1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <net/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
49 #include <linux/bits.h>
51 #include <linux/sched/signal.h>
53 #include <linux/file.h>
54 #include <linux/fdtable.h>
56 #include <linux/mman.h>
57 #include <linux/percpu.h>
58 #include <linux/slab.h>
59 #include <linux/bvec.h>
60 #include <linux/net.h>
62 #include <net/af_unix.h>
64 #include <linux/anon_inodes.h>
65 #include <linux/sched/mm.h>
66 #include <linux/uaccess.h>
67 #include <linux/nospec.h>
68 #include <linux/highmem.h>
69 #include <linux/fsnotify.h>
70 #include <linux/fadvise.h>
71 #include <linux/task_work.h>
72 #include <linux/io_uring.h>
73 #include <linux/audit.h>
74 #include <linux/security.h>
76 #define CREATE_TRACE_POINTS
77 #include <trace/events/io_uring.h>
79 #include <uapi/linux/io_uring.h>
97 #include "alloc_cache.h"
99 #define IORING_MAX_ENTRIES 32768
100 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
102 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
103 IORING_REGISTER_LAST + IORING_OP_LAST)
105 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
106 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
108 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
109 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
111 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
112 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
115 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
118 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
120 #define IO_COMPL_BATCH 32
121 #define IO_REQ_ALLOC_BATCH 8
124 IO_CHECK_CQ_OVERFLOW_BIT,
125 IO_CHECK_CQ_DROPPED_BIT,
129 IO_EVENTFD_OP_SIGNAL_BIT,
130 IO_EVENTFD_OP_FREE_BIT,
133 struct io_defer_entry {
134 struct list_head list;
135 struct io_kiocb *req;
139 /* requests with any of those set should undergo io_disarm_next() */
140 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
141 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
143 static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
144 struct task_struct *task,
147 static void io_dismantle_req(struct io_kiocb *req);
148 static void io_clean_op(struct io_kiocb *req);
149 static void io_queue_sqe(struct io_kiocb *req);
150 static void io_move_task_work_from_local(struct io_ring_ctx *ctx);
151 static void __io_submit_flush_completions(struct io_ring_ctx *ctx);
153 static struct kmem_cache *req_cachep;
155 struct sock *io_uring_get_socket(struct file *file)
157 #if defined(CONFIG_UNIX)
158 if (io_is_uring_fops(file)) {
159 struct io_ring_ctx *ctx = file->private_data;
161 return ctx->ring_sock->sk;
166 EXPORT_SYMBOL(io_uring_get_socket);
168 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
170 if (!wq_list_empty(&ctx->submit_state.compl_reqs))
171 __io_submit_flush_completions(ctx);
174 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
176 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
179 static bool io_match_linked(struct io_kiocb *head)
181 struct io_kiocb *req;
183 io_for_each_link(req, head) {
184 if (req->flags & REQ_F_INFLIGHT)
191 * As io_match_task() but protected against racing with linked timeouts.
192 * User must not hold timeout_lock.
194 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
199 if (task && head->task != task)
204 if (head->flags & REQ_F_LINK_TIMEOUT) {
205 struct io_ring_ctx *ctx = head->ctx;
207 /* protect against races with linked timeouts */
208 spin_lock_irq(&ctx->timeout_lock);
209 matched = io_match_linked(head);
210 spin_unlock_irq(&ctx->timeout_lock);
212 matched = io_match_linked(head);
217 static inline void req_fail_link_node(struct io_kiocb *req, int res)
220 io_req_set_res(req, res, 0);
223 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
225 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
228 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
230 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
232 complete(&ctx->ref_comp);
235 static __cold void io_fallback_req_func(struct work_struct *work)
237 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
239 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
240 struct io_kiocb *req, *tmp;
243 percpu_ref_get(&ctx->refs);
244 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
245 req->io_task_work.func(req, &locked);
248 io_submit_flush_completions(ctx);
249 mutex_unlock(&ctx->uring_lock);
251 percpu_ref_put(&ctx->refs);
254 static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
256 unsigned hash_buckets = 1U << bits;
257 size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
259 table->hbs = kmalloc(hash_size, GFP_KERNEL);
263 table->hash_bits = bits;
264 init_hash_table(table, hash_buckets);
268 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
270 struct io_ring_ctx *ctx;
273 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
277 xa_init(&ctx->io_bl_xa);
280 * Use 5 bits less than the max cq entries, that should give us around
281 * 32 entries per hash list if totally full and uniformly spread, but
282 * don't keep too many buckets to not overconsume memory.
