Error messages are exempt from the 80 columns rule. Move them
onto one line.
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
Reviewed-by: Pavel Hrdina <phrdina@redhat.com>
Before a PCI device can be assigned to a guest with VFIO, that device
must be bound to the vfio-pci driver rather than to the device's
normal host driver. The vfio-pci driver provides APIs that permit QEMU
to perform all the necessary operations to make the device accessible
to the guest.
In the past vfio-pci was the only driver that supplied these APIs, but
there are now vendor/device-specific "VFIO variant" drivers that
provide the basic vfio-pci driver functionality/API while adding
support for device-specific operations (for example these
device-specific drivers may support live migration of certain
devices). All that is needed to make this functionality available is
to bind the vendor-specific "VFIO variant" driver to the device
(rather than the generic vfio-pci driver, which will continue to work,
just without the extra functionality).
But until now libvirt has required that all PCI devices being assigned
to a guest with VFIO specifically have the "vfio-pci" driver bound to
the device. So even if the user manually binds a shiny new
vendor-specific VFIO variant driver to the device (and puts
"managed='no'" in the config to prevent libvirt from changing the
binding), libvirt will just fail during startup of the guest (or
during hotplug) because the driver bound to the device isn't exactly
"vfio-pci".
Beginning with kernel 6.1, it's possible to determine from the sysfs
directory for a device whether the currently-bound driver is the
vfio-pci driver or a VFIO variant - the device directory will have a
subdirectory called "vfio-dev". We can use that to appropriately widen
the list of drivers that libvirt will allow for VFIO device
assignment.
This patch doesn't remove the explicit check for the exact "vfio-pci"
driver (since that would cause systems with pre-6.1 kernels to behave
incorrectly), but adds an additional check for the vfio-dev directory,
so that any VFIO variant driver is acceptable for libvirt to continue
setting up for VFIO device assignment.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
Instead, call it virPCIDeviceGetCurrentDriverPathAndName() to avoid
confusion with the device name that is stored in the virPCIDevice
object - that one is not necessarily the name of the current driver
for the device, but could instead be the driver that we want to be
bound to the device in the future.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
In the past we just kept track of the type of the "stub driver" (the
driver that is bound to a device in order to assign it to a
guest). The next commit will add a stubDriverName to go along with
type, so lets use stubDriverType for the existing enum to make it
easier to keep track of whether we're talking about the name or the
type.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
Refactor the code to use a bitmap with an enum.
Signed-off-by: Peter Krempa <pkrempa@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
Reviewed-by: Martin Kletzander <mkletzan@redhat.com>
There should be a way to show no intent in programming a VLAN at all
(including clearing it). This allows handling error conditions
differently when VLAN clearing is explicit (vlan id == 0) vs implicit
(vlanid == NULL - try to clear it if possible).
Signed-off-by: Dmitrii Shcherbakov <dmitrii.shcherbakov@canonical.com>
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
There are few places where the g_steal_pointer() is open coded.
Switch them to calling the g_steal_pointer() function instead.
Generated by the following spatch:
@ rule1 @
expression a, b;
@@
<...
- b = a;
... when != b
- a = NULL;
+ b = g_steal_pointer(&a);
...>
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
Reviewed-by: Erik Skultety <eskultet@redhat.com>
In some cases we have a label that contains nothing but a return
statement. The amount of such labels rises as we use automagic
cleanup. Anyway, such labels are pointless and can be dropped.
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
Reviewed-by: Daniel Henrique Barboza <danielhb413@gmail.com>
Historically, we declared pointer type to our types:
typedef struct _virXXX virXXX;
typedef virXXX *virXXXPtr;
But usefulness of such declaration is questionable, at best.
Unfortunately, we can't drop every such declaration - we have to
carry some over, because they are part of public API (e.g.
virDomainPtr). But for internal types - we can do drop them and
use what every other C project uses 'virXXX *'.
This change was generated by a very ugly shell script that
generated sed script which was then called over each file in the
repository. For the shell script refer to the cover letter:
https://listman.redhat.com/archives/libvir-list/2021-March/msg00537.html
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
Reviewed-by: Peter Krempa <pkrempa@redhat.com>
In short, virXXXPtr type is going away. With big bang. And to
help us rewrite the code with a sed script, it's better if each
variable is declared on its own line.
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
Generated using the following spatch:
@@
expression path;
@@
- virFileMakePath(path)
+ g_mkdir_with_parents(path, 0777)
However, 14 occurrences were not replaced, e.g. in
virHostdevManagerNew(). I don't really understand why.
Fixed by hand afterwards.
