1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
5 *
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
8 *
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
16 * CQ entries.
17 *
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
23 * head will do).
24 *
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
28 * between.
29 *
30 * Also see the examples in the liburing library:
31 *
32 * git://git.kernel.dk/liburing
33 *
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
38 *
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
41 */
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <net/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
49 #include <linux/bits.h>
50
51 #include <linux/sched/signal.h>
52 #include <linux/fs.h>
53 #include <linux/file.h>
54 #include <linux/mm.h>
55 #include <linux/mman.h>
56 #include <linux/percpu.h>
57 #include <linux/slab.h>
58 #include <linux/bvec.h>
59 #include <linux/net.h>
60 #include <net/sock.h>
61 #include <linux/anon_inodes.h>
62 #include <linux/sched/mm.h>
63 #include <linux/uaccess.h>
64 #include <linux/nospec.h>
65 #include <linux/fsnotify.h>
66 #include <linux/fadvise.h>
67 #include <linux/task_work.h>
68 #include <linux/io_uring.h>
69 #include <linux/io_uring/cmd.h>
70 #include <linux/audit.h>
71 #include <linux/security.h>
72 #include <linux/jump_label.h>
73 #include <asm/shmparam.h>
74
75 #define CREATE_TRACE_POINTS
76 #include <trace/events/io_uring.h>
77
78 #include <uapi/linux/io_uring.h>
79
80 #include "io-wq.h"
81
82 #include "io_uring.h"
83 #include "opdef.h"
84 #include "refs.h"
85 #include "tctx.h"
86 #include "register.h"
87 #include "sqpoll.h"
88 #include "fdinfo.h"
89 #include "kbuf.h"
90 #include "rsrc.h"
91 #include "cancel.h"
92 #include "net.h"
93 #include "notif.h"
94 #include "waitid.h"
95 #include "futex.h"
96 #include "napi.h"
97 #include "uring_cmd.h"
98 #include "msg_ring.h"
99 #include "memmap.h"
100
101 #include "timeout.h"
102 #include "poll.h"
103 #include "rw.h"
104 #include "alloc_cache.h"
105 #include "eventfd.h"
106
107 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
108 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
109
110 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
111 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
112
113 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
114 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
115 REQ_F_ASYNC_DATA)
116
117 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
118 IO_REQ_CLEAN_FLAGS)
119
120 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
121
122 #define IO_COMPL_BATCH 32
123 #define IO_REQ_ALLOC_BATCH 8
124 #define IO_LOCAL_TW_DEFAULT_MAX 20
125
126 struct io_defer_entry {
127 struct list_head list;
128 struct io_kiocb *req;
129 u32 seq;
130 };
131
132 /* requests with any of those set should undergo io_disarm_next() */
133 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
134 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
135
136 /*
137 * No waiters. It's larger than any valid value of the tw counter
138 * so that tests against ->cq_wait_nr would fail and skip wake_up().
139 */
140 #define IO_CQ_WAKE_INIT (-1U)
141 /* Forced wake up if there is a waiter regardless of ->cq_wait_nr */
142 #define IO_CQ_WAKE_FORCE (IO_CQ_WAKE_INIT >> 1)
143
144 static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
145 struct io_uring_task *tctx,
146 bool cancel_all);
147
148 static void io_queue_sqe(struct io_kiocb *req);
149
150 static __read_mostly DEFINE_STATIC_KEY_FALSE(io_key_has_sqarray);
151
152 struct kmem_cache *req_cachep;
153 static struct workqueue_struct *iou_wq __ro_after_init;
154
155 static int __read_mostly sysctl_io_uring_disabled;
156 static int __read_mostly sysctl_io_uring_group = -1;
157
158 #ifdef CONFIG_SYSCTL
159 static struct ctl_table kernel_io_uring_disabled_table[] = {
160 {
161 .procname = "io_uring_disabled",
162 .data = &sysctl_io_uring_disabled,
163 .maxlen = sizeof(sysctl_io_uring_disabled),
164 .mode = 0644,
165 .proc_handler = proc_dointvec_minmax,
166 .extra1 = SYSCTL_ZERO,
167 .extra2 = SYSCTL_TWO,
168 },
169 {
170 .procname = "io_uring_group",
171 .data = &sysctl_io_uring_group,
172 .maxlen = sizeof(gid_t),
173 .mode = 0644,
174 .proc_handler = proc_dointvec,
175 },
176 };
177 #endif
178
__io_cqring_events(struct io_ring_ctx * ctx)179 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
180 {
181 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
182 }
183
__io_cqring_events_user(struct io_ring_ctx * ctx)184 static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx)
185 {
186 return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head);
187 }
188
io_match_linked(struct io_kiocb * head)189 static bool io_match_linked(struct io_kiocb *head)
190 {
191 struct io_kiocb *req;
192
193 io_for_each_link(req, head) {
194 if (req->flags & REQ_F_INFLIGHT)
195 return true;
196 }
197 return false;
198 }
199
200 /*
201 * As io_match_task() but protected against racing with linked timeouts.
202 * User must not hold timeout_lock.
203 */
io_match_task_safe(struct io_kiocb * head,struct io_uring_task * tctx,bool cancel_all)204 bool io_match_task_safe(struct io_kiocb *head, struct io_uring_task *tctx,
205 bool cancel_all)
206 {
207 bool matched;
208
209 if (tctx && head->tctx != tctx)
210 return false;
211 if (cancel_all)
212 return true;
213
214 if (head->flags & REQ_F_LINK_TIMEOUT) {
215 struct io_ring_ctx *ctx = head->ctx;
216
217 /* protect against races with linked timeouts */
218 raw_spin_lock_irq(&ctx->timeout_lock);
219 matched = io_match_linked(head);
220 raw_spin_unlock_irq(&ctx->timeout_lock);
221 } else {
222 matched = io_match_linked(head);
223 }
224 return matched;
225 }
226
req_fail_link_node(struct io_kiocb * req,int res)227 static inline void req_fail_link_node(struct io_kiocb *req, int res)
228 {
229 req_set_fail(req);
230 io_req_set_res(req, res, 0);
231 }
232
io_req_add_to_cache(struct io_kiocb * req,struct io_ring_ctx * ctx)233 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
234 {
235 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
236 }
237
io_ring_ctx_ref_free(struct percpu_ref * ref)238 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
239 {
240 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
241
242 complete(&ctx->ref_comp);
243 }
244
io_fallback_req_func(struct work_struct * work)245 static __cold void io_fallback_req_func(struct work_struct *work)
246 {
247 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
248 fallback_work.work);
249 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
250 struct io_kiocb *req, *tmp;
251 struct io_tw_state ts = {};
252
253 percpu_ref_get(&ctx->refs);
254 mutex_lock(&ctx->uring_lock);
255 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
256 req->io_task_work.func(req, &ts);
257 io_submit_flush_completions(ctx);
258 mutex_unlock(&ctx->uring_lock);
259 percpu_ref_put(&ctx->refs);
260 }
261
io_alloc_hash_table(struct io_hash_table * table,unsigned bits)262 static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
263 {
264 unsigned int hash_buckets;
265 int i;
266
267 do {
268 hash_buckets = 1U << bits;
269 table->hbs = kvmalloc_array(hash_buckets, sizeof(table->hbs[0]),
270 GFP_KERNEL_ACCOUNT);
271 if (table->hbs)
272 break;
273 if (bits == 1)
274 return -ENOMEM;
275 bits--;
276 } while (1);
277
278 table->hash_bits = bits;
279 for (i = 0; i < hash_buckets; i++)
280 INIT_HLIST_HEAD(&table->hbs[i].list);
281 return 0;
282 }
283
io_ring_ctx_alloc(struct io_uring_params * p)284 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
285 {
286 struct io_ring_ctx *ctx;
287 int hash_bits;
288 bool ret;
289
290 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
291 if (!ctx)
292 return NULL;
293
294 xa_init(&ctx->io_bl_xa);
295
296 /*
297 * Use 5 bits less than the max cq entries, that should give us around
298 * 32 entries per hash list if totally full and uniformly spread, but
299 * don't keep too many buckets to not overconsume memory.
300 */
301 hash_bits = ilog2(p->cq_entries) - 5;
302 hash_bits = clamp(hash_bits, 1, 8);
303 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
304 goto err;
305 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
306 0, GFP_KERNEL))
307 goto err;
308
309 ctx->flags = p->flags;
310 ctx->hybrid_poll_time = LLONG_MAX;
311 atomic_set(&ctx->cq_wait_nr, IO_CQ_WAKE_INIT);
312 init_waitqueue_head(&ctx->sqo_sq_wait);
313 INIT_LIST_HEAD(&ctx->sqd_list);
314 INIT_LIST_HEAD(&ctx->cq_overflow_list);
315 INIT_LIST_HEAD(&ctx->io_buffers_cache);
316 ret = io_alloc_cache_init(&ctx->apoll_cache, IO_POLL_ALLOC_CACHE_MAX,
317 sizeof(struct async_poll));
318 ret |= io_alloc_cache_init(&ctx->netmsg_cache, IO_ALLOC_CACHE_MAX,
319 sizeof(struct io_async_msghdr));
320 ret |= io_alloc_cache_init(&ctx->rw_cache, IO_ALLOC_CACHE_MAX,
321 sizeof(struct io_async_rw));
322 ret |= io_alloc_cache_init(&ctx->uring_cache, IO_ALLOC_CACHE_MAX,
323 sizeof(struct uring_cache));
324 spin_lock_init(&ctx->msg_lock);
325 ret |= io_alloc_cache_init(&ctx->msg_cache, IO_ALLOC_CACHE_MAX,
326 sizeof(struct io_kiocb));
327 ret |= io_futex_cache_init(ctx);
328 if (ret)
329 goto free_ref;
330 init_completion(&ctx->ref_comp);
331 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
332 mutex_init(&ctx->uring_lock);
333 init_waitqueue_head(&ctx->cq_wait);
334 init_waitqueue_head(&ctx->poll_wq);
335 spin_lock_init(&ctx->completion_lock);
336 raw_spin_lock_init(&ctx->timeout_lock);
337 INIT_WQ_LIST(&ctx->iopoll_list);
338 INIT_LIST_HEAD(&ctx->io_buffers_comp);
339 INIT_LIST_HEAD(&ctx->defer_list);
340 INIT_LIST_HEAD(&ctx->timeout_list);
341 INIT_LIST_HEAD(&ctx->ltimeout_list);
342 init_llist_head(&ctx->work_llist);
343 INIT_LIST_HEAD(&ctx->tctx_list);
344 ctx->submit_state.free_list.next = NULL;
345 INIT_HLIST_HEAD(&ctx->waitid_list);
346 #ifdef CONFIG_FUTEX
347 INIT_HLIST_HEAD(&ctx->futex_list);
348 #endif
349 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
350 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
351 INIT_HLIST_HEAD(&ctx->cancelable_uring_cmd);
352 io_napi_init(ctx);
353 mutex_init(&ctx->resize_lock);
354
355 return ctx;
356
357 free_ref:
358 percpu_ref_exit(&ctx->refs);
359 err:
360 io_alloc_cache_free(&ctx->apoll_cache, kfree);
361 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
362 io_alloc_cache_free(&ctx->rw_cache, io_rw_cache_free);
363 io_alloc_cache_free(&ctx->uring_cache, kfree);
364 io_alloc_cache_free(&ctx->msg_cache, io_msg_cache_free);
365 io_futex_cache_free(ctx);
366 kvfree(ctx->cancel_table.hbs);
367 xa_destroy(&ctx->io_bl_xa);
368 kfree(ctx);
369 return NULL;
370 }
371
io_account_cq_overflow(struct io_ring_ctx * ctx)372 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
373 {
374 struct io_rings *r = ctx->rings;
375
376 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
377 ctx->cq_extra--;
378 }
379
req_need_defer(struct io_kiocb * req,u32 seq)380 static bool req_need_defer(struct io_kiocb *req, u32 seq)
381 {
382 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
383 struct io_ring_ctx *ctx = req->ctx;
384
385 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
386 }
387
388 return false;
389 }
390
io_clean_op(struct io_kiocb * req)391 static void io_clean_op(struct io_kiocb *req)
392 {
393 if (req->flags & REQ_F_BUFFER_SELECTED) {
394 spin_lock(&req->ctx->completion_lock);
395 io_kbuf_drop(req);
396 spin_unlock(&req->ctx->completion_lock);
397 }
398
399 if (req->flags & REQ_F_NEED_CLEANUP) {
400 const struct io_cold_def *def = &io_cold_defs[req->opcode];
401
402 if (def->cleanup)
403 def->cleanup(req);
404 }
405 if ((req->flags & REQ_F_POLLED) && req->apoll) {
406 kfree(req->apoll->double_poll);
407 kfree(req->apoll);
408 req->apoll = NULL;
409 }
410 if (req->flags & REQ_F_INFLIGHT)
411 atomic_dec(&req->tctx->inflight_tracked);
412 if (req->flags & REQ_F_CREDS)
413 put_cred(req->creds);
414 if (req->flags & REQ_F_ASYNC_DATA) {
415 kfree(req->async_data);
416 req->async_data = NULL;
417 }
418 req->flags &= ~IO_REQ_CLEAN_FLAGS;
419 }
420
io_req_track_inflight(struct io_kiocb * req)421 static inline void io_req_track_inflight(struct io_kiocb *req)
422 {
423 if (!(req->flags & REQ_F_INFLIGHT)) {
424 req->flags |= REQ_F_INFLIGHT;
425 atomic_inc(&req->tctx->inflight_tracked);
426 }
427 }
428
__io_prep_linked_timeout(struct io_kiocb * req)429 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
430 {
431 if (WARN_ON_ONCE(!req->link))
432 return NULL;
433
434 req->flags &= ~REQ_F_ARM_LTIMEOUT;
435 req->flags |= REQ_F_LINK_TIMEOUT;
436
437 /* linked timeouts should have two refs once prep'ed */
438 io_req_set_refcount(req);
439 __io_req_set_refcount(req->link, 2);
440 return req->link;
441 }
442
io_prep_linked_timeout(struct io_kiocb * req)443 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
444 {
445 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
446 return NULL;
447 return __io_prep_linked_timeout(req);
448 }
449
__io_arm_ltimeout(struct io_kiocb * req)450 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
451 {
452 io_queue_linked_timeout(__io_prep_linked_timeout(req));
453 }
454
io_arm_ltimeout(struct io_kiocb * req)455 static inline void io_arm_ltimeout(struct io_kiocb *req)
456 {
457 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
458 __io_arm_ltimeout(req);
459 }
460
io_prep_async_work(struct io_kiocb * req)461 static void io_prep_async_work(struct io_kiocb *req)
462 {
463 const struct io_issue_def *def = &io_issue_defs[req->opcode];
464 struct io_ring_ctx *ctx = req->ctx;
465
466 if (!(req->flags & REQ_F_CREDS)) {
467 req->flags |= REQ_F_CREDS;
468 req->creds = get_current_cred();
469 }
470
471 req->work.list.next = NULL;
472 atomic_set(&req->work.flags, 0);
473 if (req->flags & REQ_F_FORCE_ASYNC)
474 atomic_or(IO_WQ_WORK_CONCURRENT, &req->work.flags);
475
476 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
477 req->flags |= io_file_get_flags(req->file);
478
479 if (req->file && (req->flags & REQ_F_ISREG)) {
480 bool should_hash = def->hash_reg_file;
481
482 /* don't serialize this request if the fs doesn't need it */
483 if (should_hash && (req->file->f_flags & O_DIRECT) &&
484 (req->file->f_op->fop_flags & FOP_DIO_PARALLEL_WRITE))
485 should_hash = false;
486 if (should_hash || (ctx->flags & IORING_SETUP_IOPOLL))
487 io_wq_hash_work(&req->work, file_inode(req->file));
488 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
489 if (def->unbound_nonreg_file)
490 atomic_or(IO_WQ_WORK_UNBOUND, &req->work.flags);
491 }
492 }
493
io_prep_async_link(struct io_kiocb * req)494 static void io_prep_async_link(struct io_kiocb *req)
495 {
496 struct io_kiocb *cur;
497
498 if (req->flags & REQ_F_LINK_TIMEOUT) {
499 struct io_ring_ctx *ctx = req->ctx;
500
501 raw_spin_lock_irq(&ctx->timeout_lock);
502 io_for_each_link(cur, req)
503 io_prep_async_work(cur);
504 raw_spin_unlock_irq(&ctx->timeout_lock);
505 } else {
506 io_for_each_link(cur, req)
507 io_prep_async_work(cur);
508 }
509 }
510
io_queue_iowq(struct io_kiocb * req)511 static void io_queue_iowq(struct io_kiocb *req)
512 {
513 struct io_kiocb *link = io_prep_linked_timeout(req);
514 struct io_uring_task *tctx = req->tctx;
515
516 BUG_ON(!tctx);
517
518 if ((current->flags & PF_KTHREAD) || !tctx->io_wq) {
519 io_req_task_queue_fail(req, -ECANCELED);
520 return;
521 }
522
523 /* init ->work of the whole link before punting */
524 io_prep_async_link(req);
525
526 /*
527 * Not expected to happen, but if we do have a bug where this _can_
528 * happen, catch it here and ensure the request is marked as
529 * canceled. That will make io-wq go through the usual work cancel
530 * procedure rather than attempt to run this request (or create a new
531 * worker for it).
