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>
99 #include "alloc_cache.h"
101 #define IORING_MAX_ENTRIES 32768
102 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
104 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
105 IORING_REGISTER_LAST + IORING_OP_LAST)
107 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
108 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
110 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
111 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
113 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
114 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
117 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
120 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
122 #define IO_COMPL_BATCH 32
123 #define IO_REQ_ALLOC_BATCH 8
126 IO_CHECK_CQ_OVERFLOW_BIT
,
127 IO_CHECK_CQ_DROPPED_BIT
,
131 IO_EVENTFD_OP_SIGNAL_BIT
,
132 IO_EVENTFD_OP_FREE_BIT
,
135 struct io_defer_entry
{
136 struct list_head list
;
137 struct io_kiocb
*req
;
141 /* requests with any of those set should undergo io_disarm_next() */
142 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
143 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
145 static bool io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
146 struct task_struct
*task
,
149 static void io_queue_sqe(struct io_kiocb
*req
);
150 static void io_move_task_work_from_local(struct io_ring_ctx
*ctx
);
151 static void __io_submit_flush_completions(struct io_ring_ctx
*ctx
);
153 struct kmem_cache
*req_cachep
;
155 struct sock
*io_uring_get_socket(struct file
*file
)
157 #if defined(CONFIG_UNIX)
158 if (io_is_uring_fops(file
)) {
159 struct io_ring_ctx
*ctx
= file
->private_data
;
161 return ctx
->ring_sock
->sk
;
166 EXPORT_SYMBOL(io_uring_get_socket
);
168 static inline void io_submit_flush_completions(struct io_ring_ctx
*ctx
)
170 if (!wq_list_empty(&ctx
->submit_state
.compl_reqs
) ||
171 ctx
->submit_state
.cqes_count
)
172 __io_submit_flush_completions(ctx
);
175 static inline unsigned int __io_cqring_events(struct io_ring_ctx
*ctx
)
177 return ctx
->cached_cq_tail
- READ_ONCE(ctx
->rings
->cq
.head
);
180 static inline unsigned int __io_cqring_events_user(struct io_ring_ctx
*ctx
)
182 return READ_ONCE(ctx
->rings
->cq
.tail
) - READ_ONCE(ctx
->rings
->cq
.head
);
185 static bool io_match_linked(struct io_kiocb
*head
)
187 struct io_kiocb
*req
;
189 io_for_each_link(req
, head
) {
190 if (req
->flags
& REQ_F_INFLIGHT
)
197 * As io_match_task() but protected against racing with linked timeouts.
198 * User must not hold timeout_lock.
200 bool io_match_task_safe(struct io_kiocb
*head
, struct task_struct
*task
,
205 if (task
&& head
->task
!= task
)
210 if (head
->flags
& REQ_F_LINK_TIMEOUT
) {
211 struct io_ring_ctx
*ctx
= head
->ctx
;
213 /* protect against races with linked timeouts */
214 spin_lock_irq(&ctx
->timeout_lock
);
215 matched
= io_match_linked(head
);
216 spin_unlock_irq(&ctx
->timeout_lock
);
218 matched
= io_match_linked(head
);
223 static inline void req_fail_link_node(struct io_kiocb
*req
, int res
)
226 io_req_set_res(req
, res
, 0);
229 static inline void io_req_add_to_cache(struct io_kiocb
*req
, struct io_ring_ctx
*ctx
)
231 wq_stack_add_head(&req
->comp_list
, &ctx
->submit_state
.free_list
);
232 kasan_poison_object_data(req_cachep
, req
);
235 static __cold
void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
237 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
239 complete(&ctx
->ref_comp
);
242 static __cold
void io_fallback_req_func(struct work_struct
*work
)
244 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
,
246 struct llist_node
*node
= llist_del_all(&ctx
->fallback_llist
);
247 struct io_kiocb
*req
, *tmp
;
248 struct io_tw_state ts
= { .locked
= true, };
250 mutex_lock(&ctx
->uring_lock
);
251 llist_for_each_entry_safe(req
, tmp
, node
, io_task_work
.node
)
252 req
->io_task_work
.func(req
, &ts
);
253 if (WARN_ON_ONCE(!ts
.locked
))
255 io_submit_flush_completions(ctx
);
256 mutex_unlock(&ctx
->uring_lock
);
259 static int io_alloc_hash_table(struct io_hash_table
*table
, unsigned bits
)
261 unsigned hash_buckets
= 1U << bits
;
262 size_t hash_size
= hash_buckets
* sizeof(table
->hbs
[0]);
264 table
->hbs
= kmalloc(hash_size
, GFP_KERNEL
);
268 table
->hash_bits
= bits
;
269 init_hash_table(table
, hash_buckets
);
273 static __cold
struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
275 struct io_ring_ctx
*ctx
;
278 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
282 xa_init(&ctx
->io_bl_xa
);
285 * Use 5 bits less than the max cq entries, that should give us around
286 * 32 entries per hash list if totally full and uniformly spread, but
287 * don't keep too many buckets to not overconsume memory.
289 hash_bits
= ilog2(p
->cq_entries
) - 5;
290 hash_bits
= clamp(hash_bits
, 1, 8);
291 if (io_alloc_hash_table(&ctx
->cancel_table
, hash_bits
))
293 if (io_alloc_hash_table(&ctx
->cancel_table_locked
, hash_bits
))
296 ctx
->dummy_ubuf
= kzalloc(sizeof(*ctx
->dummy_ubuf
), GFP_KERNEL
);
297 if (!ctx
->dummy_ubuf
)
299 /* set invalid range, so io_import_fixed() fails meeting it */
300 ctx
->dummy_ubuf
->ubuf
= -1UL;
302 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
,
306 ctx
->flags
= p
->flags
;
307 init_waitqueue_head(&ctx
->sqo_sq_wait
);
308 INIT_LIST_HEAD(&ctx
->sqd_list
);
309 INIT_LIST_HEAD(&ctx
->cq_overflow_list
);
310 INIT_LIST_HEAD(&ctx
->io_buffers_cache
);
311 io_alloc_cache_init(&ctx
->rsrc_node_cache
, IO_NODE_ALLOC_CACHE_MAX
,
312 sizeof(struct io_rsrc_node
));
313 io_alloc_cache_init(&ctx
->apoll_cache
, IO_ALLOC_CACHE_MAX
,
314 sizeof(struct async_poll
));
315 io_alloc_cache_init(&ctx
->netmsg_cache
, IO_ALLOC_CACHE_MAX
,
316 sizeof(struct io_async_msghdr
));
317 init_completion(&ctx
->ref_comp
);
318 xa_init_flags(&ctx
->personalities
, XA_FLAGS_ALLOC1
);
319 mutex_init(&ctx
->uring_lock
);
320 init_waitqueue_head(&ctx
->cq_wait
);
321 init_waitqueue_head(&ctx
->poll_wq
);
322 init_waitqueue_head(&ctx
->rsrc_quiesce_wq
);
323 spin_lock_init(&ctx
->completion_lock
);
324 spin_lock_init(&ctx
->timeout_lock
);
325 INIT_WQ_LIST(&ctx
->iopoll_list
);
326 INIT_LIST_HEAD(&ctx
->io_buffers_pages
);
327 INIT_LIST_HEAD(&ctx
->io_buffers_comp
);
328 INIT_LIST_HEAD(&ctx
->defer_list
);
329 INIT_LIST_HEAD(&ctx
->timeout_list
);
330 INIT_LIST_HEAD(&ctx
->ltimeout_list
);
331 INIT_LIST_HEAD(&ctx
->rsrc_ref_list
);
332 init_llist_head(&ctx
->work_llist
);
333 INIT_LIST_HEAD(&ctx
->tctx_list
);
334 ctx
->submit_state
.free_list
.next
= NULL
;
335 INIT_WQ_LIST(&ctx
->locked_free_list
);
336 INIT_DELAYED_WORK(&ctx
->fallback_work
, io_fallback_req_func
);
337 INIT_WQ_LIST(&ctx
->submit_state
.compl_reqs
);
340 kfree(ctx
->dummy_ubuf
);
341 kfree(ctx
->cancel_table
.hbs
);
342 kfree(ctx
->cancel_table_locked
.hbs
);
344 xa_destroy(&ctx
->io_bl_xa
);
349 static void io_account_cq_overflow(struct io_ring_ctx
*ctx
)
351 struct io_rings
*r
= ctx
->rings
;
353 WRITE_ONCE(r
->cq_overflow
, READ_ONCE(r
->cq_overflow
) + 1);
357 static bool req_need_defer(struct io_kiocb
*req
, u32 seq
)
359 if (unlikely(req
->flags
& REQ_F_IO_DRAIN
)) {
360 struct io_ring_ctx
*ctx
= req
->ctx
;
362 return seq
+ READ_ONCE(ctx
->cq_extra
) != ctx
->cached_cq_tail
;
368 static void io_clean_op(struct io_kiocb
*req
)
370 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
371 spin_lock(&req
->ctx
->completion_lock
);
372 io_put_kbuf_comp(req
);
373 spin_unlock(&req
->ctx
->completion_lock
);
376 if (req
->flags
& REQ_F_NEED_CLEANUP
) {
377 const struct io_cold_def
*def
= &io_cold_defs
[req
->opcode
];
382 if ((req
->flags
& REQ_F_POLLED
) && req
->apoll
) {
383 kfree(req
->apoll
->double_poll
);
387 if (req
->flags
& REQ_F_INFLIGHT
) {
388 struct io_uring_task
*tctx
= req
->task
->io_uring
;
390 atomic_dec(&tctx
->inflight_tracked
);
392 if (req
->flags
& REQ_F_CREDS
)
393 put_cred(req
->creds
);
394 if (req
->flags
& REQ_F_ASYNC_DATA
) {
395 kfree(req
->async_data
);
396 req
->async_data
= NULL
;
398 req
->flags
&= ~IO_REQ_CLEAN_FLAGS
;
401 static inline void io_req_track_inflight(struct io_kiocb
*req
)
403 if (!(req
->flags
& REQ_F_INFLIGHT
)) {
404 req
->flags
|= REQ_F_INFLIGHT
;
405 atomic_inc(&req
->task
->io_uring
->inflight_tracked
);
409 static struct io_kiocb
*__io_prep_linked_timeout(struct io_kiocb
*req
)
411 if (WARN_ON_ONCE(!req
->link
))
414 req
->flags
&= ~REQ_F_ARM_LTIMEOUT
;
415 req
->flags
|= REQ_F_LINK_TIMEOUT
;
417 /* linked timeouts should have two refs once prep'ed */
418 io_req_set_refcount(req
);
419 __io_req_set_refcount(req
->link
, 2);
423 static inline struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
)
425 if (likely(!(req
->flags
& REQ_F_ARM_LTIMEOUT
)))
427 return __io_prep_linked_timeout(req
);
430 static noinline
void __io_arm_ltimeout(struct io_kiocb
*req
)
432 io_queue_linked_timeout(__io_prep_linked_timeout(req
));
435 static inline void io_arm_ltimeout(struct io_kiocb
*req
)
437 if (unlikely(req
->flags
& REQ_F_ARM_LTIMEOUT
))
438 __io_arm_ltimeout(req
);
441 static void io_prep_async_work(struct io_kiocb
*req
)
443 const struct io_issue_def
*def
= &io_issue_defs
[req
->opcode
];
444 struct io_ring_ctx
*ctx
= req
->ctx
;
446 if (!(req
->flags
& REQ_F_CREDS
)) {
447 req
->flags
|= REQ_F_CREDS
;
448 req
->creds
= get_current_cred();
451 req
->work
.list
.next
= NULL
;
453 req
->work
.cancel_seq
= atomic_read(&ctx
->cancel_seq
);
454 if (req
->flags
& REQ_F_FORCE_ASYNC
)
455 req
->work
.flags
|= IO_WQ_WORK_CONCURRENT
;
457 if (req
->file
&& !(req
->flags
& REQ_F_FIXED_FILE
))
458 req
->flags
|= io_file_get_flags(req
->file
);
460 if (req
->file
&& (req
->flags
& REQ_F_ISREG
)) {
461 bool should_hash
= def
->hash_reg_file
;
463 /* don't serialize this request if the fs doesn't need it */
464 if (should_hash
&& (req
->file
->f_flags
& O_DIRECT
) &&
465 (req
->file
->f_mode
& FMODE_DIO_PARALLEL_WRITE
))
467 if (should_hash
|| (ctx
->flags
& IORING_SETUP_IOPOLL
))
468 io_wq_hash_work(&req
->work
, file_inode(req
->file
));
469 } else if (!req
->file
|| !S_ISBLK(file_inode(req
->file
)->i_mode
)) {
470 if (def
->unbound_nonreg_file
)
471 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
475 static void io_prep_async_link(struct io_kiocb
*req
)
477 struct io_kiocb
*cur
;
479 if (req
->flags
& REQ_F_LINK_TIMEOUT
) {
480 struct io_ring_ctx
*ctx
= req
->ctx
;
482 spin_lock_irq(&ctx
->timeout_lock
);
483 io_for_each_link(cur
, req
)
484 io_prep_async_work(cur
);
485 spin_unlock_irq(&ctx
->timeout_lock
);
487 io_for_each_link(cur
, req
)
488 io_prep_async_work(cur
);
492 void io_queue_iowq(struct io_kiocb
*req
, struct io_tw_state
*ts_dont_use
)
494 struct io_kiocb
*link
= io_prep_linked_timeout(req
);
495 struct io_uring_task
*tctx
= req
->task
->io_uring
;
498 BUG_ON(!tctx
->io_wq
);
500 /* init ->work of the whole link before punting */
501 io_prep_async_link(req
);
504 * Not expected to happen, but if we do have a bug where this _can_
505 * happen, catch it here and ensure the request is marked as
506 * canceled. That will make io-wq go through the usual work cancel
507 * procedure rather than attempt to run this request (or create a new
510 if (WARN_ON_ONCE(!same_thread_group(req
->task
, current
)))
511 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
513 trace_io_uring_queue_async_work(req
, io_wq_is_hashed(&req
->work
));
514 io_wq_enqueue(tctx
->io_wq
, &req
->work
);
516 io_queue_linked_timeout(link
);
519 static __cold
void io_queue_deferred(struct io_ring_ctx
*ctx
)
521 while (!list_empty(&ctx
->defer_list
)) {
522 struct io_defer_entry
*de
= list_first_entry(&ctx
->defer_list
,
523 struct io_defer_entry
, list
);
525 if (req_need_defer(de
->req
, de
->seq
))
527 list_del_init(&de
->list
);
528 io_req_task_queue(de
->req
);
534 static void io_eventfd_ops(struct rcu_head
*rcu
)
536 struct io_ev_fd
*ev_fd
= container_of(rcu
, struct io_ev_fd
, rcu
);
537 int ops
= atomic_xchg(&ev_fd
->ops
, 0);
539 if (ops
& BIT(IO_EVENTFD_OP_SIGNAL_BIT
))
540 eventfd_signal_mask(ev_fd
->cq_ev_fd
, 1, EPOLL_URING_WAKE
);
542 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
543 * ordering in a race but if references are 0 we know we have to free
546 if (atomic_dec_and_test(&ev_fd
->refs
)) {
547 eventfd_ctx_put(ev_fd
->cq_ev_fd
);
552 static void io_eventfd_signal(struct io_ring_ctx
*ctx
)
554 struct io_ev_fd
*ev_fd
= NULL
;
558 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
561 ev_fd
= rcu_dereference(ctx
->io_ev_fd
);
564 * Check again if ev_fd exists incase an io_eventfd_unregister call
565 * completed between the NULL check of ctx->io_ev_fd at the start of
566 * the function and rcu_read_lock.
568 if (unlikely(!ev_fd
))
570 if (READ_ONCE(ctx
->rings
->cq_flags
) & IORING_CQ_EVENTFD_DISABLED
)
572 if (ev_fd
->eventfd_async
&& !io_wq_current_is_worker())
575 if (likely(eventfd_signal_allowed())) {
576 eventfd_signal_mask(ev_fd
->cq_ev_fd
, 1, EPOLL_URING_WAKE
);
578 atomic_inc(&ev_fd
->refs
);
579 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT
), &ev_fd
->ops
))
580 call_rcu_hurry(&ev_fd
->rcu
, io_eventfd_ops
);
582 atomic_dec(&ev_fd
->refs
);
589 static void io_eventfd_flush_signal(struct io_ring_ctx
*ctx
)
593 spin_lock(&ctx
->completion_lock
);
596 * Eventfd should only get triggered when at least one event has been
597 * posted. Some applications rely on the eventfd notification count
598 * only changing IFF a new CQE has been added to the CQ ring. There's
599 * no depedency on 1:1 relationship between how many times this
600 * function is called (and hence the eventfd count) and number of CQEs
601 * posted to the CQ ring.
603 skip
= ctx
->cached_cq_tail
== ctx
->evfd_last_cq_tail
;
604 ctx
->evfd_last_cq_tail
= ctx
->cached_cq_tail
;
605 spin_unlock(&ctx
->completion_lock
);
609 io_eventfd_signal(ctx
);
612 void __io_commit_cqring_flush(struct io_ring_ctx
*ctx
)
614 if (ctx
->poll_activated
)
615 io_poll_wq_wake(ctx
);
616 if (ctx
->off_timeout_used
)
617 io_flush_timeouts(ctx
);
618 if (ctx
->drain_active
) {
619 spin_lock(&ctx
->completion_lock
);
620 io_queue_deferred(ctx
);
621 spin_unlock(&ctx
->completion_lock
);
624 io_eventfd_flush_signal(ctx
);
627 static inline void __io_cq_lock(struct io_ring_ctx
*ctx
)
629 if (!ctx
->task_complete
)
630 spin_lock(&ctx
->completion_lock
);
633 static inline void io_cq_lock(struct io_ring_ctx
*ctx
)
634 __acquires(ctx
->completion_lock
)
636 spin_lock(&ctx
->completion_lock
);
639 static inline void __io_cq_unlock_post(struct io_ring_ctx
*ctx
)
641 io_commit_cqring(ctx
);
643 if (ctx
->task_complete
) {
645 * ->task_complete implies that only current might be waiting
646 * for CQEs, and obviously, we currently don't. No one is
647 * waiting, wakeups are futile, skip them.
