On 6/19/22 12:26, Pavel Begunkov wrote:
Apart from this patches removing some implicit assumptions, which we had problems with before, and making code cleaner, they and especially 6-7 are also needed to push for synchronisation optimisations later, lile [1] or removing spinlocking with SINGLE_ISSUER. The downside is that we add additional lock/unlock into eventfd path, but I don't think we care about it.
just in case there conflicts, it's based on top of "[PATCH for-next 0/4] simple cleanups" from yesterday
The series also exposes a minor issue with cancellations, which for some reason calls io_kill_timeouts() and io_poll_remove_all() too many times on task exit. That makes poll-cancel to timeout on sigalarm, though usually is fine if given 3-5 sec instead of 1. We'll investigate it later. [1] https://github.com/isilence/linux/commit/6224f58bf7b542e6aed1eed44ee6bd5b5f706437 Pavel Begunkov (7): io_uring: remove extra io_commit_cqring() io_uring: reshuffle io_uring/io_uring.h io_uring: move io_eventfd_signal() io_uring: hide eventfd assumptions in evenfd paths io_uring: remove ->flush_cqes optimisation io_uring: introduce locking helpers for CQE posting io_uring: add io_commit_cqring_flush() include/linux/io_uring_types.h | 2 + io_uring/io_uring.c | 144 ++++++++++++++++----------------- io_uring/io_uring.h | 108 ++++++++++++++----------- io_uring/rw.c | 5 +- io_uring/timeout.c | 7 +- 5 files changed, 133 insertions(+), 133 deletions(-)
-- Pavel Begunkov
