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>
99 #include "alloc_cache.h"
101 #define IORING_MAX_ENTRIES 32768
102 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
104 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
105 IORING_REGISTER_LAST + IORING_OP_LAST)
107 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
108 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
110 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
111 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
113 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
114 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
117 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
120 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
122 #define IO_COMPL_BATCH 32
123 #define IO_REQ_ALLOC_BATCH 8
126 IO_CHECK_CQ_OVERFLOW_BIT,
127 IO_CHECK_CQ_DROPPED_BIT,
131 IO_EVENTFD_OP_SIGNAL_BIT,
132 IO_EVENTFD_OP_FREE_BIT,
135 struct io_defer_entry {
136 struct list_head list;
137 struct io_kiocb *req;
141 /* requests with any of those set should undergo io_disarm_next() */
142 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
143 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
145 static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
146 struct task_struct *task,
149 static void io_queue_sqe(struct io_kiocb *req);
150 static void io_move_task_work_from_local(struct io_ring_ctx *ctx);
151 static void __io_submit_flush_completions(struct io_ring_ctx *ctx);
152 static __cold void io_fallback_tw(struct io_uring_task *tctx);
154 struct kmem_cache *req_cachep;
156 struct sock *io_uring_get_socket(struct file *file)
158 #if defined(CONFIG_UNIX)
159 if (io_is_uring_fops(file)) {
160 struct io_ring_ctx *ctx = file->private_data;
162 return ctx->ring_sock->sk;
167 EXPORT_SYMBOL(io_uring_get_socket);
169 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
171 if (!wq_list_empty(&ctx->submit_state.compl_reqs) ||
172 ctx->submit_state.cqes_count)
173 __io_submit_flush_completions(ctx);
176 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
178 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
181 static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx)
183 return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head);
186 static bool io_match_linked(struct io_kiocb *head)
188 struct io_kiocb *req;
190 io_for_each_link(req, head) {
191 if (req->flags & REQ_F_INFLIGHT)
198 * As io_match_task() but protected against racing with linked timeouts.
199 * User must not hold timeout_lock.
201 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
206 if (task && head->task != task)
211 if (head->flags & REQ_F_LINK_TIMEOUT) {
212 struct io_ring_ctx *ctx = head->ctx;
214 /* protect against races with linked timeouts */
215 spin_lock_irq(&ctx->timeout_lock);
216 matched = io_match_linked(head);
217 spin_unlock_irq(&ctx->timeout_lock);
219 matched = io_match_linked(head);
224 static inline void req_fail_link_node(struct io_kiocb *req, int res)
227 io_req_set_res(req, res, 0);
230 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
232 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
233 kasan_poison_object_data(req_cachep, req);
236 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
238 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
240 complete(&ctx->ref_comp);
243 static __cold void io_fallback_req_func(struct work_struct *work)
245 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
247 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
248 struct io_kiocb *req, *tmp;
249 struct io_tw_state ts = { .locked = true, };
251 mutex_lock(&ctx->uring_lock);
252 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
253 req->io_task_work.func(req, &ts);
254 if (WARN_ON_ONCE(!ts.locked))
256 io_submit_flush_completions(ctx);
257 mutex_unlock(&ctx->uring_lock);
260 static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
262 unsigned hash_buckets = 1U << bits;
263 size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
265 table->hbs = kmalloc(hash_size, GFP_KERNEL);
269 table->hash_bits = bits;
270 init_hash_table(table, hash_buckets);
274 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
276 struct io_ring_ctx *ctx;
279 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
283 xa_init(&ctx->io_bl_xa);
286 * Use 5 bits less than the max cq entries, that should give us around
287 * 32 entries per hash list if totally full and uniformly spread, but
288 * don't keep too many buckets to not overconsume memory.
290 hash_bits = ilog2(p->cq_entries) - 5;
291 hash_bits = clamp(hash_bits, 1, 8);
292 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
294 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
297 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
298 if (!ctx->dummy_ubuf)
300 /* set invalid range, so io_import_fixed() fails meeting it */
301 ctx->dummy_ubuf->ubuf = -1UL;
303 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
307 ctx->flags = p->flags;
308 init_waitqueue_head(&ctx->sqo_sq_wait);
309 INIT_LIST_HEAD(&ctx->sqd_list);
310 INIT_LIST_HEAD(&ctx->cq_overflow_list);
311 INIT_LIST_HEAD(&ctx->io_buffers_cache);
312 io_alloc_cache_init(&ctx->rsrc_node_cache, IO_NODE_ALLOC_CACHE_MAX,
313 sizeof(struct io_rsrc_node));
314 io_alloc_cache_init(&ctx->apoll_cache, IO_ALLOC_CACHE_MAX,
315 sizeof(struct async_poll));
316 io_alloc_cache_init(&ctx->netmsg_cache, IO_ALLOC_CACHE_MAX,
317 sizeof(struct io_async_msghdr));
318 init_completion(&ctx->ref_comp);
319 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
320 mutex_init(&ctx->uring_lock);
321 init_waitqueue_head(&ctx->cq_wait);
322 init_waitqueue_head(&ctx->poll_wq);
323 init_waitqueue_head(&ctx->rsrc_quiesce_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 void io_clean_op(struct io_kiocb *req)
371 if (req->flags & REQ_F_BUFFER_SELECTED) {
372 spin_lock(&req->ctx->completion_lock);
373 io_put_kbuf_comp(req);
374 spin_unlock(&req->ctx->completion_lock);
377 if (req->flags & REQ_F_NEED_CLEANUP) {
378 const struct io_cold_def *def = &io_cold_defs[req->opcode];
383 if ((req->flags & REQ_F_POLLED) && req->apoll) {
384 kfree(req->apoll->double_poll);
388 if (req->flags & REQ_F_INFLIGHT) {
389 struct io_uring_task *tctx = req->task->io_uring;
391 atomic_dec(&tctx->inflight_tracked);
393 if (req->flags & REQ_F_CREDS)
394 put_cred(req->creds);
395 if (req->flags & REQ_F_ASYNC_DATA) {
396 kfree(req->async_data);
397 req->async_data = NULL;
399 req->flags &= ~IO_REQ_CLEAN_FLAGS;
402 static inline void io_req_track_inflight(struct io_kiocb *req)
404 if (!(req->flags & REQ_F_INFLIGHT)) {
405 req->flags |= REQ_F_INFLIGHT;
406 atomic_inc(&req->task->io_uring->inflight_tracked);
410 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
412 if (WARN_ON_ONCE(!req->link))
415 req->flags &= ~REQ_F_ARM_LTIMEOUT;
416 req->flags |= REQ_F_LINK_TIMEOUT;
418 /* linked timeouts should have two refs once prep'ed */
419 io_req_set_refcount(req);
420 __io_req_set_refcount(req->link, 2);
424 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
426 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
428 return __io_prep_linked_timeout(req);
431 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
433 io_queue_linked_timeout(__io_prep_linked_timeout(req));
436 static inline void io_arm_ltimeout(struct io_kiocb *req)
438 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
439 __io_arm_ltimeout(req);
442 static void io_prep_async_work(struct io_kiocb *req)
444 const struct io_issue_def *def = &io_issue_defs[req->opcode];
445 struct io_ring_ctx *ctx = req->ctx;
447 if (!(req->flags & REQ_F_CREDS)) {
448 req->flags |= REQ_F_CREDS;
449 req->creds = get_current_cred();
452 req->work.list.next = NULL;
454 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
455 if (req->flags & REQ_F_FORCE_ASYNC)
456 req->work.flags |= IO_WQ_WORK_CONCURRENT;
458 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
459 req->flags |= io_file_get_flags(req->file);
461 if (req->file && (req->flags & REQ_F_ISREG)) {
462 bool should_hash = def->hash_reg_file;
464 /* don't serialize this request if the fs doesn't need it */
465 if (should_hash && (req->file->f_flags & O_DIRECT) &&
466 (req->file->f_mode & FMODE_DIO_PARALLEL_WRITE))
468 if (should_hash || (ctx->flags & IORING_SETUP_IOPOLL))
469 io_wq_hash_work(&req->work, file_inode(req->file));
470 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
471 if (def->unbound_nonreg_file)
472 req->work.flags |= IO_WQ_WORK_UNBOUND;
476 static void io_prep_async_link(struct io_kiocb *req)
478 struct io_kiocb *cur;
480 if (req->flags & REQ_F_LINK_TIMEOUT) {
481 struct io_ring_ctx *ctx = req->ctx;
483 spin_lock_irq(&ctx->timeout_lock);
484 io_for_each_link(cur, req)
485 io_prep_async_work(cur);
486 spin_unlock_irq(&ctx->timeout_lock);
488 io_for_each_link(cur, req)
489 io_prep_async_work(cur);
493 void io_queue_iowq(struct io_kiocb *req, struct io_tw_state *ts_dont_use)
495 struct io_kiocb *link = io_prep_linked_timeout(req);
496 struct io_uring_task *tctx = req->task->io_uring;
499 BUG_ON(!tctx->io_wq);
501 /* init ->work of the whole link before punting */
502 io_prep_async_link(req);
505 * Not expected to happen, but if we do have a bug where this _can_
506 * happen, catch it here and ensure the request is marked as
507 * canceled. That will make io-wq go through the usual work cancel
508 * procedure rather than attempt to run this request (or create a new
511 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
512 req->work.flags |= IO_WQ_WORK_CANCEL;
514 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
515 io_wq_enqueue(tctx->io_wq, &req->work);
517 io_queue_linked_timeout(link);
520 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
522 while (!list_empty(&ctx->defer_list)) {
523 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
524 struct io_defer_entry, list);
526 if (req_need_defer(de->req, de->seq))
528 list_del_init(&de->list);
529 io_req_task_queue(de->req);
535 static void io_eventfd_ops(struct rcu_head *rcu)
537 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
538 int ops = atomic_xchg(&ev_fd->ops, 0);
540 if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT))
541 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
543 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
544 * ordering in a race but if references are 0 we know we have to free
547 if (atomic_dec_and_test(&ev_fd->refs)) {
548 eventfd_ctx_put(ev_fd->cq_ev_fd);
553 static void io_eventfd_signal(struct io_ring_ctx *ctx)
555 struct io_ev_fd *ev_fd = NULL;
559 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
562 ev_fd = rcu_dereference(ctx->io_ev_fd);
565 * Check again if ev_fd exists incase an io_eventfd_unregister call
566 * completed between the NULL check of ctx->io_ev_fd at the start of
567 * the function and rcu_read_lock.
569 if (unlikely(!ev_fd))
571 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
573 if (ev_fd->eventfd_async && !io_wq_current_is_worker())
576 if (likely(eventfd_signal_allowed())) {
577 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
579 atomic_inc(&ev_fd->refs);
580 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops))
581 call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops);
583 atomic_dec(&ev_fd->refs);
590 static void io_eventfd_flush_signal(struct io_ring_ctx *ctx)
594 spin_lock(&ctx->completion_lock);
597 * Eventfd should only get triggered when at least one event has been
598 * posted. Some applications rely on the eventfd notification count
599 * only changing IFF a new CQE has been added to the CQ ring. There's
600 * no depedency on 1:1 relationship between how many times this
601 * function is called (and hence the eventfd count) and number of CQEs
602 * posted to the CQ ring.
604 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
605 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
606 spin_unlock(&ctx->completion_lock);
610 io_eventfd_signal(ctx);
613 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
615 if (ctx->poll_activated)
616 io_poll_wq_wake(ctx);
617 if (ctx->off_timeout_used)
618 io_flush_timeouts(ctx);
619 if (ctx->drain_active) {
620 spin_lock(&ctx->completion_lock);
621 io_queue_deferred(ctx);
622 spin_unlock(&ctx->completion_lock);
625 io_eventfd_flush_signal(ctx);
628 static inline void __io_cq_lock(struct io_ring_ctx *ctx)
630 if (!ctx->task_complete)
631 spin_lock(&ctx->completion_lock);
634 static inline void io_cq_lock(struct io_ring_ctx *ctx)
635 __acquires(ctx->completion_lock)
637 spin_lock(&ctx->completion_lock);
640 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
642 io_commit_cqring(ctx);
644 if (ctx->task_complete) {
646 * ->task_complete implies that only current might be waiting
647 * for CQEs, and obviously, we currently don't. No one is
648 * waiting, wakeups are futile, skip them.
