Use io_wq_work_list for task_work
Jens Axboe
maintainercover Use io_wq_work_list for task_work
Hi,
This converts the deferred, normal, and fallback task_work to use a normal io_wq_work_list, rather than an llist.
The main motivation behind this is to get rid of the need to reverse the list once it's deleted and run. I tested this basic conversion of just switching it from an llist to an io_wq_work_list with a spinlock, and I don't see any benefits from the lockless list. And for cases where we get a bursty addition of task_work, this approach is faster as it avoids the need to iterate the list upfront while reversing it.
And this is less code and simpler, so I'd prefer to go that route.
include/linux/io_uring_types.h | 13 +-- io_uring/io_uring.c | 175 ++++++++++++++++----------------- io_uring/io_uring.h | 8 +- io_uring/sqpoll.c | 8 +- io_uring/tctx.c | 3 +- 5 files changed, 102 insertions(+), 105 deletions(-)
Signature · 1 line
Jens Axboe
Jens Axboe
maintainerPATCH 1/4 io_uring: use the right type for work_llist empty check
io_task_work_pending() uses wq_list_empty() on ctx->work_llist, but it's not an io_wq_work_list, it's a struct llist_head. They both have ->first as head-of-list, and it turns out the checks are identical. But be proper and use the right helper.
--- io_uring/io_uring.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
io_uring/io_uring.h +1 -1
@@ -343,7 +343,7 @@ static inline int io_run_task_work(void)static inline bool io_task_work_pending (struct io_ring_ctx *ctx) { -return task_work_pending (current) || !wq_list_empty (&ctx->work_llist); +return task_work_pending (current) || !llist_empty (&ctx->work_llist); }static inline void io_tw_lock (struct io_ring_ctx *ctx,struct io_tw_state *ts)
Signature · 1 line
2.43.0
Jens Axboe
maintainerPATCH 2/4 io_uring: switch deferred task_work to an io_wq_work_list
In reply to Jens AxboeLockless lists may be handy for some things, but they mean that items are in the reverse order as we can only add to the head of the list. That in turn means that iterating items on the list needs to reverse it first, if it's sensitive to ordering between items on the list.
Switch the DEFER_TASKRUN work list from an llist to a normal io_wq_work_list, and protect it with a lock. Then we can get rid of the manual reversing of the list when running it, which takes considerable cycles particularly for bursty task_work additions.
--- include/linux/io_uring_types.h | 11 ++-- io_uring/io_uring.c | 117 ++++++++++++--------------------- io_uring/io_uring.h | 4 +- 3 files changed, 51 insertions(+), 81 deletions(-)
include/linux/io_uring_types.h +7 -4
@@ -329,7 +329,9 @@ struct io_ring_ctx {* regularly bounce b/w CPUs. */ struct { -struct llist_head work_llist; +struct io_wq_work_list work_list; +spinlock_t work_lock; +int work_items;unsigned long check_cq;atomic_t cq_wait_nr;atomic_t cq_timeouts;
@@ -559,7 +561,10 @@ enum {typedef void (*io_req_tw_func_t )(struct io_kiocb *req,struct io_tw_state *ts);struct io_task_work { -struct llist_node node; +union { +struct io_wq_work_node node; +struct llist_node llist_node; + };io_req_tw_func_t func; };
@@ -615,8 +620,6 @@ struct io_kiocb {*/ u16 buf_index; -unsigned nr_tw; -/* REQ_F_* flags */ io_req_flags_t flags;
io_uring/io_uring.c +42 -75
250 lines not on this page · raw message
io_uring/io_uring.h +2 -2
@@ -343,7 +343,7 @@ static inline int io_run_task_work(void)static inline bool io_task_work_pending (struct io_ring_ctx *ctx) { -return task_work_pending (current) || !llist_empty (&ctx->work_llist); +return task_work_pending (current) || !wq_list_empty (&ctx->work_list); }static inline void io_tw_lock (struct io_ring_ctx *ctx,struct io_tw_state *ts)
@@ -457,6 +457,6 @@ enum {static inline bool io_has_work (struct io_ring_ctx *ctx) {return test_bit (IO_CHECK_CQ_OVERFLOW_BIT , &ctx->check_cq) || - !llist_empty (&ctx->work_llist); + !wq_list_empty (&ctx->work_list); }#endif
Signature · 1 line
2.43.0
Re: [PATCH 2/4] io_uring: switch deferred task_work to an io_wq_work_list
Quoted 23 lines · Jens Axboe
Lockless lists may be handy for some things, but they mean that items are in the reverse order as we can only add to the head of the list. That in turn means that iterating items on the list needs to reverse it first, if it's sensitive to ordering between items on the list.
