readahead: return early when readahead is disabled
[linux-2.6.git] / mm / oom_kill.c
index c48c5ef..e4b0991 100644 (file)
 #include <linux/memcontrol.h>
 #include <linux/mempolicy.h>
 #include <linux/security.h>
+#include <linux/ptrace.h>
 
 int sysctl_panic_on_oom;
 int sysctl_oom_kill_allocating_task;
 int sysctl_oom_dump_tasks = 1;
 static DEFINE_SPINLOCK(zone_scan_lock);
 
+/**
+ * test_set_oom_score_adj() - set current's oom_score_adj and return old value
+ * @new_val: new oom_score_adj value
+ *
+ * Sets the oom_score_adj value for current to @new_val with proper
+ * synchronization and returns the old value.  Usually used to temporarily
+ * set a value, save the old value in the caller, and then reinstate it later.
+ */
+int test_set_oom_score_adj(int new_val)
+{
+       struct sighand_struct *sighand = current->sighand;
+       int old_val;
+
+       spin_lock_irq(&sighand->siglock);
+       old_val = current->signal->oom_score_adj;
+       if (new_val != old_val) {
+               if (new_val == OOM_SCORE_ADJ_MIN)
+                       atomic_inc(&current->mm->oom_disable_count);
+               else if (old_val == OOM_SCORE_ADJ_MIN)
+                       atomic_dec(&current->mm->oom_disable_count);
+               current->signal->oom_score_adj = new_val;
+       }
+       spin_unlock_irq(&sighand->siglock);
+
+       return old_val;
+}
+
 #ifdef CONFIG_NUMA
 /**
  * has_intersects_mems_allowed() - check task eligiblity for kill
@@ -83,24 +111,6 @@ static bool has_intersects_mems_allowed(struct task_struct *tsk,
 #endif /* CONFIG_NUMA */
 
 /*
- * If this is a system OOM (not a memcg OOM) and the task selected to be
- * killed is not already running at high (RT) priorities, speed up the
- * recovery by boosting the dying task to the lowest FIFO priority.
- * That helps with the recovery and avoids interfering with RT tasks.
- */
-static void boost_dying_task_prio(struct task_struct *p,
-                                 struct mem_cgroup *mem)
-{
-       struct sched_param param = { .sched_priority = 1 };
-
-       if (mem)
-               return;
-
-       if (!rt_task(p))
-               sched_setscheduler_nocheck(p, SCHED_FIFO, &param);
-}
-
-/*
  * The process p may have detached its own ->mm while exiting or through
  * use_mm(), but one or more of its subthreads may still have a valid
  * pointer.  Return p, or any of its subthreads with a valid ->mm, with
@@ -121,8 +131,8 @@ struct task_struct *find_lock_task_mm(struct task_struct *p)
 }
 
 /* return true if the task is not adequate as candidate victim task. */
-static bool oom_unkillable_task(struct task_struct *p, struct mem_cgroup *mem,
-                          const nodemask_t *nodemask)
+static bool oom_unkillable_task(struct task_struct *p,
+               const struct mem_cgroup *mem, const nodemask_t *nodemask)
 {
        if (is_global_init(p))
                return true;
@@ -162,24 +172,16 @@ unsigned int oom_badness(struct task_struct *p, struct mem_cgroup *mem,
                return 0;
 
        /*
-        * Shortcut check for OOM_SCORE_ADJ_MIN so the entire heuristic doesn't
-        * need to be executed for something that cannot be killed.
+        * Shortcut check for a thread sharing p->mm that is OOM_SCORE_ADJ_MIN
+        * so the entire heuristic doesn't need to be executed for something
+        * that cannot be killed.
         */
-       if (p->signal->oom_score_adj == OOM_SCORE_ADJ_MIN) {
+       if (atomic_read(&p->mm->oom_disable_count)) {
                task_unlock(p);
                return 0;
        }
 
        /*
-        * When the PF_OOM_ORIGIN bit is set, it indicates the task should have
-        * priority for oom killing.
-        */
-       if (p->flags & PF_OOM_ORIGIN) {
-               task_unlock(p);
-               return 1000;
-       }
-
-       /*
         * The memory controller may have a limit of 0 bytes, so avoid a divide
         * by zero, if necessary.
         */
@@ -188,10 +190,13 @@ unsigned int oom_badness(struct task_struct *p, struct mem_cgroup *mem,
 