650 io_commit_cqring_flush(ctx);
652 spin_unlock(&ctx->completion_lock);
653 io_commit_cqring_flush(ctx);
658 static void io_cq_unlock_post(struct io_ring_ctx *ctx)
659 __releases(ctx->completion_lock)
661 io_commit_cqring(ctx);
662 spin_unlock(&ctx->completion_lock);
663 io_commit_cqring_flush(ctx);
667 /* Returns true if there are no backlogged entries after the flush */
668 static void io_cqring_overflow_kill(struct io_ring_ctx *ctx)
670 struct io_overflow_cqe *ocqe;
673 spin_lock(&ctx->completion_lock);
674 list_splice_init(&ctx->cq_overflow_list, &list);
675 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
676 spin_unlock(&ctx->completion_lock);
678 while (!list_empty(&list)) {
679 ocqe = list_first_entry(&list, struct io_overflow_cqe, list);
680 list_del(&ocqe->list);
685 static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx)
687 size_t cqe_size = sizeof(struct io_uring_cqe);
689 if (__io_cqring_events(ctx) == ctx->cq_entries)
692 if (ctx->flags & IORING_SETUP_CQE32)
696 while (!list_empty(&ctx->cq_overflow_list)) {
697 struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true);
698 struct io_overflow_cqe *ocqe;
702 ocqe = list_first_entry(&ctx->cq_overflow_list,
703 struct io_overflow_cqe, list);
704 memcpy(cqe, &ocqe->cqe, cqe_size);
705 list_del(&ocqe->list);
709 if (list_empty(&ctx->cq_overflow_list)) {
710 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
711 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
713 io_cq_unlock_post(ctx);
716 static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx)
718 /* iopoll syncs against uring_lock, not completion_lock */
719 if (ctx->flags & IORING_SETUP_IOPOLL)
720 mutex_lock(&ctx->uring_lock);
721 __io_cqring_overflow_flush(ctx);
722 if (ctx->flags & IORING_SETUP_IOPOLL)
723 mutex_unlock(&ctx->uring_lock);
726 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx)
728 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
729 io_cqring_do_overflow_flush(ctx);
732 /* can be called by any task */
733 static void io_put_task_remote(struct task_struct *task)
735 struct io_uring_task *tctx = task->io_uring;
737 percpu_counter_sub(&tctx->inflight, 1);
738 if (unlikely(atomic_read(&tctx->in_cancel)))
739 wake_up(&tctx->wait);
740 put_task_struct(task);
743 /* used by a task to put its own references */
744 static void io_put_task_local(struct task_struct *task)
746 task->io_uring->cached_refs++;
749 /* must to be called somewhat shortly after putting a request */
750 static inline void io_put_task(struct task_struct *task)
752 if (likely(task == current))
753 io_put_task_local(task);
755 io_put_task_remote(task);
758 void io_task_refs_refill(struct io_uring_task *tctx)
760 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
762 percpu_counter_add(&tctx->inflight, refill);
763 refcount_add(refill, ¤t->usage);
764 tctx->cached_refs += refill;
767 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
769 struct io_uring_task *tctx = task->io_uring;
770 unsigned int refs = tctx->cached_refs;
773 tctx->cached_refs = 0;
774 percpu_counter_sub(&tctx->inflight, refs);
775 put_task_struct_many(task, refs);
779 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
780 s32 res, u32 cflags, u64 extra1, u64 extra2)
782 struct io_overflow_cqe *ocqe;
783 size_t ocq_size = sizeof(struct io_overflow_cqe);
784 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
786 lockdep_assert_held(&ctx->completion_lock);
789 ocq_size += sizeof(struct io_uring_cqe);
791 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
792 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
795 * If we're in ring overflow flush mode, or in task cancel mode,
796 * or cannot allocate an overflow entry, then we need to drop it
799 io_account_cq_overflow(ctx);
800 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
803 if (list_empty(&ctx->cq_overflow_list)) {
804 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
805 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
808 ocqe->cqe.user_data = user_data;
810 ocqe->cqe.flags = cflags;
812 ocqe->cqe.big_cqe[0] = extra1;
813 ocqe->cqe.big_cqe[1] = extra2;
815 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
819 bool io_req_cqe_overflow(struct io_kiocb *req)
821 if (!(req->flags & REQ_F_CQE32_INIT)) {
825 return io_cqring_event_overflow(req->ctx, req->cqe.user_data,
826 req->cqe.res, req->cqe.flags,
827 req->extra1, req->extra2);
831 * writes to the cq entry need to come after reading head; the
832 * control dependency is enough as we're using WRITE_ONCE to
835 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow)
837 struct io_rings *rings = ctx->rings;
838 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
839 unsigned int free, queued, len;
842 * Posting into the CQ when there are pending overflowed CQEs may break
843 * ordering guarantees, which will affect links, F_MORE users and more.
844 * Force overflow the completion.
846 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
849 /* userspace may cheat modifying the tail, be safe and do min */
850 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
851 free = ctx->cq_entries - queued;
852 /* we need a contiguous range, limit based on the current array offset */
853 len = min(free, ctx->cq_entries - off);
857 if (ctx->flags & IORING_SETUP_CQE32) {
862 ctx->cqe_cached = &rings->cqes[off];
863 ctx->cqe_sentinel = ctx->cqe_cached + len;
865 ctx->cached_cq_tail++;
867 if (ctx->flags & IORING_SETUP_CQE32)
869 return &rings->cqes[off];
872 static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
875 struct io_uring_cqe *cqe;
880 * If we can't get a cq entry, userspace overflowed the
881 * submission (by quite a lot). Increment the overflow count in
884 cqe = io_get_cqe(ctx);
886 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
888 WRITE_ONCE(cqe->user_data, user_data);
889 WRITE_ONCE(cqe->res, res);
890 WRITE_ONCE(cqe->flags, cflags);
892 if (ctx->flags & IORING_SETUP_CQE32) {
893 WRITE_ONCE(cqe->big_cqe[0], 0);
894 WRITE_ONCE(cqe->big_cqe[1], 0);
901 static void __io_flush_post_cqes(struct io_ring_ctx *ctx)
902 __must_hold(&ctx->uring_lock)
904 struct io_submit_state *state = &ctx->submit_state;
907 lockdep_assert_held(&ctx->uring_lock);
908 for (i = 0; i < state->cqes_count; i++) {
909 struct io_uring_cqe *cqe = &state->cqes[i];
911 if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) {
912 if (ctx->task_complete) {
913 spin_lock(&ctx->completion_lock);
914 io_cqring_event_overflow(ctx, cqe->user_data,
915 cqe->res, cqe->flags, 0, 0);
916 spin_unlock(&ctx->completion_lock);
918 io_cqring_event_overflow(ctx, cqe->user_data,
919 cqe->res, cqe->flags, 0, 0);
923 state->cqes_count = 0;
926 static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
932 filled = io_fill_cqe_aux(ctx, user_data, res, cflags);
933 if (!filled && allow_overflow)
934 filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
936 io_cq_unlock_post(ctx);
940 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags)
942 return __io_post_aux_cqe(ctx, user_data, res, cflags, true);
945 bool io_aux_cqe(const struct io_kiocb *req, bool defer, s32 res, u32 cflags,
948 struct io_ring_ctx *ctx = req->ctx;
949 u64 user_data = req->cqe.user_data;
950 struct io_uring_cqe *cqe;
953 return __io_post_aux_cqe(ctx, user_data, res, cflags, allow_overflow);
955 lockdep_assert_held(&ctx->uring_lock);
957 if (ctx->submit_state.cqes_count == ARRAY_SIZE(ctx->submit_state.cqes)) {
959 __io_flush_post_cqes(ctx);
960 /* no need to flush - flush is deferred */
961 __io_cq_unlock_post(ctx);
964 /* For defered completions this is not as strict as it is otherwise,
965 * however it's main job is to prevent unbounded posted completions,
966 * and in that it works just as well.
968 if (!allow_overflow && test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
971 cqe = &ctx->submit_state.cqes[ctx->submit_state.cqes_count++];
972 cqe->user_data = user_data;
978 static void __io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
980 struct io_ring_ctx *ctx = req->ctx;
981 struct io_rsrc_node *rsrc_node = NULL;
984 if (!(req->flags & REQ_F_CQE_SKIP))
985 io_fill_cqe_req(ctx, req);
988 * If we're the last reference to this request, add to our locked
991 if (req_ref_put_and_test(req)) {
992 if (req->flags & IO_REQ_LINK_FLAGS) {
993 if (req->flags & IO_DISARM_MASK)
996 io_req_task_queue(req->link);
1000 io_put_kbuf_comp(req);
1001 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1003 if (!(req->flags & REQ_F_FIXED_FILE))
1004 io_put_file(req->file);
1006 rsrc_node = req->rsrc_node;
1008 * Selected buffer deallocation in io_clean_op() assumes that
1009 * we don't hold ->completion_lock. Clean them here to avoid
1012 io_put_task_remote(req->task);
1013 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
1014 ctx->locked_free_nr++;
1016 io_cq_unlock_post(ctx);
1019 io_ring_submit_lock(ctx, issue_flags);
1020 io_put_rsrc_node(ctx, rsrc_node);
1021 io_ring_submit_unlock(ctx, issue_flags);
1025 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
1027 if (req->ctx->task_complete && req->ctx->submitter_task != current) {
1028 req->io_task_work.func = io_req_task_complete;
1029 io_req_task_work_add(req);
1030 } else if (!(issue_flags & IO_URING_F_UNLOCKED) ||
1031 !(req->ctx->flags & IORING_SETUP_IOPOLL)) {
1032 __io_req_complete_post(req, issue_flags);
1034 struct io_ring_ctx *ctx = req->ctx;
1036 mutex_lock(&ctx->uring_lock);
1037 __io_req_complete_post(req, issue_flags & ~IO_URING_F_UNLOCKED);
1038 mutex_unlock(&ctx->uring_lock);
1042 void io_req_defer_failed(struct io_kiocb *req, s32 res)
1043 __must_hold(&ctx->uring_lock)
1045 const struct io_cold_def *def = &io_cold_defs[req->opcode];
1047 lockdep_assert_held(&req->ctx->uring_lock);
1050 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
1053 io_req_complete_defer(req);
1057 * Don't initialise the fields below on every allocation, but do that in
1058 * advance and keep them valid across allocations.
1060 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1064 req->async_data = NULL;
1065 /* not necessary, but safer to zero */
1069 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1070 struct io_submit_state *state)
1072 spin_lock(&ctx->completion_lock);
1073 wq_list_splice(&ctx->locked_free_list, &state->free_list);
1074 ctx->locked_free_nr = 0;
1075 spin_unlock(&ctx->completion_lock);
1079 * A request might get retired back into the request caches even before opcode
1080 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1081 * Because of that, io_alloc_req() should be called only under ->uring_lock
1082 * and with extra caution to not get a request that is still worked on.
1084 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
1085 __must_hold(&ctx->uring_lock)
1087 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1088 void *reqs[IO_REQ_ALLOC_BATCH];
1092 * If we have more than a batch's worth of requests in our IRQ side
1093 * locked cache, grab the lock and move them over to our submission
1096 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
1097 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
1098 if (!io_req_cache_empty(ctx))
1102 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
1105 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1106 * retry single alloc to be on the safe side.
1108 if (unlikely(ret <= 0)) {
1109 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1115 percpu_ref_get_many(&ctx->refs, ret);
1116 for (i = 0; i < ret; i++) {
1117 struct io_kiocb *req = reqs[i];
1119 io_preinit_req(req, ctx);
1120 io_req_add_to_cache(req, ctx);
1125 __cold void io_free_req(struct io_kiocb *req)
1127 /* refs were already put, restore them for io_req_task_complete() */
1128 req->flags &= ~REQ_F_REFCOUNT;
1129 /* we only want to free it, don't post CQEs */
1130 req->flags |= REQ_F_CQE_SKIP;
1131 req->io_task_work.func = io_req_task_complete;
1132 io_req_task_work_add(req);
1135 static void __io_req_find_next_prep(struct io_kiocb *req)
1137 struct io_ring_ctx *ctx = req->ctx;
1139 spin_lock(&ctx->completion_lock);
1140 io_disarm_next(req);
1141 spin_unlock(&ctx->completion_lock);
1144 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1146 struct io_kiocb *nxt;
1149 * If LINK is set, we have dependent requests in this chain. If we
1150 * didn't fail this request, queue the first one up, moving any other
1151 * dependencies to the next request. In case of failure, fail the rest
1154 if (unlikely(req->flags & IO_DISARM_MASK))
1155 __io_req_find_next_prep(req);
1161 static void ctx_flush_and_put(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1165 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1166 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1168 io_submit_flush_completions(ctx);
1169 mutex_unlock(&ctx->uring_lock);
1172 percpu_ref_put(&ctx->refs);
1175 static unsigned int handle_tw_list(struct llist_node *node,
1176 struct io_ring_ctx **ctx,
1177 struct io_tw_state *ts,
1178 struct llist_node *last)
1180 unsigned int count = 0;
1182 while (node && node != last) {
1183 struct llist_node *next = node->next;
1184 struct io_kiocb *req = container_of(node, struct io_kiocb,
1187 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1189 if (req->ctx != *ctx) {
1190 ctx_flush_and_put(*ctx, ts);
1192 /* if not contended, grab and improve batching */
1193 ts->locked = mutex_trylock(&(*ctx)->uring_lock);
1194 percpu_ref_get(&(*ctx)->refs);
1196 INDIRECT_CALL_2(req->io_task_work.func,
1197 io_poll_task_func, io_req_rw_complete,
1201 if (unlikely(need_resched())) {
1202 ctx_flush_and_put(*ctx, ts);
1212 * io_llist_xchg - swap all entries in a lock-less list
1213 * @head: the head of lock-less list to delete all entries
1214 * @new: new entry as the head of the list
1216 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1217 * The order of entries returned is from the newest to the oldest added one.