Switch the
DEFER_TASKRUNwork list from an llist to a normalio_wq_work_list, and protect it with a lock. Then we can get rid of the manual reversing of the list when running it, which takes considerable cycles particularly for burstytask_workadditions.Signed-off-by Jens Axboe <[email protected]>--- include/linux/io_uring_types.h | 11 ++-- io_uring/io_uring.c | 117 ++++++++++++--------------------- io_uring/io_uring.h | 4 +- 3 files changed, 51 insertions(+), 81 deletions(-)include/linux/io_uring_types.h +0
-0@@ -329,7 +329,9 @@ struct io_ring_ctx {* regularly bounce b/w CPUs.
...
Quoted 31 lines
include/linux/io_uring_types.h +4
-17-static inline void io_req_local_work_add (struct io_kiocb *req,unsigned flags) +static inline void io_req_local_work_add (struct io_kiocb *req,unsigned tw_flags) {struct io_ring_ctx *ctx = req->ctx; -unsigned nr_wait, nr_tw, nr_tw_prev; -struct llist_node *head; +unsigned nr_wait, nr_tw; +unsigned long flags;/* See comment above IO_CQ_WAKE_INIT */ BUILD_BUG_ON (IO_CQ_WAKE_FORCE <=IORING_MAX_CQ_ENTRIES );/* -* We don't know how many reuqests is there in the link and whether +* We don't know how many requests is there in the link and whether * they can even be queued lazily, fall back to non-lazy. */ if (req->flags & (REQ_F_LINK |REQ_F_HARDLINK )) - flags &= ~IOU_F_TWQ_LAZY_WAKE ; - - head =READ_ONCE (ctx->work_llist.first); - do { - nr_tw_prev = 0; - if (head) { - struct io_kiocb *first_req = container_of(head, - struct io_kiocb, - io_task_work.node); - /* - * Might be executed at any moment, rely on - * SLAB_TYPESAFE_BY_RCU to keep it alive. - */
include/linux/io_uring_types.h
Quoted 10 lines
include/linux/io_uring_types.h +0
-10- nr_tw_prev =READ_ONCE (first_req->nr_tw); - } - -/* -* Theoretically, it can overflow, but that's fine as one of -* previous adds should've tried to wake the task. -*/ - nr_tw = nr_tw_prev +1 ; -if (!(flags &IOU_F_TWQ_LAZY_WAKE )) - nr_tw =IO_CQ_WAKE_FORCE ;
Aren't you just killing the entire IOU_F_TWQ_LAZY_WAKE handling? It's assigned to IO_CQ_WAKE_FORCE so that it passes the check before wake_up below.
include/linux/io_uring_types.h
Quoted 10 lines
include/linux/io_uring_types.h +5
-4+ tw_flags &= ~IOU_F_TWQ_LAZY_WAKE ; - req->nr_tw = nr_tw; - req->io_task_work.node.next = head; - }while (!try_cmpxchg (&ctx->work_llist.first, &head, - &req->io_task_work.node)); +spin_lock_irqsave (&ctx->work_lock, flags); +wq_list_add_tail (&req->io_task_work.node, &ctx->work_list); + nr_tw = ++ctx->work_items; +spin_unlock_irqrestore (&ctx->work_lock, flags);
smp_mb(), see the comment below, and fwiw "_after_atomic" would not work.
