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_defer_failed(struct io_kiocb
*req
, s32 res
);
34 void io_req_complete_post(struct io_kiocb
*req
, unsigned issue_flags
);
35 bool io_post_aux_cqe(struct io_ring_ctx
*ctx
, u64 user_data
, s32 res
, u32 cflags
);
36 bool io_aux_cqe(struct io_ring_ctx
*ctx
, bool defer
, u64 user_data
, s32 res
, u32 cflags
,
38 void __io_commit_cqring_flush(struct io_ring_ctx
*ctx
);
40 struct page
**io_pin_pages(unsigned long ubuf
, unsigned long len
, int *npages
);
42 struct file
*io_file_get_normal(struct io_kiocb
*req
, int fd
);
43 struct file
*io_file_get_fixed(struct io_kiocb
*req
, int fd
,
44 unsigned issue_flags
);
46 static inline bool io_req_ffs_set(struct io_kiocb
*req
)
48 return req
->flags
& REQ_F_FIXED_FILE
;
51 void __io_req_task_work_add(struct io_kiocb
*req
, bool allow_local
);
52 bool io_is_uring_fops(struct file
*file
);
53 bool io_alloc_async_data(struct io_kiocb
*req
);
54 void io_req_task_queue(struct io_kiocb
*req
);
55 void io_queue_iowq(struct io_kiocb
*req
, bool *dont_use
);
56 void io_req_task_complete(struct io_kiocb
*req
, bool *locked
);
57 void io_req_task_queue_fail(struct io_kiocb
*req
, int ret
);
58 void io_req_task_submit(struct io_kiocb
*req
, bool *locked
);
59 void tctx_task_work(struct callback_head
*cb
);
60 __cold
void io_uring_cancel_generic(bool cancel_all
, struct io_sq_data
*sqd
);
61 int io_uring_alloc_task_context(struct task_struct
*task
,
62 struct io_ring_ctx
*ctx
);
64 int io_poll_issue(struct io_kiocb
*req
, bool *locked
);
65 int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
);
66 int io_do_iopoll(struct io_ring_ctx
*ctx
, bool force_nonspin
);
67 void io_free_batch_list(struct io_ring_ctx
*ctx
, struct io_wq_work_node
*node
);
68 int io_req_prep_async(struct io_kiocb
*req
);
70 struct io_wq_work
*io_wq_free_work(struct io_wq_work
*work
);
71 void io_wq_submit_work(struct io_wq_work
*work
);
73 void io_free_req(struct io_kiocb
*req
);
74 void io_queue_next(struct io_kiocb
*req
);
75 void __io_put_task(struct task_struct
*task
, int nr
);
76 void io_task_refs_refill(struct io_uring_task
*tctx
);
77 bool __io_alloc_req_refill(struct io_ring_ctx
*ctx
);
79 bool io_match_task_safe(struct io_kiocb
*head
, struct task_struct
*task
,
82 #define io_lockdep_assert_cq_locked(ctx) \
84 if (ctx->flags & IORING_SETUP_IOPOLL) { \
85 lockdep_assert_held(&ctx->uring_lock); \
86 } else if (!ctx->task_complete) { \
87 lockdep_assert_held(&ctx->completion_lock); \
88 } else if (ctx->submitter_task->flags & PF_EXITING) { \
89 lockdep_assert(current_work()); \
91 lockdep_assert(current == ctx->submitter_task); \
95 static inline void io_req_task_work_add(struct io_kiocb
*req
)
97 __io_req_task_work_add(req
, true);
100 #define io_for_each_link(pos, head) \
101 for (pos = (head); pos; pos = pos->link)
103 void io_cq_unlock_post(struct io_ring_ctx
*ctx
);
105 static inline struct io_uring_cqe
*io_get_cqe_overflow(struct io_ring_ctx
*ctx
,
108 io_lockdep_assert_cq_locked(ctx
);
110 if (likely(ctx
->cqe_cached
< ctx
->cqe_sentinel
)) {
111 struct io_uring_cqe
*cqe
= ctx
->cqe_cached
;
113 ctx
->cached_cq_tail