1219 static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1220 struct llist_node *new)
1222 return xchg(&head->first, new);
1226 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1227 * @head: the head of lock-less list to delete all entries
1228 * @old: expected old value of the first entry of the list
1229 * @new: new entry as the head of the list
1231 * perform a cmpxchg on the first entry of the list.
1234 static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1235 struct llist_node *old,
1236 struct llist_node *new)
1238 return cmpxchg(&head->first, old, new);
1241 void tctx_task_work(struct callback_head *cb)
1243 struct io_tw_state ts = {};
1244 struct io_ring_ctx *ctx = NULL;
1245 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1247 struct llist_node fake = {};
1248 struct llist_node *node;
1249 unsigned int loops = 0;
1250 unsigned int count = 0;
1252 if (unlikely(current->flags & PF_EXITING)) {
1253 io_fallback_tw(tctx);
1259 node = io_llist_xchg(&tctx->task_list, &fake);
1260 count += handle_tw_list(node, &ctx, &ts, &fake);
1262 /* skip expensive cmpxchg if there are items in the list */
1263 if (READ_ONCE(tctx->task_list.first) != &fake)
1265 if (ts.locked && !wq_list_empty(&ctx->submit_state.compl_reqs)) {
1266 io_submit_flush_completions(ctx);
1267 if (READ_ONCE(tctx->task_list.first) != &fake)
1270 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1271 } while (node != &fake);
1273 ctx_flush_and_put(ctx, &ts);
1275 /* relaxed read is enough as only the task itself sets ->in_cancel */
1276 if (unlikely(atomic_read(&tctx->in_cancel)))
1277 io_uring_drop_tctx_refs(current);
1279 trace_io_uring_task_work_run(tctx, count, loops);
1282 static __cold void io_fallback_tw(struct io_uring_task *tctx)
1284 struct llist_node *node = llist_del_all(&tctx->task_list);
1285 struct io_kiocb *req;
1288 req = container_of(node, struct io_kiocb, io_task_work.node);
1290 if (llist_add(&req->io_task_work.node,
1291 &req->ctx->fallback_llist))
1292 schedule_delayed_work(&req->ctx->fallback_work, 1);
1296 static inline void io_req_local_work_add(struct io_kiocb *req, unsigned flags)
1298 struct io_ring_ctx *ctx = req->ctx;
1299 unsigned nr_wait, nr_tw, nr_tw_prev;
1300 struct llist_node *first;
1302 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK))
1303 flags &= ~IOU_F_TWQ_LAZY_WAKE;
1305 first = READ_ONCE(ctx->work_llist.first);
1309 struct io_kiocb *first_req = container_of(first,
1313 * Might be executed at any moment, rely on
1314 * SLAB_TYPESAFE_BY_RCU to keep it alive.
1316 nr_tw_prev = READ_ONCE(first_req->nr_tw);
1318 nr_tw = nr_tw_prev + 1;
1319 /* Large enough to fail the nr_wait comparison below */
1320 if (!(flags & IOU_F_TWQ_LAZY_WAKE))
1324 req->io_task_work.node.next = first;
1325 } while (!try_cmpxchg(&ctx->work_llist.first, &first,
1326 &req->io_task_work.node));
1329 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1330 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1332 io_eventfd_signal(ctx);
1335 nr_wait = atomic_read(&ctx->cq_wait_nr);
1336 /* no one is waiting */
1339 /* either not enough or the previous add has already woken it up */
1340 if (nr_wait > nr_tw || nr_tw_prev >= nr_wait)
1342 /* pairs with set_current_state() in io_cqring_wait() */
1343 smp_mb__after_atomic();
1344 wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE);
1347 static void io_req_normal_work_add(struct io_kiocb *req)
1349 struct io_uring_task *tctx = req->task->io_uring;
1350 struct io_ring_ctx *ctx = req->ctx;
1352 /* task_work already pending, we're done */
1353 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1356 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1357 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1359 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1362 io_fallback_tw(tctx);
1365 void __io_req_task_work_add(struct io_kiocb *req, unsigned flags)
1367 if (req->ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
1369 io_req_local_work_add(req, flags);
1372 io_req_normal_work_add(req);
1376 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1378 struct llist_node *node;
1380 node = llist_del_all(&ctx->work_llist);
1382 struct io_kiocb *req = container_of(node, struct io_kiocb,
1386 io_req_normal_work_add(req);
1390 static int __io_run_local_work(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1392 struct llist_node *node;
1393 unsigned int loops = 0;
1396 if (WARN_ON_ONCE(ctx->submitter_task != current))
1398 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1399 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1402 * llists are in reverse order, flip it back the right way before
1403 * running the pending items.
1405 node = llist_reverse_order(io_llist_xchg(&ctx->work_llist, NULL));
1407 struct llist_node *next = node->next;
1408 struct io_kiocb *req = container_of(node, struct io_kiocb,
1410 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1411 INDIRECT_CALL_2(req->io_task_work.func,
1412 io_poll_task_func, io_req_rw_complete,
1419 if (!llist_empty(&ctx->work_llist))
1422 io_submit_flush_completions(ctx);
1423 if (!llist_empty(&ctx->work_llist))
1426 trace_io_uring_local_work_run(ctx, ret, loops);
1430 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx)
1432 struct io_tw_state ts = { .locked = true, };
1435 if (llist_empty(&ctx->work_llist))
1438 ret = __io_run_local_work(ctx, &ts);
1439 /* shouldn't happen! */
1440 if (WARN_ON_ONCE(!ts.locked))
1441 mutex_lock(&ctx->uring_lock);
1445 static int io_run_local_work(struct io_ring_ctx *ctx)
1447 struct io_tw_state ts = {};
1450 ts.locked = mutex_trylock(&ctx->uring_lock);
1451 ret = __io_run_local_work(ctx, &ts);
1453 mutex_unlock(&ctx->uring_lock);
1458 static void io_req_task_cancel(struct io_kiocb *req, struct io_tw_state *ts)
1460 io_tw_lock(req->ctx, ts);
1461 io_req_defer_failed(req, req->cqe.res);
1464 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts)
1466 io_tw_lock(req->ctx, ts);
1467 /* req->task == current here, checking PF_EXITING is safe */
1468 if (unlikely(req->task->flags & PF_EXITING))
1469 io_req_defer_failed(req, -EFAULT);
1470 else if (req->flags & REQ_F_FORCE_ASYNC)
1471 io_queue_iowq(req, ts);
1476 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1478 io_req_set_res(req, ret, 0);
1479 req->io_task_work.func = io_req_task_cancel;
1480 io_req_task_work_add(req);
1483 void io_req_task_queue(struct io_kiocb *req)
1485 req->io_task_work.func = io_req_task_submit;
1486 io_req_task_work_add(req);
1489 void io_queue_next(struct io_kiocb *req)
1491 struct io_kiocb *nxt = io_req_find_next(req);
1494 io_req_task_queue(nxt);
1497 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1498 __must_hold(&ctx->uring_lock)
1501 struct io_kiocb *req = container_of(node, struct io_kiocb,
1504 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1505 if (req->flags & REQ_F_REFCOUNT) {
1506 node = req->comp_list.next;
1507 if (!req_ref_put_and_test(req))
1510 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1511 struct async_poll *apoll = req->apoll;
1513 if (apoll->double_poll)
1514 kfree(apoll->double_poll);
1515 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1517 req->flags &= ~REQ_F_POLLED;
1519 if (req->flags & IO_REQ_LINK_FLAGS)
1521 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1524 if (!(req->flags & REQ_F_FIXED_FILE))
1525 io_put_file(req->file);
1527 io_req_put_rsrc_locked(req, ctx);
1529 io_put_task(req->task);
1530 node = req->comp_list.next;
1531 io_req_add_to_cache(req, ctx);
1535 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1536 __must_hold(&ctx->uring_lock)
1538 struct io_submit_state *state = &ctx->submit_state;
1539 struct io_wq_work_node *node;
1542 /* must come first to preserve CQE ordering in failure cases */
1543 if (state->cqes_count)
1544 __io_flush_post_cqes(ctx);
1545 __wq_list_for_each(node, &state->compl_reqs) {
1546 struct io_kiocb *req = container_of(node, struct io_kiocb,
1549 if (!(req->flags & REQ_F_CQE_SKIP) &&
1550 unlikely(!__io_fill_cqe_req(ctx, req))) {
1551 if (ctx->task_complete) {
1552 spin_lock(&ctx->completion_lock);
1553 io_req_cqe_overflow(req);
1554 spin_unlock(&ctx->completion_lock);
1556 io_req_cqe_overflow(req);
1560 __io_cq_unlock_post(ctx);
1562 if (!wq_list_empty(&ctx->submit_state.compl_reqs)) {
1563 io_free_batch_list(ctx, state->compl_reqs.first);
1564 INIT_WQ_LIST(&state->compl_reqs);
1568 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1570 /* See comment at the top of this file */
1572 return __io_cqring_events(ctx);
1576 * We can't just wait for polled events to come to us, we have to actively
1577 * find and complete them.
1579 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1581 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1584 mutex_lock(&ctx->uring_lock);
1585 while (!wq_list_empty(&ctx->iopoll_list)) {
1586 /* let it sleep and repeat later if can't complete a request */
1587 if (io_do_iopoll(ctx, true) == 0)
1590 * Ensure we allow local-to-the-cpu processing to take place,
1591 * in this case we need to ensure that we reap all events.
1592 * Also let task_work, etc. to progress by releasing the mutex
1594 if (need_resched()) {
1595 mutex_unlock(&ctx->uring_lock);
1597 mutex_lock(&ctx->uring_lock);
1600 mutex_unlock(&ctx->uring_lock);
1603 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1605 unsigned int nr_events = 0;
1607 unsigned long check_cq;
1609 if (!io_allowed_run_tw(ctx))
1612 check_cq = READ_ONCE(ctx->check_cq);
1613 if (unlikely(check_cq)) {
1614 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1615 __io_cqring_overflow_flush(ctx);
1617 * Similarly do not spin if we have not informed the user of any
1620 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1624 * Don't enter poll loop if we already have events pending.
1625 * If we do, we can potentially be spinning for commands that
1626 * already triggered a CQE (eg in error).
1628 if (io_cqring_events(ctx))
1633 * If a submit got punted to a workqueue, we can have the
1634 * application entering polling for a command before it gets
1635 * issued. That app will hold the uring_lock for the duration
1636 * of the poll right here, so we need to take a breather every
1637 * now and then to ensure that the issue has a chance to add
1638 * the poll to the issued list. Otherwise we can spin here
1639 * forever, while the workqueue is stuck trying to acquire the
1642 if (wq_list_empty(&ctx->iopoll_list) ||
1643 io_task_work_pending(ctx)) {
1644 u32 tail = ctx->cached_cq_tail;
1646 (void) io_run_local_work_locked(ctx);
1648 if (task_work_pending(current) ||
1649 wq_list_empty(&ctx->iopoll_list)) {
1650 mutex_unlock(&ctx->uring_lock);
1652 mutex_lock(&ctx->uring_lock);
1654 /* some requests don't go through iopoll_list */
1655 if (tail != ctx->cached_cq_tail ||
1656 wq_list_empty(&ctx->iopoll_list))
1659 ret = io_do_iopoll(ctx, !min);
1664 } while (nr_events < min && !need_resched());
1669 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts)
1672 io_req_complete_defer(req);
1674 io_req_complete_post(req, IO_URING_F_UNLOCKED);
1678 * After the iocb has been issued, it's safe to be found on the poll list.
1679 * Adding the kiocb to the list AFTER submission ensures that we don't
1680 * find it from a io_do_iopoll() thread before the issuer is done
1681 * accessing the kiocb cookie.
1683 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1685 struct io_ring_ctx *ctx = req->ctx;
1686 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1688 /* workqueue context doesn't hold uring_lock, grab it now */
1689 if (unlikely(needs_lock))
1690 mutex_lock(&ctx->uring_lock);
1693 * Track whether we have multiple files in our lists. This will impact
1694 * how we do polling eventually, not spinning if we're on potentially
1695 * different devices.
1697 if (wq_list_empty(&ctx->iopoll_list)) {
1698 ctx->poll_multi_queue = false;
1699 } else if (!ctx->poll_multi_queue) {
1700 struct io_kiocb *list_req;
1702 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1704 if (list_req->file != req->file)
1705 ctx->poll_multi_queue = true;
1709 * For fast devices, IO may have already completed. If it has, add
1710 * it to the front so we find it first.
