]>
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 | static inline bool is_memcg_oom(struct oom_control *oc) | |
136 | { | |
137 | return oc->memcg != NULL; | |
138 | } | |
139 | ||
140 | /* return true if the task is not adequate as candidate victim task. */ | |
141 | static bool oom_unkillable_task(struct task_struct *p, | |
142 | struct mem_cgroup *memcg, const nodemask_t *nodemask) | |
143 | { | |
144 | if (is_global_init(p)) | |
145 | return true; | |
146 | if (p->flags & PF_KTHREAD) | |
147 | return true; | |
148 | ||
149 | /* When mem_cgroup_out_of_memory() and p is not member of the group */ | |
150 | if (memcg && !task_in_mem_cgroup(p, memcg)) | |
151 | return true; | |
152 | ||
153 | /* p may not have freeable memory in nodemask */ | |
154 | if (!has_intersects_mems_allowed(p, nodemask)) | |
155 | return true; | |
156 | ||
157 | return false; | |
158 | } | |
159 | ||
160 | /** | |
161 | * oom_badness - heuristic function to determine which candidate task to kill | |
162 | * @p: task struct of which task we should calculate | |
163 | * @totalpages: total present RAM allowed for page allocation | |
164 | * | |
165 | * The heuristic for determining which task to kill is made to be as simple and | |
166 | * predictable as possible. The goal is to return the highest value for the | |
167 | * task consuming the most memory to avoid subsequent oom failures. | |
168 | */ | |
169 | unsigned long oom_badness(struct task_struct *p, struct mem_cgroup *memcg, | |
170 | const nodemask_t *nodemask, unsigned long totalpages) | |
171 | { | |
172 | long points; | |
173 | long adj; | |
174 | ||
175 | if (oom_unkillable_task(p, memcg, nodemask)) | |
176 | return 0; | |
177 | ||
178 | p = find_lock_task_mm(p); | |
179 | if (!p) | |
180 | return 0; | |
181 | ||
182 | /* | |
183 | * Do not even consider tasks which are explicitly marked oom | |
184 | * unkillable or have been already oom reaped or the are in | |
185 | * the middle of vfork | |
186 | */ | |
187 | adj = (long)p->signal->oom_score_adj; | |
188 | if (adj == OOM_SCORE_ADJ_MIN || | |
189 | test_bit(MMF_OOM_SKIP, &p->mm->flags) || | |
190 | in_vfork(p)) { | |
191 | task_unlock(p); | |
192 | return 0; | |
193 | } | |
194 | ||
195 | /* | |
196 | * The baseline for the badness score is the proportion of RAM that each | |
197 | * task's rss, pagetable and swap space use. | |
198 | */ | |
199 | points = get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS) + | |
200 | atomic_long_read(&p->mm->nr_ptes) + mm_nr_pmds(p->mm); | |
201 | task_unlock(p); | |
202 | ||
203 | /* | |
204 | * Root processes get 3% bonus, just like the __vm_enough_memory() | |
205 | * implementation used by LSMs. | |
206 | */ | |
207 | if (has_capability_noaudit(p, CAP_SYS_ADMIN)) | |
208 | points -= (points * 3) / 100; | |
209 | ||
210 | /* Normalize to oom_score_adj units */ | |
211 | adj *= totalpages / 1000; | |
212 | points += adj; | |
213 | ||
214 | /* | |
215 | * Never return 0 for an eligible task regardless of the root bonus and | |
216 | * oom_score_adj (oom_score_adj can't be OOM_SCORE_ADJ_MIN here). | |
217 | */ | |
218 | return points > 0 ? points : 1; | |
219 | } | |
220 | ||
221 | enum oom_constraint { | |
222 | CONSTRAINT_NONE, | |
223 | CONSTRAINT_CPUSET, | |
224 | CONSTRAINT_MEMORY_POLICY, | |
225 | CONSTRAINT_MEMCG, | |
226 | }; | |
227 | ||
228 | /* | |
229 | * Determine the type of allocation constraint. | |
230 | */ | |
231 | static enum oom_constraint constrained_alloc(struct oom_control *oc) | |
232 | { | |
233 | struct zone *zone; | |
234 | struct zoneref *z; | |
235 | enum zone_type high_zoneidx = gfp_zone(oc->gfp_mask); | |
236 | bool cpuset_limited = false; | |
237 | int nid; | |
238 | ||
239 | if (is_memcg_oom(oc)) { | |
240 | oc->totalpages = mem_cgroup_get_limit(oc->memcg) ?: 1; | |
241 | return CONSTRAINT_MEMCG; | |
242 | } | |
243 | ||
244 | /* Default to all available memory */ | |
245 | oc->totalpages = totalram_pages + total_swap_pages; | |
246 | ||
247 | if (!IS_ENABLED(CONFIG_NUMA)) | |
248 | return CONSTRAINT_NONE; | |
249 | ||
250 | if (!oc->zonelist) | |
251 | return CONSTRAINT_NONE; | |
252 | /* | |
253 | * Reach here only when __GFP_NOFAIL is used. So, we should avoid | |
