oom_kill.c revision bbe373f2c60b2aa36c3231734a5afc5271a06718
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/oom.h> 19#include <linux/mm.h> 20#include <linux/err.h> 21#include <linux/sched.h> 22#include <linux/swap.h> 23#include <linux/timex.h> 24#include <linux/jiffies.h> 25#include <linux/cpuset.h> 26#include <linux/module.h> 27#include <linux/notifier.h> 28 29int sysctl_panic_on_oom; 30int sysctl_oom_kill_allocating_task; 31static DEFINE_MUTEX(zone_scan_mutex); 32/* #define DEBUG */ 33 34/** 35 * badness - calculate a numeric value for how bad this task has been 36 * @p: task struct of which task we should calculate 37 * @uptime: current uptime in seconds 38 * 39 * The formula used is relatively simple and documented inline in the 40 * function. The main rationale is that we want to select a good task 41 * to kill when we run out of memory. 42 * 43 * Good in this context means that: 44 * 1) we lose the minimum amount of work done 45 * 2) we recover a large amount of memory 46 * 3) we don't kill anything innocent of eating tons of memory 47 * 4) we want to kill the minimum amount of processes (one) 48 * 5) we try to kill the process the user expects us to kill, this 49 * algorithm has been meticulously tuned to meet the principle 50 * of least surprise ... (be careful when you change it) 51 */ 52 53unsigned long badness(struct task_struct *p, unsigned long uptime) 54{ 55 unsigned long points, cpu_time, run_time, s; 56 struct mm_struct *mm; 57 struct task_struct *child; 58 59 task_lock(p); 60 mm = p->mm; 61 if (!mm) { 62 task_unlock(p); 63 return 0; 64 } 65 66 /* 67 * The memory size of the process is the basis for the badness. 68 */ 69 points = mm->total_vm; 70 71 /* 72 * After this unlock we can no longer dereference local variable `mm' 73 */ 74 task_unlock(p); 75 76 /* 77 * swapoff can easily use up all memory, so kill those first. 78 */ 79 if (p->flags & PF_SWAPOFF) 80 return ULONG_MAX; 81 82 /* 83 * Processes which fork a lot of child processes are likely 84 * a good choice. We add half the vmsize of the children if they 85 * have an own mm. This prevents forking servers to flood the 86 * machine with an endless amount of children. In case a single 87 * child is eating the vast majority of memory, adding only half 88 * to the parents will make the child our kill candidate of choice. 89 */ 90 list_for_each_entry(child, &p->children, sibling) { 91 task_lock(child); 92 if (child->mm != mm && child->mm) 93 points += child->mm->total_vm/2 + 1; 94 task_unlock(child); 95 } 96 97 /* 98 * CPU time is in tens of seconds and run time is in thousands 99 * of seconds. There is no particular reason for this other than 100 * that it turned out to work very well in practice. 101 */ 102 cpu_time = (cputime_to_jiffies(p->utime) + cputime_to_jiffies(p->stime)) 103 >> (SHIFT_HZ + 3); 104 105 if (uptime >= p->start_time.tv_sec) 106 run_time = (uptime - p->start_time.tv_sec) >> 10; 107 else 108 run_time = 0; 109 110 s = int_sqrt(cpu_time); 111 if (s) 112 points /= s; 113 s = int_sqrt(int_sqrt(run_time)); 114 if (s) 115 points /= s; 116 117 /* 118 * Niced processes are most likely less important, so double 119 * their badness points. 120 */ 121 if (task_nice(p) > 0) 122 points *= 2; 123 124 /* 125 * Superuser processes are usually more important, so we make it 126 * less likely that we kill those. 127 */ 128 if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_ADMIN) || 129 p->uid == 0 || p->euid == 0) 130 points /= 4; 131 132 /* 133 * We don't want to kill a process with direct hardware access. 134 * Not only could that mess up the hardware, but usually users 135 * tend to only have this flag set on applications they think 136 * of as important. 137 */ 138 if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_RAWIO)) 139 points /= 4; 140 141 /* 142 * If p's nodes don't overlap ours, it may still help to kill p 143 * because p may have allocated or otherwise mapped memory on 144 * this node before. However it will be less likely. 145 */ 146 if (!cpuset_mems_allowed_intersects(current, p)) 147 points /= 8; 148 149 /* 150 * Adjust the score by oomkilladj. 151 */ 152 if (p->oomkilladj) { 153 if (p->oomkilladj > 0) { 154 if (!points) 155 points = 1; 156 points <<= p->oomkilladj; 157 } else 158 points >>= -(p->oomkilladj); 159 } 160 161#ifdef DEBUG 162 printk(KERN_DEBUG "OOMkill: task %d (%s) got %lu points\n", 163 p->pid, p->comm, points); 164#endif 165 return points; 166} 167 168/* 169 * Determine the type of allocation constraint. 