Quoted 14 lines
/*
* cmpxchg implies a full barrier, which pairs with the barrierinclude/linux/io_uring_types.h +1
-1@@ -1289,7 +1254,7 @@ static inline void io_req_local_work_add(struct io_kiocb *req, unsigned flags)* is similar to the wait/wawke task state sync. */ -if (!head) { +if (nr_tw ==1 ) {if (ctx->flags &IORING_SETUP_TASKRUN_FLAG )atomic_or (IORING_SQ_TASKRUN , &ctx->rings->sq_flags);if (ctx->has_evfd)@@ -1297,13 +1262,8 @@ static inline void io_req_local_work_add(struct io_kiocb *req, unsigned flags) }
include/linux/io_uring_types.h static inline void io_req_local_work_add(struct io_kiocb *req, unsigned flags) · line 1266
Quoted 10 lines
include/linux/io_uring_types.h +2
-7nr_wait =atomic_read (&ctx->cq_wait_nr); -/* not enough or no one is waiting */ -if (nr_tw < nr_wait) -return ; -/* the previous add has already woken it up */ -if (nr_tw_prev >= nr_wait) -return ; -wake_up_state (ctx->submitter_task,TASK_INTERRUPTIBLE ); +if (nr_tw >= nr_wait) +wake_up_state (ctx->submitter_task,TASK_INTERRUPTIBLE );
IIUC, you're removing a very important optimisation, and I don't understand why'd you do that. It was waking up only when it's changing from "don't need to wake" to "have to be woken up", just once per splicing the list on the waiting side.
It seems like this one will keep pounding with wake_up_state for every single tw queued after @nr_wait, which quite often just 1.
Signature · 1 line
Pavel Begunkov
Jens Axboe
maintainerRe: [PATCH 2/4] io_uring: switch deferred task_work to an io_wq_work_list
Quoted 62 lines · Pavel Begunkov
Quoted 23 lines · Jens Axboe
Lockless lists may be handy for some things, but they mean that items are in the reverse order as we can only add to the head of the list. That in turn means that iterating items on the list needs to reverse it first, if it's sensitive to ordering between items on the list.
Switch the
DEFER_TASKRUNwork list from an llist to a normalio_wq_work_list, and protect it with a lock. Then we can get rid of the manual reversing of the list when running it, which takes considerable cycles particularly for burstytask_workadditions.Signed-off-by Jens Axboe <[email protected]>--- include/linux/io_uring_types.h | 11 ++-- io_uring/io_uring.c | 117 ++++++++++++--------------------- io_uring/io_uring.h | 4 +- 3 files changed, 51 insertions(+), 81 deletions(-)include/linux/io_uring_types.h +0
-0@@ -329,7 +329,9 @@ struct io_ring_ctx {* regularly bounce b/w CPUs. ...
Quoted 35 lines
include/linux/io_uring_types.h +4
-23-static inline void io_req_local_work_add (struct io_kiocb *req,unsigned flags) +static inline void io_req_local_work_add (struct io_kiocb *req,unsigned tw_flags) {struct io_ring_ctx *ctx = req->ctx; -unsigned nr_wait, nr_tw, nr_tw_prev; -struct llist_node *head; +unsigned nr_wait, nr_tw; +unsigned long flags;/* See comment above IO_CQ_WAKE_INIT */ BUILD_BUG_ON (IO_CQ_WAKE_FORCE <=IORING_MAX_CQ_ENTRIES );/* -* We don't know how many reuqests is there in the link and whether +* We don't know how many requests is there in the link and whether * they can even be queued lazily, fall back to non-lazy. */ if (req->flags & (REQ_F_LINK |REQ_F_HARDLINK )) - flags &= ~IOU_F_TWQ_LAZY_WAKE ; - - head =READ_ONCE (ctx->work_llist.first); - do { - nr_tw_prev = 0; - if (head) { - struct io_kiocb *first_req = container_of(head, - struct io_kiocb, - io_task_work.node); - /* - * Might be executed at any moment, rely on - * SLAB_TYPESAFE_BY_RCU to keep it alive. - */ - nr_tw_prev = READ_ONCE(first_req->nr_tw); - } - - /* - * Theoretically, it can overflow, but that's fine as one of - * previous adds should've tried to wake the task.
Quoted 8 lines
include/linux/io_uring_types.h
Quoted 4 lines
include/linux/io_uring_types.h +0
-4-*/ - nr_tw = nr_tw_prev +1 ; -if (!(flags &IOU_F_TWQ_LAZY_WAKE )) - nr_tw =IO_CQ_WAKE_FORCE ;Aren't you just killing the entire
IOU_F_TWQ_LAZY_WAKEhandling? It's assigned toIO_CQ_WAKE_FORCEso that it passes the check beforewake_upbelow.
Yeah I messed that one up, did fix that one yesterday before sending it out.
