mm: nommu: find vma using the sorted vma list
[linux-2.6.git] / mm / oom_kill.c
index ef5084d..f52e85c 100644 (file)
@@ -4,6 +4,8 @@
  *  Copyright (C)  1998,2000  Rik van Riel
  *     Thanks go out to Claus Fischer for some serious inspiration and
  *     for goading me into coding this file...
+ *  Copyright (C)  2010  Google, Inc.
+ *     Rewritten by David Rientjes
  *
  *  The routines in this file are used to kill a process when
  *  we're seriously out of memory. This gets called from __alloc_pages()
@@ -18,6 +20,7 @@
 #include <linux/oom.h>
 #include <linux/mm.h>
 #include <linux/err.h>
+#include <linux/gfp.h>
 #include <linux/sched.h>
 #include <linux/swap.h>
 #include <linux/timex.h>
 #include <linux/module.h>
 #include <linux/notifier.h>
 #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;
-static DEFINE_SPINLOCK(zone_scan_mutex);
-/* #define DEBUG */
+int sysctl_oom_dump_tasks = 1;
+static DEFINE_SPINLOCK(zone_scan_lock);
 
+#ifdef CONFIG_NUMA
 /**
- * badness - calculate a numeric value for how bad this task has been
- * @p: task struct of which task we should calculate
- * @uptime: current uptime in seconds
+ * has_intersects_mems_allowed() - check task eligiblity for kill
+ * @tsk: task struct of which task to consider
+ * @mask: nodemask passed to page allocator for mempolicy ooms
  *
- * The formula used is relatively simple and documented inline in the
- * function. The main rationale is that we want to select a good task
- * to kill when we run out of memory.
- *
- * Good in this context means that:
- * 1) we lose the minimum amount of work done
- * 2) we recover a large amount of memory
- * 3) we don't kill anything innocent of eating tons of memory
- * 4) we want to kill the minimum amount of processes (one)
- * 5) we try to kill the process the user expects us to kill, this
- *    algorithm has been meticulously tuned to meet the principle
- *    of least surprise ... (be careful when you change it)
+ * Task eligibility is determined by whether or not a candidate task, @tsk,
+ * shares the same mempolicy nodes as current if it is bound by such a policy
+ * and whether or not it has the same set of allowed cpuset nodes.
  */
+static bool has_intersects_mems_allowed(struct task_struct *tsk,
+                                       const nodemask_t *mask)
+{
+       struct task_struct *start = tsk;
+
+       do {
+               if (mask) {
+                       /*
+                        * If this is a mempolicy constrained oom, tsk's
+                        * cpuset is irrelevant.  Only return true if its
+                        * mempolicy intersects current, otherwise it may be
+                        * needlessly killed.
+                        */
+                       if (mempolicy_nodemask_intersects(tsk, mask))
+                               return true;
+               } else {
+                       /*
+                        * This is not a mempolicy constrained oom, so only
+                        * check the mems of tsk's cpuset.
+                        */
+                       if (cpuset_mems_allowed_intersects(current, tsk))
+                               return true;
+               }
+       } while_each_thread(start, tsk);
 
-unsigned long badness(struct task_struct *p, unsigned long uptime,
-                       struct mem_cgroup *mem)
+       return false;
+}
+#else
+static bool has_intersects_mems_allowed(struct task_struct *tsk,
+                                       const nodemask_t *mask)
 {
-       unsigned long points, cpu_time, run_time, s;
-       struct mm_struct *mm;
-       struct task_struct *child;
+       return true;
+}
+#endif /* CONFIG_NUMA */
 
-       task_lock(p);
-       mm = p->mm;
-       if (!mm) {
-               task_unlock(p);
-               return 0;
-       }
+/*
+ * 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
+ * task_lock() held.
+ */
+struct task_struct *find_lock_task_mm(struct task_struct *p)
+{
+       struct task_struct *t = p;
 
-       /*
-        * The memory size of the process is the basis for the badness.
-        */
-       points = mm->total_vm;
+       do {
+               task_lock(t);
+               if (likely(t->mm))
+                       return t;
+               task_unlock(t);
+       } while_each_thread(p, t);
 
-       /*
-        * After this unlock we can no longer dereference local variable `mm'
-        */
-       task_unlock(p);
+       return NULL;
+}
 
