1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
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
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
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
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
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.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
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>
51 #include <linux/sched/signal.h>
53 #include <linux/file.h>
54 #include <linux/fdtable.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>
62 #include <net/af_unix.h>
64 #include <linux/anon_inodes.h>
65 #include <linux/sched/mm.h>
66 #include <linux/uaccess.h>
67 #include <linux/nospec.h>
68 #include <linux/highmem.h>
69 #include <linux/fsnotify.h>
70 #include <linux/fadvise.h>
71 #include <linux/task_work.h>
72 #include <linux/io_uring.h>
73 #include <linux/audit.h>
74 #include <linux/security.h>
75 #include <asm/shmparam.h>
77 #define CREATE_TRACE_POINTS
78 #include <trace/events/io_uring.h>
80 #include <uapi/linux/io_uring.h>
98 #include "alloc_cache.h"
100 #define IORING_MAX_ENTRIES 32768
101 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
103 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
104 IORING_REGISTER_LAST + IORING_OP_LAST)
106 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
107 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
109 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
110 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
112 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
113 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
116 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
119 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
121 #define IO_COMPL_BATCH 32
122 #define IO_REQ_ALLOC_BATCH 8
125 IO_CHECK_CQ_OVERFLOW_BIT
,
126 IO_CHECK_CQ_DROPPED_BIT
,
130 IO_EVENTFD_OP_SIGNAL_BIT
,
131 IO_EVENTFD_OP_FREE_BIT
,
134 struct io_defer_entry
{
135 struct list_head list
;
136 struct io_kiocb
*req
;
140 /* requests with any of those set should undergo io_disarm_next() */
141 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
142 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
144 static bool io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
145 struct task_struct
*task
,
148 static void io_dismantle_req(struct io_kiocb
*req
);
149 static void io_clean_op(struct io_kiocb
*req
);
150 static void io_queue_sqe(struct io_kiocb
*req
);
151 static void io_move_task_work_from_local(struct io_ring_ctx
*ctx
);
152 static void __io_submit_flush_completions(struct io_ring_ctx
*ctx
);
153 static __cold
void io_fallback_tw(struct io_uring_task
*tctx
);
155 struct kmem_cache
*req_cachep
;
157 struct sock
*io_uring_get_socket(struct file
*file
)
159 #if defined(CONFIG_UNIX)
160 if (io_is_uring_fops(file
)) {
161 struct io_ring_ctx
*ctx
= file
->private_data
;
163 return ctx
->ring_sock
->sk
;
168 EXPORT_SYMBOL(io_uring_get_socket
);
170 static inline void io_submit_flush_completions(struct io_ring_ctx
*ctx
)
172 if (!wq_list_empty(&ctx
->submit_state
.compl_reqs
) ||
173 ctx
->submit_state
.cqes_count
)
174 __io_submit_flush_completions(ctx
);
177 static inline unsigned int __io_cqring_events(struct io_ring_ctx
*ctx
)
179 return ctx
->cached_cq_tail
- READ_ONCE(ctx
->rings
->cq
.head
);
182 static inline unsigned int __io_cqring_events_user(struct io_ring_ctx
*ctx
)
184 return READ_ONCE(ctx
->rings
->cq
.tail
) - READ_ONCE(ctx
->rings
->cq
.head
);
187 static bool io_match_linked(struct io_kiocb
*head
)
189 struct io_kiocb
*req
;
191 io_for_each_link(req
, head
) {
192 if (req
->flags
& REQ_F_INFLIGHT
)
199 * As io_match_task() but protected against racing with linked timeouts.
200 * User must not hold timeout_lock.
202 bool io_match_task_safe(struct io_kiocb
*head
, struct task_struct
*task
,
207 if (task
&& head
->task
!= task
)
212 if (head
->flags
& REQ_F_LINK_TIMEOUT
) {
213 struct io_ring_ctx
*ctx
= head
->ctx
;
215 /* protect against races with linked timeouts */
216 spin_lock_irq(&ctx
->timeout_lock
);
217 matched
= io_match_linked(head
);
218 spin_unlock_irq(&ctx
->timeout_lock
);
220 matched
= io_match_linked(head
);
225 static inline void req_fail_link_node(struct io_kiocb
*req
, int res
)
228 io_req_set_res(req
, res
, 0);
231 static inline void io_req_add_to_cache(struct io_kiocb
*req
, struct io_ring_ctx
*ctx
)
233 wq_stack_add_head(&req
->comp_list
, &ctx
->submit_state
.free_list
);
234 kasan_poison_object_data(req_cachep
, req
);
237 static __cold
void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
239 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
241 complete(&ctx
->ref_comp
);
244 static __cold
void io_fallback_req_func(struct work_struct
*work
)
246 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
,
248 struct llist_node
*node
= llist_del_all(&ctx
->fallback_llist
);
249 struct io_kiocb
*req
, *tmp
;
250 struct io_tw_state ts
= { .locked
= true, };
252 mutex_lock(&ctx
->uring_lock
);
253 llist_for_each_entry_safe(req
, tmp
, node
, io_task_work
.node
)
254 req
->io_task_work
.func(req
, &ts
);
255 if (WARN_ON_ONCE(!ts
.locked
))
257 io_submit_flush_completions(ctx
);
258 mutex_unlock(&ctx
->uring_lock
);
261 static int io_alloc_hash_table(struct io_hash_table
*table
, unsigned bits
)
263 unsigned hash_buckets
= 1U << bits
;
264 size_t hash_size
= hash_buckets
* sizeof(table
->hbs
[0]);
266 table
->hbs
= kmalloc(hash_size
, GFP_KERNEL
);
270 table
->hash_bits
= bits
;
271 init_hash_table(table
, hash_buckets
);
275 static __cold
struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
277 struct io_ring_ctx
*ctx
;
280 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
284 xa_init(&ctx
->io_bl_xa
);
287 * Use 5 bits less than the max cq entries, that should give us around
288 * 32 entries per hash list if totally full and uniformly spread, but
289 * don't keep too many buckets to not overconsume memory.
291 hash_bits
= ilog2(p
->cq_entries
) - 5;
292 hash_bits
= clamp(hash_bits
, 1, 8);
293 if (io_alloc_hash_table(&ctx
->cancel_table
, hash_bits
))
295 if (io_alloc_hash_table(&ctx
->cancel_table_locked
, hash_bits
))
298 ctx
->dummy_ubuf
= kzalloc(sizeof(*ctx
->dummy_ubuf
), GFP_KERNEL
);
299 if (!ctx
->dummy_ubuf
)
301 /* set invalid range, so io_import_fixed() fails meeting it */
302 ctx
->dummy_ubuf
->ubuf
= -1UL;
304 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
,
308 ctx
->flags
= p
->flags
;
309 init_waitqueue_head(&ctx
->sqo_sq_wait
);
310 INIT_LIST_HEAD(&ctx
->sqd_list
);
311 INIT_LIST_HEAD(&ctx
->cq_overflow_list
);
312 INIT_LIST_HEAD(&ctx
->io_buffers_cache
);
313 io_alloc_cache_init(&ctx
->rsrc_node_cache
, IO_NODE_ALLOC_CACHE_MAX
,
314 sizeof(struct io_rsrc_node
));
315 io_alloc_cache_init(&ctx
->apoll_cache
, IO_ALLOC_CACHE_MAX
,
316 sizeof(struct async_poll
));
317 io_alloc_cache_init(&ctx
->netmsg_cache
, IO_ALLOC_CACHE_MAX
,
318 sizeof(struct io_async_msghdr
));
319 init_completion(&ctx
->ref_comp
);
320 xa_init_flags(&ctx
->personalities
, XA_FLAGS_ALLOC1
);
321 mutex_init(&ctx
->uring_lock
);
322 init_waitqueue_head(&ctx
->cq_wait
);
323 init_waitqueue_head(&ctx
->poll_wq
);
324 init_waitqueue_head(&ctx
->rsrc_quiesce_wq
);
325 spin_lock_init(&ctx
->completion_lock
);
326 spin_lock_init(&ctx
->timeout_lock
);
327 INIT_WQ_LIST(&ctx
->iopoll_list
);
328 INIT_LIST_HEAD(&ctx
->io_buffers_pages
);
329 INIT_LIST_HEAD(&ctx
->io_buffers_comp
);
330 INIT_LIST_HEAD(&ctx
->defer_list
);
331 INIT_LIST_HEAD(&ctx
->timeout_list
);
332 INIT_LIST_HEAD(&ctx
->ltimeout_list
);
333 INIT_LIST_HEAD(&ctx
->rsrc_ref_list
);
334 init_llist_head(&ctx
->work_llist
);
335 INIT_LIST_HEAD(&ctx
->tctx_list
);
336 ctx
->submit_state
.free_list
.next
= NULL
;
337 INIT_WQ_LIST(&ctx
->locked_free_list
);
338 INIT_DELAYED_WORK(&ctx
->fallback_work
, io_fallback_req_func
);
339 INIT_WQ_LIST(&ctx
->submit_state
.compl_reqs
);
342 kfree(ctx
->dummy_ubuf
);
343 kfree(ctx
->cancel_table
.hbs
);
344 kfree(ctx
->cancel_table_locked
.hbs
);
346 xa_destroy(&ctx
->io_bl_xa
);
351 static void io_account_cq_overflow(struct io_ring_ctx
*ctx
)
353 struct io_rings
*r
= ctx
->rings
;
355 WRITE_ONCE(r
->cq_overflow
, READ_ONCE(r
->cq_overflow
) + 1);
359 static bool req_need_defer(struct io_kiocb
*req
, u32 seq
)
361 if (unlikely(req
->flags
& REQ_F_IO_DRAIN
)) {
362 struct io_ring_ctx
*ctx
= req
->ctx
;
364 return seq
+ READ_ONCE(ctx
->cq_extra
) != ctx
->cached_cq_tail
;
370 static inline void io_req_track_inflight(struct io_kiocb
*req
)
372 if (!(req
->flags
& REQ_F_INFLIGHT
)) {
373 req
->flags
|= REQ_F_INFLIGHT
;
374 atomic_inc(&req
->task
->io_uring
->inflight_tracked
);
378 static struct io_kiocb
*__io_prep_linked_timeout(struct io_kiocb
*req
)
380 if (WARN_ON_ONCE(!req
->link
))
383 req
->flags
&= ~REQ_F_ARM_LTIMEOUT
;
384 req
->flags
|= REQ_F_LINK_TIMEOUT
;
386 /* linked timeouts should have two refs once prep'ed */
387 io_req_set_refcount(req
);
388 __io_req_set_refcount(req
->link
, 2);
392 static inline struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
)
394 if (likely(!(req
->flags
& REQ_F_ARM_LTIMEOUT
)))
396 return __io_prep_linked_timeout(req
);
399 static noinline
void __io_arm_ltimeout(struct io_kiocb
*req
)
401 io_queue_linked_timeout(__io_prep_linked_timeout(req
));
404 static inline void io_arm_ltimeout(struct io_kiocb
*req
)
406 if (unlikely(req
->flags
& REQ_F_ARM_LTIMEOUT
))
407 __io_arm_ltimeout(req
);
410 static void io_prep_async_work(struct io_kiocb
*req
)
412 const struct io_issue_def
*def
= &io_issue_defs
[req
->opcode
];
413 struct io_ring_ctx
*ctx
= req
->ctx
;
415 if (!(req
->flags
& REQ_F_CREDS
)) {
416 req
->flags
|= REQ_F_CREDS
;
417 req
->creds
= get_current_cred();
420 req
->work
.list
.next
= NULL
;
422 req
->work
.cancel_seq
= atomic_read(&ctx
->cancel_seq
);
423 if (req
->flags
& REQ_F_FORCE_ASYNC
)
424 req
->work
.flags
|= IO_WQ_WORK_CONCURRENT
;
426 if (req
->file
&& !io_req_ffs_set(req
))
427 req
->flags
|= io_file_get_flags(req
->file
) << REQ_F_SUPPORT_NOWAIT_BIT
;
429 if (req
->file
&& (req
->flags
& REQ_F_ISREG
)) {
430 bool should_hash
= def
->hash_reg_file
;
432 /* don't serialize this request if the fs doesn't need it */
433 if (should_hash
&& (req
->file
->f_flags
& O_DIRECT
) &&
434 (req
->file
->f_mode
& FMODE_DIO_PARALLEL_WRITE
))
436 if (should_hash
|| (ctx
->flags
& IORING_SETUP_IOPOLL
))
437 io_wq_hash_work(&req
->work
, file_inode(req
->file
));
438 } else if (!req
->file
|| !S_ISBLK(file_inode(req
->file
)->i_mode
)) {
439 if (def
->unbound_nonreg_file
)
440 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
444 static void io_prep_async_link(struct io_kiocb
*req
)
446 struct io_kiocb
*cur
;
448 if (req
->flags
& REQ_F_LINK_TIMEOUT
) {
449 struct io_ring_ctx
*ctx
= req
->ctx
;
451 spin_lock_irq(&ctx
->timeout_lock
);
452 io_for_each_link(cur
, req
)
453 io_prep_async_work(cur
);
454 spin_unlock_irq(&ctx
->timeout_lock
);
456 io_for_each_link(cur
, req
)
457 io_prep_async_work(cur
);
461 void io_queue_iowq(struct io_kiocb
*req
, struct io_tw_state
*ts_dont_use
)
463 struct io_kiocb
*link
= io_prep_linked_timeout(req
);
464 struct io_uring_task
*tctx
= req
->task
->io_uring
;
467 BUG_ON(!tctx
->io_wq
);
469 /* init ->work of the whole link before punting */
470 io_prep_async_link(req
);
473 * Not expected to happen, but if we do have a bug where this _can_
474 * happen, catch it here and ensure the request is marked as
475 * canceled. That will make io-wq go through the usual work cancel
476 * procedure rather than attempt to run this request (or create a new
479 if (WARN_ON_ONCE(!same_thread_group(req
->task
, current
)))
480 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
482 trace_io_uring_queue_async_work(req
, io_wq_is_hashed(&req
->work
));
483 io_wq_enqueue(tctx
->io_wq
, &req
->work
);
485 io_queue_linked_timeout(link
);
488 static __cold
void io_queue_deferred(struct io_ring_ctx
*ctx
)
490 while (!list_empty(&ctx
->defer_list
)) {
491 struct io_defer_entry
*de
= list_first_entry(&ctx
->defer_list
,
492 struct io_defer_entry
, list
);
494 if (req_need_defer(de
->req
, de
->seq
))
496 list_del_init(&de
->list
);
497 io_req_task_queue(de
->req
);
503 static void io_eventfd_ops(struct rcu_head
*rcu
)
505 struct io_ev_fd
*ev_fd
= container_of(rcu
, struct io_ev_fd
, rcu
);
506 int ops
= atomic_xchg(&ev_fd
->ops
, 0);
508 if (ops
& BIT(IO_EVENTFD_OP_SIGNAL_BIT
))
509 eventfd_signal_mask(ev_fd
->cq_ev_fd
, 1, EPOLL_URING_WAKE
);
511 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
512 * ordering in a race but if references are 0 we know we have to free
515 if (atomic_dec_and_test(&ev_fd
->refs
)) {
516 eventfd_ctx_put(ev_fd
->cq_ev_fd
);
521 static void io_eventfd_signal(struct io_ring_ctx
*ctx
)
523 struct io_ev_fd
*ev_fd
= NULL
;
527 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
530 ev_fd
= rcu_dereference(ctx
->io_ev_fd
);
533 * Check again if ev_fd exists incase an io_eventfd_unregister call
534 * completed between the NULL check of ctx->io_ev_fd at the start of
535 * the function and rcu_read_lock.
537 if (unlikely(!ev_fd
))
539 if (READ_ONCE(ctx
->rings
->cq_flags
) & IORING_CQ_EVENTFD_DISABLED
)
541 if (ev_fd
->eventfd_async
&& !io_wq_current_is_worker())
544 if (likely(eventfd_signal_allowed())) {
545 eventfd_signal_mask(ev_fd
->cq_ev_fd
, 1, EPOLL_URING_WAKE
);
547 atomic_inc(&ev_fd
->refs
);
548 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT
), &ev_fd
->ops
))
549 call_rcu_hurry(&ev_fd
->rcu
, io_eventfd_ops
);
551 atomic_dec(&ev_fd
->refs
);
558 static void io_eventfd_flush_signal(struct io_ring_ctx
*ctx
)
562 spin_lock(&ctx
->completion_lock
);
565 * Eventfd should only get triggered when at least one event has been
566 * posted. Some applications rely on the eventfd notification count
567 * only changing IFF a new CQE has been added to the CQ ring. There's
568 * no depedency on 1:1 relationship between how many times this
569 * function is called (and hence the eventfd count) and number of CQEs
570 * posted to the CQ ring.
