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>
63 #include <linux/anon_inodes.h>
64 #include <linux/sched/mm.h>
65 #include <linux/uaccess.h>
66 #include <linux/nospec.h>
67 #include <linux/highmem.h>
68 #include <linux/fsnotify.h>
69 #include <linux/fadvise.h>
70 #include <linux/task_work.h>
71 #include <linux/io_uring.h>
72 #include <linux/io_uring/cmd.h>
73 #include <linux/audit.h>
74 #include <linux/security.h>
75 #include <asm/shmparam.h>
77 #define CREATE_TRACE_POINTS
78 #include <trace/events/io_uring.h>
80 #include <uapi/linux/io_uring.h>
102 #include "alloc_cache.h"
104 #define IORING_MAX_ENTRIES 32768
105 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
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
,
130 struct io_defer_entry
{
131 struct list_head list
;
132 struct io_kiocb
*req
;
136 /* requests with any of those set should undergo io_disarm_next() */
137 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
138 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
141 * No waiters. It's larger than any valid value of the tw counter
142 * so that tests against ->cq_wait_nr would fail and skip wake_up().
144 #define IO_CQ_WAKE_INIT (-1U)
145 /* Forced wake up if there is a waiter regardless of ->cq_wait_nr */
146 #define IO_CQ_WAKE_FORCE (IO_CQ_WAKE_INIT >> 1)
148 static bool io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
149 struct task_struct
*task
,
152 static void io_queue_sqe(struct io_kiocb
*req
);
154 struct kmem_cache
*req_cachep
;
156 static int __read_mostly sysctl_io_uring_disabled
;
157 static int __read_mostly sysctl_io_uring_group
= -1;
160 static struct ctl_table kernel_io_uring_disabled_table
[] = {
162 .procname
= "io_uring_disabled",
163 .data
= &sysctl_io_uring_disabled
,
164 .maxlen
= sizeof(sysctl_io_uring_disabled
),
166 .proc_handler
= proc_dointvec_minmax
,
167 .extra1
= SYSCTL_ZERO
,
168 .extra2
= SYSCTL_TWO
,
171 .procname
= "io_uring_group",
172 .data
= &sysctl_io_uring_group
,
173 .maxlen
= sizeof(gid_t
),
175 .proc_handler
= proc_dointvec
,
181 static inline void io_submit_flush_completions(struct io_ring_ctx
*ctx
)
183 if (!wq_list_empty(&ctx
->submit_state
.compl_reqs
) ||
184 ctx
->submit_state
.cqes_count
)
185 __io_submit_flush_completions(ctx
);
188 static inline unsigned int __io_cqring_events(struct io_ring_ctx
*ctx
)
190 return ctx
->cached_cq_tail
- READ_ONCE(ctx
->rings
->cq
.head
);
193 static inline unsigned int __io_cqring_events_user(struct io_ring_ctx
*ctx
)
195 return READ_ONCE(ctx
->rings
->cq
.tail
) - READ_ONCE(ctx
->rings
->cq
.head
);
198 static bool io_match_linked(struct io_kiocb
*head
)
200 struct io_kiocb
*req
;
202 io_for_each_link(req
, head
) {
203 if (req
->flags
& REQ_F_INFLIGHT
)
210 * As io_match_task() but protected against racing with linked timeouts.
211 * User must not hold timeout_lock.
213 bool io_match_task_safe(struct io_kiocb
*head
, struct task_struct
*task
,
218 if (task
&& head
->task
!= task
)
223 if (head
->flags
& REQ_F_LINK_TIMEOUT
) {
224 struct io_ring_ctx
*ctx
= head
->ctx
;
226 /* protect against races with linked timeouts */
227 spin_lock_irq(&ctx
->timeout_lock
);
228 matched
= io_match_linked(head
);
229 spin_unlock_irq(&ctx
->timeout_lock
);
231 matched
= io_match_linked(head
);
236 static inline void req_fail_link_node(struct io_kiocb
*req
, int res
)
239 io_req_set_res(req
, res
, 0);
242 static inline void io_req_add_to_cache(struct io_kiocb
*req
, struct io_ring_ctx
*ctx
)
244 wq_stack_add_head(&req
->comp_list
, &ctx
->submit_state
.free_list
);
247 static __cold
void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
249 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
251 complete(&ctx
->ref_comp
);
254 static __cold
void io_fallback_req_func(struct work_struct
*work
)
256 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
,
258 struct llist_node
*node
= llist_del_all(&ctx
->fallback_llist
);
259 struct io_kiocb
*req
, *tmp
;
260 struct io_tw_state ts
= { .locked
= true, };
262 percpu_ref_get(&ctx
->refs
);
263 mutex_lock(&ctx
->uring_lock
);
264 llist_for_each_entry_safe(req
, tmp
, node
, io_task_work
.node
)
265 req
->io_task_work
.func(req
, &ts
);
266 if (WARN_ON_ONCE(!ts
.locked
))
268 io_submit_flush_completions(ctx
);
269 mutex_unlock(&ctx
->uring_lock
);
270 percpu_ref_put(&ctx
->refs
);
273 static int io_alloc_hash_table(struct io_hash_table
*table
, unsigned bits
)
275 unsigned hash_buckets
= 1U << bits
;
276 size_t hash_size
= hash_buckets
* sizeof(table
->hbs
[0]);
278 table
->hbs
= kmalloc(hash_size
, GFP_KERNEL
);
282 table
->hash_bits
= bits
;
283 init_hash_table(table
, hash_buckets
);
287 static __cold
struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
289 struct io_ring_ctx
*ctx
;
292 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
296 xa_init(&ctx
->io_bl_xa
);
299 * Use 5 bits less than the max cq entries, that should give us around
300 * 32 entries per hash list if totally full and uniformly spread, but
301 * don't keep too many buckets to not overconsume memory.
303 hash_bits
= ilog2(p
->cq_entries
) - 5;
304 hash_bits
= clamp(hash_bits
, 1, 8);
305 if (io_alloc_hash_table(&ctx
->cancel_table
, hash_bits
))
307 if (io_alloc_hash_table(&ctx
->cancel_table_locked
, hash_bits
))
309 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
,
313 ctx
->flags
= p
->flags
;
314 atomic_set(&ctx
->cq_wait_nr
, IO_CQ_WAKE_INIT
);
315 init_waitqueue_head(&ctx
->sqo_sq_wait
);
316 INIT_LIST_HEAD(&ctx
->sqd_list
);
317 INIT_LIST_HEAD(&ctx
->cq_overflow_list
);
318 INIT_LIST_HEAD(&ctx
->io_buffers_cache
);
319 INIT_HLIST_HEAD(&ctx
->io_buf_list
);
320 io_alloc_cache_init(&ctx
->rsrc_node_cache
, IO_NODE_ALLOC_CACHE_MAX
,
321 sizeof(struct io_rsrc_node
));
322 io_alloc_cache_init(&ctx
->apoll_cache
, IO_ALLOC_CACHE_MAX
,
323 sizeof(struct async_poll
));
324 io_alloc_cache_init(&ctx
->netmsg_cache
, IO_ALLOC_CACHE_MAX
,
325 sizeof(struct io_async_msghdr
));
326 io_futex_cache_init(ctx
);
327 init_completion(&ctx
->ref_comp
);
328 xa_init_flags(&ctx
->personalities
, XA_FLAGS_ALLOC1
);
329 mutex_init(&ctx
->uring_lock
);
330 init_waitqueue_head(&ctx
->cq_wait
);
331 init_waitqueue_head(&ctx
->poll_wq
);
332 init_waitqueue_head(&ctx
->rsrc_quiesce_wq
);
333 spin_lock_init(&ctx
->completion_lock
);
334 spin_lock_init(&ctx
->timeout_lock
);
335 INIT_WQ_LIST(&ctx
->iopoll_list
);
336 INIT_LIST_HEAD(&ctx
->io_buffers_comp
);
337 INIT_LIST_HEAD(&ctx
->defer_list
);
338 INIT_LIST_HEAD(&ctx
->timeout_list
);
339 INIT_LIST_HEAD(&ctx
->ltimeout_list
);
340 INIT_LIST_HEAD(&ctx
->rsrc_ref_list
);
341 init_llist_head(&ctx
->work_llist
);
342 INIT_LIST_HEAD(&ctx
->tctx_list
);
343 ctx
->submit_state
.free_list
.next
= NULL
;
344 INIT_WQ_LIST(&ctx
->locked_free_list
);
345 INIT_HLIST_HEAD(&ctx
->waitid_list
);
347 INIT_HLIST_HEAD(&ctx
->futex_list
);
349 INIT_DELAYED_WORK(&ctx
->fallback_work
, io_fallback_req_func
);
350 INIT_WQ_LIST(&ctx
->submit_state
.compl_reqs
);
351 INIT_HLIST_HEAD(&ctx
->cancelable_uring_cmd
);
354 kfree(ctx
->cancel_table
.hbs
);
355 kfree(ctx
->cancel_table_locked
.hbs
);
357 xa_destroy(&ctx
->io_bl_xa
);
362 static void io_account_cq_overflow(struct io_ring_ctx
*ctx
)
364 struct io_rings
*r
= ctx
->rings
;
366 WRITE_ONCE(r
->cq_overflow
, READ_ONCE(r
->cq_overflow
) + 1);
370 static bool req_need_defer(struct io_kiocb
*req
, u32 seq
)
372 if (unlikely(req
->flags
& REQ_F_IO_DRAIN
)) {
373 struct io_ring_ctx
*ctx
= req
->ctx
;
375 return seq
+ READ_ONCE(ctx
->cq_extra
) != ctx
->cached_cq_tail
;
381 static void io_clean_op(struct io_kiocb
*req
)
383 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
384 spin_lock(&req
->ctx
->completion_lock
);
385 io_put_kbuf_comp(req
);
386 spin_unlock(&req
->ctx
->completion_lock
);
389 if (req
->flags
& REQ_F_NEED_CLEANUP
) {
390 const struct io_cold_def
*def
= &io_cold_defs
[req
->opcode
];
395 if ((req
->flags
& REQ_F_POLLED
) && req
->apoll
) {
396 kfree(req
->apoll
->double_poll
);
400 if (req
->flags
& REQ_F_INFLIGHT
) {
401 struct io_uring_task
*tctx
= req
->task
->io_uring
;
403 atomic_dec(&tctx
->inflight_tracked
);
405 if (req
->flags
& REQ_F_CREDS
)
406 put_cred(req
->creds
);
407 if (req
->flags
& REQ_F_ASYNC_DATA
) {
408 kfree(req
->async_data
);
409 req
->async_data
= NULL
;
411 req
->flags
&= ~IO_REQ_CLEAN_FLAGS
;
414 static inline void io_req_track_inflight(struct io_kiocb
*req
)
416 if (!(req
->flags
& REQ_F_INFLIGHT
)) {
417 req
->flags
|= REQ_F_INFLIGHT
;
418 atomic_inc(&req
->task
->io_uring
->inflight_tracked
);
422 static struct io_kiocb
*__io_prep_linked_timeout(struct io_kiocb
*req
)
424 if (WARN_ON_ONCE(!req
->link
))
427 req
->flags
&= ~REQ_F_ARM_LTIMEOUT
;
428 req
->flags
|= REQ_F_LINK_TIMEOUT
;
430 /* linked timeouts should have two refs once prep'ed */
431 io_req_set_refcount(req
);
432 __io_req_set_refcount(req
->link
, 2);
436 static inline struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
)
438 if (likely(!(req
->flags
& REQ_F_ARM_LTIMEOUT
)))
440 return __io_prep_linked_timeout(req
);
443 static noinline
void __io_arm_ltimeout(struct io_kiocb
*req
)
445 io_queue_linked_timeout(__io_prep_linked_timeout(req
));
448 static inline void io_arm_ltimeout(struct io_kiocb
*req
)
450 if (unlikely(req
->flags
& REQ_F_ARM_LTIMEOUT
))
451 __io_arm_ltimeout(req
);
454 static void io_prep_async_work(struct io_kiocb
*req
)
456 const struct io_issue_def
*def
= &io_issue_defs
[req
->opcode
];
457 struct io_ring_ctx
*ctx
= req
->ctx
;
459 if (!(req
->flags
& REQ_F_CREDS
)) {
460 req
->flags
|= REQ_F_CREDS
;
461 req
->creds
= get_current_cred();
464 req
->work
.list
.next
= NULL
;
466 req
->work
.cancel_seq
= atomic_read(&ctx
->cancel_seq
);
467 if (req
->flags
& REQ_F_FORCE_ASYNC
)
468 req
->work
.flags
|= IO_WQ_WORK_CONCURRENT
;
470 if (req
->file
&& !(req
->flags
& REQ_F_FIXED_FILE
))
471 req
->flags
|= io_file_get_flags(req
->file
);
473 if (req
->file
&& (req
->flags
& REQ_F_ISREG
)) {
474 bool should_hash
= def
->hash_reg_file
;
476 /* don't serialize this request if the fs doesn't need it */
477 if (should_hash
&& (req
->file
->f_flags
& O_DIRECT
) &&
478 (req
->file
->f_mode
& FMODE_DIO_PARALLEL_WRITE
))
480 if (should_hash
|| (ctx
->flags
& IORING_SETUP_IOPOLL
))
481 io_wq_hash_work(&req
->work
, file_inode(req
->file
));
482 } else if (!req
->file
|| !S_ISBLK(file_inode(req
->file
)->i_mode
)) {
483 if (def
->unbound_nonreg_file
)
484 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
488 static void io_prep_async_link(struct io_kiocb
*req
)
490 struct io_kiocb
*cur
;
492 if (req
->flags
& REQ_F_LINK_TIMEOUT
) {
493 struct io_ring_ctx
*ctx
= req
->ctx
;
495 spin_lock_irq(&ctx
->timeout_lock
);
496 io_for_each_link(cur
, req
)
497 io_prep_async_work(cur
);
498 spin_unlock_irq(&ctx
->timeout_lock
);
500 io_for_each_link(cur
, req
)
501 io_prep_async_work(cur
);
505 void io_queue_iowq(struct io_kiocb
*req
, struct io_tw_state
*ts_dont_use
)
507 struct io_kiocb
*link
= io_prep_linked_timeout(req
);
508 struct io_uring_task
*tctx
= req
->task
->io_uring
;
511 BUG_ON(!tctx
->io_wq
);
513 /* init ->work of the whole link before punting */
514 io_prep_async_link(req
);
517 * Not expected to happen, but if we do have a bug where this _can_
518 * happen, catch it here and ensure the request is marked as
519 * canceled. That will make io-wq go through the usual work cancel
520 * procedure rather than attempt to run this request (or create a new
523 if (WARN_ON_ONCE(!same_thread_group(req
->task
, current
)))
524 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
526 trace_io_uring_queue_async_work(req
, io_wq_is_hashed(&req
->work
));
527 io_wq_enqueue(tctx
->io_wq
, &req
->work
);
529 io_queue_linked_timeout(link
);
532 static __cold
void io_queue_deferred(struct io_ring_ctx
*ctx
)
534 while (!list_empty(&ctx
->defer_list
)) {
535 struct io_defer_entry
*de
= list_first_entry(&ctx
->defer_list
,
536 struct io_defer_entry
, list
);
538 if (req_need_defer(de
->req
, de
->seq
))
540 list_del_init(&de
->list
);
541 io_req_task_queue(de
->req
);
546 void io_eventfd_ops(struct rcu_head
*rcu
)
548 struct io_ev_fd
*ev_fd
= container_of(rcu
, struct io_ev_fd
, rcu
);
549 int ops
= atomic_xchg(&ev_fd
->ops
, 0);
551 if (ops
& BIT(IO_EVENTFD_OP_SIGNAL_BIT
))
552 eventfd_signal_mask(ev_fd
->cq_ev_fd
, EPOLL_URING_WAKE
);
554 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
555 * ordering in a race but if references are 0 we know we have to free
558 if (atomic_dec_and_test(&ev_fd
->refs
)) {
559 eventfd_ctx_put(ev_fd
->cq_ev_fd
);
564 static void io_eventfd_signal(struct io_ring_ctx
*ctx
)
566 struct io_ev_fd
*ev_fd
= NULL
;
570 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
573 ev_fd
= rcu_dereference(ctx
->io_ev_fd
);
576 * Check again if ev_fd exists incase an io_eventfd_unregister call
577 * completed between the NULL check of ctx->io_ev_fd at the start of
578 * the function and rcu_read_lock.
