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>
11 #include "alloc_cache.h"
14 #include "filetable.h"
17 #ifndef CREATE_TRACE_POINTS
18 #include <trace/events/io_uring.h>
24 IOU_ISSUE_SKIP_COMPLETE
= -EIOCBQUEUED
,
27 * The request has more work to do and should be retried. io_uring will
28 * attempt to wait on the file for eligible opcodes, but otherwise
29 * it'll be handed to iowq for blocking execution. It works for normal
30 * requests as well as for the multi shot mode.
35 * Requeue the task_work to restart operations on this request. The
36 * actual value isn't important, should just be not an otherwise
37 * valid error code, yet less than -MAX_ERRNO and valid internally.
42 struct io_wait_queue
{
43 struct wait_queue_entry wq
;
44 struct io_ring_ctx
*ctx
;
53 #ifdef CONFIG_NET_RX_BUSY_POLL
54 ktime_t napi_busy_poll_dt
;
55 bool napi_prefer_busy_poll
;
59 static inline bool io_should_wake(struct io_wait_queue
*iowq
)
61 struct io_ring_ctx
*ctx
= iowq
->ctx
;
62 int dist
= READ_ONCE(ctx
->rings
->cq
.tail
) - (int) iowq
->cq_tail
;
65 * Wake up if we have enough events, or if a timeout occurred since we
66 * started waiting. For timeouts, we always want to return to userspace,
67 * regardless of event count.
69 return dist
>= 0 || atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
72 #define IORING_MAX_ENTRIES 32768
73 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
75 unsigned long rings_size(unsigned int flags
, unsigned int sq_entries
,
76 unsigned int cq_entries
, size_t *sq_offset
);
77 int io_uring_fill_params(unsigned entries
, struct io_uring_params
*p
);
78 bool io_cqe_cache_refill(struct io_ring_ctx
*ctx
, bool overflow
);
79 int io_run_task_work_sig(struct io_ring_ctx
*ctx
);
80 void io_req_defer_failed(struct io_kiocb
*req
, s32 res
);
81 bool io_post_aux_cqe(struct io_ring_ctx
*ctx
, u64 user_data
, s32 res
, u32 cflags
);
82 void io_add_aux_cqe(struct io_ring_ctx
*ctx
, u64 user_data
, s32 res
, u32 cflags
);
83 bool io_req_post_cqe(struct io_kiocb
*req
, s32 res
, u32 cflags
);
84 void __io_commit_cqring_flush(struct io_ring_ctx
*ctx
);
86 void io_req_track_inflight(struct io_kiocb
*req
);
87 struct file
*io_file_get_normal(struct io_kiocb
*req
, int fd
);
88 struct file
*io_file_get_fixed(struct io_kiocb
*req
, int fd
,
89 unsigned issue_flags
);
91 void __io_req_task_work_add(struct io_kiocb
*req
, unsigned flags
);
92 void io_req_task_work_add_remote(struct io_kiocb
*req
, unsigned flags
);
93 void io_req_task_queue(struct io_kiocb
*req
);
94 void io_req_task_complete(struct io_kiocb
*req
, io_tw_token_t tw
);
95 void io_req_task_queue_fail(struct io_kiocb
*req
, int ret
);
96 void io_req_task_submit(struct io_kiocb
*req
, io_tw_token_t tw
);
97 struct llist_node
*io_handle_tw_list(struct llist_node
*node
, unsigned int *count
, unsigned int max_entries
);
98 struct llist_node
*tctx_task_work_run(struct io_uring_task
*tctx
, unsigned int max_entries
, unsigned int *count
);
99 void tctx_task_work(struct callback_head
*cb
);
100 __cold
void io_uring_cancel_generic(bool cancel_all
, struct io_sq_data
*sqd
);
102 int io_ring_add_registered_file(struct io_uring_task
*tctx
, struct file
*file
,
104 void io_req_queue_iowq(struct io_kiocb
*req
);
106 int io_poll_issue(struct io_kiocb
*req
, io_tw_token_t tw
);
107 int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
);
108 int io_do_iopoll(struct io_ring_ctx
*ctx
, bool force_nonspin
);
109 void __io_submit_flush_completions(struct io_ring_ctx
*ctx
);
111 struct io_wq_work
*io_wq_free_work(struct io_wq_work
*work
);
112 void io_wq_submit_work(struct io_wq_work
*work
);
114 void io_free_req(struct io_kiocb
*req
);
115 void io_queue_next(struct io_kiocb
*req
);
116 void io_task_refs_refill(struct io_uring_task
*tctx
);
117 bool __io_alloc_req_refill(struct io_ring_ctx
*ctx
);
119 bool io_match_task_safe(struct io_kiocb
*head
, struct io_uring_task
*tctx
,
122 void io_activate_pollwq(struct io_ring_ctx
*ctx
);
124 static inline void io_lockdep_assert_cq_locked(struct io_ring_ctx
*ctx
)
126 #if defined(CONFIG_PROVE_LOCKING)
127 lockdep_assert(in_task());
129 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
)
130 lockdep_assert_held(&ctx
->uring_lock
);
132 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
133 lockdep_assert_held(&ctx
->uring_lock
);
134 } else if (!ctx
->task_complete
) {
135 lockdep_assert_held(&ctx
->completion_lock
);
136 } else if (ctx
->submitter_task
) {
138 * ->submitter_task may be NULL and we can still post a CQE,
139 * if the ring has been setup with IORING_SETUP_R_DISABLED.
