This series adds the -device virtio-blk-pci,x-data-plane=on property that
enables a high performance I/O codepath.  A dedicated thread is used to process
virtio-blk requests outside the global mutex and without going through the QEMU
block layer.

Khoa Huynh <k...@us.ibm.com> reported an increase from 140,000 IOPS to 600,000
IOPS for a single VM using virtio-blk-data-plane in July:

  http://comments.gmane.org/gmane.comp.emulators.kvm.devel/94580

The virtio-blk-data-plane approach was originally presented at Linux Plumbers
Conference 2010.  The following slides contain a brief overview:

  
http://linuxplumbersconf.org/2010/ocw/system/presentations/651/original/Optimizing_the_QEMU_Storage_Stack.pdf

The basic approach is:
1. Each virtio-blk device has a thread dedicated to handling ioeventfd
   signalling when the guest kicks the virtqueue.
2. Requests are processed without going through the QEMU block layer using
   Linux AIO directly.
3. Completion interrupts are injected via irqfd from the dedicated thread.

To try it out:

  qemu -drive if=none,id=drive0,cache=none,aio=native,format=raw,file=...
       -device 
virtio-blk-pci,drive=drive0,scsi=off,config-wce=off,x-data-plane=on

Limitations:
 * Only format=raw is supported
 * Live migration is not supported
 * Block jobs, hot unplug, and other operations fail with -EBUSY
 * I/O throttling limits are ignored
 * Only Linux hosts are supported due to Linux AIO usage

The code has reached a stage where I feel it is ready to merge.  Users have
been playing with it for some time and want the significant performance boost.

We are refactoring QEMU to get rid of the global mutex.  I believe that
virtio-blk-data-plane can eventually become the default mode of operation.

Instead of waiting for global mutex removal efforts to finish, I want to use
virtio-blk-data-plane as an example device for AioContext and threaded hw
dispatch refactoring.  This means:

1. When the block layer can bind to an AioContext and execute I/O outside the
   global mutex, virtio-blk-data-plane can use this (and gain image format
   support).

2. When hw dispatch no longer needs the global mutex we can use hw/virtio.c
   again and perhaps run a pool of iothreads instead of dedicated data plane
   threads.

But in the meantime, I have cleaned up the virtio-blk-data-plane code so that
it can be merged as an experimental feature.

v7:
 * VIRTIO_BLK_T_GET_ID support
 * Replace lock/condvar with drain operation that stops data plane thread 
[Michael, Paolo, Laszlo]
 * Add vring_pop() TODO about crossing memory region boundaries [Michael]
 * Move #ifdef CONFIG_VIRTIO_BLK_DATA_PLANE to hw/virtio-blk.c [Michael]
 * Typo s/there is/there is no/ in hostmem.h [Paolo]
 * Avoid potential integer overflow in hostmem.c [Laszlo]
 * Retry epoll_wait() on EINTR so gdb works

v6:
 * Move hw/Makefile.objs dataplane/ inclusion from Patch 4 to Patch 3 [Kevin]
 * Split discard() with front/back and switch ssize_t to size_t [Michael]
 * Disable WCE config feature [Michael]
 * Assert on ioq underflow/overflow, it can never happen [Kevin]
 * Propagate fdatasync() errors [Kevin]
 * Remember to init/destroy hostmem mutex
 * Declare VirtIOBlkConf->data_plane in the right patch so building works

v5:
 * Omit memory regions with dirty logging enabled from hostmem [Michael]
 * Add doc comment about quiescing requests across memory hot unplug [Michael]
 * Clarify which Linux vhost version the vring code originates from [Michael]
 * Break up indirect vring buffer into 1 hostmem_lookup() per descriptor 
[Michael]
 * Barriers in hw/dataplane/vring.c to force fields to be loaded [Michael]
 * split vring_set_notification() into enable/disable [Paolo]
 * barriers in vring.c instead of virtio-blk.c [Michael]
 * move setup code from hw/virtio-blk.c into hw/dataplane/virtio-blk.c [Michael]

