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, unsigned flags)
1305 struct io_ring_ctx *ctx = req->ctx;
1306 unsigned nr_wait, nr_tw, nr_tw_prev;
1307 struct llist_node *first;
1309 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK))
1310 flags &= ~IOU_F_TWQ_LAZY_WAKE;
1312 first = READ_ONCE(ctx->work_llist.first);
1316 struct io_kiocb *first_req = container_of(first,
1320 * Might be executed at any moment, rely on
1321 * SLAB_TYPESAFE_BY_RCU to keep it alive.
1323 nr_tw_prev = READ_ONCE(first_req->nr_tw);
1325 nr_tw = nr_tw_prev + 1;
1326 /* Large enough to fail the nr_wait comparison below */
1327 if (!(flags & IOU_F_TWQ_LAZY_WAKE))
1331 req->io_task_work.node.next = first;
1332 } while (!try_cmpxchg(&ctx->work_llist.first, &first,
1333 &req->io_task_work.node));
1336 if (unlikely(atomic_read(&req->task->io_uring->in_cancel))) {
1337 io_move_task_work_from_local(ctx);
1340 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1341 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1343 io_eventfd_signal(ctx);
1346 nr_wait = atomic_read(&ctx->cq_wait_nr);
1347 /* no one is waiting */
1350 /* either not enough or the previous add has already woken it up */
1351 if (nr_wait > nr_tw || nr_tw_prev >= nr_wait)
1353 /* pairs with set_current_state() in io_cqring_wait() */
1354 smp_mb__after_atomic();
1355 wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE);
1358 void __io_req_task_work_add(struct io_kiocb *req, unsigned flags)
1360 struct io_uring_task *tctx = req->task->io_uring;
1361 struct io_ring_ctx *ctx = req->ctx;
1363 if (!(flags & IOU_F_TWQ_FORCE_NORMAL) &&
1364 (ctx->flags & IORING_SETUP_DEFER_TASKRUN)) {
1366 io_req_local_work_add(req, flags);
1371 /* task_work already pending, we're done */
1372 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1375 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1376 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1378 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1381 io_fallback_tw(tctx);
1384 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1386 struct llist_node *node;
1388 node = llist_del_all(&ctx->work_llist);
1390 struct io_kiocb *req = container_of(node, struct io_kiocb,
1394 __io_req_task_work_add(req, IOU_F_TWQ_FORCE_NORMAL);
1398 static int __io_run_local_work(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1400 struct llist_node *node;
1401 unsigned int loops = 0;
1404 if (WARN_ON_ONCE(ctx->submitter_task != current))
1406 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1407 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1409 node = io_llist_xchg(&ctx->work_llist, NULL);
1411 struct llist_node *next = node->next;
1412 struct io_kiocb *req = container_of(node, struct io_kiocb,
1414 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1415 req->io_task_work.func(req, ts);
1421 if (!llist_empty(&ctx->work_llist))
1424 io_submit_flush_completions(ctx);
1425 if (!llist_empty(&ctx->work_llist))
1428 trace_io_uring_local_work_run(ctx, ret, loops);
1432 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx)
1434 struct io_tw_state ts = { .locked = true, };
1437 if (llist_empty(&ctx->work_llist))
1440 ret = __io_run_local_work(ctx, &ts);
1441 /* shouldn't happen! */
1442 if (WARN_ON_ONCE(!ts.locked))
1443 mutex_lock(&ctx->uring_lock);
1447 static int io_run_local_work(struct io_ring_ctx *ctx)
1449 struct io_tw_state ts = {};
1452 ts.locked = mutex_trylock(&ctx->uring_lock);
1453 ret = __io_run_local_work(ctx, &ts);
1455 mutex_unlock(&ctx->uring_lock);
1460 static void io_req_task_cancel(struct io_kiocb *req, struct io_tw_state *ts)
1462 io_tw_lock(req->ctx, ts);
1463 io_req_defer_failed(req, req->cqe.res);
1466 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts)
1468 io_tw_lock(req->ctx, ts);
1469 /* req->task == current here, checking PF_EXITING is safe */
1470 if (unlikely(req->task->flags & PF_EXITING))
1471 io_req_defer_failed(req, -EFAULT);
1472 else if (req->flags & REQ_F_FORCE_ASYNC)
1473 io_queue_iowq(req, ts);
1478 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1480 io_req_set_res(req, ret, 0);
1481 req->io_task_work.func = io_req_task_cancel;
1482 io_req_task_work_add(req);
1485 void io_req_task_queue(struct io_kiocb *req)
1487 req->io_task_work.func = io_req_task_submit;
1488 io_req_task_work_add(req);
1491 void io_queue_next(struct io_kiocb *req)
1493 struct io_kiocb *nxt = io_req_find_next(req);
1496 io_req_task_queue(nxt);
1499 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1500 __must_hold(&ctx->uring_lock)
1502 struct task_struct *task = NULL;
1506 struct io_kiocb *req = container_of(node, struct io_kiocb,
1509 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1510 if (req->flags & REQ_F_REFCOUNT) {
1511 node = req->comp_list.next;
1512 if (!req_ref_put_and_test(req))
1515 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1516 struct async_poll *apoll = req->apoll;
1518 if (apoll->double_poll)
1519 kfree(apoll->double_poll);
1520 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1522 req->flags &= ~REQ_F_POLLED;
1524 if (req->flags & IO_REQ_LINK_FLAGS)
1526 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1529 if (!(req->flags & REQ_F_FIXED_FILE))
1530 io_put_file(req->file);
1532 io_req_put_rsrc_locked(req, ctx);
1534 if (req->task != task) {
1536 io_put_task(task, task_refs);
1541 node = req->comp_list.next;
1542 io_req_add_to_cache(req, ctx);
1546 io_put_task(task, task_refs);
1549 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1550 __must_hold(&ctx->uring_lock)
1552 struct io_submit_state *state = &ctx->submit_state;
1553 struct io_wq_work_node *node;
1556 /* must come first to preserve CQE ordering in failure cases */
1557 if (state->cqes_count)
1558 __io_flush_post_cqes(ctx);
1559 __wq_list_for_each(node, &state->compl_reqs) {
1560 struct io_kiocb *req = container_of(node, struct io_kiocb,
1563 if (!(req->flags & REQ_F_CQE_SKIP) &&
1564 unlikely(!__io_fill_cqe_req(ctx, req))) {
1565 if (ctx->task_complete) {
1566 spin_lock(&ctx->completion_lock);
1567 io_req_cqe_overflow(req);
1568 spin_unlock(&ctx->completion_lock);
1570 io_req_cqe_overflow(req);
1574 __io_cq_unlock_post_flush(ctx);
1576 if (!wq_list_empty(&ctx->submit_state.compl_reqs)) {
1577 io_free_batch_list(ctx, state->compl_reqs.first);
1578 INIT_WQ_LIST(&state->compl_reqs);
1583 * Drop reference to request, return next in chain (if there is one) if this
1584 * was the last reference to this request.
1586 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1588 struct io_kiocb *nxt = NULL;
1590 if (req_ref_put_and_test(req)) {
1591 if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
1592 nxt = io_req_find_next(req);
1598 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1600 /* See comment at the top of this file */
1602 return __io_cqring_events(ctx);
1606 * We can't just wait for polled events to come to us, we have to actively
1607 * find and complete them.
1609 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1611 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1614 mutex_lock(&ctx->uring_lock);
1615 while (!wq_list_empty(&ctx->iopoll_list)) {
1616 /* let it sleep and repeat later if can't complete a request */
1617 if (io_do_iopoll(ctx, true) == 0)
1620 * Ensure we allow local-to-the-cpu processing to take place,
1621 * in this case we need to ensure that we reap all events.
1622 * Also let task_work, etc. to progress by releasing the mutex
1624 if (need_resched()) {
1625 mutex_unlock(&ctx->uring_lock);
1627 mutex_lock(&ctx->uring_lock);
1630 mutex_unlock(&ctx->uring_lock);
1633 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1635 unsigned int nr_events = 0;
1637 unsigned long check_cq;
1639 if (!io_allowed_run_tw(ctx))
1642 check_cq = READ_ONCE(ctx->check_cq);
1643 if (unlikely(check_cq)) {
1644 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1645 __io_cqring_overflow_flush(ctx);
1647 * Similarly do not spin if we have not informed the user of any
1650 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1654 * Don't enter poll loop if we already have events pending.
1655 * If we do, we can potentially be spinning for commands that
1656 * already triggered a CQE (eg in error).
1658 if (io_cqring_events(ctx))
1663 * If a submit got punted to a workqueue, we can have the
1664 * application entering polling for a command before it gets
1665 * issued. That app will hold the uring_lock for the duration
1666 * of the poll right here, so we need to take a breather every
1667 * now and then to ensure that the issue has a chance to add
1668 * the poll to the issued list. Otherwise we can spin here
1669 * forever, while the workqueue is stuck trying to acquire the
1672 if (wq_list_empty(&ctx->iopoll_list) ||
1673 io_task_work_pending(ctx)) {
1674 u32 tail = ctx->cached_cq_tail;
1676 (void) io_run_local_work_locked(ctx);
1678 if (task_work_pending(current) ||
1679 wq_list_empty(&ctx->iopoll_list)) {
1680 mutex_unlock(&ctx->uring_lock);
1682 mutex_lock(&ctx->uring_lock);
1684 /* some requests don't go through iopoll_list */
1685 if (tail != ctx->cached_cq_tail ||
1686 wq_list_empty(&ctx->iopoll_list))
1689 ret = io_do_iopoll(ctx, !min);
1694 } while (nr_events < min && !need_resched());
1699 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts)
1702 io_req_complete_defer(req);
1704 io_req_complete_post(req, IO_URING_F_UNLOCKED);
1708 * After the iocb has been issued, it's safe to be found on the poll list.
