xref: /linux/io_uring/io_uring.h (revision c1e822754cc7f28b98c6897d62e8b47b4001e422)
1 #ifndef IOU_CORE_H
2 #define IOU_CORE_H
3 
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 "io-wq.h"
12 #include "slist.h"
13 #include "filetable.h"
14 
15 #ifndef CREATE_TRACE_POINTS
16 #include <trace/events/io_uring.h>
17 #endif
18 
19 enum {
20 	IOU_OK			= 0,
21 	IOU_ISSUE_SKIP_COMPLETE	= -EIOCBQUEUED,
22 
23 	/*
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.
27 	 */
28 	IOU_REQUEUE		= -3072,
29 
30 	/*
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.
34 	 */
35 	IOU_STOP_MULTISHOT	= -ECANCELED,
36 };
37 
38 struct io_wait_queue {
39 	struct wait_queue_entry wq;
40 	struct io_ring_ctx *ctx;
41 	unsigned cq_tail;
42 	unsigned cq_min_tail;
43 	unsigned nr_timeouts;
44 	int hit_timeout;
45 	ktime_t min_timeout;
46 	ktime_t timeout;
47 	struct hrtimer t;
48 
49 #ifdef CONFIG_NET_RX_BUSY_POLL
50 	ktime_t napi_busy_poll_dt;
51 	bool napi_prefer_busy_poll;
52 #endif
53 };
54 
55 static inline bool io_should_wake(struct io_wait_queue *iowq)
56 {
57 	struct io_ring_ctx *ctx = iowq->ctx;
58 	int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
59 
60 	/*
61 	 * Wake up if we have enough events, or if a timeout occurred since we
62 	 * started waiting. For timeouts, we always want to return to userspace,
63 	 * regardless of event count.
64 	 */
65 	return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
66 }
67 
68 bool io_cqe_cache_refill(struct io_ring_ctx *ctx, bool overflow);
69 int io_run_task_work_sig(struct io_ring_ctx *ctx);
70 void io_req_defer_failed(struct io_kiocb *req, s32 res);
71 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags);
72 void io_add_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags);
73 bool io_req_post_cqe(struct io_kiocb *req, s32 res, u32 cflags);
74 void __io_commit_cqring_flush(struct io_ring_ctx *ctx);
75 
76 struct file *io_file_get_normal(struct io_kiocb *req, int fd);
77 struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
78 			       unsigned issue_flags);
79 
80 void __io_req_task_work_add(struct io_kiocb *req, unsigned flags);
81 void io_req_task_work_add_remote(struct io_kiocb *req, struct io_ring_ctx *ctx,
82 				 unsigned flags);
83 bool io_alloc_async_data(struct io_kiocb *req);
84 void io_req_task_queue(struct io_kiocb *req);
85 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts);
86 void io_req_task_queue_fail(struct io_kiocb *req, int ret);
87 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts);
88 struct llist_node *io_handle_tw_list(struct llist_node *node, unsigned int *count, unsigned int max_entries);
89 struct llist_node *tctx_task_work_run(struct io_uring_task *tctx, unsigned int max_entries, unsigned int *count);
90 void tctx_task_work(struct callback_head *cb);
91 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
92 int io_uring_alloc_task_context(struct task_struct *task,
93 				struct io_ring_ctx *ctx);
94 
95 int io_ring_add_registered_file(struct io_uring_task *tctx, struct file *file,
96 				     int start, int end);
97 void io_req_queue_iowq(struct io_kiocb *req);
98 
99 int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts);
100 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr);
101 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin);
102 void __io_submit_flush_completions(struct io_ring_ctx *ctx);
103 
104 struct io_wq_work *io_wq_free_work(struct io_wq_work *work);
105 void io_wq_submit_work(struct io_wq_work *work);
106 
107 void io_free_req(struct io_kiocb *req);
108 void io_queue_next(struct io_kiocb *req);
109 void io_task_refs_refill(struct io_uring_task *tctx);
110 bool __io_alloc_req_refill(struct io_ring_ctx *ctx);
111 
112 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
113 			bool cancel_all);
114 
115 void io_activate_pollwq(struct io_ring_ctx *ctx);
116 
117 static inline void io_lockdep_assert_cq_locked(struct io_ring_ctx *ctx)
118 {
119 #if defined(CONFIG_PROVE_LOCKING)
120 	lockdep_assert(in_task());
121 
122 	if (ctx->flags & IORING_SETUP_IOPOLL) {
123 		lockdep_assert_held(&ctx->uring_lock);
124 	} else if (!ctx->task_complete) {
125 		lockdep_assert_held(&ctx->completion_lock);
126 	} else if (ctx->submitter_task) {
127 		/*
128 		 * ->submitter_task may be NULL and we can still post a CQE,
129 		 * if the ring has been setup with IORING_SETUP_R_DISABLED.
