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->rsrc_node_cache, IO_NODE_ALLOC_CACHE_MAX,
314 sizeof(struct io_rsrc_node));
315 io_alloc_cache_init(&ctx->apoll_cache, IO_ALLOC_CACHE_MAX,
316 sizeof(struct async_poll));
317 io_alloc_cache_init(&ctx->netmsg_cache, IO_ALLOC_CACHE_MAX,
318 sizeof(struct io_async_msghdr));
319 init_completion(&ctx->ref_comp);
320 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
321 mutex_init(&ctx->uring_lock);
322 init_waitqueue_head(&ctx->cq_wait);
323 init_waitqueue_head(&ctx->poll_wq);
324 spin_lock_init(&ctx->completion_lock);
325 spin_lock_init(&ctx->timeout_lock);
326 INIT_WQ_LIST(&ctx->iopoll_list);
327 INIT_LIST_HEAD(&ctx->io_buffers_pages);
328 INIT_LIST_HEAD(&ctx->io_buffers_comp);
329 INIT_LIST_HEAD(&ctx->defer_list);
330 INIT_LIST_HEAD(&ctx->timeout_list);
331 INIT_LIST_HEAD(&ctx->ltimeout_list);
332 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
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))
647 void io_cq_unlock_post(struct io_ring_ctx *ctx)
648 __releases(ctx->completion_lock)
650 io_commit_cqring(ctx);
651 spin_unlock(&ctx->completion_lock);
652 io_commit_cqring_flush(ctx);
656 /* Returns true if there are no backlogged entries after the flush */
657 static void io_cqring_overflow_kill(struct io_ring_ctx *ctx)
659 struct io_overflow_cqe *ocqe;
663 list_splice_init(&ctx->cq_overflow_list, &list);
664 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
667 while (!list_empty(&list)) {
668 ocqe = list_first_entry(&list, struct io_overflow_cqe, list);
669 list_del(&ocqe->list);
674 static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx)
676 size_t cqe_size = sizeof(struct io_uring_cqe);
678 if (__io_cqring_events(ctx) == ctx->cq_entries)
681 if (ctx->flags & IORING_SETUP_CQE32)
685 while (!list_empty(&ctx->cq_overflow_list)) {
686 struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true);
687 struct io_overflow_cqe *ocqe;
691 ocqe = list_first_entry(&ctx->cq_overflow_list,
692 struct io_overflow_cqe, list);
693 memcpy(cqe, &ocqe->cqe, cqe_size);
694 list_del(&ocqe->list);
698 if (list_empty(&ctx->cq_overflow_list)) {
699 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
700 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
702 io_cq_unlock_post(ctx);
705 static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx)
707 /* iopoll syncs against uring_lock, not completion_lock */
708 if (ctx->flags & IORING_SETUP_IOPOLL)
709 mutex_lock(&ctx->uring_lock);
710 __io_cqring_overflow_flush(ctx);
711 if (ctx->flags & IORING_SETUP_IOPOLL)
712 mutex_unlock(&ctx->uring_lock);
715 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx)
717 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
718 io_cqring_do_overflow_flush(ctx);
721 /* can be called by any task */
722 static void io_put_task_remote(struct task_struct *task, int nr)
724 struct io_uring_task *tctx = task->io_uring;
726 percpu_counter_sub(&tctx->inflight, nr);
727 if (unlikely(atomic_read(&tctx->in_cancel)))
728 wake_up(&tctx->wait);
729 put_task_struct_many(task, nr);
732 /* used by a task to put its own references */
733 static void io_put_task_local(struct task_struct *task, int nr)
735 task->io_uring->cached_refs += nr;
738 /* must to be called somewhat shortly after putting a request */
739 static inline void io_put_task(struct task_struct *task, int nr)
741 if (likely(task == current))
742 io_put_task_local(task, nr);
744 io_put_task_remote(task, nr);
747 void io_task_refs_refill(struct io_uring_task *tctx)
749 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
751 percpu_counter_add(&tctx->inflight, refill);
752 refcount_add(refill, ¤t->usage);
753 tctx->cached_refs += refill;
756 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
758 struct io_uring_task *tctx = task->io_uring;
759 unsigned int refs = tctx->cached_refs;
762 tctx->cached_refs = 0;
763 percpu_counter_sub(&tctx->inflight, refs);
764 put_task_struct_many(task, refs);
768 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
769 s32 res, u32 cflags, u64 extra1, u64 extra2)
771 struct io_overflow_cqe *ocqe;
772 size_t ocq_size = sizeof(struct io_overflow_cqe);
773 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
775 lockdep_assert_held(&ctx->completion_lock);
778 ocq_size += sizeof(struct io_uring_cqe);
780 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
781 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
784 * If we're in ring overflow flush mode, or in task cancel mode,
785 * or cannot allocate an overflow entry, then we need to drop it
788 io_account_cq_overflow(ctx);
789 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
792 if (list_empty(&ctx->cq_overflow_list)) {
793 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
794 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
797 ocqe->cqe.user_data = user_data;
799 ocqe->cqe.flags = cflags;
801 ocqe->cqe.big_cqe[0] = extra1;
802 ocqe->cqe.big_cqe[1] = extra2;
804 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
808 bool io_req_cqe_overflow(struct io_kiocb *req)
810 if (!(req->flags & REQ_F_CQE32_INIT)) {
814 return io_cqring_event_overflow(req->ctx, req->cqe.user_data,
815 req->cqe.res, req->cqe.flags,
816 req->extra1, req->extra2);
820 * writes to the cq entry need to come after reading head; the
821 * control dependency is enough as we're using WRITE_ONCE to
824 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow)
826 struct io_rings *rings = ctx->rings;
827 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
828 unsigned int free, queued, len;
831 * Posting into the CQ when there are pending overflowed CQEs may break
832 * ordering guarantees, which will affect links, F_MORE users and more.
833 * Force overflow the completion.
835 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
838 /* userspace may cheat modifying the tail, be safe and do min */
839 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
840 free = ctx->cq_entries - queued;
841 /* we need a contiguous range, limit based on the current array offset */
842 len = min(free, ctx->cq_entries - off);
846 if (ctx->flags & IORING_SETUP_CQE32) {
851 ctx->cqe_cached = &rings->cqes[off];
852 ctx->cqe_sentinel = ctx->cqe_cached + len;
854 ctx->cached_cq_tail++;
856 if (ctx->flags & IORING_SETUP_CQE32)
858 return &rings->cqes[off];
861 static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
864 struct io_uring_cqe *cqe;
869 * If we can't get a cq entry, userspace overflowed the
870 * submission (by quite a lot). Increment the overflow count in
873 cqe = io_get_cqe(ctx);
875 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
877 WRITE_ONCE(cqe->user_data, user_data);
878 WRITE_ONCE(cqe->res, res);
879 WRITE_ONCE(cqe->flags, cflags);
881 if (ctx->flags & IORING_SETUP_CQE32) {
882 WRITE_ONCE(cqe->big_cqe[0], 0);
883 WRITE_ONCE(cqe->big_cqe[1], 0);
890 static void __io_flush_post_cqes(struct io_ring_ctx *ctx)
891 __must_hold(&ctx->uring_lock)
893 struct io_submit_state *state = &ctx->submit_state;
896 lockdep_assert_held(&ctx->uring_lock);
897 for (i = 0; i < state->cqes_count; i++) {
898 struct io_uring_cqe *cqe = &state->cqes[i];
900 if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) {
901 if (ctx->task_complete) {
902 spin_lock(&ctx->completion_lock);
903 io_cqring_event_overflow(ctx, cqe->user_data,
904 cqe->res, cqe->flags, 0, 0);
905 spin_unlock(&ctx->completion_lock);
907 io_cqring_event_overflow(ctx, cqe->user_data,
908 cqe->res, cqe->flags, 0, 0);
912 state->cqes_count = 0;
915 static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
921 filled = io_fill_cqe_aux(ctx, user_data, res, cflags);
922 if (!filled && allow_overflow)
923 filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
925 io_cq_unlock_post(ctx);
929 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags)
931 return __io_post_aux_cqe(ctx, user_data, res, cflags, true);
934 bool io_aux_cqe(struct io_ring_ctx *ctx, bool defer, u64 user_data, s32 res, u32 cflags,
937 struct io_uring_cqe *cqe;
941 return __io_post_aux_cqe(ctx, user_data, res, cflags, allow_overflow);
943 length = ARRAY_SIZE(ctx->submit_state.cqes);
945 lockdep_assert_held(&ctx->uring_lock);
947 if (ctx->submit_state.cqes_count == length) {
949 __io_flush_post_cqes(ctx);
950 /* no need to flush - flush is deferred */
951 __io_cq_unlock_post(ctx);
954 /* For defered completions this is not as strict as it is otherwise,
955 * however it's main job is to prevent unbounded posted completions,
956 * and in that it works just as well.
958 if (!allow_overflow && test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
961 cqe = &ctx->submit_state.cqes[ctx->submit_state.cqes_count++];
962 cqe->user_data = user_data;
968 static void __io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
970 struct io_ring_ctx *ctx = req->ctx;
971 struct io_rsrc_node *rsrc_node = NULL;
974 if (!(req->flags & REQ_F_CQE_SKIP))
975 io_fill_cqe_req(ctx, req);
978 * If we're the last reference to this request, add to our locked
981 if (req_ref_put_and_test(req)) {
982 if (req->flags & IO_REQ_LINK_FLAGS) {
983 if (req->flags & IO_DISARM_MASK)
986 io_req_task_queue(req->link);
990 io_put_kbuf_comp(req);
991 io_dismantle_req(req);
992 rsrc_node = req->rsrc_node;
994 * Selected buffer deallocation in io_clean_op() assumes that
995 * we don't hold ->completion_lock. Clean them here to avoid
998 io_put_task_remote(req->task, 1);
999 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
1000 ctx->locked_free_nr++;
1002 io_cq_unlock_post(ctx);
1005 io_ring_submit_lock(ctx, issue_flags);
1006 io_put_rsrc_node(ctx, rsrc_node);
1007 io_ring_submit_unlock(ctx, issue_flags);
1011 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
1013 if (req->ctx->task_complete && (issue_flags & IO_URING_F_IOWQ)) {
1014 req->io_task_work.func = io_req_task_complete;
1015 io_req_task_work_add(req);
1016 } else if (!(issue_flags & IO_URING_F_UNLOCKED) ||
1017 !(req->ctx->flags & IORING_SETUP_IOPOLL)) {
1018 __io_req_complete_post(req, issue_flags);
1020 struct io_ring_ctx *ctx = req->ctx;
1022 mutex_lock(&ctx->uring_lock);
1023 __io_req_complete_post(req, issue_flags & ~IO_URING_F_UNLOCKED);
1024 mutex_unlock(&ctx->uring_lock);
1028 void io_req_defer_failed(struct io_kiocb *req, s32 res)
1029 __must_hold(&ctx->uring_lock)
1031 const struct io_cold_def *def = &io_cold_defs[req->opcode];
1033 lockdep_assert_held(&req->ctx->uring_lock);
1036 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
1039 io_req_complete_defer(req);
1043 * Don't initialise the fields below on every allocation, but do that in
1044 * advance and keep them valid across allocations.