580 if (unlikely(!ev_fd
))
582 if (READ_ONCE(ctx
->rings
->cq_flags
) & IORING_CQ_EVENTFD_DISABLED
)
584 if (ev_fd
->eventfd_async
&& !io_wq_current_is_worker())
587 if (likely(eventfd_signal_allowed())) {
588 eventfd_signal_mask(ev_fd
->cq_ev_fd
, EPOLL_URING_WAKE
);
590 atomic_inc(&ev_fd
->refs
);
591 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT
), &ev_fd
->ops
))
592 call_rcu_hurry(&ev_fd
->rcu
, io_eventfd_ops
);
594 atomic_dec(&ev_fd
->refs
);
601 static void io_eventfd_flush_signal(struct io_ring_ctx
*ctx
)
605 spin_lock(&ctx
->completion_lock
);
608 * Eventfd should only get triggered when at least one event has been
609 * posted. Some applications rely on the eventfd notification count
610 * only changing IFF a new CQE has been added to the CQ ring. There's
611 * no depedency on 1:1 relationship between how many times this
612 * function is called (and hence the eventfd count) and number of CQEs
613 * posted to the CQ ring.
615 skip
= ctx
->cached_cq_tail
== ctx
->evfd_last_cq_tail
;
616 ctx
->evfd_last_cq_tail
= ctx
->cached_cq_tail
;
617 spin_unlock(&ctx
->completion_lock
);
621 io_eventfd_signal(ctx
);
624 void __io_commit_cqring_flush(struct io_ring_ctx
*ctx
)
626 if (ctx
->poll_activated
)
627 io_poll_wq_wake(ctx
);
628 if (ctx
->off_timeout_used
)
629 io_flush_timeouts(ctx
);
630 if (ctx
->drain_active
) {
631 spin_lock(&ctx
->completion_lock
);
632 io_queue_deferred(ctx
);
633 spin_unlock(&ctx
->completion_lock
);
636 io_eventfd_flush_signal(ctx
);
639 static inline void __io_cq_lock(struct io_ring_ctx
*ctx
)
641 if (!ctx
->lockless_cq
)
642 spin_lock(&ctx
->completion_lock
);
645 static inline void io_cq_lock(struct io_ring_ctx
*ctx
)
646 __acquires(ctx
->completion_lock
)
648 spin_lock(&ctx
->completion_lock
);
651 static inline void __io_cq_unlock_post(struct io_ring_ctx
*ctx
)
653 io_commit_cqring(ctx
);
654 if (!ctx
->task_complete
) {
655 if (!ctx
->lockless_cq
)
656 spin_unlock(&ctx
->completion_lock
);
657 /* IOPOLL rings only need to wake up if it's also SQPOLL */
658 if (!ctx
->syscall_iopoll
)
661 io_commit_cqring_flush(ctx
);
664 static void io_cq_unlock_post(struct io_ring_ctx
*ctx
)
665 __releases(ctx
->completion_lock
)
667 io_commit_cqring(ctx
);
668 spin_unlock(&ctx
->completion_lock
);
670 io_commit_cqring_flush(ctx
);
673 /* Returns true if there are no backlogged entries after the flush */
674 static void io_cqring_overflow_kill(struct io_ring_ctx
*ctx
)
676 struct io_overflow_cqe
*ocqe
;
679 spin_lock(&ctx
->completion_lock
);
680 list_splice_init(&ctx
->cq_overflow_list
, &list
);
681 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
);
682 spin_unlock(&ctx
->completion_lock
);
684 while (!list_empty(&list
)) {
685 ocqe
= list_first_entry(&list
, struct io_overflow_cqe
, list
);
686 list_del(&ocqe
->list
);
691 static void __io_cqring_overflow_flush(struct io_ring_ctx
*ctx
)
693 size_t cqe_size
= sizeof(struct io_uring_cqe
);
695 if (__io_cqring_events(ctx
) == ctx
->cq_entries
)
698 if (ctx
->flags
& IORING_SETUP_CQE32
)
702 while (!list_empty(&ctx
->cq_overflow_list
)) {
703 struct io_uring_cqe
*cqe
;
704 struct io_overflow_cqe
*ocqe
;
706 if (!io_get_cqe_overflow(ctx
, &cqe
, true))
708 ocqe
= list_first_entry(&ctx
->cq_overflow_list
,
709 struct io_overflow_cqe
, list
);
710 memcpy(cqe
, &ocqe
->cqe
, cqe_size
);
711 list_del(&ocqe
->list
);
715 if (list_empty(&ctx
->cq_overflow_list
)) {
716 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
);
717 atomic_andnot(IORING_SQ_CQ_OVERFLOW
, &ctx
->rings
->sq_flags
);
719 io_cq_unlock_post(ctx
);
722 static void io_cqring_do_overflow_flush(struct io_ring_ctx
*ctx
)
724 /* iopoll syncs against uring_lock, not completion_lock */
725 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
726 mutex_lock(&ctx
->uring_lock
);
727 __io_cqring_overflow_flush(ctx
);
728 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
729 mutex_unlock(&ctx
->uring_lock
);
732 static void io_cqring_overflow_flush(struct io_ring_ctx
*ctx
)
734 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
))
735 io_cqring_do_overflow_flush(ctx
);
738 /* can be called by any task */
739 static void io_put_task_remote(struct task_struct
*task
)
741 struct io_uring_task
*tctx
= task
->io_uring
;
743 percpu_counter_sub(&tctx
->inflight
, 1);
744 if (unlikely(atomic_read(&tctx
->in_cancel
)))
745 wake_up(&tctx
->wait
);
746 put_task_struct(task
);
749 /* used by a task to put its own references */
750 static void io_put_task_local(struct task_struct
*task
)
752 task
->io_uring
->cached_refs
++;
755 /* must to be called somewhat shortly after putting a request */
756 static inline void io_put_task(struct task_struct
*task
)
758 if (likely(task
== current
))
759 io_put_task_local(task
);
761 io_put_task_remote(task
);
764 void io_task_refs_refill(struct io_uring_task
*tctx
)
766 unsigned int refill
= -tctx
->cached_refs
+ IO_TCTX_REFS_CACHE_NR
;
768 percpu_counter_add(&tctx
->inflight
, refill
);
769 refcount_add(refill
, ¤t
->usage
);
770 tctx
->cached_refs
+= refill
;
773 static __cold
void io_uring_drop_tctx_refs(struct task_struct
*task
)
775 struct io_uring_task
*tctx
= task
->io_uring
;
776 unsigned int refs
= tctx
->cached_refs
;
779 tctx
->cached_refs
= 0;
780 percpu_counter_sub(&tctx
->inflight
, refs
);
781 put_task_struct_many(task
, refs
);
785 static bool io_cqring_event_overflow(struct io_ring_ctx
*ctx
, u64 user_data
,
786 s32 res
, u32 cflags
, u64 extra1
, u64 extra2
)
788 struct io_overflow_cqe
*ocqe
;
789 size_t ocq_size
= sizeof(struct io_overflow_cqe
);
790 bool is_cqe32
= (ctx
->flags
& IORING_SETUP_CQE32
);
792 lockdep_assert_held(&ctx
->completion_lock
);
795 ocq_size
+= sizeof(struct io_uring_cqe
);
797 ocqe
= kmalloc(ocq_size
, GFP_ATOMIC
| __GFP_ACCOUNT
);
798 trace_io_uring_cqe_overflow(ctx
, user_data
, res
, cflags
, ocqe
);
801 * If we're in ring overflow flush mode, or in task cancel mode,
802 * or cannot allocate an overflow entry, then we need to drop it
805 io_account_cq_overflow(ctx
);
806 set_bit(IO_CHECK_CQ_DROPPED_BIT
, &ctx
->check_cq
);
809 if (list_empty(&ctx
->cq_overflow_list
)) {
810 set_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
);
811 atomic_or(IORING_SQ_CQ_OVERFLOW
, &ctx
->rings
->sq_flags
);
814 ocqe
->cqe
.user_data
= user_data
;
816 ocqe
->cqe
.flags
= cflags
;
818 ocqe
->cqe
.big_cqe
[0] = extra1
;
819 ocqe
->cqe
.big_cqe
[1] = extra2
;
821 list_add_tail(&ocqe
->list
, &ctx
->cq_overflow_list
);
825 void io_req_cqe_overflow(struct io_kiocb
*req
)
827 io_cqring_event_overflow(req
->ctx
, req
->cqe
.user_data
,
828 req
->cqe
.res
, req
->cqe
.flags
,
829 req
->big_cqe
.extra1
, req
->big_cqe
.extra2
);
830 memset(&req
->big_cqe
, 0, sizeof(req
->big_cqe
));
834 * writes to the cq entry need to come after reading head; the
835 * control dependency is enough as we're using WRITE_ONCE to
838 bool io_cqe_cache_refill(struct io_ring_ctx
*ctx
, bool overflow
)
840 struct io_rings
*rings
= ctx
->rings
;
841 unsigned int off
= ctx
->cached_cq_tail
& (ctx
->cq_entries
- 1);
842 unsigned int free
, queued
, len
;
845 * Posting into the CQ when there are pending overflowed CQEs may break
846 * ordering guarantees, which will affect links, F_MORE users and more.
847 * Force overflow the completion.
849 if (!overflow
&& (ctx
->check_cq
& BIT(IO_CHECK_CQ_OVERFLOW_BIT
)))
852 /* userspace may cheat modifying the tail, be safe and do min */
853 queued
= min(__io_cqring_events(ctx
), ctx
->cq_entries
);
854 free
= ctx
->cq_entries
- queued
;
855 /* we need a contiguous range, limit based on the current array offset */
856 len
= min(free
, ctx
->cq_entries
- off
);
860 if (ctx
->flags
& IORING_SETUP_CQE32
) {
865 ctx
->cqe_cached
= &rings
->cqes
[off
];
866 ctx
->cqe_sentinel
= ctx
->cqe_cached
+ len
;
870 static bool io_fill_cqe_aux(struct io_ring_ctx
*ctx
, u64 user_data
, s32 res
,
873 struct io_uring_cqe
*cqe
;
878 * If we can't get a cq entry, userspace overflowed the
879 * submission (by quite a lot). Increment the overflow count in
882 if (likely(io_get_cqe(ctx
, &cqe
))) {
883 trace_io_uring_complete(ctx
, NULL
, user_data
, res
, cflags
, 0, 0);
885 WRITE_ONCE(cqe
->user_data
, user_data
);
886 WRITE_ONCE(cqe
->res
, res
);
887 WRITE_ONCE(cqe
->flags
, cflags
);
889 if (ctx
->flags
& IORING_SETUP_CQE32
) {
890 WRITE_ONCE(cqe
->big_cqe
[0], 0);
891 WRITE_ONCE(cqe
->big_cqe
[1], 0);
898 static void __io_flush_post_cqes(struct io_ring_ctx
*ctx
)
899 __must_hold(&ctx
->uring_lock
)
901 struct io_submit_state
*state
= &ctx
->submit_state
;
904 lockdep_assert_held(&ctx
->uring_lock
);
905 for (i
= 0; i
< state
->cqes_count
; i
++) {
906 struct io_uring_cqe
*cqe
= &ctx
->completion_cqes
[i
];
908 if (!io_fill_cqe_aux(ctx
, cqe
->user_data
, cqe
->res
, cqe
->flags
)) {
909 if (ctx
->lockless_cq
) {
910 spin_lock(&ctx
->completion_lock
);
911 io_cqring_event_overflow(ctx
, cqe
->user_data
,
912 cqe
->res
, cqe
->flags
, 0, 0);
913 spin_unlock(&ctx
->completion_lock
);
915 io_cqring_event_overflow(ctx
, cqe
->user_data
,
916 cqe
->res
, cqe
->flags
, 0, 0);
920 state
->cqes_count
= 0;
923 static bool __io_post_aux_cqe(struct io_ring_ctx
*ctx
, u64 user_data
, s32 res
, u32 cflags
,
929 filled
= io_fill_cqe_aux(ctx
, user_data
, res
, cflags
);
930 if (!filled
&& allow_overflow
)
931 filled
= io_cqring_event_overflow(ctx
, user_data
, res
, cflags
, 0, 0);
933 io_cq_unlock_post(ctx
);
937 bool io_post_aux_cqe(struct io_ring_ctx
*ctx
, u64 user_data
, s32 res
, u32 cflags
)
939 return __io_post_aux_cqe(ctx
, user_data
, res
, cflags
, true);
943 * A helper for multishot requests posting additional CQEs.
944 * Should only be used from a task_work including IO_URING_F_MULTISHOT.
946 bool io_fill_cqe_req_aux(struct io_kiocb
*req
, bool defer
, s32 res
, u32 cflags
)
948 struct io_ring_ctx
*ctx
= req
->ctx
;
949 u64 user_data
= req
->cqe
.user_data
;
950 struct io_uring_cqe
*cqe
;
953 return __io_post_aux_cqe(ctx
, user_data
, res
, cflags
, false);
955 lockdep_assert_held(&ctx
->uring_lock
);
957 if (ctx
->submit_state
.cqes_count
== ARRAY_SIZE(ctx
->completion_cqes
)) {
959 __io_flush_post_cqes(ctx
);
960 /* no need to flush - flush is deferred */
961 __io_cq_unlock_post(ctx
);
964 /* For defered completions this is not as strict as it is otherwise,
965 * however it's main job is to prevent unbounded posted completions,
966 * and in that it works just as well.
968 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
))
971 cqe
= &ctx
->completion_cqes
[ctx
->submit_state
.cqes_count
++];
972 cqe
->user_data
= user_data
;
978 static void __io_req_complete_post(struct io_kiocb
*req
, unsigned issue_flags
)
980 struct io_ring_ctx
*ctx
= req
->ctx
;
981 struct io_rsrc_node
*rsrc_node
= NULL
;
984 if (!(req
->flags
& REQ_F_CQE_SKIP
)) {
985 if (!io_fill_cqe_req(ctx
, req
))
986 io_req_cqe_overflow(req
);
990 * If we're the last reference to this request, add to our locked
993 if (req_ref_put_and_test(req
)) {
994 if (req
->flags
& IO_REQ_LINK_FLAGS
) {
995 if (req
->flags
& IO_DISARM_MASK
)
998 io_req_task_queue(req
->link
);
1002 io_put_kbuf_comp(req
);
1003 if (unlikely(req
->flags
& IO_REQ_CLEAN_FLAGS
))
1007 rsrc_node
= req
->rsrc_node
;
1009 * Selected buffer deallocation in io_clean_op() assumes that
1010 * we don't hold ->completion_lock. Clean them here to avoid
1013 io_put_task_remote(req
->task
);
1014 wq_list_add_head(&req
->comp_list
, &ctx
->locked_free_list
);
1015 ctx
->locked_free_nr
++;
1017 io_cq_unlock_post(ctx
);
1020 io_ring_submit_lock(ctx
, issue_flags
);
1021 io_put_rsrc_node(ctx
, rsrc_node
);
1022 io_ring_submit_unlock(ctx
, issue_flags
);
1026 void io_req_complete_post(struct io_kiocb
*req
, unsigned issue_flags
)
1028 if (req
->ctx
->task_complete
&& req
->ctx
->submitter_task
!= current
) {
1029 req
->io_task_work
.func
= io_req_task_complete
;
1030 io_req_task_work_add(req
);
1031 } else if (!(issue_flags
& IO_URING_F_UNLOCKED
) ||
1032 !(req
->ctx
->flags
& IORING_SETUP_IOPOLL
)) {
1033 __io_req_complete_post(req
, issue_flags
);
1035 struct io_ring_ctx
*ctx
= req
->ctx
;
1037 mutex_lock(&ctx
->uring_lock
);
1038 __io_req_complete_post(req
, issue_flags
& ~IO_URING_F_UNLOCKED
);
1039 mutex_unlock(&ctx
->uring_lock
);
1043 void io_req_defer_failed(struct io_kiocb
*req
, s32 res
)
1044 __must_hold(&ctx
->uring_lock
)
1046 const struct io_cold_def
*def
= &io_cold_defs
[req
->opcode
];
1048 lockdep_assert_held(&req
->ctx
->uring_lock
);
1051 io_req_set_res(req
, res
, io_put_kbuf(req
, IO_URING_F_UNLOCKED
));
1054 io_req_complete_defer(req
);
1058 * Don't initialise the fields below on every allocation, but do that in
1059 * advance and keep them valid across allocations.
1061 static void io_preinit_req(struct io_kiocb
*req
, struct io_ring_ctx
*ctx
)
1065 req
->async_data
= NULL
;
1066 /* not necessary, but safer to zero */
1067 memset(&req
->cqe
, 0, sizeof(req
->cqe
));
1068 memset(&req
->big_cqe
, 0, sizeof(req
->big_cqe
));
1071 static void io_flush_cached_locked_reqs(struct io_ring_ctx
*ctx
,
1072 struct io_submit_state
*state
)
1074 spin_lock(&ctx
->completion_lock
);
1075 wq_list_splice(&ctx
->locked_free_list
, &state
->free_list
);
1076 ctx
->locked_free_nr
= 0;
1077 spin_unlock(&ctx
->completion_lock
);
1081 * A request might get retired back into the request caches even before opcode
1082 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1083 * Because of that, io_alloc_req() should be called only under ->uring_lock
1084 * and with extra caution to not get a request that is still worked on.
