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 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 void __io_req_task_work_add(struct io_kiocb *req, bool allow_local);
54 bool io_is_uring_fops(struct file *file);
55 bool io_alloc_async_data(struct io_kiocb *req);
56 void io_req_task_queue(struct io_kiocb *req);
57 void io_queue_iowq(struct io_kiocb *req, bool *dont_use);
58 void io_req_task_complete(struct io_kiocb *req, bool *locked);
59 void io_req_task_queue_fail(struct io_kiocb *req, int ret);
60 void io_req_task_submit(struct io_kiocb *req, bool *locked);
61 void tctx_task_work(struct callback_head *cb);
62 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
63 int io_uring_alloc_task_context(struct task_struct *task,
64 struct io_ring_ctx *ctx);
66 int io_poll_issue(struct io_kiocb *req, bool *locked);
67 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr);
68 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin);
69 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node);
70 int io_req_prep_async(struct io_kiocb *req);
72 struct io_wq_work *io_wq_free_work(struct io_wq_work *work);
73 void io_wq_submit_work(struct io_wq_work *work);
75 void io_free_req(struct io_kiocb *req);
76 void io_queue_next(struct io_kiocb *req);
77 void __io_put_task(struct task_struct *task, int nr);
78 void io_task_refs_refill(struct io_uring_task *tctx);
79 bool __io_alloc_req_refill(struct io_ring_ctx *ctx);
81 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
84 static inline void io_req_task_work_add(struct io_kiocb *req)
86 __io_req_task_work_add(req, true);
89 #define io_for_each_link(pos, head) \
90 for (pos = (head); pos; pos = pos->link)
92 static inline void io_cq_lock(struct io_ring_ctx *ctx)
93 __acquires(ctx->completion_lock)
95 spin_lock(&ctx->completion_lock);
98 void io_cq_unlock_post(struct io_ring_ctx *ctx);
100 static inline struct io_uring_cqe *io_get_cqe_overflow(struct io_ring_ctx *ctx,
103 if (likely(ctx->cqe_cached < ctx->cqe_sentinel)) {
104 struct io_uring_cqe *cqe = ctx->cqe_cached;
106 ctx->cached_cq_tail++;
108 if (ctx->flags & IORING_SETUP_CQE32)
113 return __io_get_cqe(ctx, overflow);
116 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
118 return io_get_cqe_overflow(ctx, false);
121 static inline bool __io_fill_cqe_req(struct io_ring_ctx *ctx,
122 struct io_kiocb *req)
124 struct io_uring_cqe *cqe;
127 * If we can't get a cq entry, userspace overflowed the
128 * submission (by quite a lot). Increment the overflow count in
131 cqe = io_get_cqe(ctx);
133 return io_req_cqe_overflow(req);
135 trace_io_uring_complete(req->ctx, req, req->cqe.user_data,
136 req->cqe.res, req->cqe.flags,
137 (req->flags & REQ_F_CQE32_INIT) ? req->extra1 : 0,
138 (req->flags & REQ_F_CQE32_INIT) ? req->extra2 : 0);
140 memcpy(cqe, &req->cqe, sizeof(*cqe));
142 if (ctx->flags & IORING_SETUP_CQE32) {
143 u64 extra1 = 0, extra2 = 0;
145 if (req->flags & REQ_F_CQE32_INIT) {
146 extra1 = req->extra1;
147 extra2 = req->extra2;
150 WRITE_ONCE(cqe->big_cqe[0], extra1);
151 WRITE_ONCE(cqe->big_cqe[1], extra2);
156 static inline void req_set_fail(struct io_kiocb *req)
158 req->flags |= REQ_F_FAIL;
159 if (req->flags & REQ_F_CQE_SKIP) {
160 req->flags &= ~REQ_F_CQE_SKIP;
161 req->flags |= REQ_F_SKIP_LINK_CQES;
165 static inline void io_req_set_res(struct io_kiocb *req, s32 res, u32 cflags)
168 req->cqe.flags = cflags;
171 static inline bool req_has_async_data(struct io_kiocb *req)
173 return req->flags & REQ_F_ASYNC_DATA;
176 static inline void io_put_file(struct file *file)
182 static inline void io_ring_submit_unlock(struct io_ring_ctx *ctx,
183 unsigned issue_flags)
185 lockdep_assert_held(&ctx->uring_lock);
186 if (issue_flags & IO_URING_F_UNLOCKED)
187 mutex_unlock(&ctx->uring_lock);
190 static inline void io_ring_submit_lock(struct io_ring_ctx *ctx,
191 unsigned issue_flags)
194 * "Normal" inline submissions always hold the uring_lock, since we
195 * grab it from the system call. Same is true for the SQPOLL offload.
196 * The only exception is when we've detached the request and issue it
197 * from an async worker thread, grab the lock for that case.
