xref: /openbmc/linux/io_uring/io_uring.h (revision 19e85d93)
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/io_uring_types.h>
9 #include <uapi/linux/eventpoll.h>
10 #include "io-wq.h"
11 #include "slist.h"
12 #include "filetable.h"
13 
14 #ifndef CREATE_TRACE_POINTS
15 #include <trace/events/io_uring.h>
16 #endif
17 
18 enum {
19 	/*
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.
22 	 *
23 	 * Must not be used wirh requests generating more than one CQE.
24 	 * It's also ignored unless IORING_SETUP_DEFER_TASKRUN is set.
25 	 */
26 	IOU_F_TWQ_LAZY_WAKE			= 1,
27 };
28 
29 enum {
30 	IOU_OK			= 0,
31 	IOU_ISSUE_SKIP_COMPLETE	= -EIOCBQUEUED,
32 
33 	/*
34 	 * Requeue the task_work to restart operations on this request. The
35 	 * actual value isn't important, should just be not an otherwise
36 	 * valid error code, yet less than -MAX_ERRNO and valid internally.
37 	 */
38 	IOU_REQUEUE		= -3072,
39 
40 	/*
41 	 * Intended only when both IO_URING_F_MULTISHOT is passed
42 	 * to indicate to the poll runner that multishot should be
43 	 * removed and the result is set on req->cqe.res.
44 	 */
45 	IOU_STOP_MULTISHOT	= -ECANCELED,
46 };
47 
48 bool io_cqe_cache_refill(struct io_ring_ctx *ctx, bool overflow);
49 void io_req_cqe_overflow(struct io_kiocb *req);
50 int io_run_task_work_sig(struct io_ring_ctx *ctx);
51 void io_req_defer_failed(struct io_kiocb *req, s32 res);
52 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags);
53 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags);
54 bool io_fill_cqe_req_aux(struct io_kiocb *req, bool defer, s32 res, u32 cflags);
55 void __io_commit_cqring_flush(struct io_ring_ctx *ctx);
56 
57 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages);
58 
59 struct file *io_file_get_normal(struct io_kiocb *req, int fd);
60 struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
61 			       unsigned issue_flags);
62 
63 void __io_req_task_work_add(struct io_kiocb *req, unsigned flags);
64 bool io_alloc_async_data(struct io_kiocb *req);
65 void io_req_task_queue(struct io_kiocb *req);
66 void io_queue_iowq(struct io_kiocb *req, struct io_tw_state *ts_dont_use);
67 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts);
68 void io_req_task_queue_fail(struct io_kiocb *req, int ret);
69 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts);
70 void tctx_task_work(struct callback_head *cb);
71 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
72 int io_uring_alloc_task_context(struct task_struct *task,
73 				struct io_ring_ctx *ctx);
74 
75 int io_ring_add_registered_file(struct io_uring_task *tctx, struct file *file,
76 				     int start, int end);
77 
78 int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts);
79 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr);
80 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin);
81 void __io_submit_flush_completions(struct io_ring_ctx *ctx);
82 int io_req_prep_async(struct io_kiocb *req);
83 
84 struct io_wq_work *io_wq_free_work(struct io_wq_work *work);
85 void io_wq_submit_work(struct io_wq_work *work);
86 
87 void io_free_req(struct io_kiocb *req);
88 void io_queue_next(struct io_kiocb *req);
89 void io_task_refs_refill(struct io_uring_task *tctx);
90 bool __io_alloc_req_refill(struct io_ring_ctx *ctx);
91 
92 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
93 			bool cancel_all);
94 
95 void *io_mem_alloc(size_t size);
96 void io_mem_free(void *ptr);
97 
98 #if defined(CONFIG_PROVE_LOCKING)
99 static inline void io_lockdep_assert_cq_locked(struct io_ring_ctx *ctx)
100 {
101 	lockdep_assert(in_task());
102 
103 	if (ctx->flags & IORING_SETUP_IOPOLL) {
104 		lockdep_assert_held(&ctx->uring_lock);
105 	} else if (!ctx->task_complete) {
106 		lockdep_assert_held(&ctx->completion_lock);
107 	} else if (ctx->submitter_task) {
108 		/*
109 		 * ->submitter_task may be NULL and we can still post a CQE,
110 		 * if the ring has been setup with IORING_SETUP_R_DISABLED.
111 		 * Not from an SQE, as those cannot be submitted, but via
112 		 * updating tagged resources.
