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);
151 struct kmem_cache *req_cachep;
153 struct sock *io_uring_get_socket(struct file *file)
155 #if defined(CONFIG_UNIX)
156 if (io_is_uring_fops(file)) {
157 struct io_ring_ctx *ctx = file->private_data;
159 return ctx->ring_sock->sk;
164 EXPORT_SYMBOL(io_uring_get_socket);
166 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
168 if (!wq_list_empty(&ctx->submit_state.compl_reqs) ||
169 ctx->submit_state.cqes_count)
170 __io_submit_flush_completions(ctx);
173 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
175 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
178 static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx)
180 return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head);
183 static bool io_match_linked(struct io_kiocb *head)
185 struct io_kiocb *req;
187 io_for_each_link(req, head) {
188 if (req->flags & REQ_F_INFLIGHT)
195 * As io_match_task() but protected against racing with linked timeouts.
196 * User must not hold timeout_lock.
198 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
203 if (task && head->task != task)
208 if (head->flags & REQ_F_LINK_TIMEOUT) {
209 struct io_ring_ctx *ctx = head->ctx;
211 /* protect against races with linked timeouts */
212 spin_lock_irq(&ctx->timeout_lock);
213 matched = io_match_linked(head);
214 spin_unlock_irq(&ctx->timeout_lock);
216 matched = io_match_linked(head);
221 static inline void req_fail_link_node(struct io_kiocb *req, int res)
224 io_req_set_res(req, res, 0);
227 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
229 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
232 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
234 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
236 complete(&ctx->ref_comp);
239 static __cold void io_fallback_req_func(struct work_struct *work)
241 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
243 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
244 struct io_kiocb *req, *tmp;
245 struct io_tw_state ts = { .locked = true, };
247 mutex_lock(&ctx->uring_lock);
248 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
249 req->io_task_work.func(req, &ts);
250 if (WARN_ON_ONCE(!ts.locked))
252 io_submit_flush_completions(ctx);
253 mutex_unlock(&ctx->uring_lock);
256 static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
258 unsigned hash_buckets = 1U << bits;
259 size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
261 table->hbs = kmalloc(hash_size, GFP_KERNEL);
265 table->hash_bits = bits;
266 init_hash_table(table, hash_buckets);
270 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
272 struct io_ring_ctx *ctx;
275 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
279 xa_init(&ctx->io_bl_xa);
282 * Use 5 bits less than the max cq entries, that should give us around
283 * 32 entries per hash list if totally full and uniformly spread, but
284 * don't keep too many buckets to not overconsume memory.
286 hash_bits = ilog2(p->cq_entries) - 5;
287 hash_bits = clamp(hash_bits, 1, 8);
288 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
290 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
292 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
296 ctx->flags = p->flags;
297 init_waitqueue_head(&ctx->sqo_sq_wait);
298 INIT_LIST_HEAD(&ctx->sqd_list);
299 INIT_LIST_HEAD(&ctx->cq_overflow_list);
300 INIT_LIST_HEAD(&ctx->io_buffers_cache);
301 io_alloc_cache_init(&ctx->rsrc_node_cache, IO_NODE_ALLOC_CACHE_MAX,
302 sizeof(struct io_rsrc_node));
303 io_alloc_cache_init(&ctx->apoll_cache, IO_ALLOC_CACHE_MAX,
304 sizeof(struct async_poll));
305 io_alloc_cache_init(&ctx->netmsg_cache, IO_ALLOC_CACHE_MAX,
306 sizeof(struct io_async_msghdr));
307 init_completion(&ctx->ref_comp);
308 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
309 mutex_init(&ctx->uring_lock);
310 init_waitqueue_head(&ctx->cq_wait);
311 init_waitqueue_head(&ctx->poll_wq);
312 init_waitqueue_head(&ctx->rsrc_quiesce_wq);
313 spin_lock_init(&ctx->completion_lock);
314 spin_lock_init(&ctx->timeout_lock);
315 INIT_WQ_LIST(&ctx->iopoll_list);
316 INIT_LIST_HEAD(&ctx->io_buffers_pages);
317 INIT_LIST_HEAD(&ctx->io_buffers_comp);
318 INIT_LIST_HEAD(&ctx->defer_list);
319 INIT_LIST_HEAD(&ctx->timeout_list);
320 INIT_LIST_HEAD(&ctx->ltimeout_list);
321 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
322 init_llist_head(&ctx->work_llist);
323 INIT_LIST_HEAD(&ctx->tctx_list);
324 ctx->submit_state.free_list.next = NULL;
325 INIT_WQ_LIST(&ctx->locked_free_list);
326 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
327 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
330 kfree(ctx->cancel_table.hbs);
331 kfree(ctx->cancel_table_locked.hbs);
333 xa_destroy(&ctx->io_bl_xa);
338 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
340 struct io_rings *r = ctx->rings;
342 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
346 static bool req_need_defer(struct io_kiocb *req, u32 seq)
348 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
349 struct io_ring_ctx *ctx = req->ctx;
351 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
357 static void io_clean_op(struct io_kiocb *req)
359 if (req->flags & REQ_F_BUFFER_SELECTED) {
360 spin_lock(&req->ctx->completion_lock);
361 io_put_kbuf_comp(req);
362 spin_unlock(&req->ctx->completion_lock);
365 if (req->flags & REQ_F_NEED_CLEANUP) {
366 const struct io_cold_def *def = &io_cold_defs[req->opcode];
371 if ((req->flags & REQ_F_POLLED) && req->apoll) {
372 kfree(req->apoll->double_poll);
376 if (req->flags & REQ_F_INFLIGHT) {
377 struct io_uring_task *tctx = req->task->io_uring;
379 atomic_dec(&tctx->inflight_tracked);
381 if (req->flags & REQ_F_CREDS)
382 put_cred(req->creds);
383 if (req->flags & REQ_F_ASYNC_DATA) {
384 kfree(req->async_data);
385 req->async_data = NULL;
387 req->flags &= ~IO_REQ_CLEAN_FLAGS;
390 static inline void io_req_track_inflight(struct io_kiocb *req)
392 if (!(req->flags & REQ_F_INFLIGHT)) {
393 req->flags |= REQ_F_INFLIGHT;
394 atomic_inc(&req->task->io_uring->inflight_tracked);
398 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
400 if (WARN_ON_ONCE(!req->link))
403 req->flags &= ~REQ_F_ARM_LTIMEOUT;
404 req->flags |= REQ_F_LINK_TIMEOUT;
406 /* linked timeouts should have two refs once prep'ed */
407 io_req_set_refcount(req);
408 __io_req_set_refcount(req->link, 2);
412 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
414 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
416 return __io_prep_linked_timeout(req);
419 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
421 io_queue_linked_timeout(__io_prep_linked_timeout(req));
424 static inline void io_arm_ltimeout(struct io_kiocb *req)
426 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
427 __io_arm_ltimeout(req);
430 static void io_prep_async_work(struct io_kiocb *req)
432 const struct io_issue_def *def = &io_issue_defs[req->opcode];
433 struct io_ring_ctx *ctx = req->ctx;
435 if (!(req->flags & REQ_F_CREDS)) {
436 req->flags |= REQ_F_CREDS;
437 req->creds = get_current_cred();
440 req->work.list.next = NULL;
442 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
443 if (req->flags & REQ_F_FORCE_ASYNC)
444 req->work.flags |= IO_WQ_WORK_CONCURRENT;
446 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
447 req->flags |= io_file_get_flags(req->file);
449 if (req->file && (req->flags & REQ_F_ISREG)) {
450 bool should_hash = def->hash_reg_file;
452 /* don't serialize this request if the fs doesn't need it */
453 if (should_hash && (req->file->f_flags & O_DIRECT) &&
454 (req->file->f_mode & FMODE_DIO_PARALLEL_WRITE))
456 if (should_hash || (ctx->flags & IORING_SETUP_IOPOLL))
457 io_wq_hash_work(&req->work, file_inode(req->file));
458 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
459 if (def->unbound_nonreg_file)
460 req->work.flags |= IO_WQ_WORK_UNBOUND;
464 static void io_prep_async_link(struct io_kiocb *req)
466 struct io_kiocb *cur;
468 if (req->flags & REQ_F_LINK_TIMEOUT) {
469 struct io_ring_ctx *ctx = req->ctx;
471 spin_lock_irq(&ctx->timeout_lock);
472 io_for_each_link(cur, req)
473 io_prep_async_work(cur);
474 spin_unlock_irq(&ctx->timeout_lock);
476 io_for_each_link(cur, req)
477 io_prep_async_work(cur);
481 void io_queue_iowq(struct io_kiocb *req, struct io_tw_state *ts_dont_use)
483 struct io_kiocb *link = io_prep_linked_timeout(req);
484 struct io_uring_task *tctx = req->task->io_uring;
487 BUG_ON(!tctx->io_wq);
489 /* init ->work of the whole link before punting */
490 io_prep_async_link(req);
493 * Not expected to happen, but if we do have a bug where this _can_
494 * happen, catch it here and ensure the request is marked as
495 * canceled. That will make io-wq go through the usual work cancel
496 * procedure rather than attempt to run this request (or create a new
499 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
500 req->work.flags |= IO_WQ_WORK_CANCEL;
502 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
503 io_wq_enqueue(tctx->io_wq, &req->work);
505 io_queue_linked_timeout(link);
508 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
510 while (!list_empty(&ctx->defer_list)) {
511 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
512 struct io_defer_entry, list);
514 if (req_need_defer(de->req, de->seq))
516 list_del_init(&de->list);
517 io_req_task_queue(de->req);
523 static void io_eventfd_ops(struct rcu_head *rcu)
525 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
526 int ops = atomic_xchg(&ev_fd->ops, 0);
528 if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT))
529 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
531 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
532 * ordering in a race but if references are 0 we know we have to free
535 if (atomic_dec_and_test(&ev_fd->refs)) {
536 eventfd_ctx_put(ev_fd->cq_ev_fd);
541 static void io_eventfd_signal(struct io_ring_ctx *ctx)
543 struct io_ev_fd *ev_fd = NULL;
547 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
550 ev_fd = rcu_dereference(ctx->io_ev_fd);
553 * Check again if ev_fd exists incase an io_eventfd_unregister call
554 * completed between the NULL check of ctx->io_ev_fd at the start of
555 * the function and rcu_read_lock.
557 if (unlikely(!ev_fd))
559 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
561 if (ev_fd->eventfd_async && !io_wq_current_is_worker())
564 if (likely(eventfd_signal_allowed())) {
565 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
567 atomic_inc(&ev_fd->refs);
568 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops))
569 call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops);
571 atomic_dec(&ev_fd->refs);
578 static void io_eventfd_flush_signal(struct io_ring_ctx *ctx)
582 spin_lock(&ctx->completion_lock);
585 * Eventfd should only get triggered when at least one event has been
586 * posted. Some applications rely on the eventfd notification count
587 * only changing IFF a new CQE has been added to the CQ ring. There's
588 * no depedency on 1:1 relationship between how many times this
589 * function is called (and hence the eventfd count) and number of CQEs
590 * posted to the CQ ring.
