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 <linux/anon_inodes.h>
63 #include <linux/sched/mm.h>
64 #include <linux/uaccess.h>
65 #include <linux/nospec.h>
66 #include <linux/highmem.h>
67 #include <linux/fsnotify.h>
68 #include <linux/fadvise.h>
69 #include <linux/task_work.h>
70 #include <linux/io_uring.h>
71 #include <linux/io_uring/cmd.h>
72 #include <linux/audit.h>
73 #include <linux/security.h>
74 #include <asm/shmparam.h>
76 #define CREATE_TRACE_POINTS
77 #include <trace/events/io_uring.h>
79 #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
125 struct io_defer_entry {
126 struct list_head list;
127 struct io_kiocb *req;
131 /* requests with any of those set should undergo io_disarm_next() */
132 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
133 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
136 * No waiters. It's larger than any valid value of the tw counter
137 * so that tests against ->cq_wait_nr would fail and skip wake_up().
139 #define IO_CQ_WAKE_INIT (-1U)
140 /* Forced wake up if there is a waiter regardless of ->cq_wait_nr */
141 #define IO_CQ_WAKE_FORCE (IO_CQ_WAKE_INIT >> 1)
143 static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
144 struct task_struct *task,
147 static void io_queue_sqe(struct io_kiocb *req);
149 struct kmem_cache *req_cachep;
150 static struct workqueue_struct *iou_wq __ro_after_init;
152 static int __read_mostly sysctl_io_uring_disabled;
153 static int __read_mostly sysctl_io_uring_group = -1;
156 static struct ctl_table kernel_io_uring_disabled_table[] = {
158 .procname = "io_uring_disabled",
159 .data = &sysctl_io_uring_disabled,
160 .maxlen = sizeof(sysctl_io_uring_disabled),
162 .proc_handler = proc_dointvec_minmax,
163 .extra1 = SYSCTL_ZERO,
164 .extra2 = SYSCTL_TWO,
167 .procname = "io_uring_group",
168 .data = &sysctl_io_uring_group,
169 .maxlen = sizeof(gid_t),
171 .proc_handler = proc_dointvec,
177 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
179 if (!wq_list_empty(&ctx->submit_state.compl_reqs) ||
180 ctx->submit_state.cqes_count)
181 __io_submit_flush_completions(ctx);
184 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
186 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
189 static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx)
191 return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head);
194 static bool io_match_linked(struct io_kiocb *head)
196 struct io_kiocb *req;
198 io_for_each_link(req, head) {
199 if (req->flags & REQ_F_INFLIGHT)
206 * As io_match_task() but protected against racing with linked timeouts.
207 * User must not hold timeout_lock.
209 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
214 if (task && head->task != task)
219 if (head->flags & REQ_F_LINK_TIMEOUT) {
220 struct io_ring_ctx *ctx = head->ctx;
222 /* protect against races with linked timeouts */
223 spin_lock_irq(&ctx->timeout_lock);
224 matched = io_match_linked(head);
225 spin_unlock_irq(&ctx->timeout_lock);
227 matched = io_match_linked(head);
232 static inline void req_fail_link_node(struct io_kiocb *req, int res)
235 io_req_set_res(req, res, 0);
238 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
240 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
243 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
245 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
247 complete(&ctx->ref_comp);
250 static __cold void io_fallback_req_func(struct work_struct *work)
252 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
254 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
255 struct io_kiocb *req, *tmp;
256 struct io_tw_state ts = { .locked = true, };
258 percpu_ref_get(&ctx->refs);
259 mutex_lock(&ctx->uring_lock);
260 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
261 req->io_task_work.func(req, &ts);
262 if (WARN_ON_ONCE(!ts.locked))
264 io_submit_flush_completions(ctx);
265 mutex_unlock(&ctx->uring_lock);
266 percpu_ref_put(&ctx->refs);
269 static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
271 unsigned hash_buckets = 1U << bits;
272 size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
274 table->hbs = kmalloc(hash_size, GFP_KERNEL);
278 table->hash_bits = bits;
279 init_hash_table(table, hash_buckets);
283 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
285 struct io_ring_ctx *ctx;
288 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
292 xa_init(&ctx->io_bl_xa);
295 * Use 5 bits less than the max cq entries, that should give us around
296 * 32 entries per hash list if totally full and uniformly spread, but
297 * don't keep too many buckets to not overconsume memory.
299 hash_bits = ilog2(p->cq_entries) - 5;
300 hash_bits = clamp(hash_bits, 1, 8);
301 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
303 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
305 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
309 ctx->flags = p->flags;
310 atomic_set(&ctx->cq_wait_nr, IO_CQ_WAKE_INIT);
311 init_waitqueue_head(&ctx->sqo_sq_wait);
312 INIT_LIST_HEAD(&ctx->sqd_list);
313 INIT_LIST_HEAD(&ctx->cq_overflow_list);
314 INIT_LIST_HEAD(&ctx->io_buffers_cache);
315 INIT_HLIST_HEAD(&ctx->io_buf_list);
316 io_alloc_cache_init(&ctx->rsrc_node_cache, IO_NODE_ALLOC_CACHE_MAX,
317 sizeof(struct io_rsrc_node));
318 io_alloc_cache_init(&ctx->apoll_cache, IO_ALLOC_CACHE_MAX,
319 sizeof(struct async_poll));
320 io_alloc_cache_init(&ctx->netmsg_cache, IO_ALLOC_CACHE_MAX,
321 sizeof(struct io_async_msghdr));
322 io_futex_cache_init(ctx);
323 init_completion(&ctx->ref_comp);
324 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
325 mutex_init(&ctx->uring_lock);
326 init_waitqueue_head(&ctx->cq_wait);
327 init_waitqueue_head(&ctx->poll_wq);
328 init_waitqueue_head(&ctx->rsrc_quiesce_wq);
329 spin_lock_init(&ctx->completion_lock);
330 spin_lock_init(&ctx->timeout_lock);
331 INIT_WQ_LIST(&ctx->iopoll_list);
332 INIT_LIST_HEAD(&ctx->io_buffers_comp);
333 INIT_LIST_HEAD(&ctx->defer_list);
334 INIT_LIST_HEAD(&ctx->timeout_list);
335 INIT_LIST_HEAD(&ctx->ltimeout_list);
336 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
337 init_llist_head(&ctx->work_llist);
338 INIT_LIST_HEAD(&ctx->tctx_list);
339 ctx->submit_state.free_list.next = NULL;
340 INIT_WQ_LIST(&ctx->locked_free_list);
341 INIT_HLIST_HEAD(&ctx->waitid_list);
343 INIT_HLIST_HEAD(&ctx->futex_list);
345 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
346 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
347 INIT_HLIST_HEAD(&ctx->cancelable_uring_cmd);
352 kfree(ctx->cancel_table.hbs);
353 kfree(ctx->cancel_table_locked.hbs);
354 xa_destroy(&ctx->io_bl_xa);
359 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
361 struct io_rings *r = ctx->rings;
363 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
367 static bool req_need_defer(struct io_kiocb *req, u32 seq)
369 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
370 struct io_ring_ctx *ctx = req->ctx;
372 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
378 static void io_clean_op(struct io_kiocb *req)
380 if (req->flags & REQ_F_BUFFER_SELECTED) {
381 spin_lock(&req->ctx->completion_lock);
382 io_put_kbuf_comp(req);
383 spin_unlock(&req->ctx->completion_lock);
386 if (req->flags & REQ_F_NEED_CLEANUP) {
387 const struct io_cold_def *def = &io_cold_defs[req->opcode];
392 if ((req->flags & REQ_F_POLLED) && req->apoll) {
393 kfree(req->apoll->double_poll);
397 if (req->flags & REQ_F_INFLIGHT) {
398 struct io_uring_task *tctx = req->task->io_uring;
400 atomic_dec(&tctx->inflight_tracked);
402 if (req->flags & REQ_F_CREDS)
403 put_cred(req->creds);
404 if (req->flags & REQ_F_ASYNC_DATA) {
405 kfree(req->async_data);
406 req->async_data = NULL;
408 req->flags &= ~IO_REQ_CLEAN_FLAGS;
411 static inline void io_req_track_inflight(struct io_kiocb *req)
413 if (!(req->flags & REQ_F_INFLIGHT)) {
414 req->flags |= REQ_F_INFLIGHT;
415 atomic_inc(&req->task->io_uring->inflight_tracked);
419 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
421 if (WARN_ON_ONCE(!req->link))
424 req->flags &= ~REQ_F_ARM_LTIMEOUT;
425 req->flags |= REQ_F_LINK_TIMEOUT;
427 /* linked timeouts should have two refs once prep'ed */
428 io_req_set_refcount(req);
429 __io_req_set_refcount(req->link, 2);
433 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
435 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
437 return __io_prep_linked_timeout(req);
440 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
442 io_queue_linked_timeout(__io_prep_linked_timeout(req));
445 static inline void io_arm_ltimeout(struct io_kiocb *req)
447 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
448 __io_arm_ltimeout(req);
451 static void io_prep_async_work(struct io_kiocb *req)
453 const struct io_issue_def *def = &io_issue_defs[req->opcode];
454 struct io_ring_ctx *ctx = req->ctx;
456 if (!(req->flags & REQ_F_CREDS)) {
457 req->flags |= REQ_F_CREDS;
458 req->creds = get_current_cred();
461 req->work.list.next = NULL;
463 if (req->flags & REQ_F_FORCE_ASYNC)
464 req->work.flags |= IO_WQ_WORK_CONCURRENT;
466 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
467 req->flags |= io_file_get_flags(req->file);
469 if (req->file && (req->flags & REQ_F_ISREG)) {
470 bool should_hash = def->hash_reg_file;
472 /* don't serialize this request if the fs doesn't need it */
473 if (should_hash && (req->file->f_flags & O_DIRECT) &&
474 (req->file->f_mode & FMODE_DIO_PARALLEL_WRITE))
476 if (should_hash || (ctx->flags & IORING_SETUP_IOPOLL))
477 io_wq_hash_work(&req->work, file_inode(req->file));
478 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
479 if (def->unbound_nonreg_file)
480 req->work.flags |= IO_WQ_WORK_UNBOUND;
484 static void io_prep_async_link(struct io_kiocb *req)
486 struct io_kiocb *cur;
488 if (req->flags & REQ_F_LINK_TIMEOUT) {
489 struct io_ring_ctx *ctx = req->ctx;
491 spin_lock_irq(&ctx->timeout_lock);
492 io_for_each_link(cur, req)
493 io_prep_async_work(cur);
494 spin_unlock_irq(&ctx->timeout_lock);
496 io_for_each_link(cur, req)
497 io_prep_async_work(cur);
501 void io_queue_iowq(struct io_kiocb *req, struct io_tw_state *ts_dont_use)
503 struct io_kiocb *link = io_prep_linked_timeout(req);
504 struct io_uring_task *tctx = req->task->io_uring;
507 BUG_ON(!tctx->io_wq);
509 /* init ->work of the whole link before punting */
510 io_prep_async_link(req);
513 * Not expected to happen, but if we do have a bug where this _can_
514 * happen, catch it here and ensure the request is marked as
515 * canceled. That will make io-wq go through the usual work cancel
516 * procedure rather than attempt to run this request (or create a new
519 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
520 req->work.flags |= IO_WQ_WORK_CANCEL;
522 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
523 io_wq_enqueue(tctx->io_wq, &req->work);
525 io_queue_linked_timeout(link);
528 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
530 while (!list_empty(&ctx->defer_list)) {
531 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
532 struct io_defer_entry, list);
534 if (req_need_defer(de->req, de->seq))
536 list_del_init(&de->list);
537 io_req_task_queue(de->req);
542 void io_eventfd_ops(struct rcu_head *rcu)
544 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
545 int ops = atomic_xchg(&ev_fd->ops, 0);
547 if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT))
548 eventfd_signal_mask(ev_fd->cq_ev_fd, EPOLL_URING_WAKE);
550 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
551 * ordering in a race but if references are 0 we know we have to free
554 if (atomic_dec_and_test(&ev_fd->refs)) {
555 eventfd_ctx_put(ev_fd->cq_ev_fd);
560 static void io_eventfd_signal(struct io_ring_ctx *ctx)
562 struct io_ev_fd *ev_fd = NULL;
566 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
569 ev_fd = rcu_dereference(ctx->io_ev_fd);
572 * Check again if ev_fd exists incase an io_eventfd_unregister call
573 * completed between the NULL check of ctx->io_ev_fd at the start of
574 * the function and rcu_read_lock.
