[RFC PATCH 0/1] fuse: per-queue background request accounting for io-uring

[Aleksander Trofimowicz <trof@n90.eu>]

fuse: per-queue background request accounting for io-uring

This RFC proposes an opt-in path to replace the connection-wide
background request accounting in fuse io-uring with per-queue
budgeting.

AFAIR the concept has already floated on the list, but nothing has
come to fruition. Nonetheless the problem is obvious: the existing
implementation relies on a single connection-wide spinlock
(fch->bg_lock) that is taken twice while servicing background
requests. That severely hampers the scalability of fuse io-uring for
certain types of workload (e.g. direct async I/O). Suffice it to say,
with this patch applied I can easily cross the 1M 4k random read IOPS
barrier, while the baseline gives a mere 600k IOPS (caeteris paribus).

I envision this enhancement being negotiated with a daemon, but for
your convenience the feature can currently be activated
unconditionally via the uring_max_bg_per_queue_force parameter. Mind
that the value of uring_max_bg_per_queue must also be increased to
obtain any meaningful results.

I am looking forward to your feedback. Thanks in advance.

Aleksander Trofimowicz (1):
  fuse: per-queue background request accounting for io-uring

 fs/fuse/control.c         |   5 +-
 fs/fuse/dev.c             |  93 ++++++++++++++++---
 fs/fuse/dev.h             |  22 ++++-
 fs/fuse/dev_uring.c       | 187 ++++++++++++++++++++++++++++++++++----
 fs/fuse/dev_uring_i.h     |  32 +++++++
 fs/fuse/fuse_dev_i.h      |  18 ++++
 fs/fuse/inode.c           |  13 ++-
 include/uapi/linux/fuse.h |  11 ++-
 8 files changed, 347 insertions(+), 34 deletions(-)


base-commit: 2f2f7eeef8b0a1a9b9ca998231419873ccae8d4e
--
2.56.0