1046 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1050 req->async_data = NULL;
1051 /* not necessary, but safer to zero */
1055 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1056 struct io_submit_state *state)
1058 spin_lock(&ctx->completion_lock);
1059 wq_list_splice(&ctx->locked_free_list, &state->free_list);
1060 ctx->locked_free_nr = 0;
1061 spin_unlock(&ctx->completion_lock);
1065 * A request might get retired back into the request caches even before opcode
1066 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1067 * Because of that, io_alloc_req() should be called only under ->uring_lock
1068 * and with extra caution to not get a request that is still worked on.
1070 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
1071 __must_hold(&ctx->uring_lock)
1073 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1074 void *reqs[IO_REQ_ALLOC_BATCH];
1078 * If we have more than a batch's worth of requests in our IRQ side
1079 * locked cache, grab the lock and move them over to our submission
1082 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
1083 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
1084 if (!io_req_cache_empty(ctx))
1088 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
1091 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1092 * retry single alloc to be on the safe side.
1094 if (unlikely(ret <= 0)) {
1095 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1101 percpu_ref_get_many(&ctx->refs, ret);
1102 for (i = 0; i < ret; i++) {
1103 struct io_kiocb *req = reqs[i];
1105 io_preinit_req(req, ctx);
1106 io_req_add_to_cache(req, ctx);
1111 static inline void io_dismantle_req(struct io_kiocb *req)
1113 unsigned int flags = req->flags;
1115 if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
1117 if (!(flags & REQ_F_FIXED_FILE))
1118 io_put_file(req->file);
1121 static __cold void io_free_req_tw(struct io_kiocb *req, struct io_tw_state *ts)
1123 struct io_ring_ctx *ctx = req->ctx;
1125 if (req->rsrc_node) {
1126 io_tw_lock(ctx, ts);
1127 io_put_rsrc_node(ctx, req->rsrc_node);
1129 io_dismantle_req(req);
1130 io_put_task_remote(req->task, 1);
1132 spin_lock(&ctx->completion_lock);
1133 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
1134 ctx->locked_free_nr++;
1135 spin_unlock(&ctx->completion_lock);
1138 __cold void io_free_req(struct io_kiocb *req)
1140 req->io_task_work.func = io_free_req_tw;
1141 io_req_task_work_add(req);
1144 static void __io_req_find_next_prep(struct io_kiocb *req)
1146 struct io_ring_ctx *ctx = req->ctx;
1148 spin_lock(&ctx->completion_lock);
1149 io_disarm_next(req);
1150 spin_unlock(&ctx->completion_lock);
1153 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1155 struct io_kiocb *nxt;
1158 * If LINK is set, we have dependent requests in this chain. If we
1159 * didn't fail this request, queue the first one up, moving any other
1160 * dependencies to the next request. In case of failure, fail the rest
1163 if (unlikely(req->flags & IO_DISARM_MASK))
1164 __io_req_find_next_prep(req);
1170 static void ctx_flush_and_put(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1174 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1175 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1177 io_submit_flush_completions(ctx);
1178 mutex_unlock(&ctx->uring_lock);
1181 percpu_ref_put(&ctx->refs);
1184 static unsigned int handle_tw_list(struct llist_node *node,
1185 struct io_ring_ctx **ctx,
1186 struct io_tw_state *ts,
1187 struct llist_node *last)
1189 unsigned int count = 0;
1191 while (node && node != last) {
1192 struct llist_node *next = node->next;
1193 struct io_kiocb *req = container_of(node, struct io_kiocb,
1196 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1198 if (req->ctx != *ctx) {
1199 ctx_flush_and_put(*ctx, ts);
1201 /* if not contended, grab and improve batching */
1202 ts->locked = mutex_trylock(&(*ctx)->uring_lock);
1203 percpu_ref_get(&(*ctx)->refs);
1205 req->io_task_work.func(req, ts);
1208 if (unlikely(need_resched())) {
1209 ctx_flush_and_put(*ctx, ts);
1219 * io_llist_xchg - swap all entries in a lock-less list
1220 * @head: the head of lock-less list to delete all entries
1221 * @new: new entry as the head of the list
1223 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1224 * The order of entries returned is from the newest to the oldest added one.
1226 static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1227 struct llist_node *new)
1229 return xchg(&head->first, new);
1233 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1234 * @head: the head of lock-less list to delete all entries
1235 * @old: expected old value of the first entry of the list
1236 * @new: new entry as the head of the list
1238 * perform a cmpxchg on the first entry of the list.
1241 static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1242 struct llist_node *old,
1243 struct llist_node *new)
1245 return cmpxchg(&head->first, old, new);
1248 void tctx_task_work(struct callback_head *cb)
1250 struct io_tw_state ts = {};
1251 struct io_ring_ctx *ctx = NULL;
1252 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1254 struct llist_node fake = {};
1255 struct llist_node *node;
1256 unsigned int loops = 0;
1257 unsigned int count = 0;
1259 if (unlikely(current->flags & PF_EXITING)) {
1260 io_fallback_tw(tctx);
1266 node = io_llist_xchg(&tctx->task_list, &fake);
1267 count += handle_tw_list(node, &ctx, &ts, &fake);
1269 /* skip expensive cmpxchg if there are items in the list */
1270 if (READ_ONCE(tctx->task_list.first) != &fake)
1272 if (ts.locked && !wq_list_empty(&ctx->submit_state.compl_reqs)) {
1273 io_submit_flush_completions(ctx);
1274 if (READ_ONCE(tctx->task_list.first) != &fake)
1277 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1278 } while (node != &fake);
1280 ctx_flush_and_put(ctx, &ts);
1282 /* relaxed read is enough as only the task itself sets ->in_cancel */
1283 if (unlikely(atomic_read(&tctx->in_cancel)))
1284 io_uring_drop_tctx_refs(current);
1286 trace_io_uring_task_work_run(tctx, count, loops);
1289 static __cold void io_fallback_tw(struct io_uring_task *tctx)
1291 struct llist_node *node = llist_del_all(&tctx->task_list);
1292 struct io_kiocb *req;
1295 req = container_of(node, struct io_kiocb, io_task_work.node);
1297 if (llist_add(&req->io_task_work.node,
1298 &req->ctx->fallback_llist))
1299 schedule_delayed_work(&req->ctx->fallback_work, 1);
1303 static void io_req_local_work_add(struct io_kiocb *req)
1305 struct io_ring_ctx *ctx = req->ctx;
1307 if (!llist_add(&req->io_task_work.node, &ctx->work_llist))
1310 /* needed for the following wake up */
1311 smp_mb__after_atomic();
1313 if (unlikely(atomic_read(&req->task->io_uring->in_cancel))) {
1314 io_move_task_work_from_local(ctx);
1318 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1319 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1321 io_eventfd_signal(ctx);
1323 if (READ_ONCE(ctx->cq_waiting))
1324 wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE);
1327 void __io_req_task_work_add(struct io_kiocb *req, unsigned flags)
1329 struct io_uring_task *tctx = req->task->io_uring;
1330 struct io_ring_ctx *ctx = req->ctx;
1332 if (!(flags & IOU_F_TWQ_FORCE_NORMAL) &&
1333 (ctx->flags & IORING_SETUP_DEFER_TASKRUN)) {
1335 io_req_local_work_add(req);
1340 /* task_work already pending, we're done */
1341 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1344 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1345 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1347 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1350 io_fallback_tw(tctx);
1353 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1355 struct llist_node *node;
1357 node = llist_del_all(&ctx->work_llist);
1359 struct io_kiocb *req = container_of(node, struct io_kiocb,
1363 __io_req_task_work_add(req, IOU_F_TWQ_FORCE_NORMAL);
1367 static int __io_run_local_work(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1369 struct llist_node *node;
1370 unsigned int loops = 0;
1373 if (WARN_ON_ONCE(ctx->submitter_task != current))
1375 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1376 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1378 node = io_llist_xchg(&ctx->work_llist, NULL);
1380 struct llist_node *next = node->next;
1381 struct io_kiocb *req = container_of(node, struct io_kiocb,
1383 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1384 req->io_task_work.func(req, ts);
1390 if (!llist_empty(&ctx->work_llist))
1393 io_submit_flush_completions(ctx);
1394 if (!llist_empty(&ctx->work_llist))
1397 trace_io_uring_local_work_run(ctx, ret, loops);
1401 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx)
1403 struct io_tw_state ts = { .locked = true, };
1406 if (llist_empty(&ctx->work_llist))
1409 ret = __io_run_local_work(ctx, &ts);
1410 /* shouldn't happen! */
1411 if (WARN_ON_ONCE(!ts.locked))
1412 mutex_lock(&ctx->uring_lock);
1416 static int io_run_local_work(struct io_ring_ctx *ctx)
1418 struct io_tw_state ts = {};
1421 ts.locked = mutex_trylock(&ctx->uring_lock);
1422 ret = __io_run_local_work(ctx, &ts);
1424 mutex_unlock(&ctx->uring_lock);
1429 static void io_req_task_cancel(struct io_kiocb *req, struct io_tw_state *ts)
1431 io_tw_lock(req->ctx, ts);
1432 io_req_defer_failed(req, req->cqe.res);
1435 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts)
1437 io_tw_lock(req->ctx, ts);
1438 /* req->task == current here, checking PF_EXITING is safe */
1439 if (unlikely(req->task->flags & PF_EXITING))
1440 io_req_defer_failed(req, -EFAULT);
1441 else if (req->flags & REQ_F_FORCE_ASYNC)
1442 io_queue_iowq(req, ts);
1447 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1449 io_req_set_res(req, ret, 0);
1450 req->io_task_work.func = io_req_task_cancel;
1451 io_req_task_work_add(req);
1454 void io_req_task_queue(struct io_kiocb *req)
1456 req->io_task_work.func = io_req_task_submit;
1457 io_req_task_work_add(req);
1460 void io_queue_next(struct io_kiocb *req)
1462 struct io_kiocb *nxt = io_req_find_next(req);
1465 io_req_task_queue(nxt);
1468 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1469 __must_hold(&ctx->uring_lock)
1471 struct task_struct *task = NULL;
1475 struct io_kiocb *req = container_of(node, struct io_kiocb,
1478 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1479 if (req->flags & REQ_F_REFCOUNT) {
1480 node = req->comp_list.next;
1481 if (!req_ref_put_and_test(req))
1484 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1485 struct async_poll *apoll = req->apoll;
1487 if (apoll->double_poll)
1488 kfree(apoll->double_poll);
1489 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1491 req->flags &= ~REQ_F_POLLED;
1493 if (req->flags & IO_REQ_LINK_FLAGS)
1495 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1498 if (!(req->flags & REQ_F_FIXED_FILE))
1499 io_put_file(req->file);
1501 io_req_put_rsrc_locked(req, ctx);
1503 if (req->task != task) {
1505 io_put_task(task, task_refs);
1510 node = req->comp_list.next;
1511 io_req_add_to_cache(req, ctx);
1515 io_put_task(task, task_refs);
1518 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1519 __must_hold(&ctx->uring_lock)
1521 struct io_submit_state *state = &ctx->submit_state;
1522 struct io_wq_work_node *node;
1525 /* must come first to preserve CQE ordering in failure cases */
1526 if (state->cqes_count)
1527 __io_flush_post_cqes(ctx);
1528 __wq_list_for_each(node, &state->compl_reqs) {
1529 struct io_kiocb *req = container_of(node, struct io_kiocb,
1532 if (!(req->flags & REQ_F_CQE_SKIP) &&
1533 unlikely(!__io_fill_cqe_req(ctx, req))) {
1534 if (ctx->task_complete) {
1535 spin_lock(&ctx->completion_lock);
1536 io_req_cqe_overflow(req);
1537 spin_unlock(&ctx->completion_lock);
1539 io_req_cqe_overflow(req);
1543 __io_cq_unlock_post_flush(ctx);
1545 if (!wq_list_empty(&ctx->submit_state.compl_reqs)) {
1546 io_free_batch_list(ctx, state->compl_reqs.first);
1547 INIT_WQ_LIST(&state->compl_reqs);
1552 * Drop reference to request, return next in chain (if there is one) if this
1553 * was the last reference to this request.
