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);
153 static struct kmem_cache *req_cachep;
155 struct sock *io_uring_get_socket(struct file *file)
157 #if defined(CONFIG_UNIX)
158 if (io_is_uring_fops(file)) {
159 struct io_ring_ctx *ctx = file->private_data;
161 return ctx->ring_sock->sk;
166 EXPORT_SYMBOL(io_uring_get_socket);
168 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
170 if (!wq_list_empty(&ctx->submit_state.compl_reqs) ||
171 ctx->submit_state.cqes_count)
172 __io_submit_flush_completions(ctx);
175 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
177 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
180 static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx)
182 return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head);
185 static bool io_match_linked(struct io_kiocb *head)
187 struct io_kiocb *req;
189 io_for_each_link(req, head) {
190 if (req->flags & REQ_F_INFLIGHT)
197 * As io_match_task() but protected against racing with linked timeouts.
198 * User must not hold timeout_lock.
200 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
205 if (task && head->task != task)
210 if (head->flags & REQ_F_LINK_TIMEOUT) {
211 struct io_ring_ctx *ctx = head->ctx;
213 /* protect against races with linked timeouts */
214 spin_lock_irq(&ctx->timeout_lock);
215 matched = io_match_linked(head);
216 spin_unlock_irq(&ctx->timeout_lock);
218 matched = io_match_linked(head);
223 static inline void req_fail_link_node(struct io_kiocb *req, int res)
226 io_req_set_res(req, res, 0);
229 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
231 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
234 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
236 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
238 complete(&ctx->ref_comp);
241 static __cold void io_fallback_req_func(struct work_struct *work)
243 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
245 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
246 struct io_kiocb *req, *tmp;
249 percpu_ref_get(&ctx->refs);
250 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
251 req->io_task_work.func(req, &locked);
254 io_submit_flush_completions(ctx);
255 mutex_unlock(&ctx->uring_lock);
257 percpu_ref_put(&ctx->refs);
260 static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
262 unsigned hash_buckets = 1U << bits;
263 size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
265 table->hbs = kmalloc(hash_size, GFP_KERNEL);
269 table->hash_bits = bits;
270 init_hash_table(table, hash_buckets);
274 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
276 struct io_ring_ctx *ctx;
279 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
283 xa_init(&ctx->io_bl_xa);
286 * Use 5 bits less than the max cq entries, that should give us around
287 * 32 entries per hash list if totally full and uniformly spread, but
288 * don't keep too many buckets to not overconsume memory.
290 hash_bits = ilog2(p->cq_entries) - 5;
291 hash_bits = clamp(hash_bits, 1, 8);
292 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
294 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
297 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
298 if (!ctx->dummy_ubuf)
300 /* set invalid range, so io_import_fixed() fails meeting it */
301 ctx->dummy_ubuf->ubuf = -1UL;
303 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
307 ctx->flags = p->flags;
308 init_waitqueue_head(&ctx->sqo_sq_wait);
309 INIT_LIST_HEAD(&ctx->sqd_list);
310 INIT_LIST_HEAD(&ctx->cq_overflow_list);
311 INIT_LIST_HEAD(&ctx->io_buffers_cache);
312 io_alloc_cache_init(&ctx->apoll_cache);
313 io_alloc_cache_init(&ctx->netmsg_cache);
314 init_completion(&ctx->ref_comp);
315 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
316 mutex_init(&ctx->uring_lock);
317 init_waitqueue_head(&ctx->cq_wait);
318 spin_lock_init(&ctx->completion_lock);
319 spin_lock_init(&ctx->timeout_lock);
320 INIT_WQ_LIST(&ctx->iopoll_list);
321 INIT_LIST_HEAD(&ctx->io_buffers_pages);
322 INIT_LIST_HEAD(&ctx->io_buffers_comp);
323 INIT_LIST_HEAD(&ctx->defer_list);
324 INIT_LIST_HEAD(&ctx->timeout_list);
325 INIT_LIST_HEAD(&ctx->ltimeout_list);
326 spin_lock_init(&ctx->rsrc_ref_lock);
327 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
328 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
329 init_llist_head(&ctx->rsrc_put_llist);
330 init_llist_head(&ctx->work_llist);
331 INIT_LIST_HEAD(&ctx->tctx_list);
332 ctx->submit_state.free_list.next = NULL;
333 INIT_WQ_LIST(&ctx->locked_free_list);
334 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
335 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
338 kfree(ctx->dummy_ubuf);
339 kfree(ctx->cancel_table.hbs);
340 kfree(ctx->cancel_table_locked.hbs);
342 xa_destroy(&ctx->io_bl_xa);
347 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
349 struct io_rings *r = ctx->rings;
351 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
355 static bool req_need_defer(struct io_kiocb *req, u32 seq)
357 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
358 struct io_ring_ctx *ctx = req->ctx;
360 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
366 static inline void io_req_track_inflight(struct io_kiocb *req)
368 if (!(req->flags & REQ_F_INFLIGHT)) {
369 req->flags |= REQ_F_INFLIGHT;
370 atomic_inc(&req->task->io_uring->inflight_tracked);
374 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
376 if (WARN_ON_ONCE(!req->link))
379 req->flags &= ~REQ_F_ARM_LTIMEOUT;
380 req->flags |= REQ_F_LINK_TIMEOUT;
382 /* linked timeouts should have two refs once prep'ed */
383 io_req_set_refcount(req);
384 __io_req_set_refcount(req->link, 2);
388 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
390 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
392 return __io_prep_linked_timeout(req);
395 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
397 io_queue_linked_timeout(__io_prep_linked_timeout(req));
400 static inline void io_arm_ltimeout(struct io_kiocb *req)
402 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
403 __io_arm_ltimeout(req);
406 static void io_prep_async_work(struct io_kiocb *req)
408 const struct io_op_def *def = &io_op_defs[req->opcode];
409 struct io_ring_ctx *ctx = req->ctx;
411 if (!(req->flags & REQ_F_CREDS)) {
412 req->flags |= REQ_F_CREDS;
413 req->creds = get_current_cred();
416 req->work.list.next = NULL;
418 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
419 if (req->flags & REQ_F_FORCE_ASYNC)
420 req->work.flags |= IO_WQ_WORK_CONCURRENT;
422 if (req->file && !io_req_ffs_set(req))
423 req->flags |= io_file_get_flags(req->file) << REQ_F_SUPPORT_NOWAIT_BIT;
425 if (req->flags & REQ_F_ISREG) {
426 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
427 io_wq_hash_work(&req->work, file_inode(req->file));
428 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
429 if (def->unbound_nonreg_file)
430 req->work.flags |= IO_WQ_WORK_UNBOUND;
434 static void io_prep_async_link(struct io_kiocb *req)
436 struct io_kiocb *cur;
438 if (req->flags & REQ_F_LINK_TIMEOUT) {
439 struct io_ring_ctx *ctx = req->ctx;
441 spin_lock_irq(&ctx->timeout_lock);
442 io_for_each_link(cur, req)
443 io_prep_async_work(cur);
444 spin_unlock_irq(&ctx->timeout_lock);
446 io_for_each_link(cur, req)
447 io_prep_async_work(cur);
451 void io_queue_iowq(struct io_kiocb *req, bool *dont_use)
453 struct io_kiocb *link = io_prep_linked_timeout(req);
454 struct io_uring_task *tctx = req->task->io_uring;
457 BUG_ON(!tctx->io_wq);
459 /* init ->work of the whole link before punting */
460 io_prep_async_link(req);
463 * Not expected to happen, but if we do have a bug where this _can_
464 * happen, catch it here and ensure the request is marked as
465 * canceled. That will make io-wq go through the usual work cancel
466 * procedure rather than attempt to run this request (or create a new
469 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
470 req->work.flags |= IO_WQ_WORK_CANCEL;
472 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
473 io_wq_enqueue(tctx->io_wq, &req->work);
475 io_queue_linked_timeout(link);
478 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
480 while (!list_empty(&ctx->defer_list)) {
481 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
482 struct io_defer_entry, list);
484 if (req_need_defer(de->req, de->seq))
486 list_del_init(&de->list);
487 io_req_task_queue(de->req);
493 static void io_eventfd_ops(struct rcu_head *rcu)
495 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
496 int ops = atomic_xchg(&ev_fd->ops, 0);
498 if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT))
499 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
501 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
502 * ordering in a race but if references are 0 we know we have to free
505 if (atomic_dec_and_test(&ev_fd->refs)) {
506 eventfd_ctx_put(ev_fd->cq_ev_fd);
511 static void io_eventfd_signal(struct io_ring_ctx *ctx)
513 struct io_ev_fd *ev_fd = NULL;
517 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
520 ev_fd = rcu_dereference(ctx->io_ev_fd);
523 * Check again if ev_fd exists incase an io_eventfd_unregister call
524 * completed between the NULL check of ctx->io_ev_fd at the start of
525 * the function and rcu_read_lock.
527 if (unlikely(!ev_fd))
529 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
531 if (ev_fd->eventfd_async && !io_wq_current_is_worker())
534 if (likely(eventfd_signal_allowed())) {
535 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
537 atomic_inc(&ev_fd->refs);
538 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops))
539 call_rcu(&ev_fd->rcu, io_eventfd_ops);
541 atomic_dec(&ev_fd->refs);
548 static void io_eventfd_flush_signal(struct io_ring_ctx *ctx)
552 spin_lock(&ctx->completion_lock);
555 * Eventfd should only get triggered when at least one event has been
556 * posted. Some applications rely on the eventfd notification count
557 * only changing IFF a new CQE has been added to the CQ ring. There's
558 * no depedency on 1:1 relationship between how many times this
559 * function is called (and hence the eventfd count) and number of CQEs
560 * posted to the CQ ring.
