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;
 181         int node;
 182
 183         nodes_clear(nodes);
 184         /* node has memory ? */
 185         for_each_online_node(node)
 186                 if (NODE_DATA(node)->node_present_pages)
 187                         node_set(node, nodes);
 188
 189         for (z = zonelist->zones; *z; z++)
 190                 if (cpuset_zone_allowed_softwall(*z, gfp_mask))
 191                         node_clear(zone_to_nid(*z), nodes);
 192                 else
 193                         return CONSTRAINT_CPUSET;
 194
 195         if (!nodes_empty(nodes))
 196                 return CONSTRAINT_MEMORY_POLICY;
 197 #endif
 198
 199         return CONSTRAINT_NONE;
 200 }
 201
 202 /*
 203  * Simple selection loop. We chose the process with the highest
 204  * number of 'points'. We expect the caller will lock the tasklist.
 205  *
 206  * (not docbooked, we don't want this one cluttering up the manual)
 207  */
 208 static struct task_struct *select_bad_process(unsigned long *ppoints)
 209 {
 210         struct task_struct *g, *p;
 211         struct task_struct *chosen = NULL;
 212         struct timespec uptime;
 213         *ppoints = 0;
 214
 215         do_posix_clock_monotonic_gettime(&uptime);
 216         do_each_thread(g, p) {
 217                 unsigned long points;
 218
 219                 /*
 220                  * skip kernel threads and tasks which have already released
 221                  * their mm.
 222                  */
 223                 if (!p->mm)
 224                         continue;
 225                 /* skip the init task */
 226                 if (is_init(p))
 227                         continue;
 228
 229                 /*
 230                  * This task already has access to memory reserves and is
 231                  * being killed. Don't allow any other task access to the
 232                  * memory reserve.
 233                  *
 234                  * Note: this may have a chance of deadlock if it gets
 235                  * blocked waiting for another task which itself is waiting
 236                  * for memory. Is there a better alternative?
 237                  */
 238                 if (test_tsk_thread_flag(p, TIF_MEMDIE))
 239                         return ERR_PTR(-1UL);
 240
 241                 /*
 242                  * This is in the process of releasing memory so wait for it
 243                  * to finish before killing some other task by mistake.
 244                  *
 245                  * However, if p is the current task, we allow the 'kill' to
 246                  * go ahead if it is exiting: this will simply set TIF_MEMDIE,
 247                  * which will allow it to gain access to memory reserves in
 248                  * the process of exiting and releasing its resources.
 249                  * Otherwise we could get an easy OOM deadlock.
 250                  */
 251                 if (p->flags & PF_EXITING) {
 252                         if (p != current)
 253                                 return ERR_PTR(-1UL);
 254
 255                         chosen = p;
 256                         *ppoints = ULONG_MAX;
 257                 }
 258
 259                 if (p->oomkilladj == OOM_DISABLE)
 260                         continue;
 261
 262                 points = badness(p, uptime.tv_sec);
 263                 if (points > *ppoints || !chosen) {
 264                         chosen = p;
 265                         *ppoints = points;
 266                 }
 267         } while_each_thread(g, p);
 268
 269         return chosen;
 270 }
 271
 272 /**
 273  * Send SIGKILL to the selected  process irrespective of  CAP_SYS_RAW_IO
 274  * flag though it's unlikely that  we select a process with CAP_SYS_RAW_IO
 275  * set.
 276  */
 277 static void __oom_kill_task(struct task_struct *p, int verbose)
 278 {
 279         if (is_init(p)) {
 280                 WARN_ON(1);
 281                 printk(KERN_WARNING "tried to kill init!\n");
 282                 return;
 283         }
 284
 285         if (!p->mm) {
 286                 WARN_ON(1);
 287                 printk(KERN_WARNING "tried to kill an mm-less task!\n");
 288                 return;
 289         }
 290
 291         if (verbose)
 292                 printk(KERN_ERR "Killed process %d (%s)\n", p->pid, p->comm);
 293
 294         /*
 295          * We give our sacrificial lamb high priority and access to
 296          * all the memory it needs. That way it should be able to
 297          * exit() and clear out its resources quickly...
 298          */
 299         p->time_slice = HZ;
 300         set_tsk_thread_flag(p, TIF_MEMDIE);
 301
 302         force_sig(SIGKILL, p);
 303 }
 304
 305 static int oom_kill_task(struct task_struct *p)
 306 {
 307         struct mm_struct *mm;
 308         struct task_struct *g, *q;
 309
 310         mm = p->mm;
 311
 312         /* WARNING: mm may not be dereferenced since we did not obtain its
 313          * value from get_task_mm(p).  This is OK since all we need to do is
 314          * compare mm to q->mm below.
 315          *
 316          * Furthermore, even if mm contains a non-NULL value, p->mm may
 317          * change to NULL at any time since we do not hold task_lock(p).
 318          * However, this is of no concern to us.