        /*
         * The baseline for the badness score is the proportion of RAM that each
-        * task's rss and swap space use.
+        * task's rss, pagetable and swap space use.
         */
-       points = (get_mm_rss(p->mm) + get_mm_counter(p->mm, MM_SWAPENTS)) * 1000 /
-                       totalpages;
+       points = get_mm_rss(p->mm) + p->mm->nr_ptes;
+       points += get_mm_counter(p->mm, MM_SWAPENTS);
+
+       points *= 1000;
+       points /= totalpages;
        task_unlock(p);
 
        /*
@@ -208,8 +213,13 @@ unsigned int oom_badness(struct task_struct *p, struct mem_cgroup *mem,
         */
        points += p->signal->oom_score_adj;
 
-       if (points < 0)
-               return 0;
+       /*
+        * Never return 0 for an eligible task that may be killed since it's
+        * possible that no single user task uses more than 0.1% of memory and
+        * no single admin tasks uses more than 3.0%.
+        */
+       if (points <= 0)
+               return 1;
        return (points < 1000) ? points : 1000;
 }
 
@@ -286,13 +296,15 @@ static struct task_struct *select_bad_process(unsigned int *ppoints,
                unsigned long totalpages, struct mem_cgroup *mem,
                const nodemask_t *nodemask)
 {
-       struct task_struct *p;
+       struct task_struct *g, *p;
        struct task_struct *chosen = NULL;
        *ppoints = 0;
 
-       for_each_process(p) {
+       do_each_thread(g, p) {
                unsigned int points;
 
+               if (!p->mm)
+                       continue;
                if (oom_unkillable_task(p, mem, nodemask))
                        continue;
 
@@ -308,22 +320,29 @@ static struct task_struct *select_bad_process(unsigned int *ppoints,
                if (test_tsk_thread_flag(p, TIF_MEMDIE))
                        return ERR_PTR(-1UL);
 
-               /*
-                * This is in the process of releasing memory so wait for it
-                * to finish before killing some other task by mistake.
-                *
-                * However, if p is the current task, we allow the 'kill' to
-                * go ahead if it is exiting: this will simply set TIF_MEMDIE,
-                * which will allow it to gain access to memory reserves in
-                * the process of exiting and releasing its resources.
-                * Otherwise we could get an easy OOM deadlock.
-                */
-               if (thread_group_empty(p) && (p->flags & PF_EXITING) && p->mm) {
-                       if (p != current)
-                               return ERR_PTR(-1UL);
-
-                       chosen = p;
-                       *ppoints = 1000;
+               if (p->flags & PF_EXITING) {
+                       /*
+                        * If p is the current task and is in the process of
+                        * releasing memory, we allow the "kill" to set
+                        * TIF_MEMDIE, which will allow it to gain access to
+                        * memory reserves.  Otherwise, it may stall forever.
+                        *
+                        * The loop isn't broken here, however, in case other
+                        * threads are found to have already been oom killed.
+                        */
+                       if (p == current) {
+                               chosen = p;
+                               *ppoints = 1000;
+                       } else {
+                               /*
+                                * If this task is not being ptraced on exit,
+                                * then wait for it to finish before killing
+                                * some other task unnecessarily.
+                                */
+                               if (!(task_ptrace(p->group_leader) &
+                                                       PT_TRACE_EXIT))
+                                       return ERR_PTR(-1UL);
+                       }
                }
 
                points = oom_badness(p, mem, nodemask, totalpages);
@@ -331,7 +350,7 @@ static struct task_struct *select_bad_process(unsigned int *ppoints,
                        chosen = p;
                        *ppoints = points;
                }
-       }
+       } while_each_thread(g, p);
 
