public inbox for [email protected]
 help / color / mirror / Atom feed
From: Pavel Begunkov <[email protected]>
To: Hao Xu <[email protected]>, Jens Axboe <[email protected]>,
	[email protected]
Subject: Re: [PATCH v2 04/24] io_uring: use slist for completion batching
Date: Tue, 28 Sep 2021 10:41:36 +0100	[thread overview]
Message-ID: <[email protected]> (raw)
In-Reply-To: <[email protected]>

On 9/26/21 7:57 AM, Hao Xu wrote:
> 在 2021/9/25 上午4:59, Pavel Begunkov 写道:
>> Currently we collect requests for completion batching in an array.
>> Replace them with a singly linked list. It's as fast as arrays but
>> doesn't take some much space in ctx, and will be used in future patches.
>>
>> Signed-off-by: Pavel Begunkov <[email protected]>
>> ---
>>   fs/io_uring.c | 52 +++++++++++++++++++++++++--------------------------
>>   1 file changed, 25 insertions(+), 27 deletions(-)
>>
>> diff --git a/fs/io_uring.c b/fs/io_uring.c
>> index 9c14e9e722ba..9a76c4f84311 100644
>> --- a/fs/io_uring.c
>> +++ b/fs/io_uring.c
>> @@ -322,8 +322,8 @@ struct io_submit_state {
>>       /*
>>        * Batch completion logic
>>        */
>> -    struct io_kiocb        *compl_reqs[IO_COMPL_BATCH];
>> -    unsigned int        compl_nr;
>> +    struct io_wq_work_list    compl_reqs;
> Will it be better to rename struct io_wq_work_list to something more
> generic, io_wq_work_list is a bit confused, we are now using this
> type of linked list (stack as well) for various aim, not just to link
> iowq works.

Was thinking about it, e.g. io_slist, but had been already late --
lots of conflicts and a good chance to add a couple of extra bugs
on rebase. I think we can do it afterward (if ever considering
it troubles backporting)


>> +
>>       /* inline/task_work completion list, under ->uring_lock */
>>       struct list_head    free_list;
>>   };
>> @@ -883,6 +883,8 @@ struct io_kiocb {
>>       struct io_wq_work        work;
>>       const struct cred        *creds;
>>   +    struct io_wq_work_node        comp_list;
>> +
>>       /* store used ubuf, so we can prevent reloading */
>>       struct io_mapped_ubuf        *imu;
>>   };
>> @@ -1169,7 +1171,7 @@ static inline void req_ref_get(struct io_kiocb *req)
>>     static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
>>   {
>> -    if (ctx->submit_state.compl_nr)
>> +    if (!wq_list_empty(&ctx->submit_state.compl_reqs))
>>           __io_submit_flush_completions(ctx);
>>   }
>>   @@ -1326,6 +1328,7 @@ static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
>>       INIT_LIST_HEAD(&ctx->submit_state.free_list);
>>       INIT_LIST_HEAD(&ctx->locked_free_list);
>>       INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
>> +    INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
>>       return ctx;
>>   err:
>>       kfree(ctx->dummy_ubuf);
>> @@ -1831,11 +1834,16 @@ static inline bool io_req_needs_clean(struct io_kiocb *req)
>>   static void io_req_complete_state(struct io_kiocb *req, long res,
>>                     unsigned int cflags)
>>   {
>> +    struct io_submit_state *state;
>> +
>>       if (io_req_needs_clean(req))
>>           io_clean_op(req);
>>       req->result = res;
>>       req->compl.cflags = cflags;
>>       req->flags |= REQ_F_COMPLETE_INLINE;
>> +
>> +    state = &req->ctx->submit_state;
>> +    wq_list_add_tail(&req->comp_list, &state->compl_reqs);
>>   }
>>     static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
>> @@ -2324,13 +2332,14 @@ static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
>>   static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
>>       __must_hold(&ctx->uring_lock)
>>   {
>> +    struct io_wq_work_node *node, *prev;
>>       struct io_submit_state *state = &ctx->submit_state;
>> -    int i, nr = state->compl_nr;
>>       struct req_batch rb;
>>         spin_lock(&ctx->completion_lock);
>> -    for (i = 0; i < nr; i++) {
>> -        struct io_kiocb *req = state->compl_reqs[i];
>> +    wq_list_for_each(node, prev, &state->compl_reqs) {
>> +        struct io_kiocb *req = container_of(node, struct io_kiocb,
>> +                            comp_list);
>>             __io_cqring_fill_event(ctx, req->user_data, req->result,
>>                       req->compl.cflags);
>> @@ -2340,15 +2349,18 @@ static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
>>       io_cqring_ev_posted(ctx);
>>         io_init_req_batch(&rb);
>> -    for (i = 0; i < nr; i++) {
>> -        struct io_kiocb *req = state->compl_reqs[i];
>> +    node = state->compl_reqs.first;
>> +    do {
>> +        struct io_kiocb *req = container_of(node, struct io_kiocb,
>> +                            comp_list);
>>   +        node = req->comp_list.next;
>>           if (req_ref_put_and_test(req))
>>               io_req_free_batch(&rb, req, &ctx->submit_state);
>> -    }
>> +    } while (node);
>>         io_req_free_batch_finish(ctx, &rb);
>> -    state->compl_nr = 0;
>> +    INIT_WQ_LIST(&state->compl_reqs);
>>   }
>>     /*
>> @@ -2668,17 +2680,10 @@ static void io_req_task_complete(struct io_kiocb *req, bool *locked)
>>       unsigned int cflags = io_put_rw_kbuf(req);
>>       long res = req->result;
>>   -    if (*locked) {
>> -        struct io_ring_ctx *ctx = req->ctx;
>> -        struct io_submit_state *state = &ctx->submit_state;
>> -
>> +    if (*locked)
>>           io_req_complete_state(req, res, cflags);
>> -        state->compl_reqs[state->compl_nr++] = req;
>> -        if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
>> -            io_submit_flush_completions(ctx);
>> -    } else {
>> +    else
>>           io_req_complete_post(req, res, cflags);
>> -    }
>>   }
>>     static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
>> @@ -6969,15 +6974,8 @@ static void __io_queue_sqe(struct io_kiocb *req)
>>        * doesn't support non-blocking read/write attempts
>>        */
>>       if (likely(!ret)) {
>> -        if (req->flags & REQ_F_COMPLETE_INLINE) {
>> -            struct io_ring_ctx *ctx = req->ctx;
>> -            struct io_submit_state *state = &ctx->submit_state;
>> -
>> -            state->compl_reqs[state->compl_nr++] = req;
>> -            if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
>> -                io_submit_flush_completions(ctx);
>> +        if (req->flags & REQ_F_COMPLETE_INLINE)
>>               return;
>> -        }
>>             linked_timeout = io_prep_linked_timeout(req);
>>           if (linked_timeout)
>>
> 