1555 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1557 struct io_kiocb *nxt = NULL;
1559 if (req_ref_put_and_test(req)) {
1560 if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
1561 nxt = io_req_find_next(req);
1567 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1569 /* See comment at the top of this file */
1571 return __io_cqring_events(ctx);
1575 * We can't just wait for polled events to come to us, we have to actively
1576 * find and complete them.
1578 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1580 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1583 mutex_lock(&ctx->uring_lock);
1584 while (!wq_list_empty(&ctx->iopoll_list)) {
1585 /* let it sleep and repeat later if can't complete a request */
1586 if (io_do_iopoll(ctx, true) == 0)
1589 * Ensure we allow local-to-the-cpu processing to take place,
1590 * in this case we need to ensure that we reap all events.
1591 * Also let task_work, etc. to progress by releasing the mutex
1593 if (need_resched()) {
1594 mutex_unlock(&ctx->uring_lock);
1596 mutex_lock(&ctx->uring_lock);
1599 mutex_unlock(&ctx->uring_lock);
1602 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1604 unsigned int nr_events = 0;
1606 unsigned long check_cq;
1608 if (!io_allowed_run_tw(ctx))
1611 check_cq = READ_ONCE(ctx->check_cq);
1612 if (unlikely(check_cq)) {
1613 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1614 __io_cqring_overflow_flush(ctx);
1616 * Similarly do not spin if we have not informed the user of any
1619 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1623 * Don't enter poll loop if we already have events pending.
1624 * If we do, we can potentially be spinning for commands that
1625 * already triggered a CQE (eg in error).
1627 if (io_cqring_events(ctx))
1632 * If a submit got punted to a workqueue, we can have the
1633 * application entering polling for a command before it gets
1634 * issued. That app will hold the uring_lock for the duration
1635 * of the poll right here, so we need to take a breather every
1636 * now and then to ensure that the issue has a chance to add
1637 * the poll to the issued list. Otherwise we can spin here
1638 * forever, while the workqueue is stuck trying to acquire the
1641 if (wq_list_empty(&ctx->iopoll_list) ||
1642 io_task_work_pending(ctx)) {
1643 u32 tail = ctx->cached_cq_tail;
1645 (void) io_run_local_work_locked(ctx);
1647 if (task_work_pending(current) ||
1648 wq_list_empty(&ctx->iopoll_list)) {
1649 mutex_unlock(&ctx->uring_lock);
1651 mutex_lock(&ctx->uring_lock);
1653 /* some requests don't go through iopoll_list */
1654 if (tail != ctx->cached_cq_tail ||
1655 wq_list_empty(&ctx->iopoll_list))
1658 ret = io_do_iopoll(ctx, !min);
1663 } while (nr_events < min && !need_resched());
1668 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts)
1671 io_req_complete_defer(req);
1673 io_req_complete_post(req, IO_URING_F_UNLOCKED);
1677 * After the iocb has been issued, it's safe to be found on the poll list.
1678 * Adding the kiocb to the list AFTER submission ensures that we don't
1679 * find it from a io_do_iopoll() thread before the issuer is done
1680 * accessing the kiocb cookie.
1682 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1684 struct io_ring_ctx *ctx = req->ctx;
1685 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1687 /* workqueue context doesn't hold uring_lock, grab it now */
1688 if (unlikely(needs_lock))
1689 mutex_lock(&ctx->uring_lock);
1692 * Track whether we have multiple files in our lists. This will impact
1693 * how we do polling eventually, not spinning if we're on potentially
1694 * different devices.
1696 if (wq_list_empty(&ctx->iopoll_list)) {
1697 ctx->poll_multi_queue = false;
1698 } else if (!ctx->poll_multi_queue) {
1699 struct io_kiocb *list_req;
1701 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1703 if (list_req->file != req->file)
1704 ctx->poll_multi_queue = true;
1708 * For fast devices, IO may have already completed. If it has, add
1709 * it to the front so we find it first.
1711 if (READ_ONCE(req->iopoll_completed))
1712 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1714 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1716 if (unlikely(needs_lock)) {
1718 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1719 * in sq thread task context or in io worker task context. If
1720 * current task context is sq thread, we don't need to check
1721 * whether should wake up sq thread.
1723 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1724 wq_has_sleeper(&ctx->sq_data->wait))
1725 wake_up(&ctx->sq_data->wait);
1727 mutex_unlock(&ctx->uring_lock);
1731 static bool io_bdev_nowait(struct block_device *bdev)
1733 return !bdev || bdev_nowait(bdev);
1737 * If we tracked the file through the SCM inflight mechanism, we could support
1738 * any file. For now, just ensure that anything potentially problematic is done
1741 static bool __io_file_supports_nowait(struct file *file, umode_t mode)
1743 if (S_ISBLK(mode)) {
1744 if (IS_ENABLED(CONFIG_BLOCK) &&
1745 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
1751 if (S_ISREG(mode)) {
1752 if (IS_ENABLED(CONFIG_BLOCK) &&
1753 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
1754 !io_is_uring_fops(file))
1759 /* any ->read/write should understand O_NONBLOCK */
1760 if (file->f_flags & O_NONBLOCK)
1762 return file->f_mode & FMODE_NOWAIT;
1766 * If we tracked the file through the SCM inflight mechanism, we could support
1767 * any file. For now, just ensure that anything potentially problematic is done
1770 unsigned int io_file_get_flags(struct file *file)
1772 umode_t mode = file_inode(file)->i_mode;
1773 unsigned int res = 0;
1777 if (__io_file_supports_nowait(file, mode))
1782 bool io_alloc_async_data(struct io_kiocb *req)
1784 WARN_ON_ONCE(!io_cold_defs[req->opcode].async_size);
1785 req->async_data = kmalloc(io_cold_defs[req->opcode].async_size, GFP_KERNEL);
1786 if (req->async_data) {
1787 req->flags |= REQ_F_ASYNC_DATA;
1793 int io_req_prep_async(struct io_kiocb *req)
1795 const struct io_cold_def *cdef = &io_cold_defs[req->opcode];
1796 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1798 /* assign early for deferred execution for non-fixed file */
1799 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE) && !req->file)
1800 req->file = io_file_get_normal(req, req->cqe.fd);
1801 if (!cdef->prep_async)
1803 if (WARN_ON_ONCE(req_has_async_data(req)))
1805 if (!def->manual_alloc) {
1806 if (io_alloc_async_data(req))
1809 return cdef->prep_async(req);
1812 static u32 io_get_sequence(struct io_kiocb *req)
1814 u32 seq = req->ctx->cached_sq_head;
1815 struct io_kiocb *cur;
1817 /* need original cached_sq_head, but it was increased for each req */
1818 io_for_each_link(cur, req)
1823 static __cold void io_drain_req(struct io_kiocb *req)
1824 __must_hold(&ctx->uring_lock)
1826 struct io_ring_ctx *ctx = req->ctx;
1827 struct io_defer_entry *de;
1829 u32 seq = io_get_sequence(req);
1831 /* Still need defer if there is pending req in defer list. */
1832 spin_lock(&ctx->completion_lock);
1833 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1834 spin_unlock(&ctx->completion_lock);
1836 ctx->drain_active = false;
1837 io_req_task_queue(req);
1840 spin_unlock(&ctx->completion_lock);
1842 io_prep_async_link(req);
1843 de = kmalloc(sizeof(*de), GFP_KERNEL);
1846 io_req_defer_failed(req, ret);
1850 spin_lock(&ctx->completion_lock);
1851 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1852 spin_unlock(&ctx->completion_lock);
1857 trace_io_uring_defer(req);
1860 list_add_tail(&de->list, &ctx->defer_list);
1861 spin_unlock(&ctx->completion_lock);
1864 static void io_clean_op(struct io_kiocb *req)
1866 if (req->flags & REQ_F_BUFFER_SELECTED) {
1867 spin_lock(&req->ctx->completion_lock);
1868 io_put_kbuf_comp(req);
1869 spin_unlock(&req->ctx->completion_lock);
1872 if (req->flags & REQ_F_NEED_CLEANUP) {
1873 const struct io_cold_def *def = &io_cold_defs[req->opcode];
1878 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1879 kfree(req->apoll->double_poll);
1883 if (req->flags & REQ_F_INFLIGHT) {
1884 struct io_uring_task *tctx = req->task->io_uring;
1886 atomic_dec(&tctx->inflight_tracked);
1888 if (req->flags & REQ_F_CREDS)
1889 put_cred(req->creds);
1890 if (req->flags & REQ_F_ASYNC_DATA) {
1891 kfree(req->async_data);
1892 req->async_data = NULL;
1894 req->flags &= ~IO_REQ_CLEAN_FLAGS;
1897 static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def,
1898 unsigned int issue_flags)
1900 if (req->file || !def->needs_file)
1903 if (req->flags & REQ_F_FIXED_FILE)
1904 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1906 req->file = io_file_get_normal(req, req->cqe.fd);
1911 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1913 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1914 const struct cred *creds = NULL;
1917 if (unlikely(!io_assign_file(req, def, issue_flags)))
1920 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1921 creds = override_creds(req->creds);
1923 if (!def->audit_skip)
1924 audit_uring_entry(req->opcode);
1926 ret = def->issue(req, issue_flags);
1928 if (!def->audit_skip)
1929 audit_uring_exit(!ret, ret);
1932 revert_creds(creds);
1934 if (ret == IOU_OK) {
1935 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1936 io_req_complete_defer(req);
1938 io_req_complete_post(req, issue_flags);
1939 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1942 /* If the op doesn't have a file, we're not polling for it */
1943 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
1944 io_iopoll_req_issued(req, issue_flags);
1949 int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts)
1951 io_tw_lock(req->ctx, ts);
1952 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT|
1953 IO_URING_F_COMPLETE_DEFER);
1956 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1958 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1960 req = io_put_req_find_next(req);
1961 return req ? &req->work : NULL;
1964 void io_wq_submit_work(struct io_wq_work *work)
1966 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1967 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1968 unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ;
1969 bool needs_poll = false;
1970 int ret = 0, err = -ECANCELED;
1972 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1973 if (!(req->flags & REQ_F_REFCOUNT))
1974 __io_req_set_refcount(req, 2);
1978 io_arm_ltimeout(req);
1980 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1981 if (work->flags & IO_WQ_WORK_CANCEL) {
1983 io_req_task_queue_fail(req, err);
1986 if (!io_assign_file(req, def, issue_flags)) {
1988 work->flags |= IO_WQ_WORK_CANCEL;
1992 if (req->flags & REQ_F_FORCE_ASYNC) {
1993 bool opcode_poll = def->pollin || def->pollout;
1995 if (opcode_poll && file_can_poll(req->file)) {
1997 issue_flags |= IO_URING_F_NONBLOCK;
2002 ret = io_issue_sqe(req, issue_flags);
2006 * We can get EAGAIN for iopolled IO even though we're
2007 * forcing a sync submission from here, since we can't
2008 * wait for request slots on the block side.