572 skip
= ctx
->cached_cq_tail
== ctx
->evfd_last_cq_tail
;
573 ctx
->evfd_last_cq_tail
= ctx
->cached_cq_tail
;
574 spin_unlock(&ctx
->completion_lock
);
578 io_eventfd_signal(ctx
);
581 void __io_commit_cqring_flush(struct io_ring_ctx
*ctx
)
583 if (ctx
->poll_activated
)
584 io_poll_wq_wake(ctx
);
585 if (ctx
->off_timeout_used
)
586 io_flush_timeouts(ctx
);
587 if (ctx
->drain_active
) {
588 spin_lock(&ctx
->completion_lock
);
589 io_queue_deferred(ctx
);
590 spin_unlock(&ctx
->completion_lock
);
593 io_eventfd_flush_signal(ctx
);
596 static inline void __io_cq_lock(struct io_ring_ctx
*ctx
)
597 __acquires(ctx
->completion_lock
)
599 if (!ctx
->task_complete
)
600 spin_lock(&ctx
->completion_lock
);
603 static inline void __io_cq_unlock(struct io_ring_ctx
*ctx
)
605 if (!ctx
->task_complete
)
606 spin_unlock(&ctx
->completion_lock
);
609 static inline void io_cq_lock(struct io_ring_ctx
*ctx
)
610 __acquires(ctx
->completion_lock
)
612 spin_lock(&ctx
->completion_lock
);
615 static inline void io_cq_unlock(struct io_ring_ctx
*ctx
)
616 __releases(ctx
->completion_lock
)
618 spin_unlock(&ctx
->completion_lock
);
621 /* keep it inlined for io_submit_flush_completions() */
622 static inline void __io_cq_unlock_post(struct io_ring_ctx
*ctx
)
623 __releases(ctx
->completion_lock
)
625 io_commit_cqring(ctx
);
627 io_commit_cqring_flush(ctx
);
631 static void __io_cq_unlock_post_flush(struct io_ring_ctx
*ctx
)
632 __releases(ctx
->completion_lock
)
634 io_commit_cqring(ctx
);
636 if (ctx
->task_complete
) {
638 * ->task_complete implies that only current might be waiting
639 * for CQEs, and obviously, we currently don't. No one is
640 * waiting, wakeups are futile, skip them.
642 io_commit_cqring_flush(ctx
);
645 io_commit_cqring_flush(ctx
);
650 void io_cq_unlock_post(struct io_ring_ctx
*ctx
)
651 __releases(ctx
->completion_lock
)
653 io_commit_cqring(ctx
);
654 spin_unlock(&ctx
->completion_lock
);
655 io_commit_cqring_flush(ctx
);
659 /* Returns true if there are no backlogged entries after the flush */
660 static void io_cqring_overflow_kill(struct io_ring_ctx
*ctx
)
662 struct io_overflow_cqe
*ocqe
;
666 list_splice_init(&ctx
->cq_overflow_list
, &list
);
667 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
);
670 while (!list_empty(&list
)) {
671 ocqe
= list_first_entry(&list
, struct io_overflow_cqe
, list
);
672 list_del(&ocqe
->list
);
677 static void __io_cqring_overflow_flush(struct io_ring_ctx
*ctx
)
679 size_t cqe_size
= sizeof(struct io_uring_cqe
);
681 if (__io_cqring_events(ctx
) == ctx
->cq_entries
)
684 if (ctx
->flags
& IORING_SETUP_CQE32
)
688 while (!list_empty(&ctx
->cq_overflow_list
)) {
689 struct io_uring_cqe
*cqe
= io_get_cqe_overflow(ctx
, true);
690 struct io_overflow_cqe
*ocqe
;
694 ocqe
= list_first_entry(&ctx
->cq_overflow_list
,
695 struct io_overflow_cqe
, list
);
696 memcpy(cqe
, &ocqe
->cqe
, cqe_size
);
697 list_del(&ocqe
->list
);
701 if (list_empty(&ctx
->cq_overflow_list
)) {
702 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
);
703 atomic_andnot(IORING_SQ_CQ_OVERFLOW
, &ctx
->rings
->sq_flags
);
705 io_cq_unlock_post(ctx
);
708 static void io_cqring_do_overflow_flush(struct io_ring_ctx
*ctx
)
710 /* iopoll syncs against uring_lock, not completion_lock */
711 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
712 mutex_lock(&ctx
->uring_lock
);
713 __io_cqring_overflow_flush(ctx
);
714 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
715 mutex_unlock(&ctx
->uring_lock
);
718 static void io_cqring_overflow_flush(struct io_ring_ctx
*ctx
)
720 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
))
721 io_cqring_do_overflow_flush(ctx
);
724 /* can be called by any task */
725 static void io_put_task_remote(struct task_struct
*task
, int nr
)
727 struct io_uring_task
*tctx
= task
->io_uring
;
729 percpu_counter_sub(&tctx
->inflight
, nr
);
730 if (unlikely(atomic_read(&tctx
->in_cancel
)))
731 wake_up(&tctx
->wait
);
732 put_task_struct_many(task
, nr
);
735 /* used by a task to put its own references */
736 static void io_put_task_local(struct task_struct
*task
, int nr
)
738 task
->io_uring
->cached_refs
+= nr
;
741 /* must to be called somewhat shortly after putting a request */
742 static inline void io_put_task(struct task_struct
*task
, int nr
)
744 if (likely(task
== current
))
745 io_put_task_local(task
, nr
);
747 io_put_task_remote(task
, nr
);
750 void io_task_refs_refill(struct io_uring_task
*tctx
)
752 unsigned int refill
= -tctx
->cached_refs
+ IO_TCTX_REFS_CACHE_NR
;
754 percpu_counter_add(&tctx
->inflight
, refill
);
755 refcount_add(refill
, ¤t
->usage
);
756 tctx
->cached_refs
+= refill
;
759 static __cold
void io_uring_drop_tctx_refs(struct task_struct
*task
)
761 struct io_uring_task
*tctx
= task
->io_uring
;
762 unsigned int refs
= tctx
->cached_refs
;
765 tctx
->cached_refs
= 0;
766 percpu_counter_sub(&tctx
->inflight
, refs
);
767 put_task_struct_many(task
, refs
);
771 static bool io_cqring_event_overflow(struct io_ring_ctx
*ctx
, u64 user_data
,
772 s32 res
, u32 cflags
, u64 extra1
, u64 extra2
)
774 struct io_overflow_cqe
*ocqe
;
775 size_t ocq_size
= sizeof(struct io_overflow_cqe
);
776 bool is_cqe32
= (ctx
->flags
& IORING_SETUP_CQE32
);
778 lockdep_assert_held(&ctx
->completion_lock
);
781 ocq_size
+= sizeof(struct io_uring_cqe
);
783 ocqe
= kmalloc(ocq_size
, GFP_ATOMIC
| __GFP_ACCOUNT
);
784 trace_io_uring_cqe_overflow(ctx
, user_data
, res
, cflags
, ocqe
);
787 * If we're in ring overflow flush mode, or in task cancel mode,
788 * or cannot allocate an overflow entry, then we need to drop it
791 io_account_cq_overflow(ctx
);
792 set_bit(IO_CHECK_CQ_DROPPED_BIT
, &ctx
->check_cq
);
795 if (list_empty(&ctx
->cq_overflow_list
)) {
796 set_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
);
797 atomic_or(IORING_SQ_CQ_OVERFLOW
, &ctx
->rings
->sq_flags
);
800 ocqe
->cqe
.user_data
= user_data
;
802 ocqe
->cqe
.flags
= cflags
;
804 ocqe
->cqe
.big_cqe
[0] = extra1
;
805 ocqe
->cqe
.big_cqe
[1] = extra2
;
807 list_add_tail(&ocqe
->list
, &ctx
->cq_overflow_list
);
811 bool io_req_cqe_overflow(struct io_kiocb
*req
)
813 if (!(req
->flags
& REQ_F_CQE32_INIT
)) {
817 return io_cqring_event_overflow(req
->ctx
, req
->cqe
.user_data
,
818 req
->cqe
.res
, req
->cqe
.flags
,
819 req
->extra1
, req
->extra2
);
823 * writes to the cq entry need to come after reading head; the
824 * control dependency is enough as we're using WRITE_ONCE to
827 struct io_uring_cqe
*__io_get_cqe(struct io_ring_ctx
*ctx
, bool overflow
)
829 struct io_rings
*rings
= ctx
->rings
;
830 unsigned int off
= ctx
->cached_cq_tail
& (ctx
->cq_entries
- 1);
831 unsigned int free
, queued
, len
;
834 * Posting into the CQ when there are pending overflowed CQEs may break
835 * ordering guarantees, which will affect links, F_MORE users and more.
836 * Force overflow the completion.
838 if (!overflow
&& (ctx
->check_cq
& BIT(IO_CHECK_CQ_OVERFLOW_BIT
)))
841 /* userspace may cheat modifying the tail, be safe and do min */
842 queued
= min(__io_cqring_events(ctx
), ctx
->cq_entries
);
843 free
= ctx
->cq_entries
- queued
;
844 /* we need a contiguous range, limit based on the current array offset */
845 len
= min(free
, ctx
->cq_entries
- off
);
849 if (ctx
->flags
& IORING_SETUP_CQE32
) {
854 ctx
->cqe_cached
= &rings
->cqes
[off
];
855 ctx
->cqe_sentinel
= ctx
->cqe_cached
+ len
;
857 ctx
->cached_cq_tail
++;
859 if (ctx
->flags
& IORING_SETUP_CQE32
)
861 return &rings
->cqes
[off
];
864 static bool io_fill_cqe_aux(struct io_ring_ctx
*ctx
, u64 user_data
, s32 res
,
867 struct io_uring_cqe
*cqe
;
872 * If we can't get a cq entry, userspace overflowed the
873 * submission (by quite a lot). Increment the overflow count in
876 cqe
= io_get_cqe(ctx
);
878 trace_io_uring_complete(ctx
, NULL
, user_data
, res
, cflags
, 0, 0);
880 WRITE_ONCE(cqe
->user_data
, user_data
);
881 WRITE_ONCE(cqe
->res
, res
);
882 WRITE_ONCE(cqe
->flags
, cflags
);
884 if (ctx
->flags
& IORING_SETUP_CQE32
) {
885 WRITE_ONCE(cqe
->big_cqe
[0], 0);
886 WRITE_ONCE(cqe
->big_cqe
[1], 0);
893 static void __io_flush_post_cqes(struct io_ring_ctx
*ctx
)
894 __must_hold(&ctx
->uring_lock
)
896 struct io_submit_state
*state
= &ctx
->submit_state
;
899 lockdep_assert_held(&ctx
->uring_lock
);
900 for (i
= 0; i
< state
->cqes_count
; i
++) {
901 struct io_uring_cqe
*cqe
= &state
->cqes
[i
];
903 if (!io_fill_cqe_aux(ctx
, cqe
->user_data
, cqe
->res
, cqe
->flags
)) {
904 if (ctx
->task_complete
) {
905 spin_lock(&ctx
->completion_lock
);
906 io_cqring_event_overflow(ctx
, cqe
->user_data
,
907 cqe
->res
, cqe
->flags
, 0, 0);
908 spin_unlock(&ctx
->completion_lock
);
910 io_cqring_event_overflow(ctx
, cqe
->user_data
,
911 cqe
->res
, cqe
->flags
, 0, 0);
915 state
->cqes_count
= 0;
918 static bool __io_post_aux_cqe(struct io_ring_ctx
*ctx
, u64 user_data
, s32 res
, u32 cflags
,
924 filled
= io_fill_cqe_aux(ctx
, user_data
, res
, cflags
);
925 if (!filled
&& allow_overflow
)
926 filled
= io_cqring_event_overflow(ctx
, user_data
, res
, cflags
, 0, 0);
928 io_cq_unlock_post(ctx
);
932 bool io_post_aux_cqe(struct io_ring_ctx
*ctx
, u64 user_data
, s32 res
, u32 cflags
)
934 return __io_post_aux_cqe(ctx
, user_data
, res
, cflags
, true);
937 bool io_aux_cqe(struct io_ring_ctx
*ctx
, bool defer
, u64 user_data
, s32 res
, u32 cflags
,
940 struct io_uring_cqe
*cqe
;
944 return __io_post_aux_cqe(ctx
, user_data
, res
, cflags
, allow_overflow
);
946 length
= ARRAY_SIZE(ctx
->submit_state
.cqes
);
948 lockdep_assert_held(&ctx
->uring_lock
);
950 if (ctx
->submit_state
.cqes_count
== length
) {
952 __io_flush_post_cqes(ctx
);
953 /* no need to flush - flush is deferred */
954 __io_cq_unlock_post(ctx
);
957 /* For defered completions this is not as strict as it is otherwise,
958 * however it's main job is to prevent unbounded posted completions,
959 * and in that it works just as well.
961 if (!allow_overflow
&& test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
))
964 cqe
= &ctx
->submit_state
.cqes
[ctx
->submit_state
.cqes_count
++];
965 cqe
->user_data
= user_data
;
971 static void __io_req_complete_post(struct io_kiocb
*req
, unsigned issue_flags
)
973 struct io_ring_ctx
*ctx
= req
->ctx
;
974 struct io_rsrc_node
*rsrc_node
= NULL
;
977 if (!(req
->flags
& REQ_F_CQE_SKIP
))
978 io_fill_cqe_req(ctx
, req
);
981 * If we're the last reference to this request, add to our locked
984 if (req_ref_put_and_test(req
)) {
985 if (req
->flags
& IO_REQ_LINK_FLAGS
) {
986 if (req
->flags
& IO_DISARM_MASK
)
989 io_req_task_queue(req
->link
);
993 io_put_kbuf_comp(req
);
994 io_dismantle_req(req
);
995 rsrc_node
= req
->rsrc_node
;
997 * Selected buffer deallocation in io_clean_op() assumes that
998 * we don't hold ->completion_lock. Clean them here to avoid
1001 io_put_task_remote(req
->task
, 1);
1002 wq_list_add_head(&req
->comp_list
, &ctx
->locked_free_list
);
1003 ctx
->locked_free_nr
++;
1005 io_cq_unlock_post(ctx
);
1008 io_ring_submit_lock(ctx
, issue_flags
);
1009 io_put_rsrc_node(ctx
, rsrc_node
);
1010 io_ring_submit_unlock(ctx
, issue_flags
);
1014 void io_req_complete_post(struct io_kiocb
*req
, unsigned issue_flags
)
1016 if (req
->ctx
->task_complete
&& req
->ctx
->submitter_task
!= current
) {
1017 req
->io_task_work
.func
= io_req_task_complete
;
1018 io_req_task_work_add(req
);
1019 } else if (!(issue_flags
& IO_URING_F_UNLOCKED
) ||
1020 !(req
->ctx
->flags
& IORING_SETUP_IOPOLL
)) {
1021 __io_req_complete_post(req
, issue_flags
);
1023 struct io_ring_ctx
*ctx
= req
->ctx
;
1025 mutex_lock(&ctx
->uring_lock
);
1026 __io_req_complete_post(req
, issue_flags
& ~IO_URING_F_UNLOCKED
);
1027 mutex_unlock(&ctx
->uring_lock
);
1031 void io_req_defer_failed(struct io_kiocb
*req
, s32 res
)
1032 __must_hold(&ctx
->uring_lock
)
1034 const struct io_cold_def
*def
= &io_cold_defs
[req
->opcode
];
1036 lockdep_assert_held(&req
->ctx
->uring_lock
);
1039 io_req_set_res(req
, res
, io_put_kbuf(req
, IO_URING_F_UNLOCKED
));
1042 io_req_complete_defer(req
);
1046 * Don't initialise the fields below on every allocation, but do that in
1047 * advance and keep them valid across allocations.
1049 static void io_preinit_req(struct io_kiocb
*req
, struct io_ring_ctx
*ctx
)
1053 req
->async_data
= NULL
;
1054 /* not necessary, but safer to zero */
1058 static void io_flush_cached_locked_reqs(struct io_ring_ctx
*ctx
,
1059 struct io_submit_state
*state
)
1061 spin_lock(&ctx
->completion_lock
);
1062 wq_list_splice(&ctx
->locked_free_list
, &state
->free_list
);
1063 ctx
->locked_free_nr
= 0;
1064 spin_unlock(&ctx
->completion_lock
);
1068 * A request might get retired back into the request caches even before opcode
1069 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1070 * Because of that, io_alloc_req() should be called only under ->uring_lock
1071 * and with extra caution to not get a request that is still worked on.
1073 __cold
bool __io_alloc_req_refill(struct io_ring_ctx
*ctx
)
1074 __must_hold(&ctx
->uring_lock
)
1076 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
1077 void *reqs
[IO_REQ_ALLOC_BATCH
];
1081 * If we have more than a batch's worth of requests in our IRQ side
1082 * locked cache, grab the lock and move them over to our submission
1085 if (data_race(ctx
->locked_free_nr
) > IO_COMPL_BATCH
) {
1086 io_flush_cached_locked_reqs(ctx
, &ctx
->submit_state
);
1087 if (!io_req_cache_empty(ctx
))
1091 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, ARRAY_SIZE(reqs
), reqs
);
1094 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1095 * retry single alloc to be on the safe side.
