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;
45 #ifdef CONFIG_NET_RX_BUSY_POLL
46 unsigned int napi_busy_poll_to;
47 bool napi_prefer_busy_poll;
51 static inline bool io_should_wake(struct io_wait_queue *iowq)
53 struct io_ring_ctx *ctx = iowq->ctx;
54 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
57 * Wake up if we have enough events, or if a timeout occurred since we
58 * started waiting. For timeouts, we always want to return to userspace,
59 * regardless of event count.
61 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
64 bool io_cqe_cache_refill(struct io_ring_ctx *ctx, bool overflow);
65 int io_run_task_work_sig(struct io_ring_ctx *ctx);
66 void io_req_defer_failed(struct io_kiocb *req, s32 res);
67 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags);
68 bool io_req_post_cqe(struct io_kiocb *req, s32 res, u32 cflags);
69 void __io_commit_cqring_flush(struct io_ring_ctx *ctx);
71 struct file *io_file_get_normal(struct io_kiocb *req, int fd);
72 struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
73 unsigned issue_flags);
75 void __io_req_task_work_add(struct io_kiocb *req, unsigned flags);
76 bool io_alloc_async_data(struct io_kiocb *req);
77 void io_req_task_queue(struct io_kiocb *req);
78 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts);
79 void io_req_task_queue_fail(struct io_kiocb *req, int ret);
80 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts);
81 struct llist_node *io_handle_tw_list(struct llist_node *node, unsigned int *count, unsigned int max_entries);
82 struct llist_node *tctx_task_work_run(struct io_uring_task *tctx, unsigned int max_entries, unsigned int *count);
83 void tctx_task_work(struct callback_head *cb);
84 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
85 int io_uring_alloc_task_context(struct task_struct *task,
86 struct io_ring_ctx *ctx);
88 int io_ring_add_registered_file(struct io_uring_task *tctx, struct file *file,
91 int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts);
92 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr);
93 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin);
94 void __io_submit_flush_completions(struct io_ring_ctx *ctx);
96 struct io_wq_work *io_wq_free_work(struct io_wq_work *work);
97 void io_wq_submit_work(struct io_wq_work *work);
99 void io_free_req(struct io_kiocb *req);
100 void io_queue_next(struct io_kiocb *req);
101 void io_task_refs_refill(struct io_uring_task *tctx);
102 bool __io_alloc_req_refill(struct io_ring_ctx *ctx);
104 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
108 IO_EVENTFD_OP_SIGNAL_BIT,
109 IO_EVENTFD_OP_FREE_BIT,
112 void io_eventfd_ops(struct rcu_head *rcu);
113 void io_activate_pollwq(struct io_ring_ctx *ctx);
115 static inline void io_lockdep_assert_cq_locked(struct io_ring_ctx *ctx)
117 #if defined(CONFIG_PROVE_LOCKING)
118 lockdep_assert(in_task());
120 if (ctx->flags & IORING_SETUP_IOPOLL) {
121 lockdep_assert_held(&ctx->uring_lock);
122 } else if (!ctx->task_complete) {
123 lockdep_assert_held(&ctx->completion_lock);
124 } else if (ctx->submitter_task) {
126 * ->submitter_task may be NULL and we can still post a CQE,
127 * if the ring has been setup with IORING_SETUP_R_DISABLED.
128 * Not from an SQE, as those cannot be submitted, but via
129 * updating tagged resources.
