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>
63 #include <linux/anon_inodes.h>
64 #include <linux/sched/mm.h>
65 #include <linux/uaccess.h>
66 #include <linux/nospec.h>
67 #include <linux/highmem.h>
68 #include <linux/fsnotify.h>
69 #include <linux/fadvise.h>
70 #include <linux/task_work.h>
71 #include <linux/io_uring.h>
72 #include <linux/io_uring/cmd.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>
102 #include "alloc_cache.h"
104 #define IORING_MAX_ENTRIES 32768
105 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
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,
130 struct io_defer_entry {
131 struct list_head list;
132 struct io_kiocb *req;
136 /* requests with any of those set should undergo io_disarm_next() */
137 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
138 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
140 static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
141 struct task_struct *task,
144 static void io_queue_sqe(struct io_kiocb *req);
146 struct kmem_cache *req_cachep;
148 static int __read_mostly sysctl_io_uring_disabled;
149 static int __read_mostly sysctl_io_uring_group = -1;
152 static struct ctl_table kernel_io_uring_disabled_table[] = {
154 .procname = "io_uring_disabled",
155 .data = &sysctl_io_uring_disabled,
156 .maxlen = sizeof(sysctl_io_uring_disabled),
158 .proc_handler = proc_dointvec_minmax,
159 .extra1 = SYSCTL_ZERO,
160 .extra2 = SYSCTL_TWO,
163 .procname = "io_uring_group",
164 .data = &sysctl_io_uring_group,
165 .maxlen = sizeof(gid_t),
167 .proc_handler = proc_dointvec,
173 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
175 if (!wq_list_empty(&ctx->submit_state.compl_reqs) ||
176 ctx->submit_state.cqes_count)
177 __io_submit_flush_completions(ctx);
180 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
182 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
185 static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx)
187 return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head);
190 static bool io_match_linked(struct io_kiocb *head)
192 struct io_kiocb *req;
194 io_for_each_link(req, head) {
195 if (req->flags & REQ_F_INFLIGHT)
202 * As io_match_task() but protected against racing with linked timeouts.
203 * User must not hold timeout_lock.
205 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
210 if (task && head->task != task)
215 if (head->flags & REQ_F_LINK_TIMEOUT) {
216 struct io_ring_ctx *ctx = head->ctx;
218 /* protect against races with linked timeouts */
219 spin_lock_irq(&ctx->timeout_lock);
220 matched = io_match_linked(head);
221 spin_unlock_irq(&ctx->timeout_lock);
223 matched = io_match_linked(head);
228 static inline void req_fail_link_node(struct io_kiocb *req, int res)
231 io_req_set_res(req, res, 0);
234 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
236 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
239 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
241 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
243 complete(&ctx->ref_comp);
246 static __cold void io_fallback_req_func(struct work_struct *work)
248 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
250 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
251 struct io_kiocb *req, *tmp;
252 struct io_tw_state ts = { .locked = true, };
254 percpu_ref_get(&ctx->refs);
255 mutex_lock(&ctx->uring_lock);
256 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
257 req->io_task_work.func(req, &ts);
258 if (WARN_ON_ONCE(!ts.locked))
260 io_submit_flush_completions(ctx);
261 mutex_unlock(&ctx->uring_lock);
262 percpu_ref_put(&ctx->refs);
265 static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
267 unsigned hash_buckets = 1U << bits;
268 size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
270 table->hbs = kmalloc(hash_size, GFP_KERNEL);
274 table->hash_bits = bits;
275 init_hash_table(table, hash_buckets);
279 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
281 struct io_ring_ctx *ctx;
284 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
288 xa_init(&ctx->io_bl_xa);
291 * Use 5 bits less than the max cq entries, that should give us around
292 * 32 entries per hash list if totally full and uniformly spread, but
293 * don't keep too many buckets to not overconsume memory.
295 hash_bits = ilog2(p->cq_entries) - 5;
296 hash_bits = clamp(hash_bits, 1, 8);
297 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
299 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
301 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
305 ctx->flags = p->flags;
306 init_waitqueue_head(&ctx->sqo_sq_wait);
307 INIT_LIST_HEAD(&ctx->sqd_list);
308 INIT_LIST_HEAD(&ctx->cq_overflow_list);
309 INIT_LIST_HEAD(&ctx->io_buffers_cache);
310 INIT_HLIST_HEAD(&ctx->io_buf_list);
311 io_alloc_cache_init(&ctx->rsrc_node_cache, IO_NODE_ALLOC_CACHE_MAX,
312 sizeof(struct io_rsrc_node));
313 io_alloc_cache_init(&ctx->apoll_cache, IO_ALLOC_CACHE_MAX,
314 sizeof(struct async_poll));
315 io_alloc_cache_init(&ctx->netmsg_cache, IO_ALLOC_CACHE_MAX,
316 sizeof(struct io_async_msghdr));
317 io_futex_cache_init(ctx);
318 init_completion(&ctx->ref_comp);
319 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
320 mutex_init(&ctx->uring_lock);
321 init_waitqueue_head(&ctx->cq_wait);
322 init_waitqueue_head(&ctx->poll_wq);
323 init_waitqueue_head(&ctx->rsrc_quiesce_wq);
324 spin_lock_init(&ctx->completion_lock);
325 spin_lock_init(&ctx->timeout_lock);
326 INIT_WQ_LIST(&ctx->iopoll_list);
327 INIT_LIST_HEAD(&ctx->io_buffers_comp);
328 INIT_LIST_HEAD(&ctx->defer_list);
329 INIT_LIST_HEAD(&ctx->timeout_list);
330 INIT_LIST_HEAD(&ctx->ltimeout_list);
331 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
332 init_llist_head(&ctx->work_llist);
333 INIT_LIST_HEAD(&ctx->tctx_list);
334 ctx->submit_state.free_list.next = NULL;
335 INIT_WQ_LIST(&ctx->locked_free_list);
336 INIT_HLIST_HEAD(&ctx->waitid_list);
338 INIT_HLIST_HEAD(&ctx->futex_list);
340 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
341 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
342 INIT_HLIST_HEAD(&ctx->cancelable_uring_cmd);
345 kfree(ctx->cancel_table.hbs);
346 kfree(ctx->cancel_table_locked.hbs);
348 xa_destroy(&ctx->io_bl_xa);
353 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
355 struct io_rings *r = ctx->rings;
357 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
361 static bool req_need_defer(struct io_kiocb *req, u32 seq)
363 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
364 struct io_ring_ctx *ctx = req->ctx;
366 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
372 static void io_clean_op(struct io_kiocb *req)
374 if (req->flags & REQ_F_BUFFER_SELECTED) {
375 spin_lock(&req->ctx->completion_lock);
376 io_put_kbuf_comp(req);
377 spin_unlock(&req->ctx->completion_lock);
380 if (req->flags & REQ_F_NEED_CLEANUP) {
381 const struct io_cold_def *def = &io_cold_defs[req->opcode];
386 if ((req->flags & REQ_F_POLLED) && req->apoll) {
387 kfree(req->apoll->double_poll);
391 if (req->flags & REQ_F_INFLIGHT) {
392 struct io_uring_task *tctx = req->task->io_uring;
394 atomic_dec(&tctx->inflight_tracked);
396 if (req->flags & REQ_F_CREDS)
397 put_cred(req->creds);
398 if (req->flags & REQ_F_ASYNC_DATA) {
399 kfree(req->async_data);
400 req->async_data = NULL;
402 req->flags &= ~IO_REQ_CLEAN_FLAGS;
405 static inline void io_req_track_inflight(struct io_kiocb *req)
407 if (!(req->flags & REQ_F_INFLIGHT)) {
408 req->flags |= REQ_F_INFLIGHT;
409 atomic_inc(&req->task->io_uring->inflight_tracked);
413 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
415 if (WARN_ON_ONCE(!req->link))
418 req->flags &= ~REQ_F_ARM_LTIMEOUT;
419 req->flags |= REQ_F_LINK_TIMEOUT;
421 /* linked timeouts should have two refs once prep'ed */
422 io_req_set_refcount(req);
423 __io_req_set_refcount(req->link, 2);
427 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
429 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
431 return __io_prep_linked_timeout(req);
434 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
436 io_queue_linked_timeout(__io_prep_linked_timeout(req));
439 static inline void io_arm_ltimeout(struct io_kiocb *req)
441 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
442 __io_arm_ltimeout(req);
445 static void io_prep_async_work(struct io_kiocb *req)
447 const struct io_issue_def *def = &io_issue_defs[req->opcode];
448 struct io_ring_ctx *ctx = req->ctx;
450 if (!(req->flags & REQ_F_CREDS)) {
451 req->flags |= REQ_F_CREDS;
452 req->creds = get_current_cred();
455 req->work.list.next = NULL;
457 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
458 if (req->flags & REQ_F_FORCE_ASYNC)
459 req->work.flags |= IO_WQ_WORK_CONCURRENT;
461 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
462 req->flags |= io_file_get_flags(req->file);
464 if (req->file && (req->flags & REQ_F_ISREG)) {
465 bool should_hash = def->hash_reg_file;
467 /* don't serialize this request if the fs doesn't need it */
468 if (should_hash && (req->file->f_flags & O_DIRECT) &&
469 (req->file->f_mode & FMODE_DIO_PARALLEL_WRITE))
471 if (should_hash || (ctx->flags & IORING_SETUP_IOPOLL))
472 io_wq_hash_work(&req->work, file_inode(req->file));
473 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
474 if (def->unbound_nonreg_file)
475 req->work.flags |= IO_WQ_WORK_UNBOUND;
479 static void io_prep_async_link(struct io_kiocb *req)
481 struct io_kiocb *cur;
483 if (req->flags & REQ_F_LINK_TIMEOUT) {
484 struct io_ring_ctx *ctx = req->ctx;
486 spin_lock_irq(&ctx->timeout_lock);
487 io_for_each_link(cur, req)
488 io_prep_async_work(cur);
489 spin_unlock_irq(&ctx->timeout_lock);
491 io_for_each_link(cur, req)
492 io_prep_async_work(cur);
496 void io_queue_iowq(struct io_kiocb *req, struct io_tw_state *ts_dont_use)
498 struct io_kiocb *link = io_prep_linked_timeout(req);
499 struct io_uring_task *tctx = req->task->io_uring;
502 BUG_ON(!tctx->io_wq);
504 /* init ->work of the whole link before punting */
505 io_prep_async_link(req);
508 * Not expected to happen, but if we do have a bug where this _can_
509 * happen, catch it here and ensure the request is marked as
510 * canceled. That will make io-wq go through the usual work cancel
511 * procedure rather than attempt to run this request (or create a new
514 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
515 req->work.flags |= IO_WQ_WORK_CANCEL;
517 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
518 io_wq_enqueue(tctx->io_wq, &req->work);
520 io_queue_linked_timeout(link);
523 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
525 while (!list_empty(&ctx->defer_list)) {
526 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
527 struct io_defer_entry, list);
529 if (req_need_defer(de->req, de->seq))
531 list_del_init(&de->list);
532 io_req_task_queue(de->req);
537 void io_eventfd_ops(struct rcu_head *rcu)
539 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
540 int ops = atomic_xchg(&ev_fd->ops, 0);
542 if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT))
543 eventfd_signal_mask(ev_fd->cq_ev_fd, EPOLL_URING_WAKE);
545 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
546 * ordering in a race but if references are 0 we know we have to free
549 if (atomic_dec_and_test(&ev_fd->refs)) {
550 eventfd_ctx_put(ev_fd->cq_ev_fd);
555 static void io_eventfd_signal(struct io_ring_ctx *ctx)
557 struct io_ev_fd *ev_fd = NULL;
561 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
564 ev_fd = rcu_dereference(ctx->io_ev_fd);
567 * Check again if ev_fd exists incase an io_eventfd_unregister call
568 * completed between the NULL check of ctx->io_ev_fd at the start of
569 * the function and rcu_read_lock.
