1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <net/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
49 #include <linux/bits.h>
51 #include <linux/sched/signal.h>
53 #include <linux/file.h>
54 #include <linux/fdtable.h>
56 #include <linux/mman.h>
57 #include <linux/percpu.h>
58 #include <linux/slab.h>
59 #include <linux/bvec.h>
60 #include <linux/net.h>
62 #include <net/af_unix.h>
64 #include <linux/anon_inodes.h>
65 #include <linux/sched/mm.h>
66 #include <linux/uaccess.h>
67 #include <linux/nospec.h>
68 #include <linux/highmem.h>
69 #include <linux/fsnotify.h>
70 #include <linux/fadvise.h>
71 #include <linux/task_work.h>
72 #include <linux/io_uring.h>
73 #include <linux/audit.h>
74 #include <linux/security.h>
75 #include <asm/shmparam.h>
77 #define CREATE_TRACE_POINTS
78 #include <trace/events/io_uring.h>
80 #include <uapi/linux/io_uring.h>
101 #include "alloc_cache.h"
103 #define IORING_MAX_ENTRIES 32768
104 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
106 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
107 IORING_REGISTER_LAST + IORING_OP_LAST)
109 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
110 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
112 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
113 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
115 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
116 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
119 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
122 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
124 #define IO_COMPL_BATCH 32
125 #define IO_REQ_ALLOC_BATCH 8
128 IO_CHECK_CQ_OVERFLOW_BIT,
129 IO_CHECK_CQ_DROPPED_BIT,
133 IO_EVENTFD_OP_SIGNAL_BIT,
134 IO_EVENTFD_OP_FREE_BIT,
137 struct io_defer_entry {
138 struct list_head list;
139 struct io_kiocb *req;
143 /* requests with any of those set should undergo io_disarm_next() */
144 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
145 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
147 static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
148 struct task_struct *task,
151 static void io_queue_sqe(struct io_kiocb *req);
153 struct kmem_cache *req_cachep;
155 static int __read_mostly sysctl_io_uring_disabled;
156 static int __read_mostly sysctl_io_uring_group = -1;
159 static struct ctl_table kernel_io_uring_disabled_table[] = {
161 .procname = "io_uring_disabled",
162 .data = &sysctl_io_uring_disabled,
163 .maxlen = sizeof(sysctl_io_uring_disabled),
165 .proc_handler = proc_dointvec_minmax,
166 .extra1 = SYSCTL_ZERO,
167 .extra2 = SYSCTL_TWO,
170 .procname = "io_uring_group",
171 .data = &sysctl_io_uring_group,
172 .maxlen = sizeof(gid_t),
174 .proc_handler = proc_dointvec,
180 struct sock *io_uring_get_socket(struct file *file)
182 #if defined(CONFIG_UNIX)
183 if (io_is_uring_fops(file)) {
184 struct io_ring_ctx *ctx = file->private_data;
186 return ctx->ring_sock->sk;
191 EXPORT_SYMBOL(io_uring_get_socket);
193 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
195 if (!wq_list_empty(&ctx->submit_state.compl_reqs) ||
196 ctx->submit_state.cqes_count)
197 __io_submit_flush_completions(ctx);
200 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
202 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
205 static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx)
207 return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head);
210 static bool io_match_linked(struct io_kiocb *head)
212 struct io_kiocb *req;
214 io_for_each_link(req, head) {
215 if (req->flags & REQ_F_INFLIGHT)
222 * As io_match_task() but protected against racing with linked timeouts.
223 * User must not hold timeout_lock.
225 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
230 if (task && head->task != task)
235 if (head->flags & REQ_F_LINK_TIMEOUT) {
236 struct io_ring_ctx *ctx = head->ctx;
238 /* protect against races with linked timeouts */
239 spin_lock_irq(&ctx->timeout_lock);
240 matched = io_match_linked(head);
241 spin_unlock_irq(&ctx->timeout_lock);
243 matched = io_match_linked(head);
248 static inline void req_fail_link_node(struct io_kiocb *req, int res)
251 io_req_set_res(req, res, 0);
254 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
256 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
259 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
261 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
263 complete(&ctx->ref_comp);
266 static __cold void io_fallback_req_func(struct work_struct *work)
268 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
270 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
271 struct io_kiocb *req, *tmp;
272 struct io_tw_state ts = { .locked = true, };
274 percpu_ref_get(&ctx->refs);
275 mutex_lock(&ctx->uring_lock);
276 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
277 req->io_task_work.func(req, &ts);
278 if (WARN_ON_ONCE(!ts.locked))
280 io_submit_flush_completions(ctx);
281 mutex_unlock(&ctx->uring_lock);
282 percpu_ref_put(&ctx->refs);
285 static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
287 unsigned hash_buckets = 1U << bits;
288 size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
290 table->hbs = kmalloc(hash_size, GFP_KERNEL);
294 table->hash_bits = bits;
295 init_hash_table(table, hash_buckets);
299 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
301 struct io_ring_ctx *ctx;
304 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
308 xa_init(&ctx->io_bl_xa);
311 * Use 5 bits less than the max cq entries, that should give us around
312 * 32 entries per hash list if totally full and uniformly spread, but
313 * don't keep too many buckets to not overconsume memory.
315 hash_bits = ilog2(p->cq_entries) - 5;
316 hash_bits = clamp(hash_bits, 1, 8);
317 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
319 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
321 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
325 ctx->flags = p->flags;
326 init_waitqueue_head(&ctx->sqo_sq_wait);
327 INIT_LIST_HEAD(&ctx->sqd_list);
328 INIT_LIST_HEAD(&ctx->cq_overflow_list);
329 INIT_LIST_HEAD(&ctx->io_buffers_cache);
330 INIT_HLIST_HEAD(&ctx->io_buf_list);
331 io_alloc_cache_init(&ctx->rsrc_node_cache, IO_NODE_ALLOC_CACHE_MAX,
332 sizeof(struct io_rsrc_node));
333 io_alloc_cache_init(&ctx->apoll_cache, IO_ALLOC_CACHE_MAX,
334 sizeof(struct async_poll));
335 io_alloc_cache_init(&ctx->netmsg_cache, IO_ALLOC_CACHE_MAX,
336 sizeof(struct io_async_msghdr));
337 io_futex_cache_init(ctx);
338 init_completion(&ctx->ref_comp);
339 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
340 mutex_init(&ctx->uring_lock);
341 init_waitqueue_head(&ctx->cq_wait);
342 init_waitqueue_head(&ctx->poll_wq);
343 init_waitqueue_head(&ctx->rsrc_quiesce_wq);
344 spin_lock_init(&ctx->completion_lock);
345 spin_lock_init(&ctx->timeout_lock);
346 INIT_WQ_LIST(&ctx->iopoll_list);
347 INIT_LIST_HEAD(&ctx->io_buffers_comp);
348 INIT_LIST_HEAD(&ctx->defer_list);
349 INIT_LIST_HEAD(&ctx->timeout_list);
350 INIT_LIST_HEAD(&ctx->ltimeout_list);
351 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
352 init_llist_head(&ctx->work_llist);
353 INIT_LIST_HEAD(&ctx->tctx_list);
354 ctx->submit_state.free_list.next = NULL;
355 INIT_WQ_LIST(&ctx->locked_free_list);
356 INIT_HLIST_HEAD(&ctx->waitid_list);
358 INIT_HLIST_HEAD(&ctx->futex_list);
360 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
361 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
362 INIT_HLIST_HEAD(&ctx->cancelable_uring_cmd);
365 kfree(ctx->cancel_table.hbs);
366 kfree(ctx->cancel_table_locked.hbs);
368 xa_destroy(&ctx->io_bl_xa);
373 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
375 struct io_rings *r = ctx->rings;
377 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
381 static bool req_need_defer(struct io_kiocb *req, u32 seq)
383 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
384 struct io_ring_ctx *ctx = req->ctx;
386 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
392 static void io_clean_op(struct io_kiocb *req)
394 if (req->flags & REQ_F_BUFFER_SELECTED) {
395 spin_lock(&req->ctx->completion_lock);
396 io_put_kbuf_comp(req);
397 spin_unlock(&req->ctx->completion_lock);
400 if (req->flags & REQ_F_NEED_CLEANUP) {
401 const struct io_cold_def *def = &io_cold_defs[req->opcode];
406 if ((req->flags & REQ_F_POLLED) && req->apoll) {
407 kfree(req->apoll->double_poll);
411 if (req->flags & REQ_F_INFLIGHT) {
412 struct io_uring_task *tctx = req->task->io_uring;
414 atomic_dec(&tctx->inflight_tracked);
416 if (req->flags & REQ_F_CREDS)
417 put_cred(req->creds);
418 if (req->flags & REQ_F_ASYNC_DATA) {
419 kfree(req->async_data);
420 req->async_data = NULL;
422 req->flags &= ~IO_REQ_CLEAN_FLAGS;
425 static inline void io_req_track_inflight(struct io_kiocb *req)
427 if (!(req->flags & REQ_F_INFLIGHT)) {
428 req->flags |= REQ_F_INFLIGHT;
429 atomic_inc(&req->task->io_uring->inflight_tracked);
433 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
435 if (WARN_ON_ONCE(!req->link))
438 req->flags &= ~REQ_F_ARM_LTIMEOUT;
439 req->flags |= REQ_F_LINK_TIMEOUT;
441 /* linked timeouts should have two refs once prep'ed */
442 io_req_set_refcount(req);
443 __io_req_set_refcount(req->link, 2);
447 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
449 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
451 return __io_prep_linked_timeout(req);
454 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
456 io_queue_linked_timeout(__io_prep_linked_timeout(req));
459 static inline void io_arm_ltimeout(struct io_kiocb *req)
461 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
462 __io_arm_ltimeout(req);
465 static void io_prep_async_work(struct io_kiocb *req)
467 const struct io_issue_def *def = &io_issue_defs[req->opcode];
468 struct io_ring_ctx *ctx = req->ctx;
470 if (!(req->flags & REQ_F_CREDS)) {
471 req->flags |= REQ_F_CREDS;
472 req->creds = get_current_cred();
475 req->work.list.next = NULL;
477 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
478 if (req->flags & REQ_F_FORCE_ASYNC)
479 req->work.flags |= IO_WQ_WORK_CONCURRENT;
481 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
482 req->flags |= io_file_get_flags(req->file);
484 if (req->file && (req->flags & REQ_F_ISREG)) {
485 bool should_hash = def->hash_reg_file;
487 /* don't serialize this request if the fs doesn't need it */
488 if (should_hash && (req->file->f_flags & O_DIRECT) &&
489 (req->file->f_mode & FMODE_DIO_PARALLEL_WRITE))
491 if (should_hash || (ctx->flags & IORING_SETUP_IOPOLL))
492 io_wq_hash_work(&req->work, file_inode(req->file));
493 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
494 if (def->unbound_nonreg_file)
495 req->work.flags |= IO_WQ_WORK_UNBOUND;
499 static void io_prep_async_link(struct io_kiocb *req)
501 struct io_kiocb *cur;
503 if (req->flags & REQ_F_LINK_TIMEOUT) {
504 struct io_ring_ctx *ctx = req->ctx;
506 spin_lock_irq(&ctx->timeout_lock);
507 io_for_each_link(cur, req)
508 io_prep_async_work(cur);
509 spin_unlock_irq(&ctx->timeout_lock);
511 io_for_each_link(cur, req)
512 io_prep_async_work(cur);
516 void io_queue_iowq(struct io_kiocb *req, struct io_tw_state *ts_dont_use)
518 struct io_kiocb *link = io_prep_linked_timeout(req);
519 struct io_uring_task *tctx = req->task->io_uring;
522 BUG_ON(!tctx->io_wq);
524 /* init ->work of the whole link before punting */
525 io_prep_async_link(req);
528 * Not expected to happen, but if we do have a bug where this _can_
529 * happen, catch it here and ensure the request is marked as
530 * canceled. That will make io-wq go through the usual work cancel
531 * procedure rather than attempt to run this request (or create a new
534 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
535 req->work.flags |= IO_WQ_WORK_CANCEL;
537 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
538 io_wq_enqueue(tctx->io_wq, &req->work);
540 io_queue_linked_timeout(link);
543 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
545 while (!list_empty(&ctx->defer_list)) {
546 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
547 struct io_defer_entry, list);
549 if (req_need_defer(de->req, de->seq))
551 list_del_init(&de->list);
552 io_req_task_queue(de->req);
558 static void io_eventfd_ops(struct rcu_head *rcu)
560 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
561 int ops = atomic_xchg(&ev_fd->ops, 0);
563 if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT))
564 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
566 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
567 * ordering in a race but if references are 0 we know we have to free
570 if (atomic_dec_and_test(&ev_fd->refs)) {
571 eventfd_ctx_put(ev_fd->cq_ev_fd);
576 static void io_eventfd_signal(struct io_ring_ctx *ctx)
578 struct io_ev_fd *ev_fd = NULL;
582 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
585 ev_fd = rcu_dereference(ctx->io_ev_fd);
588 * Check again if ev_fd exists incase an io_eventfd_unregister call
589 * completed between the NULL check of ctx->io_ev_fd at the start of
590 * the function and rcu_read_lock.
592 if (unlikely(!ev_fd))
594 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
596 if (ev_fd->eventfd_async && !io_wq_current_is_worker())
599 if (likely(eventfd_signal_allowed())) {
600 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
602 atomic_inc(&ev_fd->refs);
603 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops))
604 call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops);
606 atomic_dec(&ev_fd->refs);
613 static void io_eventfd_flush_signal(struct io_ring_ctx *ctx)
617 spin_lock(&ctx->completion_lock);
620 * Eventfd should only get triggered when at least one event has been
621 * posted. Some applications rely on the eventfd notification count
622 * only changing IFF a new CQE has been added to the CQ ring. There's
623 * no depedency on 1:1 relationship between how many times this
624 * function is called (and hence the eventfd count) and number of CQEs
625 * posted to the CQ ring.
