On 3/10/22 6:53 AM, Pavel Begunkov wrote:
> On 3/10/22 02:33, Jens Axboe wrote:
>> On 3/9/22 6:55 PM, Jens Axboe wrote:
>>> On 3/9/22 6:36 PM, Jens Axboe wrote:
>>>> On 3/9/22 4:49 PM, Artyom Pavlov wrote:
>>>>> Greetings!
>>>>>
>>>>> A common approach for multi-threaded servers is to have a number of
>>>>> threads equal to a number of cores and launch a separate ring in each
>>>>> one. AFAIK currently if we want to send an event to a different ring,
>>>>> we have to write-lock this ring, create SQE, and update the index
>>>>> ring. Alternatively, we could use some kind of user-space message
>>>>> passing.
>>>>>
>>>>> Such approaches are somewhat inefficient and I think it can be solved
>>>>> elegantly by updating the io_uring_sqe type to allow accepting fd of a
>>>>> ring to which CQE must be sent by kernel. It can be done by
>>>>> introducing an IOSQE_ flag and using one of currently unused padding
>>>>> u64s.
>>>>>
>>>>> Such feature could be useful for load balancing and message passing
>>>>> between threads which would ride on top of io-uring, i.e. you could
>>>>> send NOP with user_data pointing to a message payload.
>>>>
>>>> So what you want is a NOP with 'fd' set to the fd of another ring, and
>>>> that nop posts a CQE on that other ring? I don't think we'd need IOSQE
>>>> flags for that, we just need a NOP that supports that. I see a few ways
>>>> of going about that:
>>>>
>>>> 1) Add a new 'NOP' that takes an fd, and validates that that fd is an
>>>> io_uring instance. It can then grab the completion lock on that ring
>>>> and post an empty CQE.
>>>>
>>>> 2) We add a FEAT flag saying NOP supports taking an 'fd' argument, where
>>>> 'fd' is another ring. Posting CQE same as above.
>>>>
>>>> 3) We add a specific opcode for this. Basically the same as #2, but
>>>> maybe with a more descriptive name than NOP.
>>>>
>>>> Might make sense to pair that with a CQE flag or something like that, as
>>>> there's no specific user_data that could be used as it doesn't match an
>>>> existing SQE that has been issued. IORING_CQE_F_WAKEUP for example.
>>>> Would be applicable to all the above cases.
>>>>
>>>> I kind of like #3 the best. Add a IORING_OP_RING_WAKEUP command, require
>>>> that sqe->fd point to a ring (could even be the ring itself, doesn't
>>>> matter). And add IORING_CQE_F_WAKEUP as a specific flag for that.
>>>
>>> Something like the below, totally untested. The request will complete on
>>> the original ring with either 0, for success, or -EOVERFLOW if the
>>> target ring was already in an overflow state. If the fd specified isn't
>>> an io_uring context, then the request will complete with -EBADFD.
>>>
>>> If you have any way of testing this, please do. I'll write a basic
>>> functionality test for it as well, but not until tomorrow.
>>>
>>> Maybe we want to include in cqe->res who the waker was? We can stuff the
>>> pid/tid in there, for example.
>>
>> Made the pid change, and also wrote a test case for it. Only change
>> otherwise is adding a completion trace event as well. Patch below
>> against for-5.18/io_uring, and attached the test case for liburing.
>>
>> diff --git a/fs/io_uring.c b/fs/io_uring.c
>> index 2e04f718319d..b21f85a48224 100644
>> --- a/fs/io_uring.c
>> +++ b/fs/io_uring.c
>> @@ -1105,6 +1105,9 @@ static const struct io_op_def io_op_defs[] = {
>> [IORING_OP_MKDIRAT] = {},
>> [IORING_OP_SYMLINKAT] = {},
>> [IORING_OP_LINKAT] = {},
>> + [IORING_OP_WAKEUP_RING] = {
>> + .needs_file = 1,
>> + },
>> };
>> /* requests with any of those set should undergo io_disarm_next() */
>> @@ -4235,6 +4238,44 @@ static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
>> return 0;
>> }
>> +static int io_wakeup_ring_prep(struct io_kiocb *req,
>> + const struct io_uring_sqe *sqe)
>> +{
>> + if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index || sqe->off ||
>> + sqe->len || sqe->rw_flags || sqe->splice_fd_in ||
>> + sqe->buf_index || sqe->personality))
>> + return -EINVAL;
>> +
>> + if (req->file->f_op != &io_uring_fops)
>> + return -EBADFD;
>> +
>> + return 0;
>> +}
>> +
>> +static int io_wakeup_ring(struct io_kiocb *req, unsigned int issue_flags)
>> +{
>> + struct io_uring_cqe *cqe;
>> + struct io_ring_ctx *ctx;
>> + int ret = 0;
>> +
>> + ctx = req->file->private_data;
>> + spin_lock(&ctx->completion_lock);
>> + cqe = io_get_cqe(ctx);
>> + if (cqe) {
>> + WRITE_ONCE(cqe->user_data, 0);
>> + WRITE_ONCE(cqe->res, 0);
>> + WRITE_ONCE(cqe->flags, IORING_CQE_F_WAKEUP);
>> + } else {
>> + ret = -EOVERFLOW;
>> + }
>
> io_fill_cqe_aux(), maybe? Handles overflows better, increments cq_extra,
> etc. Might also make sense to kick cq_timeouts, so waiters are forced to
> wake up.
I think the main question here is if we want to handle overflows at all,
I deliberately didn't do that. But apart from that io_fill_cqe_aux()
does to everything we need.
I guess the nice thing about actually allocating an overflow entry is
that there's no weird error handling on the submitter side. Let's go
with that.
--
Jens Axboe
