4 #include <linux/errno.h>
5 #include <linux/lockdep.h>
6 #include "io_uring_types.h"
8 #ifndef CREATE_TRACE_POINTS
9 #include <trace/events/io_uring.h>
14 IOU_ISSUE_SKIP_COMPLETE = -EIOCBQUEUED,
17 bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data, s32 res,
18 u32 cflags, u64 extra1, u64 extra2);
20 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
22 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
26 * writes to the cq entry need to come after reading head; the
27 * control dependency is enough as we're using WRITE_ONCE to
30 static inline struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx)
32 struct io_rings *rings = ctx->rings;
33 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
34 unsigned int shift = 0;
35 unsigned int free, queued, len;
37 if (ctx->flags & IORING_SETUP_CQE32)
40 /* userspace may cheat modifying the tail, be safe and do min */
41 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
42 free = ctx->cq_entries - queued;
43 /* we need a contiguous range, limit based on the current array offset */
44 len = min(free, ctx->cq_entries - off);
48 ctx->cached_cq_tail++;
49 ctx->cqe_cached = &rings->cqes[off];
50 ctx->cqe_sentinel = ctx->cqe_cached + len;
52 return &rings->cqes[off << shift];
55 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
57 if (likely(ctx->cqe_cached < ctx->cqe_sentinel)) {
58 struct io_uring_cqe *cqe = ctx->cqe_cached;
60 if (ctx->flags & IORING_SETUP_CQE32) {
61 unsigned int off = ctx->cqe_cached - ctx->rings->cqes;
66 ctx->cached_cq_tail++;
71 return __io_get_cqe(ctx);
74 static inline bool __io_fill_cqe_req(struct io_ring_ctx *ctx,
77 struct io_uring_cqe *cqe;
79 if (!(ctx->flags & IORING_SETUP_CQE32)) {
80 trace_io_uring_complete(req->ctx, req, req->cqe.user_data,
81 req->cqe.res, req->cqe.flags, 0, 0);
84 * If we can't get a cq entry, userspace overflowed the
85 * submission (by quite a lot). Increment the overflow count in
88 cqe = io_get_cqe(ctx);
90 memcpy(cqe, &req->cqe, sizeof(*cqe));
94 return io_cqring_event_overflow(ctx, req->cqe.user_data,
95 req->cqe.res, req->cqe.flags,
98 u64 extra1 = 0, extra2 = 0;
100 if (req->flags & REQ_F_CQE32_INIT) {
101 extra1 = req->extra1;
102 extra2 = req->extra2;
105 trace_io_uring_complete(req->ctx, req, req->cqe.user_data,
106 req->cqe.res, req->cqe.flags, extra1, extra2);
109 * If we can't get a cq entry, userspace overflowed the
110 * submission (by quite a lot). Increment the overflow count in
113 cqe = io_get_cqe(ctx);
115 memcpy(cqe, &req->cqe, sizeof(struct io_uring_cqe));
116 WRITE_ONCE(cqe->big_cqe[0], extra1);
117 WRITE_ONCE(cqe->big_cqe[1], extra2);
121 return io_cqring_event_overflow(ctx, req->cqe.user_data,
122 req->cqe.res, req->cqe.flags,
127 static inline void req_set_fail(struct io_kiocb *req)
129 req->flags |= REQ_F_FAIL;
130 if (req->flags & REQ_F_CQE_SKIP) {
131 req->flags &= ~REQ_F_CQE_SKIP;
132 req->flags |= REQ_F_SKIP_LINK_CQES;
136 static inline void io_req_set_res(struct io_kiocb *req, s32 res, u32 cflags)
139 req->cqe.flags = cflags;
142 static inline bool req_has_async_data(struct io_kiocb *req)
144 return req->flags & REQ_F_ASYNC_DATA;
147 static inline void io_put_file(struct file *file)
153 static inline void io_ring_submit_unlock(struct io_ring_ctx *ctx,
154 unsigned issue_flags)
156 lockdep_assert_held(&ctx->uring_lock);
157 if (issue_flags & IO_URING_F_UNLOCKED)
158 mutex_unlock(&ctx->uring_lock);
161 static inline void io_ring_submit_lock(struct io_ring_ctx *ctx,
162 unsigned issue_flags)
165 * "Normal" inline submissions always hold the uring_lock, since we
166 * grab it from the system call. Same is true for the SQPOLL offload.
167 * The only exception is when we've detached the request and issue it
168 * from an async worker thread, grab the lock for that case.
170 if (issue_flags & IO_URING_F_UNLOCKED)
171 mutex_lock(&ctx->uring_lock);
172 lockdep_assert_held(&ctx->uring_lock);
175 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
177 /* order cqe stores with ring update */
178 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
181 static inline void io_cqring_wake(struct io_ring_ctx *ctx)
184 * wake_up_all() may seem excessive, but io_wake_function() and
185 * io_should_wake() handle the termination of the loop and only
186 * wake as many waiters as we need to.
188 if (wq_has_sleeper(&ctx->cq_wait))
189 wake_up_all(&ctx->cq_wait);
192 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
194 struct io_rings *r = ctx->rings;
196 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
199 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
201 struct io_rings *rings = ctx->rings;
203 /* make sure SQ entry isn't read before tail */
204 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
207 static inline bool io_run_task_work(void)
209 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || task_work_pending(current)) {
210 __set_current_state(TASK_RUNNING);
211 clear_notify_signal();
212 if (task_work_pending(current))
220 int io_run_task_work_sig(void);
221 void io_req_complete_failed(struct io_kiocb *req, s32 res);
222 void __io_req_complete(struct io_kiocb *req, unsigned issue_flags);
223 void io_req_complete_post(struct io_kiocb *req);
224 void __io_req_complete_post(struct io_kiocb *req);
225 bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
227 void io_cqring_ev_posted(struct io_ring_ctx *ctx);
228 void __io_commit_cqring_flush(struct io_ring_ctx *ctx);
230 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages);
232 struct file *io_file_get_normal(struct io_kiocb *req, int fd);
233 struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
234 unsigned issue_flags);
236 bool io_is_uring_fops(struct file *file);
237 bool io_alloc_async_data(struct io_kiocb *req);
238 void io_req_task_work_add(struct io_kiocb *req);
239 void io_req_task_prio_work_add(struct io_kiocb *req);
240 void io_req_tw_post_queue(struct io_kiocb *req, s32 res, u32 cflags);
241 void io_req_task_queue(struct io_kiocb *req);
242 void io_queue_iowq(struct io_kiocb *req, bool *dont_use);
243 void io_req_task_complete(struct io_kiocb *req, bool *locked);
244 void io_req_task_queue_fail(struct io_kiocb *req, int ret);
245 void io_req_task_submit(struct io_kiocb *req, bool *locked);
246 void tctx_task_work(struct callback_head *cb);
247 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
248 int io_uring_alloc_task_context(struct task_struct *task,
249 struct io_ring_ctx *ctx);
251 int io_poll_issue(struct io_kiocb *req, bool *locked);
252 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr);
253 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin);
254 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node);
255 int io_req_prep_async(struct io_kiocb *req);
257 struct io_wq_work *io_wq_free_work(struct io_wq_work *work);
258 void io_wq_submit_work(struct io_wq_work *work);
260 void io_free_req(struct io_kiocb *req);
261 void io_queue_next(struct io_kiocb *req);
263 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
266 #define io_for_each_link(pos, head) \
267 for (pos = (head); pos; pos = pos->link)