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
virHostdevReAttachPCIDevices() is called when we want to re-attach
a list of hostdevs back to the host, either on the shutdown path or
via a 'virsh detach-device' call. This function always count on the
existence of the device in the host to work, but this can lead to
problems. For example, a SR-IOV device can be removed via an admin
"echo 0 > /sys/bus/pci/devices/<addr>/sriov_numvfs", making the kernel
fire up and eventfd_signal() to the process, asking for the process to
release the device. The result might vary depending on the device driver
and OS/arch, but two possible outcomes are:
1) the hypervisor driver will detach the device from the VM, issuing a
delete event to Libvirt. This can be observed in QEMU;
2) the 'echo 0 > ...' will hang waiting for the device to be unplugged.
This means that the VM process failed/refused to release the hostdev back
to the host, and the hostdev will be detached during VM shutdown.
Today we don't behave well for both cases. We'll fail to remove the PCI device
reference from mgr->activePCIHostdevs and mgr->inactivePCIHostdevs because
we rely on the existence of the PCI device conf file in the sysfs. Attempting
to re-utilize the same device (assuming it is now present back in the host)
can result in an error like this:
$ ./run tools/virsh start vm1-sriov --console
error: Failed to start domain vm1-sriov
error: Requested operation is not valid: PCI device 0000:01:00.2 is in use by driver QEMU, domain vm1-sriov
For (1), a VM destroy/start cycle is needed to re-use the VF in the guest.
For (2), the effect is more nefarious, requiring a Libvirtd daemon restart
to use the VF again in any guest.
We can make it a bit better by checking, during virHostdevReAttachPCIDevices(),
if there is any missing PCI device that will be left behind in activePCIHostdevs
and inactivePCIHostdevs lists. Remove any missing device found from both lists,
unconditionally, matching the current state of the host. This change affects
the code path in (1) (processDeviceDeletedEvent into qemuDomainRemoveDevice, all
the way back to qemuHostdevReAttachPCIDevices) and also in (b) (qemuProcessStop
into qemuHostdevReAttachDomainDevices).
NB: Although this patch enables the possibility of 'outside Libvirt' SR-IOV
hotunplug of PCI devices, if the hypervisor and the PCI driver copes with it,
our goal is to mitigate what it is still considered a user oopsie. For all
supported purposes, the admin must remove the SR-IOV VFs from all running domains
before removing the VFs from the host.
Resolves: https://gitlab.com/libvirt/libvirt/-/issues/72
Reviewed-by: Laine Stump <laine@redhat.com>
Signed-off-by: Daniel Henrique Barboza <danielhb413@gmail.com>
This change will allow us to remove PCI devices from a list
without the need of a PCI Device object, which will be need
in the next patch.
Reviewed-by: Laine Stump <laine@redhat.com>
Signed-off-by: Daniel Henrique Barboza <danielhb413@gmail.com>
Add a helper to quickly determine if a hostdev is a PCI device,
instead of doing a tedious 'if' check with hostdev mode and
subsys type.
Reviewed-by: Laine Stump <laine@redhat.com>
Signed-off-by: Daniel Henrique Barboza <danielhb413@gmail.com>
Gitlab issue #72 [1] reports that removing SR-IOVs VFs before
removing the devices from the running domains can have strange
consequences. QEMU might be able to hotunplug the device inside the
guest, but Libvirt will not be aware of that, and then the guest is
now inconsistent with the domain definition.
There's also the possibility of the VFs removal not succeeding
while the domain is running but then, as soon as the domain
is shutdown, all the VFs are removed. Libvirt can't handle
the removal of the PCI devices while trying to reattach the
hostdevs, and the Libvirt daemon can be left in an inconsistent
state (see [2]).
This patch starts to address the issue related in Gitlab #72, most
notably the issue described in [2]. When shutting down a domain
with SR-IOV hostdevs that got missing, virHostdevReAttachPCIDevices()
is failing the whole process and failing to reattach all the
PCI devices, including the ones that aren't related to the VFs that
went missing. Let's make it more resilient with host changes by
changing virHostdevGetPCIHostDevice() to return an exclusive error
code '-2' for this case. virHostdevGetPCIHostDeviceList() can then
tell when virHostdevGetPCIHostDevice() failed to find the PCI
device of a hostdev and continue to make the list of PCI devices.
virHostdevReAttachPCIDevices() will now be able to proceed reattaching
all other valid PCI devices, at least. The 'ghost hostdevs' will be
handled later on.