170 */ 171static inline enum oom_constraint constrained_alloc(struct zonelist *zonelist, 172 gfp_t gfp_mask) 173{ 174#ifdef CONFIG_NUMA 175 struct zone **z; 176 nodemask_t nodes = node_states[N_HIGH_MEMORY]; 177 178 for (z = zonelist->zones; *z; z++) 179 if (cpuset_zone_allowed_softwall(*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 */ 197static 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 task already has access to memory reserves and is 220 * being killed. Don't allow any other task access to the 221 * memory reserve. 222 * 223 * Note: this may have a chance of deadlock if it gets 224 * blocked waiting for another task which itself is waiting 225 * for memory. Is there a better alternative? 226 */ 227 if (test_tsk_thread_flag(p, TIF_MEMDIE)) 228 return ERR_PTR(-1UL); 229 230 /* 231 * This is in the process of releasing memory so wait for it 232 * to finish before killing some other task by mistake. 233 * 234 * However, if p is the current task, we allow the 'kill' to 235 * go ahead if it is exiting: this will simply set TIF_MEMDIE, 236 * which will allow it to gain access to memory reserves in 237 * the process of exiting and releasing its resources. 238 * Otherwise we could get an easy OOM deadlock. 239 */ 240 if (p->flags & PF_EXITING) { 241 if (p != current) 242 return ERR_PTR(-1UL); 243 244 chosen = p; 245 *ppoints = ULONG_MAX; 246 } 247 248 if (p->oomkilladj == OOM_DISABLE) 249 continue; 250 251 points = badness(p, uptime.tv_sec); 252 if (points > *ppoints || !chosen) { 253 chosen = p; 254 *ppoints = points; 255 } 256 } while_each_thread(g, p); 257 258 return chosen; 259} 260 261/** 262 * Send SIGKILL to the selected process irrespective of CAP_SYS_RAW_IO 263 * flag though it's unlikely that we select a process with CAP_SYS_RAW_IO 264 * set. 265 */ 266static void __oom_kill_task(struct task_struct *p, int verbose) 267{ 268 if (is_init(p)) { 269 WARN_ON(1); 270 printk(KERN_WARNING "tried to kill init!\n"); 271 return; 272 } 273 274 if (!p->mm) { 275 WARN_ON(1); 276 printk(KERN_WARNING "tried to kill an mm-less task!\n"); 277 return; 278 } 279 280 if (verbose) 281 printk(KERN_ERR "Killed process %d (%s)\n", p->pid, p->comm); 282 283 /* 284 * We give our sacrificial lamb high priority and access to 285 * all the memory it needs. That way it should be able to 286 * exit() and clear out its resources quickly... 287 */ 288 p->time_slice = HZ; 289 set_tsk_thread_flag(p, TIF_MEMDIE); 290 291 force_sig(SIGKILL, p); 292} 293 294static int oom_kill_task(struct task_struct *p) 295{ 296 struct mm_struct *mm; 297 struct task_struct *g, *q; 298 299 mm = p->mm; 300 301 /* WARNING: mm may not be dereferenced since we did not obtain its 302 * value from get_task_mm(p). This is OK since all we need to do is 303 * compare mm to q->mm below. 304 * 305 * Furthermore, even if mm contains a non-NULL value, p->mm may 306 * change to NULL at any time since we do not hold task_lock(p). 307 * However, this is of no concern to us. 308 */ 309 310 if (mm == NULL) 311 return 1; 312 313 /* 314 * Don't kill the process if any threads are set to OOM_DISABLE 315 */ 316 do_each_thread(g, q) { 317 if (q->mm == mm && q->oomkilladj == OOM_DISABLE) 318 return 1; 319 } while_each_thread(g, q); 320 321 __oom_kill_task(p, 1); 322 323 /* 324 * kill all processes that share the ->mm (i.e. all threads), 325 * but are in a different thread group. Don't let them have access 326 * to memory reserves though, otherwise we might deplete all memory. 327 */ 328 do_each_thread(g, q) { 329 if (q->mm == mm && q->tgid != p->tgid) 330 force_sig(SIGKILL, q); 331 } while_each_thread(g, q); 332 333 return 0; 334} 335 336static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order, 337 unsigned long points, const char *message) 338{ 339 struct task_struct *c; 340 struct list_head *tsk; 341 342 if (printk_ratelimit()) { 343 printk(KERN_WARNING "%s invoked oom-killer: " 344 "gfp_mask=0x%x, order=%d, oomkilladj=%d\n", 345 current->comm, gfp_mask, order, current->oomkilladj); 346 dump_stack(); 347 show_mem(); 348 } 349 350 /* 351 * If the task is already exiting, don't alarm the sysadmin or kill 352 * its children or threads, just set TIF_MEMDIE so it can die quickly 353 */ 354 if (p->flags & PF_EXITING) { 355 __oom_kill_task(p, 0); 356 return 0; 357 } 358 359 printk(KERN_ERR "%s: