| 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/module.h> |
| 30 | #include <linux/notifier.h> |
| 31 | #include <linux/memcontrol.h> |
| 32 | #include <linux/mempolicy.h> |
| 33 | #include <linux/security.h> |
| 34 | |
| 35 | int sysctl_panic_on_oom; |
| 36 | int sysctl_oom_kill_allocating_task; |
| 37 | int sysctl_oom_dump_tasks = 1; |
| 38 | static DEFINE_SPINLOCK(zone_scan_lock); |
| 39 | |
| 40 | #ifdef CONFIG_NUMA |
| 41 | /** |
| 42 | * has_intersects_mems_allowed() - check task eligiblity for kill |
| 43 | * @tsk: task struct of which task to consider |
| 44 | * @mask: nodemask passed to page allocator for mempolicy ooms |
| 45 | * |
| 46 | * Task eligibility is determined by whether or not a candidate task, @tsk, |
| 47 | * shares the same mempolicy nodes as current if it is bound by such a policy |
| 48 | * and whether or not it has the same set of allowed cpuset nodes. |
| 49 | */ |
| 50 | static bool has_intersects_mems_allowed(struct task_struct *tsk, |
| 51 | const nodemask_t *mask) |
| 52 | { |
| 53 | struct task_struct *start = tsk; |
| 54 | |
| 55 | do { |
| 56 | if (mask) { |
| 57 | /* |
| 58 | * If this is a mempolicy constrained oom, tsk's |
| 59 | * cpuset is irrelevant. Only return true if its |
| 60 | * mempolicy intersects current, otherwise it may be |
| 61 | * needlessly killed. |
| 62 | */ |
| 63 | if (mempolicy_nodemask_intersects(tsk, mask)) |
| 64 | return true; |
| 65 | } else { |
| 66 | /* |
| 67 | * This is not a mempolicy constrained oom, so only |
| 68 | * check the mems of tsk's cpuset. |
| 69 | */ |
| 70 | if (cpuset_mems_allowed_intersects(current, tsk)) |
| 71 | return true; |
| 72 | } |
| 73 | } while_each_thread(start, tsk); |
| 74 | |
| 75 | return false; |
| 76 | } |
| 77 | #else |
| 78 | static bool has_intersects_mems_allowed(struct task_struct *tsk, |
| 79 | const nodemask_t *mask) |
| 80 | { |
| 81 | return true; |
| 82 | } |
| 83 | #endif /* CONFIG_NUMA */ |
| 84 | |
| 85 | /* |
| 86 | * If this is a system OOM (not a memcg OOM) and the task selected to be |
| 87 | * killed is not already running at high (RT) priorities, speed up the |
| 88 | * recovery by boosting the dying task to the lowest FIFO priority. |
| 89 | * That helps with the recovery and avoids interfering with RT tasks. |
| 90 | */ |
| 91 | static void boost_dying_task_prio(struct task_struct *p, |
| 92 | struct mem_cgroup *mem) |
| 93 | { |
| 94 | struct sched_param param = { .sched_priority = 1 }; |
| 95 | |
| 96 | if (mem) |
| 97 | return; |
| 98 | |
| 99 | if (!rt_task(p)) |
| 100 | sched_setscheduler_nocheck(p, SCHED_FIFO, ¶m); |
| 101 | } |
| 102 | |
| 103 | /* |
| 104 | * The process p may have detached its own ->mm while exiting or through |
| 105 | * use_mm(), but one or more of its subthreads may still have a valid |
| 106 | * pointer. Return p, or any of its subthreads with a valid ->mm, with |
| 107 | * task_lock() held. |
| 108 | */ |
| 109 | struct task_struct *find_lock_task_mm(struct task_struct *p) |
| 110 | { |
| 111 | struct task_struct *t = p; |
| 112 | |
| 113 | do { |
| 114 | task_lock(t); |
| 115 | if (likely(t->mm)) |
| 116 | return t; |
| 117 | task_unlock(t); |
| 118 | } while_each_thread(p, t); |
| 119 | |
| 120 | return NULL; |
| 121 | } |
| 122 | |
| 123 | /* return true if the task is not adequate as candidate victim task. */ |
| 124 | static bool oom_unkillable_task(struct task_struct *p, struct mem_cgroup *mem, |
| 125 | const nodemask_t *nodemask) |
| 126 | { |
| 127 | if (is_global_init(p)) |
| 128 | return true; |
| 129 | if (p->flags & PF_KTHREAD) |
