The problem resides in virHostdevUpdateActiveMediatedDevices which gets
called during qemuProcessReconnect. The issue here is that
virMediatedDeviceListAdd takes a pointer to the item to be added to the
list to which VIR_APPEND_ELEMENT is used, which also clears the pointer.
However, in this case only the local copy of the pointer got cleared,
leaving the original pointing to valid memory. To sum it up, during
cleanup phase, the original pointer is freed and the daemon crashes
basically any time it would access it.
Backtrace:
0x00007ffff3ccdeba in __strcmp_sse2_unaligned
0x00007ffff72a444a in virMediatedDeviceListFindIndex
0x00007ffff7241446 in virHostdevReAttachMediatedDevices
0x00007fffc60215d9 in qemuHostdevReAttachMediatedDevices
0x00007fffc60216dc in qemuHostdevReAttachDomainDevices
0x00007fffc6046e6f in qemuProcessStop
0x00007fffc6091596 in processMonitorEOFEvent
0x00007fffc6091793 in qemuProcessEventHandler
0x00007ffff7294bf5 in virThreadPoolWorker
0x00007ffff7294184 in virThreadHelper
0x00007ffff3fdc3c4 in start_thread () from /lib64/libpthread.so.0
0x00007ffff3d269cf in clone () from /lib64/libc.so.6
Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1446455
Signed-off-by: Erik Skultety <eskultet@redhat.com>
Reviewed-by: Laine Stump <laine@laine.org>
It takes longer to explain this than to fix it...
In the past we weren't able to save the VF's own MAC address *at all*
when using it for hostdev assignment, because we had already unbound
the VF from the host net driver prior to saving its config. With the
previous patch, that problem has been solved, so we now have the VF's
MAC address saved and can move on to the *next* problem, which is twofold:
1) during teardown we restore the config before we've re-bound, so the
VF doesn't have a net driver, and thus we can't set its MAC address
directly.
2) even if we delay restoring the config until the VF is bound to a
net driver, the request to set its MAC address would fail, since
(during device setup) we had set the "admin MAC" for the VF via an
RTM_SETLINK to the PF - once you've set the admin MAC for a VF, the
VF driver (either on host or on guest) is not allowed to change the
VF's MAC address "forever" (well, until you reload the PF driver,
but that requires destroying and recreating every single VF, which
isn't something you can require).
The solution is to keep the restoration of config at the same place,
but to set the *admin MAC* to the address you want the VF to have -
when the VF net driver is later initialized (as a part of re-binding
to the VF net driver) its MAC will be initialized to the current value
of the admin MAC.
In order to properly restore the original state of an SRIOV VF when
we're finished with it, we need to save the MAC address of the VF
itself (not just the admin MAC address for the VF that is stored in
the PF). But that can only be done when the VF is still bound to the
host's netdev driver, and we have always done the saving of device
config after the VF is already bound to vfio-pci. This patch prepares
us for adding a save of the VF's MAC by calling the function that
saves netconfig earlier in the device preparation, before we've
unbound it from the host netdev driver.
These two operations will need to be separated so that saving of the
original config is done before detaching the host net driver, and
setting the new config is done after attaching vfio-pci. This patch
splits the single function into two, but for now calls them together
(to make bisecting easier if there is a regression).
virHostdevNetConfigReplace() and virHostdevNetConfigRestore() are
modified to use the new virNetDev*NetConfig() functions.
Note that due to the VF's original MAC addresses being saved after it
has already been un-bound from the host net driver, the actual current
VF MAC address won't be saved (because it no longer exists) - only the
"admin MAC" will be saved. This reflects existing behavior that will
be fixed in an upcoming patch.
Keep track of the assigned mediated devices the same way we do it for
the rest of hostdevs. Methods like 'Prepare', 'Update', and 'ReAttach'
are introduced by this patch.
Signed-off-by: Erik Skultety <eskultet@redhat.com>
...and cleanup the callers to report it when it *is* an error.
In many cases It's useful for virPCIGetNetName() to not log an error
and simply return a NULL pointer when the given device isn't bound to
a net driver (e.g. we're looking at a VF that is permanently bound to
vfio-pci). The existing code would silently return an error in this
case, which could eventually lead to the dreaded "An error occurred
but the cause is unknown" log message.
This patch changes virPCIGetNetName() to still return success if the
device simply isn't bound to a net driver, and adjusts all the callers
that require a non-null netname to check for that condition and log an
error when it happens.