627 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
628 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
629 spin_unlock(&ctx->completion_lock);
633 io_eventfd_signal(ctx);
636 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
638 if (ctx->poll_activated)
639 io_poll_wq_wake(ctx);
640 if (ctx->off_timeout_used)
641 io_flush_timeouts(ctx);
642 if (ctx->drain_active) {
643 spin_lock(&ctx->completion_lock);
644 io_queue_deferred(ctx);
645 spin_unlock(&ctx->completion_lock);
648 io_eventfd_flush_signal(ctx);
651 static inline void __io_cq_lock(struct io_ring_ctx *ctx)
653 if (!ctx->lockless_cq)
654 spin_lock(&ctx->completion_lock);
657 static inline void io_cq_lock(struct io_ring_ctx *ctx)
658 __acquires(ctx->completion_lock)
660 spin_lock(&ctx->completion_lock);
663 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
665 io_commit_cqring(ctx);
666 if (!ctx->task_complete) {
667 if (!ctx->lockless_cq)
668 spin_unlock(&ctx->completion_lock);
669 /* IOPOLL rings only need to wake up if it's also SQPOLL */
670 if (!ctx->syscall_iopoll)
673 io_commit_cqring_flush(ctx);
676 static void io_cq_unlock_post(struct io_ring_ctx *ctx)
677 __releases(ctx->completion_lock)
679 io_commit_cqring(ctx);
680 spin_unlock(&ctx->completion_lock);
682 io_commit_cqring_flush(ctx);
685 /* Returns true if there are no backlogged entries after the flush */
686 static void io_cqring_overflow_kill(struct io_ring_ctx *ctx)
688 struct io_overflow_cqe *ocqe;
691 spin_lock(&ctx->completion_lock);
692 list_splice_init(&ctx->cq_overflow_list, &list);
693 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
694 spin_unlock(&ctx->completion_lock);
696 while (!list_empty(&list)) {
697 ocqe = list_first_entry(&list, struct io_overflow_cqe, list);
698 list_del(&ocqe->list);
703 static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx)
705 size_t cqe_size = sizeof(struct io_uring_cqe);
707 if (__io_cqring_events(ctx) == ctx->cq_entries)
710 if (ctx->flags & IORING_SETUP_CQE32)
714 while (!list_empty(&ctx->cq_overflow_list)) {
715 struct io_uring_cqe *cqe;
716 struct io_overflow_cqe *ocqe;
718 if (!io_get_cqe_overflow(ctx, &cqe, true))
720 ocqe = list_first_entry(&ctx->cq_overflow_list,
721 struct io_overflow_cqe, list);
722 memcpy(cqe, &ocqe->cqe, cqe_size);
723 list_del(&ocqe->list);
727 if (list_empty(&ctx->cq_overflow_list)) {
728 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
729 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
731 io_cq_unlock_post(ctx);
734 static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx)
736 /* iopoll syncs against uring_lock, not completion_lock */
737 if (ctx->flags & IORING_SETUP_IOPOLL)
738 mutex_lock(&ctx->uring_lock);
739 __io_cqring_overflow_flush(ctx);
740 if (ctx->flags & IORING_SETUP_IOPOLL)
741 mutex_unlock(&ctx->uring_lock);
744 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx)
746 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
747 io_cqring_do_overflow_flush(ctx);
750 /* can be called by any task */
751 static void io_put_task_remote(struct task_struct *task)
753 struct io_uring_task *tctx = task->io_uring;
755 percpu_counter_sub(&tctx->inflight, 1);
756 if (unlikely(atomic_read(&tctx->in_cancel)))
757 wake_up(&tctx->wait);
758 put_task_struct(task);
761 /* used by a task to put its own references */
762 static void io_put_task_local(struct task_struct *task)
764 task->io_uring->cached_refs++;
767 /* must to be called somewhat shortly after putting a request */
768 static inline void io_put_task(struct task_struct *task)
770 if (likely(task == current))
771 io_put_task_local(task);
773 io_put_task_remote(task);
776 void io_task_refs_refill(struct io_uring_task *tctx)
778 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
780 percpu_counter_add(&tctx->inflight, refill);
781 refcount_add(refill, ¤t->usage);
782 tctx->cached_refs += refill;
785 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
787 struct io_uring_task *tctx = task->io_uring;
788 unsigned int refs = tctx->cached_refs;
791 tctx->cached_refs = 0;
792 percpu_counter_sub(&tctx->inflight, refs);
793 put_task_struct_many(task, refs);
797 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
798 s32 res, u32 cflags, u64 extra1, u64 extra2)
800 struct io_overflow_cqe *ocqe;
801 size_t ocq_size = sizeof(struct io_overflow_cqe);
802 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
804 lockdep_assert_held(&ctx->completion_lock);
807 ocq_size += sizeof(struct io_uring_cqe);
809 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
810 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
813 * If we're in ring overflow flush mode, or in task cancel mode,
814 * or cannot allocate an overflow entry, then we need to drop it
817 io_account_cq_overflow(ctx);
818 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
821 if (list_empty(&ctx->cq_overflow_list)) {
822 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
823 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
826 ocqe->cqe.user_data = user_data;
828 ocqe->cqe.flags = cflags;
830 ocqe->cqe.big_cqe[0] = extra1;
831 ocqe->cqe.big_cqe[1] = extra2;
833 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
837 void io_req_cqe_overflow(struct io_kiocb *req)
839 io_cqring_event_overflow(req->ctx, req->cqe.user_data,
840 req->cqe.res, req->cqe.flags,
841 req->big_cqe.extra1, req->big_cqe.extra2);
842 memset(&req->big_cqe, 0, sizeof(req->big_cqe));
846 * writes to the cq entry need to come after reading head; the
847 * control dependency is enough as we're using WRITE_ONCE to
850 bool io_cqe_cache_refill(struct io_ring_ctx *ctx, bool overflow)
852 struct io_rings *rings = ctx->rings;
853 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
854 unsigned int free, queued, len;
857 * Posting into the CQ when there are pending overflowed CQEs may break
858 * ordering guarantees, which will affect links, F_MORE users and more.
859 * Force overflow the completion.
861 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
864 /* userspace may cheat modifying the tail, be safe and do min */
865 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
866 free = ctx->cq_entries - queued;
867 /* we need a contiguous range, limit based on the current array offset */
868 len = min(free, ctx->cq_entries - off);
872 if (ctx->flags & IORING_SETUP_CQE32) {
877 ctx->cqe_cached = &rings->cqes[off];
878 ctx->cqe_sentinel = ctx->cqe_cached + len;
882 static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
885 struct io_uring_cqe *cqe;
890 * If we can't get a cq entry, userspace overflowed the
891 * submission (by quite a lot). Increment the overflow count in
894 if (likely(io_get_cqe(ctx, &cqe))) {
895 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
897 WRITE_ONCE(cqe->user_data, user_data);
898 WRITE_ONCE(cqe->res, res);
899 WRITE_ONCE(cqe->flags, cflags);
901 if (ctx->flags & IORING_SETUP_CQE32) {
902 WRITE_ONCE(cqe->big_cqe[0], 0);
903 WRITE_ONCE(cqe->big_cqe[1], 0);
910 static void __io_flush_post_cqes(struct io_ring_ctx *ctx)
911 __must_hold(&ctx->uring_lock)
913 struct io_submit_state *state = &ctx->submit_state;
916 lockdep_assert_held(&ctx->uring_lock);
917 for (i = 0; i < state->cqes_count; i++) {
918 struct io_uring_cqe *cqe = &ctx->completion_cqes[i];
920 if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) {
921 if (ctx->lockless_cq) {
922 spin_lock(&ctx->completion_lock);
923 io_cqring_event_overflow(ctx, cqe->user_data,
924 cqe->res, cqe->flags, 0, 0);
925 spin_unlock(&ctx->completion_lock);
927 io_cqring_event_overflow(ctx, cqe->user_data,
928 cqe->res, cqe->flags, 0, 0);
932 state->cqes_count = 0;
935 static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
941 filled = io_fill_cqe_aux(ctx, user_data, res, cflags);
942 if (!filled && allow_overflow)
943 filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
945 io_cq_unlock_post(ctx);
949 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags)
951 return __io_post_aux_cqe(ctx, user_data, res, cflags, true);
955 * A helper for multishot requests posting additional CQEs.
956 * Should only be used from a task_work including IO_URING_F_MULTISHOT.
958 bool io_fill_cqe_req_aux(struct io_kiocb *req, bool defer, s32 res, u32 cflags)
960 struct io_ring_ctx *ctx = req->ctx;
961 u64 user_data = req->cqe.user_data;
962 struct io_uring_cqe *cqe;
965 return __io_post_aux_cqe(ctx, user_data, res, cflags, false);
967 lockdep_assert_held(&ctx->uring_lock);
969 if (ctx->submit_state.cqes_count == ARRAY_SIZE(ctx->completion_cqes)) {
971 __io_flush_post_cqes(ctx);
972 /* no need to flush - flush is deferred */
973 __io_cq_unlock_post(ctx);
976 /* For defered completions this is not as strict as it is otherwise,
977 * however it's main job is to prevent unbounded posted completions,
978 * and in that it works just as well.
980 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
983 cqe = &ctx->completion_cqes[ctx->submit_state.cqes_count++];
984 cqe->user_data = user_data;
990 static void __io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
992 struct io_ring_ctx *ctx = req->ctx;
993 struct io_rsrc_node *rsrc_node = NULL;
996 if (!(req->flags & REQ_F_CQE_SKIP)) {
997 if (!io_fill_cqe_req(ctx, req))
998 io_req_cqe_overflow(req);
1002 * If we're the last reference to this request, add to our locked
1005 if (req_ref_put_and_test(req)) {
1006 if (req->flags & IO_REQ_LINK_FLAGS) {
1007 if (req->flags & IO_DISARM_MASK)
1008 io_disarm_next(req);
1010 io_req_task_queue(req->link);
1014 io_put_kbuf_comp(req);
1015 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1019 rsrc_node = req->rsrc_node;
1021 * Selected buffer deallocation in io_clean_op() assumes that
1022 * we don't hold ->completion_lock. Clean them here to avoid
1025 io_put_task_remote(req->task);
1026 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
1027 ctx->locked_free_nr++;
1029 io_cq_unlock_post(ctx);
1032 io_ring_submit_lock(ctx, issue_flags);
1033 io_put_rsrc_node(ctx, rsrc_node);
1034 io_ring_submit_unlock(ctx, issue_flags);
1038 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
1040 if (req->ctx->task_complete && req->ctx->submitter_task != current) {
1041 req->io_task_work.func = io_req_task_complete;
1042 io_req_task_work_add(req);
1043 } else if (!(issue_flags & IO_URING_F_UNLOCKED) ||
1044 !(req->ctx->flags & IORING_SETUP_IOPOLL)) {
1045 __io_req_complete_post(req, issue_flags);
1047 struct io_ring_ctx *ctx = req->ctx;
1049 mutex_lock(&ctx->uring_lock);
1050 __io_req_complete_post(req, issue_flags & ~IO_URING_F_UNLOCKED);
1051 mutex_unlock(&ctx->uring_lock);
1055 void io_req_defer_failed(struct io_kiocb *req, s32 res)
1056 __must_hold(&ctx->uring_lock)
1058 const struct io_cold_def *def = &io_cold_defs[req->opcode];
1060 lockdep_assert_held(&req->ctx->uring_lock);
1063 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
1066 io_req_complete_defer(req);
1070 * Don't initialise the fields below on every allocation, but do that in
1071 * advance and keep them valid across allocations.
1073 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1077 req->async_data = NULL;
1078 /* not necessary, but safer to zero */
1079 memset(&req->cqe, 0, sizeof(req->cqe));
1080 memset(&req->big_cqe, 0, sizeof(req->big_cqe));
1083 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1084 struct io_submit_state *state)
1086 spin_lock(&ctx->completion_lock);
1087 wq_list_splice(&ctx->locked_free_list, &state->free_list);
1088 ctx->locked_free_nr = 0;
1089 spin_unlock(&ctx->completion_lock);
1093 * A request might get retired back into the request caches even before opcode
1094 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1095 * Because of that, io_alloc_req() should be called only under ->uring_lock
1096 * and with extra caution to not get a request that is still worked on.
1098 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
1099 __must_hold(&ctx->uring_lock)
1101 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1102 void *reqs[IO_REQ_ALLOC_BATCH];
1106 * If we have more than a batch's worth of requests in our IRQ side
1107 * locked cache, grab the lock and move them over to our submission
1110 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
1111 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
1112 if (!io_req_cache_empty(ctx))
1116 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
1119 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1120 * retry single alloc to be on the safe side.
