]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * linux/mm/oom_kill.c | |
3 | * | |
4 | * Copyright (C) 1998,2000 Rik van Riel | |
5 | * Thanks go out to Claus Fischer for some serious inspiration and | |
6 | * for goading me into coding this file... | |
7 | * Copyright (C) 2010 Google, Inc. | |
8 | * Rewritten by David Rientjes | |
9 | * | |
10 | * The routines in this file are used to kill a process when | |
11 | * we're seriously out of memory. This gets called from __alloc_pages() | |
12 | * in mm/page_alloc.c when we really run out of memory. | |
13 | * | |
14 | * Since we won't call these routines often (on a well-configured | |
15 | * machine) this file will double as a 'coding guide' and a signpost | |
16 | * for newbie kernel hackers. It features several pointers to major | |
17 | * kernel subsystems and hints as to where to find out what things do. | |
18 | */ | |
19 | ||
20 | #include <linux/oom.h> | |
21 | #include <linux/mm.h> | |
22 | #include <linux/err.h> | |
23 | #include <linux/gfp.h> | |
24 | #include <linux/sched.h> | |
25 | #include <linux/swap.h> | |
26 | #include <linux/timex.h> | |
27 | #include <linux/jiffies.h> | |
28 | #include <linux/cpuset.h> | |
29 | #include <linux/export.h> | |
30 | #include <linux/notifier.h> | |
31 | #include <linux/memcontrol.h> | |
32 | #include <linux/mempolicy.h> | |
33 | #include <linux/security.h> | |
34 | #include <linux/ptrace.h> | |
35 | #include <linux/freezer.h> | |
36 | #include <linux/ftrace.h> | |
37 | #include <linux/ratelimit.h> | |
38 | #include <linux/kthread.h> | |
39 | #include <linux/init.h> | |
40 | ||
41 | #include <asm/tlb.h> | |
42 | #include "internal.h" | |
43 | ||
44 | #define CREATE_TRACE_POINTS | |
45 | #include <trace/events/oom.h> | |
46 | ||
47 | int sysctl_panic_on_oom; | |
48 | int sysctl_oom_kill_allocating_task; | |
49 | int sysctl_oom_dump_tasks = 1; | |
50 | ||
51 | DEFINE_MUTEX(oom_lock); | |
52 | ||
53 | #ifdef CONFIG_NUMA | |
54 | /** | |
55 | * has_intersects_mems_allowed() - check task eligiblity for kill | |
56 | * @start: task struct of which task to consider | |
57 | * @mask: nodemask passed to page allocator for mempolicy ooms | |
58 | * | |
59 | * Task eligibility is determined by whether or not a candidate task, @tsk, | |
60 | * shares the same mempolicy nodes as current if it is bound by such a policy | |
61 | * and whether or not it has the same set of allowed cpuset nodes. | |
62 | */ | |
63 | static bool has_intersects_mems_allowed(struct task_struct *start, | |
64 | const nodemask_t *mask) | |
65 | { | |
66 | struct task_struct *tsk; | |
67 | bool ret = false; | |
68 | ||
69 | rcu_read_lock(); | |
70 | for_each_thread(start, tsk) { | |
71 | if (mask) { | |
72 | /* | |
73 | * If this is a mempolicy constrained oom, tsk's | |
74 | * cpuset is irrelevant. Only return true if its | |
75 | * mempolicy intersects current, otherwise it may be | |
76 | * needlessly killed. | |
77 | */ | |
78 | ret = mempolicy_nodemask_intersects(tsk, mask); | |
79 | } else { | |
80 | /* | |
81 | * This is not a mempolicy constrained oom, so only | |
82 | * check the mems of tsk's cpuset. | |
83 | */ | |
84 | ret = cpuset_mems_allowed_intersects(current, tsk); | |
85 | } | |
86 | if (ret) | |
87 | break; | |
88 | } | |
89 | rcu_read_unlock(); | |
90 | ||
91 | return ret; | |
92 | } | |
93 | #else | |
94 | static bool has_intersects_mems_allowed(struct task_struct *tsk, | |
95 | const nodemask_t *mask) | |
96 | { | |
97 | return true; | |
98 | } | |
99 | #endif /* CONFIG_NUMA */ | |
100 | ||
101 | /* | |
102 | * The process p may have detached its own ->mm while exiting or through | |
103 | * use_mm(), but one or more of its subthreads may still have a valid | |
104 | * pointer. Return p, or any of its subthreads with a valid ->mm, with | |
105 | * task_lock() held. | |
106 | */ | |
107 | struct task_struct *find_lock_task_mm(struct task_struct *p) | |
108 | { | |
109 | struct task_struct *t; | |
110 | ||
111 | rcu_read_lock(); | |
112 | ||
113 | for_each_thread(p, t) { | |
114 | task_lock(t); | |
115 | if (likely(t->mm)) | |
116 | goto found; | |
117 | task_unlock(t); | |
118 | } | |
119 | t = NULL; | |
120 | found: | |
121 | rcu_read_unlock(); | |
122 | ||
123 | return t; | |
124 | } | |
125 | ||
126 | /* | |
127 | * order == -1 means the oom kill is required by sysrq, otherwise only | |
128 | * for display purposes. | |
129 | */ | |
130 | static inline bool is_sysrq_oom(struct oom_control *oc) | |
131 | { | |
132 | return oc->order == -1; | |
133 | } | |