Quoted 10 lines
include/linux/io_uring_types.h
Quoted 9 lines
include/linux/io_uring_types.h +5
-3+ tw_flags &= ~IOU_F_TWQ_LAZY_WAKE ; - req->nr_tw = nr_tw; - req->io_task_work.node.next = head; - }while (!try_cmpxchg (&ctx->work_llist.first, &head, - &req->io_task_work.node)); +spin_lock_irqsave (&ctx->work_lock, flags); +wq_list_add_tail (&req->io_task_work.node, &ctx->work_list); + nr_tw = ++ctx->work_items; +spin_unlock_irqrestore (&ctx->work_lock, flags);
Quoted 2 lines
smp_mb(), see the comment below, and fwiw "_after_atomic" would not work.
For this one, I think all we need to do is have the wq_list_empty() check be fully stable. If we read:
nr_wait = atomic_read(&ctx->cq_wait_nr);
right before a waiter does:
atomic_set(&ctx->cq_wait_nr, foo);set_current_state(TASK_INTERRUPTIBLE);
then we need to ensure that the "I have work" check in io_cqring_wait_schedule() sees the work. The spin_unlock() has release semantics, and the current READ_ONCE() for work check sbould be enough, no?
Quoted 28 lines
Quoted 12 lines
/*
* cmpxchg implies a full barrier, which pairs with the barrierinclude/linux/io_uring_types.h +1
-0@@ -1289,7 +1254,7 @@ static inline void io_req_local_work_add(struct io_kiocb *req, unsigned flags)* is similar to the wait/wawke task state sync. */ - if (!head) { +if (nr_tw ==1 ) {if (ctx->flags &IORING_SETUP_TASKRUN_FLAG )atomic_or (IORING_SQ_TASKRUN , &ctx->rings->sq_flags);if (ctx->has_evfd)@@ -1297,13 +1262,8 @@ static inline void io_req_local_work_add(struct io_kiocb *req, unsigned flags) }include/linux/io_uring_types.h static inline void io_req_local_work_add(struct io_kiocb *req, unsigned flags) · line 1265
Quoted 10 lines
include/linux/io_uring_types.h +2
-7nr_wait =atomic_read (&ctx->cq_wait_nr); -/* not enough or no one is waiting */ -if (nr_tw < nr_wait) -return ; -/* the previous add has already woken it up */ -if (nr_tw_prev >= nr_wait) -return ; -wake_up_state (ctx->submitter_task,TASK_INTERRUPTIBLE ); +if (nr_tw >= nr_wait) +wake_up_state (ctx->submitter_task,TASK_INTERRUPTIBLE );IIUC, you're removing a very important optimisation, and I don't understand why'd you do that. It was waking up only when it's changing from "don't need to wake" to "have to be woken up", just once per splicing the list on the waiting side.
Quoted 2 lines
It seems like this one will keep pounding with
wake_up_statefor every single tw queued after @nr_wait, which quite often just 1.
Agree, that was just a silly oversight. I brought that back now. Apparently doesn't hit anything here, at least not to the extent that I saw it in testing. But it is a good idea and we should keep that, so only the first one over the threshold attempts the wake.
Signature · 1 line
Jens Axboe
Jens Axboe
maintainerRe: [PATCH 2/4] io_uring: switch deferred task_work to an io_wq_work_list
Quoted 10 lines · Jens Axboe
Quoted 2 lines
smp_mb(), see the comment below, and fwiw "_after_atomic" would not work.For this one, I think all we need to do is have the
wq_list_empty()check be fully stable. If we read:
nr_wait=atomic_read(&ctx->cq_wait_nr);right before a waiter does:
Quoted 7 lines
atomic_set(&ctx->cq_wait_nr, foo);set_current_state(TASK_INTERRUPTIBLE);then we need to ensure that the "I have work" check in
io_cqring_wait_schedule()sees the work. Thespin_unlock()has release semantics, and the currentREAD_ONCE()for work check sbould be enough, no?
To answer my own question - no, it's not enough. Let me think about this a bit.
Signature · 1 line
Jens Axboe
Re: [PATCH 2/4] io_uring: switch deferred task_work to an io_wq_work_list
Quoted 14 lines · Jens Axboe
Quoted 13 lines · Jens Axboe
Quoted 2 lines
smp_mb(), see the comment below, and fwiw "_after_atomic" would not work.For this one, I think all we need to do is have the
wq_list_empty()check be fully stable. If we read:
nr_wait=atomic_read(&ctx->cq_wait_nr);right before a waiter does:
atomic_set(&ctx->cq_wait_nr, foo);set_current_state(TASK_INTERRUPTIBLE);
Quoted 7 lines
Quoted 4 lines
then we need to ensure that the "I have work" check in
io_cqring_wait_schedule()sees the work. Thespin_unlock()has release semantics, and the currentREAD_ONCE()for work check sbould be enough, no?To answer my own question - no, it's not enough. Let me think about this a bit.