562 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
563 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
564 spin_unlock(&ctx->completion_lock);
568 io_eventfd_signal(ctx);
571 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
573 if (ctx->off_timeout_used || ctx->drain_active) {
574 spin_lock(&ctx->completion_lock);
575 if (ctx->off_timeout_used)
576 io_flush_timeouts(ctx);
577 if (ctx->drain_active)
578 io_queue_deferred(ctx);
579 spin_unlock(&ctx->completion_lock);
582 io_eventfd_flush_signal(ctx);
585 /* keep it inlined for io_submit_flush_completions() */
586 static inline void io_cq_unlock_post_inline(struct io_ring_ctx *ctx)
587 __releases(ctx->completion_lock)
589 io_commit_cqring(ctx);
590 spin_unlock(&ctx->completion_lock);
592 io_commit_cqring_flush(ctx);
596 void io_cq_unlock_post(struct io_ring_ctx *ctx)
597 __releases(ctx->completion_lock)
599 io_cq_unlock_post_inline(ctx);
602 /* Returns true if there are no backlogged entries after the flush */
603 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
606 size_t cqe_size = sizeof(struct io_uring_cqe);
608 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
611 if (ctx->flags & IORING_SETUP_CQE32)
615 while (!list_empty(&ctx->cq_overflow_list)) {
616 struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true);
617 struct io_overflow_cqe *ocqe;
621 ocqe = list_first_entry(&ctx->cq_overflow_list,
622 struct io_overflow_cqe, list);
624 memcpy(cqe, &ocqe->cqe, cqe_size);
626 io_account_cq_overflow(ctx);
628 list_del(&ocqe->list);
632 all_flushed = list_empty(&ctx->cq_overflow_list);
634 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
635 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
638 io_cq_unlock_post(ctx);
642 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
646 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
647 /* iopoll syncs against uring_lock, not completion_lock */
648 if (ctx->flags & IORING_SETUP_IOPOLL)
649 mutex_lock(&ctx->uring_lock);
650 ret = __io_cqring_overflow_flush(ctx, false);
651 if (ctx->flags & IORING_SETUP_IOPOLL)
652 mutex_unlock(&ctx->uring_lock);
658 void __io_put_task(struct task_struct *task, int nr)
660 struct io_uring_task *tctx = task->io_uring;
662 percpu_counter_sub(&tctx->inflight, nr);
663 if (unlikely(atomic_read(&tctx->in_idle)))
664 wake_up(&tctx->wait);
665 put_task_struct_many(task, nr);
668 void io_task_refs_refill(struct io_uring_task *tctx)
670 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
672 percpu_counter_add(&tctx->inflight, refill);
673 refcount_add(refill, ¤t->usage);
674 tctx->cached_refs += refill;
677 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
679 struct io_uring_task *tctx = task->io_uring;
680 unsigned int refs = tctx->cached_refs;
683 tctx->cached_refs = 0;
684 percpu_counter_sub(&tctx->inflight, refs);
685 put_task_struct_many(task, refs);
689 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
690 s32 res, u32 cflags, u64 extra1, u64 extra2)
692 struct io_overflow_cqe *ocqe;
693 size_t ocq_size = sizeof(struct io_overflow_cqe);
694 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
697 ocq_size += sizeof(struct io_uring_cqe);
699 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
700 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
703 * If we're in ring overflow flush mode, or in task cancel mode,
704 * or cannot allocate an overflow entry, then we need to drop it
707 io_account_cq_overflow(ctx);
708 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
711 if (list_empty(&ctx->cq_overflow_list)) {
712 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
713 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
716 ocqe->cqe.user_data = user_data;
718 ocqe->cqe.flags = cflags;
720 ocqe->cqe.big_cqe[0] = extra1;
721 ocqe->cqe.big_cqe[1] = extra2;
723 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
727 bool io_req_cqe_overflow(struct io_kiocb *req)
729 if (!(req->flags & REQ_F_CQE32_INIT)) {
733 return io_cqring_event_overflow(req->ctx, req->cqe.user_data,
734 req->cqe.res, req->cqe.flags,
735 req->extra1, req->extra2);
739 * writes to the cq entry need to come after reading head; the
740 * control dependency is enough as we're using WRITE_ONCE to
743 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow)
745 struct io_rings *rings = ctx->rings;
746 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
747 unsigned int free, queued, len;
750 * Posting into the CQ when there are pending overflowed CQEs may break
751 * ordering guarantees, which will affect links, F_MORE users and more.
752 * Force overflow the completion.
754 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
757 /* userspace may cheat modifying the tail, be safe and do min */
758 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
759 free = ctx->cq_entries - queued;
760 /* we need a contiguous range, limit based on the current array offset */
761 len = min(free, ctx->cq_entries - off);
765 if (ctx->flags & IORING_SETUP_CQE32) {
770 ctx->cqe_cached = &rings->cqes[off];
771 ctx->cqe_sentinel = ctx->cqe_cached + len;
773 ctx->cached_cq_tail++;
775 if (ctx->flags & IORING_SETUP_CQE32)
777 return &rings->cqes[off];
780 static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
783 struct io_uring_cqe *cqe;
785 lockdep_assert_held(&ctx->completion_lock);
790 * If we can't get a cq entry, userspace overflowed the
791 * submission (by quite a lot). Increment the overflow count in
794 cqe = io_get_cqe(ctx);
796 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
798 WRITE_ONCE(cqe->user_data, user_data);
799 WRITE_ONCE(cqe->res, res);
800 WRITE_ONCE(cqe->flags, cflags);
802 if (ctx->flags & IORING_SETUP_CQE32) {
803 WRITE_ONCE(cqe->big_cqe[0], 0);
804 WRITE_ONCE(cqe->big_cqe[1], 0);
810 return io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
815 static void __io_flush_post_cqes(struct io_ring_ctx *ctx)
816 __must_hold(&ctx->uring_lock)
818 struct io_submit_state *state = &ctx->submit_state;
821 lockdep_assert_held(&ctx->uring_lock);
822 for (i = 0; i < state->cqes_count; i++) {
823 struct io_uring_cqe *cqe = &state->cqes[i];
825 io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags, true);
827 state->cqes_count = 0;
830 static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
836 filled = io_fill_cqe_aux(ctx, user_data, res, cflags, allow_overflow);
837 io_cq_unlock_post(ctx);
841 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags)
843 return __io_post_aux_cqe(ctx, user_data, res, cflags, true);
846 bool io_aux_cqe(struct io_ring_ctx *ctx, bool defer, u64 user_data, s32 res, u32 cflags,
849 struct io_uring_cqe *cqe;
853 return __io_post_aux_cqe(ctx, user_data, res, cflags, allow_overflow);
855 length = ARRAY_SIZE(ctx->submit_state.cqes);
857 lockdep_assert_held(&ctx->uring_lock);
859 if (ctx->submit_state.cqes_count == length) {
861 __io_flush_post_cqes(ctx);
862 /* no need to flush - flush is deferred */
863 spin_unlock(&ctx->completion_lock);
866 /* For defered completions this is not as strict as it is otherwise,
867 * however it's main job is to prevent unbounded posted completions,
868 * and in that it works just as well.
870 if (!allow_overflow && test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
873 cqe = &ctx->submit_state.cqes[ctx->submit_state.cqes_count++];
874 cqe->user_data = user_data;
880 static void __io_req_complete_post(struct io_kiocb *req)
882 struct io_ring_ctx *ctx = req->ctx;
885 if (!(req->flags & REQ_F_CQE_SKIP))
886 __io_fill_cqe_req(ctx, req);
889 * If we're the last reference to this request, add to our locked
892 if (req_ref_put_and_test(req)) {
893 if (req->flags & IO_REQ_LINK_FLAGS) {
894 if (req->flags & IO_DISARM_MASK)
897 io_req_task_queue(req->link);
901 io_req_put_rsrc(req);
903 * Selected buffer deallocation in io_clean_op() assumes that
904 * we don't hold ->completion_lock. Clean them here to avoid
907 io_put_kbuf_comp(req);
908 io_dismantle_req(req);
909 io_put_task(req->task, 1);
910 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
911 ctx->locked_free_nr++;
913 io_cq_unlock_post(ctx);
916 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
918 if (!(issue_flags & IO_URING_F_UNLOCKED) ||
919 !(req->ctx->flags & IORING_SETUP_IOPOLL)) {
920 __io_req_complete_post(req);
922 struct io_ring_ctx *ctx = req->ctx;
924 mutex_lock(&ctx->uring_lock);
925 __io_req_complete_post(req);
926 mutex_unlock(&ctx->uring_lock);
930 void io_req_defer_failed(struct io_kiocb *req, s32 res)
931 __must_hold(&ctx->uring_lock)
933 const struct io_op_def *def = &io_op_defs[req->opcode];
935 lockdep_assert_held(&req->ctx->uring_lock);
938 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
941 io_req_complete_defer(req);
945 * Don't initialise the fields below on every allocation, but do that in
946 * advance and keep them valid across allocations.
948 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
952 req->async_data = NULL;
953 /* not necessary, but safer to zero */
957 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
958 struct io_submit_state *state)
960 spin_lock(&ctx->completion_lock);
961 wq_list_splice(&ctx->locked_free_list, &state->free_list);
962 ctx->locked_free_nr = 0;
963 spin_unlock(&ctx->completion_lock);
967 * A request might get retired back into the request caches even before opcode
968 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
969 * Because of that, io_alloc_req() should be called only under ->uring_lock
970 * and with extra caution to not get a request that is still worked on.
972 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
973 __must_hold(&ctx->uring_lock)
975 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
976 void *reqs[IO_REQ_ALLOC_BATCH];
980 * If we have more than a batch's worth of requests in our IRQ side
981 * locked cache, grab the lock and move them over to our submission
984 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
985 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
986 if (!io_req_cache_empty(ctx))
990 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
993 * Bulk alloc is all-or-nothing. If we fail to get a batch,
994 * retry single alloc to be on the safe side.
996 if (unlikely(ret <= 0)) {
997 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1003 percpu_ref_get_many(&ctx->refs, ret);
1004 for (i = 0; i < ret; i++) {
1005 struct io_kiocb *req = reqs[i];
1007 io_preinit_req(req, ctx);
1008 io_req_add_to_cache(req, ctx);
1013 static inline void io_dismantle_req(struct io_kiocb *req)
1015 unsigned int flags = req->flags;
1017 if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
1019 if (!(flags & REQ_F_FIXED_FILE))
1020 io_put_file(req->file);
1023 __cold void io_free_req(struct io_kiocb *req)
1025 struct io_ring_ctx *ctx = req->ctx;
1027 io_req_put_rsrc(req);
1028 io_dismantle_req(req);
1029 io_put_task(req->task, 1);
1031 spin_lock(&ctx->completion_lock);
1032 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
1033 ctx->locked_free_nr++;
1034 spin_unlock(&ctx->completion_lock);
1037 static void __io_req_find_next_prep(struct io_kiocb *req)
1039 struct io_ring_ctx *ctx = req->ctx;
1042 io_disarm_next(req);
1043 io_cq_unlock_post(ctx);
1046 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1048 struct io_kiocb *nxt;
1051 * If LINK is set, we have dependent requests in this chain. If we
1052 * didn't fail this request, queue the first one up, moving any other
1053 * dependencies to the next request. In case of failure, fail the rest
1056 if (unlikely(req->flags & IO_DISARM_MASK))
1057 __io_req_find_next_prep(req);
1063 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
1067 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1068 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1070 io_submit_flush_completions(ctx);
1071 mutex_unlock(&ctx->uring_lock);
1074 percpu_ref_put(&ctx->refs);
1077 static unsigned int handle_tw_list(struct llist_node *node,
1078 struct io_ring_ctx **ctx, bool *locked,
1079 struct llist_node *last)
1081 unsigned int count = 0;
1083 while (node != last) {
1084 struct llist_node *next = node->next;
1085 struct io_kiocb *req = container_of(node, struct io_kiocb,
1088 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1090 if (req->ctx != *ctx) {
1091 ctx_flush_and_put(*ctx, locked);
1093 /* if not contended, grab and improve batching */
1094 *locked = mutex_trylock(&(*ctx)->uring_lock);
1095 percpu_ref_get(&(*ctx)->refs);
1097 req->io_task_work.func(req, locked);
1106 * io_llist_xchg - swap all entries in a lock-less list
1107 * @head: the head of lock-less list to delete all entries
1108 * @new: new entry as the head of the list
1110 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1111 * The order of entries returned is from the newest to the oldest added one.
1113 static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1114 struct llist_node *new)
1116 return xchg(&head->first, new);
1120 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1121 * @head: the head of lock-less list to delete all entries
1122 * @old: expected old value of the first entry of the list
1123 * @new: new entry as the head of the list
1125 * perform a cmpxchg on the first entry of the list.