284 hash_bits = ilog2(p->cq_entries) - 5;
285 hash_bits = clamp(hash_bits, 1, 8);
286 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
288 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
291 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
292 if (!ctx->dummy_ubuf)
294 /* set invalid range, so io_import_fixed() fails meeting it */
295 ctx->dummy_ubuf->ubuf = -1UL;
297 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
301 ctx->flags = p->flags;
302 init_waitqueue_head(&ctx->sqo_sq_wait);
303 INIT_LIST_HEAD(&ctx->sqd_list);
304 INIT_LIST_HEAD(&ctx->cq_overflow_list);
305 INIT_LIST_HEAD(&ctx->io_buffers_cache);
306 io_alloc_cache_init(&ctx->apoll_cache);
307 io_alloc_cache_init(&ctx->netmsg_cache);
308 init_completion(&ctx->ref_comp);
309 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
310 mutex_init(&ctx->uring_lock);
311 init_waitqueue_head(&ctx->cq_wait);
312 spin_lock_init(&ctx->completion_lock);
313 spin_lock_init(&ctx->timeout_lock);
314 INIT_WQ_LIST(&ctx->iopoll_list);
315 INIT_LIST_HEAD(&ctx->io_buffers_pages);
316 INIT_LIST_HEAD(&ctx->io_buffers_comp);
317 INIT_LIST_HEAD(&ctx->defer_list);
318 INIT_LIST_HEAD(&ctx->timeout_list);
319 INIT_LIST_HEAD(&ctx->ltimeout_list);
320 spin_lock_init(&ctx->rsrc_ref_lock);
321 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
322 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
323 init_llist_head(&ctx->rsrc_put_llist);
324 init_llist_head(&ctx->work_llist);
325 INIT_LIST_HEAD(&ctx->tctx_list);
326 ctx->submit_state.free_list.next = NULL;
327 INIT_WQ_LIST(&ctx->locked_free_list);
328 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
329 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
332 kfree(ctx->dummy_ubuf);
333 kfree(ctx->cancel_table.hbs);
334 kfree(ctx->cancel_table_locked.hbs);
336 xa_destroy(&ctx->io_bl_xa);
341 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
343 struct io_rings *r = ctx->rings;
345 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
349 static bool req_need_defer(struct io_kiocb *req, u32 seq)
351 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
352 struct io_ring_ctx *ctx = req->ctx;
354 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
360 static inline void io_req_track_inflight(struct io_kiocb *req)
362 if (!(req->flags & REQ_F_INFLIGHT)) {
363 req->flags |= REQ_F_INFLIGHT;
364 atomic_inc(&req->task->io_uring->inflight_tracked);
368 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
370 if (WARN_ON_ONCE(!req->link))
373 req->flags &= ~REQ_F_ARM_LTIMEOUT;
374 req->flags |= REQ_F_LINK_TIMEOUT;
376 /* linked timeouts should have two refs once prep'ed */
377 io_req_set_refcount(req);
378 __io_req_set_refcount(req->link, 2);
382 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
384 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
386 return __io_prep_linked_timeout(req);
389 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
391 io_queue_linked_timeout(__io_prep_linked_timeout(req));
394 static inline void io_arm_ltimeout(struct io_kiocb *req)
396 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
397 __io_arm_ltimeout(req);
400 static void io_prep_async_work(struct io_kiocb *req)
402 const struct io_op_def *def = &io_op_defs[req->opcode];
403 struct io_ring_ctx *ctx = req->ctx;
405 if (!(req->flags & REQ_F_CREDS)) {
406 req->flags |= REQ_F_CREDS;
407 req->creds = get_current_cred();
410 req->work.list.next = NULL;
412 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
413 if (req->flags & REQ_F_FORCE_ASYNC)
414 req->work.flags |= IO_WQ_WORK_CONCURRENT;
416 if (req->file && !io_req_ffs_set(req))
417 req->flags |= io_file_get_flags(req->file) << REQ_F_SUPPORT_NOWAIT_BIT;
419 if (req->flags & REQ_F_ISREG) {
420 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
421 io_wq_hash_work(&req->work, file_inode(req->file));
422 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
423 if (def->unbound_nonreg_file)
424 req->work.flags |= IO_WQ_WORK_UNBOUND;
428 static void io_prep_async_link(struct io_kiocb *req)
430 struct io_kiocb *cur;
432 if (req->flags & REQ_F_LINK_TIMEOUT) {
433 struct io_ring_ctx *ctx = req->ctx;
435 spin_lock_irq(&ctx->timeout_lock);
436 io_for_each_link(cur, req)
437 io_prep_async_work(cur);
438 spin_unlock_irq(&ctx->timeout_lock);
440 io_for_each_link(cur, req)