Right, to my knowledge release does nothing for write; read; ordering, and all ops after can leak before the barrier.
Signature · 1 line
Pavel Begunkov
Jens Axboe
maintainerRe: [PATCH 2/4] io_uring: switch deferred task_work to an io_wq_work_list
Quoted 20 lines · Pavel Begunkov
Quoted 19 lines · Jens Axboe
Quoted 17 lines · Jens Axboe
Quoted 2 lines
smp_mb(), see the comment below, and fwiw "_after_atomic" would not work.For this one, I think all we need to do is have the
wq_list_empty()check be fully stable. If we read:
nr_wait=atomic_read(&ctx->cq_wait_nr);right before a waiter does:
atomic_set(&ctx->cq_wait_nr, foo);set_current_state(TASK_INTERRUPTIBLE);then we need to ensure that the "I have work" check in
io_cqring_wait_schedule()sees the work. Thespin_unlock()has release semantics, and the currentREAD_ONCE()for work check sbould be enough, no?
Quoted 5 lines
Quoted 2 lines
To answer my own question - no, it's not enough. Let me think about this a bit.
Right, to my knowledge release does nothing for write; read; ordering, and all ops after can leak before the barrier.
Yeah, it needs an smp_mb() before that atomic_read() on the task work add side.
Signature · 1 line
Jens Axboe
Jens Axboe
maintainerPATCH 3/4 io_uring: switch fallback work to io_wq_work_list
In reply to Jens AxboeJust like what was done for deferred task_work, convert the fallback task_work to a normal io_wq_work_list.
--- include/linux/io_uring_types.h | 2 +- io_uring/io_uring.c | 24 +++++++++++++++++++----- 2 files changed, 20 insertions(+), 6 deletions(-)
include/linux/io_uring_types.h +1 -1
@@ -400,7 +400,7 @@ struct io_ring_ctx {struct mm_struct *mm_account;/* ctx exit and cancelation */ -struct llist_head fallback_llist; +struct io_wq_work_list fallback_list;struct delayed_work fallback_work;struct work_struct exit_work;struct list_head tctx_list;
io_uring/io_uring.c +19 -5
@@ -243,14 +243,22 @@ static __cold void io_fallback_req_func(struct work_struct *work) { struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, fallback_work.work); -struct llist_node *node =llist_del_all (&ctx->fallback_llist); -struct io_kiocb *req, *tmp; +struct io_wq_work_node *node;struct io_tw_state ts = {}; +struct io_kiocb *req; + +spin_lock_irq (&ctx->work_lock); + node = ctx->fallback_list.first; +INIT_WQ_LIST (&ctx->fallback_list); +spin_unlock_irq (&ctx->work_lock);percpu_ref_get (&ctx->refs);mutex_lock (&ctx->uring_lock); -llist_for_each_entry_safe (req, tmp, node, io_task_work.llist_node) +while (node) { + req = container_of(node, struct io_kiocb, io_task_work.node); + node = node->next; req->io_task_work.func (req, &ts); + }io_submit_flush_completions (ctx);mutex_unlock (&ctx->uring_lock);percpu_ref_put (&ctx->refs);
@@ -1167,6 +1175,9 @@ static __cold void io_fallback_tw(struct io_uring_task *tctx, bool sync)struct io_kiocb *req;while (node) { +unsigned long flags; +bool do_wake; + req = container_of(node, struct io_kiocb, io_task_work.llist_node); node = node->next;if (sync && last_ctx != req->ctx) {
@@ -1177,8 +1188,11 @@ static __cold void io_fallback_tw(struct io_uring_task *tctx, bool sync) last_ctx = req->ctx;percpu_ref_get (&last_ctx->refs); } -if (llist_add (&req->io_task_work.llist_node, - &req->ctx->fallback_llist)) +spin_lock_irqsave (&req->ctx->work_lock, flags); + do_wake =wq_list_empty (&req->ctx->fallback_list); +wq_list_add_tail (&req->io_task_work.node, &req->ctx->fallback_list); +spin_unlock_irqrestore (&req->ctx->work_lock, flags); +if (do_wake)schedule_delayed_work (&req->ctx->fallback_work,1 ); }
Signature · 1 line
2.43.0
Jens Axboe
maintainerPATCH 4/4 io_uring: switch normal task_work to io_wq_work_list
In reply to Jens AxboeThis concludes the transition, now all times of task_work are using the same mechanism.
