1de23077eSJens Axboe #ifndef IOU_CORE_H 2de23077eSJens Axboe #define IOU_CORE_H 3de23077eSJens Axboe 4de23077eSJens Axboe #include <linux/errno.h> 5cd40cae2SJens Axboe #include <linux/lockdep.h> 6ab1c84d8SPavel Begunkov #include <linux/io_uring_types.h> 7ab1c84d8SPavel Begunkov #include "io-wq.h" 8a6b21fbbSPavel Begunkov #include "slist.h" 9ab1c84d8SPavel Begunkov #include "filetable.h" 10de23077eSJens Axboe 11f3b44f92SJens Axboe #ifndef CREATE_TRACE_POINTS 12f3b44f92SJens Axboe #include <trace/events/io_uring.h> 13f3b44f92SJens Axboe #endif 14f3b44f92SJens Axboe 1597b388d7SJens Axboe enum { 1697b388d7SJens Axboe IOU_OK = 0, 1797b388d7SJens Axboe IOU_ISSUE_SKIP_COMPLETE = -EIOCBQUEUED, 18*114eccdfSDylan Yudaken 19*114eccdfSDylan Yudaken /* 20*114eccdfSDylan Yudaken * Intended only when both REQ_F_POLLED and REQ_F_APOLL_MULTISHOT 21*114eccdfSDylan Yudaken * are set to indicate to the poll runner that multishot should be 22*114eccdfSDylan Yudaken * removed and the result is set on req->cqe.res. 23*114eccdfSDylan Yudaken */ 24*114eccdfSDylan Yudaken IOU_STOP_MULTISHOT = -ECANCELED, 2597b388d7SJens Axboe }; 2697b388d7SJens Axboe 27faf88ddeSPavel Begunkov struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx); 2868494a65SPavel Begunkov bool io_req_cqe_overflow(struct io_kiocb *req); 299046c641SPavel Begunkov int io_run_task_work_sig(void); 309046c641SPavel Begunkov void io_req_complete_failed(struct io_kiocb *req, s32 res); 319046c641SPavel Begunkov void __io_req_complete(struct io_kiocb *req, unsigned issue_flags); 329046c641SPavel Begunkov void io_req_complete_post(struct io_kiocb *req); 339046c641SPavel Begunkov void __io_req_complete_post(struct io_kiocb *req); 349046c641SPavel Begunkov bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags); 359046c641SPavel Begunkov void __io_commit_cqring_flush(struct io_ring_ctx *ctx); 369046c641SPavel Begunkov 379046c641SPavel Begunkov struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages); 389046c641SPavel Begunkov 399046c641SPavel Begunkov struct file *io_file_get_normal(struct io_kiocb *req, int fd); 409046c641SPavel Begunkov struct file *io_file_get_fixed(struct io_kiocb *req, int fd, 419046c641SPavel Begunkov unsigned issue_flags); 429046c641SPavel Begunkov 439046c641SPavel Begunkov bool io_is_uring_fops(struct file *file); 449046c641SPavel Begunkov bool io_alloc_async_data(struct io_kiocb *req); 459046c641SPavel Begunkov void io_req_task_work_add(struct io_kiocb *req); 469046c641SPavel Begunkov void io_req_tw_post_queue(struct io_kiocb *req, s32 res, u32 cflags); 479046c641SPavel Begunkov void io_req_task_queue(struct io_kiocb *req); 489046c641SPavel Begunkov void io_queue_iowq(struct io_kiocb *req, bool *dont_use); 499046c641SPavel Begunkov void io_req_task_complete(struct io_kiocb *req, bool *locked); 509046c641SPavel Begunkov void io_req_task_queue_fail(struct io_kiocb *req, int ret); 519046c641SPavel Begunkov void io_req_task_submit(struct io_kiocb *req, bool *locked); 529046c641SPavel Begunkov void tctx_task_work(struct callback_head *cb); 539046c641SPavel Begunkov __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd); 549046c641SPavel Begunkov int io_uring_alloc_task_context(struct task_struct *task, 559046c641SPavel