571 if (unlikely(!ev_fd))
573 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
575 if (ev_fd->eventfd_async && !io_wq_current_is_worker())
578 if (likely(eventfd_signal_allowed())) {
579 eventfd_signal_mask(ev_fd->cq_ev_fd, EPOLL_URING_WAKE);
581 atomic_inc(&ev_fd->refs);
582 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops))
583 call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops);
585 atomic_dec(&ev_fd->refs);
592 static void io_eventfd_flush_signal(struct io_ring_ctx *ctx)
596 spin_lock(&ctx->completion_lock);
599 * Eventfd should only get triggered when at least one event has been
600 * posted. Some applications rely on the eventfd notification count
601 * only changing IFF a new CQE has been added to the CQ ring. There's
602 * no depedency on 1:1 relationship between how many times this
603 * function is called (and hence the eventfd count) and number of CQEs
604 * posted to the CQ ring.
606 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
607 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
608 spin_unlock(&ctx->completion_lock);
612 io_eventfd_signal(ctx);
615 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
617 if (ctx->poll_activated)
618 io_poll_wq_wake(ctx);
619 if (ctx->off_timeout_used)
620 io_flush_timeouts(ctx);
621 if (ctx->drain_active) {
622 spin_lock(&ctx->completion_lock);
623 io_queue_deferred(ctx);
624 spin_unlock(&ctx->completion_lock);
627 io_eventfd_flush_signal(ctx);
630 static inline void __io_cq_lock(struct io_ring_ctx *ctx)
632 if (!ctx->lockless_cq)
633 spin_lock(&ctx->completion_lock);
636 static inline void io_cq_lock(struct io_ring_ctx *ctx)
637 __acquires(ctx->completion_lock)
639 spin_lock(&ctx->completion_lock);
642 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
644 io_commit_cqring(ctx);
645 if (!ctx->task_complete) {
646 if (!ctx->lockless_cq)
647 spin_unlock(&ctx->completion_lock);
648 /* IOPOLL rings only need to wake up if it's also SQPOLL */
649 if (!ctx->syscall_iopoll)
652 io_commit_cqring_flush(ctx);
655 static void io_cq_unlock_post(struct io_ring_ctx *ctx)
656 __releases(ctx->completion_lock)
658 io_commit_cqring(ctx);
659 spin_unlock(&ctx->completion_lock);
661 io_commit_cqring_flush(ctx);
664 /* Returns true if there are no backlogged entries after the flush */
665 static void io_cqring_overflow_kill(struct io_ring_ctx *ctx)
667 struct io_overflow_cqe *ocqe;
670 spin_lock(&ctx->completion_lock);
671 list_splice_init(&ctx->cq_overflow_list, &list);
672 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
673 spin_unlock(&ctx->completion_lock);
675 while (!list_empty(&list)) {
676 ocqe = list_first_entry(&list, struct io_overflow_cqe, list);
677 list_del(&ocqe->list);
682 static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx)
684 size_t cqe_size = sizeof(struct io_uring_cqe);
686 if (__io_cqring_events(ctx) == ctx->cq_entries)
689 if (ctx->flags & IORING_SETUP_CQE32)
693 while (!list_empty(&ctx->cq_overflow_list)) {
694 struct io_uring_cqe *cqe;
695 struct io_overflow_cqe *ocqe;
697 if (!io_get_cqe_overflow(ctx, &cqe, true))
699 ocqe = list_first_entry(&ctx->cq_overflow_list,
700 struct io_overflow_cqe, list);
701 memcpy(cqe, &ocqe->cqe, cqe_size);
702 list_del(&ocqe->list);
706 if (list_empty(&ctx->cq_overflow_list)) {
707 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
708 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
710 io_cq_unlock_post(ctx);
713 static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx)
715 /* iopoll syncs against uring_lock, not completion_lock */
716 if (ctx->flags & IORING_SETUP_IOPOLL)
717 mutex_lock(&ctx->uring_lock);
718 __io_cqring_overflow_flush(ctx);
719 if (ctx->flags & IORING_SETUP_IOPOLL)
720 mutex_unlock(&ctx->uring_lock);
723 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx)
725 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
726 io_cqring_do_overflow_flush(ctx);
729 /* can be called by any task */
730 static void io_put_task_remote(struct task_struct *task)
732 struct io_uring_task *tctx = task->io_uring;
734 percpu_counter_sub(&tctx->inflight, 1);
735 if (unlikely(atomic_read(&tctx->in_cancel)))
736 wake_up(&tctx->wait);
737 put_task_struct(task);
740 /* used by a task to put its own references */
741 static void io_put_task_local(struct task_struct *task)
743 task->io_uring->cached_refs++;
746 /* must to be called somewhat shortly after putting a request */
747 static inline void io_put_task(struct task_struct *task)
749 if (likely(task == current))
750 io_put_task_local(task);
752 io_put_task_remote(task);
755 void io_task_refs_refill(struct io_uring_task *tctx)
757 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
759 percpu_counter_add(&tctx->inflight, refill);
760 refcount_add(refill, ¤t->usage);
761 tctx->cached_refs += refill;
764 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
766 struct io_uring_task *tctx = task->io_uring;
767 unsigned int refs = tctx->cached_refs;
770 tctx->cached_refs = 0;
771 percpu_counter_sub(&tctx->inflight, refs);
772 put_task_struct_many(task, refs);
776 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
777 s32 res, u32 cflags, u64 extra1, u64 extra2)
779 struct io_overflow_cqe *ocqe;
780 size_t ocq_size = sizeof(struct io_overflow_cqe);
781 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
783 lockdep_assert_held(&ctx->completion_lock);
786 ocq_size += sizeof(struct io_uring_cqe);
788 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
789 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
792 * If we're in ring overflow flush mode, or in task cancel mode,
793 * or cannot allocate an overflow entry, then we need to drop it
796 io_account_cq_overflow(ctx);
797 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
800 if (list_empty(&ctx->cq_overflow_list)) {
801 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
802 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
805 ocqe->cqe.user_data = user_data;
807 ocqe->cqe.flags = cflags;
809 ocqe->cqe.big_cqe[0] = extra1;
810 ocqe->cqe.big_cqe[1] = extra2;
812 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
816 void io_req_cqe_overflow(struct io_kiocb *req)
818 io_cqring_event_overflow(req->ctx, req->cqe.user_data,
819 req->cqe.res, req->cqe.flags,
820 req->big_cqe.extra1, req->big_cqe.extra2);
821 memset(&req->big_cqe, 0, sizeof(req->big_cqe));
825 * writes to the cq entry need to come after reading head; the
826 * control dependency is enough as we're using WRITE_ONCE to
829 bool io_cqe_cache_refill(struct io_ring_ctx *ctx, bool overflow)
831 struct io_rings *rings = ctx->rings;
832 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
833 unsigned int free, queued, len;
836 * Posting into the CQ when there are pending overflowed CQEs may break
837 * ordering guarantees, which will affect links, F_MORE users and more.
838 * Force overflow the completion.
840 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
843 /* userspace may cheat modifying the tail, be safe and do min */
844 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
845 free = ctx->cq_entries - queued;
846 /* we need a contiguous range, limit based on the current array offset */
847 len = min(free, ctx->cq_entries - off);
851 if (ctx->flags & IORING_SETUP_CQE32) {
856 ctx->cqe_cached = &rings->cqes[off];
857 ctx->cqe_sentinel = ctx->cqe_cached + len;
861 static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
864 struct io_uring_cqe *cqe;
869 * If we can't get a cq entry, userspace overflowed the
870 * submission (by quite a lot). Increment the overflow count in
873 if (likely(io_get_cqe(ctx, &cqe))) {
874 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
876 WRITE_ONCE(cqe->user_data, user_data);
877 WRITE_ONCE(cqe->res, res);
878 WRITE_ONCE(cqe->flags, cflags);
880 if (ctx->flags & IORING_SETUP_CQE32) {
881 WRITE_ONCE(cqe->big_cqe[0], 0);
882 WRITE_ONCE(cqe->big_cqe[1], 0);
889 static void __io_flush_post_cqes(struct io_ring_ctx *ctx)
890 __must_hold(&ctx->uring_lock)
892 struct io_submit_state *state = &ctx->submit_state;
895 lockdep_assert_held(&ctx->uring_lock);
896 for (i = 0; i < state->cqes_count; i++) {
897 struct io_uring_cqe *cqe = &ctx->completion_cqes[i];
899 if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) {
900 if (ctx->lockless_cq) {
901 spin_lock(&ctx->completion_lock);
902 io_cqring_event_overflow(ctx, cqe->user_data,
903 cqe->res, cqe->flags, 0, 0);
904 spin_unlock(&ctx->completion_lock);
906 io_cqring_event_overflow(ctx, cqe->user_data,
907 cqe->res, cqe->flags, 0, 0);
911 state->cqes_count = 0;
914 static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
920 filled = io_fill_cqe_aux(ctx, user_data, res, cflags);
921 if (!filled && allow_overflow)
922 filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
924 io_cq_unlock_post(ctx);
928 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags)
930 return __io_post_aux_cqe(ctx, user_data, res, cflags, true);
934 * A helper for multishot requests posting additional CQEs.
935 * Should only be used from a task_work including IO_URING_F_MULTISHOT.
937 bool io_fill_cqe_req_aux(struct io_kiocb *req, bool defer, s32 res, u32 cflags)
939 struct io_ring_ctx *ctx = req->ctx;
940 u64 user_data = req->cqe.user_data;
941 struct io_uring_cqe *cqe;
944 return __io_post_aux_cqe(ctx, user_data, res, cflags, false);
946 lockdep_assert_held(&ctx->uring_lock);
948 if (ctx->submit_state.cqes_count == ARRAY_SIZE(ctx->completion_cqes)) {
950 __io_flush_post_cqes(ctx);
951 /* no need to flush - flush is deferred */
952 __io_cq_unlock_post(ctx);
955 /* For defered completions this is not as strict as it is otherwise,
956 * however it's main job is to prevent unbounded posted completions,
957 * and in that it works just as well.
959 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
962 cqe = &ctx->completion_cqes[ctx->submit_state.cqes_count++];
963 cqe->user_data = user_data;
969 static void __io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
971 struct io_ring_ctx *ctx = req->ctx;
972 struct io_rsrc_node *rsrc_node = NULL;
975 if (!(req->flags & REQ_F_CQE_SKIP)) {
976 if (!io_fill_cqe_req(ctx, req))
977 io_req_cqe_overflow(req);
981 * If we're the last reference to this request, add to our locked
984 if (req_ref_put_and_test(req)) {
985 if (req->flags & IO_REQ_LINK_FLAGS) {
986 if (req->flags & IO_DISARM_MASK)
989 io_req_task_queue(req->link);
993 io_put_kbuf_comp(req);
994 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
998 rsrc_node = req->rsrc_node;
1000 * Selected buffer deallocation in io_clean_op() assumes that
1001 * we don't hold ->completion_lock. Clean them here to avoid
1004 io_put_task_remote(req->task);
1005 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
1006 ctx->locked_free_nr++;
1008 io_cq_unlock_post(ctx);
1011 io_ring_submit_lock(ctx, issue_flags);
1012 io_put_rsrc_node(ctx, rsrc_node);
1013 io_ring_submit_unlock(ctx, issue_flags);
1017 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
1019 if (req->ctx->task_complete && req->ctx->submitter_task != current) {
1020 req->io_task_work.func = io_req_task_complete;
1021 io_req_task_work_add(req);
1022 } else if (!(issue_flags & IO_URING_F_UNLOCKED) ||
1023 !(req->ctx->flags & IORING_SETUP_IOPOLL)) {
1024 __io_req_complete_post(req, issue_flags);
1026 struct io_ring_ctx *ctx = req->ctx;
1028 mutex_lock(&ctx->uring_lock);
1029 __io_req_complete_post(req, issue_flags & ~IO_URING_F_UNLOCKED);
1030 mutex_unlock(&ctx->uring_lock);
1034 void io_req_defer_failed(struct io_kiocb *req, s32 res)
1035 __must_hold(&ctx->uring_lock)
1037 const struct io_cold_def *def = &io_cold_defs[req->opcode];
1039 lockdep_assert_held(&req->ctx->uring_lock);
1042 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
1045 io_req_complete_defer(req);
1049 * Don't initialise the fields below on every allocation, but do that in
1050 * advance and keep them valid across allocations.
