1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <net/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
49 #include <linux/bits.h>
51 #include <linux/sched/signal.h>
53 #include <linux/file.h>
54 #include <linux/fdtable.h>
56 #include <linux/mman.h>
57 #include <linux/percpu.h>
58 #include <linux/slab.h>
59 #include <linux/bvec.h>
60 #include <linux/net.h>
62 #include <net/af_unix.h>
64 #include <linux/anon_inodes.h>
65 #include <linux/sched/mm.h>
66 #include <linux/uaccess.h>
67 #include <linux/nospec.h>
68 #include <linux/highmem.h>
69 #include <linux/fsnotify.h>
70 #include <linux/fadvise.h>
71 #include <linux/task_work.h>
72 #include <linux/io_uring.h>
73 #include <linux/audit.h>
74 #include <linux/security.h>
75 #include <asm/shmparam.h>
77 #define CREATE_TRACE_POINTS
78 #include <trace/events/io_uring.h>
80 #include <uapi/linux/io_uring.h>
98 #include "alloc_cache.h"
100 #define IORING_MAX_ENTRIES 32768
101 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
103 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
104 IORING_REGISTER_LAST + IORING_OP_LAST)
106 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
107 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
109 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
110 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
112 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
113 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
116 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
119 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
121 #define IO_COMPL_BATCH 32
122 #define IO_REQ_ALLOC_BATCH 8
125 IO_CHECK_CQ_OVERFLOW_BIT,
126 IO_CHECK_CQ_DROPPED_BIT,
130 IO_EVENTFD_OP_SIGNAL_BIT,
131 IO_EVENTFD_OP_FREE_BIT,
134 struct io_defer_entry {
135 struct list_head list;
136 struct io_kiocb *req;
140 /* requests with any of those set should undergo io_disarm_next() */
141 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
142 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
144 static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
145 struct task_struct *task,
148 static void io_dismantle_req(struct io_kiocb *req);
149 static void io_clean_op(struct io_kiocb *req);
150 static void io_queue_sqe(struct io_kiocb *req);
151 static void io_move_task_work_from_local(struct io_ring_ctx *ctx);
152 static void __io_submit_flush_completions(struct io_ring_ctx *ctx);
154 static struct kmem_cache *req_cachep;
156 struct sock *io_uring_get_socket(struct file *file)
158 #if defined(CONFIG_UNIX)
159 if (io_is_uring_fops(file)) {
160 struct io_ring_ctx *ctx = file->private_data;
162 return ctx->ring_sock->sk;
167 EXPORT_SYMBOL(io_uring_get_socket);
169 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
171 if (!wq_list_empty(&ctx->submit_state.compl_reqs))
172 __io_submit_flush_completions(ctx);
175 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
177 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
180 static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx)
182 return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head);
185 static bool io_match_linked(struct io_kiocb *head)
187 struct io_kiocb *req;
189 io_for_each_link(req, head) {
190 if (req->flags & REQ_F_INFLIGHT)
197 * As io_match_task() but protected against racing with linked timeouts.
198 * User must not hold timeout_lock.
200 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
205 if (task && head->task != task)
210 if (head->flags & REQ_F_LINK_TIMEOUT) {
211 struct io_ring_ctx *ctx = head->ctx;
213 /* protect against races with linked timeouts */
214 spin_lock_irq(&ctx->timeout_lock);
215 matched = io_match_linked(head);
216 spin_unlock_irq(&ctx->timeout_lock);
218 matched = io_match_linked(head);
223 static inline void req_fail_link_node(struct io_kiocb *req, int res)
226 io_req_set_res(req, res, 0);
229 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
231 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
234 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
236 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
238 complete(&ctx->ref_comp);
241 static __cold void io_fallback_req_func(struct work_struct *work)
243 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
245 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
246 struct io_kiocb *req, *tmp;
249 percpu_ref_get(&ctx->refs);
250 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
251 req->io_task_work.func(req, &locked);
254 io_submit_flush_completions(ctx);
255 mutex_unlock(&ctx->uring_lock);
257 percpu_ref_put(&ctx->refs);
260 static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
262 unsigned hash_buckets = 1U << bits;
263 size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
265 table->hbs = kmalloc(hash_size, GFP_KERNEL);
269 table->hash_bits = bits;
270 init_hash_table(table, hash_buckets);
274 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
276 struct io_ring_ctx *ctx;
279 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
283 xa_init(&ctx->io_bl_xa);
286 * Use 5 bits less than the max cq entries, that should give us around
287 * 32 entries per hash list if totally full and uniformly spread, but
288 * don't keep too many buckets to not overconsume memory.
290 hash_bits = ilog2(p->cq_entries) - 5;
291 hash_bits = clamp(hash_bits, 1, 8);
292 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
294 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
297 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
298 if (!ctx->dummy_ubuf)
300 /* set invalid range, so io_import_fixed() fails meeting it */
301 ctx->dummy_ubuf->ubuf = -1UL;
303 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
307 ctx->flags = p->flags;
308 init_waitqueue_head(&ctx->sqo_sq_wait);
309 INIT_LIST_HEAD(&ctx->sqd_list);
310 INIT_LIST_HEAD(&ctx->cq_overflow_list);
311 INIT_LIST_HEAD(&ctx->io_buffers_cache);
312 io_alloc_cache_init(&ctx->apoll_cache);
313 io_alloc_cache_init(&ctx->netmsg_cache);
314 init_completion(&ctx->ref_comp);
315 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
316 mutex_init(&ctx->uring_lock);
317 init_waitqueue_head(&ctx->cq_wait);
318 spin_lock_init(&ctx->completion_lock);
319 spin_lock_init(&ctx->timeout_lock);
320 INIT_WQ_LIST(&ctx->iopoll_list);
321 INIT_LIST_HEAD(&ctx->io_buffers_pages);
322 INIT_LIST_HEAD(&ctx->io_buffers_comp);
323 INIT_LIST_HEAD(&ctx->defer_list);
324 INIT_LIST_HEAD(&ctx->timeout_list);
325 INIT_LIST_HEAD(&ctx->ltimeout_list);
326 spin_lock_init(&ctx->rsrc_ref_lock);
327 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
328 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
329 init_llist_head(&ctx->rsrc_put_llist);
330 init_llist_head(&ctx->work_llist);
331 INIT_LIST_HEAD(&ctx->tctx_list);
332 ctx->submit_state.free_list.next = NULL;
333 INIT_WQ_LIST(&ctx->locked_free_list);
334 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
335 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
338 kfree(ctx->dummy_ubuf);
339 kfree(ctx->cancel_table.hbs);
340 kfree(ctx->cancel_table_locked.hbs);
342 xa_destroy(&ctx->io_bl_xa);
347 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
349 struct io_rings *r = ctx->rings;
351 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
355 static bool req_need_defer(struct io_kiocb *req, u32 seq)
357 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
358 struct io_ring_ctx *ctx = req->ctx;
360 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
366 static inline void io_req_track_inflight(struct io_kiocb *req)
368 if (!(req->flags & REQ_F_INFLIGHT)) {
369 req->flags |= REQ_F_INFLIGHT;
370 atomic_inc(&req->task->io_uring->inflight_tracked);
374 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
376 if (WARN_ON_ONCE(!req->link))
379 req->flags &= ~REQ_F_ARM_LTIMEOUT;
380 req->flags |= REQ_F_LINK_TIMEOUT;
382 /* linked timeouts should have two refs once prep'ed */
383 io_req_set_refcount(req);
384 __io_req_set_refcount(req->link, 2);
388 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
390 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
392 return __io_prep_linked_timeout(req);
395 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
397 io_queue_linked_timeout(__io_prep_linked_timeout(req));
400 static inline void io_arm_ltimeout(struct io_kiocb *req)
402 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
403 __io_arm_ltimeout(req);
406 static void io_prep_async_work(struct io_kiocb *req)
408 const struct io_op_def *def = &io_op_defs[req->opcode];
409 struct io_ring_ctx *ctx = req->ctx;
411 if (!(req->flags & REQ_F_CREDS)) {
412 req->flags |= REQ_F_CREDS;
413 req->creds = get_current_cred();
416 req->work.list.next = NULL;
418 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
419 if (req->flags & REQ_F_FORCE_ASYNC)
420 req->work.flags |= IO_WQ_WORK_CONCURRENT;
422 if (req->file && !io_req_ffs_set(req))
423 req->flags |= io_file_get_flags(req->file) << REQ_F_SUPPORT_NOWAIT_BIT;
425 if (req->flags & REQ_F_ISREG) {
426 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
427 io_wq_hash_work(&req->work, file_inode(req->file));
428 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
429 if (def->unbound_nonreg_file)
430 req->work.flags |= IO_WQ_WORK_UNBOUND;
434 static void io_prep_async_link(struct io_kiocb *req)
436 struct io_kiocb *cur;
438 if (req->flags & REQ_F_LINK_TIMEOUT) {
439 struct io_ring_ctx *ctx = req->ctx;
441 spin_lock_irq(&ctx->timeout_lock);
442 io_for_each_link(cur, req)
443 io_prep_async_work(cur);
444 spin_unlock_irq(&ctx->timeout_lock);
446 io_for_each_link(cur, req)
447 io_prep_async_work(cur);
451 void io_queue_iowq(struct io_kiocb *req, bool *dont_use)
453 struct io_kiocb *link = io_prep_linked_timeout(req);
454 struct io_uring_task *tctx = req->task->io_uring;
457 BUG_ON(!tctx->io_wq);
459 /* init ->work of the whole link before punting */
460 io_prep_async_link(req);
463 * Not expected to happen, but if we do have a bug where this _can_
464 * happen, catch it here and ensure the request is marked as
465 * canceled. That will make io-wq go through the usual work cancel
466 * procedure rather than attempt to run this request (or create a new
469 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
470 req->work.flags |= IO_WQ_WORK_CANCEL;
472 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
473 io_wq_enqueue(tctx->io_wq, &req->work);
475 io_queue_linked_timeout(link);
478 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
480 while (!list_empty(&ctx->defer_list)) {
481 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
482 struct io_defer_entry, list);
484 if (req_need_defer(de->req, de->seq))
486 list_del_init(&de->list);
487 io_req_task_queue(de->req);
493 static void io_eventfd_ops(struct rcu_head *rcu)
495 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
496 int ops = atomic_xchg(&ev_fd->ops, 0);
498 if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT))
499 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
501 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
502 * ordering in a race but if references are 0 we know we have to free
505 if (atomic_dec_and_test(&ev_fd->refs)) {
506 eventfd_ctx_put(ev_fd->cq_ev_fd);
511 static void io_eventfd_signal(struct io_ring_ctx *ctx)
513 struct io_ev_fd *ev_fd = NULL;
517 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
520 ev_fd = rcu_dereference(ctx->io_ev_fd);
523 * Check again if ev_fd exists incase an io_eventfd_unregister call
524 * completed between the NULL check of ctx->io_ev_fd at the start of
525 * the function and rcu_read_lock.
527 if (unlikely(!ev_fd))
529 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
531 if (ev_fd->eventfd_async && !io_wq_current_is_worker())
534 if (likely(eventfd_signal_allowed())) {
535 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
537 atomic_inc(&ev_fd->refs);
538 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops))
539 call_rcu(&ev_fd->rcu, io_eventfd_ops);
541 atomic_dec(&ev_fd->refs);
548 static void io_eventfd_flush_signal(struct io_ring_ctx *ctx)
552 spin_lock(&ctx->completion_lock);
555 * Eventfd should only get triggered when at least one event has been
556 * posted. Some applications rely on the eventfd notification count
557 * only changing IFF a new CQE has been added to the CQ ring. There's
558 * no depedency on 1:1 relationship between how many times this
559 * function is called (and hence the eventfd count) and number of CQEs
560 * posted to the CQ ring.
