4 #include <linux/errno.h>
5 #include <linux/lockdep.h>
6 #include <linux/resume_user_mode.h>
7 #include <linux/kasan.h>
8 #include <linux/poll.h>
9 #include <linux/io_uring_types.h>
10 #include <uapi/linux/eventpoll.h>
13 #include "filetable.h"
15 #ifndef CREATE_TRACE_POINTS
16 #include <trace/events/io_uring.h>
21 IOU_ISSUE_SKIP_COMPLETE
= -EIOCBQUEUED
,
24 * Requeue the task_work to restart operations on this request. The
25 * actual value isn't important, should just be not an otherwise
26 * valid error code, yet less than -MAX_ERRNO and valid internally.
31 * Intended only when both IO_URING_F_MULTISHOT is passed
32 * to indicate to the poll runner that multishot should be
33 * removed and the result is set on req->cqe.res.
35 IOU_STOP_MULTISHOT
= -ECANCELED
,
38 struct io_wait_queue
{
39 struct wait_queue_entry wq
;
40 struct io_ring_ctx
*ctx
;
49 #ifdef CONFIG_NET_RX_BUSY_POLL
50 ktime_t napi_busy_poll_dt
;
51 bool napi_prefer_busy_poll
;
55 static inline bool io_should_wake(struct io_wait_queue
*iowq
)
57 struct io_ring_ctx
*ctx
= iowq
->ctx
;
58 int dist
= READ_ONCE(ctx
->rings
->cq
.tail
) - (int) iowq
->cq_tail
;
61 * Wake up if we have enough events, or if a timeout occurred since we
62 * started waiting. For timeouts, we always want to return to userspace,
63 * regardless of event count.
65 return dist
>= 0 || atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
68 bool io_cqe_cache_refill(struct io_ring_ctx
*ctx
, bool overflow
);
69 int io_run_task_work_sig(struct io_ring_ctx
*ctx
);
70 void io_req_defer_failed(struct io_kiocb
*req
, s32 res
);
71 bool io_post_aux_cqe(struct io_ring_ctx
*ctx
, u64 user_data
, s32 res
, u32 cflags
);
72 void io_add_aux_cqe(struct io_ring_ctx
*ctx
, u64 user_data
, s32 res
, u32 cflags
);
73 bool io_req_post_cqe(struct io_kiocb
*req
, s32 res
, u32 cflags
);
74 void __io_commit_cqring_flush(struct io_ring_ctx
*ctx
);
76 struct file
*io_file_get_normal(struct io_kiocb
*req
, int fd
);
77 struct file
*io_file_get_fixed(struct io_kiocb
*req
, int fd
,
78 unsigned issue_flags
);
80 void __io_req_task_work_add(struct io_kiocb
*req
, unsigned flags
);
81 void io_req_task_work_add_remote(struct io_kiocb
*req
, struct io_ring_ctx
*ctx
,
83 bool io_alloc_async_data(struct io_kiocb
*req
);
84 void io_req_task_queue(struct io_kiocb
*req
);
85 void io_req_task_complete(struct io_kiocb
*req
, struct io_tw_state
*ts
);
86 void io_req_task_queue_fail(struct io_kiocb
*req
, int ret
);
87 void io_req_task_submit(struct io_kiocb
*req
, struct io_tw_state
*ts
);
88 struct llist_node
*io_handle_tw_list(struct llist_node
*node
, unsigned int *count
, unsigned int max_entries
);
89 struct llist_node
*tctx_task_work_run(struct io_uring_task
*tctx
, unsigned int max_entries
, unsigned int *count
);
90 void tctx_task_work(struct callback_head
*cb
);
91 __cold
void io_uring_cancel_generic(bool cancel_all
, struct io_sq_data
*sqd
);
92 int io_uring_alloc_task_context(struct task_struct
*task
,
93 struct io_ring_ctx
*ctx
);
95 int io_ring_add_registered_file(struct io_uring_task
*tctx
, struct file
*file
,
97 void io_req_queue_iowq(struct io_kiocb
*req
);
99 int io_poll_issue(struct io_kiocb
*req
, struct io_tw_state
*ts
);
100 int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
);
101 int io_do_iopoll(struct io_ring_ctx
*ctx
, bool force_nonspin
);
102 void __io_submit_flush_completions(struct io_ring_ctx
*ctx
);
104 struct io_wq_work
*io_wq_free_work(struct io_wq_work
*work
);
105 void io_wq_submit_work(struct io_wq_work
*work
);
107 void io_free_req(struct io_kiocb
*req
);
108 void io_queue_next(struct io_kiocb
*req
);