140 * Not from an SQE, as those cannot be submitted, but via
141 * updating tagged resources.
143 if (!percpu_ref_is_dying(&ctx
->refs
))
144 lockdep_assert(current
== ctx
->submitter_task
);
149 static inline bool io_is_compat(struct io_ring_ctx
*ctx
)
151 return IS_ENABLED(CONFIG_COMPAT
) && unlikely(ctx
->compat
);
154 static inline void io_req_task_work_add(struct io_kiocb
*req
)
156 __io_req_task_work_add(req
, 0);
159 static inline void io_submit_flush_completions(struct io_ring_ctx
*ctx
)
161 if (!wq_list_empty(&ctx
->submit_state
.compl_reqs
) ||
162 ctx
->submit_state
.cq_flush
)
163 __io_submit_flush_completions(ctx
);
166 #define io_for_each_link(pos, head) \
167 for (pos = (head); pos; pos = pos->link)
169 static inline bool io_get_cqe_overflow(struct io_ring_ctx
*ctx
,
170 struct io_uring_cqe
**ret
,
173 io_lockdep_assert_cq_locked(ctx
);
175 if (unlikely(ctx
->cqe_cached
>= ctx
->cqe_sentinel
)) {
176 if (unlikely(!io_cqe_cache_refill(ctx
, overflow
)))
179 *ret
= ctx
->cqe_cached
;
180 ctx
->cached_cq_tail
++;
182 if (ctx
->flags
& IORING_SETUP_CQE32
)
187 static inline bool io_get_cqe(struct io_ring_ctx
*ctx
, struct io_uring_cqe
**ret
)
189 return io_get_cqe_overflow(ctx
, ret
, false);
192 static inline bool io_defer_get_uncommited_cqe(struct io_ring_ctx
*ctx
,
193 struct io_uring_cqe
**cqe_ret
)
195 io_lockdep_assert_cq_locked(ctx
);
197 ctx
->submit_state
.cq_flush
= true;
198 return io_get_cqe(ctx
, cqe_ret
);
201 static __always_inline
bool io_fill_cqe_req(struct io_ring_ctx
*ctx
,
202 struct io_kiocb
*req
)
204 struct io_uring_cqe
*cqe
;
207 * If we can't get a cq entry, userspace overflowed the
208 * submission (by quite a lot). Increment the overflow count in
211 if (unlikely(!io_get_cqe(ctx
, &cqe
)))
215 memcpy(cqe
, &req
->cqe
, sizeof(*cqe
));
216 if (ctx
->flags
& IORING_SETUP_CQE32
) {
217 memcpy(cqe
->big_cqe
, &req
->big_cqe
, sizeof(*cqe
));
218 memset(&req
->big_cqe
, 0, sizeof(req
->big_cqe
));
221 if (trace_io_uring_complete_enabled())
222 trace_io_uring_complete(req
->ctx
, req
, cqe
);
226 static inline void req_set_fail(struct io_kiocb
*req
)
228 req
->flags
|= REQ_F_FAIL
;
229 if (req
->flags
& REQ_F_CQE_SKIP
) {
230 req
->flags
&= ~REQ_F_CQE_SKIP
;
231 req
->flags
|= REQ_F_SKIP_LINK_CQES
;
235 static inline void io_req_set_res(struct io_kiocb
*req
, s32 res
, u32 cflags
)
238 req
->cqe
.flags
= cflags
;
241 static inline void *io_uring_alloc_async_data(struct io_alloc_cache
*cache
,
242 struct io_kiocb
*req
)
245 req
->async_data
= io_cache_alloc(cache
, GFP_KERNEL
);
247 const struct io_issue_def
*def
= &io_issue_defs
[req
->opcode
];
249 WARN_ON_ONCE(!def
->async_size
);
250 req
->async_data
= kmalloc(def
->async_size
, GFP_KERNEL
);
253 req
->flags
|= REQ_F_ASYNC_DATA
;
254 return req
->async_data
;
257 static inline bool req_has_async_data(struct io_kiocb
*req
)
259 return req
->flags
& REQ_F_ASYNC_DATA
;
262 static inline void io_put_file(struct io_kiocb
*req
)
264 if (!(req
->flags
& REQ_F_FIXED_FILE
) && req
->file
)
268 static inline void io_ring_submit_unlock(struct io_ring_ctx
*ctx
,
269 unsigned issue_flags
)
271 lockdep_assert_held(&ctx
->uring_lock
);
272 if (unlikely(issue_flags
& IO_URING_F_UNLOCKED
))
273 mutex_unlock(&ctx
->uring_lock
);
276 static inline void io_ring_submit_lock(struct io_ring_ctx
*ctx
,
277 unsigned issue_flags
)
280 * "Normal" inline submissions always hold the uring_lock, since we
281 * grab it from the system call. Same is true for the SQPOLL offload.
