1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <net/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
49 #include <linux/bits.h>
51 #include <linux/sched/signal.h>
53 #include <linux/file.h>
54 #include <linux/fdtable.h>
56 #include <linux/mman.h>
57 #include <linux/percpu.h>
58 #include <linux/slab.h>
59 #include <linux/bvec.h>
60 #include <linux/net.h>
62 #include <net/af_unix.h>
64 #include <linux/anon_inodes.h>
65 #include <linux/sched/mm.h>
66 #include <linux/uaccess.h>
67 #include <linux/nospec.h>
68 #include <linux/highmem.h>
69 #include <linux/fsnotify.h>
70 #include <linux/fadvise.h>
71 #include <linux/task_work.h>
72 #include <linux/io_uring.h>
73 #include <linux/audit.h>
74 #include <linux/security.h>
76 #define CREATE_TRACE_POINTS
77 #include <trace/events/io_uring.h>
79 #include <uapi/linux/io_uring.h>
97 #include "alloc_cache.h"
99 #define IORING_MAX_ENTRIES 32768
100 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
102 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
103 IORING_REGISTER_LAST + IORING_OP_LAST)
105 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
106 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
108 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
109 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
111 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
112 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
115 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
118 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
120 #define IO_COMPL_BATCH 32
121 #define IO_REQ_ALLOC_BATCH 8
124 IO_CHECK_CQ_OVERFLOW_BIT
,
125 IO_CHECK_CQ_DROPPED_BIT
,
128 struct io_defer_entry
{
129 struct list_head list
;
130 struct io_kiocb
*req
;
134 /* requests with any of those set should undergo io_disarm_next() */
135 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
136 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
138 static bool io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
139 struct task_struct
*task
,
142 static void io_dismantle_req(struct io_kiocb
*req
);
143 static void io_clean_op(struct io_kiocb
*req
);
144 static void io_queue_sqe(struct io_kiocb
*req
);
146 static void __io_submit_flush_completions(struct io_ring_ctx
*ctx
);
148 static struct kmem_cache
*req_cachep
;
150 struct sock
*io_uring_get_socket(struct file
*file
)
152 #if defined(CONFIG_UNIX)
153 if (io_is_uring_fops(file
)) {
154 struct io_ring_ctx
*ctx
= file
->private_data
;
156 return ctx
->ring_sock
->sk
;
161 EXPORT_SYMBOL(io_uring_get_socket
);
163 static inline void io_submit_flush_completions(struct io_ring_ctx
*ctx
)
165 if (!wq_list_empty(&ctx
->submit_state
.compl_reqs
))
166 __io_submit_flush_completions(ctx
);
169 static inline unsigned int __io_cqring_events(struct io_ring_ctx
*ctx
)
171 return ctx
->cached_cq_tail
- READ_ONCE(ctx
->rings
->cq
.head
);
174 static bool io_match_linked(struct io_kiocb
*head
)
176 struct io_kiocb
*req
;
178 io_for_each_link(req
, head
) {
179 if (req
->flags
& REQ_F_INFLIGHT
)
186 * As io_match_task() but protected against racing with linked timeouts.
187 * User must not hold timeout_lock.
189 bool io_match_task_safe(struct io_kiocb
*head
, struct task_struct
*task
,
194 if (task
&& head
->task
!= task
)
199 if (head
->flags
& REQ_F_LINK_TIMEOUT
) {
200 struct io_ring_ctx
*ctx
= head
->ctx
;
202 /* protect against races with linked timeouts */
203 spin_lock_irq(&ctx
->timeout_lock
);
204 matched
= io_match_linked(head
);
205 spin_unlock_irq(&ctx
->timeout_lock
);
207 matched
= io_match_linked(head
);
212 static inline void req_fail_link_node(struct io_kiocb
*req
, int res
)
215 io_req_set_res(req
, res
, 0);
218 static inline void io_req_add_to_cache(struct io_kiocb
*req
, struct io_ring_ctx
*ctx
)
220 wq_stack_add_head(&req
->comp_list
, &ctx
->submit_state
.free_list
);
223 static __cold
void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
225 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
227 complete(&ctx
->ref_comp
);
230 static __cold
void io_fallback_req_func(struct work_struct
*work
)
232 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
,
234 struct llist_node
*node
= llist_del_all(&ctx
->fallback_llist
);
235 struct io_kiocb
*req
, *tmp
;
238 percpu_ref_get(&ctx
->refs
);
239 llist_for_each_entry_safe(req
, tmp
, node
, io_task_work
.node
)
240 req
->io_task_work
.func(req
, &locked
);
243 io_submit_flush_completions(ctx
);
244 mutex_unlock(&ctx
->uring_lock
);
246 percpu_ref_put(&ctx
->refs
);
249 static int io_alloc_hash_table(struct io_hash_table
*table
, unsigned bits
)
251 unsigned hash_buckets
= 1U << bits
;
252 size_t hash_size
= hash_buckets
* sizeof(table
->hbs
[0]);
254 table
->hbs
= kmalloc(hash_size
, GFP_KERNEL
);
258 table
->hash_bits
= bits
;
259 init_hash_table(table
, hash_buckets
);
263 static __cold
struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
265 struct io_ring_ctx
*ctx
;
268 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
272 xa_init(&ctx
->io_bl_xa
);
275 * Use 5 bits less than the max cq entries, that should give us around
276 * 32 entries per hash list if totally full and uniformly spread, but
277 * don't keep too many buckets to not overconsume memory.
279 hash_bits
= ilog2(p
->cq_entries
) - 5;
280 hash_bits
= clamp(hash_bits
, 1, 8);
281 if (io_alloc_hash_table(&ctx
->cancel_table
, hash_bits
))
283 if (io_alloc_hash_table(&ctx
->cancel_table_locked
, hash_bits
))
286 ctx
->dummy_ubuf
= kzalloc(sizeof(*ctx
->dummy_ubuf
), GFP_KERNEL
);
287 if (!ctx
->dummy_ubuf
)
289 /* set invalid range, so io_import_fixed() fails meeting it */
290 ctx
->dummy_ubuf
->ubuf
= -1UL;
292 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
,
296 ctx
->flags
= p
->flags
;
297 init_waitqueue_head(&ctx
->sqo_sq_wait
);
298 INIT_LIST_HEAD(&ctx
->sqd_list
);
299 INIT_LIST_HEAD(&ctx
->cq_overflow_list
);
300 INIT_LIST_HEAD(&ctx
->io_buffers_cache
);
301 io_alloc_cache_init(&ctx
->apoll_cache
);
302 io_alloc_cache_init(&ctx
->netmsg_cache
);
303 init_completion(&ctx
->ref_comp
);
304 xa_init_flags(&ctx
->personalities
, XA_FLAGS_ALLOC1
);
305 mutex_init(&ctx
->uring_lock
);
306 init_waitqueue_head(&ctx
->cq_wait
);
307 spin_lock_init(&ctx
->completion_lock
);
308 spin_lock_init(&ctx
->timeout_lock
);
309 INIT_WQ_LIST(&ctx
->iopoll_list
);
310 INIT_LIST_HEAD(&ctx
->io_buffers_pages
);
311 INIT_LIST_HEAD(&ctx
->io_buffers_comp
);
312 INIT_LIST_HEAD(&ctx
->defer_list
);
313 INIT_LIST_HEAD(&ctx
->timeout_list
);
314 INIT_LIST_HEAD(&ctx
->ltimeout_list
);
315 spin_lock_init(&ctx
->rsrc_ref_lock
);
316 INIT_LIST_HEAD(&ctx
->rsrc_ref_list
);
317 INIT_DELAYED_WORK(&ctx
->rsrc_put_work
, io_rsrc_put_work
);
318 init_llist_head(&ctx
->rsrc_put_llist
);
319 INIT_LIST_HEAD(&ctx
->tctx_list
);
320 ctx
->submit_state
.free_list
.next
= NULL
;
321 INIT_WQ_LIST(&ctx
->locked_free_list
);
322 INIT_DELAYED_WORK(&ctx
->fallback_work
, io_fallback_req_func
);
323 INIT_WQ_LIST(&ctx
->submit_state
.compl_reqs
);
326 kfree(ctx
->dummy_ubuf
);
327 kfree(ctx
->cancel_table
.hbs
);
328 kfree(ctx
->cancel_table_locked
.hbs
);
330 xa_destroy(&ctx
->io_bl_xa
);
335 static void io_account_cq_overflow(struct io_ring_ctx
*ctx
)
337 struct io_rings
*r
= ctx
->rings
;
339 WRITE_ONCE(r
->cq_overflow
, READ_ONCE(r
->cq_overflow
) + 1);
343 static bool req_need_defer(struct io_kiocb
*req
, u32 seq
)
345 if (unlikely(req
->flags
& REQ_F_IO_DRAIN
)) {
346 struct io_ring_ctx
*ctx
= req
->ctx
;
348 return seq
+ READ_ONCE(ctx
->cq_extra
) != ctx
->cached_cq_tail
;
354 static inline void io_req_track_inflight(struct io_kiocb
*req
)
356 if (!(req
->flags
& REQ_F_INFLIGHT
)) {
357 req
->flags
|= REQ_F_INFLIGHT
;
358 atomic_inc(&req
->task
->io_uring
->inflight_tracked
);
362 static struct io_kiocb
*__io_prep_linked_timeout(struct io_kiocb
*req
)
364 if (WARN_ON_ONCE(!req
->link
))
367 req
->flags
&= ~REQ_F_ARM_LTIMEOUT
;
368 req
->flags
|= REQ_F_LINK_TIMEOUT
;
370 /* linked timeouts should have two refs once prep'ed */
371 io_req_set_refcount(req
);
372 __io_req_set_refcount(req
->link
, 2);
376 static inline struct io_kiocb
*io_prep_linked_timeout(struct io_kiocb
*req
)
378 if (likely(!(req
->flags
& REQ_F_ARM_LTIMEOUT
)))
380 return __io_prep_linked_timeout(req
);
383 static noinline
void __io_arm_ltimeout(struct io_kiocb
*req
)
385 io_queue_linked_timeout(__io_prep_linked_timeout(req
));
388 static inline void io_arm_ltimeout(struct io_kiocb
*req
)
390 if (unlikely(req
->flags
& REQ_F_ARM_LTIMEOUT
))
391 __io_arm_ltimeout(req
);
394 static void io_prep_async_work(struct io_kiocb
*req
)
396 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
397 struct io_ring_ctx
*ctx
= req
->ctx
;
399 if (!(req
->flags
& REQ_F_CREDS
)) {
400 req
->flags
|= REQ_F_CREDS
;
401 req
->creds
= get_current_cred();
404 req
->work
.list
.next
= NULL
;
406 req
->work
.cancel_seq
= atomic_read(&ctx
->cancel_seq
);
407 if (req
->flags
& REQ_F_FORCE_ASYNC
)
408 req
->work
.flags
|= IO_WQ_WORK_CONCURRENT
;
410 if (req
->file
&& !io_req_ffs_set(req
))
411 req
->flags
|= io_file_get_flags(req
->file
) << REQ_F_SUPPORT_NOWAIT_BIT
;
413 if (req
->flags
& REQ_F_ISREG
) {
414 if (def
->hash_reg_file
|| (ctx
->flags
& IORING_SETUP_IOPOLL
))
415 io_wq_hash_work(&req
->work
, file_inode(req
->file
));
416 } else if (!req
->file
|| !S_ISBLK(file_inode(req
->file
)->i_mode
)) {
417 if (def
->unbound_nonreg_file
)
418 req
->work
.flags
|= IO_WQ_WORK_UNBOUND
;
422 static void io_prep_async_link(struct io_kiocb
*req
)
424 struct io_kiocb
*cur
;
426 if (req
->flags
& REQ_F_LINK_TIMEOUT
) {
427 struct io_ring_ctx
*ctx
= req
->ctx
;
429 spin_lock_irq(&ctx
->timeout_lock
);
430 io_for_each_link(cur
, req
)
431 io_prep_async_work(cur
);
432 spin_unlock_irq(&ctx
->timeout_lock
);
434 io_for_each_link(cur
, req
)
435 io_prep_async_work(cur
);
439 void io_queue_iowq(struct io_kiocb
*req
, bool *dont_use
)
441 struct io_kiocb
*link
= io_prep_linked_timeout(req
);
442 struct io_uring_task
*tctx
= req
->task
->io_uring
;
445 BUG_ON(!tctx
->io_wq
);
447 /* init ->work of the whole link before punting */
448 io_prep_async_link(req
);
451 * Not expected to happen, but if we do have a bug where this _can_
452 * happen, catch it here and ensure the request is marked as
453 * canceled. That will make io-wq go through the usual work cancel
454 * procedure rather than attempt to run this request (or create a new
457 if (WARN_ON_ONCE(!same_thread_group(req
->task
, current
)))
458 req
->work
.flags
|= IO_WQ_WORK_CANCEL
;
460 trace_io_uring_queue_async_work(req
, io_wq_is_hashed(&req
->work
));
461 io_wq_enqueue(tctx
->io_wq
, &req
->work
);
463 io_queue_linked_timeout(link
);
466 static __cold
void io_queue_deferred(struct io_ring_ctx
*ctx
)
468 while (!list_empty(&ctx
->defer_list
)) {
469 struct io_defer_entry
*de
= list_first_entry(&ctx
->defer_list
,
470 struct io_defer_entry
, list
);
472 if (req_need_defer(de
->req
, de
->seq
))
474 list_del_init(&de
->list
);
475 io_req_task_queue(de
->req
);
480 static void io_eventfd_signal(struct io_ring_ctx
*ctx
)
482 struct io_ev_fd
*ev_fd
;
485 spin_lock(&ctx
->completion_lock
);
487 * Eventfd should only get triggered when at least one event has been
488 * posted. Some applications rely on the eventfd notification count only
489 * changing IFF a new CQE has been added to the CQ ring. There's no
490 * depedency on 1:1 relationship between how many times this function is
491 * called (and hence the eventfd count) and number of CQEs posted to the
494 skip
= ctx
->cached_cq_tail
== ctx
->evfd_last_cq_tail
;
495 ctx
->evfd_last_cq_tail
= ctx
->cached_cq_tail
;
496 spin_unlock(&ctx
->completion_lock
);
502 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
505 ev_fd
= rcu_dereference(ctx
->io_ev_fd
);
508 * Check again if ev_fd exists incase an io_eventfd_unregister call
509 * completed between the NULL check of ctx->io_ev_fd at the start of
510 * the function and rcu_read_lock.