254 | * to kill current.We have to random task kill in this case. | |
255 | * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now. | |
256 | */ | |
257 | if (oc->gfp_mask & __GFP_THISNODE) | |
258 | return CONSTRAINT_NONE; | |
259 | ||
260 | /* | |
261 | * This is not a __GFP_THISNODE allocation, so a truncated nodemask in | |
262 | * the page allocator means a mempolicy is in effect. Cpuset policy | |
263 | * is enforced in get_page_from_freelist(). | |
264 | */ | |
265 | if (oc->nodemask && | |
266 | !nodes_subset(node_states[N_MEMORY], *oc->nodemask)) { | |
267 | oc->totalpages = total_swap_pages; | |
268 | for_each_node_mask(nid, *oc->nodemask) | |
269 | oc->totalpages += node_spanned_pages(nid); | |
270 | return CONSTRAINT_MEMORY_POLICY; | |
271 | } | |
272 | ||
273 | /* Check this allocation failure is caused by cpuset's wall function */ | |
274 | for_each_zone_zonelist_nodemask(zone, z, oc->zonelist, | |
275 | high_zoneidx, oc->nodemask) | |
276 | if (!cpuset_zone_allowed(zone, oc->gfp_mask)) | |
277 | cpuset_limited = true; | |
278 | ||
279 | if (cpuset_limited) { | |
280 | oc->totalpages = total_swap_pages; | |
281 | for_each_node_mask(nid, cpuset_current_mems_allowed) | |
282 | oc->totalpages += node_spanned_pages(nid); | |
283 | return CONSTRAINT_CPUSET; | |
284 | } | |
285 | return CONSTRAINT_NONE; | |
286 | } | |
287 | ||
288 | static int oom_evaluate_task(struct task_struct *task, void *arg) | |
289 | { | |
290 | struct oom_control *oc = arg; | |
291 | unsigned long points; | |
292 | ||
293 | if (oom_unkillable_task(task, NULL, oc->nodemask)) | |
294 | goto next; | |
295 | ||
296 | /* | |
297 | * This task already has access to memory reserves and is being killed. | |
298 | * Don't allow any other task to have access to the reserves unless | |
299 | * the task has MMF_OOM_SKIP because chances that it would release | |
300 | * any memory is quite low. | |
301 | */ | |
302 | if (!is_sysrq_oom(oc) && tsk_is_oom_victim(task)) { | |
303 | if (test_bit(MMF_OOM_SKIP, &task->signal->oom_mm->flags)) | |
304 | goto next; | |
305 | goto abort; | |
306 | } | |
307 | ||
308 | /* | |
309 | * If task is allocating a lot of memory and has been marked to be | |
310 | * killed first if it triggers an oom, then select it. | |
311 | */ | |
312 | if (oom_task_origin(task)) { | |
313 | points = ULONG_MAX; | |
314 | goto select; | |
315 | } | |
316 | ||
317 | points = oom_badness(task, NULL, oc->nodemask, oc->totalpages); | |
318 | if (!points || points < oc->chosen_points) | |
319 | goto next; | |
320 | ||
321 | /* Prefer thread group leaders for display purposes */ | |
322 | if (points == oc->chosen_points && thread_group_leader(oc->chosen)) | |
323 | goto next; | |
324 | select: | |
325 | if (oc->chosen) | |
326 | put_task_struct(oc->chosen); | |
327 | get_task_struct(task); | |
328 | oc->chosen = task; | |
329 | oc->chosen_points = points; | |
330 | next: | |
331 | return 0; | |
332 | abort: | |
333 | if (oc->chosen) | |
334 | put_task_struct(oc->chosen); | |
335 | oc->chosen = (void *)-1UL; | |
336 | return 1; | |
337 | } | |
338 | ||
339 | /* | |
340 | * Simple selection loop. We choose the process with the highest number of | |
341 | * 'points'. In case scan was aborted, oc->chosen is set to -1. | |
342 | */ | |
343 | static void select_bad_process(struct oom_control *oc) | |
344 | { | |
345 | if (is_memcg_oom(oc)) | |
346 | mem_cgroup_scan_tasks(oc->memcg, oom_evaluate_task, oc); | |
347 | else { | |
348 | struct task_struct *p; | |
349 | ||
350 | rcu_read_lock(); | |
351 | for_each_process(p) | |
352 | if (oom_evaluate_task(p, oc)) | |
353 | break; | |
354 | rcu_read_unlock(); | |
355 | } | |
356 | ||
357 | oc->chosen_points = oc->chosen_points * 1000 / oc->totalpages; | |
358 | } | |
359 | ||
360 | /** | |
361 | * dump_tasks - dump current memory state of all system tasks | |
362 | * @memcg: current's memory controller, if constrained | |
363 | * @nodemask: nodemask passed to page allocator for mempolicy ooms | |
364 | * | |
365 | * Dumps the current memory state of all eligible tasks. Tasks not in the same | |
366 | * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes | |
367 | * are not shown. | |
