memcg: fix swap accounting
[safe/jmp/linux-2.6] / mm / oom_kill.c
index e5f50cf..175a67a 100644 (file)
@@ -55,9 +55,10 @@ static DEFINE_SPINLOCK(zone_scan_lock);
 
 unsigned long badness(struct task_struct *p, unsigned long uptime)
 {
-       unsigned long points, cpu_time, run_time, s;
+       unsigned long points, cpu_time, run_time;
        struct mm_struct *mm;
        struct task_struct *child;
+       int oom_adj;
 
        task_lock(p);
        mm = p->mm;
@@ -65,6 +66,11 @@ unsigned long badness(struct task_struct *p, unsigned long uptime)
                task_unlock(p);
                return 0;
        }
+       oom_adj = mm->oom_adj;
+       if (oom_adj == OOM_DISABLE) {
+               task_unlock(p);
+               return 0;
+       }
 
        /*
         * The memory size of the process is the basis for the badness.
@@ -110,12 +116,10 @@ unsigned long badness(struct task_struct *p, unsigned long uptime)
        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 (cpu_time)
+               points /= int_sqrt(cpu_time);
+       if (run_time)
+               points /= int_sqrt(int_sqrt(run_time));
 
        /*
         * Niced processes are most likely less important, so double
@@ -150,15 +154,15 @@ unsigned long badness(struct task_struct *p, unsigned long uptime)
                points /= 8;
 
        /*
-        * Adjust the score by oomkilladj.
+        * Adjust the score by oom_adj.
         */
-       if (p->oomkilladj) {
-               if (p->oomkilladj > 0) {
+       if (oom_adj) {
+               if (oom_adj > 0) {
                        if (!points)
                                points = 1;
-                       points <<= p->oomkilladj;
+                       points <<= oom_adj;
                } else
-                       points >>= -(p->oomkilladj);
+                       points >>= -(oom_adj);
        }
 
 #ifdef DEBUG
@@ -253,11 +257,8 @@ static struct task_struct *select_bad_process(unsigned long *ppoints,
                        *ppoints = ULONG_MAX;
                }
 
-               if (p->oomkilladj == OOM_DISABLE)
-                       continue;
-
                points = badness(p, uptime.tv_sec);
-               if (points > *ppoints || !chosen) {
+               if (points > *ppoints) {
                        chosen = p;
                        *ppoints = points;
                }
@@ -286,22 +287,27 @@ static void dump_tasks(const struct mem_cgroup *mem)
        printk(KERN_INFO "[ pid ]   uid  tgid total_vm      rss cpu oom_adj "
               "name\n");
        do_each_thread(g, p) {
-               /*
-                * total_vm and rss sizes do not exist for tasks with a
-                * detached mm so there's no need to report them.
-                */
-               if (!p->mm)
-                       continue;
+               struct mm_struct *mm;
+
                if (mem && !task_in_mem_cgroup(p, mem))
                        continue;
                if (!thread_group_leader(p))
                        continue;
 
                task_lock(p);
+               mm = p->mm;
+               if (!mm) {
+                       /*
+                        * total_vm and rss sizes do not exist for tasks with no
+                        * mm so there's no need to report them; they can't be
+                        * oom killed anyway.
+                        */
+                       task_unlock(p);
+                       continue;
+               }
                printk(KERN_INFO "[%5d] %5d %5d %8lu %8lu %3d     %3d %s\n",
-                      p->pid, __task_cred(p)->uid, p->tgid,
-                      p->mm->total_vm, get_mm_rss(p->mm), (int)task_cpu(p),
-                      p->oomkilladj, p->comm);
+                      p->pid, __task_cred(p)->uid, p->tgid, mm->total_vm,
+                      get_mm_rss(mm), (int)task_cpu(p), mm->oom_adj, p->comm);
                task_unlock(p);
        } while_each_thread(g, p);
 }
@@ -319,11 +325,8 @@ static void __oom_kill_task(struct task_struct *p, int verbose)
                return;
        }
 
-       if (!p->mm) {
-               WARN_ON(1);
-               printk(KERN_WARNING "tried to kill an mm-less task!\n");
+       if (!p->mm)
                return;
-       }
 
        if (verbose)
                printk(KERN_ERR "Killed process %d (%s)\n",
@@ -345,28 +348,13 @@ static int oom_kill_task(struct task_struct *p)
        struct mm_struct *mm;
        struct task_struct *g, *q;
 