649 io_commit_cqring_flush(ctx
);
651 spin_unlock(&ctx
->completion_lock
);
652 io_commit_cqring_flush(ctx
);
657 static void io_cq_unlock_post(struct io_ring_ctx
*ctx
)
658 __releases(ctx
->completion_lock
)
660 io_commit_cqring(ctx
);
661 spin_unlock(&ctx
->completion_lock
);
662 io_commit_cqring_flush(ctx
);
666 /* Returns true if there are no backlogged entries after the flush */
667 static void io_cqring_overflow_kill(struct io_ring_ctx
*ctx
)
669 struct io_overflow_cqe
*ocqe
;
672 spin_lock(&ctx
->completion_lock
);
673 list_splice_init(&ctx
->cq_overflow_list
, &list
);
674 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
);
675 spin_unlock(&ctx
->completion_lock
);
677 while (!list_empty(&list
)) {
678 ocqe
= list_first_entry(&list
, struct io_overflow_cqe
, list
);
679 list_del(&ocqe
->list
);
684 static void __io_cqring_overflow_flush(struct io_ring_ctx
*ctx
)
686 size_t cqe_size
= sizeof(struct io_uring_cqe
);
688 if (__io_cqring_events(ctx
) == ctx
->cq_entries
)
691 if (ctx
->flags
& IORING_SETUP_CQE32
)
695 while (!list_empty(&ctx
->cq_overflow_list
)) {
696 struct io_uring_cqe
*cqe
= io_get_cqe_overflow(ctx
, true);
697 struct io_overflow_cqe
*ocqe
;
701 ocqe
= list_first_entry(&ctx
->cq_overflow_list
,
702 struct io_overflow_cqe
, list
);
703 memcpy(cqe
, &ocqe
->cqe
, cqe_size
);
704 list_del(&ocqe
->list
);
708 if (list_empty(&ctx
->cq_overflow_list
)) {
709 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
);
710 atomic_andnot(IORING_SQ_CQ_OVERFLOW
, &ctx
->rings
->sq_flags
);
712 io_cq_unlock_post(ctx
);
715 static void io_cqring_do_overflow_flush(struct io_ring_ctx
*ctx
)
717 /* iopoll syncs against uring_lock, not completion_lock */
718 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
719 mutex_lock(&ctx
->uring_lock
);
720 __io_cqring_overflow_flush(ctx
);
721 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
722 mutex_unlock(&ctx
->uring_lock
);
725 static void io_cqring_overflow_flush(struct io_ring_ctx
*ctx
)
727 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
))
728 io_cqring_do_overflow_flush(ctx
);
731 /* can be called by any task */
732 static void io_put_task_remote(struct task_struct
*task
)
734 struct io_uring_task
*tctx
= task
->io_uring
;
736 percpu_counter_sub(&tctx
->inflight
, 1);
737 if (unlikely(atomic_read(&tctx
->in_cancel
)))
738 wake_up(&tctx
->wait
);
739 put_task_struct(task
);
742 /* used by a task to put its own references */
743 static void io_put_task_local(struct task_struct
*task
)
745 task
->io_uring
->cached_refs
++;
748 /* must to be called somewhat shortly after putting a request */
749 static inline void io_put_task(struct task_struct
*task
)
751 if (likely(task
== current
))
752 io_put_task_local(task
);
754 io_put_task_remote(task
);
757 void io_task_refs_refill(struct io_uring_task
*tctx
)
759 unsigned int refill
= -tctx
->cached_refs
+ IO_TCTX_REFS_CACHE_NR
;
761 percpu_counter_add(&tctx
->inflight
, refill
);
762 refcount_add(refill
, ¤t
->usage
);
763 tctx
->cached_refs
+= refill
;
766 static __cold
void io_uring_drop_tctx_refs(struct task_struct
*task
)
768 struct io_uring_task
*tctx
= task
->io_uring
;
769 unsigned int refs
= tctx
->cached_refs
;
772 tctx
->cached_refs
= 0;
773 percpu_counter_sub(&tctx
->inflight
, refs
);
774 put_task_struct_many(task
, refs
);
778 static bool io_cqring_event_overflow(struct io_ring_ctx
*ctx
, u64 user_data
,
779 s32 res
, u32 cflags
, u64 extra1
, u64 extra2
)
781 struct io_overflow_cqe
*ocqe
;
782 size_t ocq_size
= sizeof(struct io_overflow_cqe
);
783 bool is_cqe32
= (ctx
->flags
& IORING_SETUP_CQE32
);
785 lockdep_assert_held(&ctx
->completion_lock
);
788 ocq_size
+= sizeof(struct io_uring_cqe
);
790 ocqe
= kmalloc(ocq_size
, GFP_ATOMIC
| __GFP_ACCOUNT
);
791 trace_io_uring_cqe_overflow(ctx
, user_data
, res
, cflags
, ocqe
);
794 * If we're in ring overflow flush mode, or in task cancel mode,
795 * or cannot allocate an overflow entry, then we need to drop it
798 io_account_cq_overflow(ctx
);
799 set_bit(IO_CHECK_CQ_DROPPED_BIT
, &ctx
->check_cq
);
802 if (list_empty(&ctx
->cq_overflow_list
)) {
803 set_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
);
804 atomic_or(IORING_SQ_CQ_OVERFLOW
, &ctx
->rings
->sq_flags
);
807 ocqe
->cqe
.user_data
= user_data
;
809 ocqe
->cqe
.flags
= cflags
;
811 ocqe
->cqe
.big_cqe
[0] = extra1
;
812 ocqe
->cqe
.big_cqe
[1] = extra2
;
814 list_add_tail(&ocqe
->list
, &ctx
->cq_overflow_list
);
818 bool io_req_cqe_overflow(struct io_kiocb
*req
)
820 if (!(req
->flags
& REQ_F_CQE32_INIT
)) {
824 return io_cqring_event_overflow(req
->ctx
, req
->cqe
.user_data
,
825 req
->cqe
.res
, req
->cqe
.flags
,
826 req
->extra1
, req
->extra2
);
830 * writes to the cq entry need to come after reading head; the
831 * control dependency is enough as we're using WRITE_ONCE to
834 struct io_uring_cqe
*__io_get_cqe(struct io_ring_ctx
*ctx
, bool overflow
)
836 struct io_rings
*rings
= ctx
->rings
;
837 unsigned int off
= ctx
->cached_cq_tail
& (ctx
->cq_entries
- 1);
838 unsigned int free
, queued
, len
;
841 * Posting into the CQ when there are pending overflowed CQEs may break
842 * ordering guarantees, which will affect links, F_MORE users and more.
843 * Force overflow the completion.
845 if (!overflow
&& (ctx
->check_cq
& BIT(IO_CHECK_CQ_OVERFLOW_BIT
)))
848 /* userspace may cheat modifying the tail, be safe and do min */
849 queued
= min(__io_cqring_events(ctx
), ctx
->cq_entries
);
850 free
= ctx
->cq_entries
- queued
;
851 /* we need a contiguous range, limit based on the current array offset */
852 len
= min(free
, ctx
->cq_entries
- off
);
856 if (ctx
->flags
& IORING_SETUP_CQE32
) {
861 ctx
->cqe_cached
= &rings
->cqes
[off
];
862 ctx
->cqe_sentinel
= ctx
->cqe_cached
+ len
;
864 ctx
->cached_cq_tail
++;
866 if (ctx
->flags
& IORING_SETUP_CQE32
)
868 return &rings
->cqes
[off
];
871 static bool io_fill_cqe_aux(struct io_ring_ctx
*ctx
, u64 user_data
, s32 res
,
874 struct io_uring_cqe
*cqe
;
879 * If we can't get a cq entry, userspace overflowed the
880 * submission (by quite a lot). Increment the overflow count in
883 cqe
= io_get_cqe(ctx
);
885 trace_io_uring_complete(ctx
, NULL
, user_data
, res
, cflags
, 0, 0);
887 WRITE_ONCE(cqe
->user_data
, user_data
);
888 WRITE_ONCE(cqe
->res
, res
);
889 WRITE_ONCE(cqe
->flags
, cflags
);
891 if (ctx
->flags
& IORING_SETUP_CQE32
) {
892 WRITE_ONCE(cqe
->big_cqe
[0], 0);
893 WRITE_ONCE(cqe
->big_cqe
[1], 0);
900 static void __io_flush_post_cqes(struct io_ring_ctx
*ctx
)
901 __must_hold(&ctx
->uring_lock
)
903 struct io_submit_state
*state
= &ctx
->submit_state
;
906 lockdep_assert_held(&ctx
->uring_lock
);
907 for (i
= 0; i
< state
->cqes_count
; i
++) {
908 struct io_uring_cqe
*cqe
= &state
->cqes
[i
];
910 if (!io_fill_cqe_aux(ctx
, cqe
->user_data
, cqe
->res
, cqe
->flags
)) {
911 if (ctx
->task_complete
) {
912 spin_lock(&ctx
->completion_lock
);
913 io_cqring_event_overflow(ctx
, cqe
->user_data
,
914 cqe
->res
, cqe
->flags
, 0, 0);
915 spin_unlock(&ctx
->completion_lock
);
917 io_cqring_event_overflow(ctx
, cqe
->user_data
,
918 cqe
->res
, cqe
->flags
, 0, 0);
922 state
->cqes_count
= 0;
925 static bool __io_post_aux_cqe(struct io_ring_ctx
*ctx
, u64 user_data
, s32 res
, u32 cflags
,
931 filled
= io_fill_cqe_aux(ctx
, user_data
, res
, cflags
);
932 if (!filled
&& allow_overflow
)
933 filled
= io_cqring_event_overflow(ctx
, user_data
, res
, cflags
, 0, 0);
935 io_cq_unlock_post(ctx
);
939 bool io_post_aux_cqe(struct io_ring_ctx
*ctx
, u64 user_data
, s32 res
, u32 cflags
)
941 return __io_post_aux_cqe(ctx
, user_data
, res
, cflags
, true);
944 bool io_aux_cqe(const struct io_kiocb
*req
, bool defer
, s32 res
, u32 cflags
,
947 struct io_ring_ctx
*ctx
= req
->ctx
;
948 u64 user_data
= req
->cqe
.user_data
;
949 struct io_uring_cqe
*cqe
;
952 return __io_post_aux_cqe(ctx
, user_data
, res
, cflags
, allow_overflow
);
954 lockdep_assert_held(&ctx
->uring_lock
);
956 if (ctx
->submit_state
.cqes_count
== ARRAY_SIZE(ctx
->submit_state
.cqes
)) {
958 __io_flush_post_cqes(ctx
);
959 /* no need to flush - flush is deferred */
960 __io_cq_unlock_post(ctx
);
963 /* For defered completions this is not as strict as it is otherwise,
964 * however it's main job is to prevent unbounded posted completions,
965 * and in that it works just as well.
967 if (!allow_overflow
&& test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
))
970 cqe
= &ctx
->submit_state
.cqes
[ctx
->submit_state
.cqes_count
++];
971 cqe
->user_data
= user_data
;
977 static void __io_req_complete_post(struct io_kiocb
*req
, unsigned issue_flags
)
979 struct io_ring_ctx
*ctx
= req
->ctx
;
980 struct io_rsrc_node
*rsrc_node
= NULL
;
983 if (!(req
->flags
& REQ_F_CQE_SKIP
))
984 io_fill_cqe_req(ctx
, req
);
987 * If we're the last reference to this request, add to our locked
990 if (req_ref_put_and_test(req
)) {
991 if (req
->flags
& IO_REQ_LINK_FLAGS
) {
992 if (req
->flags
& IO_DISARM_MASK
)
995 io_req_task_queue(req
->link
);
999 io_put_kbuf_comp(req
);
1000 if (unlikely(req
->flags
& IO_REQ_CLEAN_FLAGS
))
1002 if (!(req
->flags
& REQ_F_FIXED_FILE
))
1003 io_put_file(req
->file
);
1005 rsrc_node
= req
->rsrc_node
;
1007 * Selected buffer deallocation in io_clean_op() assumes that
1008 * we don't hold ->completion_lock. Clean them here to avoid
1011 io_put_task_remote(req
->task
);
1012 wq_list_add_head(&req
->comp_list
, &ctx
->locked_free_list
);
1013 ctx
->locked_free_nr
++;
1015 io_cq_unlock_post(ctx
);
1018 io_ring_submit_lock(ctx
, issue_flags
);
1019 io_put_rsrc_node(ctx
, rsrc_node
);
1020 io_ring_submit_unlock(ctx
, issue_flags
);
1024 void io_req_complete_post(struct io_kiocb
*req
, unsigned issue_flags
)
1026 if (req
->ctx
->task_complete
&& req
->ctx
->submitter_task
!= current
) {
1027 req
->io_task_work
.func
= io_req_task_complete
;
1028 io_req_task_work_add(req
);
1029 } else if (!(issue_flags
& IO_URING_F_UNLOCKED
) ||
1030 !(req
->ctx
->flags
& IORING_SETUP_IOPOLL
)) {
1031 __io_req_complete_post(req
, issue_flags
);
1033 struct io_ring_ctx
*ctx
= req
->ctx
;
1035 mutex_lock(&ctx
->uring_lock
);
1036 __io_req_complete_post(req
, issue_flags
& ~IO_URING_F_UNLOCKED
);
1037 mutex_unlock(&ctx
->uring_lock
);
1041 void io_req_defer_failed(struct io_kiocb
*req
, s32 res
)
1042 __must_hold(&ctx
->uring_lock
)
1044 const struct io_cold_def
*def
= &io_cold_defs
[req
->opcode
];
1046 lockdep_assert_held(&req
->ctx
->uring_lock
);
1049 io_req_set_res(req
, res
, io_put_kbuf(req
, IO_URING_F_UNLOCKED
));
1052 io_req_complete_defer(req
);
1056 * Don't initialise the fields below on every allocation, but do that in
1057 * advance and keep them valid across allocations.
1059 static void io_preinit_req(struct io_kiocb
*req
, struct io_ring_ctx
*ctx
)
1063 req
->async_data
= NULL
;
1064 /* not necessary, but safer to zero */
1068 static void io_flush_cached_locked_reqs(struct io_ring_ctx
*ctx
,
1069 struct io_submit_state
*state
)
1071 spin_lock(&ctx
->completion_lock
);
1072 wq_list_splice(&ctx
->locked_free_list
, &state
->free_list
);
1073 ctx
->locked_free_nr
= 0;
1074 spin_unlock(&ctx
->completion_lock
);
1078 * A request might get retired back into the request caches even before opcode
1079 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1080 * Because of that, io_alloc_req() should be called only under ->uring_lock
1081 * and with extra caution to not get a request that is still worked on.
1083 __cold
bool __io_alloc_req_refill(struct io_ring_ctx
*ctx
)
1084 __must_hold(&ctx
->uring_lock
)
1086 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
1087 void *reqs
[IO_REQ_ALLOC_BATCH
];
1091 * If we have more than a batch's worth of requests in our IRQ side
1092 * locked cache, grab the lock and move them over to our submission
1095 if (data_race(ctx
->locked_free_nr
) > IO_COMPL_BATCH
) {
1096 io_flush_cached_locked_reqs(ctx
, &ctx
->submit_state
);
1097 if (!io_req_cache_empty(ctx
))
1101 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, ARRAY_SIZE(reqs
), reqs
);
1104 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1105 * retry single alloc to be on the safe side.
1107 if (unlikely(ret
<= 0)) {
1108 reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
1114 percpu_ref_get_many(&ctx
->refs
, ret
);
1115 for (i
= 0; i
< ret
; i
++) {
1116 struct io_kiocb
*req
= reqs
[i
];
1118 io_preinit_req(req
, ctx
);
1119 io_req_add_to_cache(req
, ctx
);
1124 __cold
void io_free_req(struct io_kiocb
*req
)
1126 /* refs were already put, restore them for io_req_task_complete() */
1127 req
->flags
&= ~REQ_F_REFCOUNT
;
1128 /* we only want to free it, don't post CQEs */
1129 req
->flags
|= REQ_F_CQE_SKIP
;
1130 req
->io_task_work
.func
= io_req_task_complete
;
1131 io_req_task_work_add(req
);
1134 static void __io_req_find_next_prep(struct io_kiocb
*req
)
1136 struct io_ring_ctx
*ctx
= req
->ctx
;
1138 spin_lock(&ctx
->completion_lock
);
1139 io_disarm_next(req
);
1140 spin_unlock(&ctx
->completion_lock
);
1143 static inline struct io_kiocb
*io_req_find_next(struct io_kiocb
*req
)
1145 struct io_kiocb
*nxt
;
1148 * If LINK is set, we have dependent requests in this chain. If we
1149 * didn't fail this request, queue the first one up, moving any other
1150 * dependencies to the next request. In case of failure, fail the rest
1153 if (unlikely(req
->flags
& IO_DISARM_MASK
))
1154 __io_req_find_next_prep(req
);
1160 static void ctx_flush_and_put(struct io_ring_ctx
*ctx
, struct io_tw_state
*ts
)
1164 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
)
1165 atomic_andnot(IORING_SQ_TASKRUN
, &ctx
->rings
->sq_flags
);
1167 io_submit_flush_completions(ctx
);
1168 mutex_unlock(&ctx
->uring_lock
);
1171 percpu_ref_put(&ctx
->refs
);
1174 static unsigned int handle_tw_list(struct llist_node
*node
,
1175 struct io_ring_ctx
**ctx
,
1176 struct io_tw_state
*ts
,
1177 struct llist_node
*last
)
1179 unsigned int count
= 0;
1181 while (node
&& node
!= last
) {
1182 struct llist_node
*next
= node
->next
;
1183 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1186 prefetch(container_of(next
, struct io_kiocb
, io_task_work
.node
));
1188 if (req
->ctx
!= *ctx
) {
1189 ctx_flush_and_put(*ctx
, ts
);
1191 /* if not contended, grab and improve batching */
1192 ts
->locked
= mutex_trylock(&(*ctx
)->uring_lock
);
1193 percpu_ref_get(&(*ctx
)->refs
);
1195 INDIRECT_CALL_2(req
->io_task_work
.func
,
1196 io_poll_task_func
, io_req_rw_complete
,
1200 if (unlikely(need_resched())) {
1201 ctx_flush_and_put(*ctx
, ts
);
1211 * io_llist_xchg - swap all entries in a lock-less list
1212 * @head: the head of lock-less list to delete all entries
1213 * @new: new entry as the head of the list
1215 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1216 * The order of entries returned is from the newest to the oldest added one.
