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>
101 #include "alloc_cache.h"
103 #define IORING_MAX_ENTRIES 32768
104 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
106 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
107 IORING_REGISTER_LAST + IORING_OP_LAST)
109 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
110 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
112 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
113 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
115 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
116 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
119 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
122 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
124 #define IO_COMPL_BATCH 32
125 #define IO_REQ_ALLOC_BATCH 8
128 IO_CHECK_CQ_OVERFLOW_BIT
,
129 IO_CHECK_CQ_DROPPED_BIT
,
133 IO_EVENTFD_OP_SIGNAL_BIT
,
134 IO_EVENTFD_OP_FREE_BIT
,
137 struct io_defer_entry
{
138 struct list_head list
;
139 struct io_kiocb
*req
;
143 /* requests with any of those set should undergo io_disarm_next() */
144 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
145 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
147 static bool io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
148 struct task_struct
*task
,
151 static void io_queue_sqe(struct io_kiocb
*req
);
153 struct kmem_cache
*req_cachep
;
155 static int __read_mostly sysctl_io_uring_disabled
;
156 static int __read_mostly sysctl_io_uring_group
= -1;
159 static struct ctl_table kernel_io_uring_disabled_table
[] = {
161 .procname
= "io_uring_disabled",
162 .data
= &sysctl_io_uring_disabled
,
163 .maxlen
= sizeof(sysctl_io_uring_disabled
),
165 .proc_handler
= proc_dointvec_minmax
,
166 .extra1
= SYSCTL_ZERO
,
167 .extra2
= SYSCTL_TWO
,
170 .procname
= "io_uring_group",
171 .data
= &sysctl_io_uring_group
,
172 .maxlen
= sizeof(gid_t
),
174 .proc_handler
= proc_dointvec
,
180 struct sock
*io_uring_get_socket(struct file
*file
)
182 #if defined(CONFIG_UNIX)
183 if (io_is_uring_fops(file
)) {
184 struct io_ring_ctx
*ctx
= file
->private_data
;
186 return ctx
->ring_sock
->sk
;
191 EXPORT_SYMBOL(io_uring_get_socket
);
193 static inline void io_submit_flush_completions(struct io_ring_ctx
*ctx
)
195 if (!wq_list_empty(&ctx
->submit_state
.compl_reqs
) ||
196 ctx
->submit_state
.cqes_count
)
197 __io_submit_flush_completions(ctx
);
200 static inline unsigned int __io_cqring_events(struct io_ring_ctx
*ctx
)
202 return ctx
->cached_cq_tail
- READ_ONCE(ctx
->rings
->cq
.head
);
205 static inline unsigned int __io_cqring_events_user(struct io_ring_ctx
*ctx
)
207 return READ_ONCE(ctx
->rings
->cq
.tail
) - READ_ONCE(ctx
->rings
->cq
.head
);
210 static bool io_match_linked(struct io_kiocb
*head
)
212 struct io_kiocb
*req
;
214 io_for_each_link(req
, head
) {
215 if (req
->flags
& REQ_F_INFLIGHT
)
222 * As io_match_task() but protected against racing with linked timeouts.
223 * User must not hold timeout_lock.
225 bool io_match_task_safe(struct io_kiocb
*head
, struct task_struct
*task
,
230 if (task
&& head
->task
!= task
)
235 if (head
->flags
& REQ_F_LINK_TIMEOUT
) {
236 struct io_ring_ctx
*ctx
= head
->ctx
;
238 /* protect against races with linked timeouts */
239 spin_lock_irq(&ctx
->timeout_lock
);
240 matched
= io_match_linked(head
);
241 spin_unlock_irq(&ctx
->timeout_lock
);
243 matched
= io_match_linked(head
);
248 static inline void req_fail_link_node(struct io_kiocb
*req
, int res
)
251 io_req_set_res(req
, res
, 0);
254 static inline void io_req_add_to_cache(struct io_kiocb
*req
, struct io_ring_ctx
*ctx
)
256 wq_stack_add_head(&req
->comp_list
, &ctx
->submit_state
.free_list
);
259 static __cold
void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
261 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
263 complete(&ctx
->ref_comp
);
266 static __cold
void io_fallback_req_func(struct work_struct
*work
)
268 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
,
270 struct llist_node
*node
= llist_del_all(&ctx
->fallback_llist
);
271 struct io_kiocb
*req
, *tmp
;
272 struct io_tw_state ts
= { .locked
= true, };
274 mutex_lock(&ctx
->uring_lock
);
275 llist_for_each_entry_safe(req
, tmp
, node
, io_task_work
.node
)
276 req
->io_task_work
.func(req
, &ts
);
277 if (WARN_ON_ONCE(!ts
.locked
))
279 io_submit_flush_completions(ctx
);
280 mutex_unlock(&ctx
->uring_lock
);
283 static int io_alloc_hash_table(struct io_hash_table
*table
, unsigned bits
)
285 unsigned hash_buckets
= 1U << bits
;
286 size_t hash_size
= hash_buckets
* sizeof(table
->hbs
[0]);
288 table
->hbs
= kmalloc(hash_size
, GFP_KERNEL
);
292 table
->hash_bits
= bits
;
293 init_hash_table(table
, hash_buckets
);
297 static __cold
struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
299 struct io_ring_ctx
*ctx
;
302 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
306 xa_init(&ctx
->io_bl_xa
);
309 * Use 5 bits less than the max cq entries, that should give us around
310 * 32 entries per hash list if totally full and uniformly spread, but
311 * don't keep too many buckets to not overconsume memory.
313 hash_bits
= ilog2(p
->cq_entries
) - 5;
314 hash_bits
= clamp(hash_bits
, 1, 8);
315 if (io_alloc_hash_table(&ctx
->cancel_table
, hash_bits
))
317 if (io_alloc_hash_table(&ctx
->cancel_table_locked
, hash_bits
))
319 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
,
323 ctx
->flags
= p
->flags
;
324 init_waitqueue_head(&ctx
->sqo_sq_wait
);
325 INIT_LIST_HEAD(&ctx
->sqd_list
);
326 INIT_LIST_HEAD(&ctx
->cq_overflow_list
);
327 INIT_LIST_HEAD(&ctx
->io_buffers_cache
);
328 INIT_HLIST_HEAD(&ctx
->io_buf_list
);
329 io_alloc_cache_init(&ctx
->rsrc_node_cache
, IO_NODE_ALLOC_CACHE_MAX
,
330 sizeof(struct io_rsrc_node
));
331 io_alloc_cache_init(&ctx
->apoll_cache
, IO_ALLOC_CACHE_MAX
,
332 sizeof(struct async_poll
));
333 io_alloc_cache_init(&ctx
->netmsg_cache
, IO_ALLOC_CACHE_MAX
,
334 sizeof(struct io_async_msghdr
));
335 io_futex_cache_init(ctx
);
336 init_completion(&ctx
->ref_comp
);
337 xa_init_flags(&ctx
->personalities
, XA_FLAGS_ALLOC1
);
338 mutex_init(&ctx
->uring_lock
);
339 init_waitqueue_head(&ctx
->cq_wait
);
340 init_waitqueue_head(&ctx
->poll_wq
);
341 init_waitqueue_head(&ctx
->rsrc_quiesce_wq
);
342 spin_lock_init(&ctx
->completion_lock
);
343 spin_lock_init(&ctx
->timeout_lock
);
344 INIT_WQ_LIST(&ctx
->iopoll_list
);
345 INIT_LIST_HEAD(&ctx
->io_buffers_comp
);
346 INIT_LIST_HEAD(&ctx
->defer_list
);
347 INIT_LIST_HEAD(&ctx
->timeout_list
);
348 INIT_LIST_HEAD(&ctx
->ltimeout_list
);
349 INIT_LIST_HEAD(&ctx
->rsrc_ref_list
);
350 init_llist_head(&ctx
->work_llist
);
351 INIT_LIST_HEAD(&ctx
->tctx_list
);
352 ctx
->submit_state
.free_list
.next
= NULL
;
353 INIT_WQ_LIST(&ctx
->locked_free_list
);
354 INIT_HLIST_HEAD(&ctx
->waitid_list
);
356 INIT_HLIST_HEAD(&ctx
->futex_list
);
358 INIT_DELAYED_WORK(&ctx
->fallback_work
, io_fallback_req_func
);
359 INIT_WQ_LIST(&ctx
->submit_state
.compl_reqs
);
360 INIT_HLIST_HEAD(&ctx
->cancelable_uring_cmd
);
363 kfree(ctx
->cancel_table
.hbs
);
364 kfree(ctx
->cancel_table_locked
.hbs
);
366 xa_destroy(&ctx
->io_bl_xa
);
371 static void io_account_cq_overflow(struct io_ring_ctx
*ctx
)
373 struct io_rings
*r
= ctx
->rings
;
375 WRITE_ONCE(r
->cq_overflow
, READ_ONCE(r
->cq_overflow
) + 1);
379 static bool req_need_defer(struct io_kiocb
*req
, u32 seq
)
381 if (unlikely(req
->flags
& REQ_F_IO_DRAIN
)) {
382 struct io_ring_ctx
*ctx
= req
->ctx
;
384 return seq
+ READ_ONCE(ctx
->cq_extra
) != ctx
->cached_cq_tail
;
390 static void io_clean_op(struct io_kiocb
*req
)
392 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
393 spin_lock(&req
->ctx
->completion_lock
);
394 io_put_kbuf_comp(req
);
395 spin_unlock(&req
->ctx
->completion_lock
);
398 if (req
->flags
& REQ_F_NEED_CLEANUP
) {
399 const struct io_cold_def
*def
= &io_cold_defs
[req
->opcode
];
404 if ((req
->flags
& REQ_F_POLLED
) && req
->apoll
) {
405 kfree(req
->apoll
->double_poll
);
409 if (req
->flags
& REQ_F_INFLIGHT
) {
410 struct io_uring_task
*tctx
= req
->task
->io_uring
;
412 atomic_dec(&tctx
->inflight_tracked
);
414 if (req
->flags
& REQ_F_CREDS
)
415 put_cred(req
->creds
);
416 if (req
->flags
& REQ_F_ASYNC_DATA
) {
417 kfree(req
->async_data
);
418 req
->async_data
= NULL
;
420 req
->flags
&= ~IO_REQ_CLEAN_FLAGS
;
423 static inline void io_req_track_inflight(struct io_kiocb
*req
)
425 if (!(req
->flags
& REQ_F_INFLIGHT
)) {
426 req
->flags
|= REQ_F_INFLIGHT
;
427 atomic_inc(&req
->task
->io_uring
->inflight_tracked
);
431 static struct io_kiocb
*__io_prep_linked_timeout(struct io_kiocb
*req
)
433 if (WARN_ON_ONCE(!req
->link
))
436 req
->flags
&= ~REQ_F_ARM_LTIMEOUT
;
437 req
->flags
|= REQ_F_LINK_TIMEOUT
;
439 /* linked timeouts should have two refs once prep'ed */
440 io_req_set_refcount(req
);
441 __io_req_set_refcount(req
->link
, 2);
445 static inline struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
)
447 if (likely(!(req
->flags
& REQ_F_ARM_LTIMEOUT
)))
449 return __io_prep_linked_timeout(req
);
452 static noinline
void __io_arm_ltimeout(struct io_kiocb
*req
)
454 io_queue_linked_timeout(__io_prep_linked_timeout(req
));
457 static inline void io_arm_ltimeout(struct io_kiocb
*req
)
459 if (unlikely(req
->flags
& REQ_F_ARM_LTIMEOUT
))
460 __io_arm_ltimeout(req
);
463 static void io_prep_async_work(struct io_kiocb
*req
)
465 const struct io_issue_def
*def
= &io_issue_defs
[req
->opcode
];
466 struct io_ring_ctx
*ctx
= req
->ctx
;
468 if (!(req
->flags
& REQ_F_CREDS
)) {
469 req
->flags
|= REQ_F_CREDS
;
470 req
->creds
= get_current_cred();
473 req
->work
.list
.next
= NULL
;
475 req
->work
.cancel_seq
= atomic_read(&ctx
->cancel_seq
);
476 if (req
->flags
& REQ_F_FORCE_ASYNC
)
477 req
->work
.flags
|= IO_WQ_WORK_CONCURRENT
;
479 if (req
->file
&& !(req
->flags
& REQ_F_FIXED_FILE
))
480 req
->flags
|= io_file_get_flags(req
->file
);
482 if (req
->file
&& (req
->flags
& REQ_F_ISREG
)) {
483 bool should_hash
= def
->hash_reg_file
;
485 /* don't serialize this request if the fs doesn't need it */
486 if (should_hash
&& (req
->file
->f_flags
& O_DIRECT
) &&
487 (req
->file
->f_mode
& FMODE_DIO_PARALLEL_WRITE
))
489 if (should_hash
|| (ctx
->flags
& IORING_SETUP_IOPOLL
))
490 io_wq_hash_work(&req
->work
, file_inode(req
->file
));
491 } else if (!req
->file
|| !S_ISBLK(file_inode(req
->file
)->i_mode
)) {
492 if (def
->unbound_nonreg_file
)
493 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
497 static void io_prep_async_link(struct io_kiocb
*req
)
499 struct io_kiocb
*cur
;
501 if (req
->flags
& REQ_F_LINK_TIMEOUT
) {
502 struct io_ring_ctx
*ctx
= req
->ctx
;
504 spin_lock_irq(&ctx
->timeout_lock
);
505 io_for_each_link(cur
, req
)
506 io_prep_async_work(cur
);
507 spin_unlock_irq(&ctx
->timeout_lock
);
509 io_for_each_link(cur
, req
)
510 io_prep_async_work(cur
);
514 void io_queue_iowq(struct io_kiocb
*req
, struct io_tw_state
*ts_dont_use
)
516 struct io_kiocb
*link
= io_prep_linked_timeout(req
);
517 struct io_uring_task
*tctx
= req
->task
->io_uring
;
520 BUG_ON(!tctx
->io_wq
);
522 /* init ->work of the whole link before punting */
523 io_prep_async_link(req
);
526 * Not expected to happen, but if we do have a bug where this _can_
527 * happen, catch it here and ensure the request is marked as
528 * canceled. That will make io-wq go through the usual work cancel
529 * procedure rather than attempt to run this request (or create a new
532 if (WARN_ON_ONCE(!same_thread_group(req
->task
, current
)))
533 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
535 trace_io_uring_queue_async_work(req
, io_wq_is_hashed(&req
->work
));
536 io_wq_enqueue(tctx
->io_wq
, &req
->work
);
538 io_queue_linked_timeout(link
);
541 static __cold
void io_queue_deferred(struct io_ring_ctx
*ctx
)
543 while (!list_empty(&ctx
->defer_list
)) {
544 struct io_defer_entry
*de
= list_first_entry(&ctx
->defer_list
,
545 struct io_defer_entry
, list
);
547 if (req_need_defer(de
->req
, de
->seq
))
549 list_del_init(&de
->list
);
550 io_req_task_queue(de
->req
);
556 static void io_eventfd_ops(struct rcu_head
*rcu
)
558 struct io_ev_fd
*ev_fd
= container_of(rcu
, struct io_ev_fd
, rcu
);
559 int ops
= atomic_xchg(&ev_fd
->ops
, 0);
561 if (ops
& BIT(IO_EVENTFD_OP_SIGNAL_BIT
))
562 eventfd_signal_mask(ev_fd
->cq_ev_fd
, 1, EPOLL_URING_WAKE
);
564 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
565 * ordering in a race but if references are 0 we know we have to free
568 if (atomic_dec_and_test(&ev_fd
->refs
)) {
569 eventfd_ctx_put(ev_fd
->cq_ev_fd
);
574 static void io_eventfd_signal(struct io_ring_ctx
*ctx
)
576 struct io_ev_fd
*ev_fd
= NULL
;
580 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
583 ev_fd
= rcu_dereference(ctx
->io_ev_fd
);
586 * Check again if ev_fd exists incase an io_eventfd_unregister call
587 * completed between the NULL check of ctx->io_ev_fd at the start of
588 * the function and rcu_read_lock.
590 if (unlikely(!ev_fd
))
592 if (READ_ONCE(ctx
->rings
->cq_flags
) & IORING_CQ_EVENTFD_DISABLED
)
594 if (ev_fd
->eventfd_async
&& !io_wq_current_is_worker())
597 if (likely(eventfd_signal_allowed())) {
598 eventfd_signal_mask(ev_fd
->cq_ev_fd
, 1, EPOLL_URING_WAKE
);
600 atomic_inc(&ev_fd
->refs
);
601 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT
), &ev_fd
->ops
))
602 call_rcu_hurry(&ev_fd
->rcu
, io_eventfd_ops
);
604 atomic_dec(&ev_fd
->refs
);
611 static void io_eventfd_flush_signal(struct io_ring_ctx
*ctx
)
615 spin_lock(&ctx
->completion_lock
);
618 * Eventfd should only get triggered when at least one event has been
619 * posted. Some applications rely on the eventfd notification count
620 * only changing IFF a new CQE has been added to the CQ ring. There's
621 * no depedency on 1:1 relationship between how many times this
622 * function is called (and hence the eventfd count) and number of CQEs
623 * posted to the CQ ring.