1086 __cold
bool __io_alloc_req_refill(struct io_ring_ctx
*ctx
)
1087 __must_hold(&ctx
->uring_lock
)
1089 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
1090 void *reqs
[IO_REQ_ALLOC_BATCH
];
1094 * If we have more than a batch's worth of requests in our IRQ side
1095 * locked cache, grab the lock and move them over to our submission
1098 if (data_race(ctx
->locked_free_nr
) > IO_COMPL_BATCH
) {
1099 io_flush_cached_locked_reqs(ctx
, &ctx
->submit_state
);
1100 if (!io_req_cache_empty(ctx
))
1104 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, ARRAY_SIZE(reqs
), reqs
);
1107 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1108 * retry single alloc to be on the safe side.
1110 if (unlikely(ret
<= 0)) {
1111 reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
1117 percpu_ref_get_many(&ctx
->refs
, ret
);
1118 for (i
= 0; i
< ret
; i
++) {
1119 struct io_kiocb
*req
= reqs
[i
];
1121 io_preinit_req(req
, ctx
);
1122 io_req_add_to_cache(req
, ctx
);
1127 __cold
void io_free_req(struct io_kiocb
*req
)
1129 /* refs were already put, restore them for io_req_task_complete() */
1130 req
->flags
&= ~REQ_F_REFCOUNT
;
1131 /* we only want to free it, don't post CQEs */
1132 req
->flags
|= REQ_F_CQE_SKIP
;
1133 req
->io_task_work
.func
= io_req_task_complete
;
1134 io_req_task_work_add(req
);
1137 static void __io_req_find_next_prep(struct io_kiocb
*req
)
1139 struct io_ring_ctx
*ctx
= req
->ctx
;
1141 spin_lock(&ctx
->completion_lock
);
1142 io_disarm_next(req
);
1143 spin_unlock(&ctx
->completion_lock
);
1146 static inline struct io_kiocb
*io_req_find_next(struct io_kiocb
*req
)
1148 struct io_kiocb
*nxt
;
1151 * If LINK is set, we have dependent requests in this chain. If we
1152 * didn't fail this request, queue the first one up, moving any other
1153 * dependencies to the next request. In case of failure, fail the rest
1156 if (unlikely(req
->flags
& IO_DISARM_MASK
))
1157 __io_req_find_next_prep(req
);
1163 static void ctx_flush_and_put(struct io_ring_ctx
*ctx
, struct io_tw_state
*ts
)
1167 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
)
1168 atomic_andnot(IORING_SQ_TASKRUN
, &ctx
->rings
->sq_flags
);
1170 io_submit_flush_completions(ctx
);
1171 mutex_unlock(&ctx
->uring_lock
);
1174 percpu_ref_put(&ctx
->refs
);
1177 static unsigned int handle_tw_list(struct llist_node
*node
,
1178 struct io_ring_ctx
**ctx
,
1179 struct io_tw_state
*ts
,
1180 struct llist_node
*last
)
1182 unsigned int count
= 0;
1184 while (node
&& node
!= last
) {
1185 struct llist_node
*next
= node
->next
;
1186 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1189 prefetch(container_of(next
, struct io_kiocb
, io_task_work
.node
));
1191 if (req
->ctx
!= *ctx
) {
1192 ctx_flush_and_put(*ctx
, ts
);
1194 /* if not contended, grab and improve batching */
1195 ts
->locked
= mutex_trylock(&(*ctx
)->uring_lock
);
1196 percpu_ref_get(&(*ctx
)->refs
);
1198 INDIRECT_CALL_2(req
->io_task_work
.func
,
1199 io_poll_task_func
, io_req_rw_complete
,
1203 if (unlikely(need_resched())) {
1204 ctx_flush_and_put(*ctx
, ts
);
1214 * io_llist_xchg - swap all entries in a lock-less list
1215 * @head: the head of lock-less list to delete all entries
1216 * @new: new entry as the head of the list
1218 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1219 * The order of entries returned is from the newest to the oldest added one.
1221 static inline struct llist_node
*io_llist_xchg(struct llist_head
*head
,
1222 struct llist_node
*new)
1224 return xchg(&head
->first
, new);
1228 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1229 * @head: the head of lock-less list to delete all entries
1230 * @old: expected old value of the first entry of the list
1231 * @new: new entry as the head of the list
1233 * perform a cmpxchg on the first entry of the list.
1236 static inline struct llist_node
*io_llist_cmpxchg(struct llist_head
*head
,
1237 struct llist_node
*old
,
1238 struct llist_node
*new)
1240 return cmpxchg(&head
->first
, old
, new);
1243 static __cold
void io_fallback_tw(struct io_uring_task
*tctx
, bool sync
)
1245 struct llist_node
*node
= llist_del_all(&tctx
->task_list
);
1246 struct io_ring_ctx
*last_ctx
= NULL
;
1247 struct io_kiocb
*req
;
1250 req
= container_of(node
, struct io_kiocb
, io_task_work
.node
);
1252 if (sync
&& last_ctx
!= req
->ctx
) {
1254 flush_delayed_work(&last_ctx
->fallback_work
);
1255 percpu_ref_put(&last_ctx
->refs
);
1257 last_ctx
= req
->ctx
;
1258 percpu_ref_get(&last_ctx
->refs
);
1260 if (llist_add(&req
->io_task_work
.node
,
1261 &req
->ctx
->fallback_llist
))
1262 schedule_delayed_work(&req
->ctx
->fallback_work
, 1);
1266 flush_delayed_work(&last_ctx
->fallback_work
);
1267 percpu_ref_put(&last_ctx
->refs
);
1271 void tctx_task_work(struct callback_head
*cb
)
1273 struct io_tw_state ts
= {};
1274 struct io_ring_ctx
*ctx
= NULL
;
1275 struct io_uring_task
*tctx
= container_of(cb
, struct io_uring_task
,
1277 struct llist_node fake
= {};
1278 struct llist_node
*node
;
1279 unsigned int loops
= 0;
1280 unsigned int count
= 0;
1282 if (unlikely(current
->flags
& PF_EXITING
)) {
1283 io_fallback_tw(tctx
, true);
1289 node
= io_llist_xchg(&tctx
->task_list
, &fake
);
1290 count
+= handle_tw_list(node
, &ctx
, &ts
, &fake
);
1292 /* skip expensive cmpxchg if there are items in the list */
1293 if (READ_ONCE(tctx
->task_list
.first
) != &fake
)
1295 if (ts
.locked
&& !wq_list_empty(&ctx
->submit_state
.compl_reqs
)) {
1296 io_submit_flush_completions(ctx
);
1297 if (READ_ONCE(tctx
->task_list
.first
) != &fake
)
1300 node
= io_llist_cmpxchg(&tctx
->task_list
, &fake
, NULL
);
1301 } while (node
!= &fake
);
1303 ctx_flush_and_put(ctx
, &ts
);
1305 /* relaxed read is enough as only the task itself sets ->in_cancel */
1306 if (unlikely(atomic_read(&tctx
->in_cancel
)))
1307 io_uring_drop_tctx_refs(current
);
1309 trace_io_uring_task_work_run(tctx
, count
, loops
);
1312 static inline void io_req_local_work_add(struct io_kiocb
*req
, unsigned flags
)
1314 struct io_ring_ctx
*ctx
= req
->ctx
;
1315 unsigned nr_wait
, nr_tw
, nr_tw_prev
;
1316 struct llist_node
*head
;
1318 /* See comment above IO_CQ_WAKE_INIT */
1319 BUILD_BUG_ON(IO_CQ_WAKE_FORCE
<= IORING_MAX_CQ_ENTRIES
);
1322 * We don't know how many reuqests is there in the link and whether
1323 * they can even be queued lazily, fall back to non-lazy.
1325 if (req
->flags
& (REQ_F_LINK
| REQ_F_HARDLINK
))
1326 flags
&= ~IOU_F_TWQ_LAZY_WAKE
;
1328 head
= READ_ONCE(ctx
->work_llist
.first
);
1332 struct io_kiocb
*first_req
= container_of(head
,
1336 * Might be executed at any moment, rely on
1337 * SLAB_TYPESAFE_BY_RCU to keep it alive.
1339 nr_tw_prev
= READ_ONCE(first_req
->nr_tw
);
1343 * Theoretically, it can overflow, but that's fine as one of
1344 * previous adds should've tried to wake the task.
1346 nr_tw
= nr_tw_prev
+ 1;
1347 if (!(flags
& IOU_F_TWQ_LAZY_WAKE
))
1348 nr_tw
= IO_CQ_WAKE_FORCE
;
1351 req
->io_task_work
.node
.next
= head
;
1352 } while (!try_cmpxchg(&ctx
->work_llist
.first
, &head
,
1353 &req
->io_task_work
.node
));
1356 * cmpxchg implies a full barrier, which pairs with the barrier
1357 * in set_current_state() on the io_cqring_wait() side. It's used
1358 * to ensure that either we see updated ->cq_wait_nr, or waiters
1359 * going to sleep will observe the work added to the list, which
1360 * is similar to the wait/wawke task state sync.
1364 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
)
1365 atomic_or(IORING_SQ_TASKRUN
, &ctx
->rings
->sq_flags
);
1367 io_eventfd_signal(ctx
);
1370 nr_wait
= atomic_read(&ctx
->cq_wait_nr
);
1371 /* not enough or no one is waiting */
1372 if (nr_tw
< nr_wait
)
1374 /* the previous add has already woken it up */
1375 if (nr_tw_prev
>= nr_wait
)
1377 wake_up_state(ctx
->submitter_task
, TASK_INTERRUPTIBLE
);
1380 static void io_req_normal_work_add(struct io_kiocb
*req
)
1382 struct io_uring_task
*tctx
= req
->task
->io_uring
;
1383 struct io_ring_ctx
*ctx
= req
->ctx
;
1385 /* task_work already pending, we're done */
1386 if (!llist_add(&req
->io_task_work
.node
, &tctx
->task_list
))
1389 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
)
1390 atomic_or(IORING_SQ_TASKRUN
, &ctx
->rings
->sq_flags
);
1392 if (likely(!task_work_add(req
->task
, &tctx
->task_work
, ctx
->notify_method
)))
1395 io_fallback_tw(tctx
, false);
1398 void __io_req_task_work_add(struct io_kiocb
*req
, unsigned flags
)
1400 if (req
->ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) {
1402 io_req_local_work_add(req
, flags
);
1405 io_req_normal_work_add(req
);
1409 static void __cold
io_move_task_work_from_local(struct io_ring_ctx
*ctx
)
1411 struct llist_node
*node
;
1413 node
= llist_del_all(&ctx
->work_llist
);
1415 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1419 io_req_normal_work_add(req
);
1423 static int __io_run_local_work(struct io_ring_ctx
*ctx
, struct io_tw_state
*ts
)
1425 struct llist_node
*node
;
1426 unsigned int loops
= 0;
1429 if (WARN_ON_ONCE(ctx
->submitter_task
!= current
))
1431 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
)
1432 atomic_andnot(IORING_SQ_TASKRUN
, &ctx
->rings
->sq_flags
);
1435 * llists are in reverse order, flip it back the right way before
1436 * running the pending items.
1438 node
= llist_reverse_order(io_llist_xchg(&ctx
->work_llist
, NULL
));
1440 struct llist_node
*next
= node
->next
;
1441 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1443 prefetch(container_of(next
, struct io_kiocb
, io_task_work
.node
));
1444 INDIRECT_CALL_2(req
->io_task_work
.func
,
1445 io_poll_task_func
, io_req_rw_complete
,
1452 if (!llist_empty(&ctx
->work_llist
))
1455 io_submit_flush_completions(ctx
);
1456 if (!llist_empty(&ctx
->work_llist
))
1459 trace_io_uring_local_work_run(ctx
, ret
, loops
);
1463 static inline int io_run_local_work_locked(struct io_ring_ctx
*ctx
)
1465 struct io_tw_state ts
= { .locked
= true, };
1468 if (llist_empty(&ctx
->work_llist
))
1471 ret
= __io_run_local_work(ctx
, &ts
);
1472 /* shouldn't happen! */
1473 if (WARN_ON_ONCE(!ts
.locked
))
1474 mutex_lock(&ctx
->uring_lock
);
1478 static int io_run_local_work(struct io_ring_ctx
*ctx
)
1480 struct io_tw_state ts
= {};
1483 ts
.locked
= mutex_trylock(&ctx
->uring_lock
);
1484 ret
= __io_run_local_work(ctx
, &ts
);
1486 mutex_unlock(&ctx
->uring_lock
);
1491 static void io_req_task_cancel(struct io_kiocb
*req
, struct io_tw_state
*ts
)
1493 io_tw_lock(req
->ctx
, ts
);
1494 io_req_defer_failed(req
, req
->cqe
.res
);
1497 void io_req_task_submit(struct io_kiocb
*req
, struct io_tw_state
*ts
)
1499 io_tw_lock(req
->ctx
, ts
);
1500 /* req->task == current here, checking PF_EXITING is safe */
1501 if (unlikely(req
->task
->flags
& PF_EXITING
))
1502 io_req_defer_failed(req
, -EFAULT
);
1503 else if (req
->flags
& REQ_F_FORCE_ASYNC
)
1504 io_queue_iowq(req
, ts
);
1509 void io_req_task_queue_fail(struct io_kiocb
*req
, int ret
)
1511 io_req_set_res(req
, ret
, 0);
1512 req
->io_task_work
.func
= io_req_task_cancel
;
1513 io_req_task_work_add(req
);
1516 void io_req_task_queue(struct io_kiocb
*req
)
1518 req
->io_task_work
.func
= io_req_task_submit
;
1519 io_req_task_work_add(req
);
1522 void io_queue_next(struct io_kiocb
*req
)
1524 struct io_kiocb
*nxt
= io_req_find_next(req
);
1527 io_req_task_queue(nxt
);
1530 static void io_free_batch_list(struct io_ring_ctx
*ctx
,
1531 struct io_wq_work_node
*node
)
1532 __must_hold(&ctx
->uring_lock
)
1535 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1538 if (unlikely(req
->flags
& IO_REQ_CLEAN_SLOW_FLAGS
)) {
1539 if (req
->flags
& REQ_F_REFCOUNT
) {
1540 node
= req
->comp_list
.next
;
1541 if (!req_ref_put_and_test(req
))
1544 if ((req
->flags
& REQ_F_POLLED
) && req
->apoll
) {
1545 struct async_poll
*apoll
= req
->apoll
;
1547 if (apoll
->double_poll
)
1548 kfree(apoll
->double_poll
);
1549 if (!io_alloc_cache_put(&ctx
->apoll_cache
, &apoll
->cache
))
1551 req
->flags
&= ~REQ_F_POLLED
;
1553 if (req
->flags
& IO_REQ_LINK_FLAGS
)
1555 if (unlikely(req
->flags
& IO_REQ_CLEAN_FLAGS
))
1560 io_req_put_rsrc_locked(req
, ctx
);
1562 io_put_task(req
->task
);
1563 node
= req
->comp_list
.next
;
1564 io_req_add_to_cache(req
, ctx
);
1568 void __io_submit_flush_completions(struct io_ring_ctx
*ctx
)
1569 __must_hold(&ctx
->uring_lock
)
1571 struct io_submit_state
*state
= &ctx
->submit_state
;
1572 struct io_wq_work_node
*node
;
1575 /* must come first to preserve CQE ordering in failure cases */
1576 if (state
->cqes_count
)
1577 __io_flush_post_cqes(ctx
);
1578 __wq_list_for_each(node
, &state
->compl_reqs
) {
1579 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1582 if (!(req
->flags
& REQ_F_CQE_SKIP
) &&
1583 unlikely(!io_fill_cqe_req(ctx
, req
))) {
1584 if (ctx
->lockless_cq
) {
1585 spin_lock(&ctx
->completion_lock
);
1586 io_req_cqe_overflow(req
);
1587 spin_unlock(&ctx
->completion_lock
);
1589 io_req_cqe_overflow(req
);
1593 __io_cq_unlock_post(ctx
);
1595 if (!wq_list_empty(&ctx
->submit_state
.compl_reqs
)) {
1596 io_free_batch_list(ctx
, state
->compl_reqs
.first
);
1597 INIT_WQ_LIST(&state
->compl_reqs
);
1601 static unsigned io_cqring_events(struct io_ring_ctx
*ctx
)
1603 /* See comment at the top of this file */
1605 return __io_cqring_events(ctx
);
1609 * We can't just wait for polled events to come to us, we have to actively
1610 * find and complete them.
