Merge tag 'for-linus' of git://git.kernel.org/pub/scm/virt/kvm/kvm
[linux-drm-fsl-dcu.git] / mm / oom_kill.c
index dff991e0681e85a5308ba097533e346e621bc954..1ecc0bcaecc518458765347f8b0fa5d5eed46f75 100644 (file)
@@ -196,27 +196,26 @@ unsigned long oom_badness(struct task_struct *p, struct mem_cgroup *memcg,
  * Determine the type of allocation constraint.
  */
 #ifdef CONFIG_NUMA
-static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
-                               gfp_t gfp_mask, nodemask_t *nodemask,
-                               unsigned long *totalpages)
+static enum oom_constraint constrained_alloc(struct oom_control *oc,
+                                            unsigned long *totalpages)
 {
        struct zone *zone;
        struct zoneref *z;
-       enum zone_type high_zoneidx = gfp_zone(gfp_mask);
+       enum zone_type high_zoneidx = gfp_zone(oc->gfp_mask);
        bool cpuset_limited = false;
        int nid;
 
        /* Default to all available memory */
        *totalpages = totalram_pages + total_swap_pages;
 
-       if (!zonelist)
+       if (!oc->zonelist)
                return CONSTRAINT_NONE;
        /*
         * Reach here only when __GFP_NOFAIL is used. So, we should avoid
         * to kill current.We have to random task kill in this case.
         * Hopefully, CONSTRAINT_THISNODE...but no way to handle it, now.
         */
-       if (gfp_mask & __GFP_THISNODE)
+       if (oc->gfp_mask & __GFP_THISNODE)
                return CONSTRAINT_NONE;
 
        /*
@@ -224,17 +223,18 @@ static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
         * the page allocator means a mempolicy is in effect.  Cpuset policy
         * is enforced in get_page_from_freelist().
         */
-       if (nodemask && !nodes_subset(node_states[N_MEMORY], *nodemask)) {
+       if (oc->nodemask &&
+           !nodes_subset(node_states[N_MEMORY], *oc->nodemask)) {
                *totalpages = total_swap_pages;
-               for_each_node_mask(nid, *nodemask)
+               for_each_node_mask(nid, *oc->nodemask)
                        *totalpages += node_spanned_pages(nid);
                return CONSTRAINT_MEMORY_POLICY;
        }
 
        /* Check this allocation failure is caused by cpuset's wall function */
-       for_each_zone_zonelist_nodemask(zone, z, zonelist,
-                       high_zoneidx, nodemask)
-               if (!cpuset_zone_allowed(zone, gfp_mask))
+       for_each_zone_zonelist_nodemask(zone, z, oc->zonelist,
+                       high_zoneidx, oc->nodemask)
+               if (!cpuset_zone_allowed(zone, oc->gfp_mask))
                        cpuset_limited = true;
 
        if (cpuset_limited) {
@@ -246,20 +246,18 @@ static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
        return CONSTRAINT_NONE;
 }
 #else
-static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
-                               gfp_t gfp_mask, nodemask_t *nodemask,
-                               unsigned long *totalpages)
+static enum oom_constraint constrained_alloc(struct oom_control *oc,
+                                            unsigned long *totalpages)
 {
        *totalpages = totalram_pages + total_swap_pages;
        return CONSTRAINT_NONE;
 }
 #endif
 
-enum oom_scan_t oom_scan_process_thread(struct task_struct *task,
-               unsigned long totalpages, const nodemask_t *nodemask,
-               bool force_kill)
+enum oom_scan_t oom_scan_process_thread(struct oom_control *oc,
+                       struct task_struct *task, unsigned long totalpages)
 {
-       if (oom_unkillable_task(task, NULL, nodemask))
+       if (oom_unkillable_task(task, NULL, oc->nodemask))
                return OOM_SCAN_CONTINUE;
 
        /*
@@ -267,7 +265,7 @@ enum oom_scan_t oom_scan_process_thread(struct task_struct *task,
         * Don't allow any other task to have access to the reserves.
         */
        if (test_tsk_thread_flag(task, TIF_MEMDIE)) {
-               if (!force_kill)
+               if (oc->order != -1)
                        return OOM_SCAN_ABORT;
        }
        if (!task->mm)
@@ -280,7 +278,7 @@ enum oom_scan_t oom_scan_process_thread(struct task_struct *task,
        if (oom_task_origin(task))
                return OOM_SCAN_SELECT;
 