1097 if (unlikely(ret
<= 0)) {
1098 reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
1104 percpu_ref_get_many(&ctx
->refs
, ret
);
1105 for (i
= 0; i
< ret
; i
++) {
1106 struct io_kiocb
*req
= reqs
[i
];
1108 io_preinit_req(req
, ctx
);
1109 io_req_add_to_cache(req
, ctx
);
1114 static inline void io_dismantle_req(struct io_kiocb
*req
)
1116 unsigned int flags
= req
->flags
;
1118 if (unlikely(flags
& IO_REQ_CLEAN_FLAGS
))
1120 if (!(flags
& REQ_F_FIXED_FILE
))
1121 io_put_file(req
->file
);
1124 static __cold
void io_free_req_tw(struct io_kiocb
*req
, struct io_tw_state
*ts
)
1126 struct io_ring_ctx
*ctx
= req
->ctx
;
1128 if (req
->rsrc_node
) {
1129 io_tw_lock(ctx
, ts
);
1130 io_put_rsrc_node(ctx
, req
->rsrc_node
);
1132 io_dismantle_req(req
);
1133 io_put_task_remote(req
->task
, 1);
1135 spin_lock(&ctx
->completion_lock
);
1136 wq_list_add_head(&req
->comp_list
, &ctx
->locked_free_list
);
1137 ctx
->locked_free_nr
++;
1138 spin_unlock(&ctx
->completion_lock
);
1141 __cold
void io_free_req(struct io_kiocb
*req
)
1143 req
->io_task_work
.func
= io_free_req_tw
;
1144 io_req_task_work_add(req
);
1147 static void __io_req_find_next_prep(struct io_kiocb
*req
)
1149 struct io_ring_ctx
*ctx
= req
->ctx
;
1151 spin_lock(&ctx
->completion_lock
);
1152 io_disarm_next(req
);
1153 spin_unlock(&ctx
->completion_lock
);
1156 static inline struct io_kiocb
*io_req_find_next(struct io_kiocb
*req
)
1158 struct io_kiocb
*nxt
;
1161 * If LINK is set, we have dependent requests in this chain. If we
1162 * didn't fail this request, queue the first one up, moving any other
1163 * dependencies to the next request. In case of failure, fail the rest
1166 if (unlikely(req
->flags
& IO_DISARM_MASK
))
1167 __io_req_find_next_prep(req
);
1173 static void ctx_flush_and_put(struct io_ring_ctx
*ctx
, struct io_tw_state
*ts
)
1177 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
)
1178 atomic_andnot(IORING_SQ_TASKRUN
, &ctx
->rings
->sq_flags
);
1180 io_submit_flush_completions(ctx
);
1181 mutex_unlock(&ctx
->uring_lock
);
1184 percpu_ref_put(&ctx
->refs
);
1187 static unsigned int handle_tw_list(struct llist_node
*node
,
1188 struct io_ring_ctx
**ctx
,
1189 struct io_tw_state
*ts
,
1190 struct llist_node
*last
)
1192 unsigned int count
= 0;
1194 while (node
&& node
!= last
) {
1195 struct llist_node
*next
= node
->next
;
1196 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1199 prefetch(container_of(next
, struct io_kiocb
, io_task_work
.node
));
1201 if (req
->ctx
!= *ctx
) {
1202 ctx_flush_and_put(*ctx
, ts
);
1204 /* if not contended, grab and improve batching */
1205 ts
->locked
= mutex_trylock(&(*ctx
)->uring_lock
);
1206 percpu_ref_get(&(*ctx
)->refs
);
1208 req
->io_task_work
.func(req
, ts
);
1211 if (unlikely(need_resched())) {
1212 ctx_flush_and_put(*ctx
, ts
);
1222 * io_llist_xchg - swap all entries in a lock-less list
1223 * @head: the head of lock-less list to delete all entries
1224 * @new: new entry as the head of the list
1226 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1227 * The order of entries returned is from the newest to the oldest added one.
1229 static inline struct llist_node
*io_llist_xchg(struct llist_head
*head
,
1230 struct llist_node
*new)
1232 return xchg(&head
->first
, new);
1236 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1237 * @head: the head of lock-less list to delete all entries
1238 * @old: expected old value of the first entry of the list
1239 * @new: new entry as the head of the list
1241 * perform a cmpxchg on the first entry of the list.
1244 static inline struct llist_node
*io_llist_cmpxchg(struct llist_head
*head
,
1245 struct llist_node
*old
,
1246 struct llist_node
*new)
1248 return cmpxchg(&head
->first
, old
, new);
1251 void tctx_task_work(struct callback_head
*cb
)
1253 struct io_tw_state ts
= {};
1254 struct io_ring_ctx
*ctx
= NULL
;
1255 struct io_uring_task
*tctx
= container_of(cb
, struct io_uring_task
,
1257 struct llist_node fake
= {};
1258 struct llist_node
*node
;
1259 unsigned int loops
= 0;
1260 unsigned int count
= 0;
1262 if (unlikely(current
->flags
& PF_EXITING
)) {
1263 io_fallback_tw(tctx
);
1269 node
= io_llist_xchg(&tctx
->task_list
, &fake
);
1270 count
+= handle_tw_list(node
, &ctx
, &ts
, &fake
);
1272 /* skip expensive cmpxchg if there are items in the list */
1273 if (READ_ONCE(tctx
->task_list
.first
) != &fake
)
1275 if (ts
.locked
&& !wq_list_empty(&ctx
->submit_state
.compl_reqs
)) {
1276 io_submit_flush_completions(ctx
);
1277 if (READ_ONCE(tctx
->task_list
.first
) != &fake
)
1280 node
= io_llist_cmpxchg(&tctx
->task_list
, &fake
, NULL
);
1281 } while (node
!= &fake
);
1283 ctx_flush_and_put(ctx
, &ts
);
1285 /* relaxed read is enough as only the task itself sets ->in_cancel */
1286 if (unlikely(atomic_read(&tctx
->in_cancel
)))
1287 io_uring_drop_tctx_refs(current
);
1289 trace_io_uring_task_work_run(tctx
, count
, loops
);
1292 static __cold
void io_fallback_tw(struct io_uring_task
*tctx
)
1294 struct llist_node
*node
= llist_del_all(&tctx
->task_list
);
1295 struct io_kiocb
*req
;
1298 req
= container_of(node
, struct io_kiocb
, io_task_work
.node
);
1300 if (llist_add(&req
->io_task_work
.node
,
1301 &req
->ctx
->fallback_llist
))
1302 schedule_delayed_work(&req
->ctx
->fallback_work
, 1);
1306 static void io_req_local_work_add(struct io_kiocb
*req
, unsigned flags
)
1308 struct io_ring_ctx
*ctx
= req
->ctx
;
1309 unsigned nr_wait
, nr_tw
, nr_tw_prev
;
1310 struct llist_node
*first
;
1312 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))
1313 flags
&= ~IOU_F_TWQ_LAZY_WAKE
;
1315 first
= READ_ONCE(ctx
->work_llist
.first
);
1319 struct io_kiocb
*first_req
= container_of(first
,
1323 * Might be executed at any moment, rely on
1324 * SLAB_TYPESAFE_BY_RCU to keep it alive.
1326 nr_tw_prev
= READ_ONCE(first_req
->nr_tw
);
1328 nr_tw
= nr_tw_prev
+ 1;
1329 /* Large enough to fail the nr_wait comparison below */
1330 if (!(flags
& IOU_F_TWQ_LAZY_WAKE
))
1334 req
->io_task_work
.node
.next
= first
;
1335 } while (!try_cmpxchg(&ctx
->work_llist
.first
, &first
,
1336 &req
->io_task_work
.node
));
1339 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
)
1340 atomic_or(IORING_SQ_TASKRUN
, &ctx
->rings
->sq_flags
);
1342 io_eventfd_signal(ctx
);
1345 nr_wait
= atomic_read(&ctx
->cq_wait_nr
);
1346 /* no one is waiting */
1349 /* either not enough or the previous add has already woken it up */
1350 if (nr_wait
> nr_tw
|| nr_tw_prev
>= nr_wait
)
1352 /* pairs with set_current_state() in io_cqring_wait() */
1353 smp_mb__after_atomic();
1354 wake_up_state(ctx
->submitter_task
, TASK_INTERRUPTIBLE
);
1357 void __io_req_task_work_add(struct io_kiocb
*req
, unsigned flags
)
1359 struct io_uring_task
*tctx
= req
->task
->io_uring
;
1360 struct io_ring_ctx
*ctx
= req
->ctx
;
1362 if (!(flags
& IOU_F_TWQ_FORCE_NORMAL
) &&
1363 (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
)) {
1365 io_req_local_work_add(req
, flags
);
1370 /* task_work already pending, we're done */
1371 if (!llist_add(&req
->io_task_work
.node
, &tctx
->task_list
))
1374 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
)
1375 atomic_or(IORING_SQ_TASKRUN
, &ctx
->rings
->sq_flags
);
1377 if (likely(!task_work_add(req
->task
, &tctx
->task_work
, ctx
->notify_method
)))
1380 io_fallback_tw(tctx
);
1383 static void __cold
io_move_task_work_from_local(struct io_ring_ctx
*ctx
)
1385 struct llist_node
*node
;
1387 node
= llist_del_all(&ctx
->work_llist
);
1389 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1393 __io_req_task_work_add(req
, IOU_F_TWQ_FORCE_NORMAL
);
1397 static int __io_run_local_work(struct io_ring_ctx
*ctx
, struct io_tw_state
*ts
)
1399 struct llist_node
*node
;
1400 unsigned int loops
= 0;
1403 if (WARN_ON_ONCE(ctx
->submitter_task
!= current
))
1405 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
)
1406 atomic_andnot(IORING_SQ_TASKRUN
, &ctx
->rings
->sq_flags
);
1408 node
= io_llist_xchg(&ctx
->work_llist
, NULL
);
1410 struct llist_node
*next
= node
->next
;
1411 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1413 prefetch(container_of(next
, struct io_kiocb
, io_task_work
.node
));
1414 req
->io_task_work
.func(req
, ts
);
1420 if (!llist_empty(&ctx
->work_llist
))
1423 io_submit_flush_completions(ctx
);
1424 if (!llist_empty(&ctx
->work_llist
))
1427 trace_io_uring_local_work_run(ctx
, ret
, loops
);
1431 static inline int io_run_local_work_locked(struct io_ring_ctx
*ctx
)
1433 struct io_tw_state ts
= { .locked
= true, };
1436 if (llist_empty(&ctx
->work_llist
))
1439 ret
= __io_run_local_work(ctx
, &ts
);
1440 /* shouldn't happen! */
1441 if (WARN_ON_ONCE(!ts
.locked
))
1442 mutex_lock(&ctx
->uring_lock
);
1446 static int io_run_local_work(struct io_ring_ctx
*ctx
)
1448 struct io_tw_state ts
= {};
1451 ts
.locked
= mutex_trylock(&ctx
->uring_lock
);
1452 ret
= __io_run_local_work(ctx
, &ts
);
1454 mutex_unlock(&ctx
->uring_lock
);
1459 static void io_req_task_cancel(struct io_kiocb
*req
, struct io_tw_state
*ts
)
1461 io_tw_lock(req
->ctx
, ts
);
1462 io_req_defer_failed(req
, req
->cqe
.res
);
1465 void io_req_task_submit(struct io_kiocb
*req
, struct io_tw_state
*ts
)
1467 io_tw_lock(req
->ctx
, ts
);
1468 /* req->task == current here, checking PF_EXITING is safe */
1469 if (unlikely(req
->task
->flags
& PF_EXITING
))
1470 io_req_defer_failed(req
, -EFAULT
);
1471 else if (req
->flags
& REQ_F_FORCE_ASYNC
)
1472 io_queue_iowq(req
, ts
);
1477 void io_req_task_queue_fail(struct io_kiocb
*req
, int ret
)
1479 io_req_set_res(req
, ret
, 0);
1480 req
->io_task_work
.func
= io_req_task_cancel
;
1481 io_req_task_work_add(req
);
1484 void io_req_task_queue(struct io_kiocb
*req
)
1486 req
->io_task_work
.func
= io_req_task_submit
;
1487 io_req_task_work_add(req
);
1490 void io_queue_next(struct io_kiocb
*req
)
1492 struct io_kiocb
*nxt
= io_req_find_next(req
);
1495 io_req_task_queue(nxt
);
1498 void io_free_batch_list(struct io_ring_ctx
*ctx
, struct io_wq_work_node
*node
)
1499 __must_hold(&ctx
->uring_lock
)
1501 struct task_struct
*task
= NULL
;
1505 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1508 if (unlikely(req
->flags
& IO_REQ_CLEAN_SLOW_FLAGS
)) {
1509 if (req
->flags
& REQ_F_REFCOUNT
) {
1510 node
= req
->comp_list
.next
;
1511 if (!req_ref_put_and_test(req
))
1514 if ((req
->flags
& REQ_F_POLLED
) && req
->apoll
) {
1515 struct async_poll
*apoll
= req
->apoll
;
1517 if (apoll
->double_poll
)
1518 kfree(apoll
->double_poll
);
1519 if (!io_alloc_cache_put(&ctx
->apoll_cache
, &apoll
->cache
))
1521 req
->flags
&= ~REQ_F_POLLED
;
1523 if (req
->flags
& IO_REQ_LINK_FLAGS
)
1525 if (unlikely(req
->flags
& IO_REQ_CLEAN_FLAGS
))
1528 if (!(req
->flags
& REQ_F_FIXED_FILE
))
1529 io_put_file(req
->file
);
1531 io_req_put_rsrc_locked(req
, ctx
);
1533 if (req
->task
!= task
) {
1535 io_put_task(task
, task_refs
);
1540 node
= req
->comp_list
.next
;
1541 io_req_add_to_cache(req
, ctx
);
1545 io_put_task(task
, task_refs
);
1548 static void __io_submit_flush_completions(struct io_ring_ctx
*ctx
)
1549 __must_hold(&ctx
->uring_lock
)
1551 struct io_submit_state
*state
= &ctx
->submit_state
;
1552 struct io_wq_work_node
*node
;
1555 /* must come first to preserve CQE ordering in failure cases */
1556 if (state
->cqes_count
)
1557 __io_flush_post_cqes(ctx
);
1558 __wq_list_for_each(node
, &state
->compl_reqs
) {
1559 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1562 if (!(req
->flags
& REQ_F_CQE_SKIP
) &&
1563 unlikely(!__io_fill_cqe_req(ctx
, req
))) {
1564 if (ctx
->task_complete
) {
1565 spin_lock(&ctx
->completion_lock
);
1566 io_req_cqe_overflow(req
);
1567 spin_unlock(&ctx
->completion_lock
);
1569 io_req_cqe_overflow(req
);
1573 __io_cq_unlock_post_flush(ctx
);
1575 if (!wq_list_empty(&ctx
->submit_state
.compl_reqs
)) {
1576 io_free_batch_list(ctx
, state
->compl_reqs
.first
);
1577 INIT_WQ_LIST(&state
->compl_reqs
);
1582 * Drop reference to request, return next in chain (if there is one) if this
1583 * was the last reference to this request.
1585 static inline struct io_kiocb
*io_put_req_find_next(struct io_kiocb
*req
)
1587 struct io_kiocb
*nxt
= NULL
;
1589 if (req_ref_put_and_test(req
)) {
1590 if (unlikely(req
->flags
& IO_REQ_LINK_FLAGS
))
1591 nxt
= io_req_find_next(req
);
1597 static unsigned io_cqring_events(struct io_ring_ctx
*ctx
)
1599 /* See comment at the top of this file */
1601 return __io_cqring_events(ctx
);
1605 * We can't just wait for polled events to come to us, we have to actively
1606 * find and complete them.
1608 static __cold
void io_iopoll_try_reap_events(struct io_ring_ctx
*ctx
)
1610 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
1613 mutex_lock(&ctx
->uring_lock
);
1614 while (!wq_list_empty(&ctx
->iopoll_list
)) {
1615 /* let it sleep and repeat later if can't complete a request */
1616 if (io_do_iopoll(ctx
, true) == 0)
1619 * Ensure we allow local-to-the-cpu processing to take place,
1620 * in this case we need to ensure that we reap all events.
1621 * Also let task_work, etc. to progress by releasing the mutex
1623 if (need_resched()) {
1624 mutex_unlock(&ctx
->uring_lock
);
1626 mutex_lock(&ctx
->uring_lock
);
1629 mutex_unlock(&ctx
->uring_lock
);
1632 static int io_iopoll_check(struct io_ring_ctx
*ctx
, long min
)
1634 unsigned int nr_events
= 0;
1636 unsigned long check_cq
;
1638 if (!io_allowed_run_tw(ctx
))
1641 check_cq
= READ_ONCE(ctx
->check_cq
);
1642 if (unlikely(check_cq
)) {
1643 if (check_cq
& BIT(IO_CHECK_CQ_OVERFLOW_BIT
))
1644 __io_cqring_overflow_flush(ctx
);
1646 * Similarly do not spin if we have not informed the user of any
1649 if (check_cq
& BIT(IO_CHECK_CQ_DROPPED_BIT
))
1653 * Don't enter poll loop if we already have events pending.
1654 * If we do, we can potentially be spinning for commands that
1655 * already triggered a CQE (eg in error).
1657 if (io_cqring_events(ctx
))
1662 * If a submit got punted to a workqueue, we can have the
1663 * application entering polling for a command before it gets
1664 * issued. That app will hold the uring_lock for the duration
1665 * of the poll right here, so we need to take a breather every
1666 * now and then to ensure that the issue has a chance to add
1667 * the poll to the issued list. Otherwise we can spin here
1668 * forever, while the workqueue is stuck trying to acquire the
1671 if (wq_list_empty(&ctx
->iopoll_list
) ||
1672 io_task_work_pending(ctx
)) {
1673 u32 tail
= ctx
->cached_cq_tail
;
1675 (void) io_run_local_work_locked(ctx
);
1677 if (task_work_pending(current
) ||
1678 wq_list_empty(&ctx
->iopoll_list
)) {
1679 mutex_unlock(&ctx
->uring_lock
);
1681 mutex_lock(&ctx
->uring_lock
);
1683 /* some requests don't go through iopoll_list */
1684 if (tail
!= ctx
->cached_cq_tail
||
1685 wq_list_empty(&ctx
->iopoll_list
))
1688 ret
= io_do_iopoll(ctx
, !min
);
1693 } while (nr_events
< min
&& !need_resched());
1698 void io_req_task_complete(struct io_kiocb
*req
, struct io_tw_state
*ts
)
1701 io_req_complete_defer(req
);
1703 io_req_complete_post(req
, IO_URING_F_UNLOCKED
);
1707 * After the iocb has been issued, it's safe to be found on the poll list.
1708 * Adding the kiocb to the list AFTER submission ensures that we don't
1709 * find it from a io_do_iopoll() thread before the issuer is done
1710 * accessing the kiocb cookie.
