4 #include <linux/errno.h>
5 #include <linux/lockdep.h>
6 #include <linux/io_uring_types.h>
7 #include <uapi/linux/eventpoll.h>
10 #include "filetable.h"
12 #ifndef CREATE_TRACE_POINTS
13 #include <trace/events/io_uring.h>
18 IOU_ISSUE_SKIP_COMPLETE = -EIOCBQUEUED,
21 * Intended only when both IO_URING_F_MULTISHOT is passed
22 * to indicate to the poll runner that multishot should be
23 * removed and the result is set on req->cqe.res.
25 IOU_STOP_MULTISHOT = -ECANCELED,
28 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow);
29 bool io_req_cqe_overflow(struct io_kiocb *req);
30 int io_run_task_work_sig(struct io_ring_ctx *ctx);
31 int __io_run_local_work(struct io_ring_ctx *ctx, bool *locked);
32 int io_run_local_work(struct io_ring_ctx *ctx);
33 void io_req_complete_failed(struct io_kiocb *req, s32 res);
34 void __io_req_complete(struct io_kiocb *req, unsigned issue_flags);
35 void io_req_complete_post(struct io_kiocb *req);
36 void __io_req_complete_post(struct io_kiocb *req);
37 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
39 bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
41 void __io_commit_cqring_flush(struct io_ring_ctx *ctx);
43 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages);
45 struct file *io_file_get_normal(struct io_kiocb *req, int fd);
46 struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
47 unsigned issue_flags);
49 static inline bool io_req_ffs_set(struct io_kiocb *req)
51 return req->flags & REQ_F_FIXED_FILE;
54 bool io_is_uring_fops(struct file *file);
55 bool io_alloc_async_data(struct io_kiocb *req);
56 void io_req_task_work_add(struct io_kiocb *req);
57 void io_req_tw_post_queue(struct io_kiocb *req, s32 res, u32 cflags);
58 void io_req_task_queue(struct io_kiocb *req);
59 void io_queue_iowq(struct io_kiocb *req, bool *dont_use);
60 void io_req_task_complete(struct io_kiocb *req, bool *locked);
61 void io_req_task_queue_fail(struct io_kiocb *req, int ret);
62 void io_req_task_submit(struct io_kiocb *req, bool *locked);
63 void tctx_task_work(struct callback_head *cb);
64 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
65 int io_uring_alloc_task_context(struct task_struct *task,
66 struct io_ring_ctx *ctx);
68 int io_poll_issue(struct io_kiocb *req, bool *locked);
69 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr);
70 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin);
71 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node);
72 int io_req_prep_async(struct io_kiocb *req);
74 struct io_wq_work *io_wq_free_work(struct io_wq_work *work);
75 void io_wq_submit_work(struct io_wq_work *work);
77 void io_free_req(struct io_kiocb *req);
78 void io_queue_next(struct io_kiocb *req);
79 void __io_put_task(struct task_struct *task, int nr);
80 void io_task_refs_refill(struct io_uring_task *tctx);
81 bool __io_alloc_req_refill(struct io_ring_ctx *ctx);
83 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
86 #define io_for_each_link(pos, head) \
87 for (pos = (head); pos; pos = pos->link)
89 static inline void io_cq_lock(struct io_ring_ctx *ctx)
90 __acquires(ctx->completion_lock)
92 spin_lock(&ctx->completion_lock);
95 void io_cq_unlock_post(struct io_ring_ctx *ctx);
97 static inline struct io_uring_cqe *io_get_cqe_overflow(struct io_ring_ctx *ctx,
100 if (likely(ctx->cqe_cached < ctx->cqe_sentinel)) {
101 struct io_uring_cqe *cqe = ctx->cqe_cached;
103 ctx->cached_cq_tail++;
105 if (ctx->flags & IORING_SETUP_CQE32)
110 return __io_get_cqe(ctx, overflow);
113 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
115 return io_get_cqe_overflow(ctx, false);
118 static inline bool __io_fill_cqe_req(struct io_ring_ctx *ctx,
119 struct io_kiocb *req)
121 struct io_uring_cqe *cqe;
124 * If we can't get a cq entry, userspace overflowed the
125 * submission (by quite a lot). Increment the overflow count in
128 cqe = io_get_cqe(ctx);
130 return io_req_cqe_overflow(req);
132 trace_io_uring_complete(req->ctx, req, req->cqe.user_data,
133 req->cqe.res, req->cqe.flags,
134 (req->flags & REQ_F_CQE32_INIT) ? req->extra1 : 0,
135 (req->flags & REQ_F_CQE32_INIT) ? req->extra2 : 0);
137 memcpy(cqe, &req->cqe, sizeof(*cqe));
139 if (ctx->flags & IORING_SETUP_CQE32) {
140 u64 extra1 = 0, extra2 = 0;
142 if (req->flags & REQ_F_CQE32_INIT) {
143 extra1 = req->extra1;
144 extra2 = req->extra2;
147 WRITE_ONCE(cqe->big_cqe[0], extra1);
148 WRITE_ONCE(cqe->big_cqe[1], extra2);
153 static inline void req_set_fail(struct io_kiocb *req)
155 req->flags |= REQ_F_FAIL;
156 if (req->flags & REQ_F_CQE_SKIP) {
157 req->flags &= ~REQ_F_CQE_SKIP;
158 req->flags |= REQ_F_SKIP_LINK_CQES;
162 static inline void io_req_set_res(struct io_kiocb *req, s32 res, u32 cflags)
165 req->cqe.flags = cflags;
168 static inline bool req_has_async_data(struct io_kiocb *req)
170 return req->flags & REQ_F_ASYNC_DATA;
173 static inline void io_put_file(struct file *file)
179 static inline void io_ring_submit_unlock(struct io_ring_ctx *ctx,
180 unsigned issue_flags)
182 lockdep_assert_held(&ctx->uring_lock);
183 if (issue_flags & IO_URING_F_UNLOCKED)
184 mutex_unlock(&ctx->uring_lock);
187 static inline void io_ring_submit_lock(struct io_ring_ctx *ctx,
188 unsigned issue_flags)
191 * "Normal" inline submissions always hold the uring_lock, since we
192 * grab it from the system call. Same is true for the SQPOLL offload.