Since the code checks 'mgr == NULL' anyway, no need for the prototype
to have the NONNULL arg check. Also add an error message to indicate what
the failure is so that there isn't a failed for some reason error.
Signed-off-by: John Ferlan <jferlan@redhat.com>
The comparison code used STREQ_NULLABLE anyway for both 'drv_name' and
'dom_name', so no need. Add a NULLSTR on the 'dom_name' too.
Signed-off-by: John Ferlan <jferlan@redhat.com>
The comparison code used STREQ_NULLABLE anyway for both 'drv_name' and
'dom_name', so no need. Add a NULLSTR on the 'dom_name' too.
Signed-off-by: John Ferlan <jferlan@redhat.com>
Non-endpoint devices like pci-bridges cannot be assigned to guests.
Prevent such attempts.
Signed-off-by: Shivaprasad G Bhat <sbhat@linux.vnet.ibm.com>
For a new hostdev type='scsi_host' we have a number of
required functions for managing, adding, and removing the
host device to/from guests. Provide the basic infrastructure
for these tasks.
The name "SCSIVHost" (and its variants) is chosen to avoid
conflicts with existing code named "SCSIHost" to refer to
a hostdev type='scsi' protcol='none'.
Signed-off-by: Eric Farman <farman@linux.vnet.ibm.com>
Use the util function virHostdevIsSCSIDevice() to simplify if
statements.
Signed-off-by: Marc Hartmayer <mhartmay@linux.vnet.ibm.com>
Reviewed-by: Bjoern Walk <bwalk@linux.vnet.ibm.com>
Reviewed-by: Boris Fiuczynski <fiuczy@linux.vnet.ibm.com>
Add the function virHostdevIsSCSIDevice() which detects whether a
hostdev is a SCSI device or not.
Signed-off-by: Marc Hartmayer <mhartmay@linux.vnet.ibm.com>
Reviewed-by: Bjoern Walk <bwalk@linux.vnet.ibm.com>
Reviewed-by: Boris Fiuczynski <fiuczy@linux.vnet.ibm.com>
Finding an USB device from the vendor/device values will be needed
by libxl driver to convert from vendor/device to bus/dev addresses.
Signed-off-by: Jim Fehlig <jfehlig@suse.com>
SRIOV VFs used in macvtap passthrough mode can take advantage of the
SRIOV card's transparent vlan tagging. All the code was there to set
the vlan tag, and it has been used for SRIOV VFs used for hostdev
interfaces for several years, but for some reason, the vlan tag for
macvtap passthrough devices was stubbed out with a -1.
This patch moves a bit of common validation down to a lower level
(virNetDevReplaceNetConfig()) so it is shared by hostdev and macvtap
modes, and updates the macvtap caller to actually send the vlan config
instead of -1.
This reverts commit ee4cfb5643.
Since we're still not persisting our bookkeeping lists across
daemon restarts, we might have lost some information
virPCIDeviceReattach() relies on, for example whether the
device needs to be unbound from the stub driver.
As a result, if the daemon has been restarted in the meantime,
the device might end up remaining bound to the stub driver even
after 'virsh nodedev-reattach' or similar has been called, with
no way of giving it back to the host short of messing with
sysfs behind libvirt's back.
Revert back to the previous behavior of always trying to bind
the device to the host driver, regardless of its status when it
was detached, until persistent bookkeeping lists have been
implemented.
Instead of forcing the values for the unbind_from_stub, remove_slot
and reprobe properties, look up the actual device and use that when
calling virPCIDeviceReattach().
This ensures the device is restored to its original state after
reattach: for example, if it was not bound to any driver before
detach, it will not be bound forcefully during reattach.
We would be just fine looking up the information in pcidevs most
of the time; however, some corner cases would not be handled
properly, so look up the actual device instead.
After this patch, ownership of virPCIDevice instances is very easy
to keep track of: for each host PCI device, the only instance that
actually matters is the one inside one of the bookkeeping list.
Whenever some operation needs to be performed on a PCI device, the
actual device is looked up first; when this is not the case, a
comment explains the reason.
Unmanaged devices, as the name suggests, are not detached
automatically from the host by libvirt before being attached to a
guest: it's the user's responsability to detach them manually
beforehand. If that preliminary step has not been performed, the
attach operation can't complete successfully.