 319          */
 320
 321         if (mm == NULL)
 322                 return 1;
 323
 324         /*
 325          * Don't kill the process if any threads are set to OOM_DISABLE
 326          */
 327         do_each_thread(g, q) {
 328                 if (q->mm == mm && q->oomkilladj == OOM_DISABLE)
 329                         return 1;
 330         } while_each_thread(g, q);
 331
 332         __oom_kill_task(p, 1);
 333
 334         /*
 335          * kill all processes that share the ->mm (i.e. all threads),
 336          * but are in a different thread group. Don't let them have access
 337          * to memory reserves though, otherwise we might deplete all memory.
 338          */
 339         do_each_thread(g, q) {
 340                 if (q->mm == mm && q->tgid != p->tgid)
 341                         force_sig(SIGKILL, q);
 342         } while_each_thread(g, q);
 343
 344         return 0;
 345 }
 346
 347 static int oom_kill_process(struct task_struct *p, unsigned long points,
 348                 const char *message)
 349 {
 350         struct task_struct *c;
 351         struct list_head *tsk;
 352
 353         /*
 354          * If the task is already exiting, don't alarm the sysadmin or kill
 355          * its children or threads, just set TIF_MEMDIE so it can die quickly
 356          */
 357         if (p->flags & PF_EXITING) {
 358                 __oom_kill_task(p, 0);
 359                 return 0;
 360         }
 361
 362         printk(KERN_ERR "%s: kill process %d (%s) score %li or a child\n",
 363                                         message, p->pid, p->comm, points);
 364
 365         /* Try to kill a child first */
 366         list_for_each(tsk, &p->children) {
 367                 c = list_entry(tsk, struct task_struct, sibling);
 368                 if (c->mm == p->mm)
 369                         continue;
 370                 if (!oom_kill_task(c))
 371                         return 0;
 372         }
 373         return oom_kill_task(p);
 374 }
 375
 376 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
 377
 378 int register_oom_notifier(struct notifier_block *nb)
 379 {
 380         return blocking_notifier_chain_register(&oom_notify_list, nb);
 381 }
 382 EXPORT_SYMBOL_GPL(register_oom_notifier);
 383
 384 int unregister_oom_notifier(struct notifier_block *nb)
 385 {
 386         return blocking_notifier_chain_unregister(&oom_notify_list, nb);
 387 }
 388 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
 389
 390 /**
 391  * out_of_memory - kill the "best" process when we run out of memory
 392  *
 393  * If we run out of memory, we have the choice between either
 394  * killing a random task (bad), letting the system crash (worse)
 395  * OR try to be smart about which process to kill. Note that we
 396  * don't have to be perfect here, we just have to be good.
 397  */
 398 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
 399 {
 400         struct task_struct *p;
 401         unsigned long points = 0;
 402         unsigned long freed = 0;
 403         int constraint;
 404
 405         blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
 406         if (freed > 0)
 407                 /* Got some memory back in the last second. */
 408                 return;
 409
 410         if (printk_ratelimit()) {
 411                 printk(KERN_WARNING "%s invoked oom-killer: "
 412                         "gfp_mask=0x%x, order=%d, oomkilladj=%d\n",
 413                         current->comm, gfp_mask, order, current->oomkilladj);
 414                 dump_stack();
 415                 show_mem();
 416         }
 417
 418         if (sysctl_panic_on_oom == 2)
 419                 panic("out of memory. Compulsory panic_on_oom is selected.\n");
 420
 421         /*
 422          * Check if there were limitations on the allocation (only relevant for
 423          * NUMA) that may require different handling.
 424          */
 425         constraint = constrained_alloc(zonelist, gfp_mask);
 426         cpuset_lock();
 427         read_lock(&tasklist_lock);
 428
 429         switch (constraint) {
 430         case CONSTRAINT_MEMORY_POLICY:
 431                 oom_kill_process(current, points,
 432                                 "No available memory (MPOL_BIND)");
 433                 break;
 434
 435         case CONSTRAINT_CPUSET:
 436                 oom_kill_process(current, points,
 437                                 "No available memory in cpuset");
 438                 break;
 439
 440         case CONSTRAINT_NONE:
 441                 if (sysctl_panic_on_oom)
 442                         panic("out of memory. panic_on_oom is selected\n");
 443 retry:
 444                 /*
 445                  * Rambo mode: Shoot down a process and hope it solves whatever
 446                  * issues we may have.
 447                  */
 448                 p = select_bad_process(&points);
 449
 450                 if (PTR_ERR(p) == -1UL)
 451                         goto out;
 452
 453                 /* Found nothing?!?! Either we hang forever, or we panic. */
 454                 if (!p) {
 455                         read_unlock(&tasklist_lock);
 456                         cpuset_unlock();
 457                         panic("Out of memory and no killable processes...\n");
 458                 }
 459
 460                 if (oom_kill_process(p, points, "Out of memory"))
 461                         goto retry;
 462
 463                 break;
 464         }
 465
 466 out:
 467         read_unlock(&tasklist_lock);
 468         cpuset_unlock();
 469
 470         /*
 471          * Give "p" a good chance of killing itself before we
 472          * retry to allocate memory unless "p" is current
 473          */
 474         if (!test_thread_flag(TIF_MEMDIE))
 475                 schedule_timeout_uninterruptible(1);
 476 }