625 skip
= ctx
->cached_cq_tail
== ctx
->evfd_last_cq_tail
;
626 ctx
->evfd_last_cq_tail
= ctx
->cached_cq_tail
;
627 spin_unlock(&ctx
->completion_lock
);
631 io_eventfd_signal(ctx
);
634 void __io_commit_cqring_flush(struct io_ring_ctx
*ctx
)
636 if (ctx
->poll_activated
)
637 io_poll_wq_wake(ctx
);
638 if (ctx
->off_timeout_used
)
639 io_flush_timeouts(ctx
);
640 if (ctx
->drain_active
) {
641 spin_lock(&ctx
->completion_lock
);
642 io_queue_deferred(ctx
);
643 spin_unlock(&ctx
->completion_lock
);
646 io_eventfd_flush_signal(ctx
);
649 static inline void __io_cq_lock(struct io_ring_ctx
*ctx
)
651 if (!ctx
->lockless_cq
)
652 spin_lock(&ctx
->completion_lock
);
655 static inline void io_cq_lock(struct io_ring_ctx
*ctx
)
656 __acquires(ctx
->completion_lock
)
658 spin_lock(&ctx
->completion_lock
);
661 static inline void __io_cq_unlock_post(struct io_ring_ctx
*ctx
)
663 io_commit_cqring(ctx
);
664 if (!ctx
->task_complete
) {
665 if (!ctx
->lockless_cq
)
666 spin_unlock(&ctx
->completion_lock
);
667 /* IOPOLL rings only need to wake up if it's also SQPOLL */
668 if (!ctx
->syscall_iopoll
)
671 io_commit_cqring_flush(ctx
);
674 static void io_cq_unlock_post(struct io_ring_ctx
*ctx
)
675 __releases(ctx
->completion_lock
)
677 io_commit_cqring(ctx
);
678 spin_unlock(&ctx
->completion_lock
);
680 io_commit_cqring_flush(ctx
);
683 /* Returns true if there are no backlogged entries after the flush */
684 static void io_cqring_overflow_kill(struct io_ring_ctx
*ctx
)
686 struct io_overflow_cqe
*ocqe
;
689 spin_lock(&ctx
->completion_lock
);
690 list_splice_init(&ctx
->cq_overflow_list
, &list
);
691 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
);
692 spin_unlock(&ctx
->completion_lock
);
694 while (!list_empty(&list
)) {
695 ocqe
= list_first_entry(&list
, struct io_overflow_cqe
, list
);
696 list_del(&ocqe
->list
);
701 static void __io_cqring_overflow_flush(struct io_ring_ctx
*ctx
)
703 size_t cqe_size
= sizeof(struct io_uring_cqe
);
705 if (__io_cqring_events(ctx
) == ctx
->cq_entries
)
708 if (ctx
->flags
& IORING_SETUP_CQE32
)
712 while (!list_empty(&ctx
->cq_overflow_list
)) {
713 struct io_uring_cqe
*cqe
;
714 struct io_overflow_cqe
*ocqe
;
716 if (!io_get_cqe_overflow(ctx
, &cqe
, true))
718 ocqe
= list_first_entry(&ctx
->cq_overflow_list
,
719 struct io_overflow_cqe
, list
);
720 memcpy(cqe
, &ocqe
->cqe
, cqe_size
);
721 list_del(&ocqe
->list
);
725 if (list_empty(&ctx
->cq_overflow_list
)) {
726 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
);
727 atomic_andnot(IORING_SQ_CQ_OVERFLOW
, &ctx
->rings
->sq_flags
);
729 io_cq_unlock_post(ctx
);
732 static void io_cqring_do_overflow_flush(struct io_ring_ctx
*ctx
)
734 /* iopoll syncs against uring_lock, not completion_lock */
735 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
736 mutex_lock(&ctx
->uring_lock
);
737 __io_cqring_overflow_flush(ctx
);
738 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
739 mutex_unlock(&ctx
->uring_lock
);
742 static void io_cqring_overflow_flush(struct io_ring_ctx
*ctx
)
744 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
))
745 io_cqring_do_overflow_flush(ctx
);
748 /* can be called by any task */
749 static void io_put_task_remote(struct task_struct
*task
)
751 struct io_uring_task
*tctx
= task
->io_uring
;
753 percpu_counter_sub(&tctx
->inflight
, 1);
754 if (unlikely(atomic_read(&tctx
->in_cancel
)))
755 wake_up(&tctx
->wait
);
756 put_task_struct(task
);
759 /* used by a task to put its own references */
760 static void io_put_task_local(struct task_struct
*task
)
762 task
->io_uring
->cached_refs
++;
765 /* must to be called somewhat shortly after putting a request */
766 static inline void io_put_task(struct task_struct
*task
)
768 if (likely(task
== current
))
769 io_put_task_local(task
);
771 io_put_task_remote(task
);
774 void io_task_refs_refill(struct io_uring_task
*tctx
)
776 unsigned int refill
= -tctx
->cached_refs
+ IO_TCTX_REFS_CACHE_NR
;
778 percpu_counter_add(&tctx
->inflight
, refill
);
779 refcount_add(refill
, ¤t
->usage
);
780 tctx
->cached_refs
+= refill
;
783 static __cold
void io_uring_drop_tctx_refs(struct task_struct
*task
)
785 struct io_uring_task
*tctx
= task
->io_uring
;
786 unsigned int refs
= tctx
->cached_refs
;
789 tctx
->cached_refs
= 0;
790 percpu_counter_sub(&tctx
->inflight
, refs
);
791 put_task_struct_many(task
, refs
);
795 static bool io_cqring_event_overflow(struct io_ring_ctx
*ctx
, u64 user_data
,
796 s32 res
, u32 cflags
, u64 extra1
, u64 extra2
)
798 struct io_overflow_cqe
*ocqe
;
799 size_t ocq_size
= sizeof(struct io_overflow_cqe
);
800 bool is_cqe32
= (ctx
->flags
& IORING_SETUP_CQE32
);
802 lockdep_assert_held(&ctx
->completion_lock
);
805 ocq_size
+= sizeof(struct io_uring_cqe
);
807 ocqe
= kmalloc(ocq_size
, GFP_ATOMIC
| __GFP_ACCOUNT
);
808 trace_io_uring_cqe_overflow(ctx
, user_data
, res
, cflags
, ocqe
);
811 * If we're in ring overflow flush mode, or in task cancel mode,
812 * or cannot allocate an overflow entry, then we need to drop it
815 io_account_cq_overflow(ctx
);
816 set_bit(IO_CHECK_CQ_DROPPED_BIT
, &ctx
->check_cq
);
819 if (list_empty(&ctx
->cq_overflow_list
)) {
820 set_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
);
821 atomic_or(IORING_SQ_CQ_OVERFLOW
, &ctx
->rings
->sq_flags
);
824 ocqe
->cqe
.user_data
= user_data
;
826 ocqe
->cqe
.flags
= cflags
;
828 ocqe
->cqe
.big_cqe
[0] = extra1
;
829 ocqe
->cqe
.big_cqe
[1] = extra2
;
831 list_add_tail(&ocqe
->list
, &ctx
->cq_overflow_list
);
835 void io_req_cqe_overflow(struct io_kiocb
*req
)
837 io_cqring_event_overflow(req
->ctx
, req
->cqe
.user_data
,
838 req
->cqe
.res
, req
->cqe
.flags
,
839 req
->big_cqe
.extra1
, req
->big_cqe
.extra2
);
840 memset(&req
->big_cqe
, 0, sizeof(req
->big_cqe
));
844 * writes to the cq entry need to come after reading head; the
845 * control dependency is enough as we're using WRITE_ONCE to
848 bool io_cqe_cache_refill(struct io_ring_ctx
*ctx
, bool overflow
)
850 struct io_rings
*rings
= ctx
->rings
;
851 unsigned int off
= ctx
->cached_cq_tail
& (ctx
->cq_entries
- 1);
852 unsigned int free
, queued
, len
;
855 * Posting into the CQ when there are pending overflowed CQEs may break
856 * ordering guarantees, which will affect links, F_MORE users and more.
857 * Force overflow the completion.
859 if (!overflow
&& (ctx
->check_cq
& BIT(IO_CHECK_CQ_OVERFLOW_BIT
)))
862 /* userspace may cheat modifying the tail, be safe and do min */
863 queued
= min(__io_cqring_events(ctx
), ctx
->cq_entries
);
864 free
= ctx
->cq_entries
- queued
;
865 /* we need a contiguous range, limit based on the current array offset */
866 len
= min(free
, ctx
->cq_entries
- off
);
870 if (ctx
->flags
& IORING_SETUP_CQE32
) {
875 ctx
->cqe_cached
= &rings
->cqes
[off
];
876 ctx
->cqe_sentinel
= ctx
->cqe_cached
+ len
;
880 static bool io_fill_cqe_aux(struct io_ring_ctx
*ctx
, u64 user_data
, s32 res
,
883 struct io_uring_cqe
*cqe
;
888 * If we can't get a cq entry, userspace overflowed the
889 * submission (by quite a lot). Increment the overflow count in
892 if (likely(io_get_cqe(ctx
, &cqe
))) {
893 trace_io_uring_complete(ctx
, NULL
, user_data
, res
, cflags
, 0, 0);
895 WRITE_ONCE(cqe
->user_data
, user_data
);
896 WRITE_ONCE(cqe
->res
, res
);
897 WRITE_ONCE(cqe
->flags
, cflags
);
899 if (ctx
->flags
& IORING_SETUP_CQE32
) {
900 WRITE_ONCE(cqe
->big_cqe
[0], 0);
901 WRITE_ONCE(cqe
->big_cqe
[1], 0);
908 static void __io_flush_post_cqes(struct io_ring_ctx
*ctx
)
909 __must_hold(&ctx
->uring_lock
)
911 struct io_submit_state
*state
= &ctx
->submit_state
;
914 lockdep_assert_held(&ctx
->uring_lock
);
915 for (i
= 0; i
< state
->cqes_count
; i
++) {
916 struct io_uring_cqe
*cqe
= &ctx
->completion_cqes
[i
];
918 if (!io_fill_cqe_aux(ctx
, cqe
->user_data
, cqe
->res
, cqe
->flags
)) {
919 if (ctx
->lockless_cq
) {
920 spin_lock(&ctx
->completion_lock
);
921 io_cqring_event_overflow(ctx
, cqe
->user_data
,
922 cqe
->res
, cqe
->flags
, 0, 0);
923 spin_unlock(&ctx
->completion_lock
);
925 io_cqring_event_overflow(ctx
, cqe
->user_data
,
926 cqe
->res
, cqe
->flags
, 0, 0);
930 state
->cqes_count
= 0;
933 static bool __io_post_aux_cqe(struct io_ring_ctx
*ctx
, u64 user_data
, s32 res
, u32 cflags
,
939 filled
= io_fill_cqe_aux(ctx
, user_data
, res
, cflags
);
940 if (!filled
&& allow_overflow
)
941 filled
= io_cqring_event_overflow(ctx
, user_data
, res
, cflags
, 0, 0);
943 io_cq_unlock_post(ctx
);
947 bool io_post_aux_cqe(struct io_ring_ctx
*ctx
, u64 user_data
, s32 res
, u32 cflags
)
949 return __io_post_aux_cqe(ctx
, user_data
, res
, cflags
, true);
953 * A helper for multishot requests posting additional CQEs.
954 * Should only be used from a task_work including IO_URING_F_MULTISHOT.
956 bool io_fill_cqe_req_aux(struct io_kiocb
*req
, bool defer
, s32 res
, u32 cflags
)
958 struct io_ring_ctx
*ctx
= req
->ctx
;
959 u64 user_data
= req
->cqe
.user_data
;
960 struct io_uring_cqe
*cqe
;
963 return __io_post_aux_cqe(ctx
, user_data
, res
, cflags
, false);
965 lockdep_assert_held(&ctx
->uring_lock
);
967 if (ctx
->submit_state
.cqes_count
== ARRAY_SIZE(ctx
->completion_cqes
)) {
969 __io_flush_post_cqes(ctx
);
970 /* no need to flush - flush is deferred */
971 __io_cq_unlock_post(ctx
);
974 /* For defered completions this is not as strict as it is otherwise,
975 * however it's main job is to prevent unbounded posted completions,
976 * and in that it works just as well.
978 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
))
981 cqe
= &ctx
->completion_cqes
[ctx
->submit_state
.cqes_count
++];
982 cqe
->user_data
= user_data
;
988 static void __io_req_complete_post(struct io_kiocb
*req
, unsigned issue_flags
)
990 struct io_ring_ctx
*ctx
= req
->ctx
;
991 struct io_rsrc_node
*rsrc_node
= NULL
;
994 if (!(req
->flags
& REQ_F_CQE_SKIP
)) {
995 if (!io_fill_cqe_req(ctx
, req
))
996 io_req_cqe_overflow(req
);
1000 * If we're the last reference to this request, add to our locked
1003 if (req_ref_put_and_test(req
)) {
1004 if (req
->flags
& IO_REQ_LINK_FLAGS
) {
1005 if (req
->flags
& IO_DISARM_MASK
)
1006 io_disarm_next(req
);
1008 io_req_task_queue(req
->link
);
1012 io_put_kbuf_comp(req
);
1013 if (unlikely(req
->flags
& IO_REQ_CLEAN_FLAGS
))
1017 rsrc_node
= req
->rsrc_node
;
1019 * Selected buffer deallocation in io_clean_op() assumes that
1020 * we don't hold ->completion_lock. Clean them here to avoid
1023 io_put_task_remote(req
->task
);
1024 wq_list_add_head(&req
->comp_list
, &ctx
->locked_free_list
);
1025 ctx
->locked_free_nr
++;
1027 io_cq_unlock_post(ctx
);
1030 io_ring_submit_lock(ctx
, issue_flags
);
1031 io_put_rsrc_node(ctx
, rsrc_node
);
1032 io_ring_submit_unlock(ctx
, issue_flags
);
1036 void io_req_complete_post(struct io_kiocb
*req
, unsigned issue_flags
)
1038 if (req
->ctx
->task_complete
&& req
->ctx
->submitter_task
!= current
) {
1039 req
->io_task_work
.func
= io_req_task_complete
;
1040 io_req_task_work_add(req
);
1041 } else if (!(issue_flags
& IO_URING_F_UNLOCKED
) ||
1042 !(req
->ctx
->flags
& IORING_SETUP_IOPOLL
)) {
1043 __io_req_complete_post(req
, issue_flags
);
1045 struct io_ring_ctx
*ctx
= req
->ctx
;
1047 mutex_lock(&ctx
->uring_lock
);
1048 __io_req_complete_post(req
, issue_flags
& ~IO_URING_F_UNLOCKED
);
1049 mutex_unlock(&ctx
->uring_lock
);
1053 void io_req_defer_failed(struct io_kiocb
*req
, s32 res
)
1054 __must_hold(&ctx
->uring_lock
)
1056 const struct io_cold_def
*def
= &io_cold_defs
[req
->opcode
];
1058 lockdep_assert_held(&req
->ctx
->uring_lock
);
1061 io_req_set_res(req
, res
, io_put_kbuf(req
, IO_URING_F_UNLOCKED
));
1064 io_req_complete_defer(req
);
1068 * Don't initialise the fields below on every allocation, but do that in
1069 * advance and keep them valid across allocations.
1071 static void io_preinit_req(struct io_kiocb
*req
, struct io_ring_ctx
*ctx
)
1075 req
->async_data
= NULL
;
1076 /* not necessary, but safer to zero */
1077 memset(&req
->cqe
, 0, sizeof(req
->cqe
));
1078 memset(&req
->big_cqe
, 0, sizeof(req
->big_cqe
));
1081 static void io_flush_cached_locked_reqs(struct io_ring_ctx
*ctx
,
1082 struct io_submit_state
*state
)
1084 spin_lock(&ctx
->completion_lock
);
1085 wq_list_splice(&ctx
->locked_free_list
, &state
->free_list
);
1086 ctx
->locked_free_nr
= 0;
1087 spin_unlock(&ctx
->completion_lock
);
1091 * A request might get retired back into the request caches even before opcode
1092 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1093 * Because of that, io_alloc_req() should be called only under ->uring_lock
1094 * and with extra caution to not get a request that is still worked on.
1096 __cold
bool __io_alloc_req_refill(struct io_ring_ctx
*ctx
)
1097 __must_hold(&ctx
->uring_lock
)
1099 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
1100 void *reqs
[IO_REQ_ALLOC_BATCH
];
1104 * If we have more than a batch's worth of requests in our IRQ side
1105 * locked cache, grab the lock and move them over to our submission
1108 if (data_race(ctx
->locked_free_nr
) > IO_COMPL_BATCH
) {
1109 io_flush_cached_locked_reqs(ctx
, &ctx
->submit_state
);
1110 if (!io_req_cache_empty(ctx
))
1114 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, ARRAY_SIZE(reqs
), reqs
);
1117 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1118 * retry single alloc to be on the safe side.
1120 if (unlikely(ret
<= 0)) {
1121 reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
1127 percpu_ref_get_many(&ctx
->refs
, ret
);
1128 for (i
= 0; i
< ret
; i
++) {
1129 struct io_kiocb
*req
= reqs
[i
];
1131 io_preinit_req(req
, ctx
);
1132 io_req_add_to_cache(req
, ctx
);
1137 __cold
void io_free_req(struct io_kiocb
*req
)
1139 /* refs were already put, restore them for io_req_task_complete() */
1140 req
->flags
&= ~REQ_F_REFCOUNT
;
1141 /* we only want to free it, don't post CQEs */
1142 req
->flags
|= REQ_F_CQE_SKIP
;
1143 req
->io_task_work
.func
= io_req_task_complete
;
1144 io_req_task_work_add(req
);
1147 static void __io_req_find_next_prep(struct io_kiocb
*req
)
1149 struct io_ring_ctx
*ctx
= req
->ctx
;
1151 spin_lock(&ctx
->completion_lock
);
1152 io_disarm_next(req
);
1153 spin_unlock(&ctx
->completion_lock
);
1156 static inline struct io_kiocb
*io_req_find_next(struct io_kiocb
*req
)
1158 struct io_kiocb
*nxt
;
1161 * If LINK is set, we have dependent requests in this chain. If we
1162 * didn't fail this request, queue the first one up, moving any other
1163 * dependencies to the next request. In case of failure, fail the rest
1166 if (unlikely(req
->flags
& IO_DISARM_MASK
))
1167 __io_req_find_next_prep(req
);
1173 static void ctx_flush_and_put(struct io_ring_ctx
*ctx
, struct io_tw_state
*ts
)
1177 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
)
1178 atomic_andnot(IORING_SQ_TASKRUN
, &ctx
->rings
->sq_flags
);
1180 io_submit_flush_completions(ctx
);
1181 mutex_unlock(&ctx
->uring_lock
);
1184 percpu_ref_put(&ctx
->refs
);
1187 static unsigned int handle_tw_list(struct llist_node
*node
,
1188 struct io_ring_ctx
**ctx
,
1189 struct io_tw_state
*ts
,
1190 struct llist_node
*last
)
1192 unsigned int count
= 0;
1194 while (node
&& node
!= last
) {
1195 struct llist_node
*next
= node
->next
;
1196 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1199 prefetch(container_of(next
, struct io_kiocb
, io_task_work
.node
));
1201 if (req
->ctx
!= *ctx
) {
1202 ctx_flush_and_put(*ctx
, ts
);
1204 /* if not contended, grab and improve batching */
1205 ts
->locked
= mutex_trylock(&(*ctx
)->uring_lock
);
1206 percpu_ref_get(&(*ctx
)->refs
);
1208 INDIRECT_CALL_2(req
->io_task_work
.func
,
1209 io_poll_task_func
, io_req_rw_complete
,
1213 if (unlikely(need_resched())) {
1214 ctx_flush_and_put(*ctx
, ts
);
1224 * io_llist_xchg - swap all entries in a lock-less list
1225 * @head: the head of lock-less list to delete all entries
1226 * @new: new entry as the head of the list
1228 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1229 * The order of entries returned is from the newest to the oldest added one.
