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 mm_struct *mm;
  50         struct task_struct *child;
  51
  52         task_lock(p);
  53         mm = p->mm;
  54         if (!mm) {
  55                 task_unlock(p);
  56                 return 0;
  57         }
  58
  59         /*
  60          * The memory size of the process is the basis for the badness.
  61          */
  62         points = mm->total_vm;
  63
  64         /*
  65          * After this unlock we can no longer dereference local variable `mm'
  66          */
  67         task_unlock(p);
  68
  69         /*
  70          * Processes which fork a lot of child processes are likely
  71          * a good choice. We add half the vmsize of the children if they
  72          * have an own mm. This prevents forking servers to flood the
  73          * machine with an endless amount of children. In case a single
  74          * child is eating the vast majority of memory, adding only half
  75          * to the parents will make the child our kill candidate of choice.
  76          */
  77         list_for_each_entry(child, &p->children, sibling) {
  78                 task_lock(child);
  79                 if (child->mm != mm && child->mm)
  80                         points += child->mm->total_vm/2 + 1;
  81                 task_unlock(child);
  82         }
  83
  84         /*
  85          * CPU time is in tens of seconds and run time is in thousands
  86          * of seconds. There is no particular reason for this other than
  87          * that it turned out to work very well in practice.
  88          */
  89         cpu_time = (cputime_to_jiffies(p->utime) + cputime_to_jiffies(p->stime))
  90                 >> (SHIFT_HZ + 3);
  91
  92         if (uptime >= p->start_time.tv_sec)
  93                 run_time = (uptime - p->start_time.tv_sec) >> 10;
  94         else
  95                 run_time = 0;
  96
  97         s = int_sqrt(cpu_time);
  98         if (s)
  99                 points /= s;
 100         s = int_sqrt(int_sqrt(run_time));
 101         if (s)
 102                 points /= s;
 103
 104         /*
 105          * Niced processes are most likely less important, so double
 106          * their badness points.
 107          */
 108         if (task_nice(p) > 0)
 109                 points *= 2;
 110
 111         /*
 112          * Superuser processes are usually more important, so we make it
 113          * less likely that we kill those.
 114          */
 115         if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_ADMIN) ||
 116                                 p->uid == 0 || p->euid == 0)
 117                 points /= 4;
 118
 119         /*
 120          * We don't want to kill a process with direct hardware access.
 121          * Not only could that mess up the hardware, but usually users
 122          * tend to only have this flag set on applications they think
 123          * of as important.
 124          */
 125         if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_RAWIO))
 126                 points /= 4;
 127
 128         /*
 129          * Adjust the score by oomkilladj.
 130          */
 131         if (p->oomkilladj) {
 132                 if (p->oomkilladj > 0)
 133                         points <<= p->oomkilladj;
 134                 else
 135                         points >>= -(p->oomkilladj);
 136         }
 137
 138 #ifdef DEBUG
 139         printk(KERN_DEBUG "OOMkill: task %d (%s) got %d points\n",
 140         p->pid, p->comm, points);
 141 #endif
 142         return points;
 143 }
 144
 145 /*
 146  * Types of limitations to the nodes from which allocations may occur
 147  */
 148 #define CONSTRAINT_NONE 1
 149 #define CONSTRAINT_MEMORY_POLICY 2
 150 #define CONSTRAINT_CPUSET 3
 151
 152 /*
 153  * Determine the type of allocation constraint.
 154  */
 155 static inline int constrained_alloc(struct zonelist *zonelist, gfp_t gfp_mask)
 156 {
 157 #ifdef CONFIG_NUMA
 158         struct zone **z;
 159         nodemask_t nodes = node_online_map;
 160
 161         for (z = zonelist->zones; *z; z++)
 162                 if (cpuset_zone_allowed(*z, gfp_mask))
 163                         node_clear((*z)->zone_pgdat->node_id,
 164                                         nodes);
 165                 else
 166                         return CONSTRAINT_CPUSET;
 167
 168         if (!nodes_empty(nodes))
 169                 return CONSTRAINT_MEMORY_POLICY;
 170 #endif
 171
 172         return CONSTRAINT_NONE;
 173 }
 174
 175 /*
 176  * Simple selection loop. We chose the process with the highest
 177  * number of 'points'. We expect the caller will lock the tasklist.
 178  *
 179  * (not docbooked, we don't want this one cluttering up the manual)
 180  */
 181 static struct task_struct *select_bad_process(unsigned long *ppoints)
 182 {
 183         struct task_struct *g, *p;
 184         struct task_struct *chosen = NULL;
 185         struct timespec uptime;
 186         *ppoints = 0;
 187
 188         do_posix_clock_monotonic_gettime(&uptime);
 189         do_each_thread(g, p) {
 190                 unsigned long points;
 191                 int releasing;
 192
 193                 /* skip the init task with pid == 1 */
 194                 if (p->pid == 1)
 195                         continue;
 196                 if (p->oomkilladj == OOM_DISABLE)
 197                         continue;
 198                 /* If p's nodes don't overlap ours, it won't help to kill p. */
 199                 if (!cpuset_excl_nodes_overlap(p))
 200                         continue;
 201
 202                 /*
 203                  * This is in the process of releasing memory so for wait it
 204                  * to finish before killing some other task by mistake.
