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)
629 __acquires(ctx->completion_lock)
631 if (!ctx->task_complete)
632 spin_lock(&ctx->completion_lock);
635 static inline void __io_cq_unlock(struct io_ring_ctx *ctx)
637 if (!ctx->task_complete)
638 spin_unlock(&ctx->completion_lock);
641 static inline void io_cq_lock(struct io_ring_ctx *ctx)
642 __acquires(ctx->completion_lock)
644 spin_lock(&ctx->completion_lock);
647 /* keep it inlined for io_submit_flush_completions() */
648 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
649 __releases(ctx->completion_lock)
651 io_commit_cqring(ctx);
653 io_commit_cqring_flush(ctx);
657 static void __io_cq_unlock_post_flush(struct io_ring_ctx *ctx)
658 __releases(ctx->completion_lock)
660 io_commit_cqring(ctx);
662 if (ctx->task_complete) {
664 * ->task_complete implies that only current might be waiting
665 * for CQEs, and obviously, we currently don't. No one is
666 * waiting, wakeups are futile, skip them.
668 io_commit_cqring_flush(ctx);
671 io_commit_cqring_flush(ctx);
676 void io_cq_unlock_post(struct io_ring_ctx *ctx)
677 __releases(ctx->completion_lock)
679 io_commit_cqring(ctx);
680 spin_unlock(&ctx->completion_lock);
681 io_commit_cqring_flush(ctx);
685 /* Returns true if there are no backlogged entries after the flush */
686 static void io_cqring_overflow_kill(struct io_ring_ctx *ctx)
688 struct io_overflow_cqe *ocqe;
691 spin_lock(&ctx->completion_lock);
692 list_splice_init(&ctx->cq_overflow_list, &list);
693 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
694 spin_unlock(&ctx->completion_lock);
696 while (!list_empty(&list)) {
697 ocqe = list_first_entry(&list, struct io_overflow_cqe, list);
698 list_del(&ocqe->list);
703 static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx)
705 size_t cqe_size = sizeof(struct io_uring_cqe);
707 if (__io_cqring_events(ctx) == ctx->cq_entries)
710 if (ctx->flags & IORING_SETUP_CQE32)
714 while (!list_empty(&ctx->cq_overflow_list)) {
715 struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true);
716 struct io_overflow_cqe *ocqe;
720 ocqe = list_first_entry(&ctx->cq_overflow_list,
721 struct io_overflow_cqe, list);
722 memcpy(cqe, &ocqe->cqe, cqe_size);
723 list_del(&ocqe->list);
727 if (list_empty(&ctx->cq_overflow_list)) {
728 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
729 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
731 io_cq_unlock_post(ctx);
734 static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx)
736 /* iopoll syncs against uring_lock, not completion_lock */
737 if (ctx->flags & IORING_SETUP_IOPOLL)
738 mutex_lock(&ctx->uring_lock);
739 __io_cqring_overflow_flush(ctx);
740 if (ctx->flags & IORING_SETUP_IOPOLL)
741 mutex_unlock(&ctx->uring_lock);
744 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx)
746 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
747 io_cqring_do_overflow_flush(ctx);
750 /* can be called by any task */
751 static void io_put_task_remote(struct task_struct *task)
753 struct io_uring_task *tctx = task->io_uring;
755 percpu_counter_sub(&tctx->inflight, 1);
756 if (unlikely(atomic_read(&tctx->in_cancel)))
757 wake_up(&tctx->wait);
758 put_task_struct(task);
761 /* used by a task to put its own references */
762 static void io_put_task_local(struct task_struct *task)
764 task->io_uring->cached_refs++;
767 /* must to be called somewhat shortly after putting a request */
768 static inline void io_put_task(struct task_struct *task)
770 if (likely(task == current))
771 io_put_task_local(task);
773 io_put_task_remote(task);
776 void io_task_refs_refill(struct io_uring_task *tctx)
778 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
780 percpu_counter_add(&tctx->inflight, refill);
781 refcount_add(refill, ¤t->usage);
782 tctx->cached_refs += refill;
785 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
787 struct io_uring_task *tctx = task->io_uring;
788 unsigned int refs = tctx->cached_refs;
791 tctx->cached_refs = 0;
792 percpu_counter_sub(&tctx->inflight, refs);
793 put_task_struct_many(task, refs);
797 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
798 s32 res, u32 cflags, u64 extra1, u64 extra2)
800 struct io_overflow_cqe *ocqe;
801 size_t ocq_size = sizeof(struct io_overflow_cqe);
802 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
804 lockdep_assert_held(&ctx->completion_lock);
807 ocq_size += sizeof(struct io_uring_cqe);
809 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
810 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
813 * If we're in ring overflow flush mode, or in task cancel mode,
814 * or cannot allocate an overflow entry, then we need to drop it
817 io_account_cq_overflow(ctx);
818 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
821 if (list_empty(&ctx->cq_overflow_list)) {
822 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
823 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
826 ocqe->cqe.user_data = user_data;
828 ocqe->cqe.flags = cflags;
830 ocqe->cqe.big_cqe[0] = extra1;
831 ocqe->cqe.big_cqe[1] = extra2;
833 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
837 bool io_req_cqe_overflow(struct io_kiocb *req)
839 if (!(req->flags & REQ_F_CQE32_INIT)) {
843 return io_cqring_event_overflow(req->ctx, req->cqe.user_data,
844 req->cqe.res, req->cqe.flags,
845 req->extra1, req->extra2);
849 * writes to the cq entry need to come after reading head; the
850 * control dependency is enough as we're using WRITE_ONCE to
853 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow)
855 struct io_rings *rings = ctx->rings;
856 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
857 unsigned int free, queued, len;
860 * Posting into the CQ when there are pending overflowed CQEs may break
861 * ordering guarantees, which will affect links, F_MORE users and more.
862 * Force overflow the completion.
864 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
867 /* userspace may cheat modifying the tail, be safe and do min */
868 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
869 free = ctx->cq_entries - queued;
870 /* we need a contiguous range, limit based on the current array offset */
871 len = min(free, ctx->cq_entries - off);
875 if (ctx->flags & IORING_SETUP_CQE32) {
880 ctx->cqe_cached = &rings->cqes[off];
881 ctx->cqe_sentinel = ctx->cqe_cached + len;
883 ctx->cached_cq_tail++;
885 if (ctx->flags & IORING_SETUP_CQE32)
887 return &rings->cqes[off];
890 static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
893 struct io_uring_cqe *cqe;
898 * If we can't get a cq entry, userspace overflowed the
899 * submission (by quite a lot). Increment the overflow count in
902 cqe = io_get_cqe(ctx);
904 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
906 WRITE_ONCE(cqe->user_data, user_data);
907 WRITE_ONCE(cqe->res, res);
908 WRITE_ONCE(cqe->flags, cflags);
910 if (ctx->flags & IORING_SETUP_CQE32) {
911 WRITE_ONCE(cqe->big_cqe[0], 0);
912 WRITE_ONCE(cqe->big_cqe[1], 0);
919 static void __io_flush_post_cqes(struct io_ring_ctx *ctx)
920 __must_hold(&ctx->uring_lock)
922 struct io_submit_state *state = &ctx->submit_state;
925 lockdep_assert_held(&ctx->uring_lock);
926 for (i = 0; i < state->cqes_count; i++) {
927 struct io_uring_cqe *cqe = &state->cqes[i];
929 if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) {
930 if (ctx->task_complete) {
931 spin_lock(&ctx->completion_lock);
932 io_cqring_event_overflow(ctx, cqe->user_data,
933 cqe->res, cqe->flags, 0, 0);
934 spin_unlock(&ctx->completion_lock);
936 io_cqring_event_overflow(ctx, cqe->user_data,
937 cqe->res, cqe->flags, 0, 0);
941 state->cqes_count = 0;
944 static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
950 filled = io_fill_cqe_aux(ctx, user_data, res, cflags);
951 if (!filled && allow_overflow)
952 filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
954 io_cq_unlock_post(ctx);
958 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags)
960 return __io_post_aux_cqe(ctx, user_data, res, cflags, true);
963 bool io_aux_cqe(const struct io_kiocb *req, bool defer, s32 res, u32 cflags,
966 struct io_ring_ctx *ctx = req->ctx;
967 u64 user_data = req->cqe.user_data;
968 struct io_uring_cqe *cqe;
971 return __io_post_aux_cqe(ctx, user_data, res, cflags, allow_overflow);
973 lockdep_assert_held(&ctx->uring_lock);
975 if (ctx->submit_state.cqes_count == ARRAY_SIZE(ctx->submit_state.cqes)) {
977 __io_flush_post_cqes(ctx);
978 /* no need to flush - flush is deferred */
979 __io_cq_unlock_post(ctx);
982 /* For defered completions this is not as strict as it is otherwise,
983 * however it's main job is to prevent unbounded posted completions,
984 * and in that it works just as well.
986 if (!allow_overflow && test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
989 cqe = &ctx->submit_state.cqes[ctx->submit_state.cqes_count++];
990 cqe->user_data = user_data;
996 static void __io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
998 struct io_ring_ctx *ctx = req->ctx;
999 struct io_rsrc_node *rsrc_node = NULL;
1002 if (!(req->flags & REQ_F_CQE_SKIP))
1003 io_fill_cqe_req(ctx, req);
1006 * If we're the last reference to this request, add to our locked
1009 if (req_ref_put_and_test(req)) {
1010 if (req->flags & IO_REQ_LINK_FLAGS) {
1011 if (req->flags & IO_DISARM_MASK)
1012 io_disarm_next(req);
1014 io_req_task_queue(req->link);
1018 io_put_kbuf_comp(req);
1019 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1021 if (!(req->flags & REQ_F_FIXED_FILE))
1022 io_put_file(req->file);
1024 rsrc_node = req->rsrc_node;
1026 * Selected buffer deallocation in io_clean_op() assumes that
1027 * we don't hold ->completion_lock. Clean them here to avoid
1030 io_put_task_remote(req->task);
1031 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
1032 ctx->locked_free_nr++;
1034 io_cq_unlock_post(ctx);
1037 io_ring_submit_lock(ctx, issue_flags);
1038 io_put_rsrc_node(ctx, rsrc_node);
1039 io_ring_submit_unlock(ctx, issue_flags);
1043 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
1045 if (req->ctx->task_complete && req->ctx->submitter_task != current) {
1046 req->io_task_work.func = io_req_task_complete;
1047 io_req_task_work_add(req);
1048 } else if (!(issue_flags & IO_URING_F_UNLOCKED) ||
1049 !(req->ctx->flags & IORING_SETUP_IOPOLL)) {
1050 __io_req_complete_post(req, issue_flags);
1052 struct io_ring_ctx *ctx = req->ctx;
1054 mutex_lock(&ctx->uring_lock);
1055 __io_req_complete_post(req, issue_flags & ~IO_URING_F_UNLOCKED);
1056 mutex_unlock(&ctx->uring_lock);
1060 void io_req_defer_failed(struct io_kiocb *req, s32 res)
1061 __must_hold(&ctx->uring_lock)
1063 const struct io_cold_def *def = &io_cold_defs[req->opcode];
1065 lockdep_assert_held(&req->ctx->uring_lock);
1068 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
1071 io_req_complete_defer(req);
1075 * Don't initialise the fields below on every allocation, but do that in
1076 * advance and keep them valid across allocations.
