| 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 | * |
| 8 | * The routines in this file are used to kill a process when |
| 9 | * we're seriously out of memory. This gets called from __alloc_pages() |
| 10 | * in mm/page_alloc.c when we really run out of memory. |
| 11 | * |
| 12 | * Since we won't call these routines often (on a well-configured |
| 13 | * machine) this file will double as a 'coding guide' and a signpost |
| 14 | * for newbie kernel hackers. It features several pointers to major |
| 15 | * kernel subsystems and hints as to where to find out what things do. |
| 16 | */ |
| 17 | |
| 18 | #include <linux/oom.h> |
| 19 | #include <linux/mm.h> |
| 20 | #include <linux/err.h> |
| 21 | #include <linux/sched.h> |
| 22 | #include <linux/swap.h> |
| 23 | #include <linux/timex.h> |
| 24 | #include <linux/jiffies.h> |
| 25 | #include <linux/cpuset.h> |
| 26 | #include <linux/module.h> |
| 27 | #include <linux/notifier.h> |
| 28 | #include <linux/memcontrol.h> |
| 29 | |
| 30 | int sysctl_panic_on_oom; |
| 31 | int sysctl_oom_kill_allocating_task; |
| 32 | static DEFINE_SPINLOCK(zone_scan_mutex); |
| 33 | /* #define DEBUG */ |
| 34 | |
| 35 | /** |
| 36 | * badness - calculate a numeric value for how bad this task has been |
| 37 | * @p: task struct of which task we should calculate |
| 38 | * @uptime: current uptime in seconds |
| 39 | * |
| 40 | * The formula used is relatively simple and documented inline in the |
| 41 | * function. The main rationale is that we want to select a good task |
| 42 | * to kill when we run out of memory. |
| 43 | * |
| 44 | * Good in this context means that: |
| 45 | * 1) we lose the minimum amount of work done |
| 46 | * 2) we recover a large amount of memory |
| 47 | * 3) we don't kill anything innocent of eating tons of memory |
| 48 | * 4) we want to kill the minimum amount of processes (one) |
| 49 | * 5) we try to kill the process the user expects us to kill, this |
| 50 | * algorithm has been meticulously tuned to meet the principle |
| 51 | * of least surprise ... (be careful when you change it) |
| 52 | */ |
| 53 | |
| 54 | unsigned long badness(struct task_struct *p, unsigned long uptime, |
| 55 | struct mem_cgroup *mem) |
| 56 | { |
| 57 | unsigned long points, cpu_time, run_time, s; |
| 58 | struct mm_struct *mm; |
| 59 | struct task_struct *child; |
| 60 | |
| 61 | task_lock(p); |
| 62 | mm = p->mm; |
| 63 | if (!mm) { |
| 64 | task_unlock(p); |
| 65 | return 0; |
| 66 | } |
| 67 | |
| 68 | #ifdef CONFIG_CGROUP_MEM_CONT |
| 69 | if (mem != NULL && mm->mem_cgroup != mem) { |
| 70 | task_unlock(p); |
| 71 | return 0; |
| 72 | } |
| 73 | #endif |
| 74 | |
| 75 | /* |
| 76 | * The memory size of the process is the basis for the badness. |
| 77 | */ |
| 78 | points = mm->total_vm; |
| 79 | |
| 80 | /* |
| 81 | * After this unlock we can no longer dereference local variable `mm' |
| 82 | */ |
| 83 | task_unlock(p); |
| 84 | |
| 85 | /* |
| 86 | * swapoff can easily use up all memory, so kill those first. |
| 87 | */ |
| 88 | if (p->flags & PF_SWAPOFF) |
| 89 | return ULONG_MAX; |
| 90 | |
| 91 | /* |
| 92 | * Processes which fork a lot of child processes are likely |
| 93 | * a good choice. We add half the vmsize of the children if they |
| 94 | * have an own mm. This prevents forking servers to flood the |
| 95 | * machine with an endless amount of children. In case a single |
| 96 | * child is eating the vast majority of memory, adding only half |
| 97 | * to the parents will make the child our kill candidate of choice. |
| 98 | */ |
| 99 | list_for_each_entry(child, &p->children, sibling) { |
| 100 | task_lock(child); |
| 101 | if (child->mm != mm && child->mm) |