++;
115 if (ctx
->flags
& IORING_SETUP_CQE32
)
120 return __io_get_cqe(ctx
, overflow
);
123 static inline struct io_uring_cqe
*io_get_cqe(struct io_ring_ctx
*ctx
)
125 return io_get_cqe_overflow(ctx
, false);
128 static inline bool __io_fill_cqe_req(struct io_ring_ctx
*ctx
,
129 struct io_kiocb
*req
)
131 struct io_uring_cqe
*cqe
;
134 * If we can't get a cq entry, userspace overflowed the
135 * submission (by quite a lot). Increment the overflow count in
138 cqe
= io_get_cqe(ctx
);
142 trace_io_uring_complete(req
->ctx
, req
, req
->cqe
.user_data
,
143 req
->cqe
.res
, req
->cqe
.flags
,
144 (req
->flags
& REQ_F_CQE32_INIT
) ? req
->extra1
: 0,
145 (req
->flags
& REQ_F_CQE32_INIT
) ? req
->extra2
: 0);
147 memcpy(cqe
, &req
->cqe
, sizeof(*cqe
));
149 if (ctx
->flags
& IORING_SETUP_CQE32
) {
150 u64 extra1
= 0, extra2
= 0;
152 if (req
->flags
& REQ_F_CQE32_INIT
) {
153 extra1
= req
->extra1
;
154 extra2
= req
->extra2
;
157 WRITE_ONCE(cqe
->big_cqe
[0], extra1
);
158 WRITE_ONCE(cqe
->big_cqe
[1], extra2
);
163 static inline bool io_fill_cqe_req(struct io_ring_ctx
*ctx
,
164 struct io_kiocb
*req
)
166 if (likely(__io_fill_cqe_req(ctx
, req
)))
168 return io_req_cqe_overflow(req
);
171 static inline void req_set_fail(struct io_kiocb
*req
)
173 req
->flags
|= REQ_F_FAIL
;
174 if (req
->flags
& REQ_F_CQE_SKIP
) {
175 req
->flags
&= ~REQ_F_CQE_SKIP
;
176 req
->flags
|= REQ_F_SKIP_LINK_CQES
;
180 static inline void io_req_set_res(struct io_kiocb
*req
, s32 res
, u32 cflags
)
183 req
->cqe
.flags
= cflags
;
186 static inline bool req_has_async_data(struct io_kiocb
*req
)
188 return req
->flags
& REQ_F_ASYNC_DATA
;
191 static inline void io_put_file(struct file
*file
)
197 static inline void io_ring_submit_unlock(struct io_ring_ctx
*ctx
,
198 unsigned issue_flags
)
200 lockdep_assert_held(&ctx
->uring_lock
);
201 if (issue_flags
& IO_URING_F_UNLOCKED
)
202 mutex_unlock(&ctx
->uring_lock
);
205 static inline void io_ring_submit_lock(struct io_ring_ctx
*ctx
,
206 unsigned issue_flags
)
209 * "Normal" inline submissions always hold the uring_lock, since we
210 * grab it from the system call. Same is true for the SQPOLL offload.
211 * The only exception is when we've detached the request and issue it
212 * from an async worker thread, grab the lock for that case.
214 if (issue_flags
& IO_URING_F_UNLOCKED
)
215 mutex_lock(&ctx
->uring_lock
);
216 lockdep_assert_held(&ctx
->uring_lock
);
219 static inline void io_commit_cqring(struct io_ring_ctx
*ctx
)
221 /* order cqe stores with ring update */
222 smp_store_release(&ctx
->rings
->cq
.tail
, ctx
->cached_cq_tail
);
225 /* requires smb_mb() prior, see wq_has_sleeper() */
226 static inline void __io_cqring_wake(struct io_ring_ctx
*ctx
)
229 * Trigger waitqueue handler on all waiters on our waitqueue. This
230 * won't necessarily wake up all the tasks, io_should_wake() will make
233 * Pass in EPOLLIN|EPOLL_URING_WAKE as the poll wakeup key. The latter
234 * set in the mask so that if we recurse back into our own poll
235 * waitqueue handlers, we know we have a dependency between eventfd or
236 * epoll and should terminate multishot poll at that point.
