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/sched.h>
  21 #include <linux/swap.h>
  22 #include <linux/timex.h>
  23 #include <linux/jiffies.h>
  24 #include <linux/cpuset.h>
  25 #include <linux/module.h>
  26 #include <linux/notifier.h>
  27
  28 int sysctl_panic_on_oom;
  29 /* #define DEBUG */
  30
  31 /**
  32  * badness - calculate a numeric value for how bad this task has been
  33  * @p: task struct of which task we should calculate
  34  * @uptime: current uptime in seconds
  35  *
  36  * The formula used is relatively simple and documented inline in the
  37  * function. The main rationale is that we want to select a good task
  38  * to kill when we run out of memory.
  39  *
  40  * Good in this context means that:
  41  * 1) we lose the minimum amount of work done
  42  * 2) we recover a large amount of memory
  43  * 3) we don't kill anything innocent of eating tons of memory
  44  * 4) we want to kill the minimum amount of processes (one)
  45  * 5) we try to kill the process the user expects us to kill, this
  46  *    algorithm has been meticulously tuned to meet the principle
  47  *    of least surprise ... (be careful when you change it)
  48  */
  49
  50 unsigned long badness(struct task_struct *p, unsigned long uptime)
  51 {
  52         unsigned long points, cpu_time, run_time, s;
  53         struct mm_struct *mm;
  54         struct task_struct *child;
  55
  56         task_lock(p);
  57         mm = p->mm;
  58         if (!mm) {
  59                 task_unlock(p);
  60                 return 0;
  61         }
  62
  63         /*
  64          * The memory size of the process is the basis for the badness.
  65          */
  66         points = mm->total_vm;
  67
  68         /*
  69          * After this unlock we can no longer dereference local variable `mm'
  70          */
  71         task_unlock(p);
  72
  73         /*
  74          * swapoff can easily use up all memory, so kill those first.
  75          */
  76         if (p->flags & PF_SWAPOFF)
  77                 return ULONG_MAX;
  78
  79         /*
  80          * Processes which fork a lot of child processes are likely
  81          * a good choice. We add half the vmsize of the children if they
  82          * have an own mm. This prevents forking servers to flood the
  83          * machine with an endless amount of children. In case a single
  84          * child is eating the vast majority of memory, adding only half
  85          * to the parents will make the child our kill candidate of choice.
  86          */
  87         list_for_each_entry(child, &p->children, sibling) {
  88                 task_lock(child);
  89                 if (child->mm != mm && child->mm)
  90                         points += child->mm->total_vm/2 + 1;
  91                 task_unlock(child);
  92         }
  93
  94         /*
  95          * CPU time is in tens of seconds and run time is in thousands
  96          * of seconds. There is no particular reason for this other than
  97          * that it turned out to work very well in practice.
  98          */
  99         cpu_time = (cputime_to_jiffies(p->utime) + cputime_to_jiffies(p->stime))
 100                 >> (SHIFT_HZ + 3);
 101
 102         if (uptime >= p->start_time.tv_sec)
 103                 run_time = (uptime - p->start_time.tv_sec) >> 10;
 104         else
 105                 run_time = 0;
 106
 107         s = int_sqrt(cpu_time);
 108         if (s)
 109                 points /= s;
 110         s = int_sqrt(int_sqrt(run_time));
 111         if (s)
 112                 points /= s;
 113
 114         /*
 115          * Niced processes are most likely less important, so double
 116          * their badness points.
 117          */
 118         if (task_nice(p) > 0)
 119                 points *= 2;
 120
 121         /*
 122          * Superuser processes are usually more important, so we make it
 123          * less likely that we kill those.
 124          */
 125         if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_ADMIN) ||
 126                                 p->uid == 0 || p->euid == 0)
 127                 points /= 4;
 128
 129         /*
 130          * We don't want to kill a process with direct hardware access.
 131          * Not only could that mess up the hardware, but usually users
 132          * tend to only have this flag set on applications they think
 133          * of as important.
 134          */
 135         if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_RAWIO))
 136                 points /= 4;
 137
 138         /*
 139          * If p's nodes don't overlap ours, it may still help to kill p
 140          * because p may have allocated or otherwise mapped memory on
 141          * this node before. However it will be less likely.
 142          */
 143         if (!cpuset_excl_nodes_overlap(p))
 144                 points /= 8;
 145
 146         /*
 147          * Adjust the score by oomkilladj.
 148          */
 149         if (p->oomkilladj) {
 150                 if (p->oomkilladj > 0) {
 151                         if (!points)
 152                                 points = 1;
 153                         points <<= p->oomkilladj;
 154                 } else
 155                         points >>= -(p->oomkilladj);
 156         }
 157
 158 #ifdef DEBUG
 159         printk(KERN_DEBUG "OOMkill: task %d (%s) got %d points\n",
 160         p->pid, p->comm, points);
 161 #endif
 162         return points;
 163 }
 164
 165 /*
 166  * Types of limitations to the nodes from which allocations may occur
 167  */
 168 #define CONSTRAINT_NONE 1
 169 #define CONSTRAINT_MEMORY_POLICY 2
 170 #define CONSTRAINT_CPUSET 3
 171
 172 /*
 173  * Determine the type of allocation constraint.
 174  */
 175 static inline int constrained_alloc(struct zonelist *zonelist, gfp_t gfp_mask)
 176 {
 177 #ifdef CONFIG_NUMA
 178         struct zone **z;
 179         nodemask_t nodes;
 180         int node;
 181
 182         nodes_clear(nodes);
 183         /* node has memory ? */
 184         for_each_online_node(node)
 185                 if (NODE_DATA(node)->node_present_pages)
 186                         node_set(node, nodes);
 187
 188         for (z = zonelist->zones; *z; z++)
 189                 if (cpuset_zone_allowed_softwall(*z, gfp_mask))
 190                         node_clear(zone_to_nid(*z), nodes);
 191                 else
 192                         return CONSTRAINT_CPUSET;
 193
 194         if (!nodes_empty(nodes))
 195                 return CONSTRAINT_MEMORY_POLICY;
 196 #endif
 197
 198         return CONSTRAINT_NONE;
 199 }
 200
 201 /*
 202  * Simple selection loop. We chose the process with the highest
 203  * number of 'points'. We expect the caller will lock the tasklist.