-       /*
-        * swapoff can easily use up all memory, so kill those first.
-        */
-       if (p->flags & PF_SWAPOFF)
-               return ULONG_MAX;
+/* return true if the task is not adequate as candidate victim task. */
+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;
+       if (p->flags & PF_KTHREAD)
+               return true;
+
+       /* When mem_cgroup_out_of_memory() and p is not member of the group */
+       if (mem && !task_in_mem_cgroup(p, mem))
+               return true;
+
+       /* p may not have freeable memory in nodemask */
+       if (!has_intersects_mems_allowed(p, nodemask))
+               return true;
+
+       return false;
+}
+
+/**
+ * oom_badness - heuristic function to determine which candidate task to kill
+ * @p: task struct of which task we should calculate
+ * @totalpages: total present RAM allowed for page allocation
+ *
+ * The heuristic for determining which task to kill is made to be as simple and
+ * predictable as possible.  The goal is to return the highest value for the
+ * task consuming the most memory to avoid subsequent oom failures.
+ */
+unsigned int oom_badness(struct task_struct *p, struct mem_cgroup *mem,
+                     const nodemask_t *nodemask, unsigned long totalpages)
+{
+       int points;
+
+       if (oom_unkillable_task(p, mem, nodemask))
+               return 0;
+
+       p = find_lock_task_mm(p);
+       if (!p)
+               return 0;
 
        /*
-        * Processes which fork a lot of child processes are likely
-        * a good choice. We add half the vmsize of the children if they
-        * have an own mm. This prevents forking servers to flood the
-        * machine with an endless amount of children. In case a single
-        * child is eating the vast majority of memory, adding only half
-        * to the parents will make the child our kill candidate of choice.
+        * 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.
         */
-       list_for_each_entry(child, &p->children, sibling) {
-               task_lock(child);
-               if (child->mm != mm && child->mm)
-                       points += child->mm->total_vm/2 + 1;
-               task_unlock(child);
+       if (atomic_read(&p->mm->oom_disable_count)) {
+               task_unlock(p);
+               return 0;
        }
 
        /*
-        * CPU time is in tens of seconds and run time is in thousands
-         * of seconds. There is no particular reason for this other than
-         * that it turned out to work very well in practice.
+        * When the PF_OOM_ORIGIN bit is set, it indicates the task should have
+        * priority for oom killing.
         */
-       cpu_time = (cputime_to_jiffies(p->utime) + cputime_to_jiffies(p->stime))
-               >> (SHIFT_HZ + 3);
-
-       if (uptime >= p->start_time.tv_sec)
-               run_time = (uptime - p->start_time.tv_sec) >> 10;
-       else
-               run_time = 0;
-
-       s = int_sqrt(cpu_time);
-       if (s)
-               points /= s;
-       s = int_sqrt(int_sqrt(run_time));
-       if (s)
-               points /= s;
+       if (p->flags & PF_OOM_ORIGIN) {
+               task_unlock(p);
+               return 1000;
+       }
 
        /*
-        * Niced processes are most likely less important, so double
-        * their badness points.
+        * The memory controller may have a limit of 0 bytes, so avoid a divide
+        * by zero, if necessary.
         */
-       if (task_nice(p) > 0)
-               points *= 2;
+       if (!totalpages)
+               totalpages = 1;
 
        /*
-        * Superuser processes are usually more important, so we make it
-        * less likely that we kill those.
+        * The baseline for the badness score is the proportion of RAM that each
+        * task's rss, pagetable and swap space use.
         */
-       if (__capable(p, CAP_SYS_ADMIN) || __capable(p, CAP_SYS_RESOURCE))
-               points /= 4;
+       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);
 
        /*
-        * We don't want to kill a process with direct hardware access.
-        * Not only could that mess up the hardware, but usually users
-        * tend to only have this flag set on applications they think
-        * of as important.
+        * Root processes get 3% bonus, just like the __vm_enough_memory()
+        * implementation used by LSMs.
         */
-       if (__capable(p, CAP_SYS_RAWIO))
-               points /= 4;
+       if (has_capability_noaudit(p, CAP_SYS_ADMIN))
+               points -= 30;
 