1231 static inline struct llist_node
*io_llist_xchg(struct llist_head
*head
,
1232 struct llist_node
*new)
1234 return xchg(&head
->first
, new);
1238 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1239 * @head: the head of lock-less list to delete all entries
1240 * @old: expected old value of the first entry of the list
1241 * @new: new entry as the head of the list
1243 * perform a cmpxchg on the first entry of the list.
1246 static inline struct llist_node
*io_llist_cmpxchg(struct llist_head
*head
,
1247 struct llist_node
*old
,
1248 struct llist_node
*new)
1250 return cmpxchg(&head
->first
, old
, new);
1253 static __cold
void io_fallback_tw(struct io_uring_task
*tctx
, bool sync
)
1255 struct llist_node
*node
= llist_del_all(&tctx
->task_list
);
1256 struct io_ring_ctx
*last_ctx
= NULL
;
1257 struct io_kiocb
*req
;
1260 req
= container_of(node
, struct io_kiocb
, io_task_work
.node
);
1262 if (sync
&& last_ctx
!= req
->ctx
) {
1264 flush_delayed_work(&last_ctx
->fallback_work
);
1265 percpu_ref_put(&last_ctx
->refs
);
1267 last_ctx
= req
->ctx
;
1268 percpu_ref_get(&last_ctx
->refs
);
1270 if (llist_add(&req
->io_task_work
.node
,
1271 &req
->ctx
->fallback_llist
))
1272 schedule_delayed_work(&req
->ctx
->fallback_work
, 1);
1276 flush_delayed_work(&last_ctx
->fallback_work
);
1277 percpu_ref_put(&last_ctx
->refs
);
1281 void tctx_task_work(struct callback_head
*cb
)
1283 struct io_tw_state ts
= {};
1284 struct io_ring_ctx
*ctx
= NULL
;
1285 struct io_uring_task
*tctx
= container_of(cb
, struct io_uring_task
,
1287 struct llist_node fake
= {};
1288 struct llist_node
*node
;
1289 unsigned int loops
= 0;
1290 unsigned int count
= 0;
1292 if (unlikely(current
->flags
& PF_EXITING
)) {
1293 io_fallback_tw(tctx
, true);
1299 node
= io_llist_xchg(&tctx
->task_list
, &fake
);
1300 count
+= handle_tw_list(node
, &ctx
, &ts
, &fake
);
1302 /* skip expensive cmpxchg if there are items in the list */
1303 if (READ_ONCE(tctx
->task_list
.first
) != &fake
)
1305 if (ts
.locked
&& !wq_list_empty(&ctx
->submit_state
.compl_reqs
)) {
1306 io_submit_flush_completions(ctx
);
1307 if (READ_ONCE(tctx
->task_list
.first
) != &fake
)
1310 node
= io_llist_cmpxchg(&tctx
->task_list
, &fake
, NULL
);
1311 } while (node
!= &fake
);
1313 ctx_flush_and_put(ctx
, &ts
);
1315 /* relaxed read is enough as only the task itself sets ->in_cancel */
1316 if (unlikely(atomic_read(&tctx
->in_cancel
)))
1317 io_uring_drop_tctx_refs(current
);
1319 trace_io_uring_task_work_run(tctx
, count
, loops
);
1322 static inline void io_req_local_work_add(struct io_kiocb
*req
, unsigned flags
)
1324 struct io_ring_ctx
*ctx
= req
->ctx
;
1325 unsigned nr_wait
, nr_tw
, nr_tw_prev
;
1326 struct llist_node
*first
;
1328 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))
1329 flags
&= ~IOU_F_TWQ_LAZY_WAKE
;
1331 first
= READ_ONCE(ctx
->work_llist
.first
);
1335 struct io_kiocb
*first_req
= container_of(first
,
1339 * Might be executed at any moment, rely on
1340 * SLAB_TYPESAFE_BY_RCU to keep it alive.
1342 nr_tw_prev
= READ_ONCE(first_req
->nr_tw
);
1344 nr_tw
= nr_tw_prev
+ 1;
1345 /* Large enough to fail the nr_wait comparison below */
1346 if (!(flags
& IOU_F_TWQ_LAZY_WAKE
))
1350 req
->io_task_work
.node
.next
= first
;
1351 } while (!try_cmpxchg(&ctx
->work_llist
.first
, &first
,
1352 &req
->io_task_work
.node
));
1355 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
)
1356 atomic_or(IORING_SQ_TASKRUN
, &ctx
->rings
->sq_flags
);
1358 io_eventfd_signal(ctx
);
1361 nr_wait
= atomic_read(&ctx
->cq_wait_nr
);
1362 /* no one is waiting */
1365 /* either not enough or the previous add has already woken it up */
1366 if (nr_wait
> nr_tw
|| nr_tw_prev
>= nr_wait
)
1368 /* pairs with set_current_state() in io_cqring_wait() */
1369 smp_mb__after_atomic();
1370 wake_up_state(ctx
->submitter_task
, TASK_INTERRUPTIBLE
);
1373 static void io_req_normal_work_add(struct io_kiocb
*req
)
1375 struct io_uring_task
*tctx
= req
->task
->io_uring
;
1376 struct io_ring_ctx
*ctx
= req
->ctx
;
1378 /* task_work already pending, we're done */
1379 if (!llist_add(&req
->io_task_work
.node
, &tctx
->task_list
))
1382 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
)
1383 atomic_or(IORING_SQ_TASKRUN
, &ctx
->rings
->sq_flags
);
1385 if (likely(!task_work_add(req
->task
, &tctx
->task_work
, ctx
->notify_method
)))
1388 io_fallback_tw(tctx
, false);
1391 void __io_req_task_work_add(struct io_kiocb
*req
, unsigned flags
)
1393 if (req
->ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) {
1395 io_req_local_work_add(req
, flags
);
1398 io_req_normal_work_add(req
);
1402 static void __cold
io_move_task_work_from_local(struct io_ring_ctx
*ctx
)
1404 struct llist_node
*node
;
1406 node
= llist_del_all(&ctx
->work_llist
);
1408 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1412 io_req_normal_work_add(req
);
1416 static int __io_run_local_work(struct io_ring_ctx
*ctx
, struct io_tw_state
*ts
)
1418 struct llist_node
*node
;
1419 unsigned int loops
= 0;
1422 if (WARN_ON_ONCE(ctx
->submitter_task
!= current
))
1424 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
)
1425 atomic_andnot(IORING_SQ_TASKRUN
, &ctx
->rings
->sq_flags
);
1428 * llists are in reverse order, flip it back the right way before
1429 * running the pending items.
1431 node
= llist_reverse_order(io_llist_xchg(&ctx
->work_llist
, NULL
));
1433 struct llist_node
*next
= node
->next
;
1434 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1436 prefetch(container_of(next
, struct io_kiocb
, io_task_work
.node
));
1437 INDIRECT_CALL_2(req
->io_task_work
.func
,
1438 io_poll_task_func
, io_req_rw_complete
,
1445 if (!llist_empty(&ctx
->work_llist
))
1448 io_submit_flush_completions(ctx
);
1449 if (!llist_empty(&ctx
->work_llist
))
1452 trace_io_uring_local_work_run(ctx
, ret
, loops
);
1456 static inline int io_run_local_work_locked(struct io_ring_ctx
*ctx
)
1458 struct io_tw_state ts
= { .locked
= true, };
1461 if (llist_empty(&ctx
->work_llist
))
1464 ret
= __io_run_local_work(ctx
, &ts
);
1465 /* shouldn't happen! */
1466 if (WARN_ON_ONCE(!ts
.locked
))
1467 mutex_lock(&ctx
->uring_lock
);
1471 static int io_run_local_work(struct io_ring_ctx
*ctx
)
1473 struct io_tw_state ts
= {};
1476 ts
.locked
= mutex_trylock(&ctx
->uring_lock
);
1477 ret
= __io_run_local_work(ctx
, &ts
);
1479 mutex_unlock(&ctx
->uring_lock
);
1484 static void io_req_task_cancel(struct io_kiocb
*req
, struct io_tw_state
*ts
)
1486 io_tw_lock(req
->ctx
, ts
);
1487 io_req_defer_failed(req
, req
->cqe
.res
);
1490 void io_req_task_submit(struct io_kiocb
*req
, struct io_tw_state
*ts
)
1492 io_tw_lock(req
->ctx
, ts
);
1493 /* req->task == current here, checking PF_EXITING is safe */
1494 if (unlikely(req
->task
->flags
& PF_EXITING
))
1495 io_req_defer_failed(req
, -EFAULT
);
1496 else if (req
->flags
& REQ_F_FORCE_ASYNC
)
1497 io_queue_iowq(req
, ts
);
1502 void io_req_task_queue_fail(struct io_kiocb
*req
, int ret
)
1504 io_req_set_res(req
, ret
, 0);
1505 req
->io_task_work
.func
= io_req_task_cancel
;
1506 io_req_task_work_add(req
);
1509 void io_req_task_queue(struct io_kiocb
*req
)
1511 req
->io_task_work
.func
= io_req_task_submit
;
1512 io_req_task_work_add(req
);
1515 void io_queue_next(struct io_kiocb
*req
)
1517 struct io_kiocb
*nxt
= io_req_find_next(req
);
1520 io_req_task_queue(nxt
);
1523 static void io_free_batch_list(struct io_ring_ctx
*ctx
,
1524 struct io_wq_work_node
*node
)
1525 __must_hold(&ctx
->uring_lock
)
1528 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1531 if (unlikely(req
->flags
& IO_REQ_CLEAN_SLOW_FLAGS
)) {
1532 if (req
->flags
& REQ_F_REFCOUNT
) {
1533 node
= req
->comp_list
.next
;
1534 if (!req_ref_put_and_test(req
))
1537 if ((req
->flags
& REQ_F_POLLED
) && req
->apoll
) {
1538 struct async_poll
*apoll
= req
->apoll
;
1540 if (apoll
->double_poll
)
1541 kfree(apoll
->double_poll
);
1542 if (!io_alloc_cache_put(&ctx
->apoll_cache
, &apoll
->cache
))
1544 req
->flags
&= ~REQ_F_POLLED
;
1546 if (req
->flags
& IO_REQ_LINK_FLAGS
)
1548 if (unlikely(req
->flags
& IO_REQ_CLEAN_FLAGS
))
1553 io_req_put_rsrc_locked(req
, ctx
);
1555 io_put_task(req
->task
);
1556 node
= req
->comp_list
.next
;
1557 io_req_add_to_cache(req
, ctx
);
1561 void __io_submit_flush_completions(struct io_ring_ctx
*ctx
)
1562 __must_hold(&ctx
->uring_lock
)
1564 struct io_submit_state
*state
= &ctx
->submit_state
;
1565 struct io_wq_work_node
*node
;
1568 /* must come first to preserve CQE ordering in failure cases */
1569 if (state
->cqes_count
)
1570 __io_flush_post_cqes(ctx
);
1571 __wq_list_for_each(node
, &state
->compl_reqs
) {
1572 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1575 if (!(req
->flags
& REQ_F_CQE_SKIP
) &&
1576 unlikely(!io_fill_cqe_req(ctx
, req
))) {
1577 if (ctx
->lockless_cq
) {
1578 spin_lock(&ctx
->completion_lock
);
1579 io_req_cqe_overflow(req
);
1580 spin_unlock(&ctx
->completion_lock
);
1582 io_req_cqe_overflow(req
);
1586 __io_cq_unlock_post(ctx
);
1588 if (!wq_list_empty(&ctx
->submit_state
.compl_reqs
)) {
1589 io_free_batch_list(ctx
, state
->compl_reqs
.first
);
1590 INIT_WQ_LIST(&state
->compl_reqs
);
1594 static unsigned io_cqring_events(struct io_ring_ctx
*ctx
)
1596 /* See comment at the top of this file */
1598 return __io_cqring_events(ctx
);
1602 * We can't just wait for polled events to come to us, we have to actively
1603 * find and complete them.
1605 static __cold
void io_iopoll_try_reap_events(struct io_ring_ctx
*ctx
)
1607 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
1610 mutex_lock(&ctx
->uring_lock
);
1611 while (!wq_list_empty(&ctx
->iopoll_list
)) {
1612 /* let it sleep and repeat later if can't complete a request */
1613 if (io_do_iopoll(ctx
, true) == 0)
1616 * Ensure we allow local-to-the-cpu processing to take place,
1617 * in this case we need to ensure that we reap all events.
1618 * Also let task_work, etc. to progress by releasing the mutex
1620 if (need_resched()) {
1621 mutex_unlock(&ctx
->uring_lock
);
1623 mutex_lock(&ctx
->uring_lock
);
1626 mutex_unlock(&ctx
->uring_lock
);
1629 static int io_iopoll_check(struct io_ring_ctx
*ctx
, long min
)
1631 unsigned int nr_events
= 0;
1632 unsigned long check_cq
;
1634 if (!io_allowed_run_tw(ctx
))
1637 check_cq
= READ_ONCE(ctx
->check_cq
);
1638 if (unlikely(check_cq
)) {
1639 if (check_cq
& BIT(IO_CHECK_CQ_OVERFLOW_BIT
))
1640 __io_cqring_overflow_flush(ctx
);
1642 * Similarly do not spin if we have not informed the user of any
1645 if (check_cq
& BIT(IO_CHECK_CQ_DROPPED_BIT
))
1649 * Don't enter poll loop if we already have events pending.
1650 * If we do, we can potentially be spinning for commands that
1651 * already triggered a CQE (eg in error).
1653 if (io_cqring_events(ctx
))
1660 * If a submit got punted to a workqueue, we can have the
1661 * application entering polling for a command before it gets
1662 * issued. That app will hold the uring_lock for the duration
1663 * of the poll right here, so we need to take a breather every
1664 * now and then to ensure that the issue has a chance to add
1665 * the poll to the issued list. Otherwise we can spin here
1666 * forever, while the workqueue is stuck trying to acquire the
1669 if (wq_list_empty(&ctx
->iopoll_list
) ||
1670 io_task_work_pending(ctx
)) {
1671 u32 tail
= ctx
->cached_cq_tail
;
1673 (void) io_run_local_work_locked(ctx
);
1675 if (task_work_pending(current
) ||
1676 wq_list_empty(&ctx
->iopoll_list
)) {
1677 mutex_unlock(&ctx
->uring_lock
);
1679 mutex_lock(&ctx
->uring_lock
);
1681 /* some requests don't go through iopoll_list */
1682 if (tail
!= ctx
->cached_cq_tail
||
1683 wq_list_empty(&ctx
->iopoll_list
))
1686 ret
= io_do_iopoll(ctx
, !min
);
1687 if (unlikely(ret
< 0))
1690 if (task_sigpending(current
))
1696 } while (nr_events
< min
);
1701 void io_req_task_complete(struct io_kiocb
*req
, struct io_tw_state
*ts
)
1704 io_req_complete_defer(req
);
1706 io_req_complete_post(req
, IO_URING_F_UNLOCKED
);
1710 * After the iocb has been issued, it's safe to be found on the poll list.
1711 * Adding the kiocb to the list AFTER submission ensures that we don't
1712 * find it from a io_do_iopoll() thread before the issuer is done
1713 * accessing the kiocb cookie.
1715 static void io_iopoll_req_issued(struct io_kiocb
*req
, unsigned int issue_flags
)
1717 struct io_ring_ctx
*ctx
= req
->ctx
;
1718 const bool needs_lock
= issue_flags
& IO_URING_F_UNLOCKED
;
1720 /* workqueue context doesn't hold uring_lock, grab it now */
1721 if (unlikely(needs_lock
))
1722 mutex_lock(&ctx
->uring_lock
);
1725 * Track whether we have multiple files in our lists. This will impact
1726 * how we do polling eventually, not spinning if we're on potentially
1727 * different devices.
1729 if (wq_list_empty(&ctx
->iopoll_list
)) {
1730 ctx
->poll_multi_queue
= false;
1731 } else if (!ctx
->poll_multi_queue
) {
1732 struct io_kiocb
*list_req
;
1734 list_req
= container_of(ctx
->iopoll_list
.first
, struct io_kiocb
,
1736 if (list_req
->file
!= req
->file
)
1737 ctx
->poll_multi_queue
= true;
1741 * For fast devices, IO may have already completed. If it has, add
1742 * it to the front so we find it first.
1744 if (READ_ONCE(req
->iopoll_completed
))
1745 wq_list_add_head(&req
->comp_list
, &ctx
->iopoll_list
);
1747 wq_list_add_tail(&req
->comp_list
, &ctx
->iopoll_list
);
1749 if (unlikely(needs_lock
)) {
1751 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1752 * in sq thread task context or in io worker task context. If
1753 * current task context is sq thread, we don't need to check
1754 * whether should wake up sq thread.