134 | ||
135 | /* return true if the task is not adequate as candidate victim task. */ | |
136 | static bool oom_unkillable_task(struct task_struct *p, | |
137 | struct mem_cgroup *memcg, const nodemask_t *nodemask) | |
138 | { | |
139 | if (is_global_init(p)) | |
140 | return true; | |
141 | if (p->flags & PF_KTHREAD) | |
142 | return true; | |
143 | ||
144 | /* When mem_cgroup_out_of_memory() and p is not member of the group */ | |
145 | if (memcg && !task_in_mem_cgroup(p, memcg)) | |
146 | return true; | |
147 | ||
148 | /* p may not have freeable memory in nodemask */ | |
149 | if (!has_intersects_mems_allowed(p, nodemask)) | |
150 | return true; | |
151 | ||
152 | return false; | |
153 | } | |
154 | ||
155 | /** | |
156 | * oom_badness - heuristic function to determine which candidate task to kill | |
157 | * @p: task struct of which task we should calculate | |
158 | * @totalpages: total present RAM allowed for page allocation | |
159 | * | |
160 | * The heuristic for determining which task to kill is made to be as simple and | |
161 | * predictable as possible. The goal is to return the highest value for the | |
162 | * task consuming the most memory to avoid subsequent oom failures. | |
163 | */ | |
164 | unsigned long oom_badness(struct task_struct *p, struct mem_cgroup *memcg, | |
165 | const nodemask_t *nodemask, unsigned long totalpages) | |
166 | { | |
167 | long points; | |
168 | long adj; | |
169 | ||
170 | if (oom_unkillable_task(p, memcg, nodemask)) | |
171 | return 0; | |
172 | ||
173 | p = find_lock_task_mm(p); | |
174 | if (!p) | |
175 | return 0; | |
176 | ||
177 | /* | |
178 | * Do not even consider tasks which are explicitly marked oom | |
179 | * unkillable or have been already oom reaped. | |
180 | */ | |
181 | adj = (long)p->signal->oom_score_adj; | |
182 | if (adj == OOM_SCORE_ADJ_MIN || | |
183 | test_bit(MMF_OOM_REAPED, &p->mm->flags)) { | |
184 | task_unlock(p); | |
185 | return 0; | |
186 | } | |
187 | ||
188 | /* | |
189 | * The baseline for the badness score is the proportion of RAM that each | |
190 | * task's rss, pagetable and swap space use. | |
191 | */ | |
192 | points = get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS) + | |
193 | atomic_long_read(&p->mm->nr_ptes) + mm_nr_pmds(p->mm); | |
194 | task_unlock(p); | |
195 | ||
196 | /* | |
197 | * Root processes get 3% bonus, just like the __vm_enough_memory() | |
198 | * implementation used by LSMs. | |
199 | */ | |
200 | if (has_capability_noaudit(p, CAP_SYS_ADMIN)) | |
201 | points -= (points * 3) / 100; | |
202 | ||
203 | /* Normalize to oom_score_adj units */ | |
204 | adj *= totalpages / 1000; | |
205 | points += adj; | |
206 | ||
207 | /* | |
208 | * Never return 0 for an eligible task regardless of the root bonus and | |
209 | * oom_score_adj (oom_score_adj can't be OOM_SCORE_ADJ_MIN here). | |
210 | */ | |
211 | return points > 0 ? points : 1; | |
212 | } | |
213 | ||
214 | /* | |
215 | * Determine the type of allocation constraint. | |
216 | */ | |
217 | #ifdef CONFIG_NUMA | |
218 | static enum oom_constraint constrained_alloc(struct oom_control *oc, | |
219 | unsigned long *totalpages) | |
220 | { | |
221 | struct zone *zone; | |
222 | struct zoneref *z; | |
223 | enum zone_type high_zoneidx = gfp_zone(oc->gfp_mask); | |
224 | bool cpuset_limited = false; | |
225 | int nid; | |
226 | ||
227 | /* Default to all available memory */ | |
228 | *totalpages = totalram_pages + total_swap_pages; | |
229 | ||
230 | if (!oc->zonelist) | |
231 | return CONSTRAINT_NONE; | |
232 | /* | |
233 | * Reach here only when __GFP_NOFAIL is used. So, we should avoid | |
234 | * to kill current.We have to random task kill in this case. | |
235 | * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now. | |
236 | */ | |
237 | if (oc->gfp_mask & __GFP_THISNODE) | |
238 | return CONSTRAINT_NONE; | |
239 | ||
240 | /* | |
241 | * This is not a __GFP_THISNODE allocation, so a truncated nodemask in | |
242 | * the page allocator means a mempolicy is in effect. Cpuset policy | |
243 | * is enforced in get_page_from_freelist(). | |
244 | */ | |
245 | if (oc->nodemask && | |
246 | !nodes_subset(node_states[N_MEMORY], *oc->nodemask)) { | |
247 | *totalpages = total_swap_pages; | |
248 | for_each_node_mask(nid, *oc->nodemask) | |
249 | *totalpages += node_spanned_pages(nid); | |
250 | return CONSTRAINT_MEMORY_POLICY; | |
251 | } | |
252 | ||
253 | /* Check this allocation failure is caused by cpuset's wall function */ | |
254 | for_each_zone_zonelist_nodemask(zone, z, oc->zonelist, | |