1612 static __cold
void io_iopoll_try_reap_events(struct io_ring_ctx
*ctx
)
1614 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
1617 mutex_lock(&ctx
->uring_lock
);
1618 while (!wq_list_empty(&ctx
->iopoll_list
)) {
1619 /* let it sleep and repeat later if can't complete a request */
1620 if (io_do_iopoll(ctx
, true) == 0)
1623 * Ensure we allow local-to-the-cpu processing to take place,
1624 * in this case we need to ensure that we reap all events.
1625 * Also let task_work, etc. to progress by releasing the mutex
1627 if (need_resched()) {
1628 mutex_unlock(&ctx
->uring_lock
);
1630 mutex_lock(&ctx
->uring_lock
);
1633 mutex_unlock(&ctx
->uring_lock
);
1636 static int io_iopoll_check(struct io_ring_ctx
*ctx
, long min
)
1638 unsigned int nr_events
= 0;
1639 unsigned long check_cq
;
1641 if (!io_allowed_run_tw(ctx
))
1644 check_cq
= READ_ONCE(ctx
->check_cq
);
1645 if (unlikely(check_cq
)) {
1646 if (check_cq
& BIT(IO_CHECK_CQ_OVERFLOW_BIT
))
1647 __io_cqring_overflow_flush(ctx
);
1649 * Similarly do not spin if we have not informed the user of any
1652 if (check_cq
& BIT(IO_CHECK_CQ_DROPPED_BIT
))
1656 * Don't enter poll loop if we already have events pending.
1657 * If we do, we can potentially be spinning for commands that
1658 * already triggered a CQE (eg in error).
1660 if (io_cqring_events(ctx
))
1667 * If a submit got punted to a workqueue, we can have the
1668 * application entering polling for a command before it gets
1669 * issued. That app will hold the uring_lock for the duration
1670 * of the poll right here, so we need to take a breather every
1671 * now and then to ensure that the issue has a chance to add
1672 * the poll to the issued list. Otherwise we can spin here
1673 * forever, while the workqueue is stuck trying to acquire the
1676 if (wq_list_empty(&ctx
->iopoll_list
) ||
1677 io_task_work_pending(ctx
)) {
1678 u32 tail
= ctx
->cached_cq_tail
;
1680 (void) io_run_local_work_locked(ctx
);
1682 if (task_work_pending(current
) ||
1683 wq_list_empty(&ctx
->iopoll_list
)) {
1684 mutex_unlock(&ctx
->uring_lock
);
1686 mutex_lock(&ctx
->uring_lock
);
1688 /* some requests don't go through iopoll_list */
1689 if (tail
!= ctx
->cached_cq_tail
||
1690 wq_list_empty(&ctx
->iopoll_list
))
1693 ret
= io_do_iopoll(ctx
, !min
);
1694 if (unlikely(ret
< 0))
1697 if (task_sigpending(current
))
1703 } while (nr_events
< min
);
1708 void io_req_task_complete(struct io_kiocb
*req
, struct io_tw_state
*ts
)
1711 io_req_complete_defer(req
);
1713 io_req_complete_post(req
, IO_URING_F_UNLOCKED
);
1717 * After the iocb has been issued, it's safe to be found on the poll list.
1718 * Adding the kiocb to the list AFTER submission ensures that we don't
1719 * find it from a io_do_iopoll() thread before the issuer is done
1720 * accessing the kiocb cookie.
1722 static void io_iopoll_req_issued(struct io_kiocb
*req
, unsigned int issue_flags
)
1724 struct io_ring_ctx
*ctx
= req
->ctx
;
1725 const bool needs_lock
= issue_flags
& IO_URING_F_UNLOCKED
;
1727 /* workqueue context doesn't hold uring_lock, grab it now */
1728 if (unlikely(needs_lock
))
1729 mutex_lock(&ctx
->uring_lock
);
1732 * Track whether we have multiple files in our lists. This will impact
1733 * how we do polling eventually, not spinning if we're on potentially
1734 * different devices.
1736 if (wq_list_empty(&ctx
->iopoll_list
)) {
1737 ctx
->poll_multi_queue
= false;
1738 } else if (!ctx
->poll_multi_queue
) {
1739 struct io_kiocb
*list_req
;
1741 list_req
= container_of(ctx
->iopoll_list
.first
, struct io_kiocb
,
1743 if (list_req
->file
!= req
->file
)
1744 ctx
->poll_multi_queue
= true;
1748 * For fast devices, IO may have already completed. If it has, add
1749 * it to the front so we find it first.
1751 if (READ_ONCE(req
->iopoll_completed
))
1752 wq_list_add_head(&req
->comp_list
, &ctx
->iopoll_list
);
1754 wq_list_add_tail(&req
->comp_list
, &ctx
->iopoll_list
);
1756 if (unlikely(needs_lock
)) {
1758 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1759 * in sq thread task context or in io worker task context. If
1760 * current task context is sq thread, we don't need to check
1761 * whether should wake up sq thread.
1763 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) &&
1764 wq_has_sleeper(&ctx
->sq_data
->wait
))
1765 wake_up(&ctx
->sq_data
->wait
);
1767 mutex_unlock(&ctx
->uring_lock
);
1771 unsigned int io_file_get_flags(struct file
*file
)
1773 unsigned int res
= 0;
1775 if (S_ISREG(file_inode(file
)->i_mode
))
1777 if ((file
->f_flags
& O_NONBLOCK
) || (file
->f_mode
& FMODE_NOWAIT
))
1778 res
|= REQ_F_SUPPORT_NOWAIT
;
1782 bool io_alloc_async_data(struct io_kiocb
*req
)
1784 WARN_ON_ONCE(!io_cold_defs
[req
->opcode
].async_size
);
1785 req
->async_data
= kmalloc(io_cold_defs
[req
->opcode
].async_size
, GFP_KERNEL
);
1786 if (req
->async_data
) {
1787 req
->flags
|= REQ_F_ASYNC_DATA
;
1793 int io_req_prep_async(struct io_kiocb
*req
)
1795 const struct io_cold_def
*cdef
= &io_cold_defs
[req
->opcode
];
1796 const struct io_issue_def
*def
= &io_issue_defs
[req
->opcode
];
1798 /* assign early for deferred execution for non-fixed file */
1799 if (def
->needs_file
&& !(req
->flags
& REQ_F_FIXED_FILE
) && !req
->file
)
1800 req
->file
= io_file_get_normal(req
, req
->cqe
.fd
);
1801 if (!cdef
->prep_async
)
1803 if (WARN_ON_ONCE(req_has_async_data(req
)))
1805 if (!def
->manual_alloc
) {
1806 if (io_alloc_async_data(req
))
1809 return cdef
->prep_async(req
);
1812 static u32
io_get_sequence(struct io_kiocb
*req
)
1814 u32 seq
= req
->ctx
->cached_sq_head
;
1815 struct io_kiocb
*cur
;
1817 /* need original cached_sq_head, but it was increased for each req */
1818 io_for_each_link(cur
, req
)
1823 static __cold
void io_drain_req(struct io_kiocb
*req
)
1824 __must_hold(&ctx
->uring_lock
)
1826 struct io_ring_ctx
*ctx
= req
->ctx
;
1827 struct io_defer_entry
*de
;
1829 u32 seq
= io_get_sequence(req
);
1831 /* Still need defer if there is pending req in defer list. */
1832 spin_lock(&ctx
->completion_lock
);
1833 if (!req_need_defer(req
, seq
) && list_empty_careful(&ctx
->defer_list
)) {
1834 spin_unlock(&ctx
->completion_lock
);
1836 ctx
->drain_active
= false;
1837 io_req_task_queue(req
);
1840 spin_unlock(&ctx
->completion_lock
);
1842 io_prep_async_link(req
);
1843 de
= kmalloc(sizeof(*de
), GFP_KERNEL
);
1846 io_req_defer_failed(req
, ret
);
1850 spin_lock(&ctx
->completion_lock
);
1851 if (!req_need_defer(req
, seq
) && list_empty(&ctx
->defer_list
)) {
1852 spin_unlock(&ctx
->completion_lock
);
1857 trace_io_uring_defer(req
);
1860 list_add_tail(&de
->list
, &ctx
->defer_list
);
1861 spin_unlock(&ctx
->completion_lock
);
1864 static bool io_assign_file(struct io_kiocb
*req
, const struct io_issue_def
*def
,
1865 unsigned int issue_flags
)
1867 if (req
->file
|| !def
->needs_file
)
1870 if (req
->flags
& REQ_F_FIXED_FILE
)
1871 req
->file
= io_file_get_fixed(req
, req
->cqe
.fd
, issue_flags
);
1873 req
->file
= io_file_get_normal(req
, req
->cqe
.fd
);
1878 static int io_issue_sqe(struct io_kiocb
*req
, unsigned int issue_flags
)
1880 const struct io_issue_def
*def
= &io_issue_defs
[req
->opcode
];
1881 const struct cred
*creds
= NULL
;
1884 if (unlikely(!io_assign_file(req
, def
, issue_flags
)))
1887 if (unlikely((req
->flags
& REQ_F_CREDS
) && req
->creds
!= current_cred()))
1888 creds
= override_creds(req
->creds
);
1890 if (!def
->audit_skip
)
1891 audit_uring_entry(req
->opcode
);
1893 ret
= def
->issue(req
, issue_flags
);
1895 if (!def
->audit_skip
)
1896 audit_uring_exit(!ret
, ret
);
1899 revert_creds(creds
);
1901 if (ret
== IOU_OK
) {
1902 if (issue_flags
& IO_URING_F_COMPLETE_DEFER
)
1903 io_req_complete_defer(req
);
1905 io_req_complete_post(req
, issue_flags
);
1910 if (ret
== IOU_ISSUE_SKIP_COMPLETE
) {
1912 io_arm_ltimeout(req
);
1914 /* If the op doesn't have a file, we're not polling for it */
1915 if ((req
->ctx
->flags
& IORING_SETUP_IOPOLL
) && def
->iopoll_queue
)
1916 io_iopoll_req_issued(req
, issue_flags
);
1921 int io_poll_issue(struct io_kiocb
*req
, struct io_tw_state
*ts
)
1923 io_tw_lock(req
->ctx
, ts
);
1924 return io_issue_sqe(req
, IO_URING_F_NONBLOCK
|IO_URING_F_MULTISHOT
|
1925 IO_URING_F_COMPLETE_DEFER
);
1928 struct io_wq_work
*io_wq_free_work(struct io_wq_work
*work
)
1930 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
1931 struct io_kiocb
*nxt
= NULL
;
1933 if (req_ref_put_and_test(req
)) {
1934 if (req
->flags
& IO_REQ_LINK_FLAGS
)
1935 nxt
= io_req_find_next(req
);
1938 return nxt
? &nxt
->work
: NULL
;
1941 void io_wq_submit_work(struct io_wq_work
*work
)
1943 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
1944 const struct io_issue_def
*def
= &io_issue_defs
[req
->opcode
];
1945 unsigned int issue_flags
= IO_URING_F_UNLOCKED
| IO_URING_F_IOWQ
;
1946 bool needs_poll
= false;
1947 int ret
= 0, err
= -ECANCELED
;
1949 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1950 if (!(req
->flags
& REQ_F_REFCOUNT
))
1951 __io_req_set_refcount(req
, 2);
1955 io_arm_ltimeout(req
);
1957 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1958 if (work
->flags
& IO_WQ_WORK_CANCEL
) {
1960 io_req_task_queue_fail(req
, err
);
1963 if (!io_assign_file(req
, def
, issue_flags
)) {
1965 work
->flags
|= IO_WQ_WORK_CANCEL
;
1969 if (req
->flags
& REQ_F_FORCE_ASYNC
) {
1970 bool opcode_poll
= def
->pollin
|| def
->pollout
;
1972 if (opcode_poll
&& file_can_poll(req
->file
)) {
1974 issue_flags
|= IO_URING_F_NONBLOCK
;
1979 ret
= io_issue_sqe(req
, issue_flags
);
1984 * If REQ_F_NOWAIT is set, then don't wait or retry with
1985 * poll. -EAGAIN is final for that case.
1987 if (req
->flags
& REQ_F_NOWAIT
)
1991 * We can get EAGAIN for iopolled IO even though we're
1992 * forcing a sync submission from here, since we can't
1993 * wait for request slots on the block side.
1996 if (!(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
1998 if (io_wq_worker_stopped())
2004 if (io_arm_poll_handler(req
, issue_flags
) == IO_APOLL_OK
)
2006 /* aborted or ready, in either case retry blocking */
2008 issue_flags
&= ~IO_URING_F_NONBLOCK
;
2011 /* avoid locking problems by failing it from a clean context */
2013 io_req_task_queue_fail(req
, ret
);
2016 inline struct file
*io_file_get_fixed(struct io_kiocb
*req
, int fd
,
2017 unsigned int issue_flags
)
2019 struct io_ring_ctx
*ctx
= req
->ctx
;
2020 struct io_fixed_file
*slot
;
2021 struct file
*file
= NULL
;
2023 io_ring_submit_lock(ctx
, issue_flags
);
2025 if (unlikely((unsigned int)fd
>= ctx
->nr_user_files
))
2027 fd
= array_index_nospec(fd
, ctx
->nr_user_files
);
2028 slot
= io_fixed_file_slot(&ctx
->file_table
, fd
);
2029 if (!req
->rsrc_node
)
2030 __io_req_set_rsrc_node(req
, ctx
);
2031 req
->flags
|= io_slot_flags(slot
);
2032 file
= io_slot_file(slot
);
2034 io_ring_submit_unlock(ctx
, issue_flags
);
2038 struct file
*io_file_get_normal(struct io_kiocb
*req
, int fd
)
2040 struct file
*file
= fget(fd
);
2042 trace_io_uring_file_get(req
, fd
);
2044 /* we don't allow fixed io_uring files */
2045 if (file
&& io_is_uring_fops(file
))
2046 io_req_track_inflight(req
);
2050 static void io_queue_async(struct io_kiocb
*req
, int ret
)
2051 __must_hold(&req
->ctx
->uring_lock
)
2053 struct io_kiocb
*linked_timeout
;
2055 if (ret
!= -EAGAIN
|| (req
->flags
& REQ_F_NOWAIT
)) {
2056 io_req_defer_failed(req
, ret
);
2060 linked_timeout
= io_prep_linked_timeout(req
);
2062 switch (io_arm_poll_handler(req
, 0)) {
2063 case IO_APOLL_READY
:
2064 io_kbuf_recycle(req
, 0);
2065 io_req_task_queue(req
);
2067 case IO_APOLL_ABORTED
:
2068 io_kbuf_recycle(req
, 0);
2069 io_queue_iowq(req
, NULL
);
2076 io_queue_linked_timeout(linked_timeout
);
2079 static inline void io_queue_sqe(struct io_kiocb
*req
)
2080 __must_hold(&req
->ctx
->uring_lock
)
2084 ret
= io_issue_sqe(req
, IO_URING_F_NONBLOCK
|IO_URING_F_COMPLETE_DEFER
);
2087 * We async punt it if the file wasn't marked NOWAIT, or if the file
2088 * doesn't support non-blocking read/write attempts
2091 io_queue_async(req
, ret
);
2094 static void io_queue_sqe_fallback(struct io_kiocb
*req
)
2095 __must_hold(&req
->ctx
->uring_lock
)
2097 if (unlikely(req
->flags
& REQ_F_FAIL
)) {
2099 * We don't submit, fail them all, for that replace hardlinks
2100 * with normal links. Extra REQ_F_LINK is tolerated.
2102 req
->flags
&= ~REQ_F_HARDLINK
;
2103 req
->flags
|= REQ_F_LINK
;
2104 io_req_defer_failed(req
, req
->cqe
.res
);
2106 int ret
= io_req_prep_async(req
);
2108 if (unlikely(ret
)) {
2109 io_req_defer_failed(req
, ret
);
2113 if (unlikely(req
->ctx
->drain_active
))
2116 io_queue_iowq(req
, NULL
);
2121 * Check SQE restrictions (opcode and flags).
2123 * Returns 'true' if SQE is allowed, 'false' otherwise.
