xref: /openbmc/linux/io_uring/io_uring.h (revision 9322af3e6aeae04c7eda3e6a0c977e97a13cf6ed)
1 #ifndef IOU_CORE_H
2 #define IOU_CORE_H
3 
4 #include <linux/errno.h>
5 #include <linux/lockdep.h>
6 #include <linux/io_uring_types.h>
7 #include <uapi/linux/eventpoll.h>
8 #include "io-wq.h"
9 #include "slist.h"
10 #include "filetable.h"
11 
12 #ifndef CREATE_TRACE_POINTS
13 #include <trace/events/io_uring.h>
14 #endif
15 
16 enum {
17 	IOU_OK			= 0,
18 	IOU_ISSUE_SKIP_COMPLETE	= -EIOCBQUEUED,
19 
20 	/*
21 	 * Intended only when both IO_URING_F_MULTISHOT is passed
22 	 * to indicate to the poll runner that multishot should be
23 	 * removed and the result is set on req->cqe.res.
24 	 */
25 	IOU_STOP_MULTISHOT	= -ECANCELED,
26 };
27 
28 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow);
29 bool io_req_cqe_overflow(struct io_kiocb *req);
30 int io_run_task_work_sig(struct io_ring_ctx *ctx);
31 int __io_run_local_work(struct io_ring_ctx *ctx, bool *locked);
32 int io_run_local_work(struct io_ring_ctx *ctx);
33 void io_req_defer_failed(struct io_kiocb *req, s32 res);
34 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags);
35 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags);
36 bool io_aux_cqe(struct io_ring_ctx *ctx, bool defer, u64 user_data, s32 res, u32 cflags,
37 		bool allow_overflow);
38 void __io_commit_cqring_flush(struct io_ring_ctx *ctx);
39 
40 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages);
41 
42 struct file *io_file_get_normal(struct io_kiocb *req, int fd);
43 struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
44 			       unsigned issue_flags);
45 
46 static inline bool io_req_ffs_set(struct io_kiocb *req)
47 {
48 	return req->flags & REQ_F_FIXED_FILE;
49 }
50 
51 void __io_req_task_work_add(struct io_kiocb *req, bool allow_local);
52 bool io_is_uring_fops(struct file *file);
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, bool *dont_use);
56 void io_req_task_complete(struct io_kiocb *req, bool *locked);
57 void io_req_task_queue_fail(struct io_kiocb *req, int ret);
58 void io_req_task_submit(struct io_kiocb *req, bool *locked);
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);
63 
64 int io_poll_issue(struct io_kiocb *req, bool *locked);
65 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr);
66 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin);
67 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node);
68 int io_req_prep_async(struct io_kiocb *req);
69 
70 struct io_wq_work *io_wq_free_work(struct io_wq_work *work);
71 void io_wq_submit_work(struct io_wq_work *work);
72 
73 void io_free_req(struct io_kiocb *req);
74 void io_queue_next(struct io_kiocb *req);
75 void __io_put_task(struct task_struct *task, int nr);
76 void io_task_refs_refill(struct io_uring_task *tctx);
77 bool __io_alloc_req_refill(struct io_ring_ctx *ctx);
78 
79 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
80 			bool cancel_all);
81 
82 static inline void io_req_task_work_add(struct io_kiocb *req)
83 {
84 	__io_req_task_work_add(req, true);
85 }
86 
87 #define io_for_each_link(pos, head) \
88 	for (pos = (head); pos; pos = pos->link)
89 
90 static inline void io_cq_lock(struct io_ring_ctx *ctx)
91 	__acquires(ctx->completion_lock)
92 {
93 	spin_lock(&ctx->completion_lock);
94 }
95 
96 static inline void io_cq_unlock(struct io_ring_ctx *ctx)
97 {
98 	spin_unlock(&ctx->completion_lock);
99 }
100 
101 void io_cq_unlock_post(struct io_ring_ctx *ctx);
102 
103 static inline struct io_uring_cqe *io_get_cqe_overflow(struct io_ring_ctx *ctx,
104 						       bool overflow)
105 {
106 	if (likely(ctx->cqe_cached < ctx->cqe_sentinel)) {
107 		struct io_uring_cqe *cqe = ctx->cqe_cached;
108 
109 		ctx->cached_cq_tail++;
110 		ctx->cqe_cached++;
111 		if (ctx->flags & IORING_SETUP_CQE32)
112 			ctx->cqe_cached++;
113 		return cqe;
114 	}
115 
116 	return __io_get_cqe(ctx, overflow);
117 }
118 
119 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
120 {
121 	return io_get_cqe_overflow(ctx, false);
122 }
123 
124 static inline bool __io_fill_cqe_req(struct io_ring_ctx *ctx,
125 				     struct io_kiocb *req)
126 {
127 	struct io_uring_cqe *cqe;
128 
129 	/*
130 	 * If we can't get a cq entry, userspace overflowed the
131 	 * submission (by quite a lot). Increment the overflow count in
132 	 * the ring.