1078 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1082 req->async_data = NULL;
1083 /* not necessary, but safer to zero */
1087 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1088 struct io_submit_state *state)
1090 spin_lock(&ctx->completion_lock);
1091 wq_list_splice(&ctx->locked_free_list, &state->free_list);
1092 ctx->locked_free_nr = 0;
1093 spin_unlock(&ctx->completion_lock);
1097 * A request might get retired back into the request caches even before opcode
1098 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1099 * Because of that, io_alloc_req() should be called only under ->uring_lock
1100 * and with extra caution to not get a request that is still worked on.
1102 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
1103 __must_hold(&ctx->uring_lock)
1105 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1106 void *reqs[IO_REQ_ALLOC_BATCH];
1110 * If we have more than a batch's worth of requests in our IRQ side
1111 * locked cache, grab the lock and move them over to our submission
1114 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
1115 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
1116 if (!io_req_cache_empty(ctx))
1120 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
1123 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1124 * retry single alloc to be on the safe side.
1126 if (unlikely(ret <= 0)) {
1127 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1133 percpu_ref_get_many(&ctx->refs, ret);
1134 for (i = 0; i < ret; i++) {
1135 struct io_kiocb *req = reqs[i];
1137 io_preinit_req(req, ctx);
1138 io_req_add_to_cache(req, ctx);
1143 __cold void io_free_req(struct io_kiocb *req)
1145 /* refs were already put, restore them for io_req_task_complete() */
1146 req->flags &= ~REQ_F_REFCOUNT;
1147 /* we only want to free it, don't post CQEs */
1148 req->flags |= REQ_F_CQE_SKIP;
1149 req->io_task_work.func = io_req_task_complete;
1150 io_req_task_work_add(req);
1153 static void __io_req_find_next_prep(struct io_kiocb *req)
1155 struct io_ring_ctx *ctx = req->ctx;
1157 spin_lock(&ctx->completion_lock);
1158 io_disarm_next(req);
1159 spin_unlock(&ctx->completion_lock);
1162 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1164 struct io_kiocb *nxt;
1167 * If LINK is set, we have dependent requests in this chain. If we
1168 * didn't fail this request, queue the first one up, moving any other
1169 * dependencies to the next request. In case of failure, fail the rest
1172 if (unlikely(req->flags & IO_DISARM_MASK))
1173 __io_req_find_next_prep(req);
1179 static void ctx_flush_and_put(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1183 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1184 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1186 io_submit_flush_completions(ctx);
1187 mutex_unlock(&ctx->uring_lock);
1190 percpu_ref_put(&ctx->refs);
1193 static unsigned int handle_tw_list(struct llist_node *node,
1194 struct io_ring_ctx **ctx,
1195 struct io_tw_state *ts,
1196 struct llist_node *last)
1198 unsigned int count = 0;
1200 while (node && node != last) {
1201 struct llist_node *next = node->next;
1202 struct io_kiocb *req = container_of(node, struct io_kiocb,
1205 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1207 if (req->ctx != *ctx) {
1208 ctx_flush_and_put(*ctx, ts);
1210 /* if not contended, grab and improve batching */
1211 ts->locked = mutex_trylock(&(*ctx)->uring_lock);
1212 percpu_ref_get(&(*ctx)->refs);
1214 INDIRECT_CALL_2(req->io_task_work.func,
1215 io_poll_task_func, io_req_rw_complete,
1219 if (unlikely(need_resched())) {
1220 ctx_flush_and_put(*ctx, ts);
1230 * io_llist_xchg - swap all entries in a lock-less list
1231 * @head: the head of lock-less list to delete all entries
1232 * @new: new entry as the head of the list
1234 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1235 * The order of entries returned is from the newest to the oldest added one.
1237 static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1238 struct llist_node *new)
1240 return xchg(&head->first, new);
1244 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1245 * @head: the head of lock-less list to delete all entries
1246 * @old: expected old value of the first entry of the list
1247 * @new: new entry as the head of the list
1249 * perform a cmpxchg on the first entry of the list.
1252 static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1253 struct llist_node *old,
1254 struct llist_node *new)
1256 return cmpxchg(&head->first, old, new);
1259 void tctx_task_work(struct callback_head *cb)
1261 struct io_tw_state ts = {};
1262 struct io_ring_ctx *ctx = NULL;
1263 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1265 struct llist_node fake = {};
1266 struct llist_node *node;
1267 unsigned int loops = 0;
1268 unsigned int count = 0;
1270 if (unlikely(current->flags & PF_EXITING)) {
1271 io_fallback_tw(tctx);
1277 node = io_llist_xchg(&tctx->task_list, &fake);
1278 count += handle_tw_list(node, &ctx, &ts, &fake);
1280 /* skip expensive cmpxchg if there are items in the list */
1281 if (READ_ONCE(tctx->task_list.first) != &fake)
1283 if (ts.locked && !wq_list_empty(&ctx->submit_state.compl_reqs)) {
1284 io_submit_flush_completions(ctx);
1285 if (READ_ONCE(tctx->task_list.first) != &fake)
1288 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1289 } while (node != &fake);
1291 ctx_flush_and_put(ctx, &ts);
1293 /* relaxed read is enough as only the task itself sets ->in_cancel */
1294 if (unlikely(atomic_read(&tctx->in_cancel)))
1295 io_uring_drop_tctx_refs(current);
1297 trace_io_uring_task_work_run(tctx, count, loops);
1300 static __cold void io_fallback_tw(struct io_uring_task *tctx)
1302 struct llist_node *node = llist_del_all(&tctx->task_list);
1303 struct io_kiocb *req;
1306 req = container_of(node, struct io_kiocb, io_task_work.node);
1308 if (llist_add(&req->io_task_work.node,
1309 &req->ctx->fallback_llist))
1310 schedule_delayed_work(&req->ctx->fallback_work, 1);
1314 static inline void io_req_local_work_add(struct io_kiocb *req, unsigned flags)
1316 struct io_ring_ctx *ctx = req->ctx;
1317 unsigned nr_wait, nr_tw, nr_tw_prev;
1318 struct llist_node *first;
1320 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK))
1321 flags &= ~IOU_F_TWQ_LAZY_WAKE;
1323 first = READ_ONCE(ctx->work_llist.first);
1327 struct io_kiocb *first_req = container_of(first,
1331 * Might be executed at any moment, rely on
1332 * SLAB_TYPESAFE_BY_RCU to keep it alive.
1334 nr_tw_prev = READ_ONCE(first_req->nr_tw);
1336 nr_tw = nr_tw_prev + 1;
1337 /* Large enough to fail the nr_wait comparison below */
1338 if (!(flags & IOU_F_TWQ_LAZY_WAKE))
1342 req->io_task_work.node.next = first;
1343 } while (!try_cmpxchg(&ctx->work_llist.first, &first,
1344 &req->io_task_work.node));
1347 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1348 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1350 io_eventfd_signal(ctx);
1353 nr_wait = atomic_read(&ctx->cq_wait_nr);
1354 /* no one is waiting */
1357 /* either not enough or the previous add has already woken it up */
1358 if (nr_wait > nr_tw || nr_tw_prev >= nr_wait)
1360 /* pairs with set_current_state() in io_cqring_wait() */
1361 smp_mb__after_atomic();
1362 wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE);
1365 static void io_req_normal_work_add(struct io_kiocb *req)
1367 struct io_uring_task *tctx = req->task->io_uring;
1368 struct io_ring_ctx *ctx = req->ctx;
1370 /* task_work already pending, we're done */
1371 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1374 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1375 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1377 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1380 io_fallback_tw(tctx);
1383 void __io_req_task_work_add(struct io_kiocb *req, unsigned flags)
1385 if (req->ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
1387 io_req_local_work_add(req, flags);
1390 io_req_normal_work_add(req);
1394 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1396 struct llist_node *node;
1398 node = llist_del_all(&ctx->work_llist);
1400 struct io_kiocb *req = container_of(node, struct io_kiocb,
1404 io_req_normal_work_add(req);
1408 static int __io_run_local_work(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1410 struct llist_node *node;
1411 unsigned int loops = 0;
1414 if (WARN_ON_ONCE(ctx->submitter_task != current))
1416 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1417 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1420 * llists are in reverse order, flip it back the right way before
1421 * running the pending items.
1423 node = llist_reverse_order(io_llist_xchg(&ctx->work_llist, NULL));
1425 struct llist_node *next = node->next;
1426 struct io_kiocb *req = container_of(node, struct io_kiocb,
1428 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1429 INDIRECT_CALL_2(req->io_task_work.func,
1430 io_poll_task_func, io_req_rw_complete,
1437 if (!llist_empty(&ctx->work_llist))
1440 io_submit_flush_completions(ctx);
1441 if (!llist_empty(&ctx->work_llist))
1444 trace_io_uring_local_work_run(ctx, ret, loops);
1448 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx)
1450 struct io_tw_state ts = { .locked = true, };
1453 if (llist_empty(&ctx->work_llist))
1456 ret = __io_run_local_work(ctx, &ts);
1457 /* shouldn't happen! */
1458 if (WARN_ON_ONCE(!ts.locked))
1459 mutex_lock(&ctx->uring_lock);
1463 static int io_run_local_work(struct io_ring_ctx *ctx)
1465 struct io_tw_state ts = {};
1468 ts.locked = mutex_trylock(&ctx->uring_lock);
1469 ret = __io_run_local_work(ctx, &ts);
1471 mutex_unlock(&ctx->uring_lock);
1476 static void io_req_task_cancel(struct io_kiocb *req, struct io_tw_state *ts)
1478 io_tw_lock(req->ctx, ts);
1479 io_req_defer_failed(req, req->cqe.res);
1482 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts)
1484 io_tw_lock(req->ctx, ts);
1485 /* req->task == current here, checking PF_EXITING is safe */
1486 if (unlikely(req->task->flags & PF_EXITING))
1487 io_req_defer_failed(req, -EFAULT);
1488 else if (req->flags & REQ_F_FORCE_ASYNC)
1489 io_queue_iowq(req, ts);
1494 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1496 io_req_set_res(req, ret, 0);
1497 req->io_task_work.func = io_req_task_cancel;
1498 io_req_task_work_add(req);
1501 void io_req_task_queue(struct io_kiocb *req)
1503 req->io_task_work.func = io_req_task_submit;
1504 io_req_task_work_add(req);
1507 void io_queue_next(struct io_kiocb *req)
1509 struct io_kiocb *nxt = io_req_find_next(req);
1512 io_req_task_queue(nxt);
1515 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1516 __must_hold(&ctx->uring_lock)
1519 struct io_kiocb *req = container_of(node, struct io_kiocb,
1522 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1523 if (req->flags & REQ_F_REFCOUNT) {
1524 node = req->comp_list.next;
1525 if (!req_ref_put_and_test(req))
1528 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1529 struct async_poll *apoll = req->apoll;
1531 if (apoll->double_poll)
1532 kfree(apoll->double_poll);
1533 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1535 req->flags &= ~REQ_F_POLLED;
1537 if (req->flags & IO_REQ_LINK_FLAGS)
1539 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1542 if (!(req->flags & REQ_F_FIXED_FILE))
1543 io_put_file(req->file);
1545 io_req_put_rsrc_locked(req, ctx);
1547 io_put_task(req->task);
1548 node = req->comp_list.next;
1549 io_req_add_to_cache(req, ctx);
1553 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1554 __must_hold(&ctx->uring_lock)
1556 struct io_submit_state *state = &ctx->submit_state;
1557 struct io_wq_work_node *node;
1560 /* must come first to preserve CQE ordering in failure cases */
1561 if (state->cqes_count)
1562 __io_flush_post_cqes(ctx);
1563 __wq_list_for_each(node, &state->compl_reqs) {
1564 struct io_kiocb *req = container_of(node, struct io_kiocb,
1567 if (!(req->flags & REQ_F_CQE_SKIP) &&
1568 unlikely(!__io_fill_cqe_req(ctx, req))) {
1569 if (ctx->task_complete) {
1570 spin_lock(&ctx->completion_lock);
1571 io_req_cqe_overflow(req);
1572 spin_unlock(&ctx->completion_lock);
1574 io_req_cqe_overflow(req);
1578 __io_cq_unlock_post_flush(ctx);
1580 if (!wq_list_empty(&ctx->submit_state.compl_reqs)) {
1581 io_free_batch_list(ctx, state->compl_reqs.first);
1582 INIT_WQ_LIST(&state->compl_reqs);
1586 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1588 /* See comment at the top of this file */
1590 return __io_cqring_events(ctx);
1594 * We can't just wait for polled events to come to us, we have to actively
1595 * find and complete them.