1052 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1056 req->async_data = NULL;
1057 /* not necessary, but safer to zero */
1058 memset(&req->cqe, 0, sizeof(req->cqe));
1059 memset(&req->big_cqe, 0, sizeof(req->big_cqe));
1062 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1063 struct io_submit_state *state)
1065 spin_lock(&ctx->completion_lock);
1066 wq_list_splice(&ctx->locked_free_list, &state->free_list);
1067 ctx->locked_free_nr = 0;
1068 spin_unlock(&ctx->completion_lock);
1072 * A request might get retired back into the request caches even before opcode
1073 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1074 * Because of that, io_alloc_req() should be called only under ->uring_lock
1075 * and with extra caution to not get a request that is still worked on.
1077 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
1078 __must_hold(&ctx->uring_lock)
1080 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1081 void *reqs[IO_REQ_ALLOC_BATCH];
1085 * If we have more than a batch's worth of requests in our IRQ side
1086 * locked cache, grab the lock and move them over to our submission
1089 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
1090 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
1091 if (!io_req_cache_empty(ctx))
1095 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
1098 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1099 * retry single alloc to be on the safe side.
1101 if (unlikely(ret <= 0)) {
1102 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1108 percpu_ref_get_many(&ctx->refs, ret);
1109 for (i = 0; i < ret; i++) {
1110 struct io_kiocb *req = reqs[i];
1112 io_preinit_req(req, ctx);
1113 io_req_add_to_cache(req, ctx);
1118 __cold void io_free_req(struct io_kiocb *req)
1120 /* refs were already put, restore them for io_req_task_complete() */
1121 req->flags &= ~REQ_F_REFCOUNT;
1122 /* we only want to free it, don't post CQEs */
1123 req->flags |= REQ_F_CQE_SKIP;
1124 req->io_task_work.func = io_req_task_complete;
1125 io_req_task_work_add(req);
1128 static void __io_req_find_next_prep(struct io_kiocb *req)
1130 struct io_ring_ctx *ctx = req->ctx;
1132 spin_lock(&ctx->completion_lock);
1133 io_disarm_next(req);
1134 spin_unlock(&ctx->completion_lock);
1137 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1139 struct io_kiocb *nxt;
1142 * If LINK is set, we have dependent requests in this chain. If we
1143 * didn't fail this request, queue the first one up, moving any other
1144 * dependencies to the next request. In case of failure, fail the rest
1147 if (unlikely(req->flags & IO_DISARM_MASK))
1148 __io_req_find_next_prep(req);
1154 static void ctx_flush_and_put(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1158 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1159 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1161 io_submit_flush_completions(ctx);
1162 mutex_unlock(&ctx->uring_lock);
1165 percpu_ref_put(&ctx->refs);
1168 static unsigned int handle_tw_list(struct llist_node *node,
1169 struct io_ring_ctx **ctx,
1170 struct io_tw_state *ts,
1171 struct llist_node *last)
1173 unsigned int count = 0;
1175 while (node && node != last) {
1176 struct llist_node *next = node->next;
1177 struct io_kiocb *req = container_of(node, struct io_kiocb,
1180 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1182 if (req->ctx != *ctx) {
1183 ctx_flush_and_put(*ctx, ts);
1185 /* if not contended, grab and improve batching */
1186 ts->locked = mutex_trylock(&(*ctx)->uring_lock);
1187 percpu_ref_get(&(*ctx)->refs);
1189 INDIRECT_CALL_2(req->io_task_work.func,
1190 io_poll_task_func, io_req_rw_complete,
1194 if (unlikely(need_resched())) {
1195 ctx_flush_and_put(*ctx, ts);
1205 * io_llist_xchg - swap all entries in a lock-less list
1206 * @head: the head of lock-less list to delete all entries
1207 * @new: new entry as the head of the list
1209 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1210 * The order of entries returned is from the newest to the oldest added one.
1212 static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1213 struct llist_node *new)
1215 return xchg(&head->first, new);
1219 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1220 * @head: the head of lock-less list to delete all entries
1221 * @old: expected old value of the first entry of the list
1222 * @new: new entry as the head of the list
1224 * perform a cmpxchg on the first entry of the list.
1227 static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1228 struct llist_node *old,
1229 struct llist_node *new)
1231 return cmpxchg(&head->first, old, new);
1234 static __cold void io_fallback_tw(struct io_uring_task *tctx, bool sync)
1236 struct llist_node *node = llist_del_all(&tctx->task_list);
1237 struct io_ring_ctx *last_ctx = NULL;
1238 struct io_kiocb *req;
1241 req = container_of(node, struct io_kiocb, io_task_work.node);
1243 if (sync && last_ctx != req->ctx) {
1245 flush_delayed_work(&last_ctx->fallback_work);
1246 percpu_ref_put(&last_ctx->refs);
1248 last_ctx = req->ctx;
1249 percpu_ref_get(&last_ctx->refs);
1251 if (llist_add(&req->io_task_work.node,
1252 &req->ctx->fallback_llist))
1253 schedule_delayed_work(&req->ctx->fallback_work, 1);
1257 flush_delayed_work(&last_ctx->fallback_work);
1258 percpu_ref_put(&last_ctx->refs);
1262 void tctx_task_work(struct callback_head *cb)
1264 struct io_tw_state ts = {};
1265 struct io_ring_ctx *ctx = NULL;
1266 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1268 struct llist_node fake = {};
1269 struct llist_node *node;
1270 unsigned int loops = 0;
1271 unsigned int count = 0;
1273 if (unlikely(current->flags & PF_EXITING)) {
1274 io_fallback_tw(tctx, true);
1280 node = io_llist_xchg(&tctx->task_list, &fake);
1281 count += handle_tw_list(node, &ctx, &ts, &fake);
1283 /* skip expensive cmpxchg if there are items in the list */
1284 if (READ_ONCE(tctx->task_list.first) != &fake)
1286 if (ts.locked && !wq_list_empty(&ctx->submit_state.compl_reqs)) {
1287 io_submit_flush_completions(ctx);
1288 if (READ_ONCE(tctx->task_list.first) != &fake)
1291 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1292 } while (node != &fake);
1294 ctx_flush_and_put(ctx, &ts);
1296 /* relaxed read is enough as only the task itself sets ->in_cancel */
1297 if (unlikely(atomic_read(&tctx->in_cancel)))
1298 io_uring_drop_tctx_refs(current);
1300 trace_io_uring_task_work_run(tctx, count, loops);
1303 static inline void io_req_local_work_add(struct io_kiocb *req, unsigned flags)
1305 struct io_ring_ctx *ctx = req->ctx;
1306 unsigned nr_wait, nr_tw, nr_tw_prev;
1307 struct llist_node *first;
1309 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK))
1310 flags &= ~IOU_F_TWQ_LAZY_WAKE;
1312 first = READ_ONCE(ctx->work_llist.first);
1316 struct io_kiocb *first_req = container_of(first,
1320 * Might be executed at any moment, rely on
1321 * SLAB_TYPESAFE_BY_RCU to keep it alive.
1323 nr_tw_prev = READ_ONCE(first_req->nr_tw);
1325 nr_tw = nr_tw_prev + 1;
1326 /* Large enough to fail the nr_wait comparison below */
1327 if (!(flags & IOU_F_TWQ_LAZY_WAKE))
1331 req->io_task_work.node.next = first;
1332 } while (!try_cmpxchg(&ctx->work_llist.first, &first,
1333 &req->io_task_work.node));
1336 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1337 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1339 io_eventfd_signal(ctx);
1342 nr_wait = atomic_read(&ctx->cq_wait_nr);
1343 /* no one is waiting */
1346 /* either not enough or the previous add has already woken it up */
1347 if (nr_wait > nr_tw || nr_tw_prev >= nr_wait)
1349 /* pairs with set_current_state() in io_cqring_wait() */
1350 smp_mb__after_atomic();
1351 wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE);
1354 static void io_req_normal_work_add(struct io_kiocb *req)
1356 struct io_uring_task *tctx = req->task->io_uring;
1357 struct io_ring_ctx *ctx = req->ctx;
1359 /* task_work already pending, we're done */
1360 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1363 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1364 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1366 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1369 io_fallback_tw(tctx, false);
1372 void __io_req_task_work_add(struct io_kiocb *req, unsigned flags)
1374 if (req->ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
1376 io_req_local_work_add(req, flags);
1379 io_req_normal_work_add(req);
1383 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1385 struct llist_node *node;
1387 node = llist_del_all(&ctx->work_llist);
1389 struct io_kiocb *req = container_of(node, struct io_kiocb,
1393 io_req_normal_work_add(req);
1397 static int __io_run_local_work(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1399 struct llist_node *node;
1400 unsigned int loops = 0;
1403 if (WARN_ON_ONCE(ctx->submitter_task != current))
1405 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1406 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1409 * llists are in reverse order, flip it back the right way before
1410 * running the pending items.
1412 node = llist_reverse_order(io_llist_xchg(&ctx->work_llist, NULL));
1414 struct llist_node *next = node->next;
1415 struct io_kiocb *req = container_of(node, struct io_kiocb,
1417 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1418 INDIRECT_CALL_2(req->io_task_work.func,
1419 io_poll_task_func, io_req_rw_complete,
1426 if (!llist_empty(&ctx->work_llist))
1429 io_submit_flush_completions(ctx);
1430 if (!llist_empty(&ctx->work_llist))
1433 trace_io_uring_local_work_run(ctx, ret, loops);
1437 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx)
1439 struct io_tw_state ts = { .locked = true, };
1442 if (llist_empty(&ctx->work_llist))
1445 ret = __io_run_local_work(ctx, &ts);
1446 /* shouldn't happen! */
1447 if (WARN_ON_ONCE(!ts.locked))
1448 mutex_lock(&ctx->uring_lock);
1452 static int io_run_local_work(struct io_ring_ctx *ctx)
1454 struct io_tw_state ts = {};
1457 ts.locked = mutex_trylock(&ctx->uring_lock);
1458 ret = __io_run_local_work(ctx, &ts);
1460 mutex_unlock(&ctx->uring_lock);
1465 static void io_req_task_cancel(struct io_kiocb *req, struct io_tw_state *ts)
1467 io_tw_lock(req->ctx, ts);
1468 io_req_defer_failed(req, req->cqe.res);
1471 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts)
1473 io_tw_lock(req->ctx, ts);
1474 /* req->task == current here, checking PF_EXITING is safe */
1475 if (unlikely(req->task->flags & PF_EXITING))
1476 io_req_defer_failed(req, -EFAULT);
1477 else if (req->flags & REQ_F_FORCE_ASYNC)
1478 io_queue_iowq(req, ts);
1483 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1485 io_req_set_res(req, ret, 0);
1486 req->io_task_work.func = io_req_task_cancel;
1487 io_req_task_work_add(req);
1490 void io_req_task_queue(struct io_kiocb *req)
1492 req->io_task_work.func = io_req_task_submit;
1493 io_req_task_work_add(req);
1496 void io_queue_next(struct io_kiocb *req)
1498 struct io_kiocb *nxt = io_req_find_next(req);
1501 io_req_task_queue(nxt);
1504 static void io_free_batch_list(struct io_ring_ctx *ctx,
1505 struct io_wq_work_node *node)
1506 __must_hold(&ctx->uring_lock)
1509 struct io_kiocb *req = container_of(node, struct io_kiocb,
1512 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1513 if (req->flags & REQ_F_REFCOUNT) {
1514 node = req->comp_list.next;
1515 if (!req_ref_put_and_test(req))
1518 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1519 struct async_poll *apoll = req->apoll;
1521 if (apoll->double_poll)
1522 kfree(apoll->double_poll);
1523 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1525 req->flags &= ~REQ_F_POLLED;
1527 if (req->flags & IO_REQ_LINK_FLAGS)
1529 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1534 io_req_put_rsrc_locked(req, ctx);
1536 io_put_task(req->task);
1537 node = req->comp_list.next;
1538 io_req_add_to_cache(req, ctx);
1542 void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1543 __must_hold(&ctx->uring_lock)
1545 struct io_submit_state *state = &ctx->submit_state;
1546 struct io_wq_work_node *node;
1549 /* must come first to preserve CQE ordering in failure cases */
1550 if (state->cqes_count)
1551 __io_flush_post_cqes(ctx);
1552 __wq_list_for_each(node, &state->compl_reqs) {
1553 struct io_kiocb *req = container_of(node, struct io_kiocb,
1556 if (!(req->flags & REQ_F_CQE_SKIP) &&
1557 unlikely(!io_fill_cqe_req(ctx, req))) {
1558 if (ctx->lockless_cq) {
1559 spin_lock(&ctx->completion_lock);
1560 io_req_cqe_overflow(req);
1561 spin_unlock(&ctx->completion_lock);
1563 io_req_cqe_overflow(req);
1567 __io_cq_unlock_post(ctx);
1569 if (!wq_list_empty(&ctx->submit_state.compl_reqs)) {
1570 io_free_batch_list(ctx, state->compl_reqs.first);
1571 INIT_WQ_LIST(&state->compl_reqs);
1575 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1577 /* See comment at the top of this file */
1579 return __io_cqring_events(ctx);
1583 * We can't just wait for polled events to come to us, we have to actively
1584 * find and complete them.
