1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <net/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
49 #include <linux/bits.h>
51 #include <linux/sched/signal.h>
53 #include <linux/file.h>
54 #include <linux/fdtable.h>
56 #include <linux/mman.h>
57 #include <linux/percpu.h>
58 #include <linux/slab.h>
59 #include <linux/bvec.h>
60 #include <linux/net.h>
62 #include <net/af_unix.h>
64 #include <linux/anon_inodes.h>
65 #include <linux/sched/mm.h>
66 #include <linux/uaccess.h>
67 #include <linux/nospec.h>
68 #include <linux/highmem.h>
69 #include <linux/fsnotify.h>
70 #include <linux/fadvise.h>
71 #include <linux/task_work.h>
72 #include <linux/io_uring.h>
73 #include <linux/audit.h>
74 #include <linux/security.h>
75 #include <asm/shmparam.h>
77 #define CREATE_TRACE_POINTS
78 #include <trace/events/io_uring.h>
80 #include <uapi/linux/io_uring.h>
99 #include "alloc_cache.h"
101 #define IORING_MAX_ENTRIES 32768
102 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
104 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
105 IORING_REGISTER_LAST + IORING_OP_LAST)
107 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
108 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
110 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
111 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
113 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
114 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
117 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
120 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
122 #define IO_COMPL_BATCH 32
123 #define IO_REQ_ALLOC_BATCH 8
126 IO_CHECK_CQ_OVERFLOW_BIT,
127 IO_CHECK_CQ_DROPPED_BIT,
131 IO_EVENTFD_OP_SIGNAL_BIT,
132 IO_EVENTFD_OP_FREE_BIT,
135 struct io_defer_entry {
136 struct list_head list;
137 struct io_kiocb *req;
141 /* requests with any of those set should undergo io_disarm_next() */
142 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
143 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
145 static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
146 struct task_struct *task,
149 static void io_queue_sqe(struct io_kiocb *req);
150 static void io_move_task_work_from_local(struct io_ring_ctx *ctx);
151 static void __io_submit_flush_completions(struct io_ring_ctx *ctx);
153 struct kmem_cache *req_cachep;
155 struct sock *io_uring_get_socket(struct file *file)
157 #if defined(CONFIG_UNIX)
158 if (io_is_uring_fops(file)) {
159 struct io_ring_ctx *ctx = file->private_data;
161 return ctx->ring_sock->sk;
166 EXPORT_SYMBOL(io_uring_get_socket);
168 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
170 if (!wq_list_empty(&ctx->submit_state.compl_reqs) ||
171 ctx->submit_state.cqes_count)
172 __io_submit_flush_completions(ctx);
175 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
177 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
180 static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx)
182 return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head);
185 static bool io_match_linked(struct io_kiocb *head)
187 struct io_kiocb *req;
189 io_for_each_link(req, head) {
190 if (req->flags & REQ_F_INFLIGHT)
197 * As io_match_task() but protected against racing with linked timeouts.
198 * User must not hold timeout_lock.
200 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
205 if (task && head->task != task)
210 if (head->flags & REQ_F_LINK_TIMEOUT) {
211 struct io_ring_ctx *ctx = head->ctx;
213 /* protect against races with linked timeouts */
214 spin_lock_irq(&ctx->timeout_lock);
215 matched = io_match_linked(head);
216 spin_unlock_irq(&ctx->timeout_lock);
218 matched = io_match_linked(head);
223 static inline void req_fail_link_node(struct io_kiocb *req, int res)
226 io_req_set_res(req, res, 0);
229 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
231 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
232 kasan_poison_object_data(req_cachep, req);
235 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
237 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
239 complete(&ctx->ref_comp);
242 static __cold void io_fallback_req_func(struct work_struct *work)
244 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
246 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
247 struct io_kiocb *req, *tmp;
248 struct io_tw_state ts = { .locked = true, };
250 mutex_lock(&ctx->uring_lock);
251 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
252 req->io_task_work.func(req, &ts);
253 if (WARN_ON_ONCE(!ts.locked))
255 io_submit_flush_completions(ctx);
256 mutex_unlock(&ctx->uring_lock);
259 static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
261 unsigned hash_buckets = 1U << bits;
262 size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
264 table->hbs = kmalloc(hash_size, GFP_KERNEL);
268 table->hash_bits = bits;
269 init_hash_table(table, hash_buckets);
273 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
275 struct io_ring_ctx *ctx;
278 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
282 xa_init(&ctx->io_bl_xa);
285 * Use 5 bits less than the max cq entries, that should give us around
286 * 32 entries per hash list if totally full and uniformly spread, but
287 * don't keep too many buckets to not overconsume memory.
289 hash_bits = ilog2(p->cq_entries) - 5;
290 hash_bits = clamp(hash_bits, 1, 8);
291 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
293 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
296 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
297 if (!ctx->dummy_ubuf)
299 /* set invalid range, so io_import_fixed() fails meeting it */
300 ctx->dummy_ubuf->ubuf = -1UL;
302 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
306 ctx->flags = p->flags;
307 init_waitqueue_head(&ctx->sqo_sq_wait);
308 INIT_LIST_HEAD(&ctx->sqd_list);
309 INIT_LIST_HEAD(&ctx->cq_overflow_list);
310 INIT_LIST_HEAD(&ctx->io_buffers_cache);
311 io_alloc_cache_init(&ctx->rsrc_node_cache, IO_NODE_ALLOC_CACHE_MAX,
312 sizeof(struct io_rsrc_node));
313 io_alloc_cache_init(&ctx->apoll_cache, IO_ALLOC_CACHE_MAX,
314 sizeof(struct async_poll));
315 io_alloc_cache_init(&ctx->netmsg_cache, IO_ALLOC_CACHE_MAX,
316 sizeof(struct io_async_msghdr));
317 init_completion(&ctx->ref_comp);
318 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
319 mutex_init(&ctx->uring_lock);
320 init_waitqueue_head(&ctx->cq_wait);
321 init_waitqueue_head(&ctx->poll_wq);
322 init_waitqueue_head(&ctx->rsrc_quiesce_wq);
323 spin_lock_init(&ctx->completion_lock);
324 spin_lock_init(&ctx->timeout_lock);
325 INIT_WQ_LIST(&ctx->iopoll_list);
326 INIT_LIST_HEAD(&ctx->io_buffers_pages);
327 INIT_LIST_HEAD(&ctx->io_buffers_comp);
328 INIT_LIST_HEAD(&ctx->defer_list);
329 INIT_LIST_HEAD(&ctx->timeout_list);
330 INIT_LIST_HEAD(&ctx->ltimeout_list);
331 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
332 init_llist_head(&ctx->work_llist);
333 INIT_LIST_HEAD(&ctx->tctx_list);
334 ctx->submit_state.free_list.next = NULL;
335 INIT_WQ_LIST(&ctx->locked_free_list);
336 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
337 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
340 kfree(ctx->dummy_ubuf);
341 kfree(ctx->cancel_table.hbs);
342 kfree(ctx->cancel_table_locked.hbs);
344 xa_destroy(&ctx->io_bl_xa);
349 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
351 struct io_rings *r = ctx->rings;
353 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
357 static bool req_need_defer(struct io_kiocb *req, u32 seq)
359 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
360 struct io_ring_ctx *ctx = req->ctx;
362 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
368 static void io_clean_op(struct io_kiocb *req)
370 if (req->flags & REQ_F_BUFFER_SELECTED) {
371 spin_lock(&req->ctx->completion_lock);
372 io_put_kbuf_comp(req);
373 spin_unlock(&req->ctx->completion_lock);
376 if (req->flags & REQ_F_NEED_CLEANUP) {
377 const struct io_cold_def *def = &io_cold_defs[req->opcode];
382 if ((req->flags & REQ_F_POLLED) && req->apoll) {
383 kfree(req->apoll->double_poll);
387 if (req->flags & REQ_F_INFLIGHT) {
388 struct io_uring_task *tctx = req->task->io_uring;
390 atomic_dec(&tctx->inflight_tracked);
392 if (req->flags & REQ_F_CREDS)
393 put_cred(req->creds);
394 if (req->flags & REQ_F_ASYNC_DATA) {
395 kfree(req->async_data);
396 req->async_data = NULL;
398 req->flags &= ~IO_REQ_CLEAN_FLAGS;
401 static inline void io_req_track_inflight(struct io_kiocb *req)
403 if (!(req->flags & REQ_F_INFLIGHT)) {
404 req->flags |= REQ_F_INFLIGHT;
405 atomic_inc(&req->task->io_uring->inflight_tracked);
409 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
411 if (WARN_ON_ONCE(!req->link))
414 req->flags &= ~REQ_F_ARM_LTIMEOUT;
415 req->flags |= REQ_F_LINK_TIMEOUT;
417 /* linked timeouts should have two refs once prep'ed */
418 io_req_set_refcount(req);
419 __io_req_set_refcount(req->link, 2);
423 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
425 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
427 return __io_prep_linked_timeout(req);
430 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
432 io_queue_linked_timeout(__io_prep_linked_timeout(req));
435 static inline void io_arm_ltimeout(struct io_kiocb *req)
437 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
438 __io_arm_ltimeout(req);
441 static void io_prep_async_work(struct io_kiocb *req)
443 const struct io_issue_def *def = &io_issue_defs[req->opcode];
444 struct io_ring_ctx *ctx = req->ctx;
446 if (!(req->flags & REQ_F_CREDS)) {
447 req->flags |= REQ_F_CREDS;
448 req->creds = get_current_cred();
451 req->work.list.next = NULL;
453 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
454 if (req->flags & REQ_F_FORCE_ASYNC)
455 req->work.flags |= IO_WQ_WORK_CONCURRENT;
457 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
458 req->flags |= io_file_get_flags(req->file);
460 if (req->file && (req->flags & REQ_F_ISREG)) {
461 bool should_hash = def->hash_reg_file;
463 /* don't serialize this request if the fs doesn't need it */
464 if (should_hash && (req->file->f_flags & O_DIRECT) &&
465 (req->file->f_mode & FMODE_DIO_PARALLEL_WRITE))
467 if (should_hash || (ctx->flags & IORING_SETUP_IOPOLL))
468 io_wq_hash_work(&req->work, file_inode(req->file));
469 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
470 if (def->unbound_nonreg_file)
471 req->work.flags |= IO_WQ_WORK_UNBOUND;
475 static void io_prep_async_link(struct io_kiocb *req)
477 struct io_kiocb *cur;
479 if (req->flags & REQ_F_LINK_TIMEOUT) {
480 struct io_ring_ctx *ctx = req->ctx;
482 spin_lock_irq(&ctx->timeout_lock);
483 io_for_each_link(cur, req)
484 io_prep_async_work(cur);
485 spin_unlock_irq(&ctx->timeout_lock);
487 io_for_each_link(cur, req)
488 io_prep_async_work(cur);
492 void io_queue_iowq(struct io_kiocb *req, struct io_tw_state *ts_dont_use)
494 struct io_kiocb *link = io_prep_linked_timeout(req);
495 struct io_uring_task *tctx = req->task->io_uring;
498 BUG_ON(!tctx->io_wq);
500 /* init ->work of the whole link before punting */
501 io_prep_async_link(req);
504 * Not expected to happen, but if we do have a bug where this _can_
505 * happen, catch it here and ensure the request is marked as
506 * canceled. That will make io-wq go through the usual work cancel
507 * procedure rather than attempt to run this request (or create a new
510 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
511 req->work.flags |= IO_WQ_WORK_CANCEL;
513 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
514 io_wq_enqueue(tctx->io_wq, &req->work);
516 io_queue_linked_timeout(link);
519 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
521 while (!list_empty(&ctx->defer_list)) {
522 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
523 struct io_defer_entry, list);
525 if (req_need_defer(de->req, de->seq))
527 list_del_init(&de->list);
528 io_req_task_queue(de->req);
534 static void io_eventfd_ops(struct rcu_head *rcu)
536 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
537 int ops = atomic_xchg(&ev_fd->ops, 0);
539 if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT))
540 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
542 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
543 * ordering in a race but if references are 0 we know we have to free
546 if (atomic_dec_and_test(&ev_fd->refs)) {
547 eventfd_ctx_put(ev_fd->cq_ev_fd);
552 static void io_eventfd_signal(struct io_ring_ctx *ctx)
554 struct io_ev_fd *ev_fd = NULL;
558 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
561 ev_fd = rcu_dereference(ctx->io_ev_fd);
564 * Check again if ev_fd exists incase an io_eventfd_unregister call
565 * completed between the NULL check of ctx->io_ev_fd at the start of
566 * the function and rcu_read_lock.
568 if (unlikely(!ev_fd))
570 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
572 if (ev_fd->eventfd_async && !io_wq_current_is_worker())
575 if (likely(eventfd_signal_allowed())) {
576 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
578 atomic_inc(&ev_fd->refs);
579 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops))
580 call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops);
582 atomic_dec(&ev_fd->refs);
589 static void io_eventfd_flush_signal(struct io_ring_ctx *ctx)
593 spin_lock(&ctx->completion_lock);
596 * Eventfd should only get triggered when at least one event has been
597 * posted. Some applications rely on the eventfd notification count
598 * only changing IFF a new CQE has been added to the CQ ring. There's
599 * no depedency on 1:1 relationship between how many times this
600 * function is called (and hence the eventfd count) and number of CQEs
601 * posted to the CQ ring.
