oom_kill.c revision 81236810226f71bd9ff77321c8e8276dae7efc61
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#include <linux/security.h> 30 31int sysctl_panic_on_oom; 32int sysctl_oom_kill_allocating_task; 33int sysctl_oom_dump_tasks; 34static DEFINE_SPINLOCK(zone_scan_lock); 35/* #define DEBUG */ 36 37/** 38 * badness - calculate a numeric value for how bad this task has been 39 * @p: task struct of which task we should calculate 40 * @uptime: current uptime in seconds 41 * 42 * The formula used is relatively simple and documented inline in the 43 * function. The main rationale is that we want to select a good task 44 * to kill when we run out of memory. 45 * 46 * Good in this context means that: 47 * 1) we lose the minimum amount of work done 48 * 2) we recover a large amount of memory 49 * 3) we don't kill anything innocent of eating tons of memory 50 * 4) we want to kill the minimum amount of processes (one) 51 * 5) we try to kill the process the user expects us to kill, this 52 * algorithm has been meticulously tuned to meet the principle 53 * of least surprise ... (be careful when you change it) 54 */ 55 56unsigned long badness(struct task_struct *p, unsigned long uptime) 57{ 58 unsigned long points, cpu_time, run_time; 59 struct mm_struct *mm; 60 struct task_struct *child; 61 int oom_adj; 62 63 task_lock(p); 64 mm = p->mm; 65 if (!mm) { 66 task_unlock(p); 67 return 0; 68 } 69 oom_adj = mm->oom_adj; 70 if (oom_adj == OOM_DISABLE) { 71 task_unlock(p); 72 return 0; 73 } 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 if (cpu_time) 120 points /= int_sqrt(cpu_time); 121 if (run_time) 122 points /= int_sqrt(int_sqrt(run_time)); 123 124 /* 125 * Niced processes are most likely less important, so double 126 * their badness points. 127 */ 128 if (task_nice(p) > 0) 129 points *= 2; 130 131 /* 132 * Superuser processes are usually more important, so we make it 133 * less likely that we kill those. 134 */ 135 if (has_capability_noaudit(p, CAP_SYS_ADMIN) || 136 has_capability_noaudit(p, CAP_SYS_RESOURCE)) 137 points /= 4; 138 139 /* 140 * We don't want to kill a process with direct hardware access. 141 * Not only could that mess up the hardware, but usually users 142 * tend to only have this flag set on applications they think 143 * of as important. 144 */ 145 if (has_capability_noaudit(p, CAP_SYS_RAWIO)) 146 points /= 4; 147 148 /* 149 * If p's nodes don't overlap ours, it may still help to kill p 150 * because p may have allocated or otherwise mapped memory on 151 * this node before. However it will be less likely. 152 */ 153 if (!cpuset_mems_allowed_intersects(current, p)) 154 points /= 8; 155 156 /* 157 * Adjust the score by oom_adj. 158 */ 159 if (oom_adj) { 160 if (oom_adj > 0) { 161 if (!points) 162 points = 1; 163 points <<= oom_adj; 164 } else 165 points >>= -(oom_adj); 166 } 167 168#ifdef DEBUG 169 printk(KERN_DEBUG "OOMkill: task %d (%s) got %lu points\n", 170 p->pid, p->comm, points); 171#endif 172 return points; 173} 174 175/* 176 * Determine the type of allocation constraint. 177 */ 178static inline enum oom_constraint constrained_alloc(struct zonelist *zonelist, 179 gfp_t gfp_mask) 180{ 181#ifdef CONFIG_NUMA 182 struct zone *zone; 183 struct zoneref *z; 184 enum zone_type high_zoneidx = gfp_zone(gfp_mask); 185 nodemask_t nodes = node_states[N_HIGH_MEMORY]; 186 187 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) 188 if (cpuset_zone_allowed_softwall(zone, gfp_mask)) 189 node_clear(zone_to_nid(zone), nodes); 190 else 191 return CONSTRAINT_CPUSET; 192 193 if (!nodes_empty(nodes)) 194 return CONSTRAINT_MEMORY_POLICY; 195#endif 196 197 return CONSTRAINT_NONE; 198} 199 200/* 201 * Simple selection loop. We chose the process with the highest 202 * number of 'points'. We expect the caller will lock the tasklist. 