1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <net/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
49 #include <linux/bits.h>
51 #include <linux/sched/signal.h>
53 #include <linux/file.h>
54 #include <linux/fdtable.h>
56 #include <linux/mman.h>
57 #include <linux/percpu.h>
58 #include <linux/slab.h>
59 #include <linux/bvec.h>
60 #include <linux/net.h>
62 #include <net/af_unix.h>
64 #include <linux/anon_inodes.h>
65 #include <linux/sched/mm.h>
66 #include <linux/uaccess.h>
67 #include <linux/nospec.h>
68 #include <linux/highmem.h>
69 #include <linux/fsnotify.h>
70 #include <linux/fadvise.h>
71 #include <linux/task_work.h>
72 #include <linux/io_uring.h>
73 #include <linux/audit.h>
74 #include <linux/security.h>
76 #define CREATE_TRACE_POINTS
77 #include <trace/events/io_uring.h>
79 #include <uapi/linux/io_uring.h>
95 #include "alloc_cache.h"
97 #define IORING_MAX_ENTRIES 32768
98 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
100 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
101 IORING_REGISTER_LAST + IORING_OP_LAST)
103 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
104 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
106 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
107 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
109 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
110 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
113 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
116 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
118 #define IO_COMPL_BATCH 32
119 #define IO_REQ_ALLOC_BATCH 8
122 IO_CHECK_CQ_OVERFLOW_BIT,
123 IO_CHECK_CQ_DROPPED_BIT,
126 struct io_defer_entry {
127 struct list_head list;
128 struct io_kiocb *req;
132 /* requests with any of those set should undergo io_disarm_next() */
133 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
134 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
136 static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
137 struct task_struct *task,
140 static void io_dismantle_req(struct io_kiocb *req);
141 static void io_clean_op(struct io_kiocb *req);
142 static void io_queue_sqe(struct io_kiocb *req);
144 static void __io_submit_flush_completions(struct io_ring_ctx *ctx);
146 static struct kmem_cache *req_cachep;
148 struct sock *io_uring_get_socket(struct file *file)
150 #if defined(CONFIG_UNIX)
151 if (io_is_uring_fops(file)) {
152 struct io_ring_ctx *ctx = file->private_data;
154 return ctx->ring_sock->sk;
159 EXPORT_SYMBOL(io_uring_get_socket);
161 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
163 if (!wq_list_empty(&ctx->submit_state.compl_reqs))
164 __io_submit_flush_completions(ctx);
167 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
169 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
172 static bool io_match_linked(struct io_kiocb *head)
174 struct io_kiocb *req;
176 io_for_each_link(req, head) {
177 if (req->flags & REQ_F_INFLIGHT)
184 * As io_match_task() but protected against racing with linked timeouts.
185 * User must not hold timeout_lock.
187 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
192 if (task && head->task != task)
197 if (head->flags & REQ_F_LINK_TIMEOUT) {
198 struct io_ring_ctx *ctx = head->ctx;
200 /* protect against races with linked timeouts */
201 spin_lock_irq(&ctx->timeout_lock);
202 matched = io_match_linked(head);
203 spin_unlock_irq(&ctx->timeout_lock);
205 matched = io_match_linked(head);
210 static inline void req_fail_link_node(struct io_kiocb *req, int res)
213 io_req_set_res(req, res, 0);
216 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
218 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
221 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
223 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
225 complete(&ctx->ref_comp);
228 static __cold void io_fallback_req_func(struct work_struct *work)
230 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
232 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
233 struct io_kiocb *req, *tmp;
236 percpu_ref_get(&ctx->refs);
237 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
238 req->io_task_work.func(req, &locked);
241 io_submit_flush_completions(ctx);
242 mutex_unlock(&ctx->uring_lock);
244 percpu_ref_put(&ctx->refs);
247 static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
249 unsigned hash_buckets = 1U << bits;
250 size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
252 table->hbs = kmalloc(hash_size, GFP_KERNEL);
256 table->hash_bits = bits;
257 init_hash_table(table, hash_buckets);
261 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
263 struct io_ring_ctx *ctx;
266 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
270 xa_init(&ctx->io_bl_xa);
273 * Use 5 bits less than the max cq entries, that should give us around
274 * 32 entries per hash list if totally full and uniformly spread, but
275 * don't keep too many buckets to not overconsume memory.
277 hash_bits = ilog2(p->cq_entries) - 5;
278 hash_bits = clamp(hash_bits, 1, 8);
279 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
281 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
284 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
285 if (!ctx->dummy_ubuf)
287 /* set invalid range, so io_import_fixed() fails meeting it */
288 ctx->dummy_ubuf->ubuf = -1UL;
290 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
291 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
294 ctx->flags = p->flags;
295 init_waitqueue_head(&ctx->sqo_sq_wait);
296 INIT_LIST_HEAD(&ctx->sqd_list);
297 INIT_LIST_HEAD(&ctx->cq_overflow_list);
298 INIT_LIST_HEAD(&ctx->io_buffers_cache);
299 io_alloc_cache_init(&ctx->apoll_cache);
300 init_completion(&ctx->ref_comp);
301 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
302 mutex_init(&ctx->uring_lock);
303 init_waitqueue_head(&ctx->cq_wait);
304 spin_lock_init(&ctx->completion_lock);
305 spin_lock_init(&ctx->timeout_lock);
306 INIT_WQ_LIST(&ctx->iopoll_list);
307 INIT_LIST_HEAD(&ctx->io_buffers_pages);
308 INIT_LIST_HEAD(&ctx->io_buffers_comp);
309 INIT_LIST_HEAD(&ctx->defer_list);
310 INIT_LIST_HEAD(&ctx->timeout_list);
311 INIT_LIST_HEAD(&ctx->ltimeout_list);
312 spin_lock_init(&ctx->rsrc_ref_lock);
313 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
314 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
315 init_llist_head(&ctx->rsrc_put_llist);
316 INIT_LIST_HEAD(&ctx->tctx_list);
317 ctx->submit_state.free_list.next = NULL;
318 INIT_WQ_LIST(&ctx->locked_free_list);
319 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
320 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
323 kfree(ctx->dummy_ubuf);
324 kfree(ctx->cancel_table.hbs);
325 kfree(ctx->cancel_table_locked.hbs);
327 xa_destroy(&ctx->io_bl_xa);
332 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
334 struct io_rings *r = ctx->rings;
336 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
340 static bool req_need_defer(struct io_kiocb *req, u32 seq)
342 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
343 struct io_ring_ctx *ctx = req->ctx;
345 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
351 static inline void io_req_track_inflight(struct io_kiocb *req)
353 if (!(req->flags & REQ_F_INFLIGHT)) {
354 req->flags |= REQ_F_INFLIGHT;
355 atomic_inc(&req->task->io_uring->inflight_tracked);
359 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
361 if (WARN_ON_ONCE(!req->link))
364 req->flags &= ~REQ_F_ARM_LTIMEOUT;
365 req->flags |= REQ_F_LINK_TIMEOUT;
367 /* linked timeouts should have two refs once prep'ed */
368 io_req_set_refcount(req);
369 __io_req_set_refcount(req->link, 2);
373 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
375 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
377 return __io_prep_linked_timeout(req);
380 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
382 io_queue_linked_timeout(__io_prep_linked_timeout(req));
385 static inline void io_arm_ltimeout(struct io_kiocb *req)
387 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
388 __io_arm_ltimeout(req);
391 static void io_prep_async_work(struct io_kiocb *req)
393 const struct io_op_def *def = &io_op_defs[req->opcode];
394 struct io_ring_ctx *ctx = req->ctx;
396 if (!(req->flags & REQ_F_CREDS)) {
397 req->flags |= REQ_F_CREDS;
398 req->creds = get_current_cred();
401 req->work.list.next = NULL;
403 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
404 if (req->flags & REQ_F_FORCE_ASYNC)
405 req->work.flags |= IO_WQ_WORK_CONCURRENT;
407 if (req->flags & REQ_F_ISREG) {
408 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
409 io_wq_hash_work(&req->work, file_inode(req->file));
410 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
411 if (def->unbound_nonreg_file)
412 req->work.flags |= IO_WQ_WORK_UNBOUND;
416 static void io_prep_async_link(struct io_kiocb *req)
418 struct io_kiocb *cur;
420 if (req->flags & REQ_F_LINK_TIMEOUT) {
421 struct io_ring_ctx *ctx = req->ctx;
423 spin_lock_irq(&ctx->timeout_lock);
424 io_for_each_link(cur, req)
425 io_prep_async_work(cur);
426 spin_unlock_irq(&ctx->timeout_lock);
428 io_for_each_link(cur, req)
429 io_prep_async_work(cur);
433 void io_queue_iowq(struct io_kiocb *req, bool *dont_use)
435 struct io_kiocb *link = io_prep_linked_timeout(req);
436 struct io_uring_task *tctx = req->task->io_uring;
439 BUG_ON(!tctx->io_wq);
441 /* init ->work of the whole link before punting */
442 io_prep_async_link(req);
445 * Not expected to happen, but if we do have a bug where this _can_
446 * happen, catch it here and ensure the request is marked as
447 * canceled. That will make io-wq go through the usual work cancel
448 * procedure rather than attempt to run this request (or create a new
451 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
452 req->work.flags |= IO_WQ_WORK_CANCEL;
454 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
455 io_wq_enqueue(tctx->io_wq, &req->work);
457 io_queue_linked_timeout(link);
460 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
462 while (!list_empty(&ctx->defer_list)) {
463 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
464 struct io_defer_entry, list);
466 if (req_need_defer(de->req, de->seq))
468 list_del_init(&de->list);
469 io_req_task_queue(de->req);
474 static void io_eventfd_signal(struct io_ring_ctx *ctx)
476 struct io_ev_fd *ev_fd;
479 spin_lock(&ctx->completion_lock);
481 * Eventfd should only get triggered when at least one event has been
482 * posted. Some applications rely on the eventfd notification count only
483 * changing IFF a new CQE has been added to the CQ ring. There's no
484 * depedency on 1:1 relationship between how many times this function is
485 * called (and hence the eventfd count) and number of CQEs posted to the
488 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
489 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
490 spin_unlock(&ctx->completion_lock);
496 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
499 ev_fd = rcu_dereference(ctx->io_ev_fd);
502 * Check again if ev_fd exists incase an io_eventfd_unregister call
503 * completed between the NULL check of ctx->io_ev_fd at the start of
504 * the function and rcu_read_lock.