576 if (unlikely(!ev_fd))
578 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
580 if (ev_fd->eventfd_async && !io_wq_current_is_worker())
583 if (likely(eventfd_signal_allowed())) {
584 eventfd_signal_mask(ev_fd->cq_ev_fd, EPOLL_URING_WAKE);
586 atomic_inc(&ev_fd->refs);
587 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops))
588 call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops);
590 atomic_dec(&ev_fd->refs);
597 static void io_eventfd_flush_signal(struct io_ring_ctx *ctx)
601 spin_lock(&ctx->completion_lock);
604 * Eventfd should only get triggered when at least one event has been
605 * posted. Some applications rely on the eventfd notification count
606 * only changing IFF a new CQE has been added to the CQ ring. There's
607 * no depedency on 1:1 relationship between how many times this
608 * function is called (and hence the eventfd count) and number of CQEs
609 * posted to the CQ ring.
611 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
612 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
613 spin_unlock(&ctx->completion_lock);
617 io_eventfd_signal(ctx);
620 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
622 if (ctx->poll_activated)
623 io_poll_wq_wake(ctx);
624 if (ctx->off_timeout_used)
625 io_flush_timeouts(ctx);
626 if (ctx->drain_active) {
627 spin_lock(&ctx->completion_lock);
628 io_queue_deferred(ctx);
629 spin_unlock(&ctx->completion_lock);
632 io_eventfd_flush_signal(ctx);
635 static inline void __io_cq_lock(struct io_ring_ctx *ctx)
637 if (!ctx->lockless_cq)
638 spin_lock(&ctx->completion_lock);
641 static inline void io_cq_lock(struct io_ring_ctx *ctx)
642 __acquires(ctx->completion_lock)
644 spin_lock(&ctx->completion_lock);
647 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
649 io_commit_cqring(ctx);
650 if (!ctx->task_complete) {
651 if (!ctx->lockless_cq)
652 spin_unlock(&ctx->completion_lock);
653 /* IOPOLL rings only need to wake up if it's also SQPOLL */
654 if (!ctx->syscall_iopoll)
657 io_commit_cqring_flush(ctx);
660 static void io_cq_unlock_post(struct io_ring_ctx *ctx)
661 __releases(ctx->completion_lock)
663 io_commit_cqring(ctx);
664 spin_unlock(&ctx->completion_lock);
666 io_commit_cqring_flush(ctx);
669 static void io_cqring_overflow_kill(struct io_ring_ctx *ctx)
671 struct io_overflow_cqe *ocqe;
674 spin_lock(&ctx->completion_lock);
675 list_splice_init(&ctx->cq_overflow_list, &list);
676 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
677 spin_unlock(&ctx->completion_lock);
679 while (!list_empty(&list)) {
680 ocqe = list_first_entry(&list, struct io_overflow_cqe, list);
681 list_del(&ocqe->list);
686 static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx)
688 size_t cqe_size = sizeof(struct io_uring_cqe);
690 if (__io_cqring_events(ctx) == ctx->cq_entries)
693 if (ctx->flags & IORING_SETUP_CQE32)
697 while (!list_empty(&ctx->cq_overflow_list)) {
698 struct io_uring_cqe *cqe;
699 struct io_overflow_cqe *ocqe;
701 if (!io_get_cqe_overflow(ctx, &cqe, true))
703 ocqe = list_first_entry(&ctx->cq_overflow_list,
704 struct io_overflow_cqe, list);
705 memcpy(cqe, &ocqe->cqe, cqe_size);
706 list_del(&ocqe->list);
710 if (list_empty(&ctx->cq_overflow_list)) {
711 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
712 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
714 io_cq_unlock_post(ctx);
717 static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx)
719 /* iopoll syncs against uring_lock, not completion_lock */
720 if (ctx->flags & IORING_SETUP_IOPOLL)
721 mutex_lock(&ctx->uring_lock);
722 __io_cqring_overflow_flush(ctx);
723 if (ctx->flags & IORING_SETUP_IOPOLL)
724 mutex_unlock(&ctx->uring_lock);
727 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx)
729 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
730 io_cqring_do_overflow_flush(ctx);
733 /* can be called by any task */
734 static void io_put_task_remote(struct task_struct *task)
736 struct io_uring_task *tctx = task->io_uring;
738 percpu_counter_sub(&tctx->inflight, 1);
739 if (unlikely(atomic_read(&tctx->in_cancel)))
740 wake_up(&tctx->wait);
741 put_task_struct(task);
744 /* used by a task to put its own references */
745 static void io_put_task_local(struct task_struct *task)
747 task->io_uring->cached_refs++;
750 /* must to be called somewhat shortly after putting a request */
751 static inline void io_put_task(struct task_struct *task)
753 if (likely(task == current))
754 io_put_task_local(task);
756 io_put_task_remote(task);
759 void io_task_refs_refill(struct io_uring_task *tctx)
761 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
763 percpu_counter_add(&tctx->inflight, refill);
764 refcount_add(refill, ¤t->usage);
765 tctx->cached_refs += refill;
768 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
770 struct io_uring_task *tctx = task->io_uring;
771 unsigned int refs = tctx->cached_refs;
774 tctx->cached_refs = 0;
775 percpu_counter_sub(&tctx->inflight, refs);
776 put_task_struct_many(task, refs);
780 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
781 s32 res, u32 cflags, u64 extra1, u64 extra2)
783 struct io_overflow_cqe *ocqe;
784 size_t ocq_size = sizeof(struct io_overflow_cqe);
785 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
787 lockdep_assert_held(&ctx->completion_lock);
790 ocq_size += sizeof(struct io_uring_cqe);
792 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
793 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
796 * If we're in ring overflow flush mode, or in task cancel mode,
797 * or cannot allocate an overflow entry, then we need to drop it
800 io_account_cq_overflow(ctx);
801 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
804 if (list_empty(&ctx->cq_overflow_list)) {
805 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
806 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
809 ocqe->cqe.user_data = user_data;
811 ocqe->cqe.flags = cflags;
813 ocqe->cqe.big_cqe[0] = extra1;
814 ocqe->cqe.big_cqe[1] = extra2;
816 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
820 void io_req_cqe_overflow(struct io_kiocb *req)
822 io_cqring_event_overflow(req->ctx, req->cqe.user_data,
823 req->cqe.res, req->cqe.flags,
824 req->big_cqe.extra1, req->big_cqe.extra2);
825 memset(&req->big_cqe, 0, sizeof(req->big_cqe));
829 * writes to the cq entry need to come after reading head; the
830 * control dependency is enough as we're using WRITE_ONCE to
833 bool io_cqe_cache_refill(struct io_ring_ctx *ctx, bool overflow)
835 struct io_rings *rings = ctx->rings;
836 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
837 unsigned int free, queued, len;
840 * Posting into the CQ when there are pending overflowed CQEs may break
841 * ordering guarantees, which will affect links, F_MORE users and more.
842 * Force overflow the completion.
844 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
847 /* userspace may cheat modifying the tail, be safe and do min */
848 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
849 free = ctx->cq_entries - queued;
850 /* we need a contiguous range, limit based on the current array offset */
851 len = min(free, ctx->cq_entries - off);
855 if (ctx->flags & IORING_SETUP_CQE32) {
860 ctx->cqe_cached = &rings->cqes[off];
861 ctx->cqe_sentinel = ctx->cqe_cached + len;
865 static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
868 struct io_uring_cqe *cqe;
873 * If we can't get a cq entry, userspace overflowed the
874 * submission (by quite a lot). Increment the overflow count in
877 if (likely(io_get_cqe(ctx, &cqe))) {
878 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
880 WRITE_ONCE(cqe->user_data, user_data);
881 WRITE_ONCE(cqe->res, res);
882 WRITE_ONCE(cqe->flags, cflags);
884 if (ctx->flags & IORING_SETUP_CQE32) {
885 WRITE_ONCE(cqe->big_cqe[0], 0);
886 WRITE_ONCE(cqe->big_cqe[1], 0);
893 static void __io_flush_post_cqes(struct io_ring_ctx *ctx)
894 __must_hold(&ctx->uring_lock)
896 struct io_submit_state *state = &ctx->submit_state;
899 lockdep_assert_held(&ctx->uring_lock);
900 for (i = 0; i < state->cqes_count; i++) {
901 struct io_uring_cqe *cqe = &ctx->completion_cqes[i];
903 if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) {
904 if (ctx->lockless_cq) {
905 spin_lock(&ctx->completion_lock);
906 io_cqring_event_overflow(ctx, cqe->user_data,
907 cqe->res, cqe->flags, 0, 0);
908 spin_unlock(&ctx->completion_lock);
910 io_cqring_event_overflow(ctx, cqe->user_data,
911 cqe->res, cqe->flags, 0, 0);
915 state->cqes_count = 0;
918 static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
924 filled = io_fill_cqe_aux(ctx, user_data, res, cflags);
925 if (!filled && allow_overflow)
926 filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
928 io_cq_unlock_post(ctx);
932 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags)
934 return __io_post_aux_cqe(ctx, user_data, res, cflags, true);
938 * A helper for multishot requests posting additional CQEs.
939 * Should only be used from a task_work including IO_URING_F_MULTISHOT.
941 bool io_fill_cqe_req_aux(struct io_kiocb *req, bool defer, s32 res, u32 cflags)
943 struct io_ring_ctx *ctx = req->ctx;
944 u64 user_data = req->cqe.user_data;
945 struct io_uring_cqe *cqe;
947 lockdep_assert(!io_wq_current_is_worker());
950 return __io_post_aux_cqe(ctx, user_data, res, cflags, false);
952 lockdep_assert_held(&ctx->uring_lock);
954 if (ctx->submit_state.cqes_count == ARRAY_SIZE(ctx->completion_cqes)) {
956 __io_flush_post_cqes(ctx);
957 /* no need to flush - flush is deferred */
958 __io_cq_unlock_post(ctx);
961 /* For defered completions this is not as strict as it is otherwise,
962 * however it's main job is to prevent unbounded posted completions,
963 * and in that it works just as well.
965 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
968 cqe = &ctx->completion_cqes[ctx->submit_state.cqes_count++];
969 cqe->user_data = user_data;
975 static void __io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
977 struct io_ring_ctx *ctx = req->ctx;
978 struct io_rsrc_node *rsrc_node = NULL;
981 if (!(req->flags & REQ_F_CQE_SKIP)) {
982 if (!io_fill_cqe_req(ctx, req))
983 io_req_cqe_overflow(req);
987 * If we're the last reference to this request, add to our locked
990 if (req_ref_put_and_test(req)) {
991 if (req->flags & IO_REQ_LINK_FLAGS) {
992 if (req->flags & IO_DISARM_MASK)
995 io_req_task_queue(req->link);
999 io_put_kbuf_comp(req);
1000 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1004 rsrc_node = req->rsrc_node;
1006 * Selected buffer deallocation in io_clean_op() assumes that
1007 * we don't hold ->completion_lock. Clean them here to avoid
1010 io_put_task_remote(req->task);
1011 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
1012 ctx->locked_free_nr++;
1014 io_cq_unlock_post(ctx);
1017 io_ring_submit_lock(ctx, issue_flags);
1018 io_put_rsrc_node(ctx, rsrc_node);
1019 io_ring_submit_unlock(ctx, issue_flags);
1023 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
1025 struct io_ring_ctx *ctx = req->ctx;
1027 if (ctx->task_complete && ctx->submitter_task != current) {
1028 req->io_task_work.func = io_req_task_complete;
1029 io_req_task_work_add(req);
1030 } else if (!(issue_flags & IO_URING_F_UNLOCKED) ||
1031 !(ctx->flags & IORING_SETUP_IOPOLL)) {
1032 __io_req_complete_post(req, issue_flags);
1034 mutex_lock(&ctx->uring_lock);
1035 __io_req_complete_post(req, issue_flags & ~IO_URING_F_UNLOCKED);
1036 mutex_unlock(&ctx->uring_lock);
1040 void io_req_defer_failed(struct io_kiocb *req, s32 res)
1041 __must_hold(&ctx->uring_lock)
1043 const struct io_cold_def *def = &io_cold_defs[req->opcode];
1045 lockdep_assert_held(&req->ctx->uring_lock);
1048 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
1051 io_req_complete_defer(req);
1055 * Don't initialise the fields below on every allocation, but do that in
1056 * advance and keep them valid across allocations.
1058 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1062 req->async_data = NULL;
1063 /* not necessary, but safer to zero */
1064 memset(&req->cqe, 0, sizeof(req->cqe));
1065 memset(&req->big_cqe, 0, sizeof(req->big_cqe));
1068 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1069 struct io_submit_state *state)
1071 spin_lock(&ctx->completion_lock);
1072 wq_list_splice(&ctx->locked_free_list, &state->free_list);
1073 ctx->locked_free_nr = 0;
1074 spin_unlock(&ctx->completion_lock);
1078 * A request might get retired back into the request caches even before opcode
1079 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1080 * Because of that, io_alloc_req() should be called only under ->uring_lock
1081 * and with extra caution to not get a request that is still worked on.
