1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <net/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
49 #include <linux/bits.h>
51 #include <linux/sched/signal.h>
53 #include <linux/file.h>
54 #include <linux/fdtable.h>
56 #include <linux/mman.h>
57 #include <linux/percpu.h>
58 #include <linux/slab.h>
59 #include <linux/bvec.h>
60 #include <linux/net.h>
62 #include <net/af_unix.h>
64 #include <linux/anon_inodes.h>
65 #include <linux/sched/mm.h>
66 #include <linux/uaccess.h>
67 #include <linux/nospec.h>
68 #include <linux/highmem.h>
69 #include <linux/fsnotify.h>
70 #include <linux/fadvise.h>
71 #include <linux/task_work.h>
72 #include <linux/io_uring.h>
73 #include <linux/audit.h>
74 #include <linux/security.h>
75 #include <asm/shmparam.h>
77 #define CREATE_TRACE_POINTS
78 #include <trace/events/io_uring.h>
80 #include <uapi/linux/io_uring.h>
98 #include "alloc_cache.h"
100 #define IORING_MAX_ENTRIES 32768
101 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
103 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
104 IORING_REGISTER_LAST + IORING_OP_LAST)
106 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
107 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
109 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
110 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
112 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
113 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
116 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
119 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
121 #define IO_COMPL_BATCH 32
122 #define IO_REQ_ALLOC_BATCH 8
125 IO_CHECK_CQ_OVERFLOW_BIT,
126 IO_CHECK_CQ_DROPPED_BIT,
130 IO_EVENTFD_OP_SIGNAL_BIT,
131 IO_EVENTFD_OP_FREE_BIT,
134 struct io_defer_entry {
135 struct list_head list;
136 struct io_kiocb *req;
140 /* requests with any of those set should undergo io_disarm_next() */
141 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
142 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
144 static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
145 struct task_struct *task,
148 static void io_dismantle_req(struct io_kiocb *req);
149 static void io_clean_op(struct io_kiocb *req);
150 static void io_queue_sqe(struct io_kiocb *req);
151 static void io_move_task_work_from_local(struct io_ring_ctx *ctx);
152 static void __io_submit_flush_completions(struct io_ring_ctx *ctx);
153 static __cold void io_fallback_tw(struct io_uring_task *tctx);
155 struct kmem_cache *req_cachep;
157 struct sock *io_uring_get_socket(struct file *file)
159 #if defined(CONFIG_UNIX)
160 if (io_is_uring_fops(file)) {
161 struct io_ring_ctx *ctx = file->private_data;
163 return ctx->ring_sock->sk;
168 EXPORT_SYMBOL(io_uring_get_socket);
170 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
172 if (!wq_list_empty(&ctx->submit_state.compl_reqs) ||
173 ctx->submit_state.cqes_count)
174 __io_submit_flush_completions(ctx);
177 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
179 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
182 static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx)
184 return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head);
187 static bool io_match_linked(struct io_kiocb *head)
189 struct io_kiocb *req;
191 io_for_each_link(req, head) {
192 if (req->flags & REQ_F_INFLIGHT)
199 * As io_match_task() but protected against racing with linked timeouts.
200 * User must not hold timeout_lock.
202 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
207 if (task && head->task != task)
212 if (head->flags & REQ_F_LINK_TIMEOUT) {
213 struct io_ring_ctx *ctx = head->ctx;
215 /* protect against races with linked timeouts */
216 spin_lock_irq(&ctx->timeout_lock);
217 matched = io_match_linked(head);
218 spin_unlock_irq(&ctx->timeout_lock);
220 matched = io_match_linked(head);
225 static inline void req_fail_link_node(struct io_kiocb *req, int res)
228 io_req_set_res(req, res, 0);
231 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
233 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
234 kasan_poison_object_data(req_cachep, req);
237 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
239 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
241 complete(&ctx->ref_comp);
244 static __cold void io_fallback_req_func(struct work_struct *work)
246 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
248 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
249 struct io_kiocb *req, *tmp;
250 struct io_tw_state ts = { .locked = true, };
252 mutex_lock(&ctx->uring_lock);
253 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
254 req->io_task_work.func(req, &ts);
255 if (WARN_ON_ONCE(!ts.locked))
257 io_submit_flush_completions(ctx);
258 mutex_unlock(&ctx->uring_lock);
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, sizeof(struct async_poll));
314 io_alloc_cache_init(&ctx->netmsg_cache, sizeof(struct io_async_msghdr));
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 init_waitqueue_head(&ctx->poll_wq);
320 spin_lock_init(&ctx->completion_lock);
321 spin_lock_init(&ctx->timeout_lock);
322 INIT_WQ_LIST(&ctx->iopoll_list);
323 INIT_LIST_HEAD(&ctx->io_buffers_pages);
324 INIT_LIST_HEAD(&ctx->io_buffers_comp);
325 INIT_LIST_HEAD(&ctx->defer_list);
326 INIT_LIST_HEAD(&ctx->timeout_list);
327 INIT_LIST_HEAD(&ctx->ltimeout_list);
328 spin_lock_init(&ctx->rsrc_ref_lock);
329 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
330 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
331 init_task_work(&ctx->rsrc_put_tw, io_rsrc_put_tw);
332 init_llist_head(&ctx->rsrc_put_llist);
333 init_llist_head(&ctx->work_llist);
334 INIT_LIST_HEAD(&ctx->tctx_list);
335 ctx->submit_state.free_list.next = NULL;
336 INIT_WQ_LIST(&ctx->locked_free_list);
337 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
338 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
341 kfree(ctx->dummy_ubuf);
342 kfree(ctx->cancel_table.hbs);
343 kfree(ctx->cancel_table_locked.hbs);
345 xa_destroy(&ctx->io_bl_xa);
350 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
352 struct io_rings *r = ctx->rings;
354 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
358 static bool req_need_defer(struct io_kiocb *req, u32 seq)
360 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
361 struct io_ring_ctx *ctx = req->ctx;
363 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
369 static inline void io_req_track_inflight(struct io_kiocb *req)
371 if (!(req->flags & REQ_F_INFLIGHT)) {
372 req->flags |= REQ_F_INFLIGHT;
373 atomic_inc(&req->task->io_uring->inflight_tracked);
377 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
379 if (WARN_ON_ONCE(!req->link))
382 req->flags &= ~REQ_F_ARM_LTIMEOUT;
383 req->flags |= REQ_F_LINK_TIMEOUT;
385 /* linked timeouts should have two refs once prep'ed */
386 io_req_set_refcount(req);
387 __io_req_set_refcount(req->link, 2);
391 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
393 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
395 return __io_prep_linked_timeout(req);
398 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
400 io_queue_linked_timeout(__io_prep_linked_timeout(req));
403 static inline void io_arm_ltimeout(struct io_kiocb *req)
405 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
406 __io_arm_ltimeout(req);
409 static void io_prep_async_work(struct io_kiocb *req)
411 const struct io_issue_def *def = &io_issue_defs[req->opcode];
412 struct io_ring_ctx *ctx = req->ctx;
414 if (!(req->flags & REQ_F_CREDS)) {
415 req->flags |= REQ_F_CREDS;
416 req->creds = get_current_cred();
419 req->work.list.next = NULL;
421 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
422 if (req->flags & REQ_F_FORCE_ASYNC)
423 req->work.flags |= IO_WQ_WORK_CONCURRENT;
425 if (req->file && !io_req_ffs_set(req))
426 req->flags |= io_file_get_flags(req->file) << REQ_F_SUPPORT_NOWAIT_BIT;
428 if (req->flags & REQ_F_ISREG) {
429 bool should_hash = def->hash_reg_file;
431 /* don't serialize this request if the fs doesn't need it */
432 if (should_hash && (req->file->f_flags & O_DIRECT) &&
433 (req->file->f_mode & FMODE_DIO_PARALLEL_WRITE))
435 if (should_hash || (ctx->flags & IORING_SETUP_IOPOLL))
436 io_wq_hash_work(&req->work, file_inode(req->file));
437 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
438 if (def->unbound_nonreg_file)
439 req->work.flags |= IO_WQ_WORK_UNBOUND;
443 static void io_prep_async_link(struct io_kiocb *req)
445 struct io_kiocb *cur;
447 if (req->flags & REQ_F_LINK_TIMEOUT) {
448 struct io_ring_ctx *ctx = req->ctx;
450 spin_lock_irq(&ctx->timeout_lock);
451 io_for_each_link(cur, req)
452 io_prep_async_work(cur);
453 spin_unlock_irq(&ctx->timeout_lock);
455 io_for_each_link(cur, req)
456 io_prep_async_work(cur);
460 void io_queue_iowq(struct io_kiocb *req, struct io_tw_state *ts_dont_use)
462 struct io_kiocb *link = io_prep_linked_timeout(req);
463 struct io_uring_task *tctx = req->task->io_uring;
466 BUG_ON(!tctx->io_wq);
468 /* init ->work of the whole link before punting */
469 io_prep_async_link(req);
472 * Not expected to happen, but if we do have a bug where this _can_
473 * happen, catch it here and ensure the request is marked as
474 * canceled. That will make io-wq go through the usual work cancel
475 * procedure rather than attempt to run this request (or create a new
478 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
479 req->work.flags |= IO_WQ_WORK_CANCEL;
481 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
482 io_wq_enqueue(tctx->io_wq, &req->work);
484 io_queue_linked_timeout(link);
487 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
489 while (!list_empty(&ctx->defer_list)) {
490 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
491 struct io_defer_entry, list);
493 if (req_need_defer(de->req, de->seq))
495 list_del_init(&de->list);
496 io_req_task_queue(de->req);
502 static void io_eventfd_ops(struct rcu_head *rcu)
504 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
505 int ops = atomic_xchg(&ev_fd->ops, 0);
507 if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT))
508 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
510 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
511 * ordering in a race but if references are 0 we know we have to free
514 if (atomic_dec_and_test(&ev_fd->refs)) {
515 eventfd_ctx_put(ev_fd->cq_ev_fd);
520 static void io_eventfd_signal(struct io_ring_ctx *ctx)
522 struct io_ev_fd *ev_fd = NULL;
526 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
529 ev_fd = rcu_dereference(ctx->io_ev_fd);
532 * Check again if ev_fd exists incase an io_eventfd_unregister call
533 * completed between the NULL check of ctx->io_ev_fd at the start of
534 * the function and rcu_read_lock.
536 if (unlikely(!ev_fd))
538 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
540 if (ev_fd->eventfd_async && !io_wq_current_is_worker())
543 if (likely(eventfd_signal_allowed())) {
544 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
546 atomic_inc(&ev_fd->refs);
547 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops))
548 call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops);
550 atomic_dec(&ev_fd->refs);
557 static void io_eventfd_flush_signal(struct io_ring_ctx *ctx)
561 spin_lock(&ctx->completion_lock);
564 * Eventfd should only get triggered when at least one event has been
565 * posted. Some applications rely on the eventfd notification count
566 * only changing IFF a new CQE has been added to the CQ ring. There's
567 * no depedency on 1:1 relationship between how many times this
568 * function is called (and hence the eventfd count) and number of CQEs
569 * posted to the CQ ring.