203 * 204 * (not docbooked, we don't want this one cluttering up the manual) 205 */ 206static struct task_struct *select_bad_process(unsigned long *ppoints, 207 struct mem_cgroup *mem) 208{ 209 struct task_struct *g, *p; 210 struct task_struct *chosen = NULL; 211 struct timespec uptime; 212 *ppoints = 0; 213 214 do_posix_clock_monotonic_gettime(&uptime); 215 do_each_thread(g, p) { 216 unsigned long points; 217 218 /* 219 * skip kernel threads and tasks which have already released 220 * their mm. 221 */ 222 if (!p->mm) 223 continue; 224 /* skip the init task */ 225 if (is_global_init(p)) 226 continue; 227 if (mem && !task_in_mem_cgroup(p, mem)) 228 continue; 229 230 /* 231 * This task already has access to memory reserves and is 232 * being killed. Don't allow any other task access to the 233 * memory reserve. 234 * 235 * Note: this may have a chance of deadlock if it gets 236 * blocked waiting for another task which itself is waiting 237 * for memory. Is there a better alternative? 238 */ 239 if (test_tsk_thread_flag(p, TIF_MEMDIE)) 240 return ERR_PTR(-1UL); 241 242 /* 243 * This is in the process of releasing memory so wait for it 244 * to finish before killing some other task by mistake. 245 * 246 * However, if p is the current task, we allow the 'kill' to 247 * go ahead if it is exiting: this will simply set TIF_MEMDIE, 248 * which will allow it to gain access to memory reserves in 249 * the process of exiting and releasing its resources. 250 * Otherwise we could get an easy OOM deadlock. 251 */ 252 if (p->flags & PF_EXITING) { 253 if (p != current) 254 return ERR_PTR(-1UL); 255 256 chosen = p; 257 *ppoints = ULONG_MAX; 258 } 259 260 points = badness(p, uptime.tv_sec); 261 if (points > *ppoints) { 262 chosen = p; 263 *ppoints = points; 264 } 265 } while_each_thread(g, p); 266 267 return chosen; 268} 269 270/** 271 * dump_tasks - dump current memory state of all system tasks 272 * @mem: target memory controller 273 * 274 * Dumps the current memory state of all system tasks, excluding kernel threads. 275 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj 276 * score, and name. 277 * 278 * If the actual is non-NULL, only tasks that are a member of the mem_cgroup are 279 * shown. 280 * 281 * Call with tasklist_lock read-locked. 282 */ 283static void dump_tasks(const struct mem_cgroup *mem) 284{ 285 struct task_struct *g, *p; 286 287 printk(KERN_INFO "[ pid ] uid tgid total_vm rss cpu oom_adj " 288 "name\n"); 289 do_each_thread(g, p) { 290 struct mm_struct *mm; 291 292 if (mem && !task_in_mem_cgroup(p, mem)) 293 continue; 294 if (!thread_group_leader(p)) 295 continue; 296 297 task_lock(p); 298 mm = p->mm; 299 if (!mm) { 300 /* 301 * total_vm and rss sizes do not exist for tasks with no 302 * mm so there's no need to report them; they can't be 303 * oom killed anyway. 304 */ 305 task_unlock(p); 306 continue; 307 } 308 printk(KERN_INFO "[%5d] %5d %5d %8lu %8lu %3d %3d %s\n", 309 p->pid, __task_cred(p)->uid, p->tgid, mm->total_vm, 310 get_mm_rss(mm), (int)task_cpu(p), mm->oom_adj, p->comm); 311 task_unlock(p); 312 } while_each_thread(g, p); 313} 314 315/* 316 * Send SIGKILL to the selected process irrespective of CAP_SYS_RAW_IO 317 * flag though it's unlikely that we select a process with CAP_SYS_RAW_IO 318 * set. 319 */ 320static void __oom_kill_task(struct task_struct *p, int verbose) 321{ 322 if (is_global_init(p)) { 323 WARN_ON(1); 324 printk(KERN_WARNING "tried to kill init!\n"); 325 return; 326 } 327 328 if (!p->mm) 329 return; 330 331 if (verbose) 332 printk(KERN_ERR "Killed process %d (%s)\n", 333 task_pid_nr(p), p->comm); 334 335 /* 336 * We give our sacrificial lamb high priority and access to 337 * all the memory it needs. That way it should be able to 338 * exit() and clear out its resources quickly... 