1128 static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1129 struct llist_node *old,
1130 struct llist_node *new)
1132 return cmpxchg(&head->first, old, new);
1135 void tctx_task_work(struct callback_head *cb)
1137 bool uring_locked = false;
1138 struct io_ring_ctx *ctx = NULL;
1139 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1141 struct llist_node fake = {};
1142 struct llist_node *node = io_llist_xchg(&tctx->task_list, &fake);
1143 unsigned int loops = 1;
1144 unsigned int count = handle_tw_list(node, &ctx, &uring_locked, NULL);
1146 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1147 while (node != &fake) {
1149 node = io_llist_xchg(&tctx->task_list, &fake);
1150 count += handle_tw_list(node, &ctx, &uring_locked, &fake);
1151 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1154 ctx_flush_and_put(ctx, &uring_locked);
1156 /* relaxed read is enough as only the task itself sets ->in_idle */
1157 if (unlikely(atomic_read(&tctx->in_idle)))
1158 io_uring_drop_tctx_refs(current);
1160 trace_io_uring_task_work_run(tctx, count, loops);
1163 static __cold void io_fallback_tw(struct io_uring_task *tctx)
1165 struct llist_node *node = llist_del_all(&tctx->task_list);
1166 struct io_kiocb *req;
1169 req = container_of(node, struct io_kiocb, io_task_work.node);
1171 if (llist_add(&req->io_task_work.node,
1172 &req->ctx->fallback_llist))
1173 schedule_delayed_work(&req->ctx->fallback_work, 1);
1177 static void io_req_local_work_add(struct io_kiocb *req)
1179 struct io_ring_ctx *ctx = req->ctx;
1181 if (!llist_add(&req->io_task_work.node, &ctx->work_llist))
1183 /* need it for the following io_cqring_wake() */
1184 smp_mb__after_atomic();
1186 if (unlikely(atomic_read(&req->task->io_uring->in_idle))) {
1187 io_move_task_work_from_local(ctx);
1191 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1192 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1195 io_eventfd_signal(ctx);
1196 __io_cqring_wake(ctx);
1199 void __io_req_task_work_add(struct io_kiocb *req, bool allow_local)
1201 struct io_uring_task *tctx = req->task->io_uring;
1202 struct io_ring_ctx *ctx = req->ctx;
1204 if (allow_local && ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
1205 io_req_local_work_add(req);
1209 /* task_work already pending, we're done */
1210 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1213 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1214 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1216 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1219 io_fallback_tw(tctx);
1222 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1224 struct llist_node *node;
1226 node = llist_del_all(&ctx->work_llist);
1228 struct io_kiocb *req = container_of(node, struct io_kiocb,
1232 __io_req_task_work_add(req, false);
1236 int __io_run_local_work(struct io_ring_ctx *ctx, bool *locked)
1238 struct llist_node *node;
1239 struct llist_node fake;
1240 struct llist_node *current_final = NULL;
1242 unsigned int loops = 1;
1244 if (unlikely(ctx->submitter_task != current))
1247 node = io_llist_xchg(&ctx->work_llist, &fake);
1250 while (node != current_final) {
1251 struct llist_node *next = node->next;
1252 struct io_kiocb *req = container_of(node, struct io_kiocb,
1254 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1255 req->io_task_work.func(req, locked);
1260 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1261 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1263 node = io_llist_cmpxchg(&ctx->work_llist, &fake, NULL);
1264 if (node != &fake) {
1266 current_final = &fake;
1267 node = io_llist_xchg(&ctx->work_llist, &fake);
1272 io_submit_flush_completions(ctx);
1273 trace_io_uring_local_work_run(ctx, ret, loops);
1278 int io_run_local_work(struct io_ring_ctx *ctx)
1283 if (llist_empty(&ctx->work_llist))
1286 __set_current_state(TASK_RUNNING);
1287 locked = mutex_trylock(&ctx->uring_lock);
1288 ret = __io_run_local_work(ctx, &locked);
1290 mutex_unlock(&ctx->uring_lock);
1295 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
1297 io_tw_lock(req->ctx, locked);
1298 io_req_defer_failed(req, req->cqe.res);
1301 void io_req_task_submit(struct io_kiocb *req, bool *locked)
1303 io_tw_lock(req->ctx, locked);
1304 /* req->task == current here, checking PF_EXITING is safe */
1305 if (likely(!(req->task->flags & PF_EXITING)))
1308 io_req_defer_failed(req, -EFAULT);
1311 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1313 io_req_set_res(req, ret, 0);
1314 req->io_task_work.func = io_req_task_cancel;
1315 io_req_task_work_add(req);
1318 void io_req_task_queue(struct io_kiocb *req)
1320 req->io_task_work.func = io_req_task_submit;
1321 io_req_task_work_add(req);
1324 void io_queue_next(struct io_kiocb *req)
1326 struct io_kiocb *nxt = io_req_find_next(req);
1329 io_req_task_queue(nxt);
1332 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1333 __must_hold(&ctx->uring_lock)
1335 struct task_struct *task = NULL;
1339 struct io_kiocb *req = container_of(node, struct io_kiocb,
1342 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1343 if (req->flags & REQ_F_REFCOUNT) {
1344 node = req->comp_list.next;
1345 if (!req_ref_put_and_test(req))
1348 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1349 struct async_poll *apoll = req->apoll;
1351 if (apoll->double_poll)
1352 kfree(apoll->double_poll);
1353 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1355 req->flags &= ~REQ_F_POLLED;
1357 if (req->flags & IO_REQ_LINK_FLAGS)
1359 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1362 if (!(req->flags & REQ_F_FIXED_FILE))
1363 io_put_file(req->file);
1365 io_req_put_rsrc_locked(req, ctx);
1367 if (req->task != task) {
1369 io_put_task(task, task_refs);
1374 node = req->comp_list.next;
1375 io_req_add_to_cache(req, ctx);
1379 io_put_task(task, task_refs);
1382 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1383 __must_hold(&ctx->uring_lock)
1385 struct io_wq_work_node *node, *prev;
1386 struct io_submit_state *state = &ctx->submit_state;
1389 /* must come first to preserve CQE ordering in failure cases */
1390 if (state->cqes_count)
1391 __io_flush_post_cqes(ctx);
1392 wq_list_for_each(node, prev, &state->compl_reqs) {
1393 struct io_kiocb *req = container_of(node, struct io_kiocb,
1396 if (!(req->flags & REQ_F_CQE_SKIP))
1397 __io_fill_cqe_req(ctx, req);
1399 io_cq_unlock_post_inline(ctx);
1401 if (!wq_list_empty(&ctx->submit_state.compl_reqs)) {
1402 io_free_batch_list(ctx, state->compl_reqs.first);
1403 INIT_WQ_LIST(&state->compl_reqs);
1408 * Drop reference to request, return next in chain (if there is one) if this
1409 * was the last reference to this request.
1411 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1413 struct io_kiocb *nxt = NULL;
1415 if (req_ref_put_and_test(req)) {
1416 if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
1417 nxt = io_req_find_next(req);
1423 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1425 /* See comment at the top of this file */
1427 return __io_cqring_events(ctx);
1431 * We can't just wait for polled events to come to us, we have to actively
1432 * find and complete them.
1434 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1436 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1439 mutex_lock(&ctx->uring_lock);
1440 while (!wq_list_empty(&ctx->iopoll_list)) {
1441 /* let it sleep and repeat later if can't complete a request */
1442 if (io_do_iopoll(ctx, true) == 0)
1445 * Ensure we allow local-to-the-cpu processing to take place,
1446 * in this case we need to ensure that we reap all events.
1447 * Also let task_work, etc. to progress by releasing the mutex
1449 if (need_resched()) {
1450 mutex_unlock(&ctx->uring_lock);
1452 mutex_lock(&ctx->uring_lock);
1455 mutex_unlock(&ctx->uring_lock);
1458 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1460 unsigned int nr_events = 0;
1462 unsigned long check_cq;
1464 if (!io_allowed_run_tw(ctx))
1467 check_cq = READ_ONCE(ctx->check_cq);
1468 if (unlikely(check_cq)) {
1469 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1470 __io_cqring_overflow_flush(ctx, false);
1472 * Similarly do not spin if we have not informed the user of any
1475 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1479 * Don't enter poll loop if we already have events pending.
1480 * If we do, we can potentially be spinning for commands that
1481 * already triggered a CQE (eg in error).
1483 if (io_cqring_events(ctx))
1488 * If a submit got punted to a workqueue, we can have the
1489 * application entering polling for a command before it gets
1490 * issued. That app will hold the uring_lock for the duration
1491 * of the poll right here, so we need to take a breather every
1492 * now and then to ensure that the issue has a chance to add
1493 * the poll to the issued list. Otherwise we can spin here
1494 * forever, while the workqueue is stuck trying to acquire the
1497 if (wq_list_empty(&ctx->iopoll_list) ||
1498 io_task_work_pending(ctx)) {
1499 u32 tail = ctx->cached_cq_tail;
1501 (void) io_run_local_work_locked(ctx);
1503 if (task_work_pending(current) ||
1504 wq_list_empty(&ctx->iopoll_list)) {
1505 mutex_unlock(&ctx->uring_lock);
1507 mutex_lock(&ctx->uring_lock);
1509 /* some requests don't go through iopoll_list */
1510 if (tail != ctx->cached_cq_tail ||
1511 wq_list_empty(&ctx->iopoll_list))
1514 ret = io_do_iopoll(ctx, !min);
1519 } while (nr_events < min && !need_resched());
1524 void io_req_task_complete(struct io_kiocb *req, bool *locked)
1527 io_req_complete_defer(req);
1529 io_req_complete_post(req, IO_URING_F_UNLOCKED);
1533 * After the iocb has been issued, it's safe to be found on the poll list.
1534 * Adding the kiocb to the list AFTER submission ensures that we don't
1535 * find it from a io_do_iopoll() thread before the issuer is done
1536 * accessing the kiocb cookie.
1538 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1540 struct io_ring_ctx *ctx = req->ctx;
1541 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1543 /* workqueue context doesn't hold uring_lock, grab it now */
1544 if (unlikely(needs_lock))
1545 mutex_lock(&ctx->uring_lock);
1548 * Track whether we have multiple files in our lists. This will impact
1549 * how we do polling eventually, not spinning if we're on potentially
1550 * different devices.
1552 if (wq_list_empty(&ctx->iopoll_list)) {
1553 ctx->poll_multi_queue = false;
1554 } else if (!ctx->poll_multi_queue) {
1555 struct io_kiocb *list_req;
1557 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1559 if (list_req->file != req->file)
1560 ctx->poll_multi_queue = true;
1564 * For fast devices, IO may have already completed. If it has, add
1565 * it to the front so we find it first.
1567 if (READ_ONCE(req->iopoll_completed))
1568 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1570 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1572 if (unlikely(needs_lock)) {
1574 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1575 * in sq thread task context or in io worker task context. If
1576 * current task context is sq thread, we don't need to check
1577 * whether should wake up sq thread.