| 102 | points += child->mm->total_vm/2 + 1; |
| 103 | task_unlock(child); |
| 104 | } |
| 105 | |
| 106 | /* |
| 107 | * CPU time is in tens of seconds and run time is in thousands |
| 108 | * of seconds. There is no particular reason for this other than |
| 109 | * that it turned out to work very well in practice. |
| 110 | */ |
| 111 | cpu_time = (cputime_to_jiffies(p->utime) + cputime_to_jiffies(p->stime)) |
| 112 | >> (SHIFT_HZ + 3); |
| 113 | |
| 114 | if (uptime >= p->start_time.tv_sec) |
| 115 | run_time = (uptime - p->start_time.tv_sec) >> 10; |
| 116 | else |
| 117 | run_time = 0; |
| 118 | |
| 119 | s = int_sqrt(cpu_time); |
| 120 | if (s) |
| 121 | points /= s; |
| 122 | s = int_sqrt(int_sqrt(run_time)); |
| 123 | if (s) |
| 124 | points /= s; |
| 125 | |
| 126 | /* |
| 127 | * Niced processes are most likely less important, so double |
| 128 | * their badness points. |
| 129 | */ |
| 130 | if (task_nice(p) > 0) |
| 131 | points *= 2; |
| 132 | |
| 133 | /* |
| 134 | * Superuser processes are usually more important, so we make it |
| 135 | * less likely that we kill those. |
| 136 | */ |
| 137 | if (__capable(p, CAP_SYS_ADMIN) || __capable(p, CAP_SYS_RESOURCE)) |
| 138 | points /= 4; |
| 139 | |
| 140 | /* |
| 141 | * We don't want to kill a process with direct hardware access. |
| 142 | * Not only could that mess up the hardware, but usually users |
| 143 | * tend to only have this flag set on applications they think |
| 144 | * of as important. |
| 145 | */ |
| 146 | if (__capable(p, CAP_SYS_RAWIO)) |
| 147 | points /= 4; |
| 148 | |
| 149 | /* |
| 150 | * If p's nodes don't overlap ours, it may still help to kill p |
| 151 | * because p may have allocated or otherwise mapped memory on |
| 152 | * this node before. However it will be less likely. |
| 153 | */ |
| 154 | if (!cpuset_mems_allowed_intersects(current, p)) |
| 155 | points /= 8; |
| 156 | |
| 157 | /* |
| 158 | * Adjust the score by oomkilladj. |
| 159 | */ |
| 160 | if (p->oomkilladj) { |
| 161 | if (p->oomkilladj > 0) { |
| 162 | if (!points) |
| 163 | points = 1; |
| 164 | points <<= p->oomkilladj; |
| 165 | } else |
| 166 | points >>= -(p->oomkilladj); |
| 167 | } |
| 168 | |
| 169 | #ifdef DEBUG |
| 170 | printk(KERN_DEBUG "OOMkill: task %d (%s) got %lu points\n", |
| 171 | p->pid, p->comm, points); |
| 172 | #endif |
| 173 | return points; |
| 174 | } |
| 175 | |
| 176 | /* |
| 177 | * Determine the type of allocation constraint. |
| 178 | */ |
| 179 | static inline enum oom_constraint constrained_alloc(struct zonelist *zonelist, |
| 180 | gfp_t gfp_mask) |
| 181 | { |
| 182 | #ifdef CONFIG_NUMA |
| 183 | struct zone **z; |
| 184 | nodemask_t nodes = node_states[N_HIGH_MEMORY]; |
| 185 | |
| 186 | for (z = zonelist->zones; *z; z++) |
| 187 | if (cpuset_zone_allowed_softwall(*z, gfp_mask)) |
| 188 | node_clear(zone_to_nid(*z), nodes); |
| 189 | else |
| 190 | return CONSTRAINT_CPUSET; |
| 191 | |
| 192 | if (!nodes_empty(nodes)) |
| 193 | return CONSTRAINT_MEMORY_POLICY; |
| 194 | #endif |
| 195 | |
| 196 | return CONSTRAINT_NONE; |
| 197 | } |
| 198 | |
| 199 | /* |
| 200 | * Simple selection loop. We chose the process with the highest |
| 201 | * number of 'points'. We expect the caller will lock the tasklist. |
| 202 | * |
| 203 | * (not docbooked, we don't want this one cluttering up the manual) |
| 204 | */ |
| 205 | static struct task_struct *select_bad_process(unsigned long *ppoints, |
| 206 | struct mem_cgroup *mem) |
| 207 | { |
| 208 | struct task_struct *g, *p; |
| 209 | struct task_struct *chosen = NULL; |
| 210 | struct timespec uptime; |
| 211 | *ppoints = 0; |
| 212 | |
| 213 | do_posix_clock_monotonic_gettime(&uptime); |