1586 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1588 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1591 mutex_lock(&ctx->uring_lock);
1592 while (!wq_list_empty(&ctx->iopoll_list)) {
1593 /* let it sleep and repeat later if can't complete a request */
1594 if (io_do_iopoll(ctx, true) == 0)
1597 * Ensure we allow local-to-the-cpu processing to take place,
1598 * in this case we need to ensure that we reap all events.
1599 * Also let task_work, etc. to progress by releasing the mutex
1601 if (need_resched()) {
1602 mutex_unlock(&ctx->uring_lock);
1604 mutex_lock(&ctx->uring_lock);
1607 mutex_unlock(&ctx->uring_lock);
1610 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1612 unsigned int nr_events = 0;
1613 unsigned long check_cq;
1615 if (!io_allowed_run_tw(ctx))
1618 check_cq = READ_ONCE(ctx->check_cq);
1619 if (unlikely(check_cq)) {
1620 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1621 __io_cqring_overflow_flush(ctx);
1623 * Similarly do not spin if we have not informed the user of any
1626 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1630 * Don't enter poll loop if we already have events pending.
1631 * If we do, we can potentially be spinning for commands that
1632 * already triggered a CQE (eg in error).
1634 if (io_cqring_events(ctx))
1641 * If a submit got punted to a workqueue, we can have the
1642 * application entering polling for a command before it gets
1643 * issued. That app will hold the uring_lock for the duration
1644 * of the poll right here, so we need to take a breather every
1645 * now and then to ensure that the issue has a chance to add
1646 * the poll to the issued list. Otherwise we can spin here
1647 * forever, while the workqueue is stuck trying to acquire the
1650 if (wq_list_empty(&ctx->iopoll_list) ||
1651 io_task_work_pending(ctx)) {
1652 u32 tail = ctx->cached_cq_tail;
1654 (void) io_run_local_work_locked(ctx);
1656 if (task_work_pending(current) ||
1657 wq_list_empty(&ctx->iopoll_list)) {
1658 mutex_unlock(&ctx->uring_lock);
1660 mutex_lock(&ctx->uring_lock);
1662 /* some requests don't go through iopoll_list */
1663 if (tail != ctx->cached_cq_tail ||
1664 wq_list_empty(&ctx->iopoll_list))
1667 ret = io_do_iopoll(ctx, !min);
1668 if (unlikely(ret < 0))
1671 if (task_sigpending(current))
1677 } while (nr_events < min);
1682 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts)
1685 io_req_complete_defer(req);
1687 io_req_complete_post(req, IO_URING_F_UNLOCKED);
1691 * After the iocb has been issued, it's safe to be found on the poll list.
1692 * Adding the kiocb to the list AFTER submission ensures that we don't
1693 * find it from a io_do_iopoll() thread before the issuer is done
1694 * accessing the kiocb cookie.
1696 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1698 struct io_ring_ctx *ctx = req->ctx;
1699 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1701 /* workqueue context doesn't hold uring_lock, grab it now */
1702 if (unlikely(needs_lock))
1703 mutex_lock(&ctx->uring_lock);
1706 * Track whether we have multiple files in our lists. This will impact
1707 * how we do polling eventually, not spinning if we're on potentially
1708 * different devices.
1710 if (wq_list_empty(&ctx->iopoll_list)) {
1711 ctx->poll_multi_queue = false;
1712 } else if (!ctx->poll_multi_queue) {
1713 struct io_kiocb *list_req;
1715 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1717 if (list_req->file != req->file)
1718 ctx->poll_multi_queue = true;
1722 * For fast devices, IO may have already completed. If it has, add
1723 * it to the front so we find it first.
1725 if (READ_ONCE(req->iopoll_completed))
1726 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1728 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1730 if (unlikely(needs_lock)) {
1732 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1733 * in sq thread task context or in io worker task context. If
1734 * current task context is sq thread, we don't need to check
1735 * whether should wake up sq thread.
1737 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1738 wq_has_sleeper(&ctx->sq_data->wait))
1739 wake_up(&ctx->sq_data->wait);
1741 mutex_unlock(&ctx->uring_lock);
1745 unsigned int io_file_get_flags(struct file *file)
1747 unsigned int res = 0;
1749 if (S_ISREG(file_inode(file)->i_mode))
1751 if ((file->f_flags & O_NONBLOCK) || (file->f_mode & FMODE_NOWAIT))
1752 res |= REQ_F_SUPPORT_NOWAIT;
1756 bool io_alloc_async_data(struct io_kiocb *req)
1758 WARN_ON_ONCE(!io_cold_defs[req->opcode].async_size);
1759 req->async_data = kmalloc(io_cold_defs[req->opcode].async_size, GFP_KERNEL);
1760 if (req->async_data) {
1761 req->flags |= REQ_F_ASYNC_DATA;
1767 int io_req_prep_async(struct io_kiocb *req)
1769 const struct io_cold_def *cdef = &io_cold_defs[req->opcode];
1770 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1772 /* assign early for deferred execution for non-fixed file */
1773 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE) && !req->file)
1774 req->file = io_file_get_normal(req, req->cqe.fd);
1775 if (!cdef->prep_async)
1777 if (WARN_ON_ONCE(req_has_async_data(req)))
1779 if (!def->manual_alloc) {
1780 if (io_alloc_async_data(req))
1783 return cdef->prep_async(req);
1786 static u32 io_get_sequence(struct io_kiocb *req)
1788 u32 seq = req->ctx->cached_sq_head;
1789 struct io_kiocb *cur;
1791 /* need original cached_sq_head, but it was increased for each req */
1792 io_for_each_link(cur, req)
1797 static __cold void io_drain_req(struct io_kiocb *req)
1798 __must_hold(&ctx->uring_lock)
1800 struct io_ring_ctx *ctx = req->ctx;
1801 struct io_defer_entry *de;
1803 u32 seq = io_get_sequence(req);
1805 /* Still need defer if there is pending req in defer list. */
1806 spin_lock(&ctx->completion_lock);
1807 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1808 spin_unlock(&ctx->completion_lock);
1810 ctx->drain_active = false;
1811 io_req_task_queue(req);
1814 spin_unlock(&ctx->completion_lock);
1816 io_prep_async_link(req);
1817 de = kmalloc(sizeof(*de), GFP_KERNEL);
1820 io_req_defer_failed(req, ret);
1824 spin_lock(&ctx->completion_lock);
1825 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1826 spin_unlock(&ctx->completion_lock);
1831 trace_io_uring_defer(req);
1834 list_add_tail(&de->list, &ctx->defer_list);
1835 spin_unlock(&ctx->completion_lock);
1838 static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def,
1839 unsigned int issue_flags)
1841 if (req->file || !def->needs_file)
1844 if (req->flags & REQ_F_FIXED_FILE)
1845 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1847 req->file = io_file_get_normal(req, req->cqe.fd);
1852 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1854 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1855 const struct cred *creds = NULL;
1858 if (unlikely(!io_assign_file(req, def, issue_flags)))
1861 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1862 creds = override_creds(req->creds);
1864 if (!def->audit_skip)
1865 audit_uring_entry(req->opcode);
1867 ret = def->issue(req, issue_flags);
1869 if (!def->audit_skip)
1870 audit_uring_exit(!ret, ret);
1873 revert_creds(creds);
1875 if (ret == IOU_OK) {
1876 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1877 io_req_complete_defer(req);
1879 io_req_complete_post(req, issue_flags);
1884 if (ret == IOU_ISSUE_SKIP_COMPLETE) {
1886 io_arm_ltimeout(req);
1888 /* If the op doesn't have a file, we're not polling for it */
1889 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
1890 io_iopoll_req_issued(req, issue_flags);
1895 int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts)
1897 io_tw_lock(req->ctx, ts);
1898 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT|
1899 IO_URING_F_COMPLETE_DEFER);
1902 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1904 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1905 struct io_kiocb *nxt = NULL;
1907 if (req_ref_put_and_test(req)) {
1908 if (req->flags & IO_REQ_LINK_FLAGS)
1909 nxt = io_req_find_next(req);
1912 return nxt ? &nxt->work : NULL;
1915 void io_wq_submit_work(struct io_wq_work *work)
1917 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1918 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1919 unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ;
1920 bool needs_poll = false;
1921 int ret = 0, err = -ECANCELED;
1923 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1924 if (!(req->flags & REQ_F_REFCOUNT))
1925 __io_req_set_refcount(req, 2);
1929 io_arm_ltimeout(req);
1931 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1932 if (work->flags & IO_WQ_WORK_CANCEL) {
1934 io_req_task_queue_fail(req, err);
1937 if (!io_assign_file(req, def, issue_flags)) {
1939 work->flags |= IO_WQ_WORK_CANCEL;
1943 if (req->flags & REQ_F_FORCE_ASYNC) {
1944 bool opcode_poll = def->pollin || def->pollout;
1946 if (opcode_poll && file_can_poll(req->file)) {
1948 issue_flags |= IO_URING_F_NONBLOCK;
1953 ret = io_issue_sqe(req, issue_flags);
1958 * If REQ_F_NOWAIT is set, then don't wait or retry with
1959 * poll. -EAGAIN is final for that case.
1961 if (req->flags & REQ_F_NOWAIT)
1965 * We can get EAGAIN for iopolled IO even though we're
1966 * forcing a sync submission from here, since we can't
1967 * wait for request slots on the block side.
1970 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1972 if (io_wq_worker_stopped())
1978 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1980 /* aborted or ready, in either case retry blocking */
1982 issue_flags &= ~IO_URING_F_NONBLOCK;
1985 /* avoid locking problems by failing it from a clean context */
1987 io_req_task_queue_fail(req, ret);
1990 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1991 unsigned int issue_flags)
1993 struct io_ring_ctx *ctx = req->ctx;
1994 struct io_fixed_file *slot;
1995 struct file *file = NULL;
1997 io_ring_submit_lock(ctx, issue_flags);
1999 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
2001 fd = array_index_nospec(fd, ctx->nr_user_files);
2002 slot = io_fixed_file_slot(&ctx->file_table, fd);
2003 file = io_slot_file(slot);
2004 req->flags |= io_slot_flags(slot);
2005 io_req_set_rsrc_node(req, ctx, 0);
2007 io_ring_submit_unlock(ctx, issue_flags);
2011 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
2013 struct file *file = fget(fd);
2015 trace_io_uring_file_get(req, fd);
2017 /* we don't allow fixed io_uring files */
2018 if (file && io_is_uring_fops(file))
2019 io_req_track_inflight(req);
2023 static void io_queue_async(struct io_kiocb *req, int ret)
2024 __must_hold(&req->ctx->uring_lock)
2026 struct io_kiocb *linked_timeout;
2028 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
2029 io_req_defer_failed(req, ret);
2033 linked_timeout = io_prep_linked_timeout(req);
2035 switch (io_arm_poll_handler(req, 0)) {
2036 case IO_APOLL_READY:
2037 io_kbuf_recycle(req, 0);
2038 io_req_task_queue(req);
2040 case IO_APOLL_ABORTED:
2041 io_kbuf_recycle(req, 0);
2042 io_queue_iowq(req, NULL);
2049 io_queue_linked_timeout(linked_timeout);
2052 static inline void io_queue_sqe(struct io_kiocb *req)
2053 __must_hold(&req->ctx->uring_lock)
2057 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
2060 * We async punt it if the file wasn't marked NOWAIT, or if the file
2061 * doesn't support non-blocking read/write attempts
2064 io_queue_async(req, ret);
2067 static void io_queue_sqe_fallback(struct io_kiocb *req)
2068 __must_hold(&req->ctx->uring_lock)
2070 if (unlikely(req->flags & REQ_F_FAIL)) {
2072 * We don't submit, fail them all, for that replace hardlinks
2073 * with normal links. Extra REQ_F_LINK is tolerated.