1712 static void io_iopoll_req_issued(struct io_kiocb
*req
, unsigned int issue_flags
)
1714 struct io_ring_ctx
*ctx
= req
->ctx
;
1715 const bool needs_lock
= issue_flags
& IO_URING_F_UNLOCKED
;
1717 /* workqueue context doesn't hold uring_lock, grab it now */
1718 if (unlikely(needs_lock
))
1719 mutex_lock(&ctx
->uring_lock
);
1722 * Track whether we have multiple files in our lists. This will impact
1723 * how we do polling eventually, not spinning if we're on potentially
1724 * different devices.
1726 if (wq_list_empty(&ctx
->iopoll_list
)) {
1727 ctx
->poll_multi_queue
= false;
1728 } else if (!ctx
->poll_multi_queue
) {
1729 struct io_kiocb
*list_req
;
1731 list_req
= container_of(ctx
->iopoll_list
.first
, struct io_kiocb
,
1733 if (list_req
->file
!= req
->file
)
1734 ctx
->poll_multi_queue
= true;
1738 * For fast devices, IO may have already completed. If it has, add
1739 * it to the front so we find it first.
1741 if (READ_ONCE(req
->iopoll_completed
))
1742 wq_list_add_head(&req
->comp_list
, &ctx
->iopoll_list
);
1744 wq_list_add_tail(&req
->comp_list
, &ctx
->iopoll_list
);
1746 if (unlikely(needs_lock
)) {
1748 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1749 * in sq thread task context or in io worker task context. If
1750 * current task context is sq thread, we don't need to check
1751 * whether should wake up sq thread.
1753 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) &&
1754 wq_has_sleeper(&ctx
->sq_data
->wait
))
1755 wake_up(&ctx
->sq_data
->wait
);
1757 mutex_unlock(&ctx
->uring_lock
);
1761 static bool io_bdev_nowait(struct block_device
*bdev
)
1763 return !bdev
|| bdev_nowait(bdev
);
1767 * If we tracked the file through the SCM inflight mechanism, we could support
1768 * any file. For now, just ensure that anything potentially problematic is done
1771 static bool __io_file_supports_nowait(struct file
*file
, umode_t mode
)
1773 if (S_ISBLK(mode
)) {
1774 if (IS_ENABLED(CONFIG_BLOCK
) &&
1775 io_bdev_nowait(I_BDEV(file
->f_mapping
->host
)))
1781 if (S_ISREG(mode
)) {
1782 if (IS_ENABLED(CONFIG_BLOCK
) &&
1783 io_bdev_nowait(file
->f_inode
->i_sb
->s_bdev
) &&
1784 !io_is_uring_fops(file
))
1789 /* any ->read/write should understand O_NONBLOCK */
1790 if (file
->f_flags
& O_NONBLOCK
)
1792 return file
->f_mode
& FMODE_NOWAIT
;
1796 * If we tracked the file through the SCM inflight mechanism, we could support
1797 * any file. For now, just ensure that anything potentially problematic is done
1800 unsigned int io_file_get_flags(struct file
*file
)
1802 umode_t mode
= file_inode(file
)->i_mode
;
1803 unsigned int res
= 0;
1807 if (__io_file_supports_nowait(file
, mode
))
1812 bool io_alloc_async_data(struct io_kiocb
*req
)
1814 WARN_ON_ONCE(!io_cold_defs
[req
->opcode
].async_size
);
1815 req
->async_data
= kmalloc(io_cold_defs
[req
->opcode
].async_size
, GFP_KERNEL
);
1816 if (req
->async_data
) {
1817 req
->flags
|= REQ_F_ASYNC_DATA
;
1823 int io_req_prep_async(struct io_kiocb
*req
)
1825 const struct io_cold_def
*cdef
= &io_cold_defs
[req
->opcode
];
1826 const struct io_issue_def
*def
= &io_issue_defs
[req
->opcode
];
1828 /* assign early for deferred execution for non-fixed file */
1829 if (def
->needs_file
&& !(req
->flags
& REQ_F_FIXED_FILE
) && !req
->file
)
1830 req
->file
= io_file_get_normal(req
, req
->cqe
.fd
);
1831 if (!cdef
->prep_async
)
1833 if (WARN_ON_ONCE(req_has_async_data(req
)))
1835 if (!def
->manual_alloc
) {
1836 if (io_alloc_async_data(req
))
1839 return cdef
->prep_async(req
);
1842 static u32
io_get_sequence(struct io_kiocb
*req
)
1844 u32 seq
= req
->ctx
->cached_sq_head
;
1845 struct io_kiocb
*cur
;
1847 /* need original cached_sq_head, but it was increased for each req */
1848 io_for_each_link(cur
, req
)
1853 static __cold
void io_drain_req(struct io_kiocb
*req
)
1854 __must_hold(&ctx
->uring_lock
)
1856 struct io_ring_ctx
*ctx
= req
->ctx
;
1857 struct io_defer_entry
*de
;
1859 u32 seq
= io_get_sequence(req
);
1861 /* Still need defer if there is pending req in defer list. */
1862 spin_lock(&ctx
->completion_lock
);
1863 if (!req_need_defer(req
, seq
) && list_empty_careful(&ctx
->defer_list
)) {
1864 spin_unlock(&ctx
->completion_lock
);
1866 ctx
->drain_active
= false;
1867 io_req_task_queue(req
);
1870 spin_unlock(&ctx
->completion_lock
);
1872 io_prep_async_link(req
);
1873 de
= kmalloc(sizeof(*de
), GFP_KERNEL
);
1876 io_req_defer_failed(req
, ret
);
1880 spin_lock(&ctx
->completion_lock
);
1881 if (!req_need_defer(req
, seq
) && list_empty(&ctx
->defer_list
)) {
1882 spin_unlock(&ctx
->completion_lock
);
1887 trace_io_uring_defer(req
);
1890 list_add_tail(&de
->list
, &ctx
->defer_list
);
1891 spin_unlock(&ctx
->completion_lock
);
1894 static void io_clean_op(struct io_kiocb
*req
)
1896 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
1897 spin_lock(&req
->ctx
->completion_lock
);
1898 io_put_kbuf_comp(req
);
1899 spin_unlock(&req
->ctx
->completion_lock
);
1902 if (req
->flags
& REQ_F_NEED_CLEANUP
) {
1903 const struct io_cold_def
*def
= &io_cold_defs
[req
->opcode
];
1908 if ((req
->flags
& REQ_F_POLLED
) && req
->apoll
) {
1909 kfree(req
->apoll
->double_poll
);
1913 if (req
->flags
& REQ_F_INFLIGHT
) {
1914 struct io_uring_task
*tctx
= req
->task
->io_uring
;
1916 atomic_dec(&tctx
->inflight_tracked
);
1918 if (req
->flags
& REQ_F_CREDS
)
1919 put_cred(req
->creds
);
1920 if (req
->flags
& REQ_F_ASYNC_DATA
) {
1921 kfree(req
->async_data
);
1922 req
->async_data
= NULL
;
1924 req
->flags
&= ~IO_REQ_CLEAN_FLAGS
;
1927 static bool io_assign_file(struct io_kiocb
*req
, const struct io_issue_def
*def
,
1928 unsigned int issue_flags
)
1930 if (req
->file
|| !def
->needs_file
)
1933 if (req
->flags
& REQ_F_FIXED_FILE
)
1934 req
->file
= io_file_get_fixed(req
, req
->cqe
.fd
, issue_flags
);
1936 req
->file
= io_file_get_normal(req
, req
->cqe
.fd
);
1941 static int io_issue_sqe(struct io_kiocb
*req
, unsigned int issue_flags
)
1943 const struct io_issue_def
*def
= &io_issue_defs
[req
->opcode
];
1944 const struct cred
*creds
= NULL
;
1947 if (unlikely(!io_assign_file(req
, def
, issue_flags
)))
1950 if (unlikely((req
->flags
& REQ_F_CREDS
) && req
->creds
!= current_cred()))
1951 creds
= override_creds(req
->creds
);
1953 if (!def
->audit_skip
)
1954 audit_uring_entry(req
->opcode
);
1956 ret
= def
->issue(req
, issue_flags
);
1958 if (!def
->audit_skip
)
1959 audit_uring_exit(!ret
, ret
);
1962 revert_creds(creds
);
1964 if (ret
== IOU_OK
) {
1965 if (issue_flags
& IO_URING_F_COMPLETE_DEFER
)
1966 io_req_complete_defer(req
);
1968 io_req_complete_post(req
, issue_flags
);
1969 } else if (ret
!= IOU_ISSUE_SKIP_COMPLETE
)
1972 /* If the op doesn't have a file, we're not polling for it */
1973 if ((req
->ctx
->flags
& IORING_SETUP_IOPOLL
) && def
->iopoll_queue
)
1974 io_iopoll_req_issued(req
, issue_flags
);
1979 int io_poll_issue(struct io_kiocb
*req
, struct io_tw_state
*ts
)
1981 io_tw_lock(req
->ctx
, ts
);
1982 return io_issue_sqe(req
, IO_URING_F_NONBLOCK
|IO_URING_F_MULTISHOT
|
1983 IO_URING_F_COMPLETE_DEFER
);
1986 struct io_wq_work
*io_wq_free_work(struct io_wq_work
*work
)
1988 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
1990 req
= io_put_req_find_next(req
);
1991 return req
? &req
->work
: NULL
;
1994 void io_wq_submit_work(struct io_wq_work
*work
)
1996 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
1997 const struct io_issue_def
*def
= &io_issue_defs
[req
->opcode
];
1998 unsigned int issue_flags
= IO_URING_F_UNLOCKED
| IO_URING_F_IOWQ
;
1999 bool needs_poll
= false;
2000 int ret
= 0, err
= -ECANCELED
;
2002 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
2003 if (!(req
->flags
& REQ_F_REFCOUNT
))
2004 __io_req_set_refcount(req
, 2);
2008 io_arm_ltimeout(req
);
2010 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
2011 if (work
->flags
& IO_WQ_WORK_CANCEL
) {
2013 io_req_task_queue_fail(req
, err
);
2016 if (!io_assign_file(req
, def
, issue_flags
)) {
2018 work
->flags
|= IO_WQ_WORK_CANCEL
;
2022 if (req
->flags
& REQ_F_FORCE_ASYNC
) {
2023 bool opcode_poll
= def
->pollin
|| def
->pollout
;
2025 if (opcode_poll
&& file_can_poll(req
->file
)) {
2027 issue_flags
|= IO_URING_F_NONBLOCK
;
2032 ret
= io_issue_sqe(req
, issue_flags
);
2036 * We can get EAGAIN for iopolled IO even though we're
2037 * forcing a sync submission from here, since we can't
2038 * wait for request slots on the block side.
2041 if (!(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
2047 if (io_arm_poll_handler(req
, issue_flags
) == IO_APOLL_OK
)
2049 /* aborted or ready, in either case retry blocking */
2051 issue_flags
&= ~IO_URING_F_NONBLOCK
;
2054 /* avoid locking problems by failing it from a clean context */
2056 io_req_task_queue_fail(req
, ret
);
2059 inline struct file
*io_file_get_fixed(struct io_kiocb
*req
, int fd
,
2060 unsigned int issue_flags
)
2062 struct io_ring_ctx
*ctx
= req
->ctx
;
2063 struct file
*file
= NULL
;
2064 unsigned long file_ptr
;
2066 io_ring_submit_lock(ctx
, issue_flags
);
2068 if (unlikely((unsigned int)fd
>= ctx
->nr_user_files
))
2070 fd
= array_index_nospec(fd
, ctx
->nr_user_files
);
2071 file_ptr
= io_fixed_file_slot(&ctx
->file_table
, fd
)->file_ptr
;
2072 file
= (struct file
*) (file_ptr
& FFS_MASK
);
2073 file_ptr
&= ~FFS_MASK
;
2074 /* mask in overlapping REQ_F and FFS bits */
2075 req
->flags
|= (file_ptr
<< REQ_F_SUPPORT_NOWAIT_BIT
);
2076 io_req_set_rsrc_node(req
, ctx
, 0);
2078 io_ring_submit_unlock(ctx
, issue_flags
);
2082 struct file
*io_file_get_normal(struct io_kiocb
*req
, int fd
)
2084 struct file
*file
= fget(fd
);
2086 trace_io_uring_file_get(req
, fd
);
2088 /* we don't allow fixed io_uring files */
2089 if (file
&& io_is_uring_fops(file
))
2090 io_req_track_inflight(req
);
2094 static void io_queue_async(struct io_kiocb
*req
, int ret
)
2095 __must_hold(&req
->ctx
->uring_lock
)
2097 struct io_kiocb
*linked_timeout
;
2099 if (ret
!= -EAGAIN
|| (req
->flags
& REQ_F_NOWAIT
)) {
2100 io_req_defer_failed(req
, ret
);
2104 linked_timeout
= io_prep_linked_timeout(req
);
2106 switch (io_arm_poll_handler(req
, 0)) {
2107 case IO_APOLL_READY
:
2108 io_kbuf_recycle(req
, 0);
2109 io_req_task_queue(req
);
2111 case IO_APOLL_ABORTED
:
2112 io_kbuf_recycle(req
, 0);
2113 io_queue_iowq(req
, NULL
);
2120 io_queue_linked_timeout(linked_timeout
);
2123 static inline void io_queue_sqe(struct io_kiocb
*req
)
2124 __must_hold(&req
->ctx
->uring_lock
)
2128 ret
= io_issue_sqe(req
, IO_URING_F_NONBLOCK
|IO_URING_F_COMPLETE_DEFER
);
2131 * We async punt it if the file wasn't marked NOWAIT, or if the file
2132 * doesn't support non-blocking read/write attempts
2135 io_arm_ltimeout(req
);
2137 io_queue_async(req
, ret
);
2140 static void io_queue_sqe_fallback(struct io_kiocb
*req
)
2141 __must_hold(&req
->ctx
->uring_lock
)
2143 if (unlikely(req
->flags
& REQ_F_FAIL
)) {
2145 * We don't submit, fail them all, for that replace hardlinks
2146 * with normal links. Extra REQ_F_LINK is tolerated.
2148 req
->flags
&= ~REQ_F_HARDLINK
;
2149 req
->flags
|= REQ_F_LINK
;
2150 io_req_defer_failed(req
, req
->cqe
.res
);
2152 int ret
= io_req_prep_async(req
);
2154 if (unlikely(ret
)) {
2155 io_req_defer_failed(req
, ret
);
2159 if (unlikely(req
->ctx
->drain_active
))
2162 io_queue_iowq(req
, NULL
);
2167 * Check SQE restrictions (opcode and flags).
2169 * Returns 'true' if SQE is allowed, 'false' otherwise.
2171 static inline bool io_check_restriction(struct io_ring_ctx
*ctx
,
2172 struct io_kiocb
*req
,
2173 unsigned int sqe_flags
)
2175 if (!test_bit(req
->opcode
, ctx
->restrictions
.sqe_op
))
2178 if ((sqe_flags
& ctx
->restrictions
.sqe_flags_required
) !=
2179 ctx
->restrictions
.sqe_flags_required
)
2182 if (sqe_flags
& ~(ctx
->restrictions
.sqe_flags_allowed
|
2183 ctx
->restrictions
.sqe_flags_required
))
2189 static void io_init_req_drain(struct io_kiocb
*req
)
2191 struct io_ring_ctx
*ctx
= req
->ctx
;
2192 struct io_kiocb
*head
= ctx
->submit_state
.link
.head
;
2194 ctx
->drain_active
= true;
2197 * If we need to drain a request in the middle of a link, drain
2198 * the head request and the next request/link after the current
2199 * link. Considering sequential execution of links,
2200 * REQ_F_IO_DRAIN will be maintained for every request of our
2203 head
->flags
|= REQ_F_IO_DRAIN
| REQ_F_FORCE_ASYNC
;
2204 ctx
->drain_next
= true;
2208 static int io_init_req(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
2209 const struct io_uring_sqe
*sqe
)
2210 __must_hold(&ctx
->uring_lock
)
2212 const struct io_issue_def
*def
;
2213 unsigned int sqe_flags
;
2217 /* req is partially pre-initialised, see io_preinit_req() */
2218 req
->opcode
= opcode
= READ_ONCE(sqe
->opcode
);
2219 /* same numerical values with corresponding REQ_F_*, safe to copy */
2220 req
->flags
= sqe_flags
= READ_ONCE(sqe
->flags
);
2221 req
->cqe
.user_data
= READ_ONCE(sqe
->user_data
);
2223 req
->rsrc_node
= NULL
;
2224 req
->task
= current
;
2226 if (unlikely(opcode
>= IORING_OP_LAST
)) {
2230 def
= &io_issue_defs
[opcode
];
2231 if (unlikely(sqe_flags
& ~SQE_COMMON_FLAGS
)) {
2232 /* enforce forwards compatibility on users */
2233 if (sqe_flags
& ~SQE_VALID_FLAGS
)
2235 if (sqe_flags
& IOSQE_BUFFER_SELECT
) {
2236 if (!def
->buffer_select
)
2238 req
->buf_index
= READ_ONCE(sqe
->buf_group
);
2240 if (sqe_flags
& IOSQE_CQE_SKIP_SUCCESS
)
2241 ctx
->drain_disabled
= true;
2242 if (sqe_flags
& IOSQE_IO_DRAIN
) {
2243 if (ctx
->drain_disabled
)
2245 io_init_req_drain(req
);
2248 if (unlikely(ctx
->restricted
|| ctx
->drain_active
|| ctx
->drain_next
)) {
2249 if (ctx
->restricted
&& !io_check_restriction(ctx
, req
, sqe_flags
))
2251 /* knock it to the slow queue path, will be drained there */
2252 if (ctx
->drain_active
)
2253 req
->flags
|= REQ_F_FORCE_ASYNC
;
2254 /* if there is no link, we're at "next" request and need to drain */
2255 if (unlikely(ctx
->drain_next
) && !ctx
->submit_state
.link
.head
) {
2256 ctx
->drain_next
= false;
2257 ctx
->drain_active
= true;
2258 req
->flags
|= REQ_F_IO_DRAIN
| REQ_F_FORCE_ASYNC
;
2262 if (!def
->ioprio
&& sqe
->ioprio
)
2264 if (!def
->iopoll
&& (ctx
->flags
& IORING_SETUP_IOPOLL
))
2267 if (def
->needs_file
) {
2268 struct io_submit_state
*state
= &ctx
->submit_state
;
2270 req
->cqe
.fd
= READ_ONCE(sqe
->fd
);
2273 * Plug now if we have more than 2 IO left after this, and the
2274 * target is potentially a read/write to block based storage.
