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 "alloc_cache.h"
12#include "io-wq.h"
13#include "slist.h"
14#include "opdef.h"
15
16#ifndef CREATE_TRACE_POINTS
17#include <trace/events/io_uring.h>
18#endif
19
20struct io_rings_layout {
21 /* size of CQ + headers + SQ offset array */
22 size_t rings_size;
23 size_t sq_size;
24
25 size_t sq_array_offset;
26};
27
28struct io_ctx_config {
29 struct io_uring_params p;
30 struct io_rings_layout layout;
31 struct io_uring_params __user *uptr;
32};
33
34#define IORING_FEAT_FLAGS (IORING_FEAT_SINGLE_MMAP |\
35 IORING_FEAT_NODROP |\
36 IORING_FEAT_SUBMIT_STABLE |\
37 IORING_FEAT_RW_CUR_POS |\
38 IORING_FEAT_CUR_PERSONALITY |\
39 IORING_FEAT_FAST_POLL |\
40 IORING_FEAT_POLL_32BITS |\
41 IORING_FEAT_SQPOLL_NONFIXED |\
42 IORING_FEAT_EXT_ARG |\
43 IORING_FEAT_NATIVE_WORKERS |\
44 IORING_FEAT_RSRC_TAGS |\
45 IORING_FEAT_CQE_SKIP |\
46 IORING_FEAT_LINKED_FILE |\
47 IORING_FEAT_REG_REG_RING |\
48 IORING_FEAT_RECVSEND_BUNDLE |\
49 IORING_FEAT_MIN_TIMEOUT |\
50 IORING_FEAT_RW_ATTR |\
51 IORING_FEAT_NO_IOWAIT)
52
53#define IORING_SETUP_FLAGS (IORING_SETUP_IOPOLL |\
54 IORING_SETUP_SQPOLL |\
55 IORING_SETUP_SQ_AFF |\
56 IORING_SETUP_CQSIZE |\
57 IORING_SETUP_CLAMP |\
58 IORING_SETUP_ATTACH_WQ |\
59 IORING_SETUP_R_DISABLED |\
60 IORING_SETUP_SUBMIT_ALL |\
61 IORING_SETUP_COOP_TASKRUN |\
62 IORING_SETUP_TASKRUN_FLAG |\
63 IORING_SETUP_SQE128 |\
64 IORING_SETUP_CQE32 |\
65 IORING_SETUP_SINGLE_ISSUER |\
66 IORING_SETUP_DEFER_TASKRUN |\
67 IORING_SETUP_NO_MMAP |\
68 IORING_SETUP_REGISTERED_FD_ONLY |\
69 IORING_SETUP_NO_SQARRAY |\
70 IORING_SETUP_HYBRID_IOPOLL |\
71 IORING_SETUP_CQE_MIXED |\
72 IORING_SETUP_SQE_MIXED)
73
74#define IORING_ENTER_FLAGS (IORING_ENTER_GETEVENTS |\
75 IORING_ENTER_SQ_WAKEUP |\
76 IORING_ENTER_SQ_WAIT |\
77 IORING_ENTER_EXT_ARG |\
78 IORING_ENTER_REGISTERED_RING |\
79 IORING_ENTER_ABS_TIMER |\
80 IORING_ENTER_EXT_ARG_REG |\
81 IORING_ENTER_NO_IOWAIT)
82
83
84#define SQE_VALID_FLAGS (IOSQE_FIXED_FILE |\
85 IOSQE_IO_DRAIN |\
86 IOSQE_IO_LINK |\
87 IOSQE_IO_HARDLINK |\
88 IOSQE_ASYNC |\
89 IOSQE_BUFFER_SELECT |\
90 IOSQE_CQE_SKIP_SUCCESS)
91
92enum {
93 IOU_COMPLETE = 0,
94
95 IOU_ISSUE_SKIP_COMPLETE = -EIOCBQUEUED,
96
97 /*
98 * The request has more work to do and should be retried. io_uring will
99 * attempt to wait on the file for eligible opcodes, but otherwise
100 * it'll be handed to iowq for blocking execution. It works for normal
101 * requests as well as for the multi shot mode.
