aboutsummaryrefslogtreecommitdiffstats
path: root/mm/oom_kill.c
diff options
context:
space:
mode:
Diffstat (limited to 'mm/oom_kill.c')
-rw-r--r--mm/oom_kill.c687
1 files changed, 371 insertions, 316 deletions
diff --git a/mm/oom_kill.c b/mm/oom_kill.c
index 709aedfaa014..fc81cb22869e 100644
--- a/mm/oom_kill.c
+++ b/mm/oom_kill.c
@@ -4,6 +4,8 @@
4 * Copyright (C) 1998,2000 Rik van Riel 4 * Copyright (C) 1998,2000 Rik van Riel
5 * Thanks go out to Claus Fischer for some serious inspiration and 5 * Thanks go out to Claus Fischer for some serious inspiration and
6 * for goading me into coding this file... 6 * for goading me into coding this file...
7 * Copyright (C) 2010 Google, Inc.
8 * Rewritten by David Rientjes
7 * 9 *
8 * The routines in this file are used to kill a process when 10 * 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() 11 * we're seriously out of memory. This gets called from __alloc_pages()
@@ -27,171 +29,188 @@
27#include <linux/module.h> 29#include <linux/module.h>
28#include <linux/notifier.h> 30#include <linux/notifier.h>
29#include <linux/memcontrol.h> 31#include <linux/memcontrol.h>
32#include <linux/mempolicy.h>
30#include <linux/security.h> 33#include <linux/security.h>
31 34
32int sysctl_panic_on_oom; 35int sysctl_panic_on_oom;
33int sysctl_oom_kill_allocating_task; 36int sysctl_oom_kill_allocating_task;
34int sysctl_oom_dump_tasks; 37int sysctl_oom_dump_tasks = 1;
35static DEFINE_SPINLOCK(zone_scan_lock); 38static DEFINE_SPINLOCK(zone_scan_lock);
36/* #define DEBUG */ 39
40#ifdef CONFIG_NUMA
41/**
42 * has_intersects_mems_allowed() - check task eligiblity for kill
43 * @tsk: task struct of which task to consider
44 * @mask: nodemask passed to page allocator for mempolicy ooms
45 *
46 * Task eligibility is determined by whether or not a candidate task, @tsk,
47 * shares the same mempolicy nodes as current if it is bound by such a policy
48 * and whether or not it has the same set of allowed cpuset nodes.
49 */
50static bool has_intersects_mems_allowed(struct task_struct *tsk,
51 const nodemask_t *mask)
52{
53 struct task_struct *start = tsk;
54
55 do {
56 if (mask) {
57 /*
58 * If this is a mempolicy constrained oom, tsk's
59 * cpuset is irrelevant. Only return true if its
60 * mempolicy intersects current, otherwise it may be
61 * needlessly killed.
62 */
63 if (mempolicy_nodemask_intersects(tsk, mask))
64 return true;
65 } else {
66 /*
67 * This is not a mempolicy constrained oom, so only
68 * check the mems of tsk's cpuset.
69 */
70 if (cpuset_mems_allowed_intersects(current, tsk))
71 return true;
72 }
73 } while_each_thread(start, tsk);
74
75 return false;
76}
77#else
78static bool has_intersects_mems_allowed(struct task_struct *tsk,
79 const nodemask_t *mask)
80{
81 return true;
82}
83#endif /* CONFIG_NUMA */
37 84
38/* 85/*
39 * Is all threads of the target process nodes overlap ours? 86 * If this is a system OOM (not a memcg OOM) and the task selected to be
87 * killed is not already running at high (RT) priorities, speed up the
88 * recovery by boosting the dying task to the lowest FIFO priority.
89 * That helps with the recovery and avoids interfering with RT tasks.
40 */ 90 */
41static int has_intersects_mems_allowed(struct task_struct *tsk) 91static void boost_dying_task_prio(struct task_struct *p,
92 struct mem_cgroup *mem)
42{ 93{
43 struct task_struct *t; 94 struct sched_param param = { .sched_priority = 1 };
95
96 if (mem)
97 return;
98
99 if (!rt_task(p))
100 sched_setscheduler_nocheck(p, SCHED_FIFO, &param);
101}
102
103/*
104 * The process p may have detached its own ->mm while exiting or through
105 * use_mm(), but one or more of its subthreads may still have a valid
106 * pointer. Return p, or any of its subthreads with a valid ->mm, with
107 * task_lock() held.
