thp: add tlb_remove_pmd_tlb_entry
[linux-2.6.git] / mm / oom_kill.c
index fcb2fa5..2958fd8 100644 (file)
 #include <linux/mempolicy.h>
 #include <linux/security.h>
 #include <linux/ptrace.h>
+#include <linux/freezer.h>
+#include <linux/ftrace.h>
+
+#define CREATE_TRACE_POINTS
+#include <trace/events/oom.h>
 
 int sysctl_panic_on_oom;
 int sysctl_oom_kill_allocating_task;
 int sysctl_oom_dump_tasks = 1;
 static DEFINE_SPINLOCK(zone_scan_lock);
 
+/*
+ * compare_swap_oom_score_adj() - compare and swap current's oom_score_adj
+ * @old_val: old oom_score_adj for compare
+ * @new_val: new oom_score_adj for swap
+ *
+ * Sets the oom_score_adj value for current to @new_val iff its present value is
+ * @old_val.  Usually used to reinstate a previous value to prevent racing with
+ * userspacing tuning the value in the interim.
+ */
+void compare_swap_oom_score_adj(int old_val, int new_val)
+{
+       struct sighand_struct *sighand = current->sighand;
+
+       spin_lock_irq(&sighand->siglock);
+       if (current->signal->oom_score_adj == old_val)
+               current->signal->oom_score_adj = new_val;
+       trace_oom_score_adj_update(current);
+       spin_unlock_irq(&sighand->siglock);
+}
+
 /**
  * test_set_oom_score_adj() - set current's oom_score_adj and return old value
  * @new_val: new oom_score_adj value
@@ -53,13 +78,8 @@ int test_set_oom_score_adj(int new_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;
-       }
+       current->signal->oom_score_adj = new_val;
+       trace_oom_score_adj_update(current);
        spin_unlock_irq(&sighand->siglock);
 
        return old_val;
@@ -132,7 +152,7 @@ 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,
-               const struct mem_cgroup *mem, const nodemask_t *nodemask)
+               const struct mem_cgroup *memcg, const nodemask_t *nodemask)
 {
        if (is_global_init(p))
                return true;
@@ -140,7 +160,7 @@ static bool oom_unkillable_task(struct task_struct *p,
                return true;
 
        /* When mem_cgroup_out_of_memory() and p is not member of the group */
-       if (mem && !task_in_mem_cgroup(p, mem))
+       if (memcg && !task_in_mem_cgroup(p, memcg))
                return true;
 
        /* p may not have freeable memory in nodemask */
@@ -159,24 +179,19 @@ static bool oom_unkillable_task(struct task_struct *p,
  * 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,
+unsigned int oom_badness(struct task_struct *p, struct mem_cgroup *memcg,
                      const nodemask_t *nodemask, unsigned long totalpages)
 {
-       int points;
+       long points;
 
-       if (oom_unkillable_task(p, mem, nodemask))
+       if (oom_unkillable_task(p, memcg, nodemask))
                return 0;
 
        p = find_lock_task_mm(p);
        if (!p)
                return 0;
 
-       /*
-        * 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 (atomic_read(&p->mm->oom_disable_count)) {
+       if (p->signal->oom_score_adj == OOM_SCORE_ADJ_MIN) {
                task_unlock(p);
                return 0;
        }
@@ -293,7 +308,7 @@ static 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 int *ppoints,
-               unsigned long totalpages, struct mem_cgroup *mem,
+               unsigned long totalpages, struct mem_cgroup *memcg,
                const nodemask_t *nodemask)
 {
        struct task_struct *g, *p;
@@ -305,7 +320,7 @@ static struct task_struct *select_bad_process(unsigned int *ppoints,
 
                if (p->exit_state)
                        continue;
-               if (oom_unkillable_task(p, mem, nodemask))
+               if (oom_unkillable_task(p, memcg, nodemask))
                        continue;
 
                /*
@@ -317,8 +332,11 @@ static struct task_struct *select_bad_process(unsigned int *ppoints,
                 * blocked waiting for another task which itself is waiting
                 * for memory. Is there a better alternative?
                 */
-               if (test_tsk_thread_flag(p, TIF_MEMDIE))
+               if (test_tsk_thread_flag(p, TIF_MEMDIE)) {
+                       if (unlikely(frozen(p)))
+                               __thaw_task(p);
                        return ERR_PTR(-1UL);
+               }
                if (!p->mm)
                        continue;
 
@@ -346,7 +364,7 @@ static struct task_struct *select_bad_process(unsigned int *ppoints,
                        }
                }
 
