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 static int __read_mostly sysctl_io_uring_disabled;
154 static int __read_mostly sysctl_io_uring_group = -1;
157 static struct ctl_table kernel_io_uring_disabled_table[] = {
159 .procname = "io_uring_disabled",
160 .data = &sysctl_io_uring_disabled,
161 .maxlen = sizeof(sysctl_io_uring_disabled),
163 .proc_handler = proc_dointvec_minmax,
164 .extra1 = SYSCTL_ZERO,
165 .extra2 = SYSCTL_TWO,
168 .procname = "io_uring_group",
169 .data = &sysctl_io_uring_group,
170 .maxlen = sizeof(gid_t),
172 .proc_handler = proc_dointvec,
178 struct sock *io_uring_get_socket(struct file *file)
180 #if defined(CONFIG_UNIX)
181 if (io_is_uring_fops(file)) {
182 struct io_ring_ctx *ctx = file->private_data;
184 return ctx->ring_sock->sk;
189 EXPORT_SYMBOL(io_uring_get_socket);
191 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
193 if (!wq_list_empty(&ctx->submit_state.compl_reqs) ||
194 ctx->submit_state.cqes_count)
195 __io_submit_flush_completions(ctx);
198 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
200 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
203 static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx)
205 return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head);
208 static bool io_match_linked(struct io_kiocb *head)
210 struct io_kiocb *req;
212 io_for_each_link(req, head) {
213 if (req->flags & REQ_F_INFLIGHT)
220 * As io_match_task() but protected against racing with linked timeouts.
221 * User must not hold timeout_lock.
223 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
228 if (task && head->task != task)
233 if (head->flags & REQ_F_LINK_TIMEOUT) {
234 struct io_ring_ctx *ctx = head->ctx;
236 /* protect against races with linked timeouts */
237 spin_lock_irq(&ctx->timeout_lock);
238 matched = io_match_linked(head);
239 spin_unlock_irq(&ctx->timeout_lock);
241 matched = io_match_linked(head);
246 static inline void req_fail_link_node(struct io_kiocb *req, int res)
249 io_req_set_res(req, res, 0);
252 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
254 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
257 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
259 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
261 complete(&ctx->ref_comp);
264 static __cold void io_fallback_req_func(struct work_struct *work)
266 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
268 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
269 struct io_kiocb *req, *tmp;
270 struct io_tw_state ts = { .locked = true, };
272 percpu_ref_get(&ctx->refs);
273 mutex_lock(&ctx->uring_lock);
274 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
275 req->io_task_work.func(req, &ts);
276 if (WARN_ON_ONCE(!ts.locked))
278 io_submit_flush_completions(ctx);
279 mutex_unlock(&ctx->uring_lock);
280 percpu_ref_put(&ctx->refs);
283 static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
285 unsigned hash_buckets = 1U << bits;
286 size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
288 table->hbs = kmalloc(hash_size, GFP_KERNEL);
292 table->hash_bits = bits;
293 init_hash_table(table, hash_buckets);
297 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
299 struct io_ring_ctx *ctx;
302 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
306 xa_init(&ctx->io_bl_xa);
309 * Use 5 bits less than the max cq entries, that should give us around
310 * 32 entries per hash list if totally full and uniformly spread, but
311 * don't keep too many buckets to not overconsume memory.
313 hash_bits = ilog2(p->cq_entries) - 5;
314 hash_bits = clamp(hash_bits, 1, 8);
315 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
317 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
319 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
323 ctx->flags = p->flags;
324 init_waitqueue_head(&ctx->sqo_sq_wait);
325 INIT_LIST_HEAD(&ctx->sqd_list);
326 INIT_LIST_HEAD(&ctx->cq_overflow_list);
327 INIT_LIST_HEAD(&ctx->io_buffers_cache);
328 INIT_HLIST_HEAD(&ctx->io_buf_list);
329 io_alloc_cache_init(&ctx->rsrc_node_cache, IO_NODE_ALLOC_CACHE_MAX,
330 sizeof(struct io_rsrc_node));
331 io_alloc_cache_init(&ctx->apoll_cache, IO_ALLOC_CACHE_MAX,
332 sizeof(struct async_poll));
333 io_alloc_cache_init(&ctx->netmsg_cache, IO_ALLOC_CACHE_MAX,
334 sizeof(struct io_async_msghdr));
335 init_completion(&ctx->ref_comp);
336 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
337 mutex_init(&ctx->uring_lock);
338 init_waitqueue_head(&ctx->cq_wait);
339 init_waitqueue_head(&ctx->poll_wq);
340 init_waitqueue_head(&ctx->rsrc_quiesce_wq);
341 spin_lock_init(&ctx->completion_lock);
342 spin_lock_init(&ctx->timeout_lock);
343 INIT_WQ_LIST(&ctx->iopoll_list);
344 INIT_LIST_HEAD(&ctx->io_buffers_pages);
345 INIT_LIST_HEAD(&ctx->io_buffers_comp);
346 INIT_LIST_HEAD(&ctx->defer_list);
347 INIT_LIST_HEAD(&ctx->timeout_list);
348 INIT_LIST_HEAD(&ctx->ltimeout_list);
349 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
350 init_llist_head(&ctx->work_llist);
351 INIT_LIST_HEAD(&ctx->tctx_list);
352 ctx->submit_state.free_list.next = NULL;
353 INIT_WQ_LIST(&ctx->locked_free_list);
354 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
355 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
358 kfree(ctx->cancel_table.hbs);
359 kfree(ctx->cancel_table_locked.hbs);
361 xa_destroy(&ctx->io_bl_xa);
366 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
368 struct io_rings *r = ctx->rings;
370 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
374 static bool req_need_defer(struct io_kiocb *req, u32 seq)
376 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
377 struct io_ring_ctx *ctx = req->ctx;
379 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
385 static void io_clean_op(struct io_kiocb *req)
387 if (req->flags & REQ_F_BUFFER_SELECTED) {
388 spin_lock(&req->ctx->completion_lock);
389 io_put_kbuf_comp(req);
390 spin_unlock(&req->ctx->completion_lock);
393 if (req->flags & REQ_F_NEED_CLEANUP) {
394 const struct io_cold_def *def = &io_cold_defs[req->opcode];
399 if ((req->flags & REQ_F_POLLED) && req->apoll) {
400 kfree(req->apoll->double_poll);
404 if (req->flags & REQ_F_INFLIGHT) {
405 struct io_uring_task *tctx = req->task->io_uring;
407 atomic_dec(&tctx->inflight_tracked);
409 if (req->flags & REQ_F_CREDS)
410 put_cred(req->creds);
411 if (req->flags & REQ_F_ASYNC_DATA) {
412 kfree(req->async_data);
413 req->async_data = NULL;
415 req->flags &= ~IO_REQ_CLEAN_FLAGS;
418 static inline void io_req_track_inflight(struct io_kiocb *req)
420 if (!(req->flags & REQ_F_INFLIGHT)) {
421 req->flags |= REQ_F_INFLIGHT;
422 atomic_inc(&req->task->io_uring->inflight_tracked);
426 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
428 if (WARN_ON_ONCE(!req->link))
431 req->flags &= ~REQ_F_ARM_LTIMEOUT;
432 req->flags |= REQ_F_LINK_TIMEOUT;
434 /* linked timeouts should have two refs once prep'ed */
435 io_req_set_refcount(req);
436 __io_req_set_refcount(req->link, 2);
440 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
442 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
444 return __io_prep_linked_timeout(req);
447 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
449 io_queue_linked_timeout(__io_prep_linked_timeout(req));
452 static inline void io_arm_ltimeout(struct io_kiocb *req)
454 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
455 __io_arm_ltimeout(req);
458 static void io_prep_async_work(struct io_kiocb *req)
460 const struct io_issue_def *def = &io_issue_defs[req->opcode];
461 struct io_ring_ctx *ctx = req->ctx;
463 if (!(req->flags & REQ_F_CREDS)) {
464 req->flags |= REQ_F_CREDS;
465 req->creds = get_current_cred();
468 req->work.list.next = NULL;
470 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
471 if (req->flags & REQ_F_FORCE_ASYNC)
472 req->work.flags |= IO_WQ_WORK_CONCURRENT;
474 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
475 req->flags |= io_file_get_flags(req->file);
477 if (req->file && (req->flags & REQ_F_ISREG)) {
478 bool should_hash = def->hash_reg_file;
480 /* don't serialize this request if the fs doesn't need it */
481 if (should_hash && (req->file->f_flags & O_DIRECT) &&
482 (req->file->f_mode & FMODE_DIO_PARALLEL_WRITE))
484 if (should_hash || (ctx->flags & IORING_SETUP_IOPOLL))
485 io_wq_hash_work(&req->work, file_inode(req->file));
486 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
487 if (def->unbound_nonreg_file)
488 req->work.flags |= IO_WQ_WORK_UNBOUND;
492 static void io_prep_async_link(struct io_kiocb *req)
494 struct io_kiocb *cur;
496 if (req->flags & REQ_F_LINK_TIMEOUT) {
497 struct io_ring_ctx *ctx = req->ctx;
499 spin_lock_irq(&ctx->timeout_lock);
500 io_for_each_link(cur, req)
501 io_prep_async_work(cur);
502 spin_unlock_irq(&ctx->timeout_lock);
504 io_for_each_link(cur, req)
505 io_prep_async_work(cur);
509 void io_queue_iowq(struct io_kiocb *req, struct io_tw_state *ts_dont_use)
511 struct io_kiocb *link = io_prep_linked_timeout(req);
512 struct io_uring_task *tctx = req->task->io_uring;
515 BUG_ON(!tctx->io_wq);
517 /* init ->work of the whole link before punting */
518 io_prep_async_link(req);
521 * Not expected to happen, but if we do have a bug where this _can_
522 * happen, catch it here and ensure the request is marked as
523 * canceled. That will make io-wq go through the usual work cancel
524 * procedure rather than attempt to run this request (or create a new
527 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
528 req->work.flags |= IO_WQ_WORK_CANCEL;
530 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
531 io_wq_enqueue(tctx->io_wq, &req->work);
533 io_queue_linked_timeout(link);
536 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
538 while (!list_empty(&ctx->defer_list)) {
539 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
540 struct io_defer_entry, list);
542 if (req_need_defer(de->req, de->seq))
544 list_del_init(&de->list);
545 io_req_task_queue(de->req);
551 static void io_eventfd_ops(struct rcu_head *rcu)
553 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
554 int ops = atomic_xchg(&ev_fd->ops, 0);
556 if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT))
557 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
559 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
560 * ordering in a race but if references are 0 we know we have to free
563 if (atomic_dec_and_test(&ev_fd->refs)) {
564 eventfd_ctx_put(ev_fd->cq_ev_fd);
569 static void io_eventfd_signal(struct io_ring_ctx *ctx)
571 struct io_ev_fd *ev_fd = NULL;
575 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
578 ev_fd = rcu_dereference(ctx->io_ev_fd);
581 * Check again if ev_fd exists incase an io_eventfd_unregister call
582 * completed between the NULL check of ctx->io_ev_fd at the start of
583 * the function and rcu_read_lock.
585 if (unlikely(!ev_fd))
587 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
589 if (ev_fd->eventfd_async && !io_wq_current_is_worker())
592 if (likely(eventfd_signal_allowed())) {
593 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
595 atomic_inc(&ev_fd->refs);
596 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops))
597 call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops);
599 atomic_dec(&ev_fd->refs);
606 static void io_eventfd_flush_signal(struct io_ring_ctx *ctx)
610 spin_lock(&ctx->completion_lock);
613 * Eventfd should only get triggered when at least one event has been
614 * posted. Some applications rely on the eventfd notification count
615 * only changing IFF a new CQE has been added to the CQ ring. There's
616 * no depedency on 1:1 relationship between how many times this
617 * function is called (and hence the eventfd count) and number of CQEs
618 * posted to the CQ ring.