255 | high_zoneidx, oc->nodemask) | |
256 | if (!cpuset_zone_allowed(zone, oc->gfp_mask)) | |
257 | cpuset_limited = true; | |
258 | ||
259 | if (cpuset_limited) { | |
260 | *totalpages = total_swap_pages; | |
261 | for_each_node_mask(nid, cpuset_current_mems_allowed) | |
262 | *totalpages += node_spanned_pages(nid); | |
263 | return CONSTRAINT_CPUSET; | |
264 | } | |
265 | return CONSTRAINT_NONE; | |
266 | } | |
267 | #else | |
268 | static enum oom_constraint constrained_alloc(struct oom_control *oc, | |
269 | unsigned long *totalpages) | |
270 | { | |
271 | *totalpages = totalram_pages + total_swap_pages; | |
272 | return CONSTRAINT_NONE; | |
273 | } | |
274 | #endif | |
275 | ||
276 | enum oom_scan_t oom_scan_process_thread(struct oom_control *oc, | |
277 | struct task_struct *task, unsigned long totalpages) | |
278 | { | |
279 | if (oom_unkillable_task(task, NULL, oc->nodemask)) | |
280 | return OOM_SCAN_CONTINUE; | |
281 | ||
282 | /* | |
283 | * This task already has access to memory reserves and is being killed. | |
284 | * Don't allow any other task to have access to the reserves. | |
285 | */ | |
286 | if (!is_sysrq_oom(oc) && atomic_read(&task->signal->oom_victims)) | |
287 | return OOM_SCAN_ABORT; | |
288 | ||
289 | /* | |
290 | * If task is allocating a lot of memory and has been marked to be | |
291 | * killed first if it triggers an oom, then select it. | |
292 | */ | |
293 | if (oom_task_origin(task)) | |
294 | return OOM_SCAN_SELECT; | |
295 | ||
296 | return OOM_SCAN_OK; | |
297 | } | |
298 | ||
299 | /* | |
300 | * Simple selection loop. We chose the process with the highest | |
301 | * number of 'points'. Returns -1 on scan abort. | |
302 | */ | |
303 | static struct task_struct *select_bad_process(struct oom_control *oc, | |
304 | unsigned int *ppoints, unsigned long totalpages) | |
305 | { | |
306 | struct task_struct *p; | |
307 | struct task_struct *chosen = NULL; | |
308 | unsigned long chosen_points = 0; | |
309 | ||
310 | rcu_read_lock(); | |
311 | for_each_process(p) { | |
312 | unsigned int points; | |
313 | ||
314 | switch (oom_scan_process_thread(oc, p, totalpages)) { | |
315 | case OOM_SCAN_SELECT: | |
316 | chosen = p; | |
317 | chosen_points = ULONG_MAX; | |
318 | /* fall through */ | |
319 | case OOM_SCAN_CONTINUE: | |
320 | continue; | |
321 | case OOM_SCAN_ABORT: | |
322 | rcu_read_unlock(); | |
323 | return (struct task_struct *)(-1UL); | |
324 | case OOM_SCAN_OK: | |
325 | break; | |
326 | }; | |
327 | points = oom_badness(p, NULL, oc->nodemask, totalpages); | |
328 | if (!points || points < chosen_points) | |
329 | continue; | |
330 | ||
331 | chosen = p; | |
332 | chosen_points = points; | |
333 | } | |
334 | if (chosen) | |
335 | get_task_struct(chosen); | |
336 | rcu_read_unlock(); | |
337 | ||
338 | *ppoints = chosen_points * 1000 / totalpages; | |
339 | return chosen; | |
340 | } | |
341 | ||
342 | /** | |
343 | * dump_tasks - dump current memory state of all system tasks | |
344 | * @memcg: current's memory controller, if constrained | |
345 | * @nodemask: nodemask passed to page allocator for mempolicy ooms | |
346 | * | |
347 | * Dumps the current memory state of all eligible tasks. Tasks not in the same | |
348 | * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes | |
349 | * are not shown. | |
350 | * State information includes task's pid, uid, tgid, vm size, rss, nr_ptes, | |
351 | * swapents, oom_score_adj value, and name. | |
352 | */ | |
353 | static void dump_tasks(struct mem_cgroup *memcg, const nodemask_t *nodemask) | |
354 | { | |
355 | struct task_struct *p; | |
356 | struct task_struct *task; | |
357 | ||
358 | pr_info("[ pid ] uid tgid total_vm rss nr_ptes nr_pmds swapents oom_score_adj name\n"); | |
359 | rcu_read_lock(); | |
360 | for_each_process(p) { | |
361 | if (oom_unkillable_task(p, memcg, nodemask)) | |
362 | continue; | |
363 | ||
364 | task = find_lock_task_mm(p); | |
365 | if (!task) { | |
366 | /* | |
367 | * This is a kthread or all of p's threads have already | |
368 | * detached their mm's. There's no need to report | |
369 | * them; they can't be oom killed anyway. | |
370 | */ | |
371 | continue; | |
372 | } | |
373 | ||
374 | pr_info("[%5d] %5d %5d %8lu %8lu %7ld %7ld %8lu %5hd %s\n", | |
375 | task->pid, from_kuid(&init_user_ns, task_uid(task)), | |
376 | task->tgid, task->mm->total_vm, get_mm_rss(task->mm), | |
377 | atomic_long_read(&task->mm->nr_ptes), | |
378 | mm_nr_pmds(task->mm), | |
379 | get_mm_counter(task->mm, MM_SWAPENTS), | |
