Message ID | 20240524230501.20178-3-axboe@kernel.dk (mailing list archive) |
---|---|
State | New |
Headers | show |
Series | Improve MSG_RING SINGLE_ISSUER performance | expand |
On 5/24/24 23:58, Jens Axboe wrote: > If IORING_SETUP_SINGLE_ISSUER is set, then we can't post CQEs remotely > to the target ring. Instead, task_work is queued for the target ring, > which is used to post the CQE. To make matters worse, once the target > CQE has been posted, task_work is then queued with the originator to > fill the completion. > > This obviously adds a bunch of overhead and latency. Instead of relying > on generic kernel task_work for this, fill an overflow entry on the > target ring and flag it as such that the target ring will flush it. This > avoids both the task_work for posting the CQE, and it means that the > originator CQE can be filled inline as well. > > In local testing, this reduces the latency on the sender side by 5-6x. > > Signed-off-by: Jens Axboe <axboe@kernel.dk> > --- > io_uring/msg_ring.c | 77 +++++++++++++++++++++++++++++++++++++++++++-- > 1 file changed, 74 insertions(+), 3 deletions(-) > > diff --git a/io_uring/msg_ring.c b/io_uring/msg_ring.c > index feff2b0822cf..3f89ff3a40ad 100644 > --- a/io_uring/msg_ring.c > +++ b/io_uring/msg_ring.c > @@ -123,6 +123,69 @@ static void io_msg_tw_complete(struct callback_head *head) > io_req_queue_tw_complete(req, ret); > } > > +static struct io_overflow_cqe *io_alloc_overflow(struct io_ring_ctx *target_ctx) > +{ > + bool is_cqe32 = target_ctx->flags & IORING_SETUP_CQE32; > + size_t cqe_size = sizeof(struct io_overflow_cqe); > + struct io_overflow_cqe *ocqe; > + > + if (is_cqe32) > + cqe_size += sizeof(struct io_uring_cqe); > + > + ocqe = kmalloc(cqe_size, GFP_ATOMIC | __GFP_ACCOUNT); > + if (!ocqe) > + return NULL; > + > + if (is_cqe32) > + ocqe->cqe.big_cqe[0] = ocqe->cqe.big_cqe[1] = 0; > + > + return ocqe; > +} > + > +/* > + * Entered with the target uring_lock held, and will drop it before > + * returning. Adds a previously allocated ocqe to the overflow list on > + * the target, and marks it appropriately for flushing. > + */ > +static void io_msg_add_overflow(struct io_msg *msg, > + struct io_ring_ctx *target_ctx, > + struct io_overflow_cqe *ocqe, int ret) > + __releases(target_ctx->uring_lock) > +{ > + spin_lock(&target_ctx->completion_lock); > + > + if (list_empty(&target_ctx->cq_overflow_list)) { > + set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &target_ctx->check_cq); > + atomic_or(IORING_SQ_TASKRUN, &target_ctx->rings->sq_flags); TASKRUN? The normal overflow path sets IORING_SQ_CQ_OVERFLOW > + } > + > + ocqe->cqe.user_data = msg->user_data; > + ocqe->cqe.res = ret; > + list_add_tail(&ocqe->list, &target_ctx->cq_overflow_list); > + spin_unlock(&target_ctx->completion_lock); > + mutex_unlock(&target_ctx->uring_lock); > + wake_up_state(target_ctx->submitter_task, TASK_INTERRUPTIBLE); > +}