1712 if (READ_ONCE(req->iopoll_completed))
1713 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1715 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1717 if (unlikely(needs_lock)) {
1719 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1720 * in sq thread task context or in io worker task context. If
1721 * current task context is sq thread, we don't need to check
1722 * whether should wake up sq thread.
1724 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1725 wq_has_sleeper(&ctx->sq_data->wait))
1726 wake_up(&ctx->sq_data->wait);
1728 mutex_unlock(&ctx->uring_lock);
1732 unsigned int io_file_get_flags(struct file *file)
1734 unsigned int res = 0;
1736 if (S_ISREG(file_inode(file)->i_mode))
1738 if ((file->f_flags & O_NONBLOCK) || (file->f_mode & FMODE_NOWAIT))
1739 res |= REQ_F_SUPPORT_NOWAIT;
1743 bool io_alloc_async_data(struct io_kiocb *req)
1745 WARN_ON_ONCE(!io_cold_defs[req->opcode].async_size);
1746 req->async_data = kmalloc(io_cold_defs[req->opcode].async_size, GFP_KERNEL);
1747 if (req->async_data) {
1748 req->flags |= REQ_F_ASYNC_DATA;
1754 int io_req_prep_async(struct io_kiocb *req)
1756 const struct io_cold_def *cdef = &io_cold_defs[req->opcode];
1757 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1759 /* assign early for deferred execution for non-fixed file */
1760 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE) && !req->file)
1761 req->file = io_file_get_normal(req, req->cqe.fd);
1762 if (!cdef->prep_async)
1764 if (WARN_ON_ONCE(req_has_async_data(req)))
1766 if (!def->manual_alloc) {
1767 if (io_alloc_async_data(req))
1770 return cdef->prep_async(req);
1773 static u32 io_get_sequence(struct io_kiocb *req)
1775 u32 seq = req->ctx->cached_sq_head;
1776 struct io_kiocb *cur;
1778 /* need original cached_sq_head, but it was increased for each req */
1779 io_for_each_link(cur, req)
1784 static __cold void io_drain_req(struct io_kiocb *req)
1785 __must_hold(&ctx->uring_lock)
1787 struct io_ring_ctx *ctx = req->ctx;
1788 struct io_defer_entry *de;
1790 u32 seq = io_get_sequence(req);
1792 /* Still need defer if there is pending req in defer list. */
1793 spin_lock(&ctx->completion_lock);
1794 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1795 spin_unlock(&ctx->completion_lock);
1797 ctx->drain_active = false;
1798 io_req_task_queue(req);
1801 spin_unlock(&ctx->completion_lock);
1803 io_prep_async_link(req);
1804 de = kmalloc(sizeof(*de), GFP_KERNEL);
1807 io_req_defer_failed(req, ret);
1811 spin_lock(&ctx->completion_lock);
1812 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1813 spin_unlock(&ctx->completion_lock);
1818 trace_io_uring_defer(req);
1821 list_add_tail(&de->list, &ctx->defer_list);
1822 spin_unlock(&ctx->completion_lock);
1825 static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def,
1826 unsigned int issue_flags)
1828 if (req->file || !def->needs_file)
1831 if (req->flags & REQ_F_FIXED_FILE)
1832 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1834 req->file = io_file_get_normal(req, req->cqe.fd);
1839 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1841 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1842 const struct cred *creds = NULL;
1845 if (unlikely(!io_assign_file(req, def, issue_flags)))
1848 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1849 creds = override_creds(req->creds);
1851 if (!def->audit_skip)
1852 audit_uring_entry(req->opcode);
1854 ret = def->issue(req, issue_flags);
1856 if (!def->audit_skip)
1857 audit_uring_exit(!ret, ret);
1860 revert_creds(creds);
1862 if (ret == IOU_OK) {
1863 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1864 io_req_complete_defer(req);
1866 io_req_complete_post(req, issue_flags);
1867 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1870 /* If the op doesn't have a file, we're not polling for it */
1871 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
1872 io_iopoll_req_issued(req, issue_flags);
1877 int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts)
1879 io_tw_lock(req->ctx, ts);
1880 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT|
1881 IO_URING_F_COMPLETE_DEFER);
1884 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1886 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1887 struct io_kiocb *nxt = NULL;
1889 if (req_ref_put_and_test(req)) {
1890 if (req->flags & IO_REQ_LINK_FLAGS)
1891 nxt = io_req_find_next(req);
1894 return nxt ? &nxt->work : NULL;
1897 void io_wq_submit_work(struct io_wq_work *work)
1899 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1900 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1901 unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ;
1902 bool needs_poll = false;
1903 int ret = 0, err = -ECANCELED;
1905 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1906 if (!(req->flags & REQ_F_REFCOUNT))
1907 __io_req_set_refcount(req, 2);
1911 io_arm_ltimeout(req);
1913 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1914 if (work->flags & IO_WQ_WORK_CANCEL) {
1916 io_req_task_queue_fail(req, err);
1919 if (!io_assign_file(req, def, issue_flags)) {
1921 work->flags |= IO_WQ_WORK_CANCEL;
1925 if (req->flags & REQ_F_FORCE_ASYNC) {
1926 bool opcode_poll = def->pollin || def->pollout;
1928 if (opcode_poll && file_can_poll(req->file)) {
1930 issue_flags |= IO_URING_F_NONBLOCK;
1935 ret = io_issue_sqe(req, issue_flags);
1939 * We can get EAGAIN for iopolled IO even though we're
1940 * forcing a sync submission from here, since we can't
1941 * wait for request slots on the block side.
1944 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1950 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1952 /* aborted or ready, in either case retry blocking */
1954 issue_flags &= ~IO_URING_F_NONBLOCK;
1957 /* avoid locking problems by failing it from a clean context */
1959 io_req_task_queue_fail(req, ret);
1962 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1963 unsigned int issue_flags)
1965 struct io_ring_ctx *ctx = req->ctx;
1966 struct io_fixed_file *slot;
1967 struct file *file = NULL;
1969 io_ring_submit_lock(ctx, issue_flags);
1971 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
1973 fd = array_index_nospec(fd, ctx->nr_user_files);
1974 slot = io_fixed_file_slot(&ctx->file_table, fd);
1975 file = io_slot_file(slot);
1976 req->flags |= io_slot_flags(slot);
1977 io_req_set_rsrc_node(req, ctx, 0);
1979 io_ring_submit_unlock(ctx, issue_flags);
1983 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
1985 struct file *file = fget(fd);
1987 trace_io_uring_file_get(req, fd);
1989 /* we don't allow fixed io_uring files */
1990 if (file && io_is_uring_fops(file))
1991 io_req_track_inflight(req);
1995 static void io_queue_async(struct io_kiocb *req, int ret)
1996 __must_hold(&req->ctx->uring_lock)
1998 struct io_kiocb *linked_timeout;
2000 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
2001 io_req_defer_failed(req, ret);
2005 linked_timeout = io_prep_linked_timeout(req);
2007 switch (io_arm_poll_handler(req, 0)) {
2008 case IO_APOLL_READY:
2009 io_kbuf_recycle(req, 0);
2010 io_req_task_queue(req);
2012 case IO_APOLL_ABORTED:
2013 io_kbuf_recycle(req, 0);
2014 io_queue_iowq(req, NULL);
2021 io_queue_linked_timeout(linked_timeout);
2024 static inline void io_queue_sqe(struct io_kiocb *req)
2025 __must_hold(&req->ctx->uring_lock)
2029 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
2032 * We async punt it if the file wasn't marked NOWAIT, or if the file
2033 * doesn't support non-blocking read/write attempts
2036 io_arm_ltimeout(req);
2038 io_queue_async(req, ret);
2041 static void io_queue_sqe_fallback(struct io_kiocb *req)
2042 __must_hold(&req->ctx->uring_lock)
2044 if (unlikely(req->flags & REQ_F_FAIL)) {
2046 * We don't submit, fail them all, for that replace hardlinks
2047 * with normal links. Extra REQ_F_LINK is tolerated.
2049 req->flags &= ~REQ_F_HARDLINK;
2050 req->flags |= REQ_F_LINK;
2051 io_req_defer_failed(req, req->cqe.res);
2053 int ret = io_req_prep_async(req);
2055 if (unlikely(ret)) {
2056 io_req_defer_failed(req, ret);
2060 if (unlikely(req->ctx->drain_active))
2063 io_queue_iowq(req, NULL);
2068 * Check SQE restrictions (opcode and flags).
2070 * Returns 'true' if SQE is allowed, 'false' otherwise.
2072 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
2073 struct io_kiocb *req,
2074 unsigned int sqe_flags)
2076 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
2079 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
2080 ctx->restrictions.sqe_flags_required)
2083 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
2084 ctx->restrictions.sqe_flags_required))
2090 static void io_init_req_drain(struct io_kiocb *req)
2092 struct io_ring_ctx *ctx = req->ctx;
2093 struct io_kiocb *head = ctx->submit_state.link.head;
2095 ctx->drain_active = true;
2098 * If we need to drain a request in the middle of a link, drain
2099 * the head request and the next request/link after the current
2100 * link. Considering sequential execution of links,
2101 * REQ_F_IO_DRAIN will be maintained for every request of our
2104 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2105 ctx->drain_next = true;
2109 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
2110 const struct io_uring_sqe *sqe)
2111 __must_hold(&ctx->uring_lock)
2113 const struct io_issue_def *def;
2114 unsigned int sqe_flags;
2118 /* req is partially pre-initialised, see io_preinit_req() */
2119 req->opcode = opcode = READ_ONCE(sqe->opcode);
2120 /* same numerical values with corresponding REQ_F_*, safe to copy */
2121 req->flags = sqe_flags = READ_ONCE(sqe->flags);
2122 req->cqe.user_data = READ_ONCE(sqe->user_data);
2124 req->rsrc_node = NULL;
2125 req->task = current;
2127 if (unlikely(opcode >= IORING_OP_LAST)) {
2131 def = &io_issue_defs[opcode];
2132 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2133 /* enforce forwards compatibility on users */
2134 if (sqe_flags & ~SQE_VALID_FLAGS)
2136 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2137 if (!def->buffer_select)
2139 req->buf_index = READ_ONCE(sqe->buf_group);
2141 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2142 ctx->drain_disabled = true;
2143 if (sqe_flags & IOSQE_IO_DRAIN) {
2144 if (ctx->drain_disabled)
2146 io_init_req_drain(req);
2149 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2150 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2152 /* knock it to the slow queue path, will be drained there */
2153 if (ctx->drain_active)
2154 req->flags |= REQ_F_FORCE_ASYNC;
2155 /* if there is no link, we're at "next" request and need to drain */
2156 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2157 ctx->drain_next = false;
2158 ctx->drain_active = true;
2159 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2163 if (!def->ioprio && sqe->ioprio)
2165 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2168 if (def->needs_file) {
2169 struct io_submit_state *state = &ctx->submit_state;
2171 req->cqe.fd = READ_ONCE(sqe->fd);
2174 * Plug now if we have more than 2 IO left after this, and the
2175 * target is potentially a read/write to block based storage.
2177 if (state->need_plug && def->plug) {
2178 state->plug_started = true;
2179 state->need_plug = false;
2180 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2184 personality = READ_ONCE(sqe->personality);
2188 req->creds = xa_load(&ctx->personalities, personality);
2191 get_cred(req->creds);
2192 ret = security_uring_override_creds(req->creds);
2194 put_cred(req->creds);
2197 req->flags |= REQ_F_CREDS;
2200 return def->prep(req, sqe);
2203 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2204 struct io_kiocb *req, int ret)
2206 struct io_ring_ctx *ctx = req->ctx;
2207 struct io_submit_link *link = &ctx->submit_state.link;
2208 struct io_kiocb *head = link->head;
2210 trace_io_uring_req_failed(sqe, req, ret);
2213 * Avoid breaking links in the middle as it renders links with SQPOLL
2214 * unusable. Instead of failing eagerly, continue assembling the link if
2215 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2216 * should find the flag and handle the rest.
2218 req_fail_link_node(req, ret);
2219 if (head && !(head->flags & REQ_F_FAIL))
2220 req_fail_link_node(head, -ECANCELED);
2222 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2224 link->last->link = req;
2228 io_queue_sqe_fallback(req);
2233 link->last->link = req;
2240 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2241 const struct io_uring_sqe *sqe)
2242 __must_hold(&ctx->uring_lock)
2244 struct io_submit_link *link = &ctx->submit_state.link;
2247 ret = io_init_req(ctx, req, sqe);
2249 return io_submit_fail_init(sqe, req, ret);
2251 trace_io_uring_submit_req(req);
2254 * If we already have a head request, queue this one for async
2255 * submittal once the head completes. If we don't have a head but
2256 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2257 * submitted sync once the chain is complete. If none of those
2258 * conditions are true (normal request), then just queue it.
