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 __io_submit_flush_completions(ctx);
174 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
176 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
179 static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx)
181 return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head);
184 static bool io_match_linked(struct io_kiocb *head)
186 struct io_kiocb *req;
188 io_for_each_link(req, head) {
189 if (req->flags & REQ_F_INFLIGHT)
196 * As io_match_task() but protected against racing with linked timeouts.
197 * User must not hold timeout_lock.
199 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
204 if (task && head->task != task)
209 if (head->flags & REQ_F_LINK_TIMEOUT) {
210 struct io_ring_ctx *ctx = head->ctx;
212 /* protect against races with linked timeouts */
213 spin_lock_irq(&ctx->timeout_lock);
214 matched = io_match_linked(head);
215 spin_unlock_irq(&ctx->timeout_lock);
217 matched = io_match_linked(head);
222 static inline void req_fail_link_node(struct io_kiocb *req, int res)
225 io_req_set_res(req, res, 0);
228 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
230 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
233 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
235 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
237 complete(&ctx->ref_comp);
240 static __cold void io_fallback_req_func(struct work_struct *work)
242 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
244 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
245 struct io_kiocb *req, *tmp;
248 percpu_ref_get(&ctx->refs);
249 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
250 req->io_task_work.func(req, &locked);
253 io_submit_flush_completions(ctx);
254 mutex_unlock(&ctx->uring_lock);
256 percpu_ref_put(&ctx->refs);
259 static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
261 unsigned hash_buckets = 1U << bits;
262 size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
264 table->hbs = kmalloc(hash_size, GFP_KERNEL);
268 table->hash_bits = bits;
269 init_hash_table(table, hash_buckets);
273 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
275 struct io_ring_ctx *ctx;
278 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
282 xa_init(&ctx->io_bl_xa);
285 * Use 5 bits less than the max cq entries, that should give us around
286 * 32 entries per hash list if totally full and uniformly spread, but
287 * don't keep too many buckets to not overconsume memory.
289 hash_bits = ilog2(p->cq_entries) - 5;
290 hash_bits = clamp(hash_bits, 1, 8);
291 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
293 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
296 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
297 if (!ctx->dummy_ubuf)
299 /* set invalid range, so io_import_fixed() fails meeting it */
300 ctx->dummy_ubuf->ubuf = -1UL;
302 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
306 ctx->flags = p->flags;
307 init_waitqueue_head(&ctx->sqo_sq_wait);
308 INIT_LIST_HEAD(&ctx->sqd_list);
309 INIT_LIST_HEAD(&ctx->cq_overflow_list);
310 INIT_LIST_HEAD(&ctx->io_buffers_cache);
311 io_alloc_cache_init(&ctx->apoll_cache);
312 io_alloc_cache_init(&ctx->netmsg_cache);
313 init_completion(&ctx->ref_comp);
314 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
315 mutex_init(&ctx->uring_lock);
316 init_waitqueue_head(&ctx->cq_wait);
317 spin_lock_init(&ctx->completion_lock);
318 spin_lock_init(&ctx->timeout_lock);
319 INIT_WQ_LIST(&ctx->iopoll_list);
320 INIT_LIST_HEAD(&ctx->io_buffers_pages);
321 INIT_LIST_HEAD(&ctx->io_buffers_comp);
322 INIT_LIST_HEAD(&ctx->defer_list);
323 INIT_LIST_HEAD(&ctx->timeout_list);
324 INIT_LIST_HEAD(&ctx->ltimeout_list);
325 spin_lock_init(&ctx->rsrc_ref_lock);
326 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
327 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
328 init_llist_head(&ctx->rsrc_put_llist);
329 init_llist_head(&ctx->work_llist);
330 INIT_LIST_HEAD(&ctx->tctx_list);
331 ctx->submit_state.free_list.next = NULL;
332 INIT_WQ_LIST(&ctx->locked_free_list);
333 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
334 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
337 kfree(ctx->dummy_ubuf);
338 kfree(ctx->cancel_table.hbs);
339 kfree(ctx->cancel_table_locked.hbs);
341 xa_destroy(&ctx->io_bl_xa);
346 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
348 struct io_rings *r = ctx->rings;
350 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
354 static bool req_need_defer(struct io_kiocb *req, u32 seq)
356 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
357 struct io_ring_ctx *ctx = req->ctx;
359 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
365 static inline void io_req_track_inflight(struct io_kiocb *req)
367 if (!(req->flags & REQ_F_INFLIGHT)) {
368 req->flags |= REQ_F_INFLIGHT;
369 atomic_inc(&req->task->io_uring->inflight_tracked);
373 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
375 if (WARN_ON_ONCE(!req->link))
378 req->flags &= ~REQ_F_ARM_LTIMEOUT;
379 req->flags |= REQ_F_LINK_TIMEOUT;
381 /* linked timeouts should have two refs once prep'ed */
382 io_req_set_refcount(req);
383 __io_req_set_refcount(req->link, 2);
387 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
389 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
391 return __io_prep_linked_timeout(req);
394 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
396 io_queue_linked_timeout(__io_prep_linked_timeout(req));
399 static inline void io_arm_ltimeout(struct io_kiocb *req)
401 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
402 __io_arm_ltimeout(req);
405 static void io_prep_async_work(struct io_kiocb *req)
407 const struct io_op_def *def = &io_op_defs[req->opcode];
408 struct io_ring_ctx *ctx = req->ctx;
410 if (!(req->flags & REQ_F_CREDS)) {
411 req->flags |= REQ_F_CREDS;
412 req->creds = get_current_cred();
415 req->work.list.next = NULL;
417 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
418 if (req->flags & REQ_F_FORCE_ASYNC)
419 req->work.flags |= IO_WQ_WORK_CONCURRENT;
421 if (req->file && !io_req_ffs_set(req))
422 req->flags |= io_file_get_flags(req->file) << REQ_F_SUPPORT_NOWAIT_BIT;
424 if (req->flags & REQ_F_ISREG) {
425 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
426 io_wq_hash_work(&req->work, file_inode(req->file));
427 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
428 if (def->unbound_nonreg_file)
429 req->work.flags |= IO_WQ_WORK_UNBOUND;
433 static void io_prep_async_link(struct io_kiocb *req)
435 struct io_kiocb *cur;
437 if (req->flags & REQ_F_LINK_TIMEOUT) {
438 struct io_ring_ctx *ctx = req->ctx;
440 spin_lock_irq(&ctx->timeout_lock);
441 io_for_each_link(cur, req)
442 io_prep_async_work(cur);
443 spin_unlock_irq(&ctx->timeout_lock);
445 io_for_each_link(cur, req)
446 io_prep_async_work(cur);
450 void io_queue_iowq(struct io_kiocb *req, bool *dont_use)
452 struct io_kiocb *link = io_prep_linked_timeout(req);
453 struct io_uring_task *tctx = req->task->io_uring;
456 BUG_ON(!tctx->io_wq);
458 /* init ->work of the whole link before punting */
459 io_prep_async_link(req);
462 * Not expected to happen, but if we do have a bug where this _can_
463 * happen, catch it here and ensure the request is marked as
464 * canceled. That will make io-wq go through the usual work cancel
465 * procedure rather than attempt to run this request (or create a new
468 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
469 req->work.flags |= IO_WQ_WORK_CANCEL;
471 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
472 io_wq_enqueue(tctx->io_wq, &req->work);
474 io_queue_linked_timeout(link);
477 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
479 while (!list_empty(&ctx->defer_list)) {
480 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
481 struct io_defer_entry, list);
483 if (req_need_defer(de->req, de->seq))
485 list_del_init(&de->list);
486 io_req_task_queue(de->req);
492 static void io_eventfd_ops(struct rcu_head *rcu)
494 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
495 int ops = atomic_xchg(&ev_fd->ops, 0);
497 if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT))
498 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
500 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
501 * ordering in a race but if references are 0 we know we have to free
504 if (atomic_dec_and_test(&ev_fd->refs)) {
505 eventfd_ctx_put(ev_fd->cq_ev_fd);
510 static void io_eventfd_signal(struct io_ring_ctx *ctx)
512 struct io_ev_fd *ev_fd = NULL;
516 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
519 ev_fd = rcu_dereference(ctx->io_ev_fd);
522 * Check again if ev_fd exists incase an io_eventfd_unregister call
523 * completed between the NULL check of ctx->io_ev_fd at the start of
524 * the function and rcu_read_lock.
526 if (unlikely(!ev_fd))
528 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
530 if (ev_fd->eventfd_async && !io_wq_current_is_worker())
533 if (likely(eventfd_signal_allowed())) {
534 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
536 atomic_inc(&ev_fd->refs);
537 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops))
538 call_rcu(&ev_fd->rcu, io_eventfd_ops);
540 atomic_dec(&ev_fd->refs);
547 static void io_eventfd_flush_signal(struct io_ring_ctx *ctx)
551 spin_lock(&ctx->completion_lock);
554 * Eventfd should only get triggered when at least one event has been
555 * posted. Some applications rely on the eventfd notification count
556 * only changing IFF a new CQE has been added to the CQ ring. There's
557 * no depedency on 1:1 relationship between how many times this
558 * function is called (and hence the eventfd count) and number of CQEs
559 * posted to the CQ ring.
