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