- 11 Apr, 2021 40 commits
-
-
Jens Axboe authored
The default io_uring poll mode is one-shot, where once the event triggers, the poll command is completed and won't trigger any further events. If we're doing repeated polling on the same file or socket, then it can be more efficient to do multishot, where we keep triggering whenever the event becomes true. This deviates from the usual norm of having one CQE per SQE submitted. Add a CQE flag, IORING_CQE_F_MORE, which tells the application to expect further completion events from the submitted SQE. Right now the only user of this is POLL_ADD in multishot mode. Since sqe->poll_events is using the space that we normally use for adding flags to commands, use sqe->len for the flag space for POLL_ADD. Multishot mode is selected by setting IORING_POLL_ADD_MULTI in sqe->len. An application should expect more CQEs for the specificed SQE if the CQE is flagged with IORING_CQE_F_MORE. In multishot mode, only cancelation or an error will terminate the poll request, in which case the flag will be cleared. Signed-off-by:Jens Axboe <axboe@kernel.dk>
-
Jens Axboe authored
We should be including the completion flags for better introspection on exactly what completion event was logged. Signed-off-by: -
Pavel Begunkov authored
Instead of using a request itself for overflowed CQE stashing, allocate a separate entry. The disadvantage is that the allocation may fail and it will be accounted as lost (see rings->cq_overflow), so we lose reliability in case of memory pressure if the application is driving the CQ ring into overflow. However, it opens a way for for multiple CQEs per an SQE and even generating SQE-less CQEs. Signed-off-by:
Pavel Begunkov <asml.silence@gmail.com> [axboe: use GFP_ATOMIC | __GFP_ACCOUNT] Signed-off-by:
-
Jens Axboe authored
Instead of masking these in as part of regular POLL_ADD prep, do it in io_init_poll_iocb(), and include NVAL as that's generally unmaskable, and RDHUP alongside the HUP that is already set. Signed-off-by: -
Pavel Begunkov authored
Expect read/write to succeed and create a hot path for this case, in particular hide all error handling with resubmission under a single check with the desired result. Signed-off-by:
-
Pavel Begunkov authored
Move iov_iter_revert() resetting iterator in case of -EIOCBQUEUED into io_resubmit_prep(), so we don't do heavy revert in hot path, also saves a couple of checks. Signed-off-by:
-
Pavel Begunkov authored
When reissue_prep failed in io_complete_rw_iopoll(), we change return code to -EIO to prevent io_iopoll_complete() from doing resubmission. Mark requests with a new flag (i.e. REQ_F_DONT_REISSUE) instead and retain the original return value. It also removes io_rw_reissue() from io_iopoll_complete() that will be used later. Signed-off-by:
-
Pavel Begunkov authored
file_end_write() is only for regular files, so the function do a couple of dereferences to get inode and check for it. However, we already have REQ_F_ISREG at hand, just use it and inline file_end_write(). Signed-off-by:
-
Pavel Begunkov authored
current->files are always valid now even for io-wq threads, so kill not used anymore REQ_F_NO_FILE_TABLE. Signed-off-by:
-
Pavel Begunkov authored
req->work is mostly unused unless it's punted, and io_init_req() is too hot for fully initialising it. Fortunately, we can skip init work.next as it's controlled by io-wq, and can not touch work.flags by moving everything related into io_prep_async_work(). The only field left is req->work.creds, but there is nothing can be done, keep maintaining it. Signed-off-by:
-
Pavel Begunkov authored
Extract a helper for io_work_get_acct() and io_wqe_get_acct() to avoid duplication. Signed-off-by:
-
Pavel Begunkov authored
struct io_uring_task::sqpoll is not used anymore, kill it Signed-off-by:
-
Pavel Begunkov authored
io_match_task() matches all requests with PF_EXITING task, even though those may be valid requests. It was necessary for SQPOLL cancellation, but now it kills all requests before exiting via io_uring_cancel_sqpoll(), so it's not needed. Signed-off-by:
-
Pavel Begunkov authored
REQ_F_LINK_TIMEOUT is a hint that to look for linked timeouts to cancel, we're leaving it even when it's already fired. Hence don't care to clear it in io_kill_linked_timeout(), it's safe and is called only once. Signed-off-by:
-
Pavel Begunkov authored
Inline io_task_work_add() into io_req_task_work_add(). They both work with a request, so keeping them separate doesn't make things much more clear, but merging allows optimise it. Apart from small wins like not reading req->ctx or not calculating @notify in the hot path, i.e. with tctx->task_state set, it avoids doing wake_up_process() for every single add, but only after actually done task_work_add(). Signed-off-by:
-
Pavel Begunkov authored
io_put_file() doesn't do a good job at generating a good code. Inline it, so we can check REQ_F_FIXED_FILE first, prioritising FIXED_FILE case over requests without files, and saving a memory load in that case. Signed-off-by:
-
Pavel Begunkov authored
Reshuffle io_dismantle_req() checks to put most of slow path stuff under a single if. Signed-off-by:
-
Pavel Begunkov authored
Inline io_clean_op(), leaving __io_clean_op() but renaming it. This will be used in following patches. Signed-off-by:
-
Pavel Begunkov authored
Both io_req_complete_failed() and __io_req_task_cancel() do the same thing: set failure flag, put both req refs and emit an CQE. The former one is a bit more advance as it puts req back into a req cache, so make it to take over __io_req_task_cancel() and remove the last one. Signed-off-by:
-
Pavel Begunkov authored
