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_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 static inline void io_req_task_work_add(struct io_kiocb *req)
87 __io_req_task_work_add(req, true);
90 #define io_for_each_link(pos, head) \
91 for (pos = (head); pos; pos = pos->link)
93 static inline void io_cq_lock(struct io_ring_ctx *ctx)
94 __acquires(ctx->completion_lock)
96 spin_lock(&ctx->completion_lock);
99 void io_cq_unlock_post(struct io_ring_ctx *ctx);
101 static inline struct io_uring_cqe *io_get_cqe_overflow(struct io_ring_ctx *ctx,
104 if (likely(ctx->cqe_cached < ctx->cqe_sentinel)) {
105 struct io_uring_cqe *cqe = ctx->cqe_cached;
107 ctx->cached_cq_tail++;
109 if (ctx->flags & IORING_SETUP_CQE32)
114 return __io_get_cqe(ctx, overflow);
117 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
119 return io_get_cqe_overflow(ctx, false);
122 static inline bool __io_fill_cqe_req(struct io_ring_ctx *ctx,
123 struct io_kiocb *req)
125 struct io_uring_cqe *cqe;
128 * If we can't get a cq entry, userspace overflowed the
129 * submission (by quite a lot). Increment the overflow count in
132 cqe = io_get_cqe(ctx);
134 return io_req_cqe_overflow(req);
136 trace_io_uring_complete(req->ctx, req, req->cqe.user_data,
137 req->cqe.res, req->cqe.flags,
138 (req->flags & REQ_F_CQE32_INIT) ? req->extra1 : 0,
139 (req->flags & REQ_F_CQE32_INIT) ? req->extra2 : 0);
141 memcpy(cqe, &req->cqe, sizeof(*cqe));
143 if (ctx->flags & IORING_SETUP_CQE32) {
144 u64 extra1 = 0, extra2 = 0;
146 if (req->flags & REQ_F_CQE32_INIT) {
147 extra1 = req->extra1;
148 extra2 = req->extra2;
151 WRITE_ONCE(cqe->big_cqe[0], extra1);
152 WRITE_ONCE(cqe->big_cqe[1], extra2);
157 static inline void req_set_fail(struct io_kiocb *req)
159 req->flags |= REQ_F_FAIL;
160 if (req->flags & REQ_F_CQE_SKIP) {
161 req->flags &= ~REQ_F_CQE_SKIP;
162 req->flags |= REQ_F_SKIP_LINK_CQES;
166 static inline void io_req_set_res(struct io_kiocb *req, s32 res, u32 cflags)
169 req->cqe.flags = cflags;
172 static inline bool req_has_async_data(struct io_kiocb *req)
174 return req->flags & REQ_F_ASYNC_DATA;
177 static inline void io_put_file(struct file *file)
183 static inline void io_ring_submit_unlock(struct io_ring_ctx *ctx,
184 unsigned issue_flags)
186 lockdep_assert_held(&ctx->uring_lock);
187 if (issue_flags & IO_URING_F_UNLOCKED)
188 mutex_unlock(&ctx->uring_lock);
191 static inline void io_ring_submit_lock(struct io_ring_ctx *ctx,
192 unsigned issue_flags)
195 * "Normal" inline submissions always hold the uring_lock, since we
196 * grab it from the system call. Same is true for the SQPOLL offload.
197 * The only exception is when we've detached the request and issue it
198 * from an async worker thread, grab the lock for that case.
200 if (issue_flags & IO_URING_F_UNLOCKED)
201 mutex_lock(&ctx->uring_lock);
202 lockdep_assert_held(&ctx->uring_lock);
205 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
207 /* order cqe stores with ring update */
208 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
211 /* requires smb_mb() prior, see wq_has_sleeper() */
212 static inline void __io_cqring_wake(struct io_ring_ctx *ctx)
215 * Trigger waitqueue handler on all waiters on our waitqueue. This
216 * won't necessarily wake up all the tasks, io_should_wake() will make
219 * Pass in EPOLLIN|EPOLL_URING_WAKE as the poll wakeup key. The latter
220 * set in the mask so that if we recurse back into our own poll
221 * waitqueue handlers, we know we have a dependency between eventfd or
222 * epoll and should terminate multishot poll at that point.