506 if (unlikely(!ev_fd))
508 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
511 if (!ev_fd->eventfd_async || io_wq_current_is_worker())
512 eventfd_signal(ev_fd->cq_ev_fd, 1);
517 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
519 if (ctx->off_timeout_used || ctx->drain_active) {
520 spin_lock(&ctx->completion_lock);
521 if (ctx->off_timeout_used)
522 io_flush_timeouts(ctx);
523 if (ctx->drain_active)
524 io_queue_deferred(ctx);
525 spin_unlock(&ctx->completion_lock);
528 io_eventfd_signal(ctx);
531 static inline void io_cqring_ev_posted(struct io_ring_ctx *ctx)
533 io_commit_cqring_flush(ctx);
537 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
538 __releases(ctx->completion_lock)
540 io_commit_cqring(ctx);
541 spin_unlock(&ctx->completion_lock);
542 io_cqring_ev_posted(ctx);
545 void io_cq_unlock_post(struct io_ring_ctx *ctx)
547 __io_cq_unlock_post(ctx);
550 /* Returns true if there are no backlogged entries after the flush */
551 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
554 size_t cqe_size = sizeof(struct io_uring_cqe);
556 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
559 if (ctx->flags & IORING_SETUP_CQE32)
563 while (!list_empty(&ctx->cq_overflow_list)) {
564 struct io_uring_cqe *cqe = io_get_cqe(ctx);
565 struct io_overflow_cqe *ocqe;
569 ocqe = list_first_entry(&ctx->cq_overflow_list,
570 struct io_overflow_cqe, list);
572 memcpy(cqe, &ocqe->cqe, cqe_size);
574 io_account_cq_overflow(ctx);
576 list_del(&ocqe->list);
580 all_flushed = list_empty(&ctx->cq_overflow_list);
582 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
583 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
586 io_cq_unlock_post(ctx);
590 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
594 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
595 /* iopoll syncs against uring_lock, not completion_lock */
596 if (ctx->flags & IORING_SETUP_IOPOLL)
597 mutex_lock(&ctx->uring_lock);
598 ret = __io_cqring_overflow_flush(ctx, false);
599 if (ctx->flags & IORING_SETUP_IOPOLL)
600 mutex_unlock(&ctx->uring_lock);
606 static void __io_put_task(struct task_struct *task, int nr)
608 struct io_uring_task *tctx = task->io_uring;
610 percpu_counter_sub(&tctx->inflight, nr);
611 if (unlikely(atomic_read(&tctx->in_idle)))
612 wake_up(&tctx->wait);
613 put_task_struct_many(task, nr);
616 /* must to be called somewhat shortly after putting a request */
617 static inline void io_put_task(struct task_struct *task, int nr)
619 if (likely(task == current))
620 task->io_uring->cached_refs += nr;
622 __io_put_task(task, nr);
625 static void io_task_refs_refill(struct io_uring_task *tctx)
627 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
629 percpu_counter_add(&tctx->inflight, refill);
630 refcount_add(refill, ¤t->usage);
631 tctx->cached_refs += refill;
634 static inline void io_get_task_refs(int nr)
636 struct io_uring_task *tctx = current->io_uring;
638 tctx->cached_refs -= nr;
639 if (unlikely(tctx->cached_refs < 0))
640 io_task_refs_refill(tctx);
643 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
645 struct io_uring_task *tctx = task->io_uring;
646 unsigned int refs = tctx->cached_refs;
649 tctx->cached_refs = 0;
650 percpu_counter_sub(&tctx->inflight, refs);
651 put_task_struct_many(task, refs);
655 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
656 s32 res, u32 cflags, u64 extra1, u64 extra2)
658 struct io_overflow_cqe *ocqe;
659 size_t ocq_size = sizeof(struct io_overflow_cqe);
660 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
663 ocq_size += sizeof(struct io_uring_cqe);
665 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
666 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
669 * If we're in ring overflow flush mode, or in task cancel mode,
670 * or cannot allocate an overflow entry, then we need to drop it
673 io_account_cq_overflow(ctx);
674 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
677 if (list_empty(&ctx->cq_overflow_list)) {
678 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
679 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
682 ocqe->cqe.user_data = user_data;
684 ocqe->cqe.flags = cflags;
686 ocqe->cqe.big_cqe[0] = extra1;
687 ocqe->cqe.big_cqe[1] = extra2;
689 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
693 bool io_req_cqe_overflow(struct io_kiocb *req)
695 if (!(req->flags & REQ_F_CQE32_INIT)) {
699 return io_cqring_event_overflow(req->ctx, req->cqe.user_data,
700 req->cqe.res, req->cqe.flags,
701 req->extra1, req->extra2);
705 * writes to the cq entry need to come after reading head; the
706 * control dependency is enough as we're using WRITE_ONCE to
709 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx)
711 struct io_rings *rings = ctx->rings;
712 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
713 unsigned int free, queued, len;
716 /* userspace may cheat modifying the tail, be safe and do min */
717 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
718 free = ctx->cq_entries - queued;
719 /* we need a contiguous range, limit based on the current array offset */
720 len = min(free, ctx->cq_entries - off);
724 if (ctx->flags & IORING_SETUP_CQE32) {
729 ctx->cqe_cached = &rings->cqes[off];
730 ctx->cqe_sentinel = ctx->cqe_cached + len;
732 ctx->cached_cq_tail++;
734 if (ctx->flags & IORING_SETUP_CQE32)
736 return &rings->cqes[off];
739 static bool io_fill_cqe_aux(struct io_ring_ctx *ctx,
740 u64 user_data, s32 res, u32 cflags,
743 struct io_uring_cqe *cqe;
748 * If we can't get a cq entry, userspace overflowed the
749 * submission (by quite a lot). Increment the overflow count in
752 cqe = io_get_cqe(ctx);
754 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
756 WRITE_ONCE(cqe->user_data, user_data);
757 WRITE_ONCE(cqe->res, res);
758 WRITE_ONCE(cqe->flags, cflags);
760 if (ctx->flags & IORING_SETUP_CQE32) {
761 WRITE_ONCE(cqe->big_cqe[0], 0);
762 WRITE_ONCE(cqe->big_cqe[1], 0);
768 return io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
773 bool io_post_aux_cqe(struct io_ring_ctx *ctx,
774 u64 user_data, s32 res, u32 cflags,
780 filled = io_fill_cqe_aux(ctx, user_data, res, cflags, allow_overflow);
781 io_cq_unlock_post(ctx);
785 static void __io_req_complete_put(struct io_kiocb *req)
788 * If we're the last reference to this request, add to our locked
791 if (req_ref_put_and_test(req)) {
792 struct io_ring_ctx *ctx = req->ctx;
794 if (req->flags & IO_REQ_LINK_FLAGS) {
795 if (req->flags & IO_DISARM_MASK)
798 io_req_task_queue(req->link);
802 io_req_put_rsrc(req);
804 * Selected buffer deallocation in io_clean_op() assumes that
805 * we don't hold ->completion_lock. Clean them here to avoid
808 io_put_kbuf_comp(req);
809 io_dismantle_req(req);
810 io_put_task(req->task, 1);
811 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
812 ctx->locked_free_nr++;
816 void __io_req_complete_post(struct io_kiocb *req)
818 if (!(req->flags & REQ_F_CQE_SKIP))
819 __io_fill_cqe_req(req->ctx, req);
820 __io_req_complete_put(req);
823 void io_req_complete_post(struct io_kiocb *req)
825 struct io_ring_ctx *ctx = req->ctx;
828 __io_req_complete_post(req);
829 io_cq_unlock_post(ctx);
832 inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags)
834 io_req_complete_post(req);
837 void io_req_complete_failed(struct io_kiocb *req, s32 res)
840 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
841 io_req_complete_post(req);
845 * Don't initialise the fields below on every allocation, but do that in
846 * advance and keep them valid across allocations.
848 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
852 req->async_data = NULL;
853 /* not necessary, but safer to zero */
857 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
858 struct io_submit_state *state)
860 spin_lock(&ctx->completion_lock);
861 wq_list_splice(&ctx->locked_free_list, &state->free_list);
862 ctx->locked_free_nr = 0;
863 spin_unlock(&ctx->completion_lock);
866 static inline bool io_req_cache_empty(struct io_ring_ctx *ctx)
868 return !ctx->submit_state.free_list.next;
872 * A request might get retired back into the request caches even before opcode
873 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
874 * Because of that, io_alloc_req() should be called only under ->uring_lock
875 * and with extra caution to not get a request that is still worked on.
877 static __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
878 __must_hold(&ctx->uring_lock)
880 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
881 void *reqs[IO_REQ_ALLOC_BATCH];
885 * If we have more than a batch's worth of requests in our IRQ side
886 * locked cache, grab the lock and move them over to our submission
889 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
890 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
891 if (!io_req_cache_empty(ctx))
895 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
898 * Bulk alloc is all-or-nothing. If we fail to get a batch,
899 * retry single alloc to be on the safe side.
901 if (unlikely(ret <= 0)) {
902 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
908 percpu_ref_get_many(&ctx->refs, ret);
909 for (i = 0; i < ret; i++) {
910 struct io_kiocb *req = reqs[i];
912 io_preinit_req(req, ctx);
913 io_req_add_to_cache(req, ctx);
918 static inline bool io_alloc_req_refill(struct io_ring_ctx *ctx)
920 if (unlikely(io_req_cache_empty(ctx)))
921 return __io_alloc_req_refill(ctx);
925 static inline struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
927 struct io_wq_work_node *node;
929 node = wq_stack_extract(&ctx->submit_state.free_list);
930 return container_of(node, struct io_kiocb, comp_list);
933 static inline void io_dismantle_req(struct io_kiocb *req)
935 unsigned int flags = req->flags;
937 if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
939 if (!(flags & REQ_F_FIXED_FILE))
940 io_put_file(req->file);
943 __cold void io_free_req(struct io_kiocb *req)
945 struct io_ring_ctx *ctx = req->ctx;
947 io_req_put_rsrc(req);
948 io_dismantle_req(req);
949 io_put_task(req->task, 1);
951 spin_lock(&ctx->completion_lock);
952 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
953 ctx->locked_free_nr++;
954 spin_unlock(&ctx->completion_lock);
957 static void __io_req_find_next_prep(struct io_kiocb *req)
959 struct io_ring_ctx *ctx = req->ctx;
963 io_cq_unlock_post(ctx);
966 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
968 struct io_kiocb *nxt;
971 * If LINK is set, we have dependent requests in this chain. If we
972 * didn't fail this request, queue the first one up, moving any other
973 * dependencies to the next request. In case of failure, fail the rest
976 if (unlikely(req->flags & IO_DISARM_MASK))
977 __io_req_find_next_prep(req);
983 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
987 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
988 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
990 io_submit_flush_completions(ctx);
991 mutex_unlock(&ctx->uring_lock);
994 percpu_ref_put(&ctx->refs);
997 static unsigned int handle_tw_list(struct llist_node *node,
998 struct io_ring_ctx **ctx, bool *locked,
999 struct llist_node *last)
1001 unsigned int count = 0;
1003 while (node != last) {
1004 struct llist_node *next = node->next;
1005 struct io_kiocb *req = container_of(node, struct io_kiocb,
1008 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1010 if (req->ctx != *ctx) {
1011 ctx_flush_and_put(*ctx, locked);
1013 /* if not contended, grab and improve batching */
1014 *locked = mutex_trylock(&(*ctx)->uring_lock);
1015 percpu_ref_get(&(*ctx)->refs);
1017 req->io_task_work.func(req, locked);
1026 * io_llist_xchg - swap all entries in a lock-less list
1027 * @head: the head of lock-less list to delete all entries
1028 * @new: new entry as the head of the list
1030 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1031 * The order of entries returned is from the newest to the oldest added one.
