1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <net/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
49 #include <linux/bits.h>
51 #include <linux/sched/signal.h>
53 #include <linux/file.h>
54 #include <linux/fdtable.h>
56 #include <linux/mman.h>
57 #include <linux/percpu.h>
58 #include <linux/slab.h>
59 #include <linux/bvec.h>
60 #include <linux/net.h>
62 #include <net/af_unix.h>
64 #include <linux/anon_inodes.h>
65 #include <linux/sched/mm.h>
66 #include <linux/uaccess.h>
67 #include <linux/nospec.h>
68 #include <linux/highmem.h>
69 #include <linux/fsnotify.h>
70 #include <linux/fadvise.h>
71 #include <linux/task_work.h>
72 #include <linux/io_uring.h>
73 #include <linux/audit.h>
74 #include <linux/security.h>
76 #define CREATE_TRACE_POINTS
77 #include <trace/events/io_uring.h>
79 #include <uapi/linux/io_uring.h>
97 #include "alloc_cache.h"
99 #define IORING_MAX_ENTRIES 32768
100 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
102 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
103 IORING_REGISTER_LAST + IORING_OP_LAST)
105 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
106 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
108 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
109 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
111 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
112 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
115 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
118 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
120 #define IO_COMPL_BATCH 32
121 #define IO_REQ_ALLOC_BATCH 8
124 IO_CHECK_CQ_OVERFLOW_BIT,
125 IO_CHECK_CQ_DROPPED_BIT,
129 IO_EVENTFD_OP_SIGNAL_BIT,
130 IO_EVENTFD_OP_FREE_BIT,
133 struct io_defer_entry {
134 struct list_head list;
135 struct io_kiocb *req;
139 /* requests with any of those set should undergo io_disarm_next() */
140 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
141 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
143 static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
144 struct task_struct *task,
147 static void io_dismantle_req(struct io_kiocb *req);
148 static void io_clean_op(struct io_kiocb *req);
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);
152 static __cold void io_fallback_tw(struct io_uring_task *tctx);
154 static struct kmem_cache *req_cachep;
156 struct sock *io_uring_get_socket(struct file *file)
158 #if defined(CONFIG_UNIX)
159 if (io_is_uring_fops(file)) {
160 struct io_ring_ctx *ctx = file->private_data;
162 return ctx->ring_sock->sk;
167 EXPORT_SYMBOL(io_uring_get_socket);
169 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
171 if (!wq_list_empty(&ctx->submit_state.compl_reqs) ||
172 ctx->submit_state.cqes_count)
173 __io_submit_flush_completions(ctx);
176 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
178 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
181 static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx)
183 return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head);
186 static bool io_match_linked(struct io_kiocb *head)
188 struct io_kiocb *req;
190 io_for_each_link(req, head) {
191 if (req->flags & REQ_F_INFLIGHT)
198 * As io_match_task() but protected against racing with linked timeouts.
199 * User must not hold timeout_lock.
201 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
206 if (task && head->task != task)
211 if (head->flags & REQ_F_LINK_TIMEOUT) {
212 struct io_ring_ctx *ctx = head->ctx;
214 /* protect against races with linked timeouts */
215 spin_lock_irq(&ctx->timeout_lock);
216 matched = io_match_linked(head);
217 spin_unlock_irq(&ctx->timeout_lock);
219 matched = io_match_linked(head);
224 static inline void req_fail_link_node(struct io_kiocb *req, int res)
227 io_req_set_res(req, res, 0);
230 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
232 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
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;
250 percpu_ref_get(&ctx->refs);
251 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
252 req->io_task_work.func(req, &locked);
255 io_submit_flush_completions(ctx);
256 mutex_unlock(&ctx->uring_lock);
258 percpu_ref_put(&ctx->refs);
261 static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
263 unsigned hash_buckets = 1U << bits;
264 size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
266 table->hbs = kmalloc(hash_size, GFP_KERNEL);
270 table->hash_bits = bits;
271 init_hash_table(table, hash_buckets);
275 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
277 struct io_ring_ctx *ctx;
280 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
284 xa_init(&ctx->io_bl_xa);
287 * Use 5 bits less than the max cq entries, that should give us around
288 * 32 entries per hash list if totally full and uniformly spread, but
289 * don't keep too many buckets to not overconsume memory.
291 hash_bits = ilog2(p->cq_entries) - 5;
292 hash_bits = clamp(hash_bits, 1, 8);
293 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
295 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
298 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
299 if (!ctx->dummy_ubuf)
301 /* set invalid range, so io_import_fixed() fails meeting it */
302 ctx->dummy_ubuf->ubuf = -1UL;
304 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
308 ctx->flags = p->flags;
309 init_waitqueue_head(&ctx->sqo_sq_wait);
310 INIT_LIST_HEAD(&ctx->sqd_list);
311 INIT_LIST_HEAD(&ctx->cq_overflow_list);
312 INIT_LIST_HEAD(&ctx->io_buffers_cache);
313 io_alloc_cache_init(&ctx->apoll_cache);
314 io_alloc_cache_init(&ctx->netmsg_cache);
315 init_completion(&ctx->ref_comp);
316 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
317 mutex_init(&ctx->uring_lock);
318 init_waitqueue_head(&ctx->cq_wait);
319 spin_lock_init(&ctx->completion_lock);
320 spin_lock_init(&ctx->timeout_lock);
321 INIT_WQ_LIST(&ctx->iopoll_list);
322 INIT_LIST_HEAD(&ctx->io_buffers_pages);
323 INIT_LIST_HEAD(&ctx->io_buffers_comp);
324 INIT_LIST_HEAD(&ctx->defer_list);
325 INIT_LIST_HEAD(&ctx->timeout_list);
326 INIT_LIST_HEAD(&ctx->ltimeout_list);
327 spin_lock_init(&ctx->rsrc_ref_lock);
328 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
329 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
330 init_task_work(&ctx->rsrc_put_tw, io_rsrc_put_tw);
331 init_llist_head(&ctx->rsrc_put_llist);
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 inline void io_req_track_inflight(struct io_kiocb *req)
370 if (!(req->flags & REQ_F_INFLIGHT)) {
371 req->flags |= REQ_F_INFLIGHT;
372 atomic_inc(&req->task->io_uring->inflight_tracked);
376 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
378 if (WARN_ON_ONCE(!req->link))
381 req->flags &= ~REQ_F_ARM_LTIMEOUT;
382 req->flags |= REQ_F_LINK_TIMEOUT;
384 /* linked timeouts should have two refs once prep'ed */
385 io_req_set_refcount(req);
386 __io_req_set_refcount(req->link, 2);
390 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
392 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
394 return __io_prep_linked_timeout(req);
397 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
399 io_queue_linked_timeout(__io_prep_linked_timeout(req));
402 static inline void io_arm_ltimeout(struct io_kiocb *req)
404 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
405 __io_arm_ltimeout(req);
408 static void io_prep_async_work(struct io_kiocb *req)
410 const struct io_op_def *def = &io_op_defs[req->opcode];
411 struct io_ring_ctx *ctx = req->ctx;
413 if (!(req->flags & REQ_F_CREDS)) {
414 req->flags |= REQ_F_CREDS;
415 req->creds = get_current_cred();
418 req->work.list.next = NULL;
420 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
421 if (req->flags & REQ_F_FORCE_ASYNC)
422 req->work.flags |= IO_WQ_WORK_CONCURRENT;
424 if (req->file && !io_req_ffs_set(req))
425 req->flags |= io_file_get_flags(req->file) << REQ_F_SUPPORT_NOWAIT_BIT;
427 if (req->flags & REQ_F_ISREG) {
428 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
429 io_wq_hash_work(&req->work, file_inode(req->file));
430 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
431 if (def->unbound_nonreg_file)
432 req->work.flags |= IO_WQ_WORK_UNBOUND;
436 static void io_prep_async_link(struct io_kiocb *req)
438 struct io_kiocb *cur;
440 if (req->flags & REQ_F_LINK_TIMEOUT) {
441 struct io_ring_ctx *ctx = req->ctx;
443 spin_lock_irq(&ctx->timeout_lock);
444 io_for_each_link(cur, req)
445 io_prep_async_work(cur);
446 spin_unlock_irq(&ctx->timeout_lock);
448 io_for_each_link(cur, req)
449 io_prep_async_work(cur);
453 void io_queue_iowq(struct io_kiocb *req, bool *dont_use)
455 struct io_kiocb *link = io_prep_linked_timeout(req);
456 struct io_uring_task *tctx = req->task->io_uring;
459 BUG_ON(!tctx->io_wq);
461 /* init ->work of the whole link before punting */
462 io_prep_async_link(req);
465 * Not expected to happen, but if we do have a bug where this _can_
466 * happen, catch it here and ensure the request is marked as
467 * canceled. That will make io-wq go through the usual work cancel
468 * procedure rather than attempt to run this request (or create a new
471 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
472 req->work.flags |= IO_WQ_WORK_CANCEL;
474 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
475 io_wq_enqueue(tctx->io_wq, &req->work);
477 io_queue_linked_timeout(link);
480 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
482 while (!list_empty(&ctx->defer_list)) {
483 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
484 struct io_defer_entry, list);
486 if (req_need_defer(de->req, de->seq))
488 list_del_init(&de->list);
489 io_req_task_queue(de->req);
495 static void io_eventfd_ops(struct rcu_head *rcu)
497 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
498 int ops = atomic_xchg(&ev_fd->ops, 0);
500 if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT))
501 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
503 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
504 * ordering in a race but if references are 0 we know we have to free
507 if (atomic_dec_and_test(&ev_fd->refs)) {
508 eventfd_ctx_put(ev_fd->cq_ev_fd);
513 static void io_eventfd_signal(struct io_ring_ctx *ctx)
515 struct io_ev_fd *ev_fd = NULL;
519 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
522 ev_fd = rcu_dereference(ctx->io_ev_fd);
525 * Check again if ev_fd exists incase an io_eventfd_unregister call
526 * completed between the NULL check of ctx->io_ev_fd at the start of
527 * the function and rcu_read_lock.
529 if (unlikely(!ev_fd))
531 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
533 if (ev_fd->eventfd_async && !io_wq_current_is_worker())
536 if (likely(eventfd_signal_allowed())) {
537 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
539 atomic_inc(&ev_fd->refs);
540 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops))
541 call_rcu(&ev_fd->rcu, io_eventfd_ops);
543 atomic_dec(&ev_fd->refs);
550 static void io_eventfd_flush_signal(struct io_ring_ctx *ctx)
554 spin_lock(&ctx->completion_lock);
557 * Eventfd should only get triggered when at least one event has been
558 * posted. Some applications rely on the eventfd notification count
559 * only changing IFF a new CQE has been added to the CQ ring. There's
560 * no depedency on 1:1 relationship between how many times this
561 * function is called (and hence the eventfd count) and number of CQEs
562 * posted to the CQ ring.
