1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <net/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
49 #include <linux/bits.h>
51 #include <linux/sched/signal.h>
53 #include <linux/file.h>
54 #include <linux/fdtable.h>
56 #include <linux/mman.h>
57 #include <linux/percpu.h>
58 #include <linux/slab.h>
59 #include <linux/bvec.h>
60 #include <linux/net.h>
62 #include <net/af_unix.h>
64 #include <linux/anon_inodes.h>
65 #include <linux/sched/mm.h>
66 #include <linux/uaccess.h>
67 #include <linux/nospec.h>
68 #include <linux/highmem.h>
69 #include <linux/fsnotify.h>
70 #include <linux/fadvise.h>
71 #include <linux/task_work.h>
72 #include <linux/io_uring.h>
73 #include <linux/audit.h>
74 #include <linux/security.h>
75 #include <asm/shmparam.h>
77 #define CREATE_TRACE_POINTS
78 #include <trace/events/io_uring.h>
80 #include <uapi/linux/io_uring.h>
99 #include "alloc_cache.h"
101 #define IORING_MAX_ENTRIES 32768
102 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
104 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
105 IORING_REGISTER_LAST + IORING_OP_LAST)
107 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
108 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
110 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
111 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
113 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
114 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
117 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
120 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
122 #define IO_COMPL_BATCH 32
123 #define IO_REQ_ALLOC_BATCH 8
126 IO_CHECK_CQ_OVERFLOW_BIT
,
127 IO_CHECK_CQ_DROPPED_BIT
,
131 IO_EVENTFD_OP_SIGNAL_BIT
,
132 IO_EVENTFD_OP_FREE_BIT
,
135 struct io_defer_entry
{
136 struct list_head list
;
137 struct io_kiocb
*req
;
141 /* requests with any of those set should undergo io_disarm_next() */
142 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
143 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
145 static bool io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
146 struct task_struct
*task
,
149 static void io_queue_sqe(struct io_kiocb
*req
);
151 struct kmem_cache
*req_cachep
;
153 static int __read_mostly sysctl_io_uring_disabled
;
154 static int __read_mostly sysctl_io_uring_group
= -1;
157 static struct ctl_table kernel_io_uring_disabled_table
[] = {
159 .procname
= "io_uring_disabled",
160 .data
= &sysctl_io_uring_disabled
,
161 .maxlen
= sizeof(sysctl_io_uring_disabled
),
163 .proc_handler
= proc_dointvec_minmax
,
164 .extra1
= SYSCTL_ZERO
,
165 .extra2
= SYSCTL_TWO
,
168 .procname
= "io_uring_group",
169 .data
= &sysctl_io_uring_group
,
170 .maxlen
= sizeof(gid_t
),
172 .proc_handler
= proc_dointvec
,
178 struct sock
*io_uring_get_socket(struct file
*file
)
180 #if defined(CONFIG_UNIX)
181 if (io_is_uring_fops(file
)) {
182 struct io_ring_ctx
*ctx
= file
->private_data
;
184 return ctx
->ring_sock
->sk
;
189 EXPORT_SYMBOL(io_uring_get_socket
);
191 static inline void io_submit_flush_completions(struct io_ring_ctx
*ctx
)
193 if (!wq_list_empty(&ctx
->submit_state
.compl_reqs
) ||
194 ctx
->submit_state
.cqes_count
)
195 __io_submit_flush_completions(ctx
);
198 static inline unsigned int __io_cqring_events(struct io_ring_ctx
*ctx
)
200 return ctx
->cached_cq_tail
- READ_ONCE(ctx
->rings
->cq
.head
);
203 static inline unsigned int __io_cqring_events_user(struct io_ring_ctx
*ctx
)
205 return READ_ONCE(ctx
->rings
->cq
.tail
) - READ_ONCE(ctx
->rings
->cq
.head
);
208 static bool io_match_linked(struct io_kiocb
*head
)
210 struct io_kiocb
*req
;
212 io_for_each_link(req
, head
) {
213 if (req
->flags
& REQ_F_INFLIGHT
)
220 * As io_match_task() but protected against racing with linked timeouts.
221 * User must not hold timeout_lock.
223 bool io_match_task_safe(struct io_kiocb
*head
, struct task_struct
*task
,
228 if (task
&& head
->task
!= task
)
233 if (head
->flags
& REQ_F_LINK_TIMEOUT
) {
234 struct io_ring_ctx
*ctx
= head
->ctx
;
236 /* protect against races with linked timeouts */
237 spin_lock_irq(&ctx
->timeout_lock
);
238 matched
= io_match_linked(head
);
239 spin_unlock_irq(&ctx
->timeout_lock
);
241 matched
= io_match_linked(head
);
246 static inline void req_fail_link_node(struct io_kiocb
*req
, int res
)
249 io_req_set_res(req
, res
, 0);
252 static inline void io_req_add_to_cache(struct io_kiocb
*req
, struct io_ring_ctx
*ctx
)
254 wq_stack_add_head(&req
->comp_list
, &ctx
->submit_state
.free_list
);
257 static __cold
void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
259 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
261 complete(&ctx
->ref_comp
);
264 static __cold
void io_fallback_req_func(struct work_struct
*work
)
266 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
,
268 struct llist_node
*node
= llist_del_all(&ctx
->fallback_llist
);
269 struct io_kiocb
*req
, *tmp
;
270 struct io_tw_state ts
= { .locked
= true, };
272 mutex_lock(&ctx
->uring_lock
);
273 llist_for_each_entry_safe(req
, tmp
, node
, io_task_work
.node
)
274 req
->io_task_work
.func(req
, &ts
);
275 if (WARN_ON_ONCE(!ts
.locked
))
277 io_submit_flush_completions(ctx
);
278 mutex_unlock(&ctx
->uring_lock
);
281 static int io_alloc_hash_table(struct io_hash_table
*table
, unsigned bits
)
283 unsigned hash_buckets
= 1U << bits
;
284 size_t hash_size
= hash_buckets
* sizeof(table
->hbs
[0]);
286 table
->hbs
= kmalloc(hash_size
, GFP_KERNEL
);
290 table
->hash_bits
= bits
;
291 init_hash_table(table
, hash_buckets
);
295 static __cold
struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
297 struct io_ring_ctx
*ctx
;
300 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
304 xa_init(&ctx
->io_bl_xa
);
307 * Use 5 bits less than the max cq entries, that should give us around
308 * 32 entries per hash list if totally full and uniformly spread, but
309 * don't keep too many buckets to not overconsume memory.
311 hash_bits
= ilog2(p
->cq_entries
) - 5;
312 hash_bits
= clamp(hash_bits
, 1, 8);
313 if (io_alloc_hash_table(&ctx
->cancel_table
, hash_bits
))
315 if (io_alloc_hash_table(&ctx
->cancel_table_locked
, hash_bits
))
317 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
,
321 ctx
->flags
= p
->flags
;
322 init_waitqueue_head(&ctx
->sqo_sq_wait
);
323 INIT_LIST_HEAD(&ctx
->sqd_list
);
324 INIT_LIST_HEAD(&ctx
->cq_overflow_list
);
325 INIT_LIST_HEAD(&ctx
->io_buffers_cache
);
326 io_alloc_cache_init(&ctx
->rsrc_node_cache
, IO_NODE_ALLOC_CACHE_MAX
,
327 sizeof(struct io_rsrc_node
));
328 io_alloc_cache_init(&ctx
->apoll_cache
, IO_ALLOC_CACHE_MAX
,
329 sizeof(struct async_poll
));
330 io_alloc_cache_init(&ctx
->netmsg_cache
, IO_ALLOC_CACHE_MAX
,
331 sizeof(struct io_async_msghdr
));
332 init_completion(&ctx
->ref_comp
);
333 xa_init_flags(&ctx
->personalities
, XA_FLAGS_ALLOC1
);
334 mutex_init(&ctx
->uring_lock
);
335 init_waitqueue_head(&ctx
->cq_wait
);
336 init_waitqueue_head(&ctx
->poll_wq
);
337 init_waitqueue_head(&ctx
->rsrc_quiesce_wq
);
338 spin_lock_init(&ctx
->completion_lock
);
339 spin_lock_init(&ctx
->timeout_lock
);
340 INIT_WQ_LIST(&ctx
->iopoll_list
);
341 INIT_LIST_HEAD(&ctx
->io_buffers_pages
);
342 INIT_LIST_HEAD(&ctx
->io_buffers_comp
);
343 INIT_LIST_HEAD(&ctx
->defer_list
);
344 INIT_LIST_HEAD(&ctx
->timeout_list
);
345 INIT_LIST_HEAD(&ctx
->ltimeout_list
);
346 INIT_LIST_HEAD(&ctx
->rsrc_ref_list
);
347 init_llist_head(&ctx
->work_llist
);
348 INIT_LIST_HEAD(&ctx
->tctx_list
);
349 ctx
->submit_state
.free_list
.next
= NULL
;
350 INIT_WQ_LIST(&ctx
->locked_free_list
);
351 INIT_DELAYED_WORK(&ctx
->fallback_work
, io_fallback_req_func
);
352 INIT_WQ_LIST(&ctx
->submit_state
.compl_reqs
);
355 kfree(ctx
->cancel_table
.hbs
);
356 kfree(ctx
->cancel_table_locked
.hbs
);
358 xa_destroy(&ctx
->io_bl_xa
);
363 static void io_account_cq_overflow(struct io_ring_ctx
*ctx
)
365 struct io_rings
*r
= ctx
->rings
;
367 WRITE_ONCE(r
->cq_overflow
, READ_ONCE(r
->cq_overflow
) + 1);
371 static bool req_need_defer(struct io_kiocb
*req
, u32 seq
)
373 if (unlikely(req
->flags
& REQ_F_IO_DRAIN
)) {
374 struct io_ring_ctx
*ctx
= req
->ctx
;
376 return seq
+ READ_ONCE(ctx
->cq_extra
) != ctx
->cached_cq_tail
;
382 static void io_clean_op(struct io_kiocb
*req
)
384 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
385 spin_lock(&req
->ctx
->completion_lock
);
386 io_put_kbuf_comp(req
);
387 spin_unlock(&req
->ctx
->completion_lock
);
390 if (req
->flags
& REQ_F_NEED_CLEANUP
) {
391 const struct io_cold_def
*def
= &io_cold_defs
[req
->opcode
];
396 if ((req
->flags
& REQ_F_POLLED
) && req
->apoll
) {
397 kfree(req
->apoll
->double_poll
);
401 if (req
->flags
& REQ_F_INFLIGHT
) {
402 struct io_uring_task
*tctx
= req
->task
->io_uring
;
404 atomic_dec(&tctx
->inflight_tracked
);
406 if (req
->flags
& REQ_F_CREDS
)
407 put_cred(req
->creds
);
408 if (req
->flags
& REQ_F_ASYNC_DATA
) {
409 kfree(req
->async_data
);
410 req
->async_data
= NULL
;
412 req
->flags
&= ~IO_REQ_CLEAN_FLAGS
;
415 static inline void io_req_track_inflight(struct io_kiocb
*req
)
417 if (!(req
->flags
& REQ_F_INFLIGHT
)) {
418 req
->flags
|= REQ_F_INFLIGHT
;
419 atomic_inc(&req
->task
->io_uring
->inflight_tracked
);
423 static struct io_kiocb
*__io_prep_linked_timeout(struct io_kiocb
*req
)
425 if (WARN_ON_ONCE(!req
->link
))
428 req
->flags
&= ~REQ_F_ARM_LTIMEOUT
;
429 req
->flags
|= REQ_F_LINK_TIMEOUT
;
431 /* linked timeouts should have two refs once prep'ed */
432 io_req_set_refcount(req
);
433 __io_req_set_refcount(req
->link
, 2);
437 static inline struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
)
439 if (likely(!(req
->flags
& REQ_F_ARM_LTIMEOUT
)))
441 return __io_prep_linked_timeout(req
);
444 static noinline
void __io_arm_ltimeout(struct io_kiocb
*req
)
446 io_queue_linked_timeout(__io_prep_linked_timeout(req
));
449 static inline void io_arm_ltimeout(struct io_kiocb
*req
)
451 if (unlikely(req
->flags
& REQ_F_ARM_LTIMEOUT
))
452 __io_arm_ltimeout(req
);
455 static void io_prep_async_work(struct io_kiocb
*req
)
457 const struct io_issue_def
*def
= &io_issue_defs
[req
->opcode
];
458 struct io_ring_ctx
*ctx
= req
->ctx
;
460 if (!(req
->flags
& REQ_F_CREDS
)) {
461 req
->flags
|= REQ_F_CREDS
;
462 req
->creds
= get_current_cred();
465 req
->work
.list
.next
= NULL
;
467 req
->work
.cancel_seq
= atomic_read(&ctx
->cancel_seq
);
468 if (req
->flags
& REQ_F_FORCE_ASYNC
)
469 req
->work
.flags
|= IO_WQ_WORK_CONCURRENT
;
471 if (req
->file
&& !(req
->flags
& REQ_F_FIXED_FILE
))
472 req
->flags
|= io_file_get_flags(req
->file
);
474 if (req
->file
&& (req
->flags
& REQ_F_ISREG
)) {
475 bool should_hash
= def
->hash_reg_file
;
477 /* don't serialize this request if the fs doesn't need it */
478 if (should_hash
&& (req
->file
->f_flags
& O_DIRECT
) &&
479 (req
->file
->f_mode
& FMODE_DIO_PARALLEL_WRITE
))
481 if (should_hash
|| (ctx
->flags
& IORING_SETUP_IOPOLL
))
482 io_wq_hash_work(&req
->work
, file_inode(req
->file
));
483 } else if (!req
->file
|| !S_ISBLK(file_inode(req
->file
)->i_mode
)) {
484 if (def
->unbound_nonreg_file
)
485 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
489 static void io_prep_async_link(struct io_kiocb
*req
)
491 struct io_kiocb
*cur
;
493 if (req
->flags
& REQ_F_LINK_TIMEOUT
) {
494 struct io_ring_ctx
*ctx
= req
->ctx
;
496 spin_lock_irq(&ctx
->timeout_lock
);
497 io_for_each_link(cur
, req
)
498 io_prep_async_work(cur
);
499 spin_unlock_irq(&ctx
->timeout_lock
);
501 io_for_each_link(cur
, req
)
502 io_prep_async_work(cur
);
506 void io_queue_iowq(struct io_kiocb
*req
, struct io_tw_state
*ts_dont_use
)
508 struct io_kiocb
*link
= io_prep_linked_timeout(req
);
509 struct io_uring_task
*tctx
= req
->task
->io_uring
;
512 BUG_ON(!tctx
->io_wq
);
514 /* init ->work of the whole link before punting */
515 io_prep_async_link(req
);
518 * Not expected to happen, but if we do have a bug where this _can_
519 * happen, catch it here and ensure the request is marked as
520 * canceled. That will make io-wq go through the usual work cancel
521 * procedure rather than attempt to run this request (or create a new
524 if (WARN_ON_ONCE(!same_thread_group(req
->task
, current
)))
525 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
527 trace_io_uring_queue_async_work(req
, io_wq_is_hashed(&req
->work
));
528 io_wq_enqueue(tctx
->io_wq
, &req
->work
);
530 io_queue_linked_timeout(link
);
533 static __cold
void io_queue_deferred(struct io_ring_ctx
*ctx
)
535 while (!list_empty(&ctx
->defer_list
)) {
536 struct io_defer_entry
*de
= list_first_entry(&ctx
->defer_list
,
537 struct io_defer_entry
, list
);
539 if (req_need_defer(de
->req
, de
->seq
))
541 list_del_init(&de
->list
);
542 io_req_task_queue(de
->req
);
548 static void io_eventfd_ops(struct rcu_head
*rcu
)
550 struct io_ev_fd
*ev_fd
= container_of(rcu
, struct io_ev_fd
, rcu
);
551 int ops
= atomic_xchg(&ev_fd
->ops
, 0);
553 if (ops
& BIT(IO_EVENTFD_OP_SIGNAL_BIT
))
554 eventfd_signal_mask(ev_fd
->cq_ev_fd
, 1, EPOLL_URING_WAKE
);
556 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
557 * ordering in a race but if references are 0 we know we have to free
560 if (atomic_dec_and_test(&ev_fd
->refs
)) {
561 eventfd_ctx_put(ev_fd
->cq_ev_fd
);
566 static void io_eventfd_signal(struct io_ring_ctx
*ctx
)
568 struct io_ev_fd
*ev_fd
= NULL
;
572 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
575 ev_fd
= rcu_dereference(ctx
->io_ev_fd
);
578 * Check again if ev_fd exists incase an io_eventfd_unregister call
579 * completed between the NULL check of ctx->io_ev_fd at the start of
580 * the function and rcu_read_lock.
582 if (unlikely(!ev_fd
))
584 if (READ_ONCE(ctx
->rings
->cq_flags
) & IORING_CQ_EVENTFD_DISABLED
)
586 if (ev_fd
->eventfd_async
&& !io_wq_current_is_worker())
589 if (likely(eventfd_signal_allowed())) {
590 eventfd_signal_mask(ev_fd
->cq_ev_fd
, 1, EPOLL_URING_WAKE
);
592 atomic_inc(&ev_fd
->refs
);
593 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT
), &ev_fd
->ops
))
594 call_rcu_hurry(&ev_fd
->rcu
, io_eventfd_ops
);
596 atomic_dec(&ev_fd
->refs
);
603 static void io_eventfd_flush_signal(struct io_ring_ctx
*ctx
)
607 spin_lock(&ctx
->completion_lock
);
610 * Eventfd should only get triggered when at least one event has been
611 * posted. Some applications rely on the eventfd notification count
612 * only changing IFF a new CQE has been added to the CQ ring. There's
613 * no depedency on 1:1 relationship between how many times this
614 * function is called (and hence the eventfd count) and number of CQEs
615 * posted to the CQ ring.