 204  *
 205  * (not docbooked, we don't want this one cluttering up the manual)
 206  */
 207 static struct task_struct *select_bad_process(unsigned long *ppoints)
 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_init(p))
 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->oomkilladj == 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  * Send SIGKILL to the selected  process irrespective of  CAP_SYS_RAW_IO
 273  * flag though it's unlikely that  we select a process with CAP_SYS_RAW_IO
 274  * set.
 275  */
 276 static void __oom_kill_task(struct task_struct *p, int verbose)
 277 {
 278         if (is_init(p)) {
 279                 WARN_ON(1);
 280                 printk(KERN_WARNING "tried to kill init!\n");
 281                 return;
 282         }
 283
 284         if (!p->mm) {
 285                 WARN_ON(1);
 286                 printk(KERN_WARNING "tried to kill an mm-less task!\n");
 287                 return;
 288         }
 289
 290         if (verbose)
 291                 printk(KERN_ERR "Killed process %d (%s)\n", p->pid, 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->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 && q->tgid != p->tgid)
 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, unsigned long points,
 347                 const char *message)
 348 {
 349         struct task_struct *c;
 350         struct list_head *tsk;
 351
 352         /*
 353          * If the task is already exiting, don't alarm the sysadmin or kill
 354          * its children or threads, just set TIF_MEMDIE so it can die quickly
 355          */
 356         if (p->flags & PF_EXITING) {
 357                 __oom_kill_task(p, 0);
 358                 return 0;
 359         }
 360
 361         printk(KERN_ERR "%s: kill process %d (%s) score %li or a child\n",
 362                                         message, p->pid, p->comm, points);
 363
 364         /* Try to kill a child first */
 365         list_for_each(tsk, &p->children) {
 366                 c = list_entry(tsk, struct task_struct, sibling);
 367                 if (c->mm == p->mm)
 368                         continue;
 369                 if (!oom_kill_task(c))
 370                         return 0;
 371         }
 372         return oom_kill_task(p);
 373 }
 374
 375 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
 376
 377 int register_oom_notifier(struct notifier_block *nb)
 378 {
 379         return blocking_notifier_chain_register(&oom_notify_list, nb);
 380 }
 381 EXPORT_SYMBOL_GPL(register_oom_notifier);
 382
 383 int unregister_oom_notifier(struct notifier_block *nb)
 384 {
 385         return blocking_notifier_chain_unregister(&oom_notify_list, nb);
 386 }
 387 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
 388
 389 /**
 390  * out_of_memory - kill the "best" process when we run out of memory
 391  *
 392  * If we run out of memory, we have the choice between either
 393  * killing a random task (bad), letting the system crash (worse)
 394  * OR try to be smart about which process to kill. Note that we
 395  * don't have to be perfect here, we just have to be good.
 396  */
 397 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
 398 {
 399         struct task_struct *p;
 400         unsigned long points = 0;
 401         unsigned long freed = 0;
 402
 403         blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
 404         if (freed > 0)
 405                 /* Got some memory back in the last second. */
 406                 return;
 407
 408         if (printk_ratelimit()) {
 409                 printk(KERN_WARNING "%s invoked oom-killer: "
 410                         "gfp_mask=0x%x, order=%d, oomkilladj=%d\n",
 411                         current->comm, gfp_mask, order, current->oomkilladj);
 412                 dump_stack();
 413                 show_mem();
 414         }
 415
 416         cpuset_lock();
 417         read_lock(&tasklist_lock);
 418
 419         /*
 420          * Check if there were limitations on the allocation (only relevant for
 421          * NUMA) that may require different handling.
 422          */
 423         switch (constrained_alloc(zonelist, gfp_mask)) {
 424         case CONSTRAINT_MEMORY_POLICY:
 425                 oom_kill_process(current, points,
 426                                 "No available memory (MPOL_BIND)");
 427                 break;
 428
 429         case CONSTRAINT_CPUSET:
 430                 oom_kill_process(current, points,
 431                                 "No available memory in cpuset");
 432                 break;
 433
 434         case CONSTRAINT_NONE:
 435                 if (sysctl_panic_on_oom)
 436                         panic("out of memory. panic_on_oom is selected\n");
 437 retry:
 438                 /*
 439                  * Rambo mode: Shoot down a process and hope it solves whatever
 440                  * issues we may have.
 441                  */
 442                 p = select_bad_process(&points);
 443
 444                 if (PTR_ERR(p) == -1UL)
 445                         goto out;
 446
 447                 /* Found nothing?!?! Either we hang forever, or we panic. */
 448                 if (!p) {
 449                         read_unlock(&tasklist_lock);
 450                         cpuset_unlock();
 451                         panic("Out of memory and no killable processes...\n");
 452                 }
 453
 454                 if (oom_kill_process(p, points, "Out of memory"))
 455                         goto retry;
 456
 457                 break;
 458         }
 459
 460 out:
 461         read_unlock(&tasklist_lock);
 462         cpuset_unlock();
 463
 464         /*
 465          * Give "p" a good chance of killing itself before we
 466          * retry to allocate memory unless "p" is current
 467          */
 468         if (!test_thread_flag(TIF_MEMDIE))
 469                 schedule_timeout_uninterruptible(1);
 470 }