592 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
593 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
594 spin_unlock(&ctx->completion_lock);
598 io_eventfd_signal(ctx);
601 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
603 if (ctx->poll_activated)
604 io_poll_wq_wake(ctx);
605 if (ctx->off_timeout_used)
606 io_flush_timeouts(ctx);
607 if (ctx->drain_active) {
608 spin_lock(&ctx->completion_lock);
609 io_queue_deferred(ctx);
610 spin_unlock(&ctx->completion_lock);
613 io_eventfd_flush_signal(ctx);
616 static inline void __io_cq_lock(struct io_ring_ctx *ctx)
618 if (!ctx->lockless_cq)
619 spin_lock(&ctx->completion_lock);
622 static inline void io_cq_lock(struct io_ring_ctx *ctx)
623 __acquires(ctx->completion_lock)
625 spin_lock(&ctx->completion_lock);
628 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
630 io_commit_cqring(ctx);
631 if (!ctx->task_complete) {
632 if (!ctx->lockless_cq)
633 spin_unlock(&ctx->completion_lock);
634 /* IOPOLL rings only need to wake up if it's also SQPOLL */
635 if (!ctx->syscall_iopoll)
638 io_commit_cqring_flush(ctx);
641 static void io_cq_unlock_post(struct io_ring_ctx *ctx)
642 __releases(ctx->completion_lock)
644 io_commit_cqring(ctx);
645 spin_unlock(&ctx->completion_lock);
647 io_commit_cqring_flush(ctx);
650 /* Returns true if there are no backlogged entries after the flush */
651 static void io_cqring_overflow_kill(struct io_ring_ctx *ctx)
653 struct io_overflow_cqe *ocqe;
656 spin_lock(&ctx->completion_lock);
657 list_splice_init(&ctx->cq_overflow_list, &list);
658 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
659 spin_unlock(&ctx->completion_lock);
661 while (!list_empty(&list)) {
662 ocqe = list_first_entry(&list, struct io_overflow_cqe, list);
663 list_del(&ocqe->list);
668 static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx)
670 size_t cqe_size = sizeof(struct io_uring_cqe);
672 if (__io_cqring_events(ctx) == ctx->cq_entries)
675 if (ctx->flags & IORING_SETUP_CQE32)
679 while (!list_empty(&ctx->cq_overflow_list)) {
680 struct io_uring_cqe *cqe;
681 struct io_overflow_cqe *ocqe;
683 if (!io_get_cqe_overflow(ctx, &cqe, true))
685 ocqe = list_first_entry(&ctx->cq_overflow_list,
686 struct io_overflow_cqe, list);
687 memcpy(cqe, &ocqe->cqe, cqe_size);
688 list_del(&ocqe->list);
692 if (list_empty(&ctx->cq_overflow_list)) {
693 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
694 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
696 io_cq_unlock_post(ctx);
699 static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx)
701 /* iopoll syncs against uring_lock, not completion_lock */
702 if (ctx->flags & IORING_SETUP_IOPOLL)
703 mutex_lock(&ctx->uring_lock);
704 __io_cqring_overflow_flush(ctx);
705 if (ctx->flags & IORING_SETUP_IOPOLL)
706 mutex_unlock(&ctx->uring_lock);
709 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx)
711 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
712 io_cqring_do_overflow_flush(ctx);
715 /* can be called by any task */
716 static void io_put_task_remote(struct task_struct *task)
718 struct io_uring_task *tctx = task->io_uring;
720 percpu_counter_sub(&tctx->inflight, 1);
721 if (unlikely(atomic_read(&tctx->in_cancel)))
722 wake_up(&tctx->wait);
723 put_task_struct(task);
726 /* used by a task to put its own references */
727 static void io_put_task_local(struct task_struct *task)
729 task->io_uring->cached_refs++;
732 /* must to be called somewhat shortly after putting a request */
733 static inline void io_put_task(struct task_struct *task)
735 if (likely(task == current))
736 io_put_task_local(task);
738 io_put_task_remote(task);
741 void io_task_refs_refill(struct io_uring_task *tctx)
743 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
745 percpu_counter_add(&tctx->inflight, refill);
746 refcount_add(refill, ¤t->usage);
747 tctx->cached_refs += refill;
750 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
752 struct io_uring_task *tctx = task->io_uring;
753 unsigned int refs = tctx->cached_refs;
756 tctx->cached_refs = 0;
757 percpu_counter_sub(&tctx->inflight, refs);
758 put_task_struct_many(task, refs);
762 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
763 s32 res, u32 cflags, u64 extra1, u64 extra2)
765 struct io_overflow_cqe *ocqe;
766 size_t ocq_size = sizeof(struct io_overflow_cqe);
767 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
769 lockdep_assert_held(&ctx->completion_lock);
772 ocq_size += sizeof(struct io_uring_cqe);
774 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
775 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
778 * If we're in ring overflow flush mode, or in task cancel mode,
779 * or cannot allocate an overflow entry, then we need to drop it
782 io_account_cq_overflow(ctx);
783 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
786 if (list_empty(&ctx->cq_overflow_list)) {
787 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
788 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
791 ocqe->cqe.user_data = user_data;
793 ocqe->cqe.flags = cflags;
795 ocqe->cqe.big_cqe[0] = extra1;
796 ocqe->cqe.big_cqe[1] = extra2;
798 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
802 void io_req_cqe_overflow(struct io_kiocb *req)
804 io_cqring_event_overflow(req->ctx, req->cqe.user_data,
805 req->cqe.res, req->cqe.flags,
806 req->big_cqe.extra1, req->big_cqe.extra2);
807 memset(&req->big_cqe, 0, sizeof(req->big_cqe));
811 * writes to the cq entry need to come after reading head; the
812 * control dependency is enough as we're using WRITE_ONCE to
815 bool io_cqe_cache_refill(struct io_ring_ctx *ctx, bool overflow)
817 struct io_rings *rings = ctx->rings;
818 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
819 unsigned int free, queued, len;
822 * Posting into the CQ when there are pending overflowed CQEs may break
823 * ordering guarantees, which will affect links, F_MORE users and more.
824 * Force overflow the completion.
826 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
829 /* userspace may cheat modifying the tail, be safe and do min */
830 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
831 free = ctx->cq_entries - queued;
832 /* we need a contiguous range, limit based on the current array offset */
833 len = min(free, ctx->cq_entries - off);
837 if (ctx->flags & IORING_SETUP_CQE32) {
842 ctx->cqe_cached = &rings->cqes[off];
843 ctx->cqe_sentinel = ctx->cqe_cached + len;
847 static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
850 struct io_uring_cqe *cqe;
855 * If we can't get a cq entry, userspace overflowed the
856 * submission (by quite a lot). Increment the overflow count in
859 if (likely(io_get_cqe(ctx, &cqe))) {
860 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
862 WRITE_ONCE(cqe->user_data, user_data);
863 WRITE_ONCE(cqe->res, res);
864 WRITE_ONCE(cqe->flags, cflags);
866 if (ctx->flags & IORING_SETUP_CQE32) {
867 WRITE_ONCE(cqe->big_cqe[0], 0);
868 WRITE_ONCE(cqe->big_cqe[1], 0);
875 static void __io_flush_post_cqes(struct io_ring_ctx *ctx)
876 __must_hold(&ctx->uring_lock)
878 struct io_submit_state *state = &ctx->submit_state;
881 lockdep_assert_held(&ctx->uring_lock);
882 for (i = 0; i < state->cqes_count; i++) {
883 struct io_uring_cqe *cqe = &state->cqes[i];
885 if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) {
886 if (ctx->task_complete) {
887 spin_lock(&ctx->completion_lock);
888 io_cqring_event_overflow(ctx, cqe->user_data,
889 cqe->res, cqe->flags, 0, 0);
890 spin_unlock(&ctx->completion_lock);
892 io_cqring_event_overflow(ctx, cqe->user_data,
893 cqe->res, cqe->flags, 0, 0);
897 state->cqes_count = 0;
900 static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
906 filled = io_fill_cqe_aux(ctx, user_data, res, cflags);
907 if (!filled && allow_overflow)
908 filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
910 io_cq_unlock_post(ctx);
914 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags)
916 return __io_post_aux_cqe(ctx, user_data, res, cflags, true);
920 * A helper for multishot requests posting additional CQEs.
921 * Should only be used from a task_work including IO_URING_F_MULTISHOT.
923 bool io_fill_cqe_req_aux(struct io_kiocb *req, bool defer, s32 res, u32 cflags)
925 struct io_ring_ctx *ctx = req->ctx;
926 u64 user_data = req->cqe.user_data;
927 struct io_uring_cqe *cqe;
930 return __io_post_aux_cqe(ctx, user_data, res, cflags, false);
932 lockdep_assert_held(&ctx->uring_lock);
934 if (ctx->submit_state.cqes_count == ARRAY_SIZE(ctx->submit_state.cqes)) {
936 __io_flush_post_cqes(ctx);
937 /* no need to flush - flush is deferred */
938 __io_cq_unlock_post(ctx);
941 /* For defered completions this is not as strict as it is otherwise,
942 * however it's main job is to prevent unbounded posted completions,
943 * and in that it works just as well.
945 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
948 cqe = &ctx->submit_state.cqes[ctx->submit_state.cqes_count++];
949 cqe->user_data = user_data;
955 static void __io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
957 struct io_ring_ctx *ctx = req->ctx;
958 struct io_rsrc_node *rsrc_node = NULL;
961 if (!(req->flags & REQ_F_CQE_SKIP)) {
962 if (!io_fill_cqe_req(ctx, req))
963 io_req_cqe_overflow(req);
967 * If we're the last reference to this request, add to our locked
970 if (req_ref_put_and_test(req)) {
971 if (req->flags & IO_REQ_LINK_FLAGS) {
972 if (req->flags & IO_DISARM_MASK)
975 io_req_task_queue(req->link);
979 io_put_kbuf_comp(req);
980 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
984 rsrc_node = req->rsrc_node;
986 * Selected buffer deallocation in io_clean_op() assumes that
987 * we don't hold ->completion_lock. Clean them here to avoid
990 io_put_task_remote(req->task);
991 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
992 ctx->locked_free_nr++;
994 io_cq_unlock_post(ctx);
997 io_ring_submit_lock(ctx, issue_flags);
998 io_put_rsrc_node(ctx, rsrc_node);
999 io_ring_submit_unlock(ctx, issue_flags);
1003 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
1005 if (req->ctx->task_complete && req->ctx->submitter_task != current) {
1006 req->io_task_work.func = io_req_task_complete;
1007 io_req_task_work_add(req);
1008 } else if (!(issue_flags & IO_URING_F_UNLOCKED) ||
1009 !(req->ctx->flags & IORING_SETUP_IOPOLL)) {
1010 __io_req_complete_post(req, issue_flags);
1012 struct io_ring_ctx *ctx = req->ctx;
1014 mutex_lock(&ctx->uring_lock);
1015 __io_req_complete_post(req, issue_flags & ~IO_URING_F_UNLOCKED);
1016 mutex_unlock(&ctx->uring_lock);
1020 void io_req_defer_failed(struct io_kiocb *req, s32 res)
1021 __must_hold(&ctx->uring_lock)
1023 const struct io_cold_def *def = &io_cold_defs[req->opcode];
1025 lockdep_assert_held(&req->ctx->uring_lock);
1028 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
1031 io_req_complete_defer(req);
1035 * Don't initialise the fields below on every allocation, but do that in
1036 * advance and keep them valid across allocations.
1038 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1042 req->async_data = NULL;
1043 /* not necessary, but safer to zero */
1044 memset(&req->cqe, 0, sizeof(req->cqe));
1045 memset(&req->big_cqe, 0, sizeof(req->big_cqe));
1048 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1049 struct io_submit_state *state)
1051 spin_lock(&ctx->completion_lock);
1052 wq_list_splice(&ctx->locked_free_list, &state->free_list);
1053 ctx->locked_free_nr = 0;
1054 spin_unlock(&ctx->completion_lock);
1058 * A request might get retired back into the request caches even before opcode
1059 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1060 * Because of that, io_alloc_req() should be called only under ->uring_lock
1061 * and with extra caution to not get a request that is still worked on.
1063 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
1064 __must_hold(&ctx->uring_lock)
1066 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1067 void *reqs[IO_REQ_ALLOC_BATCH];
1071 * If we have more than a batch's worth of requests in our IRQ side
1072 * locked cache, grab the lock and move them over to our submission
1075 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
1076 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
1077 if (!io_req_cache_empty(ctx))
1081 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
1084 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1085 * retry single alloc to be on the safe side.
1087 if (unlikely(ret <= 0)) {
1088 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1094 percpu_ref_get_many(&ctx->refs, ret);
1095 for (i = 0; i < ret; i++) {
1096 struct io_kiocb *req = reqs[i];
1098 io_preinit_req(req, ctx);
1099 io_req_add_to_cache(req, ctx);
1104 __cold void io_free_req(struct io_kiocb *req)
1106 /* refs were already put, restore them for io_req_task_complete() */
1107 req->flags &= ~REQ_F_REFCOUNT;
1108 /* we only want to free it, don't post CQEs */
1109 req->flags |= REQ_F_CQE_SKIP;
1110 req->io_task_work.func = io_req_task_complete;
1111 io_req_task_work_add(req);
1114 static void __io_req_find_next_prep(struct io_kiocb *req)
1116 struct io_ring_ctx *ctx = req->ctx;
1118 spin_lock(&ctx->completion_lock);
1119 io_disarm_next(req);
1120 spin_unlock(&ctx->completion_lock);
1123 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1125 struct io_kiocb *nxt;
1128 * If LINK is set, we have dependent requests in this chain. If we
1129 * didn't fail this request, queue the first one up, moving any other
1130 * dependencies to the next request. In case of failure, fail the rest
1133 if (unlikely(req->flags & IO_DISARM_MASK))
1134 __io_req_find_next_prep(req);
1140 static void ctx_flush_and_put(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1144 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1145 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1147 io_submit_flush_completions(ctx);
1148 mutex_unlock(&ctx->uring_lock);
1151 percpu_ref_put(&ctx->refs);
1154 static unsigned int handle_tw_list(struct llist_node *node,
1155 struct io_ring_ctx **ctx,
1156 struct io_tw_state *ts,
1157 struct llist_node *last)
1159 unsigned int count = 0;
1161 while (node && node != last) {
1162 struct llist_node *next = node->next;
1163 struct io_kiocb *req = container_of(node, struct io_kiocb,
1166 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1168 if (req->ctx != *ctx) {
1169 ctx_flush_and_put(*ctx, ts);
1171 /* if not contended, grab and improve batching */
1172 ts->locked = mutex_trylock(&(*ctx)->uring_lock);
1173 percpu_ref_get(&(*ctx)->refs);
1175 INDIRECT_CALL_2(req->io_task_work.func,
1176 io_poll_task_func, io_req_rw_complete,
1180 if (unlikely(need_resched())) {
1181 ctx_flush_and_put(*ctx, ts);
1191 * io_llist_xchg - swap all entries in a lock-less list
1192 * @head: the head of lock-less list to delete all entries
1193 * @new: new entry as the head of the list
1195 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1196 * The order of entries returned is from the newest to the oldest added one.