2260 if (unlikely(link->head)) {
2261 ret = io_req_prep_async(req);
2263 return io_submit_fail_init(sqe, req, ret);
2265 trace_io_uring_link(req, link->head);
2266 link->last->link = req;
2269 if (req->flags & IO_REQ_LINK_FLAGS)
2271 /* last request of the link, flush it */
2274 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2277 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2278 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2279 if (req->flags & IO_REQ_LINK_FLAGS) {
2284 io_queue_sqe_fallback(req);
2294 * Batched submission is done, ensure local IO is flushed out.
2296 static void io_submit_state_end(struct io_ring_ctx *ctx)
2298 struct io_submit_state *state = &ctx->submit_state;
2300 if (unlikely(state->link.head))
2301 io_queue_sqe_fallback(state->link.head);
2302 /* flush only after queuing links as they can generate completions */
2303 io_submit_flush_completions(ctx);
2304 if (state->plug_started)
2305 blk_finish_plug(&state->plug);
2309 * Start submission side cache.
2311 static void io_submit_state_start(struct io_submit_state *state,
2312 unsigned int max_ios)
2314 state->plug_started = false;
2315 state->need_plug = max_ios > 2;
2316 state->submit_nr = max_ios;
2317 /* set only head, no need to init link_last in advance */
2318 state->link.head = NULL;
2321 static void io_commit_sqring(struct io_ring_ctx *ctx)
2323 struct io_rings *rings = ctx->rings;
2326 * Ensure any loads from the SQEs are done at this point,
2327 * since once we write the new head, the application could
2328 * write new data to them.
2330 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2334 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2335 * that is mapped by userspace. This means that care needs to be taken to
2336 * ensure that reads are stable, as we cannot rely on userspace always
2337 * being a good citizen. If members of the sqe are validated and then later
2338 * used, it's important that those reads are done through READ_ONCE() to
2339 * prevent a re-load down the line.
2341 static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe)
2343 unsigned head, mask = ctx->sq_entries - 1;
2344 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2347 * The cached sq head (or cq tail) serves two purposes:
2349 * 1) allows us to batch the cost of updating the user visible
2351 * 2) allows the kernel side to track the head on its own, even
2352 * though the application is the one updating it.
2354 head = READ_ONCE(ctx->sq_array[sq_idx]);
2355 if (likely(head < ctx->sq_entries)) {
2356 /* double index for 128-byte SQEs, twice as long */
2357 if (ctx->flags & IORING_SETUP_SQE128)
2359 *sqe = &ctx->sq_sqes[head];
2363 /* drop invalid entries */
2365 WRITE_ONCE(ctx->rings->sq_dropped,
2366 READ_ONCE(ctx->rings->sq_dropped) + 1);
2370 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2371 __must_hold(&ctx->uring_lock)
2373 unsigned int entries = io_sqring_entries(ctx);
2377 if (unlikely(!entries))
2379 /* make sure SQ entry isn't read before tail */
2380 ret = left = min(nr, entries);
2381 io_get_task_refs(left);
2382 io_submit_state_start(&ctx->submit_state, left);
2385 const struct io_uring_sqe *sqe;
2386 struct io_kiocb *req;
2388 if (unlikely(!io_alloc_req(ctx, &req)))
2390 if (unlikely(!io_get_sqe(ctx, &sqe))) {
2391 io_req_add_to_cache(req, ctx);
2396 * Continue submitting even for sqe failure if the
2397 * ring was setup with IORING_SETUP_SUBMIT_ALL
2399 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2400 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2406 if (unlikely(left)) {
2408 /* try again if it submitted nothing and can't allocate a req */
2409 if (!ret && io_req_cache_empty(ctx))
2411 current->io_uring->cached_refs += left;
2414 io_submit_state_end(ctx);
2415 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2416 io_commit_sqring(ctx);
2420 struct io_wait_queue {
2421 struct wait_queue_entry wq;
2422 struct io_ring_ctx *ctx;
2424 unsigned nr_timeouts;
2428 static inline bool io_has_work(struct io_ring_ctx *ctx)
2430 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
2431 !llist_empty(&ctx->work_llist);
2434 static inline bool io_should_wake(struct io_wait_queue *iowq)
2436 struct io_ring_ctx *ctx = iowq->ctx;
2437 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
2440 * Wake up if we have enough events, or if a timeout occurred since we
2441 * started waiting. For timeouts, we always want to return to userspace,
2442 * regardless of event count.
2444 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2447 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2448 int wake_flags, void *key)
2450 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq);
2453 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2454 * the task, and the next invocation will do it.
2456 if (io_should_wake(iowq) || io_has_work(iowq->ctx))
2457 return autoremove_wake_function(curr, mode, wake_flags, key);
2461 int io_run_task_work_sig(struct io_ring_ctx *ctx)
2463 if (!llist_empty(&ctx->work_llist)) {
2464 __set_current_state(TASK_RUNNING);
2465 if (io_run_local_work(ctx) > 0)
2468 if (io_run_task_work() > 0)
2470 if (task_sigpending(current))
2475 /* when returns >0, the caller should retry */
2476 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2477 struct io_wait_queue *iowq)
2479 if (unlikely(READ_ONCE(ctx->check_cq)))
2481 if (unlikely(!llist_empty(&ctx->work_llist)))
2483 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL)))
2485 if (unlikely(task_sigpending(current)))
2487 if (unlikely(io_should_wake(iowq)))
2489 if (iowq->timeout == KTIME_MAX)
2491 else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS))
2497 * Wait until events become available, if we don't already have some. The
2498 * application must reap them itself, as they reside on the shared cq ring.
2500 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2501 const sigset_t __user *sig, size_t sigsz,
2502 struct __kernel_timespec __user *uts)
2504 struct io_wait_queue iowq;
2505 struct io_rings *rings = ctx->rings;
2508 if (!io_allowed_run_tw(ctx))
2510 if (!llist_empty(&ctx->work_llist))
2511 io_run_local_work(ctx);
2513 io_cqring_overflow_flush(ctx);
2514 /* if user messes with these they will just get an early return */
2515 if (__io_cqring_events_user(ctx) >= min_events)
2519 #ifdef CONFIG_COMPAT
2520 if (in_compat_syscall())
2521 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2525 ret = set_user_sigmask(sig, sigsz);
2531 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2532 iowq.wq.private = current;
2533 INIT_LIST_HEAD(&iowq.wq.entry);
2535 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2536 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2537 iowq.timeout = KTIME_MAX;
2540 struct timespec64 ts;
2542 if (get_timespec64(&ts, uts))
2544 iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2547 trace_io_uring_cqring_wait(ctx, min_events);
2549 unsigned long check_cq;
2551 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
2552 int nr_wait = (int) iowq.cq_tail - READ_ONCE(ctx->rings->cq.tail);
2554 atomic_set(&ctx->cq_wait_nr, nr_wait);
2555 set_current_state(TASK_INTERRUPTIBLE);
2557 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2558 TASK_INTERRUPTIBLE);
2561 ret = io_cqring_wait_schedule(ctx, &iowq);
2562 __set_current_state(TASK_RUNNING);
2563 atomic_set(&ctx->cq_wait_nr, 0);
2568 * Run task_work after scheduling and before io_should_wake().
2569 * If we got woken because of task_work being processed, run it
2570 * now rather than let the caller do another wait loop.
2573 if (!llist_empty(&ctx->work_llist))
2574 io_run_local_work(ctx);
2576 check_cq = READ_ONCE(ctx->check_cq);
2577 if (unlikely(check_cq)) {
2578 /* let the caller flush overflows, retry */
2579 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2580 io_cqring_do_overflow_flush(ctx);
2581 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) {
2587 if (io_should_wake(&iowq)) {
2594 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
2595 finish_wait(&ctx->cq_wait, &iowq.wq);
2596 restore_saved_sigmask_unless(ret == -EINTR);
2598 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2601 static void io_mem_free(void *ptr)
2608 page = virt_to_head_page(ptr);
2609 if (put_page_testzero(page))
2610 free_compound_page(page);
2613 static void io_pages_free(struct page ***pages, int npages)
2615 struct page **page_array;
2620 page_array = *pages;
2621 for (i = 0; i < npages; i++)
2622 unpin_user_page(page_array[i]);
2627 static void *__io_uaddr_map(struct page ***pages, unsigned short *npages,
2628 unsigned long uaddr, size_t size)
2630 struct page **page_array;
2631 unsigned int nr_pages;
2636 if (uaddr & (PAGE_SIZE - 1) || !size)
2637 return ERR_PTR(-EINVAL);
2639 nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
2640 if (nr_pages > USHRT_MAX)
2641 return ERR_PTR(-EINVAL);
2642 page_array = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
2644 return ERR_PTR(-ENOMEM);
2646 ret = pin_user_pages_fast(uaddr, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
2648 if (ret != nr_pages) {
2650 io_pages_free(&page_array, ret > 0 ? ret : 0);
2651 return ret < 0 ? ERR_PTR(ret) : ERR_PTR(-EFAULT);
2654 * Should be a single page. If the ring is small enough that we can
2655 * use a normal page, that is fine. If we need multiple pages, then
2656 * userspace should use a huge page. That's the only way to guarantee
2657 * that we get contigious memory, outside of just being lucky or
2658 * (currently) having low memory fragmentation.
2660 if (page_array[0] != page_array[ret - 1])
2662 *pages = page_array;
2664 return page_to_virt(page_array[0]);
2667 static void *io_rings_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2670 return __io_uaddr_map(&ctx->ring_pages, &ctx->n_ring_pages, uaddr,
2674 static void *io_sqes_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2677 return __io_uaddr_map(&ctx->sqe_pages, &ctx->n_sqe_pages, uaddr,
2681 static void io_rings_free(struct io_ring_ctx *ctx)
2683 if (!(ctx->flags & IORING_SETUP_NO_MMAP)) {
2684 io_mem_free(ctx->rings);
2685 io_mem_free(ctx->sq_sqes);
2687 ctx->sq_sqes = NULL;
2689 io_pages_free(&ctx->ring_pages, ctx->n_ring_pages);
2690 io_pages_free(&ctx->sqe_pages, ctx->n_sqe_pages);
2694 static void *io_mem_alloc(size_t size)
2696 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2699 ret = (void *) __get_free_pages(gfp, get_order(size));
2702 return ERR_PTR(-ENOMEM);
2705 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2706 unsigned int cq_entries, size_t *sq_offset)
2708 struct io_rings *rings;
2709 size_t off, sq_array_size;
2711 off = struct_size(rings, cqes, cq_entries);
2712 if (off == SIZE_MAX)
2714 if (ctx->flags & IORING_SETUP_CQE32) {
2715 if (check_shl_overflow(off, 1, &off))
2720 off = ALIGN(off, SMP_CACHE_BYTES);
2728 sq_array_size = array_size(sizeof(u32), sq_entries);
2729 if (sq_array_size == SIZE_MAX)
2732 if (check_add_overflow(off, sq_array_size, &off))
2738 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2739 unsigned int eventfd_async)
2741 struct io_ev_fd *ev_fd;
2742 __s32 __user *fds = arg;
2745 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2746 lockdep_is_held(&ctx->uring_lock));
2750 if (copy_from_user(&fd, fds, sizeof(*fds)))
2753 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2757 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2758 if (IS_ERR(ev_fd->cq_ev_fd)) {
2759 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2764 spin_lock(&ctx->completion_lock);
2765 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2766 spin_unlock(&ctx->completion_lock);
2768 ev_fd->eventfd_async = eventfd_async;
2769 ctx->has_evfd = true;
2770 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2771 atomic_set(&ev_fd->refs, 1);
2772 atomic_set(&ev_fd->ops, 0);
2776 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2778 struct io_ev_fd *ev_fd;
2780 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2781 lockdep_is_held(&ctx->uring_lock));
2783 ctx->has_evfd = false;
2784 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2785 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops))
2786 call_rcu(&ev_fd->rcu, io_eventfd_ops);
2793 static void io_req_caches_free(struct io_ring_ctx *ctx)
2795 struct io_kiocb *req;
2798 mutex_lock(&ctx->uring_lock);
2799 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
2801 while (!io_req_cache_empty(ctx)) {
2802 req = io_extract_req(ctx);
2803 kmem_cache_free(req_cachep, req);
2807 percpu_ref_put_many(&ctx->refs, nr);
2808 mutex_unlock(&ctx->uring_lock);
2811 static void io_rsrc_node_cache_free(struct io_cache_entry *entry)
2813 kfree(container_of(entry, struct io_rsrc_node, cache));
2816 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2818 io_sq_thread_finish(ctx);
2819 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2820 if (WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)))
2823 mutex_lock(&ctx->uring_lock);
2825 __io_sqe_buffers_unregister(ctx);
2827 __io_sqe_files_unregister(ctx);
2828 io_cqring_overflow_kill(ctx);
2829 io_eventfd_unregister(ctx);
2830 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2831 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2832 io_destroy_buffers(ctx);
2833 mutex_unlock(&ctx->uring_lock);
2835 put_cred(ctx->sq_creds);
2836 if (ctx->submitter_task)
2837 put_task_struct(ctx->submitter_task);
2839 /* there are no registered resources left, nobody uses it */
2841 io_rsrc_node_destroy(ctx, ctx->rsrc_node);
2843 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2845 #if defined(CONFIG_UNIX)
2846 if (ctx->ring_sock) {
2847 ctx->ring_sock->file = NULL; /* so that iput() is called */
2848 sock_release(ctx->ring_sock);
2851 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2853 io_alloc_cache_free(&ctx->rsrc_node_cache, io_rsrc_node_cache_free);
2854 if (ctx->mm_account) {
2855 mmdrop(ctx->mm_account);
2856 ctx->mm_account = NULL;
2860 percpu_ref_exit(&ctx->refs);
2861 free_uid(ctx->user);
2862 io_req_caches_free(ctx);
2864 io_wq_put_hash(ctx->hash_map);
2865 kfree(ctx->cancel_table.hbs);
2866 kfree(ctx->cancel_table_locked.hbs);
2867 kfree(ctx->dummy_ubuf);
2869 xa_destroy(&ctx->io_bl_xa);
2873 static __cold void io_activate_pollwq_cb(struct callback_head *cb)
2875 struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx,
2878 mutex_lock(&ctx->uring_lock);
2879 ctx->poll_activated = true;
2880 mutex_unlock(&ctx->uring_lock);
2883 * Wake ups for some events between start of polling and activation
2884 * might've been lost due to loose synchronisation.