1756 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) &&
1757 wq_has_sleeper(&ctx
->sq_data
->wait
))
1758 wake_up(&ctx
->sq_data
->wait
);
1760 mutex_unlock(&ctx
->uring_lock
);
1764 unsigned int io_file_get_flags(struct file
*file
)
1766 unsigned int res
= 0;
1768 if (S_ISREG(file_inode(file
)->i_mode
))
1770 if ((file
->f_flags
& O_NONBLOCK
) || (file
->f_mode
& FMODE_NOWAIT
))
1771 res
|= REQ_F_SUPPORT_NOWAIT
;
1775 bool io_alloc_async_data(struct io_kiocb
*req
)
1777 WARN_ON_ONCE(!io_cold_defs
[req
->opcode
].async_size
);
1778 req
->async_data
= kmalloc(io_cold_defs
[req
->opcode
].async_size
, GFP_KERNEL
);
1779 if (req
->async_data
) {
1780 req
->flags
|= REQ_F_ASYNC_DATA
;
1786 int io_req_prep_async(struct io_kiocb
*req
)
1788 const struct io_cold_def
*cdef
= &io_cold_defs
[req
->opcode
];
1789 const struct io_issue_def
*def
= &io_issue_defs
[req
->opcode
];
1791 /* assign early for deferred execution for non-fixed file */
1792 if (def
->needs_file
&& !(req
->flags
& REQ_F_FIXED_FILE
) && !req
->file
)
1793 req
->file
= io_file_get_normal(req
, req
->cqe
.fd
);
1794 if (!cdef
->prep_async
)
1796 if (WARN_ON_ONCE(req_has_async_data(req
)))
1798 if (!def
->manual_alloc
) {
1799 if (io_alloc_async_data(req
))
1802 return cdef
->prep_async(req
);
1805 static u32
io_get_sequence(struct io_kiocb
*req
)
1807 u32 seq
= req
->ctx
->cached_sq_head
;
1808 struct io_kiocb
*cur
;
1810 /* need original cached_sq_head, but it was increased for each req */
1811 io_for_each_link(cur
, req
)
1816 static __cold
void io_drain_req(struct io_kiocb
*req
)
1817 __must_hold(&ctx
->uring_lock
)
1819 struct io_ring_ctx
*ctx
= req
->ctx
;
1820 struct io_defer_entry
*de
;
1822 u32 seq
= io_get_sequence(req
);
1824 /* Still need defer if there is pending req in defer list. */
1825 spin_lock(&ctx
->completion_lock
);
1826 if (!req_need_defer(req
, seq
) && list_empty_careful(&ctx
->defer_list
)) {
1827 spin_unlock(&ctx
->completion_lock
);
1829 ctx
->drain_active
= false;
1830 io_req_task_queue(req
);
1833 spin_unlock(&ctx
->completion_lock
);
1835 io_prep_async_link(req
);
1836 de
= kmalloc(sizeof(*de
), GFP_KERNEL
);
1839 io_req_defer_failed(req
, ret
);
1843 spin_lock(&ctx
->completion_lock
);
1844 if (!req_need_defer(req
, seq
) && list_empty(&ctx
->defer_list
)) {
1845 spin_unlock(&ctx
->completion_lock
);
1850 trace_io_uring_defer(req
);
1853 list_add_tail(&de
->list
, &ctx
->defer_list
);
1854 spin_unlock(&ctx
->completion_lock
);
1857 static bool io_assign_file(struct io_kiocb
*req
, const struct io_issue_def
*def
,
1858 unsigned int issue_flags
)
1860 if (req
->file
|| !def
->needs_file
)
1863 if (req
->flags
& REQ_F_FIXED_FILE
)
1864 req
->file
= io_file_get_fixed(req
, req
->cqe
.fd
, issue_flags
);
1866 req
->file
= io_file_get_normal(req
, req
->cqe
.fd
);
1871 static int io_issue_sqe(struct io_kiocb
*req
, unsigned int issue_flags
)
1873 const struct io_issue_def
*def
= &io_issue_defs
[req
->opcode
];
1874 const struct cred
*creds
= NULL
;
1877 if (unlikely(!io_assign_file(req
, def
, issue_flags
)))
1880 if (unlikely((req
->flags
& REQ_F_CREDS
) && req
->creds
!= current_cred()))
1881 creds
= override_creds(req
->creds
);
1883 if (!def
->audit_skip
)
1884 audit_uring_entry(req
->opcode
);
1886 ret
= def
->issue(req
, issue_flags
);
1888 if (!def
->audit_skip
)
1889 audit_uring_exit(!ret
, ret
);
1892 revert_creds(creds
);
1894 if (ret
== IOU_OK
) {
1895 if (issue_flags
& IO_URING_F_COMPLETE_DEFER
)
1896 io_req_complete_defer(req
);
1898 io_req_complete_post(req
, issue_flags
);
1899 } else if (ret
!= IOU_ISSUE_SKIP_COMPLETE
)
1902 /* If the op doesn't have a file, we're not polling for it */
1903 if ((req
->ctx
->flags
& IORING_SETUP_IOPOLL
) && def
->iopoll_queue
)
1904 io_iopoll_req_issued(req
, issue_flags
);
1909 int io_poll_issue(struct io_kiocb
*req
, struct io_tw_state
*ts
)
1911 io_tw_lock(req
->ctx
, ts
);
1912 return io_issue_sqe(req
, IO_URING_F_NONBLOCK
|IO_URING_F_MULTISHOT
|
1913 IO_URING_F_COMPLETE_DEFER
);
1916 struct io_wq_work
*io_wq_free_work(struct io_wq_work
*work
)
1918 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
1919 struct io_kiocb
*nxt
= NULL
;
1921 if (req_ref_put_and_test(req
)) {
1922 if (req
->flags
& IO_REQ_LINK_FLAGS
)
1923 nxt
= io_req_find_next(req
);
1926 return nxt
? &nxt
->work
: NULL
;
1929 void io_wq_submit_work(struct io_wq_work
*work
)
1931 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
1932 const struct io_issue_def
*def
= &io_issue_defs
[req
->opcode
];
1933 unsigned int issue_flags
= IO_URING_F_UNLOCKED
| IO_URING_F_IOWQ
;
1934 bool needs_poll
= false;
1935 int ret
= 0, err
= -ECANCELED
;
1937 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1938 if (!(req
->flags
& REQ_F_REFCOUNT
))
1939 __io_req_set_refcount(req
, 2);
1943 io_arm_ltimeout(req
);
1945 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1946 if (work
->flags
& IO_WQ_WORK_CANCEL
) {
1948 io_req_task_queue_fail(req
, err
);
1951 if (!io_assign_file(req
, def
, issue_flags
)) {
1953 work
->flags
|= IO_WQ_WORK_CANCEL
;
1957 if (req
->flags
& REQ_F_FORCE_ASYNC
) {
1958 bool opcode_poll
= def
->pollin
|| def
->pollout
;
1960 if (opcode_poll
&& file_can_poll(req
->file
)) {
1962 issue_flags
|= IO_URING_F_NONBLOCK
;
1967 ret
= io_issue_sqe(req
, issue_flags
);
1972 * If REQ_F_NOWAIT is set, then don't wait or retry with
1973 * poll. -EAGAIN is final for that case.
1975 if (req
->flags
& REQ_F_NOWAIT
)
1979 * We can get EAGAIN for iopolled IO even though we're
1980 * forcing a sync submission from here, since we can't
1981 * wait for request slots on the block side.
1984 if (!(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
1986 if (io_wq_worker_stopped())
1992 if (io_arm_poll_handler(req
, issue_flags
) == IO_APOLL_OK
)
1994 /* aborted or ready, in either case retry blocking */
1996 issue_flags
&= ~IO_URING_F_NONBLOCK
;
1999 /* avoid locking problems by failing it from a clean context */
2001 io_req_task_queue_fail(req
, ret
);
2004 inline struct file
*io_file_get_fixed(struct io_kiocb
*req
, int fd
,
2005 unsigned int issue_flags
)
2007 struct io_ring_ctx
*ctx
= req
->ctx
;
2008 struct io_fixed_file
*slot
;
2009 struct file
*file
= NULL
;
2011 io_ring_submit_lock(ctx
, issue_flags
);
2013 if (unlikely((unsigned int)fd
>= ctx
->nr_user_files
))
2015 fd
= array_index_nospec(fd
, ctx
->nr_user_files
);
2016 slot
= io_fixed_file_slot(&ctx
->file_table
, fd
);
2017 file
= io_slot_file(slot
);
2018 req
->flags
|= io_slot_flags(slot
);
2019 io_req_set_rsrc_node(req
, ctx
, 0);
2021 io_ring_submit_unlock(ctx
, issue_flags
);
2025 struct file
*io_file_get_normal(struct io_kiocb
*req
, int fd
)
2027 struct file
*file
= fget(fd
);
2029 trace_io_uring_file_get(req
, fd
);
2031 /* we don't allow fixed io_uring files */
2032 if (file
&& io_is_uring_fops(file
))
2033 io_req_track_inflight(req
);
2037 static void io_queue_async(struct io_kiocb
*req
, int ret
)
2038 __must_hold(&req
->ctx
->uring_lock
)
2040 struct io_kiocb
*linked_timeout
;
2042 if (ret
!= -EAGAIN
|| (req
->flags
& REQ_F_NOWAIT
)) {
2043 io_req_defer_failed(req
, ret
);
2047 linked_timeout
= io_prep_linked_timeout(req
);
2049 switch (io_arm_poll_handler(req
, 0)) {
2050 case IO_APOLL_READY
:
2051 io_kbuf_recycle(req
, 0);
2052 io_req_task_queue(req
);
2054 case IO_APOLL_ABORTED
:
2055 io_kbuf_recycle(req
, 0);
2056 io_queue_iowq(req
, NULL
);
2063 io_queue_linked_timeout(linked_timeout
);
2066 static inline void io_queue_sqe(struct io_kiocb
*req
)
2067 __must_hold(&req
->ctx
->uring_lock
)
2071 ret
= io_issue_sqe(req
, IO_URING_F_NONBLOCK
|IO_URING_F_COMPLETE_DEFER
);
2074 * We async punt it if the file wasn't marked NOWAIT, or if the file
2075 * doesn't support non-blocking read/write attempts
2078 io_arm_ltimeout(req
);
2080 io_queue_async(req
, ret
);
2083 static void io_queue_sqe_fallback(struct io_kiocb
*req
)
2084 __must_hold(&req
->ctx
->uring_lock
)
2086 if (unlikely(req
->flags
& REQ_F_FAIL
)) {
2088 * We don't submit, fail them all, for that replace hardlinks
2089 * with normal links. Extra REQ_F_LINK is tolerated.
2091 req
->flags
&= ~REQ_F_HARDLINK
;
2092 req
->flags
|= REQ_F_LINK
;
2093 io_req_defer_failed(req
, req
->cqe
.res
);
2095 int ret
= io_req_prep_async(req
);
2097 if (unlikely(ret
)) {
2098 io_req_defer_failed(req
, ret
);
2102 if (unlikely(req
->ctx
->drain_active
))
2105 io_queue_iowq(req
, NULL
);
2110 * Check SQE restrictions (opcode and flags).
2112 * Returns 'true' if SQE is allowed, 'false' otherwise.
2114 static inline bool io_check_restriction(struct io_ring_ctx
*ctx
,
2115 struct io_kiocb
*req
,
2116 unsigned int sqe_flags
)
2118 if (!test_bit(req
->opcode
, ctx
->restrictions
.sqe_op
))
2121 if ((sqe_flags
& ctx
->restrictions
.sqe_flags_required
) !=
2122 ctx
->restrictions
.sqe_flags_required
)
2125 if (sqe_flags
& ~(ctx
->restrictions
.sqe_flags_allowed
|
2126 ctx
->restrictions
.sqe_flags_required
))
2132 static void io_init_req_drain(struct io_kiocb
*req
)
2134 struct io_ring_ctx
*ctx
= req
->ctx
;
2135 struct io_kiocb
*head
= ctx
->submit_state
.link
.head
;
2137 ctx
->drain_active
= true;
2140 * If we need to drain a request in the middle of a link, drain
2141 * the head request and the next request/link after the current
2142 * link. Considering sequential execution of links,
2143 * REQ_F_IO_DRAIN will be maintained for every request of our
2146 head
->flags
|= REQ_F_IO_DRAIN
| REQ_F_FORCE_ASYNC
;
2147 ctx
->drain_next
= true;
2151 static int io_init_req(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
2152 const struct io_uring_sqe
*sqe
)
2153 __must_hold(&ctx
->uring_lock
)
2155 const struct io_issue_def
*def
;
2156 unsigned int sqe_flags
;
2160 /* req is partially pre-initialised, see io_preinit_req() */
2161 req
->opcode
= opcode
= READ_ONCE(sqe
->opcode
);
2162 /* same numerical values with corresponding REQ_F_*, safe to copy */
2163 req
->flags
= sqe_flags
= READ_ONCE(sqe
->flags
);
2164 req
->cqe
.user_data
= READ_ONCE(sqe
->user_data
);
2166 req
->rsrc_node
= NULL
;
2167 req
->task
= current
;
2169 if (unlikely(opcode
>= IORING_OP_LAST
)) {
2173 def
= &io_issue_defs
[opcode
];
2174 if (unlikely(sqe_flags
& ~SQE_COMMON_FLAGS
)) {
2175 /* enforce forwards compatibility on users */
2176 if (sqe_flags
& ~SQE_VALID_FLAGS
)
2178 if (sqe_flags
& IOSQE_BUFFER_SELECT
) {
2179 if (!def
->buffer_select
)
2181 req
->buf_index
= READ_ONCE(sqe
->buf_group
);
2183 if (sqe_flags
& IOSQE_CQE_SKIP_SUCCESS
)
2184 ctx
->drain_disabled
= true;
2185 if (sqe_flags
& IOSQE_IO_DRAIN
) {
2186 if (ctx
->drain_disabled
)
2188 io_init_req_drain(req
);
2191 if (unlikely(ctx
->restricted
|| ctx
->drain_active
|| ctx
->drain_next
)) {
2192 if (ctx
->restricted
&& !io_check_restriction(ctx
, req
, sqe_flags
))
2194 /* knock it to the slow queue path, will be drained there */
2195 if (ctx
->drain_active
)
2196 req
->flags
|= REQ_F_FORCE_ASYNC
;
2197 /* if there is no link, we're at "next" request and need to drain */
2198 if (unlikely(ctx
->drain_next
) && !ctx
->submit_state
.link
.head
) {
2199 ctx
->drain_next
= false;
2200 ctx
->drain_active
= true;
2201 req
->flags
|= REQ_F_IO_DRAIN
| REQ_F_FORCE_ASYNC
;
2205 if (!def
->ioprio
&& sqe
->ioprio
)
2207 if (!def
->iopoll
&& (ctx
->flags
& IORING_SETUP_IOPOLL
))
2210 if (def
->needs_file
) {
2211 struct io_submit_state
*state
= &ctx
->submit_state
;
2213 req
->cqe
.fd
= READ_ONCE(sqe
->fd
);
2216 * Plug now if we have more than 2 IO left after this, and the
2217 * target is potentially a read/write to block based storage.
2219 if (state
->need_plug
&& def
->plug
) {
2220 state
->plug_started
= true;
2221 state
->need_plug
= false;
2222 blk_start_plug_nr_ios(&state
->plug
, state
->submit_nr
);
2226 personality
= READ_ONCE(sqe
->personality
);
2230 req
->creds
= xa_load(&ctx
->personalities
, personality
);
2233 get_cred(req
->creds
);
2234 ret
= security_uring_override_creds(req
->creds
);
2236 put_cred(req
->creds
);
2239 req
->flags
|= REQ_F_CREDS
;
2242 return def
->prep(req
, sqe
);
2245 static __cold
int io_submit_fail_init(const struct io_uring_sqe
*sqe
,
2246 struct io_kiocb
*req
, int ret
)
2248 struct io_ring_ctx
*ctx
= req
->ctx
;
2249 struct io_submit_link
*link
= &ctx
->submit_state
.link
;
2250 struct io_kiocb
*head
= link
->head
;
2252 trace_io_uring_req_failed(sqe
, req
, ret
);
2255 * Avoid breaking links in the middle as it renders links with SQPOLL
2256 * unusable. Instead of failing eagerly, continue assembling the link if
2257 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2258 * should find the flag and handle the rest.
2260 req_fail_link_node(req
, ret
);
2261 if (head
&& !(head
->flags
& REQ_F_FAIL
))
2262 req_fail_link_node(head
, -ECANCELED
);
2264 if (!(req
->flags
& IO_REQ_LINK_FLAGS
)) {
2266 link
->last
->link
= req
;
2270 io_queue_sqe_fallback(req
);
2275 link
->last
->link
= req
;
2282 static inline int io_submit_sqe(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
2283 const struct io_uring_sqe
*sqe
)
2284 __must_hold(&ctx
->uring_lock
)
2286 struct io_submit_link
*link
= &ctx
->submit_state
.link
;
2289 ret
= io_init_req(ctx
, req
, sqe
);
2291 return io_submit_fail_init(sqe
, req
, ret
);
2293 trace_io_uring_submit_req(req
);
2296 * If we already have a head request, queue this one for async
2297 * submittal once the head completes. If we don't have a head but
2298 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2299 * submitted sync once the chain is complete. If none of those
2300 * conditions are true (normal request), then just queue it.
2302 if (unlikely(link
->head
)) {
2303 ret
= io_req_prep_async(req
);
2305 return io_submit_fail_init(sqe
, req
, ret
);
2307 trace_io_uring_link(req
, link
->head
);
2308 link
->last
->link
= req
;
2311 if (req
->flags
& IO_REQ_LINK_FLAGS
)
2313 /* last request of the link, flush it */
2316 if (req
->flags
& (REQ_F_FORCE_ASYNC
| REQ_F_FAIL
))
2319 } else if (unlikely(req
->flags
& (IO_REQ_LINK_FLAGS
|
2320 REQ_F_FORCE_ASYNC
| REQ_F_FAIL
))) {
2321 if (req
->flags
& IO_REQ_LINK_FLAGS
) {
2326 io_queue_sqe_fallback(req
);
2336 * Batched submission is done, ensure local IO is flushed out.
2338 static void io_submit_state_end(struct io_ring_ctx
*ctx
)
2340 struct io_submit_state
*state
= &ctx
->submit_state
;
2342 if (unlikely(state
->link
.head
))
2343 io_queue_sqe_fallback(state
->link
.head
);
2344 /* flush only after queuing links as they can generate completions */
2345 io_submit_flush_completions(ctx
);
2346 if (state
->plug_started
)
2347 blk_finish_plug(&state
->plug
);
2351 * Start submission side cache.
2353 static void io_submit_state_start(struct io_submit_state
*state
,
2354 unsigned int max_ios
)
2356 state
->plug_started
= false;
2357 state
->need_plug
= max_ios
> 2;
2358 state
->submit_nr
= max_ios
;
2359 /* set only head, no need to init link_last in advance */
2360 state
->link
.head
= NULL
;
2363 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
2365 struct io_rings
*rings
= ctx
->rings
;
2368 * Ensure any loads from the SQEs are done at this point,
2369 * since once we write the new head, the application could
2370 * write new data to them.
2372 smp_store_release(&rings
->sq
.head
, ctx
->cached_sq_head
);
2376 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2377 * that is mapped by userspace. This means that care needs to be taken to
2378 * ensure that reads are stable, as we cannot rely on userspace always
2379 * being a good citizen. If members of the sqe are validated and then later
2380 * used, it's important that those reads are done through READ_ONCE() to
2381 * prevent a re-load down the line.