1198 static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1199 struct llist_node *new)
1201 return xchg(&head->first, new);
1205 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1206 * @head: the head of lock-less list to delete all entries
1207 * @old: expected old value of the first entry of the list
1208 * @new: new entry as the head of the list
1210 * perform a cmpxchg on the first entry of the list.
1213 static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1214 struct llist_node *old,
1215 struct llist_node *new)
1217 return cmpxchg(&head->first, old, new);
1220 static __cold void io_fallback_tw(struct io_uring_task *tctx, bool sync)
1222 struct llist_node *node = llist_del_all(&tctx->task_list);
1223 struct io_ring_ctx *last_ctx = NULL;
1224 struct io_kiocb *req;
1227 req = container_of(node, struct io_kiocb, io_task_work.node);
1229 if (sync && last_ctx != req->ctx) {
1231 flush_delayed_work(&last_ctx->fallback_work);
1232 percpu_ref_put(&last_ctx->refs);
1234 last_ctx = req->ctx;
1235 percpu_ref_get(&last_ctx->refs);
1237 if (llist_add(&req->io_task_work.node,
1238 &req->ctx->fallback_llist))
1239 schedule_delayed_work(&req->ctx->fallback_work, 1);
1243 flush_delayed_work(&last_ctx->fallback_work);
1244 percpu_ref_put(&last_ctx->refs);
1248 void tctx_task_work(struct callback_head *cb)
1250 struct io_tw_state ts = {};
1251 struct io_ring_ctx *ctx = NULL;
1252 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1254 struct llist_node fake = {};
1255 struct llist_node *node;
1256 unsigned int loops = 0;
1257 unsigned int count = 0;
1259 if (unlikely(current->flags & PF_EXITING)) {
1260 io_fallback_tw(tctx, true);
1266 node = io_llist_xchg(&tctx->task_list, &fake);
1267 count += handle_tw_list(node, &ctx, &ts, &fake);
1269 /* skip expensive cmpxchg if there are items in the list */
1270 if (READ_ONCE(tctx->task_list.first) != &fake)
1272 if (ts.locked && !wq_list_empty(&ctx->submit_state.compl_reqs)) {
1273 io_submit_flush_completions(ctx);
1274 if (READ_ONCE(tctx->task_list.first) != &fake)
1277 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1278 } while (node != &fake);
1280 ctx_flush_and_put(ctx, &ts);
1282 /* relaxed read is enough as only the task itself sets ->in_cancel */
1283 if (unlikely(atomic_read(&tctx->in_cancel)))
1284 io_uring_drop_tctx_refs(current);
1286 trace_io_uring_task_work_run(tctx, count, loops);
1289 static inline void io_req_local_work_add(struct io_kiocb *req, unsigned flags)
1291 struct io_ring_ctx *ctx = req->ctx;
1292 unsigned nr_wait, nr_tw, nr_tw_prev;
1293 struct llist_node *first;
1295 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK))
1296 flags &= ~IOU_F_TWQ_LAZY_WAKE;
1298 first = READ_ONCE(ctx->work_llist.first);
1302 struct io_kiocb *first_req = container_of(first,
1306 * Might be executed at any moment, rely on
1307 * SLAB_TYPESAFE_BY_RCU to keep it alive.
1309 nr_tw_prev = READ_ONCE(first_req->nr_tw);
1311 nr_tw = nr_tw_prev + 1;
1312 /* Large enough to fail the nr_wait comparison below */
1313 if (!(flags & IOU_F_TWQ_LAZY_WAKE))
1317 req->io_task_work.node.next = first;
1318 } while (!try_cmpxchg(&ctx->work_llist.first, &first,
1319 &req->io_task_work.node));
1322 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1323 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1325 io_eventfd_signal(ctx);
1328 nr_wait = atomic_read(&ctx->cq_wait_nr);
1329 /* no one is waiting */
1332 /* either not enough or the previous add has already woken it up */
1333 if (nr_wait > nr_tw || nr_tw_prev >= nr_wait)
1335 /* pairs with set_current_state() in io_cqring_wait() */
1336 smp_mb__after_atomic();
1337 wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE);
1340 static void io_req_normal_work_add(struct io_kiocb *req)
1342 struct io_uring_task *tctx = req->task->io_uring;
1343 struct io_ring_ctx *ctx = req->ctx;
1345 /* task_work already pending, we're done */
1346 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1349 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1350 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1352 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1355 io_fallback_tw(tctx, false);
1358 void __io_req_task_work_add(struct io_kiocb *req, unsigned flags)
1360 if (req->ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
1362 io_req_local_work_add(req, flags);
1365 io_req_normal_work_add(req);
1369 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1371 struct llist_node *node;
1373 node = llist_del_all(&ctx->work_llist);
1375 struct io_kiocb *req = container_of(node, struct io_kiocb,
1379 io_req_normal_work_add(req);
1383 static int __io_run_local_work(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1385 struct llist_node *node;
1386 unsigned int loops = 0;
1389 if (WARN_ON_ONCE(ctx->submitter_task != current))
1391 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1392 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1395 * llists are in reverse order, flip it back the right way before
1396 * running the pending items.
1398 node = llist_reverse_order(io_llist_xchg(&ctx->work_llist, NULL));
1400 struct llist_node *next = node->next;
1401 struct io_kiocb *req = container_of(node, struct io_kiocb,
1403 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1404 INDIRECT_CALL_2(req->io_task_work.func,
1405 io_poll_task_func, io_req_rw_complete,
1412 if (!llist_empty(&ctx->work_llist))
1415 io_submit_flush_completions(ctx);
1416 if (!llist_empty(&ctx->work_llist))
1419 trace_io_uring_local_work_run(ctx, ret, loops);
1423 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx)
1425 struct io_tw_state ts = { .locked = true, };
1428 if (llist_empty(&ctx->work_llist))
1431 ret = __io_run_local_work(ctx, &ts);
1432 /* shouldn't happen! */
1433 if (WARN_ON_ONCE(!ts.locked))
1434 mutex_lock(&ctx->uring_lock);
1438 static int io_run_local_work(struct io_ring_ctx *ctx)
1440 struct io_tw_state ts = {};
1443 ts.locked = mutex_trylock(&ctx->uring_lock);
1444 ret = __io_run_local_work(ctx, &ts);
1446 mutex_unlock(&ctx->uring_lock);
1451 static void io_req_task_cancel(struct io_kiocb *req, struct io_tw_state *ts)
1453 io_tw_lock(req->ctx, ts);
1454 io_req_defer_failed(req, req->cqe.res);
1457 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts)
1459 io_tw_lock(req->ctx, ts);
1460 /* req->task == current here, checking PF_EXITING is safe */
1461 if (unlikely(req->task->flags & PF_EXITING))
1462 io_req_defer_failed(req, -EFAULT);
1463 else if (req->flags & REQ_F_FORCE_ASYNC)
1464 io_queue_iowq(req, ts);
1469 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1471 io_req_set_res(req, ret, 0);
1472 req->io_task_work.func = io_req_task_cancel;
1473 io_req_task_work_add(req);
1476 void io_req_task_queue(struct io_kiocb *req)
1478 req->io_task_work.func = io_req_task_submit;
1479 io_req_task_work_add(req);
1482 void io_queue_next(struct io_kiocb *req)
1484 struct io_kiocb *nxt = io_req_find_next(req);
1487 io_req_task_queue(nxt);
1490 static void io_free_batch_list(struct io_ring_ctx *ctx,
1491 struct io_wq_work_node *node)
1492 __must_hold(&ctx->uring_lock)
1495 struct io_kiocb *req = container_of(node, struct io_kiocb,
1498 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1499 if (req->flags & REQ_F_REFCOUNT) {
1500 node = req->comp_list.next;
1501 if (!req_ref_put_and_test(req))
1504 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1505 struct async_poll *apoll = req->apoll;
1507 if (apoll->double_poll)
1508 kfree(apoll->double_poll);
1509 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1511 req->flags &= ~REQ_F_POLLED;
1513 if (req->flags & IO_REQ_LINK_FLAGS)
1515 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1520 io_req_put_rsrc_locked(req, ctx);
1522 io_put_task(req->task);
1523 node = req->comp_list.next;
1524 io_req_add_to_cache(req, ctx);
1528 void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1529 __must_hold(&ctx->uring_lock)
1531 struct io_submit_state *state = &ctx->submit_state;
1532 struct io_wq_work_node *node;
1535 /* must come first to preserve CQE ordering in failure cases */
1536 if (state->cqes_count)
1537 __io_flush_post_cqes(ctx);
1538 __wq_list_for_each(node, &state->compl_reqs) {
1539 struct io_kiocb *req = container_of(node, struct io_kiocb,
1542 if (!(req->flags & REQ_F_CQE_SKIP) &&
1543 unlikely(!io_fill_cqe_req(ctx, req))) {
1544 if (ctx->task_complete) {
1545 spin_lock(&ctx->completion_lock);
1546 io_req_cqe_overflow(req);
1547 spin_unlock(&ctx->completion_lock);
1549 io_req_cqe_overflow(req);
1553 __io_cq_unlock_post(ctx);
1555 if (!wq_list_empty(&ctx->submit_state.compl_reqs)) {
1556 io_free_batch_list(ctx, state->compl_reqs.first);
1557 INIT_WQ_LIST(&state->compl_reqs);
1561 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1563 /* See comment at the top of this file */
1565 return __io_cqring_events(ctx);
1569 * We can't just wait for polled events to come to us, we have to actively
1570 * find and complete them.
1572 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1574 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1577 mutex_lock(&ctx->uring_lock);
1578 while (!wq_list_empty(&ctx->iopoll_list)) {
1579 /* let it sleep and repeat later if can't complete a request */
1580 if (io_do_iopoll(ctx, true) == 0)
1583 * Ensure we allow local-to-the-cpu processing to take place,
1584 * in this case we need to ensure that we reap all events.
1585 * Also let task_work, etc. to progress by releasing the mutex
1587 if (need_resched()) {
1588 mutex_unlock(&ctx->uring_lock);
1590 mutex_lock(&ctx->uring_lock);
1593 mutex_unlock(&ctx->uring_lock);
1596 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1598 unsigned int nr_events = 0;
1599 unsigned long check_cq;
1601 if (!io_allowed_run_tw(ctx))
1604 check_cq = READ_ONCE(ctx->check_cq);
1605 if (unlikely(check_cq)) {
1606 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1607 __io_cqring_overflow_flush(ctx);
1609 * Similarly do not spin if we have not informed the user of any
1612 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1616 * Don't enter poll loop if we already have events pending.
1617 * If we do, we can potentially be spinning for commands that
1618 * already triggered a CQE (eg in error).
1620 if (io_cqring_events(ctx))
1627 * If a submit got punted to a workqueue, we can have the
1628 * application entering polling for a command before it gets
1629 * issued. That app will hold the uring_lock for the duration
1630 * of the poll right here, so we need to take a breather every
1631 * now and then to ensure that the issue has a chance to add
1632 * the poll to the issued list. Otherwise we can spin here
1633 * forever, while the workqueue is stuck trying to acquire the
1636 if (wq_list_empty(&ctx->iopoll_list) ||
1637 io_task_work_pending(ctx)) {
1638 u32 tail = ctx->cached_cq_tail;
1640 (void) io_run_local_work_locked(ctx);
1642 if (task_work_pending(current) ||
1643 wq_list_empty(&ctx->iopoll_list)) {
1644 mutex_unlock(&ctx->uring_lock);
1646 mutex_lock(&ctx->uring_lock);
1648 /* some requests don't go through iopoll_list */
1649 if (tail != ctx->cached_cq_tail ||
1650 wq_list_empty(&ctx->iopoll_list))
1653 ret = io_do_iopoll(ctx, !min);
1654 if (unlikely(ret < 0))
1657 if (task_sigpending(current))
1663 } while (nr_events < min);
1668 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts)
1671 io_req_complete_defer(req);
1673 io_req_complete_post(req, IO_URING_F_UNLOCKED);
1677 * After the iocb has been issued, it's safe to be found on the poll list.
1678 * Adding the kiocb to the list AFTER submission ensures that we don't
1679 * find it from a io_do_iopoll() thread before the issuer is done
1680 * accessing the kiocb cookie.
1682 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1684 struct io_ring_ctx *ctx = req->ctx;
1685 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1687 /* workqueue context doesn't hold uring_lock, grab it now */
1688 if (unlikely(needs_lock))
1689 mutex_lock(&ctx->uring_lock);
1692 * Track whether we have multiple files in our lists. This will impact
1693 * how we do polling eventually, not spinning if we're on potentially
1694 * different devices.
1696 if (wq_list_empty(&ctx->iopoll_list)) {
1697 ctx->poll_multi_queue = false;
1698 } else if (!ctx->poll_multi_queue) {
1699 struct io_kiocb *list_req;
1701 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1703 if (list_req->file != req->file)
1704 ctx->poll_multi_queue = true;
1708 * For fast devices, IO may have already completed. If it has, add
1709 * it to the front so we find it first.