532 */
533 if (WARN_ON_ONCE(!same_thread_group(tctx->task, current)))
534 atomic_or(IO_WQ_WORK_CANCEL, &req->work.flags);
535
536 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
537 io_wq_enqueue(tctx->io_wq, &req->work);
538 if (link)
539 io_queue_linked_timeout(link);
540 }
541
io_req_queue_iowq_tw(struct io_kiocb * req,struct io_tw_state * ts)542 static void io_req_queue_iowq_tw(struct io_kiocb *req, struct io_tw_state *ts)
543 {
544 io_queue_iowq(req);
545 }
546
io_req_queue_iowq(struct io_kiocb * req)547 void io_req_queue_iowq(struct io_kiocb *req)
548 {
549 req->io_task_work.func = io_req_queue_iowq_tw;
550 io_req_task_work_add(req);
551 }
552
io_queue_deferred(struct io_ring_ctx * ctx)553 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
554 {
555 while (!list_empty(&ctx->defer_list)) {
556 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
557 struct io_defer_entry, list);
558
559 if (req_need_defer(de->req, de->seq))
560 break;
561 list_del_init(&de->list);
562 io_req_task_queue(de->req);
563 kfree(de);
564 }
565 }
566
__io_commit_cqring_flush(struct io_ring_ctx * ctx)567 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
568 {
569 if (ctx->poll_activated)
570 io_poll_wq_wake(ctx);
571 if (ctx->off_timeout_used)
572 io_flush_timeouts(ctx);
573 if (ctx->drain_active) {
574 spin_lock(&ctx->completion_lock);
575 io_queue_deferred(ctx);
576 spin_unlock(&ctx->completion_lock);
577 }
578 if (ctx->has_evfd)
579 io_eventfd_flush_signal(ctx);
580 }
581
__io_cq_lock(struct io_ring_ctx * ctx)582 static inline void __io_cq_lock(struct io_ring_ctx *ctx)
583 {
584 if (!ctx->lockless_cq)
585 spin_lock(&ctx->completion_lock);
586 }
587
io_cq_lock(struct io_ring_ctx * ctx)588 static inline void io_cq_lock(struct io_ring_ctx *ctx)
589 __acquires(ctx->completion_lock)
590 {
591 spin_lock(&ctx->completion_lock);
592 }
593
__io_cq_unlock_post(struct io_ring_ctx * ctx)594 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
595 {
596 io_commit_cqring(ctx);
597 if (!ctx->task_complete) {
598 if (!ctx->lockless_cq)
599 spin_unlock(&ctx->completion_lock);
600 /* IOPOLL rings only need to wake up if it's also SQPOLL */
601 if (!ctx->syscall_iopoll)
602 io_cqring_wake(ctx);
603 }
604 io_commit_cqring_flush(ctx);
605 }
606
io_cq_unlock_post(struct io_ring_ctx * ctx)607 static void io_cq_unlock_post(struct io_ring_ctx *ctx)
608 __releases(ctx->completion_lock)
609 {
610 io_commit_cqring(ctx);
611 spin_unlock(&ctx->completion_lock);
612 io_cqring_wake(ctx);
613 io_commit_cqring_flush(ctx);
614 }
615
__io_cqring_overflow_flush(struct io_ring_ctx * ctx,bool dying)616 static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool dying)
617 {
618 size_t cqe_size = sizeof(struct io_uring_cqe);
619
620 lockdep_assert_held(&ctx->uring_lock);
621
622 /* don't abort if we're dying, entries must get freed */
623 if (!dying && __io_cqring_events(ctx) == ctx->cq_entries)
624 return;
625
626 if (ctx->flags & IORING_SETUP_CQE32)
627 cqe_size <<= 1;
628
629 io_cq_lock(ctx);
630 while (!list_empty(&ctx->cq_overflow_list)) {
631 struct io_uring_cqe *cqe;
632 struct io_overflow_cqe *ocqe;
633
634 ocqe = list_first_entry(&ctx->cq_overflow_list,
635 struct io_overflow_cqe, list);
636
637 if (!dying) {
638 if (!io_get_cqe_overflow(ctx, &cqe, true))
639 break;
640 memcpy(cqe, &ocqe->cqe, cqe_size);
641 }
642 list_del(&ocqe->list);
643 kfree(ocqe);
644
645 /*
646 * For silly syzbot cases that deliberately overflow by huge
647 * amounts, check if we need to resched and drop and
648 * reacquire the locks if so. Nothing real would ever hit this.
649 * Ideally we'd have a non-posting unlock for this, but hard
650 * to care for a non-real case.
651 */
652 if (need_resched()) {
653 io_cq_unlock_post(ctx);
654 mutex_unlock(&ctx->uring_lock);
655 cond_resched();
656 mutex_lock(&ctx->uring_lock);
657 io_cq_lock(ctx);
658 }
659 }
660
661 if (list_empty(&ctx->cq_overflow_list)) {
662 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
663 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
664 }
665 io_cq_unlock_post(ctx);
666 }
667
io_cqring_overflow_kill(struct io_ring_ctx * ctx)668 static void io_cqring_overflow_kill(struct io_ring_ctx *ctx)
669 {
670 if (ctx->rings)
671 __io_cqring_overflow_flush(ctx, true);
672 }
673
io_cqring_do_overflow_flush(struct io_ring_ctx * ctx)674 static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx)
675 {
676 mutex_lock(&ctx->uring_lock);
677 __io_cqring_overflow_flush(ctx, false);
678 mutex_unlock(&ctx->uring_lock);
679 }
680
681 /* must to be called somewhat shortly after putting a request */
io_put_task(struct io_kiocb * req)682 static inline void io_put_task(struct io_kiocb *req)
683 {
684 struct io_uring_task *tctx = req->tctx;
685
686 if (likely(tctx->task == current)) {
687 tctx->cached_refs++;
688 } else {
689 percpu_counter_sub(&tctx->inflight, 1);
690 if (unlikely(atomic_read(&tctx->in_cancel)))
691 wake_up(&tctx->wait);
692 put_task_struct(tctx->task);
693 }
694 }
695
io_task_refs_refill(struct io_uring_task * tctx)696 void io_task_refs_refill(struct io_uring_task *tctx)
697 {
698 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
699
700 percpu_counter_add(&tctx->inflight, refill);
701 refcount_add(refill, ¤t->usage);
702 tctx->cached_refs += refill;
703 }
704
io_uring_drop_tctx_refs(struct task_struct * task)705 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
706 {
707 struct io_uring_task *tctx = task->io_uring;
708 unsigned int refs = tctx->cached_refs;
709
710 if (refs) {
711 tctx->cached_refs = 0;
712 percpu_counter_sub(&tctx->inflight, refs);
713 put_task_struct_many(task, refs);
714 }
715 }
716
io_cqring_event_overflow(struct io_ring_ctx * ctx,u64 user_data,s32 res,u32 cflags,u64 extra1,u64 extra2)717 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
718 s32 res, u32 cflags, u64 extra1, u64 extra2)
719 {
720 struct io_overflow_cqe *ocqe;
721 size_t ocq_size = sizeof(struct io_overflow_cqe);
722 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
723
724 lockdep_assert_held(&ctx->completion_lock);
725
726 if (is_cqe32)
727 ocq_size += sizeof(struct io_uring_cqe);
728
729 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
730 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
731 if (!ocqe) {
732 /*
733 * If we're in ring overflow flush mode, or in task cancel mode,
734 * or cannot allocate an overflow entry, then we need to drop it
735 * on the floor.
736 */
737 io_account_cq_overflow(ctx);
738 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
739 return false;
740 }
741 if (list_empty(&ctx->cq_overflow_list)) {
742 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
743 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
744
745 }
746 ocqe->cqe.user_data = user_data;
747 ocqe->cqe.res = res;
748 ocqe->cqe.flags = cflags;
749 if (is_cqe32) {
750 ocqe->cqe.big_cqe[0] = extra1;
751 ocqe->cqe.big_cqe[1] = extra2;
752 }
753 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
754 return true;
755 }
756
io_req_cqe_overflow(struct io_kiocb * req)757 static void io_req_cqe_overflow(struct io_kiocb *req)
758 {
759 io_cqring_event_overflow(req->ctx, req->cqe.user_data,
760 req->cqe.res, req->cqe.flags,
761 req->big_cqe.extra1, req->big_cqe.extra2);
762 memset(&req->big_cqe, 0, sizeof(req->big_cqe));
763 }
764
765 /*
766 * writes to the cq entry need to come after reading head; the
767 * control dependency is enough as we're using WRITE_ONCE to
768 * fill the cq entry
769 */
io_cqe_cache_refill(struct io_ring_ctx * ctx,bool overflow)770 bool io_cqe_cache_refill(struct io_ring_ctx *ctx, bool overflow)
771 {
772 struct io_rings *rings = ctx->rings;
773 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
774 unsigned int free, queued, len;
775
776 /*
777 * Posting into the CQ when there are pending overflowed CQEs may break
778 * ordering guarantees, which will affect links, F_MORE users and more.
779 * Force overflow the completion.
780 */
781 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
782 return false;
783
784 /* userspace may cheat modifying the tail, be safe and do min */
785 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
786 free = ctx->cq_entries - queued;
787 /* we need a contiguous range, limit based on the current array offset */
788 len = min(free, ctx->cq_entries - off);
789 if (!len)
790 return false;
791
792 if (ctx->flags & IORING_SETUP_CQE32) {
793 off <<= 1;
794 len <<= 1;
795 }
796
797 ctx->cqe_cached = &rings->cqes[off];
798 ctx->cqe_sentinel = ctx->cqe_cached + len;
799 return true;
800 }
801
io_fill_cqe_aux(struct io_ring_ctx * ctx,u64 user_data,s32 res,u32 cflags)802 static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
803 u32 cflags)
804 {
805 struct io_uring_cqe *cqe;
806
807 ctx->cq_extra++;
808
809 /*
810 * If we can't get a cq entry, userspace overflowed the
811 * submission (by quite a lot). Increment the overflow count in
812 * the ring.
813 */
814 if (likely(io_get_cqe(ctx, &cqe))) {
815 WRITE_ONCE(cqe->user_data, user_data);
816 WRITE_ONCE(cqe->res, res);
817 WRITE_ONCE(cqe->flags, cflags);
818
819 if (ctx->flags & IORING_SETUP_CQE32) {
820 WRITE_ONCE(cqe->big_cqe[0], 0);
821 WRITE_ONCE(cqe->big_cqe[1], 0);
822 }
823
824 trace_io_uring_complete(ctx, NULL, cqe);
825 return true;
826 }
827 return false;
828 }
829
__io_post_aux_cqe(struct io_ring_ctx * ctx,u64 user_data,s32 res,u32 cflags)830 static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res,
831 u32 cflags)
832 {
833 bool filled;
834
835 filled = io_fill_cqe_aux(ctx, user_data, res, cflags);
836 if (!filled)
837 filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
838
839 return filled;
840 }
841
io_post_aux_cqe(struct io_ring_ctx * ctx,u64 user_data,s32 res,u32 cflags)842 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags)
843 {
844 bool filled;
845
846 io_cq_lock(ctx);
847 filled = __io_post_aux_cqe(ctx, user_data, res, cflags);
848 io_cq_unlock_post(ctx);
849 return filled;
850 }
851
852 /*
853 * Must be called from inline task_work so we now a flush will happen later,
854 * and obviously with ctx->uring_lock held (tw always has that).
855 */
io_add_aux_cqe(struct io_ring_ctx * ctx,u64 user_data,s32 res,u32 cflags)856 void io_add_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags)
857 {
858 if (!io_fill_cqe_aux(ctx, user_data, res, cflags)) {
859 spin_lock(&ctx->completion_lock);
860 io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
861 spin_unlock(&ctx->completion_lock);
862 }
863 ctx->submit_state.cq_flush = true;
864 }
865
866 /*
867 * A helper for multishot requests posting additional CQEs.
868 * Should only be used from a task_work including IO_URING_F_MULTISHOT.
869 */
io_req_post_cqe(struct io_kiocb * req,s32 res,u32 cflags)870 bool io_req_post_cqe(struct io_kiocb *req, s32 res, u32 cflags)
871 {
872 struct io_ring_ctx *ctx = req->ctx;
873 bool posted;
874
875 lockdep_assert(!io_wq_current_is_worker());
876 lockdep_assert_held(&ctx->uring_lock);
877
878 __io_cq_lock(ctx);
879 posted = io_fill_cqe_aux(ctx, req->cqe.user_data, res, cflags);
880 ctx->submit_state.cq_flush = true;
881 __io_cq_unlock_post(ctx);
882 return posted;
883 }
884
io_req_complete_post(struct io_kiocb * req,unsigned issue_flags)885 static void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
886 {
887 struct io_ring_ctx *ctx = req->ctx;
888
889 /*
890 * All execution paths but io-wq use the deferred completions by
891 * passing IO_URING_F_COMPLETE_DEFER and thus should not end up here.
892 */
893 if (WARN_ON_ONCE(!(issue_flags & IO_URING_F_IOWQ)))
894 return;
895
896 /*
897 * Handle special CQ sync cases via task_work. DEFER_TASKRUN requires
898 * the submitter task context, IOPOLL protects with uring_lock.
899 */
900 if (ctx->task_complete || (ctx->flags & IORING_SETUP_IOPOLL)) {
901 req->io_task_work.func = io_req_task_complete;
902 io_req_task_work_add(req);
903 return;
904 }
905
906 io_cq_lock(ctx);
907 if (!(req->flags & REQ_F_CQE_SKIP)) {
908 if (!io_fill_cqe_req(ctx, req))
909 io_req_cqe_overflow(req);
910 }
911 io_cq_unlock_post(ctx);
912
913 /*
914 * We don't free the request here because we know it's called from
915 * io-wq only, which holds a reference, so it cannot be the last put.
916 */
917 req_ref_put(req);
918 }
919
io_req_defer_failed(struct io_kiocb * req,s32 res)920 void io_req_defer_failed(struct io_kiocb *req, s32 res)
921 __must_hold(&ctx->uring_lock)
922 {
923 const struct io_cold_def *def = &io_cold_defs[req->opcode];
924
925 lockdep_assert_held(&req->ctx->uring_lock);
926
927 req_set_fail(req);
928 io_req_set_res(req, res, io_put_kbuf(req, res, IO_URING_F_UNLOCKED));
929 if (def->fail)
930 def->fail(req);
931 io_req_complete_defer(req);
932 }
933
934 /*
935 * Don't initialise the fields below on every allocation, but do that in
936 * advance and keep them valid across allocations.
937 */
io_preinit_req(struct io_kiocb * req,struct io_ring_ctx * ctx)938 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
939 {
940 req->ctx = ctx;
941 req->buf_node = NULL;
942 req->file_node = NULL;
943 req->link = NULL;
944 req->async_data = NULL;
945 /* not necessary, but safer to zero */
946 memset(&req->cqe, 0, sizeof(req->cqe));
947 memset(&req->big_cqe, 0, sizeof(req->big_cqe));
948 }
949
950 /*
951 * A request might get retired back into the request caches even before opcode
952 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
953 * Because of that, io_alloc_req() should be called only under ->uring_lock
954 * and with extra caution to not get a request that is still worked on.
955 */
__io_alloc_req_refill(struct io_ring_ctx * ctx)956 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
957 __must_hold(&ctx->uring_lock)
958 {
959 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
960 void *reqs[IO_REQ_ALLOC_BATCH];
961 int ret;
962
963 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
964
965 /*
966 * Bulk alloc is all-or-nothing. If we fail to get a batch,
967 * retry single alloc to be on the safe side.
968 */
969 if (unlikely(ret <= 0)) {
970 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
971 if (!reqs[0])
972 return false;
973 ret = 1;
974 }
975
976 percpu_ref_get_many(&ctx->refs, ret);
977 while (ret--) {
978 struct io_kiocb *req = reqs[ret];
979
980 io_preinit_req(req, ctx);
981 io_req_add_to_cache(req, ctx);
982 }
983 return true;
984 }
985
io_free_req(struct io_kiocb * req)986 __cold void io_free_req(struct io_kiocb *req)
987 {
988 /* refs were already put, restore them for io_req_task_complete() */
989 req->flags &= ~REQ_F_REFCOUNT;
990 /* we only want to free it, don't post CQEs */
991 req->flags |= REQ_F_CQE_SKIP;
992 req->io_task_work.func = io_req_task_complete;
993 io_req_task_work_add(req);
994 }
995
__io_req_find_next_prep(struct io_kiocb * req)996 static void __io_req_find_next_prep(struct io_kiocb *req)
997 {
998 struct io_ring_ctx *ctx = req->ctx;
999
1000 spin_lock(&ctx->completion_lock);
1001 io_disarm_next(req);
1002 spin_unlock(&ctx->completion_lock);
1003 }
1004
io_req_find_next(struct io_kiocb * req)1005 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1006 {
1007 struct io_kiocb *nxt;
1008
1009 /*
1010 * If LINK is set, we have dependent requests in this chain. If we
1011 * didn't fail this request, queue the first one up, moving any other
1012 * dependencies to the next request. In case of failure, fail the rest
1013 * of the chain.
