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
  25 /* #define DEBUG */
  26
  27 /**
  28  * oom_badness - calculate a numeric value for how bad this task has been
  29  * @p: task struct of which task we should calculate
  30  * @uptime: current uptime in seconds
  31  *
  32  * The formula used is relatively simple and documented inline in the
  33  * function. The main rationale is that we want to select a good task
  34  * to kill when we run out of memory.
  35  *
  36  * Good in this context means that:
  37  * 1) we lose the minimum amount of work done
  38  * 2) we recover a large amount of memory
  39  * 3) we don't kill anything innocent of eating tons of memory
  40  * 4) we want to kill the minimum amount of processes (one)
  41  * 5) we try to kill the process the user expects us to kill, this
  42  *    algorithm has been meticulously tuned to meet the principle
  43  *    of least surprise ... (be careful when you change it)
  44  */
  45
  46 unsigned long badness(struct task_struct *p, unsigned long uptime)
  47 {
  48         unsigned long points, cpu_time, run_time, s;
  49         struct list_head *tsk;
  50
  51         if (!p->mm)
  52                 return 0;
  53
  54         /*
  55          * The memory size of the process is the basis for the badness.
  56          */
  57         points = p->mm->total_vm;
  58
  59         /*
  60          * Processes which fork a lot of child processes are likely
  61          * a good choice. We add half the vmsize of the children if they
  62          * have an own mm. This prevents forking servers to flood the
  63          * machine with an endless amount of children. In case a single
  64          * child is eating the vast majority of memory, adding only half
  65          * to the parents will make the child our kill candidate of choice.
  66          */
  67         list_for_each(tsk, &p->children) {
  68                 struct task_struct *chld;
  69                 chld = list_entry(tsk, struct task_struct, sibling);
  70                 if (chld->mm != p->mm && chld->mm)
  71                         points += chld->mm->total_vm/2 + 1;
  72         }
  73
  74         /*
  75          * CPU time is in tens of seconds and run time is in thousands
  76          * of seconds. There is no particular reason for this other than
  77          * that it turned out to work very well in practice.
  78          */
  79         cpu_time = (cputime_to_jiffies(p->utime) + cputime_to_jiffies(p->stime))
  80                 >> (SHIFT_HZ + 3);
  81
  82         if (uptime >= p->start_time.tv_sec)
  83                 run_time = (uptime - p->start_time.tv_sec) >> 10;
  84         else
  85                 run_time = 0;
  86
  87         s = int_sqrt(cpu_time);
  88         if (s)
  89                 points /= s;
  90         s = int_sqrt(int_sqrt(run_time));
  91         if (s)
  92                 points /= s;
  93
  94         /*
  95          * Niced processes are most likely less important, so double
  96          * their badness points.
  97          */
  98         if (task_nice(p) > 0)
  99                 points *= 2;
 100
 101         /*
 102          * Superuser processes are usually more important, so we make it
 103          * less likely that we kill those.
 104          */
 105         if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_ADMIN) ||
 106                                 p->uid == 0 || p->euid == 0)
 107                 points /= 4;
 108
 109         /*
 110          * We don't want to kill a process with direct hardware access.
 111          * Not only could that mess up the hardware, but usually users
 112          * tend to only have this flag set on applications they think
 113          * of as important.
 114          */
 115         if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_RAWIO))
 116                 points /= 4;
 117
 118         /*
 119          * Adjust the score by oomkilladj.
 120          */
 121         if (p->oomkilladj) {
 122                 if (p->oomkilladj > 0)
 123                         points <<= p->oomkilladj;
 124                 else
 125                         points >>= -(p->oomkilladj);
 126         }
 127
 128 #ifdef DEBUG
 129         printk(KERN_DEBUG "OOMkill: task %d (%s) got %d points\n",
 130         p->pid, p->comm, points);
 131 #endif
 132         return points;
 133 }
 134
 135 /*
 136  * Simple selection loop. We chose the process with the highest
 137  * number of 'points'. We expect the caller will lock the tasklist.
 138  *
 139  * (not docbooked, we don't want this one cluttering up the manual)
 140  */
 141 static struct task_struct *select_bad_process(unsigned long *ppoints)
 142 {
 143         struct task_struct *g, *p;
 144         struct task_struct *chosen = NULL;
 145         struct timespec uptime;
 146         *ppoints = 0;
 147
 148         do_posix_clock_monotonic_gettime(&uptime);
 149         do_each_thread(g, p) {
 150                 unsigned long points;
 151                 int releasing;
 152
 153                 /* skip the init task with pid == 1 */
 154                 if (p->pid == 1)
 155                         continue;
 156                 if (p->oomkilladj == OOM_DISABLE)
 157                         continue;
 158                 /* If p's nodes don't overlap ours, it won't help to kill p. */
 159                 if (!cpuset_excl_nodes_overlap(p))
 160                         continue;
 161
 162                 /*
 163                  * This is in the process of releasing memory so for wait it
 164                  * to finish before killing some other task by mistake.