603 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
604 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
605 spin_unlock(&ctx->completion_lock);
609 io_eventfd_signal(ctx);
612 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
614 if (ctx->poll_activated)
615 io_poll_wq_wake(ctx);
616 if (ctx->off_timeout_used)
617 io_flush_timeouts(ctx);
618 if (ctx->drain_active) {
619 spin_lock(&ctx->completion_lock);
620 io_queue_deferred(ctx);
621 spin_unlock(&ctx->completion_lock);
624 io_eventfd_flush_signal(ctx);
627 static inline void __io_cq_lock(struct io_ring_ctx *ctx)
629 if (!ctx->task_complete)
630 spin_lock(&ctx->completion_lock);
633 static inline void io_cq_lock(struct io_ring_ctx *ctx)
634 __acquires(ctx->completion_lock)
636 spin_lock(&ctx->completion_lock);
639 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
641 io_commit_cqring(ctx);
643 if (ctx->task_complete) {
645 * ->task_complete implies that only current might be waiting
646 * for CQEs, and obviously, we currently don't. No one is
647 * waiting, wakeups are futile, skip them.
649 io_commit_cqring_flush(ctx);
651 spin_unlock(&ctx->completion_lock);
652 io_commit_cqring_flush(ctx);
657 static void io_cq_unlock_post(struct io_ring_ctx *ctx)
658 __releases(ctx->completion_lock)
660 io_commit_cqring(ctx);
661 spin_unlock(&ctx->completion_lock);
662 io_commit_cqring_flush(ctx);
666 /* Returns true if there are no backlogged entries after the flush */
667 static void io_cqring_overflow_kill(struct io_ring_ctx *ctx)
669 struct io_overflow_cqe *ocqe;
672 spin_lock(&ctx->completion_lock);
673 list_splice_init(&ctx->cq_overflow_list, &list);
674 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
675 spin_unlock(&ctx->completion_lock);
677 while (!list_empty(&list)) {
678 ocqe = list_first_entry(&list, struct io_overflow_cqe, list);
679 list_del(&ocqe->list);
684 static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx)
686 size_t cqe_size = sizeof(struct io_uring_cqe);
688 if (__io_cqring_events(ctx) == ctx->cq_entries)
691 if (ctx->flags & IORING_SETUP_CQE32)
695 while (!list_empty(&ctx->cq_overflow_list)) {
696 struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true);
697 struct io_overflow_cqe *ocqe;
701 ocqe = list_first_entry(&ctx->cq_overflow_list,
702 struct io_overflow_cqe, list);
703 memcpy(cqe, &ocqe->cqe, cqe_size);
704 list_del(&ocqe->list);
708 if (list_empty(&ctx->cq_overflow_list)) {
709 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
710 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
712 io_cq_unlock_post(ctx);
715 static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx)
717 /* iopoll syncs against uring_lock, not completion_lock */
718 if (ctx->flags & IORING_SETUP_IOPOLL)
719 mutex_lock(&ctx->uring_lock);
720 __io_cqring_overflow_flush(ctx);
721 if (ctx->flags & IORING_SETUP_IOPOLL)
722 mutex_unlock(&ctx->uring_lock);
725 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx)
727 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
728 io_cqring_do_overflow_flush(ctx);
731 /* can be called by any task */
732 static void io_put_task_remote(struct task_struct *task)
734 struct io_uring_task *tctx = task->io_uring;
736 percpu_counter_sub(&tctx->inflight, 1);
737 if (unlikely(atomic_read(&tctx->in_cancel)))
738 wake_up(&tctx->wait);
739 put_task_struct(task);
742 /* used by a task to put its own references */
743 static void io_put_task_local(struct task_struct *task)
745 task->io_uring->cached_refs++;
748 /* must to be called somewhat shortly after putting a request */
749 static inline void io_put_task(struct task_struct *task)
751 if (likely(task == current))
752 io_put_task_local(task);
754 io_put_task_remote(task);
757 void io_task_refs_refill(struct io_uring_task *tctx)
759 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
761 percpu_counter_add(&tctx->inflight, refill);
762 refcount_add(refill, ¤t->usage);
763 tctx->cached_refs += refill;
766 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
768 struct io_uring_task *tctx = task->io_uring;
769 unsigned int refs = tctx->cached_refs;
772 tctx->cached_refs = 0;
773 percpu_counter_sub(&tctx->inflight, refs);
774 put_task_struct_many(task, refs);
778 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
779 s32 res, u32 cflags, u64 extra1, u64 extra2)
781 struct io_overflow_cqe *ocqe;
782 size_t ocq_size = sizeof(struct io_overflow_cqe);
783 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
785 lockdep_assert_held(&ctx->completion_lock);
788 ocq_size += sizeof(struct io_uring_cqe);
790 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
791 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
794 * If we're in ring overflow flush mode, or in task cancel mode,
795 * or cannot allocate an overflow entry, then we need to drop it
798 io_account_cq_overflow(ctx);
799 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
802 if (list_empty(&ctx->cq_overflow_list)) {
803 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
804 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
807 ocqe->cqe.user_data = user_data;
809 ocqe->cqe.flags = cflags;
811 ocqe->cqe.big_cqe[0] = extra1;
812 ocqe->cqe.big_cqe[1] = extra2;
814 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
818 bool io_req_cqe_overflow(struct io_kiocb *req)
820 if (!(req->flags & REQ_F_CQE32_INIT)) {
824 return io_cqring_event_overflow(req->ctx, req->cqe.user_data,
825 req->cqe.res, req->cqe.flags,
826 req->extra1, req->extra2);
830 * writes to the cq entry need to come after reading head; the
831 * control dependency is enough as we're using WRITE_ONCE to
834 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow)
836 struct io_rings *rings = ctx->rings;
837 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
838 unsigned int free, queued, len;
841 * Posting into the CQ when there are pending overflowed CQEs may break
842 * ordering guarantees, which will affect links, F_MORE users and more.
843 * Force overflow the completion.
845 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
848 /* userspace may cheat modifying the tail, be safe and do min */
849 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
850 free = ctx->cq_entries - queued;
851 /* we need a contiguous range, limit based on the current array offset */
852 len = min(free, ctx->cq_entries - off);
856 if (ctx->flags & IORING_SETUP_CQE32) {
861 ctx->cqe_cached = &rings->cqes[off];
862 ctx->cqe_sentinel = ctx->cqe_cached + len;
864 ctx->cached_cq_tail++;
866 if (ctx->flags & IORING_SETUP_CQE32)
868 return &rings->cqes[off];
871 static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
874 struct io_uring_cqe *cqe;
879 * If we can't get a cq entry, userspace overflowed the
880 * submission (by quite a lot). Increment the overflow count in
883 cqe = io_get_cqe(ctx);
885 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
887 WRITE_ONCE(cqe->user_data, user_data);
888 WRITE_ONCE(cqe->res, res);
889 WRITE_ONCE(cqe->flags, cflags);
891 if (ctx->flags & IORING_SETUP_CQE32) {
892 WRITE_ONCE(cqe->big_cqe[0], 0);
893 WRITE_ONCE(cqe->big_cqe[1], 0);
900 static void __io_flush_post_cqes(struct io_ring_ctx *ctx)
901 __must_hold(&ctx->uring_lock)
903 struct io_submit_state *state = &ctx->submit_state;
906 lockdep_assert_held(&ctx->uring_lock);
907 for (i = 0; i < state->cqes_count; i++) {
908 struct io_uring_cqe *cqe = &state->cqes[i];
910 if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) {
911 if (ctx->task_complete) {
912 spin_lock(&ctx->completion_lock);
913 io_cqring_event_overflow(ctx, cqe->user_data,
914 cqe->res, cqe->flags, 0, 0);
915 spin_unlock(&ctx->completion_lock);
917 io_cqring_event_overflow(ctx, cqe->user_data,
918 cqe->res, cqe->flags, 0, 0);
922 state->cqes_count = 0;
925 static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
931 filled = io_fill_cqe_aux(ctx, user_data, res, cflags);
932 if (!filled && allow_overflow)
933 filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
935 io_cq_unlock_post(ctx);
939 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags)
941 return __io_post_aux_cqe(ctx, user_data, res, cflags, true);
944 bool io_aux_cqe(const struct io_kiocb *req, bool defer, s32 res, u32 cflags,
947 struct io_ring_ctx *ctx = req->ctx;
948 u64 user_data = req->cqe.user_data;
949 struct io_uring_cqe *cqe;
952 return __io_post_aux_cqe(ctx, user_data, res, cflags, allow_overflow);
954 lockdep_assert_held(&ctx->uring_lock);
956 if (ctx->submit_state.cqes_count == ARRAY_SIZE(ctx->submit_state.cqes)) {
958 __io_flush_post_cqes(ctx);
959 /* no need to flush - flush is deferred */
960 __io_cq_unlock_post(ctx);
963 /* For defered completions this is not as strict as it is otherwise,
964 * however it's main job is to prevent unbounded posted completions,
965 * and in that it works just as well.
967 if (!allow_overflow && test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
970 cqe = &ctx->submit_state.cqes[ctx->submit_state.cqes_count++];
971 cqe->user_data = user_data;
977 static void __io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
979 struct io_ring_ctx *ctx = req->ctx;
980 struct io_rsrc_node *rsrc_node = NULL;
983 if (!(req->flags & REQ_F_CQE_SKIP))
984 io_fill_cqe_req(ctx, req);
987 * If we're the last reference to this request, add to our locked
990 if (req_ref_put_and_test(req)) {
991 if (req->flags & IO_REQ_LINK_FLAGS) {
992 if (req->flags & IO_DISARM_MASK)
995 io_req_task_queue(req->link);
999 io_put_kbuf_comp(req);
1000 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1002 if (!(req->flags & REQ_F_FIXED_FILE))
1003 io_put_file(req->file);
1005 rsrc_node = req->rsrc_node;
1007 * Selected buffer deallocation in io_clean_op() assumes that
1008 * we don't hold ->completion_lock. Clean them here to avoid
1011 io_put_task_remote(req->task);
1012 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
1013 ctx->locked_free_nr++;
1015 io_cq_unlock_post(ctx);
1018 io_ring_submit_lock(ctx, issue_flags);
1019 io_put_rsrc_node(ctx, rsrc_node);
1020 io_ring_submit_unlock(ctx, issue_flags);
1024 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
1026 if (req->ctx->task_complete && req->ctx->submitter_task != current) {
1027 req->io_task_work.func = io_req_task_complete;
1028 io_req_task_work_add(req);
1029 } else if (!(issue_flags & IO_URING_F_UNLOCKED) ||
1030 !(req->ctx->flags & IORING_SETUP_IOPOLL)) {
1031 __io_req_complete_post(req, issue_flags);
1033 struct io_ring_ctx *ctx = req->ctx;
1035 mutex_lock(&ctx->uring_lock);
1036 __io_req_complete_post(req, issue_flags & ~IO_URING_F_UNLOCKED);
1037 mutex_unlock(&ctx->uring_lock);
1041 void io_req_defer_failed(struct io_kiocb *req, s32 res)
1042 __must_hold(&ctx->uring_lock)
1044 const struct io_cold_def *def = &io_cold_defs[req->opcode];
1046 lockdep_assert_held(&req->ctx->uring_lock);
1049 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
1052 io_req_complete_defer(req);
1056 * Don't initialise the fields below on every allocation, but do that in
1057 * advance and keep them valid across allocations.
1059 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1063 req->async_data = NULL;
1064 /* not necessary, but safer to zero */
1068 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1069 struct io_submit_state *state)
1071 spin_lock(&ctx->completion_lock);
1072 wq_list_splice(&ctx->locked_free_list, &state->free_list);
1073 ctx->locked_free_nr = 0;
1074 spin_unlock(&ctx->completion_lock);
1078 * A request might get retired back into the request caches even before opcode
1079 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1080 * Because of that, io_alloc_req() should be called only under ->uring_lock
1081 * and with extra caution to not get a request that is still worked on.
1083 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
1084 __must_hold(&ctx->uring_lock)
1086 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1087 void *reqs[IO_REQ_ALLOC_BATCH];
1091 * If we have more than a batch's worth of requests in our IRQ side
1092 * locked cache, grab the lock and move them over to our submission
1095 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
1096 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
1097 if (!io_req_cache_empty(ctx))
1101 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
1104 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1105 * retry single alloc to be on the safe side.