-- 
Pavel Begunkov

  reply	other threads:[~2021-09-28  9:42 UTC|newest]

Thread overview: 32+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-09-24 20:59 [PATCH v2 00/24] rework and optimise submission+completion paths Pavel Begunkov
2021-09-24 20:59 ` [PATCH v2 01/24] io_uring: mark having different creds unlikely Pavel Begunkov
2021-09-24 20:59 ` [PATCH v2 02/24] io_uring: force_nonspin Pavel Begunkov
2021-09-24 20:59 ` [PATCH v2 03/24] io_uring: make io_do_iopoll return number of reqs Pavel Begunkov
2021-09-24 20:59 ` [PATCH v2 04/24] io_uring: use slist for completion batching Pavel Begunkov
2021-09-26  6:57   ` Hao Xu
2021-09-28  9:41     ` Pavel Begunkov [this message]
2021-09-28 15:32       ` Jens Axboe
2021-09-24 20:59 ` [PATCH v2 05/24] io_uring: remove allocation cache array Pavel Begunkov
2021-09-24 20:59 ` [PATCH v2 06/24] io-wq: add io_wq_work_node based stack Pavel Begunkov
2021-09-24 20:59 ` [PATCH v2 07/24] io_uring: replace list with stack for req caches Pavel Begunkov
2021-09-24 20:59 ` [PATCH v2 08/24] io_uring: split iopoll loop Pavel Begunkov
2021-09-24 20:59 ` [PATCH v2 09/24] io_uring: use single linked list for iopoll Pavel Begunkov
2021-09-24 20:59 ` [PATCH v2 10/24] io_uring: add a helper for batch free Pavel Begunkov
2021-09-26  3:36   ` Hao Xu
2021-09-28  9:33     ` Pavel Begunkov
2021-09-24 20:59 ` [PATCH v2 11/24] io_uring: convert iopoll_completed to store_release Pavel Begunkov
2021-09-24 20:59 ` [PATCH v2 12/24] io_uring: optimise batch completion Pavel Begunkov
     [not found]   ` <CAFUsyfLSXMvd_MBAp83qriW7LD=bg2=25TC4e_X4oMO1atoPYg@mail.gmail.com>
2021-09-28  9:35     ` Pavel Begunkov
2021-09-24 20:59 ` [PATCH v2 13/24] io_uring: inline completion batching helpers Pavel Begunkov
2021-09-24 20:59 ` [PATCH v2 14/24] io_uring: don't pass tail into io_free_batch_list Pavel Begunkov
2021-09-24 20:59 ` [PATCH v2 15/24] io_uring: don't pass state to io_submit_state_end Pavel Begunkov
2021-09-24 20:59 ` [PATCH v2 16/24] io_uring: deduplicate io_queue_sqe() call sites Pavel Begunkov
2021-09-24 20:59 ` [PATCH v2 17/24] io_uring: remove drain_active check from hot path Pavel Begunkov
2021-09-24 20:59 ` [PATCH v2 18/24] io_uring: split slow path from io_queue_sqe Pavel Begunkov
2021-09-24 20:59 ` [PATCH v2 19/24] io_uring: inline hot path of __io_queue_sqe() Pavel Begunkov
2021-09-24 21:00 ` [PATCH v2 20/24] io_uring: reshuffle queue_sqe completion handling Pavel Begunkov
2021-09-24 21:00 ` [PATCH v2 21/24] io_uring: restructure submit sqes to_submit checks Pavel Begunkov
2021-09-24 21:00 ` [PATCH v2 22/24] io_uring: kill off ->inflight_entry field Pavel Begunkov
2021-09-24 21:00 ` [PATCH v2 23/24] io_uring: comment why inline complete calls io_clean_op() Pavel Begunkov
2021-09-24 21:00 ` [PATCH v2 24/24] io_uring: disable draining earlier Pavel Begunkov
2021-09-30 16:04 ` [PATCH v2 00/24] rework and optimise submission+completion paths Jens Axboe

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    [email protected] \
    [email protected] \
    [email protected] \
    [email protected] \
    [email protected] \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox