Loading...
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
126struct 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
144static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
145 struct io_uring_task *tctx,
146 bool cancel_all);
147
148static void io_queue_sqe(struct io_kiocb *req);
149
150static __read_mostly DEFINE_STATIC_KEY_FALSE(io_key_has_sqarray);
151
152struct kmem_cache *req_cachep;
153static struct workqueue_struct *iou_wq __ro_after_init;
154
155static int __read_mostly sysctl_io_uring_disabled;
156static int __read_mostly sysctl_io_uring_group = -1;
157
158#ifdef CONFIG_SYSCTL
159static 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
179static 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
184static 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
189static 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 */
204bool 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
227static 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
233static 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
238static __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
245static __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
262static 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
284static __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 io_uring_cmd_data));
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
357free_ref:
358 percpu_ref_exit(&ctx->refs);
359err:
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
372static 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
380static 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
391static 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
421static 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
429static 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
443static 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
450static noinline void __io_arm_ltimeout(struct io_kiocb *req)
451{
452 io_queue_linked_timeout(__io_prep_linked_timeout(req));
453}
454
455static 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
461static 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
494static 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
511static 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
542static void io_req_queue_iowq_tw(struct io_kiocb *req, struct io_tw_state *ts)
543{
544 io_queue_iowq(req);
545}
546
547void 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
553static __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
567void __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
582static inline void __io_cq_lock(struct io_ring_ctx *ctx)
583{
584 if (!ctx->lockless_cq)
585 spin_lock(&ctx->completion_lock);
586}
587
588static inline void io_cq_lock(struct io_ring_ctx *ctx)
589 __acquires(ctx->completion_lock)
590{
591 spin_lock(&ctx->completion_lock);
592}
593
594static 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
607static 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
616static 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
668static void io_cqring_overflow_kill(struct io_ring_ctx *ctx)
669{
670 if (ctx->rings)
671 __io_cqring_overflow_flush(ctx, true);
672}
673
674static 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 */
682static 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
696void 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
705static __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
717static 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
757static 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 */
770bool 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
802static 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
830static 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
842bool 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 */
856void 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 */
870bool 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
885static 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
920void 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 */
938static 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 */
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
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
996static 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
1005static 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
1022static 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 */
1039struct 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
1073static __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
1100static 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
1107struct 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
1132void 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
1144static 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
1215static 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 __set_notify_signal(tctx->task);
1230 return;
1231 }
1232
1233 if (likely(!task_work_add(tctx->task, &tctx->task_work, ctx->notify_method)))
1234 return;
1235
1236 io_fallback_tw(tctx, false);
1237}
1238
1239void __io_req_task_work_add(struct io_kiocb *req, unsigned flags)
1240{
1241 if (req->ctx->flags & IORING_SETUP_DEFER_TASKRUN)
1242 io_req_local_work_add(req, req->ctx, flags);
1243 else
1244 io_req_normal_work_add(req);
1245}
1246
1247void io_req_task_work_add_remote(struct io_kiocb *req, struct io_ring_ctx *ctx,
1248 unsigned flags)
1249{
1250 if (WARN_ON_ONCE(!(ctx->flags & IORING_SETUP_DEFER_TASKRUN)))
1251 return;
1252 io_req_local_work_add(req, ctx, flags);
1253}
1254
1255static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1256{
1257 struct llist_node *node = llist_del_all(&ctx->work_llist);
1258
1259 __io_fallback_tw(node, false);
1260 node = llist_del_all(&ctx->retry_llist);
1261 __io_fallback_tw(node, false);
1262}
1263
1264static bool io_run_local_work_continue(struct io_ring_ctx *ctx, int events,
1265 int min_events)
1266{
1267 if (!io_local_work_pending(ctx))
1268 return false;
1269 if (events < min_events)
1270 return true;
1271 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1272 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1273 return false;
1274}
1275
1276static int __io_run_local_work_loop(struct llist_node **node,
1277 struct io_tw_state *ts,
1278 int events)
1279{
1280 int ret = 0;
1281
1282 while (*node) {
1283 struct llist_node *next = (*node)->next;
1284 struct io_kiocb *req = container_of(*node, struct io_kiocb,
1285 io_task_work.node);
1286 INDIRECT_CALL_2(req->io_task_work.func,
1287 io_poll_task_func, io_req_rw_complete,
1288 req, ts);
1289 *node = next;
1290 if (++ret >= events)
1291 break;
1292 }
1293
1294 return ret;
1295}
1296
1297static int __io_run_local_work(struct io_ring_ctx *ctx, struct io_tw_state *ts,
1298 int min_events, int max_events)
1299{
1300 struct llist_node *node;
1301 unsigned int loops = 0;
1302 int ret = 0;
1303
1304 if (WARN_ON_ONCE(ctx->submitter_task != current))
1305 return -EEXIST;
1306 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1307 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1308again:
1309 min_events -= ret;
1310 ret = __io_run_local_work_loop(&ctx->retry_llist.first, ts, max_events);
1311 if (ctx->retry_llist.first)
1312 goto retry_done;
1313
1314 /*
1315 * llists are in reverse order, flip it back the right way before
1316 * running the pending items.
1317 */
1318 node = llist_reverse_order(llist_del_all(&ctx->work_llist));
1319 ret += __io_run_local_work_loop(&node, ts, max_events - ret);
1320 ctx->retry_llist.first = node;
1321 loops++;
1322
1323 if (io_run_local_work_continue(ctx, ret, min_events))
1324 goto again;
1325retry_done:
1326 io_submit_flush_completions(ctx);
1327 if (io_run_local_work_continue(ctx, ret, min_events))
1328 goto again;
1329
1330 trace_io_uring_local_work_run(ctx, ret, loops);
1331 return ret;
1332}
1333
1334static inline int io_run_local_work_locked(struct io_ring_ctx *ctx,
1335 int min_events)
1336{
1337 struct io_tw_state ts = {};
1338
1339 if (!io_local_work_pending(ctx))
1340 return 0;
1341 return __io_run_local_work(ctx, &ts, min_events,
1342 max(IO_LOCAL_TW_DEFAULT_MAX, min_events));
1343}
1344
1345static int io_run_local_work(struct io_ring_ctx *ctx, int min_events,
1346 int max_events)
1347{
1348 struct io_tw_state ts = {};
1349 int ret;
1350
1351 mutex_lock(&ctx->uring_lock);
1352 ret = __io_run_local_work(ctx, &ts, min_events, max_events);
1353 mutex_unlock(&ctx->uring_lock);
1354 return ret;
1355}
1356
1357static void io_req_task_cancel(struct io_kiocb *req, struct io_tw_state *ts)
1358{
1359 io_tw_lock(req->ctx, ts);
1360 io_req_defer_failed(req, req->cqe.res);
1361}
1362
1363void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts)
1364{
1365 io_tw_lock(req->ctx, ts);
1366 if (unlikely(io_should_terminate_tw()))
1367 io_req_defer_failed(req, -EFAULT);
1368 else if (req->flags & REQ_F_FORCE_ASYNC)
1369 io_queue_iowq(req);
1370 else
1371 io_queue_sqe(req);
1372}
1373
1374void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1375{
1376 io_req_set_res(req, ret, 0);
1377 req->io_task_work.func = io_req_task_cancel;
1378 io_req_task_work_add(req);
1379}
1380
1381void io_req_task_queue(struct io_kiocb *req)
1382{
1383 req->io_task_work.func = io_req_task_submit;
1384 io_req_task_work_add(req);
1385}
1386
1387void io_queue_next(struct io_kiocb *req)
1388{
1389 struct io_kiocb *nxt = io_req_find_next(req);
1390
1391 if (nxt)
1392 io_req_task_queue(nxt);
1393}
1394
1395static void io_free_batch_list(struct io_ring_ctx *ctx,
1396 struct io_wq_work_node *node)
1397 __must_hold(&ctx->uring_lock)
1398{
1399 do {
1400 struct io_kiocb *req = container_of(node, struct io_kiocb,
1401 comp_list);
1402
1403 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1404 if (req->flags & REQ_F_REFCOUNT) {
1405 node = req->comp_list.next;
1406 if (!req_ref_put_and_test(req))
1407 continue;
1408 }
1409 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1410 struct async_poll *apoll = req->apoll;
1411
1412 if (apoll->double_poll)
1413 kfree(apoll->double_poll);
1414 if (!io_alloc_cache_put(&ctx->apoll_cache, apoll))
1415 kfree(apoll);
1416 req->flags &= ~REQ_F_POLLED;
1417 }
1418 if (req->flags & IO_REQ_LINK_FLAGS)
1419 io_queue_next(req);
1420 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1421 io_clean_op(req);
1422 }
1423 io_put_file(req);
1424 io_req_put_rsrc_nodes(req);
1425 io_put_task(req);
1426
1427 node = req->comp_list.next;
1428 io_req_add_to_cache(req, ctx);
1429 } while (node);
1430}
1431
1432void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1433 __must_hold(&ctx->uring_lock)
1434{
1435 struct io_submit_state *state = &ctx->submit_state;
1436 struct io_wq_work_node *node;
1437
1438 __io_cq_lock(ctx);
1439 __wq_list_for_each(node, &state->compl_reqs) {
1440 struct io_kiocb *req = container_of(node, struct io_kiocb,
1441 comp_list);
1442
1443 if (!(req->flags & REQ_F_CQE_SKIP) &&
1444 unlikely(!io_fill_cqe_req(ctx, req))) {
1445 if (ctx->lockless_cq) {
1446 spin_lock(&ctx->completion_lock);
1447 io_req_cqe_overflow(req);
1448 spin_unlock(&ctx->completion_lock);
1449 } else {
1450 io_req_cqe_overflow(req);
1451 }
1452 }
1453 }
1454 __io_cq_unlock_post(ctx);
1455
1456 if (!wq_list_empty(&state->compl_reqs)) {
1457 io_free_batch_list(ctx, state->compl_reqs.first);
1458 INIT_WQ_LIST(&state->compl_reqs);
1459 }
1460 ctx->submit_state.cq_flush = false;
1461}
1462
1463static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1464{
1465 /* See comment at the top of this file */
1466 smp_rmb();
1467 return __io_cqring_events(ctx);
1468}
1469
1470/*
1471 * We can't just wait for polled events to come to us, we have to actively
1472 * find and complete them.
1473 */
1474static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1475{
1476 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1477 return;
1478
1479 mutex_lock(&ctx->uring_lock);
1480 while (!wq_list_empty(&ctx->iopoll_list)) {
1481 /* let it sleep and repeat later if can't complete a request */
1482 if (io_do_iopoll(ctx, true) == 0)
1483 break;
1484 /*
1485 * Ensure we allow local-to-the-cpu processing to take place,
1486 * in this case we need to ensure that we reap all events.
1487 * Also let task_work, etc. to progress by releasing the mutex
1488 */
1489 if (need_resched()) {
1490 mutex_unlock(&ctx->uring_lock);
1491 cond_resched();
1492 mutex_lock(&ctx->uring_lock);
1493 }
1494 }
1495 mutex_unlock(&ctx->uring_lock);
1496}
1497
1498static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1499{
1500 unsigned int nr_events = 0;
1501 unsigned long check_cq;
1502
1503 lockdep_assert_held(&ctx->uring_lock);
1504
1505 if (!io_allowed_run_tw(ctx))
1506 return -EEXIST;
1507
1508 check_cq = READ_ONCE(ctx->check_cq);
1509 if (unlikely(check_cq)) {
1510 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1511 __io_cqring_overflow_flush(ctx, false);
1512 /*
1513 * Similarly do not spin if we have not informed the user of any
1514 * dropped CQE.
1515 */
1516 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1517 return -EBADR;
1518 }
1519 /*
1520 * Don't enter poll loop if we already have events pending.
1521 * If we do, we can potentially be spinning for commands that
1522 * already triggered a CQE (eg in error).
1523 */
1524 if (io_cqring_events(ctx))
1525 return 0;
1526
1527 do {
1528 int ret = 0;
1529
1530 /*
1531 * If a submit got punted to a workqueue, we can have the
1532 * application entering polling for a command before it gets
1533 * issued. That app will hold the uring_lock for the duration
1534 * of the poll right here, so we need to take a breather every
1535 * now and then to ensure that the issue has a chance to add
1536 * the poll to the issued list. Otherwise we can spin here
1537 * forever, while the workqueue is stuck trying to acquire the
1538 * very same mutex.
1539 */
1540 if (wq_list_empty(&ctx->iopoll_list) ||
1541 io_task_work_pending(ctx)) {
1542 u32 tail = ctx->cached_cq_tail;
1543
1544 (void) io_run_local_work_locked(ctx, min);
1545
1546 if (task_work_pending(current) ||
1547 wq_list_empty(&ctx->iopoll_list)) {
1548 mutex_unlock(&ctx->uring_lock);
1549 io_run_task_work();
1550 mutex_lock(&ctx->uring_lock);
1551 }
1552 /* some requests don't go through iopoll_list */
1553 if (tail != ctx->cached_cq_tail ||
1554 wq_list_empty(&ctx->iopoll_list))
1555 break;
1556 }
1557 ret = io_do_iopoll(ctx, !min);
1558 if (unlikely(ret < 0))
1559 return ret;
1560
1561 if (task_sigpending(current))
1562 return -EINTR;
1563 if (need_resched())
1564 break;
1565
1566 nr_events += ret;
1567 } while (nr_events < min);
1568
1569 return 0;
1570}
1571
1572void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts)
1573{
1574 io_req_complete_defer(req);
1575}
1576
1577/*
1578 * After the iocb has been issued, it's safe to be found on the poll list.
1579 * Adding the kiocb to the list AFTER submission ensures that we don't
1580 * find it from a io_do_iopoll() thread before the issuer is done
1581 * accessing the kiocb cookie.
1582 */
1583static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1584{
1585 struct io_ring_ctx *ctx = req->ctx;
1586 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1587
1588 /* workqueue context doesn't hold uring_lock, grab it now */
1589 if (unlikely(needs_lock))
1590 mutex_lock(&ctx->uring_lock);
1591
1592 /*
1593 * Track whether we have multiple files in our lists. This will impact
1594 * how we do polling eventually, not spinning if we're on potentially
1595 * different devices.
1596 */
1597 if (wq_list_empty(&ctx->iopoll_list)) {
1598 ctx->poll_multi_queue = false;
1599 } else if (!ctx->poll_multi_queue) {
1600 struct io_kiocb *list_req;
1601
1602 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1603 comp_list);
1604 if (list_req->file != req->file)
1605 ctx->poll_multi_queue = true;
1606 }
1607
1608 /*
1609 * For fast devices, IO may have already completed. If it has, add
1610 * it to the front so we find it first.
1611 */
1612 if (READ_ONCE(req->iopoll_completed))
1613 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1614 else
1615 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1616
1617 if (unlikely(needs_lock)) {
1618 /*
1619 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1620 * in sq thread task context or in io worker task context. If
1621 * current task context is sq thread, we don't need to check
1622 * whether should wake up sq thread.
1623 */
1624 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1625 wq_has_sleeper(&ctx->sq_data->wait))
1626 wake_up(&ctx->sq_data->wait);
1627
1628 mutex_unlock(&ctx->uring_lock);
1629 }
1630}
1631
1632io_req_flags_t io_file_get_flags(struct file *file)
1633{
1634 io_req_flags_t res = 0;
1635
1636 if (S_ISREG(file_inode(file)->i_mode))
1637 res |= REQ_F_ISREG;
1638 if ((file->f_flags & O_NONBLOCK) || (file->f_mode & FMODE_NOWAIT))
1639 res |= REQ_F_SUPPORT_NOWAIT;
1640 return res;
1641}
1642
1643bool io_alloc_async_data(struct io_kiocb *req)
1644{
1645 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1646
1647 WARN_ON_ONCE(!def->async_size);
1648 req->async_data = kmalloc(def->async_size, GFP_KERNEL);
1649 if (req->async_data) {
1650 req->flags |= REQ_F_ASYNC_DATA;
1651 return false;
1652 }
1653 return true;
1654}
1655
1656static u32 io_get_sequence(struct io_kiocb *req)
1657{
1658 u32 seq = req->ctx->cached_sq_head;
1659 struct io_kiocb *cur;
1660
1661 /* need original cached_sq_head, but it was increased for each req */
1662 io_for_each_link(cur, req)
1663 seq--;
1664 return seq;
1665}
1666
1667static __cold void io_drain_req(struct io_kiocb *req)
1668 __must_hold(&ctx->uring_lock)
1669{
1670 struct io_ring_ctx *ctx = req->ctx;
1671 struct io_defer_entry *de;
1672 int ret;
1673 u32 seq = io_get_sequence(req);
1674
1675 /* Still need defer if there is pending req in defer list. */
1676 spin_lock(&ctx->completion_lock);
1677 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1678 spin_unlock(&ctx->completion_lock);
1679queue:
1680 ctx->drain_active = false;
1681 io_req_task_queue(req);
1682 return;
1683 }
1684 spin_unlock(&ctx->completion_lock);
1685
1686 io_prep_async_link(req);
1687 de = kmalloc(sizeof(*de), GFP_KERNEL);
1688 if (!de) {
1689 ret = -ENOMEM;
1690 io_req_defer_failed(req, ret);
1691 return;
1692 }
1693
1694 spin_lock(&ctx->completion_lock);
1695 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1696 spin_unlock(&ctx->completion_lock);
1697 kfree(de);
1698 goto queue;
1699 }
1700
1701 trace_io_uring_defer(req);
1702 de->req = req;
1703 de->seq = seq;
1704 list_add_tail(&de->list, &ctx->defer_list);
1705 spin_unlock(&ctx->completion_lock);
1706}
1707
1708static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def,
1709 unsigned int issue_flags)
1710{
1711 if (req->file || !def->needs_file)
1712 return true;
1713
1714 if (req->flags & REQ_F_FIXED_FILE)
1715 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1716 else
1717 req->file = io_file_get_normal(req, req->cqe.fd);
1718
1719 return !!req->file;
1720}
1721
1722static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1723{
1724 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1725 const struct cred *creds = NULL;
1726 int ret;
1727
1728 if (unlikely(!io_assign_file(req, def, issue_flags)))
1729 return -EBADF;
1730
1731 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1732 creds = override_creds(req->creds);
1733
1734 if (!def->audit_skip)
1735 audit_uring_entry(req->opcode);
1736
1737 ret = def->issue(req, issue_flags);
1738
1739 if (!def->audit_skip)
1740 audit_uring_exit(!ret, ret);
1741
1742 if (creds)
1743 revert_creds(creds);
1744
1745 if (ret == IOU_OK) {
1746 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1747 io_req_complete_defer(req);
1748 else
1749 io_req_complete_post(req, issue_flags);
1750
1751 return 0;
1752 }
1753
1754 if (ret == IOU_ISSUE_SKIP_COMPLETE) {
1755 ret = 0;
1756 io_arm_ltimeout(req);
1757
1758 /* If the op doesn't have a file, we're not polling for it */
1759 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
1760 io_iopoll_req_issued(req, issue_flags);
1761 }
1762 return ret;
1763}
1764
1765int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts)
1766{
1767 io_tw_lock(req->ctx, ts);
1768 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT|
1769 IO_URING_F_COMPLETE_DEFER);
1770}
1771
1772struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1773{
1774 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1775 struct io_kiocb *nxt = NULL;
1776
1777 if (req_ref_put_and_test(req)) {
1778 if (req->flags & IO_REQ_LINK_FLAGS)
1779 nxt = io_req_find_next(req);
1780 io_free_req(req);
1781 }
1782 return nxt ? &nxt->work : NULL;
1783}
1784
1785void io_wq_submit_work(struct io_wq_work *work)
1786{
1787 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1788 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1789 unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ;
1790 bool needs_poll = false;
1791 int ret = 0, err = -ECANCELED;
1792
1793 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1794 if (!(req->flags & REQ_F_REFCOUNT))
1795 __io_req_set_refcount(req, 2);
1796 else
1797 req_ref_get(req);
1798
1799 io_arm_ltimeout(req);
1800
1801 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1802 if (atomic_read(&work->flags) & IO_WQ_WORK_CANCEL) {
1803fail:
1804 io_req_task_queue_fail(req, err);
1805 return;
1806 }
1807 if (!io_assign_file(req, def, issue_flags)) {
1808 err = -EBADF;
1809 atomic_or(IO_WQ_WORK_CANCEL, &work->flags);
1810 goto fail;
1811 }
1812
1813 /*
1814 * If DEFER_TASKRUN is set, it's only allowed to post CQEs from the
1815 * submitter task context. Final request completions are handed to the
1816 * right context, however this is not the case of auxiliary CQEs,
1817 * which is the main mean of operation for multishot requests.
1818 * Don't allow any multishot execution from io-wq. It's more restrictive
1819 * than necessary and also cleaner.
1820 */
1821 if (req->flags & REQ_F_APOLL_MULTISHOT) {
1822 err = -EBADFD;
1823 if (!io_file_can_poll(req))
1824 goto fail;
1825 if (req->file->f_flags & O_NONBLOCK ||
1826 req->file->f_mode & FMODE_NOWAIT) {
1827 err = -ECANCELED;
1828 if (io_arm_poll_handler(req, issue_flags) != IO_APOLL_OK)
1829 goto fail;
1830 return;
1831 } else {
1832 req->flags &= ~REQ_F_APOLL_MULTISHOT;
1833 }
1834 }
1835
1836 if (req->flags & REQ_F_FORCE_ASYNC) {
1837 bool opcode_poll = def->pollin || def->pollout;
1838
1839 if (opcode_poll && io_file_can_poll(req)) {
1840 needs_poll = true;
1841 issue_flags |= IO_URING_F_NONBLOCK;
1842 }
1843 }
1844
1845 do {
1846 ret = io_issue_sqe(req, issue_flags);
1847 if (ret != -EAGAIN)
1848 break;
1849
1850 /*
1851 * If REQ_F_NOWAIT is set, then don't wait or retry with
1852 * poll. -EAGAIN is final for that case.
1853 */
1854 if (req->flags & REQ_F_NOWAIT)
1855 break;
1856
1857 /*
1858 * We can get EAGAIN for iopolled IO even though we're
1859 * forcing a sync submission from here, since we can't
1860 * wait for request slots on the block side.
1861 */
1862 if (!needs_poll) {
1863 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1864 break;
1865 if (io_wq_worker_stopped())
1866 break;
1867 cond_resched();
1868 continue;
1869 }
1870
1871 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1872 return;
1873 /* aborted or ready, in either case retry blocking */
1874 needs_poll = false;
1875 issue_flags &= ~IO_URING_F_NONBLOCK;
1876 } while (1);
1877
1878 /* avoid locking problems by failing it from a clean context */
1879 if (ret)
1880 io_req_task_queue_fail(req, ret);
1881}
1882
1883inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1884 unsigned int issue_flags)
1885{
1886 struct io_ring_ctx *ctx = req->ctx;
1887 struct io_rsrc_node *node;
1888 struct file *file = NULL;
1889
1890 io_ring_submit_lock(ctx, issue_flags);
1891 node = io_rsrc_node_lookup(&ctx->file_table.data, fd);
1892 if (node) {
1893 io_req_assign_rsrc_node(&req->file_node, node);
1894 req->flags |= io_slot_flags(node);
1895 file = io_slot_file(node);
1896 }
1897 io_ring_submit_unlock(ctx, issue_flags);
1898 return file;
1899}
1900
1901struct file *io_file_get_normal(struct io_kiocb *req, int fd)
1902{
1903 struct file *file = fget(fd);
1904
1905 trace_io_uring_file_get(req, fd);
1906
1907 /* we don't allow fixed io_uring files */
1908 if (file && io_is_uring_fops(file))
1909 io_req_track_inflight(req);
1910 return file;
1911}
1912
1913static void io_queue_async(struct io_kiocb *req, int ret)
1914 __must_hold(&req->ctx->uring_lock)
1915{
1916 struct io_kiocb *linked_timeout;
1917
1918 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
1919 io_req_defer_failed(req, ret);
1920 return;
1921 }
1922
1923 linked_timeout = io_prep_linked_timeout(req);
1924
1925 switch (io_arm_poll_handler(req, 0)) {
1926 case IO_APOLL_READY:
1927 io_kbuf_recycle(req, 0);
1928 io_req_task_queue(req);
1929 break;
1930 case IO_APOLL_ABORTED:
1931 io_kbuf_recycle(req, 0);
1932 io_queue_iowq(req);
1933 break;
1934 case IO_APOLL_OK:
1935 break;
1936 }
1937
1938 if (linked_timeout)
1939 io_queue_linked_timeout(linked_timeout);
1940}
1941
1942static inline void io_queue_sqe(struct io_kiocb *req)
1943 __must_hold(&req->ctx->uring_lock)
1944{
1945 int ret;
1946
1947 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
1948
1949 /*
1950 * We async punt it if the file wasn't marked NOWAIT, or if the file
1951 * doesn't support non-blocking read/write attempts
1952 */
1953 if (unlikely(ret))
1954 io_queue_async(req, ret);
1955}
1956
1957static void io_queue_sqe_fallback(struct io_kiocb *req)
1958 __must_hold(&req->ctx->uring_lock)
1959{
1960 if (unlikely(req->flags & REQ_F_FAIL)) {
1961 /*
1962 * We don't submit, fail them all, for that replace hardlinks
1963 * with normal links. Extra REQ_F_LINK is tolerated.
1964 */
1965 req->flags &= ~REQ_F_HARDLINK;
1966 req->flags |= REQ_F_LINK;
1967 io_req_defer_failed(req, req->cqe.res);
1968 } else {
1969 if (unlikely(req->ctx->drain_active))
1970 io_drain_req(req);
1971 else
1972 io_queue_iowq(req);
1973 }
1974}
1975
1976/*
1977 * Check SQE restrictions (opcode and flags).
1978 *
1979 * Returns 'true' if SQE is allowed, 'false' otherwise.
1980 */
1981static inline bool io_check_restriction(struct io_ring_ctx *ctx,
1982 struct io_kiocb *req,
1983 unsigned int sqe_flags)
1984{
1985 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
1986 return false;
1987
1988 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
1989 ctx->restrictions.sqe_flags_required)
1990 return false;
1991
1992 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
1993 ctx->restrictions.sqe_flags_required))
1994 return false;
1995
1996 return true;
1997}
1998
1999static void io_init_req_drain(struct io_kiocb *req)
2000{
2001 struct io_ring_ctx *ctx = req->ctx;
2002 struct io_kiocb *head = ctx->submit_state.link.head;
2003
2004 ctx->drain_active = true;
2005 if (head) {
2006 /*
2007 * If we need to drain a request in the middle of a link, drain
2008 * the head request and the next request/link after the current
2009 * link. Considering sequential execution of links,
2010 * REQ_F_IO_DRAIN will be maintained for every request of our
2011 * link.
2012 */
2013 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2014 ctx->drain_next = true;
2015 }
2016}
2017
2018static __cold int io_init_fail_req(struct io_kiocb *req, int err)
2019{
2020 /* ensure per-opcode data is cleared if we fail before prep */
2021 memset(&req->cmd.data, 0, sizeof(req->cmd.data));
2022 return err;
2023}
2024
2025static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
2026 const struct io_uring_sqe *sqe)
2027 __must_hold(&ctx->uring_lock)
2028{
2029 const struct io_issue_def *def;
2030 unsigned int sqe_flags;
2031 int personality;
2032 u8 opcode;
2033
2034 /* req is partially pre-initialised, see io_preinit_req() */
2035 req->opcode = opcode = READ_ONCE(sqe->opcode);
2036 /* same numerical values with corresponding REQ_F_*, safe to copy */
2037 sqe_flags = READ_ONCE(sqe->flags);
2038 req->flags = (__force io_req_flags_t) sqe_flags;
2039 req->cqe.user_data = READ_ONCE(sqe->user_data);
2040 req->file = NULL;
2041 req->tctx = current->io_uring;
2042 req->cancel_seq_set = false;
2043
2044 if (unlikely(opcode >= IORING_OP_LAST)) {
2045 req->opcode = 0;
2046 return io_init_fail_req(req, -EINVAL);
2047 }
2048 opcode = array_index_nospec(opcode, IORING_OP_LAST);
2049
2050 def = &io_issue_defs[opcode];
2051 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2052 /* enforce forwards compatibility on users */
2053 if (sqe_flags & ~SQE_VALID_FLAGS)
2054 return io_init_fail_req(req, -EINVAL);
2055 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2056 if (!def->buffer_select)
2057 return io_init_fail_req(req, -EOPNOTSUPP);
2058 req->buf_index = READ_ONCE(sqe->buf_group);
2059 }
2060 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2061 ctx->drain_disabled = true;
2062 if (sqe_flags & IOSQE_IO_DRAIN) {
2063 if (ctx->drain_disabled)
2064 return io_init_fail_req(req, -EOPNOTSUPP);
2065 io_init_req_drain(req);
2066 }
2067 }
2068 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2069 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2070 return io_init_fail_req(req, -EACCES);
2071 /* knock it to the slow queue path, will be drained there */
2072 if (ctx->drain_active)
2073 req->flags |= REQ_F_FORCE_ASYNC;
2074 /* if there is no link, we're at "next" request and need to drain */
2075 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2076 ctx->drain_next = false;
2077 ctx->drain_active = true;
2078 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2079 }
2080 }
2081
2082 if (!def->ioprio && sqe->ioprio)
2083 return io_init_fail_req(req, -EINVAL);
2084 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2085 return io_init_fail_req(req, -EINVAL);
2086
2087 if (def->needs_file) {
2088 struct io_submit_state *state = &ctx->submit_state;
2089
2090 req->cqe.fd = READ_ONCE(sqe->fd);
2091
2092 /*
2093 * Plug now if we have more than 2 IO left after this, and the
2094 * target is potentially a read/write to block based storage.
2095 */
2096 if (state->need_plug && def->plug) {
2097 state->plug_started = true;
2098 state->need_plug = false;
2099 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2100 }
2101 }
2102
2103 personality = READ_ONCE(sqe->personality);
2104 if (personality) {
2105 int ret;
2106
2107 req->creds = xa_load(&ctx->personalities, personality);
2108 if (!req->creds)
2109 return io_init_fail_req(req, -EINVAL);
2110 get_cred(req->creds);
2111 ret = security_uring_override_creds(req->creds);
2112 if (ret) {
2113 put_cred(req->creds);
2114 return io_init_fail_req(req, ret);
2115 }
2116 req->flags |= REQ_F_CREDS;
2117 }
2118
2119 return def->prep(req, sqe);
2120}
2121
2122static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2123 struct io_kiocb *req, int ret)
2124{
2125 struct io_ring_ctx *ctx = req->ctx;
2126 struct io_submit_link *link = &ctx->submit_state.link;
2127 struct io_kiocb *head = link->head;
2128
2129 trace_io_uring_req_failed(sqe, req, ret);
2130
2131 /*
2132 * Avoid breaking links in the middle as it renders links with SQPOLL
2133 * unusable. Instead of failing eagerly, continue assembling the link if
2134 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2135 * should find the flag and handle the rest.
2136 */
2137 req_fail_link_node(req, ret);
2138 if (head && !(head->flags & REQ_F_FAIL))
2139 req_fail_link_node(head, -ECANCELED);
2140
2141 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2142 if (head) {
2143 link->last->link = req;
2144 link->head = NULL;
2145 req = head;
2146 }
2147 io_queue_sqe_fallback(req);
2148 return ret;
2149 }
2150
2151 if (head)
2152 link->last->link = req;
2153 else
2154 link->head = req;
2155 link->last = req;
2156 return 0;
2157}
2158
2159static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2160 const struct io_uring_sqe *sqe)
2161 __must_hold(&ctx->uring_lock)
2162{
2163 struct io_submit_link *link = &ctx->submit_state.link;
2164 int ret;
2165
2166 ret = io_init_req(ctx, req, sqe);
2167 if (unlikely(ret))
2168 return io_submit_fail_init(sqe, req, ret);
2169
2170 trace_io_uring_submit_req(req);
2171
2172 /*
2173 * If we already have a head request, queue this one for async
2174 * submittal once the head completes. If we don't have a head but
2175 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2176 * submitted sync once the chain is complete. If none of those
2177 * conditions are true (normal request), then just queue it.
2178 */
2179 if (unlikely(link->head)) {
2180 trace_io_uring_link(req, link->last);
2181 link->last->link = req;
2182 link->last = req;
2183
2184 if (req->flags & IO_REQ_LINK_FLAGS)
2185 return 0;
2186 /* last request of the link, flush it */
2187 req = link->head;
2188 link->head = NULL;
2189 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2190 goto fallback;
2191
2192 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2193 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2194 if (req->flags & IO_REQ_LINK_FLAGS) {
2195 link->head = req;
2196 link->last = req;
2197 } else {
2198fallback:
2199 io_queue_sqe_fallback(req);
2200 }
2201 return 0;
2202 }
2203
2204 io_queue_sqe(req);
2205 return 0;
2206}
2207
2208/*
2209 * Batched submission is done, ensure local IO is flushed out.
2210 */
2211static void io_submit_state_end(struct io_ring_ctx *ctx)
2212{
2213 struct io_submit_state *state = &ctx->submit_state;
2214
2215 if (unlikely(state->link.head))
2216 io_queue_sqe_fallback(state->link.head);
2217 /* flush only after queuing links as they can generate completions */
2218 io_submit_flush_completions(ctx);
2219 if (state->plug_started)
2220 blk_finish_plug(&state->plug);
2221}
2222
2223/*
2224 * Start submission side cache.
2225 */
2226static void io_submit_state_start(struct io_submit_state *state,
2227 unsigned int max_ios)
2228{
2229 state->plug_started = false;
2230 state->need_plug = max_ios > 2;
2231 state->submit_nr = max_ios;
2232 /* set only head, no need to init link_last in advance */
2233 state->link.head = NULL;
2234}
2235
2236static void io_commit_sqring(struct io_ring_ctx *ctx)
2237{
2238 struct io_rings *rings = ctx->rings;
2239
2240 /*
2241 * Ensure any loads from the SQEs are done at this point,
2242 * since once we write the new head, the application could
2243 * write new data to them.
2244 */
2245 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2246}
2247
2248/*
2249 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2250 * that is mapped by userspace. This means that care needs to be taken to
2251 * ensure that reads are stable, as we cannot rely on userspace always
2252 * being a good citizen. If members of the sqe are validated and then later
2253 * used, it's important that those reads are done through READ_ONCE() to
2254 * prevent a re-load down the line.
2255 */
2256static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe)
2257{
2258 unsigned mask = ctx->sq_entries - 1;
2259 unsigned head = ctx->cached_sq_head++ & mask;
2260
2261 if (static_branch_unlikely(&io_key_has_sqarray) &&
2262 (!(ctx->flags & IORING_SETUP_NO_SQARRAY))) {
2263 head = READ_ONCE(ctx->sq_array[head]);
2264 if (unlikely(head >= ctx->sq_entries)) {
2265 /* drop invalid entries */
2266 spin_lock(&ctx->completion_lock);
2267 ctx->cq_extra--;
2268 spin_unlock(&ctx->completion_lock);
2269 WRITE_ONCE(ctx->rings->sq_dropped,
2270 READ_ONCE(ctx->rings->sq_dropped) + 1);
2271 return false;
2272 }
2273 head = array_index_nospec(head, ctx->sq_entries);
2274 }
2275
2276 /*
2277 * The cached sq head (or cq tail) serves two purposes:
2278 *
2279 * 1) allows us to batch the cost of updating the user visible
2280 * head updates.
2281 * 2) allows the kernel side to track the head on its own, even
2282 * though the application is the one updating it.
2283 */
2284
2285 /* double index for 128-byte SQEs, twice as long */
2286 if (ctx->flags & IORING_SETUP_SQE128)
2287 head <<= 1;
2288 *sqe = &ctx->sq_sqes[head];
2289 return true;
2290}
2291
2292int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2293 __must_hold(&ctx->uring_lock)
2294{
2295 unsigned int entries = io_sqring_entries(ctx);
2296 unsigned int left;
2297 int ret;
2298
2299 if (unlikely(!entries))
2300 return 0;
2301 /* make sure SQ entry isn't read before tail */
2302 ret = left = min(nr, entries);
2303 io_get_task_refs(left);
2304 io_submit_state_start(&ctx->submit_state, left);
2305
2306 do {
2307 const struct io_uring_sqe *sqe;
2308 struct io_kiocb *req;
2309
2310 if (unlikely(!io_alloc_req(ctx, &req)))
2311 break;
2312 if (unlikely(!io_get_sqe(ctx, &sqe))) {
2313 io_req_add_to_cache(req, ctx);
2314 break;
2315 }
2316
2317 /*
2318 * Continue submitting even for sqe failure if the
2319 * ring was setup with IORING_SETUP_SUBMIT_ALL
2320 */
2321 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2322 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2323 left--;
2324 break;
2325 }
2326 } while (--left);
2327
2328 if (unlikely(left)) {
2329 ret -= left;
2330 /* try again if it submitted nothing and can't allocate a req */
2331 if (!ret && io_req_cache_empty(ctx))
2332 ret = -EAGAIN;
2333 current->io_uring->cached_refs += left;
2334 }
2335
2336 io_submit_state_end(ctx);
2337 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2338 io_commit_sqring(ctx);
2339 return ret;
2340}
2341
2342static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2343 int wake_flags, void *key)
2344{
2345 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq);
2346
2347 /*
2348 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2349 * the task, and the next invocation will do it.
2350 */
2351 if (io_should_wake(iowq) || io_has_work(iowq->ctx))
2352 return autoremove_wake_function(curr, mode, wake_flags, key);
2353 return -1;
2354}
2355
2356int io_run_task_work_sig(struct io_ring_ctx *ctx)
2357{
2358 if (io_local_work_pending(ctx)) {
2359 __set_current_state(TASK_RUNNING);
2360 if (io_run_local_work(ctx, INT_MAX, IO_LOCAL_TW_DEFAULT_MAX) > 0)
2361 return 0;
2362 }
2363 if (io_run_task_work() > 0)
2364 return 0;
2365 if (task_sigpending(current))
2366 return -EINTR;
2367 return 0;
2368}
2369
2370static bool current_pending_io(void)
2371{
2372 struct io_uring_task *tctx = current->io_uring;
2373
2374 if (!tctx)
2375 return false;
2376 return percpu_counter_read_positive(&tctx->inflight);
2377}
2378
2379static enum hrtimer_restart io_cqring_timer_wakeup(struct hrtimer *timer)
2380{
2381 struct io_wait_queue *iowq = container_of(timer, struct io_wait_queue, t);
2382
2383 WRITE_ONCE(iowq->hit_timeout, 1);
2384 iowq->min_timeout = 0;
2385 wake_up_process(iowq->wq.private);
2386 return HRTIMER_NORESTART;
2387}
2388
2389/*
2390 * Doing min_timeout portion. If we saw any timeouts, events, or have work,
2391 * wake up. If not, and we have a normal timeout, switch to that and keep
2392 * sleeping.
2393 */
2394static enum hrtimer_restart io_cqring_min_timer_wakeup(struct hrtimer *timer)
2395{
2396 struct io_wait_queue *iowq = container_of(timer, struct io_wait_queue, t);
2397 struct io_ring_ctx *ctx = iowq->ctx;
2398
2399 /* no general timeout, or shorter (or equal), we are done */
2400 if (iowq->timeout == KTIME_MAX ||
2401 ktime_compare(iowq->min_timeout, iowq->timeout) >= 0)
2402 goto out_wake;
2403 /* work we may need to run, wake function will see if we need to wake */
2404 if (io_has_work(ctx))
2405 goto out_wake;
2406 /* got events since we started waiting, min timeout is done */
2407 if (iowq->cq_min_tail != READ_ONCE(ctx->rings->cq.tail))
2408 goto out_wake;
2409 /* if we have any events and min timeout expired, we're done */
2410 if (io_cqring_events(ctx))
2411 goto out_wake;
2412
2413 /*
2414 * If using deferred task_work running and application is waiting on
2415 * more than one request, ensure we reset it now where we are switching
2416 * to normal sleeps. Any request completion post min_wait should wake
2417 * the task and return.
2418 */
2419 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
2420 atomic_set(&ctx->cq_wait_nr, 1);
2421 smp_mb();
2422 if (!llist_empty(&ctx->work_llist))
2423 goto out_wake;
2424 }
2425
2426 iowq->t.function = io_cqring_timer_wakeup;
2427 hrtimer_set_expires(timer, iowq->timeout);
2428 return HRTIMER_RESTART;
2429out_wake:
2430 return io_cqring_timer_wakeup(timer);
2431}
2432
2433static int io_cqring_schedule_timeout(struct io_wait_queue *iowq,
2434 clockid_t clock_id, ktime_t start_time)
2435{
2436 ktime_t timeout;
2437
2438 if (iowq->min_timeout) {
2439 timeout = ktime_add_ns(iowq->min_timeout, start_time);
2440 hrtimer_setup_on_stack(&iowq->t, io_cqring_min_timer_wakeup, clock_id,
2441 HRTIMER_MODE_ABS);
2442 } else {
2443 timeout = iowq->timeout;
2444 hrtimer_setup_on_stack(&iowq->t, io_cqring_timer_wakeup, clock_id,
2445 HRTIMER_MODE_ABS);
2446 }
2447
2448 hrtimer_set_expires_range_ns(&iowq->t, timeout, 0);
2449 hrtimer_start_expires(&iowq->t, HRTIMER_MODE_ABS);
2450
2451 if (!READ_ONCE(iowq->hit_timeout))
2452 schedule();
2453
2454 hrtimer_cancel(&iowq->t);
2455 destroy_hrtimer_on_stack(&iowq->t);
2456 __set_current_state(TASK_RUNNING);
2457
2458 return READ_ONCE(iowq->hit_timeout) ? -ETIME : 0;
2459}
2460
2461static int __io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2462 struct io_wait_queue *iowq,
2463 ktime_t start_time)
2464{
2465 int ret = 0;
2466
2467 /*
2468 * Mark us as being in io_wait if we have pending requests, so cpufreq
2469 * can take into account that the task is waiting for IO - turns out
2470 * to be important for low QD IO.
2471 */
2472 if (current_pending_io())
2473 current->in_iowait = 1;
2474 if (iowq->timeout != KTIME_MAX || iowq->min_timeout)
2475 ret = io_cqring_schedule_timeout(iowq, ctx->clockid, start_time);
2476 else
2477 schedule();
2478 current->in_iowait = 0;
2479 return ret;
2480}
2481
2482/* If this returns > 0, the caller should retry */
2483static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2484 struct io_wait_queue *iowq,
2485 ktime_t start_time)
2486{
2487 if (unlikely(READ_ONCE(ctx->check_cq)))
2488 return 1;
2489 if (unlikely(io_local_work_pending(ctx)))
2490 return 1;
2491 if (unlikely(task_work_pending(current)))
2492 return 1;
2493 if (unlikely(task_sigpending(current)))
2494 return -EINTR;
2495 if (unlikely(io_should_wake(iowq)))
2496 return 0;
2497
2498 return __io_cqring_wait_schedule(ctx, iowq, start_time);
2499}
2500
2501struct ext_arg {
2502 size_t argsz;
2503 struct timespec64 ts;
2504 const sigset_t __user *sig;
2505 ktime_t min_time;
2506 bool ts_set;
2507};
2508
2509/*
2510 * Wait until events become available, if we don't already have some. The
2511 * application must reap them itself, as they reside on the shared cq ring.
2512 */
2513static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, u32 flags,
2514 struct ext_arg *ext_arg)
2515{
2516 struct io_wait_queue iowq;
2517 struct io_rings *rings = ctx->rings;
2518 ktime_t start_time;
2519 int ret;
2520
2521 if (!io_allowed_run_tw(ctx))
2522 return -EEXIST;
2523 if (io_local_work_pending(ctx))
2524 io_run_local_work(ctx, min_events,
2525 max(IO_LOCAL_TW_DEFAULT_MAX, min_events));
2526 io_run_task_work();
2527
2528 if (unlikely(test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)))
2529 io_cqring_do_overflow_flush(ctx);
2530 if (__io_cqring_events_user(ctx) >= min_events)
2531 return 0;
2532
2533 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2534 iowq.wq.private = current;
2535 INIT_LIST_HEAD(&iowq.wq.entry);
2536 iowq.ctx = ctx;
2537 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2538 iowq.cq_min_tail = READ_ONCE(ctx->rings->cq.tail);
2539 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2540 iowq.hit_timeout = 0;
2541 iowq.min_timeout = ext_arg->min_time;
2542 iowq.timeout = KTIME_MAX;
2543 start_time = io_get_time(ctx);
2544
2545 if (ext_arg->ts_set) {
2546 iowq.timeout = timespec64_to_ktime(ext_arg->ts);
2547 if (!(flags & IORING_ENTER_ABS_TIMER))
2548 iowq.timeout = ktime_add(iowq.timeout, start_time);
2549 }
2550
2551 if (ext_arg->sig) {
2552#ifdef CONFIG_COMPAT
2553 if (in_compat_syscall())
2554 ret = set_compat_user_sigmask((const compat_sigset_t __user *)ext_arg->sig,
2555 ext_arg->argsz);
2556 else
2557#endif
2558 ret = set_user_sigmask(ext_arg->sig, ext_arg->argsz);
2559
2560 if (ret)
2561 return ret;
2562 }
2563
2564 io_napi_busy_loop(ctx, &iowq);
2565
2566 trace_io_uring_cqring_wait(ctx, min_events);
2567 do {
2568 unsigned long check_cq;
2569 int nr_wait;
2570
2571 /* if min timeout has been hit, don't reset wait count */
2572 if (!iowq.hit_timeout)
2573 nr_wait = (int) iowq.cq_tail -
2574 READ_ONCE(ctx->rings->cq.tail);
2575 else
2576 nr_wait = 1;
2577
2578 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
2579 atomic_set(&ctx->cq_wait_nr, nr_wait);
2580 set_current_state(TASK_INTERRUPTIBLE);
2581 } else {
2582 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2583 TASK_INTERRUPTIBLE);
2584 }
2585
2586 ret = io_cqring_wait_schedule(ctx, &iowq, start_time);
2587 __set_current_state(TASK_RUNNING);
2588 atomic_set(&ctx->cq_wait_nr, IO_CQ_WAKE_INIT);
2589
2590 /*
2591 * Run task_work after scheduling and before io_should_wake().
2592 * If we got woken because of task_work being processed, run it
2593 * now rather than let the caller do another wait loop.
2594 */
2595 if (io_local_work_pending(ctx))
2596 io_run_local_work(ctx, nr_wait, nr_wait);
2597 io_run_task_work();
2598
2599 /*
2600 * Non-local task_work will be run on exit to userspace, but
2601 * if we're using DEFER_TASKRUN, then we could have waited
2602 * with a timeout for a number of requests. If the timeout
2603 * hits, we could have some requests ready to process. Ensure
2604 * this break is _after_ we have run task_work, to avoid
2605 * deferring running potentially pending requests until the
2606 * next time we wait for events.
2607 */
2608 if (ret < 0)
2609 break;
2610
2611 check_cq = READ_ONCE(ctx->check_cq);
2612 if (unlikely(check_cq)) {
2613 /* let the caller flush overflows, retry */
2614 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2615 io_cqring_do_overflow_flush(ctx);
2616 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) {
2617 ret = -EBADR;
2618 break;
2619 }
2620 }
2621
2622 if (io_should_wake(&iowq)) {
2623 ret = 0;
2624 break;
2625 }
2626 cond_resched();
2627 } while (1);
2628
2629 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
2630 finish_wait(&ctx->cq_wait, &iowq.wq);
2631 restore_saved_sigmask_unless(ret == -EINTR);
2632
2633 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2634}
2635
2636static void *io_rings_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2637 size_t size)
2638{
2639 return __io_uaddr_map(&ctx->ring_pages, &ctx->n_ring_pages, uaddr,
2640 size);
2641}
2642
2643static void *io_sqes_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2644 size_t size)
2645{
2646 return __io_uaddr_map(&ctx->sqe_pages, &ctx->n_sqe_pages, uaddr,
2647 size);
2648}
2649
2650static void io_rings_free(struct io_ring_ctx *ctx)
2651{
2652 if (!(ctx->flags & IORING_SETUP_NO_MMAP)) {
2653 io_pages_unmap(ctx->rings, &ctx->ring_pages, &ctx->n_ring_pages,
2654 true);
2655 io_pages_unmap(ctx->sq_sqes, &ctx->sqe_pages, &ctx->n_sqe_pages,
2656 true);
2657 } else {
2658 io_pages_free(&ctx->ring_pages, ctx->n_ring_pages);
2659 ctx->n_ring_pages = 0;
2660 io_pages_free(&ctx->sqe_pages, ctx->n_sqe_pages);
2661 ctx->n_sqe_pages = 0;
2662 vunmap(ctx->rings);
2663 vunmap(ctx->sq_sqes);
2664 }
2665
2666 ctx->rings = NULL;
2667 ctx->sq_sqes = NULL;
2668}
2669
2670unsigned long rings_size(unsigned int flags, unsigned int sq_entries,
2671 unsigned int cq_entries, size_t *sq_offset)
2672{
2673 struct io_rings *rings;
2674 size_t off, sq_array_size;
2675
2676 off = struct_size(rings, cqes, cq_entries);
2677 if (off == SIZE_MAX)
2678 return SIZE_MAX;
2679 if (flags & IORING_SETUP_CQE32) {
2680 if (check_shl_overflow(off, 1, &off))
2681 return SIZE_MAX;
2682 }
2683
2684#ifdef CONFIG_SMP
2685 off = ALIGN(off, SMP_CACHE_BYTES);
2686 if (off == 0)
2687 return SIZE_MAX;
2688#endif
2689
2690 if (flags & IORING_SETUP_NO_SQARRAY) {
2691 *sq_offset = SIZE_MAX;
2692 return off;
2693 }
2694
2695 *sq_offset = off;
2696
2697 sq_array_size = array_size(sizeof(u32), sq_entries);
2698 if (sq_array_size == SIZE_MAX)
2699 return SIZE_MAX;
2700
2701 if (check_add_overflow(off, sq_array_size, &off))
2702 return SIZE_MAX;
2703
2704 return off;
2705}
2706
2707static void io_req_caches_free(struct io_ring_ctx *ctx)
2708{
2709 struct io_kiocb *req;
2710 int nr = 0;
2711
2712 mutex_lock(&ctx->uring_lock);
2713
2714 while (!io_req_cache_empty(ctx)) {
2715 req = io_extract_req(ctx);
2716 kmem_cache_free(req_cachep, req);
2717 nr++;
2718 }
2719 if (nr)
2720 percpu_ref_put_many(&ctx->refs, nr);
2721 mutex_unlock(&ctx->uring_lock);
2722}
2723
2724static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2725{
2726 io_sq_thread_finish(ctx);
2727
2728 mutex_lock(&ctx->uring_lock);
2729 io_sqe_buffers_unregister(ctx);
2730 io_sqe_files_unregister(ctx);
2731 io_cqring_overflow_kill(ctx);
2732 io_eventfd_unregister(ctx);
2733 io_alloc_cache_free(&ctx->apoll_cache, kfree);
2734 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2735 io_alloc_cache_free(&ctx->rw_cache, io_rw_cache_free);
2736 io_alloc_cache_free(&ctx->uring_cache, kfree);
2737 io_alloc_cache_free(&ctx->msg_cache, io_msg_cache_free);
2738 io_futex_cache_free(ctx);
2739 io_destroy_buffers(ctx);
2740 io_free_region(ctx, &ctx->param_region);
2741 mutex_unlock(&ctx->uring_lock);
2742 if (ctx->sq_creds)
2743 put_cred(ctx->sq_creds);
2744 if (ctx->submitter_task)
2745 put_task_struct(ctx->submitter_task);
2746
2747 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2748
2749 if (ctx->mm_account) {
2750 mmdrop(ctx->mm_account);
2751 ctx->mm_account = NULL;
2752 }
2753 io_rings_free(ctx);
2754
2755 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY))
2756 static_branch_dec(&io_key_has_sqarray);
2757
2758 percpu_ref_exit(&ctx->refs);
2759 free_uid(ctx->user);
2760 io_req_caches_free(ctx);
2761 if (ctx->hash_map)
2762 io_wq_put_hash(ctx->hash_map);
2763 io_napi_free(ctx);
2764 kvfree(ctx->cancel_table.hbs);
2765 xa_destroy(&ctx->io_bl_xa);
2766 kfree(ctx);
2767}
2768
2769static __cold void io_activate_pollwq_cb(struct callback_head *cb)
2770{
2771 struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx,
2772 poll_wq_task_work);
2773
2774 mutex_lock(&ctx->uring_lock);
2775 ctx->poll_activated = true;
2776 mutex_unlock(&ctx->uring_lock);
2777
2778 /*
2779 * Wake ups for some events between start of polling and activation
2780 * might've been lost due to loose synchronisation.
2781 */
2782 wake_up_all(&ctx->poll_wq);
2783 percpu_ref_put(&ctx->refs);
2784}
2785
2786__cold void io_activate_pollwq(struct io_ring_ctx *ctx)
2787{
2788 spin_lock(&ctx->completion_lock);
2789 /* already activated or in progress */
2790 if (ctx->poll_activated || ctx->poll_wq_task_work.func)
2791 goto out;
2792 if (WARN_ON_ONCE(!ctx->task_complete))
2793 goto out;
2794 if (!ctx->submitter_task)
2795 goto out;
2796 /*
2797 * with ->submitter_task only the submitter task completes requests, we
2798 * only need to sync with it, which is done by injecting a tw
2799 */
2800 init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb);
2801 percpu_ref_get(&ctx->refs);
2802 if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL))
2803 percpu_ref_put(&ctx->refs);
2804out:
2805 spin_unlock(&ctx->completion_lock);
2806}
2807
2808static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2809{
2810 struct io_ring_ctx *ctx = file->private_data;
2811 __poll_t mask = 0;
2812
2813 if (unlikely(!ctx->poll_activated))
2814 io_activate_pollwq(ctx);
2815 /*
2816 * provides mb() which pairs with barrier from wq_has_sleeper
2817 * call in io_commit_cqring
2818 */
2819 poll_wait(file, &ctx->poll_wq, wait);
2820
2821 if (!io_sqring_full(ctx))
2822 mask |= EPOLLOUT | EPOLLWRNORM;
2823
2824 /*
2825 * Don't flush cqring overflow list here, just do a simple check.
2826 * Otherwise there could possible be ABBA deadlock:
2827 * CPU0 CPU1
2828 * ---- ----
2829 * lock(&ctx->uring_lock);
2830 * lock(&ep->mtx);
2831 * lock(&ctx->uring_lock);
2832 * lock(&ep->mtx);
2833 *
2834 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2835 * pushes them to do the flush.
2836 */
2837
2838 if (__io_cqring_events_user(ctx) || io_has_work(ctx))
2839 mask |= EPOLLIN | EPOLLRDNORM;
2840
2841 return mask;
2842}
2843
2844struct io_tctx_exit {
2845 struct callback_head task_work;
2846 struct completion completion;
2847 struct io_ring_ctx *ctx;
2848};
2849
2850static __cold void io_tctx_exit_cb(struct callback_head *cb)
2851{
2852 struct io_uring_task *tctx = current->io_uring;
2853 struct io_tctx_exit *work;
2854
2855 work = container_of(cb, struct io_tctx_exit, task_work);
2856 /*
2857 * When @in_cancel, we're in cancellation and it's racy to remove the
2858 * node. It'll be removed by the end of cancellation, just ignore it.
2859 * tctx can be NULL if the queueing of this task_work raced with
2860 * work cancelation off the exec path.
2861 */
2862 if (tctx && !atomic_read(&tctx->in_cancel))
2863 io_uring_del_tctx_node((unsigned long)work->ctx);
2864 complete(&work->completion);
2865}
2866
2867static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2868{
2869 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2870
2871 return req->ctx == data;
2872}
2873
2874static __cold void io_ring_exit_work(struct work_struct *work)
2875{
2876 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2877 unsigned long timeout = jiffies + HZ * 60 * 5;
2878 unsigned long interval = HZ / 20;
2879 struct io_tctx_exit exit;
2880 struct io_tctx_node *node;
2881 int ret;
2882
2883 /*
2884 * If we're doing polled IO and end up having requests being
2885 * submitted async (out-of-line), then completions can come in while
2886 * we're waiting for refs to drop. We need to reap these manually,
2887 * as nobody else will be looking for them.
2888 */
2889 do {
2890 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
2891 mutex_lock(&ctx->uring_lock);
2892 io_cqring_overflow_kill(ctx);
2893 mutex_unlock(&ctx->uring_lock);
2894 }
2895
2896 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
2897 io_move_task_work_from_local(ctx);
2898
2899 while (io_uring_try_cancel_requests(ctx, NULL, true))
2900 cond_resched();
2901
2902 if (ctx->sq_data) {
2903 struct io_sq_data *sqd = ctx->sq_data;
2904 struct task_struct *tsk;
2905
2906 io_sq_thread_park(sqd);
2907 tsk = sqd->thread;
2908 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
2909 io_wq_cancel_cb(tsk->io_uring->io_wq,
2910 io_cancel_ctx_cb, ctx, true);
2911 io_sq_thread_unpark(sqd);
2912 }
2913
2914 io_req_caches_free(ctx);
2915
2916 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
2917 /* there is little hope left, don't run it too often */
2918 interval = HZ * 60;
2919 }
2920 /*
2921 * This is really an uninterruptible wait, as it has to be
2922 * complete. But it's also run from a kworker, which doesn't
2923 * take signals, so it's fine to make it interruptible. This
2924 * avoids scenarios where we knowingly can wait much longer
2925 * on completions, for example if someone does a SIGSTOP on
2926 * a task that needs to finish task_work to make this loop
2927 * complete. That's a synthetic situation that should not
2928 * cause a stuck task backtrace, and hence a potential panic
2929 * on stuck tasks if that is enabled.
2930 */
2931 } while (!wait_for_completion_interruptible_timeout(&ctx->ref_comp, interval));
2932
2933 init_completion(&exit.completion);
2934 init_task_work(&exit.task_work, io_tctx_exit_cb);
2935 exit.ctx = ctx;
2936
2937 mutex_lock(&ctx->uring_lock);
2938 while (!list_empty(&ctx->tctx_list)) {
2939 WARN_ON_ONCE(time_after(jiffies, timeout));
2940
2941 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
2942 ctx_node);
2943 /* don't spin on a single task if cancellation failed */
2944 list_rotate_left(&ctx->tctx_list);
2945 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
2946 if (WARN_ON_ONCE(ret))
2947 continue;
2948
2949 mutex_unlock(&ctx->uring_lock);
2950 /*
2951 * See comment above for
2952 * wait_for_completion_interruptible_timeout() on why this
2953 * wait is marked as interruptible.
2954 */
2955 wait_for_completion_interruptible(&exit.completion);
2956 mutex_lock(&ctx->uring_lock);
2957 }
2958 mutex_unlock(&ctx->uring_lock);
2959 spin_lock(&ctx->completion_lock);
2960 spin_unlock(&ctx->completion_lock);
2961
2962 /* pairs with RCU read section in io_req_local_work_add() */
2963 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
2964 synchronize_rcu();
2965
2966 io_ring_ctx_free(ctx);
2967}
2968
2969static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
2970{
2971 unsigned long index;
2972 struct creds *creds;
2973
2974 mutex_lock(&ctx->uring_lock);
2975 percpu_ref_kill(&ctx->refs);
2976 xa_for_each(&ctx->personalities, index, creds)
2977 io_unregister_personality(ctx, index);
2978 mutex_unlock(&ctx->uring_lock);
2979
2980 flush_delayed_work(&ctx->fallback_work);
2981
2982 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
2983 /*
2984 * Use system_unbound_wq to avoid spawning tons of event kworkers
2985 * if we're exiting a ton of rings at the same time. It just adds
2986 * noise and overhead, there's no discernable change in runtime
2987 * over using system_wq.
2988 */
2989 queue_work(iou_wq, &ctx->exit_work);
2990}
2991
2992static int io_uring_release(struct inode *inode, struct file *file)
2993{
2994 struct io_ring_ctx *ctx = file->private_data;
2995
2996 file->private_data = NULL;
2997 io_ring_ctx_wait_and_kill(ctx);
2998 return 0;
2999}
3000
3001struct io_task_cancel {
3002 struct io_uring_task *tctx;
3003 bool all;
3004};
3005
3006static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
3007{
3008 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3009 struct io_task_cancel *cancel = data;
3010
3011 return io_match_task_safe(req, cancel->tctx, cancel->all);
3012}
3013
3014static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
3015 struct io_uring_task *tctx,
3016 bool cancel_all)
3017{
3018 struct io_defer_entry *de;
3019 LIST_HEAD(list);
3020
3021 spin_lock(&ctx->completion_lock);
3022 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
3023 if (io_match_task_safe(de->req, tctx, cancel_all)) {
3024 list_cut_position(&list, &ctx->defer_list, &de->list);
3025 break;
3026 }
3027 }
3028 spin_unlock(&ctx->completion_lock);
3029 if (list_empty(&list))
3030 return false;
3031
3032 while (!list_empty(&list)) {
3033 de = list_first_entry(&list, struct io_defer_entry, list);
3034 list_del_init(&de->list);
3035 io_req_task_queue_fail(de->req, -ECANCELED);
3036 kfree(de);
3037 }
3038 return true;
3039}
3040
3041static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
3042{
3043 struct io_tctx_node *node;
3044 enum io_wq_cancel cret;
3045 bool ret = false;
3046
3047 mutex_lock(&ctx->uring_lock);
3048 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3049 struct io_uring_task *tctx = node->task->io_uring;
3050
3051 /*
3052 * io_wq will stay alive while we hold uring_lock, because it's
3053 * killed after ctx nodes, which requires to take the lock.
3054 */
3055 if (!tctx || !tctx->io_wq)
3056 continue;
3057 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
3058 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3059 }
3060 mutex_unlock(&ctx->uring_lock);
3061
3062 return ret;
3063}
3064
3065static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
3066 struct io_uring_task *tctx,
3067 bool cancel_all)
3068{
3069 struct io_task_cancel cancel = { .tctx = tctx, .all = cancel_all, };
3070 enum io_wq_cancel cret;
3071 bool ret = false;
3072
3073 /* set it so io_req_local_work_add() would wake us up */
3074 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
3075 atomic_set(&ctx->cq_wait_nr, 1);
3076 smp_mb();
3077 }
3078
3079 /* failed during ring init, it couldn't have issued any requests */
3080 if (!ctx->rings)
3081 return false;
3082
3083 if (!tctx) {
3084 ret |= io_uring_try_cancel_iowq(ctx);
3085 } else if (tctx->io_wq) {
3086 /*
3087 * Cancels requests of all rings, not only @ctx, but
3088 * it's fine as the task is in exit/exec.
3089 */
3090 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
3091 &cancel, true);
3092 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3093 }
3094
3095 /* SQPOLL thread does its own polling */
3096 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
3097 (ctx->sq_data && ctx->sq_data->thread == current)) {
3098 while (!wq_list_empty(&ctx->iopoll_list)) {
3099 io_iopoll_try_reap_events(ctx);
3100 ret = true;
3101 cond_resched();
3102 }
3103 }
3104
3105 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3106 io_allowed_defer_tw_run(ctx))
3107 ret |= io_run_local_work(ctx, INT_MAX, INT_MAX) > 0;
3108 ret |= io_cancel_defer_files(ctx, tctx, cancel_all);
3109 mutex_lock(&ctx->uring_lock);
3110 ret |= io_poll_remove_all(ctx, tctx, cancel_all);
3111 ret |= io_waitid_remove_all(ctx, tctx, cancel_all);
3112 ret |= io_futex_remove_all(ctx, tctx, cancel_all);
3113 ret |= io_uring_try_cancel_uring_cmd(ctx, tctx, cancel_all);
3114 mutex_unlock(&ctx->uring_lock);
3115 ret |= io_kill_timeouts(ctx, tctx, cancel_all);
3116 if (tctx)
3117 ret |= io_run_task_work() > 0;
3118 else
3119 ret |= flush_delayed_work(&ctx->fallback_work);
3120 return ret;
3121}
3122
3123static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
3124{
3125 if (tracked)
3126 return atomic_read(&tctx->inflight_tracked);
3127 return percpu_counter_sum(&tctx->inflight);
3128}
3129
3130/*
3131 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3132 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3133 */
3134__cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
3135{
3136 struct io_uring_task *tctx = current->io_uring;
3137 struct io_ring_ctx *ctx;
3138 struct io_tctx_node *node;
3139 unsigned long index;
3140 s64 inflight;
3141 DEFINE_WAIT(wait);
3142
3143 WARN_ON_ONCE(sqd && sqd->thread != current);
3144
3145 if (!current->io_uring)
3146 return;
3147 if (tctx->io_wq)
3148 io_wq_exit_start(tctx->io_wq);
3149
3150 atomic_inc(&tctx->in_cancel);
3151 do {
3152 bool loop = false;
3153
3154 io_uring_drop_tctx_refs(current);
3155 if (!tctx_inflight(tctx, !cancel_all))
3156 break;
3157
3158 /* read completions before cancelations */
3159 inflight = tctx_inflight(tctx, false);
3160 if (!inflight)
3161 break;
3162
3163 if (!sqd) {
3164 xa_for_each(&tctx->xa, index, node) {
3165 /* sqpoll task will cancel all its requests */
3166 if (node->ctx->sq_data)
3167 continue;
3168 loop |= io_uring_try_cancel_requests(node->ctx,
3169 current->io_uring,
3170 cancel_all);
3171 }
3172 } else {
3173 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3174 loop |= io_uring_try_cancel_requests(ctx,
3175 current->io_uring,
3176 cancel_all);
3177 }
3178
3179 if (loop) {
3180 cond_resched();
3181 continue;
3182 }
3183
3184 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3185 io_run_task_work();
3186 io_uring_drop_tctx_refs(current);
3187 xa_for_each(&tctx->xa, index, node) {
3188 if (io_local_work_pending(node->ctx)) {
3189 WARN_ON_ONCE(node->ctx->submitter_task &&
3190 node->ctx->submitter_task != current);
3191 goto end_wait;
3192 }
3193 }
3194 /*
3195 * If we've seen completions, retry without waiting. This
3196 * avoids a race where a completion comes in before we did
3197 * prepare_to_wait().
3198 */
3199 if (inflight == tctx_inflight(tctx, !cancel_all))
3200 schedule();
3201end_wait:
3202 finish_wait(&tctx->wait, &wait);
3203 } while (1);
3204
3205 io_uring_clean_tctx(tctx);
3206 if (cancel_all) {
3207 /*
3208 * We shouldn't run task_works after cancel, so just leave
3209 * ->in_cancel set for normal exit.
3210 */
3211 atomic_dec(&tctx->in_cancel);
3212 /* for exec all current's requests should be gone, kill tctx */
3213 __io_uring_free(current);
3214 }
3215}
3216
3217void __io_uring_cancel(bool cancel_all)
3218{
3219 io_uring_unreg_ringfd();
3220 io_uring_cancel_generic(cancel_all, NULL);
3221}
3222
3223static struct io_uring_reg_wait *io_get_ext_arg_reg(struct io_ring_ctx *ctx,
3224 const struct io_uring_getevents_arg __user *uarg)
3225{
3226 unsigned long size = sizeof(struct io_uring_reg_wait);
3227 unsigned long offset = (uintptr_t)uarg;
3228 unsigned long end;
3229
3230 if (unlikely(offset % sizeof(long)))
3231 return ERR_PTR(-EFAULT);
3232
3233 /* also protects from NULL ->cq_wait_arg as the size would be 0 */
3234 if (unlikely(check_add_overflow(offset, size, &end) ||
3235 end > ctx->cq_wait_size))
3236 return ERR_PTR(-EFAULT);
3237
3238 offset = array_index_nospec(offset, ctx->cq_wait_size - size);
3239 return ctx->cq_wait_arg + offset;
3240}
3241
3242static int io_validate_ext_arg(struct io_ring_ctx *ctx, unsigned flags,
3243 const void __user *argp, size_t argsz)
3244{
3245 struct io_uring_getevents_arg arg;
3246
3247 if (!(flags & IORING_ENTER_EXT_ARG))
3248 return 0;
3249 if (flags & IORING_ENTER_EXT_ARG_REG)
3250 return -EINVAL;
3251 if (argsz != sizeof(arg))
3252 return -EINVAL;
3253 if (copy_from_user(&arg, argp, sizeof(arg)))
3254 return -EFAULT;
3255 return 0;
3256}
3257
3258static int io_get_ext_arg(struct io_ring_ctx *ctx, unsigned flags,
3259 const void __user *argp, struct ext_arg *ext_arg)
3260{
3261 const struct io_uring_getevents_arg __user *uarg = argp;
3262 struct io_uring_getevents_arg arg;
3263
3264 /*
3265 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3266 * is just a pointer to the sigset_t.
3267 */
3268 if (!(flags & IORING_ENTER_EXT_ARG)) {
3269 ext_arg->sig = (const sigset_t __user *) argp;
3270 return 0;
3271 }
3272
3273 if (flags & IORING_ENTER_EXT_ARG_REG) {
3274 struct io_uring_reg_wait *w;
3275
3276 if (ext_arg->argsz != sizeof(struct io_uring_reg_wait))
3277 return -EINVAL;
3278 w = io_get_ext_arg_reg(ctx, argp);
3279 if (IS_ERR(w))
3280 return PTR_ERR(w);
3281
3282 if (w->flags & ~IORING_REG_WAIT_TS)
3283 return -EINVAL;
3284 ext_arg->min_time = READ_ONCE(w->min_wait_usec) * NSEC_PER_USEC;
3285 ext_arg->sig = u64_to_user_ptr(READ_ONCE(w->sigmask));
3286 ext_arg->argsz = READ_ONCE(w->sigmask_sz);
3287 if (w->flags & IORING_REG_WAIT_TS) {
3288 ext_arg->ts.tv_sec = READ_ONCE(w->ts.tv_sec);
3289 ext_arg->ts.tv_nsec = READ_ONCE(w->ts.tv_nsec);
3290 ext_arg->ts_set = true;
3291 }
3292 return 0;
3293 }
3294
3295 /*
3296 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3297 * timespec and sigset_t pointers if good.
3298 */
3299 if (ext_arg->argsz != sizeof(arg))
3300 return -EINVAL;
3301#ifdef CONFIG_64BIT
3302 if (!user_access_begin(uarg, sizeof(*uarg)))
3303 return -EFAULT;
3304 unsafe_get_user(arg.sigmask, &uarg->sigmask, uaccess_end);
3305 unsafe_get_user(arg.sigmask_sz, &uarg->sigmask_sz, uaccess_end);
3306 unsafe_get_user(arg.min_wait_usec, &uarg->min_wait_usec, uaccess_end);
3307 unsafe_get_user(arg.ts, &uarg->ts, uaccess_end);
3308 user_access_end();
3309#else
3310 if (copy_from_user(&arg, uarg, sizeof(arg)))
3311 return -EFAULT;
3312#endif
3313 ext_arg->min_time = arg.min_wait_usec * NSEC_PER_USEC;
3314 ext_arg->sig = u64_to_user_ptr(arg.sigmask);
3315 ext_arg->argsz = arg.sigmask_sz;
3316 if (arg.ts) {
3317 if (get_timespec64(&ext_arg->ts, u64_to_user_ptr(arg.ts)))
3318 return -EFAULT;
3319 ext_arg->ts_set = true;
3320 }
3321 return 0;
3322#ifdef CONFIG_64BIT
3323uaccess_end:
3324 user_access_end();
3325 return -EFAULT;
3326#endif
3327}
3328
3329SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3330 u32, min_complete, u32, flags, const void __user *, argp,
3331 size_t, argsz)
3332{
3333 struct io_ring_ctx *ctx;
3334 struct file *file;
3335 long ret;
3336
3337 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3338 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3339 IORING_ENTER_REGISTERED_RING |
3340 IORING_ENTER_ABS_TIMER |
3341 IORING_ENTER_EXT_ARG_REG)))
3342 return -EINVAL;
3343
3344 /*
3345 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3346 * need only dereference our task private array to find it.
3347 */
3348 if (flags & IORING_ENTER_REGISTERED_RING) {
3349 struct io_uring_task *tctx = current->io_uring;
3350
3351 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3352 return -EINVAL;
3353 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3354 file = tctx->registered_rings[fd];
3355 if (unlikely(!file))
3356 return -EBADF;
3357 } else {
3358 file = fget(fd);
3359 if (unlikely(!file))
3360 return -EBADF;
3361 ret = -EOPNOTSUPP;
3362 if (unlikely(!io_is_uring_fops(file)))
3363 goto out;
3364 }
3365
3366 ctx = file->private_data;
3367 ret = -EBADFD;
3368 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3369 goto out;
3370
3371 /*
3372 * For SQ polling, the thread will do all submissions and completions.
3373 * Just return the requested submit count, and wake the thread if
3374 * we were asked to.
3375 */
3376 ret = 0;
3377 if (ctx->flags & IORING_SETUP_SQPOLL) {
3378 if (unlikely(ctx->sq_data->thread == NULL)) {
3379 ret = -EOWNERDEAD;
3380 goto out;
3381 }
3382 if (flags & IORING_ENTER_SQ_WAKEUP)
3383 wake_up(&ctx->sq_data->wait);
3384 if (flags & IORING_ENTER_SQ_WAIT)
3385 io_sqpoll_wait_sq(ctx);
3386
3387 ret = to_submit;
3388 } else if (to_submit) {
3389 ret = io_uring_add_tctx_node(ctx);
3390 if (unlikely(ret))
3391 goto out;
3392
3393 mutex_lock(&ctx->uring_lock);
3394 ret = io_submit_sqes(ctx, to_submit);
3395 if (ret != to_submit) {
3396 mutex_unlock(&ctx->uring_lock);
3397 goto out;
3398 }
3399 if (flags & IORING_ENTER_GETEVENTS) {
3400 if (ctx->syscall_iopoll)
3401 goto iopoll_locked;
3402 /*
3403 * Ignore errors, we'll soon call io_cqring_wait() and
3404 * it should handle ownership problems if any.
3405 */
3406 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3407 (void)io_run_local_work_locked(ctx, min_complete);
3408 }
3409 mutex_unlock(&ctx->uring_lock);
3410 }
3411
3412 if (flags & IORING_ENTER_GETEVENTS) {
3413 int ret2;
3414
3415 if (ctx->syscall_iopoll) {
3416 /*
3417 * We disallow the app entering submit/complete with
3418 * polling, but we still need to lock the ring to
3419 * prevent racing with polled issue that got punted to
3420 * a workqueue.
3421 */
3422 mutex_lock(&ctx->uring_lock);
3423iopoll_locked:
3424 ret2 = io_validate_ext_arg(ctx, flags, argp, argsz);
3425 if (likely(!ret2)) {
3426 min_complete = min(min_complete,
3427 ctx->cq_entries);
3428 ret2 = io_iopoll_check(ctx, min_complete);
3429 }
3430 mutex_unlock(&ctx->uring_lock);
3431 } else {
3432 struct ext_arg ext_arg = { .argsz = argsz };
3433
3434 ret2 = io_get_ext_arg(ctx, flags, argp, &ext_arg);
3435 if (likely(!ret2)) {
3436 min_complete = min(min_complete,
3437 ctx->cq_entries);
3438 ret2 = io_cqring_wait(ctx, min_complete, flags,
3439 &ext_arg);
3440 }
3441 }
3442
3443 if (!ret) {
3444 ret = ret2;
3445
3446 /*
3447 * EBADR indicates that one or more CQE were dropped.
3448 * Once the user has been informed we can clear the bit
3449 * as they are obviously ok with those drops.
3450 */
3451 if (unlikely(ret2 == -EBADR))
3452 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3453 &ctx->check_cq);
3454 }
3455 }
3456out:
3457 if (!(flags & IORING_ENTER_REGISTERED_RING))
3458 fput(file);
3459 return ret;
3460}
3461
3462static const struct file_operations io_uring_fops = {
3463 .release = io_uring_release,
3464 .mmap = io_uring_mmap,
3465 .get_unmapped_area = io_uring_get_unmapped_area,
3466#ifndef CONFIG_MMU
3467 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3468#endif
3469 .poll = io_uring_poll,
3470#ifdef CONFIG_PROC_FS
3471 .show_fdinfo = io_uring_show_fdinfo,
3472#endif
3473};
3474
3475bool io_is_uring_fops(struct file *file)
3476{
3477 return file->f_op == &io_uring_fops;
3478}
3479
3480static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3481 struct io_uring_params *p)
3482{
3483 struct io_rings *rings;
3484 size_t size, sq_array_offset;
3485 void *ptr;
3486
3487 /* make sure these are sane, as we already accounted them */
3488 ctx->sq_entries = p->sq_entries;
3489 ctx->cq_entries = p->cq_entries;
3490
3491 size = rings_size(ctx->flags, p->sq_entries, p->cq_entries,
3492 &sq_array_offset);
3493 if (size == SIZE_MAX)
3494 return -EOVERFLOW;
3495
3496 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3497 rings = io_pages_map(&ctx->ring_pages, &ctx->n_ring_pages, size);
3498 else
3499 rings = io_rings_map(ctx, p->cq_off.user_addr, size);
3500
3501 if (IS_ERR(rings))
3502 return PTR_ERR(rings);
3503
3504 ctx->rings = rings;
3505 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY))
3506 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3507 rings->sq_ring_mask = p->sq_entries - 1;
3508 rings->cq_ring_mask = p->cq_entries - 1;
3509 rings->sq_ring_entries = p->sq_entries;
3510 rings->cq_ring_entries = p->cq_entries;
3511
3512 if (p->flags & IORING_SETUP_SQE128)
3513 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3514 else
3515 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3516 if (size == SIZE_MAX) {
3517 io_rings_free(ctx);
3518 return -EOVERFLOW;
3519 }
3520
3521 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3522 ptr = io_pages_map(&ctx->sqe_pages, &ctx->n_sqe_pages, size);
3523 else
3524 ptr = io_sqes_map(ctx, p->sq_off.user_addr, size);
3525
3526 if (IS_ERR(ptr)) {
3527 io_rings_free(ctx);
3528 return PTR_ERR(ptr);
3529 }
3530
3531 ctx->sq_sqes = ptr;
3532 return 0;
3533}
3534
3535static int io_uring_install_fd(struct file *file)
3536{
3537 int fd;
3538
3539 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3540 if (fd < 0)
3541 return fd;
3542 fd_install(fd, file);
3543 return fd;
3544}
3545
3546/*
3547 * Allocate an anonymous fd, this is what constitutes the application
3548 * visible backing of an io_uring instance. The application mmaps this
3549 * fd to gain access to the SQ/CQ ring details.
3550 */
3551static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3552{
3553 /* Create a new inode so that the LSM can block the creation. */
3554 return anon_inode_create_getfile("[io_uring]", &io_uring_fops, ctx,
3555 O_RDWR | O_CLOEXEC, NULL);
3556}
3557
3558int io_uring_fill_params(unsigned entries, struct io_uring_params *p)
3559{
3560 if (!entries)
3561 return -EINVAL;
3562 if (entries > IORING_MAX_ENTRIES) {
3563 if (!(p->flags & IORING_SETUP_CLAMP))
3564 return -EINVAL;
3565 entries = IORING_MAX_ENTRIES;
3566 }
3567
3568 if ((p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
3569 && !(p->flags & IORING_SETUP_NO_MMAP))
3570 return -EINVAL;
3571
3572 /*
3573 * Use twice as many entries for the CQ ring. It's possible for the
3574 * application to drive a higher depth than the size of the SQ ring,
3575 * since the sqes are only used at submission time. This allows for
3576 * some flexibility in overcommitting a bit. If the application has
3577 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3578 * of CQ ring entries manually.
3579 */
3580 p->sq_entries = roundup_pow_of_two(entries);
3581 if (p->flags & IORING_SETUP_CQSIZE) {
3582 /*
3583 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3584 * to a power-of-two, if it isn't already. We do NOT impose
3585 * any cq vs sq ring sizing.
3586 */
3587 if (!p->cq_entries)
3588 return -EINVAL;
3589 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3590 if (!(p->flags & IORING_SETUP_CLAMP))
3591 return -EINVAL;
3592 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3593 }
3594 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3595 if (p->cq_entries < p->sq_entries)
3596 return -EINVAL;
3597 } else {
3598 p->cq_entries = 2 * p->sq_entries;
3599 }
3600
3601 p->sq_off.head = offsetof(struct io_rings, sq.head);
3602 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3603 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3604 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3605 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3606 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3607 p->sq_off.resv1 = 0;
3608 if (!(p->flags & IORING_SETUP_NO_MMAP))
3609 p->sq_off.user_addr = 0;
3610
3611 p->cq_off.head = offsetof(struct io_rings, cq.head);
3612 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3613 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3614 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3615 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3616 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3617 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3618 p->cq_off.resv1 = 0;
3619 if (!(p->flags & IORING_SETUP_NO_MMAP))
3620 p->cq_off.user_addr = 0;
3621
3622 return 0;
3623}
3624
3625static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3626 struct io_uring_params __user *params)
3627{
3628 struct io_ring_ctx *ctx;
3629 struct io_uring_task *tctx;
3630 struct file *file;
3631 int ret;
3632
3633 ret = io_uring_fill_params(entries, p);
3634 if (unlikely(ret))
3635 return ret;
3636
3637 ctx = io_ring_ctx_alloc(p);
3638 if (!ctx)
3639 return -ENOMEM;
3640
3641 ctx->clockid = CLOCK_MONOTONIC;
3642 ctx->clock_offset = 0;
3643
3644 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY))
3645 static_branch_inc(&io_key_has_sqarray);
3646
3647 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3648 !(ctx->flags & IORING_SETUP_IOPOLL) &&
3649 !(ctx->flags & IORING_SETUP_SQPOLL))
3650 ctx->task_complete = true;
3651
3652 if (ctx->task_complete || (ctx->flags & IORING_SETUP_IOPOLL))
3653 ctx->lockless_cq = true;
3654
3655 /*
3656 * lazy poll_wq activation relies on ->task_complete for synchronisation
3657 * purposes, see io_activate_pollwq()
3658 */
3659 if (!ctx->task_complete)
3660 ctx->poll_activated = true;
3661
3662 /*
3663 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3664 * space applications don't need to do io completion events
3665 * polling again, they can rely on io_sq_thread to do polling
3666 * work, which can reduce cpu usage and uring_lock contention.
3667 */
3668 if (ctx->flags & IORING_SETUP_IOPOLL &&
3669 !(ctx->flags & IORING_SETUP_SQPOLL))
3670 ctx->syscall_iopoll = 1;
3671
3672 ctx->compat = in_compat_syscall();
3673 if (!ns_capable_noaudit(&init_user_ns, CAP_IPC_LOCK))
3674 ctx->user = get_uid(current_user());
3675
3676 /*
3677 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3678 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3679 */
3680 ret = -EINVAL;
3681 if (ctx->flags & IORING_SETUP_SQPOLL) {
3682 /* IPI related flags don't make sense with SQPOLL */
3683 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3684 IORING_SETUP_TASKRUN_FLAG |
3685 IORING_SETUP_DEFER_TASKRUN))
3686 goto err;
3687 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3688 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3689 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3690 } else {
3691 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3692 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3693 goto err;
3694 ctx->notify_method = TWA_SIGNAL;
3695 }
3696
3697 /* HYBRID_IOPOLL only valid with IOPOLL */
3698 if ((ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_HYBRID_IOPOLL)) ==
3699 IORING_SETUP_HYBRID_IOPOLL)
3700 goto err;
3701
3702 /*
3703 * For DEFER_TASKRUN we require the completion task to be the same as the
3704 * submission task. This implies that there is only one submitter, so enforce
3705 * that.
3706 */
3707 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3708 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3709 goto err;
3710 }
3711
3712 /*
3713 * This is just grabbed for accounting purposes. When a process exits,
3714 * the mm is exited and dropped before the files, hence we need to hang
3715 * on to this mm purely for the purposes of being able to unaccount
3716 * memory (locked/pinned vm). It's not used for anything else.
3717 */
3718 mmgrab(current->mm);
3719 ctx->mm_account = current->mm;
3720
3721 ret = io_allocate_scq_urings(ctx, p);
3722 if (ret)
3723 goto err;
3724
3725 if (!(p->flags & IORING_SETUP_NO_SQARRAY))
3726 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3727
3728 ret = io_sq_offload_create(ctx, p);
3729 if (ret)
3730 goto err;
3731
3732 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3733 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3734 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3735 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3736 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3737 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3738 IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING |
3739 IORING_FEAT_RECVSEND_BUNDLE | IORING_FEAT_MIN_TIMEOUT;
3740
3741 if (copy_to_user(params, p, sizeof(*p))) {
3742 ret = -EFAULT;
3743 goto err;
3744 }
3745
3746 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
3747 && !(ctx->flags & IORING_SETUP_R_DISABLED))
3748 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
3749
3750 file = io_uring_get_file(ctx);
3751 if (IS_ERR(file)) {
3752 ret = PTR_ERR(file);
3753 goto err;
3754 }
3755
3756 ret = __io_uring_add_tctx_node(ctx);
3757 if (ret)
3758 goto err_fput;
3759 tctx = current->io_uring;
3760
3761 /*
3762 * Install ring fd as the very last thing, so we don't risk someone
3763 * having closed it before we finish setup
3764 */
3765 if (p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
3766 ret = io_ring_add_registered_file(tctx, file, 0, IO_RINGFD_REG_MAX);
3767 else
3768 ret = io_uring_install_fd(file);
3769 if (ret < 0)
3770 goto err_fput;
3771
3772 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3773 return ret;
3774err:
3775 io_ring_ctx_wait_and_kill(ctx);
3776 return ret;
3777err_fput:
3778 fput(file);
3779 return ret;
3780}
3781
3782/*
3783 * Sets up an aio uring context, and returns the fd. Applications asks for a
3784 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3785 * params structure passed in.
3786 */
3787static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3788{
3789 struct io_uring_params p;
3790 int i;
3791
3792 if (copy_from_user(&p, params, sizeof(p)))
3793 return -EFAULT;
3794 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3795 if (p.resv[i])
3796 return -EINVAL;
3797 }
3798
3799 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3800 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3801 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3802 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3803 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3804 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3805 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN |
3806 IORING_SETUP_NO_MMAP | IORING_SETUP_REGISTERED_FD_ONLY |
3807 IORING_SETUP_NO_SQARRAY | IORING_SETUP_HYBRID_IOPOLL))
3808 return -EINVAL;
3809
3810 return io_uring_create(entries, &p, params);
3811}
3812
3813static inline bool io_uring_allowed(void)
3814{
3815 int disabled = READ_ONCE(sysctl_io_uring_disabled);
3816 kgid_t io_uring_group;
3817
3818 if (disabled == 2)
3819 return false;
3820
3821 if (disabled == 0 || capable(CAP_SYS_ADMIN))
3822 return true;
3823
3824 io_uring_group = make_kgid(&init_user_ns, sysctl_io_uring_group);
3825 if (!gid_valid(io_uring_group))
3826 return false;
3827
3828 return in_group_p(io_uring_group);
3829}
3830
3831SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3832 struct io_uring_params __user *, params)
3833{
3834 if (!io_uring_allowed())
3835 return -EPERM;
3836
3837 return io_uring_setup(entries, params);
3838}
3839
3840static int __init io_uring_init(void)
3841{
3842 struct kmem_cache_args kmem_args = {
3843 .useroffset = offsetof(struct io_kiocb, cmd.data),
3844 .usersize = sizeof_field(struct io_kiocb, cmd.data),
3845 .freeptr_offset = offsetof(struct io_kiocb, work),
3846 .use_freeptr_offset = true,
3847 };
3848
3849#define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
3850 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
3851 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
3852} while (0)
3853
3854#define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
3855 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
3856#define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
3857 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
3858 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
3859 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
3860 BUILD_BUG_SQE_ELEM(1, __u8, flags);
3861 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
3862 BUILD_BUG_SQE_ELEM(4, __s32, fd);
3863 BUILD_BUG_SQE_ELEM(8, __u64, off);
3864 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
3865 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
3866 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
3867 BUILD_BUG_SQE_ELEM(16, __u64, addr);
3868 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
3869 BUILD_BUG_SQE_ELEM(24, __u32, len);
3870 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
3871 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
3872 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
3873 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
3874 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
3875 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
3876 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
3877 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
3878 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
3879 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
3880 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
3881 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
3882 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
3883 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
3884 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
3885 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
3886 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
3887 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
3888 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
3889 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
3890 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
3891 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
3892 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
3893 BUILD_BUG_SQE_ELEM(42, __u16, personality);
3894 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
3895 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
3896 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
3897 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
3898 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
3899 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
3900 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
3901
3902 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
3903 sizeof(struct io_uring_rsrc_update));
3904 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
3905 sizeof(struct io_uring_rsrc_update2));
3906
3907 /* ->buf_index is u16 */
3908 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
3909 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
3910 offsetof(struct io_uring_buf_ring, tail));
3911
3912 /* should fit into one byte */
3913 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
3914 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
3915 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
3916
3917 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof_field(struct io_kiocb, flags));
3918
3919 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
3920
3921 /* top 8bits are for internal use */
3922 BUILD_BUG_ON((IORING_URING_CMD_MASK & 0xff000000) != 0);
3923
3924 io_uring_optable_init();
3925
3926 /*
3927 * Allow user copy in the per-command field, which starts after the
3928 * file in io_kiocb and until the opcode field. The openat2 handling
3929 * requires copying in user memory into the io_kiocb object in that
3930 * range, and HARDENED_USERCOPY will complain if we haven't
3931 * correctly annotated this range.
3932 */
3933 req_cachep = kmem_cache_create("io_kiocb", sizeof(struct io_kiocb), &kmem_args,
3934 SLAB_HWCACHE_ALIGN | SLAB_PANIC | SLAB_ACCOUNT |
3935 SLAB_TYPESAFE_BY_RCU);
3936 io_buf_cachep = KMEM_CACHE(io_buffer,
3937 SLAB_HWCACHE_ALIGN | SLAB_PANIC | SLAB_ACCOUNT);
3938
3939 iou_wq = alloc_workqueue("iou_exit", WQ_UNBOUND, 64);
3940
3941#ifdef CONFIG_SYSCTL
3942 register_sysctl_init("kernel", kernel_io_uring_disabled_table);
3943#endif
3944
3945 return 0;
3946};
3947__initcall(io_uring_init);
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/fdtable.h>
55#include <linux/mm.h>
56#include <linux/mman.h>
57#include <linux/percpu.h>
58#include <linux/slab.h>
59#include <linux/bvec.h>
60#include <linux/net.h>
61#include <net/sock.h>
62#include <net/af_unix.h>
63#include <linux/anon_inodes.h>
64#include <linux/sched/mm.h>
65#include <linux/uaccess.h>
66#include <linux/nospec.h>
67#include <linux/highmem.h>
68#include <linux/fsnotify.h>
69#include <linux/fadvise.h>
70#include <linux/task_work.h>
71#include <linux/io_uring.h>
72#include <linux/io_uring/cmd.h>
73#include <linux/audit.h>
74#include <linux/security.h>
75#include <asm/shmparam.h>
76
77#define CREATE_TRACE_POINTS
78#include <trace/events/io_uring.h>
79
80#include <uapi/linux/io_uring.h>
81
82#include "io-wq.h"
83
84#include "io_uring.h"
85#include "opdef.h"
86#include "refs.h"
87#include "tctx.h"
88#include "register.h"
89#include "sqpoll.h"
90#include "fdinfo.h"
91#include "kbuf.h"
92#include "rsrc.h"
93#include "cancel.h"
94#include "net.h"
95#include "notif.h"
96#include "waitid.h"
97#include "futex.h"
98
99#include "timeout.h"
100#include "poll.h"
101#include "rw.h"
102#include "alloc_cache.h"
103
104#define IORING_MAX_ENTRIES 32768
105#define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
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
125enum {
126 IO_CHECK_CQ_OVERFLOW_BIT,
127 IO_CHECK_CQ_DROPPED_BIT,
128};
129
130struct io_defer_entry {
131 struct list_head list;
132 struct io_kiocb *req;
133 u32 seq;
134};
135
136/* requests with any of those set should undergo io_disarm_next() */
137#define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
138#define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
139
140/*
141 * No waiters. It's larger than any valid value of the tw counter
142 * so that tests against ->cq_wait_nr would fail and skip wake_up().
143 */
144#define IO_CQ_WAKE_INIT (-1U)
145/* Forced wake up if there is a waiter regardless of ->cq_wait_nr */
146#define IO_CQ_WAKE_FORCE (IO_CQ_WAKE_INIT >> 1)
147
148static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
149 struct task_struct *task,
150 bool cancel_all);
151
152static void io_queue_sqe(struct io_kiocb *req);
153
154struct kmem_cache *req_cachep;
155
156static int __read_mostly sysctl_io_uring_disabled;
157static int __read_mostly sysctl_io_uring_group = -1;
158
159#ifdef CONFIG_SYSCTL
160static struct ctl_table kernel_io_uring_disabled_table[] = {
161 {
162 .procname = "io_uring_disabled",
163 .data = &sysctl_io_uring_disabled,
164 .maxlen = sizeof(sysctl_io_uring_disabled),
165 .mode = 0644,
166 .proc_handler = proc_dointvec_minmax,
167 .extra1 = SYSCTL_ZERO,
168 .extra2 = SYSCTL_TWO,
169 },
170 {
171 .procname = "io_uring_group",
172 .data = &sysctl_io_uring_group,
173 .maxlen = sizeof(gid_t),
174 .mode = 0644,
175 .proc_handler = proc_dointvec,
176 },
177 {},
178};
179#endif
180
181static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
182{
183 if (!wq_list_empty(&ctx->submit_state.compl_reqs) ||
184 ctx->submit_state.cqes_count)
185 __io_submit_flush_completions(ctx);
186}
187
188static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
189{
190 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
191}
192
193static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx)
194{
195 return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head);
196}
197
198static bool io_match_linked(struct io_kiocb *head)
199{
200 struct io_kiocb *req;
201
202 io_for_each_link(req, head) {
203 if (req->flags & REQ_F_INFLIGHT)
204 return true;
205 }
206 return false;
207}
208
209/*
210 * As io_match_task() but protected against racing with linked timeouts.
211 * User must not hold timeout_lock.
212 */
213bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
214 bool cancel_all)
215{
216 bool matched;
217
218 if (task && head->task != task)
219 return false;
220 if (cancel_all)
221 return true;
222
223 if (head->flags & REQ_F_LINK_TIMEOUT) {
224 struct io_ring_ctx *ctx = head->ctx;
225
226 /* protect against races with linked timeouts */
227 spin_lock_irq(&ctx->timeout_lock);
228 matched = io_match_linked(head);
229 spin_unlock_irq(&ctx->timeout_lock);
230 } else {
231 matched = io_match_linked(head);
232 }
233 return matched;
234}
235
236static inline void req_fail_link_node(struct io_kiocb *req, int res)
237{
238 req_set_fail(req);
239 io_req_set_res(req, res, 0);
240}
241
242static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
243{
244 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
245}
246
247static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
248{
249 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
250
251 complete(&ctx->ref_comp);
252}
253
254static __cold void io_fallback_req_func(struct work_struct *work)
255{
256 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
257 fallback_work.work);
258 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
259 struct io_kiocb *req, *tmp;
260 struct io_tw_state ts = { .locked = true, };
261
262 percpu_ref_get(&ctx->refs);
263 mutex_lock(&ctx->uring_lock);
264 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
265 req->io_task_work.func(req, &ts);
266 if (WARN_ON_ONCE(!ts.locked))
267 return;
268 io_submit_flush_completions(ctx);
269 mutex_unlock(&ctx->uring_lock);
270 percpu_ref_put(&ctx->refs);
271}
272
273static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
274{
275 unsigned hash_buckets = 1U << bits;
276 size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
277
278 table->hbs = kmalloc(hash_size, GFP_KERNEL);
279 if (!table->hbs)
280 return -ENOMEM;
281
282 table->hash_bits = bits;
283 init_hash_table(table, hash_buckets);
284 return 0;
285}
286
287static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
288{
289 struct io_ring_ctx *ctx;
290 int hash_bits;
291
292 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
293 if (!ctx)
294 return NULL;
295
296 xa_init(&ctx->io_bl_xa);
297
298 /*
299 * Use 5 bits less than the max cq entries, that should give us around
300 * 32 entries per hash list if totally full and uniformly spread, but
301 * don't keep too many buckets to not overconsume memory.
302 */
303 hash_bits = ilog2(p->cq_entries) - 5;
304 hash_bits = clamp(hash_bits, 1, 8);
305 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
306 goto err;
307 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
308 goto err;
309 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
310 0, GFP_KERNEL))
311 goto err;
312
313 ctx->flags = p->flags;
314 atomic_set(&ctx->cq_wait_nr, IO_CQ_WAKE_INIT);
315 init_waitqueue_head(&ctx->sqo_sq_wait);
316 INIT_LIST_HEAD(&ctx->sqd_list);
317 INIT_LIST_HEAD(&ctx->cq_overflow_list);
318 INIT_LIST_HEAD(&ctx->io_buffers_cache);
319 INIT_HLIST_HEAD(&ctx->io_buf_list);
320 io_alloc_cache_init(&ctx->rsrc_node_cache, IO_NODE_ALLOC_CACHE_MAX,
321 sizeof(struct io_rsrc_node));
322 io_alloc_cache_init(&ctx->apoll_cache, IO_ALLOC_CACHE_MAX,
323 sizeof(struct async_poll));
324 io_alloc_cache_init(&ctx->netmsg_cache, IO_ALLOC_CACHE_MAX,
325 sizeof(struct io_async_msghdr));
326 io_futex_cache_init(ctx);
327 init_completion(&ctx->ref_comp);
328 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
329 mutex_init(&ctx->uring_lock);
330 init_waitqueue_head(&ctx->cq_wait);
331 init_waitqueue_head(&ctx->poll_wq);
332 init_waitqueue_head(&ctx->rsrc_quiesce_wq);
333 spin_lock_init(&ctx->completion_lock);
334 spin_lock_init(&ctx->timeout_lock);
335 INIT_WQ_LIST(&ctx->iopoll_list);
336 INIT_LIST_HEAD(&ctx->io_buffers_comp);
337 INIT_LIST_HEAD(&ctx->defer_list);
338 INIT_LIST_HEAD(&ctx->timeout_list);
339 INIT_LIST_HEAD(&ctx->ltimeout_list);
340 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
341 init_llist_head(&ctx->work_llist);
342 INIT_LIST_HEAD(&ctx->tctx_list);
343 ctx->submit_state.free_list.next = NULL;
344 INIT_WQ_LIST(&ctx->locked_free_list);
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 return ctx;
353err:
354 kfree(ctx->cancel_table.hbs);
355 kfree(ctx->cancel_table_locked.hbs);
356 kfree(ctx->io_bl);
357 xa_destroy(&ctx->io_bl_xa);
358 kfree(ctx);
359 return NULL;
360}
361
362static void io_account_cq_overflow(struct io_ring_ctx *ctx)
363{
364 struct io_rings *r = ctx->rings;
365
366 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
367 ctx->cq_extra--;
368}
369
370static bool req_need_defer(struct io_kiocb *req, u32 seq)
371{
372 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
373 struct io_ring_ctx *ctx = req->ctx;
374
375 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
376 }
377
378 return false;
379}
380
381static void io_clean_op(struct io_kiocb *req)
382{
383 if (req->flags & REQ_F_BUFFER_SELECTED) {
384 spin_lock(&req->ctx->completion_lock);
385 io_put_kbuf_comp(req);
386 spin_unlock(&req->ctx->completion_lock);
387 }
388
389 if (req->flags & REQ_F_NEED_CLEANUP) {
390 const struct io_cold_def *def = &io_cold_defs[req->opcode];
391
392 if (def->cleanup)
393 def->cleanup(req);
394 }
395 if ((req->flags & REQ_F_POLLED) && req->apoll) {
396 kfree(req->apoll->double_poll);
397 kfree(req->apoll);
398 req->apoll = NULL;
399 }
400 if (req->flags & REQ_F_INFLIGHT) {
401 struct io_uring_task *tctx = req->task->io_uring;
402
403 atomic_dec(&tctx->inflight_tracked);
404 }
405 if (req->flags & REQ_F_CREDS)
406 put_cred(req->creds);
407 if (req->flags & REQ_F_ASYNC_DATA) {
408 kfree(req->async_data);
409 req->async_data = NULL;
410 }
411 req->flags &= ~IO_REQ_CLEAN_FLAGS;
412}
413
414static inline void io_req_track_inflight(struct io_kiocb *req)
415{
416 if (!(req->flags & REQ_F_INFLIGHT)) {
417 req->flags |= REQ_F_INFLIGHT;
418 atomic_inc(&req->task->io_uring->inflight_tracked);
419 }
420}
421
422static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
423{
424 if (WARN_ON_ONCE(!req->link))
425 return NULL;
426
427 req->flags &= ~REQ_F_ARM_LTIMEOUT;
428 req->flags |= REQ_F_LINK_TIMEOUT;
429
430 /* linked timeouts should have two refs once prep'ed */
431 io_req_set_refcount(req);
432 __io_req_set_refcount(req->link, 2);
433 return req->link;
434}
435
436static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
437{
438 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
439 return NULL;
440 return __io_prep_linked_timeout(req);
441}
442
443static noinline void __io_arm_ltimeout(struct io_kiocb *req)
444{
445 io_queue_linked_timeout(__io_prep_linked_timeout(req));
446}
447
448static inline void io_arm_ltimeout(struct io_kiocb *req)
449{
450 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
451 __io_arm_ltimeout(req);
452}
453
454static void io_prep_async_work(struct io_kiocb *req)
455{
456 const struct io_issue_def *def = &io_issue_defs[req->opcode];
457 struct io_ring_ctx *ctx = req->ctx;
458
459 if (!(req->flags & REQ_F_CREDS)) {
460 req->flags |= REQ_F_CREDS;
461 req->creds = get_current_cred();
462 }
463
464 req->work.list.next = NULL;
465 req->work.flags = 0;
466 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
467 if (req->flags & REQ_F_FORCE_ASYNC)
468 req->work.flags |= IO_WQ_WORK_CONCURRENT;
469
470 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
471 req->flags |= io_file_get_flags(req->file);
472
473 if (req->file && (req->flags & REQ_F_ISREG)) {
474 bool should_hash = def->hash_reg_file;
475
476 /* don't serialize this request if the fs doesn't need it */
477 if (should_hash && (req->file->f_flags & O_DIRECT) &&
478 (req->file->f_mode & FMODE_DIO_PARALLEL_WRITE))
479 should_hash = false;
480 if (should_hash || (ctx->flags & IORING_SETUP_IOPOLL))
481 io_wq_hash_work(&req->work, file_inode(req->file));
482 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
483 if (def->unbound_nonreg_file)
484 req->work.flags |= IO_WQ_WORK_UNBOUND;
485 }
486}
487
488static void io_prep_async_link(struct io_kiocb *req)
489{
490 struct io_kiocb *cur;
491
492 if (req->flags & REQ_F_LINK_TIMEOUT) {
493 struct io_ring_ctx *ctx = req->ctx;
494
495 spin_lock_irq(&ctx->timeout_lock);
496 io_for_each_link(cur, req)
497 io_prep_async_work(cur);
498 spin_unlock_irq(&ctx->timeout_lock);
499 } else {
500 io_for_each_link(cur, req)
501 io_prep_async_work(cur);
502 }
503}
504
505void io_queue_iowq(struct io_kiocb *req, struct io_tw_state *ts_dont_use)
506{
507 struct io_kiocb *link = io_prep_linked_timeout(req);
508 struct io_uring_task *tctx = req->task->io_uring;
509
510 BUG_ON(!tctx);
511 BUG_ON(!tctx->io_wq);
512
513 /* init ->work of the whole link before punting */
514 io_prep_async_link(req);
515
516 /*
517 * Not expected to happen, but if we do have a bug where this _can_
518 * happen, catch it here and ensure the request is marked as
519 * canceled. That will make io-wq go through the usual work cancel
520 * procedure rather than attempt to run this request (or create a new
521 * worker for it).
522 */
523 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
524 req->work.flags |= IO_WQ_WORK_CANCEL;
525
526 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
527 io_wq_enqueue(tctx->io_wq, &req->work);
528 if (link)
529 io_queue_linked_timeout(link);
530}
531
532static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
533{
534 while (!list_empty(&ctx->defer_list)) {
535 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
536 struct io_defer_entry, list);
537
538 if (req_need_defer(de->req, de->seq))
539 break;
540 list_del_init(&de->list);
541 io_req_task_queue(de->req);
542 kfree(de);
543 }
544}
545
546void io_eventfd_ops(struct rcu_head *rcu)
547{
548 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
549 int ops = atomic_xchg(&ev_fd->ops, 0);
550
551 if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT))
552 eventfd_signal_mask(ev_fd->cq_ev_fd, EPOLL_URING_WAKE);
553
554 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
555 * ordering in a race but if references are 0 we know we have to free
556 * it regardless.
557 */
558 if (atomic_dec_and_test(&ev_fd->refs)) {
559 eventfd_ctx_put(ev_fd->cq_ev_fd);
560 kfree(ev_fd);
561 }
562}
563
564static void io_eventfd_signal(struct io_ring_ctx *ctx)
565{
566 struct io_ev_fd *ev_fd = NULL;
567
568 rcu_read_lock();
569 /*
570 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
571 * and eventfd_signal
572 */
573 ev_fd = rcu_dereference(ctx->io_ev_fd);
574
575 /*
576 * Check again if ev_fd exists incase an io_eventfd_unregister call
577 * completed between the NULL check of ctx->io_ev_fd at the start of
578 * the function and rcu_read_lock.
579 */
580 if (unlikely(!ev_fd))
581 goto out;
582 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
583 goto out;
584 if (ev_fd->eventfd_async && !io_wq_current_is_worker())
585 goto out;
586
587 if (likely(eventfd_signal_allowed())) {
588 eventfd_signal_mask(ev_fd->cq_ev_fd, EPOLL_URING_WAKE);
589 } else {
590 atomic_inc(&ev_fd->refs);
591 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops))
592 call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops);
593 else
594 atomic_dec(&ev_fd->refs);
595 }
596
597out:
598 rcu_read_unlock();
599}
600
601static void io_eventfd_flush_signal(struct io_ring_ctx *ctx)
602{
603 bool skip;
604
605 spin_lock(&ctx->completion_lock);
606
607 /*
608 * Eventfd should only get triggered when at least one event has been
609 * posted. Some applications rely on the eventfd notification count
610 * only changing IFF a new CQE has been added to the CQ ring. There's
611 * no depedency on 1:1 relationship between how many times this
612 * function is called (and hence the eventfd count) and number of CQEs
613 * posted to the CQ ring.
614 */
615 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
616 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
617 spin_unlock(&ctx->completion_lock);
618 if (skip)
619 return;
620
621 io_eventfd_signal(ctx);
622}
623
624void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
625{
626 if (ctx->poll_activated)
627 io_poll_wq_wake(ctx);
628 if (ctx->off_timeout_used)
629 io_flush_timeouts(ctx);
630 if (ctx->drain_active) {
631 spin_lock(&ctx->completion_lock);
632 io_queue_deferred(ctx);
633 spin_unlock(&ctx->completion_lock);
634 }
635 if (ctx->has_evfd)
636 io_eventfd_flush_signal(ctx);
637}
638
639static inline void __io_cq_lock(struct io_ring_ctx *ctx)
640{
641 if (!ctx->lockless_cq)
642 spin_lock(&ctx->completion_lock);
643}
644
645static inline void io_cq_lock(struct io_ring_ctx *ctx)
646 __acquires(ctx->completion_lock)
647{
648 spin_lock(&ctx->completion_lock);
649}
650
651static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
652{
653 io_commit_cqring(ctx);
654 if (!ctx->task_complete) {
655 if (!ctx->lockless_cq)
656 spin_unlock(&ctx->completion_lock);
657 /* IOPOLL rings only need to wake up if it's also SQPOLL */
658 if (!ctx->syscall_iopoll)
659 io_cqring_wake(ctx);
660 }
661 io_commit_cqring_flush(ctx);
662}
663
664static void io_cq_unlock_post(struct io_ring_ctx *ctx)
665 __releases(ctx->completion_lock)
666{
667 io_commit_cqring(ctx);
668 spin_unlock(&ctx->completion_lock);
669 io_cqring_wake(ctx);
670 io_commit_cqring_flush(ctx);
671}
672
673/* Returns true if there are no backlogged entries after the flush */
674static void io_cqring_overflow_kill(struct io_ring_ctx *ctx)
675{
676 struct io_overflow_cqe *ocqe;
677 LIST_HEAD(list);
678
679 spin_lock(&ctx->completion_lock);
680 list_splice_init(&ctx->cq_overflow_list, &list);
681 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
682 spin_unlock(&ctx->completion_lock);
683
684 while (!list_empty(&list)) {
685 ocqe = list_first_entry(&list, struct io_overflow_cqe, list);
686 list_del(&ocqe->list);
687 kfree(ocqe);
688 }
689}
690
691static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx)
692{
693 size_t cqe_size = sizeof(struct io_uring_cqe);
694
695 if (__io_cqring_events(ctx) == ctx->cq_entries)
696 return;
697
698 if (ctx->flags & IORING_SETUP_CQE32)
699 cqe_size <<= 1;
700
701 io_cq_lock(ctx);
702 while (!list_empty(&ctx->cq_overflow_list)) {
703 struct io_uring_cqe *cqe;
704 struct io_overflow_cqe *ocqe;
705
706 if (!io_get_cqe_overflow(ctx, &cqe, true))
707 break;
708 ocqe = list_first_entry(&ctx->cq_overflow_list,
709 struct io_overflow_cqe, list);
710 memcpy(cqe, &ocqe->cqe, cqe_size);
711 list_del(&ocqe->list);
712 kfree(ocqe);
713 }
714
715 if (list_empty(&ctx->cq_overflow_list)) {
716 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
717 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
718 }
719 io_cq_unlock_post(ctx);
720}
721
722static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx)
723{
724 /* iopoll syncs against uring_lock, not completion_lock */
725 if (ctx->flags & IORING_SETUP_IOPOLL)
726 mutex_lock(&ctx->uring_lock);
727 __io_cqring_overflow_flush(ctx);
728 if (ctx->flags & IORING_SETUP_IOPOLL)
729 mutex_unlock(&ctx->uring_lock);
730}
731
732static void io_cqring_overflow_flush(struct io_ring_ctx *ctx)
733{
734 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
735 io_cqring_do_overflow_flush(ctx);
736}
737
738/* can be called by any task */
739static void io_put_task_remote(struct task_struct *task)
740{
741 struct io_uring_task *tctx = task->io_uring;
742
743 percpu_counter_sub(&tctx->inflight, 1);
744 if (unlikely(atomic_read(&tctx->in_cancel)))
745 wake_up(&tctx->wait);
746 put_task_struct(task);
747}
748
749/* used by a task to put its own references */
750static void io_put_task_local(struct task_struct *task)
751{
752 task->io_uring->cached_refs++;
753}
754
755/* must to be called somewhat shortly after putting a request */
756static inline void io_put_task(struct task_struct *task)
757{
758 if (likely(task == current))
759 io_put_task_local(task);
760 else
761 io_put_task_remote(task);
762}
763
764void io_task_refs_refill(struct io_uring_task *tctx)
765{
766 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
767
768 percpu_counter_add(&tctx->inflight, refill);
769 refcount_add(refill, ¤t->usage);
770 tctx->cached_refs += refill;
771}
772
773static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
774{
775 struct io_uring_task *tctx = task->io_uring;
776 unsigned int refs = tctx->cached_refs;
777
778 if (refs) {
779 tctx->cached_refs = 0;
780 percpu_counter_sub(&tctx->inflight, refs);
781 put_task_struct_many(task, refs);
782 }
783}
784
785static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
786 s32 res, u32 cflags, u64 extra1, u64 extra2)
787{
788 struct io_overflow_cqe *ocqe;
789 size_t ocq_size = sizeof(struct io_overflow_cqe);
790 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
791
792 lockdep_assert_held(&ctx->completion_lock);
793
794 if (is_cqe32)
795 ocq_size += sizeof(struct io_uring_cqe);
796
797 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
798 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
799 if (!ocqe) {
800 /*
801 * If we're in ring overflow flush mode, or in task cancel mode,
802 * or cannot allocate an overflow entry, then we need to drop it
803 * on the floor.
804 */
805 io_account_cq_overflow(ctx);
806 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
807 return false;
808 }
809 if (list_empty(&ctx->cq_overflow_list)) {
810 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
811 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
812
813 }
814 ocqe->cqe.user_data = user_data;
815 ocqe->cqe.res = res;
816 ocqe->cqe.flags = cflags;
817 if (is_cqe32) {
818 ocqe->cqe.big_cqe[0] = extra1;
819 ocqe->cqe.big_cqe[1] = extra2;
820 }
821 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
822 return true;
823}
824
825void io_req_cqe_overflow(struct io_kiocb *req)
826{
827 io_cqring_event_overflow(req->ctx, req->cqe.user_data,
828 req->cqe.res, req->cqe.flags,
829 req->big_cqe.extra1, req->big_cqe.extra2);
830 memset(&req->big_cqe, 0, sizeof(req->big_cqe));
831}
832
833/*
834 * writes to the cq entry need to come after reading head; the
835 * control dependency is enough as we're using WRITE_ONCE to
836 * fill the cq entry
837 */
838bool io_cqe_cache_refill(struct io_ring_ctx *ctx, bool overflow)
839{
840 struct io_rings *rings = ctx->rings;
841 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
842 unsigned int free, queued, len;
843
844 /*
845 * Posting into the CQ when there are pending overflowed CQEs may break
846 * ordering guarantees, which will affect links, F_MORE users and more.
847 * Force overflow the completion.
848 */
849 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
850 return false;
851
852 /* userspace may cheat modifying the tail, be safe and do min */
853 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
854 free = ctx->cq_entries - queued;
855 /* we need a contiguous range, limit based on the current array offset */
856 len = min(free, ctx->cq_entries - off);
857 if (!len)
858 return false;
859
860 if (ctx->flags & IORING_SETUP_CQE32) {
861 off <<= 1;
862 len <<= 1;
863 }
864
865 ctx->cqe_cached = &rings->cqes[off];
866 ctx->cqe_sentinel = ctx->cqe_cached + len;
867 return true;
868}
869
870static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
871 u32 cflags)
872{
873 struct io_uring_cqe *cqe;
874
875 ctx->cq_extra++;
876
877 /*
878 * If we can't get a cq entry, userspace overflowed the
879 * submission (by quite a lot). Increment the overflow count in
880 * the ring.
881 */
882 if (likely(io_get_cqe(ctx, &cqe))) {
883 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
884
885 WRITE_ONCE(cqe->user_data, user_data);
886 WRITE_ONCE(cqe->res, res);
887 WRITE_ONCE(cqe->flags, cflags);
888
889 if (ctx->flags & IORING_SETUP_CQE32) {
890 WRITE_ONCE(cqe->big_cqe[0], 0);
891 WRITE_ONCE(cqe->big_cqe[1], 0);
892 }
893 return true;
894 }
895 return false;
896}
897
898static void __io_flush_post_cqes(struct io_ring_ctx *ctx)
899 __must_hold(&ctx->uring_lock)
900{
901 struct io_submit_state *state = &ctx->submit_state;
902 unsigned int i;
903
904 lockdep_assert_held(&ctx->uring_lock);
905 for (i = 0; i < state->cqes_count; i++) {
906 struct io_uring_cqe *cqe = &ctx->completion_cqes[i];
907
908 if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) {
909 if (ctx->lockless_cq) {
910 spin_lock(&ctx->completion_lock);
911 io_cqring_event_overflow(ctx, cqe->user_data,
912 cqe->res, cqe->flags, 0, 0);
913 spin_unlock(&ctx->completion_lock);
914 } else {
915 io_cqring_event_overflow(ctx, cqe->user_data,
916 cqe->res, cqe->flags, 0, 0);
917 }
918 }
919 }
920 state->cqes_count = 0;
921}
922
923static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
924 bool allow_overflow)
925{
926 bool filled;
927
928 io_cq_lock(ctx);
929 filled = io_fill_cqe_aux(ctx, user_data, res, cflags);
930 if (!filled && allow_overflow)
931 filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
932
933 io_cq_unlock_post(ctx);
934 return filled;
935}
936
937bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags)
938{
939 return __io_post_aux_cqe(ctx, user_data, res, cflags, true);
940}
941
942/*
943 * A helper for multishot requests posting additional CQEs.
944 * Should only be used from a task_work including IO_URING_F_MULTISHOT.
945 */
946bool io_fill_cqe_req_aux(struct io_kiocb *req, bool defer, s32 res, u32 cflags)
947{
948 struct io_ring_ctx *ctx = req->ctx;
949 u64 user_data = req->cqe.user_data;
950 struct io_uring_cqe *cqe;
951
952 if (!defer)
953 return __io_post_aux_cqe(ctx, user_data, res, cflags, false);
954
955 lockdep_assert_held(&ctx->uring_lock);
956
957 if (ctx->submit_state.cqes_count == ARRAY_SIZE(ctx->completion_cqes)) {
958 __io_cq_lock(ctx);
959 __io_flush_post_cqes(ctx);
960 /* no need to flush - flush is deferred */
961 __io_cq_unlock_post(ctx);
962 }
963
964 /* For defered completions this is not as strict as it is otherwise,
965 * however it's main job is to prevent unbounded posted completions,
966 * and in that it works just as well.
967 */
968 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
969 return false;
970
971 cqe = &ctx->completion_cqes[ctx->submit_state.cqes_count++];
972 cqe->user_data = user_data;
973 cqe->res = res;
974 cqe->flags = cflags;
975 return true;
976}
977
978static void __io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
979{
980 struct io_ring_ctx *ctx = req->ctx;
981 struct io_rsrc_node *rsrc_node = NULL;
982
983 io_cq_lock(ctx);
984 if (!(req->flags & REQ_F_CQE_SKIP)) {
985 if (!io_fill_cqe_req(ctx, req))
986 io_req_cqe_overflow(req);
987 }
988
989 /*
990 * If we're the last reference to this request, add to our locked
991 * free_list cache.
992 */
993 if (req_ref_put_and_test(req)) {
994 if (req->flags & IO_REQ_LINK_FLAGS) {
995 if (req->flags & IO_DISARM_MASK)
996 io_disarm_next(req);
997 if (req->link) {
998 io_req_task_queue(req->link);
999 req->link = NULL;
1000 }
1001 }
1002 io_put_kbuf_comp(req);
1003 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1004 io_clean_op(req);
1005 io_put_file(req);
1006
1007 rsrc_node = req->rsrc_node;
1008 /*
1009 * Selected buffer deallocation in io_clean_op() assumes that
1010 * we don't hold ->completion_lock. Clean them here to avoid
1011 * deadlocks.
1012 */
1013 io_put_task_remote(req->task);
1014 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
1015 ctx->locked_free_nr++;
1016 }
1017 io_cq_unlock_post(ctx);
1018
1019 if (rsrc_node) {
1020 io_ring_submit_lock(ctx, issue_flags);
1021 io_put_rsrc_node(ctx, rsrc_node);
1022 io_ring_submit_unlock(ctx, issue_flags);
1023 }
1024}
1025
1026void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
1027{
1028 if (req->ctx->task_complete && req->ctx->submitter_task != current) {
1029 req->io_task_work.func = io_req_task_complete;
1030 io_req_task_work_add(req);
1031 } else if (!(issue_flags & IO_URING_F_UNLOCKED) ||
1032 !(req->ctx->flags & IORING_SETUP_IOPOLL)) {
1033 __io_req_complete_post(req, issue_flags);
1034 } else {
1035 struct io_ring_ctx *ctx = req->ctx;
1036
1037 mutex_lock(&ctx->uring_lock);
1038 __io_req_complete_post(req, issue_flags & ~IO_URING_F_UNLOCKED);
1039 mutex_unlock(&ctx->uring_lock);
1040 }
1041}
1042
1043void io_req_defer_failed(struct io_kiocb *req, s32 res)
1044 __must_hold(&ctx->uring_lock)
1045{
1046 const struct io_cold_def *def = &io_cold_defs[req->opcode];
1047
1048 lockdep_assert_held(&req->ctx->uring_lock);
1049
1050 req_set_fail(req);
1051 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
1052 if (def->fail)
1053 def->fail(req);
1054 io_req_complete_defer(req);
1055}
1056
1057/*
1058 * Don't initialise the fields below on every allocation, but do that in
1059 * advance and keep them valid across allocations.
1060 */
1061static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1062{
1063 req->ctx = ctx;
1064 req->link = NULL;
1065 req->async_data = NULL;
1066 /* not necessary, but safer to zero */
1067 memset(&req->cqe, 0, sizeof(req->cqe));
1068 memset(&req->big_cqe, 0, sizeof(req->big_cqe));
1069}
1070
1071static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1072 struct io_submit_state *state)
1073{
1074 spin_lock(&ctx->completion_lock);
1075 wq_list_splice(&ctx->locked_free_list, &state->free_list);
1076 ctx->locked_free_nr = 0;
1077 spin_unlock(&ctx->completion_lock);
1078}
1079
1080/*
1081 * A request might get retired back into the request caches even before opcode
1082 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1083 * Because of that, io_alloc_req() should be called only under ->uring_lock
1084 * and with extra caution to not get a request that is still worked on.
1085 */
1086__cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
1087 __must_hold(&ctx->uring_lock)
1088{
1089 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1090 void *reqs[IO_REQ_ALLOC_BATCH];
1091 int ret, i;
1092
1093 /*
1094 * If we have more than a batch's worth of requests in our IRQ side
1095 * locked cache, grab the lock and move them over to our submission
1096 * side cache.
1097 */
1098 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
1099 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
1100 if (!io_req_cache_empty(ctx))
1101 return true;
1102 }
1103
1104 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
1105
1106 /*
1107 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1108 * retry single alloc to be on the safe side.
1109 */
1110 if (unlikely(ret <= 0)) {
1111 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1112 if (!reqs[0])
1113 return false;
1114 ret = 1;
1115 }
1116
1117 percpu_ref_get_many(&ctx->refs, ret);
1118 for (i = 0; i < ret; i++) {
1119 struct io_kiocb *req = reqs[i];
1120
1121 io_preinit_req(req, ctx);
1122 io_req_add_to_cache(req, ctx);
1123 }
1124 return true;
1125}
1126
1127__cold void io_free_req(struct io_kiocb *req)
1128{
1129 /* refs were already put, restore them for io_req_task_complete() */
1130 req->flags &= ~REQ_F_REFCOUNT;
1131 /* we only want to free it, don't post CQEs */
1132 req->flags |= REQ_F_CQE_SKIP;
1133 req->io_task_work.func = io_req_task_complete;
1134 io_req_task_work_add(req);
1135}
1136
1137static void __io_req_find_next_prep(struct io_kiocb *req)
1138{
1139 struct io_ring_ctx *ctx = req->ctx;
1140
1141 spin_lock(&ctx->completion_lock);
1142 io_disarm_next(req);
1143 spin_unlock(&ctx->completion_lock);
1144}
1145
1146static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1147{
1148 struct io_kiocb *nxt;
1149
1150 /*
1151 * If LINK is set, we have dependent requests in this chain. If we
1152 * didn't fail this request, queue the first one up, moving any other
1153 * dependencies to the next request. In case of failure, fail the rest
1154 * of the chain.
1155 */
1156 if (unlikely(req->flags & IO_DISARM_MASK))
1157 __io_req_find_next_prep(req);
1158 nxt = req->link;
1159 req->link = NULL;
1160 return nxt;
1161}
1162
1163static void ctx_flush_and_put(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1164{
1165 if (!ctx)
1166 return;
1167 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1168 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1169 if (ts->locked) {
1170 io_submit_flush_completions(ctx);
1171 mutex_unlock(&ctx->uring_lock);
1172 ts->locked = false;
1173 }
1174 percpu_ref_put(&ctx->refs);
1175}
1176
1177static unsigned int handle_tw_list(struct llist_node *node,
1178 struct io_ring_ctx **ctx,
1179 struct io_tw_state *ts,
1180 struct llist_node *last)
1181{
1182 unsigned int count = 0;
1183
1184 while (node && node != last) {
1185 struct llist_node *next = node->next;
1186 struct io_kiocb *req = container_of(node, struct io_kiocb,
1187 io_task_work.node);
1188
1189 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1190
1191 if (req->ctx != *ctx) {
1192 ctx_flush_and_put(*ctx, ts);
1193 *ctx = req->ctx;
1194 /* if not contended, grab and improve batching */
1195 ts->locked = mutex_trylock(&(*ctx)->uring_lock);
1196 percpu_ref_get(&(*ctx)->refs);
1197 }
1198 INDIRECT_CALL_2(req->io_task_work.func,
1199 io_poll_task_func, io_req_rw_complete,
1200 req, ts);
1201 node = next;
1202 count++;
1203 if (unlikely(need_resched())) {
1204 ctx_flush_and_put(*ctx, ts);
1205 *ctx = NULL;
1206 cond_resched();
1207 }
1208 }
1209
1210 return count;
1211}
1212
1213/**
1214 * io_llist_xchg - swap all entries in a lock-less list
1215 * @head: the head of lock-less list to delete all entries
1216 * @new: new entry as the head of the list
1217 *
1218 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1219 * The order of entries returned is from the newest to the oldest added one.
1220 */
1221static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1222 struct llist_node *new)
1223{
1224 return xchg(&head->first, new);
1225}
1226
1227/**
1228 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1229 * @head: the head of lock-less list to delete all entries
1230 * @old: expected old value of the first entry of the list
1231 * @new: new entry as the head of the list
1232 *
1233 * perform a cmpxchg on the first entry of the list.
1234 */
1235
1236static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1237 struct llist_node *old,
1238 struct llist_node *new)
1239{
1240 return cmpxchg(&head->first, old, new);
1241}
1242
1243static __cold void io_fallback_tw(struct io_uring_task *tctx, bool sync)
1244{
1245 struct llist_node *node = llist_del_all(&tctx->task_list);
1246 struct io_ring_ctx *last_ctx = NULL;
1247 struct io_kiocb *req;
1248
1249 while (node) {
1250 req = container_of(node, struct io_kiocb, io_task_work.node);
1251 node = node->next;
1252 if (sync && last_ctx != req->ctx) {
1253 if (last_ctx) {
1254 flush_delayed_work(&last_ctx->fallback_work);
1255 percpu_ref_put(&last_ctx->refs);
1256 }
1257 last_ctx = req->ctx;
1258 percpu_ref_get(&last_ctx->refs);
1259 }
1260 if (llist_add(&req->io_task_work.node,
1261 &req->ctx->fallback_llist))
1262 schedule_delayed_work(&req->ctx->fallback_work, 1);
1263 }
1264
1265 if (last_ctx) {
1266 flush_delayed_work(&last_ctx->fallback_work);
1267 percpu_ref_put(&last_ctx->refs);
1268 }
1269}
1270
1271void tctx_task_work(struct callback_head *cb)
1272{
1273 struct io_tw_state ts = {};
1274 struct io_ring_ctx *ctx = NULL;
1275 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1276 task_work);
1277 struct llist_node fake = {};
1278 struct llist_node *node;
1279 unsigned int loops = 0;
1280 unsigned int count = 0;
1281
1282 if (unlikely(current->flags & PF_EXITING)) {
1283 io_fallback_tw(tctx, true);
1284 return;
1285 }
1286
1287 do {
1288 loops++;
1289 node = io_llist_xchg(&tctx->task_list, &fake);
1290 count += handle_tw_list(node, &ctx, &ts, &fake);
1291
1292 /* skip expensive cmpxchg if there are items in the list */
1293 if (READ_ONCE(tctx->task_list.first) != &fake)
1294 continue;
1295 if (ts.locked && !wq_list_empty(&ctx->submit_state.compl_reqs)) {
1296 io_submit_flush_completions(ctx);
1297 if (READ_ONCE(tctx->task_list.first) != &fake)
1298 continue;
1299 }
1300 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1301 } while (node != &fake);
1302
1303 ctx_flush_and_put(ctx, &ts);
1304
1305 /* relaxed read is enough as only the task itself sets ->in_cancel */
1306 if (unlikely(atomic_read(&tctx->in_cancel)))
1307 io_uring_drop_tctx_refs(current);
1308
1309 trace_io_uring_task_work_run(tctx, count, loops);
1310}
1311
1312static inline void io_req_local_work_add(struct io_kiocb *req, unsigned flags)
1313{
1314 struct io_ring_ctx *ctx = req->ctx;
1315 unsigned nr_wait, nr_tw, nr_tw_prev;
1316 struct llist_node *head;
1317
1318 /* See comment above IO_CQ_WAKE_INIT */
1319 BUILD_BUG_ON(IO_CQ_WAKE_FORCE <= IORING_MAX_CQ_ENTRIES);
1320
1321 /*
1322 * We don't know how many reuqests is there in the link and whether
1323 * they can even be queued lazily, fall back to non-lazy.
1324 */
1325 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK))
1326 flags &= ~IOU_F_TWQ_LAZY_WAKE;
1327
1328 head = READ_ONCE(ctx->work_llist.first);
1329 do {
1330 nr_tw_prev = 0;
1331 if (head) {
1332 struct io_kiocb *first_req = container_of(head,
1333 struct io_kiocb,
1334 io_task_work.node);
1335 /*
1336 * Might be executed at any moment, rely on
1337 * SLAB_TYPESAFE_BY_RCU to keep it alive.
1338 */
1339 nr_tw_prev = READ_ONCE(first_req->nr_tw);
1340 }
1341
1342 /*
1343 * Theoretically, it can overflow, but that's fine as one of
1344 * previous adds should've tried to wake the task.
1345 */
1346 nr_tw = nr_tw_prev + 1;
1347 if (!(flags & IOU_F_TWQ_LAZY_WAKE))
1348 nr_tw = IO_CQ_WAKE_FORCE;
1349
1350 req->nr_tw = nr_tw;
1351 req->io_task_work.node.next = head;
1352 } while (!try_cmpxchg(&ctx->work_llist.first, &head,
1353 &req->io_task_work.node));
1354
1355 /*
1356 * cmpxchg implies a full barrier, which pairs with the barrier
1357 * in set_current_state() on the io_cqring_wait() side. It's used
1358 * to ensure that either we see updated ->cq_wait_nr, or waiters
1359 * going to sleep will observe the work added to the list, which
1360 * is similar to the wait/wawke task state sync.
1361 */
1362
1363 if (!head) {
1364 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1365 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1366 if (ctx->has_evfd)
1367 io_eventfd_signal(ctx);
1368 }
1369
1370 nr_wait = atomic_read(&ctx->cq_wait_nr);
1371 /* not enough or no one is waiting */
1372 if (nr_tw < nr_wait)
1373 return;
1374 /* the previous add has already woken it up */
1375 if (nr_tw_prev >= nr_wait)
1376 return;
1377 wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE);
1378}
1379
1380static void io_req_normal_work_add(struct io_kiocb *req)
1381{
1382 struct io_uring_task *tctx = req->task->io_uring;
1383 struct io_ring_ctx *ctx = req->ctx;
1384
1385 /* task_work already pending, we're done */
1386 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1387 return;
1388
1389 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1390 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1391
1392 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1393 return;
1394
1395 io_fallback_tw(tctx, false);
1396}
1397
1398void __io_req_task_work_add(struct io_kiocb *req, unsigned flags)
1399{
1400 if (req->ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
1401 rcu_read_lock();
1402 io_req_local_work_add(req, flags);
1403 rcu_read_unlock();
1404 } else {
1405 io_req_normal_work_add(req);
1406 }
1407}
1408
1409static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1410{
1411 struct llist_node *node;
1412
1413 node = llist_del_all(&ctx->work_llist);
1414 while (node) {
1415 struct io_kiocb *req = container_of(node, struct io_kiocb,
1416 io_task_work.node);
1417
1418 node = node->next;
1419 io_req_normal_work_add(req);
1420 }
1421}
1422
1423static int __io_run_local_work(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1424{
1425 struct llist_node *node;
1426 unsigned int loops = 0;
1427 int ret = 0;
1428
1429 if (WARN_ON_ONCE(ctx->submitter_task != current))
1430 return -EEXIST;
1431 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1432 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1433again:
1434 /*
1435 * llists are in reverse order, flip it back the right way before
1436 * running the pending items.
1437 */
1438 node = llist_reverse_order(io_llist_xchg(&ctx->work_llist, NULL));
1439 while (node) {
1440 struct llist_node *next = node->next;
1441 struct io_kiocb *req = container_of(node, struct io_kiocb,
1442 io_task_work.node);
1443 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1444 INDIRECT_CALL_2(req->io_task_work.func,
1445 io_poll_task_func, io_req_rw_complete,
1446 req, ts);
1447 ret++;
1448 node = next;
1449 }
1450 loops++;
1451
1452 if (!llist_empty(&ctx->work_llist))
1453 goto again;
1454 if (ts->locked) {
1455 io_submit_flush_completions(ctx);
1456 if (!llist_empty(&ctx->work_llist))
1457 goto again;
1458 }
1459 trace_io_uring_local_work_run(ctx, ret, loops);
1460 return ret;
1461}
1462
1463static inline int io_run_local_work_locked(struct io_ring_ctx *ctx)
1464{
1465 struct io_tw_state ts = { .locked = true, };
1466 int ret;
1467
1468 if (llist_empty(&ctx->work_llist))
1469 return 0;
1470
1471 ret = __io_run_local_work(ctx, &ts);
1472 /* shouldn't happen! */
1473 if (WARN_ON_ONCE(!ts.locked))
1474 mutex_lock(&ctx->uring_lock);
1475 return ret;
1476}
1477
1478static int io_run_local_work(struct io_ring_ctx *ctx)
1479{
1480 struct io_tw_state ts = {};
1481 int ret;
1482
1483 ts.locked = mutex_trylock(&ctx->uring_lock);
1484 ret = __io_run_local_work(ctx, &ts);
1485 if (ts.locked)
1486 mutex_unlock(&ctx->uring_lock);
1487
1488 return ret;
1489}
1490
1491static void io_req_task_cancel(struct io_kiocb *req, struct io_tw_state *ts)
1492{
1493 io_tw_lock(req->ctx, ts);
1494 io_req_defer_failed(req, req->cqe.res);
1495}
1496
1497void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts)
1498{
1499 io_tw_lock(req->ctx, ts);
1500 /* req->task == current here, checking PF_EXITING is safe */
1501 if (unlikely(req->task->flags & PF_EXITING))
1502 io_req_defer_failed(req, -EFAULT);
1503 else if (req->flags & REQ_F_FORCE_ASYNC)
1504 io_queue_iowq(req, ts);
1505 else
1506 io_queue_sqe(req);
1507}
1508
1509void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1510{
1511 io_req_set_res(req, ret, 0);
1512 req->io_task_work.func = io_req_task_cancel;
1513 io_req_task_work_add(req);
1514}
1515
1516void io_req_task_queue(struct io_kiocb *req)
1517{
1518 req->io_task_work.func = io_req_task_submit;
1519 io_req_task_work_add(req);
1520}
1521
1522void io_queue_next(struct io_kiocb *req)
1523{
1524 struct io_kiocb *nxt = io_req_find_next(req);
1525
1526 if (nxt)
1527 io_req_task_queue(nxt);
1528}
1529
1530static void io_free_batch_list(struct io_ring_ctx *ctx,
1531 struct io_wq_work_node *node)
1532 __must_hold(&ctx->uring_lock)
1533{
1534 do {
1535 struct io_kiocb *req = container_of(node, struct io_kiocb,
1536 comp_list);
1537
1538 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1539 if (req->flags & REQ_F_REFCOUNT) {
1540 node = req->comp_list.next;
1541 if (!req_ref_put_and_test(req))
1542 continue;
1543 }
1544 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1545 struct async_poll *apoll = req->apoll;
1546
1547 if (apoll->double_poll)
1548 kfree(apoll->double_poll);
1549 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1550 kfree(apoll);
1551 req->flags &= ~REQ_F_POLLED;
1552 }
1553 if (req->flags & IO_REQ_LINK_FLAGS)
1554 io_queue_next(req);
1555 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1556 io_clean_op(req);
1557 }
1558 io_put_file(req);
1559
1560 io_req_put_rsrc_locked(req, ctx);
1561
1562 io_put_task(req->task);
1563 node = req->comp_list.next;
1564 io_req_add_to_cache(req, ctx);
1565 } while (node);
1566}
1567
1568void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1569 __must_hold(&ctx->uring_lock)
1570{
1571 struct io_submit_state *state = &ctx->submit_state;
1572 struct io_wq_work_node *node;
1573
1574 __io_cq_lock(ctx);
1575 /* must come first to preserve CQE ordering in failure cases */
1576 if (state->cqes_count)
1577 __io_flush_post_cqes(ctx);
1578 __wq_list_for_each(node, &state->compl_reqs) {
1579 struct io_kiocb *req = container_of(node, struct io_kiocb,
1580 comp_list);
1581
1582 if (!(req->flags & REQ_F_CQE_SKIP) &&
1583 unlikely(!io_fill_cqe_req(ctx, req))) {
1584 if (ctx->lockless_cq) {
1585 spin_lock(&ctx->completion_lock);
1586 io_req_cqe_overflow(req);
1587 spin_unlock(&ctx->completion_lock);
1588 } else {
1589 io_req_cqe_overflow(req);
1590 }
1591 }
1592 }
1593 __io_cq_unlock_post(ctx);
1594
1595 if (!wq_list_empty(&ctx->submit_state.compl_reqs)) {
1596 io_free_batch_list(ctx, state->compl_reqs.first);
1597 INIT_WQ_LIST(&state->compl_reqs);
1598 }
1599}
1600
1601static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1602{
1603 /* See comment at the top of this file */
1604 smp_rmb();
1605 return __io_cqring_events(ctx);
1606}
1607
1608/*
1609 * We can't just wait for polled events to come to us, we have to actively
1610 * find and complete them.