-               points = oom_badness(p, mem, nodemask, totalpages);
+               points = oom_badness(p, memcg, nodemask, totalpages);
                if (points > *ppoints) {
                        chosen = p;
                        *ppoints = points;
@@ -369,14 +387,14 @@ static struct task_struct *select_bad_process(unsigned int *ppoints,
  *
  * Call with tasklist_lock read-locked.
  */
-static void dump_tasks(const struct mem_cgroup *mem, const nodemask_t *nodemask)
+static void dump_tasks(const struct mem_cgroup *memcg, 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 (oom_unkillable_task(p, mem, nodemask))
+               if (oom_unkillable_task(p, memcg, nodemask))
                        continue;
 
                task = find_lock_task_mm(p);
@@ -399,7 +417,7 @@ static void dump_tasks(const struct mem_cgroup *mem, const nodemask_t *nodemask)
 }
 
 static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
-                       struct mem_cgroup *mem, const nodemask_t *nodemask)
+                       struct mem_cgroup *memcg, const nodemask_t *nodemask)
 {
        task_lock(current);
        pr_warning("%s invoked oom-killer: gfp_mask=0x%x, order=%d, "
@@ -409,14 +427,14 @@ static void dump_header(struct task_struct *p, gfp_t gfp_mask, int order,
        cpuset_print_task_mems_allowed(current);
        task_unlock(current);
        dump_stack();
-       mem_cgroup_print_oom_info(mem, p);
+       mem_cgroup_print_oom_info(memcg, p);
        show_mem(SHOW_MEM_FILTER_NODES);
        if (sysctl_oom_dump_tasks)
-               dump_tasks(mem, nodemask);
+               dump_tasks(memcg, nodemask);
 }
 
 #define K(x) ((x) << (PAGE_SHIFT-10))
-static int oom_kill_task(struct task_struct *p, struct mem_cgroup *mem)
+static int oom_kill_task(struct task_struct *p)
 {
        struct task_struct *q;
        struct mm_struct *mm;
@@ -435,7 +453,7 @@ static int oom_kill_task(struct task_struct *p, struct mem_cgroup *mem)
        task_unlock(p);
 
        /*
-        * Kill all processes sharing p->mm in other thread groups, if any.
+        * Kill all user 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
@@ -445,7 +463,11 @@ static int oom_kill_task(struct task_struct *p, struct mem_cgroup *mem)
         * signal.
         */
        for_each_process(q)
-               if (q->mm == mm && !same_thread_group(q, p)) {
+               if (q->mm == mm && !same_thread_group(q, p) &&
+                   !(q->flags & PF_KTHREAD)) {
+                       if (q->signal->oom_score_adj == OOM_SCORE_ADJ_MIN)
+                               continue;
+
                        task_lock(q);   /* Protect ->comm from prctl() */
                        pr_err("Kill process %d (%s) sharing same memory\n",
                                task_pid_nr(q), q->comm);
@@ -462,7 +484,7 @@ static int oom_kill_task(struct task_struct *p, struct mem_cgroup *mem)
 
 static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
                            unsigned int points, unsigned long totalpages,
-                           struct mem_cgroup *mem, nodemask_t *nodemask,
+                           struct mem_cgroup *memcg, nodemask_t *nodemask,
                            const char *message)
 {
        struct task_struct *victim = p;
@@ -471,7 +493,7 @@ 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, nodemask);
+               dump_header(p, gfp_mask, order, memcg, nodemask);
 
        /*
         * If the task is already exiting, don't alarm the sysadmin or kill
@@ -502,7 +524,7 @@ static int 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, mem, nodemask,
+                       child_points = oom_badness(child, memcg, nodemask,
                                                                totalpages);
                        if (child_points > victim_points) {
                                victim = child;
@@ -511,7 +533,7 @@ static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
                }
        } while_each_thread(p, t);
 
-       return oom_kill_task(victim, mem);
+       return oom_kill_task(victim);
 }
 
 /*
@@ -539,7 +561,7 @@ static void check_panic_on_oom(enum oom_constraint constraint, gfp_t gfp_mask,
 }
 
 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
-void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
+void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask)
 {
        unsigned long limit;
        unsigned int points = 0;
@@ -556,14 +578,14 @@ void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
        }
 
        check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, 0, NULL);
-       limit = mem_cgroup_get_limit(mem) >> PAGE_SHIFT;
+       limit = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT;
        read_lock(&tasklist_lock);
 retry:
-       p = select_bad_process(&points, limit, mem, NULL);
+       p = select_bad_process(&points, limit, memcg, NULL);
        if (!p || PTR_ERR(p) == -1UL)
                goto out;
 
-       if (oom_kill_process(p, gfp_mask, 0, points, limit, mem, NULL,
+       if (oom_kill_process(p, gfp_mask, 0, points, limit, memcg, NULL,
                                "Memory cgroup out of memory"))
                goto retry;
 out:
@@ -722,7 +744,7 @@ void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
        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)) {
+           current->mm) {
                /*
                 * oom_kill_process() needs tasklist_lock held.  If it returns
                 * non-zero, current could not be killed so we must fallback to