1107 if (unlikely(ret <= 0)) {
1108 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1114 percpu_ref_get_many(&ctx->refs, ret);
1115 for (i = 0; i < ret; i++) {
1116 struct io_kiocb *req = reqs[i];
1118 io_preinit_req(req, ctx);
1119 io_req_add_to_cache(req, ctx);
1124 __cold void io_free_req(struct io_kiocb *req)
1126 /* refs were already put, restore them for io_req_task_complete() */
1127 req->flags &= ~REQ_F_REFCOUNT;
1128 /* we only want to free it, don't post CQEs */
1129 req->flags |= REQ_F_CQE_SKIP;
1130 req->io_task_work.func = io_req_task_complete;
1131 io_req_task_work_add(req);
1134 static void __io_req_find_next_prep(struct io_kiocb *req)
1136 struct io_ring_ctx *ctx = req->ctx;
1138 spin_lock(&ctx->completion_lock);
1139 io_disarm_next(req);
1140 spin_unlock(&ctx->completion_lock);
1143 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1145 struct io_kiocb *nxt;
1148 * If LINK is set, we have dependent requests in this chain. If we
1149 * didn't fail this request, queue the first one up, moving any other
1150 * dependencies to the next request. In case of failure, fail the rest
1153 if (unlikely(req->flags & IO_DISARM_MASK))
1154 __io_req_find_next_prep(req);
1160 static void ctx_flush_and_put(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1164 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1165 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1167 io_submit_flush_completions(ctx);
1168 mutex_unlock(&ctx->uring_lock);
1171 percpu_ref_put(&ctx->refs);
1174 static unsigned int handle_tw_list(struct llist_node *node,
1175 struct io_ring_ctx **ctx,
1176 struct io_tw_state *ts,
1177 struct llist_node *last)
1179 unsigned int count = 0;
1181 while (node && node != last) {
1182 struct llist_node *next = node->next;
1183 struct io_kiocb *req = container_of(node, struct io_kiocb,
1186 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1188 if (req->ctx != *ctx) {
1189 ctx_flush_and_put(*ctx, ts);
1191 /* if not contended, grab and improve batching */
1192 ts->locked = mutex_trylock(&(*ctx)->uring_lock);
1193 percpu_ref_get(&(*ctx)->refs);
1195 INDIRECT_CALL_2(req->io_task_work.func,
1196 io_poll_task_func, io_req_rw_complete,
1200 if (unlikely(need_resched())) {
1201 ctx_flush_and_put(*ctx, ts);
1211 * io_llist_xchg - swap all entries in a lock-less list
1212 * @head: the head of lock-less list to delete all entries
1213 * @new: new entry as the head of the list
1215 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1216 * The order of entries returned is from the newest to the oldest added one.
1218 static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1219 struct llist_node *new)
1221 return xchg(&head->first, new);
1225 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1226 * @head: the head of lock-less list to delete all entries
1227 * @old: expected old value of the first entry of the list
1228 * @new: new entry as the head of the list
1230 * perform a cmpxchg on the first entry of the list.
1233 static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1234 struct llist_node *old,
1235 struct llist_node *new)
1237 return cmpxchg(&head->first, old, new);
1240 static __cold void io_fallback_tw(struct io_uring_task *tctx)
1242 struct llist_node *node = llist_del_all(&tctx->task_list);
1243 struct io_kiocb *req;
1246 req = container_of(node, struct io_kiocb, io_task_work.node);
1248 if (llist_add(&req->io_task_work.node,
1249 &req->ctx->fallback_llist))
1250 schedule_delayed_work(&req->ctx->fallback_work, 1);
1254 void tctx_task_work(struct callback_head *cb)
1256 struct io_tw_state ts = {};
1257 struct io_ring_ctx *ctx = NULL;
1258 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1260 struct llist_node fake = {};
1261 struct llist_node *node;
1262 unsigned int loops = 0;
1263 unsigned int count = 0;
1265 if (unlikely(current->flags & PF_EXITING)) {
1266 io_fallback_tw(tctx);
1272 node = io_llist_xchg(&tctx->task_list, &fake);
1273 count += handle_tw_list(node, &ctx, &ts, &fake);
1275 /* skip expensive cmpxchg if there are items in the list */
1276 if (READ_ONCE(tctx->task_list.first) != &fake)
1278 if (ts.locked && !wq_list_empty(&ctx->submit_state.compl_reqs)) {
1279 io_submit_flush_completions(ctx);
1280 if (READ_ONCE(tctx->task_list.first) != &fake)
1283 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1284 } while (node != &fake);
1286 ctx_flush_and_put(ctx, &ts);
1288 /* relaxed read is enough as only the task itself sets ->in_cancel */
1289 if (unlikely(atomic_read(&tctx->in_cancel)))
1290 io_uring_drop_tctx_refs(current);
1292 trace_io_uring_task_work_run(tctx, count, loops);
1295 static inline void io_req_local_work_add(struct io_kiocb *req, unsigned flags)
1297 struct io_ring_ctx *ctx = req->ctx;
1298 unsigned nr_wait, nr_tw, nr_tw_prev;
1299 struct llist_node *first;
1301 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK))
1302 flags &= ~IOU_F_TWQ_LAZY_WAKE;
1304 first = READ_ONCE(ctx->work_llist.first);
1308 struct io_kiocb *first_req = container_of(first,
1312 * Might be executed at any moment, rely on
1313 * SLAB_TYPESAFE_BY_RCU to keep it alive.
1315 nr_tw_prev = READ_ONCE(first_req->nr_tw);
1317 nr_tw = nr_tw_prev + 1;
1318 /* Large enough to fail the nr_wait comparison below */
1319 if (!(flags & IOU_F_TWQ_LAZY_WAKE))
1323 req->io_task_work.node.next = first;
1324 } while (!try_cmpxchg(&ctx->work_llist.first, &first,
1325 &req->io_task_work.node));
1328 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1329 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1331 io_eventfd_signal(ctx);
1334 nr_wait = atomic_read(&ctx->cq_wait_nr);
1335 /* no one is waiting */
1338 /* either not enough or the previous add has already woken it up */
1339 if (nr_wait > nr_tw || nr_tw_prev >= nr_wait)
1341 /* pairs with set_current_state() in io_cqring_wait() */
1342 smp_mb__after_atomic();
1343 wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE);
1346 static void io_req_normal_work_add(struct io_kiocb *req)
1348 struct io_uring_task *tctx = req->task->io_uring;
1349 struct io_ring_ctx *ctx = req->ctx;
1351 /* task_work already pending, we're done */
1352 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1355 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1356 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1358 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1361 io_fallback_tw(tctx);
1364 void __io_req_task_work_add(struct io_kiocb *req, unsigned flags)
1366 if (req->ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
1368 io_req_local_work_add(req, flags);
1371 io_req_normal_work_add(req);
1375 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1377 struct llist_node *node;
1379 node = llist_del_all(&ctx->work_llist);
1381 struct io_kiocb *req = container_of(node, struct io_kiocb,
1385 io_req_normal_work_add(req);
1389 static int __io_run_local_work(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1391 struct llist_node *node;
1392 unsigned int loops = 0;
1395 if (WARN_ON_ONCE(ctx->submitter_task != current))
1397 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1398 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1401 * llists are in reverse order, flip it back the right way before
1402 * running the pending items.
1404 node = llist_reverse_order(io_llist_xchg(&ctx->work_llist, NULL));
1406 struct llist_node *next = node->next;
1407 struct io_kiocb *req = container_of(node, struct io_kiocb,
1409 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1410 INDIRECT_CALL_2(req->io_task_work.func,
1411 io_poll_task_func, io_req_rw_complete,
1418 if (!llist_empty(&ctx->work_llist))
1421 io_submit_flush_completions(ctx);
1422 if (!llist_empty(&ctx->work_llist))
1425 trace_io_uring_local_work_run(ctx, ret, loops);
1429 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx)
1431 struct io_tw_state ts = { .locked = true, };
1434 if (llist_empty(&ctx->work_llist))
1437 ret = __io_run_local_work(ctx, &ts);
1438 /* shouldn't happen! */
1439 if (WARN_ON_ONCE(!ts.locked))
1440 mutex_lock(&ctx->uring_lock);
1444 static int io_run_local_work(struct io_ring_ctx *ctx)
1446 struct io_tw_state ts = {};
1449 ts.locked = mutex_trylock(&ctx->uring_lock);
1450 ret = __io_run_local_work(ctx, &ts);
1452 mutex_unlock(&ctx->uring_lock);
1457 static void io_req_task_cancel(struct io_kiocb *req, struct io_tw_state *ts)
1459 io_tw_lock(req->ctx, ts);
1460 io_req_defer_failed(req, req->cqe.res);
1463 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts)
1465 io_tw_lock(req->ctx, ts);
1466 /* req->task == current here, checking PF_EXITING is safe */
1467 if (unlikely(req->task->flags & PF_EXITING))
1468 io_req_defer_failed(req, -EFAULT);
1469 else if (req->flags & REQ_F_FORCE_ASYNC)
1470 io_queue_iowq(req, ts);
1475 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1477 io_req_set_res(req, ret, 0);
1478 req->io_task_work.func = io_req_task_cancel;
1479 io_req_task_work_add(req);
1482 void io_req_task_queue(struct io_kiocb *req)
1484 req->io_task_work.func = io_req_task_submit;
1485 io_req_task_work_add(req);
1488 void io_queue_next(struct io_kiocb *req)
1490 struct io_kiocb *nxt = io_req_find_next(req);
1493 io_req_task_queue(nxt);
1496 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1497 __must_hold(&ctx->uring_lock)
1500 struct io_kiocb *req = container_of(node, struct io_kiocb,
1503 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1504 if (req->flags & REQ_F_REFCOUNT) {
1505 node = req->comp_list.next;
1506 if (!req_ref_put_and_test(req))
1509 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1510 struct async_poll *apoll = req->apoll;
1512 if (apoll->double_poll)
1513 kfree(apoll->double_poll);
1514 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1516 req->flags &= ~REQ_F_POLLED;
1518 if (req->flags & IO_REQ_LINK_FLAGS)
1520 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1523 if (!(req->flags & REQ_F_FIXED_FILE))
1524 io_put_file(req->file);
1526 io_req_put_rsrc_locked(req, ctx);
1528 io_put_task(req->task);
1529 node = req->comp_list.next;
1530 io_req_add_to_cache(req, ctx);
1534 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1535 __must_hold(&ctx->uring_lock)
1537 struct io_submit_state *state = &ctx->submit_state;
1538 struct io_wq_work_node *node;
1541 /* must come first to preserve CQE ordering in failure cases */
1542 if (state->cqes_count)
1543 __io_flush_post_cqes(ctx);
1544 __wq_list_for_each(node, &state->compl_reqs) {
1545 struct io_kiocb *req = container_of(node, struct io_kiocb,
1548 if (!(req->flags & REQ_F_CQE_SKIP) &&
1549 unlikely(!__io_fill_cqe_req(ctx, req))) {
1550 if (ctx->task_complete) {
1551 spin_lock(&ctx->completion_lock);
1552 io_req_cqe_overflow(req);
1553 spin_unlock(&ctx->completion_lock);
1555 io_req_cqe_overflow(req);
1559 __io_cq_unlock_post(ctx);
1561 if (!wq_list_empty(&ctx->submit_state.compl_reqs)) {
1562 io_free_batch_list(ctx, state->compl_reqs.first);
1563 INIT_WQ_LIST(&state->compl_reqs);
1567 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1569 /* See comment at the top of this file */
1571 return __io_cqring_events(ctx);
1575 * We can't just wait for polled events to come to us, we have to actively
1576 * find and complete them.
1578 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1580 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1583 mutex_lock(&ctx->uring_lock);
1584 while (!wq_list_empty(&ctx->iopoll_list)) {
1585 /* let it sleep and repeat later if can't complete a request */
1586 if (io_do_iopoll(ctx, true) == 0)
1589 * Ensure we allow local-to-the-cpu processing to take place,
1590 * in this case we need to ensure that we reap all events.
1591 * Also let task_work, etc. to progress by releasing the mutex
1593 if (need_resched()) {
1594 mutex_unlock(&ctx->uring_lock);
1596 mutex_lock(&ctx->uring_lock);
1599 mutex_unlock(&ctx->uring_lock);
1602 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1604 unsigned int nr_events = 0;
1606 unsigned long check_cq;
1608 if (!io_allowed_run_tw(ctx))
1611 check_cq = READ_ONCE(ctx->check_cq);
1612 if (unlikely(check_cq)) {
1613 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1614 __io_cqring_overflow_flush(ctx);
1616 * Similarly do not spin if we have not informed the user of any
1619 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1623 * Don't enter poll loop if we already have events pending.
1624 * If we do, we can potentially be spinning for commands that
1625 * already triggered a CQE (eg in error).
1627 if (io_cqring_events(ctx))
1632 * If a submit got punted to a workqueue, we can have the
1633 * application entering polling for a command before it gets
1634 * issued. That app will hold the uring_lock for the duration
1635 * of the poll right here, so we need to take a breather every
1636 * now and then to ensure that the issue has a chance to add
1637 * the poll to the issued list. Otherwise we can spin here
1638 * forever, while the workqueue is stuck trying to acquire the
1641 if (wq_list_empty(&ctx->iopoll_list) ||
1642 io_task_work_pending(ctx)) {
1643 u32 tail = ctx->cached_cq_tail;
1645 (void) io_run_local_work_locked(ctx);
1647 if (task_work_pending(current) ||
1648 wq_list_empty(&ctx->iopoll_list)) {
1649 mutex_unlock(&ctx->uring_lock);
1651 mutex_lock(&ctx->uring_lock);
1653 /* some requests don't go through iopoll_list */
1654 if (tail != ctx->cached_cq_tail ||
1655 wq_list_empty(&ctx->iopoll_list))
1658 ret = io_do_iopoll(ctx, !min);
1663 } while (nr_events < min && !need_resched());
1668 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts)
1671 io_req_complete_defer(req);
1673 io_req_complete_post(req, IO_URING_F_UNLOCKED);
1677 * After the iocb has been issued, it's safe to be found on the poll list.