571 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
572 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
573 spin_unlock(&ctx->completion_lock);
577 io_eventfd_signal(ctx);
580 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
582 if (ctx->poll_activated)
583 io_poll_wq_wake(ctx);
584 if (ctx->off_timeout_used)
585 io_flush_timeouts(ctx);
586 if (ctx->drain_active) {
587 spin_lock(&ctx->completion_lock);
588 io_queue_deferred(ctx);
589 spin_unlock(&ctx->completion_lock);
592 io_eventfd_flush_signal(ctx);
595 static inline void __io_cq_lock(struct io_ring_ctx *ctx)
596 __acquires(ctx->completion_lock)
598 if (!ctx->task_complete)
599 spin_lock(&ctx->completion_lock);
602 static inline void __io_cq_unlock(struct io_ring_ctx *ctx)
604 if (!ctx->task_complete)
605 spin_unlock(&ctx->completion_lock);
608 static inline void io_cq_lock(struct io_ring_ctx *ctx)
609 __acquires(ctx->completion_lock)
611 spin_lock(&ctx->completion_lock);
614 static inline void io_cq_unlock(struct io_ring_ctx *ctx)
615 __releases(ctx->completion_lock)
617 spin_unlock(&ctx->completion_lock);
620 /* keep it inlined for io_submit_flush_completions() */
621 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
622 __releases(ctx->completion_lock)
624 io_commit_cqring(ctx);
626 io_commit_cqring_flush(ctx);
630 static inline void __io_cq_unlock_post_flush(struct io_ring_ctx *ctx)
631 __releases(ctx->completion_lock)
633 io_commit_cqring(ctx);
635 io_commit_cqring_flush(ctx);
638 * As ->task_complete implies that the ring is single tasked, cq_wait
639 * may only be waited on by the current in io_cqring_wait(), but since
640 * it will re-check the wakeup conditions once we return we can safely
643 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) {
645 __io_cqring_wake(ctx);
649 void io_cq_unlock_post(struct io_ring_ctx *ctx)
650 __releases(ctx->completion_lock)
652 io_commit_cqring(ctx);
653 spin_unlock(&ctx->completion_lock);
654 io_commit_cqring_flush(ctx);
658 /* Returns true if there are no backlogged entries after the flush */
659 static void io_cqring_overflow_kill(struct io_ring_ctx *ctx)
661 struct io_overflow_cqe *ocqe;
665 list_splice_init(&ctx->cq_overflow_list, &list);
666 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
669 while (!list_empty(&list)) {
670 ocqe = list_first_entry(&list, struct io_overflow_cqe, list);
671 list_del(&ocqe->list);
676 static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx)
678 size_t cqe_size = sizeof(struct io_uring_cqe);
680 if (__io_cqring_events(ctx) == ctx->cq_entries)
683 if (ctx->flags & IORING_SETUP_CQE32)
687 while (!list_empty(&ctx->cq_overflow_list)) {
688 struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true);
689 struct io_overflow_cqe *ocqe;
693 ocqe = list_first_entry(&ctx->cq_overflow_list,
694 struct io_overflow_cqe, list);
695 memcpy(cqe, &ocqe->cqe, cqe_size);
696 list_del(&ocqe->list);
700 if (list_empty(&ctx->cq_overflow_list)) {
701 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
702 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
704 io_cq_unlock_post(ctx);
707 static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx)
709 /* iopoll syncs against uring_lock, not completion_lock */
710 if (ctx->flags & IORING_SETUP_IOPOLL)
711 mutex_lock(&ctx->uring_lock);
712 __io_cqring_overflow_flush(ctx);
713 if (ctx->flags & IORING_SETUP_IOPOLL)
714 mutex_unlock(&ctx->uring_lock);
717 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx)
719 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
720 io_cqring_do_overflow_flush(ctx);
723 /* can be called by any task */
724 static void io_put_task_remote(struct task_struct *task, int nr)
726 struct io_uring_task *tctx = task->io_uring;
728 percpu_counter_sub(&tctx->inflight, nr);
729 if (unlikely(atomic_read(&tctx->in_cancel)))
730 wake_up(&tctx->wait);
731 put_task_struct_many(task, nr);
734 /* used by a task to put its own references */
735 static void io_put_task_local(struct task_struct *task, int nr)
737 task->io_uring->cached_refs += nr;
740 /* must to be called somewhat shortly after putting a request */
741 static inline void io_put_task(struct task_struct *task, int nr)
743 if (likely(task == current))
744 io_put_task_local(task, nr);
746 io_put_task_remote(task, nr);
749 void io_task_refs_refill(struct io_uring_task *tctx)
751 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
753 percpu_counter_add(&tctx->inflight, refill);
754 refcount_add(refill, ¤t->usage);
755 tctx->cached_refs += refill;
758 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
760 struct io_uring_task *tctx = task->io_uring;
761 unsigned int refs = tctx->cached_refs;
764 tctx->cached_refs = 0;
765 percpu_counter_sub(&tctx->inflight, refs);
766 put_task_struct_many(task, refs);
770 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
771 s32 res, u32 cflags, u64 extra1, u64 extra2)
773 struct io_overflow_cqe *ocqe;
774 size_t ocq_size = sizeof(struct io_overflow_cqe);
775 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
777 lockdep_assert_held(&ctx->completion_lock);
780 ocq_size += sizeof(struct io_uring_cqe);
782 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
783 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
786 * If we're in ring overflow flush mode, or in task cancel mode,
787 * or cannot allocate an overflow entry, then we need to drop it
790 io_account_cq_overflow(ctx);
791 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
794 if (list_empty(&ctx->cq_overflow_list)) {
795 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
796 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
799 ocqe->cqe.user_data = user_data;
801 ocqe->cqe.flags = cflags;
803 ocqe->cqe.big_cqe[0] = extra1;
804 ocqe->cqe.big_cqe[1] = extra2;
806 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
810 bool io_req_cqe_overflow(struct io_kiocb *req)
812 if (!(req->flags & REQ_F_CQE32_INIT)) {
816 return io_cqring_event_overflow(req->ctx, req->cqe.user_data,
817 req->cqe.res, req->cqe.flags,
818 req->extra1, req->extra2);
822 * writes to the cq entry need to come after reading head; the
823 * control dependency is enough as we're using WRITE_ONCE to
826 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow)
828 struct io_rings *rings = ctx->rings;
829 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
830 unsigned int free, queued, len;
833 * Posting into the CQ when there are pending overflowed CQEs may break
834 * ordering guarantees, which will affect links, F_MORE users and more.
835 * Force overflow the completion.
837 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
840 /* userspace may cheat modifying the tail, be safe and do min */
841 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
842 free = ctx->cq_entries - queued;
843 /* we need a contiguous range, limit based on the current array offset */
844 len = min(free, ctx->cq_entries - off);
848 if (ctx->flags & IORING_SETUP_CQE32) {
853 ctx->cqe_cached = &rings->cqes[off];
854 ctx->cqe_sentinel = ctx->cqe_cached + len;
856 ctx->cached_cq_tail++;
858 if (ctx->flags & IORING_SETUP_CQE32)
860 return &rings->cqes[off];
863 static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
866 struct io_uring_cqe *cqe;
871 * If we can't get a cq entry, userspace overflowed the
872 * submission (by quite a lot). Increment the overflow count in
875 cqe = io_get_cqe(ctx);
877 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
879 WRITE_ONCE(cqe->user_data, user_data);
880 WRITE_ONCE(cqe->res, res);
881 WRITE_ONCE(cqe->flags, cflags);
883 if (ctx->flags & IORING_SETUP_CQE32) {
884 WRITE_ONCE(cqe->big_cqe[0], 0);
885 WRITE_ONCE(cqe->big_cqe[1], 0);
892 static void __io_flush_post_cqes(struct io_ring_ctx *ctx)
893 __must_hold(&ctx->uring_lock)
895 struct io_submit_state *state = &ctx->submit_state;
898 lockdep_assert_held(&ctx->uring_lock);
899 for (i = 0; i < state->cqes_count; i++) {
900 struct io_uring_cqe *cqe = &state->cqes[i];
902 if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) {
903 if (ctx->task_complete) {
904 spin_lock(&ctx->completion_lock);
905 io_cqring_event_overflow(ctx, cqe->user_data,
906 cqe->res, cqe->flags, 0, 0);
907 spin_unlock(&ctx->completion_lock);
909 io_cqring_event_overflow(ctx, cqe->user_data,
910 cqe->res, cqe->flags, 0, 0);
914 state->cqes_count = 0;
917 static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
923 filled = io_fill_cqe_aux(ctx, user_data, res, cflags);
924 if (!filled && allow_overflow)
925 filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
927 io_cq_unlock_post(ctx);
931 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags)
933 return __io_post_aux_cqe(ctx, user_data, res, cflags, true);
936 bool io_aux_cqe(struct io_ring_ctx *ctx, bool defer, u64 user_data, s32 res, u32 cflags,
939 struct io_uring_cqe *cqe;
943 return __io_post_aux_cqe(ctx, user_data, res, cflags, allow_overflow);
945 length = ARRAY_SIZE(ctx->submit_state.cqes);
947 lockdep_assert_held(&ctx->uring_lock);
949 if (ctx->submit_state.cqes_count == length) {
951 __io_flush_post_cqes(ctx);
952 /* no need to flush - flush is deferred */
953 __io_cq_unlock_post(ctx);
956 /* For defered completions this is not as strict as it is otherwise,
957 * however it's main job is to prevent unbounded posted completions,
958 * and in that it works just as well.
960 if (!allow_overflow && test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
963 cqe = &ctx->submit_state.cqes[ctx->submit_state.cqes_count++];
964 cqe->user_data = user_data;
970 static void __io_req_complete_post(struct io_kiocb *req)
972 struct io_ring_ctx *ctx = req->ctx;
973 struct io_rsrc_node *rsrc_node = NULL;
976 if (!(req->flags & REQ_F_CQE_SKIP))
977 io_fill_cqe_req(ctx, req);
980 * If we're the last reference to this request, add to our locked
983 if (req_ref_put_and_test(req)) {
984 if (req->flags & IO_REQ_LINK_FLAGS) {
985 if (req->flags & IO_DISARM_MASK)
988 io_req_task_queue(req->link);
992 io_put_kbuf_comp(req);
993 io_dismantle_req(req);
994 rsrc_node = req->rsrc_node;
996 * Selected buffer deallocation in io_clean_op() assumes that
997 * we don't hold ->completion_lock. Clean them here to avoid
1000 io_put_task_remote(req->task, 1);
1001 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
1002 ctx->locked_free_nr++;
1004 io_cq_unlock_post(ctx);
1006 io_put_rsrc_node(rsrc_node);
1009 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
1011 if (req->ctx->task_complete && (issue_flags & IO_URING_F_IOWQ)) {
1012 req->io_task_work.func = io_req_task_complete;
1013 io_req_task_work_add(req);
1014 } else if (!(issue_flags & IO_URING_F_UNLOCKED) ||
1015 !(req->ctx->flags & IORING_SETUP_IOPOLL)) {
1016 __io_req_complete_post(req);
1018 struct io_ring_ctx *ctx = req->ctx;
1020 mutex_lock(&ctx->uring_lock);
1021 __io_req_complete_post(req);
1022 mutex_unlock(&ctx->uring_lock);
1026 void io_req_defer_failed(struct io_kiocb *req, s32 res)
1027 __must_hold(&ctx->uring_lock)
1029 const struct io_cold_def *def = &io_cold_defs[req->opcode];
1031 lockdep_assert_held(&req->ctx->uring_lock);
1034 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
1037 io_req_complete_defer(req);
1041 * Don't initialise the fields below on every allocation, but do that in
1042 * advance and keep them valid across allocations.
1044 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1048 req->async_data = NULL;
1049 /* not necessary, but safer to zero */
1053 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1054 struct io_submit_state *state)
1056 spin_lock(&ctx->completion_lock);
1057 wq_list_splice(&ctx->locked_free_list, &state->free_list);
1058 ctx->locked_free_nr = 0;
1059 spin_unlock(&ctx->completion_lock);
1063 * A request might get retired back into the request caches even before opcode
1064 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1065 * Because of that, io_alloc_req() should be called only under ->uring_lock
1066 * and with extra caution to not get a request that is still worked on.
1068 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
1069 __must_hold(&ctx->uring_lock)
1071 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1072 void *reqs[IO_REQ_ALLOC_BATCH];
1076 * If we have more than a batch's worth of requests in our IRQ side
1077 * locked cache, grab the lock and move them over to our submission
1080 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
1081 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
1082 if (!io_req_cache_empty(ctx))
1086 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
1089 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1090 * retry single alloc to be on the safe side.
