If panic_on_oops is not set and oops happens inside workqueue kthread,
kernel kills this kthread. Current patch fixes recursive GPF which
happens when wq_worker_sleeping() function unconditionally accesses
the NULL kthread->vfork_done ptr thru kthread_data() -> to_kthread().
The stack is the following:
[<ffffffff81397f75>] dump_stack+0x68/0x93
[<ffffffff8106954b>] ? do_exit+0x7ab/0xc10
[<ffffffff8108fd73>] __schedule_bug+0x83/0xe0
[<ffffffff81716d5a>] __schedule+0x7ea/0xba0
[<ffffffff810c864f>] ? vprintk_default+0x1f/0x30
[<ffffffff8116a63c>] ? printk+0x48/0x50
[<ffffffff81717150>] schedule+0x40/0x90
[<ffffffff8106976a>] do_exit+0x9ca/0xc10
[<ffffffff810c8e3d>] ? kmsg_dump+0x11d/0x190
[<ffffffff810c8d37>] ? kmsg_dump+0x17/0x190
[<ffffffff81021ee9>] oops_end+0x99/0xd0
[<ffffffff81052da5>] no_context+0x185/0x3e0
[<ffffffff81053083>] __bad_area_nosemaphore+0x83/0x1c0
[<ffffffff810c820e>] ? vprintk_emit+0x25e/0x530
[<ffffffff810531d4>] bad_area_nosemaphore+0x14/0x20
[<ffffffff8105355c>] __do_page_fault+0xac/0x570
[<ffffffff810c66fe>] ? console_trylock+0x1e/0xe0
[<ffffffff81002036>] ? trace_hardirqs_off_thunk+0x1a/0x1c
[<ffffffff81053a2c>] do_page_fault+0xc/0x10
[<ffffffff8171f812>] page_fault+0x22/0x30
[<ffffffff81089bc3>] ? kthread_data+0x33/0x40
[<ffffffff8108427e>] ? wq_worker_sleeping+0xe/0x80
[<ffffffff817169eb>] __schedule+0x47b/0xba0
[<ffffffff81717150>] schedule+0x40/0x90
[<ffffffff8106957d>] do_exit+0x7dd/0xc10
[<ffffffff81021ee9>] oops_end+0x99/0xd0
kthread->vfork_done is zeroed out on the following path:
do_exit()
exit_mm()
mm_release()
complete_vfork_done()
In order to fix a bug dead tasks must be ignored.
Signed-off-by: Roman Pen <[email protected]>
Cc: Andy Lutomirski <[email protected]>
Cc: Oleg Nesterov <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: Thomas Gleixner <[email protected]>
Cc: Ingo Molnar <[email protected]>
Cc: Tejun Heo <[email protected]>
Cc: [email protected]
---
v4:
o instead of TASK_DEAD state use more generic PF_EXITING flag.
o same dead task check should be also done for wk_worker_waking_up().
With this we try to avoid a case, when we scheduled back to a task,
which was just in do_exit and have set the PF_EXITING flag.
v3:
o minor comment and coding style fixes.
v2:
o put a task->state check directly into a wq_worker_sleeping() function
instead of changing the __schedule().
kernel/workqueue.c | 24 ++++++++++++++++++++++--
1 file changed, 22 insertions(+), 2 deletions(-)
diff --git a/kernel/workqueue.c b/kernel/workqueue.c
index 9dc7ac5101e0..23f2d764cebf 100644
--- a/kernel/workqueue.c
+++ b/kernel/workqueue.c
@@ -851,7 +851,17 @@ static void wake_up_worker(struct worker_pool *pool)
*/
void wq_worker_waking_up(struct task_struct *task, int cpu)
{
- struct worker *worker = kthread_data(task);
+ struct worker *worker;
+
+ if (task->flags & PF_EXITING) {
+ /*
+ * Careful here, t->vfork_done is zeroed out for
+ * almost dead tasks, do not touch kthread_data().