1218 static inline struct llist_node
*io_llist_xchg(struct llist_head
*head
,
1219 struct llist_node
*new)
1221 return xchg(&head
->first
, new);
1225 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1226 * @head: the head of lock-less list to delete all entries
1227 * @old: expected old value of the first entry of the list
1228 * @new: new entry as the head of the list
1230 * perform a cmpxchg on the first entry of the list.
1233 static inline struct llist_node
*io_llist_cmpxchg(struct llist_head
*head
,
1234 struct llist_node
*old
,
1235 struct llist_node
*new)
1237 return cmpxchg(&head
->first
, old
, new);
1240 static __cold
void io_fallback_tw(struct io_uring_task
*tctx
, bool sync
)
1242 struct llist_node
*node
= llist_del_all(&tctx
->task_list
);
1243 struct io_ring_ctx
*last_ctx
= NULL
;
1244 struct io_kiocb
*req
;
1247 req
= container_of(node
, struct io_kiocb
, io_task_work
.node
);
1249 if (sync
&& last_ctx
!= req
->ctx
) {
1251 flush_delayed_work(&last_ctx
->fallback_work
);
1252 percpu_ref_put(&last_ctx
->refs
);
1254 last_ctx
= req
->ctx
;
1255 percpu_ref_get(&last_ctx
->refs
);
1257 if (llist_add(&req
->io_task_work
.node
,
1258 &req
->ctx
->fallback_llist
))
1259 schedule_delayed_work(&req
->ctx
->fallback_work
, 1);
1263 flush_delayed_work(&last_ctx
->fallback_work
);
1264 percpu_ref_put(&last_ctx
->refs
);
1268 void tctx_task_work(struct callback_head
*cb
)
1270 struct io_tw_state ts
= {};
1271 struct io_ring_ctx
*ctx
= NULL
;
1272 struct io_uring_task
*tctx
= container_of(cb
, struct io_uring_task
,
1274 struct llist_node fake
= {};
1275 struct llist_node
*node
;
1276 unsigned int loops
= 0;
1277 unsigned int count
= 0;
1279 if (unlikely(current
->flags
& PF_EXITING
)) {
1280 io_fallback_tw(tctx
, true);
1286 node
= io_llist_xchg(&tctx
->task_list
, &fake
);
1287 count
+= handle_tw_list(node
, &ctx
, &ts
, &fake
);
1289 /* skip expensive cmpxchg if there are items in the list */
1290 if (READ_ONCE(tctx
->task_list
.first
) != &fake
)
1292 if (ts
.locked
&& !wq_list_empty(&ctx
->submit_state
.compl_reqs
)) {
1293 io_submit_flush_completions(ctx
);
1294 if (READ_ONCE(tctx
->task_list
.first
) != &fake
)
1297 node
= io_llist_cmpxchg(&tctx
->task_list
, &fake
, NULL
);
1298 } while (node
!= &fake
);
1300 ctx_flush_and_put(ctx
, &ts
);
1302 /* relaxed read is enough as only the task itself sets ->in_cancel */
1303 if (unlikely(atomic_read(&tctx
->in_cancel
)))
1304 io_uring_drop_tctx_refs(current
);
1306 trace_io_uring_task_work_run(tctx
, count
, loops
);
1309 static inline void io_req_local_work_add(struct io_kiocb
*req
, unsigned flags
)
1311 struct io_ring_ctx
*ctx
= req
->ctx
;
1312 unsigned nr_wait
, nr_tw
, nr_tw_prev
;
1313 struct llist_node
*first
;
1315 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))
1316 flags
&= ~IOU_F_TWQ_LAZY_WAKE
;
1318 first
= READ_ONCE(ctx
->work_llist
.first
);
1322 struct io_kiocb
*first_req
= container_of(first
,
1326 * Might be executed at any moment, rely on
1327 * SLAB_TYPESAFE_BY_RCU to keep it alive.
1329 nr_tw_prev
= READ_ONCE(first_req
->nr_tw
);
1331 nr_tw
= nr_tw_prev
+ 1;
1332 /* Large enough to fail the nr_wait comparison below */
1333 if (!(flags
& IOU_F_TWQ_LAZY_WAKE
))
1337 req
->io_task_work
.node
.next
= first
;
1338 } while (!try_cmpxchg(&ctx
->work_llist
.first
, &first
,
1339 &req
->io_task_work
.node
));
1342 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
)
1343 atomic_or(IORING_SQ_TASKRUN
, &ctx
->rings
->sq_flags
);
1345 io_eventfd_signal(ctx
);
1348 nr_wait
= atomic_read(&ctx
->cq_wait_nr
);
1349 /* no one is waiting */
1352 /* either not enough or the previous add has already woken it up */
1353 if (nr_wait
> nr_tw
|| nr_tw_prev
>= nr_wait
)
1355 /* pairs with set_current_state() in io_cqring_wait() */
1356 smp_mb__after_atomic();
1357 wake_up_state(ctx
->submitter_task
, TASK_INTERRUPTIBLE
);
1360 static void io_req_normal_work_add(struct io_kiocb
*req
)
1362 struct io_uring_task
*tctx
= req
->task
->io_uring
;
1363 struct io_ring_ctx
*ctx
= req
->ctx
;
1365 /* task_work already pending, we're done */
1366 if (!llist_add(&req
->io_task_work
.node
, &tctx
->task_list
))
1369 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
)
1370 atomic_or(IORING_SQ_TASKRUN
, &ctx
->rings
->sq_flags
);
1372 if (likely(!task_work_add(req
->task
, &tctx
->task_work
, ctx
->notify_method
)))
1375 io_fallback_tw(tctx
, false);
1378 void __io_req_task_work_add(struct io_kiocb
*req
, unsigned flags
)
1380 if (req
->ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) {
1382 io_req_local_work_add(req
, flags
);
1385 io_req_normal_work_add(req
);
1389 static void __cold
io_move_task_work_from_local(struct io_ring_ctx
*ctx
)
1391 struct llist_node
*node
;
1393 node
= llist_del_all(&ctx
->work_llist
);
1395 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1399 io_req_normal_work_add(req
);
1403 static int __io_run_local_work(struct io_ring_ctx
*ctx
, struct io_tw_state
*ts
)
1405 struct llist_node
*node
;
1406 unsigned int loops
= 0;
1409 if (WARN_ON_ONCE(ctx
->submitter_task
!= current
))
1411 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
)
1412 atomic_andnot(IORING_SQ_TASKRUN
, &ctx
->rings
->sq_flags
);
1415 * llists are in reverse order, flip it back the right way before
1416 * running the pending items.
1418 node
= llist_reverse_order(io_llist_xchg(&ctx
->work_llist
, NULL
));
1420 struct llist_node
*next
= node
->next
;
1421 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1423 prefetch(container_of(next
, struct io_kiocb
, io_task_work
.node
));
1424 INDIRECT_CALL_2(req
->io_task_work
.func
,
1425 io_poll_task_func
, io_req_rw_complete
,
1432 if (!llist_empty(&ctx
->work_llist
))
1435 io_submit_flush_completions(ctx
);
1436 if (!llist_empty(&ctx
->work_llist
))
1439 trace_io_uring_local_work_run(ctx
, ret
, loops
);
1443 static inline int io_run_local_work_locked(struct io_ring_ctx
*ctx
)
1445 struct io_tw_state ts
= { .locked
= true, };
1448 if (llist_empty(&ctx
->work_llist
))
1451 ret
= __io_run_local_work(ctx
, &ts
);
1452 /* shouldn't happen! */
1453 if (WARN_ON_ONCE(!ts
.locked
))
1454 mutex_lock(&ctx
->uring_lock
);
1458 static int io_run_local_work(struct io_ring_ctx
*ctx
)
1460 struct io_tw_state ts
= {};
1463 ts
.locked
= mutex_trylock(&ctx
->uring_lock
);
1464 ret
= __io_run_local_work(ctx
, &ts
);
1466 mutex_unlock(&ctx
->uring_lock
);
1471 static void io_req_task_cancel(struct io_kiocb
*req
, struct io_tw_state
*ts
)
1473 io_tw_lock(req
->ctx
, ts
);
1474 io_req_defer_failed(req
, req
->cqe
.res
);
1477 void io_req_task_submit(struct io_kiocb
*req
, struct io_tw_state
*ts
)
1479 io_tw_lock(req
->ctx
, ts
);
1480 /* req->task == current here, checking PF_EXITING is safe */
1481 if (unlikely(req
->task
->flags
& PF_EXITING
))
1482 io_req_defer_failed(req
, -EFAULT
);
1483 else if (req
->flags
& REQ_F_FORCE_ASYNC
)
1484 io_queue_iowq(req
, ts
);
1489 void io_req_task_queue_fail(struct io_kiocb
*req
, int ret
)
1491 io_req_set_res(req
, ret
, 0);
1492 req
->io_task_work
.func
= io_req_task_cancel
;
1493 io_req_task_work_add(req
);
1496 void io_req_task_queue(struct io_kiocb
*req
)
1498 req
->io_task_work
.func
= io_req_task_submit
;
1499 io_req_task_work_add(req
);
1502 void io_queue_next(struct io_kiocb
*req
)
1504 struct io_kiocb
*nxt
= io_req_find_next(req
);
1507 io_req_task_queue(nxt
);
1510 void io_free_batch_list(struct io_ring_ctx
*ctx
, struct io_wq_work_node
*node
)
1511 __must_hold(&ctx
->uring_lock
)
1514 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1517 if (unlikely(req
->flags
& IO_REQ_CLEAN_SLOW_FLAGS
)) {
1518 if (req
->flags
& REQ_F_REFCOUNT
) {
1519 node
= req
->comp_list
.next
;
1520 if (!req_ref_put_and_test(req
))
1523 if ((req
->flags
& REQ_F_POLLED
) && req
->apoll
) {
1524 struct async_poll
*apoll
= req
->apoll
;
1526 if (apoll
->double_poll
)
1527 kfree(apoll
->double_poll
);
1528 if (!io_alloc_cache_put(&ctx
->apoll_cache
, &apoll
->cache
))
1530 req
->flags
&= ~REQ_F_POLLED
;
1532 if (req
->flags
& IO_REQ_LINK_FLAGS
)
1534 if (unlikely(req
->flags
& IO_REQ_CLEAN_FLAGS
))
1537 if (!(req
->flags
& REQ_F_FIXED_FILE
))
1538 io_put_file(req
->file
);
1540 io_req_put_rsrc_locked(req
, ctx
);
1542 io_put_task(req
->task
);
1543 node
= req
->comp_list
.next
;
1544 io_req_add_to_cache(req
, ctx
);
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(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
);
1581 static unsigned io_cqring_events(struct io_ring_ctx
*ctx
)
1583 /* See comment at the top of this file */
1585 return __io_cqring_events(ctx
);
1589 * We can't just wait for polled events to come to us, we have to actively
1590 * find and complete them.
1592 static __cold
void io_iopoll_try_reap_events(struct io_ring_ctx
*ctx
)
1594 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
1597 mutex_lock(&ctx
->uring_lock
);
1598 while (!wq_list_empty(&ctx
->iopoll_list
)) {
1599 /* let it sleep and repeat later if can't complete a request */
1600 if (io_do_iopoll(ctx
, true) == 0)
1603 * Ensure we allow local-to-the-cpu processing to take place,
1604 * in this case we need to ensure that we reap all events.
1605 * Also let task_work, etc. to progress by releasing the mutex
1607 if (need_resched()) {
1608 mutex_unlock(&ctx
->uring_lock
);
1610 mutex_lock(&ctx
->uring_lock
);
1613 mutex_unlock(&ctx
->uring_lock
);
1616 static int io_iopoll_check(struct io_ring_ctx
*ctx
, long min
)
1618 unsigned int nr_events
= 0;
1620 unsigned long check_cq
;
1622 if (!io_allowed_run_tw(ctx
))
1625 check_cq
= READ_ONCE(ctx
->check_cq
);
1626 if (unlikely(check_cq
)) {
1627 if (check_cq
& BIT(IO_CHECK_CQ_OVERFLOW_BIT
))
1628 __io_cqring_overflow_flush(ctx
);
1630 * Similarly do not spin if we have not informed the user of any
1633 if (check_cq
& BIT(IO_CHECK_CQ_DROPPED_BIT
))
1637 * Don't enter poll loop if we already have events pending.
1638 * If we do, we can potentially be spinning for commands that
1639 * already triggered a CQE (eg in error).
1641 if (io_cqring_events(ctx
))
1646 * If a submit got punted to a workqueue, we can have the
1647 * application entering polling for a command before it gets
1648 * issued. That app will hold the uring_lock for the duration
1649 * of the poll right here, so we need to take a breather every
1650 * now and then to ensure that the issue has a chance to add
1651 * the poll to the issued list. Otherwise we can spin here
1652 * forever, while the workqueue is stuck trying to acquire the
1655 if (wq_list_empty(&ctx
->iopoll_list
) ||
1656 io_task_work_pending(ctx
)) {
1657 u32 tail
= ctx
->cached_cq_tail
;
1659 (void) io_run_local_work_locked(ctx
);
1661 if (task_work_pending(current
) ||
1662 wq_list_empty(&ctx
->iopoll_list
)) {
1663 mutex_unlock(&ctx
->uring_lock
);
1665 mutex_lock(&ctx
->uring_lock
);
1667 /* some requests don't go through iopoll_list */
1668 if (tail
!= ctx
->cached_cq_tail
||
1669 wq_list_empty(&ctx
->iopoll_list
))
1672 ret
= io_do_iopoll(ctx
, !min
);
1677 } while (nr_events
< min
&& !need_resched());
1682 void io_req_task_complete(struct io_kiocb
*req
, struct io_tw_state
*ts
)
1685 io_req_complete_defer(req
);
1687 io_req_complete_post(req
, IO_URING_F_UNLOCKED
);
1691 * After the iocb has been issued, it's safe to be found on the poll list.
1692 * Adding the kiocb to the list AFTER submission ensures that we don't
1693 * find it from a io_do_iopoll() thread before the issuer is done
1694 * accessing the kiocb cookie.
1696 static void io_iopoll_req_issued(struct io_kiocb
*req
, unsigned int issue_flags
)
1698 struct io_ring_ctx
*ctx
= req
->ctx
;
1699 const bool needs_lock
= issue_flags
& IO_URING_F_UNLOCKED
;
1701 /* workqueue context doesn't hold uring_lock, grab it now */
1702 if (unlikely(needs_lock
))
1703 mutex_lock(&ctx
->uring_lock
);
1706 * Track whether we have multiple files in our lists. This will impact
1707 * how we do polling eventually, not spinning if we're on potentially
1708 * different devices.
1710 if (wq_list_empty(&ctx
->iopoll_list
)) {
1711 ctx
->poll_multi_queue
= false;
1712 } else if (!ctx
->poll_multi_queue
) {
1713 struct io_kiocb
*list_req
;
1715 list_req
= container_of(ctx
->iopoll_list
.first
, struct io_kiocb
,
1717 if (list_req
->file
!= req
->file
)
1718 ctx
->poll_multi_queue
= true;
1722 * For fast devices, IO may have already completed. If it has, add
1723 * it to the front so we find it first.
1725 if (READ_ONCE(req
->iopoll_completed
))
1726 wq_list_add_head(&req
->comp_list
, &ctx
->iopoll_list
);
1728 wq_list_add_tail(&req
->comp_list
, &ctx
->iopoll_list
);
1730 if (unlikely(needs_lock
)) {
1732 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1733 * in sq thread task context or in io worker task context. If
1734 * current task context is sq thread, we don't need to check
1735 * whether should wake up sq thread.
