On Fri, Jul 8, 2016 at 10:30 PM, Patrick McLean <patrickm@xxxxxxxxxx> wrote:
> On Fri, Jul 8, 2016 at 11:46 AM, Ilya Dryomov <idryomov@xxxxxxxxx> wrote:
>> On Fri, Jul 8, 2016 at 7:51 PM, Patrick McLean <patrickm@xxxxxxxxxx> wrote:
>>> On Fri, Jul 8, 2016 at 4:40 AM, Ilya Dryomov <idryomov@xxxxxxxxx> wrote:
>>>> On Fri, Jul 8, 2016 at 3:28 AM, Patrick McLean <patrickm@xxxxxxxxxx> wrote:
>>>>> This is on linus git master as of 2016/07/01
>>>>>
>>>>> These appear to be two separate deadlocks, one on a a map operation,
>>>>> and one on an unmap operation. We can reproduce these pretty
>>>>> regularly, but it seems like there is some sort of race condition, as
>>>>> it happens no where near every time.
>>>>>
>>>
>>>>
>>>> It's a single deadlock between "rbd map" and a kworker thread, a later
>>>> "rbd unmap" is just a victim.
>>>>
>>>> Are you mapping the same image more than once?
>>>>
>>>
>>> We shouldn't be, there is a higher-level locking system that is
>>> supposed to prevent that.
>>
>> It's actually allowed, I'm just trying to get an idea of what was going
>> on.
>>
>> I spoke too soon. The trace of pid 14109 is inconsistent - the entries
>> in the middle don't make sense. Do you have a different set of traces
>> or are they all the same?
>
> Here is another backtrace. It seems that once it happens to a
> particular image once, it always does the same thing. We have not
> figured out what triggers the first hang.
>
> # grep -C 40 rbd /proc/*/stack
> /proc/5295/stack-[<ffffffff8111c7c4>] __queue_work+0x144/0x420
> /proc/5295/stack-[<ffffffff8112c1c0>] default_wake_function+0x0/0x10
> /proc/5295/stack:[<ffffffffa01621ee>]
> rbd_obj_method_sync.constprop.48+0x1be/0x290 [rbd]
> /proc/5295/stack:[<ffffffffa0162aea>] rbd_dev_header_info+0x15a/0x930 [rbd]
> /proc/5295/stack:[<ffffffffa0163450>] rbd_watch_cb+0x0/0xa0 [rbd]
> /proc/5295/stack:[<ffffffffa0163586>]
> rbd_dev_image_probe.part.42+0x96/0x910 [rbd]
> /proc/5295/stack:[<ffffffffa016240e>] rbd_dev_image_id+0x14e/0x1b0 [rbd]
> /proc/5295/stack:[<ffffffffa016428f>] do_rbd_add.isra.43+0x48f/0xbb0 [rbd]
> /proc/5295/stack-[<ffffffff8123ba27>] __kmalloc+0x27/0x170
> /proc/5295/stack-[<ffffffff812b7f3a>] kernfs_fop_write+0x10a/0x190
> /proc/5295/stack-[<ffffffff8124dc63>] __vfs_write+0x23/0x120
> /proc/5295/stack-[<ffffffff8124e8f3>] vfs_write+0xb3/0x1a0
> /proc/5295/stack-[<ffffffff8124fbd2>] SyS_write+0x42/0xa0
> /proc/5295/stack-[<ffffffff81816f72>] entry_SYSCALL_64_fastpath+0x1a/0xa4
> /proc/5295/stack-[<ffffffffffffffff>] 0xffffffffffffffff
>
> # ps aux | grep 5295
> root 5295 0.0 0.0 246704 10352 pts/4 Sl+ 20:20 0:00 rbd
> map --pool XXXXXXX XXXXXXXX
Anything in dmesg? It looks like rbd map is waiting for a reply to one
of the header read-in requests and is never woken up. rbd map holds
a semaphore, blocking the kworker, which blocks rbd unmap.
Can you set "debug ms = 1" on all OSDs and try to reproduce?
I'd also like to see the content of /sys/kernel/debug/ceph/*/osdc when
it happens again.
How long does it take to hit it on average?
Thanks,
Ilya
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