1579 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1580 wq_has_sleeper(&ctx->sq_data->wait))
1581 wake_up(&ctx->sq_data->wait);
1583 mutex_unlock(&ctx->uring_lock);
1587 static bool io_bdev_nowait(struct block_device *bdev)
1589 return !bdev || bdev_nowait(bdev);
1593 * If we tracked the file through the SCM inflight mechanism, we could support
1594 * any file. For now, just ensure that anything potentially problematic is done
1597 static bool __io_file_supports_nowait(struct file *file, umode_t mode)
1599 if (S_ISBLK(mode)) {
1600 if (IS_ENABLED(CONFIG_BLOCK) &&
1601 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
1607 if (S_ISREG(mode)) {
1608 if (IS_ENABLED(CONFIG_BLOCK) &&
1609 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
1610 !io_is_uring_fops(file))
1615 /* any ->read/write should understand O_NONBLOCK */
1616 if (file->f_flags & O_NONBLOCK)
1618 return file->f_mode & FMODE_NOWAIT;
1622 * If we tracked the file through the SCM inflight mechanism, we could support
1623 * any file. For now, just ensure that anything potentially problematic is done
1626 unsigned int io_file_get_flags(struct file *file)
1628 umode_t mode = file_inode(file)->i_mode;
1629 unsigned int res = 0;
1633 if (__io_file_supports_nowait(file, mode))
1638 bool io_alloc_async_data(struct io_kiocb *req)
1640 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
1641 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
1642 if (req->async_data) {
1643 req->flags |= REQ_F_ASYNC_DATA;
1649 int io_req_prep_async(struct io_kiocb *req)
1651 const struct io_op_def *def = &io_op_defs[req->opcode];
1653 /* assign early for deferred execution for non-fixed file */
1654 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE))
1655 req->file = io_file_get_normal(req, req->cqe.fd);
1656 if (!def->prep_async)
1658 if (WARN_ON_ONCE(req_has_async_data(req)))
1660 if (!io_op_defs[req->opcode].manual_alloc) {
1661 if (io_alloc_async_data(req))
1664 return def->prep_async(req);
1667 static u32 io_get_sequence(struct io_kiocb *req)
1669 u32 seq = req->ctx->cached_sq_head;
1670 struct io_kiocb *cur;
1672 /* need original cached_sq_head, but it was increased for each req */
1673 io_for_each_link(cur, req)
1678 static __cold void io_drain_req(struct io_kiocb *req)
1679 __must_hold(&ctx->uring_lock)
1681 struct io_ring_ctx *ctx = req->ctx;
1682 struct io_defer_entry *de;
1684 u32 seq = io_get_sequence(req);
1686 /* Still need defer if there is pending req in defer list. */
1687 spin_lock(&ctx->completion_lock);
1688 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1689 spin_unlock(&ctx->completion_lock);
1691 ctx->drain_active = false;
1692 io_req_task_queue(req);
1695 spin_unlock(&ctx->completion_lock);
1697 ret = io_req_prep_async(req);
1700 io_req_defer_failed(req, ret);
1703 io_prep_async_link(req);
1704 de = kmalloc(sizeof(*de), GFP_KERNEL);
1710 spin_lock(&ctx->completion_lock);
1711 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1712 spin_unlock(&ctx->completion_lock);
1717 trace_io_uring_defer(req);
1720 list_add_tail(&de->list, &ctx->defer_list);
1721 spin_unlock(&ctx->completion_lock);
1724 static void io_clean_op(struct io_kiocb *req)
1726 if (req->flags & REQ_F_BUFFER_SELECTED) {
1727 spin_lock(&req->ctx->completion_lock);
1728 io_put_kbuf_comp(req);
1729 spin_unlock(&req->ctx->completion_lock);
1732 if (req->flags & REQ_F_NEED_CLEANUP) {
1733 const struct io_op_def *def = &io_op_defs[req->opcode];
1738 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1739 kfree(req->apoll->double_poll);
1743 if (req->flags & REQ_F_INFLIGHT) {
1744 struct io_uring_task *tctx = req->task->io_uring;
1746 atomic_dec(&tctx->inflight_tracked);
1748 if (req->flags & REQ_F_CREDS)
1749 put_cred(req->creds);
1750 if (req->flags & REQ_F_ASYNC_DATA) {
1751 kfree(req->async_data);
1752 req->async_data = NULL;
1754 req->flags &= ~IO_REQ_CLEAN_FLAGS;
1757 static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags)
1759 if (req->file || !io_op_defs[req->opcode].needs_file)
1762 if (req->flags & REQ_F_FIXED_FILE)
1763 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1765 req->file = io_file_get_normal(req, req->cqe.fd);
1770 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1772 const struct io_op_def *def = &io_op_defs[req->opcode];
1773 const struct cred *creds = NULL;
1776 if (unlikely(!io_assign_file(req, issue_flags)))
1779 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1780 creds = override_creds(req->creds);
1782 if (!def->audit_skip)
1783 audit_uring_entry(req->opcode);
1785 ret = def->issue(req, issue_flags);
1787 if (!def->audit_skip)
1788 audit_uring_exit(!ret, ret);
1791 revert_creds(creds);
1793 if (ret == IOU_OK) {
1794 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1795 io_req_complete_defer(req);
1797 io_req_complete_post(req, issue_flags);
1798 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1801 /* If the op doesn't have a file, we're not polling for it */
1802 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && req->file)
1803 io_iopoll_req_issued(req, issue_flags);
1808 int io_poll_issue(struct io_kiocb *req, bool *locked)
1810 io_tw_lock(req->ctx, locked);
1811 if (unlikely(req->task->flags & PF_EXITING))
1813 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT|
1814 IO_URING_F_COMPLETE_DEFER);
1817 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1819 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1821 req = io_put_req_find_next(req);
1822 return req ? &req->work : NULL;
1825 void io_wq_submit_work(struct io_wq_work *work)
1827 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1828 const struct io_op_def *def = &io_op_defs[req->opcode];
1829 unsigned int issue_flags = IO_URING_F_UNLOCKED;
1830 bool needs_poll = false;
1831 int ret = 0, err = -ECANCELED;
1833 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1834 if (!(req->flags & REQ_F_REFCOUNT))
1835 __io_req_set_refcount(req, 2);
1839 io_arm_ltimeout(req);
1841 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1842 if (work->flags & IO_WQ_WORK_CANCEL) {
1844 io_req_task_queue_fail(req, err);
1847 if (!io_assign_file(req, issue_flags)) {
1849 work->flags |= IO_WQ_WORK_CANCEL;
1853 if (req->flags & REQ_F_FORCE_ASYNC) {
1854 bool opcode_poll = def->pollin || def->pollout;
1856 if (opcode_poll && file_can_poll(req->file)) {
1858 issue_flags |= IO_URING_F_NONBLOCK;
1863 ret = io_issue_sqe(req, issue_flags);
1867 * We can get EAGAIN for iopolled IO even though we're
1868 * forcing a sync submission from here, since we can't
1869 * wait for request slots on the block side.
1872 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1878 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1880 /* aborted or ready, in either case retry blocking */
1882 issue_flags &= ~IO_URING_F_NONBLOCK;
1885 /* avoid locking problems by failing it from a clean context */
1887 io_req_task_queue_fail(req, ret);
1890 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1891 unsigned int issue_flags)
1893 struct io_ring_ctx *ctx = req->ctx;
1894 struct file *file = NULL;
1895 unsigned long file_ptr;
1897 io_ring_submit_lock(ctx, issue_flags);
1899 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
1901 fd = array_index_nospec(fd, ctx->nr_user_files);
1902 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
1903 file = (struct file *) (file_ptr & FFS_MASK);
1904 file_ptr &= ~FFS_MASK;
1905 /* mask in overlapping REQ_F and FFS bits */
1906 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
1907 io_req_set_rsrc_node(req, ctx, 0);
1909 io_ring_submit_unlock(ctx, issue_flags);
1913 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
1915 struct file *file = fget(fd);
1917 trace_io_uring_file_get(req, fd);
1919 /* we don't allow fixed io_uring files */
1920 if (file && io_is_uring_fops(file))
1921 io_req_track_inflight(req);
1925 static void io_queue_async(struct io_kiocb *req, int ret)
1926 __must_hold(&req->ctx->uring_lock)
1928 struct io_kiocb *linked_timeout;
1930 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
1931 io_req_defer_failed(req, ret);
1935 linked_timeout = io_prep_linked_timeout(req);
1937 switch (io_arm_poll_handler(req, 0)) {
1938 case IO_APOLL_READY:
1939 io_kbuf_recycle(req, 0);
1940 io_req_task_queue(req);
1942 case IO_APOLL_ABORTED:
1943 io_kbuf_recycle(req, 0);
1944 io_queue_iowq(req, NULL);
1951 io_queue_linked_timeout(linked_timeout);
1954 static inline void io_queue_sqe(struct io_kiocb *req)
1955 __must_hold(&req->ctx->uring_lock)
1959 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
1962 * We async punt it if the file wasn't marked NOWAIT, or if the file
1963 * doesn't support non-blocking read/write attempts
1966 io_arm_ltimeout(req);
1968 io_queue_async(req, ret);
1971 static void io_queue_sqe_fallback(struct io_kiocb *req)
1972 __must_hold(&req->ctx->uring_lock)
1974 if (unlikely(req->flags & REQ_F_FAIL)) {
1976 * We don't submit, fail them all, for that replace hardlinks
1977 * with normal links. Extra REQ_F_LINK is tolerated.
1979 req->flags &= ~REQ_F_HARDLINK;
1980 req->flags |= REQ_F_LINK;
1981 io_req_defer_failed(req, req->cqe.res);
1982 } else if (unlikely(req->ctx->drain_active)) {
1985 int ret = io_req_prep_async(req);
1988 io_req_defer_failed(req, ret);
1990 io_queue_iowq(req, NULL);
1995 * Check SQE restrictions (opcode and flags).
1997 * Returns 'true' if SQE is allowed, 'false' otherwise.
1999 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
2000 struct io_kiocb *req,
2001 unsigned int sqe_flags)
2003 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
2006 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
2007 ctx->restrictions.sqe_flags_required)
2010 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
2011 ctx->restrictions.sqe_flags_required))
2017 static void io_init_req_drain(struct io_kiocb *req)
2019 struct io_ring_ctx *ctx = req->ctx;
2020 struct io_kiocb *head = ctx->submit_state.link.head;
2022 ctx->drain_active = true;
2025 * If we need to drain a request in the middle of a link, drain
2026 * the head request and the next request/link after the current
2027 * link. Considering sequential execution of links,
2028 * REQ_F_IO_DRAIN will be maintained for every request of our
2031 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2032 ctx->drain_next = true;
2036 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
2037 const struct io_uring_sqe *sqe)
2038 __must_hold(&ctx->uring_lock)
2040 const struct io_op_def *def;
2041 unsigned int sqe_flags;
2045 /* req is partially pre-initialised, see io_preinit_req() */
2046 req->opcode = opcode = READ_ONCE(sqe->opcode);
2047 /* same numerical values with corresponding REQ_F_*, safe to copy */
2048 req->flags = sqe_flags = READ_ONCE(sqe->flags);
2049 req->cqe.user_data = READ_ONCE(sqe->user_data);
2051 req->rsrc_node = NULL;
2052 req->task = current;
2054 if (unlikely(opcode >= IORING_OP_LAST)) {
2058 def = &io_op_defs[opcode];
2059 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2060 /* enforce forwards compatibility on users */
2061 if (sqe_flags & ~SQE_VALID_FLAGS)
2063 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2064 if (!def->buffer_select)
2066 req->buf_index = READ_ONCE(sqe->buf_group);
2068 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2069 ctx->drain_disabled = true;
2070 if (sqe_flags & IOSQE_IO_DRAIN) {
2071 if (ctx->drain_disabled)
2073 io_init_req_drain(req);
2076 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2077 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2079 /* knock it to the slow queue path, will be drained there */
2080 if (ctx->drain_active)
2081 req->flags |= REQ_F_FORCE_ASYNC;
2082 /* if there is no link, we're at "next" request and need to drain */
2083 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2084 ctx->drain_next = false;
2085 ctx->drain_active = true;
2086 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2090 if (!def->ioprio && sqe->ioprio)
2092 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2095 if (def->needs_file) {
2096 struct io_submit_state *state = &ctx->submit_state;
2098 req->cqe.fd = READ_ONCE(sqe->fd);
2101 * Plug now if we have more than 2 IO left after this, and the
2102 * target is potentially a read/write to block based storage.