        /*
-        * If p's nodes don't overlap ours, it may still help to kill p
-        * because p may have allocated or otherwise mapped memory on
-        * this node before. However it will be less likely.
+        * /proc/pid/oom_score_adj ranges from -1000 to +1000 such that it may
+        * either completely disable oom killing or always prefer a certain
+        * task.
         */
-       if (!cpuset_mems_allowed_intersects(current, p))
-               points /= 8;
+       points += p->signal->oom_score_adj;
 
        /*
-        * Adjust the score by oomkilladj.
+        * 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 (p->oomkilladj) {
-               if (p->oomkilladj > 0) {
-                       if (!points)
-                               points = 1;
-                       points <<= p->oomkilladj;
-               } else
-                       points >>= -(p->oomkilladj);
-       }
-
-#ifdef DEBUG
-       printk(KERN_DEBUG "OOMkill: task %d (%s) got %lu points\n",
-       p->pid, p->comm, points);
-#endif
-       return points;
+       if (points <= 0)
+               return 1;
+       return (points < 1000) ? points : 1000;
 }
 
 /*
  * Determine the type of allocation constraint.
  */
-static inline enum oom_constraint constrained_alloc(struct zonelist *zonelist,
-                                                   gfp_t gfp_mask)
-{
 #ifdef CONFIG_NUMA
-       struct zone **z;
-       nodemask_t nodes = node_states[N_HIGH_MEMORY];
+static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
+                               gfp_t gfp_mask, nodemask_t *nodemask,
+                               unsigned long *totalpages)
+{
+       struct zone *zone;
+       struct zoneref *z;
+       enum zone_type high_zoneidx = gfp_zone(gfp_mask);
+       bool cpuset_limited = false;
+       int nid;
 
-       for (z = zonelist->zones; *z; z++)
-               if (cpuset_zone_allowed_softwall(*z, gfp_mask))
-                       node_clear(zone_to_nid(*z), nodes);
-               else
-                       return CONSTRAINT_CPUSET;
+       /* Default to all available memory */
+       *totalpages = totalram_pages + total_swap_pages;
+
+       if (!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)
+               return CONSTRAINT_NONE;
 
-       if (!nodes_empty(nodes))
+       /*
+        * This is not a __GFP_THISNODE allocation, so a truncated nodemask in
+        * 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_HIGH_MEMORY], *nodemask)) {
+               *totalpages = total_swap_pages;
+               for_each_node_mask(nid, *nodemask)
+                       *totalpages += node_spanned_pages(nid);
                return CONSTRAINT_MEMORY_POLICY;
-#endif
+       }
 
+       /* 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_softwall(zone, gfp_mask))
+                       cpuset_limited = true;
+
+       if (cpuset_limited) {
+               *totalpages = total_swap_pages;
+               for_each_node_mask(nid, cpuset_current_mems_allowed)
+                       *totalpages += node_spanned_pages(nid);
+               return CONSTRAINT_CPUSET;
+       }
+       return CONSTRAINT_NONE;
+}
+#else
+static enum oom_constraint constrained_alloc(struct zonelist *zonelist,
+                               gfp_t gfp_mask, nodemask_t *nodemask,
+                               unsigned long *totalpages)
+{
+       *totalpages = totalram_pages + total_swap_pages;
        return CONSTRAINT_NONE;
 }
+#endif
 
 /*
  * Simple selection loop. We chose the process with the highest
@@ -195,28 +274,20 @@ static inline enum oom_constraint constrained_alloc(struct zonelist *zonelist,
  *
  * (not docbooked, we don't want this one cluttering up the manual)
  */
-static struct task_struct *select_bad_process(unsigned long *ppoints,
-                                               struct mem_cgroup *mem)
+static struct task_struct *select_bad_process(unsigned int *ppoints,
+               unsigned long totalpages, struct mem_cgroup *mem,
+               const nodemask_t *nodemask)
 {
        struct task_struct *g, *p;
        struct task_struct *chosen = NULL;
-       struct timespec uptime;
        *ppoints = 0;
 
-       do_posix_clock_monotonic_gettime(&uptime);
        do_each_thread(g, p) {
-               unsigned long points;
+               unsigned int points;
 