2125 static inline bool io_check_restriction(struct io_ring_ctx
*ctx
,
2126 struct io_kiocb
*req
,
2127 unsigned int sqe_flags
)
2129 if (!test_bit(req
->opcode
, ctx
->restrictions
.sqe_op
))
2132 if ((sqe_flags
& ctx
->restrictions
.sqe_flags_required
) !=
2133 ctx
->restrictions
.sqe_flags_required
)
2136 if (sqe_flags
& ~(ctx
->restrictions
.sqe_flags_allowed
|
2137 ctx
->restrictions
.sqe_flags_required
))
2143 static void io_init_req_drain(struct io_kiocb
*req
)
2145 struct io_ring_ctx
*ctx
= req
->ctx
;
2146 struct io_kiocb
*head
= ctx
->submit_state
.link
.head
;
2148 ctx
->drain_active
= true;
2151 * If we need to drain a request in the middle of a link, drain
2152 * the head request and the next request/link after the current
2153 * link. Considering sequential execution of links,
2154 * REQ_F_IO_DRAIN will be maintained for every request of our
2157 head
->flags
|= REQ_F_IO_DRAIN
| REQ_F_FORCE_ASYNC
;
2158 ctx
->drain_next
= true;
2162 static int io_init_req(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
2163 const struct io_uring_sqe
*sqe
)
2164 __must_hold(&ctx
->uring_lock
)
2166 const struct io_issue_def
*def
;
2167 unsigned int sqe_flags
;
2171 /* req is partially pre-initialised, see io_preinit_req() */
2172 req
->opcode
= opcode
= READ_ONCE(sqe
->opcode
);
2173 /* same numerical values with corresponding REQ_F_*, safe to copy */
2174 req
->flags
= sqe_flags
= READ_ONCE(sqe
->flags
);
2175 req
->cqe
.user_data
= READ_ONCE(sqe
->user_data
);
2177 req
->rsrc_node
= NULL
;
2178 req
->task
= current
;
2180 if (unlikely(opcode
>= IORING_OP_LAST
)) {
2184 def
= &io_issue_defs
[opcode
];
2185 if (unlikely(sqe_flags
& ~SQE_COMMON_FLAGS
)) {
2186 /* enforce forwards compatibility on users */
2187 if (sqe_flags
& ~SQE_VALID_FLAGS
)
2189 if (sqe_flags
& IOSQE_BUFFER_SELECT
) {
2190 if (!def
->buffer_select
)
2192 req
->buf_index
= READ_ONCE(sqe
->buf_group
);
2194 if (sqe_flags
& IOSQE_CQE_SKIP_SUCCESS
)
2195 ctx
->drain_disabled
= true;
2196 if (sqe_flags
& IOSQE_IO_DRAIN
) {
2197 if (ctx
->drain_disabled
)
2199 io_init_req_drain(req
);
2202 if (unlikely(ctx
->restricted
|| ctx
->drain_active
|| ctx
->drain_next
)) {
2203 if (ctx
->restricted
&& !io_check_restriction(ctx
, req
, sqe_flags
))
2205 /* knock it to the slow queue path, will be drained there */
2206 if (ctx
->drain_active
)
2207 req
->flags
|= REQ_F_FORCE_ASYNC
;
2208 /* if there is no link, we're at "next" request and need to drain */
2209 if (unlikely(ctx
->drain_next
) && !ctx
->submit_state
.link
.head
) {
2210 ctx
->drain_next
= false;
2211 ctx
->drain_active
= true;
2212 req
->flags
|= REQ_F_IO_DRAIN
| REQ_F_FORCE_ASYNC
;
2216 if (!def
->ioprio
&& sqe
->ioprio
)
2218 if (!def
->iopoll
&& (ctx
->flags
& IORING_SETUP_IOPOLL
))
2221 if (def
->needs_file
) {
2222 struct io_submit_state
*state
= &ctx
->submit_state
;
2224 req
->cqe
.fd
= READ_ONCE(sqe
->fd
);
2227 * Plug now if we have more than 2 IO left after this, and the
2228 * target is potentially a read/write to block based storage.
2230 if (state
->need_plug
&& def
->plug
) {
2231 state
->plug_started
= true;
2232 state
->need_plug
= false;
2233 blk_start_plug_nr_ios(&state
->plug
, state
->submit_nr
);
2237 personality
= READ_ONCE(sqe
->personality
);
2241 req
->creds
= xa_load(&ctx
->personalities
, personality
);
2244 get_cred(req
->creds
);
2245 ret
= security_uring_override_creds(req
->creds
);
2247 put_cred(req
->creds
);
2250 req
->flags
|= REQ_F_CREDS
;
2253 return def
->prep(req
, sqe
);
2256 static __cold
int io_submit_fail_init(const struct io_uring_sqe
*sqe
,
2257 struct io_kiocb
*req
, int ret
)
2259 struct io_ring_ctx
*ctx
= req
->ctx
;
2260 struct io_submit_link
*link
= &ctx
->submit_state
.link
;
2261 struct io_kiocb
*head
= link
->head
;
2263 trace_io_uring_req_failed(sqe
, req
, ret
);
2266 * Avoid breaking links in the middle as it renders links with SQPOLL
2267 * unusable. Instead of failing eagerly, continue assembling the link if
2268 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2269 * should find the flag and handle the rest.
2271 req_fail_link_node(req
, ret
);
2272 if (head
&& !(head
->flags
& REQ_F_FAIL
))
2273 req_fail_link_node(head
, -ECANCELED
);
2275 if (!(req
->flags
& IO_REQ_LINK_FLAGS
)) {
2277 link
->last
->link
= req
;
2281 io_queue_sqe_fallback(req
);
2286 link
->last
->link
= req
;
2293 static inline int io_submit_sqe(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
2294 const struct io_uring_sqe
*sqe
)
2295 __must_hold(&ctx
->uring_lock
)
2297 struct io_submit_link
*link
= &ctx
->submit_state
.link
;
2300 ret
= io_init_req(ctx
, req
, sqe
);
2302 return io_submit_fail_init(sqe
, req
, ret
);
2304 trace_io_uring_submit_req(req
);
2307 * If we already have a head request, queue this one for async
2308 * submittal once the head completes. If we don't have a head but
2309 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2310 * submitted sync once the chain is complete. If none of those
2311 * conditions are true (normal request), then just queue it.
2313 if (unlikely(link
->head
)) {
2314 ret
= io_req_prep_async(req
);
2316 return io_submit_fail_init(sqe
, req
, ret
);
2318 trace_io_uring_link(req
, link
->head
);
2319 link
->last
->link
= req
;
2322 if (req
->flags
& IO_REQ_LINK_FLAGS
)
2324 /* last request of the link, flush it */
2327 if (req
->flags
& (REQ_F_FORCE_ASYNC
| REQ_F_FAIL
))
2330 } else if (unlikely(req
->flags
& (IO_REQ_LINK_FLAGS
|
2331 REQ_F_FORCE_ASYNC
| REQ_F_FAIL
))) {
2332 if (req
->flags
& IO_REQ_LINK_FLAGS
) {
2337 io_queue_sqe_fallback(req
);
2347 * Batched submission is done, ensure local IO is flushed out.
2349 static void io_submit_state_end(struct io_ring_ctx
*ctx
)
2351 struct io_submit_state
*state
= &ctx
->submit_state
;
2353 if (unlikely(state
->link
.head
))
2354 io_queue_sqe_fallback(state
->link
.head
);
2355 /* flush only after queuing links as they can generate completions */
2356 io_submit_flush_completions(ctx
);
2357 if (state
->plug_started
)
2358 blk_finish_plug(&state
->plug
);
2362 * Start submission side cache.
2364 static void io_submit_state_start(struct io_submit_state
*state
,
2365 unsigned int max_ios
)
2367 state
->plug_started
= false;
2368 state
->need_plug
= max_ios
> 2;
2369 state
->submit_nr
= max_ios
;
2370 /* set only head, no need to init link_last in advance */
2371 state
->link
.head
= NULL
;
2374 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
2376 struct io_rings
*rings
= ctx
->rings
;
2379 * Ensure any loads from the SQEs are done at this point,
2380 * since once we write the new head, the application could
2381 * write new data to them.
2383 smp_store_release(&rings
->sq
.head
, ctx
->cached_sq_head
);
2387 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2388 * that is mapped by userspace. This means that care needs to be taken to
2389 * ensure that reads are stable, as we cannot rely on userspace always
2390 * being a good citizen. If members of the sqe are validated and then later
2391 * used, it's important that those reads are done through READ_ONCE() to
2392 * prevent a re-load down the line.
2394 static bool io_get_sqe(struct io_ring_ctx
*ctx
, const struct io_uring_sqe
**sqe
)
2396 unsigned mask
= ctx
->sq_entries
- 1;
2397 unsigned head
= ctx
->cached_sq_head
++ & mask
;
2399 if (!(ctx
->flags
& IORING_SETUP_NO_SQARRAY
)) {
2400 head
= READ_ONCE(ctx
->sq_array
[head
]);
2401 if (unlikely(head
>= ctx
->sq_entries
)) {
2402 /* drop invalid entries */
2403 spin_lock(&ctx
->completion_lock
);
2405 spin_unlock(&ctx
->completion_lock
);
2406 WRITE_ONCE(ctx
->rings
->sq_dropped
,
2407 READ_ONCE(ctx
->rings
->sq_dropped
) + 1);
2413 * The cached sq head (or cq tail) serves two purposes:
2415 * 1) allows us to batch the cost of updating the user visible
2417 * 2) allows the kernel side to track the head on its own, even
2418 * though the application is the one updating it.
2421 /* double index for 128-byte SQEs, twice as long */
2422 if (ctx
->flags
& IORING_SETUP_SQE128
)
2424 *sqe
= &ctx
->sq_sqes
[head
];
2428 int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
)
2429 __must_hold(&ctx
->uring_lock
)
2431 unsigned int entries
= io_sqring_entries(ctx
);
2435 if (unlikely(!entries
))
2437 /* make sure SQ entry isn't read before tail */
2438 ret
= left
= min(nr
, entries
);
2439 io_get_task_refs(left
);
2440 io_submit_state_start(&ctx
->submit_state
, left
);
2443 const struct io_uring_sqe
*sqe
;
2444 struct io_kiocb
*req
;
2446 if (unlikely(!io_alloc_req(ctx
, &req
)))
2448 if (unlikely(!io_get_sqe(ctx
, &sqe
))) {
2449 io_req_add_to_cache(req
, ctx
);
2454 * Continue submitting even for sqe failure if the
2455 * ring was setup with IORING_SETUP_SUBMIT_ALL
2457 if (unlikely(io_submit_sqe(ctx
, req
, sqe
)) &&
2458 !(ctx
->flags
& IORING_SETUP_SUBMIT_ALL
)) {
2464 if (unlikely(left
)) {
2466 /* try again if it submitted nothing and can't allocate a req */
2467 if (!ret
&& io_req_cache_empty(ctx
))
2469 current
->io_uring
->cached_refs
+= left
;
2472 io_submit_state_end(ctx
);
2473 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2474 io_commit_sqring(ctx
);
2478 struct io_wait_queue
{
2479 struct wait_queue_entry wq
;
2480 struct io_ring_ctx
*ctx
;
2482 unsigned nr_timeouts
;
2486 static inline bool io_has_work(struct io_ring_ctx
*ctx
)
2488 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
) ||
2489 !llist_empty(&ctx
->work_llist
);
2492 static inline bool io_should_wake(struct io_wait_queue
*iowq
)
2494 struct io_ring_ctx
*ctx
= iowq
->ctx
;
2495 int dist
= READ_ONCE(ctx
->rings
->cq
.tail
) - (int) iowq
->cq_tail
;
2498 * Wake up if we have enough events, or if a timeout occurred since we
2499 * started waiting. For timeouts, we always want to return to userspace,
2500 * regardless of event count.
2502 return dist
>= 0 || atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
2505 static int io_wake_function(struct wait_queue_entry
*curr
, unsigned int mode
,
2506 int wake_flags
, void *key
)
2508 struct io_wait_queue
*iowq
= container_of(curr
, struct io_wait_queue
, wq
);
2511 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2512 * the task, and the next invocation will do it.
2514 if (io_should_wake(iowq
) || io_has_work(iowq
->ctx
))
2515 return autoremove_wake_function(curr
, mode
, wake_flags
, key
);
2519 int io_run_task_work_sig(struct io_ring_ctx
*ctx
)
2521 if (!llist_empty(&ctx
->work_llist
)) {
2522 __set_current_state(TASK_RUNNING
);
2523 if (io_run_local_work(ctx
) > 0)
2526 if (io_run_task_work() > 0)
2528 if (task_sigpending(current
))
2533 static bool current_pending_io(void)
2535 struct io_uring_task
*tctx
= current
->io_uring
;
2539 return percpu_counter_read_positive(&tctx
->inflight
);
2542 /* when returns >0, the caller should retry */
2543 static inline int io_cqring_wait_schedule(struct io_ring_ctx
*ctx
,
2544 struct io_wait_queue
*iowq
)
2548 if (unlikely(READ_ONCE(ctx
->check_cq
)))
2550 if (unlikely(!llist_empty(&ctx
->work_llist
)))
2552 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL
)))
2554 if (unlikely(task_sigpending(current
)))
2556 if (unlikely(io_should_wake(iowq
)))
2560 * Mark us as being in io_wait if we have pending requests, so cpufreq
2561 * can take into account that the task is waiting for IO - turns out
2562 * to be important for low QD IO.
2564 io_wait
= current
->in_iowait
;
2565 if (current_pending_io())
2566 current
->in_iowait
= 1;
2568 if (iowq
->timeout
== KTIME_MAX
)
2570 else if (!schedule_hrtimeout(&iowq
->timeout
, HRTIMER_MODE_ABS
))
2572 current
->in_iowait
= io_wait
;
2577 * Wait until events become available, if we don't already have some. The
2578 * application must reap them itself, as they reside on the shared cq ring.
2580 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
2581 const sigset_t __user
*sig
, size_t sigsz
,
2582 struct __kernel_timespec __user
*uts
)
2584 struct io_wait_queue iowq
;
2585 struct io_rings
*rings
= ctx
->rings
;
2588 if (!io_allowed_run_tw(ctx
))
2590 if (!llist_empty(&ctx
->work_llist
))
2591 io_run_local_work(ctx
);
2593 io_cqring_overflow_flush(ctx
);
2594 /* if user messes with these they will just get an early return */
2595 if (__io_cqring_events_user(ctx
) >= min_events
)
2599 #ifdef CONFIG_COMPAT
2600 if (in_compat_syscall())
2601 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
2605 ret
= set_user_sigmask(sig
, sigsz
);
2611 init_waitqueue_func_entry(&iowq
.wq
, io_wake_function
);
2612 iowq
.wq
.private = current
;
2613 INIT_LIST_HEAD(&iowq
.wq
.entry
);
2615 iowq
.nr_timeouts
= atomic_read(&ctx
->cq_timeouts
);
2616 iowq
.cq_tail
= READ_ONCE(ctx
->rings
->cq
.head
) + min_events
;
2617 iowq
.timeout
= KTIME_MAX
;
2620 struct timespec64 ts
;
2622 if (get_timespec64(&ts
, uts
))
2624 iowq
.timeout
= ktime_add_ns(timespec64_to_ktime(ts
), ktime_get_ns());
2627 trace_io_uring_cqring_wait(ctx
, min_events
);
2629 unsigned long check_cq
;
2631 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) {
2632 int nr_wait
= (int) iowq
.cq_tail
- READ_ONCE(ctx
->rings
->cq
.tail
);
2634 atomic_set(&ctx
->cq_wait_nr
, nr_wait
);
2635 set_current_state(TASK_INTERRUPTIBLE
);
2637 prepare_to_wait_exclusive(&ctx
->cq_wait
, &iowq
.wq
,
2638 TASK_INTERRUPTIBLE
);
2641 ret
= io_cqring_wait_schedule(ctx
, &iowq
);
2642 __set_current_state(TASK_RUNNING
);
2643 atomic_set(&ctx
->cq_wait_nr
, IO_CQ_WAKE_INIT
);
2646 * Run task_work after scheduling and before io_should_wake().
2647 * If we got woken because of task_work being processed, run it
2648 * now rather than let the caller do another wait loop.
2651 if (!llist_empty(&ctx
->work_llist
))
2652 io_run_local_work(ctx
);
2655 * Non-local task_work will be run on exit to userspace, but
2656 * if we're using DEFER_TASKRUN, then we could have waited
2657 * with a timeout for a number of requests. If the timeout
2658 * hits, we could have some requests ready to process. Ensure
2659 * this break is _after_ we have run task_work, to avoid
2660 * deferring running potentially pending requests until the
2661 * next time we wait for events.
