
@@ -3319,7 +3828,6 @@ mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
3319 INIT_WORK(&stock->work, drain_local_stock); 3828 INIT_WORK(&stock->work, drain_local_stock);
3320 } 3829 }
3321 hotcpu_notifier(memcg_stock_cpu_callback, 0); 3830 hotcpu_notifier(memcg_stock_cpu_callback, 0);
3322
3323 } else { 3831 } else {
3324 parent = mem_cgroup_from_cont(cont->parent); 3832 parent = mem_cgroup_from_cont(cont->parent);
3325 mem->use_hierarchy = parent->use_hierarchy; 3833 mem->use_hierarchy = parent->use_hierarchy;
@@ -3345,6 +3853,8 @@ mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
3345 if (parent) 3853 if (parent)
3346 mem->swappiness = get_swappiness(parent); 3854 mem->swappiness = get_swappiness(parent);
3347 atomic_set(&mem->refcnt, 1); 3855 atomic_set(&mem->refcnt, 1);
3856 mem->move_charge_at_immigrate = 0;
3857 mutex_init(&mem->thresholds_lock);
3348 return &mem->css; 3858 return &mem->css;
3349free_out: 3859free_out:
3350 __mem_cgroup_free(mem); 3860 __mem_cgroup_free(mem);
@@ -3381,16 +3891,444 @@ static int mem_cgroup_populate(struct cgroup_subsys *ss,
3381 return ret; 3891 return ret;
3382} 3892}
3383 3893
3894#ifdef CONFIG_MMU
3895/* Handlers for move charge at task migration. */
3896#define PRECHARGE_COUNT_AT_ONCE 256
3897static int mem_cgroup_do_precharge(unsigned long count)
3898{
3899 int ret = 0;
3900 int batch_count = PRECHARGE_COUNT_AT_ONCE;
3901 struct mem_cgroup *mem = mc.to;
3902
3903 if (mem_cgroup_is_root(mem)) {
3904 mc.precharge += count;
3905 /* we don't need css_get for root */
3906 return ret;
3907 }
3908 /* try to charge at once */
3909 if (count > 1) {
3910 struct res_counter *dummy;
3911 /*
3912 * "mem" cannot be under rmdir() because we've already checked
3913 * by cgroup_lock_live_cgroup() that it is not removed and we
3914 * are still under the same cgroup_mutex. So we can postpone
3915 * css_get().
3916 */
3917 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
3918 goto one_by_one;
3919 if (do_swap_account && res_counter_charge(&mem->memsw,
3920 PAGE_SIZE * count, &dummy)) {
3921 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
3922 goto one_by_one;
3923 }
3924 mc.precharge += count;
3925 VM_BUG_ON(test_bit(CSS_ROOT, &mem->css.flags));
3926 WARN_ON_ONCE(count > INT_MAX);
3927 __css_get(&mem->css, (int)count);
3928 return ret;
3929 }
3930one_by_one:
3931 /* fall back to one by one charge */
3932 while (count--) {
3933 if (signal_pending(current)) {
3934 ret = -EINTR;
3935 break;
3936 }
3937 if (!batch_count--) {
3938 batch_count = PRECHARGE_COUNT_AT_ONCE;
3939 cond_resched();
3940 }
3941 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false);
3942 if (ret || !mem)
3943 /* mem_cgroup_clear_mc() will do uncharge later */
3944 return -ENOMEM;
3945 mc.precharge++;
3946 }
3947 return ret;
3948}
3949#else /* !CONFIG_MMU */
3950static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
3951 struct cgroup *cgroup,
3952 struct task_struct *p,
3953 bool threadgroup)
3954{
3955 return 0;
3956}
3957static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
3958 struct cgroup *cgroup,
3959 struct task_struct *p,
3960 bool threadgroup)
3961{
3962}
3384static void mem_cgroup_move_task(struct cgroup_subsys *ss, 3963static void mem_cgroup_move_task(struct cgroup_subsys *ss,
3385 struct cgroup *cont, 3964 struct cgroup *cont,
3386 struct cgroup *old_cont, 3965 struct cgroup *old_cont,
3387 struct task_struct *p, 3966 struct task_struct *p,
3388 bool threadgroup) 3967 bool threadgroup)
3389{ 3968{
3969}
3970#endif
3971
3972/**
3973 * is_target_pte_for_mc - check a pte whether it is valid for move charge
3974 * @vma: the vma the pte to be checked belongs
3975 * @addr: the address corresponding to the pte to be checked
3976 * @ptent: the pte to be checked
3977 * @target: the pointer the target page or swap ent will be stored(can be NULL)
3978 *
3979 * Returns
3980 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
3981 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
3982 * move charge. if @target is not NULL, the page is stored in target->page
3983 * with extra refcnt got(Callers should handle it).