224 if (waitqueue_active(&ctx->cq_wait))
225 __wake_up(&ctx->cq_wait, TASK_NORMAL, 0,
226 poll_to_key(EPOLL_URING_WAKE | EPOLLIN));
229 static inline void io_cqring_wake(struct io_ring_ctx *ctx)
232 __io_cqring_wake(ctx);
235 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
237 struct io_rings *r = ctx->rings;
239 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
242 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
244 struct io_rings *rings = ctx->rings;
246 /* make sure SQ entry isn't read before tail */
247 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
250 static inline int io_run_task_work(void)
253 * Always check-and-clear the task_work notification signal. With how
254 * signaling works for task_work, we can find it set with nothing to
255 * run. We need to clear it for that case, like get_signal() does.
257 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
258 clear_notify_signal();
259 if (task_work_pending(current)) {
260 __set_current_state(TASK_RUNNING);
268 static inline bool io_task_work_pending(struct io_ring_ctx *ctx)
270 return test_thread_flag(TIF_NOTIFY_SIGNAL) ||
271 !wq_list_empty(&ctx->work_llist);
274 static inline int io_run_task_work_ctx(struct io_ring_ctx *ctx)
279 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
280 ret = io_run_local_work(ctx);
282 /* want to run this after in case more is added */
283 ret2 = io_run_task_work();
285 /* Try propagate error in favour of if tasks were run,
286 * but still make sure to run them if requested
294 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx)
299 if (llist_empty(&ctx->work_llist))
303 ret = __io_run_local_work(ctx, &locked);
304 /* shouldn't happen! */
305 if (WARN_ON_ONCE(!locked))
306 mutex_lock(&ctx->uring_lock);
310 static inline void io_tw_lock(struct io_ring_ctx *ctx, bool *locked)
313 mutex_lock(&ctx->uring_lock);
319 * Don't complete immediately but use deferred completion infrastructure.
320 * Protected by ->uring_lock and can only be used either with
321 * IO_URING_F_COMPLETE_DEFER or inside a tw handler holding the mutex.
323 static inline void io_req_complete_defer(struct io_kiocb *req)
324 __must_hold(&req->ctx->uring_lock)
326 struct io_submit_state *state = &req->ctx->submit_state;
328 lockdep_assert_held(&req->ctx->uring_lock);
330 wq_list_add_tail(&req->comp_list, &state->compl_reqs);
333 static inline void io_commit_cqring_flush(struct io_ring_ctx *ctx)
335 if (unlikely(ctx->off_timeout_used || ctx->drain_active || ctx->has_evfd))
336 __io_commit_cqring_flush(ctx);
339 /* must to be called somewhat shortly after putting a request */
340 static inline void io_put_task(struct task_struct *task, int nr)
342 if (likely(task == current))
343 task->io_uring->cached_refs += nr;
345 __io_put_task(task, nr);
348 static inline void io_get_task_refs(int nr)
350 struct io_uring_task *tctx = current->io_uring;
352 tctx->cached_refs -= nr;
353 if (unlikely(tctx->cached_refs < 0))
354 io_task_refs_refill(tctx);
357 static inline bool io_req_cache_empty(struct io_ring_ctx *ctx)
359 return !ctx->submit_state.free_list.next;
362 static inline bool io_alloc_req_refill(struct io_ring_ctx *ctx)
364 if (unlikely(io_req_cache_empty(ctx)))
365 return __io_alloc_req_refill(ctx);
369 static inline struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
371 struct io_wq_work_node *node;
373 node = wq_stack_extract(&ctx->submit_state.free_list);
374 return container_of(node, struct io_kiocb, comp_list);
377 static inline bool io_allowed_run_tw(struct io_ring_ctx *ctx)
379 return likely(!(ctx->flags & IORING_SETUP_DEFER_TASKRUN) ||
380 ctx->submitter_task == current);