193 * The only exception is when we've detached the request and issue it
194 * from an async worker thread, grab the lock for that case.
196 if (issue_flags & IO_URING_F_UNLOCKED)
197 mutex_lock(&ctx->uring_lock);
198 lockdep_assert_held(&ctx->uring_lock);
201 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
203 /* order cqe stores with ring update */
204 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
207 /* requires smb_mb() prior, see wq_has_sleeper() */
208 static inline void __io_cqring_wake(struct io_ring_ctx *ctx)
211 * Trigger waitqueue handler on all waiters on our waitqueue. This
212 * won't necessarily wake up all the tasks, io_should_wake() will make
215 * Pass in EPOLLIN|EPOLL_URING_WAKE as the poll wakeup key. The latter
216 * set in the mask so that if we recurse back into our own poll
217 * waitqueue handlers, we know we have a dependency between eventfd or
218 * epoll and should terminate multishot poll at that point.
220 if (waitqueue_active(&ctx->cq_wait))
221 __wake_up(&ctx->cq_wait, TASK_NORMAL, 0,
222 poll_to_key(EPOLL_URING_WAKE | EPOLLIN));
225 static inline void io_cqring_wake(struct io_ring_ctx *ctx)
228 __io_cqring_wake(ctx);
231 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
233 struct io_rings *r = ctx->rings;
235 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
238 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
240 struct io_rings *rings = ctx->rings;
242 /* make sure SQ entry isn't read before tail */
243 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
246 static inline int io_run_task_work(void)
249 * Always check-and-clear the task_work notification signal. With how
250 * signaling works for task_work, we can find it set with nothing to
251 * run. We need to clear it for that case, like get_signal() does.
253 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
254 clear_notify_signal();
255 if (task_work_pending(current)) {
256 __set_current_state(TASK_RUNNING);
264 static inline bool io_task_work_pending(struct io_ring_ctx *ctx)
266 return test_thread_flag(TIF_NOTIFY_SIGNAL) ||
267 !wq_list_empty(&ctx->work_llist);
270 static inline int io_run_task_work_ctx(struct io_ring_ctx *ctx)
275 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
276 ret = io_run_local_work(ctx);
278 /* want to run this after in case more is added */
279 ret2 = io_run_task_work();
281 /* Try propagate error in favour of if tasks were run,
282 * but still make sure to run them if requested
290 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx)
295 if (llist_empty(&ctx->work_llist))
299 ret = __io_run_local_work(ctx, &locked);
300 /* shouldn't happen! */
301 if (WARN_ON_ONCE(!locked))
302 mutex_lock(&ctx->uring_lock);
306 static inline void io_tw_lock(struct io_ring_ctx *ctx, bool *locked)
309 mutex_lock(&ctx->uring_lock);
315 * Don't complete immediately but use deferred completion infrastructure.
316 * Protected by ->uring_lock and can only be used either with
317 * IO_URING_F_COMPLETE_DEFER or inside a tw handler holding the mutex.
319 static inline void io_req_complete_defer(struct io_kiocb *req)
320 __must_hold(&req->ctx->uring_lock)
322 struct io_submit_state *state = &req->ctx->submit_state;
324 lockdep_assert_held(&req->ctx->uring_lock);
326 wq_list_add_tail(&req->comp_list, &state->compl_reqs);
329 static inline void io_commit_cqring_flush(struct io_ring_ctx *ctx)
331 if (unlikely(ctx->off_timeout_used || ctx->drain_active || ctx->has_evfd))
332 __io_commit_cqring_flush(ctx);
335 /* must to be called somewhat shortly after putting a request */
336 static inline void io_put_task(struct task_struct *task, int nr)
338 if (likely(task == current))
339 task->io_uring->cached_refs += nr;
341 __io_put_task(task, nr);
344 static inline void io_get_task_refs(int nr)
346 struct io_uring_task *tctx = current->io_uring;
348 tctx->cached_refs -= nr;
349 if (unlikely(tctx->cached_refs < 0))
350 io_task_refs_refill(tctx);
353 static inline bool io_req_cache_empty(struct io_ring_ctx *ctx)
355 return !ctx->submit_state.free_list.next;
358 static inline bool io_alloc_req_refill(struct io_ring_ctx *ctx)
360 if (unlikely(io_req_cache_empty(ctx)))
361 return __io_alloc_req_refill(ctx);
365 static inline struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
367 struct io_wq_work_node *node;
369 node = wq_stack_extract(&ctx->submit_state.free_list);
370 return container_of(node, struct io_kiocb, comp_list);
373 static inline bool io_allowed_run_tw(struct io_ring_ctx *ctx)
375 return likely(!(ctx->flags & IORING_SETUP_DEFER_TASKRUN) ||
376 ctx->submitter_task == current);