security: Fix setting of PF_SUPERPRIV by __capable()
[linux-2.6.git] / mm / oom_kill.c
index 183a964..64e5b4b 100644 (file)
 #include <linux/cpuset.h>
 #include <linux/module.h>
 #include <linux/notifier.h>
+#include <linux/memcontrol.h>
+#include <linux/security.h>
 
 int sysctl_panic_on_oom;
 int sysctl_oom_kill_allocating_task;
-static DEFINE_MUTEX(zone_scan_mutex);
+int sysctl_oom_dump_tasks;
+static DEFINE_SPINLOCK(zone_scan_mutex);
 /* #define DEBUG */
 
 /**
  * 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
+ * @mem: target memory controller
  *
  * The formula used is relatively simple and documented inline in the
  * function. The main rationale is that we want to select a good task
@@ -125,8 +129,8 @@ unsigned long badness(struct task_struct *p, unsigned long uptime)
         * Superuser processes are usually more important, so we make it
         * less likely that we kill those.
         */
-       if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_ADMIN) ||
-                               p->uid == 0 || p->euid == 0)
+       if (has_capability(p, CAP_SYS_ADMIN) ||
+           has_capability(p, CAP_SYS_RESOURCE))
                points /= 4;
 
        /*
@@ -135,7 +139,7 @@ unsigned long badness(struct task_struct *p, unsigned long uptime)
         * tend to only have this flag set on applications they think
         * of as important.
         */
-       if (cap_t(p->cap_effective) & CAP_TO_MASK(CAP_SYS_RAWIO))
+       if (has_capability(p, CAP_SYS_RAWIO))
                points /= 4;
 
        /*
@@ -172,12 +176,14 @@ static inline enum oom_constraint constrained_alloc(struct zonelist *zonelist,
                                                    gfp_t gfp_mask)
 {
 #ifdef CONFIG_NUMA
-       struct zone **z;
+       struct zone *zone;
+       struct zoneref *z;
+       enum zone_type high_zoneidx = gfp_zone(gfp_mask);
        nodemask_t nodes = node_states[N_HIGH_MEMORY];
 
-       for (z = zonelist->zones; *z; z++)
-               if (cpuset_zone_allowed_softwall(*z, gfp_mask))
-                       node_clear(zone_to_nid(*z), nodes);
+       for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
+               if (cpuset_zone_allowed_softwall(zone, gfp_mask))
+                       node_clear(zone_to_nid(zone), nodes);
                else
                        return CONSTRAINT_CPUSET;
 
@@ -194,7 +200,8 @@ 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)
+static struct task_struct *select_bad_process(unsigned long *ppoints,
+                                               struct mem_cgroup *mem)
 {
        struct task_struct *g, *p;
        struct task_struct *chosen = NULL;
@@ -212,7 +219,9 @@ static struct task_struct *select_bad_process(unsigned long *ppoints)
                if (!p->mm)
                        continue;
                /* skip the init task */
-               if (is_init(p))
+               if (is_global_init(p))
+                       continue;
+               if (mem && !task_in_mem_cgroup(p, mem))
                        continue;
 
                /*
@@ -259,13 +268,51 @@ static struct task_struct *select_bad_process(unsigned long *ppoints)
 }
 
 /**
+ * dump_tasks - dump current memory state of all system tasks
+ * @mem: target memory controller
+ *
+ * Dumps the current memory state of all system tasks, excluding kernel threads.
+ * State information includes task's pid, uid, tgid, vm size, rss, cpu, oom_adj
+ * score, 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)
+{
+       struct task_struct *g, *p;
+
+       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;
+               if (mem && !task_in_mem_cgroup(p, mem))
+                       continue;
+
+               task_lock(p);
+               printk(KERN_INFO "[%5d] %5d %5d %8lu %8lu %3d     %3d %s\n",
+                      p->pid, p->uid, p->tgid, p->mm->total_vm,
+                      get_mm_rss(p->mm), (int)task_cpu(p), p->oomkilladj,
+                      p->comm);
+               task_unlock(p);
+       } while_each_thread(g, p);
+}
+
+/*
  * 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.
  */
 static void __oom_kill_task(struct task_struct *p, int verbose)
 {
-       if (is_init(p)) {
+       if (is_global_init(p)) {
                WARN_ON(1);
                printk(KERN_WARNING "tried to kill init!\n");
                return;
@@ -278,14 +325,15 @@ static void __oom_kill_task(struct task_struct *p, int verbose)
        }
 
        if (verbose)
-               printk(KERN_ERR "Killed process %d (%s)\n", p->pid, p->comm);
+               printk(KERN_ERR "Killed process %d (%s)\n",
+                               task_pid_nr(p), p->comm);
 