512 if (unlikely(!ev_fd
))
514 if (READ_ONCE(ctx
->rings
->cq_flags
) & IORING_CQ_EVENTFD_DISABLED
)
517 if (!ev_fd
->eventfd_async
|| io_wq_current_is_worker())
518 eventfd_signal(ev_fd
->cq_ev_fd
, 1);
523 void __io_commit_cqring_flush(struct io_ring_ctx
*ctx
)
525 if (ctx
->off_timeout_used
|| ctx
->drain_active
) {
526 spin_lock(&ctx
->completion_lock
);
527 if (ctx
->off_timeout_used
)
528 io_flush_timeouts(ctx
);
529 if (ctx
->drain_active
)
530 io_queue_deferred(ctx
);
531 spin_unlock(&ctx
->completion_lock
);
534 io_eventfd_signal(ctx
);
537 static inline void io_cqring_ev_posted(struct io_ring_ctx
*ctx
)
539 io_commit_cqring_flush(ctx
);
543 static inline void __io_cq_unlock_post(struct io_ring_ctx
*ctx
)
544 __releases(ctx
->completion_lock
)
546 io_commit_cqring(ctx
);
547 spin_unlock(&ctx
->completion_lock
);
548 io_cqring_ev_posted(ctx
);
551 void io_cq_unlock_post(struct io_ring_ctx
*ctx
)
553 __io_cq_unlock_post(ctx
);
556 /* Returns true if there are no backlogged entries after the flush */
557 static bool __io_cqring_overflow_flush(struct io_ring_ctx
*ctx
, bool force
)
560 size_t cqe_size
= sizeof(struct io_uring_cqe
);
562 if (!force
&& __io_cqring_events(ctx
) == ctx
->cq_entries
)
565 if (ctx
->flags
& IORING_SETUP_CQE32
)
569 while (!list_empty(&ctx
->cq_overflow_list
)) {
570 struct io_uring_cqe
*cqe
= io_get_cqe(ctx
);
571 struct io_overflow_cqe
*ocqe
;
575 ocqe
= list_first_entry(&ctx
->cq_overflow_list
,
576 struct io_overflow_cqe
, list
);
578 memcpy(cqe
, &ocqe
->cqe
, cqe_size
);
580 io_account_cq_overflow(ctx
);
582 list_del(&ocqe
->list
);
586 all_flushed
= list_empty(&ctx
->cq_overflow_list
);
588 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
);
589 atomic_andnot(IORING_SQ_CQ_OVERFLOW
, &ctx
->rings
->sq_flags
);
592 io_cq_unlock_post(ctx
);
596 static bool io_cqring_overflow_flush(struct io_ring_ctx
*ctx
)
600 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
)) {
601 /* iopoll syncs against uring_lock, not completion_lock */
602 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
603 mutex_lock(&ctx
->uring_lock
);
604 ret
= __io_cqring_overflow_flush(ctx
, false);
605 if (ctx
->flags
& IORING_SETUP_IOPOLL
)
606 mutex_unlock(&ctx
->uring_lock
);
612 void __io_put_task(struct task_struct
*task
, int nr
)
614 struct io_uring_task
*tctx
= task
->io_uring
;
616 percpu_counter_sub(&tctx
->inflight
, nr
);
617 if (unlikely(atomic_read(&tctx
->in_idle
)))
618 wake_up(&tctx
->wait
);
619 put_task_struct_many(task
, nr
);
622 void io_task_refs_refill(struct io_uring_task
*tctx
)
624 unsigned int refill
= -tctx
->cached_refs
+ IO_TCTX_REFS_CACHE_NR
;
626 percpu_counter_add(&tctx
->inflight
, refill
);
627 refcount_add(refill
, ¤t
->usage
);
628 tctx
->cached_refs
+= refill
;
631 static __cold
void io_uring_drop_tctx_refs(struct task_struct
*task
)
633 struct io_uring_task
*tctx
= task
->io_uring
;
634 unsigned int refs
= tctx
->cached_refs
;
637 tctx
->cached_refs
= 0;
638 percpu_counter_sub(&tctx
->inflight
, refs
);
639 put_task_struct_many(task
, refs
);
643 static bool io_cqring_event_overflow(struct io_ring_ctx
*ctx
, u64 user_data
,
644 s32 res
, u32 cflags
, u64 extra1
, u64 extra2
)
646 struct io_overflow_cqe
*ocqe
;
647 size_t ocq_size
= sizeof(struct io_overflow_cqe
);
648 bool is_cqe32
= (ctx
->flags
& IORING_SETUP_CQE32
);
651 ocq_size
+= sizeof(struct io_uring_cqe
);
653 ocqe
= kmalloc(ocq_size
, GFP_ATOMIC
| __GFP_ACCOUNT
);
654 trace_io_uring_cqe_overflow(ctx
, user_data
, res
, cflags
, ocqe
);
657 * If we're in ring overflow flush mode, or in task cancel mode,
658 * or cannot allocate an overflow entry, then we need to drop it
661 io_account_cq_overflow(ctx
);
662 set_bit(IO_CHECK_CQ_DROPPED_BIT
, &ctx
->check_cq
);
665 if (list_empty(&ctx
->cq_overflow_list
)) {
666 set_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
);
667 atomic_or(IORING_SQ_CQ_OVERFLOW
, &ctx
->rings
->sq_flags
);
670 ocqe
->cqe
.user_data
= user_data
;
672 ocqe
->cqe
.flags
= cflags
;
674 ocqe
->cqe
.big_cqe
[0] = extra1
;
675 ocqe
->cqe
.big_cqe
[1] = extra2
;
677 list_add_tail(&ocqe
->list
, &ctx
->cq_overflow_list
);
681 bool io_req_cqe_overflow(struct io_kiocb
*req
)
683 if (!(req
->flags
& REQ_F_CQE32_INIT
)) {
687 return io_cqring_event_overflow(req
->ctx
, req
->cqe
.user_data
,
688 req
->cqe
.res
, req
->cqe
.flags
,
689 req
->extra1
, req
->extra2
);
693 * writes to the cq entry need to come after reading head; the
694 * control dependency is enough as we're using WRITE_ONCE to
697 struct io_uring_cqe
*__io_get_cqe(struct io_ring_ctx
*ctx
)
699 struct io_rings
*rings
= ctx
->rings
;
700 unsigned int off
= ctx
->cached_cq_tail
& (ctx
->cq_entries
- 1);
701 unsigned int free
, queued
, len
;
704 /* userspace may cheat modifying the tail, be safe and do min */
705 queued
= min(__io_cqring_events(ctx
), ctx
->cq_entries
);
706 free
= ctx
->cq_entries
- queued
;
707 /* we need a contiguous range, limit based on the current array offset */
708 len
= min(free
, ctx
->cq_entries
- off
);
712 if (ctx
->flags
& IORING_SETUP_CQE32
) {
717 ctx
->cqe_cached
= &rings
->cqes
[off
];
718 ctx
->cqe_sentinel
= ctx
->cqe_cached
+ len
;
720 ctx
->cached_cq_tail
++;
722 if (ctx
->flags
& IORING_SETUP_CQE32
)
724 return &rings
->cqes
[off
];
727 bool io_fill_cqe_aux(struct io_ring_ctx
*ctx
, u64 user_data
, s32 res
, u32 cflags
,
730 struct io_uring_cqe
*cqe
;
735 * If we can't get a cq entry, userspace overflowed the
736 * submission (by quite a lot). Increment the overflow count in
739 cqe
= io_get_cqe(ctx
);
741 trace_io_uring_complete(ctx
, NULL
, user_data
, res
, cflags
, 0, 0);
743 WRITE_ONCE(cqe
->user_data
, user_data
);
744 WRITE_ONCE(cqe
->res
, res
);
745 WRITE_ONCE(cqe
->flags
, cflags
);
747 if (ctx
->flags
& IORING_SETUP_CQE32
) {
748 WRITE_ONCE(cqe
->big_cqe
[0], 0);
749 WRITE_ONCE(cqe
->big_cqe
[1], 0);
755 return io_cqring_event_overflow(ctx
, user_data
, res
, cflags
, 0, 0);
760 bool io_post_aux_cqe(struct io_ring_ctx
*ctx
,
761 u64 user_data
, s32 res
, u32 cflags
,
767 filled
= io_fill_cqe_aux(ctx
, user_data
, res
, cflags
, allow_overflow
);
768 io_cq_unlock_post(ctx
);
772 static void __io_req_complete_put(struct io_kiocb
*req
)
775 * If we're the last reference to this request, add to our locked
778 if (req_ref_put_and_test(req
)) {
779 struct io_ring_ctx
*ctx
= req
->ctx
;
781 if (req
->flags
& IO_REQ_LINK_FLAGS
) {
782 if (req
->flags
& IO_DISARM_MASK
)
785 io_req_task_queue(req
->link
);
789 io_req_put_rsrc(req
);
791 * Selected buffer deallocation in io_clean_op() assumes that
792 * we don't hold ->completion_lock. Clean them here to avoid
795 io_put_kbuf_comp(req
);
796 io_dismantle_req(req
);
797 io_put_task(req
->task
, 1);
798 wq_list_add_head(&req
->comp_list
, &ctx
->locked_free_list
);
799 ctx
->locked_free_nr
++;
803 void __io_req_complete_post(struct io_kiocb
*req
)
805 if (!(req
->flags
& REQ_F_CQE_SKIP
))
806 __io_fill_cqe_req(req
->ctx
, req
);
807 __io_req_complete_put(req
);
810 void io_req_complete_post(struct io_kiocb
*req
)
812 struct io_ring_ctx
*ctx
= req
->ctx
;
815 __io_req_complete_post(req
);
816 io_cq_unlock_post(ctx
);
819 inline void __io_req_complete(struct io_kiocb
*req
, unsigned issue_flags
)
821 io_req_complete_post(req
);
824 void io_req_complete_failed(struct io_kiocb
*req
, s32 res
)
827 io_req_set_res(req
, res
, io_put_kbuf(req
, IO_URING_F_UNLOCKED
));
828 io_req_complete_post(req
);
832 * Don't initialise the fields below on every allocation, but do that in
833 * advance and keep them valid across allocations.
835 static void io_preinit_req(struct io_kiocb
*req
, struct io_ring_ctx
*ctx
)
839 req
->async_data
= NULL
;
840 /* not necessary, but safer to zero */
844 static void io_flush_cached_locked_reqs(struct io_ring_ctx
*ctx
,
845 struct io_submit_state
*state
)
847 spin_lock(&ctx
->completion_lock
);
848 wq_list_splice(&ctx
->locked_free_list
, &state
->free_list
);
849 ctx
->locked_free_nr
= 0;
850 spin_unlock(&ctx
->completion_lock
);
854 * A request might get retired back into the request caches even before opcode
855 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
856 * Because of that, io_alloc_req() should be called only under ->uring_lock
857 * and with extra caution to not get a request that is still worked on.
859 __cold
bool __io_alloc_req_refill(struct io_ring_ctx
*ctx
)
860 __must_hold(&ctx
->uring_lock
)
862 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
863 void *reqs
[IO_REQ_ALLOC_BATCH
];
867 * If we have more than a batch's worth of requests in our IRQ side
868 * locked cache, grab the lock and move them over to our submission
871 if (data_race(ctx
->locked_free_nr
) > IO_COMPL_BATCH
) {
872 io_flush_cached_locked_reqs(ctx
, &ctx
->submit_state
);
873 if (!io_req_cache_empty(ctx
))
877 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, ARRAY_SIZE(reqs
), reqs
);
880 * Bulk alloc is all-or-nothing. If we fail to get a batch,
881 * retry single alloc to be on the safe side.
883 if (unlikely(ret
<= 0)) {
884 reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
890 percpu_ref_get_many(&ctx
->refs
, ret
);
891 for (i
= 0; i
< ret
; i
++) {
892 struct io_kiocb
*req
= reqs
[i
];
894 io_preinit_req(req
, ctx
);
895 io_req_add_to_cache(req
, ctx
);
900 static inline void io_dismantle_req(struct io_kiocb
*req
)
902 unsigned int flags
= req
->flags
;
904 if (unlikely(flags
& IO_REQ_CLEAN_FLAGS
))
906 if (!(flags
& REQ_F_FIXED_FILE
))
907 io_put_file(req
->file
);
910 __cold
void io_free_req(struct io_kiocb
*req
)
912 struct io_ring_ctx
*ctx
= req
->ctx
;
914 io_req_put_rsrc(req
);
915 io_dismantle_req(req
);
916 io_put_task(req
->task
, 1);
918 spin_lock(&ctx
->completion_lock
);
919 wq_list_add_head(&req
->comp_list
, &ctx
->locked_free_list
);
920 ctx
->locked_free_nr
++;
921 spin_unlock(&ctx
->completion_lock
);
924 static void __io_req_find_next_prep(struct io_kiocb
*req
)
926 struct io_ring_ctx
*ctx
= req
->ctx
;
930 io_cq_unlock_post(ctx
);
933 static inline struct io_kiocb
*io_req_find_next(struct io_kiocb
*req
)
935 struct io_kiocb
*nxt
;
938 * If LINK is set, we have dependent requests in this chain. If we
939 * didn't fail this request, queue the first one up, moving any other
940 * dependencies to the next request. In case of failure, fail the rest
943 if (unlikely(req
->flags
& IO_DISARM_MASK
))
944 __io_req_find_next_prep(req
);
950 static void ctx_flush_and_put(struct io_ring_ctx
*ctx
, bool *locked
)
954 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
)
955 atomic_andnot(IORING_SQ_TASKRUN
, &ctx
->rings
->sq_flags
);
957 io_submit_flush_completions(ctx
);
958 mutex_unlock(&ctx
->uring_lock
);
961 percpu_ref_put(&ctx
->refs
);
964 static unsigned int handle_tw_list(struct llist_node
*node
,
965 struct io_ring_ctx
**ctx
, bool *locked
,
966 struct llist_node
*last
)
968 unsigned int count
= 0;
970 while (node
!= last
) {
971 struct llist_node
*next
= node
->next
;
972 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
975 prefetch(container_of(next
, struct io_kiocb
, io_task_work
.node
));
977 if (req
->ctx
!= *ctx
) {
978 ctx_flush_and_put(*ctx
, locked
);
980 /* if not contended, grab and improve batching */
981 *locked
= mutex_trylock(&(*ctx
)->uring_lock
);
982 percpu_ref_get(&(*ctx
)->refs
);
984 req
->io_task_work
.func(req
, locked
);
993 * io_llist_xchg - swap all entries in a lock-less list
994 * @head: the head of lock-less list to delete all entries
995 * @new: new entry as the head of the list
997 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
998 * The order of entries returned is from the newest to the oldest added one.
1000 static inline struct llist_node
*io_llist_xchg(struct llist_head
*head
,
1001 struct llist_node
*new)
1003 return xchg(&head
->first
, new);
1007 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1008 * @head: the head of lock-less list to delete all entries
1009 * @old: expected old value of the first entry of the list
1010 * @new: new entry as the head of the list
1012 * perform a cmpxchg on the first entry of the list.
