1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <net/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
49 #include <linux/bits.h>
51 #include <linux/sched/signal.h>
53 #include <linux/file.h>
54 #include <linux/fdtable.h>
56 #include <linux/mman.h>
57 #include <linux/percpu.h>
58 #include <linux/slab.h>
59 #include <linux/bvec.h>
60 #include <linux/net.h>
62 #include <net/af_unix.h>
64 #include <linux/anon_inodes.h>
65 #include <linux/sched/mm.h>
66 #include <linux/uaccess.h>
67 #include <linux/nospec.h>
68 #include <linux/highmem.h>
69 #include <linux/fsnotify.h>
70 #include <linux/fadvise.h>
71 #include <linux/task_work.h>
72 #include <linux/io_uring.h>
73 #include <linux/audit.h>
74 #include <linux/security.h>
76 #define CREATE_TRACE_POINTS
77 #include <trace/events/io_uring.h>
79 #include <uapi/linux/io_uring.h>
97 #include "alloc_cache.h"
99 #define IORING_MAX_ENTRIES 32768
100 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
102 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
103 IORING_REGISTER_LAST + IORING_OP_LAST)
105 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
106 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
108 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
109 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
111 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
112 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
115 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
118 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
120 #define IO_COMPL_BATCH 32
121 #define IO_REQ_ALLOC_BATCH 8
124 IO_CHECK_CQ_OVERFLOW_BIT,
125 IO_CHECK_CQ_DROPPED_BIT,
129 IO_EVENTFD_OP_SIGNAL_BIT,
130 IO_EVENTFD_OP_FREE_BIT,
133 struct io_defer_entry {
134 struct list_head list;
135 struct io_kiocb *req;
139 /* requests with any of those set should undergo io_disarm_next() */
140 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
141 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
143 static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
144 struct task_struct *task,
147 static void io_dismantle_req(struct io_kiocb *req);
148 static void io_clean_op(struct io_kiocb *req);
149 static void io_queue_sqe(struct io_kiocb *req);
150 static void io_move_task_work_from_local(struct io_ring_ctx *ctx);
151 static void __io_submit_flush_completions(struct io_ring_ctx *ctx);
152 static __cold void io_fallback_tw(struct io_uring_task *tctx);
154 static struct kmem_cache *req_cachep;
156 struct sock *io_uring_get_socket(struct file *file)
158 #if defined(CONFIG_UNIX)
159 if (io_is_uring_fops(file)) {
160 struct io_ring_ctx *ctx = file->private_data;
162 return ctx->ring_sock->sk;
167 EXPORT_SYMBOL(io_uring_get_socket);
169 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
171 if (!wq_list_empty(&ctx->submit_state.compl_reqs) ||
172 ctx->submit_state.cqes_count)
173 __io_submit_flush_completions(ctx);
176 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
178 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
181 static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx)
183 return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head);
186 static bool io_match_linked(struct io_kiocb *head)
188 struct io_kiocb *req;
190 io_for_each_link(req, head) {
191 if (req->flags & REQ_F_INFLIGHT)
198 * As io_match_task() but protected against racing with linked timeouts.
199 * User must not hold timeout_lock.
201 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
206 if (task && head->task != task)
211 if (head->flags & REQ_F_LINK_TIMEOUT) {
212 struct io_ring_ctx *ctx = head->ctx;
214 /* protect against races with linked timeouts */
215 spin_lock_irq(&ctx->timeout_lock);
216 matched = io_match_linked(head);
217 spin_unlock_irq(&ctx->timeout_lock);
219 matched = io_match_linked(head);
224 static inline void req_fail_link_node(struct io_kiocb *req, int res)
227 io_req_set_res(req, res, 0);
230 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
232 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
235 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
237 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
239 complete(&ctx->ref_comp);
242 static __cold void io_fallback_req_func(struct work_struct *work)
244 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
246 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
247 struct io_kiocb *req, *tmp;
250 percpu_ref_get(&ctx->refs);
251 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
252 req->io_task_work.func(req, &locked);
255 io_submit_flush_completions(ctx);
256 mutex_unlock(&ctx->uring_lock);
258 percpu_ref_put(&ctx->refs);
261 static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
263 unsigned hash_buckets = 1U << bits;
264 size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
266 table->hbs = kmalloc(hash_size, GFP_KERNEL);
270 table->hash_bits = bits;
271 init_hash_table(table, hash_buckets);
275 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
277 struct io_ring_ctx *ctx;
280 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
284 xa_init(&ctx->io_bl_xa);
287 * Use 5 bits less than the max cq entries, that should give us around
288 * 32 entries per hash list if totally full and uniformly spread, but
289 * don't keep too many buckets to not overconsume memory.
291 hash_bits = ilog2(p->cq_entries) - 5;
292 hash_bits = clamp(hash_bits, 1, 8);
293 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
295 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
298 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
299 if (!ctx->dummy_ubuf)
301 /* set invalid range, so io_import_fixed() fails meeting it */
302 ctx->dummy_ubuf->ubuf = -1UL;
304 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
308 ctx->flags = p->flags;
309 init_waitqueue_head(&ctx->sqo_sq_wait);
310 INIT_LIST_HEAD(&ctx->sqd_list);
311 INIT_LIST_HEAD(&ctx->cq_overflow_list);
312 INIT_LIST_HEAD(&ctx->io_buffers_cache);
313 io_alloc_cache_init(&ctx->apoll_cache);
314 io_alloc_cache_init(&ctx->netmsg_cache);
315 init_completion(&ctx->ref_comp);
316 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
317 mutex_init(&ctx->uring_lock);
318 init_waitqueue_head(&ctx->cq_wait);
319 spin_lock_init(&ctx->completion_lock);
320 spin_lock_init(&ctx->timeout_lock);
321 INIT_WQ_LIST(&ctx->iopoll_list);
322 INIT_LIST_HEAD(&ctx->io_buffers_pages);
323 INIT_LIST_HEAD(&ctx->io_buffers_comp);
324 INIT_LIST_HEAD(&ctx->defer_list);
325 INIT_LIST_HEAD(&ctx->timeout_list);
326 INIT_LIST_HEAD(&ctx->ltimeout_list);
327 spin_lock_init(&ctx->rsrc_ref_lock);
328 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
329 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
330 init_task_work(&ctx->rsrc_put_tw, io_rsrc_put_tw);
331 init_llist_head(&ctx->rsrc_put_llist);
332 init_llist_head(&ctx->work_llist);
333 INIT_LIST_HEAD(&ctx->tctx_list);
334 ctx->submit_state.free_list.next = NULL;
335 INIT_WQ_LIST(&ctx->locked_free_list);
336 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
337 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
340 kfree(ctx->dummy_ubuf);
341 kfree(ctx->cancel_table.hbs);
342 kfree(ctx->cancel_table_locked.hbs);
344 xa_destroy(&ctx->io_bl_xa);
349 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
351 struct io_rings *r = ctx->rings;
353 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
357 static bool req_need_defer(struct io_kiocb *req, u32 seq)
359 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
360 struct io_ring_ctx *ctx = req->ctx;
362 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
368 static inline void io_req_track_inflight(struct io_kiocb *req)
370 if (!(req->flags & REQ_F_INFLIGHT)) {
371 req->flags |= REQ_F_INFLIGHT;
372 atomic_inc(&req->task->io_uring->inflight_tracked);
376 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
378 if (WARN_ON_ONCE(!req->link))
381 req->flags &= ~REQ_F_ARM_LTIMEOUT;
382 req->flags |= REQ_F_LINK_TIMEOUT;
384 /* linked timeouts should have two refs once prep'ed */
385 io_req_set_refcount(req);
386 __io_req_set_refcount(req->link, 2);
390 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
392 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
394 return __io_prep_linked_timeout(req);
397 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
399 io_queue_linked_timeout(__io_prep_linked_timeout(req));
402 static inline void io_arm_ltimeout(struct io_kiocb *req)
404 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
405 __io_arm_ltimeout(req);
408 static void io_prep_async_work(struct io_kiocb *req)
410 const struct io_op_def *def = &io_op_defs[req->opcode];
411 struct io_ring_ctx *ctx = req->ctx;
413 if (!(req->flags & REQ_F_CREDS)) {
414 req->flags |= REQ_F_CREDS;
415 req->creds = get_current_cred();
418 req->work.list.next = NULL;
420 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
421 if (req->flags & REQ_F_FORCE_ASYNC)
422 req->work.flags |= IO_WQ_WORK_CONCURRENT;
424 if (req->file && !io_req_ffs_set(req))
425 req->flags |= io_file_get_flags(req->file) << REQ_F_SUPPORT_NOWAIT_BIT;
427 if (req->flags & REQ_F_ISREG) {
428 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
429 io_wq_hash_work(&req->work, file_inode(req->file));
430 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
431 if (def->unbound_nonreg_file)
432 req->work.flags |= IO_WQ_WORK_UNBOUND;
436 static void io_prep_async_link(struct io_kiocb *req)
438 struct io_kiocb *cur;
440 if (req->flags & REQ_F_LINK_TIMEOUT) {
441 struct io_ring_ctx *ctx = req->ctx;
443 spin_lock_irq(&ctx->timeout_lock);
444 io_for_each_link(cur, req)
445 io_prep_async_work(cur);
446 spin_unlock_irq(&ctx->timeout_lock);
448 io_for_each_link(cur, req)
449 io_prep_async_work(cur);
453 void io_queue_iowq(struct io_kiocb *req, bool *dont_use)
455 struct io_kiocb *link = io_prep_linked_timeout(req);
456 struct io_uring_task *tctx = req->task->io_uring;
459 BUG_ON(!tctx->io_wq);
461 /* init ->work of the whole link before punting */
462 io_prep_async_link(req);
465 * Not expected to happen, but if we do have a bug where this _can_
466 * happen, catch it here and ensure the request is marked as
467 * canceled. That will make io-wq go through the usual work cancel
468 * procedure rather than attempt to run this request (or create a new
471 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
472 req->work.flags |= IO_WQ_WORK_CANCEL;
474 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
475 io_wq_enqueue(tctx->io_wq, &req->work);
477 io_queue_linked_timeout(link);
480 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
482 while (!list_empty(&ctx->defer_list)) {
483 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
484 struct io_defer_entry, list);
486 if (req_need_defer(de->req, de->seq))
488 list_del_init(&de->list);
489 io_req_task_queue(de->req);
495 static void io_eventfd_ops(struct rcu_head *rcu)
497 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
498 int ops = atomic_xchg(&ev_fd->ops, 0);
500 if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT))
501 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
503 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
504 * ordering in a race but if references are 0 we know we have to free
507 if (atomic_dec_and_test(&ev_fd->refs)) {
508 eventfd_ctx_put(ev_fd->cq_ev_fd);
513 static void io_eventfd_signal(struct io_ring_ctx *ctx)
515 struct io_ev_fd *ev_fd = NULL;
519 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
522 ev_fd = rcu_dereference(ctx->io_ev_fd);
525 * Check again if ev_fd exists incase an io_eventfd_unregister call
526 * completed between the NULL check of ctx->io_ev_fd at the start of
527 * the function and rcu_read_lock.
529 if (unlikely(!ev_fd))
531 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
533 if (ev_fd->eventfd_async && !io_wq_current_is_worker())
536 if (likely(eventfd_signal_allowed())) {
537 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
539 atomic_inc(&ev_fd->refs);
540 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops))
541 call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops);
543 atomic_dec(&ev_fd->refs);
550 static void io_eventfd_flush_signal(struct io_ring_ctx *ctx)
554 spin_lock(&ctx->completion_lock);
557 * Eventfd should only get triggered when at least one event has been
558 * posted. Some applications rely on the eventfd notification count
559 * only changing IFF a new CQE has been added to the CQ ring. There's
560 * no depedency on 1:1 relationship between how many times this
561 * function is called (and hence the eventfd count) and number of CQEs
562 * posted to the CQ ring.
