mm/oom_kill.c

   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
  31 int sysctl_panic_on_oom;
  32 int sysctl_oom_kill_allocating_task;
  33 int sysctl_oom_dump_tasks;
  34 static 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
  56 unsigned 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 = p->signal->oom_adj;
  62
  63         if (oom_adj == OOM_DISABLE)
  64                 return 0;
  65
  66         task_lock(p);
  67         mm = p->mm;
  68         if (!mm) {
  69                 task_unlock(p);
  70                 return 0;
  71         }
  72
  73         /*
  74          * The memory size of the process is the basis for the badness.
  75          */
  76         points = mm->total_vm;
  77
  78         /*
  79          * After this unlock we can no longer dereference local variable `mm'
  80          */
  81         task_unlock(p);
  82
  83         /*
  84          * swapoff can easily use up all memory, so kill those first.
  85          */
  86         if (p->flags & PF_OOM_ORIGIN)
  87                 return ULONG_MAX;
  88
  89         /*
  90          * Processes which fork a lot of child processes are likely
  91          * a good choice. We add half the vmsize of the children if they
  92          * have an own mm. This prevents forking servers to flood the
  93          * machine with an endless amount of children. In case a single
  94          * child is eating the vast majority of memory, adding only half
  95          * to the parents will make the child our kill candidate of choice.
  96          */
  97         list_for_each_entry(child, &p->children, sibling) {
  98                 task_lock(child);
  99                 if (child->mm != mm && child->mm)
 100                         points += child->mm->total_vm/2 + 1;
 101                 task_unlock(child);
 102         }
 103
 104         /*
 105          * CPU time is in tens of seconds and run time is in thousands
 106          * of seconds. There is no particular reason for this other than
 107          * that it turned out to work very well in practice.
 108          */
 109         cpu_time = (cputime_to_jiffies(p->utime) + cputime_to_jiffies(p->stime))
 110                 >> (SHIFT_HZ + 3);
 111
 112         if (uptime >= p->start_time.tv_sec)
 113                 run_time = (uptime - p->start_time.tv_sec) >> 10;
 114         else
 115                 run_time = 0;
 116
 117         if (cpu_time)
 118                 points /= int_sqrt(cpu_time);
 119         if (run_time)
 120                 points /= int_sqrt(int_sqrt(run_time));
 121
 122         /*
 123          * Niced processes are most likely less important, so double
 124          * their badness points.
 125          */
 126         if (task_nice(p) > 0)
 127                 points *= 2;
 128
 129         /*
 130          * Superuser processes are usually more important, so we make it
 131          * less likely that we kill those.
 132          */
 133         if (has_capability_noaudit(p, CAP_SYS_ADMIN) ||
 134             has_capability_noaudit(p, CAP_SYS_RESOURCE))
 135                 points /= 4;
 136
 137         /*
 138          * We don't want to kill a process with direct hardware access.
 139          * Not only could that mess up the hardware, but usually users
 140          * tend to only have this flag set on applications they think
 141          * of as important.
 142          */
 143         if (has_capability_noaudit(p, CAP_SYS_RAWIO))
 144                 points /= 4;
 145
 146         /*
 147          * If p's nodes don't overlap ours, it may still help to kill p
 148          * because p may have allocated or otherwise mapped memory on
 149          * this node before. However it will be less likely.
 150          */
 151         if (!cpuset_mems_allowed_intersects(current, p))
 152                 points /= 8;
 153
 154         /*
 155          * Adjust the score by oom_adj.
 156          */
 157         if (oom_adj) {
 158                 if (oom_adj > 0) {
 159                         if (!points)
 160                                 points = 1;
 161                         points <<= oom_adj;
 162                 } else
 163                         points >>= -(oom_adj);
 164         }
 165
 166 #ifdef DEBUG
 167         printk(KERN_DEBUG "OOMkill: task %d (%s) got %lu points\n",
 168         p->pid, p->comm, points);
 169 #endif
 170         return points;
 171 }
 172
 173 /*
 174  * Determine the type of allocation constraint.