1611 */
1612static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1613{
1614 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1615 return;
1616
1617 mutex_lock(&ctx->uring_lock);
1618 while (!wq_list_empty(&ctx->iopoll_list)) {
1619 /* let it sleep and repeat later if can't complete a request */
1620 if (io_do_iopoll(ctx, true) == 0)
1621 break;
1622 /*
1623 * Ensure we allow local-to-the-cpu processing to take place,
1624 * in this case we need to ensure that we reap all events.
1625 * Also let task_work, etc. to progress by releasing the mutex
1626 */
1627 if (need_resched()) {
1628 mutex_unlock(&ctx->uring_lock);
1629 cond_resched();
1630 mutex_lock(&ctx->uring_lock);
1631 }
1632 }
1633 mutex_unlock(&ctx->uring_lock);
1634}
1635
1636static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1637{
1638 unsigned int nr_events = 0;
1639 unsigned long check_cq;
1640
1641 if (!io_allowed_run_tw(ctx))
1642 return -EEXIST;
1643
1644 check_cq = READ_ONCE(ctx->check_cq);
1645 if (unlikely(check_cq)) {
1646 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1647 __io_cqring_overflow_flush(ctx);
1648 /*
1649 * Similarly do not spin if we have not informed the user of any
1650 * dropped CQE.
1651 */
1652 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1653 return -EBADR;
1654 }
1655 /*
1656 * Don't enter poll loop if we already have events pending.
1657 * If we do, we can potentially be spinning for commands that
1658 * already triggered a CQE (eg in error).
1659 */
1660 if (io_cqring_events(ctx))
1661 return 0;
1662
1663 do {
1664 int ret = 0;
1665
1666 /*
1667 * If a submit got punted to a workqueue, we can have the
1668 * application entering polling for a command before it gets
1669 * issued. That app will hold the uring_lock for the duration
1670 * of the poll right here, so we need to take a breather every
1671 * now and then to ensure that the issue has a chance to add
1672 * the poll to the issued list. Otherwise we can spin here
1673 * forever, while the workqueue is stuck trying to acquire the
1674 * very same mutex.
1675 */
1676 if (wq_list_empty(&ctx->iopoll_list) ||
1677 io_task_work_pending(ctx)) {
1678 u32 tail = ctx->cached_cq_tail;
1679
1680 (void) io_run_local_work_locked(ctx);
1681
1682 if (task_work_pending(current) ||
1683 wq_list_empty(&ctx->iopoll_list)) {
1684 mutex_unlock(&ctx->uring_lock);
1685 io_run_task_work();
1686 mutex_lock(&ctx->uring_lock);
1687 }
1688 /* some requests don't go through iopoll_list */
1689 if (tail != ctx->cached_cq_tail ||
1690 wq_list_empty(&ctx->iopoll_list))
1691 break;
1692 }
1693 ret = io_do_iopoll(ctx, !min);
1694 if (unlikely(ret < 0))
1695 return ret;
1696
1697 if (task_sigpending(current))
1698 return -EINTR;
1699 if (need_resched())
1700 break;
1701
1702 nr_events += ret;
1703 } while (nr_events < min);
1704
1705 return 0;
1706}
1707
1708void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts)
1709{
1710 if (ts->locked)
1711 io_req_complete_defer(req);
1712 else
1713 io_req_complete_post(req, IO_URING_F_UNLOCKED);
1714}
1715
1716/*
1717 * After the iocb has been issued, it's safe to be found on the poll list.
1718 * Adding the kiocb to the list AFTER submission ensures that we don't
1719 * find it from a io_do_iopoll() thread before the issuer is done
1720 * accessing the kiocb cookie.
1721 */
1722static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1723{
1724 struct io_ring_ctx *ctx = req->ctx;
1725 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1726
1727 /* workqueue context doesn't hold uring_lock, grab it now */
1728 if (unlikely(needs_lock))
1729 mutex_lock(&ctx->uring_lock);
1730
1731 /*
1732 * Track whether we have multiple files in our lists. This will impact
1733 * how we do polling eventually, not spinning if we're on potentially
1734 * different devices.
1735 */
1736 if (wq_list_empty(&ctx->iopoll_list)) {
1737 ctx->poll_multi_queue = false;
1738 } else if (!ctx->poll_multi_queue) {
1739 struct io_kiocb *list_req;
1740
1741 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1742 comp_list);
1743 if (list_req->file != req->file)
1744 ctx->poll_multi_queue = true;
1745 }
1746
1747 /*
1748 * For fast devices, IO may have already completed. If it has, add
1749 * it to the front so we find it first.
1750 */
1751 if (READ_ONCE(req->iopoll_completed))
1752 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1753 else
1754 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1755
1756 if (unlikely(needs_lock)) {
1757 /*
1758 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1759 * in sq thread task context or in io worker task context. If
1760 * current task context is sq thread, we don't need to check
1761 * whether should wake up sq thread.
1762 */
1763 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1764 wq_has_sleeper(&ctx->sq_data->wait))
1765 wake_up(&ctx->sq_data->wait);
1766
1767 mutex_unlock(&ctx->uring_lock);
1768 }
1769}
1770
1771unsigned int io_file_get_flags(struct file *file)
1772{
1773 unsigned int res = 0;
1774
1775 if (S_ISREG(file_inode(file)->i_mode))
1776 res |= REQ_F_ISREG;
1777 if ((file->f_flags & O_NONBLOCK) || (file->f_mode & FMODE_NOWAIT))
1778 res |= REQ_F_SUPPORT_NOWAIT;
1779 return res;
1780}
1781
1782bool io_alloc_async_data(struct io_kiocb *req)
1783{
1784 WARN_ON_ONCE(!io_cold_defs[req->opcode].async_size);
1785 req->async_data = kmalloc(io_cold_defs[req->opcode].async_size, GFP_KERNEL);
1786 if (req->async_data) {
1787 req->flags |= REQ_F_ASYNC_DATA;
1788 return false;
1789 }
1790 return true;
1791}
1792
1793int io_req_prep_async(struct io_kiocb *req)
1794{
1795 const struct io_cold_def *cdef = &io_cold_defs[req->opcode];
1796 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1797
1798 /* assign early for deferred execution for non-fixed file */
1799 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE) && !req->file)
1800 req->file = io_file_get_normal(req, req->cqe.fd);
1801 if (!cdef->prep_async)
1802 return 0;
1803 if (WARN_ON_ONCE(req_has_async_data(req)))
1804 return -EFAULT;
1805 if (!def->manual_alloc) {
1806 if (io_alloc_async_data(req))
1807 return -EAGAIN;
1808 }
1809 return cdef->prep_async(req);
1810}
1811
1812static u32 io_get_sequence(struct io_kiocb *req)
1813{
1814 u32 seq = req->ctx->cached_sq_head;
1815 struct io_kiocb *cur;
1816
1817 /* need original cached_sq_head, but it was increased for each req */
1818 io_for_each_link(cur, req)
1819 seq--;
1820 return seq;
1821}
1822
1823static __cold void io_drain_req(struct io_kiocb *req)
1824 __must_hold(&ctx->uring_lock)
1825{
1826 struct io_ring_ctx *ctx = req->ctx;
1827 struct io_defer_entry *de;
1828 int ret;
1829 u32 seq = io_get_sequence(req);
1830
1831 /* Still need defer if there is pending req in defer list. */
1832 spin_lock(&ctx->completion_lock);
1833 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1834 spin_unlock(&ctx->completion_lock);
1835queue:
1836 ctx->drain_active = false;
1837 io_req_task_queue(req);
1838 return;
1839 }
1840 spin_unlock(&ctx->completion_lock);
1841
1842 io_prep_async_link(req);
1843 de = kmalloc(sizeof(*de), GFP_KERNEL);
1844 if (!de) {
1845 ret = -ENOMEM;
1846 io_req_defer_failed(req, ret);
1847 return;
1848 }
1849
1850 spin_lock(&ctx->completion_lock);
1851 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1852 spin_unlock(&ctx->completion_lock);
1853 kfree(de);
1854 goto queue;
1855 }
1856
1857 trace_io_uring_defer(req);
1858 de->req = req;
1859 de->seq = seq;
1860 list_add_tail(&de->list, &ctx->defer_list);
1861 spin_unlock(&ctx->completion_lock);
1862}
1863
1864static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def,
1865 unsigned int issue_flags)
1866{
1867 if (req->file || !def->needs_file)
1868 return true;
1869
1870 if (req->flags & REQ_F_FIXED_FILE)
1871 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1872 else
1873 req->file = io_file_get_normal(req, req->cqe.fd);
1874
1875 return !!req->file;
1876}
1877
1878static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1879{
1880 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1881 const struct cred *creds = NULL;
1882 int ret;
1883
1884 if (unlikely(!io_assign_file(req, def, issue_flags)))
1885 return -EBADF;
1886
1887 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1888 creds = override_creds(req->creds);
1889
1890 if (!def->audit_skip)
1891 audit_uring_entry(req->opcode);
1892
1893 ret = def->issue(req, issue_flags);
1894
1895 if (!def->audit_skip)
1896 audit_uring_exit(!ret, ret);
1897
1898 if (creds)
1899 revert_creds(creds);
1900
1901 if (ret == IOU_OK) {
1902 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1903 io_req_complete_defer(req);
1904 else
1905 io_req_complete_post(req, issue_flags);
1906
1907 return 0;
1908 }
1909
1910 if (ret == IOU_ISSUE_SKIP_COMPLETE) {
1911 ret = 0;
1912 io_arm_ltimeout(req);
1913
1914 /* If the op doesn't have a file, we're not polling for it */
1915 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
1916 io_iopoll_req_issued(req, issue_flags);
1917 }
1918 return ret;
1919}
1920
1921int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts)
1922{
1923 io_tw_lock(req->ctx, ts);
1924 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT|
1925 IO_URING_F_COMPLETE_DEFER);
1926}
1927
1928struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1929{
1930 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1931 struct io_kiocb *nxt = NULL;
1932
1933 if (req_ref_put_and_test(req)) {
1934 if (req->flags & IO_REQ_LINK_FLAGS)
1935 nxt = io_req_find_next(req);
1936 io_free_req(req);
1937 }
1938 return nxt ? &nxt->work : NULL;
1939}
1940
1941void io_wq_submit_work(struct io_wq_work *work)
1942{
1943 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1944 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1945 unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ;
1946 bool needs_poll = false;
1947 int ret = 0, err = -ECANCELED;
1948
1949 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1950 if (!(req->flags & REQ_F_REFCOUNT))
1951 __io_req_set_refcount(req, 2);
1952 else
1953 req_ref_get(req);
1954
1955 io_arm_ltimeout(req);
1956
1957 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1958 if (work->flags & IO_WQ_WORK_CANCEL) {
1959fail:
1960 io_req_task_queue_fail(req, err);
1961 return;
1962 }
1963 if (!io_assign_file(req, def, issue_flags)) {
1964 err = -EBADF;
1965 work->flags |= IO_WQ_WORK_CANCEL;
1966 goto fail;
1967 }
1968
1969 if (req->flags & REQ_F_FORCE_ASYNC) {
1970 bool opcode_poll = def->pollin || def->pollout;
1971
1972 if (opcode_poll && file_can_poll(req->file)) {
1973 needs_poll = true;
1974 issue_flags |= IO_URING_F_NONBLOCK;
1975 }
1976 }
1977
1978 do {
1979 ret = io_issue_sqe(req, issue_flags);
1980 if (ret != -EAGAIN)
1981 break;
1982
1983 /*
1984 * If REQ_F_NOWAIT is set, then don't wait or retry with
1985 * poll. -EAGAIN is final for that case.
1986 */
1987 if (req->flags & REQ_F_NOWAIT)
1988 break;
1989
1990 /*
1991 * We can get EAGAIN for iopolled IO even though we're
1992 * forcing a sync submission from here, since we can't
1993 * wait for request slots on the block side.
1994 */
1995 if (!needs_poll) {
1996 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1997 break;
1998 if (io_wq_worker_stopped())
1999 break;
2000 cond_resched();
2001 continue;
2002 }
2003
2004 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
2005 return;
2006 /* aborted or ready, in either case retry blocking */
2007 needs_poll = false;
2008 issue_flags &= ~IO_URING_F_NONBLOCK;
2009 } while (1);
2010
2011 /* avoid locking problems by failing it from a clean context */
2012 if (ret < 0)
2013 io_req_task_queue_fail(req, ret);
2014}
2015
2016inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
2017 unsigned int issue_flags)
2018{
2019 struct io_ring_ctx *ctx = req->ctx;
2020 struct io_fixed_file *slot;
2021 struct file *file = NULL;
2022
2023 io_ring_submit_lock(ctx, issue_flags);
2024
2025 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
2026 goto out;
2027 fd = array_index_nospec(fd, ctx->nr_user_files);
2028 slot = io_fixed_file_slot(&ctx->file_table, fd);
2029 if (!req->rsrc_node)
2030 __io_req_set_rsrc_node(req, ctx);
2031 req->flags |= io_slot_flags(slot);
2032 file = io_slot_file(slot);
2033out:
2034 io_ring_submit_unlock(ctx, issue_flags);
2035 return file;
2036}
2037
2038struct file *io_file_get_normal(struct io_kiocb *req, int fd)
2039{
2040 struct file *file = fget(fd);
2041
2042 trace_io_uring_file_get(req, fd);
2043
2044 /* we don't allow fixed io_uring files */
2045 if (file && io_is_uring_fops(file))
2046 io_req_track_inflight(req);
2047 return file;
2048}
2049
2050static void io_queue_async(struct io_kiocb *req, int ret)
2051 __must_hold(&req->ctx->uring_lock)
2052{
2053 struct io_kiocb *linked_timeout;
2054
2055 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
2056 io_req_defer_failed(req, ret);
2057 return;
2058 }
2059
2060 linked_timeout = io_prep_linked_timeout(req);
2061
2062 switch (io_arm_poll_handler(req, 0)) {
2063 case IO_APOLL_READY:
2064 io_kbuf_recycle(req, 0);
2065 io_req_task_queue(req);
2066 break;
2067 case IO_APOLL_ABORTED:
2068 io_kbuf_recycle(req, 0);
2069 io_queue_iowq(req, NULL);
2070 break;
2071 case IO_APOLL_OK:
2072 break;
2073 }
2074
2075 if (linked_timeout)
2076 io_queue_linked_timeout(linked_timeout);
2077}
2078
2079static inline void io_queue_sqe(struct io_kiocb *req)
2080 __must_hold(&req->ctx->uring_lock)
2081{
2082 int ret;
2083
2084 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
2085
2086 /*
2087 * We async punt it if the file wasn't marked NOWAIT, or if the file
2088 * doesn't support non-blocking read/write attempts
2089 */
2090 if (unlikely(ret))
2091 io_queue_async(req, ret);
2092}
2093
2094static void io_queue_sqe_fallback(struct io_kiocb *req)
2095 __must_hold(&req->ctx->uring_lock)
2096{
2097 if (unlikely(req->flags & REQ_F_FAIL)) {
2098 /*
2099 * We don't submit, fail them all, for that replace hardlinks
2100 * with normal links. Extra REQ_F_LINK is tolerated.
2101 */
2102 req->flags &= ~REQ_F_HARDLINK;
2103 req->flags |= REQ_F_LINK;
2104 io_req_defer_failed(req, req->cqe.res);
2105 } else {
2106 int ret = io_req_prep_async(req);
2107
2108 if (unlikely(ret)) {
2109 io_req_defer_failed(req, ret);
2110 return;
2111 }
2112
2113 if (unlikely(req->ctx->drain_active))
2114 io_drain_req(req);
2115 else
2116 io_queue_iowq(req, NULL);
2117 }
2118}
2119
2120/*
2121 * Check SQE restrictions (opcode and flags).
2122 *
2123 * Returns 'true' if SQE is allowed, 'false' otherwise.
2124 */
2125static inline bool io_check_restriction(struct io_ring_ctx *ctx,
2126 struct io_kiocb *req,
2127 unsigned int sqe_flags)
2128{
2129 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
2130 return false;
2131
2132 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
2133 ctx->restrictions.sqe_flags_required)
2134 return false;
2135
2136 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
2137 ctx->restrictions.sqe_flags_required))
2138 return false;
2139
2140 return true;
2141}
2142
2143static void io_init_req_drain(struct io_kiocb *req)
2144{
2145 struct io_ring_ctx *ctx = req->ctx;
2146 struct io_kiocb *head = ctx->submit_state.link.head;
2147
2148 ctx->drain_active = true;
2149 if (head) {
2150 /*
2151 * If we need to drain a request in the middle of a link, drain
2152 * the head request and the next request/link after the current
2153 * link. Considering sequential execution of links,
2154 * REQ_F_IO_DRAIN will be maintained for every request of our
2155 * link.
2156 */
2157 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2158 ctx->drain_next = true;
2159 }
2160}
2161
2162static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
2163 const struct io_uring_sqe *sqe)
2164 __must_hold(&ctx->uring_lock)
2165{
2166 const struct io_issue_def *def;
2167 unsigned int sqe_flags;
2168 int personality;
2169 u8 opcode;
2170
2171 /* req is partially pre-initialised, see io_preinit_req() */
2172 req->opcode = opcode = READ_ONCE(sqe->opcode);
2173 /* same numerical values with corresponding REQ_F_*, safe to copy */
2174 req->flags = sqe_flags = READ_ONCE(sqe->flags);
2175 req->cqe.user_data = READ_ONCE(sqe->user_data);
2176 req->file = NULL;
2177 req->rsrc_node = NULL;
2178 req->task = current;
2179
2180 if (unlikely(opcode >= IORING_OP_LAST)) {
2181 req->opcode = 0;
2182 return -EINVAL;
2183 }
2184 def = &io_issue_defs[opcode];
2185 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2186 /* enforce forwards compatibility on users */
2187 if (sqe_flags & ~SQE_VALID_FLAGS)
2188 return -EINVAL;
2189 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2190 if (!def->buffer_select)
2191 return -EOPNOTSUPP;
2192 req->buf_index = READ_ONCE(sqe->buf_group);
2193 }
2194 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2195 ctx->drain_disabled = true;
2196 if (sqe_flags & IOSQE_IO_DRAIN) {
2197 if (ctx->drain_disabled)
2198 return -EOPNOTSUPP;
2199 io_init_req_drain(req);
2200 }
2201 }
2202 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2203 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2204 return -EACCES;
2205 /* knock it to the slow queue path, will be drained there */
2206 if (ctx->drain_active)
2207 req->flags |= REQ_F_FORCE_ASYNC;
2208 /* if there is no link, we're at "next" request and need to drain */
2209 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2210 ctx->drain_next = false;
2211 ctx->drain_active = true;
2212 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2213 }
2214 }
2215
2216 if (!def->ioprio && sqe->ioprio)
2217 return -EINVAL;
2218 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2219 return -EINVAL;
2220
2221 if (def->needs_file) {
2222 struct io_submit_state *state = &ctx->submit_state;
2223
2224 req->cqe.fd = READ_ONCE(sqe->fd);
2225
2226 /*
2227 * Plug now if we have more than 2 IO left after this, and the
2228 * target is potentially a read/write to block based storage.
2229 */
2230 if (state->need_plug && def->plug) {
2231 state->plug_started = true;
2232 state->need_plug = false;
2233 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2234 }
2235 }
2236
2237 personality = READ_ONCE(sqe->personality);
2238 if (personality) {
2239 int ret;
2240
2241 req->creds = xa_load(&ctx->personalities, personality);
2242 if (!req->creds)
2243 return -EINVAL;
2244 get_cred(req->creds);
2245 ret = security_uring_override_creds(req->creds);
2246 if (ret) {
2247 put_cred(req->creds);
2248 return ret;
2249 }
2250 req->flags |= REQ_F_CREDS;
2251 }
2252
2253 return def->prep(req, sqe);
2254}
2255
2256static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2257 struct io_kiocb *req, int ret)
2258{
2259 struct io_ring_ctx *ctx = req->ctx;
2260 struct io_submit_link *link = &ctx->submit_state.link;
2261 struct io_kiocb *head = link->head;
2262
2263 trace_io_uring_req_failed(sqe, req, ret);
2264
2265 /*
2266 * Avoid breaking links in the middle as it renders links with SQPOLL
2267 * unusable. Instead of failing eagerly, continue assembling the link if
2268 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2269 * should find the flag and handle the rest.
2270 */
2271 req_fail_link_node(req, ret);
2272 if (head && !(head->flags & REQ_F_FAIL))
2273 req_fail_link_node(head, -ECANCELED);
2274
2275 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2276 if (head) {
2277 link->last->link = req;
2278 link->head = NULL;
2279 req = head;
2280 }
2281 io_queue_sqe_fallback(req);
2282 return ret;
2283 }
2284
2285 if (head)
2286 link->last->link = req;
2287 else
2288 link->head = req;
2289 link->last = req;
2290 return 0;
2291}
2292
2293static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2294 const struct io_uring_sqe *sqe)
2295 __must_hold(&ctx->uring_lock)
2296{
2297 struct io_submit_link *link = &ctx->submit_state.link;
2298 int ret;
2299
2300 ret = io_init_req(ctx, req, sqe);
2301 if (unlikely(ret))
2302 return io_submit_fail_init(sqe, req, ret);
2303
2304 trace_io_uring_submit_req(req);
2305
2306 /*
2307 * If we already have a head request, queue this one for async
2308 * submittal once the head completes. If we don't have a head but
2309 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2310 * submitted sync once the chain is complete. If none of those
2311 * conditions are true (normal request), then just queue it.
2312 */
2313 if (unlikely(link->head)) {
2314 ret = io_req_prep_async(req);
2315 if (unlikely(ret))
2316 return io_submit_fail_init(sqe, req, ret);
2317
2318 trace_io_uring_link(req, link->head);
2319 link->last->link = req;
2320 link->last = req;
2321
2322 if (req->flags & IO_REQ_LINK_FLAGS)
2323 return 0;
2324 /* last request of the link, flush it */
2325 req = link->head;
2326 link->head = NULL;
2327 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2328 goto fallback;
2329
2330 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2331 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2332 if (req->flags & IO_REQ_LINK_FLAGS) {
2333 link->head = req;
2334 link->last = req;
2335 } else {
2336fallback:
2337 io_queue_sqe_fallback(req);
2338 }
2339 return 0;
2340 }
2341
2342 io_queue_sqe(req);
2343 return 0;
2344}
2345
2346/*
2347 * Batched submission is done, ensure local IO is flushed out.
2348 */
2349static void io_submit_state_end(struct io_ring_ctx *ctx)
2350{
2351 struct io_submit_state *state = &ctx->submit_state;
2352
2353 if (unlikely(state->link.head))
2354 io_queue_sqe_fallback(state->link.head);
2355 /* flush only after queuing links as they can generate completions */
2356 io_submit_flush_completions(ctx);
2357 if (state->plug_started)
2358 blk_finish_plug(&state->plug);
2359}
2360
2361/*
2362 * Start submission side cache.
2363 */
2364static void io_submit_state_start(struct io_submit_state *state,
2365 unsigned int max_ios)
2366{
2367 state->plug_started = false;
2368 state->need_plug = max_ios > 2;
2369 state->submit_nr = max_ios;
2370 /* set only head, no need to init link_last in advance */
2371 state->link.head = NULL;
2372}
2373
2374static void io_commit_sqring(struct io_ring_ctx *ctx)
2375{
2376 struct io_rings *rings = ctx->rings;
2377
2378 /*
2379 * Ensure any loads from the SQEs are done at this point,
2380 * since once we write the new head, the application could
2381 * write new data to them.
2382 */
2383 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2384}
2385
2386/*
2387 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2388 * that is mapped by userspace. This means that care needs to be taken to
2389 * ensure that reads are stable, as we cannot rely on userspace always
2390 * being a good citizen. If members of the sqe are validated and then later
2391 * used, it's important that those reads are done through READ_ONCE() to
2392 * prevent a re-load down the line.
2393 */
2394static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe)
2395{
2396 unsigned mask = ctx->sq_entries - 1;
2397 unsigned head = ctx->cached_sq_head++ & mask;
2398
2399 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY)) {
2400 head = READ_ONCE(ctx->sq_array[head]);
2401 if (unlikely(head >= ctx->sq_entries)) {
2402 /* drop invalid entries */
2403 spin_lock(&ctx->completion_lock);
2404 ctx->cq_extra--;
2405 spin_unlock(&ctx->completion_lock);
2406 WRITE_ONCE(ctx->rings->sq_dropped,
2407 READ_ONCE(ctx->rings->sq_dropped) + 1);
2408 return false;
2409 }
2410 }
2411
2412 /*
2413 * The cached sq head (or cq tail) serves two purposes:
2414 *
2415 * 1) allows us to batch the cost of updating the user visible
2416 * head updates.