On 5/24/24 23:58, Jens Axboe wrote: > If IORING_SETUP_SINGLE_ISSUER is set, then we can't post CQEs remotely > to the target ring. Instead, task_work is queued for the target ring, > which is used to post the CQE. To make matters worse, once the target > CQE has been posted, task_work is then queued with the originator to > fill the completion. > > This obviously adds a bunch of overhead and latency. Instead of relying > on generic kernel task_work for this, fill an overflow entry on the > target ring and flag it as such that the target ring will flush it. This > avoids both the task_work for posting the CQE, and it means that the > originator CQE can be filled inline as well. > > In local testing, this reduces the latency on the sender side by 5-6x. > > Signed-off-by: Jens Axboe <axboe@kernel.dk> > --- > io_uring/msg_ring.c | 77 +++++++++++++++++++++++++++++++++++++++++++-- > 1 file changed, 74 insertions(+), 3 deletions(-) > > diff --git a/io_uring/msg_ring.c b/io_uring/msg_ring.c > index feff2b0822cf..3f89ff3a40ad 100644 > --- a/io_uring/msg_ring.c > +++ b/io_uring/msg_ring.c > @@ -123,6 +123,69 @@ static void io_msg_tw_complete(struct callback_head *head) > io_req_queue_tw_complete(req, ret); > } > > +static struct io_overflow_cqe *io_alloc_overflow(struct io_ring_ctx *target_ctx) > +{ > + bool is_cqe32 = target_ctx->flags & IORING_SETUP_CQE32; > + size_t cqe_size = sizeof(struct io_overflow_cqe); > + struct io_overflow_cqe *ocqe; > + > + if (is_cqe32) > + cqe_size += sizeof(struct io_uring_cqe); > + > + ocqe = kmalloc(cqe_size, GFP_ATOMIC | __GFP_ACCOUNT); __GFP_ACCOUNT looks painful > + if (!ocqe) > + return NULL; > + > + if (is_cqe32) > + ocqe->cqe.big_cqe[0] = ocqe->cqe.big_cqe[1] = 0; > + > + return ocqe; > +} > + ...
On 5/28/24 7:32 AM, Pavel Begunkov wrote: > On 5/24/24 23:58, Jens Axboe wrote: >> If IORING_SETUP_SINGLE_ISSUER is set, then we can't post CQEs remotely >> to the target ring. Instead, task_work is queued for the target ring, >> which is used to post the CQE. To make matters worse, once the target >> CQE has been posted, task_work is then queued with the originator to >> fill the completion. >> >> This obviously adds a bunch of overhead and latency. Instead of relying >> on generic kernel task_work for this, fill an overflow entry on the >> target ring and flag it as such that the target ring will flush it. This >> avoids both the task_work for posting the CQE, and it means that the >> originator CQE can be filled inline as well. >> >> In local testing, this reduces the latency on the sender side by 5-6x. >> >> Signed-off-by: Jens Axboe <axboe@kernel.dk> >> --- >> io_uring/msg_ring.c | 77 +++++++++++++++++++++++++++++++++++++++++++-- >> 1 file changed, 74 insertions(+), 3 deletions(-) >> >> diff --git a/io_uring/msg_ring.c b/io_uring/msg_ring.c >> index feff2b0822cf..3f89ff3a40ad 100644 >> --- a/io_uring/msg_ring.c >> +++ b/io_uring/msg_ring.c >> @@ -123,6 +123,69 @@ static void io_msg_tw_complete(struct callback_head *head) >> io_req_queue_tw_complete(req, ret); >> } >> +static struct io_overflow_cqe *io_alloc_overflow(struct io_ring_ctx *target_ctx) >> +{ >> + bool is_cqe32 = target_ctx->flags & IORING_SETUP_CQE32; >> + size_t cqe_size = sizeof(struct io_overflow_cqe); >> + struct io_overflow_cqe *ocqe; >> + >> + if (is_cqe32) >> + cqe_size += sizeof(struct io_uring_cqe); >> + >> + ocqe = kmalloc(cqe_size, GFP_ATOMIC | __GFP_ACCOUNT); > > __GFP_ACCOUNT looks painful It always is - I did add the usual alloc cache for this after posting this series, which makes it a no-op basically: https://git.kernel.dk/cgit/linux/commit/?h=io_uring-msg_ring&id=c39ead262b60872d6d7daf55e9fc7d76dc09b29d Just haven't posted a v2 yet.