1014 */
1015 if (unlikely(req->flags & IO_DISARM_MASK))
1016 __io_req_find_next_prep(req);
1017 nxt = req->link;
1018 req->link = NULL;
1019 return nxt;
1020 }
1021
ctx_flush_and_put(struct io_ring_ctx * ctx,struct io_tw_state * ts)1022 static void ctx_flush_and_put(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1023 {
1024 if (!ctx)
1025 return;
1026 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1027 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1028
1029 io_submit_flush_completions(ctx);
1030 mutex_unlock(&ctx->uring_lock);
1031 percpu_ref_put(&ctx->refs);
1032 }
1033
1034 /*
1035 * Run queued task_work, returning the number of entries processed in *count.
1036 * If more entries than max_entries are available, stop processing once this
1037 * is reached and return the rest of the list.
1038 */
io_handle_tw_list(struct llist_node * node,unsigned int * count,unsigned int max_entries)1039 struct llist_node *io_handle_tw_list(struct llist_node *node,
1040 unsigned int *count,
1041 unsigned int max_entries)
1042 {
1043 struct io_ring_ctx *ctx = NULL;
1044 struct io_tw_state ts = { };
1045
1046 do {
1047 struct llist_node *next = node->next;
1048 struct io_kiocb *req = container_of(node, struct io_kiocb,
1049 io_task_work.node);
1050
1051 if (req->ctx != ctx) {
1052 ctx_flush_and_put(ctx, &ts);
1053 ctx = req->ctx;
1054 mutex_lock(&ctx->uring_lock);
1055 percpu_ref_get(&ctx->refs);
1056 }
1057 INDIRECT_CALL_2(req->io_task_work.func,
1058 io_poll_task_func, io_req_rw_complete,
1059 req, &ts);
1060 node = next;
1061 (*count)++;
1062 if (unlikely(need_resched())) {
1063 ctx_flush_and_put(ctx, &ts);
1064 ctx = NULL;
1065 cond_resched();
1066 }
1067 } while (node && *count < max_entries);
1068
1069 ctx_flush_and_put(ctx, &ts);
1070 return node;
1071 }
1072
__io_fallback_tw(struct llist_node * node,bool sync)1073 static __cold void __io_fallback_tw(struct llist_node *node, bool sync)
1074 {
1075 struct io_ring_ctx *last_ctx = NULL;
1076 struct io_kiocb *req;
1077
1078 while (node) {
1079 req = container_of(node, struct io_kiocb, io_task_work.node);
1080 node = node->next;
1081 if (sync && last_ctx != req->ctx) {
1082 if (last_ctx) {
1083 flush_delayed_work(&last_ctx->fallback_work);
1084 percpu_ref_put(&last_ctx->refs);
1085 }
1086 last_ctx = req->ctx;
1087 percpu_ref_get(&last_ctx->refs);
1088 }
1089 if (llist_add(&req->io_task_work.node,
1090 &req->ctx->fallback_llist))
1091 schedule_delayed_work(&req->ctx->fallback_work, 1);
1092 }
1093
1094 if (last_ctx) {
1095 flush_delayed_work(&last_ctx->fallback_work);
1096 percpu_ref_put(&last_ctx->refs);
1097 }
1098 }
1099
io_fallback_tw(struct io_uring_task * tctx,bool sync)1100 static void io_fallback_tw(struct io_uring_task *tctx, bool sync)
1101 {
1102 struct llist_node *node = llist_del_all(&tctx->task_list);
1103
1104 __io_fallback_tw(node, sync);
1105 }
1106
tctx_task_work_run(struct io_uring_task * tctx,unsigned int max_entries,unsigned int * count)1107 struct llist_node *tctx_task_work_run(struct io_uring_task *tctx,
1108 unsigned int max_entries,
1109 unsigned int *count)
1110 {
1111 struct llist_node *node;
1112
1113 if (unlikely(current->flags & PF_EXITING)) {
1114 io_fallback_tw(tctx, true);
1115 return NULL;
1116 }
1117
1118 node = llist_del_all(&tctx->task_list);
1119 if (node) {
1120 node = llist_reverse_order(node);
1121 node = io_handle_tw_list(node, count, max_entries);
1122 }
1123
1124 /* relaxed read is enough as only the task itself sets ->in_cancel */
1125 if (unlikely(atomic_read(&tctx->in_cancel)))
1126 io_uring_drop_tctx_refs(current);
1127
1128 trace_io_uring_task_work_run(tctx, *count);
1129 return node;
1130 }
1131
tctx_task_work(struct callback_head * cb)1132 void tctx_task_work(struct callback_head *cb)
1133 {
1134 struct io_uring_task *tctx;
1135 struct llist_node *ret;
1136 unsigned int count = 0;
1137
1138 tctx = container_of(cb, struct io_uring_task, task_work);
1139 ret = tctx_task_work_run(tctx, UINT_MAX, &count);
1140 /* can't happen */
1141 WARN_ON_ONCE(ret);
1142 }
1143
io_req_local_work_add(struct io_kiocb * req,struct io_ring_ctx * ctx,unsigned flags)1144 static inline void io_req_local_work_add(struct io_kiocb *req,
1145 struct io_ring_ctx *ctx,
1146 unsigned flags)
1147 {
1148 unsigned nr_wait, nr_tw, nr_tw_prev;
1149 struct llist_node *head;
1150
1151 /* See comment above IO_CQ_WAKE_INIT */
1152 BUILD_BUG_ON(IO_CQ_WAKE_FORCE <= IORING_MAX_CQ_ENTRIES);
1153
1154 /*
1155 * We don't know how many reuqests is there in the link and whether
1156 * they can even be queued lazily, fall back to non-lazy.
1157 */
1158 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK))
1159 flags &= ~IOU_F_TWQ_LAZY_WAKE;
1160
1161 guard(rcu)();
1162
1163 head = READ_ONCE(ctx->work_llist.first);
1164 do {
1165 nr_tw_prev = 0;
1166 if (head) {
1167 struct io_kiocb *first_req = container_of(head,
1168 struct io_kiocb,
1169 io_task_work.node);
1170 /*
1171 * Might be executed at any moment, rely on
1172 * SLAB_TYPESAFE_BY_RCU to keep it alive.
1173 */
1174 nr_tw_prev = READ_ONCE(first_req->nr_tw);
1175 }
1176
1177 /*
1178 * Theoretically, it can overflow, but that's fine as one of
1179 * previous adds should've tried to wake the task.
1180 */
1181 nr_tw = nr_tw_prev + 1;
1182 if (!(flags & IOU_F_TWQ_LAZY_WAKE))
1183 nr_tw = IO_CQ_WAKE_FORCE;
1184
1185 req->nr_tw = nr_tw;
1186 req->io_task_work.node.next = head;
1187 } while (!try_cmpxchg(&ctx->work_llist.first, &head,
1188 &req->io_task_work.node));
1189
1190 /*
1191 * cmpxchg implies a full barrier, which pairs with the barrier
1192 * in set_current_state() on the io_cqring_wait() side. It's used
1193 * to ensure that either we see updated ->cq_wait_nr, or waiters
1194 * going to sleep will observe the work added to the list, which
1195 * is similar to the wait/wawke task state sync.
1196 */
1197
1198 if (!head) {
1199 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1200 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1201 if (ctx->has_evfd)
1202 io_eventfd_signal(ctx);
1203 }
1204
1205 nr_wait = atomic_read(&ctx->cq_wait_nr);
1206 /* not enough or no one is waiting */
1207 if (nr_tw < nr_wait)
1208 return;
1209 /* the previous add has already woken it up */
1210 if (nr_tw_prev >= nr_wait)
1211 return;
1212 wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE);
1213 }
1214
io_req_normal_work_add(struct io_kiocb * req)1215 static void io_req_normal_work_add(struct io_kiocb *req)
1216 {
1217 struct io_uring_task *tctx = req->tctx;
1218 struct io_ring_ctx *ctx = req->ctx;
1219
1220 /* task_work already pending, we're done */
1221 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1222 return;
1223
1224 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1225 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1226
1227 /* SQPOLL doesn't need the task_work added, it'll run it itself */
1228 if (ctx->flags & IORING_SETUP_SQPOLL) {
1229 struct io_sq_data *sqd = ctx->sq_data;
1230
1231 if (sqd->thread)
1232 __set_notify_signal(sqd->thread);
1233 return;
1234 }
1235
1236 if (likely(!task_work_add(tctx->task, &tctx->task_work, ctx->notify_method)))
1237 return;
1238
1239 io_fallback_tw(tctx, false);
1240 }
1241
__io_req_task_work_add(struct io_kiocb * req,unsigned flags)1242 void __io_req_task_work_add(struct io_kiocb *req, unsigned flags)
1243 {
1244 if (req->ctx->flags & IORING_SETUP_DEFER_TASKRUN)
1245 io_req_local_work_add(req, req->ctx, flags);
1246 else
1247 io_req_normal_work_add(req);
1248 }
1249
io_req_task_work_add_remote(struct io_kiocb * req,struct io_ring_ctx * ctx,unsigned flags)1250 void io_req_task_work_add_remote(struct io_kiocb *req, struct io_ring_ctx *ctx,
1251 unsigned flags)
1252 {
1253 if (WARN_ON_ONCE(!(ctx->flags & IORING_SETUP_DEFER_TASKRUN)))
1254 return;
1255 io_req_local_work_add(req, ctx, flags);
1256 }
1257
io_move_task_work_from_local(struct io_ring_ctx * ctx)1258 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1259 {
1260 struct llist_node *node = llist_del_all(&ctx->work_llist);
1261
1262 __io_fallback_tw(node, false);
1263 node = llist_del_all(&ctx->retry_llist);
1264 __io_fallback_tw(node, false);
1265 }
1266
io_run_local_work_continue(struct io_ring_ctx * ctx,int events,int min_events)1267 static bool io_run_local_work_continue(struct io_ring_ctx *ctx, int events,
1268 int min_events)
1269 {
1270 if (!io_local_work_pending(ctx))
1271 return false;
1272 if (events < min_events)
1273 return true;
1274 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1275 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1276 return false;
1277 }
1278
__io_run_local_work_loop(struct llist_node ** node,struct io_tw_state * ts,int events)1279 static int __io_run_local_work_loop(struct llist_node **node,
1280 struct io_tw_state *ts,
1281 int events)
1282 {
1283 int ret = 0;
1284
1285 while (*node) {
1286 struct llist_node *next = (*node)->next;
1287 struct io_kiocb *req = container_of(*node, struct io_kiocb,
1288 io_task_work.node);
1289 INDIRECT_CALL_2(req->io_task_work.func,
1290 io_poll_task_func, io_req_rw_complete,
1291 req, ts);
1292 *node = next;
1293 if (++ret >= events)
1294 break;
1295 }
1296
1297 return ret;
1298 }
1299
__io_run_local_work(struct io_ring_ctx * ctx,struct io_tw_state * ts,int min_events,int max_events)1300 static int __io_run_local_work(struct io_ring_ctx *ctx, struct io_tw_state *ts,
1301 int min_events, int max_events)
1302 {
1303 struct llist_node *node;
1304 unsigned int loops = 0;
1305 int ret = 0;
1306
1307 if (WARN_ON_ONCE(ctx->submitter_task != current))
1308 return -EEXIST;
1309 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1310 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1311 again:
1312 min_events -= ret;
1313 ret = __io_run_local_work_loop(&ctx->retry_llist.first, ts, max_events);
1314 if (ctx->retry_llist.first)
1315 goto retry_done;
1316
1317 /*
1318 * llists are in reverse order, flip it back the right way before
1319 * running the pending items.
1320 */
1321 node = llist_reverse_order(llist_del_all(&ctx->work_llist));
1322 ret += __io_run_local_work_loop(&node, ts, max_events - ret);
1323 ctx->retry_llist.first = node;
1324 loops++;
1325
1326 if (io_run_local_work_continue(ctx, ret, min_events))
1327 goto again;
1328 retry_done:
1329 io_submit_flush_completions(ctx);
1330 if (io_run_local_work_continue(ctx, ret, min_events))
1331 goto again;
1332
1333 trace_io_uring_local_work_run(ctx, ret, loops);
1334 return ret;
1335 }
1336
io_run_local_work_locked(struct io_ring_ctx * ctx,int min_events)1337 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx,
1338 int min_events)
1339 {
1340 struct io_tw_state ts = {};
1341
1342 if (!io_local_work_pending(ctx))
1343 return 0;
1344 return __io_run_local_work(ctx, &ts, min_events,
1345 max(IO_LOCAL_TW_DEFAULT_MAX, min_events));
1346 }
1347
io_run_local_work(struct io_ring_ctx * ctx,int min_events,int max_events)1348 static int io_run_local_work(struct io_ring_ctx *ctx, int min_events,
1349 int max_events)
1350 {
1351 struct io_tw_state ts = {};
1352 int ret;
1353
1354 mutex_lock(&ctx->uring_lock);
1355 ret = __io_run_local_work(ctx, &ts, min_events, max_events);
1356 mutex_unlock(&ctx->uring_lock);
1357 return ret;
1358 }
1359
io_req_task_cancel(struct io_kiocb * req,struct io_tw_state * ts)1360 static void io_req_task_cancel(struct io_kiocb *req, struct io_tw_state *ts)
1361 {
1362 io_tw_lock(req->ctx, ts);
1363 io_req_defer_failed(req, req->cqe.res);
1364 }
1365
io_req_task_submit(struct io_kiocb * req,struct io_tw_state * ts)1366 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts)
1367 {
1368 io_tw_lock(req->ctx, ts);
1369 if (unlikely(io_should_terminate_tw()))
1370 io_req_defer_failed(req, -EFAULT);
1371 else if (req->flags & REQ_F_FORCE_ASYNC)
1372 io_queue_iowq(req);
1373 else
1374 io_queue_sqe(req);
1375 }
1376
io_req_task_queue_fail(struct io_kiocb * req,int ret)1377 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1378 {
1379 io_req_set_res(req, ret, 0);
1380 req->io_task_work.func = io_req_task_cancel;
1381 io_req_task_work_add(req);
1382 }
1383
io_req_task_queue(struct io_kiocb * req)1384 void io_req_task_queue(struct io_kiocb *req)
1385 {
1386 req->io_task_work.func = io_req_task_submit;
1387 io_req_task_work_add(req);
1388 }
1389
io_queue_next(struct io_kiocb * req)1390 void io_queue_next(struct io_kiocb *req)
1391 {
1392 struct io_kiocb *nxt = io_req_find_next(req);
1393
1394 if (nxt)
1395 io_req_task_queue(nxt);
1396 }
1397
io_free_batch_list(struct io_ring_ctx * ctx,struct io_wq_work_node * node)1398 static void io_free_batch_list(struct io_ring_ctx *ctx,
1399 struct io_wq_work_node *node)
1400 __must_hold(&ctx->uring_lock)
1401 {
1402 do {
1403 struct io_kiocb *req = container_of(node, struct io_kiocb,
1404 comp_list);
1405
1406 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1407 if (req->flags & REQ_F_REFCOUNT) {
1408 node = req->comp_list.next;
1409 if (!req_ref_put_and_test(req))
1410 continue;
1411 }
1412 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1413 struct async_poll *apoll = req->apoll;
1414
1415 if (apoll->double_poll)
1416 kfree(apoll->double_poll);
1417 if (!io_alloc_cache_put(&ctx->apoll_cache, apoll))
1418 kfree(apoll);
1419 req->flags &= ~REQ_F_POLLED;
1420 }
1421 if (req->flags & IO_REQ_LINK_FLAGS)
1422 io_queue_next(req);
1423 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1424 io_clean_op(req);
1425 }
1426 io_put_file(req);
1427 io_req_put_rsrc_nodes(req);
1428 io_put_task(req);
1429
1430 node = req->comp_list.next;
1431 io_req_add_to_cache(req, ctx);
1432 } while (node);
1433 }
1434
__io_submit_flush_completions(struct io_ring_ctx * ctx)1435 void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1436 __must_hold(&ctx->uring_lock)
1437 {
1438 struct io_submit_state *state = &ctx->submit_state;
1439 struct io_wq_work_node *node;
1440
1441 __io_cq_lock(ctx);
1442 __wq_list_for_each(node, &state->compl_reqs) {
1443 struct io_kiocb *req = container_of(node, struct io_kiocb,
1444 comp_list);
1445
1446 if (!(req->flags & REQ_F_CQE_SKIP) &&
1447 unlikely(!io_fill_cqe_req(ctx, req))) {
1448 if (ctx->lockless_cq) {
1449 spin_lock(&ctx->completion_lock);
1450 io_req_cqe_overflow(req);
1451 spin_unlock(&ctx->completion_lock);
1452 } else {
1453 io_req_cqe_overflow(req);
1454 }
1455 }
1456 }
1457 __io_cq_unlock_post(ctx);
1458
1459 if (!wq_list_empty(&state->compl_reqs)) {
1460 io_free_batch_list(ctx, state->compl_reqs.first);
1461 INIT_WQ_LIST(&state->compl_reqs);
1462 }
1463 ctx->submit_state.cq_flush = false;
1464 }
1465
io_cqring_events(struct io_ring_ctx * ctx)1466 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1467 {
1468 /* See comment at the top of this file */
1469 smp_rmb();
1470 return __io_cqring_events(ctx);
1471 }
1472
1473 /*
1474 * We can't just wait for polled events to come to us, we have to actively
1475 * find and complete them.