1678 * Adding the kiocb to the list AFTER submission ensures that we don't
1679 * find it from a io_do_iopoll() thread before the issuer is done
1680 * accessing the kiocb cookie.
1682 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1684 struct io_ring_ctx *ctx = req->ctx;
1685 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1687 /* workqueue context doesn't hold uring_lock, grab it now */
1688 if (unlikely(needs_lock))
1689 mutex_lock(&ctx->uring_lock);
1692 * Track whether we have multiple files in our lists. This will impact
1693 * how we do polling eventually, not spinning if we're on potentially
1694 * different devices.
1696 if (wq_list_empty(&ctx->iopoll_list)) {
1697 ctx->poll_multi_queue = false;
1698 } else if (!ctx->poll_multi_queue) {
1699 struct io_kiocb *list_req;
1701 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1703 if (list_req->file != req->file)
1704 ctx->poll_multi_queue = true;
1708 * For fast devices, IO may have already completed. If it has, add
1709 * it to the front so we find it first.
1711 if (READ_ONCE(req->iopoll_completed))
1712 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1714 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1716 if (unlikely(needs_lock)) {
1718 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1719 * in sq thread task context or in io worker task context. If
1720 * current task context is sq thread, we don't need to check
1721 * whether should wake up sq thread.
1723 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1724 wq_has_sleeper(&ctx->sq_data->wait))
1725 wake_up(&ctx->sq_data->wait);
1727 mutex_unlock(&ctx->uring_lock);
1731 unsigned int io_file_get_flags(struct file *file)
1733 unsigned int res = 0;
1735 if (S_ISREG(file_inode(file)->i_mode))
1737 if ((file->f_flags & O_NONBLOCK) || (file->f_mode & FMODE_NOWAIT))
1738 res |= REQ_F_SUPPORT_NOWAIT;
1742 bool io_alloc_async_data(struct io_kiocb *req)
1744 WARN_ON_ONCE(!io_cold_defs[req->opcode].async_size);
1745 req->async_data = kmalloc(io_cold_defs[req->opcode].async_size, GFP_KERNEL);
1746 if (req->async_data) {
1747 req->flags |= REQ_F_ASYNC_DATA;
1753 int io_req_prep_async(struct io_kiocb *req)
1755 const struct io_cold_def *cdef = &io_cold_defs[req->opcode];
1756 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1758 /* assign early for deferred execution for non-fixed file */
1759 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE) && !req->file)
1760 req->file = io_file_get_normal(req, req->cqe.fd);
1761 if (!cdef->prep_async)
1763 if (WARN_ON_ONCE(req_has_async_data(req)))
1765 if (!def->manual_alloc) {
1766 if (io_alloc_async_data(req))
1769 return cdef->prep_async(req);
1772 static u32 io_get_sequence(struct io_kiocb *req)
1774 u32 seq = req->ctx->cached_sq_head;
1775 struct io_kiocb *cur;
1777 /* need original cached_sq_head, but it was increased for each req */
1778 io_for_each_link(cur, req)
1783 static __cold void io_drain_req(struct io_kiocb *req)
1784 __must_hold(&ctx->uring_lock)
1786 struct io_ring_ctx *ctx = req->ctx;
1787 struct io_defer_entry *de;
1789 u32 seq = io_get_sequence(req);
1791 /* Still need defer if there is pending req in defer list. */
1792 spin_lock(&ctx->completion_lock);
1793 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1794 spin_unlock(&ctx->completion_lock);
1796 ctx->drain_active = false;
1797 io_req_task_queue(req);
1800 spin_unlock(&ctx->completion_lock);
1802 io_prep_async_link(req);
1803 de = kmalloc(sizeof(*de), GFP_KERNEL);
1806 io_req_defer_failed(req, ret);
1810 spin_lock(&ctx->completion_lock);
1811 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1812 spin_unlock(&ctx->completion_lock);
1817 trace_io_uring_defer(req);
1820 list_add_tail(&de->list, &ctx->defer_list);
1821 spin_unlock(&ctx->completion_lock);
1824 static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def,
1825 unsigned int issue_flags)
1827 if (req->file || !def->needs_file)
1830 if (req->flags & REQ_F_FIXED_FILE)
1831 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1833 req->file = io_file_get_normal(req, req->cqe.fd);
1838 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1840 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1841 const struct cred *creds = NULL;
1844 if (unlikely(!io_assign_file(req, def, issue_flags)))
1847 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1848 creds = override_creds(req->creds);
1850 if (!def->audit_skip)
1851 audit_uring_entry(req->opcode);
1853 ret = def->issue(req, issue_flags);
1855 if (!def->audit_skip)
1856 audit_uring_exit(!ret, ret);
1859 revert_creds(creds);
1861 if (ret == IOU_OK) {
1862 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1863 io_req_complete_defer(req);
1865 io_req_complete_post(req, issue_flags);
1866 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1869 /* If the op doesn't have a file, we're not polling for it */
1870 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
1871 io_iopoll_req_issued(req, issue_flags);
1876 int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts)
1878 io_tw_lock(req->ctx, ts);
1879 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT|
1880 IO_URING_F_COMPLETE_DEFER);
1883 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1885 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1886 struct io_kiocb *nxt = NULL;
1888 if (req_ref_put_and_test(req)) {
1889 if (req->flags & IO_REQ_LINK_FLAGS)
1890 nxt = io_req_find_next(req);
1893 return nxt ? &nxt->work : NULL;
1896 void io_wq_submit_work(struct io_wq_work *work)
1898 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1899 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1900 unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ;
1901 bool needs_poll = false;
1902 int ret = 0, err = -ECANCELED;
1904 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1905 if (!(req->flags & REQ_F_REFCOUNT))
1906 __io_req_set_refcount(req, 2);
1910 io_arm_ltimeout(req);
1912 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1913 if (work->flags & IO_WQ_WORK_CANCEL) {
1915 io_req_task_queue_fail(req, err);
1918 if (!io_assign_file(req, def, issue_flags)) {
1920 work->flags |= IO_WQ_WORK_CANCEL;
1924 if (req->flags & REQ_F_FORCE_ASYNC) {
1925 bool opcode_poll = def->pollin || def->pollout;
1927 if (opcode_poll && file_can_poll(req->file)) {
1929 issue_flags |= IO_URING_F_NONBLOCK;
1934 ret = io_issue_sqe(req, issue_flags);
1938 * We can get EAGAIN for iopolled IO even though we're
1939 * forcing a sync submission from here, since we can't
1940 * wait for request slots on the block side.
1943 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1949 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1951 /* aborted or ready, in either case retry blocking */
1953 issue_flags &= ~IO_URING_F_NONBLOCK;
1956 /* avoid locking problems by failing it from a clean context */
1958 io_req_task_queue_fail(req, ret);
1961 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1962 unsigned int issue_flags)
1964 struct io_ring_ctx *ctx = req->ctx;
1965 struct io_fixed_file *slot;
1966 struct file *file = NULL;
1968 io_ring_submit_lock(ctx, issue_flags);
1970 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
1972 fd = array_index_nospec(fd, ctx->nr_user_files);
1973 slot = io_fixed_file_slot(&ctx->file_table, fd);
1974 file = io_slot_file(slot);
1975 req->flags |= io_slot_flags(slot);
1976 io_req_set_rsrc_node(req, ctx, 0);
1978 io_ring_submit_unlock(ctx, issue_flags);
1982 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
1984 struct file *file = fget(fd);
1986 trace_io_uring_file_get(req, fd);
1988 /* we don't allow fixed io_uring files */
1989 if (file && io_is_uring_fops(file))
1990 io_req_track_inflight(req);
1994 static void io_queue_async(struct io_kiocb *req, int ret)
1995 __must_hold(&req->ctx->uring_lock)
1997 struct io_kiocb *linked_timeout;
1999 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
2000 io_req_defer_failed(req, ret);
2004 linked_timeout = io_prep_linked_timeout(req);
2006 switch (io_arm_poll_handler(req, 0)) {
2007 case IO_APOLL_READY:
2008 io_kbuf_recycle(req, 0);
2009 io_req_task_queue(req);
2011 case IO_APOLL_ABORTED:
2012 io_kbuf_recycle(req, 0);
2013 io_queue_iowq(req, NULL);
2020 io_queue_linked_timeout(linked_timeout);
2023 static inline void io_queue_sqe(struct io_kiocb *req)
2024 __must_hold(&req->ctx->uring_lock)
2028 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
2031 * We async punt it if the file wasn't marked NOWAIT, or if the file
2032 * doesn't support non-blocking read/write attempts
2035 io_arm_ltimeout(req);
2037 io_queue_async(req, ret);
2040 static void io_queue_sqe_fallback(struct io_kiocb *req)
2041 __must_hold(&req->ctx->uring_lock)
2043 if (unlikely(req->flags & REQ_F_FAIL)) {
2045 * We don't submit, fail them all, for that replace hardlinks
2046 * with normal links. Extra REQ_F_LINK is tolerated.
2048 req->flags &= ~REQ_F_HARDLINK;
2049 req->flags |= REQ_F_LINK;
2050 io_req_defer_failed(req, req->cqe.res);
2052 int ret = io_req_prep_async(req);
2054 if (unlikely(ret)) {
2055 io_req_defer_failed(req, ret);
2059 if (unlikely(req->ctx->drain_active))
2062 io_queue_iowq(req, NULL);
2067 * Check SQE restrictions (opcode and flags).
2069 * Returns 'true' if SQE is allowed, 'false' otherwise.
2071 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
2072 struct io_kiocb *req,
2073 unsigned int sqe_flags)
2075 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
2078 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
2079 ctx->restrictions.sqe_flags_required)
2082 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
2083 ctx->restrictions.sqe_flags_required))
2089 static void io_init_req_drain(struct io_kiocb *req)
2091 struct io_ring_ctx *ctx = req->ctx;
2092 struct io_kiocb *head = ctx->submit_state.link.head;
2094 ctx->drain_active = true;
2097 * If we need to drain a request in the middle of a link, drain
2098 * the head request and the next request/link after the current
2099 * link. Considering sequential execution of links,
2100 * REQ_F_IO_DRAIN will be maintained for every request of our
2103 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2104 ctx->drain_next = true;
2108 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
2109 const struct io_uring_sqe *sqe)
2110 __must_hold(&ctx->uring_lock)
2112 const struct io_issue_def *def;
2113 unsigned int sqe_flags;
2117 /* req is partially pre-initialised, see io_preinit_req() */
2118 req->opcode = opcode = READ_ONCE(sqe->opcode);
2119 /* same numerical values with corresponding REQ_F_*, safe to copy */
2120 req->flags = sqe_flags = READ_ONCE(sqe->flags);
2121 req->cqe.user_data = READ_ONCE(sqe->user_data);
2123 req->rsrc_node = NULL;
2124 req->task = current;
2126 if (unlikely(opcode >= IORING_OP_LAST)) {
2130 def = &io_issue_defs[opcode];
2131 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2132 /* enforce forwards compatibility on users */
2133 if (sqe_flags & ~SQE_VALID_FLAGS)
2135 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2136 if (!def->buffer_select)
2138 req->buf_index = READ_ONCE(sqe->buf_group);
2140 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2141 ctx->drain_disabled = true;
2142 if (sqe_flags & IOSQE_IO_DRAIN) {
2143 if (ctx->drain_disabled)
2145 io_init_req_drain(req);
2148 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2149 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2151 /* knock it to the slow queue path, will be drained there */
2152 if (ctx->drain_active)
2153 req->flags |= REQ_F_FORCE_ASYNC;
2154 /* if there is no link, we're at "next" request and need to drain */
2155 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2156 ctx->drain_next = false;
2157 ctx->drain_active = true;
2158 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2162 if (!def->ioprio && sqe->ioprio)
2164 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2167 if (def->needs_file) {
2168 struct io_submit_state *state = &ctx->submit_state;
2170 req->cqe.fd = READ_ONCE(sqe->fd);
2173 * Plug now if we have more than 2 IO left after this, and the
2174 * target is potentially a read/write to block based storage.
2176 if (state->need_plug && def->plug) {
2177 state->plug_started = true;
2178 state->need_plug = false;
2179 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2183 personality = READ_ONCE(sqe->personality);
2187 req->creds = xa_load(&ctx->personalities, personality);
2190 get_cred(req->creds);
2191 ret = security_uring_override_creds(req->creds);
2193 put_cred(req->creds);
2196 req->flags |= REQ_F_CREDS;
2199 return def->prep(req, sqe);
2202 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2203 struct io_kiocb *req, int ret)
2205 struct io_ring_ctx *ctx = req->ctx;
2206 struct io_submit_link *link = &ctx->submit_state.link;
2207 struct io_kiocb *head = link->head;
2209 trace_io_uring_req_failed(sqe, req, ret);
2212 * Avoid breaking links in the middle as it renders links with SQPOLL
2213 * unusable. Instead of failing eagerly, continue assembling the link if
2214 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2215 * should find the flag and handle the rest.
