* Chris Lalancette (clalance@xxxxxxxxxx) wrote:
> When trying to assign a PCI device to a guest, we have
> to check that all bridges upstream of that device support
> ACS. That means that we have to find the parent bridge of
> the current device, check for ACS, then find the parent bridge
> of that device, check for ACS, etc. As it currently stands,
> the code to do this iterates through all PCI devices on the
> system, looking for a device that has a range of busses that
> included the current device's bus.
>
> That check is not restrictive enough, though. Depending on
> how we iterated through the list of PCI devices, we could first
> find the *topmost* bridge in the system; since it necessarily had
> a range of busses including the current device's bus, we
> would only ever check the topmost bridge, and not check
> any of the intermediate bridges.
>
> Note that this also caused a fairly serious bug in the
> secondary bus reset code, where we could erroneously
> find and reset the topmost bus instead of the inner bus.
>
> This patch changes pciGetParentDevice() so that it first
> checks if a bridge device's secondary bus exactly matches
> the bus of the device we are looking for. If it does, we've
> found the correct parent bridge and we are done. If it does not,
> then we check to see if this bridge device's busses *include* the
> bus of the device we care about. If so, we mark this bridge device
> as best, and go on. If we later find another bridge device whose
> busses include this device, but is more restrictive, then we
> free up the previous best and mark the new one as best. This
> algorithm ensures that in the normal case we find the direct
> parent, but in the case that the parent bridge secondary bus
> is not exactly the same as the device, we still find the
> correct bridge.
>
> This patch was tested by me on a 4-port NIC with a
> bridge without ACS (where assignment failed), a 4-port
> NIC with a bridge with ACS (where assignment succeeded),
> and a 2-port NIC with no bridges (where assignment
> succeeded).
>
> Signed-off-by: Chris Lalancette <clalance@xxxxxxxxxx>
> ---
> src/util/pci.c | 78 +++++++++++++++++++++++++++++++++++++++++++++++--------
> 1 files changed, 66 insertions(+), 12 deletions(-)
>
> diff --git a/src/util/pci.c b/src/util/pci.c
> index 26d55b8..f2890bd 100644
> --- a/src/util/pci.c
> +++ b/src/util/pci.c
> @@ -283,6 +283,7 @@ pciIterDevices(pciIterPredicate predicate,
> DIR *dir;
> struct dirent *entry;
> int ret = 0;
> + int rc;
>
> *matched = NULL;
>
> @@ -322,11 +323,20 @@ pciIterDevices(pciIterPredicate predicate,
> break;
> }
>
> - if (predicate(dev, check, data)) {
> + rc = predicate(dev, check, data);
> + if (rc < 0) {
> + /* the predicate returned an error, bail */
> + pciFreeDevice(check);
> + ret = -1;
> + break;
> + }
> + else if (rc == 1) {
CodingStyle?
$ git grep "else {" | wc -l
1150
$ git grep "else {" | grep -v } | wc -l
67
Guess it's not unprecedented, just rare ;)
> VIR_DEBUG("%s %s: iter matched on %s", dev->id, dev->name, check->name);
> *matched = check;
> + ret = 1;
> break;
> }
> +
> pciFreeDevice(check);
> }
> closedir(dir);
> @@ -510,10 +520,11 @@ pciBusContainsActiveDevices(pciDevice *dev,
>
> /* Is @check the parent of @dev ? */
> static int
> -pciIsParent(pciDevice *dev, pciDevice *check, void *data ATTRIBUTE_UNUSED)
> +pciIsParent(pciDevice *dev, pciDevice *check, void *data)
> {
> uint16_t device_class;
> uint8_t header_type, secondary, subordinate;
> + pciDevice **best = data;
>
> if (dev->domain != check->domain)
> return 0;
> @@ -533,16 +544,54 @@ pciIsParent(pciDevice *dev, pciDevice *check, void *data ATTRIBUTE_UNUSED)
>
> VIR_DEBUG("%s %s: found parent device %s", dev->id, dev->name, check->name);
>
> - /* No, it's superman! */
> - return (dev->bus >= secondary && dev->bus <= subordinate);
> + /* if the secondary bus exactly equals the device's bus, then we found
> + * the direct parent. No further work is necessary
> + */
> + if (dev->bus == secondary)
> + return 1;
> +
> + /* otherwise, SRIOV allows VFs to be on different busses then their PFs.