102 */
103 IOU_RETRY = -EAGAIN,
104
105 /*
106 * Requeue the task_work to restart operations on this request. The
107 * actual value isn't important, should just be not an otherwise
108 * valid error code, yet less than -MAX_ERRNO and valid internally.
109 */
110 IOU_REQUEUE = -3072,
111};
112
113struct io_defer_entry {
114 struct list_head list;
115 struct io_kiocb *req;
116};
117
118struct io_wait_queue {
119 struct wait_queue_entry wq;
120 struct io_ring_ctx *ctx;
121 unsigned cq_tail;
122 unsigned cq_min_tail;
123 unsigned nr_timeouts;
124 int hit_timeout;
125 ktime_t min_timeout;
126 ktime_t timeout;
127 struct hrtimer t;
128
129#ifdef CONFIG_NET_RX_BUSY_POLL
130 ktime_t napi_busy_poll_dt;
131 bool napi_prefer_busy_poll;
132#endif
133};
134
135static inline bool io_should_wake(struct io_wait_queue *iowq)
136{
137 struct io_ring_ctx *ctx = iowq->ctx;
138 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
139
140 /*
141 * Wake up if we have enough events, or if a timeout occurred since we
142 * started waiting. For timeouts, we always want to return to userspace,
143 * regardless of event count.
144 */
145 return dist >= 0 || atomic_read(v: &ctx->cq_timeouts) != iowq->nr_timeouts;
146}
147
148#define IORING_MAX_ENTRIES 32768
149#define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
150
151int io_prepare_config(struct io_ctx_config *config);
152
153bool io_cqe_cache_refill(struct io_ring_ctx *ctx, bool overflow, bool cqe32);
154int io_run_task_work_sig(struct io_ring_ctx *ctx);
155int io_run_local_work(struct io_ring_ctx *ctx, int min_events, int max_events);
156void io_req_defer_failed(struct io_kiocb *req, s32 res);
157bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags);
158void io_add_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags);
159bool io_req_post_cqe(struct io_kiocb *req, s32 res, u32 cflags);
160bool io_req_post_cqe32(struct io_kiocb *req, struct io_uring_cqe src_cqe[2]);
161void __io_commit_cqring_flush(struct io_ring_ctx *ctx);
162
163unsigned io_linked_nr(struct io_kiocb *req);
164void io_req_track_inflight(struct io_kiocb *req);
165struct file *io_file_get_normal(struct io_kiocb *req, int fd);
166struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
167 unsigned issue_flags);
168
169void __io_req_task_work_add(struct io_kiocb *req, unsigned flags);
170void io_req_task_work_add_remote(struct io_kiocb *req, unsigned flags);
171void io_req_task_queue(struct io_kiocb *req);
172void io_req_task_complete(struct io_tw_req tw_req, io_tw_token_t tw);
173void io_req_task_queue_fail(struct io_kiocb *req, int ret);
174void io_req_task_submit(struct io_tw_req tw_req, io_tw_token_t tw);
175struct llist_node *io_handle_tw_list(struct llist_node *node, unsigned int *count, unsigned int max_entries);
176struct llist_node *tctx_task_work_run(struct io_uring_task *tctx, unsigned int max_entries, unsigned int *count);
177void tctx_task_work(struct callback_head *cb);
178__cold void io_uring_drop_tctx_refs(struct task_struct *task);
179
180int io_ring_add_registered_file(struct io_uring_task *tctx, struct file *file,
181 int start, int end);
182void io_req_queue_iowq(struct io_kiocb *req);
183
184int io_poll_issue(struct io_kiocb *req, io_tw_token_t tw);
185int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr);
186int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin);