238 if (waitqueue_active(&ctx
->cq_wait
))
239 __wake_up(&ctx
->cq_wait
, TASK_NORMAL
, 0,
240 poll_to_key(EPOLL_URING_WAKE
| EPOLLIN
));
243 static inline void io_cqring_wake(struct io_ring_ctx
*ctx
)
246 __io_cqring_wake(ctx
);
249 static inline bool io_sqring_full(struct io_ring_ctx
*ctx
)
251 struct io_rings
*r
= ctx
->rings
;
253 return READ_ONCE(r
->sq
.tail
) - ctx
->cached_sq_head
== ctx
->sq_entries
;
256 static inline unsigned int io_sqring_entries(struct io_ring_ctx
*ctx
)
258 struct io_rings
*rings
= ctx
->rings
;
260 /* make sure SQ entry isn't read before tail */
261 return smp_load_acquire(&rings
->sq
.tail
) - ctx
->cached_sq_head
;
264 static inline int io_run_task_work(void)
267 * Always check-and-clear the task_work notification signal. With how
268 * signaling works for task_work, we can find it set with nothing to
269 * run. We need to clear it for that case, like get_signal() does.
271 if (test_thread_flag(TIF_NOTIFY_SIGNAL
))
272 clear_notify_signal();
273 if (task_work_pending(current
)) {
274 __set_current_state(TASK_RUNNING
);
282 static inline bool io_task_work_pending(struct io_ring_ctx
*ctx
)
284 return task_work_pending(current
) || !wq_list_empty(&ctx
->work_llist
);
287 static inline int io_run_task_work_ctx(struct io_ring_ctx
*ctx
)
292 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
)
293 ret
= io_run_local_work(ctx
);
295 /* want to run this after in case more is added */
296 ret2
= io_run_task_work();
298 /* Try propagate error in favour of if tasks were run,
299 * but still make sure to run them if requested
307 static inline int io_run_local_work_locked(struct io_ring_ctx
*ctx
)
312 if (llist_empty(&ctx
->work_llist
))
316 ret
= __io_run_local_work(ctx
, &locked
);
317 /* shouldn't happen! */
318 if (WARN_ON_ONCE(!locked
))
319 mutex_lock(&ctx
->uring_lock
);
323 static inline void io_tw_lock(struct io_ring_ctx
*ctx
, bool *locked
)
326 mutex_lock(&ctx
->uring_lock
);
332 * Don't complete immediately but use deferred completion infrastructure.
333 * Protected by ->uring_lock and can only be used either with
334 * IO_URING_F_COMPLETE_DEFER or inside a tw handler holding the mutex.
336 static inline void io_req_complete_defer(struct io_kiocb
*req
)
337 __must_hold(&req
->ctx
->uring_lock
)
339 struct io_submit_state
*state
= &req
->ctx
->submit_state
;
341 lockdep_assert_held(&req
->ctx
->uring_lock
);
343 wq_list_add_tail(&req
->comp_list
, &state
->compl_reqs
);
346 static inline void io_commit_cqring_flush(struct io_ring_ctx
*ctx
)
348 if (unlikely(ctx
->off_timeout_used
|| ctx
->drain_active
|| ctx
->has_evfd
))
349 __io_commit_cqring_flush(ctx
);
352 /* must to be called somewhat shortly after putting a request */
353 static inline void io_put_task(struct task_struct
*task
, int nr
)
355 if (likely(task
== current
))
356 task
->io_uring
->cached_refs
+= nr
;
358 __io_put_task(task
, nr
);
361 static inline void io_get_task_refs(int nr
)
363 struct io_uring_task
*tctx
= current
->io_uring
;
365 tctx
->cached_refs
-= nr
;
366 if (unlikely(tctx
->cached_refs
< 0))
367 io_task_refs_refill(tctx
);
370 static inline bool io_req_cache_empty(struct io_ring_ctx
*ctx
)
372 return !ctx
->submit_state
.free_list
.next
;
375 static inline bool io_alloc_req_refill(struct io_ring_ctx
*ctx
)
377 if (unlikely(io_req_cache_empty(ctx
)))
378 return __io_alloc_req_refill(ctx
);
382 static inline struct io_kiocb
*io_alloc_req(struct io_ring_ctx
*ctx
)
384 struct io_wq_work_node
*node
;
386 node
= wq_stack_extract(&ctx
->submit_state
.free_list
);
387 return container_of(node
, struct io_kiocb
, comp_list
);
390 static inline bool io_allowed_run_tw(struct io_ring_ctx
*ctx
)
392 return likely(!(ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) ||
393 ctx
->submitter_task
== current
);
396 static inline void io_req_queue_tw_complete(struct io_kiocb
*req
, s32 res
)
398 io_req_set_res(req
, res
, 0);
399 req
->io_task_work
.func
= io_req_task_complete
;
400 io_req_task_work_add(req
);