368 | * State information includes task's pid, uid, tgid, vm size, rss, nr_ptes, | |
369 | * swapents, oom_score_adj value, and name. | |
370 | */ | |
371 | static void dump_tasks(struct mem_cgroup *memcg, const nodemask_t *nodemask) | |
372 | { | |
373 | struct task_struct *p; | |
374 | struct task_struct *task; | |
375 | ||
376 | pr_info("[ pid ] uid tgid total_vm rss nr_ptes nr_pmds swapents oom_score_adj name\n"); | |
377 | rcu_read_lock(); | |
378 | for_each_process(p) { | |
379 | if (oom_unkillable_task(p, memcg, nodemask)) | |
380 | continue; | |
381 | ||
382 | task = find_lock_task_mm(p); | |
383 | if (!task) { | |
384 | /* | |
385 | * This is a kthread or all of p's threads have already | |
386 | * detached their mm's. There's no need to report | |
387 | * them; they can't be oom killed anyway. | |
388 | */ | |
389 | continue; | |
390 | } | |
391 | ||
392 | pr_info("[%5d] %5d %5d %8lu %8lu %7ld %7ld %8lu %5hd %s\n", | |
393 | task->pid, from_kuid(&init_user_ns, task_uid(task)), | |
394 | task->tgid, task->mm->total_vm, get_mm_rss(task->mm), | |
395 | atomic_long_read(&task->mm->nr_ptes), | |
396 | mm_nr_pmds(task->mm), | |
397 | get_mm_counter(task->mm, MM_SWAPENTS), | |
398 | task->signal->oom_score_adj, task->comm); | |
399 | task_unlock(task); | |
400 | } | |
401 | rcu_read_unlock(); | |
402 | } | |
403 | ||
404 | static void dump_header(struct oom_control *oc, struct task_struct *p) | |
405 | { | |
406 | nodemask_t *nm = (oc->nodemask) ? oc->nodemask : &cpuset_current_mems_allowed; | |
407 | ||
408 | pr_warn("%s invoked oom-killer: gfp_mask=%#x(%pGg), nodemask=%*pbl, order=%d, oom_score_adj=%hd\n", | |
409 | current->comm, oc->gfp_mask, &oc->gfp_mask, | |
410 | nodemask_pr_args(nm), oc->order, | |
411 | current->signal->oom_score_adj); | |
412 | if (!IS_ENABLED(CONFIG_COMPACTION) && oc->order) | |
413 | pr_warn("COMPACTION is disabled!!!\n"); | |
414 | ||
415 | cpuset_print_current_mems_allowed(); | |
416 | dump_stack(); | |
417 | if (oc->memcg) | |
418 | mem_cgroup_print_oom_info(oc->memcg, p); | |
419 | else | |
420 | show_mem(SHOW_MEM_FILTER_NODES, nm); | |
421 | if (sysctl_oom_dump_tasks) | |
422 | dump_tasks(oc->memcg, oc->nodemask); | |
423 | } | |
424 | ||
425 | /* | |
426 | * Number of OOM victims in flight | |
427 | */ | |
428 | static atomic_t oom_victims = ATOMIC_INIT(0); | |
429 | static DECLARE_WAIT_QUEUE_HEAD(oom_victims_wait); | |
430 | ||
431 | static bool oom_killer_disabled __read_mostly; | |
432 | ||
433 | #define K(x) ((x) << (PAGE_SHIFT-10)) | |
434 | ||
435 | /* | |
436 | * task->mm can be NULL if the task is the exited group leader. So to | |
437 | * determine whether the task is using a particular mm, we examine all the | |
438 | * task's threads: if one of those is using this mm then this task was also | |
439 | * using it. | |
440 | */ | |
441 | bool process_shares_mm(struct task_struct *p, struct mm_struct *mm) | |
442 | { | |
443 | struct task_struct *t; | |
444 | ||
445 | for_each_thread(p, t) { | |
446 | struct mm_struct *t_mm = READ_ONCE(t->mm); | |
447 | if (t_mm) | |
448 | return t_mm == mm; | |
449 | } | |
450 | return false; | |
451 | } | |
452 | ||
453 | ||
454 | #ifdef CONFIG_MMU | |
455 | /* | |
456 | * OOM Reaper kernel thread which tries to reap the memory used by the OOM | |
457 | * victim (if that is possible) to help the OOM killer to move on. | |
458 | */ | |
459 | static struct task_struct *oom_reaper_th; | |
460 | static DECLARE_WAIT_QUEUE_HEAD(oom_reaper_wait); | |
461 | static struct task_struct *oom_reaper_list; | |
462 | static DEFINE_SPINLOCK(oom_reaper_lock); | |
463 | ||
464 | static bool __oom_reap_task_mm(struct task_struct *tsk, struct mm_struct *mm) | |
465 | { | |
466 | struct mmu_gather tlb; | |
467 | struct vm_area_struct *vma; | |
468 | bool ret = true; | |
469 | ||
470 | /* | |
471 | * We have to make sure to not race with the victim exit path | |
472 | * and cause premature new oom victim selection: | |
473 | * __oom_reap_task_mm exit_mm | |
474 | * mmget_not_zero | |
475 | * mmput | |
476 | * atomic_dec_and_test | |
477 | * exit_oom_victim | |
478 | * [...] | |
479 | * out_of_memory | |
480 | * select_bad_process | |
481 | * # no TIF_MEMDIE task selects new victim | |
482 | * unmap_page_range # frees some memory | |