133 	 */
134 	cqe = io_get_cqe(ctx);
135 	if (unlikely(!cqe))
136 		return false;
137 
138 	trace_io_uring_complete(req->ctx, req, req->cqe.user_data,
139 				req->cqe.res, req->cqe.flags,
140 				(req->flags & REQ_F_CQE32_INIT) ? req->extra1 : 0,
141 				(req->flags & REQ_F_CQE32_INIT) ? req->extra2 : 0);
142 
143 	memcpy(cqe, &req->cqe, sizeof(*cqe));
144 
145 	if (ctx->flags & IORING_SETUP_CQE32) {
146 		u64 extra1 = 0, extra2 = 0;
147 
148 		if (req->flags & REQ_F_CQE32_INIT) {
149 			extra1 = req->extra1;
150 			extra2 = req->extra2;
151 		}
152 
153 		WRITE_ONCE(cqe->big_cqe[0], extra1);
154 		WRITE_ONCE(cqe->big_cqe[1], extra2);
155 	}
156 	return true;
157 }
158 
159 static inline bool io_fill_cqe_req(struct io_ring_ctx *ctx,
160 				   struct io_kiocb *req)
161 {
162 	if (likely(__io_fill_cqe_req(ctx, req)))
163 		return true;
164 	return io_req_cqe_overflow(req);
165 }
166 
167 static inline void req_set_fail(struct io_kiocb *req)
168 {
169 	req->flags |= REQ_F_FAIL;
170 	if (req->flags & REQ_F_CQE_SKIP) {
171 		req->flags &= ~REQ_F_CQE_SKIP;
172 		req->flags |= REQ_F_SKIP_LINK_CQES;
173 	}
174 }
175 
176 static inline void io_req_set_res(struct io_kiocb *req, s32 res, u32 cflags)
177 {
178 	req->cqe.res = res;
179 	req->cqe.flags = cflags;
180 }
181 
182 static inline bool req_has_async_data(struct io_kiocb *req)
183 {
184 	return req->flags & REQ_F_ASYNC_DATA;
185 }
186 
187 static inline void io_put_file(struct file *file)
188 {
189 	if (file)
190 		fput(file);
191 }
192 
193 static inline void io_ring_submit_unlock(struct io_ring_ctx *ctx,
194 					 unsigned issue_flags)
195 {
196 	lockdep_assert_held(&ctx->uring_lock);
197 	if (issue_flags & IO_URING_F_UNLOCKED)
198 		mutex_unlock(&ctx->uring_lock);
199 }
200 
201 static inline void io_ring_submit_lock(struct io_ring_ctx *ctx,
202 				       unsigned issue_flags)
203 {
204 	/*
205 	 * "Normal" inline submissions always hold the uring_lock, since we
206 	 * grab it from the system call. Same is true for the SQPOLL offload.
207 	 * The only exception is when we've detached the request and issue it
208 	 * from an async worker thread, grab the lock for that case.
209 	 */
210 	if (issue_flags & IO_URING_F_UNLOCKED)
211 		mutex_lock(&ctx->uring_lock);
212 	lockdep_assert_held(&ctx->uring_lock);
213 }
214 
215 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
216 {
217 	/* order cqe stores with ring update */
218 	smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
219 }
220 
221 /* requires smb_mb() prior, see wq_has_sleeper() */
222 static inline void __io_cqring_wake(struct io_ring_ctx *ctx)
223 {
224 	/*
225 	 * Trigger waitqueue handler on all waiters on our waitqueue. This
226 	 * won't necessarily wake up all the tasks, io_should_wake() will make
227 	 * that decision.
228 	 *
229 	 * Pass in EPOLLIN|EPOLL_URING_WAKE as the poll wakeup key. The latter
230 	 * set in the mask so that if we recurse back into our own poll
231 	 * waitqueue handlers, we know we have a dependency between eventfd or
232 	 * epoll and should terminate multishot poll at that point.
233 	 */
234 	if (waitqueue_active(&ctx->cq_wait))
235 		__wake_up(&ctx->cq_wait, TASK_NORMAL, 0,
236 				poll_to_key(EPOLL_URING_WAKE | EPOLLIN));
237 }
238 
239 static inline void io_cqring_wake(struct io_ring_ctx *ctx)
240 {
241 	smp_mb();
242 	__io_cqring_wake(ctx);
243 }
244 
245 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
246 {
247 	struct io_rings *r = ctx->rings;
248 
249 	return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
250 }
251 
252 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
253 {
254 	struct io_rings *rings = ctx->rings;
255 
256 	/* make sure SQ entry isn't read before tail */
257 	return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
258 }
259 
260 static inline int io_run_task_work(void)
261 {
262 	/*
263 	 * Always check-and-clear the task_work notification signal. With how
264 	 * signaling works for task_work, we can find it set with nothing to
265 	 * run. We need to clear it for that case, like get_signal() does.