1033 static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1034 struct llist_node *new)
1036 return xchg(&head->first, new);
1040 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1041 * @head: the head of lock-less list to delete all entries
1042 * @old: expected old value of the first entry of the list
1043 * @new: new entry as the head of the list
1045 * perform a cmpxchg on the first entry of the list.
1048 static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1049 struct llist_node *old,
1050 struct llist_node *new)
1052 return cmpxchg(&head->first, old, new);
1055 void tctx_task_work(struct callback_head *cb)
1057 bool uring_locked = false;
1058 struct io_ring_ctx *ctx = NULL;
1059 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1061 struct llist_node fake = {};
1062 struct llist_node *node = io_llist_xchg(&tctx->task_list, &fake);
1063 unsigned int loops = 1;
1064 unsigned int count = handle_tw_list(node, &ctx, &uring_locked, NULL);
1066 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1067 while (node != &fake) {
1069 node = io_llist_xchg(&tctx->task_list, &fake);
1070 count += handle_tw_list(node, &ctx, &uring_locked, &fake);
1071 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1074 ctx_flush_and_put(ctx, &uring_locked);
1076 /* relaxed read is enough as only the task itself sets ->in_idle */
1077 if (unlikely(atomic_read(&tctx->in_idle)))
1078 io_uring_drop_tctx_refs(current);
1080 trace_io_uring_task_work_run(tctx, count, loops);
1083 void io_req_task_work_add(struct io_kiocb *req)
1085 struct io_uring_task *tctx = req->task->io_uring;
1086 struct io_ring_ctx *ctx = req->ctx;
1087 struct llist_node *node;
1090 running = !llist_add(&req->io_task_work.node, &tctx->task_list);
1092 /* task_work already pending, we're done */
1096 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1097 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1099 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1102 node = llist_del_all(&tctx->task_list);
1105 req = container_of(node, struct io_kiocb, io_task_work.node);
1107 if (llist_add(&req->io_task_work.node,
1108 &req->ctx->fallback_llist))
1109 schedule_delayed_work(&req->ctx->fallback_work, 1);
1113 static void io_req_tw_post(struct io_kiocb *req, bool *locked)
1115 io_req_complete_post(req);
1118 void io_req_tw_post_queue(struct io_kiocb *req, s32 res, u32 cflags)
1120 io_req_set_res(req, res, cflags);
1121 req->io_task_work.func = io_req_tw_post;
1122 io_req_task_work_add(req);
1125 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
1127 /* not needed for normal modes, but SQPOLL depends on it */
1128 io_tw_lock(req->ctx, locked);
1129 io_req_complete_failed(req, req->cqe.res);
1132 void io_req_task_submit(struct io_kiocb *req, bool *locked)
1134 io_tw_lock(req->ctx, locked);
1135 /* req->task == current here, checking PF_EXITING is safe */
1136 if (likely(!(req->task->flags & PF_EXITING)))
1139 io_req_complete_failed(req, -EFAULT);
1142 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1144 io_req_set_res(req, ret, 0);
1145 req->io_task_work.func = io_req_task_cancel;
1146 io_req_task_work_add(req);
1149 void io_req_task_queue(struct io_kiocb *req)
1151 req->io_task_work.func = io_req_task_submit;
1152 io_req_task_work_add(req);
1155 void io_queue_next(struct io_kiocb *req)
1157 struct io_kiocb *nxt = io_req_find_next(req);
1160 io_req_task_queue(nxt);
1163 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1164 __must_hold(&ctx->uring_lock)
1166 struct task_struct *task = NULL;
1170 struct io_kiocb *req = container_of(node, struct io_kiocb,
1173 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1174 if (req->flags & REQ_F_REFCOUNT) {
1175 node = req->comp_list.next;
1176 if (!req_ref_put_and_test(req))
1179 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1180 struct async_poll *apoll = req->apoll;
1182 if (apoll->double_poll)
1183 kfree(apoll->double_poll);
1184 io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache);
1185 req->flags &= ~REQ_F_POLLED;
1187 if (req->flags & IO_REQ_LINK_FLAGS)
1189 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1192 if (!(req->flags & REQ_F_FIXED_FILE))
1193 io_put_file(req->file);
1195 io_req_put_rsrc_locked(req, ctx);
1197 if (req->task != task) {
1199 io_put_task(task, task_refs);
1204 node = req->comp_list.next;
1205 io_req_add_to_cache(req, ctx);
1209 io_put_task(task, task_refs);
1212 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1213 __must_hold(&ctx->uring_lock)
1215 struct io_wq_work_node *node, *prev;
1216 struct io_submit_state *state = &ctx->submit_state;
1218 spin_lock(&ctx->completion_lock);
1219 wq_list_for_each(node, prev, &state->compl_reqs) {
1220 struct io_kiocb *req = container_of(node, struct io_kiocb,
1223 if (!(req->flags & REQ_F_CQE_SKIP))
1224 __io_fill_cqe_req(ctx, req);
1226 __io_cq_unlock_post(ctx);
1228 io_free_batch_list(ctx, state->compl_reqs.first);
1229 INIT_WQ_LIST(&state->compl_reqs);
1233 * Drop reference to request, return next in chain (if there is one) if this
1234 * was the last reference to this request.
1236 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1238 struct io_kiocb *nxt = NULL;
1240 if (req_ref_put_and_test(req)) {
1241 if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
1242 nxt = io_req_find_next(req);
1248 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1250 /* See comment at the top of this file */
1252 return __io_cqring_events(ctx);
1256 * We can't just wait for polled events to come to us, we have to actively
1257 * find and complete them.
1259 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1261 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1264 mutex_lock(&ctx->uring_lock);
1265 while (!wq_list_empty(&ctx->iopoll_list)) {
1266 /* let it sleep and repeat later if can't complete a request */
1267 if (io_do_iopoll(ctx, true) == 0)
1270 * Ensure we allow local-to-the-cpu processing to take place,
1271 * in this case we need to ensure that we reap all events.
1272 * Also let task_work, etc. to progress by releasing the mutex
1274 if (need_resched()) {
1275 mutex_unlock(&ctx->uring_lock);
1277 mutex_lock(&ctx->uring_lock);
1280 mutex_unlock(&ctx->uring_lock);
1283 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1285 unsigned int nr_events = 0;
1287 unsigned long check_cq;
1289 check_cq = READ_ONCE(ctx->check_cq);
1290 if (unlikely(check_cq)) {
1291 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1292 __io_cqring_overflow_flush(ctx, false);
1294 * Similarly do not spin if we have not informed the user of any
1297 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1301 * Don't enter poll loop if we already have events pending.
1302 * If we do, we can potentially be spinning for commands that
1303 * already triggered a CQE (eg in error).
1305 if (io_cqring_events(ctx))
1310 * If a submit got punted to a workqueue, we can have the
1311 * application entering polling for a command before it gets
1312 * issued. That app will hold the uring_lock for the duration
1313 * of the poll right here, so we need to take a breather every
1314 * now and then to ensure that the issue has a chance to add
1315 * the poll to the issued list. Otherwise we can spin here
1316 * forever, while the workqueue is stuck trying to acquire the
1319 if (wq_list_empty(&ctx->iopoll_list)) {
1320 u32 tail = ctx->cached_cq_tail;
1322 mutex_unlock(&ctx->uring_lock);
1324 mutex_lock(&ctx->uring_lock);
1326 /* some requests don't go through iopoll_list */
1327 if (tail != ctx->cached_cq_tail ||
1328 wq_list_empty(&ctx->iopoll_list))
1331 ret = io_do_iopoll(ctx, !min);
1336 } while (nr_events < min && !need_resched());
1341 void io_req_task_complete(struct io_kiocb *req, bool *locked)
1343 if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) {
1344 unsigned issue_flags = *locked ? 0 : IO_URING_F_UNLOCKED;
1346 req->cqe.flags |= io_put_kbuf(req, issue_flags);
1350 io_req_complete_defer(req);
1352 io_req_complete_post(req);
1356 * After the iocb has been issued, it's safe to be found on the poll list.
1357 * Adding the kiocb to the list AFTER submission ensures that we don't
1358 * find it from a io_do_iopoll() thread before the issuer is done
1359 * accessing the kiocb cookie.
1361 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1363 struct io_ring_ctx *ctx = req->ctx;
1364 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1366 /* workqueue context doesn't hold uring_lock, grab it now */
1367 if (unlikely(needs_lock))
1368 mutex_lock(&ctx->uring_lock);
1371 * Track whether we have multiple files in our lists. This will impact
1372 * how we do polling eventually, not spinning if we're on potentially
1373 * different devices.
1375 if (wq_list_empty(&ctx->iopoll_list)) {
1376 ctx->poll_multi_queue = false;
1377 } else if (!ctx->poll_multi_queue) {
1378 struct io_kiocb *list_req;
1380 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1382 if (list_req->file != req->file)
1383 ctx->poll_multi_queue = true;
1387 * For fast devices, IO may have already completed. If it has, add
1388 * it to the front so we find it first.
1390 if (READ_ONCE(req->iopoll_completed))
1391 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1393 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1395 if (unlikely(needs_lock)) {
1397 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1398 * in sq thread task context or in io worker task context. If
1399 * current task context is sq thread, we don't need to check
1400 * whether should wake up sq thread.