| 214 | do_each_thread(g, p) { |
| 215 | unsigned long points; |
| 216 | |
| 217 | /* |
| 218 | * skip kernel threads and tasks which have already released |
| 219 | * their mm. |
| 220 | */ |
| 221 | if (!p->mm) |
| 222 | continue; |
| 223 | /* skip the init task */ |
| 224 | if (is_global_init(p)) |
| 225 | continue; |
| 226 | |
| 227 | /* |
| 228 | * This task already has access to memory reserves and is |
| 229 | * being killed. Don't allow any other task access to the |
| 230 | * memory reserve. |
| 231 | * |
| 232 | * Note: this may have a chance of deadlock if it gets |
| 233 | * blocked waiting for another task which itself is waiting |
| 234 | * for memory. Is there a better alternative? |
| 235 | */ |
| 236 | if (test_tsk_thread_flag(p, TIF_MEMDIE)) |
| 237 | return ERR_PTR(-1UL); |
| 238 | |
| 239 | /* |
| 240 | * This is in the process of releasing memory so wait for it |
| 241 | * to finish before killing some other task by mistake. |
| 242 | * |
| 243 | * However, if p is the current task, we allow the 'kill' to |
| 244 | * go ahead if it is exiting: this will simply set TIF_MEMDIE, |
| 245 | * which will allow it to gain access to memory reserves in |
| 246 | * the process of exiting and releasing its resources. |
| 247 | * Otherwise we could get an easy OOM deadlock. |
| 248 | */ |
| 249 | if (p->flags & PF_EXITING) { |
| 250 | if (p != current) |
| 251 | return ERR_PTR(-1UL); |
| 252 | |
| 253 | chosen = p; |
| 254 | *ppoints = ULONG_MAX; |
| 255 | } |
| 256 | |
| 257 | if (p->oomkilladj == OOM_DISABLE) |
| 258 | continue; |
| 259 | |
| 260 | points = badness(p, uptime.tv_sec, mem); |
| 261 | if (points > *ppoints || !chosen) { |
| 262 | chosen = p; |
| 263 | *ppoints = points; |
| 264 | } |
| 265 | } while_each_thread(g, p); |
| 266 | |
| 267 | return chosen; |
| 268 | } |
| 269 | |
| 270 | /** |
| 271 | * Send SIGKILL to the selected process irrespective of CAP_SYS_RAW_IO |
| 272 | * flag though it's unlikely that we select a process with CAP_SYS_RAW_IO |
| 273 | * set. |
| 274 | */ |
| 275 | static void __oom_kill_task(struct task_struct *p, int verbose) |
| 276 | { |
| 277 | if (is_global_init(p)) { |
| 278 | WARN_ON(1); |
| 279 | printk(KERN_WARNING "tried to kill init!\n"); |
| 280 | return; |
| 281 | } |
| 282 | |
| 283 | if (!p->mm) { |
| 284 | WARN_ON(1); |
| 285 | printk(KERN_WARNING "tried to kill an mm-less task!\n"); |
| 286 | return; |
| 287 | } |
| 288 | |
| 289 | if (verbose) |
| 290 | printk(KERN_ERR "Killed process %d (%s)\n", |
| 291 | task_pid_nr(p), p->comm); |
| 292 | |
| 293 | /* |
| 294 | * We give our sacrificial lamb high priority and access to |
| 295 | * all the memory it needs. That way it should be able to |
| 296 | * exit() and clear out its resources quickly... |
| 297 | */ |
| 298 | p->rt.time_slice = HZ; |
| 299 | set_tsk_thread_flag(p, TIF_MEMDIE); |
| 300 | |
| 301 | force_sig(SIGKILL, p); |
| 302 | } |
| 303 | |
| 304 | static int oom_kill_task(struct task_struct *p) |
| 305 | { |
| 306 | struct mm_struct *mm; |
| 307 | struct task_struct *g, *q; |
| 308 | |
| 309 | mm = p->mm; |
| 310 | |
| 311 | /* WARNING: mm may not be dereferenced since we did not obtain its |
| 312 | * value from get_task_mm(p). This is OK since all we need to do is |
| 313 | * compare mm to q->mm below. |
| 314 | * |
| 315 | * Furthermore, even if mm contains a non-NULL value, p->mm may |
| 316 | * change to NULL at any time since we do not hold task_lock(p). |
| 317 | * However, this is of no concern to us. |
| 318 | */ |
| 319 | |
| 320 | if (mm == NULL) |
| 321 | return 1; |
| 322 | |
| 323 | /* |
| 324 | * Don't kill the process if any threads are set to OOM_DISABLE |
| 325 | */ |