564 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
565 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
566 spin_unlock(&ctx->completion_lock);
570 io_eventfd_signal(ctx);
573 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
575 if (ctx->off_timeout_used)
576 io_flush_timeouts(ctx);
577 if (ctx->drain_active) {
578 spin_lock(&ctx->completion_lock);
579 io_queue_deferred(ctx);
580 spin_unlock(&ctx->completion_lock);
583 io_eventfd_flush_signal(ctx);
586 static inline void __io_cq_lock(struct io_ring_ctx *ctx)
587 __acquires(ctx->completion_lock)
589 if (!ctx->task_complete)
590 spin_lock(&ctx->completion_lock);
593 static inline void __io_cq_unlock(struct io_ring_ctx *ctx)
595 if (!ctx->task_complete)
596 spin_unlock(&ctx->completion_lock);
599 static inline void io_cq_lock(struct io_ring_ctx *ctx)
600 __acquires(ctx->completion_lock)
602 spin_lock(&ctx->completion_lock);
605 static inline void io_cq_unlock(struct io_ring_ctx *ctx)
606 __releases(ctx->completion_lock)
608 spin_unlock(&ctx->completion_lock);
611 /* keep it inlined for io_submit_flush_completions() */
612 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
613 __releases(ctx->completion_lock)
615 io_commit_cqring(ctx);
617 io_commit_cqring_flush(ctx);
621 void io_cq_unlock_post(struct io_ring_ctx *ctx)
622 __releases(ctx->completion_lock)
624 io_commit_cqring(ctx);
625 spin_unlock(&ctx->completion_lock);
626 io_commit_cqring_flush(ctx);
630 /* Returns true if there are no backlogged entries after the flush */
631 static void io_cqring_overflow_kill(struct io_ring_ctx *ctx)
633 struct io_overflow_cqe *ocqe;
637 list_splice_init(&ctx->cq_overflow_list, &list);
638 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
641 while (!list_empty(&list)) {
642 ocqe = list_first_entry(&list, struct io_overflow_cqe, list);
643 list_del(&ocqe->list);
648 /* Returns true if there are no backlogged entries after the flush */
649 static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx)
651 size_t cqe_size = sizeof(struct io_uring_cqe);
653 if (__io_cqring_events(ctx) == ctx->cq_entries)
656 if (ctx->flags & IORING_SETUP_CQE32)
660 while (!list_empty(&ctx->cq_overflow_list)) {
661 struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true);
662 struct io_overflow_cqe *ocqe;
666 ocqe = list_first_entry(&ctx->cq_overflow_list,
667 struct io_overflow_cqe, list);
668 memcpy(cqe, &ocqe->cqe, cqe_size);
669 list_del(&ocqe->list);
673 if (list_empty(&ctx->cq_overflow_list)) {
674 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
675 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
677 io_cq_unlock_post(ctx);
680 static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx)
682 /* iopoll syncs against uring_lock, not completion_lock */
683 if (ctx->flags & IORING_SETUP_IOPOLL)
684 mutex_lock(&ctx->uring_lock);
685 __io_cqring_overflow_flush(ctx);
686 if (ctx->flags & IORING_SETUP_IOPOLL)
687 mutex_unlock(&ctx->uring_lock);
690 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx)
692 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
693 io_cqring_do_overflow_flush(ctx);
696 void __io_put_task(struct task_struct *task, int nr)
698 struct io_uring_task *tctx = task->io_uring;
700 percpu_counter_sub(&tctx->inflight, nr);
701 if (unlikely(atomic_read(&tctx->in_idle)))
702 wake_up(&tctx->wait);
703 put_task_struct_many(task, nr);
706 void io_task_refs_refill(struct io_uring_task *tctx)
708 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
710 percpu_counter_add(&tctx->inflight, refill);
711 refcount_add(refill, ¤t->usage);
712 tctx->cached_refs += refill;
715 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
717 struct io_uring_task *tctx = task->io_uring;
718 unsigned int refs = tctx->cached_refs;
721 tctx->cached_refs = 0;
722 percpu_counter_sub(&tctx->inflight, refs);
723 put_task_struct_many(task, refs);
727 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
728 s32 res, u32 cflags, u64 extra1, u64 extra2)
730 struct io_overflow_cqe *ocqe;
731 size_t ocq_size = sizeof(struct io_overflow_cqe);
732 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
734 lockdep_assert_held(&ctx->completion_lock);
737 ocq_size += sizeof(struct io_uring_cqe);
739 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
740 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
743 * If we're in ring overflow flush mode, or in task cancel mode,
744 * or cannot allocate an overflow entry, then we need to drop it
747 io_account_cq_overflow(ctx);
748 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
751 if (list_empty(&ctx->cq_overflow_list)) {
752 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
753 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
756 ocqe->cqe.user_data = user_data;
758 ocqe->cqe.flags = cflags;
760 ocqe->cqe.big_cqe[0] = extra1;
761 ocqe->cqe.big_cqe[1] = extra2;
763 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
767 bool io_req_cqe_overflow(struct io_kiocb *req)
769 if (!(req->flags & REQ_F_CQE32_INIT)) {
773 return io_cqring_event_overflow(req->ctx, req->cqe.user_data,
774 req->cqe.res, req->cqe.flags,
775 req->extra1, req->extra2);
779 * writes to the cq entry need to come after reading head; the
780 * control dependency is enough as we're using WRITE_ONCE to
783 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow)
785 struct io_rings *rings = ctx->rings;
786 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
787 unsigned int free, queued, len;
790 * Posting into the CQ when there are pending overflowed CQEs may break
791 * ordering guarantees, which will affect links, F_MORE users and more.
792 * Force overflow the completion.
794 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
797 /* userspace may cheat modifying the tail, be safe and do min */
798 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
799 free = ctx->cq_entries - queued;
800 /* we need a contiguous range, limit based on the current array offset */
801 len = min(free, ctx->cq_entries - off);
805 if (ctx->flags & IORING_SETUP_CQE32) {
810 ctx->cqe_cached = &rings->cqes[off];
811 ctx->cqe_sentinel = ctx->cqe_cached + len;
813 ctx->cached_cq_tail++;
815 if (ctx->flags & IORING_SETUP_CQE32)
817 return &rings->cqes[off];
820 static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
823 struct io_uring_cqe *cqe;
828 * If we can't get a cq entry, userspace overflowed the
829 * submission (by quite a lot). Increment the overflow count in
832 cqe = io_get_cqe(ctx);
834 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
836 WRITE_ONCE(cqe->user_data, user_data);
837 WRITE_ONCE(cqe->res, res);
838 WRITE_ONCE(cqe->flags, cflags);
840 if (ctx->flags & IORING_SETUP_CQE32) {
841 WRITE_ONCE(cqe->big_cqe[0], 0);
842 WRITE_ONCE(cqe->big_cqe[1], 0);
849 static void __io_flush_post_cqes(struct io_ring_ctx *ctx)
850 __must_hold(&ctx->uring_lock)
852 struct io_submit_state *state = &ctx->submit_state;
855 lockdep_assert_held(&ctx->uring_lock);
856 for (i = 0; i < state->cqes_count; i++) {
857 struct io_uring_cqe *cqe = &state->cqes[i];
859 if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) {
860 if (ctx->task_complete) {
861 spin_lock(&ctx->completion_lock);
862 io_cqring_event_overflow(ctx, cqe->user_data,
863 cqe->res, cqe->flags, 0, 0);
864 spin_unlock(&ctx->completion_lock);
866 io_cqring_event_overflow(ctx, cqe->user_data,
867 cqe->res, cqe->flags, 0, 0);
871 state->cqes_count = 0;
874 static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
880 filled = io_fill_cqe_aux(ctx, user_data, res, cflags);
881 if (!filled && allow_overflow)
882 filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
884 io_cq_unlock_post(ctx);
888 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags)
890 return __io_post_aux_cqe(ctx, user_data, res, cflags, true);
893 bool io_aux_cqe(struct io_ring_ctx *ctx, bool defer, u64 user_data, s32 res, u32 cflags,
896 struct io_uring_cqe *cqe;
900 return __io_post_aux_cqe(ctx, user_data, res, cflags, allow_overflow);
902 length = ARRAY_SIZE(ctx->submit_state.cqes);
904 lockdep_assert_held(&ctx->uring_lock);
906 if (ctx->submit_state.cqes_count == length) {
908 __io_flush_post_cqes(ctx);
909 /* no need to flush - flush is deferred */
910 __io_cq_unlock_post(ctx);
913 /* For defered completions this is not as strict as it is otherwise,
914 * however it's main job is to prevent unbounded posted completions,
915 * and in that it works just as well.
917 if (!allow_overflow && test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
920 cqe = &ctx->submit_state.cqes[ctx->submit_state.cqes_count++];
921 cqe->user_data = user_data;
927 static void __io_req_complete_post(struct io_kiocb *req)
929 struct io_ring_ctx *ctx = req->ctx;
932 if (!(req->flags & REQ_F_CQE_SKIP))
933 io_fill_cqe_req(ctx, req);
936 * If we're the last reference to this request, add to our locked
939 if (req_ref_put_and_test(req)) {
940 if (req->flags & IO_REQ_LINK_FLAGS) {
941 if (req->flags & IO_DISARM_MASK)
944 io_req_task_queue(req->link);
948 io_req_put_rsrc(req);
950 * Selected buffer deallocation in io_clean_op() assumes that
951 * we don't hold ->completion_lock. Clean them here to avoid
954 io_put_kbuf_comp(req);
955 io_dismantle_req(req);
956 io_put_task(req->task, 1);
957 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
958 ctx->locked_free_nr++;
960 io_cq_unlock_post(ctx);
963 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
965 if (req->ctx->task_complete && (issue_flags & IO_URING_F_IOWQ)) {
966 req->io_task_work.func = io_req_task_complete;
967 io_req_task_work_add(req);
968 } else if (!(issue_flags & IO_URING_F_UNLOCKED) ||
969 !(req->ctx->flags & IORING_SETUP_IOPOLL)) {
970 __io_req_complete_post(req);
972 struct io_ring_ctx *ctx = req->ctx;
974 mutex_lock(&ctx->uring_lock);
975 __io_req_complete_post(req);
976 mutex_unlock(&ctx->uring_lock);
980 void io_req_defer_failed(struct io_kiocb *req, s32 res)
981 __must_hold(&ctx->uring_lock)
983 const struct io_op_def *def = &io_op_defs[req->opcode];
985 lockdep_assert_held(&req->ctx->uring_lock);
988 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
991 io_req_complete_defer(req);
995 * Don't initialise the fields below on every allocation, but do that in
996 * advance and keep them valid across allocations.
998 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1002 req->async_data = NULL;
1003 /* not necessary, but safer to zero */
1007 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1008 struct io_submit_state *state)
1010 spin_lock(&ctx->completion_lock);
1011 wq_list_splice(&ctx->locked_free_list, &state->free_list);
1012 ctx->locked_free_nr = 0;
1013 spin_unlock(&ctx->completion_lock);
1017 * A request might get retired back into the request caches even before opcode
1018 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1019 * Because of that, io_alloc_req() should be called only under ->uring_lock
1020 * and with extra caution to not get a request that is still worked on.
1022 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
1023 __must_hold(&ctx->uring_lock)
1025 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1026 void *reqs[IO_REQ_ALLOC_BATCH];
1030 * If we have more than a batch's worth of requests in our IRQ side
1031 * locked cache, grab the lock and move them over to our submission
1034 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
1035 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
1036 if (!io_req_cache_empty(ctx))
1040 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
1043 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1044 * retry single alloc to be on the safe side.
