On Tue, Jun 04, 2019 at 02:36:17PM -0600, Alex Williamson wrote:
> On Tue, 04 Jun 2019 11:26:42 -0600
> Alex Williamson <alex.williamson@xxxxxxxxxx> wrote:
>
> > Commit 975bb8b4dc93 ("PCI/IOV: Use VF0 cached config space size for
> > other VFs") attempts to cache the config space size of VF0 to re-use
> > for all other VFs, but the cache is setup before the call to
> > pci_setup_device(), where we use set_pcie_port_type() to setup the
> > pcie_cap field on the struct pci_dev. Without pcie_cap configured,
> > pci_cfg_space_size() returns PCI_CFG_SPACE_SIZE for the size. VF0
> > has a bypass through pci_cfg_space_size(), so its size is reported
> > correctly, but all subsequent VFs incorrectly report 256 bytes of
> > config space.
> >
> > Resolve by delaying pci_read_vf_config_common() until after
> > pci_setup_device().
> >
> > Fixes: 975bb8b4dc93 ("PCI/IOV: Use VF0 cached config space size for other VFs")
> > Link: https://bugzilla.redhat.com/show_bug.cgi?id=1714978
> > Cc: KarimAllah Ahmed <karahmed@xxxxxxxxx>
> > Signed-off-by: Alex Williamson <alex.williamson@xxxxxxxxxx>
> > ---
> > drivers/pci/iov.c | 6 +++---
> > 1 file changed, 3 insertions(+), 3 deletions(-)
> >
> > diff --git a/drivers/pci/iov.c b/drivers/pci/iov.c
> > index 3aa115ed3a65..34b1f78f4d31 100644
> > --- a/drivers/pci/iov.c
> > +++ b/drivers/pci/iov.c
> > @@ -161,13 +161,13 @@ int pci_iov_add_virtfn(struct pci_dev *dev, int id)
> > virtfn->is_virtfn = 1;
> > virtfn->physfn = pci_dev_get(dev);
> >
> > - if (id == 0)
> > - pci_read_vf_config_common(virtfn);
> > -
> > rc = pci_setup_device(virtfn);
> > if (rc)
> > goto failed1;
> >
> > + if (id == 0)
> > + pci_read_vf_config_common(virtfn);
> > +
> > virtfn->dev.parent = dev->dev.parent;
> > virtfn->multifunction = 0;
>
> Would it actually make more sense to revert 975bb8b4dc93 and just
> assume any is_virtfn device has PCI_CFG_SPACE_EXP_SIZE for cfg_size?
> Per the SR-IOV spec, VFs are required to implement a PCIe capability,
> which should imply 4K of config space. The reachability of that
> extended config space seems unnecessary to test if we assume that it
> has the same characteristics as the PF, which must be reachable if
> we're able to enable SR-IOV. Thoughts? Thanks,
I like this idea.
I first thought maybe we'd still be susceptible to the gotchas
described in the pci_cfg_space_size_ext() comment, i.e., we might not
have a way to generate extended config space accesses, or the device
might be behind a reverse Express bridge.
But as you say, SR-IOV is an extended capability that must be located
at config offset 0x100 or greater, so the fact that we have a VF at
all means we must be able to reach it.
Bjorn