617 skip
= ctx
->cached_cq_tail
== ctx
->evfd_last_cq_tail
;
618 ctx
->evfd_last_cq_tail
= ctx
->cached_cq_tail
;
619 spin_unlock(&ctx
->completion_lock
);
623 io_eventfd_signal(ctx
);
626 void __io_commit_cqring_flush(struct io_ring_ctx
*ctx
)
628 if (ctx
->poll_activated
)
629 io_poll_wq_wake(ctx
);
630 if (ctx
->off_timeout_used
)
631 io_flush_timeouts(ctx
);
632 if (ctx
->drain_active
) {
633 spin_lock(&ctx
->completion_lock
);
634 io_queue_deferred(ctx
);
635 spin_unlock(&ctx
->completion_lock
);
638 io_eventfd_flush_signal(ctx
);
641 static inline void __io_cq_lock(struct io_ring_ctx
*ctx
)
643 if (!ctx
->lockless_cq
)
644 spin_lock(&ctx
->completion_lock
);
647 static inline void io_cq_lock(struct io_ring_ctx
*ctx
)
648 __acquires(ctx
->completion_lock
)
650 spin_lock(&ctx
->completion_lock
);
653 static inline void __io_cq_unlock_post(struct io_ring_ctx
*ctx
)
655 io_commit_cqring(ctx
);
656 if (!ctx
->task_complete
) {
657 if (!ctx
->lockless_cq
)
658 spin_unlock(&ctx
->completion_lock
);
659 /* IOPOLL rings only need to wake up if it's also SQPOLL */
660 if (!ctx
->syscall_iopoll
)
663 io_commit_cqring_flush(ctx
);
666 static void io_cq_unlock_post(struct io_ring_ctx
*ctx
)
667 __releases(ctx
->completion_lock
)
669 io_commit_cqring(ctx
);
670 spin_unlock(&ctx
->completion_lock
);
672 io_commit_cqring_flush(ctx
);
675 /* Returns true if there are no backlogged entries after the flush */
676 static void io_cqring_overflow_kill(struct io_ring_ctx
*ctx
)
678 struct io_overflow_cqe
*ocqe
;
681 spin_lock(&ctx
->completion_lock
);
682 list_splice_init(&ctx
->cq_overflow_list
, &list
);
683 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
);
684 spin_unlock(&ctx
->completion_lock
);
686 while (!list_empty(&list
)) {
687 ocqe
= list_first_entry(&list
, struct io_overflow_cqe
, list
);
688 list_del(&ocqe
->list
);
693 static void __io_cqring_overflow_flush(struct io_ring_ctx
*ctx
)
695 size_t cqe_size
= sizeof(struct io_uring_cqe
);
697 if (__io_cqring_events(ctx
) == ctx
->cq_entries
)
700 if (ctx
->flags
& IORING_SETUP_CQE32
)
704 while (!list_empty(&ctx
->cq_overflow_list
)) {
705 struct io_uring_cqe
*cqe
;
706 struct io_overflow_cqe
*ocqe
;
708 if (!io_get_cqe_overflow(ctx
, &cqe
, true))
710 ocqe
= list_first_entry(&ctx
->cq_overflow_list
,
711 struct io_overflow_cqe
, list
);
712 memcpy(cqe
, &ocqe
->cqe
, cqe_size
);
713 list_del(&ocqe
->list
);
717 if (list_empty(&ctx
->cq_overflow_list
)) {
718 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
);
719 atomic_andnot(IORING_SQ_CQ_OVERFLOW
, &ctx
->rings
->sq_flags
);
721 io_cq_unlock_post(ctx
);
724 static void io_cqring_do_overflow_flush(struct io_ring_ctx
*ctx
)
726 /* iopoll syncs against uring_lock, not completion_lock */
727 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
728 mutex_lock(&ctx
->uring_lock
);
729 __io_cqring_overflow_flush(ctx
);
730 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
731 mutex_unlock(&ctx
->uring_lock
);
734 static void io_cqring_overflow_flush(struct io_ring_ctx
*ctx
)
736 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
))
737 io_cqring_do_overflow_flush(ctx
);
740 /* can be called by any task */
741 static void io_put_task_remote(struct task_struct
*task
)
743 struct io_uring_task
*tctx
= task
->io_uring
;
745 percpu_counter_sub(&tctx
->inflight
, 1);
746 if (unlikely(atomic_read(&tctx
->in_cancel
)))
747 wake_up(&tctx
->wait
);
748 put_task_struct(task
);
751 /* used by a task to put its own references */
752 static void io_put_task_local(struct task_struct
*task
)
754 task
->io_uring
->cached_refs
++;
757 /* must to be called somewhat shortly after putting a request */
758 static inline void io_put_task(struct task_struct
*task
)
760 if (likely(task
== current
))
761 io_put_task_local(task
);
763 io_put_task_remote(task
);
766 void io_task_refs_refill(struct io_uring_task
*tctx
)
768 unsigned int refill
= -tctx
->cached_refs
+ IO_TCTX_REFS_CACHE_NR
;
770 percpu_counter_add(&tctx
->inflight
, refill
);
771 refcount_add(refill
, ¤t
->usage
);
772 tctx
->cached_refs
+= refill
;
775 static __cold
void io_uring_drop_tctx_refs(struct task_struct
*task
)
777 struct io_uring_task
*tctx
= task
->io_uring
;
778 unsigned int refs
= tctx
->cached_refs
;
781 tctx
->cached_refs
= 0;
782 percpu_counter_sub(&tctx
->inflight
, refs
);
783 put_task_struct_many(task
, refs
);
787 static bool io_cqring_event_overflow(struct io_ring_ctx
*ctx
, u64 user_data
,
788 s32 res
, u32 cflags
, u64 extra1
, u64 extra2
)
790 struct io_overflow_cqe
*ocqe
;
791 size_t ocq_size
= sizeof(struct io_overflow_cqe
);
792 bool is_cqe32
= (ctx
->flags
& IORING_SETUP_CQE32
);
794 lockdep_assert_held(&ctx
->completion_lock
);
797 ocq_size
+= sizeof(struct io_uring_cqe
);
799 ocqe
= kmalloc(ocq_size
, GFP_ATOMIC
| __GFP_ACCOUNT
);
800 trace_io_uring_cqe_overflow(ctx
, user_data
, res
, cflags
, ocqe
);
803 * If we're in ring overflow flush mode, or in task cancel mode,
804 * or cannot allocate an overflow entry, then we need to drop it
807 io_account_cq_overflow(ctx
);
808 set_bit(IO_CHECK_CQ_DROPPED_BIT
, &ctx
->check_cq
);
811 if (list_empty(&ctx
->cq_overflow_list
)) {
812 set_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
);
813 atomic_or(IORING_SQ_CQ_OVERFLOW
, &ctx
->rings
->sq_flags
);
816 ocqe
->cqe
.user_data
= user_data
;
818 ocqe
->cqe
.flags
= cflags
;
820 ocqe
->cqe
.big_cqe
[0] = extra1
;
821 ocqe
->cqe
.big_cqe
[1] = extra2
;
823 list_add_tail(&ocqe
->list
, &ctx
->cq_overflow_list
);
827 void io_req_cqe_overflow(struct io_kiocb
*req
)
829 io_cqring_event_overflow(req
->ctx
, req
->cqe
.user_data
,
830 req
->cqe
.res
, req
->cqe
.flags
,
831 req
->big_cqe
.extra1
, req
->big_cqe
.extra2
);
832 memset(&req
->big_cqe
, 0, sizeof(req
->big_cqe
));
836 * writes to the cq entry need to come after reading head; the
837 * control dependency is enough as we're using WRITE_ONCE to
840 bool io_cqe_cache_refill(struct io_ring_ctx
*ctx
, bool overflow
)
842 struct io_rings
*rings
= ctx
->rings
;
843 unsigned int off
= ctx
->cached_cq_tail
& (ctx
->cq_entries
- 1);
844 unsigned int free
, queued
, len
;
847 * Posting into the CQ when there are pending overflowed CQEs may break
848 * ordering guarantees, which will affect links, F_MORE users and more.
849 * Force overflow the completion.
851 if (!overflow
&& (ctx
->check_cq
& BIT(IO_CHECK_CQ_OVERFLOW_BIT
)))
854 /* userspace may cheat modifying the tail, be safe and do min */
855 queued
= min(__io_cqring_events(ctx
), ctx
->cq_entries
);
856 free
= ctx
->cq_entries
- queued
;
857 /* we need a contiguous range, limit based on the current array offset */
858 len
= min(free
, ctx
->cq_entries
- off
);
862 if (ctx
->flags
& IORING_SETUP_CQE32
) {
867 ctx
->cqe_cached
= &rings
->cqes
[off
];
868 ctx
->cqe_sentinel
= ctx
->cqe_cached
+ len
;
872 static bool io_fill_cqe_aux(struct io_ring_ctx
*ctx
, u64 user_data
, s32 res
,
875 struct io_uring_cqe
*cqe
;
880 * If we can't get a cq entry, userspace overflowed the
881 * submission (by quite a lot). Increment the overflow count in
884 if (likely(io_get_cqe(ctx
, &cqe
))) {
885 trace_io_uring_complete(ctx
, NULL
, user_data
, res
, cflags
, 0, 0);
887 WRITE_ONCE(cqe
->user_data
, user_data
);
888 WRITE_ONCE(cqe
->res
, res
);
889 WRITE_ONCE(cqe
->flags
, cflags
);
891 if (ctx
->flags
& IORING_SETUP_CQE32
) {
892 WRITE_ONCE(cqe
->big_cqe
[0], 0);
893 WRITE_ONCE(cqe
->big_cqe
[1], 0);
900 static void __io_flush_post_cqes(struct io_ring_ctx
*ctx
)
901 __must_hold(&ctx
->uring_lock
)
903 struct io_submit_state
*state
= &ctx
->submit_state
;
906 lockdep_assert_held(&ctx
->uring_lock
);
907 for (i
= 0; i
< state
->cqes_count
; i
++) {
908 struct io_uring_cqe
*cqe
= &ctx
->completion_cqes
[i
];
910 if (!io_fill_cqe_aux(ctx
, cqe
->user_data
, cqe
->res
, cqe
->flags
)) {
911 if (ctx
->lockless_cq
) {
912 spin_lock(&ctx
->completion_lock
);
913 io_cqring_event_overflow(ctx
, cqe
->user_data
,
914 cqe
->res
, cqe
->flags
, 0, 0);
915 spin_unlock(&ctx
->completion_lock
);
917 io_cqring_event_overflow(ctx
, cqe
->user_data
,
918 cqe
->res
, cqe
->flags
, 0, 0);
922 state
->cqes_count
= 0;
925 static bool __io_post_aux_cqe(struct io_ring_ctx
*ctx
, u64 user_data
, s32 res
, u32 cflags
,
931 filled
= io_fill_cqe_aux(ctx
, user_data
, res
, cflags
);
932 if (!filled
&& allow_overflow
)
933 filled
= io_cqring_event_overflow(ctx
, user_data
, res
, cflags
, 0, 0);
935 io_cq_unlock_post(ctx
);
939 bool io_post_aux_cqe(struct io_ring_ctx
*ctx
, u64 user_data
, s32 res
, u32 cflags
)
941 return __io_post_aux_cqe(ctx
, user_data
, res
, cflags
, true);
945 * A helper for multishot requests posting additional CQEs.
946 * Should only be used from a task_work including IO_URING_F_MULTISHOT.
948 bool io_fill_cqe_req_aux(struct io_kiocb
*req
, bool defer
, s32 res
, u32 cflags
)
950 struct io_ring_ctx
*ctx
= req
->ctx
;
951 u64 user_data
= req
->cqe
.user_data
;
952 struct io_uring_cqe
*cqe
;
955 return __io_post_aux_cqe(ctx
, user_data
, res
, cflags
, false);
957 lockdep_assert_held(&ctx
->uring_lock
);
959 if (ctx
->submit_state
.cqes_count
== ARRAY_SIZE(ctx
->completion_cqes
)) {
961 __io_flush_post_cqes(ctx
);
962 /* no need to flush - flush is deferred */
963 __io_cq_unlock_post(ctx
);
966 /* For defered completions this is not as strict as it is otherwise,
967 * however it's main job is to prevent unbounded posted completions,
968 * and in that it works just as well.
970 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
))
973 cqe
= &ctx
->completion_cqes
[ctx
->submit_state
.cqes_count
++];
974 cqe
->user_data
= user_data
;
980 static void __io_req_complete_post(struct io_kiocb
*req
, unsigned issue_flags
)
982 struct io_ring_ctx
*ctx
= req
->ctx
;
983 struct io_rsrc_node
*rsrc_node
= NULL
;
986 if (!(req
->flags
& REQ_F_CQE_SKIP
)) {
987 if (!io_fill_cqe_req(ctx
, req
))
988 io_req_cqe_overflow(req
);
992 * If we're the last reference to this request, add to our locked
995 if (req_ref_put_and_test(req
)) {
996 if (req
->flags
& IO_REQ_LINK_FLAGS
) {
997 if (req
->flags
& IO_DISARM_MASK
)
1000 io_req_task_queue(req
->link
);
1004 io_put_kbuf_comp(req
);
1005 if (unlikely(req
->flags
& IO_REQ_CLEAN_FLAGS
))
1009 rsrc_node
= req
->rsrc_node
;
1011 * Selected buffer deallocation in io_clean_op() assumes that
1012 * we don't hold ->completion_lock. Clean them here to avoid
1015 io_put_task_remote(req
->task
);
1016 wq_list_add_head(&req
->comp_list
, &ctx
->locked_free_list
);
1017 ctx
->locked_free_nr
++;
1019 io_cq_unlock_post(ctx
);
1022 io_ring_submit_lock(ctx
, issue_flags
);
1023 io_put_rsrc_node(ctx
, rsrc_node
);
1024 io_ring_submit_unlock(ctx
, issue_flags
);
1028 void io_req_complete_post(struct io_kiocb
*req
, unsigned issue_flags
)
1030 if (req
->ctx
->task_complete
&& req
->ctx
->submitter_task
!= current
) {
1031 req
->io_task_work
.func
= io_req_task_complete
;
1032 io_req_task_work_add(req
);
1033 } else if (!(issue_flags
& IO_URING_F_UNLOCKED
) ||
1034 !(req
->ctx
->flags
& IORING_SETUP_IOPOLL
)) {
1035 __io_req_complete_post(req
, issue_flags
);
1037 struct io_ring_ctx
*ctx
= req
->ctx
;
1039 mutex_lock(&ctx
->uring_lock
);
1040 __io_req_complete_post(req
, issue_flags
& ~IO_URING_F_UNLOCKED
);
1041 mutex_unlock(&ctx
->uring_lock
);
1045 void io_req_defer_failed(struct io_kiocb
*req
, s32 res
)
1046 __must_hold(&ctx
->uring_lock
)
1048 const struct io_cold_def
*def
= &io_cold_defs
[req
->opcode
];
1050 lockdep_assert_held(&req
->ctx
->uring_lock
);
1053 io_req_set_res(req
, res
, io_put_kbuf(req
, IO_URING_F_UNLOCKED
));
1056 io_req_complete_defer(req
);
1060 * Don't initialise the fields below on every allocation, but do that in
1061 * advance and keep them valid across allocations.
1063 static void io_preinit_req(struct io_kiocb
*req
, struct io_ring_ctx
*ctx
)
1067 req
->async_data
= NULL
;
1068 /* not necessary, but safer to zero */
1069 memset(&req
->cqe
, 0, sizeof(req
->cqe
));
1070 memset(&req
->big_cqe
, 0, sizeof(req
->big_cqe
));
1073 static void io_flush_cached_locked_reqs(struct io_ring_ctx
*ctx
,
1074 struct io_submit_state
*state
)
1076 spin_lock(&ctx
->completion_lock
);
1077 wq_list_splice(&ctx
->locked_free_list
, &state
->free_list
);
1078 ctx
->locked_free_nr
= 0;
1079 spin_unlock(&ctx
->completion_lock
);
1083 * A request might get retired back into the request caches even before opcode
1084 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1085 * Because of that, io_alloc_req() should be called only under ->uring_lock
1086 * and with extra caution to not get a request that is still worked on.
1088 __cold
bool __io_alloc_req_refill(struct io_ring_ctx
*ctx
)
1089 __must_hold(&ctx
->uring_lock
)
1091 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
1092 void *reqs
[IO_REQ_ALLOC_BATCH
];
1096 * If we have more than a batch's worth of requests in our IRQ side
1097 * locked cache, grab the lock and move them over to our submission
1100 if (data_race(ctx
->locked_free_nr
) > IO_COMPL_BATCH
) {
1101 io_flush_cached_locked_reqs(ctx
, &ctx
->submit_state
);
1102 if (!io_req_cache_empty(ctx
))
1106 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, ARRAY_SIZE(reqs
), reqs
);
1109 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1110 * retry single alloc to be on the safe side.
1112 if (unlikely(ret
<= 0)) {
1113 reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
1119 percpu_ref_get_many(&ctx
->refs
, ret
);
1120 for (i
= 0; i
< ret
; i
++) {
1121 struct io_kiocb
*req
= reqs
[i
];
1123 io_preinit_req(req
, ctx
);
1124 io_req_add_to_cache(req
, ctx
);
1129 __cold
void io_free_req(struct io_kiocb
*req
)
1131 /* refs were already put, restore them for io_req_task_complete() */
1132 req
->flags
&= ~REQ_F_REFCOUNT
;
1133 /* we only want to free it, don't post CQEs */
1134 req
->flags
|= REQ_F_CQE_SKIP
;
1135 req
->io_task_work
.func
= io_req_task_complete
;
1136 io_req_task_work_add(req
);
1139 static void __io_req_find_next_prep(struct io_kiocb
*req
)
1141 struct io_ring_ctx
*ctx
= req
->ctx
;
1143 spin_lock(&ctx
->completion_lock
);
1144 io_disarm_next(req
);
1145 spin_unlock(&ctx
->completion_lock
);
1148 static inline struct io_kiocb
*io_req_find_next(struct io_kiocb
*req
)
1150 struct io_kiocb
*nxt
;
1153 * If LINK is set, we have dependent requests in this chain. If we
1154 * didn't fail this request, queue the first one up, moving any other
1155 * dependencies to the next request. In case of failure, fail the rest
1158 if (unlikely(req
->flags
& IO_DISARM_MASK
))
1159 __io_req_find_next_prep(req
);
1165 static void ctx_flush_and_put(struct io_ring_ctx
*ctx
, struct io_tw_state
*ts
)
1169 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
)
1170 atomic_andnot(IORING_SQ_TASKRUN
, &ctx
->rings
->sq_flags
);
1172 io_submit_flush_completions(ctx
);
1173 mutex_unlock(&ctx
->uring_lock
);
1176 percpu_ref_put(&ctx
->refs
);
1179 static unsigned int handle_tw_list(struct llist_node
*node
,
1180 struct io_ring_ctx
**ctx
,
1181 struct io_tw_state
*ts
,
1182 struct llist_node
*last
)
1184 unsigned int count
= 0;
1186 while (node
&& node
!= last
) {
1187 struct llist_node
*next
= node
->next
;
1188 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1191 prefetch(container_of(next
, struct io_kiocb
, io_task_work
.node
));
1193 if (req
->ctx
!= *ctx
) {
1194 ctx_flush_and_put(*ctx
, ts
);
1196 /* if not contended, grab and improve batching */
1197 ts
->locked
= mutex_trylock(&(*ctx
)->uring_lock
);
1198 percpu_ref_get(&(*ctx
)->refs
);
1200 INDIRECT_CALL_2(req
->io_task_work
.func
,
1201 io_poll_task_func
, io_req_rw_complete
,
1205 if (unlikely(need_resched())) {
1206 ctx_flush_and_put(*ctx
, ts
);
1216 * io_llist_xchg - swap all entries in a lock-less list
1217 * @head: the head of lock-less list to delete all entries
1218 * @new: new entry as the head of the list
1220 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1221 * The order of entries returned is from the newest to the oldest added one.