1046 if (unlikely(ret <= 0)) {
1047 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1053 percpu_ref_get_many(&ctx->refs, ret);
1054 for (i = 0; i < ret; i++) {
1055 struct io_kiocb *req = reqs[i];
1057 io_preinit_req(req, ctx);
1058 io_req_add_to_cache(req, ctx);
1063 static inline void io_dismantle_req(struct io_kiocb *req)
1065 unsigned int flags = req->flags;
1067 if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
1069 if (!(flags & REQ_F_FIXED_FILE))
1070 io_put_file(req->file);
1073 __cold void io_free_req(struct io_kiocb *req)
1075 struct io_ring_ctx *ctx = req->ctx;
1077 io_req_put_rsrc(req);
1078 io_dismantle_req(req);
1079 io_put_task(req->task, 1);
1081 spin_lock(&ctx->completion_lock);
1082 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
1083 ctx->locked_free_nr++;
1084 spin_unlock(&ctx->completion_lock);
1087 static void __io_req_find_next_prep(struct io_kiocb *req)
1089 struct io_ring_ctx *ctx = req->ctx;
1091 spin_lock(&ctx->completion_lock);
1092 io_disarm_next(req);
1093 spin_unlock(&ctx->completion_lock);
1096 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1098 struct io_kiocb *nxt;
1101 * If LINK is set, we have dependent requests in this chain. If we
1102 * didn't fail this request, queue the first one up, moving any other
1103 * dependencies to the next request. In case of failure, fail the rest
1106 if (unlikely(req->flags & IO_DISARM_MASK))
1107 __io_req_find_next_prep(req);
1113 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
1117 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1118 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1120 io_submit_flush_completions(ctx);
1121 mutex_unlock(&ctx->uring_lock);
1124 percpu_ref_put(&ctx->refs);
1127 static unsigned int handle_tw_list(struct llist_node *node,
1128 struct io_ring_ctx **ctx, bool *locked,
1129 struct llist_node *last)
1131 unsigned int count = 0;
1133 while (node != last) {
1134 struct llist_node *next = node->next;
1135 struct io_kiocb *req = container_of(node, struct io_kiocb,
1138 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1140 if (req->ctx != *ctx) {
1141 ctx_flush_and_put(*ctx, locked);
1143 /* if not contended, grab and improve batching */
1144 *locked = mutex_trylock(&(*ctx)->uring_lock);
1145 percpu_ref_get(&(*ctx)->refs);
1147 req->io_task_work.func(req, locked);
1156 * io_llist_xchg - swap all entries in a lock-less list
1157 * @head: the head of lock-less list to delete all entries
1158 * @new: new entry as the head of the list
1160 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1161 * The order of entries returned is from the newest to the oldest added one.
1163 static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1164 struct llist_node *new)
1166 return xchg(&head->first, new);
1170 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1171 * @head: the head of lock-less list to delete all entries
1172 * @old: expected old value of the first entry of the list
1173 * @new: new entry as the head of the list
1175 * perform a cmpxchg on the first entry of the list.
1178 static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1179 struct llist_node *old,
1180 struct llist_node *new)
1182 return cmpxchg(&head->first, old, new);
1185 void tctx_task_work(struct callback_head *cb)
1187 bool uring_locked = false;
1188 struct io_ring_ctx *ctx = NULL;
1189 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1191 struct llist_node fake = {};
1192 struct llist_node *node;
1193 unsigned int loops = 1;
1196 if (unlikely(current->flags & PF_EXITING)) {
1197 io_fallback_tw(tctx);
1201 node = io_llist_xchg(&tctx->task_list, &fake);
1202 count = handle_tw_list(node, &ctx, &uring_locked, NULL);
1203 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1204 while (node != &fake) {
1206 node = io_llist_xchg(&tctx->task_list, &fake);
1207 count += handle_tw_list(node, &ctx, &uring_locked, &fake);
1208 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1211 ctx_flush_and_put(ctx, &uring_locked);
1213 /* relaxed read is enough as only the task itself sets ->in_idle */
1214 if (unlikely(atomic_read(&tctx->in_idle)))
1215 io_uring_drop_tctx_refs(current);
1217 trace_io_uring_task_work_run(tctx, count, loops);
1220 static __cold void io_fallback_tw(struct io_uring_task *tctx)
1222 struct llist_node *node = llist_del_all(&tctx->task_list);
1223 struct io_kiocb *req;
1226 req = container_of(node, struct io_kiocb, io_task_work.node);
1228 if (llist_add(&req->io_task_work.node,
1229 &req->ctx->fallback_llist))
1230 schedule_delayed_work(&req->ctx->fallback_work, 1);
1234 static void io_req_local_work_add(struct io_kiocb *req)
1236 struct io_ring_ctx *ctx = req->ctx;
1238 percpu_ref_get(&ctx->refs);
1240 if (!llist_add(&req->io_task_work.node, &ctx->work_llist)) {
1241 percpu_ref_put(&ctx->refs);
1244 /* need it for the following io_cqring_wake() */
1245 smp_mb__after_atomic();
1247 if (unlikely(atomic_read(&req->task->io_uring->in_idle))) {
1248 io_move_task_work_from_local(ctx);
1249 percpu_ref_put(&ctx->refs);
1253 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1254 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1257 io_eventfd_signal(ctx);
1258 __io_cqring_wake(ctx);
1259 percpu_ref_put(&ctx->refs);
1262 void __io_req_task_work_add(struct io_kiocb *req, bool allow_local)
1264 struct io_uring_task *tctx = req->task->io_uring;
1265 struct io_ring_ctx *ctx = req->ctx;
1267 if (allow_local && ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
1268 io_req_local_work_add(req);
1272 /* task_work already pending, we're done */
1273 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1276 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1277 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1279 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1282 io_fallback_tw(tctx);
1285 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1287 struct llist_node *node;
1289 node = llist_del_all(&ctx->work_llist);
1291 struct io_kiocb *req = container_of(node, struct io_kiocb,
1295 __io_req_task_work_add(req, false);
1299 int __io_run_local_work(struct io_ring_ctx *ctx, bool *locked)
1301 struct llist_node *node;
1302 struct llist_node fake;
1303 struct llist_node *current_final = NULL;
1305 unsigned int loops = 1;
1307 if (unlikely(ctx->submitter_task != current))
1310 node = io_llist_xchg(&ctx->work_llist, &fake);
1313 while (node != current_final) {
1314 struct llist_node *next = node->next;
1315 struct io_kiocb *req = container_of(node, struct io_kiocb,
1317 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1318 req->io_task_work.func(req, locked);
1323 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1324 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1326 node = io_llist_cmpxchg(&ctx->work_llist, &fake, NULL);
1327 if (node != &fake) {
1329 current_final = &fake;
1330 node = io_llist_xchg(&ctx->work_llist, &fake);
1335 io_submit_flush_completions(ctx);
1336 trace_io_uring_local_work_run(ctx, ret, loops);
1341 int io_run_local_work(struct io_ring_ctx *ctx)
1346 __set_current_state(TASK_RUNNING);
1347 locked = mutex_trylock(&ctx->uring_lock);
1348 ret = __io_run_local_work(ctx, &locked);
1350 mutex_unlock(&ctx->uring_lock);
1355 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
1357 io_tw_lock(req->ctx, locked);
1358 io_req_defer_failed(req, req->cqe.res);
1361 void io_req_task_submit(struct io_kiocb *req, bool *locked)
1363 io_tw_lock(req->ctx, locked);
1364 /* req->task == current here, checking PF_EXITING is safe */
1365 if (likely(!(req->task->flags & PF_EXITING)))
1368 io_req_defer_failed(req, -EFAULT);
1371 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1373 io_req_set_res(req, ret, 0);
1374 req->io_task_work.func = io_req_task_cancel;
1375 io_req_task_work_add(req);
1378 void io_req_task_queue(struct io_kiocb *req)
1380 req->io_task_work.func = io_req_task_submit;
1381 io_req_task_work_add(req);
1384 void io_queue_next(struct io_kiocb *req)
1386 struct io_kiocb *nxt = io_req_find_next(req);
1389 io_req_task_queue(nxt);
1392 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1393 __must_hold(&ctx->uring_lock)
1395 struct task_struct *task = NULL;
1399 struct io_kiocb *req = container_of(node, struct io_kiocb,
1402 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1403 if (req->flags & REQ_F_REFCOUNT) {
1404 node = req->comp_list.next;
1405 if (!req_ref_put_and_test(req))
1408 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1409 struct async_poll *apoll = req->apoll;
1411 if (apoll->double_poll)
1412 kfree(apoll->double_poll);
1413 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1415 req->flags &= ~REQ_F_POLLED;
1417 if (req->flags & IO_REQ_LINK_FLAGS)
1419 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1422 if (!(req->flags & REQ_F_FIXED_FILE))
1423 io_put_file(req->file);
1425 io_req_put_rsrc_locked(req, ctx);
1427 if (req->task != task) {
1429 io_put_task(task, task_refs);
1434 node = req->comp_list.next;
1435 io_req_add_to_cache(req, ctx);
1439 io_put_task(task, task_refs);
1442 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1443 __must_hold(&ctx->uring_lock)
1445 struct io_wq_work_node *node, *prev;
1446 struct io_submit_state *state = &ctx->submit_state;
1449 /* must come first to preserve CQE ordering in failure cases */
1450 if (state->cqes_count)
1451 __io_flush_post_cqes(ctx);
1452 wq_list_for_each(node, prev, &state->compl_reqs) {
1453 struct io_kiocb *req = container_of(node, struct io_kiocb,
1456 if (!(req->flags & REQ_F_CQE_SKIP) &&
1457 unlikely(!__io_fill_cqe_req(ctx, req))) {
1458 if (ctx->task_complete) {
1459 spin_lock(&ctx->completion_lock);
1460 io_req_cqe_overflow(req);
1461 spin_unlock(&ctx->completion_lock);
1463 io_req_cqe_overflow(req);
1467 __io_cq_unlock_post(ctx);
1469 if (!wq_list_empty(&ctx->submit_state.compl_reqs)) {
1470 io_free_batch_list(ctx, state->compl_reqs.first);
1471 INIT_WQ_LIST(&state->compl_reqs);
1476 * Drop reference to request, return next in chain (if there is one) if this
1477 * was the last reference to this request.
1479 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1481 struct io_kiocb *nxt = NULL;
1483 if (req_ref_put_and_test(req)) {
1484 if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
1485 nxt = io_req_find_next(req);
1491 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1493 /* See comment at the top of this file */
1495 return __io_cqring_events(ctx);
1499 * We can't just wait for polled events to come to us, we have to actively
1500 * find and complete them.
1502 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1504 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1507 mutex_lock(&ctx->uring_lock);
1508 while (!wq_list_empty(&ctx->iopoll_list)) {
1509 /* let it sleep and repeat later if can't complete a request */
1510 if (io_do_iopoll(ctx, true) == 0)
1513 * Ensure we allow local-to-the-cpu processing to take place,
1514 * in this case we need to ensure that we reap all events.
1515 * Also let task_work, etc. to progress by releasing the mutex
1517 if (need_resched()) {
1518 mutex_unlock(&ctx->uring_lock);
1520 mutex_lock(&ctx->uring_lock);
1523 mutex_unlock(&ctx->uring_lock);
1526 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1528 unsigned int nr_events = 0;
1530 unsigned long check_cq;
1532 if (!io_allowed_run_tw(ctx))
1535 check_cq = READ_ONCE(ctx->check_cq);
1536 if (unlikely(check_cq)) {
1537 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1538 __io_cqring_overflow_flush(ctx);
1540 * Similarly do not spin if we have not informed the user of any
1543 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1547 * Don't enter poll loop if we already have events pending.
1548 * If we do, we can potentially be spinning for commands that
1549 * already triggered a CQE (eg in error).