2276 if (state
->need_plug
&& def
->plug
) {
2277 state
->plug_started
= true;
2278 state
->need_plug
= false;
2279 blk_start_plug_nr_ios(&state
->plug
, state
->submit_nr
);
2283 personality
= READ_ONCE(sqe
->personality
);
2287 req
->creds
= xa_load(&ctx
->personalities
, personality
);
2290 get_cred(req
->creds
);
2291 ret
= security_uring_override_creds(req
->creds
);
2293 put_cred(req
->creds
);
2296 req
->flags
|= REQ_F_CREDS
;
2299 return def
->prep(req
, sqe
);
2302 static __cold
int io_submit_fail_init(const struct io_uring_sqe
*sqe
,
2303 struct io_kiocb
*req
, int ret
)
2305 struct io_ring_ctx
*ctx
= req
->ctx
;
2306 struct io_submit_link
*link
= &ctx
->submit_state
.link
;
2307 struct io_kiocb
*head
= link
->head
;
2309 trace_io_uring_req_failed(sqe
, req
, ret
);
2312 * Avoid breaking links in the middle as it renders links with SQPOLL
2313 * unusable. Instead of failing eagerly, continue assembling the link if
2314 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2315 * should find the flag and handle the rest.
2317 req_fail_link_node(req
, ret
);
2318 if (head
&& !(head
->flags
& REQ_F_FAIL
))
2319 req_fail_link_node(head
, -ECANCELED
);
2321 if (!(req
->flags
& IO_REQ_LINK_FLAGS
)) {
2323 link
->last
->link
= req
;
2327 io_queue_sqe_fallback(req
);
2332 link
->last
->link
= req
;
2339 static inline int io_submit_sqe(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
2340 const struct io_uring_sqe
*sqe
)
2341 __must_hold(&ctx
->uring_lock
)
2343 struct io_submit_link
*link
= &ctx
->submit_state
.link
;
2346 ret
= io_init_req(ctx
, req
, sqe
);
2348 return io_submit_fail_init(sqe
, req
, ret
);
2350 trace_io_uring_submit_req(req
);
2353 * If we already have a head request, queue this one for async
2354 * submittal once the head completes. If we don't have a head but
2355 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2356 * submitted sync once the chain is complete. If none of those
2357 * conditions are true (normal request), then just queue it.
2359 if (unlikely(link
->head
)) {
2360 ret
= io_req_prep_async(req
);
2362 return io_submit_fail_init(sqe
, req
, ret
);
2364 trace_io_uring_link(req
, link
->head
);
2365 link
->last
->link
= req
;
2368 if (req
->flags
& IO_REQ_LINK_FLAGS
)
2370 /* last request of the link, flush it */
2373 if (req
->flags
& (REQ_F_FORCE_ASYNC
| REQ_F_FAIL
))
2376 } else if (unlikely(req
->flags
& (IO_REQ_LINK_FLAGS
|
2377 REQ_F_FORCE_ASYNC
| REQ_F_FAIL
))) {
2378 if (req
->flags
& IO_REQ_LINK_FLAGS
) {
2383 io_queue_sqe_fallback(req
);
2393 * Batched submission is done, ensure local IO is flushed out.
2395 static void io_submit_state_end(struct io_ring_ctx
*ctx
)
2397 struct io_submit_state
*state
= &ctx
->submit_state
;
2399 if (unlikely(state
->link
.head
))
2400 io_queue_sqe_fallback(state
->link
.head
);
2401 /* flush only after queuing links as they can generate completions */
2402 io_submit_flush_completions(ctx
);
2403 if (state
->plug_started
)
2404 blk_finish_plug(&state
->plug
);
2408 * Start submission side cache.
2410 static void io_submit_state_start(struct io_submit_state
*state
,
2411 unsigned int max_ios
)
2413 state
->plug_started
= false;
2414 state
->need_plug
= max_ios
> 2;
2415 state
->submit_nr
= max_ios
;
2416 /* set only head, no need to init link_last in advance */
2417 state
->link
.head
= NULL
;
2420 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
2422 struct io_rings
*rings
= ctx
->rings
;
2425 * Ensure any loads from the SQEs are done at this point,
2426 * since once we write the new head, the application could
2427 * write new data to them.
2429 smp_store_release(&rings
->sq
.head
, ctx
->cached_sq_head
);
2433 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2434 * that is mapped by userspace. This means that care needs to be taken to
2435 * ensure that reads are stable, as we cannot rely on userspace always
2436 * being a good citizen. If members of the sqe are validated and then later
2437 * used, it's important that those reads are done through READ_ONCE() to
2438 * prevent a re-load down the line.
2440 static bool io_get_sqe(struct io_ring_ctx
*ctx
, const struct io_uring_sqe
**sqe
)
2442 unsigned head
, mask
= ctx
->sq_entries
- 1;
2443 unsigned sq_idx
= ctx
->cached_sq_head
++ & mask
;
2446 * The cached sq head (or cq tail) serves two purposes:
2448 * 1) allows us to batch the cost of updating the user visible
2450 * 2) allows the kernel side to track the head on its own, even
2451 * though the application is the one updating it.
2453 head
= READ_ONCE(ctx
->sq_array
[sq_idx
]);
2454 if (likely(head
< ctx
->sq_entries
)) {
2455 /* double index for 128-byte SQEs, twice as long */
2456 if (ctx
->flags
& IORING_SETUP_SQE128
)
2458 *sqe
= &ctx
->sq_sqes
[head
];
2462 /* drop invalid entries */
2464 WRITE_ONCE(ctx
->rings
->sq_dropped
,
2465 READ_ONCE(ctx
->rings
->sq_dropped
) + 1);
2469 int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
)
2470 __must_hold(&ctx
->uring_lock
)
2472 unsigned int entries
= io_sqring_entries(ctx
);
2476 if (unlikely(!entries
))
2478 /* make sure SQ entry isn't read before tail */
2479 ret
= left
= min(nr
, entries
);
2480 io_get_task_refs(left
);
2481 io_submit_state_start(&ctx
->submit_state
, left
);
2484 const struct io_uring_sqe
*sqe
;
2485 struct io_kiocb
*req
;
2487 if (unlikely(!io_alloc_req(ctx
, &req
)))
2489 if (unlikely(!io_get_sqe(ctx
, &sqe
))) {
2490 io_req_add_to_cache(req
, ctx
);
2495 * Continue submitting even for sqe failure if the
2496 * ring was setup with IORING_SETUP_SUBMIT_ALL
2498 if (unlikely(io_submit_sqe(ctx
, req
, sqe
)) &&
2499 !(ctx
->flags
& IORING_SETUP_SUBMIT_ALL
)) {
2505 if (unlikely(left
)) {
2507 /* try again if it submitted nothing and can't allocate a req */
2508 if (!ret
&& io_req_cache_empty(ctx
))
2510 current
->io_uring
->cached_refs
+= left
;
2513 io_submit_state_end(ctx
);
2514 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2515 io_commit_sqring(ctx
);
2519 struct io_wait_queue
{
2520 struct wait_queue_entry wq
;
2521 struct io_ring_ctx
*ctx
;
2523 unsigned nr_timeouts
;
2527 static inline bool io_has_work(struct io_ring_ctx
*ctx
)
2529 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
) ||
2530 !llist_empty(&ctx
->work_llist
);
2533 static inline bool io_should_wake(struct io_wait_queue
*iowq
)
2535 struct io_ring_ctx
*ctx
= iowq
->ctx
;
2536 int dist
= READ_ONCE(ctx
->rings
->cq
.tail
) - (int) iowq
->cq_tail
;
2539 * Wake up if we have enough events, or if a timeout occurred since we
2540 * started waiting. For timeouts, we always want to return to userspace,
2541 * regardless of event count.
2543 return dist
>= 0 || atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
2546 static int io_wake_function(struct wait_queue_entry
*curr
, unsigned int mode
,
2547 int wake_flags
, void *key
)
2549 struct io_wait_queue
*iowq
= container_of(curr
, struct io_wait_queue
, wq
);
2552 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2553 * the task, and the next invocation will do it.
2555 if (io_should_wake(iowq
) || io_has_work(iowq
->ctx
))
2556 return autoremove_wake_function(curr
, mode
, wake_flags
, key
);
2560 int io_run_task_work_sig(struct io_ring_ctx
*ctx
)
2562 if (!llist_empty(&ctx
->work_llist
)) {
2563 __set_current_state(TASK_RUNNING
);
2564 if (io_run_local_work(ctx
) > 0)
2567 if (io_run_task_work() > 0)
2569 if (task_sigpending(current
))
2574 /* when returns >0, the caller should retry */
2575 static inline int io_cqring_wait_schedule(struct io_ring_ctx
*ctx
,
2576 struct io_wait_queue
*iowq
)
2578 if (unlikely(READ_ONCE(ctx
->check_cq
)))
2580 if (unlikely(!llist_empty(&ctx
->work_llist
)))
2582 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL
)))
2584 if (unlikely(task_sigpending(current
)))
2586 if (unlikely(io_should_wake(iowq
)))
2588 if (iowq
->timeout
== KTIME_MAX
)
2590 else if (!schedule_hrtimeout(&iowq
->timeout
, HRTIMER_MODE_ABS
))
2596 * Wait until events become available, if we don't already have some. The
2597 * application must reap them itself, as they reside on the shared cq ring.
2599 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
2600 const sigset_t __user
*sig
, size_t sigsz
,
2601 struct __kernel_timespec __user
*uts
)
2603 struct io_wait_queue iowq
;
2604 struct io_rings
*rings
= ctx
->rings
;
2607 if (!io_allowed_run_tw(ctx
))
2609 if (!llist_empty(&ctx
->work_llist
))
2610 io_run_local_work(ctx
);
2612 io_cqring_overflow_flush(ctx
);
2613 /* if user messes with these they will just get an early return */
2614 if (__io_cqring_events_user(ctx
) >= min_events
)
2618 #ifdef CONFIG_COMPAT
2619 if (in_compat_syscall())
2620 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
2624 ret
= set_user_sigmask(sig
, sigsz
);
2630 init_waitqueue_func_entry(&iowq
.wq
, io_wake_function
);
2631 iowq
.wq
.private = current
;
2632 INIT_LIST_HEAD(&iowq
.wq
.entry
);
2634 iowq
.nr_timeouts
= atomic_read(&ctx
->cq_timeouts
);
2635 iowq
.cq_tail
= READ_ONCE(ctx
->rings
->cq
.head
) + min_events
;
2636 iowq
.timeout
= KTIME_MAX
;
2639 struct timespec64 ts
;
2641 if (get_timespec64(&ts
, uts
))
2643 iowq
.timeout
= ktime_add_ns(timespec64_to_ktime(ts
), ktime_get_ns());
2646 trace_io_uring_cqring_wait(ctx
, min_events
);
2648 unsigned long check_cq
;
2650 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) {
2651 int nr_wait
= (int) iowq
.cq_tail
- READ_ONCE(ctx
->rings
->cq
.tail
);
2653 atomic_set(&ctx
->cq_wait_nr
, nr_wait
);
2654 set_current_state(TASK_INTERRUPTIBLE
);
2656 prepare_to_wait_exclusive(&ctx
->cq_wait
, &iowq
.wq
,
2657 TASK_INTERRUPTIBLE
);
2660 ret
= io_cqring_wait_schedule(ctx
, &iowq
);
2661 __set_current_state(TASK_RUNNING
);
2662 atomic_set(&ctx
->cq_wait_nr
, 0);
2667 * Run task_work after scheduling and before io_should_wake().
2668 * If we got woken because of task_work being processed, run it
2669 * now rather than let the caller do another wait loop.
2672 if (!llist_empty(&ctx
->work_llist
))
2673 io_run_local_work(ctx
);
2675 check_cq
= READ_ONCE(ctx
->check_cq
);
2676 if (unlikely(check_cq
)) {
2677 /* let the caller flush overflows, retry */
2678 if (check_cq
& BIT(IO_CHECK_CQ_OVERFLOW_BIT
))
2679 io_cqring_do_overflow_flush(ctx
);
2680 if (check_cq
& BIT(IO_CHECK_CQ_DROPPED_BIT
)) {
2686 if (io_should_wake(&iowq
)) {
2693 if (!(ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
))
2694 finish_wait(&ctx
->cq_wait
, &iowq
.wq
);
2695 restore_saved_sigmask_unless(ret
== -EINTR
);
2697 return READ_ONCE(rings
->cq
.head
) == READ_ONCE(rings
->cq
.tail
) ? ret
: 0;
2700 static void io_mem_free(void *ptr
)
2707 page
= virt_to_head_page(ptr
);
2708 if (put_page_testzero(page
))
2709 free_compound_page(page
);
2712 static void *io_mem_alloc(size_t size
)
2714 gfp_t gfp
= GFP_KERNEL_ACCOUNT
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
;
2716 return (void *) __get_free_pages(gfp
, get_order(size
));
2719 static unsigned long rings_size(struct io_ring_ctx
*ctx
, unsigned int sq_entries
,
2720 unsigned int cq_entries
, size_t *sq_offset
)
2722 struct io_rings
*rings
;
2723 size_t off
, sq_array_size
;
2725 off
= struct_size(rings
, cqes
, cq_entries
);
2726 if (off
== SIZE_MAX
)
2728 if (ctx
->flags
& IORING_SETUP_CQE32
) {
2729 if (check_shl_overflow(off
, 1, &off
))
2734 off
= ALIGN(off
, SMP_CACHE_BYTES
);
2742 sq_array_size
= array_size(sizeof(u32
), sq_entries
);
2743 if (sq_array_size
== SIZE_MAX
)
2746 if (check_add_overflow(off
, sq_array_size
, &off
))
2752 static int io_eventfd_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
2753 unsigned int eventfd_async
)
2755 struct io_ev_fd
*ev_fd
;
2756 __s32 __user
*fds
= arg
;
2759 ev_fd
= rcu_dereference_protected(ctx
->io_ev_fd
,
2760 lockdep_is_held(&ctx
->uring_lock
));
2764 if (copy_from_user(&fd
, fds
, sizeof(*fds
)))
2767 ev_fd
= kmalloc(sizeof(*ev_fd
), GFP_KERNEL
);
2771 ev_fd
->cq_ev_fd
= eventfd_ctx_fdget(fd
);
2772 if (IS_ERR(ev_fd
->cq_ev_fd
)) {
2773 int ret
= PTR_ERR(ev_fd
->cq_ev_fd
);
2778 spin_lock(&ctx
->completion_lock
);
2779 ctx
->evfd_last_cq_tail
= ctx
->cached_cq_tail
;
2780 spin_unlock(&ctx
->completion_lock
);
2782 ev_fd
->eventfd_async
= eventfd_async
;
2783 ctx
->has_evfd
= true;
2784 rcu_assign_pointer(ctx
->io_ev_fd
, ev_fd
);
2785 atomic_set(&ev_fd
->refs
, 1);
2786 atomic_set(&ev_fd
->ops
, 0);
2790 static int io_eventfd_unregister(struct io_ring_ctx
*ctx
)
2792 struct io_ev_fd
*ev_fd
;
2794 ev_fd
= rcu_dereference_protected(ctx
->io_ev_fd
,
2795 lockdep_is_held(&ctx
->uring_lock
));
2797 ctx
->has_evfd
= false;
2798 rcu_assign_pointer(ctx
->io_ev_fd
, NULL
);
2799 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT
), &ev_fd
->ops
))
2800 call_rcu(&ev_fd
->rcu
, io_eventfd_ops
);
2807 static void io_req_caches_free(struct io_ring_ctx
*ctx
)
2809 struct io_kiocb
*req
;
2812 mutex_lock(&ctx
->uring_lock
);
2813 io_flush_cached_locked_reqs(ctx
, &ctx
->submit_state
);
2815 while (!io_req_cache_empty(ctx
)) {
2816 req
= io_extract_req(ctx
);
2817 kmem_cache_free(req_cachep
, req
);
2821 percpu_ref_put_many(&ctx
->refs
, nr
);
2822 mutex_unlock(&ctx
->uring_lock
);
2825 static void io_rsrc_node_cache_free(struct io_cache_entry
*entry
)
2827 kfree(container_of(entry
, struct io_rsrc_node
, cache
));
2830 static __cold
void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
2832 io_sq_thread_finish(ctx
);
2833 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2834 if (WARN_ON_ONCE(!list_empty(&ctx
->rsrc_ref_list
)))
2837 mutex_lock(&ctx
->uring_lock
);
2839 __io_sqe_buffers_unregister(ctx
);
2841 __io_sqe_files_unregister(ctx
);
2842 io_cqring_overflow_kill(ctx
);
2843 io_eventfd_unregister(ctx
);
2844 io_alloc_cache_free(&ctx
->apoll_cache
, io_apoll_cache_free
);
2845 io_alloc_cache_free(&ctx
->netmsg_cache
, io_netmsg_cache_free
);
2846 io_destroy_buffers(ctx
);
2847 mutex_unlock(&ctx
->uring_lock
);
2849 put_cred(ctx
->sq_creds
);
2850 if (ctx
->submitter_task
)
2851 put_task_struct(ctx
->submitter_task
);
2853 /* there are no registered resources left, nobody uses it */
2855 io_rsrc_node_destroy(ctx
, ctx
->rsrc_node
);
2857 WARN_ON_ONCE(!list_empty(&ctx
->rsrc_ref_list
));
2859 #if defined(CONFIG_UNIX)
2860 if (ctx
->ring_sock
) {
2861 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
2862 sock_release(ctx
->ring_sock
);
2865 WARN_ON_ONCE(!list_empty(&ctx
->ltimeout_list
));
2867 io_alloc_cache_free(&ctx
->rsrc_node_cache
, io_rsrc_node_cache_free
);
2868 if (ctx
->mm_account
) {
2869 mmdrop(ctx
->mm_account
);
2870 ctx
->mm_account
= NULL
;
2872 io_mem_free(ctx
->rings
);
2873 io_mem_free(ctx
->sq_sqes
);
2875 percpu_ref_exit(&ctx
->refs
);
2876 free_uid(ctx
->user
);
2877 io_req_caches_free(ctx
);
2879 io_wq_put_hash(ctx
->hash_map
);
2880 kfree(ctx
->cancel_table
.hbs
);
2881 kfree(ctx
->cancel_table_locked
.hbs
);
2882 kfree(ctx
->dummy_ubuf
);
2884 xa_destroy(&ctx
->io_bl_xa
);
2888 static __cold
void io_activate_pollwq_cb(struct callback_head
*cb
)
2890 struct io_ring_ctx
*ctx
= container_of(cb
, struct io_ring_ctx
,
2893 mutex_lock(&ctx
->uring_lock
);
2894 ctx
->poll_activated
= true;
2895 mutex_unlock(&ctx
->uring_lock
);
2898 * Wake ups for some events between start of polling and activation
2899 * might've been lost due to loose synchronisation.