130 		 * Not from an SQE, as those cannot be submitted, but via
131 		 * updating tagged resources.
132 		 */
133 		if (ctx->submitter_task->flags & PF_EXITING)
134 			lockdep_assert(current_work());
135 		else
136 			lockdep_assert(current == ctx->submitter_task);
137 	}
138 #endif
139 }
140 
141 static inline void io_req_task_work_add(struct io_kiocb *req)
142 {
143 	__io_req_task_work_add(req, 0);
144 }
145 
146 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
147 {
148 	if (!wq_list_empty(&ctx->submit_state.compl_reqs) ||
149 	    ctx->submit_state.cq_flush)
150 		__io_submit_flush_completions(ctx);
151 }
152 
153 #define io_for_each_link(pos, head) \
154 	for (pos = (head); pos; pos = pos->link)
155 
156 static inline bool io_get_cqe_overflow(struct io_ring_ctx *ctx,
157 					struct io_uring_cqe **ret,
158 					bool overflow)
159 {
160 	io_lockdep_assert_cq_locked(ctx);
161 
162 	if (unlikely(ctx->cqe_cached >= ctx->cqe_sentinel)) {
163 		if (unlikely(!io_cqe_cache_refill(ctx, overflow)))
164 			return false;
165 	}
166 	*ret = ctx->cqe_cached;
167 	ctx->cached_cq_tail++;
168 	ctx->cqe_cached++;
169 	if (ctx->flags & IORING_SETUP_CQE32)
170 		ctx->cqe_cached++;
171 	return true;
172 }
173 
174 static inline bool io_get_cqe(struct io_ring_ctx *ctx, struct io_uring_cqe **ret)
175 {
176 	return io_get_cqe_overflow(ctx, ret, false);
177 }
178 
179 static __always_inline bool io_fill_cqe_req(struct io_ring_ctx *ctx,
180 					    struct io_kiocb *req)
181 {
182 	struct io_uring_cqe *cqe;
183 
184 	/*
185 	 * If we can't get a cq entry, userspace overflowed the
186 	 * submission (by quite a lot). Increment the overflow count in
187 	 * the ring.
188 	 */
189 	if (unlikely(!io_get_cqe(ctx, &cqe)))
190 		return false;
191 
192 	if (trace_io_uring_complete_enabled())
193 		trace_io_uring_complete(req->ctx, req, req->cqe.user_data,
194 					req->cqe.res, req->cqe.flags,
195 					req->big_cqe.extra1, req->big_cqe.extra2);
196 
197 	memcpy(cqe, &req->cqe, sizeof(*cqe));
198 	if (ctx->flags & IORING_SETUP_CQE32) {
199 		memcpy(cqe->big_cqe, &req->big_cqe, sizeof(*cqe));
200 		memset(&req->big_cqe, 0, sizeof(req->big_cqe));
201 	}
202 	return true;
203 }
204 
205 static inline void req_set_fail(struct io_kiocb *req)
206 {
207 	req->flags |= REQ_F_FAIL;
208 	if (req->flags & REQ_F_CQE_SKIP) {
209 		req->flags &= ~REQ_F_CQE_SKIP;
210 		req->flags |= REQ_F_SKIP_LINK_CQES;
211 	}
212 }
213 
214 static inline void io_req_set_res(struct io_kiocb *req, s32 res, u32 cflags)
215 {
216 	req->cqe.res = res;
217 	req->cqe.flags = cflags;
218 }
219 
220 static inline bool req_has_async_data(struct io_kiocb *req)
221 {
222 	return req->flags & REQ_F_ASYNC_DATA;
223 }
224 
225 static inline void io_put_file(struct io_kiocb *req)
226 {
227 	if (!(req->flags & REQ_F_FIXED_FILE) && req->file)
228 		fput(req->file);
229 }
230 
231 static inline void io_ring_submit_unlock(struct io_ring_ctx *ctx,
232 					 unsigned issue_flags)
233 {
234 	lockdep_assert_held(&ctx->uring_lock);
235 	if (unlikely(issue_flags & IO_URING_F_UNLOCKED))
236 		mutex_unlock(&ctx->uring_lock);
237 }
238 
239 static inline void io_ring_submit_lock(struct io_ring_ctx *ctx,
240 				       unsigned issue_flags)
241 {
242 	/*
243 	 * "Normal" inline submissions always hold the uring_lock, since we
244 	 * grab it from the system call. Same is true for the SQPOLL offload.