441 io_prep_async_work(cur);
445 void io_queue_iowq(struct io_kiocb *req, bool *dont_use)
447 struct io_kiocb *link = io_prep_linked_timeout(req);
448 struct io_uring_task *tctx = req->task->io_uring;
451 BUG_ON(!tctx->io_wq);
453 /* init ->work of the whole link before punting */
454 io_prep_async_link(req);
457 * Not expected to happen, but if we do have a bug where this _can_
458 * happen, catch it here and ensure the request is marked as
459 * canceled. That will make io-wq go through the usual work cancel
460 * procedure rather than attempt to run this request (or create a new
463 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
464 req->work.flags |= IO_WQ_WORK_CANCEL;
466 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
467 io_wq_enqueue(tctx->io_wq, &req->work);
469 io_queue_linked_timeout(link);
472 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
474 while (!list_empty(&ctx->defer_list)) {
475 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
476 struct io_defer_entry, list);
478 if (req_need_defer(de->req, de->seq))
480 list_del_init(&de->list);
481 io_req_task_queue(de->req);
487 static void io_eventfd_ops(struct rcu_head *rcu)
489 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
490 int ops = atomic_xchg(&ev_fd->ops, 0);
492 if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT))
493 eventfd_signal(ev_fd->cq_ev_fd, 1);
495 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
496 * ordering in a race but if references are 0 we know we have to free
499 if (atomic_dec_and_test(&ev_fd->refs)) {
500 eventfd_ctx_put(ev_fd->cq_ev_fd);
505 static void io_eventfd_signal(struct io_ring_ctx *ctx)
507 struct io_ev_fd *ev_fd = NULL;
511 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
514 ev_fd = rcu_dereference(ctx->io_ev_fd);
517 * Check again if ev_fd exists incase an io_eventfd_unregister call
518 * completed between the NULL check of ctx->io_ev_fd at the start of
519 * the function and rcu_read_lock.
521 if (unlikely(!ev_fd))
523 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
525 if (ev_fd->eventfd_async && !io_wq_current_is_worker())
528 if (likely(eventfd_signal_allowed())) {
529 eventfd_signal(ev_fd->cq_ev_fd, 1);
531 atomic_inc(&ev_fd->refs);
532 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops))
533 call_rcu(&ev_fd->rcu, io_eventfd_ops);
535 atomic_dec(&ev_fd->refs);
542 static void io_eventfd_flush_signal(struct io_ring_ctx *ctx)
546 spin_lock(&ctx->completion_lock);
549 * Eventfd should only get triggered when at least one event has been
550 * posted. Some applications rely on the eventfd notification count
551 * only changing IFF a new CQE has been added to the CQ ring. There's
552 * no depedency on 1:1 relationship between how many times this
553 * function is called (and hence the eventfd count) and number of CQEs
554 * posted to the CQ ring.
556 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
557 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
558 spin_unlock(&ctx->completion_lock);
562 io_eventfd_signal(ctx);
565 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
567 if (ctx->off_timeout_used || ctx->drain_active) {
568 spin_lock(&ctx->completion_lock);
569 if (ctx->off_timeout_used)
570 io_flush_timeouts(ctx);
571 if (ctx->drain_active)
572 io_queue_deferred(ctx);
573 spin_unlock(&ctx->completion_lock);
576 io_eventfd_flush_signal(ctx);
579 static inline void io_cqring_ev_posted(struct io_ring_ctx *ctx)
581 io_commit_cqring_flush(ctx);
585 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
586 __releases(ctx->completion_lock)
588 io_commit_cqring(ctx);
589 spin_unlock(&ctx->completion_lock);
590 io_cqring_ev_posted(ctx);
593 void io_cq_unlock_post(struct io_ring_ctx *ctx)
595 __io_cq_unlock_post(ctx);
598 /* Returns true if there are no backlogged entries after the flush */
599 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
602 size_t cqe_size = sizeof(struct io_uring_cqe);
604 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
607 if (ctx->flags & IORING_SETUP_CQE32)
611 while (!list_empty(&ctx->cq_overflow_list)) {
612 struct io_uring_cqe *cqe = io_get_cqe(ctx);
613 struct io_overflow_cqe *ocqe;
617 ocqe = list_first_entry(&ctx->cq_overflow_list,
618 struct io_overflow_cqe, list);
620 memcpy(cqe, &ocqe->cqe, cqe_size);