483 | */ | |
484 | mutex_lock(&oom_lock); | |
485 | ||
486 | if (!down_read_trylock(&mm->mmap_sem)) { | |
487 | ret = false; | |
488 | goto unlock_oom; | |
489 | } | |
490 | ||
491 | /* | |
492 | * increase mm_users only after we know we will reap something so | |
493 | * that the mmput_async is called only when we have reaped something | |
494 | * and delayed __mmput doesn't matter that much | |
495 | */ | |
496 | if (!mmget_not_zero(mm)) { | |
497 | up_read(&mm->mmap_sem); | |
498 | goto unlock_oom; | |
499 | } | |
500 | ||
501 | /* | |
502 | * Tell all users of get_user/copy_from_user etc... that the content | |
503 | * is no longer stable. No barriers really needed because unmapping | |
504 | * should imply barriers already and the reader would hit a page fault | |
505 | * if it stumbled over a reaped memory. | |
506 | */ | |
507 | set_bit(MMF_UNSTABLE, &mm->flags); | |
508 | ||
509 | tlb_gather_mmu(&tlb, mm, 0, -1); | |
510 | for (vma = mm->mmap ; vma; vma = vma->vm_next) { | |
511 | if (is_vm_hugetlb_page(vma)) | |
512 | continue; | |
513 | ||
514 | /* | |
515 | * mlocked VMAs require explicit munlocking before unmap. | |
516 | * Let's keep it simple here and skip such VMAs. | |
517 | */ | |
518 | if (vma->vm_flags & VM_LOCKED) | |
519 | continue; | |
520 | ||
521 | /* | |
522 | * Only anonymous pages have a good chance to be dropped | |
523 | * without additional steps which we cannot afford as we | |
524 | * are OOM already. | |
525 | * | |
526 | * We do not even care about fs backed pages because all | |
527 | * which are reclaimable have already been reclaimed and | |
528 | * we do not want to block exit_mmap by keeping mm ref | |
529 | * count elevated without a good reason. | |
530 | */ | |
531 | if (vma_is_anonymous(vma) || !(vma->vm_flags & VM_SHARED)) | |
532 | unmap_page_range(&tlb, vma, vma->vm_start, vma->vm_end, | |
533 | NULL); | |
534 | } | |
535 | tlb_finish_mmu(&tlb, 0, -1); | |
536 | pr_info("oom_reaper: reaped process %d (%s), now anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n", | |
537 | task_pid_nr(tsk), tsk->comm, | |
538 | K(get_mm_counter(mm, MM_ANONPAGES)), | |
539 | K(get_mm_counter(mm, MM_FILEPAGES)), | |
540 | K(get_mm_counter(mm, MM_SHMEMPAGES))); | |
541 | up_read(&mm->mmap_sem); | |
542 | ||
543 | /* | |
544 | * Drop our reference but make sure the mmput slow path is called from a | |
545 | * different context because we shouldn't risk we get stuck there and | |
546 | * put the oom_reaper out of the way. | |
547 | */ | |
548 | mmput_async(mm); | |
549 | unlock_oom: | |
550 | mutex_unlock(&oom_lock); | |
551 | return ret; | |
552 | } | |
553 | ||
554 | #define MAX_OOM_REAP_RETRIES 10 | |
555 | static void oom_reap_task(struct task_struct *tsk) | |
556 | { | |
557 | int attempts = 0; | |
558 | struct mm_struct *mm = tsk->signal->oom_mm; | |
559 | ||
560 | /* Retry the down_read_trylock(mmap_sem) a few times */ | |
561 | while (attempts++ < MAX_OOM_REAP_RETRIES && !__oom_reap_task_mm(tsk, mm)) | |
562 | schedule_timeout_idle(HZ/10); | |
563 | ||
564 | if (attempts <= MAX_OOM_REAP_RETRIES) | |
565 | goto done; | |
566 | ||
567 | ||
568 | pr_info("oom_reaper: unable to reap pid:%d (%s)\n", | |
569 | task_pid_nr(tsk), tsk->comm); | |
570 | debug_show_all_locks(); | |
571 | ||
572 | done: | |
573 | tsk->oom_reaper_list = NULL; | |
574 | ||
575 | /* | |
576 | * Hide this mm from OOM killer because it has been either reaped or | |
577 | * somebody can't call up_write(mmap_sem). | |
578 | */ | |
579 | set_bit(MMF_OOM_SKIP, &mm->flags); | |
580 | ||
581 | /* Drop a reference taken by wake_oom_reaper */ | |
582 | put_task_struct(tsk); | |
583 | } | |
584 | ||
585 | static int oom_reaper(void *unused) | |
586 | { | |
587 | while (true) { | |
588 | struct task_struct *tsk = NULL; | |
589 | ||
590 | wait_event_freezable(oom_reaper_wait, oom_reaper_list != NULL); | |
591 | spin_lock(&oom_reaper_lock); | |
592 | if (oom_reaper_list != NULL) { | |
593 | tsk = oom_reaper_list; | |
594 | oom_reaper_list = tsk->oom_reaper_list; | |
595 | } | |
596 | spin_unlock(&oom_reaper_lock); | |
597 | ||
598 | if (tsk) | |
599 | oom_reap_task(tsk); | |
600 | } | |
601 | ||
602 | return 0; | |
603 | } | |