| 326 | do_each_thread(g, q) { |
| 327 | if (q->mm == mm && q->oomkilladj == OOM_DISABLE) |
| 328 | return 1; |
| 329 | } while_each_thread(g, q); |
| 330 | |
| 331 | __oom_kill_task(p, 1); |
| 332 | |
| 333 | /* |
| 334 | * kill all processes that share the ->mm (i.e. all threads), |
| 335 | * but are in a different thread group. Don't let them have access |
| 336 | * to memory reserves though, otherwise we might deplete all memory. |
| 337 | */ |
| 338 | do_each_thread(g, q) { |
| 339 | if (q->mm == mm && !same_thread_group(q, p)) |
| 340 | force_sig(SIGKILL, q); |
| 341 | } while_each_thread(g, q); |
| 342 | |
| 343 | return 0; |
| 344 | } |
| 345 | |
| 346 | static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order, |
| 347 | unsigned long points, const char *message) |
| 348 | { |
| 349 | struct task_struct *c; |
| 350 | |
| 351 | if (printk_ratelimit()) { |
| 352 | printk(KERN_WARNING "%s invoked oom-killer: " |
| 353 | "gfp_mask=0x%x, order=%d, oomkilladj=%d\n", |
| 354 | current->comm, gfp_mask, order, current->oomkilladj); |
| 355 | dump_stack(); |
| 356 | show_mem(); |
| 357 | } |
| 358 | |
| 359 | /* |
| 360 | * If the task is already exiting, don't alarm the sysadmin or kill |
| 361 | * its children or threads, just set TIF_MEMDIE so it can die quickly |
| 362 | */ |
| 363 | if (p->flags & PF_EXITING) { |
| 364 | __oom_kill_task(p, 0); |
| 365 | return 0; |
| 366 | } |
| 367 | |
| 368 | printk(KERN_ERR "%s: kill process %d (%s) score %li or a child\n", |
| 369 | message, task_pid_nr(p), p->comm, points); |
| 370 | |
| 371 | /* Try to kill a child first */ |
| 372 | list_for_each_entry(c, &p->children, sibling) { |
| 373 | if (c->mm == p->mm) |
| 374 | continue; |
| 375 | if (!oom_kill_task(c)) |
| 376 | return 0; |
| 377 | } |
| 378 | return oom_kill_task(p); |
| 379 | } |
| 380 | |
| 381 | #ifdef CONFIG_CGROUP_MEM_CONT |
| 382 | void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask) |
| 383 | { |
| 384 | unsigned long points = 0; |
| 385 | struct task_struct *p; |
| 386 | |
| 387 | cgroup_lock(); |
| 388 | rcu_read_lock(); |
| 389 | retry: |
| 390 | p = select_bad_process(&points, mem); |
| 391 | if (PTR_ERR(p) == -1UL) |
| 392 | goto out; |
| 393 | |
| 394 | if (!p) |
| 395 | p = current; |
| 396 | |
| 397 | if (oom_kill_process(p, gfp_mask, 0, points, |
| 398 | "Memory cgroup out of memory")) |
| 399 | goto retry; |
| 400 | out: |
| 401 | rcu_read_unlock(); |
| 402 | cgroup_unlock(); |
| 403 | } |
| 404 | #endif |
| 405 | |
| 406 | static BLOCKING_NOTIFIER_HEAD(oom_notify_list); |
| 407 | |
| 408 | int register_oom_notifier(struct notifier_block *nb) |
| 409 | { |
| 410 | return blocking_notifier_chain_register(&oom_notify_list, nb); |
| 411 | } |
| 412 | EXPORT_SYMBOL_GPL(register_oom_notifier); |
| 413 | |
| 414 | int unregister_oom_notifier(struct notifier_block *nb) |
| 415 | { |
| 416 | return blocking_notifier_chain_unregister(&oom_notify_list, nb); |
| 417 | } |
| 418 | EXPORT_SYMBOL_GPL(unregister_oom_notifier); |
| 419 | |
| 420 | /* |
| 421 | * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero |
| 422 | * if a parallel OOM killing is already taking place that includes a zone in |
| 423 | * the zonelist. Otherwise, locks all zones in the zonelist and returns 1. |
| 424 | */ |
| 425 | int try_set_zone_oom(struct zonelist *zonelist) |
| 426 | { |
| 427 | struct zone **z; |
| 428 | int ret = 1; |
| 429 | |
| 430 | z = zonelist->zones; |
| 431 | |
| 432 | spin_lock(&zone_scan_mutex); |
| 433 | do { |
| 434 | if (zone_is_oom_locked(*z)) { |
| 435 | ret = 0; |
| 436 | goto out; |
| 437 | } |
| 438 | } while (*(++z) != NULL); |
| 439 | |
| 440 | /* |
| 441 | * Lock each zone in the zonelist under zone_scan_mutex so a parallel |
