security: Fix setting of PF_SUPERPRIV by __capable()
[linux-2.6.git] / mm / oom_kill.c
index 78747af..64e5b4b 100644 (file)
  *  kernel subsystems and hints as to where to find out what things do.
  */
 
+#include <linux/oom.h>
 #include <linux/mm.h>
+#include <linux/err.h>
 #include <linux/sched.h>
 #include <linux/swap.h>
 #include <linux/timex.h>
 #include <linux/jiffies.h>
 #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;
+int sysctl_oom_dump_tasks;
+static DEFINE_SPINLOCK(zone_scan_mutex);
 /* #define DEBUG */
 
 /**
- * oom_badness - calculate a numeric value for how bad this task has been
+ * 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
 unsigned long badness(struct task_struct *p, unsigned long uptime)
 {
        unsigned long points, cpu_time, run_time, s;
-       struct list_head *tsk;
+       struct mm_struct *mm;
+       struct task_struct *child;
 
-       if (!p->mm)
+       task_lock(p);
+       mm = p->mm;
+       if (!mm) {
+               task_unlock(p);
                return 0;
+       }
 
        /*
         * The memory size of the process is the basis for the badness.
         */
-       points = p->mm->total_vm;
+       points = mm->total_vm;
+
+       /*
+        * After this unlock we can no longer dereference local variable `mm'
+        */
+       task_unlock(p);
+
+       /*
+        * swapoff can easily use up all memory, so kill those first.
+        */
+       if (p->flags & PF_SWAPOFF)
+               return ULONG_MAX;
 
        /*
         * Processes which fork a lot of child processes are likely
@@ -64,11 +91,11 @@ unsigned long badness(struct task_struct *p, unsigned long uptime)
         * child is eating the vast majority of memory, adding only half
         * to the parents will make the child our kill candidate of choice.
         */
-       list_for_each(tsk, &p->children) {
-               struct task_struct *chld;
-               chld = list_entry(tsk, struct task_struct, sibling);
-               if (chld->mm != p->mm && chld->mm)
-                       points += chld->mm->total_vm/2 + 1;
+       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);
        }
 
        /*
@@ -102,8 +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;
 
        /*
@@ -112,46 +139,51 @@ 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;
 
        /*
+        * 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.
+        */
+       if (!cpuset_mems_allowed_intersects(current, p))
+               points /= 8;
+
+       /*
         * Adjust the score by oomkilladj.
         */
        if (p->oomkilladj) {
-               if (p->oomkilladj > 0)
+               if (p->oomkilladj > 0) {
+                       if (!points)
+                               points = 1;
                        points <<= p->oomkilladj;
-               else
+               } else
                        points >>= -(p->oomkilladj);
        }
 
 #ifdef DEBUG
-       printk(KERN_DEBUG "OOMkill: task %d (%s) got %d points\n",
+       printk(KERN_DEBUG "OOMkill: task %d (%s) got %lu points\n",
        p->pid, p->comm, points);
 #endif
        return points;
 }
 
 /*
- * Types of limitations to the nodes from which allocations may occur
- */
-#define CONSTRAINT_NONE 1
-#define CONSTRAINT_MEMORY_POLICY 2
-#define CONSTRAINT_CPUSET 3
-
-/*
  * Determine the type of allocation constraint.
  */
-static inline int constrained_alloc(struct zonelist *zonelist, gfp_t gfp_mask)
+static inline enum oom_constraint constrained_alloc(struct zonelist *zonelist,
+                                                   gfp_t gfp_mask)
 {
 #ifdef CONFIG_NUMA
-       struct zone **z;
-       nodemask_t nodes = node_online_map;
-
-       for (z = zonelist->zones; *z; z++)
-               if (cpuset_zone_allowed(*z, gfp_mask))
-                       node_clear((*z)->zone_pgdat->node_id,
-                                       nodes);
+       struct zone *zone;
+       struct zoneref *z;
+       enum zone_type high_zoneidx = gfp_zone(gfp_mask);
+       nodemask_t nodes = node_states[N_HIGH_MEMORY];
+
+       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;
 
@@ -168,7 +200,8 @@ static inline int constrained_alloc(struct zonelist *zonelist, gfp_t gfp_mask)
  *
  * (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;
@@ -178,27 +211,51 @@ static struct task_struct *select_bad_process(unsigned long *ppoints)
        do_posix_clock_monotonic_gettime(&uptime);
        do_each_thread(g, p) {
                unsigned long points;
-               int releasing;
 