604 | ||
605 | static void wake_oom_reaper(struct task_struct *tsk) | |
606 | { | |
607 | if (!oom_reaper_th) | |
608 | return; | |
609 | ||
610 | /* tsk is already queued? */ | |
611 | if (tsk == oom_reaper_list || tsk->oom_reaper_list) | |
612 | return; | |
613 | ||
614 | get_task_struct(tsk); | |
615 | ||
616 | spin_lock(&oom_reaper_lock); | |
617 | tsk->oom_reaper_list = oom_reaper_list; | |
618 | oom_reaper_list = tsk; | |
619 | spin_unlock(&oom_reaper_lock); | |
620 | wake_up(&oom_reaper_wait); | |
621 | } | |
622 | ||
623 | static int __init oom_init(void) | |
624 | { | |
625 | oom_reaper_th = kthread_run(oom_reaper, NULL, "oom_reaper"); | |
626 | if (IS_ERR(oom_reaper_th)) { | |
627 | pr_err("Unable to start OOM reaper %ld. Continuing regardless\n", | |
628 | PTR_ERR(oom_reaper_th)); | |
629 | oom_reaper_th = NULL; | |
630 | } | |
631 | return 0; | |
632 | } | |
633 | subsys_initcall(oom_init) | |
634 | #else | |
635 | static inline void wake_oom_reaper(struct task_struct *tsk) | |
636 | { | |
637 | } | |
638 | #endif /* CONFIG_MMU */ | |
639 | ||
640 | /** | |
641 | * mark_oom_victim - mark the given task as OOM victim | |
642 | * @tsk: task to mark | |
643 | * | |
644 | * Has to be called with oom_lock held and never after | |
645 | * oom has been disabled already. | |
646 | * | |
647 | * tsk->mm has to be non NULL and caller has to guarantee it is stable (either | |
648 | * under task_lock or operate on the current). | |
649 | */ | |
650 | static void mark_oom_victim(struct task_struct *tsk) | |
651 | { | |
652 | struct mm_struct *mm = tsk->mm; | |
653 | ||
654 | WARN_ON(oom_killer_disabled); | |
655 | /* OOM killer might race with memcg OOM */ | |
656 | if (test_and_set_tsk_thread_flag(tsk, TIF_MEMDIE)) | |
657 | return; | |
658 | ||
659 | /* oom_mm is bound to the signal struct life time. */ | |
660 | if (!cmpxchg(&tsk->signal->oom_mm, NULL, mm)) | |
661 | atomic_inc(&tsk->signal->oom_mm->mm_count); | |
662 | ||
663 | /* | |
664 | * Make sure that the task is woken up from uninterruptible sleep | |
665 | * if it is frozen because OOM killer wouldn't be able to free | |
666 | * any memory and livelock. freezing_slow_path will tell the freezer | |
667 | * that TIF_MEMDIE tasks should be ignored. | |
668 | */ | |
669 | __thaw_task(tsk); | |
670 | atomic_inc(&oom_victims); | |
671 | } | |
672 | ||
673 | /** | |
674 | * exit_oom_victim - note the exit of an OOM victim | |
675 | */ | |
676 | void exit_oom_victim(void) | |
677 | { | |
678 | clear_thread_flag(TIF_MEMDIE); | |
679 | ||
680 | if (!atomic_dec_return(&oom_victims)) | |
681 | wake_up_all(&oom_victims_wait); | |
682 | } | |
683 | ||
684 | /** | |
685 | * oom_killer_enable - enable OOM killer | |
686 | */ | |
687 | void oom_killer_enable(void) | |
688 | { | |
689 | oom_killer_disabled = false; | |
690 | } | |
691 | ||
692 | /** | |
693 | * oom_killer_disable - disable OOM killer | |
694 | * @timeout: maximum timeout to wait for oom victims in jiffies | |
695 | * | |
696 | * Forces all page allocations to fail rather than trigger OOM killer. | |
697 | * Will block and wait until all OOM victims are killed or the given | |
698 | * timeout expires. | |
699 | * | |
700 | * The function cannot be called when there are runnable user tasks because | |
701 | * the userspace would see unexpected allocation failures as a result. Any | |
702 | * new usage of this function should be consulted with MM people. | |
703 | * | |
704 | * Returns true if successful and false if the OOM killer cannot be | |
705 | * disabled. | |
706 | */ | |
707 | bool oom_killer_disable(signed long timeout) | |
708 | { | |
709 | signed long ret; | |
710 | ||
711 | /* | |
712 | * Make sure to not race with an ongoing OOM killer. Check that the | |
713 | * current is not killed (possibly due to sharing the victim's memory). | |
714 | */ | |
715 | if (mutex_lock_killable(&oom_lock)) | |
716 | return false; | |
717 | oom_killer_disabled = true; | |
718 | mutex_unlock(&oom_lock); | |
719 | ||
720 | ret = wait_event_interruptible_timeout(oom_victims_wait, | |
721 | !atomic_read(&oom_victims), timeout); | |
722 | if (ret <= 0) { | |
723 | oom_killer_enable(); | |
724 | return false; | |
725 | } | |
726 | ||