| 442 | * invocation of try_set_zone_oom() doesn't succeed when it shouldn't. |
| 443 | */ |
| 444 | z = zonelist->zones; |
| 445 | do { |
| 446 | zone_set_flag(*z, ZONE_OOM_LOCKED); |
| 447 | } while (*(++z) != NULL); |
| 448 | out: |
| 449 | spin_unlock(&zone_scan_mutex); |
| 450 | return ret; |
| 451 | } |
| 452 | |
| 453 | /* |
| 454 | * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed |
| 455 | * allocation attempts with zonelists containing them may now recall the OOM |
| 456 | * killer, if necessary. |
| 457 | */ |
| 458 | void clear_zonelist_oom(struct zonelist *zonelist) |
| 459 | { |
| 460 | struct zone **z; |
| 461 | |
| 462 | z = zonelist->zones; |
| 463 | |
| 464 | spin_lock(&zone_scan_mutex); |
| 465 | do { |
| 466 | zone_clear_flag(*z, ZONE_OOM_LOCKED); |
| 467 | } while (*(++z) != NULL); |
| 468 | spin_unlock(&zone_scan_mutex); |
| 469 | } |
| 470 | |
| 471 | /** |
| 472 | * out_of_memory - kill the "best" process when we run out of memory |
| 473 | * |
| 474 | * If we run out of memory, we have the choice between either |
| 475 | * killing a random task (bad), letting the system crash (worse) |
| 476 | * OR try to be smart about which process to kill. Note that we |
| 477 | * don't have to be perfect here, we just have to be good. |
| 478 | */ |
| 479 | void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order) |
| 480 | { |
| 481 | struct task_struct *p; |
| 482 | unsigned long points = 0; |
| 483 | unsigned long freed = 0; |
| 484 | enum oom_constraint constraint; |
| 485 | |
| 486 | blocking_notifier_call_chain(&oom_notify_list, 0, &freed); |
| 487 | if (freed > 0) |
| 488 | /* Got some memory back in the last second. */ |
| 489 | return; |
| 490 | |
| 491 | if (sysctl_panic_on_oom == 2) |
| 492 | panic("out of memory. Compulsory panic_on_oom is selected.\n"); |
| 493 | |
| 494 | /* |
| 495 | * Check if there were limitations on the allocation (only relevant for |
| 496 | * NUMA) that may require different handling. |
| 497 | */ |
| 498 | constraint = constrained_alloc(zonelist, gfp_mask); |
| 499 | read_lock(&tasklist_lock); |
| 500 | |
| 501 | switch (constraint) { |
| 502 | case CONSTRAINT_MEMORY_POLICY: |
| 503 | oom_kill_process(current, gfp_mask, order, points, |
| 504 | "No available memory (MPOL_BIND)"); |
| 505 | break; |
| 506 | |
| 507 | case CONSTRAINT_NONE: |
| 508 | if (sysctl_panic_on_oom) |
| 509 | panic("out of memory. panic_on_oom is selected\n"); |
| 510 | /* Fall-through */ |
| 511 | case CONSTRAINT_CPUSET: |
| 512 | if (sysctl_oom_kill_allocating_task) { |
| 513 | oom_kill_process(current, gfp_mask, order, points, |
| 514 | "Out of memory (oom_kill_allocating_task)"); |
| 515 | break; |
| 516 | } |
| 517 | retry: |
| 518 | /* |
| 519 | * Rambo mode: Shoot down a process and hope it solves whatever |
| 520 | * issues we may have. |
| 521 | */ |
| 522 | p = select_bad_process(&points, NULL); |
| 523 | |
| 524 | if (PTR_ERR(p) == -1UL) |
| 525 | goto out; |
| 526 | |
| 527 | /* Found nothing?!?! Either we hang forever, or we panic. */ |
| 528 | if (!p) { |
| 529 | read_unlock(&tasklist_lock); |
| 530 | panic("Out of memory and no killable processes...\n"); |
| 531 | } |
| 532 | |
| 533 | if (oom_kill_process(p, gfp_mask, order, points, |
| 534 | "Out of memory")) |
| 535 | goto retry; |
| 536 | |
| 537 | break; |
| 538 | } |
| 539 | |
| 540 | out: |
| 541 | read_unlock(&tasklist_lock); |
| 542 | |
| 543 | /* |
| 544 | * Give "p" a good chance of killing itself before we |
| 545 | * retry to allocate memory unless "p" is current |
| 546 | */ |
| 547 | if (!test_thread_flag(TIF_MEMDIE)) |
| 548 | schedule_timeout_uninterruptible(1); |
| 549 | } |