1709 * Adding the kiocb to the list AFTER submission ensures that we don't
1710 * find it from a io_do_iopoll() thread before the issuer is done
1711 * accessing the kiocb cookie.
1713 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1715 struct io_ring_ctx *ctx = req->ctx;
1716 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1718 /* workqueue context doesn't hold uring_lock, grab it now */
1719 if (unlikely(needs_lock))
1720 mutex_lock(&ctx->uring_lock);
1723 * Track whether we have multiple files in our lists. This will impact
1724 * how we do polling eventually, not spinning if we're on potentially
1725 * different devices.
1727 if (wq_list_empty(&ctx->iopoll_list)) {
1728 ctx->poll_multi_queue = false;
1729 } else if (!ctx->poll_multi_queue) {
1730 struct io_kiocb *list_req;
1732 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1734 if (list_req->file != req->file)
1735 ctx->poll_multi_queue = true;
1739 * For fast devices, IO may have already completed. If it has, add
1740 * it to the front so we find it first.
1742 if (READ_ONCE(req->iopoll_completed))
1743 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1745 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1747 if (unlikely(needs_lock)) {
1749 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1750 * in sq thread task context or in io worker task context. If
1751 * current task context is sq thread, we don't need to check
1752 * whether should wake up sq thread.
1754 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1755 wq_has_sleeper(&ctx->sq_data->wait))
1756 wake_up(&ctx->sq_data->wait);
1758 mutex_unlock(&ctx->uring_lock);
1762 static bool io_bdev_nowait(struct block_device *bdev)
1764 return !bdev || bdev_nowait(bdev);
1768 * If we tracked the file through the SCM inflight mechanism, we could support
1769 * any file. For now, just ensure that anything potentially problematic is done
1772 static bool __io_file_supports_nowait(struct file *file, umode_t mode)
1774 if (S_ISBLK(mode)) {
1775 if (IS_ENABLED(CONFIG_BLOCK) &&
1776 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
1782 if (S_ISREG(mode)) {
1783 if (IS_ENABLED(CONFIG_BLOCK) &&
1784 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
1785 !io_is_uring_fops(file))
1790 /* any ->read/write should understand O_NONBLOCK */
1791 if (file->f_flags & O_NONBLOCK)
1793 return file->f_mode & FMODE_NOWAIT;
1797 * If we tracked the file through the SCM inflight mechanism, we could support
1798 * any file. For now, just ensure that anything potentially problematic is done
1801 unsigned int io_file_get_flags(struct file *file)
1803 umode_t mode = file_inode(file)->i_mode;
1804 unsigned int res = 0;
1808 if (__io_file_supports_nowait(file, mode))
1813 bool io_alloc_async_data(struct io_kiocb *req)
1815 WARN_ON_ONCE(!io_cold_defs[req->opcode].async_size);
1816 req->async_data = kmalloc(io_cold_defs[req->opcode].async_size, GFP_KERNEL);
1817 if (req->async_data) {
1818 req->flags |= REQ_F_ASYNC_DATA;
1824 int io_req_prep_async(struct io_kiocb *req)
1826 const struct io_cold_def *cdef = &io_cold_defs[req->opcode];
1827 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1829 /* assign early for deferred execution for non-fixed file */
1830 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE) && !req->file)
1831 req->file = io_file_get_normal(req, req->cqe.fd);
1832 if (!cdef->prep_async)
1834 if (WARN_ON_ONCE(req_has_async_data(req)))
1836 if (!def->manual_alloc) {
1837 if (io_alloc_async_data(req))
1840 return cdef->prep_async(req);
1843 static u32 io_get_sequence(struct io_kiocb *req)
1845 u32 seq = req->ctx->cached_sq_head;
1846 struct io_kiocb *cur;
1848 /* need original cached_sq_head, but it was increased for each req */
1849 io_for_each_link(cur, req)
1854 static __cold void io_drain_req(struct io_kiocb *req)
1855 __must_hold(&ctx->uring_lock)
1857 struct io_ring_ctx *ctx = req->ctx;
1858 struct io_defer_entry *de;
1860 u32 seq = io_get_sequence(req);
1862 /* Still need defer if there is pending req in defer list. */
1863 spin_lock(&ctx->completion_lock);
1864 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1865 spin_unlock(&ctx->completion_lock);
1867 ctx->drain_active = false;
1868 io_req_task_queue(req);
1871 spin_unlock(&ctx->completion_lock);
1873 io_prep_async_link(req);
1874 de = kmalloc(sizeof(*de), GFP_KERNEL);
1877 io_req_defer_failed(req, ret);
1881 spin_lock(&ctx->completion_lock);
1882 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1883 spin_unlock(&ctx->completion_lock);
1888 trace_io_uring_defer(req);
1891 list_add_tail(&de->list, &ctx->defer_list);
1892 spin_unlock(&ctx->completion_lock);
1895 static void io_clean_op(struct io_kiocb *req)
1897 if (req->flags & REQ_F_BUFFER_SELECTED) {
1898 spin_lock(&req->ctx->completion_lock);
1899 io_put_kbuf_comp(req);
1900 spin_unlock(&req->ctx->completion_lock);
1903 if (req->flags & REQ_F_NEED_CLEANUP) {
1904 const struct io_cold_def *def = &io_cold_defs[req->opcode];
1909 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1910 kfree(req->apoll->double_poll);
1914 if (req->flags & REQ_F_INFLIGHT) {
1915 struct io_uring_task *tctx = req->task->io_uring;
1917 atomic_dec(&tctx->inflight_tracked);
1919 if (req->flags & REQ_F_CREDS)
1920 put_cred(req->creds);
1921 if (req->flags & REQ_F_ASYNC_DATA) {
1922 kfree(req->async_data);
1923 req->async_data = NULL;
1925 req->flags &= ~IO_REQ_CLEAN_FLAGS;
1928 static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def,
1929 unsigned int issue_flags)
1931 if (req->file || !def->needs_file)
1934 if (req->flags & REQ_F_FIXED_FILE)
1935 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1937 req->file = io_file_get_normal(req, req->cqe.fd);
1942 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1944 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1945 const struct cred *creds = NULL;
1948 if (unlikely(!io_assign_file(req, def, issue_flags)))
1951 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1952 creds = override_creds(req->creds);
1954 if (!def->audit_skip)
1955 audit_uring_entry(req->opcode);
1957 ret = def->issue(req, issue_flags);
1959 if (!def->audit_skip)
1960 audit_uring_exit(!ret, ret);
1963 revert_creds(creds);
1965 if (ret == IOU_OK) {
1966 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1967 io_req_complete_defer(req);
1969 io_req_complete_post(req, issue_flags);
1970 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1973 /* If the op doesn't have a file, we're not polling for it */
1974 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
1975 io_iopoll_req_issued(req, issue_flags);
1980 int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts)
1982 io_tw_lock(req->ctx, ts);
1983 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT|
1984 IO_URING_F_COMPLETE_DEFER);
1987 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1989 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1991 req = io_put_req_find_next(req);
1992 return req ? &req->work : NULL;
1995 void io_wq_submit_work(struct io_wq_work *work)
1997 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1998 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1999 unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ;
2000 bool needs_poll = false;
2001 int ret = 0, err = -ECANCELED;
2003 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
2004 if (!(req->flags & REQ_F_REFCOUNT))
2005 __io_req_set_refcount(req, 2);
2009 io_arm_ltimeout(req);
2011 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
2012 if (work->flags & IO_WQ_WORK_CANCEL) {
2014 io_req_task_queue_fail(req, err);
2017 if (!io_assign_file(req, def, issue_flags)) {
2019 work->flags |= IO_WQ_WORK_CANCEL;
2023 if (req->flags & REQ_F_FORCE_ASYNC) {
2024 bool opcode_poll = def->pollin || def->pollout;
2026 if (opcode_poll && file_can_poll(req->file)) {
2028 issue_flags |= IO_URING_F_NONBLOCK;
2033 ret = io_issue_sqe(req, issue_flags);
2037 * We can get EAGAIN for iopolled IO even though we're
2038 * forcing a sync submission from here, since we can't
2039 * wait for request slots on the block side.