1092 if (unlikely(ret <= 0)) {
1093 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1099 percpu_ref_get_many(&ctx->refs, ret);
1100 for (i = 0; i < ret; i++) {
1101 struct io_kiocb *req = reqs[i];
1103 io_preinit_req(req, ctx);
1104 io_req_add_to_cache(req, ctx);
1109 static inline void io_dismantle_req(struct io_kiocb *req)
1111 unsigned int flags = req->flags;
1113 if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
1115 if (!(flags & REQ_F_FIXED_FILE))
1116 io_put_file(req->file);
1119 static __cold void io_free_req_tw(struct io_kiocb *req, struct io_tw_state *ts)
1121 struct io_ring_ctx *ctx = req->ctx;
1123 io_put_rsrc_node(req->rsrc_node);
1124 io_dismantle_req(req);
1125 io_put_task_remote(req->task, 1);
1127 spin_lock(&ctx->completion_lock);
1128 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
1129 ctx->locked_free_nr++;
1130 spin_unlock(&ctx->completion_lock);
1133 __cold void io_free_req(struct io_kiocb *req)
1135 req->io_task_work.func = io_free_req_tw;
1136 io_req_task_work_add(req);
1139 static void __io_req_find_next_prep(struct io_kiocb *req)
1141 struct io_ring_ctx *ctx = req->ctx;
1143 spin_lock(&ctx->completion_lock);
1144 io_disarm_next(req);
1145 spin_unlock(&ctx->completion_lock);
1148 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1150 struct io_kiocb *nxt;
1153 * If LINK is set, we have dependent requests in this chain. If we
1154 * didn't fail this request, queue the first one up, moving any other
1155 * dependencies to the next request. In case of failure, fail the rest
1158 if (unlikely(req->flags & IO_DISARM_MASK))
1159 __io_req_find_next_prep(req);
1165 static void ctx_flush_and_put(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1169 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1170 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1172 io_submit_flush_completions(ctx);
1173 mutex_unlock(&ctx->uring_lock);
1176 percpu_ref_put(&ctx->refs);
1179 static unsigned int handle_tw_list(struct llist_node *node,
1180 struct io_ring_ctx **ctx,
1181 struct io_tw_state *ts,
1182 struct llist_node *last)
1184 unsigned int count = 0;
1186 while (node && node != last) {
1187 struct llist_node *next = node->next;
1188 struct io_kiocb *req = container_of(node, struct io_kiocb,
1191 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1193 if (req->ctx != *ctx) {
1194 ctx_flush_and_put(*ctx, ts);
1196 /* if not contended, grab and improve batching */
1197 ts->locked = mutex_trylock(&(*ctx)->uring_lock);
1198 percpu_ref_get(&(*ctx)->refs);
1200 req->io_task_work.func(req, ts);
1203 if (unlikely(need_resched())) {
1204 ctx_flush_and_put(*ctx, ts);
1214 * io_llist_xchg - swap all entries in a lock-less list
1215 * @head: the head of lock-less list to delete all entries
1216 * @new: new entry as the head of the list
1218 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1219 * The order of entries returned is from the newest to the oldest added one.
1221 static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1222 struct llist_node *new)
1224 return xchg(&head->first, new);
1228 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1229 * @head: the head of lock-less list to delete all entries
1230 * @old: expected old value of the first entry of the list
1231 * @new: new entry as the head of the list
1233 * perform a cmpxchg on the first entry of the list.
1236 static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1237 struct llist_node *old,
1238 struct llist_node *new)
1240 return cmpxchg(&head->first, old, new);
1243 void tctx_task_work(struct callback_head *cb)
1245 struct io_tw_state ts = {};
1246 struct io_ring_ctx *ctx = NULL;
1247 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1249 struct llist_node fake = {};
1250 struct llist_node *node;
1251 unsigned int loops = 0;
1252 unsigned int count = 0;
1254 if (unlikely(current->flags & PF_EXITING)) {
1255 io_fallback_tw(tctx);
1261 node = io_llist_xchg(&tctx->task_list, &fake);
1262 count += handle_tw_list(node, &ctx, &ts, &fake);
1264 /* skip expensive cmpxchg if there are items in the list */
1265 if (READ_ONCE(tctx->task_list.first) != &fake)
1267 if (ts.locked && !wq_list_empty(&ctx->submit_state.compl_reqs)) {
1268 io_submit_flush_completions(ctx);
1269 if (READ_ONCE(tctx->task_list.first) != &fake)
1272 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1273 } while (node != &fake);
1275 ctx_flush_and_put(ctx, &ts);
1277 /* relaxed read is enough as only the task itself sets ->in_cancel */
1278 if (unlikely(atomic_read(&tctx->in_cancel)))
1279 io_uring_drop_tctx_refs(current);
1281 trace_io_uring_task_work_run(tctx, count, loops);
1284 static __cold void io_fallback_tw(struct io_uring_task *tctx)
1286 struct llist_node *node = llist_del_all(&tctx->task_list);
1287 struct io_kiocb *req;
1290 req = container_of(node, struct io_kiocb, io_task_work.node);
1292 if (llist_add(&req->io_task_work.node,
1293 &req->ctx->fallback_llist))
1294 schedule_delayed_work(&req->ctx->fallback_work, 1);
1298 static void io_req_local_work_add(struct io_kiocb *req)
1300 struct io_ring_ctx *ctx = req->ctx;
1302 percpu_ref_get(&ctx->refs);
1304 if (!llist_add(&req->io_task_work.node, &ctx->work_llist))
1307 /* needed for the following wake up */
1308 smp_mb__after_atomic();
1310 if (unlikely(atomic_read(&req->task->io_uring->in_cancel))) {
1311 io_move_task_work_from_local(ctx);
1315 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1316 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1318 io_eventfd_signal(ctx);
1320 if (READ_ONCE(ctx->cq_waiting))
1321 wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE);
1324 percpu_ref_put(&ctx->refs);
1327 void __io_req_task_work_add(struct io_kiocb *req, bool allow_local)
1329 struct io_uring_task *tctx = req->task->io_uring;
1330 struct io_ring_ctx *ctx = req->ctx;
1332 if (allow_local && ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
1333 io_req_local_work_add(req);
1337 /* task_work already pending, we're done */
1338 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1341 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1342 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1344 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1347 io_fallback_tw(tctx);
1350 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1352 struct llist_node *node;
1354 node = llist_del_all(&ctx->work_llist);
1356 struct io_kiocb *req = container_of(node, struct io_kiocb,
1360 __io_req_task_work_add(req, false);
1364 static int __io_run_local_work(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1366 struct llist_node *node;
1367 unsigned int loops = 0;
1370 if (WARN_ON_ONCE(ctx->submitter_task != current))
1372 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1373 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1375 node = io_llist_xchg(&ctx->work_llist, NULL);
1377 struct llist_node *next = node->next;
1378 struct io_kiocb *req = container_of(node, struct io_kiocb,
1380 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1381 req->io_task_work.func(req, ts);
1387 if (!llist_empty(&ctx->work_llist))
1390 io_submit_flush_completions(ctx);
1391 if (!llist_empty(&ctx->work_llist))
1394 trace_io_uring_local_work_run(ctx, ret, loops);
1398 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx)
1400 struct io_tw_state ts = { .locked = true, };
1403 if (llist_empty(&ctx->work_llist))
1406 ret = __io_run_local_work(ctx, &ts);
1407 /* shouldn't happen! */
1408 if (WARN_ON_ONCE(!ts.locked))
1409 mutex_lock(&ctx->uring_lock);
1413 static int io_run_local_work(struct io_ring_ctx *ctx)
1415 struct io_tw_state ts = {};
1418 ts.locked = mutex_trylock(&ctx->uring_lock);
1419 ret = __io_run_local_work(ctx, &ts);
1421 mutex_unlock(&ctx->uring_lock);
1426 static void io_req_task_cancel(struct io_kiocb *req, struct io_tw_state *ts)
1428 io_tw_lock(req->ctx, ts);
1429 io_req_defer_failed(req, req->cqe.res);
1432 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts)
1434 io_tw_lock(req->ctx, ts);
1435 /* req->task == current here, checking PF_EXITING is safe */
1436 if (unlikely(req->task->flags & PF_EXITING))
1437 io_req_defer_failed(req, -EFAULT);
1438 else if (req->flags & REQ_F_FORCE_ASYNC)
1439 io_queue_iowq(req, ts);
1444 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1446 io_req_set_res(req, ret, 0);
1447 req->io_task_work.func = io_req_task_cancel;
1448 io_req_task_work_add(req);
1451 void io_req_task_queue(struct io_kiocb *req)
1453 req->io_task_work.func = io_req_task_submit;
1454 io_req_task_work_add(req);
1457 void io_queue_next(struct io_kiocb *req)
1459 struct io_kiocb *nxt = io_req_find_next(req);
1462 io_req_task_queue(nxt);
1465 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1466 __must_hold(&ctx->uring_lock)
1468 struct task_struct *task = NULL;
1472 struct io_kiocb *req = container_of(node, struct io_kiocb,
1475 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1476 if (req->flags & REQ_F_REFCOUNT) {
1477 node = req->comp_list.next;
1478 if (!req_ref_put_and_test(req))
1481 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1482 struct async_poll *apoll = req->apoll;
1484 if (apoll->double_poll)
1485 kfree(apoll->double_poll);
1486 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1488 req->flags &= ~REQ_F_POLLED;
1490 if (req->flags & IO_REQ_LINK_FLAGS)
1492 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1495 if (!(req->flags & REQ_F_FIXED_FILE))
1496 io_put_file(req->file);
1498 io_req_put_rsrc_locked(req, ctx);
1500 if (req->task != task) {
1502 io_put_task(task, task_refs);
1507 node = req->comp_list.next;
1508 io_req_add_to_cache(req, ctx);
1512 io_put_task(task, task_refs);
1515 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1516 __must_hold(&ctx->uring_lock)
1518 struct io_submit_state *state = &ctx->submit_state;
1519 struct io_wq_work_node *node;
1522 /* must come first to preserve CQE ordering in failure cases */
1523 if (state->cqes_count)
1524 __io_flush_post_cqes(ctx);
1525 __wq_list_for_each(node, &state->compl_reqs) {
1526 struct io_kiocb *req = container_of(node, struct io_kiocb,
1529 if (!(req->flags & REQ_F_CQE_SKIP) &&
1530 unlikely(!__io_fill_cqe_req(ctx, req))) {
1531 if (ctx->task_complete) {
1532 spin_lock(&ctx->completion_lock);
1533 io_req_cqe_overflow(req);
1534 spin_unlock(&ctx->completion_lock);
1536 io_req_cqe_overflow(req);
1540 __io_cq_unlock_post_flush(ctx);
1542 if (!wq_list_empty(&ctx->submit_state.compl_reqs)) {
1543 io_free_batch_list(ctx, state->compl_reqs.first);
1544 INIT_WQ_LIST(&state->compl_reqs);
1549 * Drop reference to request, return next in chain (if there is one) if this
1550 * was the last reference to this request.
1552 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1554 struct io_kiocb *nxt = NULL;
1556 if (req_ref_put_and_test(req)) {
1557 if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
1558 nxt = io_req_find_next(req);
1564 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1566 /* See comment at the top of this file */
1568 return __io_cqring_events(ctx);
1572 * We can't just wait for polled events to come to us, we have to actively
1573 * find and complete them.
1575 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1577 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1580 mutex_lock(&ctx->uring_lock);
1581 while (!wq_list_empty(&ctx->iopoll_list)) {
1582 /* let it sleep and repeat later if can't complete a request */
1583 if (io_do_iopoll(ctx, true) == 0)
1586 * Ensure we allow local-to-the-cpu processing to take place,
1587 * in this case we need to ensure that we reap all events.
1588 * Also let task_work, etc. to progress by releasing the mutex
1590 if (need_resched()) {
1591 mutex_unlock(&ctx->uring_lock);
1593 mutex_lock(&ctx->uring_lock);
1596 mutex_unlock(&ctx->uring_lock);
1599 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1601 unsigned int nr_events = 0;
1603 unsigned long check_cq;
1605 if (!io_allowed_run_tw(ctx))
1608 check_cq = READ_ONCE(ctx->check_cq);
1609 if (unlikely(check_cq)) {
1610 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1611 __io_cqring_overflow_flush(ctx);
1613 * Similarly do not spin if we have not informed the user of any
1616 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1620 * Don't enter poll loop if we already have events pending.
1621 * If we do, we can potentially be spinning for commands that
1622 * already triggered a CQE (eg in error).
1624 if (io_cqring_events(ctx))
1629 * If a submit got punted to a workqueue, we can have the
1630 * application entering polling for a command before it gets
1631 * issued. That app will hold the uring_lock for the duration
1632 * of the poll right here, so we need to take a breather every
1633 * now and then to ensure that the issue has a chance to add
1634 * the poll to the issued list. Otherwise we can spin here
1635 * forever, while the workqueue is stuck trying to acquire the
1638 if (wq_list_empty(&ctx->iopoll_list) ||
1639 io_task_work_pending(ctx)) {
1640 u32 tail = ctx->cached_cq_tail;
1642 (void) io_run_local_work_locked(ctx);
1644 if (task_work_pending(current) ||
1645 wq_list_empty(&ctx->iopoll_list)) {
1646 mutex_unlock(&ctx->uring_lock);
1648 mutex_lock(&ctx->uring_lock);
1650 /* some requests don't go through iopoll_list */
1651 if (tail != ctx->cached_cq_tail ||
1652 wq_list_empty(&ctx->iopoll_list))
1655 ret = io_do_iopoll(ctx, !min);
1660 } while (nr_events < min && !need_resched());
1665 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts)
1668 io_req_complete_defer(req);
1670 io_req_complete_post(req, IO_URING_F_UNLOCKED);
1674 * After the iocb has been issued, it's safe to be found on the poll list.