1551 if (io_cqring_events(ctx))
1556 * If a submit got punted to a workqueue, we can have the
1557 * application entering polling for a command before it gets
1558 * issued. That app will hold the uring_lock for the duration
1559 * of the poll right here, so we need to take a breather every
1560 * now and then to ensure that the issue has a chance to add
1561 * the poll to the issued list. Otherwise we can spin here
1562 * forever, while the workqueue is stuck trying to acquire the
1565 if (wq_list_empty(&ctx->iopoll_list) ||
1566 io_task_work_pending(ctx)) {
1567 u32 tail = ctx->cached_cq_tail;
1569 (void) io_run_local_work_locked(ctx);
1571 if (task_work_pending(current) ||
1572 wq_list_empty(&ctx->iopoll_list)) {
1573 mutex_unlock(&ctx->uring_lock);
1575 mutex_lock(&ctx->uring_lock);
1577 /* some requests don't go through iopoll_list */
1578 if (tail != ctx->cached_cq_tail ||
1579 wq_list_empty(&ctx->iopoll_list))
1582 ret = io_do_iopoll(ctx, !min);
1587 } while (nr_events < min && !need_resched());
1592 void io_req_task_complete(struct io_kiocb *req, bool *locked)
1595 io_req_complete_defer(req);
1597 io_req_complete_post(req, IO_URING_F_UNLOCKED);
1601 * After the iocb has been issued, it's safe to be found on the poll list.
1602 * Adding the kiocb to the list AFTER submission ensures that we don't
1603 * find it from a io_do_iopoll() thread before the issuer is done
1604 * accessing the kiocb cookie.
1606 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1608 struct io_ring_ctx *ctx = req->ctx;
1609 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1611 /* workqueue context doesn't hold uring_lock, grab it now */
1612 if (unlikely(needs_lock))
1613 mutex_lock(&ctx->uring_lock);
1616 * Track whether we have multiple files in our lists. This will impact
1617 * how we do polling eventually, not spinning if we're on potentially
1618 * different devices.
1620 if (wq_list_empty(&ctx->iopoll_list)) {
1621 ctx->poll_multi_queue = false;
1622 } else if (!ctx->poll_multi_queue) {
1623 struct io_kiocb *list_req;
1625 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1627 if (list_req->file != req->file)
1628 ctx->poll_multi_queue = true;
1632 * For fast devices, IO may have already completed. If it has, add
1633 * it to the front so we find it first.
1635 if (READ_ONCE(req->iopoll_completed))
1636 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1638 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1640 if (unlikely(needs_lock)) {
1642 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1643 * in sq thread task context or in io worker task context. If
1644 * current task context is sq thread, we don't need to check
1645 * whether should wake up sq thread.
1647 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1648 wq_has_sleeper(&ctx->sq_data->wait))
1649 wake_up(&ctx->sq_data->wait);
1651 mutex_unlock(&ctx->uring_lock);
1655 static bool io_bdev_nowait(struct block_device *bdev)
1657 return !bdev || bdev_nowait(bdev);
1661 * If we tracked the file through the SCM inflight mechanism, we could support
1662 * any file. For now, just ensure that anything potentially problematic is done
1665 static bool __io_file_supports_nowait(struct file *file, umode_t mode)
1667 if (S_ISBLK(mode)) {
1668 if (IS_ENABLED(CONFIG_BLOCK) &&
1669 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
1675 if (S_ISREG(mode)) {
1676 if (IS_ENABLED(CONFIG_BLOCK) &&
1677 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
1678 !io_is_uring_fops(file))
1683 /* any ->read/write should understand O_NONBLOCK */
1684 if (file->f_flags & O_NONBLOCK)
1686 return file->f_mode & FMODE_NOWAIT;
1690 * If we tracked the file through the SCM inflight mechanism, we could support
1691 * any file. For now, just ensure that anything potentially problematic is done
1694 unsigned int io_file_get_flags(struct file *file)
1696 umode_t mode = file_inode(file)->i_mode;
1697 unsigned int res = 0;
1701 if (__io_file_supports_nowait(file, mode))
1706 bool io_alloc_async_data(struct io_kiocb *req)
1708 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
1709 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
1710 if (req->async_data) {
1711 req->flags |= REQ_F_ASYNC_DATA;
1717 int io_req_prep_async(struct io_kiocb *req)
1719 const struct io_op_def *def = &io_op_defs[req->opcode];
1721 /* assign early for deferred execution for non-fixed file */
1722 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE))
1723 req->file = io_file_get_normal(req, req->cqe.fd);
1724 if (!def->prep_async)
1726 if (WARN_ON_ONCE(req_has_async_data(req)))
1728 if (!io_op_defs[req->opcode].manual_alloc) {
1729 if (io_alloc_async_data(req))
1732 return def->prep_async(req);
1735 static u32 io_get_sequence(struct io_kiocb *req)
1737 u32 seq = req->ctx->cached_sq_head;
1738 struct io_kiocb *cur;
1740 /* need original cached_sq_head, but it was increased for each req */
1741 io_for_each_link(cur, req)
1746 static __cold void io_drain_req(struct io_kiocb *req)
1747 __must_hold(&ctx->uring_lock)
1749 struct io_ring_ctx *ctx = req->ctx;
1750 struct io_defer_entry *de;
1752 u32 seq = io_get_sequence(req);
1754 /* Still need defer if there is pending req in defer list. */
1755 spin_lock(&ctx->completion_lock);
1756 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1757 spin_unlock(&ctx->completion_lock);
1759 ctx->drain_active = false;
1760 io_req_task_queue(req);
1763 spin_unlock(&ctx->completion_lock);
1765 io_prep_async_link(req);
1766 de = kmalloc(sizeof(*de), GFP_KERNEL);
1769 io_req_defer_failed(req, ret);
1773 spin_lock(&ctx->completion_lock);
1774 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1775 spin_unlock(&ctx->completion_lock);
1780 trace_io_uring_defer(req);
1783 list_add_tail(&de->list, &ctx->defer_list);
1784 spin_unlock(&ctx->completion_lock);
1787 static void io_clean_op(struct io_kiocb *req)
1789 if (req->flags & REQ_F_BUFFER_SELECTED) {
1790 spin_lock(&req->ctx->completion_lock);
1791 io_put_kbuf_comp(req);
1792 spin_unlock(&req->ctx->completion_lock);
1795 if (req->flags & REQ_F_NEED_CLEANUP) {
1796 const struct io_op_def *def = &io_op_defs[req->opcode];
1801 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1802 kfree(req->apoll->double_poll);
1806 if (req->flags & REQ_F_INFLIGHT) {
1807 struct io_uring_task *tctx = req->task->io_uring;
1809 atomic_dec(&tctx->inflight_tracked);
1811 if (req->flags & REQ_F_CREDS)
1812 put_cred(req->creds);
1813 if (req->flags & REQ_F_ASYNC_DATA) {
1814 kfree(req->async_data);
1815 req->async_data = NULL;
1817 req->flags &= ~IO_REQ_CLEAN_FLAGS;
1820 static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags)
1822 if (req->file || !io_op_defs[req->opcode].needs_file)
1825 if (req->flags & REQ_F_FIXED_FILE)
1826 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1828 req->file = io_file_get_normal(req, req->cqe.fd);
1833 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1835 const struct io_op_def *def = &io_op_defs[req->opcode];
1836 const struct cred *creds = NULL;
1839 if (unlikely(!io_assign_file(req, issue_flags)))
1842 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1843 creds = override_creds(req->creds);
1845 if (!def->audit_skip)
1846 audit_uring_entry(req->opcode);
1848 ret = def->issue(req, issue_flags);
1850 if (!def->audit_skip)
1851 audit_uring_exit(!ret, ret);
1854 revert_creds(creds);
1856 if (ret == IOU_OK) {
1857 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1858 io_req_complete_defer(req);
1860 io_req_complete_post(req, issue_flags);
1861 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1864 /* If the op doesn't have a file, we're not polling for it */
1865 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
1866 io_iopoll_req_issued(req, issue_flags);
1871 int io_poll_issue(struct io_kiocb *req, bool *locked)
1873 io_tw_lock(req->ctx, locked);
1874 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT|
1875 IO_URING_F_COMPLETE_DEFER);
1878 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1880 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1882 req = io_put_req_find_next(req);
1883 return req ? &req->work : NULL;
1886 void io_wq_submit_work(struct io_wq_work *work)
1888 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1889 const struct io_op_def *def = &io_op_defs[req->opcode];
1890 unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ;
1891 bool needs_poll = false;
1892 int ret = 0, err = -ECANCELED;
1894 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1895 if (!(req->flags & REQ_F_REFCOUNT))
1896 __io_req_set_refcount(req, 2);
1900 io_arm_ltimeout(req);
1902 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1903 if (work->flags & IO_WQ_WORK_CANCEL) {
1905 io_req_task_queue_fail(req, err);
1908 if (!io_assign_file(req, issue_flags)) {
1910 work->flags |= IO_WQ_WORK_CANCEL;
1914 if (req->flags & REQ_F_FORCE_ASYNC) {
1915 bool opcode_poll = def->pollin || def->pollout;
1917 if (opcode_poll && file_can_poll(req->file)) {
1919 issue_flags |= IO_URING_F_NONBLOCK;
1924 ret = io_issue_sqe(req, issue_flags);
1928 * We can get EAGAIN for iopolled IO even though we're
1929 * forcing a sync submission from here, since we can't
1930 * wait for request slots on the block side.
1933 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1939 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1941 /* aborted or ready, in either case retry blocking */
1943 issue_flags &= ~IO_URING_F_NONBLOCK;
1946 /* avoid locking problems by failing it from a clean context */
1948 io_req_task_queue_fail(req, ret);
1951 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1952 unsigned int issue_flags)
1954 struct io_ring_ctx *ctx = req->ctx;
1955 struct file *file = NULL;
1956 unsigned long file_ptr;
1958 io_ring_submit_lock(ctx, issue_flags);
1960 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
1962 fd = array_index_nospec(fd, ctx->nr_user_files);
1963 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
1964 file = (struct file *) (file_ptr & FFS_MASK);
1965 file_ptr &= ~FFS_MASK;
1966 /* mask in overlapping REQ_F and FFS bits */
1967 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
1968 io_req_set_rsrc_node(req, ctx, 0);
1970 io_ring_submit_unlock(ctx, issue_flags);
1974 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
1976 struct file *file = fget(fd);
1978 trace_io_uring_file_get(req, fd);
1980 /* we don't allow fixed io_uring files */
1981 if (file && io_is_uring_fops(file))
1982 io_req_track_inflight(req);
1986 static void io_queue_async(struct io_kiocb *req, int ret)
1987 __must_hold(&req->ctx->uring_lock)
1989 struct io_kiocb *linked_timeout;
1991 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
1992 io_req_defer_failed(req, ret);
1996 linked_timeout = io_prep_linked_timeout(req);
1998 switch (io_arm_poll_handler(req, 0)) {
1999 case IO_APOLL_READY:
2000 io_kbuf_recycle(req, 0);
2001 io_req_task_queue(req);
2003 case IO_APOLL_ABORTED:
2004 io_kbuf_recycle(req, 0);
2005 io_queue_iowq(req, NULL);
2012 io_queue_linked_timeout(linked_timeout);
2015 static inline void io_queue_sqe(struct io_kiocb *req)
2016 __must_hold(&req->ctx->uring_lock)
2020 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
2023 * We async punt it if the file wasn't marked NOWAIT, or if the file
2024 * doesn't support non-blocking read/write attempts
2027 io_arm_ltimeout(req);
2029 io_queue_async(req, ret);
2032 static void io_queue_sqe_fallback(struct io_kiocb *req)
2033 __must_hold(&req->ctx->uring_lock)
2035 if (unlikely(req->flags & REQ_F_FAIL)) {
2037 * We don't submit, fail them all, for that replace hardlinks
2038 * with normal links. Extra REQ_F_LINK is tolerated.
2040 req->flags &= ~REQ_F_HARDLINK;
2041 req->flags |= REQ_F_LINK;
2042 io_req_defer_failed(req, req->cqe.res);
2044 int ret = io_req_prep_async(req);
2046 if (unlikely(ret)) {
2047 io_req_defer_failed(req, ret);
2051 if (unlikely(req->ctx->drain_active))
2054 io_queue_iowq(req, NULL);
2059 * Check SQE restrictions (opcode and flags).