-               /*
-                * skip kernel threads and tasks which have already released
-                * their mm.
-                */
                if (!p->mm)
                        continue;
-               /* skip the init task */
-               if (is_global_init(p))
-                       continue;
-               if (mem && !task_in_mem_cgroup(p, mem))
+               if (oom_unkillable_task(p, mem, nodemask))
                        continue;
 
                /*
@@ -231,29 +302,33 @@ static struct task_struct *select_bad_process(unsigned long *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 (p->flags & PF_EXITING) {
-                       if (p != current)
-                               return ERR_PTR(-1UL);
-
-                       chosen = p;
-                       *ppoints = ULONG_MAX;
+                       /*
+                        * 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);
+                       }
                }
 
-               if (p->oomkilladj == OOM_DISABLE)
-                       continue;
-
-               points = badness(p, uptime.tv_sec, mem);
-               if (points > *ppoints || !chosen) {
+               points = oom_badness(p, mem, nodemask, totalpages);
+               if (points > *ppoints) {
                        chosen = p;
                        *ppoints = points;
                }
@@ -263,138 +338,217 @@ static struct task_struct *select_bad_process(unsigned long *ppoints,
 }
 
 /**
- * Send SIGKILL to the selected  process irrespective of  CAP_SYS_RAW_IO
- * flag though it's unlikely that  we select a process with CAP_SYS_RAW_IO
- * set.
+ * 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 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.
+ *
+ * Call with tasklist_lock read-locked.
  */
-static void __oom_kill_task(struct task_struct *p, int verbose)
+static void dump_tasks(const struct mem_cgroup *mem, const nodemask_t *nodemask)
 {
-       if (is_global_init(p)) {
-               WARN_ON(1);
-               printk(KERN_WARNING "tried to kill init!\n");
-               return;
-       }
+       struct task_struct *p;
+       struct task_struct *task;
 
-       if (!p->mm) {
-               WARN_ON(1);
-               printk(KERN_WARNING "tried to kill an mm-less task!\n");
-               return;
-       }
+       pr_info("[ pid ]   uid  tgid total_vm      rss cpu oom_adj oom_score_adj name\n");
+       for_each_process(p) {
+               if (oom_unkillable_task(p, mem, nodemask))
+                       continue;
 
-       if (verbose)
-               printk(KERN_ERR "Killed process %d (%s)\n",
-                               task_pid_nr(p), p->comm);
+               task = find_lock_task_mm(p);
+               if (!task) {
+                       /*
+                        * This is a kthread or all of p's threads have already
+                        * detached their mm's.  There's no need to report
+                        * them; they can't be oom killed anyway.
+                        */
+                       continue;
+               }
 
-       /*
-        * 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...
-        */
-       p->rt.time_slice = HZ;
-       set_tsk_thread_flag(p, TIF_MEMDIE);
+               pr_info("[%5d] %5d %5d %8lu %8lu %3u     %3d         %5d %s\n",
+                       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);
+               task_unlock(task);
+       }
+}
 
-       force_sig(SIGKILL, p);
+static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
+                       struct mem_cgroup *mem, const nodemask_t *nodemask)
+{
+       task_lock(current);
+       pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
+               "oom_adj=%d, oom_score_adj=%d\n",
+               current->comm, gfp_mask, order, current->signal->oom_adj,
+               current->signal->oom_score_adj);
+       cpuset_print_task_mems_allowed(current);
+       task_unlock(current);
+       dump_stack();
+       mem_cgroup_print_oom_info(mem, p);
+       show_mem(SHOW_MEM_FILTER_NODES);
+       if (sysctl_oom_dump_tasks)
+               dump_tasks(mem, nodemask);
 }
 
-static int oom_kill_task(struct task_struct *p)
+#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;
-       struct task_struct *g, *q;
 
-       mm = p->mm;
-
-       /* WARNING: mm may not be dereferenced since we did not obtain its
-        * value from get_task_mm(p).  This is OK since all we need to do is
-        * compare mm to q->mm below.
-        *
-        * Furthermore, even if mm contains a non-NULL value, p->mm may
-        * change to NULL at any time since we do not hold task_lock(p).
-        * However, this is of no concern to us.
-        */
-
-       if (mm == NULL)
+       p = find_lock_task_mm(p);
+       if (!p)
                return 1;
 