339 */ 340 p->rt.time_slice = HZ; 341 set_tsk_thread_flag(p, TIF_MEMDIE); 342 343 force_sig(SIGKILL, p); 344} 345 346static int oom_kill_task(struct task_struct *p) 347{ 348 struct mm_struct *mm; 349 struct task_struct *g, *q; 350 351 task_lock(p); 352 mm = p->mm; 353 if (!mm || mm->oom_adj == OOM_DISABLE) { 354 task_unlock(p); 355 return 1; 356 } 357 task_unlock(p); 358 __oom_kill_task(p, 1); 359 360 /* 361 * kill all processes that share the ->mm (i.e. all threads), 362 * but are in a different thread group. Don't let them have access 363 * to memory reserves though, otherwise we might deplete all memory. 364 */ 365 do_each_thread(g, q) { 366 if (q->mm == mm && !same_thread_group(q, p)) 367 force_sig(SIGKILL, q); 368 } while_each_thread(g, q); 369 370 return 0; 371} 372 373static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order, 374 unsigned long points, struct mem_cgroup *mem, 375 const char *message) 376{ 377 struct task_struct *c; 378 379 if (printk_ratelimit()) { 380 task_lock(current); 381 printk(KERN_WARNING "%s invoked oom-killer: " 382 "gfp_mask=0x%x, order=%d, oom_adj=%d\n", 383 current->comm, gfp_mask, order, 384 current->mm ? current->mm->oom_adj : OOM_DISABLE); 385 cpuset_print_task_mems_allowed(current); 386 task_unlock(current); 387 dump_stack(); 388 mem_cgroup_print_oom_info(mem, current); 389 show_mem(); 390 if (sysctl_oom_dump_tasks) 391 dump_tasks(mem); 392 } 393 394 /* 395 * If the task is already exiting, don't alarm the sysadmin or kill 396 * its children or threads, just set TIF_MEMDIE so it can die quickly 397 * if its mm is still attached. 398 */ 399 if (p->mm && (p->flags & PF_EXITING)) { 400 __oom_kill_task(p, 0); 401 return 0; 402 } 403 404 printk(KERN_ERR "%s: kill process %d (%s) score %li or a child\n", 405 message, task_pid_nr(p), p->comm, points); 406 407 /* Try to kill a child first */ 408 list_for_each_entry(c, &p->children, sibling) { 409 if (c->mm == p->mm) 410 continue; 411 if (!oom_kill_task(c)) 412 return 0; 413 } 414 return oom_kill_task(p); 415} 416 417#ifdef CONFIG_CGROUP_MEM_RES_CTLR 418void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask) 419{ 420 unsigned long points = 0; 421 struct task_struct *p; 422 423 read_lock(&tasklist_lock); 424retry: 425 p = select_bad_process(&points, mem); 426 if (PTR_ERR(p) == -1UL) 427 goto out; 428 429 if (!p) 430 p = current; 431 432 if (oom_kill_process(p, gfp_mask, 0, points, mem, 433 "Memory cgroup out of memory")) 434 goto retry; 435out: 436 read_unlock(&tasklist_lock); 437} 438#endif 439 440static BLOCKING_NOTIFIER_HEAD(oom_notify_list); 441 442int register_oom_notifier(struct notifier_block *nb) 443{ 444 return blocking_notifier_chain_register(&oom_notify_list, nb); 445} 446EXPORT_SYMBOL_GPL(register_oom_notifier); 447 448int unregister_oom_notifier(struct notifier_block *nb) 449{ 450 return blocking_notifier_chain_unregister(&oom_notify_list, nb); 451} 452EXPORT_SYMBOL_GPL(unregister_oom_notifier); 453 454/* 455 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero 456 * if a parallel OOM killing is already taking place that includes a zone in 457 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1. 458 */ 459int try_set_zone_oom(struct zonelist *zonelist, gfp_t gfp_mask) 460{ 461 struct zoneref *z; 462 struct zone *zone; 463 int ret = 1; 464 465 spin_lock(&zone_scan_lock); 466 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) { 467 if (zone_is_oom_locked(zone)) { 468 ret = 0; 469 goto out; 470 } 471 } 472 473 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) { 474 /* 475 * Lock each zone in the zonelist under zone_scan_lock so a 476 * parallel invocation of try_set_zone_oom() doesn't succeed 477 * when it shouldn't. 478 */ 479 zone_set_flag(zone, ZONE_OOM_LOCKED); 480 } 481 482out: 483 spin_unlock(&zone_scan_lock); 484 return ret; 485} 486 487/* 488 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed 489 * allocation attempts with zonelists containing them may now recall the OOM 490 * killer, if necessary. 