Begunkov struct io_ring_ctx *ctx); 569046c641SPavel Begunkov 579046c641SPavel Begunkov int io_poll_issue(struct io_kiocb *req, bool *locked); 589046c641SPavel Begunkov int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr); 599046c641SPavel Begunkov int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin); 609046c641SPavel Begunkov void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node); 619046c641SPavel Begunkov int io_req_prep_async(struct io_kiocb *req); 629046c641SPavel Begunkov 639046c641SPavel Begunkov struct io_wq_work *io_wq_free_work(struct io_wq_work *work); 649046c641SPavel Begunkov void io_wq_submit_work(struct io_wq_work *work); 659046c641SPavel Begunkov 669046c641SPavel Begunkov void io_free_req(struct io_kiocb *req); 679046c641SPavel Begunkov void io_queue_next(struct io_kiocb *req); 689046c641SPavel Begunkov 699046c641SPavel Begunkov bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task, 709046c641SPavel Begunkov bool cancel_all); 719046c641SPavel Begunkov 729046c641SPavel Begunkov #define io_for_each_link(pos, head) \ 739046c641SPavel Begunkov for (pos = (head); pos; pos = pos->link) 74f3b44f92SJens Axboe 7525399321SPavel Begunkov static inline void io_cq_lock(struct io_ring_ctx *ctx) 7625399321SPavel Begunkov __acquires(ctx->completion_lock) 7725399321SPavel Begunkov { 7825399321SPavel Begunkov spin_lock(&ctx->completion_lock); 7925399321SPavel Begunkov } 8025399321SPavel Begunkov 8125399321SPavel Begunkov void io_cq_unlock_post(struct io_ring_ctx *ctx); 8225399321SPavel Begunkov 83f3b44f92SJens Axboe static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx) 84f3b44f92SJens Axboe { 85f3b44f92SJens Axboe if (likely(ctx->cqe_cached < ctx->cqe_sentinel)) { 86f3b44f92SJens Axboe struct io_uring_cqe *cqe = ctx->cqe_cached; 87f3b44f92SJens Axboe 88f3b44f92SJens Axboe ctx->cached_cq_tail++; 89f3b44f92SJens Axboe ctx->cqe_cached++; 90b3659a65SPavel Begunkov if (ctx->flags & IORING_SETUP_CQE32) 91b3659a65SPavel Begunkov ctx->cqe_cached++; 92f3b44f92SJens Axboe return cqe; 93f3b44f92SJens Axboe } 94f3b44f92SJens Axboe 95f3b44f92SJens Axboe return __io_get_cqe(ctx); 96f3b44f92SJens Axboe } 97f3b44f92SJens Axboe 98f3b44f92SJens Axboe static inline bool __io_fill_cqe_req(struct io_ring_ctx *ctx, 99f3b44f92SJens Axboe struct io_kiocb *req) 100f3b44f92SJens Axboe { 101f3b44f92SJens Axboe struct io_uring_cqe *cqe; 102f3b44f92SJens Axboe 103f3b44f92SJens Axboe trace_io_uring_complete(req->ctx, req, req->cqe.user_data, 104ae5735c6SPavel Begunkov req->cqe.res, req->cqe.flags, 105ae5735c6SPavel Begunkov (req->flags & REQ_F_CQE32_INIT) ? req->extra1 : 0, 106ae5735c6SPavel Begunkov (req->flags & REQ_F_CQE32_INIT) ? req->extra2 : 0); 107f3b44f92SJens Axboe /* 108f3b44f92SJens Axboe * If we can't get a cq entry, userspace overflowed the 109f3b44f92SJens Axboe * submission (by quite a lot). Increment the overflow count in 110f3b44f92SJens Axboe * the ring. 111f3b44f92SJens Axboe */ 112f3b44f92SJens Axboe cqe = io_get_cqe(ctx); 113e8c328c3SPavel Begunkov if (unlikely(!cqe)) 114e8c328c3SPavel Begunkov return io_req_cqe_overflow(req); 115f3b44f92SJens Axboe memcpy(cqe, &req->cqe, sizeof(*cqe)); 116e8c328c3SPavel Begunkov 117e8c328c3SPavel Begunkov if (ctx->flags & IORING_SETUP_CQE32) { 118f3b44f92SJens Axboe u64 extra1 = 0, extra2 = 0; 119f3b44f92SJens Axboe 120f3b44f92SJens Axboe if (req->flags & REQ_F_CQE32_INIT) { 121f3b44f92SJens Axboe extra1 = req->extra1; 122f3b44f92SJens Axboe extra2 = req->extra2; 123f3b44f92SJens Axboe } 124f3b44f92SJens Axboe 125f3b44f92SJens Axboe WRITE_ONCE(cqe->big_cqe[0], extra1); 126f3b44f92SJens Axboe WRITE_ONCE(cqe->big_cqe[1], extra2); 127e8c328c3SPavel Begunkov } 128f3b44f92SJens Axboe return true; 129f3b44f92SJens Axboe } 130f3b44f92SJens Axboe 131531113bbSJens Axboe static inline void req_set_fail(struct io_kiocb *req) 132531113bbSJens Axboe { 133531113bbSJens Axboe req->flags |= REQ_F_FAIL; 134531113bbSJens Axboe if (req->flags & REQ_F_CQE_SKIP) { 135531113bbSJens Axboe req->flags &= ~REQ_F_CQE_SKIP; 136531113bbSJens Axboe req->flags |= REQ_F_SKIP_LINK_CQES; 137531113bbSJens Axboe } 138531113bbSJens Axboe } 139531113bbSJens Axboe 140de23077eSJens Axboe static inline void io_req_set_res(struct io_kiocb *req, s32 res, u32 cflags) 141de23077eSJens Axboe { 142de23077eSJens Axboe req->cqe.res = res; 143de23077eSJens Axboe req->cqe.flags = cflags; 144de23077eSJens Axboe } 145de23077eSJens Axboe 14699f15d8dSJens Axboe static inline bool req_has_async_data(struct io_kiocb *req) 14799f15d8dSJens Axboe { 14899f15d8dSJens Axboe return req->flags & REQ_F_ASYNC_DATA; 14999f15d8dSJens Axboe } 15099f15d8dSJens Axboe 151531113bbSJens Axboe static inline void io_put_file(struct file *file) 152531113bbSJens Axboe { 153531113bbSJens Axboe if (file) 154531113bbSJens Axboe fput(file); 155531113bbSJens Axboe } 156531113bbSJens Axboe 157cd40cae2SJens Axboe static inline void io_ring_submit_unlock(struct io_ring_ctx *ctx, 158cd40cae2SJens Axboe unsigned issue_flags) 159cd40cae2SJens Axboe { 160cd40cae2SJens Axboe lockdep_assert_held(&ctx->uring_lock); 161cd40cae2SJens Axboe if (issue_flags & IO_URING_F_UNLOCKED) 162cd40cae2SJens Axboe mutex_unlock(&ctx->uring_lock); 163cd40cae2SJens Axboe } 164cd40cae2SJens Axboe 165cd40cae2SJens Axboe static inline void io_ring_submit_lock(struct io_ring_ctx *ctx, 166cd40cae2SJens Axboe unsigned issue_flags) 167cd40cae2SJens Axboe { 168cd40cae2SJens Axboe /* 169cd40cae2SJens Axboe * "Normal" inline submissions always hold the uring_lock, since we 170cd40cae2SJens Axboe * grab it from the system call. Same is true for the SQPOLL offload. 171cd40cae2SJens Axboe * The only exception is when we've detached the request and issue it 172cd40cae2SJens Axboe * from an async worker thread, grab the lock for that case. 173cd40cae2SJens Axboe */ 174cd40cae2SJens Axboe if (issue_flags & IO_URING_F_UNLOCKED) 175cd40cae2SJens Axboe mutex_lock(&ctx->uring_lock); 176cd40cae2SJens Axboe lockdep_assert_held(&ctx->uring_lock); 177cd40cae2SJens Axboe } 178cd40cae2SJens Axboe 179f9ead18cSJens Axboe static inline void io_commit_cqring(struct io_ring_ctx *ctx) 180f9ead18cSJens Axboe { 181f9ead18cSJens Axboe /* order cqe stores with ring update */ 182f9ead18cSJens Axboe smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail); 183f9ead18cSJens Axboe } 184f9ead18cSJens Axboe 185f3b44f92SJens Axboe static inline void io_cqring_wake(struct io_ring_ctx *ctx) 186f3b44f92SJens Axboe { 187f3b44f92SJens Axboe /* 188f3b44f92SJens