562 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
563 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
564 spin_unlock(&ctx->completion_lock);
568 io_eventfd_signal(ctx);
571 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
573 if (ctx->off_timeout_used || ctx->drain_active) {
574 spin_lock(&ctx->completion_lock);
575 if (ctx->off_timeout_used)
576 io_flush_timeouts(ctx);
577 if (ctx->drain_active)
578 io_queue_deferred(ctx);
579 spin_unlock(&ctx->completion_lock);
582 io_eventfd_flush_signal(ctx);
585 static inline void io_cqring_ev_posted(struct io_ring_ctx *ctx)
587 io_commit_cqring_flush(ctx);
591 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
592 __releases(ctx->completion_lock)
594 io_commit_cqring(ctx);
595 spin_unlock(&ctx->completion_lock);
596 io_cqring_ev_posted(ctx);
599 void io_cq_unlock_post(struct io_ring_ctx *ctx)
601 __io_cq_unlock_post(ctx);
604 /* Returns true if there are no backlogged entries after the flush */
605 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
608 size_t cqe_size = sizeof(struct io_uring_cqe);
610 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
613 if (ctx->flags & IORING_SETUP_CQE32)
617 while (!list_empty(&ctx->cq_overflow_list)) {
618 struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true);
619 struct io_overflow_cqe *ocqe;
623 ocqe = list_first_entry(&ctx->cq_overflow_list,
624 struct io_overflow_cqe, list);
626 memcpy(cqe, &ocqe->cqe, cqe_size);
628 io_account_cq_overflow(ctx);
630 list_del(&ocqe->list);
634 all_flushed = list_empty(&ctx->cq_overflow_list);
636 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
637 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
640 io_cq_unlock_post(ctx);
644 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
648 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
649 /* iopoll syncs against uring_lock, not completion_lock */
650 if (ctx->flags & IORING_SETUP_IOPOLL)
651 mutex_lock(&ctx->uring_lock);
652 ret = __io_cqring_overflow_flush(ctx, false);
653 if (ctx->flags & IORING_SETUP_IOPOLL)
654 mutex_unlock(&ctx->uring_lock);
660 void __io_put_task(struct task_struct *task, int nr)
662 struct io_uring_task *tctx = task->io_uring;
664 percpu_counter_sub(&tctx->inflight, nr);
665 if (unlikely(atomic_read(&tctx->in_idle)))
666 wake_up(&tctx->wait);
667 put_task_struct_many(task, nr);
670 void io_task_refs_refill(struct io_uring_task *tctx)
672 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
674 percpu_counter_add(&tctx->inflight, refill);
675 refcount_add(refill, ¤t->usage);
676 tctx->cached_refs += refill;
679 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
681 struct io_uring_task *tctx = task->io_uring;
682 unsigned int refs = tctx->cached_refs;
685 tctx->cached_refs = 0;
686 percpu_counter_sub(&tctx->inflight, refs);
687 put_task_struct_many(task, refs);
691 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
692 s32 res, u32 cflags, u64 extra1, u64 extra2)
694 struct io_overflow_cqe *ocqe;
695 size_t ocq_size = sizeof(struct io_overflow_cqe);
696 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
699 ocq_size += sizeof(struct io_uring_cqe);
701 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
702 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
705 * If we're in ring overflow flush mode, or in task cancel mode,
706 * or cannot allocate an overflow entry, then we need to drop it
709 io_account_cq_overflow(ctx);
710 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
713 if (list_empty(&ctx->cq_overflow_list)) {
714 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
715 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
718 ocqe->cqe.user_data = user_data;
720 ocqe->cqe.flags = cflags;
722 ocqe->cqe.big_cqe[0] = extra1;
723 ocqe->cqe.big_cqe[1] = extra2;
725 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
729 bool io_req_cqe_overflow(struct io_kiocb *req)
731 if (!(req->flags & REQ_F_CQE32_INIT)) {
735 return io_cqring_event_overflow(req->ctx, req->cqe.user_data,
736 req->cqe.res, req->cqe.flags,
737 req->extra1, req->extra2);
741 * writes to the cq entry need to come after reading head; the
742 * control dependency is enough as we're using WRITE_ONCE to
745 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow)
747 struct io_rings *rings = ctx->rings;
748 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
749 unsigned int free, queued, len;
752 * Posting into the CQ when there are pending overflowed CQEs may break
753 * ordering guarantees, which will affect links, F_MORE users and more.
754 * Force overflow the completion.
756 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
759 /* userspace may cheat modifying the tail, be safe and do min */
760 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
761 free = ctx->cq_entries - queued;
762 /* we need a contiguous range, limit based on the current array offset */
763 len = min(free, ctx->cq_entries - off);
767 if (ctx->flags & IORING_SETUP_CQE32) {
772 ctx->cqe_cached = &rings->cqes[off];
773 ctx->cqe_sentinel = ctx->cqe_cached + len;
775 ctx->cached_cq_tail++;
777 if (ctx->flags & IORING_SETUP_CQE32)
779 return &rings->cqes[off];
782 bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
785 struct io_uring_cqe *cqe;
790 * If we can't get a cq entry, userspace overflowed the
791 * submission (by quite a lot). Increment the overflow count in
794 cqe = io_get_cqe(ctx);
796 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
798 WRITE_ONCE(cqe->user_data, user_data);
799 WRITE_ONCE(cqe->res, res);
800 WRITE_ONCE(cqe->flags, cflags);
802 if (ctx->flags & IORING_SETUP_CQE32) {
803 WRITE_ONCE(cqe->big_cqe[0], 0);
804 WRITE_ONCE(cqe->big_cqe[1], 0);
810 return io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
815 bool io_post_aux_cqe(struct io_ring_ctx *ctx,
816 u64 user_data, s32 res, u32 cflags,
822 filled = io_fill_cqe_aux(ctx, user_data, res, cflags, allow_overflow);
823 io_cq_unlock_post(ctx);
827 void io_req_complete_post(struct io_kiocb *req)
829 struct io_ring_ctx *ctx = req->ctx;
832 if (!(req->flags & REQ_F_CQE_SKIP))
833 __io_fill_cqe_req(ctx, req);
836 * If we're the last reference to this request, add to our locked
839 if (req_ref_put_and_test(req)) {
840 if (req->flags & IO_REQ_LINK_FLAGS) {
841 if (req->flags & IO_DISARM_MASK)
844 io_req_task_queue(req->link);
848 io_req_put_rsrc(req);
850 * Selected buffer deallocation in io_clean_op() assumes that
851 * we don't hold ->completion_lock. Clean them here to avoid
854 io_put_kbuf_comp(req);
855 io_dismantle_req(req);
856 io_put_task(req->task, 1);
857 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
858 ctx->locked_free_nr++;
860 io_cq_unlock_post(ctx);
863 inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags)
865 io_req_complete_post(req);
868 void io_req_complete_failed(struct io_kiocb *req, s32 res)
869 __must_hold(&ctx->uring_lock)
871 const struct io_op_def *def = &io_op_defs[req->opcode];
873 lockdep_assert_held(&req->ctx->uring_lock);
876 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
879 io_req_complete_post(req);
883 * Don't initialise the fields below on every allocation, but do that in
884 * advance and keep them valid across allocations.
886 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
890 req->async_data = NULL;
891 /* not necessary, but safer to zero */
895 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
896 struct io_submit_state *state)
898 spin_lock(&ctx->completion_lock);
899 wq_list_splice(&ctx->locked_free_list, &state->free_list);
900 ctx->locked_free_nr = 0;
901 spin_unlock(&ctx->completion_lock);
905 * A request might get retired back into the request caches even before opcode
906 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
907 * Because of that, io_alloc_req() should be called only under ->uring_lock
908 * and with extra caution to not get a request that is still worked on.
910 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
911 __must_hold(&ctx->uring_lock)
913 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
914 void *reqs[IO_REQ_ALLOC_BATCH];
918 * If we have more than a batch's worth of requests in our IRQ side
919 * locked cache, grab the lock and move them over to our submission
922 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
923 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
924 if (!io_req_cache_empty(ctx))
928 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
931 * Bulk alloc is all-or-nothing. If we fail to get a batch,
932 * retry single alloc to be on the safe side.
934 if (unlikely(ret <= 0)) {
935 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
941 percpu_ref_get_many(&ctx->refs, ret);
942 for (i = 0; i < ret; i++) {
943 struct io_kiocb *req = reqs[i];
945 io_preinit_req(req, ctx);
946 io_req_add_to_cache(req, ctx);
951 static inline void io_dismantle_req(struct io_kiocb *req)
953 unsigned int flags = req->flags;
955 if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
957 if (!(flags & REQ_F_FIXED_FILE))
958 io_put_file(req->file);
961 __cold void io_free_req(struct io_kiocb *req)
963 struct io_ring_ctx *ctx = req->ctx;
965 io_req_put_rsrc(req);
966 io_dismantle_req(req);
967 io_put_task(req->task, 1);
969 spin_lock(&ctx->completion_lock);
970 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
971 ctx->locked_free_nr++;
972 spin_unlock(&ctx->completion_lock);
975 static void __io_req_find_next_prep(struct io_kiocb *req)
977 struct io_ring_ctx *ctx = req->ctx;
981 io_cq_unlock_post(ctx);
984 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
986 struct io_kiocb *nxt;
989 * If LINK is set, we have dependent requests in this chain. If we
990 * didn't fail this request, queue the first one up, moving any other
991 * dependencies to the next request. In case of failure, fail the rest
994 if (unlikely(req->flags & IO_DISARM_MASK))
995 __io_req_find_next_prep(req);
1001 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
1005 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1006 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1008 io_submit_flush_completions(ctx);
1009 mutex_unlock(&ctx->uring_lock);
1012 percpu_ref_put(&ctx->refs);
1015 static unsigned int handle_tw_list(struct llist_node *node,
1016 struct io_ring_ctx **ctx, bool *locked,
1017 struct llist_node *last)
1019 unsigned int count = 0;
1021 while (node != last) {
1022 struct llist_node *next = node->next;
1023 struct io_kiocb *req = container_of(node, struct io_kiocb,
1026 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1028 if (req->ctx != *ctx) {
1029 ctx_flush_and_put(*ctx, locked);
1031 /* if not contended, grab and improve batching */
1032 *locked = mutex_trylock(&(*ctx)->uring_lock);
1033 percpu_ref_get(&(*ctx)->refs);
1034 } else if (!*locked)
1035 *locked = mutex_trylock(&(*ctx)->uring_lock);
1036 req->io_task_work.func(req, locked);
1039 if (unlikely(need_resched())) {
1040 ctx_flush_and_put(*ctx, locked);
1050 * io_llist_xchg - swap all entries in a lock-less list
1051 * @head: the head of lock-less list to delete all entries
1052 * @new: new entry as the head of the list
1054 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1055 * The order of entries returned is from the newest to the oldest added one.
1057 static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1058 struct llist_node *new)
1060 return xchg(&head->first, new);
1064 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1065 * @head: the head of lock-less list to delete all entries
1066 * @old: expected old value of the first entry of the list
1067 * @new: new entry as the head of the list
1069 * perform a cmpxchg on the first entry of the list.