3984 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
3985 * target for charge migration. if @target is not NULL, the entry is stored
3986 * in target->ent.
3987 *
3988 * Called with pte lock held.
3989 */
3990union mc_target {
3991 struct page *page;
3992 swp_entry_t ent;
3993};
3994
3995enum mc_target_type {
3996 MC_TARGET_NONE, /* not used */
3997 MC_TARGET_PAGE,
3998 MC_TARGET_SWAP,
3999};
4000
4001static int is_target_pte_for_mc(struct vm_area_struct *vma,
4002 unsigned long addr, pte_t ptent, union mc_target *target)
4003{
4004 struct page *page = NULL;
4005 struct page_cgroup *pc;
4006 int ret = 0;
4007 swp_entry_t ent = { .val = 0 };
4008 int usage_count = 0;
4009 bool move_anon = test_bit(MOVE_CHARGE_TYPE_ANON,
4010 &mc.to->move_charge_at_immigrate);
4011
4012 if (!pte_present(ptent)) {
4013 /* TODO: handle swap of shmes/tmpfs */
4014 if (pte_none(ptent) || pte_file(ptent))
4015 return 0;
4016 else if (is_swap_pte(ptent)) {
4017 ent = pte_to_swp_entry(ptent);
4018 if (!move_anon || non_swap_entry(ent))
4019 return 0;
4020 usage_count = mem_cgroup_count_swap_user(ent, &page);
4021 }
4022 } else {
4023 page = vm_normal_page(vma, addr, ptent);
4024 if (!page || !page_mapped(page))
4025 return 0;
4026 /*
4027 * TODO: We don't move charges of file(including shmem/tmpfs)
4028 * pages for now.
4029 */
4030 if (!move_anon || !PageAnon(page))
4031 return 0;
4032 if (!get_page_unless_zero(page))
4033 return 0;
4034 usage_count = page_mapcount(page);
4035 }
4036 if (usage_count > 1) {
4037 /*
4038 * TODO: We don't move charges of shared(used by multiple
4039 * processes) pages for now.
4040 */
4041 if (page)
4042 put_page(page);
4043 return 0;
4044 }
4045 if (page) {
4046 pc = lookup_page_cgroup(page);
4047 /*
4048 * Do only loose check w/o page_cgroup lock.
4049 * mem_cgroup_move_account() checks the pc is valid or not under
4050 * the lock.
4051 */
4052 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
4053 ret = MC_TARGET_PAGE;
4054 if (target)
4055 target->page = page;
4056 }
4057 if (!ret || !target)
4058 put_page(page);
4059 }
4060 /* throught */
4061 if (ent.val && do_swap_account && !ret &&
4062 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
4063 ret = MC_TARGET_SWAP;
4064 if (target)
4065 target->ent = ent;
4066 }
4067 return ret;
4068}
4069
4070static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4071 unsigned long addr, unsigned long end,
4072 struct mm_walk *walk)
4073{
4074 struct vm_area_struct *vma = walk->private;
4075 pte_t *pte;
4076 spinlock_t *ptl;
4077
4078 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4079 for (; addr != end; pte++, addr += PAGE_SIZE)
4080 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
4081 mc.precharge++; /* increment precharge temporarily */
4082 pte_unmap_unlock(pte - 1, ptl);
4083 cond_resched();
4084
4085 return 0;
4086}
4087
4088static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4089{
4090 unsigned long precharge;
4091 struct vm_area_struct *vma;
4092
4093 down_read(&mm->mmap_sem);
4094 for (vma = mm->mmap; vma; vma = vma->vm_next) {
4095 struct mm_walk mem_cgroup_count_precharge_walk = {
4096 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4097 .mm = mm,
4098 .private = vma,
4099 };
4100 if (is_vm_hugetlb_page(vma))
4101 continue;
4102 /* TODO: We don't move charges of shmem/tmpfs pages for now. */
4103 if (vma->vm_flags & VM_SHARED)
4104 continue;
4105 walk_page_range(vma->vm_start, vma->vm_end,
4106 &mem_cgroup_count_precharge_walk);
4107 }
4108 up_read(&mm->mmap_sem);
4109
4110 precharge = mc.precharge;
4111 mc.precharge = 0;
4112
4113 return precharge;
4114}
4115
4116static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4117{
4118 return mem_cgroup_do_precharge(mem_cgroup_count_precharge(mm));
4119}
4120
4121static void mem_cgroup_clear_mc(void)
4122{
4123 /* we must uncharge all the leftover precharges from mc.to */
4124 if (mc.precharge) {
4125 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
4126 mc.precharge = 0;
4127 }
3390 /* 4128 /*
3391 * FIXME: It's better to move charges of this process from old 4129 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
3392 * memcg to new memcg. But it's just on TODO-List now. 4130 * we must uncharge here.