727 | return true; | |
728 | } | |
729 | ||
730 | static inline bool __task_will_free_mem(struct task_struct *task) | |
731 | { | |
732 | struct signal_struct *sig = task->signal; | |
733 | ||
734 | /* | |
735 | * A coredumping process may sleep for an extended period in exit_mm(), | |
736 | * so the oom killer cannot assume that the process will promptly exit | |
737 | * and release memory. | |
738 | */ | |
739 | if (sig->flags & SIGNAL_GROUP_COREDUMP) | |
740 | return false; | |
741 | ||
742 | if (sig->flags & SIGNAL_GROUP_EXIT) | |
743 | return true; | |
744 | ||
745 | if (thread_group_empty(task) && (task->flags & PF_EXITING)) | |
746 | return true; | |
747 | ||
748 | return false; | |
749 | } | |
750 | ||
751 | /* | |
752 | * Checks whether the given task is dying or exiting and likely to | |
753 | * release its address space. This means that all threads and processes | |
754 | * sharing the same mm have to be killed or exiting. | |
755 | * Caller has to make sure that task->mm is stable (hold task_lock or | |
756 | * it operates on the current). | |
757 | */ | |
758 | static bool task_will_free_mem(struct task_struct *task) | |
759 | { | |
760 | struct mm_struct *mm = task->mm; | |
761 | struct task_struct *p; | |
762 | bool ret = true; | |
763 | ||
764 | /* | |
765 | * Skip tasks without mm because it might have passed its exit_mm and | |
766 | * exit_oom_victim. oom_reaper could have rescued that but do not rely | |
767 | * on that for now. We can consider find_lock_task_mm in future. | |
768 | */ | |
769 | if (!mm) | |
770 | return false; | |
771 | ||
772 | if (!__task_will_free_mem(task)) | |
773 | return false; | |
774 | ||
775 | /* | |
776 | * This task has already been drained by the oom reaper so there are | |
777 | * only small chances it will free some more | |
778 | */ | |
779 | if (test_bit(MMF_OOM_SKIP, &mm->flags)) | |
780 | return false; | |
781 | ||
782 | if (atomic_read(&mm->mm_users) <= 1) | |
783 | return true; | |
784 | ||
785 | /* | |
786 | * Make sure that all tasks which share the mm with the given tasks | |
787 | * are dying as well to make sure that a) nobody pins its mm and | |
788 | * b) the task is also reapable by the oom reaper. | |
789 | */ | |
790 | rcu_read_lock(); | |
791 | for_each_process(p) { | |
792 | if (!process_shares_mm(p, mm)) | |
793 | continue; | |
794 | if (same_thread_group(task, p)) | |
795 | continue; | |
796 | ret = __task_will_free_mem(p); | |
797 | if (!ret) | |
798 | break; | |
799 | } | |
800 | rcu_read_unlock(); | |
801 | ||
802 | return ret; | |
803 | } | |
804 | ||
805 | static void oom_kill_process(struct oom_control *oc, const char *message) | |
806 | { | |
807 | struct task_struct *p = oc->chosen; | |
808 | unsigned int points = oc->chosen_points; | |
809 | struct task_struct *victim = p; | |
810 | struct task_struct *child; | |
811 | struct task_struct *t; | |
812 | struct mm_struct *mm; | |
813 | unsigned int victim_points = 0; | |
814 | static DEFINE_RATELIMIT_STATE(oom_rs, DEFAULT_RATELIMIT_INTERVAL, | |
815 | DEFAULT_RATELIMIT_BURST); | |
816 | bool can_oom_reap = true; | |
817 | ||
818 | /* | |
819 | * If the task is already exiting, don't alarm the sysadmin or kill | |
820 | * its children or threads, just set TIF_MEMDIE so it can die quickly | |
821 | */ | |
822 | task_lock(p); | |
823 | if (task_will_free_mem(p)) { | |
824 | mark_oom_victim(p); | |
825 | wake_oom_reaper(p); | |
826 | task_unlock(p); | |
827 | put_task_struct(p); | |
828 | return; | |
829 | } | |
830 | task_unlock(p); | |
831 | ||
832 | if (__ratelimit(&oom_rs)) | |
833 | dump_header(oc, p); | |
834 | ||
835 | pr_err("%s: Kill process %d (%s) score %u or sacrifice child\n", | |
836 | message, task_pid_nr(p), p->comm, points); | |
837 | ||
838 | /* | |
839 | * If any of p's children has a different mm and is eligible for kill, | |
840 | * the one with the highest oom_badness() score is sacrificed for its | |
841 | * parent. This attempts to lose the minimal amount of work done while | |
842 | * still freeing memory. | |
843 | */ | |
844 | read_lock(&tasklist_lock); | |
845 | for_each_thread(p, t) { | |
846 | list_for_each_entry(child, &t->children, sibling) { | |
847 | unsigned int child_points; | |
848 | ||
