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          * swapoff can easily use up all memory, so kill those first.
  65          */
  66         if (p->flags & PF_SWAPOFF)
  67                 return ULONG_MAX;
  68
  69         /*
  70          * The memory size of the process is the basis for the badness.
  71          */
  72         points = mm->total_vm;
  73
  74         /*
  75          * After this unlock we can no longer dereference local variable `mm'
  76          */
  77         task_unlock(p);
  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                         points <<= p->oomkilladj;
 152                 else
 153                         points >>= -(p->oomkilladj);
 154         }
 155
 156 #ifdef DEBUG
 157         printk(KERN_DEBUG "OOMkill: task %d (%s) got %d points\n",
 158         p->pid, p->comm, points);
 159 #endif
 160         return points;
 161 }
 162
 163 /*
 164  * Types of limitations to the nodes from which allocations may occur
 165  */
 166 #define CONSTRAINT_NONE 1
 167 #define CONSTRAINT_MEMORY_POLICY 2
 168 #define CONSTRAINT_CPUSET 3
 169
 170 /*
 171  * Determine the type of allocation constraint.
 172  */
 173 static inline int constrained_alloc(struct zonelist *zonelist, gfp_t gfp_mask)
 174 {
 175 #ifdef CONFIG_NUMA
 176         struct zone **z;
 177         nodemask_t nodes;
 178         int node;
 179         /* node has memory ? */
 180         for_each_online_node(node)
 181                 if (NODE_DATA(node)->node_present_pages)
 182                         node_set(node, nodes);
 183
 184         for (z = zonelist->zones; *z; z++)
 185                 if (cpuset_zone_allowed_softwall(*z, gfp_mask))
 186                         node_clear(zone_to_nid(*z), nodes);
 187                 else
 188                         return CONSTRAINT_CPUSET;
 189
 190         if (!nodes_empty(nodes))
 191                 return CONSTRAINT_MEMORY_POLICY;
 192 #endif
 193
 194         return CONSTRAINT_NONE;
 195 }
 196
 197 /*
 198  * Simple selection loop. We chose the process with the highest
 199  * number of 'points'. We expect the caller will lock the tasklist.
 200  *
 201  * (not docbooked, we don't want this one cluttering up the manual)
 202  */
 203 static struct task_struct *select_bad_process(unsigned long *ppoints)
 204 {
 205         struct task_struct *g, *p;
 206         struct task_struct *chosen = NULL;
 207         struct timespec uptime;
 208         *ppoints = 0;
 209
 210         do_posix_clock_monotonic_gettime(&uptime);
 211         do_each_thread(g, p) {
 212                 unsigned long points;
 213
 214                 /*
 215                  * skip kernel threads and tasks which have already released
 216                  * their mm.
 217                  */
 218                 if (!p->mm)
 219                         continue;
 220                 /* skip the init task */
 221                 if (is_init(p))
 222                         continue;
 223
 224                 /*
 225                  * This task already has access to memory reserves and is
 226                  * being killed. Don't allow any other task access to the
 227                  * memory reserve.
 228                  *
 229                  * Note: this may have a chance of deadlock if it gets
 230                  * blocked waiting for another task which itself is waiting
 231                  * for memory. Is there a better alternative?
 232                  */
 233                 if (test_tsk_thread_flag(p, TIF_MEMDIE))
 234                         return ERR_PTR(-1UL);
 235
 236                 /*
 237                  * This is in the process of releasing memory so wait for it
 238                  * to finish before killing some other task by mistake.
 239                  *
 240                  * However, if p is the current task, we allow the 'kill' to
 241                  * go ahead if it is exiting: this will simply set TIF_MEMDIE,
 242                  * which will allow it to gain access to memory reserves in
 243                  * the process of exiting and releasing its resources.
 244                  * Otherwise we could get an easy OOM deadlock.
 245                  */
 246                 if (p->flags & PF_EXITING) {
 247                         if (p != current)
 248                                 return ERR_PTR(-1UL);
 249
 250                         chosen = p;
 251                         *ppoints = ULONG_MAX;
 252                 }
 253
 254                 if (p->oomkilladj == OOM_DISABLE)
 255                         continue;
 256
 257                 points = badness(p, uptime.tv_sec);
 258                 if (points > *ppoints || !chosen) {
 259                         chosen = p;
 260                         *ppoints = points;
 261                 }
 262         } while_each_thread(g, p);
 263
 264         return chosen;
 265 }
 266
 267 /**
 268  * Send SIGKILL to the selected  process irrespective of  CAP_SYS_RAW_IO
 269  * flag though it's unlikely that  we select a process with CAP_SYS_RAW_IO
 270  * set.
 271  */
 272 static void __oom_kill_task(struct task_struct *p, int verbose)
 273 {
 274         if (is_init(p)) {
 275                 WARN_ON(1);
 276                 printk(KERN_WARNING "tried to kill init!\n");
 277                 return;
 278         }
 279
 280         if (!p->mm) {
 281                 WARN_ON(1);
 282                 printk(KERN_WARNING "tried to kill an mm-less task!\n");
 283                 return;
 284         }
 285
 286         if (verbose)
 287                 printk(KERN_ERR "Killed process %d (%s)\n", p->pid, p->comm);
 288
 289         /*
 290          * We give our sacrificial lamb high priority and access to
 291          * all the memory it needs. That way it should be able to
 292          * exit() and clear out its resources quickly...