2886 wake_up_all(&ctx->poll_wq);
2887 percpu_ref_put(&ctx->refs);
2890 static __cold void io_activate_pollwq(struct io_ring_ctx *ctx)
2892 spin_lock(&ctx->completion_lock);
2893 /* already activated or in progress */
2894 if (ctx->poll_activated || ctx->poll_wq_task_work.func)
2896 if (WARN_ON_ONCE(!ctx->task_complete))
2898 if (!ctx->submitter_task)
2901 * with ->submitter_task only the submitter task completes requests, we
2902 * only need to sync with it, which is done by injecting a tw
2904 init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb);
2905 percpu_ref_get(&ctx->refs);
2906 if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL))
2907 percpu_ref_put(&ctx->refs);
2909 spin_unlock(&ctx->completion_lock);
2912 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2914 struct io_ring_ctx *ctx = file->private_data;
2917 if (unlikely(!ctx->poll_activated))
2918 io_activate_pollwq(ctx);
2920 poll_wait(file, &ctx->poll_wq, wait);
2922 * synchronizes with barrier from wq_has_sleeper call in
2926 if (!io_sqring_full(ctx))
2927 mask |= EPOLLOUT | EPOLLWRNORM;
2930 * Don't flush cqring overflow list here, just do a simple check.
2931 * Otherwise there could possible be ABBA deadlock:
2934 * lock(&ctx->uring_lock);
2936 * lock(&ctx->uring_lock);
2939 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2940 * pushes them to do the flush.
2943 if (__io_cqring_events_user(ctx) || io_has_work(ctx))
2944 mask |= EPOLLIN | EPOLLRDNORM;
2949 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2951 const struct cred *creds;
2953 creds = xa_erase(&ctx->personalities, id);
2962 struct io_tctx_exit {
2963 struct callback_head task_work;
2964 struct completion completion;
2965 struct io_ring_ctx *ctx;
2968 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2970 struct io_uring_task *tctx = current->io_uring;
2971 struct io_tctx_exit *work;
2973 work = container_of(cb, struct io_tctx_exit, task_work);
2975 * When @in_cancel, we're in cancellation and it's racy to remove the
2976 * node. It'll be removed by the end of cancellation, just ignore it.
2977 * tctx can be NULL if the queueing of this task_work raced with
2978 * work cancelation off the exec path.
2980 if (tctx && !atomic_read(&tctx->in_cancel))
2981 io_uring_del_tctx_node((unsigned long)work->ctx);
2982 complete(&work->completion);
2985 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2987 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2989 return req->ctx == data;
2992 static __cold void io_ring_exit_work(struct work_struct *work)
2994 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2995 unsigned long timeout = jiffies + HZ * 60 * 5;
2996 unsigned long interval = HZ / 20;
2997 struct io_tctx_exit exit;
2998 struct io_tctx_node *node;
3002 * If we're doing polled IO and end up having requests being
3003 * submitted async (out-of-line), then completions can come in while
3004 * we're waiting for refs to drop. We need to reap these manually,
3005 * as nobody else will be looking for them.
3008 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
3009 mutex_lock(&ctx->uring_lock);
3010 io_cqring_overflow_kill(ctx);
3011 mutex_unlock(&ctx->uring_lock);
3014 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3015 io_move_task_work_from_local(ctx);
3017 while (io_uring_try_cancel_requests(ctx, NULL, true))
3021 struct io_sq_data *sqd = ctx->sq_data;
3022 struct task_struct *tsk;
3024 io_sq_thread_park(sqd);
3026 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
3027 io_wq_cancel_cb(tsk->io_uring->io_wq,
3028 io_cancel_ctx_cb, ctx, true);
3029 io_sq_thread_unpark(sqd);
3032 io_req_caches_free(ctx);
3034 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
3035 /* there is little hope left, don't run it too often */
3039 * This is really an uninterruptible wait, as it has to be
3040 * complete. But it's also run from a kworker, which doesn't
3041 * take signals, so it's fine to make it interruptible. This
3042 * avoids scenarios where we knowingly can wait much longer
3043 * on completions, for example if someone does a SIGSTOP on
3044 * a task that needs to finish task_work to make this loop
3045 * complete. That's a synthetic situation that should not
3046 * cause a stuck task backtrace, and hence a potential panic
3047 * on stuck tasks if that is enabled.
3049 } while (!wait_for_completion_interruptible_timeout(&ctx->ref_comp, interval));
3051 init_completion(&exit.completion);
3052 init_task_work(&exit.task_work, io_tctx_exit_cb);
3055 * Some may use context even when all refs and requests have been put,
3056 * and they are free to do so while still holding uring_lock or
3057 * completion_lock, see io_req_task_submit(). Apart from other work,
3058 * this lock/unlock section also waits them to finish.
3060 mutex_lock(&ctx->uring_lock);
3061 while (!list_empty(&ctx->tctx_list)) {
3062 WARN_ON_ONCE(time_after(jiffies, timeout));
3064 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
3066 /* don't spin on a single task if cancellation failed */
3067 list_rotate_left(&ctx->tctx_list);
3068 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
3069 if (WARN_ON_ONCE(ret))
3072 mutex_unlock(&ctx->uring_lock);
3074 * See comment above for
3075 * wait_for_completion_interruptible_timeout() on why this
3076 * wait is marked as interruptible.
3078 wait_for_completion_interruptible(&exit.completion);
3079 mutex_lock(&ctx->uring_lock);
3081 mutex_unlock(&ctx->uring_lock);
3082 spin_lock(&ctx->completion_lock);
3083 spin_unlock(&ctx->completion_lock);
3085 /* pairs with RCU read section in io_req_local_work_add() */
3086 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3089 io_ring_ctx_free(ctx);
3092 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
3094 unsigned long index;
3095 struct creds *creds;
3097 mutex_lock(&ctx->uring_lock);
3098 percpu_ref_kill(&ctx->refs);
3099 xa_for_each(&ctx->personalities, index, creds)
3100 io_unregister_personality(ctx, index);
3102 io_poll_remove_all(ctx, NULL, true);
3103 mutex_unlock(&ctx->uring_lock);
3106 * If we failed setting up the ctx, we might not have any rings
3107 * and therefore did not submit any requests
3110 io_kill_timeouts(ctx, NULL, true);
3112 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
3114 * Use system_unbound_wq to avoid spawning tons of event kworkers
3115 * if we're exiting a ton of rings at the same time. It just adds
3116 * noise and overhead, there's no discernable change in runtime
3117 * over using system_wq.
3119 queue_work(system_unbound_wq, &ctx->exit_work);
3122 static int io_uring_release(struct inode *inode, struct file *file)
3124 struct io_ring_ctx *ctx = file->private_data;
3126 file->private_data = NULL;
3127 io_ring_ctx_wait_and_kill(ctx);
3131 struct io_task_cancel {
3132 struct task_struct *task;
3136 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
3138 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3139 struct io_task_cancel *cancel = data;
3141 return io_match_task_safe(req, cancel->task, cancel->all);
3144 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
3145 struct task_struct *task,
3148 struct io_defer_entry *de;
3151 spin_lock(&ctx->completion_lock);
3152 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
3153 if (io_match_task_safe(de->req, task, cancel_all)) {
3154 list_cut_position(&list, &ctx->defer_list, &de->list);
3158 spin_unlock(&ctx->completion_lock);
3159 if (list_empty(&list))
3162 while (!list_empty(&list)) {
3163 de = list_first_entry(&list, struct io_defer_entry, list);
3164 list_del_init(&de->list);
3165 io_req_task_queue_fail(de->req, -ECANCELED);
3171 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
3173 struct io_tctx_node *node;
3174 enum io_wq_cancel cret;
3177 mutex_lock(&ctx->uring_lock);
3178 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3179 struct io_uring_task *tctx = node->task->io_uring;
3182 * io_wq will stay alive while we hold uring_lock, because it's
3183 * killed after ctx nodes, which requires to take the lock.
3185 if (!tctx || !tctx->io_wq)
3187 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
3188 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3190 mutex_unlock(&ctx->uring_lock);
3195 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
3196 struct task_struct *task,
3199 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
3200 struct io_uring_task *tctx = task ? task->io_uring : NULL;
3201 enum io_wq_cancel cret;
3204 /* set it so io_req_local_work_add() would wake us up */
3205 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
3206 atomic_set(&ctx->cq_wait_nr, 1);
3210 /* failed during ring init, it couldn't have issued any requests */
3215 ret |= io_uring_try_cancel_iowq(ctx);
3216 } else if (tctx && tctx->io_wq) {
3218 * Cancels requests of all rings, not only @ctx, but
3219 * it's fine as the task is in exit/exec.
3221 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
3223 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3226 /* SQPOLL thread does its own polling */
3227 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
3228 (ctx->sq_data && ctx->sq_data->thread == current)) {
3229 while (!wq_list_empty(&ctx->iopoll_list)) {
3230 io_iopoll_try_reap_events(ctx);
3236 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3237 io_allowed_defer_tw_run(ctx))
3238 ret |= io_run_local_work(ctx) > 0;
3239 ret |= io_cancel_defer_files(ctx, task, cancel_all);
3240 mutex_lock(&ctx->uring_lock);
3241 ret |= io_poll_remove_all(ctx, task, cancel_all);
3242 mutex_unlock(&ctx->uring_lock);
3243 ret |= io_kill_timeouts(ctx, task, cancel_all);
3245 ret |= io_run_task_work() > 0;
3249 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
3252 return atomic_read(&tctx->inflight_tracked);
3253 return percpu_counter_sum(&tctx->inflight);
3257 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3258 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3260 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
3262 struct io_uring_task *tctx = current->io_uring;
3263 struct io_ring_ctx *ctx;
3264 struct io_tctx_node *node;
3265 unsigned long index;
3269 WARN_ON_ONCE(sqd && sqd->thread != current);
3271 if (!current->io_uring)
3274 io_wq_exit_start(tctx->io_wq);
3276 atomic_inc(&tctx->in_cancel);
3280 io_uring_drop_tctx_refs(current);
3281 /* read completions before cancelations */
3282 inflight = tctx_inflight(tctx, !cancel_all);
3287 xa_for_each(&tctx->xa, index, node) {
3288 /* sqpoll task will cancel all its requests */
3289 if (node->ctx->sq_data)
3291 loop |= io_uring_try_cancel_requests(node->ctx,
3292 current, cancel_all);
3295 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3296 loop |= io_uring_try_cancel_requests(ctx,
3306 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3308 io_uring_drop_tctx_refs(current);
3309 xa_for_each(&tctx->xa, index, node) {
3310 if (!llist_empty(&node->ctx->work_llist)) {
3311 WARN_ON_ONCE(node->ctx->submitter_task &&
3312 node->ctx->submitter_task != current);
3317 * If we've seen completions, retry without waiting. This
3318 * avoids a race where a completion comes in before we did
3319 * prepare_to_wait().
3321 if (inflight == tctx_inflight(tctx, !cancel_all))
3324 finish_wait(&tctx->wait, &wait);
3327 io_uring_clean_tctx(tctx);
3330 * We shouldn't run task_works after cancel, so just leave
3331 * ->in_cancel set for normal exit.
