mbox series

[v7,0/8] ublk: decouple server threads from ublk_queues/hctxs

Message ID 20250527-ublk_task_per_io-v7-0-cbdbaf283baa@purestorage.com
Headers show
Series ublk: decouple server threads from ublk_queues/hctxs | expand

Message

Uday Shankar May 27, 2025, 11:01 p.m. UTC 
This patch set aims to allow ublk server threads to better balance load
amongst themselves by decoupling server threads from ublk_queues/hctxs,
so that multiple threads can service I/Os that are issued from a single
CPU. This can improve performance for workloads in which ublk server CPU
is a bottleneck, and for which load is issued from CPUs which are not
balanced across ublk_queues/hctxs.

Performance
-----------

First create two ublk devices with:

ublkb0: ./kublk add -t null -q 2 --nthreads 2
ublkb1: ./kublk add -t null -q 2 --nthreads 2 --per_io_tasks

Then run load with:

taskset -c 1 fio/t/io_uring -r5 -p0 /dev/ublkb0: 1.90M IOPS
taskset -c 1 fio/t/io_uring -r5 -p0 /dev/ublkb1: 2.18M IOPS

Since ublkb1 has per-io-tasks, the second command is able to make use of
both ublk server worker threads and therefore has increased max
throughput.

Caveats:
- This testing was done on a system with 2 numa nodes, but the penalty
  of having I/O cross a numa (or LLC) boundary in the per_io_tasks case
  is quite high. So these numbers were obtained after moving all ublk
  server threads and the application threads to CPUs on the same numa
  node/LLC.
- One might expect the scaling to be linear - because ublkb1 can make
  use of twice as many ublk server threads, it should be able to drive
  twice the throughput. However this is not true (the improvement is
  ~15%), and needs further investigation.

Signed-off-by: Uday Shankar <ushankar@purestorage.com>
---
Changes in v7:
- Fix queue_rqs batch dispatch for per-io daemons
- Kick round-robin tag allocation changes to a followup
- Add explicit feature flag for per-task daemons (Ming Lei, Caleb Sander
  Mateos)
- Move some variable assignments to avoid redundant computation (Caleb
  Sander Mateos)
- Switch from storing pointers in ublk_io to computing based on address
  with container_of in a couple places (Ming Lei)
- Link to v6: https://lore.kernel.org/r/20250507-ublk_task_per_io-v6-0-a2a298783c01@purestorage.com

Changes in v6:
- Add a feature flag for this feature, called UBLK_F_RR_TAGS (Ming Lei)
- Add test for this feature (Ming Lei)
- Add documentation for this feature (Ming Lei)
- Link to v5: https://lore.kernel.org/r/20250416-ublk_task_per_io-v5-0-9261ad7bff20@purestorage.com

Changes in v5:
- Set io->task before ublk_mark_io_ready (Caleb Sander Mateos)
- Set io->task atomically, read it atomically when needed
- Return 0 on success from command-specific helpers in
  __ublk_ch_uring_cmd (Caleb Sander Mateos)
- Rename ublk_handle_need_get_data to ublk_get_data (Caleb Sander
  Mateos)
- Link to v4: https://lore.kernel.org/r/20250415-ublk_task_per_io-v4-0-54210b91a46f@purestorage.com

Changes in v4:
- Drop "ublk: properly serialize all FETCH_REQs" since Ming is taking it
  in another set
- Prevent data races by marking data structures which should be
  read-only in the I/O path as const (Ming Lei)
- Link to v3: https://lore.kernel.org/r/20250410-ublk_task_per_io-v3-0-b811e8f4554a@purestorage.com

Changes in v3:
- Check for UBLK_IO_FLAG_ACTIVE on I/O again after taking lock to ensure
  that two concurrent FETCH_REQs on the same I/O can't succeed (Caleb
  Sander Mateos)
- Link to v2: https://lore.kernel.org/r/20250408-ublk_task_per_io-v2-0-b97877e6fd50@purestorage.com

Changes in v2:
- Remove changes split into other patches
- To ease error handling/synchronization, associate each I/O (instead of
  each queue) to the last task that issues a FETCH_REQ against it. Only
  that task is allowed to operate on the I/O.
- Link to v1: https://lore.kernel.org/r/20241002224437.3088981-1-ushankar@purestorage.com

---
Uday Shankar (8):
      ublk: have a per-io daemon instead of a per-queue daemon
      selftests: ublk: kublk: plumb q_id in io_uring user_data
      selftests: ublk: kublk: tie sqe allocation to io instead of queue
      selftests: ublk: kublk: lift queue initialization out of thread
      selftests: ublk: kublk: move per-thread data out of ublk_queue
      selftests: ublk: kublk: decouple ublk_queues from ublk server threads
      selftests: ublk: add test for per io daemons
      Documentation: ublk: document UBLK_F_PER_IO_DAEMON

 Documentation/block/ublk.rst                       |  35 ++-
 drivers/block/ublk_drv.c                           | 108 +++----
 include/uapi/linux/ublk_cmd.h                      |   9 +
 tools/testing/selftests/ublk/Makefile              |   1 +
 tools/testing/selftests/ublk/fault_inject.c        |   4 +-
 tools/testing/selftests/ublk/file_backed.c         |  20 +-
 tools/testing/selftests/ublk/kublk.c               | 345 ++++++++++++++-------
 tools/testing/selftests/ublk/kublk.h               |  73 +++--
 tools/testing/selftests/ublk/null.c                |  22 +-
 tools/testing/selftests/ublk/stripe.c              |  17 +-
 tools/testing/selftests/ublk/test_generic_12.sh    |  55 ++++
 .../selftests/ublk/trace/count_ios_per_tid.bt      |  11 +
 12 files changed, 470 insertions(+), 230 deletions(-)
---
base-commit: 533c87e2ed742454957f14d7bef9f48d5a72e72d
change-id: 20250408-ublk_task_per_io-c693cf608d7a

Best regards,