1737 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) &&
1738 wq_has_sleeper(&ctx
->sq_data
->wait
))
1739 wake_up(&ctx
->sq_data
->wait
);
1741 mutex_unlock(&ctx
->uring_lock
);
1745 unsigned int io_file_get_flags(struct file
*file
)
1747 unsigned int res
= 0;
1749 if (S_ISREG(file_inode(file
)->i_mode
))
1751 if ((file
->f_flags
& O_NONBLOCK
) || (file
->f_mode
& FMODE_NOWAIT
))
1752 res
|= REQ_F_SUPPORT_NOWAIT
;
1756 bool io_alloc_async_data(struct io_kiocb
*req
)
1758 WARN_ON_ONCE(!io_cold_defs
[req
->opcode
].async_size
);
1759 req
->async_data
= kmalloc(io_cold_defs
[req
->opcode
].async_size
, GFP_KERNEL
);
1760 if (req
->async_data
) {
1761 req
->flags
|= REQ_F_ASYNC_DATA
;
1767 int io_req_prep_async(struct io_kiocb
*req
)
1769 const struct io_cold_def
*cdef
= &io_cold_defs
[req
->opcode
];
1770 const struct io_issue_def
*def
= &io_issue_defs
[req
->opcode
];
1772 /* assign early for deferred execution for non-fixed file */
1773 if (def
->needs_file
&& !(req
->flags
& REQ_F_FIXED_FILE
) && !req
->file
)
1774 req
->file
= io_file_get_normal(req
, req
->cqe
.fd
);
1775 if (!cdef
->prep_async
)
1777 if (WARN_ON_ONCE(req_has_async_data(req
)))
1779 if (!def
->manual_alloc
) {
1780 if (io_alloc_async_data(req
))
1783 return cdef
->prep_async(req
);
1786 static u32
io_get_sequence(struct io_kiocb
*req
)
1788 u32 seq
= req
->ctx
->cached_sq_head
;
1789 struct io_kiocb
*cur
;
1791 /* need original cached_sq_head, but it was increased for each req */
1792 io_for_each_link(cur
, req
)
1797 static __cold
void io_drain_req(struct io_kiocb
*req
)
1798 __must_hold(&ctx
->uring_lock
)
1800 struct io_ring_ctx
*ctx
= req
->ctx
;
1801 struct io_defer_entry
*de
;
1803 u32 seq
= io_get_sequence(req
);
1805 /* Still need defer if there is pending req in defer list. */
1806 spin_lock(&ctx
->completion_lock
);
1807 if (!req_need_defer(req
, seq
) && list_empty_careful(&ctx
->defer_list
)) {
1808 spin_unlock(&ctx
->completion_lock
);
1810 ctx
->drain_active
= false;
1811 io_req_task_queue(req
);
1814 spin_unlock(&ctx
->completion_lock
);
1816 io_prep_async_link(req
);
1817 de
= kmalloc(sizeof(*de
), GFP_KERNEL
);
1820 io_req_defer_failed(req
, ret
);
1824 spin_lock(&ctx
->completion_lock
);
1825 if (!req_need_defer(req
, seq
) && list_empty(&ctx
->defer_list
)) {
1826 spin_unlock(&ctx
->completion_lock
);
1831 trace_io_uring_defer(req
);
1834 list_add_tail(&de
->list
, &ctx
->defer_list
);
1835 spin_unlock(&ctx
->completion_lock
);
1838 static bool io_assign_file(struct io_kiocb
*req
, const struct io_issue_def
*def
,
1839 unsigned int issue_flags
)
1841 if (req
->file
|| !def
->needs_file
)
1844 if (req
->flags
& REQ_F_FIXED_FILE
)
1845 req
->file
= io_file_get_fixed(req
, req
->cqe
.fd
, issue_flags
);
1847 req
->file
= io_file_get_normal(req
, req
->cqe
.fd
);
1852 static int io_issue_sqe(struct io_kiocb
*req
, unsigned int issue_flags
)
1854 const struct io_issue_def
*def
= &io_issue_defs
[req
->opcode
];
1855 const struct cred
*creds
= NULL
;
1858 if (unlikely(!io_assign_file(req
, def
, issue_flags
)))
1861 if (unlikely((req
->flags
& REQ_F_CREDS
) && req
->creds
!= current_cred()))
1862 creds
= override_creds(req
->creds
);
1864 if (!def
->audit_skip
)
1865 audit_uring_entry(req
->opcode
);
1867 ret
= def
->issue(req
, issue_flags
);
1869 if (!def
->audit_skip
)
1870 audit_uring_exit(!ret
, ret
);
1873 revert_creds(creds
);
1875 if (ret
== IOU_OK
) {
1876 if (issue_flags
& IO_URING_F_COMPLETE_DEFER
)
1877 io_req_complete_defer(req
);
1879 io_req_complete_post(req
, issue_flags
);
1880 } else if (ret
!= IOU_ISSUE_SKIP_COMPLETE
)
1883 /* If the op doesn't have a file, we're not polling for it */
1884 if ((req
->ctx
->flags
& IORING_SETUP_IOPOLL
) && def
->iopoll_queue
)
1885 io_iopoll_req_issued(req
, issue_flags
);
1890 int io_poll_issue(struct io_kiocb
*req
, struct io_tw_state
*ts
)
1892 io_tw_lock(req
->ctx
, ts
);
1893 return io_issue_sqe(req
, IO_URING_F_NONBLOCK
|IO_URING_F_MULTISHOT
|
1894 IO_URING_F_COMPLETE_DEFER
);
1897 struct io_wq_work
*io_wq_free_work(struct io_wq_work
*work
)
1899 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
1900 struct io_kiocb
*nxt
= NULL
;
1902 if (req_ref_put_and_test(req
)) {
1903 if (req
->flags
& IO_REQ_LINK_FLAGS
)
1904 nxt
= io_req_find_next(req
);
1907 return nxt
? &nxt
->work
: NULL
;
1910 void io_wq_submit_work(struct io_wq_work
*work
)
1912 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
1913 const struct io_issue_def
*def
= &io_issue_defs
[req
->opcode
];
1914 unsigned int issue_flags
= IO_URING_F_UNLOCKED
| IO_URING_F_IOWQ
;
1915 bool needs_poll
= false;
1916 int ret
= 0, err
= -ECANCELED
;
1918 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1919 if (!(req
->flags
& REQ_F_REFCOUNT
))
1920 __io_req_set_refcount(req
, 2);
1924 io_arm_ltimeout(req
);
1926 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1927 if (work
->flags
& IO_WQ_WORK_CANCEL
) {
1929 io_req_task_queue_fail(req
, err
);
1932 if (!io_assign_file(req
, def
, issue_flags
)) {
1934 work
->flags
|= IO_WQ_WORK_CANCEL
;
1938 if (req
->flags
& REQ_F_FORCE_ASYNC
) {
1939 bool opcode_poll
= def
->pollin
|| def
->pollout
;
1941 if (opcode_poll
&& file_can_poll(req
->file
)) {
1943 issue_flags
|= IO_URING_F_NONBLOCK
;
1948 ret
= io_issue_sqe(req
, issue_flags
);
1952 * We can get EAGAIN for iopolled IO even though we're
1953 * forcing a sync submission from here, since we can't
1954 * wait for request slots on the block side.
1957 if (!(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
1963 if (io_arm_poll_handler(req
, issue_flags
) == IO_APOLL_OK
)
1965 /* aborted or ready, in either case retry blocking */
1967 issue_flags
&= ~IO_URING_F_NONBLOCK
;
1970 /* avoid locking problems by failing it from a clean context */
1972 io_req_task_queue_fail(req
, ret
);
1975 inline struct file
*io_file_get_fixed(struct io_kiocb
*req
, int fd
,
1976 unsigned int issue_flags
)
1978 struct io_ring_ctx
*ctx
= req
->ctx
;
1979 struct io_fixed_file
*slot
;
1980 struct file
*file
= NULL
;
1982 io_ring_submit_lock(ctx
, issue_flags
);
1984 if (unlikely((unsigned int)fd
>= ctx
->nr_user_files
))
1986 fd
= array_index_nospec(fd
, ctx
->nr_user_files
);
1987 slot
= io_fixed_file_slot(&ctx
->file_table
, fd
);
1988 file
= io_slot_file(slot
);
1989 req
->flags
|= io_slot_flags(slot
);
1990 io_req_set_rsrc_node(req
, ctx
, 0);
1992 io_ring_submit_unlock(ctx
, issue_flags
);
1996 struct file
*io_file_get_normal(struct io_kiocb
*req
, int fd
)
1998 struct file
*file
= fget(fd
);
2000 trace_io_uring_file_get(req
, fd
);
2002 /* we don't allow fixed io_uring files */
2003 if (file
&& io_is_uring_fops(file
))
2004 io_req_track_inflight(req
);
2008 static void io_queue_async(struct io_kiocb
*req
, int ret
)
2009 __must_hold(&req
->ctx
->uring_lock
)
2011 struct io_kiocb
*linked_timeout
;
2013 if (ret
!= -EAGAIN
|| (req
->flags
& REQ_F_NOWAIT
)) {
2014 io_req_defer_failed(req
, ret
);
2018 linked_timeout
= io_prep_linked_timeout(req
);
2020 switch (io_arm_poll_handler(req
, 0)) {
2021 case IO_APOLL_READY
:
2022 io_kbuf_recycle(req
, 0);
2023 io_req_task_queue(req
);
2025 case IO_APOLL_ABORTED
:
2026 io_kbuf_recycle(req
, 0);
2027 io_queue_iowq(req
, NULL
);
2034 io_queue_linked_timeout(linked_timeout
);
2037 static inline void io_queue_sqe(struct io_kiocb
*req
)
2038 __must_hold(&req
->ctx
->uring_lock
)
2042 ret
= io_issue_sqe(req
, IO_URING_F_NONBLOCK
|IO_URING_F_COMPLETE_DEFER
);
2045 * We async punt it if the file wasn't marked NOWAIT, or if the file
2046 * doesn't support non-blocking read/write attempts
2049 io_arm_ltimeout(req
);
2051 io_queue_async(req
, ret
);
2054 static void io_queue_sqe_fallback(struct io_kiocb
*req
)
2055 __must_hold(&req
->ctx
->uring_lock
)
2057 if (unlikely(req
->flags
& REQ_F_FAIL
)) {
2059 * We don't submit, fail them all, for that replace hardlinks
2060 * with normal links. Extra REQ_F_LINK is tolerated.
2062 req
->flags
&= ~REQ_F_HARDLINK
;
2063 req
->flags
|= REQ_F_LINK
;
2064 io_req_defer_failed(req
, req
->cqe
.res
);
2066 int ret
= io_req_prep_async(req
);
2068 if (unlikely(ret
)) {
2069 io_req_defer_failed(req
, ret
);
2073 if (unlikely(req
->ctx
->drain_active
))
2076 io_queue_iowq(req
, NULL
);
2081 * Check SQE restrictions (opcode and flags).
2083 * Returns 'true' if SQE is allowed, 'false' otherwise.
2085 static inline bool io_check_restriction(struct io_ring_ctx
*ctx
,
2086 struct io_kiocb
*req
,
2087 unsigned int sqe_flags
)
2089 if (!test_bit(req
->opcode
, ctx
->restrictions
.sqe_op
))
2092 if ((sqe_flags
& ctx
->restrictions
.sqe_flags_required
) !=
2093 ctx
->restrictions
.sqe_flags_required
)
2096 if (sqe_flags
& ~(ctx
->restrictions
.sqe_flags_allowed
|
2097 ctx
->restrictions
.sqe_flags_required
))
2103 static void io_init_req_drain(struct io_kiocb
*req
)
2105 struct io_ring_ctx
*ctx
= req
->ctx
;
2106 struct io_kiocb
*head
= ctx
->submit_state
.link
.head
;
2108 ctx
->drain_active
= true;
2111 * If we need to drain a request in the middle of a link, drain
2112 * the head request and the next request/link after the current
2113 * link. Considering sequential execution of links,
2114 * REQ_F_IO_DRAIN will be maintained for every request of our
2117 head
->flags
|= REQ_F_IO_DRAIN
| REQ_F_FORCE_ASYNC
;
2118 ctx
->drain_next
= true;
2122 static int io_init_req(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
2123 const struct io_uring_sqe
*sqe
)
2124 __must_hold(&ctx
->uring_lock
)
2126 const struct io_issue_def
*def
;
2127 unsigned int sqe_flags
;
2131 /* req is partially pre-initialised, see io_preinit_req() */
2132 req
->opcode
= opcode
= READ_ONCE(sqe
->opcode
);
2133 /* same numerical values with corresponding REQ_F_*, safe to copy */
2134 req
->flags
= sqe_flags
= READ_ONCE(sqe
->flags
);
2135 req
->cqe
.user_data
= READ_ONCE(sqe
->user_data
);
2137 req
->rsrc_node
= NULL
;
2138 req
->task
= current
;
2140 if (unlikely(opcode
>= IORING_OP_LAST
)) {
2144 def
= &io_issue_defs
[opcode
];
2145 if (unlikely(sqe_flags
& ~SQE_COMMON_FLAGS
)) {
2146 /* enforce forwards compatibility on users */
2147 if (sqe_flags
& ~SQE_VALID_FLAGS
)
2149 if (sqe_flags
& IOSQE_BUFFER_SELECT
) {
2150 if (!def
->buffer_select
)
2152 req
->buf_index
= READ_ONCE(sqe
->buf_group
);
2154 if (sqe_flags
& IOSQE_CQE_SKIP_SUCCESS
)
2155 ctx
->drain_disabled
= true;
2156 if (sqe_flags
& IOSQE_IO_DRAIN
) {
2157 if (ctx
->drain_disabled
)
2159 io_init_req_drain(req
);
2162 if (unlikely(ctx
->restricted
|| ctx
->drain_active
|| ctx
->drain_next
)) {
2163 if (ctx
->restricted
&& !io_check_restriction(ctx
, req
, sqe_flags
))
2165 /* knock it to the slow queue path, will be drained there */
2166 if (ctx
->drain_active
)
2167 req
->flags
|= REQ_F_FORCE_ASYNC
;
2168 /* if there is no link, we're at "next" request and need to drain */
2169 if (unlikely(ctx
->drain_next
) && !ctx
->submit_state
.link
.head
) {
2170 ctx
->drain_next
= false;
2171 ctx
->drain_active
= true;
2172 req
->flags
|= REQ_F_IO_DRAIN
| REQ_F_FORCE_ASYNC
;
2176 if (!def
->ioprio
&& sqe
->ioprio
)
2178 if (!def
->iopoll
&& (ctx
->flags
& IORING_SETUP_IOPOLL
))
2181 if (def
->needs_file
) {
2182 struct io_submit_state
*state
= &ctx
->submit_state
;
2184 req
->cqe
.fd
= READ_ONCE(sqe
->fd
);
2187 * Plug now if we have more than 2 IO left after this, and the
2188 * target is potentially a read/write to block based storage.
2190 if (state
->need_plug
&& def
->plug
) {
2191 state
->plug_started
= true;
2192 state
->need_plug
= false;
2193 blk_start_plug_nr_ios(&state
->plug
, state
->submit_nr
);
2197 personality
= READ_ONCE(sqe
->personality
);
2201 req
->creds
= xa_load(&ctx
->personalities
, personality
);
2204 get_cred(req
->creds
);
2205 ret
= security_uring_override_creds(req
->creds
);
2207 put_cred(req
->creds
);
2210 req
->flags
|= REQ_F_CREDS
;
2213 return def
->prep(req
, sqe
);
2216 static __cold
int io_submit_fail_init(const struct io_uring_sqe
*sqe
,
2217 struct io_kiocb
*req
, int ret
)
2219 struct io_ring_ctx
*ctx
= req
->ctx
;
2220 struct io_submit_link
*link
= &ctx
->submit_state
.link
;
2221 struct io_kiocb
*head
= link
->head
;
2223 trace_io_uring_req_failed(sqe
, req
, ret
);
2226 * Avoid breaking links in the middle as it renders links with SQPOLL
2227 * unusable. Instead of failing eagerly, continue assembling the link if
2228 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2229 * should find the flag and handle the rest.
2231 req_fail_link_node(req
, ret
);
2232 if (head
&& !(head
->flags
& REQ_F_FAIL
))
2233 req_fail_link_node(head
, -ECANCELED
);
2235 if (!(req
->flags
& IO_REQ_LINK_FLAGS
)) {
2237 link
->last
->link
= req
;
2241 io_queue_sqe_fallback(req
);
2246 link
->last
->link
= req
;
2253 static inline int io_submit_sqe(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
2254 const struct io_uring_sqe
*sqe
)
2255 __must_hold(&ctx
->uring_lock
)
2257 struct io_submit_link
*link
= &ctx
->submit_state
.link
;
2260 ret
= io_init_req(ctx
, req
, sqe
);
2262 return io_submit_fail_init(sqe
, req
, ret
);
2264 trace_io_uring_submit_req(req
);
2267 * If we already have a head request, queue this one for async
2268 * submittal once the head completes. If we don't have a head but
2269 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2270 * submitted sync once the chain is complete. If none of those
2271 * conditions are true (normal request), then just queue it.
2273 if (unlikely(link
->head
)) {
2274 ret
= io_req_prep_async(req
);
2276 return io_submit_fail_init(sqe
, req
, ret
);
2278 trace_io_uring_link(req
, link
->head
);
2279 link
->last
->link
= req
;
2282 if (req
->flags
& IO_REQ_LINK_FLAGS
)
2284 /* last request of the link, flush it */
2287 if (req
->flags
& (REQ_F_FORCE_ASYNC
| REQ_F_FAIL
))
2290 } else if (unlikely(req
->flags
& (IO_REQ_LINK_FLAGS
|
2291 REQ_F_FORCE_ASYNC
| REQ_F_FAIL
))) {
2292 if (req
->flags
& IO_REQ_LINK_FLAGS
) {
2297 io_queue_sqe_fallback(req
);
2307 * Batched submission is done, ensure local IO is flushed out.
2309 static void io_submit_state_end(struct io_ring_ctx
*ctx
)
2311 struct io_submit_state
*state
= &ctx
->submit_state
;
2313 if (unlikely(state
->link
.head
))
2314 io_queue_sqe_fallback(state
->link
.head
);
2315 /* flush only after queuing links as they can generate completions */
2316 io_submit_flush_completions(ctx
);
2317 if (state
->plug_started
)
2318 blk_finish_plug(&state
->plug
);
2322 * Start submission side cache.