| 130 | return true; |
| 131 | |
| 132 | /* When mem_cgroup_out_of_memory() and p is not member of the group */ |
| 133 | if (mem && !task_in_mem_cgroup(p, mem)) |
| 134 | return true; |
| 135 | |
| 136 | /* p may not have freeable memory in nodemask */ |
| 137 | if (!has_intersects_mems_allowed(p, nodemask)) |
| 138 | return true; |
| 139 | |
| 140 | return false; |
| 141 | } |
| 142 | |
| 143 | /** |
| 144 | * oom_badness - heuristic function to determine which candidate task to kill |
| 145 | * @p: task struct of which task we should calculate |
| 146 | * @totalpages: total present RAM allowed for page allocation |
| 147 | * |
| 148 | * The heuristic for determining which task to kill is made to be as simple and |
| 149 | * predictable as possible. The goal is to return the highest value for the |
| 150 | * task consuming the most memory to avoid subsequent oom failures. |
| 151 | */ |
| 152 | unsigned int oom_badness(struct task_struct *p, struct mem_cgroup *mem, |
| 153 | const nodemask_t *nodemask, unsigned long totalpages) |
| 154 | { |
| 155 | int points; |
| 156 | |
| 157 | if (oom_unkillable_task(p, mem, nodemask)) |
| 158 | return 0; |
| 159 | |
| 160 | p = find_lock_task_mm(p); |
| 161 | if (!p) |
| 162 | return 0; |
| 163 | |
| 164 | /* |
| 165 | * Shortcut check for OOM_SCORE_ADJ_MIN so the entire heuristic doesn't |
| 166 | * need to be executed for something that cannot be killed. |
| 167 | */ |
| 168 | if (p->signal->oom_score_adj == OOM_SCORE_ADJ_MIN) { |
| 169 | task_unlock(p); |
| 170 | return 0; |
| 171 | } |
| 172 | |
| 173 | /* |
| 174 | * When the PF_OOM_ORIGIN bit is set, it indicates the task should have |
| 175 | * priority for oom killing. |
| 176 | */ |
| 177 | if (p->flags & PF_OOM_ORIGIN) { |
| 178 | task_unlock(p); |
| 179 | return 1000; |
| 180 | } |
| 181 | |
| 182 | /* |
| 183 | * The memory controller may have a limit of 0 bytes, so avoid a divide |
| 184 | * by zero, if necessary. |
| 185 | */ |
| 186 | if (!totalpages) |
| 187 | totalpages = 1; |
| 188 | |
| 189 | /* |
| 190 | * The baseline for the badness score is the proportion of RAM that each |
| 191 | * task's rss and swap space use. |
| 192 | */ |
| 193 | points = (get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS)) * 1000 / |
| 194 | totalpages; |
| 195 | task_unlock(p); |
| 196 | |
| 197 | /* |
| 198 | * Root processes get 3% bonus, just like the __vm_enough_memory() |
| 199 | * implementation used by LSMs. |
| 200 | */ |
| 201 | if (has_capability_noaudit(p, CAP_SYS_ADMIN)) |
| 202 | points -= 30; |
| 203 | |
| 204 | /* |
| 205 | * /proc/pid/oom_score_adj ranges from -1000 to +1000 such that it may |
| 206 | * either completely disable oom killing or always prefer a certain |
| 207 | * task. |
| 208 | */ |
| 209 | points += p->signal->oom_score_adj; |
| 210 | |
| 211 | if (points < 0) |
| 212 | return 0; |
| 213 | return (points < 1000) ? points : 1000; |
| 214 | } |
| 215 | |
| 216 | /* |
| 217 | * Determine the type of allocation constraint. |
| 218 | */ |
| 219 | #ifdef CONFIG_NUMA |
| 220 | static enum oom_constraint constrained_alloc(struct zonelist *zonelist, |
| 221 | gfp_t gfp_mask, nodemask_t *nodemask, |
| 222 | unsigned long *totalpages) |
| 223 | { |
| 224 | struct zone *zone; |
| 225 | struct zoneref *z; |
| 226 | enum zone_type high_zoneidx = gfp_zone(gfp_mask); |
| 227 | bool cpuset_limited = false; |
| 228 | int nid; |
| 229 | |
| 230 | /* Default to all available memory */ |
| 231 | *totalpages = totalram_pages + total_swap_pages; |
| 232 | |
| 233 | if (!zonelist) |