108 */
109struct task_struct *find_lock_task_mm(struct task_struct *p)
110{
111 struct task_struct *t = p;
44 112
45 t = tsk;
46 do { 113 do {
47 if (cpuset_mems_allowed_intersects(current, t)) 114 task_lock(t);
48 return 1; 115 if (likely(t->mm))
49 t = next_thread(t); 116 return t;
50 } while (t != tsk); 117 task_unlock(t);
118 } while_each_thread(p, t);
51 119
52 return 0; 120 return NULL;
121}
122
123/* return true if the task is not adequate as candidate victim task. */
124static bool oom_unkillable_task(struct task_struct *p, struct mem_cgroup *mem,
125 const nodemask_t *nodemask)
126{
127 if (is_global_init(p))
128 return true;
129 if (p->flags & PF_KTHREAD)
130 return true;
131
132 /* When mem_cgroup_out_of_memory() and p is not member of the group */
133 if (mem && !task_in_mem_cgroup(p, mem))
134 return true;
135
136 /* p may not have freeable memory in nodemask */
137 if (!has_intersects_mems_allowed(p, nodemask))
138 return true;
139
140 return false;
53} 141}
54 142
55/** 143/**
56 * badness - calculate a numeric value for how bad this task has been 144 * oom_badness - heuristic function to determine which candidate task to kill
57 * @p: task struct of which task we should calculate 145 * @p: task struct of which task we should calculate
58 * @uptime: current uptime in seconds 146 * @totalpages: total present RAM allowed for page allocation
59 *
60 * The formula used is relatively simple and documented inline in the
61 * function. The main rationale is that we want to select a good task
62 * to kill when we run out of memory.
63 * 147 *
64 * Good in this context means that: 148 * The heuristic for determining which task to kill is made to be as simple and
65 * 1) we lose the minimum amount of work done 149 * predictable as possible. The goal is to return the highest value for the
66 * 2) we recover a large amount of memory 150 * task consuming the most memory to avoid subsequent oom failures.
67 * 3) we don't kill anything innocent of eating tons of memory
68 * 4) we want to kill the minimum amount of processes (one)
69 * 5) we try to kill the process the user expects us to kill, this
70 * algorithm has been meticulously tuned to meet the principle
71 * of least surprise ... (be careful when you change it)
72 */ 151 */
73 152unsigned int oom_badness(struct task_struct *p, struct mem_cgroup *mem,
74unsigned long badness(struct task_struct *p, unsigned long uptime) 153 const nodemask_t *nodemask, unsigned long totalpages)
75{ 154{
76 unsigned long points, cpu_time, run_time; 155 int points;
77 struct mm_struct *mm;
78 struct task_struct *child;
79 int oom_adj = p->signal->oom_adj;
80 struct task_cputime task_time;
81 unsigned long utime;
82 unsigned long stime;
83 156
84 if (oom_adj == OOM_DISABLE) 157 if (oom_unkillable_task(p, mem, nodemask))
85 return 0; 158 return 0;
86 159
87 task_lock(p); 160 p = find_lock_task_mm(p);
88 mm = p->mm; 161 if (!p)
89 if (!mm) {
90 task_unlock(p);
91 return 0; 162 return 0;
92 }
93
94 /*
95 * The memory size of the process is the basis for the badness.
96 */
97 points = mm->total_vm;
98 163
99 /* 164 /*
100 * After this unlock we can no longer dereference local variable `mm' 165 * Shortcut check for OOM_SCORE_ADJ_MIN so the entire heuristic doesn't
166 * need to be executed for something that cannot be killed.
101 */ 167 */
102 task_unlock(p); 168 if (p->signal->oom_score_adj == OOM_SCORE_ADJ_MIN) {
103 169 task_unlock(p);
104 /* 170 return 0;
105 * swapoff can easily use up all memory, so kill those first.
106 */
107 if (p->flags & PF_OOM_ORIGIN)
108 return ULONG_MAX;
109
110 /*
111 * Processes which fork a lot of child processes are likely
112 * a good choice. We add half the vmsize of the children if they
113 * have an own mm. This prevents forking servers to flood the
114 * machine with an endless amount of children. In case a single
115 * child is eating the vast majority of memory, adding only half
116 * to the parents will make the child our kill candidate of choice.
117 */
118 list_for_each_entry(child, &p->children, sibling) {
119 task_lock(child);
120 if (child->mm != mm && child->mm)
121 points += child->mm->total_vm/2 + 1;
122 task_unlock(child);
123 } 171 }
124 172
125 /* 173 /*
126 * CPU time is in tens of seconds and run time is in thousands 174 * When the PF_OOM_ORIGIN bit is set, it indicates the task should have
127 * of seconds. There is no particular reason for this other than 175 * priority for oom killing.
128 * that it turned out to work very well in practice.
129 */
130 thread_group_cputime(p, &task_time);
131 utime = cputime_to_jiffies(task_time.utime);
132 stime = cputime_to_jiffies(task_time.stime);
133 cpu_time = (utime + stime) >> (SHIFT_HZ + 3);
134
135
136 if (uptime >= p->start_time.tv_sec)
137 run_time = (uptime - p->start_time.tv_sec) >> 10;
138 else
139 run_time = 0;
140
141 if (cpu_time)
142 points /= int_sqrt(cpu_time);
143 if (run_time)
144 points /= int_sqrt(int_sqrt(run_time));
145
146 /*
147 * Niced processes are most likely less important, so double
148 * their badness points.
149 */ 176 */
150 if (task_nice(p) > 0) 177 if (p->flags & PF_OOM_ORIGIN) {
151 points *= 2; 178 task_unlock(p);
179 return 1000;
180 }
152 181
153 /* 182 /*
154 * Superuser processes are usually more important, so we make it 183 * The memory controller may have a limit of 0 bytes, so avoid a divide
155 * less likely that we kill those. 184 * by zero, if necessary.