1223 static inline struct llist_node
*io_llist_xchg(struct llist_head
*head
,
1224 struct llist_node
*new)
1226 return xchg(&head
->first
, new);
1230 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1231 * @head: the head of lock-less list to delete all entries
1232 * @old: expected old value of the first entry of the list
1233 * @new: new entry as the head of the list
1235 * perform a cmpxchg on the first entry of the list.
1238 static inline struct llist_node
*io_llist_cmpxchg(struct llist_head
*head
,
1239 struct llist_node
*old
,
1240 struct llist_node
*new)
1242 return cmpxchg(&head
->first
, old
, new);
1245 static __cold
void io_fallback_tw(struct io_uring_task
*tctx
, bool sync
)
1247 struct llist_node
*node
= llist_del_all(&tctx
->task_list
);
1248 struct io_ring_ctx
*last_ctx
= NULL
;
1249 struct io_kiocb
*req
;
1252 req
= container_of(node
, struct io_kiocb
, io_task_work
.node
);
1254 if (sync
&& last_ctx
!= req
->ctx
) {
1256 flush_delayed_work(&last_ctx
->fallback_work
);
1257 percpu_ref_put(&last_ctx
->refs
);
1259 last_ctx
= req
->ctx
;
1260 percpu_ref_get(&last_ctx
->refs
);
1262 if (llist_add(&req
->io_task_work
.node
,
1263 &req
->ctx
->fallback_llist
))
1264 schedule_delayed_work(&req
->ctx
->fallback_work
, 1);
1268 flush_delayed_work(&last_ctx
->fallback_work
);
1269 percpu_ref_put(&last_ctx
->refs
);
1273 void tctx_task_work(struct callback_head
*cb
)
1275 struct io_tw_state ts
= {};
1276 struct io_ring_ctx
*ctx
= NULL
;
1277 struct io_uring_task
*tctx
= container_of(cb
, struct io_uring_task
,
1279 struct llist_node fake
= {};
1280 struct llist_node
*node
;
1281 unsigned int loops
= 0;
1282 unsigned int count
= 0;
1284 if (unlikely(current
->flags
& PF_EXITING
)) {
1285 io_fallback_tw(tctx
, true);
1291 node
= io_llist_xchg(&tctx
->task_list
, &fake
);
1292 count
+= handle_tw_list(node
, &ctx
, &ts
, &fake
);
1294 /* skip expensive cmpxchg if there are items in the list */
1295 if (READ_ONCE(tctx
->task_list
.first
) != &fake
)
1297 if (ts
.locked
&& !wq_list_empty(&ctx
->submit_state
.compl_reqs
)) {
1298 io_submit_flush_completions(ctx
);
1299 if (READ_ONCE(tctx
->task_list
.first
) != &fake
)
1302 node
= io_llist_cmpxchg(&tctx
->task_list
, &fake
, NULL
);
1303 } while (node
!= &fake
);
1305 ctx_flush_and_put(ctx
, &ts
);
1307 /* relaxed read is enough as only the task itself sets ->in_cancel */
1308 if (unlikely(atomic_read(&tctx
->in_cancel
)))
1309 io_uring_drop_tctx_refs(current
);
1311 trace_io_uring_task_work_run(tctx
, count
, loops
);
1314 static inline void io_req_local_work_add(struct io_kiocb
*req
, unsigned flags
)
1316 struct io_ring_ctx
*ctx
= req
->ctx
;
1317 unsigned nr_wait
, nr_tw
, nr_tw_prev
;
1318 struct llist_node
*first
;
1320 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))
1321 flags
&= ~IOU_F_TWQ_LAZY_WAKE
;
1323 first
= READ_ONCE(ctx
->work_llist
.first
);
1327 struct io_kiocb
*first_req
= container_of(first
,
1331 * Might be executed at any moment, rely on
1332 * SLAB_TYPESAFE_BY_RCU to keep it alive.
1334 nr_tw_prev
= READ_ONCE(first_req
->nr_tw
);
1336 nr_tw
= nr_tw_prev
+ 1;
1337 /* Large enough to fail the nr_wait comparison below */
1338 if (!(flags
& IOU_F_TWQ_LAZY_WAKE
))
1342 req
->io_task_work
.node
.next
= first
;
1343 } while (!try_cmpxchg(&ctx
->work_llist
.first
, &first
,
1344 &req
->io_task_work
.node
));
1347 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
)
1348 atomic_or(IORING_SQ_TASKRUN
, &ctx
->rings
->sq_flags
);
1350 io_eventfd_signal(ctx
);
1353 nr_wait
= atomic_read(&ctx
->cq_wait_nr
);
1354 /* no one is waiting */
1357 /* either not enough or the previous add has already woken it up */
1358 if (nr_wait
> nr_tw
|| nr_tw_prev
>= nr_wait
)
1360 /* pairs with set_current_state() in io_cqring_wait() */
1361 smp_mb__after_atomic();
1362 wake_up_state(ctx
->submitter_task
, TASK_INTERRUPTIBLE
);
1365 static void io_req_normal_work_add(struct io_kiocb
*req
)
1367 struct io_uring_task
*tctx
= req
->task
->io_uring
;
1368 struct io_ring_ctx
*ctx
= req
->ctx
;
1370 /* task_work already pending, we're done */
1371 if (!llist_add(&req
->io_task_work
.node
, &tctx
->task_list
))
1374 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
)
1375 atomic_or(IORING_SQ_TASKRUN
, &ctx
->rings
->sq_flags
);
1377 if (likely(!task_work_add(req
->task
, &tctx
->task_work
, ctx
->notify_method
)))
1380 io_fallback_tw(tctx
, false);
1383 void __io_req_task_work_add(struct io_kiocb
*req
, unsigned flags
)
1385 if (req
->ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) {
1387 io_req_local_work_add(req
, flags
);
1390 io_req_normal_work_add(req
);
1394 static void __cold
io_move_task_work_from_local(struct io_ring_ctx
*ctx
)
1396 struct llist_node
*node
;
1398 node
= llist_del_all(&ctx
->work_llist
);
1400 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1404 io_req_normal_work_add(req
);
1408 static int __io_run_local_work(struct io_ring_ctx
*ctx
, struct io_tw_state
*ts
)
1410 struct llist_node
*node
;
1411 unsigned int loops
= 0;
1414 if (WARN_ON_ONCE(ctx
->submitter_task
!= current
))
1416 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
)
1417 atomic_andnot(IORING_SQ_TASKRUN
, &ctx
->rings
->sq_flags
);
1420 * llists are in reverse order, flip it back the right way before
1421 * running the pending items.
1423 node
= llist_reverse_order(io_llist_xchg(&ctx
->work_llist
, NULL
));
1425 struct llist_node
*next
= node
->next
;
1426 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1428 prefetch(container_of(next
, struct io_kiocb
, io_task_work
.node
));
1429 INDIRECT_CALL_2(req
->io_task_work
.func
,
1430 io_poll_task_func
, io_req_rw_complete
,
1437 if (!llist_empty(&ctx
->work_llist
))
1440 io_submit_flush_completions(ctx
);
1441 if (!llist_empty(&ctx
->work_llist
))
1444 trace_io_uring_local_work_run(ctx
, ret
, loops
);
1448 static inline int io_run_local_work_locked(struct io_ring_ctx
*ctx
)
1450 struct io_tw_state ts
= { .locked
= true, };
1453 if (llist_empty(&ctx
->work_llist
))
1456 ret
= __io_run_local_work(ctx
, &ts
);
1457 /* shouldn't happen! */
1458 if (WARN_ON_ONCE(!ts
.locked
))
1459 mutex_lock(&ctx
->uring_lock
);
1463 static int io_run_local_work(struct io_ring_ctx
*ctx
)
1465 struct io_tw_state ts
= {};
1468 ts
.locked
= mutex_trylock(&ctx
->uring_lock
);
1469 ret
= __io_run_local_work(ctx
, &ts
);
1471 mutex_unlock(&ctx
->uring_lock
);
1476 static void io_req_task_cancel(struct io_kiocb
*req
, struct io_tw_state
*ts
)
1478 io_tw_lock(req
->ctx
, ts
);
1479 io_req_defer_failed(req
, req
->cqe
.res
);
1482 void io_req_task_submit(struct io_kiocb
*req
, struct io_tw_state
*ts
)
1484 io_tw_lock(req
->ctx
, ts
);
1485 /* req->task == current here, checking PF_EXITING is safe */
1486 if (unlikely(req
->task
->flags
& PF_EXITING
))
1487 io_req_defer_failed(req
, -EFAULT
);
1488 else if (req
->flags
& REQ_F_FORCE_ASYNC
)
1489 io_queue_iowq(req
, ts
);
1494 void io_req_task_queue_fail(struct io_kiocb
*req
, int ret
)
1496 io_req_set_res(req
, ret
, 0);
1497 req
->io_task_work
.func
= io_req_task_cancel
;
1498 io_req_task_work_add(req
);
1501 void io_req_task_queue(struct io_kiocb
*req
)
1503 req
->io_task_work
.func
= io_req_task_submit
;
1504 io_req_task_work_add(req
);
1507 void io_queue_next(struct io_kiocb
*req
)
1509 struct io_kiocb
*nxt
= io_req_find_next(req
);
1512 io_req_task_queue(nxt
);
1515 static void io_free_batch_list(struct io_ring_ctx
*ctx
,
1516 struct io_wq_work_node
*node
)
1517 __must_hold(&ctx
->uring_lock
)
1520 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1523 if (unlikely(req
->flags
& IO_REQ_CLEAN_SLOW_FLAGS
)) {
1524 if (req
->flags
& REQ_F_REFCOUNT
) {
1525 node
= req
->comp_list
.next
;
1526 if (!req_ref_put_and_test(req
))
1529 if ((req
->flags
& REQ_F_POLLED
) && req
->apoll
) {
1530 struct async_poll
*apoll
= req
->apoll
;
1532 if (apoll
->double_poll
)
1533 kfree(apoll
->double_poll
);
1534 if (!io_alloc_cache_put(&ctx
->apoll_cache
, &apoll
->cache
))
1536 req
->flags
&= ~REQ_F_POLLED
;
1538 if (req
->flags
& IO_REQ_LINK_FLAGS
)
1540 if (unlikely(req
->flags
& IO_REQ_CLEAN_FLAGS
))
1545 io_req_put_rsrc_locked(req
, ctx
);
1547 io_put_task(req
->task
);
1548 node
= req
->comp_list
.next
;
1549 io_req_add_to_cache(req
, ctx
);
1553 void __io_submit_flush_completions(struct io_ring_ctx
*ctx
)
1554 __must_hold(&ctx
->uring_lock
)
1556 struct io_submit_state
*state
= &ctx
->submit_state
;
1557 struct io_wq_work_node
*node
;
1560 /* must come first to preserve CQE ordering in failure cases */
1561 if (state
->cqes_count
)
1562 __io_flush_post_cqes(ctx
);
1563 __wq_list_for_each(node
, &state
->compl_reqs
) {
1564 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1567 if (!(req
->flags
& REQ_F_CQE_SKIP
) &&
1568 unlikely(!io_fill_cqe_req(ctx
, req
))) {
1569 if (ctx
->lockless_cq
) {
1570 spin_lock(&ctx
->completion_lock
);
1571 io_req_cqe_overflow(req
);
1572 spin_unlock(&ctx
->completion_lock
);
1574 io_req_cqe_overflow(req
);
1578 __io_cq_unlock_post(ctx
);
1580 if (!wq_list_empty(&ctx
->submit_state
.compl_reqs
)) {
1581 io_free_batch_list(ctx
, state
->compl_reqs
.first
);
1582 INIT_WQ_LIST(&state
->compl_reqs
);
1586 static unsigned io_cqring_events(struct io_ring_ctx
*ctx
)
1588 /* See comment at the top of this file */
1590 return __io_cqring_events(ctx
);
1594 * We can't just wait for polled events to come to us, we have to actively
1595 * find and complete them.
1597 static __cold
void io_iopoll_try_reap_events(struct io_ring_ctx
*ctx
)
1599 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
1602 mutex_lock(&ctx
->uring_lock
);
1603 while (!wq_list_empty(&ctx
->iopoll_list
)) {
1604 /* let it sleep and repeat later if can't complete a request */
1605 if (io_do_iopoll(ctx
, true) == 0)
1608 * Ensure we allow local-to-the-cpu processing to take place,
1609 * in this case we need to ensure that we reap all events.
1610 * Also let task_work, etc. to progress by releasing the mutex
1612 if (need_resched()) {
1613 mutex_unlock(&ctx
->uring_lock
);
1615 mutex_lock(&ctx
->uring_lock
);
1618 mutex_unlock(&ctx
->uring_lock
);
1621 static int io_iopoll_check(struct io_ring_ctx
*ctx
, long min
)
1623 unsigned int nr_events
= 0;
1624 unsigned long check_cq
;
1626 if (!io_allowed_run_tw(ctx
))
1629 check_cq
= READ_ONCE(ctx
->check_cq
);
1630 if (unlikely(check_cq
)) {
1631 if (check_cq
& BIT(IO_CHECK_CQ_OVERFLOW_BIT
))
1632 __io_cqring_overflow_flush(ctx
);
1634 * Similarly do not spin if we have not informed the user of any
1637 if (check_cq
& BIT(IO_CHECK_CQ_DROPPED_BIT
))
1641 * Don't enter poll loop if we already have events pending.
1642 * If we do, we can potentially be spinning for commands that
1643 * already triggered a CQE (eg in error).
1645 if (io_cqring_events(ctx
))
1652 * If a submit got punted to a workqueue, we can have the
1653 * application entering polling for a command before it gets
1654 * issued. That app will hold the uring_lock for the duration
1655 * of the poll right here, so we need to take a breather every
1656 * now and then to ensure that the issue has a chance to add
1657 * the poll to the issued list. Otherwise we can spin here
1658 * forever, while the workqueue is stuck trying to acquire the
1661 if (wq_list_empty(&ctx
->iopoll_list
) ||
1662 io_task_work_pending(ctx
)) {
1663 u32 tail
= ctx
->cached_cq_tail
;
1665 (void) io_run_local_work_locked(ctx
);
1667 if (task_work_pending(current
) ||
1668 wq_list_empty(&ctx
->iopoll_list
)) {
1669 mutex_unlock(&ctx
->uring_lock
);
1671 mutex_lock(&ctx
->uring_lock
);
1673 /* some requests don't go through iopoll_list */
1674 if (tail
!= ctx
->cached_cq_tail
||
1675 wq_list_empty(&ctx
->iopoll_list
))
1678 ret
= io_do_iopoll(ctx
, !min
);
1679 if (unlikely(ret
< 0))
1682 if (task_sigpending(current
))
1688 } while (nr_events
< min
);
1693 void io_req_task_complete(struct io_kiocb
*req
, struct io_tw_state
*ts
)
1696 io_req_complete_defer(req
);
1698 io_req_complete_post(req
, IO_URING_F_UNLOCKED
);
1702 * After the iocb has been issued, it's safe to be found on the poll list.
1703 * Adding the kiocb to the list AFTER submission ensures that we don't
1704 * find it from a io_do_iopoll() thread before the issuer is done
1705 * accessing the kiocb cookie.
1707 static void io_iopoll_req_issued(struct io_kiocb
*req
, unsigned int issue_flags
)
1709 struct io_ring_ctx
*ctx
= req
->ctx
;
1710 const bool needs_lock
= issue_flags
& IO_URING_F_UNLOCKED
;
1712 /* workqueue context doesn't hold uring_lock, grab it now */
1713 if (unlikely(needs_lock
))
1714 mutex_lock(&ctx
->uring_lock
);
1717 * Track whether we have multiple files in our lists. This will impact
1718 * how we do polling eventually, not spinning if we're on potentially
1719 * different devices.
1721 if (wq_list_empty(&ctx
->iopoll_list
)) {
1722 ctx
->poll_multi_queue
= false;
1723 } else if (!ctx
->poll_multi_queue
) {
1724 struct io_kiocb
*list_req
;
1726 list_req
= container_of(ctx
->iopoll_list
.first
, struct io_kiocb
,
1728 if (list_req
->file
!= req
->file
)
1729 ctx
->poll_multi_queue
= true;
1733 * For fast devices, IO may have already completed. If it has, add
1734 * it to the front so we find it first.
1736 if (READ_ONCE(req
->iopoll_completed
))
1737 wq_list_add_head(&req
->comp_list
, &ctx
->iopoll_list
);
1739 wq_list_add_tail(&req
->comp_list
, &ctx
->iopoll_list
);
1741 if (unlikely(needs_lock
)) {
1743 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1744 * in sq thread task context or in io worker task context. If
1745 * current task context is sq thread, we don't need to check
1746 * whether should wake up sq thread.