1711 if (READ_ONCE(req->iopoll_completed))
1712 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1714 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1716 if (unlikely(needs_lock)) {
1718 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1719 * in sq thread task context or in io worker task context. If
1720 * current task context is sq thread, we don't need to check
1721 * whether should wake up sq thread.
1723 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1724 wq_has_sleeper(&ctx->sq_data->wait))
1725 wake_up(&ctx->sq_data->wait);
1727 mutex_unlock(&ctx->uring_lock);
1731 unsigned int io_file_get_flags(struct file *file)
1733 unsigned int res = 0;
1735 if (S_ISREG(file_inode(file)->i_mode))
1737 if ((file->f_flags & O_NONBLOCK) || (file->f_mode & FMODE_NOWAIT))
1738 res |= REQ_F_SUPPORT_NOWAIT;
1742 bool io_alloc_async_data(struct io_kiocb *req)
1744 WARN_ON_ONCE(!io_cold_defs[req->opcode].async_size);
1745 req->async_data = kmalloc(io_cold_defs[req->opcode].async_size, GFP_KERNEL);
1746 if (req->async_data) {
1747 req->flags |= REQ_F_ASYNC_DATA;
1753 int io_req_prep_async(struct io_kiocb *req)
1755 const struct io_cold_def *cdef = &io_cold_defs[req->opcode];
1756 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1758 /* assign early for deferred execution for non-fixed file */
1759 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE) && !req->file)
1760 req->file = io_file_get_normal(req, req->cqe.fd);
1761 if (!cdef->prep_async)
1763 if (WARN_ON_ONCE(req_has_async_data(req)))
1765 if (!def->manual_alloc) {
1766 if (io_alloc_async_data(req))
1769 return cdef->prep_async(req);
1772 static u32 io_get_sequence(struct io_kiocb *req)
1774 u32 seq = req->ctx->cached_sq_head;
1775 struct io_kiocb *cur;
1777 /* need original cached_sq_head, but it was increased for each req */
1778 io_for_each_link(cur, req)
1783 static __cold void io_drain_req(struct io_kiocb *req)
1784 __must_hold(&ctx->uring_lock)
1786 struct io_ring_ctx *ctx = req->ctx;
1787 struct io_defer_entry *de;
1789 u32 seq = io_get_sequence(req);
1791 /* Still need defer if there is pending req in defer list. */
1792 spin_lock(&ctx->completion_lock);
1793 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1794 spin_unlock(&ctx->completion_lock);
1796 ctx->drain_active = false;
1797 io_req_task_queue(req);
1800 spin_unlock(&ctx->completion_lock);
1802 io_prep_async_link(req);
1803 de = kmalloc(sizeof(*de), GFP_KERNEL);
1806 io_req_defer_failed(req, ret);
1810 spin_lock(&ctx->completion_lock);
1811 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1812 spin_unlock(&ctx->completion_lock);
1817 trace_io_uring_defer(req);
1820 list_add_tail(&de->list, &ctx->defer_list);
1821 spin_unlock(&ctx->completion_lock);
1824 static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def,
1825 unsigned int issue_flags)
1827 if (req->file || !def->needs_file)
1830 if (req->flags & REQ_F_FIXED_FILE)
1831 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1833 req->file = io_file_get_normal(req, req->cqe.fd);
1838 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1840 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1841 const struct cred *creds = NULL;
1844 if (unlikely(!io_assign_file(req, def, issue_flags)))
1847 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1848 creds = override_creds(req->creds);
1850 if (!def->audit_skip)
1851 audit_uring_entry(req->opcode);
1853 ret = def->issue(req, issue_flags);
1855 if (!def->audit_skip)
1856 audit_uring_exit(!ret, ret);
1859 revert_creds(creds);
1861 if (ret == IOU_OK) {
1862 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1863 io_req_complete_defer(req);
1865 io_req_complete_post(req, issue_flags);
1866 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1869 /* If the op doesn't have a file, we're not polling for it */
1870 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
1871 io_iopoll_req_issued(req, issue_flags);
1876 int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts)
1878 io_tw_lock(req->ctx, ts);
1879 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT|
1880 IO_URING_F_COMPLETE_DEFER);
1883 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1885 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1886 struct io_kiocb *nxt = NULL;
1888 if (req_ref_put_and_test(req)) {
1889 if (req->flags & IO_REQ_LINK_FLAGS)
1890 nxt = io_req_find_next(req);
1893 return nxt ? &nxt->work : NULL;
1896 void io_wq_submit_work(struct io_wq_work *work)
1898 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1899 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1900 unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ;
1901 bool needs_poll = false;
1902 int ret = 0, err = -ECANCELED;
1904 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1905 if (!(req->flags & REQ_F_REFCOUNT))
1906 __io_req_set_refcount(req, 2);
1910 io_arm_ltimeout(req);
1912 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1913 if (work->flags & IO_WQ_WORK_CANCEL) {
1915 io_req_task_queue_fail(req, err);
1918 if (!io_assign_file(req, def, issue_flags)) {
1920 work->flags |= IO_WQ_WORK_CANCEL;
1924 if (req->flags & REQ_F_FORCE_ASYNC) {
1925 bool opcode_poll = def->pollin || def->pollout;
1927 if (opcode_poll && file_can_poll(req->file)) {
1929 issue_flags |= IO_URING_F_NONBLOCK;
1934 ret = io_issue_sqe(req, issue_flags);
1938 * We can get EAGAIN for iopolled IO even though we're
1939 * forcing a sync submission from here, since we can't
1940 * wait for request slots on the block side.
1943 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1949 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1951 /* aborted or ready, in either case retry blocking */
1953 issue_flags &= ~IO_URING_F_NONBLOCK;
1956 /* avoid locking problems by failing it from a clean context */
1958 io_req_task_queue_fail(req, ret);
1961 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1962 unsigned int issue_flags)
1964 struct io_ring_ctx *ctx = req->ctx;
1965 struct io_fixed_file *slot;
1966 struct file *file = NULL;
1968 io_ring_submit_lock(ctx, issue_flags);
1970 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
1972 fd = array_index_nospec(fd, ctx->nr_user_files);
1973 slot = io_fixed_file_slot(&ctx->file_table, fd);
1974 file = io_slot_file(slot);
1975 req->flags |= io_slot_flags(slot);
1976 io_req_set_rsrc_node(req, ctx, 0);
1978 io_ring_submit_unlock(ctx, issue_flags);
1982 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
1984 struct file *file = fget(fd);
1986 trace_io_uring_file_get(req, fd);
1988 /* we don't allow fixed io_uring files */
1989 if (file && io_is_uring_fops(file))
1990 io_req_track_inflight(req);
1994 static void io_queue_async(struct io_kiocb *req, int ret)
1995 __must_hold(&req->ctx->uring_lock)
1997 struct io_kiocb *linked_timeout;
1999 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
2000 io_req_defer_failed(req, ret);
2004 linked_timeout = io_prep_linked_timeout(req);
2006 switch (io_arm_poll_handler(req, 0)) {
2007 case IO_APOLL_READY:
2008 io_kbuf_recycle(req, 0);
2009 io_req_task_queue(req);
2011 case IO_APOLL_ABORTED:
2012 io_kbuf_recycle(req, 0);
2013 io_queue_iowq(req, NULL);
2020 io_queue_linked_timeout(linked_timeout);
2023 static inline void io_queue_sqe(struct io_kiocb *req)
2024 __must_hold(&req->ctx->uring_lock)
2028 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
2031 * We async punt it if the file wasn't marked NOWAIT, or if the file
2032 * doesn't support non-blocking read/write attempts
2035 io_arm_ltimeout(req);
2037 io_queue_async(req, ret);
2040 static void io_queue_sqe_fallback(struct io_kiocb *req)
2041 __must_hold(&req->ctx->uring_lock)
2043 if (unlikely(req->flags & REQ_F_FAIL)) {
2045 * We don't submit, fail them all, for that replace hardlinks
2046 * with normal links. Extra REQ_F_LINK is tolerated.
2048 req->flags &= ~REQ_F_HARDLINK;
2049 req->flags |= REQ_F_LINK;
2050 io_req_defer_failed(req, req->cqe.res);
2052 int ret = io_req_prep_async(req);
2054 if (unlikely(ret)) {
2055 io_req_defer_failed(req, ret);
2059 if (unlikely(req->ctx->drain_active))
2062 io_queue_iowq(req, NULL);
2067 * Check SQE restrictions (opcode and flags).
2069 * Returns 'true' if SQE is allowed, 'false' otherwise.
2071 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
2072 struct io_kiocb *req,
2073 unsigned int sqe_flags)
2075 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
2078 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
2079 ctx->restrictions.sqe_flags_required)
2082 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
2083 ctx->restrictions.sqe_flags_required))
2089 static void io_init_req_drain(struct io_kiocb *req)
2091 struct io_ring_ctx *ctx = req->ctx;
2092 struct io_kiocb *head = ctx->submit_state.link.head;
2094 ctx->drain_active = true;
2097 * If we need to drain a request in the middle of a link, drain
2098 * the head request and the next request/link after the current
2099 * link. Considering sequential execution of links,
2100 * REQ_F_IO_DRAIN will be maintained for every request of our
2103 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2104 ctx->drain_next = true;
2108 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
2109 const struct io_uring_sqe *sqe)
2110 __must_hold(&ctx->uring_lock)
2112 const struct io_issue_def *def;
2113 unsigned int sqe_flags;
2117 /* req is partially pre-initialised, see io_preinit_req() */
2118 req->opcode = opcode = READ_ONCE(sqe->opcode);
2119 /* same numerical values with corresponding REQ_F_*, safe to copy */
2120 req->flags = sqe_flags = READ_ONCE(sqe->flags);
2121 req->cqe.user_data = READ_ONCE(sqe->user_data);
2123 req->rsrc_node = NULL;
2124 req->task = current;
2126 if (unlikely(opcode >= IORING_OP_LAST)) {
2130 def = &io_issue_defs[opcode];
2131 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2132 /* enforce forwards compatibility on users */
2133 if (sqe_flags & ~SQE_VALID_FLAGS)
2135 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2136 if (!def->buffer_select)
2138 req->buf_index = READ_ONCE(sqe->buf_group);
2140 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2141 ctx->drain_disabled = true;
2142 if (sqe_flags & IOSQE_IO_DRAIN) {
2143 if (ctx->drain_disabled)
2145 io_init_req_drain(req);
2148 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2149 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2151 /* knock it to the slow queue path, will be drained there */
2152 if (ctx->drain_active)
2153 req->flags |= REQ_F_FORCE_ASYNC;
2154 /* if there is no link, we're at "next" request and need to drain */
2155 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2156 ctx->drain_next = false;
2157 ctx->drain_active = true;
2158 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2162 if (!def->ioprio && sqe->ioprio)
2164 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2167 if (def->needs_file) {
2168 struct io_submit_state *state = &ctx->submit_state;
2170 req->cqe.fd = READ_ONCE(sqe->fd);
2173 * Plug now if we have more than 2 IO left after this, and the
2174 * target is potentially a read/write to block based storage.
2176 if (state->need_plug && def->plug) {
2177 state->plug_started = true;
2178 state->need_plug = false;
2179 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2183 personality = READ_ONCE(sqe->personality);
2187 req->creds = xa_load(&ctx->personalities, personality);
2190 get_cred(req->creds);
2191 ret = security_uring_override_creds(req->creds);
2193 put_cred(req->creds);
2196 req->flags |= REQ_F_CREDS;
2199 return def->prep(req, sqe);
2202 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2203 struct io_kiocb *req, int ret)
2205 struct io_ring_ctx *ctx = req->ctx;
2206 struct io_submit_link *link = &ctx->submit_state.link;
2207 struct io_kiocb *head = link->head;
2209 trace_io_uring_req_failed(sqe, req, ret);
2212 * Avoid breaking links in the middle as it renders links with SQPOLL
2213 * unusable. Instead of failing eagerly, continue assembling the link if
2214 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2215 * should find the flag and handle the rest.
2217 req_fail_link_node(req, ret);
2218 if (head && !(head->flags & REQ_F_FAIL))
2219 req_fail_link_node(head, -ECANCELED);
2221 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2223 link->last->link = req;
2227 io_queue_sqe_fallback(req);
2232 link->last->link = req;
2239 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2240 const struct io_uring_sqe *sqe)
2241 __must_hold(&ctx->uring_lock)
2243 struct io_submit_link *link = &ctx->submit_state.link;
2246 ret = io_init_req(ctx, req, sqe);
2248 return io_submit_fail_init(sqe, req, ret);
2250 trace_io_uring_submit_req(req);
2253 * If we already have a head request, queue this one for async
2254 * submittal once the head completes. If we don't have a head but
2255 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2256 * submitted sync once the chain is complete. If none of those
2257 * conditions are true (normal request), then just queue it.
