On Wed, Feb 14, 2024 at 05:04:08PM +0000, Jim Harris wrote:
> On Wed, Feb 14, 2024 at 09:16:18AM +0200, Leon Romanovsky wrote:
> > On Tue, Feb 13, 2024 at 01:45:56PM -0600, Bjorn Helgaas wrote:
> > > On Tue, Feb 13, 2024 at 07:46:02PM +0200, Leon Romanovsky wrote:
> > > > On Tue, Feb 13, 2024 at 09:59:54AM -0600, Bjorn Helgaas wrote:
> > > > ...
> > >
> > > > > 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.
> > >
> > > I like the idea of refactoring it so as to preserve the existing
> > > ordering while also fixing the deadlock.
> >
> > I think something like this will be enough (not tested). It will et the number of VFs
> > before we make VFs visible to probe:
>
> I'll push a v3, replacing the second patch with this one instead. Although
> based on this discussion it seems we're moving towards squashing the revert
> with Leon's suggested patch. Bjorn, I'll assume you're still OK with just
> squashing these on your end.
Yep.
> I would like some input on how to actually test this though.
> Presumably we see some event on device PF and we want to make sure
> if we read PF/device/sriov_numvfs that we see the updated value. But
> the only type of event I think we can expect is the PF's
> sriov_numvfs CHANGE event.
>
> Is there any way for VFs to be created outside of writing to the
> sriov_numvfs sysfs file? My understanding is some older
> devices/drivers will auto-create VFs when the PF is initialized, but
> it wasn't clear from the bug report whether that was part of the
> configuration here. Pierre, do you have any recollection on this?
>
> Or maybe testing for this case just means compile and verify with
> udevadm monitor that we see the CHANGE event before any of the VFs
> are actually created...
I just want to make sure that Pierre's existing code continues to work
unchanged.
Ideally we could revert 35ff867b7657 ("PCI/IOV: Serialize sysfs
sriov_numvfs reads vs writes"), reproduce the problem with the shell
script attached to https://bugzilla.kernel.org/show_bug.cgi?id=202991
(I assume Pierre used a /sys/.../sriov_numvfs write to trigger the
change).
Then we could verify that with 35ff867b7657 still reverted but the
change below added, the problem is no longer reproducible.
> > diff --git a/drivers/pci/iov.c b/drivers/pci/iov.c
> > index aaa33e8dc4c9..0cdfaae80594 100644
> > --- a/drivers/pci/iov.c
> > +++ b/drivers/pci/iov.c
> > @@ -679,12 +679,14 @@ static int sriov_enable(struct pci_dev *dev, int nr_virtfn)
> > msleep(100);
> > pci_cfg_access_unlock(dev);
> >
> > + iov->num_VFs = nr_virtfn;
> > rc = sriov_add_vfs(dev, initial);
> > - if (rc)
> > + if (rc) {
> > + iov->num_VFs = 0;
> > goto err_pcibios;
> > + }
> >
> > kobject_uevent(&dev->dev.kobj, KOBJ_CHANGE);
> > - iov->num_VFs = nr_virtfn;
> >
> > return 0;
> >