2383 static bool io_get_sqe(struct io_ring_ctx
*ctx
, const struct io_uring_sqe
**sqe
)
2385 unsigned mask
= ctx
->sq_entries
- 1;
2386 unsigned head
= ctx
->cached_sq_head
++ & mask
;
2388 if (!(ctx
->flags
& IORING_SETUP_NO_SQARRAY
)) {
2389 head
= READ_ONCE(ctx
->sq_array
[head
]);
2390 if (unlikely(head
>= ctx
->sq_entries
)) {
2391 /* drop invalid entries */
2392 spin_lock(&ctx
->completion_lock
);
2394 spin_unlock(&ctx
->completion_lock
);
2395 WRITE_ONCE(ctx
->rings
->sq_dropped
,
2396 READ_ONCE(ctx
->rings
->sq_dropped
) + 1);
2402 * The cached sq head (or cq tail) serves two purposes:
2404 * 1) allows us to batch the cost of updating the user visible
2406 * 2) allows the kernel side to track the head on its own, even
2407 * though the application is the one updating it.
2410 /* double index for 128-byte SQEs, twice as long */
2411 if (ctx
->flags
& IORING_SETUP_SQE128
)
2413 *sqe
= &ctx
->sq_sqes
[head
];
2417 int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
)
2418 __must_hold(&ctx
->uring_lock
)
2420 unsigned int entries
= io_sqring_entries(ctx
);
2424 if (unlikely(!entries
))
2426 /* make sure SQ entry isn't read before tail */
2427 ret
= left
= min(nr
, entries
);
2428 io_get_task_refs(left
);
2429 io_submit_state_start(&ctx
->submit_state
, left
);
2432 const struct io_uring_sqe
*sqe
;
2433 struct io_kiocb
*req
;
2435 if (unlikely(!io_alloc_req(ctx
, &req
)))
2437 if (unlikely(!io_get_sqe(ctx
, &sqe
))) {
2438 io_req_add_to_cache(req
, ctx
);
2443 * Continue submitting even for sqe failure if the
2444 * ring was setup with IORING_SETUP_SUBMIT_ALL
2446 if (unlikely(io_submit_sqe(ctx
, req
, sqe
)) &&
2447 !(ctx
->flags
& IORING_SETUP_SUBMIT_ALL
)) {
2453 if (unlikely(left
)) {
2455 /* try again if it submitted nothing and can't allocate a req */
2456 if (!ret
&& io_req_cache_empty(ctx
))
2458 current
->io_uring
->cached_refs
+= left
;
2461 io_submit_state_end(ctx
);
2462 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2463 io_commit_sqring(ctx
);
2467 struct io_wait_queue
{
2468 struct wait_queue_entry wq
;
2469 struct io_ring_ctx
*ctx
;
2471 unsigned nr_timeouts
;
2475 static inline bool io_has_work(struct io_ring_ctx
*ctx
)
2477 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
) ||
2478 !llist_empty(&ctx
->work_llist
);
2481 static inline bool io_should_wake(struct io_wait_queue
*iowq
)
2483 struct io_ring_ctx
*ctx
= iowq
->ctx
;
2484 int dist
= READ_ONCE(ctx
->rings
->cq
.tail
) - (int) iowq
->cq_tail
;
2487 * Wake up if we have enough events, or if a timeout occurred since we
2488 * started waiting. For timeouts, we always want to return to userspace,
2489 * regardless of event count.
2491 return dist
>= 0 || atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
2494 static int io_wake_function(struct wait_queue_entry
*curr
, unsigned int mode
,
2495 int wake_flags
, void *key
)
2497 struct io_wait_queue
*iowq
= container_of(curr
, struct io_wait_queue
, wq
);
2500 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2501 * the task, and the next invocation will do it.
2503 if (io_should_wake(iowq
) || io_has_work(iowq
->ctx
))
2504 return autoremove_wake_function(curr
, mode
, wake_flags
, key
);
2508 int io_run_task_work_sig(struct io_ring_ctx
*ctx
)
2510 if (!llist_empty(&ctx
->work_llist
)) {
2511 __set_current_state(TASK_RUNNING
);
2512 if (io_run_local_work(ctx
) > 0)
2515 if (io_run_task_work() > 0)
2517 if (task_sigpending(current
))
2522 static bool current_pending_io(void)
2524 struct io_uring_task
*tctx
= current
->io_uring
;
2528 return percpu_counter_read_positive(&tctx
->inflight
);
2531 /* when returns >0, the caller should retry */
2532 static inline int io_cqring_wait_schedule(struct io_ring_ctx
*ctx
,
2533 struct io_wait_queue
*iowq
)
2537 if (unlikely(READ_ONCE(ctx
->check_cq
)))
2539 if (unlikely(!llist_empty(&ctx
->work_llist
)))
2541 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL
)))
2543 if (unlikely(task_sigpending(current
)))
2545 if (unlikely(io_should_wake(iowq
)))
2549 * Mark us as being in io_wait if we have pending requests, so cpufreq
2550 * can take into account that the task is waiting for IO - turns out
2551 * to be important for low QD IO.
2553 io_wait
= current
->in_iowait
;
2554 if (current_pending_io())
2555 current
->in_iowait
= 1;
2557 if (iowq
->timeout
== KTIME_MAX
)
2559 else if (!schedule_hrtimeout(&iowq
->timeout
, HRTIMER_MODE_ABS
))
2561 current
->in_iowait
= io_wait
;
2566 * Wait until events become available, if we don't already have some. The
2567 * application must reap them itself, as they reside on the shared cq ring.
2569 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
2570 const sigset_t __user
*sig
, size_t sigsz
,
2571 struct __kernel_timespec __user
*uts
)
2573 struct io_wait_queue iowq
;
2574 struct io_rings
*rings
= ctx
->rings
;
2577 if (!io_allowed_run_tw(ctx
))
2579 if (!llist_empty(&ctx
->work_llist
))
2580 io_run_local_work(ctx
);
2582 io_cqring_overflow_flush(ctx
);
2583 /* if user messes with these they will just get an early return */
2584 if (__io_cqring_events_user(ctx
) >= min_events
)
2588 #ifdef CONFIG_COMPAT
2589 if (in_compat_syscall())
2590 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
2594 ret
= set_user_sigmask(sig
, sigsz
);
2600 init_waitqueue_func_entry(&iowq
.wq
, io_wake_function
);
2601 iowq
.wq
.private = current
;
2602 INIT_LIST_HEAD(&iowq
.wq
.entry
);
2604 iowq
.nr_timeouts
= atomic_read(&ctx
->cq_timeouts
);
2605 iowq
.cq_tail
= READ_ONCE(ctx
->rings
->cq
.head
) + min_events
;
2606 iowq
.timeout
= KTIME_MAX
;
2609 struct timespec64 ts
;
2611 if (get_timespec64(&ts
, uts
))
2613 iowq
.timeout
= ktime_add_ns(timespec64_to_ktime(ts
), ktime_get_ns());
2616 trace_io_uring_cqring_wait(ctx
, min_events
);
2618 unsigned long check_cq
;
2620 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) {
2621 int nr_wait
= (int) iowq
.cq_tail
- READ_ONCE(ctx
->rings
->cq
.tail
);
2623 atomic_set(&ctx
->cq_wait_nr
, nr_wait
);
2624 set_current_state(TASK_INTERRUPTIBLE
);
2626 prepare_to_wait_exclusive(&ctx
->cq_wait
, &iowq
.wq
,
2627 TASK_INTERRUPTIBLE
);
2630 ret
= io_cqring_wait_schedule(ctx
, &iowq
);
2631 __set_current_state(TASK_RUNNING
);
2632 atomic_set(&ctx
->cq_wait_nr
, 0);
2637 * Run task_work after scheduling and before io_should_wake().
2638 * If we got woken because of task_work being processed, run it
2639 * now rather than let the caller do another wait loop.
2642 if (!llist_empty(&ctx
->work_llist
))
2643 io_run_local_work(ctx
);
2645 check_cq
= READ_ONCE(ctx
->check_cq
);
2646 if (unlikely(check_cq
)) {
2647 /* let the caller flush overflows, retry */
2648 if (check_cq
& BIT(IO_CHECK_CQ_OVERFLOW_BIT
))
2649 io_cqring_do_overflow_flush(ctx
);
2650 if (check_cq
& BIT(IO_CHECK_CQ_DROPPED_BIT
)) {
2656 if (io_should_wake(&iowq
)) {
2663 if (!(ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
))
2664 finish_wait(&ctx
->cq_wait
, &iowq
.wq
);
2665 restore_saved_sigmask_unless(ret
== -EINTR
);
2667 return READ_ONCE(rings
->cq
.head
) == READ_ONCE(rings
->cq
.tail
) ? ret
: 0;
2670 void io_mem_free(void *ptr
)
2675 folio_put(virt_to_folio(ptr
));
2678 static void io_pages_free(struct page
***pages
, int npages
)
2680 struct page
**page_array
;
2686 page_array
= *pages
;
2690 for (i
= 0; i
< npages
; i
++)
2691 unpin_user_page(page_array
[i
]);
2696 static void *__io_uaddr_map(struct page
***pages
, unsigned short *npages
,
2697 unsigned long uaddr
, size_t size
)
2699 struct page
**page_array
;
2700 unsigned int nr_pages
;
2706 if (uaddr
& (PAGE_SIZE
- 1) || !size
)
2707 return ERR_PTR(-EINVAL
);
2709 nr_pages
= (size
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2710 if (nr_pages
> USHRT_MAX
)
2711 return ERR_PTR(-EINVAL
);
2712 page_array
= kvmalloc_array(nr_pages
, sizeof(struct page
*), GFP_KERNEL
);
2714 return ERR_PTR(-ENOMEM
);
2716 ret
= pin_user_pages_fast(uaddr
, nr_pages
, FOLL_WRITE
| FOLL_LONGTERM
,
2718 if (ret
!= nr_pages
) {
2720 io_pages_free(&page_array
, ret
> 0 ? ret
: 0);
2721 return ret
< 0 ? ERR_PTR(ret
) : ERR_PTR(-EFAULT
);
2724 page_addr
= page_address(page_array
[0]);
2725 for (i
= 0; i
< nr_pages
; i
++) {
2729 * Can't support mapping user allocated ring memory on 32-bit
2730 * archs where it could potentially reside in highmem. Just
2731 * fail those with -EINVAL, just like we did on kernels that
2732 * didn't support this feature.
2734 if (PageHighMem(page_array
[i
]))
2738 * No support for discontig pages for now, should either be a
2739 * single normal page, or a huge page. Later on we can add
2740 * support for remapping discontig pages, for now we will
2741 * just fail them with EINVAL.
2743 if (page_address(page_array
[i
]) != page_addr
)
2745 page_addr
+= PAGE_SIZE
;
2748 *pages
= page_array
;
2750 return page_to_virt(page_array
[0]);
2753 static void *io_rings_map(struct io_ring_ctx
*ctx
, unsigned long uaddr
,
2756 return __io_uaddr_map(&ctx
->ring_pages
, &ctx
->n_ring_pages
, uaddr
,
2760 static void *io_sqes_map(struct io_ring_ctx
*ctx
, unsigned long uaddr
,
2763 return __io_uaddr_map(&ctx
->sqe_pages
, &ctx
->n_sqe_pages
, uaddr
,
2767 static void io_rings_free(struct io_ring_ctx
*ctx
)
2769 if (!(ctx
->flags
& IORING_SETUP_NO_MMAP
)) {
2770 io_mem_free(ctx
->rings
);
2771 io_mem_free(ctx
->sq_sqes
);
2773 ctx
->sq_sqes
= NULL
;
2775 io_pages_free(&ctx
->ring_pages
, ctx
->n_ring_pages
);
2776 ctx
->n_ring_pages
= 0;
2777 io_pages_free(&ctx
->sqe_pages
, ctx
->n_sqe_pages
);
2778 ctx
->n_sqe_pages
= 0;
2782 void *io_mem_alloc(size_t size
)
2784 gfp_t gfp
= GFP_KERNEL_ACCOUNT
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
;
2787 ret
= (void *) __get_free_pages(gfp
, get_order(size
));
2790 return ERR_PTR(-ENOMEM
);
2793 static unsigned long rings_size(struct io_ring_ctx
*ctx
, unsigned int sq_entries
,
2794 unsigned int cq_entries
, size_t *sq_offset
)
2796 struct io_rings
*rings
;
2797 size_t off
, sq_array_size
;
2799 off
= struct_size(rings
, cqes
, cq_entries
);
2800 if (off
== SIZE_MAX
)
2802 if (ctx
->flags
& IORING_SETUP_CQE32
) {
2803 if (check_shl_overflow(off
, 1, &off
))
2808 off
= ALIGN(off
, SMP_CACHE_BYTES
);
2813 if (ctx
->flags
& IORING_SETUP_NO_SQARRAY
) {
2815 *sq_offset
= SIZE_MAX
;
2822 sq_array_size
= array_size(sizeof(u32
), sq_entries
);
2823 if (sq_array_size
== SIZE_MAX
)
2826 if (check_add_overflow(off
, sq_array_size
, &off
))
2832 static int io_eventfd_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
2833 unsigned int eventfd_async
)
2835 struct io_ev_fd
*ev_fd
;
2836 __s32 __user
*fds
= arg
;
2839 ev_fd
= rcu_dereference_protected(ctx
->io_ev_fd
,
2840 lockdep_is_held(&ctx
->uring_lock
));
2844 if (copy_from_user(&fd
, fds
, sizeof(*fds
)))
2847 ev_fd
= kmalloc(sizeof(*ev_fd
), GFP_KERNEL
);
2851 ev_fd
->cq_ev_fd
= eventfd_ctx_fdget(fd
);
2852 if (IS_ERR(ev_fd
->cq_ev_fd
)) {
2853 int ret
= PTR_ERR(ev_fd
->cq_ev_fd
);
2858 spin_lock(&ctx
->completion_lock
);
2859 ctx
->evfd_last_cq_tail
= ctx
->cached_cq_tail
;
2860 spin_unlock(&ctx
->completion_lock
);
2862 ev_fd
->eventfd_async
= eventfd_async
;
2863 ctx
->has_evfd
= true;
2864 rcu_assign_pointer(ctx
->io_ev_fd
, ev_fd
);
2865 atomic_set(&ev_fd
->refs
, 1);
2866 atomic_set(&ev_fd
->ops
, 0);
2870 static int io_eventfd_unregister(struct io_ring_ctx
*ctx
)
2872 struct io_ev_fd
*ev_fd
;
2874 ev_fd
= rcu_dereference_protected(ctx
->io_ev_fd
,
2875 lockdep_is_held(&ctx
->uring_lock
));
2877 ctx
->has_evfd
= false;
2878 rcu_assign_pointer(ctx
->io_ev_fd
, NULL
);
2879 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT
), &ev_fd
->ops
))
2880 call_rcu(&ev_fd
->rcu
, io_eventfd_ops
);
2887 static void io_req_caches_free(struct io_ring_ctx
*ctx
)
2889 struct io_kiocb
*req
;
2892 mutex_lock(&ctx
->uring_lock
);
2893 io_flush_cached_locked_reqs(ctx
, &ctx
->submit_state
);
2895 while (!io_req_cache_empty(ctx
)) {
2896 req
= io_extract_req(ctx
);
2897 kmem_cache_free(req_cachep
, req
);
2901 percpu_ref_put_many(&ctx
->refs
, nr
);
2902 mutex_unlock(&ctx
->uring_lock
);
2905 static void io_rsrc_node_cache_free(struct io_cache_entry
*entry
)
2907 kfree(container_of(entry
, struct io_rsrc_node
, cache
));
2910 static __cold
void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
2912 io_sq_thread_finish(ctx
);
2913 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2914 if (WARN_ON_ONCE(!list_empty(&ctx
->rsrc_ref_list
)))
2917 mutex_lock(&ctx
->uring_lock
);
2919 __io_sqe_buffers_unregister(ctx
);
2921 __io_sqe_files_unregister(ctx
);
2922 io_cqring_overflow_kill(ctx
);
2923 io_eventfd_unregister(ctx
);
2924 io_alloc_cache_free(&ctx
->apoll_cache
, io_apoll_cache_free
);
2925 io_alloc_cache_free(&ctx
->netmsg_cache
, io_netmsg_cache_free
);
2926 io_futex_cache_free(ctx
);
2927 io_destroy_buffers(ctx
);
2928 mutex_unlock(&ctx
->uring_lock
);
2930 put_cred(ctx
->sq_creds
);
2931 if (ctx
->submitter_task
)
2932 put_task_struct(ctx
->submitter_task
);
2934 /* there are no registered resources left, nobody uses it */
2936 io_rsrc_node_destroy(ctx
, ctx
->rsrc_node
);
2938 WARN_ON_ONCE(!list_empty(&ctx
->rsrc_ref_list
));
2940 #if defined(CONFIG_UNIX)
2941 if (ctx
->ring_sock
) {
2942 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
2943 sock_release(ctx
->ring_sock
);
2946 WARN_ON_ONCE(!list_empty(&ctx
->ltimeout_list
));
2948 io_alloc_cache_free(&ctx
->rsrc_node_cache
, io_rsrc_node_cache_free
);
2949 if (ctx
->mm_account
) {
2950 mmdrop(ctx
->mm_account
);
2951 ctx
->mm_account
= NULL
;
2954 io_kbuf_mmap_list_free(ctx
);
2956 percpu_ref_exit(&ctx
->refs
);
2957 free_uid(ctx
->user
);
2958 io_req_caches_free(ctx
);
2960 io_wq_put_hash(ctx
->hash_map
);
2961 kfree(ctx
->cancel_table
.hbs
);
2962 kfree(ctx
->cancel_table_locked
.hbs
);
2964 xa_destroy(&ctx
->io_bl_xa
);
2968 static __cold
void io_activate_pollwq_cb(struct callback_head
*cb
)
2970 struct io_ring_ctx
*ctx
= container_of(cb
, struct io_ring_ctx
,
2973 mutex_lock(&ctx
->uring_lock
);
2974 ctx
->poll_activated
= true;
2975 mutex_unlock(&ctx
->uring_lock
);
2978 * Wake ups for some events between start of polling and activation
2979 * might've been lost due to loose synchronisation.