+ */
+ return;
+ }
+
+ worker = kthread_data(task);
if (!(worker->flags & WORKER_NOT_RUNNING)) {
WARN_ON_ONCE(worker->pool->cpu != cpu);
@@ -875,9 +885,19 @@ void wq_worker_waking_up(struct task_struct *task, int cpu)
*/
struct task_struct *wq_worker_sleeping(struct task_struct *task)
{
- struct worker *worker = kthread_data(task), *to_wakeup = NULL;
+ struct worker *worker, *to_wakeup = NULL;
struct worker_pool *pool;
+ if (task->flags & PF_EXITING) {
+ /*
+ * Careful here, t->vfork_done is zeroed out for
+ * almost dead tasks, do not touch kthread_data().
+ */
+ return NULL;
+ }
+
+ worker = kthread_data(task);
+
/*
* Rescuers, which may not have all the fields set up like normal
* workers, also reach here, let's not access anything before
--
2.9.3
On 10/25, Roman Pen wrote:
>
> void wq_worker_waking_up(struct task_struct *task, int cpu)
> {
> - struct worker *worker = kthread_data(task);
> + struct worker *worker;
> +
> + if (task->flags & PF_EXITING) {
> + /*
> + * Careful here, t->vfork_done is zeroed out for
> + * almost dead tasks, do not touch kthread_data().
> + */
> + return;
> + }
> +
> + worker = kthread_data(task);
>
> if (!(worker->flags & WORKER_NOT_RUNNING)) {
> WARN_ON_ONCE(worker->pool->cpu != cpu);
> @@ -875,9 +885,19 @@ void wq_worker_waking_up(struct task_struct *task, int cpu)
> */
> struct task_struct *wq_worker_sleeping(struct task_struct *task)
> {
> - struct worker *worker = kthread_data(task), *to_wakeup = NULL;
> + struct worker *worker, *to_wakeup = NULL;
> struct worker_pool *pool;
>
> + if (task->flags & PF_EXITING) {
> + /*
> + * Careful here, t->vfork_done is zeroed out for
> + * almost dead tasks, do not touch kthread_data().
> + */
> + return NULL;
> + }
> +
> + worker = kthread_data(task);
> +
Please see the patch I send a minute ago. With that patch we do not need
this fix, kthread_data() no longer uses ->vfork_done.
Oleg.
On Wed, Oct 26, 2016 at 4:18 PM, Oleg Nesterov <[email protected]> wrote:
> On 10/25, Roman Pen wrote:
>>
>> void wq_worker_waking_up(struct task_struct *task, int cpu)
>> {
>> - struct worker *worker = kthread_data(task);
>> + struct worker *worker;
>> +
>> + if (task->flags & PF_EXITING) {
>> + /*
>> + * Careful here, t->vfork_done is zeroed out for
>> + * almost dead tasks, do not touch kthread_data().
>> + */
>> + return;
>> + }
>> +
>> + worker = kthread_data(task);
>>
>> if (!(worker->flags & WORKER_NOT_RUNNING)) {
>> WARN_ON_ONCE(worker->pool->cpu != cpu);
>> @@ -875,9 +885,19 @@ void wq_worker_waking_up(struct task_struct *task, int cpu)
>> */
>> struct task_struct *wq_worker_sleeping(struct task_struct *task)
>> {
>> - struct worker *worker = kthread_data(task), *to_wakeup = NULL;
>> + struct worker *worker, *to_wakeup = NULL;
>> struct worker_pool *pool;
>>
>> + if (task->flags & PF_EXITING) {
>> + /*
>> + * Careful here, t->vfork_done is zeroed out for
>> + * almost dead tasks, do not touch kthread_data().
>> + */
>> + return NULL;
>> + }
>> +
>> + worker = kthread_data(task);
>> +
>
> Please see the patch I send a minute ago. With that patch we do not need
> this fix, kthread_data() no longer uses ->vfork_done.
Perfect. Thanks. Oleg, could you please add me to CC when you finally send
a patch.
--
Roman