Re: [PATCH] scsi: storvsc: Fix a panic in the hibernation procedure

[Ming Lei]

On Tue, Apr 21, 2020 at 08:08:07PM -0700, Paul E. McKenney wrote:
> On Wed, Apr 22, 2020 at 10:01:34AM +0800, Ming Lei wrote:
> > On Wed, Apr 22, 2020 at 01:48:25AM +0000, Dexuan Cui wrote:
> > > > From: Ming Lei <ming.lei@xxxxxxxxxx>
> > > > Sent: Tuesday, April 21, 2020 6:28 PM
> > > > To: Dexuan Cui <decui@xxxxxxxxxxxxx>
> > > > 
> > > > On Tue, Apr 21, 2020 at 05:17:24PM -0700, Dexuan Cui wrote:
> > > > > During hibernation, the sdevs are suspended automatically in
> > > > > drivers/scsi/scsi_pm.c before storvsc_suspend(), so after
> > > > > storvsc_suspend(), there is no disk I/O from the file systems, but there
> > > > > can still be disk I/O from the kernel space, e.g. disk_check_events() ->
> > > > > sr_block_check_events() -> cdrom_check_events() can still submit I/O
> > > > > to the storvsc driver, which causes a paic of NULL pointer dereference,
> > > > > since storvsc has closed the vmbus channel in storvsc_suspend(): refer
> > > > > to the below links for more info:
> > > > >
> > > > > Fix the panic by blocking/unblocking all the I/O queues properly.
> > > > >
> > > > > Note: this patch depends on another patch "scsi: core: Allow the state
> > > > > change from SDEV_QUIESCE to SDEV_BLOCK" (refer to the second link
> > > > above).
> > > > >
> > > > > Fixes: 56fb10585934 ("scsi: storvsc: Add the support of hibernation")
> > > > > Signed-off-by: Dexuan Cui <decui@xxxxxxxxxxxxx>
> > > > > ---
> > > > >  drivers/scsi/storvsc_drv.c | 10 ++++++++++
> > > > >  1 file changed, 10 insertions(+)
> > > > >
> > > > > diff --git a/drivers/scsi/storvsc_drv.c b/drivers/scsi/storvsc_drv.c
> > > > > index fb41636519ee..fd51d2f03778 100644
> > > > > --- a/drivers/scsi/storvsc_drv.c
> > > > > +++ b/drivers/scsi/storvsc_drv.c
> > > > > @@ -1948,6 +1948,11 @@ static int storvsc_suspend(struct hv_device
> > > > *hv_dev)
> > > > >  	struct storvsc_device *stor_device = hv_get_drvdata(hv_dev);
> > > > >  	struct Scsi_Host *host = stor_device->host;
> > > > >  	struct hv_host_device *host_dev = shost_priv(host);
> > > > > +	int ret;
> > > > > +
> > > > > +	ret = scsi_host_block(host);
> > > > > +	if (ret)
> > > > > +		return ret;
> > > > >
> > > > >  	storvsc_wait_to_drain(stor_device);
> > > > >
> > > > > @@ -1968,10 +1973,15 @@ static int storvsc_suspend(struct hv_device
> > > > *hv_dev)
> > > > >
> > > > >  static int storvsc_resume(struct hv_device *hv_dev)
> > > > >  {
> > > > > +	struct storvsc_device *stor_device = hv_get_drvdata(hv_dev);
> > > > > +	struct Scsi_Host *host = stor_device->host;
> > > > >  	int ret;
> > > > >
> > > > >  	ret = storvsc_connect_to_vsp(hv_dev, storvsc_ringbuffer_size,
> > > > >  				     hv_dev_is_fc(hv_dev));
> > > > > +	if (!ret)
> > > > > +		ret = scsi_host_unblock(host, SDEV_RUNNING);
> > > > > +
> > > > >  	return ret;
> > > > >  }
> > > > 
> > > > scsi_host_block() is actually too heavy for just avoiding
> > > > scsi internal command, which can be done simply by one atomic
> > > > variable.
> > > > 
> > > > Not mention scsi_host_block() is implemented too clumsy because
> > > > nr_luns * synchronize_rcu() are required in scsi_host_block(),
> > > > which should have been optimized to just one.
> > > > 
> > > > Also scsi_device_quiesce() is heavy too, still takes 2
> > > > synchronize_rcu() for one LUN.
> > > > 
> > > > That is said SCSI suspend may take (3 * nr_luns) sysnchronize_rcu() in
> > > > case that the HBA's suspend handler needs scsi_host_block().
> > > > 
> > > > Thanks,
> > > > Ming
> > > 
> > > When we're in storvsc_suspend(), all the userspace processes have been
> > > frozen and all the file systems have been flushed, and there should not
> > > be too much I/O from the kernel space, so IMO scsi_host_block() should be
> > > pretty fast here. 
> > 
> > I guess it depends on RCU's implementation, so CC RCU guys.
> > 
> > Hello Paul & Josh,
> > 
> > Could you clarify that if sysnchronize_rcu becomes quickly during
> > system suspend?
> 
> Once you have all but one CPU offlined, it becomes extremely fast, as
> in roughly a no-op (which is an idea of Josh's from back in the day).
> But if there is more than one CPU online, then synchronize_rcu() still
> takes on the order of several to several tens of jiffies.
> 
> So, yes, in some portions of system suspend, synchronize_rcu() becomes
> very fast indeed.

Hi Paul,

Thanks for your clarification.

In system suspend path, device is suspended before suspend_disable_secondary_cpus(), 
so I guess synchronize_rcu() is not quick enough even though user space
processes and some kernel threads are frozen.

Thanks,
Ming