-       /*
-        * Don't kill the process if any threads are set to OOM_DISABLE
-        */
-       do_each_thread(g, q) {
-               if (q->mm == mm && q->oomkilladj == OOM_DISABLE)
-                       return 1;
-       } while_each_thread(g, q);
+       /* mm cannot be safely dereferenced after task_unlock(p) */
+       mm = p->mm;
 
-       __oom_kill_task(p, 1);
+       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 that share the ->mm (i.e. all threads),
-        * but are in a different thread group. Don't let them have access
-        * to memory reserves though, otherwise we might deplete all memory.
+        * 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.
         */
-       do_each_thread(g, q) {
-               if (q->mm == mm && !same_thread_group(q, p))
+       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);
-       } while_each_thread(g, q);
+               }
+
+       set_tsk_thread_flag(p, TIF_MEMDIE);
+       force_sig(SIGKILL, p);
 
        return 0;
 }
+#undef K
 
 static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
-                           unsigned long points, const char *message)
+                           unsigned int points, unsigned long totalpages,
+                           struct mem_cgroup *mem, nodemask_t *nodemask,
+                           const char *message)
 {
-       struct task_struct *c;
-
-       if (printk_ratelimit()) {
-               printk(KERN_WARNING "%s invoked oom-killer: "
-                       "gfp_mask=0x%x, order=%d, oomkilladj=%d\n",
-                       current->comm, gfp_mask, order, current->oomkilladj);
-               dump_stack();
-               show_mem();
-       }
+       struct task_struct *victim = p;
+       struct task_struct *child;
+       struct task_struct *t = p;
+       unsigned int victim_points = 0;
+
+       if (printk_ratelimit())
+               dump_header(p, gfp_mask, order, mem, nodemask);
 
        /*
         * If the task is already exiting, don't alarm the sysadmin or kill
         * its children or threads, just set TIF_MEMDIE so it can die quickly
         */
        if (p->flags & PF_EXITING) {
-               __oom_kill_task(p, 0);
+               set_tsk_thread_flag(p, TIF_MEMDIE);
                return 0;
        }
 
-       printk(KERN_ERR "%s: kill process %d (%s) score %li or a child\n",
-                                       message, task_pid_nr(p), p->comm, points);
+       task_lock(p);
+       pr_err("%s: Kill process %d (%s) score %d or sacrifice child\n",
+               message, task_pid_nr(p), p->comm, points);
+       task_unlock(p);
+
+       /*
+        * If any of p's children has a different mm and is eligible for kill,
+        * the one with the highest badness() score is sacrificed for its
+        * parent.  This attempts to lose the minimal amount of work done while
+        * still freeing memory.
+        */
+       do {
+               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
+                        */
+                       child_points = oom_badness(child, mem, nodemask,
+                                                               totalpages);
+                       if (child_points > victim_points) {
+                               victim = child;
+                               victim_points = child_points;
+                       }
+               }
+       } while_each_thread(p, t);
 
-       /* Try to kill a child first */
-       list_for_each_entry(c, &p->children, sibling) {
-               if (c->mm == p->mm)
-                       continue;
-               if (!oom_kill_task(c))
-                       return 0;
+       return oom_kill_task(victim, mem);
+}
+
+/*
+ * 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, const nodemask_t *nodemask)
+{
+       if (likely(!sysctl_panic_on_oom))
+               return;
+       if (sysctl_panic_on_oom != 2) {
+               /*
+                * panic_on_oom == 1 only affects CONSTRAINT_NONE, the kernel
+                * does not panic for cpuset, mempolicy, or memcg allocation
+                * failures.
+                */
+               if (constraint != CONSTRAINT_NONE)
+                       return;
        }
-       return oom_kill_task(p);
+       read_lock(&tasklist_lock);
+       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");
 }
 
-#ifdef CONFIG_CGROUP_MEM_CONT
+#ifdef CONFIG_CGROUP_MEM_RES_CTLR
 void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
 {
-       unsigned long points = 0;
+       unsigned long limit;
+       unsigned int points = 0;
        struct task_struct *p;
 