2259 if (unlikely(link->head)) {
2260 ret = io_req_prep_async(req);
2262 return io_submit_fail_init(sqe, req, ret);
2264 trace_io_uring_link(req, link->head);
2265 link->last->link = req;
2268 if (req->flags & IO_REQ_LINK_FLAGS)
2270 /* last request of the link, flush it */
2273 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2276 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2277 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2278 if (req->flags & IO_REQ_LINK_FLAGS) {
2283 io_queue_sqe_fallback(req);
2293 * Batched submission is done, ensure local IO is flushed out.
2295 static void io_submit_state_end(struct io_ring_ctx *ctx)
2297 struct io_submit_state *state = &ctx->submit_state;
2299 if (unlikely(state->link.head))
2300 io_queue_sqe_fallback(state->link.head);
2301 /* flush only after queuing links as they can generate completions */
2302 io_submit_flush_completions(ctx);
2303 if (state->plug_started)
2304 blk_finish_plug(&state->plug);
2308 * Start submission side cache.
2310 static void io_submit_state_start(struct io_submit_state *state,
2311 unsigned int max_ios)
2313 state->plug_started = false;
2314 state->need_plug = max_ios > 2;
2315 state->submit_nr = max_ios;
2316 /* set only head, no need to init link_last in advance */
2317 state->link.head = NULL;
2320 static void io_commit_sqring(struct io_ring_ctx *ctx)
2322 struct io_rings *rings = ctx->rings;
2325 * Ensure any loads from the SQEs are done at this point,
2326 * since once we write the new head, the application could
2327 * write new data to them.
2329 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2333 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2334 * that is mapped by userspace. This means that care needs to be taken to
2335 * ensure that reads are stable, as we cannot rely on userspace always
2336 * being a good citizen. If members of the sqe are validated and then later
2337 * used, it's important that those reads are done through READ_ONCE() to
2338 * prevent a re-load down the line.
2340 static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe)
2342 unsigned head, mask = ctx->sq_entries - 1;
2343 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2346 * The cached sq head (or cq tail) serves two purposes:
2348 * 1) allows us to batch the cost of updating the user visible
2350 * 2) allows the kernel side to track the head on its own, even
2351 * though the application is the one updating it.
2353 head = READ_ONCE(ctx->sq_array[sq_idx]);
2354 if (likely(head < ctx->sq_entries)) {
2355 /* double index for 128-byte SQEs, twice as long */
2356 if (ctx->flags & IORING_SETUP_SQE128)
2358 *sqe = &ctx->sq_sqes[head];
2362 /* drop invalid entries */
2363 spin_lock(&ctx->completion_lock);
2365 spin_unlock(&ctx->completion_lock);
2366 WRITE_ONCE(ctx->rings->sq_dropped,
2367 READ_ONCE(ctx->rings->sq_dropped) + 1);
2371 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2372 __must_hold(&ctx->uring_lock)
2374 unsigned int entries = io_sqring_entries(ctx);
2378 if (unlikely(!entries))
2380 /* make sure SQ entry isn't read before tail */
2381 ret = left = min(nr, entries);
2382 io_get_task_refs(left);
2383 io_submit_state_start(&ctx->submit_state, left);
2386 const struct io_uring_sqe *sqe;
2387 struct io_kiocb *req;
2389 if (unlikely(!io_alloc_req(ctx, &req)))
2391 if (unlikely(!io_get_sqe(ctx, &sqe))) {
2392 io_req_add_to_cache(req, ctx);
2397 * Continue submitting even for sqe failure if the
2398 * ring was setup with IORING_SETUP_SUBMIT_ALL
2400 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2401 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2407 if (unlikely(left)) {
2409 /* try again if it submitted nothing and can't allocate a req */
2410 if (!ret && io_req_cache_empty(ctx))
2412 current->io_uring->cached_refs += left;
2415 io_submit_state_end(ctx);
2416 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2417 io_commit_sqring(ctx);
2421 struct io_wait_queue {
2422 struct wait_queue_entry wq;
2423 struct io_ring_ctx *ctx;
2425 unsigned nr_timeouts;
2429 static inline bool io_has_work(struct io_ring_ctx *ctx)
2431 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
2432 !llist_empty(&ctx->work_llist);
2435 static inline bool io_should_wake(struct io_wait_queue *iowq)
2437 struct io_ring_ctx *ctx = iowq->ctx;
2438 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
2441 * Wake up if we have enough events, or if a timeout occurred since we
2442 * started waiting. For timeouts, we always want to return to userspace,
2443 * regardless of event count.
2445 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2448 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2449 int wake_flags, void *key)
2451 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq);
2454 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2455 * the task, and the next invocation will do it.
2457 if (io_should_wake(iowq) || io_has_work(iowq->ctx))
2458 return autoremove_wake_function(curr, mode, wake_flags, key);
2462 int io_run_task_work_sig(struct io_ring_ctx *ctx)
2464 if (!llist_empty(&ctx->work_llist)) {
2465 __set_current_state(TASK_RUNNING);
2466 if (io_run_local_work(ctx) > 0)
2469 if (io_run_task_work() > 0)
2471 if (task_sigpending(current))
2476 /* when returns >0, the caller should retry */
2477 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2478 struct io_wait_queue *iowq)
2482 if (unlikely(READ_ONCE(ctx->check_cq)))
2484 if (unlikely(!llist_empty(&ctx->work_llist)))
2486 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL)))
2488 if (unlikely(task_sigpending(current)))
2490 if (unlikely(io_should_wake(iowq)))
2494 * Use io_schedule_prepare/finish, so cpufreq can take into account
2495 * that the task is waiting for IO - turns out to be important for low
2498 token = io_schedule_prepare();
2500 if (iowq->timeout == KTIME_MAX)
2502 else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS))
2504 io_schedule_finish(token);
2509 * Wait until events become available, if we don't already have some. The
2510 * application must reap them itself, as they reside on the shared cq ring.
2512 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2513 const sigset_t __user *sig, size_t sigsz,
2514 struct __kernel_timespec __user *uts)
2516 struct io_wait_queue iowq;
2517 struct io_rings *rings = ctx->rings;
2520 if (!io_allowed_run_tw(ctx))
2522 if (!llist_empty(&ctx->work_llist))
2523 io_run_local_work(ctx);
2525 io_cqring_overflow_flush(ctx);
2526 /* if user messes with these they will just get an early return */
2527 if (__io_cqring_events_user(ctx) >= min_events)
2531 #ifdef CONFIG_COMPAT
2532 if (in_compat_syscall())
2533 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2537 ret = set_user_sigmask(sig, sigsz);
2543 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2544 iowq.wq.private = current;
2545 INIT_LIST_HEAD(&iowq.wq.entry);
2547 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2548 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2549 iowq.timeout = KTIME_MAX;
2552 struct timespec64 ts;
2554 if (get_timespec64(&ts, uts))
2556 iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2559 trace_io_uring_cqring_wait(ctx, min_events);
2561 unsigned long check_cq;
2563 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
2564 int nr_wait = (int) iowq.cq_tail - READ_ONCE(ctx->rings->cq.tail);
2566 atomic_set(&ctx->cq_wait_nr, nr_wait);
2567 set_current_state(TASK_INTERRUPTIBLE);
2569 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2570 TASK_INTERRUPTIBLE);
2573 ret = io_cqring_wait_schedule(ctx, &iowq);
2574 __set_current_state(TASK_RUNNING);
2575 atomic_set(&ctx->cq_wait_nr, 0);
2580 * Run task_work after scheduling and before io_should_wake().
2581 * If we got woken because of task_work being processed, run it
2582 * now rather than let the caller do another wait loop.
2585 if (!llist_empty(&ctx->work_llist))
2586 io_run_local_work(ctx);
2588 check_cq = READ_ONCE(ctx->check_cq);
2589 if (unlikely(check_cq)) {
2590 /* let the caller flush overflows, retry */
2591 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2592 io_cqring_do_overflow_flush(ctx);
2593 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) {
2599 if (io_should_wake(&iowq)) {
2606 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
2607 finish_wait(&ctx->cq_wait, &iowq.wq);
2608 restore_saved_sigmask_unless(ret == -EINTR);
2610 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2613 static void io_mem_free(void *ptr)
2620 page = virt_to_head_page(ptr);
2621 if (put_page_testzero(page))
2622 free_compound_page(page);
2625 static void io_pages_free(struct page ***pages, int npages)
2627 struct page **page_array;
2632 page_array = *pages;
2633 for (i = 0; i < npages; i++)
2634 unpin_user_page(page_array[i]);
2639 static void *__io_uaddr_map(struct page ***pages, unsigned short *npages,
2640 unsigned long uaddr, size_t size)
2642 struct page **page_array;
2643 unsigned int nr_pages;
2648 if (uaddr & (PAGE_SIZE - 1) || !size)
2649 return ERR_PTR(-EINVAL);
2651 nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
2652 if (nr_pages > USHRT_MAX)
2653 return ERR_PTR(-EINVAL);
2654 page_array = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
2656 return ERR_PTR(-ENOMEM);
2658 ret = pin_user_pages_fast(uaddr, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
2660 if (ret != nr_pages) {
2662 io_pages_free(&page_array, ret > 0 ? ret : 0);
2663 return ret < 0 ? ERR_PTR(ret) : ERR_PTR(-EFAULT);
2666 * Should be a single page. If the ring is small enough that we can
2667 * use a normal page, that is fine. If we need multiple pages, then
2668 * userspace should use a huge page. That's the only way to guarantee
2669 * that we get contigious memory, outside of just being lucky or
2670 * (currently) having low memory fragmentation.
2672 if (page_array[0] != page_array[ret - 1])
2674 *pages = page_array;
2676 return page_to_virt(page_array[0]);
2679 static void *io_rings_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2682 return __io_uaddr_map(&ctx->ring_pages, &ctx->n_ring_pages, uaddr,
2686 static void *io_sqes_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2689 return __io_uaddr_map(&ctx->sqe_pages, &ctx->n_sqe_pages, uaddr,
2693 static void io_rings_free(struct io_ring_ctx *ctx)
2695 if (!(ctx->flags & IORING_SETUP_NO_MMAP)) {
2696 io_mem_free(ctx->rings);
2697 io_mem_free(ctx->sq_sqes);
2699 ctx->sq_sqes = NULL;
2701 io_pages_free(&ctx->ring_pages, ctx->n_ring_pages);
2702 io_pages_free(&ctx->sqe_pages, ctx->n_sqe_pages);
2706 static void *io_mem_alloc(size_t size)
2708 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2711 ret = (void *) __get_free_pages(gfp, get_order(size));
2714 return ERR_PTR(-ENOMEM);
2717 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2718 unsigned int cq_entries, size_t *sq_offset)
2720 struct io_rings *rings;
2721 size_t off, sq_array_size;
2723 off = struct_size(rings, cqes, cq_entries);
2724 if (off == SIZE_MAX)
2726 if (ctx->flags & IORING_SETUP_CQE32) {
2727 if (check_shl_overflow(off, 1, &off))
2732 off = ALIGN(off, SMP_CACHE_BYTES);
2740 sq_array_size = array_size(sizeof(u32), sq_entries);
2741 if (sq_array_size == SIZE_MAX)
2744 if (check_add_overflow(off, sq_array_size, &off))
2750 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2751 unsigned int eventfd_async)
2753 struct io_ev_fd *ev_fd;
2754 __s32 __user *fds = arg;
2757 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2758 lockdep_is_held(&ctx->uring_lock));
2762 if (copy_from_user(&fd, fds, sizeof(*fds)))
2765 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2769 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2770 if (IS_ERR(ev_fd->cq_ev_fd)) {
2771 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2776 spin_lock(&ctx->completion_lock);
2777 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2778 spin_unlock(&ctx->completion_lock);
2780 ev_fd->eventfd_async = eventfd_async;
2781 ctx->has_evfd = true;
2782 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2783 atomic_set(&ev_fd->refs, 1);
2784 atomic_set(&ev_fd->ops, 0);
2788 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2790 struct io_ev_fd *ev_fd;
2792 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2793 lockdep_is_held(&ctx->uring_lock));
2795 ctx->has_evfd = false;
2796 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2797 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops))
2798 call_rcu(&ev_fd->rcu, io_eventfd_ops);
2805 static void io_req_caches_free(struct io_ring_ctx *ctx)
2807 struct io_kiocb *req;
2810 mutex_lock(&ctx->uring_lock);
2811 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
2813 while (!io_req_cache_empty(ctx)) {
2814 req = io_extract_req(ctx);
2815 kmem_cache_free(req_cachep, req);
2819 percpu_ref_put_many(&ctx->refs, nr);
2820 mutex_unlock(&ctx->uring_lock);
2823 static void io_rsrc_node_cache_free(struct io_cache_entry *entry)
2825 kfree(container_of(entry, struct io_rsrc_node, cache));
2828 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2830 io_sq_thread_finish(ctx);
2831 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2832 if (WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)))
2835 mutex_lock(&ctx->uring_lock);
2837 __io_sqe_buffers_unregister(ctx);
2839 __io_sqe_files_unregister(ctx);
2840 io_cqring_overflow_kill(ctx);
2841 io_eventfd_unregister(ctx);
2842 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2843 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2844 io_destroy_buffers(ctx);
2845 mutex_unlock(&ctx->uring_lock);
2847 put_cred(ctx->sq_creds);
2848 if (ctx->submitter_task)
2849 put_task_struct(ctx->submitter_task);
2851 /* there are no registered resources left, nobody uses it */
2853 io_rsrc_node_destroy(ctx, ctx->rsrc_node);
2855 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2857 #if defined(CONFIG_UNIX)
2858 if (ctx->ring_sock) {
2859 ctx->ring_sock->file = NULL; /* so that iput() is called */
2860 sock_release(ctx->ring_sock);
2863 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2865 io_alloc_cache_free(&ctx->rsrc_node_cache, io_rsrc_node_cache_free);
2866 if (ctx->mm_account) {
2867 mmdrop(ctx->mm_account);
2868 ctx->mm_account = NULL;
2872 percpu_ref_exit(&ctx->refs);
2873 free_uid(ctx->user);
2874 io_req_caches_free(ctx);
2876 io_wq_put_hash(ctx->hash_map);
2877 kfree(ctx->cancel_table.hbs);
2878 kfree(ctx->cancel_table_locked.hbs);
2880 xa_destroy(&ctx->io_bl_xa);
2884 static __cold void io_activate_pollwq_cb(struct callback_head *cb)
2886 struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx,
2889 mutex_lock(&ctx->uring_lock);
2890 ctx->poll_activated = true;
2891 mutex_unlock(&ctx->uring_lock);
2894 * Wake ups for some events between start of polling and activation
2895 * might've been lost due to loose synchronisation.