1476 */
io_iopoll_try_reap_events(struct io_ring_ctx * ctx)1477 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1478 {
1479 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1480 return;
1481
1482 mutex_lock(&ctx->uring_lock);
1483 while (!wq_list_empty(&ctx->iopoll_list)) {
1484 /* let it sleep and repeat later if can't complete a request */
1485 if (io_do_iopoll(ctx, true) == 0)
1486 break;
1487 /*
1488 * Ensure we allow local-to-the-cpu processing to take place,
1489 * in this case we need to ensure that we reap all events.
1490 * Also let task_work, etc. to progress by releasing the mutex
1491 */
1492 if (need_resched()) {
1493 mutex_unlock(&ctx->uring_lock);
1494 cond_resched();
1495 mutex_lock(&ctx->uring_lock);
1496 }
1497 }
1498 mutex_unlock(&ctx->uring_lock);
1499 }
1500
io_iopoll_check(struct io_ring_ctx * ctx,long min)1501 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1502 {
1503 unsigned int nr_events = 0;
1504 unsigned long check_cq;
1505
1506 lockdep_assert_held(&ctx->uring_lock);
1507
1508 if (!io_allowed_run_tw(ctx))
1509 return -EEXIST;
1510
1511 check_cq = READ_ONCE(ctx->check_cq);
1512 if (unlikely(check_cq)) {
1513 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1514 __io_cqring_overflow_flush(ctx, false);
1515 /*
1516 * Similarly do not spin if we have not informed the user of any
1517 * dropped CQE.
1518 */
1519 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1520 return -EBADR;
1521 }
1522 /*
1523 * Don't enter poll loop if we already have events pending.
1524 * If we do, we can potentially be spinning for commands that
1525 * already triggered a CQE (eg in error).
1526 */
1527 if (io_cqring_events(ctx))
1528 return 0;
1529
1530 do {
1531 int ret = 0;
1532
1533 /*
1534 * If a submit got punted to a workqueue, we can have the
1535 * application entering polling for a command before it gets
1536 * issued. That app will hold the uring_lock for the duration
1537 * of the poll right here, so we need to take a breather every
1538 * now and then to ensure that the issue has a chance to add
1539 * the poll to the issued list. Otherwise we can spin here
1540 * forever, while the workqueue is stuck trying to acquire the
1541 * very same mutex.
1542 */
1543 if (wq_list_empty(&ctx->iopoll_list) ||
1544 io_task_work_pending(ctx)) {
1545 u32 tail = ctx->cached_cq_tail;
1546
1547 (void) io_run_local_work_locked(ctx, min);
1548
1549 if (task_work_pending(current) ||
1550 wq_list_empty(&ctx->iopoll_list)) {
1551 mutex_unlock(&ctx->uring_lock);
1552 io_run_task_work();
1553 mutex_lock(&ctx->uring_lock);
1554 }
1555 /* some requests don't go through iopoll_list */
1556 if (tail != ctx->cached_cq_tail ||
1557 wq_list_empty(&ctx->iopoll_list))
1558 break;
1559 }
1560 ret = io_do_iopoll(ctx, !min);
1561 if (unlikely(ret < 0))
1562 return ret;
1563
1564 if (task_sigpending(current))
1565 return -EINTR;
1566 if (need_resched())
1567 break;
1568
1569 nr_events += ret;
1570 } while (nr_events < min);
1571
1572 return 0;
1573 }
1574
io_req_task_complete(struct io_kiocb * req,struct io_tw_state * ts)1575 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts)
1576 {
1577 io_req_complete_defer(req);
1578 }
1579
1580 /*
1581 * After the iocb has been issued, it's safe to be found on the poll list.
1582 * Adding the kiocb to the list AFTER submission ensures that we don't
1583 * find it from a io_do_iopoll() thread before the issuer is done
1584 * accessing the kiocb cookie.
1585 */
io_iopoll_req_issued(struct io_kiocb * req,unsigned int issue_flags)1586 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1587 {
1588 struct io_ring_ctx *ctx = req->ctx;
1589 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1590
1591 /* workqueue context doesn't hold uring_lock, grab it now */
1592 if (unlikely(needs_lock))
1593 mutex_lock(&ctx->uring_lock);
1594
1595 /*
1596 * Track whether we have multiple files in our lists. This will impact
1597 * how we do polling eventually, not spinning if we're on potentially
1598 * different devices.
1599 */
1600 if (wq_list_empty(&ctx->iopoll_list)) {
1601 ctx->poll_multi_queue = false;
1602 } else if (!ctx->poll_multi_queue) {
1603 struct io_kiocb *list_req;
1604
1605 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1606 comp_list);
1607 if (list_req->file != req->file)
1608 ctx->poll_multi_queue = true;
1609 }
1610
1611 /*
1612 * For fast devices, IO may have already completed. If it has, add
1613 * it to the front so we find it first.
1614 */
1615 if (READ_ONCE(req->iopoll_completed))
1616 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1617 else
1618 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1619
1620 if (unlikely(needs_lock)) {
1621 /*
1622 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1623 * in sq thread task context or in io worker task context. If
1624 * current task context is sq thread, we don't need to check
1625 * whether should wake up sq thread.
1626 */
1627 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1628 wq_has_sleeper(&ctx->sq_data->wait))
1629 wake_up(&ctx->sq_data->wait);
1630
1631 mutex_unlock(&ctx->uring_lock);
1632 }
1633 }
1634
io_file_get_flags(struct file * file)1635 io_req_flags_t io_file_get_flags(struct file *file)
1636 {
1637 io_req_flags_t res = 0;
1638
1639 if (S_ISREG(file_inode(file)->i_mode))
1640 res |= REQ_F_ISREG;
1641 if ((file->f_flags & O_NONBLOCK) || (file->f_mode & FMODE_NOWAIT))
1642 res |= REQ_F_SUPPORT_NOWAIT;
1643 return res;
1644 }
1645
io_alloc_async_data(struct io_kiocb * req)1646 bool io_alloc_async_data(struct io_kiocb *req)
1647 {
1648 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1649
1650 WARN_ON_ONCE(!def->async_size);
1651 req->async_data = kmalloc(def->async_size, GFP_KERNEL);
1652 if (req->async_data) {
1653 req->flags |= REQ_F_ASYNC_DATA;
1654 return false;
1655 }
1656 return true;
1657 }
1658
io_get_sequence(struct io_kiocb * req)1659 static u32 io_get_sequence(struct io_kiocb *req)
1660 {
1661 u32 seq = req->ctx->cached_sq_head;
1662 struct io_kiocb *cur;
1663
1664 /* need original cached_sq_head, but it was increased for each req */
1665 io_for_each_link(cur, req)
1666 seq--;
1667 return seq;
1668 }
1669
io_drain_req(struct io_kiocb * req)1670 static __cold void io_drain_req(struct io_kiocb *req)
1671 __must_hold(&ctx->uring_lock)
1672 {
1673 struct io_ring_ctx *ctx = req->ctx;
1674 struct io_defer_entry *de;
1675 int ret;
1676 u32 seq = io_get_sequence(req);
1677
1678 /* Still need defer if there is pending req in defer list. */
1679 spin_lock(&ctx->completion_lock);
1680 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1681 spin_unlock(&ctx->completion_lock);
1682 queue:
1683 ctx->drain_active = false;
1684 io_req_task_queue(req);
1685 return;
1686 }
1687 spin_unlock(&ctx->completion_lock);
1688
1689 io_prep_async_link(req);
1690 de = kmalloc(sizeof(*de), GFP_KERNEL);
1691 if (!de) {
1692 ret = -ENOMEM;
1693 io_req_defer_failed(req, ret);
1694 return;
1695 }
1696
1697 spin_lock(&ctx->completion_lock);
1698 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1699 spin_unlock(&ctx->completion_lock);
1700 kfree(de);
1701 goto queue;
1702 }
1703
1704 trace_io_uring_defer(req);
1705 de->req = req;
1706 de->seq = seq;
1707 list_add_tail(&de->list, &ctx->defer_list);
1708 spin_unlock(&ctx->completion_lock);
1709 }
1710
io_assign_file(struct io_kiocb * req,const struct io_issue_def * def,unsigned int issue_flags)1711 static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def,
1712 unsigned int issue_flags)
1713 {
1714 if (req->file || !def->needs_file)
1715 return true;
1716
1717 if (req->flags & REQ_F_FIXED_FILE)
1718 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1719 else
1720 req->file = io_file_get_normal(req, req->cqe.fd);
1721
1722 return !!req->file;
1723 }
1724
io_issue_sqe(struct io_kiocb * req,unsigned int issue_flags)1725 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1726 {
1727 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1728 const struct cred *creds = NULL;
1729 int ret;
1730
1731 if (unlikely(!io_assign_file(req, def, issue_flags)))
1732 return -EBADF;
1733
1734 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1735 creds = override_creds(req->creds);
1736
1737 if (!def->audit_skip)
1738 audit_uring_entry(req->opcode);
1739
1740 ret = def->issue(req, issue_flags);
1741
1742 if (!def->audit_skip)
1743 audit_uring_exit(!ret, ret);
1744
1745 if (creds)
1746 revert_creds(creds);
1747
1748 if (ret == IOU_OK) {
1749 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1750 io_req_complete_defer(req);
1751 else
1752 io_req_complete_post(req, issue_flags);
1753
1754 return 0;
1755 }
1756
1757 if (ret == IOU_ISSUE_SKIP_COMPLETE) {
1758 ret = 0;
1759 io_arm_ltimeout(req);
1760
1761 /* If the op doesn't have a file, we're not polling for it */
1762 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
1763 io_iopoll_req_issued(req, issue_flags);
1764 }
1765 return ret;
1766 }
1767
io_poll_issue(struct io_kiocb * req,struct io_tw_state * ts)1768 int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts)
1769 {
1770 io_tw_lock(req->ctx, ts);
1771 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT|
1772 IO_URING_F_COMPLETE_DEFER);
1773 }
1774
io_wq_free_work(struct io_wq_work * work)1775 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1776 {
1777 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1778 struct io_kiocb *nxt = NULL;
1779
1780 if (req_ref_put_and_test(req)) {
1781 if (req->flags & IO_REQ_LINK_FLAGS)
1782 nxt = io_req_find_next(req);
1783 io_free_req(req);
1784 }
1785 return nxt ? &nxt->work : NULL;
1786 }
1787
io_wq_submit_work(struct io_wq_work * work)1788 void io_wq_submit_work(struct io_wq_work *work)
1789 {
1790 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1791 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1792 unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ;
1793 bool needs_poll = false;
1794 int ret = 0, err = -ECANCELED;
1795
1796 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1797 if (!(req->flags & REQ_F_REFCOUNT))
1798 __io_req_set_refcount(req, 2);
1799 else
1800 req_ref_get(req);
1801
1802 io_arm_ltimeout(req);
1803
1804 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1805 if (atomic_read(&work->flags) & IO_WQ_WORK_CANCEL) {
1806 fail:
1807 io_req_task_queue_fail(req, err);
1808 return;
1809 }
1810 if (!io_assign_file(req, def, issue_flags)) {
1811 err = -EBADF;
1812 atomic_or(IO_WQ_WORK_CANCEL, &work->flags);
1813 goto fail;
1814 }
1815
1816 /*
1817 * If DEFER_TASKRUN is set, it's only allowed to post CQEs from the
1818 * submitter task context. Final request completions are handed to the
1819 * right context, however this is not the case of auxiliary CQEs,
1820 * which is the main mean of operation for multishot requests.
1821 * Don't allow any multishot execution from io-wq. It's more restrictive
1822 * than necessary and also cleaner.
1823 */
1824 if (req->flags & REQ_F_APOLL_MULTISHOT) {
1825 err = -EBADFD;
1826 if (!io_file_can_poll(req))
1827 goto fail;
1828 if (req->file->f_flags & O_NONBLOCK ||
1829 req->file->f_mode & FMODE_NOWAIT) {
1830 err = -ECANCELED;
1831 if (io_arm_poll_handler(req, issue_flags) != IO_APOLL_OK)
1832 goto fail;
1833 return;
1834 } else {
1835 req->flags &= ~REQ_F_APOLL_MULTISHOT;
1836 }
1837 }
1838
1839 if (req->flags & REQ_F_FORCE_ASYNC) {
1840 bool opcode_poll = def->pollin || def->pollout;
1841
1842 if (opcode_poll && io_file_can_poll(req)) {
1843 needs_poll = true;
1844 issue_flags |= IO_URING_F_NONBLOCK;
1845 }
1846 }
1847
1848 do {
1849 ret = io_issue_sqe(req, issue_flags);
1850 if (ret != -EAGAIN)
1851 break;
1852
1853 /*
1854 * If REQ_F_NOWAIT is set, then don't wait or retry with
1855 * poll. -EAGAIN is final for that case.
1856 */
1857 if (req->flags & REQ_F_NOWAIT)
1858 break;
1859
1860 /*
1861 * We can get EAGAIN for iopolled IO even though we're
1862 * forcing a sync submission from here, since we can't
1863 * wait for request slots on the block side.
1864 */
1865 if (!needs_poll) {
1866 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1867 break;
1868 if (io_wq_worker_stopped())
1869 break;
1870 cond_resched();
1871 continue;
1872 }
1873
1874 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1875 return;
1876 /* aborted or ready, in either case retry blocking */
1877 needs_poll = false;
1878 issue_flags &= ~IO_URING_F_NONBLOCK;
1879 } while (1);
1880
1881 /* avoid locking problems by failing it from a clean context */
1882 if (ret)
1883 io_req_task_queue_fail(req, ret);
1884 }
1885
io_file_get_fixed(struct io_kiocb * req,int fd,unsigned int issue_flags)1886 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1887 unsigned int issue_flags)
1888 {
1889 struct io_ring_ctx *ctx = req->ctx;
1890 struct io_rsrc_node *node;
1891 struct file *file = NULL;
1892
1893 io_ring_submit_lock(ctx, issue_flags);
1894 node = io_rsrc_node_lookup(&ctx->file_table.data, fd);
1895 if (node) {
1896 io_req_assign_rsrc_node(&req->file_node, node);
1897 req->flags |= io_slot_flags(node);
1898 file = io_slot_file(node);
1899 }
1900 io_ring_submit_unlock(ctx, issue_flags);
1901 return file;
1902 }
1903
io_file_get_normal(struct io_kiocb * req,int fd)1904 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
1905 {
1906 struct file *file = fget(fd);
1907
1908 trace_io_uring_file_get(req, fd);
1909
1910 /* we don't allow fixed io_uring files */
1911 if (file && io_is_uring_fops(file))
1912 io_req_track_inflight(req);
1913 return file;
1914 }
1915
io_queue_async(struct io_kiocb * req,int ret)1916 static void io_queue_async(struct io_kiocb *req, int ret)
1917 __must_hold(&req->ctx->uring_lock)
1918 {
1919 struct io_kiocb *linked_timeout;
1920
1921 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
1922 io_req_defer_failed(req, ret);
1923 return;
1924 }
1925
1926 linked_timeout = io_prep_linked_timeout(req);
1927
1928 switch (io_arm_poll_handler(req, 0)) {
1929 case IO_APOLL_READY:
1930 io_kbuf_recycle(req, 0);
1931 io_req_task_queue(req);
1932 break;
1933 case IO_APOLL_ABORTED:
1934 io_kbuf_recycle(req, 0);
1935 io_queue_iowq(req);
1936 break;
1937 case IO_APOLL_OK:
1938 break;
1939 }
1940
1941 if (linked_timeout)
1942 io_queue_linked_timeout(linked_timeout);
1943 }
1944
io_queue_sqe(struct io_kiocb * req)1945 static inline void io_queue_sqe(struct io_kiocb *req)
1946 __must_hold(&req->ctx->uring_lock)
1947 {
1948 int ret;
1949
1950 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
1951
1952 /*
1953 * We async punt it if the file wasn't marked NOWAIT, or if the file
1954 * doesn't support non-blocking read/write attempts
1955 */
1956 if (unlikely(ret))
1957 io_queue_async(req, ret);
1958 }
1959
io_queue_sqe_fallback(struct io_kiocb * req)1960 static void io_queue_sqe_fallback(struct io_kiocb *req)
1961 __must_hold(&req->ctx->uring_lock)
1962 {
1963 if (unlikely(req->flags & REQ_F_FAIL)) {
1964 /*
1965 * We don't submit, fail them all, for that replace hardlinks
1966 * with normal links. Extra REQ_F_LINK is tolerated.
1967 */
1968 req->flags &= ~REQ_F_HARDLINK;
1969 req->flags |= REQ_F_LINK;
1970 io_req_defer_failed(req, req->cqe.res);
1971 } else {
1972 if (unlikely(req->ctx->drain_active))
1973 io_drain_req(req);
1974 else
1975 io_queue_iowq(req);
1976 }
1977 }
1978
1979 /*
1980 * Check SQE restrictions (opcode and flags).
1981 *
1982 * Returns 'true' if SQE is allowed, 'false' otherwise.
1983 */
io_check_restriction(struct io_ring_ctx * ctx,struct io_kiocb * req,unsigned int sqe_flags)1984 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
1985 struct io_kiocb *req,
1986 unsigned int sqe_flags)
1987 {
1988 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
1989 return false;
1990
1991 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
1992 ctx->restrictions.sqe_flags_required)
1993 return false;
1994
1995 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
1996 ctx->restrictions.sqe_flags_required))
1997 return false;
1998
1999 return true;
2000 }
2001
io_init_req_drain(struct io_kiocb * req)2002 static void io_init_req_drain(struct io_kiocb *req)
2003 {
2004 struct io_ring_ctx *ctx = req->ctx;
2005 struct io_kiocb *head = ctx->submit_state.link.head;
2006
2007 ctx->drain_active = true;
2008 if (head) {
2009 /*
2010 * If we need to drain a request in the middle of a link, drain
2011 * the head request and the next request/link after the current
2012 * link. Considering sequential execution of links,
2013 * REQ_F_IO_DRAIN will be maintained for every request of our
2014 * link.