561 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
562 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
563 spin_unlock(&ctx->completion_lock);
567 io_eventfd_signal(ctx);
570 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
572 if (ctx->off_timeout_used || ctx->drain_active) {
573 spin_lock(&ctx->completion_lock);
574 if (ctx->off_timeout_used)
575 io_flush_timeouts(ctx);
576 if (ctx->drain_active)
577 io_queue_deferred(ctx);
578 spin_unlock(&ctx->completion_lock);
581 io_eventfd_flush_signal(ctx);
584 static inline void io_cqring_ev_posted(struct io_ring_ctx *ctx)
586 io_commit_cqring_flush(ctx);
590 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
591 __releases(ctx->completion_lock)
593 io_commit_cqring(ctx);
594 spin_unlock(&ctx->completion_lock);
595 io_cqring_ev_posted(ctx);
598 void io_cq_unlock_post(struct io_ring_ctx *ctx)
600 __io_cq_unlock_post(ctx);
603 /* Returns true if there are no backlogged entries after the flush */
604 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
607 size_t cqe_size = sizeof(struct io_uring_cqe);
609 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
612 if (ctx->flags & IORING_SETUP_CQE32)
616 while (!list_empty(&ctx->cq_overflow_list)) {
617 struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true);
618 struct io_overflow_cqe *ocqe;
622 ocqe = list_first_entry(&ctx->cq_overflow_list,
623 struct io_overflow_cqe, list);
625 memcpy(cqe, &ocqe->cqe, cqe_size);
627 io_account_cq_overflow(ctx);
629 list_del(&ocqe->list);
633 all_flushed = list_empty(&ctx->cq_overflow_list);
635 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
636 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
639 io_cq_unlock_post(ctx);
643 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
647 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
648 /* iopoll syncs against uring_lock, not completion_lock */
649 if (ctx->flags & IORING_SETUP_IOPOLL)
650 mutex_lock(&ctx->uring_lock);
651 ret = __io_cqring_overflow_flush(ctx, false);
652 if (ctx->flags & IORING_SETUP_IOPOLL)
653 mutex_unlock(&ctx->uring_lock);
659 void __io_put_task(struct task_struct *task, int nr)
661 struct io_uring_task *tctx = task->io_uring;
663 percpu_counter_sub(&tctx->inflight, nr);
664 if (unlikely(atomic_read(&tctx->in_idle)))
665 wake_up(&tctx->wait);
666 put_task_struct_many(task, nr);
669 void io_task_refs_refill(struct io_uring_task *tctx)
671 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
673 percpu_counter_add(&tctx->inflight, refill);
674 refcount_add(refill, ¤t->usage);
675 tctx->cached_refs += refill;
678 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
680 struct io_uring_task *tctx = task->io_uring;
681 unsigned int refs = tctx->cached_refs;
684 tctx->cached_refs = 0;
685 percpu_counter_sub(&tctx->inflight, refs);
686 put_task_struct_many(task, refs);
690 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
691 s32 res, u32 cflags, u64 extra1, u64 extra2)
693 struct io_overflow_cqe *ocqe;
694 size_t ocq_size = sizeof(struct io_overflow_cqe);
695 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
698 ocq_size += sizeof(struct io_uring_cqe);
700 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
701 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
704 * If we're in ring overflow flush mode, or in task cancel mode,
705 * or cannot allocate an overflow entry, then we need to drop it
708 io_account_cq_overflow(ctx);
709 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
712 if (list_empty(&ctx->cq_overflow_list)) {
713 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
714 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
717 ocqe->cqe.user_data = user_data;
719 ocqe->cqe.flags = cflags;
721 ocqe->cqe.big_cqe[0] = extra1;
722 ocqe->cqe.big_cqe[1] = extra2;
724 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
728 bool io_req_cqe_overflow(struct io_kiocb *req)
730 if (!(req->flags & REQ_F_CQE32_INIT)) {
734 return io_cqring_event_overflow(req->ctx, req->cqe.user_data,
735 req->cqe.res, req->cqe.flags,
736 req->extra1, req->extra2);
740 * writes to the cq entry need to come after reading head; the
741 * control dependency is enough as we're using WRITE_ONCE to
744 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow)
746 struct io_rings *rings = ctx->rings;
747 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
748 unsigned int free, queued, len;
751 * Posting into the CQ when there are pending overflowed CQEs may break
752 * ordering guarantees, which will affect links, F_MORE users and more.
753 * Force overflow the completion.
755 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
758 /* userspace may cheat modifying the tail, be safe and do min */
759 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
760 free = ctx->cq_entries - queued;
761 /* we need a contiguous range, limit based on the current array offset */
762 len = min(free, ctx->cq_entries - off);
766 if (ctx->flags & IORING_SETUP_CQE32) {
771 ctx->cqe_cached = &rings->cqes[off];
772 ctx->cqe_sentinel = ctx->cqe_cached + len;
774 ctx->cached_cq_tail++;
776 if (ctx->flags & IORING_SETUP_CQE32)
778 return &rings->cqes[off];
781 bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
784 struct io_uring_cqe *cqe;
789 * If we can't get a cq entry, userspace overflowed the
790 * submission (by quite a lot). Increment the overflow count in
793 cqe = io_get_cqe(ctx);
795 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
797 WRITE_ONCE(cqe->user_data, user_data);
798 WRITE_ONCE(cqe->res, res);
799 WRITE_ONCE(cqe->flags, cflags);
801 if (ctx->flags & IORING_SETUP_CQE32) {
802 WRITE_ONCE(cqe->big_cqe[0], 0);
803 WRITE_ONCE(cqe->big_cqe[1], 0);
809 return io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
814 bool io_post_aux_cqe(struct io_ring_ctx *ctx,
815 u64 user_data, s32 res, u32 cflags,
821 filled = io_fill_cqe_aux(ctx, user_data, res, cflags, allow_overflow);
822 io_cq_unlock_post(ctx);
826 void io_req_complete_post(struct io_kiocb *req)
828 struct io_ring_ctx *ctx = req->ctx;
831 if (!(req->flags & REQ_F_CQE_SKIP))
832 __io_fill_cqe_req(ctx, req);
835 * If we're the last reference to this request, add to our locked
838 if (req_ref_put_and_test(req)) {
839 if (req->flags & IO_REQ_LINK_FLAGS) {
840 if (req->flags & IO_DISARM_MASK)
843 io_req_task_queue(req->link);
847 io_req_put_rsrc(req);
849 * Selected buffer deallocation in io_clean_op() assumes that
850 * we don't hold ->completion_lock. Clean them here to avoid
853 io_put_kbuf_comp(req);
854 io_dismantle_req(req);
855 io_put_task(req->task, 1);
856 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
857 ctx->locked_free_nr++;
859 io_cq_unlock_post(ctx);
862 inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags)
864 io_req_complete_post(req);
867 void io_req_complete_failed(struct io_kiocb *req, s32 res)
868 __must_hold(&ctx->uring_lock)
870 const struct io_op_def *def = &io_op_defs[req->opcode];
872 lockdep_assert_held(&req->ctx->uring_lock);
875 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
878 io_req_complete_post(req);
882 * Don't initialise the fields below on every allocation, but do that in
883 * advance and keep them valid across allocations.
885 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
889 req->async_data = NULL;
890 /* not necessary, but safer to zero */
894 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
895 struct io_submit_state *state)
897 spin_lock(&ctx->completion_lock);
898 wq_list_splice(&ctx->locked_free_list, &state->free_list);
899 ctx->locked_free_nr = 0;
900 spin_unlock(&ctx->completion_lock);
904 * A request might get retired back into the request caches even before opcode
905 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
906 * Because of that, io_alloc_req() should be called only under ->uring_lock
907 * and with extra caution to not get a request that is still worked on.
909 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
910 __must_hold(&ctx->uring_lock)
912 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
913 void *reqs[IO_REQ_ALLOC_BATCH];
917 * If we have more than a batch's worth of requests in our IRQ side
918 * locked cache, grab the lock and move them over to our submission
921 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
922 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
923 if (!io_req_cache_empty(ctx))
927 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
930 * Bulk alloc is all-or-nothing. If we fail to get a batch,
931 * retry single alloc to be on the safe side.
933 if (unlikely(ret <= 0)) {
934 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
940 percpu_ref_get_many(&ctx->refs, ret);
941 for (i = 0; i < ret; i++) {
942 struct io_kiocb *req = reqs[i];
944 io_preinit_req(req, ctx);
945 io_req_add_to_cache(req, ctx);
950 static inline void io_dismantle_req(struct io_kiocb *req)
952 unsigned int flags = req->flags;
954 if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
956 if (!(flags & REQ_F_FIXED_FILE))
957 io_put_file(req->file);
960 __cold void io_free_req(struct io_kiocb *req)
962 struct io_ring_ctx *ctx = req->ctx;
964 io_req_put_rsrc(req);
965 io_dismantle_req(req);
966 io_put_task(req->task, 1);
968 spin_lock(&ctx->completion_lock);
969 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
970 ctx->locked_free_nr++;
971 spin_unlock(&ctx->completion_lock);
974 static void __io_req_find_next_prep(struct io_kiocb *req)
976 struct io_ring_ctx *ctx = req->ctx;
980 io_cq_unlock_post(ctx);
983 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
985 struct io_kiocb *nxt;
988 * If LINK is set, we have dependent requests in this chain. If we
989 * didn't fail this request, queue the first one up, moving any other
990 * dependencies to the next request. In case of failure, fail the rest
993 if (unlikely(req->flags & IO_DISARM_MASK))
994 __io_req_find_next_prep(req);
1000 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
1004 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1005 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1007 io_submit_flush_completions(ctx);
1008 mutex_unlock(&ctx->uring_lock);
1011 percpu_ref_put(&ctx->refs);
1014 static unsigned int handle_tw_list(struct llist_node *node,
1015 struct io_ring_ctx **ctx, bool *locked,
1016 struct llist_node *last)
1018 unsigned int count = 0;
1020 while (node != last) {
1021 struct llist_node *next = node->next;
1022 struct io_kiocb *req = container_of(node, struct io_kiocb,
1025 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1027 if (req->ctx != *ctx) {
1028 ctx_flush_and_put(*ctx, locked);
1030 /* if not contended, grab and improve batching */
1031 *locked = mutex_trylock(&(*ctx)->uring_lock);
1032 percpu_ref_get(&(*ctx)->refs);
1033 } else if (!*locked)
1034 *locked = mutex_trylock(&(*ctx)->uring_lock);
1035 req->io_task_work.func(req, locked);
1038 if (unlikely(need_resched())) {
1039 ctx_flush_and_put(*ctx, locked);
1049 * io_llist_xchg - swap all entries in a lock-less list
1050 * @head: the head of lock-less list to delete all entries
1051 * @new: new entry as the head of the list
1053 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1054 * The order of entries returned is from the newest to the oldest added one.
1056 static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1057 struct llist_node *new)
1059 return xchg(&head->first, new);
1063 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1064 * @head: the head of lock-less list to delete all entries
1065 * @old: expected old value of the first entry of the list
1066 * @new: new entry as the head of the list
1068 * perform a cmpxchg on the first entry of the list.