564 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
565 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
566 spin_unlock(&ctx->completion_lock);
570 io_eventfd_signal(ctx);
573 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
575 if (ctx->off_timeout_used || ctx->drain_active) {
576 spin_lock(&ctx->completion_lock);
577 if (ctx->off_timeout_used)
578 io_flush_timeouts(ctx);
579 if (ctx->drain_active)
580 io_queue_deferred(ctx);
581 spin_unlock(&ctx->completion_lock);
584 io_eventfd_flush_signal(ctx);
587 static inline void __io_cq_lock(struct io_ring_ctx *ctx)
588 __acquires(ctx->completion_lock)
590 if (!ctx->task_complete)
591 spin_lock(&ctx->completion_lock);
594 static inline void __io_cq_unlock(struct io_ring_ctx *ctx)
596 if (!ctx->task_complete)
597 spin_unlock(&ctx->completion_lock);
600 /* keep it inlined for io_submit_flush_completions() */
601 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
602 __releases(ctx->completion_lock)
604 io_commit_cqring(ctx);
606 io_commit_cqring_flush(ctx);
610 void io_cq_unlock_post(struct io_ring_ctx *ctx)
611 __releases(ctx->completion_lock)
613 io_commit_cqring(ctx);
614 spin_unlock(&ctx->completion_lock);
615 io_commit_cqring_flush(ctx);
619 /* Returns true if there are no backlogged entries after the flush */
620 static void io_cqring_overflow_kill(struct io_ring_ctx *ctx)
622 struct io_overflow_cqe *ocqe;
626 list_splice_init(&ctx->cq_overflow_list, &list);
627 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
630 while (!list_empty(&list)) {
631 ocqe = list_first_entry(&list, struct io_overflow_cqe, list);
632 list_del(&ocqe->list);
637 /* Returns true if there are no backlogged entries after the flush */
638 static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx)
640 size_t cqe_size = sizeof(struct io_uring_cqe);
642 if (__io_cqring_events(ctx) == ctx->cq_entries)
645 if (ctx->flags & IORING_SETUP_CQE32)
649 while (!list_empty(&ctx->cq_overflow_list)) {
650 struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true);
651 struct io_overflow_cqe *ocqe;
655 ocqe = list_first_entry(&ctx->cq_overflow_list,
656 struct io_overflow_cqe, list);
657 memcpy(cqe, &ocqe->cqe, cqe_size);
658 list_del(&ocqe->list);
662 if (list_empty(&ctx->cq_overflow_list)) {
663 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
664 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
666 io_cq_unlock_post(ctx);
669 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx)
671 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
672 /* iopoll syncs against uring_lock, not completion_lock */
673 if (ctx->flags & IORING_SETUP_IOPOLL)
674 mutex_lock(&ctx->uring_lock);
675 __io_cqring_overflow_flush(ctx);
676 if (ctx->flags & IORING_SETUP_IOPOLL)
677 mutex_unlock(&ctx->uring_lock);
681 void __io_put_task(struct task_struct *task, int nr)
683 struct io_uring_task *tctx = task->io_uring;
685 percpu_counter_sub(&tctx->inflight, nr);
686 if (unlikely(atomic_read(&tctx->in_idle)))
687 wake_up(&tctx->wait);
688 put_task_struct_many(task, nr);
691 void io_task_refs_refill(struct io_uring_task *tctx)
693 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
695 percpu_counter_add(&tctx->inflight, refill);
696 refcount_add(refill, ¤t->usage);
697 tctx->cached_refs += refill;
700 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
702 struct io_uring_task *tctx = task->io_uring;
703 unsigned int refs = tctx->cached_refs;
706 tctx->cached_refs = 0;
707 percpu_counter_sub(&tctx->inflight, refs);
708 put_task_struct_many(task, refs);
712 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
713 s32 res, u32 cflags, u64 extra1, u64 extra2)
715 struct io_overflow_cqe *ocqe;
716 size_t ocq_size = sizeof(struct io_overflow_cqe);
717 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
720 ocq_size += sizeof(struct io_uring_cqe);
722 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
723 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
726 * If we're in ring overflow flush mode, or in task cancel mode,
727 * or cannot allocate an overflow entry, then we need to drop it
730 io_account_cq_overflow(ctx);
731 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
734 if (list_empty(&ctx->cq_overflow_list)) {
735 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
736 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
739 ocqe->cqe.user_data = user_data;
741 ocqe->cqe.flags = cflags;
743 ocqe->cqe.big_cqe[0] = extra1;
744 ocqe->cqe.big_cqe[1] = extra2;
746 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
750 bool io_req_cqe_overflow(struct io_kiocb *req)
752 if (!(req->flags & REQ_F_CQE32_INIT)) {
756 return io_cqring_event_overflow(req->ctx, req->cqe.user_data,
757 req->cqe.res, req->cqe.flags,
758 req->extra1, req->extra2);
762 * writes to the cq entry need to come after reading head; the
763 * control dependency is enough as we're using WRITE_ONCE to
766 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow)
768 struct io_rings *rings = ctx->rings;
769 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
770 unsigned int free, queued, len;
773 * Posting into the CQ when there are pending overflowed CQEs may break
774 * ordering guarantees, which will affect links, F_MORE users and more.
775 * Force overflow the completion.
777 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
780 /* userspace may cheat modifying the tail, be safe and do min */
781 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
782 free = ctx->cq_entries - queued;
783 /* we need a contiguous range, limit based on the current array offset */
784 len = min(free, ctx->cq_entries - off);
788 if (ctx->flags & IORING_SETUP_CQE32) {
793 ctx->cqe_cached = &rings->cqes[off];
794 ctx->cqe_sentinel = ctx->cqe_cached + len;
796 ctx->cached_cq_tail++;
798 if (ctx->flags & IORING_SETUP_CQE32)
800 return &rings->cqes[off];
803 static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
806 struct io_uring_cqe *cqe;
808 if (!ctx->task_complete)
809 lockdep_assert_held(&ctx->completion_lock);
814 * If we can't get a cq entry, userspace overflowed the
815 * submission (by quite a lot). Increment the overflow count in
818 cqe = io_get_cqe(ctx);
820 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
822 WRITE_ONCE(cqe->user_data, user_data);
823 WRITE_ONCE(cqe->res, res);
824 WRITE_ONCE(cqe->flags, cflags);
826 if (ctx->flags & IORING_SETUP_CQE32) {
827 WRITE_ONCE(cqe->big_cqe[0], 0);
828 WRITE_ONCE(cqe->big_cqe[1], 0);
835 static void __io_flush_post_cqes(struct io_ring_ctx *ctx)
836 __must_hold(&ctx->uring_lock)
838 struct io_submit_state *state = &ctx->submit_state;
841 lockdep_assert_held(&ctx->uring_lock);
842 for (i = 0; i < state->cqes_count; i++) {
843 struct io_uring_cqe *cqe = &state->cqes[i];
845 if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) {
846 if (ctx->task_complete) {
847 spin_lock(&ctx->completion_lock);
848 io_cqring_event_overflow(ctx, cqe->user_data,
849 cqe->res, cqe->flags, 0, 0);
850 spin_unlock(&ctx->completion_lock);
852 io_cqring_event_overflow(ctx, cqe->user_data,
853 cqe->res, cqe->flags, 0, 0);
857 state->cqes_count = 0;
860 static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
866 filled = io_fill_cqe_aux(ctx, user_data, res, cflags);
867 if (!filled && allow_overflow)
868 filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
870 io_cq_unlock_post(ctx);
874 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags)
876 return __io_post_aux_cqe(ctx, user_data, res, cflags, true);
879 bool io_aux_cqe(struct io_ring_ctx *ctx, bool defer, u64 user_data, s32 res, u32 cflags,
882 struct io_uring_cqe *cqe;
886 return __io_post_aux_cqe(ctx, user_data, res, cflags, allow_overflow);
888 length = ARRAY_SIZE(ctx->submit_state.cqes);
890 lockdep_assert_held(&ctx->uring_lock);
892 if (ctx->submit_state.cqes_count == length) {
894 __io_flush_post_cqes(ctx);
895 /* no need to flush - flush is deferred */
896 __io_cq_unlock_post(ctx);
899 /* For defered completions this is not as strict as it is otherwise,
900 * however it's main job is to prevent unbounded posted completions,
901 * and in that it works just as well.
903 if (!allow_overflow && test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
906 cqe = &ctx->submit_state.cqes[ctx->submit_state.cqes_count++];
907 cqe->user_data = user_data;
913 static void __io_req_complete_post(struct io_kiocb *req)
915 struct io_ring_ctx *ctx = req->ctx;
918 if (!(req->flags & REQ_F_CQE_SKIP))
919 __io_fill_cqe_req(ctx, req);
922 * If we're the last reference to this request, add to our locked
925 if (req_ref_put_and_test(req)) {
926 if (req->flags & IO_REQ_LINK_FLAGS) {
927 if (req->flags & IO_DISARM_MASK)
930 io_req_task_queue(req->link);
934 io_req_put_rsrc(req);
936 * Selected buffer deallocation in io_clean_op() assumes that
937 * we don't hold ->completion_lock. Clean them here to avoid
940 io_put_kbuf_comp(req);
941 io_dismantle_req(req);
942 io_put_task(req->task, 1);
943 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
944 ctx->locked_free_nr++;
946 io_cq_unlock_post(ctx);
949 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
951 if (req->ctx->task_complete && (issue_flags & IO_URING_F_IOWQ)) {
952 req->io_task_work.func = io_req_task_complete;
953 io_req_task_work_add(req);
954 } else if (!(issue_flags & IO_URING_F_UNLOCKED) ||
955 !(req->ctx->flags & IORING_SETUP_IOPOLL)) {
956 __io_req_complete_post(req);
958 struct io_ring_ctx *ctx = req->ctx;
960 mutex_lock(&ctx->uring_lock);
961 __io_req_complete_post(req);
962 mutex_unlock(&ctx->uring_lock);
966 void io_req_defer_failed(struct io_kiocb *req, s32 res)
967 __must_hold(&ctx->uring_lock)
969 const struct io_op_def *def = &io_op_defs[req->opcode];
971 lockdep_assert_held(&req->ctx->uring_lock);
974 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
977 io_req_complete_defer(req);
981 * Don't initialise the fields below on every allocation, but do that in
982 * advance and keep them valid across allocations.
984 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
988 req->async_data = NULL;
989 /* not necessary, but safer to zero */
993 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
994 struct io_submit_state *state)
996 spin_lock(&ctx->completion_lock);
997 wq_list_splice(&ctx->locked_free_list, &state->free_list);
998 ctx->locked_free_nr = 0;
999 spin_unlock(&ctx->completion_lock);
1003 * A request might get retired back into the request caches even before opcode
1004 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1005 * Because of that, io_alloc_req() should be called only under ->uring_lock
1006 * and with extra caution to not get a request that is still worked on.
1008 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
1009 __must_hold(&ctx->uring_lock)
1011 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1012 void *reqs[IO_REQ_ALLOC_BATCH];
1016 * If we have more than a batch's worth of requests in our IRQ side
1017 * locked cache, grab the lock and move them over to our submission
1020 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
1021 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
1022 if (!io_req_cache_empty(ctx))
1026 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
1029 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1030 * retry single alloc to be on the safe side.
