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/mm.h>
  19 #include <linux/sched.h>
  20 #include <linux/swap.h>
  21 #include <linux/timex.h>
  22 #include <linux/jiffies.h>
  23 #include <linux/cpuset.h>
  24 #include <linux/module.h>
  25 #include <linux/notifier.h>
  26
  27 int sysctl_panic_on_oom;
  28 /* #define DEBUG */
  29
  30 /**
  31  * badness - calculate a numeric value for how bad this task has been
  32  * @p: task struct of which task we should calculate
  33  * @uptime: current uptime in seconds
  34  *
  35  * The formula used is relatively simple and documented inline in the
  36  * function. The main rationale is that we want to select a good task
  37  * to kill when we run out of memory.
  38  *
  39  * Good in this context means that:
  40  * 1) we lose the minimum amount of work done
  41  * 2) we recover a large amount of memory
  42  * 3) we don't kill anything innocent of eating tons of memory
  43  * 4) we want to kill the minimum amount of processes (one)
  44  * 5) we try to kill the process the user expects us to kill, this
  45  *    algorithm has been meticulously tuned to meet the principle
  46  *    of least surprise ... (be careful when you change it)
  47  */
  48
  49 unsigned long badness(struct task_struct *p, unsigned long uptime)
  50 {
  51         unsigned long points, cpu_time, run_time, s;
  52         struct mm_struct *mm;
  53         struct task_struct *child;
  54
  55         task_lock(p);
  56         mm = p->mm;
  57         if (!mm) {
  58                 task_unlock(p);
  59                 return 0;
  60         }
  61
  62         /*
  63          * swapoff can easily use up all memory, so kill those first.
  64          */
  65         if (p->flags & PF_SWAPOFF)
  66                 return ULONG_MAX;
  67
  68         /*
  69          * The memory size of the process is the basis for the badness.
  70          */
  71         points = mm->total_vm;
  72
  73         /*
  74          * After this unlock we can no longer dereference local variable `mm'
  75          */
  76         task_unlock(p);
  77
  78         /*
  79          * Processes which fork a lot of child processes are likely
  80          * a good choice. We add half the vmsize of the children if they
  81          * have an own mm. This prevents forking servers to flood the
  82          * machine with an endless amount of children. In case a single
  83          * child is eating the vast majority of memory, adding only half
  84          * to the parents will make the child our kill candidate of choice.
  85          */
  86         list_for_each_entry(child, &p->children, sibling) {
  87                 task_lock(child);
  88                 if (child->mm != mm && child->mm)
  89                         points += child->mm->total_vm/2 + 1;
  90                 task_unlock(child);
  91         }
  92
  93         /*
  94          * CPU time is in tens of seconds and run time is in thousands
  95          * of seconds. There is no particular reason for this other than
  96          * that it turned out to work very well in practice.
  97          */
  98         cpu_time = (cputime_to_jiffies(p->utime) + cputime_to_jiffies(p->stime))
  99                 >> (SHIFT_HZ + 3);
 100
 101         if (uptime >= p->start_time.tv_sec)
 102                 run_time = (uptime - p->start_time.tv_sec) >> 10;
 103         else
 104                 run_time = 0;
 105
 106         s = int_sqrt(cpu_time);
 107         if (s)
 108                 points /= s;
 109         s = int_sqrt(int_sqrt(run_time));
 110         if (s)
 111                 points /= s;
 112
 113         /*
 114          * Niced processes are most likely less important, so double
 115          * their badness points.
 116          */
 117         if (task_nice(p) > 0)
 118                 points *= 2;
 119
 120         /*
 121          * Superuser processes are usually more important, so we make it
 122          * less likely that we kill those.
 123          */
 124         if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_ADMIN) ||
 125                                 p->uid == 0 || p->euid == 0)
 126                 points /= 4;
 127
 128         /*
 129          * We don't want to kill a process with direct hardware access.
 130          * Not only could that mess up the hardware, but usually users
 131          * tend to only have this flag set on applications they think
 132          * of as important.
 133          */
 134         if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_RAWIO))
 135                 points /= 4;
 136
 137         /*
 138          * If p's nodes don't overlap ours, it may still help to kill p
 139          * because p may have allocated or otherwise mapped memory on
 140          * this node before. However it will be less likely.
 141          */
 142         if (!cpuset_excl_nodes_overlap(p))
 143                 points /= 8;
 144
 145         /*
 146          * Adjust the score by oomkilladj.
 147          */
 148         if (p->oomkilladj) {
 149                 if (p->oomkilladj > 0)
 150                         points <<= p->oomkilladj;
 151                 else
 152                         points >>= -(p->oomkilladj);
 153         }
 154
 155 #ifdef DEBUG
 156         printk(KERN_DEBUG "OOMkill: task %d (%s) got %d points\n",
 157         p->pid, p->comm, points);
 158 #endif
 159         return points;
 160 }
 161
 162 /*
 163  * Types of limitations to the nodes from which allocations may occur
 164  */
 165 #define CONSTRAINT_NONE 1
 166 #define CONSTRAINT_MEMORY_POLICY 2
 167 #define CONSTRAINT_CPUSET 3
 168
 169 /*
 170  * Determine the type of allocation constraint.
