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 kswapd()
  10  *  in linux/mm/vmscan.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
  24 /* #define DEBUG */
  25
  26 /**
  27  * int_sqrt - oom_kill.c internal function, rough approximation to sqrt
  28  * @x: integer of which to calculate the sqrt
  29  *
  30  * A very rough approximation to the sqrt() function.
  31  */
  32 static unsigned int int_sqrt(unsigned int x)
  33 {
  34         unsigned int out = x;
  35         while (x & ~(unsigned int)1) x >>=2, out >>=1;
  36         if (x) out -= out >> 2;
  37         return (out ? out : 1);
  38 }
  39
  40 /**
  41  * oom_badness - calculate a numeric value for how bad this task has been
  42  * @p: task struct of which task we should calculate
  43  *
  44  * The formula used is relatively simple and documented inline in the
  45  * function. The main rationale is that we want to select a good task
  46  * to kill when we run out of memory.
  47  *
  48  * Good in this context means that:
  49  * 1) we lose the minimum amount of work done
  50  * 2) we recover a large amount of memory
  51  * 3) we don't kill anything innocent of eating tons of memory
  52  * 4) we want to kill the minimum amount of processes (one)
  53  * 5) we try to kill the process the user expects us to kill, this
  54  *    algorithm has been meticulously tuned to meet the principle
  55  *    of least surprise ... (be careful when you change it)
  56  */
  57
  58 static int badness(struct task_struct *p)
  59 {
  60         int points, cpu_time, run_time;
  61
  62         if (!p->mm)
  63                 return 0;
  64
  65         if (p->flags & PF_MEMDIE)
  66                 return 0;
  67         /*
  68          * The memory size of the process is the basis for the badness.
  69          */
  70         points = p->mm->total_vm;
  71
  72         /*
  73          * CPU time is in seconds and run time is in minutes. There is no
  74          * particular reason for this other than that it turned out to work
  75          * very well in practice.
  76          */
  77         cpu_time = (p->utime + p->stime) >> (SHIFT_HZ + 3);
  78         run_time = (get_jiffies_64() - p->start_time) >> (SHIFT_HZ + 10);
  79
  80         points /= int_sqrt(cpu_time);
  81         points /= int_sqrt(int_sqrt(run_time));
  82
  83         /*
  84          * Niced processes are most likely less important, so double
  85          * their badness points.
  86          */
  87         if (task_nice(p) > 0)
  88                 points *= 2;
  89
  90         /*
  91          * Superuser processes are usually more important, so we make it
  92          * less likely that we kill those.
  93          */
  94         if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_ADMIN) ||
  95                                 p->uid == 0 || p->euid == 0)
  96                 points /= 4;
  97
  98         /*
  99          * We don't want to kill a process with direct hardware access.
 100          * Not only could that mess up the hardware, but usually users
 101          * tend to only have this flag set on applications they think
 102          * of as important.
 103          */
 104         if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_RAWIO))
 105                 points /= 4;
 106 #ifdef DEBUG
 107         printk(KERN_DEBUG "OOMkill: task %d (%s) got %d points\n",
 108         p->pid, p->comm, points);
 109 #endif
 110         return points;
 111 }
 112
 113 /*
 114  * Simple selection loop. We chose the process with the highest
 115  * number of 'points'. We expect the caller will lock the tasklist.
 116  *
 117  * (not docbooked, we don't want this one cluttering up the manual)
 118  */
 119 static struct task_struct * select_bad_process(void)
 120 {
 121         int maxpoints = 0;
 122         struct task_struct *g, *p;
 123         struct task_struct *chosen = NULL;
 124
 125         do_each_thread(g, p)
 126                 if (p->pid) {
 127                         int points = badness(p);
 128                         if (points > maxpoints) {
 129                                 chosen = p;
 130                                 maxpoints = points;
 131                         }
 132                         if (p->flags & PF_SWAPOFF)
 133                                 return p;
 134                 }
 135         while_each_thread(g, p);
 136         return chosen;
 137 }
 138
 139 /**
 140  * We must be careful though to never send SIGKILL a process with
 141  * CAP_SYS_RAW_IO set, send SIGTERM instead (but it's unlikely that
 142  * we select a process with CAP_SYS_RAW_IO set).
 143  */
 144 static void __oom_kill_task(task_t *p)
 145 {
 146         task_lock(p);
 147         if (!p->mm || p->mm == &init_mm) {
 148                 WARN_ON(1);
 149                 printk(KERN_WARNING "tried to kill an mm-less task!\n");
 150                 task_unlock(p);
 151                 return;
 152         }
 153         task_unlock(p);
 154         printk(KERN_ERR "Out of Memory: Killed process %d (%s).\n", p->pid, p->comm);
 155
 156         /*
 157          * We give our sacrificial lamb high priority and access to
 158          * all the memory it needs. That way it should be able to
 159          * exit() and clear out its resources quickly...