1032 if (unlikely(ret <= 0)) {
1033 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1039 percpu_ref_get_many(&ctx->refs, ret);
1040 for (i = 0; i < ret; i++) {
1041 struct io_kiocb *req = reqs[i];
1043 io_preinit_req(req, ctx);
1044 io_req_add_to_cache(req, ctx);
1049 static inline void io_dismantle_req(struct io_kiocb *req)
1051 unsigned int flags = req->flags;
1053 if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
1055 if (!(flags & REQ_F_FIXED_FILE))
1056 io_put_file(req->file);
1059 __cold void io_free_req(struct io_kiocb *req)
1061 struct io_ring_ctx *ctx = req->ctx;
1063 io_req_put_rsrc(req);
1064 io_dismantle_req(req);
1065 io_put_task(req->task, 1);
1067 spin_lock(&ctx->completion_lock);
1068 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
1069 ctx->locked_free_nr++;
1070 spin_unlock(&ctx->completion_lock);
1073 static void __io_req_find_next_prep(struct io_kiocb *req)
1075 struct io_ring_ctx *ctx = req->ctx;
1078 io_disarm_next(req);
1079 io_cq_unlock_post(ctx);
1082 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1084 struct io_kiocb *nxt;
1087 * If LINK is set, we have dependent requests in this chain. If we
1088 * didn't fail this request, queue the first one up, moving any other
1089 * dependencies to the next request. In case of failure, fail the rest
1092 if (unlikely(req->flags & IO_DISARM_MASK))
1093 __io_req_find_next_prep(req);
1099 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
1103 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1104 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1106 io_submit_flush_completions(ctx);
1107 mutex_unlock(&ctx->uring_lock);
1110 percpu_ref_put(&ctx->refs);
1113 static unsigned int handle_tw_list(struct llist_node *node,
1114 struct io_ring_ctx **ctx, bool *locked,
1115 struct llist_node *last)
1117 unsigned int count = 0;
1119 while (node != last) {
1120 struct llist_node *next = node->next;
1121 struct io_kiocb *req = container_of(node, struct io_kiocb,
1124 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1126 if (req->ctx != *ctx) {
1127 ctx_flush_and_put(*ctx, locked);
1129 /* if not contended, grab and improve batching */
1130 *locked = mutex_trylock(&(*ctx)->uring_lock);
1131 percpu_ref_get(&(*ctx)->refs);
1133 req->io_task_work.func(req, locked);
1142 * io_llist_xchg - swap all entries in a lock-less list
1143 * @head: the head of lock-less list to delete all entries
1144 * @new: new entry as the head of the list
1146 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1147 * The order of entries returned is from the newest to the oldest added one.
1149 static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1150 struct llist_node *new)
1152 return xchg(&head->first, new);
1156 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1157 * @head: the head of lock-less list to delete all entries
1158 * @old: expected old value of the first entry of the list
1159 * @new: new entry as the head of the list
1161 * perform a cmpxchg on the first entry of the list.
1164 static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1165 struct llist_node *old,
1166 struct llist_node *new)
1168 return cmpxchg(&head->first, old, new);
1171 void tctx_task_work(struct callback_head *cb)
1173 bool uring_locked = false;
1174 struct io_ring_ctx *ctx = NULL;
1175 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1177 struct llist_node fake = {};
1178 struct llist_node *node;
1179 unsigned int loops = 1;
1182 if (unlikely(current->flags & PF_EXITING)) {
1183 io_fallback_tw(tctx);
1187 node = io_llist_xchg(&tctx->task_list, &fake);
1188 count = handle_tw_list(node, &ctx, &uring_locked, NULL);
1189 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1190 while (node != &fake) {
1192 node = io_llist_xchg(&tctx->task_list, &fake);
1193 count += handle_tw_list(node, &ctx, &uring_locked, &fake);
1194 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1197 ctx_flush_and_put(ctx, &uring_locked);
1199 /* relaxed read is enough as only the task itself sets ->in_idle */
1200 if (unlikely(atomic_read(&tctx->in_idle)))
1201 io_uring_drop_tctx_refs(current);
1203 trace_io_uring_task_work_run(tctx, count, loops);
1206 static __cold void io_fallback_tw(struct io_uring_task *tctx)
1208 struct llist_node *node = llist_del_all(&tctx->task_list);
1209 struct io_kiocb *req;
1212 req = container_of(node, struct io_kiocb, io_task_work.node);
1214 if (llist_add(&req->io_task_work.node,
1215 &req->ctx->fallback_llist))
1216 schedule_delayed_work(&req->ctx->fallback_work, 1);
1220 static void io_req_local_work_add(struct io_kiocb *req)
1222 struct io_ring_ctx *ctx = req->ctx;
1224 if (!llist_add(&req->io_task_work.node, &ctx->work_llist))
1226 /* need it for the following io_cqring_wake() */
1227 smp_mb__after_atomic();
1229 if (unlikely(atomic_read(&req->task->io_uring->in_idle))) {
1230 io_move_task_work_from_local(ctx);
1234 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1235 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1238 io_eventfd_signal(ctx);
1239 __io_cqring_wake(ctx);
1242 void __io_req_task_work_add(struct io_kiocb *req, bool allow_local)
1244 struct io_uring_task *tctx = req->task->io_uring;
1245 struct io_ring_ctx *ctx = req->ctx;
1247 if (allow_local && ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
1248 io_req_local_work_add(req);
1252 /* task_work already pending, we're done */
1253 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1256 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1257 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1259 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1262 io_fallback_tw(tctx);
1265 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1267 struct llist_node *node;
1269 node = llist_del_all(&ctx->work_llist);
1271 struct io_kiocb *req = container_of(node, struct io_kiocb,
1275 __io_req_task_work_add(req, false);
1279 int __io_run_local_work(struct io_ring_ctx *ctx, bool *locked)
1281 struct llist_node *node;
1282 struct llist_node fake;
1283 struct llist_node *current_final = NULL;
1285 unsigned int loops = 1;
1287 if (unlikely(ctx->submitter_task != current))
1290 node = io_llist_xchg(&ctx->work_llist, &fake);
1293 while (node != current_final) {
1294 struct llist_node *next = node->next;
1295 struct io_kiocb *req = container_of(node, struct io_kiocb,
1297 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1298 req->io_task_work.func(req, locked);
1303 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1304 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1306 node = io_llist_cmpxchg(&ctx->work_llist, &fake, NULL);
1307 if (node != &fake) {
1309 current_final = &fake;
1310 node = io_llist_xchg(&ctx->work_llist, &fake);
1315 io_submit_flush_completions(ctx);
1316 trace_io_uring_local_work_run(ctx, ret, loops);
1321 int io_run_local_work(struct io_ring_ctx *ctx)
1326 if (llist_empty(&ctx->work_llist))
1329 __set_current_state(TASK_RUNNING);
1330 locked = mutex_trylock(&ctx->uring_lock);
1331 ret = __io_run_local_work(ctx, &locked);
1333 mutex_unlock(&ctx->uring_lock);
1338 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
1340 io_tw_lock(req->ctx, locked);
1341 io_req_defer_failed(req, req->cqe.res);
1344 void io_req_task_submit(struct io_kiocb *req, bool *locked)
1346 io_tw_lock(req->ctx, locked);
1347 /* req->task == current here, checking PF_EXITING is safe */
1348 if (likely(!(req->task->flags & PF_EXITING)))
1351 io_req_defer_failed(req, -EFAULT);
1354 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1356 io_req_set_res(req, ret, 0);
1357 req->io_task_work.func = io_req_task_cancel;
1358 io_req_task_work_add(req);
1361 void io_req_task_queue(struct io_kiocb *req)
1363 req->io_task_work.func = io_req_task_submit;
1364 io_req_task_work_add(req);
1367 void io_queue_next(struct io_kiocb *req)
1369 struct io_kiocb *nxt = io_req_find_next(req);
1372 io_req_task_queue(nxt);
1375 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1376 __must_hold(&ctx->uring_lock)
1378 struct task_struct *task = NULL;
1382 struct io_kiocb *req = container_of(node, struct io_kiocb,
1385 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1386 if (req->flags & REQ_F_REFCOUNT) {
1387 node = req->comp_list.next;
1388 if (!req_ref_put_and_test(req))
1391 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1392 struct async_poll *apoll = req->apoll;
1394 if (apoll->double_poll)
1395 kfree(apoll->double_poll);
1396 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1398 req->flags &= ~REQ_F_POLLED;
1400 if (req->flags & IO_REQ_LINK_FLAGS)
1402 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1405 if (!(req->flags & REQ_F_FIXED_FILE))
1406 io_put_file(req->file);
1408 io_req_put_rsrc_locked(req, ctx);
1410 if (req->task != task) {
1412 io_put_task(task, task_refs);
1417 node = req->comp_list.next;
1418 io_req_add_to_cache(req, ctx);
1422 io_put_task(task, task_refs);
1425 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1426 __must_hold(&ctx->uring_lock)
1428 struct io_wq_work_node *node, *prev;
1429 struct io_submit_state *state = &ctx->submit_state;
1432 /* must come first to preserve CQE ordering in failure cases */
1433 if (state->cqes_count)
1434 __io_flush_post_cqes(ctx);
1435 wq_list_for_each(node, prev, &state->compl_reqs) {
1436 struct io_kiocb *req = container_of(node, struct io_kiocb,
1439 if (!(req->flags & REQ_F_CQE_SKIP) &&
1440 unlikely(!__io_fill_cqe_req(ctx, req))) {
1441 if (ctx->task_complete) {
1442 spin_lock(&ctx->completion_lock);
1443 io_req_cqe_overflow(req);
1444 spin_unlock(&ctx->completion_lock);
1446 io_req_cqe_overflow(req);
1450 __io_cq_unlock_post(ctx);
1452 if (!wq_list_empty(&ctx->submit_state.compl_reqs)) {
1453 io_free_batch_list(ctx, state->compl_reqs.first);
1454 INIT_WQ_LIST(&state->compl_reqs);
1459 * Drop reference to request, return next in chain (if there is one) if this
1460 * was the last reference to this request.
1462 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1464 struct io_kiocb *nxt = NULL;
1466 if (req_ref_put_and_test(req)) {
1467 if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
1468 nxt = io_req_find_next(req);
1474 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1476 /* See comment at the top of this file */
1478 return __io_cqring_events(ctx);
1482 * We can't just wait for polled events to come to us, we have to actively
1483 * find and complete them.
1485 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1487 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1490 mutex_lock(&ctx->uring_lock);
1491 while (!wq_list_empty(&ctx->iopoll_list)) {
1492 /* let it sleep and repeat later if can't complete a request */
1493 if (io_do_iopoll(ctx, true) == 0)
1496 * Ensure we allow local-to-the-cpu processing to take place,
1497 * in this case we need to ensure that we reap all events.
1498 * Also let task_work, etc. to progress by releasing the mutex
1500 if (need_resched()) {
1501 mutex_unlock(&ctx->uring_lock);
1503 mutex_lock(&ctx->uring_lock);
1506 mutex_unlock(&ctx->uring_lock);
1509 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1511 unsigned int nr_events = 0;
1513 unsigned long check_cq;
1515 if (!io_allowed_run_tw(ctx))
1518 check_cq = READ_ONCE(ctx->check_cq);
1519 if (unlikely(check_cq)) {
1520 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1521 __io_cqring_overflow_flush(ctx);
1523 * Similarly do not spin if we have not informed the user of any
1526 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1530 * Don't enter poll loop if we already have events pending.
1531 * If we do, we can potentially be spinning for commands that
1532 * already triggered a CQE (eg in error).
1534 if (io_cqring_events(ctx))
1539 * If a submit got punted to a workqueue, we can have the
1540 * application entering polling for a command before it gets
1541 * issued. That app will hold the uring_lock for the duration
1542 * of the poll right here, so we need to take a breather every
1543 * now and then to ensure that the issue has a chance to add
1544 * the poll to the issued list. Otherwise we can spin here
1545 * forever, while the workqueue is stuck trying to acquire the
1548 if (wq_list_empty(&ctx->iopoll_list) ||
1549 io_task_work_pending(ctx)) {
1550 u32 tail = ctx->cached_cq_tail;
1552 (void) io_run_local_work_locked(ctx);
1554 if (task_work_pending(current) ||
1555 wq_list_empty(&ctx->iopoll_list)) {
1556 mutex_unlock(&ctx->uring_lock);
1558 mutex_lock(&ctx->uring_lock);
1560 /* some requests don't go through iopoll_list */
1561 if (tail != ctx->cached_cq_tail ||
1562 wq_list_empty(&ctx->iopoll_list))
1565 ret = io_do_iopoll(ctx, !min);
1570 } while (nr_events < min && !need_resched());
1575 void io_req_task_complete(struct io_kiocb *req, bool *locked)
1578 io_req_complete_defer(req);
1580 io_req_complete_post(req, IO_URING_F_UNLOCKED);
1584 * After the iocb has been issued, it's safe to be found on the poll list.