2104 if (state->need_plug && def->plug) {
2105 state->plug_started = true;
2106 state->need_plug = false;
2107 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2111 personality = READ_ONCE(sqe->personality);
2115 req->creds = xa_load(&ctx->personalities, personality);
2118 get_cred(req->creds);
2119 ret = security_uring_override_creds(req->creds);
2121 put_cred(req->creds);
2124 req->flags |= REQ_F_CREDS;
2127 return def->prep(req, sqe);
2130 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2131 struct io_kiocb *req, int ret)
2133 struct io_ring_ctx *ctx = req->ctx;
2134 struct io_submit_link *link = &ctx->submit_state.link;
2135 struct io_kiocb *head = link->head;
2137 trace_io_uring_req_failed(sqe, req, ret);
2140 * Avoid breaking links in the middle as it renders links with SQPOLL
2141 * unusable. Instead of failing eagerly, continue assembling the link if
2142 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2143 * should find the flag and handle the rest.
2145 req_fail_link_node(req, ret);
2146 if (head && !(head->flags & REQ_F_FAIL))
2147 req_fail_link_node(head, -ECANCELED);
2149 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2151 link->last->link = req;
2155 io_queue_sqe_fallback(req);
2160 link->last->link = req;
2167 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2168 const struct io_uring_sqe *sqe)
2169 __must_hold(&ctx->uring_lock)
2171 struct io_submit_link *link = &ctx->submit_state.link;
2174 ret = io_init_req(ctx, req, sqe);
2176 return io_submit_fail_init(sqe, req, ret);
2178 /* don't need @sqe from now on */
2179 trace_io_uring_submit_sqe(req, true);
2182 * If we already have a head request, queue this one for async
2183 * submittal once the head completes. If we don't have a head but
2184 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2185 * submitted sync once the chain is complete. If none of those
2186 * conditions are true (normal request), then just queue it.
2188 if (unlikely(link->head)) {
2189 ret = io_req_prep_async(req);
2191 return io_submit_fail_init(sqe, req, ret);
2193 trace_io_uring_link(req, link->head);
2194 link->last->link = req;
2197 if (req->flags & IO_REQ_LINK_FLAGS)
2199 /* last request of the link, flush it */
2202 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2205 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2206 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2207 if (req->flags & IO_REQ_LINK_FLAGS) {
2212 io_queue_sqe_fallback(req);
2222 * Batched submission is done, ensure local IO is flushed out.
2224 static void io_submit_state_end(struct io_ring_ctx *ctx)
2226 struct io_submit_state *state = &ctx->submit_state;
2228 if (unlikely(state->link.head))
2229 io_queue_sqe_fallback(state->link.head);
2230 /* flush only after queuing links as they can generate completions */
2231 io_submit_flush_completions(ctx);
2232 if (state->plug_started)
2233 blk_finish_plug(&state->plug);
2237 * Start submission side cache.
2239 static void io_submit_state_start(struct io_submit_state *state,
2240 unsigned int max_ios)
2242 state->plug_started = false;
2243 state->need_plug = max_ios > 2;
2244 state->submit_nr = max_ios;
2245 /* set only head, no need to init link_last in advance */
2246 state->link.head = NULL;
2249 static void io_commit_sqring(struct io_ring_ctx *ctx)
2251 struct io_rings *rings = ctx->rings;
2254 * Ensure any loads from the SQEs are done at this point,
2255 * since once we write the new head, the application could
2256 * write new data to them.
2258 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2262 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2263 * that is mapped by userspace. This means that care needs to be taken to
2264 * ensure that reads are stable, as we cannot rely on userspace always
2265 * being a good citizen. If members of the sqe are validated and then later
2266 * used, it's important that those reads are done through READ_ONCE() to
2267 * prevent a re-load down the line.
2269 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
2271 unsigned head, mask = ctx->sq_entries - 1;
2272 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2275 * The cached sq head (or cq tail) serves two purposes:
2277 * 1) allows us to batch the cost of updating the user visible
2279 * 2) allows the kernel side to track the head on its own, even
2280 * though the application is the one updating it.
2282 head = READ_ONCE(ctx->sq_array[sq_idx]);
2283 if (likely(head < ctx->sq_entries)) {
2284 /* double index for 128-byte SQEs, twice as long */
2285 if (ctx->flags & IORING_SETUP_SQE128)
2287 return &ctx->sq_sqes[head];
2290 /* drop invalid entries */
2292 WRITE_ONCE(ctx->rings->sq_dropped,
2293 READ_ONCE(ctx->rings->sq_dropped) + 1);
2297 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2298 __must_hold(&ctx->uring_lock)
2300 unsigned int entries = io_sqring_entries(ctx);
2304 if (unlikely(!entries))
2306 /* make sure SQ entry isn't read before tail */
2307 ret = left = min3(nr, ctx->sq_entries, entries);
2308 io_get_task_refs(left);
2309 io_submit_state_start(&ctx->submit_state, left);
2312 const struct io_uring_sqe *sqe;
2313 struct io_kiocb *req;
2315 if (unlikely(!io_alloc_req_refill(ctx)))
2317 req = io_alloc_req(ctx);
2318 sqe = io_get_sqe(ctx);
2319 if (unlikely(!sqe)) {
2320 io_req_add_to_cache(req, ctx);
2325 * Continue submitting even for sqe failure if the
2326 * ring was setup with IORING_SETUP_SUBMIT_ALL
2328 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2329 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2335 if (unlikely(left)) {
2337 /* try again if it submitted nothing and can't allocate a req */
2338 if (!ret && io_req_cache_empty(ctx))
2340 current->io_uring->cached_refs += left;
2343 io_submit_state_end(ctx);
2344 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2345 io_commit_sqring(ctx);
2349 struct io_wait_queue {
2350 struct wait_queue_entry wq;
2351 struct io_ring_ctx *ctx;
2353 unsigned nr_timeouts;
2356 static inline bool io_has_work(struct io_ring_ctx *ctx)
2358 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
2359 ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
2360 !llist_empty(&ctx->work_llist));
2363 static inline bool io_should_wake(struct io_wait_queue *iowq)
2365 struct io_ring_ctx *ctx = iowq->ctx;
2366 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
2369 * Wake up if we have enough events, or if a timeout occurred since we
2370 * started waiting. For timeouts, we always want to return to userspace,
2371 * regardless of event count.
2373 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2376 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2377 int wake_flags, void *key)
2379 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
2381 struct io_ring_ctx *ctx = iowq->ctx;
2384 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2385 * the task, and the next invocation will do it.
2387 if (io_should_wake(iowq) || io_has_work(ctx))
2388 return autoremove_wake_function(curr, mode, wake_flags, key);
2392 int io_run_task_work_sig(struct io_ring_ctx *ctx)
2394 if (io_run_task_work_ctx(ctx) > 0)
2396 if (task_sigpending(current))
2401 /* when returns >0, the caller should retry */
2402 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2403 struct io_wait_queue *iowq,
2407 unsigned long check_cq;
2409 /* make sure we run task_work before checking for signals */
2410 ret = io_run_task_work_sig(ctx);
2411 if (ret || io_should_wake(iowq))
2414 check_cq = READ_ONCE(ctx->check_cq);
2415 if (unlikely(check_cq)) {
2416 /* let the caller flush overflows, retry */
2417 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2419 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
2422 if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS))
2428 * Wait until events become available, if we don't already have some. The
2429 * application must reap them itself, as they reside on the shared cq ring.
2431 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2432 const sigset_t __user *sig, size_t sigsz,
2433 struct __kernel_timespec __user *uts)
2435 struct io_wait_queue iowq;
2436 struct io_rings *rings = ctx->rings;
2437 ktime_t timeout = KTIME_MAX;
2440 if (!io_allowed_run_tw(ctx))
2444 /* always run at least 1 task work to process local work */
2445 ret = io_run_task_work_ctx(ctx);
2448 io_cqring_overflow_flush(ctx);
2450 /* if user messes with these they will just get an early return */
2451 if (__io_cqring_events_user(ctx) >= min_events)
2456 #ifdef CONFIG_COMPAT
2457 if (in_compat_syscall())
2458 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2462 ret = set_user_sigmask(sig, sigsz);
2469 struct timespec64 ts;
2471 if (get_timespec64(&ts, uts))
2473 timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2476 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2477 iowq.wq.private = current;
2478 INIT_LIST_HEAD(&iowq.wq.entry);
2480 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2481 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2483 trace_io_uring_cqring_wait(ctx, min_events);
2485 /* if we can't even flush overflow, don't wait for more */
2486 if (!io_cqring_overflow_flush(ctx)) {
2490 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2491 TASK_INTERRUPTIBLE);
2492 ret = io_cqring_wait_schedule(ctx, &iowq, timeout);
2496 finish_wait(&ctx->cq_wait, &iowq.wq);
2497 restore_saved_sigmask_unless(ret == -EINTR);
2499 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2502 static void io_mem_free(void *ptr)
2509 page = virt_to_head_page(ptr);
2510 if (put_page_testzero(page))
2511 free_compound_page(page);
2514 static void *io_mem_alloc(size_t size)
2516 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2518 return (void *) __get_free_pages(gfp, get_order(size));
2521 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2522 unsigned int cq_entries, size_t *sq_offset)
2524 struct io_rings *rings;
2525 size_t off, sq_array_size;
2527 off = struct_size(rings, cqes, cq_entries);
2528 if (off == SIZE_MAX)
2530 if (ctx->flags & IORING_SETUP_CQE32) {
2531 if (check_shl_overflow(off, 1, &off))
2536 off = ALIGN(off, SMP_CACHE_BYTES);
2544 sq_array_size = array_size(sizeof(u32), sq_entries);
2545 if (sq_array_size == SIZE_MAX)
2548 if (check_add_overflow(off, sq_array_size, &off))
2554 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2555 unsigned int eventfd_async)
2557 struct io_ev_fd *ev_fd;
2558 __s32 __user *fds = arg;
2561 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2562 lockdep_is_held(&ctx->uring_lock));
2566 if (copy_from_user(&fd, fds, sizeof(*fds)))
2569 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2573 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2574 if (IS_ERR(ev_fd->cq_ev_fd)) {
2575 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2580 spin_lock(&ctx->completion_lock);
2581 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2582 spin_unlock(&ctx->completion_lock);
2584 ev_fd->eventfd_async = eventfd_async;
2585 ctx->has_evfd = true;
2586 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2587 atomic_set(&ev_fd->refs, 1);
2588 atomic_set(&ev_fd->ops, 0);
2592 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2594 struct io_ev_fd *ev_fd;
2596 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2597 lockdep_is_held(&ctx->uring_lock));
2599 ctx->has_evfd = false;
2600 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2601 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops))
2602 call_rcu(&ev_fd->rcu, io_eventfd_ops);
2609 static void io_req_caches_free(struct io_ring_ctx *ctx)
2613 mutex_lock(&ctx->uring_lock);
2614 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
2616 while (!io_req_cache_empty(ctx)) {
2617 struct io_kiocb *req = io_alloc_req(ctx);
2619 kmem_cache_free(req_cachep, req);
2623 percpu_ref_put_many(&ctx->refs, nr);
2624 mutex_unlock(&ctx->uring_lock);
2627 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2629 io_sq_thread_finish(ctx);
2630 io_rsrc_refs_drop(ctx);
2631 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2632 io_wait_rsrc_data(ctx->buf_data);
2633 io_wait_rsrc_data(ctx->file_data);
2635 mutex_lock(&ctx->uring_lock);
2637 __io_sqe_buffers_unregister(ctx);
2639 __io_sqe_files_unregister(ctx);
2641 __io_cqring_overflow_flush(ctx, true);
2642 io_eventfd_unregister(ctx);
2643 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2644 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2645 mutex_unlock(&ctx->uring_lock);
2646 io_destroy_buffers(ctx);
2648 put_cred(ctx->sq_creds);
2649 if (ctx->submitter_task)
2650 put_task_struct(ctx->submitter_task);
2652 /* there are no registered resources left, nobody uses it */
2654 io_rsrc_node_destroy(ctx->rsrc_node);
2655 if (ctx->rsrc_backup_node)
2656 io_rsrc_node_destroy(ctx->rsrc_backup_node);
2657 flush_delayed_work(&ctx->rsrc_put_work);
2658 flush_delayed_work(&ctx->fallback_work);
2660 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2661 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
2663 #if defined(CONFIG_UNIX)
2664 if (ctx->ring_sock) {
2665 ctx->ring_sock->file = NULL; /* so that iput() is called */
2666 sock_release(ctx->ring_sock);
2669 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2671 if (ctx->mm_account) {
2672 mmdrop(ctx->mm_account);
2673 ctx->mm_account = NULL;
2675 io_mem_free(ctx->rings);
2676 io_mem_free(ctx->sq_sqes);
2678 percpu_ref_exit(&ctx->refs);
2679 free_uid(ctx->user);
2680 io_req_caches_free(ctx);
2682 io_wq_put_hash(ctx->hash_map);
2683 kfree(ctx->cancel_table.hbs);
2684 kfree(ctx->cancel_table_locked.hbs);
2685 kfree(ctx->dummy_ubuf);
2687 xa_destroy(&ctx->io_bl_xa);
2691 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2693 struct io_ring_ctx *ctx = file->private_data;
2696 poll_wait(file, &ctx->cq_wait, wait);
2698 * synchronizes with barrier from wq_has_sleeper call in
2702 if (!io_sqring_full(ctx))
2703 mask |= EPOLLOUT | EPOLLWRNORM;
2706 * Don't flush cqring overflow list here, just do a simple check.