1675 * Adding the kiocb to the list AFTER submission ensures that we don't
1676 * find it from a io_do_iopoll() thread before the issuer is done
1677 * accessing the kiocb cookie.
1679 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1681 struct io_ring_ctx *ctx = req->ctx;
1682 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1684 /* workqueue context doesn't hold uring_lock, grab it now */
1685 if (unlikely(needs_lock))
1686 mutex_lock(&ctx->uring_lock);
1689 * Track whether we have multiple files in our lists. This will impact
1690 * how we do polling eventually, not spinning if we're on potentially
1691 * different devices.
1693 if (wq_list_empty(&ctx->iopoll_list)) {
1694 ctx->poll_multi_queue = false;
1695 } else if (!ctx->poll_multi_queue) {
1696 struct io_kiocb *list_req;
1698 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1700 if (list_req->file != req->file)
1701 ctx->poll_multi_queue = true;
1705 * For fast devices, IO may have already completed. If it has, add
1706 * it to the front so we find it first.
1708 if (READ_ONCE(req->iopoll_completed))
1709 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1711 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1713 if (unlikely(needs_lock)) {
1715 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1716 * in sq thread task context or in io worker task context. If
1717 * current task context is sq thread, we don't need to check
1718 * whether should wake up sq thread.
1720 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1721 wq_has_sleeper(&ctx->sq_data->wait))
1722 wake_up(&ctx->sq_data->wait);
1724 mutex_unlock(&ctx->uring_lock);
1728 static bool io_bdev_nowait(struct block_device *bdev)
1730 return !bdev || bdev_nowait(bdev);
1734 * If we tracked the file through the SCM inflight mechanism, we could support
1735 * any file. For now, just ensure that anything potentially problematic is done
1738 static bool __io_file_supports_nowait(struct file *file, umode_t mode)
1740 if (S_ISBLK(mode)) {
1741 if (IS_ENABLED(CONFIG_BLOCK) &&
1742 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
1748 if (S_ISREG(mode)) {
1749 if (IS_ENABLED(CONFIG_BLOCK) &&
1750 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
1751 !io_is_uring_fops(file))
1756 /* any ->read/write should understand O_NONBLOCK */
1757 if (file->f_flags & O_NONBLOCK)
1759 return file->f_mode & FMODE_NOWAIT;
1763 * If we tracked the file through the SCM inflight mechanism, we could support
1764 * any file. For now, just ensure that anything potentially problematic is done
1767 unsigned int io_file_get_flags(struct file *file)
1769 umode_t mode = file_inode(file)->i_mode;
1770 unsigned int res = 0;
1774 if (__io_file_supports_nowait(file, mode))
1779 bool io_alloc_async_data(struct io_kiocb *req)
1781 WARN_ON_ONCE(!io_cold_defs[req->opcode].async_size);
1782 req->async_data = kmalloc(io_cold_defs[req->opcode].async_size, GFP_KERNEL);
1783 if (req->async_data) {
1784 req->flags |= REQ_F_ASYNC_DATA;
1790 int io_req_prep_async(struct io_kiocb *req)
1792 const struct io_cold_def *cdef = &io_cold_defs[req->opcode];
1793 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1795 /* assign early for deferred execution for non-fixed file */
1796 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE) && !req->file)
1797 req->file = io_file_get_normal(req, req->cqe.fd);
1798 if (!cdef->prep_async)
1800 if (WARN_ON_ONCE(req_has_async_data(req)))
1802 if (!def->manual_alloc) {
1803 if (io_alloc_async_data(req))
1806 return cdef->prep_async(req);
1809 static u32 io_get_sequence(struct io_kiocb *req)
1811 u32 seq = req->ctx->cached_sq_head;
1812 struct io_kiocb *cur;
1814 /* need original cached_sq_head, but it was increased for each req */
1815 io_for_each_link(cur, req)
1820 static __cold void io_drain_req(struct io_kiocb *req)
1821 __must_hold(&ctx->uring_lock)
1823 struct io_ring_ctx *ctx = req->ctx;
1824 struct io_defer_entry *de;
1826 u32 seq = io_get_sequence(req);
1828 /* Still need defer if there is pending req in defer list. */
1829 spin_lock(&ctx->completion_lock);
1830 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1831 spin_unlock(&ctx->completion_lock);
1833 ctx->drain_active = false;
1834 io_req_task_queue(req);
1837 spin_unlock(&ctx->completion_lock);
1839 io_prep_async_link(req);
1840 de = kmalloc(sizeof(*de), GFP_KERNEL);
1843 io_req_defer_failed(req, ret);
1847 spin_lock(&ctx->completion_lock);
1848 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1849 spin_unlock(&ctx->completion_lock);
1854 trace_io_uring_defer(req);
1857 list_add_tail(&de->list, &ctx->defer_list);
1858 spin_unlock(&ctx->completion_lock);
1861 static void io_clean_op(struct io_kiocb *req)
1863 if (req->flags & REQ_F_BUFFER_SELECTED) {
1864 spin_lock(&req->ctx->completion_lock);
1865 io_put_kbuf_comp(req);
1866 spin_unlock(&req->ctx->completion_lock);
1869 if (req->flags & REQ_F_NEED_CLEANUP) {
1870 const struct io_cold_def *def = &io_cold_defs[req->opcode];
1875 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1876 kfree(req->apoll->double_poll);
1880 if (req->flags & REQ_F_INFLIGHT) {
1881 struct io_uring_task *tctx = req->task->io_uring;
1883 atomic_dec(&tctx->inflight_tracked);
1885 if (req->flags & REQ_F_CREDS)
1886 put_cred(req->creds);
1887 if (req->flags & REQ_F_ASYNC_DATA) {
1888 kfree(req->async_data);
1889 req->async_data = NULL;
1891 req->flags &= ~IO_REQ_CLEAN_FLAGS;
1894 static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def,
1895 unsigned int issue_flags)
1897 if (req->file || !def->needs_file)
1900 if (req->flags & REQ_F_FIXED_FILE)
1901 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1903 req->file = io_file_get_normal(req, req->cqe.fd);
1908 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1910 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1911 const struct cred *creds = NULL;
1914 if (unlikely(!io_assign_file(req, def, issue_flags)))
1917 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1918 creds = override_creds(req->creds);
1920 if (!def->audit_skip)
1921 audit_uring_entry(req->opcode);
1923 ret = def->issue(req, issue_flags);
1925 if (!def->audit_skip)
1926 audit_uring_exit(!ret, ret);
1929 revert_creds(creds);
1931 if (ret == IOU_OK) {
1932 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1933 io_req_complete_defer(req);
1935 io_req_complete_post(req, issue_flags);
1936 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1939 /* If the op doesn't have a file, we're not polling for it */
1940 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
1941 io_iopoll_req_issued(req, issue_flags);
1946 int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts)
1948 io_tw_lock(req->ctx, ts);
1949 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT|
1950 IO_URING_F_COMPLETE_DEFER);
1953 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1955 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1957 req = io_put_req_find_next(req);
1958 return req ? &req->work : NULL;
1961 void io_wq_submit_work(struct io_wq_work *work)
1963 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1964 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1965 unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ;
1966 bool needs_poll = false;
1967 int ret = 0, err = -ECANCELED;
1969 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1970 if (!(req->flags & REQ_F_REFCOUNT))
1971 __io_req_set_refcount(req, 2);
1975 io_arm_ltimeout(req);
1977 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1978 if (work->flags & IO_WQ_WORK_CANCEL) {
1980 io_req_task_queue_fail(req, err);
1983 if (!io_assign_file(req, def, issue_flags)) {
1985 work->flags |= IO_WQ_WORK_CANCEL;
1989 if (req->flags & REQ_F_FORCE_ASYNC) {
1990 bool opcode_poll = def->pollin || def->pollout;
1992 if (opcode_poll && file_can_poll(req->file)) {
1994 issue_flags |= IO_URING_F_NONBLOCK;
1999 ret = io_issue_sqe(req, issue_flags);
2003 * We can get EAGAIN for iopolled IO even though we're
2004 * forcing a sync submission from here, since we can't
2005 * wait for request slots on the block side.
2008 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
2014 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
2016 /* aborted or ready, in either case retry blocking */
2018 issue_flags &= ~IO_URING_F_NONBLOCK;
2021 /* avoid locking problems by failing it from a clean context */
2023 io_req_task_queue_fail(req, ret);
2026 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
2027 unsigned int issue_flags)
2029 struct io_ring_ctx *ctx = req->ctx;
2030 struct file *file = NULL;
2031 unsigned long file_ptr;
2033 io_ring_submit_lock(ctx, issue_flags);
2035 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
2037 fd = array_index_nospec(fd, ctx->nr_user_files);
2038 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
2039 file = (struct file *) (file_ptr & FFS_MASK);
2040 file_ptr &= ~FFS_MASK;
2041 /* mask in overlapping REQ_F and FFS bits */
2042 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
2043 io_req_set_rsrc_node(req, ctx, 0);
2045 io_ring_submit_unlock(ctx, issue_flags);
2049 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
2051 struct file *file = fget(fd);
2053 trace_io_uring_file_get(req, fd);
2055 /* we don't allow fixed io_uring files */
2056 if (file && io_is_uring_fops(file))
2057 io_req_track_inflight(req);
2061 static void io_queue_async(struct io_kiocb *req, int ret)
2062 __must_hold(&req->ctx->uring_lock)
2064 struct io_kiocb *linked_timeout;
2066 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
2067 io_req_defer_failed(req, ret);
2071 linked_timeout = io_prep_linked_timeout(req);
2073 switch (io_arm_poll_handler(req, 0)) {
2074 case IO_APOLL_READY:
2075 io_kbuf_recycle(req, 0);
2076 io_req_task_queue(req);
2078 case IO_APOLL_ABORTED:
2079 io_kbuf_recycle(req, 0);
2080 io_queue_iowq(req, NULL);
2087 io_queue_linked_timeout(linked_timeout);
2090 static inline void io_queue_sqe(struct io_kiocb *req)
2091 __must_hold(&req->ctx->uring_lock)
2095 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
2098 * We async punt it if the file wasn't marked NOWAIT, or if the file
2099 * doesn't support non-blocking read/write attempts
2102 io_arm_ltimeout(req);
2104 io_queue_async(req, ret);
2107 static void io_queue_sqe_fallback(struct io_kiocb *req)
2108 __must_hold(&req->ctx->uring_lock)
2110 if (unlikely(req->flags & REQ_F_FAIL)) {
2112 * We don't submit, fail them all, for that replace hardlinks
2113 * with normal links. Extra REQ_F_LINK is tolerated.
2115 req->flags &= ~REQ_F_HARDLINK;
2116 req->flags |= REQ_F_LINK;
2117 io_req_defer_failed(req, req->cqe.res);
2119 int ret = io_req_prep_async(req);
2121 if (unlikely(ret)) {
2122 io_req_defer_failed(req, ret);
2126 if (unlikely(req->ctx->drain_active))
2129 io_queue_iowq(req, NULL);
2134 * Check SQE restrictions (opcode and flags).
2136 * Returns 'true' if SQE is allowed, 'false' otherwise.
2138 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
2139 struct io_kiocb *req,
2140 unsigned int sqe_flags)
2142 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
2145 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
2146 ctx->restrictions.sqe_flags_required)
2149 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
2150 ctx->restrictions.sqe_flags_required))
2156 static void io_init_req_drain(struct io_kiocb *req)
2158 struct io_ring_ctx *ctx = req->ctx;
2159 struct io_kiocb *head = ctx->submit_state.link.head;
2161 ctx->drain_active = true;
2164 * If we need to drain a request in the middle of a link, drain
2165 * the head request and the next request/link after the current
2166 * link. Considering sequential execution of links,
2167 * REQ_F_IO_DRAIN will be maintained for every request of our
2170 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2171 ctx->drain_next = true;
2175 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
2176 const struct io_uring_sqe *sqe)
2177 __must_hold(&ctx->uring_lock)
2179 const struct io_issue_def *def;
2180 unsigned int sqe_flags;
2184 /* req is partially pre-initialised, see io_preinit_req() */
2185 req->opcode = opcode = READ_ONCE(sqe->opcode);
2186 /* same numerical values with corresponding REQ_F_*, safe to copy */
2187 req->flags = sqe_flags = READ_ONCE(sqe->flags);
2188 req->cqe.user_data = READ_ONCE(sqe->user_data);
2190 req->rsrc_node = NULL;
2191 req->task = current;
2193 if (unlikely(opcode >= IORING_OP_LAST)) {
2197 def = &io_issue_defs[opcode];
2198 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2199 /* enforce forwards compatibility on users */
2200 if (sqe_flags & ~SQE_VALID_FLAGS)
2202 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2203 if (!def->buffer_select)
2205 req->buf_index = READ_ONCE(sqe->buf_group);
2207 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2208 ctx->drain_disabled = true;
2209 if (sqe_flags & IOSQE_IO_DRAIN) {
2210 if (ctx->drain_disabled)
2212 io_init_req_drain(req);
2215 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2216 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2218 /* knock it to the slow queue path, will be drained there */
2219 if (ctx->drain_active)
2220 req->flags |= REQ_F_FORCE_ASYNC;
2221 /* if there is no link, we're at "next" request and need to drain */
2222 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2223 ctx->drain_next = false;
2224 ctx->drain_active = true;
2225 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2229 if (!def->ioprio && sqe->ioprio)
2231 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2234 if (def->needs_file) {
2235 struct io_submit_state *state = &ctx->submit_state;
2237 req->cqe.fd = READ_ONCE(sqe->fd);
2240 * Plug now if we have more than 2 IO left after this, and the
2241 * target is potentially a read/write to block based storage.