 293          */
 294         p->time_slice = HZ;
 295         set_tsk_thread_flag(p, TIF_MEMDIE);
 296
 297         force_sig(SIGKILL, p);
 298 }
 299
 300 static int oom_kill_task(struct task_struct *p)
 301 {
 302         struct mm_struct *mm;
 303         struct task_struct *g, *q;
 304
 305         mm = p->mm;
 306
 307         /* WARNING: mm may not be dereferenced since we did not obtain its
 308          * value from get_task_mm(p).  This is OK since all we need to do is
 309          * compare mm to q->mm below.
 310          *
 311          * Furthermore, even if mm contains a non-NULL value, p->mm may
 312          * change to NULL at any time since we do not hold task_lock(p).
 313          * However, this is of no concern to us.
 314          */
 315
 316         if (mm == NULL)
 317                 return 1;
 318
 319         /*
 320          * Don't kill the process if any threads are set to OOM_DISABLE
 321          */
 322         do_each_thread(g, q) {
 323                 if (q->mm == mm && p->oomkilladj == OOM_DISABLE)
 324                         return 1;
 325         } while_each_thread(g, q);
 326
 327         __oom_kill_task(p, 1);
 328
 329         /*
 330          * kill all processes that share the ->mm (i.e. all threads),
 331          * but are in a different thread group. Don't let them have access
 332          * to memory reserves though, otherwise we might deplete all memory.
 333          */
 334         do_each_thread(g, q) {
 335                 if (q->mm == mm && q->tgid != p->tgid)
 336                         force_sig(SIGKILL, p);
 337         } while_each_thread(g, q);
 338
 339         return 0;
 340 }
 341
 342 static int oom_kill_process(struct task_struct *p, unsigned long points,
 343                 const char *message)
 344 {
 345         struct task_struct *c;
 346         struct list_head *tsk;
 347
 348         /*
 349          * If the task is already exiting, don't alarm the sysadmin or kill
 350          * its children or threads, just set TIF_MEMDIE so it can die quickly
 351          */
 352         if (p->flags & PF_EXITING) {
 353                 __oom_kill_task(p, 0);
 354                 return 0;
 355         }
 356
 357         printk(KERN_ERR "%s: kill process %d (%s) score %li or a child\n",
 358                                         message, p->pid, p->comm, points);
 359
 360         /* Try to kill a child first */
 361         list_for_each(tsk, &p->children) {
 362                 c = list_entry(tsk, struct task_struct, sibling);
 363                 if (c->mm == p->mm)
 364                         continue;
 365                 if (!oom_kill_task(c))
 366                         return 0;
 367         }
 368         return oom_kill_task(p);
 369 }
 370
 371 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
 372
 373 int register_oom_notifier(struct notifier_block *nb)
 374 {
 375         return blocking_notifier_chain_register(&oom_notify_list, nb);
 376 }
 377 EXPORT_SYMBOL_GPL(register_oom_notifier);
 378
 379 int unregister_oom_notifier(struct notifier_block *nb)
 380 {
 381         return blocking_notifier_chain_unregister(&oom_notify_list, nb);
 382 }
 383 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
 384
 385 /**
 386  * out_of_memory - kill the "best" process when we run out of memory
 387  *
 388  * If we run out of memory, we have the choice between either
 389  * killing a random task (bad), letting the system crash (worse)
 390  * OR try to be smart about which process to kill. Note that we
 391  * don't have to be perfect here, we just have to be good.
 392  */
 393 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
 394 {
 395         struct task_struct *p;
 396         unsigned long points = 0;
 397         unsigned long freed = 0;
 398
 399         blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
 400         if (freed > 0)
 401                 /* Got some memory back in the last second. */
 402                 return;
 403
 404         if (printk_ratelimit()) {
 405                 printk(KERN_WARNING "%s invoked oom-killer: "
 406                         "gfp_mask=0x%x, order=%d, oomkilladj=%d\n",
 407                         current->comm, gfp_mask, order, current->oomkilladj);
 408                 dump_stack();
 409                 show_mem();
 410         }
 411
 412         cpuset_lock();
 413         read_lock(&tasklist_lock);
 414
 415         /*
 416          * Check if there were limitations on the allocation (only relevant for
 417          * NUMA) that may require different handling.
 418          */
 419         switch (constrained_alloc(zonelist, gfp_mask)) {
 420         case CONSTRAINT_MEMORY_POLICY:
 421                 oom_kill_process(current, points,
 422                                 "No available memory (MPOL_BIND)");
 423                 break;
 424
 425         case CONSTRAINT_CPUSET:
 426                 oom_kill_process(current, points,
 427                                 "No available memory in cpuset");
 428                 break;
 429
 430         case CONSTRAINT_NONE:
 431                 if (sysctl_panic_on_oom)
 432                         panic("out of memory. panic_on_oom is selected\n");
 433 retry:
 434                 /*
 435                  * Rambo mode: Shoot down a process and hope it solves whatever
 436                  * issues we may have.
 437                  */
 438                 p = select_bad_process(&points);
 439
 440                 if (PTR_ERR(p) == -1UL)
 441                         goto out;
 442
 443                 /* Found nothing?!?! Either we hang forever, or we panic. */
 444                 if (!p) {
 445                         read_unlock(&tasklist_lock);
 446                         cpuset_unlock();
 447                         panic("Out of memory and no killable processes...\n");
 448                 }
 449
 450                 if (oom_kill_process(p, points, "Out of memory"))
 451                         goto retry;
 452
 453                 break;
 454         }
 455
 456 out:
 457         read_unlock(&tasklist_lock);
 458         cpuset_unlock();
 459
 460         /*
 461          * Give "p" a good chance of killing itself before we
 462          * retry to allocate memory unless "p" is current
 463          */
 464         if (!test_thread_flag(TIF_MEMDIE))
 465                 schedule_timeout_uninterruptible(1);
 466 }