1122 if (unlikely(ret <= 0)) {
1123 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1129 percpu_ref_get_many(&ctx->refs, ret);
1130 for (i = 0; i < ret; i++) {
1131 struct io_kiocb *req = reqs[i];
1133 io_preinit_req(req, ctx);
1134 io_req_add_to_cache(req, ctx);
1139 __cold void io_free_req(struct io_kiocb *req)
1141 /* refs were already put, restore them for io_req_task_complete() */
1142 req->flags &= ~REQ_F_REFCOUNT;
1143 /* we only want to free it, don't post CQEs */
1144 req->flags |= REQ_F_CQE_SKIP;
1145 req->io_task_work.func = io_req_task_complete;
1146 io_req_task_work_add(req);
1149 static void __io_req_find_next_prep(struct io_kiocb *req)
1151 struct io_ring_ctx *ctx = req->ctx;
1153 spin_lock(&ctx->completion_lock);
1154 io_disarm_next(req);
1155 spin_unlock(&ctx->completion_lock);
1158 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1160 struct io_kiocb *nxt;
1163 * If LINK is set, we have dependent requests in this chain. If we
1164 * didn't fail this request, queue the first one up, moving any other
1165 * dependencies to the next request. In case of failure, fail the rest
1168 if (unlikely(req->flags & IO_DISARM_MASK))
1169 __io_req_find_next_prep(req);
1175 static void ctx_flush_and_put(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1179 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1180 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1182 io_submit_flush_completions(ctx);
1183 mutex_unlock(&ctx->uring_lock);
1186 percpu_ref_put(&ctx->refs);
1189 static unsigned int handle_tw_list(struct llist_node *node,
1190 struct io_ring_ctx **ctx,
1191 struct io_tw_state *ts,
1192 struct llist_node *last)
1194 unsigned int count = 0;
1196 while (node && node != last) {
1197 struct llist_node *next = node->next;
1198 struct io_kiocb *req = container_of(node, struct io_kiocb,
1201 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1203 if (req->ctx != *ctx) {
1204 ctx_flush_and_put(*ctx, ts);
1206 /* if not contended, grab and improve batching */
1207 ts->locked = mutex_trylock(&(*ctx)->uring_lock);
1208 percpu_ref_get(&(*ctx)->refs);
1210 INDIRECT_CALL_2(req->io_task_work.func,
1211 io_poll_task_func, io_req_rw_complete,
1215 if (unlikely(need_resched())) {
1216 ctx_flush_and_put(*ctx, ts);
1226 * io_llist_xchg - swap all entries in a lock-less list
1227 * @head: the head of lock-less list to delete all entries
1228 * @new: new entry as the head of the list
1230 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1231 * The order of entries returned is from the newest to the oldest added one.
1233 static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1234 struct llist_node *new)
1236 return xchg(&head->first, new);
1240 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1241 * @head: the head of lock-less list to delete all entries
1242 * @old: expected old value of the first entry of the list
1243 * @new: new entry as the head of the list
1245 * perform a cmpxchg on the first entry of the list.
1248 static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1249 struct llist_node *old,
1250 struct llist_node *new)
1252 return cmpxchg(&head->first, old, new);
1255 static __cold void io_fallback_tw(struct io_uring_task *tctx, bool sync)
1257 struct llist_node *node = llist_del_all(&tctx->task_list);
1258 struct io_ring_ctx *last_ctx = NULL;
1259 struct io_kiocb *req;
1262 req = container_of(node, struct io_kiocb, io_task_work.node);
1264 if (sync && last_ctx != req->ctx) {
1266 flush_delayed_work(&last_ctx->fallback_work);
1267 percpu_ref_put(&last_ctx->refs);
1269 last_ctx = req->ctx;
1270 percpu_ref_get(&last_ctx->refs);
1272 if (llist_add(&req->io_task_work.node,
1273 &req->ctx->fallback_llist))
1274 schedule_delayed_work(&req->ctx->fallback_work, 1);
1278 flush_delayed_work(&last_ctx->fallback_work);
1279 percpu_ref_put(&last_ctx->refs);
1283 void tctx_task_work(struct callback_head *cb)
1285 struct io_tw_state ts = {};
1286 struct io_ring_ctx *ctx = NULL;
1287 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1289 struct llist_node fake = {};
1290 struct llist_node *node;
1291 unsigned int loops = 0;
1292 unsigned int count = 0;
1294 if (unlikely(current->flags & PF_EXITING)) {
1295 io_fallback_tw(tctx, true);
1301 node = io_llist_xchg(&tctx->task_list, &fake);
1302 count += handle_tw_list(node, &ctx, &ts, &fake);
1304 /* skip expensive cmpxchg if there are items in the list */
1305 if (READ_ONCE(tctx->task_list.first) != &fake)
1307 if (ts.locked && !wq_list_empty(&ctx->submit_state.compl_reqs)) {
1308 io_submit_flush_completions(ctx);
1309 if (READ_ONCE(tctx->task_list.first) != &fake)
1312 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1313 } while (node != &fake);
1315 ctx_flush_and_put(ctx, &ts);
1317 /* relaxed read is enough as only the task itself sets ->in_cancel */
1318 if (unlikely(atomic_read(&tctx->in_cancel)))
1319 io_uring_drop_tctx_refs(current);
1321 trace_io_uring_task_work_run(tctx, count, loops);
1324 static inline void io_req_local_work_add(struct io_kiocb *req, unsigned flags)
1326 struct io_ring_ctx *ctx = req->ctx;
1327 unsigned nr_wait, nr_tw, nr_tw_prev;
1328 struct llist_node *first;
1330 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK))
1331 flags &= ~IOU_F_TWQ_LAZY_WAKE;
1333 first = READ_ONCE(ctx->work_llist.first);
1337 struct io_kiocb *first_req = container_of(first,
1341 * Might be executed at any moment, rely on
1342 * SLAB_TYPESAFE_BY_RCU to keep it alive.
1344 nr_tw_prev = READ_ONCE(first_req->nr_tw);
1346 nr_tw = nr_tw_prev + 1;
1347 /* Large enough to fail the nr_wait comparison below */
1348 if (!(flags & IOU_F_TWQ_LAZY_WAKE))
1352 req->io_task_work.node.next = first;
1353 } while (!try_cmpxchg(&ctx->work_llist.first, &first,
1354 &req->io_task_work.node));
1357 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1358 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1360 io_eventfd_signal(ctx);
1363 nr_wait = atomic_read(&ctx->cq_wait_nr);
1364 /* no one is waiting */
1367 /* either not enough or the previous add has already woken it up */
1368 if (nr_wait > nr_tw || nr_tw_prev >= nr_wait)
1370 /* pairs with set_current_state() in io_cqring_wait() */
1371 smp_mb__after_atomic();
1372 wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE);
1375 static void io_req_normal_work_add(struct io_kiocb *req)
1377 struct io_uring_task *tctx = req->task->io_uring;
1378 struct io_ring_ctx *ctx = req->ctx;
1380 /* task_work already pending, we're done */
1381 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1384 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1385 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1387 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1390 io_fallback_tw(tctx, false);
1393 void __io_req_task_work_add(struct io_kiocb *req, unsigned flags)
1395 if (req->ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
1397 io_req_local_work_add(req, flags);
1400 io_req_normal_work_add(req);
1404 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1406 struct llist_node *node;
1408 node = llist_del_all(&ctx->work_llist);
1410 struct io_kiocb *req = container_of(node, struct io_kiocb,
1414 io_req_normal_work_add(req);
1418 static int __io_run_local_work(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1420 struct llist_node *node;
1421 unsigned int loops = 0;
1424 if (WARN_ON_ONCE(ctx->submitter_task != current))
1426 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1427 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1430 * llists are in reverse order, flip it back the right way before
1431 * running the pending items.
1433 node = llist_reverse_order(io_llist_xchg(&ctx->work_llist, NULL));
1435 struct llist_node *next = node->next;
1436 struct io_kiocb *req = container_of(node, struct io_kiocb,
1438 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1439 INDIRECT_CALL_2(req->io_task_work.func,
1440 io_poll_task_func, io_req_rw_complete,
1447 if (!llist_empty(&ctx->work_llist))
1450 io_submit_flush_completions(ctx);
1451 if (!llist_empty(&ctx->work_llist))
1454 trace_io_uring_local_work_run(ctx, ret, loops);
1458 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx)
1460 struct io_tw_state ts = { .locked = true, };
1463 if (llist_empty(&ctx->work_llist))
1466 ret = __io_run_local_work(ctx, &ts);
1467 /* shouldn't happen! */
1468 if (WARN_ON_ONCE(!ts.locked))
1469 mutex_lock(&ctx->uring_lock);
1473 static int io_run_local_work(struct io_ring_ctx *ctx)
1475 struct io_tw_state ts = {};
1478 ts.locked = mutex_trylock(&ctx->uring_lock);
1479 ret = __io_run_local_work(ctx, &ts);
1481 mutex_unlock(&ctx->uring_lock);
1486 static void io_req_task_cancel(struct io_kiocb *req, struct io_tw_state *ts)
1488 io_tw_lock(req->ctx, ts);
1489 io_req_defer_failed(req, req->cqe.res);
1492 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts)
1494 io_tw_lock(req->ctx, ts);
1495 /* req->task == current here, checking PF_EXITING is safe */
1496 if (unlikely(req->task->flags & PF_EXITING))
1497 io_req_defer_failed(req, -EFAULT);
1498 else if (req->flags & REQ_F_FORCE_ASYNC)
1499 io_queue_iowq(req, ts);
1504 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1506 io_req_set_res(req, ret, 0);
1507 req->io_task_work.func = io_req_task_cancel;
1508 io_req_task_work_add(req);
1511 void io_req_task_queue(struct io_kiocb *req)
1513 req->io_task_work.func = io_req_task_submit;
1514 io_req_task_work_add(req);
1517 void io_queue_next(struct io_kiocb *req)
1519 struct io_kiocb *nxt = io_req_find_next(req);
1522 io_req_task_queue(nxt);
1525 static void io_free_batch_list(struct io_ring_ctx *ctx,
1526 struct io_wq_work_node *node)
1527 __must_hold(&ctx->uring_lock)
1530 struct io_kiocb *req = container_of(node, struct io_kiocb,
1533 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1534 if (req->flags & REQ_F_REFCOUNT) {
1535 node = req->comp_list.next;
1536 if (!req_ref_put_and_test(req))
1539 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1540 struct async_poll *apoll = req->apoll;
1542 if (apoll->double_poll)
1543 kfree(apoll->double_poll);
1544 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1546 req->flags &= ~REQ_F_POLLED;
1548 if (req->flags & IO_REQ_LINK_FLAGS)
1550 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1555 io_req_put_rsrc_locked(req, ctx);
1557 io_put_task(req->task);
1558 node = req->comp_list.next;
1559 io_req_add_to_cache(req, ctx);
1563 void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1564 __must_hold(&ctx->uring_lock)
1566 struct io_submit_state *state = &ctx->submit_state;
1567 struct io_wq_work_node *node;
1570 /* must come first to preserve CQE ordering in failure cases */
1571 if (state->cqes_count)
1572 __io_flush_post_cqes(ctx);
1573 __wq_list_for_each(node, &state->compl_reqs) {
1574 struct io_kiocb *req = container_of(node, struct io_kiocb,
1577 if (!(req->flags & REQ_F_CQE_SKIP) &&
1578 unlikely(!io_fill_cqe_req(ctx, req))) {
1579 if (ctx->lockless_cq) {
1580 spin_lock(&ctx->completion_lock);
1581 io_req_cqe_overflow(req);
1582 spin_unlock(&ctx->completion_lock);
1584 io_req_cqe_overflow(req);
1588 __io_cq_unlock_post(ctx);
1590 if (!wq_list_empty(&ctx->submit_state.compl_reqs)) {
1591 io_free_batch_list(ctx, state->compl_reqs.first);
1592 INIT_WQ_LIST(&state->compl_reqs);
1596 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1598 /* See comment at the top of this file */
1600 return __io_cqring_events(ctx);
1604 * We can't just wait for polled events to come to us, we have to actively
1605 * find and complete them.
1607 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1609 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1612 mutex_lock(&ctx->uring_lock);
1613 while (!wq_list_empty(&ctx->iopoll_list)) {
1614 /* let it sleep and repeat later if can't complete a request */
1615 if (io_do_iopoll(ctx, true) == 0)
1618 * Ensure we allow local-to-the-cpu processing to take place,
1619 * in this case we need to ensure that we reap all events.
1620 * Also let task_work, etc. to progress by releasing the mutex
1622 if (need_resched()) {
1623 mutex_unlock(&ctx->uring_lock);
1625 mutex_lock(&ctx->uring_lock);
1628 mutex_unlock(&ctx->uring_lock);
1631 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1633 unsigned int nr_events = 0;
1634 unsigned long check_cq;
1636 if (!io_allowed_run_tw(ctx))
1639 check_cq = READ_ONCE(ctx->check_cq);
1640 if (unlikely(check_cq)) {
1641 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1642 __io_cqring_overflow_flush(ctx);
1644 * Similarly do not spin if we have not informed the user of any
1647 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1651 * Don't enter poll loop if we already have events pending.
1652 * If we do, we can potentially be spinning for commands that
1653 * already triggered a CQE (eg in error).
1655 if (io_cqring_events(ctx))
1662 * If a submit got punted to a workqueue, we can have the
1663 * application entering polling for a command before it gets
1664 * issued. That app will hold the uring_lock for the duration
1665 * of the poll right here, so we need to take a breather every
1666 * now and then to ensure that the issue has a chance to add
1667 * the poll to the issued list. Otherwise we can spin here
1668 * forever, while the workqueue is stuck trying to acquire the
1671 if (wq_list_empty(&ctx->iopoll_list) ||
1672 io_task_work_pending(ctx)) {
1673 u32 tail = ctx->cached_cq_tail;
1675 (void) io_run_local_work_locked(ctx);
1677 if (task_work_pending(current) ||
1678 wq_list_empty(&ctx->iopoll_list)) {
1679 mutex_unlock(&ctx->uring_lock);
1681 mutex_lock(&ctx->uring_lock);
1683 /* some requests don't go through iopoll_list */
1684 if (tail != ctx->cached_cq_tail ||
1685 wq_list_empty(&ctx->iopoll_list))
1688 ret = io_do_iopoll(ctx, !min);
1689 if (unlikely(ret < 0))
1692 if (task_sigpending(current))
1698 } while (nr_events < min);
1703 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts)
1706 io_req_complete_defer(req);
1708 io_req_complete_post(req, IO_URING_F_UNLOCKED);
1712 * After the iocb has been issued, it's safe to be found on the poll list.
1713 * Adding the kiocb to the list AFTER submission ensures that we don't
1714 * find it from a io_do_iopoll() thread before the issuer is done
1715 * accessing the kiocb cookie.
1717 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1719 struct io_ring_ctx *ctx = req->ctx;
1720 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1722 /* workqueue context doesn't hold uring_lock, grab it now */
1723 if (unlikely(needs_lock))
1724 mutex_lock(&ctx->uring_lock);
1727 * Track whether we have multiple files in our lists. This will impact
1728 * how we do polling eventually, not spinning if we're on potentially
1729 * different devices.
