On 10/28/2015 11:43 AM, Bjorn Helgaas wrote:
On Wed, Oct 28, 2015 at 11:32:16AM -0500, Bjorn Helgaas wrote:
Hi Alex,
Thanks a lot for cleaning this up. I think this is a great
improvement over what I did.
On Tue, Oct 27, 2015 at 01:52:15PM -0700, Alexander Duyck wrote:
This patch pulls the validation of offset and stride into virtfn_max_buses.
The general idea is to validate offset and stride for each possible value
of numvfs in addition to still determining the maximum bus value for the
VFs.
I also reversed the loop as the most likely maximum will be when numvfs is
set to total_VFs. In addition this makes it so that we loop down to a
value of 0 for numvfs which should be the resting state for the register.
Fixes: 8e20e89658f2 ("PCI: Set SR-IOV NumVFs to zero after enumeration")
Signed-off-by: Alexander Duyck <aduyck@xxxxxxxxxxxx>
I'd like to squash this together with my patch instead of having fixes
on top of fixes. What do you think of the following? (This applies
on top of 70675e0b6a1a ("PCI: Don't try to restore VF BARs")).
commit c20e11b572c5d4e4f01c86580a133122fbd13cfa
Author: Alexander Duyck <aduyck@xxxxxxxxxxxx>
Date: Wed Oct 28 10:54:32 2015 -0500
PCI: Set SR-IOV NumVFs to zero after enumeration
The enumeration path should leave NumVFs set to zero. But after
4449f079722c ("PCI: Calculate maximum number of buses required for VFs"),
we call virtfn_max_buses() in the enumeration path, which changes NumVFs.
This NumVFs change is visible via lspci and sysfs until a driver enables
SR-IOV.
Iterate from TotalVFs down to zero so NumVFs is zero when we're finished
computing the maximum number of buses. Validate offset and stride in
the loop, so we can test it at every possible NumVFs setting. Rename
virtfn_max_buses() to compute_max_vf_buses() to hint that it does have a
side effect of updating iov->max_VF_buses.
[bhelgaas: changelog, rename, reverse sense of error path]
Fixes: 4449f079722c ("PCI: Calculate maximum number of buses required for VFs")
Based-on-patch-by: Ethan Zhao <ethan.zhao@xxxxxxxxxx>
Signed-off-by: Alexander Duyck <aduyck@xxxxxxxxxxxx>
Signed-off-by: Bjorn Helgaas <bhelgaas@xxxxxxxxxx>
diff --git a/drivers/pci/iov.c b/drivers/pci/iov.c
index ee0ebff..120cfb3 100644
--- a/drivers/pci/iov.c
+++ b/drivers/pci/iov.c
@@ -54,24 +54,33 @@ static inline void pci_iov_set_numvfs(struct pci_dev *dev, int nr_virtfn)
* The PF consumes one bus number. NumVFs, First VF Offset, and VF Stride
* determine how many additional bus numbers will be consumed by VFs.
*
- * Iterate over all valid NumVFs and calculate the maximum number of bus
- * numbers that could ever be required.
+ * Iterate over all valid NumVFs, validate offset and stride, and calculate
+ * the maximum number of bus numbers that could ever be required.
*/
-static inline u8 virtfn_max_buses(struct pci_dev *dev)
+static int compute_max_vf_buses(struct pci_dev *dev)
{
struct pci_sriov *iov = dev->sriov;
- int nr_virtfn;
- u8 max = 0;
+ int nr_virtfn = iov->total_VFs;
int busnr;
- for (nr_virtfn = 1; nr_virtfn <= iov->total_VFs; nr_virtfn++) {
- pci_iov_set_numvfs(dev, nr_virtfn);
+ pci_iov_set_numvfs(dev, nr_virtfn);
+
+ while (nr_virtfn--) {
+ if (!iov->offset || !iov->stride)
+ goto err;
I think we have a minor problem here. In sriov_enable(), we return an
error if "nr_virtfn > 1 && !iov->stride", so it's legal for stride to
be zero if NumVF is 1. Here we don't allow that. Sec 3.3.10 says:
Note: VF Stride is unused if NumVFs is 0 or 1. If NumVFs is greater
than 1, VF Stride must not be zero."
So I think we should allow "stride == 0" here when NumVFs is 1.
Right, we shouldn't be testing it if NumVFs is 1 or less.
+
busnr = pci_iov_virtfn_bus(dev, nr_virtfn - 1);
I think this loop management is slightly wrong: I don't think we ever
compute busnr for the highest VF because we always decrement nr_virtfn
after calling pci_iov_set_numvfs(), and then we subtract one again.
E.g., if Total VFs is 8, the VFs are numbered VF0..VF7, and we have
this, which doesn't check VF7:
nr_virtfn = iov->total_VFs # nr_virtfn == 8
pci_iov_set_numvfs(..., nr_virtfn) # passes 8 (correct)
while (nr_virtfn--) {
# nr_virtfn == 7 in loop body
pci_iov_virtfn_bus(..., nr_virtfn - 1) # passes 6 (wrong)
Yeah, that was supposed to just be nr_virtfn.
- if (busnr > max)
- max = busnr;
+ if (busnr > iov->max_VF_buses)
+ iov->max_VF_buses = busnr;
+
+ pci_iov_set_numvfs(dev, nr_virtfn);
}
- return max;
+ return 0;
+
+err:
+ pci_iov_set_numvfs(dev, 0);
+ return -EIO;
}
Here's my new proposal:
static int compute_max_vf_buses(struct pci_dev *dev)
{
struct pci_sriov *iov = dev->sriov;
int nr_virtfn, busnr, rc = 0;
for (nr_virtfn = iov->total_VFs; nr_virtfn; nr_virtfn--) {
pci_iov_set_numvfs(dev, nr_virtfn);
if (!iov->offset || (nr_virtfn > 1 && !iov->stride)) {
rc = -EIO;
goto out;
}
busnr = pci_iov_virtfn_bus(dev, nr_virtfn - 1);
if (busnr > iov->max_VF_buses)
iov->max_VF_buses = busnr;
}
out:
pci_iov_set_numvfs(dev, 0);
return rc;
}
This looks good to me. In theory you could save yourself a pair of MMIO
reads at the end of the loop by just writing numvfs without the offset
and stride read, but this should work.
- Alex
--
To unsubscribe from this list: send the line "unsubscribe linux-pci" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html