1072 static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1073 struct llist_node *old,
1074 struct llist_node *new)
1076 return cmpxchg(&head->first, old, new);
1079 void tctx_task_work(struct callback_head *cb)
1081 bool uring_locked = false;
1082 struct io_ring_ctx *ctx = NULL;
1083 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1085 struct llist_node fake = {};
1086 struct llist_node *node = io_llist_xchg(&tctx->task_list, &fake);
1087 unsigned int loops = 1;
1088 unsigned int count = handle_tw_list(node, &ctx, &uring_locked, NULL);
1090 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1091 while (node != &fake) {
1093 node = io_llist_xchg(&tctx->task_list, &fake);
1094 count += handle_tw_list(node, &ctx, &uring_locked, &fake);
1095 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1098 ctx_flush_and_put(ctx, &uring_locked);
1100 /* relaxed read is enough as only the task itself sets ->in_idle */
1101 if (unlikely(atomic_read(&tctx->in_idle)))
1102 io_uring_drop_tctx_refs(current);
1104 trace_io_uring_task_work_run(tctx, count, loops);
1107 static void io_req_local_work_add(struct io_kiocb *req)
1109 struct io_ring_ctx *ctx = req->ctx;
1111 percpu_ref_get(&ctx->refs);
1113 if (!llist_add(&req->io_task_work.node, &ctx->work_llist)) {
1114 percpu_ref_put(&ctx->refs);
1117 /* need it for the following io_cqring_wake() */
1118 smp_mb__after_atomic();
1120 if (unlikely(atomic_read(&req->task->io_uring->in_idle))) {
1121 io_move_task_work_from_local(ctx);
1122 percpu_ref_put(&ctx->refs);
1126 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1127 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1130 io_eventfd_signal(ctx);
1131 __io_cqring_wake(ctx);
1132 percpu_ref_put(&ctx->refs);
1135 void __io_req_task_work_add(struct io_kiocb *req, bool allow_local)
1137 struct io_uring_task *tctx = req->task->io_uring;
1138 struct io_ring_ctx *ctx = req->ctx;
1139 struct llist_node *node;
1141 if (allow_local && ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
1142 io_req_local_work_add(req);
1146 /* task_work already pending, we're done */
1147 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1150 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1151 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1153 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1156 node = llist_del_all(&tctx->task_list);
1159 req = container_of(node, struct io_kiocb, io_task_work.node);
1161 if (llist_add(&req->io_task_work.node,
1162 &req->ctx->fallback_llist))
1163 schedule_delayed_work(&req->ctx->fallback_work, 1);
1167 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1169 struct llist_node *node;
1171 node = llist_del_all(&ctx->work_llist);
1173 struct io_kiocb *req = container_of(node, struct io_kiocb,
1177 __io_req_task_work_add(req, false);
1181 int __io_run_local_work(struct io_ring_ctx *ctx, bool *locked)
1183 struct llist_node *node;
1184 struct llist_node fake;
1185 struct llist_node *current_final = NULL;
1187 unsigned int loops = 1;
1189 if (unlikely(ctx->submitter_task != current))
1192 node = io_llist_xchg(&ctx->work_llist, &fake);
1195 while (node != current_final) {
1196 struct llist_node *next = node->next;
1197 struct io_kiocb *req = container_of(node, struct io_kiocb,
1199 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1200 req->io_task_work.func(req, locked);
1205 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1206 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1208 node = io_llist_cmpxchg(&ctx->work_llist, &fake, NULL);
1209 if (node != &fake) {
1211 current_final = &fake;
1212 node = io_llist_xchg(&ctx->work_llist, &fake);
1217 io_submit_flush_completions(ctx);
1218 trace_io_uring_local_work_run(ctx, ret, loops);
1223 int io_run_local_work(struct io_ring_ctx *ctx)
1228 if (llist_empty(&ctx->work_llist))
1231 __set_current_state(TASK_RUNNING);
1232 locked = mutex_trylock(&ctx->uring_lock);
1233 ret = __io_run_local_work(ctx, &locked);
1235 mutex_unlock(&ctx->uring_lock);
1240 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
1242 /* not needed for normal modes, but SQPOLL depends on it */
1243 io_tw_lock(req->ctx, locked);
1244 io_req_complete_failed(req, req->cqe.res);
1247 void io_req_task_submit(struct io_kiocb *req, bool *locked)
1249 io_tw_lock(req->ctx, locked);
1250 /* req->task == current here, checking PF_EXITING is safe */
1251 if (likely(!(req->task->flags & PF_EXITING)))
1254 io_req_complete_failed(req, -EFAULT);
1257 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1259 io_req_set_res(req, ret, 0);
1260 req->io_task_work.func = io_req_task_cancel;
1261 io_req_task_work_add(req);
1264 void io_req_task_queue(struct io_kiocb *req)
1266 req->io_task_work.func = io_req_task_submit;
1267 io_req_task_work_add(req);
1270 void io_queue_next(struct io_kiocb *req)
1272 struct io_kiocb *nxt = io_req_find_next(req);
1275 io_req_task_queue(nxt);
1278 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1279 __must_hold(&ctx->uring_lock)
1281 struct task_struct *task = NULL;
1285 struct io_kiocb *req = container_of(node, struct io_kiocb,
1288 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1289 if (req->flags & REQ_F_REFCOUNT) {
1290 node = req->comp_list.next;
1291 if (!req_ref_put_and_test(req))
1294 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1295 struct async_poll *apoll = req->apoll;
1297 if (apoll->double_poll)
1298 kfree(apoll->double_poll);
1299 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1301 req->flags &= ~REQ_F_POLLED;
1303 if (req->flags & IO_REQ_LINK_FLAGS)
1305 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1308 if (!(req->flags & REQ_F_FIXED_FILE))
1309 io_put_file(req->file);
1311 io_req_put_rsrc_locked(req, ctx);
1313 if (req->task != task) {
1315 io_put_task(task, task_refs);
1320 node = req->comp_list.next;
1321 io_req_add_to_cache(req, ctx);
1325 io_put_task(task, task_refs);
1328 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1329 __must_hold(&ctx->uring_lock)
1331 struct io_wq_work_node *node, *prev;
1332 struct io_submit_state *state = &ctx->submit_state;
1335 wq_list_for_each(node, prev, &state->compl_reqs) {
1336 struct io_kiocb *req = container_of(node, struct io_kiocb,
1339 if (!(req->flags & REQ_F_CQE_SKIP))
1340 __io_fill_cqe_req(ctx, req);
1342 __io_cq_unlock_post(ctx);
1344 io_free_batch_list(ctx, state->compl_reqs.first);
1345 INIT_WQ_LIST(&state->compl_reqs);
1349 * Drop reference to request, return next in chain (if there is one) if this
1350 * was the last reference to this request.
1352 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1354 struct io_kiocb *nxt = NULL;
1356 if (req_ref_put_and_test(req)) {
1357 if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
1358 nxt = io_req_find_next(req);
1364 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1366 /* See comment at the top of this file */
1368 return __io_cqring_events(ctx);
1372 * We can't just wait for polled events to come to us, we have to actively
1373 * find and complete them.
1375 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1377 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1380 mutex_lock(&ctx->uring_lock);
1381 while (!wq_list_empty(&ctx->iopoll_list)) {
1382 /* let it sleep and repeat later if can't complete a request */
1383 if (io_do_iopoll(ctx, true) == 0)
1386 * Ensure we allow local-to-the-cpu processing to take place,
1387 * in this case we need to ensure that we reap all events.
1388 * Also let task_work, etc. to progress by releasing the mutex
1390 if (need_resched()) {
1391 mutex_unlock(&ctx->uring_lock);
1393 mutex_lock(&ctx->uring_lock);
1396 mutex_unlock(&ctx->uring_lock);
1399 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1401 unsigned int nr_events = 0;
1403 unsigned long check_cq;
1405 if (!io_allowed_run_tw(ctx))
1408 check_cq = READ_ONCE(ctx->check_cq);
1409 if (unlikely(check_cq)) {
1410 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1411 __io_cqring_overflow_flush(ctx, false);
1413 * Similarly do not spin if we have not informed the user of any
1416 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1420 * Don't enter poll loop if we already have events pending.
1421 * If we do, we can potentially be spinning for commands that
1422 * already triggered a CQE (eg in error).
1424 if (io_cqring_events(ctx))
1429 * If a submit got punted to a workqueue, we can have the
1430 * application entering polling for a command before it gets
1431 * issued. That app will hold the uring_lock for the duration
1432 * of the poll right here, so we need to take a breather every
1433 * now and then to ensure that the issue has a chance to add
1434 * the poll to the issued list. Otherwise we can spin here
1435 * forever, while the workqueue is stuck trying to acquire the
1438 if (wq_list_empty(&ctx->iopoll_list) ||
1439 io_task_work_pending(ctx)) {
1440 u32 tail = ctx->cached_cq_tail;
1442 (void) io_run_local_work_locked(ctx);
1444 if (task_work_pending(current) ||
1445 wq_list_empty(&ctx->iopoll_list)) {
1446 mutex_unlock(&ctx->uring_lock);
1448 mutex_lock(&ctx->uring_lock);
1450 /* some requests don't go through iopoll_list */
1451 if (tail != ctx->cached_cq_tail ||
1452 wq_list_empty(&ctx->iopoll_list))
1455 ret = io_do_iopoll(ctx, !min);
1460 } while (nr_events < min && !need_resched());
1465 void io_req_task_complete(struct io_kiocb *req, bool *locked)
1467 if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) {
1468 unsigned issue_flags = *locked ? 0 : IO_URING_F_UNLOCKED;
1470 req->cqe.flags |= io_put_kbuf(req, issue_flags);
1474 io_req_complete_defer(req);
1476 io_req_complete_post(req);
1480 * After the iocb has been issued, it's safe to be found on the poll list.
1481 * Adding the kiocb to the list AFTER submission ensures that we don't
1482 * find it from a io_do_iopoll() thread before the issuer is done
1483 * accessing the kiocb cookie.
1485 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1487 struct io_ring_ctx *ctx = req->ctx;
1488 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1490 /* workqueue context doesn't hold uring_lock, grab it now */
1491 if (unlikely(needs_lock))
1492 mutex_lock(&ctx->uring_lock);
1495 * Track whether we have multiple files in our lists. This will impact
1496 * how we do polling eventually, not spinning if we're on potentially
1497 * different devices.
1499 if (wq_list_empty(&ctx->iopoll_list)) {
1500 ctx->poll_multi_queue = false;
1501 } else if (!ctx->poll_multi_queue) {
1502 struct io_kiocb *list_req;
1504 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1506 if (list_req->file != req->file)
1507 ctx->poll_multi_queue = true;
1511 * For fast devices, IO may have already completed. If it has, add
1512 * it to the front so we find it first.
1514 if (READ_ONCE(req->iopoll_completed))
1515 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1517 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1519 if (unlikely(needs_lock)) {
1521 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1522 * in sq thread task context or in io worker task context. If
1523 * current task context is sq thread, we don't need to check
1524 * whether should wake up sq thread.