849 | if (process_shares_mm(child, p->mm)) | |
850 | continue; | |
851 | /* | |
852 | * oom_badness() returns 0 if the thread is unkillable | |
853 | */ | |
854 | child_points = oom_badness(child, | |
855 | oc->memcg, oc->nodemask, oc->totalpages); | |
856 | if (child_points > victim_points) { | |
857 | put_task_struct(victim); | |
858 | victim = child; | |
859 | victim_points = child_points; | |
860 | get_task_struct(victim); | |
861 | } | |
862 | } | |
863 | } | |
864 | read_unlock(&tasklist_lock); | |
865 | ||
866 | p = find_lock_task_mm(victim); | |
867 | if (!p) { | |
868 | put_task_struct(victim); | |
869 | return; | |
870 | } else if (victim != p) { | |
871 | get_task_struct(p); | |
872 | put_task_struct(victim); | |
873 | victim = p; | |
874 | } | |
875 | ||
876 | /* Get a reference to safely compare mm after task_unlock(victim) */ | |
877 | mm = victim->mm; | |
878 | atomic_inc(&mm->mm_count); | |
879 | /* | |
880 | * We should send SIGKILL before setting TIF_MEMDIE in order to prevent | |
881 | * the OOM victim from depleting the memory reserves from the user | |
882 | * space under its control. | |
883 | */ | |
884 | do_send_sig_info(SIGKILL, SEND_SIG_FORCED, victim, true); | |
885 | mark_oom_victim(victim); | |
886 | pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB, shmem-rss:%lukB\n", | |
887 | task_pid_nr(victim), victim->comm, K(victim->mm->total_vm), | |
888 | K(get_mm_counter(victim->mm, MM_ANONPAGES)), | |
889 | K(get_mm_counter(victim->mm, MM_FILEPAGES)), | |
890 | K(get_mm_counter(victim->mm, MM_SHMEMPAGES))); | |
891 | task_unlock(victim); | |
892 | ||
893 | /* | |
894 | * Kill all user processes sharing victim->mm in other thread groups, if | |
895 | * any. They don't get access to memory reserves, though, to avoid | |
896 | * depletion of all memory. This prevents mm->mmap_sem livelock when an | |
897 | * oom killed thread cannot exit because it requires the semaphore and | |
898 | * its contended by another thread trying to allocate memory itself. | |
899 | * That thread will now get access to memory reserves since it has a | |
900 | * pending fatal signal. | |
901 | */ | |
902 | rcu_read_lock(); | |
903 | for_each_process(p) { | |
904 | if (!process_shares_mm(p, mm)) | |
905 | continue; | |
906 | if (same_thread_group(p, victim)) | |
907 | continue; | |
908 | if (is_global_init(p)) { | |
909 | can_oom_reap = false; | |
910 | set_bit(MMF_OOM_SKIP, &mm->flags); | |
911 | pr_info("oom killer %d (%s) has mm pinned by %d (%s)\n", | |
912 | task_pid_nr(victim), victim->comm, | |
913 | task_pid_nr(p), p->comm); | |
914 | continue; | |
915 | } | |
916 | /* | |
917 | * No use_mm() user needs to read from the userspace so we are | |
918 | * ok to reap it. | |
919 | */ | |
920 | if (unlikely(p->flags & PF_KTHREAD)) | |
921 | continue; | |
922 | do_send_sig_info(SIGKILL, SEND_SIG_FORCED, p, true); | |
923 | } | |
924 | rcu_read_unlock(); | |
925 | ||
926 | if (can_oom_reap) | |
927 | wake_oom_reaper(victim); | |
928 | ||
929 | mmdrop(mm); | |
930 | put_task_struct(victim); | |
931 | } | |
932 | #undef K | |
933 | ||
934 | /* | |
935 | * Determines whether the kernel must panic because of the panic_on_oom sysctl. | |
936 | */ | |
937 | static void check_panic_on_oom(struct oom_control *oc, | |
938 | enum oom_constraint constraint) | |
939 | { | |
940 | if (likely(!sysctl_panic_on_oom)) | |
941 | return; | |
942 | if (sysctl_panic_on_oom != 2) { | |
943 | /* | |
944 | * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel | |
945 | * does not panic for cpuset, mempolicy, or memcg allocation | |
946 | * failures. | |
947 | */ | |
948 | if (constraint != CONSTRAINT_NONE) | |
949 | return; | |
950 | } | |
951 | /* Do not panic for oom kills triggered by sysrq */ | |
952 | if (is_sysrq_oom(oc)) | |
953 | return; | |
954 | dump_header(oc, NULL); | |
955 | panic("Out of memory: %s panic_on_oom is enabled\n", | |
956 | sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide"); | |
957 | } | |
958 | ||
959 | static BLOCKING_NOTIFIER_HEAD(oom_notify_list); | |
960 | ||
961 | int register_oom_notifier(struct notifier_block *nb) | |
962 | { | |
963 | return blocking_notifier_chain_register(&oom_notify_list, nb); | |
964 | } | |