        return chosen;
 }
@@ -339,26 +358,24 @@ static struct task_struct *select_bad_process(unsigned int *ppoints,
 /**
  * dump_tasks - dump current memory state of all system tasks
  * @mem: current's memory controller, if constrained
+ * @nodemask: nodemask passed to page allocator for mempolicy ooms
  *
- * Dumps the current memory state of all system tasks, excluding kernel threads.
+ * Dumps the current memory state of all eligible tasks.  Tasks not in the same
+ * memcg, not in the same cpuset, or bound to a disjoint set of mempolicy nodes
+ * are not shown.
  * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
  * value, oom_score_adj value, and name.
  *
- * If the actual is non-NULL, only tasks that are a member of the mem_cgroup are
- * shown.
- *
  * Call with tasklist_lock read-locked.
  */
-static void dump_tasks(const struct mem_cgroup *mem)
+static void dump_tasks(const struct mem_cgroup *mem, const nodemask_t *nodemask)
 {
        struct task_struct *p;
        struct task_struct *task;
 
        pr_info("[ pid ]   uid  tgid total_vm      rss cpu oom_adj oom_score_adj name\n");
        for_each_process(p) {
-               if (p->flags & PF_KTHREAD)
-                       continue;
-               if (mem && !task_in_mem_cgroup(p, mem))
+               if (oom_unkillable_task(p, mem, nodemask))
                        continue;
 
                task = find_lock_task_mm(p);
@@ -372,7 +389,7 @@ static void dump_tasks(const struct mem_cgroup *mem)
                }
 
                pr_info("[%5d] %5d %5d %8lu %8lu %3u     %3d         %5d %s\n",
-                       task->pid, __task_cred(task)->uid, task->tgid,
+                       task->pid, task_uid(task), task->tgid,
                        task->mm->total_vm, get_mm_rss(task->mm),
                        task_cpu(task), task->signal->oom_adj,
                        task->signal->oom_score_adj, task->comm);
@@ -381,7 +398,7 @@ static void dump_tasks(const struct mem_cgroup *mem)
 }
 
 static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
-                                                       struct mem_cgroup *mem)
+                       struct mem_cgroup *mem, const nodemask_t *nodemask)
 {
        task_lock(current);
        pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
@@ -392,35 +409,52 @@ static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
        task_unlock(current);
        dump_stack();
        mem_cgroup_print_oom_info(mem, p);
-       show_mem();
+       show_mem(SHOW_MEM_FILTER_NODES);
        if (sysctl_oom_dump_tasks)
-               dump_tasks(mem);
+               dump_tasks(mem, nodemask);
 }
 
 #define K(x) ((x) << (PAGE_SHIFT-10))
 static int oom_kill_task(struct task_struct *p, struct mem_cgroup *mem)
 {
+       struct task_struct *q;
+       struct mm_struct *mm;
+
        p = find_lock_task_mm(p);
        if (!p)
                return 1;
 
+       /* mm cannot be safely dereferenced after task_unlock(p) */
+       mm = p->mm;
+
        pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
                task_pid_nr(p), p->comm, K(p->mm->total_vm),
                K(get_mm_counter(p->mm, MM_ANONPAGES)),
                K(get_mm_counter(p->mm, MM_FILEPAGES)));
        task_unlock(p);
 
+       /*
+        * Kill all processes sharing p->mm in other thread groups, if any.
+        * They don't get access to memory reserves or a higher scheduler
+        * priority, though, to avoid depletion of all memory or task
+        * starvation.  This prevents mm->mmap_sem livelock when an oom killed
+        * task cannot exit because it requires the semaphore and its contended
+        * by another thread trying to allocate memory itself.  That thread will
+        * now get access to memory reserves since it has a pending fatal
+        * signal.
+        */
+       for_each_process(q)
+               if (q->mm == mm && !same_thread_group(q, p)) {
+                       task_lock(q);   /* Protect ->comm from prctl() */
+                       pr_err("Kill process %d (%s) sharing same memory\n",
+                               task_pid_nr(q), q->comm);
+                       task_unlock(q);
+                       force_sig(SIGKILL, q);
+               }
 
        set_tsk_thread_flag(p, TIF_MEMDIE);
        force_sig(SIGKILL, p);
 
-       /*
-        * We give our sacrificial lamb high priority and access to
-        * all the memory it needs. That way it should be able to
-        * exit() and clear out its resources quickly...
-        */
-       boost_dying_task_prio(p, mem);
-
        return 0;
 }
 #undef K
@@ -436,7 +470,7 @@ static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
        unsigned int victim_points = 0;
 
        if (printk_ratelimit())
-               dump_header(p, gfp_mask, order, mem);
+               dump_header(p, gfp_mask, order, mem, nodemask);
 