2042 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
2048 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
2050 /* aborted or ready, in either case retry blocking */
2052 issue_flags &= ~IO_URING_F_NONBLOCK;
2055 /* avoid locking problems by failing it from a clean context */
2057 io_req_task_queue_fail(req, ret);
2060 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
2061 unsigned int issue_flags)
2063 struct io_ring_ctx *ctx = req->ctx;
2064 struct file *file = NULL;
2065 unsigned long file_ptr;
2067 io_ring_submit_lock(ctx, issue_flags);
2069 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
2071 fd = array_index_nospec(fd, ctx->nr_user_files);
2072 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
2073 file = (struct file *) (file_ptr & FFS_MASK);
2074 file_ptr &= ~FFS_MASK;
2075 /* mask in overlapping REQ_F and FFS bits */
2076 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
2077 io_req_set_rsrc_node(req, ctx, 0);
2079 io_ring_submit_unlock(ctx, issue_flags);
2083 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
2085 struct file *file = fget(fd);
2087 trace_io_uring_file_get(req, fd);
2089 /* we don't allow fixed io_uring files */
2090 if (file && io_is_uring_fops(file))
2091 io_req_track_inflight(req);
2095 static void io_queue_async(struct io_kiocb *req, int ret)
2096 __must_hold(&req->ctx->uring_lock)
2098 struct io_kiocb *linked_timeout;
2100 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
2101 io_req_defer_failed(req, ret);
2105 linked_timeout = io_prep_linked_timeout(req);
2107 switch (io_arm_poll_handler(req, 0)) {
2108 case IO_APOLL_READY:
2109 io_kbuf_recycle(req, 0);
2110 io_req_task_queue(req);
2112 case IO_APOLL_ABORTED:
2113 io_kbuf_recycle(req, 0);
2114 io_queue_iowq(req, NULL);
2121 io_queue_linked_timeout(linked_timeout);
2124 static inline void io_queue_sqe(struct io_kiocb *req)
2125 __must_hold(&req->ctx->uring_lock)
2129 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
2132 * We async punt it if the file wasn't marked NOWAIT, or if the file
2133 * doesn't support non-blocking read/write attempts
2136 io_arm_ltimeout(req);
2138 io_queue_async(req, ret);
2141 static void io_queue_sqe_fallback(struct io_kiocb *req)
2142 __must_hold(&req->ctx->uring_lock)
2144 if (unlikely(req->flags & REQ_F_FAIL)) {
2146 * We don't submit, fail them all, for that replace hardlinks
2147 * with normal links. Extra REQ_F_LINK is tolerated.
2149 req->flags &= ~REQ_F_HARDLINK;
2150 req->flags |= REQ_F_LINK;
2151 io_req_defer_failed(req, req->cqe.res);
2153 int ret = io_req_prep_async(req);
2155 if (unlikely(ret)) {
2156 io_req_defer_failed(req, ret);
2160 if (unlikely(req->ctx->drain_active))
2163 io_queue_iowq(req, NULL);
2168 * Check SQE restrictions (opcode and flags).
2170 * Returns 'true' if SQE is allowed, 'false' otherwise.
2172 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
2173 struct io_kiocb *req,
2174 unsigned int sqe_flags)
2176 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
2179 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
2180 ctx->restrictions.sqe_flags_required)
2183 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
2184 ctx->restrictions.sqe_flags_required))
2190 static void io_init_req_drain(struct io_kiocb *req)
2192 struct io_ring_ctx *ctx = req->ctx;
2193 struct io_kiocb *head = ctx->submit_state.link.head;
2195 ctx->drain_active = true;
2198 * If we need to drain a request in the middle of a link, drain
2199 * the head request and the next request/link after the current
2200 * link. Considering sequential execution of links,
2201 * REQ_F_IO_DRAIN will be maintained for every request of our
2204 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2205 ctx->drain_next = true;
2209 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
2210 const struct io_uring_sqe *sqe)
2211 __must_hold(&ctx->uring_lock)
2213 const struct io_issue_def *def;
2214 unsigned int sqe_flags;
2218 /* req is partially pre-initialised, see io_preinit_req() */
2219 req->opcode = opcode = READ_ONCE(sqe->opcode);
2220 /* same numerical values with corresponding REQ_F_*, safe to copy */
2221 req->flags = sqe_flags = READ_ONCE(sqe->flags);
2222 req->cqe.user_data = READ_ONCE(sqe->user_data);
2224 req->rsrc_node = NULL;
2225 req->task = current;
2227 if (unlikely(opcode >= IORING_OP_LAST)) {
2231 def = &io_issue_defs[opcode];
2232 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2233 /* enforce forwards compatibility on users */
2234 if (sqe_flags & ~SQE_VALID_FLAGS)
2236 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2237 if (!def->buffer_select)
2239 req->buf_index = READ_ONCE(sqe->buf_group);
2241 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2242 ctx->drain_disabled = true;
2243 if (sqe_flags & IOSQE_IO_DRAIN) {
2244 if (ctx->drain_disabled)
2246 io_init_req_drain(req);
2249 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2250 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2252 /* knock it to the slow queue path, will be drained there */
2253 if (ctx->drain_active)
2254 req->flags |= REQ_F_FORCE_ASYNC;
2255 /* if there is no link, we're at "next" request and need to drain */
2256 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2257 ctx->drain_next = false;
2258 ctx->drain_active = true;
2259 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2263 if (!def->ioprio && sqe->ioprio)
2265 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2268 if (def->needs_file) {
2269 struct io_submit_state *state = &ctx->submit_state;
2271 req->cqe.fd = READ_ONCE(sqe->fd);
2274 * Plug now if we have more than 2 IO left after this, and the
2275 * target is potentially a read/write to block based storage.
2277 if (state->need_plug && def->plug) {
2278 state->plug_started = true;
2279 state->need_plug = false;
2280 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2284 personality = READ_ONCE(sqe->personality);
2288 req->creds = xa_load(&ctx->personalities, personality);
2291 get_cred(req->creds);
2292 ret = security_uring_override_creds(req->creds);
2294 put_cred(req->creds);
2297 req->flags |= REQ_F_CREDS;
2300 return def->prep(req, sqe);
2303 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2304 struct io_kiocb *req, int ret)
2306 struct io_ring_ctx *ctx = req->ctx;
2307 struct io_submit_link *link = &ctx->submit_state.link;
2308 struct io_kiocb *head = link->head;
2310 trace_io_uring_req_failed(sqe, req, ret);
2313 * Avoid breaking links in the middle as it renders links with SQPOLL
2314 * unusable. Instead of failing eagerly, continue assembling the link if
2315 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2316 * should find the flag and handle the rest.
2318 req_fail_link_node(req, ret);
2319 if (head && !(head->flags & REQ_F_FAIL))
2320 req_fail_link_node(head, -ECANCELED);
2322 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2324 link->last->link = req;
2328 io_queue_sqe_fallback(req);
2333 link->last->link = req;
2340 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2341 const struct io_uring_sqe *sqe)
2342 __must_hold(&ctx->uring_lock)
2344 struct io_submit_link *link = &ctx->submit_state.link;
2347 ret = io_init_req(ctx, req, sqe);
2349 return io_submit_fail_init(sqe, req, ret);
2351 trace_io_uring_submit_req(req);
2354 * If we already have a head request, queue this one for async
2355 * submittal once the head completes. If we don't have a head but
2356 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2357 * submitted sync once the chain is complete. If none of those
2358 * conditions are true (normal request), then just queue it.
2360 if (unlikely(link->head)) {
2361 ret = io_req_prep_async(req);
2363 return io_submit_fail_init(sqe, req, ret);
2365 trace_io_uring_link(req, link->head);
2366 link->last->link = req;
2369 if (req->flags & IO_REQ_LINK_FLAGS)
2371 /* last request of the link, flush it */
2374 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2377 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2378 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2379 if (req->flags & IO_REQ_LINK_FLAGS) {
2384 io_queue_sqe_fallback(req);
2394 * Batched submission is done, ensure local IO is flushed out.
2396 static void io_submit_state_end(struct io_ring_ctx *ctx)
2398 struct io_submit_state *state = &ctx->submit_state;
2400 if (unlikely(state->link.head))
2401 io_queue_sqe_fallback(state->link.head);
2402 /* flush only after queuing links as they can generate completions */
2403 io_submit_flush_completions(ctx);
2404 if (state->plug_started)
2405 blk_finish_plug(&state->plug);
2409 * Start submission side cache.
2411 static void io_submit_state_start(struct io_submit_state *state,
2412 unsigned int max_ios)
2414 state->plug_started = false;
2415 state->need_plug = max_ios > 2;
2416 state->submit_nr = max_ios;
2417 /* set only head, no need to init link_last in advance */
2418 state->link.head = NULL;
2421 static void io_commit_sqring(struct io_ring_ctx *ctx)
2423 struct io_rings *rings = ctx->rings;
2426 * Ensure any loads from the SQEs are done at this point,
2427 * since once we write the new head, the application could
2428 * write new data to them.
2430 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2434 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2435 * that is mapped by userspace. This means that care needs to be taken to
2436 * ensure that reads are stable, as we cannot rely on userspace always
2437 * being a good citizen. If members of the sqe are validated and then later
2438 * used, it's important that those reads are done through READ_ONCE() to
2439 * prevent a re-load down the line.
2441 static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe)
2443 unsigned head, mask = ctx->sq_entries - 1;
2444 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2447 * The cached sq head (or cq tail) serves two purposes:
2449 * 1) allows us to batch the cost of updating the user visible
2451 * 2) allows the kernel side to track the head on its own, even
2452 * though the application is the one updating it.