2217 req_fail_link_node(req, ret);
2218 if (head && !(head->flags & REQ_F_FAIL))
2219 req_fail_link_node(head, -ECANCELED);
2221 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2223 link->last->link = req;
2227 io_queue_sqe_fallback(req);
2232 link->last->link = req;
2239 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2240 const struct io_uring_sqe *sqe)
2241 __must_hold(&ctx->uring_lock)
2243 struct io_submit_link *link = &ctx->submit_state.link;
2246 ret = io_init_req(ctx, req, sqe);
2248 return io_submit_fail_init(sqe, req, ret);
2250 trace_io_uring_submit_req(req);
2253 * If we already have a head request, queue this one for async
2254 * submittal once the head completes. If we don't have a head but
2255 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2256 * submitted sync once the chain is complete. If none of those
2257 * conditions are true (normal request), then just queue it.
2259 if (unlikely(link->head)) {
2260 ret = io_req_prep_async(req);
2262 return io_submit_fail_init(sqe, req, ret);
2264 trace_io_uring_link(req, link->head);
2265 link->last->link = req;
2268 if (req->flags & IO_REQ_LINK_FLAGS)
2270 /* last request of the link, flush it */
2273 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2276 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2277 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2278 if (req->flags & IO_REQ_LINK_FLAGS) {
2283 io_queue_sqe_fallback(req);
2293 * Batched submission is done, ensure local IO is flushed out.
2295 static void io_submit_state_end(struct io_ring_ctx *ctx)
2297 struct io_submit_state *state = &ctx->submit_state;
2299 if (unlikely(state->link.head))
2300 io_queue_sqe_fallback(state->link.head);
2301 /* flush only after queuing links as they can generate completions */
2302 io_submit_flush_completions(ctx);
2303 if (state->plug_started)
2304 blk_finish_plug(&state->plug);
2308 * Start submission side cache.
2310 static void io_submit_state_start(struct io_submit_state *state,
2311 unsigned int max_ios)
2313 state->plug_started = false;
2314 state->need_plug = max_ios > 2;
2315 state->submit_nr = max_ios;
2316 /* set only head, no need to init link_last in advance */
2317 state->link.head = NULL;
2320 static void io_commit_sqring(struct io_ring_ctx *ctx)
2322 struct io_rings *rings = ctx->rings;
2325 * Ensure any loads from the SQEs are done at this point,
2326 * since once we write the new head, the application could
2327 * write new data to them.
2329 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2333 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2334 * that is mapped by userspace. This means that care needs to be taken to
2335 * ensure that reads are stable, as we cannot rely on userspace always
2336 * being a good citizen. If members of the sqe are validated and then later
2337 * used, it's important that those reads are done through READ_ONCE() to
2338 * prevent a re-load down the line.
2340 static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe)
2342 unsigned head, mask = ctx->sq_entries - 1;
2343 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2346 * The cached sq head (or cq tail) serves two purposes:
2348 * 1) allows us to batch the cost of updating the user visible
2350 * 2) allows the kernel side to track the head on its own, even
2351 * though the application is the one updating it.
2353 head = READ_ONCE(ctx->sq_array[sq_idx]);
2354 if (likely(head < ctx->sq_entries)) {
2355 /* double index for 128-byte SQEs, twice as long */
2356 if (ctx->flags & IORING_SETUP_SQE128)
2358 *sqe = &ctx->sq_sqes[head];
2362 /* drop invalid entries */
2364 WRITE_ONCE(ctx->rings->sq_dropped,
2365 READ_ONCE(ctx->rings->sq_dropped) + 1);
2369 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2370 __must_hold(&ctx->uring_lock)
2372 unsigned int entries = io_sqring_entries(ctx);
2376 if (unlikely(!entries))
2378 /* make sure SQ entry isn't read before tail */
2379 ret = left = min(nr, entries);
2380 io_get_task_refs(left);
2381 io_submit_state_start(&ctx->submit_state, left);
2384 const struct io_uring_sqe *sqe;
2385 struct io_kiocb *req;
2387 if (unlikely(!io_alloc_req(ctx, &req)))
2389 if (unlikely(!io_get_sqe(ctx, &sqe))) {
2390 io_req_add_to_cache(req, ctx);
2395 * Continue submitting even for sqe failure if the
2396 * ring was setup with IORING_SETUP_SUBMIT_ALL
2398 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2399 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2405 if (unlikely(left)) {
2407 /* try again if it submitted nothing and can't allocate a req */
2408 if (!ret && io_req_cache_empty(ctx))
2410 current->io_uring->cached_refs += left;
2413 io_submit_state_end(ctx);
2414 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2415 io_commit_sqring(ctx);
2419 struct io_wait_queue {
2420 struct wait_queue_entry wq;
2421 struct io_ring_ctx *ctx;
2423 unsigned nr_timeouts;
2427 static inline bool io_has_work(struct io_ring_ctx *ctx)
2429 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
2430 !llist_empty(&ctx->work_llist);
2433 static inline bool io_should_wake(struct io_wait_queue *iowq)
2435 struct io_ring_ctx *ctx = iowq->ctx;
2436 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
2439 * Wake up if we have enough events, or if a timeout occurred since we
2440 * started waiting. For timeouts, we always want to return to userspace,
2441 * regardless of event count.
2443 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2446 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2447 int wake_flags, void *key)
2449 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq);
2452 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2453 * the task, and the next invocation will do it.
2455 if (io_should_wake(iowq) || io_has_work(iowq->ctx))
2456 return autoremove_wake_function(curr, mode, wake_flags, key);
2460 int io_run_task_work_sig(struct io_ring_ctx *ctx)
2462 if (!llist_empty(&ctx->work_llist)) {
2463 __set_current_state(TASK_RUNNING);
2464 if (io_run_local_work(ctx) > 0)
2467 if (io_run_task_work() > 0)
2469 if (task_sigpending(current))
2474 /* when returns >0, the caller should retry */
2475 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2476 struct io_wait_queue *iowq)
2478 if (unlikely(READ_ONCE(ctx->check_cq)))
2480 if (unlikely(!llist_empty(&ctx->work_llist)))
2482 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL)))
2484 if (unlikely(task_sigpending(current)))
2486 if (unlikely(io_should_wake(iowq)))
2488 if (iowq->timeout == KTIME_MAX)
2490 else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS))
2496 * Wait until events become available, if we don't already have some. The
2497 * application must reap them itself, as they reside on the shared cq ring.
2499 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2500 const sigset_t __user *sig, size_t sigsz,
2501 struct __kernel_timespec __user *uts)
2503 struct io_wait_queue iowq;
2504 struct io_rings *rings = ctx->rings;
2507 if (!io_allowed_run_tw(ctx))
2509 if (!llist_empty(&ctx->work_llist))
2510 io_run_local_work(ctx);
2512 io_cqring_overflow_flush(ctx);
2513 /* if user messes with these they will just get an early return */
2514 if (__io_cqring_events_user(ctx) >= min_events)
2518 #ifdef CONFIG_COMPAT
2519 if (in_compat_syscall())
2520 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2524 ret = set_user_sigmask(sig, sigsz);
2530 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2531 iowq.wq.private = current;
2532 INIT_LIST_HEAD(&iowq.wq.entry);
2534 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2535 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2536 iowq.timeout = KTIME_MAX;
2539 struct timespec64 ts;
2541 if (get_timespec64(&ts, uts))
2543 iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2546 trace_io_uring_cqring_wait(ctx, min_events);
2548 unsigned long check_cq;
2550 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
2551 int nr_wait = (int) iowq.cq_tail - READ_ONCE(ctx->rings->cq.tail);
2553 atomic_set(&ctx->cq_wait_nr, nr_wait);
2554 set_current_state(TASK_INTERRUPTIBLE);
2556 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2557 TASK_INTERRUPTIBLE);
2560 ret = io_cqring_wait_schedule(ctx, &iowq);
2561 __set_current_state(TASK_RUNNING);
2562 atomic_set(&ctx->cq_wait_nr, 0);
2567 * Run task_work after scheduling and before io_should_wake().
2568 * If we got woken because of task_work being processed, run it
2569 * now rather than let the caller do another wait loop.
2572 if (!llist_empty(&ctx->work_llist))
2573 io_run_local_work(ctx);
2575 check_cq = READ_ONCE(ctx->check_cq);
2576 if (unlikely(check_cq)) {
2577 /* let the caller flush overflows, retry */
2578 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2579 io_cqring_do_overflow_flush(ctx);
2580 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) {
2586 if (io_should_wake(&iowq)) {
2593 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
2594 finish_wait(&ctx->cq_wait, &iowq.wq);
2595 restore_saved_sigmask_unless(ret == -EINTR);
2597 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2600 static void io_mem_free(void *ptr)
2607 page = virt_to_head_page(ptr);
2608 if (put_page_testzero(page))
2609 free_compound_page(page);
2612 static void io_pages_free(struct page ***pages, int npages)
2614 struct page **page_array;
2619 page_array = *pages;
2620 for (i = 0; i < npages; i++)
2621 unpin_user_page(page_array[i]);
2626 static void *__io_uaddr_map(struct page ***pages, unsigned short *npages,
2627 unsigned long uaddr, size_t size)
2629 struct page **page_array;
2630 unsigned int nr_pages;
2635 if (uaddr & (PAGE_SIZE - 1) || !size)
2636 return ERR_PTR(-EINVAL);
2638 nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
2639 if (nr_pages > USHRT_MAX)
2640 return ERR_PTR(-EINVAL);
2641 page_array = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
2643 return ERR_PTR(-ENOMEM);
2645 ret = pin_user_pages_fast(uaddr, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
2647 if (ret != nr_pages) {
2649 io_pages_free(&page_array, ret > 0 ? ret : 0);
2650 return ret < 0 ? ERR_PTR(ret) : ERR_PTR(-EFAULT);
2653 * Should be a single page. If the ring is small enough that we can
2654 * use a normal page, that is fine. If we need multiple pages, then
2655 * userspace should use a huge page. That's the only way to guarantee
2656 * that we get contigious memory, outside of just being lucky or
2657 * (currently) having low memory fragmentation.
2659 if (page_array[0] != page_array[ret - 1])
2661 *pages = page_array;
2663 return page_to_virt(page_array[0]);
2666 static void *io_rings_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2669 return __io_uaddr_map(&ctx->ring_pages, &ctx->n_ring_pages, uaddr,
2673 static void *io_sqes_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2676 return __io_uaddr_map(&ctx->sqe_pages, &ctx->n_sqe_pages, uaddr,
2680 static void io_rings_free(struct io_ring_ctx *ctx)
2682 if (!(ctx->flags & IORING_SETUP_NO_MMAP)) {
2683 io_mem_free(ctx->rings);
2684 io_mem_free(ctx->sq_sqes);
2686 ctx->sq_sqes = NULL;
2688 io_pages_free(&ctx->ring_pages, ctx->n_ring_pages);
2689 io_pages_free(&ctx->sqe_pages, ctx->n_sqe_pages);
2693 static void *io_mem_alloc(size_t size)
2695 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2698 ret = (void *) __get_free_pages(gfp, get_order(size));
2701 return ERR_PTR(-ENOMEM);
2704 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2705 unsigned int cq_entries, size_t *sq_offset)
2707 struct io_rings *rings;
2708 size_t off, sq_array_size;
2710 off = struct_size(rings, cqes, cq_entries);
2711 if (off == SIZE_MAX)
2713 if (ctx->flags & IORING_SETUP_CQE32) {
2714 if (check_shl_overflow(off, 1, &off))
2719 off = ALIGN(off, SMP_CACHE_BYTES);
2727 sq_array_size = array_size(sizeof(u32), sq_entries);
2728 if (sq_array_size == SIZE_MAX)
2731 if (check_add_overflow(off, sq_array_size, &off))
2737 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2738 unsigned int eventfd_async)
2740 struct io_ev_fd *ev_fd;
2741 __s32 __user *fds = arg;
2744 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2745 lockdep_is_held(&ctx->uring_lock));
2749 if (copy_from_user(&fd, fds, sizeof(*fds)))
2752 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2756 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2757 if (IS_ERR(ev_fd->cq_ev_fd)) {
2758 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2763 spin_lock(&ctx->completion_lock);
2764 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2765 spin_unlock(&ctx->completion_lock);
2767 ev_fd->eventfd_async = eventfd_async;
2768 ctx->has_evfd = true;
2769 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2770 atomic_set(&ev_fd->refs, 1);
2771 atomic_set(&ev_fd->ops, 0);
2775 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2777 struct io_ev_fd *ev_fd;
2779 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2780 lockdep_is_held(&ctx->uring_lock));
2782 ctx->has_evfd = false;
2783 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2784 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops))
2785 call_rcu(&ev_fd->rcu, io_eventfd_ops);
2792 static void io_req_caches_free(struct io_ring_ctx *ctx)
2794 struct io_kiocb *req;
2797 mutex_lock(&ctx->uring_lock);
2798 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
2800 while (!io_req_cache_empty(ctx)) {
2801 req = io_extract_req(ctx);
2802 kmem_cache_free(req_cachep, req);
2806 percpu_ref_put_many(&ctx->refs, nr);
2807 mutex_unlock(&ctx->uring_lock);
2810 static void io_rsrc_node_cache_free(struct io_cache_entry *entry)
2812 kfree(container_of(entry, struct io_rsrc_node, cache));
2815 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2817 io_sq_thread_finish(ctx);
2818 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2819 if (WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)))
2822 mutex_lock(&ctx->uring_lock);
2824 __io_sqe_buffers_unregister(ctx);
2826 __io_sqe_files_unregister(ctx);
2827 io_cqring_overflow_kill(ctx);
2828 io_eventfd_unregister(ctx);
2829 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2830 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2831 io_destroy_buffers(ctx);
2832 mutex_unlock(&ctx->uring_lock);
2834 put_cred(ctx->sq_creds);
2835 if (ctx->submitter_task)
2836 put_task_struct(ctx->submitter_task);
2838 /* there are no registered resources left, nobody uses it */
2840 io_rsrc_node_destroy(ctx, ctx->rsrc_node);
2842 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2844 #if defined(CONFIG_UNIX)
2845 if (ctx->ring_sock) {
2846 ctx->ring_sock->file = NULL; /* so that iput() is called */
2847 sock_release(ctx->ring_sock);
2850 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2852 io_alloc_cache_free(&ctx->rsrc_node_cache, io_rsrc_node_cache_free);
2853 if (ctx->mm_account) {
2854 mmdrop(ctx->mm_account);
2855 ctx->mm_account = NULL;
2859 percpu_ref_exit(&ctx->refs);
2860 free_uid(ctx->user);
2861 io_req_caches_free(ctx);
2863 io_wq_put_hash(ctx->hash_map);
2864 kfree(ctx->cancel_table.hbs);
2865 kfree(ctx->cancel_table_locked.hbs);
2866 kfree(ctx->dummy_ubuf);
2868 xa_destroy(&ctx->io_bl_xa);
2872 static __cold void io_activate_pollwq_cb(struct callback_head *cb)
2874 struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx,
2877 mutex_lock(&ctx->uring_lock);
2878 ctx->poll_activated = true;
2879 mutex_unlock(&ctx->uring_lock);
2882 * Wake ups for some events between start of polling and activation
2883 * might've been lost due to loose synchronisation.