245 	 * The only exception is when we've detached the request and issue it
246 	 * from an async worker thread, grab the lock for that case.
247 	 */
248 	if (unlikely(issue_flags & IO_URING_F_UNLOCKED))
249 		mutex_lock(&ctx->uring_lock);
250 	lockdep_assert_held(&ctx->uring_lock);
251 }
252 
253 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
254 {
255 	/* order cqe stores with ring update */
256 	smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
257 }
258 
259 static inline void io_poll_wq_wake(struct io_ring_ctx *ctx)
260 {
261 	if (wq_has_sleeper(&ctx->poll_wq))
262 		__wake_up(&ctx->poll_wq, TASK_NORMAL, 0,
263 				poll_to_key(EPOLL_URING_WAKE | EPOLLIN));
264 }
265 
266 static inline void io_cqring_wake(struct io_ring_ctx *ctx)
267 {
268 	/*
269 	 * Trigger waitqueue handler on all waiters on our waitqueue. This
270 	 * won't necessarily wake up all the tasks, io_should_wake() will make
271 	 * that decision.
272 	 *
273 	 * Pass in EPOLLIN|EPOLL_URING_WAKE as the poll wakeup key. The latter
274 	 * set in the mask so that if we recurse back into our own poll
275 	 * waitqueue handlers, we know we have a dependency between eventfd or
276 	 * epoll and should terminate multishot poll at that point.
277 	 */
278 	if (wq_has_sleeper(&ctx->cq_wait))
279 		__wake_up(&ctx->cq_wait, TASK_NORMAL, 0,
280 				poll_to_key(EPOLL_URING_WAKE | EPOLLIN));
281 }
282 
283 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
284 {
285 	struct io_rings *r = ctx->rings;
286 
287 	/*
288 	 * SQPOLL must use the actual sqring head, as using the cached_sq_head
289 	 * is race prone if the SQPOLL thread has grabbed entries but not yet
290 	 * committed them to the ring. For !SQPOLL, this doesn't matter, but
291 	 * since this helper is just used for SQPOLL sqring waits (or POLLOUT),
292 	 * just read the actual sqring head unconditionally.
293 	 */
294 	return READ_ONCE(r->sq.tail) - READ_ONCE(r->sq.head) == ctx->sq_entries;
295 }
296 
297 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
298 {
299 	struct io_rings *rings = ctx->rings;
300 	unsigned int entries;
301 
302 	/* make sure SQ entry isn't read before tail */
303 	entries = smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
304 	return min(entries, ctx->sq_entries);
305 }
306 
307 static inline int io_run_task_work(void)
308 {
309 	bool ret = false;
310 
311 	/*
312 	 * Always check-and-clear the task_work notification signal. With how
313 	 * signaling works for task_work, we can find it set with nothing to
314 	 * run. We need to clear it for that case, like get_signal() does.
315 	 */
316 	if (test_thread_flag(TIF_NOTIFY_SIGNAL))
317 		clear_notify_signal();
318 	/*
319 	 * PF_IO_WORKER never returns to userspace, so check here if we have
320 	 * notify work that needs processing.
321 	 */
322 	if (current->flags & PF_IO_WORKER) {
323 		if (test_thread_flag(TIF_NOTIFY_RESUME)) {
324 			__set_current_state(TASK_RUNNING);
325 			resume_user_mode_work(NULL);
326 		}
327 		if (current->io_uring) {
328 			unsigned int count = 0;
329 
330 			__set_current_state(TASK_RUNNING);
331 			tctx_task_work_run(current->io_uring, UINT_MAX, &count);
332 			if (count)
333 				ret = true;
334 		}
335 	}
336 	if (task_work_pending(current)) {
337 		__set_current_state(TASK_RUNNING);
338 		task_work_run();
339 		ret = true;
340 	}
341 
342 	return ret;
343 }
344 
345 static inline bool io_task_work_pending(struct io_ring_ctx *ctx)
346 {
347 	return task_work_pending(current) || !llist_empty(&ctx->work_llist);
348 }
349 
350 static inline void io_tw_lock(struct io_ring_ctx *ctx, struct io_tw_state *ts)
351 {
352 	lockdep_assert_held(&ctx->uring_lock);
353 }
354 
355 /*
356  * Don't complete immediately but use deferred completion infrastructure.