1402 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1403 wq_has_sleeper(&ctx->sq_data->wait))
1404 wake_up(&ctx->sq_data->wait);
1406 mutex_unlock(&ctx->uring_lock);
1410 static bool io_bdev_nowait(struct block_device *bdev)
1412 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
1416 * If we tracked the file through the SCM inflight mechanism, we could support
1417 * any file. For now, just ensure that anything potentially problematic is done
1420 static bool __io_file_supports_nowait(struct file *file, umode_t mode)
1422 if (S_ISBLK(mode)) {
1423 if (IS_ENABLED(CONFIG_BLOCK) &&
1424 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
1430 if (S_ISREG(mode)) {
1431 if (IS_ENABLED(CONFIG_BLOCK) &&
1432 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
1433 !io_is_uring_fops(file))
1438 /* any ->read/write should understand O_NONBLOCK */
1439 if (file->f_flags & O_NONBLOCK)
1441 return file->f_mode & FMODE_NOWAIT;
1445 * If we tracked the file through the SCM inflight mechanism, we could support
1446 * any file. For now, just ensure that anything potentially problematic is done
1449 unsigned int io_file_get_flags(struct file *file)
1451 umode_t mode = file_inode(file)->i_mode;
1452 unsigned int res = 0;
1456 if (__io_file_supports_nowait(file, mode))
1458 if (io_file_need_scm(file))
1463 bool io_alloc_async_data(struct io_kiocb *req)
1465 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
1466 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
1467 if (req->async_data) {
1468 req->flags |= REQ_F_ASYNC_DATA;
1474 int io_req_prep_async(struct io_kiocb *req)
1476 const struct io_op_def *def = &io_op_defs[req->opcode];
1478 /* assign early for deferred execution for non-fixed file */
1479 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE))
1480 req->file = io_file_get_normal(req, req->cqe.fd);
1481 if (!def->prep_async)
1483 if (WARN_ON_ONCE(req_has_async_data(req)))
1485 if (io_alloc_async_data(req))
1488 return def->prep_async(req);
1491 static u32 io_get_sequence(struct io_kiocb *req)
1493 u32 seq = req->ctx->cached_sq_head;
1494 struct io_kiocb *cur;
1496 /* need original cached_sq_head, but it was increased for each req */
1497 io_for_each_link(cur, req)
1502 static __cold void io_drain_req(struct io_kiocb *req)
1504 struct io_ring_ctx *ctx = req->ctx;
1505 struct io_defer_entry *de;
1507 u32 seq = io_get_sequence(req);
1509 /* Still need defer if there is pending req in defer list. */
1510 spin_lock(&ctx->completion_lock);
1511 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1512 spin_unlock(&ctx->completion_lock);
1514 ctx->drain_active = false;
1515 io_req_task_queue(req);
1518 spin_unlock(&ctx->completion_lock);
1520 ret = io_req_prep_async(req);
1523 io_req_complete_failed(req, ret);
1526 io_prep_async_link(req);
1527 de = kmalloc(sizeof(*de), GFP_KERNEL);
1533 spin_lock(&ctx->completion_lock);
1534 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1535 spin_unlock(&ctx->completion_lock);
1540 trace_io_uring_defer(req);
1543 list_add_tail(&de->list, &ctx->defer_list);
1544 spin_unlock(&ctx->completion_lock);
1547 static void io_clean_op(struct io_kiocb *req)
1549 if (req->flags & REQ_F_BUFFER_SELECTED) {
1550 spin_lock(&req->ctx->completion_lock);
1551 io_put_kbuf_comp(req);
1552 spin_unlock(&req->ctx->completion_lock);
1555 if (req->flags & REQ_F_NEED_CLEANUP) {
1556 const struct io_op_def *def = &io_op_defs[req->opcode];
1561 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1562 kfree(req->apoll->double_poll);
1566 if (req->flags & REQ_F_INFLIGHT) {
1567 struct io_uring_task *tctx = req->task->io_uring;
1569 atomic_dec(&tctx->inflight_tracked);
1571 if (req->flags & REQ_F_CREDS)
1572 put_cred(req->creds);
1573 if (req->flags & REQ_F_ASYNC_DATA) {
1574 kfree(req->async_data);
1575 req->async_data = NULL;
1577 req->flags &= ~IO_REQ_CLEAN_FLAGS;
1580 static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags)
1582 if (req->file || !io_op_defs[req->opcode].needs_file)
1585 if (req->flags & REQ_F_FIXED_FILE)
1586 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1588 req->file = io_file_get_normal(req, req->cqe.fd);
1593 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1595 const struct io_op_def *def = &io_op_defs[req->opcode];
1596 const struct cred *creds = NULL;
1599 if (unlikely(!io_assign_file(req, issue_flags)))
1602 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1603 creds = override_creds(req->creds);
1605 if (!def->audit_skip)
1606 audit_uring_entry(req->opcode);
1608 ret = def->issue(req, issue_flags);
1610 if (!def->audit_skip)
1611 audit_uring_exit(!ret, ret);
1614 revert_creds(creds);
1616 if (ret == IOU_OK) {
1617 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1618 io_req_complete_defer(req);
1620 io_req_complete_post(req);
1621 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1624 /* If the op doesn't have a file, we're not polling for it */
1625 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && req->file)
1626 io_iopoll_req_issued(req, issue_flags);
1631 int io_poll_issue(struct io_kiocb *req, bool *locked)
1633 io_tw_lock(req->ctx, locked);
1634 if (unlikely(req->task->flags & PF_EXITING))
1636 return io_issue_sqe(req, IO_URING_F_NONBLOCK);
1639 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1641 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1643 req = io_put_req_find_next(req);
1644 return req ? &req->work : NULL;
1647 void io_wq_submit_work(struct io_wq_work *work)
1649 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1650 const struct io_op_def *def = &io_op_defs[req->opcode];
1651 unsigned int issue_flags = IO_URING_F_UNLOCKED;
1652 bool needs_poll = false;
1653 int ret = 0, err = -ECANCELED;
1655 /* one will be dropped by ->io_free_work() after returning to io-wq */
1656 if (!(req->flags & REQ_F_REFCOUNT))
1657 __io_req_set_refcount(req, 2);
1661 io_arm_ltimeout(req);
1663 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1664 if (work->flags & IO_WQ_WORK_CANCEL) {
1666 io_req_task_queue_fail(req, err);
1669 if (!io_assign_file(req, issue_flags)) {
1671 work->flags |= IO_WQ_WORK_CANCEL;
1675 if (req->flags & REQ_F_FORCE_ASYNC) {
1676 bool opcode_poll = def->pollin || def->pollout;
1678 if (opcode_poll && file_can_poll(req->file)) {
1680 issue_flags |= IO_URING_F_NONBLOCK;
1685 ret = io_issue_sqe(req, issue_flags);
1689 * We can get EAGAIN for iopolled IO even though we're
1690 * forcing a sync submission from here, since we can't
1691 * wait for request slots on the block side.
1694 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1700 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1702 /* aborted or ready, in either case retry blocking */
1704 issue_flags &= ~IO_URING_F_NONBLOCK;
1707 /* avoid locking problems by failing it from a clean context */
1709 io_req_task_queue_fail(req, ret);
1712 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1713 unsigned int issue_flags)
1715 struct io_ring_ctx *ctx = req->ctx;
1716 struct file *file = NULL;
1717 unsigned long file_ptr;
1719 io_ring_submit_lock(ctx, issue_flags);
1721 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
1723 fd = array_index_nospec(fd, ctx->nr_user_files);
1724 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
1725 file = (struct file *) (file_ptr & FFS_MASK);
1726 file_ptr &= ~FFS_MASK;
1727 /* mask in overlapping REQ_F and FFS bits */
1728 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
1729 io_req_set_rsrc_node(req, ctx, 0);
1730 WARN_ON_ONCE(file && !test_bit(fd, ctx->file_table.bitmap));
1732 io_ring_submit_unlock(ctx, issue_flags);
1736 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
1738 struct file *file = fget(fd);
1740 trace_io_uring_file_get(req, fd);
1742 /* we don't allow fixed io_uring files */
1743 if (file && io_is_uring_fops(file))
1744 io_req_track_inflight(req);
1748 static void io_queue_async(struct io_kiocb *req, int ret)
1749 __must_hold(&req->ctx->uring_lock)
1751 struct io_kiocb *linked_timeout;
1753 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
1754 io_req_complete_failed(req, ret);
1758 linked_timeout = io_prep_linked_timeout(req);
1760 switch (io_arm_poll_handler(req, 0)) {
1761 case IO_APOLL_READY:
1762 io_req_task_queue(req);
1764 case IO_APOLL_ABORTED:
1766 * Queued up for async execution, worker will release
1767 * submit reference when the iocb is actually submitted.
1769 io_kbuf_recycle(req, 0);
1770 io_queue_iowq(req, NULL);
1777 io_queue_linked_timeout(linked_timeout);
1780 static inline void io_queue_sqe(struct io_kiocb *req)
1781 __must_hold(&req->ctx->uring_lock)
1785 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
1788 * We async punt it if the file wasn't marked NOWAIT, or if the file
1789 * doesn't support non-blocking read/write attempts
1792 io_arm_ltimeout(req);
1794 io_queue_async(req, ret);
1797 static void io_queue_sqe_fallback(struct io_kiocb *req)
1798 __must_hold(&req->ctx->uring_lock)
1800 if (unlikely(req->flags & REQ_F_FAIL)) {
1802 * We don't submit, fail them all, for that replace hardlinks
1803 * with normal links. Extra REQ_F_LINK is tolerated.
1805 req->flags &= ~REQ_F_HARDLINK;
1806 req->flags |= REQ_F_LINK;
1807 io_req_complete_failed(req, req->cqe.res);
1808 } else if (unlikely(req->ctx->drain_active)) {
1811 int ret = io_req_prep_async(req);
1814 io_req_complete_failed(req, ret);
1816 io_queue_iowq(req, NULL);
1821 * Check SQE restrictions (opcode and flags).
1823 * Returns 'true' if SQE is allowed, 'false' otherwise.
1825 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
1826 struct io_kiocb *req,
1827 unsigned int sqe_flags)
1829 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
1832 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
1833 ctx->restrictions.sqe_flags_required)
1836 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
1837 ctx->restrictions.sqe_flags_required))
1843 static void io_init_req_drain(struct io_kiocb *req)
1845 struct io_ring_ctx *ctx = req->ctx;
1846 struct io_kiocb *head = ctx->submit_state.link.head;
1848 ctx->drain_active = true;
1851 * If we need to drain a request in the middle of a link, drain
1852 * the head request and the next request/link after the current
1853 * link. Considering sequential execution of links,
1854 * REQ_F_IO_DRAIN will be maintained for every request of our
1857 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
1858 ctx->drain_next = true;
1862 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
1863 const struct io_uring_sqe *sqe)
1864 __must_hold(&ctx->uring_lock)
1866 const struct io_op_def *def;
1867 unsigned int sqe_flags;
1871 /* req is partially pre-initialised, see io_preinit_req() */
1872 req->opcode = opcode = READ_ONCE(sqe->opcode);
1873 /* same numerical values with corresponding REQ_F_*, safe to copy */
1874 req->flags = sqe_flags = READ_ONCE(sqe->flags);
1875 req->cqe.user_data = READ_ONCE(sqe->user_data);
1877 req->rsrc_node = NULL;
1878 req->task = current;
1880 if (unlikely(opcode >= IORING_OP_LAST)) {
1884 def = &io_op_defs[opcode];
1885 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
1886 /* enforce forwards compatibility on users */
1887 if (sqe_flags & ~SQE_VALID_FLAGS)
1889 if (sqe_flags & IOSQE_BUFFER_SELECT) {
1890 if (!def->buffer_select)
1892 req->buf_index = READ_ONCE(sqe->buf_group);
1894 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
1895 ctx->drain_disabled = true;
1896 if (sqe_flags & IOSQE_IO_DRAIN) {
1897 if (ctx->drain_disabled)
1899 io_init_req_drain(req);
1902 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
1903 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
1905 /* knock it to the slow queue path, will be drained there */
1906 if (ctx->drain_active)
1907 req->flags |= REQ_F_FORCE_ASYNC;
1908 /* if there is no link, we're at "next" request and need to drain */
1909 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
1910 ctx->drain_next = false;
1911 ctx->drain_active = true;
1912 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
1916 if (!def->ioprio && sqe->ioprio)
1918 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
1921 if (def->needs_file) {
1922 struct io_submit_state *state = &ctx->submit_state;
1924 req->cqe.fd = READ_ONCE(sqe->fd);
1927 * Plug now if we have more than 2 IO left after this, and the
1928 * target is potentially a read/write to block based storage.