2666 check_cq
= READ_ONCE(ctx
->check_cq
);
2667 if (unlikely(check_cq
)) {
2668 /* let the caller flush overflows, retry */
2669 if (check_cq
& BIT(IO_CHECK_CQ_OVERFLOW_BIT
))
2670 io_cqring_do_overflow_flush(ctx
);
2671 if (check_cq
& BIT(IO_CHECK_CQ_DROPPED_BIT
)) {
2677 if (io_should_wake(&iowq
)) {
2684 if (!(ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
))
2685 finish_wait(&ctx
->cq_wait
, &iowq
.wq
);
2686 restore_saved_sigmask_unless(ret
== -EINTR
);
2688 return READ_ONCE(rings
->cq
.head
) == READ_ONCE(rings
->cq
.tail
) ? ret
: 0;
2691 void io_mem_free(void *ptr
)
2696 folio_put(virt_to_folio(ptr
));
2699 static void io_pages_free(struct page
***pages
, int npages
)
2701 struct page
**page_array
;
2707 page_array
= *pages
;
2711 for (i
= 0; i
< npages
; i
++)
2712 unpin_user_page(page_array
[i
]);
2717 static void *__io_uaddr_map(struct page
***pages
, unsigned short *npages
,
2718 unsigned long uaddr
, size_t size
)
2720 struct page
**page_array
;
2721 unsigned int nr_pages
;
2727 if (uaddr
& (PAGE_SIZE
- 1) || !size
)
2728 return ERR_PTR(-EINVAL
);
2730 nr_pages
= (size
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2731 if (nr_pages
> USHRT_MAX
)
2732 return ERR_PTR(-EINVAL
);
2733 page_array
= kvmalloc_array(nr_pages
, sizeof(struct page
*), GFP_KERNEL
);
2735 return ERR_PTR(-ENOMEM
);
2737 ret
= pin_user_pages_fast(uaddr
, nr_pages
, FOLL_WRITE
| FOLL_LONGTERM
,
2739 if (ret
!= nr_pages
) {
2741 io_pages_free(&page_array
, ret
> 0 ? ret
: 0);
2742 return ret
< 0 ? ERR_PTR(ret
) : ERR_PTR(-EFAULT
);
2745 page_addr
= page_address(page_array
[0]);
2746 for (i
= 0; i
< nr_pages
; i
++) {
2750 * Can't support mapping user allocated ring memory on 32-bit
2751 * archs where it could potentially reside in highmem. Just
2752 * fail those with -EINVAL, just like we did on kernels that
2753 * didn't support this feature.
2755 if (PageHighMem(page_array
[i
]))
2759 * No support for discontig pages for now, should either be a
2760 * single normal page, or a huge page. Later on we can add
2761 * support for remapping discontig pages, for now we will
2762 * just fail them with EINVAL.
2764 if (page_address(page_array
[i
]) != page_addr
)
2766 page_addr
+= PAGE_SIZE
;
2769 *pages
= page_array
;
2771 return page_to_virt(page_array
[0]);
2774 static void *io_rings_map(struct io_ring_ctx
*ctx
, unsigned long uaddr
,
2777 return __io_uaddr_map(&ctx
->ring_pages
, &ctx
->n_ring_pages
, uaddr
,
2781 static void *io_sqes_map(struct io_ring_ctx
*ctx
, unsigned long uaddr
,
2784 return __io_uaddr_map(&ctx
->sqe_pages
, &ctx
->n_sqe_pages
, uaddr
,
2788 static void io_rings_free(struct io_ring_ctx
*ctx
)
2790 if (!(ctx
->flags
& IORING_SETUP_NO_MMAP
)) {
2791 io_mem_free(ctx
->rings
);
2792 io_mem_free(ctx
->sq_sqes
);
2794 ctx
->sq_sqes
= NULL
;
2796 io_pages_free(&ctx
->ring_pages
, ctx
->n_ring_pages
);
2797 ctx
->n_ring_pages
= 0;
2798 io_pages_free(&ctx
->sqe_pages
, ctx
->n_sqe_pages
);
2799 ctx
->n_sqe_pages
= 0;
2803 void *io_mem_alloc(size_t size
)
2805 gfp_t gfp
= GFP_KERNEL_ACCOUNT
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
;
2808 ret
= (void *) __get_free_pages(gfp
, get_order(size
));
2811 return ERR_PTR(-ENOMEM
);
2814 static unsigned long rings_size(struct io_ring_ctx
*ctx
, unsigned int sq_entries
,
2815 unsigned int cq_entries
, size_t *sq_offset
)
2817 struct io_rings
*rings
;
2818 size_t off
, sq_array_size
;
2820 off
= struct_size(rings
, cqes
, cq_entries
);
2821 if (off
== SIZE_MAX
)
2823 if (ctx
->flags
& IORING_SETUP_CQE32
) {
2824 if (check_shl_overflow(off
, 1, &off
))
2829 off
= ALIGN(off
, SMP_CACHE_BYTES
);
2834 if (ctx
->flags
& IORING_SETUP_NO_SQARRAY
) {
2836 *sq_offset
= SIZE_MAX
;
2843 sq_array_size
= array_size(sizeof(u32
), sq_entries
);
2844 if (sq_array_size
== SIZE_MAX
)
2847 if (check_add_overflow(off
, sq_array_size
, &off
))
2853 static void io_req_caches_free(struct io_ring_ctx
*ctx
)
2855 struct io_kiocb
*req
;
2858 mutex_lock(&ctx
->uring_lock
);
2859 io_flush_cached_locked_reqs(ctx
, &ctx
->submit_state
);
2861 while (!io_req_cache_empty(ctx
)) {
2862 req
= io_extract_req(ctx
);
2863 kmem_cache_free(req_cachep
, req
);
2867 percpu_ref_put_many(&ctx
->refs
, nr
);
2868 mutex_unlock(&ctx
->uring_lock
);
2871 static void io_rsrc_node_cache_free(struct io_cache_entry
*entry
)
2873 kfree(container_of(entry
, struct io_rsrc_node
, cache
));
2876 static __cold
void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
2878 io_sq_thread_finish(ctx
);
2879 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2880 if (WARN_ON_ONCE(!list_empty(&ctx
->rsrc_ref_list
)))
2883 mutex_lock(&ctx
->uring_lock
);
2885 __io_sqe_buffers_unregister(ctx
);
2887 __io_sqe_files_unregister(ctx
);
2888 io_cqring_overflow_kill(ctx
);
2889 io_eventfd_unregister(ctx
);
2890 io_alloc_cache_free(&ctx
->apoll_cache
, io_apoll_cache_free
);
2891 io_alloc_cache_free(&ctx
->netmsg_cache
, io_netmsg_cache_free
);
2892 io_futex_cache_free(ctx
);
2893 io_destroy_buffers(ctx
);
2894 mutex_unlock(&ctx
->uring_lock
);
2896 put_cred(ctx
->sq_creds
);
2897 if (ctx
->submitter_task
)
2898 put_task_struct(ctx
->submitter_task
);
2900 /* there are no registered resources left, nobody uses it */
2902 io_rsrc_node_destroy(ctx
, ctx
->rsrc_node
);
2904 WARN_ON_ONCE(!list_empty(&ctx
->rsrc_ref_list
));
2905 WARN_ON_ONCE(!list_empty(&ctx
->ltimeout_list
));
2907 io_alloc_cache_free(&ctx
->rsrc_node_cache
, io_rsrc_node_cache_free
);
2908 if (ctx
->mm_account
) {
2909 mmdrop(ctx
->mm_account
);
2910 ctx
->mm_account
= NULL
;
2913 io_kbuf_mmap_list_free(ctx
);
2915 percpu_ref_exit(&ctx
->refs
);
2916 free_uid(ctx
->user
);
2917 io_req_caches_free(ctx
);
2919 io_wq_put_hash(ctx
->hash_map
);
2920 kfree(ctx
->cancel_table
.hbs
);
2921 kfree(ctx
->cancel_table_locked
.hbs
);
2923 xa_destroy(&ctx
->io_bl_xa
);
2927 static __cold
void io_activate_pollwq_cb(struct callback_head
*cb
)
2929 struct io_ring_ctx
*ctx
= container_of(cb
, struct io_ring_ctx
,
2932 mutex_lock(&ctx
->uring_lock
);
2933 ctx
->poll_activated
= true;
2934 mutex_unlock(&ctx
->uring_lock
);
2937 * Wake ups for some events between start of polling and activation
2938 * might've been lost due to loose synchronisation.
2940 wake_up_all(&ctx
->poll_wq
);
2941 percpu_ref_put(&ctx
->refs
);
2944 __cold
void io_activate_pollwq(struct io_ring_ctx
*ctx
)
2946 spin_lock(&ctx
->completion_lock
);
2947 /* already activated or in progress */
2948 if (ctx
->poll_activated
|| ctx
->poll_wq_task_work
.func
)
2950 if (WARN_ON_ONCE(!ctx
->task_complete
))
2952 if (!ctx
->submitter_task
)
2955 * with ->submitter_task only the submitter task completes requests, we
2956 * only need to sync with it, which is done by injecting a tw
2958 init_task_work(&ctx
->poll_wq_task_work
, io_activate_pollwq_cb
);
2959 percpu_ref_get(&ctx
->refs
);
2960 if (task_work_add(ctx
->submitter_task
, &ctx
->poll_wq_task_work
, TWA_SIGNAL
))
2961 percpu_ref_put(&ctx
->refs
);
2963 spin_unlock(&ctx
->completion_lock
);
2966 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
2968 struct io_ring_ctx
*ctx
= file
->private_data
;
2971 if (unlikely(!ctx
->poll_activated
))
2972 io_activate_pollwq(ctx
);
2974 poll_wait(file
, &ctx
->poll_wq
, wait
);
2976 * synchronizes with barrier from wq_has_sleeper call in
2980 if (!io_sqring_full(ctx
))
2981 mask
|= EPOLLOUT
| EPOLLWRNORM
;
2984 * Don't flush cqring overflow list here, just do a simple check.
2985 * Otherwise there could possible be ABBA deadlock:
2988 * lock(&ctx->uring_lock);
2990 * lock(&ctx->uring_lock);
2993 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2994 * pushes them to do the flush.
2997 if (__io_cqring_events_user(ctx
) || io_has_work(ctx
))
2998 mask
|= EPOLLIN
| EPOLLRDNORM
;
3003 struct io_tctx_exit
{
3004 struct callback_head task_work
;
3005 struct completion completion
;
3006 struct io_ring_ctx
*ctx
;
3009 static __cold
void io_tctx_exit_cb(struct callback_head
*cb
)
3011 struct io_uring_task
*tctx
= current
->io_uring
;
3012 struct io_tctx_exit
*work
;
3014 work
= container_of(cb
, struct io_tctx_exit
, task_work
);
3016 * When @in_cancel, we're in cancellation and it's racy to remove the
3017 * node. It'll be removed by the end of cancellation, just ignore it.
3018 * tctx can be NULL if the queueing of this task_work raced with
3019 * work cancelation off the exec path.
3021 if (tctx
&& !atomic_read(&tctx
->in_cancel
))
3022 io_uring_del_tctx_node((unsigned long)work
->ctx
);
3023 complete(&work
->completion
);
3026 static __cold
bool io_cancel_ctx_cb(struct io_wq_work
*work
, void *data
)
3028 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
3030 return req
->ctx
== data
;
3033 static __cold
void io_ring_exit_work(struct work_struct
*work
)
3035 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
, exit_work
);
3036 unsigned long timeout
= jiffies
+ HZ
* 60 * 5;
3037 unsigned long interval
= HZ
/ 20;
3038 struct io_tctx_exit exit
;
3039 struct io_tctx_node
*node
;
3043 * If we're doing polled IO and end up having requests being
3044 * submitted async (out-of-line), then completions can come in while
3045 * we're waiting for refs to drop. We need to reap these manually,
3046 * as nobody else will be looking for them.
3049 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
)) {
3050 mutex_lock(&ctx
->uring_lock
);
3051 io_cqring_overflow_kill(ctx
);
3052 mutex_unlock(&ctx
->uring_lock
);
3055 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
)
3056 io_move_task_work_from_local(ctx
);
3058 while (io_uring_try_cancel_requests(ctx
, NULL
, true))
3062 struct io_sq_data
*sqd
= ctx
->sq_data
;
3063 struct task_struct
*tsk
;
3065 io_sq_thread_park(sqd
);
3067 if (tsk
&& tsk
->io_uring
&& tsk
->io_uring
->io_wq
)
3068 io_wq_cancel_cb(tsk
->io_uring
->io_wq
,
3069 io_cancel_ctx_cb
, ctx
, true);
3070 io_sq_thread_unpark(sqd
);
3073 io_req_caches_free(ctx
);
3075 if (WARN_ON_ONCE(time_after(jiffies
, timeout
))) {
3076 /* there is little hope left, don't run it too often */
3080 * This is really an uninterruptible wait, as it has to be
3081 * complete. But it's also run from a kworker, which doesn't
3082 * take signals, so it's fine to make it interruptible. This
3083 * avoids scenarios where we knowingly can wait much longer
3084 * on completions, for example if someone does a SIGSTOP on
3085 * a task that needs to finish task_work to make this loop
3086 * complete. That's a synthetic situation that should not
3087 * cause a stuck task backtrace, and hence a potential panic
3088 * on stuck tasks if that is enabled.
3090 } while (!wait_for_completion_interruptible_timeout(&ctx
->ref_comp
, interval
));
3092 init_completion(&exit
.completion
);
3093 init_task_work(&exit
.task_work
, io_tctx_exit_cb
);
3096 mutex_lock(&ctx
->uring_lock
);
3097 while (!list_empty(&ctx
->tctx_list
)) {
3098 WARN_ON_ONCE(time_after(jiffies
, timeout
));
3100 node
= list_first_entry(&ctx
->tctx_list
, struct io_tctx_node
,
3102 /* don't spin on a single task if cancellation failed */
3103 list_rotate_left(&ctx
->tctx_list
);
3104 ret
= task_work_add(node
->task
, &exit
.task_work
, TWA_SIGNAL
);
3105 if (WARN_ON_ONCE(ret
))
3108 mutex_unlock(&ctx
->uring_lock
);
3110 * See comment above for
3111 * wait_for_completion_interruptible_timeout() on why this
3112 * wait is marked as interruptible.
3114 wait_for_completion_interruptible(&exit
.completion
);
3115 mutex_lock(&ctx
->uring_lock
);
3117 mutex_unlock(&ctx
->uring_lock
);
3118 spin_lock(&ctx
->completion_lock
);
3119 spin_unlock(&ctx
->completion_lock
);
3121 /* pairs with RCU read section in io_req_local_work_add() */
3122 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
)
3125 io_ring_ctx_free(ctx
);
3128 static __cold
void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
3130 unsigned long index
;
3131 struct creds
*creds
;
3133 mutex_lock(&ctx
->uring_lock
);
3134 percpu_ref_kill(&ctx
->refs
);
3135 xa_for_each(&ctx
->personalities
, index
, creds
)
3136 io_unregister_personality(ctx
, index
);
3138 io_poll_remove_all(ctx
, NULL
, true);
3139 mutex_unlock(&ctx
->uring_lock
);
3142 * If we failed setting up the ctx, we might not have any rings
3143 * and therefore did not submit any requests
3146 io_kill_timeouts(ctx
, NULL
, true);
3148 flush_delayed_work(&ctx
->fallback_work
);
3150 INIT_WORK(&ctx
->exit_work
, io_ring_exit_work
);
3152 * Use system_unbound_wq to avoid spawning tons of event kworkers
3153 * if we're exiting a ton of rings at the same time. It just adds
3154 * noise and overhead, there's no discernable change in runtime
3155 * over using system_wq.
3157 queue_work(system_unbound_wq
, &ctx
->exit_work
);
3160 static int io_uring_release(struct inode
*inode
, struct file
*file
)
3162 struct io_ring_ctx
*ctx
= file
->private_data
;
3164 file
->private_data
= NULL
;
3165 io_ring_ctx_wait_and_kill(ctx
);
3169 struct io_task_cancel
{
3170 struct task_struct
*task
;
3174 static bool io_cancel_task_cb(struct io_wq_work
*work
, void *data
)
3176 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
3177 struct io_task_cancel
*cancel
= data
;
3179 return io_match_task_safe(req
, cancel
->task
, cancel
->all
);
3182 static __cold
bool io_cancel_defer_files(struct io_ring_ctx
*ctx
,
3183 struct task_struct
*task
,
3186 struct io_defer_entry
*de
;
3189 spin_lock(&ctx
->completion_lock
);
3190 list_for_each_entry_reverse(de
, &ctx
->defer_list
, list
) {
3191 if (io_match_task_safe(de
->req
, task
, cancel_all
)) {
3192 list_cut_position(&list
, &ctx
->defer_list
, &de
->list
);
3196 spin_unlock(&ctx
->completion_lock
);
3197 if (list_empty(&list
))
3200 while (!list_empty(&list
)) {
3201 de
= list_first_entry(&list
, struct io_defer_entry
, list
);
3202 list_del_init(&de
->list
);
3203 io_req_task_queue_fail(de
->req
, -ECANCELED
);
3209 static __cold
bool io_uring_try_cancel_iowq(struct io_ring_ctx
*ctx
)
3211 struct io_tctx_node
*node
;
3212 enum io_wq_cancel cret
;
3215 mutex_lock(&ctx
->uring_lock
);
3216 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
3217 struct io_uring_task
*tctx
= node
->task
->io_uring
;
3220 * io_wq will stay alive while we hold uring_lock, because it's
3221 * killed after ctx nodes, which requires to take the lock.