-       if (task_will_free_mem(task) && !force_kill)
+       if (task_will_free_mem(task) && oc->order != -1)
                return OOM_SCAN_ABORT;
 
        return OOM_SCAN_OK;
@@ -289,12 +287,9 @@ enum oom_scan_t oom_scan_process_thread(struct task_struct *task,
 /*
  * Simple selection loop. We chose the process with the highest
  * number of 'points'.  Returns -1 on scan abort.
- *
- * (not docbooked, we don't want this one cluttering up the manual)
  */
-static struct task_struct *select_bad_process(unsigned int *ppoints,
-               unsigned long totalpages, const nodemask_t *nodemask,
-               bool force_kill)
+static struct task_struct *select_bad_process(struct oom_control *oc,
+               unsigned int *ppoints, unsigned long totalpages)
 {
        struct task_struct *g, *p;
        struct task_struct *chosen = NULL;
@@ -304,8 +299,7 @@ static struct task_struct *select_bad_process(unsigned int *ppoints,
        for_each_process_thread(g, p) {
                unsigned int points;
 
-               switch (oom_scan_process_thread(p, totalpages, nodemask,
-                                               force_kill)) {
+               switch (oom_scan_process_thread(oc, p, totalpages)) {
                case OOM_SCAN_SELECT:
                        chosen = p;
                        chosen_points = ULONG_MAX;
@@ -318,7 +312,7 @@ static struct task_struct *select_bad_process(unsigned int *ppoints,
                case OOM_SCAN_OK:
                        break;
                };
-               points = oom_badness(p, NULL, nodemask, totalpages);
+               points = oom_badness(p, NULL, oc->nodemask, totalpages);
                if (!points || points < chosen_points)
                        continue;
                /* Prefer thread group leaders for display purposes */
@@ -380,13 +374,13 @@ static void dump_tasks(struct mem_cgroup *memcg, const nodemask_t *nodemask)
        rcu_read_unlock();
 }
 
-static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
-                       struct mem_cgroup *memcg, const nodemask_t *nodemask)
+static void dump_header(struct oom_control *oc, struct task_struct *p,
+                       struct mem_cgroup *memcg)
 {
        task_lock(current);
        pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
                "oom_score_adj=%hd\n",
-               current->comm, gfp_mask, order,
+               current->comm, oc->gfp_mask, oc->order,
                current->signal->oom_score_adj);
        cpuset_print_task_mems_allowed(current);
        task_unlock(current);
@@ -396,7 +390,7 @@ static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
        else
                show_mem(SHOW_MEM_FILTER_NODES);
        if (sysctl_oom_dump_tasks)
-               dump_tasks(memcg, nodemask);
+               dump_tasks(memcg, oc->nodemask);
 }
 
 /*
@@ -487,10 +481,9 @@ void oom_killer_enable(void)
  * Must be called while holding a reference to p, which will be released upon
  * returning.
  */
-void oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
+void oom_kill_process(struct oom_control *oc, struct task_struct *p,
                      unsigned int points, unsigned long totalpages,
-                     struct mem_cgroup *memcg, nodemask_t *nodemask,
-                     const char *message)
+                     struct mem_cgroup *memcg, const char *message)
 {
        struct task_struct *victim = p;
        struct task_struct *child;
@@ -514,7 +507,7 @@ void oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
        task_unlock(p);
 
        if (__ratelimit(&oom_rs))
-               dump_header(p, gfp_mask, order, memcg, nodemask);
+               dump_header(oc, p, memcg);
 
        task_lock(p);
        pr_err("%s: Kill process %d (%s) score %u or sacrifice child\n",
@@ -537,7 +530,7 @@ void oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
                        /*
                         * oom_badness() returns 0 if the thread is unkillable
                         */
-                       child_points = oom_badness(child, memcg, nodemask,
+                       child_points = oom_badness(child, memcg, oc->nodemask,
                                                                totalpages);
                        if (child_points > victim_points) {
                                put_task_struct(victim);
@@ -600,8 +593,7 @@ void 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.
  */
-void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
-                       int order, const nodemask_t *nodemask,
+void check_panic_on_oom(struct oom_control *oc, enum oom_constraint constraint,
                        struct mem_cgroup *memcg)
 {
        if (likely(!sysctl_panic_on_oom))
@@ -615,7 +607,10 @@ void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
                if (constraint != CONSTRAINT_NONE)
                        return;
        }
-       dump_header(NULL, gfp_mask, order, memcg, nodemask);
+       /* Do not panic for oom kills triggered by sysrq */
+       if (oc->order == -1)
+               return;
+       dump_header(oc, NULL, memcg);
        panic("Out of memory: %s panic_on_oom is enabled\n",
                sysctl_panic_on_oom == 2 ? "compulsory" : "system-wide");
 }
@@ -635,28 +630,21 @@ int unregister_oom_notifier(struct notifier_block *nb)
 EXPORT_SYMBOL_GPL(unregister_oom_notifier);
 