1597 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1599 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1602 mutex_lock(&ctx->uring_lock);
1603 while (!wq_list_empty(&ctx->iopoll_list)) {
1604 /* let it sleep and repeat later if can't complete a request */
1605 if (io_do_iopoll(ctx, true) == 0)
1608 * Ensure we allow local-to-the-cpu processing to take place,
1609 * in this case we need to ensure that we reap all events.
1610 * Also let task_work, etc. to progress by releasing the mutex
1612 if (need_resched()) {
1613 mutex_unlock(&ctx->uring_lock);
1615 mutex_lock(&ctx->uring_lock);
1618 mutex_unlock(&ctx->uring_lock);
1621 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1623 unsigned int nr_events = 0;
1625 unsigned long check_cq;
1627 if (!io_allowed_run_tw(ctx))
1630 check_cq = READ_ONCE(ctx->check_cq);
1631 if (unlikely(check_cq)) {
1632 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1633 __io_cqring_overflow_flush(ctx);
1635 * Similarly do not spin if we have not informed the user of any
1638 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1642 * Don't enter poll loop if we already have events pending.
1643 * If we do, we can potentially be spinning for commands that
1644 * already triggered a CQE (eg in error).
1646 if (io_cqring_events(ctx))
1651 * If a submit got punted to a workqueue, we can have the
1652 * application entering polling for a command before it gets
1653 * issued. That app will hold the uring_lock for the duration
1654 * of the poll right here, so we need to take a breather every
1655 * now and then to ensure that the issue has a chance to add
1656 * the poll to the issued list. Otherwise we can spin here
1657 * forever, while the workqueue is stuck trying to acquire the
1660 if (wq_list_empty(&ctx->iopoll_list) ||
1661 io_task_work_pending(ctx)) {
1662 u32 tail = ctx->cached_cq_tail;
1664 (void) io_run_local_work_locked(ctx);
1666 if (task_work_pending(current) ||
1667 wq_list_empty(&ctx->iopoll_list)) {
1668 mutex_unlock(&ctx->uring_lock);
1670 mutex_lock(&ctx->uring_lock);
1672 /* some requests don't go through iopoll_list */
1673 if (tail != ctx->cached_cq_tail ||
1674 wq_list_empty(&ctx->iopoll_list))
1677 ret = io_do_iopoll(ctx, !min);
1682 } while (nr_events < min && !need_resched());
1687 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts)
1690 io_req_complete_defer(req);
1692 io_req_complete_post(req, IO_URING_F_UNLOCKED);
1696 * After the iocb has been issued, it's safe to be found on the poll list.
1697 * Adding the kiocb to the list AFTER submission ensures that we don't
1698 * find it from a io_do_iopoll() thread before the issuer is done
1699 * accessing the kiocb cookie.
1701 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1703 struct io_ring_ctx *ctx = req->ctx;
1704 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1706 /* workqueue context doesn't hold uring_lock, grab it now */
1707 if (unlikely(needs_lock))
1708 mutex_lock(&ctx->uring_lock);
1711 * Track whether we have multiple files in our lists. This will impact
1712 * how we do polling eventually, not spinning if we're on potentially
1713 * different devices.
1715 if (wq_list_empty(&ctx->iopoll_list)) {
1716 ctx->poll_multi_queue = false;
1717 } else if (!ctx->poll_multi_queue) {
1718 struct io_kiocb *list_req;
1720 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1722 if (list_req->file != req->file)
1723 ctx->poll_multi_queue = true;
1727 * For fast devices, IO may have already completed. If it has, add
1728 * it to the front so we find it first.
1730 if (READ_ONCE(req->iopoll_completed))
1731 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1733 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1735 if (unlikely(needs_lock)) {
1737 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1738 * in sq thread task context or in io worker task context. If
1739 * current task context is sq thread, we don't need to check
1740 * whether should wake up sq thread.
1742 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1743 wq_has_sleeper(&ctx->sq_data->wait))
1744 wake_up(&ctx->sq_data->wait);
1746 mutex_unlock(&ctx->uring_lock);
1750 unsigned int io_file_get_flags(struct file *file)
1752 unsigned int res = 0;
1754 if (S_ISREG(file_inode(file)->i_mode))
1756 if ((file->f_flags & O_NONBLOCK) || (file->f_mode & FMODE_NOWAIT))
1757 res |= REQ_F_SUPPORT_NOWAIT;
1761 bool io_alloc_async_data(struct io_kiocb *req)
1763 WARN_ON_ONCE(!io_cold_defs[req->opcode].async_size);
1764 req->async_data = kmalloc(io_cold_defs[req->opcode].async_size, GFP_KERNEL);
1765 if (req->async_data) {
1766 req->flags |= REQ_F_ASYNC_DATA;
1772 int io_req_prep_async(struct io_kiocb *req)
1774 const struct io_cold_def *cdef = &io_cold_defs[req->opcode];
1775 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1777 /* assign early for deferred execution for non-fixed file */
1778 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE) && !req->file)
1779 req->file = io_file_get_normal(req, req->cqe.fd);
1780 if (!cdef->prep_async)
1782 if (WARN_ON_ONCE(req_has_async_data(req)))
1784 if (!def->manual_alloc) {
1785 if (io_alloc_async_data(req))
1788 return cdef->prep_async(req);
1791 static u32 io_get_sequence(struct io_kiocb *req)
1793 u32 seq = req->ctx->cached_sq_head;
1794 struct io_kiocb *cur;
1796 /* need original cached_sq_head, but it was increased for each req */
1797 io_for_each_link(cur, req)
1802 static __cold void io_drain_req(struct io_kiocb *req)
1803 __must_hold(&ctx->uring_lock)
1805 struct io_ring_ctx *ctx = req->ctx;
1806 struct io_defer_entry *de;
1808 u32 seq = io_get_sequence(req);
1810 /* Still need defer if there is pending req in defer list. */
1811 spin_lock(&ctx->completion_lock);
1812 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1813 spin_unlock(&ctx->completion_lock);
1815 ctx->drain_active = false;
1816 io_req_task_queue(req);
1819 spin_unlock(&ctx->completion_lock);
1821 io_prep_async_link(req);
1822 de = kmalloc(sizeof(*de), GFP_KERNEL);
1825 io_req_defer_failed(req, ret);
1829 spin_lock(&ctx->completion_lock);
1830 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1831 spin_unlock(&ctx->completion_lock);
1836 trace_io_uring_defer(req);
1839 list_add_tail(&de->list, &ctx->defer_list);
1840 spin_unlock(&ctx->completion_lock);
1843 static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def,
1844 unsigned int issue_flags)
1846 if (req->file || !def->needs_file)
1849 if (req->flags & REQ_F_FIXED_FILE)
1850 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1852 req->file = io_file_get_normal(req, req->cqe.fd);
1857 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1859 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1860 const struct cred *creds = NULL;
1863 if (unlikely(!io_assign_file(req, def, issue_flags)))
1866 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1867 creds = override_creds(req->creds);
1869 if (!def->audit_skip)
1870 audit_uring_entry(req->opcode);
1872 ret = def->issue(req, issue_flags);
1874 if (!def->audit_skip)
1875 audit_uring_exit(!ret, ret);
1878 revert_creds(creds);
1880 if (ret == IOU_OK) {
1881 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1882 io_req_complete_defer(req);
1884 io_req_complete_post(req, issue_flags);
1885 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1888 /* If the op doesn't have a file, we're not polling for it */
1889 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
1890 io_iopoll_req_issued(req, issue_flags);
1895 int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts)
1897 io_tw_lock(req->ctx, ts);
1898 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT|
1899 IO_URING_F_COMPLETE_DEFER);
1902 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1904 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1905 struct io_kiocb *nxt = NULL;
1907 if (req_ref_put_and_test(req)) {
1908 if (req->flags & IO_REQ_LINK_FLAGS)
1909 nxt = io_req_find_next(req);
1912 return nxt ? &nxt->work : NULL;
1915 void io_wq_submit_work(struct io_wq_work *work)
1917 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1918 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1919 unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ;
1920 bool needs_poll = false;
1921 int ret = 0, err = -ECANCELED;
1923 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1924 if (!(req->flags & REQ_F_REFCOUNT))
1925 __io_req_set_refcount(req, 2);
1929 io_arm_ltimeout(req);
1931 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1932 if (work->flags & IO_WQ_WORK_CANCEL) {
1934 io_req_task_queue_fail(req, err);
1937 if (!io_assign_file(req, def, issue_flags)) {
1939 work->flags |= IO_WQ_WORK_CANCEL;
1943 if (req->flags & REQ_F_FORCE_ASYNC) {
1944 bool opcode_poll = def->pollin || def->pollout;
1946 if (opcode_poll && file_can_poll(req->file)) {
1948 issue_flags |= IO_URING_F_NONBLOCK;
1953 ret = io_issue_sqe(req, issue_flags);
1957 * We can get EAGAIN for iopolled IO even though we're
1958 * forcing a sync submission from here, since we can't
1959 * wait for request slots on the block side.
1962 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1968 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1970 /* aborted or ready, in either case retry blocking */
1972 issue_flags &= ~IO_URING_F_NONBLOCK;
1975 /* avoid locking problems by failing it from a clean context */
1977 io_req_task_queue_fail(req, ret);
1980 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1981 unsigned int issue_flags)
1983 struct io_ring_ctx *ctx = req->ctx;
1984 struct io_fixed_file *slot;
1985 struct file *file = NULL;
1987 io_ring_submit_lock(ctx, issue_flags);
1989 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
1991 fd = array_index_nospec(fd, ctx->nr_user_files);
1992 slot = io_fixed_file_slot(&ctx->file_table, fd);
1993 file = io_slot_file(slot);
1994 req->flags |= io_slot_flags(slot);
1995 io_req_set_rsrc_node(req, ctx, 0);
1997 io_ring_submit_unlock(ctx, issue_flags);
2001 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
2003 struct file *file = fget(fd);
2005 trace_io_uring_file_get(req, fd);
2007 /* we don't allow fixed io_uring files */
2008 if (file && io_is_uring_fops(file))
2009 io_req_track_inflight(req);
2013 static void io_queue_async(struct io_kiocb *req, int ret)
2014 __must_hold(&req->ctx->uring_lock)
2016 struct io_kiocb *linked_timeout;
2018 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
2019 io_req_defer_failed(req, ret);
2023 linked_timeout = io_prep_linked_timeout(req);
2025 switch (io_arm_poll_handler(req, 0)) {
2026 case IO_APOLL_READY:
2027 io_kbuf_recycle(req, 0);
2028 io_req_task_queue(req);
2030 case IO_APOLL_ABORTED:
2031 io_kbuf_recycle(req, 0);
2032 io_queue_iowq(req, NULL);
2039 io_queue_linked_timeout(linked_timeout);
2042 static inline void io_queue_sqe(struct io_kiocb *req)
2043 __must_hold(&req->ctx->uring_lock)
2047 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
2050 * We async punt it if the file wasn't marked NOWAIT, or if the file
2051 * doesn't support non-blocking read/write attempts
2054 io_arm_ltimeout(req);
2056 io_queue_async(req, ret);
2059 static void io_queue_sqe_fallback(struct io_kiocb *req)
2060 __must_hold(&req->ctx->uring_lock)
2062 if (unlikely(req->flags & REQ_F_FAIL)) {
2064 * We don't submit, fail them all, for that replace hardlinks
2065 * with normal links. Extra REQ_F_LINK is tolerated.
