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