Add a new helper io_flush_cached_locked_reqs() that splices locked_free_list to free_list, and does it right doing all sync and invariant reinit. Signed-off-by:
-
Pavel Begunkov authored
We don't care about ret value in io_free_req_deferred(), make the code a bit more concise. Signed-off-by:
-
Pavel Begunkov authored
One big omission is that io_put_req() haven't been marked inline, and at least gcc 9 doesn't inline it, not to mention that it's really hot and extra function call is intolerable, especially when it doesn't put a final ref. Signed-off-by:
-
Pavel Begunkov authored
There are two problems: 1) we always allocate refnodes in advance and free them if those haven't been used. It's expensive, takes two allocations, where one of them is percpu. And it may be pretty common not actually using them. 2) Current API with allocating a refnode and setting some of the fields is error prone, we don't ever want to have a file node runninng fixed buffer callback... Solve both with pre-init/get API. Pre-init just leaves the node for later if not used, and for get (i.e. io_rsrc_refnode_get()), you need to explicitly pass all arguments setting callbacks/etc., so it's more resilient. Signed-off-by:
-
Pavel Begunkov authored
Emphasize that return value of io_flush_cached_reqs() depends on number of requests in the cache. It looks nicer and might help tools from false-negative analyses. Signed-off-by:
-
Pavel Begunkov authored
Move the case of successfully issued request by doing that check first. It's not much of a difference, just generates slightly better code for me. Signed-off-by:
-
Pavel Begunkov authored
Inline __io_queue_linked_timeout(), we don't need it Signed-off-by:
-
Pavel Begunkov authored
Don't do a function call (io_dismantle_req()) in the middle and place it to near other function calls, otherwise may lead to excessive register spilling. Signed-off-by:
-
Pavel Begunkov authored
First of all, w need to set tctx->sqpoll only when we add a new entry into ->xa, so move it from the hot path. Also extract a hot path for io_uring_add_task_file() as an inline helper. Signed-off-by:
-
Pavel Begunkov authored
Add unlikely annotations, because my compiler pretty much mispredicts every first check, and apart jumping around in the fast path, it also generates extra instructions, like in advance setting ret value. Signed-off-by:
-
Pavel Begunkov authored
__tctx_task_work() guarantees that ctx won't be killed while running task_works, so we can remove now unnecessary ctx pinning for internally armed polling. Signed-off-by:
-
Jens Axboe authored
We can set canceled == true and complete out-of-line, ensure that we catch that and correctly return -ECANCELED if the poll operation got canceled. Signed-off-by: -
Jens Axboe authored
The correct function is io_iopoll_complete(), which deals with completions of IOPOLL requests, not io_poll_complete(). Signed-off-by: -
Jens Axboe authored
We have to dig quite deep to check for particularly whether or not a file supports a fast-path nonblock attempt. For fixed files, we can do this lookup once and cache the state instead. This adds two new bits to track whether we support async read/write attempt, and lines up the REQ_F_ISREG bit with those two. The file slot re-uses the last 3 (or 2, for 32-bit) of the file pointer to cache that state, and then we mask it in when we go and use a fixed file. Signed-off-by: -
Jens Axboe authored
We don't allow them at registration time, so limit the check for needing inflight tracking in io_file_get() to the non-fixed path. Signed-off-by: -
Pavel Begunkov authored
Use a more comprehensible() max instead of hand coding it with ifs in io_sqd_update_thread_idle(). Signed-off-by:
-
Jens Axboe authored
io_uring manipulates references twice for each request, and hence is very sensitive to performance of the reference count. This commit borrows a trick from: commit f958d7b5 Author: Linus Torvalds <torvalds@linux-foundation.org> Date: Thu Apr 11 10:06:20 2019 -0700 mm: make page ref count overflow check tighter and more explicit and switches to atomic_t for references, while still retaining overflow and underflow checks. This is good for a 2-3% increase in peak IOPS on a single core. Before: IOPS=2970879, IOS/call=31/31, inflight=128 (128) IOPS=2952597, IOS/call=31/31, inflight=128 (128) IOPS=2943904, IOS/call=31/31, inflight=128 (128) IOPS=2930006, IOS/call=31/31, inflight=96 (96) and after: IOPS=3054354, IOS/call=31/31, inflight=128 (128) IOPS=3059038, IOS/call=31/31, inflight=128 (128) IOPS=3060320, IOS/call=31/31, inflight=128 (128) IOPS=3068256, IOS/call=31/31, inflight=96 (96) Signed-off-by:
-
Jens Axboe authored
No functional changes in this patch, just in preparation for handling the references a bit more efficiently. Signed-off-by: -
Pavel Begunkov authored
If not for async_data NULL check, io_resubmit_prep() is already an rw specific version of io_req_prep_async(), but slower because 1) it always goes through io_import_iovec() even if following io_setup_async_rw() the result 2) instead of initialising iovec/iter in-place it does it on-stack and then copies with io_setup_async_rw(). Signed-off-by:
-
Pavel Begunkov authored
Merge two function and do renaming in favour of the second one, it relays the meaning better. Signed-off-by:
-
Pavel Begunkov authored
needs_async_data controls allocation of async_data, and used in two cases. 1) when async setup requires it (by io_req_prep_async() or handler themselves), and 2) when op always needs additional space to operate, like timeouts do. Opcode preps already don't bother about the second case and do allocation unconditionally, restrict needs_async_data to the first case only and rename it into needs_async_setup. Signed-off-by:
-