2981 wake_up_all(&ctx
->poll_wq
);
2982 percpu_ref_put(&ctx
->refs
);
2985 static __cold
void io_activate_pollwq(struct io_ring_ctx
*ctx
)
2987 spin_lock(&ctx
->completion_lock
);
2988 /* already activated or in progress */
2989 if (ctx
->poll_activated
|| ctx
->poll_wq_task_work
.func
)
2991 if (WARN_ON_ONCE(!ctx
->task_complete
))
2993 if (!ctx
->submitter_task
)
2996 * with ->submitter_task only the submitter task completes requests, we
2997 * only need to sync with it, which is done by injecting a tw
2999 init_task_work(&ctx
->poll_wq_task_work
, io_activate_pollwq_cb
);
3000 percpu_ref_get(&ctx
->refs
);
3001 if (task_work_add(ctx
->submitter_task
, &ctx
->poll_wq_task_work
, TWA_SIGNAL
))
3002 percpu_ref_put(&ctx
->refs
);
3004 spin_unlock(&ctx
->completion_lock
);
3007 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
3009 struct io_ring_ctx
*ctx
= file
->private_data
;
3012 if (unlikely(!ctx
->poll_activated
))
3013 io_activate_pollwq(ctx
);
3015 poll_wait(file
, &ctx
->poll_wq
, wait
);
3017 * synchronizes with barrier from wq_has_sleeper call in
3021 if (!io_sqring_full(ctx
))
3022 mask
|= EPOLLOUT
| EPOLLWRNORM
;
3025 * Don't flush cqring overflow list here, just do a simple check.
3026 * Otherwise there could possible be ABBA deadlock:
3029 * lock(&ctx->uring_lock);
3031 * lock(&ctx->uring_lock);
3034 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
3035 * pushes them to do the flush.
3038 if (__io_cqring_events_user(ctx
) || io_has_work(ctx
))
3039 mask
|= EPOLLIN
| EPOLLRDNORM
;
3044 static int io_unregister_personality(struct io_ring_ctx
*ctx
, unsigned id
)
3046 const struct cred
*creds
;
3048 creds
= xa_erase(&ctx
->personalities
, id
);
3057 struct io_tctx_exit
{
3058 struct callback_head task_work
;
3059 struct completion completion
;
3060 struct io_ring_ctx
*ctx
;
3063 static __cold
void io_tctx_exit_cb(struct callback_head
*cb
)
3065 struct io_uring_task
*tctx
= current
->io_uring
;
3066 struct io_tctx_exit
*work
;
3068 work
= container_of(cb
, struct io_tctx_exit
, task_work
);
3070 * When @in_cancel, we're in cancellation and it's racy to remove the
3071 * node. It'll be removed by the end of cancellation, just ignore it.
3072 * tctx can be NULL if the queueing of this task_work raced with
3073 * work cancelation off the exec path.
3075 if (tctx
&& !atomic_read(&tctx
->in_cancel
))
3076 io_uring_del_tctx_node((unsigned long)work
->ctx
);
3077 complete(&work
->completion
);
3080 static __cold
bool io_cancel_ctx_cb(struct io_wq_work
*work
, void *data
)
3082 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
3084 return req
->ctx
== data
;
3087 static __cold
void io_ring_exit_work(struct work_struct
*work
)
3089 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
, exit_work
);
3090 unsigned long timeout
= jiffies
+ HZ
* 60 * 5;
3091 unsigned long interval
= HZ
/ 20;
3092 struct io_tctx_exit exit
;
3093 struct io_tctx_node
*node
;
3097 * If we're doing polled IO and end up having requests being
3098 * submitted async (out-of-line), then completions can come in while
3099 * we're waiting for refs to drop. We need to reap these manually,
3100 * as nobody else will be looking for them.
3103 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
)) {
3104 mutex_lock(&ctx
->uring_lock
);
3105 io_cqring_overflow_kill(ctx
);
3106 mutex_unlock(&ctx
->uring_lock
);
3109 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
)
3110 io_move_task_work_from_local(ctx
);
3112 while (io_uring_try_cancel_requests(ctx
, NULL
, true))
3116 struct io_sq_data
*sqd
= ctx
->sq_data
;
3117 struct task_struct
*tsk
;
3119 io_sq_thread_park(sqd
);
3121 if (tsk
&& tsk
->io_uring
&& tsk
->io_uring
->io_wq
)
3122 io_wq_cancel_cb(tsk
->io_uring
->io_wq
,
3123 io_cancel_ctx_cb
, ctx
, true);
3124 io_sq_thread_unpark(sqd
);
3127 io_req_caches_free(ctx
);
3129 if (WARN_ON_ONCE(time_after(jiffies
, timeout
))) {
3130 /* there is little hope left, don't run it too often */
3134 * This is really an uninterruptible wait, as it has to be
3135 * complete. But it's also run from a kworker, which doesn't
3136 * take signals, so it's fine to make it interruptible. This
3137 * avoids scenarios where we knowingly can wait much longer
3138 * on completions, for example if someone does a SIGSTOP on
3139 * a task that needs to finish task_work to make this loop
3140 * complete. That's a synthetic situation that should not
3141 * cause a stuck task backtrace, and hence a potential panic
3142 * on stuck tasks if that is enabled.
3144 } while (!wait_for_completion_interruptible_timeout(&ctx
->ref_comp
, interval
));
3146 init_completion(&exit
.completion
);
3147 init_task_work(&exit
.task_work
, io_tctx_exit_cb
);
3150 * Some may use context even when all refs and requests have been put,
3151 * and they are free to do so while still holding uring_lock or
3152 * completion_lock, see io_req_task_submit(). Apart from other work,
3153 * this lock/unlock section also waits them to finish.
3155 mutex_lock(&ctx
->uring_lock
);
3156 while (!list_empty(&ctx
->tctx_list
)) {
3157 WARN_ON_ONCE(time_after(jiffies
, timeout
));
3159 node
= list_first_entry(&ctx
->tctx_list
, struct io_tctx_node
,
3161 /* don't spin on a single task if cancellation failed */
3162 list_rotate_left(&ctx
->tctx_list
);
3163 ret
= task_work_add(node
->task
, &exit
.task_work
, TWA_SIGNAL
);
3164 if (WARN_ON_ONCE(ret
))
3167 mutex_unlock(&ctx
->uring_lock
);
3169 * See comment above for
3170 * wait_for_completion_interruptible_timeout() on why this
3171 * wait is marked as interruptible.
3173 wait_for_completion_interruptible(&exit
.completion
);
3174 mutex_lock(&ctx
->uring_lock
);
3176 mutex_unlock(&ctx
->uring_lock
);
3177 spin_lock(&ctx
->completion_lock
);
3178 spin_unlock(&ctx
->completion_lock
);
3180 /* pairs with RCU read section in io_req_local_work_add() */
3181 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
)
3184 io_ring_ctx_free(ctx
);
3187 static __cold
void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
3189 unsigned long index
;
3190 struct creds
*creds
;
3192 mutex_lock(&ctx
->uring_lock
);
3193 percpu_ref_kill(&ctx
->refs
);
3194 xa_for_each(&ctx
->personalities
, index
, creds
)
3195 io_unregister_personality(ctx
, index
);
3197 io_poll_remove_all(ctx
, NULL
, true);
3198 mutex_unlock(&ctx
->uring_lock
);
3201 * If we failed setting up the ctx, we might not have any rings
3202 * and therefore did not submit any requests
3205 io_kill_timeouts(ctx
, NULL
, true);
3207 flush_delayed_work(&ctx
->fallback_work
);
3209 INIT_WORK(&ctx
->exit_work
, io_ring_exit_work
);
3211 * Use system_unbound_wq to avoid spawning tons of event kworkers
3212 * if we're exiting a ton of rings at the same time. It just adds
3213 * noise and overhead, there's no discernable change in runtime
3214 * over using system_wq.
3216 queue_work(system_unbound_wq
, &ctx
->exit_work
);
3219 static int io_uring_release(struct inode
*inode
, struct file
*file
)
3221 struct io_ring_ctx
*ctx
= file
->private_data
;
3223 file
->private_data
= NULL
;
3224 io_ring_ctx_wait_and_kill(ctx
);
3228 struct io_task_cancel
{
3229 struct task_struct
*task
;
3233 static bool io_cancel_task_cb(struct io_wq_work
*work
, void *data
)
3235 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
3236 struct io_task_cancel
*cancel
= data
;
3238 return io_match_task_safe(req
, cancel
->task
, cancel
->all
);
3241 static __cold
bool io_cancel_defer_files(struct io_ring_ctx
*ctx
,
3242 struct task_struct
*task
,
3245 struct io_defer_entry
*de
;
3248 spin_lock(&ctx
->completion_lock
);
3249 list_for_each_entry_reverse(de
, &ctx
->defer_list
, list
) {
3250 if (io_match_task_safe(de
->req
, task
, cancel_all
)) {
3251 list_cut_position(&list
, &ctx
->defer_list
, &de
->list
);
3255 spin_unlock(&ctx
->completion_lock
);
3256 if (list_empty(&list
))
3259 while (!list_empty(&list
)) {
3260 de
= list_first_entry(&list
, struct io_defer_entry
, list
);
3261 list_del_init(&de
->list
);
3262 io_req_task_queue_fail(de
->req
, -ECANCELED
);
3268 static __cold
bool io_uring_try_cancel_iowq(struct io_ring_ctx
*ctx
)
3270 struct io_tctx_node
*node
;
3271 enum io_wq_cancel cret
;
3274 mutex_lock(&ctx
->uring_lock
);
3275 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
3276 struct io_uring_task
*tctx
= node
->task
->io_uring
;
3279 * io_wq will stay alive while we hold uring_lock, because it's
3280 * killed after ctx nodes, which requires to take the lock.
3282 if (!tctx
|| !tctx
->io_wq
)
3284 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_ctx_cb
, ctx
, true);
3285 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
3287 mutex_unlock(&ctx
->uring_lock
);
3292 static bool io_uring_try_cancel_uring_cmd(struct io_ring_ctx
*ctx
,
3293 struct task_struct
*task
, bool cancel_all
)
3295 struct hlist_node
*tmp
;
3296 struct io_kiocb
*req
;
3299 lockdep_assert_held(&ctx
->uring_lock
);
3301 hlist_for_each_entry_safe(req
, tmp
, &ctx
->cancelable_uring_cmd
,
3303 struct io_uring_cmd
*cmd
= io_kiocb_to_cmd(req
,
3304 struct io_uring_cmd
);
3305 struct file
*file
= req
->file
;
3307 if (!cancel_all
&& req
->task
!= task
)
3310 if (cmd
->flags
& IORING_URING_CMD_CANCELABLE
) {
3311 /* ->sqe isn't available if no async data */
3312 if (!req_has_async_data(req
))
3314 file
->f_op
->uring_cmd(cmd
, IO_URING_F_CANCEL
);
3318 io_submit_flush_completions(ctx
);
3323 static __cold
bool io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
3324 struct task_struct
*task
,
3327 struct io_task_cancel cancel
= { .task
= task
, .all
= cancel_all
, };
3328 struct io_uring_task
*tctx
= task
? task
->io_uring
: NULL
;
3329 enum io_wq_cancel cret
;
3332 /* set it so io_req_local_work_add() would wake us up */
3333 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) {
3334 atomic_set(&ctx
->cq_wait_nr
, 1);
3338 /* failed during ring init, it couldn't have issued any requests */
3343 ret
|= io_uring_try_cancel_iowq(ctx
);
3344 } else if (tctx
&& tctx
->io_wq
) {
3346 * Cancels requests of all rings, not only @ctx, but
3347 * it's fine as the task is in exit/exec.
3349 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_task_cb
,
3351 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
3354 /* SQPOLL thread does its own polling */
3355 if ((!(ctx
->flags
& IORING_SETUP_SQPOLL
) && cancel_all
) ||
3356 (ctx
->sq_data
&& ctx
->sq_data
->thread
== current
)) {
3357 while (!wq_list_empty(&ctx
->iopoll_list
)) {
3358 io_iopoll_try_reap_events(ctx
);
3364 if ((ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) &&
3365 io_allowed_defer_tw_run(ctx
))
3366 ret
|= io_run_local_work(ctx
) > 0;
3367 ret
|= io_cancel_defer_files(ctx
, task
, cancel_all
);
3368 mutex_lock(&ctx
->uring_lock
);
3369 ret
|= io_poll_remove_all(ctx
, task
, cancel_all
);
3370 ret
|= io_waitid_remove_all(ctx
, task
, cancel_all
);
3371 ret
|= io_futex_remove_all(ctx
, task
, cancel_all
);
3372 ret
|= io_uring_try_cancel_uring_cmd(ctx
, task
, cancel_all
);
3373 mutex_unlock(&ctx
->uring_lock
);
3374 ret
|= io_kill_timeouts(ctx
, task
, cancel_all
);
3376 ret
|= io_run_task_work() > 0;
3380 static s64
tctx_inflight(struct io_uring_task
*tctx
, bool tracked
)
3383 return atomic_read(&tctx
->inflight_tracked
);
3384 return percpu_counter_sum(&tctx
->inflight
);
3388 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3389 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3391 __cold
void io_uring_cancel_generic(bool cancel_all
, struct io_sq_data
*sqd
)
3393 struct io_uring_task
*tctx
= current
->io_uring
;
3394 struct io_ring_ctx
*ctx
;
3395 struct io_tctx_node
*node
;
3396 unsigned long index
;
3400 WARN_ON_ONCE(sqd
&& sqd
->thread
!= current
);
3402 if (!current
->io_uring
)
3405 io_wq_exit_start(tctx
->io_wq
);
3407 atomic_inc(&tctx
->in_cancel
);
3411 io_uring_drop_tctx_refs(current
);
3412 /* read completions before cancelations */
3413 inflight
= tctx_inflight(tctx
, !cancel_all
);
3418 xa_for_each(&tctx
->xa
, index
, node
) {
3419 /* sqpoll task will cancel all its requests */
3420 if (node
->ctx
->sq_data
)
3422 loop
|= io_uring_try_cancel_requests(node
->ctx
,
3423 current
, cancel_all
);
3426 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
3427 loop
|= io_uring_try_cancel_requests(ctx
,
3437 prepare_to_wait(&tctx
->wait
, &wait
, TASK_INTERRUPTIBLE
);
3439 io_uring_drop_tctx_refs(current
);
3440 xa_for_each(&tctx
->xa
, index
, node
) {
3441 if (!llist_empty(&node
->ctx
->work_llist
)) {
3442 WARN_ON_ONCE(node
->ctx
->submitter_task
&&
3443 node
->ctx
->submitter_task
!= current
);
3448 * If we've seen completions, retry without waiting. This
3449 * avoids a race where a completion comes in before we did
3450 * prepare_to_wait().
3452 if (inflight
== tctx_inflight(tctx
, !cancel_all
))
3455 finish_wait(&tctx
->wait
, &wait
);
3458 io_uring_clean_tctx(tctx
);
3461 * We shouldn't run task_works after cancel, so just leave
3462 * ->in_cancel set for normal exit.
3464 atomic_dec(&tctx
->in_cancel
);
3465 /* for exec all current's requests should be gone, kill tctx */
3466 __io_uring_free(current
);
3470 void __io_uring_cancel(bool cancel_all
)
3472 io_uring_cancel_generic(cancel_all
, NULL
);
3475 static void *io_uring_validate_mmap_request(struct file
*file
,
3476 loff_t pgoff
, size_t sz
)
3478 struct io_ring_ctx
*ctx
= file
->private_data
;
3479 loff_t offset
= pgoff
<< PAGE_SHIFT
;
3483 switch (offset
& IORING_OFF_MMAP_MASK
) {
3484 case IORING_OFF_SQ_RING
:
3485 case IORING_OFF_CQ_RING
:
3486 /* Don't allow mmap if the ring was setup without it */
3487 if (ctx
->flags
& IORING_SETUP_NO_MMAP
)
3488 return ERR_PTR(-EINVAL
);
3491 case IORING_OFF_SQES
:
3492 /* Don't allow mmap if the ring was setup without it */
3493 if (ctx
->flags
& IORING_SETUP_NO_MMAP
)
3494 return ERR_PTR(-EINVAL
);
3497 case IORING_OFF_PBUF_RING
: {
3500 bgid
= (offset
& ~IORING_OFF_MMAP_MASK
) >> IORING_OFF_PBUF_SHIFT
;
3502 ptr
= io_pbuf_get_address(ctx
, bgid
);
3505 return ERR_PTR(-EINVAL
);
3509 return ERR_PTR(-EINVAL
);
3512 page
= virt_to_head_page(ptr
);
3513 if (sz
> page_size(page
))
3514 return ERR_PTR(-EINVAL
);
3521 static __cold
int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3523 size_t sz
= vma
->vm_end
- vma
->vm_start
;
3527 ptr
= io_uring_validate_mmap_request(file
, vma
->vm_pgoff
, sz
);
3529 return PTR_ERR(ptr
);
3531 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
3532 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
3535 static unsigned long io_uring_mmu_get_unmapped_area(struct file
*filp
,
3536 unsigned long addr
, unsigned long len
,
3537 unsigned long pgoff
, unsigned long flags
)
3542 * Do not allow to map to user-provided address to avoid breaking the
3543 * aliasing rules. Userspace is not able to guess the offset address of
3544 * kernel kmalloc()ed memory area.
3549 ptr
= io_uring_validate_mmap_request(filp
, pgoff
, len
);
3554 * Some architectures have strong cache aliasing requirements.
3555 * For such architectures we need a coherent mapping which aliases
3556 * kernel memory *and* userspace memory. To achieve that:
3557 * - use a NULL file pointer to reference physical memory, and
3558 * - use the kernel virtual address of the shared io_uring context
3559 * (instead of the userspace-provided address, which has to be 0UL
3561 * - use the same pgoff which the get_unmapped_area() uses to
3562 * calculate the page colouring.
3563 * For architectures without such aliasing requirements, the
3564 * architecture will return any suitable mapping because addr is 0.
3567 flags
|= MAP_SHARED
;
3568 pgoff
= 0; /* has been translated to ptr above */
3570 addr
= (uintptr_t) ptr
;
3571 pgoff
= addr
>> PAGE_SHIFT
;
3575 return current
->mm
->get_unmapped_area(filp
, addr
, len
, pgoff
, flags
);
3578 #else /* !CONFIG_MMU */
3580 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3582 return is_nommu_shared_mapping(vma
->vm_flags
) ? 0 : -EINVAL
;
3585 static unsigned int io_uring_nommu_mmap_capabilities(struct file
*file
)
3587 return NOMMU_MAP_DIRECT
| NOMMU_MAP_READ
| NOMMU_MAP_WRITE
;
3590 static unsigned long io_uring_nommu_get_unmapped_area(struct file
*file
,
3591 unsigned long addr
, unsigned long len
,
3592 unsigned long pgoff
, unsigned long flags
)
3596 ptr
= io_uring_validate_mmap_request(file
, pgoff
, len
);
3598 return PTR_ERR(ptr
);
3600 return (unsigned long) ptr
;
3603 #endif /* !CONFIG_MMU */
3605 static int io_validate_ext_arg(unsigned flags
, const void __user
*argp
, size_t argsz
)
3607 if (flags
& IORING_ENTER_EXT_ARG
) {
3608 struct io_uring_getevents_arg arg
;
3610 if (argsz
!= sizeof(arg
))
3612 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
3618 static int io_get_ext_arg(unsigned flags
, const void __user
*argp
, size_t *argsz
,
3619 struct __kernel_timespec __user
**ts
,
3620 const sigset_t __user
**sig
)
3622 struct io_uring_getevents_arg arg
;
3625 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3626 * is just a pointer to the sigset_t.