-               /* skip the init task with pid == 1 */
-               if (p->pid == 1)
+               /*
+                * skip kernel threads and tasks which have already released
+                * their mm.
+                */
+               if (!p->mm)
                        continue;
-               if (p->oomkilladj == OOM_DISABLE)
+               /* skip the init task */
+               if (is_global_init(p))
                        continue;
-               /* If p's nodes don't overlap ours, it won't help to kill p. */
-               if (!cpuset_excl_nodes_overlap(p))
+               if (mem && !task_in_mem_cgroup(p, mem))
                        continue;
 
                /*
-                * This is in the process of releasing memory so for wait it
-                * to finish before killing some other task by mistake.
+                * This task already has access to memory reserves and is
+                * being killed. Don't allow any other task access to the
+                * memory reserve.
+                *
+                * Note: this may have a chance of deadlock if it gets
+                * blocked waiting for another task which itself is waiting
+                * for memory. Is there a better alternative?
                 */
-               releasing = test_tsk_thread_flag(p, TIF_MEMDIE) ||
-                                               p->flags & PF_EXITING;
-               if (releasing && !(p->flags & PF_DEAD))
+               if (test_tsk_thread_flag(p, TIF_MEMDIE))
                        return ERR_PTR(-1UL);
-               if (p->flags & PF_SWAPOFF)
-                       return p;
+
+               /*
+                * 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->oomkilladj == OOM_DISABLE)
+                       continue;
 
                points = badness(p, uptime.tv_sec);
                if (points > *ppoints || !chosen) {
@@ -206,92 +263,256 @@ static struct task_struct *select_bad_process(unsigned long *ppoints)
                        *ppoints = points;
                }
        } while_each_thread(g, p);
+
        return chosen;
 }
 
 /**
- * We must be careful though to never send SIGKILL a process with
- * CAP_SYS_RAW_IO set, send SIGTERM instead (but 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: 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(task_t *p, const char *message)
+static void __oom_kill_task(struct task_struct *p, int verbose)
 {
-       if (p->pid == 1) {
+       if (is_global_init(p)) {
                WARN_ON(1);
                printk(KERN_WARNING "tried to kill init!\n");
                return;
        }
 
-       task_lock(p);
-       if (!p->mm || p->mm == &init_mm) {
+       if (!p->mm) {
                WARN_ON(1);
                printk(KERN_WARNING "tried to kill an mm-less task!\n");
-               task_unlock(p);
                return;
        }
-       task_unlock(p);
-       printk(KERN_ERR "%s: Killed process %d (%s).\n",
-                               message, p->pid, p->comm);
+
+       if (verbose)
+               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);
 }
 
-static struct mm_struct *oom_kill_task(task_t *p, const char *message)
+static int oom_kill_task(struct task_struct *p)
 {
-       struct mm_struct *mm = get_task_mm(p);
-       task_t * g, * q;
-
-       if (!mm)
-               return NULL;
-       if (mm == &init_mm) {
-               mmput(mm);
-               return NULL;
-       }
+       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)
+               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);
+
+       __oom_kill_task(p, 1);
 
-       __oom_kill_task(p, message);
        /*
         * kill all processes that share the ->mm (i.e. all threads),
-        * but are in a different thread group
+        * but are in a different thread group. Don't let them have access
+        * to memory reserves though, otherwise we might deplete all memory.
         */
-       do_each_thread(g, q)
-               if (q->mm == mm && q->tgid != p->tgid)
-                       __oom_kill_task(q, message);
-       while_each_thread(g, q);
+       do_each_thread(g, q) {
+               if (q->mm == mm && !same_thread_group(q, p))
+                       force_sig(SIGKILL, q);
+       } while_each_thread(g, q);
 
-       return mm;
+       return 0;
 }
 
-static struct mm_struct *oom_kill_process(struct task_struct *p,
-                               unsigned long points, const char *message)
+static int oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
+                           unsigned long points, struct mem_cgroup *mem,
+                           const char *message)
 {
-       struct mm_struct *mm;
        struct task_struct *c;
-       struct list_head *tsk;
 
