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 void io_req_defer_failed(struct io_kiocb *req, s32 res);
32 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags);
33 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags);
34 bool io_aux_cqe(struct io_ring_ctx *ctx, bool defer, u64 user_data, s32 res, u32 cflags,
36 void __io_commit_cqring_flush(struct io_ring_ctx *ctx);
38 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages);
40 struct file *io_file_get_normal(struct io_kiocb *req, int fd);
41 struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
42 unsigned issue_flags);
44 static inline bool io_req_ffs_set(struct io_kiocb *req)
46 return req->flags & REQ_F_FIXED_FILE;
49 void __io_req_task_work_add(struct io_kiocb *req, bool allow_local);
50 bool io_is_uring_fops(struct file *file);
51 bool io_alloc_async_data(struct io_kiocb *req);
52 void io_req_task_queue(struct io_kiocb *req);
53 void io_queue_iowq(struct io_kiocb *req, bool *dont_use);
54 void io_req_task_complete(struct io_kiocb *req, bool *locked);
55 void io_req_task_queue_fail(struct io_kiocb *req, int ret);
56 void io_req_task_submit(struct io_kiocb *req, bool *locked);
57 void tctx_task_work(struct callback_head *cb);
58 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
59 int io_uring_alloc_task_context(struct task_struct *task,
60 struct io_ring_ctx *ctx);
62 int io_poll_issue(struct io_kiocb *req, bool *locked);
63 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr);
64 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin);
65 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node);
66 int io_req_prep_async(struct io_kiocb *req);
68 struct io_wq_work *io_wq_free_work(struct io_wq_work *work);
69 void io_wq_submit_work(struct io_wq_work *work);
71 void io_free_req(struct io_kiocb *req);
72 void io_queue_next(struct io_kiocb *req);
73 void io_task_refs_refill(struct io_uring_task *tctx);
74 bool __io_alloc_req_refill(struct io_ring_ctx *ctx);
76 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
79 #define io_lockdep_assert_cq_locked(ctx) \
81 if (ctx->flags & IORING_SETUP_IOPOLL) { \
82 lockdep_assert_held(&ctx->uring_lock); \
83 } else if (!ctx->task_complete) { \
84 lockdep_assert_held(&ctx->completion_lock); \
85 } else if (ctx->submitter_task->flags & PF_EXITING) { \
86 lockdep_assert(current_work()); \
88 lockdep_assert(current == ctx->submitter_task); \
92 static inline void io_req_task_work_add(struct io_kiocb *req)
94 __io_req_task_work_add(req, true);
97 #define io_for_each_link(pos, head) \
98 for (pos = (head); pos; pos = pos->link)
100 void io_cq_unlock_post(struct io_ring_ctx *ctx);
102 static inline struct io_uring_cqe *io_get_cqe_overflow(struct io_ring_ctx *ctx,
105 io_lockdep_assert_cq_locked(ctx);
107 if (likely(ctx->cqe_cached < ctx->cqe_sentinel)) {
108 struct io_uring_cqe *cqe = ctx->cqe_cached;
110 ctx->cached_cq_tail++;
112 if (ctx->flags & IORING_SETUP_CQE32)
117 return __io_get_cqe(ctx, overflow);
120 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
122 return io_get_cqe_overflow(ctx, false);
125 static inline bool __io_fill_cqe_req(struct io_ring_ctx *ctx,
126 struct io_kiocb *req)
128 struct io_uring_cqe *cqe;
131 * If we can't get a cq entry, userspace overflowed the
132 * submission (by quite a lot). Increment the overflow count in
135 cqe = io_get_cqe(ctx);
139 trace_io_uring_complete(req->ctx, req, req->cqe.user_data,
140 req->cqe.res, req->cqe.flags,
141 (req->flags & REQ_F_CQE32_INIT) ? req->extra1 : 0,
142 (req->flags & REQ_F_CQE32_INIT) ? req->extra2 : 0);
144 memcpy(cqe, &req->cqe, sizeof(*cqe));
146 if (ctx->flags & IORING_SETUP_CQE32) {
147 u64 extra1 = 0, extra2 = 0;
149 if (req->flags & REQ_F_CQE32_INIT) {
150 extra1 = req->extra1;
151 extra2 = req->extra2;
154 WRITE_ONCE(cqe->big_cqe[0], extra1);
155 WRITE_ONCE(cqe->big_cqe[1], extra2);
160 static inline bool io_fill_cqe_req(struct io_ring_ctx *ctx,
161 struct io_kiocb *req)
163 if (likely(__io_fill_cqe_req(ctx, req)))
165 return io_req_cqe_overflow(req);
168 static inline void req_set_fail(struct io_kiocb *req)
170 req->flags |= REQ_F_FAIL;
171 if (req->flags & REQ_F_CQE_SKIP) {
172 req->flags &= ~REQ_F_CQE_SKIP;
173 req->flags |= REQ_F_SKIP_LINK_CQES;
177 static inline void io_req_set_res(struct io_kiocb *req, s32 res, u32 cflags)
180 req->cqe.flags = cflags;
183 static inline bool req_has_async_data(struct io_kiocb *req)
185 return req->flags & REQ_F_ASYNC_DATA;
188 static inline void io_put_file(struct file *file)
194 static inline void io_ring_submit_unlock(struct io_ring_ctx *ctx,
195 unsigned issue_flags)
197 lockdep_assert_held(&ctx->uring_lock);
198 if (issue_flags & IO_URING_F_UNLOCKED)
199 mutex_unlock(&ctx->uring_lock);
202 static inline void io_ring_submit_lock(struct io_ring_ctx *ctx,
203 unsigned issue_flags)
206 * "Normal" inline submissions always hold the uring_lock, since we
207 * grab it from the system call. Same is true for the SQPOLL offload.