620 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
621 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
622 spin_unlock(&ctx->completion_lock);
626 io_eventfd_signal(ctx);
629 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
631 if (ctx->poll_activated)
632 io_poll_wq_wake(ctx);
633 if (ctx->off_timeout_used)
634 io_flush_timeouts(ctx);
635 if (ctx->drain_active) {
636 spin_lock(&ctx->completion_lock);
637 io_queue_deferred(ctx);
638 spin_unlock(&ctx->completion_lock);
641 io_eventfd_flush_signal(ctx);
644 static inline void __io_cq_lock(struct io_ring_ctx *ctx)
646 if (!ctx->lockless_cq)
647 spin_lock(&ctx->completion_lock);
650 static inline void io_cq_lock(struct io_ring_ctx *ctx)
651 __acquires(ctx->completion_lock)
653 spin_lock(&ctx->completion_lock);
656 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
658 io_commit_cqring(ctx);
659 if (!ctx->task_complete) {
660 if (!ctx->lockless_cq)
661 spin_unlock(&ctx->completion_lock);
662 /* IOPOLL rings only need to wake up if it's also SQPOLL */
663 if (!ctx->syscall_iopoll)
666 io_commit_cqring_flush(ctx);
669 static void io_cq_unlock_post(struct io_ring_ctx *ctx)
670 __releases(ctx->completion_lock)
672 io_commit_cqring(ctx);
673 spin_unlock(&ctx->completion_lock);
675 io_commit_cqring_flush(ctx);
678 /* Returns true if there are no backlogged entries after the flush */
679 static void io_cqring_overflow_kill(struct io_ring_ctx *ctx)
681 struct io_overflow_cqe *ocqe;
684 spin_lock(&ctx->completion_lock);
685 list_splice_init(&ctx->cq_overflow_list, &list);
686 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
687 spin_unlock(&ctx->completion_lock);
689 while (!list_empty(&list)) {
690 ocqe = list_first_entry(&list, struct io_overflow_cqe, list);
691 list_del(&ocqe->list);
696 static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx)
698 size_t cqe_size = sizeof(struct io_uring_cqe);
700 if (__io_cqring_events(ctx) == ctx->cq_entries)
703 if (ctx->flags & IORING_SETUP_CQE32)
707 while (!list_empty(&ctx->cq_overflow_list)) {
708 struct io_uring_cqe *cqe;
709 struct io_overflow_cqe *ocqe;
711 if (!io_get_cqe_overflow(ctx, &cqe, true))
713 ocqe = list_first_entry(&ctx->cq_overflow_list,
714 struct io_overflow_cqe, list);
715 memcpy(cqe, &ocqe->cqe, cqe_size);
716 list_del(&ocqe->list);
720 if (list_empty(&ctx->cq_overflow_list)) {
721 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
722 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
724 io_cq_unlock_post(ctx);
727 static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx)
729 /* iopoll syncs against uring_lock, not completion_lock */
730 if (ctx->flags & IORING_SETUP_IOPOLL)
731 mutex_lock(&ctx->uring_lock);
732 __io_cqring_overflow_flush(ctx);
733 if (ctx->flags & IORING_SETUP_IOPOLL)
734 mutex_unlock(&ctx->uring_lock);
737 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx)
739 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
740 io_cqring_do_overflow_flush(ctx);
743 /* can be called by any task */
744 static void io_put_task_remote(struct task_struct *task)
746 struct io_uring_task *tctx = task->io_uring;
748 percpu_counter_sub(&tctx->inflight, 1);
749 if (unlikely(atomic_read(&tctx->in_cancel)))
750 wake_up(&tctx->wait);
751 put_task_struct(task);
754 /* used by a task to put its own references */
755 static void io_put_task_local(struct task_struct *task)
757 task->io_uring->cached_refs++;
760 /* must to be called somewhat shortly after putting a request */
761 static inline void io_put_task(struct task_struct *task)
763 if (likely(task == current))
764 io_put_task_local(task);
766 io_put_task_remote(task);
769 void io_task_refs_refill(struct io_uring_task *tctx)
771 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
773 percpu_counter_add(&tctx->inflight, refill);
774 refcount_add(refill, ¤t->usage);
775 tctx->cached_refs += refill;
778 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
780 struct io_uring_task *tctx = task->io_uring;
781 unsigned int refs = tctx->cached_refs;
784 tctx->cached_refs = 0;
785 percpu_counter_sub(&tctx->inflight, refs);
786 put_task_struct_many(task, refs);
790 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
791 s32 res, u32 cflags, u64 extra1, u64 extra2)
793 struct io_overflow_cqe *ocqe;
794 size_t ocq_size = sizeof(struct io_overflow_cqe);
795 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
797 lockdep_assert_held(&ctx->completion_lock);
800 ocq_size += sizeof(struct io_uring_cqe);
802 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
803 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
806 * If we're in ring overflow flush mode, or in task cancel mode,
807 * or cannot allocate an overflow entry, then we need to drop it
810 io_account_cq_overflow(ctx);
811 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
814 if (list_empty(&ctx->cq_overflow_list)) {
815 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
816 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
819 ocqe->cqe.user_data = user_data;
821 ocqe->cqe.flags = cflags;
823 ocqe->cqe.big_cqe[0] = extra1;
824 ocqe->cqe.big_cqe[1] = extra2;
826 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
830 void io_req_cqe_overflow(struct io_kiocb *req)
832 io_cqring_event_overflow(req->ctx, req->cqe.user_data,
833 req->cqe.res, req->cqe.flags,
834 req->big_cqe.extra1, req->big_cqe.extra2);
835 memset(&req->big_cqe, 0, sizeof(req->big_cqe));
839 * writes to the cq entry need to come after reading head; the
840 * control dependency is enough as we're using WRITE_ONCE to
843 bool io_cqe_cache_refill(struct io_ring_ctx *ctx, bool overflow)
845 struct io_rings *rings = ctx->rings;
846 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
847 unsigned int free, queued, len;
850 * Posting into the CQ when there are pending overflowed CQEs may break
851 * ordering guarantees, which will affect links, F_MORE users and more.
852 * Force overflow the completion.
854 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
857 /* userspace may cheat modifying the tail, be safe and do min */
858 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
859 free = ctx->cq_entries - queued;
860 /* we need a contiguous range, limit based on the current array offset */
861 len = min(free, ctx->cq_entries - off);
865 if (ctx->flags & IORING_SETUP_CQE32) {
870 ctx->cqe_cached = &rings->cqes[off];
871 ctx->cqe_sentinel = ctx->cqe_cached + len;
875 static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
878 struct io_uring_cqe *cqe;
883 * If we can't get a cq entry, userspace overflowed the
884 * submission (by quite a lot). Increment the overflow count in
887 if (likely(io_get_cqe(ctx, &cqe))) {
888 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
890 WRITE_ONCE(cqe->user_data, user_data);
891 WRITE_ONCE(cqe->res, res);
892 WRITE_ONCE(cqe->flags, cflags);
894 if (ctx->flags & IORING_SETUP_CQE32) {
895 WRITE_ONCE(cqe->big_cqe[0], 0);
896 WRITE_ONCE(cqe->big_cqe[1], 0);
903 static void __io_flush_post_cqes(struct io_ring_ctx *ctx)
904 __must_hold(&ctx->uring_lock)
906 struct io_submit_state *state = &ctx->submit_state;
909 lockdep_assert_held(&ctx->uring_lock);
910 for (i = 0; i < state->cqes_count; i++) {
911 struct io_uring_cqe *cqe = &ctx->completion_cqes[i];
913 if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) {
914 if (ctx->lockless_cq) {
915 spin_lock(&ctx->completion_lock);
916 io_cqring_event_overflow(ctx, cqe->user_data,
917 cqe->res, cqe->flags, 0, 0);
918 spin_unlock(&ctx->completion_lock);
920 io_cqring_event_overflow(ctx, cqe->user_data,
921 cqe->res, cqe->flags, 0, 0);
925 state->cqes_count = 0;
928 static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
934 filled = io_fill_cqe_aux(ctx, user_data, res, cflags);
935 if (!filled && allow_overflow)
936 filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
938 io_cq_unlock_post(ctx);
942 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags)
944 return __io_post_aux_cqe(ctx, user_data, res, cflags, true);
948 * A helper for multishot requests posting additional CQEs.
949 * Should only be used from a task_work including IO_URING_F_MULTISHOT.
951 bool io_fill_cqe_req_aux(struct io_kiocb *req, bool defer, s32 res, u32 cflags)
953 struct io_ring_ctx *ctx = req->ctx;
954 u64 user_data = req->cqe.user_data;
955 struct io_uring_cqe *cqe;
958 return __io_post_aux_cqe(ctx, user_data, res, cflags, false);
960 lockdep_assert_held(&ctx->uring_lock);
962 if (ctx->submit_state.cqes_count == ARRAY_SIZE(ctx->completion_cqes)) {
964 __io_flush_post_cqes(ctx);
965 /* no need to flush - flush is deferred */
966 __io_cq_unlock_post(ctx);
969 /* For defered completions this is not as strict as it is otherwise,
970 * however it's main job is to prevent unbounded posted completions,
971 * and in that it works just as well.
973 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
976 cqe = &ctx->completion_cqes[ctx->submit_state.cqes_count++];
977 cqe->user_data = user_data;
983 static void __io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
985 struct io_ring_ctx *ctx = req->ctx;
986 struct io_rsrc_node *rsrc_node = NULL;
989 if (!(req->flags & REQ_F_CQE_SKIP)) {
990 if (!io_fill_cqe_req(ctx, req))
991 io_req_cqe_overflow(req);
995 * If we're the last reference to this request, add to our locked
998 if (req_ref_put_and_test(req)) {
999 if (req->flags & IO_REQ_LINK_FLAGS) {
1000 if (req->flags & IO_DISARM_MASK)
1001 io_disarm_next(req);
1003 io_req_task_queue(req->link);
1007 io_put_kbuf_comp(req);
1008 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1012 rsrc_node = req->rsrc_node;
1014 * Selected buffer deallocation in io_clean_op() assumes that
1015 * we don't hold ->completion_lock. Clean them here to avoid
1018 io_put_task_remote(req->task);
1019 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
1020 ctx->locked_free_nr++;
1022 io_cq_unlock_post(ctx);
1025 io_ring_submit_lock(ctx, issue_flags);
1026 io_put_rsrc_node(ctx, rsrc_node);
1027 io_ring_submit_unlock(ctx, issue_flags);
1031 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
1033 if (req->ctx->task_complete && req->ctx->submitter_task != current) {
1034 req->io_task_work.func = io_req_task_complete;
1035 io_req_task_work_add(req);
1036 } else if (!(issue_flags & IO_URING_F_UNLOCKED) ||
1037 !(req->ctx->flags & IORING_SETUP_IOPOLL)) {
1038 __io_req_complete_post(req, issue_flags);
1040 struct io_ring_ctx *ctx = req->ctx;
1042 mutex_lock(&ctx->uring_lock);
1043 __io_req_complete_post(req, issue_flags & ~IO_URING_F_UNLOCKED);
1044 mutex_unlock(&ctx->uring_lock);
1048 void io_req_defer_failed(struct io_kiocb *req, s32 res)
1049 __must_hold(&ctx->uring_lock)
1051 const struct io_cold_def *def = &io_cold_defs[req->opcode];
1053 lockdep_assert_held(&req->ctx->uring_lock);
1056 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
1059 io_req_complete_defer(req);
1063 * Don't initialise the fields below on every allocation, but do that in
1064 * advance and keep them valid across allocations.
1066 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1070 req->async_data = NULL;
1071 /* not necessary, but safer to zero */
1072 memset(&req->cqe, 0, sizeof(req->cqe));
1073 memset(&req->big_cqe, 0, sizeof(req->big_cqe));
1076 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1077 struct io_submit_state *state)
1079 spin_lock(&ctx->completion_lock);
1080 wq_list_splice(&ctx->locked_free_list, &state->free_list);
1081 ctx->locked_free_nr = 0;
1082 spin_unlock(&ctx->completion_lock);
1086 * A request might get retired back into the request caches even before opcode
1087 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1088 * Because of that, io_alloc_req() should be called only under ->uring_lock
1089 * and with extra caution to not get a request that is still worked on.
1091 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
1092 __must_hold(&ctx->uring_lock)
1094 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1095 void *reqs[IO_REQ_ALLOC_BATCH];
1099 * If we have more than a batch's worth of requests in our IRQ side
1100 * locked cache, grab the lock and move them over to our submission
1103 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
1104 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
1105 if (!io_req_cache_empty(ctx))
1109 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
1112 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1113 * retry single alloc to be on the safe side.
1115 if (unlikely(ret <= 0)) {
1116 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1122 percpu_ref_get_many(&ctx->refs, ret);
1123 for (i = 0; i < ret; i++) {
1124 struct io_kiocb *req = reqs[i];
1126 io_preinit_req(req, ctx);
1127 io_req_add_to_cache(req, ctx);
1132 __cold void io_free_req(struct io_kiocb *req)
1134 /* refs were already put, restore them for io_req_task_complete() */
1135 req->flags &= ~REQ_F_REFCOUNT;
1136 /* we only want to free it, don't post CQEs */
1137 req->flags |= REQ_F_CQE_SKIP;
1138 req->io_task_work.func = io_req_task_complete;
1139 io_req_task_work_add(req);
1142 static void __io_req_find_next_prep(struct io_kiocb *req)
1144 struct io_ring_ctx *ctx = req->ctx;
1146 spin_lock(&ctx->completion_lock);
1147 io_disarm_next(req);
1148 spin_unlock(&ctx->completion_lock);
1151 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1153 struct io_kiocb *nxt;
1156 * If LINK is set, we have dependent requests in this chain. If we
1157 * didn't fail this request, queue the first one up, moving any other
1158 * dependencies to the next request. In case of failure, fail the rest
1161 if (unlikely(req->flags & IO_DISARM_MASK))
1162 __io_req_find_next_prep(req);
1168 static void ctx_flush_and_put(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1172 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1173 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1175 io_submit_flush_completions(ctx);
1176 mutex_unlock(&ctx->uring_lock);
1179 percpu_ref_put(&ctx->refs);
1182 static unsigned int handle_tw_list(struct llist_node *node,
1183 struct io_ring_ctx **ctx,
1184 struct io_tw_state *ts,
1185 struct llist_node *last)
1187 unsigned int count = 0;
1189 while (node && node != last) {
1190 struct llist_node *next = node->next;
1191 struct io_kiocb *req = container_of(node, struct io_kiocb,
1194 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1196 if (req->ctx != *ctx) {
1197 ctx_flush_and_put(*ctx, ts);
1199 /* if not contended, grab and improve batching */
1200 ts->locked = mutex_trylock(&(*ctx)->uring_lock);
1201 percpu_ref_get(&(*ctx)->refs);
1203 INDIRECT_CALL_2(req->io_task_work.func,
1204 io_poll_task_func, io_req_rw_complete,
1208 if (unlikely(need_resched())) {
1209 ctx_flush_and_put(*ctx, ts);
1219 * io_llist_xchg - swap all entries in a lock-less list
1220 * @head: the head of lock-less list to delete all entries
1221 * @new: new entry as the head of the list
1223 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1224 * The order of entries returned is from the newest to the oldest added one.