[1] https://gitlab.com/libvirt/libvirt/-/issues/72
[2] https://gitlab.com/libvirt/libvirt/-/issues/72#note_459032148
Reviewed-by: Laine Stump <laine@redhat.com>
Signed-off-by: Daniel Henrique Barboza <danielhb413@gmail.com>
Back in commit 2c71d3826, which appeared in libvirt-1.2.3 in April
2014, the location used to store saved MAC addresses and vlan tags of
SRIOV VFs was changed from /var/run/libvirt/qemu to
/var/run/libvirt/hostdevmgr. For backward compatibility the code was
made to continue looking in the old location for the files when it
didn't find them in the new location.
It's now been 6 years, and even if there was somebody still running
libvirt-1.2.3 on their system, that system would now be out of support
for libvirt, so there would be no way for them to upgrade to a new
libvirt that no longer looks in "oldStateDir" for the files. So
let's no longer look in "oldStateDir" for the files!
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
The current virPCIDeviceNew() signature, receiving 4 uints in sequence
(domain, bus, slot, function), is not neat.
We already have a way to represent a PCI address in virPCIDeviceAddress
that is used in the code. Aside from the test files, most of
virPCIDeviceNew() callers have access to a virPCIDeviceAddress reference,
but then we need to retrieve the 4 required uints (addr.domain, addr.bus,
addr.slot, addr.function) to satisfy virPCIDeviceNew(). The result is
that we have extra verbosity/boilerplate to retrieve an information that
is already available in virPCIDeviceAddress.
A better way is presented by virNVMEDeviceNew(), where the caller just
supplies a virPCIDeviceAddress pointer and the function handles the
details internally.
This patch changes virPCIDeviceNew() to receive a virPCIDeviceAddress
pointer instead of 4 uints.
Reviewed-by: Laine Stump <laine@redhat.com>
Signed-off-by: Daniel Henrique Barboza <danielhb413@gmail.com>
The lookup didn't do anything apart from comparing the sysfs paths
anyway since that's what makes each mdev unique.
The most ridiculous usage of the old logic was in
virHostdevReAttachMediatedDevices where in order to drop an mdev
hostdev from the list of active devices we first had to create a new
mdev and use it in the lookup call. Why couldn't we have used the
hostdev directly? Because the hostdev and mdev structures are
incompatible.
The way mdevs are currently removed is via a write to a specific sysfs
attribute. If you do it while the machine which has the mdev assigned
is running, the write call may block (with a new enough kernel, with
older kernels it would return a write error!) until the device
is no longer in use which is when the QEMU process exits.
The interesting part here comes afterwards when we're cleaning up and
call virHostdevReAttachMediatedDevices. The domain doesn't exist
anymore, so the list of active hostdevs needs to be updated and the
respective hostdevs removed from the list, but remember we had to
create an mdev object in the memory in order to find it in the list
first which will fail because the write to sysfs had already removed
the mdev instance from the host system.
And so the next time you try to start the same domain you'll get:
"Requested operation is not valid: mediated device <path> is in use by
driver QEMU, domain <name>"
Fixes: https://gitlab.com/libvirt/libvirt/-/issues/119
Signed-off-by: Erik Skultety <eskultet@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
We set the pointer to some garbage packed structure data without
knowing whether we were actually handling the type of device we
expected to be handling. On its own, this was harmless, because we'd
never use the pointer as we'd skip the device if it were not the
expected type. However, it's better to make the logic even more
explicit - we first check the device and only when we're sure we have
the expected type we then update the pointer shortcut.
Signed-off-by: Erik Skultety <eskultet@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
Apparently at some point in the past, when there were multiple types
to represent PCI addresses, the function
virPCIDeviceAddressGetSysfsFile() used one of those types, while
virDomainHostDevDef used another. It's been quite awhile since we
reduced the number of different representations of PCI address, but
this function was still creating a temporary virPCIDeviceAddress, then
copying the individual elements into this temporary object from the
same type of object in the virDomainHostDevDef.
This patch just eliminates that pointless copy.
Signed-off-by: Laine Stump <laine@redhat.com>
Reviewed-by: Erik Skultety <eskultet@redhat.com>
Reviewed-by: Peter Krempa <pkrempa@redhat.com>
While 'namespace' is not a reserved word in C, it is in C++. Our
compilers are happy with it but syntax-hilighting in some editors
hilights is as a keyword. Rename it to prevent confusion.
Signed-off-by: Peter Krempa <pkrempa@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
This module depends on domain_conf and is used directly by various
hypervisor drivers.
Move it to src/hypervisor.
Signed-off-by: Ján Tomko <jtomko@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