208 * The only exception is when we've detached the request and issue it
209 * from an async worker thread, grab the lock for that case.
211 if (issue_flags & IO_URING_F_UNLOCKED)
212 mutex_lock(&ctx->uring_lock);
213 lockdep_assert_held(&ctx->uring_lock);
216 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
218 /* order cqe stores with ring update */
219 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
222 static inline void io_poll_wq_wake(struct io_ring_ctx *ctx)
224 if (wq_has_sleeper(&ctx->poll_wq))
225 __wake_up(&ctx->poll_wq, TASK_NORMAL, 0,
226 poll_to_key(EPOLL_URING_WAKE | EPOLLIN));
229 /* requires smb_mb() prior, see wq_has_sleeper() */
230 static inline void __io_cqring_wake(struct io_ring_ctx *ctx)
233 * Trigger waitqueue handler on all waiters on our waitqueue. This
234 * won't necessarily wake up all the tasks, io_should_wake() will make
237 * Pass in EPOLLIN|EPOLL_URING_WAKE as the poll wakeup key. The latter
238 * set in the mask so that if we recurse back into our own poll
239 * waitqueue handlers, we know we have a dependency between eventfd or
240 * epoll and should terminate multishot poll at that point.
242 if (waitqueue_active(&ctx->cq_wait))
243 __wake_up(&ctx->cq_wait, TASK_NORMAL, 0,
244 poll_to_key(EPOLL_URING_WAKE | EPOLLIN));
247 static inline void io_cqring_wake(struct io_ring_ctx *ctx)
250 __io_cqring_wake(ctx);
253 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
255 struct io_rings *r = ctx->rings;
257 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
260 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
262 struct io_rings *rings = ctx->rings;
264 /* make sure SQ entry isn't read before tail */
265 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
268 static inline int io_run_task_work(void)
271 * Always check-and-clear the task_work notification signal. With how
272 * signaling works for task_work, we can find it set with nothing to
273 * run. We need to clear it for that case, like get_signal() does.
275 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
276 clear_notify_signal();
277 if (task_work_pending(current)) {
278 __set_current_state(TASK_RUNNING);
286 static inline bool io_task_work_pending(struct io_ring_ctx *ctx)
288 return task_work_pending(current) || !wq_list_empty(&ctx->work_llist);
291 static inline void io_tw_lock(struct io_ring_ctx *ctx, bool *locked)
294 mutex_lock(&ctx->uring_lock);
300 * Don't complete immediately but use deferred completion infrastructure.
301 * Protected by ->uring_lock and can only be used either with
302 * IO_URING_F_COMPLETE_DEFER or inside a tw handler holding the mutex.
304 static inline void io_req_complete_defer(struct io_kiocb *req)
305 __must_hold(&req->ctx->uring_lock)
307 struct io_submit_state *state = &req->ctx->submit_state;
309 lockdep_assert_held(&req->ctx->uring_lock);
311 wq_list_add_tail(&req->comp_list, &state->compl_reqs);
314 static inline void io_commit_cqring_flush(struct io_ring_ctx *ctx)
316 if (unlikely(ctx->off_timeout_used || ctx->drain_active ||
317 ctx->has_evfd || ctx->poll_activated))
318 __io_commit_cqring_flush(ctx);
321 static inline void io_get_task_refs(int nr)
323 struct io_uring_task *tctx = current->io_uring;
325 tctx->cached_refs -= nr;
326 if (unlikely(tctx->cached_refs < 0))
327 io_task_refs_refill(tctx);
330 static inline bool io_req_cache_empty(struct io_ring_ctx *ctx)
332 return !ctx->submit_state.free_list.next;
335 static inline bool io_alloc_req_refill(struct io_ring_ctx *ctx)
337 if (unlikely(io_req_cache_empty(ctx)))
338 return __io_alloc_req_refill(ctx);
342 static inline struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
344 struct io_wq_work_node *node;
346 node = wq_stack_extract(&ctx->submit_state.free_list);
347 return container_of(node, struct io_kiocb, comp_list);
350 static inline bool io_allowed_defer_tw_run(struct io_ring_ctx *ctx)
352 return likely(ctx->submitter_task == current);
355 static inline bool io_allowed_run_tw(struct io_ring_ctx *ctx)
357 return likely(!(ctx->flags & IORING_SETUP_DEFER_TASKRUN) ||
358 ctx->submitter_task == current);
361 static inline void io_req_queue_tw_complete(struct io_kiocb *req, s32 res)
363 io_req_set_res(req, res, 0);
364 req->io_task_work.func = io_req_task_complete;
365 io_req_task_work_add(req);