1226 static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1227 struct llist_node *new)
1229 return xchg(&head->first, new);
1233 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1234 * @head: the head of lock-less list to delete all entries
1235 * @old: expected old value of the first entry of the list
1236 * @new: new entry as the head of the list
1238 * perform a cmpxchg on the first entry of the list.
1241 static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1242 struct llist_node *old,
1243 struct llist_node *new)
1245 return cmpxchg(&head->first, old, new);
1248 static __cold void io_fallback_tw(struct io_uring_task *tctx, bool sync)
1250 struct llist_node *node = llist_del_all(&tctx->task_list);
1251 struct io_ring_ctx *last_ctx = NULL;
1252 struct io_kiocb *req;
1255 req = container_of(node, struct io_kiocb, io_task_work.node);
1257 if (sync && last_ctx != req->ctx) {
1259 flush_delayed_work(&last_ctx->fallback_work);
1260 percpu_ref_put(&last_ctx->refs);
1262 last_ctx = req->ctx;
1263 percpu_ref_get(&last_ctx->refs);
1265 if (llist_add(&req->io_task_work.node,
1266 &req->ctx->fallback_llist))
1267 schedule_delayed_work(&req->ctx->fallback_work, 1);
1271 flush_delayed_work(&last_ctx->fallback_work);
1272 percpu_ref_put(&last_ctx->refs);
1276 void tctx_task_work(struct callback_head *cb)
1278 struct io_tw_state ts = {};
1279 struct io_ring_ctx *ctx = NULL;
1280 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1282 struct llist_node fake = {};
1283 struct llist_node *node;
1284 unsigned int loops = 0;
1285 unsigned int count = 0;
1287 if (unlikely(current->flags & PF_EXITING)) {
1288 io_fallback_tw(tctx, true);
1294 node = io_llist_xchg(&tctx->task_list, &fake);
1295 count += handle_tw_list(node, &ctx, &ts, &fake);
1297 /* skip expensive cmpxchg if there are items in the list */
1298 if (READ_ONCE(tctx->task_list.first) != &fake)
1300 if (ts.locked && !wq_list_empty(&ctx->submit_state.compl_reqs)) {
1301 io_submit_flush_completions(ctx);
1302 if (READ_ONCE(tctx->task_list.first) != &fake)
1305 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1306 } while (node != &fake);
1308 ctx_flush_and_put(ctx, &ts);
1310 /* relaxed read is enough as only the task itself sets ->in_cancel */
1311 if (unlikely(atomic_read(&tctx->in_cancel)))
1312 io_uring_drop_tctx_refs(current);
1314 trace_io_uring_task_work_run(tctx, count, loops);
1317 static inline void io_req_local_work_add(struct io_kiocb *req, unsigned flags)
1319 struct io_ring_ctx *ctx = req->ctx;
1320 unsigned nr_wait, nr_tw, nr_tw_prev;
1321 struct llist_node *first;
1323 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK))
1324 flags &= ~IOU_F_TWQ_LAZY_WAKE;
1326 first = READ_ONCE(ctx->work_llist.first);
1330 struct io_kiocb *first_req = container_of(first,
1334 * Might be executed at any moment, rely on
1335 * SLAB_TYPESAFE_BY_RCU to keep it alive.
1337 nr_tw_prev = READ_ONCE(first_req->nr_tw);
1339 nr_tw = nr_tw_prev + 1;
1340 /* Large enough to fail the nr_wait comparison below */
1341 if (!(flags & IOU_F_TWQ_LAZY_WAKE))
1345 req->io_task_work.node.next = first;
1346 } while (!try_cmpxchg(&ctx->work_llist.first, &first,
1347 &req->io_task_work.node));
1350 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1351 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1353 io_eventfd_signal(ctx);
1356 nr_wait = atomic_read(&ctx->cq_wait_nr);
1357 /* no one is waiting */
1360 /* either not enough or the previous add has already woken it up */
1361 if (nr_wait > nr_tw || nr_tw_prev >= nr_wait)
1363 /* pairs with set_current_state() in io_cqring_wait() */
1364 smp_mb__after_atomic();
1365 wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE);
1368 static void io_req_normal_work_add(struct io_kiocb *req)
1370 struct io_uring_task *tctx = req->task->io_uring;
1371 struct io_ring_ctx *ctx = req->ctx;
1373 /* task_work already pending, we're done */
1374 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1377 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1378 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1380 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1383 io_fallback_tw(tctx, false);
1386 void __io_req_task_work_add(struct io_kiocb *req, unsigned flags)
1388 if (req->ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
1390 io_req_local_work_add(req, flags);
1393 io_req_normal_work_add(req);
1397 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1399 struct llist_node *node;
1401 node = llist_del_all(&ctx->work_llist);
1403 struct io_kiocb *req = container_of(node, struct io_kiocb,
1407 io_req_normal_work_add(req);
1411 static int __io_run_local_work(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1413 struct llist_node *node;
1414 unsigned int loops = 0;
1417 if (WARN_ON_ONCE(ctx->submitter_task != current))
1419 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1420 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1423 * llists are in reverse order, flip it back the right way before
1424 * running the pending items.
1426 node = llist_reverse_order(io_llist_xchg(&ctx->work_llist, NULL));
1428 struct llist_node *next = node->next;
1429 struct io_kiocb *req = container_of(node, struct io_kiocb,
1431 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1432 INDIRECT_CALL_2(req->io_task_work.func,
1433 io_poll_task_func, io_req_rw_complete,
1440 if (!llist_empty(&ctx->work_llist))
1443 io_submit_flush_completions(ctx);
1444 if (!llist_empty(&ctx->work_llist))
1447 trace_io_uring_local_work_run(ctx, ret, loops);
1451 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx)
1453 struct io_tw_state ts = { .locked = true, };
1456 if (llist_empty(&ctx->work_llist))
1459 ret = __io_run_local_work(ctx, &ts);
1460 /* shouldn't happen! */
1461 if (WARN_ON_ONCE(!ts.locked))
1462 mutex_lock(&ctx->uring_lock);
1466 static int io_run_local_work(struct io_ring_ctx *ctx)
1468 struct io_tw_state ts = {};
1471 ts.locked = mutex_trylock(&ctx->uring_lock);
1472 ret = __io_run_local_work(ctx, &ts);
1474 mutex_unlock(&ctx->uring_lock);
1479 static void io_req_task_cancel(struct io_kiocb *req, struct io_tw_state *ts)
1481 io_tw_lock(req->ctx, ts);
1482 io_req_defer_failed(req, req->cqe.res);
1485 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts)
1487 io_tw_lock(req->ctx, ts);
1488 /* req->task == current here, checking PF_EXITING is safe */
1489 if (unlikely(req->task->flags & PF_EXITING))
1490 io_req_defer_failed(req, -EFAULT);
1491 else if (req->flags & REQ_F_FORCE_ASYNC)
1492 io_queue_iowq(req, ts);
1497 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1499 io_req_set_res(req, ret, 0);
1500 req->io_task_work.func = io_req_task_cancel;
1501 io_req_task_work_add(req);
1504 void io_req_task_queue(struct io_kiocb *req)
1506 req->io_task_work.func = io_req_task_submit;
1507 io_req_task_work_add(req);
1510 void io_queue_next(struct io_kiocb *req)
1512 struct io_kiocb *nxt = io_req_find_next(req);
1515 io_req_task_queue(nxt);
1518 static void io_free_batch_list(struct io_ring_ctx *ctx,
1519 struct io_wq_work_node *node)
1520 __must_hold(&ctx->uring_lock)
1523 struct io_kiocb *req = container_of(node, struct io_kiocb,
1526 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1527 if (req->flags & REQ_F_REFCOUNT) {
1528 node = req->comp_list.next;
1529 if (!req_ref_put_and_test(req))
1532 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1533 struct async_poll *apoll = req->apoll;
1535 if (apoll->double_poll)
1536 kfree(apoll->double_poll);
1537 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1539 req->flags &= ~REQ_F_POLLED;
1541 if (req->flags & IO_REQ_LINK_FLAGS)
1543 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1548 io_req_put_rsrc_locked(req, ctx);
1550 io_put_task(req->task);
1551 node = req->comp_list.next;
1552 io_req_add_to_cache(req, ctx);
1556 void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1557 __must_hold(&ctx->uring_lock)
1559 struct io_submit_state *state = &ctx->submit_state;
1560 struct io_wq_work_node *node;
1563 /* must come first to preserve CQE ordering in failure cases */
1564 if (state->cqes_count)
1565 __io_flush_post_cqes(ctx);
1566 __wq_list_for_each(node, &state->compl_reqs) {
1567 struct io_kiocb *req = container_of(node, struct io_kiocb,
1570 if (!(req->flags & REQ_F_CQE_SKIP) &&
1571 unlikely(!io_fill_cqe_req(ctx, req))) {
1572 if (ctx->lockless_cq) {
1573 spin_lock(&ctx->completion_lock);
1574 io_req_cqe_overflow(req);
1575 spin_unlock(&ctx->completion_lock);
1577 io_req_cqe_overflow(req);
1581 __io_cq_unlock_post(ctx);
1583 if (!wq_list_empty(&ctx->submit_state.compl_reqs)) {
1584 io_free_batch_list(ctx, state->compl_reqs.first);
1585 INIT_WQ_LIST(&state->compl_reqs);
1589 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1591 /* See comment at the top of this file */
1593 return __io_cqring_events(ctx);
1597 * We can't just wait for polled events to come to us, we have to actively
1598 * find and complete them.
1600 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1602 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1605 mutex_lock(&ctx->uring_lock);
1606 while (!wq_list_empty(&ctx->iopoll_list)) {
1607 /* let it sleep and repeat later if can't complete a request */
1608 if (io_do_iopoll(ctx, true) == 0)
1611 * Ensure we allow local-to-the-cpu processing to take place,
1612 * in this case we need to ensure that we reap all events.
1613 * Also let task_work, etc. to progress by releasing the mutex
1615 if (need_resched()) {
1616 mutex_unlock(&ctx->uring_lock);
1618 mutex_lock(&ctx->uring_lock);
1621 mutex_unlock(&ctx->uring_lock);
1624 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1626 unsigned int nr_events = 0;
1627 unsigned long check_cq;
1629 if (!io_allowed_run_tw(ctx))
1632 check_cq = READ_ONCE(ctx->check_cq);
1633 if (unlikely(check_cq)) {
1634 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1635 __io_cqring_overflow_flush(ctx);
1637 * Similarly do not spin if we have not informed the user of any
1640 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1644 * Don't enter poll loop if we already have events pending.
1645 * If we do, we can potentially be spinning for commands that
1646 * already triggered a CQE (eg in error).
1648 if (io_cqring_events(ctx))
1655 * If a submit got punted to a workqueue, we can have the
1656 * application entering polling for a command before it gets
1657 * issued. That app will hold the uring_lock for the duration
1658 * of the poll right here, so we need to take a breather every
1659 * now and then to ensure that the issue has a chance to add
1660 * the poll to the issued list. Otherwise we can spin here
1661 * forever, while the workqueue is stuck trying to acquire the
1664 if (wq_list_empty(&ctx->iopoll_list) ||
1665 io_task_work_pending(ctx)) {
1666 u32 tail = ctx->cached_cq_tail;
1668 (void) io_run_local_work_locked(ctx);
1670 if (task_work_pending(current) ||
1671 wq_list_empty(&ctx->iopoll_list)) {
1672 mutex_unlock(&ctx->uring_lock);
1674 mutex_lock(&ctx->uring_lock);
1676 /* some requests don't go through iopoll_list */
1677 if (tail != ctx->cached_cq_tail ||
1678 wq_list_empty(&ctx->iopoll_list))
1681 ret = io_do_iopoll(ctx, !min);
1682 if (unlikely(ret < 0))
1685 if (task_sigpending(current))
1691 } while (nr_events < min);
1696 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts)
1699 io_req_complete_defer(req);
1701 io_req_complete_post(req, IO_URING_F_UNLOCKED);
1705 * After the iocb has been issued, it's safe to be found on the poll list.
1706 * Adding the kiocb to the list AFTER submission ensures that we don't
1707 * find it from a io_do_iopoll() thread before the issuer is done
1708 * accessing the kiocb cookie.
1710 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1712 struct io_ring_ctx *ctx = req->ctx;
1713 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1715 /* workqueue context doesn't hold uring_lock, grab it now */
1716 if (unlikely(needs_lock))
1717 mutex_lock(&ctx->uring_lock);
1720 * Track whether we have multiple files in our lists. This will impact
1721 * how we do polling eventually, not spinning if we're on potentially
1722 * different devices.
1724 if (wq_list_empty(&ctx->iopoll_list)) {
1725 ctx->poll_multi_queue = false;
1726 } else if (!ctx->poll_multi_queue) {
1727 struct io_kiocb *list_req;
1729 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1731 if (list_req->file != req->file)
1732 ctx->poll_multi_queue = true;
1736 * For fast devices, IO may have already completed. If it has, add
1737 * it to the front so we find it first.
1739 if (READ_ONCE(req->iopoll_completed))
1740 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1742 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1744 if (unlikely(needs_lock)) {
1746 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1747 * in sq thread task context or in io worker task context. If
1748 * current task context is sq thread, we don't need to check
1749 * whether should wake up sq thread.