On 5/28/24 7:18 AM, Pavel Begunkov wrote: > On 5/24/24 23:58, Jens Axboe wrote: >> If IORING_SETUP_SINGLE_ISSUER is set, then we can't post CQEs remotely >> to the target ring. Instead, task_work is queued for the target ring, >> which is used to post the CQE. To make matters worse, once the target >> CQE has been posted, task_work is then queued with the originator to >> fill the completion. >> >> This obviously adds a bunch of overhead and latency. Instead of relying >> on generic kernel task_work for this, fill an overflow entry on the >> target ring and flag it as such that the target ring will flush it. This >> avoids both the task_work for posting the CQE, and it means that the >> originator CQE can be filled inline as well. >> >> In local testing, this reduces the latency on the sender side by 5-6x. >> >> Signed-off-by: Jens Axboe <axboe@kernel.dk> >> --- >> io_uring/msg_ring.c | 77 +++++++++++++++++++++++++++++++++++++++++++-- >> 1 file changed, 74 insertions(+), 3 deletions(-) >> >> diff --git a/io_uring/msg_ring.c b/io_uring/msg_ring.c >> index feff2b0822cf..3f89ff3a40ad 100644 >> --- a/io_uring/msg_ring.c >> +++ b/io_uring/msg_ring.c >> @@ -123,6 +123,69 @@ static void io_msg_tw_complete(struct callback_head *head) >> io_req_queue_tw_complete(req, ret); >> } >> +static struct io_overflow_cqe *io_alloc_overflow(struct io_ring_ctx *target_ctx) >> +{ >> + bool is_cqe32 = target_ctx->flags & IORING_SETUP_CQE32; >> + size_t cqe_size = sizeof(struct io_overflow_cqe); >> + struct io_overflow_cqe *ocqe; >> + >> + if (is_cqe32) >> + cqe_size += sizeof(struct io_uring_cqe); >> + >> + ocqe = kmalloc(cqe_size, GFP_ATOMIC | __GFP_ACCOUNT); >> + if (!ocqe) >> + return NULL; >> + >> + if (is_cqe32) >> + ocqe->cqe.big_cqe[0] = ocqe->cqe.big_cqe[1] = 0; >> + >> + return ocqe; >> +} >> + >> +/* >> + * Entered with the target uring_lock held, and will drop it before >> + * returning. Adds a previously allocated ocqe to the overflow list on >> + * the target, and marks it appropriately for flushing. >> + */ >> +static void io_msg_add_overflow(struct io_msg *msg, >> + struct io_ring_ctx *target_ctx, >> + struct io_overflow_cqe *ocqe, int ret) >> + __releases(target_ctx->uring_lock) >> +{ >> + spin_lock(&target_ctx->completion_lock); >> + >> + if (list_empty(&target_ctx->cq_overflow_list)) { >> + set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &target_ctx->check_cq); >> + atomic_or(IORING_SQ_TASKRUN, &target_ctx->rings->sq_flags); > > TASKRUN? The normal overflow path sets IORING_SQ_CQ_OVERFLOW Was a bit split on it - we want it run as part of waiting, but I also wasn't super interested in exposing it as an overflow condition since it is now. It's more of an internal implementation detail.