1015 static inline struct llist_node
*io_llist_cmpxchg(struct llist_head
*head
,
1016 struct llist_node
*old
,
1017 struct llist_node
*new)
1019 return cmpxchg(&head
->first
, old
, new);
1022 void tctx_task_work(struct callback_head
*cb
)
1024 bool uring_locked
= false;
1025 struct io_ring_ctx
*ctx
= NULL
;
1026 struct io_uring_task
*tctx
= container_of(cb
, struct io_uring_task
,
1028 struct llist_node fake
= {};
1029 struct llist_node
*node
= io_llist_xchg(&tctx
->task_list
, &fake
);
1030 unsigned int loops
= 1;
1031 unsigned int count
= handle_tw_list(node
, &ctx
, &uring_locked
, NULL
);
1033 node
= io_llist_cmpxchg(&tctx
->task_list
, &fake
, NULL
);
1034 while (node
!= &fake
) {
1036 node
= io_llist_xchg(&tctx
->task_list
, &fake
);
1037 count
+= handle_tw_list(node
, &ctx
, &uring_locked
, &fake
);
1038 node
= io_llist_cmpxchg(&tctx
->task_list
, &fake
, NULL
);
1041 ctx_flush_and_put(ctx
, &uring_locked
);
1043 /* relaxed read is enough as only the task itself sets ->in_idle */
1044 if (unlikely(atomic_read(&tctx
->in_idle
)))
1045 io_uring_drop_tctx_refs(current
);
1047 trace_io_uring_task_work_run(tctx
, count
, loops
);
1050 void io_req_task_work_add(struct io_kiocb
*req
)
1052 struct io_uring_task
*tctx
= req
->task
->io_uring
;
1053 struct io_ring_ctx
*ctx
= req
->ctx
;
1054 struct llist_node
*node
;
1057 running
= !llist_add(&req
->io_task_work
.node
, &tctx
->task_list
);
1059 /* task_work already pending, we're done */
1063 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
)
1064 atomic_or(IORING_SQ_TASKRUN
, &ctx
->rings
->sq_flags
);
1066 if (likely(!task_work_add(req
->task
, &tctx
->task_work
, ctx
->notify_method
)))
1069 node
= llist_del_all(&tctx
->task_list
);
1072 req
= container_of(node
, struct io_kiocb
, io_task_work
.node
);
1074 if (llist_add(&req
->io_task_work
.node
,
1075 &req
->ctx
->fallback_llist
))
1076 schedule_delayed_work(&req
->ctx
->fallback_work
, 1);
1080 static void io_req_tw_post(struct io_kiocb
*req
, bool *locked
)
1082 io_req_complete_post(req
);
1085 void io_req_tw_post_queue(struct io_kiocb
*req
, s32 res
, u32 cflags
)
1087 io_req_set_res(req
, res
, cflags
);
1088 req
->io_task_work
.func
= io_req_tw_post
;
1089 io_req_task_work_add(req
);
1092 static void io_req_task_cancel(struct io_kiocb
*req
, bool *locked
)
1094 /* not needed for normal modes, but SQPOLL depends on it */
1095 io_tw_lock(req
->ctx
, locked
);
1096 io_req_complete_failed(req
, req
->cqe
.res
);
1099 void io_req_task_submit(struct io_kiocb
*req
, bool *locked
)
1101 io_tw_lock(req
->ctx
, locked
);
1102 /* req->task == current here, checking PF_EXITING is safe */
1103 if (likely(!(req
->task
->flags
& PF_EXITING
)))
1106 io_req_complete_failed(req
, -EFAULT
);
1109 void io_req_task_queue_fail(struct io_kiocb
*req
, int ret
)
1111 io_req_set_res(req
, ret
, 0);
1112 req
->io_task_work
.func
= io_req_task_cancel
;
1113 io_req_task_work_add(req
);
1116 void io_req_task_queue(struct io_kiocb
*req
)
1118 req
->io_task_work
.func
= io_req_task_submit
;
1119 io_req_task_work_add(req
);
1122 void io_queue_next(struct io_kiocb
*req
)
1124 struct io_kiocb
*nxt
= io_req_find_next(req
);
1127 io_req_task_queue(nxt
);
1130 void io_free_batch_list(struct io_ring_ctx
*ctx
, struct io_wq_work_node
*node
)
1131 __must_hold(&ctx
->uring_lock
)
1133 struct task_struct
*task
= NULL
;
1137 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1140 if (unlikely(req
->flags
& IO_REQ_CLEAN_SLOW_FLAGS
)) {
1141 if (req
->flags
& REQ_F_REFCOUNT
) {
1142 node
= req
->comp_list
.next
;
1143 if (!req_ref_put_and_test(req
))
1146 if ((req
->flags
& REQ_F_POLLED
) && req
->apoll
) {
1147 struct async_poll
*apoll
= req
->apoll
;
1149 if (apoll
->double_poll
)
1150 kfree(apoll
->double_poll
);
1151 if (!io_alloc_cache_put(&ctx
->apoll_cache
, &apoll
->cache
))
1153 req
->flags
&= ~REQ_F_POLLED
;
1155 if (req
->flags
& IO_REQ_LINK_FLAGS
)
1157 if (unlikely(req
->flags
& IO_REQ_CLEAN_FLAGS
))
1160 if (!(req
->flags
& REQ_F_FIXED_FILE
))
1161 io_put_file(req
->file
);
1163 io_req_put_rsrc_locked(req
, ctx
);
1165 if (req
->task
!= task
) {
1167 io_put_task(task
, task_refs
);
1172 node
= req
->comp_list
.next
;
1173 io_req_add_to_cache(req
, ctx
);
1177 io_put_task(task
, task_refs
);
1180 static void __io_submit_flush_completions(struct io_ring_ctx
*ctx
)
1181 __must_hold(&ctx
->uring_lock
)
1183 struct io_wq_work_node
*node
, *prev
;
1184 struct io_submit_state
*state
= &ctx
->submit_state
;
1186 spin_lock(&ctx
->completion_lock
);
1187 wq_list_for_each(node
, prev
, &state
->compl_reqs
) {
1188 struct io_kiocb
*req
= container_of(node
, struct io_kiocb
,
1191 if (!(req
->flags
& REQ_F_CQE_SKIP
))
1192 __io_fill_cqe_req(ctx
, req
);
1194 __io_cq_unlock_post(ctx
);
1196 io_free_batch_list(ctx
, state
->compl_reqs
.first
);
1197 INIT_WQ_LIST(&state
->compl_reqs
);
1201 * Drop reference to request, return next in chain (if there is one) if this
1202 * was the last reference to this request.
1204 static inline struct io_kiocb
*io_put_req_find_next(struct io_kiocb
*req
)
1206 struct io_kiocb
*nxt
= NULL
;
1208 if (req_ref_put_and_test(req
)) {
1209 if (unlikely(req
->flags
& IO_REQ_LINK_FLAGS
))
1210 nxt
= io_req_find_next(req
);
1216 static unsigned io_cqring_events(struct io_ring_ctx
*ctx
)
1218 /* See comment at the top of this file */
1220 return __io_cqring_events(ctx
);
1224 * We can't just wait for polled events to come to us, we have to actively
1225 * find and complete them.
1227 static __cold
void io_iopoll_try_reap_events(struct io_ring_ctx
*ctx
)
1229 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
1232 mutex_lock(&ctx
->uring_lock
);
1233 while (!wq_list_empty(&ctx
->iopoll_list
)) {
1234 /* let it sleep and repeat later if can't complete a request */
1235 if (io_do_iopoll(ctx
, true) == 0)
1238 * Ensure we allow local-to-the-cpu processing to take place,
1239 * in this case we need to ensure that we reap all events.
1240 * Also let task_work, etc. to progress by releasing the mutex
1242 if (need_resched()) {
1243 mutex_unlock(&ctx
->uring_lock
);
1245 mutex_lock(&ctx
->uring_lock
);
1248 mutex_unlock(&ctx
->uring_lock
);
1251 static int io_iopoll_check(struct io_ring_ctx
*ctx
, long min
)
1253 unsigned int nr_events
= 0;
1255 unsigned long check_cq
;
1257 check_cq
= READ_ONCE(ctx
->check_cq
);
1258 if (unlikely(check_cq
)) {
1259 if (check_cq
& BIT(IO_CHECK_CQ_OVERFLOW_BIT
))
1260 __io_cqring_overflow_flush(ctx
, false);
1262 * Similarly do not spin if we have not informed the user of any
1265 if (check_cq
& BIT(IO_CHECK_CQ_DROPPED_BIT
))
1269 * Don't enter poll loop if we already have events pending.
1270 * If we do, we can potentially be spinning for commands that
1271 * already triggered a CQE (eg in error).
1273 if (io_cqring_events(ctx
))
1278 * If a submit got punted to a workqueue, we can have the
1279 * application entering polling for a command before it gets
1280 * issued. That app will hold the uring_lock for the duration
1281 * of the poll right here, so we need to take a breather every
1282 * now and then to ensure that the issue has a chance to add
1283 * the poll to the issued list. Otherwise we can spin here
1284 * forever, while the workqueue is stuck trying to acquire the
1287 if (wq_list_empty(&ctx
->iopoll_list
)) {
1288 u32 tail
= ctx
->cached_cq_tail
;
1290 mutex_unlock(&ctx
->uring_lock
);
1292 mutex_lock(&ctx
->uring_lock
);
1294 /* some requests don't go through iopoll_list */
1295 if (tail
!= ctx
->cached_cq_tail
||
1296 wq_list_empty(&ctx
->iopoll_list
))
1299 ret
= io_do_iopoll(ctx
, !min
);
1304 } while (nr_events
< min
&& !need_resched());
1309 void io_req_task_complete(struct io_kiocb
*req
, bool *locked
)
1311 if (req
->flags
& (REQ_F_BUFFER_SELECTED
|REQ_F_BUFFER_RING
)) {
1312 unsigned issue_flags
= *locked
? 0 : IO_URING_F_UNLOCKED
;
1314 req
->cqe
.flags
|= io_put_kbuf(req
, issue_flags
);
1318 io_req_complete_defer(req
);
1320 io_req_complete_post(req
);
1324 * After the iocb has been issued, it's safe to be found on the poll list.
1325 * Adding the kiocb to the list AFTER submission ensures that we don't
1326 * find it from a io_do_iopoll() thread before the issuer is done
1327 * accessing the kiocb cookie.
1329 static void io_iopoll_req_issued(struct io_kiocb
*req
, unsigned int issue_flags
)
1331 struct io_ring_ctx
*ctx
= req
->ctx
;
1332 const bool needs_lock
= issue_flags
& IO_URING_F_UNLOCKED
;
1334 /* workqueue context doesn't hold uring_lock, grab it now */
1335 if (unlikely(needs_lock
))
1336 mutex_lock(&ctx
->uring_lock
);
1339 * Track whether we have multiple files in our lists. This will impact
1340 * how we do polling eventually, not spinning if we're on potentially
1341 * different devices.
1343 if (wq_list_empty(&ctx
->iopoll_list
)) {
1344 ctx
->poll_multi_queue
= false;
1345 } else if (!ctx
->poll_multi_queue
) {
1346 struct io_kiocb
*list_req
;
1348 list_req
= container_of(ctx
->iopoll_list
.first
, struct io_kiocb
,
1350 if (list_req
->file
!= req
->file
)
1351 ctx
->poll_multi_queue
= true;
1355 * For fast devices, IO may have already completed. If it has, add
1356 * it to the front so we find it first.
1358 if (READ_ONCE(req
->iopoll_completed
))
1359 wq_list_add_head(&req
->comp_list
, &ctx
->iopoll_list
);
1361 wq_list_add_tail(&req
->comp_list
, &ctx
->iopoll_list
);
1363 if (unlikely(needs_lock
)) {
1365 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1366 * in sq thread task context or in io worker task context. If
1367 * current task context is sq thread, we don't need to check
1368 * whether should wake up sq thread.