2067 req->flags &= ~REQ_F_HARDLINK;
2068 req->flags |= REQ_F_LINK;
2069 io_req_defer_failed(req, req->cqe.res);
2071 int ret = io_req_prep_async(req);
2073 if (unlikely(ret)) {
2074 io_req_defer_failed(req, ret);
2078 if (unlikely(req->ctx->drain_active))
2081 io_queue_iowq(req, NULL);
2086 * Check SQE restrictions (opcode and flags).
2088 * Returns 'true' if SQE is allowed, 'false' otherwise.
2090 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
2091 struct io_kiocb *req,
2092 unsigned int sqe_flags)
2094 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
2097 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
2098 ctx->restrictions.sqe_flags_required)
2101 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
2102 ctx->restrictions.sqe_flags_required))
2108 static void io_init_req_drain(struct io_kiocb *req)
2110 struct io_ring_ctx *ctx = req->ctx;
2111 struct io_kiocb *head = ctx->submit_state.link.head;
2113 ctx->drain_active = true;
2116 * If we need to drain a request in the middle of a link, drain
2117 * the head request and the next request/link after the current
2118 * link. Considering sequential execution of links,
2119 * REQ_F_IO_DRAIN will be maintained for every request of our
2122 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2123 ctx->drain_next = true;
2127 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
2128 const struct io_uring_sqe *sqe)
2129 __must_hold(&ctx->uring_lock)
2131 const struct io_issue_def *def;
2132 unsigned int sqe_flags;
2136 /* req is partially pre-initialised, see io_preinit_req() */
2137 req->opcode = opcode = READ_ONCE(sqe->opcode);
2138 /* same numerical values with corresponding REQ_F_*, safe to copy */
2139 req->flags = sqe_flags = READ_ONCE(sqe->flags);
2140 req->cqe.user_data = READ_ONCE(sqe->user_data);
2142 req->rsrc_node = NULL;
2143 req->task = current;
2145 if (unlikely(opcode >= IORING_OP_LAST)) {
2149 def = &io_issue_defs[opcode];
2150 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2151 /* enforce forwards compatibility on users */
2152 if (sqe_flags & ~SQE_VALID_FLAGS)
2154 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2155 if (!def->buffer_select)
2157 req->buf_index = READ_ONCE(sqe->buf_group);
2159 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2160 ctx->drain_disabled = true;
2161 if (sqe_flags & IOSQE_IO_DRAIN) {
2162 if (ctx->drain_disabled)
2164 io_init_req_drain(req);
2167 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2168 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2170 /* knock it to the slow queue path, will be drained there */
2171 if (ctx->drain_active)
2172 req->flags |= REQ_F_FORCE_ASYNC;
2173 /* if there is no link, we're at "next" request and need to drain */
2174 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2175 ctx->drain_next = false;
2176 ctx->drain_active = true;
2177 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2181 if (!def->ioprio && sqe->ioprio)
2183 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2186 if (def->needs_file) {
2187 struct io_submit_state *state = &ctx->submit_state;
2189 req->cqe.fd = READ_ONCE(sqe->fd);
2192 * Plug now if we have more than 2 IO left after this, and the
2193 * target is potentially a read/write to block based storage.
2195 if (state->need_plug && def->plug) {
2196 state->plug_started = true;
2197 state->need_plug = false;
2198 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2202 personality = READ_ONCE(sqe->personality);
2206 req->creds = xa_load(&ctx->personalities, personality);
2209 get_cred(req->creds);
2210 ret = security_uring_override_creds(req->creds);
2212 put_cred(req->creds);
2215 req->flags |= REQ_F_CREDS;
2218 return def->prep(req, sqe);
2221 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2222 struct io_kiocb *req, int ret)
2224 struct io_ring_ctx *ctx = req->ctx;
2225 struct io_submit_link *link = &ctx->submit_state.link;
2226 struct io_kiocb *head = link->head;
2228 trace_io_uring_req_failed(sqe, req, ret);
2231 * Avoid breaking links in the middle as it renders links with SQPOLL
2232 * unusable. Instead of failing eagerly, continue assembling the link if
2233 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2234 * should find the flag and handle the rest.
2236 req_fail_link_node(req, ret);
2237 if (head && !(head->flags & REQ_F_FAIL))
2238 req_fail_link_node(head, -ECANCELED);
2240 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2242 link->last->link = req;
2246 io_queue_sqe_fallback(req);
2251 link->last->link = req;
2258 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2259 const struct io_uring_sqe *sqe)
2260 __must_hold(&ctx->uring_lock)
2262 struct io_submit_link *link = &ctx->submit_state.link;
2265 ret = io_init_req(ctx, req, sqe);
2267 return io_submit_fail_init(sqe, req, ret);
2269 trace_io_uring_submit_req(req);
2272 * If we already have a head request, queue this one for async
2273 * submittal once the head completes. If we don't have a head but
2274 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2275 * submitted sync once the chain is complete. If none of those
2276 * conditions are true (normal request), then just queue it.
2278 if (unlikely(link->head)) {
2279 ret = io_req_prep_async(req);
2281 return io_submit_fail_init(sqe, req, ret);
2283 trace_io_uring_link(req, link->head);
2284 link->last->link = req;
2287 if (req->flags & IO_REQ_LINK_FLAGS)
2289 /* last request of the link, flush it */
2292 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2295 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2296 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2297 if (req->flags & IO_REQ_LINK_FLAGS) {
2302 io_queue_sqe_fallback(req);
2312 * Batched submission is done, ensure local IO is flushed out.
2314 static void io_submit_state_end(struct io_ring_ctx *ctx)
2316 struct io_submit_state *state = &ctx->submit_state;
2318 if (unlikely(state->link.head))
2319 io_queue_sqe_fallback(state->link.head);
2320 /* flush only after queuing links as they can generate completions */
2321 io_submit_flush_completions(ctx);
2322 if (state->plug_started)
2323 blk_finish_plug(&state->plug);
2327 * Start submission side cache.
2329 static void io_submit_state_start(struct io_submit_state *state,
2330 unsigned int max_ios)
2332 state->plug_started = false;
2333 state->need_plug = max_ios > 2;
2334 state->submit_nr = max_ios;
2335 /* set only head, no need to init link_last in advance */
2336 state->link.head = NULL;
2339 static void io_commit_sqring(struct io_ring_ctx *ctx)
2341 struct io_rings *rings = ctx->rings;
2344 * Ensure any loads from the SQEs are done at this point,
2345 * since once we write the new head, the application could
2346 * write new data to them.
2348 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2352 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2353 * that is mapped by userspace. This means that care needs to be taken to
2354 * ensure that reads are stable, as we cannot rely on userspace always
2355 * being a good citizen. If members of the sqe are validated and then later
2356 * used, it's important that those reads are done through READ_ONCE() to
2357 * prevent a re-load down the line.
2359 static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe)
2361 unsigned head, mask = ctx->sq_entries - 1;
2362 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2365 * The cached sq head (or cq tail) serves two purposes:
2367 * 1) allows us to batch the cost of updating the user visible
2369 * 2) allows the kernel side to track the head on its own, even
2370 * though the application is the one updating it.
2372 head = READ_ONCE(ctx->sq_array[sq_idx]);
2373 if (likely(head < ctx->sq_entries)) {
2374 /* double index for 128-byte SQEs, twice as long */
2375 if (ctx->flags & IORING_SETUP_SQE128)
2377 *sqe = &ctx->sq_sqes[head];
2381 /* drop invalid entries */
2383 WRITE_ONCE(ctx->rings->sq_dropped,
2384 READ_ONCE(ctx->rings->sq_dropped) + 1);
2388 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2389 __must_hold(&ctx->uring_lock)
2391 unsigned int entries = io_sqring_entries(ctx);
2395 if (unlikely(!entries))
2397 /* make sure SQ entry isn't read before tail */
2398 ret = left = min(nr, entries);
2399 io_get_task_refs(left);
2400 io_submit_state_start(&ctx->submit_state, left);
2403 const struct io_uring_sqe *sqe;
2404 struct io_kiocb *req;
2406 if (unlikely(!io_alloc_req(ctx, &req)))
2408 if (unlikely(!io_get_sqe(ctx, &sqe))) {
2409 io_req_add_to_cache(req, ctx);
2414 * Continue submitting even for sqe failure if the
2415 * ring was setup with IORING_SETUP_SUBMIT_ALL
2417 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2418 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2424 if (unlikely(left)) {
2426 /* try again if it submitted nothing and can't allocate a req */
2427 if (!ret && io_req_cache_empty(ctx))
2429 current->io_uring->cached_refs += left;
2432 io_submit_state_end(ctx);
2433 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2434 io_commit_sqring(ctx);
2438 struct io_wait_queue {
2439 struct wait_queue_entry wq;
2440 struct io_ring_ctx *ctx;
2442 unsigned nr_timeouts;
2446 static inline bool io_has_work(struct io_ring_ctx *ctx)
2448 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
2449 !llist_empty(&ctx->work_llist);
2452 static inline bool io_should_wake(struct io_wait_queue *iowq)
2454 struct io_ring_ctx *ctx = iowq->ctx;
2455 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
2458 * Wake up if we have enough events, or if a timeout occurred since we
2459 * started waiting. For timeouts, we always want to return to userspace,
2460 * regardless of event count.
2462 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2465 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2466 int wake_flags, void *key)
2468 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq);
2471 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2472 * the task, and the next invocation will do it.
2474 if (io_should_wake(iowq) || io_has_work(iowq->ctx))
2475 return autoremove_wake_function(curr, mode, wake_flags, key);
2479 int io_run_task_work_sig(struct io_ring_ctx *ctx)
2481 if (!llist_empty(&ctx->work_llist)) {
2482 __set_current_state(TASK_RUNNING);
2483 if (io_run_local_work(ctx) > 0)
2486 if (io_run_task_work() > 0)
2488 if (task_sigpending(current))
2493 /* when returns >0, the caller should retry */
2494 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2495 struct io_wait_queue *iowq)
2497 if (unlikely(READ_ONCE(ctx->check_cq)))
2499 if (unlikely(!llist_empty(&ctx->work_llist)))
2501 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL)))
2503 if (unlikely(task_sigpending(current)))
2505 if (unlikely(io_should_wake(iowq)))
2507 if (iowq->timeout == KTIME_MAX)
2509 else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS))
2515 * Wait until events become available, if we don't already have some. The
2516 * application must reap them itself, as they reside on the shared cq ring.
