[Qemu-block] [PATCH v2 00/22] Extended I/O accounting

[Alberto Garcia]

Hi,

A few months ago I announced that I was planning to extend the I/O
accounting in QEMU based on the previous plans by Benoît and his
discussions in the mailing list in 2014.

Here are the links for reference:

   https://lists.nongnu.org/archive/html/qemu-devel/2014-08/msg04954.html
   https://lists.nongnu.org/archive/html/qemu-devel/2014-09/msg00080.html
   https://lists.nongnu.org/archive/html/qemu-block/2015-06/msg00071.html

In June I sent a patch series with the infrastructure that I was going
to use to compute averages. After that the 2.4 release happened and I
had some time to continue working on this, so this new series contains
the complete implementation of the new statistics, plus tests and a
couple of bug fixes.

The series is long but most patches are quite simple so they should be
easy to understand. It applies on top of Max's "BlockBackend and media
v5" that moves BlockAcctStats to BlockBackend.

Here's the summary of what's new:

 - New block_acct_failed() and block_acct_invalid() calls.
   We keep track now of the number of successful, failed and invalid
   operations (each one separated into read, write and flush). So from
   the API point of view, BlockDeviceStats contains 6 new fields for
   those.

 - idle_time_ns: time since the last I/O operation.

 - New BlockDeviceTimedStats struct: it has statistics for the I/O
   during a given interval of time. It keeps minimum, maximum and
   average latencies for read, write and flush operations.

   It also keeps the average read and write queue depths.

 - New 'stats-intervals' option that allows the user to define the
   intervals used to keep the aforementioned statistics. An arbitrary
   number of intervals can be specified, the length of each one is in
   seconds.

   For the API I opted for a colon-separated list of numbers,

      stats-intervals=60:3600:86400

   I also considered something a different syntax,

      stats-intervals.0.length=60,
      stats-intervals.1.length=3600,
      stats-intervals.2.length=86400

   This one could be useful if we want to specify any other attribute
   for each interval, but I couldn't come up with any, so I chose the
   simpler solution.

 - Two new options, stats-account-invalid and stats-account-failed,
   which allow the user to decide whether to count invalid and failed
   operations when computing the idle time and total latency.

 - 'supports_stats': a new field for the BlockStats structure that
   tells you whether the BDS supports statistics or not. This one can
   probably be improved by asking the device model.

I think that's all. I'm sure there will be questions and rough edges
to discuss, so I'm all yours.

Regards,

Berto

v2:
- First complete implementation of the new statistics

v1: https://lists.gnu.org/archive/html/qemu-devel/2015-06/msg03321.html
- Initial series containing only the timed average infrastructure.

Alberto Garcia (22):
  xen_disk: Account for flush operations
  ide: Account for write operations correctly
  block: define 'clock_type' for the accounting code
  util: Infrastructure for computing recent averages
  block: Add idle_time_ns to BlockDeviceStats
  block: Add "supports_stats" field to BlockStats
  block: Add statistics for failed and invalid I/O operations
  block: Allow configuring whether to account failed and invalid ops
  block: Compute minimum, maximum and average I/O latencies
  block: Add average I/O queue depth to BlockDeviceTimedStats
  block: New option to define the intervals for collecting I/O
    statistics
  qemu-io: Account for failed, invalid and flush operations
  block: Use QEMU_CLOCK_VIRTUAL for the accounting code in qtest mode
  iotests: Add test for the block device statistics
  nvme: Account for failed and invalid operations
  virtio-blk: Account for failed and invalid operations
  xen_disk: Account for failed and invalid operations
  atapi: Account for failed and invalid operations
  ide: Account for failed and invalid operations
  macio: Account for failed operations
  scsi-disk: Account for failed operations
  block: Update copyright of the accounting code

 block/accounting.c           | 118 ++++++++++++++-
 block/block-backend.c        |   1 +
 block/qapi.c                 |  53 +++++++
 blockdev.c                   |  53 +++++++
 hmp.c                        |   4 +-
 hw/block/nvme.c              |  11 +-
 hw/block/virtio-blk.c        |   4 +-
 hw/block/xen_disk.c          |  27 +++-
 hw/ide/atapi.c               |  31 ++--
 hw/ide/core.c                |  12 +-
 hw/ide/macio.c               |  12 +-
 hw/scsi/scsi-disk.c          |  46 ++++--
 include/block/accounting.h   |  28 ++++
 include/qemu/timed-average.h |  64 ++++++++
 qapi/block-core.json         | 106 ++++++++++++-
 qemu-io-cmds.c               |   9 ++
 qmp-commands.hx              |  86 ++++++++++-
 tests/Makefile               |   4 +
 tests/qemu-iotests/136       | 349 +++++++++++++++++++++++++++++++++++++++++++
 tests/qemu-iotests/136.out   |   5 +
 tests/qemu-iotests/group     |   1 +
 tests/test-timed-average.c   |  89 +++++++++++
 util/Makefile.objs           |   1 +
 util/timed-average.c         | 227 ++++++++++++++++++++++++++++
 24 files changed, 1293 insertions(+), 48 deletions(-)
 create mode 100644 include/qemu/timed-average.h
 create mode 100644 tests/qemu-iotests/136
 create mode 100644 tests/qemu-iotests/136.out
 create mode 100644 tests/test-timed-average.c
 create mode 100644 util/timed-average.c

-- 
2.5.3