 165                  */
 166                 releasing = test_tsk_thread_flag(p, TIF_MEMDIE) ||
 167                                                 p->flags & PF_EXITING;
 168                 if (releasing && !(p->flags & PF_DEAD))
 169                         return ERR_PTR(-1UL);
 170                 if (p->flags & PF_SWAPOFF)
 171                         return p;
 172
 173                 points = badness(p, uptime.tv_sec);
 174                 if (points > *ppoints || !chosen) {
 175                         chosen = p;
 176                         *ppoints = points;
 177                 }
 178         } while_each_thread(g, p);
 179         return chosen;
 180 }
 181
 182 /**
 183  * We must be careful though to never send SIGKILL a process with
 184  * CAP_SYS_RAW_IO set, send SIGTERM instead (but it's unlikely that
 185  * we select a process with CAP_SYS_RAW_IO set).
 186  */
 187 static void __oom_kill_task(task_t *p)
 188 {
 189         if (p->pid == 1) {
 190                 WARN_ON(1);
 191                 printk(KERN_WARNING "tried to kill init!\n");
 192                 return;
 193         }
 194
 195         task_lock(p);
 196         if (!p->mm || p->mm == &init_mm) {
 197                 WARN_ON(1);
 198                 printk(KERN_WARNING "tried to kill an mm-less task!\n");
 199                 task_unlock(p);
 200                 return;
 201         }
 202         task_unlock(p);
 203         printk(KERN_ERR "Out of Memory: Killed process %d (%s).\n",
 204                                                         p->pid, p->comm);
 205
 206         /*
 207          * We give our sacrificial lamb high priority and access to
 208          * all the memory it needs. That way it should be able to
 209          * exit() and clear out its resources quickly...
 210          */
 211         p->time_slice = HZ;
 212         set_tsk_thread_flag(p, TIF_MEMDIE);
 213
 214         force_sig(SIGKILL, p);
 215 }
 216
 217 static struct mm_struct *oom_kill_task(task_t *p)
 218 {
 219         struct mm_struct *mm = get_task_mm(p);
 220         task_t * g, * q;
 221
 222         if (!mm)
 223                 return NULL;
 224         if (mm == &init_mm) {
 225                 mmput(mm);
 226                 return NULL;
 227         }
 228
 229         __oom_kill_task(p);
 230         /*
 231          * kill all processes that share the ->mm (i.e. all threads),
 232          * but are in a different thread group
 233          */
 234         do_each_thread(g, q)
 235                 if (q->mm == mm && q->tgid != p->tgid)
 236                         __oom_kill_task(q);
 237         while_each_thread(g, q);
 238
 239         return mm;
 240 }
 241
 242 static struct mm_struct *oom_kill_process(struct task_struct *p,
 243                                           unsigned long points)
 244 {
 245         struct mm_struct *mm;
 246         struct task_struct *c;
 247         struct list_head *tsk;
 248
 249         printk(KERN_ERR "Out of Memory: Kill process %d (%s) score %li and "
 250                 "children.\n", p->pid, p->comm, points);
 251         /* Try to kill a child first */
 252         list_for_each(tsk, &p->children) {
 253                 c = list_entry(tsk, struct task_struct, sibling);
 254                 if (c->mm == p->mm)
 255                         continue;
 256                 mm = oom_kill_task(c);
 257                 if (mm)
 258                         return mm;
 259         }
 260         return oom_kill_task(p);
 261 }
 262
 263 /**
 264  * oom_kill - kill the "best" process when we run out of memory
 265  *
 266  * If we run out of memory, we have the choice between either
 267  * killing a random task (bad), letting the system crash (worse)
 268  * OR try to be smart about which process to kill. Note that we
 269  * don't have to be perfect here, we just have to be good.
 270  */
 271 void out_of_memory(gfp_t gfp_mask, int order)
 272 {
 273         struct mm_struct *mm = NULL;
 274         task_t * p;
 275         unsigned long points;
 276
 277         if (printk_ratelimit()) {
 278                 printk("oom-killer: gfp_mask=0x%x, order=%d\n",
 279                         gfp_mask, order);
 280                 dump_stack();
 281                 show_mem();
 282         }
 283
 284         cpuset_lock();
 285         read_lock(&tasklist_lock);
 286 retry:
 287         p = select_bad_process(&points);
 288
 289         if (PTR_ERR(p) == -1UL)
 290                 goto out;
 291
 292         /* Found nothing?!?! Either we hang forever, or we panic. */
 293         if (!p) {
 294                 read_unlock(&tasklist_lock);
 295                 cpuset_unlock();
 296                 panic("Out of memory and no killable processes...\n");
 297         }
 298
 299         mm = oom_kill_process(p, points);
 300         if (!mm)
 301                 goto retry;
 302
 303  out:
 304         read_unlock(&tasklist_lock);
 305         cpuset_unlock();
 306         if (mm)
 307                 mmput(mm);
 308
 309         /*
 310          * Give "p" a good chance of killing itself before we
 311          * retry to allocate memory unless "p" is current
 312          */
 313         if (!test_thread_flag(TIF_MEMDIE))
 314                 schedule_timeout_interruptible(1);
 315 }