1370 if ((ctx
->flags
& IORING_SETUP_SQPOLL
) &&
1371 wq_has_sleeper(&ctx
->sq_data
->wait
))
1372 wake_up(&ctx
->sq_data
->wait
);
1374 mutex_unlock(&ctx
->uring_lock
);
1378 static bool io_bdev_nowait(struct block_device
*bdev
)
1380 return !bdev
|| blk_queue_nowait(bdev_get_queue(bdev
));
1384 * If we tracked the file through the SCM inflight mechanism, we could support
1385 * any file. For now, just ensure that anything potentially problematic is done
1388 static bool __io_file_supports_nowait(struct file
*file
, umode_t mode
)
1390 if (S_ISBLK(mode
)) {
1391 if (IS_ENABLED(CONFIG_BLOCK
) &&
1392 io_bdev_nowait(I_BDEV(file
->f_mapping
->host
)))
1398 if (S_ISREG(mode
)) {
1399 if (IS_ENABLED(CONFIG_BLOCK
) &&
1400 io_bdev_nowait(file
->f_inode
->i_sb
->s_bdev
) &&
1401 !io_is_uring_fops(file
))
1406 /* any ->read/write should understand O_NONBLOCK */
1407 if (file
->f_flags
& O_NONBLOCK
)
1409 return file
->f_mode
& FMODE_NOWAIT
;
1413 * If we tracked the file through the SCM inflight mechanism, we could support
1414 * any file. For now, just ensure that anything potentially problematic is done
1417 unsigned int io_file_get_flags(struct file
*file
)
1419 umode_t mode
= file_inode(file
)->i_mode
;
1420 unsigned int res
= 0;
1424 if (__io_file_supports_nowait(file
, mode
))
1426 if (io_file_need_scm(file
))
1431 bool io_alloc_async_data(struct io_kiocb
*req
)
1433 WARN_ON_ONCE(!io_op_defs
[req
->opcode
].async_size
);
1434 req
->async_data
= kmalloc(io_op_defs
[req
->opcode
].async_size
, GFP_KERNEL
);
1435 if (req
->async_data
) {
1436 req
->flags
|= REQ_F_ASYNC_DATA
;
1442 int io_req_prep_async(struct io_kiocb
*req
)
1444 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1446 /* assign early for deferred execution for non-fixed file */
1447 if (def
->needs_file
&& !(req
->flags
& REQ_F_FIXED_FILE
))
1448 req
->file
= io_file_get_normal(req
, req
->cqe
.fd
);
1449 if (!def
->prep_async
)
1451 if (WARN_ON_ONCE(req_has_async_data(req
)))
1453 if (!io_op_defs
[req
->opcode
].manual_alloc
) {
1454 if (io_alloc_async_data(req
))
1457 return def
->prep_async(req
);
1460 static u32
io_get_sequence(struct io_kiocb
*req
)
1462 u32 seq
= req
->ctx
->cached_sq_head
;
1463 struct io_kiocb
*cur
;
1465 /* need original cached_sq_head, but it was increased for each req */
1466 io_for_each_link(cur
, req
)
1471 static __cold
void io_drain_req(struct io_kiocb
*req
)
1473 struct io_ring_ctx
*ctx
= req
->ctx
;
1474 struct io_defer_entry
*de
;
1476 u32 seq
= io_get_sequence(req
);
1478 /* Still need defer if there is pending req in defer list. */
1479 spin_lock(&ctx
->completion_lock
);
1480 if (!req_need_defer(req
, seq
) && list_empty_careful(&ctx
->defer_list
)) {
1481 spin_unlock(&ctx
->completion_lock
);
1483 ctx
->drain_active
= false;
1484 io_req_task_queue(req
);
1487 spin_unlock(&ctx
->completion_lock
);
1489 ret
= io_req_prep_async(req
);
1492 io_req_complete_failed(req
, ret
);
1495 io_prep_async_link(req
);
1496 de
= kmalloc(sizeof(*de
), GFP_KERNEL
);
1502 spin_lock(&ctx
->completion_lock
);
1503 if (!req_need_defer(req
, seq
) && list_empty(&ctx
->defer_list
)) {
1504 spin_unlock(&ctx
->completion_lock
);
1509 trace_io_uring_defer(req
);
1512 list_add_tail(&de
->list
, &ctx
->defer_list
);
1513 spin_unlock(&ctx
->completion_lock
);
1516 static void io_clean_op(struct io_kiocb
*req
)
1518 if (req
->flags
& REQ_F_BUFFER_SELECTED
) {
1519 spin_lock(&req
->ctx
->completion_lock
);
1520 io_put_kbuf_comp(req
);
1521 spin_unlock(&req
->ctx
->completion_lock
);
1524 if (req
->flags
& REQ_F_NEED_CLEANUP
) {
1525 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1530 if ((req
->flags
& REQ_F_POLLED
) && req
->apoll
) {
1531 kfree(req
->apoll
->double_poll
);
1535 if (req
->flags
& REQ_F_INFLIGHT
) {
1536 struct io_uring_task
*tctx
= req
->task
->io_uring
;
1538 atomic_dec(&tctx
->inflight_tracked
);
1540 if (req
->flags
& REQ_F_CREDS
)
1541 put_cred(req
->creds
);
1542 if (req
->flags
& REQ_F_ASYNC_DATA
) {
1543 kfree(req
->async_data
);
1544 req
->async_data
= NULL
;
1546 req
->flags
&= ~IO_REQ_CLEAN_FLAGS
;
1549 static bool io_assign_file(struct io_kiocb
*req
, unsigned int issue_flags
)
1551 if (req
->file
|| !io_op_defs
[req
->opcode
].needs_file
)
1554 if (req
->flags
& REQ_F_FIXED_FILE
)
1555 req
->file
= io_file_get_fixed(req
, req
->cqe
.fd
, issue_flags
);
1557 req
->file
= io_file_get_normal(req
, req
->cqe
.fd
);
1562 static int io_issue_sqe(struct io_kiocb
*req
, unsigned int issue_flags
)
1564 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1565 const struct cred
*creds
= NULL
;
1568 if (unlikely(!io_assign_file(req
, issue_flags
)))
1571 if (unlikely((req
->flags
& REQ_F_CREDS
) && req
->creds
!= current_cred()))
1572 creds
= override_creds(req
->creds
);
1574 if (!def
->audit_skip
)
1575 audit_uring_entry(req
->opcode
);
1577 ret
= def
->issue(req
, issue_flags
);
1579 if (!def
->audit_skip
)
1580 audit_uring_exit(!ret
, ret
);
1583 revert_creds(creds
);
1585 if (ret
== IOU_OK
) {
1586 if (issue_flags
& IO_URING_F_COMPLETE_DEFER
)
1587 io_req_complete_defer(req
);
1589 io_req_complete_post(req
);
1590 } else if (ret
!= IOU_ISSUE_SKIP_COMPLETE
)
1593 /* If the op doesn't have a file, we're not polling for it */
1594 if ((req
->ctx
->flags
& IORING_SETUP_IOPOLL
) && req
->file
)
1595 io_iopoll_req_issued(req
, issue_flags
);
1600 int io_poll_issue(struct io_kiocb
*req
, bool *locked
)
1602 io_tw_lock(req
->ctx
, locked
);
1603 if (unlikely(req
->task
->flags
& PF_EXITING
))
1605 return io_issue_sqe(req
, IO_URING_F_NONBLOCK
);
1608 struct io_wq_work
*io_wq_free_work(struct io_wq_work
*work
)
1610 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
1612 req
= io_put_req_find_next(req
);
1613 return req
? &req
->work
: NULL
;
1616 void io_wq_submit_work(struct io_wq_work
*work
)
1618 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
1619 const struct io_op_def
*def
= &io_op_defs
[req
->opcode
];
1620 unsigned int issue_flags
= IO_URING_F_UNLOCKED
;
1621 bool needs_poll
= false;
1622 int ret
= 0, err
= -ECANCELED
;
1624 /* one will be dropped by ->io_free_work() after returning to io-wq */
1625 if (!(req
->flags
& REQ_F_REFCOUNT
))
1626 __io_req_set_refcount(req
, 2);
1630 io_arm_ltimeout(req
);
1632 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1633 if (work
->flags
& IO_WQ_WORK_CANCEL
) {
1635 io_req_task_queue_fail(req
, err
);
1638 if (!io_assign_file(req
, issue_flags
)) {
1640 work
->flags
|= IO_WQ_WORK_CANCEL
;
1644 if (req
->flags
& REQ_F_FORCE_ASYNC
) {
1645 bool opcode_poll
= def
->pollin
|| def
->pollout
;
1647 if (opcode_poll
&& file_can_poll(req
->file
)) {
1649 issue_flags
|= IO_URING_F_NONBLOCK
;
1654 ret
= io_issue_sqe(req
, issue_flags
);
1658 * We can get EAGAIN for iopolled IO even though we're
1659 * forcing a sync submission from here, since we can't
1660 * wait for request slots on the block side.
1663 if (!(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
1669 if (io_arm_poll_handler(req
, issue_flags
) == IO_APOLL_OK
)
1671 /* aborted or ready, in either case retry blocking */
1673 issue_flags
&= ~IO_URING_F_NONBLOCK
;
1676 /* avoid locking problems by failing it from a clean context */
1678 io_req_task_queue_fail(req
, ret
);
1681 inline struct file
*io_file_get_fixed(struct io_kiocb
*req
, int fd
,
1682 unsigned int issue_flags
)
1684 struct io_ring_ctx
*ctx
= req
->ctx
;
1685 struct file
*file
= NULL
;
1686 unsigned long file_ptr
;
1688 io_ring_submit_lock(ctx
, issue_flags
);
1690 if (unlikely((unsigned int)fd
>= ctx
->nr_user_files
))
1692 fd
= array_index_nospec(fd
, ctx
->nr_user_files
);
1693 file_ptr
= io_fixed_file_slot(&ctx
->file_table
, fd
)->file_ptr
;
1694 file
= (struct file
*) (file_ptr
& FFS_MASK
);
1695 file_ptr
&= ~FFS_MASK
;
1696 /* mask in overlapping REQ_F and FFS bits */
1697 req
->flags
|= (file_ptr
<< REQ_F_SUPPORT_NOWAIT_BIT
);
1698 io_req_set_rsrc_node(req
, ctx
, 0);
1699 WARN_ON_ONCE(file
&& !test_bit(fd
, ctx
->file_table
.bitmap
));
1701 io_ring_submit_unlock(ctx
, issue_flags
);
1705 struct file
*io_file_get_normal(struct io_kiocb
*req
, int fd
)
1707 struct file
*file
= fget(fd
);
1709 trace_io_uring_file_get(req
, fd
);
1711 /* we don't allow fixed io_uring files */
1712 if (file
&& io_is_uring_fops(file
))
1713 io_req_track_inflight(req
);
1717 static void io_queue_async(struct io_kiocb
*req
, int ret
)
1718 __must_hold(&req
->ctx
->uring_lock
)
1720 struct io_kiocb
*linked_timeout
;
1722 if (ret
!= -EAGAIN
|| (req
->flags
& REQ_F_NOWAIT
)) {
1723 io_req_complete_failed(req
, ret
);
1727 linked_timeout
= io_prep_linked_timeout(req
);
1729 switch (io_arm_poll_handler(req
, 0)) {
1730 case IO_APOLL_READY
:
1731 io_req_task_queue(req
);
1733 case IO_APOLL_ABORTED
:
1735 * Queued up for async execution, worker will release
1736 * submit reference when the iocb is actually submitted.
1738 io_kbuf_recycle(req
, 0);
1739 io_queue_iowq(req
, NULL
);
1746 io_queue_linked_timeout(linked_timeout
);
1749 static inline void io_queue_sqe(struct io_kiocb
*req
)
1750 __must_hold(&req
->ctx
->uring_lock
)
1754 ret
= io_issue_sqe(req
, IO_URING_F_NONBLOCK
|IO_URING_F_COMPLETE_DEFER
);
1757 * We async punt it if the file wasn't marked NOWAIT, or if the file
1758 * doesn't support non-blocking read/write attempts
1761 io_arm_ltimeout(req
);
1763 io_queue_async(req
, ret
);
1766 static void io_queue_sqe_fallback(struct io_kiocb
*req
)
1767 __must_hold(&req
->ctx
->uring_lock
)
1769 if (unlikely(req
->flags
& REQ_F_FAIL
)) {
1771 * We don't submit, fail them all, for that replace hardlinks
1772 * with normal links. Extra REQ_F_LINK is tolerated.
1774 req
->flags
&= ~REQ_F_HARDLINK
;
1775 req
->flags
|= REQ_F_LINK
;
1776 io_req_complete_failed(req
, req
->cqe
.res
);
1777 } else if (unlikely(req
->ctx
->drain_active
)) {
1780 int ret
= io_req_prep_async(req
);
1783 io_req_complete_failed(req
, ret
);
1785 io_queue_iowq(req
, NULL
);
1790 * Check SQE restrictions (opcode and flags).
1792 * Returns 'true' if SQE is allowed, 'false' otherwise.
1794 static inline bool io_check_restriction(struct io_ring_ctx
*ctx
,
1795 struct io_kiocb
*req
,
1796 unsigned int sqe_flags
)
1798 if (!test_bit(req
->opcode
, ctx
->restrictions
.sqe_op
))
1801 if ((sqe_flags
& ctx
->restrictions
.sqe_flags_required
) !=
1802 ctx
->restrictions
.sqe_flags_required
)
1805 if (sqe_flags
& ~(ctx
->restrictions
.sqe_flags_allowed
|
1806 ctx
->restrictions
.sqe_flags_required
))
1812 static void io_init_req_drain(struct io_kiocb
*req
)
1814 struct io_ring_ctx
*ctx
= req
->ctx
;
1815 struct io_kiocb
*head
= ctx
->submit_state
.link
.head
;
1817 ctx
->drain_active
= true;
1820 * If we need to drain a request in the middle of a link, drain
1821 * the head request and the next request/link after the current
1822 * link. Considering sequential execution of links,
1823 * REQ_F_IO_DRAIN will be maintained for every request of our
1826 head
->flags
|= REQ_F_IO_DRAIN
| REQ_F_FORCE_ASYNC
;
1827 ctx
->drain_next
= true;
1831 static int io_init_req(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
1832 const struct io_uring_sqe
*sqe
)
1833 __must_hold(&ctx
->uring_lock
)
1835 const struct io_op_def
*def
;
1836 unsigned int sqe_flags
;
1840 /* req is partially pre-initialised, see io_preinit_req() */
1841 req
->opcode
= opcode
= READ_ONCE(sqe
->opcode
);
1842 /* same numerical values with corresponding REQ_F_*, safe to copy */
1843 req
->flags
= sqe_flags
= READ_ONCE(sqe
->flags
);
1844 req
->cqe
.user_data
= READ_ONCE(sqe
->user_data
);
1846 req
->rsrc_node
= NULL
;
1847 req
->task
= current
;
1849 if (unlikely(opcode
>= IORING_OP_LAST
)) {
1853 def
= &io_op_defs
[opcode
];
1854 if (unlikely(sqe_flags
& ~SQE_COMMON_FLAGS
)) {
1855 /* enforce forwards compatibility on users */
1856 if (sqe_flags
& ~SQE_VALID_FLAGS
)
1858 if (sqe_flags
& IOSQE_BUFFER_SELECT
) {
1859 if (!def
->buffer_select
)
1861 req
->buf_index
= READ_ONCE(sqe
->buf_group
);
1863 if (sqe_flags
& IOSQE_CQE_SKIP_SUCCESS
)
1864 ctx
->drain_disabled
= true;
1865 if (sqe_flags
& IOSQE_IO_DRAIN
) {
1866 if (ctx
->drain_disabled
)
1868 io_init_req_drain(req
);
1871 if (unlikely(ctx
->restricted
|| ctx
->drain_active
|| ctx
->drain_next
)) {
1872 if (ctx
->restricted
&& !io_check_restriction(ctx
, req
, sqe_flags
))
1874 /* knock it to the slow queue path, will be drained there */
1875 if (ctx
->drain_active
)
1876 req
->flags
|= REQ_F_FORCE_ASYNC
;
1877 /* if there is no link, we're at "next" request and need to drain */
1878 if (unlikely(ctx
->drain_next
) && !ctx
->submit_state
.link
.head
) {
1879 ctx
->drain_next
= false;
1880 ctx
->drain_active
= true;
1881 req
->flags
|= REQ_F_IO_DRAIN
| REQ_F_FORCE_ASYNC
;
1885 if (!def
->ioprio
&& sqe
->ioprio
)
1887 if (!def
->iopoll
&& (ctx
->flags
& IORING_SETUP_IOPOLL
))
1890 if (def
->needs_file
) {
1891 struct io_submit_state
*state
= &ctx
->submit_state
;
1893 req
->cqe
.fd
= READ_ONCE(sqe
->fd
);
1896 * Plug now if we have more than 2 IO left after this, and the
1897 * target is potentially a read/write to block based storage.
1899 if (state
->need_plug
&& def
->plug
) {
1900 state
->plug_started
= true;
1901 state
->need_plug
= false;
1902 blk_start_plug_nr_ios(&state
->plug
, state
->submit_nr
);
1906 personality
= READ_ONCE(sqe
->personality
);
1910 req
->creds
= xa_load(&ctx
->personalities
, personality
);
1913 get_cred(req
->creds
);
1914 ret
= security_uring_override_creds(req
->creds
);
1916 put_cred(req
->creds
);
1919 req
->flags
|= REQ_F_CREDS
;
1922 return def
->prep(req
, sqe
);
1925 static __cold
int io_submit_fail_init(const struct io_uring_sqe
*sqe
,
1926 struct io_kiocb
*req
, int ret
)
1928 struct io_ring_ctx
*ctx
= req
->ctx
;
1929 struct io_submit_link
*link
= &ctx
->submit_state
.link
;
1930 struct io_kiocb
*head
= link
->head
;
1932 trace_io_uring_req_failed(sqe
, req
, ret
);
1935 * Avoid breaking links in the middle as it renders links with SQPOLL
1936 * unusable. Instead of failing eagerly, continue assembling the link if
1937 * applicable and mark the head with REQ_F_FAIL. The link flushing code
1938 * should find the flag and handle the rest.
1940 req_fail_link_node(req
, ret
);
1941 if (head
&& !(head
->flags
& REQ_F_FAIL
))
1942 req_fail_link_node(head
, -ECANCELED
);
1944 if (!(req
->flags
& IO_REQ_LINK_FLAGS
)) {
1946 link
->last
->link
= req
;
1950 io_queue_sqe_fallback(req
);
1955 link
->last
->link
= req
;
1962 static inline int io_submit_sqe(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
1963 const struct io_uring_sqe
*sqe
)
1964 __must_hold(&ctx
->uring_lock
)
1966 struct io_submit_link
*link
= &ctx
->submit_state
.link
;
1969 ret
= io_init_req(ctx
, req
, sqe
);
1971 return io_submit_fail_init(sqe
, req
, ret
);
1973 /* don't need @sqe from now on */
1974 trace_io_uring_submit_sqe(req
, true);
1977 * If we already have a head request, queue this one for async
1978 * submittal once the head completes. If we don't have a head but
1979 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
1980 * submitted sync once the chain is complete. If none of those
1981 * conditions are true (normal request), then just queue it.