113 		 */
114 		if (ctx->submitter_task->flags & PF_EXITING)
115 			lockdep_assert(current_work());
116 		else
117 			lockdep_assert(current == ctx->submitter_task);
118 	}
119 }
120 #else
121 static inline void io_lockdep_assert_cq_locked(struct io_ring_ctx *ctx)
122 {
123 }
124 #endif
125 
126 static inline void io_req_task_work_add(struct io_kiocb *req)
127 {
128 	__io_req_task_work_add(req, 0);
129 }
130 
131 #define io_for_each_link(pos, head) \
132 	for (pos = (head); pos; pos = pos->link)
133 
134 static inline bool io_get_cqe_overflow(struct io_ring_ctx *ctx,
135 					struct io_uring_cqe **ret,
136 					bool overflow)
137 {
138 	io_lockdep_assert_cq_locked(ctx);
139 
140 	if (unlikely(ctx->cqe_cached >= ctx->cqe_sentinel)) {
141 		if (unlikely(!io_cqe_cache_refill(ctx, overflow)))
142 			return false;
143 	}
144 	*ret = ctx->cqe_cached;
145 	ctx->cached_cq_tail++;
146 	ctx->cqe_cached++;
147 	if (ctx->flags & IORING_SETUP_CQE32)
148 		ctx->cqe_cached++;
149 	return true;
150 }
151 
152 static inline bool io_get_cqe(struct io_ring_ctx *ctx, struct io_uring_cqe **ret)
153 {
154 	return io_get_cqe_overflow(ctx, ret, false);
155 }
156 
157 static __always_inline bool io_fill_cqe_req(struct io_ring_ctx *ctx,
158 					    struct io_kiocb *req)
159 {
160 	struct io_uring_cqe *cqe;
161 
162 	/*
163 	 * If we can't get a cq entry, userspace overflowed the
164 	 * submission (by quite a lot). Increment the overflow count in
165 	 * the ring.
166 	 */
167 	if (unlikely(!io_get_cqe(ctx, &cqe)))
168 		return false;
169 
170 	if (trace_io_uring_complete_enabled())
171 		trace_io_uring_complete(req->ctx, req, req->cqe.user_data,
172 					req->cqe.res, req->cqe.flags,
173 					req->big_cqe.extra1, req->big_cqe.extra2);
174 
175 	memcpy(cqe, &req->cqe, sizeof(*cqe));
176 	if (ctx->flags & IORING_SETUP_CQE32) {
177 		memcpy(cqe->big_cqe, &req->big_cqe, sizeof(*cqe));
178 		memset(&req->big_cqe, 0, sizeof(req->big_cqe));
179 	}
180 	return true;
181 }
182 
183 static inline void req_set_fail(struct io_kiocb *req)
184 {
185 	req->flags |= REQ_F_FAIL;
186 	if (req->flags & REQ_F_CQE_SKIP) {
187 		req->flags &= ~REQ_F_CQE_SKIP;
188 		req->flags |= REQ_F_SKIP_LINK_CQES;
189 	}
190 }
191 
192 static inline void io_req_set_res(struct io_kiocb *req, s32 res, u32 cflags)
193 {
194 	req->cqe.res = res;
195 	req->cqe.flags = cflags;
196 }
197 
198 static inline bool req_has_async_data(struct io_kiocb *req)
199 {
200 	return req->flags & REQ_F_ASYNC_DATA;
201 }
202 
203 static inline void io_put_file(struct io_kiocb *req)
204 {
205 	if (!(req->flags & REQ_F_FIXED_FILE) && req->file)
206 		fput(req->file);
207 }
208 
209 static inline void io_ring_submit_unlock(struct io_ring_ctx *ctx,
210 					 unsigned issue_flags)
211 {
212 	lockdep_assert_held(&ctx->uring_lock);
213 	if (issue_flags & IO_URING_F_UNLOCKED)
214 		mutex_unlock(&ctx->uring_lock);
215 }
216 
217 static inline void io_ring_submit_lock(struct io_ring_ctx *ctx,
218 				       unsigned issue_flags)
219 {
220 	/*
221 	 * "Normal" inline submissions always hold the uring_lock, since we
222 	 * grab it from the system call. Same is true for the SQPOLL offload.
223 	 * The only exception is when we've detached the request and issue it
224 	 * from an async worker thread, grab the lock for that case.
225 	 */
226 	if (issue_flags & IO_URING_F_UNLOCKED)
227 		mutex_lock(&ctx->uring_lock);
228 	lockdep_assert_held(&ctx->uring_lock);
229 }
230 
231 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
232 {
233 	/* order cqe stores with ring update */
234 	smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
235 }
236 
237 static inline void io_poll_wq_wake(struct io_ring_ctx *ctx)
238 {
239 	if (wq_has_sleeper(&ctx->poll_wq))
240 		__wake_up(&ctx->poll_wq, TASK_NORMAL, 0,
241 				poll_to_key(EPOLL_URING_WAKE | EPOLLIN));
242 }
243 
244 static inline void io_cqring_wake(struct io_ring_ctx *ctx)
245 {
246 	/*
247 	 * Trigger waitqueue handler on all waiters on our waitqueue. This
248 	 * won't necessarily wake up all the tasks, io_should_wake() will make
249 	 * that decision.
250 	 *
251 	 * Pass in EPOLLIN|EPOLL_URING_WAKE as the poll wakeup key. The latter
252 	 * set in the mask so that if we recurse back into our own poll
253 	 * waitqueue handlers, we know we have a dependency between eventfd or
254 	 * epoll and should terminate multishot poll at that point.
