]> git.karo-electronics.de Git - karo-tx-linux.git/commitdiff
freezer: change ptrace_stop/do_signal_stop to use freezable_schedule()
authorOleg Nesterov <oleg@redhat.com>
Fri, 26 Oct 2012 17:46:06 +0000 (19:46 +0200)
committerTejun Heo <tj@kernel.org>
Fri, 26 Oct 2012 21:27:49 +0000 (14:27 -0700)
try_to_freeze_tasks() and cgroup_freezer rely on scheduler locks
to ensure that a task doing STOPPED/TRACED -> RUNNING transition
can't escape freezing. This mostly works, but ptrace_stop() does
not necessarily call schedule(), it can change task->state back to
RUNNING and check freezing() without any lock/barrier in between.

We could add the necessary barrier, but this patch changes
ptrace_stop() and do_signal_stop() to use freezable_schedule().
This fixes the race, freezer_count() and freezer_should_skip()
carefully avoid the race.

And this simplifies the code, try_to_freeze_tasks/update_if_frozen
no longer need to use task_is_stopped_or_traced() checks with the
non trivial assumptions. We can rely on the mechanism which was
specially designed to mark the sleeping task as "frozen enough".

v2: As Tejun pointed out, we can also change get_signal_to_deliver()
and move try_to_freeze() up before 'relock' label.

Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
include/linux/freezer.h
kernel/cgroup_freezer.c
kernel/freezer.c
kernel/power/process.c
kernel/signal.c

index ee899329e65a6e43320ddfbedd38f427f6d39125..8039893bc3ec205698d5c24e3c16435887d4f1ad 100644 (file)
@@ -134,10 +134,9 @@ static inline bool freezer_should_skip(struct task_struct *p)
 }
 
 /*
- * These macros are intended to be used whenever you want allow a task that's
- * sleeping in TASK_UNINTERRUPTIBLE or TASK_KILLABLE state to be frozen. Note
- * that neither return any clear indication of whether a freeze event happened
- * while in this function.
+ * These macros are intended to be used whenever you want allow a sleeping
+ * task to be frozen. Note that neither return any clear indication of
+ * whether a freeze event happened while in this function.
  */
 
 /* Like schedule(), but should not block the freezer. */
index 8a92b0e52099185136ace406dcefa7129b1ba5e1..bedefd9a22df477f25501663e3c05e40dee353b1 100644 (file)
@@ -198,8 +198,7 @@ static void update_if_frozen(struct cgroup *cgroup, struct freezer *freezer)
                         * completion.  Consider it frozen in addition to
                         * the usual frozen condition.
                         */
-                       if (!frozen(task) && !task_is_stopped_or_traced(task) &&
-                           !freezer_should_skip(task))
+                       if (!frozen(task) && !freezer_should_skip(task))
                                goto notyet;
                }
        }
index 11f82a4d4eae07b4d98595852ea50dcf1c028a6b..c38893b0efbaa39d770b11546dba9e97ec53ff51 100644 (file)
@@ -116,17 +116,10 @@ bool freeze_task(struct task_struct *p)
                return false;
        }
 
-       if (!(p->flags & PF_KTHREAD)) {
+       if (!(p->flags & PF_KTHREAD))
                fake_signal_wake_up(p);
-               /*
-                * fake_signal_wake_up() goes through p's scheduler
-                * lock and guarantees that TASK_STOPPED/TRACED ->
-                * TASK_RUNNING transition can't race with task state
-                * testing in try_to_freeze_tasks().
-                */
-       } else {
+       else
                wake_up_state(p, TASK_INTERRUPTIBLE);
-       }
 
        spin_unlock_irqrestore(&freezer_lock, flags);
        return true;
index 87da817f9e132204add38dc73685962081de9dd8..d5a258b60c6fd71aed065a0dd2ea5acadddc991f 100644 (file)
@@ -48,18 +48,7 @@ static int try_to_freeze_tasks(bool user_only)
                        if (p == current || !freeze_task(p))
                                continue;
 
-                       /*
-                        * Now that we've done set_freeze_flag, don't
-                        * perturb a task in TASK_STOPPED or TASK_TRACED.
-                        * It is "frozen enough".  If the task does wake
-                        * up, it will immediately call try_to_freeze.
-                        *
-                        * Because freeze_task() goes through p's scheduler lock, it's
-                        * guaranteed that TASK_STOPPED/TRACED -> TASK_RUNNING
-                        * transition can't race with task state testing here.
-                        */
-                       if (!task_is_stopped_or_traced(p) &&
-                           !freezer_should_skip(p))
+                       if (!freezer_should_skip(p))
                                todo++;
                } while_each_thread(g, p);
                read_unlock(&tasklist_lock);
index 0af8868525d6d1853acfa77eb9f224566ccfb313..5ffb5626e0721d52e7230e02a809d904fb2c2666 100644 (file)
@@ -1908,7 +1908,7 @@ static void ptrace_stop(int exit_code, int why, int clear_code, siginfo_t *info)
                preempt_disable();
                read_unlock(&tasklist_lock);
                preempt_enable_no_resched();
-               schedule();
+               freezable_schedule();
        } else {
                /*
                 * By the time we got the lock, our tracer went away.
@@ -1929,13 +1929,6 @@ static void ptrace_stop(int exit_code, int why, int clear_code, siginfo_t *info)
                read_unlock(&tasklist_lock);
        }
 
-       /*
-        * While in TASK_TRACED, we were considered "frozen enough".
-        * Now that we woke up, it's crucial if we're supposed to be
-        * frozen that we freeze now before running anything substantial.
-        */
-       try_to_freeze();
-
        /*
         * We are back.  Now reacquire the siglock before touching
         * last_siginfo, so that we are sure to have synchronized with
@@ -2092,7 +2085,7 @@ static bool do_signal_stop(int signr)
                }
 
                /* Now we don't run again until woken by SIGCONT or SIGKILL */
-               schedule();
+               freezable_schedule();
                return true;
        } else {
                /*
@@ -2200,15 +2193,14 @@ int get_signal_to_deliver(siginfo_t *info, struct k_sigaction *return_ka,
        if (unlikely(uprobe_deny_signal()))
                return 0;
 
-relock:
        /*
-        * We'll jump back here after any time we were stopped in TASK_STOPPED.
-        * While in TASK_STOPPED, we were considered "frozen enough".
-        * Now that we woke up, it's crucial if we're supposed to be
-        * frozen that we freeze now before running anything substantial.
+        * Do this once, we can't return to user-mode if freezing() == T.
+        * do_signal_stop() and ptrace_stop() do freezable_schedule() and
+        * thus do not need another check after return.
         */
        try_to_freeze();
 
+relock:
        spin_lock_irq(&sighand->siglock);
        /*
         * Every stopped thread goes here after wakeup. Check to see if