2518 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2519 const sigset_t __user *sig, size_t sigsz,
2520 struct __kernel_timespec __user *uts)
2522 struct io_wait_queue iowq;
2523 struct io_rings *rings = ctx->rings;
2526 if (!io_allowed_run_tw(ctx))
2528 if (!llist_empty(&ctx->work_llist))
2529 io_run_local_work(ctx);
2531 io_cqring_overflow_flush(ctx);
2532 /* if user messes with these they will just get an early return */
2533 if (__io_cqring_events_user(ctx) >= min_events)
2537 #ifdef CONFIG_COMPAT
2538 if (in_compat_syscall())
2539 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2543 ret = set_user_sigmask(sig, sigsz);
2549 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2550 iowq.wq.private = current;
2551 INIT_LIST_HEAD(&iowq.wq.entry);
2553 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2554 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2555 iowq.timeout = KTIME_MAX;
2558 struct timespec64 ts;
2560 if (get_timespec64(&ts, uts))
2562 iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2565 trace_io_uring_cqring_wait(ctx, min_events);
2567 unsigned long check_cq;
2569 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
2570 int nr_wait = (int) iowq.cq_tail - READ_ONCE(ctx->rings->cq.tail);
2572 atomic_set(&ctx->cq_wait_nr, nr_wait);
2573 set_current_state(TASK_INTERRUPTIBLE);
2575 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2576 TASK_INTERRUPTIBLE);
2579 ret = io_cqring_wait_schedule(ctx, &iowq);
2580 __set_current_state(TASK_RUNNING);
2581 atomic_set(&ctx->cq_wait_nr, 0);
2586 * Run task_work after scheduling and before io_should_wake().
2587 * If we got woken because of task_work being processed, run it
2588 * now rather than let the caller do another wait loop.
2591 if (!llist_empty(&ctx->work_llist))
2592 io_run_local_work(ctx);
2594 check_cq = READ_ONCE(ctx->check_cq);
2595 if (unlikely(check_cq)) {
2596 /* let the caller flush overflows, retry */
2597 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2598 io_cqring_do_overflow_flush(ctx);
2599 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) {
2605 if (io_should_wake(&iowq)) {
2612 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
2613 finish_wait(&ctx->cq_wait, &iowq.wq);
2614 restore_saved_sigmask_unless(ret == -EINTR);
2616 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2619 static void io_mem_free(void *ptr)
2626 page = virt_to_head_page(ptr);
2627 if (put_page_testzero(page))
2628 free_compound_page(page);
2631 static void io_pages_free(struct page ***pages, int npages)
2633 struct page **page_array;
2638 page_array = *pages;
2639 for (i = 0; i < npages; i++)
2640 unpin_user_page(page_array[i]);
2645 static void *__io_uaddr_map(struct page ***pages, unsigned short *npages,
2646 unsigned long uaddr, size_t size)
2648 struct page **page_array;
2649 unsigned int nr_pages;
2654 if (uaddr & (PAGE_SIZE - 1) || !size)
2655 return ERR_PTR(-EINVAL);
2657 nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
2658 if (nr_pages > USHRT_MAX)
2659 return ERR_PTR(-EINVAL);
2660 page_array = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
2662 return ERR_PTR(-ENOMEM);
2664 ret = pin_user_pages_fast(uaddr, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
2666 if (ret != nr_pages) {
2668 io_pages_free(&page_array, ret > 0 ? ret : 0);
2669 return ret < 0 ? ERR_PTR(ret) : ERR_PTR(-EFAULT);
2672 * Should be a single page. If the ring is small enough that we can
2673 * use a normal page, that is fine. If we need multiple pages, then
2674 * userspace should use a huge page. That's the only way to guarantee
2675 * that we get contigious memory, outside of just being lucky or
2676 * (currently) having low memory fragmentation.
2678 if (page_array[0] != page_array[ret - 1])
2680 *pages = page_array;
2682 return page_to_virt(page_array[0]);
2685 static void *io_rings_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2688 return __io_uaddr_map(&ctx->ring_pages, &ctx->n_ring_pages, uaddr,
2692 static void *io_sqes_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2695 return __io_uaddr_map(&ctx->sqe_pages, &ctx->n_sqe_pages, uaddr,
2699 static void io_rings_free(struct io_ring_ctx *ctx)
2701 if (!(ctx->flags & IORING_SETUP_NO_MMAP)) {
2702 io_mem_free(ctx->rings);
2703 io_mem_free(ctx->sq_sqes);
2705 ctx->sq_sqes = NULL;
2707 io_pages_free(&ctx->ring_pages, ctx->n_ring_pages);
2708 io_pages_free(&ctx->sqe_pages, ctx->n_sqe_pages);
2712 static void *io_mem_alloc(size_t size)
2714 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2717 ret = (void *) __get_free_pages(gfp, get_order(size));
2720 return ERR_PTR(-ENOMEM);
2723 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2724 unsigned int cq_entries, size_t *sq_offset)
2726 struct io_rings *rings;
2727 size_t off, sq_array_size;
2729 off = struct_size(rings, cqes, cq_entries);
2730 if (off == SIZE_MAX)
2732 if (ctx->flags & IORING_SETUP_CQE32) {
2733 if (check_shl_overflow(off, 1, &off))
2738 off = ALIGN(off, SMP_CACHE_BYTES);
2746 sq_array_size = array_size(sizeof(u32), sq_entries);
2747 if (sq_array_size == SIZE_MAX)
2750 if (check_add_overflow(off, sq_array_size, &off))
2756 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2757 unsigned int eventfd_async)
2759 struct io_ev_fd *ev_fd;
2760 __s32 __user *fds = arg;
2763 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2764 lockdep_is_held(&ctx->uring_lock));
2768 if (copy_from_user(&fd, fds, sizeof(*fds)))
2771 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2775 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2776 if (IS_ERR(ev_fd->cq_ev_fd)) {
2777 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2782 spin_lock(&ctx->completion_lock);
2783 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2784 spin_unlock(&ctx->completion_lock);
2786 ev_fd->eventfd_async = eventfd_async;
2787 ctx->has_evfd = true;
2788 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2789 atomic_set(&ev_fd->refs, 1);
2790 atomic_set(&ev_fd->ops, 0);
2794 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2796 struct io_ev_fd *ev_fd;
2798 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2799 lockdep_is_held(&ctx->uring_lock));
2801 ctx->has_evfd = false;
2802 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2803 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops))
2804 call_rcu(&ev_fd->rcu, io_eventfd_ops);
2811 static void io_req_caches_free(struct io_ring_ctx *ctx)
2813 struct io_kiocb *req;
2816 mutex_lock(&ctx->uring_lock);
2817 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
2819 while (!io_req_cache_empty(ctx)) {
2820 req = io_extract_req(ctx);
2821 kmem_cache_free(req_cachep, req);
2825 percpu_ref_put_many(&ctx->refs, nr);
2826 mutex_unlock(&ctx->uring_lock);
2829 static void io_rsrc_node_cache_free(struct io_cache_entry *entry)
2831 kfree(container_of(entry, struct io_rsrc_node, cache));
2834 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2836 io_sq_thread_finish(ctx);
2837 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2838 if (WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)))
2841 mutex_lock(&ctx->uring_lock);
2843 __io_sqe_buffers_unregister(ctx);
2845 __io_sqe_files_unregister(ctx);
2846 io_cqring_overflow_kill(ctx);
2847 io_eventfd_unregister(ctx);
2848 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2849 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2850 io_destroy_buffers(ctx);
2851 mutex_unlock(&ctx->uring_lock);
2853 put_cred(ctx->sq_creds);
2854 if (ctx->submitter_task)
2855 put_task_struct(ctx->submitter_task);
2857 /* there are no registered resources left, nobody uses it */
2859 io_rsrc_node_destroy(ctx, ctx->rsrc_node);
2861 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2863 #if defined(CONFIG_UNIX)
2864 if (ctx->ring_sock) {
2865 ctx->ring_sock->file = NULL; /* so that iput() is called */
2866 sock_release(ctx->ring_sock);
2869 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2871 io_alloc_cache_free(&ctx->rsrc_node_cache, io_rsrc_node_cache_free);
2872 if (ctx->mm_account) {
2873 mmdrop(ctx->mm_account);
2874 ctx->mm_account = NULL;
2878 percpu_ref_exit(&ctx->refs);
2879 free_uid(ctx->user);
2880 io_req_caches_free(ctx);
2882 io_wq_put_hash(ctx->hash_map);
2883 kfree(ctx->cancel_table.hbs);
2884 kfree(ctx->cancel_table_locked.hbs);
2885 kfree(ctx->dummy_ubuf);
2887 xa_destroy(&ctx->io_bl_xa);
2891 static __cold void io_activate_pollwq_cb(struct callback_head *cb)
2893 struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx,
2896 mutex_lock(&ctx->uring_lock);
2897 ctx->poll_activated = true;
2898 mutex_unlock(&ctx->uring_lock);
2901 * Wake ups for some events between start of polling and activation
2902 * might've been lost due to loose synchronisation.
2904 wake_up_all(&ctx->poll_wq);
2905 percpu_ref_put(&ctx->refs);
2908 static __cold void io_activate_pollwq(struct io_ring_ctx *ctx)
2910 spin_lock(&ctx->completion_lock);
2911 /* already activated or in progress */
2912 if (ctx->poll_activated || ctx->poll_wq_task_work.func)
2914 if (WARN_ON_ONCE(!ctx->task_complete))
2916 if (!ctx->submitter_task)
2919 * with ->submitter_task only the submitter task completes requests, we
2920 * only need to sync with it, which is done by injecting a tw
2922 init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb);
2923 percpu_ref_get(&ctx->refs);
2924 if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL))
2925 percpu_ref_put(&ctx->refs);
2927 spin_unlock(&ctx->completion_lock);
2930 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2932 struct io_ring_ctx *ctx = file->private_data;
2935 if (unlikely(!ctx->poll_activated))
2936 io_activate_pollwq(ctx);
2938 poll_wait(file, &ctx->poll_wq, wait);
2940 * synchronizes with barrier from wq_has_sleeper call in
2944 if (!io_sqring_full(ctx))
2945 mask |= EPOLLOUT | EPOLLWRNORM;
2948 * Don't flush cqring overflow list here, just do a simple check.
2949 * Otherwise there could possible be ABBA deadlock:
2952 * lock(&ctx->uring_lock);
2954 * lock(&ctx->uring_lock);
2957 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2958 * pushes them to do the flush.
2961 if (__io_cqring_events_user(ctx) || io_has_work(ctx))
2962 mask |= EPOLLIN | EPOLLRDNORM;
2967 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2969 const struct cred *creds;
2971 creds = xa_erase(&ctx->personalities, id);
2980 struct io_tctx_exit {
2981 struct callback_head task_work;
2982 struct completion completion;
2983 struct io_ring_ctx *ctx;
2986 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2988 struct io_uring_task *tctx = current->io_uring;
2989 struct io_tctx_exit *work;
2991 work = container_of(cb, struct io_tctx_exit, task_work);
2993 * When @in_cancel, we're in cancellation and it's racy to remove the
2994 * node. It'll be removed by the end of cancellation, just ignore it.
2995 * tctx can be NULL if the queueing of this task_work raced with
2996 * work cancelation off the exec path.
