4 #include <linux/errno.h>
5 #include <linux/lockdep.h>
6 #include <linux/io_uring_types.h>
11 #ifndef CREATE_TRACE_POINTS
12 #include <trace/events/io_uring.h>
17 IOU_ISSUE_SKIP_COMPLETE = -EIOCBQUEUED,
20 * Intended only when both IO_URING_F_MULTISHOT is passed
21 * to indicate to the poll runner that multishot should be
22 * removed and the result is set on req->cqe.res.
24 IOU_STOP_MULTISHOT = -ECANCELED,
27 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow);
28 bool io_req_cqe_overflow(struct io_kiocb *req);
29 int io_run_task_work_sig(struct io_ring_ctx *ctx);
30 int __io_run_local_work(struct io_ring_ctx *ctx, bool *locked);
31 int io_run_local_work(struct io_ring_ctx *ctx);
32 void io_req_complete_failed(struct io_kiocb *req, s32 res);
33 void __io_req_complete(struct io_kiocb *req, unsigned issue_flags);
34 void io_req_complete_post(struct io_kiocb *req);
35 void __io_req_complete_post(struct io_kiocb *req);
36 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
38 bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
40 void __io_commit_cqring_flush(struct io_ring_ctx *ctx);
42 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages);
44 struct file *io_file_get_normal(struct io_kiocb *req, int fd);
45 struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
46 unsigned issue_flags);
48 static inline bool io_req_ffs_set(struct io_kiocb *req)
50 return req->flags & REQ_F_FIXED_FILE;
53 bool io_is_uring_fops(struct file *file);
54 bool io_alloc_async_data(struct io_kiocb *req);
55 void io_req_task_work_add(struct io_kiocb *req);
56 void io_req_tw_post_queue(struct io_kiocb *req, s32 res, u32 cflags);
57 void io_req_task_queue(struct io_kiocb *req);
58 void io_queue_iowq(struct io_kiocb *req, bool *dont_use);
59 void io_req_task_complete(struct io_kiocb *req, bool *locked);
60 void io_req_task_queue_fail(struct io_kiocb *req, int ret);
61 void io_req_task_submit(struct io_kiocb *req, bool *locked);
62 void tctx_task_work(struct callback_head *cb);
63 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
64 int io_uring_alloc_task_context(struct task_struct *task,
65 struct io_ring_ctx *ctx);
67 int io_poll_issue(struct io_kiocb *req, bool *locked);
68 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr);
69 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin);
70 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node);
71 int io_req_prep_async(struct io_kiocb *req);
73 struct io_wq_work *io_wq_free_work(struct io_wq_work *work);
74 void io_wq_submit_work(struct io_wq_work *work);
76 void io_free_req(struct io_kiocb *req);
77 void io_queue_next(struct io_kiocb *req);
78 void __io_put_task(struct task_struct *task, int nr);
79 void io_task_refs_refill(struct io_uring_task *tctx);
80 bool __io_alloc_req_refill(struct io_ring_ctx *ctx);
82 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
85 #define io_for_each_link(pos, head) \
86 for (pos = (head); pos; pos = pos->link)
88 static inline void io_cq_lock(struct io_ring_ctx *ctx)
89 __acquires(ctx->completion_lock)
91 spin_lock(&ctx->completion_lock);
94 void io_cq_unlock_post(struct io_ring_ctx *ctx);
96 static inline struct io_uring_cqe *io_get_cqe_overflow(struct io_ring_ctx *ctx,
99 if (likely(ctx->cqe_cached < ctx->cqe_sentinel)) {
100 struct io_uring_cqe *cqe = ctx->cqe_cached;
102 ctx->cached_cq_tail++;
104 if (ctx->flags & IORING_SETUP_CQE32)
109 return __io_get_cqe(ctx, overflow);
112 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
114 return io_get_cqe_overflow(ctx, false);
117 static inline bool __io_fill_cqe_req(struct io_ring_ctx *ctx,
118 struct io_kiocb *req)
120 struct io_uring_cqe *cqe;
123 * If we can't get a cq entry, userspace overflowed the
124 * submission (by quite a lot). Increment the overflow count in
127 cqe = io_get_cqe(ctx);
129 return io_req_cqe_overflow(req);
131 trace_io_uring_complete(req->ctx, req, req->cqe.user_data,
132 req->cqe.res, req->cqe.flags,
133 (req->flags & REQ_F_CQE32_INIT) ? req->extra1 : 0,
134 (req->flags & REQ_F_CQE32_INIT) ? req->extra2 : 0);
136 memcpy(cqe, &req->cqe, sizeof(*cqe));
138 if (ctx->flags & IORING_SETUP_CQE32) {
139 u64 extra1 = 0, extra2 = 0;
141 if (req->flags & REQ_F_CQE32_INIT) {
142 extra1 = req->extra1;
143 extra2 = req->extra2;
146 WRITE_ONCE(cqe->big_cqe[0], extra1);
147 WRITE_ONCE(cqe->big_cqe[1], extra2);
152 static inline void req_set_fail(struct io_kiocb *req)
154 req->flags |= REQ_F_FAIL;
155 if (req->flags & REQ_F_CQE_SKIP) {
156 req->flags &= ~REQ_F_CQE_SKIP;
157 req->flags |= REQ_F_SKIP_LINK_CQES;
161 static inline void io_req_set_res(struct io_kiocb *req, s32 res, u32 cflags)
164 req->cqe.flags = cflags;
167 static inline bool req_has_async_data(struct io_kiocb *req)
169 return req->flags & REQ_F_ASYNC_DATA;
172 static inline void io_put_file(struct file *file)
178 static inline void io_ring_submit_unlock(struct io_ring_ctx *ctx,
179 unsigned issue_flags)
181 lockdep_assert_held(&ctx->uring_lock);
182 if (issue_flags & IO_URING_F_UNLOCKED)
183 mutex_unlock(&ctx->uring_lock);
186 static inline void io_ring_submit_lock(struct io_ring_ctx *ctx,
187 unsigned issue_flags)
190 * "Normal" inline submissions always hold the uring_lock, since we
191 * grab it from the system call. Same is true for the SQPOLL offload.