        /*
         * 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->time_slice = HZ;
+       p->rt.time_slice = HZ;
        set_tsk_thread_flag(p, TIF_MEMDIE);
 
        force_sig(SIGKILL, p);
@@ -326,7 +374,7 @@ static int oom_kill_task(struct task_struct *p)
         * to memory reserves though, otherwise we might deplete all memory.
         */
        do_each_thread(g, q) {
-               if (q->mm == mm && q->tgid != p->tgid)
+               if (q->mm == mm && !same_thread_group(q, p))
                        force_sig(SIGKILL, q);
        } while_each_thread(g, q);
 
@@ -334,10 +382,10 @@ static int oom_kill_task(struct task_struct *p)
 }
 
 static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
-                           unsigned long points, const char *message)
+                           unsigned long points, struct mem_cgroup *mem,
+                           const char *message)
 {
        struct task_struct *c;
-       struct list_head *tsk;
 
        if (printk_ratelimit()) {
                printk(KERN_WARNING "%s invoked oom-killer: "
@@ -345,6 +393,8 @@ static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
                        current->comm, gfp_mask, order, current->oomkilladj);
                dump_stack();
                show_mem();
+               if (sysctl_oom_dump_tasks)
+                       dump_tasks(mem);
        }
 
        /*
@@ -357,11 +407,10 @@ static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
        }
 
        printk(KERN_ERR "%s: kill process %d (%s) score %li or a child\n",
-                                       message, p->pid, p->comm, points);
+                                       message, task_pid_nr(p), p->comm, points);
 
        /* Try to kill a child first */
-       list_for_each(tsk, &p->children) {
-               c = list_entry(tsk, struct task_struct, sibling);
+       list_for_each_entry(c, &p->children, sibling) {
                if (c->mm == p->mm)
                        continue;
                if (!oom_kill_task(c))
@@ -370,6 +419,31 @@ static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
        return oom_kill_task(p);
 }
 
+#ifdef CONFIG_CGROUP_MEM_RES_CTLR
+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);
+       if (PTR_ERR(p) == -1UL)
+               goto out;
+
+       if (!p)
+               p = current;
+
+       if (oom_kill_process(p, gfp_mask, 0, points, mem,
+                               "Memory cgroup out of memory"))
+               goto retry;
+out:
+       read_unlock(&tasklist_lock);
+       cgroup_unlock();
+}
+#endif
+
 static BLOCKING_NOTIFIER_HEAD(oom_notify_list);
 
 int register_oom_notifier(struct notifier_block *nb)
@@ -389,31 +463,31 @@ 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_zone_oom(struct zonelist *zonelist, gfp_t gfp_mask)
 {
-       struct zone **z;
+       struct zoneref *z;
+       struct zone *zone;
        int ret = 1;
 
-       z = zonelist->zones;
-
-       mutex_lock(&zone_scan_mutex);
-       do {
-               if (zone_is_oom_locked(*z)) {
+       spin_lock(&zone_scan_mutex);
+       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_mutex so a
+                * parallel invocation of try_set_zone_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:
-       mutex_unlock(&zone_scan_mutex);
+       spin_unlock(&zone_scan_mutex);
        return ret;
 }
 
@@ -422,21 +496,23 @@ 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;
-
-       mutex_lock(&zone_scan_mutex);
-       do {
-               zone_clear_flag(*z, ZONE_OOM_LOCKED);
-       } while (*(++z) != NULL);
-       mutex_unlock(&zone_scan_mutex);
+       spin_lock(&zone_scan_mutex);
+       for_each_zone_zonelist(zone, z, zonelist, gfp_zone(gfp_mask)) {
+               zone_clear_flag(zone, ZONE_OOM_LOCKED);
+       }
+       spin_unlock(&zone_scan_mutex);
 }
 
 /**
  * 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
  *
  * If we run out of memory, we have the choice between either
  * killing a random task (bad), letting the system crash (worse)
@@ -467,7 +543,7 @@ void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
 
        switch (constraint) {
        case CONSTRAINT_MEMORY_POLICY:
-               oom_kill_process(current, gfp_mask, order, points,
+               oom_kill_process(current, gfp_mask, order, points, NULL,
                                "No available memory (MPOL_BIND)");
                break;
 
@@ -477,7 +553,7 @@ void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
                /* Fall-through */
        case CONSTRAINT_CPUSET:
                if (sysctl_oom_kill_allocating_task) {
-                       oom_kill_process(current, gfp_mask, order, points,
+                       oom_kill_process(current, gfp_mask, order, points, NULL,
                                        "Out of memory (oom_kill_allocating_task)");
                        break;
                }
@@ -486,7 +562,7 @@ retry:
                 * Rambo mode: Shoot down a process and hope it solves whatever
                 * issues we may have.
                 */
-               p = select_bad_process(&points);
+               p = select_bad_process(&points, NULL);
 
                if (PTR_ERR(p) == -1UL)
                        goto out;
@@ -497,7 +573,7 @@ retry:
                        panic("Out of memory and no killable processes...\n");
                }
 
-               if (oom_kill_process(p, points, gfp_mask, order,
+               if (oom_kill_process(p, gfp_mask, order, points, NULL,
                                     "Out of memory"))
                        goto retry;