2998 if (tctx && !atomic_read(&tctx->in_cancel))
2999 io_uring_del_tctx_node((unsigned long)work->ctx);
3000 complete(&work->completion);
3003 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
3005 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3007 return req->ctx == data;
3010 static __cold void io_ring_exit_work(struct work_struct *work)
3012 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
3013 unsigned long timeout = jiffies + HZ * 60 * 5;
3014 unsigned long interval = HZ / 20;
3015 struct io_tctx_exit exit;
3016 struct io_tctx_node *node;
3020 * If we're doing polled IO and end up having requests being
3021 * submitted async (out-of-line), then completions can come in while
3022 * we're waiting for refs to drop. We need to reap these manually,
3023 * as nobody else will be looking for them.
3026 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
3027 mutex_lock(&ctx->uring_lock);
3028 io_cqring_overflow_kill(ctx);
3029 mutex_unlock(&ctx->uring_lock);
3032 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3033 io_move_task_work_from_local(ctx);
3035 while (io_uring_try_cancel_requests(ctx, NULL, true))
3039 struct io_sq_data *sqd = ctx->sq_data;
3040 struct task_struct *tsk;
3042 io_sq_thread_park(sqd);
3044 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
3045 io_wq_cancel_cb(tsk->io_uring->io_wq,
3046 io_cancel_ctx_cb, ctx, true);
3047 io_sq_thread_unpark(sqd);
3050 io_req_caches_free(ctx);
3052 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
3053 /* there is little hope left, don't run it too often */
3057 * This is really an uninterruptible wait, as it has to be
3058 * complete. But it's also run from a kworker, which doesn't
3059 * take signals, so it's fine to make it interruptible. This
3060 * avoids scenarios where we knowingly can wait much longer
3061 * on completions, for example if someone does a SIGSTOP on
3062 * a task that needs to finish task_work to make this loop
3063 * complete. That's a synthetic situation that should not
3064 * cause a stuck task backtrace, and hence a potential panic
3065 * on stuck tasks if that is enabled.
3067 } while (!wait_for_completion_interruptible_timeout(&ctx->ref_comp, interval));
3069 init_completion(&exit.completion);
3070 init_task_work(&exit.task_work, io_tctx_exit_cb);
3073 * Some may use context even when all refs and requests have been put,
3074 * and they are free to do so while still holding uring_lock or
3075 * completion_lock, see io_req_task_submit(). Apart from other work,
3076 * this lock/unlock section also waits them to finish.
3078 mutex_lock(&ctx->uring_lock);
3079 while (!list_empty(&ctx->tctx_list)) {
3080 WARN_ON_ONCE(time_after(jiffies, timeout));
3082 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
3084 /* don't spin on a single task if cancellation failed */
3085 list_rotate_left(&ctx->tctx_list);
3086 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
3087 if (WARN_ON_ONCE(ret))
3090 mutex_unlock(&ctx->uring_lock);
3092 * See comment above for
3093 * wait_for_completion_interruptible_timeout() on why this
3094 * wait is marked as interruptible.
3096 wait_for_completion_interruptible(&exit.completion);
3097 mutex_lock(&ctx->uring_lock);
3099 mutex_unlock(&ctx->uring_lock);
3100 spin_lock(&ctx->completion_lock);
3101 spin_unlock(&ctx->completion_lock);
3103 /* pairs with RCU read section in io_req_local_work_add() */
3104 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3107 io_ring_ctx_free(ctx);
3110 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
3112 unsigned long index;
3113 struct creds *creds;
3115 mutex_lock(&ctx->uring_lock);
3116 percpu_ref_kill(&ctx->refs);
3117 xa_for_each(&ctx->personalities, index, creds)
3118 io_unregister_personality(ctx, index);
3120 io_poll_remove_all(ctx, NULL, true);
3121 mutex_unlock(&ctx->uring_lock);
3124 * If we failed setting up the ctx, we might not have any rings
3125 * and therefore did not submit any requests
3128 io_kill_timeouts(ctx, NULL, true);
3130 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
3132 * Use system_unbound_wq to avoid spawning tons of event kworkers
3133 * if we're exiting a ton of rings at the same time. It just adds
3134 * noise and overhead, there's no discernable change in runtime
3135 * over using system_wq.
3137 queue_work(system_unbound_wq, &ctx->exit_work);
3140 static int io_uring_release(struct inode *inode, struct file *file)
3142 struct io_ring_ctx *ctx = file->private_data;
3144 file->private_data = NULL;
3145 io_ring_ctx_wait_and_kill(ctx);
3149 struct io_task_cancel {
3150 struct task_struct *task;
3154 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
3156 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3157 struct io_task_cancel *cancel = data;
3159 return io_match_task_safe(req, cancel->task, cancel->all);
3162 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
3163 struct task_struct *task,
3166 struct io_defer_entry *de;
3169 spin_lock(&ctx->completion_lock);
3170 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
3171 if (io_match_task_safe(de->req, task, cancel_all)) {
3172 list_cut_position(&list, &ctx->defer_list, &de->list);
3176 spin_unlock(&ctx->completion_lock);
3177 if (list_empty(&list))
3180 while (!list_empty(&list)) {
3181 de = list_first_entry(&list, struct io_defer_entry, list);
3182 list_del_init(&de->list);
3183 io_req_task_queue_fail(de->req, -ECANCELED);
3189 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
3191 struct io_tctx_node *node;
3192 enum io_wq_cancel cret;
3195 mutex_lock(&ctx->uring_lock);
3196 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3197 struct io_uring_task *tctx = node->task->io_uring;
3200 * io_wq will stay alive while we hold uring_lock, because it's
3201 * killed after ctx nodes, which requires to take the lock.
3203 if (!tctx || !tctx->io_wq)
3205 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
3206 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3208 mutex_unlock(&ctx->uring_lock);
3213 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
3214 struct task_struct *task,
3217 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
3218 struct io_uring_task *tctx = task ? task->io_uring : NULL;
3219 enum io_wq_cancel cret;
3222 /* set it so io_req_local_work_add() would wake us up */
3223 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
3224 atomic_set(&ctx->cq_wait_nr, 1);
3228 /* failed during ring init, it couldn't have issued any requests */
3233 ret |= io_uring_try_cancel_iowq(ctx);
3234 } else if (tctx && tctx->io_wq) {
3236 * Cancels requests of all rings, not only @ctx, but
3237 * it's fine as the task is in exit/exec.
3239 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
3241 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3244 /* SQPOLL thread does its own polling */
3245 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
3246 (ctx->sq_data && ctx->sq_data->thread == current)) {
3247 while (!wq_list_empty(&ctx->iopoll_list)) {
3248 io_iopoll_try_reap_events(ctx);
3254 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3255 io_allowed_defer_tw_run(ctx))
3256 ret |= io_run_local_work(ctx) > 0;
3257 ret |= io_cancel_defer_files(ctx, task, cancel_all);
3258 mutex_lock(&ctx->uring_lock);
3259 ret |= io_poll_remove_all(ctx, task, cancel_all);
3260 mutex_unlock(&ctx->uring_lock);
3261 ret |= io_kill_timeouts(ctx, task, cancel_all);
3263 ret |= io_run_task_work() > 0;
3267 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
3270 return atomic_read(&tctx->inflight_tracked);
3271 return percpu_counter_sum(&tctx->inflight);
3275 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3276 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3278 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
3280 struct io_uring_task *tctx = current->io_uring;
3281 struct io_ring_ctx *ctx;
3282 struct io_tctx_node *node;
3283 unsigned long index;
3287 WARN_ON_ONCE(sqd && sqd->thread != current);
3289 if (!current->io_uring)
3292 io_wq_exit_start(tctx->io_wq);
3294 atomic_inc(&tctx->in_cancel);
3298 io_uring_drop_tctx_refs(current);
3299 /* read completions before cancelations */
3300 inflight = tctx_inflight(tctx, !cancel_all);
3305 xa_for_each(&tctx->xa, index, node) {
3306 /* sqpoll task will cancel all its requests */
3307 if (node->ctx->sq_data)
3309 loop |= io_uring_try_cancel_requests(node->ctx,
3310 current, cancel_all);
3313 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3314 loop |= io_uring_try_cancel_requests(ctx,
3324 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3326 io_uring_drop_tctx_refs(current);
3327 xa_for_each(&tctx->xa, index, node) {
3328 if (!llist_empty(&node->ctx->work_llist)) {
3329 WARN_ON_ONCE(node->ctx->submitter_task &&
3330 node->ctx->submitter_task != current);
3335 * If we've seen completions, retry without waiting. This
3336 * avoids a race where a completion comes in before we did
3337 * prepare_to_wait().
3339 if (inflight == tctx_inflight(tctx, !cancel_all))
3342 finish_wait(&tctx->wait, &wait);
3345 io_uring_clean_tctx(tctx);
3348 * We shouldn't run task_works after cancel, so just leave
3349 * ->in_cancel set for normal exit.
3351 atomic_dec(&tctx->in_cancel);
3352 /* for exec all current's requests should be gone, kill tctx */
3353 __io_uring_free(current);
3357 void __io_uring_cancel(bool cancel_all)
3359 io_uring_cancel_generic(cancel_all, NULL);
3362 static void *io_uring_validate_mmap_request(struct file *file,
3363 loff_t pgoff, size_t sz)
3365 struct io_ring_ctx *ctx = file->private_data;
3366 loff_t offset = pgoff << PAGE_SHIFT;
3370 /* Don't allow mmap if the ring was setup without it */
3371 if (ctx->flags & IORING_SETUP_NO_MMAP)
3372 return ERR_PTR(-EINVAL);
3374 switch (offset & IORING_OFF_MMAP_MASK) {
3375 case IORING_OFF_SQ_RING:
3376 case IORING_OFF_CQ_RING:
3379 case IORING_OFF_SQES:
3382 case IORING_OFF_PBUF_RING: {
3385 bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT;
3386 mutex_lock(&ctx->uring_lock);
3387 ptr = io_pbuf_get_address(ctx, bgid);
3388 mutex_unlock(&ctx->uring_lock);
3390 return ERR_PTR(-EINVAL);
3394 return ERR_PTR(-EINVAL);
3397 page = virt_to_head_page(ptr);
3398 if (sz > page_size(page))
3399 return ERR_PTR(-EINVAL);
3406 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3408 size_t sz = vma->vm_end - vma->vm_start;
3412 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3414 return PTR_ERR(ptr);
3416 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3417 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3420 static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp,
3421 unsigned long addr, unsigned long len,
3422 unsigned long pgoff, unsigned long flags)
3424 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
3425 struct vm_unmapped_area_info info;
3429 * Do not allow to map to user-provided address to avoid breaking the
3430 * aliasing rules. Userspace is not able to guess the offset address of
3431 * kernel kmalloc()ed memory area.
3436 ptr = io_uring_validate_mmap_request(filp, pgoff, len);
3440 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
3442 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
3443 info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
3445 info.align_mask = PAGE_MASK & (SHM_COLOUR - 1UL);
3447 info.align_mask = PAGE_MASK & (SHMLBA - 1UL);
3449 info.align_offset = (unsigned long) ptr;
3452 * A failed mmap() very likely causes application failure,
3453 * so fall back to the bottom-up function here. This scenario
3454 * can happen with large stack limits and large mmap()
3457 addr = vm_unmapped_area(&info);
3458 if (offset_in_page(addr)) {
3460 info.low_limit = TASK_UNMAPPED_BASE;
3461 info.high_limit = mmap_end;
3462 addr = vm_unmapped_area(&info);
3468 #else /* !CONFIG_MMU */
3470 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3472 return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL;
3475 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3477 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3480 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3481 unsigned long addr, unsigned long len,
3482 unsigned long pgoff, unsigned long flags)
3486 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3488 return PTR_ERR(ptr);
3490 return (unsigned long) ptr;
3493 #endif /* !CONFIG_MMU */
3495 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3497 if (flags & IORING_ENTER_EXT_ARG) {
3498 struct io_uring_getevents_arg arg;
3500 if (argsz != sizeof(arg))
3502 if (copy_from_user(&arg, argp, sizeof(arg)))
3508 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3509 struct __kernel_timespec __user **ts,
3510 const sigset_t __user **sig)
3512 struct io_uring_getevents_arg arg;
3515 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3516 * is just a pointer to the sigset_t.