 160          */
 161         p->time_slice = HZ;
 162         p->flags |= PF_MEMALLOC | PF_MEMDIE;
 163
 164         /* This process has hardware access, be more careful. */
 165         if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_RAWIO)) {
 166                 force_sig(SIGTERM, p);
 167         } else {
 168                 force_sig(SIGKILL, p);
 169         }
 170 }
 171
 172 static struct mm_struct *oom_kill_task(task_t *p)
 173 {
 174         struct mm_struct *mm = get_task_mm(p);
 175         if (!mm || mm == &init_mm)
 176                 return NULL;
 177         __oom_kill_task(p);
 178         return mm;
 179 }
 180
 181
 182 /**
 183  * oom_kill - kill the "best" process when we run out of memory
 184  *
 185  * If we run out of memory, we have the choice between either
 186  * killing a random task (bad), letting the system crash (worse)
 187  * OR try to be smart about which process to kill. Note that we
 188  * don't have to be perfect here, we just have to be good.
 189  */
 190 static void oom_kill(void)
 191 {
 192         struct mm_struct *mm;
 193         struct task_struct *g, *p, *q;
 194
 195         read_lock(&tasklist_lock);
 196 retry:
 197         p = select_bad_process();
 198
 199         /* Found nothing?!?! Either we hang forever, or we panic. */
 200         if (!p) {
 201                 show_free_areas();
 202                 panic("Out of memory and no killable processes...\n");
 203         }
 204
 205         mm = oom_kill_task(p);
 206         if (!mm)
 207                 goto retry;
 208         /*
 209          * kill all processes that share the ->mm (i.e. all threads),
 210          * but are in a different thread group
 211          */
 212         do_each_thread(g, q)
 213                 if (q->mm == mm && q->tgid != p->tgid)
 214                         __oom_kill_task(q);
 215         while_each_thread(g, q);
 216         if (!p->mm)
 217                 printk(KERN_INFO "Fixed up OOM kill of mm-less task\n");
 218         read_unlock(&tasklist_lock);
 219         mmput(mm);
 220
 221         /*
 222          * Make kswapd go out of the way, so "p" has a good chance of
 223          * killing itself before someone else gets the chance to ask
 224          * for more memory.
 225          */
 226         yield();
 227         return;
 228 }
 229
 230 /**
 231  * out_of_memory - is the system out of memory?
 232  */
 233 void out_of_memory(void)
 234 {
 235         /*
 236          * oom_lock protects out_of_memory()'s static variables.
 237          * It's a global lock; this is not performance-critical.
 238          */
 239         static spinlock_t oom_lock = SPIN_LOCK_UNLOCKED;
 240         static unsigned long first, last, count, lastkill;
 241         unsigned long now, since;
 242
 243         /*
 244          * Enough swap space left?  Not OOM.
 245          */
 246         if (nr_swap_pages > 0)
 247                 return;
 248
 249         spin_lock(&oom_lock);
 250         now = jiffies;
 251         since = now - last;
 252         last = now;
 253
 254         /*
 255          * If it's been a long time since last failure,
 256          * we're not oom.
 257          */
 258         last = now;
 259         if (since > 5*HZ)
 260                 goto reset;
 261
 262         /*
 263          * If we haven't tried for at least one second,
 264          * we're not really oom.
 265          */
 266         since = now - first;
 267         if (since < HZ)
 268                 goto out_unlock;
 269
 270         /*
 271          * If we have gotten only a few failures,
 272          * we're not really oom.
 273          */
 274         if (++count < 10)
 275                 goto out_unlock;
 276
 277         /*
 278          * If we just killed a process, wait a while
 279          * to give that task a chance to exit. This
 280          * avoids killing multiple processes needlessly.
 281          */
 282         since = now - lastkill;
 283         if (since < HZ*5)
 284                 goto out_unlock;
 285
 286         /*
 287          * Ok, really out of memory. Kill something.
 288          */
 289         lastkill = now;
 290
 291         /* oom_kill() sleeps */
 292         spin_unlock(&oom_lock);
 293         oom_kill();
 294         spin_lock(&oom_lock);
 295
 296 reset:
 297         /*
 298          * We dropped the lock above, so check to be sure the variable
 299          * first only ever increases to prevent false OOM's.
 300          */
 301         if (time_after(now, first))
 302                 first = now;
 303         count = 0;
 304
 305 out_unlock:
 306         spin_unlock(&oom_lock);
 307 }