1585 * Adding the kiocb to the list AFTER submission ensures that we don't
1586 * find it from a io_do_iopoll() thread before the issuer is done
1587 * accessing the kiocb cookie.
1589 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1591 struct io_ring_ctx *ctx = req->ctx;
1592 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1594 /* workqueue context doesn't hold uring_lock, grab it now */
1595 if (unlikely(needs_lock))
1596 mutex_lock(&ctx->uring_lock);
1599 * Track whether we have multiple files in our lists. This will impact
1600 * how we do polling eventually, not spinning if we're on potentially
1601 * different devices.
1603 if (wq_list_empty(&ctx->iopoll_list)) {
1604 ctx->poll_multi_queue = false;
1605 } else if (!ctx->poll_multi_queue) {
1606 struct io_kiocb *list_req;
1608 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1610 if (list_req->file != req->file)
1611 ctx->poll_multi_queue = true;
1615 * For fast devices, IO may have already completed. If it has, add
1616 * it to the front so we find it first.
1618 if (READ_ONCE(req->iopoll_completed))
1619 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1621 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1623 if (unlikely(needs_lock)) {
1625 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1626 * in sq thread task context or in io worker task context. If
1627 * current task context is sq thread, we don't need to check
1628 * whether should wake up sq thread.
1630 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1631 wq_has_sleeper(&ctx->sq_data->wait))
1632 wake_up(&ctx->sq_data->wait);
1634 mutex_unlock(&ctx->uring_lock);
1638 static bool io_bdev_nowait(struct block_device *bdev)
1640 return !bdev || bdev_nowait(bdev);
1644 * If we tracked the file through the SCM inflight mechanism, we could support
1645 * any file. For now, just ensure that anything potentially problematic is done
1648 static bool __io_file_supports_nowait(struct file *file, umode_t mode)
1650 if (S_ISBLK(mode)) {
1651 if (IS_ENABLED(CONFIG_BLOCK) &&
1652 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
1658 if (S_ISREG(mode)) {
1659 if (IS_ENABLED(CONFIG_BLOCK) &&
1660 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
1661 !io_is_uring_fops(file))
1666 /* any ->read/write should understand O_NONBLOCK */
1667 if (file->f_flags & O_NONBLOCK)
1669 return file->f_mode & FMODE_NOWAIT;
1673 * If we tracked the file through the SCM inflight mechanism, we could support
1674 * any file. For now, just ensure that anything potentially problematic is done
1677 unsigned int io_file_get_flags(struct file *file)
1679 umode_t mode = file_inode(file)->i_mode;
1680 unsigned int res = 0;
1684 if (__io_file_supports_nowait(file, mode))
1689 bool io_alloc_async_data(struct io_kiocb *req)
1691 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
1692 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
1693 if (req->async_data) {
1694 req->flags |= REQ_F_ASYNC_DATA;
1700 int io_req_prep_async(struct io_kiocb *req)
1702 const struct io_op_def *def = &io_op_defs[req->opcode];
1704 /* assign early for deferred execution for non-fixed file */
1705 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE))
1706 req->file = io_file_get_normal(req, req->cqe.fd);
1707 if (!def->prep_async)
1709 if (WARN_ON_ONCE(req_has_async_data(req)))
1711 if (!io_op_defs[req->opcode].manual_alloc) {
1712 if (io_alloc_async_data(req))
1715 return def->prep_async(req);
1718 static u32 io_get_sequence(struct io_kiocb *req)
1720 u32 seq = req->ctx->cached_sq_head;
1721 struct io_kiocb *cur;
1723 /* need original cached_sq_head, but it was increased for each req */
1724 io_for_each_link(cur, req)
1729 static __cold void io_drain_req(struct io_kiocb *req)
1730 __must_hold(&ctx->uring_lock)
1732 struct io_ring_ctx *ctx = req->ctx;
1733 struct io_defer_entry *de;
1735 u32 seq = io_get_sequence(req);
1737 /* Still need defer if there is pending req in defer list. */
1738 spin_lock(&ctx->completion_lock);
1739 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1740 spin_unlock(&ctx->completion_lock);
1742 ctx->drain_active = false;
1743 io_req_task_queue(req);
1746 spin_unlock(&ctx->completion_lock);
1748 ret = io_req_prep_async(req);
1751 io_req_defer_failed(req, ret);
1754 io_prep_async_link(req);
1755 de = kmalloc(sizeof(*de), GFP_KERNEL);
1761 spin_lock(&ctx->completion_lock);
1762 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1763 spin_unlock(&ctx->completion_lock);
1768 trace_io_uring_defer(req);
1771 list_add_tail(&de->list, &ctx->defer_list);
1772 spin_unlock(&ctx->completion_lock);
1775 static void io_clean_op(struct io_kiocb *req)
1777 if (req->flags & REQ_F_BUFFER_SELECTED) {
1778 spin_lock(&req->ctx->completion_lock);
1779 io_put_kbuf_comp(req);
1780 spin_unlock(&req->ctx->completion_lock);
1783 if (req->flags & REQ_F_NEED_CLEANUP) {
1784 const struct io_op_def *def = &io_op_defs[req->opcode];
1789 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1790 kfree(req->apoll->double_poll);
1794 if (req->flags & REQ_F_INFLIGHT) {
1795 struct io_uring_task *tctx = req->task->io_uring;
1797 atomic_dec(&tctx->inflight_tracked);
1799 if (req->flags & REQ_F_CREDS)
1800 put_cred(req->creds);
1801 if (req->flags & REQ_F_ASYNC_DATA) {
1802 kfree(req->async_data);
1803 req->async_data = NULL;
1805 req->flags &= ~IO_REQ_CLEAN_FLAGS;
1808 static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags)
1810 if (req->file || !io_op_defs[req->opcode].needs_file)
1813 if (req->flags & REQ_F_FIXED_FILE)
1814 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1816 req->file = io_file_get_normal(req, req->cqe.fd);
1821 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1823 const struct io_op_def *def = &io_op_defs[req->opcode];
1824 const struct cred *creds = NULL;
1827 if (unlikely(!io_assign_file(req, issue_flags)))
1830 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1831 creds = override_creds(req->creds);
1833 if (!def->audit_skip)
1834 audit_uring_entry(req->opcode);
1836 ret = def->issue(req, issue_flags);
1838 if (!def->audit_skip)
1839 audit_uring_exit(!ret, ret);
1842 revert_creds(creds);
1844 if (ret == IOU_OK) {
1845 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1846 io_req_complete_defer(req);
1848 io_req_complete_post(req, issue_flags);
1849 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1852 /* If the op doesn't have a file, we're not polling for it */
1853 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
1854 io_iopoll_req_issued(req, issue_flags);
1859 int io_poll_issue(struct io_kiocb *req, bool *locked)
1861 io_tw_lock(req->ctx, locked);
1862 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT|
1863 IO_URING_F_COMPLETE_DEFER);
1866 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1868 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1870 req = io_put_req_find_next(req);
1871 return req ? &req->work : NULL;
1874 void io_wq_submit_work(struct io_wq_work *work)
1876 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1877 const struct io_op_def *def = &io_op_defs[req->opcode];
1878 unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ;
1879 bool needs_poll = false;
1880 int ret = 0, err = -ECANCELED;
1882 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1883 if (!(req->flags & REQ_F_REFCOUNT))
1884 __io_req_set_refcount(req, 2);
1888 io_arm_ltimeout(req);
1890 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1891 if (work->flags & IO_WQ_WORK_CANCEL) {
1893 io_req_task_queue_fail(req, err);
1896 if (!io_assign_file(req, issue_flags)) {
1898 work->flags |= IO_WQ_WORK_CANCEL;
1902 if (req->flags & REQ_F_FORCE_ASYNC) {
1903 bool opcode_poll = def->pollin || def->pollout;
1905 if (opcode_poll && file_can_poll(req->file)) {
1907 issue_flags |= IO_URING_F_NONBLOCK;
1912 ret = io_issue_sqe(req, issue_flags);
1916 * We can get EAGAIN for iopolled IO even though we're
1917 * forcing a sync submission from here, since we can't
1918 * wait for request slots on the block side.
1921 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1927 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1929 /* aborted or ready, in either case retry blocking */
1931 issue_flags &= ~IO_URING_F_NONBLOCK;
1934 /* avoid locking problems by failing it from a clean context */
1936 io_req_task_queue_fail(req, ret);
1939 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1940 unsigned int issue_flags)
1942 struct io_ring_ctx *ctx = req->ctx;
1943 struct file *file = NULL;
1944 unsigned long file_ptr;
1946 io_ring_submit_lock(ctx, issue_flags);
1948 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
1950 fd = array_index_nospec(fd, ctx->nr_user_files);
1951 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
1952 file = (struct file *) (file_ptr & FFS_MASK);
1953 file_ptr &= ~FFS_MASK;
1954 /* mask in overlapping REQ_F and FFS bits */
1955 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
1956 io_req_set_rsrc_node(req, ctx, 0);
1958 io_ring_submit_unlock(ctx, issue_flags);
1962 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
1964 struct file *file = fget(fd);
1966 trace_io_uring_file_get(req, fd);
1968 /* we don't allow fixed io_uring files */
1969 if (file && io_is_uring_fops(file))
1970 io_req_track_inflight(req);
1974 static void io_queue_async(struct io_kiocb *req, int ret)
1975 __must_hold(&req->ctx->uring_lock)
1977 struct io_kiocb *linked_timeout;
1979 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
1980 io_req_defer_failed(req, ret);
1984 linked_timeout = io_prep_linked_timeout(req);
1986 switch (io_arm_poll_handler(req, 0)) {
1987 case IO_APOLL_READY:
1988 io_kbuf_recycle(req, 0);
1989 io_req_task_queue(req);
1991 case IO_APOLL_ABORTED:
1992 io_kbuf_recycle(req, 0);
1993 io_queue_iowq(req, NULL);
2000 io_queue_linked_timeout(linked_timeout);
2003 static inline void io_queue_sqe(struct io_kiocb *req)
2004 __must_hold(&req->ctx->uring_lock)
2008 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
2011 * We async punt it if the file wasn't marked NOWAIT, or if the file
2012 * doesn't support non-blocking read/write attempts
2015 io_arm_ltimeout(req);
2017 io_queue_async(req, ret);
2020 static void io_queue_sqe_fallback(struct io_kiocb *req)
2021 __must_hold(&req->ctx->uring_lock)
2023 if (unlikely(req->flags & REQ_F_FAIL)) {
2025 * We don't submit, fail them all, for that replace hardlinks
2026 * with normal links. Extra REQ_F_LINK is tolerated.
2028 req->flags &= ~REQ_F_HARDLINK;
2029 req->flags |= REQ_F_LINK;
2030 io_req_defer_failed(req, req->cqe.res);
2031 } else if (unlikely(req->ctx->drain_active)) {
2034 int ret = io_req_prep_async(req);
2037 io_req_defer_failed(req, ret);
2039 io_queue_iowq(req, NULL);
2044 * Check SQE restrictions (opcode and flags).
2046 * Returns 'true' if SQE is allowed, 'false' otherwise.