 175  */
 176 static inline enum oom_constraint constrained_alloc(struct zonelist *zonelist,
 177                                                     gfp_t gfp_mask)
 178 {
 179 #ifdef CONFIG_NUMA
 180         struct zone *zone;
 181         struct zoneref *z;
 182         enum zone_type high_zoneidx = gfp_zone(gfp_mask);
 183         nodemask_t nodes = node_states[N_HIGH_MEMORY];
 184
 185         for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
 186                 if (cpuset_zone_allowed_softwall(zone, gfp_mask))
 187                         node_clear(zone_to_nid(zone), nodes);
 188                 else
 189                         return CONSTRAINT_CPUSET;
 190
 191         if (!nodes_empty(nodes))
 192                 return CONSTRAINT_MEMORY_POLICY;
 193 #endif
 194
 195         return CONSTRAINT_NONE;
 196 }
 197
 198 /*
 199  * Simple selection loop. We chose the process with the highest
 200  * number of 'points'. We expect the caller will lock the tasklist.
 201  *
 202  * (not docbooked, we don't want this one cluttering up the manual)
 203  */
 204 static struct task_struct *select_bad_process(unsigned long *ppoints,
 205                                                 struct mem_cgroup *mem)
 206 {
 207         struct task_struct *g, *p;
 208         struct task_struct *chosen = NULL;
 209         struct timespec uptime;
 210         *ppoints = 0;
 211
 212         do_posix_clock_monotonic_gettime(&uptime);
 213         do_each_thread(g, p) {
 214                 unsigned long points;
 215
 216                 /*
 217                  * skip kernel threads and tasks which have already released
 218                  * their mm.
 219                  */
 220                 if (!p->mm)
 221                         continue;
 222                 /* skip the init task */
 223                 if (is_global_init(p))
 224                         continue;
 225                 if (mem && !task_in_mem_cgroup(p, mem))
 226                         continue;
 227
 228                 /*
 229                  * This task already has access to memory reserves and is
 230                  * being killed. Don't allow any other task access to the
 231                  * memory reserve.
 232                  *
 233                  * Note: this may have a chance of deadlock if it gets
 234                  * blocked waiting for another task which itself is waiting
 235                  * for memory. Is there a better alternative?
 236                  */
 237                 if (test_tsk_thread_flag(p, TIF_MEMDIE))
 238                         return ERR_PTR(-1UL);
 239
 240                 /*
 241                  * This is in the process of releasing memory so wait for it
 242                  * to finish before killing some other task by mistake.
 243                  *
 244                  * However, if p is the current task, we allow the 'kill' to
 245                  * go ahead if it is exiting: this will simply set TIF_MEMDIE,
 246                  * which will allow it to gain access to memory reserves in
 247                  * the process of exiting and releasing its resources.
 248                  * Otherwise we could get an easy OOM deadlock.
 249                  */
 250                 if (p->flags & PF_EXITING) {
 251                         if (p != current)
 252                                 return ERR_PTR(-1UL);
 253
 254                         chosen = p;
 255                         *ppoints = ULONG_MAX;
 256                 }
 257
 258                 if (p->signal->oom_adj == OOM_DISABLE)
 259                         continue;
 260
 261                 points = badness(p, uptime.tv_sec);
 262                 if (points > *ppoints || !chosen) {
 263                         chosen = p;
 264                         *ppoints = points;
 265                 }
 266         } while_each_thread(g, p);
 267
 268         return chosen;
 269 }
 270
 271 /**
 272  * dump_tasks - dump current memory state of all system tasks
 273  * @mem: target memory controller
 274  *
 275  * Dumps the current memory state of all system tasks, excluding kernel threads.
 276  * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
 277  * score, and name.
 278  *
 279  * If the actual is non-NULL, only tasks that are a member of the mem_cgroup are
 280  * shown.
 281  *
 282  * Call with tasklist_lock read-locked.