2454 head = READ_ONCE(ctx->sq_array[sq_idx]);
2455 if (likely(head < ctx->sq_entries)) {
2456 /* double index for 128-byte SQEs, twice as long */
2457 if (ctx->flags & IORING_SETUP_SQE128)
2459 *sqe = &ctx->sq_sqes[head];
2463 /* drop invalid entries */
2465 WRITE_ONCE(ctx->rings->sq_dropped,
2466 READ_ONCE(ctx->rings->sq_dropped) + 1);
2470 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2471 __must_hold(&ctx->uring_lock)
2473 unsigned int entries = io_sqring_entries(ctx);
2477 if (unlikely(!entries))
2479 /* make sure SQ entry isn't read before tail */
2480 ret = left = min(nr, entries);
2481 io_get_task_refs(left);
2482 io_submit_state_start(&ctx->submit_state, left);
2485 const struct io_uring_sqe *sqe;
2486 struct io_kiocb *req;
2488 if (unlikely(!io_alloc_req(ctx, &req)))
2490 if (unlikely(!io_get_sqe(ctx, &sqe))) {
2491 io_req_add_to_cache(req, ctx);
2496 * Continue submitting even for sqe failure if the
2497 * ring was setup with IORING_SETUP_SUBMIT_ALL
2499 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2500 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2506 if (unlikely(left)) {
2508 /* try again if it submitted nothing and can't allocate a req */
2509 if (!ret && io_req_cache_empty(ctx))
2511 current->io_uring->cached_refs += left;
2514 io_submit_state_end(ctx);
2515 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2516 io_commit_sqring(ctx);
2520 struct io_wait_queue {
2521 struct wait_queue_entry wq;
2522 struct io_ring_ctx *ctx;
2524 unsigned nr_timeouts;
2528 static inline bool io_has_work(struct io_ring_ctx *ctx)
2530 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
2531 !llist_empty(&ctx->work_llist);
2534 static inline bool io_should_wake(struct io_wait_queue *iowq)
2536 struct io_ring_ctx *ctx = iowq->ctx;
2537 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
2540 * Wake up if we have enough events, or if a timeout occurred since we
2541 * started waiting. For timeouts, we always want to return to userspace,
2542 * regardless of event count.
2544 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2547 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2548 int wake_flags, void *key)
2550 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq);
2553 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2554 * the task, and the next invocation will do it.
2556 if (io_should_wake(iowq) || io_has_work(iowq->ctx))
2557 return autoremove_wake_function(curr, mode, wake_flags, key);
2561 int io_run_task_work_sig(struct io_ring_ctx *ctx)
2563 if (!llist_empty(&ctx->work_llist)) {
2564 __set_current_state(TASK_RUNNING);
2565 if (io_run_local_work(ctx) > 0)
2568 if (io_run_task_work() > 0)
2570 if (task_sigpending(current))
2575 /* when returns >0, the caller should retry */
2576 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2577 struct io_wait_queue *iowq)
2579 if (unlikely(READ_ONCE(ctx->check_cq)))
2581 if (unlikely(!llist_empty(&ctx->work_llist)))
2583 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL)))
2585 if (unlikely(task_sigpending(current)))
2587 if (unlikely(io_should_wake(iowq)))
2589 if (iowq->timeout == KTIME_MAX)
2591 else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS))
2597 * Wait until events become available, if we don't already have some. The
2598 * application must reap them itself, as they reside on the shared cq ring.
2600 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2601 const sigset_t __user *sig, size_t sigsz,
2602 struct __kernel_timespec __user *uts)
2604 struct io_wait_queue iowq;
2605 struct io_rings *rings = ctx->rings;
2608 if (!io_allowed_run_tw(ctx))
2610 if (!llist_empty(&ctx->work_llist))
2611 io_run_local_work(ctx);
2613 io_cqring_overflow_flush(ctx);
2614 /* if user messes with these they will just get an early return */
2615 if (__io_cqring_events_user(ctx) >= min_events)
2619 #ifdef CONFIG_COMPAT
2620 if (in_compat_syscall())
2621 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2625 ret = set_user_sigmask(sig, sigsz);
2631 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2632 iowq.wq.private = current;
2633 INIT_LIST_HEAD(&iowq.wq.entry);
2635 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2636 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2637 iowq.timeout = KTIME_MAX;
2640 struct timespec64 ts;
2642 if (get_timespec64(&ts, uts))
2644 iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2647 trace_io_uring_cqring_wait(ctx, min_events);
2649 unsigned long check_cq;
2651 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
2652 int nr_wait = (int) iowq.cq_tail - READ_ONCE(ctx->rings->cq.tail);
2654 atomic_set(&ctx->cq_wait_nr, nr_wait);
2655 set_current_state(TASK_INTERRUPTIBLE);
2657 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2658 TASK_INTERRUPTIBLE);
2661 ret = io_cqring_wait_schedule(ctx, &iowq);
2662 __set_current_state(TASK_RUNNING);
2663 atomic_set(&ctx->cq_wait_nr, 0);
2668 * Run task_work after scheduling and before io_should_wake().
2669 * If we got woken because of task_work being processed, run it
2670 * now rather than let the caller do another wait loop.
2673 if (!llist_empty(&ctx->work_llist))
2674 io_run_local_work(ctx);
2676 check_cq = READ_ONCE(ctx->check_cq);
2677 if (unlikely(check_cq)) {
2678 /* let the caller flush overflows, retry */
2679 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2680 io_cqring_do_overflow_flush(ctx);
2681 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) {
2687 if (io_should_wake(&iowq)) {
2694 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
2695 finish_wait(&ctx->cq_wait, &iowq.wq);
2696 restore_saved_sigmask_unless(ret == -EINTR);
2698 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2701 static void io_mem_free(void *ptr)
2708 page = virt_to_head_page(ptr);
2709 if (put_page_testzero(page))
2710 free_compound_page(page);
2713 static void *io_mem_alloc(size_t size)
2715 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2717 return (void *) __get_free_pages(gfp, get_order(size));
2720 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2721 unsigned int cq_entries, size_t *sq_offset)
2723 struct io_rings *rings;
2724 size_t off, sq_array_size;
2726 off = struct_size(rings, cqes, cq_entries);
2727 if (off == SIZE_MAX)
2729 if (ctx->flags & IORING_SETUP_CQE32) {
2730 if (check_shl_overflow(off, 1, &off))
2735 off = ALIGN(off, SMP_CACHE_BYTES);
2743 sq_array_size = array_size(sizeof(u32), sq_entries);
2744 if (sq_array_size == SIZE_MAX)
2747 if (check_add_overflow(off, sq_array_size, &off))
2753 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2754 unsigned int eventfd_async)
2756 struct io_ev_fd *ev_fd;
2757 __s32 __user *fds = arg;
2760 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2761 lockdep_is_held(&ctx->uring_lock));
2765 if (copy_from_user(&fd, fds, sizeof(*fds)))
2768 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2772 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2773 if (IS_ERR(ev_fd->cq_ev_fd)) {
2774 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2779 spin_lock(&ctx->completion_lock);
2780 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2781 spin_unlock(&ctx->completion_lock);
2783 ev_fd->eventfd_async = eventfd_async;
2784 ctx->has_evfd = true;
2785 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2786 atomic_set(&ev_fd->refs, 1);
2787 atomic_set(&ev_fd->ops, 0);
2791 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2793 struct io_ev_fd *ev_fd;
2795 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2796 lockdep_is_held(&ctx->uring_lock));
2798 ctx->has_evfd = false;
2799 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2800 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops))
2801 call_rcu(&ev_fd->rcu, io_eventfd_ops);
2808 static void io_req_caches_free(struct io_ring_ctx *ctx)
2810 struct io_kiocb *req;
2813 mutex_lock(&ctx->uring_lock);
2814 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
2816 while (!io_req_cache_empty(ctx)) {
2817 req = io_extract_req(ctx);
2818 kmem_cache_free(req_cachep, req);
2822 percpu_ref_put_many(&ctx->refs, nr);
2823 mutex_unlock(&ctx->uring_lock);
2826 static void io_rsrc_node_cache_free(struct io_cache_entry *entry)
2828 kfree(container_of(entry, struct io_rsrc_node, cache));
2831 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2833 io_sq_thread_finish(ctx);
2834 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2835 io_wait_rsrc_data(ctx->buf_data);
2836 io_wait_rsrc_data(ctx->file_data);
2838 mutex_lock(&ctx->uring_lock);
2840 __io_sqe_buffers_unregister(ctx);
2842 __io_sqe_files_unregister(ctx);
2843 io_cqring_overflow_kill(ctx);
2844 io_eventfd_unregister(ctx);
2845 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2846 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2847 mutex_unlock(&ctx->uring_lock);
2848 io_destroy_buffers(ctx);
2850 put_cred(ctx->sq_creds);
2851 if (ctx->submitter_task)
2852 put_task_struct(ctx->submitter_task);
2854 /* there are no registered resources left, nobody uses it */
2856 io_rsrc_node_destroy(ctx, ctx->rsrc_node);
2857 if (ctx->rsrc_backup_node)
2858 io_rsrc_node_destroy(ctx, ctx->rsrc_backup_node);
2860 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2862 #if defined(CONFIG_UNIX)
2863 if (ctx->ring_sock) {
2864 ctx->ring_sock->file = NULL; /* so that iput() is called */
2865 sock_release(ctx->ring_sock);
2868 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2870 io_alloc_cache_free(&ctx->rsrc_node_cache, io_rsrc_node_cache_free);
2871 if (ctx->mm_account) {
2872 mmdrop(ctx->mm_account);
2873 ctx->mm_account = NULL;
2875 io_mem_free(ctx->rings);
2876 io_mem_free(ctx->sq_sqes);
2878 percpu_ref_exit(&ctx->refs);
2879 free_uid(ctx->user);
2880 io_req_caches_free(ctx);
2882 io_wq_put_hash(ctx->hash_map);
2883 kfree(ctx->cancel_table.hbs);
2884 kfree(ctx->cancel_table_locked.hbs);
2885 kfree(ctx->dummy_ubuf);
2887 xa_destroy(&ctx->io_bl_xa);
2891 static __cold void io_activate_pollwq_cb(struct callback_head *cb)
2893 struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx,
2896 mutex_lock(&ctx->uring_lock);
2897 ctx->poll_activated = true;
2898 mutex_unlock(&ctx->uring_lock);
2901 * Wake ups for some events between start of polling and activation
2902 * might've been lost due to loose synchronisation.