2885 wake_up_all(&ctx->poll_wq);
2886 percpu_ref_put(&ctx->refs);
2889 static __cold void io_activate_pollwq(struct io_ring_ctx *ctx)
2891 spin_lock(&ctx->completion_lock);
2892 /* already activated or in progress */
2893 if (ctx->poll_activated || ctx->poll_wq_task_work.func)
2895 if (WARN_ON_ONCE(!ctx->task_complete))
2897 if (!ctx->submitter_task)
2900 * with ->submitter_task only the submitter task completes requests, we
2901 * only need to sync with it, which is done by injecting a tw
2903 init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb);
2904 percpu_ref_get(&ctx->refs);
2905 if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL))
2906 percpu_ref_put(&ctx->refs);
2908 spin_unlock(&ctx->completion_lock);
2911 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2913 struct io_ring_ctx *ctx = file->private_data;
2916 if (unlikely(!ctx->poll_activated))
2917 io_activate_pollwq(ctx);
2919 poll_wait(file, &ctx->poll_wq, wait);
2921 * synchronizes with barrier from wq_has_sleeper call in
2925 if (!io_sqring_full(ctx))
2926 mask |= EPOLLOUT | EPOLLWRNORM;
2929 * Don't flush cqring overflow list here, just do a simple check.
2930 * Otherwise there could possible be ABBA deadlock:
2933 * lock(&ctx->uring_lock);
2935 * lock(&ctx->uring_lock);
2938 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2939 * pushes them to do the flush.
2942 if (__io_cqring_events_user(ctx) || io_has_work(ctx))
2943 mask |= EPOLLIN | EPOLLRDNORM;
2948 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2950 const struct cred *creds;
2952 creds = xa_erase(&ctx->personalities, id);
2961 struct io_tctx_exit {
2962 struct callback_head task_work;
2963 struct completion completion;
2964 struct io_ring_ctx *ctx;
2967 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2969 struct io_uring_task *tctx = current->io_uring;
2970 struct io_tctx_exit *work;
2972 work = container_of(cb, struct io_tctx_exit, task_work);
2974 * When @in_cancel, we're in cancellation and it's racy to remove the
2975 * node. It'll be removed by the end of cancellation, just ignore it.
2976 * tctx can be NULL if the queueing of this task_work raced with
2977 * work cancelation off the exec path.
2979 if (tctx && !atomic_read(&tctx->in_cancel))
2980 io_uring_del_tctx_node((unsigned long)work->ctx);
2981 complete(&work->completion);
2984 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2986 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2988 return req->ctx == data;
2991 static __cold void io_ring_exit_work(struct work_struct *work)
2993 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2994 unsigned long timeout = jiffies + HZ * 60 * 5;
2995 unsigned long interval = HZ / 20;
2996 struct io_tctx_exit exit;
2997 struct io_tctx_node *node;
3001 * If we're doing polled IO and end up having requests being
3002 * submitted async (out-of-line), then completions can come in while
3003 * we're waiting for refs to drop. We need to reap these manually,
3004 * as nobody else will be looking for them.
3007 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
3008 mutex_lock(&ctx->uring_lock);
3009 io_cqring_overflow_kill(ctx);
3010 mutex_unlock(&ctx->uring_lock);
3013 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3014 io_move_task_work_from_local(ctx);
3016 while (io_uring_try_cancel_requests(ctx, NULL, true))
3020 struct io_sq_data *sqd = ctx->sq_data;
3021 struct task_struct *tsk;
3023 io_sq_thread_park(sqd);
3025 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
3026 io_wq_cancel_cb(tsk->io_uring->io_wq,
3027 io_cancel_ctx_cb, ctx, true);
3028 io_sq_thread_unpark(sqd);
3031 io_req_caches_free(ctx);
3033 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
3034 /* there is little hope left, don't run it too often */
3038 * This is really an uninterruptible wait, as it has to be
3039 * complete. But it's also run from a kworker, which doesn't
3040 * take signals, so it's fine to make it interruptible. This
3041 * avoids scenarios where we knowingly can wait much longer
3042 * on completions, for example if someone does a SIGSTOP on
3043 * a task that needs to finish task_work to make this loop
3044 * complete. That's a synthetic situation that should not
3045 * cause a stuck task backtrace, and hence a potential panic
3046 * on stuck tasks if that is enabled.
3048 } while (!wait_for_completion_interruptible_timeout(&ctx->ref_comp, interval));
3050 init_completion(&exit.completion);
3051 init_task_work(&exit.task_work, io_tctx_exit_cb);
3054 * Some may use context even when all refs and requests have been put,
3055 * and they are free to do so while still holding uring_lock or
3056 * completion_lock, see io_req_task_submit(). Apart from other work,
3057 * this lock/unlock section also waits them to finish.
3059 mutex_lock(&ctx->uring_lock);
3060 while (!list_empty(&ctx->tctx_list)) {
3061 WARN_ON_ONCE(time_after(jiffies, timeout));
3063 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
3065 /* don't spin on a single task if cancellation failed */
3066 list_rotate_left(&ctx->tctx_list);
3067 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
3068 if (WARN_ON_ONCE(ret))
3071 mutex_unlock(&ctx->uring_lock);
3073 * See comment above for
3074 * wait_for_completion_interruptible_timeout() on why this
3075 * wait is marked as interruptible.
3077 wait_for_completion_interruptible(&exit.completion);
3078 mutex_lock(&ctx->uring_lock);
3080 mutex_unlock(&ctx->uring_lock);
3081 spin_lock(&ctx->completion_lock);
3082 spin_unlock(&ctx->completion_lock);
3084 /* pairs with RCU read section in io_req_local_work_add() */
3085 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3088 io_ring_ctx_free(ctx);
3091 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
3093 unsigned long index;
3094 struct creds *creds;
3096 mutex_lock(&ctx->uring_lock);
3097 percpu_ref_kill(&ctx->refs);
3098 xa_for_each(&ctx->personalities, index, creds)
3099 io_unregister_personality(ctx, index);
3101 io_poll_remove_all(ctx, NULL, true);
3102 mutex_unlock(&ctx->uring_lock);
3105 * If we failed setting up the ctx, we might not have any rings
3106 * and therefore did not submit any requests
3109 io_kill_timeouts(ctx, NULL, true);
3111 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
3113 * Use system_unbound_wq to avoid spawning tons of event kworkers
3114 * if we're exiting a ton of rings at the same time. It just adds
3115 * noise and overhead, there's no discernable change in runtime
3116 * over using system_wq.
3118 queue_work(system_unbound_wq, &ctx->exit_work);
3121 static int io_uring_release(struct inode *inode, struct file *file)
3123 struct io_ring_ctx *ctx = file->private_data;
3125 file->private_data = NULL;
3126 io_ring_ctx_wait_and_kill(ctx);
3130 struct io_task_cancel {
3131 struct task_struct *task;
3135 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
3137 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3138 struct io_task_cancel *cancel = data;
3140 return io_match_task_safe(req, cancel->task, cancel->all);
3143 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
3144 struct task_struct *task,
3147 struct io_defer_entry *de;
3150 spin_lock(&ctx->completion_lock);
3151 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
3152 if (io_match_task_safe(de->req, task, cancel_all)) {
3153 list_cut_position(&list, &ctx->defer_list, &de->list);
3157 spin_unlock(&ctx->completion_lock);
3158 if (list_empty(&list))
3161 while (!list_empty(&list)) {
3162 de = list_first_entry(&list, struct io_defer_entry, list);
3163 list_del_init(&de->list);
3164 io_req_task_queue_fail(de->req, -ECANCELED);
3170 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
3172 struct io_tctx_node *node;
3173 enum io_wq_cancel cret;
3176 mutex_lock(&ctx->uring_lock);
3177 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3178 struct io_uring_task *tctx = node->task->io_uring;
3181 * io_wq will stay alive while we hold uring_lock, because it's
3182 * killed after ctx nodes, which requires to take the lock.
3184 if (!tctx || !tctx->io_wq)
3186 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
3187 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3189 mutex_unlock(&ctx->uring_lock);
3194 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
3195 struct task_struct *task,
3198 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
3199 struct io_uring_task *tctx = task ? task->io_uring : NULL;
3200 enum io_wq_cancel cret;
3203 /* set it so io_req_local_work_add() would wake us up */
3204 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
3205 atomic_set(&ctx->cq_wait_nr, 1);
3209 /* failed during ring init, it couldn't have issued any requests */
3214 ret |= io_uring_try_cancel_iowq(ctx);
3215 } else if (tctx && tctx->io_wq) {
3217 * Cancels requests of all rings, not only @ctx, but
3218 * it's fine as the task is in exit/exec.
3220 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
3222 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3225 /* SQPOLL thread does its own polling */
3226 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
3227 (ctx->sq_data && ctx->sq_data->thread == current)) {
3228 while (!wq_list_empty(&ctx->iopoll_list)) {
3229 io_iopoll_try_reap_events(ctx);
3235 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3236 io_allowed_defer_tw_run(ctx))
3237 ret |= io_run_local_work(ctx) > 0;
3238 ret |= io_cancel_defer_files(ctx, task, cancel_all);
3239 mutex_lock(&ctx->uring_lock);
3240 ret |= io_poll_remove_all(ctx, task, cancel_all);
3241 mutex_unlock(&ctx->uring_lock);
3242 ret |= io_kill_timeouts(ctx, task, cancel_all);
3244 ret |= io_run_task_work() > 0;
3248 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
3251 return atomic_read(&tctx->inflight_tracked);
3252 return percpu_counter_sum(&tctx->inflight);
3256 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3257 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3259 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
3261 struct io_uring_task *tctx = current->io_uring;
3262 struct io_ring_ctx *ctx;
3263 struct io_tctx_node *node;
3264 unsigned long index;
3268 WARN_ON_ONCE(sqd && sqd->thread != current);
3270 if (!current->io_uring)
3273 io_wq_exit_start(tctx->io_wq);
3275 atomic_inc(&tctx->in_cancel);
3279 io_uring_drop_tctx_refs(current);
3280 /* read completions before cancelations */
3281 inflight = tctx_inflight(tctx, !cancel_all);
3286 xa_for_each(&tctx->xa, index, node) {
3287 /* sqpoll task will cancel all its requests */
3288 if (node->ctx->sq_data)
3290 loop |= io_uring_try_cancel_requests(node->ctx,
3291 current, cancel_all);
3294 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3295 loop |= io_uring_try_cancel_requests(ctx,
3305 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3307 io_uring_drop_tctx_refs(current);
3308 xa_for_each(&tctx->xa, index, node) {
3309 if (!llist_empty(&node->ctx->work_llist)) {
3310 WARN_ON_ONCE(node->ctx->submitter_task &&
3311 node->ctx->submitter_task != current);
3316 * If we've seen completions, retry without waiting. This
3317 * avoids a race where a completion comes in before we did
3318 * prepare_to_wait().