kill process %d (%s) score %li or a child\n", 360 message, p->pid, p->comm, points); 361 362 /* Try to kill a child first */ 363 list_for_each(tsk, &p->children) { 364 c = list_entry(tsk, struct task_struct, sibling); 365 if (c->mm == p->mm) 366 continue; 367 if (!oom_kill_task(c)) 368 return 0; 369 } 370 return oom_kill_task(p); 371} 372 373static BLOCKING_NOTIFIER_HEAD(oom_notify_list); 374 375int register_oom_notifier(struct notifier_block *nb) 376{ 377 return blocking_notifier_chain_register(&oom_notify_list, nb); 378} 379EXPORT_SYMBOL_GPL(register_oom_notifier); 380 381int unregister_oom_notifier(struct notifier_block *nb) 382{ 383 return blocking_notifier_chain_unregister(&oom_notify_list, nb); 384} 385EXPORT_SYMBOL_GPL(unregister_oom_notifier); 386 387/* 388 * Try to acquire the OOM killer lock for the zones in zonelist. Returns zero 389 * if a parallel OOM killing is already taking place that includes a zone in 390 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1. 391 */ 392int try_set_zone_oom(struct zonelist *zonelist) 393{ 394 struct zone **z; 395 int ret = 1; 396 397 z = zonelist->zones; 398 399 mutex_lock(&zone_scan_mutex); 400 do { 401 if (zone_is_oom_locked(*z)) { 402 ret = 0; 403 goto out; 404 } 405 } while (*(++z) != NULL); 406 407 /* 408 * Lock each zone in the zonelist under zone_scan_mutex so a parallel 409 * invocation of try_set_zone_oom() doesn't succeed when it shouldn't. 410 */ 411 z = zonelist->zones; 412 do { 413 zone_set_flag(*z, ZONE_OOM_LOCKED); 414 } while (*(++z) != NULL); 415out: 416 mutex_unlock(&zone_scan_mutex); 417 return ret; 418} 419 420/* 421 * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed 422 * allocation attempts with zonelists containing them may now recall the OOM 423 * killer, if necessary. 424 */ 425void clear_zonelist_oom(struct zonelist *zonelist) 426{ 427 struct zone **z; 428 429 z = zonelist->zones; 430 431 mutex_lock(&zone_scan_mutex); 432 do { 433 zone_clear_flag(*z, ZONE_OOM_LOCKED); 434 } while (*(++z) != NULL); 435 mutex_unlock(&zone_scan_mutex); 436} 437 438/** 439 * out_of_memory - kill the "best" process when we run out of memory 440 * 441 * If we run out of memory, we have the choice between either 442 * killing a random task (bad), letting the system crash (worse) 443 * OR try to be smart about which process to kill. Note that we 444 * don't have to be perfect here, we just have to be good. 445 */ 446void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order) 447{ 448 struct task_struct *p; 449 unsigned long points = 0; 450 unsigned long freed = 0; 451 enum oom_constraint constraint; 452 453 blocking_notifier_call_chain(&oom_notify_list, 0, &freed); 454 if (freed > 0) 455 /* Got some memory back in the last second. */ 456 return; 457 458 if (sysctl_panic_on_oom == 2) 459 panic("out of memory. Compulsory panic_on_oom is selected.\n"); 460 461 /* 462 * Check if there were limitations on the allocation (only relevant for 463 * NUMA) that may require different handling. 464 */ 465 constraint = constrained_alloc(zonelist, gfp_mask); 466 cpuset_lock(); 467 read_lock(&tasklist_lock); 468 469 switch (constraint) { 470 case CONSTRAINT_MEMORY_POLICY: 471 oom_kill_process(current, gfp_mask, order, points, 472 "No available memory (MPOL_BIND)"); 473 break; 474 475 case CONSTRAINT_NONE: 476 if (sysctl_panic_on_oom) 477 panic("out of memory. panic_on_oom is selected\n"); 478 /* Fall-through */ 479 case CONSTRAINT_CPUSET: 480 if (sysctl_oom_kill_allocating_task) { 481 oom_kill_process(current, gfp_mask, order, points, 482 "Out of memory (oom_kill_allocating_task)"); 483 break; 484 } 485retry: 486 /* 487 * Rambo mode: Shoot down a process and hope it solves whatever 488 * issues we may have. 489 */ 490 p = select_bad_process(&points); 491 492 if (PTR_ERR(p) == -1UL) 493 goto out; 494 495 /* Found nothing?!?! Either we hang forever, or we panic. */ 496 if (!p) { 497 read_unlock(&tasklist_lock); 498 cpuset_unlock(); 499 panic("Out of memory and no killable processes...\n"); 500 } 501 502 if (oom_kill_process(p, points, gfp_mask, order, 503 "Out of memory")) 504 goto retry; 505 506 break; 507 } 508 509out: 510 read_unlock(&tasklist_lock); 511 cpuset_unlock(); 512 513 /* 514 * Give "p" a good chance of killing itself before we 515 * retry to allocate memory unless "p" is current 516 */ 517 if (!test_thread_flag(TIF_MEMDIE)) 518 schedule_timeout_uninterruptible(1); 519} 520