| 234 | return CONSTRAINT_NONE; |
| 235 | /* |
| 236 | * Reach here only when __GFP_NOFAIL is used. So, we should avoid |
| 237 | * to kill current.We have to random task kill in this case. |
| 238 | * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now. |
| 239 | */ |
| 240 | if (gfp_mask & __GFP_THISNODE) |
| 241 | return CONSTRAINT_NONE; |
| 242 | |
| 243 | /* |
| 244 | * This is not a __GFP_THISNODE allocation, so a truncated nodemask in |
| 245 | * the page allocator means a mempolicy is in effect. Cpuset policy |
| 246 | * is enforced in get_page_from_freelist(). |
| 247 | */ |
| 248 | if (nodemask && !nodes_subset(node_states[N_HIGH_MEMORY], *nodemask)) { |
| 249 | *totalpages = total_swap_pages; |
| 250 | for_each_node_mask(nid, *nodemask) |
| 251 | *totalpages += node_spanned_pages(nid); |
| 252 | return CONSTRAINT_MEMORY_POLICY; |
| 253 | } |
| 254 | |
| 255 | /* Check this allocation failure is caused by cpuset's wall function */ |
| 256 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
| 257 | high_zoneidx, nodemask) |
| 258 | if (!cpuset_zone_allowed_softwall(zone, gfp_mask)) |
| 259 | cpuset_limited = true; |
| 260 | |
| 261 | if (cpuset_limited) { |
| 262 | *totalpages = total_swap_pages; |
| 263 | for_each_node_mask(nid, cpuset_current_mems_allowed) |
| 264 | *totalpages += node_spanned_pages(nid); |
| 265 | return CONSTRAINT_CPUSET; |
| 266 | } |
| 267 | return CONSTRAINT_NONE; |
| 268 | } |
| 269 | #else |
| 270 | static enum oom_constraint constrained_alloc(struct zonelist *zonelist, |
| 271 | gfp_t gfp_mask, nodemask_t *nodemask, |
| 272 | unsigned long *totalpages) |
| 273 | { |
| 274 | *totalpages = totalram_pages + total_swap_pages; |
| 275 | return CONSTRAINT_NONE; |
| 276 | } |
| 277 | #endif |
| 278 | |
| 279 | /* |
| 280 | * Simple selection loop. We chose the process with the highest |
| 281 | * number of 'points'. We expect the caller will lock the tasklist. |
| 282 | * |
| 283 | * (not docbooked, we don't want this one cluttering up the manual) |
| 284 | */ |
| 285 | static struct task_struct *select_bad_process(unsigned int *ppoints, |
| 286 | unsigned long totalpages, struct mem_cgroup *mem, |
| 287 | const nodemask_t *nodemask) |
| 288 | { |
| 289 | struct task_struct *p; |
| 290 | struct task_struct *chosen = NULL; |
| 291 | *ppoints = 0; |
| 292 | |
| 293 | for_each_process(p) { |
| 294 | unsigned int points; |
| 295 | |
| 296 | if (oom_unkillable_task(p, mem, nodemask)) |
| 297 | continue; |
| 298 | |
| 299 | /* |
| 300 | * This task already has access to memory reserves and is |
| 301 | * being killed. Don't allow any other task access to the |
| 302 | * memory reserve. |
| 303 | * |
| 304 | * Note: this may have a chance of deadlock if it gets |
| 305 | * blocked waiting for another task which itself is waiting |
| 306 | * for memory. Is there a better alternative? |
| 307 | */ |
| 308 | if (test_tsk_thread_flag(p, TIF_MEMDIE)) |
| 309 | return ERR_PTR(-1UL); |
| 310 | |
| 311 | /* |
| 312 | * This is in the process of releasing memory so wait for it |
| 313 | * to finish before killing some other task by mistake. |
| 314 | * |
| 315 | * However, if p is the current task, we allow the 'kill' to |
| 316 | * go ahead if it is exiting: this will simply set TIF_MEMDIE, |
| 317 | * which will allow it to gain access to memory reserves in |
| 318 | * the process of exiting and releasing its resources. |
| 319 | * Otherwise we could get an easy OOM deadlock. |
| 320 | */ |
| 321 | if (thread_group_empty(p) && (p->flags & PF_EXITING) && p->mm) { |