3628 if (!(flags
& IORING_ENTER_EXT_ARG
)) {
3629 *sig
= (const sigset_t __user
*) argp
;
3635 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3636 * timespec and sigset_t pointers if good.
3638 if (*argsz
!= sizeof(arg
))
3640 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
3644 *sig
= u64_to_user_ptr(arg
.sigmask
);
3645 *argsz
= arg
.sigmask_sz
;
3646 *ts
= u64_to_user_ptr(arg
.ts
);
3650 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
3651 u32
, min_complete
, u32
, flags
, const void __user
*, argp
,
3654 struct io_ring_ctx
*ctx
;
3658 if (unlikely(flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
|
3659 IORING_ENTER_SQ_WAIT
| IORING_ENTER_EXT_ARG
|
3660 IORING_ENTER_REGISTERED_RING
)))
3664 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3665 * need only dereference our task private array to find it.
3667 if (flags
& IORING_ENTER_REGISTERED_RING
) {
3668 struct io_uring_task
*tctx
= current
->io_uring
;
3670 if (unlikely(!tctx
|| fd
>= IO_RINGFD_REG_MAX
))
3672 fd
= array_index_nospec(fd
, IO_RINGFD_REG_MAX
);
3673 file
= tctx
->registered_rings
[fd
];
3674 if (unlikely(!file
))
3678 if (unlikely(!file
))
3681 if (unlikely(!io_is_uring_fops(file
)))
3685 ctx
= file
->private_data
;
3687 if (unlikely(ctx
->flags
& IORING_SETUP_R_DISABLED
))
3691 * For SQ polling, the thread will do all submissions and completions.
3692 * Just return the requested submit count, and wake the thread if
3696 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
3697 io_cqring_overflow_flush(ctx
);
3699 if (unlikely(ctx
->sq_data
->thread
== NULL
)) {
3703 if (flags
& IORING_ENTER_SQ_WAKEUP
)
3704 wake_up(&ctx
->sq_data
->wait
);
3705 if (flags
& IORING_ENTER_SQ_WAIT
)
3706 io_sqpoll_wait_sq(ctx
);
3709 } else if (to_submit
) {
3710 ret
= io_uring_add_tctx_node(ctx
);
3714 mutex_lock(&ctx
->uring_lock
);
3715 ret
= io_submit_sqes(ctx
, to_submit
);
3716 if (ret
!= to_submit
) {
3717 mutex_unlock(&ctx
->uring_lock
);
3720 if (flags
& IORING_ENTER_GETEVENTS
) {
3721 if (ctx
->syscall_iopoll
)
3724 * Ignore errors, we'll soon call io_cqring_wait() and
3725 * it should handle ownership problems if any.
3727 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
)
3728 (void)io_run_local_work_locked(ctx
);
3730 mutex_unlock(&ctx
->uring_lock
);
3733 if (flags
& IORING_ENTER_GETEVENTS
) {
3736 if (ctx
->syscall_iopoll
) {
3738 * We disallow the app entering submit/complete with
3739 * polling, but we still need to lock the ring to
3740 * prevent racing with polled issue that got punted to
3743 mutex_lock(&ctx
->uring_lock
);
3745 ret2
= io_validate_ext_arg(flags
, argp
, argsz
);
3746 if (likely(!ret2
)) {
3747 min_complete
= min(min_complete
,
3749 ret2
= io_iopoll_check(ctx
, min_complete
);
3751 mutex_unlock(&ctx
->uring_lock
);
3753 const sigset_t __user
*sig
;
3754 struct __kernel_timespec __user
*ts
;
3756 ret2
= io_get_ext_arg(flags
, argp
, &argsz
, &ts
, &sig
);
3757 if (likely(!ret2
)) {
3758 min_complete
= min(min_complete
,
3760 ret2
= io_cqring_wait(ctx
, min_complete
, sig
,
3769 * EBADR indicates that one or more CQE were dropped.
3770 * Once the user has been informed we can clear the bit
3771 * as they are obviously ok with those drops.
3773 if (unlikely(ret2
== -EBADR
))
3774 clear_bit(IO_CHECK_CQ_DROPPED_BIT
,
3779 if (!(flags
& IORING_ENTER_REGISTERED_RING
))
3784 static const struct file_operations io_uring_fops
= {
3785 .release
= io_uring_release
,
3786 .mmap
= io_uring_mmap
,
3788 .get_unmapped_area
= io_uring_nommu_get_unmapped_area
,
3789 .mmap_capabilities
= io_uring_nommu_mmap_capabilities
,
3791 .get_unmapped_area
= io_uring_mmu_get_unmapped_area
,
3793 .poll
= io_uring_poll
,
3794 #ifdef CONFIG_PROC_FS
3795 .show_fdinfo
= io_uring_show_fdinfo
,
3799 bool io_is_uring_fops(struct file
*file
)
3801 return file
->f_op
== &io_uring_fops
;
3804 static __cold
int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
3805 struct io_uring_params
*p
)
3807 struct io_rings
*rings
;
3808 size_t size
, sq_array_offset
;
3811 /* make sure these are sane, as we already accounted them */
3812 ctx
->sq_entries
= p
->sq_entries
;
3813 ctx
->cq_entries
= p
->cq_entries
;
3815 size
= rings_size(ctx
, p
->sq_entries
, p
->cq_entries
, &sq_array_offset
);
3816 if (size
== SIZE_MAX
)
3819 if (!(ctx
->flags
& IORING_SETUP_NO_MMAP
))
3820 rings
= io_mem_alloc(size
);
3822 rings
= io_rings_map(ctx
, p
->cq_off
.user_addr
, size
);
3825 return PTR_ERR(rings
);
3828 if (!(ctx
->flags
& IORING_SETUP_NO_SQARRAY
))
3829 ctx
->sq_array
= (u32
*)((char *)rings
+ sq_array_offset
);
3830 rings
->sq_ring_mask
= p
->sq_entries
- 1;
3831 rings
->cq_ring_mask
= p
->cq_entries
- 1;
3832 rings
->sq_ring_entries
= p
->sq_entries
;
3833 rings
->cq_ring_entries
= p
->cq_entries
;
3835 if (p
->flags
& IORING_SETUP_SQE128
)
3836 size
= array_size(2 * sizeof(struct io_uring_sqe
), p
->sq_entries
);
3838 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
3839 if (size
== SIZE_MAX
) {
3844 if (!(ctx
->flags
& IORING_SETUP_NO_MMAP
))
3845 ptr
= io_mem_alloc(size
);
3847 ptr
= io_sqes_map(ctx
, p
->sq_off
.user_addr
, size
);
3851 return PTR_ERR(ptr
);
3858 static int io_uring_install_fd(struct file
*file
)
3862 fd
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
3865 fd_install(fd
, file
);
3870 * Allocate an anonymous fd, this is what constitutes the application
3871 * visible backing of an io_uring instance. The application mmaps this
3872 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3873 * we have to tie this fd to a socket for file garbage collection purposes.
3875 static struct file
*io_uring_get_file(struct io_ring_ctx
*ctx
)
3878 #if defined(CONFIG_UNIX)
3881 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
3884 return ERR_PTR(ret
);
3887 file
= anon_inode_getfile_secure("[io_uring]", &io_uring_fops
, ctx
,
3888 O_RDWR
| O_CLOEXEC
, NULL
);
3889 #if defined(CONFIG_UNIX)
3891 sock_release(ctx
->ring_sock
);
3892 ctx
->ring_sock
= NULL
;
3894 ctx
->ring_sock
->file
= file
;
3900 static __cold
int io_uring_create(unsigned entries
, struct io_uring_params
*p
,
3901 struct io_uring_params __user
*params
)
3903 struct io_ring_ctx
*ctx
;
3904 struct io_uring_task
*tctx
;
3910 if (entries
> IORING_MAX_ENTRIES
) {
3911 if (!(p
->flags
& IORING_SETUP_CLAMP
))
3913 entries
= IORING_MAX_ENTRIES
;
3916 if ((p
->flags
& IORING_SETUP_REGISTERED_FD_ONLY
)
3917 && !(p
->flags
& IORING_SETUP_NO_MMAP
))
3921 * Use twice as many entries for the CQ ring. It's possible for the
3922 * application to drive a higher depth than the size of the SQ ring,
3923 * since the sqes are only used at submission time. This allows for
3924 * some flexibility in overcommitting a bit. If the application has
3925 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3926 * of CQ ring entries manually.
3928 p
->sq_entries
= roundup_pow_of_two(entries
);
3929 if (p
->flags
& IORING_SETUP_CQSIZE
) {
3931 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3932 * to a power-of-two, if it isn't already. We do NOT impose
3933 * any cq vs sq ring sizing.
3937 if (p
->cq_entries
> IORING_MAX_CQ_ENTRIES
) {
3938 if (!(p
->flags
& IORING_SETUP_CLAMP
))
3940 p
->cq_entries
= IORING_MAX_CQ_ENTRIES
;
3942 p
->cq_entries
= roundup_pow_of_two(p
->cq_entries
);
3943 if (p
->cq_entries
< p
->sq_entries
)
3946 p
->cq_entries
= 2 * p
->sq_entries
;
3949 ctx
= io_ring_ctx_alloc(p
);
3953 if ((ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) &&
3954 !(ctx
->flags
& IORING_SETUP_IOPOLL
) &&
3955 !(ctx
->flags
& IORING_SETUP_SQPOLL
))
3956 ctx
->task_complete
= true;
3958 if (ctx
->task_complete
|| (ctx
->flags
& IORING_SETUP_IOPOLL
))
3959 ctx
->lockless_cq
= true;
3962 * lazy poll_wq activation relies on ->task_complete for synchronisation
3963 * purposes, see io_activate_pollwq()
3965 if (!ctx
->task_complete
)
3966 ctx
->poll_activated
= true;
3969 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3970 * space applications don't need to do io completion events
3971 * polling again, they can rely on io_sq_thread to do polling
3972 * work, which can reduce cpu usage and uring_lock contention.
3974 if (ctx
->flags
& IORING_SETUP_IOPOLL
&&
3975 !(ctx
->flags
& IORING_SETUP_SQPOLL
))
3976 ctx
->syscall_iopoll
= 1;
3978 ctx
->compat
= in_compat_syscall();
3979 if (!ns_capable_noaudit(&init_user_ns
, CAP_IPC_LOCK
))
3980 ctx
->user
= get_uid(current_user());
3983 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3984 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3987 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
3988 /* IPI related flags don't make sense with SQPOLL */
3989 if (ctx
->flags
& (IORING_SETUP_COOP_TASKRUN
|
3990 IORING_SETUP_TASKRUN_FLAG
|
3991 IORING_SETUP_DEFER_TASKRUN
))
3993 ctx
->notify_method
= TWA_SIGNAL_NO_IPI
;
3994 } else if (ctx
->flags
& IORING_SETUP_COOP_TASKRUN
) {
3995 ctx
->notify_method
= TWA_SIGNAL_NO_IPI
;
3997 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
&&
3998 !(ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
))
4000 ctx
->notify_method
= TWA_SIGNAL
;
4004 * For DEFER_TASKRUN we require the completion task to be the same as the
4005 * submission task. This implies that there is only one submitter, so enforce
4008 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
&&
4009 !(ctx
->flags
& IORING_SETUP_SINGLE_ISSUER
)) {
4014 * This is just grabbed for accounting purposes. When a process exits,
4015 * the mm is exited and dropped before the files, hence we need to hang
4016 * on to this mm purely for the purposes of being able to unaccount
4017 * memory (locked/pinned vm). It's not used for anything else.
4019 mmgrab(current
->mm
);
4020 ctx
->mm_account
= current
->mm
;
4022 ret
= io_allocate_scq_urings(ctx
, p
);
4026 ret
= io_sq_offload_create(ctx
, p
);
4030 ret
= io_rsrc_init(ctx
);
4034 p
->sq_off
.head
= offsetof(struct io_rings
, sq
.head
);
4035 p
->sq_off
.tail
= offsetof(struct io_rings
, sq
.tail
);
4036 p
->sq_off
.ring_mask
= offsetof(struct io_rings
, sq_ring_mask
);
4037 p
->sq_off
.ring_entries
= offsetof(struct io_rings
, sq_ring_entries
);
4038 p
->sq_off
.flags
= offsetof(struct io_rings
, sq_flags
);
4039 p
->sq_off
.dropped
= offsetof(struct io_rings
, sq_dropped
);
4040 if (!(ctx
->flags
& IORING_SETUP_NO_SQARRAY
))
4041 p
->sq_off
.array
= (char *)ctx
->sq_array
- (char *)ctx
->rings
;
4042 p
->sq_off
.resv1
= 0;
4043 if (!(ctx
->flags
& IORING_SETUP_NO_MMAP
))
4044 p
->sq_off
.user_addr
= 0;
4046 p
->cq_off
.head
= offsetof(struct io_rings
, cq
.head
);
4047 p
->cq_off
.tail
= offsetof(struct io_rings
, cq
.tail
);
4048 p
->cq_off
.ring_mask
= offsetof(struct io_rings
, cq_ring_mask
);
4049 p
->cq_off
.ring_entries
= offsetof(struct io_rings
, cq_ring_entries
);
4050 p
->cq_off
.overflow
= offsetof(struct io_rings
, cq_overflow
);
4051 p
->cq_off
.cqes
= offsetof(struct io_rings
, cqes
);
4052 p
->cq_off
.flags
= offsetof(struct io_rings
, cq_flags
);
4053 p
->cq_off
.resv1
= 0;
4054 if (!(ctx
->flags
& IORING_SETUP_NO_MMAP
))
4055 p
->cq_off
.user_addr
= 0;
4057 p
->features
= IORING_FEAT_SINGLE_MMAP
| IORING_FEAT_NODROP
|
4058 IORING_FEAT_SUBMIT_STABLE
| IORING_FEAT_RW_CUR_POS
|
4059 IORING_FEAT_CUR_PERSONALITY
| IORING_FEAT_FAST_POLL
|
4060 IORING_FEAT_POLL_32BITS
| IORING_FEAT_SQPOLL_NONFIXED
|
4061 IORING_FEAT_EXT_ARG
| IORING_FEAT_NATIVE_WORKERS
|
4062 IORING_FEAT_RSRC_TAGS
| IORING_FEAT_CQE_SKIP
|
4063 IORING_FEAT_LINKED_FILE
| IORING_FEAT_REG_REG_RING
;
4065 if (copy_to_user(params
, p
, sizeof(*p
))) {
4070 if (ctx
->flags
& IORING_SETUP_SINGLE_ISSUER
4071 && !(ctx
->flags
& IORING_SETUP_R_DISABLED
))
4072 WRITE_ONCE(ctx
->submitter_task
, get_task_struct(current
));
4074 file
= io_uring_get_file(ctx
);
4076 ret
= PTR_ERR(file
);
4080 ret
= __io_uring_add_tctx_node(ctx
);
4083 tctx
= current
->io_uring
;
4086 * Install ring fd as the very last thing, so we don't risk someone
4087 * having closed it before we finish setup
4089 if (p
->flags
& IORING_SETUP_REGISTERED_FD_ONLY
)
4090 ret
= io_ring_add_registered_file(tctx
, file
, 0, IO_RINGFD_REG_MAX
);
4092 ret
= io_uring_install_fd(file
);
4096 trace_io_uring_create(ret
, ctx
, p
->sq_entries
, p
->cq_entries
, p
->flags
);
4099 io_ring_ctx_wait_and_kill(ctx
);
4107 * Sets up an aio uring context, and returns the fd. Applications asks for a
4108 * ring size, we return the actual sq/cq ring sizes (among other things) in the
4109 * params structure passed in.