109 void io_task_refs_refill(struct io_uring_task
*tctx
);
110 bool __io_alloc_req_refill(struct io_ring_ctx
*ctx
);
112 bool io_match_task_safe(struct io_kiocb
*head
, struct task_struct
*task
,
115 void io_activate_pollwq(struct io_ring_ctx
*ctx
);
117 static inline void io_lockdep_assert_cq_locked(struct io_ring_ctx
*ctx
)
119 #if defined(CONFIG_PROVE_LOCKING)
120 lockdep_assert(in_task());
122 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
123 lockdep_assert_held(&ctx
->uring_lock
);
124 } else if (!ctx
->task_complete
) {
125 lockdep_assert_held(&ctx
->completion_lock
);
126 } else if (ctx
->submitter_task
) {
128 * ->submitter_task may be NULL and we can still post a CQE,
129 * if the ring has been setup with IORING_SETUP_R_DISABLED.
130 * Not from an SQE, as those cannot be submitted, but via
131 * updating tagged resources.
133 if (ctx
->submitter_task
->flags
& PF_EXITING
)
134 lockdep_assert(current_work());
136 lockdep_assert(current
== ctx
->submitter_task
);
141 static inline void io_req_task_work_add(struct io_kiocb
*req
)
143 __io_req_task_work_add(req
, 0);
146 static inline void io_submit_flush_completions(struct io_ring_ctx
*ctx
)
148 if (!wq_list_empty(&ctx
->submit_state
.compl_reqs
) ||
149 ctx
->submit_state
.cq_flush
)
150 __io_submit_flush_completions(ctx
);
153 #define io_for_each_link(pos, head) \
154 for (pos = (head); pos; pos = pos->link)
156 static inline bool io_get_cqe_overflow(struct io_ring_ctx
*ctx
,
157 struct io_uring_cqe
**ret
,
160 io_lockdep_assert_cq_locked(ctx
);
162 if (unlikely(ctx
->cqe_cached
>= ctx
->cqe_sentinel
)) {
163 if (unlikely(!io_cqe_cache_refill(ctx
, overflow
)))
166 *ret
= ctx
->cqe_cached
;
167 ctx
->cached_cq_tail
++;
169 if (ctx
->flags
& IORING_SETUP_CQE32
)
174 static inline bool io_get_cqe(struct io_ring_ctx
*ctx
, struct io_uring_cqe
**ret
)
176 return io_get_cqe_overflow(ctx
, ret
, false);
179 static __always_inline
bool io_fill_cqe_req(struct io_ring_ctx
*ctx
,
180 struct io_kiocb
*req
)
182 struct io_uring_cqe
*cqe
;
185 * If we can't get a cq entry, userspace overflowed the
186 * submission (by quite a lot). Increment the overflow count in
189 if (unlikely(!io_get_cqe(ctx
, &cqe
)))
192 if (trace_io_uring_complete_enabled())
193 trace_io_uring_complete(req
->ctx
, req
, req
->cqe
.user_data
,
194 req
->cqe
.res
, req
->cqe
.flags
,
195 req
->big_cqe
.extra1
, req
->big_cqe
.extra2
);
197 memcpy(cqe
, &req
->cqe
, sizeof(*cqe
));
198 if (ctx
->flags
& IORING_SETUP_CQE32
) {
199 memcpy(cqe
->big_cqe
, &req
->big_cqe
, sizeof(*cqe
));
200 memset(&req
->big_cqe
, 0, sizeof(req
->big_cqe
));
205 static inline void req_set_fail(struct io_kiocb
*req
)
207 req
->flags
|= REQ_F_FAIL
;
208 if (req
->flags
& REQ_F_CQE_SKIP
) {
209 req
->flags
&= ~REQ_F_CQE_SKIP
;
210 req
->flags
|= REQ_F_SKIP_LINK_CQES
;
214 static inline void io_req_set_res(struct io_kiocb
*req
, s32 res
, u32 cflags
)
217 req
->cqe
.flags
= cflags
;
220 static inline bool req_has_async_data(struct io_kiocb
*req
)
222 return req
->flags
& REQ_F_ASYNC_DATA
;
225 static inline void io_put_file(struct io_kiocb
*req
)
227 if (!(req
->flags
& REQ_F_FIXED_FILE
) && req
->file
)
231 static inline void io_ring_submit_unlock(struct io_ring_ctx
*ctx
,
232 unsigned issue_flags
)
234 lockdep_assert_held(&ctx
->uring_lock
);
235 if (unlikely(issue_flags
& IO_URING_F_UNLOCKED
))
236 mutex_unlock(&ctx
->uring_lock
);
239 static inline void io_ring_submit_lock(struct io_ring_ctx
*ctx
,
240 unsigned issue_flags
)
243 * "Normal" inline submissions always hold the uring_lock, since we
244 * grab it from the system call. Same is true for the SQPOLL offload.