3223 if (!tctx
|| !tctx
->io_wq
)
3225 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_ctx_cb
, ctx
, true);
3226 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
3228 mutex_unlock(&ctx
->uring_lock
);
3233 static bool io_uring_try_cancel_uring_cmd(struct io_ring_ctx
*ctx
,
3234 struct task_struct
*task
, bool cancel_all
)
3236 struct hlist_node
*tmp
;
3237 struct io_kiocb
*req
;
3240 lockdep_assert_held(&ctx
->uring_lock
);
3242 hlist_for_each_entry_safe(req
, tmp
, &ctx
->cancelable_uring_cmd
,
3244 struct io_uring_cmd
*cmd
= io_kiocb_to_cmd(req
,
3245 struct io_uring_cmd
);
3246 struct file
*file
= req
->file
;
3248 if (!cancel_all
&& req
->task
!= task
)
3251 if (cmd
->flags
& IORING_URING_CMD_CANCELABLE
) {
3252 /* ->sqe isn't available if no async data */
3253 if (!req_has_async_data(req
))
3255 file
->f_op
->uring_cmd(cmd
, IO_URING_F_CANCEL
);
3259 io_submit_flush_completions(ctx
);
3264 static __cold
bool io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
3265 struct task_struct
*task
,
3268 struct io_task_cancel cancel
= { .task
= task
, .all
= cancel_all
, };
3269 struct io_uring_task
*tctx
= task
? task
->io_uring
: NULL
;
3270 enum io_wq_cancel cret
;
3273 /* set it so io_req_local_work_add() would wake us up */
3274 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) {
3275 atomic_set(&ctx
->cq_wait_nr
, 1);
3279 /* failed during ring init, it couldn't have issued any requests */
3284 ret
|= io_uring_try_cancel_iowq(ctx
);
3285 } else if (tctx
&& tctx
->io_wq
) {
3287 * Cancels requests of all rings, not only @ctx, but
3288 * it's fine as the task is in exit/exec.
3290 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_task_cb
,
3292 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
3295 /* SQPOLL thread does its own polling */
3296 if ((!(ctx
->flags
& IORING_SETUP_SQPOLL
) && cancel_all
) ||
3297 (ctx
->sq_data
&& ctx
->sq_data
->thread
== current
)) {
3298 while (!wq_list_empty(&ctx
->iopoll_list
)) {
3299 io_iopoll_try_reap_events(ctx
);
3305 if ((ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) &&
3306 io_allowed_defer_tw_run(ctx
))
3307 ret
|= io_run_local_work(ctx
) > 0;
3308 ret
|= io_cancel_defer_files(ctx
, task
, cancel_all
);
3309 mutex_lock(&ctx
->uring_lock
);
3310 ret
|= io_poll_remove_all(ctx
, task
, cancel_all
);
3311 ret
|= io_waitid_remove_all(ctx
, task
, cancel_all
);
3312 ret
|= io_futex_remove_all(ctx
, task
, cancel_all
);
3313 ret
|= io_uring_try_cancel_uring_cmd(ctx
, task
, cancel_all
);
3314 mutex_unlock(&ctx
->uring_lock
);
3315 ret
|= io_kill_timeouts(ctx
, task
, cancel_all
);
3317 ret
|= io_run_task_work() > 0;
3321 static s64
tctx_inflight(struct io_uring_task
*tctx
, bool tracked
)
3324 return atomic_read(&tctx
->inflight_tracked
);
3325 return percpu_counter_sum(&tctx
->inflight
);
3329 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3330 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3332 __cold
void io_uring_cancel_generic(bool cancel_all
, struct io_sq_data
*sqd
)
3334 struct io_uring_task
*tctx
= current
->io_uring
;
3335 struct io_ring_ctx
*ctx
;
3336 struct io_tctx_node
*node
;
3337 unsigned long index
;
3341 WARN_ON_ONCE(sqd
&& sqd
->thread
!= current
);
3343 if (!current
->io_uring
)
3346 io_wq_exit_start(tctx
->io_wq
);
3348 atomic_inc(&tctx
->in_cancel
);
3352 io_uring_drop_tctx_refs(current
);
3353 /* read completions before cancelations */
3354 inflight
= tctx_inflight(tctx
, !cancel_all
);
3359 xa_for_each(&tctx
->xa
, index
, node
) {
3360 /* sqpoll task will cancel all its requests */
3361 if (node
->ctx
->sq_data
)
3363 loop
|= io_uring_try_cancel_requests(node
->ctx
,
3364 current
, cancel_all
);
3367 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
3368 loop
|= io_uring_try_cancel_requests(ctx
,
3378 prepare_to_wait(&tctx
->wait
, &wait
, TASK_INTERRUPTIBLE
);
3380 io_uring_drop_tctx_refs(current
);
3381 xa_for_each(&tctx
->xa
, index
, node
) {
3382 if (!llist_empty(&node
->ctx
->work_llist
)) {
3383 WARN_ON_ONCE(node
->ctx
->submitter_task
&&
3384 node
->ctx
->submitter_task
!= current
);
3389 * If we've seen completions, retry without waiting. This
3390 * avoids a race where a completion comes in before we did
3391 * prepare_to_wait().
3393 if (inflight
== tctx_inflight(tctx
, !cancel_all
))
3396 finish_wait(&tctx
->wait
, &wait
);
3399 io_uring_clean_tctx(tctx
);
3402 * We shouldn't run task_works after cancel, so just leave
3403 * ->in_cancel set for normal exit.
3405 atomic_dec(&tctx
->in_cancel
);
3406 /* for exec all current's requests should be gone, kill tctx */
3407 __io_uring_free(current
);
3411 void __io_uring_cancel(bool cancel_all
)
3413 io_uring_cancel_generic(cancel_all
, NULL
);
3416 static void *io_uring_validate_mmap_request(struct file
*file
,
3417 loff_t pgoff
, size_t sz
)
3419 struct io_ring_ctx
*ctx
= file
->private_data
;
3420 loff_t offset
= pgoff
<< PAGE_SHIFT
;
3424 switch (offset
& IORING_OFF_MMAP_MASK
) {
3425 case IORING_OFF_SQ_RING
:
3426 case IORING_OFF_CQ_RING
:
3427 /* Don't allow mmap if the ring was setup without it */
3428 if (ctx
->flags
& IORING_SETUP_NO_MMAP
)
3429 return ERR_PTR(-EINVAL
);
3432 case IORING_OFF_SQES
:
3433 /* Don't allow mmap if the ring was setup without it */
3434 if (ctx
->flags
& IORING_SETUP_NO_MMAP
)
3435 return ERR_PTR(-EINVAL
);
3438 case IORING_OFF_PBUF_RING
: {
3441 bgid
= (offset
& ~IORING_OFF_MMAP_MASK
) >> IORING_OFF_PBUF_SHIFT
;
3443 ptr
= io_pbuf_get_address(ctx
, bgid
);
3446 return ERR_PTR(-EINVAL
);
3450 return ERR_PTR(-EINVAL
);
3453 page
= virt_to_head_page(ptr
);
3454 if (sz
> page_size(page
))
3455 return ERR_PTR(-EINVAL
);
3462 static __cold
int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3464 size_t sz
= vma
->vm_end
- vma
->vm_start
;
3468 ptr
= io_uring_validate_mmap_request(file
, vma
->vm_pgoff
, sz
);
3470 return PTR_ERR(ptr
);
3472 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
3473 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
3476 static unsigned long io_uring_mmu_get_unmapped_area(struct file
*filp
,
3477 unsigned long addr
, unsigned long len
,
3478 unsigned long pgoff
, unsigned long flags
)
3483 * Do not allow to map to user-provided address to avoid breaking the
3484 * aliasing rules. Userspace is not able to guess the offset address of
3485 * kernel kmalloc()ed memory area.
3490 ptr
= io_uring_validate_mmap_request(filp
, pgoff
, len
);
3495 * Some architectures have strong cache aliasing requirements.
3496 * For such architectures we need a coherent mapping which aliases
3497 * kernel memory *and* userspace memory. To achieve that:
3498 * - use a NULL file pointer to reference physical memory, and
3499 * - use the kernel virtual address of the shared io_uring context
3500 * (instead of the userspace-provided address, which has to be 0UL
3502 * - use the same pgoff which the get_unmapped_area() uses to
3503 * calculate the page colouring.
3504 * For architectures without such aliasing requirements, the
3505 * architecture will return any suitable mapping because addr is 0.
3508 flags
|= MAP_SHARED
;
3509 pgoff
= 0; /* has been translated to ptr above */
3511 addr
= (uintptr_t) ptr
;
3512 pgoff
= addr
>> PAGE_SHIFT
;
3516 return current
->mm
->get_unmapped_area(filp
, addr
, len
, pgoff
, flags
);
3519 #else /* !CONFIG_MMU */
3521 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3523 return is_nommu_shared_mapping(vma
->vm_flags
) ? 0 : -EINVAL
;
3526 static unsigned int io_uring_nommu_mmap_capabilities(struct file
*file
)
3528 return NOMMU_MAP_DIRECT
| NOMMU_MAP_READ
| NOMMU_MAP_WRITE
;
3531 static unsigned long io_uring_nommu_get_unmapped_area(struct file
*file
,
3532 unsigned long addr
, unsigned long len
,
3533 unsigned long pgoff
, unsigned long flags
)
3537 ptr
= io_uring_validate_mmap_request(file
, pgoff
, len
);
3539 return PTR_ERR(ptr
);
3541 return (unsigned long) ptr
;
3544 #endif /* !CONFIG_MMU */
3546 static int io_validate_ext_arg(unsigned flags
, const void __user
*argp
, size_t argsz
)
3548 if (flags
& IORING_ENTER_EXT_ARG
) {
3549 struct io_uring_getevents_arg arg
;
3551 if (argsz
!= sizeof(arg
))
3553 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
3559 static int io_get_ext_arg(unsigned flags
, const void __user
*argp
, size_t *argsz
,
3560 struct __kernel_timespec __user
**ts
,
3561 const sigset_t __user
**sig
)
3563 struct io_uring_getevents_arg arg
;
3566 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3567 * is just a pointer to the sigset_t.
3569 if (!(flags
& IORING_ENTER_EXT_ARG
)) {
3570 *sig
= (const sigset_t __user
*) argp
;
3576 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3577 * timespec and sigset_t pointers if good.
3579 if (*argsz
!= sizeof(arg
))
3581 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
3585 *sig
= u64_to_user_ptr(arg
.sigmask
);
3586 *argsz
= arg
.sigmask_sz
;
3587 *ts
= u64_to_user_ptr(arg
.ts
);
3591 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
3592 u32
, min_complete
, u32
, flags
, const void __user
*, argp
,
3595 struct io_ring_ctx
*ctx
;
3599 if (unlikely(flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
|
3600 IORING_ENTER_SQ_WAIT
| IORING_ENTER_EXT_ARG
|
3601 IORING_ENTER_REGISTERED_RING
)))
3605 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3606 * need only dereference our task private array to find it.
3608 if (flags
& IORING_ENTER_REGISTERED_RING
) {
3609 struct io_uring_task
*tctx
= current
->io_uring
;
3611 if (unlikely(!tctx
|| fd
>= IO_RINGFD_REG_MAX
))
3613 fd
= array_index_nospec(fd
, IO_RINGFD_REG_MAX
);
3614 file
= tctx
->registered_rings
[fd
];
3615 if (unlikely(!file
))
3619 if (unlikely(!file
))
3622 if (unlikely(!io_is_uring_fops(file
)))
3626 ctx
= file
->private_data
;
3628 if (unlikely(ctx
->flags
& IORING_SETUP_R_DISABLED
))
3632 * For SQ polling, the thread will do all submissions and completions.
3633 * Just return the requested submit count, and wake the thread if
3637 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
3638 io_cqring_overflow_flush(ctx
);
3640 if (unlikely(ctx
->sq_data
->thread
== NULL
)) {
3644 if (flags
& IORING_ENTER_SQ_WAKEUP
)
3645 wake_up(&ctx
->sq_data
->wait
);
3646 if (flags
& IORING_ENTER_SQ_WAIT
)
3647 io_sqpoll_wait_sq(ctx
);
3650 } else if (to_submit
) {
3651 ret
= io_uring_add_tctx_node(ctx
);
3655 mutex_lock(&ctx
->uring_lock
);
3656 ret
= io_submit_sqes(ctx
, to_submit
);
3657 if (ret
!= to_submit
) {
3658 mutex_unlock(&ctx
->uring_lock
);
3661 if (flags
& IORING_ENTER_GETEVENTS
) {
3662 if (ctx
->syscall_iopoll
)
3665 * Ignore errors, we'll soon call io_cqring_wait() and
3666 * it should handle ownership problems if any.
3668 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
)
3669 (void)io_run_local_work_locked(ctx
);
3671 mutex_unlock(&ctx
->uring_lock
);
3674 if (flags
& IORING_ENTER_GETEVENTS
) {
3677 if (ctx
->syscall_iopoll
) {
3679 * We disallow the app entering submit/complete with
3680 * polling, but we still need to lock the ring to
3681 * prevent racing with polled issue that got punted to
3684 mutex_lock(&ctx
->uring_lock
);
3686 ret2
= io_validate_ext_arg(flags
, argp
, argsz
);
3687 if (likely(!ret2
)) {
3688 min_complete
= min(min_complete
,
3690 ret2
= io_iopoll_check(ctx
, min_complete
);
3692 mutex_unlock(&ctx
->uring_lock
);
3694 const sigset_t __user
*sig
;
3695 struct __kernel_timespec __user
*ts
;
3697 ret2
= io_get_ext_arg(flags
, argp
, &argsz
, &ts
, &sig
);
3698 if (likely(!ret2
)) {
3699 min_complete
= min(min_complete
,
3701 ret2
= io_cqring_wait(ctx
, min_complete
, sig
,
3710 * EBADR indicates that one or more CQE were dropped.
3711 * Once the user has been informed we can clear the bit
3712 * as they are obviously ok with those drops.
3714 if (unlikely(ret2
== -EBADR
))
3715 clear_bit(IO_CHECK_CQ_DROPPED_BIT
,
3720 if (!(flags
& IORING_ENTER_REGISTERED_RING
))
3725 static const struct file_operations io_uring_fops
= {
3726 .release
= io_uring_release
,
3727 .mmap
= io_uring_mmap
,
3729 .get_unmapped_area
= io_uring_nommu_get_unmapped_area
,
3730 .mmap_capabilities
= io_uring_nommu_mmap_capabilities
,
3732 .get_unmapped_area
= io_uring_mmu_get_unmapped_area
,
3734 .poll
= io_uring_poll
,
3735 #ifdef CONFIG_PROC_FS
3736 .show_fdinfo
= io_uring_show_fdinfo
,
3740 bool io_is_uring_fops(struct file
*file
)
3742 return file
->f_op
== &io_uring_fops
;
3745 static __cold
int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
3746 struct io_uring_params
*p
)
3748 struct io_rings
*rings
;
3749 size_t size
, sq_array_offset
;
3752 /* make sure these are sane, as we already accounted them */
3753 ctx
->sq_entries
= p
->sq_entries
;
3754 ctx
->cq_entries
= p
->cq_entries
;
3756 size
= rings_size(ctx
, p
->sq_entries
, p
->cq_entries
, &sq_array_offset
);
3757 if (size
== SIZE_MAX
)
3760 if (!(ctx
->flags
& IORING_SETUP_NO_MMAP
))
3761 rings
= io_mem_alloc(size
);
3763 rings
= io_rings_map(ctx
, p
->cq_off
.user_addr
, size
);
3766 return PTR_ERR(rings
);
3769 if (!(ctx
->flags
& IORING_SETUP_NO_SQARRAY
))
3770 ctx
->sq_array
= (u32
*)((char *)rings
+ sq_array_offset
);
3771 rings
->sq_ring_mask
= p
->sq_entries
- 1;
3772 rings
->cq_ring_mask
= p
->cq_entries
- 1;
3773 rings
->sq_ring_entries
= p
->sq_entries
;
3774 rings
->cq_ring_entries
= p
->cq_entries
;
3776 if (p
->flags
& IORING_SETUP_SQE128
)
3777 size
= array_size(2 * sizeof(struct io_uring_sqe
), p
->sq_entries
);
3779 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
3780 if (size
== SIZE_MAX
) {
3785 if (!(ctx
->flags
& IORING_SETUP_NO_MMAP
))
3786 ptr
= io_mem_alloc(size
);
3788 ptr
= io_sqes_map(ctx
, p
->sq_off
.user_addr
, size
);
3792 return PTR_ERR(ptr
);
3799 static int io_uring_install_fd(struct file
*file
)
3803 fd
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
3806 fd_install(fd
, file
);
3811 * Allocate an anonymous fd, this is what constitutes the application
3812 * visible backing of an io_uring instance. The application mmaps this
3813 * fd to gain access to the SQ/CQ ring details.