2324 static void io_submit_state_start(struct io_submit_state
*state
,
2325 unsigned int max_ios
)
2327 state
->plug_started
= false;
2328 state
->need_plug
= max_ios
> 2;
2329 state
->submit_nr
= max_ios
;
2330 /* set only head, no need to init link_last in advance */
2331 state
->link
.head
= NULL
;
2334 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
2336 struct io_rings
*rings
= ctx
->rings
;
2339 * Ensure any loads from the SQEs are done at this point,
2340 * since once we write the new head, the application could
2341 * write new data to them.
2343 smp_store_release(&rings
->sq
.head
, ctx
->cached_sq_head
);
2347 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2348 * that is mapped by userspace. This means that care needs to be taken to
2349 * ensure that reads are stable, as we cannot rely on userspace always
2350 * being a good citizen. If members of the sqe are validated and then later
2351 * used, it's important that those reads are done through READ_ONCE() to
2352 * prevent a re-load down the line.
2354 static bool io_get_sqe(struct io_ring_ctx
*ctx
, const struct io_uring_sqe
**sqe
)
2356 unsigned head
, mask
= ctx
->sq_entries
- 1;
2357 unsigned sq_idx
= ctx
->cached_sq_head
++ & mask
;
2360 * The cached sq head (or cq tail) serves two purposes:
2362 * 1) allows us to batch the cost of updating the user visible
2364 * 2) allows the kernel side to track the head on its own, even
2365 * though the application is the one updating it.
2367 head
= READ_ONCE(ctx
->sq_array
[sq_idx
]);
2368 if (likely(head
< ctx
->sq_entries
)) {
2369 /* double index for 128-byte SQEs, twice as long */
2370 if (ctx
->flags
& IORING_SETUP_SQE128
)
2372 *sqe
= &ctx
->sq_sqes
[head
];
2376 /* drop invalid entries */
2378 WRITE_ONCE(ctx
->rings
->sq_dropped
,
2379 READ_ONCE(ctx
->rings
->sq_dropped
) + 1);
2383 int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
)
2384 __must_hold(&ctx
->uring_lock
)
2386 unsigned int entries
= io_sqring_entries(ctx
);
2390 if (unlikely(!entries
))
2392 /* make sure SQ entry isn't read before tail */
2393 ret
= left
= min(nr
, entries
);
2394 io_get_task_refs(left
);
2395 io_submit_state_start(&ctx
->submit_state
, left
);
2398 const struct io_uring_sqe
*sqe
;
2399 struct io_kiocb
*req
;
2401 if (unlikely(!io_alloc_req(ctx
, &req
)))
2403 if (unlikely(!io_get_sqe(ctx
, &sqe
))) {
2404 io_req_add_to_cache(req
, ctx
);
2409 * Continue submitting even for sqe failure if the
2410 * ring was setup with IORING_SETUP_SUBMIT_ALL
2412 if (unlikely(io_submit_sqe(ctx
, req
, sqe
)) &&
2413 !(ctx
->flags
& IORING_SETUP_SUBMIT_ALL
)) {
2419 if (unlikely(left
)) {
2421 /* try again if it submitted nothing and can't allocate a req */
2422 if (!ret
&& io_req_cache_empty(ctx
))
2424 current
->io_uring
->cached_refs
+= left
;
2427 io_submit_state_end(ctx
);
2428 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2429 io_commit_sqring(ctx
);
2433 struct io_wait_queue
{
2434 struct wait_queue_entry wq
;
2435 struct io_ring_ctx
*ctx
;
2437 unsigned nr_timeouts
;
2441 static inline bool io_has_work(struct io_ring_ctx
*ctx
)
2443 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
) ||
2444 !llist_empty(&ctx
->work_llist
);
2447 static inline bool io_should_wake(struct io_wait_queue
*iowq
)
2449 struct io_ring_ctx
*ctx
= iowq
->ctx
;
2450 int dist
= READ_ONCE(ctx
->rings
->cq
.tail
) - (int) iowq
->cq_tail
;
2453 * Wake up if we have enough events, or if a timeout occurred since we
2454 * started waiting. For timeouts, we always want to return to userspace,
2455 * regardless of event count.
2457 return dist
>= 0 || atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
2460 static int io_wake_function(struct wait_queue_entry
*curr
, unsigned int mode
,
2461 int wake_flags
, void *key
)
2463 struct io_wait_queue
*iowq
= container_of(curr
, struct io_wait_queue
, wq
);
2466 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2467 * the task, and the next invocation will do it.
2469 if (io_should_wake(iowq
) || io_has_work(iowq
->ctx
))
2470 return autoremove_wake_function(curr
, mode
, wake_flags
, key
);
2474 int io_run_task_work_sig(struct io_ring_ctx
*ctx
)
2476 if (!llist_empty(&ctx
->work_llist
)) {
2477 __set_current_state(TASK_RUNNING
);
2478 if (io_run_local_work(ctx
) > 0)
2481 if (io_run_task_work() > 0)
2483 if (task_sigpending(current
))
2488 /* when returns >0, the caller should retry */
2489 static inline int io_cqring_wait_schedule(struct io_ring_ctx
*ctx
,
2490 struct io_wait_queue
*iowq
)
2492 if (unlikely(READ_ONCE(ctx
->check_cq
)))
2494 if (unlikely(!llist_empty(&ctx
->work_llist
)))
2496 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL
)))
2498 if (unlikely(task_sigpending(current
)))
2500 if (unlikely(io_should_wake(iowq
)))
2502 if (iowq
->timeout
== KTIME_MAX
)
2504 else if (!schedule_hrtimeout(&iowq
->timeout
, HRTIMER_MODE_ABS
))
2510 * Wait until events become available, if we don't already have some. The
2511 * application must reap them itself, as they reside on the shared cq ring.
2513 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
2514 const sigset_t __user
*sig
, size_t sigsz
,
2515 struct __kernel_timespec __user
*uts
)
2517 struct io_wait_queue iowq
;
2518 struct io_rings
*rings
= ctx
->rings
;
2521 if (!io_allowed_run_tw(ctx
))
2523 if (!llist_empty(&ctx
->work_llist
))
2524 io_run_local_work(ctx
);
2526 io_cqring_overflow_flush(ctx
);
2527 /* if user messes with these they will just get an early return */
2528 if (__io_cqring_events_user(ctx
) >= min_events
)
2532 #ifdef CONFIG_COMPAT
2533 if (in_compat_syscall())
2534 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
2538 ret
= set_user_sigmask(sig
, sigsz
);
2544 init_waitqueue_func_entry(&iowq
.wq
, io_wake_function
);
2545 iowq
.wq
.private = current
;
2546 INIT_LIST_HEAD(&iowq
.wq
.entry
);
2548 iowq
.nr_timeouts
= atomic_read(&ctx
->cq_timeouts
);
2549 iowq
.cq_tail
= READ_ONCE(ctx
->rings
->cq
.head
) + min_events
;
2550 iowq
.timeout
= KTIME_MAX
;
2553 struct timespec64 ts
;
2555 if (get_timespec64(&ts
, uts
))
2557 iowq
.timeout
= ktime_add_ns(timespec64_to_ktime(ts
), ktime_get_ns());
2560 trace_io_uring_cqring_wait(ctx
, min_events
);
2562 unsigned long check_cq
;
2564 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) {
2565 int nr_wait
= (int) iowq
.cq_tail
- READ_ONCE(ctx
->rings
->cq
.tail
);
2567 atomic_set(&ctx
->cq_wait_nr
, nr_wait
);
2568 set_current_state(TASK_INTERRUPTIBLE
);
2570 prepare_to_wait_exclusive(&ctx
->cq_wait
, &iowq
.wq
,
2571 TASK_INTERRUPTIBLE
);
2574 ret
= io_cqring_wait_schedule(ctx
, &iowq
);
2575 __set_current_state(TASK_RUNNING
);
2576 atomic_set(&ctx
->cq_wait_nr
, 0);
2581 * Run task_work after scheduling and before io_should_wake().
2582 * If we got woken because of task_work being processed, run it
2583 * now rather than let the caller do another wait loop.
2586 if (!llist_empty(&ctx
->work_llist
))
2587 io_run_local_work(ctx
);
2589 check_cq
= READ_ONCE(ctx
->check_cq
);
2590 if (unlikely(check_cq
)) {
2591 /* let the caller flush overflows, retry */
2592 if (check_cq
& BIT(IO_CHECK_CQ_OVERFLOW_BIT
))
2593 io_cqring_do_overflow_flush(ctx
);
2594 if (check_cq
& BIT(IO_CHECK_CQ_DROPPED_BIT
)) {
2600 if (io_should_wake(&iowq
)) {
2607 if (!(ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
))
2608 finish_wait(&ctx
->cq_wait
, &iowq
.wq
);
2609 restore_saved_sigmask_unless(ret
== -EINTR
);
2611 return READ_ONCE(rings
->cq
.head
) == READ_ONCE(rings
->cq
.tail
) ? ret
: 0;
2614 static void io_mem_free(void *ptr
)
2621 page
= virt_to_head_page(ptr
);
2622 if (put_page_testzero(page
))
2623 free_compound_page(page
);
2626 static void io_pages_free(struct page
***pages
, int npages
)
2628 struct page
**page_array
;
2633 page_array
= *pages
;
2634 for (i
= 0; i
< npages
; i
++)
2635 unpin_user_page(page_array
[i
]);
2640 static void *__io_uaddr_map(struct page
***pages
, unsigned short *npages
,
2641 unsigned long uaddr
, size_t size
)
2643 struct page
**page_array
;
2644 unsigned int nr_pages
;
2649 if (uaddr
& (PAGE_SIZE
- 1) || !size
)
2650 return ERR_PTR(-EINVAL
);
2652 nr_pages
= (size
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2653 if (nr_pages
> USHRT_MAX
)
2654 return ERR_PTR(-EINVAL
);
2655 page_array
= kvmalloc_array(nr_pages
, sizeof(struct page
*), GFP_KERNEL
);
2657 return ERR_PTR(-ENOMEM
);
2659 ret
= pin_user_pages_fast(uaddr
, nr_pages
, FOLL_WRITE
| FOLL_LONGTERM
,
2661 if (ret
!= nr_pages
) {
2663 io_pages_free(&page_array
, ret
> 0 ? ret
: 0);
2664 return ret
< 0 ? ERR_PTR(ret
) : ERR_PTR(-EFAULT
);
2667 * Should be a single page. If the ring is small enough that we can
2668 * use a normal page, that is fine. If we need multiple pages, then
2669 * userspace should use a huge page. That's the only way to guarantee
2670 * that we get contigious memory, outside of just being lucky or
2671 * (currently) having low memory fragmentation.
2673 if (page_array
[0] != page_array
[ret
- 1])
2675 *pages
= page_array
;
2677 return page_to_virt(page_array
[0]);
2680 static void *io_rings_map(struct io_ring_ctx
*ctx
, unsigned long uaddr
,
2683 return __io_uaddr_map(&ctx
->ring_pages
, &ctx
->n_ring_pages
, uaddr
,
2687 static void *io_sqes_map(struct io_ring_ctx
*ctx
, unsigned long uaddr
,
2690 return __io_uaddr_map(&ctx
->sqe_pages
, &ctx
->n_sqe_pages
, uaddr
,
2694 static void io_rings_free(struct io_ring_ctx
*ctx
)
2696 if (!(ctx
->flags
& IORING_SETUP_NO_MMAP
)) {
2697 io_mem_free(ctx
->rings
);
2698 io_mem_free(ctx
->sq_sqes
);
2700 ctx
->sq_sqes
= NULL
;
2702 io_pages_free(&ctx
->ring_pages
, ctx
->n_ring_pages
);
2703 io_pages_free(&ctx
->sqe_pages
, ctx
->n_sqe_pages
);
2707 static void *io_mem_alloc(size_t size
)
2709 gfp_t gfp
= GFP_KERNEL_ACCOUNT
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
;
2712 ret
= (void *) __get_free_pages(gfp
, get_order(size
));
2715 return ERR_PTR(-ENOMEM
);
2718 static unsigned long rings_size(struct io_ring_ctx
*ctx
, unsigned int sq_entries
,
2719 unsigned int cq_entries
, size_t *sq_offset
)
2721 struct io_rings
*rings
;
2722 size_t off
, sq_array_size
;
2724 off
= struct_size(rings
, cqes
, cq_entries
);
2725 if (off
== SIZE_MAX
)
2727 if (ctx
->flags
& IORING_SETUP_CQE32
) {
2728 if (check_shl_overflow(off
, 1, &off
))
2733 off
= ALIGN(off
, SMP_CACHE_BYTES
);
2741 sq_array_size
= array_size(sizeof(u32
), sq_entries
);
2742 if (sq_array_size
== SIZE_MAX
)
2745 if (check_add_overflow(off
, sq_array_size
, &off
))
2751 static int io_eventfd_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
2752 unsigned int eventfd_async
)
2754 struct io_ev_fd
*ev_fd
;
2755 __s32 __user
*fds
= arg
;
2758 ev_fd
= rcu_dereference_protected(ctx
->io_ev_fd
,
2759 lockdep_is_held(&ctx
->uring_lock
));
2763 if (copy_from_user(&fd
, fds
, sizeof(*fds
)))
2766 ev_fd
= kmalloc(sizeof(*ev_fd
), GFP_KERNEL
);
2770 ev_fd
->cq_ev_fd
= eventfd_ctx_fdget(fd
);
2771 if (IS_ERR(ev_fd
->cq_ev_fd
)) {
2772 int ret
= PTR_ERR(ev_fd
->cq_ev_fd
);
2777 spin_lock(&ctx
->completion_lock
);
2778 ctx
->evfd_last_cq_tail
= ctx
->cached_cq_tail
;
2779 spin_unlock(&ctx
->completion_lock
);
2781 ev_fd
->eventfd_async
= eventfd_async
;
2782 ctx
->has_evfd
= true;
2783 rcu_assign_pointer(ctx
->io_ev_fd
, ev_fd
);
2784 atomic_set(&ev_fd
->refs
, 1);
2785 atomic_set(&ev_fd
->ops
, 0);
2789 static int io_eventfd_unregister(struct io_ring_ctx
*ctx
)
2791 struct io_ev_fd
*ev_fd
;
2793 ev_fd
= rcu_dereference_protected(ctx
->io_ev_fd
,
2794 lockdep_is_held(&ctx
->uring_lock
));
2796 ctx
->has_evfd
= false;
2797 rcu_assign_pointer(ctx
->io_ev_fd
, NULL
);
2798 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT
), &ev_fd
->ops
))
2799 call_rcu(&ev_fd
->rcu
, io_eventfd_ops
);
2806 static void io_req_caches_free(struct io_ring_ctx
*ctx
)
2808 struct io_kiocb
*req
;
2811 mutex_lock(&ctx
->uring_lock
);
2812 io_flush_cached_locked_reqs(ctx
, &ctx
->submit_state
);
2814 while (!io_req_cache_empty(ctx
)) {
2815 req
= io_extract_req(ctx
);
2816 kmem_cache_free(req_cachep
, req
);
2820 percpu_ref_put_many(&ctx
->refs
, nr
);
2821 mutex_unlock(&ctx
->uring_lock
);
2824 static void io_rsrc_node_cache_free(struct io_cache_entry
*entry
)
2826 kfree(container_of(entry
, struct io_rsrc_node
, cache
));
2829 static __cold
void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
2831 io_sq_thread_finish(ctx
);
2832 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2833 if (WARN_ON_ONCE(!list_empty(&ctx
->rsrc_ref_list
)))
2836 mutex_lock(&ctx
->uring_lock
);
2838 __io_sqe_buffers_unregister(ctx
);
2840 __io_sqe_files_unregister(ctx
);
2841 io_cqring_overflow_kill(ctx
);
2842 io_eventfd_unregister(ctx
);
2843 io_alloc_cache_free(&ctx
->apoll_cache
, io_apoll_cache_free
);
2844 io_alloc_cache_free(&ctx
->netmsg_cache
, io_netmsg_cache_free
);
2845 io_destroy_buffers(ctx
);
2846 mutex_unlock(&ctx
->uring_lock
);
2848 put_cred(ctx
->sq_creds
);
2849 if (ctx
->submitter_task
)
2850 put_task_struct(ctx
->submitter_task
);
2852 /* there are no registered resources left, nobody uses it */
2854 io_rsrc_node_destroy(ctx
, ctx
->rsrc_node
);
2856 WARN_ON_ONCE(!list_empty(&ctx
->rsrc_ref_list
));
2858 #if defined(CONFIG_UNIX)
2859 if (ctx
->ring_sock
) {
2860 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
2861 sock_release(ctx
->ring_sock
);
2864 WARN_ON_ONCE(!list_empty(&ctx
->ltimeout_list
));
2866 io_alloc_cache_free(&ctx
->rsrc_node_cache
, io_rsrc_node_cache_free
);
2867 if (ctx
->mm_account
) {
2868 mmdrop(ctx
->mm_account
);
2869 ctx
->mm_account
= NULL
;
2873 percpu_ref_exit(&ctx
->refs
);
2874 free_uid(ctx
->user
);
2875 io_req_caches_free(ctx
);
2877 io_wq_put_hash(ctx
->hash_map
);
2878 kfree(ctx
->cancel_table
.hbs
);
2879 kfree(ctx
->cancel_table_locked
.hbs
);
2880 kfree(ctx
->dummy_ubuf
);
2882 xa_destroy(&ctx
->io_bl_xa
);
2886 static __cold
void io_activate_pollwq_cb(struct callback_head
*cb
)
2888 struct io_ring_ctx
*ctx
= container_of(cb
, struct io_ring_ctx
,
2891 mutex_lock(&ctx
->uring_lock
);
2892 ctx
->poll_activated
= true;
2893 mutex_unlock(&ctx
->uring_lock
);
2896 * Wake ups for some events between start of polling and activation
2897 * might've been lost due to loose synchronisation.