3320 if (inflight == tctx_inflight(tctx, !cancel_all))
3323 finish_wait(&tctx->wait, &wait);
3326 io_uring_clean_tctx(tctx);
3329 * We shouldn't run task_works after cancel, so just leave
3330 * ->in_cancel set for normal exit.
3332 atomic_dec(&tctx->in_cancel);
3333 /* for exec all current's requests should be gone, kill tctx */
3334 __io_uring_free(current);
3338 void __io_uring_cancel(bool cancel_all)
3340 io_uring_cancel_generic(cancel_all, NULL);
3343 static void *io_uring_validate_mmap_request(struct file *file,
3344 loff_t pgoff, size_t sz)
3346 struct io_ring_ctx *ctx = file->private_data;
3347 loff_t offset = pgoff << PAGE_SHIFT;
3351 /* Don't allow mmap if the ring was setup without it */
3352 if (ctx->flags & IORING_SETUP_NO_MMAP)
3353 return ERR_PTR(-EINVAL);
3355 switch (offset & IORING_OFF_MMAP_MASK) {
3356 case IORING_OFF_SQ_RING:
3357 case IORING_OFF_CQ_RING:
3360 case IORING_OFF_SQES:
3363 case IORING_OFF_PBUF_RING: {
3366 bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT;
3367 mutex_lock(&ctx->uring_lock);
3368 ptr = io_pbuf_get_address(ctx, bgid);
3369 mutex_unlock(&ctx->uring_lock);
3371 return ERR_PTR(-EINVAL);
3375 return ERR_PTR(-EINVAL);
3378 page = virt_to_head_page(ptr);
3379 if (sz > page_size(page))
3380 return ERR_PTR(-EINVAL);
3387 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3389 size_t sz = vma->vm_end - vma->vm_start;
3393 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3395 return PTR_ERR(ptr);
3397 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3398 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3401 static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp,
3402 unsigned long addr, unsigned long len,
3403 unsigned long pgoff, unsigned long flags)
3405 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
3406 struct vm_unmapped_area_info info;
3410 * Do not allow to map to user-provided address to avoid breaking the
3411 * aliasing rules. Userspace is not able to guess the offset address of
3412 * kernel kmalloc()ed memory area.
3417 ptr = io_uring_validate_mmap_request(filp, pgoff, len);
3421 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
3423 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
3424 info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
3426 info.align_mask = PAGE_MASK & (SHM_COLOUR - 1UL);
3428 info.align_mask = PAGE_MASK & (SHMLBA - 1UL);
3430 info.align_offset = (unsigned long) ptr;
3433 * A failed mmap() very likely causes application failure,
3434 * so fall back to the bottom-up function here. This scenario
3435 * can happen with large stack limits and large mmap()
3438 addr = vm_unmapped_area(&info);
3439 if (offset_in_page(addr)) {
3441 info.low_limit = TASK_UNMAPPED_BASE;
3442 info.high_limit = mmap_end;
3443 addr = vm_unmapped_area(&info);
3449 #else /* !CONFIG_MMU */
3451 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3453 return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL;
3456 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3458 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3461 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3462 unsigned long addr, unsigned long len,
3463 unsigned long pgoff, unsigned long flags)
3467 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3469 return PTR_ERR(ptr);
3471 return (unsigned long) ptr;
3474 #endif /* !CONFIG_MMU */
3476 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3478 if (flags & IORING_ENTER_EXT_ARG) {
3479 struct io_uring_getevents_arg arg;
3481 if (argsz != sizeof(arg))
3483 if (copy_from_user(&arg, argp, sizeof(arg)))
3489 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3490 struct __kernel_timespec __user **ts,
3491 const sigset_t __user **sig)
3493 struct io_uring_getevents_arg arg;
3496 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3497 * is just a pointer to the sigset_t.
3499 if (!(flags & IORING_ENTER_EXT_ARG)) {
3500 *sig = (const sigset_t __user *) argp;
3506 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3507 * timespec and sigset_t pointers if good.
3509 if (*argsz != sizeof(arg))
3511 if (copy_from_user(&arg, argp, sizeof(arg)))
3515 *sig = u64_to_user_ptr(arg.sigmask);
3516 *argsz = arg.sigmask_sz;
3517 *ts = u64_to_user_ptr(arg.ts);
3521 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3522 u32, min_complete, u32, flags, const void __user *, argp,
3525 struct io_ring_ctx *ctx;
3529 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3530 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3531 IORING_ENTER_REGISTERED_RING)))
3535 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3536 * need only dereference our task private array to find it.
3538 if (flags & IORING_ENTER_REGISTERED_RING) {
3539 struct io_uring_task *tctx = current->io_uring;
3541 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3543 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3544 f.file = tctx->registered_rings[fd];
3546 if (unlikely(!f.file))
3550 if (unlikely(!f.file))
3553 if (unlikely(!io_is_uring_fops(f.file)))
3557 ctx = f.file->private_data;
3559 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3563 * For SQ polling, the thread will do all submissions and completions.
3564 * Just return the requested submit count, and wake the thread if
3568 if (ctx->flags & IORING_SETUP_SQPOLL) {
3569 io_cqring_overflow_flush(ctx);
3571 if (unlikely(ctx->sq_data->thread == NULL)) {
3575 if (flags & IORING_ENTER_SQ_WAKEUP)
3576 wake_up(&ctx->sq_data->wait);
3577 if (flags & IORING_ENTER_SQ_WAIT)
3578 io_sqpoll_wait_sq(ctx);
3581 } else if (to_submit) {
3582 ret = io_uring_add_tctx_node(ctx);
3586 mutex_lock(&ctx->uring_lock);
3587 ret = io_submit_sqes(ctx, to_submit);
3588 if (ret != to_submit) {
3589 mutex_unlock(&ctx->uring_lock);
3592 if (flags & IORING_ENTER_GETEVENTS) {
3593 if (ctx->syscall_iopoll)
3596 * Ignore errors, we'll soon call io_cqring_wait() and
3597 * it should handle ownership problems if any.
3599 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3600 (void)io_run_local_work_locked(ctx);
3602 mutex_unlock(&ctx->uring_lock);
3605 if (flags & IORING_ENTER_GETEVENTS) {
3608 if (ctx->syscall_iopoll) {
3610 * We disallow the app entering submit/complete with
3611 * polling, but we still need to lock the ring to
3612 * prevent racing with polled issue that got punted to
3615 mutex_lock(&ctx->uring_lock);
3617 ret2 = io_validate_ext_arg(flags, argp, argsz);
3618 if (likely(!ret2)) {
3619 min_complete = min(min_complete,
3621 ret2 = io_iopoll_check(ctx, min_complete);
3623 mutex_unlock(&ctx->uring_lock);
3625 const sigset_t __user *sig;
3626 struct __kernel_timespec __user *ts;
3628 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3629 if (likely(!ret2)) {
3630 min_complete = min(min_complete,
3632 ret2 = io_cqring_wait(ctx, min_complete, sig,
3641 * EBADR indicates that one or more CQE were dropped.
3642 * Once the user has been informed we can clear the bit
3643 * as they are obviously ok with those drops.
3645 if (unlikely(ret2 == -EBADR))
3646 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3655 static const struct file_operations io_uring_fops = {
3656 .release = io_uring_release,
3657 .mmap = io_uring_mmap,
3659 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3660 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3662 .get_unmapped_area = io_uring_mmu_get_unmapped_area,
3664 .poll = io_uring_poll,
3665 #ifdef CONFIG_PROC_FS
3666 .show_fdinfo = io_uring_show_fdinfo,
3670 bool io_is_uring_fops(struct file *file)
3672 return file->f_op == &io_uring_fops;
3675 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3676 struct io_uring_params *p)
3678 struct io_rings *rings;
3679 size_t size, sq_array_offset;
3682 /* make sure these are sane, as we already accounted them */
3683 ctx->sq_entries = p->sq_entries;
3684 ctx->cq_entries = p->cq_entries;
3686 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3687 if (size == SIZE_MAX)
3690 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3691 rings = io_mem_alloc(size);
3693 rings = io_rings_map(ctx, p->cq_off.user_addr, size);
3696 return PTR_ERR(rings);
3699 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3700 rings->sq_ring_mask = p->sq_entries - 1;
3701 rings->cq_ring_mask = p->cq_entries - 1;
3702 rings->sq_ring_entries = p->sq_entries;
3703 rings->cq_ring_entries = p->cq_entries;
3705 if (p->flags & IORING_SETUP_SQE128)
3706 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3708 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3709 if (size == SIZE_MAX) {
3714 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3715 ptr = io_mem_alloc(size);
3717 ptr = io_sqes_map(ctx, p->sq_off.user_addr, size);
3721 return PTR_ERR(ptr);
3728 static int io_uring_install_fd(struct file *file)
3732 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3735 fd_install(fd, file);
3740 * Allocate an anonymous fd, this is what constitutes the application
3741 * visible backing of an io_uring instance. The application mmaps this
3742 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3743 * we have to tie this fd to a socket for file garbage collection purposes.
3745 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3748 #if defined(CONFIG_UNIX)
3751 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3754 return ERR_PTR(ret);
3757 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3758 O_RDWR | O_CLOEXEC, NULL);
3759 #if defined(CONFIG_UNIX)
3761 sock_release(ctx->ring_sock);
3762 ctx->ring_sock = NULL;
3764 ctx->ring_sock->file = file;
3770 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3771 struct io_uring_params __user *params)
3773 struct io_ring_ctx *ctx;
3774 struct io_uring_task *tctx;
3780 if (entries > IORING_MAX_ENTRIES) {
3781 if (!(p->flags & IORING_SETUP_CLAMP))
3783 entries = IORING_MAX_ENTRIES;
3786 if ((p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
3787 && !(p->flags & IORING_SETUP_NO_MMAP))
3791 * Use twice as many entries for the CQ ring. It's possible for the
3792 * application to drive a higher depth than the size of the SQ ring,
3793 * since the sqes are only used at submission time. This allows for
3794 * some flexibility in overcommitting a bit. If the application has
3795 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3796 * of CQ ring entries manually.
3798 p->sq_entries = roundup_pow_of_two(entries);
3799 if (p->flags & IORING_SETUP_CQSIZE) {
3801 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3802 * to a power-of-two, if it isn't already. We do NOT impose
3803 * any cq vs sq ring sizing.
3807 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3808 if (!(p->flags & IORING_SETUP_CLAMP))
3810 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3812 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3813 if (p->cq_entries < p->sq_entries)
3816 p->cq_entries = 2 * p->sq_entries;
3819 ctx = io_ring_ctx_alloc(p);
3823 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3824 !(ctx->flags & IORING_SETUP_IOPOLL) &&
3825 !(ctx->flags & IORING_SETUP_SQPOLL))
3826 ctx->task_complete = true;
3829 * lazy poll_wq activation relies on ->task_complete for synchronisation
3830 * purposes, see io_activate_pollwq()
3832 if (!ctx->task_complete)
3833 ctx->poll_activated = true;
3836 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3837 * space applications don't need to do io completion events
3838 * polling again, they can rely on io_sq_thread to do polling
3839 * work, which can reduce cpu usage and uring_lock contention.
3841 if (ctx->flags & IORING_SETUP_IOPOLL &&
3842 !(ctx->flags & IORING_SETUP_SQPOLL))
3843 ctx->syscall_iopoll = 1;
3845 ctx->compat = in_compat_syscall();
3846 if (!capable(CAP_IPC_LOCK))
3847 ctx->user = get_uid(current_user());
3850 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3851 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3854 if (ctx->flags & IORING_SETUP_SQPOLL) {
3855 /* IPI related flags don't make sense with SQPOLL */
3856 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3857 IORING_SETUP_TASKRUN_FLAG |
3858 IORING_SETUP_DEFER_TASKRUN))
3860 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3861 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3862 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3864 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3865 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3867 ctx->notify_method = TWA_SIGNAL;
3871 * For DEFER_TASKRUN we require the completion task to be the same as the
3872 * submission task. This implies that there is only one submitter, so enforce
3875 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3876 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3881 * This is just grabbed for accounting purposes. When a process exits,
3882 * the mm is exited and dropped before the files, hence we need to hang
3883 * on to this mm purely for the purposes of being able to unaccount
3884 * memory (locked/pinned vm). It's not used for anything else.
