io_uring/cancel: futex and waitid cancel results
cover io_uring/cancel: futex and waitid cancel results
io_futex_cancel() and io_waitid_cancel() return the number of requests they canceled, where the rest of what io_try_cancel() calls returns 0. Commit c74193430feb ("io_uring/timeout: don't end a linked timeout with 1") took care of what that did to a linked timeout.
liburing's futex, waitid, io-cancel, sync-cancel, poll-cancel-all, cancel-race, futex-kill, link-timeout, timeout and poll-cancel tests pass on both patched kernels. Unchanged liburing tests pass with patch 1 and fail with patch 2, where they check for the 1.
agxr (2):
io_uring/cancel: count all canceled futex and waitid requests on
CANCEL_ALL
io_uring/cancel: complete single futex or waitid cancel with 0
io_uring/cancel.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
base-commit: ecb6817b769fd7d71af2c9f446188649020b2e53
Signature · 1 line
2.56.0
PATCH 1/2 io_uring/cancel: count all canceled futex and waitid requests on CANCEL_ALL
With IORING_ASYNC_CANCEL_ALL, __io_async_cancel() calls io_try_cancel() until it finds nothing more and adds 1 to the count for each call that found something, which is correct for poll, io-wq and timeouts, which cancel one request per call. io_futex_cancel() and io_waitid_cancel() cancel all matches in a single call and return how many they canceled. Three futex waits with the same user_data are all canceled by the first call, which counts as one, and the cancel completes with 1, not 3.
Sum the return values of io_try_cancel() instead, so that the count is the number of canceled requests, as io_uring_prep_cancel(3) documents.
--- io_uring/cancel.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
io_uring/cancel.c +1 -1
@@ -198,7 +198,7 @@ static int __io_async_cancel(struct io_cancel_data *cd,break ;if (!all)return ret; - nr++; + nr += ret >0 ? ret :1 ; }while (1 );/* slow path, try all io-wq's */
Signature · 1 line
2.56.0
PATCH 2/2 io_uring/cancel: complete single futex or waitid cancel with 0
In reply to agxrA cancel of one request completes with 0, as io_uring_prep_cancel(3) documents and as poll, io-wq and timeouts do. For a futex wait or a waitid it completes with 1. io_futex_cancel() and io_waitid_cancel() return the number canceled, and __io_async_cancel() hands it on, for IORING_OP_ASYNC_CANCEL and IORING_REGISTER_SYNC_CANCEL alike.
Turn that count into 0, as io_req_task_link_timeout() does since commit c74193430feb ("io_uring/timeout: don't end a linked timeout with 1").
liburing's futex and waitid tests expect the 1, and a liburing change makes them accept 0 as well.
--- io_uring/cancel.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
io_uring/cancel.c +1 -1
@@ -197,7 +197,7 @@ static int __io_async_cancel(struct io_cancel_data *cd,if (ret == -ENOENT )break ;if (!all) -return ret; +return ret >0 ?0 : ret; nr += ret >0 ? ret :1 ; }while (1 );
Signature · 1 line
2.56.0
Jens Axboe
maintainerQuoted 10 lines · agxr
io_futex_cancel()andio_waitid_cancel()return the number of requests they canceled, where the rest of whatio_try_cancel()calls returns 0. Commitc74193430feb("io_uring/timeout: don't end a linked timeout with 1") took care of what that did to a linked timeout.liburing's futex, waitid, io-cancel, sync-cancel, poll-cancel-all, cancel-race, futex-kill, link-timeout, timeout and poll-cancel tests pass on both patched kernels. Unchanged liburing tests pass with patch 1 and fail with patch 2, where they check for the 1.
Quoted 7 lines
agxr (2):
io_uring/cancel: count all canceled futex and waitid requests on
CANCEL_ALL
io_uring/cancel: complete single futex or waitid cancel with 0io_uring/cancel.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
I'll take a look at these - but please note, that:
is not going to be adequate for the kernel, we generally ask for a known identity. That's either a real name or at least a well known identity, above is neither of those.
Signature · 1 line
Jens Axboe