3393 */ 4131 */
4132 if (mc.moved_charge) {
4133 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
4134 mc.moved_charge = 0;
4135 }
4136 /* we must fixup refcnts and charges */
4137 if (mc.moved_swap) {
4138 WARN_ON_ONCE(mc.moved_swap > INT_MAX);
4139 /* uncharge swap account from the old cgroup */
4140 if (!mem_cgroup_is_root(mc.from))
4141 res_counter_uncharge(&mc.from->memsw,
4142 PAGE_SIZE * mc.moved_swap);
4143 __mem_cgroup_put(mc.from, mc.moved_swap);
4144
4145 if (!mem_cgroup_is_root(mc.to)) {
4146 /*
4147 * we charged both to->res and to->memsw, so we should
4148 * uncharge to->res.
4149 */
4150 res_counter_uncharge(&mc.to->res,
4151 PAGE_SIZE * mc.moved_swap);
4152 VM_BUG_ON(test_bit(CSS_ROOT, &mc.to->css.flags));
4153 __css_put(&mc.to->css, mc.moved_swap);
4154 }
4155 /* we've already done mem_cgroup_get(mc.to) */
4156
4157 mc.moved_swap = 0;
4158 }
4159 mc.from = NULL;
4160 mc.to = NULL;
4161 mc.moving_task = NULL;
4162 wake_up_all(&mc.waitq);
4163}
4164
4165static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4166 struct cgroup *cgroup,
4167 struct task_struct *p,
4168 bool threadgroup)
4169{
4170 int ret = 0;
4171 struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
4172
4173 if (mem->move_charge_at_immigrate) {
4174 struct mm_struct *mm;
4175 struct mem_cgroup *from = mem_cgroup_from_task(p);
4176
4177 VM_BUG_ON(from == mem);
4178
4179 mm = get_task_mm(p);
4180 if (!mm)
4181 return 0;
4182 /* We move charges only when we move a owner of the mm */
4183 if (mm->owner == p) {
4184 VM_BUG_ON(mc.from);
4185 VM_BUG_ON(mc.to);
4186 VM_BUG_ON(mc.precharge);
4187 VM_BUG_ON(mc.moved_charge);
4188 VM_BUG_ON(mc.moved_swap);
4189 VM_BUG_ON(mc.moving_task);
4190 mc.from = from;
4191 mc.to = mem;
4192 mc.precharge = 0;
4193 mc.moved_charge = 0;
4194 mc.moved_swap = 0;
4195 mc.moving_task = current;
4196
4197 ret = mem_cgroup_precharge_mc(mm);
4198 if (ret)
4199 mem_cgroup_clear_mc();
4200 }
4201 mmput(mm);
4202 }
4203 return ret;
4204}
4205
4206static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4207 struct cgroup *cgroup,
4208 struct task_struct *p,
4209 bool threadgroup)
4210{
4211 mem_cgroup_clear_mc();
4212}
4213
4214static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4215 unsigned long addr, unsigned long end,
4216 struct mm_walk *walk)
4217{
4218 int ret = 0;
4219 struct vm_area_struct *vma = walk->private;
4220 pte_t *pte;
4221 spinlock_t *ptl;
4222
4223retry:
4224 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4225 for (; addr != end; addr += PAGE_SIZE) {
4226 pte_t ptent = *(pte++);
4227 union mc_target target;
4228 int type;
4229 struct page *page;
4230 struct page_cgroup *pc;
4231 swp_entry_t ent;
4232
4233 if (!mc.precharge)
4234 break;
4235
4236 type = is_target_pte_for_mc(vma, addr, ptent, &target);
4237 switch (type) {
4238 case MC_TARGET_PAGE:
4239 page = target.page;
4240 if (isolate_lru_page(page))
4241 goto put;
4242 pc = lookup_page_cgroup(page);
4243 if (!mem_cgroup_move_account(pc,
4244 mc.from, mc.to, false)) {
4245 mc.precharge--;
4246 /* we uncharge from mc.from later. */
4247 mc.moved_charge++;
4248 }
4249 putback_lru_page(page);
4250put: /* is_target_pte_for_mc() gets the page */
4251 put_page(page);
4252 break;
4253 case MC_TARGET_SWAP:
4254 ent = target.ent;
4255 if (!mem_cgroup_move_swap_account(ent,
4256 mc.from, mc.to, false)) {
4257 mc.precharge--;
4258 /* we fixup refcnts and charges later. */
4259 mc.moved_swap++;
4260 }
4261 break;
4262 default:
4263 break;
4264 }
4265 }
4266 pte_unmap_unlock(pte - 1, ptl);
4267 cond_resched();
4268
4269 if (addr != end) {
4270 /*
4271 * We have consumed all precharges we got in can_attach().