| 322 | if (p != current) |
| 323 | return ERR_PTR(-1UL); |
| 324 | |
| 325 | chosen = p; |
| 326 | *ppoints = 1000; |
| 327 | } |
| 328 | |
| 329 | points = oom_badness(p, mem, nodemask, totalpages); |
| 330 | if (points > *ppoints) { |
| 331 | chosen = p; |
| 332 | *ppoints = points; |
| 333 | } |
| 334 | } |
| 335 | |
| 336 | return chosen; |
| 337 | } |
| 338 | |
| 339 | /** |
| 340 | * dump_tasks - dump current memory state of all system tasks |
| 341 | * @mem: current's memory controller, if constrained |
| 342 | * |
| 343 | * Dumps the current memory state of all system tasks, excluding kernel threads. |
| 344 | * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj |
| 345 | * value, oom_score_adj value, and name. |
| 346 | * |
| 347 | * If the actual is non-NULL, only tasks that are a member of the mem_cgroup are |
| 348 | * shown. |
| 349 | * |
| 350 | * Call with tasklist_lock read-locked. |
| 351 | */ |
| 352 | static void dump_tasks(const struct mem_cgroup *mem) |
| 353 | { |
| 354 | struct task_struct *p; |
| 355 | struct task_struct *task; |
| 356 | |
| 357 | pr_info("[ pid ] uid tgid total_vm rss cpu oom_adj oom_score_adj name\n"); |
| 358 | for_each_process(p) { |
| 359 | if (p->flags & PF_KTHREAD) |
| 360 | continue; |
| 361 | if (mem && !task_in_mem_cgroup(p, mem)) |
| 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 %3u %3d %5d %s\n", |
| 375 | task->pid, task_uid(task), task->tgid, |
| 376 | task->mm->total_vm, get_mm_rss(task->mm), |
| 377 | task_cpu(task), task->signal->oom_adj, |
| 378 | task->signal->oom_score_adj, task->comm); |
| 379 | task_unlock(task); |
| 380 | } |
| 381 | } |
| 382 | |
| 383 | static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order, |
| 384 | struct mem_cgroup *mem) |
| 385 | { |
| 386 | task_lock(current); |
| 387 | pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, " |
| 388 | "oom_adj=%d, oom_score_adj=%d\n", |
| 389 | current->comm, gfp_mask, order, current->signal->oom_adj, |
| 390 | current->signal->oom_score_adj); |
| 391 | cpuset_print_task_mems_allowed(current); |
| 392 | task_unlock(current); |
| 393 | dump_stack(); |
| 394 | mem_cgroup_print_oom_info(mem, p); |
| 395 | show_mem(); |
| 396 | if (sysctl_oom_dump_tasks) |
| 397 | dump_tasks(mem); |
| 398 | } |
| 399 | |
| 400 | #define K(x) ((x) << (PAGE_SHIFT-10)) |
| 401 | static int oom_kill_task(struct task_struct *p, struct mem_cgroup *mem) |
| 402 | { |
| 403 | p = find_lock_task_mm(p); |
| 404 | if (!p) |
| 405 | return 1; |
| 406 | |
| 407 | pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB\n", |
| 408 | task_pid_nr(p), p->comm, K(p->mm->total_vm), |
| 409 | K(get_mm_counter(p->mm, MM_ANONPAGES)), |
| 410 | K(get_mm_counter(p->mm, MM_FILEPAGES))); |
| 411 | task_unlock(p); |
| 412 | |
| 413 | |
| 414 | set_tsk_thread_flag(p, TIF_MEMDIE); |
| 415 | force_sig(SIGKILL, p); |
| 416 | |
| 417 | /* |
| 418 | * We give our sacrificial lamb high priority and access to |
| 419 | * all the memory it needs. That way it should be able to |
| 420 | * exit() and clear out its resources quickly... |
| 421 | */ |
| 422 | boost_dying_task_prio(p, mem); |
| 423 | |
| 424 | return 0; |
| 425 | } |
| 426 | #undef K |
| 427 | |
| 428 | static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order, |
| 429 | unsigned int points, unsigned long totalpages, |
| 430 | struct mem_cgroup *mem, nodemask_t *nodemask, |
| 431 | const char *message) |
| 432 | { |
| 433 | struct task_struct *victim = p; |
| 434 | struct task_struct *child; |
| 435 | struct task_struct *t = p; |