2048 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
2049 struct io_kiocb *req,
2050 unsigned int sqe_flags)
2052 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
2055 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
2056 ctx->restrictions.sqe_flags_required)
2059 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
2060 ctx->restrictions.sqe_flags_required))
2066 static void io_init_req_drain(struct io_kiocb *req)
2068 struct io_ring_ctx *ctx = req->ctx;
2069 struct io_kiocb *head = ctx->submit_state.link.head;
2071 ctx->drain_active = true;
2074 * If we need to drain a request in the middle of a link, drain
2075 * the head request and the next request/link after the current
2076 * link. Considering sequential execution of links,
2077 * REQ_F_IO_DRAIN will be maintained for every request of our
2080 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2081 ctx->drain_next = true;
2085 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
2086 const struct io_uring_sqe *sqe)
2087 __must_hold(&ctx->uring_lock)
2089 const struct io_op_def *def;
2090 unsigned int sqe_flags;
2094 /* req is partially pre-initialised, see io_preinit_req() */
2095 req->opcode = opcode = READ_ONCE(sqe->opcode);
2096 /* same numerical values with corresponding REQ_F_*, safe to copy */
2097 req->flags = sqe_flags = READ_ONCE(sqe->flags);
2098 req->cqe.user_data = READ_ONCE(sqe->user_data);
2100 req->rsrc_node = NULL;
2101 req->task = current;
2103 if (unlikely(opcode >= IORING_OP_LAST)) {
2107 def = &io_op_defs[opcode];
2108 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2109 /* enforce forwards compatibility on users */
2110 if (sqe_flags & ~SQE_VALID_FLAGS)
2112 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2113 if (!def->buffer_select)
2115 req->buf_index = READ_ONCE(sqe->buf_group);
2117 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2118 ctx->drain_disabled = true;
2119 if (sqe_flags & IOSQE_IO_DRAIN) {
2120 if (ctx->drain_disabled)
2122 io_init_req_drain(req);
2125 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2126 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2128 /* knock it to the slow queue path, will be drained there */
2129 if (ctx->drain_active)
2130 req->flags |= REQ_F_FORCE_ASYNC;
2131 /* if there is no link, we're at "next" request and need to drain */
2132 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2133 ctx->drain_next = false;
2134 ctx->drain_active = true;
2135 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2139 if (!def->ioprio && sqe->ioprio)
2141 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2144 if (def->needs_file) {
2145 struct io_submit_state *state = &ctx->submit_state;
2147 req->cqe.fd = READ_ONCE(sqe->fd);
2150 * Plug now if we have more than 2 IO left after this, and the
2151 * target is potentially a read/write to block based storage.
2153 if (state->need_plug && def->plug) {
2154 state->plug_started = true;
2155 state->need_plug = false;
2156 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2160 personality = READ_ONCE(sqe->personality);
2164 req->creds = xa_load(&ctx->personalities, personality);
2167 get_cred(req->creds);
2168 ret = security_uring_override_creds(req->creds);
2170 put_cred(req->creds);
2173 req->flags |= REQ_F_CREDS;
2176 return def->prep(req, sqe);
2179 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2180 struct io_kiocb *req, int ret)
2182 struct io_ring_ctx *ctx = req->ctx;
2183 struct io_submit_link *link = &ctx->submit_state.link;
2184 struct io_kiocb *head = link->head;
2186 trace_io_uring_req_failed(sqe, req, ret);
2189 * Avoid breaking links in the middle as it renders links with SQPOLL
2190 * unusable. Instead of failing eagerly, continue assembling the link if
2191 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2192 * should find the flag and handle the rest.
2194 req_fail_link_node(req, ret);
2195 if (head && !(head->flags & REQ_F_FAIL))
2196 req_fail_link_node(head, -ECANCELED);
2198 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2200 link->last->link = req;
2204 io_queue_sqe_fallback(req);
2209 link->last->link = req;
2216 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2217 const struct io_uring_sqe *sqe)
2218 __must_hold(&ctx->uring_lock)
2220 struct io_submit_link *link = &ctx->submit_state.link;
2223 ret = io_init_req(ctx, req, sqe);
2225 return io_submit_fail_init(sqe, req, ret);
2227 /* don't need @sqe from now on */
2228 trace_io_uring_submit_sqe(req, true);
2231 * If we already have a head request, queue this one for async
2232 * submittal once the head completes. If we don't have a head but
2233 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2234 * submitted sync once the chain is complete. If none of those
2235 * conditions are true (normal request), then just queue it.
2237 if (unlikely(link->head)) {
2238 ret = io_req_prep_async(req);
2240 return io_submit_fail_init(sqe, req, ret);
2242 trace_io_uring_link(req, link->head);
2243 link->last->link = req;
2246 if (req->flags & IO_REQ_LINK_FLAGS)
2248 /* last request of the link, flush it */
2251 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2254 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2255 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2256 if (req->flags & IO_REQ_LINK_FLAGS) {
2261 io_queue_sqe_fallback(req);
2271 * Batched submission is done, ensure local IO is flushed out.
2273 static void io_submit_state_end(struct io_ring_ctx *ctx)
2275 struct io_submit_state *state = &ctx->submit_state;
2277 if (unlikely(state->link.head))
2278 io_queue_sqe_fallback(state->link.head);
2279 /* flush only after queuing links as they can generate completions */
2280 io_submit_flush_completions(ctx);
2281 if (state->plug_started)
2282 blk_finish_plug(&state->plug);
2286 * Start submission side cache.
2288 static void io_submit_state_start(struct io_submit_state *state,
2289 unsigned int max_ios)
2291 state->plug_started = false;
2292 state->need_plug = max_ios > 2;
2293 state->submit_nr = max_ios;
2294 /* set only head, no need to init link_last in advance */
2295 state->link.head = NULL;
2298 static void io_commit_sqring(struct io_ring_ctx *ctx)
2300 struct io_rings *rings = ctx->rings;
2303 * Ensure any loads from the SQEs are done at this point,
2304 * since once we write the new head, the application could
2305 * write new data to them.
2307 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2311 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2312 * that is mapped by userspace. This means that care needs to be taken to
2313 * ensure that reads are stable, as we cannot rely on userspace always
2314 * being a good citizen. If members of the sqe are validated and then later
2315 * used, it's important that those reads are done through READ_ONCE() to
2316 * prevent a re-load down the line.
2318 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
2320 unsigned head, mask = ctx->sq_entries - 1;
2321 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2324 * The cached sq head (or cq tail) serves two purposes:
2326 * 1) allows us to batch the cost of updating the user visible
2328 * 2) allows the kernel side to track the head on its own, even
2329 * though the application is the one updating it.
2331 head = READ_ONCE(ctx->sq_array[sq_idx]);
2332 if (likely(head < ctx->sq_entries)) {
2333 /* double index for 128-byte SQEs, twice as long */
2334 if (ctx->flags & IORING_SETUP_SQE128)
2336 return &ctx->sq_sqes[head];
2339 /* drop invalid entries */
2341 WRITE_ONCE(ctx->rings->sq_dropped,
2342 READ_ONCE(ctx->rings->sq_dropped) + 1);
2346 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2347 __must_hold(&ctx->uring_lock)
2349 unsigned int entries = io_sqring_entries(ctx);
2353 if (unlikely(!entries))
2355 /* make sure SQ entry isn't read before tail */
2356 ret = left = min3(nr, ctx->sq_entries, entries);
2357 io_get_task_refs(left);
2358 io_submit_state_start(&ctx->submit_state, left);
2361 const struct io_uring_sqe *sqe;
2362 struct io_kiocb *req;
2364 if (unlikely(!io_alloc_req_refill(ctx)))
2366 req = io_alloc_req(ctx);
2367 sqe = io_get_sqe(ctx);
2368 if (unlikely(!sqe)) {
2369 io_req_add_to_cache(req, ctx);
2374 * Continue submitting even for sqe failure if the
2375 * ring was setup with IORING_SETUP_SUBMIT_ALL
2377 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2378 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2384 if (unlikely(left)) {
2386 /* try again if it submitted nothing and can't allocate a req */
2387 if (!ret && io_req_cache_empty(ctx))
2389 current->io_uring->cached_refs += left;
2392 io_submit_state_end(ctx);
2393 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2394 io_commit_sqring(ctx);
2398 struct io_wait_queue {
2399 struct wait_queue_entry wq;
2400 struct io_ring_ctx *ctx;
2402 unsigned nr_timeouts;
2405 static inline bool io_has_work(struct io_ring_ctx *ctx)
2407 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
2408 ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
2409 !llist_empty(&ctx->work_llist));
2412 static inline bool io_should_wake(struct io_wait_queue *iowq)
2414 struct io_ring_ctx *ctx = iowq->ctx;
2415 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
2418 * Wake up if we have enough events, or if a timeout occurred since we
2419 * started waiting. For timeouts, we always want to return to userspace,
2420 * regardless of event count.
2422 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2425 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2426 int wake_flags, void *key)
2428 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
2430 struct io_ring_ctx *ctx = iowq->ctx;
2433 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2434 * the task, and the next invocation will do it.
2436 if (io_should_wake(iowq) || io_has_work(ctx))
2437 return autoremove_wake_function(curr, mode, wake_flags, key);
2441 int io_run_task_work_sig(struct io_ring_ctx *ctx)
2443 if (io_run_task_work_ctx(ctx) > 0)
2445 if (task_sigpending(current))
2450 /* when returns >0, the caller should retry */
2451 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2452 struct io_wait_queue *iowq,
2456 unsigned long check_cq;
2458 /* make sure we run task_work before checking for signals */
2459 ret = io_run_task_work_sig(ctx);
2460 if (ret || io_should_wake(iowq))
2463 check_cq = READ_ONCE(ctx->check_cq);
2464 if (unlikely(check_cq)) {
2465 /* let the caller flush overflows, retry */
2466 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2468 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
2471 if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS))
2477 * Wait until events become available, if we don't already have some. The
2478 * application must reap them itself, as they reside on the shared cq ring.