3886 mmgrab(current->mm);
3887 ctx->mm_account = current->mm;
3889 ret = io_allocate_scq_urings(ctx, p);
3893 ret = io_sq_offload_create(ctx, p);
3897 ret = io_rsrc_init(ctx);
3901 p->sq_off.head = offsetof(struct io_rings, sq.head);
3902 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3903 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3904 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3905 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3906 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3907 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3908 p->sq_off.resv1 = 0;
3909 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3910 p->sq_off.user_addr = 0;
3912 p->cq_off.head = offsetof(struct io_rings, cq.head);
3913 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3914 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3915 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3916 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3917 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3918 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3919 p->cq_off.resv1 = 0;
3920 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3921 p->cq_off.user_addr = 0;
3923 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3924 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3925 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3926 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3927 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3928 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3929 IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING;
3931 if (copy_to_user(params, p, sizeof(*p))) {
3936 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
3937 && !(ctx->flags & IORING_SETUP_R_DISABLED))
3938 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
3940 file = io_uring_get_file(ctx);
3942 ret = PTR_ERR(file);
3946 ret = __io_uring_add_tctx_node(ctx);
3949 tctx = current->io_uring;
3952 * Install ring fd as the very last thing, so we don't risk someone
3953 * having closed it before we finish setup
3955 if (p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
3956 ret = io_ring_add_registered_file(tctx, file, 0, IO_RINGFD_REG_MAX);
3958 ret = io_uring_install_fd(file);
3962 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3965 io_ring_ctx_wait_and_kill(ctx);
3973 * Sets up an aio uring context, and returns the fd. Applications asks for a
3974 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3975 * params structure passed in.
3977 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3979 struct io_uring_params p;
3982 if (copy_from_user(&p, params, sizeof(p)))
3984 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3989 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3990 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3991 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3992 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3993 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3994 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3995 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN |
3996 IORING_SETUP_NO_MMAP | IORING_SETUP_REGISTERED_FD_ONLY))
3999 return io_uring_create(entries, &p, params);
4002 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
4003 struct io_uring_params __user *, params)
4005 return io_uring_setup(entries, params);
4008 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
4011 struct io_uring_probe *p;
4015 size = struct_size(p, ops, nr_args);
4016 if (size == SIZE_MAX)
4018 p = kzalloc(size, GFP_KERNEL);
4023 if (copy_from_user(p, arg, size))
4026 if (memchr_inv(p, 0, size))
4029 p->last_op = IORING_OP_LAST - 1;
4030 if (nr_args > IORING_OP_LAST)
4031 nr_args = IORING_OP_LAST;
4033 for (i = 0; i < nr_args; i++) {
4035 if (!io_issue_defs[i].not_supported)
4036 p->ops[i].flags = IO_URING_OP_SUPPORTED;
4041 if (copy_to_user(arg, p, size))
4048 static int io_register_personality(struct io_ring_ctx *ctx)
4050 const struct cred *creds;
4054 creds = get_current_cred();
4056 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
4057 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
4065 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
4066 void __user *arg, unsigned int nr_args)
4068 struct io_uring_restriction *res;
4072 /* Restrictions allowed only if rings started disabled */
4073 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
4076 /* We allow only a single restrictions registration */
4077 if (ctx->restrictions.registered)
4080 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
4083 size = array_size(nr_args, sizeof(*res));
4084 if (size == SIZE_MAX)
4087 res = memdup_user(arg, size);
4089 return PTR_ERR(res);
4093 for (i = 0; i < nr_args; i++) {
4094 switch (res[i].opcode) {
4095 case IORING_RESTRICTION_REGISTER_OP:
4096 if (res[i].register_op >= IORING_REGISTER_LAST) {
4101 __set_bit(res[i].register_op,
4102 ctx->restrictions.register_op);
4104 case IORING_RESTRICTION_SQE_OP:
4105 if (res[i].sqe_op >= IORING_OP_LAST) {
4110 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
4112 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
4113 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
4115 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
4116 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
4125 /* Reset all restrictions if an error happened */
4127 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
4129 ctx->restrictions.registered = true;
4135 static int io_register_enable_rings(struct io_ring_ctx *ctx)
4137 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
4140 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task) {
4141 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
4143 * Lazy activation attempts would fail if it was polled before
4144 * submitter_task is set.
4146 if (wq_has_sleeper(&ctx->poll_wq))
4147 io_activate_pollwq(ctx);
4150 if (ctx->restrictions.registered)
4151 ctx->restricted = 1;
4153 ctx->flags &= ~IORING_SETUP_R_DISABLED;
4154 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
4155 wake_up(&ctx->sq_data->wait);
4159 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
4160 void __user *arg, unsigned len)
4162 struct io_uring_task *tctx = current->io_uring;
4163 cpumask_var_t new_mask;
4166 if (!tctx || !tctx->io_wq)
4169 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4172 cpumask_clear(new_mask);
4173 if (len > cpumask_size())
4174 len = cpumask_size();
4176 if (in_compat_syscall()) {
4177 ret = compat_get_bitmap(cpumask_bits(new_mask),
4178 (const compat_ulong_t __user *)arg,
4179 len * 8 /* CHAR_BIT */);
4181 ret = copy_from_user(new_mask, arg, len);
4185 free_cpumask_var(new_mask);
4189 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
4190 free_cpumask_var(new_mask);
4194 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
4196 struct io_uring_task *tctx = current->io_uring;
4198 if (!tctx || !tctx->io_wq)
4201 return io_wq_cpu_affinity(tctx->io_wq, NULL);
4204 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
4206 __must_hold(&ctx->uring_lock)
4208 struct io_tctx_node *node;
4209 struct io_uring_task *tctx = NULL;
4210 struct io_sq_data *sqd = NULL;
4214 if (copy_from_user(new_count, arg, sizeof(new_count)))
4216 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4217 if (new_count[i] > INT_MAX)
4220 if (ctx->flags & IORING_SETUP_SQPOLL) {
4224 * Observe the correct sqd->lock -> ctx->uring_lock
4225 * ordering. Fine to drop uring_lock here, we hold
4228 refcount_inc(&sqd->refs);
4229 mutex_unlock(&ctx->uring_lock);
4230 mutex_lock(&sqd->lock);
4231 mutex_lock(&ctx->uring_lock);
4233 tctx = sqd->thread->io_uring;
4236 tctx = current->io_uring;
4239 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
4241 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4243 ctx->iowq_limits[i] = new_count[i];
4244 ctx->iowq_limits_set = true;
4246 if (tctx && tctx->io_wq) {
4247 ret = io_wq_max_workers(tctx->io_wq, new_count);
4251 memset(new_count, 0, sizeof(new_count));
4255 mutex_unlock(&sqd->lock);
4256 io_put_sq_data(sqd);
4259 if (copy_to_user(arg, new_count, sizeof(new_count)))
4262 /* that's it for SQPOLL, only the SQPOLL task creates requests */
4266 /* now propagate the restriction to all registered users */
4267 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
4268 struct io_uring_task *tctx = node->task->io_uring;
4270 if (WARN_ON_ONCE(!tctx->io_wq))
4273 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4274 new_count[i] = ctx->iowq_limits[i];
4275 /* ignore errors, it always returns zero anyway */
4276 (void)io_wq_max_workers(tctx->io_wq, new_count);
4281 mutex_unlock(&sqd->lock);
4282 io_put_sq_data(sqd);
4287 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
4288 void __user *arg, unsigned nr_args)
4289 __releases(ctx->uring_lock)
4290 __acquires(ctx->uring_lock)
4295 * We don't quiesce the refs for register anymore and so it can't be
4296 * dying as we're holding a file ref here.
4298 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
4301 if (ctx->submitter_task && ctx->submitter_task != current)
4304 if (ctx->restricted) {
4305 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
4306 if (!test_bit(opcode, ctx->restrictions.register_op))
4311 case IORING_REGISTER_BUFFERS:
4315 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
4317 case IORING_UNREGISTER_BUFFERS:
4321 ret = io_sqe_buffers_unregister(ctx);
4323 case IORING_REGISTER_FILES:
4327 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
4329 case IORING_UNREGISTER_FILES:
4333 ret = io_sqe_files_unregister(ctx);
4335 case IORING_REGISTER_FILES_UPDATE:
4336 ret = io_register_files_update(ctx, arg, nr_args);
4338 case IORING_REGISTER_EVENTFD:
4342 ret = io_eventfd_register(ctx, arg, 0);
4344 case IORING_REGISTER_EVENTFD_ASYNC:
4348 ret = io_eventfd_register(ctx, arg, 1);
4350 case IORING_UNREGISTER_EVENTFD:
4354 ret = io_eventfd_unregister(ctx);
4356 case IORING_REGISTER_PROBE:
4358 if (!arg || nr_args > 256)
4360 ret = io_probe(ctx, arg, nr_args);
4362 case IORING_REGISTER_PERSONALITY:
4366 ret = io_register_personality(ctx);
4368 case IORING_UNREGISTER_PERSONALITY:
4372 ret = io_unregister_personality(ctx, nr_args);
4374 case IORING_REGISTER_ENABLE_RINGS:
4378 ret = io_register_enable_rings(ctx);
4380 case IORING_REGISTER_RESTRICTIONS:
4381 ret = io_register_restrictions(ctx, arg, nr_args);
4383 case IORING_REGISTER_FILES2:
4384 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
4386 case IORING_REGISTER_FILES_UPDATE2:
4387 ret = io_register_rsrc_update(ctx, arg, nr_args,
4390 case IORING_REGISTER_BUFFERS2:
4391 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
4393 case IORING_REGISTER_BUFFERS_UPDATE:
4394 ret = io_register_rsrc_update(ctx, arg, nr_args,
4395 IORING_RSRC_BUFFER);
4397 case IORING_REGISTER_IOWQ_AFF:
4399 if (!arg || !nr_args)
4401 ret = io_register_iowq_aff(ctx, arg, nr_args);
4403 case IORING_UNREGISTER_IOWQ_AFF:
4407 ret = io_unregister_iowq_aff(ctx);
4409 case IORING_REGISTER_IOWQ_MAX_WORKERS:
4411 if (!arg || nr_args != 2)
4413 ret = io_register_iowq_max_workers(ctx, arg);
4415 case IORING_REGISTER_RING_FDS:
4416 ret = io_ringfd_register(ctx, arg, nr_args);
4418 case IORING_UNREGISTER_RING_FDS:
4419 ret = io_ringfd_unregister(ctx, arg, nr_args);
4421 case IORING_REGISTER_PBUF_RING:
4423 if (!arg || nr_args != 1)
4425 ret = io_register_pbuf_ring(ctx, arg);
4427 case IORING_UNREGISTER_PBUF_RING:
4429 if (!arg || nr_args != 1)
4431 ret = io_unregister_pbuf_ring(ctx, arg);
4433 case IORING_REGISTER_SYNC_CANCEL:
4435 if (!arg || nr_args != 1)
4437 ret = io_sync_cancel(ctx, arg);
4439 case IORING_REGISTER_FILE_ALLOC_RANGE:
4441 if (!arg || nr_args)
4443 ret = io_register_file_alloc_range(ctx, arg);
4453 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
4454 void __user *, arg, unsigned int, nr_args)
4456 struct io_ring_ctx *ctx;
4459 bool use_registered_ring;
4461 use_registered_ring = !!(opcode & IORING_REGISTER_USE_REGISTERED_RING);
4462 opcode &= ~IORING_REGISTER_USE_REGISTERED_RING;
4464 if (opcode >= IORING_REGISTER_LAST)
4467 if (use_registered_ring) {
4469 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
4470 * need only dereference our task private array to find it.
4472 struct io_uring_task *tctx = current->io_uring;
4474 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
4476 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
4477 f.file = tctx->registered_rings[fd];
4479 if (unlikely(!f.file))
4483 if (unlikely(!f.file))
4486 if (!io_is_uring_fops(f.file))
4490 ctx = f.file->private_data;
4492 mutex_lock(&ctx->uring_lock);
4493 ret = __io_uring_register(ctx, opcode, arg, nr_args);
4494 mutex_unlock(&ctx->uring_lock);
4495 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
4501 static int __init io_uring_init(void)
4503 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4504 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4505 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4508 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4509 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4510 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4511 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4512 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4513 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
4514 BUILD_BUG_SQE_ELEM(1, __u8, flags);
4515 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
4516 BUILD_BUG_SQE_ELEM(4, __s32, fd);
4517 BUILD_BUG_SQE_ELEM(8, __u64, off);
4518 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
4519 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
4520 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
4521 BUILD_BUG_SQE_ELEM(16, __u64, addr);
4522 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
4523 BUILD_BUG_SQE_ELEM(24, __u32, len);
4524 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
4525 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
4526 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4527 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
4528 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
4529 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
4530 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
4531 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
4532 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
4533 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
4534 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
4535 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
4536 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
4537 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
4538 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
4539 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
4540 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
4541 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
4542 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
4543 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
4544 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
4545 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
4546 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
4547 BUILD_BUG_SQE_ELEM(42, __u16, personality);
4548 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
4549 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
4550 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
4551 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
4552 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
4553 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4554 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
4556 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4557 sizeof(struct io_uring_rsrc_update));
4558 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4559 sizeof(struct io_uring_rsrc_update2));
4561 /* ->buf_index is u16 */
4562 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4563 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4564 offsetof(struct io_uring_buf_ring, tail));
4566 /* should fit into one byte */
4567 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
4568 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4569 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4571 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
4573 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4575 io_uring_optable_init();
4577 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4578 SLAB_ACCOUNT | SLAB_TYPESAFE_BY_RCU);
4581 __initcall(io_uring_init);