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 * A hint to not wake right away but delay until there are enough of
21 * tw's queued to match the number of CQEs the task is waiting for.
23 * Must not be used wirh requests generating more than one CQE.
24 * It's also ignored unless IORING_SETUP_DEFER_TASKRUN is set.
26 IOU_F_TWQ_LAZY_WAKE = 1,
31 IOU_ISSUE_SKIP_COMPLETE = -EIOCBQUEUED,
34 * Intended only when both IO_URING_F_MULTISHOT is passed
35 * to indicate to the poll runner that multishot should be
36 * removed and the result is set on req->cqe.res.
38 IOU_STOP_MULTISHOT = -ECANCELED,
41 bool io_cqe_cache_refill(struct io_ring_ctx *ctx, bool overflow);
42 void io_req_cqe_overflow(struct io_kiocb *req);
43 int io_run_task_work_sig(struct io_ring_ctx *ctx);
44 void io_req_defer_failed(struct io_kiocb *req, s32 res);
45 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags);
46 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags);
47 bool io_fill_cqe_req_aux(struct io_kiocb *req, bool defer, s32 res, u32 cflags);
48 void __io_commit_cqring_flush(struct io_ring_ctx *ctx);
50 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages);
52 struct file *io_file_get_normal(struct io_kiocb *req, int fd);
53 struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
54 unsigned issue_flags);
56 void __io_req_task_work_add(struct io_kiocb *req, unsigned flags);
57 bool io_is_uring_fops(struct file *file);
58 bool io_alloc_async_data(struct io_kiocb *req);
59 void io_req_task_queue(struct io_kiocb *req);
60 void io_queue_iowq(struct io_kiocb *req, struct io_tw_state *ts_dont_use);
61 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts);
62 void io_req_task_queue_fail(struct io_kiocb *req, int ret);
63 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts);
64 void tctx_task_work(struct callback_head *cb);
65 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
66 int io_uring_alloc_task_context(struct task_struct *task,
67 struct io_ring_ctx *ctx);
69 int io_ring_add_registered_file(struct io_uring_task *tctx, struct file *file,
72 int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts);
73 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr);
74 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin);
75 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node);
76 int io_req_prep_async(struct io_kiocb *req);
78 struct io_wq_work *io_wq_free_work(struct io_wq_work *work);
79 void io_wq_submit_work(struct io_wq_work *work);
81 void io_free_req(struct io_kiocb *req);
82 void io_queue_next(struct io_kiocb *req);
83 void io_task_refs_refill(struct io_uring_task *tctx);
84 bool __io_alloc_req_refill(struct io_ring_ctx *ctx);
86 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
89 #define io_lockdep_assert_cq_locked(ctx) \
91 lockdep_assert(in_task()); \
93 if (ctx->flags & IORING_SETUP_IOPOLL) { \
94 lockdep_assert_held(&ctx->uring_lock); \
95 } else if (!ctx->task_complete) { \
96 lockdep_assert_held(&ctx->completion_lock); \
97 } else if (ctx->submitter_task->flags & PF_EXITING) { \
98 lockdep_assert(current_work()); \
100 lockdep_assert(current == ctx->submitter_task); \
104 static inline void io_req_task_work_add(struct io_kiocb *req)
106 __io_req_task_work_add(req, 0);
109 #define io_for_each_link(pos, head) \
110 for (pos = (head); pos; pos = pos->link)
112 static inline struct io_uring_cqe *io_get_cqe_overflow(struct io_ring_ctx *ctx,
115 struct io_uring_cqe *cqe;
117 io_lockdep_assert_cq_locked(ctx);
119 if (unlikely(ctx->cqe_cached >= ctx->cqe_sentinel)) {
120 if (unlikely(!io_cqe_cache_refill(ctx, overflow)))
123 cqe = ctx->cqe_cached;
124 ctx->cached_cq_tail++;
126 if (ctx->flags & IORING_SETUP_CQE32)
131 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
133 return io_get_cqe_overflow(ctx, false);
136 static inline bool io_fill_cqe_req(struct io_ring_ctx *ctx, struct io_kiocb *req)
138 struct io_uring_cqe *cqe;
141 * If we can't get a cq entry, userspace overflowed the
142 * submission (by quite a lot). Increment the overflow count in
145 cqe = io_get_cqe(ctx);
149 if (trace_io_uring_complete_enabled())
150 trace_io_uring_complete(req->ctx, req, req->cqe.user_data,
151 req->cqe.res, req->cqe.flags,
152 req->big_cqe.extra1, req->big_cqe.extra2);
154 memcpy(cqe, &req->cqe, sizeof(*cqe));
155 if (ctx->flags & IORING_SETUP_CQE32) {
156 memcpy(cqe->big_cqe, &req->big_cqe, sizeof(*cqe));
157 memset(&req->big_cqe, 0, sizeof(req->big_cqe));
162 static inline void req_set_fail(struct io_kiocb *req)
164 req->flags |= REQ_F_FAIL;
165 if (req->flags & REQ_F_CQE_SKIP) {
166 req->flags &= ~REQ_F_CQE_SKIP;
167 req->flags |= REQ_F_SKIP_LINK_CQES;
171 static inline void io_req_set_res(struct io_kiocb *req, s32 res, u32 cflags)
174 req->cqe.flags = cflags;
177 static inline bool req_has_async_data(struct io_kiocb *req)
179 return req->flags & REQ_F_ASYNC_DATA;
182 static inline void io_put_file(struct io_kiocb *req)
184 if (!(req->flags & REQ_F_FIXED_FILE) && req->file)
188 static inline void io_ring_submit_unlock(struct io_ring_ctx *ctx,
189 unsigned issue_flags)
191 lockdep_assert_held(&ctx->uring_lock);
192 if (issue_flags & IO_URING_F_UNLOCKED)
193 mutex_unlock(&ctx->uring_lock);
196 static inline void io_ring_submit_lock(struct io_ring_ctx *ctx,
197 unsigned issue_flags)
200 * "Normal" inline submissions always hold the uring_lock, since we
201 * grab it from the system call. Same is true for the SQPOLL offload.