1083 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
1084 __must_hold(&ctx->uring_lock)
1086 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1087 void *reqs[IO_REQ_ALLOC_BATCH];
1091 * If we have more than a batch's worth of requests in our IRQ side
1092 * locked cache, grab the lock and move them over to our submission
1095 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
1096 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
1097 if (!io_req_cache_empty(ctx))
1101 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
1104 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1105 * retry single alloc to be on the safe side.
1107 if (unlikely(ret <= 0)) {
1108 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1114 percpu_ref_get_many(&ctx->refs, ret);
1115 for (i = 0; i < ret; i++) {
1116 struct io_kiocb *req = reqs[i];
1118 io_preinit_req(req, ctx);
1119 io_req_add_to_cache(req, ctx);
1124 __cold void io_free_req(struct io_kiocb *req)
1126 /* refs were already put, restore them for io_req_task_complete() */
1127 req->flags &= ~REQ_F_REFCOUNT;
1128 /* we only want to free it, don't post CQEs */
1129 req->flags |= REQ_F_CQE_SKIP;
1130 req->io_task_work.func = io_req_task_complete;
1131 io_req_task_work_add(req);
1134 static void __io_req_find_next_prep(struct io_kiocb *req)
1136 struct io_ring_ctx *ctx = req->ctx;
1138 spin_lock(&ctx->completion_lock);
1139 io_disarm_next(req);
1140 spin_unlock(&ctx->completion_lock);
1143 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1145 struct io_kiocb *nxt;
1148 * If LINK is set, we have dependent requests in this chain. If we
1149 * didn't fail this request, queue the first one up, moving any other
1150 * dependencies to the next request. In case of failure, fail the rest
1153 if (unlikely(req->flags & IO_DISARM_MASK))
1154 __io_req_find_next_prep(req);
1160 static void ctx_flush_and_put(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1164 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1165 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1167 io_submit_flush_completions(ctx);
1168 mutex_unlock(&ctx->uring_lock);
1171 percpu_ref_put(&ctx->refs);
1175 * Run queued task_work, returning the number of entries processed in *count.
1176 * If more entries than max_entries are available, stop processing once this
1177 * is reached and return the rest of the list.
1179 struct llist_node *io_handle_tw_list(struct llist_node *node,
1180 unsigned int *count,
1181 unsigned int max_entries)
1183 struct io_ring_ctx *ctx = NULL;
1184 struct io_tw_state ts = { };
1187 struct llist_node *next = node->next;
1188 struct io_kiocb *req = container_of(node, struct io_kiocb,
1191 if (req->ctx != ctx) {
1192 ctx_flush_and_put(ctx, &ts);
1194 /* if not contended, grab and improve batching */
1195 ts.locked = mutex_trylock(&ctx->uring_lock);
1196 percpu_ref_get(&ctx->refs);
1198 INDIRECT_CALL_2(req->io_task_work.func,
1199 io_poll_task_func, io_req_rw_complete,
1203 if (unlikely(need_resched())) {
1204 ctx_flush_and_put(ctx, &ts);
1208 } while (node && *count < max_entries);
1210 ctx_flush_and_put(ctx, &ts);
1215 * io_llist_xchg - swap all entries in a lock-less list
1216 * @head: the head of lock-less list to delete all entries
1217 * @new: new entry as the head of the list
1219 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1220 * The order of entries returned is from the newest to the oldest added one.
1222 static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1223 struct llist_node *new)
1225 return xchg(&head->first, new);
1228 static __cold void io_fallback_tw(struct io_uring_task *tctx, bool sync)
1230 struct llist_node *node = llist_del_all(&tctx->task_list);
1231 struct io_ring_ctx *last_ctx = NULL;
1232 struct io_kiocb *req;
1235 req = container_of(node, struct io_kiocb, io_task_work.node);
1237 if (sync && last_ctx != req->ctx) {
1239 flush_delayed_work(&last_ctx->fallback_work);
1240 percpu_ref_put(&last_ctx->refs);
1242 last_ctx = req->ctx;
1243 percpu_ref_get(&last_ctx->refs);
1245 if (llist_add(&req->io_task_work.node,
1246 &req->ctx->fallback_llist))
1247 schedule_delayed_work(&req->ctx->fallback_work, 1);
1251 flush_delayed_work(&last_ctx->fallback_work);
1252 percpu_ref_put(&last_ctx->refs);
1256 struct llist_node *tctx_task_work_run(struct io_uring_task *tctx,
1257 unsigned int max_entries,
1258 unsigned int *count)
1260 struct llist_node *node;
1262 if (unlikely(current->flags & PF_EXITING)) {
1263 io_fallback_tw(tctx, true);
1267 node = llist_del_all(&tctx->task_list);
1269 node = llist_reverse_order(node);
1270 node = io_handle_tw_list(node, count, max_entries);
1273 /* relaxed read is enough as only the task itself sets ->in_cancel */
1274 if (unlikely(atomic_read(&tctx->in_cancel)))
1275 io_uring_drop_tctx_refs(current);
1277 trace_io_uring_task_work_run(tctx, *count);
1281 void tctx_task_work(struct callback_head *cb)
1283 struct io_uring_task *tctx;
1284 struct llist_node *ret;
1285 unsigned int count = 0;
1287 tctx = container_of(cb, struct io_uring_task, task_work);
1288 ret = tctx_task_work_run(tctx, UINT_MAX, &count);
1293 static inline void io_req_local_work_add(struct io_kiocb *req, unsigned flags)
1295 struct io_ring_ctx *ctx = req->ctx;
1296 unsigned nr_wait, nr_tw, nr_tw_prev;
1297 struct llist_node *head;
1299 /* See comment above IO_CQ_WAKE_INIT */
1300 BUILD_BUG_ON(IO_CQ_WAKE_FORCE <= IORING_MAX_CQ_ENTRIES);
1303 * We don't know how many reuqests is there in the link and whether
1304 * they can even be queued lazily, fall back to non-lazy.
1306 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK))
1307 flags &= ~IOU_F_TWQ_LAZY_WAKE;
1309 head = READ_ONCE(ctx->work_llist.first);
1313 struct io_kiocb *first_req = container_of(head,
1317 * Might be executed at any moment, rely on
1318 * SLAB_TYPESAFE_BY_RCU to keep it alive.
1320 nr_tw_prev = READ_ONCE(first_req->nr_tw);
1324 * Theoretically, it can overflow, but that's fine as one of
1325 * previous adds should've tried to wake the task.
1327 nr_tw = nr_tw_prev + 1;
1328 if (!(flags & IOU_F_TWQ_LAZY_WAKE))
1329 nr_tw = IO_CQ_WAKE_FORCE;
1332 req->io_task_work.node.next = head;
1333 } while (!try_cmpxchg(&ctx->work_llist.first, &head,
1334 &req->io_task_work.node));
1337 * cmpxchg implies a full barrier, which pairs with the barrier
1338 * in set_current_state() on the io_cqring_wait() side. It's used
1339 * to ensure that either we see updated ->cq_wait_nr, or waiters
1340 * going to sleep will observe the work added to the list, which
1341 * is similar to the wait/wawke task state sync.
1345 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1346 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1348 io_eventfd_signal(ctx);
1351 nr_wait = atomic_read(&ctx->cq_wait_nr);
1352 /* not enough or no one is waiting */
1353 if (nr_tw < nr_wait)
1355 /* the previous add has already woken it up */
1356 if (nr_tw_prev >= nr_wait)
1358 wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE);
1361 static void io_req_normal_work_add(struct io_kiocb *req)
1363 struct io_uring_task *tctx = req->task->io_uring;
1364 struct io_ring_ctx *ctx = req->ctx;
1366 /* task_work already pending, we're done */
1367 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1370 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1371 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1373 /* SQPOLL doesn't need the task_work added, it'll run it itself */
1374 if (ctx->flags & IORING_SETUP_SQPOLL) {
1375 struct io_sq_data *sqd = ctx->sq_data;
1377 if (wq_has_sleeper(&sqd->wait))
1378 wake_up(&sqd->wait);
1382 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1385 io_fallback_tw(tctx, false);
1388 void __io_req_task_work_add(struct io_kiocb *req, unsigned flags)
1390 if (req->ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
1392 io_req_local_work_add(req, flags);
1395 io_req_normal_work_add(req);
1399 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1401 struct llist_node *node;
1403 node = llist_del_all(&ctx->work_llist);
1405 struct io_kiocb *req = container_of(node, struct io_kiocb,
1409 io_req_normal_work_add(req);
1413 static bool io_run_local_work_continue(struct io_ring_ctx *ctx, int events,
1416 if (llist_empty(&ctx->work_llist))
1418 if (events < min_events)
1420 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1421 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1425 static int __io_run_local_work(struct io_ring_ctx *ctx, struct io_tw_state *ts,
1428 struct llist_node *node;
1429 unsigned int loops = 0;
1432 if (WARN_ON_ONCE(ctx->submitter_task != current))
1434 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1435 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1438 * llists are in reverse order, flip it back the right way before
1439 * running the pending items.
1441 node = llist_reverse_order(io_llist_xchg(&ctx->work_llist, NULL));
1443 struct llist_node *next = node->next;
1444 struct io_kiocb *req = container_of(node, struct io_kiocb,
1446 INDIRECT_CALL_2(req->io_task_work.func,
1447 io_poll_task_func, io_req_rw_complete,
1454 if (io_run_local_work_continue(ctx, ret, min_events))
1457 io_submit_flush_completions(ctx);
1458 if (io_run_local_work_continue(ctx, ret, min_events))
1462 trace_io_uring_local_work_run(ctx, ret, loops);
1466 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx,
1469 struct io_tw_state ts = { .locked = true, };
1472 if (llist_empty(&ctx->work_llist))
1475 ret = __io_run_local_work(ctx, &ts, min_events);
1476 /* shouldn't happen! */
1477 if (WARN_ON_ONCE(!ts.locked))
1478 mutex_lock(&ctx->uring_lock);
1482 static int io_run_local_work(struct io_ring_ctx *ctx, int min_events)
1484 struct io_tw_state ts = {};
1487 ts.locked = mutex_trylock(&ctx->uring_lock);
1488 ret = __io_run_local_work(ctx, &ts, min_events);
1490 mutex_unlock(&ctx->uring_lock);
1495 static void io_req_task_cancel(struct io_kiocb *req, struct io_tw_state *ts)
1497 io_tw_lock(req->ctx, ts);
1498 io_req_defer_failed(req, req->cqe.res);
1501 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts)
1503 io_tw_lock(req->ctx, ts);
1504 /* req->task == current here, checking PF_EXITING is safe */
1505 if (unlikely(req->task->flags & PF_EXITING))
1506 io_req_defer_failed(req, -EFAULT);
1507 else if (req->flags & REQ_F_FORCE_ASYNC)
1508 io_queue_iowq(req, ts);
1513 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1515 io_req_set_res(req, ret, 0);
1516 req->io_task_work.func = io_req_task_cancel;
1517 io_req_task_work_add(req);
1520 void io_req_task_queue(struct io_kiocb *req)
1522 req->io_task_work.func = io_req_task_submit;
1523 io_req_task_work_add(req);
1526 void io_queue_next(struct io_kiocb *req)
1528 struct io_kiocb *nxt = io_req_find_next(req);
1531 io_req_task_queue(nxt);
1534 static void io_free_batch_list(struct io_ring_ctx *ctx,
1535 struct io_wq_work_node *node)
1536 __must_hold(&ctx->uring_lock)
1539 struct io_kiocb *req = container_of(node, struct io_kiocb,
1542 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1543 if (req->flags & REQ_F_REFCOUNT) {
1544 node = req->comp_list.next;
1545 if (!req_ref_put_and_test(req))
1548 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1549 struct async_poll *apoll = req->apoll;
1551 if (apoll->double_poll)
1552 kfree(apoll->double_poll);
1553 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1555 req->flags &= ~REQ_F_POLLED;
1557 if (req->flags & IO_REQ_LINK_FLAGS)
1559 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1564 io_req_put_rsrc_locked(req, ctx);
1566 io_put_task(req->task);
1567 node = req->comp_list.next;
1568 io_req_add_to_cache(req, ctx);
1572 void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1573 __must_hold(&ctx->uring_lock)
1575 struct io_submit_state *state = &ctx->submit_state;
1576 struct io_wq_work_node *node;
1579 /* must come first to preserve CQE ordering in failure cases */
1580 if (state->cqes_count)
1581 __io_flush_post_cqes(ctx);
1582 __wq_list_for_each(node, &state->compl_reqs) {
1583 struct io_kiocb *req = container_of(node, struct io_kiocb,
1586 if (!(req->flags & REQ_F_CQE_SKIP) &&
1587 unlikely(!io_fill_cqe_req(ctx, req))) {
1588 if (ctx->lockless_cq) {
1589 spin_lock(&ctx->completion_lock);
1590 io_req_cqe_overflow(req);
1591 spin_unlock(&ctx->completion_lock);
1593 io_req_cqe_overflow(req);
1597 __io_cq_unlock_post(ctx);
1599 if (!wq_list_empty(&ctx->submit_state.compl_reqs)) {
1600 io_free_batch_list(ctx, state->compl_reqs.first);
1601 INIT_WQ_LIST(&state->compl_reqs);
1605 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1607 /* See comment at the top of this file */
1609 return __io_cqring_events(ctx);
1613 * We can't just wait for polled events to come to us, we have to actively
1614 * find and complete them.