2015 */
2016 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2017 ctx->drain_next = true;
2018 }
2019 }
2020
io_init_fail_req(struct io_kiocb * req,int err)2021 static __cold int io_init_fail_req(struct io_kiocb *req, int err)
2022 {
2023 /* ensure per-opcode data is cleared if we fail before prep */
2024 memset(&req->cmd.data, 0, sizeof(req->cmd.data));
2025 return err;
2026 }
2027
io_init_req(struct io_ring_ctx * ctx,struct io_kiocb * req,const struct io_uring_sqe * sqe)2028 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
2029 const struct io_uring_sqe *sqe)
2030 __must_hold(&ctx->uring_lock)
2031 {
2032 const struct io_issue_def *def;
2033 unsigned int sqe_flags;
2034 int personality;
2035 u8 opcode;
2036
2037 /* req is partially pre-initialised, see io_preinit_req() */
2038 req->opcode = opcode = READ_ONCE(sqe->opcode);
2039 /* same numerical values with corresponding REQ_F_*, safe to copy */
2040 sqe_flags = READ_ONCE(sqe->flags);
2041 req->flags = (__force io_req_flags_t) sqe_flags;
2042 req->cqe.user_data = READ_ONCE(sqe->user_data);
2043 req->file = NULL;
2044 req->tctx = current->io_uring;
2045 req->cancel_seq_set = false;
2046
2047 if (unlikely(opcode >= IORING_OP_LAST)) {
2048 req->opcode = 0;
2049 return io_init_fail_req(req, -EINVAL);
2050 }
2051 def = &io_issue_defs[opcode];
2052 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2053 /* enforce forwards compatibility on users */
2054 if (sqe_flags & ~SQE_VALID_FLAGS)
2055 return io_init_fail_req(req, -EINVAL);
2056 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2057 if (!def->buffer_select)
2058 return io_init_fail_req(req, -EOPNOTSUPP);
2059 req->buf_index = READ_ONCE(sqe->buf_group);
2060 }
2061 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2062 ctx->drain_disabled = true;
2063 if (sqe_flags & IOSQE_IO_DRAIN) {
2064 if (ctx->drain_disabled)
2065 return io_init_fail_req(req, -EOPNOTSUPP);
2066 io_init_req_drain(req);
2067 }
2068 }
2069 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2070 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2071 return io_init_fail_req(req, -EACCES);
2072 /* knock it to the slow queue path, will be drained there */
2073 if (ctx->drain_active)
2074 req->flags |= REQ_F_FORCE_ASYNC;
2075 /* if there is no link, we're at "next" request and need to drain */
2076 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2077 ctx->drain_next = false;
2078 ctx->drain_active = true;
2079 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2080 }
2081 }
2082
2083 if (!def->ioprio && sqe->ioprio)
2084 return io_init_fail_req(req, -EINVAL);
2085 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2086 return io_init_fail_req(req, -EINVAL);
2087
2088 if (def->needs_file) {
2089 struct io_submit_state *state = &ctx->submit_state;
2090
2091 req->cqe.fd = READ_ONCE(sqe->fd);
2092
2093 /*
2094 * Plug now if we have more than 2 IO left after this, and the
2095 * target is potentially a read/write to block based storage.
2096 */
2097 if (state->need_plug && def->plug) {
2098 state->plug_started = true;
2099 state->need_plug = false;
2100 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2101 }
2102 }
2103
2104 personality = READ_ONCE(sqe->personality);
2105 if (personality) {
2106 int ret;
2107
2108 req->creds = xa_load(&ctx->personalities, personality);
2109 if (!req->creds)
2110 return io_init_fail_req(req, -EINVAL);
2111 get_cred(req->creds);
2112 ret = security_uring_override_creds(req->creds);
2113 if (ret) {
2114 put_cred(req->creds);
2115 return io_init_fail_req(req, ret);
2116 }
2117 req->flags |= REQ_F_CREDS;
2118 }
2119
2120 return def->prep(req, sqe);
2121 }
2122
io_submit_fail_init(const struct io_uring_sqe * sqe,struct io_kiocb * req,int ret)2123 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2124 struct io_kiocb *req, int ret)
2125 {
2126 struct io_ring_ctx *ctx = req->ctx;
2127 struct io_submit_link *link = &ctx->submit_state.link;
2128 struct io_kiocb *head = link->head;
2129
2130 trace_io_uring_req_failed(sqe, req, ret);
2131
2132 /*
2133 * Avoid breaking links in the middle as it renders links with SQPOLL
2134 * unusable. Instead of failing eagerly, continue assembling the link if
2135 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2136 * should find the flag and handle the rest.
2137 */
2138 req_fail_link_node(req, ret);
2139 if (head && !(head->flags & REQ_F_FAIL))
2140 req_fail_link_node(head, -ECANCELED);
2141
2142 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2143 if (head) {
2144 link->last->link = req;
2145 link->head = NULL;
2146 req = head;
2147 }
2148 io_queue_sqe_fallback(req);
2149 return ret;
2150 }
2151
2152 if (head)
2153 link->last->link = req;
2154 else
2155 link->head = req;
2156 link->last = req;
2157 return 0;
2158 }
2159
io_submit_sqe(struct io_ring_ctx * ctx,struct io_kiocb * req,const struct io_uring_sqe * sqe)2160 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2161 const struct io_uring_sqe *sqe)
2162 __must_hold(&ctx->uring_lock)
2163 {
2164 struct io_submit_link *link = &ctx->submit_state.link;
2165 int ret;
2166
2167 ret = io_init_req(ctx, req, sqe);
2168 if (unlikely(ret))
2169 return io_submit_fail_init(sqe, req, ret);
2170
2171 trace_io_uring_submit_req(req);
2172
2173 /*
2174 * If we already have a head request, queue this one for async
2175 * submittal once the head completes. If we don't have a head but
2176 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2177 * submitted sync once the chain is complete. If none of those
2178 * conditions are true (normal request), then just queue it.
2179 */
2180 if (unlikely(link->head)) {
2181 trace_io_uring_link(req, link->last);
2182 link->last->link = req;
2183 link->last = req;
2184
2185 if (req->flags & IO_REQ_LINK_FLAGS)
2186 return 0;
2187 /* last request of the link, flush it */
2188 req = link->head;
2189 link->head = NULL;
2190 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2191 goto fallback;
2192
2193 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2194 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2195 if (req->flags & IO_REQ_LINK_FLAGS) {
2196 link->head = req;
2197 link->last = req;
2198 } else {
2199 fallback:
2200 io_queue_sqe_fallback(req);
2201 }
2202 return 0;
2203 }
2204
2205 io_queue_sqe(req);
2206 return 0;
2207 }
2208
2209 /*
2210 * Batched submission is done, ensure local IO is flushed out.
2211 */
io_submit_state_end(struct io_ring_ctx * ctx)2212 static void io_submit_state_end(struct io_ring_ctx *ctx)
2213 {
2214 struct io_submit_state *state = &ctx->submit_state;
2215
2216 if (unlikely(state->link.head))
2217 io_queue_sqe_fallback(state->link.head);
2218 /* flush only after queuing links as they can generate completions */
2219 io_submit_flush_completions(ctx);
2220 if (state->plug_started)
2221 blk_finish_plug(&state->plug);
2222 }
2223
2224 /*
2225 * Start submission side cache.
2226 */
io_submit_state_start(struct io_submit_state * state,unsigned int max_ios)2227 static void io_submit_state_start(struct io_submit_state *state,
2228 unsigned int max_ios)
2229 {
2230 state->plug_started = false;
2231 state->need_plug = max_ios > 2;
2232 state->submit_nr = max_ios;
2233 /* set only head, no need to init link_last in advance */
2234 state->link.head = NULL;
2235 }
2236
io_commit_sqring(struct io_ring_ctx * ctx)2237 static void io_commit_sqring(struct io_ring_ctx *ctx)
2238 {
2239 struct io_rings *rings = ctx->rings;
2240
2241 /*
2242 * Ensure any loads from the SQEs are done at this point,
2243 * since once we write the new head, the application could
2244 * write new data to them.
2245 */
2246 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2247 }
2248
2249 /*
2250 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2251 * that is mapped by userspace. This means that care needs to be taken to
2252 * ensure that reads are stable, as we cannot rely on userspace always
2253 * being a good citizen. If members of the sqe are validated and then later
2254 * used, it's important that those reads are done through READ_ONCE() to
2255 * prevent a re-load down the line.
2256 */
io_get_sqe(struct io_ring_ctx * ctx,const struct io_uring_sqe ** sqe)2257 static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe)
2258 {
2259 unsigned mask = ctx->sq_entries - 1;
2260 unsigned head = ctx->cached_sq_head++ & mask;
2261
2262 if (static_branch_unlikely(&io_key_has_sqarray) &&
2263 (!(ctx->flags & IORING_SETUP_NO_SQARRAY))) {
2264 head = READ_ONCE(ctx->sq_array[head]);
2265 if (unlikely(head >= ctx->sq_entries)) {
2266 /* drop invalid entries */
2267 spin_lock(&ctx->completion_lock);
2268 ctx->cq_extra--;
2269 spin_unlock(&ctx->completion_lock);
2270 WRITE_ONCE(ctx->rings->sq_dropped,
2271 READ_ONCE(ctx->rings->sq_dropped) + 1);
2272 return false;
2273 }
2274 head = array_index_nospec(head, ctx->sq_entries);
2275 }
2276
2277 /*
2278 * The cached sq head (or cq tail) serves two purposes:
2279 *
2280 * 1) allows us to batch the cost of updating the user visible
2281 * head updates.
2282 * 2) allows the kernel side to track the head on its own, even
2283 * though the application is the one updating it.
2284 */
2285
2286 /* double index for 128-byte SQEs, twice as long */
2287 if (ctx->flags & IORING_SETUP_SQE128)
2288 head <<= 1;
2289 *sqe = &ctx->sq_sqes[head];
2290 return true;
2291 }
2292
io_submit_sqes(struct io_ring_ctx * ctx,unsigned int nr)2293 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2294 __must_hold(&ctx->uring_lock)
2295 {
2296 unsigned int entries = io_sqring_entries(ctx);
2297 unsigned int left;
2298 int ret;
2299
2300 if (unlikely(!entries))
2301 return 0;
2302 /* make sure SQ entry isn't read before tail */
2303 ret = left = min(nr, entries);
2304 io_get_task_refs(left);
2305 io_submit_state_start(&ctx->submit_state, left);
2306
2307 do {
2308 const struct io_uring_sqe *sqe;
2309 struct io_kiocb *req;
2310
2311 if (unlikely(!io_alloc_req(ctx, &req)))
2312 break;
2313 if (unlikely(!io_get_sqe(ctx, &sqe))) {
2314 io_req_add_to_cache(req, ctx);
2315 break;
2316 }
2317
2318 /*
2319 * Continue submitting even for sqe failure if the
2320 * ring was setup with IORING_SETUP_SUBMIT_ALL
2321 */
2322 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2323 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2324 left--;
2325 break;
2326 }
2327 } while (--left);
2328
2329 if (unlikely(left)) {
2330 ret -= left;
2331 /* try again if it submitted nothing and can't allocate a req */
2332 if (!ret && io_req_cache_empty(ctx))
2333 ret = -EAGAIN;
2334 current->io_uring->cached_refs += left;
2335 }
2336
2337 io_submit_state_end(ctx);
2338 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2339 io_commit_sqring(ctx);
2340 return ret;
2341 }
2342
io_wake_function(struct wait_queue_entry * curr,unsigned int mode,int wake_flags,void * key)2343 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2344 int wake_flags, void *key)
2345 {
2346 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq);
2347
2348 /*
2349 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2350 * the task, and the next invocation will do it.
2351 */
2352 if (io_should_wake(iowq) || io_has_work(iowq->ctx))
2353 return autoremove_wake_function(curr, mode, wake_flags, key);
2354 return -1;
2355 }
2356
io_run_task_work_sig(struct io_ring_ctx * ctx)2357 int io_run_task_work_sig(struct io_ring_ctx *ctx)
2358 {
2359 if (io_local_work_pending(ctx)) {
2360 __set_current_state(TASK_RUNNING);
2361 if (io_run_local_work(ctx, INT_MAX, IO_LOCAL_TW_DEFAULT_MAX) > 0)
2362 return 0;
2363 }
2364 if (io_run_task_work() > 0)
2365 return 0;
2366 if (task_sigpending(current))
2367 return -EINTR;
2368 return 0;
2369 }
2370
current_pending_io(void)2371 static bool current_pending_io(void)
2372 {
2373 struct io_uring_task *tctx = current->io_uring;
2374
2375 if (!tctx)
2376 return false;
2377 return percpu_counter_read_positive(&tctx->inflight);
2378 }
2379
io_cqring_timer_wakeup(struct hrtimer * timer)2380 static enum hrtimer_restart io_cqring_timer_wakeup(struct hrtimer *timer)
2381 {
2382 struct io_wait_queue *iowq = container_of(timer, struct io_wait_queue, t);
2383
2384 WRITE_ONCE(iowq->hit_timeout, 1);
2385 iowq->min_timeout = 0;
2386 wake_up_process(iowq->wq.private);
2387 return HRTIMER_NORESTART;
2388 }
2389
2390 /*
2391 * Doing min_timeout portion. If we saw any timeouts, events, or have work,
2392 * wake up. If not, and we have a normal timeout, switch to that and keep
2393 * sleeping.
2394 */
io_cqring_min_timer_wakeup(struct hrtimer * timer)2395 static enum hrtimer_restart io_cqring_min_timer_wakeup(struct hrtimer *timer)
2396 {
2397 struct io_wait_queue *iowq = container_of(timer, struct io_wait_queue, t);
2398 struct io_ring_ctx *ctx = iowq->ctx;
2399
2400 /* no general timeout, or shorter (or equal), we are done */
2401 if (iowq->timeout == KTIME_MAX ||
2402 ktime_compare(iowq->min_timeout, iowq->timeout) >= 0)
2403 goto out_wake;
2404 /* work we may need to run, wake function will see if we need to wake */
2405 if (io_has_work(ctx))
2406 goto out_wake;
2407 /* got events since we started waiting, min timeout is done */
2408 if (iowq->cq_min_tail != READ_ONCE(ctx->rings->cq.tail))
2409 goto out_wake;
2410 /* if we have any events and min timeout expired, we're done */
2411 if (io_cqring_events(ctx))
2412 goto out_wake;
2413
2414 /*
2415 * If using deferred task_work running and application is waiting on
2416 * more than one request, ensure we reset it now where we are switching
2417 * to normal sleeps. Any request completion post min_wait should wake
2418 * the task and return.
2419 */
2420 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
2421 atomic_set(&ctx->cq_wait_nr, 1);
2422 smp_mb();
2423 if (!llist_empty(&ctx->work_llist))
2424 goto out_wake;
2425 }
2426
2427 iowq->t.function = io_cqring_timer_wakeup;
2428 hrtimer_set_expires(timer, iowq->timeout);
2429 return HRTIMER_RESTART;
2430 out_wake:
2431 return io_cqring_timer_wakeup(timer);
2432 }
2433
io_cqring_schedule_timeout(struct io_wait_queue * iowq,clockid_t clock_id,ktime_t start_time)2434 static int io_cqring_schedule_timeout(struct io_wait_queue *iowq,
2435 clockid_t clock_id, ktime_t start_time)
2436 {
2437 ktime_t timeout;
2438
2439 if (iowq->min_timeout) {
2440 timeout = ktime_add_ns(iowq->min_timeout, start_time);
2441 hrtimer_setup_on_stack(&iowq->t, io_cqring_min_timer_wakeup, clock_id,
2442 HRTIMER_MODE_ABS);
2443 } else {
2444 timeout = iowq->timeout;
2445 hrtimer_setup_on_stack(&iowq->t, io_cqring_timer_wakeup, clock_id,
2446 HRTIMER_MODE_ABS);
2447 }
2448
2449 hrtimer_set_expires_range_ns(&iowq->t, timeout, 0);
2450 hrtimer_start_expires(&iowq->t, HRTIMER_MODE_ABS);
2451
2452 if (!READ_ONCE(iowq->hit_timeout))
2453 schedule();
2454
2455 hrtimer_cancel(&iowq->t);
2456 destroy_hrtimer_on_stack(&iowq->t);
2457 __set_current_state(TASK_RUNNING);
2458
2459 return READ_ONCE(iowq->hit_timeout) ? -ETIME : 0;
2460 }
2461
__io_cqring_wait_schedule(struct io_ring_ctx * ctx,struct io_wait_queue * iowq,ktime_t start_time)2462 static int __io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2463 struct io_wait_queue *iowq,
2464 ktime_t start_time)
2465 {
2466 int ret = 0;
2467
2468 /*
2469 * Mark us as being in io_wait if we have pending requests, so cpufreq
2470 * can take into account that the task is waiting for IO - turns out
2471 * to be important for low QD IO.
