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/io_uring_types.h>
9 #include <uapi/linux/eventpoll.h>
12 #include "filetable.h"
14 #ifndef CREATE_TRACE_POINTS
15 #include <trace/events/io_uring.h>
20 IOU_ISSUE_SKIP_COMPLETE = -EIOCBQUEUED,
23 * Requeue the task_work to restart operations on this request. The
24 * actual value isn't important, should just be not an otherwise
25 * valid error code, yet less than -MAX_ERRNO and valid internally.
30 * Intended only when both IO_URING_F_MULTISHOT is passed
31 * to indicate to the poll runner that multishot should be
32 * removed and the result is set on req->cqe.res.
34 IOU_STOP_MULTISHOT = -ECANCELED,
37 bool io_cqe_cache_refill(struct io_ring_ctx *ctx, bool overflow);
38 void io_req_cqe_overflow(struct io_kiocb *req);
39 int io_run_task_work_sig(struct io_ring_ctx *ctx);
40 void io_req_defer_failed(struct io_kiocb *req, s32 res);
41 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags);
42 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags);
43 bool io_fill_cqe_req_aux(struct io_kiocb *req, bool defer, s32 res, u32 cflags);
44 void __io_commit_cqring_flush(struct io_ring_ctx *ctx);
46 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages);
48 struct file *io_file_get_normal(struct io_kiocb *req, int fd);
49 struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
50 unsigned issue_flags);
52 void __io_req_task_work_add(struct io_kiocb *req, unsigned flags);
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, struct io_tw_state *ts_dont_use);
56 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts);
57 void io_req_task_queue_fail(struct io_kiocb *req, int ret);
58 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts);
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_ring_add_registered_file(struct io_uring_task *tctx, struct file *file,
67 int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts);
68 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr);
69 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin);
70 void __io_submit_flush_completions(struct io_ring_ctx *ctx);
71 int io_req_prep_async(struct io_kiocb *req);
73 struct io_wq_work *io_wq_free_work(struct io_wq_work *work);
74 void io_wq_submit_work(struct io_wq_work *work);
76 void io_free_req(struct io_kiocb *req);
77 void io_queue_next(struct io_kiocb *req);
78 void io_task_refs_refill(struct io_uring_task *tctx);
79 bool __io_alloc_req_refill(struct io_ring_ctx *ctx);
81 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
84 void *io_mem_alloc(size_t size);
85 void io_mem_free(void *ptr);
88 IO_EVENTFD_OP_SIGNAL_BIT,
89 IO_EVENTFD_OP_FREE_BIT,
92 void io_eventfd_ops(struct rcu_head *rcu);
93 void io_activate_pollwq(struct io_ring_ctx *ctx);
95 #if defined(CONFIG_PROVE_LOCKING)
96 static inline void io_lockdep_assert_cq_locked(struct io_ring_ctx *ctx)
98 lockdep_assert(in_task());
100 if (ctx->flags & IORING_SETUP_IOPOLL) {
101 lockdep_assert_held(&ctx->uring_lock);
102 } else if (!ctx->task_complete) {
103 lockdep_assert_held(&ctx->completion_lock);
104 } else if (ctx->submitter_task) {
106 * ->submitter_task may be NULL and we can still post a CQE,
107 * if the ring has been setup with IORING_SETUP_R_DISABLED.
108 * Not from an SQE, as those cannot be submitted, but via
109 * updating tagged resources.