380 | task->signal->oom_score_adj, task->comm); | |
381 | task_unlock(task); | |
382 | } | |
383 | rcu_read_unlock(); | |
384 | } | |
385 | ||
386 | static void dump_header(struct oom_control *oc, struct task_struct *p, | |
387 | struct mem_cgroup *memcg) | |
388 | { | |
389 | pr_warn("%s invoked oom-killer: gfp_mask=%#x(%pGg), order=%d, oom_score_adj=%hd\n", | |
390 | current->comm, oc->gfp_mask, &oc->gfp_mask, oc->order, | |
391 | current->signal->oom_score_adj); | |
392 | ||
393 | cpuset_print_current_mems_allowed(); | |
394 | dump_stack(); | |
395 | if (memcg) | |
396 | mem_cgroup_print_oom_info(memcg, p); | |
397 | else | |
398 | show_mem(SHOW_MEM_FILTER_NODES); | |
399 | if (sysctl_oom_dump_tasks) | |
400 | dump_tasks(memcg, oc->nodemask); | |
401 | } | |
402 | ||
403 | /* | |
404 | * Number of OOM victims in flight | |
405 | */ | |
406 | static atomic_t oom_victims = ATOMIC_INIT(0); | |
407 | static DECLARE_WAIT_QUEUE_HEAD(oom_victims_wait); | |
408 | ||
409 | bool oom_killer_disabled __read_mostly; | |
410 | ||
411 | #define K(x) ((x) << (PAGE_SHIFT-10)) | |
412 | ||
413 | /* | |
414 | * task->mm can be NULL if the task is the exited group leader. So to | |
415 | * determine whether the task is using a particular mm, we examine all the | |
416 | * task's threads: if one of those is using this mm then this task was also | |
417 | * using it. | |
418 | */ | |
419 | static bool process_shares_mm(struct task_struct *p, struct mm_struct *mm) | |
420 | { | |
421 | struct task_struct *t; | |
422 | ||
423 | for_each_thread(p, t) { | |
424 | struct mm_struct *t_mm = READ_ONCE(t->mm); | |
425 | if (t_mm) | |
426 | return t_mm == mm; | |
427 | } | |
428 | return false; | |
429 | } | |
430 | ||
431 | ||
432 | #ifdef CONFIG_MMU | |
433 | /* | |
434 | * OOM Reaper kernel thread which tries to reap the memory used by the OOM | |
435 | * victim (if that is possible) to help the OOM killer to move on. | |
436 | */ | |
437 | static struct task_struct *oom_reaper_th; | |
438 | static DECLARE_WAIT_QUEUE_HEAD(oom_reaper_wait); | |
439 | static struct task_struct *oom_reaper_list; | |
440 | static DEFINE_SPINLOCK(oom_reaper_lock); | |
441 | ||
442 | static bool __oom_reap_task(struct task_struct *tsk) | |
443 | { | |
444 | struct mmu_gather tlb; | |
445 | struct vm_area_struct *vma; | |
446 | struct mm_struct *mm = NULL; | |
447 | struct task_struct *p; | |
448 | struct zap_details details = {.check_swap_entries = true, | |
449 | .ignore_dirty = true}; | |
450 | bool ret = true; | |
451 | ||
452 | /* | |
453 | * We have to make sure to not race with the victim exit path | |
454 | * and cause premature new oom victim selection: | |
455 | * __oom_reap_task exit_mm | |
456 | * atomic_inc_not_zero | |
457 | * mmput | |
458 | * atomic_dec_and_test | |
459 | * exit_oom_victim | |
460 | * [...] | |
461 | * out_of_memory | |
462 | * select_bad_process | |
463 | * # no TIF_MEMDIE task selects new victim | |
464 | * unmap_page_range # frees some memory | |
465 | */ | |
466 | mutex_lock(&oom_lock); | |
467 | ||
468 | /* | |
469 | * Make sure we find the associated mm_struct even when the particular | |
470 | * thread has already terminated and cleared its mm. | |
471 | * We might have race with exit path so consider our work done if there | |
472 | * is no mm. | |
473 | */ | |
474 | p = find_lock_task_mm(tsk); | |
475 | if (!p) | |
476 | goto unlock_oom; | |
477 | mm = p->mm; | |
478 | atomic_inc(&mm->mm_users); | |
479 | task_unlock(p); | |
480 | ||
481 | if (!down_read_trylock(&mm->mmap_sem)) { | |
482 | ret = false; | |
483 | goto unlock_oom; | |
484 | } | |
485 | ||
486 | tlb_gather_mmu(&tlb, mm, 0, -1); | |
487 | for (vma = mm->mmap ; vma; vma = vma->vm_next) { | |
488 | if (is_vm_hugetlb_page(vma)) | |
489 | continue; | |
490 | ||
491 | /* | |
492 | * mlocked VMAs require explicit munlocking before unmap. | |
493 | * Let's keep it simple here and skip such VMAs. | |
494 | */ | |
495 | if (vma->vm_flags & VM_LOCKED) | |
496 | continue; | |
497 | ||
498 | /* | |
499 | * Only anonymous pages have a good chance to be dropped | |
500 | * without additional steps which we cannot afford as we | |
501 | * are OOM already. | |
502 | * | |
503 | * We do not even care about fs backed pages because all | |
504 | * which are reclaimable have already been reclaimed and | |
505 | * we do not want to block exit_mmap by keeping mm ref | |
506 | * count elevated without a good reason. | |
507 | */ | |
508 | if (vma_is_anonymous(vma) || !(vma->vm_flags & VM_SHARED)) | |