2904 wake_up_all(&ctx->poll_wq);
2905 percpu_ref_put(&ctx->refs);
2908 static __cold void io_activate_pollwq(struct io_ring_ctx *ctx)
2910 spin_lock(&ctx->completion_lock);
2911 /* already activated or in progress */
2912 if (ctx->poll_activated || ctx->poll_wq_task_work.func)
2914 if (WARN_ON_ONCE(!ctx->task_complete))
2916 if (!ctx->submitter_task)
2919 * with ->submitter_task only the submitter task completes requests, we
2920 * only need to sync with it, which is done by injecting a tw
2922 init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb);
2923 percpu_ref_get(&ctx->refs);
2924 if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL))
2925 percpu_ref_put(&ctx->refs);
2927 spin_unlock(&ctx->completion_lock);
2930 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2932 struct io_ring_ctx *ctx = file->private_data;
2935 if (unlikely(!ctx->poll_activated))
2936 io_activate_pollwq(ctx);
2938 poll_wait(file, &ctx->poll_wq, wait);
2940 * synchronizes with barrier from wq_has_sleeper call in
2944 if (!io_sqring_full(ctx))
2945 mask |= EPOLLOUT | EPOLLWRNORM;
2948 * Don't flush cqring overflow list here, just do a simple check.
2949 * Otherwise there could possible be ABBA deadlock:
2952 * lock(&ctx->uring_lock);
2954 * lock(&ctx->uring_lock);
2957 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2958 * pushes them to do the flush.
2961 if (__io_cqring_events_user(ctx) || io_has_work(ctx))
2962 mask |= EPOLLIN | EPOLLRDNORM;
2967 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2969 const struct cred *creds;
2971 creds = xa_erase(&ctx->personalities, id);
2980 struct io_tctx_exit {
2981 struct callback_head task_work;
2982 struct completion completion;
2983 struct io_ring_ctx *ctx;
2986 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2988 struct io_uring_task *tctx = current->io_uring;
2989 struct io_tctx_exit *work;
2991 work = container_of(cb, struct io_tctx_exit, task_work);
2993 * When @in_cancel, we're in cancellation and it's racy to remove the
2994 * node. It'll be removed by the end of cancellation, just ignore it.
2995 * tctx can be NULL if the queueing of this task_work raced with
2996 * work cancelation off the exec path.
2998 if (tctx && !atomic_read(&tctx->in_cancel))
2999 io_uring_del_tctx_node((unsigned long)work->ctx);
3000 complete(&work->completion);
3003 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
3005 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3007 return req->ctx == data;
3010 static __cold void io_ring_exit_work(struct work_struct *work)
3012 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
3013 unsigned long timeout = jiffies + HZ * 60 * 5;
3014 unsigned long interval = HZ / 20;
3015 struct io_tctx_exit exit;
3016 struct io_tctx_node *node;
3020 * If we're doing polled IO and end up having requests being
3021 * submitted async (out-of-line), then completions can come in while
3022 * we're waiting for refs to drop. We need to reap these manually,
3023 * as nobody else will be looking for them.
3026 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
3027 mutex_lock(&ctx->uring_lock);
3028 io_cqring_overflow_kill(ctx);
3029 mutex_unlock(&ctx->uring_lock);
3032 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3033 io_move_task_work_from_local(ctx);
3035 while (io_uring_try_cancel_requests(ctx, NULL, true))
3039 struct io_sq_data *sqd = ctx->sq_data;
3040 struct task_struct *tsk;
3042 io_sq_thread_park(sqd);
3044 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
3045 io_wq_cancel_cb(tsk->io_uring->io_wq,
3046 io_cancel_ctx_cb, ctx, true);
3047 io_sq_thread_unpark(sqd);
3050 io_req_caches_free(ctx);
3052 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
3053 /* there is little hope left, don't run it too often */
3056 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
3058 init_completion(&exit.completion);
3059 init_task_work(&exit.task_work, io_tctx_exit_cb);
3062 * Some may use context even when all refs and requests have been put,
3063 * and they are free to do so while still holding uring_lock or
3064 * completion_lock, see io_req_task_submit(). Apart from other work,
3065 * this lock/unlock section also waits them to finish.
3067 mutex_lock(&ctx->uring_lock);
3068 while (!list_empty(&ctx->tctx_list)) {
3069 WARN_ON_ONCE(time_after(jiffies, timeout));
3071 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
3073 /* don't spin on a single task if cancellation failed */
3074 list_rotate_left(&ctx->tctx_list);
3075 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
3076 if (WARN_ON_ONCE(ret))
3079 mutex_unlock(&ctx->uring_lock);
3080 wait_for_completion(&exit.completion);
3081 mutex_lock(&ctx->uring_lock);
3083 mutex_unlock(&ctx->uring_lock);
3084 spin_lock(&ctx->completion_lock);
3085 spin_unlock(&ctx->completion_lock);
3087 /* pairs with RCU read section in io_req_local_work_add() */
3088 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3091 io_ring_ctx_free(ctx);
3094 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
3096 unsigned long index;
3097 struct creds *creds;
3099 mutex_lock(&ctx->uring_lock);
3100 percpu_ref_kill(&ctx->refs);
3101 xa_for_each(&ctx->personalities, index, creds)
3102 io_unregister_personality(ctx, index);
3104 io_poll_remove_all(ctx, NULL, true);
3105 mutex_unlock(&ctx->uring_lock);
3108 * If we failed setting up the ctx, we might not have any rings
3109 * and therefore did not submit any requests
3112 io_kill_timeouts(ctx, NULL, true);
3114 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
3116 * Use system_unbound_wq to avoid spawning tons of event kworkers
3117 * if we're exiting a ton of rings at the same time. It just adds
3118 * noise and overhead, there's no discernable change in runtime
3119 * over using system_wq.
3121 queue_work(system_unbound_wq, &ctx->exit_work);
3124 static int io_uring_release(struct inode *inode, struct file *file)
3126 struct io_ring_ctx *ctx = file->private_data;
3128 file->private_data = NULL;
3129 io_ring_ctx_wait_and_kill(ctx);
3133 struct io_task_cancel {
3134 struct task_struct *task;
3138 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
3140 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3141 struct io_task_cancel *cancel = data;
3143 return io_match_task_safe(req, cancel->task, cancel->all);
3146 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
3147 struct task_struct *task,
3150 struct io_defer_entry *de;
3153 spin_lock(&ctx->completion_lock);
3154 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
3155 if (io_match_task_safe(de->req, task, cancel_all)) {
3156 list_cut_position(&list, &ctx->defer_list, &de->list);
3160 spin_unlock(&ctx->completion_lock);
3161 if (list_empty(&list))
3164 while (!list_empty(&list)) {
3165 de = list_first_entry(&list, struct io_defer_entry, list);
3166 list_del_init(&de->list);
3167 io_req_task_queue_fail(de->req, -ECANCELED);
3173 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
3175 struct io_tctx_node *node;
3176 enum io_wq_cancel cret;
3179 mutex_lock(&ctx->uring_lock);
3180 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3181 struct io_uring_task *tctx = node->task->io_uring;
3184 * io_wq will stay alive while we hold uring_lock, because it's
3185 * killed after ctx nodes, which requires to take the lock.
3187 if (!tctx || !tctx->io_wq)
3189 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
3190 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3192 mutex_unlock(&ctx->uring_lock);
3197 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
3198 struct task_struct *task,
3201 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
3202 struct io_uring_task *tctx = task ? task->io_uring : NULL;
3203 enum io_wq_cancel cret;
3206 /* failed during ring init, it couldn't have issued any requests */
3211 ret |= io_uring_try_cancel_iowq(ctx);
3212 } else if (tctx && tctx->io_wq) {
3214 * Cancels requests of all rings, not only @ctx, but
3215 * it's fine as the task is in exit/exec.