1526 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1527 wq_has_sleeper(&ctx->sq_data->wait))
1528 wake_up(&ctx->sq_data->wait);
1530 mutex_unlock(&ctx->uring_lock);
1534 static bool io_bdev_nowait(struct block_device *bdev)
1536 return !bdev || bdev_nowait(bdev);
1540 * If we tracked the file through the SCM inflight mechanism, we could support
1541 * any file. For now, just ensure that anything potentially problematic is done
1544 static bool __io_file_supports_nowait(struct file *file, umode_t mode)
1546 if (S_ISBLK(mode)) {
1547 if (IS_ENABLED(CONFIG_BLOCK) &&
1548 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
1554 if (S_ISREG(mode)) {
1555 if (IS_ENABLED(CONFIG_BLOCK) &&
1556 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
1557 !io_is_uring_fops(file))
1562 /* any ->read/write should understand O_NONBLOCK */
1563 if (file->f_flags & O_NONBLOCK)
1565 return file->f_mode & FMODE_NOWAIT;
1569 * If we tracked the file through the SCM inflight mechanism, we could support
1570 * any file. For now, just ensure that anything potentially problematic is done
1573 unsigned int io_file_get_flags(struct file *file)
1575 umode_t mode = file_inode(file)->i_mode;
1576 unsigned int res = 0;
1580 if (__io_file_supports_nowait(file, mode))
1585 bool io_alloc_async_data(struct io_kiocb *req)
1587 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
1588 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
1589 if (req->async_data) {
1590 req->flags |= REQ_F_ASYNC_DATA;
1596 int io_req_prep_async(struct io_kiocb *req)
1598 const struct io_op_def *def = &io_op_defs[req->opcode];
1600 /* assign early for deferred execution for non-fixed file */
1601 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE) && !req->file)
1602 req->file = io_file_get_normal(req, req->cqe.fd);
1603 if (!def->prep_async)
1605 if (WARN_ON_ONCE(req_has_async_data(req)))
1607 if (!io_op_defs[req->opcode].manual_alloc) {
1608 if (io_alloc_async_data(req))
1611 return def->prep_async(req);
1614 static u32 io_get_sequence(struct io_kiocb *req)
1616 u32 seq = req->ctx->cached_sq_head;
1617 struct io_kiocb *cur;
1619 /* need original cached_sq_head, but it was increased for each req */
1620 io_for_each_link(cur, req)
1625 static __cold void io_drain_req(struct io_kiocb *req)
1626 __must_hold(&ctx->uring_lock)
1628 struct io_ring_ctx *ctx = req->ctx;
1629 struct io_defer_entry *de;
1631 u32 seq = io_get_sequence(req);
1633 /* Still need defer if there is pending req in defer list. */
1634 spin_lock(&ctx->completion_lock);
1635 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1636 spin_unlock(&ctx->completion_lock);
1638 ctx->drain_active = false;
1639 io_req_task_queue(req);
1642 spin_unlock(&ctx->completion_lock);
1644 io_prep_async_link(req);
1645 de = kmalloc(sizeof(*de), GFP_KERNEL);
1648 io_req_complete_failed(req, ret);
1652 spin_lock(&ctx->completion_lock);
1653 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1654 spin_unlock(&ctx->completion_lock);
1659 trace_io_uring_defer(req);
1662 list_add_tail(&de->list, &ctx->defer_list);
1663 spin_unlock(&ctx->completion_lock);
1666 static void io_clean_op(struct io_kiocb *req)
1668 if (req->flags & REQ_F_BUFFER_SELECTED) {
1669 spin_lock(&req->ctx->completion_lock);
1670 io_put_kbuf_comp(req);
1671 spin_unlock(&req->ctx->completion_lock);
1674 if (req->flags & REQ_F_NEED_CLEANUP) {
1675 const struct io_op_def *def = &io_op_defs[req->opcode];
1680 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1681 kfree(req->apoll->double_poll);
1685 if (req->flags & REQ_F_INFLIGHT) {
1686 struct io_uring_task *tctx = req->task->io_uring;
1688 atomic_dec(&tctx->inflight_tracked);
1690 if (req->flags & REQ_F_CREDS)
1691 put_cred(req->creds);
1692 if (req->flags & REQ_F_ASYNC_DATA) {
1693 kfree(req->async_data);
1694 req->async_data = NULL;
1696 req->flags &= ~IO_REQ_CLEAN_FLAGS;
1699 static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags)
1701 if (req->file || !io_op_defs[req->opcode].needs_file)
1704 if (req->flags & REQ_F_FIXED_FILE)
1705 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1707 req->file = io_file_get_normal(req, req->cqe.fd);
1712 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1714 const struct io_op_def *def = &io_op_defs[req->opcode];
1715 const struct cred *creds = NULL;
1718 if (unlikely(!io_assign_file(req, issue_flags)))
1721 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1722 creds = override_creds(req->creds);
1724 if (!def->audit_skip)
1725 audit_uring_entry(req->opcode);
1727 ret = def->issue(req, issue_flags);
1729 if (!def->audit_skip)
1730 audit_uring_exit(!ret, ret);
1733 revert_creds(creds);
1735 if (ret == IOU_OK) {
1736 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1737 io_req_complete_defer(req);
1739 io_req_complete_post(req);
1740 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1743 /* If the op doesn't have a file, we're not polling for it */
1744 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
1745 io_iopoll_req_issued(req, issue_flags);
1750 int io_poll_issue(struct io_kiocb *req, bool *locked)
1752 io_tw_lock(req->ctx, locked);
1753 if (unlikely(req->task->flags & PF_EXITING))
1755 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT);
1758 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1760 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1762 req = io_put_req_find_next(req);
1763 return req ? &req->work : NULL;
1766 void io_wq_submit_work(struct io_wq_work *work)
1768 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1769 const struct io_op_def *def = &io_op_defs[req->opcode];
1770 unsigned int issue_flags = IO_URING_F_UNLOCKED;
1771 bool needs_poll = false;
1772 int ret = 0, err = -ECANCELED;
1774 /* one will be dropped by ->io_free_work() after returning to io-wq */
1775 if (!(req->flags & REQ_F_REFCOUNT))
1776 __io_req_set_refcount(req, 2);
1780 io_arm_ltimeout(req);
1782 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1783 if (work->flags & IO_WQ_WORK_CANCEL) {
1785 io_req_task_queue_fail(req, err);
1788 if (!io_assign_file(req, issue_flags)) {
1790 work->flags |= IO_WQ_WORK_CANCEL;
1794 if (req->flags & REQ_F_FORCE_ASYNC) {
1795 bool opcode_poll = def->pollin || def->pollout;
1797 if (opcode_poll && file_can_poll(req->file)) {
1799 issue_flags |= IO_URING_F_NONBLOCK;
1804 ret = io_issue_sqe(req, issue_flags);
1809 * If REQ_F_NOWAIT is set, then don't wait or retry with
1810 * poll. -EAGAIN is final for that case.
1812 if (req->flags & REQ_F_NOWAIT)
1816 * We can get EAGAIN for iopolled IO even though we're
1817 * forcing a sync submission from here, since we can't
1818 * wait for request slots on the block side.
1821 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1827 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1829 /* aborted or ready, in either case retry blocking */
1831 issue_flags &= ~IO_URING_F_NONBLOCK;
1834 /* avoid locking problems by failing it from a clean context */
1836 io_req_task_queue_fail(req, ret);
1839 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1840 unsigned int issue_flags)
1842 struct io_ring_ctx *ctx = req->ctx;
1843 struct file *file = NULL;
1844 unsigned long file_ptr;
1846 io_ring_submit_lock(ctx, issue_flags);
1848 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
1850 fd = array_index_nospec(fd, ctx->nr_user_files);
1851 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
1852 file = (struct file *) (file_ptr & FFS_MASK);
1853 file_ptr &= ~FFS_MASK;
1854 /* mask in overlapping REQ_F and FFS bits */
1855 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
1856 io_req_set_rsrc_node(req, ctx, 0);
1858 io_ring_submit_unlock(ctx, issue_flags);
1862 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
1864 struct file *file = fget(fd);
1866 trace_io_uring_file_get(req, fd);
1868 /* we don't allow fixed io_uring files */
1869 if (file && io_is_uring_fops(file))
1870 io_req_track_inflight(req);
1874 static void io_queue_async(struct io_kiocb *req, int ret)
1875 __must_hold(&req->ctx->uring_lock)
1877 struct io_kiocb *linked_timeout;
1879 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
1880 io_req_complete_failed(req, ret);
1884 linked_timeout = io_prep_linked_timeout(req);
1886 switch (io_arm_poll_handler(req, 0)) {
1887 case IO_APOLL_READY:
1888 io_kbuf_recycle(req, 0);
1889 io_req_task_queue(req);
1891 case IO_APOLL_ABORTED:
1892 io_kbuf_recycle(req, 0);
1893 io_queue_iowq(req, NULL);
1900 io_queue_linked_timeout(linked_timeout);
1903 static inline void io_queue_sqe(struct io_kiocb *req)
1904 __must_hold(&req->ctx->uring_lock)
1908 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
1911 * We async punt it if the file wasn't marked NOWAIT, or if the file
1912 * doesn't support non-blocking read/write attempts
1915 io_arm_ltimeout(req);
1917 io_queue_async(req, ret);
1920 static void io_queue_sqe_fallback(struct io_kiocb *req)
1921 __must_hold(&req->ctx->uring_lock)
1923 if (unlikely(req->flags & REQ_F_FAIL)) {
1925 * We don't submit, fail them all, for that replace hardlinks
1926 * with normal links. Extra REQ_F_LINK is tolerated.
1928 req->flags &= ~REQ_F_HARDLINK;
1929 req->flags |= REQ_F_LINK;
1930 io_req_complete_failed(req, req->cqe.res);
1932 int ret = io_req_prep_async(req);
1934 if (unlikely(ret)) {
1935 io_req_complete_failed(req, ret);
1939 if (unlikely(req->ctx->drain_active))
1942 io_queue_iowq(req, NULL);
1947 * Check SQE restrictions (opcode and flags).
1949 * Returns 'true' if SQE is allowed, 'false' otherwise.
1951 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
1952 struct io_kiocb *req,
1953 unsigned int sqe_flags)
1955 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
1958 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
1959 ctx->restrictions.sqe_flags_required)
1962 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
1963 ctx->restrictions.sqe_flags_required))
1969 static void io_init_req_drain(struct io_kiocb *req)
1971 struct io_ring_ctx *ctx = req->ctx;
1972 struct io_kiocb *head = ctx->submit_state.link.head;
1974 ctx->drain_active = true;
1977 * If we need to drain a request in the middle of a link, drain
1978 * the head request and the next request/link after the current
1979 * link. Considering sequential execution of links,
1980 * REQ_F_IO_DRAIN will be maintained for every request of our
1983 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
1984 ctx->drain_next = true;
1988 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
1989 const struct io_uring_sqe *sqe)
1990 __must_hold(&ctx->uring_lock)
1992 const struct io_op_def *def;
1993 unsigned int sqe_flags;
1997 /* req is partially pre-initialised, see io_preinit_req() */
1998 req->opcode = opcode = READ_ONCE(sqe->opcode);
1999 /* same numerical values with corresponding REQ_F_*, safe to copy */
2000 req->flags = sqe_flags = READ_ONCE(sqe->flags);
2001 req->cqe.user_data = READ_ONCE(sqe->user_data);
2003 req->rsrc_node = NULL;
2004 req->task = current;
2006 if (unlikely(opcode >= IORING_OP_LAST)) {
2010 def = &io_op_defs[opcode];
2011 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2012 /* enforce forwards compatibility on users */
2013 if (sqe_flags & ~SQE_VALID_FLAGS)
2015 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2016 if (!def->buffer_select)
2018 req->buf_index = READ_ONCE(sqe->buf_group);
2020 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2021 ctx->drain_disabled = true;
2022 if (sqe_flags & IOSQE_IO_DRAIN) {
2023 if (ctx->drain_disabled)
2025 io_init_req_drain(req);
2028 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2029 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2031 /* knock it to the slow queue path, will be drained there */
2032 if (ctx->drain_active)
2033 req->flags |= REQ_F_FORCE_ASYNC;
2034 /* if there is no link, we're at "next" request and need to drain */
2035 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2036 ctx->drain_next = false;
2037 ctx->drain_active = true;
2038 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2042 if (!def->ioprio && sqe->ioprio)
2044 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2047 if (def->needs_file) {
2048 struct io_submit_state *state = &ctx->submit_state;
2050 req->cqe.fd = READ_ONCE(sqe->fd);
2053 * Plug now if we have more than 2 IO left after this, and the
2054 * target is potentially a read/write to block based storage.