1751 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1752 wq_has_sleeper(&ctx->sq_data->wait))
1753 wake_up(&ctx->sq_data->wait);
1755 mutex_unlock(&ctx->uring_lock);
1759 unsigned int io_file_get_flags(struct file *file)
1761 unsigned int res = 0;
1763 if (S_ISREG(file_inode(file)->i_mode))
1765 if ((file->f_flags & O_NONBLOCK) || (file->f_mode & FMODE_NOWAIT))
1766 res |= REQ_F_SUPPORT_NOWAIT;
1770 bool io_alloc_async_data(struct io_kiocb *req)
1772 WARN_ON_ONCE(!io_cold_defs[req->opcode].async_size);
1773 req->async_data = kmalloc(io_cold_defs[req->opcode].async_size, GFP_KERNEL);
1774 if (req->async_data) {
1775 req->flags |= REQ_F_ASYNC_DATA;
1781 int io_req_prep_async(struct io_kiocb *req)
1783 const struct io_cold_def *cdef = &io_cold_defs[req->opcode];
1784 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1786 /* assign early for deferred execution for non-fixed file */
1787 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE) && !req->file)
1788 req->file = io_file_get_normal(req, req->cqe.fd);
1789 if (!cdef->prep_async)
1791 if (WARN_ON_ONCE(req_has_async_data(req)))
1793 if (!def->manual_alloc) {
1794 if (io_alloc_async_data(req))
1797 return cdef->prep_async(req);
1800 static u32 io_get_sequence(struct io_kiocb *req)
1802 u32 seq = req->ctx->cached_sq_head;
1803 struct io_kiocb *cur;
1805 /* need original cached_sq_head, but it was increased for each req */
1806 io_for_each_link(cur, req)
1811 static __cold void io_drain_req(struct io_kiocb *req)
1812 __must_hold(&ctx->uring_lock)
1814 struct io_ring_ctx *ctx = req->ctx;
1815 struct io_defer_entry *de;
1817 u32 seq = io_get_sequence(req);
1819 /* Still need defer if there is pending req in defer list. */
1820 spin_lock(&ctx->completion_lock);
1821 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1822 spin_unlock(&ctx->completion_lock);
1824 ctx->drain_active = false;
1825 io_req_task_queue(req);
1828 spin_unlock(&ctx->completion_lock);
1830 io_prep_async_link(req);
1831 de = kmalloc(sizeof(*de), GFP_KERNEL);
1834 io_req_defer_failed(req, ret);
1838 spin_lock(&ctx->completion_lock);
1839 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1840 spin_unlock(&ctx->completion_lock);
1845 trace_io_uring_defer(req);
1848 list_add_tail(&de->list, &ctx->defer_list);
1849 spin_unlock(&ctx->completion_lock);
1852 static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def,
1853 unsigned int issue_flags)
1855 if (req->file || !def->needs_file)
1858 if (req->flags & REQ_F_FIXED_FILE)
1859 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1861 req->file = io_file_get_normal(req, req->cqe.fd);
1866 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1868 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1869 const struct cred *creds = NULL;
1872 if (unlikely(!io_assign_file(req, def, issue_flags)))
1875 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1876 creds = override_creds(req->creds);
1878 if (!def->audit_skip)
1879 audit_uring_entry(req->opcode);
1881 ret = def->issue(req, issue_flags);
1883 if (!def->audit_skip)
1884 audit_uring_exit(!ret, ret);
1887 revert_creds(creds);
1889 if (ret == IOU_OK) {
1890 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1891 io_req_complete_defer(req);
1893 io_req_complete_post(req, issue_flags);
1898 if (ret != IOU_ISSUE_SKIP_COMPLETE)
1901 /* If the op doesn't have a file, we're not polling for it */
1902 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
1903 io_iopoll_req_issued(req, issue_flags);
1908 int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts)
1910 io_tw_lock(req->ctx, ts);
1911 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT|
1912 IO_URING_F_COMPLETE_DEFER);
1915 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1917 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1918 struct io_kiocb *nxt = NULL;
1920 if (req_ref_put_and_test(req)) {
1921 if (req->flags & IO_REQ_LINK_FLAGS)
1922 nxt = io_req_find_next(req);
1925 return nxt ? &nxt->work : NULL;
1928 void io_wq_submit_work(struct io_wq_work *work)
1930 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1931 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1932 unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ;
1933 bool needs_poll = false;
1934 int ret = 0, err = -ECANCELED;
1936 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1937 if (!(req->flags & REQ_F_REFCOUNT))
1938 __io_req_set_refcount(req, 2);
1942 io_arm_ltimeout(req);
1944 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1945 if (work->flags & IO_WQ_WORK_CANCEL) {
1947 io_req_task_queue_fail(req, err);
1950 if (!io_assign_file(req, def, issue_flags)) {
1952 work->flags |= IO_WQ_WORK_CANCEL;
1956 if (req->flags & REQ_F_FORCE_ASYNC) {
1957 bool opcode_poll = def->pollin || def->pollout;
1959 if (opcode_poll && file_can_poll(req->file)) {
1961 issue_flags |= IO_URING_F_NONBLOCK;
1966 ret = io_issue_sqe(req, issue_flags);
1971 * If REQ_F_NOWAIT is set, then don't wait or retry with
1972 * poll. -EAGAIN is final for that case.
1974 if (req->flags & REQ_F_NOWAIT)
1978 * We can get EAGAIN for iopolled IO even though we're
1979 * forcing a sync submission from here, since we can't
1980 * wait for request slots on the block side.
1983 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1985 if (io_wq_worker_stopped())
1991 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1993 /* aborted or ready, in either case retry blocking */
1995 issue_flags &= ~IO_URING_F_NONBLOCK;
1998 /* avoid locking problems by failing it from a clean context */
2000 io_req_task_queue_fail(req, ret);
2003 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
2004 unsigned int issue_flags)
2006 struct io_ring_ctx *ctx = req->ctx;
2007 struct io_fixed_file *slot;
2008 struct file *file = NULL;
2010 io_ring_submit_lock(ctx, issue_flags);
2012 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
2014 fd = array_index_nospec(fd, ctx->nr_user_files);
2015 slot = io_fixed_file_slot(&ctx->file_table, fd);
2016 file = io_slot_file(slot);
2017 req->flags |= io_slot_flags(slot);
2018 io_req_set_rsrc_node(req, ctx, 0);
2020 io_ring_submit_unlock(ctx, issue_flags);
2024 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
2026 struct file *file = fget(fd);
2028 trace_io_uring_file_get(req, fd);
2030 /* we don't allow fixed io_uring files */
2031 if (file && io_is_uring_fops(file))
2032 io_req_track_inflight(req);
2036 static void io_queue_async(struct io_kiocb *req, int ret)
2037 __must_hold(&req->ctx->uring_lock)
2039 struct io_kiocb *linked_timeout;
2041 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
2042 io_req_defer_failed(req, ret);
2046 linked_timeout = io_prep_linked_timeout(req);
2048 switch (io_arm_poll_handler(req, 0)) {
2049 case IO_APOLL_READY:
2050 io_kbuf_recycle(req, 0);
2051 io_req_task_queue(req);
2053 case IO_APOLL_ABORTED:
2054 io_kbuf_recycle(req, 0);
2055 io_queue_iowq(req, NULL);
2062 io_queue_linked_timeout(linked_timeout);
2065 static inline void io_queue_sqe(struct io_kiocb *req)
2066 __must_hold(&req->ctx->uring_lock)
2070 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
2073 * We async punt it if the file wasn't marked NOWAIT, or if the file
2074 * doesn't support non-blocking read/write attempts
2077 io_arm_ltimeout(req);
2079 io_queue_async(req, ret);
2082 static void io_queue_sqe_fallback(struct io_kiocb *req)
2083 __must_hold(&req->ctx->uring_lock)
2085 if (unlikely(req->flags & REQ_F_FAIL)) {
2087 * We don't submit, fail them all, for that replace hardlinks
2088 * with normal links. Extra REQ_F_LINK is tolerated.
2090 req->flags &= ~REQ_F_HARDLINK;
2091 req->flags |= REQ_F_LINK;
2092 io_req_defer_failed(req, req->cqe.res);
2094 int ret = io_req_prep_async(req);
2096 if (unlikely(ret)) {
2097 io_req_defer_failed(req, ret);
2101 if (unlikely(req->ctx->drain_active))
2104 io_queue_iowq(req, NULL);
2109 * Check SQE restrictions (opcode and flags).
2111 * Returns 'true' if SQE is allowed, 'false' otherwise.
2113 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
2114 struct io_kiocb *req,
2115 unsigned int sqe_flags)
2117 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
2120 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
2121 ctx->restrictions.sqe_flags_required)
2124 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
2125 ctx->restrictions.sqe_flags_required))
2131 static void io_init_req_drain(struct io_kiocb *req)
2133 struct io_ring_ctx *ctx = req->ctx;
2134 struct io_kiocb *head = ctx->submit_state.link.head;
2136 ctx->drain_active = true;
2139 * If we need to drain a request in the middle of a link, drain
2140 * the head request and the next request/link after the current
2141 * link. Considering sequential execution of links,
2142 * REQ_F_IO_DRAIN will be maintained for every request of our
2145 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2146 ctx->drain_next = true;
2150 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
2151 const struct io_uring_sqe *sqe)
2152 __must_hold(&ctx->uring_lock)
2154 const struct io_issue_def *def;
2155 unsigned int sqe_flags;
2159 /* req is partially pre-initialised, see io_preinit_req() */
2160 req->opcode = opcode = READ_ONCE(sqe->opcode);
2161 /* same numerical values with corresponding REQ_F_*, safe to copy */
2162 req->flags = sqe_flags = READ_ONCE(sqe->flags);
2163 req->cqe.user_data = READ_ONCE(sqe->user_data);
2165 req->rsrc_node = NULL;
2166 req->task = current;
2168 if (unlikely(opcode >= IORING_OP_LAST)) {
2172 def = &io_issue_defs[opcode];
2173 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2174 /* enforce forwards compatibility on users */
2175 if (sqe_flags & ~SQE_VALID_FLAGS)
2177 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2178 if (!def->buffer_select)
2180 req->buf_index = READ_ONCE(sqe->buf_group);
2182 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2183 ctx->drain_disabled = true;
2184 if (sqe_flags & IOSQE_IO_DRAIN) {
2185 if (ctx->drain_disabled)
2187 io_init_req_drain(req);
2190 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2191 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2193 /* knock it to the slow queue path, will be drained there */
2194 if (ctx->drain_active)
2195 req->flags |= REQ_F_FORCE_ASYNC;
2196 /* if there is no link, we're at "next" request and need to drain */
2197 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2198 ctx->drain_next = false;
2199 ctx->drain_active = true;
2200 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2204 if (!def->ioprio && sqe->ioprio)
2206 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2209 if (def->needs_file) {
2210 struct io_submit_state *state = &ctx->submit_state;
2212 req->cqe.fd = READ_ONCE(sqe->fd);
2215 * Plug now if we have more than 2 IO left after this, and the
2216 * target is potentially a read/write to block based storage.
2218 if (state->need_plug && def->plug) {
2219 state->plug_started = true;
2220 state->need_plug = false;
2221 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2225 personality = READ_ONCE(sqe->personality);
2229 req->creds = xa_load(&ctx->personalities, personality);
2232 get_cred(req->creds);
2233 ret = security_uring_override_creds(req->creds);
2235 put_cred(req->creds);
2238 req->flags |= REQ_F_CREDS;
2241 return def->prep(req, sqe);
2244 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2245 struct io_kiocb *req, int ret)
2247 struct io_ring_ctx *ctx = req->ctx;
2248 struct io_submit_link *link = &ctx->submit_state.link;
2249 struct io_kiocb *head = link->head;
2251 trace_io_uring_req_failed(sqe, req, ret);
2254 * Avoid breaking links in the middle as it renders links with SQPOLL
2255 * unusable. Instead of failing eagerly, continue assembling the link if
2256 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2257 * should find the flag and handle the rest.
2259 req_fail_link_node(req, ret);
2260 if (head && !(head->flags & REQ_F_FAIL))
2261 req_fail_link_node(head, -ECANCELED);
2263 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2265 link->last->link = req;
2269 io_queue_sqe_fallback(req);
2274 link->last->link = req;
2281 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2282 const struct io_uring_sqe *sqe)
2283 __must_hold(&ctx->uring_lock)
2285 struct io_submit_link *link = &ctx->submit_state.link;
2288 ret = io_init_req(ctx, req, sqe);
2290 return io_submit_fail_init(sqe, req, ret);
2292 trace_io_uring_submit_req(req);
2295 * If we already have a head request, queue this one for async
2296 * submittal once the head completes. If we don't have a head but
2297 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2298 * submitted sync once the chain is complete. If none of those
2299 * conditions are true (normal request), then just queue it.
2301 if (unlikely(link->head)) {
2302 ret = io_req_prep_async(req);
2304 return io_submit_fail_init(sqe, req, ret);
2306 trace_io_uring_link(req, link->head);
2307 link->last->link = req;
2310 if (req->flags & IO_REQ_LINK_FLAGS)
2312 /* last request of the link, flush it */
2315 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2318 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2319 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2320 if (req->flags & IO_REQ_LINK_FLAGS) {
2325 io_queue_sqe_fallback(req);
2335 * Batched submission is done, ensure local IO is flushed out.