3815 static struct file
*io_uring_get_file(struct io_ring_ctx
*ctx
)
3817 /* Create a new inode so that the LSM can block the creation. */
3818 return anon_inode_create_getfile("[io_uring]", &io_uring_fops
, ctx
,
3819 O_RDWR
| O_CLOEXEC
, NULL
);
3822 static __cold
int io_uring_create(unsigned entries
, struct io_uring_params
*p
,
3823 struct io_uring_params __user
*params
)
3825 struct io_ring_ctx
*ctx
;
3826 struct io_uring_task
*tctx
;
3832 if (entries
> IORING_MAX_ENTRIES
) {
3833 if (!(p
->flags
& IORING_SETUP_CLAMP
))
3835 entries
= IORING_MAX_ENTRIES
;
3838 if ((p
->flags
& IORING_SETUP_REGISTERED_FD_ONLY
)
3839 && !(p
->flags
& IORING_SETUP_NO_MMAP
))
3843 * Use twice as many entries for the CQ ring. It's possible for the
3844 * application to drive a higher depth than the size of the SQ ring,
3845 * since the sqes are only used at submission time. This allows for
3846 * some flexibility in overcommitting a bit. If the application has
3847 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3848 * of CQ ring entries manually.
3850 p
->sq_entries
= roundup_pow_of_two(entries
);
3851 if (p
->flags
& IORING_SETUP_CQSIZE
) {
3853 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3854 * to a power-of-two, if it isn't already. We do NOT impose
3855 * any cq vs sq ring sizing.
3859 if (p
->cq_entries
> IORING_MAX_CQ_ENTRIES
) {
3860 if (!(p
->flags
& IORING_SETUP_CLAMP
))
3862 p
->cq_entries
= IORING_MAX_CQ_ENTRIES
;
3864 p
->cq_entries
= roundup_pow_of_two(p
->cq_entries
);
3865 if (p
->cq_entries
< p
->sq_entries
)
3868 p
->cq_entries
= 2 * p
->sq_entries
;
3871 ctx
= io_ring_ctx_alloc(p
);
3875 if ((ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
) &&
3876 !(ctx
->flags
& IORING_SETUP_IOPOLL
) &&
3877 !(ctx
->flags
& IORING_SETUP_SQPOLL
))
3878 ctx
->task_complete
= true;
3880 if (ctx
->task_complete
|| (ctx
->flags
& IORING_SETUP_IOPOLL
))
3881 ctx
->lockless_cq
= true;
3884 * lazy poll_wq activation relies on ->task_complete for synchronisation
3885 * purposes, see io_activate_pollwq()
3887 if (!ctx
->task_complete
)
3888 ctx
->poll_activated
= true;
3891 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3892 * space applications don't need to do io completion events
3893 * polling again, they can rely on io_sq_thread to do polling
3894 * work, which can reduce cpu usage and uring_lock contention.
3896 if (ctx
->flags
& IORING_SETUP_IOPOLL
&&
3897 !(ctx
->flags
& IORING_SETUP_SQPOLL
))
3898 ctx
->syscall_iopoll
= 1;
3900 ctx
->compat
= in_compat_syscall();
3901 if (!ns_capable_noaudit(&init_user_ns
, CAP_IPC_LOCK
))
3902 ctx
->user
= get_uid(current_user());
3905 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3906 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3909 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
3910 /* IPI related flags don't make sense with SQPOLL */
3911 if (ctx
->flags
& (IORING_SETUP_COOP_TASKRUN
|
3912 IORING_SETUP_TASKRUN_FLAG
|
3913 IORING_SETUP_DEFER_TASKRUN
))
3915 ctx
->notify_method
= TWA_SIGNAL_NO_IPI
;
3916 } else if (ctx
->flags
& IORING_SETUP_COOP_TASKRUN
) {
3917 ctx
->notify_method
= TWA_SIGNAL_NO_IPI
;
3919 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
&&
3920 !(ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
))
3922 ctx
->notify_method
= TWA_SIGNAL
;
3926 * For DEFER_TASKRUN we require the completion task to be the same as the
3927 * submission task. This implies that there is only one submitter, so enforce
3930 if (ctx
->flags
& IORING_SETUP_DEFER_TASKRUN
&&
3931 !(ctx
->flags
& IORING_SETUP_SINGLE_ISSUER
)) {
3936 * This is just grabbed for accounting purposes. When a process exits,
3937 * the mm is exited and dropped before the files, hence we need to hang
3938 * on to this mm purely for the purposes of being able to unaccount
3939 * memory (locked/pinned vm). It's not used for anything else.
3941 mmgrab(current
->mm
);
3942 ctx
->mm_account
= current
->mm
;
3944 ret
= io_allocate_scq_urings(ctx
, p
);
3948 ret
= io_sq_offload_create(ctx
, p
);
3952 ret
= io_rsrc_init(ctx
);
3956 p
->sq_off
.head
= offsetof(struct io_rings
, sq
.head
);
3957 p
->sq_off
.tail
= offsetof(struct io_rings
, sq
.tail
);
3958 p
->sq_off
.ring_mask
= offsetof(struct io_rings
, sq_ring_mask
);
3959 p
->sq_off
.ring_entries
= offsetof(struct io_rings
, sq_ring_entries
);
3960 p
->sq_off
.flags
= offsetof(struct io_rings
, sq_flags
);
3961 p
->sq_off
.dropped
= offsetof(struct io_rings
, sq_dropped
);
3962 if (!(ctx
->flags
& IORING_SETUP_NO_SQARRAY
))
3963 p
->sq_off
.array
= (char *)ctx
->sq_array
- (char *)ctx
->rings
;
3964 p
->sq_off
.resv1
= 0;
3965 if (!(ctx
->flags
& IORING_SETUP_NO_MMAP
))
3966 p
->sq_off
.user_addr
= 0;
3968 p
->cq_off
.head
= offsetof(struct io_rings
, cq
.head
);
3969 p
->cq_off
.tail
= offsetof(struct io_rings
, cq
.tail
);
3970 p
->cq_off
.ring_mask
= offsetof(struct io_rings
, cq_ring_mask
);
3971 p
->cq_off
.ring_entries
= offsetof(struct io_rings
, cq_ring_entries
);
3972 p
->cq_off
.overflow
= offsetof(struct io_rings
, cq_overflow
);
3973 p
->cq_off
.cqes
= offsetof(struct io_rings
, cqes
);
3974 p
->cq_off
.flags
= offsetof(struct io_rings
, cq_flags
);
3975 p
->cq_off
.resv1
= 0;
3976 if (!(ctx
->flags
& IORING_SETUP_NO_MMAP
))
3977 p
->cq_off
.user_addr
= 0;
3979 p
->features
= IORING_FEAT_SINGLE_MMAP
| IORING_FEAT_NODROP
|
3980 IORING_FEAT_SUBMIT_STABLE
| IORING_FEAT_RW_CUR_POS
|
3981 IORING_FEAT_CUR_PERSONALITY
| IORING_FEAT_FAST_POLL
|
3982 IORING_FEAT_POLL_32BITS
| IORING_FEAT_SQPOLL_NONFIXED
|
3983 IORING_FEAT_EXT_ARG
| IORING_FEAT_NATIVE_WORKERS
|
3984 IORING_FEAT_RSRC_TAGS
| IORING_FEAT_CQE_SKIP
|
3985 IORING_FEAT_LINKED_FILE
| IORING_FEAT_REG_REG_RING
;
3987 if (copy_to_user(params
, p
, sizeof(*p
))) {
3992 if (ctx
->flags
& IORING_SETUP_SINGLE_ISSUER
3993 && !(ctx
->flags
& IORING_SETUP_R_DISABLED
))
3994 WRITE_ONCE(ctx
->submitter_task
, get_task_struct(current
));
3996 file
= io_uring_get_file(ctx
);
3998 ret
= PTR_ERR(file
);
4002 ret
= __io_uring_add_tctx_node(ctx
);
4005 tctx
= current
->io_uring
;
4008 * Install ring fd as the very last thing, so we don't risk someone
4009 * having closed it before we finish setup
4011 if (p
->flags
& IORING_SETUP_REGISTERED_FD_ONLY
)
4012 ret
= io_ring_add_registered_file(tctx
, file
, 0, IO_RINGFD_REG_MAX
);
4014 ret
= io_uring_install_fd(file
);
4018 trace_io_uring_create(ret
, ctx
, p
->sq_entries
, p
->cq_entries
, p
->flags
);
4021 io_ring_ctx_wait_and_kill(ctx
);
4029 * Sets up an aio uring context, and returns the fd. Applications asks for a
4030 * ring size, we return the actual sq/cq ring sizes (among other things) in the
4031 * params structure passed in.
4033 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
4035 struct io_uring_params p
;
4038 if (copy_from_user(&p
, params
, sizeof(p
)))
4040 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
4045 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
4046 IORING_SETUP_SQ_AFF
| IORING_SETUP_CQSIZE
|
4047 IORING_SETUP_CLAMP
| IORING_SETUP_ATTACH_WQ
|
4048 IORING_SETUP_R_DISABLED
| IORING_SETUP_SUBMIT_ALL
|
4049 IORING_SETUP_COOP_TASKRUN
| IORING_SETUP_TASKRUN_FLAG
|
4050 IORING_SETUP_SQE128
| IORING_SETUP_CQE32
|
4051 IORING_SETUP_SINGLE_ISSUER
| IORING_SETUP_DEFER_TASKRUN
|
4052 IORING_SETUP_NO_MMAP
| IORING_SETUP_REGISTERED_FD_ONLY
|
4053 IORING_SETUP_NO_SQARRAY
))
4056 return io_uring_create(entries
, &p
, params
);
4059 static inline bool io_uring_allowed(void)
4061 int disabled
= READ_ONCE(sysctl_io_uring_disabled
);
4062 kgid_t io_uring_group
;
4067 if (disabled
== 0 || capable(CAP_SYS_ADMIN
))
4070 io_uring_group
= make_kgid(&init_user_ns
, sysctl_io_uring_group
);
4071 if (!gid_valid(io_uring_group
))
4074 return in_group_p(io_uring_group
);
4077 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
4078 struct io_uring_params __user
*, params
)
4080 if (!io_uring_allowed())
4083 return io_uring_setup(entries
, params
);
4086 static int __init
io_uring_init(void)
4088 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4089 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4090 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4093 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4094 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4095 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4096 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4097 BUILD_BUG_ON(sizeof(struct io_uring_sqe
) != 64);
4098 BUILD_BUG_SQE_ELEM(0, __u8
, opcode
);
4099 BUILD_BUG_SQE_ELEM(1, __u8
, flags
);
4100 BUILD_BUG_SQE_ELEM(2, __u16
, ioprio
);
4101 BUILD_BUG_SQE_ELEM(4, __s32
, fd
);
4102 BUILD_BUG_SQE_ELEM(8, __u64
, off
);
4103 BUILD_BUG_SQE_ELEM(8, __u64
, addr2
);
4104 BUILD_BUG_SQE_ELEM(8, __u32
, cmd_op
);
4105 BUILD_BUG_SQE_ELEM(12, __u32
, __pad1
);
4106 BUILD_BUG_SQE_ELEM(16, __u64
, addr
);
4107 BUILD_BUG_SQE_ELEM(16, __u64
, splice_off_in
);
4108 BUILD_BUG_SQE_ELEM(24, __u32
, len
);
4109 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t
, rw_flags
);
4110 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags
);
4111 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32
, rw_flags
);
4112 BUILD_BUG_SQE_ELEM(28, __u32
, fsync_flags
);
4113 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16
, poll_events
);
4114 BUILD_BUG_SQE_ELEM(28, __u32
, poll32_events
);
4115 BUILD_BUG_SQE_ELEM(28, __u32
, sync_range_flags
);
4116 BUILD_BUG_SQE_ELEM(28, __u32
, msg_flags
);
4117 BUILD_BUG_SQE_ELEM(28, __u32
, timeout_flags
);
4118 BUILD_BUG_SQE_ELEM(28, __u32
, accept_flags
);
4119 BUILD_BUG_SQE_ELEM(28, __u32
, cancel_flags
);
4120 BUILD_BUG_SQE_ELEM(28, __u32
, open_flags
);
4121 BUILD_BUG_SQE_ELEM(28, __u32
, statx_flags
);
4122 BUILD_BUG_SQE_ELEM(28, __u32
, fadvise_advice
);
4123 BUILD_BUG_SQE_ELEM(28, __u32
, splice_flags
);
4124 BUILD_BUG_SQE_ELEM(28, __u32
, rename_flags
);
4125 BUILD_BUG_SQE_ELEM(28, __u32
, unlink_flags
);
4126 BUILD_BUG_SQE_ELEM(28, __u32
, hardlink_flags
);
4127 BUILD_BUG_SQE_ELEM(28, __u32
, xattr_flags
);
4128 BUILD_BUG_SQE_ELEM(28, __u32
, msg_ring_flags
);
4129 BUILD_BUG_SQE_ELEM(32, __u64
, user_data
);
4130 BUILD_BUG_SQE_ELEM(40, __u16
, buf_index
);
4131 BUILD_BUG_SQE_ELEM(40, __u16
, buf_group
);
4132 BUILD_BUG_SQE_ELEM(42, __u16
, personality
);
4133 BUILD_BUG_SQE_ELEM(44, __s32
, splice_fd_in
);
4134 BUILD_BUG_SQE_ELEM(44, __u32
, file_index
);
4135 BUILD_BUG_SQE_ELEM(44, __u16
, addr_len
);
4136 BUILD_BUG_SQE_ELEM(46, __u16
, __pad3
[0]);
4137 BUILD_BUG_SQE_ELEM(48, __u64
, addr3
);
4138 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd
);
4139 BUILD_BUG_SQE_ELEM(56, __u64
, __pad2
);
4141 BUILD_BUG_ON(sizeof(struct io_uring_files_update
) !=
4142 sizeof(struct io_uring_rsrc_update
));
4143 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update
) >
4144 sizeof(struct io_uring_rsrc_update2
));
4146 /* ->buf_index is u16 */
4147 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring
, bufs
) != 0);
4148 BUILD_BUG_ON(offsetof(struct io_uring_buf
, resv
) !=
4149 offsetof(struct io_uring_buf_ring
, tail
));
4151 /* should fit into one byte */
4152 BUILD_BUG_ON(SQE_VALID_FLAGS
>= (1 << 8));
4153 BUILD_BUG_ON(SQE_COMMON_FLAGS
>= (1 << 8));
4154 BUILD_BUG_ON((SQE_VALID_FLAGS
| SQE_COMMON_FLAGS
) != SQE_VALID_FLAGS
);
4156 BUILD_BUG_ON(__REQ_F_LAST_BIT
> 8 * sizeof(int));
4158 BUILD_BUG_ON(sizeof(atomic_t
) != sizeof(u32
));
4160 /* top 8bits are for internal use */
4161 BUILD_BUG_ON((IORING_URING_CMD_MASK
& 0xff000000) != 0);
4163 io_uring_optable_init();
4166 * Allow user copy in the per-command field, which starts after the
4167 * file in io_kiocb and until the opcode field. The openat2 handling
4168 * requires copying in user memory into the io_kiocb object in that
4169 * range, and HARDENED_USERCOPY will complain if we haven't
4170 * correctly annotated this range.
4172 req_cachep
= kmem_cache_create_usercopy("io_kiocb",
4173 sizeof(struct io_kiocb
), 0,
4174 SLAB_HWCACHE_ALIGN
| SLAB_PANIC
|
4175 SLAB_ACCOUNT
| SLAB_TYPESAFE_BY_RCU
,
4176 offsetof(struct io_kiocb
, cmd
.data
),
4177 sizeof_field(struct io_kiocb
, cmd
.data
), NULL
);
4178 io_buf_cachep
= kmem_cache_create("io_buffer", sizeof(struct io_buffer
), 0,
4179 SLAB_HWCACHE_ALIGN
| SLAB_PANIC
| SLAB_ACCOUNT
,
4182 #ifdef CONFIG_SYSCTL
4183 register_sysctl_init("kernel", kernel_io_uring_disabled_table
);
4188 __initcall(io_uring_init
);