491 */ 492void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask) 493{ 494 struct zoneref *z; 495 struct zone *zone; 496 497 spin_lock(&zone_scan_lock); 498 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) { 499 zone_clear_flag(zone, ZONE_OOM_LOCKED); 500 } 501 spin_unlock(&zone_scan_lock); 502} 503 504/* 505 * Must be called with tasklist_lock held for read. 506 */ 507static void __out_of_memory(gfp_t gfp_mask, int order) 508{ 509 struct task_struct *p; 510 unsigned long points; 511 512 if (sysctl_oom_kill_allocating_task) 513 if (!oom_kill_process(current, gfp_mask, order, 0, NULL, 514 "Out of memory (oom_kill_allocating_task)")) 515 return; 516retry: 517 /* 518 * Rambo mode: Shoot down a process and hope it solves whatever 519 * issues we may have. 520 */ 521 p = select_bad_process(&points, NULL); 522 523 if (PTR_ERR(p) == -1UL) 524 return; 525 526 /* Found nothing?!?! Either we hang forever, or we panic. */ 527 if (!p) { 528 read_unlock(&tasklist_lock); 529 panic("Out of memory and no killable processes...\n"); 530 } 531 532 if (oom_kill_process(p, gfp_mask, order, points, NULL, 533 "Out of memory")) 534 goto retry; 535} 536 537/* 538 * pagefault handler calls into here because it is out of memory but 539 * doesn't know exactly how or why. 540 */ 541void pagefault_out_of_memory(void) 542{ 543 unsigned long freed = 0; 544 545 blocking_notifier_call_chain(&oom_notify_list, 0, &freed); 546 if (freed > 0) 547 /* Got some memory back in the last second. */ 548 return; 549 550 /* 551 * If this is from memcg, oom-killer is already invoked. 552 * and not worth to go system-wide-oom. 553 */ 554 if (mem_cgroup_oom_called(current)) 555 goto rest_and_return; 556 557 if (sysctl_panic_on_oom) 558 panic("out of memory from page fault. panic_on_oom is selected.\n"); 559 560 read_lock(&tasklist_lock); 561 __out_of_memory(0, 0); /* unknown gfp_mask and order */ 562 read_unlock(&tasklist_lock); 563 564 /* 565 * Give "p" a good chance of killing itself before we 566 * retry to allocate memory. 567 */ 568rest_and_return: 569 if (!test_thread_flag(TIF_MEMDIE)) 570 schedule_timeout_uninterruptible(1); 571} 572 573/** 574 * out_of_memory - kill the "best" process when we run out of memory 575 * @zonelist: zonelist pointer 576 * @gfp_mask: memory allocation flags 577 * @order: amount of memory being requested as a power of 2 578 * 579 * If we run out of memory, we have the choice between either 580 * killing a random task (bad), letting the system crash (worse) 581 * OR try to be smart about which process to kill. Note that we 582 * don't have to be perfect here, we just have to be good. 583 */ 584void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order) 585{ 586 unsigned long freed = 0; 587 enum oom_constraint constraint; 588 589 blocking_notifier_call_chain(&oom_notify_list, 0, &freed); 590 if (freed > 0) 591 /* Got some memory back in the last second. */ 592 return; 593 594 if (sysctl_panic_on_oom == 2) 595 panic("out of memory. Compulsory panic_on_oom is selected.\n"); 596 597 /* 598 * Check if there were limitations on the allocation (only relevant for 599 * NUMA) that may require different handling. 600 */ 601 constraint = constrained_alloc(zonelist, gfp_mask); 602 read_lock(&tasklist_lock); 603 604 switch (constraint) { 605 case CONSTRAINT_MEMORY_POLICY: 606 oom_kill_process(current, gfp_mask, order, 0, NULL, 607 "No available memory (MPOL_BIND)"); 608 break; 609 610 case CONSTRAINT_NONE: 611 if (sysctl_panic_on_oom) 612 panic("out of memory. panic_on_oom is selected\n"); 613 /* Fall-through */ 614 case CONSTRAINT_CPUSET: 615 __out_of_memory(gfp_mask, order); 616 break; 617 } 618 619 read_unlock(&tasklist_lock); 620 621 /* 622 * Give "p" a good chance of killing itself before we 623 * retry to allocate memory unless "p" is current 624 */ 625 if (!test_thread_flag(TIF_MEMDIE)) 626 schedule_timeout_uninterruptible(1); 627} 628