2337 static void io_submit_state_end(struct io_ring_ctx *ctx)
2339 struct io_submit_state *state = &ctx->submit_state;
2341 if (unlikely(state->link.head))
2342 io_queue_sqe_fallback(state->link.head);
2343 /* flush only after queuing links as they can generate completions */
2344 io_submit_flush_completions(ctx);
2345 if (state->plug_started)
2346 blk_finish_plug(&state->plug);
2350 * Start submission side cache.
2352 static void io_submit_state_start(struct io_submit_state *state,
2353 unsigned int max_ios)
2355 state->plug_started = false;
2356 state->need_plug = max_ios > 2;
2357 state->submit_nr = max_ios;
2358 /* set only head, no need to init link_last in advance */
2359 state->link.head = NULL;
2362 static void io_commit_sqring(struct io_ring_ctx *ctx)
2364 struct io_rings *rings = ctx->rings;
2367 * Ensure any loads from the SQEs are done at this point,
2368 * since once we write the new head, the application could
2369 * write new data to them.
2371 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2375 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2376 * that is mapped by userspace. This means that care needs to be taken to
2377 * ensure that reads are stable, as we cannot rely on userspace always
2378 * being a good citizen. If members of the sqe are validated and then later
2379 * used, it's important that those reads are done through READ_ONCE() to
2380 * prevent a re-load down the line.
2382 static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe)
2384 unsigned mask = ctx->sq_entries - 1;
2385 unsigned head = ctx->cached_sq_head++ & mask;
2387 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY)) {
2388 head = READ_ONCE(ctx->sq_array[head]);
2389 if (unlikely(head >= ctx->sq_entries)) {
2390 /* drop invalid entries */
2391 spin_lock(&ctx->completion_lock);
2393 spin_unlock(&ctx->completion_lock);
2394 WRITE_ONCE(ctx->rings->sq_dropped,
2395 READ_ONCE(ctx->rings->sq_dropped) + 1);
2401 * The cached sq head (or cq tail) serves two purposes:
2403 * 1) allows us to batch the cost of updating the user visible
2405 * 2) allows the kernel side to track the head on its own, even
2406 * though the application is the one updating it.
2409 /* double index for 128-byte SQEs, twice as long */
2410 if (ctx->flags & IORING_SETUP_SQE128)
2412 *sqe = &ctx->sq_sqes[head];
2416 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2417 __must_hold(&ctx->uring_lock)
2419 unsigned int entries = io_sqring_entries(ctx);
2423 if (unlikely(!entries))
2425 /* make sure SQ entry isn't read before tail */
2426 ret = left = min(nr, entries);
2427 io_get_task_refs(left);
2428 io_submit_state_start(&ctx->submit_state, left);
2431 const struct io_uring_sqe *sqe;
2432 struct io_kiocb *req;
2434 if (unlikely(!io_alloc_req(ctx, &req)))
2436 if (unlikely(!io_get_sqe(ctx, &sqe))) {
2437 io_req_add_to_cache(req, ctx);
2442 * Continue submitting even for sqe failure if the
2443 * ring was setup with IORING_SETUP_SUBMIT_ALL
2445 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2446 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2452 if (unlikely(left)) {
2454 /* try again if it submitted nothing and can't allocate a req */
2455 if (!ret && io_req_cache_empty(ctx))
2457 current->io_uring->cached_refs += left;
2460 io_submit_state_end(ctx);
2461 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2462 io_commit_sqring(ctx);
2466 struct io_wait_queue {
2467 struct wait_queue_entry wq;
2468 struct io_ring_ctx *ctx;
2470 unsigned nr_timeouts;
2474 static inline bool io_has_work(struct io_ring_ctx *ctx)
2476 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
2477 !llist_empty(&ctx->work_llist);
2480 static inline bool io_should_wake(struct io_wait_queue *iowq)
2482 struct io_ring_ctx *ctx = iowq->ctx;
2483 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
2486 * Wake up if we have enough events, or if a timeout occurred since we
2487 * started waiting. For timeouts, we always want to return to userspace,
2488 * regardless of event count.
2490 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2493 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2494 int wake_flags, void *key)
2496 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq);
2499 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2500 * the task, and the next invocation will do it.
2502 if (io_should_wake(iowq) || io_has_work(iowq->ctx))
2503 return autoremove_wake_function(curr, mode, wake_flags, key);
2507 int io_run_task_work_sig(struct io_ring_ctx *ctx)
2509 if (!llist_empty(&ctx->work_llist)) {
2510 __set_current_state(TASK_RUNNING);
2511 if (io_run_local_work(ctx) > 0)
2514 if (io_run_task_work() > 0)
2516 if (task_sigpending(current))
2521 static bool current_pending_io(void)
2523 struct io_uring_task *tctx = current->io_uring;
2527 return percpu_counter_read_positive(&tctx->inflight);
2530 /* when returns >0, the caller should retry */
2531 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2532 struct io_wait_queue *iowq)
2536 if (unlikely(READ_ONCE(ctx->check_cq)))
2538 if (unlikely(!llist_empty(&ctx->work_llist)))
2540 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL)))
2542 if (unlikely(task_sigpending(current)))
2544 if (unlikely(io_should_wake(iowq)))
2548 * Mark us as being in io_wait if we have pending requests, so cpufreq
2549 * can take into account that the task is waiting for IO - turns out
2550 * to be important for low QD IO.
2552 io_wait = current->in_iowait;
2553 if (current_pending_io())
2554 current->in_iowait = 1;
2556 if (iowq->timeout == KTIME_MAX)
2558 else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS))
2560 current->in_iowait = io_wait;
2565 * Wait until events become available, if we don't already have some. The
2566 * application must reap them itself, as they reside on the shared cq ring.
2568 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2569 const sigset_t __user *sig, size_t sigsz,
2570 struct __kernel_timespec __user *uts)
2572 struct io_wait_queue iowq;
2573 struct io_rings *rings = ctx->rings;
2576 if (!io_allowed_run_tw(ctx))
2578 if (!llist_empty(&ctx->work_llist))
2579 io_run_local_work(ctx);
2581 io_cqring_overflow_flush(ctx);
2582 /* if user messes with these they will just get an early return */
2583 if (__io_cqring_events_user(ctx) >= min_events)
2587 #ifdef CONFIG_COMPAT
2588 if (in_compat_syscall())
2589 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2593 ret = set_user_sigmask(sig, sigsz);
2599 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2600 iowq.wq.private = current;
2601 INIT_LIST_HEAD(&iowq.wq.entry);
2603 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2604 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2605 iowq.timeout = KTIME_MAX;
2608 struct timespec64 ts;
2610 if (get_timespec64(&ts, uts))
2612 iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2615 trace_io_uring_cqring_wait(ctx, min_events);
2617 unsigned long check_cq;
2619 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
2620 int nr_wait = (int) iowq.cq_tail - READ_ONCE(ctx->rings->cq.tail);
2622 atomic_set(&ctx->cq_wait_nr, nr_wait);
2623 set_current_state(TASK_INTERRUPTIBLE);
2625 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2626 TASK_INTERRUPTIBLE);
2629 ret = io_cqring_wait_schedule(ctx, &iowq);
2630 __set_current_state(TASK_RUNNING);
2631 atomic_set(&ctx->cq_wait_nr, 0);
2636 * Run task_work after scheduling and before io_should_wake().
2637 * If we got woken because of task_work being processed, run it
2638 * now rather than let the caller do another wait loop.
2641 if (!llist_empty(&ctx->work_llist))
2642 io_run_local_work(ctx);
2644 check_cq = READ_ONCE(ctx->check_cq);
2645 if (unlikely(check_cq)) {
2646 /* let the caller flush overflows, retry */
2647 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2648 io_cqring_do_overflow_flush(ctx);
2649 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) {
2655 if (io_should_wake(&iowq)) {
2662 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
2663 finish_wait(&ctx->cq_wait, &iowq.wq);
2664 restore_saved_sigmask_unless(ret == -EINTR);
2666 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2669 void io_mem_free(void *ptr)
2674 folio_put(virt_to_folio(ptr));
2677 static void io_pages_free(struct page ***pages, int npages)
2679 struct page **page_array;
2685 page_array = *pages;
2689 for (i = 0; i < npages; i++)
2690 unpin_user_page(page_array[i]);
2695 static void *__io_uaddr_map(struct page ***pages, unsigned short *npages,
2696 unsigned long uaddr, size_t size)
2698 struct page **page_array;
2699 unsigned int nr_pages;
2705 if (uaddr & (PAGE_SIZE - 1) || !size)
2706 return ERR_PTR(-EINVAL);
2708 nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
2709 if (nr_pages > USHRT_MAX)
2710 return ERR_PTR(-EINVAL);
2711 page_array = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
2713 return ERR_PTR(-ENOMEM);
2715 ret = pin_user_pages_fast(uaddr, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
2717 if (ret != nr_pages) {
2719 io_pages_free(&page_array, ret > 0 ? ret : 0);
2720 return ret < 0 ? ERR_PTR(ret) : ERR_PTR(-EFAULT);
2723 page_addr = page_address(page_array[0]);
2724 for (i = 0; i < nr_pages; i++) {
2728 * Can't support mapping user allocated ring memory on 32-bit
2729 * archs where it could potentially reside in highmem. Just
2730 * fail those with -EINVAL, just like we did on kernels that
2731 * didn't support this feature.
2733 if (PageHighMem(page_array[i]))
2737 * No support for discontig pages for now, should either be a
2738 * single normal page, or a huge page. Later on we can add
2739 * support for remapping discontig pages, for now we will
2740 * just fail them with EINVAL.
2742 if (page_address(page_array[i]) != page_addr)
2744 page_addr += PAGE_SIZE;
2747 *pages = page_array;
2749 return page_to_virt(page_array[0]);
2752 static void *io_rings_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2755 return __io_uaddr_map(&ctx->ring_pages, &ctx->n_ring_pages, uaddr,
2759 static void *io_sqes_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2762 return __io_uaddr_map(&ctx->sqe_pages, &ctx->n_sqe_pages, uaddr,
2766 static void io_rings_free(struct io_ring_ctx *ctx)
2768 if (!(ctx->flags & IORING_SETUP_NO_MMAP)) {
2769 io_mem_free(ctx->rings);
2770 io_mem_free(ctx->sq_sqes);
2772 ctx->sq_sqes = NULL;
2774 io_pages_free(&ctx->ring_pages, ctx->n_ring_pages);
2775 ctx->n_ring_pages = 0;
2776 io_pages_free(&ctx->sqe_pages, ctx->n_sqe_pages);
2777 ctx->n_sqe_pages = 0;
2781 void *io_mem_alloc(size_t size)
2783 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2786 ret = (void *) __get_free_pages(gfp, get_order(size));
2789 return ERR_PTR(-ENOMEM);
2792 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2793 unsigned int cq_entries, size_t *sq_offset)
2795 struct io_rings *rings;
2796 size_t off, sq_array_size;
2798 off = struct_size(rings, cqes, cq_entries);
2799 if (off == SIZE_MAX)
2801 if (ctx->flags & IORING_SETUP_CQE32) {
2802 if (check_shl_overflow(off, 1, &off))
2807 off = ALIGN(off, SMP_CACHE_BYTES);
2812 if (ctx->flags & IORING_SETUP_NO_SQARRAY) {
2814 *sq_offset = SIZE_MAX;
2821 sq_array_size = array_size(sizeof(u32), sq_entries);
2822 if (sq_array_size == SIZE_MAX)
2825 if (check_add_overflow(off, sq_array_size, &off))
2831 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2832 unsigned int eventfd_async)
2834 struct io_ev_fd *ev_fd;
2835 __s32 __user *fds = arg;
2838 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2839 lockdep_is_held(&ctx->uring_lock));
2843 if (copy_from_user(&fd, fds, sizeof(*fds)))
2846 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2850 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2851 if (IS_ERR(ev_fd->cq_ev_fd)) {
2852 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2857 spin_lock(&ctx->completion_lock);
2858 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2859 spin_unlock(&ctx->completion_lock);
2861 ev_fd->eventfd_async = eventfd_async;
2862 ctx->has_evfd = true;
2863 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2864 atomic_set(&ev_fd->refs, 1);
2865 atomic_set(&ev_fd->ops, 0);
2869 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2871 struct io_ev_fd *ev_fd;
2873 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2874 lockdep_is_held(&ctx->uring_lock));
2876 ctx->has_evfd = false;
2877 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2878 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops))
2879 call_rcu(&ev_fd->rcu, io_eventfd_ops);
2886 static void io_req_caches_free(struct io_ring_ctx *ctx)
2888 struct io_kiocb *req;
2891 mutex_lock(&ctx->uring_lock);
2892 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
2894 while (!io_req_cache_empty(ctx)) {
2895 req = io_extract_req(ctx);
2896 kmem_cache_free(req_cachep, req);
2900 percpu_ref_put_many(&ctx->refs, nr);
2901 mutex_unlock(&ctx->uring_lock);
2904 static void io_rsrc_node_cache_free(struct io_cache_entry *entry)
2906 kfree(container_of(entry, struct io_rsrc_node, cache));
2909 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2911 io_sq_thread_finish(ctx);
2912 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2913 if (WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)))
2916 mutex_lock(&ctx->uring_lock);
2918 __io_sqe_buffers_unregister(ctx);
2920 __io_sqe_files_unregister(ctx);
2921 io_cqring_overflow_kill(ctx);
2922 io_eventfd_unregister(ctx);
2923 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2924 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2925 io_destroy_buffers(ctx);
2926 mutex_unlock(&ctx->uring_lock);
2928 put_cred(ctx->sq_creds);
2929 if (ctx->submitter_task)
2930 put_task_struct(ctx->submitter_task);
2932 /* there are no registered resources left, nobody uses it */
2934 io_rsrc_node_destroy(ctx, ctx->rsrc_node);
2936 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2938 #if defined(CONFIG_UNIX)
2939 if (ctx->ring_sock) {
2940 ctx->ring_sock->file = NULL; /* so that iput() is called */
2941 sock_release(ctx->ring_sock);
2944 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2946 io_alloc_cache_free(&ctx->rsrc_node_cache, io_rsrc_node_cache_free);
2947 if (ctx->mm_account) {
2948 mmdrop(ctx->mm_account);
2949 ctx->mm_account = NULL;
2952 io_kbuf_mmap_list_free(ctx);
2954 percpu_ref_exit(&ctx->refs);
2955 free_uid(ctx->user);
2956 io_req_caches_free(ctx);
2958 io_wq_put_hash(ctx->hash_map);
2959 kfree(ctx->cancel_table.hbs);
2960 kfree(ctx->cancel_table_locked.hbs);
2962 xa_destroy(&ctx->io_bl_xa);
2966 static __cold void io_activate_pollwq_cb(struct callback_head *cb)
2968 struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx,
2971 mutex_lock(&ctx->uring_lock);
2972 ctx->poll_activated = true;
2973 mutex_unlock(&ctx->uring_lock);
2976 * Wake ups for some events between start of polling and activation
2977 * might've been lost due to loose synchronisation.