2243 if (state->need_plug && def->plug) {
2244 state->plug_started = true;
2245 state->need_plug = false;
2246 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2250 personality = READ_ONCE(sqe->personality);
2254 req->creds = xa_load(&ctx->personalities, personality);
2257 get_cred(req->creds);
2258 ret = security_uring_override_creds(req->creds);
2260 put_cred(req->creds);
2263 req->flags |= REQ_F_CREDS;
2266 return def->prep(req, sqe);
2269 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2270 struct io_kiocb *req, int ret)
2272 struct io_ring_ctx *ctx = req->ctx;
2273 struct io_submit_link *link = &ctx->submit_state.link;
2274 struct io_kiocb *head = link->head;
2276 trace_io_uring_req_failed(sqe, req, ret);
2279 * Avoid breaking links in the middle as it renders links with SQPOLL
2280 * unusable. Instead of failing eagerly, continue assembling the link if
2281 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2282 * should find the flag and handle the rest.
2284 req_fail_link_node(req, ret);
2285 if (head && !(head->flags & REQ_F_FAIL))
2286 req_fail_link_node(head, -ECANCELED);
2288 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2290 link->last->link = req;
2294 io_queue_sqe_fallback(req);
2299 link->last->link = req;
2306 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2307 const struct io_uring_sqe *sqe)
2308 __must_hold(&ctx->uring_lock)
2310 struct io_submit_link *link = &ctx->submit_state.link;
2313 ret = io_init_req(ctx, req, sqe);
2315 return io_submit_fail_init(sqe, req, ret);
2317 trace_io_uring_submit_req(req);
2320 * If we already have a head request, queue this one for async
2321 * submittal once the head completes. If we don't have a head but
2322 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2323 * submitted sync once the chain is complete. If none of those
2324 * conditions are true (normal request), then just queue it.
2326 if (unlikely(link->head)) {
2327 ret = io_req_prep_async(req);
2329 return io_submit_fail_init(sqe, req, ret);
2331 trace_io_uring_link(req, link->head);
2332 link->last->link = req;
2335 if (req->flags & IO_REQ_LINK_FLAGS)
2337 /* last request of the link, flush it */
2340 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2343 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2344 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2345 if (req->flags & IO_REQ_LINK_FLAGS) {
2350 io_queue_sqe_fallback(req);
2360 * Batched submission is done, ensure local IO is flushed out.
2362 static void io_submit_state_end(struct io_ring_ctx *ctx)
2364 struct io_submit_state *state = &ctx->submit_state;
2366 if (unlikely(state->link.head))
2367 io_queue_sqe_fallback(state->link.head);
2368 /* flush only after queuing links as they can generate completions */
2369 io_submit_flush_completions(ctx);
2370 if (state->plug_started)
2371 blk_finish_plug(&state->plug);
2375 * Start submission side cache.
2377 static void io_submit_state_start(struct io_submit_state *state,
2378 unsigned int max_ios)
2380 state->plug_started = false;
2381 state->need_plug = max_ios > 2;
2382 state->submit_nr = max_ios;
2383 /* set only head, no need to init link_last in advance */
2384 state->link.head = NULL;
2387 static void io_commit_sqring(struct io_ring_ctx *ctx)
2389 struct io_rings *rings = ctx->rings;
2392 * Ensure any loads from the SQEs are done at this point,
2393 * since once we write the new head, the application could
2394 * write new data to them.
2396 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2400 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2401 * that is mapped by userspace. This means that care needs to be taken to
2402 * ensure that reads are stable, as we cannot rely on userspace always
2403 * being a good citizen. If members of the sqe are validated and then later
2404 * used, it's important that those reads are done through READ_ONCE() to
2405 * prevent a re-load down the line.
2407 static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe)
2409 unsigned head, mask = ctx->sq_entries - 1;
2410 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2413 * The cached sq head (or cq tail) serves two purposes:
2415 * 1) allows us to batch the cost of updating the user visible
2417 * 2) allows the kernel side to track the head on its own, even
2418 * though the application is the one updating it.
2420 head = READ_ONCE(ctx->sq_array[sq_idx]);
2421 if (likely(head < ctx->sq_entries)) {
2422 /* double index for 128-byte SQEs, twice as long */
2423 if (ctx->flags & IORING_SETUP_SQE128)
2425 *sqe = &ctx->sq_sqes[head];
2429 /* drop invalid entries */
2431 WRITE_ONCE(ctx->rings->sq_dropped,
2432 READ_ONCE(ctx->rings->sq_dropped) + 1);
2436 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2437 __must_hold(&ctx->uring_lock)
2439 unsigned int entries = io_sqring_entries(ctx);
2443 if (unlikely(!entries))
2445 /* make sure SQ entry isn't read before tail */
2446 ret = left = min(nr, entries);
2447 io_get_task_refs(left);
2448 io_submit_state_start(&ctx->submit_state, left);
2451 const struct io_uring_sqe *sqe;
2452 struct io_kiocb *req;
2454 if (unlikely(!io_alloc_req(ctx, &req)))
2456 if (unlikely(!io_get_sqe(ctx, &sqe))) {
2457 io_req_add_to_cache(req, ctx);
2462 * Continue submitting even for sqe failure if the
2463 * ring was setup with IORING_SETUP_SUBMIT_ALL
2465 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2466 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2472 if (unlikely(left)) {
2474 /* try again if it submitted nothing and can't allocate a req */
2475 if (!ret && io_req_cache_empty(ctx))
2477 current->io_uring->cached_refs += left;
2480 io_submit_state_end(ctx);
2481 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2482 io_commit_sqring(ctx);
2486 struct io_wait_queue {
2487 struct wait_queue_entry wq;
2488 struct io_ring_ctx *ctx;
2490 unsigned nr_timeouts;
2494 static inline bool io_has_work(struct io_ring_ctx *ctx)
2496 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
2497 !llist_empty(&ctx->work_llist);
2500 static inline bool io_should_wake(struct io_wait_queue *iowq)
2502 struct io_ring_ctx *ctx = iowq->ctx;
2503 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
2506 * Wake up if we have enough events, or if a timeout occurred since we
2507 * started waiting. For timeouts, we always want to return to userspace,
2508 * regardless of event count.
2510 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2513 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2514 int wake_flags, void *key)
2516 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq);
2519 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2520 * the task, and the next invocation will do it.
2522 if (io_should_wake(iowq) || io_has_work(iowq->ctx))
2523 return autoremove_wake_function(curr, mode, wake_flags, key);
2527 int io_run_task_work_sig(struct io_ring_ctx *ctx)
2529 if (!llist_empty(&ctx->work_llist)) {
2530 __set_current_state(TASK_RUNNING);
2531 if (io_run_local_work(ctx) > 0)
2534 if (io_run_task_work() > 0)
2536 if (task_sigpending(current))
2541 /* when returns >0, the caller should retry */
2542 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2543 struct io_wait_queue *iowq)
2545 if (unlikely(READ_ONCE(ctx->check_cq)))
2547 if (unlikely(!llist_empty(&ctx->work_llist)))
2549 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL)))
2551 if (unlikely(task_sigpending(current)))
2553 if (unlikely(io_should_wake(iowq)))
2555 if (iowq->timeout == KTIME_MAX)
2557 else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS))
2563 * Wait until events become available, if we don't already have some. The
2564 * application must reap them itself, as they reside on the shared cq ring.
2566 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2567 const sigset_t __user *sig, size_t sigsz,
2568 struct __kernel_timespec __user *uts)
2570 struct io_wait_queue iowq;
2571 struct io_rings *rings = ctx->rings;
2574 if (!io_allowed_run_tw(ctx))
2576 if (!llist_empty(&ctx->work_llist))
2577 io_run_local_work(ctx);
2579 io_cqring_overflow_flush(ctx);
2580 /* if user messes with these they will just get an early return */
2581 if (__io_cqring_events_user(ctx) >= min_events)
2585 #ifdef CONFIG_COMPAT
2586 if (in_compat_syscall())
2587 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2591 ret = set_user_sigmask(sig, sigsz);
2597 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2598 iowq.wq.private = current;
2599 INIT_LIST_HEAD(&iowq.wq.entry);
2601 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2602 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2603 iowq.timeout = KTIME_MAX;
2606 struct timespec64 ts;
2608 if (get_timespec64(&ts, uts))
2610 iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2613 trace_io_uring_cqring_wait(ctx, min_events);
2615 unsigned long check_cq;
2617 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
2618 WRITE_ONCE(ctx->cq_waiting, 1);
2619 set_current_state(TASK_INTERRUPTIBLE);
2621 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2622 TASK_INTERRUPTIBLE);
2625 ret = io_cqring_wait_schedule(ctx, &iowq);
2626 __set_current_state(TASK_RUNNING);
2627 WRITE_ONCE(ctx->cq_waiting, 0);
2632 * Run task_work after scheduling and before io_should_wake().
2633 * If we got woken because of task_work being processed, run it
2634 * now rather than let the caller do another wait loop.