Instead of relying on the lower layers to error out with cryptic
messages such as
error: Failed to attach device from /tmp/hostdev.xml
error: Path '/dev/vfio/12' is not accessible: No such file or directory
prevent the situation altogether and provide the user with a more
useful error message.
Unmanaged devices are attached to guests in two steps: first,
the device is detached from the host and marked as inactive;
subsequently, it is marked as active and attached to the guest.
If the daemon is restarted between these two operations, we lose
track of the inactive device.
Steps 5 and 6 of virHostdevPreparePCIDevices() already subtly
take care of this situation, but some planned changes will make
it so that's no longer the case. Plus, explicit is always better
than implicit.
These comments explain the difference between a virPCIDevice
instance used for lookups and an actual device instance; some
information is also provided for specific uses.
virHostdevGetPCIHostDeviceList() is similar but does not filter out
devices that are not in the active list; that said, we are looking
up the device in the active list just a few lines after anyway, so
we might as well just keep a single function around.
This also helps stress the fact the objects contained in pcidevs are
only for looking up the actual devices, which is something later
commits will make even more explicit.
We're in the hostdev module, so mgr is not an ambiguous name, and
in fact it's already used in some cases. Switch all the code over.
Take the chance to shorten declaration of
virHostdevIsPCINodeDeviceUsedData structures.
When we want to look up a device in a device list and we already
have the IDs from another source, we can simply use
virPCIDeviceListFindByIDs() instead of creating a temporary device
object.
If 'last_processed_hostdev_vf != -1' is false then, since the
loop counter 'i' starts at 0, 'i <= last_processed_hostdev_vf'
can't possibly be true and the loop body will never be executed.
However, since 'i' is unsigned and 'last_processed_hostdev_vf'
is signed, we can't just get rid of the check completely; what
we can do is move it outside of the loop to avoid checking its
value on every iteration and cluttering the actual loop
condition.
The comment claimed that virPCIDeviceReattach() does not reattach
a device to the host driver; except it actually does, so the
comment is just confusing and we're better off removing it.
Replace the term "loop" with the more generic "step". This allows us
to be more flexible and eg. have a step that consists in a single
function call.
Don't include the number of steps in the first comment of the
function, so that we can add or remove steps without having to worry
about keeping that comment in sync.
For the same reason, remove the summary contained in that comment.
Clean up some weird vertical spacing while we're at it.
The virHostdevIsVirtualFunction() was called exactly twice, and in
both cases the return value was saved to a temporary variable before
being checked. This would be okay if it improved readability, but in
this case is pretty pointless.
Get rid of the temporary variable and check the return value
directly; while at it, change the check from '<= 0' to '!= 1' to
align it with the way other similar *IsVirtualFunction() functions
are used thorough the code.
The name is confusing, and there are just two uses: one is a test case,
and the other will be removed as part of an upcoming refactoring of
the hostdev code.
We want to eventually factor out the code dealing with device detaching
and reattaching, so that we can share it and make sure it's called eg.
when 'virsh nodedev-detach' is used.
For that to happen, it's important that the lists of active and inactive
PCI devices are updated every time a device changes its state.
Instead of passing NULL as the last argument of virPCIDeviceDetach() and
virPCIDeviceReattach(), pass the proper list so that it can be updated.
This replaces the virPCIKnownStubs string array that was used
internally for stub driver validation.
Advantages:
* possible values are well-defined
* typos in driver names will be detected at compile time
* avoids having several copies of the same string around
* no error checking required when setting / getting value
The names used mirror those in the
virDomainHostdevSubsysPCIBackendType enumeration.
Instead of replicating the information (domain, bus, slot, function)
inside the virPCIDevice structure, use the already-existing
virPCIDeviceAddress structure.
For users of the module, this means that the object returned by
virPCIDeviceGetAddress() can no longer be NULL and must no longer
be freed by the caller.
The ACS checks are meaningless when using the more modern VFIO driver
for device assignment since VFIO has its own more complete and exact
checks, but I didn't realize that when I added support for VFIO. This
patch eliminates the ACS check when preparing PCI devices for
assignment if VFIO is being used.
This resolves:
https://bugzilla.redhat.com/show_bug.cgi?id=1256486
The commit 7e72de4 didn't consider the hotplug scenarios. The patch addresses
the hotplug case whereby if atleast one of the pci function is owned by a
guest, the hotplug of other functions/devices in the same iommu group to the
same guest goes through successfully.
Signed-off-by: Shivaprasad G Bhat <sbhat@linux.vnet.ibm.com>