 171  */
 172 static inline int constrained_alloc(struct zonelist *zonelist, gfp_t gfp_mask)
 173 {
 174 #ifdef CONFIG_NUMA
 175         struct zone **z;
 176         nodemask_t nodes = node_online_map;
 177
 178         for (z = zonelist->zones; *z; z++)
 179                 if (cpuset_zone_allowed(*z, gfp_mask))
 180                         node_clear(zone_to_nid(*z), nodes);
 181                 else
 182                         return CONSTRAINT_CPUSET;
 183
 184         if (!nodes_empty(nodes))
 185                 return CONSTRAINT_MEMORY_POLICY;
 186 #endif
 187
 188         return CONSTRAINT_NONE;
 189 }
 190
 191 /*
 192  * Simple selection loop. We chose the process with the highest
 193  * number of 'points'. We expect the caller will lock the tasklist.
 194  *
 195  * (not docbooked, we don't want this one cluttering up the manual)
 196  */
 197 static struct task_struct *select_bad_process(unsigned long *ppoints)
 198 {
 199         struct task_struct *g, *p;
 200         struct task_struct *chosen = NULL;
 201         struct timespec uptime;
 202         *ppoints = 0;
 203
 204         do_posix_clock_monotonic_gettime(&uptime);
 205         do_each_thread(g, p) {
 206                 unsigned long points;
 207
 208                 /*
 209                  * skip kernel threads and tasks which have already released
 210                  * their mm.
 211                  */
 212                 if (!p->mm)
 213                         continue;
 214                 /* skip the init task */
 215                 if (is_init(p))
 216                         continue;
 217
 218                 /*
 219                  * This is in the process of releasing memory so wait for it
 220                  * to finish before killing some other task by mistake.
 221                  *
 222                  * However, if p is the current task, we allow the 'kill' to
 223                  * go ahead if it is exiting: this will simply set TIF_MEMDIE,
 224                  * which will allow it to gain access to memory reserves in
 225                  * the process of exiting and releasing its resources.
 226                  * Otherwise we could get an OOM deadlock.
 227                  */
 228                 if ((p->flags & PF_EXITING) && p == current) {
 229                         chosen = p;
 230                         *ppoints = ULONG_MAX;
 231                         break;
 232                 }
 233                 if ((p->flags & PF_EXITING) ||
 234                                 test_tsk_thread_flag(p, TIF_MEMDIE))
 235                         return ERR_PTR(-1UL);
 236
 237                 if (p->oomkilladj == OOM_DISABLE)
 238                         continue;
 239
 240                 points = badness(p, uptime.tv_sec);
 241                 if (points > *ppoints || !chosen) {
 242                         chosen = p;
 243                         *ppoints = points;
 244                 }
 245         } while_each_thread(g, p);
 246
 247         return chosen;
 248 }
 249
 250 /**
 251  * Send SIGKILL to the selected  process irrespective of  CAP_SYS_RAW_IO
 252  * flag though it's unlikely that  we select a process with CAP_SYS_RAW_IO
 253  * set.
 254  */
 255 static void __oom_kill_task(struct task_struct *p, const char *message)
 256 {
 257         if (is_init(p)) {
 258                 WARN_ON(1);
 259                 printk(KERN_WARNING "tried to kill init!\n");
 260                 return;
 261         }
 262
 263         task_lock(p);
 264         if (!p->mm || p->mm == &init_mm) {
 265                 WARN_ON(1);
 266                 printk(KERN_WARNING "tried to kill an mm-less task!\n");
 267                 task_unlock(p);
 268                 return;
 269         }
 270         task_unlock(p);
 271
 272         if (message) {
 273                 printk(KERN_ERR "%s: Killed process %d (%s).\n",
 274                                 message, p->pid, p->comm);
 275         }
 276
 277         /*
 278          * We give our sacrificial lamb high priority and access to
 279          * all the memory it needs. That way it should be able to
 280          * exit() and clear out its resources quickly...
 281          */
 282         p->time_slice = HZ;
 283         set_tsk_thread_flag(p, TIF_MEMDIE);
 284
 285         force_sig(SIGKILL, p);
 286 }
 287
 288 static int oom_kill_task(struct task_struct *p, const char *message)
 289 {
 290         struct mm_struct *mm;
 291         struct task_struct *g, *q;
 292
 293         mm = p->mm;
 294
 295         /* WARNING: mm may not be dereferenced since we did not obtain its
 296          * value from get_task_mm(p).  This is OK since all we need to do is
 297          * compare mm to q->mm below.
