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 bool io_cqe_cache_refill(struct io_ring_ctx *ctx, bool overflow);
39 void io_req_cqe_overflow(struct io_kiocb *req);
40 int io_run_task_work_sig(struct io_ring_ctx *ctx);
41 void io_req_defer_failed(struct io_kiocb *req, s32 res);
42 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags);
43 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags);
44 bool io_fill_cqe_req_aux(struct io_kiocb *req, bool defer, s32 res, u32 cflags);
45 void __io_commit_cqring_flush(struct io_ring_ctx *ctx);
47 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages);
49 struct file *io_file_get_normal(struct io_kiocb *req, int fd);
50 struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
51 unsigned issue_flags);
53 void __io_req_task_work_add(struct io_kiocb *req, unsigned flags);
54 bool io_alloc_async_data(struct io_kiocb *req);
55 void io_req_task_queue(struct io_kiocb *req);
56 void io_queue_iowq(struct io_kiocb *req, struct io_tw_state *ts_dont_use);
57 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts);
58 void io_req_task_queue_fail(struct io_kiocb *req, int ret);
59 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts);
60 void tctx_task_work(struct callback_head *cb);
61 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
62 int io_uring_alloc_task_context(struct task_struct *task,
63 struct io_ring_ctx *ctx);
65 int io_ring_add_registered_file(struct io_uring_task *tctx, struct file *file,
68 int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts);
69 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr);
70 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin);
71 void __io_submit_flush_completions(struct io_ring_ctx *ctx);
72 int io_req_prep_async(struct io_kiocb *req);
74 struct io_wq_work *io_wq_free_work(struct io_wq_work *work);
75 void io_wq_submit_work(struct io_wq_work *work);
77 void io_free_req(struct io_kiocb *req);
78 void io_queue_next(struct io_kiocb *req);
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 void *io_mem_alloc(size_t size);
86 void io_mem_free(void *ptr);
89 IO_EVENTFD_OP_SIGNAL_BIT,
90 IO_EVENTFD_OP_FREE_BIT,
93 void io_eventfd_ops(struct rcu_head *rcu);
94 void io_activate_pollwq(struct io_ring_ctx *ctx);
96 #if defined(CONFIG_PROVE_LOCKING)
97 static inline void io_lockdep_assert_cq_locked(struct io_ring_ctx *ctx)
99 lockdep_assert(in_task());
101 if (ctx->flags & IORING_SETUP_IOPOLL) {
102 lockdep_assert_held(&ctx->uring_lock);
103 } else if (!ctx->task_complete) {
104 lockdep_assert_held(&ctx->completion_lock);
105 } else if (ctx->submitter_task) {
107 * ->submitter_task may be NULL and we can still post a CQE,
108 * if the ring has been setup with IORING_SETUP_R_DISABLED.
109 * Not from an SQE, as those cannot be submitted, but via
110 * updating tagged resources.