4111 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
4113 struct io_uring_params p
;
4116 if (copy_from_user(&p
, params
, sizeof(p
)))
4118 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
4123 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
4124 IORING_SETUP_SQ_AFF
| IORING_SETUP_CQSIZE
|
4125 IORING_SETUP_CLAMP
| IORING_SETUP_ATTACH_WQ
|
4126 IORING_SETUP_R_DISABLED
| IORING_SETUP_SUBMIT_ALL
|
4127 IORING_SETUP_COOP_TASKRUN
| IORING_SETUP_TASKRUN_FLAG
|
4128 IORING_SETUP_SQE128
| IORING_SETUP_CQE32
|
4129 IORING_SETUP_SINGLE_ISSUER
| IORING_SETUP_DEFER_TASKRUN
|
4130 IORING_SETUP_NO_MMAP
| IORING_SETUP_REGISTERED_FD_ONLY
|
4131 IORING_SETUP_NO_SQARRAY
))
4134 return io_uring_create(entries
, &p
, params
);
4137 static inline bool io_uring_allowed(void)
4139 int disabled
= READ_ONCE(sysctl_io_uring_disabled
);
4140 kgid_t io_uring_group
;
4145 if (disabled
== 0 || capable(CAP_SYS_ADMIN
))
4148 io_uring_group
= make_kgid(&init_user_ns
, sysctl_io_uring_group
);
4149 if (!gid_valid(io_uring_group
))
4152 return in_group_p(io_uring_group
);
4155 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
4156 struct io_uring_params __user
*, params
)
4158 if (!io_uring_allowed())
4161 return io_uring_setup(entries
, params
);
4164 static __cold
int io_probe(struct io_ring_ctx
*ctx
, void __user
*arg
,
4167 struct io_uring_probe
*p
;
4171 size
= struct_size(p
, ops
, nr_args
);
4172 if (size
== SIZE_MAX
)
4174 p
= kzalloc(size
, GFP_KERNEL
);
4179 if (copy_from_user(p
, arg
, size
))
4182 if (memchr_inv(p
, 0, size
))
4185 p
->last_op
= IORING_OP_LAST
- 1;
4186 if (nr_args
> IORING_OP_LAST
)
4187 nr_args
= IORING_OP_LAST
;
4189 for (i
= 0; i
< nr_args
; i
++) {
4191 if (!io_issue_defs
[i
].not_supported
)
4192 p
->ops
[i
].flags
= IO_URING_OP_SUPPORTED
;
4197 if (copy_to_user(arg
, p
, size
))
4204 static int io_register_personality(struct io_ring_ctx
*ctx
)
4206 const struct cred
*creds
;
4210 creds
= get_current_cred();
4212 ret
= xa_alloc_cyclic(&ctx
->personalities
, &id
, (void *)creds
,
4213 XA_LIMIT(0, USHRT_MAX
), &ctx
->pers_next
, GFP_KERNEL
);
4221 static __cold
int io_register_restrictions(struct io_ring_ctx
*ctx
,
4222 void __user
*arg
, unsigned int nr_args
)
4224 struct io_uring_restriction
*res
;
4228 /* Restrictions allowed only if rings started disabled */
4229 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
4232 /* We allow only a single restrictions registration */
4233 if (ctx
->restrictions
.registered
)
4236 if (!arg
|| nr_args
> IORING_MAX_RESTRICTIONS
)
4239 size
= array_size(nr_args
, sizeof(*res
));
4240 if (size
== SIZE_MAX
)
4243 res
= memdup_user(arg
, size
);
4245 return PTR_ERR(res
);
4249 for (i
= 0; i
< nr_args
; i
++) {
4250 switch (res
[i
].opcode
) {
4251 case IORING_RESTRICTION_REGISTER_OP
:
4252 if (res
[i
].register_op
>= IORING_REGISTER_LAST
) {
4257 __set_bit(res
[i
].register_op
,
4258 ctx
->restrictions
.register_op
);
4260 case IORING_RESTRICTION_SQE_OP
:
4261 if (res
[i
].sqe_op
>= IORING_OP_LAST
) {
4266 __set_bit(res
[i
].sqe_op
, ctx
->restrictions
.sqe_op
);
4268 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED
:
4269 ctx
->restrictions
.sqe_flags_allowed
= res
[i
].sqe_flags
;
4271 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED
:
4272 ctx
->restrictions
.sqe_flags_required
= res
[i
].sqe_flags
;
4281 /* Reset all restrictions if an error happened */
4283 memset(&ctx
->restrictions
, 0, sizeof(ctx
->restrictions
));
4285 ctx
->restrictions
.registered
= true;
4291 static int io_register_enable_rings(struct io_ring_ctx
*ctx
)
4293 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
4296 if (ctx
->flags
& IORING_SETUP_SINGLE_ISSUER
&& !ctx
->submitter_task
) {
4297 WRITE_ONCE(ctx
->submitter_task
, get_task_struct(current
));
4299 * Lazy activation attempts would fail if it was polled before
4300 * submitter_task is set.
4302 if (wq_has_sleeper(&ctx
->poll_wq
))
4303 io_activate_pollwq(ctx
);
4306 if (ctx
->restrictions
.registered
)
4307 ctx
->restricted
= 1;
4309 ctx
->flags
&= ~IORING_SETUP_R_DISABLED
;
4310 if (ctx
->sq_data
&& wq_has_sleeper(&ctx
->sq_data
->wait
))
4311 wake_up(&ctx
->sq_data
->wait
);
4315 static __cold
int __io_register_iowq_aff(struct io_ring_ctx
*ctx
,
4316 cpumask_var_t new_mask
)
4320 if (!(ctx
->flags
& IORING_SETUP_SQPOLL
)) {
4321 ret
= io_wq_cpu_affinity(current
->io_uring
, new_mask
);
4323 mutex_unlock(&ctx
->uring_lock
);
4324 ret
= io_sqpoll_wq_cpu_affinity(ctx
, new_mask
);
4325 mutex_lock(&ctx
->uring_lock
);
4331 static __cold
int io_register_iowq_aff(struct io_ring_ctx
*ctx
,
4332 void __user
*arg
, unsigned len
)
4334 cpumask_var_t new_mask
;
4337 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4340 cpumask_clear(new_mask
);
4341 if (len
> cpumask_size())
4342 len
= cpumask_size();
4344 if (in_compat_syscall()) {
4345 ret
= compat_get_bitmap(cpumask_bits(new_mask
),
4346 (const compat_ulong_t __user
*)arg
,
4347 len
* 8 /* CHAR_BIT */);
4349 ret
= copy_from_user(new_mask
, arg
, len
);
4353 free_cpumask_var(new_mask
);
4357 ret
= __io_register_iowq_aff(ctx
, new_mask
);
4358 free_cpumask_var(new_mask
);
4362 static __cold
int io_unregister_iowq_aff(struct io_ring_ctx
*ctx
)
4364 return __io_register_iowq_aff(ctx
, NULL
);
4367 static __cold
int io_register_iowq_max_workers(struct io_ring_ctx
*ctx
,
4369 __must_hold(&ctx
->uring_lock
)
4371 struct io_tctx_node
*node
;
4372 struct io_uring_task
*tctx
= NULL
;
4373 struct io_sq_data
*sqd
= NULL
;
4377 if (copy_from_user(new_count
, arg
, sizeof(new_count
)))
4379 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
4380 if (new_count
[i
] > INT_MAX
)
4383 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
4387 * Observe the correct sqd->lock -> ctx->uring_lock
4388 * ordering. Fine to drop uring_lock here, we hold
4391 refcount_inc(&sqd
->refs
);
4392 mutex_unlock(&ctx
->uring_lock
);
4393 mutex_lock(&sqd
->lock
);
4394 mutex_lock(&ctx
->uring_lock
);
4396 tctx
= sqd
->thread
->io_uring
;
4399 tctx
= current
->io_uring
;
4402 BUILD_BUG_ON(sizeof(new_count
) != sizeof(ctx
->iowq_limits
));
4404 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
4406 ctx
->iowq_limits
[i
] = new_count
[i
];
4407 ctx
->iowq_limits_set
= true;
4409 if (tctx
&& tctx
->io_wq
) {
4410 ret
= io_wq_max_workers(tctx
->io_wq
, new_count
);
4414 memset(new_count
, 0, sizeof(new_count
));
4418 mutex_unlock(&sqd
->lock
);
4419 io_put_sq_data(sqd
);
4422 if (copy_to_user(arg
, new_count
, sizeof(new_count
)))
4425 /* that's it for SQPOLL, only the SQPOLL task creates requests */
4429 /* now propagate the restriction to all registered users */
4430 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
4431 struct io_uring_task
*tctx
= node
->task
->io_uring
;
4433 if (WARN_ON_ONCE(!tctx
->io_wq
))
4436 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
4437 new_count
[i
] = ctx
->iowq_limits
[i
];
4438 /* ignore errors, it always returns zero anyway */
4439 (void)io_wq_max_workers(tctx
->io_wq
, new_count
);
4444 mutex_unlock(&sqd
->lock
);
4445 io_put_sq_data(sqd
);
4450 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
4451 void __user
*arg
, unsigned nr_args
)
4452 __releases(ctx
->uring_lock
)
4453 __acquires(ctx
->uring_lock
)
4458 * We don't quiesce the refs for register anymore and so it can't be
4459 * dying as we're holding a file ref here.
4461 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx
->refs
)))
4464 if (ctx
->submitter_task
&& ctx
->submitter_task
!= current
)
4467 if (ctx
->restricted
) {
4468 opcode
= array_index_nospec(opcode
, IORING_REGISTER_LAST
);
4469 if (!test_bit(opcode
, ctx
->restrictions
.register_op
))
4474 case IORING_REGISTER_BUFFERS
:
4478 ret
= io_sqe_buffers_register(ctx
, arg
, nr_args
, NULL
);
4480 case IORING_UNREGISTER_BUFFERS
:
4484 ret
= io_sqe_buffers_unregister(ctx
);
4486 case IORING_REGISTER_FILES
:
4490 ret
= io_sqe_files_register(ctx
, arg
, nr_args
, NULL
);
4492 case IORING_UNREGISTER_FILES
:
4496 ret
= io_sqe_files_unregister(ctx
);
4498 case IORING_REGISTER_FILES_UPDATE
:
4499 ret
= io_register_files_update(ctx
, arg
, nr_args
);
4501 case IORING_REGISTER_EVENTFD
:
4505 ret
= io_eventfd_register(ctx
, arg
, 0);
4507 case IORING_REGISTER_EVENTFD_ASYNC
:
4511 ret
= io_eventfd_register(ctx
, arg
, 1);
4513 case IORING_UNREGISTER_EVENTFD
:
4517 ret
= io_eventfd_unregister(ctx
);
4519 case IORING_REGISTER_PROBE
:
4521 if (!arg
|| nr_args
> 256)
4523 ret
= io_probe(ctx
, arg
, nr_args
);
4525 case IORING_REGISTER_PERSONALITY
:
4529 ret
= io_register_personality(ctx
);
4531 case IORING_UNREGISTER_PERSONALITY
:
4535 ret
= io_unregister_personality(ctx
, nr_args
);
4537 case IORING_REGISTER_ENABLE_RINGS
:
4541 ret
= io_register_enable_rings(ctx
);
4543 case IORING_REGISTER_RESTRICTIONS
:
4544 ret
= io_register_restrictions(ctx
, arg
, nr_args
);
4546 case IORING_REGISTER_FILES2
:
4547 ret
= io_register_rsrc(ctx
, arg
, nr_args
, IORING_RSRC_FILE
);
4549 case IORING_REGISTER_FILES_UPDATE2
:
4550 ret
= io_register_rsrc_update(ctx
, arg
, nr_args
,
4553 case IORING_REGISTER_BUFFERS2
:
4554 ret
= io_register_rsrc(ctx
, arg
, nr_args
, IORING_RSRC_BUFFER
);
4556 case IORING_REGISTER_BUFFERS_UPDATE
:
4557 ret
= io_register_rsrc_update(ctx
, arg
, nr_args
,
4558 IORING_RSRC_BUFFER
);
4560 case IORING_REGISTER_IOWQ_AFF
:
4562 if (!arg
|| !nr_args
)
4564 ret
= io_register_iowq_aff(ctx
, arg
, nr_args
);
4566 case IORING_UNREGISTER_IOWQ_AFF
:
4570 ret
= io_unregister_iowq_aff(ctx
);
4572 case IORING_REGISTER_IOWQ_MAX_WORKERS
:
4574 if (!arg
|| nr_args
!= 2)
4576 ret
= io_register_iowq_max_workers(ctx
, arg
);
4578 case IORING_REGISTER_RING_FDS
:
4579 ret
= io_ringfd_register(ctx
, arg
, nr_args
);
4581 case IORING_UNREGISTER_RING_FDS
:
4582 ret
= io_ringfd_unregister(ctx
, arg
, nr_args
);
4584 case IORING_REGISTER_PBUF_RING
:
4586 if (!arg
|| nr_args
!= 1)
4588 ret
= io_register_pbuf_ring(ctx
, arg
);
4590 case IORING_UNREGISTER_PBUF_RING
:
4592 if (!arg
|| nr_args
!= 1)
4594 ret
= io_unregister_pbuf_ring(ctx
, arg
);
4596 case IORING_REGISTER_SYNC_CANCEL
:
4598 if (!arg
|| nr_args
!= 1)
4600 ret
= io_sync_cancel(ctx
, arg
);
4602 case IORING_REGISTER_FILE_ALLOC_RANGE
:
4604 if (!arg
|| nr_args
)
4606 ret
= io_register_file_alloc_range(ctx
, arg
);
4616 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
4617 void __user
*, arg
, unsigned int, nr_args
)
4619 struct io_ring_ctx
*ctx
;
4622 bool use_registered_ring
;
4624 use_registered_ring
= !!(opcode
& IORING_REGISTER_USE_REGISTERED_RING
);
4625 opcode
&= ~IORING_REGISTER_USE_REGISTERED_RING
;
4627 if (opcode
>= IORING_REGISTER_LAST
)
4630 if (use_registered_ring
) {
4632 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
4633 * need only dereference our task private array to find it.
4635 struct io_uring_task
*tctx
= current
->io_uring
;
4637 if (unlikely(!tctx
|| fd
>= IO_RINGFD_REG_MAX
))
4639 fd
= array_index_nospec(fd
, IO_RINGFD_REG_MAX
);
4640 file
= tctx
->registered_rings
[fd
];
4641 if (unlikely(!file
))
4645 if (unlikely(!file
))
4648 if (!io_is_uring_fops(file
))
4652 ctx
= file
->private_data
;
4654 mutex_lock(&ctx
->uring_lock
);
4655 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
4656 mutex_unlock(&ctx
->uring_lock
);
4657 trace_io_uring_register(ctx
, opcode
, ctx
->nr_user_files
, ctx
->nr_user_bufs
, ret
);
4659 if (!use_registered_ring
)
4664 static int __init
io_uring_init(void)
4666 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4667 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4668 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4671 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4672 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4673 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4674 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4675 BUILD_BUG_ON(sizeof(struct io_uring_sqe
) != 64);
4676 BUILD_BUG_SQE_ELEM(0, __u8
, opcode
);
4677 BUILD_BUG_SQE_ELEM(1, __u8
, flags
);
4678 BUILD_BUG_SQE_ELEM(2, __u16
, ioprio
);
4679 BUILD_BUG_SQE_ELEM(4, __s32
, fd
);
4680 BUILD_BUG_SQE_ELEM(8, __u64
, off
);
4681 BUILD_BUG_SQE_ELEM(8, __u64
, addr2
);
4682 BUILD_BUG_SQE_ELEM(8, __u32
, cmd_op
);
4683 BUILD_BUG_SQE_ELEM(12, __u32
, __pad1
);
4684 BUILD_BUG_SQE_ELEM(16, __u64
, addr
);
4685 BUILD_BUG_SQE_ELEM(16, __u64
, splice_off_in
);
4686 BUILD_BUG_SQE_ELEM(24, __u32
, len
);
4687 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t
, rw_flags
);
4688 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags
);
4689 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32
, rw_flags
);
4690 BUILD_BUG_SQE_ELEM(28, __u32
, fsync_flags
);
4691 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16
, poll_events
);
4692 BUILD_BUG_SQE_ELEM(28, __u32
, poll32_events
);
4693 BUILD_BUG_SQE_ELEM(28, __u32
, sync_range_flags
);
4694 BUILD_BUG_SQE_ELEM(28, __u32
, msg_flags
);
4695 BUILD_BUG_SQE_ELEM(28, __u32
, timeout_flags
);
4696 BUILD_BUG_SQE_ELEM(28, __u32
, accept_flags
);
4697 BUILD_BUG_SQE_ELEM(28, __u32
, cancel_flags
);
4698 BUILD_BUG_SQE_ELEM(28, __u32
, open_flags
);
4699 BUILD_BUG_SQE_ELEM(28, __u32
, statx_flags
);
4700 BUILD_BUG_SQE_ELEM(28, __u32
, fadvise_advice
);
4701 BUILD_BUG_SQE_ELEM(28, __u32
, splice_flags
);
4702 BUILD_BUG_SQE_ELEM(28, __u32
, rename_flags
);
4703 BUILD_BUG_SQE_ELEM(28, __u32
, unlink_flags
);
4704 BUILD_BUG_SQE_ELEM(28, __u32
, hardlink_flags
);
4705 BUILD_BUG_SQE_ELEM(28, __u32
, xattr_flags
);
4706 BUILD_BUG_SQE_ELEM(28, __u32
, msg_ring_flags
);
4707 BUILD_BUG_SQE_ELEM(32, __u64
, user_data
);
4708 BUILD_BUG_SQE_ELEM(40, __u16
, buf_index
);
4709 BUILD_BUG_SQE_ELEM(40, __u16
, buf_group
);
4710 BUILD_BUG_SQE_ELEM(42, __u16
, personality
);
4711 BUILD_BUG_SQE_ELEM(44, __s32
, splice_fd_in
);
4712 BUILD_BUG_SQE_ELEM(44, __u32
, file_index
);
4713 BUILD_BUG_SQE_ELEM(44, __u16
, addr_len
);
4714 BUILD_BUG_SQE_ELEM(46, __u16
, __pad3
[0]);
4715 BUILD_BUG_SQE_ELEM(48, __u64
, addr3
);
4716 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd
);
4717 BUILD_BUG_SQE_ELEM(56, __u64
, __pad2
);
4719 BUILD_BUG_ON(sizeof(struct io_uring_files_update
) !=
4720 sizeof(struct io_uring_rsrc_update
));
4721 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update
) >
4722 sizeof(struct io_uring_rsrc_update2
));
4724 /* ->buf_index is u16 */
4725 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring
, bufs
) != 0);
4726 BUILD_BUG_ON(offsetof(struct io_uring_buf
, resv
) !=
4727 offsetof(struct io_uring_buf_ring
, tail
));
4729 /* should fit into one byte */
4730 BUILD_BUG_ON(SQE_VALID_FLAGS
>= (1 << 8));
4731 BUILD_BUG_ON(SQE_COMMON_FLAGS
>= (1 << 8));
4732 BUILD_BUG_ON((SQE_VALID_FLAGS
| SQE_COMMON_FLAGS
) != SQE_VALID_FLAGS
);
4734 BUILD_BUG_ON(__REQ_F_LAST_BIT
> 8 * sizeof(int));
4736 BUILD_BUG_ON(sizeof(atomic_t
) != sizeof(u32
));
4738 /* top 8bits are for internal use */
4739 BUILD_BUG_ON((IORING_URING_CMD_MASK
& 0xff000000) != 0);
4741 io_uring_optable_init();
4744 * Allow user copy in the per-command field, which starts after the
4745 * file in io_kiocb and until the opcode field. The openat2 handling
4746 * requires copying in user memory into the io_kiocb object in that
4747 * range, and HARDENED_USERCOPY will complain if we haven't
4748 * correctly annotated this range.
4750 req_cachep
= kmem_cache_create_usercopy("io_kiocb",
4751 sizeof(struct io_kiocb
), 0,
4752 SLAB_HWCACHE_ALIGN
| SLAB_PANIC
|
4753 SLAB_ACCOUNT
| SLAB_TYPESAFE_BY_RCU
,
4754 offsetof(struct io_kiocb
, cmd
.data
),
4755 sizeof_field(struct io_kiocb
, cmd
.data
), NULL
);
4756 io_buf_cachep
= kmem_cache_create("io_buffer", sizeof(struct io_buffer
), 0,
4757 SLAB_HWCACHE_ALIGN
| SLAB_PANIC
| SLAB_ACCOUNT
,
4760 #ifdef CONFIG_SYSCTL
4761 register_sysctl_init("kernel", kernel_io_uring_disabled_table
);
4766 __initcall(io_uring_init
);