        /*
         * If the task is already exiting, don't alarm the sysadmin or kill
@@ -444,7 +478,6 @@ static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
         */
        if (p->flags & PF_EXITING) {
                set_tsk_thread_flag(p, TIF_MEMDIE);
-               boost_dying_task_prio(p, mem);
                return 0;
        }
 
@@ -463,6 +496,8 @@ static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
                list_for_each_entry(child, &t->children, sibling) {
                        unsigned int child_points;
 
+                       if (child->mm == p->mm)
+                               continue;
                        /*
                         * oom_badness() returns 0 if the thread is unkillable
                         */
@@ -482,7 +517,7 @@ static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
  * Determines whether the kernel must panic because of the panic_on_oom sysctl.
  */
 static void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
-                               int order)
+                               int order, const nodemask_t *nodemask)
 {
        if (likely(!sysctl_panic_on_oom))
                return;
@@ -496,7 +531,7 @@ static void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
                        return;
        }
        read_lock(&tasklist_lock);
-       dump_header(NULL, gfp_mask, order, NULL);
+       dump_header(NULL, gfp_mask, order, NULL, nodemask);
        read_unlock(&tasklist_lock);
        panic("Out of memory: %s panic_on_oom is enabled\n",
                sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
@@ -509,7 +544,17 @@ void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
        unsigned int points = 0;
        struct task_struct *p;
 
-       check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, 0);
+       /*
+        * If current has a pending SIGKILL, then automatically select it.  The
+        * goal is to allow it to allocate so that it may quickly exit and free
+        * its memory.
+        */
+       if (fatal_signal_pending(current)) {
+               set_thread_flag(TIF_MEMDIE);
+               return;
+       }
+
+       check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, 0, NULL);
        limit = mem_cgroup_get_limit(mem) >> PAGE_SHIFT;
        read_lock(&tasklist_lock);
 retry:
@@ -641,6 +686,7 @@ static void clear_system_oom(void)
 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
                int order, nodemask_t *nodemask)
 {
+       const nodemask_t *mpol_mask;
        struct task_struct *p;
        unsigned long totalpages;
        unsigned long freed = 0;
@@ -660,7 +706,6 @@ void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
         */
        if (fatal_signal_pending(current)) {
                set_thread_flag(TIF_MEMDIE);
-               boost_dying_task_prio(current, NULL);
                return;
        }
 
@@ -670,12 +715,13 @@ void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
         */
        constraint = constrained_alloc(zonelist, gfp_mask, nodemask,
                                                &totalpages);
-       check_panic_on_oom(constraint, gfp_mask, order);
+       mpol_mask = (constraint == CONSTRAINT_MEMORY_POLICY) ? nodemask : NULL;
+       check_panic_on_oom(constraint, gfp_mask, order, mpol_mask);
 
        read_lock(&tasklist_lock);
        if (sysctl_oom_kill_allocating_task &&
            !oom_unkillable_task(current, NULL, nodemask) &&
-           (current->signal->oom_adj != OOM_DISABLE)) {
+           current->mm && !atomic_read(&current->mm->oom_disable_count)) {
                /*
                 * oom_kill_process() needs tasklist_lock held.  If it returns
                 * non-zero, current could not be killed so we must fallback to
@@ -688,15 +734,13 @@ void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
        }
 
 retry:
-       p = select_bad_process(&points, totalpages, NULL,
-                       constraint == CONSTRAINT_MEMORY_POLICY ? nodemask :
-                                                                NULL);
+       p = select_bad_process(&points, totalpages, NULL, mpol_mask);
        if (PTR_ERR(p) == -1UL)
                goto out;
 
        /* Found nothing?!?! Either we hang forever, or we panic. */
        if (!p) {
-               dump_header(NULL, gfp_mask, order, NULL);
+               dump_header(NULL, gfp_mask, order, NULL, mpol_mask);
                read_unlock(&tasklist_lock);
                panic("Out of memory and no killable processes...\n");
        }