2472 */
2473 if (current_pending_io())
2474 current->in_iowait = 1;
2475 if (iowq->timeout != KTIME_MAX || iowq->min_timeout)
2476 ret = io_cqring_schedule_timeout(iowq, ctx->clockid, start_time);
2477 else
2478 schedule();
2479 current->in_iowait = 0;
2480 return ret;
2481 }
2482
2483 /* If this returns > 0, the caller should retry */
io_cqring_wait_schedule(struct io_ring_ctx * ctx,struct io_wait_queue * iowq,ktime_t start_time)2484 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2485 struct io_wait_queue *iowq,
2486 ktime_t start_time)
2487 {
2488 if (unlikely(READ_ONCE(ctx->check_cq)))
2489 return 1;
2490 if (unlikely(io_local_work_pending(ctx)))
2491 return 1;
2492 if (unlikely(task_work_pending(current)))
2493 return 1;
2494 if (unlikely(task_sigpending(current)))
2495 return -EINTR;
2496 if (unlikely(io_should_wake(iowq)))
2497 return 0;
2498
2499 return __io_cqring_wait_schedule(ctx, iowq, start_time);
2500 }
2501
2502 struct ext_arg {
2503 size_t argsz;
2504 struct timespec64 ts;
2505 const sigset_t __user *sig;
2506 ktime_t min_time;
2507 bool ts_set;
2508 };
2509
2510 /*
2511 * Wait until events become available, if we don't already have some. The
2512 * application must reap them itself, as they reside on the shared cq ring.
2513 */
io_cqring_wait(struct io_ring_ctx * ctx,int min_events,u32 flags,struct ext_arg * ext_arg)2514 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, u32 flags,
2515 struct ext_arg *ext_arg)
2516 {
2517 struct io_wait_queue iowq;
2518 struct io_rings *rings = ctx->rings;
2519 ktime_t start_time;
2520 int ret;
2521
2522 if (!io_allowed_run_tw(ctx))
2523 return -EEXIST;
2524 if (io_local_work_pending(ctx))
2525 io_run_local_work(ctx, min_events,
2526 max(IO_LOCAL_TW_DEFAULT_MAX, min_events));
2527 io_run_task_work();
2528
2529 if (unlikely(test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)))
2530 io_cqring_do_overflow_flush(ctx);
2531 if (__io_cqring_events_user(ctx) >= min_events)
2532 return 0;
2533
2534 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2535 iowq.wq.private = current;
2536 INIT_LIST_HEAD(&iowq.wq.entry);
2537 iowq.ctx = ctx;
2538 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2539 iowq.cq_min_tail = READ_ONCE(ctx->rings->cq.tail);
2540 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2541 iowq.hit_timeout = 0;
2542 iowq.min_timeout = ext_arg->min_time;
2543 iowq.timeout = KTIME_MAX;
2544 start_time = io_get_time(ctx);
2545
2546 if (ext_arg->ts_set) {
2547 iowq.timeout = timespec64_to_ktime(ext_arg->ts);
2548 if (!(flags & IORING_ENTER_ABS_TIMER))
2549 iowq.timeout = ktime_add(iowq.timeout, start_time);
2550 }
2551
2552 if (ext_arg->sig) {
2553 #ifdef CONFIG_COMPAT
2554 if (in_compat_syscall())
2555 ret = set_compat_user_sigmask((const compat_sigset_t __user *)ext_arg->sig,
2556 ext_arg->argsz);
2557 else
2558 #endif
2559 ret = set_user_sigmask(ext_arg->sig, ext_arg->argsz);
2560
2561 if (ret)
2562 return ret;
2563 }
2564
2565 io_napi_busy_loop(ctx, &iowq);
2566
2567 trace_io_uring_cqring_wait(ctx, min_events);
2568 do {
2569 unsigned long check_cq;
2570 int nr_wait;
2571
2572 /* if min timeout has been hit, don't reset wait count */
2573 if (!iowq.hit_timeout)
2574 nr_wait = (int) iowq.cq_tail -
2575 READ_ONCE(ctx->rings->cq.tail);
2576 else
2577 nr_wait = 1;
2578
2579 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
2580 atomic_set(&ctx->cq_wait_nr, nr_wait);
2581 set_current_state(TASK_INTERRUPTIBLE);
2582 } else {
2583 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2584 TASK_INTERRUPTIBLE);
2585 }
2586
2587 ret = io_cqring_wait_schedule(ctx, &iowq, start_time);
2588 __set_current_state(TASK_RUNNING);
2589 atomic_set(&ctx->cq_wait_nr, IO_CQ_WAKE_INIT);
2590
2591 /*
2592 * Run task_work after scheduling and before io_should_wake().
2593 * If we got woken because of task_work being processed, run it
2594 * now rather than let the caller do another wait loop.
2595 */
2596 if (io_local_work_pending(ctx))
2597 io_run_local_work(ctx, nr_wait, nr_wait);
2598 io_run_task_work();
2599
2600 /*
2601 * Non-local task_work will be run on exit to userspace, but
2602 * if we're using DEFER_TASKRUN, then we could have waited
2603 * with a timeout for a number of requests. If the timeout
2604 * hits, we could have some requests ready to process. Ensure
2605 * this break is _after_ we have run task_work, to avoid
2606 * deferring running potentially pending requests until the
2607 * next time we wait for events.
2608 */
2609 if (ret < 0)
2610 break;
2611
2612 check_cq = READ_ONCE(ctx->check_cq);
2613 if (unlikely(check_cq)) {
2614 /* let the caller flush overflows, retry */
2615 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2616 io_cqring_do_overflow_flush(ctx);
2617 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) {
2618 ret = -EBADR;
2619 break;
2620 }
2621 }
2622
2623 if (io_should_wake(&iowq)) {
2624 ret = 0;
2625 break;
2626 }
2627 cond_resched();
2628 } while (1);
2629
2630 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
2631 finish_wait(&ctx->cq_wait, &iowq.wq);
2632 restore_saved_sigmask_unless(ret == -EINTR);
2633
2634 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2635 }
2636
io_rings_map(struct io_ring_ctx * ctx,unsigned long uaddr,size_t size)2637 static void *io_rings_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2638 size_t size)
2639 {
2640 return __io_uaddr_map(&ctx->ring_pages, &ctx->n_ring_pages, uaddr,
2641 size);
2642 }
2643
io_sqes_map(struct io_ring_ctx * ctx,unsigned long uaddr,size_t size)2644 static void *io_sqes_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2645 size_t size)
2646 {
2647 return __io_uaddr_map(&ctx->sqe_pages, &ctx->n_sqe_pages, uaddr,
2648 size);
2649 }
2650
io_rings_free(struct io_ring_ctx * ctx)2651 static void io_rings_free(struct io_ring_ctx *ctx)
2652 {
2653 if (!(ctx->flags & IORING_SETUP_NO_MMAP)) {
2654 io_pages_unmap(ctx->rings, &ctx->ring_pages, &ctx->n_ring_pages,
2655 true);
2656 io_pages_unmap(ctx->sq_sqes, &ctx->sqe_pages, &ctx->n_sqe_pages,
2657 true);
2658 } else {
2659 io_pages_free(&ctx->ring_pages, ctx->n_ring_pages);
2660 ctx->n_ring_pages = 0;
2661 io_pages_free(&ctx->sqe_pages, ctx->n_sqe_pages);
2662 ctx->n_sqe_pages = 0;
2663 vunmap(ctx->rings);
2664 vunmap(ctx->sq_sqes);
2665 }
2666
2667 ctx->rings = NULL;
2668 ctx->sq_sqes = NULL;
2669 }
2670
rings_size(unsigned int flags,unsigned int sq_entries,unsigned int cq_entries,size_t * sq_offset)2671 unsigned long rings_size(unsigned int flags, unsigned int sq_entries,
2672 unsigned int cq_entries, size_t *sq_offset)
2673 {
2674 struct io_rings *rings;
2675 size_t off, sq_array_size;
2676
2677 off = struct_size(rings, cqes, cq_entries);
2678 if (off == SIZE_MAX)
2679 return SIZE_MAX;
2680 if (flags & IORING_SETUP_CQE32) {
2681 if (check_shl_overflow(off, 1, &off))
2682 return SIZE_MAX;
2683 }
2684
2685 #ifdef CONFIG_SMP
2686 off = ALIGN(off, SMP_CACHE_BYTES);
2687 if (off == 0)
2688 return SIZE_MAX;
2689 #endif
2690
2691 if (flags & IORING_SETUP_NO_SQARRAY) {
2692 *sq_offset = SIZE_MAX;
2693 return off;
2694 }
2695
2696 *sq_offset = off;
2697
2698 sq_array_size = array_size(sizeof(u32), sq_entries);
2699 if (sq_array_size == SIZE_MAX)
2700 return SIZE_MAX;
2701
2702 if (check_add_overflow(off, sq_array_size, &off))
2703 return SIZE_MAX;
2704
2705 return off;
2706 }
2707
io_req_caches_free(struct io_ring_ctx * ctx)2708 static void io_req_caches_free(struct io_ring_ctx *ctx)
2709 {
2710 struct io_kiocb *req;
2711 int nr = 0;
2712
2713 mutex_lock(&ctx->uring_lock);
2714
2715 while (!io_req_cache_empty(ctx)) {
2716 req = io_extract_req(ctx);
2717 kmem_cache_free(req_cachep, req);
2718 nr++;
2719 }
2720 if (nr)
2721 percpu_ref_put_many(&ctx->refs, nr);
2722 mutex_unlock(&ctx->uring_lock);
2723 }
2724
io_ring_ctx_free(struct io_ring_ctx * ctx)2725 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2726 {
2727 io_sq_thread_finish(ctx);
2728
2729 mutex_lock(&ctx->uring_lock);
2730 io_sqe_buffers_unregister(ctx);
2731 io_sqe_files_unregister(ctx);
2732 io_cqring_overflow_kill(ctx);
2733 io_eventfd_unregister(ctx);
2734 io_alloc_cache_free(&ctx->apoll_cache, kfree);
2735 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2736 io_alloc_cache_free(&ctx->rw_cache, io_rw_cache_free);
2737 io_alloc_cache_free(&ctx->uring_cache, kfree);
2738 io_alloc_cache_free(&ctx->msg_cache, io_msg_cache_free);
2739 io_futex_cache_free(ctx);
2740 io_destroy_buffers(ctx);
2741 io_free_region(ctx, &ctx->param_region);
2742 mutex_unlock(&ctx->uring_lock);
2743 if (ctx->sq_creds)
2744 put_cred(ctx->sq_creds);
2745 if (ctx->submitter_task)
2746 put_task_struct(ctx->submitter_task);
2747
2748 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2749
2750 if (ctx->mm_account) {
2751 mmdrop(ctx->mm_account);
2752 ctx->mm_account = NULL;
2753 }
2754 io_rings_free(ctx);
2755
2756 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY))
2757 static_branch_dec(&io_key_has_sqarray);
2758
2759 percpu_ref_exit(&ctx->refs);
2760 free_uid(ctx->user);
2761 io_req_caches_free(ctx);
2762 if (ctx->hash_map)
2763 io_wq_put_hash(ctx->hash_map);
2764 io_napi_free(ctx);
2765 kvfree(ctx->cancel_table.hbs);
2766 xa_destroy(&ctx->io_bl_xa);
2767 kfree(ctx);
2768 }
2769
io_activate_pollwq_cb(struct callback_head * cb)2770 static __cold void io_activate_pollwq_cb(struct callback_head *cb)
2771 {
2772 struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx,
2773 poll_wq_task_work);
2774
2775 mutex_lock(&ctx->uring_lock);
2776 ctx->poll_activated = true;
2777 mutex_unlock(&ctx->uring_lock);
2778
2779 /*
2780 * Wake ups for some events between start of polling and activation
2781 * might've been lost due to loose synchronisation.
2782 */
2783 wake_up_all(&ctx->poll_wq);
2784 percpu_ref_put(&ctx->refs);
2785 }
2786
io_activate_pollwq(struct io_ring_ctx * ctx)2787 __cold void io_activate_pollwq(struct io_ring_ctx *ctx)
2788 {
2789 spin_lock(&ctx->completion_lock);
2790 /* already activated or in progress */
2791 if (ctx->poll_activated || ctx->poll_wq_task_work.func)
2792 goto out;
2793 if (WARN_ON_ONCE(!ctx->task_complete))
2794 goto out;
2795 if (!ctx->submitter_task)
2796 goto out;
2797 /*
2798 * with ->submitter_task only the submitter task completes requests, we
2799 * only need to sync with it, which is done by injecting a tw
2800 */
2801 init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb);
2802 percpu_ref_get(&ctx->refs);
2803 if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL))
2804 percpu_ref_put(&ctx->refs);
2805 out:
2806 spin_unlock(&ctx->completion_lock);
2807 }
2808
io_uring_poll(struct file * file,poll_table * wait)2809 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2810 {
2811 struct io_ring_ctx *ctx = file->private_data;
2812 __poll_t mask = 0;
2813
2814 if (unlikely(!ctx->poll_activated))
2815 io_activate_pollwq(ctx);
2816
2817 poll_wait(file, &ctx->poll_wq, wait);
2818 /*
2819 * synchronizes with barrier from wq_has_sleeper call in
2820 * io_commit_cqring
2821 */
2822 smp_rmb();
2823 if (!io_sqring_full(ctx))
2824 mask |= EPOLLOUT | EPOLLWRNORM;
2825
2826 /*
2827 * Don't flush cqring overflow list here, just do a simple check.
2828 * Otherwise there could possible be ABBA deadlock:
2829 * CPU0 CPU1
2830 * ---- ----
2831 * lock(&ctx->uring_lock);
2832 * lock(&ep->mtx);
2833 * lock(&ctx->uring_lock);
2834 * lock(&ep->mtx);
2835 *
2836 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2837 * pushes them to do the flush.
2838 */
2839
2840 if (__io_cqring_events_user(ctx) || io_has_work(ctx))
2841 mask |= EPOLLIN | EPOLLRDNORM;
2842
2843 return mask;
2844 }
2845
2846 struct io_tctx_exit {
2847 struct callback_head task_work;
2848 struct completion completion;
2849 struct io_ring_ctx *ctx;
2850 };
2851
io_tctx_exit_cb(struct callback_head * cb)2852 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2853 {
2854 struct io_uring_task *tctx = current->io_uring;
2855 struct io_tctx_exit *work;
2856
2857 work = container_of(cb, struct io_tctx_exit, task_work);
2858 /*
2859 * When @in_cancel, we're in cancellation and it's racy to remove the
2860 * node. It'll be removed by the end of cancellation, just ignore it.
2861 * tctx can be NULL if the queueing of this task_work raced with
2862 * work cancelation off the exec path.
2863 */
2864 if (tctx && !atomic_read(&tctx->in_cancel))
2865 io_uring_del_tctx_node((unsigned long)work->ctx);
2866 complete(&work->completion);
2867 }
2868
io_cancel_ctx_cb(struct io_wq_work * work,void * data)2869 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2870 {
2871 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2872
2873 return req->ctx == data;
2874 }
2875
io_ring_exit_work(struct work_struct * work)2876 static __cold void io_ring_exit_work(struct work_struct *work)
2877 {
2878 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2879 unsigned long timeout = jiffies + HZ * 60 * 5;
2880 unsigned long interval = HZ / 20;
2881 struct io_tctx_exit exit;
2882 struct io_tctx_node *node;
2883 int ret;
2884
2885 /*
2886 * If we're doing polled IO and end up having requests being
2887 * submitted async (out-of-line), then completions can come in while
2888 * we're waiting for refs to drop. We need to reap these manually,
2889 * as nobody else will be looking for them.
2890 */
2891 do {
2892 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
2893 mutex_lock(&ctx->uring_lock);
2894 io_cqring_overflow_kill(ctx);
2895 mutex_unlock(&ctx->uring_lock);
2896 }
2897
2898 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
2899 io_move_task_work_from_local(ctx);
2900
2901 while (io_uring_try_cancel_requests(ctx, NULL, true))
2902 cond_resched();
2903
2904 if (ctx->sq_data) {
2905 struct io_sq_data *sqd = ctx->sq_data;
2906 struct task_struct *tsk;
2907
2908 io_sq_thread_park(sqd);
2909 tsk = sqd->thread;
2910 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
2911 io_wq_cancel_cb(tsk->io_uring->io_wq,
2912 io_cancel_ctx_cb, ctx, true);
2913 io_sq_thread_unpark(sqd);
2914 }
2915
2916 io_req_caches_free(ctx);
2917
2918 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
2919 /* there is little hope left, don't run it too often */
2920 interval = HZ * 60;
2921 }
2922 /*
2923 * This is really an uninterruptible wait, as it has to be
2924 * complete. But it's also run from a kworker, which doesn't
2925 * take signals, so it's fine to make it interruptible. This
2926 * avoids scenarios where we knowingly can wait much longer
2927 * on completions, for example if someone does a SIGSTOP on
2928 * a task that needs to finish task_work to make this loop
2929 * complete. That's a synthetic situation that should not
2930 * cause a stuck task backtrace, and hence a potential panic
2931 * on stuck tasks if that is enabled.