2417 * 2) allows the kernel side to track the head on its own, even
2418 * though the application is the one updating it.
2419 */
2420
2421 /* double index for 128-byte SQEs, twice as long */
2422 if (ctx->flags & IORING_SETUP_SQE128)
2423 head <<= 1;
2424 *sqe = &ctx->sq_sqes[head];
2425 return true;
2426}
2427
2428int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2429 __must_hold(&ctx->uring_lock)
2430{
2431 unsigned int entries = io_sqring_entries(ctx);
2432 unsigned int left;
2433 int ret;
2434
2435 if (unlikely(!entries))
2436 return 0;
2437 /* make sure SQ entry isn't read before tail */
2438 ret = left = min(nr, entries);
2439 io_get_task_refs(left);
2440 io_submit_state_start(&ctx->submit_state, left);
2441
2442 do {
2443 const struct io_uring_sqe *sqe;
2444 struct io_kiocb *req;
2445
2446 if (unlikely(!io_alloc_req(ctx, &req)))
2447 break;
2448 if (unlikely(!io_get_sqe(ctx, &sqe))) {
2449 io_req_add_to_cache(req, ctx);
2450 break;
2451 }
2452
2453 /*
2454 * Continue submitting even for sqe failure if the
2455 * ring was setup with IORING_SETUP_SUBMIT_ALL
2456 */
2457 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2458 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2459 left--;
2460 break;
2461 }
2462 } while (--left);
2463
2464 if (unlikely(left)) {
2465 ret -= left;
2466 /* try again if it submitted nothing and can't allocate a req */
2467 if (!ret && io_req_cache_empty(ctx))
2468 ret = -EAGAIN;
2469 current->io_uring->cached_refs += left;
2470 }
2471
2472 io_submit_state_end(ctx);
2473 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2474 io_commit_sqring(ctx);
2475 return ret;
2476}
2477
2478struct io_wait_queue {
2479 struct wait_queue_entry wq;
2480 struct io_ring_ctx *ctx;
2481 unsigned cq_tail;
2482 unsigned nr_timeouts;
2483 ktime_t timeout;
2484};
2485
2486static inline bool io_has_work(struct io_ring_ctx *ctx)
2487{
2488 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
2489 !llist_empty(&ctx->work_llist);
2490}
2491
2492static inline bool io_should_wake(struct io_wait_queue *iowq)
2493{
2494 struct io_ring_ctx *ctx = iowq->ctx;
2495 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
2496
2497 /*
2498 * Wake up if we have enough events, or if a timeout occurred since we
2499 * started waiting. For timeouts, we always want to return to userspace,
2500 * regardless of event count.
2501 */
2502 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2503}
2504
2505static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2506 int wake_flags, void *key)
2507{
2508 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq);
2509
2510 /*
2511 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2512 * the task, and the next invocation will do it.
2513 */
2514 if (io_should_wake(iowq) || io_has_work(iowq->ctx))
2515 return autoremove_wake_function(curr, mode, wake_flags, key);
2516 return -1;
2517}
2518
2519int io_run_task_work_sig(struct io_ring_ctx *ctx)
2520{
2521 if (!llist_empty(&ctx->work_llist)) {
2522 __set_current_state(TASK_RUNNING);
2523 if (io_run_local_work(ctx) > 0)
2524 return 0;
2525 }
2526 if (io_run_task_work() > 0)
2527 return 0;
2528 if (task_sigpending(current))
2529 return -EINTR;
2530 return 0;
2531}
2532
2533static bool current_pending_io(void)
2534{
2535 struct io_uring_task *tctx = current->io_uring;
2536
2537 if (!tctx)
2538 return false;
2539 return percpu_counter_read_positive(&tctx->inflight);
2540}
2541
2542/* when returns >0, the caller should retry */
2543static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2544 struct io_wait_queue *iowq)
2545{
2546 int io_wait, ret;
2547
2548 if (unlikely(READ_ONCE(ctx->check_cq)))
2549 return 1;
2550 if (unlikely(!llist_empty(&ctx->work_llist)))
2551 return 1;
2552 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL)))
2553 return 1;
2554 if (unlikely(task_sigpending(current)))
2555 return -EINTR;
2556 if (unlikely(io_should_wake(iowq)))
2557 return 0;
2558
2559 /*
2560 * Mark us as being in io_wait if we have pending requests, so cpufreq
2561 * can take into account that the task is waiting for IO - turns out
2562 * to be important for low QD IO.
2563 */
2564 io_wait = current->in_iowait;
2565 if (current_pending_io())
2566 current->in_iowait = 1;
2567 ret = 0;
2568 if (iowq->timeout == KTIME_MAX)
2569 schedule();
2570 else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS))
2571 ret = -ETIME;
2572 current->in_iowait = io_wait;
2573 return ret;
2574}
2575
2576/*
2577 * Wait until events become available, if we don't already have some. The
2578 * application must reap them itself, as they reside on the shared cq ring.
2579 */
2580static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2581 const sigset_t __user *sig, size_t sigsz,
2582 struct __kernel_timespec __user *uts)
2583{
2584 struct io_wait_queue iowq;
2585 struct io_rings *rings = ctx->rings;
2586 int ret;
2587
2588 if (!io_allowed_run_tw(ctx))
2589 return -EEXIST;
2590 if (!llist_empty(&ctx->work_llist))
2591 io_run_local_work(ctx);
2592 io_run_task_work();
2593 io_cqring_overflow_flush(ctx);
2594 /* if user messes with these they will just get an early return */
2595 if (__io_cqring_events_user(ctx) >= min_events)
2596 return 0;
2597
2598 if (sig) {
2599#ifdef CONFIG_COMPAT
2600 if (in_compat_syscall())
2601 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2602 sigsz);
2603 else
2604#endif
2605 ret = set_user_sigmask(sig, sigsz);
2606
2607 if (ret)
2608 return ret;
2609 }
2610
2611 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2612 iowq.wq.private = current;
2613 INIT_LIST_HEAD(&iowq.wq.entry);
2614 iowq.ctx = ctx;
2615 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2616 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2617 iowq.timeout = KTIME_MAX;
2618
2619 if (uts) {
2620 struct timespec64 ts;
2621
2622 if (get_timespec64(&ts, uts))
2623 return -EFAULT;
2624 iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2625 }
2626
2627 trace_io_uring_cqring_wait(ctx, min_events);
2628 do {
2629 unsigned long check_cq;
2630
2631 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
2632 int nr_wait = (int) iowq.cq_tail - READ_ONCE(ctx->rings->cq.tail);
2633
2634 atomic_set(&ctx->cq_wait_nr, nr_wait);
2635 set_current_state(TASK_INTERRUPTIBLE);
2636 } else {
2637 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2638 TASK_INTERRUPTIBLE);
2639 }
2640
2641 ret = io_cqring_wait_schedule(ctx, &iowq);
2642 __set_current_state(TASK_RUNNING);
2643 atomic_set(&ctx->cq_wait_nr, IO_CQ_WAKE_INIT);
2644
2645 /*
2646 * Run task_work after scheduling and before io_should_wake().
2647 * If we got woken because of task_work being processed, run it
2648 * now rather than let the caller do another wait loop.
2649 */
2650 io_run_task_work();
2651 if (!llist_empty(&ctx->work_llist))
2652 io_run_local_work(ctx);
2653
2654 /*
2655 * Non-local task_work will be run on exit to userspace, but
2656 * if we're using DEFER_TASKRUN, then we could have waited
2657 * with a timeout for a number of requests. If the timeout
2658 * hits, we could have some requests ready to process. Ensure
2659 * this break is _after_ we have run task_work, to avoid
2660 * deferring running potentially pending requests until the
2661 * next time we wait for events.
2662 */
2663 if (ret < 0)
2664 break;
2665
2666 check_cq = READ_ONCE(ctx->check_cq);
2667 if (unlikely(check_cq)) {
2668 /* let the caller flush overflows, retry */
2669 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2670 io_cqring_do_overflow_flush(ctx);
2671 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) {
2672 ret = -EBADR;
2673 break;
2674 }
2675 }
2676
2677 if (io_should_wake(&iowq)) {
2678 ret = 0;
2679 break;
2680 }
2681 cond_resched();
2682 } while (1);
2683
2684 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
2685 finish_wait(&ctx->cq_wait, &iowq.wq);
2686 restore_saved_sigmask_unless(ret == -EINTR);
2687
2688 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2689}
2690
2691void io_mem_free(void *ptr)
2692{
2693 if (!ptr)
2694 return;
2695
2696 folio_put(virt_to_folio(ptr));
2697}
2698
2699static void io_pages_free(struct page ***pages, int npages)
2700{
2701 struct page **page_array;
2702 int i;
2703
2704 if (!pages)
2705 return;
2706
2707 page_array = *pages;
2708 if (!page_array)
2709 return;
2710
2711 for (i = 0; i < npages; i++)
2712 unpin_user_page(page_array[i]);
2713 kvfree(page_array);
2714 *pages = NULL;
2715}
2716
2717static void *__io_uaddr_map(struct page ***pages, unsigned short *npages,
2718 unsigned long uaddr, size_t size)
2719{
2720 struct page **page_array;
2721 unsigned int nr_pages;
2722 void *page_addr;
2723 int ret, i;
2724
2725 *npages = 0;
2726
2727 if (uaddr & (PAGE_SIZE - 1) || !size)
2728 return ERR_PTR(-EINVAL);
2729
2730 nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
2731 if (nr_pages > USHRT_MAX)
2732 return ERR_PTR(-EINVAL);
2733 page_array = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
2734 if (!page_array)
2735 return ERR_PTR(-ENOMEM);
2736
2737 ret = pin_user_pages_fast(uaddr, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
2738 page_array);
2739 if (ret != nr_pages) {
2740err:
2741 io_pages_free(&page_array, ret > 0 ? ret : 0);
2742 return ret < 0 ? ERR_PTR(ret) : ERR_PTR(-EFAULT);
2743 }
2744
2745 page_addr = page_address(page_array[0]);
2746 for (i = 0; i < nr_pages; i++) {
2747 ret = -EINVAL;
2748
2749 /*
2750 * Can't support mapping user allocated ring memory on 32-bit
2751 * archs where it could potentially reside in highmem. Just
2752 * fail those with -EINVAL, just like we did on kernels that
2753 * didn't support this feature.
2754 */
2755 if (PageHighMem(page_array[i]))
2756 goto err;
2757
2758 /*
2759 * No support for discontig pages for now, should either be a
2760 * single normal page, or a huge page. Later on we can add
2761 * support for remapping discontig pages, for now we will
2762 * just fail them with EINVAL.
2763 */
2764 if (page_address(page_array[i]) != page_addr)
2765 goto err;
2766 page_addr += PAGE_SIZE;
2767 }
2768
2769 *pages = page_array;
2770 *npages = nr_pages;
2771 return page_to_virt(page_array[0]);
2772}
2773
2774static void *io_rings_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2775 size_t size)
2776{
2777 return __io_uaddr_map(&ctx->ring_pages, &ctx->n_ring_pages, uaddr,
2778 size);
2779}
2780
2781static void *io_sqes_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2782 size_t size)
2783{
2784 return __io_uaddr_map(&ctx->sqe_pages, &ctx->n_sqe_pages, uaddr,
2785 size);
2786}
2787
2788static void io_rings_free(struct io_ring_ctx *ctx)
2789{
2790 if (!(ctx->flags & IORING_SETUP_NO_MMAP)) {
2791 io_mem_free(ctx->rings);
2792 io_mem_free(ctx->sq_sqes);
2793 ctx->rings = NULL;
2794 ctx->sq_sqes = NULL;
2795 } else {
2796 io_pages_free(&ctx->ring_pages, ctx->n_ring_pages);
2797 ctx->n_ring_pages = 0;
2798 io_pages_free(&ctx->sqe_pages, ctx->n_sqe_pages);
2799 ctx->n_sqe_pages = 0;
2800 }
2801}
2802
2803void *io_mem_alloc(size_t size)
2804{
2805 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2806 void *ret;
2807
2808 ret = (void *) __get_free_pages(gfp, get_order(size));
2809 if (ret)
2810 return ret;
2811 return ERR_PTR(-ENOMEM);
2812}
2813
2814static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2815 unsigned int cq_entries, size_t *sq_offset)
2816{
2817 struct io_rings *rings;
2818 size_t off, sq_array_size;
2819
2820 off = struct_size(rings, cqes, cq_entries);
2821 if (off == SIZE_MAX)
2822 return SIZE_MAX;
2823 if (ctx->flags & IORING_SETUP_CQE32) {
2824 if (check_shl_overflow(off, 1, &off))
2825 return SIZE_MAX;
2826 }
2827
2828#ifdef CONFIG_SMP
2829 off = ALIGN(off, SMP_CACHE_BYTES);
2830 if (off == 0)
2831 return SIZE_MAX;
2832#endif
2833
2834 if (ctx->flags & IORING_SETUP_NO_SQARRAY) {
2835 if (sq_offset)
2836 *sq_offset = SIZE_MAX;
2837 return off;
2838 }
2839
2840 if (sq_offset)
2841 *sq_offset = off;
2842
2843 sq_array_size = array_size(sizeof(u32), sq_entries);
2844 if (sq_array_size == SIZE_MAX)
2845 return SIZE_MAX;
2846
2847 if (check_add_overflow(off, sq_array_size, &off))
2848 return SIZE_MAX;
2849
2850 return off;
2851}
2852
2853static void io_req_caches_free(struct io_ring_ctx *ctx)
2854{
2855 struct io_kiocb *req;
2856 int nr = 0;
2857
2858 mutex_lock(&ctx->uring_lock);
2859 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
2860
2861 while (!io_req_cache_empty(ctx)) {
2862 req = io_extract_req(ctx);
2863 kmem_cache_free(req_cachep, req);
2864 nr++;
2865 }
2866 if (nr)
2867 percpu_ref_put_many(&ctx->refs, nr);
2868 mutex_unlock(&ctx->uring_lock);
2869}
2870
2871static void io_rsrc_node_cache_free(struct io_cache_entry *entry)
2872{
2873 kfree(container_of(entry, struct io_rsrc_node, cache));
2874}
2875
2876static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2877{
2878 io_sq_thread_finish(ctx);
2879 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2880 if (WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)))
2881 return;
2882
2883 mutex_lock(&ctx->uring_lock);
2884 if (ctx->buf_data)
2885 __io_sqe_buffers_unregister(ctx);
2886 if (ctx->file_data)
2887 __io_sqe_files_unregister(ctx);
2888 io_cqring_overflow_kill(ctx);
2889 io_eventfd_unregister(ctx);
2890 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2891 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2892 io_futex_cache_free(ctx);
2893 io_destroy_buffers(ctx);
2894 mutex_unlock(&ctx->uring_lock);
2895 if (ctx->sq_creds)
2896 put_cred(ctx->sq_creds);
2897 if (ctx->submitter_task)
2898 put_task_struct(ctx->submitter_task);
2899
2900 /* there are no registered resources left, nobody uses it */
2901 if (ctx->rsrc_node)
2902 io_rsrc_node_destroy(ctx, ctx->rsrc_node);
2903
2904 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2905 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2906
2907 io_alloc_cache_free(&ctx->rsrc_node_cache, io_rsrc_node_cache_free);
2908 if (ctx->mm_account) {
2909 mmdrop(ctx->mm_account);
2910 ctx->mm_account = NULL;
2911 }
2912 io_rings_free(ctx);
2913 io_kbuf_mmap_list_free(ctx);
2914
2915 percpu_ref_exit(&ctx->refs);
2916 free_uid(ctx->user);
2917 io_req_caches_free(ctx);
2918 if (ctx->hash_map)
2919 io_wq_put_hash(ctx->hash_map);
2920 kfree(ctx->cancel_table.hbs);
2921 kfree(ctx->cancel_table_locked.hbs);
2922 kfree(ctx->io_bl);
2923 xa_destroy(&ctx->io_bl_xa);
2924 kfree(ctx);
2925}
2926
2927static __cold void io_activate_pollwq_cb(struct callback_head *cb)
2928{
2929 struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx,
2930 poll_wq_task_work);
2931
2932 mutex_lock(&ctx->uring_lock);
2933 ctx->poll_activated = true;
2934 mutex_unlock(&ctx->uring_lock);
2935
2936 /*
2937 * Wake ups for some events between start of polling and activation
2938 * might've been lost due to loose synchronisation.
2939 */
2940 wake_up_all(&ctx->poll_wq);
2941 percpu_ref_put(&ctx->refs);
2942}
2943
2944__cold void io_activate_pollwq(struct io_ring_ctx *ctx)
2945{
2946 spin_lock(&ctx->completion_lock);
2947 /* already activated or in progress */
2948 if (ctx->poll_activated || ctx->poll_wq_task_work.func)
2949 goto out;
2950 if (WARN_ON_ONCE(!ctx->task_complete))
2951 goto out;
2952 if (!ctx->submitter_task)
2953 goto out;
2954 /*
2955 * with ->submitter_task only the submitter task completes requests, we
2956 * only need to sync with it, which is done by injecting a tw
2957 */
2958 init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb);
2959 percpu_ref_get(&ctx->refs);
2960 if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL))
2961 percpu_ref_put(&ctx->refs);
2962out:
2963 spin_unlock(&ctx->completion_lock);
2964}
2965
2966static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2967{
2968 struct io_ring_ctx *ctx = file->private_data;
2969 __poll_t mask = 0;
2970
2971 if (unlikely(!ctx->poll_activated))
2972 io_activate_pollwq(ctx);
2973
2974 poll_wait(file, &ctx->poll_wq, wait);
2975 /*
2976 * synchronizes with barrier from wq_has_sleeper call in
2977 * io_commit_cqring
2978 */
2979 smp_rmb();
2980 if (!io_sqring_full(ctx))
2981 mask |= EPOLLOUT | EPOLLWRNORM;
2982
2983 /*
2984 * Don't flush cqring overflow list here, just do a simple check.
2985 * Otherwise there could possible be ABBA deadlock:
2986 * CPU0 CPU1
2987 * ---- ----
2988 * lock(&ctx->uring_lock);
2989 * lock(&ep->mtx);
2990 * lock(&ctx->uring_lock);
2991 * lock(&ep->mtx);
2992 *
2993 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2994 * pushes them to do the flush.
2995 */
2996
2997 if (__io_cqring_events_user(ctx) || io_has_work(ctx))
2998 mask |= EPOLLIN | EPOLLRDNORM;
2999
3000 return mask;
3001}
3002
3003struct io_tctx_exit {
3004 struct callback_head task_work;
3005 struct completion completion;
3006 struct io_ring_ctx *ctx;
3007};
3008
3009static __cold void io_tctx_exit_cb(struct callback_head *cb)
3010{
3011 struct io_uring_task *tctx = current->io_uring;
3012 struct io_tctx_exit *work;
3013
3014 work = container_of(cb, struct io_tctx_exit, task_work);
3015 /*
3016 * When @in_cancel, we're in cancellation and it's racy to remove the
3017 * node. It'll be removed by the end of cancellation, just ignore it.
3018 * tctx can be NULL if the queueing of this task_work raced with
3019 * work cancelation off the exec path.
3020 */
3021 if (tctx && !atomic_read(&tctx->in_cancel))
3022 io_uring_del_tctx_node((unsigned long)work->ctx);
3023 complete(&work->completion);
3024}
3025
3026static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
3027{
3028 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3029
3030 return req->ctx == data;
3031}
3032
3033static __cold void io_ring_exit_work(struct work_struct *work)
3034{
3035 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
3036 unsigned long timeout = jiffies + HZ * 60 * 5;
3037 unsigned long interval = HZ / 20;
3038 struct io_tctx_exit exit;
3039 struct io_tctx_node *node;
3040 int ret;
3041
3042 /*
3043 * If we're doing polled IO and end up having requests being
3044 * submitted async (out-of-line), then completions can come in while
3045 * we're waiting for refs to drop. We need to reap these manually,
3046 * as nobody else will be looking for them.
3047 */
3048 do {
3049 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
3050 mutex_lock(&ctx->uring_lock);
3051 io_cqring_overflow_kill(ctx);
3052 mutex_unlock(&ctx->uring_lock);
3053 }
3054
3055 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3056 io_move_task_work_from_local(ctx);
3057
3058 while (io_uring_try_cancel_requests(ctx, NULL, true))
3059 cond_resched();
3060
3061 if (ctx->sq_data) {
3062 struct io_sq_data *sqd = ctx->sq_data;
3063 struct task_struct *tsk;
3064
3065 io_sq_thread_park(sqd);
3066 tsk = sqd->thread;
3067 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
3068 io_wq_cancel_cb(tsk->io_uring->io_wq,
3069 io_cancel_ctx_cb, ctx, true);
3070 io_sq_thread_unpark(sqd);
3071 }
3072
3073 io_req_caches_free(ctx);
3074
3075 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
3076 /* there is little hope left, don't run it too often */
3077 interval = HZ * 60;
3078 }
3079 /*
3080 * This is really an uninterruptible wait, as it has to be
3081 * complete. But it's also run from a kworker, which doesn't
3082 * take signals, so it's fine to make it interruptible. This
3083 * avoids scenarios where we knowingly can wait much longer
3084 * on completions, for example if someone does a SIGSTOP on
3085 * a task that needs to finish task_work to make this loop
3086 * complete. That's a synthetic situation that should not
3087 * cause a stuck task backtrace, and hence a potential panic
3088 * on stuck tasks if that is enabled.
3089 */
3090 } while (!wait_for_completion_interruptible_timeout(&ctx->ref_comp, interval));
3091
3092 init_completion(&exit.completion);
3093 init_task_work(&exit.task_work, io_tctx_exit_cb);
3094 exit.ctx = ctx;
3095
3096 mutex_lock(&ctx->uring_lock);
3097 while (!list_empty(&ctx->tctx_list)) {
3098 WARN_ON_ONCE(time_after(jiffies, timeout));
3099
3100 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
3101 ctx_node);
3102 /* don't spin on a single task if cancellation failed */
3103 list_rotate_left(&ctx->tctx_list);
3104 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
3105 if (WARN_ON_ONCE(ret))
3106 continue;
3107
3108 mutex_unlock(&ctx->uring_lock);
3109 /*
3110 * See comment above for
3111 * wait_for_completion_interruptible_timeout() on why this
3112 * wait is marked as interruptible.
3113 */
3114 wait_for_completion_interruptible(&exit.completion);
3115 mutex_lock(&ctx->uring_lock);
3116 }
3117 mutex_unlock(&ctx->uring_lock);
3118 spin_lock(&ctx->completion_lock);
3119 spin_unlock(&ctx->completion_lock);
3120
3121 /* pairs with RCU read section in io_req_local_work_add() */
3122 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3123 synchronize_rcu();
3124
3125 io_ring_ctx_free(ctx);
3126}
3127
3128static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
3129{
3130 unsigned long index;
3131 struct creds *creds;
3132
3133 mutex_lock(&ctx->uring_lock);
3134 percpu_ref_kill(&ctx->refs);
3135 xa_for_each(&ctx->personalities, index, creds)
3136 io_unregister_personality(ctx, index);
3137 if (ctx->rings)
3138 io_poll_remove_all(ctx, NULL, true);
3139 mutex_unlock(&ctx->uring_lock);
3140
3141 /*
3142 * If we failed setting up the ctx, we might not have any rings
3143 * and therefore did not submit any requests
3144 */
3145 if (ctx->rings)
3146 io_kill_timeouts(ctx, NULL, true);
3147
3148 flush_delayed_work(&ctx->fallback_work);
3149
3150 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
3151 /*
3152 * Use system_unbound_wq to avoid spawning tons of event kworkers
3153 * if we're exiting a ton of rings at the same time. It just adds
3154 * noise and overhead, there's no discernable change in runtime
3155 * over using system_wq.
3156 */
3157 queue_work(system_unbound_wq, &ctx->exit_work);
3158}
3159
3160static int io_uring_release(struct inode *inode, struct file *file)
3161{
3162 struct io_ring_ctx *ctx = file->private_data;
3163
3164 file->private_data = NULL;
3165 io_ring_ctx_wait_and_kill(ctx);
3166 return 0;
3167}
3168
3169struct io_task_cancel {
3170 struct task_struct *task;
3171 bool all;
3172};
3173
3174static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
3175{
3176 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3177 struct io_task_cancel *cancel = data;
3178
3179 return io_match_task_safe(req, cancel->task, cancel->all);
3180}
3181
3182static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
3183 struct task_struct *task,
3184 bool cancel_all)
3185{
3186 struct io_defer_entry *de;
3187 LIST_HEAD(list);
3188
3189 spin_lock(&ctx->completion_lock);
3190 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
3191 if (io_match_task_safe(de->req, task, cancel_all)) {
3192 list_cut_position(&list, &ctx->defer_list, &de->list);
3193 break;
3194 }
3195 }
3196 spin_unlock(&ctx->completion_lock);
3197 if (list_empty(&list))
3198 return false;
3199
3200 while (!list_empty(&list)) {
3201 de = list_first_entry(&list, struct io_defer_entry, list);
3202 list_del_init(&de->list);
3203 io_req_task_queue_fail(de->req, -ECANCELED);
3204 kfree(de);
3205 }
3206 return true;
3207}
3208
3209static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
3210{
3211 struct io_tctx_node *node;
3212 enum io_wq_cancel cret;
3213 bool ret = false;
3214
3215 mutex_lock(&ctx->uring_lock);
3216 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3217 struct io_uring_task *tctx = node->task->io_uring;
3218
3219 /*
3220 * io_wq will stay alive while we hold uring_lock, because it's
3221 * killed after ctx nodes, which requires to take the lock.
3222 */
3223 if (!tctx || !tctx->io_wq)
3224 continue;
3225 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
3226 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3227 }
3228 mutex_unlock(&ctx->uring_lock);
3229
3230 return ret;
3231}
3232
3233static bool io_uring_try_cancel_uring_cmd(struct io_ring_ctx *ctx,
3234 struct task_struct *task, bool cancel_all)
3235{
3236 struct hlist_node *tmp;
3237 struct io_kiocb *req;
3238 bool ret = false;
3239
3240 lockdep_assert_held(&ctx->uring_lock);
3241
3242 hlist_for_each_entry_safe(req, tmp, &ctx->cancelable_uring_cmd,
3243 hash_node) {
3244 struct io_uring_cmd *cmd = io_kiocb_to_cmd(req,
3245 struct io_uring_cmd);
3246 struct file *file = req->file;
3247
3248 if (!cancel_all && req->task != task)
3249 continue;
3250
3251 if (cmd->flags & IORING_URING_CMD_CANCELABLE) {
3252 /* ->sqe isn't available if no async data */
3253 if (!req_has_async_data(req))
3254 cmd->sqe = NULL;
3255 file->f_op->uring_cmd(cmd, IO_URING_F_CANCEL);
3256 ret = true;
3257 }
3258 }
3259 io_submit_flush_completions(ctx);
3260
3261 return ret;
3262}
3263
3264static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
3265 struct task_struct *task,
3266 bool cancel_all)
3267{
3268 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
3269 struct io_uring_task *tctx = task ? task->io_uring : NULL;
3270 enum io_wq_cancel cret;
3271 bool ret = false;
3272
3273 /* set it so io_req_local_work_add() would wake us up */
3274 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
3275 atomic_set(&ctx->cq_wait_nr, 1);
3276 smp_mb();
3277 }
3278
3279 /* failed during ring init, it couldn't have issued any requests */
3280 if (!ctx->rings)
3281 return false;
3282
3283 if (!task) {
3284 ret |= io_uring_try_cancel_iowq(ctx);
3285 } else if (tctx && tctx->io_wq) {
3286 /*
3287 * Cancels requests of all rings, not only @ctx, but
3288 * it's fine as the task is in exit/exec.