111 if (ctx->submitter_task->flags & PF_EXITING)
112 lockdep_assert(current_work());
114 lockdep_assert(current == ctx->submitter_task);
118 static inline void io_lockdep_assert_cq_locked(struct io_ring_ctx *ctx)
123 static inline void io_req_task_work_add(struct io_kiocb *req)
125 __io_req_task_work_add(req, 0);
128 #define io_for_each_link(pos, head) \
129 for (pos = (head); pos; pos = pos->link)
131 static inline bool io_get_cqe_overflow(struct io_ring_ctx *ctx,
132 struct io_uring_cqe **ret,
135 io_lockdep_assert_cq_locked(ctx);
137 if (unlikely(ctx->cqe_cached >= ctx->cqe_sentinel)) {
138 if (unlikely(!io_cqe_cache_refill(ctx, overflow)))
141 *ret = ctx->cqe_cached;
142 ctx->cached_cq_tail++;
144 if (ctx->flags & IORING_SETUP_CQE32)
149 static inline bool io_get_cqe(struct io_ring_ctx *ctx, struct io_uring_cqe **ret)
151 return io_get_cqe_overflow(ctx, ret, false);
154 static __always_inline bool io_fill_cqe_req(struct io_ring_ctx *ctx,
155 struct io_kiocb *req)
157 struct io_uring_cqe *cqe;
160 * If we can't get a cq entry, userspace overflowed the
161 * submission (by quite a lot). Increment the overflow count in
164 if (unlikely(!io_get_cqe(ctx, &cqe)))
167 if (trace_io_uring_complete_enabled())
168 trace_io_uring_complete(req->ctx, req, req->cqe.user_data,
169 req->cqe.res, req->cqe.flags,
170 req->big_cqe.extra1, req->big_cqe.extra2);
172 memcpy(cqe, &req->cqe, sizeof(*cqe));
173 if (ctx->flags & IORING_SETUP_CQE32) {
174 memcpy(cqe->big_cqe, &req->big_cqe, sizeof(*cqe));
175 memset(&req->big_cqe, 0, sizeof(req->big_cqe));
180 static inline void req_set_fail(struct io_kiocb *req)
182 req->flags |= REQ_F_FAIL;
183 if (req->flags & REQ_F_CQE_SKIP) {
184 req->flags &= ~REQ_F_CQE_SKIP;
185 req->flags |= REQ_F_SKIP_LINK_CQES;
189 static inline void io_req_set_res(struct io_kiocb *req, s32 res, u32 cflags)
192 req->cqe.flags = cflags;
195 static inline bool req_has_async_data(struct io_kiocb *req)
197 return req->flags & REQ_F_ASYNC_DATA;
200 static inline void io_put_file(struct io_kiocb *req)
202 if (!(req->flags & REQ_F_FIXED_FILE) && req->file)
206 static inline void io_ring_submit_unlock(struct io_ring_ctx *ctx,
207 unsigned issue_flags)
209 lockdep_assert_held(&ctx->uring_lock);
210 if (issue_flags & IO_URING_F_UNLOCKED)
211 mutex_unlock(&ctx->uring_lock);
214 static inline void io_ring_submit_lock(struct io_ring_ctx *ctx,
215 unsigned issue_flags)
218 * "Normal" inline submissions always hold the uring_lock, since we
219 * grab it from the system call. Same is true for the SQPOLL offload.
220 * The only exception is when we've detached the request and issue it
221 * from an async worker thread, grab the lock for that case.
223 if (issue_flags & IO_URING_F_UNLOCKED)
224 mutex_lock(&ctx->uring_lock);
225 lockdep_assert_held(&ctx->uring_lock);
228 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
230 /* order cqe stores with ring update */
231 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
234 static inline void io_poll_wq_wake(struct io_ring_ctx *ctx)
236 if (wq_has_sleeper(&ctx->poll_wq))
237 __wake_up(&ctx->poll_wq, TASK_NORMAL, 0,
238 poll_to_key(EPOLL_URING_WAKE | EPOLLIN));
241 static inline void io_cqring_wake(struct io_ring_ctx *ctx)
244 * Trigger waitqueue handler on all waiters on our waitqueue. This
245 * won't necessarily wake up all the tasks, io_should_wake() will make
248 * Pass in EPOLLIN|EPOLL_URING_WAKE as the poll wakeup key. The latter
249 * set in the mask so that if we recurse back into our own poll
250 * waitqueue handlers, we know we have a dependency between eventfd or
251 * epoll and should terminate multishot poll at that point.
