If this is set, io_uring will never use an IPI to deliver a task_work notification. This can be used in the common case where a single task or thread communicates with the ring, and doesn't rely on io_uring_cqe_peek(). This provides a noticeable win in performance, both from eliminating the IPI itself, but also from avoiding interrupting the submitting task unnecessarily.
Unified · Split
| @@ -11327,12 +11327,20 @@ static __cold int io_uring_create(unsigned entries, struct io_uring_params *p, · 11326 unmodified lines | ||
| 11327 | 11327 | ctx->user = get_uid(current_user()); |
| 11328 | 11328 | |
| 11329 | 11329 | /* |
| 11330 | − * For SQPOLL, we just need a wakeup, always. | |
| 11330 | + * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if | |
| 11331 | + * COOP_TASKRUN is set, then IPIs are never needed by the app. | |
| 11331 | 11332 | */ |
| 11332 | − if (ctx->flags & IORING_SETUP_SQPOLL) | |
| 11333 | + ret = -EINVAL; | |
| 11334 | + if (ctx->flags & IORING_SETUP_SQPOLL) { | |
| 11335 | + /* IPI related flags don't make sense with SQPOLL */ | |
| 11336 | + if (ctx->flags & IORING_SETUP_COOP_TASKRUN) | |
| 11337 | + goto err; | |
| 11333 | 11338 | ctx->notify_method = TWA_SIGNAL_NO_IPI; |
| 11334 | − else | |
| 11339 | + } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) { | |
| 11340 | + ctx->notify_method = TWA_SIGNAL_NO_IPI; | |
| 11341 | + } else { | |
| 11335 | 11342 | ctx->notify_method = TWA_SIGNAL; |
| 11343 | + } | |
| 11336 | 11344 | |
| 11337 | 11345 | /* |
| 11338 | 11346 | * This is just grabbed for accounting purposes. When a process exits, |
| @@ -11431,7 +11439,8 @@ static long io_uring_setup(u32 entries, struct io_uring_params __user *params) · 92 unmodified lines | ||
| 11431 | 11439 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
| 11432 | 11440 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | |
| 11433 | 11441 | IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ | |
| 11434 | − IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL)) | |
| 11442 | + IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL | | |
| 11443 | + IORING_SETUP_COOP_TASKRUN)) | |
| 11435 | 11444 | return -EINVAL; |
| 11436 | 11445 | |
| 11437 | 11446 | return io_uring_create(entries, &p, params); |
| @@ -102,6 +102,14 @@ enum { · 101 unmodified lines | ||
| 102 | 102 | #define IORING_SETUP_ATTACH_WQ (1U << 5) /* attach to existing wq */ |
| 103 | 103 | #define IORING_SETUP_R_DISABLED (1U << 6) /* start with ring disabled */ |
| 104 | 104 | #define IORING_SETUP_SUBMIT_ALL (1U << 7) /* continue submit on error */ |
| 105 | +/* | |
| 106 | + * Cooperative task running. When requests complete, they often require | |
| 107 | + * forcing the submitter to transition to the kernel to complete. If this | |
| 108 | + * flag is set, work will be done when the task transitions anyway, rather | |
| 109 | + * than force an inter-processor interrupt reschedule. This avoids interrupting | |
| 110 | + * a task running in userspace, and saves an IPI. | |
| 111 | + */ | |
| 112 | +#define IORING_SETUP_COOP_TASKRUN (1U << 8) | |
| 105 | 113 | |
| 106 | 114 | enum { |
| 107 | 115 | IORING_OP_NOP, |