3289 */
3290 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
3291 &cancel, true);
3292 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3293 }
3294
3295 /* SQPOLL thread does its own polling */
3296 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
3297 (ctx->sq_data && ctx->sq_data->thread == current)) {
3298 while (!wq_list_empty(&ctx->iopoll_list)) {
3299 io_iopoll_try_reap_events(ctx);
3300 ret = true;
3301 cond_resched();
3302 }
3303 }
3304
3305 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3306 io_allowed_defer_tw_run(ctx))
3307 ret |= io_run_local_work(ctx) > 0;
3308 ret |= io_cancel_defer_files(ctx, task, cancel_all);
3309 mutex_lock(&ctx->uring_lock);
3310 ret |= io_poll_remove_all(ctx, task, cancel_all);
3311 ret |= io_waitid_remove_all(ctx, task, cancel_all);
3312 ret |= io_futex_remove_all(ctx, task, cancel_all);
3313 ret |= io_uring_try_cancel_uring_cmd(ctx, task, cancel_all);
3314 mutex_unlock(&ctx->uring_lock);
3315 ret |= io_kill_timeouts(ctx, task, cancel_all);
3316 if (task)
3317 ret |= io_run_task_work() > 0;
3318 return ret;
3319}
3320
3321static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
3322{
3323 if (tracked)
3324 return atomic_read(&tctx->inflight_tracked);
3325 return percpu_counter_sum(&tctx->inflight);
3326}
3327
3328/*
3329 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3330 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3331 */
3332__cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
3333{
3334 struct io_uring_task *tctx = current->io_uring;
3335 struct io_ring_ctx *ctx;
3336 struct io_tctx_node *node;
3337 unsigned long index;
3338 s64 inflight;
3339 DEFINE_WAIT(wait);
3340
3341 WARN_ON_ONCE(sqd && sqd->thread != current);
3342
3343 if (!current->io_uring)
3344 return;
3345 if (tctx->io_wq)
3346 io_wq_exit_start(tctx->io_wq);
3347
3348 atomic_inc(&tctx->in_cancel);
3349 do {
3350 bool loop = false;
3351
3352 io_uring_drop_tctx_refs(current);
3353 /* read completions before cancelations */
3354 inflight = tctx_inflight(tctx, !cancel_all);
3355 if (!inflight)
3356 break;
3357
3358 if (!sqd) {
3359 xa_for_each(&tctx->xa, index, node) {
3360 /* sqpoll task will cancel all its requests */
3361 if (node->ctx->sq_data)
3362 continue;
3363 loop |= io_uring_try_cancel_requests(node->ctx,
3364 current, cancel_all);
3365 }
3366 } else {
3367 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3368 loop |= io_uring_try_cancel_requests(ctx,
3369 current,
3370 cancel_all);
3371 }
3372
3373 if (loop) {
3374 cond_resched();
3375 continue;
3376 }
3377
3378 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3379 io_run_task_work();
3380 io_uring_drop_tctx_refs(current);
3381 xa_for_each(&tctx->xa, index, node) {
3382 if (!llist_empty(&node->ctx->work_llist)) {
3383 WARN_ON_ONCE(node->ctx->submitter_task &&
3384 node->ctx->submitter_task != current);
3385 goto end_wait;
3386 }
3387 }
3388 /*
3389 * If we've seen completions, retry without waiting. This
3390 * avoids a race where a completion comes in before we did
3391 * prepare_to_wait().
3392 */
3393 if (inflight == tctx_inflight(tctx, !cancel_all))
3394 schedule();
3395end_wait:
3396 finish_wait(&tctx->wait, &wait);
3397 } while (1);
3398
3399 io_uring_clean_tctx(tctx);
3400 if (cancel_all) {
3401 /*
3402 * We shouldn't run task_works after cancel, so just leave
3403 * ->in_cancel set for normal exit.
3404 */
3405 atomic_dec(&tctx->in_cancel);
3406 /* for exec all current's requests should be gone, kill tctx */
3407 __io_uring_free(current);
3408 }
3409}
3410
3411void __io_uring_cancel(bool cancel_all)
3412{
3413 io_uring_cancel_generic(cancel_all, NULL);
3414}
3415
3416static void *io_uring_validate_mmap_request(struct file *file,
3417 loff_t pgoff, size_t sz)
3418{
3419 struct io_ring_ctx *ctx = file->private_data;
3420 loff_t offset = pgoff << PAGE_SHIFT;
3421 struct page *page;
3422 void *ptr;
3423
3424 switch (offset & IORING_OFF_MMAP_MASK) {
3425 case IORING_OFF_SQ_RING:
3426 case IORING_OFF_CQ_RING:
3427 /* Don't allow mmap if the ring was setup without it */
3428 if (ctx->flags & IORING_SETUP_NO_MMAP)
3429 return ERR_PTR(-EINVAL);
3430 ptr = ctx->rings;
3431 break;
3432 case IORING_OFF_SQES:
3433 /* Don't allow mmap if the ring was setup without it */
3434 if (ctx->flags & IORING_SETUP_NO_MMAP)
3435 return ERR_PTR(-EINVAL);
3436 ptr = ctx->sq_sqes;
3437 break;
3438 case IORING_OFF_PBUF_RING: {
3439 unsigned int bgid;
3440
3441 bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT;
3442 rcu_read_lock();
3443 ptr = io_pbuf_get_address(ctx, bgid);
3444 rcu_read_unlock();
3445 if (!ptr)
3446 return ERR_PTR(-EINVAL);
3447 break;
3448 }
3449 default:
3450 return ERR_PTR(-EINVAL);
3451 }
3452
3453 page = virt_to_head_page(ptr);
3454 if (sz > page_size(page))
3455 return ERR_PTR(-EINVAL);
3456
3457 return ptr;
3458}
3459
3460#ifdef CONFIG_MMU
3461
3462static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3463{
3464 size_t sz = vma->vm_end - vma->vm_start;
3465 unsigned long pfn;
3466 void *ptr;
3467
3468 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3469 if (IS_ERR(ptr))
3470 return PTR_ERR(ptr);
3471
3472 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3473 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3474}
3475
3476static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp,
3477 unsigned long addr, unsigned long len,
3478 unsigned long pgoff, unsigned long flags)
3479{
3480 void *ptr;
3481
3482 /*
3483 * Do not allow to map to user-provided address to avoid breaking the
3484 * aliasing rules. Userspace is not able to guess the offset address of
3485 * kernel kmalloc()ed memory area.
3486 */
3487 if (addr)
3488 return -EINVAL;
3489
3490 ptr = io_uring_validate_mmap_request(filp, pgoff, len);
3491 if (IS_ERR(ptr))
3492 return -ENOMEM;
3493
3494 /*
3495 * Some architectures have strong cache aliasing requirements.
3496 * For such architectures we need a coherent mapping which aliases
3497 * kernel memory *and* userspace memory. To achieve that:
3498 * - use a NULL file pointer to reference physical memory, and
3499 * - use the kernel virtual address of the shared io_uring context
3500 * (instead of the userspace-provided address, which has to be 0UL
3501 * anyway).
3502 * - use the same pgoff which the get_unmapped_area() uses to
3503 * calculate the page colouring.
3504 * For architectures without such aliasing requirements, the
3505 * architecture will return any suitable mapping because addr is 0.
3506 */
3507 filp = NULL;
3508 flags |= MAP_SHARED;
3509 pgoff = 0; /* has been translated to ptr above */
3510#ifdef SHM_COLOUR
3511 addr = (uintptr_t) ptr;
3512 pgoff = addr >> PAGE_SHIFT;
3513#else
3514 addr = 0UL;
3515#endif
3516 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
3517}
3518
3519#else /* !CONFIG_MMU */
3520
3521static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3522{
3523 return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL;
3524}
3525
3526static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3527{
3528 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3529}
3530
3531static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3532 unsigned long addr, unsigned long len,
3533 unsigned long pgoff, unsigned long flags)
3534{
3535 void *ptr;
3536
3537 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3538 if (IS_ERR(ptr))
3539 return PTR_ERR(ptr);
3540
3541 return (unsigned long) ptr;
3542}
3543
3544#endif /* !CONFIG_MMU */
3545
3546static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3547{
3548 if (flags & IORING_ENTER_EXT_ARG) {
3549 struct io_uring_getevents_arg arg;
3550
3551 if (argsz != sizeof(arg))
3552 return -EINVAL;
3553 if (copy_from_user(&arg, argp, sizeof(arg)))
3554 return -EFAULT;
3555 }
3556 return 0;
3557}
3558
3559static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3560 struct __kernel_timespec __user **ts,
3561 const sigset_t __user **sig)
3562{
3563 struct io_uring_getevents_arg arg;
3564
3565 /*
3566 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3567 * is just a pointer to the sigset_t.
3568 */
3569 if (!(flags & IORING_ENTER_EXT_ARG)) {
3570 *sig = (const sigset_t __user *) argp;
3571 *ts = NULL;
3572 return 0;
3573 }
3574
3575 /*
3576 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3577 * timespec and sigset_t pointers if good.
3578 */
3579 if (*argsz != sizeof(arg))
3580 return -EINVAL;
3581 if (copy_from_user(&arg, argp, sizeof(arg)))
3582 return -EFAULT;
3583 if (arg.pad)
3584 return -EINVAL;
3585 *sig = u64_to_user_ptr(arg.sigmask);
3586 *argsz = arg.sigmask_sz;
3587 *ts = u64_to_user_ptr(arg.ts);
3588 return 0;
3589}
3590
3591SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3592 u32, min_complete, u32, flags, const void __user *, argp,
3593 size_t, argsz)
3594{
3595 struct io_ring_ctx *ctx;
3596 struct file *file;
3597 long ret;
3598
3599 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3600 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3601 IORING_ENTER_REGISTERED_RING)))
3602 return -EINVAL;
3603
3604 /*
3605 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3606 * need only dereference our task private array to find it.
3607 */
3608 if (flags & IORING_ENTER_REGISTERED_RING) {
3609 struct io_uring_task *tctx = current->io_uring;
3610
3611 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3612 return -EINVAL;
3613 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3614 file = tctx->registered_rings[fd];
3615 if (unlikely(!file))
3616 return -EBADF;
3617 } else {
3618 file = fget(fd);
3619 if (unlikely(!file))
3620 return -EBADF;
3621 ret = -EOPNOTSUPP;
3622 if (unlikely(!io_is_uring_fops(file)))
3623 goto out;
3624 }
3625
3626 ctx = file->private_data;
3627 ret = -EBADFD;
3628 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3629 goto out;
3630
3631 /*
3632 * For SQ polling, the thread will do all submissions and completions.
3633 * Just return the requested submit count, and wake the thread if
3634 * we were asked to.
3635 */
3636 ret = 0;
3637 if (ctx->flags & IORING_SETUP_SQPOLL) {
3638 io_cqring_overflow_flush(ctx);
3639
3640 if (unlikely(ctx->sq_data->thread == NULL)) {
3641 ret = -EOWNERDEAD;
3642 goto out;
3643 }
3644 if (flags & IORING_ENTER_SQ_WAKEUP)
3645 wake_up(&ctx->sq_data->wait);
3646 if (flags & IORING_ENTER_SQ_WAIT)
3647 io_sqpoll_wait_sq(ctx);
3648
3649 ret = to_submit;
3650 } else if (to_submit) {
3651 ret = io_uring_add_tctx_node(ctx);
3652 if (unlikely(ret))
3653 goto out;
3654
3655 mutex_lock(&ctx->uring_lock);
3656 ret = io_submit_sqes(ctx, to_submit);
3657 if (ret != to_submit) {
3658 mutex_unlock(&ctx->uring_lock);
3659 goto out;
3660 }
3661 if (flags & IORING_ENTER_GETEVENTS) {
3662 if (ctx->syscall_iopoll)
3663 goto iopoll_locked;
3664 /*
3665 * Ignore errors, we'll soon call io_cqring_wait() and
3666 * it should handle ownership problems if any.
3667 */
3668 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3669 (void)io_run_local_work_locked(ctx);
3670 }
3671 mutex_unlock(&ctx->uring_lock);
3672 }
3673
3674 if (flags & IORING_ENTER_GETEVENTS) {
3675 int ret2;
3676
3677 if (ctx->syscall_iopoll) {
3678 /*
3679 * We disallow the app entering submit/complete with
3680 * polling, but we still need to lock the ring to
3681 * prevent racing with polled issue that got punted to
3682 * a workqueue.
3683 */
3684 mutex_lock(&ctx->uring_lock);
3685iopoll_locked:
3686 ret2 = io_validate_ext_arg(flags, argp, argsz);
3687 if (likely(!ret2)) {
3688 min_complete = min(min_complete,
3689 ctx->cq_entries);
3690 ret2 = io_iopoll_check(ctx, min_complete);
3691 }
3692 mutex_unlock(&ctx->uring_lock);
3693 } else {
3694 const sigset_t __user *sig;
3695 struct __kernel_timespec __user *ts;
3696
3697 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3698 if (likely(!ret2)) {
3699 min_complete = min(min_complete,
3700 ctx->cq_entries);
3701 ret2 = io_cqring_wait(ctx, min_complete, sig,
3702 argsz, ts);
3703 }
3704 }
3705
3706 if (!ret) {
3707 ret = ret2;
3708
3709 /*
3710 * EBADR indicates that one or more CQE were dropped.
3711 * Once the user has been informed we can clear the bit
3712 * as they are obviously ok with those drops.
3713 */
3714 if (unlikely(ret2 == -EBADR))
3715 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3716 &ctx->check_cq);
3717 }
3718 }
3719out:
3720 if (!(flags & IORING_ENTER_REGISTERED_RING))
3721 fput(file);
3722 return ret;
3723}
3724
3725static const struct file_operations io_uring_fops = {
3726 .release = io_uring_release,
3727 .mmap = io_uring_mmap,
3728#ifndef CONFIG_MMU
3729 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3730 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3731#else
3732 .get_unmapped_area = io_uring_mmu_get_unmapped_area,
3733#endif
3734 .poll = io_uring_poll,
3735#ifdef CONFIG_PROC_FS
3736 .show_fdinfo = io_uring_show_fdinfo,
3737#endif
3738};
3739
3740bool io_is_uring_fops(struct file *file)
3741{
3742 return file->f_op == &io_uring_fops;
3743}
3744
3745static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3746 struct io_uring_params *p)
3747{
3748 struct io_rings *rings;
3749 size_t size, sq_array_offset;
3750 void *ptr;
3751
3752 /* make sure these are sane, as we already accounted them */
3753 ctx->sq_entries = p->sq_entries;
3754 ctx->cq_entries = p->cq_entries;
3755
3756 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3757 if (size == SIZE_MAX)
3758 return -EOVERFLOW;
3759
3760 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3761 rings = io_mem_alloc(size);
3762 else
3763 rings = io_rings_map(ctx, p->cq_off.user_addr, size);
3764
3765 if (IS_ERR(rings))
3766 return PTR_ERR(rings);
3767
3768 ctx->rings = rings;
3769 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY))
3770 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3771 rings->sq_ring_mask = p->sq_entries - 1;
3772 rings->cq_ring_mask = p->cq_entries - 1;
3773 rings->sq_ring_entries = p->sq_entries;
3774 rings->cq_ring_entries = p->cq_entries;
3775
3776 if (p->flags & IORING_SETUP_SQE128)
3777 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3778 else
3779 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3780 if (size == SIZE_MAX) {
3781 io_rings_free(ctx);
3782 return -EOVERFLOW;
3783 }
3784
3785 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3786 ptr = io_mem_alloc(size);
3787 else
3788 ptr = io_sqes_map(ctx, p->sq_off.user_addr, size);
3789
3790 if (IS_ERR(ptr)) {
3791 io_rings_free(ctx);
3792 return PTR_ERR(ptr);
3793 }
3794
3795 ctx->sq_sqes = ptr;
3796 return 0;
3797}
3798
3799static int io_uring_install_fd(struct file *file)
3800{
3801 int fd;
3802
3803 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3804 if (fd < 0)
3805 return fd;
3806 fd_install(fd, file);
3807 return fd;
3808}
3809
3810/*
3811 * Allocate an anonymous fd, this is what constitutes the application
3812 * visible backing of an io_uring instance. The application mmaps this
3813 * fd to gain access to the SQ/CQ ring details.
3814 */
3815static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3816{
3817 /* Create a new inode so that the LSM can block the creation. */
3818 return anon_inode_create_getfile("[io_uring]", &io_uring_fops, ctx,
3819 O_RDWR | O_CLOEXEC, NULL);
3820}
3821
3822static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3823 struct io_uring_params __user *params)
3824{
3825 struct io_ring_ctx *ctx;
3826 struct io_uring_task *tctx;
3827 struct file *file;
3828 int ret;
3829
3830 if (!entries)
3831 return -EINVAL;
3832 if (entries > IORING_MAX_ENTRIES) {
3833 if (!(p->flags & IORING_SETUP_CLAMP))
3834 return -EINVAL;
3835 entries = IORING_MAX_ENTRIES;
3836 }
3837
3838 if ((p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
3839 && !(p->flags & IORING_SETUP_NO_MMAP))
3840 return -EINVAL;
3841
3842 /*
3843 * Use twice as many entries for the CQ ring. It's possible for the
3844 * application to drive a higher depth than the size of the SQ ring,
3845 * since the sqes are only used at submission time. This allows for
3846 * some flexibility in overcommitting a bit. If the application has
3847 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3848 * of CQ ring entries manually.
3849 */
3850 p->sq_entries = roundup_pow_of_two(entries);
3851 if (p->flags & IORING_SETUP_CQSIZE) {
3852 /*
3853 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3854 * to a power-of-two, if it isn't already. We do NOT impose
3855 * any cq vs sq ring sizing.
3856 */
3857 if (!p->cq_entries)
3858 return -EINVAL;
3859 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3860 if (!(p->flags & IORING_SETUP_CLAMP))
3861 return -EINVAL;
3862 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3863 }
3864 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3865 if (p->cq_entries < p->sq_entries)
3866 return -EINVAL;
3867 } else {
3868 p->cq_entries = 2 * p->sq_entries;
3869 }
3870
3871 ctx = io_ring_ctx_alloc(p);
3872 if (!ctx)
3873 return -ENOMEM;
3874
3875 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3876 !(ctx->flags & IORING_SETUP_IOPOLL) &&
3877 !(ctx->flags & IORING_SETUP_SQPOLL))
3878 ctx->task_complete = true;
3879
3880 if (ctx->task_complete || (ctx->flags & IORING_SETUP_IOPOLL))
3881 ctx->lockless_cq = true;
3882
3883 /*
3884 * lazy poll_wq activation relies on ->task_complete for synchronisation
3885 * purposes, see io_activate_pollwq()
3886 */
3887 if (!ctx->task_complete)
3888 ctx->poll_activated = true;
3889
3890 /*
3891 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3892 * space applications don't need to do io completion events
3893 * polling again, they can rely on io_sq_thread to do polling
3894 * work, which can reduce cpu usage and uring_lock contention.
3895 */
3896 if (ctx->flags & IORING_SETUP_IOPOLL &&
3897 !(ctx->flags & IORING_SETUP_SQPOLL))
3898 ctx->syscall_iopoll = 1;
3899
3900 ctx->compat = in_compat_syscall();
3901 if (!ns_capable_noaudit(&init_user_ns, CAP_IPC_LOCK))
3902 ctx->user = get_uid(current_user());
3903
3904 /*
3905 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3906 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3907 */
3908 ret = -EINVAL;
3909 if (ctx->flags & IORING_SETUP_SQPOLL) {
3910 /* IPI related flags don't make sense with SQPOLL */
3911 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3912 IORING_SETUP_TASKRUN_FLAG |
3913 IORING_SETUP_DEFER_TASKRUN))
3914 goto err;
3915 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3916 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3917 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3918 } else {
3919 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3920 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3921 goto err;
3922 ctx->notify_method = TWA_SIGNAL;
3923 }
3924
3925 /*
3926 * For DEFER_TASKRUN we require the completion task to be the same as the
3927 * submission task. This implies that there is only one submitter, so enforce
3928 * that.
3929 */
3930 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3931 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3932 goto err;
3933 }
3934
3935 /*
3936 * This is just grabbed for accounting purposes. When a process exits,
3937 * the mm is exited and dropped before the files, hence we need to hang
3938 * on to this mm purely for the purposes of being able to unaccount
3939 * memory (locked/pinned vm). It's not used for anything else.
3940 */
3941 mmgrab(current->mm);
3942 ctx->mm_account = current->mm;
3943
3944 ret = io_allocate_scq_urings(ctx, p);
3945 if (ret)
3946 goto err;
3947
3948 ret = io_sq_offload_create(ctx, p);
3949 if (ret)
3950 goto err;
3951
3952 ret = io_rsrc_init(ctx);
3953 if (ret)
3954 goto err;
3955
3956 p->sq_off.head = offsetof(struct io_rings, sq.head);
3957 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3958 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3959 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3960 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3961 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3962 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY))
3963 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3964 p->sq_off.resv1 = 0;
3965 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3966 p->sq_off.user_addr = 0;
3967
3968 p->cq_off.head = offsetof(struct io_rings, cq.head);
3969 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3970 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3971 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3972 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3973 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3974 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3975 p->cq_off.resv1 = 0;
3976 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3977 p->cq_off.user_addr = 0;
3978
3979 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3980 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3981 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3982 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3983 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3984 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3985 IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING;
3986
3987 if (copy_to_user(params, p, sizeof(*p))) {
3988 ret = -EFAULT;
3989 goto err;
3990 }
3991
3992 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
3993 && !(ctx->flags & IORING_SETUP_R_DISABLED))
3994 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
3995
3996 file = io_uring_get_file(ctx);
3997 if (IS_ERR(file)) {
3998 ret = PTR_ERR(file);
3999 goto err;
4000 }
4001
4002 ret = __io_uring_add_tctx_node(ctx);
4003 if (ret)
4004 goto err_fput;
4005 tctx = current->io_uring;
4006
4007 /*
4008 * Install ring fd as the very last thing, so we don't risk someone
4009 * having closed it before we finish setup
4010 */
4011 if (p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
4012 ret = io_ring_add_registered_file(tctx, file, 0, IO_RINGFD_REG_MAX);
4013 else
4014 ret = io_uring_install_fd(file);
4015 if (ret < 0)
4016 goto err_fput;
4017
4018 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
4019 return ret;
4020err:
4021 io_ring_ctx_wait_and_kill(ctx);
4022 return ret;
4023err_fput:
4024 fput(file);
4025 return ret;
4026}
4027
4028/*
4029 * Sets up an aio uring context, and returns the fd. Applications asks for a
4030 * ring size, we return the actual sq/cq ring sizes (among other things) in the
4031 * params structure passed in.
4032 */
4033static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
4034{
4035 struct io_uring_params p;
4036 int i;
4037
4038 if (copy_from_user(&p, params, sizeof(p)))
4039 return -EFAULT;
4040 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
4041 if (p.resv[i])
4042 return -EINVAL;
4043 }
4044
4045 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
4046 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
4047 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
4048 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
4049 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
4050 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
4051 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN |
4052 IORING_SETUP_NO_MMAP | IORING_SETUP_REGISTERED_FD_ONLY |
4053 IORING_SETUP_NO_SQARRAY))
4054 return -EINVAL;
4055
4056 return io_uring_create(entries, &p, params);
4057}
4058
4059static inline bool io_uring_allowed(void)
4060{
4061 int disabled = READ_ONCE(sysctl_io_uring_disabled);
4062 kgid_t io_uring_group;
4063
4064 if (disabled == 2)
4065 return false;
4066
4067 if (disabled == 0 || capable(CAP_SYS_ADMIN))
4068 return true;
4069
4070 io_uring_group = make_kgid(&init_user_ns, sysctl_io_uring_group);
4071 if (!gid_valid(io_uring_group))
4072 return false;
4073
4074 return in_group_p(io_uring_group);
4075}
4076
4077SYSCALL_DEFINE2(io_uring_setup, u32, entries,
4078 struct io_uring_params __user *, params)
4079{
4080 if (!io_uring_allowed())
4081 return -EPERM;
4082
4083 return io_uring_setup(entries, params);
4084}
4085
4086static int __init io_uring_init(void)
4087{
4088#define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4089 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4090 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4091} while (0)
4092
4093#define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4094 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4095#define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4096 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4097 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4098 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
4099 BUILD_BUG_SQE_ELEM(1, __u8, flags);
4100 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
4101 BUILD_BUG_SQE_ELEM(4, __s32, fd);
4102 BUILD_BUG_SQE_ELEM(8, __u64, off);
4103 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
4104 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
4105 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
4106 BUILD_BUG_SQE_ELEM(16, __u64, addr);
4107 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
4108 BUILD_BUG_SQE_ELEM(24, __u32, len);
4109 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
4110 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
4111 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4112 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
4113 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
4114 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
4115 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
4116 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
4117 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
4118 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
4119 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
4120 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
4121 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
4122 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
4123 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
4124 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
4125 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
4126 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
4127 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
4128 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
4129 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
4130 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
4131 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
4132 BUILD_BUG_SQE_ELEM(42, __u16, personality);
4133 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
4134 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
4135 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
4136 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
4137 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
4138 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4139 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
4140
4141 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4142 sizeof(struct io_uring_rsrc_update));
4143 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4144 sizeof(struct io_uring_rsrc_update2));
4145
4146 /* ->buf_index is u16 */
4147 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4148 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4149 offsetof(struct io_uring_buf_ring, tail));
4150
4151 /* should fit into one byte */
4152 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
4153 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4154 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4155
4156 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
4157
4158 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4159
4160 /* top 8bits are for internal use */
4161 BUILD_BUG_ON((IORING_URING_CMD_MASK & 0xff000000) != 0);
4162
4163 io_uring_optable_init();
4164
4165 /*
4166 * Allow user copy in the per-command field, which starts after the
4167 * file in io_kiocb and until the opcode field. The openat2 handling
4168 * requires copying in user memory into the io_kiocb object in that
4169 * range, and HARDENED_USERCOPY will complain if we haven't
4170 * correctly annotated this range.
4171 */
4172 req_cachep = kmem_cache_create_usercopy("io_kiocb",
4173 sizeof(struct io_kiocb), 0,
4174 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4175 SLAB_ACCOUNT | SLAB_TYPESAFE_BY_RCU,
4176 offsetof(struct io_kiocb, cmd.data),
4177 sizeof_field(struct io_kiocb, cmd.data), NULL);
4178 io_buf_cachep = kmem_cache_create("io_buffer", sizeof(struct io_buffer), 0,
4179 SLAB_HWCACHE_ALIGN | SLAB_PANIC | SLAB_ACCOUNT,
4180 NULL);
4181
4182#ifdef CONFIG_SYSCTL
4183 register_sysctl_init("kernel", kernel_io_uring_disabled_table);
4184#endif
4185
4186 return 0;
4187};
4188__initcall(io_uring_init);