2932 */
2933 } while (!wait_for_completion_interruptible_timeout(&ctx->ref_comp, interval));
2934
2935 init_completion(&exit.completion);
2936 init_task_work(&exit.task_work, io_tctx_exit_cb);
2937 exit.ctx = ctx;
2938
2939 mutex_lock(&ctx->uring_lock);
2940 while (!list_empty(&ctx->tctx_list)) {
2941 WARN_ON_ONCE(time_after(jiffies, timeout));
2942
2943 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
2944 ctx_node);
2945 /* don't spin on a single task if cancellation failed */
2946 list_rotate_left(&ctx->tctx_list);
2947 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
2948 if (WARN_ON_ONCE(ret))
2949 continue;
2950
2951 mutex_unlock(&ctx->uring_lock);
2952 /*
2953 * See comment above for
2954 * wait_for_completion_interruptible_timeout() on why this
2955 * wait is marked as interruptible.
2956 */
2957 wait_for_completion_interruptible(&exit.completion);
2958 mutex_lock(&ctx->uring_lock);
2959 }
2960 mutex_unlock(&ctx->uring_lock);
2961 spin_lock(&ctx->completion_lock);
2962 spin_unlock(&ctx->completion_lock);
2963
2964 /* pairs with RCU read section in io_req_local_work_add() */
2965 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
2966 synchronize_rcu();
2967
2968 io_ring_ctx_free(ctx);
2969 }
2970
io_ring_ctx_wait_and_kill(struct io_ring_ctx * ctx)2971 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
2972 {
2973 unsigned long index;
2974 struct creds *creds;
2975
2976 mutex_lock(&ctx->uring_lock);
2977 percpu_ref_kill(&ctx->refs);
2978 xa_for_each(&ctx->personalities, index, creds)
2979 io_unregister_personality(ctx, index);
2980 mutex_unlock(&ctx->uring_lock);
2981
2982 flush_delayed_work(&ctx->fallback_work);
2983
2984 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
2985 /*
2986 * Use system_unbound_wq to avoid spawning tons of event kworkers
2987 * if we're exiting a ton of rings at the same time. It just adds
2988 * noise and overhead, there's no discernable change in runtime
2989 * over using system_wq.
2990 */
2991 queue_work(iou_wq, &ctx->exit_work);
2992 }
2993
io_uring_release(struct inode * inode,struct file * file)2994 static int io_uring_release(struct inode *inode, struct file *file)
2995 {
2996 struct io_ring_ctx *ctx = file->private_data;
2997
2998 file->private_data = NULL;
2999 io_ring_ctx_wait_and_kill(ctx);
3000 return 0;
3001 }
3002
3003 struct io_task_cancel {
3004 struct io_uring_task *tctx;
3005 bool all;
3006 };
3007
io_cancel_task_cb(struct io_wq_work * work,void * data)3008 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
3009 {
3010 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3011 struct io_task_cancel *cancel = data;
3012
3013 return io_match_task_safe(req, cancel->tctx, cancel->all);
3014 }
3015
io_cancel_defer_files(struct io_ring_ctx * ctx,struct io_uring_task * tctx,bool cancel_all)3016 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
3017 struct io_uring_task *tctx,
3018 bool cancel_all)
3019 {
3020 struct io_defer_entry *de;
3021 LIST_HEAD(list);
3022
3023 spin_lock(&ctx->completion_lock);
3024 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
3025 if (io_match_task_safe(de->req, tctx, cancel_all)) {
3026 list_cut_position(&list, &ctx->defer_list, &de->list);
3027 break;
3028 }
3029 }
3030 spin_unlock(&ctx->completion_lock);
3031 if (list_empty(&list))
3032 return false;
3033
3034 while (!list_empty(&list)) {
3035 de = list_first_entry(&list, struct io_defer_entry, list);
3036 list_del_init(&de->list);
3037 io_req_task_queue_fail(de->req, -ECANCELED);
3038 kfree(de);
3039 }
3040 return true;
3041 }
3042
io_uring_try_cancel_iowq(struct io_ring_ctx * ctx)3043 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
3044 {
3045 struct io_tctx_node *node;
3046 enum io_wq_cancel cret;
3047 bool ret = false;
3048
3049 mutex_lock(&ctx->uring_lock);
3050 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3051 struct io_uring_task *tctx = node->task->io_uring;
3052
3053 /*
3054 * io_wq will stay alive while we hold uring_lock, because it's
3055 * killed after ctx nodes, which requires to take the lock.
3056 */
3057 if (!tctx || !tctx->io_wq)
3058 continue;
3059 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
3060 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3061 }
3062 mutex_unlock(&ctx->uring_lock);
3063
3064 return ret;
3065 }
3066
io_uring_try_cancel_requests(struct io_ring_ctx * ctx,struct io_uring_task * tctx,bool cancel_all)3067 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
3068 struct io_uring_task *tctx,
3069 bool cancel_all)
3070 {
3071 struct io_task_cancel cancel = { .tctx = tctx, .all = cancel_all, };
3072 enum io_wq_cancel cret;
3073 bool ret = false;
3074
3075 /* set it so io_req_local_work_add() would wake us up */
3076 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
3077 atomic_set(&ctx->cq_wait_nr, 1);
3078 smp_mb();
3079 }
3080
3081 /* failed during ring init, it couldn't have issued any requests */
3082 if (!ctx->rings)
3083 return false;
3084
3085 if (!tctx) {
3086 ret |= io_uring_try_cancel_iowq(ctx);
3087 } else if (tctx->io_wq) {
3088 /*
3089 * Cancels requests of all rings, not only @ctx, but
3090 * it's fine as the task is in exit/exec.
3091 */
3092 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
3093 &cancel, true);
3094 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3095 }
3096
3097 /* SQPOLL thread does its own polling */
3098 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
3099 (ctx->sq_data && ctx->sq_data->thread == current)) {
3100 while (!wq_list_empty(&ctx->iopoll_list)) {
3101 io_iopoll_try_reap_events(ctx);
3102 ret = true;
3103 cond_resched();
3104 }
3105 }
3106
3107 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3108 io_allowed_defer_tw_run(ctx))
3109 ret |= io_run_local_work(ctx, INT_MAX, INT_MAX) > 0;
3110 ret |= io_cancel_defer_files(ctx, tctx, cancel_all);
3111 mutex_lock(&ctx->uring_lock);
3112 ret |= io_poll_remove_all(ctx, tctx, cancel_all);
3113 ret |= io_waitid_remove_all(ctx, tctx, cancel_all);
3114 ret |= io_futex_remove_all(ctx, tctx, cancel_all);
3115 ret |= io_uring_try_cancel_uring_cmd(ctx, tctx, cancel_all);
3116 mutex_unlock(&ctx->uring_lock);
3117 ret |= io_kill_timeouts(ctx, tctx, cancel_all);
3118 if (tctx)
3119 ret |= io_run_task_work() > 0;
3120 else
3121 ret |= flush_delayed_work(&ctx->fallback_work);
3122 return ret;
3123 }
3124
tctx_inflight(struct io_uring_task * tctx,bool tracked)3125 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
3126 {
3127 if (tracked)
3128 return atomic_read(&tctx->inflight_tracked);
3129 return percpu_counter_sum(&tctx->inflight);
3130 }
3131
3132 /*
3133 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3134 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3135 */
io_uring_cancel_generic(bool cancel_all,struct io_sq_data * sqd)3136 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
3137 {
3138 struct io_uring_task *tctx = current->io_uring;
3139 struct io_ring_ctx *ctx;
3140 struct io_tctx_node *node;
3141 unsigned long index;
3142 s64 inflight;
3143 DEFINE_WAIT(wait);
3144
3145 WARN_ON_ONCE(sqd && sqd->thread != current);
3146
3147 if (!current->io_uring)
3148 return;
3149 if (tctx->io_wq)
3150 io_wq_exit_start(tctx->io_wq);
3151
3152 atomic_inc(&tctx->in_cancel);
3153 do {
3154 bool loop = false;
3155
3156 io_uring_drop_tctx_refs(current);
3157 if (!tctx_inflight(tctx, !cancel_all))
3158 break;
3159
3160 /* read completions before cancelations */
3161 inflight = tctx_inflight(tctx, false);
3162 if (!inflight)
3163 break;
3164
3165 if (!sqd) {
3166 xa_for_each(&tctx->xa, index, node) {
3167 /* sqpoll task will cancel all its requests */
3168 if (node->ctx->sq_data)
3169 continue;
3170 loop |= io_uring_try_cancel_requests(node->ctx,
3171 current->io_uring,
3172 cancel_all);
3173 }
3174 } else {
3175 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3176 loop |= io_uring_try_cancel_requests(ctx,
3177 current->io_uring,
3178 cancel_all);
3179 }
3180
3181 if (loop) {
3182 cond_resched();
3183 continue;
3184 }
3185
3186 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3187 io_run_task_work();
3188 io_uring_drop_tctx_refs(current);
3189 xa_for_each(&tctx->xa, index, node) {
3190 if (io_local_work_pending(node->ctx)) {
3191 WARN_ON_ONCE(node->ctx->submitter_task &&
3192 node->ctx->submitter_task != current);
3193 goto end_wait;
3194 }
3195 }
3196 /*
3197 * If we've seen completions, retry without waiting. This
3198 * avoids a race where a completion comes in before we did
3199 * prepare_to_wait().
3200 */
3201 if (inflight == tctx_inflight(tctx, !cancel_all))
3202 schedule();
3203 end_wait:
3204 finish_wait(&tctx->wait, &wait);
3205 } while (1);
3206
3207 io_uring_clean_tctx(tctx);
3208 if (cancel_all) {
3209 /*
3210 * We shouldn't run task_works after cancel, so just leave
3211 * ->in_cancel set for normal exit.
3212 */
3213 atomic_dec(&tctx->in_cancel);
3214 /* for exec all current's requests should be gone, kill tctx */
3215 __io_uring_free(current);
3216 }
3217 }
3218
__io_uring_cancel(bool cancel_all)3219 void __io_uring_cancel(bool cancel_all)
3220 {
3221 io_uring_unreg_ringfd();
3222 io_uring_cancel_generic(cancel_all, NULL);
3223 }
3224
io_get_ext_arg_reg(struct io_ring_ctx * ctx,const struct io_uring_getevents_arg __user * uarg)3225 static struct io_uring_reg_wait *io_get_ext_arg_reg(struct io_ring_ctx *ctx,
3226 const struct io_uring_getevents_arg __user *uarg)
3227 {
3228 unsigned long size = sizeof(struct io_uring_reg_wait);
3229 unsigned long offset = (uintptr_t)uarg;
3230 unsigned long end;
3231
3232 if (unlikely(offset % sizeof(long)))
3233 return ERR_PTR(-EFAULT);
3234
3235 /* also protects from NULL ->cq_wait_arg as the size would be 0 */
3236 if (unlikely(check_add_overflow(offset, size, &end) ||
3237 end > ctx->cq_wait_size))
3238 return ERR_PTR(-EFAULT);
3239
3240 return ctx->cq_wait_arg + offset;
3241 }
3242
io_validate_ext_arg(struct io_ring_ctx * ctx,unsigned flags,const void __user * argp,size_t argsz)3243 static int io_validate_ext_arg(struct io_ring_ctx *ctx, unsigned flags,
3244 const void __user *argp, size_t argsz)
3245 {
3246 struct io_uring_getevents_arg arg;
3247
3248 if (!(flags & IORING_ENTER_EXT_ARG))
3249 return 0;
3250 if (flags & IORING_ENTER_EXT_ARG_REG)
3251 return -EINVAL;
3252 if (argsz != sizeof(arg))
3253 return -EINVAL;
3254 if (copy_from_user(&arg, argp, sizeof(arg)))
3255 return -EFAULT;
3256 return 0;
3257 }
3258
io_get_ext_arg(struct io_ring_ctx * ctx,unsigned flags,const void __user * argp,struct ext_arg * ext_arg)3259 static int io_get_ext_arg(struct io_ring_ctx *ctx, unsigned flags,
3260 const void __user *argp, struct ext_arg *ext_arg)
3261 {
3262 const struct io_uring_getevents_arg __user *uarg = argp;
3263 struct io_uring_getevents_arg arg;
3264
3265 /*
3266 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3267 * is just a pointer to the sigset_t.
3268 */
3269 if (!(flags & IORING_ENTER_EXT_ARG)) {
3270 ext_arg->sig = (const sigset_t __user *) argp;
3271 return 0;
3272 }
3273
3274 if (flags & IORING_ENTER_EXT_ARG_REG) {
3275 struct io_uring_reg_wait *w;
3276
3277 if (ext_arg->argsz != sizeof(struct io_uring_reg_wait))
3278 return -EINVAL;
3279 w = io_get_ext_arg_reg(ctx, argp);
3280 if (IS_ERR(w))
3281 return PTR_ERR(w);
3282
3283 if (w->flags & ~IORING_REG_WAIT_TS)
3284 return -EINVAL;
3285 ext_arg->min_time = READ_ONCE(w->min_wait_usec) * NSEC_PER_USEC;
3286 ext_arg->sig = u64_to_user_ptr(READ_ONCE(w->sigmask));
3287 ext_arg->argsz = READ_ONCE(w->sigmask_sz);
3288 if (w->flags & IORING_REG_WAIT_TS) {
3289 ext_arg->ts.tv_sec = READ_ONCE(w->ts.tv_sec);
3290 ext_arg->ts.tv_nsec = READ_ONCE(w->ts.tv_nsec);
3291 ext_arg->ts_set = true;
3292 }
3293 return 0;
3294 }
3295
3296 /*
3297 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3298 * timespec and sigset_t pointers if good.
3299 */
3300 if (ext_arg->argsz != sizeof(arg))
3301 return -EINVAL;
3302 #ifdef CONFIG_64BIT
3303 if (!user_access_begin(uarg, sizeof(*uarg)))
3304 return -EFAULT;
3305 unsafe_get_user(arg.sigmask, &uarg->sigmask, uaccess_end);
3306 unsafe_get_user(arg.sigmask_sz, &uarg->sigmask_sz, uaccess_end);
3307 unsafe_get_user(arg.min_wait_usec, &uarg->min_wait_usec, uaccess_end);
3308 unsafe_get_user(arg.ts, &uarg->ts, uaccess_end);
3309 user_access_end();
3310 #else
3311 if (copy_from_user(&arg, uarg, sizeof(arg)))
3312 return -EFAULT;
3313 #endif
3314 ext_arg->min_time = arg.min_wait_usec * NSEC_PER_USEC;
3315 ext_arg->sig = u64_to_user_ptr(arg.sigmask);
3316 ext_arg->argsz = arg.sigmask_sz;
3317 if (arg.ts) {
3318 if (get_timespec64(&ext_arg->ts, u64_to_user_ptr(arg.ts)))
3319 return -EFAULT;
3320 ext_arg->ts_set = true;
3321 }
3322 return 0;
3323 #ifdef CONFIG_64BIT
3324 uaccess_end:
3325 user_access_end();
3326 return -EFAULT;
3327 #endif
3328 }
3329
SYSCALL_DEFINE6(io_uring_enter,unsigned int,fd,u32,to_submit,u32,min_complete,u32,flags,const void __user *,argp,size_t,argsz)3330 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3331 u32, min_complete, u32, flags, const void __user *, argp,
3332 size_t, argsz)
3333 {
3334 struct io_ring_ctx *ctx;
3335 struct file *file;
3336 long ret;
3337
3338 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3339 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3340 IORING_ENTER_REGISTERED_RING |
3341 IORING_ENTER_ABS_TIMER |
3342 IORING_ENTER_EXT_ARG_REG)))
3343 return -EINVAL;
3344
3345 /*
3346 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3347 * need only dereference our task private array to find it.
3348 */
3349 if (flags & IORING_ENTER_REGISTERED_RING) {
3350 struct io_uring_task *tctx = current->io_uring;
3351
3352 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3353 return -EINVAL;
3354 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3355 file = tctx->registered_rings[fd];
3356 if (unlikely(!file))
3357 return -EBADF;
3358 } else {
3359 file = fget(fd);
3360 if (unlikely(!file))
3361 return -EBADF;
3362 ret = -EOPNOTSUPP;
3363 if (unlikely(!io_is_uring_fops(file)))
3364 goto out;
3365 }
3366
3367 ctx = file->private_data;
3368 ret = -EBADFD;
3369 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3370 goto out;
3371
3372 /*
3373 * For SQ polling, the thread will do all submissions and completions.
3374 * Just return the requested submit count, and wake the thread if
3375 * we were asked to.
3376 */
3377 ret = 0;
3378 if (ctx->flags & IORING_SETUP_SQPOLL) {
3379 if (unlikely(ctx->sq_data->thread == NULL)) {
3380 ret = -EOWNERDEAD;
3381 goto out;
3382 }
3383 if (flags & IORING_ENTER_SQ_WAKEUP)
3384 wake_up(&ctx->sq_data->wait);
3385 if (flags & IORING_ENTER_SQ_WAIT)
3386 io_sqpoll_wait_sq(ctx);
3387
3388 ret = to_submit;
3389 } else if (to_submit) {
3390 ret = io_uring_add_tctx_node(ctx);
3391 if (unlikely(ret))
3392 goto out;
3393
3394 mutex_lock(&ctx->uring_lock);
3395 ret = io_submit_sqes(ctx, to_submit);
3396 if (ret != to_submit) {
3397 mutex_unlock(&ctx->uring_lock);
3398 goto out;
3399 }
3400 if (flags & IORING_ENTER_GETEVENTS) {
3401 if (ctx->syscall_iopoll)
3402 goto iopoll_locked;
3403 /*
3404 * Ignore errors, we'll soon call io_cqring_wait() and
3405 * it should handle ownership problems if any.
