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