 /**
- * __out_of_memory - kill the "best" process when we run out of memory
- * @zonelist: zonelist pointer
- * @gfp_mask: memory allocation flags
- * @order: amount of memory being requested as a power of 2
- * @nodemask: nodemask passed to page allocator
- * @force_kill: true if a task must be killed, even if others are exiting
+ * out_of_memory - kill the "best" process when we run out of memory
+ * @oc: pointer to struct oom_control
  *
  * If we run out of memory, we have the choice between either
  * killing a random task (bad), letting the system crash (worse)
  * OR try to be smart about which process to kill. Note that we
  * don't have to be perfect here, we just have to be good.
  */
-bool out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
-                  int order, nodemask_t *nodemask, bool force_kill)
+bool out_of_memory(struct oom_control *oc)
 {
-       const nodemask_t *mpol_mask;
        struct task_struct *p;
        unsigned long totalpages;
        unsigned long freed = 0;
        unsigned int uninitialized_var(points);
        enum oom_constraint constraint = CONSTRAINT_NONE;
-       int killed = 0;
 
        if (oom_killer_disabled)
                return false;
@@ -664,7 +652,7 @@ bool out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
        blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
        if (freed > 0)
                /* Got some memory back in the last second. */
-               goto out;
+               return true;
 
        /*
         * If current has a pending SIGKILL or is exiting, then automatically
@@ -677,47 +665,42 @@ bool out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
        if (current->mm &&
            (fatal_signal_pending(current) || task_will_free_mem(current))) {
                mark_oom_victim(current);
-               goto out;
+               return true;
        }
 
        /*
         * Check if there were limitations on the allocation (only relevant for
         * NUMA) that may require different handling.
         */
-       constraint = constrained_alloc(zonelist, gfp_mask, nodemask,
-                                               &totalpages);
-       mpol_mask = (constraint == CONSTRAINT_MEMORY_POLICY) ? nodemask : NULL;
-       check_panic_on_oom(constraint, gfp_mask, order, mpol_mask, NULL);
+       constraint = constrained_alloc(oc, &totalpages);
+       if (constraint != CONSTRAINT_MEMORY_POLICY)
+               oc->nodemask = NULL;
+       check_panic_on_oom(oc, constraint, NULL);
 
        if (sysctl_oom_kill_allocating_task && current->mm &&
-           !oom_unkillable_task(current, NULL, nodemask) &&
+           !oom_unkillable_task(current, NULL, oc->nodemask) &&
            current->signal->oom_score_adj != OOM_SCORE_ADJ_MIN) {
                get_task_struct(current);
-               oom_kill_process(current, gfp_mask, order, 0, totalpages, NULL,
-                                nodemask,
+               oom_kill_process(oc, current, 0, totalpages, NULL,
                                 "Out of memory (oom_kill_allocating_task)");
-               goto out;
+               return true;
        }
 
-       p = select_bad_process(&points, totalpages, mpol_mask, force_kill);
+       p = select_bad_process(oc, &points, totalpages);
        /* Found nothing?!?! Either we hang forever, or we panic. */
-       if (!p) {
-               dump_header(NULL, gfp_mask, order, NULL, mpol_mask);
+       if (!p && oc->order != -1) {
+               dump_header(oc, NULL, NULL);
                panic("Out of memory and no killable processes...\n");
        }
-       if (p != (void *)-1UL) {
-               oom_kill_process(p, gfp_mask, order, points, totalpages, NULL,
-                                nodemask, "Out of memory");
-               killed = 1;
-       }
-out:
-       /*
-        * Give the killed threads a good chance of exiting before trying to
-        * allocate memory again.
-        */
-       if (killed)
+       if (p && p != (void *)-1UL) {
+               oom_kill_process(oc, p, points, totalpages, NULL,
+                                "Out of memory");
+               /*
+                * Give the killed process a good chance to exit before trying
+                * to allocate memory again.
+                */
                schedule_timeout_killable(1);
-
+       }
        return true;
 }
 
@@ -728,13 +711,20 @@ out:
  */
 void pagefault_out_of_memory(void)
 {
+       struct oom_control oc = {
+               .zonelist = NULL,
+               .nodemask = NULL,
+               .gfp_mask = 0,
+               .order = 0,
+       };
+
        if (mem_cgroup_oom_synchronize(true))
                return;
 
        if (!mutex_trylock(&oom_lock))
                return;
 
-       if (!out_of_memory(NULL, 0, 0, NULL, false)) {
+       if (!out_of_memory(&oc)) {
                /*
                 * There shouldn't be any user tasks runnable while the
                 * OOM killer is disabled, so the current task has to