2056 if (state->need_plug && def->plug) {
2057 state->plug_started = true;
2058 state->need_plug = false;
2059 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2063 personality = READ_ONCE(sqe->personality);
2067 req->creds = xa_load(&ctx->personalities, personality);
2070 get_cred(req->creds);
2071 ret = security_uring_override_creds(req->creds);
2073 put_cred(req->creds);
2076 req->flags |= REQ_F_CREDS;
2079 return def->prep(req, sqe);
2082 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2083 struct io_kiocb *req, int ret)
2085 struct io_ring_ctx *ctx = req->ctx;
2086 struct io_submit_link *link = &ctx->submit_state.link;
2087 struct io_kiocb *head = link->head;
2089 trace_io_uring_req_failed(sqe, req, ret);
2092 * Avoid breaking links in the middle as it renders links with SQPOLL
2093 * unusable. Instead of failing eagerly, continue assembling the link if
2094 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2095 * should find the flag and handle the rest.
2097 req_fail_link_node(req, ret);
2098 if (head && !(head->flags & REQ_F_FAIL))
2099 req_fail_link_node(head, -ECANCELED);
2101 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2103 link->last->link = req;
2107 io_queue_sqe_fallback(req);
2112 link->last->link = req;
2119 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2120 const struct io_uring_sqe *sqe)
2121 __must_hold(&ctx->uring_lock)
2123 struct io_submit_link *link = &ctx->submit_state.link;
2126 ret = io_init_req(ctx, req, sqe);
2128 return io_submit_fail_init(sqe, req, ret);
2130 /* don't need @sqe from now on */
2131 trace_io_uring_submit_sqe(req, true);
2134 * If we already have a head request, queue this one for async
2135 * submittal once the head completes. If we don't have a head but
2136 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2137 * submitted sync once the chain is complete. If none of those
2138 * conditions are true (normal request), then just queue it.
2140 if (unlikely(link->head)) {
2141 ret = io_req_prep_async(req);
2143 return io_submit_fail_init(sqe, req, ret);
2145 trace_io_uring_link(req, link->head);
2146 link->last->link = req;
2149 if (req->flags & IO_REQ_LINK_FLAGS)
2151 /* last request of the link, flush it */
2154 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2157 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2158 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2159 if (req->flags & IO_REQ_LINK_FLAGS) {
2164 io_queue_sqe_fallback(req);
2174 * Batched submission is done, ensure local IO is flushed out.
2176 static void io_submit_state_end(struct io_ring_ctx *ctx)
2178 struct io_submit_state *state = &ctx->submit_state;
2180 if (unlikely(state->link.head))
2181 io_queue_sqe_fallback(state->link.head);
2182 /* flush only after queuing links as they can generate completions */
2183 io_submit_flush_completions(ctx);
2184 if (state->plug_started)
2185 blk_finish_plug(&state->plug);
2189 * Start submission side cache.
2191 static void io_submit_state_start(struct io_submit_state *state,
2192 unsigned int max_ios)
2194 state->plug_started = false;
2195 state->need_plug = max_ios > 2;
2196 state->submit_nr = max_ios;
2197 /* set only head, no need to init link_last in advance */
2198 state->link.head = NULL;
2201 static void io_commit_sqring(struct io_ring_ctx *ctx)
2203 struct io_rings *rings = ctx->rings;
2206 * Ensure any loads from the SQEs are done at this point,
2207 * since once we write the new head, the application could
2208 * write new data to them.
2210 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2214 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2215 * that is mapped by userspace. This means that care needs to be taken to
2216 * ensure that reads are stable, as we cannot rely on userspace always
2217 * being a good citizen. If members of the sqe are validated and then later
2218 * used, it's important that those reads are done through READ_ONCE() to
2219 * prevent a re-load down the line.
2221 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
2223 unsigned head, mask = ctx->sq_entries - 1;
2224 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2227 * The cached sq head (or cq tail) serves two purposes:
2229 * 1) allows us to batch the cost of updating the user visible
2231 * 2) allows the kernel side to track the head on its own, even
2232 * though the application is the one updating it.
2234 head = READ_ONCE(ctx->sq_array[sq_idx]);
2235 if (likely(head < ctx->sq_entries)) {
2236 /* double index for 128-byte SQEs, twice as long */
2237 if (ctx->flags & IORING_SETUP_SQE128)
2239 return &ctx->sq_sqes[head];
2242 /* drop invalid entries */
2243 spin_lock(&ctx->completion_lock);
2245 spin_unlock(&ctx->completion_lock);
2246 WRITE_ONCE(ctx->rings->sq_dropped,
2247 READ_ONCE(ctx->rings->sq_dropped) + 1);
2251 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2252 __must_hold(&ctx->uring_lock)
2254 unsigned int entries = io_sqring_entries(ctx);
2258 if (unlikely(!entries))
2260 /* make sure SQ entry isn't read before tail */
2261 ret = left = min3(nr, ctx->sq_entries, entries);
2262 io_get_task_refs(left);
2263 io_submit_state_start(&ctx->submit_state, left);
2266 const struct io_uring_sqe *sqe;
2267 struct io_kiocb *req;
2269 if (unlikely(!io_alloc_req_refill(ctx)))
2271 req = io_alloc_req(ctx);
2272 sqe = io_get_sqe(ctx);
2273 if (unlikely(!sqe)) {
2274 io_req_add_to_cache(req, ctx);
2279 * Continue submitting even for sqe failure if the
2280 * ring was setup with IORING_SETUP_SUBMIT_ALL
2282 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2283 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2289 if (unlikely(left)) {
2291 /* try again if it submitted nothing and can't allocate a req */
2292 if (!ret && io_req_cache_empty(ctx))
2294 current->io_uring->cached_refs += left;
2297 io_submit_state_end(ctx);
2298 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2299 io_commit_sqring(ctx);
2303 struct io_wait_queue {
2304 struct wait_queue_entry wq;
2305 struct io_ring_ctx *ctx;
2307 unsigned nr_timeouts;
2310 static inline bool io_has_work(struct io_ring_ctx *ctx)
2312 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
2313 ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
2314 !llist_empty(&ctx->work_llist));
2317 static inline bool io_should_wake(struct io_wait_queue *iowq)
2319 struct io_ring_ctx *ctx = iowq->ctx;
2320 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
2323 * Wake up if we have enough events, or if a timeout occurred since we
2324 * started waiting. For timeouts, we always want to return to userspace,
2325 * regardless of event count.
2327 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2330 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2331 int wake_flags, void *key)
2333 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
2335 struct io_ring_ctx *ctx = iowq->ctx;
2338 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2339 * the task, and the next invocation will do it.
2341 if (io_should_wake(iowq) || io_has_work(ctx))
2342 return autoremove_wake_function(curr, mode, wake_flags, key);
2346 int io_run_task_work_sig(struct io_ring_ctx *ctx)
2348 if (io_run_task_work_ctx(ctx) > 0)
2350 if (task_sigpending(current))
2355 static bool current_pending_io(void)
2357 struct io_uring_task *tctx = current->io_uring;
2361 return percpu_counter_read_positive(&tctx->inflight);
2364 /* when returns >0, the caller should retry */
2365 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2366 struct io_wait_queue *iowq,
2370 unsigned long check_cq;
2372 /* make sure we run task_work before checking for signals */
2373 ret = io_run_task_work_sig(ctx);
2374 if (ret || io_should_wake(iowq))
2377 check_cq = READ_ONCE(ctx->check_cq);
2378 if (unlikely(check_cq)) {
2379 /* let the caller flush overflows, retry */
2380 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2382 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
2387 * Mark us as being in io_wait if we have pending requests, so cpufreq
2388 * can take into account that the task is waiting for IO - turns out
2389 * to be important for low QD IO.
2391 io_wait = current->in_iowait;
2392 if (current_pending_io())
2393 current->in_iowait = 1;
2395 if (!schedule_hrtimeout(timeout, HRTIMER_MODE_ABS))
2397 current->in_iowait = io_wait;
2402 * Wait until events become available, if we don't already have some. The
2403 * application must reap them itself, as they reside on the shared cq ring.
2405 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2406 const sigset_t __user *sig, size_t sigsz,
2407 struct __kernel_timespec __user *uts)
2409 struct io_wait_queue iowq;
2410 struct io_rings *rings = ctx->rings;
2411 ktime_t timeout = KTIME_MAX;
2414 if (!io_allowed_run_tw(ctx))
2418 /* always run at least 1 task work to process local work */
2419 ret = io_run_task_work_ctx(ctx);
2422 io_cqring_overflow_flush(ctx);
2424 /* if user messes with these they will just get an early return */
2425 if (__io_cqring_events_user(ctx) >= min_events)
2430 #ifdef CONFIG_COMPAT
2431 if (in_compat_syscall())
2432 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2436 ret = set_user_sigmask(sig, sigsz);
2443 struct timespec64 ts;
2445 if (get_timespec64(&ts, uts))
2447 timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2450 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2451 iowq.wq.private = current;
2452 INIT_LIST_HEAD(&iowq.wq.entry);
2454 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2455 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2457 trace_io_uring_cqring_wait(ctx, min_events);
2459 /* if we can't even flush overflow, don't wait for more */
2460 if (!io_cqring_overflow_flush(ctx)) {
2464 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2465 TASK_INTERRUPTIBLE);
2466 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
2470 finish_wait(&ctx->cq_wait, &iowq.wq);
2471 restore_saved_sigmask_unless(ret == -EINTR);
2473 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2476 static void io_mem_free(void *ptr)
2483 page = virt_to_head_page(ptr);
2484 if (put_page_testzero(page))
2485 free_compound_page(page);
2488 static void *io_mem_alloc(size_t size)
2490 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2492 return (void *) __get_free_pages(gfp, get_order(size));
2495 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2496 unsigned int cq_entries, size_t *sq_offset)
2498 struct io_rings *rings;
2499 size_t off, sq_array_size;
2501 off = struct_size(rings, cqes, cq_entries);
2502 if (off == SIZE_MAX)
2504 if (ctx->flags & IORING_SETUP_CQE32) {
2505 if (check_shl_overflow(off, 1, &off))
2510 off = ALIGN(off, SMP_CACHE_BYTES);
2518 sq_array_size = array_size(sizeof(u32), sq_entries);
2519 if (sq_array_size == SIZE_MAX)
2522 if (check_add_overflow(off, sq_array_size, &off))
2528 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2529 unsigned int eventfd_async)
2531 struct io_ev_fd *ev_fd;
2532 __s32 __user *fds = arg;
2535 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2536 lockdep_is_held(&ctx->uring_lock));
2540 if (copy_from_user(&fd, fds, sizeof(*fds)))
2543 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2547 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2548 if (IS_ERR(ev_fd->cq_ev_fd)) {
2549 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2554 spin_lock(&ctx->completion_lock);
2555 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2556 spin_unlock(&ctx->completion_lock);
2558 ev_fd->eventfd_async = eventfd_async;
2559 ctx->has_evfd = true;
2560 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2561 atomic_set(&ev_fd->refs, 1);
2562 atomic_set(&ev_fd->ops, 0);
2566 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2568 struct io_ev_fd *ev_fd;
2570 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2571 lockdep_is_held(&ctx->uring_lock));
2573 ctx->has_evfd = false;
2574 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2575 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops))
2576 call_rcu(&ev_fd->rcu, io_eventfd_ops);
2583 static void io_req_caches_free(struct io_ring_ctx *ctx)
2587 mutex_lock(&ctx->uring_lock);
2588 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
2590 while (!io_req_cache_empty(ctx)) {
2591 struct io_kiocb *req = io_alloc_req(ctx);
2593 kmem_cache_free(req_cachep, req);
2597 percpu_ref_put_many(&ctx->refs, nr);
2598 mutex_unlock(&ctx->uring_lock);
2601 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2603 io_sq_thread_finish(ctx);
2604 io_rsrc_refs_drop(ctx);
2605 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2606 io_wait_rsrc_data(ctx->buf_data);
2607 io_wait_rsrc_data(ctx->file_data);
2609 mutex_lock(&ctx->uring_lock);
2611 __io_sqe_buffers_unregister(ctx);
2613 __io_sqe_files_unregister(ctx);
2615 __io_cqring_overflow_flush(ctx, true);
2616 io_eventfd_unregister(ctx);
2617 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2618 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2619 io_destroy_buffers(ctx);
2620 mutex_unlock(&ctx->uring_lock);
2622 put_cred(ctx->sq_creds);
2623 if (ctx->submitter_task)
2624 put_task_struct(ctx->submitter_task);
2626 /* there are no registered resources left, nobody uses it */
2628 io_rsrc_node_destroy(ctx->rsrc_node);
2629 if (ctx->rsrc_backup_node)
2630 io_rsrc_node_destroy(ctx->rsrc_backup_node);
2631 flush_delayed_work(&ctx->rsrc_put_work);
2632 flush_delayed_work(&ctx->fallback_work);
2634 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2635 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
2637 #if defined(CONFIG_UNIX)
2638 if (ctx->ring_sock) {
2639 ctx->ring_sock->file = NULL; /* so that iput() is called */
2640 sock_release(ctx->ring_sock);
2643 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2645 if (ctx->mm_account) {
2646 mmdrop(ctx->mm_account);
2647 ctx->mm_account = NULL;
2649 io_mem_free(ctx->rings);
2650 io_mem_free(ctx->sq_sqes);
2652 percpu_ref_exit(&ctx->refs);
2653 free_uid(ctx->user);
2654 io_req_caches_free(ctx);
2656 io_wq_put_hash(ctx->hash_map);
2657 kfree(ctx->cancel_table.hbs);
2658 kfree(ctx->cancel_table_locked.hbs);
2659 kfree(ctx->dummy_ubuf);
2661 xa_destroy(&ctx->io_bl_xa);
2665 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2667 struct io_ring_ctx *ctx = file->private_data;
2670 poll_wait(file, &ctx->cq_wait, wait);
2672 * synchronizes with barrier from wq_has_sleeper call in
2676 if (!io_sqring_full(ctx))
2677 mask |= EPOLLOUT | EPOLLWRNORM;
2680 * Don't flush cqring overflow list here, just do a simple check.