131 if (ctx->submitter_task->flags & PF_EXITING)
132 lockdep_assert(current_work());
134 lockdep_assert(current == ctx->submitter_task);
139 static inline void io_req_task_work_add(struct io_kiocb *req)
141 __io_req_task_work_add(req, 0);
144 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
146 if (!wq_list_empty(&ctx->submit_state.compl_reqs) ||
147 ctx->submit_state.cq_flush)
148 __io_submit_flush_completions(ctx);
151 #define io_for_each_link(pos, head) \
152 for (pos = (head); pos; pos = pos->link)
154 static inline bool io_get_cqe_overflow(struct io_ring_ctx *ctx,
155 struct io_uring_cqe **ret,
158 io_lockdep_assert_cq_locked(ctx);
160 if (unlikely(ctx->cqe_cached >= ctx->cqe_sentinel)) {
161 if (unlikely(!io_cqe_cache_refill(ctx, overflow)))
164 *ret = ctx->cqe_cached;
165 ctx->cached_cq_tail++;
167 if (ctx->flags & IORING_SETUP_CQE32)
172 static inline bool io_get_cqe(struct io_ring_ctx *ctx, struct io_uring_cqe **ret)
174 return io_get_cqe_overflow(ctx, ret, false);
177 static __always_inline bool io_fill_cqe_req(struct io_ring_ctx *ctx,
178 struct io_kiocb *req)
180 struct io_uring_cqe *cqe;
183 * If we can't get a cq entry, userspace overflowed the
184 * submission (by quite a lot). Increment the overflow count in
187 if (unlikely(!io_get_cqe(ctx, &cqe)))
190 if (trace_io_uring_complete_enabled())
191 trace_io_uring_complete(req->ctx, req, req->cqe.user_data,
192 req->cqe.res, req->cqe.flags,
193 req->big_cqe.extra1, req->big_cqe.extra2);
195 memcpy(cqe, &req->cqe, sizeof(*cqe));
196 if (ctx->flags & IORING_SETUP_CQE32) {
197 memcpy(cqe->big_cqe, &req->big_cqe, sizeof(*cqe));
198 memset(&req->big_cqe, 0, sizeof(req->big_cqe));
203 static inline void req_set_fail(struct io_kiocb *req)
205 req->flags |= REQ_F_FAIL;
206 if (req->flags & REQ_F_CQE_SKIP) {
207 req->flags &= ~REQ_F_CQE_SKIP;
208 req->flags |= REQ_F_SKIP_LINK_CQES;
212 static inline void io_req_set_res(struct io_kiocb *req, s32 res, u32 cflags)
215 req->cqe.flags = cflags;
218 static inline bool req_has_async_data(struct io_kiocb *req)
220 return req->flags & REQ_F_ASYNC_DATA;
223 static inline void io_put_file(struct io_kiocb *req)
225 if (!(req->flags & REQ_F_FIXED_FILE) && req->file)
229 static inline void io_ring_submit_unlock(struct io_ring_ctx *ctx,
230 unsigned issue_flags)
232 lockdep_assert_held(&ctx->uring_lock);
233 if (unlikely(issue_flags & IO_URING_F_UNLOCKED))
234 mutex_unlock(&ctx->uring_lock);
237 static inline void io_ring_submit_lock(struct io_ring_ctx *ctx,
238 unsigned issue_flags)
241 * "Normal" inline submissions always hold the uring_lock, since we
242 * grab it from the system call. Same is true for the SQPOLL offload.
243 * The only exception is when we've detached the request and issue it
244 * from an async worker thread, grab the lock for that case.
246 if (unlikely(issue_flags & IO_URING_F_UNLOCKED))
247 mutex_lock(&ctx->uring_lock);
248 lockdep_assert_held(&ctx->uring_lock);
251 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
253 /* order cqe stores with ring update */
254 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
257 static inline void io_poll_wq_wake(struct io_ring_ctx *ctx)
259 if (wq_has_sleeper(&ctx->poll_wq))
260 __wake_up(&ctx->poll_wq, TASK_NORMAL, 0,
261 poll_to_key(EPOLL_URING_WAKE | EPOLLIN));
264 static inline void io_cqring_wake(struct io_ring_ctx *ctx)
267 * Trigger waitqueue handler on all waiters on our waitqueue. This
268 * won't necessarily wake up all the tasks, io_should_wake() will make
271 * Pass in EPOLLIN|EPOLL_URING_WAKE as the poll wakeup key. The latter
272 * set in the mask so that if we recurse back into our own poll
273 * waitqueue handlers, we know we have a dependency between eventfd or
274 * epoll and should terminate multishot poll at that point.
276 if (wq_has_sleeper(&ctx->cq_wait))
277 __wake_up(&ctx->cq_wait, TASK_NORMAL, 0,
278 poll_to_key(EPOLL_URING_WAKE | EPOLLIN));
281 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
283 struct io_rings *r = ctx->rings;
285 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
288 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
290 struct io_rings *rings = ctx->rings;
291 unsigned int entries;
293 /* make sure SQ entry isn't read before tail */
294 entries = smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
295 return min(entries, ctx->sq_entries);
298 static inline int io_run_task_work(void)
303 * Always check-and-clear the task_work notification signal. With how
304 * signaling works for task_work, we can find it set with nothing to
305 * run. We need to clear it for that case, like get_signal() does.