187__cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx);
188void __io_submit_flush_completions(struct io_ring_ctx *ctx);
189
190struct io_wq_work *io_wq_free_work(struct io_wq_work *work);
191void io_wq_submit_work(struct io_wq_work *work);
192
193void io_free_req(struct io_kiocb *req);
194void io_queue_next(struct io_kiocb *req);
195void io_task_refs_refill(struct io_uring_task *tctx);
196bool __io_alloc_req_refill(struct io_ring_ctx *ctx);
197
198void io_activate_pollwq(struct io_ring_ctx *ctx);
199
200static inline void io_lockdep_assert_cq_locked(struct io_ring_ctx *ctx)
201{
202#if defined(CONFIG_PROVE_LOCKING)
203 lockdep_assert(in_task());
204
205 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
206 lockdep_assert_held(&ctx->uring_lock);
207
208 if (ctx->flags & IORING_SETUP_IOPOLL) {
209 lockdep_assert_held(&ctx->uring_lock);
210 } else if (!ctx->task_complete) {
211 lockdep_assert_held(&ctx->completion_lock);
212 } else if (ctx->submitter_task) {
213 /*
214 * ->submitter_task may be NULL and we can still post a CQE,
215 * if the ring has been setup with IORING_SETUP_R_DISABLED.
216 * Not from an SQE, as those cannot be submitted, but via
217 * updating tagged resources.
218 */
219 if (!percpu_ref_is_dying(ref: &ctx->refs))
220 lockdep_assert(current == ctx->submitter_task);
221 }
222#endif
223}
224
225static inline bool io_is_compat(struct io_ring_ctx *ctx)
226{
227 return IS_ENABLED(CONFIG_COMPAT) && unlikely(ctx->compat);
228}
229
230static inline void io_req_task_work_add(struct io_kiocb *req)
231{
232 __io_req_task_work_add(req, flags: 0);
233}
234
235static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
236{
237 if (!wq_list_empty(&ctx->submit_state.compl_reqs) ||
238 ctx->submit_state.cq_flush)
239 __io_submit_flush_completions(ctx);
240}
241
242#define io_for_each_link(pos, head) \
243 for (pos = (head); pos; pos = pos->link)
244
245static inline bool io_get_cqe_overflow(struct io_ring_ctx *ctx,
246 struct io_uring_cqe **ret,
247 bool overflow, bool cqe32)
248{
249 io_lockdep_assert_cq_locked(ctx);
250
251 if (unlikely(ctx->cqe_sentinel - ctx->cqe_cached < (cqe32 + 1))) {
252 if (unlikely(!io_cqe_cache_refill(ctx, overflow, cqe32)))
253 return false;
254 }
255 *ret = ctx->cqe_cached;
256 ctx->cached_cq_tail++;
257 ctx->cqe_cached++;
258 if (ctx->flags & IORING_SETUP_CQE32) {
259 ctx->cqe_cached++;
260 } else if (cqe32 && ctx->flags & IORING_SETUP_CQE_MIXED) {
261 ctx->cqe_cached++;
262 ctx->cached_cq_tail++;
263 }
264 WARN_ON_ONCE(ctx->cqe_cached > ctx->cqe_sentinel);
265 return true;
266}
267
268static inline bool io_get_cqe(struct io_ring_ctx *ctx, struct io_uring_cqe **ret,
269 bool cqe32)
270{
271 return io_get_cqe_overflow(ctx, ret, overflow: false, cqe32);
272}
273
274static inline bool io_defer_get_uncommited_cqe(struct io_ring_ctx *ctx,
275 struct io_uring_cqe **cqe_ret)
276{
277 io_lockdep_assert_cq_locked(ctx);
278
279 ctx->submit_state.cq_flush = true;
280 return io_get_cqe(ctx, ret: cqe_ret, cqe32: ctx->flags & IORING_SETUP_CQE_MIXED);
281}
282
283static __always_inline bool io_fill_cqe_req(struct io_ring_ctx *ctx,
284 struct io_kiocb *req)
285{
286 bool is_cqe32 = req->cqe.flags & IORING_CQE_F_32;
287 struct io_uring_cqe *cqe;
288
289 /*
290 * If we can't get a cq entry, userspace overflowed the submission
291 * (by quite a lot).