2681 * Otherwise there could possible be ABBA deadlock:
2684 * lock(&ctx->uring_lock);
2686 * lock(&ctx->uring_lock);
2689 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2690 * pushs them to do the flush.
2693 if (__io_cqring_events_user(ctx) || io_has_work(ctx))
2694 mask |= EPOLLIN | EPOLLRDNORM;
2699 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2701 const struct cred *creds;
2703 creds = xa_erase(&ctx->personalities, id);
2712 struct io_tctx_exit {
2713 struct callback_head task_work;
2714 struct completion completion;
2715 struct io_ring_ctx *ctx;
2718 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2720 struct io_uring_task *tctx = current->io_uring;
2721 struct io_tctx_exit *work;
2723 work = container_of(cb, struct io_tctx_exit, task_work);
2725 * When @in_idle, we're in cancellation and it's racy to remove the
2726 * node. It'll be removed by the end of cancellation, just ignore it.
2727 * tctx can be NULL if the queueing of this task_work raced with
2728 * work cancelation off the exec path.
2730 if (tctx && !atomic_read(&tctx->in_idle))
2731 io_uring_del_tctx_node((unsigned long)work->ctx);
2732 complete(&work->completion);
2735 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2737 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2739 return req->ctx == data;
2742 static __cold void io_ring_exit_work(struct work_struct *work)
2744 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2745 unsigned long timeout = jiffies + HZ * 60 * 5;
2746 unsigned long interval = HZ / 20;
2747 struct io_tctx_exit exit;
2748 struct io_tctx_node *node;
2752 * If we're doing polled IO and end up having requests being
2753 * submitted async (out-of-line), then completions can come in while
2754 * we're waiting for refs to drop. We need to reap these manually,
2755 * as nobody else will be looking for them.
2758 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
2759 io_move_task_work_from_local(ctx);
2761 while (io_uring_try_cancel_requests(ctx, NULL, true))
2765 struct io_sq_data *sqd = ctx->sq_data;
2766 struct task_struct *tsk;
2768 io_sq_thread_park(sqd);
2770 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
2771 io_wq_cancel_cb(tsk->io_uring->io_wq,
2772 io_cancel_ctx_cb, ctx, true);
2773 io_sq_thread_unpark(sqd);
2776 io_req_caches_free(ctx);
2778 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
2779 /* there is little hope left, don't run it too often */
2783 * This is really an uninterruptible wait, as it has to be
2784 * complete. But it's also run from a kworker, which doesn't
2785 * take signals, so it's fine to make it interruptible. This
2786 * avoids scenarios where we knowingly can wait much longer
2787 * on completions, for example if someone does a SIGSTOP on
2788 * a task that needs to finish task_work to make this loop
2789 * complete. That's a synthetic situation that should not
2790 * cause a stuck task backtrace, and hence a potential panic
2791 * on stuck tasks if that is enabled.
2793 } while (!wait_for_completion_interruptible_timeout(&ctx->ref_comp, interval));
2795 init_completion(&exit.completion);
2796 init_task_work(&exit.task_work, io_tctx_exit_cb);
2799 * Some may use context even when all refs and requests have been put,
2800 * and they are free to do so while still holding uring_lock or
2801 * completion_lock, see io_req_task_submit(). Apart from other work,
2802 * this lock/unlock section also waits them to finish.
2804 mutex_lock(&ctx->uring_lock);
2805 while (!list_empty(&ctx->tctx_list)) {
2806 WARN_ON_ONCE(time_after(jiffies, timeout));
2808 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
2810 /* don't spin on a single task if cancellation failed */
2811 list_rotate_left(&ctx->tctx_list);
2812 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
2813 if (WARN_ON_ONCE(ret))
2816 mutex_unlock(&ctx->uring_lock);
2818 * See comment above for
2819 * wait_for_completion_interruptible_timeout() on why this
2820 * wait is marked as interruptible.
2822 wait_for_completion_interruptible(&exit.completion);
2823 mutex_lock(&ctx->uring_lock);
2825 mutex_unlock(&ctx->uring_lock);
2826 spin_lock(&ctx->completion_lock);
2827 spin_unlock(&ctx->completion_lock);
2829 io_ring_ctx_free(ctx);
2832 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
2834 unsigned long index;
2835 struct creds *creds;
2837 mutex_lock(&ctx->uring_lock);
2838 percpu_ref_kill(&ctx->refs);
2840 __io_cqring_overflow_flush(ctx, true);
2841 xa_for_each(&ctx->personalities, index, creds)
2842 io_unregister_personality(ctx, index);
2844 io_poll_remove_all(ctx, NULL, true);
2845 mutex_unlock(&ctx->uring_lock);
2848 * If we failed setting up the ctx, we might not have any rings
2849 * and therefore did not submit any requests
2852 io_kill_timeouts(ctx, NULL, true);
2854 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
2856 * Use system_unbound_wq to avoid spawning tons of event kworkers
2857 * if we're exiting a ton of rings at the same time. It just adds
2858 * noise and overhead, there's no discernable change in runtime
2859 * over using system_wq.
2861 queue_work(system_unbound_wq, &ctx->exit_work);
2864 static int io_uring_release(struct inode *inode, struct file *file)
2866 struct io_ring_ctx *ctx = file->private_data;
2868 file->private_data = NULL;
2869 io_ring_ctx_wait_and_kill(ctx);
2873 struct io_task_cancel {
2874 struct task_struct *task;
2878 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
2880 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2881 struct io_task_cancel *cancel = data;
2883 return io_match_task_safe(req, cancel->task, cancel->all);
2886 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
2887 struct task_struct *task,
2890 struct io_defer_entry *de;
2893 spin_lock(&ctx->completion_lock);
2894 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
2895 if (io_match_task_safe(de->req, task, cancel_all)) {
2896 list_cut_position(&list, &ctx->defer_list, &de->list);
2900 spin_unlock(&ctx->completion_lock);
2901 if (list_empty(&list))
2904 while (!list_empty(&list)) {
2905 de = list_first_entry(&list, struct io_defer_entry, list);
2906 list_del_init(&de->list);
2907 io_req_task_queue_fail(de->req, -ECANCELED);
2913 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
2915 struct io_tctx_node *node;
2916 enum io_wq_cancel cret;
2919 mutex_lock(&ctx->uring_lock);
2920 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
2921 struct io_uring_task *tctx = node->task->io_uring;
2924 * io_wq will stay alive while we hold uring_lock, because it's
2925 * killed after ctx nodes, which requires to take the lock.
2927 if (!tctx || !tctx->io_wq)
2929 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
2930 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2932 mutex_unlock(&ctx->uring_lock);
2937 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
2938 struct task_struct *task,
2941 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
2942 struct io_uring_task *tctx = task ? task->io_uring : NULL;
2943 enum io_wq_cancel cret;
2946 /* failed during ring init, it couldn't have issued any requests */
2951 ret |= io_uring_try_cancel_iowq(ctx);
2952 } else if (tctx && tctx->io_wq) {
2954 * Cancels requests of all rings, not only @ctx, but
2955 * it's fine as the task is in exit/exec.
2957 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
2959 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2962 /* SQPOLL thread does its own polling */
2963 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
2964 (ctx->sq_data && ctx->sq_data->thread == current)) {
2965 while (!wq_list_empty(&ctx->iopoll_list)) {
2966 io_iopoll_try_reap_events(ctx);
2972 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
2973 ret |= io_run_local_work(ctx) > 0;
2974 ret |= io_cancel_defer_files(ctx, task, cancel_all);
2975 mutex_lock(&ctx->uring_lock);
2976 ret |= io_poll_remove_all(ctx, task, cancel_all);
2977 mutex_unlock(&ctx->uring_lock);
2978 ret |= io_kill_timeouts(ctx, task, cancel_all);
2980 ret |= io_run_task_work() > 0;
2984 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
2987 return atomic_read(&tctx->inflight_tracked);
2988 return percpu_counter_sum(&tctx->inflight);
2992 * Find any io_uring ctx that this task has registered or done IO on, and cancel
2993 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
2995 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
2997 struct io_uring_task *tctx = current->io_uring;
2998 struct io_ring_ctx *ctx;
3002 WARN_ON_ONCE(sqd && sqd->thread != current);
3004 if (!current->io_uring)
3007 io_wq_exit_start(tctx->io_wq);
3009 atomic_inc(&tctx->in_idle);
3013 io_uring_drop_tctx_refs(current);
3014 /* read completions before cancelations */
3015 inflight = tctx_inflight(tctx, !cancel_all);
3020 struct io_tctx_node *node;
3021 unsigned long index;
3023 xa_for_each(&tctx->xa, index, node) {
3024 /* sqpoll task will cancel all its requests */
3025 if (node->ctx->sq_data)
3027 loop |= io_uring_try_cancel_requests(node->ctx,
3028 current, cancel_all);
3031 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3032 loop |= io_uring_try_cancel_requests(ctx,
3042 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3044 io_uring_drop_tctx_refs(current);
3047 * If we've seen completions, retry without waiting. This
3048 * avoids a race where a completion comes in before we did
3049 * prepare_to_wait().
3051 if (inflight == tctx_inflight(tctx, !cancel_all))
3053 finish_wait(&tctx->wait, &wait);
3056 io_uring_clean_tctx(tctx);
3059 * We shouldn't run task_works after cancel, so just leave
3060 * ->in_idle set for normal exit.