156 */ 185 */
157 if (has_capability_noaudit(p, CAP_SYS_ADMIN) || 186 if (!totalpages)
158 has_capability_noaudit(p, CAP_SYS_RESOURCE)) 187 totalpages = 1;
159 points /= 4;
160 188
161 /* 189 /*
162 * We don't want to kill a process with direct hardware access. 190 * The baseline for the badness score is the proportion of RAM that each
163 * Not only could that mess up the hardware, but usually users 191 * task's rss and swap space use.
164 * tend to only have this flag set on applications they think
165 * of as important.
166 */ 192 */
167 if (has_capability_noaudit(p, CAP_SYS_RAWIO)) 193 points = (get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS)) * 1000 /
168 points /= 4; 194 totalpages;
195 task_unlock(p);
169 196
170 /* 197 /*
171 * If p's nodes don't overlap ours, it may still help to kill p 198 * Root processes get 3% bonus, just like the __vm_enough_memory()
172 * because p may have allocated or otherwise mapped memory on 199 * implementation used by LSMs.
173 * this node before. However it will be less likely.
174 */ 200 */
175 if (!has_intersects_mems_allowed(p)) 201 if (has_capability_noaudit(p, CAP_SYS_ADMIN))
176 points /= 8; 202 points -= 30;
177 203
178 /* 204 /*
179 * Adjust the score by oom_adj. 205 * /proc/pid/oom_score_adj ranges from -1000 to +1000 such that it may
206 * either completely disable oom killing or always prefer a certain
207 * task.
180 */ 208 */
181 if (oom_adj) { 209 points += p->signal->oom_score_adj;
182 if (oom_adj > 0) {
183 if (!points)
184 points = 1;
185 points <<= oom_adj;
186 } else
187 points >>= -(oom_adj);
188 }
189 210
190#ifdef DEBUG 211 if (points < 0)
191 printk(KERN_DEBUG "OOMkill: task %d (%s) got %lu points\n", 212 return 0;
192 p->pid, p->comm, points); 213 return (points < 1000) ? points : 1000;
193#endif
194 return points;
195} 214}
196 215
197/* 216/*
@@ -199,12 +218,20 @@ unsigned long badness(struct task_struct *p, unsigned long uptime)
199 */ 218 */
200#ifdef CONFIG_NUMA 219#ifdef CONFIG_NUMA
201static enum oom_constraint constrained_alloc(struct zonelist *zonelist, 220static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
202 gfp_t gfp_mask, nodemask_t *nodemask) 221 gfp_t gfp_mask, nodemask_t *nodemask,
222 unsigned long *totalpages)
203{ 223{
204 struct zone *zone; 224 struct zone *zone;
205 struct zoneref *z; 225 struct zoneref *z;
206 enum zone_type high_zoneidx = gfp_zone(gfp_mask); 226 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
227 bool cpuset_limited = false;
228 int nid;
229
230 /* Default to all available memory */
231 *totalpages = totalram_pages + total_swap_pages;
207 232
233 if (!zonelist)
234 return CONSTRAINT_NONE;
208 /* 235 /*
209 * Reach here only when __GFP_NOFAIL is used. So, we should avoid 236 * Reach here only when __GFP_NOFAIL is used. So, we should avoid
210 * to kill current.We have to random task kill in this case. 237 * to kill current.We have to random task kill in this case.
@@ -214,26 +241,37 @@ static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
214 return CONSTRAINT_NONE; 241 return CONSTRAINT_NONE;
215 242
216 /* 243 /*
217 * The nodemask here is a nodemask passed to alloc_pages(). Now, 244 * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
218 * cpuset doesn't use this nodemask for its hardwall/softwall/hierarchy 245 * the page allocator means a mempolicy is in effect. Cpuset policy
219 * feature. mempolicy is an only user of nodemask here. 246 * is enforced in get_page_from_freelist().