2061 * Returns 'true' if SQE is allowed, 'false' otherwise.
2063 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
2064 struct io_kiocb *req,
2065 unsigned int sqe_flags)
2067 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
2070 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
2071 ctx->restrictions.sqe_flags_required)
2074 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
2075 ctx->restrictions.sqe_flags_required))
2081 static void io_init_req_drain(struct io_kiocb *req)
2083 struct io_ring_ctx *ctx = req->ctx;
2084 struct io_kiocb *head = ctx->submit_state.link.head;
2086 ctx->drain_active = true;
2089 * If we need to drain a request in the middle of a link, drain
2090 * the head request and the next request/link after the current
2091 * link. Considering sequential execution of links,
2092 * REQ_F_IO_DRAIN will be maintained for every request of our
2095 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2096 ctx->drain_next = true;
2100 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
2101 const struct io_uring_sqe *sqe)
2102 __must_hold(&ctx->uring_lock)
2104 const struct io_op_def *def;
2105 unsigned int sqe_flags;
2109 /* req is partially pre-initialised, see io_preinit_req() */
2110 req->opcode = opcode = READ_ONCE(sqe->opcode);
2111 /* same numerical values with corresponding REQ_F_*, safe to copy */
2112 req->flags = sqe_flags = READ_ONCE(sqe->flags);
2113 req->cqe.user_data = READ_ONCE(sqe->user_data);
2115 req->rsrc_node = NULL;
2116 req->task = current;
2118 if (unlikely(opcode >= IORING_OP_LAST)) {
2122 def = &io_op_defs[opcode];
2123 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2124 /* enforce forwards compatibility on users */
2125 if (sqe_flags & ~SQE_VALID_FLAGS)
2127 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2128 if (!def->buffer_select)
2130 req->buf_index = READ_ONCE(sqe->buf_group);
2132 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2133 ctx->drain_disabled = true;
2134 if (sqe_flags & IOSQE_IO_DRAIN) {
2135 if (ctx->drain_disabled)
2137 io_init_req_drain(req);
2140 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2141 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2143 /* knock it to the slow queue path, will be drained there */
2144 if (ctx->drain_active)
2145 req->flags |= REQ_F_FORCE_ASYNC;
2146 /* if there is no link, we're at "next" request and need to drain */
2147 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2148 ctx->drain_next = false;
2149 ctx->drain_active = true;
2150 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2154 if (!def->ioprio && sqe->ioprio)
2156 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2159 if (def->needs_file) {
2160 struct io_submit_state *state = &ctx->submit_state;
2162 req->cqe.fd = READ_ONCE(sqe->fd);
2165 * Plug now if we have more than 2 IO left after this, and the
2166 * target is potentially a read/write to block based storage.
2168 if (state->need_plug && def->plug) {
2169 state->plug_started = true;
2170 state->need_plug = false;
2171 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2175 personality = READ_ONCE(sqe->personality);
2179 req->creds = xa_load(&ctx->personalities, personality);
2182 get_cred(req->creds);
2183 ret = security_uring_override_creds(req->creds);
2185 put_cred(req->creds);
2188 req->flags |= REQ_F_CREDS;
2191 return def->prep(req, sqe);
2194 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2195 struct io_kiocb *req, int ret)
2197 struct io_ring_ctx *ctx = req->ctx;
2198 struct io_submit_link *link = &ctx->submit_state.link;
2199 struct io_kiocb *head = link->head;
2201 trace_io_uring_req_failed(sqe, req, ret);
2204 * Avoid breaking links in the middle as it renders links with SQPOLL
2205 * unusable. Instead of failing eagerly, continue assembling the link if
2206 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2207 * should find the flag and handle the rest.
2209 req_fail_link_node(req, ret);
2210 if (head && !(head->flags & REQ_F_FAIL))
2211 req_fail_link_node(head, -ECANCELED);
2213 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2215 link->last->link = req;
2219 io_queue_sqe_fallback(req);
2224 link->last->link = req;
2231 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2232 const struct io_uring_sqe *sqe)
2233 __must_hold(&ctx->uring_lock)
2235 struct io_submit_link *link = &ctx->submit_state.link;
2238 ret = io_init_req(ctx, req, sqe);
2240 return io_submit_fail_init(sqe, req, ret);
2242 /* don't need @sqe from now on */
2243 trace_io_uring_submit_sqe(req, true);
2246 * If we already have a head request, queue this one for async
2247 * submittal once the head completes. If we don't have a head but
2248 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2249 * submitted sync once the chain is complete. If none of those
2250 * conditions are true (normal request), then just queue it.
2252 if (unlikely(link->head)) {
2253 ret = io_req_prep_async(req);
2255 return io_submit_fail_init(sqe, req, ret);
2257 trace_io_uring_link(req, link->head);
2258 link->last->link = req;
2261 if (req->flags & IO_REQ_LINK_FLAGS)
2263 /* last request of the link, flush it */
2266 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2269 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2270 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2271 if (req->flags & IO_REQ_LINK_FLAGS) {
2276 io_queue_sqe_fallback(req);
2286 * Batched submission is done, ensure local IO is flushed out.
2288 static void io_submit_state_end(struct io_ring_ctx *ctx)
2290 struct io_submit_state *state = &ctx->submit_state;
2292 if (unlikely(state->link.head))
2293 io_queue_sqe_fallback(state->link.head);
2294 /* flush only after queuing links as they can generate completions */
2295 io_submit_flush_completions(ctx);
2296 if (state->plug_started)
2297 blk_finish_plug(&state->plug);
2301 * Start submission side cache.
2303 static void io_submit_state_start(struct io_submit_state *state,
2304 unsigned int max_ios)
2306 state->plug_started = false;
2307 state->need_plug = max_ios > 2;
2308 state->submit_nr = max_ios;
2309 /* set only head, no need to init link_last in advance */
2310 state->link.head = NULL;
2313 static void io_commit_sqring(struct io_ring_ctx *ctx)
2315 struct io_rings *rings = ctx->rings;
2318 * Ensure any loads from the SQEs are done at this point,
2319 * since once we write the new head, the application could
2320 * write new data to them.
2322 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2326 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2327 * that is mapped by userspace. This means that care needs to be taken to
2328 * ensure that reads are stable, as we cannot rely on userspace always
2329 * being a good citizen. If members of the sqe are validated and then later
2330 * used, it's important that those reads are done through READ_ONCE() to
2331 * prevent a re-load down the line.
2333 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
2335 unsigned head, mask = ctx->sq_entries - 1;
2336 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2339 * The cached sq head (or cq tail) serves two purposes:
2341 * 1) allows us to batch the cost of updating the user visible
2343 * 2) allows the kernel side to track the head on its own, even
2344 * though the application is the one updating it.
2346 head = READ_ONCE(ctx->sq_array[sq_idx]);
2347 if (likely(head < ctx->sq_entries)) {
2348 /* double index for 128-byte SQEs, twice as long */
2349 if (ctx->flags & IORING_SETUP_SQE128)
2351 return &ctx->sq_sqes[head];
2354 /* drop invalid entries */
2356 WRITE_ONCE(ctx->rings->sq_dropped,
2357 READ_ONCE(ctx->rings->sq_dropped) + 1);
2361 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2362 __must_hold(&ctx->uring_lock)
2364 unsigned int entries = io_sqring_entries(ctx);
2368 if (unlikely(!entries))
2370 /* make sure SQ entry isn't read before tail */
2371 ret = left = min3(nr, ctx->sq_entries, entries);
2372 io_get_task_refs(left);
2373 io_submit_state_start(&ctx->submit_state, left);
2376 const struct io_uring_sqe *sqe;
2377 struct io_kiocb *req;
2379 if (unlikely(!io_alloc_req_refill(ctx)))
2381 req = io_alloc_req(ctx);
2382 sqe = io_get_sqe(ctx);
2383 if (unlikely(!sqe)) {
2384 io_req_add_to_cache(req, ctx);
2389 * Continue submitting even for sqe failure if the
2390 * ring was setup with IORING_SETUP_SUBMIT_ALL
2392 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2393 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2399 if (unlikely(left)) {
2401 /* try again if it submitted nothing and can't allocate a req */
2402 if (!ret && io_req_cache_empty(ctx))
2404 current->io_uring->cached_refs += left;
2407 io_submit_state_end(ctx);
2408 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2409 io_commit_sqring(ctx);
2413 struct io_wait_queue {
2414 struct wait_queue_entry wq;
2415 struct io_ring_ctx *ctx;
2417 unsigned nr_timeouts;
2420 static inline bool io_has_work(struct io_ring_ctx *ctx)
2422 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
2423 ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
2424 !llist_empty(&ctx->work_llist));
2427 static inline bool io_should_wake(struct io_wait_queue *iowq)
2429 struct io_ring_ctx *ctx = iowq->ctx;
2430 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
2433 * Wake up if we have enough events, or if a timeout occurred since we
2434 * started waiting. For timeouts, we always want to return to userspace,
2435 * regardless of event count.
2437 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2440 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2441 int wake_flags, void *key)
2443 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
2445 struct io_ring_ctx *ctx = iowq->ctx;
2448 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2449 * the task, and the next invocation will do it.
2451 if (io_should_wake(iowq) || io_has_work(ctx))
2452 return autoremove_wake_function(curr, mode, wake_flags, key);
2456 int io_run_task_work_sig(struct io_ring_ctx *ctx)
2458 if (io_run_task_work_ctx(ctx) > 0)
2460 if (task_sigpending(current))
2465 /* when returns >0, the caller should retry */
2466 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2467 struct io_wait_queue *iowq,
2471 unsigned long check_cq;
2473 /* make sure we run task_work before checking for signals */
2474 ret = io_run_task_work_sig(ctx);
2475 if (ret || io_should_wake(iowq))
2478 check_cq = READ_ONCE(ctx->check_cq);
2479 if (unlikely(check_cq)) {
2480 /* let the caller flush overflows, retry */
2481 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2483 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
2486 if (!schedule_hrtimeout(timeout, HRTIMER_MODE_ABS))
2490 * Run task_work after scheduling. If we got woken because of
2491 * task_work being processed, run it now rather than let the caller
2492 * do another wait loop.
2494 ret = io_run_task_work_sig(ctx);
2495 return ret < 0 ? ret : 1;
2499 * Wait until events become available, if we don't already have some. The
2500 * application must reap them itself, as they reside on the shared cq ring.