1983 if (unlikely(link
->head
)) {
1984 ret
= io_req_prep_async(req
);
1986 return io_submit_fail_init(sqe
, req
, ret
);
1988 trace_io_uring_link(req
, link
->head
);
1989 link
->last
->link
= req
;
1992 if (req
->flags
& IO_REQ_LINK_FLAGS
)
1994 /* last request of the link, flush it */
1997 if (req
->flags
& (REQ_F_FORCE_ASYNC
| REQ_F_FAIL
))
2000 } else if (unlikely(req
->flags
& (IO_REQ_LINK_FLAGS
|
2001 REQ_F_FORCE_ASYNC
| REQ_F_FAIL
))) {
2002 if (req
->flags
& IO_REQ_LINK_FLAGS
) {
2007 io_queue_sqe_fallback(req
);
2017 * Batched submission is done, ensure local IO is flushed out.
2019 static void io_submit_state_end(struct io_ring_ctx
*ctx
)
2021 struct io_submit_state
*state
= &ctx
->submit_state
;
2023 if (unlikely(state
->link
.head
))
2024 io_queue_sqe_fallback(state
->link
.head
);
2025 /* flush only after queuing links as they can generate completions */
2026 io_submit_flush_completions(ctx
);
2027 if (state
->plug_started
)
2028 blk_finish_plug(&state
->plug
);
2032 * Start submission side cache.
2034 static void io_submit_state_start(struct io_submit_state
*state
,
2035 unsigned int max_ios
)
2037 state
->plug_started
= false;
2038 state
->need_plug
= max_ios
> 2;
2039 state
->submit_nr
= max_ios
;
2040 /* set only head, no need to init link_last in advance */
2041 state
->link
.head
= NULL
;
2044 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
2046 struct io_rings
*rings
= ctx
->rings
;
2049 * Ensure any loads from the SQEs are done at this point,
2050 * since once we write the new head, the application could
2051 * write new data to them.
2053 smp_store_release(&rings
->sq
.head
, ctx
->cached_sq_head
);
2057 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2058 * that is mapped by userspace. This means that care needs to be taken to
2059 * ensure that reads are stable, as we cannot rely on userspace always
2060 * being a good citizen. If members of the sqe are validated and then later
2061 * used, it's important that those reads are done through READ_ONCE() to
2062 * prevent a re-load down the line.
2064 static const struct io_uring_sqe
*io_get_sqe(struct io_ring_ctx
*ctx
)
2066 unsigned head
, mask
= ctx
->sq_entries
- 1;
2067 unsigned sq_idx
= ctx
->cached_sq_head
++ & mask
;
2070 * The cached sq head (or cq tail) serves two purposes:
2072 * 1) allows us to batch the cost of updating the user visible
2074 * 2) allows the kernel side to track the head on its own, even
2075 * though the application is the one updating it.
2077 head
= READ_ONCE(ctx
->sq_array
[sq_idx
]);
2078 if (likely(head
< ctx
->sq_entries
)) {
2079 /* double index for 128-byte SQEs, twice as long */
2080 if (ctx
->flags
& IORING_SETUP_SQE128
)
2082 return &ctx
->sq_sqes
[head
];
2085 /* drop invalid entries */
2087 WRITE_ONCE(ctx
->rings
->sq_dropped
,
2088 READ_ONCE(ctx
->rings
->sq_dropped
) + 1);
2092 int io_submit_sqes(struct io_ring_ctx
*ctx
, unsigned int nr
)
2093 __must_hold(&ctx
->uring_lock
)
2095 unsigned int entries
= io_sqring_entries(ctx
);
2099 if (unlikely(!entries
))
2101 /* make sure SQ entry isn't read before tail */
2102 ret
= left
= min3(nr
, ctx
->sq_entries
, entries
);
2103 io_get_task_refs(left
);
2104 io_submit_state_start(&ctx
->submit_state
, left
);
2107 const struct io_uring_sqe
*sqe
;
2108 struct io_kiocb
*req
;
2110 if (unlikely(!io_alloc_req_refill(ctx
)))
2112 req
= io_alloc_req(ctx
);
2113 sqe
= io_get_sqe(ctx
);
2114 if (unlikely(!sqe
)) {
2115 io_req_add_to_cache(req
, ctx
);
2120 * Continue submitting even for sqe failure if the
2121 * ring was setup with IORING_SETUP_SUBMIT_ALL
2123 if (unlikely(io_submit_sqe(ctx
, req
, sqe
)) &&
2124 !(ctx
->flags
& IORING_SETUP_SUBMIT_ALL
)) {
2130 if (unlikely(left
)) {
2132 /* try again if it submitted nothing and can't allocate a req */
2133 if (!ret
&& io_req_cache_empty(ctx
))
2135 current
->io_uring
->cached_refs
+= left
;
2138 io_submit_state_end(ctx
);
2139 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2140 io_commit_sqring(ctx
);
2144 struct io_wait_queue
{
2145 struct wait_queue_entry wq
;
2146 struct io_ring_ctx
*ctx
;
2148 unsigned nr_timeouts
;
2151 static inline bool io_should_wake(struct io_wait_queue
*iowq
)
2153 struct io_ring_ctx
*ctx
= iowq
->ctx
;
2154 int dist
= ctx
->cached_cq_tail
- (int) iowq
->cq_tail
;
2157 * Wake up if we have enough events, or if a timeout occurred since we
2158 * started waiting. For timeouts, we always want to return to userspace,
2159 * regardless of event count.
2161 return dist
>= 0 || atomic_read(&ctx
->cq_timeouts
) != iowq
->nr_timeouts
;
2164 static int io_wake_function(struct wait_queue_entry
*curr
, unsigned int mode
,
2165 int wake_flags
, void *key
)
2167 struct io_wait_queue
*iowq
= container_of(curr
, struct io_wait_queue
,
2171 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2172 * the task, and the next invocation will do it.
2174 if (io_should_wake(iowq
) ||
2175 test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &iowq
->ctx
->check_cq
))
2176 return autoremove_wake_function(curr
, mode
, wake_flags
, key
);
2180 int io_run_task_work_sig(void)
2182 if (io_run_task_work())
2184 if (task_sigpending(current
))
2189 /* when returns >0, the caller should retry */
2190 static inline int io_cqring_wait_schedule(struct io_ring_ctx
*ctx
,
2191 struct io_wait_queue
*iowq
,
2195 unsigned long check_cq
;
2197 /* make sure we run task_work before checking for signals */
2198 ret
= io_run_task_work_sig();
2199 if (ret
|| io_should_wake(iowq
))
2202 check_cq
= READ_ONCE(ctx
->check_cq
);
2203 if (unlikely(check_cq
)) {
2204 /* let the caller flush overflows, retry */
2205 if (check_cq
& BIT(IO_CHECK_CQ_OVERFLOW_BIT
))
2207 if (check_cq
& BIT(IO_CHECK_CQ_DROPPED_BIT
))
2210 if (!schedule_hrtimeout(&timeout
, HRTIMER_MODE_ABS
))
2216 * Wait until events become available, if we don't already have some. The
2217 * application must reap them itself, as they reside on the shared cq ring.
2219 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
2220 const sigset_t __user
*sig
, size_t sigsz
,
2221 struct __kernel_timespec __user
*uts
)
2223 struct io_wait_queue iowq
;
2224 struct io_rings
*rings
= ctx
->rings
;
2225 ktime_t timeout
= KTIME_MAX
;
2229 io_cqring_overflow_flush(ctx
);
2230 if (io_cqring_events(ctx
) >= min_events
)
2232 if (!io_run_task_work())
2237 #ifdef CONFIG_COMPAT
2238 if (in_compat_syscall())
2239 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
2243 ret
= set_user_sigmask(sig
, sigsz
);
2250 struct timespec64 ts
;
2252 if (get_timespec64(&ts
, uts
))
2254 timeout
= ktime_add_ns(timespec64_to_ktime(ts
), ktime_get_ns());
2257 init_waitqueue_func_entry(&iowq
.wq
, io_wake_function
);
2258 iowq
.wq
.private = current
;
2259 INIT_LIST_HEAD(&iowq
.wq
.entry
);
2261 iowq
.nr_timeouts
= atomic_read(&ctx
->cq_timeouts
);
2262 iowq
.cq_tail
= READ_ONCE(ctx
->rings
->cq
.head
) + min_events
;
2264 trace_io_uring_cqring_wait(ctx
, min_events
);
2266 /* if we can't even flush overflow, don't wait for more */
2267 if (!io_cqring_overflow_flush(ctx
)) {
2271 prepare_to_wait_exclusive(&ctx
->cq_wait
, &iowq
.wq
,
2272 TASK_INTERRUPTIBLE
);
2273 ret
= io_cqring_wait_schedule(ctx
, &iowq
, timeout
);
2277 finish_wait(&ctx
->cq_wait
, &iowq
.wq
);
2278 restore_saved_sigmask_unless(ret
== -EINTR
);
2280 return READ_ONCE(rings
->cq
.head
) == READ_ONCE(rings
->cq
.tail
) ? ret
: 0;
2283 static void io_mem_free(void *ptr
)
2290 page
= virt_to_head_page(ptr
);
2291 if (put_page_testzero(page
))
2292 free_compound_page(page
);
2295 static void *io_mem_alloc(size_t size
)
2297 gfp_t gfp
= GFP_KERNEL_ACCOUNT
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
;
2299 return (void *) __get_free_pages(gfp
, get_order(size
));
2302 static unsigned long rings_size(struct io_ring_ctx
*ctx
, unsigned int sq_entries
,
2303 unsigned int cq_entries
, size_t *sq_offset
)
2305 struct io_rings
*rings
;
2306 size_t off
, sq_array_size
;
2308 off
= struct_size(rings
, cqes
, cq_entries
);
2309 if (off
== SIZE_MAX
)
2311 if (ctx
->flags
& IORING_SETUP_CQE32
) {
2312 if (check_shl_overflow(off
, 1, &off
))
2317 off
= ALIGN(off
, SMP_CACHE_BYTES
);
2325 sq_array_size
= array_size(sizeof(u32
), sq_entries
);
2326 if (sq_array_size
== SIZE_MAX
)
2329 if (check_add_overflow(off
, sq_array_size
, &off
))
2335 static int io_eventfd_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
2336 unsigned int eventfd_async
)
2338 struct io_ev_fd
*ev_fd
;
2339 __s32 __user
*fds
= arg
;
2342 ev_fd
= rcu_dereference_protected(ctx
->io_ev_fd
,
2343 lockdep_is_held(&ctx
->uring_lock
));
2347 if (copy_from_user(&fd
, fds
, sizeof(*fds
)))
2350 ev_fd
= kmalloc(sizeof(*ev_fd
), GFP_KERNEL
);
2354 ev_fd
->cq_ev_fd
= eventfd_ctx_fdget(fd
);
2355 if (IS_ERR(ev_fd
->cq_ev_fd
)) {
2356 int ret
= PTR_ERR(ev_fd
->cq_ev_fd
);
2361 spin_lock(&ctx
->completion_lock
);
2362 ctx
->evfd_last_cq_tail
= ctx
->cached_cq_tail
;
2363 spin_unlock(&ctx
->completion_lock
);
2365 ev_fd
->eventfd_async
= eventfd_async
;
2366 ctx
->has_evfd
= true;
2367 rcu_assign_pointer(ctx
->io_ev_fd
, ev_fd
);
2371 static void io_eventfd_put(struct rcu_head
*rcu
)
2373 struct io_ev_fd
*ev_fd
= container_of(rcu
, struct io_ev_fd
, rcu
);
2375 eventfd_ctx_put(ev_fd
->cq_ev_fd
);
2379 static int io_eventfd_unregister(struct io_ring_ctx
*ctx
)
2381 struct io_ev_fd
*ev_fd
;
2383 ev_fd
= rcu_dereference_protected(ctx
->io_ev_fd
,
2384 lockdep_is_held(&ctx
->uring_lock
));
2386 ctx
->has_evfd
= false;
2387 rcu_assign_pointer(ctx
->io_ev_fd
, NULL
);
2388 call_rcu(&ev_fd
->rcu
, io_eventfd_put
);
2395 static void io_req_caches_free(struct io_ring_ctx
*ctx
)
2397 struct io_submit_state
*state
= &ctx
->submit_state
;
2400 mutex_lock(&ctx
->uring_lock
);
2401 io_flush_cached_locked_reqs(ctx
, state
);
2403 while (!io_req_cache_empty(ctx
)) {
2404 struct io_wq_work_node
*node
;
2405 struct io_kiocb
*req
;
2407 node
= wq_stack_extract(&state
->free_list
);
2408 req
= container_of(node
, struct io_kiocb
, comp_list
);
2409 kmem_cache_free(req_cachep
, req
);
2413 percpu_ref_put_many(&ctx
->refs
, nr
);
2414 mutex_unlock(&ctx
->uring_lock
);
2417 static __cold
void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
2419 io_sq_thread_finish(ctx
);
2421 if (ctx
->mm_account
) {
2422 mmdrop(ctx
->mm_account
);
2423 ctx
->mm_account
= NULL
;
2426 io_rsrc_refs_drop(ctx
);
2427 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2428 io_wait_rsrc_data(ctx
->buf_data
);
2429 io_wait_rsrc_data(ctx
->file_data
);
2431 mutex_lock(&ctx
->uring_lock
);
2433 __io_sqe_buffers_unregister(ctx
);
2435 __io_sqe_files_unregister(ctx
);
2437 __io_cqring_overflow_flush(ctx
, true);
2438 io_eventfd_unregister(ctx
);
2439 io_alloc_cache_free(&ctx
->apoll_cache
, io_apoll_cache_free
);
2440 io_alloc_cache_free(&ctx
->netmsg_cache
, io_netmsg_cache_free
);
2441 mutex_unlock(&ctx
->uring_lock
);
2442 io_destroy_buffers(ctx
);
2444 put_cred(ctx
->sq_creds
);
2445 if (ctx
->submitter_task
)
2446 put_task_struct(ctx
->submitter_task
);
2448 /* there are no registered resources left, nobody uses it */
2450 io_rsrc_node_destroy(ctx
->rsrc_node
);
2451 if (ctx
->rsrc_backup_node
)
2452 io_rsrc_node_destroy(ctx
->rsrc_backup_node
);
2453 flush_delayed_work(&ctx
->rsrc_put_work
);
2454 flush_delayed_work(&ctx
->fallback_work
);
2456 WARN_ON_ONCE(!list_empty(&ctx
->rsrc_ref_list
));
2457 WARN_ON_ONCE(!llist_empty(&ctx
->rsrc_put_llist
));
2459 #if defined(CONFIG_UNIX)
2460 if (ctx
->ring_sock
) {
2461 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
2462 sock_release(ctx
->ring_sock
);
2465 WARN_ON_ONCE(!list_empty(&ctx
->ltimeout_list
));
2466 WARN_ON_ONCE(ctx
->notif_slots
|| ctx
->nr_notif_slots
);
2468 io_mem_free(ctx
->rings
);
2469 io_mem_free(ctx
->sq_sqes
);
2471 percpu_ref_exit(&ctx
->refs
);
2472 free_uid(ctx
->user
);
2473 io_req_caches_free(ctx
);
2475 io_wq_put_hash(ctx
->hash_map
);
2476 kfree(ctx
->cancel_table
.hbs
);
2477 kfree(ctx
->cancel_table_locked
.hbs
);
2478 kfree(ctx
->dummy_ubuf
);
2480 xa_destroy(&ctx
->io_bl_xa
);
2484 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
2486 struct io_ring_ctx
*ctx
= file
->private_data
;
2489 poll_wait(file
, &ctx
->cq_wait
, wait
);
2491 * synchronizes with barrier from wq_has_sleeper call in
2495 if (!io_sqring_full(ctx
))
2496 mask
|= EPOLLOUT
| EPOLLWRNORM
;
2499 * Don't flush cqring overflow list here, just do a simple check.