4272 * We try charge one by one, but don't do any additional
4273 * charges to mc.to if we have failed in charge once in attach()
4274 * phase.
4275 */
4276 ret = mem_cgroup_do_precharge(1);
4277 if (!ret)
4278 goto retry;
4279 }
4280
4281 return ret;
4282}
4283
4284static void mem_cgroup_move_charge(struct mm_struct *mm)
4285{
4286 struct vm_area_struct *vma;
4287
4288 lru_add_drain_all();
4289 down_read(&mm->mmap_sem);
4290 for (vma = mm->mmap; vma; vma = vma->vm_next) {
4291 int ret;
4292 struct mm_walk mem_cgroup_move_charge_walk = {
4293 .pmd_entry = mem_cgroup_move_charge_pte_range,
4294 .mm = mm,
4295 .private = vma,
4296 };
4297 if (is_vm_hugetlb_page(vma))
4298 continue;
4299 /* TODO: We don't move charges of shmem/tmpfs pages for now. */
4300 if (vma->vm_flags & VM_SHARED)
4301 continue;
4302 ret = walk_page_range(vma->vm_start, vma->vm_end,
4303 &mem_cgroup_move_charge_walk);
4304 if (ret)
4305 /*
4306 * means we have consumed all precharges and failed in
4307 * doing additional charge. Just abandon here.
4308 */
4309 break;
4310 }
4311 up_read(&mm->mmap_sem);
4312}
4313
4314static void mem_cgroup_move_task(struct cgroup_subsys *ss,
4315 struct cgroup *cont,
4316 struct cgroup *old_cont,
4317 struct task_struct *p,
4318 bool threadgroup)
4319{
4320 struct mm_struct *mm;
4321
4322 if (!mc.to)
4323 /* no need to move charge */
4324 return;
4325
4326 mm = get_task_mm(p);
4327 if (mm) {
4328 mem_cgroup_move_charge(mm);
4329 mmput(mm);
4330 }
4331 mem_cgroup_clear_mc();
3394} 4332}
3395 4333
3396struct cgroup_subsys mem_cgroup_subsys = { 4334struct cgroup_subsys mem_cgroup_subsys = {
@@ -3400,6 +4338,8 @@ struct cgroup_subsys mem_cgroup_subsys = {
3400 .pre_destroy = mem_cgroup_pre_destroy, 4338 .pre_destroy = mem_cgroup_pre_destroy,
3401 .destroy = mem_cgroup_destroy, 4339 .destroy = mem_cgroup_destroy,
3402 .populate = mem_cgroup_populate, 4340 .populate = mem_cgroup_populate,
4341 .can_attach = mem_cgroup_can_attach,
4342 .cancel_attach = mem_cgroup_cancel_attach,
3403 .attach = mem_cgroup_move_task, 4343 .attach = mem_cgroup_move_task,
3404 .early_init = 0, 4344 .early_init = 0,
3405 .use_id = 1, 4345 .use_id = 1,
generated by cgit v1.2.2 (git 2.25.0) at 2025-02-18 06:23:15 -0500