1748 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) &&
1749 wq_has_sleeper(&ctx
->sq_data
->wait
))
1750 wake_up(&ctx
->sq_data
->wait
);
1752 mutex_unlock(&ctx
->uring_lock
);
1756 unsigned int io_file_get_flags(struct file
*file
)
1758 unsigned int res
= 0;
1760 if (S_ISREG(file_inode(file
)->i_mode
))
1762 if ((file
->f_flags
& O_NONBLOCK
) || (file
->f_mode
& FMODE_NOWAIT
))
1763 res
|= REQ_F_SUPPORT_NOWAIT
;
1767 bool io_alloc_async_data(struct io_kiocb
*req
)
1769 WARN_ON_ONCE(!io_cold_defs
[req
->opcode
].async_size
);
1770 req
->async_data
= kmalloc(io_cold_defs
[req
->opcode
].async_size
, GFP_KERNEL
);
1771 if (req
->async_data
) {
1772 req
->flags
|= REQ_F_ASYNC_DATA
;
1778 int io_req_prep_async(struct io_kiocb
*req
)
1780 const struct io_cold_def
*cdef
= &io_cold_defs
[req
->opcode
];
1781 const struct io_issue_def
*def
= &io_issue_defs
[req
->opcode
];
1783 /* assign early for deferred execution for non-fixed file */
1784 if (def
->needs_file
&& !(req
->flags
& REQ_F_FIXED_FILE
) && !req
->file
)
1785 req
->file
= io_file_get_normal(req
, req
->cqe
.fd
);
1786 if (!cdef
->prep_async
)
1788 if (WARN_ON_ONCE(req_has_async_data(req
)))
1790 if (!def
->manual_alloc
) {
1791 if (io_alloc_async_data(req
))
1794 return cdef
->prep_async(req
);
1797 static u32
io_get_sequence(struct io_kiocb
*req
)
1799 u32 seq
= req
->ctx
->cached_sq_head
;
1800 struct io_kiocb
*cur
;
1802 /* need original cached_sq_head, but it was increased for each req */
1803 io_for_each_link(cur
, req
)
1808 static __cold
void io_drain_req(struct io_kiocb
*req
)
1809 __must_hold(&ctx
->uring_lock
)
1811 struct io_ring_ctx
*ctx
= req
->ctx
;
1812 struct io_defer_entry
*de
;
1814 u32 seq
= io_get_sequence(req
);
1816 /* Still need defer if there is pending req in defer list. */
1817 spin_lock(&ctx
->completion_lock
);
1818 if (!req_need_defer(req
, seq
) && list_empty_careful(&ctx
->defer_list
)) {
1819 spin_unlock(&ctx
->completion_lock
);
1821 ctx
->drain_active
= false;
1822 io_req_task_queue(req
);
1825 spin_unlock(&ctx
->completion_lock
);
1827 io_prep_async_link(req
);
1828 de
= kmalloc(sizeof(*de
), GFP_KERNEL
);
1831 io_req_defer_failed(req
, ret
);
1835 spin_lock(&ctx
->completion_lock
);
1836 if (!req_need_defer(req
, seq
) && list_empty(&ctx
->defer_list
)) {
1837 spin_unlock(&ctx
->completion_lock
);
1842 trace_io_uring_defer(req
);
1845 list_add_tail(&de
->list
, &ctx
->defer_list
);
1846 spin_unlock(&ctx
->completion_lock
);
1849 static bool io_assign_file(struct io_kiocb
*req
, const struct io_issue_def
*def
,
1850 unsigned int issue_flags
)
1852 if (req
->file
|| !def
->needs_file
)
1855 if (req
->flags
& REQ_F_FIXED_FILE
)
1856 req
->file
= io_file_get_fixed(req
, req
->cqe
.fd
, issue_flags
);
1858 req
->file
= io_file_get_normal(req
, req
->cqe
.fd
);
1863 static int io_issue_sqe(struct io_kiocb
*req
, unsigned int issue_flags
)
1865 const struct io_issue_def
*def
= &io_issue_defs
[req
->opcode
];
1866 const struct cred
*creds
= NULL
;
1869 if (unlikely(!io_assign_file(req
, def
, issue_flags
)))
1872 if (unlikely((req
->flags
& REQ_F_CREDS
) && req
->creds
!= current_cred()))
1873 creds
= override_creds(req
->creds
);
1875 if (!def
->audit_skip
)
1876 audit_uring_entry(req
->opcode
);
1878 ret
= def
->issue(req
, issue_flags
);
1880 if (!def
->audit_skip
)
1881 audit_uring_exit(!ret
, ret
);
1884 revert_creds(creds
);
1886 if (ret
== IOU_OK
) {
1887 if (issue_flags
& IO_URING_F_COMPLETE_DEFER
)
1888 io_req_complete_defer(req
);
1890 io_req_complete_post(req
, issue_flags
);
1891 } else if (ret
!= IOU_ISSUE_SKIP_COMPLETE
)
1894 /* If the op doesn't have a file, we're not polling for it */
1895 if ((req
->ctx
->flags
& IORING_SETUP_IOPOLL
) && def
->iopoll_queue
)
1896 io_iopoll_req_issued(req
, issue_flags
);
1901 int io_poll_issue(struct io_kiocb
*req
, struct io_tw_state
*ts
)
1903 io_tw_lock(req
->ctx
, ts
);
1904 return io_issue_sqe(req
, IO_URING_F_NONBLOCK
|IO_URING_F_MULTISHOT
|
1905 IO_URING_F_COMPLETE_DEFER
);
1908 struct io_wq_work
*io_wq_free_work(struct io_wq_work
*work
)
1910 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
1911 struct io_kiocb
*nxt
= NULL
;
1913 if (req_ref_put_and_test(req
)) {
1914 if (req
->flags
& IO_REQ_LINK_FLAGS
)
1915 nxt
= io_req_find_next(req
);
1918 return nxt
? &nxt
->work
: NULL
;
1921 void io_wq_submit_work(struct io_wq_work
*work
)
1923 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
1924 const struct io_issue_def
*def
= &io_issue_defs
[req
->opcode
];
1925 unsigned int issue_flags
= IO_URING_F_UNLOCKED
| IO_URING_F_IOWQ
;
1926 bool needs_poll
= false;
1927 int ret
= 0, err
= -ECANCELED
;
1929 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1930 if (!(req
->flags
& REQ_F_REFCOUNT
))
1931 __io_req_set_refcount(req
, 2);
1935 io_arm_ltimeout(req
);
1937 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1938 if (work
->flags
& IO_WQ_WORK_CANCEL
) {
1940 io_req_task_queue_fail(req
, err
);
1943 if (!io_assign_file(req
, def
, issue_flags
)) {
1945 work
->flags
|= IO_WQ_WORK_CANCEL
;
1949 if (req
->flags
& REQ_F_FORCE_ASYNC
) {
1950 bool opcode_poll
= def
->pollin
|| def
->pollout
;
1952 if (opcode_poll
&& file_can_poll(req
->file
)) {
1954 issue_flags
|= IO_URING_F_NONBLOCK
;
1959 ret
= io_issue_sqe(req
, issue_flags
);
1964 * If REQ_F_NOWAIT is set, then don't wait or retry with
1965 * poll. -EAGAIN is final for that case.
1967 if (req
->flags
& REQ_F_NOWAIT
)
1971 * We can get EAGAIN for iopolled IO even though we're
1972 * forcing a sync submission from here, since we can't
1973 * wait for request slots on the block side.
1976 if (!(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
1978 if (io_wq_worker_stopped())
1984 if (io_arm_poll_handler(req
, issue_flags
) == IO_APOLL_OK
)
1986 /* aborted or ready, in either case retry blocking */
1988 issue_flags
&= ~IO_URING_F_NONBLOCK
;
1991 /* avoid locking problems by failing it from a clean context */
1993 io_req_task_queue_fail(req
, ret
);
1996 inline struct file
*io_file_get_fixed(struct io_kiocb
*req
, int fd
,
1997 unsigned int issue_flags
)
1999 struct io_ring_ctx
*ctx
= req
->ctx
;
2000 struct io_fixed_file
*slot
;
2001 struct file
*file
= NULL
;
2003 io_ring_submit_lock(ctx
, issue_flags
);
2005 if (unlikely((unsigned int)fd
>= ctx
->nr_user_files
))
2007 fd
= array_index_nospec(fd
, ctx
->nr_user_files
);
2008 slot
= io_fixed_file_slot(&ctx
->file_table
, fd
);
2009 file
= io_slot_file(slot
);
2010 req
->flags
|= io_slot_flags(slot
);
2011 io_req_set_rsrc_node(req
, ctx
, 0);
2013 io_ring_submit_unlock(ctx
, issue_flags
);
2017 struct file
*io_file_get_normal(struct io_kiocb
*req
, int fd
)
2019 struct file
*file
= fget(fd
);
2021 trace_io_uring_file_get(req
, fd
);
2023 /* we don't allow fixed io_uring files */
2024 if (file
&& io_is_uring_fops(file
))
2025 io_req_track_inflight(req
);
2029 static void io_queue_async(struct io_kiocb
*req
, int ret
)
2030 __must_hold(&req
->ctx
->uring_lock
)
2032 struct io_kiocb
*linked_timeout
;
2034 if (ret
!= -EAGAIN
|| (req
->flags
& REQ_F_NOWAIT
)) {
2035 io_req_defer_failed(req
, ret
);
2039 linked_timeout
= io_prep_linked_timeout(req
);
2041 switch (io_arm_poll_handler(req
, 0)) {
2042 case IO_APOLL_READY
:
2043 io_kbuf_recycle(req
, 0);
2044 io_req_task_queue(req
);
2046 case IO_APOLL_ABORTED
:
2047 io_kbuf_recycle(req
, 0);
2048 io_queue_iowq(req
, NULL
);
2055 io_queue_linked_timeout(linked_timeout
);
2058 static inline void io_queue_sqe(struct io_kiocb
*req
)
2059 __must_hold(&req
->ctx
->uring_lock
)
2063 ret
= io_issue_sqe(req
, IO_URING_F_NONBLOCK
|IO_URING_F_COMPLETE_DEFER
);
2066 * We async punt it if the file wasn't marked NOWAIT, or if the file
2067 * doesn't support non-blocking read/write attempts
2070 io_arm_ltimeout(req
);
2072 io_queue_async(req
, ret
);
2075 static void io_queue_sqe_fallback(struct io_kiocb
*req
)
2076 __must_hold(&req
->ctx
->uring_lock
)
2078 if (unlikely(req
->flags
& REQ_F_FAIL
)) {
2080 * We don't submit, fail them all, for that replace hardlinks
2081 * with normal links. Extra REQ_F_LINK is tolerated.
2083 req
->flags
&= ~REQ_F_HARDLINK
;
2084 req
->flags
|= REQ_F_LINK
;
2085 io_req_defer_failed(req
, req
->cqe
.res
);
2087 int ret
= io_req_prep_async(req
);
2089 if (unlikely(ret
)) {
2090 io_req_defer_failed(req
, ret
);
2094 if (unlikely(req
->ctx
->drain_active
))
2097 io_queue_iowq(req
, NULL
);
2102 * Check SQE restrictions (opcode and flags).
2104 * Returns 'true' if SQE is allowed, 'false' otherwise.
2106 static inline bool io_check_restriction(struct io_ring_ctx
*ctx
,
2107 struct io_kiocb
*req
,
2108 unsigned int sqe_flags
)
2110 if (!test_bit(req
->opcode
, ctx
->restrictions
.sqe_op
))
2113 if ((sqe_flags
& ctx
->restrictions
.sqe_flags_required
) !=
2114 ctx
->restrictions
.sqe_flags_required
)
2117 if (sqe_flags
& ~(ctx
->restrictions
.sqe_flags_allowed
|
2118 ctx
->restrictions
.sqe_flags_required
))
2124 static void io_init_req_drain(struct io_kiocb
*req
)
2126 struct io_ring_ctx
*ctx
= req
->ctx
;
2127 struct io_kiocb
*head
= ctx
->submit_state
.link
.head
;
2129 ctx
->drain_active
= true;
2132 * If we need to drain a request in the middle of a link, drain
2133 * the head request and the next request/link after the current
2134 * link. Considering sequential execution of links,
2135 * REQ_F_IO_DRAIN will be maintained for every request of our
2138 head
->flags
|= REQ_F_IO_DRAIN
| REQ_F_FORCE_ASYNC
;
2139 ctx
->drain_next
= true;
2143 static int io_init_req(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
2144 const struct io_uring_sqe
*sqe
)
2145 __must_hold(&ctx
->uring_lock
)
2147 const struct io_issue_def
*def
;
2148 unsigned int sqe_flags
;
2152 /* req is partially pre-initialised, see io_preinit_req() */
2153 req
->opcode
= opcode
= READ_ONCE(sqe
->opcode
);
2154 /* same numerical values with corresponding REQ_F_*, safe to copy */
2155 req
->flags
= sqe_flags
= READ_ONCE(sqe
->flags
);
2156 req
->cqe
.user_data
= READ_ONCE(sqe
->user_data
);
2158 req
->rsrc_node
= NULL
;
2159 req
->task
= current
;
2161 if (unlikely(opcode
>= IORING_OP_LAST
)) {
2165 def
= &io_issue_defs
[opcode
];
2166 if (unlikely(sqe_flags
& ~SQE_COMMON_FLAGS
)) {
2167 /* enforce forwards compatibility on users */
2168 if (sqe_flags
& ~SQE_VALID_FLAGS
)
2170 if (sqe_flags
& IOSQE_BUFFER_SELECT
) {
2171 if (!def
->buffer_select
)
2173 req
->buf_index
= READ_ONCE(sqe
->buf_group
);
2175 if (sqe_flags
& IOSQE_CQE_SKIP_SUCCESS
)
2176 ctx
->drain_disabled
= true;
2177 if (sqe_flags
& IOSQE_IO_DRAIN
) {
2178 if (ctx
->drain_disabled
)
2180 io_init_req_drain(req
);
2183 if (unlikely(ctx
->restricted
|| ctx
->drain_active
|| ctx
->drain_next
)) {
2184 if (ctx
->restricted
&& !io_check_restriction(ctx
, req
, sqe_flags
))
2186 /* knock it to the slow queue path, will be drained there */
2187 if (ctx
->drain_active
)
2188 req
->flags
|= REQ_F_FORCE_ASYNC
;
2189 /* if there is no link, we're at "next" request and need to drain */
2190 if (unlikely(ctx
->drain_next
) && !ctx
->submit_state
.link
.head
) {
2191 ctx
->drain_next
= false;
2192 ctx
->drain_active
= true;
2193 req
->flags
|= REQ_F_IO_DRAIN
| REQ_F_FORCE_ASYNC
;
2197 if (!def
->ioprio
&& sqe
->ioprio
)
2199 if (!def
->iopoll
&& (ctx
->flags
& IORING_SETUP_IOPOLL
))
2202 if (def
->needs_file
) {
2203 struct io_submit_state
*state
= &ctx
->submit_state
;
2205 req
->cqe
.fd
= READ_ONCE(sqe
->fd
);
2208 * Plug now if we have more than 2 IO left after this, and the
2209 * target is potentially a read/write to block based storage.
2211 if (state
->need_plug
&& def
->plug
) {
2212 state
->plug_started
= true;
2213 state
->need_plug
= false;
2214 blk_start_plug_nr_ios(&state
->plug
, state
->submit_nr
);
2218 personality
= READ_ONCE(sqe
->personality
);
2222 req
->creds
= xa_load(&ctx
->personalities
, personality
);
2225 get_cred(req
->creds
);
2226 ret
= security_uring_override_creds(req
->creds
);
2228 put_cred(req
->creds
);
2231 req
->flags
|= REQ_F_CREDS
;
2234 return def
->prep(req
, sqe
);
2237 static __cold
int io_submit_fail_init(const struct io_uring_sqe
*sqe
,
2238 struct io_kiocb
*req
, int ret
)
2240 struct io_ring_ctx
*ctx
= req
->ctx
;
2241 struct io_submit_link
*link
= &ctx
->submit_state
.link
;
2242 struct io_kiocb
*head
= link
->head
;
2244 trace_io_uring_req_failed(sqe
, req
, ret
);
2247 * Avoid breaking links in the middle as it renders links with SQPOLL
2248 * unusable. Instead of failing eagerly, continue assembling the link if
2249 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2250 * should find the flag and handle the rest.
2252 req_fail_link_node(req
, ret
);
2253 if (head
&& !(head
->flags
& REQ_F_FAIL
))
2254 req_fail_link_node(head
, -ECANCELED
);
2256 if (!(req
->flags
& IO_REQ_LINK_FLAGS
)) {
2258 link
->last
->link
= req
;
2262 io_queue_sqe_fallback(req
);
2267 link
->last
->link
= req
;
2274 static inline int io_submit_sqe(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
2275 const struct io_uring_sqe
*sqe
)
2276 __must_hold(&ctx
->uring_lock
)
2278 struct io_submit_link
*link
= &ctx
->submit_state
.link
;
2281 ret
= io_init_req(ctx
, req
, sqe
);
2283 return io_submit_fail_init(sqe
, req
, ret
);
2285 trace_io_uring_submit_req(req
);
2288 * If we already have a head request, queue this one for async
2289 * submittal once the head completes. If we don't have a head but
2290 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2291 * submitted sync once the chain is complete. If none of those
2292 * conditions are true (normal request), then just queue it.
2294 if (unlikely(link
->head
)) {
2295 ret
= io_req_prep_async(req
);
2297 return io_submit_fail_init(sqe
, req
, ret
);
2299 trace_io_uring_link(req
, link
->head
);
2300 link
->last
->link
= req
;
2303 if (req
->flags
& IO_REQ_LINK_FLAGS
)
2305 /* last request of the link, flush it */
2308 if (req
->flags
& (REQ_F_FORCE_ASYNC
| REQ_F_FAIL
))
2311 } else if (unlikely(req
->flags
& (IO_REQ_LINK_FLAGS
|
2312 REQ_F_FORCE_ASYNC
| REQ_F_FAIL
))) {
2313 if (req
->flags
& IO_REQ_LINK_FLAGS
) {
2318 io_queue_sqe_fallback(req
);
2328 * Batched submission is done, ensure local IO is flushed out.
2330 static void io_submit_state_end(struct io_ring_ctx
*ctx
)
2332 struct io_submit_state
*state
= &ctx
->submit_state
;
2334 if (unlikely(state
->link
.head
))
2335 io_queue_sqe_fallback(state
->link
.head
);
2336 /* flush only after queuing links as they can generate completions */
2337 io_submit_flush_completions(ctx
);
2338 if (state
->plug_started
)
2339 blk_finish_plug(&state
->plug
);
2343 * Start submission side cache.