2979 wake_up_all(&ctx->poll_wq);
2980 percpu_ref_put(&ctx->refs);
2983 static __cold void io_activate_pollwq(struct io_ring_ctx *ctx)
2985 spin_lock(&ctx->completion_lock);
2986 /* already activated or in progress */
2987 if (ctx->poll_activated || ctx->poll_wq_task_work.func)
2989 if (WARN_ON_ONCE(!ctx->task_complete))
2991 if (!ctx->submitter_task)
2994 * with ->submitter_task only the submitter task completes requests, we
2995 * only need to sync with it, which is done by injecting a tw
2997 init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb);
2998 percpu_ref_get(&ctx->refs);
2999 if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL))
3000 percpu_ref_put(&ctx->refs);
3002 spin_unlock(&ctx->completion_lock);
3005 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
3007 struct io_ring_ctx *ctx = file->private_data;
3010 if (unlikely(!ctx->poll_activated))
3011 io_activate_pollwq(ctx);
3013 poll_wait(file, &ctx->poll_wq, wait);
3015 * synchronizes with barrier from wq_has_sleeper call in
3019 if (!io_sqring_full(ctx))
3020 mask |= EPOLLOUT | EPOLLWRNORM;
3023 * Don't flush cqring overflow list here, just do a simple check.
3024 * Otherwise there could possible be ABBA deadlock:
3027 * lock(&ctx->uring_lock);
3029 * lock(&ctx->uring_lock);
3032 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
3033 * pushes them to do the flush.
3036 if (__io_cqring_events_user(ctx) || io_has_work(ctx))
3037 mask |= EPOLLIN | EPOLLRDNORM;
3042 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
3044 const struct cred *creds;
3046 creds = xa_erase(&ctx->personalities, id);
3055 struct io_tctx_exit {
3056 struct callback_head task_work;
3057 struct completion completion;
3058 struct io_ring_ctx *ctx;
3061 static __cold void io_tctx_exit_cb(struct callback_head *cb)
3063 struct io_uring_task *tctx = current->io_uring;
3064 struct io_tctx_exit *work;
3066 work = container_of(cb, struct io_tctx_exit, task_work);
3068 * When @in_cancel, we're in cancellation and it's racy to remove the
3069 * node. It'll be removed by the end of cancellation, just ignore it.
3070 * tctx can be NULL if the queueing of this task_work raced with
3071 * work cancelation off the exec path.
3073 if (tctx && !atomic_read(&tctx->in_cancel))
3074 io_uring_del_tctx_node((unsigned long)work->ctx);
3075 complete(&work->completion);
3078 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
3080 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3082 return req->ctx == data;
3085 static __cold void io_ring_exit_work(struct work_struct *work)
3087 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
3088 unsigned long timeout = jiffies + HZ * 60 * 5;
3089 unsigned long interval = HZ / 20;
3090 struct io_tctx_exit exit;
3091 struct io_tctx_node *node;
3095 * If we're doing polled IO and end up having requests being
3096 * submitted async (out-of-line), then completions can come in while
3097 * we're waiting for refs to drop. We need to reap these manually,
3098 * as nobody else will be looking for them.
3101 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
3102 mutex_lock(&ctx->uring_lock);
3103 io_cqring_overflow_kill(ctx);
3104 mutex_unlock(&ctx->uring_lock);
3107 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3108 io_move_task_work_from_local(ctx);
3110 while (io_uring_try_cancel_requests(ctx, NULL, true))
3114 struct io_sq_data *sqd = ctx->sq_data;
3115 struct task_struct *tsk;
3117 io_sq_thread_park(sqd);
3119 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
3120 io_wq_cancel_cb(tsk->io_uring->io_wq,
3121 io_cancel_ctx_cb, ctx, true);
3122 io_sq_thread_unpark(sqd);
3125 io_req_caches_free(ctx);
3127 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
3128 /* there is little hope left, don't run it too often */
3132 * This is really an uninterruptible wait, as it has to be
3133 * complete. But it's also run from a kworker, which doesn't
3134 * take signals, so it's fine to make it interruptible. This
3135 * avoids scenarios where we knowingly can wait much longer
3136 * on completions, for example if someone does a SIGSTOP on
3137 * a task that needs to finish task_work to make this loop
3138 * complete. That's a synthetic situation that should not
3139 * cause a stuck task backtrace, and hence a potential panic
3140 * on stuck tasks if that is enabled.
3142 } while (!wait_for_completion_interruptible_timeout(&ctx->ref_comp, interval));
3144 init_completion(&exit.completion);
3145 init_task_work(&exit.task_work, io_tctx_exit_cb);
3148 mutex_lock(&ctx->uring_lock);
3149 while (!list_empty(&ctx->tctx_list)) {
3150 WARN_ON_ONCE(time_after(jiffies, timeout));
3152 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
3154 /* don't spin on a single task if cancellation failed */
3155 list_rotate_left(&ctx->tctx_list);
3156 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
3157 if (WARN_ON_ONCE(ret))
3160 mutex_unlock(&ctx->uring_lock);
3162 * See comment above for
3163 * wait_for_completion_interruptible_timeout() on why this
3164 * wait is marked as interruptible.
3166 wait_for_completion_interruptible(&exit.completion);
3167 mutex_lock(&ctx->uring_lock);
3169 mutex_unlock(&ctx->uring_lock);
3170 spin_lock(&ctx->completion_lock);
3171 spin_unlock(&ctx->completion_lock);
3173 /* pairs with RCU read section in io_req_local_work_add() */
3174 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3177 io_ring_ctx_free(ctx);
3180 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
3182 unsigned long index;
3183 struct creds *creds;
3185 mutex_lock(&ctx->uring_lock);
3186 percpu_ref_kill(&ctx->refs);
3187 xa_for_each(&ctx->personalities, index, creds)
3188 io_unregister_personality(ctx, index);
3190 io_poll_remove_all(ctx, NULL, true);
3191 mutex_unlock(&ctx->uring_lock);
3194 * If we failed setting up the ctx, we might not have any rings
3195 * and therefore did not submit any requests
3198 io_kill_timeouts(ctx, NULL, true);
3200 flush_delayed_work(&ctx->fallback_work);
3202 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
3204 * Use system_unbound_wq to avoid spawning tons of event kworkers
3205 * if we're exiting a ton of rings at the same time. It just adds
3206 * noise and overhead, there's no discernable change in runtime
3207 * over using system_wq.
3209 queue_work(system_unbound_wq, &ctx->exit_work);
3212 static int io_uring_release(struct inode *inode, struct file *file)
3214 struct io_ring_ctx *ctx = file->private_data;
3216 file->private_data = NULL;
3217 io_ring_ctx_wait_and_kill(ctx);
3221 struct io_task_cancel {
3222 struct task_struct *task;
3226 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
3228 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3229 struct io_task_cancel *cancel = data;
3231 return io_match_task_safe(req, cancel->task, cancel->all);
3234 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
3235 struct task_struct *task,
3238 struct io_defer_entry *de;
3241 spin_lock(&ctx->completion_lock);
3242 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
3243 if (io_match_task_safe(de->req, task, cancel_all)) {
3244 list_cut_position(&list, &ctx->defer_list, &de->list);
3248 spin_unlock(&ctx->completion_lock);
3249 if (list_empty(&list))
3252 while (!list_empty(&list)) {
3253 de = list_first_entry(&list, struct io_defer_entry, list);
3254 list_del_init(&de->list);
3255 io_req_task_queue_fail(de->req, -ECANCELED);
3261 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
3263 struct io_tctx_node *node;
3264 enum io_wq_cancel cret;
3267 mutex_lock(&ctx->uring_lock);
3268 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3269 struct io_uring_task *tctx = node->task->io_uring;
3272 * io_wq will stay alive while we hold uring_lock, because it's
3273 * killed after ctx nodes, which requires to take the lock.
3275 if (!tctx || !tctx->io_wq)
3277 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
3278 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3280 mutex_unlock(&ctx->uring_lock);
3285 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
3286 struct task_struct *task,
3289 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
3290 struct io_uring_task *tctx = task ? task->io_uring : NULL;
3291 enum io_wq_cancel cret;
3294 /* set it so io_req_local_work_add() would wake us up */
3295 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
3296 atomic_set(&ctx->cq_wait_nr, 1);
3300 /* failed during ring init, it couldn't have issued any requests */
3305 ret |= io_uring_try_cancel_iowq(ctx);
3306 } else if (tctx && tctx->io_wq) {
3308 * Cancels requests of all rings, not only @ctx, but
3309 * it's fine as the task is in exit/exec.
3311 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
3313 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3316 /* SQPOLL thread does its own polling */
3317 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
3318 (ctx->sq_data && ctx->sq_data->thread == current)) {
3319 while (!wq_list_empty(&ctx->iopoll_list)) {
3320 io_iopoll_try_reap_events(ctx);
3326 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3327 io_allowed_defer_tw_run(ctx))
3328 ret |= io_run_local_work(ctx) > 0;
3329 ret |= io_cancel_defer_files(ctx, task, cancel_all);
3330 mutex_lock(&ctx->uring_lock);
3331 ret |= io_poll_remove_all(ctx, task, cancel_all);
3332 mutex_unlock(&ctx->uring_lock);
3333 ret |= io_kill_timeouts(ctx, task, cancel_all);
3335 ret |= io_run_task_work() > 0;
3339 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
3342 return atomic_read(&tctx->inflight_tracked);
3343 return percpu_counter_sum(&tctx->inflight);
3347 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3348 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3350 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
3352 struct io_uring_task *tctx = current->io_uring;
3353 struct io_ring_ctx *ctx;
3354 struct io_tctx_node *node;
3355 unsigned long index;
3359 WARN_ON_ONCE(sqd && sqd->thread != current);
3361 if (!current->io_uring)
3364 io_wq_exit_start(tctx->io_wq);
3366 atomic_inc(&tctx->in_cancel);
3370 io_uring_drop_tctx_refs(current);
3371 /* read completions before cancelations */
3372 inflight = tctx_inflight(tctx, !cancel_all);
3377 xa_for_each(&tctx->xa, index, node) {
3378 /* sqpoll task will cancel all its requests */
3379 if (node->ctx->sq_data)
3381 loop |= io_uring_try_cancel_requests(node->ctx,
3382 current, cancel_all);
3385 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3386 loop |= io_uring_try_cancel_requests(ctx,
3396 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3398 io_uring_drop_tctx_refs(current);
3399 xa_for_each(&tctx->xa, index, node) {
3400 if (!llist_empty(&node->ctx->work_llist)) {
3401 WARN_ON_ONCE(node->ctx->submitter_task &&
3402 node->ctx->submitter_task != current);
3407 * If we've seen completions, retry without waiting. This
3408 * avoids a race where a completion comes in before we did
3409 * prepare_to_wait().
3411 if (inflight == tctx_inflight(tctx, !cancel_all))
3414 finish_wait(&tctx->wait, &wait);
3417 io_uring_clean_tctx(tctx);
3420 * We shouldn't run task_works after cancel, so just leave
3421 * ->in_cancel set for normal exit.