1930 if (state->need_plug && def->plug) {
1931 state->plug_started = true;
1932 state->need_plug = false;
1933 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
1937 personality = READ_ONCE(sqe->personality);
1941 req->creds = xa_load(&ctx->personalities, personality);
1944 get_cred(req->creds);
1945 ret = security_uring_override_creds(req->creds);
1947 put_cred(req->creds);
1950 req->flags |= REQ_F_CREDS;
1953 return def->prep(req, sqe);
1956 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
1957 struct io_kiocb *req, int ret)
1959 struct io_ring_ctx *ctx = req->ctx;
1960 struct io_submit_link *link = &ctx->submit_state.link;
1961 struct io_kiocb *head = link->head;
1963 trace_io_uring_req_failed(sqe, req, ret);
1966 * Avoid breaking links in the middle as it renders links with SQPOLL
1967 * unusable. Instead of failing eagerly, continue assembling the link if
1968 * applicable and mark the head with REQ_F_FAIL. The link flushing code
1969 * should find the flag and handle the rest.
1971 req_fail_link_node(req, ret);
1972 if (head && !(head->flags & REQ_F_FAIL))
1973 req_fail_link_node(head, -ECANCELED);
1975 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
1977 link->last->link = req;
1981 io_queue_sqe_fallback(req);
1986 link->last->link = req;
1993 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
1994 const struct io_uring_sqe *sqe)
1995 __must_hold(&ctx->uring_lock)
1997 struct io_submit_link *link = &ctx->submit_state.link;
2000 ret = io_init_req(ctx, req, sqe);
2002 return io_submit_fail_init(sqe, req, ret);
2004 /* don't need @sqe from now on */
2005 trace_io_uring_submit_sqe(req, true);
2008 * If we already have a head request, queue this one for async
2009 * submittal once the head completes. If we don't have a head but
2010 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2011 * submitted sync once the chain is complete. If none of those
2012 * conditions are true (normal request), then just queue it.
2014 if (unlikely(link->head)) {
2015 ret = io_req_prep_async(req);
2017 return io_submit_fail_init(sqe, req, ret);
2019 trace_io_uring_link(req, link->head);
2020 link->last->link = req;
2023 if (req->flags & IO_REQ_LINK_FLAGS)
2025 /* last request of the link, flush it */
2028 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2031 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2032 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2033 if (req->flags & IO_REQ_LINK_FLAGS) {
2038 io_queue_sqe_fallback(req);
2048 * Batched submission is done, ensure local IO is flushed out.
2050 static void io_submit_state_end(struct io_ring_ctx *ctx)
2052 struct io_submit_state *state = &ctx->submit_state;
2054 if (unlikely(state->link.head))
2055 io_queue_sqe_fallback(state->link.head);
2056 /* flush only after queuing links as they can generate completions */
2057 io_submit_flush_completions(ctx);
2058 if (state->plug_started)
2059 blk_finish_plug(&state->plug);
2063 * Start submission side cache.
2065 static void io_submit_state_start(struct io_submit_state *state,
2066 unsigned int max_ios)
2068 state->plug_started = false;
2069 state->need_plug = max_ios > 2;
2070 state->submit_nr = max_ios;
2071 /* set only head, no need to init link_last in advance */
2072 state->link.head = NULL;
2075 static void io_commit_sqring(struct io_ring_ctx *ctx)
2077 struct io_rings *rings = ctx->rings;
2080 * Ensure any loads from the SQEs are done at this point,
2081 * since once we write the new head, the application could
2082 * write new data to them.
2084 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2088 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2089 * that is mapped by userspace. This means that care needs to be taken to
2090 * ensure that reads are stable, as we cannot rely on userspace always
2091 * being a good citizen. If members of the sqe are validated and then later
2092 * used, it's important that those reads are done through READ_ONCE() to
2093 * prevent a re-load down the line.
2095 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
2097 unsigned head, mask = ctx->sq_entries - 1;
2098 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2101 * The cached sq head (or cq tail) serves two purposes:
2103 * 1) allows us to batch the cost of updating the user visible
2105 * 2) allows the kernel side to track the head on its own, even
2106 * though the application is the one updating it.
2108 head = READ_ONCE(ctx->sq_array[sq_idx]);
2109 if (likely(head < ctx->sq_entries)) {
2110 /* double index for 128-byte SQEs, twice as long */
2111 if (ctx->flags & IORING_SETUP_SQE128)
2113 return &ctx->sq_sqes[head];
2116 /* drop invalid entries */
2118 WRITE_ONCE(ctx->rings->sq_dropped,
2119 READ_ONCE(ctx->rings->sq_dropped) + 1);
2123 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2124 __must_hold(&ctx->uring_lock)
2126 unsigned int entries = io_sqring_entries(ctx);
2130 if (unlikely(!entries))
2132 /* make sure SQ entry isn't read before tail */
2133 ret = left = min3(nr, ctx->sq_entries, entries);
2134 io_get_task_refs(left);
2135 io_submit_state_start(&ctx->submit_state, left);
2138 const struct io_uring_sqe *sqe;
2139 struct io_kiocb *req;
2141 if (unlikely(!io_alloc_req_refill(ctx)))
2143 req = io_alloc_req(ctx);
2144 sqe = io_get_sqe(ctx);
2145 if (unlikely(!sqe)) {
2146 io_req_add_to_cache(req, ctx);
2151 * Continue submitting even for sqe failure if the
2152 * ring was setup with IORING_SETUP_SUBMIT_ALL
2154 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2155 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2161 if (unlikely(left)) {
2163 /* try again if it submitted nothing and can't allocate a req */
2164 if (!ret && io_req_cache_empty(ctx))
2166 current->io_uring->cached_refs += left;
2169 io_submit_state_end(ctx);
2170 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2171 io_commit_sqring(ctx);
2175 struct io_wait_queue {
2176 struct wait_queue_entry wq;
2177 struct io_ring_ctx *ctx;
2179 unsigned nr_timeouts;
2182 static inline bool io_should_wake(struct io_wait_queue *iowq)
2184 struct io_ring_ctx *ctx = iowq->ctx;
2185 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
2188 * Wake up if we have enough events, or if a timeout occurred since we
2189 * started waiting. For timeouts, we always want to return to userspace,
2190 * regardless of event count.
2192 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2195 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2196 int wake_flags, void *key)
2198 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
2202 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2203 * the task, and the next invocation will do it.
2205 if (io_should_wake(iowq) ||
2206 test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &iowq->ctx->check_cq))
2207 return autoremove_wake_function(curr, mode, wake_flags, key);
2211 int io_run_task_work_sig(void)
2213 if (io_run_task_work())
2215 if (task_sigpending(current))
2220 /* when returns >0, the caller should retry */
2221 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2222 struct io_wait_queue *iowq,
2226 unsigned long check_cq;
2228 /* make sure we run task_work before checking for signals */
2229 ret = io_run_task_work_sig();
2230 if (ret || io_should_wake(iowq))
2233 check_cq = READ_ONCE(ctx->check_cq);
2234 if (unlikely(check_cq)) {
2235 /* let the caller flush overflows, retry */
2236 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2238 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
2241 if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS))
2247 * Wait until events become available, if we don't already have some. The
2248 * application must reap them itself, as they reside on the shared cq ring.
2250 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2251 const sigset_t __user *sig, size_t sigsz,
2252 struct __kernel_timespec __user *uts)
2254 struct io_wait_queue iowq;
2255 struct io_rings *rings = ctx->rings;
2256 ktime_t timeout = KTIME_MAX;
2260 io_cqring_overflow_flush(ctx);
2261 if (io_cqring_events(ctx) >= min_events)
2263 if (!io_run_task_work())
2268 #ifdef CONFIG_COMPAT
2269 if (in_compat_syscall())
2270 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2274 ret = set_user_sigmask(sig, sigsz);
2281 struct timespec64 ts;
2283 if (get_timespec64(&ts, uts))
2285 timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2288 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2289 iowq.wq.private = current;
2290 INIT_LIST_HEAD(&iowq.wq.entry);
2292 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2293 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2295 trace_io_uring_cqring_wait(ctx, min_events);
2297 /* if we can't even flush overflow, don't wait for more */
2298 if (!io_cqring_overflow_flush(ctx)) {
2302 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2303 TASK_INTERRUPTIBLE);
2304 ret = io_cqring_wait_schedule(ctx, &iowq, timeout);
2308 finish_wait(&ctx->cq_wait, &iowq.wq);
2309 restore_saved_sigmask_unless(ret == -EINTR);
2311 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2314 static void io_mem_free(void *ptr)
2321 page = virt_to_head_page(ptr);
2322 if (put_page_testzero(page))
2323 free_compound_page(page);
2326 static void *io_mem_alloc(size_t size)
2328 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2330 return (void *) __get_free_pages(gfp, get_order(size));
2333 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2334 unsigned int cq_entries, size_t *sq_offset)
2336 struct io_rings *rings;
2337 size_t off, sq_array_size;
2339 off = struct_size(rings, cqes, cq_entries);
2340 if (off == SIZE_MAX)
2342 if (ctx->flags & IORING_SETUP_CQE32) {
2343 if (check_shl_overflow(off, 1, &off))
2348 off = ALIGN(off, SMP_CACHE_BYTES);
2356 sq_array_size = array_size(sizeof(u32), sq_entries);
2357 if (sq_array_size == SIZE_MAX)
2360 if (check_add_overflow(off, sq_array_size, &off))
2366 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2367 unsigned int eventfd_async)
2369 struct io_ev_fd *ev_fd;
2370 __s32 __user *fds = arg;
2373 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2374 lockdep_is_held(&ctx->uring_lock));
2378 if (copy_from_user(&fd, fds, sizeof(*fds)))
2381 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2385 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2386 if (IS_ERR(ev_fd->cq_ev_fd)) {
2387 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2392 spin_lock(&ctx->completion_lock);
2393 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2394 spin_unlock(&ctx->completion_lock);
2396 ev_fd->eventfd_async = eventfd_async;
2397 ctx->has_evfd = true;
2398 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2402 static void io_eventfd_put(struct rcu_head *rcu)
2404 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
2406 eventfd_ctx_put(ev_fd->cq_ev_fd);
2410 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2412 struct io_ev_fd *ev_fd;
2414 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2415 lockdep_is_held(&ctx->uring_lock));
2417 ctx->has_evfd = false;
2418 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2419 call_rcu(&ev_fd->rcu, io_eventfd_put);
2426 static void io_req_caches_free(struct io_ring_ctx *ctx)
2428 struct io_submit_state *state = &ctx->submit_state;
2431 mutex_lock(&ctx->uring_lock);
2432 io_flush_cached_locked_reqs(ctx, state);
2434 while (!io_req_cache_empty(ctx)) {
2435 struct io_wq_work_node *node;
2436 struct io_kiocb *req;
2438 node = wq_stack_extract(&state->free_list);
2439 req = container_of(node, struct io_kiocb, comp_list);
2440 kmem_cache_free(req_cachep, req);
2444 percpu_ref_put_many(&ctx->refs, nr);
2445 mutex_unlock(&ctx->uring_lock);
2448 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2450 io_sq_thread_finish(ctx);
2452 if (ctx->mm_account) {
2453 mmdrop(ctx->mm_account);
2454 ctx->mm_account = NULL;
2457 io_rsrc_refs_drop(ctx);
2458 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2459 io_wait_rsrc_data(ctx->buf_data);
2460 io_wait_rsrc_data(ctx->file_data);
2462 mutex_lock(&ctx->uring_lock);
2464 __io_sqe_buffers_unregister(ctx);
2466 __io_sqe_files_unregister(ctx);
2468 __io_cqring_overflow_flush(ctx, true);
2469 io_eventfd_unregister(ctx);
2470 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2471 mutex_unlock(&ctx->uring_lock);
2472 io_destroy_buffers(ctx);
2474 put_cred(ctx->sq_creds);
2475 if (ctx->submitter_task)
2476 put_task_struct(ctx->submitter_task);
2478 /* there are no registered resources left, nobody uses it */
2480 io_rsrc_node_destroy(ctx->rsrc_node);
2481 if (ctx->rsrc_backup_node)
2482 io_rsrc_node_destroy(ctx->rsrc_backup_node);
2483 flush_delayed_work(&ctx->rsrc_put_work);
2484 flush_delayed_work(&ctx->fallback_work);
2486 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2487 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
2489 #if defined(CONFIG_UNIX)
2490 if (ctx->ring_sock) {
2491 ctx->ring_sock->file = NULL; /* so that iput() is called */
2492 sock_release(ctx->ring_sock);
2495 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2497 io_mem_free(ctx->rings);
2498 io_mem_free(ctx->sq_sqes);
2500 percpu_ref_exit(&ctx->refs);
2501 free_uid(ctx->user);
2502 io_req_caches_free(ctx);
2504 io_wq_put_hash(ctx->hash_map);
2505 kfree(ctx->cancel_table.hbs);
2506 kfree(ctx->cancel_table_locked.hbs);
2507 kfree(ctx->dummy_ubuf);
2509 xa_destroy(&ctx->io_bl_xa);
2513 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2515 struct io_ring_ctx *ctx = file->private_data;
2518 poll_wait(file, &ctx->cq_wait, wait);
2520 * synchronizes with barrier from wq_has_sleeper call in
2524 if (!io_sqring_full(ctx))
2525 mask |= EPOLLOUT | EPOLLWRNORM;
2528 * Don't flush cqring overflow list here, just do a simple check.