266 	 */
267 	if (test_thread_flag(TIF_NOTIFY_SIGNAL))
268 		clear_notify_signal();
269 	if (task_work_pending(current)) {
270 		__set_current_state(TASK_RUNNING);
271 		task_work_run();
272 		return 1;
273 	}
274 
275 	return 0;
276 }
277 
278 static inline bool io_task_work_pending(struct io_ring_ctx *ctx)
279 {
280 	return test_thread_flag(TIF_NOTIFY_SIGNAL) ||
281 		!wq_list_empty(&ctx->work_llist);
282 }
283 
284 static inline int io_run_task_work_ctx(struct io_ring_ctx *ctx)
285 {
286 	int ret = 0;
287 	int ret2;
288 
289 	if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
290 		ret = io_run_local_work(ctx);
291 
292 	/* want to run this after in case more is added */
293 	ret2 = io_run_task_work();
294 
295 	/* Try propagate error in favour of if tasks were run,
296 	 * but still make sure to run them if requested
297 	 */
298 	if (ret >= 0)
299 		ret += ret2;
300 
301 	return ret;
302 }
303 
304 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx)
305 {
306 	bool locked;
307 	int ret;
308 
309 	if (llist_empty(&ctx->work_llist))
310 		return 0;
311 
312 	locked = true;
313 	ret = __io_run_local_work(ctx, &locked);
314 	/* shouldn't happen! */
315 	if (WARN_ON_ONCE(!locked))
316 		mutex_lock(&ctx->uring_lock);
317 	return ret;
318 }
319 
320 static inline void io_tw_lock(struct io_ring_ctx *ctx, bool *locked)
321 {
322 	if (!*locked) {
323 		mutex_lock(&ctx->uring_lock);
324 		*locked = true;
325 	}
326 }
327 
328 /*
329  * Don't complete immediately but use deferred completion infrastructure.
330  * Protected by ->uring_lock and can only be used either with
331  * IO_URING_F_COMPLETE_DEFER or inside a tw handler holding the mutex.
332  */
333 static inline void io_req_complete_defer(struct io_kiocb *req)
334 	__must_hold(&req->ctx->uring_lock)
335 {
336 	struct io_submit_state *state = &req->ctx->submit_state;
337 
338 	lockdep_assert_held(&req->ctx->uring_lock);
339 
340 	wq_list_add_tail(&req->comp_list, &state->compl_reqs);
341 }
342 
343 static inline void io_commit_cqring_flush(struct io_ring_ctx *ctx)
344 {
345 	if (unlikely(ctx->off_timeout_used || ctx->drain_active || ctx->has_evfd))
346 		__io_commit_cqring_flush(ctx);
347 }
348 
349 /* must to be called somewhat shortly after putting a request */
350 static inline void io_put_task(struct task_struct *task, int nr)
351 {
352 	if (likely(task == current))
353 		task->io_uring->cached_refs += nr;
354 	else
355 		__io_put_task(task, nr);
356 }
357 
358 static inline void io_get_task_refs(int nr)
359 {
360 	struct io_uring_task *tctx = current->io_uring;
361 
362 	tctx->cached_refs -= nr;
363 	if (unlikely(tctx->cached_refs < 0))
364 		io_task_refs_refill(tctx);
365 }
366 
367 static inline bool io_req_cache_empty(struct io_ring_ctx *ctx)
368 {
369 	return !ctx->submit_state.free_list.next;
370 }
371 
372 static inline bool io_alloc_req_refill(struct io_ring_ctx *ctx)
373 {
374 	if (unlikely(io_req_cache_empty(ctx)))
375 		return __io_alloc_req_refill(ctx);
376 	return true;
377 }
378 
379 static inline struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
380 {
381 	struct io_wq_work_node *node;
382 
383 	node = wq_stack_extract(&ctx->submit_state.free_list);
384 	return container_of(node, struct io_kiocb, comp_list);
385 }
386 
387 static inline bool io_allowed_run_tw(struct io_ring_ctx *ctx)
388 {
389 	return likely(!(ctx->flags & IORING_SETUP_DEFER_TASKRUN) ||
390 		      ctx->submitter_task == current);
391 }
392 
393 static inline void io_req_queue_tw_complete(struct io_kiocb *req, s32 res)
394 {
395 	io_req_set_res(req, res, 0);
396 	req->io_task_work.func = io_req_task_complete;
397 	io_req_task_work_add(req);
398 }
399 
400 #endif
401