3423 atomic_dec(&tctx->in_cancel);
3424 /* for exec all current's requests should be gone, kill tctx */
3425 __io_uring_free(current);
3429 void __io_uring_cancel(bool cancel_all)
3431 io_uring_cancel_generic(cancel_all, NULL);
3434 static void *io_uring_validate_mmap_request(struct file *file,
3435 loff_t pgoff, size_t sz)
3437 struct io_ring_ctx *ctx = file->private_data;
3438 loff_t offset = pgoff << PAGE_SHIFT;
3442 switch (offset & IORING_OFF_MMAP_MASK) {
3443 case IORING_OFF_SQ_RING:
3444 case IORING_OFF_CQ_RING:
3445 /* Don't allow mmap if the ring was setup without it */
3446 if (ctx->flags & IORING_SETUP_NO_MMAP)
3447 return ERR_PTR(-EINVAL);
3450 case IORING_OFF_SQES:
3451 /* Don't allow mmap if the ring was setup without it */
3452 if (ctx->flags & IORING_SETUP_NO_MMAP)
3453 return ERR_PTR(-EINVAL);
3456 case IORING_OFF_PBUF_RING: {
3459 bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT;
3461 ptr = io_pbuf_get_address(ctx, bgid);
3464 return ERR_PTR(-EINVAL);
3468 return ERR_PTR(-EINVAL);
3471 page = virt_to_head_page(ptr);
3472 if (sz > page_size(page))
3473 return ERR_PTR(-EINVAL);
3480 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3482 size_t sz = vma->vm_end - vma->vm_start;
3486 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3488 return PTR_ERR(ptr);
3490 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3491 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3494 static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp,
3495 unsigned long addr, unsigned long len,
3496 unsigned long pgoff, unsigned long flags)
3501 * Do not allow to map to user-provided address to avoid breaking the
3502 * aliasing rules. Userspace is not able to guess the offset address of
3503 * kernel kmalloc()ed memory area.
3508 ptr = io_uring_validate_mmap_request(filp, pgoff, len);
3513 * Some architectures have strong cache aliasing requirements.
3514 * For such architectures we need a coherent mapping which aliases
3515 * kernel memory *and* userspace memory. To achieve that:
3516 * - use a NULL file pointer to reference physical memory, and
3517 * - use the kernel virtual address of the shared io_uring context
3518 * (instead of the userspace-provided address, which has to be 0UL
3520 * - use the same pgoff which the get_unmapped_area() uses to
3521 * calculate the page colouring.
3522 * For architectures without such aliasing requirements, the
3523 * architecture will return any suitable mapping because addr is 0.
3526 flags |= MAP_SHARED;
3527 pgoff = 0; /* has been translated to ptr above */
3529 addr = (uintptr_t) ptr;
3530 pgoff = addr >> PAGE_SHIFT;
3534 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
3537 #else /* !CONFIG_MMU */
3539 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3541 return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL;
3544 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3546 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3549 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3550 unsigned long addr, unsigned long len,
3551 unsigned long pgoff, unsigned long flags)
3555 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3557 return PTR_ERR(ptr);
3559 return (unsigned long) ptr;
3562 #endif /* !CONFIG_MMU */
3564 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3566 if (flags & IORING_ENTER_EXT_ARG) {
3567 struct io_uring_getevents_arg arg;
3569 if (argsz != sizeof(arg))
3571 if (copy_from_user(&arg, argp, sizeof(arg)))
3577 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3578 struct __kernel_timespec __user **ts,
3579 const sigset_t __user **sig)
3581 struct io_uring_getevents_arg arg;
3584 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3585 * is just a pointer to the sigset_t.
3587 if (!(flags & IORING_ENTER_EXT_ARG)) {
3588 *sig = (const sigset_t __user *) argp;
3594 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3595 * timespec and sigset_t pointers if good.
3597 if (*argsz != sizeof(arg))
3599 if (copy_from_user(&arg, argp, sizeof(arg)))
3603 *sig = u64_to_user_ptr(arg.sigmask);
3604 *argsz = arg.sigmask_sz;
3605 *ts = u64_to_user_ptr(arg.ts);
3609 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3610 u32, min_complete, u32, flags, const void __user *, argp,
3613 struct io_ring_ctx *ctx;
3617 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3618 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3619 IORING_ENTER_REGISTERED_RING)))
3623 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3624 * need only dereference our task private array to find it.
3626 if (flags & IORING_ENTER_REGISTERED_RING) {
3627 struct io_uring_task *tctx = current->io_uring;
3629 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3631 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3632 file = tctx->registered_rings[fd];
3633 if (unlikely(!file))
3637 if (unlikely(!file))
3640 if (unlikely(!io_is_uring_fops(file)))
3644 ctx = file->private_data;
3646 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3650 * For SQ polling, the thread will do all submissions and completions.
3651 * Just return the requested submit count, and wake the thread if
3655 if (ctx->flags & IORING_SETUP_SQPOLL) {
3656 io_cqring_overflow_flush(ctx);
3658 if (unlikely(ctx->sq_data->thread == NULL)) {
3662 if (flags & IORING_ENTER_SQ_WAKEUP)
3663 wake_up(&ctx->sq_data->wait);
3664 if (flags & IORING_ENTER_SQ_WAIT)
3665 io_sqpoll_wait_sq(ctx);
3668 } else if (to_submit) {
3669 ret = io_uring_add_tctx_node(ctx);
3673 mutex_lock(&ctx->uring_lock);
3674 ret = io_submit_sqes(ctx, to_submit);
3675 if (ret != to_submit) {
3676 mutex_unlock(&ctx->uring_lock);
3679 if (flags & IORING_ENTER_GETEVENTS) {
3680 if (ctx->syscall_iopoll)
3683 * Ignore errors, we'll soon call io_cqring_wait() and
3684 * it should handle ownership problems if any.
3686 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3687 (void)io_run_local_work_locked(ctx);
3689 mutex_unlock(&ctx->uring_lock);
3692 if (flags & IORING_ENTER_GETEVENTS) {
3695 if (ctx->syscall_iopoll) {
3697 * We disallow the app entering submit/complete with
3698 * polling, but we still need to lock the ring to
3699 * prevent racing with polled issue that got punted to
3702 mutex_lock(&ctx->uring_lock);
3704 ret2 = io_validate_ext_arg(flags, argp, argsz);
3705 if (likely(!ret2)) {
3706 min_complete = min(min_complete,
3708 ret2 = io_iopoll_check(ctx, min_complete);
3710 mutex_unlock(&ctx->uring_lock);
3712 const sigset_t __user *sig;
3713 struct __kernel_timespec __user *ts;
3715 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3716 if (likely(!ret2)) {
3717 min_complete = min(min_complete,
3719 ret2 = io_cqring_wait(ctx, min_complete, sig,
3728 * EBADR indicates that one or more CQE were dropped.
3729 * Once the user has been informed we can clear the bit
3730 * as they are obviously ok with those drops.
3732 if (unlikely(ret2 == -EBADR))
3733 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3738 if (!(flags & IORING_ENTER_REGISTERED_RING))
3743 static const struct file_operations io_uring_fops = {
3744 .release = io_uring_release,
3745 .mmap = io_uring_mmap,
3747 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3748 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3750 .get_unmapped_area = io_uring_mmu_get_unmapped_area,
3752 .poll = io_uring_poll,
3753 #ifdef CONFIG_PROC_FS
3754 .show_fdinfo = io_uring_show_fdinfo,
3758 bool io_is_uring_fops(struct file *file)
3760 return file->f_op == &io_uring_fops;
3763 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3764 struct io_uring_params *p)
3766 struct io_rings *rings;
3767 size_t size, sq_array_offset;
3770 /* make sure these are sane, as we already accounted them */
3771 ctx->sq_entries = p->sq_entries;
3772 ctx->cq_entries = p->cq_entries;
3774 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3775 if (size == SIZE_MAX)
3778 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3779 rings = io_mem_alloc(size);
3781 rings = io_rings_map(ctx, p->cq_off.user_addr, size);
3784 return PTR_ERR(rings);
3787 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY))
3788 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3789 rings->sq_ring_mask = p->sq_entries - 1;
3790 rings->cq_ring_mask = p->cq_entries - 1;
3791 rings->sq_ring_entries = p->sq_entries;
3792 rings->cq_ring_entries = p->cq_entries;
3794 if (p->flags & IORING_SETUP_SQE128)
3795 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3797 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3798 if (size == SIZE_MAX) {
3803 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3804 ptr = io_mem_alloc(size);
3806 ptr = io_sqes_map(ctx, p->sq_off.user_addr, size);
3810 return PTR_ERR(ptr);
3817 static int io_uring_install_fd(struct file *file)
3821 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3824 fd_install(fd, file);
3829 * Allocate an anonymous fd, this is what constitutes the application
3830 * visible backing of an io_uring instance. The application mmaps this
3831 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3832 * we have to tie this fd to a socket for file garbage collection purposes.
3834 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3837 #if defined(CONFIG_UNIX)
3840 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3843 return ERR_PTR(ret);
3846 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3847 O_RDWR | O_CLOEXEC, NULL);
3848 #if defined(CONFIG_UNIX)
3850 sock_release(ctx->ring_sock);
3851 ctx->ring_sock = NULL;
3853 ctx->ring_sock->file = file;
3859 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3860 struct io_uring_params __user *params)
3862 struct io_ring_ctx *ctx;
3863 struct io_uring_task *tctx;
3869 if (entries > IORING_MAX_ENTRIES) {
3870 if (!(p->flags & IORING_SETUP_CLAMP))
3872 entries = IORING_MAX_ENTRIES;
3875 if ((p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
3876 && !(p->flags & IORING_SETUP_NO_MMAP))
3880 * Use twice as many entries for the CQ ring. It's possible for the
3881 * application to drive a higher depth than the size of the SQ ring,
3882 * since the sqes are only used at submission time. This allows for
3883 * some flexibility in overcommitting a bit. If the application has
3884 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3885 * of CQ ring entries manually.
3887 p->sq_entries = roundup_pow_of_two(entries);
3888 if (p->flags & IORING_SETUP_CQSIZE) {
3890 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3891 * to a power-of-two, if it isn't already. We do NOT impose
3892 * any cq vs sq ring sizing.
3896 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3897 if (!(p->flags & IORING_SETUP_CLAMP))
3899 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3901 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3902 if (p->cq_entries < p->sq_entries)
3905 p->cq_entries = 2 * p->sq_entries;
3908 ctx = io_ring_ctx_alloc(p);
3912 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3913 !(ctx->flags & IORING_SETUP_IOPOLL) &&
3914 !(ctx->flags & IORING_SETUP_SQPOLL))
3915 ctx->task_complete = true;
3917 if (ctx->task_complete || (ctx->flags & IORING_SETUP_IOPOLL))
3918 ctx->lockless_cq = true;
3921 * lazy poll_wq activation relies on ->task_complete for synchronisation
3922 * purposes, see io_activate_pollwq()
3924 if (!ctx->task_complete)
3925 ctx->poll_activated = true;
3928 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3929 * space applications don't need to do io completion events
3930 * polling again, they can rely on io_sq_thread to do polling
3931 * work, which can reduce cpu usage and uring_lock contention.
3933 if (ctx->flags & IORING_SETUP_IOPOLL &&
3934 !(ctx->flags & IORING_SETUP_SQPOLL))
3935 ctx->syscall_iopoll = 1;
3937 ctx->compat = in_compat_syscall();
3938 if (!ns_capable_noaudit(&init_user_ns, CAP_IPC_LOCK))
3939 ctx->user = get_uid(current_user());
3942 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3943 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3946 if (ctx->flags & IORING_SETUP_SQPOLL) {
3947 /* IPI related flags don't make sense with SQPOLL */
3948 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3949 IORING_SETUP_TASKRUN_FLAG |
3950 IORING_SETUP_DEFER_TASKRUN))
3952 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3953 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3954 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3956 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3957 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3959 ctx->notify_method = TWA_SIGNAL;
3963 * For DEFER_TASKRUN we require the completion task to be the same as the
3964 * submission task. This implies that there is only one submitter, so enforce
3967 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3968 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3973 * This is just grabbed for accounting purposes. When a process exits,
3974 * the mm is exited and dropped before the files, hence we need to hang
3975 * on to this mm purely for the purposes of being able to unaccount
3976 * memory (locked/pinned vm). It's not used for anything else.
3978 mmgrab(current->mm);
3979 ctx->mm_account = current->mm;
3981 ret = io_allocate_scq_urings(ctx, p);
3985 ret = io_sq_offload_create(ctx, p);
3989 ret = io_rsrc_init(ctx);
3993 p->sq_off.head = offsetof(struct io_rings, sq.head);
3994 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3995 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3996 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3997 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3998 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3999 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY))
4000 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
4001 p->sq_off.resv1 = 0;
4002 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
4003 p->sq_off.user_addr = 0;
4005 p->cq_off.head = offsetof(struct io_rings, cq.head);
4006 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
4007 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
4008 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
4009 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
4010 p->cq_off.cqes = offsetof(struct io_rings, cqes);
4011 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
4012 p->cq_off.resv1 = 0;
4013 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
4014 p->cq_off.user_addr = 0;
4016 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
4017 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
4018 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
4019 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
4020 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
4021 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
4022 IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING;
4024 if (copy_to_user(params, p, sizeof(*p))) {
4029 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
4030 && !(ctx->flags & IORING_SETUP_R_DISABLED))
4031 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
4033 file = io_uring_get_file(ctx);
4035 ret = PTR_ERR(file);
4039 ret = __io_uring_add_tctx_node(ctx);
4042 tctx = current->io_uring;
4045 * Install ring fd as the very last thing, so we don't risk someone
4046 * having closed it before we finish setup
4048 if (p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
4049 ret = io_ring_add_registered_file(tctx, file, 0, IO_RINGFD_REG_MAX);
4051 ret = io_uring_install_fd(file);
4055 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
4058 io_ring_ctx_wait_and_kill(ctx);
4066 * Sets up an aio uring context, and returns the fd. Applications asks for a
4067 * ring size, we return the actual sq/cq ring sizes (among other things) in the
4068 * params structure passed in.
