coredump: ensure the fpu state is flushed for proper multi-threaded core dump

Nalluru reported hitting the BUG_ON(__thread_has_fpu(tsk)) in
arch/x86/kernel/xsave.c:__sanitize_i387_state() during the coredump
of a multi-threaded application.

A look at the exit seqeuence shows that other threads can still be on the
runqueue potentially at the below shown exit_mm() code snippet:

		if (atomic_dec_and_test(&core_state->nr_threads))
			complete(&core_state->startup);

===> other threads can still be active here, but we notify the thread
===> dumping core to wakeup from the coredump_wait() after the last thread
===> joins this point. Core dumping thread will continue dumping
===> all the threads state to the core file.

		for (;;) {
			set_task_state(tsk, TASK_UNINTERRUPTIBLE);
			if (!self.task) /* see coredump_finish() */
				break;
			schedule();
		}

As some of those threads are on the runqueue and didn't call schedule() yet,
their fpu state is still active in the live registers and the thread
proceeding with the coredump will hit the above mentioned BUG_ON while
trying to dump other threads fpustate to the coredump file.

BUG_ON() in arch/x86/kernel/xsave.c:__sanitize_i387_state() is
in the code paths for processors supporting xsaveopt. With or without
xsaveopt, multi-threaded coredump is broken and maynot contain
the correct fpustate at the time of exit.

In coredump_wait(), wait for all the threads to be come inactive, so
that we are sure all the extended register state is flushed to
the memory, so that it can be reliably copied to the core file.

Reported-by: Suresh Nalluru <suresh@aristanetworks.com>
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Link: http://lkml.kernel.org/r/1336692811-30576-2-git-send-email-suresh.b.siddha@intel.com
Acked-by: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>

diff --git a/fs/exec.c b/fs/exec.c
index b1fd2025e59a1e0ea0a2f475b812987dac176e9f..8e2ddebaccddb0547a9c6122a8d89429a56bdf99 100644 (file)
--- a/fs/exec.c
+++ b/fs/exec.c
@@ -1930,8 +1930,21 @@ static int coredump_wait(int exit_code, struct core_state *core_state)
                core_waiters = zap_threads(tsk, mm, core_state, exit_code);
        up_write(&mm->mmap_sem);
 
-       if (core_waiters > 0)
+       if (core_waiters > 0) {
+               struct core_thread *ptr;
+
                wait_for_completion(&core_state->startup);
+               /*
+                * Wait for all the threads to become inactive, so that
+                * all the thread context (extended register state, like
+                * fpu etc) gets copied to the memory.
+                */
+               ptr = core_state->dumper.next;
+               while (ptr != NULL) {
+                       wait_task_inactive(ptr->task, 0);
+                       ptr = ptr->next;
+               }
+       }
 
        return core_waiters;
 }