1616 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1618 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1621 mutex_lock(&ctx->uring_lock);
1622 while (!wq_list_empty(&ctx->iopoll_list)) {
1623 /* let it sleep and repeat later if can't complete a request */
1624 if (io_do_iopoll(ctx, true) == 0)
1627 * Ensure we allow local-to-the-cpu processing to take place,
1628 * in this case we need to ensure that we reap all events.
1629 * Also let task_work, etc. to progress by releasing the mutex
1631 if (need_resched()) {
1632 mutex_unlock(&ctx->uring_lock);
1634 mutex_lock(&ctx->uring_lock);
1637 mutex_unlock(&ctx->uring_lock);
1640 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1642 unsigned int nr_events = 0;
1643 unsigned long check_cq;
1645 if (!io_allowed_run_tw(ctx))
1648 check_cq = READ_ONCE(ctx->check_cq);
1649 if (unlikely(check_cq)) {
1650 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1651 __io_cqring_overflow_flush(ctx);
1653 * Similarly do not spin if we have not informed the user of any
1656 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1660 * Don't enter poll loop if we already have events pending.
1661 * If we do, we can potentially be spinning for commands that
1662 * already triggered a CQE (eg in error).
1664 if (io_cqring_events(ctx))
1671 * If a submit got punted to a workqueue, we can have the
1672 * application entering polling for a command before it gets
1673 * issued. That app will hold the uring_lock for the duration
1674 * of the poll right here, so we need to take a breather every
1675 * now and then to ensure that the issue has a chance to add
1676 * the poll to the issued list. Otherwise we can spin here
1677 * forever, while the workqueue is stuck trying to acquire the
1680 if (wq_list_empty(&ctx->iopoll_list) ||
1681 io_task_work_pending(ctx)) {
1682 u32 tail = ctx->cached_cq_tail;
1684 (void) io_run_local_work_locked(ctx, min);
1686 if (task_work_pending(current) ||
1687 wq_list_empty(&ctx->iopoll_list)) {
1688 mutex_unlock(&ctx->uring_lock);
1690 mutex_lock(&ctx->uring_lock);
1692 /* some requests don't go through iopoll_list */
1693 if (tail != ctx->cached_cq_tail ||
1694 wq_list_empty(&ctx->iopoll_list))
1697 ret = io_do_iopoll(ctx, !min);
1698 if (unlikely(ret < 0))
1701 if (task_sigpending(current))
1707 } while (nr_events < min);
1712 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts)
1715 io_req_complete_defer(req);
1717 io_req_complete_post(req, IO_URING_F_UNLOCKED);
1721 * After the iocb has been issued, it's safe to be found on the poll list.
1722 * Adding the kiocb to the list AFTER submission ensures that we don't
1723 * find it from a io_do_iopoll() thread before the issuer is done
1724 * accessing the kiocb cookie.
1726 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1728 struct io_ring_ctx *ctx = req->ctx;
1729 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1731 /* workqueue context doesn't hold uring_lock, grab it now */
1732 if (unlikely(needs_lock))
1733 mutex_lock(&ctx->uring_lock);
1736 * Track whether we have multiple files in our lists. This will impact
1737 * how we do polling eventually, not spinning if we're on potentially
1738 * different devices.
1740 if (wq_list_empty(&ctx->iopoll_list)) {
1741 ctx->poll_multi_queue = false;
1742 } else if (!ctx->poll_multi_queue) {
1743 struct io_kiocb *list_req;
1745 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1747 if (list_req->file != req->file)
1748 ctx->poll_multi_queue = true;
1752 * For fast devices, IO may have already completed. If it has, add
1753 * it to the front so we find it first.
1755 if (READ_ONCE(req->iopoll_completed))
1756 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1758 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1760 if (unlikely(needs_lock)) {
1762 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1763 * in sq thread task context or in io worker task context. If
1764 * current task context is sq thread, we don't need to check
1765 * whether should wake up sq thread.
1767 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1768 wq_has_sleeper(&ctx->sq_data->wait))
1769 wake_up(&ctx->sq_data->wait);
1771 mutex_unlock(&ctx->uring_lock);
1775 io_req_flags_t io_file_get_flags(struct file *file)
1777 io_req_flags_t res = 0;
1779 if (S_ISREG(file_inode(file)->i_mode))
1781 if ((file->f_flags & O_NONBLOCK) || (file->f_mode & FMODE_NOWAIT))
1782 res |= REQ_F_SUPPORT_NOWAIT;
1786 bool io_alloc_async_data(struct io_kiocb *req)
1788 WARN_ON_ONCE(!io_cold_defs[req->opcode].async_size);
1789 req->async_data = kmalloc(io_cold_defs[req->opcode].async_size, GFP_KERNEL);
1790 if (req->async_data) {
1791 req->flags |= REQ_F_ASYNC_DATA;
1797 int io_req_prep_async(struct io_kiocb *req)
1799 const struct io_cold_def *cdef = &io_cold_defs[req->opcode];
1800 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1802 /* assign early for deferred execution for non-fixed file */
1803 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE) && !req->file)
1804 req->file = io_file_get_normal(req, req->cqe.fd);
1805 if (!cdef->prep_async)
1807 if (WARN_ON_ONCE(req_has_async_data(req)))
1809 if (!def->manual_alloc) {
1810 if (io_alloc_async_data(req))
1813 return cdef->prep_async(req);
1816 static u32 io_get_sequence(struct io_kiocb *req)
1818 u32 seq = req->ctx->cached_sq_head;
1819 struct io_kiocb *cur;
1821 /* need original cached_sq_head, but it was increased for each req */
1822 io_for_each_link(cur, req)
1827 static __cold void io_drain_req(struct io_kiocb *req)
1828 __must_hold(&ctx->uring_lock)
1830 struct io_ring_ctx *ctx = req->ctx;
1831 struct io_defer_entry *de;
1833 u32 seq = io_get_sequence(req);
1835 /* Still need defer if there is pending req in defer list. */
1836 spin_lock(&ctx->completion_lock);
1837 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1838 spin_unlock(&ctx->completion_lock);
1840 ctx->drain_active = false;
1841 io_req_task_queue(req);
1844 spin_unlock(&ctx->completion_lock);
1846 io_prep_async_link(req);
1847 de = kmalloc(sizeof(*de), GFP_KERNEL);
1850 io_req_defer_failed(req, ret);
1854 spin_lock(&ctx->completion_lock);
1855 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1856 spin_unlock(&ctx->completion_lock);
1861 trace_io_uring_defer(req);
1864 list_add_tail(&de->list, &ctx->defer_list);
1865 spin_unlock(&ctx->completion_lock);
1868 static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def,
1869 unsigned int issue_flags)
1871 if (req->file || !def->needs_file)
1874 if (req->flags & REQ_F_FIXED_FILE)
1875 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1877 req->file = io_file_get_normal(req, req->cqe.fd);
1882 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1884 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1885 const struct cred *creds = NULL;
1888 if (unlikely(!io_assign_file(req, def, issue_flags)))
1891 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1892 creds = override_creds(req->creds);
1894 if (!def->audit_skip)
1895 audit_uring_entry(req->opcode);
1897 ret = def->issue(req, issue_flags);
1899 if (!def->audit_skip)
1900 audit_uring_exit(!ret, ret);
1903 revert_creds(creds);
1905 if (ret == IOU_OK) {
1906 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1907 io_req_complete_defer(req);
1909 io_req_complete_post(req, issue_flags);
1914 if (ret == IOU_ISSUE_SKIP_COMPLETE) {
1916 io_arm_ltimeout(req);
1918 /* If the op doesn't have a file, we're not polling for it */
1919 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
1920 io_iopoll_req_issued(req, issue_flags);
1925 int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts)
1927 io_tw_lock(req->ctx, ts);
1928 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT|
1929 IO_URING_F_COMPLETE_DEFER);
1932 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1934 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1935 struct io_kiocb *nxt = NULL;
1937 if (req_ref_put_and_test(req)) {
1938 if (req->flags & IO_REQ_LINK_FLAGS)
1939 nxt = io_req_find_next(req);
1942 return nxt ? &nxt->work : NULL;
1945 void io_wq_submit_work(struct io_wq_work *work)
1947 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1948 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1949 unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ;
1950 bool needs_poll = false;
1951 int ret = 0, err = -ECANCELED;
1953 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1954 if (!(req->flags & REQ_F_REFCOUNT))
1955 __io_req_set_refcount(req, 2);
1959 io_arm_ltimeout(req);
1961 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1962 if (work->flags & IO_WQ_WORK_CANCEL) {
1964 io_req_task_queue_fail(req, err);
1967 if (!io_assign_file(req, def, issue_flags)) {
1969 work->flags |= IO_WQ_WORK_CANCEL;
1974 * If DEFER_TASKRUN is set, it's only allowed to post CQEs from the
1975 * submitter task context. Final request completions are handed to the
1976 * right context, however this is not the case of auxiliary CQEs,
1977 * which is the main mean of operation for multishot requests.
1978 * Don't allow any multishot execution from io-wq. It's more restrictive
1979 * than necessary and also cleaner.
1981 if (req->flags & REQ_F_APOLL_MULTISHOT) {
1983 if (!io_file_can_poll(req))
1985 if (req->file->f_flags & O_NONBLOCK ||
1986 req->file->f_mode & FMODE_NOWAIT) {
1988 if (io_arm_poll_handler(req, issue_flags) != IO_APOLL_OK)
1992 req->flags &= ~REQ_F_APOLL_MULTISHOT;
1996 if (req->flags & REQ_F_FORCE_ASYNC) {
1997 bool opcode_poll = def->pollin || def->pollout;
1999 if (opcode_poll && io_file_can_poll(req)) {
2001 issue_flags |= IO_URING_F_NONBLOCK;
2006 ret = io_issue_sqe(req, issue_flags);
2011 * If REQ_F_NOWAIT is set, then don't wait or retry with
2012 * poll. -EAGAIN is final for that case.
2014 if (req->flags & REQ_F_NOWAIT)
2018 * We can get EAGAIN for iopolled IO even though we're
2019 * forcing a sync submission from here, since we can't
2020 * wait for request slots on the block side.
2023 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
2025 if (io_wq_worker_stopped())
2031 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
2033 /* aborted or ready, in either case retry blocking */
2035 issue_flags &= ~IO_URING_F_NONBLOCK;
2038 /* avoid locking problems by failing it from a clean context */
2040 io_req_task_queue_fail(req, ret);
2043 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
2044 unsigned int issue_flags)
2046 struct io_ring_ctx *ctx = req->ctx;
2047 struct io_fixed_file *slot;
2048 struct file *file = NULL;
2050 io_ring_submit_lock(ctx, issue_flags);
2052 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
2054 fd = array_index_nospec(fd, ctx->nr_user_files);
2055 slot = io_fixed_file_slot(&ctx->file_table, fd);
2056 if (!req->rsrc_node)
2057 __io_req_set_rsrc_node(req, ctx);
2058 req->flags |= io_slot_flags(slot);
2059 file = io_slot_file(slot);
2061 io_ring_submit_unlock(ctx, issue_flags);
2065 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
2067 struct file *file = fget(fd);
2069 trace_io_uring_file_get(req, fd);
2071 /* we don't allow fixed io_uring files */
2072 if (file && io_is_uring_fops(file))
2073 io_req_track_inflight(req);
2077 static void io_queue_async(struct io_kiocb *req, int ret)
2078 __must_hold(&req->ctx->uring_lock)
2080 struct io_kiocb *linked_timeout;
2082 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
2083 io_req_defer_failed(req, ret);
2087 linked_timeout = io_prep_linked_timeout(req);
2089 switch (io_arm_poll_handler(req, 0)) {
2090 case IO_APOLL_READY:
2091 io_kbuf_recycle(req, 0);
2092 io_req_task_queue(req);
2094 case IO_APOLL_ABORTED:
2095 io_kbuf_recycle(req, 0);
2096 io_queue_iowq(req, NULL);
2103 io_queue_linked_timeout(linked_timeout);
2106 static inline void io_queue_sqe(struct io_kiocb *req)
2107 __must_hold(&req->ctx->uring_lock)
2111 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
2114 * We async punt it if the file wasn't marked NOWAIT, or if the file
2115 * doesn't support non-blocking read/write attempts
2118 io_queue_async(req, ret);
2121 static void io_queue_sqe_fallback(struct io_kiocb *req)
2122 __must_hold(&req->ctx->uring_lock)
2124 if (unlikely(req->flags & REQ_F_FAIL)) {
2126 * We don't submit, fail them all, for that replace hardlinks
2127 * with normal links. Extra REQ_F_LINK is tolerated.