2345 static void io_submit_state_start(struct io_submit_state
*state
,
2346 unsigned int max_ios
)
2348 state
->plug_started
= false;
2349 state
->need_plug
= max_ios
> 2;
2350 state
->submit_nr
= max_ios
;
2351 /* set only head, no need to init link_last in advance */
2352 state
->link
.head
= NULL
;
2355 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
2357 struct io_rings
*rings
= ctx
->rings
;
2360 * Ensure any loads from the SQEs are done at this point,
2361 * since once we write the new head, the application could
2362 * write new data to them.
2364 smp_store_release(&rings
->sq
.head
, ctx
->cached_sq_head
);
2368 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2369 * that is mapped by userspace. This means that care needs to be taken to
2370 * ensure that reads are stable, as we cannot rely on userspace always
2371 * being a good citizen. If members of the sqe are validated and then later
2372 * used, it's important that those reads are done through READ_ONCE() to
2373 * prevent a re-load down the line.
2375 static bool io_get_sqe(struct io_ring_ctx
*ctx
, const struct io_uring_sqe
**sqe
)
2377 unsigned mask
= ctx
->sq_entries
- 1;
2378 unsigned head
= ctx
->cached_sq_head
++ & mask
;
2380 if (!(ctx
->flags
& IORING_SETUP_NO_SQARRAY
)) {
2381 head
= READ_ONCE(ctx
->sq_array
[head
]);
2382 if (unlikely(head
>= ctx
->sq_entries
)) {
2383 /* drop invalid entries */
2384 spin_lock(&ctx
->completion_lock
);
2386 spin_unlock(&ctx
->completion_lock
);
2387 WRITE_ONCE(ctx
->rings
->sq_dropped
,
2388 READ_ONCE(ctx
->rings
->sq_dropped
) + 1);
2394 * The cached sq head (or cq tail) serves two purposes:
2396 * 1) allows us to batch the cost of updating the user visible
2398 * 2) allows the kernel side to track the head on its own, even
2399 * though the application is the one updating it.
2402 /* double index for 128-byte SQEs, twice as long */
2403 if (ctx
->flags
& IORING_SETUP_SQE128
)
2405 *sqe
= &ctx
->sq_sqes
[head
];
2409 int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
)
2410 __must_hold(&ctx
->uring_lock
)
2412 unsigned int entries
= io_sqring_entries(ctx
);
2416 if (unlikely(!entries
))
2418 /* make sure SQ entry isn't read before tail */
2419 ret
= left
= min(nr
, entries
);
2420 io_get_task_refs(left
);
2421 io_submit_state_start(&ctx
->submit_state
, left
);
2424 const struct io_uring_sqe
*sqe
;
2425 struct io_kiocb
*req
;
2427 if (unlikely(!io_alloc_req(ctx
, &req
)))
2429 if (unlikely(!io_get_sqe(ctx
, &sqe
))) {
2430 io_req_add_to_cache(req
, ctx
);
2435 * Continue submitting even for sqe failure if the
2436 * ring was setup with IORING_SETUP_SUBMIT_ALL
2438 if (unlikely(io_submit_sqe(ctx
, req
, sqe
)) &&
2439 !(ctx
->flags
& IORING_SETUP_SUBMIT_ALL
)) {
2445 if (unlikely(left
)) {
2447 /* try again if it submitted nothing and can't allocate a req */
2448 if (!ret
&& io_req_cache_empty(ctx
))
2450 current
->io_uring
->cached_refs
+= left
;
2453 io_submit_state_end(ctx
);
2454 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2455 io_commit_sqring(ctx
);
2459 struct io_wait_queue
{
2460 struct wait_queue_entry wq
;
2461 struct io_ring_ctx
*ctx
;
2463 unsigned nr_timeouts
;
2467 static inline bool io_has_work(struct io_ring_ctx
*ctx
)
2469 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
) ||
2470 !llist_empty(&ctx
->work_llist
);
2473 static inline bool io_should_wake(struct io_wait_queue
*iowq
)
2475 struct io_ring_ctx
*ctx
= iowq
->ctx
;
2476 int dist
= READ_ONCE(ctx
->rings
->cq
.tail
) - (int) iowq
->cq_tail
;
2479 * Wake up if we have enough events, or if a timeout occurred since we
2480 * started waiting. For timeouts, we always want to return to userspace,
2481 * regardless of event count.
2483 return dist
>= 0 || atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
2486 static int io_wake_function(struct wait_queue_entry
*curr
, unsigned int mode
,
2487 int wake_flags
, void *key
)
2489 struct io_wait_queue
*iowq
= container_of(curr
, struct io_wait_queue
, wq
);
2492 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2493 * the task, and the next invocation will do it.
2495 if (io_should_wake(iowq
) || io_has_work(iowq
->ctx
))
2496 return autoremove_wake_function(curr
, mode
, wake_flags
, key
);
2500 int io_run_task_work_sig(struct io_ring_ctx
*ctx
)
2502 if (!llist_empty(&ctx
->work_llist
)) {
2503 __set_current_state(TASK_RUNNING
);
2504 if (io_run_local_work(ctx
) > 0)
2507 if (io_run_task_work() > 0)
2509 if (task_sigpending(current
))
2514 static bool current_pending_io(void)
2516 struct io_uring_task
*tctx
= current
->io_uring
;
2520 return percpu_counter_read_positive(&tctx
->inflight
);
2523 /* when returns >0, the caller should retry */
2524 static inline int io_cqring_wait_schedule(struct io_ring_ctx
*ctx
,
2525 struct io_wait_queue
*iowq
)
2529 if (unlikely(READ_ONCE(ctx
->check_cq
)))
2531 if (unlikely(!llist_empty(&ctx
->work_llist
)))
2533 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL
)))
2535 if (unlikely(task_sigpending(current
)))
2537 if (unlikely(io_should_wake(iowq
)))
2541 * Mark us as being in io_wait if we have pending requests, so cpufreq
2542 * can take into account that the task is waiting for IO - turns out
2543 * to be important for low QD IO.
2545 io_wait
= current
->in_iowait
;
2546 if (current_pending_io())
2547 current
->in_iowait
= 1;
2549 if (iowq
->timeout
== KTIME_MAX
)
2551 else if (!schedule_hrtimeout(&iowq
->timeout
, HRTIMER_MODE_ABS
))
2553 current
->in_iowait
= io_wait
;
2558 * Wait until events become available, if we don't already have some. The
2559 * application must reap them itself, as they reside on the shared cq ring.
2561 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
2562 const sigset_t __user
*sig
, size_t sigsz
,
2563 struct __kernel_timespec __user
*uts
)
2565 struct io_wait_queue iowq
;
2566 struct io_rings
*rings
= ctx
->rings
;
2569 if (!io_allowed_run_tw(ctx
))
2571 if (!llist_empty(&ctx
->work_llist
))
2572 io_run_local_work(ctx
);
2574 io_cqring_overflow_flush(ctx
);
2575 /* if user messes with these they will just get an early return */
2576 if (__io_cqring_events_user(ctx
) >= min_events
)
2580 #ifdef CONFIG_COMPAT
2581 if (in_compat_syscall())
2582 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
2586 ret
= set_user_sigmask(sig
, sigsz
);
2592 init_waitqueue_func_entry(&iowq
.wq
, io_wake_function
);
2593 iowq
.wq
.private = current
;
2594 INIT_LIST_HEAD(&iowq
.wq
.entry
);
2596 iowq
.nr_timeouts
= atomic_read(&ctx
->cq_timeouts
);
2597 iowq
.cq_tail
= READ_ONCE(ctx
->rings
->cq
.head
) + min_events
;
2598 iowq
.timeout
= KTIME_MAX
;
2601 struct timespec64 ts
;
2603 if (get_timespec64(&ts
, uts
))
2605 iowq
.timeout
= ktime_add_ns(timespec64_to_ktime(ts
), ktime_get_ns());
2608 trace_io_uring_cqring_wait(ctx
, min_events
);
2610 unsigned long check_cq
;
2612 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) {
2613 int nr_wait
= (int) iowq
.cq_tail
- READ_ONCE(ctx
->rings
->cq
.tail
);
2615 atomic_set(&ctx
->cq_wait_nr
, nr_wait
);
2616 set_current_state(TASK_INTERRUPTIBLE
);
2618 prepare_to_wait_exclusive(&ctx
->cq_wait
, &iowq
.wq
,
2619 TASK_INTERRUPTIBLE
);
2622 ret
= io_cqring_wait_schedule(ctx
, &iowq
);
2623 __set_current_state(TASK_RUNNING
);
2624 atomic_set(&ctx
->cq_wait_nr
, 0);
2629 * Run task_work after scheduling and before io_should_wake().
2630 * If we got woken because of task_work being processed, run it
2631 * now rather than let the caller do another wait loop.
2634 if (!llist_empty(&ctx
->work_llist
))
2635 io_run_local_work(ctx
);
2637 check_cq
= READ_ONCE(ctx
->check_cq
);
2638 if (unlikely(check_cq
)) {
2639 /* let the caller flush overflows, retry */
2640 if (check_cq
& BIT(IO_CHECK_CQ_OVERFLOW_BIT
))
2641 io_cqring_do_overflow_flush(ctx
);
2642 if (check_cq
& BIT(IO_CHECK_CQ_DROPPED_BIT
)) {
2648 if (io_should_wake(&iowq
)) {
2655 if (!(ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
))
2656 finish_wait(&ctx
->cq_wait
, &iowq
.wq
);
2657 restore_saved_sigmask_unless(ret
== -EINTR
);
2659 return READ_ONCE(rings
->cq
.head
) == READ_ONCE(rings
->cq
.tail
) ? ret
: 0;
2662 static void io_mem_free(void *ptr
)
2667 folio_put(virt_to_folio(ptr
));
2670 static void io_pages_free(struct page
***pages
, int npages
)
2672 struct page
**page_array
;
2677 page_array
= *pages
;
2678 for (i
= 0; i
< npages
; i
++)
2679 unpin_user_page(page_array
[i
]);
2684 static void *__io_uaddr_map(struct page
***pages
, unsigned short *npages
,
2685 unsigned long uaddr
, size_t size
)
2687 struct page
**page_array
;
2688 unsigned int nr_pages
;
2693 if (uaddr
& (PAGE_SIZE
- 1) || !size
)
2694 return ERR_PTR(-EINVAL
);
2696 nr_pages
= (size
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2697 if (nr_pages
> USHRT_MAX
)
2698 return ERR_PTR(-EINVAL
);
2699 page_array
= kvmalloc_array(nr_pages
, sizeof(struct page
*), GFP_KERNEL
);
2701 return ERR_PTR(-ENOMEM
);
2703 ret
= pin_user_pages_fast(uaddr
, nr_pages
, FOLL_WRITE
| FOLL_LONGTERM
,
2705 if (ret
!= nr_pages
) {
2707 io_pages_free(&page_array
, ret
> 0 ? ret
: 0);
2708 return ret
< 0 ? ERR_PTR(ret
) : ERR_PTR(-EFAULT
);
2711 * Should be a single page. If the ring is small enough that we can
2712 * use a normal page, that is fine. If we need multiple pages, then
2713 * userspace should use a huge page. That's the only way to guarantee
2714 * that we get contigious memory, outside of just being lucky or
2715 * (currently) having low memory fragmentation.
2717 if (page_array
[0] != page_array
[ret
- 1])
2719 *pages
= page_array
;
2721 return page_to_virt(page_array
[0]);
2724 static void *io_rings_map(struct io_ring_ctx
*ctx
, unsigned long uaddr
,
2727 return __io_uaddr_map(&ctx
->ring_pages
, &ctx
->n_ring_pages
, uaddr
,
2731 static void *io_sqes_map(struct io_ring_ctx
*ctx
, unsigned long uaddr
,
2734 return __io_uaddr_map(&ctx
->sqe_pages
, &ctx
->n_sqe_pages
, uaddr
,
2738 static void io_rings_free(struct io_ring_ctx
*ctx
)
2740 if (!(ctx
->flags
& IORING_SETUP_NO_MMAP
)) {
2741 io_mem_free(ctx
->rings
);
2742 io_mem_free(ctx
->sq_sqes
);
2744 ctx
->sq_sqes
= NULL
;
2746 io_pages_free(&ctx
->ring_pages
, ctx
->n_ring_pages
);
2747 io_pages_free(&ctx
->sqe_pages
, ctx
->n_sqe_pages
);
2751 static void *io_mem_alloc(size_t size
)
2753 gfp_t gfp
= GFP_KERNEL_ACCOUNT
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
;
2756 ret
= (void *) __get_free_pages(gfp
, get_order(size
));
2759 return ERR_PTR(-ENOMEM
);
2762 static unsigned long rings_size(struct io_ring_ctx
*ctx
, unsigned int sq_entries
,
2763 unsigned int cq_entries
, size_t *sq_offset
)
2765 struct io_rings
*rings
;
2766 size_t off
, sq_array_size
;
2768 off
= struct_size(rings
, cqes
, cq_entries
);
2769 if (off
== SIZE_MAX
)
2771 if (ctx
->flags
& IORING_SETUP_CQE32
) {
2772 if (check_shl_overflow(off
, 1, &off
))
2777 off
= ALIGN(off
, SMP_CACHE_BYTES
);
2782 if (ctx
->flags
& IORING_SETUP_NO_SQARRAY
) {
2784 *sq_offset
= SIZE_MAX
;
2791 sq_array_size
= array_size(sizeof(u32
), sq_entries
);
2792 if (sq_array_size
== SIZE_MAX
)
2795 if (check_add_overflow(off
, sq_array_size
, &off
))
2801 static int io_eventfd_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
2802 unsigned int eventfd_async
)
2804 struct io_ev_fd
*ev_fd
;
2805 __s32 __user
*fds
= arg
;
2808 ev_fd
= rcu_dereference_protected(ctx
->io_ev_fd
,
2809 lockdep_is_held(&ctx
->uring_lock
));
2813 if (copy_from_user(&fd
, fds
, sizeof(*fds
)))
2816 ev_fd
= kmalloc(sizeof(*ev_fd
), GFP_KERNEL
);
2820 ev_fd
->cq_ev_fd
= eventfd_ctx_fdget(fd
);
2821 if (IS_ERR(ev_fd
->cq_ev_fd
)) {
2822 int ret
= PTR_ERR(ev_fd
->cq_ev_fd
);
2827 spin_lock(&ctx
->completion_lock
);
2828 ctx
->evfd_last_cq_tail
= ctx
->cached_cq_tail
;
2829 spin_unlock(&ctx
->completion_lock
);
2831 ev_fd
->eventfd_async
= eventfd_async
;
2832 ctx
->has_evfd
= true;
2833 rcu_assign_pointer(ctx
->io_ev_fd
, ev_fd
);
2834 atomic_set(&ev_fd
->refs
, 1);
2835 atomic_set(&ev_fd
->ops
, 0);
2839 static int io_eventfd_unregister(struct io_ring_ctx
*ctx
)
2841 struct io_ev_fd
*ev_fd
;
2843 ev_fd
= rcu_dereference_protected(ctx
->io_ev_fd
,
2844 lockdep_is_held(&ctx
->uring_lock
));
2846 ctx
->has_evfd
= false;
2847 rcu_assign_pointer(ctx
->io_ev_fd
, NULL
);
2848 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT
), &ev_fd
->ops
))
2849 call_rcu(&ev_fd
->rcu
, io_eventfd_ops
);
2856 static void io_req_caches_free(struct io_ring_ctx
*ctx
)
2858 struct io_kiocb
*req
;
2861 mutex_lock(&ctx
->uring_lock
);
2862 io_flush_cached_locked_reqs(ctx
, &ctx
->submit_state
);
2864 while (!io_req_cache_empty(ctx
)) {
2865 req
= io_extract_req(ctx
);
2866 kmem_cache_free(req_cachep
, req
);
2870 percpu_ref_put_many(&ctx
->refs
, nr
);
2871 mutex_unlock(&ctx
->uring_lock
);
2874 static void io_rsrc_node_cache_free(struct io_cache_entry
*entry
)
2876 kfree(container_of(entry
, struct io_rsrc_node
, cache
));
2879 static __cold
void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
2881 io_sq_thread_finish(ctx
);
2882 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2883 if (WARN_ON_ONCE(!list_empty(&ctx
->rsrc_ref_list
)))
2886 mutex_lock(&ctx
->uring_lock
);
2888 __io_sqe_buffers_unregister(ctx
);
2890 __io_sqe_files_unregister(ctx
);
2891 io_cqring_overflow_kill(ctx
);
2892 io_eventfd_unregister(ctx
);
2893 io_alloc_cache_free(&ctx
->apoll_cache
, io_apoll_cache_free
);
2894 io_alloc_cache_free(&ctx
->netmsg_cache
, io_netmsg_cache_free
);
2895 io_destroy_buffers(ctx
);
2896 mutex_unlock(&ctx
->uring_lock
);
2898 put_cred(ctx
->sq_creds
);
2899 if (ctx
->submitter_task
)
2900 put_task_struct(ctx
->submitter_task
);
2902 /* there are no registered resources left, nobody uses it */
2904 io_rsrc_node_destroy(ctx
, ctx
->rsrc_node
);
2906 WARN_ON_ONCE(!list_empty(&ctx
->rsrc_ref_list
));
2908 #if defined(CONFIG_UNIX)
2909 if (ctx
->ring_sock
) {
2910 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
2911 sock_release(ctx
->ring_sock
);
2914 WARN_ON_ONCE(!list_empty(&ctx
->ltimeout_list
));
2916 io_alloc_cache_free(&ctx
->rsrc_node_cache
, io_rsrc_node_cache_free
);
2917 if (ctx
->mm_account
) {
2918 mmdrop(ctx
->mm_account
);
2919 ctx
->mm_account
= NULL
;
2923 percpu_ref_exit(&ctx
->refs
);
2924 free_uid(ctx
->user
);
2925 io_req_caches_free(ctx
);
2927 io_wq_put_hash(ctx
->hash_map
);
2928 kfree(ctx
->cancel_table
.hbs
);
2929 kfree(ctx
->cancel_table_locked
.hbs
);
2931 xa_destroy(&ctx
->io_bl_xa
);
2935 static __cold
void io_activate_pollwq_cb(struct callback_head
*cb
)
2937 struct io_ring_ctx
*ctx
= container_of(cb
, struct io_ring_ctx
,
2940 mutex_lock(&ctx
->uring_lock
);
2941 ctx
->poll_activated
= true;
2942 mutex_unlock(&ctx
->uring_lock
);
2945 * Wake ups for some events between start of polling and activation
2946 * might've been lost due to loose synchronisation.