3217 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
3219 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3222 /* SQPOLL thread does its own polling */
3223 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
3224 (ctx->sq_data && ctx->sq_data->thread == current)) {
3225 while (!wq_list_empty(&ctx->iopoll_list)) {
3226 io_iopoll_try_reap_events(ctx);
3232 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3233 io_allowed_defer_tw_run(ctx))
3234 ret |= io_run_local_work(ctx) > 0;
3235 ret |= io_cancel_defer_files(ctx, task, cancel_all);
3236 mutex_lock(&ctx->uring_lock);
3237 ret |= io_poll_remove_all(ctx, task, cancel_all);
3238 mutex_unlock(&ctx->uring_lock);
3239 ret |= io_kill_timeouts(ctx, task, cancel_all);
3241 ret |= io_run_task_work() > 0;
3245 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
3248 return atomic_read(&tctx->inflight_tracked);
3249 return percpu_counter_sum(&tctx->inflight);
3253 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3254 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3256 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
3258 struct io_uring_task *tctx = current->io_uring;
3259 struct io_ring_ctx *ctx;
3263 WARN_ON_ONCE(sqd && sqd->thread != current);
3265 if (!current->io_uring)
3268 io_wq_exit_start(tctx->io_wq);
3270 atomic_inc(&tctx->in_cancel);
3274 io_uring_drop_tctx_refs(current);
3275 /* read completions before cancelations */
3276 inflight = tctx_inflight(tctx, !cancel_all);
3281 struct io_tctx_node *node;
3282 unsigned long index;
3284 xa_for_each(&tctx->xa, index, node) {
3285 /* sqpoll task will cancel all its requests */
3286 if (node->ctx->sq_data)
3288 loop |= io_uring_try_cancel_requests(node->ctx,
3289 current, cancel_all);
3292 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3293 loop |= io_uring_try_cancel_requests(ctx,
3303 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3305 io_uring_drop_tctx_refs(current);
3308 * If we've seen completions, retry without waiting. This
3309 * avoids a race where a completion comes in before we did
3310 * prepare_to_wait().
3312 if (inflight == tctx_inflight(tctx, !cancel_all))
3314 finish_wait(&tctx->wait, &wait);
3317 io_uring_clean_tctx(tctx);
3320 * We shouldn't run task_works after cancel, so just leave
3321 * ->in_cancel set for normal exit.
3323 atomic_dec(&tctx->in_cancel);
3324 /* for exec all current's requests should be gone, kill tctx */
3325 __io_uring_free(current);
3329 void __io_uring_cancel(bool cancel_all)
3331 io_uring_cancel_generic(cancel_all, NULL);
3334 static void *io_uring_validate_mmap_request(struct file *file,
3335 loff_t pgoff, size_t sz)
3337 struct io_ring_ctx *ctx = file->private_data;
3338 loff_t offset = pgoff << PAGE_SHIFT;
3342 switch (offset & IORING_OFF_MMAP_MASK) {
3343 case IORING_OFF_SQ_RING:
3344 case IORING_OFF_CQ_RING:
3347 case IORING_OFF_SQES:
3350 case IORING_OFF_PBUF_RING: {
3353 bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT;
3354 mutex_lock(&ctx->uring_lock);
3355 ptr = io_pbuf_get_address(ctx, bgid);
3356 mutex_unlock(&ctx->uring_lock);
3358 return ERR_PTR(-EINVAL);
3362 return ERR_PTR(-EINVAL);
3365 page = virt_to_head_page(ptr);
3366 if (sz > page_size(page))
3367 return ERR_PTR(-EINVAL);
3374 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3376 size_t sz = vma->vm_end - vma->vm_start;
3380 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3382 return PTR_ERR(ptr);
3384 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3385 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3388 static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp,
3389 unsigned long addr, unsigned long len,
3390 unsigned long pgoff, unsigned long flags)
3392 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
3393 struct vm_unmapped_area_info info;
3397 * Do not allow to map to user-provided address to avoid breaking the
3398 * aliasing rules. Userspace is not able to guess the offset address of
3399 * kernel kmalloc()ed memory area.
3404 ptr = io_uring_validate_mmap_request(filp, pgoff, len);
3408 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
3410 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
3411 info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
3413 info.align_mask = PAGE_MASK & (SHM_COLOUR - 1UL);
3415 info.align_mask = PAGE_MASK & (SHMLBA - 1UL);
3417 info.align_offset = (unsigned long) ptr;
3420 * A failed mmap() very likely causes application failure,
3421 * so fall back to the bottom-up function here. This scenario
3422 * can happen with large stack limits and large mmap()
3425 addr = vm_unmapped_area(&info);
3426 if (offset_in_page(addr)) {
3428 info.low_limit = TASK_UNMAPPED_BASE;
3429 info.high_limit = mmap_end;
3430 addr = vm_unmapped_area(&info);
3436 #else /* !CONFIG_MMU */
3438 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3440 return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL;
3443 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3445 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3448 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3449 unsigned long addr, unsigned long len,
3450 unsigned long pgoff, unsigned long flags)
3454 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3456 return PTR_ERR(ptr);
3458 return (unsigned long) ptr;
3461 #endif /* !CONFIG_MMU */
3463 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3465 if (flags & IORING_ENTER_EXT_ARG) {
3466 struct io_uring_getevents_arg arg;
3468 if (argsz != sizeof(arg))
3470 if (copy_from_user(&arg, argp, sizeof(arg)))
3476 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3477 struct __kernel_timespec __user **ts,
3478 const sigset_t __user **sig)
3480 struct io_uring_getevents_arg arg;
3483 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3484 * is just a pointer to the sigset_t.
3486 if (!(flags & IORING_ENTER_EXT_ARG)) {
3487 *sig = (const sigset_t __user *) argp;
3493 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3494 * timespec and sigset_t pointers if good.
3496 if (*argsz != sizeof(arg))
3498 if (copy_from_user(&arg, argp, sizeof(arg)))
3502 *sig = u64_to_user_ptr(arg.sigmask);
3503 *argsz = arg.sigmask_sz;
3504 *ts = u64_to_user_ptr(arg.ts);
3508 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3509 u32, min_complete, u32, flags, const void __user *, argp,
3512 struct io_ring_ctx *ctx;
3516 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3517 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3518 IORING_ENTER_REGISTERED_RING)))
3522 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3523 * need only dereference our task private array to find it.
3525 if (flags & IORING_ENTER_REGISTERED_RING) {
3526 struct io_uring_task *tctx = current->io_uring;
3528 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3530 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3531 f.file = tctx->registered_rings[fd];
3533 if (unlikely(!f.file))
3537 if (unlikely(!f.file))
3540 if (unlikely(!io_is_uring_fops(f.file)))
3544 ctx = f.file->private_data;
3546 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3550 * For SQ polling, the thread will do all submissions and completions.
3551 * Just return the requested submit count, and wake the thread if
3555 if (ctx->flags & IORING_SETUP_SQPOLL) {
3556 io_cqring_overflow_flush(ctx);
3558 if (unlikely(ctx->sq_data->thread == NULL)) {
3562 if (flags & IORING_ENTER_SQ_WAKEUP)
3563 wake_up(&ctx->sq_data->wait);
3564 if (flags & IORING_ENTER_SQ_WAIT)
3565 io_sqpoll_wait_sq(ctx);
3568 } else if (to_submit) {
3569 ret = io_uring_add_tctx_node(ctx);
3573 mutex_lock(&ctx->uring_lock);
3574 ret = io_submit_sqes(ctx, to_submit);
3575 if (ret != to_submit) {
3576 mutex_unlock(&ctx->uring_lock);
3579 if (flags & IORING_ENTER_GETEVENTS) {
3580 if (ctx->syscall_iopoll)
3583 * Ignore errors, we'll soon call io_cqring_wait() and
3584 * it should handle ownership problems if any.
3586 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3587 (void)io_run_local_work_locked(ctx);
3589 mutex_unlock(&ctx->uring_lock);
3592 if (flags & IORING_ENTER_GETEVENTS) {
3595 if (ctx->syscall_iopoll) {
3597 * We disallow the app entering submit/complete with
3598 * polling, but we still need to lock the ring to
3599 * prevent racing with polled issue that got punted to
3602 mutex_lock(&ctx->uring_lock);
3604 ret2 = io_validate_ext_arg(flags, argp, argsz);
3605 if (likely(!ret2)) {
3606 min_complete = min(min_complete,
3608 ret2 = io_iopoll_check(ctx, min_complete);
3610 mutex_unlock(&ctx->uring_lock);
3612 const sigset_t __user *sig;
3613 struct __kernel_timespec __user *ts;
3615 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3616 if (likely(!ret2)) {
3617 min_complete = min(min_complete,
3619 ret2 = io_cqring_wait(ctx, min_complete, sig,
3628 * EBADR indicates that one or more CQE were dropped.
3629 * Once the user has been informed we can clear the bit
3630 * as they are obviously ok with those drops.