3333 atomic_dec(&tctx->in_cancel);
3334 /* for exec all current's requests should be gone, kill tctx */
3335 __io_uring_free(current);
3339 void __io_uring_cancel(bool cancel_all)
3341 io_uring_cancel_generic(cancel_all, NULL);
3344 static void *io_uring_validate_mmap_request(struct file *file,
3345 loff_t pgoff, size_t sz)
3347 struct io_ring_ctx *ctx = file->private_data;
3348 loff_t offset = pgoff << PAGE_SHIFT;
3352 /* Don't allow mmap if the ring was setup without it */
3353 if (ctx->flags & IORING_SETUP_NO_MMAP)
3354 return ERR_PTR(-EINVAL);
3356 switch (offset & IORING_OFF_MMAP_MASK) {
3357 case IORING_OFF_SQ_RING:
3358 case IORING_OFF_CQ_RING:
3361 case IORING_OFF_SQES:
3364 case IORING_OFF_PBUF_RING: {
3367 bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT;
3368 mutex_lock(&ctx->uring_lock);
3369 ptr = io_pbuf_get_address(ctx, bgid);
3370 mutex_unlock(&ctx->uring_lock);
3372 return ERR_PTR(-EINVAL);
3376 return ERR_PTR(-EINVAL);
3379 page = virt_to_head_page(ptr);
3380 if (sz > page_size(page))
3381 return ERR_PTR(-EINVAL);
3388 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3390 size_t sz = vma->vm_end - vma->vm_start;
3394 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3396 return PTR_ERR(ptr);
3398 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3399 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3402 static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp,
3403 unsigned long addr, unsigned long len,
3404 unsigned long pgoff, unsigned long flags)
3406 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
3407 struct vm_unmapped_area_info info;
3411 * Do not allow to map to user-provided address to avoid breaking the
3412 * aliasing rules. Userspace is not able to guess the offset address of
3413 * kernel kmalloc()ed memory area.
3418 ptr = io_uring_validate_mmap_request(filp, pgoff, len);
3422 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
3424 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
3425 info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
3427 info.align_mask = PAGE_MASK & (SHM_COLOUR - 1UL);
3429 info.align_mask = PAGE_MASK & (SHMLBA - 1UL);
3431 info.align_offset = (unsigned long) ptr;
3434 * A failed mmap() very likely causes application failure,
3435 * so fall back to the bottom-up function here. This scenario
3436 * can happen with large stack limits and large mmap()
3439 addr = vm_unmapped_area(&info);
3440 if (offset_in_page(addr)) {
3442 info.low_limit = TASK_UNMAPPED_BASE;
3443 info.high_limit = mmap_end;
3444 addr = vm_unmapped_area(&info);
3450 #else /* !CONFIG_MMU */
3452 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3454 return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL;
3457 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3459 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3462 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3463 unsigned long addr, unsigned long len,
3464 unsigned long pgoff, unsigned long flags)
3468 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3470 return PTR_ERR(ptr);
3472 return (unsigned long) ptr;
3475 #endif /* !CONFIG_MMU */
3477 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3479 if (flags & IORING_ENTER_EXT_ARG) {
3480 struct io_uring_getevents_arg arg;
3482 if (argsz != sizeof(arg))
3484 if (copy_from_user(&arg, argp, sizeof(arg)))
3490 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3491 struct __kernel_timespec __user **ts,
3492 const sigset_t __user **sig)
3494 struct io_uring_getevents_arg arg;
3497 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3498 * is just a pointer to the sigset_t.
3500 if (!(flags & IORING_ENTER_EXT_ARG)) {
3501 *sig = (const sigset_t __user *) argp;
3507 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3508 * timespec and sigset_t pointers if good.
3510 if (*argsz != sizeof(arg))
3512 if (copy_from_user(&arg, argp, sizeof(arg)))
3516 *sig = u64_to_user_ptr(arg.sigmask);
3517 *argsz = arg.sigmask_sz;
3518 *ts = u64_to_user_ptr(arg.ts);
3522 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3523 u32, min_complete, u32, flags, const void __user *, argp,
3526 struct io_ring_ctx *ctx;
3530 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3531 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3532 IORING_ENTER_REGISTERED_RING)))
3536 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3537 * need only dereference our task private array to find it.
3539 if (flags & IORING_ENTER_REGISTERED_RING) {
3540 struct io_uring_task *tctx = current->io_uring;
3542 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3544 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3545 f.file = tctx->registered_rings[fd];
3547 if (unlikely(!f.file))
3551 if (unlikely(!f.file))
3554 if (unlikely(!io_is_uring_fops(f.file)))
3558 ctx = f.file->private_data;
3560 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3564 * For SQ polling, the thread will do all submissions and completions.
3565 * Just return the requested submit count, and wake the thread if
3569 if (ctx->flags & IORING_SETUP_SQPOLL) {
3570 io_cqring_overflow_flush(ctx);
3572 if (unlikely(ctx->sq_data->thread == NULL)) {
3576 if (flags & IORING_ENTER_SQ_WAKEUP)
3577 wake_up(&ctx->sq_data->wait);
3578 if (flags & IORING_ENTER_SQ_WAIT)
3579 io_sqpoll_wait_sq(ctx);
3582 } else if (to_submit) {
3583 ret = io_uring_add_tctx_node(ctx);
3587 mutex_lock(&ctx->uring_lock);
3588 ret = io_submit_sqes(ctx, to_submit);
3589 if (ret != to_submit) {
3590 mutex_unlock(&ctx->uring_lock);
3593 if (flags & IORING_ENTER_GETEVENTS) {
3594 if (ctx->syscall_iopoll)
3597 * Ignore errors, we'll soon call io_cqring_wait() and
3598 * it should handle ownership problems if any.
3600 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3601 (void)io_run_local_work_locked(ctx);
3603 mutex_unlock(&ctx->uring_lock);
3606 if (flags & IORING_ENTER_GETEVENTS) {
3609 if (ctx->syscall_iopoll) {
3611 * We disallow the app entering submit/complete with
3612 * polling, but we still need to lock the ring to
3613 * prevent racing with polled issue that got punted to
3616 mutex_lock(&ctx->uring_lock);
3618 ret2 = io_validate_ext_arg(flags, argp, argsz);
3619 if (likely(!ret2)) {
3620 min_complete = min(min_complete,
3622 ret2 = io_iopoll_check(ctx, min_complete);
3624 mutex_unlock(&ctx->uring_lock);
3626 const sigset_t __user *sig;
3627 struct __kernel_timespec __user *ts;
3629 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3630 if (likely(!ret2)) {
3631 min_complete = min(min_complete,
3633 ret2 = io_cqring_wait(ctx, min_complete, sig,
3642 * EBADR indicates that one or more CQE were dropped.
3643 * Once the user has been informed we can clear the bit
3644 * as they are obviously ok with those drops.
3646 if (unlikely(ret2 == -EBADR))
3647 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3656 static const struct file_operations io_uring_fops = {
3657 .release = io_uring_release,
3658 .mmap = io_uring_mmap,
3660 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3661 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3663 .get_unmapped_area = io_uring_mmu_get_unmapped_area,
3665 .poll = io_uring_poll,
3666 #ifdef CONFIG_PROC_FS
3667 .show_fdinfo = io_uring_show_fdinfo,
3671 bool io_is_uring_fops(struct file *file)
3673 return file->f_op == &io_uring_fops;
3676 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3677 struct io_uring_params *p)
3679 struct io_rings *rings;
3680 size_t size, sq_array_offset;
3683 /* make sure these are sane, as we already accounted them */
3684 ctx->sq_entries = p->sq_entries;
3685 ctx->cq_entries = p->cq_entries;
3687 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3688 if (size == SIZE_MAX)
3691 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3692 rings = io_mem_alloc(size);
3694 rings = io_rings_map(ctx, p->cq_off.user_addr, size);
3697 return PTR_ERR(rings);
3700 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3701 rings->sq_ring_mask = p->sq_entries - 1;
3702 rings->cq_ring_mask = p->cq_entries - 1;
3703 rings->sq_ring_entries = p->sq_entries;
3704 rings->cq_ring_entries = p->cq_entries;
3706 if (p->flags & IORING_SETUP_SQE128)
3707 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3709 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3710 if (size == SIZE_MAX) {
3715 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3716 ptr = io_mem_alloc(size);
3718 ptr = io_sqes_map(ctx, p->sq_off.user_addr, size);
3722 return PTR_ERR(ptr);
3729 static int io_uring_install_fd(struct file *file)
3733 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3736 fd_install(fd, file);
3741 * Allocate an anonymous fd, this is what constitutes the application
3742 * visible backing of an io_uring instance. The application mmaps this
3743 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3744 * we have to tie this fd to a socket for file garbage collection purposes.
3746 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3749 #if defined(CONFIG_UNIX)
3752 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3755 return ERR_PTR(ret);
3758 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3759 O_RDWR | O_CLOEXEC, NULL);
3760 #if defined(CONFIG_UNIX)
3762 sock_release(ctx->ring_sock);
3763 ctx->ring_sock = NULL;
3765 ctx->ring_sock->file = file;
3771 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3772 struct io_uring_params __user *params)
3774 struct io_ring_ctx *ctx;
3775 struct io_uring_task *tctx;
3781 if (entries > IORING_MAX_ENTRIES) {
3782 if (!(p->flags & IORING_SETUP_CLAMP))
3784 entries = IORING_MAX_ENTRIES;
3787 if ((p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
3788 && !(p->flags & IORING_SETUP_NO_MMAP))
3792 * Use twice as many entries for the CQ ring. It's possible for the
3793 * application to drive a higher depth than the size of the SQ ring,
3794 * since the sqes are only used at submission time. This allows for
3795 * some flexibility in overcommitting a bit. If the application has
3796 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3797 * of CQ ring entries manually.
3799 p->sq_entries = roundup_pow_of_two(entries);
3800 if (p->flags & IORING_SETUP_CQSIZE) {
3802 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3803 * to a power-of-two, if it isn't already. We do NOT impose
3804 * any cq vs sq ring sizing.
3808 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3809 if (!(p->flags & IORING_SETUP_CLAMP))
3811 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3813 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3814 if (p->cq_entries < p->sq_entries)
3817 p->cq_entries = 2 * p->sq_entries;
3820 ctx = io_ring_ctx_alloc(p);
3824 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3825 !(ctx->flags & IORING_SETUP_IOPOLL) &&
3826 !(ctx->flags & IORING_SETUP_SQPOLL))
3827 ctx->task_complete = true;
3830 * lazy poll_wq activation relies on ->task_complete for synchronisation
3831 * purposes, see io_activate_pollwq()
3833 if (!ctx->task_complete)
3834 ctx->poll_activated = true;
3837 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3838 * space applications don't need to do io completion events
3839 * polling again, they can rely on io_sq_thread to do polling
3840 * work, which can reduce cpu usage and uring_lock contention.
3842 if (ctx->flags & IORING_SETUP_IOPOLL &&
3843 !(ctx->flags & IORING_SETUP_SQPOLL))
3844 ctx->syscall_iopoll = 1;
3846 ctx->compat = in_compat_syscall();
3847 if (!capable(CAP_IPC_LOCK))
3848 ctx->user = get_uid(current_user());
3851 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3852 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3855 if (ctx->flags & IORING_SETUP_SQPOLL) {
3856 /* IPI related flags don't make sense with SQPOLL */
3857 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3858 IORING_SETUP_TASKRUN_FLAG |
3859 IORING_SETUP_DEFER_TASKRUN))
3861 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3862 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3863 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3865 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3866 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3868 ctx->notify_method = TWA_SIGNAL;
3872 * For DEFER_TASKRUN we require the completion task to be the same as the
3873 * submission task. This implies that there is only one submitter, so enforce
3876 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3877 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3882 * This is just grabbed for accounting purposes. When a process exits,
3883 * the mm is exited and dropped before the files, hence we need to hang
3884 * on to this mm purely for the purposes of being able to unaccount
3885 * memory (locked/pinned vm). It's not used for anything else.