1071 static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1072 struct llist_node *old,
1073 struct llist_node *new)
1075 return cmpxchg(&head->first, old, new);
1078 void tctx_task_work(struct callback_head *cb)
1080 bool uring_locked = false;
1081 struct io_ring_ctx *ctx = NULL;
1082 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1084 struct llist_node fake = {};
1085 struct llist_node *node = io_llist_xchg(&tctx->task_list, &fake);
1086 unsigned int loops = 1;
1087 unsigned int count = handle_tw_list(node, &ctx, &uring_locked, NULL);
1089 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1090 while (node != &fake) {
1092 node = io_llist_xchg(&tctx->task_list, &fake);
1093 count += handle_tw_list(node, &ctx, &uring_locked, &fake);
1094 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1097 ctx_flush_and_put(ctx, &uring_locked);
1099 /* relaxed read is enough as only the task itself sets ->in_idle */
1100 if (unlikely(atomic_read(&tctx->in_idle)))
1101 io_uring_drop_tctx_refs(current);
1103 trace_io_uring_task_work_run(tctx, count, loops);
1106 static void io_req_local_work_add(struct io_kiocb *req)
1108 struct io_ring_ctx *ctx = req->ctx;
1110 percpu_ref_get(&ctx->refs);
1112 if (!llist_add(&req->io_task_work.node, &ctx->work_llist)) {
1113 percpu_ref_put(&ctx->refs);
1116 /* need it for the following io_cqring_wake() */
1117 smp_mb__after_atomic();
1119 if (unlikely(atomic_read(&req->task->io_uring->in_idle))) {
1120 io_move_task_work_from_local(ctx);
1121 percpu_ref_put(&ctx->refs);
1125 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1126 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1129 io_eventfd_signal(ctx);
1130 __io_cqring_wake(ctx);
1131 percpu_ref_put(&ctx->refs);
1134 void __io_req_task_work_add(struct io_kiocb *req, bool allow_local)
1136 struct io_uring_task *tctx = req->task->io_uring;
1137 struct io_ring_ctx *ctx = req->ctx;
1138 struct llist_node *node;
1140 if (allow_local && ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
1141 io_req_local_work_add(req);
1145 /* task_work already pending, we're done */
1146 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1149 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1150 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1152 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1155 node = llist_del_all(&tctx->task_list);
1158 req = container_of(node, struct io_kiocb, io_task_work.node);
1160 if (llist_add(&req->io_task_work.node,
1161 &req->ctx->fallback_llist))
1162 schedule_delayed_work(&req->ctx->fallback_work, 1);
1166 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1168 struct llist_node *node;
1170 node = llist_del_all(&ctx->work_llist);
1172 struct io_kiocb *req = container_of(node, struct io_kiocb,
1176 __io_req_task_work_add(req, false);
1180 int __io_run_local_work(struct io_ring_ctx *ctx, bool *locked)
1182 struct llist_node *node;
1183 struct llist_node fake;
1184 struct llist_node *current_final = NULL;
1186 unsigned int loops = 1;
1188 if (unlikely(ctx->submitter_task != current))
1191 node = io_llist_xchg(&ctx->work_llist, &fake);
1194 while (node != current_final) {
1195 struct llist_node *next = node->next;
1196 struct io_kiocb *req = container_of(node, struct io_kiocb,
1198 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1199 req->io_task_work.func(req, locked);
1204 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1205 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1207 node = io_llist_cmpxchg(&ctx->work_llist, &fake, NULL);
1208 if (node != &fake) {
1210 current_final = &fake;
1211 node = io_llist_xchg(&ctx->work_llist, &fake);
1216 io_submit_flush_completions(ctx);
1217 trace_io_uring_local_work_run(ctx, ret, loops);
1222 int io_run_local_work(struct io_ring_ctx *ctx)
1227 if (llist_empty(&ctx->work_llist))
1230 __set_current_state(TASK_RUNNING);
1231 locked = mutex_trylock(&ctx->uring_lock);
1232 ret = __io_run_local_work(ctx, &locked);
1234 mutex_unlock(&ctx->uring_lock);
1239 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
1241 /* not needed for normal modes, but SQPOLL depends on it */
1242 io_tw_lock(req->ctx, locked);
1243 io_req_complete_failed(req, req->cqe.res);
1246 void io_req_task_submit(struct io_kiocb *req, bool *locked)
1248 io_tw_lock(req->ctx, locked);
1249 /* req->task == current here, checking PF_EXITING is safe */
1250 if (likely(!(req->task->flags & PF_EXITING)))
1253 io_req_complete_failed(req, -EFAULT);
1256 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1258 io_req_set_res(req, ret, 0);
1259 req->io_task_work.func = io_req_task_cancel;
1260 io_req_task_work_add(req);
1263 void io_req_task_queue(struct io_kiocb *req)
1265 req->io_task_work.func = io_req_task_submit;
1266 io_req_task_work_add(req);
1269 void io_queue_next(struct io_kiocb *req)
1271 struct io_kiocb *nxt = io_req_find_next(req);
1274 io_req_task_queue(nxt);
1277 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1278 __must_hold(&ctx->uring_lock)
1280 struct task_struct *task = NULL;
1284 struct io_kiocb *req = container_of(node, struct io_kiocb,
1287 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1288 if (req->flags & REQ_F_REFCOUNT) {
1289 node = req->comp_list.next;
1290 if (!req_ref_put_and_test(req))
1293 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1294 struct async_poll *apoll = req->apoll;
1296 if (apoll->double_poll)
1297 kfree(apoll->double_poll);
1298 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1300 req->flags &= ~REQ_F_POLLED;
1302 if (req->flags & IO_REQ_LINK_FLAGS)
1304 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1307 if (!(req->flags & REQ_F_FIXED_FILE))
1308 io_put_file(req->file);
1310 io_req_put_rsrc_locked(req, ctx);
1312 if (req->task != task) {
1314 io_put_task(task, task_refs);
1319 node = req->comp_list.next;
1320 io_req_add_to_cache(req, ctx);
1324 io_put_task(task, task_refs);
1327 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1328 __must_hold(&ctx->uring_lock)
1330 struct io_wq_work_node *node, *prev;
1331 struct io_submit_state *state = &ctx->submit_state;
1334 wq_list_for_each(node, prev, &state->compl_reqs) {
1335 struct io_kiocb *req = container_of(node, struct io_kiocb,
1338 if (!(req->flags & REQ_F_CQE_SKIP))
1339 __io_fill_cqe_req(ctx, req);
1341 __io_cq_unlock_post(ctx);
1343 io_free_batch_list(ctx, state->compl_reqs.first);
1344 INIT_WQ_LIST(&state->compl_reqs);
1348 * Drop reference to request, return next in chain (if there is one) if this
1349 * was the last reference to this request.
1351 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1353 struct io_kiocb *nxt = NULL;
1355 if (req_ref_put_and_test(req)) {
1356 if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
1357 nxt = io_req_find_next(req);
1363 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1365 /* See comment at the top of this file */
1367 return __io_cqring_events(ctx);
1371 * We can't just wait for polled events to come to us, we have to actively
1372 * find and complete them.
1374 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1376 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1379 mutex_lock(&ctx->uring_lock);
1380 while (!wq_list_empty(&ctx->iopoll_list)) {
1381 /* let it sleep and repeat later if can't complete a request */
1382 if (io_do_iopoll(ctx, true) == 0)
1385 * Ensure we allow local-to-the-cpu processing to take place,
1386 * in this case we need to ensure that we reap all events.
1387 * Also let task_work, etc. to progress by releasing the mutex
1389 if (need_resched()) {
1390 mutex_unlock(&ctx->uring_lock);
1392 mutex_lock(&ctx->uring_lock);
1395 mutex_unlock(&ctx->uring_lock);
1398 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1400 unsigned int nr_events = 0;
1402 unsigned long check_cq;
1404 if (!io_allowed_run_tw(ctx))
1407 check_cq = READ_ONCE(ctx->check_cq);
1408 if (unlikely(check_cq)) {
1409 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1410 __io_cqring_overflow_flush(ctx, false);
1412 * Similarly do not spin if we have not informed the user of any
1415 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1419 * Don't enter poll loop if we already have events pending.
1420 * If we do, we can potentially be spinning for commands that
1421 * already triggered a CQE (eg in error).
1423 if (io_cqring_events(ctx))
1428 * If a submit got punted to a workqueue, we can have the
1429 * application entering polling for a command before it gets
1430 * issued. That app will hold the uring_lock for the duration
1431 * of the poll right here, so we need to take a breather every
1432 * now and then to ensure that the issue has a chance to add
1433 * the poll to the issued list. Otherwise we can spin here
1434 * forever, while the workqueue is stuck trying to acquire the
1437 if (wq_list_empty(&ctx->iopoll_list) ||
1438 io_task_work_pending(ctx)) {
1439 u32 tail = ctx->cached_cq_tail;
1441 (void) io_run_local_work_locked(ctx);
1443 if (task_work_pending(current) ||
1444 wq_list_empty(&ctx->iopoll_list)) {
1445 mutex_unlock(&ctx->uring_lock);
1447 mutex_lock(&ctx->uring_lock);
1449 /* some requests don't go through iopoll_list */
1450 if (tail != ctx->cached_cq_tail ||
1451 wq_list_empty(&ctx->iopoll_list))
1454 ret = io_do_iopoll(ctx, !min);
1459 } while (nr_events < min && !need_resched());
1464 void io_req_task_complete(struct io_kiocb *req, bool *locked)
1466 if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) {
1467 unsigned issue_flags = *locked ? 0 : IO_URING_F_UNLOCKED;
1469 req->cqe.flags |= io_put_kbuf(req, issue_flags);
1473 io_req_complete_defer(req);
1475 io_req_complete_post(req);
1479 * After the iocb has been issued, it's safe to be found on the poll list.
1480 * Adding the kiocb to the list AFTER submission ensures that we don't
1481 * find it from a io_do_iopoll() thread before the issuer is done
1482 * accessing the kiocb cookie.
1484 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1486 struct io_ring_ctx *ctx = req->ctx;
1487 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1489 /* workqueue context doesn't hold uring_lock, grab it now */
1490 if (unlikely(needs_lock))
1491 mutex_lock(&ctx->uring_lock);
1494 * Track whether we have multiple files in our lists. This will impact
1495 * how we do polling eventually, not spinning if we're on potentially
1496 * different devices.
1498 if (wq_list_empty(&ctx->iopoll_list)) {
1499 ctx->poll_multi_queue = false;
1500 } else if (!ctx->poll_multi_queue) {
1501 struct io_kiocb *list_req;
1503 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1505 if (list_req->file != req->file)
1506 ctx->poll_multi_queue = true;
1510 * For fast devices, IO may have already completed. If it has, add
1511 * it to the front so we find it first.
1513 if (READ_ONCE(req->iopoll_completed))
1514 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1516 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1518 if (unlikely(needs_lock)) {
1520 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1521 * in sq thread task context or in io worker task context. If
1522 * current task context is sq thread, we don't need to check
1523 * whether should wake up sq thread.