282 * The only exception is when we've detached the request and issue it
283 * from an async worker thread, grab the lock for that case.
285 if (unlikely(issue_flags
& IO_URING_F_UNLOCKED
))
286 mutex_lock(&ctx
->uring_lock
);
287 lockdep_assert_held(&ctx
->uring_lock
);
290 static inline void io_commit_cqring(struct io_ring_ctx
*ctx
)
292 /* order cqe stores with ring update */
293 smp_store_release(&ctx
->rings
->cq
.tail
, ctx
->cached_cq_tail
);
296 static inline void io_poll_wq_wake(struct io_ring_ctx
*ctx
)
298 if (wq_has_sleeper(&ctx
->poll_wq
))
299 __wake_up(&ctx
->poll_wq
, TASK_NORMAL
, 0,
300 poll_to_key(EPOLL_URING_WAKE
| EPOLLIN
));
303 static inline void io_cqring_wake(struct io_ring_ctx
*ctx
)
306 * Trigger waitqueue handler on all waiters on our waitqueue. This
307 * won't necessarily wake up all the tasks, io_should_wake() will make
310 * Pass in EPOLLIN|EPOLL_URING_WAKE as the poll wakeup key. The latter
311 * set in the mask so that if we recurse back into our own poll
312 * waitqueue handlers, we know we have a dependency between eventfd or
313 * epoll and should terminate multishot poll at that point.
315 if (wq_has_sleeper(&ctx
->cq_wait
))
316 __wake_up(&ctx
->cq_wait
, TASK_NORMAL
, 0,
317 poll_to_key(EPOLL_URING_WAKE
| EPOLLIN
));
320 static inline bool io_sqring_full(struct io_ring_ctx
*ctx
)
322 struct io_rings
*r
= ctx
->rings
;
325 * SQPOLL must use the actual sqring head, as using the cached_sq_head
326 * is race prone if the SQPOLL thread has grabbed entries but not yet
327 * committed them to the ring. For !SQPOLL, this doesn't matter, but
328 * since this helper is just used for SQPOLL sqring waits (or POLLOUT),
329 * just read the actual sqring head unconditionally.
331 return READ_ONCE(r
->sq
.tail
) - READ_ONCE(r
->sq
.head
) == ctx
->sq_entries
;
334 static inline unsigned int io_sqring_entries(struct io_ring_ctx
*ctx
)
336 struct io_rings
*rings
= ctx
->rings
;
337 unsigned int entries
;
339 /* make sure SQ entry isn't read before tail */
340 entries
= smp_load_acquire(&rings
->sq
.tail
) - ctx
->cached_sq_head
;
341 return min(entries
, ctx
->sq_entries
);
344 static inline int io_run_task_work(void)
349 * Always check-and-clear the task_work notification signal. With how
350 * signaling works for task_work, we can find it set with nothing to
351 * run. We need to clear it for that case, like get_signal() does.
353 if (test_thread_flag(TIF_NOTIFY_SIGNAL
))
354 clear_notify_signal();
356 * PF_IO_WORKER never returns to userspace, so check here if we have
357 * notify work that needs processing.
