Re: [PATCH v2 1/2] PCI/IOV: Revert "PCI/IOV: Serialize sysfs sriov_numvfs reads vs writes"
From: Leon Romanovsky
Date: Tue Feb 13 2024 - 12:47:05 EST
On Tue, Feb 13, 2024 at 09:59:54AM -0600, Bjorn Helgaas wrote:
> On Tue, Feb 13, 2024 at 09:34:50AM +0200, Leon Romanovsky wrote:
> > On Mon, Feb 12, 2024 at 02:27:14PM -0600, Bjorn Helgaas wrote:
> > > On Sun, Feb 11, 2024 at 10:48:44AM +0200, Leon Romanovsky wrote:
> > > > On Fri, Feb 09, 2024 at 07:20:28PM -0800, Kuppuswamy Sathyanarayanan wrote:
> > > > > On 2/9/24 3:52 PM, Jim Harris wrote:
> > > > > > If an SR-IOV enabled device is held by vfio, and the device
> > > > > > is removed, vfio will hold device lock and notify userspace
> > > > > > of the removal. If userspace reads the sriov_numvfs sysfs
> > > > > > entry, that thread will be blocked since sriov_numvfs_show()
> > > > > > also tries to acquire the device lock. If that same thread
> > > > > > is responsible for releasing the device to vfio, it results
> > > > > > in a deadlock.
> > > > > >
> > > > > > The proper way to detect a change to the num_VFs value is to
> > > > > > listen for a sysfs event, not to add a device_lock() on the
> > > > > > attribute _show() in the kernel.
> > >
> > > The lock was not about detecting a change; Pierre did this:
> > >
> > > ip monitor dev ${DEVICE} | grep --line-buffered "^${id}:" | while read line; do \
> > > cat ${path}/device/sriov_numvfs; \
> > >
> > > which I assume works by listening for sysfs events.
> >
> > It is not, "ip monitor ..." listens to netlink events emitted by
> > netdev core and not sysfs events. Sysfs events are not involved in
> > this case.
>
> Thanks for correcting my hasty assumption!
>
> > > The problem was that after the event occurred, the sriov_numvfs
> > > read got a stale value (see https://bugzilla.kernel.org/show_bug.cgi?id=202991).
> >
> > Yes, and it is outcome of such cross-subsytem involvement, which
> > is racy by definition. Someone can come with even simpler example of why
> > locking sysfs read and write is not a good idea.
> >
> > For example, let's consider the following scenario with two CPUs and
> > locks on sysfs read and write:
> >
> > CPU1 CPU2
> > echo 1 > ${path}/device/sriov_numvfs
> > context_switch ->
> > cat ${path}/device/sriov_numvfs
> > lock
> > return 0
> > unlock
> > context_switch <-
> > lock
> > set 1
> > unlock
> >
> > CPU1 CPU2
> > echo 1 > ${path}/device/sriov_numvfs
> > lock
> > set 1
> > unlock
> > context_switch ->
> > cat ${path}/device/sriov_numvfs
> > lock
> > return 1
> > unlock
> >
> > So same scenario will return different values if user doesn't protect
> > such case with external to the kernel lock.
> >
> > But if we return back to Pierre report and if you want to provide
> > completely bullet proof solution to solve cross-subsystem interaction,
> > you will need to prohibit device probe till sriov_numvfs update is completed.
> > However, it is overkill for something that is not a real issue.
>
> Pierre wanted to detect the configuration change and learn the new
> num_vfs, which seems like a reasonable thing to do. Is there a way to
> do both via netlink or some other mechanism?
Please pay attention that Pierre listened to specific netdevice and not
to something general. After patch #2 in Jim's series, he will be able to
rely on "udevadm monitor" instead of "ip monitor".
>
> > > So I would drop this sentence because I don't think it accurately
> > > reflects the reason for 35ff867b7657.
> > >
> > > > > Since you are reverting a commit that synchronizes SysFS read
> > > > > /write, please add some comments about why it is not an
> > > > > issue anymore.
> > > >
> > > > It was never an issue, the idea that sysfs read and write should be
> > > > serialized by kernel is not correct by definition.
> > >
> > > I think it *was* an issue. The behavior Pierre observed at was
> > > clearly wrong,
> >
> > I disagree with this sentence.
> >
> > > and we added 35ff867b7657 ("PCI/IOV: Serialize sysfs
> > > sriov_numvfs reads vs writes") to resolve it.
> > >
> > > We should try to avoid reintroducing the problem, so I think we should
> > > probably squash these two patches and describe it as a deadlock fix
> > > instead of dismissing 35ff867b7657 as being based on false premises.
> > >
> > > It would be awesome if you had time to verify that these patches also
> > > resolve the problem you saw, Pierre.
> >
> > They won't resolve his problem, because he is not listening to sysfs
> > events, but rely on something from netdev side.
>
> I guess that means that if we apply this revert, the problem Pierre
> reported will return. Obviously the deadlock is more important than
> the inconsistency Pierre observed, but from the user's point of view
> this will look like a regression.
>
> Maybe listening to netlink and then looking at sysfs isn't the
> "correct" way to do this, but I don't want to just casually break
> existing user code. If we do contemplate doing the revert, at the
> very least we should include specific details about what the user code
> *should* do instead, at the level of the actual commands to use
> instead of "ip monitor dev; cat ${path}/device/sriov_numvfs".
udevadm monitor will do the trick.
Another possible solution is to refactor the code to make sure that
probe on VFs happens only after sriov_numvfs is updated.
Thanks
>
> Bjorn
>