3887 mmgrab(current->mm);
3888 ctx->mm_account = current->mm;
3890 ret = io_allocate_scq_urings(ctx, p);
3894 ret = io_sq_offload_create(ctx, p);
3898 ret = io_rsrc_init(ctx);
3902 p->sq_off.head = offsetof(struct io_rings, sq.head);
3903 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3904 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3905 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3906 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3907 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3908 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3909 p->sq_off.resv1 = 0;
3910 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3911 p->sq_off.user_addr = 0;
3913 p->cq_off.head = offsetof(struct io_rings, cq.head);
3914 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3915 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3916 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3917 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3918 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3919 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3920 p->cq_off.resv1 = 0;
3921 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3922 p->cq_off.user_addr = 0;
3924 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3925 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3926 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3927 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3928 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3929 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3930 IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING;
3932 if (copy_to_user(params, p, sizeof(*p))) {
3937 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
3938 && !(ctx->flags & IORING_SETUP_R_DISABLED))
3939 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
3941 file = io_uring_get_file(ctx);
3943 ret = PTR_ERR(file);
3947 ret = __io_uring_add_tctx_node(ctx);
3950 tctx = current->io_uring;
3953 * Install ring fd as the very last thing, so we don't risk someone
3954 * having closed it before we finish setup
3956 if (p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
3957 ret = io_ring_add_registered_file(tctx, file, 0, IO_RINGFD_REG_MAX);
3959 ret = io_uring_install_fd(file);
3963 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3966 io_ring_ctx_wait_and_kill(ctx);
3974 * Sets up an aio uring context, and returns the fd. Applications asks for a
3975 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3976 * params structure passed in.
3978 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3980 struct io_uring_params p;
3983 if (copy_from_user(&p, params, sizeof(p)))
3985 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3990 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3991 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3992 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3993 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3994 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3995 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3996 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN |
3997 IORING_SETUP_NO_MMAP | IORING_SETUP_REGISTERED_FD_ONLY))
4000 return io_uring_create(entries, &p, params);
4003 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
4004 struct io_uring_params __user *, params)
4006 return io_uring_setup(entries, params);
4009 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
4012 struct io_uring_probe *p;
4016 size = struct_size(p, ops, nr_args);
4017 if (size == SIZE_MAX)
4019 p = kzalloc(size, GFP_KERNEL);
4024 if (copy_from_user(p, arg, size))
4027 if (memchr_inv(p, 0, size))
4030 p->last_op = IORING_OP_LAST - 1;
4031 if (nr_args > IORING_OP_LAST)
4032 nr_args = IORING_OP_LAST;
4034 for (i = 0; i < nr_args; i++) {
4036 if (!io_issue_defs[i].not_supported)
4037 p->ops[i].flags = IO_URING_OP_SUPPORTED;
4042 if (copy_to_user(arg, p, size))
4049 static int io_register_personality(struct io_ring_ctx *ctx)
4051 const struct cred *creds;
4055 creds = get_current_cred();
4057 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
4058 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
4066 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
4067 void __user *arg, unsigned int nr_args)
4069 struct io_uring_restriction *res;
4073 /* Restrictions allowed only if rings started disabled */
4074 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
4077 /* We allow only a single restrictions registration */
4078 if (ctx->restrictions.registered)
4081 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
4084 size = array_size(nr_args, sizeof(*res));
4085 if (size == SIZE_MAX)
4088 res = memdup_user(arg, size);
4090 return PTR_ERR(res);
4094 for (i = 0; i < nr_args; i++) {
4095 switch (res[i].opcode) {
4096 case IORING_RESTRICTION_REGISTER_OP:
4097 if (res[i].register_op >= IORING_REGISTER_LAST) {
4102 __set_bit(res[i].register_op,
4103 ctx->restrictions.register_op);
4105 case IORING_RESTRICTION_SQE_OP:
4106 if (res[i].sqe_op >= IORING_OP_LAST) {
4111 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
4113 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
4114 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
4116 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
4117 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
4126 /* Reset all restrictions if an error happened */
4128 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
4130 ctx->restrictions.registered = true;
4136 static int io_register_enable_rings(struct io_ring_ctx *ctx)
4138 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
4141 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task) {
4142 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
4144 * Lazy activation attempts would fail if it was polled before
4145 * submitter_task is set.
4147 if (wq_has_sleeper(&ctx->poll_wq))
4148 io_activate_pollwq(ctx);
4151 if (ctx->restrictions.registered)
4152 ctx->restricted = 1;
4154 ctx->flags &= ~IORING_SETUP_R_DISABLED;
4155 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
4156 wake_up(&ctx->sq_data->wait);
4160 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
4161 void __user *arg, unsigned len)
4163 struct io_uring_task *tctx = current->io_uring;
4164 cpumask_var_t new_mask;
4167 if (!tctx || !tctx->io_wq)
4170 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4173 cpumask_clear(new_mask);
4174 if (len > cpumask_size())
4175 len = cpumask_size();
4177 if (in_compat_syscall()) {
4178 ret = compat_get_bitmap(cpumask_bits(new_mask),
4179 (const compat_ulong_t __user *)arg,
4180 len * 8 /* CHAR_BIT */);
4182 ret = copy_from_user(new_mask, arg, len);
4186 free_cpumask_var(new_mask);
4190 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
4191 free_cpumask_var(new_mask);
4195 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
4197 struct io_uring_task *tctx = current->io_uring;
4199 if (!tctx || !tctx->io_wq)
4202 return io_wq_cpu_affinity(tctx->io_wq, NULL);
4205 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
4207 __must_hold(&ctx->uring_lock)
4209 struct io_tctx_node *node;
4210 struct io_uring_task *tctx = NULL;
4211 struct io_sq_data *sqd = NULL;
4215 if (copy_from_user(new_count, arg, sizeof(new_count)))
4217 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4218 if (new_count[i] > INT_MAX)
4221 if (ctx->flags & IORING_SETUP_SQPOLL) {
4225 * Observe the correct sqd->lock -> ctx->uring_lock
4226 * ordering. Fine to drop uring_lock here, we hold
4229 refcount_inc(&sqd->refs);
4230 mutex_unlock(&ctx->uring_lock);
4231 mutex_lock(&sqd->lock);
4232 mutex_lock(&ctx->uring_lock);
4234 tctx = sqd->thread->io_uring;
4237 tctx = current->io_uring;
4240 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
4242 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4244 ctx->iowq_limits[i] = new_count[i];
4245 ctx->iowq_limits_set = true;
4247 if (tctx && tctx->io_wq) {
4248 ret = io_wq_max_workers(tctx->io_wq, new_count);
4252 memset(new_count, 0, sizeof(new_count));
4256 mutex_unlock(&sqd->lock);
4257 io_put_sq_data(sqd);
4260 if (copy_to_user(arg, new_count, sizeof(new_count)))
4263 /* that's it for SQPOLL, only the SQPOLL task creates requests */
4267 /* now propagate the restriction to all registered users */
4268 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
4269 struct io_uring_task *tctx = node->task->io_uring;
4271 if (WARN_ON_ONCE(!tctx->io_wq))
4274 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4275 new_count[i] = ctx->iowq_limits[i];
4276 /* ignore errors, it always returns zero anyway */
4277 (void)io_wq_max_workers(tctx->io_wq, new_count);
4282 mutex_unlock(&sqd->lock);
4283 io_put_sq_data(sqd);
4288 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
4289 void __user *arg, unsigned nr_args)
4290 __releases(ctx->uring_lock)
4291 __acquires(ctx->uring_lock)
4296 * We don't quiesce the refs for register anymore and so it can't be
4297 * dying as we're holding a file ref here.
4299 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
4302 if (ctx->submitter_task && ctx->submitter_task != current)
4305 if (ctx->restricted) {
4306 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
4307 if (!test_bit(opcode, ctx->restrictions.register_op))
4312 case IORING_REGISTER_BUFFERS:
4316 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
4318 case IORING_UNREGISTER_BUFFERS:
4322 ret = io_sqe_buffers_unregister(ctx);
4324 case IORING_REGISTER_FILES:
4328 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
4330 case IORING_UNREGISTER_FILES:
4334 ret = io_sqe_files_unregister(ctx);
4336 case IORING_REGISTER_FILES_UPDATE:
4337 ret = io_register_files_update(ctx, arg, nr_args);
4339 case IORING_REGISTER_EVENTFD:
4343 ret = io_eventfd_register(ctx, arg, 0);
4345 case IORING_REGISTER_EVENTFD_ASYNC:
4349 ret = io_eventfd_register(ctx, arg, 1);
4351 case IORING_UNREGISTER_EVENTFD:
4355 ret = io_eventfd_unregister(ctx);
4357 case IORING_REGISTER_PROBE:
4359 if (!arg || nr_args > 256)
4361 ret = io_probe(ctx, arg, nr_args);
4363 case IORING_REGISTER_PERSONALITY:
4367 ret = io_register_personality(ctx);
4369 case IORING_UNREGISTER_PERSONALITY:
4373 ret = io_unregister_personality(ctx, nr_args);
4375 case IORING_REGISTER_ENABLE_RINGS:
4379 ret = io_register_enable_rings(ctx);
4381 case IORING_REGISTER_RESTRICTIONS:
4382 ret = io_register_restrictions(ctx, arg, nr_args);
4384 case IORING_REGISTER_FILES2:
4385 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
4387 case IORING_REGISTER_FILES_UPDATE2:
4388 ret = io_register_rsrc_update(ctx, arg, nr_args,
4391 case IORING_REGISTER_BUFFERS2:
4392 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
4394 case IORING_REGISTER_BUFFERS_UPDATE:
4395 ret = io_register_rsrc_update(ctx, arg, nr_args,
4396 IORING_RSRC_BUFFER);
4398 case IORING_REGISTER_IOWQ_AFF:
4400 if (!arg || !nr_args)
4402 ret = io_register_iowq_aff(ctx, arg, nr_args);
4404 case IORING_UNREGISTER_IOWQ_AFF:
4408 ret = io_unregister_iowq_aff(ctx);
4410 case IORING_REGISTER_IOWQ_MAX_WORKERS:
4412 if (!arg || nr_args != 2)
4414 ret = io_register_iowq_max_workers(ctx, arg);
4416 case IORING_REGISTER_RING_FDS:
4417 ret = io_ringfd_register(ctx, arg, nr_args);
4419 case IORING_UNREGISTER_RING_FDS:
4420 ret = io_ringfd_unregister(ctx, arg, nr_args);
4422 case IORING_REGISTER_PBUF_RING:
4424 if (!arg || nr_args != 1)
4426 ret = io_register_pbuf_ring(ctx, arg);
4428 case IORING_UNREGISTER_PBUF_RING:
4430 if (!arg || nr_args != 1)
4432 ret = io_unregister_pbuf_ring(ctx, arg);
4434 case IORING_REGISTER_SYNC_CANCEL:
4436 if (!arg || nr_args != 1)
4438 ret = io_sync_cancel(ctx, arg);
4440 case IORING_REGISTER_FILE_ALLOC_RANGE:
4442 if (!arg || nr_args)
4444 ret = io_register_file_alloc_range(ctx, arg);
4454 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
4455 void __user *, arg, unsigned int, nr_args)
4457 struct io_ring_ctx *ctx;
4460 bool use_registered_ring;
4462 use_registered_ring = !!(opcode & IORING_REGISTER_USE_REGISTERED_RING);
4463 opcode &= ~IORING_REGISTER_USE_REGISTERED_RING;
4465 if (opcode >= IORING_REGISTER_LAST)
4468 if (use_registered_ring) {
4470 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
4471 * need only dereference our task private array to find it.
4473 struct io_uring_task *tctx = current->io_uring;
4475 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
4477 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
4478 f.file = tctx->registered_rings[fd];
4480 if (unlikely(!f.file))
4484 if (unlikely(!f.file))
4487 if (!io_is_uring_fops(f.file))
4491 ctx = f.file->private_data;
4493 mutex_lock(&ctx->uring_lock);
4494 ret = __io_uring_register(ctx, opcode, arg, nr_args);
4495 mutex_unlock(&ctx->uring_lock);
4496 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
4502 static int __init io_uring_init(void)
4504 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4505 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4506 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4509 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4510 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4511 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4512 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4513 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4514 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
4515 BUILD_BUG_SQE_ELEM(1, __u8, flags);
4516 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
4517 BUILD_BUG_SQE_ELEM(4, __s32, fd);
4518 BUILD_BUG_SQE_ELEM(8, __u64, off);
4519 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
4520 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
4521 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
4522 BUILD_BUG_SQE_ELEM(16, __u64, addr);
4523 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
4524 BUILD_BUG_SQE_ELEM(24, __u32, len);
4525 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
4526 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
4527 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4528 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
4529 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
4530 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
4531 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
4532 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
4533 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
4534 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
4535 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
4536 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
4537 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
4538 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
4539 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
4540 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
4541 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
4542 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
4543 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
4544 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
4545 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
4546 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
4547 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
4548 BUILD_BUG_SQE_ELEM(42, __u16, personality);
4549 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
4550 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
4551 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
4552 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
4553 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
4554 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4555 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
4557 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4558 sizeof(struct io_uring_rsrc_update));
4559 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4560 sizeof(struct io_uring_rsrc_update2));
4562 /* ->buf_index is u16 */
4563 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4564 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4565 offsetof(struct io_uring_buf_ring, tail));
4567 /* should fit into one byte */
4568 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
4569 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4570 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4572 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
4574 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4576 io_uring_optable_init();
4578 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4579 SLAB_ACCOUNT | SLAB_TYPESAFE_BY_RCU);
4582 __initcall(io_uring_init);