1525 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1526 wq_has_sleeper(&ctx->sq_data->wait))
1527 wake_up(&ctx->sq_data->wait);
1529 mutex_unlock(&ctx->uring_lock);
1533 static bool io_bdev_nowait(struct block_device *bdev)
1535 return !bdev || bdev_nowait(bdev);
1539 * If we tracked the file through the SCM inflight mechanism, we could support
1540 * any file. For now, just ensure that anything potentially problematic is done
1543 static bool __io_file_supports_nowait(struct file *file, umode_t mode)
1545 if (S_ISBLK(mode)) {
1546 if (IS_ENABLED(CONFIG_BLOCK) &&
1547 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
1553 if (S_ISREG(mode)) {
1554 if (IS_ENABLED(CONFIG_BLOCK) &&
1555 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
1556 !io_is_uring_fops(file))
1561 /* any ->read/write should understand O_NONBLOCK */
1562 if (file->f_flags & O_NONBLOCK)
1564 return file->f_mode & FMODE_NOWAIT;
1568 * If we tracked the file through the SCM inflight mechanism, we could support
1569 * any file. For now, just ensure that anything potentially problematic is done
1572 unsigned int io_file_get_flags(struct file *file)
1574 umode_t mode = file_inode(file)->i_mode;
1575 unsigned int res = 0;
1579 if (__io_file_supports_nowait(file, mode))
1584 bool io_alloc_async_data(struct io_kiocb *req)
1586 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
1587 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
1588 if (req->async_data) {
1589 req->flags |= REQ_F_ASYNC_DATA;
1595 int io_req_prep_async(struct io_kiocb *req)
1597 const struct io_op_def *def = &io_op_defs[req->opcode];
1599 /* assign early for deferred execution for non-fixed file */
1600 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE) && !req->file)
1601 req->file = io_file_get_normal(req, req->cqe.fd);
1602 if (!def->prep_async)
1604 if (WARN_ON_ONCE(req_has_async_data(req)))
1606 if (!io_op_defs[req->opcode].manual_alloc) {
1607 if (io_alloc_async_data(req))
1610 return def->prep_async(req);
1613 static u32 io_get_sequence(struct io_kiocb *req)
1615 u32 seq = req->ctx->cached_sq_head;
1616 struct io_kiocb *cur;
1618 /* need original cached_sq_head, but it was increased for each req */
1619 io_for_each_link(cur, req)
1624 static __cold void io_drain_req(struct io_kiocb *req)
1625 __must_hold(&ctx->uring_lock)
1627 struct io_ring_ctx *ctx = req->ctx;
1628 struct io_defer_entry *de;
1630 u32 seq = io_get_sequence(req);
1632 /* Still need defer if there is pending req in defer list. */
1633 spin_lock(&ctx->completion_lock);
1634 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1635 spin_unlock(&ctx->completion_lock);
1637 ctx->drain_active = false;
1638 io_req_task_queue(req);
1641 spin_unlock(&ctx->completion_lock);
1643 io_prep_async_link(req);
1644 de = kmalloc(sizeof(*de), GFP_KERNEL);
1647 io_req_complete_failed(req, ret);
1651 spin_lock(&ctx->completion_lock);
1652 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1653 spin_unlock(&ctx->completion_lock);
1658 trace_io_uring_defer(req);
1661 list_add_tail(&de->list, &ctx->defer_list);
1662 spin_unlock(&ctx->completion_lock);
1665 static void io_clean_op(struct io_kiocb *req)
1667 if (req->flags & REQ_F_BUFFER_SELECTED) {
1668 spin_lock(&req->ctx->completion_lock);
1669 io_put_kbuf_comp(req);
1670 spin_unlock(&req->ctx->completion_lock);
1673 if (req->flags & REQ_F_NEED_CLEANUP) {
1674 const struct io_op_def *def = &io_op_defs[req->opcode];
1679 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1680 kfree(req->apoll->double_poll);
1684 if (req->flags & REQ_F_INFLIGHT) {
1685 struct io_uring_task *tctx = req->task->io_uring;
1687 atomic_dec(&tctx->inflight_tracked);
1689 if (req->flags & REQ_F_CREDS)
1690 put_cred(req->creds);
1691 if (req->flags & REQ_F_ASYNC_DATA) {
1692 kfree(req->async_data);
1693 req->async_data = NULL;
1695 req->flags &= ~IO_REQ_CLEAN_FLAGS;
1698 static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags)
1700 if (req->file || !io_op_defs[req->opcode].needs_file)
1703 if (req->flags & REQ_F_FIXED_FILE)
1704 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1706 req->file = io_file_get_normal(req, req->cqe.fd);
1711 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1713 const struct io_op_def *def = &io_op_defs[req->opcode];
1714 const struct cred *creds = NULL;
1717 if (unlikely(!io_assign_file(req, issue_flags)))
1720 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1721 creds = override_creds(req->creds);
1723 if (!def->audit_skip)
1724 audit_uring_entry(req->opcode);
1726 ret = def->issue(req, issue_flags);
1728 if (!def->audit_skip)
1729 audit_uring_exit(!ret, ret);
1732 revert_creds(creds);
1734 if (ret == IOU_OK) {
1735 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1736 io_req_complete_defer(req);
1738 io_req_complete_post(req);
1739 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1742 /* If the op doesn't have a file, we're not polling for it */
1743 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
1744 io_iopoll_req_issued(req, issue_flags);
1749 int io_poll_issue(struct io_kiocb *req, bool *locked)
1751 io_tw_lock(req->ctx, locked);
1752 if (unlikely(req->task->flags & PF_EXITING))
1754 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT);
1757 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1759 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1761 req = io_put_req_find_next(req);
1762 return req ? &req->work : NULL;
1765 void io_wq_submit_work(struct io_wq_work *work)
1767 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1768 const struct io_op_def *def = &io_op_defs[req->opcode];
1769 unsigned int issue_flags = IO_URING_F_UNLOCKED;
1770 bool needs_poll = false;
1771 int ret = 0, err = -ECANCELED;
1773 /* one will be dropped by ->io_free_work() after returning to io-wq */
1774 if (!(req->flags & REQ_F_REFCOUNT))
1775 __io_req_set_refcount(req, 2);
1779 io_arm_ltimeout(req);
1781 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1782 if (work->flags & IO_WQ_WORK_CANCEL) {
1784 io_req_task_queue_fail(req, err);
1787 if (!io_assign_file(req, issue_flags)) {
1789 work->flags |= IO_WQ_WORK_CANCEL;
1793 if (req->flags & REQ_F_FORCE_ASYNC) {
1794 bool opcode_poll = def->pollin || def->pollout;
1796 if (opcode_poll && file_can_poll(req->file)) {
1798 issue_flags |= IO_URING_F_NONBLOCK;
1803 ret = io_issue_sqe(req, issue_flags);
1808 * If REQ_F_NOWAIT is set, then don't wait or retry with
1809 * poll. -EAGAIN is final for that case.
1811 if (req->flags & REQ_F_NOWAIT)
1815 * We can get EAGAIN for iopolled IO even though we're
1816 * forcing a sync submission from here, since we can't
1817 * wait for request slots on the block side.
1820 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1826 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1828 /* aborted or ready, in either case retry blocking */
1830 issue_flags &= ~IO_URING_F_NONBLOCK;
1833 /* avoid locking problems by failing it from a clean context */
1835 io_req_task_queue_fail(req, ret);
1838 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1839 unsigned int issue_flags)
1841 struct io_ring_ctx *ctx = req->ctx;
1842 struct file *file = NULL;
1843 unsigned long file_ptr;
1845 io_ring_submit_lock(ctx, issue_flags);
1847 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
1849 fd = array_index_nospec(fd, ctx->nr_user_files);
1850 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
1851 file = (struct file *) (file_ptr & FFS_MASK);
1852 file_ptr &= ~FFS_MASK;
1853 /* mask in overlapping REQ_F and FFS bits */
1854 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
1855 io_req_set_rsrc_node(req, ctx, 0);
1857 io_ring_submit_unlock(ctx, issue_flags);
1861 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
1863 struct file *file = fget(fd);
1865 trace_io_uring_file_get(req, fd);
1867 /* we don't allow fixed io_uring files */
1868 if (file && io_is_uring_fops(file))
1869 io_req_track_inflight(req);
1873 static void io_queue_async(struct io_kiocb *req, int ret)
1874 __must_hold(&req->ctx->uring_lock)
1876 struct io_kiocb *linked_timeout;
1878 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
1879 io_req_complete_failed(req, ret);
1883 linked_timeout = io_prep_linked_timeout(req);
1885 switch (io_arm_poll_handler(req, 0)) {
1886 case IO_APOLL_READY:
1887 io_kbuf_recycle(req, 0);
1888 io_req_task_queue(req);
1890 case IO_APOLL_ABORTED:
1891 io_kbuf_recycle(req, 0);
1892 io_queue_iowq(req, NULL);
1899 io_queue_linked_timeout(linked_timeout);
1902 static inline void io_queue_sqe(struct io_kiocb *req)
1903 __must_hold(&req->ctx->uring_lock)
1907 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
1910 * We async punt it if the file wasn't marked NOWAIT, or if the file
1911 * doesn't support non-blocking read/write attempts
1914 io_arm_ltimeout(req);
1916 io_queue_async(req, ret);
1919 static void io_queue_sqe_fallback(struct io_kiocb *req)
1920 __must_hold(&req->ctx->uring_lock)
1922 if (unlikely(req->flags & REQ_F_FAIL)) {
1924 * We don't submit, fail them all, for that replace hardlinks
1925 * with normal links. Extra REQ_F_LINK is tolerated.
1927 req->flags &= ~REQ_F_HARDLINK;
1928 req->flags |= REQ_F_LINK;
1929 io_req_complete_failed(req, req->cqe.res);
1931 int ret = io_req_prep_async(req);
1933 if (unlikely(ret)) {
1934 io_req_complete_failed(req, ret);
1938 if (unlikely(req->ctx->drain_active))
1941 io_queue_iowq(req, NULL);
1946 * Check SQE restrictions (opcode and flags).
1948 * Returns 'true' if SQE is allowed, 'false' otherwise.
1950 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
1951 struct io_kiocb *req,
1952 unsigned int sqe_flags)
1954 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
1957 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
1958 ctx->restrictions.sqe_flags_required)
1961 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
1962 ctx->restrictions.sqe_flags_required))
1968 static void io_init_req_drain(struct io_kiocb *req)
1970 struct io_ring_ctx *ctx = req->ctx;
1971 struct io_kiocb *head = ctx->submit_state.link.head;
1973 ctx->drain_active = true;
1976 * If we need to drain a request in the middle of a link, drain
1977 * the head request and the next request/link after the current
1978 * link. Considering sequential execution of links,
1979 * REQ_F_IO_DRAIN will be maintained for every request of our
1982 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
1983 ctx->drain_next = true;
1987 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
1988 const struct io_uring_sqe *sqe)
1989 __must_hold(&ctx->uring_lock)
1991 const struct io_op_def *def;
1992 unsigned int sqe_flags;
1996 /* req is partially pre-initialised, see io_preinit_req() */
1997 req->opcode = opcode = READ_ONCE(sqe->opcode);
1998 /* same numerical values with corresponding REQ_F_*, safe to copy */
1999 req->flags = sqe_flags = READ_ONCE(sqe->flags);
2000 req->cqe.user_data = READ_ONCE(sqe->user_data);
2002 req->rsrc_node = NULL;
2003 req->task = current;
2005 if (unlikely(opcode >= IORING_OP_LAST)) {
2009 def = &io_op_defs[opcode];
2010 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2011 /* enforce forwards compatibility on users */
2012 if (sqe_flags & ~SQE_VALID_FLAGS)
2014 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2015 if (!def->buffer_select)
2017 req->buf_index = READ_ONCE(sqe->buf_group);
2019 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2020 ctx->drain_disabled = true;
2021 if (sqe_flags & IOSQE_IO_DRAIN) {
2022 if (ctx->drain_disabled)
2024 io_init_req_drain(req);
2027 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2028 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2030 /* knock it to the slow queue path, will be drained there */
2031 if (ctx->drain_active)
2032 req->flags |= REQ_F_FORCE_ASYNC;
2033 /* if there is no link, we're at "next" request and need to drain */
2034 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2035 ctx->drain_next = false;
2036 ctx->drain_active = true;
2037 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2041 if (!def->ioprio && sqe->ioprio)
2043 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2046 if (def->needs_file) {
2047 struct io_submit_state *state = &ctx->submit_state;
2049 req->cqe.fd = READ_ONCE(sqe->fd);
2052 * Plug now if we have more than 2 IO left after this, and the
2053 * target is potentially a read/write to block based storage.
2055 if (state->need_plug && def->plug) {
2056 state->plug_started = true;
2057 state->need_plug = false;
2058 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2062 personality = READ_ONCE(sqe->personality);
2066 req->creds = xa_load(&ctx->personalities, personality);
2069 get_cred(req->creds);
2070 ret = security_uring_override_creds(req->creds);
2072 put_cred(req->creds);
2075 req->flags |= REQ_F_CREDS;
2078 return def->prep(req, sqe);
2081 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2082 struct io_kiocb *req, int ret)
2084 struct io_ring_ctx *ctx = req->ctx;
2085 struct io_submit_link *link = &ctx->submit_state.link;
2086 struct io_kiocb *head = link->head;
2088 trace_io_uring_req_failed(sqe, req, ret);
2091 * Avoid breaking links in the middle as it renders links with SQPOLL
2092 * unusable. Instead of failing eagerly, continue assembling the link if
2093 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2094 * should find the flag and handle the rest.