220 * check mempolicy's nodemask contains all N_HIGH_MEMORY
221 */ 247 */
222 if (nodemask && !nodes_subset(node_states[N_HIGH_MEMORY], *nodemask)) 248 if (nodemask && !nodes_subset(node_states[N_HIGH_MEMORY], *nodemask)) {
249 *totalpages = total_swap_pages;
250 for_each_node_mask(nid, *nodemask)
251 *totalpages += node_spanned_pages(nid);
223 return CONSTRAINT_MEMORY_POLICY; 252 return CONSTRAINT_MEMORY_POLICY;
253 }
224 254
225 /* Check this allocation failure is caused by cpuset's wall function */ 255 /* Check this allocation failure is caused by cpuset's wall function */
226 for_each_zone_zonelist_nodemask(zone, z, zonelist, 256 for_each_zone_zonelist_nodemask(zone, z, zonelist,
227 high_zoneidx, nodemask) 257 high_zoneidx, nodemask)
228 if (!cpuset_zone_allowed_softwall(zone, gfp_mask)) 258 if (!cpuset_zone_allowed_softwall(zone, gfp_mask))
229 return CONSTRAINT_CPUSET; 259 cpuset_limited = true;
230 260
261 if (cpuset_limited) {
262 *totalpages = total_swap_pages;
263 for_each_node_mask(nid, cpuset_current_mems_allowed)
264 *totalpages += node_spanned_pages(nid);
265 return CONSTRAINT_CPUSET;
266 }
231 return CONSTRAINT_NONE; 267 return CONSTRAINT_NONE;
232} 268}
233#else 269#else
234static enum oom_constraint constrained_alloc(struct zonelist *zonelist, 270static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
235 gfp_t gfp_mask, nodemask_t *nodemask) 271 gfp_t gfp_mask, nodemask_t *nodemask,
272 unsigned long *totalpages)
236{ 273{
274 *totalpages = totalram_pages + total_swap_pages;
237 return CONSTRAINT_NONE; 275 return CONSTRAINT_NONE;
238} 276}
239#endif 277#endif
@@ -244,28 +282,18 @@ static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
244 * 282 *
245 * (not docbooked, we don't want this one cluttering up the manual) 283 * (not docbooked, we don't want this one cluttering up the manual)
246 */ 284 */
247static struct task_struct *select_bad_process(unsigned long *ppoints, 285static struct task_struct *select_bad_process(unsigned int *ppoints,
248 struct mem_cgroup *mem) 286 unsigned long totalpages, struct mem_cgroup *mem,
287 const nodemask_t *nodemask)
249{ 288{
250 struct task_struct *p; 289 struct task_struct *p;
251 struct task_struct *chosen = NULL; 290 struct task_struct *chosen = NULL;
252 struct timespec uptime;
253 *ppoints = 0; 291 *ppoints = 0;
254 292
255 do_posix_clock_monotonic_gettime(&uptime);
256 for_each_process(p) { 293 for_each_process(p) {
257 unsigned long points; 294 unsigned int points;
258 295
259 /* 296 if (oom_unkillable_task(p, mem, nodemask))
260 * skip kernel threads and tasks which have already released
261 * their mm.
262 */
263 if (!p->mm)
264 continue;
265 /* skip the init task */
266 if (is_global_init(p))
267 continue;
268 if (mem && !task_in_mem_cgroup(p, mem))
269 continue; 297 continue;
270 298
271 /* 299 /*
@@ -290,19 +318,16 @@ static struct task_struct *select_bad_process(unsigned long *ppoints,
290 * the process of exiting and releasing its resources. 318 * the process of exiting and releasing its resources.
291 * Otherwise we could get an easy OOM deadlock. 319 * Otherwise we could get an easy OOM deadlock.
292 */ 320 */
293 if (p->flags & PF_EXITING) { 321 if (thread_group_empty(p) && (p->flags & PF_EXITING) && p->mm) {
294 if (p != current) 322 if (p != current)
295 return ERR_PTR(-1UL); 323 return ERR_PTR(-1UL);
296 324
297 chosen = p; 325 chosen = p;
298 *ppoints = ULONG_MAX; 326 *ppoints = 1000;
299 } 327 }
300 328
301 if (p->signal->oom_adj == OOM_DISABLE) 329 points = oom_badness(p, mem, nodemask, totalpages);
302 continue; 330 if (points > *ppoints) {
303
304 points = badness(p, uptime.tv_sec);
305 if (points > *ppoints || !chosen) {
306 chosen = p; 331 chosen = p;
307 *ppoints = points; 332 *ppoints = points;
308 } 333 }
@@ -313,11 +338,11 @@ static struct task_struct *select_bad_process(unsigned long *ppoints,
313 338
314/** 339/**
315 * dump_tasks - dump current memory state of all system tasks 340 * dump_tasks - dump current memory state of all system tasks
316 * @mem: target memory controller 341 * @mem: current's memory controller, if constrained
317 * 342 *
318 * Dumps the current memory state of all system tasks, excluding kernel threads. 343 * Dumps the current memory state of all system tasks, excluding kernel threads.
319 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj 344 * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
320 * score, and name. 345 * value, oom_score_adj value, and name.
321 * 346 *
322 * If the actual is non-NULL, only tasks that are a member of the mem_cgroup are 347 * If the actual is non-NULL, only tasks that are a member of the mem_cgroup are
323 * shown. 348 * shown.