| 436 | unsigned int victim_points = 0; |
| 437 | |
| 438 | if (printk_ratelimit()) |
| 439 | dump_header(p, gfp_mask, order, mem); |
| 440 | |
| 441 | /* |
| 442 | * If the task is already exiting, don't alarm the sysadmin or kill |
| 443 | * its children or threads, just set TIF_MEMDIE so it can die quickly |
| 444 | */ |
| 445 | if (p->flags & PF_EXITING) { |
| 446 | set_tsk_thread_flag(p, TIF_MEMDIE); |
| 447 | boost_dying_task_prio(p, mem); |
| 448 | return 0; |
| 449 | } |
| 450 | |
| 451 | task_lock(p); |
| 452 | pr_err("%s: Kill process %d (%s) score %d or sacrifice child\n", |
| 453 | message, task_pid_nr(p), p->comm, points); |
| 454 | task_unlock(p); |
| 455 | |
| 456 | /* |
| 457 | * If any of p's children has a different mm and is eligible for kill, |
| 458 | * the one with the highest badness() score is sacrificed for its |
| 459 | * parent. This attempts to lose the minimal amount of work done while |
| 460 | * still freeing memory. |
| 461 | */ |
| 462 | do { |
| 463 | list_for_each_entry(child, &t->children, sibling) { |
| 464 | unsigned int child_points; |
| 465 | |
| 466 | /* |
| 467 | * oom_badness() returns 0 if the thread is unkillable |
| 468 | */ |
| 469 | child_points = oom_badness(child, mem, nodemask, |
| 470 | totalpages); |
| 471 | if (child_points > victim_points) { |
| 472 | victim = child; |
| 473 | victim_points = child_points; |
| 474 | } |
| 475 | } |
| 476 | } while_each_thread(p, t); |
| 477 | |
| 478 | return oom_kill_task(victim, mem); |
| 479 | } |
| 480 | |
| 481 | /* |
| 482 | * Determines whether the kernel must panic because of the panic_on_oom sysctl. |
| 483 | */ |
| 484 | static void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask, |
| 485 | int order) |
| 486 | { |
| 487 | if (likely(!sysctl_panic_on_oom)) |
| 488 | return; |
| 489 | if (sysctl_panic_on_oom != 2) { |
| 490 | /* |
| 491 | * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel |
| 492 | * does not panic for cpuset, mempolicy, or memcg allocation |
| 493 | * failures. |
| 494 | */ |
| 495 | if (constraint != CONSTRAINT_NONE) |
| 496 | return; |
| 497 | } |
| 498 | read_lock(&tasklist_lock); |
| 499 | dump_header(NULL, gfp_mask, order, NULL); |
| 500 | read_unlock(&tasklist_lock); |
| 501 | panic("Out of memory: %s panic_on_oom is enabled\n", |
| 502 | sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide"); |
| 503 | } |
| 504 | |
| 505 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR |
| 506 | void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask) |
| 507 | { |
| 508 | unsigned long limit; |
| 509 | unsigned int points = 0; |
| 510 | struct task_struct *p; |
| 511 | |
| 512 | check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, 0); |
| 513 | limit = mem_cgroup_get_limit(mem) >> PAGE_SHIFT; |
| 514 | read_lock(&tasklist_lock); |
| 515 | retry: |
| 516 | p = select_bad_process(&points, limit, mem, NULL); |
| 517 | if (!p || PTR_ERR(p) == -1UL) |
| 518 | goto out; |
| 519 | |
| 520 | if (oom_kill_process(p, gfp_mask, 0, points, limit, mem, NULL, |
| 521 | "Memory cgroup out of memory")) |
| 522 | goto retry; |
| 523 | out: |
| 524 | read_unlock(&tasklist_lock); |
| 525 | } |
| 526 | #endif |
| 527 | |
| 528 | static BLOCKING_NOTIFIER_HEAD(oom_notify_list); |
| 529 | |
| 530 | int register_oom_notifier(struct notifier_block *nb) |
| 531 | { |
| 532 | return blocking_notifier_chain_register(&oom_notify_list, nb); |
| 533 | } |
| 534 | EXPORT_SYMBOL_GPL(register_oom_notifier); |
| 535 | |
| 536 | int unregister_oom_notifier(struct notifier_block *nb) |
| 537 | { |