On 5/28/24 15:23, Jens Axboe wrote: > On 5/28/24 7:32 AM, Pavel Begunkov wrote: >> On 5/24/24 23:58, Jens Axboe wrote: >>> If IORING_SETUP_SINGLE_ISSUER is set, then we can't post CQEs remotely >>> to the target ring. Instead, task_work is queued for the target ring, >>> which is used to post the CQE. To make matters worse, once the target >>> CQE has been posted, task_work is then queued with the originator to >>> fill the completion. >>> >>> This obviously adds a bunch of overhead and latency. Instead of relying >>> on generic kernel task_work for this, fill an overflow entry on the >>> target ring and flag it as such that the target ring will flush it. This >>> avoids both the task_work for posting the CQE, and it means that the >>> originator CQE can be filled inline as well. >>> >>> In local testing, this reduces the latency on the sender side by 5-6x. >>> >>> Signed-off-by: Jens Axboe <axboe@kernel.dk> >>> --- >>> io_uring/msg_ring.c | 77 +++++++++++++++++++++++++++++++++++++++++++-- >>> 1 file changed, 74 insertions(+), 3 deletions(-) >>> >>> diff --git a/io_uring/msg_ring.c b/io_uring/msg_ring.c >>> index feff2b0822cf..3f89ff3a40ad 100644 >>> --- a/io_uring/msg_ring.c >>> +++ b/io_uring/msg_ring.c >>> @@ -123,6 +123,69 @@ static void io_msg_tw_complete(struct callback_head *head) >>> io_req_queue_tw_complete(req, ret); >>> } >>> +static struct io_overflow_cqe *io_alloc_overflow(struct io_ring_ctx *target_ctx) >>> +{ >>> + bool is_cqe32 = target_ctx->flags & IORING_SETUP_CQE32; >>> + size_t cqe_size = sizeof(struct io_overflow_cqe); >>> + struct io_overflow_cqe *ocqe; >>> + >>> + if (is_cqe32) >>> + cqe_size += sizeof(struct io_uring_cqe); >>> + >>> + ocqe = kmalloc(cqe_size, GFP_ATOMIC | __GFP_ACCOUNT); >> >> __GFP_ACCOUNT looks painful > > It always is - I did add the usual alloc cache for this after posting > this series, which makes it a no-op basically: Simple ring private cache wouldn't work so well with non uniform transfer distributions. One way messaging, userspace level batching, etc., but the main question is in the other email, i.e. maybe it's better to go with the 2 tw hop model, which returns memory back where it came from. > https://git.kernel.dk/cgit/linux/commit/?h=io_uring-msg_ring&id=c39ead262b60872d6d7daf55e9fc7d76dc09b29d > > Just haven't posted a v2 yet. >
On 5/28/24 10:23 AM, Pavel Begunkov wrote: > On 5/28/24 15:23, Jens Axboe wrote: >> On 5/28/24 7:32 AM, Pavel Begunkov wrote: >>> On 5/24/24 23:58, Jens Axboe wrote: >>>> If IORING_SETUP_SINGLE_ISSUER is set, then we can't post CQEs remotely >>>> to the target ring. Instead, task_work is queued for the target ring, >>>> which is used to post the CQE. To make matters worse, once the target >>>> CQE has been posted, task_work is then queued with the originator to >>>> fill the completion. >>>> >>>> This obviously adds a bunch of overhead and latency. Instead of relying >>>> on generic kernel task_work for this, fill an overflow entry on the >>>> target ring and flag it as such that the target ring will flush it. This >>>> avoids both the task_work for posting the CQE, and it means that the >>>> originator CQE can be filled inline as well. >>>> >>>> In local testing, this reduces the latency on the sender side by 5-6x. >>>> >>>> Signed-off-by: Jens Axboe <axboe@kernel.dk> >>>> --- >>>> io_uring/msg_ring.c | 77 +++++++++++++++++++++++++++++++++++++++++++-- >>>> 1 file changed, 74 insertions(+), 3 deletions(-) >>>> >>>> diff --git a/io_uring/msg_ring.c b/io_uring/msg_ring.c >>>> index feff2b0822cf..3f89ff3a40ad 100644 >>>> --- a/io_uring/msg_ring.c >>>> +++ b/io_uring/msg_ring.c >>>> @@ -123,6 +123,69 @@ static void io_msg_tw_complete(struct callback_head *head) >>>> io_req_queue_tw_complete(req, ret); >>>> } >>>> +static struct io_overflow_cqe *io_alloc_overflow(struct io_ring_ctx *target_ctx) >>>> +{ >>>> + bool is_cqe32 = target_ctx->flags & IORING_SETUP_CQE32; >>>> + size_t cqe_size = sizeof(struct io_overflow_cqe); >>>> + struct io_overflow_cqe *ocqe; >>>> + >>>> + if (is_cqe32) >>>> + cqe_size += sizeof(struct io_uring_cqe); >>>> + >>>> + ocqe = kmalloc(cqe_size, GFP_ATOMIC | __GFP_ACCOUNT); >>> >>> __GFP_ACCOUNT looks painful >> >> It always is - I did add the usual alloc cache for this after posting >> this series, which makes it a no-op basically: > > Simple ring private cache wouldn't work so well with non > uniform transfer distributions. One way messaging, userspace > level batching, etc., but the main question is in the other > email, i.e. maybe it's better to go with the 2 tw hop model, > which returns memory back where it came from. The cache is local to the ring, so anyone that sends messages to that ring gets to use it. So I believe it should in fact work really well. If messaging is bidirectional, then caching on the target will apply in both directions.
