This series adds a new protocol driver that is intended to achieve about 20%
better performance for latency bound workloads (i.e. synchronous I/O) than
linux-aio when guest is exclusively accessing a NVMe device, by talking to the
device directly instead of through kernel file system layers and its NVMe
driver.
This applies on top of Stefan's block-next tree which has the busy polling
patches - the new driver also supports it.
A git branch is also available as:
https://github.com/famz/qemu nvme
See patch 4 for benchmark numbers.
Tests are done on QEMU's NVMe emulation and a real Intel P3700 SSD NVMe card.
Most of dd/fio/mkfs/kernel build and OS installation testings work well, but an
weird write fault looking similar to [1] is consistently seen when installing
RHEL 7.3 guest, which is still under investigation.
[1]: http://lists.infradead.org/pipermail/linux-nvme/2015-May/001840.html
Also, the ram notifier is not enough for hot plugged block device because in
that case the notifier is installed _after_ ram blocks are added so it won't
get the events.
Fam Zheng (3):
util: Add a notifier list for qemu_vfree()
util: Add VFIO helper library
block: Add VFIO based NVMe driver
Paolo Bonzini (1):
ramblock-notifier: new
block/Makefile.objs | 1 +
block/nvme.c | 1026 +++++++++++++++++++++++++++++++++++++++++++
exec.c | 5 +
include/exec/memory.h | 6 +-
include/exec/ram_addr.h | 46 +-
include/exec/ramlist.h | 72 +++
include/qemu/notify.h | 1 +
include/qemu/vfio-helpers.h | 29 ++
numa.c | 29 ++
stubs/Makefile.objs | 1 +
stubs/ram.c | 10 +
util/Makefile.objs | 1 +
util/oslib-posix.c | 9 +
util/vfio-helpers.c | 713 ++++++++++++++++++++++++++++++
xen-mapcache.c | 3 +
15 files changed, 1903 insertions(+), 49 deletions(-)
create mode 100644 block/nvme.c
create mode 100644 include/exec/ramlist.h
create mode 100644 include/qemu/vfio-helpers.h
create mode 100644 stubs/ram.c
create mode 100644 util/vfio-helpers.c
--
2.9.3