2707 * Otherwise there could possible be ABBA deadlock:
2710 * lock(&ctx->uring_lock);
2712 * lock(&ctx->uring_lock);
2715 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2716 * pushes them to do the flush.
2719 if (io_cqring_events(ctx) || io_has_work(ctx))
2720 mask |= EPOLLIN | EPOLLRDNORM;
2725 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2727 const struct cred *creds;
2729 creds = xa_erase(&ctx->personalities, id);
2738 struct io_tctx_exit {
2739 struct callback_head task_work;
2740 struct completion completion;
2741 struct io_ring_ctx *ctx;
2744 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2746 struct io_uring_task *tctx = current->io_uring;
2747 struct io_tctx_exit *work;
2749 work = container_of(cb, struct io_tctx_exit, task_work);
2751 * When @in_idle, we're in cancellation and it's racy to remove the
2752 * node. It'll be removed by the end of cancellation, just ignore it.
2753 * tctx can be NULL if the queueing of this task_work raced with
2754 * work cancelation off the exec path.
2756 if (tctx && !atomic_read(&tctx->in_idle))
2757 io_uring_del_tctx_node((unsigned long)work->ctx);
2758 complete(&work->completion);
2761 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2763 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2765 return req->ctx == data;
2768 static __cold void io_ring_exit_work(struct work_struct *work)
2770 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2771 unsigned long timeout = jiffies + HZ * 60 * 5;
2772 unsigned long interval = HZ / 20;
2773 struct io_tctx_exit exit;
2774 struct io_tctx_node *node;
2778 * If we're doing polled IO and end up having requests being
2779 * submitted async (out-of-line), then completions can come in while
2780 * we're waiting for refs to drop. We need to reap these manually,
2781 * as nobody else will be looking for them.
2784 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
2785 io_move_task_work_from_local(ctx);
2787 while (io_uring_try_cancel_requests(ctx, NULL, true))
2791 struct io_sq_data *sqd = ctx->sq_data;
2792 struct task_struct *tsk;
2794 io_sq_thread_park(sqd);
2796 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
2797 io_wq_cancel_cb(tsk->io_uring->io_wq,
2798 io_cancel_ctx_cb, ctx, true);
2799 io_sq_thread_unpark(sqd);
2802 io_req_caches_free(ctx);
2804 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
2805 /* there is little hope left, don't run it too often */
2808 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
2810 init_completion(&exit.completion);
2811 init_task_work(&exit.task_work, io_tctx_exit_cb);
2814 * Some may use context even when all refs and requests have been put,
2815 * and they are free to do so while still holding uring_lock or
2816 * completion_lock, see io_req_task_submit(). Apart from other work,
2817 * this lock/unlock section also waits them to finish.
2819 mutex_lock(&ctx->uring_lock);
2820 while (!list_empty(&ctx->tctx_list)) {
2821 WARN_ON_ONCE(time_after(jiffies, timeout));
2823 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
2825 /* don't spin on a single task if cancellation failed */
2826 list_rotate_left(&ctx->tctx_list);
2827 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
2828 if (WARN_ON_ONCE(ret))
2831 mutex_unlock(&ctx->uring_lock);
2832 wait_for_completion(&exit.completion);
2833 mutex_lock(&ctx->uring_lock);
2835 mutex_unlock(&ctx->uring_lock);
2836 spin_lock(&ctx->completion_lock);
2837 spin_unlock(&ctx->completion_lock);
2839 io_ring_ctx_free(ctx);
2842 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
2844 unsigned long index;
2845 struct creds *creds;
2847 mutex_lock(&ctx->uring_lock);
2848 percpu_ref_kill(&ctx->refs);
2850 __io_cqring_overflow_flush(ctx, true);
2851 xa_for_each(&ctx->personalities, index, creds)
2852 io_unregister_personality(ctx, index);
2854 io_poll_remove_all(ctx, NULL, true);
2855 mutex_unlock(&ctx->uring_lock);
2858 * If we failed setting up the ctx, we might not have any rings
2859 * and therefore did not submit any requests
2862 io_kill_timeouts(ctx, NULL, true);
2864 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
2866 * Use system_unbound_wq to avoid spawning tons of event kworkers
2867 * if we're exiting a ton of rings at the same time. It just adds
2868 * noise and overhead, there's no discernable change in runtime
2869 * over using system_wq.
2871 queue_work(system_unbound_wq, &ctx->exit_work);
2874 static int io_uring_release(struct inode *inode, struct file *file)
2876 struct io_ring_ctx *ctx = file->private_data;
2878 file->private_data = NULL;
2879 io_ring_ctx_wait_and_kill(ctx);
2883 struct io_task_cancel {
2884 struct task_struct *task;
2888 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
2890 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2891 struct io_task_cancel *cancel = data;
2893 return io_match_task_safe(req, cancel->task, cancel->all);
2896 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
2897 struct task_struct *task,
2900 struct io_defer_entry *de;
2903 spin_lock(&ctx->completion_lock);
2904 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
2905 if (io_match_task_safe(de->req, task, cancel_all)) {
2906 list_cut_position(&list, &ctx->defer_list, &de->list);
2910 spin_unlock(&ctx->completion_lock);
2911 if (list_empty(&list))
2914 while (!list_empty(&list)) {
2915 de = list_first_entry(&list, struct io_defer_entry, list);
2916 list_del_init(&de->list);
2917 io_req_task_queue_fail(de->req, -ECANCELED);
2923 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
2925 struct io_tctx_node *node;
2926 enum io_wq_cancel cret;
2929 mutex_lock(&ctx->uring_lock);
2930 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
2931 struct io_uring_task *tctx = node->task->io_uring;
2934 * io_wq will stay alive while we hold uring_lock, because it's
2935 * killed after ctx nodes, which requires to take the lock.
2937 if (!tctx || !tctx->io_wq)
2939 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
2940 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2942 mutex_unlock(&ctx->uring_lock);
2947 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
2948 struct task_struct *task,
2951 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
2952 struct io_uring_task *tctx = task ? task->io_uring : NULL;
2953 enum io_wq_cancel cret;
2956 /* failed during ring init, it couldn't have issued any requests */
2961 ret |= io_uring_try_cancel_iowq(ctx);
2962 } else if (tctx && tctx->io_wq) {
2964 * Cancels requests of all rings, not only @ctx, but
2965 * it's fine as the task is in exit/exec.
2967 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
2969 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2972 /* SQPOLL thread does its own polling */
2973 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
2974 (ctx->sq_data && ctx->sq_data->thread == current)) {
2975 while (!wq_list_empty(&ctx->iopoll_list)) {
2976 io_iopoll_try_reap_events(ctx);
2981 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
2982 ret |= io_run_local_work(ctx) > 0;
2983 ret |= io_cancel_defer_files(ctx, task, cancel_all);
2984 mutex_lock(&ctx->uring_lock);
2985 ret |= io_poll_remove_all(ctx, task, cancel_all);
2986 mutex_unlock(&ctx->uring_lock);
2987 ret |= io_kill_timeouts(ctx, task, cancel_all);
2989 ret |= io_run_task_work() > 0;
2993 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
2996 return atomic_read(&tctx->inflight_tracked);
2997 return percpu_counter_sum(&tctx->inflight);
3001 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3002 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3004 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
3006 struct io_uring_task *tctx = current->io_uring;
3007 struct io_ring_ctx *ctx;
3011 WARN_ON_ONCE(sqd && sqd->thread != current);
3013 if (!current->io_uring)
3016 io_wq_exit_start(tctx->io_wq);
3018 atomic_inc(&tctx->in_idle);
3022 io_uring_drop_tctx_refs(current);
3023 /* read completions before cancelations */
3024 inflight = tctx_inflight(tctx, !cancel_all);
3029 struct io_tctx_node *node;
3030 unsigned long index;
3032 xa_for_each(&tctx->xa, index, node) {
3033 /* sqpoll task will cancel all its requests */
3034 if (node->ctx->sq_data)
3036 loop |= io_uring_try_cancel_requests(node->ctx,
3037 current, cancel_all);
3040 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3041 loop |= io_uring_try_cancel_requests(ctx,
3051 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3053 io_uring_drop_tctx_refs(current);
3056 * If we've seen completions, retry without waiting. This
3057 * avoids a race where a completion comes in before we did
3058 * prepare_to_wait().
