perf: Fix find_get_context() vs perf_event_exit_task() race
authorOleg Nesterov <oleg@redhat.com>
Wed, 19 Jan 2011 18:22:07 +0000 (19:22 +0100)
committerIngo Molnar <mingo@elte.hu>
Wed, 19 Jan 2011 19:04:27 +0000 (20:04 +0100)
find_get_context() must not install the new perf_event_context
if the task has already passed perf_event_exit_task().

If nothing else, this means the memory leak. Initially
ctx->refcount == 2, it is supposed that
perf_event_exit_task_context() should participate and do the
necessary put_ctx().

find_lively_task_by_vpid() checks PF_EXITING but this buys
nothing, by the time we call find_get_context() this task can be
already dead. To the point, cmpxchg() can succeed when the task
has already done the last schedule().

Change find_get_context() to populate task->perf_event_ctxp[]
under task->perf_event_mutex, this way we can trust PF_EXITING
because perf_event_exit_task() takes the same mutex.

Also, change perf_event_exit_task_context() to use
rcu_dereference(). Probably this is not strictly needed, but
with or without this change find_get_context() can race with
setup_new_exec()->perf_event_exit_task(), rcu_dereference()
looks better.

Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Alan Stern <stern@rowland.harvard.edu>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Prasad <prasad@linux.vnet.ibm.com>
Cc: Roland McGrath <roland@redhat.com>
LKML-Reference: <20110119182207.GB12183@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
kernel/perf_event.c

index 84522c7969870fa3a6024086c6a3559546b0e9db..4ec55ef5810cd98a8fdfff3a85c0d02e4de39c20 100644 (file)
@@ -2201,13 +2201,6 @@ find_lively_task_by_vpid(pid_t vpid)
        if (!task)
                return ERR_PTR(-ESRCH);
 
-       /*
-        * Can't attach events to a dying task.
-        */
-       err = -ESRCH;
-       if (task->flags & PF_EXITING)
-               goto errout;
-
        /* Reuse ptrace permission checks for now. */
        err = -EACCES;
        if (!ptrace_may_access(task, PTRACE_MODE_READ))
@@ -2268,14 +2261,27 @@ retry:
 
                get_ctx(ctx);
 
-               if (cmpxchg(&task->perf_event_ctxp[ctxn], NULL, ctx)) {
-                       /*
-                        * We raced with some other task; use
-                        * the context they set.
-                        */
+               err = 0;
+               mutex_lock(&task->perf_event_mutex);
+               /*
+                * If it has already passed perf_event_exit_task().
+                * we must see PF_EXITING, it takes this mutex too.
+                */
+               if (task->flags & PF_EXITING)
+                       err = -ESRCH;
+               else if (task->perf_event_ctxp[ctxn])
+                       err = -EAGAIN;
+               else
+                       rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
+               mutex_unlock(&task->perf_event_mutex);
+
+               if (unlikely(err)) {
                        put_task_struct(task);
                        kfree(ctx);
-                       goto retry;
+
+                       if (err == -EAGAIN)
+                               goto retry;
+                       goto errout;
                }
        }
 
@@ -6127,7 +6133,7 @@ static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
         * scheduled, so we are now safe from rescheduling changing
         * our context.
         */
-       child_ctx = child->perf_event_ctxp[ctxn];
+       child_ctx = rcu_dereference(child->perf_event_ctxp[ctxn]);
        task_ctx_sched_out(child_ctx, EVENT_ALL);
 
        /*