2899 wake_up_all(&ctx
->poll_wq
);
2900 percpu_ref_put(&ctx
->refs
);
2903 static __cold
void io_activate_pollwq(struct io_ring_ctx
*ctx
)
2905 spin_lock(&ctx
->completion_lock
);
2906 /* already activated or in progress */
2907 if (ctx
->poll_activated
|| ctx
->poll_wq_task_work
.func
)
2909 if (WARN_ON_ONCE(!ctx
->task_complete
))
2911 if (!ctx
->submitter_task
)
2914 * with ->submitter_task only the submitter task completes requests, we
2915 * only need to sync with it, which is done by injecting a tw
2917 init_task_work(&ctx
->poll_wq_task_work
, io_activate_pollwq_cb
);
2918 percpu_ref_get(&ctx
->refs
);
2919 if (task_work_add(ctx
->submitter_task
, &ctx
->poll_wq_task_work
, TWA_SIGNAL
))
2920 percpu_ref_put(&ctx
->refs
);
2922 spin_unlock(&ctx
->completion_lock
);
2925 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
2927 struct io_ring_ctx
*ctx
= file
->private_data
;
2930 if (unlikely(!ctx
->poll_activated
))
2931 io_activate_pollwq(ctx
);
2933 poll_wait(file
, &ctx
->poll_wq
, wait
);
2935 * synchronizes with barrier from wq_has_sleeper call in
2939 if (!io_sqring_full(ctx
))
2940 mask
|= EPOLLOUT
| EPOLLWRNORM
;
2943 * Don't flush cqring overflow list here, just do a simple check.
2944 * Otherwise there could possible be ABBA deadlock:
2947 * lock(&ctx->uring_lock);
2949 * lock(&ctx->uring_lock);
2952 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2953 * pushes them to do the flush.
2956 if (__io_cqring_events_user(ctx
) || io_has_work(ctx
))
2957 mask
|= EPOLLIN
| EPOLLRDNORM
;
2962 static int io_unregister_personality(struct io_ring_ctx
*ctx
, unsigned id
)
2964 const struct cred
*creds
;
2966 creds
= xa_erase(&ctx
->personalities
, id
);
2975 struct io_tctx_exit
{
2976 struct callback_head task_work
;
2977 struct completion completion
;
2978 struct io_ring_ctx
*ctx
;
2981 static __cold
void io_tctx_exit_cb(struct callback_head
*cb
)
2983 struct io_uring_task
*tctx
= current
->io_uring
;
2984 struct io_tctx_exit
*work
;
2986 work
= container_of(cb
, struct io_tctx_exit
, task_work
);
2988 * When @in_cancel, we're in cancellation and it's racy to remove the
2989 * node. It'll be removed by the end of cancellation, just ignore it.
2990 * tctx can be NULL if the queueing of this task_work raced with
2991 * work cancelation off the exec path.
2993 if (tctx
&& !atomic_read(&tctx
->in_cancel
))
2994 io_uring_del_tctx_node((unsigned long)work
->ctx
);
2995 complete(&work
->completion
);
2998 static __cold
bool io_cancel_ctx_cb(struct io_wq_work
*work
, void *data
)
3000 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
3002 return req
->ctx
== data
;
3005 static __cold
void io_ring_exit_work(struct work_struct
*work
)
3007 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
, exit_work
);
3008 unsigned long timeout
= jiffies
+ HZ
* 60 * 5;
3009 unsigned long interval
= HZ
/ 20;
3010 struct io_tctx_exit exit
;
3011 struct io_tctx_node
*node
;
3015 * If we're doing polled IO and end up having requests being
3016 * submitted async (out-of-line), then completions can come in while
3017 * we're waiting for refs to drop. We need to reap these manually,
3018 * as nobody else will be looking for them.
3021 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
)) {
3022 mutex_lock(&ctx
->uring_lock
);
3023 io_cqring_overflow_kill(ctx
);
3024 mutex_unlock(&ctx
->uring_lock
);
3027 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
)
3028 io_move_task_work_from_local(ctx
);
3030 while (io_uring_try_cancel_requests(ctx
, NULL
, true))
3034 struct io_sq_data
*sqd
= ctx
->sq_data
;
3035 struct task_struct
*tsk
;
3037 io_sq_thread_park(sqd
);
3039 if (tsk
&& tsk
->io_uring
&& tsk
->io_uring
->io_wq
)
3040 io_wq_cancel_cb(tsk
->io_uring
->io_wq
,
3041 io_cancel_ctx_cb
, ctx
, true);
3042 io_sq_thread_unpark(sqd
);
3045 io_req_caches_free(ctx
);
3047 if (WARN_ON_ONCE(time_after(jiffies
, timeout
))) {
3048 /* there is little hope left, don't run it too often */
3052 * This is really an uninterruptible wait, as it has to be
3053 * complete. But it's also run from a kworker, which doesn't
3054 * take signals, so it's fine to make it interruptible. This
3055 * avoids scenarios where we knowingly can wait much longer
3056 * on completions, for example if someone does a SIGSTOP on
3057 * a task that needs to finish task_work to make this loop
3058 * complete. That's a synthetic situation that should not
3059 * cause a stuck task backtrace, and hence a potential panic
3060 * on stuck tasks if that is enabled.
3062 } while (!wait_for_completion_interruptible_timeout(&ctx
->ref_comp
, interval
));
3064 init_completion(&exit
.completion
);
3065 init_task_work(&exit
.task_work
, io_tctx_exit_cb
);
3068 * Some may use context even when all refs and requests have been put,
3069 * and they are free to do so while still holding uring_lock or
3070 * completion_lock, see io_req_task_submit(). Apart from other work,
3071 * this lock/unlock section also waits them to finish.
3073 mutex_lock(&ctx
->uring_lock
);
3074 while (!list_empty(&ctx
->tctx_list
)) {
3075 WARN_ON_ONCE(time_after(jiffies
, timeout
));
3077 node
= list_first_entry(&ctx
->tctx_list
, struct io_tctx_node
,
3079 /* don't spin on a single task if cancellation failed */
3080 list_rotate_left(&ctx
->tctx_list
);
3081 ret
= task_work_add(node
->task
, &exit
.task_work
, TWA_SIGNAL
);
3082 if (WARN_ON_ONCE(ret
))
3085 mutex_unlock(&ctx
->uring_lock
);
3087 * See comment above for
3088 * wait_for_completion_interruptible_timeout() on why this
3089 * wait is marked as interruptible.
3091 wait_for_completion_interruptible(&exit
.completion
);
3092 mutex_lock(&ctx
->uring_lock
);
3094 mutex_unlock(&ctx
->uring_lock
);
3095 spin_lock(&ctx
->completion_lock
);
3096 spin_unlock(&ctx
->completion_lock
);
3098 /* pairs with RCU read section in io_req_local_work_add() */
3099 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
)
3102 io_ring_ctx_free(ctx
);
3105 static __cold
void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
3107 unsigned long index
;
3108 struct creds
*creds
;
3110 mutex_lock(&ctx
->uring_lock
);
3111 percpu_ref_kill(&ctx
->refs
);
3112 xa_for_each(&ctx
->personalities
, index
, creds
)
3113 io_unregister_personality(ctx
, index
);
3115 io_poll_remove_all(ctx
, NULL
, true);
3116 mutex_unlock(&ctx
->uring_lock
);
3119 * If we failed setting up the ctx, we might not have any rings
3120 * and therefore did not submit any requests
3123 io_kill_timeouts(ctx
, NULL
, true);
3125 flush_delayed_work(&ctx
->fallback_work
);
3127 INIT_WORK(&ctx
->exit_work
, io_ring_exit_work
);
3129 * Use system_unbound_wq to avoid spawning tons of event kworkers
3130 * if we're exiting a ton of rings at the same time. It just adds
3131 * noise and overhead, there's no discernable change in runtime
3132 * over using system_wq.
3134 queue_work(system_unbound_wq
, &ctx
->exit_work
);
3137 static int io_uring_release(struct inode
*inode
, struct file
*file
)
3139 struct io_ring_ctx
*ctx
= file
->private_data
;
3141 file
->private_data
= NULL
;
3142 io_ring_ctx_wait_and_kill(ctx
);
3146 struct io_task_cancel
{
3147 struct task_struct
*task
;
3151 static bool io_cancel_task_cb(struct io_wq_work
*work
, void *data
)
3153 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
3154 struct io_task_cancel
*cancel
= data
;
3156 return io_match_task_safe(req
, cancel
->task
, cancel
->all
);
3159 static __cold
bool io_cancel_defer_files(struct io_ring_ctx
*ctx
,
3160 struct task_struct
*task
,
3163 struct io_defer_entry
*de
;
3166 spin_lock(&ctx
->completion_lock
);
3167 list_for_each_entry_reverse(de
, &ctx
->defer_list
, list
) {
3168 if (io_match_task_safe(de
->req
, task
, cancel_all
)) {
3169 list_cut_position(&list
, &ctx
->defer_list
, &de
->list
);
3173 spin_unlock(&ctx
->completion_lock
);
3174 if (list_empty(&list
))
3177 while (!list_empty(&list
)) {
3178 de
= list_first_entry(&list
, struct io_defer_entry
, list
);
3179 list_del_init(&de
->list
);
3180 io_req_task_queue_fail(de
->req
, -ECANCELED
);
3186 static __cold
bool io_uring_try_cancel_iowq(struct io_ring_ctx
*ctx
)
3188 struct io_tctx_node
*node
;
3189 enum io_wq_cancel cret
;
3192 mutex_lock(&ctx
->uring_lock
);
3193 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
3194 struct io_uring_task
*tctx
= node
->task
->io_uring
;
3197 * io_wq will stay alive while we hold uring_lock, because it's
3198 * killed after ctx nodes, which requires to take the lock.
3200 if (!tctx
|| !tctx
->io_wq
)
3202 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_ctx_cb
, ctx
, true);
3203 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
3205 mutex_unlock(&ctx
->uring_lock
);
3210 static __cold
bool io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
3211 struct task_struct
*task
,
3214 struct io_task_cancel cancel
= { .task
= task
, .all
= cancel_all
, };
3215 struct io_uring_task
*tctx
= task
? task
->io_uring
: NULL
;
3216 enum io_wq_cancel cret
;
3219 /* set it so io_req_local_work_add() would wake us up */
3220 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) {
3221 atomic_set(&ctx
->cq_wait_nr
, 1);
3225 /* failed during ring init, it couldn't have issued any requests */
3230 ret
|= io_uring_try_cancel_iowq(ctx
);
3231 } else if (tctx
&& tctx
->io_wq
) {
3233 * Cancels requests of all rings, not only @ctx, but
3234 * it's fine as the task is in exit/exec.
3236 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_task_cb
,
3238 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
3241 /* SQPOLL thread does its own polling */
3242 if ((!(ctx
->flags
& IORING_SETUP_SQPOLL
) && cancel_all
) ||
3243 (ctx
->sq_data
&& ctx
->sq_data
->thread
== current
)) {
3244 while (!wq_list_empty(&ctx
->iopoll_list
)) {
3245 io_iopoll_try_reap_events(ctx
);
3251 if ((ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) &&
3252 io_allowed_defer_tw_run(ctx
))
3253 ret
|= io_run_local_work(ctx
) > 0;
3254 ret
|= io_cancel_defer_files(ctx
, task
, cancel_all
);
3255 mutex_lock(&ctx
->uring_lock
);
3256 ret
|= io_poll_remove_all(ctx
, task
, cancel_all
);
3257 mutex_unlock(&ctx
->uring_lock
);
3258 ret
|= io_kill_timeouts(ctx
, task
, cancel_all
);
3260 ret
|= io_run_task_work() > 0;
3264 static s64
tctx_inflight(struct io_uring_task
*tctx
, bool tracked
)
3267 return atomic_read(&tctx
->inflight_tracked
);
3268 return percpu_counter_sum(&tctx
->inflight
);
3272 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3273 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3275 __cold
void io_uring_cancel_generic(bool cancel_all
, struct io_sq_data
*sqd
)
3277 struct io_uring_task
*tctx
= current
->io_uring
;
3278 struct io_ring_ctx
*ctx
;
3279 struct io_tctx_node
*node
;
3280 unsigned long index
;
3284 WARN_ON_ONCE(sqd
&& sqd
->thread
!= current
);
3286 if (!current
->io_uring
)
3289 io_wq_exit_start(tctx
->io_wq
);
3291 atomic_inc(&tctx
->in_cancel
);
3295 io_uring_drop_tctx_refs(current
);
3296 /* read completions before cancelations */
3297 inflight
= tctx_inflight(tctx
, !cancel_all
);
3302 xa_for_each(&tctx
->xa
, index
, node
) {
3303 /* sqpoll task will cancel all its requests */
3304 if (node
->ctx
->sq_data
)
3306 loop
|= io_uring_try_cancel_requests(node
->ctx
,
3307 current
, cancel_all
);
3310 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
3311 loop
|= io_uring_try_cancel_requests(ctx
,
3321 prepare_to_wait(&tctx
->wait
, &wait
, TASK_INTERRUPTIBLE
);
3323 io_uring_drop_tctx_refs(current
);
3324 xa_for_each(&tctx
->xa
, index
, node
) {
3325 if (!llist_empty(&node
->ctx
->work_llist
)) {
3326 WARN_ON_ONCE(node
->ctx
->submitter_task
&&
3327 node
->ctx
->submitter_task
!= current
);
3332 * If we've seen completions, retry without waiting. This
3333 * avoids a race where a completion comes in before we did
3334 * prepare_to_wait().
3336 if (inflight
== tctx_inflight(tctx
, !cancel_all
))
3339 finish_wait(&tctx
->wait
, &wait
);
3342 io_uring_clean_tctx(tctx
);
3345 * We shouldn't run task_works after cancel, so just leave
3346 * ->in_cancel set for normal exit.
3348 atomic_dec(&tctx
->in_cancel
);
3349 /* for exec all current's requests should be gone, kill tctx */
3350 __io_uring_free(current
);
3354 void __io_uring_cancel(bool cancel_all
)
3356 io_uring_cancel_generic(cancel_all
, NULL
);
3359 static void *io_uring_validate_mmap_request(struct file
*file
,
3360 loff_t pgoff
, size_t sz
)
3362 struct io_ring_ctx
*ctx
= file
->private_data
;
3363 loff_t offset
= pgoff
<< PAGE_SHIFT
;
3367 /* Don't allow mmap if the ring was setup without it */
3368 if (ctx
->flags
& IORING_SETUP_NO_MMAP
)
3369 return ERR_PTR(-EINVAL
);
3371 switch (offset
& IORING_OFF_MMAP_MASK
) {
3372 case IORING_OFF_SQ_RING
:
3373 case IORING_OFF_CQ_RING
:
3376 case IORING_OFF_SQES
:
3379 case IORING_OFF_PBUF_RING
: {
3382 bgid
= (offset
& ~IORING_OFF_MMAP_MASK
) >> IORING_OFF_PBUF_SHIFT
;
3383 mutex_lock(&ctx
->uring_lock
);
3384 ptr
= io_pbuf_get_address(ctx
, bgid
);
3385 mutex_unlock(&ctx
->uring_lock
);
3387 return ERR_PTR(-EINVAL
);
3391 return ERR_PTR(-EINVAL
);
3394 page
= virt_to_head_page(ptr
);
3395 if (sz
> page_size(page
))
3396 return ERR_PTR(-EINVAL
);
3403 static __cold
int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3405 size_t sz
= vma
->vm_end
- vma
->vm_start
;
3409 ptr
= io_uring_validate_mmap_request(file
, vma
->vm_pgoff
, sz
);
3411 return PTR_ERR(ptr
);
3413 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
3414 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
3417 static unsigned long io_uring_mmu_get_unmapped_area(struct file
*filp
,
3418 unsigned long addr
, unsigned long len
,
3419 unsigned long pgoff
, unsigned long flags
)
3421 const unsigned long mmap_end
= arch_get_mmap_end(addr
, len
, flags
);
3422 struct vm_unmapped_area_info info
;
3426 * Do not allow to map to user-provided address to avoid breaking the
3427 * aliasing rules. Userspace is not able to guess the offset address of
3428 * kernel kmalloc()ed memory area.