3518 if (!(flags & IORING_ENTER_EXT_ARG)) {
3519 *sig = (const sigset_t __user *) argp;
3525 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3526 * timespec and sigset_t pointers if good.
3528 if (*argsz != sizeof(arg))
3530 if (copy_from_user(&arg, argp, sizeof(arg)))
3534 *sig = u64_to_user_ptr(arg.sigmask);
3535 *argsz = arg.sigmask_sz;
3536 *ts = u64_to_user_ptr(arg.ts);
3540 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3541 u32, min_complete, u32, flags, const void __user *, argp,
3544 struct io_ring_ctx *ctx;
3548 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3549 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3550 IORING_ENTER_REGISTERED_RING)))
3554 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3555 * need only dereference our task private array to find it.
3557 if (flags & IORING_ENTER_REGISTERED_RING) {
3558 struct io_uring_task *tctx = current->io_uring;
3560 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3562 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3563 f.file = tctx->registered_rings[fd];
3565 if (unlikely(!f.file))
3569 if (unlikely(!f.file))
3572 if (unlikely(!io_is_uring_fops(f.file)))
3576 ctx = f.file->private_data;
3578 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3582 * For SQ polling, the thread will do all submissions and completions.
3583 * Just return the requested submit count, and wake the thread if
3587 if (ctx->flags & IORING_SETUP_SQPOLL) {
3588 io_cqring_overflow_flush(ctx);
3590 if (unlikely(ctx->sq_data->thread == NULL)) {
3594 if (flags & IORING_ENTER_SQ_WAKEUP)
3595 wake_up(&ctx->sq_data->wait);
3596 if (flags & IORING_ENTER_SQ_WAIT)
3597 io_sqpoll_wait_sq(ctx);
3600 } else if (to_submit) {
3601 ret = io_uring_add_tctx_node(ctx);
3605 mutex_lock(&ctx->uring_lock);
3606 ret = io_submit_sqes(ctx, to_submit);
3607 if (ret != to_submit) {
3608 mutex_unlock(&ctx->uring_lock);
3611 if (flags & IORING_ENTER_GETEVENTS) {
3612 if (ctx->syscall_iopoll)
3615 * Ignore errors, we'll soon call io_cqring_wait() and
3616 * it should handle ownership problems if any.
3618 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3619 (void)io_run_local_work_locked(ctx);
3621 mutex_unlock(&ctx->uring_lock);
3624 if (flags & IORING_ENTER_GETEVENTS) {
3627 if (ctx->syscall_iopoll) {
3629 * We disallow the app entering submit/complete with
3630 * polling, but we still need to lock the ring to
3631 * prevent racing with polled issue that got punted to
3634 mutex_lock(&ctx->uring_lock);
3636 ret2 = io_validate_ext_arg(flags, argp, argsz);
3637 if (likely(!ret2)) {
3638 min_complete = min(min_complete,
3640 ret2 = io_iopoll_check(ctx, min_complete);
3642 mutex_unlock(&ctx->uring_lock);
3644 const sigset_t __user *sig;
3645 struct __kernel_timespec __user *ts;
3647 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3648 if (likely(!ret2)) {
3649 min_complete = min(min_complete,
3651 ret2 = io_cqring_wait(ctx, min_complete, sig,
3660 * EBADR indicates that one or more CQE were dropped.
3661 * Once the user has been informed we can clear the bit
3662 * as they are obviously ok with those drops.
3664 if (unlikely(ret2 == -EBADR))
3665 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3674 static const struct file_operations io_uring_fops = {
3675 .release = io_uring_release,
3676 .mmap = io_uring_mmap,
3678 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3679 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3681 .get_unmapped_area = io_uring_mmu_get_unmapped_area,
3683 .poll = io_uring_poll,
3684 #ifdef CONFIG_PROC_FS
3685 .show_fdinfo = io_uring_show_fdinfo,
3689 bool io_is_uring_fops(struct file *file)
3691 return file->f_op == &io_uring_fops;
3694 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3695 struct io_uring_params *p)
3697 struct io_rings *rings;
3698 size_t size, sq_array_offset;
3701 /* make sure these are sane, as we already accounted them */
3702 ctx->sq_entries = p->sq_entries;
3703 ctx->cq_entries = p->cq_entries;
3705 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3706 if (size == SIZE_MAX)
3709 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3710 rings = io_mem_alloc(size);
3712 rings = io_rings_map(ctx, p->cq_off.user_addr, size);
3715 return PTR_ERR(rings);
3718 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3719 rings->sq_ring_mask = p->sq_entries - 1;
3720 rings->cq_ring_mask = p->cq_entries - 1;
3721 rings->sq_ring_entries = p->sq_entries;
3722 rings->cq_ring_entries = p->cq_entries;
3724 if (p->flags & IORING_SETUP_SQE128)
3725 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3727 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3728 if (size == SIZE_MAX) {
3733 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3734 ptr = io_mem_alloc(size);
3736 ptr = io_sqes_map(ctx, p->sq_off.user_addr, size);
3740 return PTR_ERR(ptr);
3747 static int io_uring_install_fd(struct file *file)
3751 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3754 fd_install(fd, file);
3759 * Allocate an anonymous fd, this is what constitutes the application
3760 * visible backing of an io_uring instance. The application mmaps this
3761 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3762 * we have to tie this fd to a socket for file garbage collection purposes.
3764 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3767 #if defined(CONFIG_UNIX)
3770 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3773 return ERR_PTR(ret);
3776 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3777 O_RDWR | O_CLOEXEC, NULL);
3778 #if defined(CONFIG_UNIX)
3780 sock_release(ctx->ring_sock);
3781 ctx->ring_sock = NULL;
3783 ctx->ring_sock->file = file;
3789 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3790 struct io_uring_params __user *params)
3792 struct io_ring_ctx *ctx;
3793 struct io_uring_task *tctx;
3799 if (entries > IORING_MAX_ENTRIES) {
3800 if (!(p->flags & IORING_SETUP_CLAMP))
3802 entries = IORING_MAX_ENTRIES;
3805 if ((p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
3806 && !(p->flags & IORING_SETUP_NO_MMAP))
3810 * Use twice as many entries for the CQ ring. It's possible for the
3811 * application to drive a higher depth than the size of the SQ ring,
3812 * since the sqes are only used at submission time. This allows for
3813 * some flexibility in overcommitting a bit. If the application has
3814 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3815 * of CQ ring entries manually.
3817 p->sq_entries = roundup_pow_of_two(entries);
3818 if (p->flags & IORING_SETUP_CQSIZE) {
3820 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3821 * to a power-of-two, if it isn't already. We do NOT impose
3822 * any cq vs sq ring sizing.
3826 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3827 if (!(p->flags & IORING_SETUP_CLAMP))
3829 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3831 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3832 if (p->cq_entries < p->sq_entries)
3835 p->cq_entries = 2 * p->sq_entries;
3838 ctx = io_ring_ctx_alloc(p);
3842 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3843 !(ctx->flags & IORING_SETUP_IOPOLL) &&
3844 !(ctx->flags & IORING_SETUP_SQPOLL))
3845 ctx->task_complete = true;
3848 * lazy poll_wq activation relies on ->task_complete for synchronisation
3849 * purposes, see io_activate_pollwq()
3851 if (!ctx->task_complete)
3852 ctx->poll_activated = true;
3855 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3856 * space applications don't need to do io completion events
3857 * polling again, they can rely on io_sq_thread to do polling
3858 * work, which can reduce cpu usage and uring_lock contention.
3860 if (ctx->flags & IORING_SETUP_IOPOLL &&
3861 !(ctx->flags & IORING_SETUP_SQPOLL))
3862 ctx->syscall_iopoll = 1;
3864 ctx->compat = in_compat_syscall();
3865 if (!capable(CAP_IPC_LOCK))
3866 ctx->user = get_uid(current_user());
3869 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3870 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3873 if (ctx->flags & IORING_SETUP_SQPOLL) {
3874 /* IPI related flags don't make sense with SQPOLL */
3875 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3876 IORING_SETUP_TASKRUN_FLAG |
3877 IORING_SETUP_DEFER_TASKRUN))
3879 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3880 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3881 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3883 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3884 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3886 ctx->notify_method = TWA_SIGNAL;
3890 * For DEFER_TASKRUN we require the completion task to be the same as the
3891 * submission task. This implies that there is only one submitter, so enforce
3894 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3895 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3900 * This is just grabbed for accounting purposes. When a process exits,
3901 * the mm is exited and dropped before the files, hence we need to hang
3902 * on to this mm purely for the purposes of being able to unaccount
3903 * memory (locked/pinned vm). It's not used for anything else.
3905 mmgrab(current->mm);
3906 ctx->mm_account = current->mm;
3908 ret = io_allocate_scq_urings(ctx, p);
3912 ret = io_sq_offload_create(ctx, p);
3916 ret = io_rsrc_init(ctx);
3920 p->sq_off.head = offsetof(struct io_rings, sq.head);
3921 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3922 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3923 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3924 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3925 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3926 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3927 p->sq_off.resv1 = 0;
3928 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3929 p->sq_off.user_addr = 0;
3931 p->cq_off.head = offsetof(struct io_rings, cq.head);
3932 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3933 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3934 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3935 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3936 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3937 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3938 p->cq_off.resv1 = 0;
3939 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3940 p->cq_off.user_addr = 0;
3942 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3943 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3944 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3945 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3946 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3947 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3948 IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING;
3950 if (copy_to_user(params, p, sizeof(*p))) {
3955 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
3956 && !(ctx->flags & IORING_SETUP_R_DISABLED))
3957 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
3959 file = io_uring_get_file(ctx);
3961 ret = PTR_ERR(file);
3965 ret = __io_uring_add_tctx_node(ctx);
3968 tctx = current->io_uring;
3971 * Install ring fd as the very last thing, so we don't risk someone
3972 * having closed it before we finish setup
3974 if (p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
3975 ret = io_ring_add_registered_file(tctx, file, 0, IO_RINGFD_REG_MAX);
3977 ret = io_uring_install_fd(file);
3981 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3984 io_ring_ctx_wait_and_kill(ctx);
3992 * Sets up an aio uring context, and returns the fd. Applications asks for a
3993 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3994 * params structure passed in.