2480 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2481 const sigset_t __user *sig, size_t sigsz,
2482 struct __kernel_timespec __user *uts)
2484 struct io_wait_queue iowq;
2485 struct io_rings *rings = ctx->rings;
2486 ktime_t timeout = KTIME_MAX;
2489 if (!io_allowed_run_tw(ctx))
2493 /* always run at least 1 task work to process local work */
2494 ret = io_run_task_work_ctx(ctx);
2497 io_cqring_overflow_flush(ctx);
2499 /* if user messes with these they will just get an early return */
2500 if (__io_cqring_events_user(ctx) >= min_events)
2505 #ifdef CONFIG_COMPAT
2506 if (in_compat_syscall())
2507 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2511 ret = set_user_sigmask(sig, sigsz);
2518 struct timespec64 ts;
2520 if (get_timespec64(&ts, uts))
2522 timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2525 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2526 iowq.wq.private = current;
2527 INIT_LIST_HEAD(&iowq.wq.entry);
2529 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2530 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2532 trace_io_uring_cqring_wait(ctx, min_events);
2534 io_cqring_overflow_flush(ctx);
2535 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2536 TASK_INTERRUPTIBLE);
2537 ret = io_cqring_wait_schedule(ctx, &iowq, timeout);
2541 finish_wait(&ctx->cq_wait, &iowq.wq);
2542 restore_saved_sigmask_unless(ret == -EINTR);
2544 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2547 static void io_mem_free(void *ptr)
2554 page = virt_to_head_page(ptr);
2555 if (put_page_testzero(page))
2556 free_compound_page(page);
2559 static void *io_mem_alloc(size_t size)
2561 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2563 return (void *) __get_free_pages(gfp, get_order(size));
2566 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2567 unsigned int cq_entries, size_t *sq_offset)
2569 struct io_rings *rings;
2570 size_t off, sq_array_size;
2572 off = struct_size(rings, cqes, cq_entries);
2573 if (off == SIZE_MAX)
2575 if (ctx->flags & IORING_SETUP_CQE32) {
2576 if (check_shl_overflow(off, 1, &off))
2581 off = ALIGN(off, SMP_CACHE_BYTES);
2589 sq_array_size = array_size(sizeof(u32), sq_entries);
2590 if (sq_array_size == SIZE_MAX)
2593 if (check_add_overflow(off, sq_array_size, &off))
2599 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2600 unsigned int eventfd_async)
2602 struct io_ev_fd *ev_fd;
2603 __s32 __user *fds = arg;
2606 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2607 lockdep_is_held(&ctx->uring_lock));
2611 if (copy_from_user(&fd, fds, sizeof(*fds)))
2614 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2618 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2619 if (IS_ERR(ev_fd->cq_ev_fd)) {
2620 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2625 spin_lock(&ctx->completion_lock);
2626 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2627 spin_unlock(&ctx->completion_lock);
2629 ev_fd->eventfd_async = eventfd_async;
2630 ctx->has_evfd = true;
2631 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2632 atomic_set(&ev_fd->refs, 1);
2633 atomic_set(&ev_fd->ops, 0);
2637 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2639 struct io_ev_fd *ev_fd;
2641 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2642 lockdep_is_held(&ctx->uring_lock));
2644 ctx->has_evfd = false;
2645 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2646 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops))
2647 call_rcu(&ev_fd->rcu, io_eventfd_ops);
2654 static void io_req_caches_free(struct io_ring_ctx *ctx)
2658 mutex_lock(&ctx->uring_lock);
2659 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
2661 while (!io_req_cache_empty(ctx)) {
2662 struct io_kiocb *req = io_alloc_req(ctx);
2664 kmem_cache_free(req_cachep, req);
2668 percpu_ref_put_many(&ctx->refs, nr);
2669 mutex_unlock(&ctx->uring_lock);
2672 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2674 io_sq_thread_finish(ctx);
2675 io_rsrc_refs_drop(ctx);
2676 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2677 io_wait_rsrc_data(ctx->buf_data);
2678 io_wait_rsrc_data(ctx->file_data);
2680 mutex_lock(&ctx->uring_lock);
2682 __io_sqe_buffers_unregister(ctx);
2684 __io_sqe_files_unregister(ctx);
2685 io_cqring_overflow_kill(ctx);
2686 io_eventfd_unregister(ctx);
2687 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2688 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2689 mutex_unlock(&ctx->uring_lock);
2690 io_destroy_buffers(ctx);
2692 put_cred(ctx->sq_creds);
2693 if (ctx->submitter_task)
2694 put_task_struct(ctx->submitter_task);
2696 /* there are no registered resources left, nobody uses it */
2698 io_rsrc_node_destroy(ctx->rsrc_node);
2699 if (ctx->rsrc_backup_node)
2700 io_rsrc_node_destroy(ctx->rsrc_backup_node);
2701 flush_delayed_work(&ctx->rsrc_put_work);
2702 flush_delayed_work(&ctx->fallback_work);
2704 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2705 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
2707 #if defined(CONFIG_UNIX)
2708 if (ctx->ring_sock) {
2709 ctx->ring_sock->file = NULL; /* so that iput() is called */
2710 sock_release(ctx->ring_sock);
2713 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2715 if (ctx->mm_account) {
2716 mmdrop(ctx->mm_account);
2717 ctx->mm_account = NULL;
2719 io_mem_free(ctx->rings);
2720 io_mem_free(ctx->sq_sqes);
2722 percpu_ref_exit(&ctx->refs);
2723 free_uid(ctx->user);
2724 io_req_caches_free(ctx);
2726 io_wq_put_hash(ctx->hash_map);
2727 kfree(ctx->cancel_table.hbs);
2728 kfree(ctx->cancel_table_locked.hbs);
2729 kfree(ctx->dummy_ubuf);
2731 xa_destroy(&ctx->io_bl_xa);
2735 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2737 struct io_ring_ctx *ctx = file->private_data;
2740 poll_wait(file, &ctx->cq_wait, wait);
2742 * synchronizes with barrier from wq_has_sleeper call in
2746 if (!io_sqring_full(ctx))
2747 mask |= EPOLLOUT | EPOLLWRNORM;
2750 * Don't flush cqring overflow list here, just do a simple check.
2751 * Otherwise there could possible be ABBA deadlock:
2754 * lock(&ctx->uring_lock);
2756 * lock(&ctx->uring_lock);
2759 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2760 * pushes them to do the flush.
2763 if (io_cqring_events(ctx) || io_has_work(ctx))
2764 mask |= EPOLLIN | EPOLLRDNORM;
2769 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2771 const struct cred *creds;
2773 creds = xa_erase(&ctx->personalities, id);
2782 struct io_tctx_exit {
2783 struct callback_head task_work;
2784 struct completion completion;
2785 struct io_ring_ctx *ctx;
2788 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2790 struct io_uring_task *tctx = current->io_uring;
2791 struct io_tctx_exit *work;
2793 work = container_of(cb, struct io_tctx_exit, task_work);
2795 * When @in_idle, we're in cancellation and it's racy to remove the
2796 * node. It'll be removed by the end of cancellation, just ignore it.
2797 * tctx can be NULL if the queueing of this task_work raced with
2798 * work cancelation off the exec path.
2800 if (tctx && !atomic_read(&tctx->in_idle))
2801 io_uring_del_tctx_node((unsigned long)work->ctx);
2802 complete(&work->completion);
2805 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2807 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2809 return req->ctx == data;
2812 static __cold void io_ring_exit_work(struct work_struct *work)
2814 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2815 unsigned long timeout = jiffies + HZ * 60 * 5;
2816 unsigned long interval = HZ / 20;
2817 struct io_tctx_exit exit;
2818 struct io_tctx_node *node;
2822 * If we're doing polled IO and end up having requests being
2823 * submitted async (out-of-line), then completions can come in while
2824 * we're waiting for refs to drop. We need to reap these manually,
2825 * as nobody else will be looking for them.
2828 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
2829 mutex_lock(&ctx->uring_lock);
2830 io_cqring_overflow_kill(ctx);
2831 mutex_unlock(&ctx->uring_lock);
2834 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
2835 io_move_task_work_from_local(ctx);
2837 while (io_uring_try_cancel_requests(ctx, NULL, true))
2841 struct io_sq_data *sqd = ctx->sq_data;
2842 struct task_struct *tsk;
2844 io_sq_thread_park(sqd);
2846 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
2847 io_wq_cancel_cb(tsk->io_uring->io_wq,
2848 io_cancel_ctx_cb, ctx, true);
2849 io_sq_thread_unpark(sqd);
2852 io_req_caches_free(ctx);
2854 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
2855 /* there is little hope left, don't run it too often */
2858 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
2860 init_completion(&exit.completion);
2861 init_task_work(&exit.task_work, io_tctx_exit_cb);
2864 * Some may use context even when all refs and requests have been put,
2865 * and they are free to do so while still holding uring_lock or
2866 * completion_lock, see io_req_task_submit(). Apart from other work,
2867 * this lock/unlock section also waits them to finish.
2869 mutex_lock(&ctx->uring_lock);
2870 while (!list_empty(&ctx->tctx_list)) {
2871 WARN_ON_ONCE(time_after(jiffies, timeout));
2873 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
2875 /* don't spin on a single task if cancellation failed */
2876 list_rotate_left(&ctx->tctx_list);
2877 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
2878 if (WARN_ON_ONCE(ret))
2881 mutex_unlock(&ctx->uring_lock);
2882 wait_for_completion(&exit.completion);
2883 mutex_lock(&ctx->uring_lock);
2885 mutex_unlock(&ctx->uring_lock);
2886 spin_lock(&ctx->completion_lock);
2887 spin_unlock(&ctx->completion_lock);
2889 io_ring_ctx_free(ctx);
2892 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
2894 unsigned long index;
2895 struct creds *creds;
2897 mutex_lock(&ctx->uring_lock);
2898 percpu_ref_kill(&ctx->refs);
2899 xa_for_each(&ctx->personalities, index, creds)
2900 io_unregister_personality(ctx, index);
2902 io_poll_remove_all(ctx, NULL, true);
2903 mutex_unlock(&ctx->uring_lock);
2906 * If we failed setting up the ctx, we might not have any rings
2907 * and therefore did not submit any requests
2910 io_kill_timeouts(ctx, NULL, true);
2912 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
2914 * Use system_unbound_wq to avoid spawning tons of event kworkers
2915 * if we're exiting a ton of rings at the same time. It just adds
2916 * noise and overhead, there's no discernable change in runtime
2917 * over using system_wq.
2919 queue_work(system_unbound_wq, &ctx->exit_work);
2922 static int io_uring_release(struct inode *inode, struct file *file)
2924 struct io_ring_ctx *ctx = file->private_data;
2926 file->private_data = NULL;
2927 io_ring_ctx_wait_and_kill(ctx);
2931 struct io_task_cancel {
2932 struct task_struct *task;
2936 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
2938 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2939 struct io_task_cancel *cancel = data;
2941 return io_match_task_safe(req, cancel->task, cancel->all);
2944 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
2945 struct task_struct *task,
2948 struct io_defer_entry *de;
2951 spin_lock(&ctx->completion_lock);
2952 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
2953 if (io_match_task_safe(de->req, task, cancel_all)) {
2954 list_cut_position(&list, &ctx->defer_list, &de->list);
2958 spin_unlock(&ctx->completion_lock);
2959 if (list_empty(&list))
2962 while (!list_empty(&list)) {
2963 de = list_first_entry(&list, struct io_defer_entry, list);
2964 list_del_init(&de->list);
2965 io_req_task_queue_fail(de->req, -ECANCELED);
2971 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
2973 struct io_tctx_node *node;
2974 enum io_wq_cancel cret;
2977 mutex_lock(&ctx->uring_lock);
2978 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
2979 struct io_uring_task *tctx = node->task->io_uring;
2982 * io_wq will stay alive while we hold uring_lock, because it's
2983 * killed after ctx nodes, which requires to take the lock.
2985 if (!tctx || !tctx->io_wq)
2987 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
2988 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2990 mutex_unlock(&ctx->uring_lock);
2995 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
2996 struct task_struct *task,
2999 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
3000 struct io_uring_task *tctx = task ? task->io_uring : NULL;
3001 enum io_wq_cancel cret;
3004 /* failed during ring init, it couldn't have issued any requests */
3009 ret |= io_uring_try_cancel_iowq(ctx);
3010 } else if (tctx && tctx->io_wq) {
3012 * Cancels requests of all rings, not only @ctx, but
3013 * it's fine as the task is in exit/exec.
3015 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
3017 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3020 /* SQPOLL thread does its own polling */
3021 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
3022 (ctx->sq_data && ctx->sq_data->thread == current)) {
3023 while (!wq_list_empty(&ctx->iopoll_list)) {
3024 io_iopoll_try_reap_events(ctx);
3029 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3030 ret |= io_run_local_work(ctx) > 0;
3031 ret |= io_cancel_defer_files(ctx, task, cancel_all);
3032 mutex_lock(&ctx->uring_lock);
3033 ret |= io_poll_remove_all(ctx, task, cancel_all);
3034 mutex_unlock(&ctx->uring_lock);
3035 ret |= io_kill_timeouts(ctx, task, cancel_all);
3037 ret |= io_run_task_work() > 0;
3041 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
3044 return atomic_read(&tctx->inflight_tracked);
3045 return percpu_counter_sum(&tctx->inflight);
3049 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3050 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3052 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
3054 struct io_uring_task *tctx = current->io_uring;
3055 struct io_ring_ctx *ctx;
3059 WARN_ON_ONCE(sqd && sqd->thread != current);
3061 if (!current->io_uring)
3064 io_wq_exit_start(tctx->io_wq);
3066 atomic_inc(&tctx->in_idle);
3070 io_uring_drop_tctx_refs(current);
3071 /* read completions before cancelations */
3072 inflight = tctx_inflight(tctx, !cancel_all);
3077 struct io_tctx_node *node;
3078 unsigned long index;
3080 xa_for_each(&tctx->xa, index, node) {
3081 /* sqpoll task will cancel all its requests */
3082 if (node->ctx->sq_data)
3084 loop |= io_uring_try_cancel_requests(node->ctx,
3085 current, cancel_all);
3088 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3089 loop |= io_uring_try_cancel_requests(ctx,
3099 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3101 io_uring_drop_tctx_refs(current);
3104 * If we've seen completions, retry without waiting. This
3105 * avoids a race where a completion comes in before we did
3106 * prepare_to_wait().