2897 wake_up_all(&ctx->poll_wq);
2898 percpu_ref_put(&ctx->refs);
2901 static __cold void io_activate_pollwq(struct io_ring_ctx *ctx)
2903 spin_lock(&ctx->completion_lock);
2904 /* already activated or in progress */
2905 if (ctx->poll_activated || ctx->poll_wq_task_work.func)
2907 if (WARN_ON_ONCE(!ctx->task_complete))
2909 if (!ctx->submitter_task)
2912 * with ->submitter_task only the submitter task completes requests, we
2913 * only need to sync with it, which is done by injecting a tw
2915 init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb);
2916 percpu_ref_get(&ctx->refs);
2917 if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL))
2918 percpu_ref_put(&ctx->refs);
2920 spin_unlock(&ctx->completion_lock);
2923 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2925 struct io_ring_ctx *ctx = file->private_data;
2928 if (unlikely(!ctx->poll_activated))
2929 io_activate_pollwq(ctx);
2931 poll_wait(file, &ctx->poll_wq, wait);
2933 * synchronizes with barrier from wq_has_sleeper call in
2937 if (!io_sqring_full(ctx))
2938 mask |= EPOLLOUT | EPOLLWRNORM;
2941 * Don't flush cqring overflow list here, just do a simple check.
2942 * Otherwise there could possible be ABBA deadlock:
2945 * lock(&ctx->uring_lock);
2947 * lock(&ctx->uring_lock);
2950 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2951 * pushes them to do the flush.
2954 if (__io_cqring_events_user(ctx) || io_has_work(ctx))
2955 mask |= EPOLLIN | EPOLLRDNORM;
2960 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2962 const struct cred *creds;
2964 creds = xa_erase(&ctx->personalities, id);
2973 struct io_tctx_exit {
2974 struct callback_head task_work;
2975 struct completion completion;
2976 struct io_ring_ctx *ctx;
2979 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2981 struct io_uring_task *tctx = current->io_uring;
2982 struct io_tctx_exit *work;
2984 work = container_of(cb, struct io_tctx_exit, task_work);
2986 * When @in_cancel, we're in cancellation and it's racy to remove the
2987 * node. It'll be removed by the end of cancellation, just ignore it.
2988 * tctx can be NULL if the queueing of this task_work raced with
2989 * work cancelation off the exec path.
2991 if (tctx && !atomic_read(&tctx->in_cancel))
2992 io_uring_del_tctx_node((unsigned long)work->ctx);
2993 complete(&work->completion);
2996 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2998 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3000 return req->ctx == data;
3003 static __cold void io_ring_exit_work(struct work_struct *work)
3005 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
3006 unsigned long timeout = jiffies + HZ * 60 * 5;
3007 unsigned long interval = HZ / 20;
3008 struct io_tctx_exit exit;
3009 struct io_tctx_node *node;
3013 * If we're doing polled IO and end up having requests being
3014 * submitted async (out-of-line), then completions can come in while
3015 * we're waiting for refs to drop. We need to reap these manually,
3016 * as nobody else will be looking for them.
3019 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
3020 mutex_lock(&ctx->uring_lock);
3021 io_cqring_overflow_kill(ctx);
3022 mutex_unlock(&ctx->uring_lock);
3025 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3026 io_move_task_work_from_local(ctx);
3028 while (io_uring_try_cancel_requests(ctx, NULL, true))
3032 struct io_sq_data *sqd = ctx->sq_data;
3033 struct task_struct *tsk;
3035 io_sq_thread_park(sqd);
3037 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
3038 io_wq_cancel_cb(tsk->io_uring->io_wq,
3039 io_cancel_ctx_cb, ctx, true);
3040 io_sq_thread_unpark(sqd);
3043 io_req_caches_free(ctx);
3045 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
3046 /* there is little hope left, don't run it too often */
3050 * This is really an uninterruptible wait, as it has to be
3051 * complete. But it's also run from a kworker, which doesn't
3052 * take signals, so it's fine to make it interruptible. This
3053 * avoids scenarios where we knowingly can wait much longer
3054 * on completions, for example if someone does a SIGSTOP on
3055 * a task that needs to finish task_work to make this loop
3056 * complete. That's a synthetic situation that should not
3057 * cause a stuck task backtrace, and hence a potential panic
3058 * on stuck tasks if that is enabled.
3060 } while (!wait_for_completion_interruptible_timeout(&ctx->ref_comp, interval));
3062 init_completion(&exit.completion);
3063 init_task_work(&exit.task_work, io_tctx_exit_cb);
3066 * Some may use context even when all refs and requests have been put,
3067 * and they are free to do so while still holding uring_lock or
3068 * completion_lock, see io_req_task_submit(). Apart from other work,
3069 * this lock/unlock section also waits them to finish.
3071 mutex_lock(&ctx->uring_lock);
3072 while (!list_empty(&ctx->tctx_list)) {
3073 WARN_ON_ONCE(time_after(jiffies, timeout));
3075 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
3077 /* don't spin on a single task if cancellation failed */
3078 list_rotate_left(&ctx->tctx_list);
3079 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
3080 if (WARN_ON_ONCE(ret))
3083 mutex_unlock(&ctx->uring_lock);
3085 * See comment above for
3086 * wait_for_completion_interruptible_timeout() on why this
3087 * wait is marked as interruptible.
3089 wait_for_completion_interruptible(&exit.completion);
3090 mutex_lock(&ctx->uring_lock);
3092 mutex_unlock(&ctx->uring_lock);
3093 spin_lock(&ctx->completion_lock);
3094 spin_unlock(&ctx->completion_lock);
3096 /* pairs with RCU read section in io_req_local_work_add() */
3097 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3100 io_ring_ctx_free(ctx);
3103 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
3105 unsigned long index;
3106 struct creds *creds;
3108 mutex_lock(&ctx->uring_lock);
3109 percpu_ref_kill(&ctx->refs);
3110 xa_for_each(&ctx->personalities, index, creds)
3111 io_unregister_personality(ctx, index);
3113 io_poll_remove_all(ctx, NULL, true);
3114 mutex_unlock(&ctx->uring_lock);
3117 * If we failed setting up the ctx, we might not have any rings
3118 * and therefore did not submit any requests
3121 io_kill_timeouts(ctx, NULL, true);
3123 flush_delayed_work(&ctx->fallback_work);
3125 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
3127 * Use system_unbound_wq to avoid spawning tons of event kworkers
3128 * if we're exiting a ton of rings at the same time. It just adds
3129 * noise and overhead, there's no discernable change in runtime
3130 * over using system_wq.
3132 queue_work(system_unbound_wq, &ctx->exit_work);
3135 static int io_uring_release(struct inode *inode, struct file *file)
3137 struct io_ring_ctx *ctx = file->private_data;
3139 file->private_data = NULL;
3140 io_ring_ctx_wait_and_kill(ctx);
3144 struct io_task_cancel {
3145 struct task_struct *task;
3149 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
3151 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3152 struct io_task_cancel *cancel = data;
3154 return io_match_task_safe(req, cancel->task, cancel->all);
3157 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
3158 struct task_struct *task,
3161 struct io_defer_entry *de;
3164 spin_lock(&ctx->completion_lock);
3165 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
3166 if (io_match_task_safe(de->req, task, cancel_all)) {
3167 list_cut_position(&list, &ctx->defer_list, &de->list);
3171 spin_unlock(&ctx->completion_lock);
3172 if (list_empty(&list))
3175 while (!list_empty(&list)) {
3176 de = list_first_entry(&list, struct io_defer_entry, list);
3177 list_del_init(&de->list);
3178 io_req_task_queue_fail(de->req, -ECANCELED);
3184 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
3186 struct io_tctx_node *node;
3187 enum io_wq_cancel cret;
3190 mutex_lock(&ctx->uring_lock);
3191 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3192 struct io_uring_task *tctx = node->task->io_uring;
3195 * io_wq will stay alive while we hold uring_lock, because it's
3196 * killed after ctx nodes, which requires to take the lock.
3198 if (!tctx || !tctx->io_wq)
3200 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
3201 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3203 mutex_unlock(&ctx->uring_lock);
3208 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
3209 struct task_struct *task,
3212 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
3213 struct io_uring_task *tctx = task ? task->io_uring : NULL;
3214 enum io_wq_cancel cret;
3217 /* set it so io_req_local_work_add() would wake us up */
3218 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
3219 atomic_set(&ctx->cq_wait_nr, 1);
3223 /* failed during ring init, it couldn't have issued any requests */
3228 ret |= io_uring_try_cancel_iowq(ctx);
3229 } else if (tctx && tctx->io_wq) {
3231 * Cancels requests of all rings, not only @ctx, but
3232 * it's fine as the task is in exit/exec.
3234 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
3236 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3239 /* SQPOLL thread does its own polling */
3240 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
3241 (ctx->sq_data && ctx->sq_data->thread == current)) {
3242 while (!wq_list_empty(&ctx->iopoll_list)) {
3243 io_iopoll_try_reap_events(ctx);
3249 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3250 io_allowed_defer_tw_run(ctx))
3251 ret |= io_run_local_work(ctx) > 0;
3252 ret |= io_cancel_defer_files(ctx, task, cancel_all);
3253 mutex_lock(&ctx->uring_lock);
3254 ret |= io_poll_remove_all(ctx, task, cancel_all);
3255 mutex_unlock(&ctx->uring_lock);
3256 ret |= io_kill_timeouts(ctx, task, cancel_all);
3258 ret |= io_run_task_work() > 0;
3262 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
3265 return atomic_read(&tctx->inflight_tracked);
3266 return percpu_counter_sum(&tctx->inflight);
3270 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3271 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3273 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
3275 struct io_uring_task *tctx = current->io_uring;
3276 struct io_ring_ctx *ctx;
3277 struct io_tctx_node *node;
3278 unsigned long index;
3282 WARN_ON_ONCE(sqd && sqd->thread != current);
3284 if (!current->io_uring)
3287 io_wq_exit_start(tctx->io_wq);
3289 atomic_inc(&tctx->in_cancel);
3293 io_uring_drop_tctx_refs(current);
3294 /* read completions before cancelations */
3295 inflight = tctx_inflight(tctx, !cancel_all);
3300 xa_for_each(&tctx->xa, index, node) {
3301 /* sqpoll task will cancel all its requests */
3302 if (node->ctx->sq_data)
3304 loop |= io_uring_try_cancel_requests(node->ctx,
3305 current, cancel_all);
3308 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3309 loop |= io_uring_try_cancel_requests(ctx,
3319 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3321 io_uring_drop_tctx_refs(current);
3322 xa_for_each(&tctx->xa, index, node) {
3323 if (!llist_empty(&node->ctx->work_llist)) {
3324 WARN_ON_ONCE(node->ctx->submitter_task &&
3325 node->ctx->submitter_task != current);
3330 * If we've seen completions, retry without waiting. This
3331 * avoids a race where a completion comes in before we did
3332 * prepare_to_wait().
3334 if (inflight == tctx_inflight(tctx, !cancel_all))
3337 finish_wait(&tctx->wait, &wait);
3340 io_uring_clean_tctx(tctx);
3343 * We shouldn't run task_works after cancel, so just leave
3344 * ->in_cancel set for normal exit.