2011 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
2017 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
2019 /* aborted or ready, in either case retry blocking */
2021 issue_flags &= ~IO_URING_F_NONBLOCK;
2024 /* avoid locking problems by failing it from a clean context */
2026 io_req_task_queue_fail(req, ret);
2029 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
2030 unsigned int issue_flags)
2032 struct io_ring_ctx *ctx = req->ctx;
2033 struct file *file = NULL;
2034 unsigned long file_ptr;
2036 io_ring_submit_lock(ctx, issue_flags);
2038 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
2040 fd = array_index_nospec(fd, ctx->nr_user_files);
2041 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
2042 file = (struct file *) (file_ptr & FFS_MASK);
2043 file_ptr &= ~FFS_MASK;
2044 /* mask in overlapping REQ_F and FFS bits */
2045 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
2046 io_req_set_rsrc_node(req, ctx, 0);
2048 io_ring_submit_unlock(ctx, issue_flags);
2052 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
2054 struct file *file = fget(fd);
2056 trace_io_uring_file_get(req, fd);
2058 /* we don't allow fixed io_uring files */
2059 if (file && io_is_uring_fops(file))
2060 io_req_track_inflight(req);
2064 static void io_queue_async(struct io_kiocb *req, int ret)
2065 __must_hold(&req->ctx->uring_lock)
2067 struct io_kiocb *linked_timeout;
2069 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
2070 io_req_defer_failed(req, ret);
2074 linked_timeout = io_prep_linked_timeout(req);
2076 switch (io_arm_poll_handler(req, 0)) {
2077 case IO_APOLL_READY:
2078 io_kbuf_recycle(req, 0);
2079 io_req_task_queue(req);
2081 case IO_APOLL_ABORTED:
2082 io_kbuf_recycle(req, 0);
2083 io_queue_iowq(req, NULL);
2090 io_queue_linked_timeout(linked_timeout);
2093 static inline void io_queue_sqe(struct io_kiocb *req)
2094 __must_hold(&req->ctx->uring_lock)
2098 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
2101 * We async punt it if the file wasn't marked NOWAIT, or if the file
2102 * doesn't support non-blocking read/write attempts
2105 io_arm_ltimeout(req);
2107 io_queue_async(req, ret);
2110 static void io_queue_sqe_fallback(struct io_kiocb *req)
2111 __must_hold(&req->ctx->uring_lock)
2113 if (unlikely(req->flags & REQ_F_FAIL)) {
2115 * We don't submit, fail them all, for that replace hardlinks
2116 * with normal links. Extra REQ_F_LINK is tolerated.
2118 req->flags &= ~REQ_F_HARDLINK;
2119 req->flags |= REQ_F_LINK;
2120 io_req_defer_failed(req, req->cqe.res);
2122 int ret = io_req_prep_async(req);
2124 if (unlikely(ret)) {
2125 io_req_defer_failed(req, ret);
2129 if (unlikely(req->ctx->drain_active))
2132 io_queue_iowq(req, NULL);
2137 * Check SQE restrictions (opcode and flags).
2139 * Returns 'true' if SQE is allowed, 'false' otherwise.
2141 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
2142 struct io_kiocb *req,
2143 unsigned int sqe_flags)
2145 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
2148 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
2149 ctx->restrictions.sqe_flags_required)
2152 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
2153 ctx->restrictions.sqe_flags_required))
2159 static void io_init_req_drain(struct io_kiocb *req)
2161 struct io_ring_ctx *ctx = req->ctx;
2162 struct io_kiocb *head = ctx->submit_state.link.head;
2164 ctx->drain_active = true;
2167 * If we need to drain a request in the middle of a link, drain
2168 * the head request and the next request/link after the current
2169 * link. Considering sequential execution of links,
2170 * REQ_F_IO_DRAIN will be maintained for every request of our
2173 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2174 ctx->drain_next = true;
2178 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
2179 const struct io_uring_sqe *sqe)
2180 __must_hold(&ctx->uring_lock)
2182 const struct io_issue_def *def;
2183 unsigned int sqe_flags;
2187 /* req is partially pre-initialised, see io_preinit_req() */
2188 req->opcode = opcode = READ_ONCE(sqe->opcode);
2189 /* same numerical values with corresponding REQ_F_*, safe to copy */
2190 req->flags = sqe_flags = READ_ONCE(sqe->flags);
2191 req->cqe.user_data = READ_ONCE(sqe->user_data);
2193 req->rsrc_node = NULL;
2194 req->task = current;
2196 if (unlikely(opcode >= IORING_OP_LAST)) {
2200 def = &io_issue_defs[opcode];
2201 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2202 /* enforce forwards compatibility on users */
2203 if (sqe_flags & ~SQE_VALID_FLAGS)
2205 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2206 if (!def->buffer_select)
2208 req->buf_index = READ_ONCE(sqe->buf_group);
2210 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2211 ctx->drain_disabled = true;
2212 if (sqe_flags & IOSQE_IO_DRAIN) {
2213 if (ctx->drain_disabled)
2215 io_init_req_drain(req);
2218 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2219 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2221 /* knock it to the slow queue path, will be drained there */
2222 if (ctx->drain_active)
2223 req->flags |= REQ_F_FORCE_ASYNC;
2224 /* if there is no link, we're at "next" request and need to drain */
2225 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2226 ctx->drain_next = false;
2227 ctx->drain_active = true;
2228 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2232 if (!def->ioprio && sqe->ioprio)
2234 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2237 if (def->needs_file) {
2238 struct io_submit_state *state = &ctx->submit_state;
2240 req->cqe.fd = READ_ONCE(sqe->fd);
2243 * Plug now if we have more than 2 IO left after this, and the
2244 * target is potentially a read/write to block based storage.
2246 if (state->need_plug && def->plug) {
2247 state->plug_started = true;
2248 state->need_plug = false;
2249 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2253 personality = READ_ONCE(sqe->personality);
2257 req->creds = xa_load(&ctx->personalities, personality);
2260 get_cred(req->creds);
2261 ret = security_uring_override_creds(req->creds);
2263 put_cred(req->creds);
2266 req->flags |= REQ_F_CREDS;
2269 return def->prep(req, sqe);
2272 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2273 struct io_kiocb *req, int ret)
2275 struct io_ring_ctx *ctx = req->ctx;
2276 struct io_submit_link *link = &ctx->submit_state.link;
2277 struct io_kiocb *head = link->head;
2279 trace_io_uring_req_failed(sqe, req, ret);
2282 * Avoid breaking links in the middle as it renders links with SQPOLL
2283 * unusable. Instead of failing eagerly, continue assembling the link if
2284 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2285 * should find the flag and handle the rest.
2287 req_fail_link_node(req, ret);
2288 if (head && !(head->flags & REQ_F_FAIL))
2289 req_fail_link_node(head, -ECANCELED);
2291 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2293 link->last->link = req;
2297 io_queue_sqe_fallback(req);
2302 link->last->link = req;
2309 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2310 const struct io_uring_sqe *sqe)
2311 __must_hold(&ctx->uring_lock)
2313 struct io_submit_link *link = &ctx->submit_state.link;
2316 ret = io_init_req(ctx, req, sqe);
2318 return io_submit_fail_init(sqe, req, ret);
2320 trace_io_uring_submit_req(req);
2323 * If we already have a head request, queue this one for async
2324 * submittal once the head completes. If we don't have a head but
2325 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2326 * submitted sync once the chain is complete. If none of those
2327 * conditions are true (normal request), then just queue it.
2329 if (unlikely(link->head)) {
2330 ret = io_req_prep_async(req);
2332 return io_submit_fail_init(sqe, req, ret);
2334 trace_io_uring_link(req, link->head);
2335 link->last->link = req;
2338 if (req->flags & IO_REQ_LINK_FLAGS)
2340 /* last request of the link, flush it */
2343 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2346 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2347 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2348 if (req->flags & IO_REQ_LINK_FLAGS) {
2353 io_queue_sqe_fallback(req);
2363 * Batched submission is done, ensure local IO is flushed out.
2365 static void io_submit_state_end(struct io_ring_ctx *ctx)
2367 struct io_submit_state *state = &ctx->submit_state;
2369 if (unlikely(state->link.head))
2370 io_queue_sqe_fallback(state->link.head);
2371 /* flush only after queuing links as they can generate completions */
2372 io_submit_flush_completions(ctx);
2373 if (state->plug_started)
2374 blk_finish_plug(&state->plug);
2378 * Start submission side cache.
2380 static void io_submit_state_start(struct io_submit_state *state,
2381 unsigned int max_ios)
2383 state->plug_started = false;
2384 state->need_plug = max_ios > 2;
2385 state->submit_nr = max_ios;
2386 /* set only head, no need to init link_last in advance */
2387 state->link.head = NULL;
2390 static void io_commit_sqring(struct io_ring_ctx *ctx)
2392 struct io_rings *rings = ctx->rings;
2395 * Ensure any loads from the SQEs are done at this point,
2396 * since once we write the new head, the application could
2397 * write new data to them.
2399 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2403 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2404 * that is mapped by userspace. This means that care needs to be taken to
2405 * ensure that reads are stable, as we cannot rely on userspace always
2406 * being a good citizen. If members of the sqe are validated and then later
2407 * used, it's important that those reads are done through READ_ONCE() to
2408 * prevent a re-load down the line.
2410 static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe)
2412 unsigned head, mask = ctx->sq_entries - 1;
2413 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2416 * The cached sq head (or cq tail) serves two purposes:
2418 * 1) allows us to batch the cost of updating the user visible
2420 * 2) allows the kernel side to track the head on its own, even
2421 * though the application is the one updating it.