3108 if (inflight == tctx_inflight(tctx, !cancel_all))
3110 finish_wait(&tctx->wait, &wait);
3113 io_uring_clean_tctx(tctx);
3116 * We shouldn't run task_works after cancel, so just leave
3117 * ->in_idle set for normal exit.
3119 atomic_dec(&tctx->in_idle);
3120 /* for exec all current's requests should be gone, kill tctx */
3121 __io_uring_free(current);
3125 void __io_uring_cancel(bool cancel_all)
3127 io_uring_cancel_generic(cancel_all, NULL);
3130 static void *io_uring_validate_mmap_request(struct file *file,
3131 loff_t pgoff, size_t sz)
3133 struct io_ring_ctx *ctx = file->private_data;
3134 loff_t offset = pgoff << PAGE_SHIFT;
3139 case IORING_OFF_SQ_RING:
3140 case IORING_OFF_CQ_RING:
3143 case IORING_OFF_SQES:
3147 return ERR_PTR(-EINVAL);
3150 page = virt_to_head_page(ptr);
3151 if (sz > page_size(page))
3152 return ERR_PTR(-EINVAL);
3159 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3161 size_t sz = vma->vm_end - vma->vm_start;
3165 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3167 return PTR_ERR(ptr);
3169 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3170 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3173 #else /* !CONFIG_MMU */
3175 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3177 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
3180 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3182 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3185 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3186 unsigned long addr, unsigned long len,
3187 unsigned long pgoff, unsigned long flags)
3191 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3193 return PTR_ERR(ptr);
3195 return (unsigned long) ptr;
3198 #endif /* !CONFIG_MMU */
3200 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3202 if (flags & IORING_ENTER_EXT_ARG) {
3203 struct io_uring_getevents_arg arg;
3205 if (argsz != sizeof(arg))
3207 if (copy_from_user(&arg, argp, sizeof(arg)))
3213 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3214 struct __kernel_timespec __user **ts,
3215 const sigset_t __user **sig)
3217 struct io_uring_getevents_arg arg;
3220 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3221 * is just a pointer to the sigset_t.
3223 if (!(flags & IORING_ENTER_EXT_ARG)) {
3224 *sig = (const sigset_t __user *) argp;
3230 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3231 * timespec and sigset_t pointers if good.
3233 if (*argsz != sizeof(arg))
3235 if (copy_from_user(&arg, argp, sizeof(arg)))
3239 *sig = u64_to_user_ptr(arg.sigmask);
3240 *argsz = arg.sigmask_sz;
3241 *ts = u64_to_user_ptr(arg.ts);
3245 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3246 u32, min_complete, u32, flags, const void __user *, argp,
3249 struct io_ring_ctx *ctx;
3253 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3254 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3255 IORING_ENTER_REGISTERED_RING)))
3259 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3260 * need only dereference our task private array to find it.
3262 if (flags & IORING_ENTER_REGISTERED_RING) {
3263 struct io_uring_task *tctx = current->io_uring;
3265 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3267 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3268 f.file = tctx->registered_rings[fd];
3270 if (unlikely(!f.file))
3274 if (unlikely(!f.file))
3277 if (unlikely(!io_is_uring_fops(f.file)))
3281 ctx = f.file->private_data;
3283 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3287 * For SQ polling, the thread will do all submissions and completions.
3288 * Just return the requested submit count, and wake the thread if
3292 if (ctx->flags & IORING_SETUP_SQPOLL) {
3293 io_cqring_overflow_flush(ctx);
3295 if (unlikely(ctx->sq_data->thread == NULL)) {
3299 if (flags & IORING_ENTER_SQ_WAKEUP)
3300 wake_up(&ctx->sq_data->wait);
3301 if (flags & IORING_ENTER_SQ_WAIT) {
3302 ret = io_sqpoll_wait_sq(ctx);
3307 } else if (to_submit) {
3308 ret = io_uring_add_tctx_node(ctx);
3312 mutex_lock(&ctx->uring_lock);
3313 ret = io_submit_sqes(ctx, to_submit);
3314 if (ret != to_submit) {
3315 mutex_unlock(&ctx->uring_lock);
3318 if (flags & IORING_ENTER_GETEVENTS) {
3319 if (ctx->syscall_iopoll)
3322 * Ignore errors, we'll soon call io_cqring_wait() and
3323 * it should handle ownership problems if any.
3325 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3326 (void)io_run_local_work_locked(ctx);
3328 mutex_unlock(&ctx->uring_lock);
3331 if (flags & IORING_ENTER_GETEVENTS) {
3334 if (ctx->syscall_iopoll) {
3336 * We disallow the app entering submit/complete with
3337 * polling, but we still need to lock the ring to
3338 * prevent racing with polled issue that got punted to
3341 mutex_lock(&ctx->uring_lock);
3343 ret2 = io_validate_ext_arg(flags, argp, argsz);
3344 if (likely(!ret2)) {
3345 min_complete = min(min_complete,
3347 ret2 = io_iopoll_check(ctx, min_complete);
3349 mutex_unlock(&ctx->uring_lock);
3351 const sigset_t __user *sig;
3352 struct __kernel_timespec __user *ts;
3354 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3355 if (likely(!ret2)) {
3356 min_complete = min(min_complete,
3358 ret2 = io_cqring_wait(ctx, min_complete, sig,
3367 * EBADR indicates that one or more CQE were dropped.
3368 * Once the user has been informed we can clear the bit
3369 * as they are obviously ok with those drops.
3371 if (unlikely(ret2 == -EBADR))
3372 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3381 static const struct file_operations io_uring_fops = {
3382 .release = io_uring_release,
3383 .mmap = io_uring_mmap,
3385 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3386 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3388 .poll = io_uring_poll,
3389 #ifdef CONFIG_PROC_FS
3390 .show_fdinfo = io_uring_show_fdinfo,
3394 bool io_is_uring_fops(struct file *file)
3396 return file->f_op == &io_uring_fops;
3399 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3400 struct io_uring_params *p)
3402 struct io_rings *rings;
3403 size_t size, sq_array_offset;
3405 /* make sure these are sane, as we already accounted them */
3406 ctx->sq_entries = p->sq_entries;
3407 ctx->cq_entries = p->cq_entries;
3409 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3410 if (size == SIZE_MAX)
3413 rings = io_mem_alloc(size);
3418 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3419 rings->sq_ring_mask = p->sq_entries - 1;
3420 rings->cq_ring_mask = p->cq_entries - 1;
3421 rings->sq_ring_entries = p->sq_entries;
3422 rings->cq_ring_entries = p->cq_entries;
3424 if (p->flags & IORING_SETUP_SQE128)
3425 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3427 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3428 if (size == SIZE_MAX) {
3429 io_mem_free(ctx->rings);
3434 ctx->sq_sqes = io_mem_alloc(size);
3435 if (!ctx->sq_sqes) {
3436 io_mem_free(ctx->rings);
3444 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3448 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3452 ret = __io_uring_add_tctx_node(ctx);
3457 fd_install(fd, file);
3462 * Allocate an anonymous fd, this is what constitutes the application
3463 * visible backing of an io_uring instance. The application mmaps this
3464 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3465 * we have to tie this fd to a socket for file garbage collection purposes.
3467 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3470 #if defined(CONFIG_UNIX)
3473 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3476 return ERR_PTR(ret);
3479 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3480 O_RDWR | O_CLOEXEC, NULL);
3481 #if defined(CONFIG_UNIX)
3483 sock_release(ctx->ring_sock);
3484 ctx->ring_sock = NULL;
3486 ctx->ring_sock->file = file;
3492 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3493 struct io_uring_params __user *params)
3495 struct io_ring_ctx *ctx;
3501 if (entries > IORING_MAX_ENTRIES) {
3502 if (!(p->flags & IORING_SETUP_CLAMP))
3504 entries = IORING_MAX_ENTRIES;
3508 * Use twice as many entries for the CQ ring. It's possible for the
3509 * application to drive a higher depth than the size of the SQ ring,
3510 * since the sqes are only used at submission time. This allows for
3511 * some flexibility in overcommitting a bit. If the application has
3512 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3513 * of CQ ring entries manually.
3515 p->sq_entries = roundup_pow_of_two(entries);
3516 if (p->flags & IORING_SETUP_CQSIZE) {
3518 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3519 * to a power-of-two, if it isn't already. We do NOT impose
3520 * any cq vs sq ring sizing.
3524 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3525 if (!(p->flags & IORING_SETUP_CLAMP))
3527 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3529 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3530 if (p->cq_entries < p->sq_entries)
3533 p->cq_entries = 2 * p->sq_entries;
3536 ctx = io_ring_ctx_alloc(p);
3540 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3541 !(ctx->flags & IORING_SETUP_IOPOLL) &&
3542 !(ctx->flags & IORING_SETUP_SQPOLL))
3543 ctx->task_complete = true;
3546 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3547 * space applications don't need to do io completion events
3548 * polling again, they can rely on io_sq_thread to do polling
3549 * work, which can reduce cpu usage and uring_lock contention.
3551 if (ctx->flags & IORING_SETUP_IOPOLL &&
3552 !(ctx->flags & IORING_SETUP_SQPOLL))
3553 ctx->syscall_iopoll = 1;
3555 ctx->compat = in_compat_syscall();
3556 if (!capable(CAP_IPC_LOCK))
3557 ctx->user = get_uid(current_user());
3560 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3561 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3564 if (ctx->flags & IORING_SETUP_SQPOLL) {
3565 /* IPI related flags don't make sense with SQPOLL */
3566 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3567 IORING_SETUP_TASKRUN_FLAG |
3568 IORING_SETUP_DEFER_TASKRUN))
3570 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3571 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3572 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3574 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3575 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3577 ctx->notify_method = TWA_SIGNAL;
3581 * For DEFER_TASKRUN we require the completion task to be the same as the
3582 * submission task. This implies that there is only one submitter, so enforce
3585 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3586 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3591 * This is just grabbed for accounting purposes. When a process exits,
3592 * the mm is exited and dropped before the files, hence we need to hang
3593 * on to this mm purely for the purposes of being able to unaccount
3594 * memory (locked/pinned vm). It's not used for anything else.
3596 mmgrab(current->mm);
3597 ctx->mm_account = current->mm;
3599 ret = io_allocate_scq_urings(ctx, p);
3603 ret = io_sq_offload_create(ctx, p);
3606 /* always set a rsrc node */
3607 ret = io_rsrc_node_switch_start(ctx);
3610 io_rsrc_node_switch(ctx, NULL);
3612 memset(&p->sq_off, 0, sizeof(p->sq_off));
3613 p->sq_off.head = offsetof(struct io_rings, sq.head);
3614 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3615 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3616 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3617 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3618 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3619 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3621 memset(&p->cq_off, 0, sizeof(p->cq_off));
3622 p->cq_off.head = offsetof(struct io_rings, cq.head);
3623 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3624 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3625 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3626 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3627 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3628 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3630 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3631 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3632 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3633 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3634 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3635 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3636 IORING_FEAT_LINKED_FILE;
3638 if (copy_to_user(params, p, sizeof(*p))) {
3643 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
3644 && !(ctx->flags & IORING_SETUP_R_DISABLED))
3645 ctx->submitter_task = get_task_struct(current);
3647 file = io_uring_get_file(ctx);
3649 ret = PTR_ERR(file);
3654 * Install ring fd as the very last thing, so we don't risk someone
3655 * having closed it before we finish setup
3657 ret = io_uring_install_fd(ctx, file);
3659 /* fput will clean it up */
3664 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3667 io_ring_ctx_wait_and_kill(ctx);
3672 * Sets up an aio uring context, and returns the fd. Applications asks for a
3673 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3674 * params structure passed in.