307 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
308 clear_notify_signal();
310 * PF_IO_WORKER never returns to userspace, so check here if we have
311 * notify work that needs processing.
313 if (current->flags & PF_IO_WORKER) {
314 if (test_thread_flag(TIF_NOTIFY_RESUME)) {
315 __set_current_state(TASK_RUNNING);
316 resume_user_mode_work(NULL);
318 if (current->io_uring) {
319 unsigned int count = 0;
321 tctx_task_work_run(current->io_uring, UINT_MAX, &count);
326 if (task_work_pending(current)) {
327 __set_current_state(TASK_RUNNING);
335 static inline bool io_task_work_pending(struct io_ring_ctx *ctx)
337 return task_work_pending(current) || !llist_empty(&ctx->work_llist);
340 static inline void io_tw_lock(struct io_ring_ctx *ctx, struct io_tw_state *ts)
342 lockdep_assert_held(&ctx->uring_lock);
346 * Don't complete immediately but use deferred completion infrastructure.
347 * Protected by ->uring_lock and can only be used either with
348 * IO_URING_F_COMPLETE_DEFER or inside a tw handler holding the mutex.
350 static inline void io_req_complete_defer(struct io_kiocb *req)
351 __must_hold(&req->ctx->uring_lock)
353 struct io_submit_state *state = &req->ctx->submit_state;
355 lockdep_assert_held(&req->ctx->uring_lock);
357 wq_list_add_tail(&req->comp_list, &state->compl_reqs);
360 static inline void io_commit_cqring_flush(struct io_ring_ctx *ctx)
362 if (unlikely(ctx->off_timeout_used || ctx->drain_active ||
363 ctx->has_evfd || ctx->poll_activated))
364 __io_commit_cqring_flush(ctx);
367 static inline void io_get_task_refs(int nr)
369 struct io_uring_task *tctx = current->io_uring;
371 tctx->cached_refs -= nr;
372 if (unlikely(tctx->cached_refs < 0))
373 io_task_refs_refill(tctx);
376 static inline bool io_req_cache_empty(struct io_ring_ctx *ctx)
378 return !ctx->submit_state.free_list.next;
381 extern struct kmem_cache *req_cachep;
382 extern struct kmem_cache *io_buf_cachep;
384 static inline struct io_kiocb *io_extract_req(struct io_ring_ctx *ctx)
386 struct io_kiocb *req;
388 req = container_of(ctx->submit_state.free_list.next, struct io_kiocb, comp_list);
389 wq_stack_extract(&ctx->submit_state.free_list);
393 static inline bool io_alloc_req(struct io_ring_ctx *ctx, struct io_kiocb **req)
395 if (unlikely(io_req_cache_empty(ctx))) {
396 if (!__io_alloc_req_refill(ctx))
399 *req = io_extract_req(ctx);
403 static inline bool io_allowed_defer_tw_run(struct io_ring_ctx *ctx)
405 return likely(ctx->submitter_task == current);
408 static inline bool io_allowed_run_tw(struct io_ring_ctx *ctx)
410 return likely(!(ctx->flags & IORING_SETUP_DEFER_TASKRUN) ||
411 ctx->submitter_task == current);
414 static inline void io_req_queue_tw_complete(struct io_kiocb *req, s32 res)
416 io_req_set_res(req, res, 0);
417 req->io_task_work.func = io_req_task_complete;
418 io_req_task_work_add(req);
422 * IORING_SETUP_SQE128 contexts allocate twice the normal SQE size for each
425 static inline size_t uring_sqe_size(struct io_ring_ctx *ctx)
427 if (ctx->flags & IORING_SETUP_SQE128)
428 return 2 * sizeof(struct io_uring_sqe);
429 return sizeof(struct io_uring_sqe);
432 static inline bool io_file_can_poll(struct io_kiocb *req)
434 if (req->flags & REQ_F_CAN_POLL)
436 if (file_can_poll(req->file)) {
437 req->flags |= REQ_F_CAN_POLL;
444 IO_CHECK_CQ_OVERFLOW_BIT,
445 IO_CHECK_CQ_DROPPED_BIT,
448 static inline bool io_has_work(struct io_ring_ctx *ctx)
450 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
451 !llist_empty(&ctx->work_llist);