3062 atomic_dec(&tctx->in_idle);
3063 /* for exec all current's requests should be gone, kill tctx */
3064 __io_uring_free(current);
3068 void __io_uring_cancel(bool cancel_all)
3070 io_uring_cancel_generic(cancel_all, NULL);
3073 static void *io_uring_validate_mmap_request(struct file *file,
3074 loff_t pgoff, size_t sz)
3076 struct io_ring_ctx *ctx = file->private_data;
3077 loff_t offset = pgoff << PAGE_SHIFT;
3082 case IORING_OFF_SQ_RING:
3083 case IORING_OFF_CQ_RING:
3086 case IORING_OFF_SQES:
3090 return ERR_PTR(-EINVAL);
3093 page = virt_to_head_page(ptr);
3094 if (sz > page_size(page))
3095 return ERR_PTR(-EINVAL);
3102 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3104 size_t sz = vma->vm_end - vma->vm_start;
3108 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3110 return PTR_ERR(ptr);
3112 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3113 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3116 static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp,
3117 unsigned long addr, unsigned long len,
3118 unsigned long pgoff, unsigned long flags)
3123 * Do not allow to map to user-provided address to avoid breaking the
3124 * aliasing rules. Userspace is not able to guess the offset address of
3125 * kernel kmalloc()ed memory area.
3130 ptr = io_uring_validate_mmap_request(filp, pgoff, len);
3135 * Some architectures have strong cache aliasing requirements.
3136 * For such architectures we need a coherent mapping which aliases
3137 * kernel memory *and* userspace memory. To achieve that:
3138 * - use a NULL file pointer to reference physical memory, and
3139 * - use the kernel virtual address of the shared io_uring context
3140 * (instead of the userspace-provided address, which has to be 0UL
3142 * - use the same pgoff which the get_unmapped_area() uses to
3143 * calculate the page colouring.
3144 * For architectures without such aliasing requirements, the
3145 * architecture will return any suitable mapping because addr is 0.
3148 flags |= MAP_SHARED;
3149 pgoff = 0; /* has been translated to ptr above */
3151 addr = (uintptr_t) ptr;
3152 pgoff = addr >> PAGE_SHIFT;
3156 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
3159 #else /* !CONFIG_MMU */
3161 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3163 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
3166 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3168 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3171 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3172 unsigned long addr, unsigned long len,
3173 unsigned long pgoff, unsigned long flags)
3177 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3179 return PTR_ERR(ptr);
3181 return (unsigned long) ptr;
3184 #endif /* !CONFIG_MMU */
3186 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3188 if (flags & IORING_ENTER_EXT_ARG) {
3189 struct io_uring_getevents_arg arg;
3191 if (argsz != sizeof(arg))
3193 if (copy_from_user(&arg, argp, sizeof(arg)))
3199 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3200 struct __kernel_timespec __user **ts,
3201 const sigset_t __user **sig)
3203 struct io_uring_getevents_arg arg;
3206 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3207 * is just a pointer to the sigset_t.
3209 if (!(flags & IORING_ENTER_EXT_ARG)) {
3210 *sig = (const sigset_t __user *) argp;
3216 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3217 * timespec and sigset_t pointers if good.
3219 if (*argsz != sizeof(arg))
3221 if (copy_from_user(&arg, argp, sizeof(arg)))
3225 *sig = u64_to_user_ptr(arg.sigmask);
3226 *argsz = arg.sigmask_sz;
3227 *ts = u64_to_user_ptr(arg.ts);
3231 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3232 u32, min_complete, u32, flags, const void __user *, argp,
3235 struct io_ring_ctx *ctx;
3239 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3240 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3241 IORING_ENTER_REGISTERED_RING)))
3245 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3246 * need only dereference our task private array to find it.
3248 if (flags & IORING_ENTER_REGISTERED_RING) {
3249 struct io_uring_task *tctx = current->io_uring;
3251 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3253 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3254 f.file = tctx->registered_rings[fd];
3256 if (unlikely(!f.file))
3260 if (unlikely(!f.file))
3263 if (unlikely(!io_is_uring_fops(f.file)))
3267 ctx = f.file->private_data;
3269 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3273 * For SQ polling, the thread will do all submissions and completions.
3274 * Just return the requested submit count, and wake the thread if
3278 if (ctx->flags & IORING_SETUP_SQPOLL) {
3279 io_cqring_overflow_flush(ctx);
3281 if (unlikely(ctx->sq_data->thread == NULL)) {
3285 if (flags & IORING_ENTER_SQ_WAKEUP)
3286 wake_up(&ctx->sq_data->wait);
3287 if (flags & IORING_ENTER_SQ_WAIT) {
3288 ret = io_sqpoll_wait_sq(ctx);
3293 } else if (to_submit) {
3294 ret = io_uring_add_tctx_node(ctx);
3298 mutex_lock(&ctx->uring_lock);
3299 ret = io_submit_sqes(ctx, to_submit);
3300 if (ret != to_submit) {
3301 mutex_unlock(&ctx->uring_lock);
3304 if (flags & IORING_ENTER_GETEVENTS) {
3305 if (ctx->syscall_iopoll)
3308 * Ignore errors, we'll soon call io_cqring_wait() and
3309 * it should handle ownership problems if any.
3311 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3312 (void)io_run_local_work_locked(ctx);
3314 mutex_unlock(&ctx->uring_lock);
3317 if (flags & IORING_ENTER_GETEVENTS) {
3320 if (ctx->syscall_iopoll) {
3322 * We disallow the app entering submit/complete with
3323 * polling, but we still need to lock the ring to
3324 * prevent racing with polled issue that got punted to
3327 mutex_lock(&ctx->uring_lock);
3329 ret2 = io_validate_ext_arg(flags, argp, argsz);
3330 if (likely(!ret2)) {
3331 min_complete = min(min_complete,
3333 ret2 = io_iopoll_check(ctx, min_complete);
3335 mutex_unlock(&ctx->uring_lock);
3337 const sigset_t __user *sig;
3338 struct __kernel_timespec __user *ts;
3340 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3341 if (likely(!ret2)) {
3342 min_complete = min(min_complete,
3344 ret2 = io_cqring_wait(ctx, min_complete, sig,
3353 * EBADR indicates that one or more CQE were dropped.
3354 * Once the user has been informed we can clear the bit
3355 * as they are obviously ok with those drops.
3357 if (unlikely(ret2 == -EBADR))
3358 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3367 static const struct file_operations io_uring_fops = {
3368 .release = io_uring_release,
3369 .mmap = io_uring_mmap,
3371 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3372 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3374 .get_unmapped_area = io_uring_mmu_get_unmapped_area,
3376 .poll = io_uring_poll,
3377 #ifdef CONFIG_PROC_FS
3378 .show_fdinfo = io_uring_show_fdinfo,
3382 bool io_is_uring_fops(struct file *file)
3384 return file->f_op == &io_uring_fops;
3387 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3388 struct io_uring_params *p)
3390 struct io_rings *rings;
3391 size_t size, sq_array_offset;
3393 /* make sure these are sane, as we already accounted them */
3394 ctx->sq_entries = p->sq_entries;
3395 ctx->cq_entries = p->cq_entries;
3397 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3398 if (size == SIZE_MAX)
3401 rings = io_mem_alloc(size);
3406 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3407 rings->sq_ring_mask = p->sq_entries - 1;
3408 rings->cq_ring_mask = p->cq_entries - 1;
3409 rings->sq_ring_entries = p->sq_entries;
3410 rings->cq_ring_entries = p->cq_entries;
3412 if (p->flags & IORING_SETUP_SQE128)
3413 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3415 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3416 if (size == SIZE_MAX) {
3417 io_mem_free(ctx->rings);
3422 ctx->sq_sqes = io_mem_alloc(size);
3423 if (!ctx->sq_sqes) {
3424 io_mem_free(ctx->rings);
3432 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3436 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3440 ret = __io_uring_add_tctx_node(ctx);
3445 fd_install(fd, file);
3450 * Allocate an anonymous fd, this is what constitutes the application
3451 * visible backing of an io_uring instance. The application mmaps this
3452 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3453 * we have to tie this fd to a socket for file garbage collection purposes.
3455 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3458 #if defined(CONFIG_UNIX)
3461 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3464 return ERR_PTR(ret);
3467 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3468 O_RDWR | O_CLOEXEC, NULL);
3469 #if defined(CONFIG_UNIX)
3471 sock_release(ctx->ring_sock);
3472 ctx->ring_sock = NULL;
3474 ctx->ring_sock->file = file;
3480 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3481 struct io_uring_params __user *params)
3483 struct io_ring_ctx *ctx;
3489 if (entries > IORING_MAX_ENTRIES) {
3490 if (!(p->flags & IORING_SETUP_CLAMP))
3492 entries = IORING_MAX_ENTRIES;
3496 * Use twice as many entries for the CQ ring. It's possible for the
3497 * application to drive a higher depth than the size of the SQ ring,
3498 * since the sqes are only used at submission time. This allows for
3499 * some flexibility in overcommitting a bit. If the application has
3500 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3501 * of CQ ring entries manually.
3503 p->sq_entries = roundup_pow_of_two(entries);
3504 if (p->flags & IORING_SETUP_CQSIZE) {
3506 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3507 * to a power-of-two, if it isn't already. We do NOT impose
3508 * any cq vs sq ring sizing.
3512 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3513 if (!(p->flags & IORING_SETUP_CLAMP))
3515 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3517 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3518 if (p->cq_entries < p->sq_entries)
3521 p->cq_entries = 2 * p->sq_entries;
3524 ctx = io_ring_ctx_alloc(p);
3529 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3530 * space applications don't need to do io completion events
3531 * polling again, they can rely on io_sq_thread to do polling
3532 * work, which can reduce cpu usage and uring_lock contention.
3534 if (ctx->flags & IORING_SETUP_IOPOLL &&
3535 !(ctx->flags & IORING_SETUP_SQPOLL))
3536 ctx->syscall_iopoll = 1;
3538 ctx->compat = in_compat_syscall();
3539 if (!ns_capable_noaudit(&init_user_ns, CAP_IPC_LOCK))
3540 ctx->user = get_uid(current_user());
3543 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3544 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3547 if (ctx->flags & IORING_SETUP_SQPOLL) {
3548 /* IPI related flags don't make sense with SQPOLL */
3549 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3550 IORING_SETUP_TASKRUN_FLAG |
3551 IORING_SETUP_DEFER_TASKRUN))
3553 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3554 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3555 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3557 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3558 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3560 ctx->notify_method = TWA_SIGNAL;
3564 * For DEFER_TASKRUN we require the completion task to be the same as the
3565 * submission task. This implies that there is only one submitter, so enforce
3568 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3569 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3574 * This is just grabbed for accounting purposes. When a process exits,
3575 * the mm is exited and dropped before the files, hence we need to hang
3576 * on to this mm purely for the purposes of being able to unaccount
3577 * memory (locked/pinned vm). It's not used for anything else.
3579 mmgrab(current->mm);
3580 ctx->mm_account = current->mm;
3582 ret = io_allocate_scq_urings(ctx, p);
3586 ret = io_sq_offload_create(ctx, p);
3589 /* always set a rsrc node */
3590 ret = io_rsrc_node_switch_start(ctx);
3593 io_rsrc_node_switch(ctx, NULL);
3595 memset(&p->sq_off, 0, sizeof(p->sq_off));
3596 p->sq_off.head = offsetof(struct io_rings, sq.head);
3597 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3598 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3599 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3600 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3601 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3602 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3604 memset(&p->cq_off, 0, sizeof(p->cq_off));
3605 p->cq_off.head = offsetof(struct io_rings, cq.head);
3606 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3607 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3608 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3609 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3610 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3611 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3613 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3614 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3615 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3616 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3617 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3618 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3619 IORING_FEAT_LINKED_FILE;
3621 if (copy_to_user(params, p, sizeof(*p))) {
3626 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
3627 && !(ctx->flags & IORING_SETUP_R_DISABLED))
3628 ctx->submitter_task = get_task_struct(current);
3630 file = io_uring_get_file(ctx);
3632 ret = PTR_ERR(file);
3637 * Install ring fd as the very last thing, so we don't risk someone
3638 * having closed it before we finish setup
3640 ret = io_uring_install_fd(ctx, file);
3642 /* fput will clean it up */
3647 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3650 io_ring_ctx_wait_and_kill(ctx);
3655 * Sets up an aio uring context, and returns the fd. Applications asks for a
3656 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3657 * params structure passed in.