2637 if (!llist_empty(&ctx->work_llist))
2638 io_run_local_work(ctx);
2640 check_cq = READ_ONCE(ctx->check_cq);
2641 if (unlikely(check_cq)) {
2642 /* let the caller flush overflows, retry */
2643 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2644 io_cqring_do_overflow_flush(ctx);
2645 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) {
2651 if (io_should_wake(&iowq)) {
2658 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
2659 finish_wait(&ctx->cq_wait, &iowq.wq);
2660 restore_saved_sigmask_unless(ret == -EINTR);
2662 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2665 static void io_mem_free(void *ptr)
2672 page = virt_to_head_page(ptr);
2673 if (put_page_testzero(page))
2674 free_compound_page(page);
2677 static void *io_mem_alloc(size_t size)
2679 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2681 return (void *) __get_free_pages(gfp, get_order(size));
2684 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2685 unsigned int cq_entries, size_t *sq_offset)
2687 struct io_rings *rings;
2688 size_t off, sq_array_size;
2690 off = struct_size(rings, cqes, cq_entries);
2691 if (off == SIZE_MAX)
2693 if (ctx->flags & IORING_SETUP_CQE32) {
2694 if (check_shl_overflow(off, 1, &off))
2699 off = ALIGN(off, SMP_CACHE_BYTES);
2707 sq_array_size = array_size(sizeof(u32), sq_entries);
2708 if (sq_array_size == SIZE_MAX)
2711 if (check_add_overflow(off, sq_array_size, &off))
2717 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2718 unsigned int eventfd_async)
2720 struct io_ev_fd *ev_fd;
2721 __s32 __user *fds = arg;
2724 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2725 lockdep_is_held(&ctx->uring_lock));
2729 if (copy_from_user(&fd, fds, sizeof(*fds)))
2732 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2736 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2737 if (IS_ERR(ev_fd->cq_ev_fd)) {
2738 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2743 spin_lock(&ctx->completion_lock);
2744 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2745 spin_unlock(&ctx->completion_lock);
2747 ev_fd->eventfd_async = eventfd_async;
2748 ctx->has_evfd = true;
2749 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2750 atomic_set(&ev_fd->refs, 1);
2751 atomic_set(&ev_fd->ops, 0);
2755 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2757 struct io_ev_fd *ev_fd;
2759 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2760 lockdep_is_held(&ctx->uring_lock));
2762 ctx->has_evfd = false;
2763 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2764 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops))
2765 call_rcu(&ev_fd->rcu, io_eventfd_ops);
2772 static void io_req_caches_free(struct io_ring_ctx *ctx)
2774 struct io_kiocb *req;
2777 mutex_lock(&ctx->uring_lock);
2778 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
2780 while (!io_req_cache_empty(ctx)) {
2781 req = io_extract_req(ctx);
2782 kmem_cache_free(req_cachep, req);
2786 percpu_ref_put_many(&ctx->refs, nr);
2787 mutex_unlock(&ctx->uring_lock);
2790 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2792 io_sq_thread_finish(ctx);
2793 io_rsrc_refs_drop(ctx);
2794 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2795 io_wait_rsrc_data(ctx->buf_data);
2796 io_wait_rsrc_data(ctx->file_data);
2798 mutex_lock(&ctx->uring_lock);
2800 __io_sqe_buffers_unregister(ctx);
2802 __io_sqe_files_unregister(ctx);
2803 io_cqring_overflow_kill(ctx);
2804 io_eventfd_unregister(ctx);
2805 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2806 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2807 mutex_unlock(&ctx->uring_lock);
2808 io_destroy_buffers(ctx);
2810 put_cred(ctx->sq_creds);
2811 if (ctx->submitter_task)
2812 put_task_struct(ctx->submitter_task);
2814 /* there are no registered resources left, nobody uses it */
2816 io_rsrc_node_destroy(ctx->rsrc_node);
2817 if (ctx->rsrc_backup_node)
2818 io_rsrc_node_destroy(ctx->rsrc_backup_node);
2819 flush_delayed_work(&ctx->rsrc_put_work);
2820 flush_delayed_work(&ctx->fallback_work);
2822 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2823 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
2825 #if defined(CONFIG_UNIX)
2826 if (ctx->ring_sock) {
2827 ctx->ring_sock->file = NULL; /* so that iput() is called */
2828 sock_release(ctx->ring_sock);
2831 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2833 if (ctx->mm_account) {
2834 mmdrop(ctx->mm_account);
2835 ctx->mm_account = NULL;
2837 io_mem_free(ctx->rings);
2838 io_mem_free(ctx->sq_sqes);
2840 percpu_ref_exit(&ctx->refs);
2841 free_uid(ctx->user);
2842 io_req_caches_free(ctx);
2844 io_wq_put_hash(ctx->hash_map);
2845 kfree(ctx->cancel_table.hbs);
2846 kfree(ctx->cancel_table_locked.hbs);
2847 kfree(ctx->dummy_ubuf);
2849 xa_destroy(&ctx->io_bl_xa);
2853 static __cold void io_activate_pollwq_cb(struct callback_head *cb)
2855 struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx,
2858 mutex_lock(&ctx->uring_lock);
2859 ctx->poll_activated = true;
2860 mutex_unlock(&ctx->uring_lock);
2863 * Wake ups for some events between start of polling and activation
2864 * might've been lost due to loose synchronisation.
2866 wake_up_all(&ctx->poll_wq);
2867 percpu_ref_put(&ctx->refs);
2870 static __cold void io_activate_pollwq(struct io_ring_ctx *ctx)
2872 spin_lock(&ctx->completion_lock);
2873 /* already activated or in progress */
2874 if (ctx->poll_activated || ctx->poll_wq_task_work.func)
2876 if (WARN_ON_ONCE(!ctx->task_complete))
2878 if (!ctx->submitter_task)
2881 * with ->submitter_task only the submitter task completes requests, we
2882 * only need to sync with it, which is done by injecting a tw
2884 init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb);
2885 percpu_ref_get(&ctx->refs);
2886 if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL))
2887 percpu_ref_put(&ctx->refs);
2889 spin_unlock(&ctx->completion_lock);
2892 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2894 struct io_ring_ctx *ctx = file->private_data;
2897 if (unlikely(!ctx->poll_activated))
2898 io_activate_pollwq(ctx);
2900 poll_wait(file, &ctx->poll_wq, wait);
2902 * synchronizes with barrier from wq_has_sleeper call in
2906 if (!io_sqring_full(ctx))
2907 mask |= EPOLLOUT | EPOLLWRNORM;
2910 * Don't flush cqring overflow list here, just do a simple check.
2911 * Otherwise there could possible be ABBA deadlock:
2914 * lock(&ctx->uring_lock);
2916 * lock(&ctx->uring_lock);
2919 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2920 * pushes them to do the flush.
2923 if (__io_cqring_events_user(ctx) || io_has_work(ctx))
2924 mask |= EPOLLIN | EPOLLRDNORM;
2929 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2931 const struct cred *creds;
2933 creds = xa_erase(&ctx->personalities, id);
2942 struct io_tctx_exit {
2943 struct callback_head task_work;
2944 struct completion completion;
2945 struct io_ring_ctx *ctx;
2948 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2950 struct io_uring_task *tctx = current->io_uring;
2951 struct io_tctx_exit *work;
2953 work = container_of(cb, struct io_tctx_exit, task_work);
2955 * When @in_cancel, we're in cancellation and it's racy to remove the
2956 * node. It'll be removed by the end of cancellation, just ignore it.
2957 * tctx can be NULL if the queueing of this task_work raced with
2958 * work cancelation off the exec path.
2960 if (tctx && !atomic_read(&tctx->in_cancel))
2961 io_uring_del_tctx_node((unsigned long)work->ctx);
2962 complete(&work->completion);
2965 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2967 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2969 return req->ctx == data;
2972 static __cold void io_ring_exit_work(struct work_struct *work)
2974 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2975 unsigned long timeout = jiffies + HZ * 60 * 5;
2976 unsigned long interval = HZ / 20;
2977 struct io_tctx_exit exit;
2978 struct io_tctx_node *node;
2982 * If we're doing polled IO and end up having requests being
2983 * submitted async (out-of-line), then completions can come in while
2984 * we're waiting for refs to drop. We need to reap these manually,
2985 * as nobody else will be looking for them.
2988 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
2989 mutex_lock(&ctx->uring_lock);
2990 io_cqring_overflow_kill(ctx);
2991 mutex_unlock(&ctx->uring_lock);
2994 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
2995 io_move_task_work_from_local(ctx);
2997 while (io_uring_try_cancel_requests(ctx, NULL, true))
3001 struct io_sq_data *sqd = ctx->sq_data;
3002 struct task_struct *tsk;
3004 io_sq_thread_park(sqd);
3006 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
3007 io_wq_cancel_cb(tsk->io_uring->io_wq,
3008 io_cancel_ctx_cb, ctx, true);
3009 io_sq_thread_unpark(sqd);
3012 io_req_caches_free(ctx);
3014 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
3015 /* there is little hope left, don't run it too often */
3018 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
3020 init_completion(&exit.completion);
3021 init_task_work(&exit.task_work, io_tctx_exit_cb);
3024 * Some may use context even when all refs and requests have been put,
3025 * and they are free to do so while still holding uring_lock or
3026 * completion_lock, see io_req_task_submit(). Apart from other work,
3027 * this lock/unlock section also waits them to finish.
3029 mutex_lock(&ctx->uring_lock);
3030 while (!list_empty(&ctx->tctx_list)) {
3031 WARN_ON_ONCE(time_after(jiffies, timeout));
3033 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
3035 /* don't spin on a single task if cancellation failed */
3036 list_rotate_left(&ctx->tctx_list);
3037 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
3038 if (WARN_ON_ONCE(ret))
3041 mutex_unlock(&ctx->uring_lock);
3042 wait_for_completion(&exit.completion);
3043 mutex_lock(&ctx->uring_lock);
3045 mutex_unlock(&ctx->uring_lock);
3046 spin_lock(&ctx->completion_lock);
3047 spin_unlock(&ctx->completion_lock);
3049 io_ring_ctx_free(ctx);
3052 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
3054 unsigned long index;
3055 struct creds *creds;
3057 mutex_lock(&ctx->uring_lock);
3058 percpu_ref_kill(&ctx->refs);
3059 xa_for_each(&ctx->personalities, index, creds)
3060 io_unregister_personality(ctx, index);
3062 io_poll_remove_all(ctx, NULL, true);
3063 mutex_unlock(&ctx->uring_lock);
3066 * If we failed setting up the ctx, we might not have any rings
3067 * and therefore did not submit any requests
3070 io_kill_timeouts(ctx, NULL, true);
3072 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
3074 * Use system_unbound_wq to avoid spawning tons of event kworkers
3075 * if we're exiting a ton of rings at the same time. It just adds
3076 * noise and overhead, there's no discernable change in runtime
3077 * over using system_wq.
3079 queue_work(system_unbound_wq, &ctx->exit_work);
3082 static int io_uring_release(struct inode *inode, struct file *file)
3084 struct io_ring_ctx *ctx = file->private_data;
3086 file->private_data = NULL;
3087 io_ring_ctx_wait_and_kill(ctx);
3091 struct io_task_cancel {
3092 struct task_struct *task;
3096 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
3098 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3099 struct io_task_cancel *cancel = data;
3101 return io_match_task_safe(req, cancel->task, cancel->all);
3104 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
3105 struct task_struct *task,
3108 struct io_defer_entry *de;
3111 spin_lock(&ctx->completion_lock);
3112 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
3113 if (io_match_task_safe(de->req, task, cancel_all)) {
3114 list_cut_position(&list, &ctx->defer_list, &de->list);
3118 spin_unlock(&ctx->completion_lock);
3119 if (list_empty(&list))
3122 while (!list_empty(&list)) {
3123 de = list_first_entry(&list, struct io_defer_entry, list);
3124 list_del_init(&de->list);
3125 io_req_task_queue_fail(de->req, -ECANCELED);
3131 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
3133 struct io_tctx_node *node;
3134 enum io_wq_cancel cret;
3137 mutex_lock(&ctx->uring_lock);
3138 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3139 struct io_uring_task *tctx = node->task->io_uring;
3142 * io_wq will stay alive while we hold uring_lock, because it's
3143 * killed after ctx nodes, which requires to take the lock.
3145 if (!tctx || !tctx->io_wq)
3147 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
3148 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3150 mutex_unlock(&ctx->uring_lock);
3155 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
3156 struct task_struct *task,
3159 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
3160 struct io_uring_task *tctx = task ? task->io_uring : NULL;
3161 enum io_wq_cancel cret;
3164 /* failed during ring init, it couldn't have issued any requests */
3169 ret |= io_uring_try_cancel_iowq(ctx);
3170 } else if (tctx && tctx->io_wq) {
3172 * Cancels requests of all rings, not only @ctx, but
3173 * it's fine as the task is in exit/exec.
3175 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
3177 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3180 /* SQPOLL thread does its own polling */
3181 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
3182 (ctx->sq_data && ctx->sq_data->thread == current)) {
3183 while (!wq_list_empty(&ctx->iopoll_list)) {
3184 io_iopoll_try_reap_events(ctx);
3190 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3191 io_allowed_defer_tw_run(ctx))
3192 ret |= io_run_local_work(ctx) > 0;
3193 ret |= io_cancel_defer_files(ctx, task, cancel_all);
3194 mutex_lock(&ctx->uring_lock);
3195 ret |= io_poll_remove_all(ctx, task, cancel_all);
3196 mutex_unlock(&ctx->uring_lock);
3197 ret |= io_kill_timeouts(ctx, task, cancel_all);
3199 ret |= io_run_task_work() > 0;
3203 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
3206 return atomic_read(&tctx->inflight_tracked);
3207 return percpu_counter_sum(&tctx->inflight);
3211 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3212 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3214 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
3216 struct io_uring_task *tctx = current->io_uring;
3217 struct io_ring_ctx *ctx;
3221 WARN_ON_ONCE(sqd && sqd->thread != current);
3223 if (!current->io_uring)
3226 io_wq_exit_start(tctx->io_wq);
3228 atomic_inc(&tctx->in_cancel);
3232 io_uring_drop_tctx_refs(current);
3233 /* read completions before cancelations */
3234 inflight = tctx_inflight(tctx, !cancel_all);
3239 struct io_tctx_node *node;
3240 unsigned long index;
3242 xa_for_each(&tctx->xa, index, node) {
3243 /* sqpoll task will cancel all its requests */
3244 if (node->ctx->sq_data)
3246 loop |= io_uring_try_cancel_requests(node->ctx,
3247 current, cancel_all);
3250 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3251 loop |= io_uring_try_cancel_requests(ctx,
3261 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3263 io_uring_drop_tctx_refs(current);
3266 * If we've seen completions, retry without waiting. This
3267 * avoids a race where a completion comes in before we did
3268 * prepare_to_wait().