253 if (wq_has_sleeper(&ctx->cq_wait))
254 __wake_up(&ctx->cq_wait, TASK_NORMAL, 0,
255 poll_to_key(EPOLL_URING_WAKE | EPOLLIN));
258 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
260 struct io_rings *r = ctx->rings;
262 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
265 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
267 struct io_rings *rings = ctx->rings;
268 unsigned int entries;
270 /* make sure SQ entry isn't read before tail */
271 entries = smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
272 return min(entries, ctx->sq_entries);
275 static inline int io_run_task_work(void)
278 * Always check-and-clear the task_work notification signal. With how
279 * signaling works for task_work, we can find it set with nothing to
280 * run. We need to clear it for that case, like get_signal() does.
282 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
283 clear_notify_signal();
285 * PF_IO_WORKER never returns to userspace, so check here if we have
286 * notify work that needs processing.
288 if (current->flags & PF_IO_WORKER &&
289 test_thread_flag(TIF_NOTIFY_RESUME)) {
290 __set_current_state(TASK_RUNNING);
291 resume_user_mode_work(NULL);
293 if (task_work_pending(current)) {
294 __set_current_state(TASK_RUNNING);
302 static inline bool io_task_work_pending(struct io_ring_ctx *ctx)
304 return task_work_pending(current) || !wq_list_empty(&ctx->work_llist);
307 static inline void io_tw_lock(struct io_ring_ctx *ctx, struct io_tw_state *ts)
310 mutex_lock(&ctx->uring_lock);
316 * Don't complete immediately but use deferred completion infrastructure.
317 * Protected by ->uring_lock and can only be used either with
318 * IO_URING_F_COMPLETE_DEFER or inside a tw handler holding the mutex.
320 static inline void io_req_complete_defer(struct io_kiocb *req)
321 __must_hold(&req->ctx->uring_lock)
323 struct io_submit_state *state = &req->ctx->submit_state;
325 lockdep_assert_held(&req->ctx->uring_lock);
327 wq_list_add_tail(&req->comp_list, &state->compl_reqs);
330 static inline void io_commit_cqring_flush(struct io_ring_ctx *ctx)
332 if (unlikely(ctx->off_timeout_used || ctx->drain_active ||
333 ctx->has_evfd || ctx->poll_activated))
334 __io_commit_cqring_flush(ctx);
337 static inline void io_get_task_refs(int nr)
339 struct io_uring_task *tctx = current->io_uring;
341 tctx->cached_refs -= nr;
342 if (unlikely(tctx->cached_refs < 0))
343 io_task_refs_refill(tctx);
346 static inline bool io_req_cache_empty(struct io_ring_ctx *ctx)
348 return !ctx->submit_state.free_list.next;
351 extern struct kmem_cache *req_cachep;
352 extern struct kmem_cache *io_buf_cachep;
354 static inline struct io_kiocb *io_extract_req(struct io_ring_ctx *ctx)
356 struct io_kiocb *req;
358 req = container_of(ctx->submit_state.free_list.next, struct io_kiocb, comp_list);
359 wq_stack_extract(&ctx->submit_state.free_list);
363 static inline bool io_alloc_req(struct io_ring_ctx *ctx, struct io_kiocb **req)
365 if (unlikely(io_req_cache_empty(ctx))) {
366 if (!__io_alloc_req_refill(ctx))
369 *req = io_extract_req(ctx);
373 static inline bool io_allowed_defer_tw_run(struct io_ring_ctx *ctx)
375 return likely(ctx->submitter_task == current);
378 static inline bool io_allowed_run_tw(struct io_ring_ctx *ctx)
380 return likely(!(ctx->flags & IORING_SETUP_DEFER_TASKRUN) ||
381 ctx->submitter_task == current);
384 static inline void io_req_queue_tw_complete(struct io_kiocb *req, s32 res)
386 io_req_set_res(req, res, 0);
387 req->io_task_work.func = io_req_task_complete;
388 io_req_task_work_add(req);
392 * IORING_SETUP_SQE128 contexts allocate twice the normal SQE size for each
395 static inline size_t uring_sqe_size(struct io_ring_ctx *ctx)
397 if (ctx->flags & IORING_SETUP_SQE128)
398 return 2 * sizeof(struct io_uring_sqe);
399 return sizeof(struct io_uring_sqe);