@@ -326,44 +351,43 @@ static struct task_struct *select_bad_process(unsigned long *ppoints,
326 */ 351 */
327static void dump_tasks(const struct mem_cgroup *mem) 352static void dump_tasks(const struct mem_cgroup *mem)
328{ 353{
329 struct task_struct *g, *p; 354 struct task_struct *p;
330 355 struct task_struct *task;
331 printk(KERN_INFO "[ pid ] uid tgid total_vm rss cpu oom_adj "
332 "name\n");
333 do_each_thread(g, p) {
334 struct mm_struct *mm;
335 356
336 if (mem && !task_in_mem_cgroup(p, mem)) 357 pr_info("[ pid ] uid tgid total_vm rss cpu oom_adj oom_score_adj name\n");
358 for_each_process(p) {
359 if (p->flags & PF_KTHREAD)
337 continue; 360 continue;
338 if (!thread_group_leader(p)) 361 if (mem && !task_in_mem_cgroup(p, mem))
339 continue; 362 continue;
340 363
341 task_lock(p); 364 task = find_lock_task_mm(p);
342 mm = p->mm; 365 if (!task) {
343 if (!mm) {
344 /* 366 /*
345 * total_vm and rss sizes do not exist for tasks with no 367 * This is a kthread or all of p's threads have already
346 * mm so there's no need to report them; they can't be 368 * detached their mm's. There's no need to report
347 * oom killed anyway. 369 * them; they can't be oom killed anyway.
348 */ 370 */
349 task_unlock(p);
350 continue; 371 continue;
351 } 372 }
352 printk(KERN_INFO "[%5d] %5d %5d %8lu %8lu %3d %3d %s\n", 373
353 p->pid, __task_cred(p)->uid, p->tgid, mm->total_vm, 374 pr_info("[%5d] %5d %5d %8lu %8lu %3u %3d %5d %s\n",
354 get_mm_rss(mm), (int)task_cpu(p), p->signal->oom_adj, 375 task->pid, task_uid(task), task->tgid,
355 p->comm); 376 task->mm->total_vm, get_mm_rss(task->mm),
356 task_unlock(p); 377 task_cpu(task), task->signal->oom_adj,
357 } while_each_thread(g, p); 378 task->signal->oom_score_adj, task->comm);
379 task_unlock(task);
380 }
358} 381}
359 382
360static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order, 383static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
361 struct mem_cgroup *mem) 384 struct mem_cgroup *mem)
362{ 385{
363 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
364 "oom_adj=%d\n",
365 current->comm, gfp_mask, order, current->signal->oom_adj);
366 task_lock(current); 386 task_lock(current);
387 pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
388 "oom_adj=%d, oom_score_adj=%d\n",
389 current->comm, gfp_mask, order, current->signal->oom_adj,
390 current->signal->oom_score_adj);
367 cpuset_print_task_mems_allowed(current); 391 cpuset_print_task_mems_allowed(current);
368 task_unlock(current); 392 task_unlock(current);
369 dump_stack(); 393 dump_stack();
@@ -374,72 +398,42 @@ static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
374} 398}
375 399
376#define K(x) ((x) << (PAGE_SHIFT-10)) 400#define K(x) ((x) << (PAGE_SHIFT-10))
377 401static int oom_kill_task(struct task_struct *p, struct mem_cgroup *mem)
378/*
379 * Send SIGKILL to the selected process irrespective of CAP_SYS_RAW_IO
380 * flag though it's unlikely that we select a process with CAP_SYS_RAW_IO
381 * set.
382 */
383static void __oom_kill_task(struct task_struct *p, int verbose)
384{ 402{
385 if (is_global_init(p)) { 403 p = find_lock_task_mm(p);
386 WARN_ON(1); 404 if (!p)
387 printk(KERN_WARNING "tried to kill init!\n"); 405 return 1;
388 return;
389 }
390
391 task_lock(p);
392 if (!p->mm) {
393 WARN_ON(1);
394 printk(KERN_WARNING "tried to kill an mm-less task %d (%s)!\n",
395 task_pid_nr(p), p->comm);
396 task_unlock(p);
397 return;
398 }
399 406
400 if (verbose) 407 pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
401 printk(KERN_ERR "Killed process %d (%s) " 408 task_pid_nr(p), p->comm, K(p->mm->total_vm),
402 "vsz:%lukB, anon-rss:%lukB, file-rss:%lukB\n", 409 K(get_mm_counter(p->mm, MM_ANONPAGES)),
403 task_pid_nr(p), p->comm, 410 K(get_mm_counter(p->mm, MM_FILEPAGES)));
404 K(p->mm->total_vm),
405 K(get_mm_counter(p->mm, MM_ANONPAGES)),
406 K(get_mm_counter(p->mm, MM_FILEPAGES)));
407 task_unlock(p); 411 task_unlock(p);
408 412
413
414 set_tsk_thread_flag(p, TIF_MEMDIE);
415 force_sig(SIGKILL, p);
416
409 /* 417 /*
410 * We give our sacrificial lamb high priority and access to 418 * We give our sacrificial lamb high priority and access to
411 * all the memory it needs. That way it should be able to 419 * all the memory it needs. That way it should be able to
412 * exit() and clear out its resources quickly... 420 * exit() and clear out its resources quickly...
413 */ 421 */
414 p->rt.time_slice = HZ; 422 boost_dying_task_prio(p, mem);
415 set_tsk_thread_flag(p, TIF_MEMDIE);
416
417 force_sig(SIGKILL, p);
418}
419
420static int oom_kill_task(struct task_struct *p)
421{
422 /* WARNING: mm may not be dereferenced since we did not obtain its
423 * value from get_task_mm(p). This is OK since all we need to do is
424 * compare mm to q->mm below.