1731 if (wq_list_empty(&ctx->iopoll_list)) {
1732 ctx->poll_multi_queue = false;
1733 } else if (!ctx->poll_multi_queue) {
1734 struct io_kiocb *list_req;
1736 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1738 if (list_req->file != req->file)
1739 ctx->poll_multi_queue = true;
1743 * For fast devices, IO may have already completed. If it has, add
1744 * it to the front so we find it first.
1746 if (READ_ONCE(req->iopoll_completed))
1747 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1749 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1751 if (unlikely(needs_lock)) {
1753 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1754 * in sq thread task context or in io worker task context. If
1755 * current task context is sq thread, we don't need to check
1756 * whether should wake up sq thread.
1758 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1759 wq_has_sleeper(&ctx->sq_data->wait))
1760 wake_up(&ctx->sq_data->wait);
1762 mutex_unlock(&ctx->uring_lock);
1766 unsigned int io_file_get_flags(struct file *file)
1768 unsigned int res = 0;
1770 if (S_ISREG(file_inode(file)->i_mode))
1772 if ((file->f_flags & O_NONBLOCK) || (file->f_mode & FMODE_NOWAIT))
1773 res |= REQ_F_SUPPORT_NOWAIT;
1777 bool io_alloc_async_data(struct io_kiocb *req)
1779 WARN_ON_ONCE(!io_cold_defs[req->opcode].async_size);
1780 req->async_data = kmalloc(io_cold_defs[req->opcode].async_size, GFP_KERNEL);
1781 if (req->async_data) {
1782 req->flags |= REQ_F_ASYNC_DATA;
1788 int io_req_prep_async(struct io_kiocb *req)
1790 const struct io_cold_def *cdef = &io_cold_defs[req->opcode];
1791 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1793 /* assign early for deferred execution for non-fixed file */
1794 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE) && !req->file)
1795 req->file = io_file_get_normal(req, req->cqe.fd);
1796 if (!cdef->prep_async)
1798 if (WARN_ON_ONCE(req_has_async_data(req)))
1800 if (!def->manual_alloc) {
1801 if (io_alloc_async_data(req))
1804 return cdef->prep_async(req);
1807 static u32 io_get_sequence(struct io_kiocb *req)
1809 u32 seq = req->ctx->cached_sq_head;
1810 struct io_kiocb *cur;
1812 /* need original cached_sq_head, but it was increased for each req */
1813 io_for_each_link(cur, req)
1818 static __cold void io_drain_req(struct io_kiocb *req)
1819 __must_hold(&ctx->uring_lock)
1821 struct io_ring_ctx *ctx = req->ctx;
1822 struct io_defer_entry *de;
1824 u32 seq = io_get_sequence(req);
1826 /* Still need defer if there is pending req in defer list. */
1827 spin_lock(&ctx->completion_lock);
1828 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1829 spin_unlock(&ctx->completion_lock);
1831 ctx->drain_active = false;
1832 io_req_task_queue(req);
1835 spin_unlock(&ctx->completion_lock);
1837 io_prep_async_link(req);
1838 de = kmalloc(sizeof(*de), GFP_KERNEL);
1841 io_req_defer_failed(req, ret);
1845 spin_lock(&ctx->completion_lock);
1846 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1847 spin_unlock(&ctx->completion_lock);
1852 trace_io_uring_defer(req);
1855 list_add_tail(&de->list, &ctx->defer_list);
1856 spin_unlock(&ctx->completion_lock);
1859 static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def,
1860 unsigned int issue_flags)
1862 if (req->file || !def->needs_file)
1865 if (req->flags & REQ_F_FIXED_FILE)
1866 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1868 req->file = io_file_get_normal(req, req->cqe.fd);
1873 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1875 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1876 const struct cred *creds = NULL;
1879 if (unlikely(!io_assign_file(req, def, issue_flags)))
1882 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1883 creds = override_creds(req->creds);
1885 if (!def->audit_skip)
1886 audit_uring_entry(req->opcode);
1888 ret = def->issue(req, issue_flags);
1890 if (!def->audit_skip)
1891 audit_uring_exit(!ret, ret);
1894 revert_creds(creds);
1896 if (ret == IOU_OK) {
1897 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1898 io_req_complete_defer(req);
1900 io_req_complete_post(req, issue_flags);
1901 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1904 /* If the op doesn't have a file, we're not polling for it */
1905 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
1906 io_iopoll_req_issued(req, issue_flags);
1911 int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts)
1913 io_tw_lock(req->ctx, ts);
1914 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT|
1915 IO_URING_F_COMPLETE_DEFER);
1918 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1920 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1921 struct io_kiocb *nxt = NULL;
1923 if (req_ref_put_and_test(req)) {
1924 if (req->flags & IO_REQ_LINK_FLAGS)
1925 nxt = io_req_find_next(req);
1928 return nxt ? &nxt->work : NULL;
1931 void io_wq_submit_work(struct io_wq_work *work)
1933 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1934 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1935 unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ;
1936 bool needs_poll = false;
1937 int ret = 0, err = -ECANCELED;
1939 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1940 if (!(req->flags & REQ_F_REFCOUNT))
1941 __io_req_set_refcount(req, 2);
1945 io_arm_ltimeout(req);
1947 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1948 if (work->flags & IO_WQ_WORK_CANCEL) {
1950 io_req_task_queue_fail(req, err);
1953 if (!io_assign_file(req, def, issue_flags)) {
1955 work->flags |= IO_WQ_WORK_CANCEL;
1959 if (req->flags & REQ_F_FORCE_ASYNC) {
1960 bool opcode_poll = def->pollin || def->pollout;
1962 if (opcode_poll && file_can_poll(req->file)) {
1964 issue_flags |= IO_URING_F_NONBLOCK;
1969 ret = io_issue_sqe(req, issue_flags);
1974 * If REQ_F_NOWAIT is set, then don't wait or retry with
1975 * poll. -EAGAIN is final for that case.
1977 if (req->flags & REQ_F_NOWAIT)
1981 * We can get EAGAIN for iopolled IO even though we're
1982 * forcing a sync submission from here, since we can't
1983 * wait for request slots on the block side.
1986 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1988 if (io_wq_worker_stopped())
1994 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1996 /* aborted or ready, in either case retry blocking */
1998 issue_flags &= ~IO_URING_F_NONBLOCK;
2001 /* avoid locking problems by failing it from a clean context */
2003 io_req_task_queue_fail(req, ret);
2006 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
2007 unsigned int issue_flags)
2009 struct io_ring_ctx *ctx = req->ctx;
2010 struct io_fixed_file *slot;
2011 struct file *file = NULL;
2013 io_ring_submit_lock(ctx, issue_flags);
2015 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
2017 fd = array_index_nospec(fd, ctx->nr_user_files);
2018 slot = io_fixed_file_slot(&ctx->file_table, fd);
2019 file = io_slot_file(slot);
2020 req->flags |= io_slot_flags(slot);
2021 io_req_set_rsrc_node(req, ctx, 0);
2023 io_ring_submit_unlock(ctx, issue_flags);
2027 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
2029 struct file *file = fget(fd);
2031 trace_io_uring_file_get(req, fd);
2033 /* we don't allow fixed io_uring files */
2034 if (file && io_is_uring_fops(file))
2035 io_req_track_inflight(req);
2039 static void io_queue_async(struct io_kiocb *req, int ret)
2040 __must_hold(&req->ctx->uring_lock)
2042 struct io_kiocb *linked_timeout;
2044 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
2045 io_req_defer_failed(req, ret);
2049 linked_timeout = io_prep_linked_timeout(req);
2051 switch (io_arm_poll_handler(req, 0)) {
2052 case IO_APOLL_READY:
2053 io_kbuf_recycle(req, 0);
2054 io_req_task_queue(req);
2056 case IO_APOLL_ABORTED:
2057 io_kbuf_recycle(req, 0);
2058 io_queue_iowq(req, NULL);
2065 io_queue_linked_timeout(linked_timeout);
2068 static inline void io_queue_sqe(struct io_kiocb *req)
2069 __must_hold(&req->ctx->uring_lock)
2073 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
2076 * We async punt it if the file wasn't marked NOWAIT, or if the file
2077 * doesn't support non-blocking read/write attempts
2080 io_arm_ltimeout(req);
2082 io_queue_async(req, ret);
2085 static void io_queue_sqe_fallback(struct io_kiocb *req)
2086 __must_hold(&req->ctx->uring_lock)
2088 if (unlikely(req->flags & REQ_F_FAIL)) {
2090 * We don't submit, fail them all, for that replace hardlinks
2091 * with normal links. Extra REQ_F_LINK is tolerated.
2093 req->flags &= ~REQ_F_HARDLINK;
2094 req->flags |= REQ_F_LINK;
2095 io_req_defer_failed(req, req->cqe.res);
2097 int ret = io_req_prep_async(req);
2099 if (unlikely(ret)) {
2100 io_req_defer_failed(req, ret);
2104 if (unlikely(req->ctx->drain_active))
2107 io_queue_iowq(req, NULL);
2112 * Check SQE restrictions (opcode and flags).
2114 * Returns 'true' if SQE is allowed, 'false' otherwise.
2116 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
2117 struct io_kiocb *req,
2118 unsigned int sqe_flags)
2120 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
2123 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
2124 ctx->restrictions.sqe_flags_required)
2127 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
2128 ctx->restrictions.sqe_flags_required))
2134 static void io_init_req_drain(struct io_kiocb *req)
2136 struct io_ring_ctx *ctx = req->ctx;
2137 struct io_kiocb *head = ctx->submit_state.link.head;
2139 ctx->drain_active = true;
2142 * If we need to drain a request in the middle of a link, drain
2143 * the head request and the next request/link after the current
2144 * link. Considering sequential execution of links,
2145 * REQ_F_IO_DRAIN will be maintained for every request of our
2148 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2149 ctx->drain_next = true;
2153 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
2154 const struct io_uring_sqe *sqe)
2155 __must_hold(&ctx->uring_lock)
2157 const struct io_issue_def *def;
2158 unsigned int sqe_flags;
2162 /* req is partially pre-initialised, see io_preinit_req() */
2163 req->opcode = opcode = READ_ONCE(sqe->opcode);
2164 /* same numerical values with corresponding REQ_F_*, safe to copy */
2165 req->flags = sqe_flags = READ_ONCE(sqe->flags);
2166 req->cqe.user_data = READ_ONCE(sqe->user_data);
2168 req->rsrc_node = NULL;
2169 req->task = current;
2171 if (unlikely(opcode >= IORING_OP_LAST)) {
2175 def = &io_issue_defs[opcode];
2176 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2177 /* enforce forwards compatibility on users */
2178 if (sqe_flags & ~SQE_VALID_FLAGS)
2180 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2181 if (!def->buffer_select)
2183 req->buf_index = READ_ONCE(sqe->buf_group);
2185 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2186 ctx->drain_disabled = true;
2187 if (sqe_flags & IOSQE_IO_DRAIN) {
2188 if (ctx->drain_disabled)
2190 io_init_req_drain(req);
2193 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2194 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2196 /* knock it to the slow queue path, will be drained there */
2197 if (ctx->drain_active)
2198 req->flags |= REQ_F_FORCE_ASYNC;
2199 /* if there is no link, we're at "next" request and need to drain */
2200 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2201 ctx->drain_next = false;
2202 ctx->drain_active = true;
2203 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2207 if (!def->ioprio && sqe->ioprio)
2209 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2212 if (def->needs_file) {
2213 struct io_submit_state *state = &ctx->submit_state;
2215 req->cqe.fd = READ_ONCE(sqe->fd);
2218 * Plug now if we have more than 2 IO left after this, and the
2219 * target is potentially a read/write to block based storage.
2221 if (state->need_plug && def->plug) {
2222 state->plug_started = true;
2223 state->need_plug = false;
2224 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2228 personality = READ_ONCE(sqe->personality);
2232 req->creds = xa_load(&ctx->personalities, personality);
2235 get_cred(req->creds);
2236 ret = security_uring_override_creds(req->creds);
2238 put_cred(req->creds);
2241 req->flags |= REQ_F_CREDS;
2244 return def->prep(req, sqe);
2247 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2248 struct io_kiocb *req, int ret)
2250 struct io_ring_ctx *ctx = req->ctx;
2251 struct io_submit_link *link = &ctx->submit_state.link;
2252 struct io_kiocb *head = link->head;
2254 trace_io_uring_req_failed(sqe, req, ret);
2257 * Avoid breaking links in the middle as it renders links with SQPOLL
2258 * unusable. Instead of failing eagerly, continue assembling the link if
2259 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2260 * should find the flag and handle the rest.
2262 req_fail_link_node(req, ret);
2263 if (head && !(head->flags & REQ_F_FAIL))
2264 req_fail_link_node(head, -ECANCELED);
2266 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2268 link->last->link = req;
2272 io_queue_sqe_fallback(req);
2277 link->last->link = req;
2284 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2285 const struct io_uring_sqe *sqe)
2286 __must_hold(&ctx->uring_lock)
2288 struct io_submit_link *link = &ctx->submit_state.link;
2291 ret = io_init_req(ctx, req, sqe);
2293 return io_submit_fail_init(sqe, req, ret);
2295 trace_io_uring_submit_req(req);
2298 * If we already have a head request, queue this one for async
2299 * submittal once the head completes. If we don't have a head but
2300 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2301 * submitted sync once the chain is complete. If none of those
2302 * conditions are true (normal request), then just queue it.
