Re: [PATCH v2 6/7] libnvdimm/bus: Fix wait_nvdimm_bus_probe_idle() ABBA deadlock

[Dan Williams]

On Wed, Jul 17, 2019 at 7:05 PM Sasha Levin <sashal@xxxxxxxxxx> wrote:
>
> On Wed, Jul 17, 2019 at 06:08:21PM -0700, Dan Williams wrote:
> >A multithreaded namespace creation/destruction stress test currently
> >deadlocks with the following lockup signature:
> >
> >    INFO: task ndctl:2924 blocked for more than 122 seconds.
> >          Tainted: G           OE     5.2.0-rc4+ #3382
> >    "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
> >    ndctl           D    0  2924   1176 0x00000000
> >    Call Trace:
> >     ? __schedule+0x27e/0x780
> >     schedule+0x30/0xb0
> >     wait_nvdimm_bus_probe_idle+0x8a/0xd0 [libnvdimm]
> >     ? finish_wait+0x80/0x80
> >     uuid_store+0xe6/0x2e0 [libnvdimm]
> >     kernfs_fop_write+0xf0/0x1a0
> >     vfs_write+0xb7/0x1b0
> >     ksys_write+0x5c/0xd0
> >     do_syscall_64+0x60/0x240
> >
> >     INFO: task ndctl:2923 blocked for more than 122 seconds.
> >           Tainted: G           OE     5.2.0-rc4+ #3382
> >     "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
> >     ndctl           D    0  2923   1175 0x00000000
> >     Call Trace:
> >      ? __schedule+0x27e/0x780
> >      ? __mutex_lock+0x489/0x910
> >      schedule+0x30/0xb0
> >      schedule_preempt_disabled+0x11/0x20
> >      __mutex_lock+0x48e/0x910
> >      ? nvdimm_namespace_common_probe+0x95/0x4d0 [libnvdimm]
> >      ? __lock_acquire+0x23f/0x1710
> >      ? nvdimm_namespace_common_probe+0x95/0x4d0 [libnvdimm]
> >      nvdimm_namespace_common_probe+0x95/0x4d0 [libnvdimm]
> >      __dax_pmem_probe+0x5e/0x210 [dax_pmem_core]
> >      ? nvdimm_bus_probe+0x1d0/0x2c0 [libnvdimm]
> >      dax_pmem_probe+0xc/0x20 [dax_pmem]
> >      nvdimm_bus_probe+0x90/0x2c0 [libnvdimm]
> >      really_probe+0xef/0x390
> >      driver_probe_device+0xb4/0x100
> >
> >In this sequence an 'nd_dax' device is being probed and trying to take
> >the lock on its backing namespace to validate that the 'nd_dax' device
> >indeed has exclusive access to the backing namespace. Meanwhile, another
> >thread is trying to update the uuid property of that same backing
> >namespace. So one thread is in the probe path trying to acquire the
> >lock, and the other thread has acquired the lock and tries to flush the
> >probe path.
> >
> >Fix this deadlock by not holding the namespace device_lock over the
> >wait_nvdimm_bus_probe_idle() synchronization step. In turn this requires
> >the device_lock to be held on entry to wait_nvdimm_bus_probe_idle() and
> >subsequently dropped internally to wait_nvdimm_bus_probe_idle().
> >
> >Cc: <stable@xxxxxxxxxxxxxxx>
> >Fixes: bf9bccc14c05 ("libnvdimm: pmem label sets and namespace instantiation")
> >Cc: Vishal Verma <vishal.l.verma@xxxxxxxxx>
> >Tested-by: Jane Chu <jane.chu@xxxxxxxxxx>
> >Signed-off-by: Dan Williams <dan.j.williams@xxxxxxxxx>
>
> Hi Dan,
>
> The way these patches are split, when we take them to stable this patch
> won't apply because it wants "libnvdimm/bus: Prepare the nd_ioctl() path
> to be re-entrant".
>
> If you were to send another iteration of this patchset, could you please
> re-order the patches so they will apply cleanly to stable? this will
> help with reducing mail exchanges later on and possibly a mis-merge into
> stable.
>
> If not, this should serve as a reference for future us to double check
> the backport.

Oh we should backport all of them. I'll tag that one for -stable as
well. It's a hard pre-requisite for the fix.