3433 ptr
= io_uring_validate_mmap_request(filp
, pgoff
, len
);
3437 info
.flags
= VM_UNMAPPED_AREA_TOPDOWN
;
3439 info
.low_limit
= max(PAGE_SIZE
, mmap_min_addr
);
3440 info
.high_limit
= arch_get_mmap_base(addr
, current
->mm
->mmap_base
);
3442 info
.align_mask
= PAGE_MASK
& (SHM_COLOUR
- 1UL);
3444 info
.align_mask
= PAGE_MASK
& (SHMLBA
- 1UL);
3446 info
.align_offset
= (unsigned long) ptr
;
3449 * A failed mmap() very likely causes application failure,
3450 * so fall back to the bottom-up function here. This scenario
3451 * can happen with large stack limits and large mmap()
3454 addr
= vm_unmapped_area(&info
);
3455 if (offset_in_page(addr
)) {
3457 info
.low_limit
= TASK_UNMAPPED_BASE
;
3458 info
.high_limit
= mmap_end
;
3459 addr
= vm_unmapped_area(&info
);
3465 #else /* !CONFIG_MMU */
3467 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3469 return is_nommu_shared_mapping(vma
->vm_flags
) ? 0 : -EINVAL
;
3472 static unsigned int io_uring_nommu_mmap_capabilities(struct file
*file
)
3474 return NOMMU_MAP_DIRECT
| NOMMU_MAP_READ
| NOMMU_MAP_WRITE
;
3477 static unsigned long io_uring_nommu_get_unmapped_area(struct file
*file
,
3478 unsigned long addr
, unsigned long len
,
3479 unsigned long pgoff
, unsigned long flags
)
3483 ptr
= io_uring_validate_mmap_request(file
, pgoff
, len
);
3485 return PTR_ERR(ptr
);
3487 return (unsigned long) ptr
;
3490 #endif /* !CONFIG_MMU */
3492 static int io_validate_ext_arg(unsigned flags
, const void __user
*argp
, size_t argsz
)
3494 if (flags
& IORING_ENTER_EXT_ARG
) {
3495 struct io_uring_getevents_arg arg
;
3497 if (argsz
!= sizeof(arg
))
3499 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
3505 static int io_get_ext_arg(unsigned flags
, const void __user
*argp
, size_t *argsz
,
3506 struct __kernel_timespec __user
**ts
,
3507 const sigset_t __user
**sig
)
3509 struct io_uring_getevents_arg arg
;
3512 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3513 * is just a pointer to the sigset_t.
3515 if (!(flags
& IORING_ENTER_EXT_ARG
)) {
3516 *sig
= (const sigset_t __user
*) argp
;
3522 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3523 * timespec and sigset_t pointers if good.
3525 if (*argsz
!= sizeof(arg
))
3527 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
3531 *sig
= u64_to_user_ptr(arg
.sigmask
);
3532 *argsz
= arg
.sigmask_sz
;
3533 *ts
= u64_to_user_ptr(arg
.ts
);
3537 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
3538 u32
, min_complete
, u32
, flags
, const void __user
*, argp
,
3541 struct io_ring_ctx
*ctx
;
3545 if (unlikely(flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
|
3546 IORING_ENTER_SQ_WAIT
| IORING_ENTER_EXT_ARG
|
3547 IORING_ENTER_REGISTERED_RING
)))
3551 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3552 * need only dereference our task private array to find it.
3554 if (flags
& IORING_ENTER_REGISTERED_RING
) {
3555 struct io_uring_task
*tctx
= current
->io_uring
;
3557 if (unlikely(!tctx
|| fd
>= IO_RINGFD_REG_MAX
))
3559 fd
= array_index_nospec(fd
, IO_RINGFD_REG_MAX
);
3560 f
.file
= tctx
->registered_rings
[fd
];
3562 if (unlikely(!f
.file
))
3566 if (unlikely(!f
.file
))
3569 if (unlikely(!io_is_uring_fops(f
.file
)))
3573 ctx
= f
.file
->private_data
;
3575 if (unlikely(ctx
->flags
& IORING_SETUP_R_DISABLED
))
3579 * For SQ polling, the thread will do all submissions and completions.
3580 * Just return the requested submit count, and wake the thread if
3584 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
3585 io_cqring_overflow_flush(ctx
);
3587 if (unlikely(ctx
->sq_data
->thread
== NULL
)) {
3591 if (flags
& IORING_ENTER_SQ_WAKEUP
)
3592 wake_up(&ctx
->sq_data
->wait
);
3593 if (flags
& IORING_ENTER_SQ_WAIT
)
3594 io_sqpoll_wait_sq(ctx
);
3597 } else if (to_submit
) {
3598 ret
= io_uring_add_tctx_node(ctx
);
3602 mutex_lock(&ctx
->uring_lock
);
3603 ret
= io_submit_sqes(ctx
, to_submit
);
3604 if (ret
!= to_submit
) {
3605 mutex_unlock(&ctx
->uring_lock
);
3608 if (flags
& IORING_ENTER_GETEVENTS
) {
3609 if (ctx
->syscall_iopoll
)
3612 * Ignore errors, we'll soon call io_cqring_wait() and
3613 * it should handle ownership problems if any.
3615 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
)
3616 (void)io_run_local_work_locked(ctx
);
3618 mutex_unlock(&ctx
->uring_lock
);
3621 if (flags
& IORING_ENTER_GETEVENTS
) {
3624 if (ctx
->syscall_iopoll
) {
3626 * We disallow the app entering submit/complete with
3627 * polling, but we still need to lock the ring to
3628 * prevent racing with polled issue that got punted to
3631 mutex_lock(&ctx
->uring_lock
);
3633 ret2
= io_validate_ext_arg(flags
, argp
, argsz
);
3634 if (likely(!ret2
)) {
3635 min_complete
= min(min_complete
,
3637 ret2
= io_iopoll_check(ctx
, min_complete
);
3639 mutex_unlock(&ctx
->uring_lock
);
3641 const sigset_t __user
*sig
;
3642 struct __kernel_timespec __user
*ts
;
3644 ret2
= io_get_ext_arg(flags
, argp
, &argsz
, &ts
, &sig
);
3645 if (likely(!ret2
)) {
3646 min_complete
= min(min_complete
,
3648 ret2
= io_cqring_wait(ctx
, min_complete
, sig
,
3657 * EBADR indicates that one or more CQE were dropped.
3658 * Once the user has been informed we can clear the bit
3659 * as they are obviously ok with those drops.
3661 if (unlikely(ret2
== -EBADR
))
3662 clear_bit(IO_CHECK_CQ_DROPPED_BIT
,
3671 static const struct file_operations io_uring_fops
= {
3672 .release
= io_uring_release
,
3673 .mmap
= io_uring_mmap
,
3675 .get_unmapped_area
= io_uring_nommu_get_unmapped_area
,
3676 .mmap_capabilities
= io_uring_nommu_mmap_capabilities
,
3678 .get_unmapped_area
= io_uring_mmu_get_unmapped_area
,
3680 .poll
= io_uring_poll
,
3681 #ifdef CONFIG_PROC_FS
3682 .show_fdinfo
= io_uring_show_fdinfo
,
3686 bool io_is_uring_fops(struct file
*file
)
3688 return file
->f_op
== &io_uring_fops
;
3691 static __cold
int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
3692 struct io_uring_params
*p
)
3694 struct io_rings
*rings
;
3695 size_t size
, sq_array_offset
;
3698 /* make sure these are sane, as we already accounted them */
3699 ctx
->sq_entries
= p
->sq_entries
;
3700 ctx
->cq_entries
= p
->cq_entries
;
3702 size
= rings_size(ctx
, p
->sq_entries
, p
->cq_entries
, &sq_array_offset
);
3703 if (size
== SIZE_MAX
)
3706 if (!(ctx
->flags
& IORING_SETUP_NO_MMAP
))
3707 rings
= io_mem_alloc(size
);
3709 rings
= io_rings_map(ctx
, p
->cq_off
.user_addr
, size
);
3712 return PTR_ERR(rings
);
3715 ctx
->sq_array
= (u32
*)((char *)rings
+ sq_array_offset
);
3716 rings
->sq_ring_mask
= p
->sq_entries
- 1;
3717 rings
->cq_ring_mask
= p
->cq_entries
- 1;
3718 rings
->sq_ring_entries
= p
->sq_entries
;
3719 rings
->cq_ring_entries
= p
->cq_entries
;
3721 if (p
->flags
& IORING_SETUP_SQE128
)
3722 size
= array_size(2 * sizeof(struct io_uring_sqe
), p
->sq_entries
);
3724 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
3725 if (size
== SIZE_MAX
) {
3730 if (!(ctx
->flags
& IORING_SETUP_NO_MMAP
))
3731 ptr
= io_mem_alloc(size
);
3733 ptr
= io_sqes_map(ctx
, p
->sq_off
.user_addr
, size
);
3737 return PTR_ERR(ptr
);
3744 static int io_uring_install_fd(struct file
*file
)
3748 fd
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
3751 fd_install(fd
, file
);
3756 * Allocate an anonymous fd, this is what constitutes the application
3757 * visible backing of an io_uring instance. The application mmaps this
3758 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3759 * we have to tie this fd to a socket for file garbage collection purposes.
3761 static struct file
*io_uring_get_file(struct io_ring_ctx
*ctx
)
3764 #if defined(CONFIG_UNIX)
3767 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
3770 return ERR_PTR(ret
);
3773 file
= anon_inode_getfile_secure("[io_uring]", &io_uring_fops
, ctx
,
3774 O_RDWR
| O_CLOEXEC
, NULL
);
3775 #if defined(CONFIG_UNIX)
3777 sock_release(ctx
->ring_sock
);
3778 ctx
->ring_sock
= NULL
;
3780 ctx
->ring_sock
->file
= file
;
3786 static __cold
int io_uring_create(unsigned entries
, struct io_uring_params
*p
,
3787 struct io_uring_params __user
*params
)
3789 struct io_ring_ctx
*ctx
;
3790 struct io_uring_task
*tctx
;
3796 if (entries
> IORING_MAX_ENTRIES
) {
3797 if (!(p
->flags
& IORING_SETUP_CLAMP
))
3799 entries
= IORING_MAX_ENTRIES
;
3802 if ((p
->flags
& IORING_SETUP_REGISTERED_FD_ONLY
)
3803 && !(p
->flags
& IORING_SETUP_NO_MMAP
))
3807 * Use twice as many entries for the CQ ring. It's possible for the
3808 * application to drive a higher depth than the size of the SQ ring,
3809 * since the sqes are only used at submission time. This allows for
3810 * some flexibility in overcommitting a bit. If the application has
3811 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3812 * of CQ ring entries manually.
3814 p
->sq_entries
= roundup_pow_of_two(entries
);
3815 if (p
->flags
& IORING_SETUP_CQSIZE
) {
3817 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3818 * to a power-of-two, if it isn't already. We do NOT impose
3819 * any cq vs sq ring sizing.
3823 if (p
->cq_entries
> IORING_MAX_CQ_ENTRIES
) {
3824 if (!(p
->flags
& IORING_SETUP_CLAMP
))
3826 p
->cq_entries
= IORING_MAX_CQ_ENTRIES
;
3828 p
->cq_entries
= roundup_pow_of_two(p
->cq_entries
);
3829 if (p
->cq_entries
< p
->sq_entries
)
3832 p
->cq_entries
= 2 * p
->sq_entries
;
3835 ctx
= io_ring_ctx_alloc(p
);
3839 if ((ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) &&
3840 !(ctx
->flags
& IORING_SETUP_IOPOLL
) &&
3841 !(ctx
->flags
& IORING_SETUP_SQPOLL
))
3842 ctx
->task_complete
= true;
3845 * lazy poll_wq activation relies on ->task_complete for synchronisation
3846 * purposes, see io_activate_pollwq()
3848 if (!ctx
->task_complete
)
3849 ctx
->poll_activated
= true;
3852 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3853 * space applications don't need to do io completion events
3854 * polling again, they can rely on io_sq_thread to do polling
3855 * work, which can reduce cpu usage and uring_lock contention.
3857 if (ctx
->flags
& IORING_SETUP_IOPOLL
&&
3858 !(ctx
->flags
& IORING_SETUP_SQPOLL
))
3859 ctx
->syscall_iopoll
= 1;
3861 ctx
->compat
= in_compat_syscall();
3862 if (!capable(CAP_IPC_LOCK
))
3863 ctx
->user
= get_uid(current_user());
3866 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3867 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3870 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
3871 /* IPI related flags don't make sense with SQPOLL */
3872 if (ctx
->flags
& (IORING_SETUP_COOP_TASKRUN
|
3873 IORING_SETUP_TASKRUN_FLAG
|
3874 IORING_SETUP_DEFER_TASKRUN
))
3876 ctx
->notify_method
= TWA_SIGNAL_NO_IPI
;
3877 } else if (ctx
->flags
& IORING_SETUP_COOP_TASKRUN
) {
3878 ctx
->notify_method
= TWA_SIGNAL_NO_IPI
;
3880 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
&&
3881 !(ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
))
3883 ctx
->notify_method
= TWA_SIGNAL
;
3887 * For DEFER_TASKRUN we require the completion task to be the same as the
3888 * submission task. This implies that there is only one submitter, so enforce
3891 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
&&
3892 !(ctx
->flags
& IORING_SETUP_SINGLE_ISSUER
)) {
3897 * This is just grabbed for accounting purposes. When a process exits,
3898 * the mm is exited and dropped before the files, hence we need to hang
3899 * on to this mm purely for the purposes of being able to unaccount
3900 * memory (locked/pinned vm). It's not used for anything else.
3902 mmgrab(current
->mm
);
3903 ctx
->mm_account
= current
->mm
;
3905 ret
= io_allocate_scq_urings(ctx
, p
);
3909 ret
= io_sq_offload_create(ctx
, p
);
3913 ret
= io_rsrc_init(ctx
);
3917 p
->sq_off
.head
= offsetof(struct io_rings
, sq
.head
);
3918 p
->sq_off
.tail
= offsetof(struct io_rings
, sq
.tail
);
3919 p
->sq_off
.ring_mask
= offsetof(struct io_rings
, sq_ring_mask
);
3920 p
->sq_off
.ring_entries
= offsetof(struct io_rings
, sq_ring_entries
);
3921 p
->sq_off
.flags
= offsetof(struct io_rings
, sq_flags
);
3922 p
->sq_off
.dropped
= offsetof(struct io_rings
, sq_dropped
);
3923 p
->sq_off
.array
= (char *)ctx
->sq_array
- (char *)ctx
->rings
;
3924 p
->sq_off
.resv1
= 0;
3925 if (!(ctx
->flags
& IORING_SETUP_NO_MMAP
))
3926 p
->sq_off
.user_addr
= 0;
3928 p
->cq_off
.head
= offsetof(struct io_rings
, cq
.head
);
3929 p
->cq_off
.tail
= offsetof(struct io_rings
, cq
.tail
);
3930 p
->cq_off
.ring_mask
= offsetof(struct io_rings
, cq_ring_mask
);
3931 p
->cq_off
.ring_entries
= offsetof(struct io_rings
, cq_ring_entries
);
3932 p
->cq_off
.overflow
= offsetof(struct io_rings
, cq_overflow
);
3933 p
->cq_off
.cqes
= offsetof(struct io_rings
, cqes
);
3934 p
->cq_off
.flags
= offsetof(struct io_rings
, cq_flags
);
3935 p
->cq_off
.resv1
= 0;
3936 if (!(ctx
->flags
& IORING_SETUP_NO_MMAP
))
3937 p
->cq_off
.user_addr
= 0;
3939 p
->features
= IORING_FEAT_SINGLE_MMAP
| IORING_FEAT_NODROP
|
3940 IORING_FEAT_SUBMIT_STABLE
| IORING_FEAT_RW_CUR_POS
|
3941 IORING_FEAT_CUR_PERSONALITY
| IORING_FEAT_FAST_POLL
|
3942 IORING_FEAT_POLL_32BITS
| IORING_FEAT_SQPOLL_NONFIXED
|
3943 IORING_FEAT_EXT_ARG
| IORING_FEAT_NATIVE_WORKERS
|
3944 IORING_FEAT_RSRC_TAGS
| IORING_FEAT_CQE_SKIP
|
3945 IORING_FEAT_LINKED_FILE
| IORING_FEAT_REG_REG_RING
;
3947 if (copy_to_user(params
, p
, sizeof(*p
))) {
3952 if (ctx
->flags
& IORING_SETUP_SINGLE_ISSUER
3953 && !(ctx
->flags
& IORING_SETUP_R_DISABLED
))
3954 WRITE_ONCE(ctx
->submitter_task
, get_task_struct(current
));
3956 file
= io_uring_get_file(ctx
);
3958 ret
= PTR_ERR(file
);
3962 ret
= __io_uring_add_tctx_node(ctx
);
3965 tctx
= current
->io_uring
;
3968 * Install ring fd as the very last thing, so we don't risk someone
3969 * having closed it before we finish setup
3971 if (p
->flags
& IORING_SETUP_REGISTERED_FD_ONLY
)
3972 ret
= io_ring_add_registered_file(tctx
, file
, 0, IO_RINGFD_REG_MAX
);
3974 ret
= io_uring_install_fd(file
);
3978 trace_io_uring_create(ret
, ctx
, p
->sq_entries
, p
->cq_entries
, p
->flags
);
3981 io_ring_ctx_wait_and_kill(ctx
);
3989 * Sets up an aio uring context, and returns the fd. Applications asks for a
3990 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3991 * params structure passed in.