2948 wake_up_all(&ctx
->poll_wq
);
2949 percpu_ref_put(&ctx
->refs
);
2952 static __cold
void io_activate_pollwq(struct io_ring_ctx
*ctx
)
2954 spin_lock(&ctx
->completion_lock
);
2955 /* already activated or in progress */
2956 if (ctx
->poll_activated
|| ctx
->poll_wq_task_work
.func
)
2958 if (WARN_ON_ONCE(!ctx
->task_complete
))
2960 if (!ctx
->submitter_task
)
2963 * with ->submitter_task only the submitter task completes requests, we
2964 * only need to sync with it, which is done by injecting a tw
2966 init_task_work(&ctx
->poll_wq_task_work
, io_activate_pollwq_cb
);
2967 percpu_ref_get(&ctx
->refs
);
2968 if (task_work_add(ctx
->submitter_task
, &ctx
->poll_wq_task_work
, TWA_SIGNAL
))
2969 percpu_ref_put(&ctx
->refs
);
2971 spin_unlock(&ctx
->completion_lock
);
2974 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
2976 struct io_ring_ctx
*ctx
= file
->private_data
;
2979 if (unlikely(!ctx
->poll_activated
))
2980 io_activate_pollwq(ctx
);
2982 poll_wait(file
, &ctx
->poll_wq
, wait
);
2984 * synchronizes with barrier from wq_has_sleeper call in
2988 if (!io_sqring_full(ctx
))
2989 mask
|= EPOLLOUT
| EPOLLWRNORM
;
2992 * Don't flush cqring overflow list here, just do a simple check.
2993 * Otherwise there could possible be ABBA deadlock:
2996 * lock(&ctx->uring_lock);
2998 * lock(&ctx->uring_lock);
3001 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
3002 * pushes them to do the flush.
3005 if (__io_cqring_events_user(ctx
) || io_has_work(ctx
))
3006 mask
|= EPOLLIN
| EPOLLRDNORM
;
3011 static int io_unregister_personality(struct io_ring_ctx
*ctx
, unsigned id
)
3013 const struct cred
*creds
;
3015 creds
= xa_erase(&ctx
->personalities
, id
);
3024 struct io_tctx_exit
{
3025 struct callback_head task_work
;
3026 struct completion completion
;
3027 struct io_ring_ctx
*ctx
;
3030 static __cold
void io_tctx_exit_cb(struct callback_head
*cb
)
3032 struct io_uring_task
*tctx
= current
->io_uring
;
3033 struct io_tctx_exit
*work
;
3035 work
= container_of(cb
, struct io_tctx_exit
, task_work
);
3037 * When @in_cancel, we're in cancellation and it's racy to remove the
3038 * node. It'll be removed by the end of cancellation, just ignore it.
3039 * tctx can be NULL if the queueing of this task_work raced with
3040 * work cancelation off the exec path.
3042 if (tctx
&& !atomic_read(&tctx
->in_cancel
))
3043 io_uring_del_tctx_node((unsigned long)work
->ctx
);
3044 complete(&work
->completion
);
3047 static __cold
bool io_cancel_ctx_cb(struct io_wq_work
*work
, void *data
)
3049 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
3051 return req
->ctx
== data
;
3054 static __cold
void io_ring_exit_work(struct work_struct
*work
)
3056 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
, exit_work
);
3057 unsigned long timeout
= jiffies
+ HZ
* 60 * 5;
3058 unsigned long interval
= HZ
/ 20;
3059 struct io_tctx_exit exit
;
3060 struct io_tctx_node
*node
;
3064 * If we're doing polled IO and end up having requests being
3065 * submitted async (out-of-line), then completions can come in while
3066 * we're waiting for refs to drop. We need to reap these manually,
3067 * as nobody else will be looking for them.
3070 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
)) {
3071 mutex_lock(&ctx
->uring_lock
);
3072 io_cqring_overflow_kill(ctx
);
3073 mutex_unlock(&ctx
->uring_lock
);
3076 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
)
3077 io_move_task_work_from_local(ctx
);
3079 while (io_uring_try_cancel_requests(ctx
, NULL
, true))
3083 struct io_sq_data
*sqd
= ctx
->sq_data
;
3084 struct task_struct
*tsk
;
3086 io_sq_thread_park(sqd
);
3088 if (tsk
&& tsk
->io_uring
&& tsk
->io_uring
->io_wq
)
3089 io_wq_cancel_cb(tsk
->io_uring
->io_wq
,
3090 io_cancel_ctx_cb
, ctx
, true);
3091 io_sq_thread_unpark(sqd
);
3094 io_req_caches_free(ctx
);
3096 if (WARN_ON_ONCE(time_after(jiffies
, timeout
))) {
3097 /* there is little hope left, don't run it too often */
3101 * This is really an uninterruptible wait, as it has to be
3102 * complete. But it's also run from a kworker, which doesn't
3103 * take signals, so it's fine to make it interruptible. This
3104 * avoids scenarios where we knowingly can wait much longer
3105 * on completions, for example if someone does a SIGSTOP on
3106 * a task that needs to finish task_work to make this loop
3107 * complete. That's a synthetic situation that should not
3108 * cause a stuck task backtrace, and hence a potential panic
3109 * on stuck tasks if that is enabled.
3111 } while (!wait_for_completion_interruptible_timeout(&ctx
->ref_comp
, interval
));
3113 init_completion(&exit
.completion
);
3114 init_task_work(&exit
.task_work
, io_tctx_exit_cb
);
3117 * Some may use context even when all refs and requests have been put,
3118 * and they are free to do so while still holding uring_lock or
3119 * completion_lock, see io_req_task_submit(). Apart from other work,
3120 * this lock/unlock section also waits them to finish.
3122 mutex_lock(&ctx
->uring_lock
);
3123 while (!list_empty(&ctx
->tctx_list
)) {
3124 WARN_ON_ONCE(time_after(jiffies
, timeout
));
3126 node
= list_first_entry(&ctx
->tctx_list
, struct io_tctx_node
,
3128 /* don't spin on a single task if cancellation failed */
3129 list_rotate_left(&ctx
->tctx_list
);
3130 ret
= task_work_add(node
->task
, &exit
.task_work
, TWA_SIGNAL
);
3131 if (WARN_ON_ONCE(ret
))
3134 mutex_unlock(&ctx
->uring_lock
);
3136 * See comment above for
3137 * wait_for_completion_interruptible_timeout() on why this
3138 * wait is marked as interruptible.
3140 wait_for_completion_interruptible(&exit
.completion
);
3141 mutex_lock(&ctx
->uring_lock
);
3143 mutex_unlock(&ctx
->uring_lock
);
3144 spin_lock(&ctx
->completion_lock
);
3145 spin_unlock(&ctx
->completion_lock
);
3147 /* pairs with RCU read section in io_req_local_work_add() */
3148 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
)
3151 io_ring_ctx_free(ctx
);
3154 static __cold
void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
3156 unsigned long index
;
3157 struct creds
*creds
;
3159 mutex_lock(&ctx
->uring_lock
);
3160 percpu_ref_kill(&ctx
->refs
);
3161 xa_for_each(&ctx
->personalities
, index
, creds
)
3162 io_unregister_personality(ctx
, index
);
3164 io_poll_remove_all(ctx
, NULL
, true);
3165 mutex_unlock(&ctx
->uring_lock
);
3168 * If we failed setting up the ctx, we might not have any rings
3169 * and therefore did not submit any requests
3172 io_kill_timeouts(ctx
, NULL
, true);
3174 flush_delayed_work(&ctx
->fallback_work
);
3176 INIT_WORK(&ctx
->exit_work
, io_ring_exit_work
);
3178 * Use system_unbound_wq to avoid spawning tons of event kworkers
3179 * if we're exiting a ton of rings at the same time. It just adds
3180 * noise and overhead, there's no discernable change in runtime
3181 * over using system_wq.
3183 queue_work(system_unbound_wq
, &ctx
->exit_work
);
3186 static int io_uring_release(struct inode
*inode
, struct file
*file
)
3188 struct io_ring_ctx
*ctx
= file
->private_data
;
3190 file
->private_data
= NULL
;
3191 io_ring_ctx_wait_and_kill(ctx
);
3195 struct io_task_cancel
{
3196 struct task_struct
*task
;
3200 static bool io_cancel_task_cb(struct io_wq_work
*work
, void *data
)
3202 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
3203 struct io_task_cancel
*cancel
= data
;
3205 return io_match_task_safe(req
, cancel
->task
, cancel
->all
);
3208 static __cold
bool io_cancel_defer_files(struct io_ring_ctx
*ctx
,
3209 struct task_struct
*task
,
3212 struct io_defer_entry
*de
;
3215 spin_lock(&ctx
->completion_lock
);
3216 list_for_each_entry_reverse(de
, &ctx
->defer_list
, list
) {
3217 if (io_match_task_safe(de
->req
, task
, cancel_all
)) {
3218 list_cut_position(&list
, &ctx
->defer_list
, &de
->list
);
3222 spin_unlock(&ctx
->completion_lock
);
3223 if (list_empty(&list
))
3226 while (!list_empty(&list
)) {
3227 de
= list_first_entry(&list
, struct io_defer_entry
, list
);
3228 list_del_init(&de
->list
);
3229 io_req_task_queue_fail(de
->req
, -ECANCELED
);
3235 static __cold
bool io_uring_try_cancel_iowq(struct io_ring_ctx
*ctx
)
3237 struct io_tctx_node
*node
;
3238 enum io_wq_cancel cret
;
3241 mutex_lock(&ctx
->uring_lock
);
3242 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
3243 struct io_uring_task
*tctx
= node
->task
->io_uring
;
3246 * io_wq will stay alive while we hold uring_lock, because it's
3247 * killed after ctx nodes, which requires to take the lock.
3249 if (!tctx
|| !tctx
->io_wq
)
3251 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_ctx_cb
, ctx
, true);
3252 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
3254 mutex_unlock(&ctx
->uring_lock
);
3259 static __cold
bool io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
3260 struct task_struct
*task
,
3263 struct io_task_cancel cancel
= { .task
= task
, .all
= cancel_all
, };
3264 struct io_uring_task
*tctx
= task
? task
->io_uring
: NULL
;
3265 enum io_wq_cancel cret
;
3268 /* set it so io_req_local_work_add() would wake us up */
3269 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) {
3270 atomic_set(&ctx
->cq_wait_nr
, 1);
3274 /* failed during ring init, it couldn't have issued any requests */
3279 ret
|= io_uring_try_cancel_iowq(ctx
);
3280 } else if (tctx
&& tctx
->io_wq
) {
3282 * Cancels requests of all rings, not only @ctx, but
3283 * it's fine as the task is in exit/exec.
3285 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_task_cb
,
3287 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
3290 /* SQPOLL thread does its own polling */
3291 if ((!(ctx
->flags
& IORING_SETUP_SQPOLL
) && cancel_all
) ||
3292 (ctx
->sq_data
&& ctx
->sq_data
->thread
== current
)) {
3293 while (!wq_list_empty(&ctx
->iopoll_list
)) {
3294 io_iopoll_try_reap_events(ctx
);
3300 if ((ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) &&
3301 io_allowed_defer_tw_run(ctx
))
3302 ret
|= io_run_local_work(ctx
) > 0;
3303 ret
|= io_cancel_defer_files(ctx
, task
, cancel_all
);
3304 mutex_lock(&ctx
->uring_lock
);
3305 ret
|= io_poll_remove_all(ctx
, task
, cancel_all
);
3306 mutex_unlock(&ctx
->uring_lock
);
3307 ret
|= io_kill_timeouts(ctx
, task
, cancel_all
);
3309 ret
|= io_run_task_work() > 0;
3313 static s64
tctx_inflight(struct io_uring_task
*tctx
, bool tracked
)
3316 return atomic_read(&tctx
->inflight_tracked
);
3317 return percpu_counter_sum(&tctx
->inflight
);
3321 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3322 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3324 __cold
void io_uring_cancel_generic(bool cancel_all
, struct io_sq_data
*sqd
)
3326 struct io_uring_task
*tctx
= current
->io_uring
;
3327 struct io_ring_ctx
*ctx
;
3328 struct io_tctx_node
*node
;
3329 unsigned long index
;
3333 WARN_ON_ONCE(sqd
&& sqd
->thread
!= current
);
3335 if (!current
->io_uring
)
3338 io_wq_exit_start(tctx
->io_wq
);
3340 atomic_inc(&tctx
->in_cancel
);
3344 io_uring_drop_tctx_refs(current
);
3345 /* read completions before cancelations */
3346 inflight
= tctx_inflight(tctx
, !cancel_all
);
3351 xa_for_each(&tctx
->xa
, index
, node
) {
3352 /* sqpoll task will cancel all its requests */
3353 if (node
->ctx
->sq_data
)
3355 loop
|= io_uring_try_cancel_requests(node
->ctx
,
3356 current
, cancel_all
);
3359 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
3360 loop
|= io_uring_try_cancel_requests(ctx
,
3370 prepare_to_wait(&tctx
->wait
, &wait
, TASK_INTERRUPTIBLE
);
3372 io_uring_drop_tctx_refs(current
);
3373 xa_for_each(&tctx
->xa
, index
, node
) {
3374 if (!llist_empty(&node
->ctx
->work_llist
)) {
3375 WARN_ON_ONCE(node
->ctx
->submitter_task
&&
3376 node
->ctx
->submitter_task
!= current
);
3381 * If we've seen completions, retry without waiting. This
3382 * avoids a race where a completion comes in before we did
3383 * prepare_to_wait().
3385 if (inflight
== tctx_inflight(tctx
, !cancel_all
))
3388 finish_wait(&tctx
->wait
, &wait
);
3391 io_uring_clean_tctx(tctx
);
3394 * We shouldn't run task_works after cancel, so just leave
3395 * ->in_cancel set for normal exit.
3397 atomic_dec(&tctx
->in_cancel
);
3398 /* for exec all current's requests should be gone, kill tctx */
3399 __io_uring_free(current
);
3403 void __io_uring_cancel(bool cancel_all
)
3405 io_uring_cancel_generic(cancel_all
, NULL
);
3408 static void *io_uring_validate_mmap_request(struct file
*file
,
3409 loff_t pgoff
, size_t sz
)
3411 struct io_ring_ctx
*ctx
= file
->private_data
;
3412 loff_t offset
= pgoff
<< PAGE_SHIFT
;
3416 /* Don't allow mmap if the ring was setup without it */
3417 if (ctx
->flags
& IORING_SETUP_NO_MMAP
)
3418 return ERR_PTR(-EINVAL
);
3420 switch (offset
& IORING_OFF_MMAP_MASK
) {
3421 case IORING_OFF_SQ_RING
:
3422 case IORING_OFF_CQ_RING
:
3425 case IORING_OFF_SQES
:
3428 case IORING_OFF_PBUF_RING
: {
3431 bgid
= (offset
& ~IORING_OFF_MMAP_MASK
) >> IORING_OFF_PBUF_SHIFT
;
3432 mutex_lock(&ctx
->uring_lock
);
3433 ptr
= io_pbuf_get_address(ctx
, bgid
);
3434 mutex_unlock(&ctx
->uring_lock
);
3436 return ERR_PTR(-EINVAL
);
3440 return ERR_PTR(-EINVAL
);
3443 page
= virt_to_head_page(ptr
);
3444 if (sz
> page_size(page
))
3445 return ERR_PTR(-EINVAL
);
3452 static __cold
int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3454 size_t sz
= vma
->vm_end
- vma
->vm_start
;
3458 ptr
= io_uring_validate_mmap_request(file
, vma
->vm_pgoff
, sz
);
3460 return PTR_ERR(ptr
);
3462 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
3463 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
3466 static unsigned long io_uring_mmu_get_unmapped_area(struct file
*filp
,
3467 unsigned long addr
, unsigned long len
,
3468 unsigned long pgoff
, unsigned long flags
)
3473 * Do not allow to map to user-provided address to avoid breaking the
3474 * aliasing rules. Userspace is not able to guess the offset address of
3475 * kernel kmalloc()ed memory area.
3480 ptr
= io_uring_validate_mmap_request(filp
, pgoff
, len
);
3485 * Some architectures have strong cache aliasing requirements.
3486 * For such architectures we need a coherent mapping which aliases
3487 * kernel memory *and* userspace memory. To achieve that:
3488 * - use a NULL file pointer to reference physical memory, and
3489 * - use the kernel virtual address of the shared io_uring context
3490 * (instead of the userspace-provided address, which has to be 0UL
3492 * - use the same pgoff which the get_unmapped_area() uses to
3493 * calculate the page colouring.
3494 * For architectures without such aliasing requirements, the
3495 * architecture will return any suitable mapping because addr is 0.
3498 flags
|= MAP_SHARED
;
3499 pgoff
= 0; /* has been translated to ptr above */
3501 addr
= (uintptr_t) ptr
;
3502 pgoff
= addr
>> PAGE_SHIFT
;
3506 return current
->mm
->get_unmapped_area(filp
, addr
, len
, pgoff
, flags
);
3509 #else /* !CONFIG_MMU */
3511 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3513 return is_nommu_shared_mapping(vma
->vm_flags
) ? 0 : -EINVAL
;
3516 static unsigned int io_uring_nommu_mmap_capabilities(struct file
*file
)
3518 return NOMMU_MAP_DIRECT
| NOMMU_MAP_READ
| NOMMU_MAP_WRITE
;
3521 static unsigned long io_uring_nommu_get_unmapped_area(struct file
*file
,
3522 unsigned long addr
, unsigned long len
,
3523 unsigned long pgoff
, unsigned long flags
)
3527 ptr
= io_uring_validate_mmap_request(file
, pgoff
, len
);
3529 return PTR_ERR(ptr
);
3531 return (unsigned long) ptr
;
3534 #endif /* !CONFIG_MMU */
3536 static int io_validate_ext_arg(unsigned flags
, const void __user
*argp
, size_t argsz
)
3538 if (flags
& IORING_ENTER_EXT_ARG
) {
3539 struct io_uring_getevents_arg arg
;
3541 if (argsz
!= sizeof(arg
))
3543 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
3549 static int io_get_ext_arg(unsigned flags
, const void __user
*argp
, size_t *argsz
,
3550 struct __kernel_timespec __user
**ts
,
3551 const sigset_t __user
**sig
)
3553 struct io_uring_getevents_arg arg
;
3556 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3557 * is just a pointer to the sigset_t.