2423 head = READ_ONCE(ctx->sq_array[sq_idx]);
2424 if (likely(head < ctx->sq_entries)) {
2425 /* double index for 128-byte SQEs, twice as long */
2426 if (ctx->flags & IORING_SETUP_SQE128)
2428 *sqe = &ctx->sq_sqes[head];
2432 /* drop invalid entries */
2434 WRITE_ONCE(ctx->rings->sq_dropped,
2435 READ_ONCE(ctx->rings->sq_dropped) + 1);
2439 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2440 __must_hold(&ctx->uring_lock)
2442 unsigned int entries = io_sqring_entries(ctx);
2446 if (unlikely(!entries))
2448 /* make sure SQ entry isn't read before tail */
2449 ret = left = min(nr, entries);
2450 io_get_task_refs(left);
2451 io_submit_state_start(&ctx->submit_state, left);
2454 const struct io_uring_sqe *sqe;
2455 struct io_kiocb *req;
2457 if (unlikely(!io_alloc_req(ctx, &req)))
2459 if (unlikely(!io_get_sqe(ctx, &sqe))) {
2460 io_req_add_to_cache(req, ctx);
2465 * Continue submitting even for sqe failure if the
2466 * ring was setup with IORING_SETUP_SUBMIT_ALL
2468 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2469 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2475 if (unlikely(left)) {
2477 /* try again if it submitted nothing and can't allocate a req */
2478 if (!ret && io_req_cache_empty(ctx))
2480 current->io_uring->cached_refs += left;
2483 io_submit_state_end(ctx);
2484 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2485 io_commit_sqring(ctx);
2489 struct io_wait_queue {
2490 struct wait_queue_entry wq;
2491 struct io_ring_ctx *ctx;
2493 unsigned nr_timeouts;
2497 static inline bool io_has_work(struct io_ring_ctx *ctx)
2499 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
2500 !llist_empty(&ctx->work_llist);
2503 static inline bool io_should_wake(struct io_wait_queue *iowq)
2505 struct io_ring_ctx *ctx = iowq->ctx;
2506 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
2509 * Wake up if we have enough events, or if a timeout occurred since we
2510 * started waiting. For timeouts, we always want to return to userspace,
2511 * regardless of event count.
2513 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2516 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2517 int wake_flags, void *key)
2519 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq);
2522 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2523 * the task, and the next invocation will do it.
2525 if (io_should_wake(iowq) || io_has_work(iowq->ctx))
2526 return autoremove_wake_function(curr, mode, wake_flags, key);
2530 int io_run_task_work_sig(struct io_ring_ctx *ctx)
2532 if (!llist_empty(&ctx->work_llist)) {
2533 __set_current_state(TASK_RUNNING);
2534 if (io_run_local_work(ctx) > 0)
2537 if (io_run_task_work() > 0)
2539 if (task_sigpending(current))
2544 /* when returns >0, the caller should retry */
2545 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2546 struct io_wait_queue *iowq)
2548 if (unlikely(READ_ONCE(ctx->check_cq)))
2550 if (unlikely(!llist_empty(&ctx->work_llist)))
2552 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL)))
2554 if (unlikely(task_sigpending(current)))
2556 if (unlikely(io_should_wake(iowq)))
2558 if (iowq->timeout == KTIME_MAX)
2560 else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS))
2566 * Wait until events become available, if we don't already have some. The
2567 * application must reap them itself, as they reside on the shared cq ring.
2569 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2570 const sigset_t __user *sig, size_t sigsz,
2571 struct __kernel_timespec __user *uts)
2573 struct io_wait_queue iowq;
2574 struct io_rings *rings = ctx->rings;
2577 if (!io_allowed_run_tw(ctx))
2579 if (!llist_empty(&ctx->work_llist))
2580 io_run_local_work(ctx);
2582 io_cqring_overflow_flush(ctx);
2583 /* if user messes with these they will just get an early return */
2584 if (__io_cqring_events_user(ctx) >= min_events)
2588 #ifdef CONFIG_COMPAT
2589 if (in_compat_syscall())
2590 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2594 ret = set_user_sigmask(sig, sigsz);
2600 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2601 iowq.wq.private = current;
2602 INIT_LIST_HEAD(&iowq.wq.entry);
2604 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2605 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2606 iowq.timeout = KTIME_MAX;
2609 struct timespec64 ts;
2611 if (get_timespec64(&ts, uts))
2613 iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2616 trace_io_uring_cqring_wait(ctx, min_events);
2618 unsigned long check_cq;
2620 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
2621 WRITE_ONCE(ctx->cq_waiting, 1);
2622 set_current_state(TASK_INTERRUPTIBLE);
2624 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2625 TASK_INTERRUPTIBLE);
2628 ret = io_cqring_wait_schedule(ctx, &iowq);
2629 __set_current_state(TASK_RUNNING);
2630 WRITE_ONCE(ctx->cq_waiting, 0);
2635 * Run task_work after scheduling and before io_should_wake().
2636 * If we got woken because of task_work being processed, run it
2637 * now rather than let the caller do another wait loop.
2640 if (!llist_empty(&ctx->work_llist))
2641 io_run_local_work(ctx);
2643 check_cq = READ_ONCE(ctx->check_cq);
2644 if (unlikely(check_cq)) {
2645 /* let the caller flush overflows, retry */
2646 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2647 io_cqring_do_overflow_flush(ctx);
2648 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) {
2654 if (io_should_wake(&iowq)) {
2661 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
2662 finish_wait(&ctx->cq_wait, &iowq.wq);
2663 restore_saved_sigmask_unless(ret == -EINTR);
2665 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2668 static void io_mem_free(void *ptr)
2675 page = virt_to_head_page(ptr);
2676 if (put_page_testzero(page))
2677 free_compound_page(page);
2680 static void *io_mem_alloc(size_t size)
2682 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2684 return (void *) __get_free_pages(gfp, get_order(size));
2687 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2688 unsigned int cq_entries, size_t *sq_offset)
2690 struct io_rings *rings;
2691 size_t off, sq_array_size;
2693 off = struct_size(rings, cqes, cq_entries);
2694 if (off == SIZE_MAX)
2696 if (ctx->flags & IORING_SETUP_CQE32) {
2697 if (check_shl_overflow(off, 1, &off))
2702 off = ALIGN(off, SMP_CACHE_BYTES);
2710 sq_array_size = array_size(sizeof(u32), sq_entries);
2711 if (sq_array_size == SIZE_MAX)
2714 if (check_add_overflow(off, sq_array_size, &off))
2720 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2721 unsigned int eventfd_async)
2723 struct io_ev_fd *ev_fd;
2724 __s32 __user *fds = arg;
2727 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2728 lockdep_is_held(&ctx->uring_lock));
2732 if (copy_from_user(&fd, fds, sizeof(*fds)))
2735 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2739 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2740 if (IS_ERR(ev_fd->cq_ev_fd)) {
2741 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2746 spin_lock(&ctx->completion_lock);
2747 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2748 spin_unlock(&ctx->completion_lock);
2750 ev_fd->eventfd_async = eventfd_async;
2751 ctx->has_evfd = true;
2752 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2753 atomic_set(&ev_fd->refs, 1);
2754 atomic_set(&ev_fd->ops, 0);
2758 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2760 struct io_ev_fd *ev_fd;
2762 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2763 lockdep_is_held(&ctx->uring_lock));
2765 ctx->has_evfd = false;
2766 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2767 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops))
2768 call_rcu(&ev_fd->rcu, io_eventfd_ops);
2775 static void io_req_caches_free(struct io_ring_ctx *ctx)
2777 struct io_kiocb *req;
2780 mutex_lock(&ctx->uring_lock);
2781 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
2783 while (!io_req_cache_empty(ctx)) {
2784 req = io_extract_req(ctx);
2785 kmem_cache_free(req_cachep, req);
2789 percpu_ref_put_many(&ctx->refs, nr);
2790 mutex_unlock(&ctx->uring_lock);
2793 static void io_rsrc_node_cache_free(struct io_cache_entry *entry)
2795 kfree(container_of(entry, struct io_rsrc_node, cache));
2798 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2800 io_sq_thread_finish(ctx);
2801 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2802 io_wait_rsrc_data(ctx->buf_data);
2803 io_wait_rsrc_data(ctx->file_data);
2805 mutex_lock(&ctx->uring_lock);
2807 __io_sqe_buffers_unregister(ctx);
2809 __io_sqe_files_unregister(ctx);
2810 io_cqring_overflow_kill(ctx);
2811 io_eventfd_unregister(ctx);
2812 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2813 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2814 mutex_unlock(&ctx->uring_lock);
2815 io_destroy_buffers(ctx);
2817 put_cred(ctx->sq_creds);
2818 if (ctx->submitter_task)
2819 put_task_struct(ctx->submitter_task);
2821 /* there are no registered resources left, nobody uses it */
2823 io_rsrc_node_destroy(ctx, ctx->rsrc_node);
2824 if (ctx->rsrc_backup_node)
2825 io_rsrc_node_destroy(ctx, ctx->rsrc_backup_node);
2827 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2829 #if defined(CONFIG_UNIX)
2830 if (ctx->ring_sock) {
2831 ctx->ring_sock->file = NULL; /* so that iput() is called */
2832 sock_release(ctx->ring_sock);
2835 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2837 io_alloc_cache_free(&ctx->rsrc_node_cache, io_rsrc_node_cache_free);
2838 if (ctx->mm_account) {
2839 mmdrop(ctx->mm_account);
2840 ctx->mm_account = NULL;
2842 io_mem_free(ctx->rings);
2843 io_mem_free(ctx->sq_sqes);
2845 percpu_ref_exit(&ctx->refs);
2846 free_uid(ctx->user);
2847 io_req_caches_free(ctx);
2849 io_wq_put_hash(ctx->hash_map);
2850 kfree(ctx->cancel_table.hbs);
2851 kfree(ctx->cancel_table_locked.hbs);
2852 kfree(ctx->dummy_ubuf);
2854 xa_destroy(&ctx->io_bl_xa);
2858 static __cold void io_activate_pollwq_cb(struct callback_head *cb)
2860 struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx,
2863 mutex_lock(&ctx->uring_lock);
2864 ctx->poll_activated = true;
2865 mutex_unlock(&ctx->uring_lock);
2868 * Wake ups for some events between start of polling and activation
2869 * might've been lost due to loose synchronisation.
2871 wake_up_all(&ctx->poll_wq);
2872 percpu_ref_put(&ctx->refs);
2875 static __cold void io_activate_pollwq(struct io_ring_ctx *ctx)
2877 spin_lock(&ctx->completion_lock);
2878 /* already activated or in progress */
2879 if (ctx->poll_activated || ctx->poll_wq_task_work.func)
2881 if (WARN_ON_ONCE(!ctx->task_complete))
2883 if (!ctx->submitter_task)
2886 * with ->submitter_task only the submitter task completes requests, we
2887 * only need to sync with it, which is done by injecting a tw
2889 init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb);
2890 percpu_ref_get(&ctx->refs);
2891 if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL))
2892 percpu_ref_put(&ctx->refs);
2894 spin_unlock(&ctx->completion_lock);
2897 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2899 struct io_ring_ctx *ctx = file->private_data;
2902 if (unlikely(!ctx->poll_activated))
2903 io_activate_pollwq(ctx);
2905 poll_wait(file, &ctx->poll_wq, wait);
2907 * synchronizes with barrier from wq_has_sleeper call in
2911 if (!io_sqring_full(ctx))
2912 mask |= EPOLLOUT | EPOLLWRNORM;
2915 * Don't flush cqring overflow list here, just do a simple check.
2916 * Otherwise there could possible be ABBA deadlock:
2919 * lock(&ctx->uring_lock);
2921 * lock(&ctx->uring_lock);
2924 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2925 * pushes them to do the flush.