245 * The only exception is when we've detached the request and issue it
246 * from an async worker thread, grab the lock for that case.
248 if (unlikely(issue_flags
& IO_URING_F_UNLOCKED
))
249 mutex_lock(&ctx
->uring_lock
);
250 lockdep_assert_held(&ctx
->uring_lock
);
253 static inline void io_commit_cqring(struct io_ring_ctx
*ctx
)
255 /* order cqe stores with ring update */
256 smp_store_release(&ctx
->rings
->cq
.tail
, ctx
->cached_cq_tail
);
259 static inline void io_poll_wq_wake(struct io_ring_ctx
*ctx
)
261 if (wq_has_sleeper(&ctx
->poll_wq
))
262 __wake_up(&ctx
->poll_wq
, TASK_NORMAL
, 0,
263 poll_to_key(EPOLL_URING_WAKE
| EPOLLIN
));
266 static inline void io_cqring_wake(struct io_ring_ctx
*ctx
)
269 * Trigger waitqueue handler on all waiters on our waitqueue. This
270 * won't necessarily wake up all the tasks, io_should_wake() will make
273 * Pass in EPOLLIN|EPOLL_URING_WAKE as the poll wakeup key. The latter
274 * set in the mask so that if we recurse back into our own poll
275 * waitqueue handlers, we know we have a dependency between eventfd or
276 * epoll and should terminate multishot poll at that point.
278 if (wq_has_sleeper(&ctx
->cq_wait
))
279 __wake_up(&ctx
->cq_wait
, TASK_NORMAL
, 0,
280 poll_to_key(EPOLL_URING_WAKE
| EPOLLIN
));
283 static inline bool io_sqring_full(struct io_ring_ctx
*ctx
)
285 struct io_rings
*r
= ctx
->rings
;
287 return READ_ONCE(r
->sq
.tail
) - ctx
->cached_sq_head
== ctx
->sq_entries
;
290 static inline unsigned int io_sqring_entries(struct io_ring_ctx
*ctx
)
292 struct io_rings
*rings
= ctx
->rings
;
293 unsigned int entries
;
295 /* make sure SQ entry isn't read before tail */
296 entries
= smp_load_acquire(&rings
->sq
.tail
) - ctx
->cached_sq_head
;
297 return min(entries
, ctx
->sq_entries
);
300 static inline int io_run_task_work(void)
305 * Always check-and-clear the task_work notification signal. With how
306 * signaling works for task_work, we can find it set with nothing to
307 * run. We need to clear it for that case, like get_signal() does.
309 if (test_thread_flag(TIF_NOTIFY_SIGNAL
))
310 clear_notify_signal();
312 * PF_IO_WORKER never returns to userspace, so check here if we have
313 * notify work that needs processing.