192 * The only exception is when we've detached the request and issue it
193 * from an async worker thread, grab the lock for that case.
195 if (issue_flags & IO_URING_F_UNLOCKED)
196 mutex_lock(&ctx->uring_lock);
197 lockdep_assert_held(&ctx->uring_lock);
200 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
202 /* order cqe stores with ring update */
203 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
206 /* requires smb_mb() prior, see wq_has_sleeper() */
207 static inline void __io_cqring_wake(struct io_ring_ctx *ctx)
210 * wake_up_all() may seem excessive, but io_wake_function() and
211 * io_should_wake() handle the termination of the loop and only
212 * wake as many waiters as we need to.
214 if (waitqueue_active(&ctx->cq_wait))
215 wake_up_all(&ctx->cq_wait);
218 static inline void io_cqring_wake(struct io_ring_ctx *ctx)
221 __io_cqring_wake(ctx);
224 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
226 struct io_rings *r = ctx->rings;
228 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
231 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
233 struct io_rings *rings = ctx->rings;
235 /* make sure SQ entry isn't read before tail */
236 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
239 static inline int io_run_task_work(void)
242 * Always check-and-clear the task_work notification signal. With how
243 * signaling works for task_work, we can find it set with nothing to
244 * run. We need to clear it for that case, like get_signal() does.
246 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
247 clear_notify_signal();
248 if (task_work_pending(current)) {
249 __set_current_state(TASK_RUNNING);
257 static inline bool io_task_work_pending(struct io_ring_ctx *ctx)
259 return test_thread_flag(TIF_NOTIFY_SIGNAL) ||
260 !wq_list_empty(&ctx->work_llist);
263 static inline int io_run_task_work_ctx(struct io_ring_ctx *ctx)
268 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
269 ret = io_run_local_work(ctx);
271 /* want to run this after in case more is added */
272 ret2 = io_run_task_work();
274 /* Try propagate error in favour of if tasks were run,
275 * but still make sure to run them if requested
283 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx)
288 if (llist_empty(&ctx->work_llist))
292 ret = __io_run_local_work(ctx, &locked);
293 /* shouldn't happen! */
294 if (WARN_ON_ONCE(!locked))
295 mutex_lock(&ctx->uring_lock);
299 static inline void io_tw_lock(struct io_ring_ctx *ctx, bool *locked)
302 mutex_lock(&ctx->uring_lock);
308 * Don't complete immediately but use deferred completion infrastructure.
309 * Protected by ->uring_lock and can only be used either with
310 * IO_URING_F_COMPLETE_DEFER or inside a tw handler holding the mutex.
312 static inline void io_req_complete_defer(struct io_kiocb *req)
313 __must_hold(&req->ctx->uring_lock)
315 struct io_submit_state *state = &req->ctx->submit_state;
317 lockdep_assert_held(&req->ctx->uring_lock);
319 wq_list_add_tail(&req->comp_list, &state->compl_reqs);
322 static inline void io_commit_cqring_flush(struct io_ring_ctx *ctx)
324 if (unlikely(ctx->off_timeout_used || ctx->drain_active || ctx->has_evfd))
325 __io_commit_cqring_flush(ctx);
328 /* must to be called somewhat shortly after putting a request */
329 static inline void io_put_task(struct task_struct *task, int nr)
331 if (likely(task == current))
332 task->io_uring->cached_refs += nr;
334 __io_put_task(task, nr);
337 static inline void io_get_task_refs(int nr)
339 struct io_uring_task *tctx = current->io_uring;
341 tctx->cached_refs -= nr;
342 if (unlikely(tctx->cached_refs < 0))
343 io_task_refs_refill(tctx);
346 static inline bool io_req_cache_empty(struct io_ring_ctx *ctx)
348 return !ctx->submit_state.free_list.next;
351 static inline bool io_alloc_req_refill(struct io_ring_ctx *ctx)
353 if (unlikely(io_req_cache_empty(ctx)))
354 return __io_alloc_req_refill(ctx);
358 static inline struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
360 struct io_wq_work_node *node;
362 node = wq_stack_extract(&ctx->submit_state.free_list);
363 return container_of(node, struct io_kiocb, comp_list);
366 static inline bool io_allowed_run_tw(struct io_ring_ctx *ctx)
368 return likely(!(ctx->flags & IORING_SETUP_DEFER_TASKRUN) ||
369 ctx->submitter_task == current);