On Wed, 1 Feb 2012 18:32:06 +0530
Vaidyanathan Srinivasan <svaidy@xxxxxxxxxxxxxxxxxx> wrote:
> * Ram Pai <linuxram@xxxxxxxxxx> [2012-02-01 14:21:45]:
>
> > On Tue, Jan 31, 2012 at 11:14:02PM +0530, Vaidyanathan Srinivasan wrote:
> > > * Ram Pai <linuxram@xxxxxxxxxx> [2012-01-30 11:18:45]:
> > >
> > > > On Sat, Jan 28, 2012 at 12:40:32AM +0530, Vaidyanathan Srinivasan wrote:
> > > > > Hi Ram and Jesse,
> > > > >
> > > > > I found a trivial issue with page size alignment check on IBM POWER
> > > > > box with 64k base page size. In sriov_init(), changing the check from
> > > > > PAGE_SIZE (arch and config dependent) to HW_PAGE_SIZE (always 4k) was
> > > > > required to use one of the sriov adapter as PF since the
> > > > > resource_size() comes up as 0x8000 and PAGE_SIZE would be 0x10000 for
> > > > > pseries boxes.
> > > > >
> > > > > I think resource_size() could be less than SystemPageSize, but I would
> > > > > like your comments/ack/nack on any consequences of checking for only
> > > > > 4k alignment here in a system with larger base page size.
> > > >
> > > > As per the SRIOV specs, the resource has to be System page size aligned.
> > > >
> > > > PFs are required to support 4-KB, 8-KB, 64-KB, 256-KB, 1-MB, and 4-MB
> > > > page sizes. In your case if your adapter's PF is not supporting 64K page size
> > > > then I think it is not conforming to the PCI SRIOV spec.
> > >
> > > Hi Ram,
> > >
> > > Thanks for the pointer. I did some more experiments and found that
> > > the card does support 64k page size, but the PCI_SRIOV_SYS_PGSIZE was
> > > set to default 4k when we do the query and check resource_size().
> > >
> > > You were correct, the resource_size() has to come up with 64k on 64k
> > > PAGE_SIZE system. We should not change that check. I was able to
> > > get a working solution by setting PCI_SRIOV_SYS_PGSIZE to 64k before
> > > we do the query.
> > >
> > > This was the case in the original code before you moved these to
> > > sriov_enable(). If it is ok to leave the SYS_PGSIZE setting in
> > > sriov_init(), then I have the following fix that works for me.
> > >
> > > Please review and let me know your comments.
> > >
> > > Thanks,
> > > Vaidy
> > > ---
> > >
> > > pci: set pci sriov page size before reading sriov bar
> > >
> > > For an SRIOV device, PCI_SRIOV_SYS_PGSIZE should be set before
> > > the PCI_SRIOV_BAR is queried. The sys pagesize defaults to 4k,
> > > so this change is required on powerpc box with 64k base page size.
> > >
> > > This is a regression caused due to moving SRIOV init to sriov_enable().
> > >
> > > | commit afd24ece5c76af87f6fc477f2747b83a764f161c
> > > | Author: Ram Pai <linuxram@xxxxxxxxxx>
> > >
> > > | PCI: delay configuration of SRIOV capability
> > > | The SRIOV capability, namely page size and total_vfs of a device are
> > > | configured during enumeration phase of the device. This can potentially
> > > | interfere with the PCI operations of the platform, if the IOV capability
> > > | of the device is not enabled.
> > >
> > > Signed-off-by: Vaidyanathan Srinivasan <svaidy@xxxxxxxxxxxxxxxxxx>
> > >
> > > diff --git a/drivers/pci/iov.c b/drivers/pci/iov.c
> > > index 0321fa3..0dab5ec 100644
> > > --- a/drivers/pci/iov.c
> > > +++ b/drivers/pci/iov.c
> > > @@ -347,8 +347,6 @@ static int sriov_enable(struct pci_dev *dev, int nr_virtfn)
> > > return rc;
> > > }
> > >
> > > - pci_write_config_dword(dev, iov->pos + PCI_SRIOV_SYS_PGSIZE, iov->pgsz);
> > > -
> > > iov->ctrl |= PCI_SRIOV_CTRL_VFE | PCI_SRIOV_CTRL_MSE;
> > > pci_cfg_access_lock(dev);
> > > pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl);
> > > @@ -466,6 +464,7 @@ found:
> > > return -EIO;
> > >
> > > pgsz &= ~(pgsz - 1);
> > > + pci_write_config_dword(dev, pos + PCI_SRIOV_SYS_PGSIZE, pgsz);
> > >
> > > nres = 0;
> > > for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
> >
> >
> > ACK. I think it is better to revert afd24ece5c76af87f6fc477f2747b83a764f161c.
>
> Hi Ram,
>
> Thanks for the ack. But afd24ece5c76af87f6fc477f2747b83a764f161c has
> one more change of moving
> pci_write_config_word(dev, pos + PCI_SRIOV_NUM_VF, total) to sriov_enable().
>
> This change is required so that we set the PCI_SRIOV_NUM_VF only
> during sriov_enable.
>
> So we should not revert the entire commit, we can just add this change.
So which is it Ram, the ack or the revert? :)
Having the right page size early seems like the right solution...
--
Jesse Barnes, Intel Open Source Technology Center
Attachment:
signature.asc
Description: PGP signature