2500 * Otherwise there could possible be ABBA deadlock:
2503 * lock(&ctx->uring_lock);
2505 * lock(&ctx->uring_lock);
2508 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2509 * pushs them to do the flush.
2511 if (io_cqring_events(ctx
) ||
2512 test_bit(IO_CHECK_CQ_OVERFLOW_BIT
, &ctx
->check_cq
))
2513 mask
|= EPOLLIN
| EPOLLRDNORM
;
2518 static int io_unregister_personality(struct io_ring_ctx
*ctx
, unsigned id
)
2520 const struct cred
*creds
;
2522 creds
= xa_erase(&ctx
->personalities
, id
);
2531 struct io_tctx_exit
{
2532 struct callback_head task_work
;
2533 struct completion completion
;
2534 struct io_ring_ctx
*ctx
;
2537 static __cold
void io_tctx_exit_cb(struct callback_head
*cb
)
2539 struct io_uring_task
*tctx
= current
->io_uring
;
2540 struct io_tctx_exit
*work
;
2542 work
= container_of(cb
, struct io_tctx_exit
, task_work
);
2544 * When @in_idle, we're in cancellation and it's racy to remove the
2545 * node. It'll be removed by the end of cancellation, just ignore it.
2547 if (!atomic_read(&tctx
->in_idle
))
2548 io_uring_del_tctx_node((unsigned long)work
->ctx
);
2549 complete(&work
->completion
);
2552 static __cold
bool io_cancel_ctx_cb(struct io_wq_work
*work
, void *data
)
2554 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
2556 return req
->ctx
== data
;
2559 static __cold
void io_ring_exit_work(struct work_struct
*work
)
2561 struct io_ring_ctx
*ctx
= container_of(work
, struct io_ring_ctx
, exit_work
);
2562 unsigned long timeout
= jiffies
+ HZ
* 60 * 5;
2563 unsigned long interval
= HZ
/ 20;
2564 struct io_tctx_exit exit
;
2565 struct io_tctx_node
*node
;
2569 * If we're doing polled IO and end up having requests being
2570 * submitted async (out-of-line), then completions can come in while
2571 * we're waiting for refs to drop. We need to reap these manually,
2572 * as nobody else will be looking for them.
2575 while (io_uring_try_cancel_requests(ctx
, NULL
, true))
2579 struct io_sq_data
*sqd
= ctx
->sq_data
;
2580 struct task_struct
*tsk
;
2582 io_sq_thread_park(sqd
);
2584 if (tsk
&& tsk
->io_uring
&& tsk
->io_uring
->io_wq
)
2585 io_wq_cancel_cb(tsk
->io_uring
->io_wq
,
2586 io_cancel_ctx_cb
, ctx
, true);
2587 io_sq_thread_unpark(sqd
);
2590 io_req_caches_free(ctx
);
2592 if (WARN_ON_ONCE(time_after(jiffies
, timeout
))) {
2593 /* there is little hope left, don't run it too often */
2596 } while (!wait_for_completion_timeout(&ctx
->ref_comp
, interval
));
2598 init_completion(&exit
.completion
);
2599 init_task_work(&exit
.task_work
, io_tctx_exit_cb
);
2602 * Some may use context even when all refs and requests have been put,
2603 * and they are free to do so while still holding uring_lock or
2604 * completion_lock, see io_req_task_submit(). Apart from other work,
2605 * this lock/unlock section also waits them to finish.
2607 mutex_lock(&ctx
->uring_lock
);
2608 while (!list_empty(&ctx
->tctx_list
)) {
2609 WARN_ON_ONCE(time_after(jiffies
, timeout
));
2611 node
= list_first_entry(&ctx
->tctx_list
, struct io_tctx_node
,
2613 /* don't spin on a single task if cancellation failed */
2614 list_rotate_left(&ctx
->tctx_list
);
2615 ret
= task_work_add(node
->task
, &exit
.task_work
, TWA_SIGNAL
);
2616 if (WARN_ON_ONCE(ret
))
2619 mutex_unlock(&ctx
->uring_lock
);
2620 wait_for_completion(&exit
.completion
);
2621 mutex_lock(&ctx
->uring_lock
);
2623 mutex_unlock(&ctx
->uring_lock
);
2624 spin_lock(&ctx
->completion_lock
);
2625 spin_unlock(&ctx
->completion_lock
);
2627 io_ring_ctx_free(ctx
);
2630 static __cold
void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
2632 unsigned long index
;
2633 struct creds
*creds
;
2635 mutex_lock(&ctx
->uring_lock
);
2636 percpu_ref_kill(&ctx
->refs
);
2638 __io_cqring_overflow_flush(ctx
, true);
2639 xa_for_each(&ctx
->personalities
, index
, creds
)
2640 io_unregister_personality(ctx
, index
);
2642 io_poll_remove_all(ctx
, NULL
, true);
2643 mutex_unlock(&ctx
->uring_lock
);
2645 /* failed during ring init, it couldn't have issued any requests */
2647 io_kill_timeouts(ctx
, NULL
, true);
2648 /* if we failed setting up the ctx, we might not have any rings */
2649 io_iopoll_try_reap_events(ctx
);
2652 INIT_WORK(&ctx
->exit_work
, io_ring_exit_work
);
2654 * Use system_unbound_wq to avoid spawning tons of event kworkers
2655 * if we're exiting a ton of rings at the same time. It just adds
2656 * noise and overhead, there's no discernable change in runtime
2657 * over using system_wq.
2659 queue_work(system_unbound_wq
, &ctx
->exit_work
);
2662 static int io_uring_release(struct inode
*inode
, struct file
*file
)
2664 struct io_ring_ctx
*ctx
= file
->private_data
;
2666 file
->private_data
= NULL
;
2667 io_ring_ctx_wait_and_kill(ctx
);
2671 struct io_task_cancel
{
2672 struct task_struct
*task
;
2676 static bool io_cancel_task_cb(struct io_wq_work
*work
, void *data
)
2678 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
2679 struct io_task_cancel
*cancel
= data
;
2681 return io_match_task_safe(req
, cancel
->task
, cancel
->all
);
2684 static __cold
bool io_cancel_defer_files(struct io_ring_ctx
*ctx
,
2685 struct task_struct
*task
,
2688 struct io_defer_entry
*de
;
2691 spin_lock(&ctx
->completion_lock
);
2692 list_for_each_entry_reverse(de
, &ctx
->defer_list
, list
) {
2693 if (io_match_task_safe(de
->req
, task
, cancel_all
)) {
2694 list_cut_position(&list
, &ctx
->defer_list
, &de
->list
);
2698 spin_unlock(&ctx
->completion_lock
);
2699 if (list_empty(&list
))
2702 while (!list_empty(&list
)) {
2703 de
= list_first_entry(&list
, struct io_defer_entry
, list
);
2704 list_del_init(&de
->list
);
2705 io_req_complete_failed(de
->req
, -ECANCELED
);
2711 static __cold
bool io_uring_try_cancel_iowq(struct io_ring_ctx
*ctx
)
2713 struct io_tctx_node
*node
;
2714 enum io_wq_cancel cret
;
2717 mutex_lock(&ctx
->uring_lock
);
2718 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
2719 struct io_uring_task
*tctx
= node
->task
->io_uring
;
2722 * io_wq will stay alive while we hold uring_lock, because it's
2723 * killed after ctx nodes, which requires to take the lock.
2725 if (!tctx
|| !tctx
->io_wq
)
2727 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_ctx_cb
, ctx
, true);
2728 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
2730 mutex_unlock(&ctx
->uring_lock
);
2735 static __cold
bool io_uring_try_cancel_requests(struct io_ring_ctx
*ctx
,
2736 struct task_struct
*task
,
2739 struct io_task_cancel cancel
= { .task
= task
, .all
= cancel_all
, };
2740 struct io_uring_task
*tctx
= task
? task
->io_uring
: NULL
;
2741 enum io_wq_cancel cret
;
2744 /* failed during ring init, it couldn't have issued any requests */
2749 ret
|= io_uring_try_cancel_iowq(ctx
);
2750 } else if (tctx
&& tctx
->io_wq
) {
2752 * Cancels requests of all rings, not only @ctx, but
2753 * it's fine as the task is in exit/exec.
2755 cret
= io_wq_cancel_cb(tctx
->io_wq
, io_cancel_task_cb
,
2757 ret
|= (cret
!= IO_WQ_CANCEL_NOTFOUND
);
2760 /* SQPOLL thread does its own polling */
2761 if ((!(ctx
->flags
& IORING_SETUP_SQPOLL
) && cancel_all
) ||
2762 (ctx
->sq_data
&& ctx
->sq_data
->thread
== current
)) {
2763 while (!wq_list_empty(&ctx
->iopoll_list
)) {
2764 io_iopoll_try_reap_events(ctx
);
2769 ret
|= io_cancel_defer_files(ctx
, task
, cancel_all
);
2770 mutex_lock(&ctx
->uring_lock
);
2771 ret
|= io_poll_remove_all(ctx
, task
, cancel_all
);
2772 mutex_unlock(&ctx
->uring_lock
);
2773 ret
|= io_kill_timeouts(ctx
, task
, cancel_all
);
2775 ret
|= io_run_task_work();
2779 static s64
tctx_inflight(struct io_uring_task
*tctx
, bool tracked
)
2782 return atomic_read(&tctx
->inflight_tracked
);
2783 return percpu_counter_sum(&tctx
->inflight
);
2787 * Find any io_uring ctx that this task has registered or done IO on, and cancel
2788 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
2790 __cold
void io_uring_cancel_generic(bool cancel_all
, struct io_sq_data
*sqd
)
2792 struct io_uring_task
*tctx
= current
->io_uring
;
2793 struct io_ring_ctx
*ctx
;
2797 WARN_ON_ONCE(sqd
&& sqd
->thread
!= current
);
2799 if (!current
->io_uring
)
2802 io_wq_exit_start(tctx
->io_wq
);
2804 atomic_inc(&tctx
->in_idle
);
2808 io_uring_drop_tctx_refs(current
);
2809 /* read completions before cancelations */
2810 inflight
= tctx_inflight(tctx
, !cancel_all
);
2815 struct io_tctx_node
*node
;
2816 unsigned long index
;
2818 xa_for_each(&tctx
->xa
, index
, node
) {
2819 /* sqpoll task will cancel all its requests */
2820 if (node
->ctx
->sq_data
)
2822 loop
|= io_uring_try_cancel_requests(node
->ctx
,
2823 current
, cancel_all
);
2826 list_for_each_entry(ctx
, &sqd
->ctx_list
, sqd_list
)
2827 loop
|= io_uring_try_cancel_requests(ctx
,
2837 prepare_to_wait(&tctx
->wait
, &wait
, TASK_INTERRUPTIBLE
);
2839 io_uring_drop_tctx_refs(current
);
2842 * If we've seen completions, retry without waiting. This
2843 * avoids a race where a completion comes in before we did
2844 * prepare_to_wait().
2846 if (inflight
== tctx_inflight(tctx
, !cancel_all
))
2848 finish_wait(&tctx
->wait
, &wait
);
2851 io_uring_clean_tctx(tctx
);
2854 * We shouldn't run task_works after cancel, so just leave
2855 * ->in_idle set for normal exit.
2857 atomic_dec(&tctx
->in_idle
);
2858 /* for exec all current's requests should be gone, kill tctx */
2859 __io_uring_free(current
);
2863 void __io_uring_cancel(bool cancel_all
)
2865 io_uring_cancel_generic(cancel_all
, NULL
);
2868 static void *io_uring_validate_mmap_request(struct file
*file
,
2869 loff_t pgoff
, size_t sz
)
2871 struct io_ring_ctx
*ctx
= file
->private_data
;
2872 loff_t offset
= pgoff
<< PAGE_SHIFT
;
2877 case IORING_OFF_SQ_RING
:
2878 case IORING_OFF_CQ_RING
:
2881 case IORING_OFF_SQES
:
2885 return ERR_PTR(-EINVAL
);
2888 page
= virt_to_head_page(ptr
);
2889 if (sz
> page_size(page
))
2890 return ERR_PTR(-EINVAL
);
2897 static __cold
int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
2899 size_t sz
= vma
->vm_end
- vma
->vm_start
;
2903 ptr
= io_uring_validate_mmap_request(file
, vma
->vm_pgoff
, sz
);
2905 return PTR_ERR(ptr
);
2907 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
2908 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
2911 #else /* !CONFIG_MMU */
2913 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
2915 return vma
->vm_flags
& (VM_SHARED
| VM_MAYSHARE
) ? 0 : -EINVAL
;
2918 static unsigned int io_uring_nommu_mmap_capabilities(struct file
*file
)
2920 return NOMMU_MAP_DIRECT
| NOMMU_MAP_READ
| NOMMU_MAP_WRITE
;
2923 static unsigned long io_uring_nommu_get_unmapped_area(struct file
*file
,
2924 unsigned long addr
, unsigned long len
,
2925 unsigned long pgoff
, unsigned long flags
)
2929 ptr
= io_uring_validate_mmap_request(file
, pgoff
, len
);
2931 return PTR_ERR(ptr
);
2933 return (unsigned long) ptr
;
2936 #endif /* !CONFIG_MMU */
2938 static int io_validate_ext_arg(unsigned flags
, const void __user
*argp
, size_t argsz
)
2940 if (flags
& IORING_ENTER_EXT_ARG
) {
2941 struct io_uring_getevents_arg arg
;
2943 if (argsz
!= sizeof(arg
))
2945 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
2951 static int io_get_ext_arg(unsigned flags
, const void __user
*argp
, size_t *argsz
,
2952 struct __kernel_timespec __user
**ts
,
2953 const sigset_t __user
**sig
)
2955 struct io_uring_getevents_arg arg
;
2958 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
2959 * is just a pointer to the sigset_t.
2961 if (!(flags
& IORING_ENTER_EXT_ARG
)) {
2962 *sig
= (const sigset_t __user
*) argp
;
2968 * EXT_ARG is set - ensure we agree on the size of it and copy in our
2969 * timespec and sigset_t pointers if good.