4070 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
4072 struct io_uring_params p;
4075 if (copy_from_user(&p, params, sizeof(p)))
4077 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
4082 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
4083 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
4084 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
4085 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
4086 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
4087 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
4088 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN |
4089 IORING_SETUP_NO_MMAP | IORING_SETUP_REGISTERED_FD_ONLY |
4090 IORING_SETUP_NO_SQARRAY))
4093 return io_uring_create(entries, &p, params);
4096 static inline bool io_uring_allowed(void)
4098 int disabled = READ_ONCE(sysctl_io_uring_disabled);
4099 kgid_t io_uring_group;
4104 if (disabled == 0 || capable(CAP_SYS_ADMIN))
4107 io_uring_group = make_kgid(&init_user_ns, sysctl_io_uring_group);
4108 if (!gid_valid(io_uring_group))
4111 return in_group_p(io_uring_group);
4114 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
4115 struct io_uring_params __user *, params)
4117 if (!io_uring_allowed())
4120 return io_uring_setup(entries, params);
4123 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
4126 struct io_uring_probe *p;
4130 size = struct_size(p, ops, nr_args);
4131 if (size == SIZE_MAX)
4133 p = kzalloc(size, GFP_KERNEL);
4138 if (copy_from_user(p, arg, size))
4141 if (memchr_inv(p, 0, size))
4144 p->last_op = IORING_OP_LAST - 1;
4145 if (nr_args > IORING_OP_LAST)
4146 nr_args = IORING_OP_LAST;
4148 for (i = 0; i < nr_args; i++) {
4150 if (!io_issue_defs[i].not_supported)
4151 p->ops[i].flags = IO_URING_OP_SUPPORTED;
4156 if (copy_to_user(arg, p, size))
4163 static int io_register_personality(struct io_ring_ctx *ctx)
4165 const struct cred *creds;
4169 creds = get_current_cred();
4171 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
4172 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
4180 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
4181 void __user *arg, unsigned int nr_args)
4183 struct io_uring_restriction *res;
4187 /* Restrictions allowed only if rings started disabled */
4188 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
4191 /* We allow only a single restrictions registration */
4192 if (ctx->restrictions.registered)
4195 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
4198 size = array_size(nr_args, sizeof(*res));
4199 if (size == SIZE_MAX)
4202 res = memdup_user(arg, size);
4204 return PTR_ERR(res);
4208 for (i = 0; i < nr_args; i++) {
4209 switch (res[i].opcode) {
4210 case IORING_RESTRICTION_REGISTER_OP:
4211 if (res[i].register_op >= IORING_REGISTER_LAST) {
4216 __set_bit(res[i].register_op,
4217 ctx->restrictions.register_op);
4219 case IORING_RESTRICTION_SQE_OP:
4220 if (res[i].sqe_op >= IORING_OP_LAST) {
4225 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
4227 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
4228 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
4230 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
4231 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
4240 /* Reset all restrictions if an error happened */
4242 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
4244 ctx->restrictions.registered = true;
4250 static int io_register_enable_rings(struct io_ring_ctx *ctx)
4252 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
4255 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task) {
4256 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
4258 * Lazy activation attempts would fail if it was polled before
4259 * submitter_task is set.
4261 if (wq_has_sleeper(&ctx->poll_wq))
4262 io_activate_pollwq(ctx);
4265 if (ctx->restrictions.registered)
4266 ctx->restricted = 1;
4268 ctx->flags &= ~IORING_SETUP_R_DISABLED;
4269 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
4270 wake_up(&ctx->sq_data->wait);
4274 static __cold int __io_register_iowq_aff(struct io_ring_ctx *ctx,
4275 cpumask_var_t new_mask)
4279 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
4280 ret = io_wq_cpu_affinity(current->io_uring, new_mask);
4282 mutex_unlock(&ctx->uring_lock);
4283 ret = io_sqpoll_wq_cpu_affinity(ctx, new_mask);
4284 mutex_lock(&ctx->uring_lock);
4290 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
4291 void __user *arg, unsigned len)
4293 cpumask_var_t new_mask;
4296 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4299 cpumask_clear(new_mask);
4300 if (len > cpumask_size())
4301 len = cpumask_size();
4303 if (in_compat_syscall()) {
4304 ret = compat_get_bitmap(cpumask_bits(new_mask),
4305 (const compat_ulong_t __user *)arg,
4306 len * 8 /* CHAR_BIT */);
4308 ret = copy_from_user(new_mask, arg, len);
4312 free_cpumask_var(new_mask);
4316 ret = __io_register_iowq_aff(ctx, new_mask);
4317 free_cpumask_var(new_mask);
4321 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
4323 return __io_register_iowq_aff(ctx, NULL);
4326 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
4328 __must_hold(&ctx->uring_lock)
4330 struct io_tctx_node *node;
4331 struct io_uring_task *tctx = NULL;
4332 struct io_sq_data *sqd = NULL;
4336 if (copy_from_user(new_count, arg, sizeof(new_count)))
4338 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4339 if (new_count[i] > INT_MAX)
4342 if (ctx->flags & IORING_SETUP_SQPOLL) {
4346 * Observe the correct sqd->lock -> ctx->uring_lock
4347 * ordering. Fine to drop uring_lock here, we hold
4350 refcount_inc(&sqd->refs);
4351 mutex_unlock(&ctx->uring_lock);
4352 mutex_lock(&sqd->lock);
4353 mutex_lock(&ctx->uring_lock);
4355 tctx = sqd->thread->io_uring;
4358 tctx = current->io_uring;
4361 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
4363 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4365 ctx->iowq_limits[i] = new_count[i];
4366 ctx->iowq_limits_set = true;
4368 if (tctx && tctx->io_wq) {
4369 ret = io_wq_max_workers(tctx->io_wq, new_count);
4373 memset(new_count, 0, sizeof(new_count));
4377 mutex_unlock(&sqd->lock);
4378 io_put_sq_data(sqd);
4381 if (copy_to_user(arg, new_count, sizeof(new_count)))
4384 /* that's it for SQPOLL, only the SQPOLL task creates requests */
4388 /* now propagate the restriction to all registered users */
4389 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
4390 struct io_uring_task *tctx = node->task->io_uring;
4392 if (WARN_ON_ONCE(!tctx->io_wq))
4395 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4396 new_count[i] = ctx->iowq_limits[i];
4397 /* ignore errors, it always returns zero anyway */
4398 (void)io_wq_max_workers(tctx->io_wq, new_count);
4403 mutex_unlock(&sqd->lock);
4404 io_put_sq_data(sqd);
4409 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
4410 void __user *arg, unsigned nr_args)
4411 __releases(ctx->uring_lock)
4412 __acquires(ctx->uring_lock)
4417 * We don't quiesce the refs for register anymore and so it can't be
4418 * dying as we're holding a file ref here.
4420 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
4423 if (ctx->submitter_task && ctx->submitter_task != current)
4426 if (ctx->restricted) {
4427 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
4428 if (!test_bit(opcode, ctx->restrictions.register_op))
4433 case IORING_REGISTER_BUFFERS:
4437 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
4439 case IORING_UNREGISTER_BUFFERS:
4443 ret = io_sqe_buffers_unregister(ctx);
4445 case IORING_REGISTER_FILES:
4449 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
4451 case IORING_UNREGISTER_FILES:
4455 ret = io_sqe_files_unregister(ctx);
4457 case IORING_REGISTER_FILES_UPDATE:
4458 ret = io_register_files_update(ctx, arg, nr_args);
4460 case IORING_REGISTER_EVENTFD:
4464 ret = io_eventfd_register(ctx, arg, 0);
4466 case IORING_REGISTER_EVENTFD_ASYNC:
4470 ret = io_eventfd_register(ctx, arg, 1);
4472 case IORING_UNREGISTER_EVENTFD:
4476 ret = io_eventfd_unregister(ctx);
4478 case IORING_REGISTER_PROBE:
4480 if (!arg || nr_args > 256)
4482 ret = io_probe(ctx, arg, nr_args);
4484 case IORING_REGISTER_PERSONALITY:
4488 ret = io_register_personality(ctx);
4490 case IORING_UNREGISTER_PERSONALITY:
4494 ret = io_unregister_personality(ctx, nr_args);
4496 case IORING_REGISTER_ENABLE_RINGS:
4500 ret = io_register_enable_rings(ctx);
4502 case IORING_REGISTER_RESTRICTIONS:
4503 ret = io_register_restrictions(ctx, arg, nr_args);
4505 case IORING_REGISTER_FILES2:
4506 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
4508 case IORING_REGISTER_FILES_UPDATE2:
4509 ret = io_register_rsrc_update(ctx, arg, nr_args,
4512 case IORING_REGISTER_BUFFERS2:
4513 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
4515 case IORING_REGISTER_BUFFERS_UPDATE:
4516 ret = io_register_rsrc_update(ctx, arg, nr_args,
4517 IORING_RSRC_BUFFER);
4519 case IORING_REGISTER_IOWQ_AFF:
4521 if (!arg || !nr_args)
4523 ret = io_register_iowq_aff(ctx, arg, nr_args);
4525 case IORING_UNREGISTER_IOWQ_AFF:
4529 ret = io_unregister_iowq_aff(ctx);
4531 case IORING_REGISTER_IOWQ_MAX_WORKERS:
4533 if (!arg || nr_args != 2)
4535 ret = io_register_iowq_max_workers(ctx, arg);
4537 case IORING_REGISTER_RING_FDS:
4538 ret = io_ringfd_register(ctx, arg, nr_args);
4540 case IORING_UNREGISTER_RING_FDS:
4541 ret = io_ringfd_unregister(ctx, arg, nr_args);
4543 case IORING_REGISTER_PBUF_RING:
4545 if (!arg || nr_args != 1)
4547 ret = io_register_pbuf_ring(ctx, arg);
4549 case IORING_UNREGISTER_PBUF_RING:
4551 if (!arg || nr_args != 1)
4553 ret = io_unregister_pbuf_ring(ctx, arg);
4555 case IORING_REGISTER_SYNC_CANCEL:
4557 if (!arg || nr_args != 1)
4559 ret = io_sync_cancel(ctx, arg);
4561 case IORING_REGISTER_FILE_ALLOC_RANGE:
4563 if (!arg || nr_args)
4565 ret = io_register_file_alloc_range(ctx, arg);
4575 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
4576 void __user *, arg, unsigned int, nr_args)
4578 struct io_ring_ctx *ctx;
4581 bool use_registered_ring;
4583 use_registered_ring = !!(opcode & IORING_REGISTER_USE_REGISTERED_RING);
4584 opcode &= ~IORING_REGISTER_USE_REGISTERED_RING;
4586 if (opcode >= IORING_REGISTER_LAST)
4589 if (use_registered_ring) {
4591 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
4592 * need only dereference our task private array to find it.
4594 struct io_uring_task *tctx = current->io_uring;
4596 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
4598 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
4599 file = tctx->registered_rings[fd];
4600 if (unlikely(!file))
4604 if (unlikely(!file))
4607 if (!io_is_uring_fops(file))
4611 ctx = file->private_data;
4613 mutex_lock(&ctx->uring_lock);
4614 ret = __io_uring_register(ctx, opcode, arg, nr_args);
4615 mutex_unlock(&ctx->uring_lock);
4616 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
4618 if (!use_registered_ring)
4623 static int __init io_uring_init(void)
4625 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4626 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4627 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4630 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4631 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4632 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4633 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4634 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4635 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
4636 BUILD_BUG_SQE_ELEM(1, __u8, flags);
4637 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
4638 BUILD_BUG_SQE_ELEM(4, __s32, fd);
4639 BUILD_BUG_SQE_ELEM(8, __u64, off);
4640 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
4641 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
4642 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
4643 BUILD_BUG_SQE_ELEM(16, __u64, addr);
4644 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
4645 BUILD_BUG_SQE_ELEM(24, __u32, len);
4646 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
4647 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
4648 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4649 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
4650 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
4651 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
4652 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
4653 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
4654 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
4655 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
4656 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
4657 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
4658 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
4659 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
4660 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
4661 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
4662 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
4663 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
4664 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
4665 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
4666 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
4667 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
4668 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
4669 BUILD_BUG_SQE_ELEM(42, __u16, personality);
4670 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
4671 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
4672 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
4673 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
4674 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
4675 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4676 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
4678 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4679 sizeof(struct io_uring_rsrc_update));
4680 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4681 sizeof(struct io_uring_rsrc_update2));
4683 /* ->buf_index is u16 */
4684 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4685 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4686 offsetof(struct io_uring_buf_ring, tail));
4688 /* should fit into one byte */
4689 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
4690 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4691 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4693 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
4695 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4697 io_uring_optable_init();
4700 * Allow user copy in the per-command field, which starts after the
4701 * file in io_kiocb and until the opcode field. The openat2 handling
4702 * requires copying in user memory into the io_kiocb object in that
4703 * range, and HARDENED_USERCOPY will complain if we haven't
4704 * correctly annotated this range.
4706 req_cachep = kmem_cache_create_usercopy("io_kiocb",
4707 sizeof(struct io_kiocb), 0,
4708 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4709 SLAB_ACCOUNT | SLAB_TYPESAFE_BY_RCU,
4710 offsetof(struct io_kiocb, cmd.data),
4711 sizeof_field(struct io_kiocb, cmd.data), NULL);
4713 #ifdef CONFIG_SYSCTL
4714 register_sysctl_init("kernel", kernel_io_uring_disabled_table);
4719 __initcall(io_uring_init);