2529 * Otherwise there could possible be ABBA deadlock:
2532 * lock(&ctx->uring_lock);
2534 * lock(&ctx->uring_lock);
2537 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2538 * pushs them to do the flush.
2540 if (io_cqring_events(ctx) ||
2541 test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
2542 mask |= EPOLLIN | EPOLLRDNORM;
2547 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2549 const struct cred *creds;
2551 creds = xa_erase(&ctx->personalities, id);
2560 struct io_tctx_exit {
2561 struct callback_head task_work;
2562 struct completion completion;
2563 struct io_ring_ctx *ctx;
2566 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2568 struct io_uring_task *tctx = current->io_uring;
2569 struct io_tctx_exit *work;
2571 work = container_of(cb, struct io_tctx_exit, task_work);
2573 * When @in_idle, we're in cancellation and it's racy to remove the
2574 * node. It'll be removed by the end of cancellation, just ignore it.
2576 if (!atomic_read(&tctx->in_idle))
2577 io_uring_del_tctx_node((unsigned long)work->ctx);
2578 complete(&work->completion);
2581 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2583 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2585 return req->ctx == data;
2588 static __cold void io_ring_exit_work(struct work_struct *work)
2590 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2591 unsigned long timeout = jiffies + HZ * 60 * 5;
2592 unsigned long interval = HZ / 20;
2593 struct io_tctx_exit exit;
2594 struct io_tctx_node *node;
2598 * If we're doing polled IO and end up having requests being
2599 * submitted async (out-of-line), then completions can come in while
2600 * we're waiting for refs to drop. We need to reap these manually,
2601 * as nobody else will be looking for them.
2604 while (io_uring_try_cancel_requests(ctx, NULL, true))
2608 struct io_sq_data *sqd = ctx->sq_data;
2609 struct task_struct *tsk;
2611 io_sq_thread_park(sqd);
2613 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
2614 io_wq_cancel_cb(tsk->io_uring->io_wq,
2615 io_cancel_ctx_cb, ctx, true);
2616 io_sq_thread_unpark(sqd);
2619 io_req_caches_free(ctx);
2621 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
2622 /* there is little hope left, don't run it too often */
2625 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
2627 init_completion(&exit.completion);
2628 init_task_work(&exit.task_work, io_tctx_exit_cb);
2631 * Some may use context even when all refs and requests have been put,
2632 * and they are free to do so while still holding uring_lock or
2633 * completion_lock, see io_req_task_submit(). Apart from other work,
2634 * this lock/unlock section also waits them to finish.
2636 mutex_lock(&ctx->uring_lock);
2637 while (!list_empty(&ctx->tctx_list)) {
2638 WARN_ON_ONCE(time_after(jiffies, timeout));
2640 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
2642 /* don't spin on a single task if cancellation failed */
2643 list_rotate_left(&ctx->tctx_list);
2644 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
2645 if (WARN_ON_ONCE(ret))
2648 mutex_unlock(&ctx->uring_lock);
2649 wait_for_completion(&exit.completion);
2650 mutex_lock(&ctx->uring_lock);
2652 mutex_unlock(&ctx->uring_lock);
2653 spin_lock(&ctx->completion_lock);
2654 spin_unlock(&ctx->completion_lock);
2656 io_ring_ctx_free(ctx);
2659 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
2661 unsigned long index;
2662 struct creds *creds;
2664 mutex_lock(&ctx->uring_lock);
2665 percpu_ref_kill(&ctx->refs);
2667 __io_cqring_overflow_flush(ctx, true);
2668 xa_for_each(&ctx->personalities, index, creds)
2669 io_unregister_personality(ctx, index);
2671 io_poll_remove_all(ctx, NULL, true);
2672 mutex_unlock(&ctx->uring_lock);
2674 /* failed during ring init, it couldn't have issued any requests */
2676 io_kill_timeouts(ctx, NULL, true);
2677 /* if we failed setting up the ctx, we might not have any rings */
2678 io_iopoll_try_reap_events(ctx);
2681 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
2683 * Use system_unbound_wq to avoid spawning tons of event kworkers
2684 * if we're exiting a ton of rings at the same time. It just adds
2685 * noise and overhead, there's no discernable change in runtime
2686 * over using system_wq.
2688 queue_work(system_unbound_wq, &ctx->exit_work);
2691 static int io_uring_release(struct inode *inode, struct file *file)
2693 struct io_ring_ctx *ctx = file->private_data;
2695 file->private_data = NULL;
2696 io_ring_ctx_wait_and_kill(ctx);
2700 struct io_task_cancel {
2701 struct task_struct *task;
2705 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
2707 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2708 struct io_task_cancel *cancel = data;
2710 return io_match_task_safe(req, cancel->task, cancel->all);
2713 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
2714 struct task_struct *task,
2717 struct io_defer_entry *de;
2720 spin_lock(&ctx->completion_lock);
2721 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
2722 if (io_match_task_safe(de->req, task, cancel_all)) {
2723 list_cut_position(&list, &ctx->defer_list, &de->list);
2727 spin_unlock(&ctx->completion_lock);
2728 if (list_empty(&list))
2731 while (!list_empty(&list)) {
2732 de = list_first_entry(&list, struct io_defer_entry, list);
2733 list_del_init(&de->list);
2734 io_req_complete_failed(de->req, -ECANCELED);
2740 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
2742 struct io_tctx_node *node;
2743 enum io_wq_cancel cret;
2746 mutex_lock(&ctx->uring_lock);
2747 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
2748 struct io_uring_task *tctx = node->task->io_uring;
2751 * io_wq will stay alive while we hold uring_lock, because it's
2752 * killed after ctx nodes, which requires to take the lock.
2754 if (!tctx || !tctx->io_wq)
2756 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
2757 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2759 mutex_unlock(&ctx->uring_lock);
2764 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
2765 struct task_struct *task,
2768 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
2769 struct io_uring_task *tctx = task ? task->io_uring : NULL;
2770 enum io_wq_cancel cret;
2773 /* failed during ring init, it couldn't have issued any requests */
2778 ret |= io_uring_try_cancel_iowq(ctx);
2779 } else if (tctx && tctx->io_wq) {
2781 * Cancels requests of all rings, not only @ctx, but
2782 * it's fine as the task is in exit/exec.
2784 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
2786 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2789 /* SQPOLL thread does its own polling */
2790 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
2791 (ctx->sq_data && ctx->sq_data->thread == current)) {
2792 while (!wq_list_empty(&ctx->iopoll_list)) {
2793 io_iopoll_try_reap_events(ctx);
2798 ret |= io_cancel_defer_files(ctx, task, cancel_all);
2799 mutex_lock(&ctx->uring_lock);
2800 ret |= io_poll_remove_all(ctx, task, cancel_all);
2801 mutex_unlock(&ctx->uring_lock);
2802 ret |= io_kill_timeouts(ctx, task, cancel_all);
2804 ret |= io_run_task_work();
2808 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
2811 return atomic_read(&tctx->inflight_tracked);
2812 return percpu_counter_sum(&tctx->inflight);
2816 * Find any io_uring ctx that this task has registered or done IO on, and cancel
2817 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
2819 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
2821 struct io_uring_task *tctx = current->io_uring;
2822 struct io_ring_ctx *ctx;
2826 WARN_ON_ONCE(sqd && sqd->thread != current);
2828 if (!current->io_uring)
2831 io_wq_exit_start(tctx->io_wq);
2833 atomic_inc(&tctx->in_idle);
2837 io_uring_drop_tctx_refs(current);
2838 /* read completions before cancelations */
2839 inflight = tctx_inflight(tctx, !cancel_all);
2844 struct io_tctx_node *node;
2845 unsigned long index;
2847 xa_for_each(&tctx->xa, index, node) {
2848 /* sqpoll task will cancel all its requests */
2849 if (node->ctx->sq_data)
2851 loop |= io_uring_try_cancel_requests(node->ctx,
2852 current, cancel_all);
2855 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
2856 loop |= io_uring_try_cancel_requests(ctx,
2866 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
2868 io_uring_drop_tctx_refs(current);
2871 * If we've seen completions, retry without waiting. This
2872 * avoids a race where a completion comes in before we did
2873 * prepare_to_wait().