+       task_lock(p);
        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)
+       if (!mm || mm->oom_adj == OOM_DISABLE) {
+               task_unlock(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);
-
+       }
+       task_unlock(p);
        __oom_kill_task(p, 1);
 
        /*
@@ -389,10 +377,15 @@ static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
        struct task_struct *c;
 
        if (printk_ratelimit()) {
+               task_lock(current);
                printk(KERN_WARNING "%s invoked oom-killer: "
-                       "gfp_mask=0x%x, order=%d, oomkilladj=%d\n",
-                       current->comm, gfp_mask, order, current->oomkilladj);
+                       "gfp_mask=0x%x, order=%d, oom_adj=%d\n",
+                       current->comm, gfp_mask, order,
+                       current->mm ? current->mm->oom_adj : OOM_DISABLE);
+               cpuset_print_task_mems_allowed(current);
+               task_unlock(current);
                dump_stack();
+               mem_cgroup_print_oom_info(mem, current);
                show_mem();
                if (sysctl_oom_dump_tasks)
                        dump_tasks(mem);
@@ -401,8 +394,9 @@ static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
        /*
         * 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 its mm is still attached.
         */
-       if (p->flags & PF_EXITING) {
+       if (p->mm && (p->flags & PF_EXITING)) {
                __oom_kill_task(p, 0);
                return 0;
        }
@@ -426,7 +420,6 @@ void mem_cgroup_out_of_memory(struct mem_cgroup *mem, gfp_t gfp_mask)
        unsigned long points = 0;
        struct task_struct *p;
 
-       cgroup_lock();
        read_lock(&tasklist_lock);
 retry:
        p = select_bad_process(&points, mem);
@@ -441,7 +434,6 @@ retry:
                goto retry;
 out:
        read_unlock(&tasklist_lock);
-       cgroup_unlock();
 }
 #endif
 
@@ -514,34 +506,32 @@ void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
  */
 static void __out_of_memory(gfp_t gfp_mask, int order)
 {
-       if (sysctl_oom_kill_allocating_task) {
-               oom_kill_process(current, gfp_mask, order, 0, NULL,
-                               "Out of memory (oom_kill_allocating_task)");
-
-       } else {
-               unsigned long points;
-               struct task_struct *p;
-
-retry:
-               /*
-                * Rambo mode: Shoot down a process and hope it solves whatever
-                * issues we may have.
-                */
-               p = select_bad_process(&points, NULL);
+       struct task_struct *p;
+       unsigned long points;
 
-               if (PTR_ERR(p) == -1UL)
+       if (sysctl_oom_kill_allocating_task)
+               if (!oom_kill_process(current, gfp_mask, order, 0, NULL,
+                               "Out of memory (oom_kill_allocating_task)"))
                        return;
+retry:
+       /*
+        * Rambo mode: Shoot down a process and hope it solves whatever
+        * issues we may have.
+        */
+       p = select_bad_process(&points, NULL);
 
-               /* 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 (PTR_ERR(p) == -1UL)
+               return;
 
-               if (oom_kill_process(p, gfp_mask, order, points, NULL,
-                                    "Out of memory"))
-                       goto retry;
+       /* 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, NULL,
+                            "Out of memory"))
+               goto retry;
 }
 
 /*
@@ -557,6 +547,13 @@ void pagefault_out_of_memory(void)
                /* Got some memory back in the last second. */
                return;
 
+       /*
+        * If this is from memcg, oom-killer is already invoked.
+        * and not worth to go system-wide-oom.
+        */
+       if (mem_cgroup_oom_called(current))
+               goto rest_and_return;
+
        if (sysctl_panic_on_oom)
                panic("out of memory from page fault. panic_on_oom is selected.\n");
 
@@ -568,6 +565,7 @@ void pagefault_out_of_memory(void)
         * Give "p" a good chance of killing itself before we
         * retry to allocate memory.
         */
+rest_and_return:
        if (!test_thread_flag(TIF_MEMDIE))
                schedule_timeout_uninterruptible(1);
 }