199 if (issue_flags & IO_URING_F_UNLOCKED)
200 mutex_lock(&ctx->uring_lock);
201 lockdep_assert_held(&ctx->uring_lock);
204 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
206 /* order cqe stores with ring update */
207 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
210 /* requires smb_mb() prior, see wq_has_sleeper() */
211 static inline void __io_cqring_wake(struct io_ring_ctx *ctx)
214 * Trigger waitqueue handler on all waiters on our waitqueue. This
215 * won't necessarily wake up all the tasks, io_should_wake() will make
218 * Pass in EPOLLIN|EPOLL_URING_WAKE as the poll wakeup key. The latter
219 * set in the mask so that if we recurse back into our own poll
220 * waitqueue handlers, we know we have a dependency between eventfd or
221 * epoll and should terminate multishot poll at that point.
223 if (waitqueue_active(&ctx->cq_wait))
224 __wake_up(&ctx->cq_wait, TASK_NORMAL, 0,
225 poll_to_key(EPOLL_URING_WAKE | EPOLLIN));
228 static inline void io_cqring_wake(struct io_ring_ctx *ctx)
231 __io_cqring_wake(ctx);
234 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
236 struct io_rings *r = ctx->rings;
238 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
241 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
243 struct io_rings *rings = ctx->rings;
245 /* make sure SQ entry isn't read before tail */
246 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
249 static inline int io_run_task_work(void)
252 * Always check-and-clear the task_work notification signal. With how
253 * signaling works for task_work, we can find it set with nothing to
254 * run. We need to clear it for that case, like get_signal() does.
256 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
257 clear_notify_signal();
258 if (task_work_pending(current)) {
259 __set_current_state(TASK_RUNNING);
267 static inline bool io_task_work_pending(struct io_ring_ctx *ctx)
269 return test_thread_flag(TIF_NOTIFY_SIGNAL) ||
270 !wq_list_empty(&ctx->work_llist);
273 static inline int io_run_task_work_ctx(struct io_ring_ctx *ctx)
278 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
279 ret = io_run_local_work(ctx);
281 /* want to run this after in case more is added */
282 ret2 = io_run_task_work();
284 /* Try propagate error in favour of if tasks were run,
285 * but still make sure to run them if requested
293 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx)
298 if (llist_empty(&ctx->work_llist))
302 ret = __io_run_local_work(ctx, &locked);
303 /* shouldn't happen! */
304 if (WARN_ON_ONCE(!locked))
305 mutex_lock(&ctx->uring_lock);
309 static inline void io_tw_lock(struct io_ring_ctx *ctx, bool *locked)
312 mutex_lock(&ctx->uring_lock);
318 * Don't complete immediately but use deferred completion infrastructure.
319 * Protected by ->uring_lock and can only be used either with
320 * IO_URING_F_COMPLETE_DEFER or inside a tw handler holding the mutex.
322 static inline void io_req_complete_defer(struct io_kiocb *req)
323 __must_hold(&req->ctx->uring_lock)
325 struct io_submit_state *state = &req->ctx->submit_state;
327 lockdep_assert_held(&req->ctx->uring_lock);
329 wq_list_add_tail(&req->comp_list, &state->compl_reqs);
332 static inline void io_commit_cqring_flush(struct io_ring_ctx *ctx)
334 if (unlikely(ctx->off_timeout_used || ctx->drain_active || ctx->has_evfd))
335 __io_commit_cqring_flush(ctx);
338 /* must to be called somewhat shortly after putting a request */
339 static inline void io_put_task(struct task_struct *task, int nr)
341 if (likely(task == current))
342 task->io_uring->cached_refs += nr;
344 __io_put_task(task, nr);
347 static inline void io_get_task_refs(int nr)
349 struct io_uring_task *tctx = current->io_uring;
351 tctx->cached_refs -= nr;
352 if (unlikely(tctx->cached_refs < 0))
353 io_task_refs_refill(tctx);
356 static inline bool io_req_cache_empty(struct io_ring_ctx *ctx)
358 return !ctx->submit_state.free_list.next;
361 static inline bool io_alloc_req_refill(struct io_ring_ctx *ctx)
363 if (unlikely(io_req_cache_empty(ctx)))
364 return __io_alloc_req_refill(ctx);
368 static inline struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
370 struct io_wq_work_node *node;
372 node = wq_stack_extract(&ctx->submit_state.free_list);
373 return container_of(node, struct io_kiocb, comp_list);
376 static inline bool io_allowed_run_tw(struct io_ring_ctx *ctx)
378 return likely(!(ctx->flags & IORING_SETUP_DEFER_TASKRUN) ||
379 ctx->submitter_task == current);
382 static inline void io_req_queue_tw_complete(struct io_kiocb *req, s32 res)
384 io_req_set_res(req, res, 0);
385 req->io_task_work.func = io_req_task_complete;
386 io_req_task_work_add(req);