2129 req->flags &= ~REQ_F_HARDLINK;
2130 req->flags |= REQ_F_LINK;
2131 io_req_defer_failed(req, req->cqe.res);
2133 int ret = io_req_prep_async(req);
2135 if (unlikely(ret)) {
2136 io_req_defer_failed(req, ret);
2140 if (unlikely(req->ctx->drain_active))
2143 io_queue_iowq(req, NULL);
2148 * Check SQE restrictions (opcode and flags).
2150 * Returns 'true' if SQE is allowed, 'false' otherwise.
2152 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
2153 struct io_kiocb *req,
2154 unsigned int sqe_flags)
2156 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
2159 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
2160 ctx->restrictions.sqe_flags_required)
2163 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
2164 ctx->restrictions.sqe_flags_required))
2170 static void io_init_req_drain(struct io_kiocb *req)
2172 struct io_ring_ctx *ctx = req->ctx;
2173 struct io_kiocb *head = ctx->submit_state.link.head;
2175 ctx->drain_active = true;
2178 * If we need to drain a request in the middle of a link, drain
2179 * the head request and the next request/link after the current
2180 * link. Considering sequential execution of links,
2181 * REQ_F_IO_DRAIN will be maintained for every request of our
2184 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2185 ctx->drain_next = true;
2189 static __cold int io_init_fail_req(struct io_kiocb *req, int err)
2191 /* ensure per-opcode data is cleared if we fail before prep */
2192 memset(&req->cmd.data, 0, sizeof(req->cmd.data));
2196 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
2197 const struct io_uring_sqe *sqe)
2198 __must_hold(&ctx->uring_lock)
2200 const struct io_issue_def *def;
2201 unsigned int sqe_flags;
2205 /* req is partially pre-initialised, see io_preinit_req() */
2206 req->opcode = opcode = READ_ONCE(sqe->opcode);
2207 /* same numerical values with corresponding REQ_F_*, safe to copy */
2208 sqe_flags = READ_ONCE(sqe->flags);
2209 req->flags = (io_req_flags_t) sqe_flags;
2210 req->cqe.user_data = READ_ONCE(sqe->user_data);
2212 req->rsrc_node = NULL;
2213 req->task = current;
2215 if (unlikely(opcode >= IORING_OP_LAST)) {
2217 return io_init_fail_req(req, -EINVAL);
2219 def = &io_issue_defs[opcode];
2220 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2221 /* enforce forwards compatibility on users */
2222 if (sqe_flags & ~SQE_VALID_FLAGS)
2223 return io_init_fail_req(req, -EINVAL);
2224 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2225 if (!def->buffer_select)
2226 return io_init_fail_req(req, -EOPNOTSUPP);
2227 req->buf_index = READ_ONCE(sqe->buf_group);
2229 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2230 ctx->drain_disabled = true;
2231 if (sqe_flags & IOSQE_IO_DRAIN) {
2232 if (ctx->drain_disabled)
2233 return io_init_fail_req(req, -EOPNOTSUPP);
2234 io_init_req_drain(req);
2237 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2238 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2239 return io_init_fail_req(req, -EACCES);
2240 /* knock it to the slow queue path, will be drained there */
2241 if (ctx->drain_active)
2242 req->flags |= REQ_F_FORCE_ASYNC;
2243 /* if there is no link, we're at "next" request and need to drain */
2244 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2245 ctx->drain_next = false;
2246 ctx->drain_active = true;
2247 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2251 if (!def->ioprio && sqe->ioprio)
2252 return io_init_fail_req(req, -EINVAL);
2253 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2254 return io_init_fail_req(req, -EINVAL);
2256 if (def->needs_file) {
2257 struct io_submit_state *state = &ctx->submit_state;
2259 req->cqe.fd = READ_ONCE(sqe->fd);
2262 * Plug now if we have more than 2 IO left after this, and the
2263 * target is potentially a read/write to block based storage.
2265 if (state->need_plug && def->plug) {
2266 state->plug_started = true;
2267 state->need_plug = false;
2268 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2272 personality = READ_ONCE(sqe->personality);
2276 req->creds = xa_load(&ctx->personalities, personality);
2278 return io_init_fail_req(req, -EINVAL);
2279 get_cred(req->creds);
2280 ret = security_uring_override_creds(req->creds);
2282 put_cred(req->creds);
2283 return io_init_fail_req(req, ret);
2285 req->flags |= REQ_F_CREDS;
2288 return def->prep(req, sqe);
2291 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2292 struct io_kiocb *req, int ret)
2294 struct io_ring_ctx *ctx = req->ctx;
2295 struct io_submit_link *link = &ctx->submit_state.link;
2296 struct io_kiocb *head = link->head;
2298 trace_io_uring_req_failed(sqe, req, ret);
2301 * Avoid breaking links in the middle as it renders links with SQPOLL
2302 * unusable. Instead of failing eagerly, continue assembling the link if
2303 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2304 * should find the flag and handle the rest.
2306 req_fail_link_node(req, ret);
2307 if (head && !(head->flags & REQ_F_FAIL))
2308 req_fail_link_node(head, -ECANCELED);
2310 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2312 link->last->link = req;
2316 io_queue_sqe_fallback(req);
2321 link->last->link = req;
2328 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2329 const struct io_uring_sqe *sqe)
2330 __must_hold(&ctx->uring_lock)
2332 struct io_submit_link *link = &ctx->submit_state.link;
2335 ret = io_init_req(ctx, req, sqe);
2337 return io_submit_fail_init(sqe, req, ret);
2339 trace_io_uring_submit_req(req);
2342 * If we already have a head request, queue this one for async
2343 * submittal once the head completes. If we don't have a head but
2344 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2345 * submitted sync once the chain is complete. If none of those
2346 * conditions are true (normal request), then just queue it.
2348 if (unlikely(link->head)) {
2349 ret = io_req_prep_async(req);
2351 return io_submit_fail_init(sqe, req, ret);
2353 trace_io_uring_link(req, link->head);
2354 link->last->link = req;
2357 if (req->flags & IO_REQ_LINK_FLAGS)
2359 /* last request of the link, flush it */
2362 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2365 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2366 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2367 if (req->flags & IO_REQ_LINK_FLAGS) {
2372 io_queue_sqe_fallback(req);
2382 * Batched submission is done, ensure local IO is flushed out.
2384 static void io_submit_state_end(struct io_ring_ctx *ctx)
2386 struct io_submit_state *state = &ctx->submit_state;
2388 if (unlikely(state->link.head))
2389 io_queue_sqe_fallback(state->link.head);
2390 /* flush only after queuing links as they can generate completions */
2391 io_submit_flush_completions(ctx);
2392 if (state->plug_started)
2393 blk_finish_plug(&state->plug);
2397 * Start submission side cache.
2399 static void io_submit_state_start(struct io_submit_state *state,
2400 unsigned int max_ios)
2402 state->plug_started = false;
2403 state->need_plug = max_ios > 2;
2404 state->submit_nr = max_ios;
2405 /* set only head, no need to init link_last in advance */
2406 state->link.head = NULL;
2409 static void io_commit_sqring(struct io_ring_ctx *ctx)
2411 struct io_rings *rings = ctx->rings;
2414 * Ensure any loads from the SQEs are done at this point,
2415 * since once we write the new head, the application could
2416 * write new data to them.
2418 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2422 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2423 * that is mapped by userspace. This means that care needs to be taken to
2424 * ensure that reads are stable, as we cannot rely on userspace always
2425 * being a good citizen. If members of the sqe are validated and then later
2426 * used, it's important that those reads are done through READ_ONCE() to
2427 * prevent a re-load down the line.
2429 static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe)
2431 unsigned mask = ctx->sq_entries - 1;
2432 unsigned head = ctx->cached_sq_head++ & mask;
2434 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY)) {
2435 head = READ_ONCE(ctx->sq_array[head]);
2436 if (unlikely(head >= ctx->sq_entries)) {
2437 /* drop invalid entries */
2438 spin_lock(&ctx->completion_lock);
2440 spin_unlock(&ctx->completion_lock);
2441 WRITE_ONCE(ctx->rings->sq_dropped,
2442 READ_ONCE(ctx->rings->sq_dropped) + 1);
2448 * The cached sq head (or cq tail) serves two purposes:
2450 * 1) allows us to batch the cost of updating the user visible
2452 * 2) allows the kernel side to track the head on its own, even
2453 * though the application is the one updating it.
2456 /* double index for 128-byte SQEs, twice as long */
2457 if (ctx->flags & IORING_SETUP_SQE128)
2459 *sqe = &ctx->sq_sqes[head];
2463 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2464 __must_hold(&ctx->uring_lock)
2466 unsigned int entries = io_sqring_entries(ctx);
2470 if (unlikely(!entries))
2472 /* make sure SQ entry isn't read before tail */
2473 ret = left = min(nr, entries);
2474 io_get_task_refs(left);
2475 io_submit_state_start(&ctx->submit_state, left);
2478 const struct io_uring_sqe *sqe;
2479 struct io_kiocb *req;
2481 if (unlikely(!io_alloc_req(ctx, &req)))
2483 if (unlikely(!io_get_sqe(ctx, &sqe))) {
2484 io_req_add_to_cache(req, ctx);
2489 * Continue submitting even for sqe failure if the
2490 * ring was setup with IORING_SETUP_SUBMIT_ALL
2492 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2493 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2499 if (unlikely(left)) {
2501 /* try again if it submitted nothing and can't allocate a req */
2502 if (!ret && io_req_cache_empty(ctx))
2504 current->io_uring->cached_refs += left;
2507 io_submit_state_end(ctx);
2508 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2509 io_commit_sqring(ctx);
2513 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2514 int wake_flags, void *key)
2516 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq);
2519 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2520 * the task, and the next invocation will do it.
2522 if (io_should_wake(iowq) || io_has_work(iowq->ctx))
2523 return autoremove_wake_function(curr, mode, wake_flags, key);
2527 int io_run_task_work_sig(struct io_ring_ctx *ctx)
2529 if (!llist_empty(&ctx->work_llist)) {
2530 __set_current_state(TASK_RUNNING);
2531 if (io_run_local_work(ctx, INT_MAX) > 0)
2534 if (io_run_task_work() > 0)
2536 if (task_sigpending(current))
2541 static bool current_pending_io(void)
2543 struct io_uring_task *tctx = current->io_uring;
2547 return percpu_counter_read_positive(&tctx->inflight);
2550 /* when returns >0, the caller should retry */
2551 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2552 struct io_wait_queue *iowq)
2556 if (unlikely(READ_ONCE(ctx->check_cq)))
2558 if (unlikely(!llist_empty(&ctx->work_llist)))
2560 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL)))
2562 if (unlikely(task_sigpending(current)))
2564 if (unlikely(io_should_wake(iowq)))
2568 * Mark us as being in io_wait if we have pending requests, so cpufreq
2569 * can take into account that the task is waiting for IO - turns out
2570 * to be important for low QD IO.
2572 if (current_pending_io())
2573 current->in_iowait = 1;
2575 if (iowq->timeout == KTIME_MAX)
2577 else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS))
2579 current->in_iowait = 0;
2584 * Wait until events become available, if we don't already have some. The
2585 * application must reap them itself, as they reside on the shared cq ring.