2928 if (__io_cqring_events_user(ctx) || io_has_work(ctx))
2929 mask |= EPOLLIN | EPOLLRDNORM;
2934 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2936 const struct cred *creds;
2938 creds = xa_erase(&ctx->personalities, id);
2947 struct io_tctx_exit {
2948 struct callback_head task_work;
2949 struct completion completion;
2950 struct io_ring_ctx *ctx;
2953 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2955 struct io_uring_task *tctx = current->io_uring;
2956 struct io_tctx_exit *work;
2958 work = container_of(cb, struct io_tctx_exit, task_work);
2960 * When @in_cancel, we're in cancellation and it's racy to remove the
2961 * node. It'll be removed by the end of cancellation, just ignore it.
2962 * tctx can be NULL if the queueing of this task_work raced with
2963 * work cancelation off the exec path.
2965 if (tctx && !atomic_read(&tctx->in_cancel))
2966 io_uring_del_tctx_node((unsigned long)work->ctx);
2967 complete(&work->completion);
2970 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2972 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2974 return req->ctx == data;
2977 static __cold void io_ring_exit_work(struct work_struct *work)
2979 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2980 unsigned long timeout = jiffies + HZ * 60 * 5;
2981 unsigned long interval = HZ / 20;
2982 struct io_tctx_exit exit;
2983 struct io_tctx_node *node;
2987 * If we're doing polled IO and end up having requests being
2988 * submitted async (out-of-line), then completions can come in while
2989 * we're waiting for refs to drop. We need to reap these manually,
2990 * as nobody else will be looking for them.
2993 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
2994 mutex_lock(&ctx->uring_lock);
2995 io_cqring_overflow_kill(ctx);
2996 mutex_unlock(&ctx->uring_lock);
2999 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3000 io_move_task_work_from_local(ctx);
3002 while (io_uring_try_cancel_requests(ctx, NULL, true))
3006 struct io_sq_data *sqd = ctx->sq_data;
3007 struct task_struct *tsk;
3009 io_sq_thread_park(sqd);
3011 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
3012 io_wq_cancel_cb(tsk->io_uring->io_wq,
3013 io_cancel_ctx_cb, ctx, true);
3014 io_sq_thread_unpark(sqd);
3017 io_req_caches_free(ctx);
3019 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
3020 /* there is little hope left, don't run it too often */
3023 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
3025 init_completion(&exit.completion);
3026 init_task_work(&exit.task_work, io_tctx_exit_cb);
3029 * Some may use context even when all refs and requests have been put,
3030 * and they are free to do so while still holding uring_lock or
3031 * completion_lock, see io_req_task_submit(). Apart from other work,
3032 * this lock/unlock section also waits them to finish.
3034 mutex_lock(&ctx->uring_lock);
3035 while (!list_empty(&ctx->tctx_list)) {
3036 WARN_ON_ONCE(time_after(jiffies, timeout));
3038 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
3040 /* don't spin on a single task if cancellation failed */
3041 list_rotate_left(&ctx->tctx_list);
3042 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
3043 if (WARN_ON_ONCE(ret))
3046 mutex_unlock(&ctx->uring_lock);
3047 wait_for_completion(&exit.completion);
3048 mutex_lock(&ctx->uring_lock);
3050 mutex_unlock(&ctx->uring_lock);
3051 spin_lock(&ctx->completion_lock);
3052 spin_unlock(&ctx->completion_lock);
3054 /* pairs with RCU read section in io_req_local_work_add() */
3055 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3058 io_ring_ctx_free(ctx);
3061 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
3063 unsigned long index;
3064 struct creds *creds;
3066 mutex_lock(&ctx->uring_lock);
3067 percpu_ref_kill(&ctx->refs);
3068 xa_for_each(&ctx->personalities, index, creds)
3069 io_unregister_personality(ctx, index);
3071 io_poll_remove_all(ctx, NULL, true);
3072 mutex_unlock(&ctx->uring_lock);
3075 * If we failed setting up the ctx, we might not have any rings
3076 * and therefore did not submit any requests
3079 io_kill_timeouts(ctx, NULL, true);
3081 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
3083 * Use system_unbound_wq to avoid spawning tons of event kworkers
3084 * if we're exiting a ton of rings at the same time. It just adds
3085 * noise and overhead, there's no discernable change in runtime
3086 * over using system_wq.
3088 queue_work(system_unbound_wq, &ctx->exit_work);
3091 static int io_uring_release(struct inode *inode, struct file *file)
3093 struct io_ring_ctx *ctx = file->private_data;
3095 file->private_data = NULL;
3096 io_ring_ctx_wait_and_kill(ctx);
3100 struct io_task_cancel {
3101 struct task_struct *task;
3105 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
3107 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3108 struct io_task_cancel *cancel = data;
3110 return io_match_task_safe(req, cancel->task, cancel->all);
3113 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
3114 struct task_struct *task,
3117 struct io_defer_entry *de;
3120 spin_lock(&ctx->completion_lock);
3121 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
3122 if (io_match_task_safe(de->req, task, cancel_all)) {
3123 list_cut_position(&list, &ctx->defer_list, &de->list);
3127 spin_unlock(&ctx->completion_lock);
3128 if (list_empty(&list))
3131 while (!list_empty(&list)) {
3132 de = list_first_entry(&list, struct io_defer_entry, list);
3133 list_del_init(&de->list);
3134 io_req_task_queue_fail(de->req, -ECANCELED);
3140 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
3142 struct io_tctx_node *node;
3143 enum io_wq_cancel cret;
3146 mutex_lock(&ctx->uring_lock);
3147 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3148 struct io_uring_task *tctx = node->task->io_uring;
3151 * io_wq will stay alive while we hold uring_lock, because it's
3152 * killed after ctx nodes, which requires to take the lock.
3154 if (!tctx || !tctx->io_wq)
3156 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
3157 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3159 mutex_unlock(&ctx->uring_lock);
3164 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
3165 struct task_struct *task,
3168 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
3169 struct io_uring_task *tctx = task ? task->io_uring : NULL;
3170 enum io_wq_cancel cret;
3173 /* failed during ring init, it couldn't have issued any requests */
3178 ret |= io_uring_try_cancel_iowq(ctx);
3179 } else if (tctx && tctx->io_wq) {
3181 * Cancels requests of all rings, not only @ctx, but
3182 * it's fine as the task is in exit/exec.
3184 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
3186 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3189 /* SQPOLL thread does its own polling */
3190 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
3191 (ctx->sq_data && ctx->sq_data->thread == current)) {
3192 while (!wq_list_empty(&ctx->iopoll_list)) {
3193 io_iopoll_try_reap_events(ctx);
3199 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3200 io_allowed_defer_tw_run(ctx))
3201 ret |= io_run_local_work(ctx) > 0;
3202 ret |= io_cancel_defer_files(ctx, task, cancel_all);
3203 mutex_lock(&ctx->uring_lock);
3204 ret |= io_poll_remove_all(ctx, task, cancel_all);
3205 mutex_unlock(&ctx->uring_lock);
3206 ret |= io_kill_timeouts(ctx, task, cancel_all);
3208 ret |= io_run_task_work() > 0;
3212 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
3215 return atomic_read(&tctx->inflight_tracked);
3216 return percpu_counter_sum(&tctx->inflight);
3220 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3221 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3223 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
3225 struct io_uring_task *tctx = current->io_uring;
3226 struct io_ring_ctx *ctx;
3230 WARN_ON_ONCE(sqd && sqd->thread != current);
3232 if (!current->io_uring)
3235 io_wq_exit_start(tctx->io_wq);
3237 atomic_inc(&tctx->in_cancel);
3241 io_uring_drop_tctx_refs(current);
3242 /* read completions before cancelations */
3243 inflight = tctx_inflight(tctx, !cancel_all);
3248 struct io_tctx_node *node;
3249 unsigned long index;
3251 xa_for_each(&tctx->xa, index, node) {
3252 /* sqpoll task will cancel all its requests */
3253 if (node->ctx->sq_data)
3255 loop |= io_uring_try_cancel_requests(node->ctx,
3256 current, cancel_all);
3259 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3260 loop |= io_uring_try_cancel_requests(ctx,
3270 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3272 io_uring_drop_tctx_refs(current);
3275 * If we've seen completions, retry without waiting. This
3276 * avoids a race where a completion comes in before we did
3277 * prepare_to_wait().
3279 if (inflight == tctx_inflight(tctx, !cancel_all))
3281 finish_wait(&tctx->wait, &wait);
3284 io_uring_clean_tctx(tctx);
3287 * We shouldn't run task_works after cancel, so just leave
3288 * ->in_cancel set for normal exit.
3290 atomic_dec(&tctx->in_cancel);
3291 /* for exec all current's requests should be gone, kill tctx */
3292 __io_uring_free(current);
3296 void __io_uring_cancel(bool cancel_all)
3298 io_uring_cancel_generic(cancel_all, NULL);
3301 static void *io_uring_validate_mmap_request(struct file *file,
3302 loff_t pgoff, size_t sz)
3304 struct io_ring_ctx *ctx = file->private_data;
3305 loff_t offset = pgoff << PAGE_SHIFT;
3309 switch (offset & IORING_OFF_MMAP_MASK) {
3310 case IORING_OFF_SQ_RING:
3311 case IORING_OFF_CQ_RING:
3314 case IORING_OFF_SQES:
3317 case IORING_OFF_PBUF_RING: {
3320 bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT;
3321 mutex_lock(&ctx->uring_lock);
3322 ptr = io_pbuf_get_address(ctx, bgid);
3323 mutex_unlock(&ctx->uring_lock);
3325 return ERR_PTR(-EINVAL);
3329 return ERR_PTR(-EINVAL);
3332 page = virt_to_head_page(ptr);
3333 if (sz > page_size(page))
3334 return ERR_PTR(-EINVAL);
3341 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3343 size_t sz = vma->vm_end - vma->vm_start;
3347 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3349 return PTR_ERR(ptr);
3351 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3352 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3355 static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp,
3356 unsigned long addr, unsigned long len,
3357 unsigned long pgoff, unsigned long flags)
3359 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
3360 struct vm_unmapped_area_info info;
3364 * Do not allow to map to user-provided address to avoid breaking the
3365 * aliasing rules. Userspace is not able to guess the offset address of
3366 * kernel kmalloc()ed memory area.
3371 ptr = io_uring_validate_mmap_request(filp, pgoff, len);
3375 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
3377 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
3378 info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
3380 info.align_mask = PAGE_MASK & (SHM_COLOUR - 1UL);
3382 info.align_mask = PAGE_MASK & (SHMLBA - 1UL);
3384 info.align_offset = (unsigned long) ptr;
3387 * A failed mmap() very likely causes application failure,
3388 * so fall back to the bottom-up function here. This scenario
3389 * can happen with large stack limits and large mmap()
3392 addr = vm_unmapped_area(&info);
3393 if (offset_in_page(addr)) {
3395 info.low_limit = TASK_UNMAPPED_BASE;
3396 info.high_limit = mmap_end;
3397 addr = vm_unmapped_area(&info);
3403 #else /* !CONFIG_MMU */
3405 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3407 return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL;
3410 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3412 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3415 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3416 unsigned long addr, unsigned long len,
3417 unsigned long pgoff, unsigned long flags)
3421 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3423 return PTR_ERR(ptr);
3425 return (unsigned long) ptr;
3428 #endif /* !CONFIG_MMU */
3430 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3432 if (flags & IORING_ENTER_EXT_ARG) {
3433 struct io_uring_getevents_arg arg;
3435 if (argsz != sizeof(arg))
3437 if (copy_from_user(&arg, argp, sizeof(arg)))
3443 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3444 struct __kernel_timespec __user **ts,
3445 const sigset_t __user **sig)
3447 struct io_uring_getevents_arg arg;
3450 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3451 * is just a pointer to the sigset_t.