3659 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3661 struct io_uring_params p;
3664 if (copy_from_user(&p, params, sizeof(p)))
3666 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3671 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3672 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3673 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3674 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3675 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3676 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3677 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN))
3680 return io_uring_create(entries, &p, params);
3683 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3684 struct io_uring_params __user *, params)
3686 return io_uring_setup(entries, params);
3689 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3692 struct io_uring_probe *p;
3696 size = struct_size(p, ops, nr_args);
3697 if (size == SIZE_MAX)
3699 p = kzalloc(size, GFP_KERNEL);
3704 if (copy_from_user(p, arg, size))
3707 if (memchr_inv(p, 0, size))
3710 p->last_op = IORING_OP_LAST - 1;
3711 if (nr_args > IORING_OP_LAST)
3712 nr_args = IORING_OP_LAST;
3714 for (i = 0; i < nr_args; i++) {
3716 if (!io_op_defs[i].not_supported)
3717 p->ops[i].flags = IO_URING_OP_SUPPORTED;
3722 if (copy_to_user(arg, p, size))
3729 static int io_register_personality(struct io_ring_ctx *ctx)
3731 const struct cred *creds;
3735 creds = get_current_cred();
3737 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
3738 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
3746 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
3747 void __user *arg, unsigned int nr_args)
3749 struct io_uring_restriction *res;
3753 /* Restrictions allowed only if rings started disabled */
3754 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3757 /* We allow only a single restrictions registration */
3758 if (ctx->restrictions.registered)
3761 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
3764 size = array_size(nr_args, sizeof(*res));
3765 if (size == SIZE_MAX)
3768 res = memdup_user(arg, size);
3770 return PTR_ERR(res);
3774 for (i = 0; i < nr_args; i++) {
3775 switch (res[i].opcode) {
3776 case IORING_RESTRICTION_REGISTER_OP:
3777 if (res[i].register_op >= IORING_REGISTER_LAST) {
3782 __set_bit(res[i].register_op,
3783 ctx->restrictions.register_op);
3785 case IORING_RESTRICTION_SQE_OP:
3786 if (res[i].sqe_op >= IORING_OP_LAST) {
3791 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
3793 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
3794 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
3796 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
3797 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
3806 /* Reset all restrictions if an error happened */
3808 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
3810 ctx->restrictions.registered = true;
3816 static int io_register_enable_rings(struct io_ring_ctx *ctx)
3818 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3821 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task)
3822 ctx->submitter_task = get_task_struct(current);
3824 if (ctx->restrictions.registered)
3825 ctx->restricted = 1;
3827 ctx->flags &= ~IORING_SETUP_R_DISABLED;
3828 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
3829 wake_up(&ctx->sq_data->wait);
3833 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
3834 void __user *arg, unsigned len)
3836 struct io_uring_task *tctx = current->io_uring;
3837 cpumask_var_t new_mask;
3840 if (!tctx || !tctx->io_wq)
3843 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3846 cpumask_clear(new_mask);
3847 if (len > cpumask_size())
3848 len = cpumask_size();
3850 if (in_compat_syscall()) {
3851 ret = compat_get_bitmap(cpumask_bits(new_mask),
3852 (const compat_ulong_t __user *)arg,
3853 len * 8 /* CHAR_BIT */);
3855 ret = copy_from_user(new_mask, arg, len);
3859 free_cpumask_var(new_mask);
3863 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
3864 free_cpumask_var(new_mask);
3868 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
3870 struct io_uring_task *tctx = current->io_uring;
3872 if (!tctx || !tctx->io_wq)
3875 return io_wq_cpu_affinity(tctx->io_wq, NULL);
3878 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
3880 __must_hold(&ctx->uring_lock)
3882 struct io_tctx_node *node;
3883 struct io_uring_task *tctx = NULL;
3884 struct io_sq_data *sqd = NULL;
3888 if (copy_from_user(new_count, arg, sizeof(new_count)))
3890 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3891 if (new_count[i] > INT_MAX)
3894 if (ctx->flags & IORING_SETUP_SQPOLL) {
3898 * Observe the correct sqd->lock -> ctx->uring_lock
3899 * ordering. Fine to drop uring_lock here, we hold
3902 refcount_inc(&sqd->refs);
3903 mutex_unlock(&ctx->uring_lock);
3904 mutex_lock(&sqd->lock);
3905 mutex_lock(&ctx->uring_lock);
3907 tctx = sqd->thread->io_uring;
3910 tctx = current->io_uring;
3913 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
3915 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3917 ctx->iowq_limits[i] = new_count[i];
3918 ctx->iowq_limits_set = true;
3920 if (tctx && tctx->io_wq) {
3921 ret = io_wq_max_workers(tctx->io_wq, new_count);
3925 memset(new_count, 0, sizeof(new_count));
3929 mutex_unlock(&sqd->lock);
3930 io_put_sq_data(sqd);
3933 if (copy_to_user(arg, new_count, sizeof(new_count)))
3936 /* that's it for SQPOLL, only the SQPOLL task creates requests */
3940 /* now propagate the restriction to all registered users */
3941 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3942 struct io_uring_task *tctx = node->task->io_uring;
3944 if (WARN_ON_ONCE(!tctx->io_wq))
3947 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3948 new_count[i] = ctx->iowq_limits[i];
3949 /* ignore errors, it always returns zero anyway */
3950 (void)io_wq_max_workers(tctx->io_wq, new_count);
3955 mutex_unlock(&sqd->lock);
3956 io_put_sq_data(sqd);
3961 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
3962 void __user *arg, unsigned nr_args)
3963 __releases(ctx->uring_lock)
3964 __acquires(ctx->uring_lock)
3969 * We don't quiesce the refs for register anymore and so it can't be
3970 * dying as we're holding a file ref here.
3972 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
3975 if (ctx->submitter_task && ctx->submitter_task != current)
3978 if (ctx->restricted) {
3979 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
3980 if (!test_bit(opcode, ctx->restrictions.register_op))
3985 case IORING_REGISTER_BUFFERS:
3989 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
3991 case IORING_UNREGISTER_BUFFERS:
3995 ret = io_sqe_buffers_unregister(ctx);
3997 case IORING_REGISTER_FILES:
4001 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
4003 case IORING_UNREGISTER_FILES:
4007 ret = io_sqe_files_unregister(ctx);
4009 case IORING_REGISTER_FILES_UPDATE:
4010 ret = io_register_files_update(ctx, arg, nr_args);
4012 case IORING_REGISTER_EVENTFD:
4016 ret = io_eventfd_register(ctx, arg, 0);
4018 case IORING_REGISTER_EVENTFD_ASYNC:
4022 ret = io_eventfd_register(ctx, arg, 1);
4024 case IORING_UNREGISTER_EVENTFD:
4028 ret = io_eventfd_unregister(ctx);
4030 case IORING_REGISTER_PROBE:
4032 if (!arg || nr_args > 256)
4034 ret = io_probe(ctx, arg, nr_args);
4036 case IORING_REGISTER_PERSONALITY:
4040 ret = io_register_personality(ctx);
4042 case IORING_UNREGISTER_PERSONALITY:
4046 ret = io_unregister_personality(ctx, nr_args);
4048 case IORING_REGISTER_ENABLE_RINGS:
4052 ret = io_register_enable_rings(ctx);
4054 case IORING_REGISTER_RESTRICTIONS:
4055 ret = io_register_restrictions(ctx, arg, nr_args);
4057 case IORING_REGISTER_FILES2:
4058 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
4060 case IORING_REGISTER_FILES_UPDATE2:
4061 ret = io_register_rsrc_update(ctx, arg, nr_args,
4064 case IORING_REGISTER_BUFFERS2:
4065 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
4067 case IORING_REGISTER_BUFFERS_UPDATE:
4068 ret = io_register_rsrc_update(ctx, arg, nr_args,
4069 IORING_RSRC_BUFFER);
4071 case IORING_REGISTER_IOWQ_AFF:
4073 if (!arg || !nr_args)
4075 ret = io_register_iowq_aff(ctx, arg, nr_args);
4077 case IORING_UNREGISTER_IOWQ_AFF:
4081 ret = io_unregister_iowq_aff(ctx);
4083 case IORING_REGISTER_IOWQ_MAX_WORKERS:
4085 if (!arg || nr_args != 2)
4087 ret = io_register_iowq_max_workers(ctx, arg);
4089 case IORING_REGISTER_RING_FDS:
4090 ret = io_ringfd_register(ctx, arg, nr_args);
4092 case IORING_UNREGISTER_RING_FDS:
4093 ret = io_ringfd_unregister(ctx, arg, nr_args);
4095 case IORING_REGISTER_PBUF_RING:
4097 if (!arg || nr_args != 1)
4099 ret = io_register_pbuf_ring(ctx, arg);
4101 case IORING_UNREGISTER_PBUF_RING:
4103 if (!arg || nr_args != 1)
4105 ret = io_unregister_pbuf_ring(ctx, arg);
4107 case IORING_REGISTER_SYNC_CANCEL:
4109 if (!arg || nr_args != 1)
4111 ret = io_sync_cancel(ctx, arg);
4113 case IORING_REGISTER_FILE_ALLOC_RANGE:
4115 if (!arg || nr_args)
4117 ret = io_register_file_alloc_range(ctx, arg);
4127 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
4128 void __user *, arg, unsigned int, nr_args)
4130 struct io_ring_ctx *ctx;
4134 if (opcode >= IORING_REGISTER_LAST)
4142 if (!io_is_uring_fops(f.file))
4145 ctx = f.file->private_data;
4147 io_run_task_work_ctx(ctx);
4149 mutex_lock(&ctx->uring_lock);
4150 ret = __io_uring_register(ctx, opcode, arg, nr_args);
4151 mutex_unlock(&ctx->uring_lock);
4152 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
4158 static int __init io_uring_init(void)
4160 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4161 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4162 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4165 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4166 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4167 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4168 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4169 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4170 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
4171 BUILD_BUG_SQE_ELEM(1, __u8, flags);
4172 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
4173 BUILD_BUG_SQE_ELEM(4, __s32, fd);
4174 BUILD_BUG_SQE_ELEM(8, __u64, off);
4175 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
4176 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
4177 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
4178 BUILD_BUG_SQE_ELEM(16, __u64, addr);
4179 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
4180 BUILD_BUG_SQE_ELEM(24, __u32, len);
4181 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
4182 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
4183 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4184 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
4185 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
4186 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
4187 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
4188 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
4189 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
4190 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
4191 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
4192 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
4193 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
4194 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
4195 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
4196 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
4197 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
4198 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
4199 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
4200 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
4201 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
4202 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
4203 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
4204 BUILD_BUG_SQE_ELEM(42, __u16, personality);
4205 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
4206 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
4207 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
4208 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
4209 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
4210 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4211 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
4213 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4214 sizeof(struct io_uring_rsrc_update));
4215 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4216 sizeof(struct io_uring_rsrc_update2));
4218 /* ->buf_index is u16 */
4219 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4220 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4221 offsetof(struct io_uring_buf_ring, tail));
4223 /* should fit into one byte */
4224 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
4225 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4226 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4228 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
4230 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4232 io_uring_optable_init();
4234 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4238 __initcall(io_uring_init);