509 | unmap_page_range(&tlb, vma, vma->vm_start, vma->vm_end, | |
510 | &details); | |
511 | } | |
512 | tlb_finish_mmu(&tlb, 0, -1); | |
513 | pr_info("oom_reaper: reaped process %d (%s), now anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n", | |
514 | task_pid_nr(tsk), tsk->comm, | |
515 | K(get_mm_counter(mm, MM_ANONPAGES)), | |
516 | K(get_mm_counter(mm, MM_FILEPAGES)), | |
517 | K(get_mm_counter(mm, MM_SHMEMPAGES))); | |
518 | up_read(&mm->mmap_sem); | |
519 | ||
520 | /* | |
521 | * This task can be safely ignored because we cannot do much more | |
522 | * to release its memory. | |
523 | */ | |
524 | set_bit(MMF_OOM_REAPED, &mm->flags); | |
525 | unlock_oom: | |
526 | mutex_unlock(&oom_lock); | |
527 | /* | |
528 | * Drop our reference but make sure the mmput slow path is called from a | |
529 | * different context because we shouldn't risk we get stuck there and | |
530 | * put the oom_reaper out of the way. | |
531 | */ | |
532 | if (mm) | |
533 | mmput_async(mm); | |
534 | return ret; | |
535 | } | |
536 | ||
537 | #define MAX_OOM_REAP_RETRIES 10 | |
538 | static void oom_reap_task(struct task_struct *tsk) | |
539 | { | |
540 | int attempts = 0; | |
541 | ||
542 | /* Retry the down_read_trylock(mmap_sem) a few times */ | |
543 | while (attempts++ < MAX_OOM_REAP_RETRIES && !__oom_reap_task(tsk)) | |
544 | schedule_timeout_idle(HZ/10); | |
545 | ||
546 | if (attempts > MAX_OOM_REAP_RETRIES) { | |
547 | pr_info("oom_reaper: unable to reap pid:%d (%s)\n", | |
548 | task_pid_nr(tsk), tsk->comm); | |
549 | debug_show_all_locks(); | |
550 | } | |
551 | ||
552 | /* | |
553 | * Clear TIF_MEMDIE because the task shouldn't be sitting on a | |
554 | * reasonably reclaimable memory anymore or it is not a good candidate | |
555 | * for the oom victim right now because it cannot release its memory | |
556 | * itself nor by the oom reaper. | |
557 | */ | |
558 | tsk->oom_reaper_list = NULL; | |
559 | exit_oom_victim(tsk); | |
560 | ||
561 | /* Drop a reference taken by wake_oom_reaper */ | |
562 | put_task_struct(tsk); | |
563 | } | |
564 | ||
565 | static int oom_reaper(void *unused) | |
566 | { | |
567 | set_freezable(); | |
568 | ||
569 | while (true) { | |
570 | struct task_struct *tsk = NULL; | |
571 | ||
572 | wait_event_freezable(oom_reaper_wait, oom_reaper_list != NULL); | |
573 | spin_lock(&oom_reaper_lock); | |
574 | if (oom_reaper_list != NULL) { | |
575 | tsk = oom_reaper_list; | |
576 | oom_reaper_list = tsk->oom_reaper_list; | |
577 | } | |
578 | spin_unlock(&oom_reaper_lock); | |
579 | ||
580 | if (tsk) | |
581 | oom_reap_task(tsk); | |
582 | } | |
583 | ||
584 | return 0; | |
585 | } | |
586 | ||
587 | static void wake_oom_reaper(struct task_struct *tsk) | |
588 | { | |
589 | if (!oom_reaper_th) | |
590 | return; | |
591 | ||
592 | /* tsk is already queued? */ | |
593 | if (tsk == oom_reaper_list || tsk->oom_reaper_list) | |
594 | return; | |
595 | ||
596 | get_task_struct(tsk); | |
597 | ||
598 | spin_lock(&oom_reaper_lock); | |
599 | tsk->oom_reaper_list = oom_reaper_list; | |
600 | oom_reaper_list = tsk; | |
601 | spin_unlock(&oom_reaper_lock); | |
602 | wake_up(&oom_reaper_wait); | |
603 | } | |
604 | ||
605 | /* Check if we can reap the given task. This has to be called with stable | |
606 | * tsk->mm | |
607 | */ | |
608 | void try_oom_reaper(struct task_struct *tsk) | |
609 | { | |
610 | struct mm_struct *mm = tsk->mm; | |
611 | struct task_struct *p; | |
612 | ||
613 | if (!mm) | |
614 | return; | |
615 | ||
616 | /* | |
617 | * There might be other threads/processes which are either not | |
618 | * dying or even not killable. | |
619 | */ | |
620 | if (atomic_read(&mm->mm_users) > 1) { | |
621 | rcu_read_lock(); | |
622 | for_each_process(p) { | |
623 | if (!process_shares_mm(p, mm)) | |
624 | continue; | |
625 | if (fatal_signal_pending(p)) | |
626 | continue; | |
627 | ||
628 | /* | |
629 | * If the task is exiting make sure the whole thread group | |
630 | * is exiting and cannot acces mm anymore. | |
631 | */ | |
632 | if (signal_group_exit(p->signal)) | |
633 | continue; | |
634 | ||
635 | /* Give up */ | |
636 | rcu_read_unlock(); | |
637 | return; | |
638 | } | |
639 | rcu_read_unlock(); | |
640 | } | |
641 | ||
642 | wake_oom_reaper(tsk); | |
643 | } | |
644 | ||
645 | static int __init oom_init(void) | |
646 | { | |
647 | oom_reaper_th = kthread_run(oom_reaper, NULL, "oom_reaper"); | |
648 | if (IS_ERR(oom_reaper_th)) { | |
649 | pr_err("Unable to start OOM reaper %ld. Continuing regardless\n", | |