-       cgroup_lock();
-       rcu_read_lock();
+       /*
+        * 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:
-       p = select_bad_process(&points, mem);
-       if (PTR_ERR(p) == -1UL)
+       p = select_bad_process(&points, limit, mem, NULL);
+       if (!p || PTR_ERR(p) == -1UL)
                goto out;
 
-       if (!p)
-               p = current;
-
-       if (oom_kill_process(p, gfp_mask, 0, points,
+       if (oom_kill_process(p, gfp_mask, 0, points, limit, mem, NULL,
                                "Memory cgroup out of memory"))
                goto retry;
 out:
-       rcu_read_unlock();
-       cgroup_unlock();
+       read_unlock(&tasklist_lock);
 }
 #endif
 
@@ -417,31 +571,31 @@ EXPORT_SYMBOL_GPL(unregister_oom_notifier);
  * if a parallel OOM killing is already taking place that includes a zone in
  * the zonelist.  Otherwise, locks all zones in the zonelist and returns 1.
  */
-int try_set_zone_oom(struct zonelist *zonelist)
+int try_set_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
 {
-       struct zone **z;
+       struct zoneref *z;
+       struct zone *zone;
        int ret = 1;
 
-       z = zonelist->zones;
-
-       spin_lock(&zone_scan_mutex);
-       do {
-               if (zone_is_oom_locked(*z)) {
+       spin_lock(&zone_scan_lock);
+       for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
+               if (zone_is_oom_locked(zone)) {
                        ret = 0;
                        goto out;
                }
-       } while (*(++z) != NULL);
+       }
+
+       for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
+               /*
+                * Lock each zone in the zonelist under zone_scan_lock so a
+                * parallel invocation of try_set_zonelist_oom() doesn't succeed
+                * when it shouldn't.
+                */
+               zone_set_flag(zone, ZONE_OOM_LOCKED);
+       }
 
-       /*
-        * Lock each zone in the zonelist under zone_scan_mutex so a parallel
-        * invocation of try_set_zone_oom() doesn't succeed when it shouldn't.
-        */
-       z = zonelist->zones;
-       do {
-               zone_set_flag(*z, ZONE_OOM_LOCKED);
-       } while (*(++z) != NULL);
 out:
-       spin_unlock(&zone_scan_mutex);
+       spin_unlock(&zone_scan_lock);
        return ret;
 }
 
@@ -450,88 +604,133 @@ out:
  * allocation attempts with zonelists containing them may now recall the OOM
  * killer, if necessary.
  */
-void clear_zonelist_oom(struct zonelist *zonelist)
+void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
 {
-       struct zone **z;
+       struct zoneref *z;
+       struct zone *zone;
 
-       z = zonelist->zones;
+       spin_lock(&zone_scan_lock);
+       for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
+               zone_clear_flag(zone, ZONE_OOM_LOCKED);
+       }
+       spin_unlock(&zone_scan_lock);
+}
 
-       spin_lock(&zone_scan_mutex);
-       do {
-               zone_clear_flag(*z, ZONE_OOM_LOCKED);
-       } while (*(++z) != NULL);
-       spin_unlock(&zone_scan_mutex);
+/*
+ * Try to acquire the oom killer lock for all system zones.  Returns zero if a
+ * parallel oom killing is taking place, otherwise locks all zones and returns
+ * non-zero.
+ */
+static int try_set_system_oom(void)
+{
+       struct zone *zone;
+       int ret = 1;
+
+       spin_lock(&zone_scan_lock);
+       for_each_populated_zone(zone)
+               if (zone_is_oom_locked(zone)) {
+                       ret = 0;
+                       goto out;
+               }
+       for_each_populated_zone(zone)
+               zone_set_flag(zone, ZONE_OOM_LOCKED);
+out:
+       spin_unlock(&zone_scan_lock);
+       return ret;
+}
+
+/*
+ * Clears ZONE_OOM_LOCKED for all system zones so that failed allocation
+ * attempts or page faults may now recall the oom killer, if necessary.
+ */
+static void clear_system_oom(void)
+{
+       struct zone *zone;
+
+       spin_lock(&zone_scan_lock);
+       for_each_populated_zone(zone)
+               zone_clear_flag(zone, ZONE_OOM_LOCKED);
+       spin_unlock(&zone_scan_lock);
 }
 