--- include/linux/io_uring_types.h | 8 ++---- io_uring/io_uring.c | 50 ++++++++++++++++++++++------------ io_uring/io_uring.h | 4 +-- io_uring/sqpoll.c | 8 +++--- io_uring/tctx.c | 3 +- 5 files changed, 43 insertions(+), 30 deletions(-)
include/linux/io_uring_types.h +3 -5
@@ -95,7 +95,8 @@ struct io_uring_task {struct percpu_counter inflight;struct {/* task_work */ -struct llist_head task_list; +struct io_wq_work_list task_list; +spinlock_t task_lock;struct callback_head task_work; }____cacheline_aligned_in_smp ; };
@@ -561,10 +562,7 @@ enum {typedef void (*io_req_tw_func_t )(struct io_kiocb *req,struct io_tw_state *ts);struct io_task_work { -union { -struct io_wq_work_node node; -struct llist_node llist_node; - }; +struct io_wq_work_node node;io_req_tw_func_t func; };
io_uring/io_uring.c +32 -18
103 lines not on this page · raw message
io_uring/io_uring.h +2 -2
@@ -87,8 +87,8 @@ void io_req_task_queue(struct io_kiocb *req);void io_req_task_complete (struct io_kiocb *req,struct io_tw_state *ts);void io_req_task_queue_fail (struct io_kiocb *req,int ret);void io_req_task_submit (struct io_kiocb *req,struct io_tw_state *ts); -struct llist_node *io_handle_tw_list (struct llist_node *node,unsigned int *count,unsigned int max_entries); -struct llist_node *tctx_task_work_run (struct io_uring_task *tctx,unsigned int max_entries,unsigned int *count); +struct io_wq_work_node *io_handle_tw_list (struct io_wq_work_node *node,unsigned int *count,unsigned int max_entries); +struct io_wq_work_node *tctx_task_work_run (struct io_uring_task *tctx,unsigned int max_entries,unsigned int *count);void tctx_task_work (struct callback_head *cb);__cold void io_uring_cancel_generic (bool cancel_all,struct io_sq_data *sqd); int io_uring_alloc_task_context(struct task_struct *task,
io_uring/sqpoll.c +4 -4
@@ -230,7 +230,7 @@ static bool io_sqd_handle_event(struct io_sq_data *sqd)* than we were asked to process. Newly queued task_work isn't run until the * retry list has been fully processed. */ -static unsigned int io_sq_tw (struct llist_node **retry_list,int max_entries) +static unsigned int io_sq_tw (struct io_wq_work_node **retry_list,int max_entries) {struct io_uring_task *tctx = current->io_uring;unsigned int count =0 ;
@@ -246,11 +246,11 @@ static unsigned int io_sq_tw(struct llist_node **retry_list, int max_entries)return count; } -static bool io_sq_tw_pending (struct llist_node *retry_list) +static bool io_sq_tw_pending (struct io_wq_work_node *retry_list) {struct io_uring_task *tctx = current->io_uring; -return retry_list || !llist_empty (&tctx->task_list); +return retry_list || !wq_list_empty (&tctx->task_list); }static void io_sq_update_worktime (struct io_sq_data *sqd,struct rusage *start)
@@ -266,7 +266,7 @@ static void io_sq_update_worktime(struct io_sq_data *sqd, struct rusage *start)static int io_sq_thread (void *data) { -struct llist_node *retry_list =NULL ; +struct io_wq_work_node *retry_list =NULL ;struct io_sq_data *sqd = data;struct io_ring_ctx *ctx;struct rusage start;
io_uring/tctx.c +2 -1
@@ -86,7 +86,8 @@ __cold int io_uring_alloc_task_context(struct task_struct *task,atomic_set (&tctx->in_cancel,0 );atomic_set (&tctx->inflight_tracked,0 ); task->io_uring = tctx; -init_llist_head (&tctx->task_list); +INIT_WQ_LIST (&tctx->task_list); +spin_lock_init (&tctx->task_lock);init_task_work (&tctx->task_work, tctx_task_work);return 0 ; }
Signature · 1 line
2.43.0
Quoted 11 lines · Jens Axboe
Hi,
This converts the deferred, normal, and fallback
task_workto use a normalio_wq_work_list, rather than an llist.The main motivation behind this is to get rid of the need to reverse the list once it's deleted and run. I tested this basic conversion of just switching it from an llist to an
io_wq_work_listwith a spinlock, and I don't see any benefits from the lockless list. And for cases where we get a bursty addition oftask_work, this approach is faster as it avoids the need to iterate the list upfront while reversing it.
