For the matching kernel patch, see:
https://lkml.org/lkml/2018/2/6/866
This series enables ioeventfd support and makes use of a proposed vfio
kernel ioeventfd interface for accelerating high frequency writes
through to the device. In the specific case addressed, the writes are
to a range of MMIO space virtualized in QEMU for NVIDIA GeForce
support, but which also hosts a register which is used to allow the
MSI interrupt for the device to re-trigger. Applications which
generate a very high interrupt rate on the GPU can see noticeable
overhead as a result of this trap through QEMU. We added an option
for users to disable these quirks entirely for non-Geforce cards[1]
for optimal performance, but for GeForce users and users that can't
tweak their VM config, this gets us to within 95% of that performance
for an interrupt intensive micro-benchmark (from 83%). I'd be
interested in more typical benchmark results to understand if there's
an improvement there as well. Thanks,
Alex
[1] https://lists.gnu.org/archive/html/qemu-devel/2018-01/msg06878.html
---
Alex Williamson (5):
vfio/quirks: Add common quirk alloc helper
vfio/quirks: Add generic support for ioveventfds
vfio/quirks: Automatic ioeventfd enabling for NVIDIA BAR0 quirks
vfio: Update linux header
vfio/quirks: Enable ioeventfd quirks to be handled by vfio directly
hw/vfio/pci-quirks.c | 184 +++++++++++++++++++++++++++++++++++++-------
hw/vfio/pci.h | 13 +++
linux-headers/linux/vfio.h | 24 ++++++
3 files changed, 192 insertions(+), 29 deletions(-)