3676 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3678 struct io_uring_params p;
3681 if (copy_from_user(&p, params, sizeof(p)))
3683 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3688 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3689 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3690 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3691 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3692 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3693 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3694 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN))
3697 return io_uring_create(entries, &p, params);
3700 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3701 struct io_uring_params __user *, params)
3703 return io_uring_setup(entries, params);
3706 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3709 struct io_uring_probe *p;
3713 size = struct_size(p, ops, nr_args);
3714 if (size == SIZE_MAX)
3716 p = kzalloc(size, GFP_KERNEL);
3721 if (copy_from_user(p, arg, size))
3724 if (memchr_inv(p, 0, size))
3727 p->last_op = IORING_OP_LAST - 1;
3728 if (nr_args > IORING_OP_LAST)
3729 nr_args = IORING_OP_LAST;
3731 for (i = 0; i < nr_args; i++) {
3733 if (!io_op_defs[i].not_supported)
3734 p->ops[i].flags = IO_URING_OP_SUPPORTED;
3739 if (copy_to_user(arg, p, size))
3746 static int io_register_personality(struct io_ring_ctx *ctx)
3748 const struct cred *creds;
3752 creds = get_current_cred();
3754 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
3755 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
3763 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
3764 void __user *arg, unsigned int nr_args)
3766 struct io_uring_restriction *res;
3770 /* Restrictions allowed only if rings started disabled */
3771 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3774 /* We allow only a single restrictions registration */
3775 if (ctx->restrictions.registered)
3778 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
3781 size = array_size(nr_args, sizeof(*res));
3782 if (size == SIZE_MAX)
3785 res = memdup_user(arg, size);
3787 return PTR_ERR(res);
3791 for (i = 0; i < nr_args; i++) {
3792 switch (res[i].opcode) {
3793 case IORING_RESTRICTION_REGISTER_OP:
3794 if (res[i].register_op >= IORING_REGISTER_LAST) {
3799 __set_bit(res[i].register_op,
3800 ctx->restrictions.register_op);
3802 case IORING_RESTRICTION_SQE_OP:
3803 if (res[i].sqe_op >= IORING_OP_LAST) {
3808 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
3810 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
3811 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
3813 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
3814 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
3823 /* Reset all restrictions if an error happened */
3825 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
3827 ctx->restrictions.registered = true;
3833 static int io_register_enable_rings(struct io_ring_ctx *ctx)
3835 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3838 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task)
3839 ctx->submitter_task = get_task_struct(current);
3841 if (ctx->restrictions.registered)
3842 ctx->restricted = 1;
3844 ctx->flags &= ~IORING_SETUP_R_DISABLED;
3845 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
3846 wake_up(&ctx->sq_data->wait);
3850 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
3851 void __user *arg, unsigned len)
3853 struct io_uring_task *tctx = current->io_uring;
3854 cpumask_var_t new_mask;
3857 if (!tctx || !tctx->io_wq)
3860 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3863 cpumask_clear(new_mask);
3864 if (len > cpumask_size())
3865 len = cpumask_size();
3867 if (in_compat_syscall()) {
3868 ret = compat_get_bitmap(cpumask_bits(new_mask),
3869 (const compat_ulong_t __user *)arg,
3870 len * 8 /* CHAR_BIT */);
3872 ret = copy_from_user(new_mask, arg, len);
3876 free_cpumask_var(new_mask);
3880 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
3881 free_cpumask_var(new_mask);
3885 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
3887 struct io_uring_task *tctx = current->io_uring;
3889 if (!tctx || !tctx->io_wq)
3892 return io_wq_cpu_affinity(tctx->io_wq, NULL);
3895 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
3897 __must_hold(&ctx->uring_lock)
3899 struct io_tctx_node *node;
3900 struct io_uring_task *tctx = NULL;
3901 struct io_sq_data *sqd = NULL;
3905 if (copy_from_user(new_count, arg, sizeof(new_count)))
3907 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3908 if (new_count[i] > INT_MAX)
3911 if (ctx->flags & IORING_SETUP_SQPOLL) {
3915 * Observe the correct sqd->lock -> ctx->uring_lock
3916 * ordering. Fine to drop uring_lock here, we hold
3919 refcount_inc(&sqd->refs);
3920 mutex_unlock(&ctx->uring_lock);
3921 mutex_lock(&sqd->lock);
3922 mutex_lock(&ctx->uring_lock);
3924 tctx = sqd->thread->io_uring;
3927 tctx = current->io_uring;
3930 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
3932 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3934 ctx->iowq_limits[i] = new_count[i];
3935 ctx->iowq_limits_set = true;
3937 if (tctx && tctx->io_wq) {
3938 ret = io_wq_max_workers(tctx->io_wq, new_count);
3942 memset(new_count, 0, sizeof(new_count));
3946 mutex_unlock(&sqd->lock);
3947 io_put_sq_data(sqd);
3950 if (copy_to_user(arg, new_count, sizeof(new_count)))
3953 /* that's it for SQPOLL, only the SQPOLL task creates requests */
3957 /* now propagate the restriction to all registered users */
3958 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3959 struct io_uring_task *tctx = node->task->io_uring;
3961 if (WARN_ON_ONCE(!tctx->io_wq))
3964 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3965 new_count[i] = ctx->iowq_limits[i];
3966 /* ignore errors, it always returns zero anyway */
3967 (void)io_wq_max_workers(tctx->io_wq, new_count);
3972 mutex_unlock(&sqd->lock);
3973 io_put_sq_data(sqd);
3978 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
3979 void __user *arg, unsigned nr_args)
3980 __releases(ctx->uring_lock)
3981 __acquires(ctx->uring_lock)
3986 * We don't quiesce the refs for register anymore and so it can't be
3987 * dying as we're holding a file ref here.
3989 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
3992 if (ctx->submitter_task && ctx->submitter_task != current)
3995 if (ctx->restricted) {
3996 if (opcode >= IORING_REGISTER_LAST)
3998 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
3999 if (!test_bit(opcode, ctx->restrictions.register_op))
4004 case IORING_REGISTER_BUFFERS:
4008 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
4010 case IORING_UNREGISTER_BUFFERS:
4014 ret = io_sqe_buffers_unregister(ctx);
4016 case IORING_REGISTER_FILES:
4020 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
4022 case IORING_UNREGISTER_FILES:
4026 ret = io_sqe_files_unregister(ctx);
4028 case IORING_REGISTER_FILES_UPDATE:
4029 ret = io_register_files_update(ctx, arg, nr_args);
4031 case IORING_REGISTER_EVENTFD:
4035 ret = io_eventfd_register(ctx, arg, 0);
4037 case IORING_REGISTER_EVENTFD_ASYNC:
4041 ret = io_eventfd_register(ctx, arg, 1);
4043 case IORING_UNREGISTER_EVENTFD:
4047 ret = io_eventfd_unregister(ctx);
4049 case IORING_REGISTER_PROBE:
4051 if (!arg || nr_args > 256)
4053 ret = io_probe(ctx, arg, nr_args);
4055 case IORING_REGISTER_PERSONALITY:
4059 ret = io_register_personality(ctx);
4061 case IORING_UNREGISTER_PERSONALITY:
4065 ret = io_unregister_personality(ctx, nr_args);
4067 case IORING_REGISTER_ENABLE_RINGS:
4071 ret = io_register_enable_rings(ctx);
4073 case IORING_REGISTER_RESTRICTIONS:
4074 ret = io_register_restrictions(ctx, arg, nr_args);
4076 case IORING_REGISTER_FILES2:
4077 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
4079 case IORING_REGISTER_FILES_UPDATE2:
4080 ret = io_register_rsrc_update(ctx, arg, nr_args,
4083 case IORING_REGISTER_BUFFERS2:
4084 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
4086 case IORING_REGISTER_BUFFERS_UPDATE:
4087 ret = io_register_rsrc_update(ctx, arg, nr_args,
4088 IORING_RSRC_BUFFER);
4090 case IORING_REGISTER_IOWQ_AFF:
4092 if (!arg || !nr_args)
4094 ret = io_register_iowq_aff(ctx, arg, nr_args);
4096 case IORING_UNREGISTER_IOWQ_AFF:
4100 ret = io_unregister_iowq_aff(ctx);
4102 case IORING_REGISTER_IOWQ_MAX_WORKERS:
4104 if (!arg || nr_args != 2)
4106 ret = io_register_iowq_max_workers(ctx, arg);
4108 case IORING_REGISTER_RING_FDS:
4109 ret = io_ringfd_register(ctx, arg, nr_args);
4111 case IORING_UNREGISTER_RING_FDS:
4112 ret = io_ringfd_unregister(ctx, arg, nr_args);
4114 case IORING_REGISTER_PBUF_RING:
4116 if (!arg || nr_args != 1)
4118 ret = io_register_pbuf_ring(ctx, arg);
4120 case IORING_UNREGISTER_PBUF_RING:
4122 if (!arg || nr_args != 1)
4124 ret = io_unregister_pbuf_ring(ctx, arg);
4126 case IORING_REGISTER_SYNC_CANCEL:
4128 if (!arg || nr_args != 1)
4130 ret = io_sync_cancel(ctx, arg);
4132 case IORING_REGISTER_FILE_ALLOC_RANGE:
4134 if (!arg || nr_args)
4136 ret = io_register_file_alloc_range(ctx, arg);
4146 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
4147 void __user *, arg, unsigned int, nr_args)
4149 struct io_ring_ctx *ctx;
4158 if (!io_is_uring_fops(f.file))
4161 ctx = f.file->private_data;
4163 mutex_lock(&ctx->uring_lock);
4164 ret = __io_uring_register(ctx, opcode, arg, nr_args);
4165 mutex_unlock(&ctx->uring_lock);
4166 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
4172 static int __init io_uring_init(void)
4174 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4175 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4176 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4179 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4180 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4181 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4182 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4183 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4184 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
4185 BUILD_BUG_SQE_ELEM(1, __u8, flags);
4186 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
4187 BUILD_BUG_SQE_ELEM(4, __s32, fd);
4188 BUILD_BUG_SQE_ELEM(8, __u64, off);
4189 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
4190 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
4191 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
4192 BUILD_BUG_SQE_ELEM(16, __u64, addr);
4193 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
4194 BUILD_BUG_SQE_ELEM(24, __u32, len);
4195 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
4196 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
4197 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4198 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
4199 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
4200 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
4201 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
4202 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
4203 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
4204 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
4205 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
4206 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
4207 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
4208 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
4209 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
4210 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
4211 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
4212 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
4213 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
4214 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
4215 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
4216 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
4217 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
4218 BUILD_BUG_SQE_ELEM(42, __u16, personality);
4219 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
4220 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
4221 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
4222 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
4223 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
4224 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4225 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
4227 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4228 sizeof(struct io_uring_rsrc_update));
4229 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4230 sizeof(struct io_uring_rsrc_update2));
4232 /* ->buf_index is u16 */
4233 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4234 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4235 offsetof(struct io_uring_buf_ring, tail));
4237 /* should fit into one byte */
4238 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
4239 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4240 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4242 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
4244 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4246 io_uring_optable_init();
4248 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4252 __initcall(io_uring_init);