3632 if (unlikely(ret2 == -EBADR))
3633 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3642 static const struct file_operations io_uring_fops = {
3643 .release = io_uring_release,
3644 .mmap = io_uring_mmap,
3646 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3647 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3649 .get_unmapped_area = io_uring_mmu_get_unmapped_area,
3651 .poll = io_uring_poll,
3652 #ifdef CONFIG_PROC_FS
3653 .show_fdinfo = io_uring_show_fdinfo,
3657 bool io_is_uring_fops(struct file *file)
3659 return file->f_op == &io_uring_fops;
3662 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3663 struct io_uring_params *p)
3665 struct io_rings *rings;
3666 size_t size, sq_array_offset;
3668 /* make sure these are sane, as we already accounted them */
3669 ctx->sq_entries = p->sq_entries;
3670 ctx->cq_entries = p->cq_entries;
3672 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3673 if (size == SIZE_MAX)
3676 rings = io_mem_alloc(size);
3681 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3682 rings->sq_ring_mask = p->sq_entries - 1;
3683 rings->cq_ring_mask = p->cq_entries - 1;
3684 rings->sq_ring_entries = p->sq_entries;
3685 rings->cq_ring_entries = p->cq_entries;
3687 if (p->flags & IORING_SETUP_SQE128)
3688 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3690 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3691 if (size == SIZE_MAX) {
3692 io_mem_free(ctx->rings);
3697 ctx->sq_sqes = io_mem_alloc(size);
3698 if (!ctx->sq_sqes) {
3699 io_mem_free(ctx->rings);
3707 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3711 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3715 ret = __io_uring_add_tctx_node(ctx);
3720 fd_install(fd, file);
3725 * Allocate an anonymous fd, this is what constitutes the application
3726 * visible backing of an io_uring instance. The application mmaps this
3727 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3728 * we have to tie this fd to a socket for file garbage collection purposes.
3730 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3733 #if defined(CONFIG_UNIX)
3736 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3739 return ERR_PTR(ret);
3742 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3743 O_RDWR | O_CLOEXEC, NULL);
3744 #if defined(CONFIG_UNIX)
3746 sock_release(ctx->ring_sock);
3747 ctx->ring_sock = NULL;
3749 ctx->ring_sock->file = file;
3755 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3756 struct io_uring_params __user *params)
3758 struct io_ring_ctx *ctx;
3764 if (entries > IORING_MAX_ENTRIES) {
3765 if (!(p->flags & IORING_SETUP_CLAMP))
3767 entries = IORING_MAX_ENTRIES;
3771 * Use twice as many entries for the CQ ring. It's possible for the
3772 * application to drive a higher depth than the size of the SQ ring,
3773 * since the sqes are only used at submission time. This allows for
3774 * some flexibility in overcommitting a bit. If the application has
3775 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3776 * of CQ ring entries manually.
3778 p->sq_entries = roundup_pow_of_two(entries);
3779 if (p->flags & IORING_SETUP_CQSIZE) {
3781 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3782 * to a power-of-two, if it isn't already. We do NOT impose
3783 * any cq vs sq ring sizing.
3787 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3788 if (!(p->flags & IORING_SETUP_CLAMP))
3790 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3792 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3793 if (p->cq_entries < p->sq_entries)
3796 p->cq_entries = 2 * p->sq_entries;
3799 ctx = io_ring_ctx_alloc(p);
3803 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3804 !(ctx->flags & IORING_SETUP_IOPOLL) &&
3805 !(ctx->flags & IORING_SETUP_SQPOLL))
3806 ctx->task_complete = true;
3809 * lazy poll_wq activation relies on ->task_complete for synchronisation
3810 * purposes, see io_activate_pollwq()
3812 if (!ctx->task_complete)
3813 ctx->poll_activated = true;
3816 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3817 * space applications don't need to do io completion events
3818 * polling again, they can rely on io_sq_thread to do polling
3819 * work, which can reduce cpu usage and uring_lock contention.
3821 if (ctx->flags & IORING_SETUP_IOPOLL &&
3822 !(ctx->flags & IORING_SETUP_SQPOLL))
3823 ctx->syscall_iopoll = 1;
3825 ctx->compat = in_compat_syscall();
3826 if (!capable(CAP_IPC_LOCK))
3827 ctx->user = get_uid(current_user());
3830 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3831 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3834 if (ctx->flags & IORING_SETUP_SQPOLL) {
3835 /* IPI related flags don't make sense with SQPOLL */
3836 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3837 IORING_SETUP_TASKRUN_FLAG |
3838 IORING_SETUP_DEFER_TASKRUN))
3840 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3841 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3842 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3844 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3845 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3847 ctx->notify_method = TWA_SIGNAL;
3851 * For DEFER_TASKRUN we require the completion task to be the same as the
3852 * submission task. This implies that there is only one submitter, so enforce
3855 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3856 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3861 * This is just grabbed for accounting purposes. When a process exits,
3862 * the mm is exited and dropped before the files, hence we need to hang
3863 * on to this mm purely for the purposes of being able to unaccount
3864 * memory (locked/pinned vm). It's not used for anything else.
3866 mmgrab(current->mm);
3867 ctx->mm_account = current->mm;
3869 ret = io_allocate_scq_urings(ctx, p);
3873 ret = io_sq_offload_create(ctx, p);
3876 /* always set a rsrc node */
3877 ret = io_rsrc_node_switch_start(ctx);
3880 io_rsrc_node_switch(ctx, NULL);
3882 memset(&p->sq_off, 0, sizeof(p->sq_off));
3883 p->sq_off.head = offsetof(struct io_rings, sq.head);
3884 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3885 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3886 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3887 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3888 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3889 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3891 memset(&p->cq_off, 0, sizeof(p->cq_off));
3892 p->cq_off.head = offsetof(struct io_rings, cq.head);
3893 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3894 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3895 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3896 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3897 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3898 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3900 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3901 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3902 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3903 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3904 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3905 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3906 IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING;
3908 if (copy_to_user(params, p, sizeof(*p))) {
3913 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
3914 && !(ctx->flags & IORING_SETUP_R_DISABLED))
3915 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
3917 file = io_uring_get_file(ctx);
3919 ret = PTR_ERR(file);
3924 * Install ring fd as the very last thing, so we don't risk someone
3925 * having closed it before we finish setup
3927 ret = io_uring_install_fd(ctx, file);
3929 /* fput will clean it up */
3934 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3937 io_ring_ctx_wait_and_kill(ctx);
3942 * Sets up an aio uring context, and returns the fd. Applications asks for a
3943 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3944 * params structure passed in.
3946 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3948 struct io_uring_params p;
3951 if (copy_from_user(&p, params, sizeof(p)))
3953 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3958 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3959 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3960 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3961 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3962 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3963 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3964 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN))
3967 return io_uring_create(entries, &p, params);
3970 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3971 struct io_uring_params __user *, params)
3973 return io_uring_setup(entries, params);
3976 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3979 struct io_uring_probe *p;
3983 size = struct_size(p, ops, nr_args);
3984 if (size == SIZE_MAX)
3986 p = kzalloc(size, GFP_KERNEL);
3991 if (copy_from_user(p, arg, size))
3994 if (memchr_inv(p, 0, size))
3997 p->last_op = IORING_OP_LAST - 1;
3998 if (nr_args > IORING_OP_LAST)
3999 nr_args = IORING_OP_LAST;
4001 for (i = 0; i < nr_args; i++) {
4003 if (!io_issue_defs[i].not_supported)
4004 p->ops[i].flags = IO_URING_OP_SUPPORTED;
4009 if (copy_to_user(arg, p, size))
4016 static int io_register_personality(struct io_ring_ctx *ctx)
4018 const struct cred *creds;
4022 creds = get_current_cred();
4024 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
4025 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
4033 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
4034 void __user *arg, unsigned int nr_args)
4036 struct io_uring_restriction *res;
4040 /* Restrictions allowed only if rings started disabled */
4041 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
4044 /* We allow only a single restrictions registration */
4045 if (ctx->restrictions.registered)
4048 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
4051 size = array_size(nr_args, sizeof(*res));
4052 if (size == SIZE_MAX)
4055 res = memdup_user(arg, size);
4057 return PTR_ERR(res);
4061 for (i = 0; i < nr_args; i++) {
4062 switch (res[i].opcode) {
4063 case IORING_RESTRICTION_REGISTER_OP:
4064 if (res[i].register_op >= IORING_REGISTER_LAST) {
4069 __set_bit(res[i].register_op,
4070 ctx->restrictions.register_op);
4072 case IORING_RESTRICTION_SQE_OP:
4073 if (res[i].sqe_op >= IORING_OP_LAST) {
4078 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
4080 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
4081 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
4083 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
4084 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
4093 /* Reset all restrictions if an error happened */
4095 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
4097 ctx->restrictions.registered = true;
4103 static int io_register_enable_rings(struct io_ring_ctx *ctx)
4105 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
4108 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task) {
4109 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
4111 * Lazy activation attempts would fail if it was polled before
4112 * submitter_task is set.