2502 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2503 const sigset_t __user *sig, size_t sigsz,
2504 struct __kernel_timespec __user *uts)
2506 struct io_wait_queue iowq;
2507 struct io_rings *rings = ctx->rings;
2508 ktime_t timeout = KTIME_MAX;
2511 if (!io_allowed_run_tw(ctx))
2513 if (!llist_empty(&ctx->work_llist)) {
2514 ret = io_run_local_work(ctx);
2519 io_cqring_overflow_flush(ctx);
2520 /* if user messes with these they will just get an early return */
2521 if (__io_cqring_events_user(ctx) >= min_events)
2525 #ifdef CONFIG_COMPAT
2526 if (in_compat_syscall())
2527 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2531 ret = set_user_sigmask(sig, sigsz);
2538 struct timespec64 ts;
2540 if (get_timespec64(&ts, uts))
2542 timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2545 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2546 iowq.wq.private = current;
2547 INIT_LIST_HEAD(&iowq.wq.entry);
2549 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2550 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2552 trace_io_uring_cqring_wait(ctx, min_events);
2554 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
2555 finish_wait(&ctx->cq_wait, &iowq.wq);
2556 io_cqring_do_overflow_flush(ctx);
2558 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2559 TASK_INTERRUPTIBLE);
2560 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
2561 if (__io_cqring_events_user(ctx) >= min_events)
2566 finish_wait(&ctx->cq_wait, &iowq.wq);
2567 restore_saved_sigmask_unless(ret == -EINTR);
2569 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2572 static void io_mem_free(void *ptr)
2579 page = virt_to_head_page(ptr);
2580 if (put_page_testzero(page))
2581 free_compound_page(page);
2584 static void *io_mem_alloc(size_t size)
2586 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2588 return (void *) __get_free_pages(gfp, get_order(size));
2591 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2592 unsigned int cq_entries, size_t *sq_offset)
2594 struct io_rings *rings;
2595 size_t off, sq_array_size;
2597 off = struct_size(rings, cqes, cq_entries);
2598 if (off == SIZE_MAX)
2600 if (ctx->flags & IORING_SETUP_CQE32) {
2601 if (check_shl_overflow(off, 1, &off))
2606 off = ALIGN(off, SMP_CACHE_BYTES);
2614 sq_array_size = array_size(sizeof(u32), sq_entries);
2615 if (sq_array_size == SIZE_MAX)
2618 if (check_add_overflow(off, sq_array_size, &off))
2624 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2625 unsigned int eventfd_async)
2627 struct io_ev_fd *ev_fd;
2628 __s32 __user *fds = arg;
2631 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2632 lockdep_is_held(&ctx->uring_lock));
2636 if (copy_from_user(&fd, fds, sizeof(*fds)))
2639 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2643 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2644 if (IS_ERR(ev_fd->cq_ev_fd)) {
2645 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2650 spin_lock(&ctx->completion_lock);
2651 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2652 spin_unlock(&ctx->completion_lock);
2654 ev_fd->eventfd_async = eventfd_async;
2655 ctx->has_evfd = true;
2656 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2657 atomic_set(&ev_fd->refs, 1);
2658 atomic_set(&ev_fd->ops, 0);
2662 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2664 struct io_ev_fd *ev_fd;
2666 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2667 lockdep_is_held(&ctx->uring_lock));
2669 ctx->has_evfd = false;
2670 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2671 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops))
2672 call_rcu(&ev_fd->rcu, io_eventfd_ops);
2679 static void io_req_caches_free(struct io_ring_ctx *ctx)
2683 mutex_lock(&ctx->uring_lock);
2684 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
2686 while (!io_req_cache_empty(ctx)) {
2687 struct io_kiocb *req = io_alloc_req(ctx);
2689 kmem_cache_free(req_cachep, req);
2693 percpu_ref_put_many(&ctx->refs, nr);
2694 mutex_unlock(&ctx->uring_lock);
2697 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2699 io_sq_thread_finish(ctx);
2700 io_rsrc_refs_drop(ctx);
2701 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2702 io_wait_rsrc_data(ctx->buf_data);
2703 io_wait_rsrc_data(ctx->file_data);
2705 mutex_lock(&ctx->uring_lock);
2707 __io_sqe_buffers_unregister(ctx);
2709 __io_sqe_files_unregister(ctx);
2710 io_cqring_overflow_kill(ctx);
2711 io_eventfd_unregister(ctx);
2712 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2713 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2714 mutex_unlock(&ctx->uring_lock);
2715 io_destroy_buffers(ctx);
2717 put_cred(ctx->sq_creds);
2718 if (ctx->submitter_task)
2719 put_task_struct(ctx->submitter_task);
2721 /* there are no registered resources left, nobody uses it */
2723 io_rsrc_node_destroy(ctx->rsrc_node);
2724 if (ctx->rsrc_backup_node)
2725 io_rsrc_node_destroy(ctx->rsrc_backup_node);
2726 flush_delayed_work(&ctx->rsrc_put_work);
2727 flush_delayed_work(&ctx->fallback_work);
2729 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2730 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
2732 #if defined(CONFIG_UNIX)
2733 if (ctx->ring_sock) {
2734 ctx->ring_sock->file = NULL; /* so that iput() is called */
2735 sock_release(ctx->ring_sock);
2738 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2740 if (ctx->mm_account) {
2741 mmdrop(ctx->mm_account);
2742 ctx->mm_account = NULL;
2744 io_mem_free(ctx->rings);
2745 io_mem_free(ctx->sq_sqes);
2747 percpu_ref_exit(&ctx->refs);
2748 free_uid(ctx->user);
2749 io_req_caches_free(ctx);
2751 io_wq_put_hash(ctx->hash_map);
2752 kfree(ctx->cancel_table.hbs);
2753 kfree(ctx->cancel_table_locked.hbs);
2754 kfree(ctx->dummy_ubuf);
2756 xa_destroy(&ctx->io_bl_xa);
2760 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2762 struct io_ring_ctx *ctx = file->private_data;
2765 poll_wait(file, &ctx->cq_wait, wait);
2767 * synchronizes with barrier from wq_has_sleeper call in
2771 if (!io_sqring_full(ctx))
2772 mask |= EPOLLOUT | EPOLLWRNORM;
2775 * Don't flush cqring overflow list here, just do a simple check.
2776 * Otherwise there could possible be ABBA deadlock:
2779 * lock(&ctx->uring_lock);
2781 * lock(&ctx->uring_lock);
2784 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2785 * pushes them to do the flush.
2788 if (io_cqring_events(ctx) || io_has_work(ctx))
2789 mask |= EPOLLIN | EPOLLRDNORM;
2794 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2796 const struct cred *creds;
2798 creds = xa_erase(&ctx->personalities, id);
2807 struct io_tctx_exit {
2808 struct callback_head task_work;
2809 struct completion completion;
2810 struct io_ring_ctx *ctx;
2813 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2815 struct io_uring_task *tctx = current->io_uring;
2816 struct io_tctx_exit *work;
2818 work = container_of(cb, struct io_tctx_exit, task_work);
2820 * When @in_idle, we're in cancellation and it's racy to remove the
2821 * node. It'll be removed by the end of cancellation, just ignore it.
2822 * tctx can be NULL if the queueing of this task_work raced with
2823 * work cancelation off the exec path.
2825 if (tctx && !atomic_read(&tctx->in_idle))
2826 io_uring_del_tctx_node((unsigned long)work->ctx);
2827 complete(&work->completion);
2830 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2832 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2834 return req->ctx == data;
2837 static __cold void io_ring_exit_work(struct work_struct *work)
2839 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2840 unsigned long timeout = jiffies + HZ * 60 * 5;
2841 unsigned long interval = HZ / 20;
2842 struct io_tctx_exit exit;
2843 struct io_tctx_node *node;
2847 * If we're doing polled IO and end up having requests being
2848 * submitted async (out-of-line), then completions can come in while
2849 * we're waiting for refs to drop. We need to reap these manually,
2850 * as nobody else will be looking for them.
2853 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
2854 mutex_lock(&ctx->uring_lock);
2855 io_cqring_overflow_kill(ctx);
2856 mutex_unlock(&ctx->uring_lock);
2859 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
2860 io_move_task_work_from_local(ctx);
2862 while (io_uring_try_cancel_requests(ctx, NULL, true))
2866 struct io_sq_data *sqd = ctx->sq_data;
2867 struct task_struct *tsk;
2869 io_sq_thread_park(sqd);
2871 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
2872 io_wq_cancel_cb(tsk->io_uring->io_wq,
2873 io_cancel_ctx_cb, ctx, true);
2874 io_sq_thread_unpark(sqd);
2877 io_req_caches_free(ctx);
2879 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
2880 /* there is little hope left, don't run it too often */
2883 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
2885 init_completion(&exit.completion);
2886 init_task_work(&exit.task_work, io_tctx_exit_cb);
2889 * Some may use context even when all refs and requests have been put,
2890 * and they are free to do so while still holding uring_lock or
2891 * completion_lock, see io_req_task_submit(). Apart from other work,
2892 * this lock/unlock section also waits them to finish.
2894 mutex_lock(&ctx->uring_lock);
2895 while (!list_empty(&ctx->tctx_list)) {
2896 WARN_ON_ONCE(time_after(jiffies, timeout));
2898 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
2900 /* don't spin on a single task if cancellation failed */
2901 list_rotate_left(&ctx->tctx_list);
2902 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
2903 if (WARN_ON_ONCE(ret))
2906 mutex_unlock(&ctx->uring_lock);
2907 wait_for_completion(&exit.completion);
2908 mutex_lock(&ctx->uring_lock);
2910 mutex_unlock(&ctx->uring_lock);
2911 spin_lock(&ctx->completion_lock);
2912 spin_unlock(&ctx->completion_lock);
2914 io_ring_ctx_free(ctx);
2917 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
2919 unsigned long index;
2920 struct creds *creds;
2922 mutex_lock(&ctx->uring_lock);
2923 percpu_ref_kill(&ctx->refs);
2924 xa_for_each(&ctx->personalities, index, creds)
2925 io_unregister_personality(ctx, index);
2927 io_poll_remove_all(ctx, NULL, true);
2928 mutex_unlock(&ctx->uring_lock);
2931 * If we failed setting up the ctx, we might not have any rings
2932 * and therefore did not submit any requests
2935 io_kill_timeouts(ctx, NULL, true);
2937 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
2939 * Use system_unbound_wq to avoid spawning tons of event kworkers
2940 * if we're exiting a ton of rings at the same time. It just adds
2941 * noise and overhead, there's no discernable change in runtime
2942 * over using system_wq.
2944 queue_work(system_unbound_wq, &ctx->exit_work);
2947 static int io_uring_release(struct inode *inode, struct file *file)
2949 struct io_ring_ctx *ctx = file->private_data;
2951 file->private_data = NULL;
2952 io_ring_ctx_wait_and_kill(ctx);
2956 struct io_task_cancel {
2957 struct task_struct *task;
2961 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
2963 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2964 struct io_task_cancel *cancel = data;
2966 return io_match_task_safe(req, cancel->task, cancel->all);
2969 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
2970 struct task_struct *task,
2973 struct io_defer_entry *de;
2976 spin_lock(&ctx->completion_lock);
2977 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
2978 if (io_match_task_safe(de->req, task, cancel_all)) {
2979 list_cut_position(&list, &ctx->defer_list, &de->list);
2983 spin_unlock(&ctx->completion_lock);
2984 if (list_empty(&list))
2987 while (!list_empty(&list)) {
2988 de = list_first_entry(&list, struct io_defer_entry, list);
2989 list_del_init(&de->list);
2990 io_req_task_queue_fail(de->req, -ECANCELED);
2996 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
2998 struct io_tctx_node *node;
2999 enum io_wq_cancel cret;
3002 mutex_lock(&ctx->uring_lock);
3003 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3004 struct io_uring_task *tctx = node->task->io_uring;
3007 * io_wq will stay alive while we hold uring_lock, because it's
3008 * killed after ctx nodes, which requires to take the lock.
3010 if (!tctx || !tctx->io_wq)
3012 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
3013 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3015 mutex_unlock(&ctx->uring_lock);
3020 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
3021 struct task_struct *task,
3024 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
3025 struct io_uring_task *tctx = task ? task->io_uring : NULL;
3026 enum io_wq_cancel cret;
3029 /* failed during ring init, it couldn't have issued any requests */
3034 ret |= io_uring_try_cancel_iowq(ctx);
3035 } else if (tctx && tctx->io_wq) {
3037 * Cancels requests of all rings, not only @ctx, but
3038 * it's fine as the task is in exit/exec.