425 *
426 * Furthermore, even if mm contains a non-NULL value, p->mm may
427 * change to NULL at any time since we do not hold task_lock(p).
428 * However, this is of no concern to us.
429 */
430 if (!p->mm || p->signal->oom_adj == OOM_DISABLE)
431 return 1;
432
433 __oom_kill_task(p, 1);
434 423
435 return 0; 424 return 0;
436} 425}
426#undef K
437 427
438static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order, 428static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
439 unsigned long points, struct mem_cgroup *mem, 429 unsigned int points, unsigned long totalpages,
430 struct mem_cgroup *mem, nodemask_t *nodemask,
440 const char *message) 431 const char *message)
441{ 432{
442 struct task_struct *c; 433 struct task_struct *victim = p;
434 struct task_struct *child;
435 struct task_struct *t = p;
436 unsigned int victim_points = 0;
443 437
444 if (printk_ratelimit()) 438 if (printk_ratelimit())
445 dump_header(p, gfp_mask, order, mem); 439 dump_header(p, gfp_mask, order, mem);
@@ -449,40 +443,81 @@ static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
449 * its children or threads, just set TIF_MEMDIE so it can die quickly 443 * its children or threads, just set TIF_MEMDIE so it can die quickly
450 */ 444 */
451 if (p->flags & PF_EXITING) { 445 if (p->flags & PF_EXITING) {
452 __oom_kill_task(p, 0); 446 set_tsk_thread_flag(p, TIF_MEMDIE);
447 boost_dying_task_prio(p, mem);
453 return 0; 448 return 0;
454 } 449 }
455 450
456 printk(KERN_ERR "%s: kill process %d (%s) score %li or a child\n", 451 task_lock(p);
457 message, task_pid_nr(p), p->comm, points); 452 pr_err("%s: Kill process %d (%s) score %d or sacrifice child\n",
453 message, task_pid_nr(p), p->comm, points);
454 task_unlock(p);
458 455
459 /* Try to kill a child first */ 456 /*
460 list_for_each_entry(c, &p->children, sibling) { 457 * If any of p's children has a different mm and is eligible for kill,
461 if (c->mm == p->mm) 458 * the one with the highest badness() score is sacrificed for its
462 continue; 459 * parent. This attempts to lose the minimal amount of work done while
463 if (mem && !task_in_mem_cgroup(c, mem)) 460 * still freeing memory.
464 continue; 461 */
465 if (!oom_kill_task(c)) 462 do {
466 return 0; 463 list_for_each_entry(child, &t->children, sibling) {
464 unsigned int child_points;
465
466 /*
467 * oom_badness() returns 0 if the thread is unkillable
468 */
469 child_points = oom_badness(child, mem, nodemask,
470 totalpages);
471 if (child_points > victim_points) {
472 victim = child;
473 victim_points = child_points;
474 }
475 }
476 } while_each_thread(p, t);
477
478 return oom_kill_task(victim, mem);
479}
480
481/*
482 * Determines whether the kernel must panic because of the panic_on_oom sysctl.
483 */
484static void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
485 int order)
486{
487 if (likely(!sysctl_panic_on_oom))
488 return;
489 if (sysctl_panic_on_oom != 2) {
490 /*
491 * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
492 * does not panic for cpuset, mempolicy, or memcg allocation
493 * failures.
494 */
495 if (constraint != CONSTRAINT_NONE)
496 return;
467 } 497 }
468 return oom_kill_task(p); 498 read_lock(&tasklist_lock);
499 dump_header(NULL, gfp_mask, order, NULL);
500 read_unlock(&tasklist_lock);
501 panic("Out of memory: %s panic_on_oom is enabled\n",
502 sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
469} 503}
470 504
471#ifdef CONFIG_CGROUP_MEM_RES_CTLR 505#ifdef CONFIG_CGROUP_MEM_RES_CTLR
472void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask) 506void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
473{ 507{
474 unsigned long points = 0; 508 unsigned long limit;
509 unsigned int points = 0;
475 struct task_struct *p; 510 struct task_struct *p;
476 511
477 if (sysctl_panic_on_oom == 2) 512 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, 0);
478 panic("out of memory(memcg). panic_on_oom is selected.\n"); 513 limit = mem_cgroup_get_limit(mem) >> PAGE_SHIFT;
479 read_lock(&tasklist_lock); 514 read_lock(&tasklist_lock);
480retry: 515retry:
481 p = select_bad_process(&points, mem); 516 p = select_bad_process(&points, limit, mem, NULL);
482 if (!p || PTR_ERR(p) == -1UL) 517 if (!p || PTR_ERR(p) == -1UL)
483 goto out; 518 goto out;
484 519
485 if (oom_kill_process(p, gfp_mask, 0, points, mem, 520 if (oom_kill_process(p, gfp_mask, 0, points, limit, mem, NULL,
486 "Memory cgroup out of memory")) 521 "Memory cgroup out of memory"))
487 goto retry; 522 goto retry;
488out: 523out:
@@ -509,7 +544,7 @@ EXPORT_SYMBOL_GPL(unregister_oom_notifier);
509 * if a parallel OOM killing is already taking place that includes a zone in 544 * if a parallel OOM killing is already taking place that includes a zone in
510 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1. 545 * the zonelist. Otherwise, locks all zones in the zonelist and returns 1.