2096 req_fail_link_node(req, ret);
2097 if (head && !(head->flags & REQ_F_FAIL))
2098 req_fail_link_node(head, -ECANCELED);
2100 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2102 link->last->link = req;
2106 io_queue_sqe_fallback(req);
2111 link->last->link = req;
2118 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2119 const struct io_uring_sqe *sqe)
2120 __must_hold(&ctx->uring_lock)
2122 struct io_submit_link *link = &ctx->submit_state.link;
2125 ret = io_init_req(ctx, req, sqe);
2127 return io_submit_fail_init(sqe, req, ret);
2129 /* don't need @sqe from now on */
2130 trace_io_uring_submit_sqe(req, true);
2133 * If we already have a head request, queue this one for async
2134 * submittal once the head completes. If we don't have a head but
2135 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2136 * submitted sync once the chain is complete. If none of those
2137 * conditions are true (normal request), then just queue it.
2139 if (unlikely(link->head)) {
2140 ret = io_req_prep_async(req);
2142 return io_submit_fail_init(sqe, req, ret);
2144 trace_io_uring_link(req, link->head);
2145 link->last->link = req;
2148 if (req->flags & IO_REQ_LINK_FLAGS)
2150 /* last request of the link, flush it */
2153 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2156 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2157 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2158 if (req->flags & IO_REQ_LINK_FLAGS) {
2163 io_queue_sqe_fallback(req);
2173 * Batched submission is done, ensure local IO is flushed out.
2175 static void io_submit_state_end(struct io_ring_ctx *ctx)
2177 struct io_submit_state *state = &ctx->submit_state;
2179 if (unlikely(state->link.head))
2180 io_queue_sqe_fallback(state->link.head);
2181 /* flush only after queuing links as they can generate completions */
2182 io_submit_flush_completions(ctx);
2183 if (state->plug_started)
2184 blk_finish_plug(&state->plug);
2188 * Start submission side cache.
2190 static void io_submit_state_start(struct io_submit_state *state,
2191 unsigned int max_ios)
2193 state->plug_started = false;
2194 state->need_plug = max_ios > 2;
2195 state->submit_nr = max_ios;
2196 /* set only head, no need to init link_last in advance */
2197 state->link.head = NULL;
2200 static void io_commit_sqring(struct io_ring_ctx *ctx)
2202 struct io_rings *rings = ctx->rings;
2205 * Ensure any loads from the SQEs are done at this point,
2206 * since once we write the new head, the application could
2207 * write new data to them.
2209 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2213 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2214 * that is mapped by userspace. This means that care needs to be taken to
2215 * ensure that reads are stable, as we cannot rely on userspace always
2216 * being a good citizen. If members of the sqe are validated and then later
2217 * used, it's important that those reads are done through READ_ONCE() to
2218 * prevent a re-load down the line.
2220 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
2222 unsigned head, mask = ctx->sq_entries - 1;
2223 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2226 * The cached sq head (or cq tail) serves two purposes:
2228 * 1) allows us to batch the cost of updating the user visible
2230 * 2) allows the kernel side to track the head on its own, even
2231 * though the application is the one updating it.
2233 head = READ_ONCE(ctx->sq_array[sq_idx]);
2234 if (likely(head < ctx->sq_entries)) {
2235 /* double index for 128-byte SQEs, twice as long */
2236 if (ctx->flags & IORING_SETUP_SQE128)
2238 return &ctx->sq_sqes[head];
2241 /* drop invalid entries */
2243 WRITE_ONCE(ctx->rings->sq_dropped,
2244 READ_ONCE(ctx->rings->sq_dropped) + 1);
2248 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2249 __must_hold(&ctx->uring_lock)
2251 unsigned int entries = io_sqring_entries(ctx);
2255 if (unlikely(!entries))
2257 /* make sure SQ entry isn't read before tail */
2258 ret = left = min3(nr, ctx->sq_entries, entries);
2259 io_get_task_refs(left);
2260 io_submit_state_start(&ctx->submit_state, left);
2263 const struct io_uring_sqe *sqe;
2264 struct io_kiocb *req;
2266 if (unlikely(!io_alloc_req_refill(ctx)))
2268 req = io_alloc_req(ctx);
2269 sqe = io_get_sqe(ctx);
2270 if (unlikely(!sqe)) {
2271 io_req_add_to_cache(req, ctx);
2276 * Continue submitting even for sqe failure if the
2277 * ring was setup with IORING_SETUP_SUBMIT_ALL
2279 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2280 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2286 if (unlikely(left)) {
2288 /* try again if it submitted nothing and can't allocate a req */
2289 if (!ret && io_req_cache_empty(ctx))
2291 current->io_uring->cached_refs += left;
2294 io_submit_state_end(ctx);
2295 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2296 io_commit_sqring(ctx);
2300 struct io_wait_queue {
2301 struct wait_queue_entry wq;
2302 struct io_ring_ctx *ctx;
2304 unsigned nr_timeouts;
2307 static inline bool io_has_work(struct io_ring_ctx *ctx)
2309 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
2310 ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
2311 !llist_empty(&ctx->work_llist));
2314 static inline bool io_should_wake(struct io_wait_queue *iowq)
2316 struct io_ring_ctx *ctx = iowq->ctx;
2317 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
2320 * Wake up if we have enough events, or if a timeout occurred since we
2321 * started waiting. For timeouts, we always want to return to userspace,
2322 * regardless of event count.
2324 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2327 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2328 int wake_flags, void *key)
2330 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
2332 struct io_ring_ctx *ctx = iowq->ctx;
2335 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2336 * the task, and the next invocation will do it.
2338 if (io_should_wake(iowq) || io_has_work(ctx))
2339 return autoremove_wake_function(curr, mode, wake_flags, key);
2343 int io_run_task_work_sig(struct io_ring_ctx *ctx)
2345 if (io_run_task_work_ctx(ctx) > 0)
2347 if (task_sigpending(current))
2352 static bool current_pending_io(void)
2354 struct io_uring_task *tctx = current->io_uring;
2358 return percpu_counter_read_positive(&tctx->inflight);
2361 /* when returns >0, the caller should retry */
2362 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2363 struct io_wait_queue *iowq,
2367 unsigned long check_cq;
2369 /* make sure we run task_work before checking for signals */
2370 ret = io_run_task_work_sig(ctx);
2371 if (ret || io_should_wake(iowq))
2374 check_cq = READ_ONCE(ctx->check_cq);
2375 if (unlikely(check_cq)) {
2376 /* let the caller flush overflows, retry */
2377 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2379 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
2384 * Mark us as being in io_wait if we have pending requests, so cpufreq
2385 * can take into account that the task is waiting for IO - turns out
2386 * to be important for low QD IO.
2388 io_wait = current->in_iowait;
2389 if (current_pending_io())
2390 current->in_iowait = 1;
2392 if (!schedule_hrtimeout(timeout, HRTIMER_MODE_ABS))
2394 current->in_iowait = io_wait;
2399 * Wait until events become available, if we don't already have some. The
2400 * application must reap them itself, as they reside on the shared cq ring.
2402 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2403 const sigset_t __user *sig, size_t sigsz,
2404 struct __kernel_timespec __user *uts)
2406 struct io_wait_queue iowq;
2407 struct io_rings *rings = ctx->rings;
2408 ktime_t timeout = KTIME_MAX;
2411 if (!io_allowed_run_tw(ctx))
2415 /* always run at least 1 task work to process local work */
2416 ret = io_run_task_work_ctx(ctx);
2419 io_cqring_overflow_flush(ctx);
2421 /* if user messes with these they will just get an early return */
2422 if (__io_cqring_events_user(ctx) >= min_events)
2427 #ifdef CONFIG_COMPAT
2428 if (in_compat_syscall())
2429 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2433 ret = set_user_sigmask(sig, sigsz);
2440 struct timespec64 ts;
2442 if (get_timespec64(&ts, uts))
2444 timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2447 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2448 iowq.wq.private = current;
2449 INIT_LIST_HEAD(&iowq.wq.entry);
2451 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2452 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2454 trace_io_uring_cqring_wait(ctx, min_events);
2456 /* if we can't even flush overflow, don't wait for more */
2457 if (!io_cqring_overflow_flush(ctx)) {
2461 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2462 TASK_INTERRUPTIBLE);
2463 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
2467 finish_wait(&ctx->cq_wait, &iowq.wq);
2468 restore_saved_sigmask_unless(ret == -EINTR);
2470 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2473 static void io_mem_free(void *ptr)
2480 page = virt_to_head_page(ptr);
2481 if (put_page_testzero(page))
2482 free_compound_page(page);
2485 static void *io_mem_alloc(size_t size)
2487 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2489 return (void *) __get_free_pages(gfp, get_order(size));
2492 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2493 unsigned int cq_entries, size_t *sq_offset)
2495 struct io_rings *rings;
2496 size_t off, sq_array_size;
2498 off = struct_size(rings, cqes, cq_entries);
2499 if (off == SIZE_MAX)
2501 if (ctx->flags & IORING_SETUP_CQE32) {
2502 if (check_shl_overflow(off, 1, &off))
2507 off = ALIGN(off, SMP_CACHE_BYTES);
2515 sq_array_size = array_size(sizeof(u32), sq_entries);
2516 if (sq_array_size == SIZE_MAX)
2519 if (check_add_overflow(off, sq_array_size, &off))
2525 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2526 unsigned int eventfd_async)
2528 struct io_ev_fd *ev_fd;
2529 __s32 __user *fds = arg;
2532 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2533 lockdep_is_held(&ctx->uring_lock));
2537 if (copy_from_user(&fd, fds, sizeof(*fds)))
2540 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2544 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2545 if (IS_ERR(ev_fd->cq_ev_fd)) {
2546 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2551 spin_lock(&ctx->completion_lock);
2552 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2553 spin_unlock(&ctx->completion_lock);
2555 ev_fd->eventfd_async = eventfd_async;
2556 ctx->has_evfd = true;
2557 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2558 atomic_set(&ev_fd->refs, 1);
2559 atomic_set(&ev_fd->ops, 0);
2563 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2565 struct io_ev_fd *ev_fd;
2567 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2568 lockdep_is_held(&ctx->uring_lock));
2570 ctx->has_evfd = false;
2571 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2572 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops))
2573 call_rcu(&ev_fd->rcu, io_eventfd_ops);
2580 static void io_req_caches_free(struct io_ring_ctx *ctx)
2584 mutex_lock(&ctx->uring_lock);
2585 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
2587 while (!io_req_cache_empty(ctx)) {
2588 struct io_kiocb *req = io_alloc_req(ctx);
2590 kmem_cache_free(req_cachep, req);
2594 percpu_ref_put_many(&ctx->refs, nr);
2595 mutex_unlock(&ctx->uring_lock);
2598 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2600 io_sq_thread_finish(ctx);
2601 io_rsrc_refs_drop(ctx);
2602 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2603 io_wait_rsrc_data(ctx->buf_data);
2604 io_wait_rsrc_data(ctx->file_data);
2606 mutex_lock(&ctx->uring_lock);
2608 __io_sqe_buffers_unregister(ctx);
2610 __io_sqe_files_unregister(ctx);
2612 __io_cqring_overflow_flush(ctx, true);
2613 io_eventfd_unregister(ctx);
2614 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2615 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2616 io_destroy_buffers(ctx);
2617 mutex_unlock(&ctx->uring_lock);
2619 put_cred(ctx->sq_creds);
2620 if (ctx->submitter_task)
2621 put_task_struct(ctx->submitter_task);
2623 /* there are no registered resources left, nobody uses it */
2625 io_rsrc_node_destroy(ctx->rsrc_node);
2626 if (ctx->rsrc_backup_node)
2627 io_rsrc_node_destroy(ctx->rsrc_backup_node);
2628 flush_delayed_work(&ctx->rsrc_put_work);
2629 flush_delayed_work(&ctx->fallback_work);
2631 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2632 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
2634 #if defined(CONFIG_UNIX)
2635 if (ctx->ring_sock) {
2636 ctx->ring_sock->file = NULL; /* so that iput() is called */
2637 sock_release(ctx->ring_sock);
2640 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2642 if (ctx->mm_account) {
2643 mmdrop(ctx->mm_account);
2644 ctx->mm_account = NULL;
2646 io_mem_free(ctx->rings);
2647 io_mem_free(ctx->sq_sqes);
2649 percpu_ref_exit(&ctx->refs);
2650 free_uid(ctx->user);
2651 io_req_caches_free(ctx);
2653 io_wq_put_hash(ctx->hash_map);
2654 kfree(ctx->cancel_table.hbs);
2655 kfree(ctx->cancel_table_locked.hbs);
2656 kfree(ctx->dummy_ubuf);
2658 xa_destroy(&ctx->io_bl_xa);
2662 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2664 struct io_ring_ctx *ctx = file->private_data;
2667 poll_wait(file, &ctx->cq_wait, wait);
2669 * synchronizes with barrier from wq_has_sleeper call in
2673 if (!io_sqring_full(ctx))
2674 mask |= EPOLLOUT | EPOLLWRNORM;
2677 * Don't flush cqring overflow list here, just do a simple check.