650 | PTR_ERR(oom_reaper_th)); | |
651 | oom_reaper_th = NULL; | |
652 | } | |
653 | return 0; | |
654 | } | |
655 | subsys_initcall(oom_init) | |
656 | #else | |
657 | static void wake_oom_reaper(struct task_struct *tsk) | |
658 | { | |
659 | } | |
660 | #endif | |
661 | ||
662 | /** | |
663 | * mark_oom_victim - mark the given task as OOM victim | |
664 | * @tsk: task to mark | |
665 | * | |
666 | * Has to be called with oom_lock held and never after | |
667 | * oom has been disabled already. | |
668 | */ | |
669 | void mark_oom_victim(struct task_struct *tsk) | |
670 | { | |
671 | WARN_ON(oom_killer_disabled); | |
672 | /* OOM killer might race with memcg OOM */ | |
673 | if (test_and_set_tsk_thread_flag(tsk, TIF_MEMDIE)) | |
674 | return; | |
675 | atomic_inc(&tsk->signal->oom_victims); | |
676 | /* | |
677 | * Make sure that the task is woken up from uninterruptible sleep | |
678 | * if it is frozen because OOM killer wouldn't be able to free | |
679 | * any memory and livelock. freezing_slow_path will tell the freezer | |
680 | * that TIF_MEMDIE tasks should be ignored. | |
681 | */ | |
682 | __thaw_task(tsk); | |
683 | atomic_inc(&oom_victims); | |
684 | } | |
685 | ||
686 | /** | |
687 | * exit_oom_victim - note the exit of an OOM victim | |
688 | */ | |
689 | void exit_oom_victim(struct task_struct *tsk) | |
690 | { | |
691 | if (!test_and_clear_tsk_thread_flag(tsk, TIF_MEMDIE)) | |
692 | return; | |
693 | atomic_dec(&tsk->signal->oom_victims); | |
694 | ||
695 | if (!atomic_dec_return(&oom_victims)) | |
696 | wake_up_all(&oom_victims_wait); | |
697 | } | |
698 | ||
699 | /** | |
700 | * oom_killer_disable - disable OOM killer | |
701 | * | |
702 | * Forces all page allocations to fail rather than trigger OOM killer. | |
703 | * Will block and wait until all OOM victims are killed. | |
704 | * | |
705 | * The function cannot be called when there are runnable user tasks because | |
706 | * the userspace would see unexpected allocation failures as a result. Any | |
707 | * new usage of this function should be consulted with MM people. | |
708 | * | |
709 | * Returns true if successful and false if the OOM killer cannot be | |
710 | * disabled. | |
711 | */ | |
712 | bool oom_killer_disable(void) | |
713 | { | |
714 | /* | |
715 | * Make sure to not race with an ongoing OOM killer. Check that the | |
716 | * current is not killed (possibly due to sharing the victim's memory). | |
717 | */ | |
718 | if (mutex_lock_killable(&oom_lock)) | |
719 | return false; | |
720 | oom_killer_disabled = true; | |
721 | mutex_unlock(&oom_lock); | |
722 | ||
723 | wait_event(oom_victims_wait, !atomic_read(&oom_victims)); | |
724 | ||
725 | return true; | |
726 | } | |
727 | ||
728 | /** | |
729 | * oom_killer_enable - enable OOM killer | |
730 | */ | |
731 | void oom_killer_enable(void) | |
732 | { | |
733 | oom_killer_disabled = false; | |
734 | } | |
735 | ||
736 | /* | |
737 | * Must be called while holding a reference to p, which will be released upon | |
738 | * returning. | |
739 | */ | |
740 | void oom_kill_process(struct oom_control *oc, struct task_struct *p, | |
741 | unsigned int points, unsigned long totalpages, | |
742 | struct mem_cgroup *memcg, const char *message) | |
743 | { | |
744 | struct task_struct *victim = p; | |
745 | struct task_struct *child; | |
746 | struct task_struct *t; | |
747 | struct mm_struct *mm; | |
748 | unsigned int victim_points = 0; | |
749 | static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL, | |
750 | DEFAULT_RATELIMIT_BURST); | |
751 | bool can_oom_reap = true; | |
752 | ||
753 | /* | |
754 | * If the task is already exiting, don't alarm the sysadmin or kill | |
755 | * its children or threads, just set TIF_MEMDIE so it can die quickly | |
756 | */ | |
757 | task_lock(p); | |
758 | if (p->mm && task_will_free_mem(p)) { | |
759 | mark_oom_victim(p); | |
760 | try_oom_reaper(p); | |
761 | task_unlock(p); | |
762 | put_task_struct(p); | |
763 | return; | |
764 | } | |
765 | task_unlock(p); | |
766 | ||
767 | if (__ratelimit(&oom_rs)) | |
768 | dump_header(oc, p, memcg); | |
769 | ||
770 | pr_err("%s: Kill process %d (%s) score %u or sacrifice child\n", | |
771 | message, task_pid_nr(p), p->comm, points); | |
772 | ||
773 | /* | |
774 | * If any of p's children has a different mm and is eligible for kill, | |
775 | * the one with the highest oom_badness() score is sacrificed for its | |
776 | * parent. This attempts to lose the minimal amount of work done while | |
777 | * still freeing memory. | |