On 5/28/24 18:59, Jens Axboe wrote: > On 5/28/24 10:23 AM, Pavel Begunkov wrote: >> On 5/28/24 15:23, Jens Axboe wrote: >>> On 5/28/24 7:32 AM, Pavel Begunkov wrote: >>>> On 5/24/24 23:58, Jens Axboe wrote: >>>>> If IORING_SETUP_SINGLE_ISSUER is set, then we can't post CQEs remotely >>>>> to the target ring. Instead, task_work is queued for the target ring, >>>>> which is used to post the CQE. To make matters worse, once the target >>>>> CQE has been posted, task_work is then queued with the originator to >>>>> fill the completion. >>>>> >>>>> This obviously adds a bunch of overhead and latency. Instead of relying >>>>> on generic kernel task_work for this, fill an overflow entry on the >>>>> target ring and flag it as such that the target ring will flush it. This >>>>> avoids both the task_work for posting the CQE, and it means that the >>>>> originator CQE can be filled inline as well. >>>>> >>>>> In local testing, this reduces the latency on the sender side by 5-6x. >>>>> >>>>> Signed-off-by: Jens Axboe <axboe@kernel.dk> >>>>> --- >>>>> io_uring/msg_ring.c | 77 +++++++++++++++++++++++++++++++++++++++++++-- >>>>> 1 file changed, 74 insertions(+), 3 deletions(-) >>>>> >>>>> diff --git a/io_uring/msg_ring.c b/io_uring/msg_ring.c >>>>> index feff2b0822cf..3f89ff3a40ad 100644 >>>>> --- a/io_uring/msg_ring.c >>>>> +++ b/io_uring/msg_ring.c >>>>> @@ -123,6 +123,69 @@ static void io_msg_tw_complete(struct callback_head *head) >>>>> io_req_queue_tw_complete(req, ret); >>>>> } >>>>> +static struct io_overflow_cqe *io_alloc_overflow(struct io_ring_ctx *target_ctx) >>>>> +{ >>>>> + bool is_cqe32 = target_ctx->flags & IORING_SETUP_CQE32; >>>>> + size_t cqe_size = sizeof(struct io_overflow_cqe); >>>>> + struct io_overflow_cqe *ocqe; >>>>> + >>>>> + if (is_cqe32) >>>>> + cqe_size += sizeof(struct io_uring_cqe); >>>>> + >>>>> + ocqe = kmalloc(cqe_size, GFP_ATOMIC | __GFP_ACCOUNT); >>>> >>>> __GFP_ACCOUNT looks painful >>> >>> It always is - I did add the usual alloc cache for this after posting >>> this series, which makes it a no-op basically: >> >> Simple ring private cache wouldn't work so well with non >> uniform transfer distributions. One way messaging, userspace >> level batching, etc., but the main question is in the other >> email, i.e. maybe it's better to go with the 2 tw hop model, >> which returns memory back where it came from. > > The cache is local to the ring, so anyone that sends messages to that > ring gets to use it. So I believe it should in fact work really well. If > messaging is bidirectional, then caching on the target will apply in > both directions. *taking a look at the patch* it gets the entry from the target's ring, so indeed not a problem. Taking the target lock for that, however, is not the best, I ranted before about inter dependencies b/w rings. E.g. requests messaging a ring run by a task CPU bound in submission / tw execution would be directed to iowq and occupy a worker thread for the time being.