2678 * Otherwise there could possible be ABBA deadlock:
2681 * lock(&ctx->uring_lock);
2683 * lock(&ctx->uring_lock);
2686 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2687 * pushs them to do the flush.
2690 if (__io_cqring_events_user(ctx) || io_has_work(ctx))
2691 mask |= EPOLLIN | EPOLLRDNORM;
2696 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2698 const struct cred *creds;
2700 creds = xa_erase(&ctx->personalities, id);
2709 struct io_tctx_exit {
2710 struct callback_head task_work;
2711 struct completion completion;
2712 struct io_ring_ctx *ctx;
2715 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2717 struct io_uring_task *tctx = current->io_uring;
2718 struct io_tctx_exit *work;
2720 work = container_of(cb, struct io_tctx_exit, task_work);
2722 * When @in_idle, we're in cancellation and it's racy to remove the
2723 * node. It'll be removed by the end of cancellation, just ignore it.
2724 * tctx can be NULL if the queueing of this task_work raced with
2725 * work cancelation off the exec path.
2727 if (tctx && !atomic_read(&tctx->in_idle))
2728 io_uring_del_tctx_node((unsigned long)work->ctx);
2729 complete(&work->completion);
2732 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2734 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2736 return req->ctx == data;
2739 static __cold void io_ring_exit_work(struct work_struct *work)
2741 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2742 unsigned long timeout = jiffies + HZ * 60 * 5;
2743 unsigned long interval = HZ / 20;
2744 struct io_tctx_exit exit;
2745 struct io_tctx_node *node;
2749 * If we're doing polled IO and end up having requests being
2750 * submitted async (out-of-line), then completions can come in while
2751 * we're waiting for refs to drop. We need to reap these manually,
2752 * as nobody else will be looking for them.
2755 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
2756 io_move_task_work_from_local(ctx);
2758 while (io_uring_try_cancel_requests(ctx, NULL, true))
2762 struct io_sq_data *sqd = ctx->sq_data;
2763 struct task_struct *tsk;
2765 io_sq_thread_park(sqd);
2767 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
2768 io_wq_cancel_cb(tsk->io_uring->io_wq,
2769 io_cancel_ctx_cb, ctx, true);
2770 io_sq_thread_unpark(sqd);
2773 io_req_caches_free(ctx);
2775 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
2776 /* there is little hope left, don't run it too often */
2780 * This is really an uninterruptible wait, as it has to be
2781 * complete. But it's also run from a kworker, which doesn't
2782 * take signals, so it's fine to make it interruptible. This
2783 * avoids scenarios where we knowingly can wait much longer
2784 * on completions, for example if someone does a SIGSTOP on
2785 * a task that needs to finish task_work to make this loop
2786 * complete. That's a synthetic situation that should not
2787 * cause a stuck task backtrace, and hence a potential panic
2788 * on stuck tasks if that is enabled.
2790 } while (!wait_for_completion_interruptible_timeout(&ctx->ref_comp, interval));
2792 init_completion(&exit.completion);
2793 init_task_work(&exit.task_work, io_tctx_exit_cb);
2796 * Some may use context even when all refs and requests have been put,
2797 * and they are free to do so while still holding uring_lock or
2798 * completion_lock, see io_req_task_submit(). Apart from other work,
2799 * this lock/unlock section also waits them to finish.
2801 mutex_lock(&ctx->uring_lock);
2802 while (!list_empty(&ctx->tctx_list)) {
2803 WARN_ON_ONCE(time_after(jiffies, timeout));
2805 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
2807 /* don't spin on a single task if cancellation failed */
2808 list_rotate_left(&ctx->tctx_list);
2809 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
2810 if (WARN_ON_ONCE(ret))
2813 mutex_unlock(&ctx->uring_lock);
2815 * See comment above for
2816 * wait_for_completion_interruptible_timeout() on why this
2817 * wait is marked as interruptible.
2819 wait_for_completion_interruptible(&exit.completion);
2820 mutex_lock(&ctx->uring_lock);
2822 mutex_unlock(&ctx->uring_lock);
2823 spin_lock(&ctx->completion_lock);
2824 spin_unlock(&ctx->completion_lock);
2826 io_ring_ctx_free(ctx);
2829 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
2831 unsigned long index;
2832 struct creds *creds;
2834 mutex_lock(&ctx->uring_lock);
2835 percpu_ref_kill(&ctx->refs);
2837 __io_cqring_overflow_flush(ctx, true);
2838 xa_for_each(&ctx->personalities, index, creds)
2839 io_unregister_personality(ctx, index);
2841 io_poll_remove_all(ctx, NULL, true);
2842 mutex_unlock(&ctx->uring_lock);
2845 * If we failed setting up the ctx, we might not have any rings
2846 * and therefore did not submit any requests
2849 io_kill_timeouts(ctx, NULL, true);
2851 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
2853 * Use system_unbound_wq to avoid spawning tons of event kworkers
2854 * if we're exiting a ton of rings at the same time. It just adds
2855 * noise and overhead, there's no discernable change in runtime
2856 * over using system_wq.
2858 queue_work(system_unbound_wq, &ctx->exit_work);
2861 static int io_uring_release(struct inode *inode, struct file *file)
2863 struct io_ring_ctx *ctx = file->private_data;
2865 file->private_data = NULL;
2866 io_ring_ctx_wait_and_kill(ctx);
2870 struct io_task_cancel {
2871 struct task_struct *task;
2875 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
2877 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2878 struct io_task_cancel *cancel = data;
2880 return io_match_task_safe(req, cancel->task, cancel->all);
2883 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
2884 struct task_struct *task,
2887 struct io_defer_entry *de;
2890 spin_lock(&ctx->completion_lock);
2891 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
2892 if (io_match_task_safe(de->req, task, cancel_all)) {
2893 list_cut_position(&list, &ctx->defer_list, &de->list);
2897 spin_unlock(&ctx->completion_lock);
2898 if (list_empty(&list))
2901 while (!list_empty(&list)) {
2902 de = list_first_entry(&list, struct io_defer_entry, list);
2903 list_del_init(&de->list);
2904 io_req_task_queue_fail(de->req, -ECANCELED);
2910 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
2912 struct io_tctx_node *node;
2913 enum io_wq_cancel cret;
2916 mutex_lock(&ctx->uring_lock);
2917 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
2918 struct io_uring_task *tctx = node->task->io_uring;
2921 * io_wq will stay alive while we hold uring_lock, because it's
2922 * killed after ctx nodes, which requires to take the lock.
2924 if (!tctx || !tctx->io_wq)
2926 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
2927 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2929 mutex_unlock(&ctx->uring_lock);
2934 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
2935 struct task_struct *task,
2938 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
2939 struct io_uring_task *tctx = task ? task->io_uring : NULL;
2940 enum io_wq_cancel cret;
2943 /* failed during ring init, it couldn't have issued any requests */
2948 ret |= io_uring_try_cancel_iowq(ctx);
2949 } else if (tctx && tctx->io_wq) {
2951 * Cancels requests of all rings, not only @ctx, but
2952 * it's fine as the task is in exit/exec.
2954 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
2956 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2959 /* SQPOLL thread does its own polling */
2960 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
2961 (ctx->sq_data && ctx->sq_data->thread == current)) {
2962 while (!wq_list_empty(&ctx->iopoll_list)) {
2963 io_iopoll_try_reap_events(ctx);
2969 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
2970 ret |= io_run_local_work(ctx) > 0;
2971 ret |= io_cancel_defer_files(ctx, task, cancel_all);
2972 mutex_lock(&ctx->uring_lock);
2973 ret |= io_poll_remove_all(ctx, task, cancel_all);
2974 mutex_unlock(&ctx->uring_lock);
2975 ret |= io_kill_timeouts(ctx, task, cancel_all);
2977 ret |= io_run_task_work() > 0;
2981 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
2984 return atomic_read(&tctx->inflight_tracked);
2985 return percpu_counter_sum(&tctx->inflight);
2989 * Find any io_uring ctx that this task has registered or done IO on, and cancel
2990 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
2992 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
2994 struct io_uring_task *tctx = current->io_uring;
2995 struct io_ring_ctx *ctx;
2999 WARN_ON_ONCE(sqd && sqd->thread != current);
3001 if (!current->io_uring)
3004 io_wq_exit_start(tctx->io_wq);
3006 atomic_inc(&tctx->in_idle);
3010 io_uring_drop_tctx_refs(current);
3011 /* read completions before cancelations */
3012 inflight = tctx_inflight(tctx, !cancel_all);
3017 struct io_tctx_node *node;
3018 unsigned long index;
3020 xa_for_each(&tctx->xa, index, node) {
3021 /* sqpoll task will cancel all its requests */
3022 if (node->ctx->sq_data)
3024 loop |= io_uring_try_cancel_requests(node->ctx,
3025 current, cancel_all);
3028 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3029 loop |= io_uring_try_cancel_requests(ctx,
3039 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3041 io_uring_drop_tctx_refs(current);
3044 * If we've seen completions, retry without waiting. This
3045 * avoids a race where a completion comes in before we did
3046 * prepare_to_wait().
3048 if (inflight == tctx_inflight(tctx, !cancel_all))
3050 finish_wait(&tctx->wait, &wait);
3053 io_uring_clean_tctx(tctx);
3056 * We shouldn't run task_works after cancel, so just leave
3057 * ->in_idle set for normal exit.