3346 atomic_dec(&tctx->in_cancel);
3347 /* for exec all current's requests should be gone, kill tctx */
3348 __io_uring_free(current);
3352 void __io_uring_cancel(bool cancel_all)
3354 io_uring_cancel_generic(cancel_all, NULL);
3357 static void *io_uring_validate_mmap_request(struct file *file,
3358 loff_t pgoff, size_t sz)
3360 struct io_ring_ctx *ctx = file->private_data;
3361 loff_t offset = pgoff << PAGE_SHIFT;
3365 /* Don't allow mmap if the ring was setup without it */
3366 if (ctx->flags & IORING_SETUP_NO_MMAP)
3367 return ERR_PTR(-EINVAL);
3369 switch (offset & IORING_OFF_MMAP_MASK) {
3370 case IORING_OFF_SQ_RING:
3371 case IORING_OFF_CQ_RING:
3374 case IORING_OFF_SQES:
3377 case IORING_OFF_PBUF_RING: {
3380 bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT;
3381 mutex_lock(&ctx->uring_lock);
3382 ptr = io_pbuf_get_address(ctx, bgid);
3383 mutex_unlock(&ctx->uring_lock);
3385 return ERR_PTR(-EINVAL);
3389 return ERR_PTR(-EINVAL);
3392 page = virt_to_head_page(ptr);
3393 if (sz > page_size(page))
3394 return ERR_PTR(-EINVAL);
3401 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3403 size_t sz = vma->vm_end - vma->vm_start;
3407 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3409 return PTR_ERR(ptr);
3411 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3412 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3415 static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp,
3416 unsigned long addr, unsigned long len,
3417 unsigned long pgoff, unsigned long flags)
3419 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
3420 struct vm_unmapped_area_info info;
3424 * Do not allow to map to user-provided address to avoid breaking the
3425 * aliasing rules. Userspace is not able to guess the offset address of
3426 * kernel kmalloc()ed memory area.
3431 ptr = io_uring_validate_mmap_request(filp, pgoff, len);
3435 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
3437 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
3438 info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
3440 info.align_mask = PAGE_MASK & (SHM_COLOUR - 1UL);
3442 info.align_mask = PAGE_MASK & (SHMLBA - 1UL);
3444 info.align_offset = (unsigned long) ptr;
3447 * A failed mmap() very likely causes application failure,
3448 * so fall back to the bottom-up function here. This scenario
3449 * can happen with large stack limits and large mmap()
3452 addr = vm_unmapped_area(&info);
3453 if (offset_in_page(addr)) {
3455 info.low_limit = TASK_UNMAPPED_BASE;
3456 info.high_limit = mmap_end;
3457 addr = vm_unmapped_area(&info);
3463 #else /* !CONFIG_MMU */
3465 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3467 return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL;
3470 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3472 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3475 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3476 unsigned long addr, unsigned long len,
3477 unsigned long pgoff, unsigned long flags)
3481 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3483 return PTR_ERR(ptr);
3485 return (unsigned long) ptr;
3488 #endif /* !CONFIG_MMU */
3490 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3492 if (flags & IORING_ENTER_EXT_ARG) {
3493 struct io_uring_getevents_arg arg;
3495 if (argsz != sizeof(arg))
3497 if (copy_from_user(&arg, argp, sizeof(arg)))
3503 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3504 struct __kernel_timespec __user **ts,
3505 const sigset_t __user **sig)
3507 struct io_uring_getevents_arg arg;
3510 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3511 * is just a pointer to the sigset_t.
3513 if (!(flags & IORING_ENTER_EXT_ARG)) {
3514 *sig = (const sigset_t __user *) argp;
3520 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3521 * timespec and sigset_t pointers if good.
3523 if (*argsz != sizeof(arg))
3525 if (copy_from_user(&arg, argp, sizeof(arg)))
3529 *sig = u64_to_user_ptr(arg.sigmask);
3530 *argsz = arg.sigmask_sz;
3531 *ts = u64_to_user_ptr(arg.ts);
3535 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3536 u32, min_complete, u32, flags, const void __user *, argp,
3539 struct io_ring_ctx *ctx;
3543 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3544 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3545 IORING_ENTER_REGISTERED_RING)))
3549 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3550 * need only dereference our task private array to find it.
3552 if (flags & IORING_ENTER_REGISTERED_RING) {
3553 struct io_uring_task *tctx = current->io_uring;
3555 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3557 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3558 f.file = tctx->registered_rings[fd];
3560 if (unlikely(!f.file))
3564 if (unlikely(!f.file))
3567 if (unlikely(!io_is_uring_fops(f.file)))
3571 ctx = f.file->private_data;
3573 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3577 * For SQ polling, the thread will do all submissions and completions.
3578 * Just return the requested submit count, and wake the thread if
3582 if (ctx->flags & IORING_SETUP_SQPOLL) {
3583 io_cqring_overflow_flush(ctx);
3585 if (unlikely(ctx->sq_data->thread == NULL)) {
3589 if (flags & IORING_ENTER_SQ_WAKEUP)
3590 wake_up(&ctx->sq_data->wait);
3591 if (flags & IORING_ENTER_SQ_WAIT)
3592 io_sqpoll_wait_sq(ctx);
3595 } else if (to_submit) {
3596 ret = io_uring_add_tctx_node(ctx);
3600 mutex_lock(&ctx->uring_lock);
3601 ret = io_submit_sqes(ctx, to_submit);
3602 if (ret != to_submit) {
3603 mutex_unlock(&ctx->uring_lock);
3606 if (flags & IORING_ENTER_GETEVENTS) {
3607 if (ctx->syscall_iopoll)
3610 * Ignore errors, we'll soon call io_cqring_wait() and
3611 * it should handle ownership problems if any.
3613 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3614 (void)io_run_local_work_locked(ctx);
3616 mutex_unlock(&ctx->uring_lock);
3619 if (flags & IORING_ENTER_GETEVENTS) {
3622 if (ctx->syscall_iopoll) {
3624 * We disallow the app entering submit/complete with
3625 * polling, but we still need to lock the ring to
3626 * prevent racing with polled issue that got punted to
3629 mutex_lock(&ctx->uring_lock);
3631 ret2 = io_validate_ext_arg(flags, argp, argsz);
3632 if (likely(!ret2)) {
3633 min_complete = min(min_complete,
3635 ret2 = io_iopoll_check(ctx, min_complete);
3637 mutex_unlock(&ctx->uring_lock);
3639 const sigset_t __user *sig;
3640 struct __kernel_timespec __user *ts;
3642 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3643 if (likely(!ret2)) {
3644 min_complete = min(min_complete,
3646 ret2 = io_cqring_wait(ctx, min_complete, sig,
3655 * EBADR indicates that one or more CQE were dropped.
3656 * Once the user has been informed we can clear the bit
3657 * as they are obviously ok with those drops.
3659 if (unlikely(ret2 == -EBADR))
3660 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3669 static const struct file_operations io_uring_fops = {
3670 .release = io_uring_release,
3671 .mmap = io_uring_mmap,
3673 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3674 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3676 .get_unmapped_area = io_uring_mmu_get_unmapped_area,
3678 .poll = io_uring_poll,
3679 #ifdef CONFIG_PROC_FS
3680 .show_fdinfo = io_uring_show_fdinfo,
3684 bool io_is_uring_fops(struct file *file)
3686 return file->f_op == &io_uring_fops;
3689 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3690 struct io_uring_params *p)
3692 struct io_rings *rings;
3693 size_t size, sq_array_offset;
3696 /* make sure these are sane, as we already accounted them */
3697 ctx->sq_entries = p->sq_entries;
3698 ctx->cq_entries = p->cq_entries;
3700 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3701 if (size == SIZE_MAX)
3704 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3705 rings = io_mem_alloc(size);
3707 rings = io_rings_map(ctx, p->cq_off.user_addr, size);
3710 return PTR_ERR(rings);
3713 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3714 rings->sq_ring_mask = p->sq_entries - 1;
3715 rings->cq_ring_mask = p->cq_entries - 1;
3716 rings->sq_ring_entries = p->sq_entries;
3717 rings->cq_ring_entries = p->cq_entries;
3719 if (p->flags & IORING_SETUP_SQE128)
3720 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3722 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3723 if (size == SIZE_MAX) {
3728 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3729 ptr = io_mem_alloc(size);
3731 ptr = io_sqes_map(ctx, p->sq_off.user_addr, size);
3735 return PTR_ERR(ptr);
3742 static int io_uring_install_fd(struct file *file)
3746 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3749 fd_install(fd, file);
3754 * Allocate an anonymous fd, this is what constitutes the application
3755 * visible backing of an io_uring instance. The application mmaps this
3756 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3757 * we have to tie this fd to a socket for file garbage collection purposes.
3759 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3762 #if defined(CONFIG_UNIX)
3765 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3768 return ERR_PTR(ret);
3771 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3772 O_RDWR | O_CLOEXEC, NULL);
3773 #if defined(CONFIG_UNIX)
3775 sock_release(ctx->ring_sock);
3776 ctx->ring_sock = NULL;
3778 ctx->ring_sock->file = file;
3784 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3785 struct io_uring_params __user *params)
3787 struct io_ring_ctx *ctx;
3788 struct io_uring_task *tctx;
3794 if (entries > IORING_MAX_ENTRIES) {
3795 if (!(p->flags & IORING_SETUP_CLAMP))
3797 entries = IORING_MAX_ENTRIES;
3800 if ((p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
3801 && !(p->flags & IORING_SETUP_NO_MMAP))
3805 * Use twice as many entries for the CQ ring. It's possible for the
3806 * application to drive a higher depth than the size of the SQ ring,
3807 * since the sqes are only used at submission time. This allows for
3808 * some flexibility in overcommitting a bit. If the application has
3809 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3810 * of CQ ring entries manually.
3812 p->sq_entries = roundup_pow_of_two(entries);
3813 if (p->flags & IORING_SETUP_CQSIZE) {
3815 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3816 * to a power-of-two, if it isn't already. We do NOT impose
3817 * any cq vs sq ring sizing.
3821 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3822 if (!(p->flags & IORING_SETUP_CLAMP))
3824 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3826 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3827 if (p->cq_entries < p->sq_entries)
3830 p->cq_entries = 2 * p->sq_entries;
3833 ctx = io_ring_ctx_alloc(p);
3837 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3838 !(ctx->flags & IORING_SETUP_IOPOLL) &&
3839 !(ctx->flags & IORING_SETUP_SQPOLL))
3840 ctx->task_complete = true;
3842 if (ctx->task_complete || (ctx->flags & IORING_SETUP_IOPOLL))
3843 ctx->lockless_cq = true;
3846 * lazy poll_wq activation relies on ->task_complete for synchronisation
3847 * purposes, see io_activate_pollwq()
3849 if (!ctx->task_complete)
3850 ctx->poll_activated = true;
3853 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3854 * space applications don't need to do io completion events
3855 * polling again, they can rely on io_sq_thread to do polling
3856 * work, which can reduce cpu usage and uring_lock contention.
3858 if (ctx->flags & IORING_SETUP_IOPOLL &&
3859 !(ctx->flags & IORING_SETUP_SQPOLL))
3860 ctx->syscall_iopoll = 1;
3862 ctx->compat = in_compat_syscall();
3863 if (!capable(CAP_IPC_LOCK))
3864 ctx->user = get_uid(current_user());
3867 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3868 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3871 if (ctx->flags & IORING_SETUP_SQPOLL) {
3872 /* IPI related flags don't make sense with SQPOLL */
3873 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3874 IORING_SETUP_TASKRUN_FLAG |
3875 IORING_SETUP_DEFER_TASKRUN))
3877 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3878 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3879 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3881 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3882 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3884 ctx->notify_method = TWA_SIGNAL;
3888 * For DEFER_TASKRUN we require the completion task to be the same as the
3889 * submission task. This implies that there is only one submitter, so enforce
3892 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3893 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3898 * This is just grabbed for accounting purposes. When a process exits,
3899 * the mm is exited and dropped before the files, hence we need to hang
3900 * on to this mm purely for the purposes of being able to unaccount
3901 * memory (locked/pinned vm). It's not used for anything else.
3903 mmgrab(current->mm);
3904 ctx->mm_account = current->mm;
3906 ret = io_allocate_scq_urings(ctx, p);
3910 ret = io_sq_offload_create(ctx, p);
3914 ret = io_rsrc_init(ctx);
3918 p->sq_off.head = offsetof(struct io_rings, sq.head);
3919 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3920 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3921 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3922 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3923 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3924 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3925 p->sq_off.resv1 = 0;
3926 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3927 p->sq_off.user_addr = 0;
3929 p->cq_off.head = offsetof(struct io_rings, cq.head);
3930 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3931 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3932 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3933 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3934 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3935 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3936 p->cq_off.resv1 = 0;
3937 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3938 p->cq_off.user_addr = 0;
3940 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3941 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3942 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3943 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3944 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3945 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3946 IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING;
3948 if (copy_to_user(params, p, sizeof(*p))) {
3953 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
3954 && !(ctx->flags & IORING_SETUP_R_DISABLED))
3955 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
3957 file = io_uring_get_file(ctx);
3959 ret = PTR_ERR(file);
3963 ret = __io_uring_add_tctx_node(ctx);
3966 tctx = current->io_uring;
3969 * Install ring fd as the very last thing, so we don't risk someone
3970 * having closed it before we finish setup
3972 if (p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
3973 ret = io_ring_add_registered_file(tctx, file, 0, IO_RINGFD_REG_MAX);
3975 ret = io_uring_install_fd(file);
3979 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3982 io_ring_ctx_wait_and_kill(ctx);
3990 * Sets up an aio uring context, and returns the fd. Applications asks for a
3991 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3992 * params structure passed in.