112 if (ctx->submitter_task->flags & PF_EXITING)
113 lockdep_assert(current_work());
115 lockdep_assert(current == ctx->submitter_task);
119 static inline void io_lockdep_assert_cq_locked(struct io_ring_ctx *ctx)
124 static inline void io_req_task_work_add(struct io_kiocb *req)
126 __io_req_task_work_add(req, 0);
129 #define io_for_each_link(pos, head) \
130 for (pos = (head); pos; pos = pos->link)
132 static inline bool io_get_cqe_overflow(struct io_ring_ctx *ctx,
133 struct io_uring_cqe **ret,
136 io_lockdep_assert_cq_locked(ctx);
138 if (unlikely(ctx->cqe_cached >= ctx->cqe_sentinel)) {
139 if (unlikely(!io_cqe_cache_refill(ctx, overflow)))
142 *ret = ctx->cqe_cached;
143 ctx->cached_cq_tail++;
145 if (ctx->flags & IORING_SETUP_CQE32)
150 static inline bool io_get_cqe(struct io_ring_ctx *ctx, struct io_uring_cqe **ret)
152 return io_get_cqe_overflow(ctx, ret, false);
155 static __always_inline bool io_fill_cqe_req(struct io_ring_ctx *ctx,
156 struct io_kiocb *req)
158 struct io_uring_cqe *cqe;
161 * If we can't get a cq entry, userspace overflowed the
162 * submission (by quite a lot). Increment the overflow count in
165 if (unlikely(!io_get_cqe(ctx, &cqe)))
168 if (trace_io_uring_complete_enabled())
169 trace_io_uring_complete(req->ctx, req, req->cqe.user_data,
170 req->cqe.res, req->cqe.flags,
171 req->big_cqe.extra1, req->big_cqe.extra2);
173 memcpy(cqe, &req->cqe, sizeof(*cqe));
174 if (ctx->flags & IORING_SETUP_CQE32) {
175 memcpy(cqe->big_cqe, &req->big_cqe, sizeof(*cqe));
176 memset(&req->big_cqe, 0, sizeof(req->big_cqe));
181 static inline void req_set_fail(struct io_kiocb *req)
183 req->flags |= REQ_F_FAIL;
184 if (req->flags & REQ_F_CQE_SKIP) {
185 req->flags &= ~REQ_F_CQE_SKIP;
186 req->flags |= REQ_F_SKIP_LINK_CQES;
190 static inline void io_req_set_res(struct io_kiocb *req, s32 res, u32 cflags)
193 req->cqe.flags = cflags;
196 static inline bool req_has_async_data(struct io_kiocb *req)
198 return req->flags & REQ_F_ASYNC_DATA;
201 static inline void io_put_file(struct io_kiocb *req)
203 if (!(req->flags & REQ_F_FIXED_FILE) && req->file)
207 static inline void io_ring_submit_unlock(struct io_ring_ctx *ctx,
208 unsigned issue_flags)
210 lockdep_assert_held(&ctx->uring_lock);
211 if (issue_flags & IO_URING_F_UNLOCKED)
212 mutex_unlock(&ctx->uring_lock);
215 static inline void io_ring_submit_lock(struct io_ring_ctx *ctx,
216 unsigned issue_flags)
219 * "Normal" inline submissions always hold the uring_lock, since we
220 * grab it from the system call. Same is true for the SQPOLL offload.
221 * The only exception is when we've detached the request and issue it
222 * from an async worker thread, grab the lock for that case.
224 if (issue_flags & IO_URING_F_UNLOCKED)
225 mutex_lock(&ctx->uring_lock);
226 lockdep_assert_held(&ctx->uring_lock);
229 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
231 /* order cqe stores with ring update */
232 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
235 static inline void io_poll_wq_wake(struct io_ring_ctx *ctx)
237 if (wq_has_sleeper(&ctx->poll_wq))
238 __wake_up(&ctx->poll_wq, TASK_NORMAL, 0,
239 poll_to_key(EPOLL_URING_WAKE | EPOLLIN));
242 static inline void io_cqring_wake(struct io_ring_ctx *ctx)
245 * Trigger waitqueue handler on all waiters on our waitqueue. This
246 * won't necessarily wake up all the tasks, io_should_wake() will make
249 * Pass in EPOLLIN|EPOLL_URING_WAKE as the poll wakeup key. The latter
250 * set in the mask so that if we recurse back into our own poll
251 * waitqueue handlers, we know we have a dependency between eventfd or
252 * epoll and should terminate multishot poll at that point.
254 if (wq_has_sleeper(&ctx->cq_wait))
255 __wake_up(&ctx->cq_wait, TASK_NORMAL, 0,
256 poll_to_key(EPOLL_URING_WAKE | EPOLLIN));
259 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
261 struct io_rings *r = ctx->rings;
263 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
266 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
268 struct io_rings *rings = ctx->rings;
269 unsigned int entries;
271 /* make sure SQ entry isn't read before tail */
272 entries = smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
273 return min(entries, ctx->sq_entries);
276 static inline int io_run_task_work(void)
279 * Always check-and-clear the task_work notification signal. With how
280 * signaling works for task_work, we can find it set with nothing to
281 * run. We need to clear it for that case, like get_signal() does.
