oom_kill.c revision 7887a3da753e1ba8244556cc9a2b38c815bfe256
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 27int 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 49unsigned 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 */ 166static 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 */ 192static 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 releasing = test_tsk_thread_flag(p, TIF_MEMDIE) || 215 p->flags & PF_EXITING; 216 if (releasing && !(p->flags & PF_DEAD)) 217 return ERR_PTR(-1UL); 218 if (p->flags & PF_SWAPOFF) 219 return p; 220 221 points = badness(p, uptime.tv_sec); 222 if (points > *ppoints || !chosen) { 223 chosen = p; 224 *ppoints = points; 225 } 226 } while_each_thread(g, p); 227 return chosen; 228} 229 230/** 231 * We must be careful though to never send SIGKILL a process with 232 * CAP_SYS_RAW_IO set, send SIGTERM instead (but it's unlikely that 233 * we select a process with CAP_SYS_RAW_IO set). 234 */ 235static void __oom_kill_task(struct task_struct *p, const char *message) 236{ 237 if (p->pid == 1) { 238 WARN_ON(1); 239 printk(KERN_WARNING "tried to kill init!\n"); 240 return; 241 } 242 243 task_lock(p); 244 if (!p->mm || p->mm == &init_mm) { 245 WARN_ON(1); 246 printk(KERN_WARNING "tried to kill an mm-less task!\n"); 247 task_unlock(p); 248 return; 249 } 250 task_unlock(p); 251 printk(KERN_ERR "%s: Killed process %d (%s).\n", 252 message, p->pid, p->comm); 253 254 /* 255 * We give our sacrificial lamb high priority and access to 256 * all the memory it needs. That way it should be able to 257 * exit() and clear out its resources quickly... 258 */ 259 p->time_slice = HZ; 260 set_tsk_thread_flag(p, TIF_MEMDIE); 261 262 force_sig(SIGKILL, p); 263} 264 265static int oom_kill_task(struct task_struct *p, const char *message) 266{ 267 struct mm_struct *mm; 268 struct task_struct *g, *q; 269 270 mm = p->mm; 271 272 /* WARNING: mm may not be dereferenced since we did not obtain its 273 * value from get_task_mm(p). This is OK since all we need to do is 274 * compare mm to q->mm below. 275 * 276 * Furthermore, even if mm contains a non-NULL value, p->mm may 277 * change to NULL at any time since we do not hold task_lock(p). 278 * However, this is of no concern to us. 279 */ 280 281 if (mm == NULL || mm == &init_mm) 282 return 1; 283 284 __oom_kill_task(p, message); 285 /* 286 * kill all processes that share the ->mm (i.e. all threads), 287 * but are in a different thread group 288 */ 289 do_each_thread(g, q) 290 if (q->mm == mm && q->tgid != p->tgid) 291 __oom_kill_task(q, message); 292 while_each_thread(g, q); 293 294 return 0; 295} 296 297static int oom_kill_process(struct task_struct *p, unsigned long points, 298 const char *message) 299{ 300 struct task_struct *c; 301 struct list_head *tsk; 302 303 printk(KERN_ERR "Out of Memory: Kill process %d (%s) score %li and " 304 "children.\n", p->pid, p->comm, points); 305 /* Try to kill a child first */ 306 list_for_each(tsk, &p->children) { 307 c = list_entry(tsk, struct task_struct, sibling); 308 if (c->mm == p->mm) 309 continue; 310 if (!oom_kill_task(c, message)) 311 return 0; 312 } 313 return oom_kill_task(p, message); 314} 315 316static BLOCKING_NOTIFIER_HEAD(oom_notify_list); 317 318int register_oom_notifier(struct notifier_block *nb) 319{ 320 return blocking_notifier_chain_register(&oom_notify_list, nb); 321} 322EXPORT_SYMBOL_GPL(register_oom_notifier); 323 324int unregister_oom_notifier(struct notifier_block *nb) 325{ 326 return blocking_notifier_chain_unregister(&oom_notify_list, nb); 327} 328EXPORT_SYMBOL_GPL(unregister_oom_notifier); 329 330/** 331 * out_of_memory - kill the "best" process when we run out of memory 332 * 333 * If we run out of memory, we have the choice between either 334 * killing a random task (bad), letting the system crash (worse) 335 * OR try to be smart about which process to kill. Note that we 336 * don't have to be perfect here, we just have to be good. 337 */ 338void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order) 339{ 340 struct task_struct *p; 341 unsigned long points = 0; 342 unsigned long freed = 0; 343 344 blocking_notifier_call_chain(&oom_notify_list, 0, &freed); 345 if (freed > 0) 346 /* Got some memory back in the last second. */ 347 return; 348 349 if (printk_ratelimit()) { 350 printk("oom-killer: gfp_mask=0x%x, order=%d\n", 351 gfp_mask, order); 352 dump_stack(); 353 show_mem(); 354 } 355 356 cpuset_lock(); 357 read_lock(&tasklist_lock); 358 359 /* 360 * Check if there were limitations on the allocation (only relevant for 361 * NUMA) that may require different handling. 362 */ 363 switch (constrained_alloc(zonelist, gfp_mask)) { 364 case CONSTRAINT_MEMORY_POLICY: 365 oom_kill_process(current, points, 366 "No available memory (MPOL_BIND)"); 367 break; 368 369 case CONSTRAINT_CPUSET: 370 oom_kill_process(current, points, 371 "No available memory in cpuset"); 372 break; 373 374 case CONSTRAINT_NONE: 375 if (sysctl_panic_on_oom) 376 panic("out of memory. panic_on_oom is selected\n"); 377retry: 378 /* 379 * Rambo mode: Shoot down a process and hope it solves whatever 380 * issues we may have. 381 */ 382 p = select_bad_process(&points); 383 384 if (PTR_ERR(p) == -1UL) 385 goto out; 386 387 /* Found nothing?!?! Either we hang forever, or we panic. */ 388 if (!p) { 389 read_unlock(&tasklist_lock); 390 cpuset_unlock(); 391 panic("Out of memory and no killable processes...\n"); 392 } 393 394 if (oom_kill_process(p, points, "Out of memory")) 395 goto retry; 396 397 break; 398 } 399 400out: 401 read_unlock(&tasklist_lock); 402 cpuset_unlock(); 403 404 /* 405 * Give "p" a good chance of killing itself before we 406 * retry to allocate memory unless "p" is current 407 */ 408 if (!test_thread_flag(TIF_MEMDIE)) 409 schedule_timeout_uninterruptible(1); 410} 411