2304 if (unlikely(link->head)) {
2305 ret = io_req_prep_async(req);
2307 return io_submit_fail_init(sqe, req, ret);
2309 trace_io_uring_link(req, link->head);
2310 link->last->link = req;
2313 if (req->flags & IO_REQ_LINK_FLAGS)
2315 /* last request of the link, flush it */
2318 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2321 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2322 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2323 if (req->flags & IO_REQ_LINK_FLAGS) {
2328 io_queue_sqe_fallback(req);
2338 * Batched submission is done, ensure local IO is flushed out.
2340 static void io_submit_state_end(struct io_ring_ctx *ctx)
2342 struct io_submit_state *state = &ctx->submit_state;
2344 if (unlikely(state->link.head))
2345 io_queue_sqe_fallback(state->link.head);
2346 /* flush only after queuing links as they can generate completions */
2347 io_submit_flush_completions(ctx);
2348 if (state->plug_started)
2349 blk_finish_plug(&state->plug);
2353 * Start submission side cache.
2355 static void io_submit_state_start(struct io_submit_state *state,
2356 unsigned int max_ios)
2358 state->plug_started = false;
2359 state->need_plug = max_ios > 2;
2360 state->submit_nr = max_ios;
2361 /* set only head, no need to init link_last in advance */
2362 state->link.head = NULL;
2365 static void io_commit_sqring(struct io_ring_ctx *ctx)
2367 struct io_rings *rings = ctx->rings;
2370 * Ensure any loads from the SQEs are done at this point,
2371 * since once we write the new head, the application could
2372 * write new data to them.
2374 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2378 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2379 * that is mapped by userspace. This means that care needs to be taken to
2380 * ensure that reads are stable, as we cannot rely on userspace always
2381 * being a good citizen. If members of the sqe are validated and then later
2382 * used, it's important that those reads are done through READ_ONCE() to
2383 * prevent a re-load down the line.
2385 static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe)
2387 unsigned mask = ctx->sq_entries - 1;
2388 unsigned head = ctx->cached_sq_head++ & mask;
2390 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY)) {
2391 head = READ_ONCE(ctx->sq_array[head]);
2392 if (unlikely(head >= ctx->sq_entries)) {
2393 /* drop invalid entries */
2394 spin_lock(&ctx->completion_lock);
2396 spin_unlock(&ctx->completion_lock);
2397 WRITE_ONCE(ctx->rings->sq_dropped,
2398 READ_ONCE(ctx->rings->sq_dropped) + 1);
2404 * The cached sq head (or cq tail) serves two purposes:
2406 * 1) allows us to batch the cost of updating the user visible
2408 * 2) allows the kernel side to track the head on its own, even
2409 * though the application is the one updating it.
2412 /* double index for 128-byte SQEs, twice as long */
2413 if (ctx->flags & IORING_SETUP_SQE128)
2415 *sqe = &ctx->sq_sqes[head];
2419 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2420 __must_hold(&ctx->uring_lock)
2422 unsigned int entries = io_sqring_entries(ctx);
2426 if (unlikely(!entries))
2428 /* make sure SQ entry isn't read before tail */
2429 ret = left = min(nr, entries);
2430 io_get_task_refs(left);
2431 io_submit_state_start(&ctx->submit_state, left);
2434 const struct io_uring_sqe *sqe;
2435 struct io_kiocb *req;
2437 if (unlikely(!io_alloc_req(ctx, &req)))
2439 if (unlikely(!io_get_sqe(ctx, &sqe))) {
2440 io_req_add_to_cache(req, ctx);
2445 * Continue submitting even for sqe failure if the
2446 * ring was setup with IORING_SETUP_SUBMIT_ALL
2448 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2449 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2455 if (unlikely(left)) {
2457 /* try again if it submitted nothing and can't allocate a req */
2458 if (!ret && io_req_cache_empty(ctx))
2460 current->io_uring->cached_refs += left;
2463 io_submit_state_end(ctx);
2464 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2465 io_commit_sqring(ctx);
2469 struct io_wait_queue {
2470 struct wait_queue_entry wq;
2471 struct io_ring_ctx *ctx;
2473 unsigned nr_timeouts;
2477 static inline bool io_has_work(struct io_ring_ctx *ctx)
2479 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
2480 !llist_empty(&ctx->work_llist);
2483 static inline bool io_should_wake(struct io_wait_queue *iowq)
2485 struct io_ring_ctx *ctx = iowq->ctx;
2486 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
2489 * Wake up if we have enough events, or if a timeout occurred since we
2490 * started waiting. For timeouts, we always want to return to userspace,
2491 * regardless of event count.
2493 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2496 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2497 int wake_flags, void *key)
2499 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq);
2502 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2503 * the task, and the next invocation will do it.
2505 if (io_should_wake(iowq) || io_has_work(iowq->ctx))
2506 return autoremove_wake_function(curr, mode, wake_flags, key);
2510 int io_run_task_work_sig(struct io_ring_ctx *ctx)
2512 if (!llist_empty(&ctx->work_llist)) {
2513 __set_current_state(TASK_RUNNING);
2514 if (io_run_local_work(ctx) > 0)
2517 if (io_run_task_work() > 0)
2519 if (task_sigpending(current))
2524 static bool current_pending_io(void)
2526 struct io_uring_task *tctx = current->io_uring;
2530 return percpu_counter_read_positive(&tctx->inflight);
2533 /* when returns >0, the caller should retry */
2534 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2535 struct io_wait_queue *iowq)
2539 if (unlikely(READ_ONCE(ctx->check_cq)))
2541 if (unlikely(!llist_empty(&ctx->work_llist)))
2543 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL)))
2545 if (unlikely(task_sigpending(current)))
2547 if (unlikely(io_should_wake(iowq)))
2551 * Mark us as being in io_wait if we have pending requests, so cpufreq
2552 * can take into account that the task is waiting for IO - turns out
2553 * to be important for low QD IO.
2555 io_wait = current->in_iowait;
2556 if (current_pending_io())
2557 current->in_iowait = 1;
2559 if (iowq->timeout == KTIME_MAX)
2561 else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS))
2563 current->in_iowait = io_wait;
2568 * Wait until events become available, if we don't already have some. The
2569 * application must reap them itself, as they reside on the shared cq ring.
2571 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2572 const sigset_t __user *sig, size_t sigsz,
2573 struct __kernel_timespec __user *uts)
2575 struct io_wait_queue iowq;
2576 struct io_rings *rings = ctx->rings;
2579 if (!io_allowed_run_tw(ctx))
2581 if (!llist_empty(&ctx->work_llist))
2582 io_run_local_work(ctx);
2584 io_cqring_overflow_flush(ctx);
2585 /* if user messes with these they will just get an early return */
2586 if (__io_cqring_events_user(ctx) >= min_events)
2590 #ifdef CONFIG_COMPAT
2591 if (in_compat_syscall())
2592 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2596 ret = set_user_sigmask(sig, sigsz);
2602 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2603 iowq.wq.private = current;
2604 INIT_LIST_HEAD(&iowq.wq.entry);
2606 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2607 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2608 iowq.timeout = KTIME_MAX;
2611 struct timespec64 ts;
2613 if (get_timespec64(&ts, uts))
2615 iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2618 trace_io_uring_cqring_wait(ctx, min_events);
2620 unsigned long check_cq;
2622 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
2623 int nr_wait = (int) iowq.cq_tail - READ_ONCE(ctx->rings->cq.tail);
2625 atomic_set(&ctx->cq_wait_nr, nr_wait);
2626 set_current_state(TASK_INTERRUPTIBLE);
2628 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2629 TASK_INTERRUPTIBLE);
2632 ret = io_cqring_wait_schedule(ctx, &iowq);
2633 __set_current_state(TASK_RUNNING);
2634 atomic_set(&ctx->cq_wait_nr, 0);
2639 * Run task_work after scheduling and before io_should_wake().
2640 * If we got woken because of task_work being processed, run it
2641 * now rather than let the caller do another wait loop.
2644 if (!llist_empty(&ctx->work_llist))
2645 io_run_local_work(ctx);
2647 check_cq = READ_ONCE(ctx->check_cq);
2648 if (unlikely(check_cq)) {
2649 /* let the caller flush overflows, retry */
2650 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2651 io_cqring_do_overflow_flush(ctx);
2652 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) {
2658 if (io_should_wake(&iowq)) {
2665 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
2666 finish_wait(&ctx->cq_wait, &iowq.wq);
2667 restore_saved_sigmask_unless(ret == -EINTR);
2669 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2672 void io_mem_free(void *ptr)
2677 folio_put(virt_to_folio(ptr));
2680 static void io_pages_free(struct page ***pages, int npages)
2682 struct page **page_array;
2688 page_array = *pages;
2692 for (i = 0; i < npages; i++)
2693 unpin_user_page(page_array[i]);
2698 static void *__io_uaddr_map(struct page ***pages, unsigned short *npages,
2699 unsigned long uaddr, size_t size)
2701 struct page **page_array;
2702 unsigned int nr_pages;
2708 if (uaddr & (PAGE_SIZE - 1) || !size)
2709 return ERR_PTR(-EINVAL);
2711 nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
2712 if (nr_pages > USHRT_MAX)
2713 return ERR_PTR(-EINVAL);
2714 page_array = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
2716 return ERR_PTR(-ENOMEM);
2718 ret = pin_user_pages_fast(uaddr, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
2720 if (ret != nr_pages) {
2722 io_pages_free(&page_array, ret > 0 ? ret : 0);
2723 return ret < 0 ? ERR_PTR(ret) : ERR_PTR(-EFAULT);
2726 page_addr = page_address(page_array[0]);
2727 for (i = 0; i < nr_pages; i++) {
2731 * Can't support mapping user allocated ring memory on 32-bit
2732 * archs where it could potentially reside in highmem. Just
2733 * fail those with -EINVAL, just like we did on kernels that
2734 * didn't support this feature.
2736 if (PageHighMem(page_array[i]))
2740 * No support for discontig pages for now, should either be a
2741 * single normal page, or a huge page. Later on we can add
2742 * support for remapping discontig pages, for now we will
2743 * just fail them with EINVAL.
2745 if (page_address(page_array[i]) != page_addr)
2747 page_addr += PAGE_SIZE;
2750 *pages = page_array;
2752 return page_to_virt(page_array[0]);
2755 static void *io_rings_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2758 return __io_uaddr_map(&ctx->ring_pages, &ctx->n_ring_pages, uaddr,
2762 static void *io_sqes_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2765 return __io_uaddr_map(&ctx->sqe_pages, &ctx->n_sqe_pages, uaddr,
2769 static void io_rings_free(struct io_ring_ctx *ctx)
2771 if (!(ctx->flags & IORING_SETUP_NO_MMAP)) {
2772 io_mem_free(ctx->rings);
2773 io_mem_free(ctx->sq_sqes);
2775 ctx->sq_sqes = NULL;
2777 io_pages_free(&ctx->ring_pages, ctx->n_ring_pages);
2778 ctx->n_ring_pages = 0;
2779 io_pages_free(&ctx->sqe_pages, ctx->n_sqe_pages);
2780 ctx->n_sqe_pages = 0;
2784 void *io_mem_alloc(size_t size)
2786 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2789 ret = (void *) __get_free_pages(gfp, get_order(size));
2792 return ERR_PTR(-ENOMEM);
2795 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2796 unsigned int cq_entries, size_t *sq_offset)
2798 struct io_rings *rings;
2799 size_t off, sq_array_size;
2801 off = struct_size(rings, cqes, cq_entries);
2802 if (off == SIZE_MAX)
2804 if (ctx->flags & IORING_SETUP_CQE32) {
2805 if (check_shl_overflow(off, 1, &off))
2810 off = ALIGN(off, SMP_CACHE_BYTES);
2815 if (ctx->flags & IORING_SETUP_NO_SQARRAY) {
2817 *sq_offset = SIZE_MAX;
2824 sq_array_size = array_size(sizeof(u32), sq_entries);
2825 if (sq_array_size == SIZE_MAX)
2828 if (check_add_overflow(off, sq_array_size, &off))
2834 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2835 unsigned int eventfd_async)
2837 struct io_ev_fd *ev_fd;
2838 __s32 __user *fds = arg;
2841 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2842 lockdep_is_held(&ctx->uring_lock));
2846 if (copy_from_user(&fd, fds, sizeof(*fds)))
2849 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2853 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2854 if (IS_ERR(ev_fd->cq_ev_fd)) {
2855 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2860 spin_lock(&ctx->completion_lock);
2861 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2862 spin_unlock(&ctx->completion_lock);
2864 ev_fd->eventfd_async = eventfd_async;
2865 ctx->has_evfd = true;
2866 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2867 atomic_set(&ev_fd->refs, 1);
2868 atomic_set(&ev_fd->ops, 0);
2872 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2874 struct io_ev_fd *ev_fd;
2876 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2877 lockdep_is_held(&ctx->uring_lock));
2879 ctx->has_evfd = false;
2880 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2881 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops))
2882 call_rcu(&ev_fd->rcu, io_eventfd_ops);
2889 static void io_req_caches_free(struct io_ring_ctx *ctx)
2891 struct io_kiocb *req;
2894 mutex_lock(&ctx->uring_lock);
2895 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
2897 while (!io_req_cache_empty(ctx)) {
2898 req = io_extract_req(ctx);
2899 kmem_cache_free(req_cachep, req);
2903 percpu_ref_put_many(&ctx->refs, nr);
2904 mutex_unlock(&ctx->uring_lock);
2907 static void io_rsrc_node_cache_free(struct io_cache_entry *entry)
2909 kfree(container_of(entry, struct io_rsrc_node, cache));
2912 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2914 io_sq_thread_finish(ctx);
2915 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2916 if (WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)))
2919 mutex_lock(&ctx->uring_lock);
2921 __io_sqe_buffers_unregister(ctx);
2923 __io_sqe_files_unregister(ctx);
2924 io_cqring_overflow_kill(ctx);
2925 io_eventfd_unregister(ctx);
2926 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2927 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2928 io_futex_cache_free(ctx);
2929 io_destroy_buffers(ctx);
2930 mutex_unlock(&ctx->uring_lock);
2932 put_cred(ctx->sq_creds);
2933 if (ctx->submitter_task)
2934 put_task_struct(ctx->submitter_task);
2936 /* there are no registered resources left, nobody uses it */
2938 io_rsrc_node_destroy(ctx, ctx->rsrc_node);
2940 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2942 #if defined(CONFIG_UNIX)
2943 if (ctx->ring_sock) {
2944 ctx->ring_sock->file = NULL; /* so that iput() is called */
2945 sock_release(ctx->ring_sock);
2948 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2950 io_alloc_cache_free(&ctx->rsrc_node_cache, io_rsrc_node_cache_free);
2951 if (ctx->mm_account) {
2952 mmdrop(ctx->mm_account);
2953 ctx->mm_account = NULL;
2956 io_kbuf_mmap_list_free(ctx);
2958 percpu_ref_exit(&ctx->refs);
2959 free_uid(ctx->user);
2960 io_req_caches_free(ctx);
2962 io_wq_put_hash(ctx->hash_map);
2963 kfree(ctx->cancel_table.hbs);
2964 kfree(ctx->cancel_table_locked.hbs);
2966 xa_destroy(&ctx->io_bl_xa);
2970 static __cold void io_activate_pollwq_cb(struct callback_head *cb)
2972 struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx,
2975 mutex_lock(&ctx->uring_lock);
2976 ctx->poll_activated = true;
2977 mutex_unlock(&ctx->uring_lock);
2980 * Wake ups for some events between start of polling and activation
2981 * might've been lost due to loose synchronisation.