965 | EXPORT_SYMBOL_GPL(register_oom_notifier); | |
966 | ||
967 | int unregister_oom_notifier(struct notifier_block *nb) | |
968 | { | |
969 | return blocking_notifier_chain_unregister(&oom_notify_list, nb); | |
970 | } | |
971 | EXPORT_SYMBOL_GPL(unregister_oom_notifier); | |
972 | ||
973 | /** | |
974 | * out_of_memory - kill the "best" process when we run out of memory | |
975 | * @oc: pointer to struct oom_control | |
976 | * | |
977 | * If we run out of memory, we have the choice between either | |
978 | * killing a random task (bad), letting the system crash (worse) | |
979 | * OR try to be smart about which process to kill. Note that we | |
980 | * don't have to be perfect here, we just have to be good. | |
981 | */ | |
982 | bool out_of_memory(struct oom_control *oc) | |
983 | { | |
984 | unsigned long freed = 0; | |
985 | enum oom_constraint constraint = CONSTRAINT_NONE; | |
986 | ||
987 | if (oom_killer_disabled) | |
988 | return false; | |
989 | ||
990 | if (!is_memcg_oom(oc)) { | |
991 | blocking_notifier_call_chain(&oom_notify_list, 0, &freed); | |
992 | if (freed > 0) | |
993 | /* Got some memory back in the last second. */ | |
994 | return true; | |
995 | } | |
996 | ||
997 | /* | |
998 | * If current has a pending SIGKILL or is exiting, then automatically | |
999 | * select it. The goal is to allow it to allocate so that it may | |
1000 | * quickly exit and free its memory. | |
1001 | */ | |
1002 | if (task_will_free_mem(current)) { | |
1003 | mark_oom_victim(current); | |
1004 | wake_oom_reaper(current); | |
1005 | return true; | |
1006 | } | |
1007 | ||
1008 | /* | |
1009 | * The OOM killer does not compensate for IO-less reclaim. | |
1010 | * pagefault_out_of_memory lost its gfp context so we have to | |
1011 | * make sure exclude 0 mask - all other users should have at least | |
1012 | * ___GFP_DIRECT_RECLAIM to get here. | |
1013 | */ | |
1014 | if (oc->gfp_mask && !(oc->gfp_mask & __GFP_FS)) | |
1015 | return true; | |
1016 | ||
1017 | /* | |
1018 | * Check if there were limitations on the allocation (only relevant for | |
1019 | * NUMA and memcg) that may require different handling. | |
1020 | */ | |
1021 | constraint = constrained_alloc(oc); | |
1022 | if (constraint != CONSTRAINT_MEMORY_POLICY) | |
1023 | oc->nodemask = NULL; | |
1024 | check_panic_on_oom(oc, constraint); | |
1025 | ||
1026 | if (!is_memcg_oom(oc) && sysctl_oom_kill_allocating_task && | |
1027 | current->mm && !oom_unkillable_task(current, NULL, oc->nodemask) && | |
1028 | current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) { | |
1029 | get_task_struct(current); | |
1030 | oc->chosen = current; | |
1031 | oom_kill_process(oc, "Out of memory (oom_kill_allocating_task)"); | |
1032 | return true; | |
1033 | } | |
1034 | ||
1035 | select_bad_process(oc); | |
1036 | /* Found nothing?!?! Either we hang forever, or we panic. */ | |
1037 | if (!oc->chosen && !is_sysrq_oom(oc) && !is_memcg_oom(oc)) { | |
1038 | dump_header(oc, NULL); | |
1039 | panic("Out of memory and no killable processes...\n"); | |
1040 | } | |
1041 | if (oc->chosen && oc->chosen != (void *)-1UL) { | |
1042 | oom_kill_process(oc, !is_memcg_oom(oc) ? "Out of memory" : | |
1043 | "Memory cgroup out of memory"); | |
1044 | /* | |
1045 | * Give the killed process a good chance to exit before trying | |
1046 | * to allocate memory again. | |
1047 | */ | |
1048 | schedule_timeout_killable(1); | |
1049 | } | |
1050 | return !!oc->chosen; | |
1051 | } | |
1052 | ||
1053 | /* | |
1054 | * The pagefault handler calls here because it is out of memory, so kill a | |
1055 | * memory-hogging task. If oom_lock is held by somebody else, a parallel oom | |
1056 | * killing is already in progress so do nothing. | |
1057 | */ | |
1058 | void pagefault_out_of_memory(void) | |
1059 | { | |
1060 | struct oom_control oc = { | |
1061 | .zonelist = NULL, | |
1062 | .nodemask = NULL, | |
1063 | .memcg = NULL, | |
1064 | .gfp_mask = 0, | |
1065 | .order = 0, | |
1066 | }; | |
1067 | ||
1068 | if (mem_cgroup_oom_synchronize(true)) | |
1069 | return; | |
1070 | ||
1071 | if (!mutex_trylock(&oom_lock)) | |
1072 | return; | |
1073 | out_of_memory(&oc); | |
1074 | mutex_unlock(&oom_lock); | |
1075 | } |