3559 if (!(flags
& IORING_ENTER_EXT_ARG
)) {
3560 *sig
= (const sigset_t __user
*) argp
;
3566 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3567 * timespec and sigset_t pointers if good.
3569 if (*argsz
!= sizeof(arg
))
3571 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
3575 *sig
= u64_to_user_ptr(arg
.sigmask
);
3576 *argsz
= arg
.sigmask_sz
;
3577 *ts
= u64_to_user_ptr(arg
.ts
);
3581 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
3582 u32
, min_complete
, u32
, flags
, const void __user
*, argp
,
3585 struct io_ring_ctx
*ctx
;
3589 if (unlikely(flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
|
3590 IORING_ENTER_SQ_WAIT
| IORING_ENTER_EXT_ARG
|
3591 IORING_ENTER_REGISTERED_RING
)))
3595 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3596 * need only dereference our task private array to find it.
3598 if (flags
& IORING_ENTER_REGISTERED_RING
) {
3599 struct io_uring_task
*tctx
= current
->io_uring
;
3601 if (unlikely(!tctx
|| fd
>= IO_RINGFD_REG_MAX
))
3603 fd
= array_index_nospec(fd
, IO_RINGFD_REG_MAX
);
3604 f
.file
= tctx
->registered_rings
[fd
];
3606 if (unlikely(!f
.file
))
3610 if (unlikely(!f
.file
))
3613 if (unlikely(!io_is_uring_fops(f
.file
)))
3617 ctx
= f
.file
->private_data
;
3619 if (unlikely(ctx
->flags
& IORING_SETUP_R_DISABLED
))
3623 * For SQ polling, the thread will do all submissions and completions.
3624 * Just return the requested submit count, and wake the thread if
3628 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
3629 io_cqring_overflow_flush(ctx
);
3631 if (unlikely(ctx
->sq_data
->thread
== NULL
)) {
3635 if (flags
& IORING_ENTER_SQ_WAKEUP
)
3636 wake_up(&ctx
->sq_data
->wait
);
3637 if (flags
& IORING_ENTER_SQ_WAIT
)
3638 io_sqpoll_wait_sq(ctx
);
3641 } else if (to_submit
) {
3642 ret
= io_uring_add_tctx_node(ctx
);
3646 mutex_lock(&ctx
->uring_lock
);
3647 ret
= io_submit_sqes(ctx
, to_submit
);
3648 if (ret
!= to_submit
) {
3649 mutex_unlock(&ctx
->uring_lock
);
3652 if (flags
& IORING_ENTER_GETEVENTS
) {
3653 if (ctx
->syscall_iopoll
)
3656 * Ignore errors, we'll soon call io_cqring_wait() and
3657 * it should handle ownership problems if any.
3659 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
)
3660 (void)io_run_local_work_locked(ctx
);
3662 mutex_unlock(&ctx
->uring_lock
);
3665 if (flags
& IORING_ENTER_GETEVENTS
) {
3668 if (ctx
->syscall_iopoll
) {
3670 * We disallow the app entering submit/complete with
3671 * polling, but we still need to lock the ring to
3672 * prevent racing with polled issue that got punted to
3675 mutex_lock(&ctx
->uring_lock
);
3677 ret2
= io_validate_ext_arg(flags
, argp
, argsz
);
3678 if (likely(!ret2
)) {
3679 min_complete
= min(min_complete
,
3681 ret2
= io_iopoll_check(ctx
, min_complete
);
3683 mutex_unlock(&ctx
->uring_lock
);
3685 const sigset_t __user
*sig
;
3686 struct __kernel_timespec __user
*ts
;
3688 ret2
= io_get_ext_arg(flags
, argp
, &argsz
, &ts
, &sig
);
3689 if (likely(!ret2
)) {
3690 min_complete
= min(min_complete
,
3692 ret2
= io_cqring_wait(ctx
, min_complete
, sig
,
3701 * EBADR indicates that one or more CQE were dropped.
3702 * Once the user has been informed we can clear the bit
3703 * as they are obviously ok with those drops.
3705 if (unlikely(ret2
== -EBADR
))
3706 clear_bit(IO_CHECK_CQ_DROPPED_BIT
,
3715 static const struct file_operations io_uring_fops
= {
3716 .release
= io_uring_release
,
3717 .mmap
= io_uring_mmap
,
3719 .get_unmapped_area
= io_uring_nommu_get_unmapped_area
,
3720 .mmap_capabilities
= io_uring_nommu_mmap_capabilities
,
3722 .get_unmapped_area
= io_uring_mmu_get_unmapped_area
,
3724 .poll
= io_uring_poll
,
3725 #ifdef CONFIG_PROC_FS
3726 .show_fdinfo
= io_uring_show_fdinfo
,
3730 bool io_is_uring_fops(struct file
*file
)
3732 return file
->f_op
== &io_uring_fops
;
3735 static __cold
int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
3736 struct io_uring_params
*p
)
3738 struct io_rings
*rings
;
3739 size_t size
, sq_array_offset
;
3742 /* make sure these are sane, as we already accounted them */
3743 ctx
->sq_entries
= p
->sq_entries
;
3744 ctx
->cq_entries
= p
->cq_entries
;
3746 size
= rings_size(ctx
, p
->sq_entries
, p
->cq_entries
, &sq_array_offset
);
3747 if (size
== SIZE_MAX
)
3750 if (!(ctx
->flags
& IORING_SETUP_NO_MMAP
))
3751 rings
= io_mem_alloc(size
);
3753 rings
= io_rings_map(ctx
, p
->cq_off
.user_addr
, size
);
3756 return PTR_ERR(rings
);
3759 if (!(ctx
->flags
& IORING_SETUP_NO_SQARRAY
))
3760 ctx
->sq_array
= (u32
*)((char *)rings
+ sq_array_offset
);
3761 rings
->sq_ring_mask
= p
->sq_entries
- 1;
3762 rings
->cq_ring_mask
= p
->cq_entries
- 1;
3763 rings
->sq_ring_entries
= p
->sq_entries
;
3764 rings
->cq_ring_entries
= p
->cq_entries
;
3766 if (p
->flags
& IORING_SETUP_SQE128
)
3767 size
= array_size(2 * sizeof(struct io_uring_sqe
), p
->sq_entries
);
3769 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
3770 if (size
== SIZE_MAX
) {
3775 if (!(ctx
->flags
& IORING_SETUP_NO_MMAP
))
3776 ptr
= io_mem_alloc(size
);
3778 ptr
= io_sqes_map(ctx
, p
->sq_off
.user_addr
, size
);
3782 return PTR_ERR(ptr
);
3789 static int io_uring_install_fd(struct file
*file
)
3793 fd
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
3796 fd_install(fd
, file
);
3801 * Allocate an anonymous fd, this is what constitutes the application
3802 * visible backing of an io_uring instance. The application mmaps this
3803 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3804 * we have to tie this fd to a socket for file garbage collection purposes.
3806 static struct file
*io_uring_get_file(struct io_ring_ctx
*ctx
)
3809 #if defined(CONFIG_UNIX)
3812 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
3815 return ERR_PTR(ret
);
3818 file
= anon_inode_getfile_secure("[io_uring]", &io_uring_fops
, ctx
,
3819 O_RDWR
| O_CLOEXEC
, NULL
);
3820 #if defined(CONFIG_UNIX)
3822 sock_release(ctx
->ring_sock
);
3823 ctx
->ring_sock
= NULL
;
3825 ctx
->ring_sock
->file
= file
;
3831 static __cold
int io_uring_create(unsigned entries
, struct io_uring_params
*p
,
3832 struct io_uring_params __user
*params
)
3834 struct io_ring_ctx
*ctx
;
3835 struct io_uring_task
*tctx
;
3841 if (entries
> IORING_MAX_ENTRIES
) {
3842 if (!(p
->flags
& IORING_SETUP_CLAMP
))
3844 entries
= IORING_MAX_ENTRIES
;
3847 if ((p
->flags
& IORING_SETUP_REGISTERED_FD_ONLY
)
3848 && !(p
->flags
& IORING_SETUP_NO_MMAP
))
3852 * Use twice as many entries for the CQ ring. It's possible for the
3853 * application to drive a higher depth than the size of the SQ ring,
3854 * since the sqes are only used at submission time. This allows for
3855 * some flexibility in overcommitting a bit. If the application has
3856 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3857 * of CQ ring entries manually.
3859 p
->sq_entries
= roundup_pow_of_two(entries
);
3860 if (p
->flags
& IORING_SETUP_CQSIZE
) {
3862 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3863 * to a power-of-two, if it isn't already. We do NOT impose
3864 * any cq vs sq ring sizing.
3868 if (p
->cq_entries
> IORING_MAX_CQ_ENTRIES
) {
3869 if (!(p
->flags
& IORING_SETUP_CLAMP
))
3871 p
->cq_entries
= IORING_MAX_CQ_ENTRIES
;
3873 p
->cq_entries
= roundup_pow_of_two(p
->cq_entries
);
3874 if (p
->cq_entries
< p
->sq_entries
)
3877 p
->cq_entries
= 2 * p
->sq_entries
;
3880 ctx
= io_ring_ctx_alloc(p
);
3884 if ((ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) &&
3885 !(ctx
->flags
& IORING_SETUP_IOPOLL
) &&
3886 !(ctx
->flags
& IORING_SETUP_SQPOLL
))
3887 ctx
->task_complete
= true;
3889 if (ctx
->task_complete
|| (ctx
->flags
& IORING_SETUP_IOPOLL
))
3890 ctx
->lockless_cq
= true;
3893 * lazy poll_wq activation relies on ->task_complete for synchronisation
3894 * purposes, see io_activate_pollwq()
3896 if (!ctx
->task_complete
)
3897 ctx
->poll_activated
= true;
3900 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3901 * space applications don't need to do io completion events
3902 * polling again, they can rely on io_sq_thread to do polling
3903 * work, which can reduce cpu usage and uring_lock contention.
3905 if (ctx
->flags
& IORING_SETUP_IOPOLL
&&
3906 !(ctx
->flags
& IORING_SETUP_SQPOLL
))
3907 ctx
->syscall_iopoll
= 1;
3909 ctx
->compat
= in_compat_syscall();
3910 if (!ns_capable_noaudit(&init_user_ns
, CAP_IPC_LOCK
))
3911 ctx
->user
= get_uid(current_user());
3914 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3915 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3918 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
3919 /* IPI related flags don't make sense with SQPOLL */
3920 if (ctx
->flags
& (IORING_SETUP_COOP_TASKRUN
|
3921 IORING_SETUP_TASKRUN_FLAG
|
3922 IORING_SETUP_DEFER_TASKRUN
))
3924 ctx
->notify_method
= TWA_SIGNAL_NO_IPI
;
3925 } else if (ctx
->flags
& IORING_SETUP_COOP_TASKRUN
) {
3926 ctx
->notify_method
= TWA_SIGNAL_NO_IPI
;
3928 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
&&
3929 !(ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
))
3931 ctx
->notify_method
= TWA_SIGNAL
;
3935 * For DEFER_TASKRUN we require the completion task to be the same as the
3936 * submission task. This implies that there is only one submitter, so enforce
3939 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
&&
3940 !(ctx
->flags
& IORING_SETUP_SINGLE_ISSUER
)) {
3945 * This is just grabbed for accounting purposes. When a process exits,
3946 * the mm is exited and dropped before the files, hence we need to hang
3947 * on to this mm purely for the purposes of being able to unaccount
3948 * memory (locked/pinned vm). It's not used for anything else.
3950 mmgrab(current
->mm
);
3951 ctx
->mm_account
= current
->mm
;
3953 ret
= io_allocate_scq_urings(ctx
, p
);
3957 ret
= io_sq_offload_create(ctx
, p
);
3961 ret
= io_rsrc_init(ctx
);
3965 p
->sq_off
.head
= offsetof(struct io_rings
, sq
.head
);
3966 p
->sq_off
.tail
= offsetof(struct io_rings
, sq
.tail
);
3967 p
->sq_off
.ring_mask
= offsetof(struct io_rings
, sq_ring_mask
);
3968 p
->sq_off
.ring_entries
= offsetof(struct io_rings
, sq_ring_entries
);
3969 p
->sq_off
.flags
= offsetof(struct io_rings
, sq_flags
);
3970 p
->sq_off
.dropped
= offsetof(struct io_rings
, sq_dropped
);
3971 if (!(ctx
->flags
& IORING_SETUP_NO_SQARRAY
))
3972 p
->sq_off
.array
= (char *)ctx
->sq_array
- (char *)ctx
->rings
;
3973 p
->sq_off
.resv1
= 0;
3974 if (!(ctx
->flags
& IORING_SETUP_NO_MMAP
))
3975 p
->sq_off
.user_addr
= 0;
3977 p
->cq_off
.head
= offsetof(struct io_rings
, cq
.head
);
3978 p
->cq_off
.tail
= offsetof(struct io_rings
, cq
.tail
);
3979 p
->cq_off
.ring_mask
= offsetof(struct io_rings
, cq_ring_mask
);
3980 p
->cq_off
.ring_entries
= offsetof(struct io_rings
, cq_ring_entries
);
3981 p
->cq_off
.overflow
= offsetof(struct io_rings
, cq_overflow
);
3982 p
->cq_off
.cqes
= offsetof(struct io_rings
, cqes
);
3983 p
->cq_off
.flags
= offsetof(struct io_rings
, cq_flags
);
3984 p
->cq_off
.resv1
= 0;
3985 if (!(ctx
->flags
& IORING_SETUP_NO_MMAP
))
3986 p
->cq_off
.user_addr
= 0;
3988 p
->features
= IORING_FEAT_SINGLE_MMAP
| IORING_FEAT_NODROP
|
3989 IORING_FEAT_SUBMIT_STABLE
| IORING_FEAT_RW_CUR_POS
|
3990 IORING_FEAT_CUR_PERSONALITY
| IORING_FEAT_FAST_POLL
|
3991 IORING_FEAT_POLL_32BITS
| IORING_FEAT_SQPOLL_NONFIXED
|
3992 IORING_FEAT_EXT_ARG
| IORING_FEAT_NATIVE_WORKERS
|
3993 IORING_FEAT_RSRC_TAGS
| IORING_FEAT_CQE_SKIP
|
3994 IORING_FEAT_LINKED_FILE
| IORING_FEAT_REG_REG_RING
;
3996 if (copy_to_user(params
, p
, sizeof(*p
))) {
4001 if (ctx
->flags
& IORING_SETUP_SINGLE_ISSUER
4002 && !(ctx
->flags
& IORING_SETUP_R_DISABLED
))
4003 WRITE_ONCE(ctx
->submitter_task
, get_task_struct(current
));
4005 file
= io_uring_get_file(ctx
);
4007 ret
= PTR_ERR(file
);
4011 ret
= __io_uring_add_tctx_node(ctx
);
4014 tctx
= current
->io_uring
;
4017 * Install ring fd as the very last thing, so we don't risk someone
4018 * having closed it before we finish setup
4020 if (p
->flags
& IORING_SETUP_REGISTERED_FD_ONLY
)
4021 ret
= io_ring_add_registered_file(tctx
, file
, 0, IO_RINGFD_REG_MAX
);
4023 ret
= io_uring_install_fd(file
);
4027 trace_io_uring_create(ret
, ctx
, p
->sq_entries
, p
->cq_entries
, p
->flags
);
4030 io_ring_ctx_wait_and_kill(ctx
);
4038 * Sets up an aio uring context, and returns the fd. Applications asks for a
4039 * ring size, we return the actual sq/cq ring sizes (among other things) in the
4040 * params structure passed in.