2075 req->flags &= ~REQ_F_HARDLINK;
2076 req->flags |= REQ_F_LINK;
2077 io_req_defer_failed(req, req->cqe.res);
2079 int ret = io_req_prep_async(req);
2081 if (unlikely(ret)) {
2082 io_req_defer_failed(req, ret);
2086 if (unlikely(req->ctx->drain_active))
2089 io_queue_iowq(req, NULL);
2094 * Check SQE restrictions (opcode and flags).
2096 * Returns 'true' if SQE is allowed, 'false' otherwise.
2098 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
2099 struct io_kiocb *req,
2100 unsigned int sqe_flags)
2102 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
2105 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
2106 ctx->restrictions.sqe_flags_required)
2109 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
2110 ctx->restrictions.sqe_flags_required))
2116 static void io_init_req_drain(struct io_kiocb *req)
2118 struct io_ring_ctx *ctx = req->ctx;
2119 struct io_kiocb *head = ctx->submit_state.link.head;
2121 ctx->drain_active = true;
2124 * If we need to drain a request in the middle of a link, drain
2125 * the head request and the next request/link after the current
2126 * link. Considering sequential execution of links,
2127 * REQ_F_IO_DRAIN will be maintained for every request of our
2130 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2131 ctx->drain_next = true;
2135 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
2136 const struct io_uring_sqe *sqe)
2137 __must_hold(&ctx->uring_lock)
2139 const struct io_issue_def *def;
2140 unsigned int sqe_flags;
2144 /* req is partially pre-initialised, see io_preinit_req() */
2145 req->opcode = opcode = READ_ONCE(sqe->opcode);
2146 /* same numerical values with corresponding REQ_F_*, safe to copy */
2147 req->flags = sqe_flags = READ_ONCE(sqe->flags);
2148 req->cqe.user_data = READ_ONCE(sqe->user_data);
2150 req->rsrc_node = NULL;
2151 req->task = current;
2153 if (unlikely(opcode >= IORING_OP_LAST)) {
2157 def = &io_issue_defs[opcode];
2158 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2159 /* enforce forwards compatibility on users */
2160 if (sqe_flags & ~SQE_VALID_FLAGS)
2162 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2163 if (!def->buffer_select)
2165 req->buf_index = READ_ONCE(sqe->buf_group);
2167 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2168 ctx->drain_disabled = true;
2169 if (sqe_flags & IOSQE_IO_DRAIN) {
2170 if (ctx->drain_disabled)
2172 io_init_req_drain(req);
2175 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2176 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2178 /* knock it to the slow queue path, will be drained there */
2179 if (ctx->drain_active)
2180 req->flags |= REQ_F_FORCE_ASYNC;
2181 /* if there is no link, we're at "next" request and need to drain */
2182 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2183 ctx->drain_next = false;
2184 ctx->drain_active = true;
2185 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2189 if (!def->ioprio && sqe->ioprio)
2191 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2194 if (def->needs_file) {
2195 struct io_submit_state *state = &ctx->submit_state;
2197 req->cqe.fd = READ_ONCE(sqe->fd);
2200 * Plug now if we have more than 2 IO left after this, and the
2201 * target is potentially a read/write to block based storage.
2203 if (state->need_plug && def->plug) {
2204 state->plug_started = true;
2205 state->need_plug = false;
2206 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2210 personality = READ_ONCE(sqe->personality);
2214 req->creds = xa_load(&ctx->personalities, personality);
2217 get_cred(req->creds);
2218 ret = security_uring_override_creds(req->creds);
2220 put_cred(req->creds);
2223 req->flags |= REQ_F_CREDS;
2226 return def->prep(req, sqe);
2229 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2230 struct io_kiocb *req, int ret)
2232 struct io_ring_ctx *ctx = req->ctx;
2233 struct io_submit_link *link = &ctx->submit_state.link;
2234 struct io_kiocb *head = link->head;
2236 trace_io_uring_req_failed(sqe, req, ret);
2239 * Avoid breaking links in the middle as it renders links with SQPOLL
2240 * unusable. Instead of failing eagerly, continue assembling the link if
2241 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2242 * should find the flag and handle the rest.
2244 req_fail_link_node(req, ret);
2245 if (head && !(head->flags & REQ_F_FAIL))
2246 req_fail_link_node(head, -ECANCELED);
2248 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2250 link->last->link = req;
2254 io_queue_sqe_fallback(req);
2259 link->last->link = req;
2266 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2267 const struct io_uring_sqe *sqe)
2268 __must_hold(&ctx->uring_lock)
2270 struct io_submit_link *link = &ctx->submit_state.link;
2273 ret = io_init_req(ctx, req, sqe);
2275 return io_submit_fail_init(sqe, req, ret);
2277 trace_io_uring_submit_req(req);
2280 * If we already have a head request, queue this one for async
2281 * submittal once the head completes. If we don't have a head but
2282 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2283 * submitted sync once the chain is complete. If none of those
2284 * conditions are true (normal request), then just queue it.
2286 if (unlikely(link->head)) {
2287 ret = io_req_prep_async(req);
2289 return io_submit_fail_init(sqe, req, ret);
2291 trace_io_uring_link(req, link->head);
2292 link->last->link = req;
2295 if (req->flags & IO_REQ_LINK_FLAGS)
2297 /* last request of the link, flush it */
2300 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2303 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2304 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2305 if (req->flags & IO_REQ_LINK_FLAGS) {
2310 io_queue_sqe_fallback(req);
2320 * Batched submission is done, ensure local IO is flushed out.
2322 static void io_submit_state_end(struct io_ring_ctx *ctx)
2324 struct io_submit_state *state = &ctx->submit_state;
2326 if (unlikely(state->link.head))
2327 io_queue_sqe_fallback(state->link.head);
2328 /* flush only after queuing links as they can generate completions */
2329 io_submit_flush_completions(ctx);
2330 if (state->plug_started)
2331 blk_finish_plug(&state->plug);
2335 * Start submission side cache.
2337 static void io_submit_state_start(struct io_submit_state *state,
2338 unsigned int max_ios)
2340 state->plug_started = false;
2341 state->need_plug = max_ios > 2;
2342 state->submit_nr = max_ios;
2343 /* set only head, no need to init link_last in advance */
2344 state->link.head = NULL;
2347 static void io_commit_sqring(struct io_ring_ctx *ctx)
2349 struct io_rings *rings = ctx->rings;
2352 * Ensure any loads from the SQEs are done at this point,
2353 * since once we write the new head, the application could
2354 * write new data to them.
2356 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2360 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2361 * that is mapped by userspace. This means that care needs to be taken to
2362 * ensure that reads are stable, as we cannot rely on userspace always
2363 * being a good citizen. If members of the sqe are validated and then later
2364 * used, it's important that those reads are done through READ_ONCE() to
2365 * prevent a re-load down the line.
2367 static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe)
2369 unsigned mask = ctx->sq_entries - 1;
2370 unsigned head = ctx->cached_sq_head++ & mask;
2372 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY)) {
2373 head = READ_ONCE(ctx->sq_array[head]);
2374 if (unlikely(head >= ctx->sq_entries)) {
2375 /* drop invalid entries */
2376 spin_lock(&ctx->completion_lock);
2378 spin_unlock(&ctx->completion_lock);
2379 WRITE_ONCE(ctx->rings->sq_dropped,
2380 READ_ONCE(ctx->rings->sq_dropped) + 1);
2386 * The cached sq head (or cq tail) serves two purposes:
2388 * 1) allows us to batch the cost of updating the user visible
2390 * 2) allows the kernel side to track the head on its own, even
2391 * though the application is the one updating it.
2394 /* double index for 128-byte SQEs, twice as long */
2395 if (ctx->flags & IORING_SETUP_SQE128)
2397 *sqe = &ctx->sq_sqes[head];
2401 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2402 __must_hold(&ctx->uring_lock)
2404 unsigned int entries = io_sqring_entries(ctx);
2408 if (unlikely(!entries))
2410 /* make sure SQ entry isn't read before tail */
2411 ret = left = min(nr, entries);
2412 io_get_task_refs(left);
2413 io_submit_state_start(&ctx->submit_state, left);
2416 const struct io_uring_sqe *sqe;
2417 struct io_kiocb *req;
2419 if (unlikely(!io_alloc_req(ctx, &req)))
2421 if (unlikely(!io_get_sqe(ctx, &sqe))) {
2422 io_req_add_to_cache(req, ctx);
2427 * Continue submitting even for sqe failure if the
2428 * ring was setup with IORING_SETUP_SUBMIT_ALL
2430 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2431 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2437 if (unlikely(left)) {
2439 /* try again if it submitted nothing and can't allocate a req */
2440 if (!ret && io_req_cache_empty(ctx))
2442 current->io_uring->cached_refs += left;
2445 io_submit_state_end(ctx);
2446 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2447 io_commit_sqring(ctx);
2451 struct io_wait_queue {
2452 struct wait_queue_entry wq;
2453 struct io_ring_ctx *ctx;
2455 unsigned nr_timeouts;
2459 static inline bool io_has_work(struct io_ring_ctx *ctx)
2461 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
2462 !llist_empty(&ctx->work_llist);
2465 static inline bool io_should_wake(struct io_wait_queue *iowq)
2467 struct io_ring_ctx *ctx = iowq->ctx;
2468 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
2471 * Wake up if we have enough events, or if a timeout occurred since we
2472 * started waiting. For timeouts, we always want to return to userspace,
2473 * regardless of event count.
2475 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2478 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2479 int wake_flags, void *key)
2481 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq);
2484 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2485 * the task, and the next invocation will do it.
2487 if (io_should_wake(iowq) || io_has_work(iowq->ctx))
2488 return autoremove_wake_function(curr, mode, wake_flags, key);
2492 int io_run_task_work_sig(struct io_ring_ctx *ctx)
2494 if (!llist_empty(&ctx->work_llist)) {
2495 __set_current_state(TASK_RUNNING);
2496 if (io_run_local_work(ctx) > 0)
2499 if (io_run_task_work() > 0)
2501 if (task_sigpending(current))
2506 static bool current_pending_io(void)
2508 struct io_uring_task *tctx = current->io_uring;
2512 return percpu_counter_read_positive(&tctx->inflight);
2515 /* when returns >0, the caller should retry */
2516 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2517 struct io_wait_queue *iowq)
2521 if (unlikely(READ_ONCE(ctx->check_cq)))
2523 if (unlikely(!llist_empty(&ctx->work_llist)))
2525 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL)))
2527 if (unlikely(task_sigpending(current)))
2529 if (unlikely(io_should_wake(iowq)))
2533 * Mark us as being in io_wait if we have pending requests, so cpufreq
2534 * can take into account that the task is waiting for IO - turns out
2535 * to be important for low QD IO.
2537 io_wait = current->in_iowait;
2538 if (current_pending_io())
2539 current->in_iowait = 1;
2541 if (iowq->timeout == KTIME_MAX)
2543 else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS))
2545 current->in_iowait = io_wait;
2550 * Wait until events become available, if we don't already have some. The
2551 * application must reap them itself, as they reside on the shared cq ring.
2553 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2554 const sigset_t __user *sig, size_t sigsz,
2555 struct __kernel_timespec __user *uts)
2557 struct io_wait_queue iowq;
2558 struct io_rings *rings = ctx->rings;
2561 if (!io_allowed_run_tw(ctx))
2563 if (!llist_empty(&ctx->work_llist))
2564 io_run_local_work(ctx);
2566 io_cqring_overflow_flush(ctx);
2567 /* if user messes with these they will just get an early return */
2568 if (__io_cqring_events_user(ctx) >= min_events)
2572 #ifdef CONFIG_COMPAT
2573 if (in_compat_syscall())
2574 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2578 ret = set_user_sigmask(sig, sigsz);
2584 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2585 iowq.wq.private = current;
2586 INIT_LIST_HEAD(&iowq.wq.entry);
2588 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2589 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2590 iowq.timeout = KTIME_MAX;
2593 struct timespec64 ts;
2595 if (get_timespec64(&ts, uts))
2597 iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2600 trace_io_uring_cqring_wait(ctx, min_events);
2602 unsigned long check_cq;
2604 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
2605 int nr_wait = (int) iowq.cq_tail - READ_ONCE(ctx->rings->cq.tail);
2607 atomic_set(&ctx->cq_wait_nr, nr_wait);
2608 set_current_state(TASK_INTERRUPTIBLE);
2610 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2611 TASK_INTERRUPTIBLE);
2614 ret = io_cqring_wait_schedule(ctx, &iowq);
2615 __set_current_state(TASK_RUNNING);
2616 atomic_set(&ctx->cq_wait_nr, 0);
2619 * Run task_work after scheduling and before io_should_wake().