2587 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2588 const sigset_t __user *sig, size_t sigsz,
2589 struct __kernel_timespec __user *uts)
2591 struct io_wait_queue iowq;
2592 struct io_rings *rings = ctx->rings;
2595 if (!io_allowed_run_tw(ctx))
2597 if (!llist_empty(&ctx->work_llist))
2598 io_run_local_work(ctx, min_events);
2600 io_cqring_overflow_flush(ctx);
2601 /* if user messes with these they will just get an early return */
2602 if (__io_cqring_events_user(ctx) >= min_events)
2605 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2606 iowq.wq.private = current;
2607 INIT_LIST_HEAD(&iowq.wq.entry);
2609 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2610 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2611 iowq.timeout = KTIME_MAX;
2614 struct timespec64 ts;
2616 if (get_timespec64(&ts, uts))
2619 iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2620 io_napi_adjust_timeout(ctx, &iowq, &ts);
2624 #ifdef CONFIG_COMPAT
2625 if (in_compat_syscall())
2626 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2630 ret = set_user_sigmask(sig, sigsz);
2636 io_napi_busy_loop(ctx, &iowq);
2638 trace_io_uring_cqring_wait(ctx, min_events);
2640 int nr_wait = (int) iowq.cq_tail - READ_ONCE(ctx->rings->cq.tail);
2641 unsigned long check_cq;
2643 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
2644 atomic_set(&ctx->cq_wait_nr, nr_wait);
2645 set_current_state(TASK_INTERRUPTIBLE);
2647 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2648 TASK_INTERRUPTIBLE);
2651 ret = io_cqring_wait_schedule(ctx, &iowq);
2652 __set_current_state(TASK_RUNNING);
2653 atomic_set(&ctx->cq_wait_nr, IO_CQ_WAKE_INIT);
2656 * Run task_work after scheduling and before io_should_wake().
2657 * If we got woken because of task_work being processed, run it
2658 * now rather than let the caller do another wait loop.
2661 if (!llist_empty(&ctx->work_llist))
2662 io_run_local_work(ctx, nr_wait);
2665 * Non-local task_work will be run on exit to userspace, but
2666 * if we're using DEFER_TASKRUN, then we could have waited
2667 * with a timeout for a number of requests. If the timeout
2668 * hits, we could have some requests ready to process. Ensure
2669 * this break is _after_ we have run task_work, to avoid
2670 * deferring running potentially pending requests until the
2671 * next time we wait for events.
2676 check_cq = READ_ONCE(ctx->check_cq);
2677 if (unlikely(check_cq)) {
2678 /* let the caller flush overflows, retry */
2679 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2680 io_cqring_do_overflow_flush(ctx);
2681 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) {
2687 if (io_should_wake(&iowq)) {
2694 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
2695 finish_wait(&ctx->cq_wait, &iowq.wq);
2696 restore_saved_sigmask_unless(ret == -EINTR);
2698 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2701 void io_mem_free(void *ptr)
2706 folio_put(virt_to_folio(ptr));
2709 static void io_pages_free(struct page ***pages, int npages)
2711 struct page **page_array = *pages;
2717 for (i = 0; i < npages; i++)
2718 unpin_user_page(page_array[i]);
2723 static void *__io_uaddr_map(struct page ***pages, unsigned short *npages,
2724 unsigned long uaddr, size_t size)
2726 struct page **page_array;
2727 unsigned int nr_pages;
2733 if (uaddr & (PAGE_SIZE - 1) || !size)
2734 return ERR_PTR(-EINVAL);
2736 nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
2737 if (nr_pages > USHRT_MAX)
2738 return ERR_PTR(-EINVAL);
2739 page_array = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
2741 return ERR_PTR(-ENOMEM);
2744 pinned = pin_user_pages_fast(uaddr, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
2746 if (pinned != nr_pages) {
2747 ret = (pinned < 0) ? pinned : -EFAULT;
2751 page_addr = page_address(page_array[0]);
2752 for (i = 0; i < nr_pages; i++) {
2756 * Can't support mapping user allocated ring memory on 32-bit
2757 * archs where it could potentially reside in highmem. Just
2758 * fail those with -EINVAL, just like we did on kernels that
2759 * didn't support this feature.
2761 if (PageHighMem(page_array[i]))
2765 * No support for discontig pages for now, should either be a
2766 * single normal page, or a huge page. Later on we can add
2767 * support for remapping discontig pages, for now we will
2768 * just fail them with EINVAL.
2770 if (page_address(page_array[i]) != page_addr)
2772 page_addr += PAGE_SIZE;
2775 *pages = page_array;
2777 return page_to_virt(page_array[0]);
2780 io_pages_free(&page_array, pinned > 0 ? pinned : 0);
2781 return ERR_PTR(ret);
2784 static void *io_rings_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2787 return __io_uaddr_map(&ctx->ring_pages, &ctx->n_ring_pages, uaddr,
2791 static void *io_sqes_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2794 return __io_uaddr_map(&ctx->sqe_pages, &ctx->n_sqe_pages, uaddr,
2798 static void io_rings_free(struct io_ring_ctx *ctx)
2800 if (!(ctx->flags & IORING_SETUP_NO_MMAP)) {
2801 io_mem_free(ctx->rings);
2802 io_mem_free(ctx->sq_sqes);
2804 io_pages_free(&ctx->ring_pages, ctx->n_ring_pages);
2805 ctx->n_ring_pages = 0;
2806 io_pages_free(&ctx->sqe_pages, ctx->n_sqe_pages);
2807 ctx->n_sqe_pages = 0;
2811 ctx->sq_sqes = NULL;
2814 void *io_mem_alloc(size_t size)
2816 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2819 ret = (void *) __get_free_pages(gfp, get_order(size));
2822 return ERR_PTR(-ENOMEM);
2825 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2826 unsigned int cq_entries, size_t *sq_offset)
2828 struct io_rings *rings;
2829 size_t off, sq_array_size;
2831 off = struct_size(rings, cqes, cq_entries);
2832 if (off == SIZE_MAX)
2834 if (ctx->flags & IORING_SETUP_CQE32) {
2835 if (check_shl_overflow(off, 1, &off))
2840 off = ALIGN(off, SMP_CACHE_BYTES);
2845 if (ctx->flags & IORING_SETUP_NO_SQARRAY) {
2847 *sq_offset = SIZE_MAX;
2854 sq_array_size = array_size(sizeof(u32), sq_entries);
2855 if (sq_array_size == SIZE_MAX)
2858 if (check_add_overflow(off, sq_array_size, &off))
2864 static void io_req_caches_free(struct io_ring_ctx *ctx)
2866 struct io_kiocb *req;
2869 mutex_lock(&ctx->uring_lock);
2870 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
2872 while (!io_req_cache_empty(ctx)) {
2873 req = io_extract_req(ctx);
2874 kmem_cache_free(req_cachep, req);
2878 percpu_ref_put_many(&ctx->refs, nr);
2879 mutex_unlock(&ctx->uring_lock);
2882 static void io_rsrc_node_cache_free(struct io_cache_entry *entry)
2884 kfree(container_of(entry, struct io_rsrc_node, cache));
2887 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2889 io_sq_thread_finish(ctx);
2890 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2891 if (WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)))
2894 mutex_lock(&ctx->uring_lock);
2896 __io_sqe_buffers_unregister(ctx);
2898 __io_sqe_files_unregister(ctx);
2899 io_cqring_overflow_kill(ctx);
2900 io_eventfd_unregister(ctx);
2901 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2902 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2903 io_futex_cache_free(ctx);
2904 io_destroy_buffers(ctx);
2905 mutex_unlock(&ctx->uring_lock);
2907 put_cred(ctx->sq_creds);
2908 if (ctx->submitter_task)
2909 put_task_struct(ctx->submitter_task);
2911 /* there are no registered resources left, nobody uses it */
2913 io_rsrc_node_destroy(ctx, ctx->rsrc_node);
2915 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2916 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2918 io_alloc_cache_free(&ctx->rsrc_node_cache, io_rsrc_node_cache_free);
2919 if (ctx->mm_account) {
2920 mmdrop(ctx->mm_account);
2921 ctx->mm_account = NULL;
2924 io_kbuf_mmap_list_free(ctx);
2926 percpu_ref_exit(&ctx->refs);
2927 free_uid(ctx->user);
2928 io_req_caches_free(ctx);
2930 io_wq_put_hash(ctx->hash_map);
2932 kfree(ctx->cancel_table.hbs);
2933 kfree(ctx->cancel_table_locked.hbs);
2934 xa_destroy(&ctx->io_bl_xa);
2938 static __cold void io_activate_pollwq_cb(struct callback_head *cb)
2940 struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx,
2943 mutex_lock(&ctx->uring_lock);
2944 ctx->poll_activated = true;
2945 mutex_unlock(&ctx->uring_lock);
2948 * Wake ups for some events between start of polling and activation
2949 * might've been lost due to loose synchronisation.
2951 wake_up_all(&ctx->poll_wq);
2952 percpu_ref_put(&ctx->refs);
2955 __cold void io_activate_pollwq(struct io_ring_ctx *ctx)
2957 spin_lock(&ctx->completion_lock);
2958 /* already activated or in progress */
2959 if (ctx->poll_activated || ctx->poll_wq_task_work.func)
2961 if (WARN_ON_ONCE(!ctx->task_complete))
2963 if (!ctx->submitter_task)
2966 * with ->submitter_task only the submitter task completes requests, we
2967 * only need to sync with it, which is done by injecting a tw
2969 init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb);
2970 percpu_ref_get(&ctx->refs);
2971 if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL))
2972 percpu_ref_put(&ctx->refs);
2974 spin_unlock(&ctx->completion_lock);
2977 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2979 struct io_ring_ctx *ctx = file->private_data;
2982 if (unlikely(!ctx->poll_activated))
2983 io_activate_pollwq(ctx);
2985 poll_wait(file, &ctx->poll_wq, wait);
2987 * synchronizes with barrier from wq_has_sleeper call in
2991 if (!io_sqring_full(ctx))
2992 mask |= EPOLLOUT | EPOLLWRNORM;
2995 * Don't flush cqring overflow list here, just do a simple check.
2996 * Otherwise there could possible be ABBA deadlock:
2999 * lock(&ctx->uring_lock);
3001 * lock(&ctx->uring_lock);
3004 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
3005 * pushes them to do the flush.
3008 if (__io_cqring_events_user(ctx) || io_has_work(ctx))
3009 mask |= EPOLLIN | EPOLLRDNORM;
3014 struct io_tctx_exit {
3015 struct callback_head task_work;
3016 struct completion completion;
3017 struct io_ring_ctx *ctx;
3020 static __cold void io_tctx_exit_cb(struct callback_head *cb)
3022 struct io_uring_task *tctx = current->io_uring;
3023 struct io_tctx_exit *work;
3025 work = container_of(cb, struct io_tctx_exit, task_work);
3027 * When @in_cancel, we're in cancellation and it's racy to remove the
3028 * node. It'll be removed by the end of cancellation, just ignore it.
3029 * tctx can be NULL if the queueing of this task_work raced with
3030 * work cancelation off the exec path.
3032 if (tctx && !atomic_read(&tctx->in_cancel))
3033 io_uring_del_tctx_node((unsigned long)work->ctx);
3034 complete(&work->completion);
3037 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
3039 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3041 return req->ctx == data;
3044 static __cold void io_ring_exit_work(struct work_struct *work)
3046 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
3047 unsigned long timeout = jiffies + HZ * 60 * 5;
3048 unsigned long interval = HZ / 20;
3049 struct io_tctx_exit exit;
3050 struct io_tctx_node *node;
3054 * If we're doing polled IO and end up having requests being
3055 * submitted async (out-of-line), then completions can come in while
3056 * we're waiting for refs to drop. We need to reap these manually,
3057 * as nobody else will be looking for them.
3060 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
3061 mutex_lock(&ctx->uring_lock);
3062 io_cqring_overflow_kill(ctx);
3063 mutex_unlock(&ctx->uring_lock);
3066 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3067 io_move_task_work_from_local(ctx);
3069 while (io_uring_try_cancel_requests(ctx, NULL, true))
3073 struct io_sq_data *sqd = ctx->sq_data;
3074 struct task_struct *tsk;
3076 io_sq_thread_park(sqd);
3078 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
3079 io_wq_cancel_cb(tsk->io_uring->io_wq,
3080 io_cancel_ctx_cb, ctx, true);
3081 io_sq_thread_unpark(sqd);
3084 io_req_caches_free(ctx);
3086 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
3087 /* there is little hope left, don't run it too often */
3091 * This is really an uninterruptible wait, as it has to be
3092 * complete. But it's also run from a kworker, which doesn't
3093 * take signals, so it's fine to make it interruptible. This
3094 * avoids scenarios where we knowingly can wait much longer
3095 * on completions, for example if someone does a SIGSTOP on
3096 * a task that needs to finish task_work to make this loop
3097 * complete. That's a synthetic situation that should not
3098 * cause a stuck task backtrace, and hence a potential panic
3099 * on stuck tasks if that is enabled.