357  * Protected by ->uring_lock and can only be used either with
358  * IO_URING_F_COMPLETE_DEFER or inside a tw handler holding the mutex.
359  */
360 static inline void io_req_complete_defer(struct io_kiocb *req)
361 	__must_hold(&req->ctx->uring_lock)
362 {
363 	struct io_submit_state *state = &req->ctx->submit_state;
364 
365 	lockdep_assert_held(&req->ctx->uring_lock);
366 
367 	wq_list_add_tail(&req->comp_list, &state->compl_reqs);
368 }
369 
370 static inline void io_commit_cqring_flush(struct io_ring_ctx *ctx)
371 {
372 	if (unlikely(ctx->off_timeout_used || ctx->drain_active ||
373 		     ctx->has_evfd || ctx->poll_activated))
374 		__io_commit_cqring_flush(ctx);
375 }
376 
377 static inline void io_get_task_refs(int nr)
378 {
379 	struct io_uring_task *tctx = current->io_uring;
380 
381 	tctx->cached_refs -= nr;
382 	if (unlikely(tctx->cached_refs < 0))
383 		io_task_refs_refill(tctx);
384 }
385 
386 static inline bool io_req_cache_empty(struct io_ring_ctx *ctx)
387 {
388 	return !ctx->submit_state.free_list.next;
389 }
390 
391 extern struct kmem_cache *req_cachep;
392 extern struct kmem_cache *io_buf_cachep;
393 
394 static inline struct io_kiocb *io_extract_req(struct io_ring_ctx *ctx)
395 {
396 	struct io_kiocb *req;
397 
398 	req = container_of(ctx->submit_state.free_list.next, struct io_kiocb, comp_list);
399 	wq_stack_extract(&ctx->submit_state.free_list);
400 	return req;
401 }
402 
403 static inline bool io_alloc_req(struct io_ring_ctx *ctx, struct io_kiocb **req)
404 {
405 	if (unlikely(io_req_cache_empty(ctx))) {
406 		if (!__io_alloc_req_refill(ctx))
407 			return false;
408 	}
409 	*req = io_extract_req(ctx);
410 	return true;
411 }
412 
413 static inline bool io_allowed_defer_tw_run(struct io_ring_ctx *ctx)
414 {
415 	return likely(ctx->submitter_task == current);
416 }
417 
418 static inline bool io_allowed_run_tw(struct io_ring_ctx *ctx)
419 {
420 	return likely(!(ctx->flags & IORING_SETUP_DEFER_TASKRUN) ||
421 		      ctx->submitter_task == current);
422 }
423 
424 static inline void io_req_queue_tw_complete(struct io_kiocb *req, s32 res)
425 {
426 	io_req_set_res(req, res, 0);
427 	req->io_task_work.func = io_req_task_complete;
428 	io_req_task_work_add(req);
429 }
430 
431 /*
432  * IORING_SETUP_SQE128 contexts allocate twice the normal SQE size for each
433  * slot.
434  */
435 static inline size_t uring_sqe_size(struct io_ring_ctx *ctx)
436 {
437 	if (ctx->flags & IORING_SETUP_SQE128)
438 		return 2 * sizeof(struct io_uring_sqe);
439 	return sizeof(struct io_uring_sqe);
440 }
441 
442 static inline bool io_file_can_poll(struct io_kiocb *req)
443 {
444 	if (req->flags & REQ_F_CAN_POLL)
445 		return true;
446 	if (req->file && file_can_poll(req->file)) {
447 		req->flags |= REQ_F_CAN_POLL;
448 		return true;
449 	}
450 	return false;
451 }
452 
453 static inline ktime_t io_get_time(struct io_ring_ctx *ctx)
454 {
455 	if (ctx->clockid == CLOCK_MONOTONIC)
456 		return ktime_get();
457 
458 	return ktime_get_with_offset(ctx->clock_offset);
459 }
460 
461 enum {
462 	IO_CHECK_CQ_OVERFLOW_BIT,
463 	IO_CHECK_CQ_DROPPED_BIT,
464 };
465 
466 static inline bool io_has_work(struct io_ring_ctx *ctx)
467 {
468 	return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
469 	       !llist_empty(&ctx->work_llist);
470 }
471 #endif
472