2971 if (*argsz
!= sizeof(arg
))
2973 if (copy_from_user(&arg
, argp
, sizeof(arg
)))
2977 *sig
= u64_to_user_ptr(arg
.sigmask
);
2978 *argsz
= arg
.sigmask_sz
;
2979 *ts
= u64_to_user_ptr(arg
.ts
);
2983 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
2984 u32
, min_complete
, u32
, flags
, const void __user
*, argp
,
2987 struct io_ring_ctx
*ctx
;
2993 if (unlikely(flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
|
2994 IORING_ENTER_SQ_WAIT
| IORING_ENTER_EXT_ARG
|
2995 IORING_ENTER_REGISTERED_RING
)))
2999 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3000 * need only dereference our task private array to find it.
3002 if (flags
& IORING_ENTER_REGISTERED_RING
) {
3003 struct io_uring_task
*tctx
= current
->io_uring
;
3005 if (unlikely(!tctx
|| fd
>= IO_RINGFD_REG_MAX
))
3007 fd
= array_index_nospec(fd
, IO_RINGFD_REG_MAX
);
3008 f
.file
= tctx
->registered_rings
[fd
];
3010 if (unlikely(!f
.file
))
3014 if (unlikely(!f
.file
))
3017 if (unlikely(!io_is_uring_fops(f
.file
)))
3021 ctx
= f
.file
->private_data
;
3023 if (unlikely(ctx
->flags
& IORING_SETUP_R_DISABLED
))
3027 * For SQ polling, the thread will do all submissions and completions.
3028 * Just return the requested submit count, and wake the thread if
3032 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
3033 io_cqring_overflow_flush(ctx
);
3035 if (unlikely(ctx
->sq_data
->thread
== NULL
)) {
3039 if (flags
& IORING_ENTER_SQ_WAKEUP
)
3040 wake_up(&ctx
->sq_data
->wait
);
3041 if (flags
& IORING_ENTER_SQ_WAIT
) {
3042 ret
= io_sqpoll_wait_sq(ctx
);
3047 } else if (to_submit
) {
3048 ret
= io_uring_add_tctx_node(ctx
);
3052 mutex_lock(&ctx
->uring_lock
);
3053 ret
= io_submit_sqes(ctx
, to_submit
);
3054 if (ret
!= to_submit
) {
3055 mutex_unlock(&ctx
->uring_lock
);
3058 if ((flags
& IORING_ENTER_GETEVENTS
) && ctx
->syscall_iopoll
)
3060 mutex_unlock(&ctx
->uring_lock
);
3062 if (flags
& IORING_ENTER_GETEVENTS
) {
3064 if (ctx
->syscall_iopoll
) {
3066 * We disallow the app entering submit/complete with
3067 * polling, but we still need to lock the ring to
3068 * prevent racing with polled issue that got punted to
3071 mutex_lock(&ctx
->uring_lock
);
3073 ret2
= io_validate_ext_arg(flags
, argp
, argsz
);
3074 if (likely(!ret2
)) {
3075 min_complete
= min(min_complete
,
3077 ret2
= io_iopoll_check(ctx
, min_complete
);
3079 mutex_unlock(&ctx
->uring_lock
);
3081 const sigset_t __user
*sig
;
3082 struct __kernel_timespec __user
*ts
;
3084 ret2
= io_get_ext_arg(flags
, argp
, &argsz
, &ts
, &sig
);
3085 if (likely(!ret2
)) {
3086 min_complete
= min(min_complete
,
3088 ret2
= io_cqring_wait(ctx
, min_complete
, sig
,
3097 * EBADR indicates that one or more CQE were dropped.
3098 * Once the user has been informed we can clear the bit
3099 * as they are obviously ok with those drops.
3101 if (unlikely(ret2
== -EBADR
))
3102 clear_bit(IO_CHECK_CQ_DROPPED_BIT
,
3111 static const struct file_operations io_uring_fops
= {
3112 .release
= io_uring_release
,
3113 .mmap
= io_uring_mmap
,
3115 .get_unmapped_area
= io_uring_nommu_get_unmapped_area
,
3116 .mmap_capabilities
= io_uring_nommu_mmap_capabilities
,
3118 .poll
= io_uring_poll
,
3119 #ifdef CONFIG_PROC_FS
3120 .show_fdinfo
= io_uring_show_fdinfo
,
3124 bool io_is_uring_fops(struct file
*file
)
3126 return file
->f_op
== &io_uring_fops
;
3129 static __cold
int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
3130 struct io_uring_params
*p
)
3132 struct io_rings
*rings
;
3133 size_t size
, sq_array_offset
;
3135 /* make sure these are sane, as we already accounted them */
3136 ctx
->sq_entries
= p
->sq_entries
;
3137 ctx
->cq_entries
= p
->cq_entries
;
3139 size
= rings_size(ctx
, p
->sq_entries
, p
->cq_entries
, &sq_array_offset
);
3140 if (size
== SIZE_MAX
)
3143 rings
= io_mem_alloc(size
);
3148 ctx
->sq_array
= (u32
*)((char *)rings
+ sq_array_offset
);
3149 rings
->sq_ring_mask
= p
->sq_entries
- 1;
3150 rings
->cq_ring_mask
= p
->cq_entries
- 1;
3151 rings
->sq_ring_entries
= p
->sq_entries
;
3152 rings
->cq_ring_entries
= p
->cq_entries
;
3154 if (p
->flags
& IORING_SETUP_SQE128
)
3155 size
= array_size(2 * sizeof(struct io_uring_sqe
), p
->sq_entries
);
3157 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
3158 if (size
== SIZE_MAX
) {
3159 io_mem_free(ctx
->rings
);
3164 ctx
->sq_sqes
= io_mem_alloc(size
);
3165 if (!ctx
->sq_sqes
) {
3166 io_mem_free(ctx
->rings
);
3174 static int io_uring_install_fd(struct io_ring_ctx
*ctx
, struct file
*file
)
3178 fd
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
3182 ret
= __io_uring_add_tctx_node(ctx
, false);
3187 fd_install(fd
, file
);
3192 * Allocate an anonymous fd, this is what constitutes the application
3193 * visible backing of an io_uring instance. The application mmaps this
3194 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3195 * we have to tie this fd to a socket for file garbage collection purposes.
3197 static struct file
*io_uring_get_file(struct io_ring_ctx
*ctx
)
3200 #if defined(CONFIG_UNIX)
3203 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
3206 return ERR_PTR(ret
);
3209 file
= anon_inode_getfile_secure("[io_uring]", &io_uring_fops
, ctx
,
3210 O_RDWR
| O_CLOEXEC
, NULL
);
3211 #if defined(CONFIG_UNIX)
3213 sock_release(ctx
->ring_sock
);
3214 ctx
->ring_sock
= NULL
;
3216 ctx
->ring_sock
->file
= file
;
3222 static __cold
int io_uring_create(unsigned entries
, struct io_uring_params
*p
,
3223 struct io_uring_params __user
*params
)
3225 struct io_ring_ctx
*ctx
;
3231 if (entries
> IORING_MAX_ENTRIES
) {
3232 if (!(p
->flags
& IORING_SETUP_CLAMP
))
3234 entries
= IORING_MAX_ENTRIES
;
3238 * Use twice as many entries for the CQ ring. It's possible for the
3239 * application to drive a higher depth than the size of the SQ ring,
3240 * since the sqes are only used at submission time. This allows for
3241 * some flexibility in overcommitting a bit. If the application has
3242 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3243 * of CQ ring entries manually.
3245 p
->sq_entries
= roundup_pow_of_two(entries
);
3246 if (p
->flags
& IORING_SETUP_CQSIZE
) {
3248 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3249 * to a power-of-two, if it isn't already. We do NOT impose
3250 * any cq vs sq ring sizing.
3254 if (p
->cq_entries
> IORING_MAX_CQ_ENTRIES
) {
3255 if (!(p
->flags
& IORING_SETUP_CLAMP
))
3257 p
->cq_entries
= IORING_MAX_CQ_ENTRIES
;
3259 p
->cq_entries
= roundup_pow_of_two(p
->cq_entries
);
3260 if (p
->cq_entries
< p
->sq_entries
)
3263 p
->cq_entries
= 2 * p
->sq_entries
;
3266 ctx
= io_ring_ctx_alloc(p
);
3271 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3272 * space applications don't need to do io completion events
3273 * polling again, they can rely on io_sq_thread to do polling
3274 * work, which can reduce cpu usage and uring_lock contention.
3276 if (ctx
->flags
& IORING_SETUP_IOPOLL
&&
3277 !(ctx
->flags
& IORING_SETUP_SQPOLL
))
3278 ctx
->syscall_iopoll
= 1;
3280 ctx
->compat
= in_compat_syscall();
3281 if (!capable(CAP_IPC_LOCK
))
3282 ctx
->user
= get_uid(current_user());
3285 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3286 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3289 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
3290 /* IPI related flags don't make sense with SQPOLL */
3291 if (ctx
->flags
& (IORING_SETUP_COOP_TASKRUN
|
3292 IORING_SETUP_TASKRUN_FLAG
))
3294 ctx
->notify_method
= TWA_SIGNAL_NO_IPI
;
3295 } else if (ctx
->flags
& IORING_SETUP_COOP_TASKRUN
) {
3296 ctx
->notify_method
= TWA_SIGNAL_NO_IPI
;
3298 if (ctx
->flags
& IORING_SETUP_TASKRUN_FLAG
)
3300 ctx
->notify_method
= TWA_SIGNAL
;
3304 * This is just grabbed for accounting purposes. When a process exits,
3305 * the mm is exited and dropped before the files, hence we need to hang
3306 * on to this mm purely for the purposes of being able to unaccount
3307 * memory (locked/pinned vm). It's not used for anything else.
3309 mmgrab(current
->mm
);
3310 ctx
->mm_account
= current
->mm
;
3312 ret
= io_allocate_scq_urings(ctx
, p
);
3316 ret
= io_sq_offload_create(ctx
, p
);
3319 /* always set a rsrc node */
3320 ret
= io_rsrc_node_switch_start(ctx
);
3323 io_rsrc_node_switch(ctx
, NULL
);
3325 memset(&p
->sq_off
, 0, sizeof(p
->sq_off
));
3326 p
->sq_off
.head
= offsetof(struct io_rings
, sq
.head
);
3327 p
->sq_off
.tail
= offsetof(struct io_rings
, sq
.tail
);
3328 p
->sq_off
.ring_mask
= offsetof(struct io_rings
, sq_ring_mask
);
3329 p
->sq_off
.ring_entries
= offsetof(struct io_rings
, sq_ring_entries
);
3330 p
->sq_off
.flags
= offsetof(struct io_rings
, sq_flags
);
3331 p
->sq_off
.dropped
= offsetof(struct io_rings
, sq_dropped
);
3332 p
->sq_off
.array
= (char *)ctx
->sq_array
- (char *)ctx
->rings
;
3334 memset(&p
->cq_off
, 0, sizeof(p
->cq_off
));
3335 p
->cq_off
.head
= offsetof(struct io_rings
, cq
.head
);
3336 p
->cq_off
.tail
= offsetof(struct io_rings
, cq
.tail
);
3337 p
->cq_off
.ring_mask
= offsetof(struct io_rings
, cq_ring_mask
);
3338 p
->cq_off
.ring_entries
= offsetof(struct io_rings
, cq_ring_entries
);
3339 p
->cq_off
.overflow
= offsetof(struct io_rings
, cq_overflow
);
3340 p
->cq_off
.cqes
= offsetof(struct io_rings
, cqes
);
3341 p
->cq_off
.flags
= offsetof(struct io_rings
, cq_flags
);
3343 p
->features
= IORING_FEAT_SINGLE_MMAP
| IORING_FEAT_NODROP
|
3344 IORING_FEAT_SUBMIT_STABLE
| IORING_FEAT_RW_CUR_POS
|
3345 IORING_FEAT_CUR_PERSONALITY
| IORING_FEAT_FAST_POLL
|
3346 IORING_FEAT_POLL_32BITS
| IORING_FEAT_SQPOLL_NONFIXED
|
3347 IORING_FEAT_EXT_ARG
| IORING_FEAT_NATIVE_WORKERS
|
3348 IORING_FEAT_RSRC_TAGS
| IORING_FEAT_CQE_SKIP
|
3349 IORING_FEAT_LINKED_FILE
;
3351 if (copy_to_user(params
, p
, sizeof(*p
))) {
3356 file
= io_uring_get_file(ctx
);
3358 ret
= PTR_ERR(file
);
3363 * Install ring fd as the very last thing, so we don't risk someone
3364 * having closed it before we finish setup
3366 ret
= io_uring_install_fd(ctx
, file
);
3368 /* fput will clean it up */
3373 trace_io_uring_create(ret
, ctx
, p
->sq_entries
, p
->cq_entries
, p
->flags
);
3376 io_ring_ctx_wait_and_kill(ctx
);
3381 * Sets up an aio uring context, and returns the fd. Applications asks for a
3382 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3383 * params structure passed in.