> + * In this case, what we need to do is look for the "best" match; i.e.
> + * the most restrictive match that still satisfies all of the conditions.
> + */
> + if (dev->bus > secondary && dev->bus <= subordinate) {
> + if (*best == NULL) {
> + *best = pciGetDevice(check->domain, check->bus, check->slot,
> + check->function);
> + if (*best == NULL)
> + return -1;
too bad it's not refcounted, then we'd never hit a failure case here
> + }
> + else {
CodingStyle?
> + /* OK, we had already recorded a previous "best" match for the
> + * parent. See if the current device is more restrictive than the
> + * best, and if so, make it the new best
> + */
> + if (secondary > pciRead8(*best, PCI_SECONDARY_BUS)) {
forgot it's not cached, at least this isn't a performance senstitive path
> + pciFreeDevice(*best);
> + *best = pciGetDevice(check->domain, check->bus, check->slot,
> + check->function);
> + if (*best == NULL)
> + return -1;
> + }
> + }
> + }
Yes, should get the right match.
> +
> + return 0;
> }
>
> -static pciDevice *
> -pciGetParentDevice(pciDevice *dev)
> +static int
> +pciGetParentDevice(pciDevice *dev, pciDevice **parent)
> {
> - pciDevice *parent = NULL;
> - pciIterDevices(pciIsParent, dev, &parent, NULL);
> - return parent;
> + pciDevice *best = NULL;
> + int ret;
> +
> + *parent = NULL;
> + ret = pciIterDevices(pciIsParent, dev, parent, &best);
> + if (ret == 1)
> + pciFreeDevice(best);
> + else if (ret == 0)
> + *parent = best;
> + return ret;
I think we could keep the old interface, but that's really just
splitting hairs, and hey...you did the work already ;)
> }
>
> /* Secondary Bus Reset is our sledgehammer - it resets all
> @@ -570,7 +619,8 @@ pciTrySecondaryBusReset(pciDevice *dev,
> }
>
> /* Find the parent bus */
> - parent = pciGetParentDevice(dev);
> + if (pciGetParentDevice(dev, &parent) < 0)
> + return -1;
> if (!parent) {
> pciReportError(VIR_ERR_NO_SUPPORT,
> _("Failed to find parent device for %s"),
> @@ -1377,7 +1427,8 @@ pciDeviceIsBehindSwitchLackingACS(pciDevice *dev)
> {
> pciDevice *parent;
>
> - parent = pciGetParentDevice(dev);
> + if (pciGetParentDevice(dev, &parent) < 0)
> + return -1;
> if (!parent) {
> /* if we have no parent, and this is the root bus, ACS doesn't come
> * into play since devices on the root bus can't P2P without going
> @@ -1400,6 +1451,7 @@ pciDeviceIsBehindSwitchLackingACS(pciDevice *dev)
> do {
> pciDevice *tmp;
> int acs;
> + int ret;
>
> acs = pciDeviceDownstreamLacksACS(parent);
>
> @@ -1412,8 +1464,10 @@ pciDeviceIsBehindSwitchLackingACS(pciDevice *dev)
> }
>
> tmp = parent;
> - parent = pciGetParentDevice(parent);
> + ret = pciGetParentDevice(parent, &parent);
> pciFreeDevice(tmp);
> + if (ret < 0)
> + return -1;
> } while (parent);
>
> return 0;
Acked-by: Chris Wright <chrisw@xxxxxxxxxx>
--
libvir-list mailing list
libvir-list@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/libvir-list