292 */
293 if (unlikely(!io_get_cqe(ctx, &cqe, is_cqe32)))
294 return false;
295
296 memcpy(cqe, &req->cqe, sizeof(*cqe));
297 if (ctx->flags & IORING_SETUP_CQE32 || is_cqe32) {
298 memcpy(cqe->big_cqe, &req->big_cqe, sizeof(*cqe));
299 memset(&req->big_cqe, 0, sizeof(req->big_cqe));
300 }
301
302 if (trace_io_uring_complete_enabled())
303 trace_io_uring_complete(ctx: req->ctx, req, cqe);
304 return true;
305}
306
307static inline void req_set_fail(struct io_kiocb *req)
308{
309 req->flags |= REQ_F_FAIL;
310 if (req->flags & REQ_F_CQE_SKIP) {
311 req->flags &= ~REQ_F_CQE_SKIP;
312 req->flags |= REQ_F_SKIP_LINK_CQES;
313 }
314}
315
316static inline void io_req_set_res(struct io_kiocb *req, s32 res, u32 cflags)
317{
318 req->cqe.res = res;
319 req->cqe.flags = cflags;
320}
321
322static inline u32 ctx_cqe32_flags(struct io_ring_ctx *ctx)
323{
324 if (ctx->flags & IORING_SETUP_CQE_MIXED)
325 return IORING_CQE_F_32;
326 return 0;
327}
328
329static inline void io_req_set_res32(struct io_kiocb *req, s32 res, u32 cflags,
330 __u64 extra1, __u64 extra2)
331{
332 req->cqe.res = res;
333 req->cqe.flags = cflags | ctx_cqe32_flags(ctx: req->ctx);
334 req->big_cqe.extra1 = extra1;
335 req->big_cqe.extra2 = extra2;
336}
337
338static inline void *io_uring_alloc_async_data(struct io_alloc_cache *cache,
339 struct io_kiocb *req)
340{
341 if (cache) {
342 req->async_data = io_cache_alloc(cache, GFP_KERNEL);
343 } else {
344 const struct io_issue_def *def = &io_issue_defs[req->opcode];
345
346 WARN_ON_ONCE(!def->async_size);
347 req->async_data = kmalloc(def->async_size, GFP_KERNEL);
348 }
349 if (req->async_data)
350 req->flags |= REQ_F_ASYNC_DATA;
351 return req->async_data;
352}
353
354static inline bool req_has_async_data(struct io_kiocb *req)
355{
356 return req->flags & REQ_F_ASYNC_DATA;
357}
358
359static inline void io_req_async_data_clear(struct io_kiocb *req,
360 io_req_flags_t extra_flags)
361{
362 req->flags &= ~(REQ_F_ASYNC_DATA|extra_flags);
363 req->async_data = NULL;
364}
365
366static inline void io_req_async_data_free(struct io_kiocb *req)
367{
368 kfree(objp: req->async_data);
369 io_req_async_data_clear(req, extra_flags: 0);
370}
371
372static inline void io_put_file(struct io_kiocb *req)
373{
374 if (!(req->flags & REQ_F_FIXED_FILE) && req->file)
375 fput(req->file);
376}
377
378static inline void io_ring_submit_unlock(struct io_ring_ctx *ctx,
379 unsigned issue_flags)
380{
381 lockdep_assert_held(&ctx->uring_lock);
382 if (unlikely(issue_flags & IO_URING_F_UNLOCKED))
383 mutex_unlock(lock: &ctx->uring_lock);
384}
385
386static inline void io_ring_submit_lock(struct io_ring_ctx *ctx,
387 unsigned issue_flags)
388{
389 /*
390 * "Normal" inline submissions always hold the uring_lock, since we
391 * grab it from the system call. Same is true for the SQPOLL offload.