| 538 | return blocking_notifier_chain_unregister(&oom_notify_list, nb); |
| 539 | } |
| 540 | EXPORT_SYMBOL_GPL(unregister_oom_notifier); |
| 541 | |
| 542 | /* |
| 543 | * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero |
| 544 | * if a parallel OOM killing is already taking place that includes a zone in |
| 545 | * the zonelist. Otherwise, locks all zones in the zonelist and returns 1. |
| 546 | */ |
| 547 | int try_set_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask) |
| 548 | { |
| 549 | struct zoneref *z; |
| 550 | struct zone *zone; |
| 551 | int ret = 1; |
| 552 | |
| 553 | spin_lock(&zone_scan_lock); |
| 554 | for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) { |
| 555 | if (zone_is_oom_locked(zone)) { |
| 556 | ret = 0; |
| 557 | goto out; |
| 558 | } |
| 559 | } |
| 560 | |
| 561 | for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) { |
| 562 | /* |
| 563 | * Lock each zone in the zonelist under zone_scan_lock so a |
| 564 | * parallel invocation of try_set_zonelist_oom() doesn't succeed |
| 565 | * when it shouldn't. |
| 566 | */ |
| 567 | zone_set_flag(zone, ZONE_OOM_LOCKED); |
| 568 | } |
| 569 | |
| 570 | out: |
| 571 | spin_unlock(&zone_scan_lock); |
| 572 | return ret; |
| 573 | } |
| 574 | |
| 575 | /* |
| 576 | * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed |
| 577 | * allocation attempts with zonelists containing them may now recall the OOM |
| 578 | * killer, if necessary. |
| 579 | */ |
| 580 | void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask) |
| 581 | { |
| 582 | struct zoneref *z; |
| 583 | struct zone *zone; |
| 584 | |
| 585 | spin_lock(&zone_scan_lock); |
| 586 | for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) { |
| 587 | zone_clear_flag(zone, ZONE_OOM_LOCKED); |
| 588 | } |
| 589 | spin_unlock(&zone_scan_lock); |
| 590 | } |
| 591 | |
| 592 | /* |
| 593 | * Try to acquire the oom killer lock for all system zones. Returns zero if a |
| 594 | * parallel oom killing is taking place, otherwise locks all zones and returns |
| 595 | * non-zero. |
| 596 | */ |
| 597 | static int try_set_system_oom(void) |
| 598 | { |
| 599 | struct zone *zone; |
| 600 | int ret = 1; |
| 601 | |
| 602 | spin_lock(&zone_scan_lock); |
| 603 | for_each_populated_zone(zone) |
| 604 | if (zone_is_oom_locked(zone)) { |
| 605 | ret = 0; |
| 606 | goto out; |
| 607 | } |
| 608 | for_each_populated_zone(zone) |
| 609 | zone_set_flag(zone, ZONE_OOM_LOCKED); |
| 610 | out: |
| 611 | spin_unlock(&zone_scan_lock); |
| 612 | return ret; |
| 613 | } |
| 614 | |
| 615 | /* |
| 616 | * Clears ZONE_OOM_LOCKED for all system zones so that failed allocation |
| 617 | * attempts or page faults may now recall the oom killer, if necessary. |
| 618 | */ |
| 619 | static void clear_system_oom(void) |
| 620 | { |
| 621 | struct zone *zone; |
| 622 | |
| 623 | spin_lock(&zone_scan_lock); |
| 624 | for_each_populated_zone(zone) |
| 625 | zone_clear_flag(zone, ZONE_OOM_LOCKED); |
| 626 | spin_unlock(&zone_scan_lock); |
| 627 | } |
| 628 | |
| 629 | /** |
| 630 | * out_of_memory - kill the "best" process when we run out of memory |
| 631 | * @zonelist: zonelist pointer |
| 632 | * @gfp_mask: memory allocation flags |
| 633 | * @order: amount of memory being requested as a power of 2 |
| 634 | * @nodemask: nodemask passed to page allocator |
| 635 | * |
| 636 | * If we run out of memory, we have the choice between either |
| 637 | * killing a random task (bad), letting the system crash (worse) |
| 638 | * OR try to be smart about which process to kill. Note that we |