3040 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
3042 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3045 /* SQPOLL thread does its own polling */
3046 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
3047 (ctx->sq_data && ctx->sq_data->thread == current)) {
3048 while (!wq_list_empty(&ctx->iopoll_list)) {
3049 io_iopoll_try_reap_events(ctx);
3054 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3055 ret |= io_run_local_work(ctx) > 0;
3056 ret |= io_cancel_defer_files(ctx, task, cancel_all);
3057 mutex_lock(&ctx->uring_lock);
3058 ret |= io_poll_remove_all(ctx, task, cancel_all);
3059 mutex_unlock(&ctx->uring_lock);
3060 ret |= io_kill_timeouts(ctx, task, cancel_all);
3062 ret |= io_run_task_work() > 0;
3066 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
3069 return atomic_read(&tctx->inflight_tracked);
3070 return percpu_counter_sum(&tctx->inflight);
3074 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3075 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3077 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
3079 struct io_uring_task *tctx = current->io_uring;
3080 struct io_ring_ctx *ctx;
3084 WARN_ON_ONCE(sqd && sqd->thread != current);
3086 if (!current->io_uring)
3089 io_wq_exit_start(tctx->io_wq);
3091 atomic_inc(&tctx->in_idle);
3095 io_uring_drop_tctx_refs(current);
3096 /* read completions before cancelations */
3097 inflight = tctx_inflight(tctx, !cancel_all);
3102 struct io_tctx_node *node;
3103 unsigned long index;
3105 xa_for_each(&tctx->xa, index, node) {
3106 /* sqpoll task will cancel all its requests */
3107 if (node->ctx->sq_data)
3109 loop |= io_uring_try_cancel_requests(node->ctx,
3110 current, cancel_all);
3113 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3114 loop |= io_uring_try_cancel_requests(ctx,
3124 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3126 io_uring_drop_tctx_refs(current);
3129 * If we've seen completions, retry without waiting. This
3130 * avoids a race where a completion comes in before we did
3131 * prepare_to_wait().
3133 if (inflight == tctx_inflight(tctx, !cancel_all))
3135 finish_wait(&tctx->wait, &wait);
3138 io_uring_clean_tctx(tctx);
3141 * We shouldn't run task_works after cancel, so just leave
3142 * ->in_idle set for normal exit.
3144 atomic_dec(&tctx->in_idle);
3145 /* for exec all current's requests should be gone, kill tctx */
3146 __io_uring_free(current);
3150 void __io_uring_cancel(bool cancel_all)
3152 io_uring_cancel_generic(cancel_all, NULL);
3155 static void *io_uring_validate_mmap_request(struct file *file,
3156 loff_t pgoff, size_t sz)
3158 struct io_ring_ctx *ctx = file->private_data;
3159 loff_t offset = pgoff << PAGE_SHIFT;
3164 case IORING_OFF_SQ_RING:
3165 case IORING_OFF_CQ_RING:
3168 case IORING_OFF_SQES:
3172 return ERR_PTR(-EINVAL);
3175 page = virt_to_head_page(ptr);
3176 if (sz > page_size(page))
3177 return ERR_PTR(-EINVAL);
3184 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3186 size_t sz = vma->vm_end - vma->vm_start;
3190 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3192 return PTR_ERR(ptr);
3194 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3195 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3198 #else /* !CONFIG_MMU */
3200 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3202 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
3205 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3207 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3210 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3211 unsigned long addr, unsigned long len,
3212 unsigned long pgoff, unsigned long flags)
3216 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3218 return PTR_ERR(ptr);
3220 return (unsigned long) ptr;
3223 #endif /* !CONFIG_MMU */
3225 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3227 if (flags & IORING_ENTER_EXT_ARG) {
3228 struct io_uring_getevents_arg arg;
3230 if (argsz != sizeof(arg))
3232 if (copy_from_user(&arg, argp, sizeof(arg)))
3238 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3239 struct __kernel_timespec __user **ts,
3240 const sigset_t __user **sig)
3242 struct io_uring_getevents_arg arg;
3245 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3246 * is just a pointer to the sigset_t.
3248 if (!(flags & IORING_ENTER_EXT_ARG)) {
3249 *sig = (const sigset_t __user *) argp;
3255 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3256 * timespec and sigset_t pointers if good.
3258 if (*argsz != sizeof(arg))
3260 if (copy_from_user(&arg, argp, sizeof(arg)))
3264 *sig = u64_to_user_ptr(arg.sigmask);
3265 *argsz = arg.sigmask_sz;
3266 *ts = u64_to_user_ptr(arg.ts);
3270 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3271 u32, min_complete, u32, flags, const void __user *, argp,
3274 struct io_ring_ctx *ctx;
3278 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3279 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3280 IORING_ENTER_REGISTERED_RING)))
3284 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3285 * need only dereference our task private array to find it.
3287 if (flags & IORING_ENTER_REGISTERED_RING) {
3288 struct io_uring_task *tctx = current->io_uring;
3290 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3292 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3293 f.file = tctx->registered_rings[fd];
3295 if (unlikely(!f.file))
3299 if (unlikely(!f.file))
3302 if (unlikely(!io_is_uring_fops(f.file)))
3306 ctx = f.file->private_data;
3308 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3312 * For SQ polling, the thread will do all submissions and completions.
3313 * Just return the requested submit count, and wake the thread if
3317 if (ctx->flags & IORING_SETUP_SQPOLL) {
3318 io_cqring_overflow_flush(ctx);
3320 if (unlikely(ctx->sq_data->thread == NULL)) {
3324 if (flags & IORING_ENTER_SQ_WAKEUP)
3325 wake_up(&ctx->sq_data->wait);
3326 if (flags & IORING_ENTER_SQ_WAIT) {
3327 ret = io_sqpoll_wait_sq(ctx);
3332 } else if (to_submit) {
3333 ret = io_uring_add_tctx_node(ctx);
3337 mutex_lock(&ctx->uring_lock);
3338 ret = io_submit_sqes(ctx, to_submit);
3339 if (ret != to_submit) {
3340 mutex_unlock(&ctx->uring_lock);
3343 if (flags & IORING_ENTER_GETEVENTS) {
3344 if (ctx->syscall_iopoll)
3347 * Ignore errors, we'll soon call io_cqring_wait() and
3348 * it should handle ownership problems if any.
3350 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3351 (void)io_run_local_work_locked(ctx);
3353 mutex_unlock(&ctx->uring_lock);
3356 if (flags & IORING_ENTER_GETEVENTS) {
3359 if (ctx->syscall_iopoll) {
3361 * We disallow the app entering submit/complete with
3362 * polling, but we still need to lock the ring to
3363 * prevent racing with polled issue that got punted to
3366 mutex_lock(&ctx->uring_lock);
3368 ret2 = io_validate_ext_arg(flags, argp, argsz);
3369 if (likely(!ret2)) {
3370 min_complete = min(min_complete,
3372 ret2 = io_iopoll_check(ctx, min_complete);
3374 mutex_unlock(&ctx->uring_lock);
3376 const sigset_t __user *sig;
3377 struct __kernel_timespec __user *ts;
3379 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3380 if (likely(!ret2)) {
3381 min_complete = min(min_complete,
3383 ret2 = io_cqring_wait(ctx, min_complete, sig,
3392 * EBADR indicates that one or more CQE were dropped.
3393 * Once the user has been informed we can clear the bit
3394 * as they are obviously ok with those drops.
3396 if (unlikely(ret2 == -EBADR))
3397 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3406 static const struct file_operations io_uring_fops = {
3407 .release = io_uring_release,
3408 .mmap = io_uring_mmap,
3410 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3411 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3413 .poll = io_uring_poll,
3414 #ifdef CONFIG_PROC_FS
3415 .show_fdinfo = io_uring_show_fdinfo,
3419 bool io_is_uring_fops(struct file *file)
3421 return file->f_op == &io_uring_fops;
3424 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3425 struct io_uring_params *p)
3427 struct io_rings *rings;
3428 size_t size, sq_array_offset;
3430 /* make sure these are sane, as we already accounted them */
3431 ctx->sq_entries = p->sq_entries;
3432 ctx->cq_entries = p->cq_entries;
3434 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3435 if (size == SIZE_MAX)
3438 rings = io_mem_alloc(size);
3443 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3444 rings->sq_ring_mask = p->sq_entries - 1;
3445 rings->cq_ring_mask = p->cq_entries - 1;
3446 rings->sq_ring_entries = p->sq_entries;
3447 rings->cq_ring_entries = p->cq_entries;
3449 if (p->flags & IORING_SETUP_SQE128)
3450 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3452 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3453 if (size == SIZE_MAX) {
3454 io_mem_free(ctx->rings);
3459 ctx->sq_sqes = io_mem_alloc(size);
3460 if (!ctx->sq_sqes) {
3461 io_mem_free(ctx->rings);
3469 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3473 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3477 ret = __io_uring_add_tctx_node(ctx);
3482 fd_install(fd, file);
3487 * Allocate an anonymous fd, this is what constitutes the application
3488 * visible backing of an io_uring instance. The application mmaps this
3489 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3490 * we have to tie this fd to a socket for file garbage collection purposes.
3492 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3495 #if defined(CONFIG_UNIX)
3498 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3501 return ERR_PTR(ret);
3504 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3505 O_RDWR | O_CLOEXEC, NULL);
3506 #if defined(CONFIG_UNIX)
3508 sock_release(ctx->ring_sock);
3509 ctx->ring_sock = NULL;
3511 ctx->ring_sock->file = file;
3517 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3518 struct io_uring_params __user *params)
3520 struct io_ring_ctx *ctx;
3526 if (entries > IORING_MAX_ENTRIES) {
3527 if (!(p->flags & IORING_SETUP_CLAMP))
3529 entries = IORING_MAX_ENTRIES;
3533 * Use twice as many entries for the CQ ring. It's possible for the
3534 * application to drive a higher depth than the size of the SQ ring,
3535 * since the sqes are only used at submission time. This allows for
3536 * some flexibility in overcommitting a bit. If the application has
3537 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3538 * of CQ ring entries manually.
3540 p->sq_entries = roundup_pow_of_two(entries);
3541 if (p->flags & IORING_SETUP_CQSIZE) {
3543 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3544 * to a power-of-two, if it isn't already. We do NOT impose
3545 * any cq vs sq ring sizing.
3549 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3550 if (!(p->flags & IORING_SETUP_CLAMP))
3552 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3554 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3555 if (p->cq_entries < p->sq_entries)
3558 p->cq_entries = 2 * p->sq_entries;
3561 ctx = io_ring_ctx_alloc(p);
3565 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3566 !(ctx->flags & IORING_SETUP_IOPOLL) &&
3567 !(ctx->flags & IORING_SETUP_SQPOLL))
3568 ctx->task_complete = true;
3571 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3572 * space applications don't need to do io completion events
3573 * polling again, they can rely on io_sq_thread to do polling
3574 * work, which can reduce cpu usage and uring_lock contention.
3576 if (ctx->flags & IORING_SETUP_IOPOLL &&
3577 !(ctx->flags & IORING_SETUP_SQPOLL))
3578 ctx->syscall_iopoll = 1;
3580 ctx->compat = in_compat_syscall();
3581 if (!capable(CAP_IPC_LOCK))
3582 ctx->user = get_uid(current_user());
3585 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3586 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3589 if (ctx->flags & IORING_SETUP_SQPOLL) {
3590 /* IPI related flags don't make sense with SQPOLL */
3591 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3592 IORING_SETUP_TASKRUN_FLAG |
3593 IORING_SETUP_DEFER_TASKRUN))
3595 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3596 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3597 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3599 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3600 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3602 ctx->notify_method = TWA_SIGNAL;
3606 * For DEFER_TASKRUN we require the completion task to be the same as the
3607 * submission task. This implies that there is only one submitter, so enforce
3610 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3611 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3616 * This is just grabbed for accounting purposes. When a process exits,
3617 * the mm is exited and dropped before the files, hence we need to hang
3618 * on to this mm purely for the purposes of being able to unaccount
3619 * memory (locked/pinned vm). It's not used for anything else.