3996 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3998 struct io_uring_params p;
4001 if (copy_from_user(&p, params, sizeof(p)))
4003 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
4008 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
4009 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
4010 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
4011 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
4012 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
4013 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
4014 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN |
4015 IORING_SETUP_NO_MMAP | IORING_SETUP_REGISTERED_FD_ONLY))
4018 return io_uring_create(entries, &p, params);
4021 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
4022 struct io_uring_params __user *, params)
4024 return io_uring_setup(entries, params);
4027 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
4030 struct io_uring_probe *p;
4034 size = struct_size(p, ops, nr_args);
4035 if (size == SIZE_MAX)
4037 p = kzalloc(size, GFP_KERNEL);
4042 if (copy_from_user(p, arg, size))
4045 if (memchr_inv(p, 0, size))
4048 p->last_op = IORING_OP_LAST - 1;
4049 if (nr_args > IORING_OP_LAST)
4050 nr_args = IORING_OP_LAST;
4052 for (i = 0; i < nr_args; i++) {
4054 if (!io_issue_defs[i].not_supported)
4055 p->ops[i].flags = IO_URING_OP_SUPPORTED;
4060 if (copy_to_user(arg, p, size))
4067 static int io_register_personality(struct io_ring_ctx *ctx)
4069 const struct cred *creds;
4073 creds = get_current_cred();
4075 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
4076 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
4084 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
4085 void __user *arg, unsigned int nr_args)
4087 struct io_uring_restriction *res;
4091 /* Restrictions allowed only if rings started disabled */
4092 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
4095 /* We allow only a single restrictions registration */
4096 if (ctx->restrictions.registered)
4099 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
4102 size = array_size(nr_args, sizeof(*res));
4103 if (size == SIZE_MAX)
4106 res = memdup_user(arg, size);
4108 return PTR_ERR(res);
4112 for (i = 0; i < nr_args; i++) {
4113 switch (res[i].opcode) {
4114 case IORING_RESTRICTION_REGISTER_OP:
4115 if (res[i].register_op >= IORING_REGISTER_LAST) {
4120 __set_bit(res[i].register_op,
4121 ctx->restrictions.register_op);
4123 case IORING_RESTRICTION_SQE_OP:
4124 if (res[i].sqe_op >= IORING_OP_LAST) {
4129 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
4131 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
4132 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
4134 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
4135 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
4144 /* Reset all restrictions if an error happened */
4146 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
4148 ctx->restrictions.registered = true;
4154 static int io_register_enable_rings(struct io_ring_ctx *ctx)
4156 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
4159 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task) {
4160 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
4162 * Lazy activation attempts would fail if it was polled before
4163 * submitter_task is set.
4165 if (wq_has_sleeper(&ctx->poll_wq))
4166 io_activate_pollwq(ctx);
4169 if (ctx->restrictions.registered)
4170 ctx->restricted = 1;
4172 ctx->flags &= ~IORING_SETUP_R_DISABLED;
4173 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
4174 wake_up(&ctx->sq_data->wait);
4178 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
4179 void __user *arg, unsigned len)
4181 struct io_uring_task *tctx = current->io_uring;
4182 cpumask_var_t new_mask;
4185 if (!tctx || !tctx->io_wq)
4188 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4191 cpumask_clear(new_mask);
4192 if (len > cpumask_size())
4193 len = cpumask_size();
4195 if (in_compat_syscall()) {
4196 ret = compat_get_bitmap(cpumask_bits(new_mask),
4197 (const compat_ulong_t __user *)arg,
4198 len * 8 /* CHAR_BIT */);
4200 ret = copy_from_user(new_mask, arg, len);
4204 free_cpumask_var(new_mask);
4208 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
4209 free_cpumask_var(new_mask);
4213 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
4215 struct io_uring_task *tctx = current->io_uring;
4217 if (!tctx || !tctx->io_wq)
4220 return io_wq_cpu_affinity(tctx->io_wq, NULL);
4223 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
4225 __must_hold(&ctx->uring_lock)
4227 struct io_tctx_node *node;
4228 struct io_uring_task *tctx = NULL;
4229 struct io_sq_data *sqd = NULL;
4233 if (copy_from_user(new_count, arg, sizeof(new_count)))
4235 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4236 if (new_count[i] > INT_MAX)
4239 if (ctx->flags & IORING_SETUP_SQPOLL) {
4243 * Observe the correct sqd->lock -> ctx->uring_lock
4244 * ordering. Fine to drop uring_lock here, we hold
4247 refcount_inc(&sqd->refs);
4248 mutex_unlock(&ctx->uring_lock);
4249 mutex_lock(&sqd->lock);
4250 mutex_lock(&ctx->uring_lock);
4252 tctx = sqd->thread->io_uring;
4255 tctx = current->io_uring;
4258 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
4260 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4262 ctx->iowq_limits[i] = new_count[i];
4263 ctx->iowq_limits_set = true;
4265 if (tctx && tctx->io_wq) {
4266 ret = io_wq_max_workers(tctx->io_wq, new_count);
4270 memset(new_count, 0, sizeof(new_count));
4274 mutex_unlock(&sqd->lock);
4275 io_put_sq_data(sqd);
4278 if (copy_to_user(arg, new_count, sizeof(new_count)))
4281 /* that's it for SQPOLL, only the SQPOLL task creates requests */
4285 /* now propagate the restriction to all registered users */
4286 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
4287 struct io_uring_task *tctx = node->task->io_uring;
4289 if (WARN_ON_ONCE(!tctx->io_wq))
4292 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4293 new_count[i] = ctx->iowq_limits[i];
4294 /* ignore errors, it always returns zero anyway */
4295 (void)io_wq_max_workers(tctx->io_wq, new_count);
4300 mutex_unlock(&sqd->lock);
4301 io_put_sq_data(sqd);
4306 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
4307 void __user *arg, unsigned nr_args)
4308 __releases(ctx->uring_lock)
4309 __acquires(ctx->uring_lock)
4314 * We don't quiesce the refs for register anymore and so it can't be
4315 * dying as we're holding a file ref here.
4317 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
4320 if (ctx->submitter_task && ctx->submitter_task != current)
4323 if (ctx->restricted) {
4324 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
4325 if (!test_bit(opcode, ctx->restrictions.register_op))
4330 case IORING_REGISTER_BUFFERS:
4334 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
4336 case IORING_UNREGISTER_BUFFERS:
4340 ret = io_sqe_buffers_unregister(ctx);
4342 case IORING_REGISTER_FILES:
4346 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
4348 case IORING_UNREGISTER_FILES:
4352 ret = io_sqe_files_unregister(ctx);
4354 case IORING_REGISTER_FILES_UPDATE:
4355 ret = io_register_files_update(ctx, arg, nr_args);
4357 case IORING_REGISTER_EVENTFD:
4361 ret = io_eventfd_register(ctx, arg, 0);
4363 case IORING_REGISTER_EVENTFD_ASYNC:
4367 ret = io_eventfd_register(ctx, arg, 1);
4369 case IORING_UNREGISTER_EVENTFD:
4373 ret = io_eventfd_unregister(ctx);
4375 case IORING_REGISTER_PROBE:
4377 if (!arg || nr_args > 256)
4379 ret = io_probe(ctx, arg, nr_args);
4381 case IORING_REGISTER_PERSONALITY:
4385 ret = io_register_personality(ctx);
4387 case IORING_UNREGISTER_PERSONALITY:
4391 ret = io_unregister_personality(ctx, nr_args);
4393 case IORING_REGISTER_ENABLE_RINGS:
4397 ret = io_register_enable_rings(ctx);
4399 case IORING_REGISTER_RESTRICTIONS:
4400 ret = io_register_restrictions(ctx, arg, nr_args);
4402 case IORING_REGISTER_FILES2:
4403 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
4405 case IORING_REGISTER_FILES_UPDATE2:
4406 ret = io_register_rsrc_update(ctx, arg, nr_args,
4409 case IORING_REGISTER_BUFFERS2:
4410 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
4412 case IORING_REGISTER_BUFFERS_UPDATE:
4413 ret = io_register_rsrc_update(ctx, arg, nr_args,
4414 IORING_RSRC_BUFFER);
4416 case IORING_REGISTER_IOWQ_AFF:
4418 if (!arg || !nr_args)
4420 ret = io_register_iowq_aff(ctx, arg, nr_args);
4422 case IORING_UNREGISTER_IOWQ_AFF:
4426 ret = io_unregister_iowq_aff(ctx);
4428 case IORING_REGISTER_IOWQ_MAX_WORKERS:
4430 if (!arg || nr_args != 2)
4432 ret = io_register_iowq_max_workers(ctx, arg);
4434 case IORING_REGISTER_RING_FDS:
4435 ret = io_ringfd_register(ctx, arg, nr_args);
4437 case IORING_UNREGISTER_RING_FDS:
4438 ret = io_ringfd_unregister(ctx, arg, nr_args);
4440 case IORING_REGISTER_PBUF_RING:
4442 if (!arg || nr_args != 1)
4444 ret = io_register_pbuf_ring(ctx, arg);
4446 case IORING_UNREGISTER_PBUF_RING:
4448 if (!arg || nr_args != 1)
4450 ret = io_unregister_pbuf_ring(ctx, arg);
4452 case IORING_REGISTER_SYNC_CANCEL:
4454 if (!arg || nr_args != 1)
4456 ret = io_sync_cancel(ctx, arg);
4458 case IORING_REGISTER_FILE_ALLOC_RANGE:
4460 if (!arg || nr_args)
4462 ret = io_register_file_alloc_range(ctx, arg);
4472 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
4473 void __user *, arg, unsigned int, nr_args)
4475 struct io_ring_ctx *ctx;
4478 bool use_registered_ring;
4480 use_registered_ring = !!(opcode & IORING_REGISTER_USE_REGISTERED_RING);
4481 opcode &= ~IORING_REGISTER_USE_REGISTERED_RING;
4483 if (opcode >= IORING_REGISTER_LAST)
4486 if (use_registered_ring) {
4488 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
4489 * need only dereference our task private array to find it.
4491 struct io_uring_task *tctx = current->io_uring;
4493 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
4495 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
4496 f.file = tctx->registered_rings[fd];
4498 if (unlikely(!f.file))
4502 if (unlikely(!f.file))
4505 if (!io_is_uring_fops(f.file))
4509 ctx = f.file->private_data;
4511 mutex_lock(&ctx->uring_lock);
4512 ret = __io_uring_register(ctx, opcode, arg, nr_args);
4513 mutex_unlock(&ctx->uring_lock);
4514 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
4520 static int __init io_uring_init(void)
4522 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4523 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4524 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4527 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4528 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4529 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4530 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4531 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4532 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
4533 BUILD_BUG_SQE_ELEM(1, __u8, flags);
4534 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
4535 BUILD_BUG_SQE_ELEM(4, __s32, fd);
4536 BUILD_BUG_SQE_ELEM(8, __u64, off);
4537 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
4538 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
4539 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
4540 BUILD_BUG_SQE_ELEM(16, __u64, addr);
4541 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
4542 BUILD_BUG_SQE_ELEM(24, __u32, len);
4543 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
4544 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
4545 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4546 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
4547 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
4548 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
4549 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
4550 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
4551 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
4552 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
4553 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
4554 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
4555 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
4556 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
4557 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
4558 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
4559 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
4560 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
4561 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
4562 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
4563 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
4564 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
4565 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
4566 BUILD_BUG_SQE_ELEM(42, __u16, personality);
4567 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
4568 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
4569 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
4570 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
4571 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
4572 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4573 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
4575 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4576 sizeof(struct io_uring_rsrc_update));
4577 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4578 sizeof(struct io_uring_rsrc_update2));
4580 /* ->buf_index is u16 */
4581 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4582 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4583 offsetof(struct io_uring_buf_ring, tail));
4585 /* should fit into one byte */
4586 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
4587 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4588 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4590 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
4592 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4594 io_uring_optable_init();
4596 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4597 SLAB_ACCOUNT | SLAB_TYPESAFE_BY_RCU);
4600 __initcall(io_uring_init);