4042 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
4044 struct io_uring_params p
;
4047 if (copy_from_user(&p
, params
, sizeof(p
)))
4049 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
4054 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
4055 IORING_SETUP_SQ_AFF
| IORING_SETUP_CQSIZE
|
4056 IORING_SETUP_CLAMP
| IORING_SETUP_ATTACH_WQ
|
4057 IORING_SETUP_R_DISABLED
| IORING_SETUP_SUBMIT_ALL
|
4058 IORING_SETUP_COOP_TASKRUN
| IORING_SETUP_TASKRUN_FLAG
|
4059 IORING_SETUP_SQE128
| IORING_SETUP_CQE32
|
4060 IORING_SETUP_SINGLE_ISSUER
| IORING_SETUP_DEFER_TASKRUN
|
4061 IORING_SETUP_NO_MMAP
| IORING_SETUP_REGISTERED_FD_ONLY
|
4062 IORING_SETUP_NO_SQARRAY
))
4065 return io_uring_create(entries
, &p
, params
);
4068 static inline bool io_uring_allowed(void)
4070 int disabled
= READ_ONCE(sysctl_io_uring_disabled
);
4071 kgid_t io_uring_group
;
4076 if (disabled
== 0 || capable(CAP_SYS_ADMIN
))
4079 io_uring_group
= make_kgid(&init_user_ns
, sysctl_io_uring_group
);
4080 if (!gid_valid(io_uring_group
))
4083 return in_group_p(io_uring_group
);
4086 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
4087 struct io_uring_params __user
*, params
)
4089 if (!io_uring_allowed())
4092 return io_uring_setup(entries
, params
);
4095 static __cold
int io_probe(struct io_ring_ctx
*ctx
, void __user
*arg
,
4098 struct io_uring_probe
*p
;
4102 size
= struct_size(p
, ops
, nr_args
);
4103 if (size
== SIZE_MAX
)
4105 p
= kzalloc(size
, GFP_KERNEL
);
4110 if (copy_from_user(p
, arg
, size
))
4113 if (memchr_inv(p
, 0, size
))
4116 p
->last_op
= IORING_OP_LAST
- 1;
4117 if (nr_args
> IORING_OP_LAST
)
4118 nr_args
= IORING_OP_LAST
;
4120 for (i
= 0; i
< nr_args
; i
++) {
4122 if (!io_issue_defs
[i
].not_supported
)
4123 p
->ops
[i
].flags
= IO_URING_OP_SUPPORTED
;
4128 if (copy_to_user(arg
, p
, size
))
4135 static int io_register_personality(struct io_ring_ctx
*ctx
)
4137 const struct cred
*creds
;
4141 creds
= get_current_cred();
4143 ret
= xa_alloc_cyclic(&ctx
->personalities
, &id
, (void *)creds
,
4144 XA_LIMIT(0, USHRT_MAX
), &ctx
->pers_next
, GFP_KERNEL
);
4152 static __cold
int io_register_restrictions(struct io_ring_ctx
*ctx
,
4153 void __user
*arg
, unsigned int nr_args
)
4155 struct io_uring_restriction
*res
;
4159 /* Restrictions allowed only if rings started disabled */
4160 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
4163 /* We allow only a single restrictions registration */
4164 if (ctx
->restrictions
.registered
)
4167 if (!arg
|| nr_args
> IORING_MAX_RESTRICTIONS
)
4170 size
= array_size(nr_args
, sizeof(*res
));
4171 if (size
== SIZE_MAX
)
4174 res
= memdup_user(arg
, size
);
4176 return PTR_ERR(res
);
4180 for (i
= 0; i
< nr_args
; i
++) {
4181 switch (res
[i
].opcode
) {
4182 case IORING_RESTRICTION_REGISTER_OP
:
4183 if (res
[i
].register_op
>= IORING_REGISTER_LAST
) {
4188 __set_bit(res
[i
].register_op
,
4189 ctx
->restrictions
.register_op
);
4191 case IORING_RESTRICTION_SQE_OP
:
4192 if (res
[i
].sqe_op
>= IORING_OP_LAST
) {
4197 __set_bit(res
[i
].sqe_op
, ctx
->restrictions
.sqe_op
);
4199 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED
:
4200 ctx
->restrictions
.sqe_flags_allowed
= res
[i
].sqe_flags
;
4202 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED
:
4203 ctx
->restrictions
.sqe_flags_required
= res
[i
].sqe_flags
;
4212 /* Reset all restrictions if an error happened */
4214 memset(&ctx
->restrictions
, 0, sizeof(ctx
->restrictions
));
4216 ctx
->restrictions
.registered
= true;
4222 static int io_register_enable_rings(struct io_ring_ctx
*ctx
)
4224 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
4227 if (ctx
->flags
& IORING_SETUP_SINGLE_ISSUER
&& !ctx
->submitter_task
) {
4228 WRITE_ONCE(ctx
->submitter_task
, get_task_struct(current
));
4230 * Lazy activation attempts would fail if it was polled before
4231 * submitter_task is set.
4233 if (wq_has_sleeper(&ctx
->poll_wq
))
4234 io_activate_pollwq(ctx
);
4237 if (ctx
->restrictions
.registered
)
4238 ctx
->restricted
= 1;
4240 ctx
->flags
&= ~IORING_SETUP_R_DISABLED
;
4241 if (ctx
->sq_data
&& wq_has_sleeper(&ctx
->sq_data
->wait
))
4242 wake_up(&ctx
->sq_data
->wait
);
4246 static __cold
int __io_register_iowq_aff(struct io_ring_ctx
*ctx
,
4247 cpumask_var_t new_mask
)
4251 if (!(ctx
->flags
& IORING_SETUP_SQPOLL
)) {
4252 ret
= io_wq_cpu_affinity(current
->io_uring
, new_mask
);
4254 mutex_unlock(&ctx
->uring_lock
);
4255 ret
= io_sqpoll_wq_cpu_affinity(ctx
, new_mask
);
4256 mutex_lock(&ctx
->uring_lock
);
4262 static __cold
int io_register_iowq_aff(struct io_ring_ctx
*ctx
,
4263 void __user
*arg
, unsigned len
)
4265 cpumask_var_t new_mask
;
4268 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4271 cpumask_clear(new_mask
);
4272 if (len
> cpumask_size())
4273 len
= cpumask_size();
4275 if (in_compat_syscall()) {
4276 ret
= compat_get_bitmap(cpumask_bits(new_mask
),
4277 (const compat_ulong_t __user
*)arg
,
4278 len
* 8 /* CHAR_BIT */);
4280 ret
= copy_from_user(new_mask
, arg
, len
);
4284 free_cpumask_var(new_mask
);
4288 ret
= __io_register_iowq_aff(ctx
, new_mask
);
4289 free_cpumask_var(new_mask
);
4293 static __cold
int io_unregister_iowq_aff(struct io_ring_ctx
*ctx
)
4295 return __io_register_iowq_aff(ctx
, NULL
);
4298 static __cold
int io_register_iowq_max_workers(struct io_ring_ctx
*ctx
,
4300 __must_hold(&ctx
->uring_lock
)
4302 struct io_tctx_node
*node
;
4303 struct io_uring_task
*tctx
= NULL
;
4304 struct io_sq_data
*sqd
= NULL
;
4308 if (copy_from_user(new_count
, arg
, sizeof(new_count
)))
4310 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
4311 if (new_count
[i
] > INT_MAX
)
4314 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
4318 * Observe the correct sqd->lock -> ctx->uring_lock
4319 * ordering. Fine to drop uring_lock here, we hold
4322 refcount_inc(&sqd
->refs
);
4323 mutex_unlock(&ctx
->uring_lock
);
4324 mutex_lock(&sqd
->lock
);
4325 mutex_lock(&ctx
->uring_lock
);
4327 tctx
= sqd
->thread
->io_uring
;
4330 tctx
= current
->io_uring
;
4333 BUILD_BUG_ON(sizeof(new_count
) != sizeof(ctx
->iowq_limits
));
4335 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
4337 ctx
->iowq_limits
[i
] = new_count
[i
];
4338 ctx
->iowq_limits_set
= true;
4340 if (tctx
&& tctx
->io_wq
) {
4341 ret
= io_wq_max_workers(tctx
->io_wq
, new_count
);
4345 memset(new_count
, 0, sizeof(new_count
));
4349 mutex_unlock(&sqd
->lock
);
4350 io_put_sq_data(sqd
);
4353 if (copy_to_user(arg
, new_count
, sizeof(new_count
)))
4356 /* that's it for SQPOLL, only the SQPOLL task creates requests */
4360 /* now propagate the restriction to all registered users */
4361 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
4362 struct io_uring_task
*tctx
= node
->task
->io_uring
;
4364 if (WARN_ON_ONCE(!tctx
->io_wq
))
4367 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
4368 new_count
[i
] = ctx
->iowq_limits
[i
];
4369 /* ignore errors, it always returns zero anyway */
4370 (void)io_wq_max_workers(tctx
->io_wq
, new_count
);
4375 mutex_unlock(&sqd
->lock
);
4376 io_put_sq_data(sqd
);
4381 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
4382 void __user
*arg
, unsigned nr_args
)
4383 __releases(ctx
->uring_lock
)
4384 __acquires(ctx
->uring_lock
)
4389 * We don't quiesce the refs for register anymore and so it can't be
4390 * dying as we're holding a file ref here.
4392 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx
->refs
)))
4395 if (ctx
->submitter_task
&& ctx
->submitter_task
!= current
)
4398 if (ctx
->restricted
) {
4399 opcode
= array_index_nospec(opcode
, IORING_REGISTER_LAST
);
4400 if (!test_bit(opcode
, ctx
->restrictions
.register_op
))
4405 case IORING_REGISTER_BUFFERS
:
4409 ret
= io_sqe_buffers_register(ctx
, arg
, nr_args
, NULL
);
4411 case IORING_UNREGISTER_BUFFERS
:
4415 ret
= io_sqe_buffers_unregister(ctx
);
4417 case IORING_REGISTER_FILES
:
4421 ret
= io_sqe_files_register(ctx
, arg
, nr_args
, NULL
);
4423 case IORING_UNREGISTER_FILES
:
4427 ret
= io_sqe_files_unregister(ctx
);
4429 case IORING_REGISTER_FILES_UPDATE
:
4430 ret
= io_register_files_update(ctx
, arg
, nr_args
);
4432 case IORING_REGISTER_EVENTFD
:
4436 ret
= io_eventfd_register(ctx
, arg
, 0);
4438 case IORING_REGISTER_EVENTFD_ASYNC
:
4442 ret
= io_eventfd_register(ctx
, arg
, 1);
4444 case IORING_UNREGISTER_EVENTFD
:
4448 ret
= io_eventfd_unregister(ctx
);
4450 case IORING_REGISTER_PROBE
:
4452 if (!arg
|| nr_args
> 256)
4454 ret
= io_probe(ctx
, arg
, nr_args
);
4456 case IORING_REGISTER_PERSONALITY
:
4460 ret
= io_register_personality(ctx
);
4462 case IORING_UNREGISTER_PERSONALITY
:
4466 ret
= io_unregister_personality(ctx
, nr_args
);
4468 case IORING_REGISTER_ENABLE_RINGS
:
4472 ret
= io_register_enable_rings(ctx
);
4474 case IORING_REGISTER_RESTRICTIONS
:
4475 ret
= io_register_restrictions(ctx
, arg
, nr_args
);
4477 case IORING_REGISTER_FILES2
:
4478 ret
= io_register_rsrc(ctx
, arg
, nr_args
, IORING_RSRC_FILE
);
4480 case IORING_REGISTER_FILES_UPDATE2
:
4481 ret
= io_register_rsrc_update(ctx
, arg
, nr_args
,
4484 case IORING_REGISTER_BUFFERS2
:
4485 ret
= io_register_rsrc(ctx
, arg
, nr_args
, IORING_RSRC_BUFFER
);
4487 case IORING_REGISTER_BUFFERS_UPDATE
:
4488 ret
= io_register_rsrc_update(ctx
, arg
, nr_args
,
4489 IORING_RSRC_BUFFER
);
4491 case IORING_REGISTER_IOWQ_AFF
:
4493 if (!arg
|| !nr_args
)
4495 ret
= io_register_iowq_aff(ctx
, arg
, nr_args
);
4497 case IORING_UNREGISTER_IOWQ_AFF
:
4501 ret
= io_unregister_iowq_aff(ctx
);
4503 case IORING_REGISTER_IOWQ_MAX_WORKERS
:
4505 if (!arg
|| nr_args
!= 2)
4507 ret
= io_register_iowq_max_workers(ctx
, arg
);
4509 case IORING_REGISTER_RING_FDS
:
4510 ret
= io_ringfd_register(ctx
, arg
, nr_args
);
4512 case IORING_UNREGISTER_RING_FDS
:
4513 ret
= io_ringfd_unregister(ctx
, arg
, nr_args
);
4515 case IORING_REGISTER_PBUF_RING
:
4517 if (!arg
|| nr_args
!= 1)
4519 ret
= io_register_pbuf_ring(ctx
, arg
);
4521 case IORING_UNREGISTER_PBUF_RING
:
4523 if (!arg
|| nr_args
!= 1)
4525 ret
= io_unregister_pbuf_ring(ctx
, arg
);
4527 case IORING_REGISTER_SYNC_CANCEL
:
4529 if (!arg
|| nr_args
!= 1)
4531 ret
= io_sync_cancel(ctx
, arg
);
4533 case IORING_REGISTER_FILE_ALLOC_RANGE
:
4535 if (!arg
|| nr_args
)
4537 ret
= io_register_file_alloc_range(ctx
, arg
);
4547 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
4548 void __user
*, arg
, unsigned int, nr_args
)
4550 struct io_ring_ctx
*ctx
;
4553 bool use_registered_ring
;
4555 use_registered_ring
= !!(opcode
& IORING_REGISTER_USE_REGISTERED_RING
);
4556 opcode
&= ~IORING_REGISTER_USE_REGISTERED_RING
;
4558 if (opcode
>= IORING_REGISTER_LAST
)
4561 if (use_registered_ring
) {
4563 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
4564 * need only dereference our task private array to find it.
4566 struct io_uring_task
*tctx
= current
->io_uring
;
4568 if (unlikely(!tctx
|| fd
>= IO_RINGFD_REG_MAX
))
4570 fd
= array_index_nospec(fd
, IO_RINGFD_REG_MAX
);
4571 f
.file
= tctx
->registered_rings
[fd
];
4573 if (unlikely(!f
.file
))
4577 if (unlikely(!f
.file
))
4580 if (!io_is_uring_fops(f
.file
))
4584 ctx
= f
.file
->private_data
;
4586 mutex_lock(&ctx
->uring_lock
);
4587 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
4588 mutex_unlock(&ctx
->uring_lock
);
4589 trace_io_uring_register(ctx
, opcode
, ctx
->nr_user_files
, ctx
->nr_user_bufs
, ret
);
4595 static int __init
io_uring_init(void)
4597 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4598 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4599 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4602 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4603 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4604 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4605 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4606 BUILD_BUG_ON(sizeof(struct io_uring_sqe
) != 64);
4607 BUILD_BUG_SQE_ELEM(0, __u8
, opcode
);
4608 BUILD_BUG_SQE_ELEM(1, __u8
, flags
);
4609 BUILD_BUG_SQE_ELEM(2, __u16
, ioprio
);
4610 BUILD_BUG_SQE_ELEM(4, __s32
, fd
);
4611 BUILD_BUG_SQE_ELEM(8, __u64
, off
);
4612 BUILD_BUG_SQE_ELEM(8, __u64
, addr2
);
4613 BUILD_BUG_SQE_ELEM(8, __u32
, cmd_op
);
4614 BUILD_BUG_SQE_ELEM(12, __u32
, __pad1
);
4615 BUILD_BUG_SQE_ELEM(16, __u64
, addr
);
4616 BUILD_BUG_SQE_ELEM(16, __u64
, splice_off_in
);
4617 BUILD_BUG_SQE_ELEM(24, __u32
, len
);
4618 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t
, rw_flags
);
4619 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags
);
4620 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32
, rw_flags
);
4621 BUILD_BUG_SQE_ELEM(28, __u32
, fsync_flags
);
4622 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16
, poll_events
);
4623 BUILD_BUG_SQE_ELEM(28, __u32
, poll32_events
);
4624 BUILD_BUG_SQE_ELEM(28, __u32
, sync_range_flags
);
4625 BUILD_BUG_SQE_ELEM(28, __u32
, msg_flags
);
4626 BUILD_BUG_SQE_ELEM(28, __u32
, timeout_flags
);
4627 BUILD_BUG_SQE_ELEM(28, __u32
, accept_flags
);
4628 BUILD_BUG_SQE_ELEM(28, __u32
, cancel_flags
);
4629 BUILD_BUG_SQE_ELEM(28, __u32
, open_flags
);
4630 BUILD_BUG_SQE_ELEM(28, __u32
, statx_flags
);
4631 BUILD_BUG_SQE_ELEM(28, __u32
, fadvise_advice
);
4632 BUILD_BUG_SQE_ELEM(28, __u32
, splice_flags
);
4633 BUILD_BUG_SQE_ELEM(28, __u32
, rename_flags
);
4634 BUILD_BUG_SQE_ELEM(28, __u32
, unlink_flags
);
4635 BUILD_BUG_SQE_ELEM(28, __u32
, hardlink_flags
);
4636 BUILD_BUG_SQE_ELEM(28, __u32
, xattr_flags
);
4637 BUILD_BUG_SQE_ELEM(28, __u32
, msg_ring_flags
);
4638 BUILD_BUG_SQE_ELEM(32, __u64
, user_data
);
4639 BUILD_BUG_SQE_ELEM(40, __u16
, buf_index
);
4640 BUILD_BUG_SQE_ELEM(40, __u16
, buf_group
);
4641 BUILD_BUG_SQE_ELEM(42, __u16
, personality
);
4642 BUILD_BUG_SQE_ELEM(44, __s32
, splice_fd_in
);
4643 BUILD_BUG_SQE_ELEM(44, __u32
, file_index
);
4644 BUILD_BUG_SQE_ELEM(44, __u16
, addr_len
);
4645 BUILD_BUG_SQE_ELEM(46, __u16
, __pad3
[0]);
4646 BUILD_BUG_SQE_ELEM(48, __u64
, addr3
);
4647 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd
);
4648 BUILD_BUG_SQE_ELEM(56, __u64
, __pad2
);
4650 BUILD_BUG_ON(sizeof(struct io_uring_files_update
) !=
4651 sizeof(struct io_uring_rsrc_update
));
4652 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update
) >
4653 sizeof(struct io_uring_rsrc_update2
));
4655 /* ->buf_index is u16 */
4656 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring
, bufs
) != 0);
4657 BUILD_BUG_ON(offsetof(struct io_uring_buf
, resv
) !=
4658 offsetof(struct io_uring_buf_ring
, tail
));
4660 /* should fit into one byte */
4661 BUILD_BUG_ON(SQE_VALID_FLAGS
>= (1 << 8));
4662 BUILD_BUG_ON(SQE_COMMON_FLAGS
>= (1 << 8));
4663 BUILD_BUG_ON((SQE_VALID_FLAGS
| SQE_COMMON_FLAGS
) != SQE_VALID_FLAGS
);
4665 BUILD_BUG_ON(__REQ_F_LAST_BIT
> 8 * sizeof(int));
4667 BUILD_BUG_ON(sizeof(atomic_t
) != sizeof(u32
));
4669 io_uring_optable_init();
4672 * Allow user copy in the per-command field, which starts after the
4673 * file in io_kiocb and until the opcode field. The openat2 handling
4674 * requires copying in user memory into the io_kiocb object in that
4675 * range, and HARDENED_USERCOPY will complain if we haven't
4676 * correctly annotated this range.
4678 req_cachep
= kmem_cache_create_usercopy("io_kiocb",
4679 sizeof(struct io_kiocb
), 0,
4680 SLAB_HWCACHE_ALIGN
| SLAB_PANIC
|
4681 SLAB_ACCOUNT
| SLAB_TYPESAFE_BY_RCU
,
4682 offsetof(struct io_kiocb
, cmd
.data
),
4683 sizeof_field(struct io_kiocb
, cmd
.data
), NULL
);
4685 #ifdef CONFIG_SYSCTL
4686 register_sysctl_init("kernel", kernel_io_uring_disabled_table
);
4691 __initcall(io_uring_init
);