255 	 */
256 	if (wq_has_sleeper(&ctx->cq_wait))
257 		__wake_up(&ctx->cq_wait, TASK_NORMAL, 0,
258 				poll_to_key(EPOLL_URING_WAKE | EPOLLIN));
259 }
260 
261 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
262 {
263 	struct io_rings *r = ctx->rings;
264 
265 	return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
266 }
267 
268 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
269 {
270 	struct io_rings *rings = ctx->rings;
271 	unsigned int entries;
272 
273 	/* make sure SQ entry isn't read before tail */
274 	entries = smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
275 	return min(entries, ctx->sq_entries);
276 }
277 
278 static inline int io_run_task_work(void)
279 {
280 	/*
281 	 * Always check-and-clear the task_work notification signal. With how
282 	 * signaling works for task_work, we can find it set with nothing to
283 	 * run. We need to clear it for that case, like get_signal() does.
284 	 */
285 	if (test_thread_flag(TIF_NOTIFY_SIGNAL))
286 		clear_notify_signal();
287 	/*
288 	 * PF_IO_WORKER never returns to userspace, so check here if we have
289 	 * notify work that needs processing.
290 	 */
291 	if (current->flags & PF_IO_WORKER &&
292 	    test_thread_flag(TIF_NOTIFY_RESUME)) {
293 		__set_current_state(TASK_RUNNING);
294 		resume_user_mode_work(NULL);
295 	}
296 	if (task_work_pending(current)) {
297 		__set_current_state(TASK_RUNNING);
298 		task_work_run();
299 		return 1;
300 	}
301 
302 	return 0;
303 }
304 
305 static inline bool io_task_work_pending(struct io_ring_ctx *ctx)
306 {
307 	return task_work_pending(current) || !llist_empty(&ctx->work_llist);
308 }
309 
310 static inline void io_tw_lock(struct io_ring_ctx *ctx, struct io_tw_state *ts)
311 {
312 	if (!ts->locked) {
313 		mutex_lock(&ctx->uring_lock);
314 		ts->locked = true;
315 	}
316 }
317 
318 /*
319  * Don't complete immediately but use deferred completion infrastructure.
320  * Protected by ->uring_lock and can only be used either with
321  * IO_URING_F_COMPLETE_DEFER or inside a tw handler holding the mutex.
322  */
323 static inline void io_req_complete_defer(struct io_kiocb *req)
324 	__must_hold(&req->ctx->uring_lock)
325 {
326 	struct io_submit_state *state = &req->ctx->submit_state;
327 
328 	lockdep_assert_held(&req->ctx->uring_lock);
329 
330 	wq_list_add_tail(&req->comp_list, &state->compl_reqs);
331 }
332 
333 static inline void io_commit_cqring_flush(struct io_ring_ctx *ctx)
334 {
335 	if (unlikely(ctx->off_timeout_used || ctx->drain_active ||
336 		     ctx->has_evfd || ctx->poll_activated))
337 		__io_commit_cqring_flush(ctx);
338 }
339 
340 static inline void io_get_task_refs(int nr)
341 {
342 	struct io_uring_task *tctx = current->io_uring;
343 
344 	tctx->cached_refs -= nr;
345 	if (unlikely(tctx->cached_refs < 0))
346 		io_task_refs_refill(tctx);
347 }
348 
349 static inline bool io_req_cache_empty(struct io_ring_ctx *ctx)
350 {
351 	return !ctx->submit_state.free_list.next;
352 }
353 
354 extern struct kmem_cache *req_cachep;
355 
356 static inline struct io_kiocb *io_extract_req(struct io_ring_ctx *ctx)
357 {
358 	struct io_kiocb *req;
359 
360 	req = container_of(ctx->submit_state.free_list.next, struct io_kiocb, comp_list);
361 	wq_stack_extract(&ctx->submit_state.free_list);
362 	return req;
363 }
364 
365 static inline bool io_alloc_req(struct io_ring_ctx *ctx, struct io_kiocb **req)
366 {
367 	if (unlikely(io_req_cache_empty(ctx))) {
368 		if (!__io_alloc_req_refill(ctx))
369 			return false;
370 	}
371 	*req = io_extract_req(ctx);
372 	return true;
373 }
374 
375 static inline bool io_allowed_defer_tw_run(struct io_ring_ctx *ctx)
376 {
377 	return likely(ctx->submitter_task == current);
378 }
379 
380 static inline bool io_allowed_run_tw(struct io_ring_ctx *ctx)
381 {
382 	return likely(!(ctx->flags & IORING_SETUP_DEFER_TASKRUN) ||
383 		      ctx->submitter_task == current);
384 }
385 
386 static inline void io_req_queue_tw_complete(struct io_kiocb *req, s32 res)
387 {
388 	io_req_set_res(req, res, 0);
389 	req->io_task_work.func = io_req_task_complete;
390 	io_req_task_work_add(req);
391 }
392 
393 /*
394  * IORING_SETUP_SQE128 contexts allocate twice the normal SQE size for each
395  * slot.
396  */
397 static inline size_t uring_sqe_size(struct io_ring_ctx *ctx)
398 {
399 	if (ctx->flags & IORING_SETUP_SQE128)
400 		return 2 * sizeof(struct io_uring_sqe);
401 	return sizeof(struct io_uring_sqe);
402 }
403 #endif
404