202 * The only exception is when we've detached the request and issue it
203 * from an async worker thread, grab the lock for that case.
205 if (issue_flags & IO_URING_F_UNLOCKED)
206 mutex_lock(&ctx->uring_lock);
207 lockdep_assert_held(&ctx->uring_lock);
210 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
212 /* order cqe stores with ring update */
213 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
216 static inline void io_poll_wq_wake(struct io_ring_ctx *ctx)
218 if (wq_has_sleeper(&ctx->poll_wq))
219 __wake_up(&ctx->poll_wq, TASK_NORMAL, 0,
220 poll_to_key(EPOLL_URING_WAKE | EPOLLIN));
223 static inline void io_cqring_wake(struct io_ring_ctx *ctx)
226 * Trigger waitqueue handler on all waiters on our waitqueue. This
227 * won't necessarily wake up all the tasks, io_should_wake() will make
230 * Pass in EPOLLIN|EPOLL_URING_WAKE as the poll wakeup key. The latter
231 * set in the mask so that if we recurse back into our own poll
232 * waitqueue handlers, we know we have a dependency between eventfd or
233 * epoll and should terminate multishot poll at that point.
235 if (wq_has_sleeper(&ctx->cq_wait))
236 __wake_up(&ctx->cq_wait, TASK_NORMAL, 0,
237 poll_to_key(EPOLL_URING_WAKE | EPOLLIN));
240 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
242 struct io_rings *r = ctx->rings;
244 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
247 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
249 struct io_rings *rings = ctx->rings;
250 unsigned int entries;
252 /* make sure SQ entry isn't read before tail */
253 entries = smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
254 return min(entries, ctx->sq_entries);
257 static inline int io_run_task_work(void)
260 * Always check-and-clear the task_work notification signal. With how
261 * signaling works for task_work, we can find it set with nothing to
262 * run. We need to clear it for that case, like get_signal() does.
264 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
265 clear_notify_signal();
267 * PF_IO_WORKER never returns to userspace, so check here if we have
268 * notify work that needs processing.
270 if (current->flags & PF_IO_WORKER &&
271 test_thread_flag(TIF_NOTIFY_RESUME)) {
272 __set_current_state(TASK_RUNNING);
273 resume_user_mode_work(NULL);
275 if (task_work_pending(current)) {
276 __set_current_state(TASK_RUNNING);
284 static inline bool io_task_work_pending(struct io_ring_ctx *ctx)
286 return task_work_pending(current) || !wq_list_empty(&ctx->work_llist);
289 static inline void io_tw_lock(struct io_ring_ctx *ctx, struct io_tw_state *ts)
292 mutex_lock(&ctx->uring_lock);
298 * Don't complete immediately but use deferred completion infrastructure.
299 * Protected by ->uring_lock and can only be used either with
300 * IO_URING_F_COMPLETE_DEFER or inside a tw handler holding the mutex.
302 static inline void io_req_complete_defer(struct io_kiocb *req)
303 __must_hold(&req->ctx->uring_lock)
305 struct io_submit_state *state = &req->ctx->submit_state;
307 lockdep_assert_held(&req->ctx->uring_lock);
309 wq_list_add_tail(&req->comp_list, &state->compl_reqs);
312 static inline void io_commit_cqring_flush(struct io_ring_ctx *ctx)
314 if (unlikely(ctx->off_timeout_used || ctx->drain_active ||
315 ctx->has_evfd || ctx->poll_activated))
316 __io_commit_cqring_flush(ctx);
319 static inline void io_get_task_refs(int nr)
321 struct io_uring_task *tctx = current->io_uring;
323 tctx->cached_refs -= nr;
324 if (unlikely(tctx->cached_refs < 0))
325 io_task_refs_refill(tctx);
328 static inline bool io_req_cache_empty(struct io_ring_ctx *ctx)
330 return !ctx->submit_state.free_list.next;
333 extern struct kmem_cache *req_cachep;
335 static inline struct io_kiocb *io_extract_req(struct io_ring_ctx *ctx)
337 struct io_kiocb *req;
339 req = container_of(ctx->submit_state.free_list.next, struct io_kiocb, comp_list);
340 wq_stack_extract(&ctx->submit_state.free_list);
344 static inline bool io_alloc_req(struct io_ring_ctx *ctx, struct io_kiocb **req)
346 if (unlikely(io_req_cache_empty(ctx))) {
347 if (!__io_alloc_req_refill(ctx))
350 *req = io_extract_req(ctx);
354 static inline bool io_allowed_defer_tw_run(struct io_ring_ctx *ctx)
356 return likely(ctx->submitter_task == current);
359 static inline bool io_allowed_run_tw(struct io_ring_ctx *ctx)
361 return likely(!(ctx->flags & IORING_SETUP_DEFER_TASKRUN) ||
362 ctx->submitter_task == current);
365 static inline void io_req_queue_tw_complete(struct io_kiocb *req, s32 res)
367 io_req_set_res(req, res, 0);
368 req->io_task_work.func = io_req_task_complete;
369 io_req_task_work_add(req);
373 * IORING_SETUP_SQE128 contexts allocate twice the normal SQE size for each
376 static inline size_t uring_sqe_size(struct io_ring_ctx *ctx)
378 if (ctx->flags & IORING_SETUP_SQE128)
379 return 2 * sizeof(struct io_uring_sqe);
380 return sizeof(struct io_uring_sqe);