The current approach is a bit naive, and hence calls the time querying way too often. Only start the "doing work" timer when there's actual work to do, and then use that information to terminate (and account) the work time once done. This greatly reduces the frequency of these calls, when they cannot have changed anyway.

Running a basic random reader that is setup to use SQPOLL, a profile before this change shows these as the top cycle consumers:

+ 32.60% iou-sqp-1074 [kernel.kallsyms] [k] thread_group_cputime_adjusted + 19.97% iou-sqp-1074 [kernel.kallsyms] [k] thread_group_cputime + 12.20% io_uring io_uring [.] submitter_uring_fn + 4.13% iou-sqp-1074 [kernel.kallsyms] [k] getrusage + 2.45% iou-sqp-1074 [kernel.kallsyms] [k] io_submit_sqes + 2.18% iou-sqp-1074 [kernel.kallsyms] [k] __pi_memset_generic + 2.09% iou-sqp-1074 [kernel.kallsyms] [k] cputime_adjust

and after this change, top of profile looks as follows:

+ 36.23% io_uring io_uring [.] submitter_uring_fn + 23.26% iou-sqp-819 [kernel.kallsyms] [k] io_sq_thread + 10.14% iou-sqp-819 [kernel.kallsyms] [k] io_sq_tw + 6.52% iou-sqp-819 [kernel.kallsyms] [k] tctx_task_work_run + 4.82% iou-sqp-819 [kernel.kallsyms] [k] nvme_submit_cmds.part.0 + 2.91% iou-sqp-819 [kernel.kallsyms] [k] io_submit_sqes [...] 0.02% iou-sqp-819 [kernel.kallsyms] [k] cputime_adjust

where it's spending the cycles on things that actually matter.

Reported-by: Fengnan Chang changfengnan@bytedance.com Cc: stable@vger.kernel.org Fixes: 3fcb9d17206e ("io_uring/sqpoll: statistics of the true utilization of sq threads") Signed-off-by: Jens Axboe axboe@kernel.dk --- io_uring/sqpoll.c | 43 ++++++++++++++++++++++++++++++++----------- 1 file changed, 32 insertions(+), 11 deletions(-)

diff --git a/io_uring/sqpoll.c b/io_uring/sqpoll.c index 2b816fdb9866..e22f072c7d5f 100644 --- a/io_uring/sqpoll.c +++ b/io_uring/sqpoll.c @@ -170,6 +170,11 @@ static inline bool io_sqd_events_pending(struct io_sq_data *sqd) return READ_ONCE(sqd->state); }

+struct io_sq_time { + bool started; + u64 usec; +}; + u64 io_sq_cpu_usec(struct task_struct *tsk) { u64 utime, stime; @@ -179,12 +184,24 @@ u64 io_sq_cpu_usec(struct task_struct *tsk) return stime; }

-static void io_sq_update_worktime(struct io_sq_data *sqd, u64 usec) +static void io_sq_update_worktime(struct io_sq_data *sqd, struct io_sq_time *ist) +{ + if (!ist->started) + return; + ist->started = false; + sqd->work_time += io_sq_cpu_usec(current) - ist->usec; +} + +static void io_sq_start_worktime(struct io_sq_time *ist) { - sqd->work_time += io_sq_cpu_usec(current) - usec; + if (ist->started) + return; + ist->started = true; + ist->usec = io_sq_cpu_usec(current); }

-static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries) +static int __io_sq_thread(struct io_ring_ctx *ctx, struct io_sq_data *sqd, + bool cap_entries, struct io_sq_time *ist) { unsigned int to_submit; int ret = 0; @@ -197,6 +214,8 @@ static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries) if (to_submit || !wq_list_empty(&ctx->iopoll_list)) { const struct cred *creds = NULL;

+ io_sq_start_worktime(ist); + if (ctx->sq_creds != current_cred()) creds = override_creds(ctx->sq_creds);

@@ -278,7 +297,6 @@ static int io_sq_thread(void *data) unsigned long timeout = 0; char buf[TASK_COMM_LEN] = {}; DEFINE_WAIT(wait); - u64 start;

/* offload context creation failed, just exit */ if (!current->io_uring) { @@ -313,6 +331,7 @@ static int io_sq_thread(void *data) mutex_lock(&sqd->lock); while (1) { bool cap_entries, sqt_spin = false; + struct io_sq_time ist = { };

if (io_sqd_events_pending(sqd) || signal_pending(current)) { if (io_sqd_handle_event(sqd)) @@ -321,9 +340,8 @@ static int io_sq_thread(void *data) }

cap_entries = !list_is_singular(&sqd->ctx_list); - start = io_sq_cpu_usec(current); list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) { - int ret = __io_sq_thread(ctx, cap_entries); + int ret = __io_sq_thread(ctx, sqd, cap_entries, &ist);

if (!sqt_spin && (ret > 0 || !wq_list_empty(&ctx->iopoll_list))) sqt_spin = true; @@ -331,15 +349,18 @@ static int io_sq_thread(void *data) if (io_sq_tw(&retry_list, IORING_TW_CAP_ENTRIES_VALUE)) sqt_spin = true;

- list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) - if (io_napi(ctx)) + list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) { + if (io_napi(ctx)) { + io_sq_start_worktime(&ist); io_napi_sqpoll_busy_poll(ctx); + } + } + + io_sq_update_worktime(sqd, &ist);

if (sqt_spin || !time_after(jiffies, timeout)) { - if (sqt_spin) { - io_sq_update_worktime(sqd, start); + if (sqt_spin) timeout = jiffies + sqd->sq_thread_idle; - } if (unlikely(need_resched())) { mutex_unlock(&sqd->lock); cond_resched();