2901 wake_up_all(&ctx
->poll_wq
);
2902 percpu_ref_put(&ctx
->refs
);
2905 static __cold
void io_activate_pollwq(struct io_ring_ctx
*ctx
)
2907 spin_lock(&ctx
->completion_lock
);
2908 /* already activated or in progress */
2909 if (ctx
->poll_activated
|| ctx
->poll_wq_task_work
.func
)
2911 if (WARN_ON_ONCE(!ctx
->task_complete
))
2913 if (!ctx
->submitter_task
)
2916 * with ->submitter_task only the submitter task completes requests, we
2917 * only need to sync with it, which is done by injecting a tw
2919 init_task_work(&ctx
->poll_wq_task_work
, io_activate_pollwq_cb
);
2920 percpu_ref_get(&ctx
->refs
);
2921 if (task_work_add(ctx
->submitter_task
, &ctx
->poll_wq_task_work
, TWA_SIGNAL
))
2922 percpu_ref_put(&ctx
->refs
);
2924 spin_unlock(&ctx
->completion_lock
);
2927 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
2929 struct io_ring_ctx
*ctx
= file
->private_data
;
2932 if (unlikely(!ctx
->poll_activated
))
2933 io_activate_pollwq(ctx
);
2935 poll_wait(file
, &ctx
->poll_wq
, wait
);
2937 * synchronizes with barrier from wq_has_sleeper call in
2941 if (!io_sqring_full(ctx
))
2942 mask
|= EPOLLOUT
| EPOLLWRNORM
;
2945 * Don't flush cqring overflow list here, just do a simple check.
2946 * Otherwise there could possible be ABBA deadlock:
2949 * lock(&ctx->uring_lock);
2951 * lock(&ctx->uring_lock);
2954 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2955 * pushes them to do the flush.
2958 if (__io_cqring_events_user(ctx
) || io_has_work(ctx
))
2959 mask
|= EPOLLIN
| EPOLLRDNORM
;
2964 static int io_unregister_personality(struct io_ring_ctx
*ctx
, unsigned id
)
2966 const struct cred
*creds
;
2968 creds
= xa_erase(&ctx
->personalities
, id
);
2977 struct io_tctx_exit
{
2978 struct callback_head task_work
;
2979 struct completion completion
;
2980 struct io_ring_ctx
*ctx
;
2983 static __cold
void io_tctx_exit_cb(struct callback_head
*cb
)
2985 struct io_uring_task
*tctx
= current
->io_uring
;
2986 struct io_tctx_exit
*work
;
2988 work
= container_of(cb
, struct io_tctx_exit
, task_work
);
2990 * When @in_cancel, we're in cancellation and it's racy to remove the
2991 * node. It'll be removed by the end of cancellation, just ignore it.
2992 * tctx can be NULL if the queueing of this task_work raced with
2993 * work cancelation off the exec path.
2995 if (tctx
&& !atomic_read(&tctx
->in_cancel
))
2996 io_uring_del_tctx_node((unsigned long)work
->ctx
);
2997 complete(&work
->completion
);
3000 static __cold
bool io_cancel_ctx_cb(struct io_wq_work
*work
, void *data
)
3002 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
3004 return req
->ctx
== data
;
3007 static __cold
void io_ring_exit_work(struct work_struct
*work
)
3009 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
, exit_work
);
3010 unsigned long timeout
= jiffies
+ HZ
* 60 * 5;
3011 unsigned long interval
= HZ
/ 20;
3012 struct io_tctx_exit exit
;
3013 struct io_tctx_node
*node
;
3017 * If we're doing polled IO and end up having requests being
3018 * submitted async (out-of-line), then completions can come in while
3019 * we're waiting for refs to drop. We need to reap these manually,
3020 * as nobody else will be looking for them.
3023 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
)) {
3024 mutex_lock(&ctx
->uring_lock
);
3025 io_cqring_overflow_kill(ctx
);
3026 mutex_unlock(&ctx
->uring_lock
);
3029 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
)
3030 io_move_task_work_from_local(ctx
);
3032 while (io_uring_try_cancel_requests(ctx
, NULL
, true))
3036 struct io_sq_data
*sqd
= ctx
->sq_data
;
3037 struct task_struct
*tsk
;
3039 io_sq_thread_park(sqd
);
3041 if (tsk
&& tsk
->io_uring
&& tsk
->io_uring
->io_wq
)
3042 io_wq_cancel_cb(tsk
->io_uring
->io_wq
,
3043 io_cancel_ctx_cb
, ctx
, true);
3044 io_sq_thread_unpark(sqd
);
3047 io_req_caches_free(ctx
);
3049 if (WARN_ON_ONCE(time_after(jiffies
, timeout
))) {
3050 /* there is little hope left, don't run it too often */
3053 } while (!wait_for_completion_timeout(&ctx
->ref_comp
, interval
));
3055 init_completion(&exit
.completion
);
3056 init_task_work(&exit
.task_work
, io_tctx_exit_cb
);
3059 * Some may use context even when all refs and requests have been put,
3060 * and they are free to do so while still holding uring_lock or
3061 * completion_lock, see io_req_task_submit(). Apart from other work,
3062 * this lock/unlock section also waits them to finish.
3064 mutex_lock(&ctx
->uring_lock
);
3065 while (!list_empty(&ctx
->tctx_list
)) {
3066 WARN_ON_ONCE(time_after(jiffies
, timeout
));
3068 node
= list_first_entry(&ctx
->tctx_list
, struct io_tctx_node
,
3070 /* don't spin on a single task if cancellation failed */
3071 list_rotate_left(&ctx
->tctx_list
);
3072 ret
= task_work_add(node
->task
, &exit
.task_work
, TWA_SIGNAL
);
3073 if (WARN_ON_ONCE(ret
))
3076 mutex_unlock(&ctx
->uring_lock
);
3077 wait_for_completion(&exit
.completion
);
3078 mutex_lock(&ctx
->uring_lock
);
3080 mutex_unlock(&ctx
->uring_lock
);
3081 spin_lock(&ctx
->completion_lock
);
3082 spin_unlock(&ctx
->completion_lock
);
3084 /* pairs with RCU read section in io_req_local_work_add() */
3085 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
)
3088 io_ring_ctx_free(ctx
);
3091 static __cold
void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
3093 unsigned long index
;
3094 struct creds
*creds
;
3096 mutex_lock(&ctx
->uring_lock
);
3097 percpu_ref_kill(&ctx
->refs
);
3098 xa_for_each(&ctx
->personalities
, index
, creds
)
3099 io_unregister_personality(ctx
, index
);
3101 io_poll_remove_all(ctx
, NULL
, true);
3102 mutex_unlock(&ctx
->uring_lock
);
3105 * If we failed setting up the ctx, we might not have any rings
3106 * and therefore did not submit any requests
3109 io_kill_timeouts(ctx
, NULL
, true);
3111 INIT_WORK(&ctx
->exit_work
, io_ring_exit_work
);
3113 * Use system_unbound_wq to avoid spawning tons of event kworkers
3114 * if we're exiting a ton of rings at the same time. It just adds
3115 * noise and overhead, there's no discernable change in runtime
3116 * over using system_wq.
3118 queue_work(system_unbound_wq
, &ctx
->exit_work
);
3121 static int io_uring_release(struct inode
*inode
, struct file
*file
)
3123 struct io_ring_ctx
*ctx
= file
->private_data
;
3125 file
->private_data
= NULL
;
3126 io_ring_ctx_wait_and_kill(ctx
);
3130 struct io_task_cancel
{
3131 struct task_struct
*task
;
3135 static bool io_cancel_task_cb(struct io_wq_work
*work
, void *data
)
3137 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
3138 struct io_task_cancel
*cancel
= data
;
3140 return io_match_task_safe(req
, cancel
->task
, cancel
->all
);
3143 static __cold
bool io_cancel_defer_files(struct io_ring_ctx
*ctx
,
3144 struct task_struct
*task
,
3147 struct io_defer_entry
*de
;
3150 spin_lock(&ctx
->completion_lock
);
3151 list_for_each_entry_reverse(de
, &ctx
->defer_list
, list
) {
3152 if (io_match_task_safe(de
->req
, task
, cancel_all
)) {
3153 list_cut_position(&list
, &ctx
->defer_list
, &de
->list
);
3157 spin_unlock(&ctx
->completion_lock
);
3158 if (list_empty(&list
))
3161 while (!list_empty(&list
)) {
3162 de
= list_first_entry(&list
, struct io_defer_entry
, list
);
3163 list_del_init(&de
->list
);
3164 io_req_task_queue_fail(de
->req
, -ECANCELED
);
3170 static __cold
bool io_uring_try_cancel_iowq(struct io_ring_ctx
*ctx
)
3172 struct io_tctx_node
*node
;
3173 enum io_wq_cancel cret
;
3176 mutex_lock(&ctx
->uring_lock
);
3177 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
3178 struct io_uring_task
*tctx
= node
->task
->io_uring
;
3181 * io_wq will stay alive while we hold uring_lock, because it's
3182 * killed after ctx nodes, which requires to take the lock.
3184 if (!tctx
|| !tctx
->io_wq
)
3186 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_ctx_cb
, ctx
, true);
3187 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
3189 mutex_unlock(&ctx
->uring_lock
);
3194 static __cold
bool io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
3195 struct task_struct
*task
,
3198 struct io_task_cancel cancel
= { .task
= task
, .all
= cancel_all
, };
3199 struct io_uring_task
*tctx
= task
? task
->io_uring
: NULL
;
3200 enum io_wq_cancel cret
;
3203 /* set it so io_req_local_work_add() would wake us up */
3204 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) {
3205 atomic_set(&ctx
->cq_wait_nr
, 1);
3209 /* failed during ring init, it couldn't have issued any requests */
3214 ret
|= io_uring_try_cancel_iowq(ctx
);
3215 } else if (tctx
&& tctx
->io_wq
) {
3217 * Cancels requests of all rings, not only @ctx, but
3218 * it's fine as the task is in exit/exec.
3220 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_task_cb
,
3222 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
3225 /* SQPOLL thread does its own polling */
3226 if ((!(ctx
->flags
& IORING_SETUP_SQPOLL
) && cancel_all
) ||
3227 (ctx
->sq_data
&& ctx
->sq_data
->thread
== current
)) {
3228 while (!wq_list_empty(&ctx
->iopoll_list
)) {
3229 io_iopoll_try_reap_events(ctx
);
3235 if ((ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) &&
3236 io_allowed_defer_tw_run(ctx
))
3237 ret
|= io_run_local_work(ctx
) > 0;
3238 ret
|= io_cancel_defer_files(ctx
, task
, cancel_all
);
3239 mutex_lock(&ctx
->uring_lock
);
3240 ret
|= io_poll_remove_all(ctx
, task
, cancel_all
);
3241 mutex_unlock(&ctx
->uring_lock
);
3242 ret
|= io_kill_timeouts(ctx
, task
, cancel_all
);
3244 ret
|= io_run_task_work() > 0;
3248 static s64
tctx_inflight(struct io_uring_task
*tctx
, bool tracked
)
3251 return atomic_read(&tctx
->inflight_tracked
);
3252 return percpu_counter_sum(&tctx
->inflight
);
3256 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3257 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3259 __cold
void io_uring_cancel_generic(bool cancel_all
, struct io_sq_data
*sqd
)
3261 struct io_uring_task
*tctx
= current
->io_uring
;
3262 struct io_ring_ctx
*ctx
;
3263 struct io_tctx_node
*node
;
3264 unsigned long index
;
3268 WARN_ON_ONCE(sqd
&& sqd
->thread
!= current
);
3270 if (!current
->io_uring
)
3273 io_wq_exit_start(tctx
->io_wq
);
3275 atomic_inc(&tctx
->in_cancel
);
3279 io_uring_drop_tctx_refs(current
);
3280 /* read completions before cancelations */
3281 inflight
= tctx_inflight(tctx
, !cancel_all
);
3286 xa_for_each(&tctx
->xa
, index
, node
) {
3287 /* sqpoll task will cancel all its requests */
3288 if (node
->ctx
->sq_data
)
3290 loop
|= io_uring_try_cancel_requests(node
->ctx
,
3291 current
, cancel_all
);
3294 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
3295 loop
|= io_uring_try_cancel_requests(ctx
,
3305 prepare_to_wait(&tctx
->wait
, &wait
, TASK_INTERRUPTIBLE
);
3307 io_uring_drop_tctx_refs(current
);
3308 xa_for_each(&tctx
->xa
, index
, node
) {
3309 if (!llist_empty(&node
->ctx
->work_llist
)) {
3310 WARN_ON_ONCE(node
->ctx
->submitter_task
&&
3311 node
->ctx
->submitter_task
!= current
);
3316 * If we've seen completions, retry without waiting. This
3317 * avoids a race where a completion comes in before we did
3318 * prepare_to_wait().
3320 if (inflight
== tctx_inflight(tctx
, !cancel_all
))
3323 finish_wait(&tctx
->wait
, &wait
);
3326 io_uring_clean_tctx(tctx
);
3329 * We shouldn't run task_works after cancel, so just leave
3330 * ->in_cancel set for normal exit.
3332 atomic_dec(&tctx
->in_cancel
);
3333 /* for exec all current's requests should be gone, kill tctx */
3334 __io_uring_free(current
);
3338 void __io_uring_cancel(bool cancel_all
)
3340 io_uring_cancel_generic(cancel_all
, NULL
);
3343 static void *io_uring_validate_mmap_request(struct file
*file
,
3344 loff_t pgoff
, size_t sz
)
3346 struct io_ring_ctx
*ctx
= file
->private_data
;
3347 loff_t offset
= pgoff
<< PAGE_SHIFT
;
3351 switch (offset
& IORING_OFF_MMAP_MASK
) {
3352 case IORING_OFF_SQ_RING
:
3353 case IORING_OFF_CQ_RING
:
3356 case IORING_OFF_SQES
:
3359 case IORING_OFF_PBUF_RING
: {
3362 bgid
= (offset
& ~IORING_OFF_MMAP_MASK
) >> IORING_OFF_PBUF_SHIFT
;
3363 mutex_lock(&ctx
->uring_lock
);
3364 ptr
= io_pbuf_get_address(ctx
, bgid
);
3365 mutex_unlock(&ctx
->uring_lock
);
3367 return ERR_PTR(-EINVAL
);
3371 return ERR_PTR(-EINVAL
);
3374 page
= virt_to_head_page(ptr
);
3375 if (sz
> page_size(page
))
3376 return ERR_PTR(-EINVAL
);
3383 static __cold
int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3385 size_t sz
= vma
->vm_end
- vma
->vm_start
;
3389 ptr
= io_uring_validate_mmap_request(file
, vma
->vm_pgoff
, sz
);
3391 return PTR_ERR(ptr
);
3393 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
3394 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
3397 static unsigned long io_uring_mmu_get_unmapped_area(struct file
*filp
,
3398 unsigned long addr
, unsigned long len
,
3399 unsigned long pgoff
, unsigned long flags
)
3401 const unsigned long mmap_end
= arch_get_mmap_end(addr
, len
, flags
);
3402 struct vm_unmapped_area_info info
;
3406 * Do not allow to map to user-provided address to avoid breaking the
3407 * aliasing rules. Userspace is not able to guess the offset address of
3408 * kernel kmalloc()ed memory area.