3621 mmgrab(current->mm);
3622 ctx->mm_account = current->mm;
3624 ret = io_allocate_scq_urings(ctx, p);
3628 ret = io_sq_offload_create(ctx, p);
3631 /* always set a rsrc node */
3632 ret = io_rsrc_node_switch_start(ctx);
3635 io_rsrc_node_switch(ctx, NULL);
3637 memset(&p->sq_off, 0, sizeof(p->sq_off));
3638 p->sq_off.head = offsetof(struct io_rings, sq.head);
3639 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3640 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3641 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3642 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3643 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3644 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3646 memset(&p->cq_off, 0, sizeof(p->cq_off));
3647 p->cq_off.head = offsetof(struct io_rings, cq.head);
3648 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3649 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3650 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3651 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3652 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3653 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3655 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3656 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3657 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3658 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3659 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3660 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3661 IORING_FEAT_LINKED_FILE;
3663 if (copy_to_user(params, p, sizeof(*p))) {
3668 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
3669 && !(ctx->flags & IORING_SETUP_R_DISABLED))
3670 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
3672 file = io_uring_get_file(ctx);
3674 ret = PTR_ERR(file);
3679 * Install ring fd as the very last thing, so we don't risk someone
3680 * having closed it before we finish setup
3682 ret = io_uring_install_fd(ctx, file);
3684 /* fput will clean it up */
3689 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3692 io_ring_ctx_wait_and_kill(ctx);
3697 * Sets up an aio uring context, and returns the fd. Applications asks for a
3698 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3699 * params structure passed in.
3701 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3703 struct io_uring_params p;
3706 if (copy_from_user(&p, params, sizeof(p)))
3708 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3713 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3714 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3715 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3716 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3717 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3718 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3719 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN))
3722 return io_uring_create(entries, &p, params);
3725 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3726 struct io_uring_params __user *, params)
3728 return io_uring_setup(entries, params);
3731 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3734 struct io_uring_probe *p;
3738 size = struct_size(p, ops, nr_args);
3739 if (size == SIZE_MAX)
3741 p = kzalloc(size, GFP_KERNEL);
3746 if (copy_from_user(p, arg, size))
3749 if (memchr_inv(p, 0, size))
3752 p->last_op = IORING_OP_LAST - 1;
3753 if (nr_args > IORING_OP_LAST)
3754 nr_args = IORING_OP_LAST;
3756 for (i = 0; i < nr_args; i++) {
3758 if (!io_op_defs[i].not_supported)
3759 p->ops[i].flags = IO_URING_OP_SUPPORTED;
3764 if (copy_to_user(arg, p, size))
3771 static int io_register_personality(struct io_ring_ctx *ctx)
3773 const struct cred *creds;
3777 creds = get_current_cred();
3779 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
3780 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
3788 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
3789 void __user *arg, unsigned int nr_args)
3791 struct io_uring_restriction *res;
3795 /* Restrictions allowed only if rings started disabled */
3796 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3799 /* We allow only a single restrictions registration */
3800 if (ctx->restrictions.registered)
3803 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
3806 size = array_size(nr_args, sizeof(*res));
3807 if (size == SIZE_MAX)
3810 res = memdup_user(arg, size);
3812 return PTR_ERR(res);
3816 for (i = 0; i < nr_args; i++) {
3817 switch (res[i].opcode) {
3818 case IORING_RESTRICTION_REGISTER_OP:
3819 if (res[i].register_op >= IORING_REGISTER_LAST) {
3824 __set_bit(res[i].register_op,
3825 ctx->restrictions.register_op);
3827 case IORING_RESTRICTION_SQE_OP:
3828 if (res[i].sqe_op >= IORING_OP_LAST) {
3833 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
3835 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
3836 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
3838 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
3839 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
3848 /* Reset all restrictions if an error happened */
3850 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
3852 ctx->restrictions.registered = true;
3858 static int io_register_enable_rings(struct io_ring_ctx *ctx)
3860 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3863 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task)
3864 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
3866 if (ctx->restrictions.registered)
3867 ctx->restricted = 1;
3869 ctx->flags &= ~IORING_SETUP_R_DISABLED;
3870 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
3871 wake_up(&ctx->sq_data->wait);
3875 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
3876 void __user *arg, unsigned len)
3878 struct io_uring_task *tctx = current->io_uring;
3879 cpumask_var_t new_mask;
3882 if (!tctx || !tctx->io_wq)
3885 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3888 cpumask_clear(new_mask);
3889 if (len > cpumask_size())
3890 len = cpumask_size();
3892 if (in_compat_syscall()) {
3893 ret = compat_get_bitmap(cpumask_bits(new_mask),
3894 (const compat_ulong_t __user *)arg,
3895 len * 8 /* CHAR_BIT */);
3897 ret = copy_from_user(new_mask, arg, len);
3901 free_cpumask_var(new_mask);
3905 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
3906 free_cpumask_var(new_mask);
3910 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
3912 struct io_uring_task *tctx = current->io_uring;
3914 if (!tctx || !tctx->io_wq)
3917 return io_wq_cpu_affinity(tctx->io_wq, NULL);
3920 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
3922 __must_hold(&ctx->uring_lock)
3924 struct io_tctx_node *node;
3925 struct io_uring_task *tctx = NULL;
3926 struct io_sq_data *sqd = NULL;
3930 if (copy_from_user(new_count, arg, sizeof(new_count)))
3932 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3933 if (new_count[i] > INT_MAX)
3936 if (ctx->flags & IORING_SETUP_SQPOLL) {
3940 * Observe the correct sqd->lock -> ctx->uring_lock
3941 * ordering. Fine to drop uring_lock here, we hold
3944 refcount_inc(&sqd->refs);
3945 mutex_unlock(&ctx->uring_lock);
3946 mutex_lock(&sqd->lock);
3947 mutex_lock(&ctx->uring_lock);
3949 tctx = sqd->thread->io_uring;
3952 tctx = current->io_uring;
3955 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
3957 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3959 ctx->iowq_limits[i] = new_count[i];
3960 ctx->iowq_limits_set = true;
3962 if (tctx && tctx->io_wq) {
3963 ret = io_wq_max_workers(tctx->io_wq, new_count);
3967 memset(new_count, 0, sizeof(new_count));
3971 mutex_unlock(&sqd->lock);
3972 io_put_sq_data(sqd);
3975 if (copy_to_user(arg, new_count, sizeof(new_count)))
3978 /* that's it for SQPOLL, only the SQPOLL task creates requests */
3982 /* now propagate the restriction to all registered users */
3983 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3984 struct io_uring_task *tctx = node->task->io_uring;
3986 if (WARN_ON_ONCE(!tctx->io_wq))
3989 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3990 new_count[i] = ctx->iowq_limits[i];
3991 /* ignore errors, it always returns zero anyway */
3992 (void)io_wq_max_workers(tctx->io_wq, new_count);
3997 mutex_unlock(&sqd->lock);
3998 io_put_sq_data(sqd);
4003 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
4004 void __user *arg, unsigned nr_args)
4005 __releases(ctx->uring_lock)
4006 __acquires(ctx->uring_lock)
4011 * We don't quiesce the refs for register anymore and so it can't be
4012 * dying as we're holding a file ref here.
4014 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
4017 if (ctx->submitter_task && ctx->submitter_task != current)
4020 if (ctx->restricted) {
4021 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
4022 if (!test_bit(opcode, ctx->restrictions.register_op))
4027 case IORING_REGISTER_BUFFERS:
4031 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
4033 case IORING_UNREGISTER_BUFFERS:
4037 ret = io_sqe_buffers_unregister(ctx);
4039 case IORING_REGISTER_FILES:
4043 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
4045 case IORING_UNREGISTER_FILES:
4049 ret = io_sqe_files_unregister(ctx);
4051 case IORING_REGISTER_FILES_UPDATE:
4052 ret = io_register_files_update(ctx, arg, nr_args);
4054 case IORING_REGISTER_EVENTFD:
4058 ret = io_eventfd_register(ctx, arg, 0);
4060 case IORING_REGISTER_EVENTFD_ASYNC:
4064 ret = io_eventfd_register(ctx, arg, 1);
4066 case IORING_UNREGISTER_EVENTFD:
4070 ret = io_eventfd_unregister(ctx);
4072 case IORING_REGISTER_PROBE:
4074 if (!arg || nr_args > 256)
4076 ret = io_probe(ctx, arg, nr_args);
4078 case IORING_REGISTER_PERSONALITY:
4082 ret = io_register_personality(ctx);
4084 case IORING_UNREGISTER_PERSONALITY:
4088 ret = io_unregister_personality(ctx, nr_args);
4090 case IORING_REGISTER_ENABLE_RINGS:
4094 ret = io_register_enable_rings(ctx);
4096 case IORING_REGISTER_RESTRICTIONS:
4097 ret = io_register_restrictions(ctx, arg, nr_args);
4099 case IORING_REGISTER_FILES2:
4100 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
4102 case IORING_REGISTER_FILES_UPDATE2:
4103 ret = io_register_rsrc_update(ctx, arg, nr_args,
4106 case IORING_REGISTER_BUFFERS2:
4107 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
4109 case IORING_REGISTER_BUFFERS_UPDATE:
4110 ret = io_register_rsrc_update(ctx, arg, nr_args,
4111 IORING_RSRC_BUFFER);
4113 case IORING_REGISTER_IOWQ_AFF:
4115 if (!arg || !nr_args)
4117 ret = io_register_iowq_aff(ctx, arg, nr_args);
4119 case IORING_UNREGISTER_IOWQ_AFF:
4123 ret = io_unregister_iowq_aff(ctx);
4125 case IORING_REGISTER_IOWQ_MAX_WORKERS:
4127 if (!arg || nr_args != 2)
4129 ret = io_register_iowq_max_workers(ctx, arg);
4131 case IORING_REGISTER_RING_FDS:
4132 ret = io_ringfd_register(ctx, arg, nr_args);
4134 case IORING_UNREGISTER_RING_FDS:
4135 ret = io_ringfd_unregister(ctx, arg, nr_args);
4137 case IORING_REGISTER_PBUF_RING:
4139 if (!arg || nr_args != 1)
4141 ret = io_register_pbuf_ring(ctx, arg);
4143 case IORING_UNREGISTER_PBUF_RING:
4145 if (!arg || nr_args != 1)
4147 ret = io_unregister_pbuf_ring(ctx, arg);
4149 case IORING_REGISTER_SYNC_CANCEL:
4151 if (!arg || nr_args != 1)
4153 ret = io_sync_cancel(ctx, arg);
4155 case IORING_REGISTER_FILE_ALLOC_RANGE:
4157 if (!arg || nr_args)
4159 ret = io_register_file_alloc_range(ctx, arg);
4169 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
4170 void __user *, arg, unsigned int, nr_args)
4172 struct io_ring_ctx *ctx;
4176 if (opcode >= IORING_REGISTER_LAST)
4184 if (!io_is_uring_fops(f.file))
4187 ctx = f.file->private_data;
4189 mutex_lock(&ctx->uring_lock);
4190 ret = __io_uring_register(ctx, opcode, arg, nr_args);
4191 mutex_unlock(&ctx->uring_lock);
4192 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
4198 static int __init io_uring_init(void)
4200 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4201 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4202 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4205 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4206 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4207 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4208 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4209 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4210 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
4211 BUILD_BUG_SQE_ELEM(1, __u8, flags);
4212 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
4213 BUILD_BUG_SQE_ELEM(4, __s32, fd);
4214 BUILD_BUG_SQE_ELEM(8, __u64, off);
4215 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
4216 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
4217 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
4218 BUILD_BUG_SQE_ELEM(16, __u64, addr);
4219 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
4220 BUILD_BUG_SQE_ELEM(24, __u32, len);
4221 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
4222 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
4223 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4224 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
4225 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
4226 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
4227 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
4228 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
4229 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
4230 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
4231 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
4232 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
4233 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
4234 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
4235 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
4236 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
4237 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
4238 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
4239 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
4240 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
4241 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
4242 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
4243 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
4244 BUILD_BUG_SQE_ELEM(42, __u16, personality);
4245 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
4246 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
4247 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
4248 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
4249 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
4250 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4251 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
4253 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4254 sizeof(struct io_uring_rsrc_update));
4255 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4256 sizeof(struct io_uring_rsrc_update2));
4258 /* ->buf_index is u16 */
4259 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4260 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4261 offsetof(struct io_uring_buf_ring, tail));
4263 /* should fit into one byte */
4264 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
4265 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4266 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4268 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
4270 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4272 io_uring_optable_init();
4274 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4278 __initcall(io_uring_init);