On 5/28/24 8:04 PM, Pavel Begunkov wrote: > On 5/28/24 18:59, Jens Axboe wrote: >> On 5/28/24 10:23 AM, Pavel Begunkov wrote: >>> On 5/28/24 15:23, Jens Axboe wrote: >>>> On 5/28/24 7:32 AM, Pavel Begunkov wrote: >>>>> On 5/24/24 23:58, Jens Axboe wrote: >>>>>> If IORING_SETUP_SINGLE_ISSUER is set, then we can't post CQEs remotely >>>>>> to the target ring. Instead, task_work is queued for the target ring, >>>>>> which is used to post the CQE. To make matters worse, once the target >>>>>> CQE has been posted, task_work is then queued with the originator to >>>>>> fill the completion. >>>>>> >>>>>> This obviously adds a bunch of overhead and latency. Instead of relying >>>>>> on generic kernel task_work for this, fill an overflow entry on the >>>>>> target ring and flag it as such that the target ring will flush it. This >>>>>> avoids both the task_work for posting the CQE, and it means that the >>>>>> originator CQE can be filled inline as well. >>>>>> >>>>>> In local testing, this reduces the latency on the sender side by 5-6x. >>>>>> >>>>>> Signed-off-by: Jens Axboe <axboe@kernel.dk> >>>>>> --- >>>>>> io_uring/msg_ring.c | 77 +++++++++++++++++++++++++++++++++++++++++++-- >>>>>> 1 file changed, 74 insertions(+), 3 deletions(-) >>>>>> >>>>>> diff --git a/io_uring/msg_ring.c b/io_uring/msg_ring.c >>>>>> index feff2b0822cf..3f89ff3a40ad 100644 >>>>>> --- a/io_uring/msg_ring.c >>>>>> +++ b/io_uring/msg_ring.c >>>>>> @@ -123,6 +123,69 @@ static void io_msg_tw_complete(struct callback_head *head) >>>>>> io_req_queue_tw_complete(req, ret); >>>>>> } >>>>>> +static struct io_overflow_cqe *io_alloc_overflow(struct io_ring_ctx *target_ctx) >>>>>> +{ >>>>>> + bool is_cqe32 = target_ctx->flags & IORING_SETUP_CQE32; >>>>>> + size_t cqe_size = sizeof(struct io_overflow_cqe); >>>>>> + struct io_overflow_cqe *ocqe; >>>>>> + >>>>>> + if (is_cqe32) >>>>>> + cqe_size += sizeof(struct io_uring_cqe); >>>>>> + >>>>>> + ocqe = kmalloc(cqe_size, GFP_ATOMIC | __GFP_ACCOUNT); >>>>> >>>>> __GFP_ACCOUNT looks painful >>>> >>>> It always is - I did add the usual alloc cache for this after posting >>>> this series, which makes it a no-op basically: >>> >>> Simple ring private cache wouldn't work so well with non >>> uniform transfer distributions. One way messaging, userspace >>> level batching, etc., but the main question is in the other >>> email, i.e. maybe it's better to go with the 2 tw hop model, >>> which returns memory back where it came from. >> >> The cache is local to the ring, so anyone that sends messages to that >> ring gets to use it. So I believe it should in fact work really well. If >> messaging is bidirectional, then caching on the target will apply in >> both directions. > > *taking a look at the patch* it gets the entry from the target's > ring, so indeed not a problem. Taking the target lock for that, > however, is not the best, I ranted before about inter dependencies > b/w rings. E.g. requests messaging a ring run by a task CPU bound > in submission / tw execution would be directed to iowq and occupy > a worker thread for the time being. I can try and do some stats on io-wq bouncing, it can indeed be a risk. Might even be possible to only retain the ring lock for flushing, which is less of an issue as it happens locally, and have the overflow entries locked separately. For now I just kept the overflow backend that we already have, and the locking that MSG_RING already does.