3270 if (inflight == tctx_inflight(tctx, !cancel_all))
3272 finish_wait(&tctx->wait, &wait);
3275 io_uring_clean_tctx(tctx);
3278 * We shouldn't run task_works after cancel, so just leave
3279 * ->in_cancel set for normal exit.
3281 atomic_dec(&tctx->in_cancel);
3282 /* for exec all current's requests should be gone, kill tctx */
3283 __io_uring_free(current);
3287 void __io_uring_cancel(bool cancel_all)
3289 io_uring_cancel_generic(cancel_all, NULL);
3292 static void *io_uring_validate_mmap_request(struct file *file,
3293 loff_t pgoff, size_t sz)
3295 struct io_ring_ctx *ctx = file->private_data;
3296 loff_t offset = pgoff << PAGE_SHIFT;
3300 switch (offset & IORING_OFF_MMAP_MASK) {
3301 case IORING_OFF_SQ_RING:
3302 case IORING_OFF_CQ_RING:
3305 case IORING_OFF_SQES:
3308 case IORING_OFF_PBUF_RING: {
3311 bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT;
3312 mutex_lock(&ctx->uring_lock);
3313 ptr = io_pbuf_get_address(ctx, bgid);
3314 mutex_unlock(&ctx->uring_lock);
3316 return ERR_PTR(-EINVAL);
3320 return ERR_PTR(-EINVAL);
3323 page = virt_to_head_page(ptr);
3324 if (sz > page_size(page))
3325 return ERR_PTR(-EINVAL);
3332 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3334 size_t sz = vma->vm_end - vma->vm_start;
3338 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3340 return PTR_ERR(ptr);
3342 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3343 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3346 static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp,
3347 unsigned long addr, unsigned long len,
3348 unsigned long pgoff, unsigned long flags)
3350 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
3351 struct vm_unmapped_area_info info;
3355 * Do not allow to map to user-provided address to avoid breaking the
3356 * aliasing rules. Userspace is not able to guess the offset address of
3357 * kernel kmalloc()ed memory area.
3362 ptr = io_uring_validate_mmap_request(filp, pgoff, len);
3366 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
3368 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
3369 info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
3371 info.align_mask = PAGE_MASK & (SHM_COLOUR - 1UL);
3373 info.align_mask = PAGE_MASK & (SHMLBA - 1UL);
3375 info.align_offset = (unsigned long) ptr;
3378 * A failed mmap() very likely causes application failure,
3379 * so fall back to the bottom-up function here. This scenario
3380 * can happen with large stack limits and large mmap()
3383 addr = vm_unmapped_area(&info);
3384 if (offset_in_page(addr)) {
3386 info.low_limit = TASK_UNMAPPED_BASE;
3387 info.high_limit = mmap_end;
3388 addr = vm_unmapped_area(&info);
3394 #else /* !CONFIG_MMU */
3396 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3398 return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL;
3401 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3403 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3406 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3407 unsigned long addr, unsigned long len,
3408 unsigned long pgoff, unsigned long flags)
3412 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3414 return PTR_ERR(ptr);
3416 return (unsigned long) ptr;
3419 #endif /* !CONFIG_MMU */
3421 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3423 if (flags & IORING_ENTER_EXT_ARG) {
3424 struct io_uring_getevents_arg arg;
3426 if (argsz != sizeof(arg))
3428 if (copy_from_user(&arg, argp, sizeof(arg)))
3434 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3435 struct __kernel_timespec __user **ts,
3436 const sigset_t __user **sig)
3438 struct io_uring_getevents_arg arg;
3441 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3442 * is just a pointer to the sigset_t.
3444 if (!(flags & IORING_ENTER_EXT_ARG)) {
3445 *sig = (const sigset_t __user *) argp;
3451 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3452 * timespec and sigset_t pointers if good.
3454 if (*argsz != sizeof(arg))
3456 if (copy_from_user(&arg, argp, sizeof(arg)))
3460 *sig = u64_to_user_ptr(arg.sigmask);
3461 *argsz = arg.sigmask_sz;
3462 *ts = u64_to_user_ptr(arg.ts);
3466 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3467 u32, min_complete, u32, flags, const void __user *, argp,
3470 struct io_ring_ctx *ctx;
3474 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3475 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3476 IORING_ENTER_REGISTERED_RING)))
3480 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3481 * need only dereference our task private array to find it.
3483 if (flags & IORING_ENTER_REGISTERED_RING) {
3484 struct io_uring_task *tctx = current->io_uring;
3486 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3488 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3489 f.file = tctx->registered_rings[fd];
3491 if (unlikely(!f.file))
3495 if (unlikely(!f.file))
3498 if (unlikely(!io_is_uring_fops(f.file)))
3502 ctx = f.file->private_data;
3504 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3508 * For SQ polling, the thread will do all submissions and completions.
3509 * Just return the requested submit count, and wake the thread if
3513 if (ctx->flags & IORING_SETUP_SQPOLL) {
3514 io_cqring_overflow_flush(ctx);
3516 if (unlikely(ctx->sq_data->thread == NULL)) {
3520 if (flags & IORING_ENTER_SQ_WAKEUP)
3521 wake_up(&ctx->sq_data->wait);
3522 if (flags & IORING_ENTER_SQ_WAIT)
3523 io_sqpoll_wait_sq(ctx);
3526 } else if (to_submit) {
3527 ret = io_uring_add_tctx_node(ctx);
3531 mutex_lock(&ctx->uring_lock);
3532 ret = io_submit_sqes(ctx, to_submit);
3533 if (ret != to_submit) {
3534 mutex_unlock(&ctx->uring_lock);
3537 if (flags & IORING_ENTER_GETEVENTS) {
3538 if (ctx->syscall_iopoll)
3541 * Ignore errors, we'll soon call io_cqring_wait() and
3542 * it should handle ownership problems if any.
3544 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3545 (void)io_run_local_work_locked(ctx);
3547 mutex_unlock(&ctx->uring_lock);
3550 if (flags & IORING_ENTER_GETEVENTS) {
3553 if (ctx->syscall_iopoll) {
3555 * We disallow the app entering submit/complete with
3556 * polling, but we still need to lock the ring to
3557 * prevent racing with polled issue that got punted to
3560 mutex_lock(&ctx->uring_lock);
3562 ret2 = io_validate_ext_arg(flags, argp, argsz);
3563 if (likely(!ret2)) {
3564 min_complete = min(min_complete,
3566 ret2 = io_iopoll_check(ctx, min_complete);
3568 mutex_unlock(&ctx->uring_lock);
3570 const sigset_t __user *sig;
3571 struct __kernel_timespec __user *ts;
3573 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3574 if (likely(!ret2)) {
3575 min_complete = min(min_complete,
3577 ret2 = io_cqring_wait(ctx, min_complete, sig,
3586 * EBADR indicates that one or more CQE were dropped.
3587 * Once the user has been informed we can clear the bit
3588 * as they are obviously ok with those drops.
3590 if (unlikely(ret2 == -EBADR))
3591 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3600 static const struct file_operations io_uring_fops = {
3601 .release = io_uring_release,
3602 .mmap = io_uring_mmap,
3604 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3605 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3607 .get_unmapped_area = io_uring_mmu_get_unmapped_area,
3609 .poll = io_uring_poll,
3610 #ifdef CONFIG_PROC_FS
3611 .show_fdinfo = io_uring_show_fdinfo,
3615 bool io_is_uring_fops(struct file *file)
3617 return file->f_op == &io_uring_fops;
3620 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3621 struct io_uring_params *p)
3623 struct io_rings *rings;
3624 size_t size, sq_array_offset;
3626 /* make sure these are sane, as we already accounted them */
3627 ctx->sq_entries = p->sq_entries;
3628 ctx->cq_entries = p->cq_entries;
3630 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3631 if (size == SIZE_MAX)
3634 rings = io_mem_alloc(size);
3639 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3640 rings->sq_ring_mask = p->sq_entries - 1;
3641 rings->cq_ring_mask = p->cq_entries - 1;
3642 rings->sq_ring_entries = p->sq_entries;
3643 rings->cq_ring_entries = p->cq_entries;
3645 if (p->flags & IORING_SETUP_SQE128)
3646 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3648 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3649 if (size == SIZE_MAX) {
3650 io_mem_free(ctx->rings);
3655 ctx->sq_sqes = io_mem_alloc(size);
3656 if (!ctx->sq_sqes) {
3657 io_mem_free(ctx->rings);
3665 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3669 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3673 ret = __io_uring_add_tctx_node(ctx);
3678 fd_install(fd, file);
3683 * Allocate an anonymous fd, this is what constitutes the application
3684 * visible backing of an io_uring instance. The application mmaps this
3685 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3686 * we have to tie this fd to a socket for file garbage collection purposes.
3688 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3691 #if defined(CONFIG_UNIX)
3694 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3697 return ERR_PTR(ret);
3700 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3701 O_RDWR | O_CLOEXEC, NULL);
3702 #if defined(CONFIG_UNIX)
3704 sock_release(ctx->ring_sock);
3705 ctx->ring_sock = NULL;
3707 ctx->ring_sock->file = file;
3713 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3714 struct io_uring_params __user *params)
3716 struct io_ring_ctx *ctx;
3722 if (entries > IORING_MAX_ENTRIES) {
3723 if (!(p->flags & IORING_SETUP_CLAMP))
3725 entries = IORING_MAX_ENTRIES;
3729 * Use twice as many entries for the CQ ring. It's possible for the
3730 * application to drive a higher depth than the size of the SQ ring,
3731 * since the sqes are only used at submission time. This allows for
3732 * some flexibility in overcommitting a bit. If the application has
3733 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3734 * of CQ ring entries manually.
3736 p->sq_entries = roundup_pow_of_two(entries);
3737 if (p->flags & IORING_SETUP_CQSIZE) {
3739 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3740 * to a power-of-two, if it isn't already. We do NOT impose
3741 * any cq vs sq ring sizing.
3745 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3746 if (!(p->flags & IORING_SETUP_CLAMP))
3748 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3750 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3751 if (p->cq_entries < p->sq_entries)
3754 p->cq_entries = 2 * p->sq_entries;
3757 ctx = io_ring_ctx_alloc(p);
3761 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3762 !(ctx->flags & IORING_SETUP_IOPOLL) &&
3763 !(ctx->flags & IORING_SETUP_SQPOLL))
3764 ctx->task_complete = true;
3767 * lazy poll_wq activation relies on ->task_complete for synchronisation
3768 * purposes, see io_activate_pollwq()
3770 if (!ctx->task_complete)
3771 ctx->poll_activated = true;
3774 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3775 * space applications don't need to do io completion events
3776 * polling again, they can rely on io_sq_thread to do polling
3777 * work, which can reduce cpu usage and uring_lock contention.
3779 if (ctx->flags & IORING_SETUP_IOPOLL &&
3780 !(ctx->flags & IORING_SETUP_SQPOLL))
3781 ctx->syscall_iopoll = 1;
3783 ctx->compat = in_compat_syscall();
3784 if (!capable(CAP_IPC_LOCK))
3785 ctx->user = get_uid(current_user());
3788 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3789 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3792 if (ctx->flags & IORING_SETUP_SQPOLL) {
3793 /* IPI related flags don't make sense with SQPOLL */
3794 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3795 IORING_SETUP_TASKRUN_FLAG |
3796 IORING_SETUP_DEFER_TASKRUN))
3798 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3799 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3800 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3802 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3803 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3805 ctx->notify_method = TWA_SIGNAL;
3809 * For DEFER_TASKRUN we require the completion task to be the same as the
3810 * submission task. This implies that there is only one submitter, so enforce
3813 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3814 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3819 * This is just grabbed for accounting purposes. When a process exits,
3820 * the mm is exited and dropped before the files, hence we need to hang
3821 * on to this mm purely for the purposes of being able to unaccount
3822 * memory (locked/pinned vm). It's not used for anything else.
3824 mmgrab(current->mm);
3825 ctx->mm_account = current->mm;
3827 ret = io_allocate_scq_urings(ctx, p);
3831 ret = io_sq_offload_create(ctx, p);
3834 /* always set a rsrc node */
3835 ret = io_rsrc_node_switch_start(ctx);
3838 io_rsrc_node_switch(ctx, NULL);
3840 memset(&p->sq_off, 0, sizeof(p->sq_off));
3841 p->sq_off.head = offsetof(struct io_rings, sq.head);
3842 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3843 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3844 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3845 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3846 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3847 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3849 memset(&p->cq_off, 0, sizeof(p->cq_off));
3850 p->cq_off.head = offsetof(struct io_rings, cq.head);
3851 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3852 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3853 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3854 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3855 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3856 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3858 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3859 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3860 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3861 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3862 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3863 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3864 IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING;
3866 if (copy_to_user(params, p, sizeof(*p))) {
3871 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
3872 && !(ctx->flags & IORING_SETUP_R_DISABLED))
3873 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
3875 file = io_uring_get_file(ctx);
3877 ret = PTR_ERR(file);
3882 * Install ring fd as the very last thing, so we don't risk someone
3883 * having closed it before we finish setup
3885 ret = io_uring_install_fd(ctx, file);
3887 /* fput will clean it up */
3892 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3895 io_ring_ctx_wait_and_kill(ctx);
3900 * Sets up an aio uring context, and returns the fd. Applications asks for a
3901 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3902 * params structure passed in.