 283  */
 284 static void dump_tasks(const struct mem_cgroup *mem)
 285 {
 286         struct task_struct *g, *p;
 287
 288         printk(KERN_INFO "[ pid ]   uid  tgid total_vm      rss cpu oom_adj "
 289                "name\n");
 290         do_each_thread(g, p) {
 291                 struct mm_struct *mm;
 292
 293                 if (mem && !task_in_mem_cgroup(p, mem))
 294                         continue;
 295                 if (!thread_group_leader(p))
 296                         continue;
 297
 298                 task_lock(p);
 299                 mm = p->mm;
 300                 if (!mm) {
 301                         /*
 302                          * total_vm and rss sizes do not exist for tasks with no
 303                          * mm so there's no need to report them; they can't be
 304                          * oom killed anyway.
 305                          */
 306                         task_unlock(p);
 307                         continue;
 308                 }
 309                 printk(KERN_INFO "[%5d] %5d %5d %8lu %8lu %3d     %3d %s\n",
 310                        p->pid, __task_cred(p)->uid, p->tgid, mm->total_vm,
 311                        get_mm_rss(mm), (int)task_cpu(p), p->signal->oom_adj,
 312                        p->comm);
 313                 task_unlock(p);
 314         } while_each_thread(g, p);
 315 }
 316
 317 /*
 318  * Send SIGKILL to the selected  process irrespective of  CAP_SYS_RAW_IO
 319  * flag though it's unlikely that  we select a process with CAP_SYS_RAW_IO
 320  * set.
 321  */
 322 static void __oom_kill_task(struct task_struct *p, int verbose)
 323 {
 324         if (is_global_init(p)) {
 325                 WARN_ON(1);
 326                 printk(KERN_WARNING "tried to kill init!\n");
 327                 return;
 328         }
 329
 330         if (!p->mm) {
 331                 WARN_ON(1);
 332                 printk(KERN_WARNING "tried to kill an mm-less task!\n");
 333                 return;
 334         }
 335
 336         if (verbose)
 337                 printk(KERN_ERR "Killed process %d (%s)\n",
 338                                 task_pid_nr(p), p->comm);
 339
 340         /*
 341          * We give our sacrificial lamb high priority and access to
 342          * all the memory it needs. That way it should be able to
 343          * exit() and clear out its resources quickly...
 344          */
 345         p->rt.time_slice = HZ;
 346         set_tsk_thread_flag(p, TIF_MEMDIE);
 347
 348         force_sig(SIGKILL, p);
 349 }
 350
 351 static int oom_kill_task(struct task_struct *p)
 352 {
 353         struct mm_struct *mm;
 354         struct task_struct *g, *q;
 355
 356         mm = p->mm;
 357
 358         /* WARNING: mm may not be dereferenced since we did not obtain its
 359          * value from get_task_mm(p).  This is OK since all we need to do is
 360          * compare mm to q->mm below.
 361          *
 362          * Furthermore, even if mm contains a non-NULL value, p->mm may
 363          * change to NULL at any time since we do not hold task_lock(p).
 364          * However, this is of no concern to us.
 365          */
 366         if (!mm || p->signal->oom_adj == OOM_DISABLE)
 367                 return 1;
 368
 369         __oom_kill_task(p, 1);
 370
 371         /*
 372          * kill all processes that share the ->mm (i.e. all threads),
 373          * but are in a different thread group. Don't let them have access
 374          * to memory reserves though, otherwise we might deplete all memory.