 298          *
 299          * Furthermore, even if mm contains a non-NULL value, p->mm may
 300          * change to NULL at any time since we do not hold task_lock(p).
 301          * However, this is of no concern to us.
 302          */
 303
 304         if (mm == NULL || mm == &init_mm)
 305                 return 1;
 306
 307         __oom_kill_task(p, message);
 308         /*
 309          * kill all processes that share the ->mm (i.e. all threads),
 310          * but are in a different thread group
 311          */
 312         do_each_thread(g, q)
 313                 if (q->mm == mm && q->tgid != p->tgid)
 314                         __oom_kill_task(q, message);
 315         while_each_thread(g, q);
 316
 317         return 0;
 318 }
 319
 320 static int oom_kill_process(struct task_struct *p, unsigned long points,
 321                 const char *message)
 322 {
 323         struct task_struct *c;
 324         struct list_head *tsk;
 325
 326         /*
 327          * If the task is already exiting, don't alarm the sysadmin or kill
 328          * its children or threads, just set TIF_MEMDIE so it can die quickly
 329          */
 330         if (p->flags & PF_EXITING) {
 331                 __oom_kill_task(p, NULL);
 332                 return 0;
 333         }
 334
 335         printk(KERN_ERR "Out of Memory: Kill process %d (%s) score %li"
 336                         " and children.\n", p->pid, p->comm, points);
 337         /* Try to kill a child first */
 338         list_for_each(tsk, &p->children) {
 339                 c = list_entry(tsk, struct task_struct, sibling);
 340                 if (c->mm == p->mm)
 341                         continue;
 342                 if (!oom_kill_task(c, message))
 343                         return 0;
 344         }
 345         return oom_kill_task(p, message);
 346 }
 347
 348 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
 349
 350 int register_oom_notifier(struct notifier_block *nb)
 351 {
 352         return blocking_notifier_chain_register(&oom_notify_list, nb);
 353 }
 354 EXPORT_SYMBOL_GPL(register_oom_notifier);
 355
 356 int unregister_oom_notifier(struct notifier_block *nb)
 357 {
 358         return blocking_notifier_chain_unregister(&oom_notify_list, nb);
 359 }
 360 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
 361
 362 /**
 363  * out_of_memory - kill the "best" process when we run out of memory
 364  *
 365  * If we run out of memory, we have the choice between either
 366  * killing a random task (bad), letting the system crash (worse)
 367  * OR try to be smart about which process to kill. Note that we
 368  * don't have to be perfect here, we just have to be good.
 369  */
 370 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
 371 {
 372         struct task_struct *p;
 373         unsigned long points = 0;
 374         unsigned long freed = 0;
 375
 376         blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
 377         if (freed > 0)
 378                 /* Got some memory back in the last second. */
 379                 return;
 380
 381         if (printk_ratelimit()) {
 382                 printk(KERN_WARNING "%s invoked oom-killer: "
 383                         "gfp_mask=0x%x, order=%d, oomkilladj=%d\n",
 384                         current->comm, gfp_mask, order, current->oomkilladj);
 385                 dump_stack();
 386                 show_mem();
 387         }
 388
 389         cpuset_lock();
 390         read_lock(&tasklist_lock);
 391
 392         /*
 393          * Check if there were limitations on the allocation (only relevant for
 394          * NUMA) that may require different handling.
 395          */
 396         switch (constrained_alloc(zonelist, gfp_mask)) {
 397         case CONSTRAINT_MEMORY_POLICY:
 398                 oom_kill_process(current, points,
 399                                 "No available memory (MPOL_BIND)");
 400                 break;
 401
 402         case CONSTRAINT_CPUSET:
 403                 oom_kill_process(current, points,
 404                                 "No available memory in cpuset");
 405                 break;
 406
 407         case CONSTRAINT_NONE:
 408                 if (sysctl_panic_on_oom)
 409                         panic("out of memory. panic_on_oom is selected\n");
 410 retry:
 411                 /*
 412                  * Rambo mode: Shoot down a process and hope it solves whatever
 413                  * issues we may have.
 414                  */
 415                 p = select_bad_process(&points);
 416
 417                 if (PTR_ERR(p) == -1UL)
 418                         goto out;
 419
 420                 /* Found nothing?!?! Either we hang forever, or we panic. */
 421                 if (!p) {
 422                         read_unlock(&tasklist_lock);
 423                         cpuset_unlock();
 424                         panic("Out of memory and no killable processes...\n");
 425                 }
 426
 427                 if (oom_kill_process(p, points, "Out of memory"))
 428                         goto retry;
 429
 430                 break;
 431         }
 432
 433 out:
 434         read_unlock(&tasklist_lock);
 435         cpuset_unlock();
 436
 437         /*
 438          * Give "p" a good chance of killing itself before we
 439          * retry to allocate memory unless "p" is current
 440          */
 441         if (!test_thread_flag(TIF_MEMDIE))
 442                 schedule_timeout_uninterruptible(1);
 443 }