2875 if (inflight == tctx_inflight(tctx, !cancel_all))
2877 finish_wait(&tctx->wait, &wait);
2880 io_uring_clean_tctx(tctx);
2883 * We shouldn't run task_works after cancel, so just leave
2884 * ->in_idle set for normal exit.
2886 atomic_dec(&tctx->in_idle);
2887 /* for exec all current's requests should be gone, kill tctx */
2888 __io_uring_free(current);
2892 void __io_uring_cancel(bool cancel_all)
2894 io_uring_cancel_generic(cancel_all, NULL);
2897 static void *io_uring_validate_mmap_request(struct file *file,
2898 loff_t pgoff, size_t sz)
2900 struct io_ring_ctx *ctx = file->private_data;
2901 loff_t offset = pgoff << PAGE_SHIFT;
2906 case IORING_OFF_SQ_RING:
2907 case IORING_OFF_CQ_RING:
2910 case IORING_OFF_SQES:
2914 return ERR_PTR(-EINVAL);
2917 page = virt_to_head_page(ptr);
2918 if (sz > page_size(page))
2919 return ERR_PTR(-EINVAL);
2926 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
2928 size_t sz = vma->vm_end - vma->vm_start;
2932 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
2934 return PTR_ERR(ptr);
2936 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
2937 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
2940 #else /* !CONFIG_MMU */
2942 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
2944 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
2947 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
2949 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
2952 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
2953 unsigned long addr, unsigned long len,
2954 unsigned long pgoff, unsigned long flags)
2958 ptr = io_uring_validate_mmap_request(file, pgoff, len);
2960 return PTR_ERR(ptr);
2962 return (unsigned long) ptr;
2965 #endif /* !CONFIG_MMU */
2967 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
2969 if (flags & IORING_ENTER_EXT_ARG) {
2970 struct io_uring_getevents_arg arg;
2972 if (argsz != sizeof(arg))
2974 if (copy_from_user(&arg, argp, sizeof(arg)))
2980 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
2981 struct __kernel_timespec __user **ts,
2982 const sigset_t __user **sig)
2984 struct io_uring_getevents_arg arg;
2987 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
2988 * is just a pointer to the sigset_t.
2990 if (!(flags & IORING_ENTER_EXT_ARG)) {
2991 *sig = (const sigset_t __user *) argp;
2997 * EXT_ARG is set - ensure we agree on the size of it and copy in our
2998 * timespec and sigset_t pointers if good.
3000 if (*argsz != sizeof(arg))
3002 if (copy_from_user(&arg, argp, sizeof(arg)))
3006 *sig = u64_to_user_ptr(arg.sigmask);
3007 *argsz = arg.sigmask_sz;
3008 *ts = u64_to_user_ptr(arg.ts);
3012 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3013 u32, min_complete, u32, flags, const void __user *, argp,
3016 struct io_ring_ctx *ctx;
3022 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3023 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3024 IORING_ENTER_REGISTERED_RING)))
3028 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3029 * need only dereference our task private array to find it.
3031 if (flags & IORING_ENTER_REGISTERED_RING) {
3032 struct io_uring_task *tctx = current->io_uring;
3034 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3036 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3037 f.file = tctx->registered_rings[fd];
3039 if (unlikely(!f.file))
3043 if (unlikely(!f.file))
3046 if (unlikely(!io_is_uring_fops(f.file)))
3050 ctx = f.file->private_data;
3052 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3056 * For SQ polling, the thread will do all submissions and completions.
3057 * Just return the requested submit count, and wake the thread if
3061 if (ctx->flags & IORING_SETUP_SQPOLL) {
3062 io_cqring_overflow_flush(ctx);
3064 if (unlikely(ctx->sq_data->thread == NULL)) {
3068 if (flags & IORING_ENTER_SQ_WAKEUP)
3069 wake_up(&ctx->sq_data->wait);
3070 if (flags & IORING_ENTER_SQ_WAIT) {
3071 ret = io_sqpoll_wait_sq(ctx);
3076 } else if (to_submit) {
3077 ret = io_uring_add_tctx_node(ctx);
3081 mutex_lock(&ctx->uring_lock);
3082 ret = io_submit_sqes(ctx, to_submit);
3083 if (ret != to_submit) {
3084 mutex_unlock(&ctx->uring_lock);
3087 if ((flags & IORING_ENTER_GETEVENTS) && ctx->syscall_iopoll)
3089 mutex_unlock(&ctx->uring_lock);
3091 if (flags & IORING_ENTER_GETEVENTS) {
3093 if (ctx->syscall_iopoll) {
3095 * We disallow the app entering submit/complete with
3096 * polling, but we still need to lock the ring to
3097 * prevent racing with polled issue that got punted to
3100 mutex_lock(&ctx->uring_lock);
3102 ret2 = io_validate_ext_arg(flags, argp, argsz);
3103 if (likely(!ret2)) {
3104 min_complete = min(min_complete,
3106 ret2 = io_iopoll_check(ctx, min_complete);
3108 mutex_unlock(&ctx->uring_lock);
3110 const sigset_t __user *sig;
3111 struct __kernel_timespec __user *ts;
3113 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3114 if (likely(!ret2)) {
3115 min_complete = min(min_complete,
3117 ret2 = io_cqring_wait(ctx, min_complete, sig,
3126 * EBADR indicates that one or more CQE were dropped.
3127 * Once the user has been informed we can clear the bit
3128 * as they are obviously ok with those drops.
3130 if (unlikely(ret2 == -EBADR))
3131 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3140 static const struct file_operations io_uring_fops = {
3141 .release = io_uring_release,
3142 .mmap = io_uring_mmap,
3144 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3145 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3147 .poll = io_uring_poll,
3148 #ifdef CONFIG_PROC_FS
3149 .show_fdinfo = io_uring_show_fdinfo,
3153 bool io_is_uring_fops(struct file *file)
3155 return file->f_op == &io_uring_fops;
3158 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3159 struct io_uring_params *p)
3161 struct io_rings *rings;
3162 size_t size, sq_array_offset;
3164 /* make sure these are sane, as we already accounted them */
3165 ctx->sq_entries = p->sq_entries;
3166 ctx->cq_entries = p->cq_entries;
3168 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3169 if (size == SIZE_MAX)
3172 rings = io_mem_alloc(size);
3177 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3178 rings->sq_ring_mask = p->sq_entries - 1;
3179 rings->cq_ring_mask = p->cq_entries - 1;
3180 rings->sq_ring_entries = p->sq_entries;
3181 rings->cq_ring_entries = p->cq_entries;
3183 if (p->flags & IORING_SETUP_SQE128)
3184 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3186 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3187 if (size == SIZE_MAX) {
3188 io_mem_free(ctx->rings);
3193 ctx->sq_sqes = io_mem_alloc(size);
3194 if (!ctx->sq_sqes) {
3195 io_mem_free(ctx->rings);
3203 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3207 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3211 ret = __io_uring_add_tctx_node(ctx, false);
3216 fd_install(fd, file);
3221 * Allocate an anonymous fd, this is what constitutes the application
3222 * visible backing of an io_uring instance. The application mmaps this
3223 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3224 * we have to tie this fd to a socket for file garbage collection purposes.
3226 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3229 #if defined(CONFIG_UNIX)
3232 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3235 return ERR_PTR(ret);
3238 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3239 O_RDWR | O_CLOEXEC, NULL);
3240 #if defined(CONFIG_UNIX)
3242 sock_release(ctx->ring_sock);
3243 ctx->ring_sock = NULL;
3245 ctx->ring_sock->file = file;
3251 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3252 struct io_uring_params __user *params)
3254 struct io_ring_ctx *ctx;
3260 if (entries > IORING_MAX_ENTRIES) {
3261 if (!(p->flags & IORING_SETUP_CLAMP))
3263 entries = IORING_MAX_ENTRIES;
3267 * Use twice as many entries for the CQ ring. It's possible for the
3268 * application to drive a higher depth than the size of the SQ ring,
3269 * since the sqes are only used at submission time. This allows for
3270 * some flexibility in overcommitting a bit. If the application has
3271 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3272 * of CQ ring entries manually.
3274 p->sq_entries = roundup_pow_of_two(entries);
3275 if (p->flags & IORING_SETUP_CQSIZE) {
3277 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3278 * to a power-of-two, if it isn't already. We do NOT impose
3279 * any cq vs sq ring sizing.
3283 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3284 if (!(p->flags & IORING_SETUP_CLAMP))
3286 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3288 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3289 if (p->cq_entries < p->sq_entries)
3292 p->cq_entries = 2 * p->sq_entries;
3295 ctx = io_ring_ctx_alloc(p);
3300 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3301 * space applications don't need to do io completion events
3302 * polling again, they can rely on io_sq_thread to do polling
3303 * work, which can reduce cpu usage and uring_lock contention.
3305 if (ctx->flags & IORING_SETUP_IOPOLL &&
3306 !(ctx->flags & IORING_SETUP_SQPOLL))
3307 ctx->syscall_iopoll = 1;
3309 ctx->compat = in_compat_syscall();
3310 if (!capable(CAP_IPC_LOCK))
3311 ctx->user = get_uid(current_user());
3314 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3315 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3318 if (ctx->flags & IORING_SETUP_SQPOLL) {
3319 /* IPI related flags don't make sense with SQPOLL */
3320 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3321 IORING_SETUP_TASKRUN_FLAG))
3323 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3324 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3325 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3327 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
3329 ctx->notify_method = TWA_SIGNAL;
3333 * This is just grabbed for accounting purposes. When a process exits,
3334 * the mm is exited and dropped before the files, hence we need to hang
3335 * on to this mm purely for the purposes of being able to unaccount
3336 * memory (locked/pinned vm). It's not used for anything else.
3338 mmgrab(current->mm);
3339 ctx->mm_account = current->mm;
3341 ret = io_allocate_scq_urings(ctx, p);
3345 ret = io_sq_offload_create(ctx, p);
3348 /* always set a rsrc node */
3349 ret = io_rsrc_node_switch_start(ctx);
3352 io_rsrc_node_switch(ctx, NULL);
3354 memset(&p->sq_off, 0, sizeof(p->sq_off));
3355 p->sq_off.head = offsetof(struct io_rings, sq.head);
3356 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3357 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3358 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3359 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3360 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3361 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3363 memset(&p->cq_off, 0, sizeof(p->cq_off));
3364 p->cq_off.head = offsetof(struct io_rings, cq.head);
3365 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3366 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3367 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3368 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3369 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3370 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3372 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3373 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3374 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3375 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3376 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3377 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3378 IORING_FEAT_LINKED_FILE;
3380 if (copy_to_user(params, p, sizeof(*p))) {
3385 file = io_uring_get_file(ctx);
3387 ret = PTR_ERR(file);
3392 * Install ring fd as the very last thing, so we don't risk someone
3393 * having closed it before we finish setup
3395 ret = io_uring_install_fd(ctx, file);
3397 /* fput will clean it up */
3402 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3405 io_ring_ctx_wait_and_kill(ctx);
3410 * Sets up an aio uring context, and returns the fd. Applications asks for a
3411 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3412 * params structure passed in.