3904 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3906 struct io_uring_params p;
3909 if (copy_from_user(&p, params, sizeof(p)))
3911 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3916 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3917 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3918 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3919 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3920 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3921 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3922 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN))
3925 return io_uring_create(entries, &p, params);
3928 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3929 struct io_uring_params __user *, params)
3931 return io_uring_setup(entries, params);
3934 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3937 struct io_uring_probe *p;
3941 size = struct_size(p, ops, nr_args);
3942 if (size == SIZE_MAX)
3944 p = kzalloc(size, GFP_KERNEL);
3949 if (copy_from_user(p, arg, size))
3952 if (memchr_inv(p, 0, size))
3955 p->last_op = IORING_OP_LAST - 1;
3956 if (nr_args > IORING_OP_LAST)
3957 nr_args = IORING_OP_LAST;
3959 for (i = 0; i < nr_args; i++) {
3961 if (!io_issue_defs[i].not_supported)
3962 p->ops[i].flags = IO_URING_OP_SUPPORTED;
3967 if (copy_to_user(arg, p, size))
3974 static int io_register_personality(struct io_ring_ctx *ctx)
3976 const struct cred *creds;
3980 creds = get_current_cred();
3982 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
3983 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
3991 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
3992 void __user *arg, unsigned int nr_args)
3994 struct io_uring_restriction *res;
3998 /* Restrictions allowed only if rings started disabled */
3999 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
4002 /* We allow only a single restrictions registration */
4003 if (ctx->restrictions.registered)
4006 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
4009 size = array_size(nr_args, sizeof(*res));
4010 if (size == SIZE_MAX)
4013 res = memdup_user(arg, size);
4015 return PTR_ERR(res);
4019 for (i = 0; i < nr_args; i++) {
4020 switch (res[i].opcode) {
4021 case IORING_RESTRICTION_REGISTER_OP:
4022 if (res[i].register_op >= IORING_REGISTER_LAST) {
4027 __set_bit(res[i].register_op,
4028 ctx->restrictions.register_op);
4030 case IORING_RESTRICTION_SQE_OP:
4031 if (res[i].sqe_op >= IORING_OP_LAST) {
4036 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
4038 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
4039 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
4041 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
4042 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
4051 /* Reset all restrictions if an error happened */
4053 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
4055 ctx->restrictions.registered = true;
4061 static int io_register_enable_rings(struct io_ring_ctx *ctx)
4063 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
4066 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task) {
4067 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
4069 * Lazy activation attempts would fail if it was polled before
4070 * submitter_task is set.
4072 if (wq_has_sleeper(&ctx->poll_wq))
4073 io_activate_pollwq(ctx);
4076 if (ctx->restrictions.registered)
4077 ctx->restricted = 1;
4079 ctx->flags &= ~IORING_SETUP_R_DISABLED;
4080 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
4081 wake_up(&ctx->sq_data->wait);
4085 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
4086 void __user *arg, unsigned len)
4088 struct io_uring_task *tctx = current->io_uring;
4089 cpumask_var_t new_mask;
4092 if (!tctx || !tctx->io_wq)
4095 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4098 cpumask_clear(new_mask);
4099 if (len > cpumask_size())
4100 len = cpumask_size();
4102 if (in_compat_syscall()) {
4103 ret = compat_get_bitmap(cpumask_bits(new_mask),
4104 (const compat_ulong_t __user *)arg,
4105 len * 8 /* CHAR_BIT */);
4107 ret = copy_from_user(new_mask, arg, len);
4111 free_cpumask_var(new_mask);
4115 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
4116 free_cpumask_var(new_mask);
4120 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
4122 struct io_uring_task *tctx = current->io_uring;
4124 if (!tctx || !tctx->io_wq)
4127 return io_wq_cpu_affinity(tctx->io_wq, NULL);
4130 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
4132 __must_hold(&ctx->uring_lock)
4134 struct io_tctx_node *node;
4135 struct io_uring_task *tctx = NULL;
4136 struct io_sq_data *sqd = NULL;
4140 if (copy_from_user(new_count, arg, sizeof(new_count)))
4142 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4143 if (new_count[i] > INT_MAX)
4146 if (ctx->flags & IORING_SETUP_SQPOLL) {
4150 * Observe the correct sqd->lock -> ctx->uring_lock
4151 * ordering. Fine to drop uring_lock here, we hold
4154 refcount_inc(&sqd->refs);
4155 mutex_unlock(&ctx->uring_lock);
4156 mutex_lock(&sqd->lock);
4157 mutex_lock(&ctx->uring_lock);
4159 tctx = sqd->thread->io_uring;
4162 tctx = current->io_uring;
4165 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
4167 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4169 ctx->iowq_limits[i] = new_count[i];
4170 ctx->iowq_limits_set = true;
4172 if (tctx && tctx->io_wq) {
4173 ret = io_wq_max_workers(tctx->io_wq, new_count);
4177 memset(new_count, 0, sizeof(new_count));
4181 mutex_unlock(&sqd->lock);
4182 io_put_sq_data(sqd);
4185 if (copy_to_user(arg, new_count, sizeof(new_count)))
4188 /* that's it for SQPOLL, only the SQPOLL task creates requests */
4192 /* now propagate the restriction to all registered users */
4193 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
4194 struct io_uring_task *tctx = node->task->io_uring;
4196 if (WARN_ON_ONCE(!tctx->io_wq))
4199 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4200 new_count[i] = ctx->iowq_limits[i];
4201 /* ignore errors, it always returns zero anyway */
4202 (void)io_wq_max_workers(tctx->io_wq, new_count);
4207 mutex_unlock(&sqd->lock);
4208 io_put_sq_data(sqd);
4213 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
4214 void __user *arg, unsigned nr_args)
4215 __releases(ctx->uring_lock)
4216 __acquires(ctx->uring_lock)
4221 * We don't quiesce the refs for register anymore and so it can't be
4222 * dying as we're holding a file ref here.
4224 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
4227 if (ctx->submitter_task && ctx->submitter_task != current)
4230 if (ctx->restricted) {
4231 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
4232 if (!test_bit(opcode, ctx->restrictions.register_op))
4237 case IORING_REGISTER_BUFFERS:
4241 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
4243 case IORING_UNREGISTER_BUFFERS:
4247 ret = io_sqe_buffers_unregister(ctx);
4249 case IORING_REGISTER_FILES:
4253 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
4255 case IORING_UNREGISTER_FILES:
4259 ret = io_sqe_files_unregister(ctx);
4261 case IORING_REGISTER_FILES_UPDATE:
4262 ret = io_register_files_update(ctx, arg, nr_args);
4264 case IORING_REGISTER_EVENTFD:
4268 ret = io_eventfd_register(ctx, arg, 0);
4270 case IORING_REGISTER_EVENTFD_ASYNC:
4274 ret = io_eventfd_register(ctx, arg, 1);
4276 case IORING_UNREGISTER_EVENTFD:
4280 ret = io_eventfd_unregister(ctx);
4282 case IORING_REGISTER_PROBE:
4284 if (!arg || nr_args > 256)
4286 ret = io_probe(ctx, arg, nr_args);
4288 case IORING_REGISTER_PERSONALITY:
4292 ret = io_register_personality(ctx);
4294 case IORING_UNREGISTER_PERSONALITY:
4298 ret = io_unregister_personality(ctx, nr_args);
4300 case IORING_REGISTER_ENABLE_RINGS:
4304 ret = io_register_enable_rings(ctx);
4306 case IORING_REGISTER_RESTRICTIONS:
4307 ret = io_register_restrictions(ctx, arg, nr_args);
4309 case IORING_REGISTER_FILES2:
4310 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
4312 case IORING_REGISTER_FILES_UPDATE2:
4313 ret = io_register_rsrc_update(ctx, arg, nr_args,
4316 case IORING_REGISTER_BUFFERS2:
4317 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
4319 case IORING_REGISTER_BUFFERS_UPDATE:
4320 ret = io_register_rsrc_update(ctx, arg, nr_args,
4321 IORING_RSRC_BUFFER);
4323 case IORING_REGISTER_IOWQ_AFF:
4325 if (!arg || !nr_args)
4327 ret = io_register_iowq_aff(ctx, arg, nr_args);
4329 case IORING_UNREGISTER_IOWQ_AFF:
4333 ret = io_unregister_iowq_aff(ctx);
4335 case IORING_REGISTER_IOWQ_MAX_WORKERS:
4337 if (!arg || nr_args != 2)
4339 ret = io_register_iowq_max_workers(ctx, arg);
4341 case IORING_REGISTER_RING_FDS:
4342 ret = io_ringfd_register(ctx, arg, nr_args);
4344 case IORING_UNREGISTER_RING_FDS:
4345 ret = io_ringfd_unregister(ctx, arg, nr_args);
4347 case IORING_REGISTER_PBUF_RING:
4349 if (!arg || nr_args != 1)
4351 ret = io_register_pbuf_ring(ctx, arg);
4353 case IORING_UNREGISTER_PBUF_RING:
4355 if (!arg || nr_args != 1)
4357 ret = io_unregister_pbuf_ring(ctx, arg);
4359 case IORING_REGISTER_SYNC_CANCEL:
4361 if (!arg || nr_args != 1)
4363 ret = io_sync_cancel(ctx, arg);
4365 case IORING_REGISTER_FILE_ALLOC_RANGE:
4367 if (!arg || nr_args)
4369 ret = io_register_file_alloc_range(ctx, arg);
4379 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
4380 void __user *, arg, unsigned int, nr_args)
4382 struct io_ring_ctx *ctx;
4385 bool use_registered_ring;
4387 use_registered_ring = !!(opcode & IORING_REGISTER_USE_REGISTERED_RING);
4388 opcode &= ~IORING_REGISTER_USE_REGISTERED_RING;
4390 if (opcode >= IORING_REGISTER_LAST)
4393 if (use_registered_ring) {
4395 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
4396 * need only dereference our task private array to find it.
4398 struct io_uring_task *tctx = current->io_uring;
4400 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
4402 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
4403 f.file = tctx->registered_rings[fd];
4405 if (unlikely(!f.file))
4409 if (unlikely(!f.file))
4412 if (!io_is_uring_fops(f.file))
4416 ctx = f.file->private_data;
4418 mutex_lock(&ctx->uring_lock);
4419 ret = __io_uring_register(ctx, opcode, arg, nr_args);
4420 mutex_unlock(&ctx->uring_lock);
4421 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
4427 static int __init io_uring_init(void)
4429 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4430 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4431 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4434 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4435 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4436 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4437 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4438 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4439 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
4440 BUILD_BUG_SQE_ELEM(1, __u8, flags);
4441 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
4442 BUILD_BUG_SQE_ELEM(4, __s32, fd);
4443 BUILD_BUG_SQE_ELEM(8, __u64, off);
4444 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
4445 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
4446 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
4447 BUILD_BUG_SQE_ELEM(16, __u64, addr);
4448 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
4449 BUILD_BUG_SQE_ELEM(24, __u32, len);
4450 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
4451 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
4452 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4453 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
4454 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
4455 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
4456 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
4457 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
4458 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
4459 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
4460 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
4461 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
4462 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
4463 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
4464 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
4465 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
4466 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
4467 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
4468 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
4469 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
4470 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
4471 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
4472 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
4473 BUILD_BUG_SQE_ELEM(42, __u16, personality);
4474 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
4475 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
4476 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
4477 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
4478 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
4479 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4480 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
4482 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4483 sizeof(struct io_uring_rsrc_update));
4484 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4485 sizeof(struct io_uring_rsrc_update2));
4487 /* ->buf_index is u16 */
4488 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4489 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4490 offsetof(struct io_uring_buf_ring, tail));
4492 /* should fit into one byte */
4493 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
4494 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4495 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4497 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
4499 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4501 io_uring_optable_init();
4503 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4507 __initcall(io_uring_init);