2983 wake_up_all(&ctx->poll_wq);
2984 percpu_ref_put(&ctx->refs);
2987 static __cold void io_activate_pollwq(struct io_ring_ctx *ctx)
2989 spin_lock(&ctx->completion_lock);
2990 /* already activated or in progress */
2991 if (ctx->poll_activated || ctx->poll_wq_task_work.func)
2993 if (WARN_ON_ONCE(!ctx->task_complete))
2995 if (!ctx->submitter_task)
2998 * with ->submitter_task only the submitter task completes requests, we
2999 * only need to sync with it, which is done by injecting a tw
3001 init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb);
3002 percpu_ref_get(&ctx->refs);
3003 if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL))
3004 percpu_ref_put(&ctx->refs);
3006 spin_unlock(&ctx->completion_lock);
3009 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
3011 struct io_ring_ctx *ctx = file->private_data;
3014 if (unlikely(!ctx->poll_activated))
3015 io_activate_pollwq(ctx);
3017 poll_wait(file, &ctx->poll_wq, wait);
3019 * synchronizes with barrier from wq_has_sleeper call in
3023 if (!io_sqring_full(ctx))
3024 mask |= EPOLLOUT | EPOLLWRNORM;
3027 * Don't flush cqring overflow list here, just do a simple check.
3028 * Otherwise there could possible be ABBA deadlock:
3031 * lock(&ctx->uring_lock);
3033 * lock(&ctx->uring_lock);
3036 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
3037 * pushes them to do the flush.
3040 if (__io_cqring_events_user(ctx) || io_has_work(ctx))
3041 mask |= EPOLLIN | EPOLLRDNORM;
3046 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
3048 const struct cred *creds;
3050 creds = xa_erase(&ctx->personalities, id);
3059 struct io_tctx_exit {
3060 struct callback_head task_work;
3061 struct completion completion;
3062 struct io_ring_ctx *ctx;
3065 static __cold void io_tctx_exit_cb(struct callback_head *cb)
3067 struct io_uring_task *tctx = current->io_uring;
3068 struct io_tctx_exit *work;
3070 work = container_of(cb, struct io_tctx_exit, task_work);
3072 * When @in_cancel, we're in cancellation and it's racy to remove the
3073 * node. It'll be removed by the end of cancellation, just ignore it.
3074 * tctx can be NULL if the queueing of this task_work raced with
3075 * work cancelation off the exec path.
3077 if (tctx && !atomic_read(&tctx->in_cancel))
3078 io_uring_del_tctx_node((unsigned long)work->ctx);
3079 complete(&work->completion);
3082 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
3084 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3086 return req->ctx == data;
3089 static __cold void io_ring_exit_work(struct work_struct *work)
3091 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
3092 unsigned long timeout = jiffies + HZ * 60 * 5;
3093 unsigned long interval = HZ / 20;
3094 struct io_tctx_exit exit;
3095 struct io_tctx_node *node;
3099 * If we're doing polled IO and end up having requests being
3100 * submitted async (out-of-line), then completions can come in while
3101 * we're waiting for refs to drop. We need to reap these manually,
3102 * as nobody else will be looking for them.
3105 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
3106 mutex_lock(&ctx->uring_lock);
3107 io_cqring_overflow_kill(ctx);
3108 mutex_unlock(&ctx->uring_lock);
3111 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3112 io_move_task_work_from_local(ctx);
3114 while (io_uring_try_cancel_requests(ctx, NULL, true))
3118 struct io_sq_data *sqd = ctx->sq_data;
3119 struct task_struct *tsk;
3121 io_sq_thread_park(sqd);
3123 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
3124 io_wq_cancel_cb(tsk->io_uring->io_wq,
3125 io_cancel_ctx_cb, ctx, true);
3126 io_sq_thread_unpark(sqd);
3129 io_req_caches_free(ctx);
3131 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
3132 /* there is little hope left, don't run it too often */
3136 * This is really an uninterruptible wait, as it has to be
3137 * complete. But it's also run from a kworker, which doesn't
3138 * take signals, so it's fine to make it interruptible. This
3139 * avoids scenarios where we knowingly can wait much longer
3140 * on completions, for example if someone does a SIGSTOP on
3141 * a task that needs to finish task_work to make this loop
3142 * complete. That's a synthetic situation that should not
3143 * cause a stuck task backtrace, and hence a potential panic
3144 * on stuck tasks if that is enabled.
3146 } while (!wait_for_completion_interruptible_timeout(&ctx->ref_comp, interval));
3148 init_completion(&exit.completion);
3149 init_task_work(&exit.task_work, io_tctx_exit_cb);
3152 mutex_lock(&ctx->uring_lock);
3153 while (!list_empty(&ctx->tctx_list)) {
3154 WARN_ON_ONCE(time_after(jiffies, timeout));
3156 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
3158 /* don't spin on a single task if cancellation failed */
3159 list_rotate_left(&ctx->tctx_list);
3160 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
3161 if (WARN_ON_ONCE(ret))
3164 mutex_unlock(&ctx->uring_lock);
3166 * See comment above for
3167 * wait_for_completion_interruptible_timeout() on why this
3168 * wait is marked as interruptible.
3170 wait_for_completion_interruptible(&exit.completion);
3171 mutex_lock(&ctx->uring_lock);
3173 mutex_unlock(&ctx->uring_lock);
3174 spin_lock(&ctx->completion_lock);
3175 spin_unlock(&ctx->completion_lock);
3177 /* pairs with RCU read section in io_req_local_work_add() */
3178 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3181 io_ring_ctx_free(ctx);
3184 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
3186 unsigned long index;
3187 struct creds *creds;
3189 mutex_lock(&ctx->uring_lock);
3190 percpu_ref_kill(&ctx->refs);
3191 xa_for_each(&ctx->personalities, index, creds)
3192 io_unregister_personality(ctx, index);
3194 io_poll_remove_all(ctx, NULL, true);
3195 mutex_unlock(&ctx->uring_lock);
3198 * If we failed setting up the ctx, we might not have any rings
3199 * and therefore did not submit any requests
3202 io_kill_timeouts(ctx, NULL, true);
3204 flush_delayed_work(&ctx->fallback_work);
3206 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
3208 * Use system_unbound_wq to avoid spawning tons of event kworkers
3209 * if we're exiting a ton of rings at the same time. It just adds
3210 * noise and overhead, there's no discernable change in runtime
3211 * over using system_wq.
3213 queue_work(system_unbound_wq, &ctx->exit_work);
3216 static int io_uring_release(struct inode *inode, struct file *file)
3218 struct io_ring_ctx *ctx = file->private_data;
3220 file->private_data = NULL;
3221 io_ring_ctx_wait_and_kill(ctx);
3225 struct io_task_cancel {
3226 struct task_struct *task;
3230 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
3232 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3233 struct io_task_cancel *cancel = data;
3235 return io_match_task_safe(req, cancel->task, cancel->all);
3238 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
3239 struct task_struct *task,
3242 struct io_defer_entry *de;
3245 spin_lock(&ctx->completion_lock);
3246 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
3247 if (io_match_task_safe(de->req, task, cancel_all)) {
3248 list_cut_position(&list, &ctx->defer_list, &de->list);
3252 spin_unlock(&ctx->completion_lock);
3253 if (list_empty(&list))
3256 while (!list_empty(&list)) {
3257 de = list_first_entry(&list, struct io_defer_entry, list);
3258 list_del_init(&de->list);
3259 io_req_task_queue_fail(de->req, -ECANCELED);
3265 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
3267 struct io_tctx_node *node;
3268 enum io_wq_cancel cret;
3271 mutex_lock(&ctx->uring_lock);
3272 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3273 struct io_uring_task *tctx = node->task->io_uring;
3276 * io_wq will stay alive while we hold uring_lock, because it's
3277 * killed after ctx nodes, which requires to take the lock.
3279 if (!tctx || !tctx->io_wq)
3281 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
3282 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3284 mutex_unlock(&ctx->uring_lock);
3289 static bool io_uring_try_cancel_uring_cmd(struct io_ring_ctx *ctx,
3290 struct task_struct *task, bool cancel_all)
3292 struct hlist_node *tmp;
3293 struct io_kiocb *req;
3296 lockdep_assert_held(&ctx->uring_lock);
3298 hlist_for_each_entry_safe(req, tmp, &ctx->cancelable_uring_cmd,
3300 struct io_uring_cmd *cmd = io_kiocb_to_cmd(req,
3301 struct io_uring_cmd);
3302 struct file *file = req->file;
3304 if (!cancel_all && req->task != task)
3307 if (cmd->flags & IORING_URING_CMD_CANCELABLE) {
3308 /* ->sqe isn't available if no async data */
3309 if (!req_has_async_data(req))
3311 file->f_op->uring_cmd(cmd, IO_URING_F_CANCEL);
3315 io_submit_flush_completions(ctx);
3320 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
3321 struct task_struct *task,
3324 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
3325 struct io_uring_task *tctx = task ? task->io_uring : NULL;
3326 enum io_wq_cancel cret;
3329 /* set it so io_req_local_work_add() would wake us up */
3330 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
3331 atomic_set(&ctx->cq_wait_nr, 1);
3335 /* failed during ring init, it couldn't have issued any requests */
3340 ret |= io_uring_try_cancel_iowq(ctx);
3341 } else if (tctx && tctx->io_wq) {
3343 * Cancels requests of all rings, not only @ctx, but
3344 * it's fine as the task is in exit/exec.
3346 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
3348 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3351 /* SQPOLL thread does its own polling */
3352 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
3353 (ctx->sq_data && ctx->sq_data->thread == current)) {
3354 while (!wq_list_empty(&ctx->iopoll_list)) {
3355 io_iopoll_try_reap_events(ctx);
3361 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3362 io_allowed_defer_tw_run(ctx))
3363 ret |= io_run_local_work(ctx) > 0;
3364 ret |= io_cancel_defer_files(ctx, task, cancel_all);
3365 mutex_lock(&ctx->uring_lock);
3366 ret |= io_poll_remove_all(ctx, task, cancel_all);
3367 ret |= io_waitid_remove_all(ctx, task, cancel_all);
3368 ret |= io_futex_remove_all(ctx, task, cancel_all);
3369 ret |= io_uring_try_cancel_uring_cmd(ctx, task, cancel_all);
3370 mutex_unlock(&ctx->uring_lock);
3371 ret |= io_kill_timeouts(ctx, task, cancel_all);
3373 ret |= io_run_task_work() > 0;
3377 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
3380 return atomic_read(&tctx->inflight_tracked);
3381 return percpu_counter_sum(&tctx->inflight);
3385 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3386 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3388 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
3390 struct io_uring_task *tctx = current->io_uring;
3391 struct io_ring_ctx *ctx;
3392 struct io_tctx_node *node;
3393 unsigned long index;
3397 WARN_ON_ONCE(sqd && sqd->thread != current);
3399 if (!current->io_uring)
3402 io_wq_exit_start(tctx->io_wq);
3404 atomic_inc(&tctx->in_cancel);
3408 io_uring_drop_tctx_refs(current);
3409 /* read completions before cancelations */
3410 inflight = tctx_inflight(tctx, !cancel_all);
3415 xa_for_each(&tctx->xa, index, node) {
3416 /* sqpoll task will cancel all its requests */
3417 if (node->ctx->sq_data)
3419 loop |= io_uring_try_cancel_requests(node->ctx,
3420 current, cancel_all);
3423 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3424 loop |= io_uring_try_cancel_requests(ctx,
3434 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3436 io_uring_drop_tctx_refs(current);
3437 xa_for_each(&tctx->xa, index, node) {
3438 if (!llist_empty(&node->ctx->work_llist)) {
3439 WARN_ON_ONCE(node->ctx->submitter_task &&
3440 node->ctx->submitter_task != current);
3445 * If we've seen completions, retry without waiting. This
3446 * avoids a race where a completion comes in before we did
3447 * prepare_to_wait().
3449 if (inflight == tctx_inflight(tctx, !cancel_all))
3452 finish_wait(&tctx->wait, &wait);
3455 io_uring_clean_tctx(tctx);
3458 * We shouldn't run task_works after cancel, so just leave
3459 * ->in_cancel set for normal exit.