778 | */ | |
779 | read_lock(&tasklist_lock); | |
780 | for_each_thread(p, t) { | |
781 | list_for_each_entry(child, &t->children, sibling) { | |
782 | unsigned int child_points; | |
783 | ||
784 | if (process_shares_mm(child, p->mm)) | |
785 | continue; | |
786 | /* | |
787 | * oom_badness() returns 0 if the thread is unkillable | |
788 | */ | |
789 | child_points = oom_badness(child, memcg, oc->nodemask, | |
790 | totalpages); | |
791 | if (child_points > victim_points) { | |
792 | put_task_struct(victim); | |
793 | victim = child; | |
794 | victim_points = child_points; | |
795 | get_task_struct(victim); | |
796 | } | |
797 | } | |
798 | } | |
799 | read_unlock(&tasklist_lock); | |
800 | ||
801 | p = find_lock_task_mm(victim); | |
802 | if (!p) { | |
803 | put_task_struct(victim); | |
804 | return; | |
805 | } else if (victim != p) { | |
806 | get_task_struct(p); | |
807 | put_task_struct(victim); | |
808 | victim = p; | |
809 | } | |
810 | ||
811 | /* Get a reference to safely compare mm after task_unlock(victim) */ | |
812 | mm = victim->mm; | |
813 | atomic_inc(&mm->mm_count); | |
814 | /* | |
815 | * We should send SIGKILL before setting TIF_MEMDIE in order to prevent | |
816 | * the OOM victim from depleting the memory reserves from the user | |
817 | * space under its control. | |
818 | */ | |
819 | do_send_sig_info(SIGKILL, SEND_SIG_FORCED, victim, true); | |
820 | mark_oom_victim(victim); | |
821 | pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n", | |
822 | task_pid_nr(victim), victim->comm, K(victim->mm->total_vm), | |
823 | K(get_mm_counter(victim->mm, MM_ANONPAGES)), | |
824 | K(get_mm_counter(victim->mm, MM_FILEPAGES)), | |
825 | K(get_mm_counter(victim->mm, MM_SHMEMPAGES))); | |
826 | task_unlock(victim); | |
827 | ||
828 | /* | |
829 | * Kill all user processes sharing victim->mm in other thread groups, if | |
830 | * any. They don't get access to memory reserves, though, to avoid | |
831 | * depletion of all memory. This prevents mm->mmap_sem livelock when an | |
832 | * oom killed thread cannot exit because it requires the semaphore and | |
833 | * its contended by another thread trying to allocate memory itself. | |
834 | * That thread will now get access to memory reserves since it has a | |
835 | * pending fatal signal. | |
836 | */ | |
837 | rcu_read_lock(); | |
838 | for_each_process(p) { | |
839 | if (!process_shares_mm(p, mm)) | |
840 | continue; | |
841 | if (same_thread_group(p, victim)) | |
842 | continue; | |
843 | if (unlikely(p->flags & PF_KTHREAD) || is_global_init(p) || | |
844 | p->signal->oom_score_adj == OOM_SCORE_ADJ_MIN) { | |
845 | /* | |
846 | * We cannot use oom_reaper for the mm shared by this | |
847 | * process because it wouldn't get killed and so the | |
848 | * memory might be still used. | |
849 | */ | |
850 | can_oom_reap = false; | |
851 | continue; | |
852 | } | |
853 | do_send_sig_info(SIGKILL, SEND_SIG_FORCED, p, true); | |
854 | } | |
855 | rcu_read_unlock(); | |
856 | ||
857 | if (can_oom_reap) | |
858 | wake_oom_reaper(victim); | |
859 | ||
860 | mmdrop(mm); | |
861 | put_task_struct(victim); | |
862 | } | |
863 | #undef K | |
864 | ||
865 | /* | |
866 | * Determines whether the kernel must panic because of the panic_on_oom sysctl. | |
867 | */ | |
868 | void check_panic_on_oom(struct oom_control *oc, enum oom_constraint constraint, | |
869 | struct mem_cgroup *memcg) | |
870 | { | |
871 | if (likely(!sysctl_panic_on_oom)) | |
872 | return; | |
873 | if (sysctl_panic_on_oom != 2) { | |
874 | /* | |
875 | * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel | |
876 | * does not panic for cpuset, mempolicy, or memcg allocation | |
877 | * failures. | |
878 | */ | |
879 | if (constraint != CONSTRAINT_NONE) | |
880 | return; | |
881 | } | |
882 | /* Do not panic for oom kills triggered by sysrq */ | |
883 | if (is_sysrq_oom(oc)) | |
884 | return; | |
885 | dump_header(oc, NULL, memcg); | |
886 | panic("Out of memory: %s panic_on_oom is enabled\n", | |
887 | sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide"); | |
888 | } | |
889 | ||
890 | static BLOCKING_NOTIFIER_HEAD(oom_notify_list); | |
891 | ||
892 | int register_oom_notifier(struct notifier_block *nb) | |
893 | { | |
894 | return blocking_notifier_chain_register(&oom_notify_list, nb); | |
895 | } | |
896 | EXPORT_SYMBOL_GPL(register_oom_notifier); | |
897 | ||
898 | int unregister_oom_notifier(struct notifier_block *nb) | |
899 | { | |
900 | return blocking_notifier_chain_unregister(&oom_notify_list, nb); | |