I'm curious how you benchmarked it including accounting of irq/softirq where tw add usually happens?
One known problem with the current list approach I mentioned several times before is that it peeks at the previous queued tw to count them. It's not nice, but that can be easily done with cmpxchg double. I wonder how much of an issue is that.
Quoted 1 line
And this is less code and simpler, so I'd prefer to go that route.
I'm not sure it's less code, if you return optimisations that I believe were killed, see comments to patch 2, it might turn out to be even bulkier and not that simpler.
Quoted 7 lines
include/linux/io_uring_types.h | 13 +-- io_uring/io_uring.c | 175 ++++++++++++++++----------------- io_uring/io_uring.h | 8 +- io_uring/sqpoll.c | 8 +- io_uring/tctx.c | 3 +- 5 files changed, 102 insertions(+), 105 deletions(-)
Signature · 1 line
Pavel Begunkov
Jens Axboe
maintainerQuoted 13 lines · Pavel Begunkov
Quoted 11 lines · Jens Axboe
Hi,
This converts the deferred, normal, and fallback
task_workto use a normalio_wq_work_list, rather than an llist.The main motivation behind this is to get rid of the need to reverse the list once it's deleted and run. I tested this basic conversion of just switching it from an llist to an
io_wq_work_listwith a spinlock, and I don't see any benefits from the lockless list. And for cases where we get a bursty addition oftask_work, this approach is faster as it avoids the need to iterate the list upfront while reversing it.
Quoted 2 lines
I'm curious how you benchmarked it including accounting of irq/softirq where tw add usually happens?
Performance based and profiles. I tested send zc with small packets, as that is task_work intensive and exhibits the bursty behavior I mentioned in the patch / cover letter. And normal storage IO, IRQ driven.
For send zc, we're spending about 2% of the time doing list reversal, and I've seen as high as 5 in other testing. And as that test is CPU bound, performance is up about 2% as well.
With the patches, task work adding accounts for about 0.25% of the cycles, before it's about 0.66%.
We're spending a bit more time in __io_run_local_work(), but I think that's deceptive as we have to disable/enable interrupts now. If an interrupt triggers on the unlock, that time tends to be attributed there in terms of cycles.
Quoted 4 lines
One known problem with the current list approach I mentioned several times before is that it peeks at the previous queued tw to count them. It's not nice, but that can be easily done with cmpxchg double. I wonder how much of an issue is that.
That's more of a wart than a real issue though, but we this approach obviously doesn't do that. And then we can drop the rcu section around adding local task_work. Not a huge deal, but still nice.
Quoted 5 lines
Quoted 1 line
And this is less code and simpler, so I'd prefer to go that route.
I'm not sure it's less code, if you return optimisations that I believe were killed, see comments to patch 2, it might turn out to be even bulkier and not that simpler.
It's still considerably less:
3 files changed, 59 insertions(+), 84 deletions(-)
thought that isn't conclusive by itself, as eg io_llist_xchg() goes away which has some comments. But I do think the resulting code is simpler and more straight forward.
Signature · 1 line
Jens Axboe
Jens Axboe
maintainerQuoted 25 lines · Jens Axboe
Quoted 13 lines · Pavel Begunkov
Quoted 11 lines · Jens Axboe
Hi,
This converts the deferred, normal, and fallback
task_workto use a normalio_wq_work_list, rather than an llist.The main motivation behind this is to get rid of the need to reverse the list once it's deleted and run. I tested this basic conversion of just switching it from an llist to an
io_wq_work_listwith a spinlock, and I don't see any benefits from the lockless list. And for cases where we get a bursty addition oftask_work, this approach is faster as it avoids the need to iterate the list upfront while reversing it.Quoted 2 lines
I'm curious how you benchmarked it including accounting of irq/softirq where tw add usually happens?