Axboe * wake_up_all() may seem excessive, but io_wake_function() and 189f3b44f92SJens Axboe * io_should_wake() handle the termination of the loop and only 190f3b44f92SJens Axboe * wake as many waiters as we need to. 191f3b44f92SJens Axboe */ 192f3b44f92SJens Axboe if (wq_has_sleeper(&ctx->cq_wait)) 193f3b44f92SJens Axboe wake_up_all(&ctx->cq_wait); 194f3b44f92SJens Axboe } 195f3b44f92SJens Axboe 19617437f31SJens Axboe static inline bool io_sqring_full(struct io_ring_ctx *ctx) 19717437f31SJens Axboe { 19817437f31SJens Axboe struct io_rings *r = ctx->rings; 19917437f31SJens Axboe 20017437f31SJens Axboe return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries; 20117437f31SJens Axboe } 20217437f31SJens Axboe 20317437f31SJens Axboe static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx) 20417437f31SJens Axboe { 20517437f31SJens Axboe struct io_rings *rings = ctx->rings; 20617437f31SJens Axboe 20717437f31SJens Axboe /* make sure SQ entry isn't read before tail */ 20817437f31SJens Axboe return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head; 20917437f31SJens Axboe } 21017437f31SJens Axboe 21117437f31SJens Axboe static inline bool io_run_task_work(void) 21217437f31SJens Axboe { 213625d38b3SPavel Begunkov if (test_thread_flag(TIF_NOTIFY_SIGNAL)) { 21417437f31SJens Axboe __set_current_state(TASK_RUNNING); 21517437f31SJens Axboe clear_notify_signal(); 21617437f31SJens Axboe if (task_work_pending(current)) 21717437f31SJens Axboe task_work_run(); 21817437f31SJens Axboe return true; 21917437f31SJens Axboe } 22017437f31SJens Axboe 22117437f31SJens Axboe return false; 22217437f31SJens Axboe } 22317437f31SJens Axboe 224aa1e90f6SPavel Begunkov static inline void io_tw_lock(struct io_ring_ctx *ctx, bool *locked) 225aa1e90f6SPavel Begunkov { 226aa1e90f6SPavel Begunkov if (!*locked) { 227aa1e90f6SPavel Begunkov mutex_lock(&ctx->uring_lock); 228aa1e90f6SPavel Begunkov *locked = true; 229aa1e90f6SPavel Begunkov } 230aa1e90f6SPavel Begunkov } 231aa1e90f6SPavel Begunkov 2329da070b1SPavel Begunkov /* 2339da070b1SPavel Begunkov * Don't complete immediately but use deferred completion infrastructure. 2349da070b1SPavel Begunkov * Protected by ->uring_lock and can only be used either with 2359da070b1SPavel Begunkov * IO_URING_F_COMPLETE_DEFER or inside a tw handler holding the mutex. 2369da070b1SPavel Begunkov */ 2379da070b1SPavel Begunkov static inline void io_req_complete_defer(struct io_kiocb *req) 2389da070b1SPavel Begunkov __must_hold(&req->ctx->uring_lock) 239aa1e90f6SPavel Begunkov { 240aa1e90f6SPavel Begunkov struct io_submit_state *state = &req->ctx->submit_state; 241aa1e90f6SPavel Begunkov 2429da070b1SPavel Begunkov lockdep_assert_held(&req->ctx->uring_lock); 2439da070b1SPavel Begunkov 244aa1e90f6SPavel Begunkov wq_list_add_tail(&req->comp_list, &state->compl_reqs); 245aa1e90f6SPavel Begunkov } 246aa1e90f6SPavel Begunkov 24746929b08SPavel Begunkov static inline void io_commit_cqring_flush(struct io_ring_ctx *ctx) 24846929b08SPavel Begunkov { 24946929b08SPavel Begunkov if (unlikely(ctx->off_timeout_used || ctx->drain_active || ctx->has_evfd)) 25046929b08SPavel Begunkov __io_commit_cqring_flush(ctx); 25146929b08SPavel Begunkov } 25246929b08SPavel Begunkov 253de23077eSJens Axboe #endif 254