3413 ptr
= io_uring_validate_mmap_request(filp
, pgoff
, len
);
3417 info
.flags
= VM_UNMAPPED_AREA_TOPDOWN
;
3419 info
.low_limit
= max(PAGE_SIZE
, mmap_min_addr
);
3420 info
.high_limit
= arch_get_mmap_base(addr
, current
->mm
->mmap_base
);
3422 info
.align_mask
= PAGE_MASK
& (SHM_COLOUR
- 1UL);
3424 info
.align_mask
= PAGE_MASK
& (SHMLBA
- 1UL);
3426 info
.align_offset
= (unsigned long) ptr
;
3429 * A failed mmap() very likely causes application failure,
3430 * so fall back to the bottom-up function here. This scenario
3431 * can happen with large stack limits and large mmap()
3434 addr
= vm_unmapped_area(&info
);
3435 if (offset_in_page(addr
)) {
3437 info
.low_limit
= TASK_UNMAPPED_BASE
;
3438 info
.high_limit
= mmap_end
;
3439 addr
= vm_unmapped_area(&info
);
3445 #else /* !CONFIG_MMU */
3447 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3449 return is_nommu_shared_mapping(vma
->vm_flags
) ? 0 : -EINVAL
;
3452 static unsigned int io_uring_nommu_mmap_capabilities(struct file
*file
)
3454 return NOMMU_MAP_DIRECT
| NOMMU_MAP_READ
| NOMMU_MAP_WRITE
;
3457 static unsigned long io_uring_nommu_get_unmapped_area(struct file
*file
,
3458 unsigned long addr
, unsigned long len
,
3459 unsigned long pgoff
, unsigned long flags
)
3463 ptr
= io_uring_validate_mmap_request(file
, pgoff
, len
);
3465 return PTR_ERR(ptr
);
3467 return (unsigned long) ptr
;
3470 #endif /* !CONFIG_MMU */
3472 static int io_validate_ext_arg(unsigned flags
, const void __user
*argp
, size_t argsz
)
3474 if (flags
& IORING_ENTER_EXT_ARG
) {
3475 struct io_uring_getevents_arg arg
;
3477 if (argsz
!= sizeof(arg
))
3479 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
3485 static int io_get_ext_arg(unsigned flags
, const void __user
*argp
, size_t *argsz
,
3486 struct __kernel_timespec __user
**ts
,
3487 const sigset_t __user
**sig
)
3489 struct io_uring_getevents_arg arg
;
3492 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3493 * is just a pointer to the sigset_t.
3495 if (!(flags
& IORING_ENTER_EXT_ARG
)) {
3496 *sig
= (const sigset_t __user
*) argp
;
3502 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3503 * timespec and sigset_t pointers if good.
3505 if (*argsz
!= sizeof(arg
))
3507 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
3511 *sig
= u64_to_user_ptr(arg
.sigmask
);
3512 *argsz
= arg
.sigmask_sz
;
3513 *ts
= u64_to_user_ptr(arg
.ts
);
3517 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
3518 u32
, min_complete
, u32
, flags
, const void __user
*, argp
,
3521 struct io_ring_ctx
*ctx
;
3525 if (unlikely(flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
|
3526 IORING_ENTER_SQ_WAIT
| IORING_ENTER_EXT_ARG
|
3527 IORING_ENTER_REGISTERED_RING
)))
3531 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3532 * need only dereference our task private array to find it.
3534 if (flags
& IORING_ENTER_REGISTERED_RING
) {
3535 struct io_uring_task
*tctx
= current
->io_uring
;
3537 if (unlikely(!tctx
|| fd
>= IO_RINGFD_REG_MAX
))
3539 fd
= array_index_nospec(fd
, IO_RINGFD_REG_MAX
);
3540 f
.file
= tctx
->registered_rings
[fd
];
3542 if (unlikely(!f
.file
))
3546 if (unlikely(!f
.file
))
3549 if (unlikely(!io_is_uring_fops(f
.file
)))
3553 ctx
= f
.file
->private_data
;
3555 if (unlikely(ctx
->flags
& IORING_SETUP_R_DISABLED
))
3559 * For SQ polling, the thread will do all submissions and completions.
3560 * Just return the requested submit count, and wake the thread if
3564 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
3565 io_cqring_overflow_flush(ctx
);
3567 if (unlikely(ctx
->sq_data
->thread
== NULL
)) {
3571 if (flags
& IORING_ENTER_SQ_WAKEUP
)
3572 wake_up(&ctx
->sq_data
->wait
);
3573 if (flags
& IORING_ENTER_SQ_WAIT
)
3574 io_sqpoll_wait_sq(ctx
);
3577 } else if (to_submit
) {
3578 ret
= io_uring_add_tctx_node(ctx
);
3582 mutex_lock(&ctx
->uring_lock
);
3583 ret
= io_submit_sqes(ctx
, to_submit
);
3584 if (ret
!= to_submit
) {
3585 mutex_unlock(&ctx
->uring_lock
);
3588 if (flags
& IORING_ENTER_GETEVENTS
) {
3589 if (ctx
->syscall_iopoll
)
3592 * Ignore errors, we'll soon call io_cqring_wait() and
3593 * it should handle ownership problems if any.
3595 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
)
3596 (void)io_run_local_work_locked(ctx
);
3598 mutex_unlock(&ctx
->uring_lock
);
3601 if (flags
& IORING_ENTER_GETEVENTS
) {
3604 if (ctx
->syscall_iopoll
) {
3606 * We disallow the app entering submit/complete with
3607 * polling, but we still need to lock the ring to
3608 * prevent racing with polled issue that got punted to
3611 mutex_lock(&ctx
->uring_lock
);
3613 ret2
= io_validate_ext_arg(flags
, argp
, argsz
);
3614 if (likely(!ret2
)) {
3615 min_complete
= min(min_complete
,
3617 ret2
= io_iopoll_check(ctx
, min_complete
);
3619 mutex_unlock(&ctx
->uring_lock
);
3621 const sigset_t __user
*sig
;
3622 struct __kernel_timespec __user
*ts
;
3624 ret2
= io_get_ext_arg(flags
, argp
, &argsz
, &ts
, &sig
);
3625 if (likely(!ret2
)) {
3626 min_complete
= min(min_complete
,
3628 ret2
= io_cqring_wait(ctx
, min_complete
, sig
,
3637 * EBADR indicates that one or more CQE were dropped.
3638 * Once the user has been informed we can clear the bit
3639 * as they are obviously ok with those drops.
3641 if (unlikely(ret2
== -EBADR
))
3642 clear_bit(IO_CHECK_CQ_DROPPED_BIT
,
3651 static const struct file_operations io_uring_fops
= {
3652 .release
= io_uring_release
,
3653 .mmap
= io_uring_mmap
,
3655 .get_unmapped_area
= io_uring_nommu_get_unmapped_area
,
3656 .mmap_capabilities
= io_uring_nommu_mmap_capabilities
,
3658 .get_unmapped_area
= io_uring_mmu_get_unmapped_area
,
3660 .poll
= io_uring_poll
,
3661 #ifdef CONFIG_PROC_FS
3662 .show_fdinfo
= io_uring_show_fdinfo
,
3666 bool io_is_uring_fops(struct file
*file
)
3668 return file
->f_op
== &io_uring_fops
;
3671 static __cold
int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
3672 struct io_uring_params
*p
)
3674 struct io_rings
*rings
;
3675 size_t size
, sq_array_offset
;
3677 /* make sure these are sane, as we already accounted them */
3678 ctx
->sq_entries
= p
->sq_entries
;
3679 ctx
->cq_entries
= p
->cq_entries
;
3681 size
= rings_size(ctx
, p
->sq_entries
, p
->cq_entries
, &sq_array_offset
);
3682 if (size
== SIZE_MAX
)
3685 rings
= io_mem_alloc(size
);
3690 ctx
->sq_array
= (u32
*)((char *)rings
+ sq_array_offset
);
3691 rings
->sq_ring_mask
= p
->sq_entries
- 1;
3692 rings
->cq_ring_mask
= p
->cq_entries
- 1;
3693 rings
->sq_ring_entries
= p
->sq_entries
;
3694 rings
->cq_ring_entries
= p
->cq_entries
;
3696 if (p
->flags
& IORING_SETUP_SQE128
)
3697 size
= array_size(2 * sizeof(struct io_uring_sqe
), p
->sq_entries
);
3699 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
3700 if (size
== SIZE_MAX
) {
3701 io_mem_free(ctx
->rings
);
3706 ctx
->sq_sqes
= io_mem_alloc(size
);
3707 if (!ctx
->sq_sqes
) {
3708 io_mem_free(ctx
->rings
);
3716 static int io_uring_install_fd(struct io_ring_ctx
*ctx
, struct file
*file
)
3720 fd
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
3724 ret
= __io_uring_add_tctx_node(ctx
);
3729 fd_install(fd
, file
);
3734 * Allocate an anonymous fd, this is what constitutes the application
3735 * visible backing of an io_uring instance. The application mmaps this
3736 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3737 * we have to tie this fd to a socket for file garbage collection purposes.
3739 static struct file
*io_uring_get_file(struct io_ring_ctx
*ctx
)
3742 #if defined(CONFIG_UNIX)
3745 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
3748 return ERR_PTR(ret
);
3751 file
= anon_inode_getfile_secure("[io_uring]", &io_uring_fops
, ctx
,
3752 O_RDWR
| O_CLOEXEC
, NULL
);
3753 #if defined(CONFIG_UNIX)
3755 sock_release(ctx
->ring_sock
);
3756 ctx
->ring_sock
= NULL
;
3758 ctx
->ring_sock
->file
= file
;
3764 static __cold
int io_uring_create(unsigned entries
, struct io_uring_params
*p
,
3765 struct io_uring_params __user
*params
)
3767 struct io_ring_ctx
*ctx
;
3773 if (entries
> IORING_MAX_ENTRIES
) {
3774 if (!(p
->flags
& IORING_SETUP_CLAMP
))
3776 entries
= IORING_MAX_ENTRIES
;
3780 * Use twice as many entries for the CQ ring. It's possible for the
3781 * application to drive a higher depth than the size of the SQ ring,
3782 * since the sqes are only used at submission time. This allows for
3783 * some flexibility in overcommitting a bit. If the application has
3784 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3785 * of CQ ring entries manually.
3787 p
->sq_entries
= roundup_pow_of_two(entries
);
3788 if (p
->flags
& IORING_SETUP_CQSIZE
) {
3790 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3791 * to a power-of-two, if it isn't already. We do NOT impose
3792 * any cq vs sq ring sizing.
3796 if (p
->cq_entries
> IORING_MAX_CQ_ENTRIES
) {
3797 if (!(p
->flags
& IORING_SETUP_CLAMP
))
3799 p
->cq_entries
= IORING_MAX_CQ_ENTRIES
;
3801 p
->cq_entries
= roundup_pow_of_two(p
->cq_entries
);
3802 if (p
->cq_entries
< p
->sq_entries
)
3805 p
->cq_entries
= 2 * p
->sq_entries
;
3808 ctx
= io_ring_ctx_alloc(p
);
3812 if ((ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) &&
3813 !(ctx
->flags
& IORING_SETUP_IOPOLL
) &&
3814 !(ctx
->flags
& IORING_SETUP_SQPOLL
))
3815 ctx
->task_complete
= true;
3818 * lazy poll_wq activation relies on ->task_complete for synchronisation
3819 * purposes, see io_activate_pollwq()
3821 if (!ctx
->task_complete
)
3822 ctx
->poll_activated
= true;
3825 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3826 * space applications don't need to do io completion events
3827 * polling again, they can rely on io_sq_thread to do polling
3828 * work, which can reduce cpu usage and uring_lock contention.
3830 if (ctx
->flags
& IORING_SETUP_IOPOLL
&&
3831 !(ctx
->flags
& IORING_SETUP_SQPOLL
))
3832 ctx
->syscall_iopoll
= 1;
3834 ctx
->compat
= in_compat_syscall();
3835 if (!capable(CAP_IPC_LOCK
))
3836 ctx
->user
= get_uid(current_user());
3839 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3840 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3843 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
3844 /* IPI related flags don't make sense with SQPOLL */
3845 if (ctx
->flags
& (IORING_SETUP_COOP_TASKRUN
|
3846 IORING_SETUP_TASKRUN_FLAG
|
3847 IORING_SETUP_DEFER_TASKRUN
))
3849 ctx
->notify_method
= TWA_SIGNAL_NO_IPI
;
3850 } else if (ctx
->flags
& IORING_SETUP_COOP_TASKRUN
) {
3851 ctx
->notify_method
= TWA_SIGNAL_NO_IPI
;
3853 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
&&
3854 !(ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
))
3856 ctx
->notify_method
= TWA_SIGNAL
;
3860 * For DEFER_TASKRUN we require the completion task to be the same as the
3861 * submission task. This implies that there is only one submitter, so enforce
3864 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
&&
3865 !(ctx
->flags
& IORING_SETUP_SINGLE_ISSUER
)) {
3870 * This is just grabbed for accounting purposes. When a process exits,
3871 * the mm is exited and dropped before the files, hence we need to hang
3872 * on to this mm purely for the purposes of being able to unaccount
3873 * memory (locked/pinned vm). It's not used for anything else.
3875 mmgrab(current
->mm
);
3876 ctx
->mm_account
= current
->mm
;
3878 ret
= io_allocate_scq_urings(ctx
, p
);
3882 ret
= io_sq_offload_create(ctx
, p
);
3886 ret
= io_rsrc_init(ctx
);
3890 memset(&p
->sq_off
, 0, sizeof(p
->sq_off
));
3891 p
->sq_off
.head
= offsetof(struct io_rings
, sq
.head
);
3892 p
->sq_off
.tail
= offsetof(struct io_rings
, sq
.tail
);
3893 p
->sq_off
.ring_mask
= offsetof(struct io_rings
, sq_ring_mask
);
3894 p
->sq_off
.ring_entries
= offsetof(struct io_rings
, sq_ring_entries
);
3895 p
->sq_off
.flags
= offsetof(struct io_rings
, sq_flags
);
3896 p
->sq_off
.dropped
= offsetof(struct io_rings
, sq_dropped
);
3897 p
->sq_off
.array
= (char *)ctx
->sq_array
- (char *)ctx
->rings
;
3899 memset(&p
->cq_off
, 0, sizeof(p
->cq_off
));
3900 p
->cq_off
.head
= offsetof(struct io_rings
, cq
.head
);
3901 p
->cq_off
.tail
= offsetof(struct io_rings
, cq
.tail
);
3902 p
->cq_off
.ring_mask
= offsetof(struct io_rings
, cq_ring_mask
);
3903 p
->cq_off
.ring_entries
= offsetof(struct io_rings
, cq_ring_entries
);
3904 p
->cq_off
.overflow
= offsetof(struct io_rings
, cq_overflow
);
3905 p
->cq_off
.cqes
= offsetof(struct io_rings
, cqes
);
3906 p
->cq_off
.flags
= offsetof(struct io_rings
, cq_flags
);
3908 p
->features
= IORING_FEAT_SINGLE_MMAP
| IORING_FEAT_NODROP
|
3909 IORING_FEAT_SUBMIT_STABLE
| IORING_FEAT_RW_CUR_POS
|
3910 IORING_FEAT_CUR_PERSONALITY
| IORING_FEAT_FAST_POLL
|
3911 IORING_FEAT_POLL_32BITS
| IORING_FEAT_SQPOLL_NONFIXED
|
3912 IORING_FEAT_EXT_ARG
| IORING_FEAT_NATIVE_WORKERS
|
3913 IORING_FEAT_RSRC_TAGS
| IORING_FEAT_CQE_SKIP
|
3914 IORING_FEAT_LINKED_FILE
| IORING_FEAT_REG_REG_RING
;
3916 if (copy_to_user(params
, p
, sizeof(*p
))) {
3921 if (ctx
->flags
& IORING_SETUP_SINGLE_ISSUER
3922 && !(ctx
->flags
& IORING_SETUP_R_DISABLED
))
3923 WRITE_ONCE(ctx
->submitter_task
, get_task_struct(current
));
3925 file
= io_uring_get_file(ctx
);
3927 ret
= PTR_ERR(file
);
3932 * Install ring fd as the very last thing, so we don't risk someone
3933 * having closed it before we finish setup
3935 ret
= io_uring_install_fd(ctx
, file
);
3937 /* fput will clean it up */
3942 trace_io_uring_create(ret
, ctx
, p
->sq_entries
, p
->cq_entries
, p
->flags
);
3945 io_ring_ctx_wait_and_kill(ctx
);
3950 * Sets up an aio uring context, and returns the fd. Applications asks for a
3951 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3952 * params structure passed in.