2620 * If we got woken because of task_work being processed, run it
2621 * now rather than let the caller do another wait loop.
2624 if (!llist_empty(&ctx->work_llist))
2625 io_run_local_work(ctx);
2628 * Non-local task_work will be run on exit to userspace, but
2629 * if we're using DEFER_TASKRUN, then we could have waited
2630 * with a timeout for a number of requests. If the timeout
2631 * hits, we could have some requests ready to process. Ensure
2632 * this break is _after_ we have run task_work, to avoid
2633 * deferring running potentially pending requests until the
2634 * next time we wait for events.
2639 check_cq = READ_ONCE(ctx->check_cq);
2640 if (unlikely(check_cq)) {
2641 /* let the caller flush overflows, retry */
2642 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2643 io_cqring_do_overflow_flush(ctx);
2644 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) {
2650 if (io_should_wake(&iowq)) {
2657 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
2658 finish_wait(&ctx->cq_wait, &iowq.wq);
2659 restore_saved_sigmask_unless(ret == -EINTR);
2661 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2664 void io_mem_free(void *ptr)
2669 folio_put(virt_to_folio(ptr));
2672 static void io_pages_free(struct page ***pages, int npages)
2674 struct page **page_array;
2680 page_array = *pages;
2684 for (i = 0; i < npages; i++)
2685 unpin_user_page(page_array[i]);
2690 static void *__io_uaddr_map(struct page ***pages, unsigned short *npages,
2691 unsigned long uaddr, size_t size)
2693 struct page **page_array;
2694 unsigned int nr_pages;
2700 if (uaddr & (PAGE_SIZE - 1) || !size)
2701 return ERR_PTR(-EINVAL);
2703 nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
2704 if (nr_pages > USHRT_MAX)
2705 return ERR_PTR(-EINVAL);
2706 page_array = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
2708 return ERR_PTR(-ENOMEM);
2710 ret = pin_user_pages_fast(uaddr, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
2712 if (ret != nr_pages) {
2714 io_pages_free(&page_array, ret > 0 ? ret : 0);
2715 return ret < 0 ? ERR_PTR(ret) : ERR_PTR(-EFAULT);
2718 page_addr = page_address(page_array[0]);
2719 for (i = 0; i < nr_pages; i++) {
2723 * Can't support mapping user allocated ring memory on 32-bit
2724 * archs where it could potentially reside in highmem. Just
2725 * fail those with -EINVAL, just like we did on kernels that
2726 * didn't support this feature.
2728 if (PageHighMem(page_array[i]))
2732 * No support for discontig pages for now, should either be a
2733 * single normal page, or a huge page. Later on we can add
2734 * support for remapping discontig pages, for now we will
2735 * just fail them with EINVAL.
2737 if (page_address(page_array[i]) != page_addr)
2739 page_addr += PAGE_SIZE;
2742 *pages = page_array;
2744 return page_to_virt(page_array[0]);
2747 static void *io_rings_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2750 return __io_uaddr_map(&ctx->ring_pages, &ctx->n_ring_pages, uaddr,
2754 static void *io_sqes_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2757 return __io_uaddr_map(&ctx->sqe_pages, &ctx->n_sqe_pages, uaddr,
2761 static void io_rings_free(struct io_ring_ctx *ctx)
2763 if (!(ctx->flags & IORING_SETUP_NO_MMAP)) {
2764 io_mem_free(ctx->rings);
2765 io_mem_free(ctx->sq_sqes);
2767 ctx->sq_sqes = NULL;
2769 io_pages_free(&ctx->ring_pages, ctx->n_ring_pages);
2770 ctx->n_ring_pages = 0;
2771 io_pages_free(&ctx->sqe_pages, ctx->n_sqe_pages);
2772 ctx->n_sqe_pages = 0;
2776 void *io_mem_alloc(size_t size)
2778 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2781 ret = (void *) __get_free_pages(gfp, get_order(size));
2784 return ERR_PTR(-ENOMEM);
2787 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2788 unsigned int cq_entries, size_t *sq_offset)
2790 struct io_rings *rings;
2791 size_t off, sq_array_size;
2793 off = struct_size(rings, cqes, cq_entries);
2794 if (off == SIZE_MAX)
2796 if (ctx->flags & IORING_SETUP_CQE32) {
2797 if (check_shl_overflow(off, 1, &off))
2802 off = ALIGN(off, SMP_CACHE_BYTES);
2807 if (ctx->flags & IORING_SETUP_NO_SQARRAY) {
2809 *sq_offset = SIZE_MAX;
2816 sq_array_size = array_size(sizeof(u32), sq_entries);
2817 if (sq_array_size == SIZE_MAX)
2820 if (check_add_overflow(off, sq_array_size, &off))
2826 static void io_req_caches_free(struct io_ring_ctx *ctx)
2828 struct io_kiocb *req;
2831 mutex_lock(&ctx->uring_lock);
2832 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
2834 while (!io_req_cache_empty(ctx)) {
2835 req = io_extract_req(ctx);
2836 kmem_cache_free(req_cachep, req);
2840 percpu_ref_put_many(&ctx->refs, nr);
2841 mutex_unlock(&ctx->uring_lock);
2844 static void io_rsrc_node_cache_free(struct io_cache_entry *entry)
2846 kfree(container_of(entry, struct io_rsrc_node, cache));
2849 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2851 io_sq_thread_finish(ctx);
2852 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2853 if (WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)))
2856 mutex_lock(&ctx->uring_lock);
2858 __io_sqe_buffers_unregister(ctx);
2860 __io_sqe_files_unregister(ctx);
2861 io_cqring_overflow_kill(ctx);
2862 io_eventfd_unregister(ctx);
2863 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2864 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2865 io_futex_cache_free(ctx);
2866 io_destroy_buffers(ctx);
2867 mutex_unlock(&ctx->uring_lock);
2869 put_cred(ctx->sq_creds);
2870 if (ctx->submitter_task)
2871 put_task_struct(ctx->submitter_task);
2873 /* there are no registered resources left, nobody uses it */
2875 io_rsrc_node_destroy(ctx, ctx->rsrc_node);
2877 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2878 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2880 io_alloc_cache_free(&ctx->rsrc_node_cache, io_rsrc_node_cache_free);
2881 if (ctx->mm_account) {
2882 mmdrop(ctx->mm_account);
2883 ctx->mm_account = NULL;
2886 io_kbuf_mmap_list_free(ctx);
2888 percpu_ref_exit(&ctx->refs);
2889 free_uid(ctx->user);
2890 io_req_caches_free(ctx);
2892 io_wq_put_hash(ctx->hash_map);
2893 kfree(ctx->cancel_table.hbs);
2894 kfree(ctx->cancel_table_locked.hbs);
2896 xa_destroy(&ctx->io_bl_xa);
2900 static __cold void io_activate_pollwq_cb(struct callback_head *cb)
2902 struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx,
2905 mutex_lock(&ctx->uring_lock);
2906 ctx->poll_activated = true;
2907 mutex_unlock(&ctx->uring_lock);
2910 * Wake ups for some events between start of polling and activation
2911 * might've been lost due to loose synchronisation.
2913 wake_up_all(&ctx->poll_wq);
2914 percpu_ref_put(&ctx->refs);
2917 __cold void io_activate_pollwq(struct io_ring_ctx *ctx)
2919 spin_lock(&ctx->completion_lock);
2920 /* already activated or in progress */
2921 if (ctx->poll_activated || ctx->poll_wq_task_work.func)
2923 if (WARN_ON_ONCE(!ctx->task_complete))
2925 if (!ctx->submitter_task)
2928 * with ->submitter_task only the submitter task completes requests, we
2929 * only need to sync with it, which is done by injecting a tw
2931 init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb);
2932 percpu_ref_get(&ctx->refs);
2933 if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL))
2934 percpu_ref_put(&ctx->refs);
2936 spin_unlock(&ctx->completion_lock);
2939 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2941 struct io_ring_ctx *ctx = file->private_data;
2944 if (unlikely(!ctx->poll_activated))
2945 io_activate_pollwq(ctx);
2947 poll_wait(file, &ctx->poll_wq, wait);
2949 * synchronizes with barrier from wq_has_sleeper call in
2953 if (!io_sqring_full(ctx))
2954 mask |= EPOLLOUT | EPOLLWRNORM;
2957 * Don't flush cqring overflow list here, just do a simple check.
2958 * Otherwise there could possible be ABBA deadlock:
2961 * lock(&ctx->uring_lock);
2963 * lock(&ctx->uring_lock);
2966 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2967 * pushes them to do the flush.
2970 if (__io_cqring_events_user(ctx) || io_has_work(ctx))
2971 mask |= EPOLLIN | EPOLLRDNORM;
2976 struct io_tctx_exit {
2977 struct callback_head task_work;
2978 struct completion completion;
2979 struct io_ring_ctx *ctx;
2982 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2984 struct io_uring_task *tctx = current->io_uring;
2985 struct io_tctx_exit *work;
2987 work = container_of(cb, struct io_tctx_exit, task_work);
2989 * When @in_cancel, we're in cancellation and it's racy to remove the
2990 * node. It'll be removed by the end of cancellation, just ignore it.
2991 * tctx can be NULL if the queueing of this task_work raced with
2992 * work cancelation off the exec path.
2994 if (tctx && !atomic_read(&tctx->in_cancel))
2995 io_uring_del_tctx_node((unsigned long)work->ctx);
2996 complete(&work->completion);
2999 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
3001 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3003 return req->ctx == data;
3006 static __cold void io_ring_exit_work(struct work_struct *work)
3008 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
3009 unsigned long timeout = jiffies + HZ * 60 * 5;
3010 unsigned long interval = HZ / 20;
3011 struct io_tctx_exit exit;
3012 struct io_tctx_node *node;
3016 * If we're doing polled IO and end up having requests being
3017 * submitted async (out-of-line), then completions can come in while
3018 * we're waiting for refs to drop. We need to reap these manually,
3019 * as nobody else will be looking for them.
3022 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
3023 mutex_lock(&ctx->uring_lock);
3024 io_cqring_overflow_kill(ctx);
3025 mutex_unlock(&ctx->uring_lock);
3028 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3029 io_move_task_work_from_local(ctx);
3031 while (io_uring_try_cancel_requests(ctx, NULL, true))
3035 struct io_sq_data *sqd = ctx->sq_data;
3036 struct task_struct *tsk;
3038 io_sq_thread_park(sqd);
3040 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
3041 io_wq_cancel_cb(tsk->io_uring->io_wq,
3042 io_cancel_ctx_cb, ctx, true);
3043 io_sq_thread_unpark(sqd);
3046 io_req_caches_free(ctx);
3048 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
3049 /* there is little hope left, don't run it too often */
3053 * This is really an uninterruptible wait, as it has to be
3054 * complete. But it's also run from a kworker, which doesn't
3055 * take signals, so it's fine to make it interruptible. This
3056 * avoids scenarios where we knowingly can wait much longer
3057 * on completions, for example if someone does a SIGSTOP on
3058 * a task that needs to finish task_work to make this loop
3059 * complete. That's a synthetic situation that should not
3060 * cause a stuck task backtrace, and hence a potential panic
3061 * on stuck tasks if that is enabled.
3063 } while (!wait_for_completion_interruptible_timeout(&ctx->ref_comp, interval));
3065 init_completion(&exit.completion);
3066 init_task_work(&exit.task_work, io_tctx_exit_cb);
3069 mutex_lock(&ctx->uring_lock);
3070 while (!list_empty(&ctx->tctx_list)) {
3071 WARN_ON_ONCE(time_after(jiffies, timeout));
3073 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
3075 /* don't spin on a single task if cancellation failed */
3076 list_rotate_left(&ctx->tctx_list);
3077 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
3078 if (WARN_ON_ONCE(ret))
3081 mutex_unlock(&ctx->uring_lock);
3083 * See comment above for
3084 * wait_for_completion_interruptible_timeout() on why this
3085 * wait is marked as interruptible.