3993 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
3995 struct io_uring_params p
;
3998 if (copy_from_user(&p
, params
, sizeof(p
)))
4000 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
4005 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
4006 IORING_SETUP_SQ_AFF
| IORING_SETUP_CQSIZE
|
4007 IORING_SETUP_CLAMP
| IORING_SETUP_ATTACH_WQ
|
4008 IORING_SETUP_R_DISABLED
| IORING_SETUP_SUBMIT_ALL
|
4009 IORING_SETUP_COOP_TASKRUN
| IORING_SETUP_TASKRUN_FLAG
|
4010 IORING_SETUP_SQE128
| IORING_SETUP_CQE32
|
4011 IORING_SETUP_SINGLE_ISSUER
| IORING_SETUP_DEFER_TASKRUN
|
4012 IORING_SETUP_NO_MMAP
| IORING_SETUP_REGISTERED_FD_ONLY
))
4015 return io_uring_create(entries
, &p
, params
);
4018 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
4019 struct io_uring_params __user
*, params
)
4021 return io_uring_setup(entries
, params
);
4024 static __cold
int io_probe(struct io_ring_ctx
*ctx
, void __user
*arg
,
4027 struct io_uring_probe
*p
;
4031 size
= struct_size(p
, ops
, nr_args
);
4032 if (size
== SIZE_MAX
)
4034 p
= kzalloc(size
, GFP_KERNEL
);
4039 if (copy_from_user(p
, arg
, size
))
4042 if (memchr_inv(p
, 0, size
))
4045 p
->last_op
= IORING_OP_LAST
- 1;
4046 if (nr_args
> IORING_OP_LAST
)
4047 nr_args
= IORING_OP_LAST
;
4049 for (i
= 0; i
< nr_args
; i
++) {
4051 if (!io_issue_defs
[i
].not_supported
)
4052 p
->ops
[i
].flags
= IO_URING_OP_SUPPORTED
;
4057 if (copy_to_user(arg
, p
, size
))
4064 static int io_register_personality(struct io_ring_ctx
*ctx
)
4066 const struct cred
*creds
;
4070 creds
= get_current_cred();
4072 ret
= xa_alloc_cyclic(&ctx
->personalities
, &id
, (void *)creds
,
4073 XA_LIMIT(0, USHRT_MAX
), &ctx
->pers_next
, GFP_KERNEL
);
4081 static __cold
int io_register_restrictions(struct io_ring_ctx
*ctx
,
4082 void __user
*arg
, unsigned int nr_args
)
4084 struct io_uring_restriction
*res
;
4088 /* Restrictions allowed only if rings started disabled */
4089 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
4092 /* We allow only a single restrictions registration */
4093 if (ctx
->restrictions
.registered
)
4096 if (!arg
|| nr_args
> IORING_MAX_RESTRICTIONS
)
4099 size
= array_size(nr_args
, sizeof(*res
));
4100 if (size
== SIZE_MAX
)
4103 res
= memdup_user(arg
, size
);
4105 return PTR_ERR(res
);
4109 for (i
= 0; i
< nr_args
; i
++) {
4110 switch (res
[i
].opcode
) {
4111 case IORING_RESTRICTION_REGISTER_OP
:
4112 if (res
[i
].register_op
>= IORING_REGISTER_LAST
) {
4117 __set_bit(res
[i
].register_op
,
4118 ctx
->restrictions
.register_op
);
4120 case IORING_RESTRICTION_SQE_OP
:
4121 if (res
[i
].sqe_op
>= IORING_OP_LAST
) {
4126 __set_bit(res
[i
].sqe_op
, ctx
->restrictions
.sqe_op
);
4128 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED
:
4129 ctx
->restrictions
.sqe_flags_allowed
= res
[i
].sqe_flags
;
4131 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED
:
4132 ctx
->restrictions
.sqe_flags_required
= res
[i
].sqe_flags
;
4141 /* Reset all restrictions if an error happened */
4143 memset(&ctx
->restrictions
, 0, sizeof(ctx
->restrictions
));
4145 ctx
->restrictions
.registered
= true;
4151 static int io_register_enable_rings(struct io_ring_ctx
*ctx
)
4153 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
4156 if (ctx
->flags
& IORING_SETUP_SINGLE_ISSUER
&& !ctx
->submitter_task
) {
4157 WRITE_ONCE(ctx
->submitter_task
, get_task_struct(current
));
4159 * Lazy activation attempts would fail if it was polled before
4160 * submitter_task is set.
4162 if (wq_has_sleeper(&ctx
->poll_wq
))
4163 io_activate_pollwq(ctx
);
4166 if (ctx
->restrictions
.registered
)
4167 ctx
->restricted
= 1;
4169 ctx
->flags
&= ~IORING_SETUP_R_DISABLED
;
4170 if (ctx
->sq_data
&& wq_has_sleeper(&ctx
->sq_data
->wait
))
4171 wake_up(&ctx
->sq_data
->wait
);
4175 static __cold
int io_register_iowq_aff(struct io_ring_ctx
*ctx
,
4176 void __user
*arg
, unsigned len
)
4178 struct io_uring_task
*tctx
= current
->io_uring
;
4179 cpumask_var_t new_mask
;
4182 if (!tctx
|| !tctx
->io_wq
)
4185 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4188 cpumask_clear(new_mask
);
4189 if (len
> cpumask_size())
4190 len
= cpumask_size();
4192 if (in_compat_syscall()) {
4193 ret
= compat_get_bitmap(cpumask_bits(new_mask
),
4194 (const compat_ulong_t __user
*)arg
,
4195 len
* 8 /* CHAR_BIT */);
4197 ret
= copy_from_user(new_mask
, arg
, len
);
4201 free_cpumask_var(new_mask
);
4205 ret
= io_wq_cpu_affinity(tctx
->io_wq
, new_mask
);
4206 free_cpumask_var(new_mask
);
4210 static __cold
int io_unregister_iowq_aff(struct io_ring_ctx
*ctx
)
4212 struct io_uring_task
*tctx
= current
->io_uring
;
4214 if (!tctx
|| !tctx
->io_wq
)
4217 return io_wq_cpu_affinity(tctx
->io_wq
, NULL
);
4220 static __cold
int io_register_iowq_max_workers(struct io_ring_ctx
*ctx
,
4222 __must_hold(&ctx
->uring_lock
)
4224 struct io_tctx_node
*node
;
4225 struct io_uring_task
*tctx
= NULL
;
4226 struct io_sq_data
*sqd
= NULL
;
4230 if (copy_from_user(new_count
, arg
, sizeof(new_count
)))
4232 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
4233 if (new_count
[i
] > INT_MAX
)
4236 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
4240 * Observe the correct sqd->lock -> ctx->uring_lock
4241 * ordering. Fine to drop uring_lock here, we hold
4244 refcount_inc(&sqd
->refs
);
4245 mutex_unlock(&ctx
->uring_lock
);
4246 mutex_lock(&sqd
->lock
);
4247 mutex_lock(&ctx
->uring_lock
);
4249 tctx
= sqd
->thread
->io_uring
;
4252 tctx
= current
->io_uring
;
4255 BUILD_BUG_ON(sizeof(new_count
) != sizeof(ctx
->iowq_limits
));
4257 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
4259 ctx
->iowq_limits
[i
] = new_count
[i
];
4260 ctx
->iowq_limits_set
= true;
4262 if (tctx
&& tctx
->io_wq
) {
4263 ret
= io_wq_max_workers(tctx
->io_wq
, new_count
);
4267 memset(new_count
, 0, sizeof(new_count
));
4271 mutex_unlock(&sqd
->lock
);
4272 io_put_sq_data(sqd
);
4275 if (copy_to_user(arg
, new_count
, sizeof(new_count
)))
4278 /* that's it for SQPOLL, only the SQPOLL task creates requests */
4282 /* now propagate the restriction to all registered users */
4283 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
4284 struct io_uring_task
*tctx
= node
->task
->io_uring
;
4286 if (WARN_ON_ONCE(!tctx
->io_wq
))
4289 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
4290 new_count
[i
] = ctx
->iowq_limits
[i
];
4291 /* ignore errors, it always returns zero anyway */
4292 (void)io_wq_max_workers(tctx
->io_wq
, new_count
);
4297 mutex_unlock(&sqd
->lock
);
4298 io_put_sq_data(sqd
);
4303 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
4304 void __user
*arg
, unsigned nr_args
)
4305 __releases(ctx
->uring_lock
)
4306 __acquires(ctx
->uring_lock
)
4311 * We don't quiesce the refs for register anymore and so it can't be
4312 * dying as we're holding a file ref here.
4314 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx
->refs
)))
4317 if (ctx
->submitter_task
&& ctx
->submitter_task
!= current
)
4320 if (ctx
->restricted
) {
4321 opcode
= array_index_nospec(opcode
, IORING_REGISTER_LAST
);
4322 if (!test_bit(opcode
, ctx
->restrictions
.register_op
))
4327 case IORING_REGISTER_BUFFERS
:
4331 ret
= io_sqe_buffers_register(ctx
, arg
, nr_args
, NULL
);
4333 case IORING_UNREGISTER_BUFFERS
:
4337 ret
= io_sqe_buffers_unregister(ctx
);
4339 case IORING_REGISTER_FILES
:
4343 ret
= io_sqe_files_register(ctx
, arg
, nr_args
, NULL
);
4345 case IORING_UNREGISTER_FILES
:
4349 ret
= io_sqe_files_unregister(ctx
);
4351 case IORING_REGISTER_FILES_UPDATE
:
4352 ret
= io_register_files_update(ctx
, arg
, nr_args
);
4354 case IORING_REGISTER_EVENTFD
:
4358 ret
= io_eventfd_register(ctx
, arg
, 0);
4360 case IORING_REGISTER_EVENTFD_ASYNC
:
4364 ret
= io_eventfd_register(ctx
, arg
, 1);
4366 case IORING_UNREGISTER_EVENTFD
:
4370 ret
= io_eventfd_unregister(ctx
);
4372 case IORING_REGISTER_PROBE
:
4374 if (!arg
|| nr_args
> 256)
4376 ret
= io_probe(ctx
, arg
, nr_args
);
4378 case IORING_REGISTER_PERSONALITY
:
4382 ret
= io_register_personality(ctx
);
4384 case IORING_UNREGISTER_PERSONALITY
:
4388 ret
= io_unregister_personality(ctx
, nr_args
);
4390 case IORING_REGISTER_ENABLE_RINGS
:
4394 ret
= io_register_enable_rings(ctx
);
4396 case IORING_REGISTER_RESTRICTIONS
:
4397 ret
= io_register_restrictions(ctx
, arg
, nr_args
);
4399 case IORING_REGISTER_FILES2
:
4400 ret
= io_register_rsrc(ctx
, arg
, nr_args
, IORING_RSRC_FILE
);
4402 case IORING_REGISTER_FILES_UPDATE2
:
4403 ret
= io_register_rsrc_update(ctx
, arg
, nr_args
,
4406 case IORING_REGISTER_BUFFERS2
:
4407 ret
= io_register_rsrc(ctx
, arg
, nr_args
, IORING_RSRC_BUFFER
);
4409 case IORING_REGISTER_BUFFERS_UPDATE
:
4410 ret
= io_register_rsrc_update(ctx
, arg
, nr_args
,
4411 IORING_RSRC_BUFFER
);
4413 case IORING_REGISTER_IOWQ_AFF
:
4415 if (!arg
|| !nr_args
)
4417 ret
= io_register_iowq_aff(ctx
, arg
, nr_args
);
4419 case IORING_UNREGISTER_IOWQ_AFF
:
4423 ret
= io_unregister_iowq_aff(ctx
);
4425 case IORING_REGISTER_IOWQ_MAX_WORKERS
:
4427 if (!arg
|| nr_args
!= 2)
4429 ret
= io_register_iowq_max_workers(ctx
, arg
);
4431 case IORING_REGISTER_RING_FDS
:
4432 ret
= io_ringfd_register(ctx
, arg
, nr_args
);
4434 case IORING_UNREGISTER_RING_FDS
:
4435 ret
= io_ringfd_unregister(ctx
, arg
, nr_args
);
4437 case IORING_REGISTER_PBUF_RING
:
4439 if (!arg
|| nr_args
!= 1)
4441 ret
= io_register_pbuf_ring(ctx
, arg
);
4443 case IORING_UNREGISTER_PBUF_RING
:
4445 if (!arg
|| nr_args
!= 1)
4447 ret
= io_unregister_pbuf_ring(ctx
, arg
);
4449 case IORING_REGISTER_SYNC_CANCEL
:
4451 if (!arg
|| nr_args
!= 1)
4453 ret
= io_sync_cancel(ctx
, arg
);
4455 case IORING_REGISTER_FILE_ALLOC_RANGE
:
4457 if (!arg
|| nr_args
)
4459 ret
= io_register_file_alloc_range(ctx
, arg
);
4469 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
4470 void __user
*, arg
, unsigned int, nr_args
)
4472 struct io_ring_ctx
*ctx
;
4475 bool use_registered_ring
;
4477 use_registered_ring
= !!(opcode
& IORING_REGISTER_USE_REGISTERED_RING
);
4478 opcode
&= ~IORING_REGISTER_USE_REGISTERED_RING
;
4480 if (opcode
>= IORING_REGISTER_LAST
)
4483 if (use_registered_ring
) {
4485 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
4486 * need only dereference our task private array to find it.
4488 struct io_uring_task
*tctx
= current
->io_uring
;
4490 if (unlikely(!tctx
|| fd
>= IO_RINGFD_REG_MAX
))
4492 fd
= array_index_nospec(fd
, IO_RINGFD_REG_MAX
);
4493 f
.file
= tctx
->registered_rings
[fd
];
4495 if (unlikely(!f
.file
))
4499 if (unlikely(!f
.file
))
4502 if (!io_is_uring_fops(f
.file
))
4506 ctx
= f
.file
->private_data
;
4508 mutex_lock(&ctx
->uring_lock
);
4509 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
4510 mutex_unlock(&ctx
->uring_lock
);
4511 trace_io_uring_register(ctx
, opcode
, ctx
->nr_user_files
, ctx
->nr_user_bufs
, ret
);
4517 static int __init
io_uring_init(void)
4519 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4520 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4521 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4524 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4525 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4526 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4527 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4528 BUILD_BUG_ON(sizeof(struct io_uring_sqe
) != 64);
4529 BUILD_BUG_SQE_ELEM(0, __u8
, opcode
);
4530 BUILD_BUG_SQE_ELEM(1, __u8
, flags
);
4531 BUILD_BUG_SQE_ELEM(2, __u16
, ioprio
);
4532 BUILD_BUG_SQE_ELEM(4, __s32
, fd
);
4533 BUILD_BUG_SQE_ELEM(8, __u64
, off
);
4534 BUILD_BUG_SQE_ELEM(8, __u64
, addr2
);
4535 BUILD_BUG_SQE_ELEM(8, __u32
, cmd_op
);
4536 BUILD_BUG_SQE_ELEM(12, __u32
, __pad1
);
4537 BUILD_BUG_SQE_ELEM(16, __u64
, addr
);
4538 BUILD_BUG_SQE_ELEM(16, __u64
, splice_off_in
);
4539 BUILD_BUG_SQE_ELEM(24, __u32
, len
);
4540 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t
, rw_flags
);
4541 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags
);
4542 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32
, rw_flags
);
4543 BUILD_BUG_SQE_ELEM(28, __u32
, fsync_flags
);
4544 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16
, poll_events
);
4545 BUILD_BUG_SQE_ELEM(28, __u32
, poll32_events
);
4546 BUILD_BUG_SQE_ELEM(28, __u32
, sync_range_flags
);
4547 BUILD_BUG_SQE_ELEM(28, __u32
, msg_flags
);
4548 BUILD_BUG_SQE_ELEM(28, __u32
, timeout_flags
);
4549 BUILD_BUG_SQE_ELEM(28, __u32
, accept_flags
);
4550 BUILD_BUG_SQE_ELEM(28, __u32
, cancel_flags
);
4551 BUILD_BUG_SQE_ELEM(28, __u32
, open_flags
);
4552 BUILD_BUG_SQE_ELEM(28, __u32
, statx_flags
);
4553 BUILD_BUG_SQE_ELEM(28, __u32
, fadvise_advice
);
4554 BUILD_BUG_SQE_ELEM(28, __u32
, splice_flags
);
4555 BUILD_BUG_SQE_ELEM(28, __u32
, rename_flags
);
4556 BUILD_BUG_SQE_ELEM(28, __u32
, unlink_flags
);
4557 BUILD_BUG_SQE_ELEM(28, __u32
, hardlink_flags
);
4558 BUILD_BUG_SQE_ELEM(28, __u32
, xattr_flags
);
4559 BUILD_BUG_SQE_ELEM(28, __u32
, msg_ring_flags
);
4560 BUILD_BUG_SQE_ELEM(32, __u64
, user_data
);
4561 BUILD_BUG_SQE_ELEM(40, __u16
, buf_index
);
4562 BUILD_BUG_SQE_ELEM(40, __u16
, buf_group
);
4563 BUILD_BUG_SQE_ELEM(42, __u16
, personality
);
4564 BUILD_BUG_SQE_ELEM(44, __s32
, splice_fd_in
);
4565 BUILD_BUG_SQE_ELEM(44, __u32
, file_index
);
4566 BUILD_BUG_SQE_ELEM(44, __u16
, addr_len
);
4567 BUILD_BUG_SQE_ELEM(46, __u16
, __pad3
[0]);
4568 BUILD_BUG_SQE_ELEM(48, __u64
, addr3
);
4569 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd
);
4570 BUILD_BUG_SQE_ELEM(56, __u64
, __pad2
);
4572 BUILD_BUG_ON(sizeof(struct io_uring_files_update
) !=
4573 sizeof(struct io_uring_rsrc_update
));
4574 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update
) >
4575 sizeof(struct io_uring_rsrc_update2
));
4577 /* ->buf_index is u16 */
4578 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring
, bufs
) != 0);
4579 BUILD_BUG_ON(offsetof(struct io_uring_buf
, resv
) !=
4580 offsetof(struct io_uring_buf_ring
, tail
));
4582 /* should fit into one byte */
4583 BUILD_BUG_ON(SQE_VALID_FLAGS
>= (1 << 8));
4584 BUILD_BUG_ON(SQE_COMMON_FLAGS
>= (1 << 8));
4585 BUILD_BUG_ON((SQE_VALID_FLAGS
| SQE_COMMON_FLAGS
) != SQE_VALID_FLAGS
);
4587 BUILD_BUG_ON(__REQ_F_LAST_BIT
> 8 * sizeof(int));
4589 BUILD_BUG_ON(sizeof(atomic_t
) != sizeof(u32
));
4591 io_uring_optable_init();
4593 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
|
4594 SLAB_ACCOUNT
| SLAB_TYPESAFE_BY_RCU
);
4597 __initcall(io_uring_init
);