3059 atomic_dec(&tctx->in_idle);
3060 /* for exec all current's requests should be gone, kill tctx */
3061 __io_uring_free(current);
3065 void __io_uring_cancel(bool cancel_all)
3067 io_uring_cancel_generic(cancel_all, NULL);
3070 static void *io_uring_validate_mmap_request(struct file *file,
3071 loff_t pgoff, size_t sz)
3073 struct io_ring_ctx *ctx = file->private_data;
3074 loff_t offset = pgoff << PAGE_SHIFT;
3079 case IORING_OFF_SQ_RING:
3080 case IORING_OFF_CQ_RING:
3083 case IORING_OFF_SQES:
3087 return ERR_PTR(-EINVAL);
3090 page = virt_to_head_page(ptr);
3091 if (sz > page_size(page))
3092 return ERR_PTR(-EINVAL);
3099 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3101 size_t sz = vma->vm_end - vma->vm_start;
3105 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3107 return PTR_ERR(ptr);
3109 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3110 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3113 #else /* !CONFIG_MMU */
3115 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3117 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
3120 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3122 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3125 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3126 unsigned long addr, unsigned long len,
3127 unsigned long pgoff, unsigned long flags)
3131 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3133 return PTR_ERR(ptr);
3135 return (unsigned long) ptr;
3138 #endif /* !CONFIG_MMU */
3140 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3142 if (flags & IORING_ENTER_EXT_ARG) {
3143 struct io_uring_getevents_arg arg;
3145 if (argsz != sizeof(arg))
3147 if (copy_from_user(&arg, argp, sizeof(arg)))
3153 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3154 struct __kernel_timespec __user **ts,
3155 const sigset_t __user **sig)
3157 struct io_uring_getevents_arg arg;
3160 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3161 * is just a pointer to the sigset_t.
3163 if (!(flags & IORING_ENTER_EXT_ARG)) {
3164 *sig = (const sigset_t __user *) argp;
3170 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3171 * timespec and sigset_t pointers if good.
3173 if (*argsz != sizeof(arg))
3175 if (copy_from_user(&arg, argp, sizeof(arg)))
3179 *sig = u64_to_user_ptr(arg.sigmask);
3180 *argsz = arg.sigmask_sz;
3181 *ts = u64_to_user_ptr(arg.ts);
3185 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3186 u32, min_complete, u32, flags, const void __user *, argp,
3189 struct io_ring_ctx *ctx;
3193 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3194 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3195 IORING_ENTER_REGISTERED_RING)))
3199 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3200 * need only dereference our task private array to find it.
3202 if (flags & IORING_ENTER_REGISTERED_RING) {
3203 struct io_uring_task *tctx = current->io_uring;
3205 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3207 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3208 f.file = tctx->registered_rings[fd];
3210 if (unlikely(!f.file))
3214 if (unlikely(!f.file))
3217 if (unlikely(!io_is_uring_fops(f.file)))
3221 ctx = f.file->private_data;
3223 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3227 * For SQ polling, the thread will do all submissions and completions.
3228 * Just return the requested submit count, and wake the thread if
3232 if (ctx->flags & IORING_SETUP_SQPOLL) {
3233 io_cqring_overflow_flush(ctx);
3235 if (unlikely(ctx->sq_data->thread == NULL)) {
3239 if (flags & IORING_ENTER_SQ_WAKEUP)
3240 wake_up(&ctx->sq_data->wait);
3241 if (flags & IORING_ENTER_SQ_WAIT) {
3242 ret = io_sqpoll_wait_sq(ctx);
3247 } else if (to_submit) {
3248 ret = io_uring_add_tctx_node(ctx);
3252 mutex_lock(&ctx->uring_lock);
3253 ret = io_submit_sqes(ctx, to_submit);
3254 if (ret != to_submit) {
3255 mutex_unlock(&ctx->uring_lock);
3258 if (flags & IORING_ENTER_GETEVENTS) {
3259 if (ctx->syscall_iopoll)
3262 * Ignore errors, we'll soon call io_cqring_wait() and
3263 * it should handle ownership problems if any.
3265 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3266 (void)io_run_local_work_locked(ctx);
3268 mutex_unlock(&ctx->uring_lock);
3271 if (flags & IORING_ENTER_GETEVENTS) {
3274 if (ctx->syscall_iopoll) {
3276 * We disallow the app entering submit/complete with
3277 * polling, but we still need to lock the ring to
3278 * prevent racing with polled issue that got punted to
3281 mutex_lock(&ctx->uring_lock);
3283 ret2 = io_validate_ext_arg(flags, argp, argsz);
3284 if (likely(!ret2)) {
3285 min_complete = min(min_complete,
3287 ret2 = io_iopoll_check(ctx, min_complete);
3289 mutex_unlock(&ctx->uring_lock);
3291 const sigset_t __user *sig;
3292 struct __kernel_timespec __user *ts;
3294 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3295 if (likely(!ret2)) {
3296 min_complete = min(min_complete,
3298 ret2 = io_cqring_wait(ctx, min_complete, sig,
3307 * EBADR indicates that one or more CQE were dropped.
3308 * Once the user has been informed we can clear the bit
3309 * as they are obviously ok with those drops.
3311 if (unlikely(ret2 == -EBADR))
3312 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3321 static const struct file_operations io_uring_fops = {
3322 .release = io_uring_release,
3323 .mmap = io_uring_mmap,
3325 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3326 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3328 .poll = io_uring_poll,
3329 #ifdef CONFIG_PROC_FS
3330 .show_fdinfo = io_uring_show_fdinfo,
3334 bool io_is_uring_fops(struct file *file)
3336 return file->f_op == &io_uring_fops;
3339 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3340 struct io_uring_params *p)
3342 struct io_rings *rings;
3343 size_t size, sq_array_offset;
3345 /* make sure these are sane, as we already accounted them */
3346 ctx->sq_entries = p->sq_entries;
3347 ctx->cq_entries = p->cq_entries;
3349 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3350 if (size == SIZE_MAX)
3353 rings = io_mem_alloc(size);
3358 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3359 rings->sq_ring_mask = p->sq_entries - 1;
3360 rings->cq_ring_mask = p->cq_entries - 1;
3361 rings->sq_ring_entries = p->sq_entries;
3362 rings->cq_ring_entries = p->cq_entries;
3364 if (p->flags & IORING_SETUP_SQE128)
3365 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3367 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3368 if (size == SIZE_MAX) {
3369 io_mem_free(ctx->rings);
3374 ctx->sq_sqes = io_mem_alloc(size);
3375 if (!ctx->sq_sqes) {
3376 io_mem_free(ctx->rings);
3384 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3388 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3392 ret = __io_uring_add_tctx_node(ctx);
3397 fd_install(fd, file);
3402 * Allocate an anonymous fd, this is what constitutes the application
3403 * visible backing of an io_uring instance. The application mmaps this
3404 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3405 * we have to tie this fd to a socket for file garbage collection purposes.
3407 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3410 #if defined(CONFIG_UNIX)
3413 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3416 return ERR_PTR(ret);
3419 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3420 O_RDWR | O_CLOEXEC, NULL);
3421 #if defined(CONFIG_UNIX)
3423 sock_release(ctx->ring_sock);
3424 ctx->ring_sock = NULL;
3426 ctx->ring_sock->file = file;
3432 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3433 struct io_uring_params __user *params)
3435 struct io_ring_ctx *ctx;
3441 if (entries > IORING_MAX_ENTRIES) {
3442 if (!(p->flags & IORING_SETUP_CLAMP))
3444 entries = IORING_MAX_ENTRIES;
3448 * Use twice as many entries for the CQ ring. It's possible for the
3449 * application to drive a higher depth than the size of the SQ ring,
3450 * since the sqes are only used at submission time. This allows for
3451 * some flexibility in overcommitting a bit. If the application has
3452 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3453 * of CQ ring entries manually.
3455 p->sq_entries = roundup_pow_of_two(entries);
3456 if (p->flags & IORING_SETUP_CQSIZE) {
3458 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3459 * to a power-of-two, if it isn't already. We do NOT impose
3460 * any cq vs sq ring sizing.
3464 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3465 if (!(p->flags & IORING_SETUP_CLAMP))
3467 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3469 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3470 if (p->cq_entries < p->sq_entries)
3473 p->cq_entries = 2 * p->sq_entries;
3476 ctx = io_ring_ctx_alloc(p);
3481 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3482 * space applications don't need to do io completion events
3483 * polling again, they can rely on io_sq_thread to do polling
3484 * work, which can reduce cpu usage and uring_lock contention.
3486 if (ctx->flags & IORING_SETUP_IOPOLL &&
3487 !(ctx->flags & IORING_SETUP_SQPOLL))
3488 ctx->syscall_iopoll = 1;
3490 ctx->compat = in_compat_syscall();
3491 if (!ns_capable_noaudit(&init_user_ns, CAP_IPC_LOCK))
3492 ctx->user = get_uid(current_user());
3495 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3496 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3499 if (ctx->flags & IORING_SETUP_SQPOLL) {
3500 /* IPI related flags don't make sense with SQPOLL */
3501 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3502 IORING_SETUP_TASKRUN_FLAG |
3503 IORING_SETUP_DEFER_TASKRUN))
3505 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3506 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3507 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3509 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3510 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3512 ctx->notify_method = TWA_SIGNAL;
3516 * For DEFER_TASKRUN we require the completion task to be the same as the
3517 * submission task. This implies that there is only one submitter, so enforce
3520 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3521 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3526 * This is just grabbed for accounting purposes. When a process exits,
3527 * the mm is exited and dropped before the files, hence we need to hang
3528 * on to this mm purely for the purposes of being able to unaccount
3529 * memory (locked/pinned vm). It's not used for anything else.
3531 mmgrab(current->mm);
3532 ctx->mm_account = current->mm;
3534 ret = io_allocate_scq_urings(ctx, p);
3538 ret = io_sq_offload_create(ctx, p);
3541 /* always set a rsrc node */
3542 ret = io_rsrc_node_switch_start(ctx);
3545 io_rsrc_node_switch(ctx, NULL);
3547 memset(&p->sq_off, 0, sizeof(p->sq_off));
3548 p->sq_off.head = offsetof(struct io_rings, sq.head);
3549 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3550 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3551 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3552 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3553 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3554 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3556 memset(&p->cq_off, 0, sizeof(p->cq_off));
3557 p->cq_off.head = offsetof(struct io_rings, cq.head);
3558 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3559 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3560 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3561 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3562 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3563 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3565 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3566 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3567 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3568 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3569 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3570 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3571 IORING_FEAT_LINKED_FILE;
3573 if (copy_to_user(params, p, sizeof(*p))) {
3578 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
3579 && !(ctx->flags & IORING_SETUP_R_DISABLED))
3580 ctx->submitter_task = get_task_struct(current);
3582 file = io_uring_get_file(ctx);
3584 ret = PTR_ERR(file);
3589 * Install ring fd as the very last thing, so we don't risk someone
3590 * having closed it before we finish setup
3592 ret = io_uring_install_fd(ctx, file);
3594 /* fput will clean it up */
3599 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3602 io_ring_ctx_wait_and_kill(ctx);
3607 * Sets up an aio uring context, and returns the fd. Applications asks for a
3608 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3609 * params structure passed in.