 375          */
 376         do_each_thread(g, q) {
 377                 if (q->mm == mm && !same_thread_group(q, p))
 378                         force_sig(SIGKILL, q);
 379         } while_each_thread(g, q);
 380
 381         return 0;
 382 }
 383
 384 static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
 385                             unsigned long points, struct mem_cgroup *mem,
 386                             const char *message)
 387 {
 388         struct task_struct *c;
 389
 390         if (printk_ratelimit()) {
 391                 printk(KERN_WARNING "%s invoked oom-killer: "
 392                         "gfp_mask=0x%x, order=%d, oom_adj=%d\n",
 393                         current->comm, gfp_mask, order,
 394                         current->signal->oom_adj);
 395                 task_lock(current);
 396                 cpuset_print_task_mems_allowed(current);
 397                 task_unlock(current);
 398                 dump_stack();
 399                 mem_cgroup_print_oom_info(mem, current);
 400                 show_mem();
 401                 if (sysctl_oom_dump_tasks)
 402                         dump_tasks(mem);
 403         }
 404
 405         /*
 406          * If the task is already exiting, don't alarm the sysadmin or kill
 407          * its children or threads, just set TIF_MEMDIE so it can die quickly
 408          */
 409         if (p->flags & PF_EXITING) {
 410                 __oom_kill_task(p, 0);
 411                 return 0;
 412         }
 413
 414         printk(KERN_ERR "%s: kill process %d (%s) score %li or a child\n",
 415                                         message, task_pid_nr(p), p->comm, points);
 416
 417         /* Try to kill a child first */
 418         list_for_each_entry(c, &p->children, sibling) {
 419                 if (c->mm == p->mm)
 420                         continue;
 421                 if (!oom_kill_task(c))
 422                         return 0;
 423         }
 424         return oom_kill_task(p);
 425 }
 426
 427 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
 428 void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
 429 {
 430         unsigned long points = 0;
 431         struct task_struct *p;
 432
 433         read_lock(&tasklist_lock);
 434 retry:
 435         p = select_bad_process(&points, mem);
 436         if (PTR_ERR(p) == -1UL)
 437                 goto out;
 438
 439         if (!p)
 440                 p = current;
 441
 442         if (oom_kill_process(p, gfp_mask, 0, points, mem,
 443                                 "Memory cgroup out of memory"))
 444                 goto retry;
 445 out:
 446         read_unlock(&tasklist_lock);
 447 }
 448 #endif
 449
 450 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
 451
 452 int register_oom_notifier(struct notifier_block *nb)
 453 {
 454         return blocking_notifier_chain_register(&oom_notify_list, nb);
 455 }
 456 EXPORT_SYMBOL_GPL(register_oom_notifier);
 457
 458 int unregister_oom_notifier(struct notifier_block *nb)
 459 {
 460         return blocking_notifier_chain_unregister(&oom_notify_list, nb);
 461 }
 462 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
 463
 464 /*
 465  * Try to acquire the OOM killer lock for the zones in zonelist.  Returns zero
 466  * if a parallel OOM killing is already taking place that includes a zone in
 467  * the zonelist.  Otherwise, locks all zones in the zonelist and returns 1.
 468  */
 469 int try_set_zone_oom(struct zonelist *zonelist, gfp_t gfp_mask)
 470 {
 471         struct zoneref *z;
 472         struct zone *zone;
 473         int ret = 1;
 474
 475         spin_lock(&zone_scan_lock);
 476         for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
 477                 if (zone_is_oom_locked(zone)) {
 478                         ret = 0;
 479                         goto out;
 480                 }
 481         }
 482
 483         for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
 484                 /*
 485                  * Lock each zone in the zonelist under zone_scan_lock so a
 486                  * parallel invocation of try_set_zone_oom() doesn't succeed
 487                  * when it shouldn't.
 488                  */
 489                 zone_set_flag(zone, ZONE_OOM_LOCKED);
 490         }
 491
 492 out:
 493         spin_unlock(&zone_scan_lock);
 494         return ret;
 495 }
 496
 497 /*
 498  * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
 499  * allocation attempts with zonelists containing them may now recall the OOM
 500  * killer, if necessary.
 501  */
 502 void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
 503 {
 504         struct zoneref *z;
 505         struct zone *zone;
 506
 507         spin_lock(&zone_scan_lock);
 508         for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
 509                 zone_clear_flag(zone, ZONE_OOM_LOCKED);
 510         }
 511         spin_unlock(&zone_scan_lock);
 512 }
 513
 514 /*
 515  * Must be called with tasklist_lock held for read.