3414 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3416 struct io_uring_params p;
3419 if (copy_from_user(&p, params, sizeof(p)))
3421 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3426 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3427 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3428 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3429 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3430 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3431 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3432 IORING_SETUP_SINGLE_ISSUER))
3435 return io_uring_create(entries, &p, params);
3438 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3439 struct io_uring_params __user *, params)
3441 return io_uring_setup(entries, params);
3444 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3447 struct io_uring_probe *p;
3451 size = struct_size(p, ops, nr_args);
3452 if (size == SIZE_MAX)
3454 p = kzalloc(size, GFP_KERNEL);
3459 if (copy_from_user(p, arg, size))
3462 if (memchr_inv(p, 0, size))
3465 p->last_op = IORING_OP_LAST - 1;
3466 if (nr_args > IORING_OP_LAST)
3467 nr_args = IORING_OP_LAST;
3469 for (i = 0; i < nr_args; i++) {
3471 if (!io_op_defs[i].not_supported)
3472 p->ops[i].flags = IO_URING_OP_SUPPORTED;
3477 if (copy_to_user(arg, p, size))
3484 static int io_register_personality(struct io_ring_ctx *ctx)
3486 const struct cred *creds;
3490 creds = get_current_cred();
3492 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
3493 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
3501 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
3502 void __user *arg, unsigned int nr_args)
3504 struct io_uring_restriction *res;
3508 /* Restrictions allowed only if rings started disabled */
3509 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3512 /* We allow only a single restrictions registration */
3513 if (ctx->restrictions.registered)
3516 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
3519 size = array_size(nr_args, sizeof(*res));
3520 if (size == SIZE_MAX)
3523 res = memdup_user(arg, size);
3525 return PTR_ERR(res);
3529 for (i = 0; i < nr_args; i++) {
3530 switch (res[i].opcode) {
3531 case IORING_RESTRICTION_REGISTER_OP:
3532 if (res[i].register_op >= IORING_REGISTER_LAST) {
3537 __set_bit(res[i].register_op,
3538 ctx->restrictions.register_op);
3540 case IORING_RESTRICTION_SQE_OP:
3541 if (res[i].sqe_op >= IORING_OP_LAST) {
3546 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
3548 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
3549 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
3551 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
3552 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
3561 /* Reset all restrictions if an error happened */
3563 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
3565 ctx->restrictions.registered = true;
3571 static int io_register_enable_rings(struct io_ring_ctx *ctx)
3573 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3576 if (ctx->restrictions.registered)
3577 ctx->restricted = 1;
3579 ctx->flags &= ~IORING_SETUP_R_DISABLED;
3580 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
3581 wake_up(&ctx->sq_data->wait);
3585 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
3586 void __user *arg, unsigned len)
3588 struct io_uring_task *tctx = current->io_uring;
3589 cpumask_var_t new_mask;
3592 if (!tctx || !tctx->io_wq)
3595 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3598 cpumask_clear(new_mask);
3599 if (len > cpumask_size())
3600 len = cpumask_size();
3602 if (in_compat_syscall()) {
3603 ret = compat_get_bitmap(cpumask_bits(new_mask),
3604 (const compat_ulong_t __user *)arg,
3605 len * 8 /* CHAR_BIT */);
3607 ret = copy_from_user(new_mask, arg, len);
3611 free_cpumask_var(new_mask);
3615 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
3616 free_cpumask_var(new_mask);
3620 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
3622 struct io_uring_task *tctx = current->io_uring;
3624 if (!tctx || !tctx->io_wq)
3627 return io_wq_cpu_affinity(tctx->io_wq, NULL);
3630 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
3632 __must_hold(&ctx->uring_lock)
3634 struct io_tctx_node *node;
3635 struct io_uring_task *tctx = NULL;
3636 struct io_sq_data *sqd = NULL;
3640 if (copy_from_user(new_count, arg, sizeof(new_count)))
3642 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3643 if (new_count[i] > INT_MAX)
3646 if (ctx->flags & IORING_SETUP_SQPOLL) {
3650 * Observe the correct sqd->lock -> ctx->uring_lock
3651 * ordering. Fine to drop uring_lock here, we hold
3654 refcount_inc(&sqd->refs);
3655 mutex_unlock(&ctx->uring_lock);
3656 mutex_lock(&sqd->lock);
3657 mutex_lock(&ctx->uring_lock);
3659 tctx = sqd->thread->io_uring;
3662 tctx = current->io_uring;
3665 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
3667 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3669 ctx->iowq_limits[i] = new_count[i];
3670 ctx->iowq_limits_set = true;
3672 if (tctx && tctx->io_wq) {
3673 ret = io_wq_max_workers(tctx->io_wq, new_count);
3677 memset(new_count, 0, sizeof(new_count));
3681 mutex_unlock(&sqd->lock);
3682 io_put_sq_data(sqd);
3685 if (copy_to_user(arg, new_count, sizeof(new_count)))
3688 /* that's it for SQPOLL, only the SQPOLL task creates requests */
3692 /* now propagate the restriction to all registered users */
3693 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3694 struct io_uring_task *tctx = node->task->io_uring;
3696 if (WARN_ON_ONCE(!tctx->io_wq))
3699 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3700 new_count[i] = ctx->iowq_limits[i];
3701 /* ignore errors, it always returns zero anyway */
3702 (void)io_wq_max_workers(tctx->io_wq, new_count);
3707 mutex_unlock(&sqd->lock);
3708 io_put_sq_data(sqd);
3713 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
3714 void __user *arg, unsigned nr_args)
3715 __releases(ctx->uring_lock)
3716 __acquires(ctx->uring_lock)
3721 * We don't quiesce the refs for register anymore and so it can't be
3722 * dying as we're holding a file ref here.
3724 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
3727 if (ctx->restricted) {
3728 if (opcode >= IORING_REGISTER_LAST)
3730 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
3731 if (!test_bit(opcode, ctx->restrictions.register_op))
3736 case IORING_REGISTER_BUFFERS:
3740 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
3742 case IORING_UNREGISTER_BUFFERS:
3746 ret = io_sqe_buffers_unregister(ctx);
3748 case IORING_REGISTER_FILES:
3752 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
3754 case IORING_UNREGISTER_FILES:
3758 ret = io_sqe_files_unregister(ctx);
3760 case IORING_REGISTER_FILES_UPDATE:
3761 ret = io_register_files_update(ctx, arg, nr_args);
3763 case IORING_REGISTER_EVENTFD:
3767 ret = io_eventfd_register(ctx, arg, 0);
3769 case IORING_REGISTER_EVENTFD_ASYNC:
3773 ret = io_eventfd_register(ctx, arg, 1);
3775 case IORING_UNREGISTER_EVENTFD:
3779 ret = io_eventfd_unregister(ctx);
3781 case IORING_REGISTER_PROBE:
3783 if (!arg || nr_args > 256)
3785 ret = io_probe(ctx, arg, nr_args);
3787 case IORING_REGISTER_PERSONALITY:
3791 ret = io_register_personality(ctx);
3793 case IORING_UNREGISTER_PERSONALITY:
3797 ret = io_unregister_personality(ctx, nr_args);
3799 case IORING_REGISTER_ENABLE_RINGS:
3803 ret = io_register_enable_rings(ctx);
3805 case IORING_REGISTER_RESTRICTIONS:
3806 ret = io_register_restrictions(ctx, arg, nr_args);
3808 case IORING_REGISTER_FILES2:
3809 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
3811 case IORING_REGISTER_FILES_UPDATE2:
3812 ret = io_register_rsrc_update(ctx, arg, nr_args,
3815 case IORING_REGISTER_BUFFERS2:
3816 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
3818 case IORING_REGISTER_BUFFERS_UPDATE:
3819 ret = io_register_rsrc_update(ctx, arg, nr_args,
3820 IORING_RSRC_BUFFER);
3822 case IORING_REGISTER_IOWQ_AFF:
3824 if (!arg || !nr_args)
3826 ret = io_register_iowq_aff(ctx, arg, nr_args);
3828 case IORING_UNREGISTER_IOWQ_AFF:
3832 ret = io_unregister_iowq_aff(ctx);
3834 case IORING_REGISTER_IOWQ_MAX_WORKERS:
3836 if (!arg || nr_args != 2)
3838 ret = io_register_iowq_max_workers(ctx, arg);
3840 case IORING_REGISTER_RING_FDS:
3841 ret = io_ringfd_register(ctx, arg, nr_args);
3843 case IORING_UNREGISTER_RING_FDS:
3844 ret = io_ringfd_unregister(ctx, arg, nr_args);
3846 case IORING_REGISTER_PBUF_RING:
3848 if (!arg || nr_args != 1)
3850 ret = io_register_pbuf_ring(ctx, arg);
3852 case IORING_UNREGISTER_PBUF_RING:
3854 if (!arg || nr_args != 1)
3856 ret = io_unregister_pbuf_ring(ctx, arg);
3858 case IORING_REGISTER_SYNC_CANCEL:
3860 if (!arg || nr_args != 1)
3862 ret = io_sync_cancel(ctx, arg);
3864 case IORING_REGISTER_FILE_ALLOC_RANGE:
3866 if (!arg || nr_args)
3868 ret = io_register_file_alloc_range(ctx, arg);
3878 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
3879 void __user *, arg, unsigned int, nr_args)
3881 struct io_ring_ctx *ctx;
3890 if (!io_is_uring_fops(f.file))
3893 ctx = f.file->private_data;
3897 mutex_lock(&ctx->uring_lock);
3898 ret = __io_uring_register(ctx, opcode, arg, nr_args);
3899 mutex_unlock(&ctx->uring_lock);
3900 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
3906 static int __init io_uring_init(void)
3908 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
3909 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
3910 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
3913 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
3914 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
3915 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
3916 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
3917 BUILD_BUG_SQE_ELEM(1, __u8, flags);
3918 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
3919 BUILD_BUG_SQE_ELEM(4, __s32, fd);
3920 BUILD_BUG_SQE_ELEM(8, __u64, off);
3921 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
3922 BUILD_BUG_SQE_ELEM(16, __u64, addr);
3923 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
3924 BUILD_BUG_SQE_ELEM(24, __u32, len);
3925 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
3926 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
3927 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
3928 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
3929 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
3930 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
3931 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
3932 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
3933 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
3934 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
3935 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
3936 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
3937 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
3938 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
3939 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
3940 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
3941 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
3942 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
3943 BUILD_BUG_SQE_ELEM(42, __u16, personality);
3944 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
3945 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
3946 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
3948 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
3949 sizeof(struct io_uring_rsrc_update));
3950 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
3951 sizeof(struct io_uring_rsrc_update2));
3953 /* ->buf_index is u16 */
3954 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
3955 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
3956 offsetof(struct io_uring_buf_ring, tail));
3958 /* should fit into one byte */
3959 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
3960 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
3961 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
3963 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
3965 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
3967 io_uring_optable_init();
3969 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
3973 __initcall(io_uring_init);