3101 } while (!wait_for_completion_interruptible_timeout(&ctx->ref_comp, interval));
3103 init_completion(&exit.completion);
3104 init_task_work(&exit.task_work, io_tctx_exit_cb);
3107 mutex_lock(&ctx->uring_lock);
3108 while (!list_empty(&ctx->tctx_list)) {
3109 WARN_ON_ONCE(time_after(jiffies, timeout));
3111 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
3113 /* don't spin on a single task if cancellation failed */
3114 list_rotate_left(&ctx->tctx_list);
3115 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
3116 if (WARN_ON_ONCE(ret))
3119 mutex_unlock(&ctx->uring_lock);
3121 * See comment above for
3122 * wait_for_completion_interruptible_timeout() on why this
3123 * wait is marked as interruptible.
3125 wait_for_completion_interruptible(&exit.completion);
3126 mutex_lock(&ctx->uring_lock);
3128 mutex_unlock(&ctx->uring_lock);
3129 spin_lock(&ctx->completion_lock);
3130 spin_unlock(&ctx->completion_lock);
3132 /* pairs with RCU read section in io_req_local_work_add() */
3133 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3136 io_ring_ctx_free(ctx);
3139 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
3141 unsigned long index;
3142 struct creds *creds;
3144 mutex_lock(&ctx->uring_lock);
3145 percpu_ref_kill(&ctx->refs);
3146 xa_for_each(&ctx->personalities, index, creds)
3147 io_unregister_personality(ctx, index);
3149 io_poll_remove_all(ctx, NULL, true);
3150 mutex_unlock(&ctx->uring_lock);
3153 * If we failed setting up the ctx, we might not have any rings
3154 * and therefore did not submit any requests
3157 io_kill_timeouts(ctx, NULL, true);
3159 flush_delayed_work(&ctx->fallback_work);
3161 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
3163 * Use system_unbound_wq to avoid spawning tons of event kworkers
3164 * if we're exiting a ton of rings at the same time. It just adds
3165 * noise and overhead, there's no discernable change in runtime
3166 * over using system_wq.
3168 queue_work(iou_wq, &ctx->exit_work);
3171 static int io_uring_release(struct inode *inode, struct file *file)
3173 struct io_ring_ctx *ctx = file->private_data;
3175 file->private_data = NULL;
3176 io_ring_ctx_wait_and_kill(ctx);
3180 struct io_task_cancel {
3181 struct task_struct *task;
3185 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
3187 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3188 struct io_task_cancel *cancel = data;
3190 return io_match_task_safe(req, cancel->task, cancel->all);
3193 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
3194 struct task_struct *task,
3197 struct io_defer_entry *de;
3200 spin_lock(&ctx->completion_lock);
3201 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
3202 if (io_match_task_safe(de->req, task, cancel_all)) {
3203 list_cut_position(&list, &ctx->defer_list, &de->list);
3207 spin_unlock(&ctx->completion_lock);
3208 if (list_empty(&list))
3211 while (!list_empty(&list)) {
3212 de = list_first_entry(&list, struct io_defer_entry, list);
3213 list_del_init(&de->list);
3214 io_req_task_queue_fail(de->req, -ECANCELED);
3220 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
3222 struct io_tctx_node *node;
3223 enum io_wq_cancel cret;
3226 mutex_lock(&ctx->uring_lock);
3227 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3228 struct io_uring_task *tctx = node->task->io_uring;
3231 * io_wq will stay alive while we hold uring_lock, because it's
3232 * killed after ctx nodes, which requires to take the lock.
3234 if (!tctx || !tctx->io_wq)
3236 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
3237 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3239 mutex_unlock(&ctx->uring_lock);
3244 static bool io_uring_try_cancel_uring_cmd(struct io_ring_ctx *ctx,
3245 struct task_struct *task, bool cancel_all)
3247 struct hlist_node *tmp;
3248 struct io_kiocb *req;
3251 lockdep_assert_held(&ctx->uring_lock);
3253 hlist_for_each_entry_safe(req, tmp, &ctx->cancelable_uring_cmd,
3255 struct io_uring_cmd *cmd = io_kiocb_to_cmd(req,
3256 struct io_uring_cmd);
3257 struct file *file = req->file;
3259 if (!cancel_all && req->task != task)
3262 if (cmd->flags & IORING_URING_CMD_CANCELABLE) {
3263 /* ->sqe isn't available if no async data */
3264 if (!req_has_async_data(req))
3266 file->f_op->uring_cmd(cmd, IO_URING_F_CANCEL);
3270 io_submit_flush_completions(ctx);
3275 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
3276 struct task_struct *task,
3279 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
3280 struct io_uring_task *tctx = task ? task->io_uring : NULL;
3281 enum io_wq_cancel cret;
3284 /* set it so io_req_local_work_add() would wake us up */
3285 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
3286 atomic_set(&ctx->cq_wait_nr, 1);
3290 /* failed during ring init, it couldn't have issued any requests */
3295 ret |= io_uring_try_cancel_iowq(ctx);
3296 } else if (tctx && tctx->io_wq) {
3298 * Cancels requests of all rings, not only @ctx, but
3299 * it's fine as the task is in exit/exec.
3301 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
3303 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3306 /* SQPOLL thread does its own polling */
3307 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
3308 (ctx->sq_data && ctx->sq_data->thread == current)) {
3309 while (!wq_list_empty(&ctx->iopoll_list)) {
3310 io_iopoll_try_reap_events(ctx);
3316 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3317 io_allowed_defer_tw_run(ctx))
3318 ret |= io_run_local_work(ctx, INT_MAX) > 0;
3319 ret |= io_cancel_defer_files(ctx, task, cancel_all);
3320 mutex_lock(&ctx->uring_lock);
3321 ret |= io_poll_remove_all(ctx, task, cancel_all);
3322 ret |= io_waitid_remove_all(ctx, task, cancel_all);
3323 ret |= io_futex_remove_all(ctx, task, cancel_all);
3324 ret |= io_uring_try_cancel_uring_cmd(ctx, task, cancel_all);
3325 mutex_unlock(&ctx->uring_lock);
3326 ret |= io_kill_timeouts(ctx, task, cancel_all);
3328 ret |= io_run_task_work() > 0;
3332 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
3335 return atomic_read(&tctx->inflight_tracked);
3336 return percpu_counter_sum(&tctx->inflight);
3340 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3341 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3343 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
3345 struct io_uring_task *tctx = current->io_uring;
3346 struct io_ring_ctx *ctx;
3347 struct io_tctx_node *node;
3348 unsigned long index;
3352 WARN_ON_ONCE(sqd && sqd->thread != current);
3354 if (!current->io_uring)
3357 io_wq_exit_start(tctx->io_wq);
3359 atomic_inc(&tctx->in_cancel);
3363 io_uring_drop_tctx_refs(current);
3364 /* read completions before cancelations */
3365 inflight = tctx_inflight(tctx, !cancel_all);
3370 xa_for_each(&tctx->xa, index, node) {
3371 /* sqpoll task will cancel all its requests */
3372 if (node->ctx->sq_data)
3374 loop |= io_uring_try_cancel_requests(node->ctx,
3375 current, cancel_all);
3378 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3379 loop |= io_uring_try_cancel_requests(ctx,
3389 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3391 io_uring_drop_tctx_refs(current);
3392 xa_for_each(&tctx->xa, index, node) {
3393 if (!llist_empty(&node->ctx->work_llist)) {
3394 WARN_ON_ONCE(node->ctx->submitter_task &&
3395 node->ctx->submitter_task != current);
3400 * If we've seen completions, retry without waiting. This
3401 * avoids a race where a completion comes in before we did
3402 * prepare_to_wait().
3404 if (inflight == tctx_inflight(tctx, !cancel_all))
3407 finish_wait(&tctx->wait, &wait);
3410 io_uring_clean_tctx(tctx);
3413 * We shouldn't run task_works after cancel, so just leave
3414 * ->in_cancel set for normal exit.
3416 atomic_dec(&tctx->in_cancel);
3417 /* for exec all current's requests should be gone, kill tctx */
3418 __io_uring_free(current);
3422 void __io_uring_cancel(bool cancel_all)
3424 io_uring_cancel_generic(cancel_all, NULL);
3427 static void *io_uring_validate_mmap_request(struct file *file,
3428 loff_t pgoff, size_t sz)
3430 struct io_ring_ctx *ctx = file->private_data;
3431 loff_t offset = pgoff << PAGE_SHIFT;
3435 switch (offset & IORING_OFF_MMAP_MASK) {
3436 case IORING_OFF_SQ_RING:
3437 case IORING_OFF_CQ_RING:
3438 /* Don't allow mmap if the ring was setup without it */
3439 if (ctx->flags & IORING_SETUP_NO_MMAP)
3440 return ERR_PTR(-EINVAL);
3443 case IORING_OFF_SQES:
3444 /* Don't allow mmap if the ring was setup without it */
3445 if (ctx->flags & IORING_SETUP_NO_MMAP)
3446 return ERR_PTR(-EINVAL);
3449 case IORING_OFF_PBUF_RING: {
3450 struct io_buffer_list *bl;
3453 bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT;
3454 bl = io_pbuf_get_bl(ctx, bgid);
3462 return ERR_PTR(-EINVAL);
3465 page = virt_to_head_page(ptr);
3466 if (sz > page_size(page))
3467 return ERR_PTR(-EINVAL);
3474 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3476 size_t sz = vma->vm_end - vma->vm_start;
3480 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3482 return PTR_ERR(ptr);
3484 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3485 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3488 static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp,
3489 unsigned long addr, unsigned long len,
3490 unsigned long pgoff, unsigned long flags)
3495 * Do not allow to map to user-provided address to avoid breaking the
3496 * aliasing rules. Userspace is not able to guess the offset address of
3497 * kernel kmalloc()ed memory area.
3502 ptr = io_uring_validate_mmap_request(filp, pgoff, len);
3507 * Some architectures have strong cache aliasing requirements.
3508 * For such architectures we need a coherent mapping which aliases
3509 * kernel memory *and* userspace memory. To achieve that:
3510 * - use a NULL file pointer to reference physical memory, and
3511 * - use the kernel virtual address of the shared io_uring context
3512 * (instead of the userspace-provided address, which has to be 0UL
3514 * - use the same pgoff which the get_unmapped_area() uses to
3515 * calculate the page colouring.
3516 * For architectures without such aliasing requirements, the
3517 * architecture will return any suitable mapping because addr is 0.
3520 flags |= MAP_SHARED;
3521 pgoff = 0; /* has been translated to ptr above */
3523 addr = (uintptr_t) ptr;
3524 pgoff = addr >> PAGE_SHIFT;
3528 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
3531 #else /* !CONFIG_MMU */
3533 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3535 return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL;
3538 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3540 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3543 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3544 unsigned long addr, unsigned long len,
3545 unsigned long pgoff, unsigned long flags)
3549 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3551 return PTR_ERR(ptr);
3553 return (unsigned long) ptr;
3556 #endif /* !CONFIG_MMU */
3558 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3560 if (flags & IORING_ENTER_EXT_ARG) {
3561 struct io_uring_getevents_arg arg;
3563 if (argsz != sizeof(arg))
3565 if (copy_from_user(&arg, argp, sizeof(arg)))
3571 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3572 struct __kernel_timespec __user **ts,
3573 const sigset_t __user **sig)
3575 struct io_uring_getevents_arg arg;
3578 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3579 * is just a pointer to the sigset_t.
3581 if (!(flags & IORING_ENTER_EXT_ARG)) {
3582 *sig = (const sigset_t __user *) argp;
3588 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3589 * timespec and sigset_t pointers if good.
3591 if (*argsz != sizeof(arg))
3593 if (copy_from_user(&arg, argp, sizeof(arg)))
3597 *sig = u64_to_user_ptr(arg.sigmask);
3598 *argsz = arg.sigmask_sz;
3599 *ts = u64_to_user_ptr(arg.ts);
3603 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3604 u32, min_complete, u32, flags, const void __user *, argp,
3607 struct io_ring_ctx *ctx;
3611 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3612 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3613 IORING_ENTER_REGISTERED_RING)))
3617 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3618 * need only dereference our task private array to find it.
3620 if (flags & IORING_ENTER_REGISTERED_RING) {
3621 struct io_uring_task *tctx = current->io_uring;
3623 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3625 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3626 file = tctx->registered_rings[fd];
3627 if (unlikely(!file))
3631 if (unlikely(!file))
3634 if (unlikely(!io_is_uring_fops(file)))
3638 ctx = file->private_data;
3640 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3644 * For SQ polling, the thread will do all submissions and completions.