392 * The only exception is when we've detached the request and issue it
393 * from an async worker thread, grab the lock for that case.
394 */
395 if (unlikely(issue_flags & IO_URING_F_UNLOCKED))
396 mutex_lock(&ctx->uring_lock);
397 lockdep_assert_held(&ctx->uring_lock);
398}
399
400static inline void io_commit_cqring(struct io_ring_ctx *ctx)
401{
402 /* order cqe stores with ring update */
403 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
404}
405
406static inline void __io_wq_wake(struct wait_queue_head *wq)
407{
408 /*
409 *
410 * Pass in EPOLLIN|EPOLL_URING_WAKE as the poll wakeup key. The latter
411 * set in the mask so that if we recurse back into our own poll
412 * waitqueue handlers, we know we have a dependency between eventfd or
413 * epoll and should terminate multishot poll at that point.
414 */
415 if (wq_has_sleeper(wq_head: wq))
416 __wake_up(wq_head: wq, TASK_NORMAL, nr: 0, poll_to_key(EPOLL_URING_WAKE | EPOLLIN));
417}
418
419static inline void io_poll_wq_wake(struct io_ring_ctx *ctx)
420{
421 __io_wq_wake(wq: &ctx->poll_wq);
422}
423
424static inline void io_cqring_wake(struct io_ring_ctx *ctx)
425{
426 /*
427 * Trigger waitqueue handler on all waiters on our waitqueue. This
428 * won't necessarily wake up all the tasks, io_should_wake() will make
429 * that decision.
430 */
431
432 __io_wq_wake(wq: &ctx->cq_wait);
433}
434
435static inline bool io_sqring_full(struct io_ring_ctx *ctx)
436{
437 struct io_rings *r = ctx->rings;
438
439 /*
440 * SQPOLL must use the actual sqring head, as using the cached_sq_head
441 * is race prone if the SQPOLL thread has grabbed entries but not yet
442 * committed them to the ring. For !SQPOLL, this doesn't matter, but
443 * since this helper is just used for SQPOLL sqring waits (or POLLOUT),
444 * just read the actual sqring head unconditionally.
445 */
446 return READ_ONCE(r->sq.tail) - READ_ONCE(r->sq.head) == ctx->sq_entries;
447}
448
449static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
450{
451 struct io_rings *rings = ctx->rings;
452 unsigned int entries;
453
454 /* make sure SQ entry isn't read before tail */
455 entries = smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
456 return min(entries, ctx->sq_entries);
457}
458
459static inline int io_run_task_work(void)
460{
461 bool ret = false;
462
463 /*
464 * Always check-and-clear the task_work notification signal. With how
465 * signaling works for task_work, we can find it set with nothing to
466 * run. We need to clear it for that case, like get_signal() does.
467 */
468 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
469 clear_notify_signal();
470 /*
471 * PF_IO_WORKER never returns to userspace, so check here if we have
472 * notify work that needs processing.
473 */
474 if (current->flags & PF_IO_WORKER) {
475 if (test_thread_flag(TIF_NOTIFY_RESUME)) {
476 __set_current_state(TASK_RUNNING);
477 resume_user_mode_work(NULL);
478 }
479 if (current->io_uring) {
480 unsigned int count = 0;
481
482 __set_current_state(TASK_RUNNING);
483 tctx_task_work_run(current->io_uring, UINT_MAX, count: &count);
484 if (count)
485 ret = true;
486 }
487 }
488 if (task_work_pending(current)) {
489 __set_current_state(TASK_RUNNING);
490 task_work_run();
491 ret = true;
492 }
493
494 return ret;
495}
496
497static inline bool io_local_work_pending(struct io_ring_ctx *ctx)
498{
499 return !llist_empty(head: &ctx->work_llist) || !llist_empty(head: &ctx->retry_llist);
500}
501
502static inline bool io_task_work_pending(struct io_ring_ctx *ctx)
503{
504 return task_work_pending(current) || io_local_work_pending(ctx);
505}
506
507static inline void io_tw_lock(struct io_ring_ctx *ctx, io_tw_token_t tw)
508{
509 lockdep_assert_held(&ctx->uring_lock);
510}
511
512/*
513 * Don't complete immediately but use deferred completion infrastructure.
