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