3461 atomic_dec(&tctx->in_cancel);
3462 /* for exec all current's requests should be gone, kill tctx */
3463 __io_uring_free(current);
3467 void __io_uring_cancel(bool cancel_all)
3469 io_uring_cancel_generic(cancel_all, NULL);
3472 static void *io_uring_validate_mmap_request(struct file *file,
3473 loff_t pgoff, size_t sz)
3475 struct io_ring_ctx *ctx = file->private_data;
3476 loff_t offset = pgoff << PAGE_SHIFT;
3480 switch (offset & IORING_OFF_MMAP_MASK) {
3481 case IORING_OFF_SQ_RING:
3482 case IORING_OFF_CQ_RING:
3483 /* Don't allow mmap if the ring was setup without it */
3484 if (ctx->flags & IORING_SETUP_NO_MMAP)
3485 return ERR_PTR(-EINVAL);
3488 case IORING_OFF_SQES:
3489 /* Don't allow mmap if the ring was setup without it */
3490 if (ctx->flags & IORING_SETUP_NO_MMAP)
3491 return ERR_PTR(-EINVAL);
3494 case IORING_OFF_PBUF_RING: {
3497 bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT;
3499 ptr = io_pbuf_get_address(ctx, bgid);
3502 return ERR_PTR(-EINVAL);
3506 return ERR_PTR(-EINVAL);
3509 page = virt_to_head_page(ptr);
3510 if (sz > page_size(page))
3511 return ERR_PTR(-EINVAL);
3518 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3520 size_t sz = vma->vm_end - vma->vm_start;
3524 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3526 return PTR_ERR(ptr);
3528 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3529 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3532 static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp,
3533 unsigned long addr, unsigned long len,
3534 unsigned long pgoff, unsigned long flags)
3539 * Do not allow to map to user-provided address to avoid breaking the
3540 * aliasing rules. Userspace is not able to guess the offset address of
3541 * kernel kmalloc()ed memory area.
3546 ptr = io_uring_validate_mmap_request(filp, pgoff, len);
3551 * Some architectures have strong cache aliasing requirements.
3552 * For such architectures we need a coherent mapping which aliases
3553 * kernel memory *and* userspace memory. To achieve that:
3554 * - use a NULL file pointer to reference physical memory, and
3555 * - use the kernel virtual address of the shared io_uring context
3556 * (instead of the userspace-provided address, which has to be 0UL
3558 * - use the same pgoff which the get_unmapped_area() uses to
3559 * calculate the page colouring.
3560 * For architectures without such aliasing requirements, the
3561 * architecture will return any suitable mapping because addr is 0.
3564 flags |= MAP_SHARED;
3565 pgoff = 0; /* has been translated to ptr above */
3567 addr = (uintptr_t) ptr;
3568 pgoff = addr >> PAGE_SHIFT;
3572 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
3575 #else /* !CONFIG_MMU */
3577 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3579 return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL;
3582 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3584 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3587 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3588 unsigned long addr, unsigned long len,
3589 unsigned long pgoff, unsigned long flags)
3593 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3595 return PTR_ERR(ptr);
3597 return (unsigned long) ptr;
3600 #endif /* !CONFIG_MMU */
3602 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3604 if (flags & IORING_ENTER_EXT_ARG) {
3605 struct io_uring_getevents_arg arg;
3607 if (argsz != sizeof(arg))
3609 if (copy_from_user(&arg, argp, sizeof(arg)))
3615 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3616 struct __kernel_timespec __user **ts,
3617 const sigset_t __user **sig)
3619 struct io_uring_getevents_arg arg;
3622 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3623 * is just a pointer to the sigset_t.
3625 if (!(flags & IORING_ENTER_EXT_ARG)) {
3626 *sig = (const sigset_t __user *) argp;
3632 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3633 * timespec and sigset_t pointers if good.
3635 if (*argsz != sizeof(arg))
3637 if (copy_from_user(&arg, argp, sizeof(arg)))
3641 *sig = u64_to_user_ptr(arg.sigmask);
3642 *argsz = arg.sigmask_sz;
3643 *ts = u64_to_user_ptr(arg.ts);
3647 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3648 u32, min_complete, u32, flags, const void __user *, argp,
3651 struct io_ring_ctx *ctx;
3655 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3656 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3657 IORING_ENTER_REGISTERED_RING)))
3661 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3662 * need only dereference our task private array to find it.
3664 if (flags & IORING_ENTER_REGISTERED_RING) {
3665 struct io_uring_task *tctx = current->io_uring;
3667 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3669 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3670 file = tctx->registered_rings[fd];
3671 if (unlikely(!file))
3675 if (unlikely(!file))
3678 if (unlikely(!io_is_uring_fops(file)))
3682 ctx = file->private_data;
3684 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3688 * For SQ polling, the thread will do all submissions and completions.
3689 * Just return the requested submit count, and wake the thread if
3693 if (ctx->flags & IORING_SETUP_SQPOLL) {
3694 io_cqring_overflow_flush(ctx);
3696 if (unlikely(ctx->sq_data->thread == NULL)) {
3700 if (flags & IORING_ENTER_SQ_WAKEUP)
3701 wake_up(&ctx->sq_data->wait);
3702 if (flags & IORING_ENTER_SQ_WAIT)
3703 io_sqpoll_wait_sq(ctx);
3706 } else if (to_submit) {
3707 ret = io_uring_add_tctx_node(ctx);
3711 mutex_lock(&ctx->uring_lock);
3712 ret = io_submit_sqes(ctx, to_submit);
3713 if (ret != to_submit) {
3714 mutex_unlock(&ctx->uring_lock);
3717 if (flags & IORING_ENTER_GETEVENTS) {
3718 if (ctx->syscall_iopoll)
3721 * Ignore errors, we'll soon call io_cqring_wait() and
3722 * it should handle ownership problems if any.
3724 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3725 (void)io_run_local_work_locked(ctx);
3727 mutex_unlock(&ctx->uring_lock);
3730 if (flags & IORING_ENTER_GETEVENTS) {
3733 if (ctx->syscall_iopoll) {
3735 * We disallow the app entering submit/complete with
3736 * polling, but we still need to lock the ring to
3737 * prevent racing with polled issue that got punted to
3740 mutex_lock(&ctx->uring_lock);
3742 ret2 = io_validate_ext_arg(flags, argp, argsz);
3743 if (likely(!ret2)) {
3744 min_complete = min(min_complete,
3746 ret2 = io_iopoll_check(ctx, min_complete);
3748 mutex_unlock(&ctx->uring_lock);
3750 const sigset_t __user *sig;
3751 struct __kernel_timespec __user *ts;
3753 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3754 if (likely(!ret2)) {
3755 min_complete = min(min_complete,
3757 ret2 = io_cqring_wait(ctx, min_complete, sig,
3766 * EBADR indicates that one or more CQE were dropped.
3767 * Once the user has been informed we can clear the bit
3768 * as they are obviously ok with those drops.
3770 if (unlikely(ret2 == -EBADR))
3771 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3776 if (!(flags & IORING_ENTER_REGISTERED_RING))
3781 static const struct file_operations io_uring_fops = {
3782 .release = io_uring_release,
3783 .mmap = io_uring_mmap,
3785 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3786 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3788 .get_unmapped_area = io_uring_mmu_get_unmapped_area,
3790 .poll = io_uring_poll,
3791 #ifdef CONFIG_PROC_FS
3792 .show_fdinfo = io_uring_show_fdinfo,
3796 bool io_is_uring_fops(struct file *file)
3798 return file->f_op == &io_uring_fops;
3801 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3802 struct io_uring_params *p)
3804 struct io_rings *rings;
3805 size_t size, sq_array_offset;
3808 /* make sure these are sane, as we already accounted them */
3809 ctx->sq_entries = p->sq_entries;
3810 ctx->cq_entries = p->cq_entries;
3812 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3813 if (size == SIZE_MAX)
3816 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3817 rings = io_mem_alloc(size);
3819 rings = io_rings_map(ctx, p->cq_off.user_addr, size);
3822 return PTR_ERR(rings);
3825 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY))
3826 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3827 rings->sq_ring_mask = p->sq_entries - 1;
3828 rings->cq_ring_mask = p->cq_entries - 1;
3829 rings->sq_ring_entries = p->sq_entries;
3830 rings->cq_ring_entries = p->cq_entries;
3832 if (p->flags & IORING_SETUP_SQE128)
3833 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3835 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3836 if (size == SIZE_MAX) {
3841 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3842 ptr = io_mem_alloc(size);
3844 ptr = io_sqes_map(ctx, p->sq_off.user_addr, size);
3848 return PTR_ERR(ptr);
3855 static int io_uring_install_fd(struct file *file)
3859 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3862 fd_install(fd, file);
3867 * Allocate an anonymous fd, this is what constitutes the application
3868 * visible backing of an io_uring instance. The application mmaps this
3869 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3870 * we have to tie this fd to a socket for file garbage collection purposes.
3872 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3875 #if defined(CONFIG_UNIX)
3878 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3881 return ERR_PTR(ret);
3884 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3885 O_RDWR | O_CLOEXEC, NULL);
3886 #if defined(CONFIG_UNIX)
3888 sock_release(ctx->ring_sock);
3889 ctx->ring_sock = NULL;
3891 ctx->ring_sock->file = file;
3897 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3898 struct io_uring_params __user *params)
3900 struct io_ring_ctx *ctx;
3901 struct io_uring_task *tctx;
3907 if (entries > IORING_MAX_ENTRIES) {
3908 if (!(p->flags & IORING_SETUP_CLAMP))
3910 entries = IORING_MAX_ENTRIES;
3913 if ((p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
3914 && !(p->flags & IORING_SETUP_NO_MMAP))
3918 * Use twice as many entries for the CQ ring. It's possible for the
3919 * application to drive a higher depth than the size of the SQ ring,
3920 * since the sqes are only used at submission time. This allows for
3921 * some flexibility in overcommitting a bit. If the application has
3922 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3923 * of CQ ring entries manually.
3925 p->sq_entries = roundup_pow_of_two(entries);
3926 if (p->flags & IORING_SETUP_CQSIZE) {
3928 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3929 * to a power-of-two, if it isn't already. We do NOT impose
3930 * any cq vs sq ring sizing.
3934 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3935 if (!(p->flags & IORING_SETUP_CLAMP))
3937 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3939 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3940 if (p->cq_entries < p->sq_entries)
3943 p->cq_entries = 2 * p->sq_entries;
3946 ctx = io_ring_ctx_alloc(p);
3950 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3951 !(ctx->flags & IORING_SETUP_IOPOLL) &&
3952 !(ctx->flags & IORING_SETUP_SQPOLL))
3953 ctx->task_complete = true;
3955 if (ctx->task_complete || (ctx->flags & IORING_SETUP_IOPOLL))
3956 ctx->lockless_cq = true;
3959 * lazy poll_wq activation relies on ->task_complete for synchronisation
3960 * purposes, see io_activate_pollwq()
3962 if (!ctx->task_complete)
3963 ctx->poll_activated = true;
3966 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3967 * space applications don't need to do io completion events
3968 * polling again, they can rely on io_sq_thread to do polling
3969 * work, which can reduce cpu usage and uring_lock contention.
3971 if (ctx->flags & IORING_SETUP_IOPOLL &&
3972 !(ctx->flags & IORING_SETUP_SQPOLL))
3973 ctx->syscall_iopoll = 1;
3975 ctx->compat = in_compat_syscall();
3976 if (!ns_capable_noaudit(&init_user_ns, CAP_IPC_LOCK))
3977 ctx->user = get_uid(current_user());
3980 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3981 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3984 if (ctx->flags & IORING_SETUP_SQPOLL) {
3985 /* IPI related flags don't make sense with SQPOLL */
3986 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3987 IORING_SETUP_TASKRUN_FLAG |
3988 IORING_SETUP_DEFER_TASKRUN))
3990 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3991 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3992 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3994 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3995 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3997 ctx->notify_method = TWA_SIGNAL;
4001 * For DEFER_TASKRUN we require the completion task to be the same as the
4002 * submission task. This implies that there is only one submitter, so enforce
4005 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
4006 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
4011 * This is just grabbed for accounting purposes. When a process exits,
4012 * the mm is exited and dropped before the files, hence we need to hang
4013 * on to this mm purely for the purposes of being able to unaccount
4014 * memory (locked/pinned vm). It's not used for anything else.
4016 mmgrab(current->mm);
4017 ctx->mm_account = current->mm;
4019 ret = io_allocate_scq_urings(ctx, p);
4023 ret = io_sq_offload_create(ctx, p);
4027 ret = io_rsrc_init(ctx);
4031 p->sq_off.head = offsetof(struct io_rings, sq.head);
4032 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
4033 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
4034 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
4035 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
4036 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
4037 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY))
4038 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
4039 p->sq_off.resv1 = 0;
4040 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
4041 p->sq_off.user_addr = 0;
4043 p->cq_off.head = offsetof(struct io_rings, cq.head);
4044 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
4045 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
4046 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
4047 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
4048 p->cq_off.cqes = offsetof(struct io_rings, cqes);
4049 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
4050 p->cq_off.resv1 = 0;
4051 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
4052 p->cq_off.user_addr = 0;
4054 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
4055 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
4056 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
4057 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
4058 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
4059 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
4060 IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING;
4062 if (copy_to_user(params, p, sizeof(*p))) {
4067 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
4068 && !(ctx->flags & IORING_SETUP_R_DISABLED))
4069 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
4071 file = io_uring_get_file(ctx);
4073 ret = PTR_ERR(file);
4077 ret = __io_uring_add_tctx_node(ctx);
4080 tctx = current->io_uring;
4083 * Install ring fd as the very last thing, so we don't risk someone
4084 * having closed it before we finish setup
4086 if (p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
4087 ret = io_ring_add_registered_file(tctx, file, 0, IO_RINGFD_REG_MAX);
4089 ret = io_uring_install_fd(file);
4093 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
4096 io_ring_ctx_wait_and_kill(ctx);
4104 * Sets up an aio uring context, and returns the fd. Applications asks for a
4105 * ring size, we return the actual sq/cq ring sizes (among other things) in the
4106 * params structure passed in.