514 * Protected by ->uring_lock and can only be used either with
515 * IO_URING_F_COMPLETE_DEFER or inside a tw handler holding the mutex.
516 */
517static inline void io_req_complete_defer(struct io_kiocb *req)
518 __must_hold(&req->ctx->uring_lock)
519{
520 struct io_submit_state *state = &req->ctx->submit_state;
521
522 lockdep_assert_held(&req->ctx->uring_lock);
523
524 wq_list_add_tail(node: &req->comp_list, list: &state->compl_reqs);
525}
526
527static inline void io_commit_cqring_flush(struct io_ring_ctx *ctx)
528{
529 if (unlikely(ctx->off_timeout_used ||
530 ctx->has_evfd || ctx->poll_activated))
531 __io_commit_cqring_flush(ctx);
532}
533
534static inline void io_get_task_refs(int nr)
535{
536 struct io_uring_task *tctx = current->io_uring;
537
538 tctx->cached_refs -= nr;
539 if (unlikely(tctx->cached_refs < 0))
540 io_task_refs_refill(tctx);
541}
542
543static inline bool io_req_cache_empty(struct io_ring_ctx *ctx)
544{
545 return !ctx->submit_state.free_list.next;
546}
547
548extern struct kmem_cache *req_cachep;
549
550static inline struct io_kiocb *io_extract_req(struct io_ring_ctx *ctx)
551{
552 struct io_kiocb *req;
553
554 req = container_of(ctx->submit_state.free_list.next, struct io_kiocb, comp_list);
555 wq_stack_extract(stack: &ctx->submit_state.free_list);
556 return req;
557}
558
559static inline bool io_alloc_req(struct io_ring_ctx *ctx, struct io_kiocb **req)
560{
561 if (unlikely(io_req_cache_empty(ctx))) {
562 if (!__io_alloc_req_refill(ctx))
563 return false;
564 }
565 *req = io_extract_req(ctx);
566 return true;
567}
568
569static inline bool io_allowed_defer_tw_run(struct io_ring_ctx *ctx)
570{
571 return likely(ctx->submitter_task == current);
572}
573
574static inline bool io_allowed_run_tw(struct io_ring_ctx *ctx)
575{
576 return likely(!(ctx->flags & IORING_SETUP_DEFER_TASKRUN) ||
577 ctx->submitter_task == current);
578}
579
580static inline void io_req_queue_tw_complete(struct io_kiocb *req, s32 res)
581{
582 io_req_set_res(req, res, cflags: 0);
583 req->io_task_work.func = io_req_task_complete;
584 io_req_task_work_add(req);
585}
586
587static inline bool io_file_can_poll(struct io_kiocb *req)
588{
589 if (req->flags & REQ_F_CAN_POLL)
590 return true;
591 if (req->file && file_can_poll(file: req->file)) {
592 req->flags |= REQ_F_CAN_POLL;
593 return true;
594 }
595 return false;
596}
597
598static inline ktime_t io_get_time(struct io_ring_ctx *ctx)
599{
600 if (ctx->clockid == CLOCK_MONOTONIC)
601 return ktime_get();
602
603 return ktime_get_with_offset(offs: ctx->clock_offset);
604}
605
606enum {
607 IO_CHECK_CQ_OVERFLOW_BIT,
608 IO_CHECK_CQ_DROPPED_BIT,
609};
610
611static inline bool io_has_work(struct io_ring_ctx *ctx)
612{
613 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
614 io_local_work_pending(ctx);
615}
616#endif
617

source code of linux/io_uring/io_uring.h