diff --git a/io_uring/msg_ring.c b/io_uring/msg_ring.c index feff2b0822cf..3f89ff3a40ad 100644 --- a/io_uring/msg_ring.c +++ b/io_uring/msg_ring.c @@ -123,6 +123,69 @@ static void io_msg_tw_complete(struct callback_head *head) io_req_queue_tw_complete(req, ret); } +static struct io_overflow_cqe *io_alloc_overflow(struct io_ring_ctx *target_ctx) +{ + bool is_cqe32 = target_ctx->flags & IORING_SETUP_CQE32; + size_t cqe_size = sizeof(struct io_overflow_cqe); + struct io_overflow_cqe *ocqe; + + if (is_cqe32) + cqe_size += sizeof(struct io_uring_cqe); + + ocqe = kmalloc(cqe_size, GFP_ATOMIC | __GFP_ACCOUNT); + if (!ocqe) + return NULL; + + if (is_cqe32) + ocqe->cqe.big_cqe[0] = ocqe->cqe.big_cqe[1] = 0; + + return ocqe; +} + +/* + * Entered with the target uring_lock held, and will drop it before + * returning. Adds a previously allocated ocqe to the overflow list on + * the target, and marks it appropriately for flushing. + */ +static void io_msg_add_overflow(struct io_msg *msg, + struct io_ring_ctx *target_ctx, + struct io_overflow_cqe *ocqe, int ret) + __releases(target_ctx->uring_lock) +{ + spin_lock(&target_ctx->completion_lock); + + if (list_empty(&target_ctx->cq_overflow_list)) { + set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &target_ctx->check_cq); + atomic_or(IORING_SQ_TASKRUN, &target_ctx->rings->sq_flags); + } + + ocqe->cqe.user_data = msg->user_data; + ocqe->cqe.res = ret; + list_add_tail(&ocqe->list, &target_ctx->cq_overflow_list); + spin_unlock(&target_ctx->completion_lock); + mutex_unlock(&target_ctx->uring_lock); + wake_up_state(target_ctx->submitter_task, TASK_INTERRUPTIBLE); +} + +static bool io_msg_fill_remote(struct io_msg *msg, unsigned int issue_flags, + struct io_ring_ctx *target_ctx, u32 flags) +{ + struct io_overflow_cqe *ocqe; + + ocqe = io_alloc_overflow(target_ctx); + if (!ocqe) + return false; + + if (unlikely(io_double_lock_ctx(target_ctx, issue_flags))) { + kfree(ocqe); + return false; + } + + ocqe->cqe.flags = flags; + io_msg_add_overflow(msg, target_ctx, ocqe, msg->len); + return true; +} + static int io_msg_ring_data(struct io_kiocb *req, unsigned int issue_flags) { struct io_ring_ctx *target_ctx = req->file->private_data; @@ -137,12 +200,20 @@ static int io_msg_ring_data(struct io_kiocb *req, unsigned int issue_flags) if (target_ctx->flags & IORING_SETUP_R_DISABLED) return -EBADFD; - if (io_msg_need_remote(target_ctx)) - return io_msg_exec_remote(req, io_msg_tw_complete); - if (msg->flags & IORING_MSG_RING_FLAGS_PASS) flags = msg->cqe_flags; + if (io_msg_need_remote(target_ctx)) { + /* + * Try adding an overflow entry to the target, and only if + * that fails, resort to using more expensive task_work to + * have the target_ctx owner fill the CQE. + */ + if (!io_msg_fill_remote(msg, issue_flags, target_ctx, flags)) + return io_msg_exec_remote(req, io_msg_tw_complete); + return 0; + } + ret = -EOVERFLOW; if (target_ctx->flags & IORING_SETUP_IOPOLL) { if (unlikely(io_double_lock_ctx(target_ctx, issue_flags)))
If IORING_SETUP_SINGLE_ISSUER is set, then we can't post CQEs remotely to the target ring. Instead, task_work is queued for the target ring, which is used to post the CQE. To make matters worse, once the target CQE has been posted, task_work is then queued with the originator to fill the completion. This obviously adds a bunch of overhead and latency. Instead of relying on generic kernel task_work for this, fill an overflow entry on the target ring and flag it as such that the target ring will flush it. This avoids both the task_work for posting the CQE, and it means that the originator CQE can be filled inline as well. In local testing, this reduces the latency on the sender side by 5-6x. Signed-off-by: Jens Axboe <axboe@kernel.dk> --- io_uring/msg_ring.c | 77 +++++++++++++++++++++++++++++++++++++++++++-- 1 file changed, 74 insertions(+), 3 deletions(-)