 516  */
 517 static void __out_of_memory(gfp_t gfp_mask, int order)
 518 {
 519         struct task_struct *p;
 520         unsigned long points;
 521
 522         if (sysctl_oom_kill_allocating_task)
 523                 if (!oom_kill_process(current, gfp_mask, order, 0, NULL,
 524                                 "Out of memory (oom_kill_allocating_task)"))
 525                         return;
 526 retry:
 527         /*
 528          * Rambo mode: Shoot down a process and hope it solves whatever
 529          * issues we may have.
 530          */
 531         p = select_bad_process(&points, NULL);
 532
 533         if (PTR_ERR(p) == -1UL)
 534                 return;
 535
 536         /* Found nothing?!?! Either we hang forever, or we panic. */
 537         if (!p) {
 538                 read_unlock(&tasklist_lock);
 539                 panic("Out of memory and no killable processes...\n");
 540         }
 541
 542         if (oom_kill_process(p, gfp_mask, order, points, NULL,
 543                              "Out of memory"))
 544                 goto retry;
 545 }
 546
 547 /*
 548  * pagefault handler calls into here because it is out of memory but
 549  * doesn't know exactly how or why.
 550  */
 551 void pagefault_out_of_memory(void)
 552 {
 553         unsigned long freed = 0;
 554
 555         blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
 556         if (freed > 0)
 557                 /* Got some memory back in the last second. */
 558                 return;
 559
 560         /*
 561          * If this is from memcg, oom-killer is already invoked.
 562          * and not worth to go system-wide-oom.
 563          */
 564         if (mem_cgroup_oom_called(current))
 565                 goto rest_and_return;
 566
 567         if (sysctl_panic_on_oom)
 568                 panic("out of memory from page fault. panic_on_oom is selected.\n");
 569
 570         read_lock(&tasklist_lock);
 571         __out_of_memory(0, 0); /* unknown gfp_mask and order */
 572         read_unlock(&tasklist_lock);
 573
 574         /*
 575          * Give "p" a good chance of killing itself before we
 576          * retry to allocate memory.
 577          */
 578 rest_and_return:
 579         if (!test_thread_flag(TIF_MEMDIE))
 580                 schedule_timeout_uninterruptible(1);
 581 }
 582
 583 /**
 584  * out_of_memory - kill the "best" process when we run out of memory
 585  * @zonelist: zonelist pointer
 586  * @gfp_mask: memory allocation flags
 587  * @order: amount of memory being requested as a power of 2
 588  *
 589  * If we run out of memory, we have the choice between either
 590  * killing a random task (bad), letting the system crash (worse)
 591  * OR try to be smart about which process to kill. Note that we
 592  * don't have to be perfect here, we just have to be good.
 593  */
 594 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
 595 {
 596         unsigned long freed = 0;
 597         enum oom_constraint constraint;
 598
 599         blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
 600         if (freed > 0)
 601                 /* Got some memory back in the last second. */
 602                 return;
 603
 604         if (sysctl_panic_on_oom == 2)
 605                 panic("out of memory. Compulsory panic_on_oom is selected.\n");
 606
 607         /*
 608          * Check if there were limitations on the allocation (only relevant for
 609          * NUMA) that may require different handling.
 610          */
 611         constraint = constrained_alloc(zonelist, gfp_mask);
 612         read_lock(&tasklist_lock);
 613
 614         switch (constraint) {
 615         case CONSTRAINT_MEMORY_POLICY:
 616                 oom_kill_process(current, gfp_mask, order, 0, NULL,
 617                                 "No available memory (MPOL_BIND)");
 618                 break;
 619
 620         case CONSTRAINT_NONE:
 621                 if (sysctl_panic_on_oom)
 622                         panic("out of memory. panic_on_oom is selected\n");
 623                 /* Fall-through */
 624         case CONSTRAINT_CPUSET:
 625                 __out_of_memory(gfp_mask, order);
 626                 break;
 627         }
 628
 629         read_unlock(&tasklist_lock);
 630
 631         /*
 632          * Give "p" a good chance of killing itself before we
 633          * retry to allocate memory unless "p" is current
 634          */
 635         if (!test_thread_flag(TIF_MEMDIE))
 636                 schedule_timeout_uninterruptible(1);
 637 }