3060 if (inflight == tctx_inflight(tctx, !cancel_all))
3062 finish_wait(&tctx->wait, &wait);
3065 io_uring_clean_tctx(tctx);
3068 * We shouldn't run task_works after cancel, so just leave
3069 * ->in_idle set for normal exit.
3071 atomic_dec(&tctx->in_idle);
3072 /* for exec all current's requests should be gone, kill tctx */
3073 __io_uring_free(current);
3077 void __io_uring_cancel(bool cancel_all)
3079 io_uring_cancel_generic(cancel_all, NULL);
3082 static void *io_uring_validate_mmap_request(struct file *file,
3083 loff_t pgoff, size_t sz)
3085 struct io_ring_ctx *ctx = file->private_data;
3086 loff_t offset = pgoff << PAGE_SHIFT;
3091 case IORING_OFF_SQ_RING:
3092 case IORING_OFF_CQ_RING:
3095 case IORING_OFF_SQES:
3099 return ERR_PTR(-EINVAL);
3102 page = virt_to_head_page(ptr);
3103 if (sz > page_size(page))
3104 return ERR_PTR(-EINVAL);
3111 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3113 size_t sz = vma->vm_end - vma->vm_start;
3117 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3119 return PTR_ERR(ptr);
3121 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3122 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3125 #else /* !CONFIG_MMU */
3127 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3129 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
3132 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3134 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3137 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3138 unsigned long addr, unsigned long len,
3139 unsigned long pgoff, unsigned long flags)
3143 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3145 return PTR_ERR(ptr);
3147 return (unsigned long) ptr;
3150 #endif /* !CONFIG_MMU */
3152 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3154 if (flags & IORING_ENTER_EXT_ARG) {
3155 struct io_uring_getevents_arg arg;
3157 if (argsz != sizeof(arg))
3159 if (copy_from_user(&arg, argp, sizeof(arg)))
3165 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3166 struct __kernel_timespec __user **ts,
3167 const sigset_t __user **sig)
3169 struct io_uring_getevents_arg arg;
3172 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3173 * is just a pointer to the sigset_t.
3175 if (!(flags & IORING_ENTER_EXT_ARG)) {
3176 *sig = (const sigset_t __user *) argp;
3182 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3183 * timespec and sigset_t pointers if good.
3185 if (*argsz != sizeof(arg))
3187 if (copy_from_user(&arg, argp, sizeof(arg)))
3191 *sig = u64_to_user_ptr(arg.sigmask);
3192 *argsz = arg.sigmask_sz;
3193 *ts = u64_to_user_ptr(arg.ts);
3197 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3198 u32, min_complete, u32, flags, const void __user *, argp,
3201 struct io_ring_ctx *ctx;
3205 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3206 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3207 IORING_ENTER_REGISTERED_RING)))
3211 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3212 * need only dereference our task private array to find it.
3214 if (flags & IORING_ENTER_REGISTERED_RING) {
3215 struct io_uring_task *tctx = current->io_uring;
3217 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3219 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3220 f.file = tctx->registered_rings[fd];
3222 if (unlikely(!f.file))
3226 if (unlikely(!f.file))
3229 if (unlikely(!io_is_uring_fops(f.file)))
3233 ctx = f.file->private_data;
3235 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3239 * For SQ polling, the thread will do all submissions and completions.
3240 * Just return the requested submit count, and wake the thread if
3244 if (ctx->flags & IORING_SETUP_SQPOLL) {
3245 io_cqring_overflow_flush(ctx);
3247 if (unlikely(ctx->sq_data->thread == NULL)) {
3251 if (flags & IORING_ENTER_SQ_WAKEUP)
3252 wake_up(&ctx->sq_data->wait);
3253 if (flags & IORING_ENTER_SQ_WAIT) {
3254 ret = io_sqpoll_wait_sq(ctx);
3259 } else if (to_submit) {
3260 ret = io_uring_add_tctx_node(ctx);
3264 mutex_lock(&ctx->uring_lock);
3265 ret = io_submit_sqes(ctx, to_submit);
3266 if (ret != to_submit) {
3267 mutex_unlock(&ctx->uring_lock);
3270 if (flags & IORING_ENTER_GETEVENTS) {
3271 if (ctx->syscall_iopoll)
3274 * Ignore errors, we'll soon call io_cqring_wait() and
3275 * it should handle ownership problems if any.
3277 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3278 (void)io_run_local_work_locked(ctx);
3280 mutex_unlock(&ctx->uring_lock);
3283 if (flags & IORING_ENTER_GETEVENTS) {
3286 if (ctx->syscall_iopoll) {
3288 * We disallow the app entering submit/complete with
3289 * polling, but we still need to lock the ring to
3290 * prevent racing with polled issue that got punted to
3293 mutex_lock(&ctx->uring_lock);
3295 ret2 = io_validate_ext_arg(flags, argp, argsz);
3296 if (likely(!ret2)) {
3297 min_complete = min(min_complete,
3299 ret2 = io_iopoll_check(ctx, min_complete);
3301 mutex_unlock(&ctx->uring_lock);
3303 const sigset_t __user *sig;
3304 struct __kernel_timespec __user *ts;
3306 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3307 if (likely(!ret2)) {
3308 min_complete = min(min_complete,
3310 ret2 = io_cqring_wait(ctx, min_complete, sig,
3319 * EBADR indicates that one or more CQE were dropped.
3320 * Once the user has been informed we can clear the bit
3321 * as they are obviously ok with those drops.
3323 if (unlikely(ret2 == -EBADR))
3324 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3333 static const struct file_operations io_uring_fops = {
3334 .release = io_uring_release,
3335 .mmap = io_uring_mmap,
3337 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3338 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3340 .poll = io_uring_poll,
3341 #ifdef CONFIG_PROC_FS
3342 .show_fdinfo = io_uring_show_fdinfo,
3346 bool io_is_uring_fops(struct file *file)
3348 return file->f_op == &io_uring_fops;
3351 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3352 struct io_uring_params *p)
3354 struct io_rings *rings;
3355 size_t size, sq_array_offset;
3357 /* make sure these are sane, as we already accounted them */
3358 ctx->sq_entries = p->sq_entries;
3359 ctx->cq_entries = p->cq_entries;
3361 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3362 if (size == SIZE_MAX)
3365 rings = io_mem_alloc(size);
3370 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3371 rings->sq_ring_mask = p->sq_entries - 1;
3372 rings->cq_ring_mask = p->cq_entries - 1;
3373 rings->sq_ring_entries = p->sq_entries;
3374 rings->cq_ring_entries = p->cq_entries;
3376 if (p->flags & IORING_SETUP_SQE128)
3377 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3379 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3380 if (size == SIZE_MAX) {
3381 io_mem_free(ctx->rings);
3386 ctx->sq_sqes = io_mem_alloc(size);
3387 if (!ctx->sq_sqes) {
3388 io_mem_free(ctx->rings);
3396 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3400 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3404 ret = __io_uring_add_tctx_node(ctx);
3409 fd_install(fd, file);
3414 * Allocate an anonymous fd, this is what constitutes the application
3415 * visible backing of an io_uring instance. The application mmaps this
3416 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3417 * we have to tie this fd to a socket for file garbage collection purposes.
3419 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3422 #if defined(CONFIG_UNIX)
3425 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3428 return ERR_PTR(ret);
3431 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3432 O_RDWR | O_CLOEXEC, NULL);
3433 #if defined(CONFIG_UNIX)
3435 sock_release(ctx->ring_sock);
3436 ctx->ring_sock = NULL;
3438 ctx->ring_sock->file = file;
3444 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3445 struct io_uring_params __user *params)
3447 struct io_ring_ctx *ctx;
3453 if (entries > IORING_MAX_ENTRIES) {
3454 if (!(p->flags & IORING_SETUP_CLAMP))
3456 entries = IORING_MAX_ENTRIES;
3460 * Use twice as many entries for the CQ ring. It's possible for the
3461 * application to drive a higher depth than the size of the SQ ring,
3462 * since the sqes are only used at submission time. This allows for
3463 * some flexibility in overcommitting a bit. If the application has
3464 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3465 * of CQ ring entries manually.
3467 p->sq_entries = roundup_pow_of_two(entries);
3468 if (p->flags & IORING_SETUP_CQSIZE) {
3470 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3471 * to a power-of-two, if it isn't already. We do NOT impose
3472 * any cq vs sq ring sizing.
3476 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3477 if (!(p->flags & IORING_SETUP_CLAMP))
3479 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3481 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3482 if (p->cq_entries < p->sq_entries)
3485 p->cq_entries = 2 * p->sq_entries;
3488 ctx = io_ring_ctx_alloc(p);
3493 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3494 * space applications don't need to do io completion events
3495 * polling again, they can rely on io_sq_thread to do polling
3496 * work, which can reduce cpu usage and uring_lock contention.
3498 if (ctx->flags & IORING_SETUP_IOPOLL &&
3499 !(ctx->flags & IORING_SETUP_SQPOLL))
3500 ctx->syscall_iopoll = 1;
3502 ctx->compat = in_compat_syscall();
3503 if (!capable(CAP_IPC_LOCK))
3504 ctx->user = get_uid(current_user());
3507 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3508 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3511 if (ctx->flags & IORING_SETUP_SQPOLL) {
3512 /* IPI related flags don't make sense with SQPOLL */
3513 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3514 IORING_SETUP_TASKRUN_FLAG |
3515 IORING_SETUP_DEFER_TASKRUN))
3517 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3518 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3519 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3521 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3522 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3524 ctx->notify_method = TWA_SIGNAL;
3528 * For DEFER_TASKRUN we require the completion task to be the same as the
3529 * submission task. This implies that there is only one submitter, so enforce
3532 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3533 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3538 * This is just grabbed for accounting purposes. When a process exits,
3539 * the mm is exited and dropped before the files, hence we need to hang
3540 * on to this mm purely for the purposes of being able to unaccount
3541 * memory (locked/pinned vm). It's not used for anything else.
3543 mmgrab(current->mm);
3544 ctx->mm_account = current->mm;
3546 ret = io_allocate_scq_urings(ctx, p);
3550 ret = io_sq_offload_create(ctx, p);
3553 /* always set a rsrc node */
3554 ret = io_rsrc_node_switch_start(ctx);
3557 io_rsrc_node_switch(ctx, NULL);
3559 memset(&p->sq_off, 0, sizeof(p->sq_off));
3560 p->sq_off.head = offsetof(struct io_rings, sq.head);
3561 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3562 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3563 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3564 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3565 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3566 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3568 memset(&p->cq_off, 0, sizeof(p->cq_off));
3569 p->cq_off.head = offsetof(struct io_rings, cq.head);
3570 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3571 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3572 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3573 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3574 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3575 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3577 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3578 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3579 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3580 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3581 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3582 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3583 IORING_FEAT_LINKED_FILE;
3585 if (copy_to_user(params, p, sizeof(*p))) {
3590 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
3591 && !(ctx->flags & IORING_SETUP_R_DISABLED))
3592 ctx->submitter_task = get_task_struct(current);
3594 file = io_uring_get_file(ctx);
3596 ret = PTR_ERR(file);
3601 * Install ring fd as the very last thing, so we don't risk someone
3602 * having closed it before we finish setup
3604 ret = io_uring_install_fd(ctx, file);
3606 /* fput will clean it up */
3611 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3614 io_ring_ctx_wait_and_kill(ctx);
3619 * Sets up an aio uring context, and returns the fd. Applications asks for a
3620 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3621 * params structure passed in.