 205                  */
 206                 releasing = test_tsk_thread_flag(p, TIF_MEMDIE) ||
 207                                                 p->flags & PF_EXITING;
 208                 if (releasing && !(p->flags & PF_DEAD))
 209                         return ERR_PTR(-1UL);
 210                 if (p->flags & PF_SWAPOFF)
 211                         return p;
 212
 213                 points = badness(p, uptime.tv_sec);
 214                 if (points > *ppoints || !chosen) {
 215                         chosen = p;
 216                         *ppoints = points;
 217                 }
 218         } while_each_thread(g, p);
 219         return chosen;
 220 }
 221
 222 /**
 223  * We must be careful though to never send SIGKILL a process with
 224  * CAP_SYS_RAW_IO set, send SIGTERM instead (but it's unlikely that
 225  * we select a process with CAP_SYS_RAW_IO set).
 226  */
 227 static void __oom_kill_task(task_t *p, const char *message)
 228 {
 229         if (p->pid == 1) {
 230                 WARN_ON(1);
 231                 printk(KERN_WARNING "tried to kill init!\n");
 232                 return;
 233         }
 234
 235         task_lock(p);
 236         if (!p->mm || p->mm == &init_mm) {
 237                 WARN_ON(1);
 238                 printk(KERN_WARNING "tried to kill an mm-less task!\n");
 239                 task_unlock(p);
 240                 return;
 241         }
 242         task_unlock(p);
 243         printk(KERN_ERR "%s: Killed process %d (%s).\n",
 244                                 message, p->pid, p->comm);
 245
 246         /*
 247          * We give our sacrificial lamb high priority and access to
 248          * all the memory it needs. That way it should be able to
 249          * exit() and clear out its resources quickly...
 250          */
 251         p->time_slice = HZ;
 252         set_tsk_thread_flag(p, TIF_MEMDIE);
 253
 254         force_sig(SIGKILL, p);
 255 }
 256
 257 static int oom_kill_task(task_t *p, const char *message)
 258 {
 259         struct mm_struct *mm;
 260         task_t * g, * q;
 261
 262         mm = p->mm;
 263
 264         /* WARNING: mm may not be dereferenced since we did not obtain its
 265          * value from get_task_mm(p).  This is OK since all we need to do is
 266          * compare mm to q->mm below.
 267          *
 268          * Furthermore, even if mm contains a non-NULL value, p->mm may
 269          * change to NULL at any time since we do not hold task_lock(p).
 270          * However, this is of no concern to us.
 271          */
 272
 273         if (mm == NULL || mm == &init_mm)
 274                 return 1;
 275
 276         __oom_kill_task(p, message);
 277         /*
 278          * kill all processes that share the ->mm (i.e. all threads),
 279          * but are in a different thread group
 280          */
 281         do_each_thread(g, q)
 282                 if (q->mm == mm && q->tgid != p->tgid)
 283                         __oom_kill_task(q, message);
 284         while_each_thread(g, q);
 285
 286         return 0;
 287 }
 288
 289 static int oom_kill_process(struct task_struct *p, unsigned long points,
 290                 const char *message)
 291 {
 292         struct task_struct *c;
 293         struct list_head *tsk;
 294
 295         printk(KERN_ERR "Out of Memory: Kill process %d (%s) score %li and "
 296                 "children.\n", p->pid, p->comm, points);
 297         /* Try to kill a child first */
 298         list_for_each(tsk, &p->children) {
 299                 c = list_entry(tsk, struct task_struct, sibling);
 300                 if (c->mm == p->mm)
 301                         continue;
 302                 if (!oom_kill_task(c, message))
 303                         return 0;
 304         }
 305         return oom_kill_task(p, message);
 306 }
 307
 308 /**
 309  * oom_kill - kill the "best" process when we run out of memory
 310  *
 311  * If we run out of memory, we have the choice between either
 312  * killing a random task (bad), letting the system crash (worse)
 313  * OR try to be smart about which process to kill. Note that we
 314  * don't have to be perfect here, we just have to be good.
 315  */
 316 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
 317 {
 318         task_t *p;
 319         unsigned long points = 0;
 320
 321         if (printk_ratelimit()) {
 322                 printk("oom-killer: gfp_mask=0x%x, order=%d\n",
 323                         gfp_mask, order);
 324                 dump_stack();
 325                 show_mem();
 326         }
 327
 328         cpuset_lock();
 329         read_lock(&tasklist_lock);
 330
 331         /*
 332          * Check if there were limitations on the allocation (only relevant for
 333          * NUMA) that may require different handling.
 334          */
 335         switch (constrained_alloc(zonelist, gfp_mask)) {
 336         case CONSTRAINT_MEMORY_POLICY:
 337                 oom_kill_process(current, points,
 338                                 "No available memory (MPOL_BIND)");
 339                 break;
 340
 341         case CONSTRAINT_CPUSET:
 342                 oom_kill_process(current, points,
 343                                 "No available memory in cpuset");
 344                 break;
 345
 346         case CONSTRAINT_NONE:
 347 retry:
 348                 /*
 349                  * Rambo mode: Shoot down a process and hope it solves whatever
 350                  * issues we may have.
 351                  */
 352                 p = select_bad_process(&points);
 353
 354                 if (PTR_ERR(p) == -1UL)
 355                         goto out;
 356
 357                 /* Found nothing?!?! Either we hang forever, or we panic. */
 358                 if (!p) {
 359                         read_unlock(&tasklist_lock);
 360                         cpuset_unlock();
 361                         panic("Out of memory and no killable processes...\n");
 362                 }
 363
 364                 if (oom_kill_process(p, points, "Out of memory"))
 365                         goto retry;
 366
 367                 break;
 368         }
 369
 370 out:
 371         read_unlock(&tasklist_lock);
 372         cpuset_unlock();
 373
 374         /*
 375          * Give "p" a good chance of killing itself before we
 376          * retry to allocate memory unless "p" is current
 377          */
 378         if (!test_thread_flag(TIF_MEMDIE))
 379                 schedule_timeout_uninterruptible(1);
 380 }