3994 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3996 struct io_uring_params p;
3999 if (copy_from_user(&p, params, sizeof(p)))
4001 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
4006 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
4007 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
4008 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
4009 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
4010 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
4011 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
4012 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN |
4013 IORING_SETUP_NO_MMAP | IORING_SETUP_REGISTERED_FD_ONLY))
4016 return io_uring_create(entries, &p, params);
4019 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
4020 struct io_uring_params __user *, params)
4022 return io_uring_setup(entries, params);
4025 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
4028 struct io_uring_probe *p;
4032 size = struct_size(p, ops, nr_args);
4033 if (size == SIZE_MAX)
4035 p = kzalloc(size, GFP_KERNEL);
4040 if (copy_from_user(p, arg, size))
4043 if (memchr_inv(p, 0, size))
4046 p->last_op = IORING_OP_LAST - 1;
4047 if (nr_args > IORING_OP_LAST)
4048 nr_args = IORING_OP_LAST;
4050 for (i = 0; i < nr_args; i++) {
4052 if (!io_issue_defs[i].not_supported)
4053 p->ops[i].flags = IO_URING_OP_SUPPORTED;
4058 if (copy_to_user(arg, p, size))
4065 static int io_register_personality(struct io_ring_ctx *ctx)
4067 const struct cred *creds;
4071 creds = get_current_cred();
4073 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
4074 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
4082 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
4083 void __user *arg, unsigned int nr_args)
4085 struct io_uring_restriction *res;
4089 /* Restrictions allowed only if rings started disabled */
4090 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
4093 /* We allow only a single restrictions registration */
4094 if (ctx->restrictions.registered)
4097 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
4100 size = array_size(nr_args, sizeof(*res));
4101 if (size == SIZE_MAX)
4104 res = memdup_user(arg, size);
4106 return PTR_ERR(res);
4110 for (i = 0; i < nr_args; i++) {
4111 switch (res[i].opcode) {
4112 case IORING_RESTRICTION_REGISTER_OP:
4113 if (res[i].register_op >= IORING_REGISTER_LAST) {
4118 __set_bit(res[i].register_op,
4119 ctx->restrictions.register_op);
4121 case IORING_RESTRICTION_SQE_OP:
4122 if (res[i].sqe_op >= IORING_OP_LAST) {
4127 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
4129 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
4130 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
4132 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
4133 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
4142 /* Reset all restrictions if an error happened */
4144 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
4146 ctx->restrictions.registered = true;
4152 static int io_register_enable_rings(struct io_ring_ctx *ctx)
4154 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
4157 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task) {
4158 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
4160 * Lazy activation attempts would fail if it was polled before
4161 * submitter_task is set.
4163 if (wq_has_sleeper(&ctx->poll_wq))
4164 io_activate_pollwq(ctx);
4167 if (ctx->restrictions.registered)
4168 ctx->restricted = 1;
4170 ctx->flags &= ~IORING_SETUP_R_DISABLED;
4171 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
4172 wake_up(&ctx->sq_data->wait);
4176 static __cold int __io_register_iowq_aff(struct io_ring_ctx *ctx,
4177 cpumask_var_t new_mask)
4181 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
4182 ret = io_wq_cpu_affinity(current->io_uring, new_mask);
4184 mutex_unlock(&ctx->uring_lock);
4185 ret = io_sqpoll_wq_cpu_affinity(ctx, new_mask);
4186 mutex_lock(&ctx->uring_lock);
4192 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
4193 void __user *arg, unsigned len)
4195 cpumask_var_t new_mask;
4198 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4201 cpumask_clear(new_mask);
4202 if (len > cpumask_size())
4203 len = cpumask_size();
4205 if (in_compat_syscall()) {
4206 ret = compat_get_bitmap(cpumask_bits(new_mask),
4207 (const compat_ulong_t __user *)arg,
4208 len * 8 /* CHAR_BIT */);
4210 ret = copy_from_user(new_mask, arg, len);
4214 free_cpumask_var(new_mask);
4218 ret = __io_register_iowq_aff(ctx, new_mask);
4219 free_cpumask_var(new_mask);
4223 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
4225 return __io_register_iowq_aff(ctx, NULL);
4228 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
4230 __must_hold(&ctx->uring_lock)
4232 struct io_tctx_node *node;
4233 struct io_uring_task *tctx = NULL;
4234 struct io_sq_data *sqd = NULL;
4238 if (copy_from_user(new_count, arg, sizeof(new_count)))
4240 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4241 if (new_count[i] > INT_MAX)
4244 if (ctx->flags & IORING_SETUP_SQPOLL) {
4248 * Observe the correct sqd->lock -> ctx->uring_lock
4249 * ordering. Fine to drop uring_lock here, we hold
4252 refcount_inc(&sqd->refs);
4253 mutex_unlock(&ctx->uring_lock);
4254 mutex_lock(&sqd->lock);
4255 mutex_lock(&ctx->uring_lock);
4257 tctx = sqd->thread->io_uring;
4260 tctx = current->io_uring;
4263 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
4265 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4267 ctx->iowq_limits[i] = new_count[i];
4268 ctx->iowq_limits_set = true;
4270 if (tctx && tctx->io_wq) {
4271 ret = io_wq_max_workers(tctx->io_wq, new_count);
4275 memset(new_count, 0, sizeof(new_count));
4279 mutex_unlock(&sqd->lock);
4280 io_put_sq_data(sqd);
4283 if (copy_to_user(arg, new_count, sizeof(new_count)))
4286 /* that's it for SQPOLL, only the SQPOLL task creates requests */
4290 /* now propagate the restriction to all registered users */
4291 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
4292 struct io_uring_task *tctx = node->task->io_uring;
4294 if (WARN_ON_ONCE(!tctx->io_wq))
4297 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4298 new_count[i] = ctx->iowq_limits[i];
4299 /* ignore errors, it always returns zero anyway */
4300 (void)io_wq_max_workers(tctx->io_wq, new_count);
4305 mutex_unlock(&sqd->lock);
4306 io_put_sq_data(sqd);
4311 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
4312 void __user *arg, unsigned nr_args)
4313 __releases(ctx->uring_lock)
4314 __acquires(ctx->uring_lock)
4319 * We don't quiesce the refs for register anymore and so it can't be
4320 * dying as we're holding a file ref here.
4322 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
4325 if (ctx->submitter_task && ctx->submitter_task != current)
4328 if (ctx->restricted) {
4329 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
4330 if (!test_bit(opcode, ctx->restrictions.register_op))
4335 case IORING_REGISTER_BUFFERS:
4339 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
4341 case IORING_UNREGISTER_BUFFERS:
4345 ret = io_sqe_buffers_unregister(ctx);
4347 case IORING_REGISTER_FILES:
4351 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
4353 case IORING_UNREGISTER_FILES:
4357 ret = io_sqe_files_unregister(ctx);
4359 case IORING_REGISTER_FILES_UPDATE:
4360 ret = io_register_files_update(ctx, arg, nr_args);
4362 case IORING_REGISTER_EVENTFD:
4366 ret = io_eventfd_register(ctx, arg, 0);
4368 case IORING_REGISTER_EVENTFD_ASYNC:
4372 ret = io_eventfd_register(ctx, arg, 1);
4374 case IORING_UNREGISTER_EVENTFD:
4378 ret = io_eventfd_unregister(ctx);
4380 case IORING_REGISTER_PROBE:
4382 if (!arg || nr_args > 256)
4384 ret = io_probe(ctx, arg, nr_args);
4386 case IORING_REGISTER_PERSONALITY:
4390 ret = io_register_personality(ctx);
4392 case IORING_UNREGISTER_PERSONALITY:
4396 ret = io_unregister_personality(ctx, nr_args);
4398 case IORING_REGISTER_ENABLE_RINGS:
4402 ret = io_register_enable_rings(ctx);
4404 case IORING_REGISTER_RESTRICTIONS:
4405 ret = io_register_restrictions(ctx, arg, nr_args);
4407 case IORING_REGISTER_FILES2:
4408 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
4410 case IORING_REGISTER_FILES_UPDATE2:
4411 ret = io_register_rsrc_update(ctx, arg, nr_args,
4414 case IORING_REGISTER_BUFFERS2:
4415 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
4417 case IORING_REGISTER_BUFFERS_UPDATE:
4418 ret = io_register_rsrc_update(ctx, arg, nr_args,
4419 IORING_RSRC_BUFFER);
4421 case IORING_REGISTER_IOWQ_AFF:
4423 if (!arg || !nr_args)
4425 ret = io_register_iowq_aff(ctx, arg, nr_args);
4427 case IORING_UNREGISTER_IOWQ_AFF:
4431 ret = io_unregister_iowq_aff(ctx);
4433 case IORING_REGISTER_IOWQ_MAX_WORKERS:
4435 if (!arg || nr_args != 2)
4437 ret = io_register_iowq_max_workers(ctx, arg);
4439 case IORING_REGISTER_RING_FDS:
4440 ret = io_ringfd_register(ctx, arg, nr_args);
4442 case IORING_UNREGISTER_RING_FDS:
4443 ret = io_ringfd_unregister(ctx, arg, nr_args);
4445 case IORING_REGISTER_PBUF_RING:
4447 if (!arg || nr_args != 1)
4449 ret = io_register_pbuf_ring(ctx, arg);
4451 case IORING_UNREGISTER_PBUF_RING:
4453 if (!arg || nr_args != 1)
4455 ret = io_unregister_pbuf_ring(ctx, arg);
4457 case IORING_REGISTER_SYNC_CANCEL:
4459 if (!arg || nr_args != 1)
4461 ret = io_sync_cancel(ctx, arg);
4463 case IORING_REGISTER_FILE_ALLOC_RANGE:
4465 if (!arg || nr_args)
4467 ret = io_register_file_alloc_range(ctx, arg);
4477 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
4478 void __user *, arg, unsigned int, nr_args)
4480 struct io_ring_ctx *ctx;
4483 bool use_registered_ring;
4485 use_registered_ring = !!(opcode & IORING_REGISTER_USE_REGISTERED_RING);
4486 opcode &= ~IORING_REGISTER_USE_REGISTERED_RING;
4488 if (opcode >= IORING_REGISTER_LAST)
4491 if (use_registered_ring) {
4493 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
4494 * need only dereference our task private array to find it.
4496 struct io_uring_task *tctx = current->io_uring;
4498 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
4500 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
4501 f.file = tctx->registered_rings[fd];
4503 if (unlikely(!f.file))
4507 if (unlikely(!f.file))
4510 if (!io_is_uring_fops(f.file))
4514 ctx = f.file->private_data;
4516 mutex_lock(&ctx->uring_lock);
4517 ret = __io_uring_register(ctx, opcode, arg, nr_args);
4518 mutex_unlock(&ctx->uring_lock);
4519 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
4525 static int __init io_uring_init(void)
4527 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4528 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4529 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4532 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4533 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4534 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4535 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4536 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4537 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
4538 BUILD_BUG_SQE_ELEM(1, __u8, flags);
4539 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
4540 BUILD_BUG_SQE_ELEM(4, __s32, fd);
4541 BUILD_BUG_SQE_ELEM(8, __u64, off);
4542 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
4543 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
4544 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
4545 BUILD_BUG_SQE_ELEM(16, __u64, addr);
4546 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
4547 BUILD_BUG_SQE_ELEM(24, __u32, len);
4548 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
4549 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
4550 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4551 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
4552 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
4553 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
4554 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
4555 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
4556 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
4557 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
4558 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
4559 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
4560 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
4561 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
4562 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
4563 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
4564 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
4565 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
4566 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
4567 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
4568 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
4569 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
4570 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
4571 BUILD_BUG_SQE_ELEM(42, __u16, personality);
4572 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
4573 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
4574 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
4575 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
4576 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
4577 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4578 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
4580 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4581 sizeof(struct io_uring_rsrc_update));
4582 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4583 sizeof(struct io_uring_rsrc_update2));
4585 /* ->buf_index is u16 */
4586 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4587 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4588 offsetof(struct io_uring_buf_ring, tail));
4590 /* should fit into one byte */
4591 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
4592 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4593 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4595 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
4597 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4599 io_uring_optable_init();
4602 * Allow user copy in the per-command field, which starts after the
4603 * file in io_kiocb and until the opcode field. The openat2 handling
4604 * requires copying in user memory into the io_kiocb object in that
4605 * range, and HARDENED_USERCOPY will complain if we haven't
4606 * correctly annotated this range.
4608 req_cachep = kmem_cache_create_usercopy("io_kiocb",
4609 sizeof(struct io_kiocb), 0,
4610 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4611 SLAB_ACCOUNT | SLAB_TYPESAFE_BY_RCU,
4612 offsetof(struct io_kiocb, cmd.data),
4613 sizeof_field(struct io_kiocb, cmd.data), NULL);
4617 __initcall(io_uring_init);