 /**
  * 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
  *
  * 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.
  */
-void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
+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 points = 0;
+       unsigned long totalpages;
        unsigned long freed = 0;
-       enum oom_constraint constraint;
+       unsigned int points;
+       enum oom_constraint constraint = CONSTRAINT_NONE;
+       int killed = 0;
 
        blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
        if (freed > 0)
                /* Got some memory back in the last second. */
                return;
 
-       if (sysctl_panic_on_oom == 2)
-               panic("out of memory. Compulsory panic_on_oom is selected.\n");
+       /*
+        * 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 if there were limitations on the allocation (only relevant for
         * NUMA) that may require different handling.
         */
-       constraint = constrained_alloc(zonelist, gfp_mask);
-       read_lock(&tasklist_lock);
+       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);
 
-       switch (constraint) {
-       case CONSTRAINT_MEMORY_POLICY:
-               oom_kill_process(current, gfp_mask, order, points,
-                               "No available memory (MPOL_BIND)");
-               break;
-
-       case CONSTRAINT_NONE:
-               if (sysctl_panic_on_oom)
-                       panic("out of memory. panic_on_oom is selected\n");
-               /* Fall-through */
-       case CONSTRAINT_CPUSET:
-               if (sysctl_oom_kill_allocating_task) {
-                       oom_kill_process(current, gfp_mask, order, points,
-                                       "Out of memory (oom_kill_allocating_task)");
-                       break;
-               }
-retry:
+       read_lock(&tasklist_lock);
+       if (sysctl_oom_kill_allocating_task &&
+           !oom_unkillable_task(current, NULL, nodemask) &&
+           current->mm && !atomic_read(&current->mm->oom_disable_count)) {
                /*
-                * Rambo mode: Shoot down a process and hope it solves whatever
-                * issues we may have.
+                * oom_kill_process() needs tasklist_lock held.  If it returns
+                * non-zero, current could not be killed so we must fallback to
+                * the tasklist scan.
                 */
-               p = select_bad_process(&points, NULL);
-
-               if (PTR_ERR(p) == -1UL)
+               if (!oom_kill_process(current, gfp_mask, order, 0, totalpages,
+                               NULL, nodemask,
+                               "Out of memory (oom_kill_allocating_task)"))
                        goto out;
+       }
 
-               /* Found nothing?!?! Either we hang forever, or we panic. */
-               if (!p) {
-                       read_unlock(&tasklist_lock);
-                       panic("Out of memory and no killable processes...\n");
-               }
-
-               if (oom_kill_process(p, gfp_mask, order, points,
-                                    "Out of memory"))
-                       goto retry;
+retry:
+       p = select_bad_process(&points, totalpages, NULL, mpol_mask);
+       if (PTR_ERR(p) == -1UL)
+               goto out;
 
-               break;
+       /* Found nothing?!?! Either we hang forever, or we panic. */
+       if (!p) {
+               dump_header(NULL, gfp_mask, order, NULL, mpol_mask);
+               read_unlock(&tasklist_lock);
+               panic("Out of memory and no killable processes...\n");
        }
 
+       if (oom_kill_process(p, gfp_mask, order, points, totalpages, NULL,
+                               nodemask, "Out of memory"))
+               goto retry;
+       killed = 1;
 out:
        read_unlock(&tasklist_lock);
 
@@ -539,6 +738,22 @@ out:
         * Give "p" a good chance of killing itself before we
         * retry to allocate memory unless "p" is current
         */
+       if (killed && !test_thread_flag(TIF_MEMDIE))
+               schedule_timeout_uninterruptible(1);
+}
+
+/*
+ * The pagefault handler calls here because it is out of memory, so kill a
+ * memory-hogging task.  If a populated zone has ZONE_OOM_LOCKED set, a parallel
+ * oom killing is already in progress so do nothing.  If a task is found with
+ * TIF_MEMDIE set, it has been killed so do nothing and allow it to exit.
+ */
+void pagefault_out_of_memory(void)
+{
+       if (try_set_system_oom()) {
+               out_of_memory(NULL, 0, 0, NULL);
+               clear_system_oom();
+       }
        if (!test_thread_flag(TIF_MEMDIE))
                schedule_timeout_uninterruptible(1);
 }