-       printk(KERN_ERR "Out of Memory: Kill process %d (%s) score %li and "
-               "children.\n", p->pid, p->comm, points);
+       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();
+               if (sysctl_oom_dump_tasks)
+                       dump_tasks(mem);
+       }
+
+       /*
+        * 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);
+               return 0;
+       }
+
+       printk(KERN_ERR "%s: kill process %d (%s) score %li or a child\n",
+                                       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;
-               mm = oom_kill_task(c, message);
-               if (mm)
-                       return mm;
+               if (!oom_kill_task(c))
+                       return 0;
        }
-       return oom_kill_task(p, message);
+       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)
+{
+       return blocking_notifier_chain_register(&oom_notify_list, nb);
+}
+EXPORT_SYMBOL_GPL(register_oom_notifier);
+
+int unregister_oom_notifier(struct notifier_block *nb)
+{
+       return blocking_notifier_chain_unregister(&oom_notify_list, nb);
+}
+EXPORT_SYMBOL_GPL(unregister_oom_notifier);
+
+/*
+ * Try to acquire the OOM killer lock for the zones in zonelist.  Returns zero
+ * 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, gfp_t gfp_mask)
+{
+       struct zoneref *z;
+       struct zone *zone;
+       int ret = 1;
+
+       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;
+               }
+       }
+
+       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);
+       }
+
+out:
+       spin_unlock(&zone_scan_mutex);
+       return ret;
+}
+
+/*
+ * Clears the ZONE_OOM_LOCKED flag for all zones in the zonelist so that failed
+ * allocation attempts with zonelists containing them may now recall the OOM
+ * killer, if necessary.
+ */
+void clear_zonelist_oom(struct zonelist *zonelist, gfp_t gfp_mask)
+{
+       struct zoneref *z;
+       struct zone *zone;
+
+       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);
 }
 
 /**
- * oom_kill - kill the "best" process when we run out of memory
+ * 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)
@@ -300,42 +521,48 @@ static struct mm_struct *oom_kill_process(struct task_struct *p,
  */
 void out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask, int order)
 {
-       struct mm_struct *mm = NULL;
-       task_t *p;
+       struct task_struct *p;
        unsigned long points = 0;
+       unsigned long freed = 0;
+       enum oom_constraint constraint;
 
-       if (printk_ratelimit()) {
-               printk("oom-killer: gfp_mask=0x%x, order=%d\n",
-                       gfp_mask, order);
-               dump_stack();
-               show_mem();
-       }
+       blocking_notifier_call_chain(&oom_notify_list, 0, &freed);
+       if (freed > 0)
+               /* Got some memory back in the last second. */
+               return;
 
-       cpuset_lock();
-       read_lock(&tasklist_lock);
+       if (sysctl_panic_on_oom == 2)
+               panic("out of memory. Compulsory panic_on_oom is selected.\n");
 
        /*
         * Check if there were limitations on the allocation (only relevant for
         * NUMA) that may require different handling.
         */
-       switch (constrained_alloc(zonelist, gfp_mask)) {
+       constraint = constrained_alloc(zonelist, gfp_mask);
+       read_lock(&tasklist_lock);
+
+       switch (constraint) {
        case CONSTRAINT_MEMORY_POLICY:
-               mm = oom_kill_process(current, points,
+               oom_kill_process(current, gfp_mask, order, points, NULL,
                                "No available memory (MPOL_BIND)");
                break;
 
-       case CONSTRAINT_CPUSET:
-               mm = oom_kill_process(current, points,
-                               "No available memory in cpuset");
-               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, NULL,
+                                       "Out of memory (oom_kill_allocating_task)");
+                       break;
+               }
 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;
@@ -343,12 +570,11 @@ retry:
                /* Found nothing?!?! Either we hang forever, or we panic. */
                if (!p) {
                        read_unlock(&tasklist_lock);
-                       cpuset_unlock();
                        panic("Out of memory and no killable processes...\n");
                }
 
-               mm = oom_kill_process(p, points, "Out of memory");
-               if (!mm)
+               if (oom_kill_process(p, gfp_mask, order, points, NULL,
+                                    "Out of memory"))
                        goto retry;
 
                break;
@@ -356,9 +582,6 @@ retry:
 
 out:
        read_unlock(&tasklist_lock);
-       cpuset_unlock();
-       if (mm)
-               mmput(mm);
 
        /*
         * Give "p" a good chance of killing itself before we