3623 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3625 struct io_uring_params p;
3628 if (copy_from_user(&p, params, sizeof(p)))
3630 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3635 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3636 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3637 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3638 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3639 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3640 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3641 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN))
3644 return io_uring_create(entries, &p, params);
3647 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3648 struct io_uring_params __user *, params)
3650 return io_uring_setup(entries, params);
3653 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3656 struct io_uring_probe *p;
3660 size = struct_size(p, ops, nr_args);
3661 if (size == SIZE_MAX)
3663 p = kzalloc(size, GFP_KERNEL);
3668 if (copy_from_user(p, arg, size))
3671 if (memchr_inv(p, 0, size))
3674 p->last_op = IORING_OP_LAST - 1;
3675 if (nr_args > IORING_OP_LAST)
3676 nr_args = IORING_OP_LAST;
3678 for (i = 0; i < nr_args; i++) {
3680 if (!io_op_defs[i].not_supported)
3681 p->ops[i].flags = IO_URING_OP_SUPPORTED;
3686 if (copy_to_user(arg, p, size))
3693 static int io_register_personality(struct io_ring_ctx *ctx)
3695 const struct cred *creds;
3699 creds = get_current_cred();
3701 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
3702 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
3710 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
3711 void __user *arg, unsigned int nr_args)
3713 struct io_uring_restriction *res;
3717 /* Restrictions allowed only if rings started disabled */
3718 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3721 /* We allow only a single restrictions registration */
3722 if (ctx->restrictions.registered)
3725 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
3728 size = array_size(nr_args, sizeof(*res));
3729 if (size == SIZE_MAX)
3732 res = memdup_user(arg, size);
3734 return PTR_ERR(res);
3738 for (i = 0; i < nr_args; i++) {
3739 switch (res[i].opcode) {
3740 case IORING_RESTRICTION_REGISTER_OP:
3741 if (res[i].register_op >= IORING_REGISTER_LAST) {
3746 __set_bit(res[i].register_op,
3747 ctx->restrictions.register_op);
3749 case IORING_RESTRICTION_SQE_OP:
3750 if (res[i].sqe_op >= IORING_OP_LAST) {
3755 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
3757 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
3758 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
3760 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
3761 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
3770 /* Reset all restrictions if an error happened */
3772 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
3774 ctx->restrictions.registered = true;
3780 static int io_register_enable_rings(struct io_ring_ctx *ctx)
3782 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3785 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task)
3786 ctx->submitter_task = get_task_struct(current);
3788 if (ctx->restrictions.registered)
3789 ctx->restricted = 1;
3791 ctx->flags &= ~IORING_SETUP_R_DISABLED;
3792 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
3793 wake_up(&ctx->sq_data->wait);
3797 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
3798 void __user *arg, unsigned len)
3800 struct io_uring_task *tctx = current->io_uring;
3801 cpumask_var_t new_mask;
3804 if (!tctx || !tctx->io_wq)
3807 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3810 cpumask_clear(new_mask);
3811 if (len > cpumask_size())
3812 len = cpumask_size();
3814 if (in_compat_syscall()) {
3815 ret = compat_get_bitmap(cpumask_bits(new_mask),
3816 (const compat_ulong_t __user *)arg,
3817 len * 8 /* CHAR_BIT */);
3819 ret = copy_from_user(new_mask, arg, len);
3823 free_cpumask_var(new_mask);
3827 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
3828 free_cpumask_var(new_mask);
3832 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
3834 struct io_uring_task *tctx = current->io_uring;
3836 if (!tctx || !tctx->io_wq)
3839 return io_wq_cpu_affinity(tctx->io_wq, NULL);
3842 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
3844 __must_hold(&ctx->uring_lock)
3846 struct io_tctx_node *node;
3847 struct io_uring_task *tctx = NULL;
3848 struct io_sq_data *sqd = NULL;
3852 if (copy_from_user(new_count, arg, sizeof(new_count)))
3854 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3855 if (new_count[i] > INT_MAX)
3858 if (ctx->flags & IORING_SETUP_SQPOLL) {
3862 * Observe the correct sqd->lock -> ctx->uring_lock
3863 * ordering. Fine to drop uring_lock here, we hold
3866 refcount_inc(&sqd->refs);
3867 mutex_unlock(&ctx->uring_lock);
3868 mutex_lock(&sqd->lock);
3869 mutex_lock(&ctx->uring_lock);
3871 tctx = sqd->thread->io_uring;
3874 tctx = current->io_uring;
3877 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
3879 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3881 ctx->iowq_limits[i] = new_count[i];
3882 ctx->iowq_limits_set = true;
3884 if (tctx && tctx->io_wq) {
3885 ret = io_wq_max_workers(tctx->io_wq, new_count);
3889 memset(new_count, 0, sizeof(new_count));
3893 mutex_unlock(&sqd->lock);
3894 io_put_sq_data(sqd);
3897 if (copy_to_user(arg, new_count, sizeof(new_count)))
3900 /* that's it for SQPOLL, only the SQPOLL task creates requests */
3904 /* now propagate the restriction to all registered users */
3905 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3906 struct io_uring_task *tctx = node->task->io_uring;
3908 if (WARN_ON_ONCE(!tctx->io_wq))
3911 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3912 new_count[i] = ctx->iowq_limits[i];
3913 /* ignore errors, it always returns zero anyway */
3914 (void)io_wq_max_workers(tctx->io_wq, new_count);
3919 mutex_unlock(&sqd->lock);
3920 io_put_sq_data(sqd);
3925 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
3926 void __user *arg, unsigned nr_args)
3927 __releases(ctx->uring_lock)
3928 __acquires(ctx->uring_lock)
3933 * We don't quiesce the refs for register anymore and so it can't be
3934 * dying as we're holding a file ref here.
3936 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
3939 if (ctx->submitter_task && ctx->submitter_task != current)
3942 if (ctx->restricted) {
3943 if (opcode >= IORING_REGISTER_LAST)
3945 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
3946 if (!test_bit(opcode, ctx->restrictions.register_op))
3951 case IORING_REGISTER_BUFFERS:
3955 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
3957 case IORING_UNREGISTER_BUFFERS:
3961 ret = io_sqe_buffers_unregister(ctx);
3963 case IORING_REGISTER_FILES:
3967 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
3969 case IORING_UNREGISTER_FILES:
3973 ret = io_sqe_files_unregister(ctx);
3975 case IORING_REGISTER_FILES_UPDATE:
3976 ret = io_register_files_update(ctx, arg, nr_args);
3978 case IORING_REGISTER_EVENTFD:
3982 ret = io_eventfd_register(ctx, arg, 0);
3984 case IORING_REGISTER_EVENTFD_ASYNC:
3988 ret = io_eventfd_register(ctx, arg, 1);
3990 case IORING_UNREGISTER_EVENTFD:
3994 ret = io_eventfd_unregister(ctx);
3996 case IORING_REGISTER_PROBE:
3998 if (!arg || nr_args > 256)
4000 ret = io_probe(ctx, arg, nr_args);
4002 case IORING_REGISTER_PERSONALITY:
4006 ret = io_register_personality(ctx);
4008 case IORING_UNREGISTER_PERSONALITY:
4012 ret = io_unregister_personality(ctx, nr_args);
4014 case IORING_REGISTER_ENABLE_RINGS:
4018 ret = io_register_enable_rings(ctx);
4020 case IORING_REGISTER_RESTRICTIONS:
4021 ret = io_register_restrictions(ctx, arg, nr_args);
4023 case IORING_REGISTER_FILES2:
4024 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
4026 case IORING_REGISTER_FILES_UPDATE2:
4027 ret = io_register_rsrc_update(ctx, arg, nr_args,
4030 case IORING_REGISTER_BUFFERS2:
4031 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
4033 case IORING_REGISTER_BUFFERS_UPDATE:
4034 ret = io_register_rsrc_update(ctx, arg, nr_args,
4035 IORING_RSRC_BUFFER);
4037 case IORING_REGISTER_IOWQ_AFF:
4039 if (!arg || !nr_args)
4041 ret = io_register_iowq_aff(ctx, arg, nr_args);
4043 case IORING_UNREGISTER_IOWQ_AFF:
4047 ret = io_unregister_iowq_aff(ctx);
4049 case IORING_REGISTER_IOWQ_MAX_WORKERS:
4051 if (!arg || nr_args != 2)
4053 ret = io_register_iowq_max_workers(ctx, arg);
4055 case IORING_REGISTER_RING_FDS:
4056 ret = io_ringfd_register(ctx, arg, nr_args);
4058 case IORING_UNREGISTER_RING_FDS:
4059 ret = io_ringfd_unregister(ctx, arg, nr_args);
4061 case IORING_REGISTER_PBUF_RING:
4063 if (!arg || nr_args != 1)
4065 ret = io_register_pbuf_ring(ctx, arg);
4067 case IORING_UNREGISTER_PBUF_RING:
4069 if (!arg || nr_args != 1)
4071 ret = io_unregister_pbuf_ring(ctx, arg);
4073 case IORING_REGISTER_SYNC_CANCEL:
4075 if (!arg || nr_args != 1)
4077 ret = io_sync_cancel(ctx, arg);
4079 case IORING_REGISTER_FILE_ALLOC_RANGE:
4081 if (!arg || nr_args)
4083 ret = io_register_file_alloc_range(ctx, arg);
4093 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
4094 void __user *, arg, unsigned int, nr_args)
4096 struct io_ring_ctx *ctx;
4105 if (!io_is_uring_fops(f.file))
4108 ctx = f.file->private_data;
4110 mutex_lock(&ctx->uring_lock);
4111 ret = __io_uring_register(ctx, opcode, arg, nr_args);
4112 mutex_unlock(&ctx->uring_lock);
4113 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
4119 static int __init io_uring_init(void)
4121 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4122 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4123 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4126 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4127 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4128 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4129 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4130 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4131 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
4132 BUILD_BUG_SQE_ELEM(1, __u8, flags);
4133 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
4134 BUILD_BUG_SQE_ELEM(4, __s32, fd);
4135 BUILD_BUG_SQE_ELEM(8, __u64, off);
4136 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
4137 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
4138 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
4139 BUILD_BUG_SQE_ELEM(16, __u64, addr);
4140 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
4141 BUILD_BUG_SQE_ELEM(24, __u32, len);
4142 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
4143 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
4144 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4145 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
4146 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
4147 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
4148 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
4149 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
4150 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
4151 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
4152 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
4153 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
4154 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
4155 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
4156 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
4157 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
4158 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
4159 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
4160 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
4161 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
4162 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
4163 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
4164 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
4165 BUILD_BUG_SQE_ELEM(42, __u16, personality);
4166 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
4167 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
4168 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
4169 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
4170 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
4171 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4172 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
4174 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4175 sizeof(struct io_uring_rsrc_update));
4176 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4177 sizeof(struct io_uring_rsrc_update2));
4179 /* ->buf_index is u16 */
4180 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4181 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4182 offsetof(struct io_uring_buf_ring, tail));
4184 /* should fit into one byte */
4185 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
4186 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4187 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4189 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
4191 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4193 io_uring_optable_init();
4195 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4199 __initcall(io_uring_init);