315 if (current
->flags
& PF_IO_WORKER
) {
316 if (test_thread_flag(TIF_NOTIFY_RESUME
)) {
317 __set_current_state(TASK_RUNNING
);
318 resume_user_mode_work(NULL
);
320 if (current
->io_uring
) {
321 unsigned int count
= 0;
323 tctx_task_work_run(current
->io_uring
, UINT_MAX
, &count
);
328 if (task_work_pending(current
)) {
329 __set_current_state(TASK_RUNNING
);
337 static inline bool io_task_work_pending(struct io_ring_ctx
*ctx
)
339 return task_work_pending(current
) || !llist_empty(&ctx
->work_llist
);
342 static inline void io_tw_lock(struct io_ring_ctx
*ctx
, struct io_tw_state
*ts
)
344 lockdep_assert_held(&ctx
->uring_lock
);
348 * Don't complete immediately but use deferred completion infrastructure.
349 * Protected by ->uring_lock and can only be used either with
350 * IO_URING_F_COMPLETE_DEFER or inside a tw handler holding the mutex.
352 static inline void io_req_complete_defer(struct io_kiocb
*req
)
353 __must_hold(&req
->ctx
->uring_lock
)
355 struct io_submit_state
*state
= &req
->ctx
->submit_state
;
357 lockdep_assert_held(&req
->ctx
->uring_lock
);
359 wq_list_add_tail(&req
->comp_list
, &state
->compl_reqs
);
362 static inline void io_commit_cqring_flush(struct io_ring_ctx
*ctx
)
364 if (unlikely(ctx
->off_timeout_used
|| ctx
->drain_active
||
365 ctx
->has_evfd
|| ctx
->poll_activated
))
366 __io_commit_cqring_flush(ctx
);
369 static inline void io_get_task_refs(int nr
)
371 struct io_uring_task
*tctx
= current
->io_uring
;
373 tctx
->cached_refs
-= nr
;
374 if (unlikely(tctx
->cached_refs
< 0))
375 io_task_refs_refill(tctx
);
378 static inline bool io_req_cache_empty(struct io_ring_ctx
*ctx
)
380 return !ctx
->submit_state
.free_list
.next
;
383 extern struct kmem_cache
*req_cachep
;
384 extern struct kmem_cache
*io_buf_cachep
;
386 static inline struct io_kiocb
*io_extract_req(struct io_ring_ctx
*ctx
)
388 struct io_kiocb
*req
;
390 req
= container_of(ctx
->submit_state
.free_list
.next
, struct io_kiocb
, comp_list
);
391 wq_stack_extract(&ctx
->submit_state
.free_list
);
395 static inline bool io_alloc_req(struct io_ring_ctx
*ctx
, struct io_kiocb
**req
)
397 if (unlikely(io_req_cache_empty(ctx
))) {
398 if (!__io_alloc_req_refill(ctx
))
401 *req
= io_extract_req(ctx
);
405 static inline bool io_allowed_defer_tw_run(struct io_ring_ctx
*ctx
)
407 return likely(ctx
->submitter_task
== current
);
410 static inline bool io_allowed_run_tw(struct io_ring_ctx
*ctx
)
412 return likely(!(ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) ||
413 ctx
->submitter_task
== current
);
416 static inline void io_req_queue_tw_complete(struct io_kiocb
*req
, s32 res
)
418 io_req_set_res(req
, res
, 0);
419 req
->io_task_work
.func
= io_req_task_complete
;
420 io_req_task_work_add(req
);
424 * IORING_SETUP_SQE128 contexts allocate twice the normal SQE size for each
427 static inline size_t uring_sqe_size(struct io_ring_ctx
*ctx
)
429 if (ctx
->flags
& IORING_SETUP_SQE128
)
430 return 2 * sizeof(struct io_uring_sqe
);
431 return sizeof(struct io_uring_sqe
);
434 static inline bool io_file_can_poll(struct io_kiocb
*req
)
436 if (req
->flags
& REQ_F_CAN_POLL
)
438 if (req
->file
&& file_can_poll(req
->file
)) {
439 req
->flags
|= REQ_F_CAN_POLL
;
445 static inline ktime_t
io_get_time(struct io_ring_ctx
*ctx
)
447 if (ctx
->clockid
== CLOCK_MONOTONIC
)
450 return ktime_get_with_offset(ctx
->clock_offset
);
454 IO_CHECK_CQ_OVERFLOW_BIT
,
455 IO_CHECK_CQ_DROPPED_BIT
,
458 static inline bool io_has_work(struct io_ring_ctx
*ctx
)
460 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
) ||
461 !llist_empty(&ctx
->work_llist
);