901 | } | |
902 | EXPORT_SYMBOL_GPL(unregister_oom_notifier); | |
903 | ||
904 | /** | |
905 | * out_of_memory - kill the "best" process when we run out of memory | |
906 | * @oc: pointer to struct oom_control | |
907 | * | |
908 | * If we run out of memory, we have the choice between either | |
909 | * killing a random task (bad), letting the system crash (worse) | |
910 | * OR try to be smart about which process to kill. Note that we | |
911 | * don't have to be perfect here, we just have to be good. | |
912 | */ | |
913 | bool out_of_memory(struct oom_control *oc) | |
914 | { | |
915 | struct task_struct *p; | |
916 | unsigned long totalpages; | |
917 | unsigned long freed = 0; | |
918 | unsigned int uninitialized_var(points); | |
919 | enum oom_constraint constraint = CONSTRAINT_NONE; | |
920 | ||
921 | if (oom_killer_disabled) | |
922 | return false; | |
923 | ||
924 | blocking_notifier_call_chain(&oom_notify_list, 0, &freed); | |
925 | if (freed > 0) | |
926 | /* Got some memory back in the last second. */ | |
927 | return true; | |
928 | ||
929 | /* | |
930 | * If current has a pending SIGKILL or is exiting, then automatically | |
931 | * select it. The goal is to allow it to allocate so that it may | |
932 | * quickly exit and free its memory. | |
933 | * | |
934 | * But don't select if current has already released its mm and cleared | |
935 | * TIF_MEMDIE flag at exit_mm(), otherwise an OOM livelock may occur. | |
936 | */ | |
937 | if (current->mm && | |
938 | (fatal_signal_pending(current) || task_will_free_mem(current))) { | |
939 | mark_oom_victim(current); | |
940 | try_oom_reaper(current); | |
941 | return true; | |
942 | } | |
943 | ||
944 | /* | |
945 | * The OOM killer does not compensate for IO-less reclaim. | |
946 | * pagefault_out_of_memory lost its gfp context so we have to | |
947 | * make sure exclude 0 mask - all other users should have at least | |
948 | * ___GFP_DIRECT_RECLAIM to get here. | |
949 | */ | |
950 | if (oc->gfp_mask && !(oc->gfp_mask & (__GFP_FS|__GFP_NOFAIL))) | |
951 | return true; | |
952 | ||
953 | /* | |
954 | * Check if there were limitations on the allocation (only relevant for | |
955 | * NUMA) that may require different handling. | |
956 | */ | |
957 | constraint = constrained_alloc(oc, &totalpages); | |
958 | if (constraint != CONSTRAINT_MEMORY_POLICY) | |
959 | oc->nodemask = NULL; | |
960 | check_panic_on_oom(oc, constraint, NULL); | |
961 | ||
962 | if (sysctl_oom_kill_allocating_task && current->mm && | |
963 | !oom_unkillable_task(current, NULL, oc->nodemask) && | |
964 | current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) { | |
965 | get_task_struct(current); | |
966 | oom_kill_process(oc, current, 0, totalpages, NULL, | |
967 | "Out of memory (oom_kill_allocating_task)"); | |
968 | return true; | |
969 | } | |
970 | ||
971 | p = select_bad_process(oc, &points, totalpages); | |
972 | /* Found nothing?!?! Either we hang forever, or we panic. */ | |
973 | if (!p && !is_sysrq_oom(oc)) { | |
974 | dump_header(oc, NULL, NULL); | |
975 | panic("Out of memory and no killable processes...\n"); | |
976 | } | |
977 | if (p && p != (void *)-1UL) { | |
978 | oom_kill_process(oc, p, points, totalpages, NULL, | |
979 | "Out of memory"); | |
980 | /* | |
981 | * Give the killed process a good chance to exit before trying | |
982 | * to allocate memory again. | |
983 | */ | |
984 | schedule_timeout_killable(1); | |
985 | } | |
986 | return true; | |
987 | } | |
988 | ||
989 | /* | |
990 | * The pagefault handler calls here because it is out of memory, so kill a | |
991 | * memory-hogging task. If oom_lock is held by somebody else, a parallel oom | |
992 | * killing is already in progress so do nothing. | |
993 | */ | |
994 | void pagefault_out_of_memory(void) | |
995 | { | |
996 | struct oom_control oc = { | |
997 | .zonelist = NULL, | |
998 | .nodemask = NULL, | |
999 | .gfp_mask = 0, | |
1000 | .order = 0, | |
1001 | }; | |
1002 | ||
1003 | if (mem_cgroup_oom_synchronize(true)) | |
1004 | return; | |
1005 | ||
1006 | if (!mutex_trylock(&oom_lock)) | |
1007 | return; | |
1008 | ||
1009 | if (!out_of_memory(&oc)) { | |
1010 | /* | |
1011 | * There shouldn't be any user tasks runnable while the | |
1012 | * OOM killer is disabled, so the current task has to | |
1013 | * be a racing OOM victim for which oom_killer_disable() | |
1014 | * is waiting for. | |
1015 | */ | |
1016 | WARN_ON(test_thread_flag(TIF_MEMDIE)); | |
1017 | } | |
1018 | ||
1019 | mutex_unlock(&oom_lock); | |
1020 | } |