3954 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
3956 struct io_uring_params p
;
3959 if (copy_from_user(&p
, params
, sizeof(p
)))
3961 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
3966 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
3967 IORING_SETUP_SQ_AFF
| IORING_SETUP_CQSIZE
|
3968 IORING_SETUP_CLAMP
| IORING_SETUP_ATTACH_WQ
|
3969 IORING_SETUP_R_DISABLED
| IORING_SETUP_SUBMIT_ALL
|
3970 IORING_SETUP_COOP_TASKRUN
| IORING_SETUP_TASKRUN_FLAG
|
3971 IORING_SETUP_SQE128
| IORING_SETUP_CQE32
|
3972 IORING_SETUP_SINGLE_ISSUER
| IORING_SETUP_DEFER_TASKRUN
))
3975 return io_uring_create(entries
, &p
, params
);
3978 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
3979 struct io_uring_params __user
*, params
)
3981 return io_uring_setup(entries
, params
);
3984 static __cold
int io_probe(struct io_ring_ctx
*ctx
, void __user
*arg
,
3987 struct io_uring_probe
*p
;
3991 size
= struct_size(p
, ops
, nr_args
);
3992 if (size
== SIZE_MAX
)
3994 p
= kzalloc(size
, GFP_KERNEL
);
3999 if (copy_from_user(p
, arg
, size
))
4002 if (memchr_inv(p
, 0, size
))
4005 p
->last_op
= IORING_OP_LAST
- 1;
4006 if (nr_args
> IORING_OP_LAST
)
4007 nr_args
= IORING_OP_LAST
;
4009 for (i
= 0; i
< nr_args
; i
++) {
4011 if (!io_issue_defs
[i
].not_supported
)
4012 p
->ops
[i
].flags
= IO_URING_OP_SUPPORTED
;
4017 if (copy_to_user(arg
, p
, size
))
4024 static int io_register_personality(struct io_ring_ctx
*ctx
)
4026 const struct cred
*creds
;
4030 creds
= get_current_cred();
4032 ret
= xa_alloc_cyclic(&ctx
->personalities
, &id
, (void *)creds
,
4033 XA_LIMIT(0, USHRT_MAX
), &ctx
->pers_next
, GFP_KERNEL
);
4041 static __cold
int io_register_restrictions(struct io_ring_ctx
*ctx
,
4042 void __user
*arg
, unsigned int nr_args
)
4044 struct io_uring_restriction
*res
;
4048 /* Restrictions allowed only if rings started disabled */
4049 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
4052 /* We allow only a single restrictions registration */
4053 if (ctx
->restrictions
.registered
)
4056 if (!arg
|| nr_args
> IORING_MAX_RESTRICTIONS
)
4059 size
= array_size(nr_args
, sizeof(*res
));
4060 if (size
== SIZE_MAX
)
4063 res
= memdup_user(arg
, size
);
4065 return PTR_ERR(res
);
4069 for (i
= 0; i
< nr_args
; i
++) {
4070 switch (res
[i
].opcode
) {
4071 case IORING_RESTRICTION_REGISTER_OP
:
4072 if (res
[i
].register_op
>= IORING_REGISTER_LAST
) {
4077 __set_bit(res
[i
].register_op
,
4078 ctx
->restrictions
.register_op
);
4080 case IORING_RESTRICTION_SQE_OP
:
4081 if (res
[i
].sqe_op
>= IORING_OP_LAST
) {
4086 __set_bit(res
[i
].sqe_op
, ctx
->restrictions
.sqe_op
);
4088 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED
:
4089 ctx
->restrictions
.sqe_flags_allowed
= res
[i
].sqe_flags
;
4091 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED
:
4092 ctx
->restrictions
.sqe_flags_required
= res
[i
].sqe_flags
;
4101 /* Reset all restrictions if an error happened */
4103 memset(&ctx
->restrictions
, 0, sizeof(ctx
->restrictions
));
4105 ctx
->restrictions
.registered
= true;
4111 static int io_register_enable_rings(struct io_ring_ctx
*ctx
)
4113 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
4116 if (ctx
->flags
& IORING_SETUP_SINGLE_ISSUER
&& !ctx
->submitter_task
) {
4117 WRITE_ONCE(ctx
->submitter_task
, get_task_struct(current
));
4119 * Lazy activation attempts would fail if it was polled before
4120 * submitter_task is set.
4122 if (wq_has_sleeper(&ctx
->poll_wq
))
4123 io_activate_pollwq(ctx
);
4126 if (ctx
->restrictions
.registered
)
4127 ctx
->restricted
= 1;
4129 ctx
->flags
&= ~IORING_SETUP_R_DISABLED
;
4130 if (ctx
->sq_data
&& wq_has_sleeper(&ctx
->sq_data
->wait
))
4131 wake_up(&ctx
->sq_data
->wait
);
4135 static __cold
int io_register_iowq_aff(struct io_ring_ctx
*ctx
,
4136 void __user
*arg
, unsigned len
)
4138 struct io_uring_task
*tctx
= current
->io_uring
;
4139 cpumask_var_t new_mask
;
4142 if (!tctx
|| !tctx
->io_wq
)
4145 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4148 cpumask_clear(new_mask
);
4149 if (len
> cpumask_size())
4150 len
= cpumask_size();
4152 if (in_compat_syscall()) {
4153 ret
= compat_get_bitmap(cpumask_bits(new_mask
),
4154 (const compat_ulong_t __user
*)arg
,
4155 len
* 8 /* CHAR_BIT */);
4157 ret
= copy_from_user(new_mask
, arg
, len
);
4161 free_cpumask_var(new_mask
);
4165 ret
= io_wq_cpu_affinity(tctx
->io_wq
, new_mask
);
4166 free_cpumask_var(new_mask
);
4170 static __cold
int io_unregister_iowq_aff(struct io_ring_ctx
*ctx
)
4172 struct io_uring_task
*tctx
= current
->io_uring
;
4174 if (!tctx
|| !tctx
->io_wq
)
4177 return io_wq_cpu_affinity(tctx
->io_wq
, NULL
);
4180 static __cold
int io_register_iowq_max_workers(struct io_ring_ctx
*ctx
,
4182 __must_hold(&ctx
->uring_lock
)
4184 struct io_tctx_node
*node
;
4185 struct io_uring_task
*tctx
= NULL
;
4186 struct io_sq_data
*sqd
= NULL
;
4190 if (copy_from_user(new_count
, arg
, sizeof(new_count
)))
4192 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
4193 if (new_count
[i
] > INT_MAX
)
4196 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
4200 * Observe the correct sqd->lock -> ctx->uring_lock
4201 * ordering. Fine to drop uring_lock here, we hold
4204 refcount_inc(&sqd
->refs
);
4205 mutex_unlock(&ctx
->uring_lock
);
4206 mutex_lock(&sqd
->lock
);
4207 mutex_lock(&ctx
->uring_lock
);
4209 tctx
= sqd
->thread
->io_uring
;
4212 tctx
= current
->io_uring
;
4215 BUILD_BUG_ON(sizeof(new_count
) != sizeof(ctx
->iowq_limits
));
4217 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
4219 ctx
->iowq_limits
[i
] = new_count
[i
];
4220 ctx
->iowq_limits_set
= true;
4222 if (tctx
&& tctx
->io_wq
) {
4223 ret
= io_wq_max_workers(tctx
->io_wq
, new_count
);
4227 memset(new_count
, 0, sizeof(new_count
));
4231 mutex_unlock(&sqd
->lock
);
4232 io_put_sq_data(sqd
);
4235 if (copy_to_user(arg
, new_count
, sizeof(new_count
)))
4238 /* that's it for SQPOLL, only the SQPOLL task creates requests */
4242 /* now propagate the restriction to all registered users */
4243 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
4244 struct io_uring_task
*tctx
= node
->task
->io_uring
;
4246 if (WARN_ON_ONCE(!tctx
->io_wq
))
4249 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
4250 new_count
[i
] = ctx
->iowq_limits
[i
];
4251 /* ignore errors, it always returns zero anyway */
4252 (void)io_wq_max_workers(tctx
->io_wq
, new_count
);
4257 mutex_unlock(&sqd
->lock
);
4258 io_put_sq_data(sqd
);
4263 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
4264 void __user
*arg
, unsigned nr_args
)
4265 __releases(ctx
->uring_lock
)
4266 __acquires(ctx
->uring_lock
)
4271 * We don't quiesce the refs for register anymore and so it can't be
4272 * dying as we're holding a file ref here.
4274 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx
->refs
)))
4277 if (ctx
->submitter_task
&& ctx
->submitter_task
!= current
)
4280 if (ctx
->restricted
) {
4281 opcode
= array_index_nospec(opcode
, IORING_REGISTER_LAST
);
4282 if (!test_bit(opcode
, ctx
->restrictions
.register_op
))
4287 case IORING_REGISTER_BUFFERS
:
4291 ret
= io_sqe_buffers_register(ctx
, arg
, nr_args
, NULL
);
4293 case IORING_UNREGISTER_BUFFERS
:
4297 ret
= io_sqe_buffers_unregister(ctx
);
4299 case IORING_REGISTER_FILES
:
4303 ret
= io_sqe_files_register(ctx
, arg
, nr_args
, NULL
);
4305 case IORING_UNREGISTER_FILES
:
4309 ret
= io_sqe_files_unregister(ctx
);
4311 case IORING_REGISTER_FILES_UPDATE
:
4312 ret
= io_register_files_update(ctx
, arg
, nr_args
);
4314 case IORING_REGISTER_EVENTFD
:
4318 ret
= io_eventfd_register(ctx
, arg
, 0);
4320 case IORING_REGISTER_EVENTFD_ASYNC
:
4324 ret
= io_eventfd_register(ctx
, arg
, 1);
4326 case IORING_UNREGISTER_EVENTFD
:
4330 ret
= io_eventfd_unregister(ctx
);
4332 case IORING_REGISTER_PROBE
:
4334 if (!arg
|| nr_args
> 256)
4336 ret
= io_probe(ctx
, arg
, nr_args
);
4338 case IORING_REGISTER_PERSONALITY
:
4342 ret
= io_register_personality(ctx
);
4344 case IORING_UNREGISTER_PERSONALITY
:
4348 ret
= io_unregister_personality(ctx
, nr_args
);
4350 case IORING_REGISTER_ENABLE_RINGS
:
4354 ret
= io_register_enable_rings(ctx
);
4356 case IORING_REGISTER_RESTRICTIONS
:
4357 ret
= io_register_restrictions(ctx
, arg
, nr_args
);
4359 case IORING_REGISTER_FILES2
:
4360 ret
= io_register_rsrc(ctx
, arg
, nr_args
, IORING_RSRC_FILE
);
4362 case IORING_REGISTER_FILES_UPDATE2
:
4363 ret
= io_register_rsrc_update(ctx
, arg
, nr_args
,
4366 case IORING_REGISTER_BUFFERS2
:
4367 ret
= io_register_rsrc(ctx
, arg
, nr_args
, IORING_RSRC_BUFFER
);
4369 case IORING_REGISTER_BUFFERS_UPDATE
:
4370 ret
= io_register_rsrc_update(ctx
, arg
, nr_args
,
4371 IORING_RSRC_BUFFER
);
4373 case IORING_REGISTER_IOWQ_AFF
:
4375 if (!arg
|| !nr_args
)
4377 ret
= io_register_iowq_aff(ctx
, arg
, nr_args
);
4379 case IORING_UNREGISTER_IOWQ_AFF
:
4383 ret
= io_unregister_iowq_aff(ctx
);
4385 case IORING_REGISTER_IOWQ_MAX_WORKERS
:
4387 if (!arg
|| nr_args
!= 2)
4389 ret
= io_register_iowq_max_workers(ctx
, arg
);
4391 case IORING_REGISTER_RING_FDS
:
4392 ret
= io_ringfd_register(ctx
, arg
, nr_args
);
4394 case IORING_UNREGISTER_RING_FDS
:
4395 ret
= io_ringfd_unregister(ctx
, arg
, nr_args
);
4397 case IORING_REGISTER_PBUF_RING
:
4399 if (!arg
|| nr_args
!= 1)
4401 ret
= io_register_pbuf_ring(ctx
, arg
);
4403 case IORING_UNREGISTER_PBUF_RING
:
4405 if (!arg
|| nr_args
!= 1)
4407 ret
= io_unregister_pbuf_ring(ctx
, arg
);
4409 case IORING_REGISTER_SYNC_CANCEL
:
4411 if (!arg
|| nr_args
!= 1)
4413 ret
= io_sync_cancel(ctx
, arg
);
4415 case IORING_REGISTER_FILE_ALLOC_RANGE
:
4417 if (!arg
|| nr_args
)
4419 ret
= io_register_file_alloc_range(ctx
, arg
);
4429 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
4430 void __user
*, arg
, unsigned int, nr_args
)
4432 struct io_ring_ctx
*ctx
;
4435 bool use_registered_ring
;
4437 use_registered_ring
= !!(opcode
& IORING_REGISTER_USE_REGISTERED_RING
);
4438 opcode
&= ~IORING_REGISTER_USE_REGISTERED_RING
;
4440 if (opcode
>= IORING_REGISTER_LAST
)
4443 if (use_registered_ring
) {
4445 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
4446 * need only dereference our task private array to find it.
4448 struct io_uring_task
*tctx
= current
->io_uring
;
4450 if (unlikely(!tctx
|| fd
>= IO_RINGFD_REG_MAX
))
4452 fd
= array_index_nospec(fd
, IO_RINGFD_REG_MAX
);
4453 f
.file
= tctx
->registered_rings
[fd
];
4455 if (unlikely(!f
.file
))
4459 if (unlikely(!f
.file
))
4462 if (!io_is_uring_fops(f
.file
))
4466 ctx
= f
.file
->private_data
;
4468 mutex_lock(&ctx
->uring_lock
);
4469 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
4470 mutex_unlock(&ctx
->uring_lock
);
4471 trace_io_uring_register(ctx
, opcode
, ctx
->nr_user_files
, ctx
->nr_user_bufs
, ret
);
4477 static int __init
io_uring_init(void)
4479 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4480 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4481 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4484 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4485 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4486 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4487 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4488 BUILD_BUG_ON(sizeof(struct io_uring_sqe
) != 64);
4489 BUILD_BUG_SQE_ELEM(0, __u8
, opcode
);
4490 BUILD_BUG_SQE_ELEM(1, __u8
, flags
);
4491 BUILD_BUG_SQE_ELEM(2, __u16
, ioprio
);
4492 BUILD_BUG_SQE_ELEM(4, __s32
, fd
);
4493 BUILD_BUG_SQE_ELEM(8, __u64
, off
);
4494 BUILD_BUG_SQE_ELEM(8, __u64
, addr2
);
4495 BUILD_BUG_SQE_ELEM(8, __u32
, cmd_op
);
4496 BUILD_BUG_SQE_ELEM(12, __u32
, __pad1
);
4497 BUILD_BUG_SQE_ELEM(16, __u64
, addr
);
4498 BUILD_BUG_SQE_ELEM(16, __u64
, splice_off_in
);
4499 BUILD_BUG_SQE_ELEM(24, __u32
, len
);
4500 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t
, rw_flags
);
4501 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags
);
4502 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32
, rw_flags
);
4503 BUILD_BUG_SQE_ELEM(28, __u32
, fsync_flags
);
4504 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16
, poll_events
);
4505 BUILD_BUG_SQE_ELEM(28, __u32
, poll32_events
);
4506 BUILD_BUG_SQE_ELEM(28, __u32
, sync_range_flags
);
4507 BUILD_BUG_SQE_ELEM(28, __u32
, msg_flags
);
4508 BUILD_BUG_SQE_ELEM(28, __u32
, timeout_flags
);
4509 BUILD_BUG_SQE_ELEM(28, __u32
, accept_flags
);
4510 BUILD_BUG_SQE_ELEM(28, __u32
, cancel_flags
);
4511 BUILD_BUG_SQE_ELEM(28, __u32
, open_flags
);
4512 BUILD_BUG_SQE_ELEM(28, __u32
, statx_flags
);
4513 BUILD_BUG_SQE_ELEM(28, __u32
, fadvise_advice
);
4514 BUILD_BUG_SQE_ELEM(28, __u32
, splice_flags
);
4515 BUILD_BUG_SQE_ELEM(28, __u32
, rename_flags
);
4516 BUILD_BUG_SQE_ELEM(28, __u32
, unlink_flags
);
4517 BUILD_BUG_SQE_ELEM(28, __u32
, hardlink_flags
);
4518 BUILD_BUG_SQE_ELEM(28, __u32
, xattr_flags
);
4519 BUILD_BUG_SQE_ELEM(28, __u32
, msg_ring_flags
);
4520 BUILD_BUG_SQE_ELEM(32, __u64
, user_data
);
4521 BUILD_BUG_SQE_ELEM(40, __u16
, buf_index
);
4522 BUILD_BUG_SQE_ELEM(40, __u16
, buf_group
);
4523 BUILD_BUG_SQE_ELEM(42, __u16
, personality
);
4524 BUILD_BUG_SQE_ELEM(44, __s32
, splice_fd_in
);
4525 BUILD_BUG_SQE_ELEM(44, __u32
, file_index
);
4526 BUILD_BUG_SQE_ELEM(44, __u16
, addr_len
);
4527 BUILD_BUG_SQE_ELEM(46, __u16
, __pad3
[0]);
4528 BUILD_BUG_SQE_ELEM(48, __u64
, addr3
);
4529 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd
);
4530 BUILD_BUG_SQE_ELEM(56, __u64
, __pad2
);
4532 BUILD_BUG_ON(sizeof(struct io_uring_files_update
) !=
4533 sizeof(struct io_uring_rsrc_update
));
4534 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update
) >
4535 sizeof(struct io_uring_rsrc_update2
));
4537 /* ->buf_index is u16 */
4538 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring
, bufs
) != 0);
4539 BUILD_BUG_ON(offsetof(struct io_uring_buf
, resv
) !=
4540 offsetof(struct io_uring_buf_ring
, tail
));
4542 /* should fit into one byte */
4543 BUILD_BUG_ON(SQE_VALID_FLAGS
>= (1 << 8));
4544 BUILD_BUG_ON(SQE_COMMON_FLAGS
>= (1 << 8));
4545 BUILD_BUG_ON((SQE_VALID_FLAGS
| SQE_COMMON_FLAGS
) != SQE_VALID_FLAGS
);
4547 BUILD_BUG_ON(__REQ_F_LAST_BIT
> 8 * sizeof(int));
4549 BUILD_BUG_ON(sizeof(atomic_t
) != sizeof(u32
));
4551 io_uring_optable_init();
4553 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
|
4554 SLAB_ACCOUNT
| SLAB_TYPESAFE_BY_RCU
);
4557 __initcall(io_uring_init
);