511 */ 546 */
512int try_set_zone_oom(struct zonelist *zonelist, gfp_t gfp_mask) 547int try_set_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
513{ 548{
514 struct zoneref *z; 549 struct zoneref *z;
515 struct zone *zone; 550 struct zone *zone;
@@ -526,7 +561,7 @@ int try_set_zone_oom(struct zonelist *zonelist, gfp_t gfp_mask)
526 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) { 561 for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
527 /* 562 /*
528 * Lock each zone in the zonelist under zone_scan_lock so a 563 * Lock each zone in the zonelist under zone_scan_lock so a
529 * parallel invocation of try_set_zone_oom() doesn't succeed 564 * parallel invocation of try_set_zonelist_oom() doesn't succeed
530 * when it shouldn't. 565 * when it shouldn't.
531 */ 566 */
532 zone_set_flag(zone, ZONE_OOM_LOCKED); 567 zone_set_flag(zone, ZONE_OOM_LOCKED);
@@ -555,65 +590,40 @@ void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
555} 590}
556 591
557/* 592/*
558 * Must be called with tasklist_lock held for read. 593 * Try to acquire the oom killer lock for all system zones. Returns zero if a
594 * parallel oom killing is taking place, otherwise locks all zones and returns
595 * non-zero.
559 */ 596 */
560static void __out_of_memory(gfp_t gfp_mask, int order) 597static int try_set_system_oom(void)
561{ 598{
562 struct task_struct *p; 599 struct zone *zone;
563 unsigned long points; 600 int ret = 1;
564
565 if (sysctl_oom_kill_allocating_task)
566 if (!oom_kill_process(current, gfp_mask, order, 0, NULL,
567 "Out of memory (oom_kill_allocating_task)"))
568 return;
569retry:
570 /*
571 * Rambo mode: Shoot down a process and hope it solves whatever
572 * issues we may have.
573 */
574 p = select_bad_process(&points, NULL);
575
576 if (PTR_ERR(p) == -1UL)
577 return;
578
579 /* Found nothing?!?! Either we hang forever, or we panic. */
580 if (!p) {
581 read_unlock(&tasklist_lock);
582 dump_header(NULL, gfp_mask, order, NULL);
583 panic("Out of memory and no killable processes...\n");
584 }
585 601
586 if (oom_kill_process(p, gfp_mask, order, points, NULL, 602 spin_lock(&zone_scan_lock);
587 "Out of memory")) 603 for_each_populated_zone(zone)
588 goto retry; 604 if (zone_is_oom_locked(zone)) {
605 ret = 0;
606 goto out;
607 }
608 for_each_populated_zone(zone)
609 zone_set_flag(zone, ZONE_OOM_LOCKED);
610out:
611 spin_unlock(&zone_scan_lock);
612 return ret;
589} 613}
590 614
591/* 615/*
592 * pagefault handler calls into here because it is out of memory but 616 * Clears ZONE_OOM_LOCKED for all system zones so that failed allocation
593 * doesn't know exactly how or why. 617 * attempts or page faults may now recall the oom killer, if necessary.
594 */ 618 */
595void pagefault_out_of_memory(void) 619static void clear_system_oom(void)
596{ 620{
597 unsigned long freed = 0; 621 struct zone *zone;
598
599 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
600 if (freed > 0)
601 /* Got some memory back in the last second. */
602 return;
603
604 if (sysctl_panic_on_oom)
605 panic("out of memory from page fault. panic_on_oom is selected.\n");
606
607 read_lock(&tasklist_lock);
608 __out_of_memory(0, 0); /* unknown gfp_mask and order */
609 read_unlock(&tasklist_lock);
610 622
611 /* 623 spin_lock(&zone_scan_lock);
612 * Give "p" a good chance of killing itself before we 624 for_each_populated_zone(zone)
613 * retry to allocate memory. 625 zone_clear_flag(zone, ZONE_OOM_LOCKED);
614 */ 626 spin_unlock(&zone_scan_lock);
615 if (!test_thread_flag(TIF_MEMDIE))
616 schedule_timeout_uninterruptible(1);
617} 627}
618 628
619/** 629/**
@@ -621,6 +631,7 @@ void pagefault_out_of_memory(void)
621 * @zonelist: zonelist pointer 631 * @zonelist: zonelist pointer
622 * @gfp_mask: memory allocation flags 632 * @gfp_mask: memory allocation flags
623 * @order: amount of memory being requested as a power of 2 633 * @order: amount of memory being requested as a power of 2
634 * @nodemask: nodemask passed to page allocator
624 * 635 *
625 * If we run out of memory, we have the choice between either 636 * If we run out of memory, we have the choice between either
626 * killing a random task (bad), letting the system crash (worse) 637 * killing a random task (bad), letting the system crash (worse)
@@ -630,49 +641,93 @@ void pagefault_out_of_memory(void)
630void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, 641void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
631 int order, nodemask_t *nodemask) 642 int order, nodemask_t *nodemask)
632{ 643{
644 struct task_struct *p;
645 unsigned long totalpages;
633 unsigned long freed = 0; 646 unsigned long freed = 0;
634 enum oom_constraint constraint; 647 unsigned int points;
648 enum oom_constraint constraint = CONSTRAINT_NONE;
649 int killed = 0;
635 650
636 blocking_notifier_call_chain(&oom_notify_list, 0, &freed); 651 blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
637 if (freed > 0) 652 if (freed > 0)
638 /* Got some memory back in the last second. */ 653 /* Got some memory back in the last second. */
639 return; 654 return;
640 655
641 if (sysctl_panic_on_oom == 2) { 656 /*
642 dump_header(NULL, gfp_mask, order, NULL); 657 * If current has a pending SIGKILL, then automatically select it. The
643 panic("out of memory. Compulsory panic_on_oom is selected.\n"); 658 * goal is to allow it to allocate so that it may quickly exit and free
659 * its memory.
