VFIO_PCI_NUM_REGIONS and VFIO_PCI_NUM_IRQS should never have been
used in this manner as it locks a specific kernel implementation.
Future features may introduce new regions or interrupt entries
(VGA may add legacy ranges, AER might add an IRQ for error
signalling). Fix this before it gets us into trouble.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Cc: qemu-stable@nongnu.org
Guests typically enable MSI-X with all of the vectors in the MSI-X
vector table masked. Only when the vector is enabled does the vector
get unmasked, resulting in a vector_use callback. These two points,
enable and unmask, correspond to pci_enable_msix() and request_irq()
for Linux guests. Some drivers rely on VF/PF or PF/fw communication
channels that expect the physical state of the device to match the
guest visible state of the device. They don't appreciate lazily
enabling MSI-X on the physical device.
To solve this, enable MSI-X with a single vector when the MSI-X
capability is enabled and immediate disable the vector. This leaves
the physical device in exactly the same state between host and guest.
Furthermore, the brief gap where we enable vector 0, it fires into
userspace, not KVM, so the guest doesn't get spurious interrupts.
Ideally we could call VFIO_DEVICE_SET_IRQS with the right parameters
to enable MSI-X with zero vectors, but this will currently return an
error as the Linux MSI-X interfaces do not allow it.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Cc: qemu-stable@nongnu.org
kvm_irqchip_in_kernel() has an architecture specific meaning, so
we shouldn't be using it to determine whether to enabled KVM INTx
bypass. kvm_irqfds_enabled() seems most appropriate. Also use this
to protect our other call to kvm_check_extension() as that explodes
when KVM isn't enabled.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Cc: qemu-stable@nongnu.org
This makes use of the new level irqfd support enabling bypass of qemu
userspace both on INTx injection and unmask. This significantly
boosts the performance of devices making use of legacy interrupts (ex.
~60% better netperf TCP_RR scores for an e1000e assigned to a Linux
guest and booted with pci=nomsi). This also avoids flipping mmaps on
and off to simulate EOIs, so greatly improves performance of device
access in addition to interrupt latency.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
target_phys_addr_t is unwieldly, violates the C standard (_t suffixes are
reserved) and its purpose doesn't match the name (most target_phys_addr_t
addresses are not target specific). Replace it with a finger-friendly,
standards conformant hwaddr.
Outstanding patchsets can be fixed up with the command
git rebase -i --exec 'find -name "*.[ch]"
| xargs s/target_phys_addr_t/hwaddr/g' origin
Signed-off-by: Avi Kivity <avi@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Using the AddressSpace type reduces confusion, as you can't accidentally
supply the MemoryRegion you're interested in.
Reviewed-by: Anthony Liguori <aliguori@us.ibm.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
DO_UPCAST is supposed to translate from the first member of a struct to
that struct, not from arbitrary ones. And it (usually) breaks the build
when neglecting this rule. Use container_of to fix the build breakage
and likely also the runtime behavior.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
aw: runtime behavior is actually the same, but clearly misuse of DO_UPCAST
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Blue Swirl reports that Clang doesn't like the structure we define to
avoid dynamic allocation for a number of calls to VFIO_DEVICE_SET_IRQS.
Adding an element after a variable sized type is a GNU extension.
Switch back to dynamic allocation, which really isn't a problem since
this is only done on interrupt setup changes.
Cc: Blue Swirl <blauwirbel@gmail.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Take what we've learned from pci-assign and apply it to vfio-pci.
On reset, disable previous interrupt config, perform a device
reset if available, re-enable INTx, and disable memory regions on
the device to prevent continuing DMA.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
This was a misinterpretation of the spec, hardware doesn't get to
specify how many were actually enabled through this field.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
We try to do lazy initialization of MSIX since we don't actually need
to setup anything until MSIX vectors start getting used. This leads
to problems if MSIX is enabled, but never used (we can end up trying
to re-enable INTx while it's still enabled). We also run into
problems trying to expand our reset function to tear down interrupts
as we can then get vector release notifications after we've released
data structures. By making explicit initialization and teardown we
can avoid both of these problems and behave more similar to bare
metal.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Occasionally we get regions added that overlap with existing mappings.
These always seems to be in the VGA ROM range. VFIO returns EBUSY
for these mapping attempts. We can try a little harder and assume
that the latest mapping is correct by removing any overlapping ranges
and retrying the original request.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
We can't afford the overhead of switching out and back into mmap mode
around each interrupt, but we can do it lazily via a timer. On INTx
interrupt, disable the mmap'd memory regions and set a timer. On
every interrupt, push the timer out. If the timer expires and the
interrupt is no longer pending, switch back to mmap mode.
This has the benefit that things like graphics cards, which rarely or
never, fire an interrupt don't need manual user intervention to add
the x-intx=off parameter. They'll just remain in mmap mode until they
trigger an interrupt, and if they don't continue to regularly fire
interrupts, they'll switch back.
The default timeout is tuned for network cards so that a ping is just
enough to keep them in non-mmap mode, where they have much better
latency. It is tunable with an experimental option,
x-intx-mmap-timeout-ms. A value of 0 keeps the device in non-mmap
mode after the first interrupt.
It's possible we could look at the class code of devices and come up
with reasonable per-class defaults based on expected interrupt
frequency and latency. None of this is used for MSI interrupts and
also won't be used if we can bypass through KVM.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
We cannot cast directly from pointer to uint64.
Cc: Alex Williamson <alex.williamson@redhat.com>
Cc: Alex Barcelo <abarcelo@ac.upc.edu>
Reported-by: Alex Barcelo <abarcelo@ac.upc.edu>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
This adds the core of the QEMU VFIO-based PCI device assignment driver.
To make use of this driver, enable CONFIG_VFIO, CONFIG_VFIO_IOMMU_TYPE1,
and CONFIG_VFIO_PCI in your host Linux kernel config. Load the vfio-pci
module. To assign device 0000:05:00.0 to a guest, do the following:
for dev in $(ls /sys/bus/pci/devices/0000:05:00.0/iommu_group/devices); do
vendor=$(cat /sys/bus/pci/devices/$dev/vendor)
device=$(cat /sys/bus/pci/devices/$dev/device)
if [ -e /sys/bus/pci/devices/$dev/driver ]; then
echo $dev > /sys/bus/pci/devices/$dev/driver/unbind
fi
echo $vendor $device > /sys/bus/pci/drivers/vfio-pci/new_id
done
See Documentation/vfio.txt in the Linux kernel tree for further
description of IOMMU groups and VFIO.
Then launch qemu including the option:
-device vfio-pci,host=0000:05:00.0
Legacy PCI interrupts (INTx) currently makes use of a kludge where we
trap BAR accesses and assume the access is in response to an interrupt,
therefore de-asserting and unmasking the interrupt. It's not quite as
targetted as using the EOI for this, but it's self contained and seems
to work across all architectures. The side-effect is a significant
performance slow-down for device in INTx mode. Some devices, like
graphics cards, don't really use their interrupt, so this can be turned
off with the x-intx=off option, which disables INTx alltogether. This
should be considered an experimental option until we refine this code.
Both MSI and MSI-X are supported and avoid these issues.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