3385 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
3387 struct io_uring_params p
;
3390 if (copy_from_user(&p
, params
, sizeof(p
)))
3392 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
3397 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
3398 IORING_SETUP_SQ_AFF
| IORING_SETUP_CQSIZE
|
3399 IORING_SETUP_CLAMP
| IORING_SETUP_ATTACH_WQ
|
3400 IORING_SETUP_R_DISABLED
| IORING_SETUP_SUBMIT_ALL
|
3401 IORING_SETUP_COOP_TASKRUN
| IORING_SETUP_TASKRUN_FLAG
|
3402 IORING_SETUP_SQE128
| IORING_SETUP_CQE32
|
3403 IORING_SETUP_SINGLE_ISSUER
))
3406 return io_uring_create(entries
, &p
, params
);
3409 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
3410 struct io_uring_params __user
*, params
)
3412 return io_uring_setup(entries
, params
);
3415 static __cold
int io_probe(struct io_ring_ctx
*ctx
, void __user
*arg
,
3418 struct io_uring_probe
*p
;
3422 size
= struct_size(p
, ops
, nr_args
);
3423 if (size
== SIZE_MAX
)
3425 p
= kzalloc(size
, GFP_KERNEL
);
3430 if (copy_from_user(p
, arg
, size
))
3433 if (memchr_inv(p
, 0, size
))
3436 p
->last_op
= IORING_OP_LAST
- 1;
3437 if (nr_args
> IORING_OP_LAST
)
3438 nr_args
= IORING_OP_LAST
;
3440 for (i
= 0; i
< nr_args
; i
++) {
3442 if (!io_op_defs
[i
].not_supported
)
3443 p
->ops
[i
].flags
= IO_URING_OP_SUPPORTED
;
3448 if (copy_to_user(arg
, p
, size
))
3455 static int io_register_personality(struct io_ring_ctx
*ctx
)
3457 const struct cred
*creds
;
3461 creds
= get_current_cred();
3463 ret
= xa_alloc_cyclic(&ctx
->personalities
, &id
, (void *)creds
,
3464 XA_LIMIT(0, USHRT_MAX
), &ctx
->pers_next
, GFP_KERNEL
);
3472 static __cold
int io_register_restrictions(struct io_ring_ctx
*ctx
,
3473 void __user
*arg
, unsigned int nr_args
)
3475 struct io_uring_restriction
*res
;
3479 /* Restrictions allowed only if rings started disabled */
3480 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
3483 /* We allow only a single restrictions registration */
3484 if (ctx
->restrictions
.registered
)
3487 if (!arg
|| nr_args
> IORING_MAX_RESTRICTIONS
)
3490 size
= array_size(nr_args
, sizeof(*res
));
3491 if (size
== SIZE_MAX
)
3494 res
= memdup_user(arg
, size
);
3496 return PTR_ERR(res
);
3500 for (i
= 0; i
< nr_args
; i
++) {
3501 switch (res
[i
].opcode
) {
3502 case IORING_RESTRICTION_REGISTER_OP
:
3503 if (res
[i
].register_op
>= IORING_REGISTER_LAST
) {
3508 __set_bit(res
[i
].register_op
,
3509 ctx
->restrictions
.register_op
);
3511 case IORING_RESTRICTION_SQE_OP
:
3512 if (res
[i
].sqe_op
>= IORING_OP_LAST
) {
3517 __set_bit(res
[i
].sqe_op
, ctx
->restrictions
.sqe_op
);
3519 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED
:
3520 ctx
->restrictions
.sqe_flags_allowed
= res
[i
].sqe_flags
;
3522 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED
:
3523 ctx
->restrictions
.sqe_flags_required
= res
[i
].sqe_flags
;
3532 /* Reset all restrictions if an error happened */
3534 memset(&ctx
->restrictions
, 0, sizeof(ctx
->restrictions
));
3536 ctx
->restrictions
.registered
= true;
3542 static int io_register_enable_rings(struct io_ring_ctx
*ctx
)
3544 if (!(ctx
->flags
& IORING_SETUP_R_DISABLED
))
3547 if (ctx
->restrictions
.registered
)
3548 ctx
->restricted
= 1;
3550 ctx
->flags
&= ~IORING_SETUP_R_DISABLED
;
3551 if (ctx
->sq_data
&& wq_has_sleeper(&ctx
->sq_data
->wait
))
3552 wake_up(&ctx
->sq_data
->wait
);
3556 static __cold
int io_register_iowq_aff(struct io_ring_ctx
*ctx
,
3557 void __user
*arg
, unsigned len
)
3559 struct io_uring_task
*tctx
= current
->io_uring
;
3560 cpumask_var_t new_mask
;
3563 if (!tctx
|| !tctx
->io_wq
)
3566 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
3569 cpumask_clear(new_mask
);
3570 if (len
> cpumask_size())
3571 len
= cpumask_size();
3573 if (in_compat_syscall()) {
3574 ret
= compat_get_bitmap(cpumask_bits(new_mask
),
3575 (const compat_ulong_t __user
*)arg
,
3576 len
* 8 /* CHAR_BIT */);
3578 ret
= copy_from_user(new_mask
, arg
, len
);
3582 free_cpumask_var(new_mask
);
3586 ret
= io_wq_cpu_affinity(tctx
->io_wq
, new_mask
);
3587 free_cpumask_var(new_mask
);
3591 static __cold
int io_unregister_iowq_aff(struct io_ring_ctx
*ctx
)
3593 struct io_uring_task
*tctx
= current
->io_uring
;
3595 if (!tctx
|| !tctx
->io_wq
)
3598 return io_wq_cpu_affinity(tctx
->io_wq
, NULL
);
3601 static __cold
int io_register_iowq_max_workers(struct io_ring_ctx
*ctx
,
3603 __must_hold(&ctx
->uring_lock
)
3605 struct io_tctx_node
*node
;
3606 struct io_uring_task
*tctx
= NULL
;
3607 struct io_sq_data
*sqd
= NULL
;
3611 if (copy_from_user(new_count
, arg
, sizeof(new_count
)))
3613 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
3614 if (new_count
[i
] > INT_MAX
)
3617 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
3621 * Observe the correct sqd->lock -> ctx->uring_lock
3622 * ordering. Fine to drop uring_lock here, we hold
3625 refcount_inc(&sqd
->refs
);
3626 mutex_unlock(&ctx
->uring_lock
);
3627 mutex_lock(&sqd
->lock
);
3628 mutex_lock(&ctx
->uring_lock
);
3630 tctx
= sqd
->thread
->io_uring
;
3633 tctx
= current
->io_uring
;
3636 BUILD_BUG_ON(sizeof(new_count
) != sizeof(ctx
->iowq_limits
));
3638 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
3640 ctx
->iowq_limits
[i
] = new_count
[i
];
3641 ctx
->iowq_limits_set
= true;
3643 if (tctx
&& tctx
->io_wq
) {
3644 ret
= io_wq_max_workers(tctx
->io_wq
, new_count
);
3648 memset(new_count
, 0, sizeof(new_count
));
3652 mutex_unlock(&sqd
->lock
);
3653 io_put_sq_data(sqd
);
3656 if (copy_to_user(arg
, new_count
, sizeof(new_count
)))
3659 /* that's it for SQPOLL, only the SQPOLL task creates requests */
3663 /* now propagate the restriction to all registered users */
3664 list_for_each_entry(node
, &ctx
->tctx_list
, ctx_node
) {
3665 struct io_uring_task
*tctx
= node
->task
->io_uring
;
3667 if (WARN_ON_ONCE(!tctx
->io_wq
))
3670 for (i
= 0; i
< ARRAY_SIZE(new_count
); i
++)
3671 new_count
[i
] = ctx
->iowq_limits
[i
];
3672 /* ignore errors, it always returns zero anyway */
3673 (void)io_wq_max_workers(tctx
->io_wq
, new_count
);
3678 mutex_unlock(&sqd
->lock
);
3679 io_put_sq_data(sqd
);
3684 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
3685 void __user
*arg
, unsigned nr_args
)
3686 __releases(ctx
->uring_lock
)
3687 __acquires(ctx
->uring_lock
)
3692 * We don't quiesce the refs for register anymore and so it can't be
3693 * dying as we're holding a file ref here.
3695 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx
->refs
)))
3698 if (ctx
->restricted
) {
3699 if (opcode
>= IORING_REGISTER_LAST
)
3701 opcode
= array_index_nospec(opcode
, IORING_REGISTER_LAST
);
3702 if (!test_bit(opcode
, ctx
->restrictions
.register_op
))
3707 case IORING_REGISTER_BUFFERS
:
3711 ret
= io_sqe_buffers_register(ctx
, arg
, nr_args
, NULL
);
3713 case IORING_UNREGISTER_BUFFERS
:
3717 ret
= io_sqe_buffers_unregister(ctx
);
3719 case IORING_REGISTER_FILES
:
3723 ret
= io_sqe_files_register(ctx
, arg
, nr_args
, NULL
);
3725 case IORING_UNREGISTER_FILES
:
3729 ret
= io_sqe_files_unregister(ctx
);
3731 case IORING_REGISTER_FILES_UPDATE
:
3732 ret
= io_register_files_update(ctx
, arg
, nr_args
);
3734 case IORING_REGISTER_EVENTFD
:
3738 ret
= io_eventfd_register(ctx
, arg
, 0);
3740 case IORING_REGISTER_EVENTFD_ASYNC
:
3744 ret
= io_eventfd_register(ctx
, arg
, 1);
3746 case IORING_UNREGISTER_EVENTFD
:
3750 ret
= io_eventfd_unregister(ctx
);
3752 case IORING_REGISTER_PROBE
:
3754 if (!arg
|| nr_args
> 256)
3756 ret
= io_probe(ctx
, arg
, nr_args
);
3758 case IORING_REGISTER_PERSONALITY
:
3762 ret
= io_register_personality(ctx
);
3764 case IORING_UNREGISTER_PERSONALITY
:
3768 ret
= io_unregister_personality(ctx
, nr_args
);
3770 case IORING_REGISTER_ENABLE_RINGS
:
3774 ret
= io_register_enable_rings(ctx
);
3776 case IORING_REGISTER_RESTRICTIONS
:
3777 ret
= io_register_restrictions(ctx
, arg
, nr_args
);
3779 case IORING_REGISTER_FILES2
:
3780 ret
= io_register_rsrc(ctx
, arg
, nr_args
, IORING_RSRC_FILE
);
3782 case IORING_REGISTER_FILES_UPDATE2
:
3783 ret
= io_register_rsrc_update(ctx
, arg
, nr_args
,
3786 case IORING_REGISTER_BUFFERS2
:
3787 ret
= io_register_rsrc(ctx
, arg
, nr_args
, IORING_RSRC_BUFFER
);
3789 case IORING_REGISTER_BUFFERS_UPDATE
:
3790 ret
= io_register_rsrc_update(ctx
, arg
, nr_args
,
3791 IORING_RSRC_BUFFER
);
3793 case IORING_REGISTER_IOWQ_AFF
:
3795 if (!arg
|| !nr_args
)
3797 ret
= io_register_iowq_aff(ctx
, arg
, nr_args
);
3799 case IORING_UNREGISTER_IOWQ_AFF
:
3803 ret
= io_unregister_iowq_aff(ctx
);
3805 case IORING_REGISTER_IOWQ_MAX_WORKERS
:
3807 if (!arg
|| nr_args
!= 2)
3809 ret
= io_register_iowq_max_workers(ctx
, arg
);
3811 case IORING_REGISTER_RING_FDS
:
3812 ret
= io_ringfd_register(ctx
, arg
, nr_args
);
3814 case IORING_UNREGISTER_RING_FDS
:
3815 ret
= io_ringfd_unregister(ctx
, arg
, nr_args
);
3817 case IORING_REGISTER_PBUF_RING
:
3819 if (!arg
|| nr_args
!= 1)
3821 ret
= io_register_pbuf_ring(ctx
, arg
);
3823 case IORING_UNREGISTER_PBUF_RING
:
3825 if (!arg
|| nr_args
!= 1)
3827 ret
= io_unregister_pbuf_ring(ctx
, arg
);
3829 case IORING_REGISTER_SYNC_CANCEL
:
3831 if (!arg
|| nr_args
!= 1)
3833 ret
= io_sync_cancel(ctx
, arg
);
3835 case IORING_REGISTER_FILE_ALLOC_RANGE
:
3837 if (!arg
|| nr_args
)
3839 ret
= io_register_file_alloc_range(ctx
, arg
);
3849 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
3850 void __user
*, arg
, unsigned int, nr_args
)
3852 struct io_ring_ctx
*ctx
;
3861 if (!io_is_uring_fops(f
.file
))
3864 ctx
= f
.file
->private_data
;
3868 mutex_lock(&ctx
->uring_lock
);
3869 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
3870 mutex_unlock(&ctx
->uring_lock
);
3871 trace_io_uring_register(ctx
, opcode
, ctx
->nr_user_files
, ctx
->nr_user_bufs
, ret
);
3877 static int __init
io_uring_init(void)
3879 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
3880 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
3881 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
3884 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
3885 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
3886 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
3887 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
3888 BUILD_BUG_ON(sizeof(struct io_uring_sqe
) != 64);
3889 BUILD_BUG_SQE_ELEM(0, __u8
, opcode
);
3890 BUILD_BUG_SQE_ELEM(1, __u8
, flags
);
3891 BUILD_BUG_SQE_ELEM(2, __u16
, ioprio
);
3892 BUILD_BUG_SQE_ELEM(4, __s32
, fd
);
3893 BUILD_BUG_SQE_ELEM(8, __u64
, off
);
3894 BUILD_BUG_SQE_ELEM(8, __u64
, addr2
);
3895 BUILD_BUG_SQE_ELEM(8, __u32
, cmd_op
);
3896 BUILD_BUG_SQE_ELEM(12, __u32
, __pad1
);
3897 BUILD_BUG_SQE_ELEM(16, __u64
, addr
);
3898 BUILD_BUG_SQE_ELEM(16, __u64
, splice_off_in
);
3899 BUILD_BUG_SQE_ELEM(24, __u32
, len
);
3900 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t
, rw_flags
);
3901 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags
);
3902 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32
, rw_flags
);
3903 BUILD_BUG_SQE_ELEM(28, __u32
, fsync_flags
);
3904 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16
, poll_events
);
3905 BUILD_BUG_SQE_ELEM(28, __u32
, poll32_events
);
3906 BUILD_BUG_SQE_ELEM(28, __u32
, sync_range_flags
);
3907 BUILD_BUG_SQE_ELEM(28, __u32
, msg_flags
);
3908 BUILD_BUG_SQE_ELEM(28, __u32
, timeout_flags
);
3909 BUILD_BUG_SQE_ELEM(28, __u32
, accept_flags
);
3910 BUILD_BUG_SQE_ELEM(28, __u32
, cancel_flags
);
3911 BUILD_BUG_SQE_ELEM(28, __u32
, open_flags
);
3912 BUILD_BUG_SQE_ELEM(28, __u32
, statx_flags
);
3913 BUILD_BUG_SQE_ELEM(28, __u32
, fadvise_advice
);
3914 BUILD_BUG_SQE_ELEM(28, __u32
, splice_flags
);
3915 BUILD_BUG_SQE_ELEM(28, __u32
, rename_flags
);
3916 BUILD_BUG_SQE_ELEM(28, __u32
, unlink_flags
);
3917 BUILD_BUG_SQE_ELEM(28, __u32
, hardlink_flags
);
3918 BUILD_BUG_SQE_ELEM(28, __u32
, xattr_flags
);
3919 BUILD_BUG_SQE_ELEM(28, __u32
, msg_ring_flags
);
3920 BUILD_BUG_SQE_ELEM(32, __u64
, user_data
);
3921 BUILD_BUG_SQE_ELEM(40, __u16
, buf_index
);
3922 BUILD_BUG_SQE_ELEM(40, __u16
, buf_group
);
3923 BUILD_BUG_SQE_ELEM(42, __u16
, personality
);
3924 BUILD_BUG_SQE_ELEM(44, __s32
, splice_fd_in
);
3925 BUILD_BUG_SQE_ELEM(44, __u32
, file_index
);
3926 BUILD_BUG_SQE_ELEM(44, __u16
, addr_len
);
3927 BUILD_BUG_SQE_ELEM(46, __u16
, __pad3
[0]);
3928 BUILD_BUG_SQE_ELEM(48, __u64
, addr3
);
3929 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd
);
3930 BUILD_BUG_SQE_ELEM(56, __u64
, __pad2
);
3932 BUILD_BUG_ON(sizeof(struct io_uring_files_update
) !=
3933 sizeof(struct io_uring_rsrc_update
));
3934 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update
) >
3935 sizeof(struct io_uring_rsrc_update2
));
3937 /* ->buf_index is u16 */
3938 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring
, bufs
) != 0);
3939 BUILD_BUG_ON(offsetof(struct io_uring_buf
, resv
) !=
3940 offsetof(struct io_uring_buf_ring
, tail
));
3942 /* should fit into one byte */
3943 BUILD_BUG_ON(SQE_VALID_FLAGS
>= (1 << 8));
3944 BUILD_BUG_ON(SQE_COMMON_FLAGS
>= (1 << 8));
3945 BUILD_BUG_ON((SQE_VALID_FLAGS
| SQE_COMMON_FLAGS
) != SQE_VALID_FLAGS
);
3947 BUILD_BUG_ON(__REQ_F_LAST_BIT
> 8 * sizeof(int));
3949 BUILD_BUG_ON(sizeof(atomic_t
) != sizeof(u32
));
3951 io_uring_optable_init();
3953 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
|
3957 __initcall(io_uring_init
);