660 */
661 if (fatal_signal_pending(current)) {
662 set_thread_flag(TIF_MEMDIE);
663 boost_dying_task_prio(current, NULL);
664 return;
644 } 665 }
645 666
646 /* 667 /*
647 * Check if there were limitations on the allocation (only relevant for 668 * Check if there were limitations on the allocation (only relevant for
648 * NUMA) that may require different handling. 669 * NUMA) that may require different handling.
649 */ 670 */
650 constraint = constrained_alloc(zonelist, gfp_mask, nodemask); 671 constraint = constrained_alloc(zonelist, gfp_mask, nodemask,
672 &totalpages);
673 check_panic_on_oom(constraint, gfp_mask, order);
674
651 read_lock(&tasklist_lock); 675 read_lock(&tasklist_lock);
676 if (sysctl_oom_kill_allocating_task &&
677 !oom_unkillable_task(current, NULL, nodemask) &&
678 (current->signal->oom_adj != OOM_DISABLE)) {
679 /*
680 * oom_kill_process() needs tasklist_lock held. If it returns
681 * non-zero, current could not be killed so we must fallback to
682 * the tasklist scan.
683 */
684 if (!oom_kill_process(current, gfp_mask, order, 0, totalpages,
685 NULL, nodemask,
686 "Out of memory (oom_kill_allocating_task)"))
687 goto out;
688 }
652 689
653 switch (constraint) { 690retry:
654 case CONSTRAINT_MEMORY_POLICY: 691 p = select_bad_process(&points, totalpages, NULL,
655 oom_kill_process(current, gfp_mask, order, 0, NULL, 692 constraint == CONSTRAINT_MEMORY_POLICY ? nodemask :
656 "No available memory (MPOL_BIND)"); 693 NULL);
657 break; 694 if (PTR_ERR(p) == -1UL)
695 goto out;
658 696
659 case CONSTRAINT_NONE: 697 /* Found nothing?!?! Either we hang forever, or we panic. */
660 if (sysctl_panic_on_oom) { 698 if (!p) {
661 dump_header(NULL, gfp_mask, order, NULL); 699 dump_header(NULL, gfp_mask, order, NULL);
662 panic("out of memory. panic_on_oom is selected\n"); 700 read_unlock(&tasklist_lock);
663 } 701 panic("Out of memory and no killable processes...\n");
664 /* Fall-through */
665 case CONSTRAINT_CPUSET:
666 __out_of_memory(gfp_mask, order);
667 break;
668 } 702 }
669 703
704 if (oom_kill_process(p, gfp_mask, order, points, totalpages, NULL,
705 nodemask, "Out of memory"))
706 goto retry;
707 killed = 1;
708out:
670 read_unlock(&tasklist_lock); 709 read_unlock(&tasklist_lock);
671 710
672 /* 711 /*
673 * Give "p" a good chance of killing itself before we 712 * Give "p" a good chance of killing itself before we
674 * retry to allocate memory unless "p" is current 713 * retry to allocate memory unless "p" is current
675 */ 714 */
715 if (killed && !test_thread_flag(TIF_MEMDIE))
716 schedule_timeout_uninterruptible(1);
717}
718
719/*
720 * The pagefault handler calls here because it is out of memory, so kill a
721 * memory-hogging task. If a populated zone has ZONE_OOM_LOCKED set, a parallel
722 * oom killing is already in progress so do nothing. If a task is found with
723 * TIF_MEMDIE set, it has been killed so do nothing and allow it to exit.
724 */
725void pagefault_out_of_memory(void)
726{
727 if (try_set_system_oom()) {
728 out_of_memory(NULL, 0, 0, NULL);
729 clear_system_oom();
730 }
676 if (!test_thread_flag(TIF_MEMDIE)) 731 if (!test_thread_flag(TIF_MEMDIE))
677 schedule_timeout_uninterruptible(1); 732 schedule_timeout_uninterruptible(1);
678} 733}
generated by cgit v1.2.2 (git 2.25.0) at 2025-03-07 10:51:31 -0500