4114 if (wq_has_sleeper(&ctx->poll_wq))
4115 io_activate_pollwq(ctx);
4118 if (ctx->restrictions.registered)
4119 ctx->restricted = 1;
4121 ctx->flags &= ~IORING_SETUP_R_DISABLED;
4122 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
4123 wake_up(&ctx->sq_data->wait);
4127 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
4128 void __user *arg, unsigned len)
4130 struct io_uring_task *tctx = current->io_uring;
4131 cpumask_var_t new_mask;
4134 if (!tctx || !tctx->io_wq)
4137 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4140 cpumask_clear(new_mask);
4141 if (len > cpumask_size())
4142 len = cpumask_size();
4144 if (in_compat_syscall()) {
4145 ret = compat_get_bitmap(cpumask_bits(new_mask),
4146 (const compat_ulong_t __user *)arg,
4147 len * 8 /* CHAR_BIT */);
4149 ret = copy_from_user(new_mask, arg, len);
4153 free_cpumask_var(new_mask);
4157 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
4158 free_cpumask_var(new_mask);
4162 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
4164 struct io_uring_task *tctx = current->io_uring;
4166 if (!tctx || !tctx->io_wq)
4169 return io_wq_cpu_affinity(tctx->io_wq, NULL);
4172 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
4174 __must_hold(&ctx->uring_lock)
4176 struct io_tctx_node *node;
4177 struct io_uring_task *tctx = NULL;
4178 struct io_sq_data *sqd = NULL;
4182 if (copy_from_user(new_count, arg, sizeof(new_count)))
4184 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4185 if (new_count[i] > INT_MAX)
4188 if (ctx->flags & IORING_SETUP_SQPOLL) {
4192 * Observe the correct sqd->lock -> ctx->uring_lock
4193 * ordering. Fine to drop uring_lock here, we hold
4196 refcount_inc(&sqd->refs);
4197 mutex_unlock(&ctx->uring_lock);
4198 mutex_lock(&sqd->lock);
4199 mutex_lock(&ctx->uring_lock);
4201 tctx = sqd->thread->io_uring;
4204 tctx = current->io_uring;
4207 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
4209 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4211 ctx->iowq_limits[i] = new_count[i];
4212 ctx->iowq_limits_set = true;
4214 if (tctx && tctx->io_wq) {
4215 ret = io_wq_max_workers(tctx->io_wq, new_count);
4219 memset(new_count, 0, sizeof(new_count));
4223 mutex_unlock(&sqd->lock);
4224 io_put_sq_data(sqd);
4227 if (copy_to_user(arg, new_count, sizeof(new_count)))
4230 /* that's it for SQPOLL, only the SQPOLL task creates requests */
4234 /* now propagate the restriction to all registered users */
4235 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
4236 struct io_uring_task *tctx = node->task->io_uring;
4238 if (WARN_ON_ONCE(!tctx->io_wq))
4241 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4242 new_count[i] = ctx->iowq_limits[i];
4243 /* ignore errors, it always returns zero anyway */
4244 (void)io_wq_max_workers(tctx->io_wq, new_count);
4249 mutex_unlock(&sqd->lock);
4250 io_put_sq_data(sqd);
4255 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
4256 void __user *arg, unsigned nr_args)
4257 __releases(ctx->uring_lock)
4258 __acquires(ctx->uring_lock)
4263 * We don't quiesce the refs for register anymore and so it can't be
4264 * dying as we're holding a file ref here.
4266 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
4269 if (ctx->submitter_task && ctx->submitter_task != current)
4272 if (ctx->restricted) {
4273 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
4274 if (!test_bit(opcode, ctx->restrictions.register_op))
4279 case IORING_REGISTER_BUFFERS:
4283 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
4285 case IORING_UNREGISTER_BUFFERS:
4289 ret = io_sqe_buffers_unregister(ctx);
4291 case IORING_REGISTER_FILES:
4295 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
4297 case IORING_UNREGISTER_FILES:
4301 ret = io_sqe_files_unregister(ctx);
4303 case IORING_REGISTER_FILES_UPDATE:
4304 ret = io_register_files_update(ctx, arg, nr_args);
4306 case IORING_REGISTER_EVENTFD:
4310 ret = io_eventfd_register(ctx, arg, 0);
4312 case IORING_REGISTER_EVENTFD_ASYNC:
4316 ret = io_eventfd_register(ctx, arg, 1);
4318 case IORING_UNREGISTER_EVENTFD:
4322 ret = io_eventfd_unregister(ctx);
4324 case IORING_REGISTER_PROBE:
4326 if (!arg || nr_args > 256)
4328 ret = io_probe(ctx, arg, nr_args);
4330 case IORING_REGISTER_PERSONALITY:
4334 ret = io_register_personality(ctx);
4336 case IORING_UNREGISTER_PERSONALITY:
4340 ret = io_unregister_personality(ctx, nr_args);
4342 case IORING_REGISTER_ENABLE_RINGS:
4346 ret = io_register_enable_rings(ctx);
4348 case IORING_REGISTER_RESTRICTIONS:
4349 ret = io_register_restrictions(ctx, arg, nr_args);
4351 case IORING_REGISTER_FILES2:
4352 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
4354 case IORING_REGISTER_FILES_UPDATE2:
4355 ret = io_register_rsrc_update(ctx, arg, nr_args,
4358 case IORING_REGISTER_BUFFERS2:
4359 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
4361 case IORING_REGISTER_BUFFERS_UPDATE:
4362 ret = io_register_rsrc_update(ctx, arg, nr_args,
4363 IORING_RSRC_BUFFER);
4365 case IORING_REGISTER_IOWQ_AFF:
4367 if (!arg || !nr_args)
4369 ret = io_register_iowq_aff(ctx, arg, nr_args);
4371 case IORING_UNREGISTER_IOWQ_AFF:
4375 ret = io_unregister_iowq_aff(ctx);
4377 case IORING_REGISTER_IOWQ_MAX_WORKERS:
4379 if (!arg || nr_args != 2)
4381 ret = io_register_iowq_max_workers(ctx, arg);
4383 case IORING_REGISTER_RING_FDS:
4384 ret = io_ringfd_register(ctx, arg, nr_args);
4386 case IORING_UNREGISTER_RING_FDS:
4387 ret = io_ringfd_unregister(ctx, arg, nr_args);
4389 case IORING_REGISTER_PBUF_RING:
4391 if (!arg || nr_args != 1)
4393 ret = io_register_pbuf_ring(ctx, arg);
4395 case IORING_UNREGISTER_PBUF_RING:
4397 if (!arg || nr_args != 1)
4399 ret = io_unregister_pbuf_ring(ctx, arg);
4401 case IORING_REGISTER_SYNC_CANCEL:
4403 if (!arg || nr_args != 1)
4405 ret = io_sync_cancel(ctx, arg);
4407 case IORING_REGISTER_FILE_ALLOC_RANGE:
4409 if (!arg || nr_args)
4411 ret = io_register_file_alloc_range(ctx, arg);
4421 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
4422 void __user *, arg, unsigned int, nr_args)
4424 struct io_ring_ctx *ctx;
4427 bool use_registered_ring;
4429 use_registered_ring = !!(opcode & IORING_REGISTER_USE_REGISTERED_RING);
4430 opcode &= ~IORING_REGISTER_USE_REGISTERED_RING;
4432 if (opcode >= IORING_REGISTER_LAST)
4435 if (use_registered_ring) {
4437 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
4438 * need only dereference our task private array to find it.
4440 struct io_uring_task *tctx = current->io_uring;
4442 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
4444 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
4445 f.file = tctx->registered_rings[fd];
4447 if (unlikely(!f.file))
4451 if (unlikely(!f.file))
4454 if (!io_is_uring_fops(f.file))
4458 ctx = f.file->private_data;
4460 mutex_lock(&ctx->uring_lock);
4461 ret = __io_uring_register(ctx, opcode, arg, nr_args);
4462 mutex_unlock(&ctx->uring_lock);
4463 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
4469 static int __init io_uring_init(void)
4471 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4472 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4473 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4476 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4477 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4478 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4479 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4480 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4481 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
4482 BUILD_BUG_SQE_ELEM(1, __u8, flags);
4483 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
4484 BUILD_BUG_SQE_ELEM(4, __s32, fd);
4485 BUILD_BUG_SQE_ELEM(8, __u64, off);
4486 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
4487 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
4488 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
4489 BUILD_BUG_SQE_ELEM(16, __u64, addr);
4490 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
4491 BUILD_BUG_SQE_ELEM(24, __u32, len);
4492 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
4493 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
4494 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4495 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
4496 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
4497 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
4498 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
4499 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
4500 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
4501 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
4502 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
4503 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
4504 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
4505 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
4506 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
4507 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
4508 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
4509 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
4510 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
4511 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
4512 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
4513 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
4514 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
4515 BUILD_BUG_SQE_ELEM(42, __u16, personality);
4516 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
4517 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
4518 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
4519 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
4520 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
4521 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4522 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
4524 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4525 sizeof(struct io_uring_rsrc_update));
4526 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4527 sizeof(struct io_uring_rsrc_update2));
4529 /* ->buf_index is u16 */
4530 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4531 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4532 offsetof(struct io_uring_buf_ring, tail));
4534 /* should fit into one byte */
4535 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
4536 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4537 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4539 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
4541 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4543 io_uring_optable_init();
4545 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4546 SLAB_ACCOUNT | SLAB_TYPESAFE_BY_RCU);
4549 __initcall(io_uring_init);