283 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
284 clear_notify_signal();
286 * PF_IO_WORKER never returns to userspace, so check here if we have
287 * notify work that needs processing.
289 if (current->flags & PF_IO_WORKER &&
290 test_thread_flag(TIF_NOTIFY_RESUME)) {
291 __set_current_state(TASK_RUNNING);
292 resume_user_mode_work(NULL);
294 if (task_work_pending(current)) {
295 __set_current_state(TASK_RUNNING);
303 static inline bool io_task_work_pending(struct io_ring_ctx *ctx)
305 return task_work_pending(current) || !wq_list_empty(&ctx->work_llist);
308 static inline void io_tw_lock(struct io_ring_ctx *ctx, struct io_tw_state *ts)
311 mutex_lock(&ctx->uring_lock);
317 * Don't complete immediately but use deferred completion infrastructure.
318 * Protected by ->uring_lock and can only be used either with
319 * IO_URING_F_COMPLETE_DEFER or inside a tw handler holding the mutex.
321 static inline void io_req_complete_defer(struct io_kiocb *req)
322 __must_hold(&req->ctx->uring_lock)
324 struct io_submit_state *state = &req->ctx->submit_state;
326 lockdep_assert_held(&req->ctx->uring_lock);
328 wq_list_add_tail(&req->comp_list, &state->compl_reqs);
331 static inline void io_commit_cqring_flush(struct io_ring_ctx *ctx)
333 if (unlikely(ctx->off_timeout_used || ctx->drain_active ||
334 ctx->has_evfd || ctx->poll_activated))
335 __io_commit_cqring_flush(ctx);
338 static inline void io_get_task_refs(int nr)
340 struct io_uring_task *tctx = current->io_uring;
342 tctx->cached_refs -= nr;
343 if (unlikely(tctx->cached_refs < 0))
344 io_task_refs_refill(tctx);
347 static inline bool io_req_cache_empty(struct io_ring_ctx *ctx)
349 return !ctx->submit_state.free_list.next;
352 extern struct kmem_cache *req_cachep;
353 extern struct kmem_cache *io_buf_cachep;
355 static inline struct io_kiocb *io_extract_req(struct io_ring_ctx *ctx)
357 struct io_kiocb *req;
359 req = container_of(ctx->submit_state.free_list.next, struct io_kiocb, comp_list);
360 wq_stack_extract(&ctx->submit_state.free_list);
364 static inline bool io_alloc_req(struct io_ring_ctx *ctx, struct io_kiocb **req)
366 if (unlikely(io_req_cache_empty(ctx))) {
367 if (!__io_alloc_req_refill(ctx))
370 *req = io_extract_req(ctx);
374 static inline bool io_allowed_defer_tw_run(struct io_ring_ctx *ctx)
376 return likely(ctx->submitter_task == current);
379 static inline bool io_allowed_run_tw(struct io_ring_ctx *ctx)
381 return likely(!(ctx->flags & IORING_SETUP_DEFER_TASKRUN) ||
382 ctx->submitter_task == current);
385 static inline void io_req_queue_tw_complete(struct io_kiocb *req, s32 res)
387 io_req_set_res(req, res, 0);
388 req->io_task_work.func = io_req_task_complete;
389 io_req_task_work_add(req);
393 * IORING_SETUP_SQE128 contexts allocate twice the normal SQE size for each
396 static inline size_t uring_sqe_size(struct io_ring_ctx *ctx)
398 if (ctx->flags & IORING_SETUP_SQE128)
399 return 2 * sizeof(struct io_uring_sqe);
400 return sizeof(struct io_uring_sqe);
403 static inline bool io_file_can_poll(struct io_kiocb *req)
405 if (req->flags & REQ_F_CAN_POLL)
407 if (file_can_poll(req->file)) {
408 req->flags |= REQ_F_CAN_POLL;