3087 wait_for_completion_interruptible(&exit.completion);
3088 mutex_lock(&ctx->uring_lock);
3090 mutex_unlock(&ctx->uring_lock);
3091 spin_lock(&ctx->completion_lock);
3092 spin_unlock(&ctx->completion_lock);
3094 /* pairs with RCU read section in io_req_local_work_add() */
3095 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3098 io_ring_ctx_free(ctx);
3101 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
3103 unsigned long index;
3104 struct creds *creds;
3106 mutex_lock(&ctx->uring_lock);
3107 percpu_ref_kill(&ctx->refs);
3108 xa_for_each(&ctx->personalities, index, creds)
3109 io_unregister_personality(ctx, index);
3111 io_poll_remove_all(ctx, NULL, true);
3112 mutex_unlock(&ctx->uring_lock);
3115 * If we failed setting up the ctx, we might not have any rings
3116 * and therefore did not submit any requests
3119 io_kill_timeouts(ctx, NULL, true);
3121 flush_delayed_work(&ctx->fallback_work);
3123 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
3125 * Use system_unbound_wq to avoid spawning tons of event kworkers
3126 * if we're exiting a ton of rings at the same time. It just adds
3127 * noise and overhead, there's no discernable change in runtime
3128 * over using system_wq.
3130 queue_work(system_unbound_wq, &ctx->exit_work);
3133 static int io_uring_release(struct inode *inode, struct file *file)
3135 struct io_ring_ctx *ctx = file->private_data;
3137 file->private_data = NULL;
3138 io_ring_ctx_wait_and_kill(ctx);
3142 struct io_task_cancel {
3143 struct task_struct *task;
3147 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
3149 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3150 struct io_task_cancel *cancel = data;
3152 return io_match_task_safe(req, cancel->task, cancel->all);
3155 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
3156 struct task_struct *task,
3159 struct io_defer_entry *de;
3162 spin_lock(&ctx->completion_lock);
3163 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
3164 if (io_match_task_safe(de->req, task, cancel_all)) {
3165 list_cut_position(&list, &ctx->defer_list, &de->list);
3169 spin_unlock(&ctx->completion_lock);
3170 if (list_empty(&list))
3173 while (!list_empty(&list)) {
3174 de = list_first_entry(&list, struct io_defer_entry, list);
3175 list_del_init(&de->list);
3176 io_req_task_queue_fail(de->req, -ECANCELED);
3182 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
3184 struct io_tctx_node *node;
3185 enum io_wq_cancel cret;
3188 mutex_lock(&ctx->uring_lock);
3189 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3190 struct io_uring_task *tctx = node->task->io_uring;
3193 * io_wq will stay alive while we hold uring_lock, because it's
3194 * killed after ctx nodes, which requires to take the lock.
3196 if (!tctx || !tctx->io_wq)
3198 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
3199 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3201 mutex_unlock(&ctx->uring_lock);
3206 static bool io_uring_try_cancel_uring_cmd(struct io_ring_ctx *ctx,
3207 struct task_struct *task, bool cancel_all)
3209 struct hlist_node *tmp;
3210 struct io_kiocb *req;
3213 lockdep_assert_held(&ctx->uring_lock);
3215 hlist_for_each_entry_safe(req, tmp, &ctx->cancelable_uring_cmd,
3217 struct io_uring_cmd *cmd = io_kiocb_to_cmd(req,
3218 struct io_uring_cmd);
3219 struct file *file = req->file;
3221 if (!cancel_all && req->task != task)
3224 if (cmd->flags & IORING_URING_CMD_CANCELABLE) {
3225 /* ->sqe isn't available if no async data */
3226 if (!req_has_async_data(req))
3228 file->f_op->uring_cmd(cmd, IO_URING_F_CANCEL);
3232 io_submit_flush_completions(ctx);
3237 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
3238 struct task_struct *task,
3241 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
3242 struct io_uring_task *tctx = task ? task->io_uring : NULL;
3243 enum io_wq_cancel cret;
3246 /* set it so io_req_local_work_add() would wake us up */
3247 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
3248 atomic_set(&ctx->cq_wait_nr, 1);
3252 /* failed during ring init, it couldn't have issued any requests */
3257 ret |= io_uring_try_cancel_iowq(ctx);
3258 } else if (tctx && tctx->io_wq) {
3260 * Cancels requests of all rings, not only @ctx, but
3261 * it's fine as the task is in exit/exec.
3263 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
3265 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3268 /* SQPOLL thread does its own polling */
3269 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
3270 (ctx->sq_data && ctx->sq_data->thread == current)) {
3271 while (!wq_list_empty(&ctx->iopoll_list)) {
3272 io_iopoll_try_reap_events(ctx);
3278 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3279 io_allowed_defer_tw_run(ctx))
3280 ret |= io_run_local_work(ctx) > 0;
3281 ret |= io_cancel_defer_files(ctx, task, cancel_all);
3282 mutex_lock(&ctx->uring_lock);
3283 ret |= io_poll_remove_all(ctx, task, cancel_all);
3284 ret |= io_waitid_remove_all(ctx, task, cancel_all);
3285 ret |= io_futex_remove_all(ctx, task, cancel_all);
3286 ret |= io_uring_try_cancel_uring_cmd(ctx, task, cancel_all);
3287 mutex_unlock(&ctx->uring_lock);
3288 ret |= io_kill_timeouts(ctx, task, cancel_all);
3290 ret |= io_run_task_work() > 0;
3294 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
3297 return atomic_read(&tctx->inflight_tracked);
3298 return percpu_counter_sum(&tctx->inflight);
3302 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3303 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3305 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
3307 struct io_uring_task *tctx = current->io_uring;
3308 struct io_ring_ctx *ctx;
3309 struct io_tctx_node *node;
3310 unsigned long index;
3314 WARN_ON_ONCE(sqd && sqd->thread != current);
3316 if (!current->io_uring)
3319 io_wq_exit_start(tctx->io_wq);
3321 atomic_inc(&tctx->in_cancel);
3325 io_uring_drop_tctx_refs(current);
3326 /* read completions before cancelations */
3327 inflight = tctx_inflight(tctx, !cancel_all);
3332 xa_for_each(&tctx->xa, index, node) {
3333 /* sqpoll task will cancel all its requests */
3334 if (node->ctx->sq_data)
3336 loop |= io_uring_try_cancel_requests(node->ctx,
3337 current, cancel_all);
3340 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3341 loop |= io_uring_try_cancel_requests(ctx,
3351 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3353 io_uring_drop_tctx_refs(current);
3354 xa_for_each(&tctx->xa, index, node) {
3355 if (!llist_empty(&node->ctx->work_llist)) {
3356 WARN_ON_ONCE(node->ctx->submitter_task &&
3357 node->ctx->submitter_task != current);
3362 * If we've seen completions, retry without waiting. This
3363 * avoids a race where a completion comes in before we did
3364 * prepare_to_wait().
3366 if (inflight == tctx_inflight(tctx, !cancel_all))
3369 finish_wait(&tctx->wait, &wait);
3372 io_uring_clean_tctx(tctx);
3375 * We shouldn't run task_works after cancel, so just leave
3376 * ->in_cancel set for normal exit.
3378 atomic_dec(&tctx->in_cancel);
3379 /* for exec all current's requests should be gone, kill tctx */
3380 __io_uring_free(current);
3384 void __io_uring_cancel(bool cancel_all)
3386 io_uring_cancel_generic(cancel_all, NULL);
3389 static void *io_uring_validate_mmap_request(struct file *file,
3390 loff_t pgoff, size_t sz)
3392 struct io_ring_ctx *ctx = file->private_data;
3393 loff_t offset = pgoff << PAGE_SHIFT;
3397 switch (offset & IORING_OFF_MMAP_MASK) {
3398 case IORING_OFF_SQ_RING:
3399 case IORING_OFF_CQ_RING:
3400 /* Don't allow mmap if the ring was setup without it */
3401 if (ctx->flags & IORING_SETUP_NO_MMAP)
3402 return ERR_PTR(-EINVAL);
3405 case IORING_OFF_SQES:
3406 /* Don't allow mmap if the ring was setup without it */
3407 if (ctx->flags & IORING_SETUP_NO_MMAP)
3408 return ERR_PTR(-EINVAL);
3411 case IORING_OFF_PBUF_RING: {
3414 bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT;
3416 ptr = io_pbuf_get_address(ctx, bgid);
3419 return ERR_PTR(-EINVAL);
3423 return ERR_PTR(-EINVAL);
3426 page = virt_to_head_page(ptr);
3427 if (sz > page_size(page))
3428 return ERR_PTR(-EINVAL);
3435 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3437 size_t sz = vma->vm_end - vma->vm_start;
3441 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3443 return PTR_ERR(ptr);
3445 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3446 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3449 static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp,
3450 unsigned long addr, unsigned long len,
3451 unsigned long pgoff, unsigned long flags)
3456 * Do not allow to map to user-provided address to avoid breaking the
3457 * aliasing rules. Userspace is not able to guess the offset address of
3458 * kernel kmalloc()ed memory area.
3463 ptr = io_uring_validate_mmap_request(filp, pgoff, len);
3468 * Some architectures have strong cache aliasing requirements.
3469 * For such architectures we need a coherent mapping which aliases
3470 * kernel memory *and* userspace memory. To achieve that:
3471 * - use a NULL file pointer to reference physical memory, and
3472 * - use the kernel virtual address of the shared io_uring context
3473 * (instead of the userspace-provided address, which has to be 0UL
3475 * - use the same pgoff which the get_unmapped_area() uses to
3476 * calculate the page colouring.
3477 * For architectures without such aliasing requirements, the
3478 * architecture will return any suitable mapping because addr is 0.
3481 flags |= MAP_SHARED;
3482 pgoff = 0; /* has been translated to ptr above */
3484 addr = (uintptr_t) ptr;
3485 pgoff = addr >> PAGE_SHIFT;
3489 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
3492 #else /* !CONFIG_MMU */
3494 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3496 return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL;
3499 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3501 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3504 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3505 unsigned long addr, unsigned long len,
3506 unsigned long pgoff, unsigned long flags)
3510 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3512 return PTR_ERR(ptr);
3514 return (unsigned long) ptr;
3517 #endif /* !CONFIG_MMU */
3519 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3521 if (flags & IORING_ENTER_EXT_ARG) {
3522 struct io_uring_getevents_arg arg;
3524 if (argsz != sizeof(arg))
3526 if (copy_from_user(&arg, argp, sizeof(arg)))
3532 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3533 struct __kernel_timespec __user **ts,
3534 const sigset_t __user **sig)
3536 struct io_uring_getevents_arg arg;
3539 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3540 * is just a pointer to the sigset_t.
3542 if (!(flags & IORING_ENTER_EXT_ARG)) {
3543 *sig = (const sigset_t __user *) argp;
3549 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3550 * timespec and sigset_t pointers if good.
3552 if (*argsz != sizeof(arg))
3554 if (copy_from_user(&arg, argp, sizeof(arg)))
3558 *sig = u64_to_user_ptr(arg.sigmask);
3559 *argsz = arg.sigmask_sz;
3560 *ts = u64_to_user_ptr(arg.ts);
3564 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3565 u32, min_complete, u32, flags, const void __user *, argp,
3568 struct io_ring_ctx *ctx;
3572 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3573 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3574 IORING_ENTER_REGISTERED_RING)))
3578 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3579 * need only dereference our task private array to find it.
3581 if (flags & IORING_ENTER_REGISTERED_RING) {
3582 struct io_uring_task *tctx = current->io_uring;
3584 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3586 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3587 file = tctx->registered_rings[fd];
3588 if (unlikely(!file))
3592 if (unlikely(!file))
3595 if (unlikely(!io_is_uring_fops(file)))
3599 ctx = file->private_data;
3601 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3605 * For SQ polling, the thread will do all submissions and completions.