3645 * Just return the requested submit count, and wake the thread if
3649 if (ctx->flags & IORING_SETUP_SQPOLL) {
3650 io_cqring_overflow_flush(ctx);
3652 if (unlikely(ctx->sq_data->thread == NULL)) {
3656 if (flags & IORING_ENTER_SQ_WAKEUP)
3657 wake_up(&ctx->sq_data->wait);
3658 if (flags & IORING_ENTER_SQ_WAIT)
3659 io_sqpoll_wait_sq(ctx);
3662 } else if (to_submit) {
3663 ret = io_uring_add_tctx_node(ctx);
3667 mutex_lock(&ctx->uring_lock);
3668 ret = io_submit_sqes(ctx, to_submit);
3669 if (ret != to_submit) {
3670 mutex_unlock(&ctx->uring_lock);
3673 if (flags & IORING_ENTER_GETEVENTS) {
3674 if (ctx->syscall_iopoll)
3677 * Ignore errors, we'll soon call io_cqring_wait() and
3678 * it should handle ownership problems if any.
3680 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3681 (void)io_run_local_work_locked(ctx, min_complete);
3683 mutex_unlock(&ctx->uring_lock);
3686 if (flags & IORING_ENTER_GETEVENTS) {
3689 if (ctx->syscall_iopoll) {
3691 * We disallow the app entering submit/complete with
3692 * polling, but we still need to lock the ring to
3693 * prevent racing with polled issue that got punted to
3696 mutex_lock(&ctx->uring_lock);
3698 ret2 = io_validate_ext_arg(flags, argp, argsz);
3699 if (likely(!ret2)) {
3700 min_complete = min(min_complete,
3702 ret2 = io_iopoll_check(ctx, min_complete);
3704 mutex_unlock(&ctx->uring_lock);
3706 const sigset_t __user *sig;
3707 struct __kernel_timespec __user *ts;
3709 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3710 if (likely(!ret2)) {
3711 min_complete = min(min_complete,
3713 ret2 = io_cqring_wait(ctx, min_complete, sig,
3722 * EBADR indicates that one or more CQE were dropped.
3723 * Once the user has been informed we can clear the bit
3724 * as they are obviously ok with those drops.
3726 if (unlikely(ret2 == -EBADR))
3727 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3732 if (!(flags & IORING_ENTER_REGISTERED_RING))
3737 static const struct file_operations io_uring_fops = {
3738 .release = io_uring_release,
3739 .mmap = io_uring_mmap,
3741 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3742 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3744 .get_unmapped_area = io_uring_mmu_get_unmapped_area,
3746 .poll = io_uring_poll,
3747 #ifdef CONFIG_PROC_FS
3748 .show_fdinfo = io_uring_show_fdinfo,
3752 bool io_is_uring_fops(struct file *file)
3754 return file->f_op == &io_uring_fops;
3757 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3758 struct io_uring_params *p)
3760 struct io_rings *rings;
3761 size_t size, sq_array_offset;
3764 /* make sure these are sane, as we already accounted them */
3765 ctx->sq_entries = p->sq_entries;
3766 ctx->cq_entries = p->cq_entries;
3768 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3769 if (size == SIZE_MAX)
3772 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3773 rings = io_mem_alloc(size);
3775 rings = io_rings_map(ctx, p->cq_off.user_addr, size);
3778 return PTR_ERR(rings);
3781 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY))
3782 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3783 rings->sq_ring_mask = p->sq_entries - 1;
3784 rings->cq_ring_mask = p->cq_entries - 1;
3785 rings->sq_ring_entries = p->sq_entries;
3786 rings->cq_ring_entries = p->cq_entries;
3788 if (p->flags & IORING_SETUP_SQE128)
3789 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3791 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3792 if (size == SIZE_MAX) {
3797 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3798 ptr = io_mem_alloc(size);
3800 ptr = io_sqes_map(ctx, p->sq_off.user_addr, size);
3804 return PTR_ERR(ptr);
3811 static int io_uring_install_fd(struct file *file)
3815 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3818 fd_install(fd, file);
3823 * Allocate an anonymous fd, this is what constitutes the application
3824 * visible backing of an io_uring instance. The application mmaps this
3825 * fd to gain access to the SQ/CQ ring details.
3827 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3829 /* Create a new inode so that the LSM can block the creation. */
3830 return anon_inode_create_getfile("[io_uring]", &io_uring_fops, ctx,
3831 O_RDWR | O_CLOEXEC, NULL);
3834 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3835 struct io_uring_params __user *params)
3837 struct io_ring_ctx *ctx;
3838 struct io_uring_task *tctx;
3844 if (entries > IORING_MAX_ENTRIES) {
3845 if (!(p->flags & IORING_SETUP_CLAMP))
3847 entries = IORING_MAX_ENTRIES;
3850 if ((p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
3851 && !(p->flags & IORING_SETUP_NO_MMAP))
3855 * Use twice as many entries for the CQ ring. It's possible for the
3856 * application to drive a higher depth than the size of the SQ ring,
3857 * since the sqes are only used at submission time. This allows for
3858 * some flexibility in overcommitting a bit. If the application has
3859 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3860 * of CQ ring entries manually.
3862 p->sq_entries = roundup_pow_of_two(entries);
3863 if (p->flags & IORING_SETUP_CQSIZE) {
3865 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3866 * to a power-of-two, if it isn't already. We do NOT impose
3867 * any cq vs sq ring sizing.
3871 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3872 if (!(p->flags & IORING_SETUP_CLAMP))
3874 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3876 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3877 if (p->cq_entries < p->sq_entries)
3880 p->cq_entries = 2 * p->sq_entries;
3883 ctx = io_ring_ctx_alloc(p);
3887 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3888 !(ctx->flags & IORING_SETUP_IOPOLL) &&
3889 !(ctx->flags & IORING_SETUP_SQPOLL))
3890 ctx->task_complete = true;
3892 if (ctx->task_complete || (ctx->flags & IORING_SETUP_IOPOLL))
3893 ctx->lockless_cq = true;
3896 * lazy poll_wq activation relies on ->task_complete for synchronisation
3897 * purposes, see io_activate_pollwq()
3899 if (!ctx->task_complete)
3900 ctx->poll_activated = true;
3903 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3904 * space applications don't need to do io completion events
3905 * polling again, they can rely on io_sq_thread to do polling
3906 * work, which can reduce cpu usage and uring_lock contention.
3908 if (ctx->flags & IORING_SETUP_IOPOLL &&
3909 !(ctx->flags & IORING_SETUP_SQPOLL))
3910 ctx->syscall_iopoll = 1;
3912 ctx->compat = in_compat_syscall();
3913 if (!ns_capable_noaudit(&init_user_ns, CAP_IPC_LOCK))
3914 ctx->user = get_uid(current_user());
3917 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3918 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3921 if (ctx->flags & IORING_SETUP_SQPOLL) {
3922 /* IPI related flags don't make sense with SQPOLL */
3923 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3924 IORING_SETUP_TASKRUN_FLAG |
3925 IORING_SETUP_DEFER_TASKRUN))
3927 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3928 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3929 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3931 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3932 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3934 ctx->notify_method = TWA_SIGNAL;
3938 * For DEFER_TASKRUN we require the completion task to be the same as the
3939 * submission task. This implies that there is only one submitter, so enforce
3942 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3943 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3948 * This is just grabbed for accounting purposes. When a process exits,
3949 * the mm is exited and dropped before the files, hence we need to hang
3950 * on to this mm purely for the purposes of being able to unaccount
3951 * memory (locked/pinned vm). It's not used for anything else.
3953 mmgrab(current->mm);
3954 ctx->mm_account = current->mm;
3956 ret = io_allocate_scq_urings(ctx, p);
3960 ret = io_sq_offload_create(ctx, p);
3964 ret = io_rsrc_init(ctx);
3968 p->sq_off.head = offsetof(struct io_rings, sq.head);
3969 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3970 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3971 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3972 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3973 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3974 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY))
3975 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3976 p->sq_off.resv1 = 0;
3977 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3978 p->sq_off.user_addr = 0;
3980 p->cq_off.head = offsetof(struct io_rings, cq.head);
3981 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3982 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3983 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3984 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3985 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3986 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3987 p->cq_off.resv1 = 0;
3988 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3989 p->cq_off.user_addr = 0;
3991 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3992 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3993 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3994 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3995 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3996 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3997 IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING;
3999 if (copy_to_user(params, p, sizeof(*p))) {
4004 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
4005 && !(ctx->flags & IORING_SETUP_R_DISABLED))
4006 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
4008 file = io_uring_get_file(ctx);
4010 ret = PTR_ERR(file);
4014 ret = __io_uring_add_tctx_node(ctx);
4017 tctx = current->io_uring;
4020 * Install ring fd as the very last thing, so we don't risk someone
4021 * having closed it before we finish setup
4023 if (p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
4024 ret = io_ring_add_registered_file(tctx, file, 0, IO_RINGFD_REG_MAX);
4026 ret = io_uring_install_fd(file);
4030 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
4033 io_ring_ctx_wait_and_kill(ctx);
4041 * Sets up an aio uring context, and returns the fd. Applications asks for a
4042 * ring size, we return the actual sq/cq ring sizes (among other things) in the
4043 * params structure passed in.
4045 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
4047 struct io_uring_params p;
4050 if (copy_from_user(&p, params, sizeof(p)))
4052 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
4057 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
4058 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
4059 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
4060 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
4061 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
4062 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
4063 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN |
4064 IORING_SETUP_NO_MMAP | IORING_SETUP_REGISTERED_FD_ONLY |
4065 IORING_SETUP_NO_SQARRAY))
4068 return io_uring_create(entries, &p, params);
4071 static inline bool io_uring_allowed(void)
4073 int disabled = READ_ONCE(sysctl_io_uring_disabled);
4074 kgid_t io_uring_group;
4079 if (disabled == 0 || capable(CAP_SYS_ADMIN))
4082 io_uring_group = make_kgid(&init_user_ns, sysctl_io_uring_group);
4083 if (!gid_valid(io_uring_group))
4086 return in_group_p(io_uring_group);
4089 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
4090 struct io_uring_params __user *, params)
4092 if (!io_uring_allowed())
4095 return io_uring_setup(entries, params);
4098 static int __init io_uring_init(void)
4100 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4101 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4102 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4105 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4106 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4107 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4108 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4109 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4110 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
4111 BUILD_BUG_SQE_ELEM(1, __u8, flags);
4112 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
4113 BUILD_BUG_SQE_ELEM(4, __s32, fd);
4114 BUILD_BUG_SQE_ELEM(8, __u64, off);
4115 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
4116 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
4117 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
4118 BUILD_BUG_SQE_ELEM(16, __u64, addr);
4119 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
4120 BUILD_BUG_SQE_ELEM(24, __u32, len);
4121 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
4122 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
4123 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4124 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
4125 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
4126 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
4127 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
4128 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
4129 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
4130 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
4131 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
4132 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
4133 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
4134 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
4135 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
4136 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
4137 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
4138 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
4139 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
4140 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
4141 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
4142 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
4143 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
4144 BUILD_BUG_SQE_ELEM(42, __u16, personality);
4145 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
4146 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
4147 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
4148 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
4149 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
4150 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4151 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
4153 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4154 sizeof(struct io_uring_rsrc_update));
4155 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4156 sizeof(struct io_uring_rsrc_update2));
4158 /* ->buf_index is u16 */
4159 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4160 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4161 offsetof(struct io_uring_buf_ring, tail));
4163 /* should fit into one byte */
4164 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
4165 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4166 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4168 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof_field(struct io_kiocb, flags));
4170 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4172 /* top 8bits are for internal use */
4173 BUILD_BUG_ON((IORING_URING_CMD_MASK & 0xff000000) != 0);
4175 io_uring_optable_init();
4178 * Allow user copy in the per-command field, which starts after the
4179 * file in io_kiocb and until the opcode field. The openat2 handling
4180 * requires copying in user memory into the io_kiocb object in that
4181 * range, and HARDENED_USERCOPY will complain if we haven't
4182 * correctly annotated this range.
4184 req_cachep = kmem_cache_create_usercopy("io_kiocb",
4185 sizeof(struct io_kiocb), 0,
4186 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4187 SLAB_ACCOUNT | SLAB_TYPESAFE_BY_RCU,
4188 offsetof(struct io_kiocb, cmd.data),
4189 sizeof_field(struct io_kiocb, cmd.data), NULL);
4190 io_buf_cachep = KMEM_CACHE(io_buffer,
4191 SLAB_HWCACHE_ALIGN | SLAB_PANIC | SLAB_ACCOUNT);
4193 iou_wq = alloc_workqueue("iou_exit", WQ_UNBOUND, 64);
4195 #ifdef CONFIG_SYSCTL
4196 register_sysctl_init("kernel", kernel_io_uring_disabled_table);
4201 __initcall(io_uring_init);