3453 if (!(flags & IORING_ENTER_EXT_ARG)) {
3454 *sig = (const sigset_t __user *) argp;
3460 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3461 * timespec and sigset_t pointers if good.
3463 if (*argsz != sizeof(arg))
3465 if (copy_from_user(&arg, argp, sizeof(arg)))
3469 *sig = u64_to_user_ptr(arg.sigmask);
3470 *argsz = arg.sigmask_sz;
3471 *ts = u64_to_user_ptr(arg.ts);
3475 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3476 u32, min_complete, u32, flags, const void __user *, argp,
3479 struct io_ring_ctx *ctx;
3483 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3484 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3485 IORING_ENTER_REGISTERED_RING)))
3489 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3490 * need only dereference our task private array to find it.
3492 if (flags & IORING_ENTER_REGISTERED_RING) {
3493 struct io_uring_task *tctx = current->io_uring;
3495 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3497 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3498 f.file = tctx->registered_rings[fd];
3500 if (unlikely(!f.file))
3504 if (unlikely(!f.file))
3507 if (unlikely(!io_is_uring_fops(f.file)))
3511 ctx = f.file->private_data;
3513 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3517 * For SQ polling, the thread will do all submissions and completions.
3518 * Just return the requested submit count, and wake the thread if
3522 if (ctx->flags & IORING_SETUP_SQPOLL) {
3523 io_cqring_overflow_flush(ctx);
3525 if (unlikely(ctx->sq_data->thread == NULL)) {
3529 if (flags & IORING_ENTER_SQ_WAKEUP)
3530 wake_up(&ctx->sq_data->wait);
3531 if (flags & IORING_ENTER_SQ_WAIT)
3532 io_sqpoll_wait_sq(ctx);
3535 } else if (to_submit) {
3536 ret = io_uring_add_tctx_node(ctx);
3540 mutex_lock(&ctx->uring_lock);
3541 ret = io_submit_sqes(ctx, to_submit);
3542 if (ret != to_submit) {
3543 mutex_unlock(&ctx->uring_lock);
3546 if (flags & IORING_ENTER_GETEVENTS) {
3547 if (ctx->syscall_iopoll)
3550 * Ignore errors, we'll soon call io_cqring_wait() and
3551 * it should handle ownership problems if any.
3553 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3554 (void)io_run_local_work_locked(ctx);
3556 mutex_unlock(&ctx->uring_lock);
3559 if (flags & IORING_ENTER_GETEVENTS) {
3562 if (ctx->syscall_iopoll) {
3564 * We disallow the app entering submit/complete with
3565 * polling, but we still need to lock the ring to
3566 * prevent racing with polled issue that got punted to
3569 mutex_lock(&ctx->uring_lock);
3571 ret2 = io_validate_ext_arg(flags, argp, argsz);
3572 if (likely(!ret2)) {
3573 min_complete = min(min_complete,
3575 ret2 = io_iopoll_check(ctx, min_complete);
3577 mutex_unlock(&ctx->uring_lock);
3579 const sigset_t __user *sig;
3580 struct __kernel_timespec __user *ts;
3582 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3583 if (likely(!ret2)) {
3584 min_complete = min(min_complete,
3586 ret2 = io_cqring_wait(ctx, min_complete, sig,
3595 * EBADR indicates that one or more CQE were dropped.
3596 * Once the user has been informed we can clear the bit
3597 * as they are obviously ok with those drops.
3599 if (unlikely(ret2 == -EBADR))
3600 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3609 static const struct file_operations io_uring_fops = {
3610 .release = io_uring_release,
3611 .mmap = io_uring_mmap,
3613 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3614 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3616 .get_unmapped_area = io_uring_mmu_get_unmapped_area,
3618 .poll = io_uring_poll,
3619 #ifdef CONFIG_PROC_FS
3620 .show_fdinfo = io_uring_show_fdinfo,
3624 bool io_is_uring_fops(struct file *file)
3626 return file->f_op == &io_uring_fops;
3629 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3630 struct io_uring_params *p)
3632 struct io_rings *rings;
3633 size_t size, sq_array_offset;
3635 /* make sure these are sane, as we already accounted them */
3636 ctx->sq_entries = p->sq_entries;
3637 ctx->cq_entries = p->cq_entries;
3639 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3640 if (size == SIZE_MAX)
3643 rings = io_mem_alloc(size);
3648 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3649 rings->sq_ring_mask = p->sq_entries - 1;
3650 rings->cq_ring_mask = p->cq_entries - 1;
3651 rings->sq_ring_entries = p->sq_entries;
3652 rings->cq_ring_entries = p->cq_entries;
3654 if (p->flags & IORING_SETUP_SQE128)
3655 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3657 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3658 if (size == SIZE_MAX) {
3659 io_mem_free(ctx->rings);
3664 ctx->sq_sqes = io_mem_alloc(size);
3665 if (!ctx->sq_sqes) {
3666 io_mem_free(ctx->rings);
3674 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3678 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3682 ret = __io_uring_add_tctx_node(ctx);
3687 fd_install(fd, file);
3692 * Allocate an anonymous fd, this is what constitutes the application
3693 * visible backing of an io_uring instance. The application mmaps this
3694 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3695 * we have to tie this fd to a socket for file garbage collection purposes.
3697 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3700 #if defined(CONFIG_UNIX)
3703 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3706 return ERR_PTR(ret);
3709 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3710 O_RDWR | O_CLOEXEC, NULL);
3711 #if defined(CONFIG_UNIX)
3713 sock_release(ctx->ring_sock);
3714 ctx->ring_sock = NULL;
3716 ctx->ring_sock->file = file;
3722 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3723 struct io_uring_params __user *params)
3725 struct io_ring_ctx *ctx;
3731 if (entries > IORING_MAX_ENTRIES) {
3732 if (!(p->flags & IORING_SETUP_CLAMP))
3734 entries = IORING_MAX_ENTRIES;
3738 * Use twice as many entries for the CQ ring. It's possible for the
3739 * application to drive a higher depth than the size of the SQ ring,
3740 * since the sqes are only used at submission time. This allows for
3741 * some flexibility in overcommitting a bit. If the application has
3742 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3743 * of CQ ring entries manually.
3745 p->sq_entries = roundup_pow_of_two(entries);
3746 if (p->flags & IORING_SETUP_CQSIZE) {
3748 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3749 * to a power-of-two, if it isn't already. We do NOT impose
3750 * any cq vs sq ring sizing.
3754 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3755 if (!(p->flags & IORING_SETUP_CLAMP))
3757 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3759 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3760 if (p->cq_entries < p->sq_entries)
3763 p->cq_entries = 2 * p->sq_entries;
3766 ctx = io_ring_ctx_alloc(p);
3770 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3771 !(ctx->flags & IORING_SETUP_IOPOLL) &&
3772 !(ctx->flags & IORING_SETUP_SQPOLL))
3773 ctx->task_complete = true;
3776 * lazy poll_wq activation relies on ->task_complete for synchronisation
3777 * purposes, see io_activate_pollwq()
3779 if (!ctx->task_complete)
3780 ctx->poll_activated = true;
3783 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3784 * space applications don't need to do io completion events
3785 * polling again, they can rely on io_sq_thread to do polling
3786 * work, which can reduce cpu usage and uring_lock contention.
3788 if (ctx->flags & IORING_SETUP_IOPOLL &&
3789 !(ctx->flags & IORING_SETUP_SQPOLL))
3790 ctx->syscall_iopoll = 1;
3792 ctx->compat = in_compat_syscall();
3793 if (!capable(CAP_IPC_LOCK))
3794 ctx->user = get_uid(current_user());
3797 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3798 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3801 if (ctx->flags & IORING_SETUP_SQPOLL) {
3802 /* IPI related flags don't make sense with SQPOLL */
3803 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3804 IORING_SETUP_TASKRUN_FLAG |
3805 IORING_SETUP_DEFER_TASKRUN))
3807 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3808 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3809 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3811 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3812 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3814 ctx->notify_method = TWA_SIGNAL;
3818 * For DEFER_TASKRUN we require the completion task to be the same as the
3819 * submission task. This implies that there is only one submitter, so enforce
3822 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3823 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3828 * This is just grabbed for accounting purposes. When a process exits,
3829 * the mm is exited and dropped before the files, hence we need to hang
3830 * on to this mm purely for the purposes of being able to unaccount
3831 * memory (locked/pinned vm). It's not used for anything else.
3833 mmgrab(current->mm);
3834 ctx->mm_account = current->mm;
3836 ret = io_allocate_scq_urings(ctx, p);
3840 ret = io_sq_offload_create(ctx, p);
3843 /* always set a rsrc node */
3844 ret = io_rsrc_node_switch_start(ctx);
3847 io_rsrc_node_switch(ctx, NULL);
3849 memset(&p->sq_off, 0, sizeof(p->sq_off));
3850 p->sq_off.head = offsetof(struct io_rings, sq.head);
3851 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3852 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3853 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3854 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3855 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3856 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3858 memset(&p->cq_off, 0, sizeof(p->cq_off));
3859 p->cq_off.head = offsetof(struct io_rings, cq.head);
3860 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3861 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3862 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3863 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3864 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3865 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3867 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3868 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3869 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3870 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3871 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3872 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3873 IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING;
3875 if (copy_to_user(params, p, sizeof(*p))) {
3880 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
3881 && !(ctx->flags & IORING_SETUP_R_DISABLED))
3882 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
3884 file = io_uring_get_file(ctx);
3886 ret = PTR_ERR(file);
3891 * Install ring fd as the very last thing, so we don't risk someone
3892 * having closed it before we finish setup
3894 ret = io_uring_install_fd(ctx, file);
3896 /* fput will clean it up */
3901 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3904 io_ring_ctx_wait_and_kill(ctx);
3909 * Sets up an aio uring context, and returns the fd. Applications asks for a
3910 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3911 * params structure passed in.