| 639 | * don't have to be perfect here, we just have to be good. |
| 640 | */ |
| 641 | void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, |
| 642 | int order, nodemask_t *nodemask) |
| 643 | { |
| 644 | struct task_struct *p; |
| 645 | unsigned long totalpages; |
| 646 | unsigned long freed = 0; |
| 647 | unsigned int points; |
| 648 | enum oom_constraint constraint = CONSTRAINT_NONE; |
| 649 | int killed = 0; |
| 650 | |
| 651 | blocking_notifier_call_chain(&oom_notify_list, 0, &freed); |
| 652 | if (freed > 0) |
| 653 | /* Got some memory back in the last second. */ |
| 654 | return; |
| 655 | |
| 656 | /* |
| 657 | * If current has a pending SIGKILL, then automatically select it. The |
| 658 | * goal is to allow it to allocate so that it may quickly exit and free |
| 659 | * its memory. |
| 660 | */ |
| 661 | if (fatal_signal_pending(current)) { |
| 662 | set_thread_flag(TIF_MEMDIE); |
| 663 | boost_dying_task_prio(current, NULL); |
| 664 | return; |
| 665 | } |
| 666 | |
| 667 | /* |
| 668 | * Check if there were limitations on the allocation (only relevant for |
| 669 | * NUMA) that may require different handling. |
| 670 | */ |
| 671 | constraint = constrained_alloc(zonelist, gfp_mask, nodemask, |
| 672 | &totalpages); |
| 673 | check_panic_on_oom(constraint, gfp_mask, order); |
| 674 | |
| 675 | read_lock(&tasklist_lock); |
| 676 | if (sysctl_oom_kill_allocating_task && |
| 677 | !oom_unkillable_task(current, NULL, nodemask) && |
| 678 | (current->signal->oom_adj != OOM_DISABLE)) { |
| 679 | /* |
| 680 | * oom_kill_process() needs tasklist_lock held. If it returns |
| 681 | * non-zero, current could not be killed so we must fallback to |
| 682 | * the tasklist scan. |
| 683 | */ |
| 684 | if (!oom_kill_process(current, gfp_mask, order, 0, totalpages, |
| 685 | NULL, nodemask, |
| 686 | "Out of memory (oom_kill_allocating_task)")) |
| 687 | goto out; |
| 688 | } |
| 689 | |
| 690 | retry: |
| 691 | p = select_bad_process(&points, totalpages, NULL, |
| 692 | constraint == CONSTRAINT_MEMORY_POLICY ? nodemask : |
| 693 | NULL); |
| 694 | if (PTR_ERR(p) == -1UL) |
| 695 | goto out; |
| 696 | |
| 697 | /* Found nothing?!?! Either we hang forever, or we panic. */ |
| 698 | if (!p) { |
| 699 | dump_header(NULL, gfp_mask, order, NULL); |
| 700 | read_unlock(&tasklist_lock); |
| 701 | panic("Out of memory and no killable processes...\n"); |
| 702 | } |
| 703 | |
| 704 | if (oom_kill_process(p, gfp_mask, order, points, totalpages, NULL, |
| 705 | nodemask, "Out of memory")) |
| 706 | goto retry; |
| 707 | killed = 1; |
| 708 | out: |
| 709 | read_unlock(&tasklist_lock); |
| 710 | |
| 711 | /* |
| 712 | * Give "p" a good chance of killing itself before we |
| 713 | * retry to allocate memory unless "p" is current |
| 714 | */ |
| 715 | if (killed && !test_thread_flag(TIF_MEMDIE)) |
| 716 | schedule_timeout_uninterruptible(1); |
| 717 | } |
| 718 | |
| 719 | /* |
| 720 | * The pagefault handler calls here because it is out of memory, so kill a |
| 721 | * memory-hogging task. If a populated zone has ZONE_OOM_LOCKED set, a parallel |
| 722 | * oom killing is already in progress so do nothing. If a task is found with |
| 723 | * TIF_MEMDIE set, it has been killed so do nothing and allow it to exit. |
| 724 | */ |
| 725 | void pagefault_out_of_memory(void) |
| 726 | { |
| 727 | if (try_set_system_oom()) { |
| 728 | out_of_memory(NULL, 0, 0, NULL); |
| 729 | clear_system_oom(); |
| 730 | } |
| 731 | if (!test_thread_flag(TIF_MEMDIE)) |
| 732 | schedule_timeout_uninterruptible(1); |
| 733 | } |