Performance based and profiles. I tested send zc with small packets, as that is
task_workintensive and exhibits the bursty behavior I mentioned in the patch / cover letter. And normal storage IO, IRQ driven.For send zc, we're spending about 2% of the time doing list reversal, and I've seen as high as 5 in other testing. And as that test is CPU bound, performance is up about 2% as well.
Quoted 7 lines
With the patches, task work adding accounts for about 0.25% of the cycles, before it's about 0.66%.
We're spending a bit more time in
__io_run_local_work(), but I think that's deceptive as we have to disable/enable interrupts now. If an interrupt triggers on the unlock, that time tends to be attributed there in terms of cycles.
Forgot to mention the storage side - profiles look eerily similar in terms of time spent in task work adding / running, the only real difference is that 1.9% of llist_reverse_list() is gone.
Signature · 1 line
Jens Axboe
Quoted 14 lines · Jens Axboe
Quoted 13 lines · Pavel Begunkov
Quoted 11 lines · Jens Axboe
Hi,
This converts the deferred, normal, and fallback
task_workto use a normalio_wq_work_list, rather than an llist.The main motivation behind this is to get rid of the need to reverse the list once it's deleted and run. I tested this basic conversion of just switching it from an llist to an
io_wq_work_listwith a spinlock, and I don't see any benefits from the lockless list. And for cases where we get a bursty addition oftask_work, this approach is faster as it avoids the need to iterate the list upfront while reversing it.
Quoted 6 lines
Quoted 2 lines
I'm curious how you benchmarked it including accounting of irq/softirq where tw add usually happens?
Performance based and profiles. I tested send zc with small packets, as that is
task_workintensive and exhibits the bursty behavior I mentioned in the patch / cover letter. And normal storage IO, IRQ driven.
I assume IRQs are firing on random CPUs then unless you configured it. In which case it should be bouncing of the cacheline + that peeking at the prev also needs to fetch it from RAM / further caches. Unless there is enough of time for TCP to batch them.
Quoted 2 lines
For send zc, we're spending about 2% of the time doing list reversal, and I've seen as high as 5 in other testing. And as that test is CPU
I've seen similar before, but for me it was overhead shifted from __io_run_local_work() fetching requests into reverse touching all of them. There should be a change in __io_run_local_work() total cycles (incl children) then I assume
Quoted 1 line
bound, performance is up about 2% as well.
Did you count by any chance how many items there was in the list? Average or so
Quoted 2 lines
With the patches, task work adding accounts for about 0.25% of the cycles, before it's about 0.66%.
i.e. spinlock is faster. How come? Same cmpxchg in spinlock with often cache misses, but with irq on/off on top. The only diff I can remember is that peek into prev req.
Quoted 4 lines
We're spending a bit more time in
__io_run_local_work(), but I think that's deceptive as we have to disable/enable interrupts now. If an interrupt triggers on the unlock, that time tends to be attributed there in terms of cycles.
Hmm, I think if run_local_work runtime doesn't change you'd statistically get same number of interrupts "items" hitting it, but the would be condensed more to irq-off. Or are you accounting for some irq delivery / hw differences?
Quoted 6 lines
Quoted 4 lines
One known problem with the current list approach I mentioned several times before is that it peeks at the previous queued tw to count them. It's not nice, but that can be easily done with cmpxchg double. I wonder how much of an issue is that.
That's more of a wart than a real issue though, but we this approach
Assuming tw add executing on random CPUs, that would be additional fetch every tw add, I wouldn't disregard it right away.
Quoted 2 lines
obviously doesn't do that. And then we can drop the rcu section around adding local
task_work. Not a huge deal, but still nice.
I don't see how. After you queue the req it might be immediately executed dropping the ctx ref, which was previously protecting ctx. The rcu section protects ctx.
Quoted 13 lines
Quoted 5 lines
Quoted 1 line
And this is less code and simpler, so I'd prefer to go that route.
I'm not sure it's less code, if you return optimisations that I believe were killed, see comments to patch 2, it might turn out to be even bulkier and not that simpler.
It's still considerably less:
3 files changed, 59 insertions(+), 84 deletions(-)thought that isn't conclusive by itself, as eg
io_llist_xchg()goes away which has some comments. But I do think the resulting code is simpler and more straight forward.
Signature · 1 line
Pavel Begunkov