359 if (current
->flags
& PF_IO_WORKER
) {
360 if (test_thread_flag(TIF_NOTIFY_RESUME
)) {
361 __set_current_state(TASK_RUNNING
);
362 resume_user_mode_work(NULL
);
364 if (current
->io_uring
) {
365 unsigned int count
= 0;
367 __set_current_state(TASK_RUNNING
);
368 tctx_task_work_run(current
->io_uring
, UINT_MAX
, &count
);
373 if (task_work_pending(current
)) {
374 __set_current_state(TASK_RUNNING
);
382 static inline bool io_local_work_pending(struct io_ring_ctx
*ctx
)
384 return !llist_empty(&ctx
->work_llist
) || !llist_empty(&ctx
->retry_llist
);
387 static inline bool io_task_work_pending(struct io_ring_ctx
*ctx
)
389 return task_work_pending(current
) || io_local_work_pending(ctx
);
392 static inline void io_tw_lock(struct io_ring_ctx
*ctx
, io_tw_token_t tw
)
394 lockdep_assert_held(&ctx
->uring_lock
);
398 * Don't complete immediately but use deferred completion infrastructure.
399 * Protected by ->uring_lock and can only be used either with
400 * IO_URING_F_COMPLETE_DEFER or inside a tw handler holding the mutex.
402 static inline void io_req_complete_defer(struct io_kiocb
*req
)
403 __must_hold(&req
->ctx
->uring_lock
)
405 struct io_submit_state
*state
= &req
->ctx
->submit_state
;
407 lockdep_assert_held(&req
->ctx
->uring_lock
);
409 wq_list_add_tail(&req
->comp_list
, &state
->compl_reqs
);
412 static inline void io_commit_cqring_flush(struct io_ring_ctx
*ctx
)
414 if (unlikely(ctx
->off_timeout_used
||
415 ctx
->has_evfd
|| ctx
->poll_activated
))
416 __io_commit_cqring_flush(ctx
);
419 static inline void io_get_task_refs(int nr
)
421 struct io_uring_task
*tctx
= current
->io_uring
;
423 tctx
->cached_refs
-= nr
;
424 if (unlikely(tctx
->cached_refs
< 0))
425 io_task_refs_refill(tctx
);
428 static inline bool io_req_cache_empty(struct io_ring_ctx
*ctx
)
430 return !ctx
->submit_state
.free_list
.next
;
433 extern struct kmem_cache
*req_cachep
;
435 static inline struct io_kiocb
*io_extract_req(struct io_ring_ctx
*ctx
)
437 struct io_kiocb
*req
;
439 req
= container_of(ctx
->submit_state
.free_list
.next
, struct io_kiocb
, comp_list
);
440 wq_stack_extract(&ctx
->submit_state
.free_list
);
444 static inline bool io_alloc_req(struct io_ring_ctx
*ctx
, struct io_kiocb
**req
)
446 if (unlikely(io_req_cache_empty(ctx
))) {
447 if (!__io_alloc_req_refill(ctx
))
450 *req
= io_extract_req(ctx
);
454 static inline bool io_allowed_defer_tw_run(struct io_ring_ctx
*ctx
)
456 return likely(ctx
->submitter_task
== current
);
459 static inline bool io_allowed_run_tw(struct io_ring_ctx
*ctx
)
461 return likely(!(ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) ||
462 ctx
->submitter_task
== current
);
466 * Terminate the request if either of these conditions are true:
468 * 1) It's being executed by the original task, but that task is marked
469 * with PF_EXITING as it's exiting.
470 * 2) PF_KTHREAD is set, in which case the invoker of the task_work is
471 * our fallback task_work.
473 static inline bool io_should_terminate_tw(void)
475 return current
->flags
& (PF_KTHREAD
| PF_EXITING
);
478 static inline void io_req_queue_tw_complete(struct io_kiocb
*req
, s32 res
)
480 io_req_set_res(req
, res
, 0);
481 req
->io_task_work
.func
= io_req_task_complete
;
482 io_req_task_work_add(req
);
486 * IORING_SETUP_SQE128 contexts allocate twice the normal SQE size for each
489 static inline size_t uring_sqe_size(struct io_ring_ctx
*ctx
)
491 if (ctx
->flags
& IORING_SETUP_SQE128
)
492 return 2 * sizeof(struct io_uring_sqe
);
493 return sizeof(struct io_uring_sqe
);
496 static inline bool io_file_can_poll(struct io_kiocb
*req
)
498 if (req
->flags
& REQ_F_CAN_POLL
)
500 if (req
->file
&& file_can_poll(req
->file
)) {
501 req
->flags
|= REQ_F_CAN_POLL
;
507 static inline ktime_t
io_get_time(struct io_ring_ctx
*ctx
)
509 if (ctx
->clockid
== CLOCK_MONOTONIC
)
512 return ktime_get_with_offset(ctx
->clock_offset
);
516 IO_CHECK_CQ_OVERFLOW_BIT
,
517 IO_CHECK_CQ_DROPPED_BIT
,
520 static inline bool io_has_work(struct io_ring_ctx
*ctx
)
522 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
) ||
523 io_local_work_pending(ctx
);