4108 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
4110 struct io_uring_params p;
4113 if (copy_from_user(&p, params, sizeof(p)))
4115 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
4120 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
4121 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
4122 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
4123 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
4124 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
4125 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
4126 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN |
4127 IORING_SETUP_NO_MMAP | IORING_SETUP_REGISTERED_FD_ONLY |
4128 IORING_SETUP_NO_SQARRAY))
4131 return io_uring_create(entries, &p, params);
4134 static inline bool io_uring_allowed(void)
4136 int disabled = READ_ONCE(sysctl_io_uring_disabled);
4137 kgid_t io_uring_group;
4142 if (disabled == 0 || capable(CAP_SYS_ADMIN))
4145 io_uring_group = make_kgid(&init_user_ns, sysctl_io_uring_group);
4146 if (!gid_valid(io_uring_group))
4149 return in_group_p(io_uring_group);
4152 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
4153 struct io_uring_params __user *, params)
4155 if (!io_uring_allowed())
4158 return io_uring_setup(entries, params);
4161 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
4164 struct io_uring_probe *p;
4168 size = struct_size(p, ops, nr_args);
4169 if (size == SIZE_MAX)
4171 p = kzalloc(size, GFP_KERNEL);
4176 if (copy_from_user(p, arg, size))
4179 if (memchr_inv(p, 0, size))
4182 p->last_op = IORING_OP_LAST - 1;
4183 if (nr_args > IORING_OP_LAST)
4184 nr_args = IORING_OP_LAST;
4186 for (i = 0; i < nr_args; i++) {
4188 if (!io_issue_defs[i].not_supported)
4189 p->ops[i].flags = IO_URING_OP_SUPPORTED;
4194 if (copy_to_user(arg, p, size))
4201 static int io_register_personality(struct io_ring_ctx *ctx)
4203 const struct cred *creds;
4207 creds = get_current_cred();
4209 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
4210 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
4218 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
4219 void __user *arg, unsigned int nr_args)
4221 struct io_uring_restriction *res;
4225 /* Restrictions allowed only if rings started disabled */
4226 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
4229 /* We allow only a single restrictions registration */
4230 if (ctx->restrictions.registered)
4233 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
4236 size = array_size(nr_args, sizeof(*res));
4237 if (size == SIZE_MAX)
4240 res = memdup_user(arg, size);
4242 return PTR_ERR(res);
4246 for (i = 0; i < nr_args; i++) {
4247 switch (res[i].opcode) {
4248 case IORING_RESTRICTION_REGISTER_OP:
4249 if (res[i].register_op >= IORING_REGISTER_LAST) {
4254 __set_bit(res[i].register_op,
4255 ctx->restrictions.register_op);
4257 case IORING_RESTRICTION_SQE_OP:
4258 if (res[i].sqe_op >= IORING_OP_LAST) {
4263 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
4265 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
4266 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
4268 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
4269 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
4278 /* Reset all restrictions if an error happened */
4280 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
4282 ctx->restrictions.registered = true;
4288 static int io_register_enable_rings(struct io_ring_ctx *ctx)
4290 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
4293 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task) {
4294 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
4296 * Lazy activation attempts would fail if it was polled before
4297 * submitter_task is set.
4299 if (wq_has_sleeper(&ctx->poll_wq))
4300 io_activate_pollwq(ctx);
4303 if (ctx->restrictions.registered)
4304 ctx->restricted = 1;
4306 ctx->flags &= ~IORING_SETUP_R_DISABLED;
4307 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
4308 wake_up(&ctx->sq_data->wait);
4312 static __cold int __io_register_iowq_aff(struct io_ring_ctx *ctx,
4313 cpumask_var_t new_mask)
4317 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
4318 ret = io_wq_cpu_affinity(current->io_uring, new_mask);
4320 mutex_unlock(&ctx->uring_lock);
4321 ret = io_sqpoll_wq_cpu_affinity(ctx, new_mask);
4322 mutex_lock(&ctx->uring_lock);
4328 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
4329 void __user *arg, unsigned len)
4331 cpumask_var_t new_mask;
4334 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4337 cpumask_clear(new_mask);
4338 if (len > cpumask_size())
4339 len = cpumask_size();
4341 if (in_compat_syscall()) {
4342 ret = compat_get_bitmap(cpumask_bits(new_mask),
4343 (const compat_ulong_t __user *)arg,
4344 len * 8 /* CHAR_BIT */);
4346 ret = copy_from_user(new_mask, arg, len);
4350 free_cpumask_var(new_mask);
4354 ret = __io_register_iowq_aff(ctx, new_mask);
4355 free_cpumask_var(new_mask);
4359 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
4361 return __io_register_iowq_aff(ctx, NULL);
4364 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
4366 __must_hold(&ctx->uring_lock)
4368 struct io_tctx_node *node;
4369 struct io_uring_task *tctx = NULL;
4370 struct io_sq_data *sqd = NULL;
4374 if (copy_from_user(new_count, arg, sizeof(new_count)))
4376 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4377 if (new_count[i] > INT_MAX)
4380 if (ctx->flags & IORING_SETUP_SQPOLL) {
4384 * Observe the correct sqd->lock -> ctx->uring_lock
4385 * ordering. Fine to drop uring_lock here, we hold
4388 refcount_inc(&sqd->refs);
4389 mutex_unlock(&ctx->uring_lock);
4390 mutex_lock(&sqd->lock);
4391 mutex_lock(&ctx->uring_lock);
4393 tctx = sqd->thread->io_uring;
4396 tctx = current->io_uring;
4399 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
4401 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4403 ctx->iowq_limits[i] = new_count[i];
4404 ctx->iowq_limits_set = true;
4406 if (tctx && tctx->io_wq) {
4407 ret = io_wq_max_workers(tctx->io_wq, new_count);
4411 memset(new_count, 0, sizeof(new_count));
4415 mutex_unlock(&sqd->lock);
4416 io_put_sq_data(sqd);
4419 if (copy_to_user(arg, new_count, sizeof(new_count)))
4422 /* that's it for SQPOLL, only the SQPOLL task creates requests */
4426 /* now propagate the restriction to all registered users */
4427 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
4428 struct io_uring_task *tctx = node->task->io_uring;
4430 if (WARN_ON_ONCE(!tctx->io_wq))
4433 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4434 new_count[i] = ctx->iowq_limits[i];
4435 /* ignore errors, it always returns zero anyway */
4436 (void)io_wq_max_workers(tctx->io_wq, new_count);
4441 mutex_unlock(&sqd->lock);
4442 io_put_sq_data(sqd);
4447 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
4448 void __user *arg, unsigned nr_args)
4449 __releases(ctx->uring_lock)
4450 __acquires(ctx->uring_lock)
4455 * We don't quiesce the refs for register anymore and so it can't be
4456 * dying as we're holding a file ref here.
4458 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
4461 if (ctx->submitter_task && ctx->submitter_task != current)
4464 if (ctx->restricted) {
4465 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
4466 if (!test_bit(opcode, ctx->restrictions.register_op))
4471 case IORING_REGISTER_BUFFERS:
4475 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
4477 case IORING_UNREGISTER_BUFFERS:
4481 ret = io_sqe_buffers_unregister(ctx);
4483 case IORING_REGISTER_FILES:
4487 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
4489 case IORING_UNREGISTER_FILES:
4493 ret = io_sqe_files_unregister(ctx);
4495 case IORING_REGISTER_FILES_UPDATE:
4496 ret = io_register_files_update(ctx, arg, nr_args);
4498 case IORING_REGISTER_EVENTFD:
4502 ret = io_eventfd_register(ctx, arg, 0);
4504 case IORING_REGISTER_EVENTFD_ASYNC:
4508 ret = io_eventfd_register(ctx, arg, 1);
4510 case IORING_UNREGISTER_EVENTFD:
4514 ret = io_eventfd_unregister(ctx);
4516 case IORING_REGISTER_PROBE:
4518 if (!arg || nr_args > 256)
4520 ret = io_probe(ctx, arg, nr_args);
4522 case IORING_REGISTER_PERSONALITY:
4526 ret = io_register_personality(ctx);
4528 case IORING_UNREGISTER_PERSONALITY:
4532 ret = io_unregister_personality(ctx, nr_args);
4534 case IORING_REGISTER_ENABLE_RINGS:
4538 ret = io_register_enable_rings(ctx);
4540 case IORING_REGISTER_RESTRICTIONS:
4541 ret = io_register_restrictions(ctx, arg, nr_args);
4543 case IORING_REGISTER_FILES2:
4544 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
4546 case IORING_REGISTER_FILES_UPDATE2:
4547 ret = io_register_rsrc_update(ctx, arg, nr_args,
4550 case IORING_REGISTER_BUFFERS2:
4551 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
4553 case IORING_REGISTER_BUFFERS_UPDATE:
4554 ret = io_register_rsrc_update(ctx, arg, nr_args,
4555 IORING_RSRC_BUFFER);
4557 case IORING_REGISTER_IOWQ_AFF:
4559 if (!arg || !nr_args)
4561 ret = io_register_iowq_aff(ctx, arg, nr_args);
4563 case IORING_UNREGISTER_IOWQ_AFF:
4567 ret = io_unregister_iowq_aff(ctx);
4569 case IORING_REGISTER_IOWQ_MAX_WORKERS:
4571 if (!arg || nr_args != 2)
4573 ret = io_register_iowq_max_workers(ctx, arg);
4575 case IORING_REGISTER_RING_FDS:
4576 ret = io_ringfd_register(ctx, arg, nr_args);
4578 case IORING_UNREGISTER_RING_FDS:
4579 ret = io_ringfd_unregister(ctx, arg, nr_args);
4581 case IORING_REGISTER_PBUF_RING:
4583 if (!arg || nr_args != 1)
4585 ret = io_register_pbuf_ring(ctx, arg);
4587 case IORING_UNREGISTER_PBUF_RING:
4589 if (!arg || nr_args != 1)
4591 ret = io_unregister_pbuf_ring(ctx, arg);
4593 case IORING_REGISTER_SYNC_CANCEL:
4595 if (!arg || nr_args != 1)
4597 ret = io_sync_cancel(ctx, arg);
4599 case IORING_REGISTER_FILE_ALLOC_RANGE:
4601 if (!arg || nr_args)
4603 ret = io_register_file_alloc_range(ctx, arg);
4613 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
4614 void __user *, arg, unsigned int, nr_args)
4616 struct io_ring_ctx *ctx;
4619 bool use_registered_ring;
4621 use_registered_ring = !!(opcode & IORING_REGISTER_USE_REGISTERED_RING);
4622 opcode &= ~IORING_REGISTER_USE_REGISTERED_RING;
4624 if (opcode >= IORING_REGISTER_LAST)
4627 if (use_registered_ring) {
4629 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
4630 * need only dereference our task private array to find it.
4632 struct io_uring_task *tctx = current->io_uring;
4634 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
4636 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
4637 file = tctx->registered_rings[fd];
4638 if (unlikely(!file))
4642 if (unlikely(!file))
4645 if (!io_is_uring_fops(file))
4649 ctx = file->private_data;
4651 mutex_lock(&ctx->uring_lock);
4652 ret = __io_uring_register(ctx, opcode, arg, nr_args);
4653 mutex_unlock(&ctx->uring_lock);
4654 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
4656 if (!use_registered_ring)
4661 static int __init io_uring_init(void)
4663 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4664 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4665 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4668 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4669 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4670 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4671 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4672 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4673 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
4674 BUILD_BUG_SQE_ELEM(1, __u8, flags);
4675 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
4676 BUILD_BUG_SQE_ELEM(4, __s32, fd);
4677 BUILD_BUG_SQE_ELEM(8, __u64, off);
4678 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
4679 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
4680 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
4681 BUILD_BUG_SQE_ELEM(16, __u64, addr);
4682 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
4683 BUILD_BUG_SQE_ELEM(24, __u32, len);
4684 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
4685 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
4686 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4687 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
4688 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
4689 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
4690 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
4691 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
4692 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
4693 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
4694 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
4695 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
4696 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
4697 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
4698 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
4699 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
4700 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
4701 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
4702 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
4703 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
4704 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
4705 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
4706 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
4707 BUILD_BUG_SQE_ELEM(42, __u16, personality);
4708 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
4709 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
4710 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
4711 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
4712 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
4713 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4714 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
4716 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4717 sizeof(struct io_uring_rsrc_update));
4718 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4719 sizeof(struct io_uring_rsrc_update2));
4721 /* ->buf_index is u16 */
4722 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4723 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4724 offsetof(struct io_uring_buf_ring, tail));
4726 /* should fit into one byte */
4727 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
4728 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4729 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4731 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
4733 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4735 /* top 8bits are for internal use */
4736 BUILD_BUG_ON((IORING_URING_CMD_MASK & 0xff000000) != 0);
4738 io_uring_optable_init();
4741 * Allow user copy in the per-command field, which starts after the
4742 * file in io_kiocb and until the opcode field. The openat2 handling
4743 * requires copying in user memory into the io_kiocb object in that
4744 * range, and HARDENED_USERCOPY will complain if we haven't
4745 * correctly annotated this range.
4747 req_cachep = kmem_cache_create_usercopy("io_kiocb",
4748 sizeof(struct io_kiocb), 0,
4749 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4750 SLAB_ACCOUNT | SLAB_TYPESAFE_BY_RCU,
4751 offsetof(struct io_kiocb, cmd.data),
4752 sizeof_field(struct io_kiocb, cmd.data), NULL);
4753 io_buf_cachep = kmem_cache_create("io_buffer", sizeof(struct io_buffer), 0,
4754 SLAB_HWCACHE_ALIGN | SLAB_PANIC | SLAB_ACCOUNT,
4757 #ifdef CONFIG_SYSCTL
4758 register_sysctl_init("kernel", kernel_io_uring_disabled_table);
4763 __initcall(io_uring_init);