 * Note I did not get rid of the mutex+condvar approach to draining requests.
   I've had good feedback on the performance of the patch series so I'm not
   worried about eliminating the lock (it's very rarely contended).  Hope
   Michael and Paolo are okay with this approach.

v4:
 * Add qemu_iovec_concat_iov() [Paolo]
 * Use QEMUIOVector to copy out virtio_blk_inhdr [Michael, Paolo]

v3:
 * Don't assume iovec layout [Michael]
 * Better naming for hostmem.c MemoryListener callbacks [Don]
 * More vring quarantining if commands are bogus instead of exiting [Blue]

v2:
 * Use MemoryListener for thread-safe memory mapping [Paolo, Anthony, and 
everyone else pointed this out ;-)]
 * Quarantine invalid vring instead of exiting [Blue]
 * Replace __u16 kernel types with uint16_t [Blue]

Changes from the RFC v9:
 * Add x-data-plane=on|off option and coexist with regular virtio-blk code
 * Create thread from BH so it inherits iothread cpusets
 * Drain requests on vm_stop() so stopped guest does not access image file
 * Add migration blocker
 * Add bdrv_in_use() to prevent block jobs and other operations that can 
interfere
 * Drop IOQueue request merging for simplicity
 * Drop ioctl interrupt injection and always use irqfd for simplicity
 * Major cleanup to split up source files
 * Rebase from qemu-kvm.git onto qemu.git
 * Address Michael Tsirkin's review comments

Stefan Hajnoczi (12):
  raw-posix: add raw_get_aio_fd() for virtio-blk-data-plane
  configure: add CONFIG_VIRTIO_BLK_DATA_PLANE
  dataplane: add host memory mapping code
  dataplane: add virtqueue vring code
  dataplane: add event loop
  dataplane: add Linux AIO request queue
  iov: add iov_discard_front/back() to remove data
  test-iov: add iov_discard_front/back() testcases
  iov: add qemu_iovec_concat_iov()
  virtio-blk: restore VirtIOBlkConf->config_wce flag
  dataplane: add virtio-blk data plane code
  virtio-blk: add x-data-plane=on|off performance feature

 block.h                    |   9 +
 block/raw-posix.c          |  34 ++++
 configure                  |  21 ++
 hw/Makefile.objs           |   2 +-
 hw/dataplane/Makefile.objs |   3 +
 hw/dataplane/event-poll.c  | 100 ++++++++++
 hw/dataplane/event-poll.h  |  40 ++++
 hw/dataplane/hostmem.c     | 176 +++++++++++++++++
 hw/dataplane/hostmem.h     |  57 ++++++
 hw/dataplane/ioq.c         | 117 ++++++++++++
 hw/dataplane/ioq.h         |  57 ++++++
 hw/dataplane/virtio-blk.c  | 465 +++++++++++++++++++++++++++++++++++++++++++++
 hw/dataplane/virtio-blk.h  |  29 +++
 hw/dataplane/vring.c       | 362 +++++++++++++++++++++++++++++++++++
 hw/dataplane/vring.h       |  63 ++++++
 hw/virtio-blk.c            |  47 ++++-
 hw/virtio-blk.h            |   5 +-
 hw/virtio-pci.c            |   4 +
 iov.c                      |  90 +++++++--
 iov.h                      |  13 ++
 qemu-common.h              |   3 +
 tests/test-iov.c           | 150 +++++++++++++++
 trace-events               |   9 +
 23 files changed, 1840 insertions(+), 16 deletions(-)
 create mode 100644 hw/dataplane/Makefile.objs
 create mode 100644 hw/dataplane/event-poll.c
 create mode 100644 hw/dataplane/event-poll.h
 create mode 100644 hw/dataplane/hostmem.c
 create mode 100644 hw/dataplane/hostmem.h
 create mode 100644 hw/dataplane/ioq.c
 create mode 100644 hw/dataplane/ioq.h
 create mode 100644 hw/dataplane/virtio-blk.c
 create mode 100644 hw/dataplane/virtio-blk.h
 create mode 100644 hw/dataplane/vring.c
 create mode 100644 hw/dataplane/vring.h

-- 
1.8.0.2


Reply via email to