3913 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3915 struct io_uring_params p;
3918 if (copy_from_user(&p, params, sizeof(p)))
3920 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3925 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3926 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3927 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3928 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3929 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3930 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3931 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN))
3934 return io_uring_create(entries, &p, params);
3937 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3938 struct io_uring_params __user *, params)
3940 return io_uring_setup(entries, params);
3943 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3946 struct io_uring_probe *p;
3950 size = struct_size(p, ops, nr_args);
3951 if (size == SIZE_MAX)
3953 p = kzalloc(size, GFP_KERNEL);
3958 if (copy_from_user(p, arg, size))
3961 if (memchr_inv(p, 0, size))
3964 p->last_op = IORING_OP_LAST - 1;
3965 if (nr_args > IORING_OP_LAST)
3966 nr_args = IORING_OP_LAST;
3968 for (i = 0; i < nr_args; i++) {
3970 if (!io_issue_defs[i].not_supported)
3971 p->ops[i].flags = IO_URING_OP_SUPPORTED;
3976 if (copy_to_user(arg, p, size))
3983 static int io_register_personality(struct io_ring_ctx *ctx)
3985 const struct cred *creds;
3989 creds = get_current_cred();
3991 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
3992 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
4000 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
4001 void __user *arg, unsigned int nr_args)
4003 struct io_uring_restriction *res;
4007 /* Restrictions allowed only if rings started disabled */
4008 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
4011 /* We allow only a single restrictions registration */
4012 if (ctx->restrictions.registered)
4015 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
4018 size = array_size(nr_args, sizeof(*res));
4019 if (size == SIZE_MAX)
4022 res = memdup_user(arg, size);
4024 return PTR_ERR(res);
4028 for (i = 0; i < nr_args; i++) {
4029 switch (res[i].opcode) {
4030 case IORING_RESTRICTION_REGISTER_OP:
4031 if (res[i].register_op >= IORING_REGISTER_LAST) {
4036 __set_bit(res[i].register_op,
4037 ctx->restrictions.register_op);
4039 case IORING_RESTRICTION_SQE_OP:
4040 if (res[i].sqe_op >= IORING_OP_LAST) {
4045 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
4047 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
4048 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
4050 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
4051 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
4060 /* Reset all restrictions if an error happened */
4062 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
4064 ctx->restrictions.registered = true;
4070 static int io_register_enable_rings(struct io_ring_ctx *ctx)
4072 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
4075 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task) {
4076 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
4078 * Lazy activation attempts would fail if it was polled before
4079 * submitter_task is set.
4081 if (wq_has_sleeper(&ctx->poll_wq))
4082 io_activate_pollwq(ctx);
4085 if (ctx->restrictions.registered)
4086 ctx->restricted = 1;
4088 ctx->flags &= ~IORING_SETUP_R_DISABLED;
4089 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
4090 wake_up(&ctx->sq_data->wait);
4094 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
4095 void __user *arg, unsigned len)
4097 struct io_uring_task *tctx = current->io_uring;
4098 cpumask_var_t new_mask;
4101 if (!tctx || !tctx->io_wq)
4104 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4107 cpumask_clear(new_mask);
4108 if (len > cpumask_size())
4109 len = cpumask_size();
4111 if (in_compat_syscall()) {
4112 ret = compat_get_bitmap(cpumask_bits(new_mask),
4113 (const compat_ulong_t __user *)arg,
4114 len * 8 /* CHAR_BIT */);
4116 ret = copy_from_user(new_mask, arg, len);
4120 free_cpumask_var(new_mask);
4124 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
4125 free_cpumask_var(new_mask);
4129 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
4131 struct io_uring_task *tctx = current->io_uring;
4133 if (!tctx || !tctx->io_wq)
4136 return io_wq_cpu_affinity(tctx->io_wq, NULL);
4139 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
4141 __must_hold(&ctx->uring_lock)
4143 struct io_tctx_node *node;
4144 struct io_uring_task *tctx = NULL;
4145 struct io_sq_data *sqd = NULL;
4149 if (copy_from_user(new_count, arg, sizeof(new_count)))
4151 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4152 if (new_count[i] > INT_MAX)
4155 if (ctx->flags & IORING_SETUP_SQPOLL) {
4159 * Observe the correct sqd->lock -> ctx->uring_lock
4160 * ordering. Fine to drop uring_lock here, we hold
4163 refcount_inc(&sqd->refs);
4164 mutex_unlock(&ctx->uring_lock);
4165 mutex_lock(&sqd->lock);
4166 mutex_lock(&ctx->uring_lock);
4168 tctx = sqd->thread->io_uring;
4171 tctx = current->io_uring;
4174 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
4176 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4178 ctx->iowq_limits[i] = new_count[i];
4179 ctx->iowq_limits_set = true;
4181 if (tctx && tctx->io_wq) {
4182 ret = io_wq_max_workers(tctx->io_wq, new_count);
4186 memset(new_count, 0, sizeof(new_count));
4190 mutex_unlock(&sqd->lock);
4191 io_put_sq_data(sqd);
4194 if (copy_to_user(arg, new_count, sizeof(new_count)))
4197 /* that's it for SQPOLL, only the SQPOLL task creates requests */
4201 /* now propagate the restriction to all registered users */
4202 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
4203 struct io_uring_task *tctx = node->task->io_uring;
4205 if (WARN_ON_ONCE(!tctx->io_wq))
4208 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4209 new_count[i] = ctx->iowq_limits[i];
4210 /* ignore errors, it always returns zero anyway */
4211 (void)io_wq_max_workers(tctx->io_wq, new_count);
4216 mutex_unlock(&sqd->lock);
4217 io_put_sq_data(sqd);
4222 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
4223 void __user *arg, unsigned nr_args)
4224 __releases(ctx->uring_lock)
4225 __acquires(ctx->uring_lock)
4230 * We don't quiesce the refs for register anymore and so it can't be
4231 * dying as we're holding a file ref here.
4233 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
4236 if (ctx->submitter_task && ctx->submitter_task != current)
4239 if (ctx->restricted) {
4240 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
4241 if (!test_bit(opcode, ctx->restrictions.register_op))
4246 case IORING_REGISTER_BUFFERS:
4250 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
4252 case IORING_UNREGISTER_BUFFERS:
4256 ret = io_sqe_buffers_unregister(ctx);
4258 case IORING_REGISTER_FILES:
4262 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
4264 case IORING_UNREGISTER_FILES:
4268 ret = io_sqe_files_unregister(ctx);
4270 case IORING_REGISTER_FILES_UPDATE:
4271 ret = io_register_files_update(ctx, arg, nr_args);
4273 case IORING_REGISTER_EVENTFD:
4277 ret = io_eventfd_register(ctx, arg, 0);
4279 case IORING_REGISTER_EVENTFD_ASYNC:
4283 ret = io_eventfd_register(ctx, arg, 1);
4285 case IORING_UNREGISTER_EVENTFD:
4289 ret = io_eventfd_unregister(ctx);
4291 case IORING_REGISTER_PROBE:
4293 if (!arg || nr_args > 256)
4295 ret = io_probe(ctx, arg, nr_args);
4297 case IORING_REGISTER_PERSONALITY:
4301 ret = io_register_personality(ctx);
4303 case IORING_UNREGISTER_PERSONALITY:
4307 ret = io_unregister_personality(ctx, nr_args);
4309 case IORING_REGISTER_ENABLE_RINGS:
4313 ret = io_register_enable_rings(ctx);
4315 case IORING_REGISTER_RESTRICTIONS:
4316 ret = io_register_restrictions(ctx, arg, nr_args);
4318 case IORING_REGISTER_FILES2:
4319 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
4321 case IORING_REGISTER_FILES_UPDATE2:
4322 ret = io_register_rsrc_update(ctx, arg, nr_args,
4325 case IORING_REGISTER_BUFFERS2:
4326 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
4328 case IORING_REGISTER_BUFFERS_UPDATE:
4329 ret = io_register_rsrc_update(ctx, arg, nr_args,
4330 IORING_RSRC_BUFFER);
4332 case IORING_REGISTER_IOWQ_AFF:
4334 if (!arg || !nr_args)
4336 ret = io_register_iowq_aff(ctx, arg, nr_args);
4338 case IORING_UNREGISTER_IOWQ_AFF:
4342 ret = io_unregister_iowq_aff(ctx);
4344 case IORING_REGISTER_IOWQ_MAX_WORKERS:
4346 if (!arg || nr_args != 2)
4348 ret = io_register_iowq_max_workers(ctx, arg);
4350 case IORING_REGISTER_RING_FDS:
4351 ret = io_ringfd_register(ctx, arg, nr_args);
4353 case IORING_UNREGISTER_RING_FDS:
4354 ret = io_ringfd_unregister(ctx, arg, nr_args);
4356 case IORING_REGISTER_PBUF_RING:
4358 if (!arg || nr_args != 1)
4360 ret = io_register_pbuf_ring(ctx, arg);
4362 case IORING_UNREGISTER_PBUF_RING:
4364 if (!arg || nr_args != 1)
4366 ret = io_unregister_pbuf_ring(ctx, arg);
4368 case IORING_REGISTER_SYNC_CANCEL:
4370 if (!arg || nr_args != 1)
4372 ret = io_sync_cancel(ctx, arg);
4374 case IORING_REGISTER_FILE_ALLOC_RANGE:
4376 if (!arg || nr_args)
4378 ret = io_register_file_alloc_range(ctx, arg);
4388 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
4389 void __user *, arg, unsigned int, nr_args)
4391 struct io_ring_ctx *ctx;
4394 bool use_registered_ring;
4396 use_registered_ring = !!(opcode & IORING_REGISTER_USE_REGISTERED_RING);
4397 opcode &= ~IORING_REGISTER_USE_REGISTERED_RING;
4399 if (opcode >= IORING_REGISTER_LAST)
4402 if (use_registered_ring) {
4404 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
4405 * need only dereference our task private array to find it.
4407 struct io_uring_task *tctx = current->io_uring;
4409 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
4411 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
4412 f.file = tctx->registered_rings[fd];
4414 if (unlikely(!f.file))
4418 if (unlikely(!f.file))
4421 if (!io_is_uring_fops(f.file))
4425 ctx = f.file->private_data;
4427 mutex_lock(&ctx->uring_lock);
4428 ret = __io_uring_register(ctx, opcode, arg, nr_args);
4429 mutex_unlock(&ctx->uring_lock);
4430 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
4436 static int __init io_uring_init(void)
4438 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4439 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4440 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4443 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4444 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4445 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4446 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4447 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4448 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
4449 BUILD_BUG_SQE_ELEM(1, __u8, flags);
4450 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
4451 BUILD_BUG_SQE_ELEM(4, __s32, fd);
4452 BUILD_BUG_SQE_ELEM(8, __u64, off);
4453 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
4454 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
4455 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
4456 BUILD_BUG_SQE_ELEM(16, __u64, addr);
4457 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
4458 BUILD_BUG_SQE_ELEM(24, __u32, len);
4459 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
4460 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
4461 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4462 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
4463 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
4464 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
4465 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
4466 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
4467 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
4468 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
4469 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
4470 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
4471 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
4472 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
4473 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
4474 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
4475 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
4476 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
4477 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
4478 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
4479 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
4480 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
4481 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
4482 BUILD_BUG_SQE_ELEM(42, __u16, personality);
4483 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
4484 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
4485 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
4486 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
4487 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
4488 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4489 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
4491 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4492 sizeof(struct io_uring_rsrc_update));
4493 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4494 sizeof(struct io_uring_rsrc_update2));
4496 /* ->buf_index is u16 */
4497 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4498 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4499 offsetof(struct io_uring_buf_ring, tail));
4501 /* should fit into one byte */
4502 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
4503 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4504 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4506 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
4508 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4510 io_uring_optable_init();
4512 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4516 __initcall(io_uring_init);