3611 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3613 struct io_uring_params p;
3616 if (copy_from_user(&p, params, sizeof(p)))
3618 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3623 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3624 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3625 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3626 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3627 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3628 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3629 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN))
3632 return io_uring_create(entries, &p, params);
3635 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3636 struct io_uring_params __user *, params)
3638 return io_uring_setup(entries, params);
3641 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3644 struct io_uring_probe *p;
3648 size = struct_size(p, ops, nr_args);
3649 if (size == SIZE_MAX)
3651 p = kzalloc(size, GFP_KERNEL);
3656 if (copy_from_user(p, arg, size))
3659 if (memchr_inv(p, 0, size))
3662 p->last_op = IORING_OP_LAST - 1;
3663 if (nr_args > IORING_OP_LAST)
3664 nr_args = IORING_OP_LAST;
3666 for (i = 0; i < nr_args; i++) {
3668 if (!io_op_defs[i].not_supported)
3669 p->ops[i].flags = IO_URING_OP_SUPPORTED;
3674 if (copy_to_user(arg, p, size))
3681 static int io_register_personality(struct io_ring_ctx *ctx)
3683 const struct cred *creds;
3687 creds = get_current_cred();
3689 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
3690 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
3698 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
3699 void __user *arg, unsigned int nr_args)
3701 struct io_uring_restriction *res;
3705 /* Restrictions allowed only if rings started disabled */
3706 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3709 /* We allow only a single restrictions registration */
3710 if (ctx->restrictions.registered)
3713 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
3716 size = array_size(nr_args, sizeof(*res));
3717 if (size == SIZE_MAX)
3720 res = memdup_user(arg, size);
3722 return PTR_ERR(res);
3726 for (i = 0; i < nr_args; i++) {
3727 switch (res[i].opcode) {
3728 case IORING_RESTRICTION_REGISTER_OP:
3729 if (res[i].register_op >= IORING_REGISTER_LAST) {
3734 __set_bit(res[i].register_op,
3735 ctx->restrictions.register_op);
3737 case IORING_RESTRICTION_SQE_OP:
3738 if (res[i].sqe_op >= IORING_OP_LAST) {
3743 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
3745 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
3746 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
3748 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
3749 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
3758 /* Reset all restrictions if an error happened */
3760 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
3762 ctx->restrictions.registered = true;
3768 static int io_register_enable_rings(struct io_ring_ctx *ctx)
3770 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3773 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task)
3774 ctx->submitter_task = get_task_struct(current);
3776 if (ctx->restrictions.registered)
3777 ctx->restricted = 1;
3779 ctx->flags &= ~IORING_SETUP_R_DISABLED;
3780 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
3781 wake_up(&ctx->sq_data->wait);
3785 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
3786 void __user *arg, unsigned len)
3788 struct io_uring_task *tctx = current->io_uring;
3789 cpumask_var_t new_mask;
3792 if (!tctx || !tctx->io_wq)
3795 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3798 cpumask_clear(new_mask);
3799 if (len > cpumask_size())
3800 len = cpumask_size();
3802 if (in_compat_syscall()) {
3803 ret = compat_get_bitmap(cpumask_bits(new_mask),
3804 (const compat_ulong_t __user *)arg,
3805 len * 8 /* CHAR_BIT */);
3807 ret = copy_from_user(new_mask, arg, len);
3811 free_cpumask_var(new_mask);
3815 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
3816 free_cpumask_var(new_mask);
3820 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
3822 struct io_uring_task *tctx = current->io_uring;
3824 if (!tctx || !tctx->io_wq)
3827 return io_wq_cpu_affinity(tctx->io_wq, NULL);
3830 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
3832 __must_hold(&ctx->uring_lock)
3834 struct io_tctx_node *node;
3835 struct io_uring_task *tctx = NULL;
3836 struct io_sq_data *sqd = NULL;
3840 if (copy_from_user(new_count, arg, sizeof(new_count)))
3842 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3843 if (new_count[i] > INT_MAX)
3846 if (ctx->flags & IORING_SETUP_SQPOLL) {
3850 * Observe the correct sqd->lock -> ctx->uring_lock
3851 * ordering. Fine to drop uring_lock here, we hold
3854 refcount_inc(&sqd->refs);
3855 mutex_unlock(&ctx->uring_lock);
3856 mutex_lock(&sqd->lock);
3857 mutex_lock(&ctx->uring_lock);
3859 tctx = sqd->thread->io_uring;
3862 tctx = current->io_uring;
3865 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
3867 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3869 ctx->iowq_limits[i] = new_count[i];
3870 ctx->iowq_limits_set = true;
3872 if (tctx && tctx->io_wq) {
3873 ret = io_wq_max_workers(tctx->io_wq, new_count);
3877 memset(new_count, 0, sizeof(new_count));
3881 mutex_unlock(&sqd->lock);
3882 io_put_sq_data(sqd);
3885 if (copy_to_user(arg, new_count, sizeof(new_count)))
3888 /* that's it for SQPOLL, only the SQPOLL task creates requests */
3892 /* now propagate the restriction to all registered users */
3893 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3894 struct io_uring_task *tctx = node->task->io_uring;
3896 if (WARN_ON_ONCE(!tctx->io_wq))
3899 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3900 new_count[i] = ctx->iowq_limits[i];
3901 /* ignore errors, it always returns zero anyway */
3902 (void)io_wq_max_workers(tctx->io_wq, new_count);
3907 mutex_unlock(&sqd->lock);
3908 io_put_sq_data(sqd);
3913 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
3914 void __user *arg, unsigned nr_args)
3915 __releases(ctx->uring_lock)
3916 __acquires(ctx->uring_lock)
3921 * We don't quiesce the refs for register anymore and so it can't be
3922 * dying as we're holding a file ref here.
3924 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
3927 if (ctx->submitter_task && ctx->submitter_task != current)
3930 if (ctx->restricted) {
3931 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
3932 if (!test_bit(opcode, ctx->restrictions.register_op))
3937 case IORING_REGISTER_BUFFERS:
3941 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
3943 case IORING_UNREGISTER_BUFFERS:
3947 ret = io_sqe_buffers_unregister(ctx);
3949 case IORING_REGISTER_FILES:
3953 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
3955 case IORING_UNREGISTER_FILES:
3959 ret = io_sqe_files_unregister(ctx);
3961 case IORING_REGISTER_FILES_UPDATE:
3962 ret = io_register_files_update(ctx, arg, nr_args);
3964 case IORING_REGISTER_EVENTFD:
3968 ret = io_eventfd_register(ctx, arg, 0);
3970 case IORING_REGISTER_EVENTFD_ASYNC:
3974 ret = io_eventfd_register(ctx, arg, 1);
3976 case IORING_UNREGISTER_EVENTFD:
3980 ret = io_eventfd_unregister(ctx);
3982 case IORING_REGISTER_PROBE:
3984 if (!arg || nr_args > 256)
3986 ret = io_probe(ctx, arg, nr_args);
3988 case IORING_REGISTER_PERSONALITY:
3992 ret = io_register_personality(ctx);
3994 case IORING_UNREGISTER_PERSONALITY:
3998 ret = io_unregister_personality(ctx, nr_args);
4000 case IORING_REGISTER_ENABLE_RINGS:
4004 ret = io_register_enable_rings(ctx);
4006 case IORING_REGISTER_RESTRICTIONS:
4007 ret = io_register_restrictions(ctx, arg, nr_args);
4009 case IORING_REGISTER_FILES2:
4010 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
4012 case IORING_REGISTER_FILES_UPDATE2:
4013 ret = io_register_rsrc_update(ctx, arg, nr_args,
4016 case IORING_REGISTER_BUFFERS2:
4017 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
4019 case IORING_REGISTER_BUFFERS_UPDATE:
4020 ret = io_register_rsrc_update(ctx, arg, nr_args,
4021 IORING_RSRC_BUFFER);
4023 case IORING_REGISTER_IOWQ_AFF:
4025 if (!arg || !nr_args)
4027 ret = io_register_iowq_aff(ctx, arg, nr_args);
4029 case IORING_UNREGISTER_IOWQ_AFF:
4033 ret = io_unregister_iowq_aff(ctx);
4035 case IORING_REGISTER_IOWQ_MAX_WORKERS:
4037 if (!arg || nr_args != 2)
4039 ret = io_register_iowq_max_workers(ctx, arg);
4041 case IORING_REGISTER_RING_FDS:
4042 ret = io_ringfd_register(ctx, arg, nr_args);
4044 case IORING_UNREGISTER_RING_FDS:
4045 ret = io_ringfd_unregister(ctx, arg, nr_args);
4047 case IORING_REGISTER_PBUF_RING:
4049 if (!arg || nr_args != 1)
4051 ret = io_register_pbuf_ring(ctx, arg);
4053 case IORING_UNREGISTER_PBUF_RING:
4055 if (!arg || nr_args != 1)
4057 ret = io_unregister_pbuf_ring(ctx, arg);
4059 case IORING_REGISTER_SYNC_CANCEL:
4061 if (!arg || nr_args != 1)
4063 ret = io_sync_cancel(ctx, arg);
4065 case IORING_REGISTER_FILE_ALLOC_RANGE:
4067 if (!arg || nr_args)
4069 ret = io_register_file_alloc_range(ctx, arg);
4079 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
4080 void __user *, arg, unsigned int, nr_args)
4082 struct io_ring_ctx *ctx;
4086 if (opcode >= IORING_REGISTER_LAST)
4094 if (!io_is_uring_fops(f.file))
4097 ctx = f.file->private_data;
4099 io_run_task_work_ctx(ctx);
4101 mutex_lock(&ctx->uring_lock);
4102 ret = __io_uring_register(ctx, opcode, arg, nr_args);
4103 mutex_unlock(&ctx->uring_lock);
4104 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
4110 static int __init io_uring_init(void)
4112 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4113 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4114 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4117 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4118 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4119 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4120 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4121 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4122 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
4123 BUILD_BUG_SQE_ELEM(1, __u8, flags);
4124 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
4125 BUILD_BUG_SQE_ELEM(4, __s32, fd);
4126 BUILD_BUG_SQE_ELEM(8, __u64, off);
4127 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
4128 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
4129 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
4130 BUILD_BUG_SQE_ELEM(16, __u64, addr);
4131 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
4132 BUILD_BUG_SQE_ELEM(24, __u32, len);
4133 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
4134 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
4135 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4136 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
4137 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
4138 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
4139 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
4140 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
4141 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
4142 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
4143 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
4144 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
4145 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
4146 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
4147 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
4148 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
4149 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
4150 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
4151 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
4152 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
4153 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
4154 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
4155 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
4156 BUILD_BUG_SQE_ELEM(42, __u16, personality);
4157 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
4158 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
4159 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
4160 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
4161 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
4162 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4163 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
4165 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4166 sizeof(struct io_uring_rsrc_update));
4167 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4168 sizeof(struct io_uring_rsrc_update2));
4170 /* ->buf_index is u16 */
4171 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4172 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4173 offsetof(struct io_uring_buf_ring, tail));
4175 /* should fit into one byte */
4176 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
4177 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4178 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4180 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
4182 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4184 io_uring_optable_init();
4186 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4190 __initcall(io_uring_init);