3606 * Just return the requested submit count, and wake the thread if
3610 if (ctx->flags & IORING_SETUP_SQPOLL) {
3611 io_cqring_overflow_flush(ctx);
3613 if (unlikely(ctx->sq_data->thread == NULL)) {
3617 if (flags & IORING_ENTER_SQ_WAKEUP)
3618 wake_up(&ctx->sq_data->wait);
3619 if (flags & IORING_ENTER_SQ_WAIT)
3620 io_sqpoll_wait_sq(ctx);
3623 } else if (to_submit) {
3624 ret = io_uring_add_tctx_node(ctx);
3628 mutex_lock(&ctx->uring_lock);
3629 ret = io_submit_sqes(ctx, to_submit);
3630 if (ret != to_submit) {
3631 mutex_unlock(&ctx->uring_lock);
3634 if (flags & IORING_ENTER_GETEVENTS) {
3635 if (ctx->syscall_iopoll)
3638 * Ignore errors, we'll soon call io_cqring_wait() and
3639 * it should handle ownership problems if any.
3641 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3642 (void)io_run_local_work_locked(ctx);
3644 mutex_unlock(&ctx->uring_lock);
3647 if (flags & IORING_ENTER_GETEVENTS) {
3650 if (ctx->syscall_iopoll) {
3652 * We disallow the app entering submit/complete with
3653 * polling, but we still need to lock the ring to
3654 * prevent racing with polled issue that got punted to
3657 mutex_lock(&ctx->uring_lock);
3659 ret2 = io_validate_ext_arg(flags, argp, argsz);
3660 if (likely(!ret2)) {
3661 min_complete = min(min_complete,
3663 ret2 = io_iopoll_check(ctx, min_complete);
3665 mutex_unlock(&ctx->uring_lock);
3667 const sigset_t __user *sig;
3668 struct __kernel_timespec __user *ts;
3670 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3671 if (likely(!ret2)) {
3672 min_complete = min(min_complete,
3674 ret2 = io_cqring_wait(ctx, min_complete, sig,
3683 * EBADR indicates that one or more CQE were dropped.
3684 * Once the user has been informed we can clear the bit
3685 * as they are obviously ok with those drops.
3687 if (unlikely(ret2 == -EBADR))
3688 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3693 if (!(flags & IORING_ENTER_REGISTERED_RING))
3698 static const struct file_operations io_uring_fops = {
3699 .release = io_uring_release,
3700 .mmap = io_uring_mmap,
3702 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3703 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3705 .get_unmapped_area = io_uring_mmu_get_unmapped_area,
3707 .poll = io_uring_poll,
3708 #ifdef CONFIG_PROC_FS
3709 .show_fdinfo = io_uring_show_fdinfo,
3713 bool io_is_uring_fops(struct file *file)
3715 return file->f_op == &io_uring_fops;
3718 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3719 struct io_uring_params *p)
3721 struct io_rings *rings;
3722 size_t size, sq_array_offset;
3725 /* make sure these are sane, as we already accounted them */
3726 ctx->sq_entries = p->sq_entries;
3727 ctx->cq_entries = p->cq_entries;
3729 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3730 if (size == SIZE_MAX)
3733 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3734 rings = io_mem_alloc(size);
3736 rings = io_rings_map(ctx, p->cq_off.user_addr, size);
3739 return PTR_ERR(rings);
3742 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY))
3743 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3744 rings->sq_ring_mask = p->sq_entries - 1;
3745 rings->cq_ring_mask = p->cq_entries - 1;
3746 rings->sq_ring_entries = p->sq_entries;
3747 rings->cq_ring_entries = p->cq_entries;
3749 if (p->flags & IORING_SETUP_SQE128)
3750 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3752 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3753 if (size == SIZE_MAX) {
3758 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3759 ptr = io_mem_alloc(size);
3761 ptr = io_sqes_map(ctx, p->sq_off.user_addr, size);
3765 return PTR_ERR(ptr);
3772 static int io_uring_install_fd(struct file *file)
3776 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3779 fd_install(fd, file);
3784 * Allocate an anonymous fd, this is what constitutes the application
3785 * visible backing of an io_uring instance. The application mmaps this
3786 * fd to gain access to the SQ/CQ ring details.
3788 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3790 return anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3791 O_RDWR | O_CLOEXEC, NULL);
3794 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3795 struct io_uring_params __user *params)
3797 struct io_ring_ctx *ctx;
3798 struct io_uring_task *tctx;
3804 if (entries > IORING_MAX_ENTRIES) {
3805 if (!(p->flags & IORING_SETUP_CLAMP))
3807 entries = IORING_MAX_ENTRIES;
3810 if ((p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
3811 && !(p->flags & IORING_SETUP_NO_MMAP))
3815 * Use twice as many entries for the CQ ring. It's possible for the
3816 * application to drive a higher depth than the size of the SQ ring,
3817 * since the sqes are only used at submission time. This allows for
3818 * some flexibility in overcommitting a bit. If the application has
3819 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3820 * of CQ ring entries manually.
3822 p->sq_entries = roundup_pow_of_two(entries);
3823 if (p->flags & IORING_SETUP_CQSIZE) {
3825 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3826 * to a power-of-two, if it isn't already. We do NOT impose
3827 * any cq vs sq ring sizing.
3831 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3832 if (!(p->flags & IORING_SETUP_CLAMP))
3834 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3836 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3837 if (p->cq_entries < p->sq_entries)
3840 p->cq_entries = 2 * p->sq_entries;
3843 ctx = io_ring_ctx_alloc(p);
3847 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3848 !(ctx->flags & IORING_SETUP_IOPOLL) &&
3849 !(ctx->flags & IORING_SETUP_SQPOLL))
3850 ctx->task_complete = true;
3852 if (ctx->task_complete || (ctx->flags & IORING_SETUP_IOPOLL))
3853 ctx->lockless_cq = true;
3856 * lazy poll_wq activation relies on ->task_complete for synchronisation
3857 * purposes, see io_activate_pollwq()
3859 if (!ctx->task_complete)
3860 ctx->poll_activated = true;
3863 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3864 * space applications don't need to do io completion events
3865 * polling again, they can rely on io_sq_thread to do polling
3866 * work, which can reduce cpu usage and uring_lock contention.
3868 if (ctx->flags & IORING_SETUP_IOPOLL &&
3869 !(ctx->flags & IORING_SETUP_SQPOLL))
3870 ctx->syscall_iopoll = 1;
3872 ctx->compat = in_compat_syscall();
3873 if (!ns_capable_noaudit(&init_user_ns, CAP_IPC_LOCK))
3874 ctx->user = get_uid(current_user());
3877 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3878 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3881 if (ctx->flags & IORING_SETUP_SQPOLL) {
3882 /* IPI related flags don't make sense with SQPOLL */
3883 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3884 IORING_SETUP_TASKRUN_FLAG |
3885 IORING_SETUP_DEFER_TASKRUN))
3887 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3888 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3889 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3891 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3892 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3894 ctx->notify_method = TWA_SIGNAL;
3898 * For DEFER_TASKRUN we require the completion task to be the same as the
3899 * submission task. This implies that there is only one submitter, so enforce
3902 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3903 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3908 * This is just grabbed for accounting purposes. When a process exits,
3909 * the mm is exited and dropped before the files, hence we need to hang
3910 * on to this mm purely for the purposes of being able to unaccount
3911 * memory (locked/pinned vm). It's not used for anything else.
3913 mmgrab(current->mm);
3914 ctx->mm_account = current->mm;
3916 ret = io_allocate_scq_urings(ctx, p);
3920 ret = io_sq_offload_create(ctx, p);
3924 ret = io_rsrc_init(ctx);
3928 p->sq_off.head = offsetof(struct io_rings, sq.head);
3929 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3930 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3931 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3932 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3933 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3934 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY))
3935 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3936 p->sq_off.resv1 = 0;
3937 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3938 p->sq_off.user_addr = 0;
3940 p->cq_off.head = offsetof(struct io_rings, cq.head);
3941 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3942 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3943 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3944 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3945 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3946 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3947 p->cq_off.resv1 = 0;
3948 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3949 p->cq_off.user_addr = 0;
3951 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3952 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3953 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3954 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3955 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3956 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3957 IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING;
3959 if (copy_to_user(params, p, sizeof(*p))) {
3964 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
3965 && !(ctx->flags & IORING_SETUP_R_DISABLED))
3966 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
3968 file = io_uring_get_file(ctx);
3970 ret = PTR_ERR(file);
3974 ret = __io_uring_add_tctx_node(ctx);
3977 tctx = current->io_uring;
3980 * Install ring fd as the very last thing, so we don't risk someone
3981 * having closed it before we finish setup
3983 if (p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
3984 ret = io_ring_add_registered_file(tctx, file, 0, IO_RINGFD_REG_MAX);
3986 ret = io_uring_install_fd(file);
3990 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3993 io_ring_ctx_wait_and_kill(ctx);
4001 * Sets up an aio uring context, and returns the fd. Applications asks for a
4002 * ring size, we return the actual sq/cq ring sizes (among other things) in the
4003 * params structure passed in.
4005 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
4007 struct io_uring_params p;
4010 if (copy_from_user(&p, params, sizeof(p)))
4012 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
4017 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
4018 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
4019 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
4020 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
4021 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
4022 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
4023 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN |
4024 IORING_SETUP_NO_MMAP | IORING_SETUP_REGISTERED_FD_ONLY |
4025 IORING_SETUP_NO_SQARRAY))
4028 return io_uring_create(entries, &p, params);
4031 static inline bool io_uring_allowed(void)
4033 int disabled = READ_ONCE(sysctl_io_uring_disabled);
4034 kgid_t io_uring_group;
4039 if (disabled == 0 || capable(CAP_SYS_ADMIN))
4042 io_uring_group = make_kgid(&init_user_ns, sysctl_io_uring_group);
4043 if (!gid_valid(io_uring_group))
4046 return in_group_p(io_uring_group);
4049 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
4050 struct io_uring_params __user *, params)
4052 if (!io_uring_allowed())
4055 return io_uring_setup(entries, params);
4058 static int __init io_uring_init(void)
4060 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4061 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4062 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4065 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4066 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4067 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4068 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4069 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4070 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
4071 BUILD_BUG_SQE_ELEM(1, __u8, flags);
4072 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
4073 BUILD_BUG_SQE_ELEM(4, __s32, fd);
4074 BUILD_BUG_SQE_ELEM(8, __u64, off);
4075 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
4076 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
4077 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
4078 BUILD_BUG_SQE_ELEM(16, __u64, addr);
4079 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
4080 BUILD_BUG_SQE_ELEM(24, __u32, len);
4081 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
4082 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
4083 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4084 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
4085 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
4086 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
4087 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
4088 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
4089 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
4090 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
4091 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
4092 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
4093 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
4094 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
4095 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
4096 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
4097 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
4098 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
4099 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
4100 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
4101 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
4102 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
4103 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
4104 BUILD_BUG_SQE_ELEM(42, __u16, personality);
4105 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
4106 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
4107 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
4108 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
4109 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
4110 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4111 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
4113 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4114 sizeof(struct io_uring_rsrc_update));
4115 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4116 sizeof(struct io_uring_rsrc_update2));
4118 /* ->buf_index is u16 */
4119 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4120 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4121 offsetof(struct io_uring_buf_ring, tail));
4123 /* should fit into one byte */
4124 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
4125 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4126 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4128 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
4130 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4132 /* top 8bits are for internal use */
4133 BUILD_BUG_ON((IORING_URING_CMD_MASK & 0xff000000) != 0);
4135 io_uring_optable_init();
4138 * Allow user copy in the per-command field, which starts after the
4139 * file in io_kiocb and until the opcode field. The openat2 handling
4140 * requires copying in user memory into the io_kiocb object in that
4141 * range, and HARDENED_USERCOPY will complain if we haven't
4142 * correctly annotated this range.
4144 req_cachep = kmem_cache_create_usercopy("io_kiocb",
4145 sizeof(struct io_kiocb), 0,
4146 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4147 SLAB_ACCOUNT | SLAB_TYPESAFE_BY_RCU,
4148 offsetof(struct io_kiocb, cmd.data),
4149 sizeof_field(struct io_kiocb, cmd.data), NULL);
4150 io_buf_cachep = kmem_cache_create("io_buffer", sizeof(struct io_buffer), 0,
4151 SLAB_HWCACHE_ALIGN | SLAB_PANIC | SLAB_ACCOUNT,
4154 #ifdef CONFIG_SYSCTL
4155 register_sysctl_init("kernel", kernel_io_uring_disabled_table);
4160 __initcall(io_uring_init);