3406 */
3407 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3408 (void)io_run_local_work_locked(ctx, min_complete);
3409 }
3410 mutex_unlock(&ctx->uring_lock);
3411 }
3412
3413 if (flags & IORING_ENTER_GETEVENTS) {
3414 int ret2;
3415
3416 if (ctx->syscall_iopoll) {
3417 /*
3418 * We disallow the app entering submit/complete with
3419 * polling, but we still need to lock the ring to
3420 * prevent racing with polled issue that got punted to
3421 * a workqueue.
3422 */
3423 mutex_lock(&ctx->uring_lock);
3424 iopoll_locked:
3425 ret2 = io_validate_ext_arg(ctx, flags, argp, argsz);
3426 if (likely(!ret2)) {
3427 min_complete = min(min_complete,
3428 ctx->cq_entries);
3429 ret2 = io_iopoll_check(ctx, min_complete);
3430 }
3431 mutex_unlock(&ctx->uring_lock);
3432 } else {
3433 struct ext_arg ext_arg = { .argsz = argsz };
3434
3435 ret2 = io_get_ext_arg(ctx, flags, argp, &ext_arg);
3436 if (likely(!ret2)) {
3437 min_complete = min(min_complete,
3438 ctx->cq_entries);
3439 ret2 = io_cqring_wait(ctx, min_complete, flags,
3440 &ext_arg);
3441 }
3442 }
3443
3444 if (!ret) {
3445 ret = ret2;
3446
3447 /*
3448 * EBADR indicates that one or more CQE were dropped.
3449 * Once the user has been informed we can clear the bit
3450 * as they are obviously ok with those drops.
3451 */
3452 if (unlikely(ret2 == -EBADR))
3453 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3454 &ctx->check_cq);
3455 }
3456 }
3457 out:
3458 if (!(flags & IORING_ENTER_REGISTERED_RING))
3459 fput(file);
3460 return ret;
3461 }
3462
3463 static const struct file_operations io_uring_fops = {
3464 .release = io_uring_release,
3465 .mmap = io_uring_mmap,
3466 .get_unmapped_area = io_uring_get_unmapped_area,
3467 #ifndef CONFIG_MMU
3468 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3469 #endif
3470 .poll = io_uring_poll,
3471 #ifdef CONFIG_PROC_FS
3472 .show_fdinfo = io_uring_show_fdinfo,
3473 #endif
3474 };
3475
io_is_uring_fops(struct file * file)3476 bool io_is_uring_fops(struct file *file)
3477 {
3478 return file->f_op == &io_uring_fops;
3479 }
3480
io_allocate_scq_urings(struct io_ring_ctx * ctx,struct io_uring_params * p)3481 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3482 struct io_uring_params *p)
3483 {
3484 struct io_rings *rings;
3485 size_t size, sq_array_offset;
3486 void *ptr;
3487
3488 /* make sure these are sane, as we already accounted them */
3489 ctx->sq_entries = p->sq_entries;
3490 ctx->cq_entries = p->cq_entries;
3491
3492 size = rings_size(ctx->flags, p->sq_entries, p->cq_entries,
3493 &sq_array_offset);
3494 if (size == SIZE_MAX)
3495 return -EOVERFLOW;
3496
3497 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3498 rings = io_pages_map(&ctx->ring_pages, &ctx->n_ring_pages, size);
3499 else
3500 rings = io_rings_map(ctx, p->cq_off.user_addr, size);
3501
3502 if (IS_ERR(rings))
3503 return PTR_ERR(rings);
3504
3505 ctx->rings = rings;
3506 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY))
3507 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3508 rings->sq_ring_mask = p->sq_entries - 1;
3509 rings->cq_ring_mask = p->cq_entries - 1;
3510 rings->sq_ring_entries = p->sq_entries;
3511 rings->cq_ring_entries = p->cq_entries;
3512
3513 if (p->flags & IORING_SETUP_SQE128)
3514 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3515 else
3516 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3517 if (size == SIZE_MAX) {
3518 io_rings_free(ctx);
3519 return -EOVERFLOW;
3520 }
3521
3522 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3523 ptr = io_pages_map(&ctx->sqe_pages, &ctx->n_sqe_pages, size);
3524 else
3525 ptr = io_sqes_map(ctx, p->sq_off.user_addr, size);
3526
3527 if (IS_ERR(ptr)) {
3528 io_rings_free(ctx);
3529 return PTR_ERR(ptr);
3530 }
3531
3532 ctx->sq_sqes = ptr;
3533 return 0;
3534 }
3535
io_uring_install_fd(struct file * file)3536 static int io_uring_install_fd(struct file *file)
3537 {
3538 int fd;
3539
3540 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3541 if (fd < 0)
3542 return fd;
3543 fd_install(fd, file);
3544 return fd;
3545 }
3546
3547 /*
3548 * Allocate an anonymous fd, this is what constitutes the application
3549 * visible backing of an io_uring instance. The application mmaps this
3550 * fd to gain access to the SQ/CQ ring details.
3551 */
io_uring_get_file(struct io_ring_ctx * ctx)3552 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3553 {
3554 /* Create a new inode so that the LSM can block the creation. */
3555 return anon_inode_create_getfile("[io_uring]", &io_uring_fops, ctx,
3556 O_RDWR | O_CLOEXEC, NULL);
3557 }
3558
io_uring_fill_params(unsigned entries,struct io_uring_params * p)3559 int io_uring_fill_params(unsigned entries, struct io_uring_params *p)
3560 {
3561 if (!entries)
3562 return -EINVAL;
3563 if (entries > IORING_MAX_ENTRIES) {
3564 if (!(p->flags & IORING_SETUP_CLAMP))
3565 return -EINVAL;
3566 entries = IORING_MAX_ENTRIES;
3567 }
3568
3569 if ((p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
3570 && !(p->flags & IORING_SETUP_NO_MMAP))
3571 return -EINVAL;
3572
3573 /*
3574 * Use twice as many entries for the CQ ring. It's possible for the
3575 * application to drive a higher depth than the size of the SQ ring,
3576 * since the sqes are only used at submission time. This allows for
3577 * some flexibility in overcommitting a bit. If the application has
3578 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3579 * of CQ ring entries manually.
3580 */
3581 p->sq_entries = roundup_pow_of_two(entries);
3582 if (p->flags & IORING_SETUP_CQSIZE) {
3583 /*
3584 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3585 * to a power-of-two, if it isn't already. We do NOT impose
3586 * any cq vs sq ring sizing.
3587 */
3588 if (!p->cq_entries)
3589 return -EINVAL;
3590 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3591 if (!(p->flags & IORING_SETUP_CLAMP))
3592 return -EINVAL;
3593 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3594 }
3595 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3596 if (p->cq_entries < p->sq_entries)
3597 return -EINVAL;
3598 } else {
3599 p->cq_entries = 2 * p->sq_entries;
3600 }
3601
3602 p->sq_off.head = offsetof(struct io_rings, sq.head);
3603 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3604 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3605 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3606 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3607 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3608 p->sq_off.resv1 = 0;
3609 if (!(p->flags & IORING_SETUP_NO_MMAP))
3610 p->sq_off.user_addr = 0;
3611
3612 p->cq_off.head = offsetof(struct io_rings, cq.head);
3613 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3614 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3615 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3616 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3617 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3618 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3619 p->cq_off.resv1 = 0;
3620 if (!(p->flags & IORING_SETUP_NO_MMAP))
3621 p->cq_off.user_addr = 0;
3622
3623 return 0;
3624 }
3625
io_uring_create(unsigned entries,struct io_uring_params * p,struct io_uring_params __user * params)3626 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3627 struct io_uring_params __user *params)
3628 {
3629 struct io_ring_ctx *ctx;
3630 struct io_uring_task *tctx;
3631 struct file *file;
3632 int ret;
3633
3634 ret = io_uring_fill_params(entries, p);
3635 if (unlikely(ret))
3636 return ret;
3637
3638 ctx = io_ring_ctx_alloc(p);
3639 if (!ctx)
3640 return -ENOMEM;
3641
3642 ctx->clockid = CLOCK_MONOTONIC;
3643 ctx->clock_offset = 0;
3644
3645 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY))
3646 static_branch_inc(&io_key_has_sqarray);
3647
3648 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3649 !(ctx->flags & IORING_SETUP_IOPOLL) &&
3650 !(ctx->flags & IORING_SETUP_SQPOLL))
3651 ctx->task_complete = true;
3652
3653 if (ctx->task_complete || (ctx->flags & IORING_SETUP_IOPOLL))
3654 ctx->lockless_cq = true;
3655
3656 /*
3657 * lazy poll_wq activation relies on ->task_complete for synchronisation
3658 * purposes, see io_activate_pollwq()
3659 */
3660 if (!ctx->task_complete)
3661 ctx->poll_activated = true;
3662
3663 /*
3664 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3665 * space applications don't need to do io completion events
3666 * polling again, they can rely on io_sq_thread to do polling
3667 * work, which can reduce cpu usage and uring_lock contention.
3668 */
3669 if (ctx->flags & IORING_SETUP_IOPOLL &&
3670 !(ctx->flags & IORING_SETUP_SQPOLL))
3671 ctx->syscall_iopoll = 1;
3672
3673 ctx->compat = in_compat_syscall();
3674 if (!ns_capable_noaudit(&init_user_ns, CAP_IPC_LOCK))
3675 ctx->user = get_uid(current_user());
3676
3677 /*
3678 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3679 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3680 */
3681 ret = -EINVAL;
3682 if (ctx->flags & IORING_SETUP_SQPOLL) {
3683 /* IPI related flags don't make sense with SQPOLL */
3684 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3685 IORING_SETUP_TASKRUN_FLAG |
3686 IORING_SETUP_DEFER_TASKRUN))
3687 goto err;
3688 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3689 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3690 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3691 } else {
3692 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3693 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3694 goto err;
3695 ctx->notify_method = TWA_SIGNAL;
3696 }
3697
3698 /* HYBRID_IOPOLL only valid with IOPOLL */
3699 if ((ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_HYBRID_IOPOLL)) ==
3700 IORING_SETUP_HYBRID_IOPOLL)
3701 goto err;
3702
3703 /*
3704 * For DEFER_TASKRUN we require the completion task to be the same as the
3705 * submission task. This implies that there is only one submitter, so enforce
3706 * that.
3707 */
3708 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3709 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3710 goto err;
3711 }
3712
3713 /*
3714 * This is just grabbed for accounting purposes. When a process exits,
3715 * the mm is exited and dropped before the files, hence we need to hang
3716 * on to this mm purely for the purposes of being able to unaccount
3717 * memory (locked/pinned vm). It's not used for anything else.
3718 */
3719 mmgrab(current->mm);
3720 ctx->mm_account = current->mm;
3721
3722 ret = io_allocate_scq_urings(ctx, p);
3723 if (ret)
3724 goto err;
3725
3726 if (!(p->flags & IORING_SETUP_NO_SQARRAY))
3727 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3728
3729 ret = io_sq_offload_create(ctx, p);
3730 if (ret)
3731 goto err;
3732
3733 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3734 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3735 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3736 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3737 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3738 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3739 IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING |
3740 IORING_FEAT_RECVSEND_BUNDLE | IORING_FEAT_MIN_TIMEOUT;
3741
3742 if (copy_to_user(params, p, sizeof(*p))) {
3743 ret = -EFAULT;
3744 goto err;
3745 }
3746
3747 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
3748 && !(ctx->flags & IORING_SETUP_R_DISABLED))
3749 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
3750
3751 file = io_uring_get_file(ctx);
3752 if (IS_ERR(file)) {
3753 ret = PTR_ERR(file);
3754 goto err;
3755 }
3756
3757 ret = __io_uring_add_tctx_node(ctx);
3758 if (ret)
3759 goto err_fput;
3760 tctx = current->io_uring;
3761
3762 /*
3763 * Install ring fd as the very last thing, so we don't risk someone
3764 * having closed it before we finish setup
3765 */
3766 if (p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
3767 ret = io_ring_add_registered_file(tctx, file, 0, IO_RINGFD_REG_MAX);
3768 else
3769 ret = io_uring_install_fd(file);
3770 if (ret < 0)
3771 goto err_fput;
3772
3773 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3774 return ret;
3775 err:
3776 io_ring_ctx_wait_and_kill(ctx);
3777 return ret;
3778 err_fput:
3779 fput(file);
3780 return ret;
3781 }
3782
3783 /*
3784 * Sets up an aio uring context, and returns the fd. Applications asks for a
3785 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3786 * params structure passed in.
3787 */
io_uring_setup(u32 entries,struct io_uring_params __user * params)3788 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3789 {
3790 struct io_uring_params p;
3791 int i;
3792
3793 if (copy_from_user(&p, params, sizeof(p)))
3794 return -EFAULT;
3795 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3796 if (p.resv[i])
3797 return -EINVAL;
3798 }
3799
3800 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3801 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3802 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3803 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3804 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3805 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3806 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN |
3807 IORING_SETUP_NO_MMAP | IORING_SETUP_REGISTERED_FD_ONLY |
3808 IORING_SETUP_NO_SQARRAY | IORING_SETUP_HYBRID_IOPOLL))
3809 return -EINVAL;
3810
3811 return io_uring_create(entries, &p, params);
3812 }
3813
io_uring_allowed(void)3814 static inline bool io_uring_allowed(void)
3815 {
3816 int disabled = READ_ONCE(sysctl_io_uring_disabled);
3817 kgid_t io_uring_group;
3818
3819 if (disabled == 2)
3820 return false;
3821
3822 if (disabled == 0 || capable(CAP_SYS_ADMIN))
3823 return true;
3824
3825 io_uring_group = make_kgid(&init_user_ns, sysctl_io_uring_group);
3826 if (!gid_valid(io_uring_group))
3827 return false;
3828
3829 return in_group_p(io_uring_group);
3830 }
3831
SYSCALL_DEFINE2(io_uring_setup,u32,entries,struct io_uring_params __user *,params)3832 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3833 struct io_uring_params __user *, params)
3834 {
3835 if (!io_uring_allowed())
3836 return -EPERM;
3837
3838 return io_uring_setup(entries, params);
3839 }
3840
io_uring_init(void)3841 static int __init io_uring_init(void)
3842 {
3843 struct kmem_cache_args kmem_args = {
3844 .useroffset = offsetof(struct io_kiocb, cmd.data),
3845 .usersize = sizeof_field(struct io_kiocb, cmd.data),
3846 .freeptr_offset = offsetof(struct io_kiocb, work),
3847 .use_freeptr_offset = true,
3848 };
3849
3850 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
3851 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
3852 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
3853 } while (0)
3854
3855 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
3856 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
3857 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
3858 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
3859 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
3860 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
3861 BUILD_BUG_SQE_ELEM(1, __u8, flags);
3862 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
3863 BUILD_BUG_SQE_ELEM(4, __s32, fd);
3864 BUILD_BUG_SQE_ELEM(8, __u64, off);
3865 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
3866 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
3867 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
3868 BUILD_BUG_SQE_ELEM(16, __u64, addr);
3869 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
3870 BUILD_BUG_SQE_ELEM(24, __u32, len);
3871 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
3872 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
3873 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
3874 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
3875 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
3876 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
3877 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
3878 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
3879 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
3880 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
3881 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
3882 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
3883 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
3884 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
3885 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
3886 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
3887 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
3888 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
3889 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
3890 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
3891 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
3892 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
3893 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
3894 BUILD_BUG_SQE_ELEM(42, __u16, personality);
3895 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
3896 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
3897 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
3898 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
3899 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
3900 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
3901 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
3902
3903 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
3904 sizeof(struct io_uring_rsrc_update));
3905 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
3906 sizeof(struct io_uring_rsrc_update2));
3907
3908 /* ->buf_index is u16 */
3909 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
3910 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
3911 offsetof(struct io_uring_buf_ring, tail));
3912
3913 /* should fit into one byte */
3914 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
3915 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
3916 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
3917
3918 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof_field(struct io_kiocb, flags));
3919
3920 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
3921
3922 /* top 8bits are for internal use */
3923 BUILD_BUG_ON((IORING_URING_CMD_MASK & 0xff000000) != 0);
3924
3925 io_uring_optable_init();
3926
3927 /*
3928 * Allow user copy in the per-command field, which starts after the
3929 * file in io_kiocb and until the opcode field. The openat2 handling
3930 * requires copying in user memory into the io_kiocb object in that
3931 * range, and HARDENED_USERCOPY will complain if we haven't
3932 * correctly annotated this range.
3933 */
3934 req_cachep = kmem_cache_create("io_kiocb", sizeof(struct io_kiocb), &kmem_args,
3935 SLAB_HWCACHE_ALIGN | SLAB_PANIC | SLAB_ACCOUNT |
3936 SLAB_TYPESAFE_BY_RCU);
3937 io_buf_cachep = KMEM_CACHE(io_buffer,
3938 SLAB_HWCACHE_ALIGN | SLAB_PANIC | SLAB_ACCOUNT);
3939
3940 iou_wq = alloc_workqueue("iou_exit", WQ_UNBOUND, 64);
3941
3942 #ifdef CONFIG_SYSCTL
3943 register_sysctl_init("kernel", kernel_io_uring_disabled_table);
3944 #endif
3945
3946 return 0;
3947 };
3948 __initcall(io_uring_init);
3949