622 io_account_cq_overflow(ctx);
624 list_del(&ocqe->list);
628 all_flushed = list_empty(&ctx->cq_overflow_list);
630 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
631 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
634 io_cq_unlock_post(ctx);
638 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
642 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
643 /* iopoll syncs against uring_lock, not completion_lock */
644 if (ctx->flags & IORING_SETUP_IOPOLL)
645 mutex_lock(&ctx->uring_lock);
646 ret = __io_cqring_overflow_flush(ctx, false);
647 if (ctx->flags & IORING_SETUP_IOPOLL)
648 mutex_unlock(&ctx->uring_lock);
654 void __io_put_task(struct task_struct *task, int nr)
656 struct io_uring_task *tctx = task->io_uring;
658 percpu_counter_sub(&tctx->inflight, nr);
659 if (unlikely(atomic_read(&tctx->in_idle)))
660 wake_up(&tctx->wait);
661 put_task_struct_many(task, nr);
664 void io_task_refs_refill(struct io_uring_task *tctx)
666 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
668 percpu_counter_add(&tctx->inflight, refill);
669 refcount_add(refill, ¤t->usage);
670 tctx->cached_refs += refill;
673 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
675 struct io_uring_task *tctx = task->io_uring;
676 unsigned int refs = tctx->cached_refs;
679 tctx->cached_refs = 0;
680 percpu_counter_sub(&tctx->inflight, refs);
681 put_task_struct_many(task, refs);
685 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
686 s32 res, u32 cflags, u64 extra1, u64 extra2)
688 struct io_overflow_cqe *ocqe;
689 size_t ocq_size = sizeof(struct io_overflow_cqe);
690 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
693 ocq_size += sizeof(struct io_uring_cqe);
695 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
696 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
699 * If we're in ring overflow flush mode, or in task cancel mode,
700 * or cannot allocate an overflow entry, then we need to drop it
703 io_account_cq_overflow(ctx);
704 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
707 if (list_empty(&ctx->cq_overflow_list)) {
708 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
709 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
712 ocqe->cqe.user_data = user_data;
714 ocqe->cqe.flags = cflags;
716 ocqe->cqe.big_cqe[0] = extra1;
717 ocqe->cqe.big_cqe[1] = extra2;
719 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
723 bool io_req_cqe_overflow(struct io_kiocb *req)
725 if (!(req->flags & REQ_F_CQE32_INIT)) {
729 return io_cqring_event_overflow(req->ctx, req->cqe.user_data,
730 req->cqe.res, req->cqe.flags,
731 req->extra1, req->extra2);
735 * writes to the cq entry need to come after reading head; the
736 * control dependency is enough as we're using WRITE_ONCE to
739 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx)
741 struct io_rings *rings = ctx->rings;
742 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
743 unsigned int free, queued, len;
746 /* userspace may cheat modifying the tail, be safe and do min */
747 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
748 free = ctx->cq_entries - queued;
749 /* we need a contiguous range, limit based on the current array offset */
750 len = min(free, ctx->cq_entries - off);
754 if (ctx->flags & IORING_SETUP_CQE32) {
759 ctx->cqe_cached = &rings->cqes[off];
760 ctx->cqe_sentinel = ctx->cqe_cached + len;
762 ctx->cached_cq_tail++;
764 if (ctx->flags & IORING_SETUP_CQE32)
766 return &rings->cqes[off];
769 bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
772 struct io_uring_cqe *cqe;
777 * If we can't get a cq entry, userspace overflowed the
778 * submission (by quite a lot). Increment the overflow count in
781 cqe = io_get_cqe(ctx);
783 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
785 WRITE_ONCE(cqe->user_data, user_data);
786 WRITE_ONCE(cqe->res, res);
787 WRITE_ONCE(cqe->flags, cflags);
789 if (ctx->flags & IORING_SETUP_CQE32) {
790 WRITE_ONCE(cqe->big_cqe[0], 0);
791 WRITE_ONCE(cqe->big_cqe[1], 0);
797 return io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
802 bool io_post_aux_cqe(struct io_ring_ctx *ctx,
803 u64 user_data, s32 res, u32 cflags,
809 filled = io_fill_cqe_aux(ctx, user_data, res, cflags, allow_overflow);
810 io_cq_unlock_post(ctx);
814 static void __io_req_complete_put(struct io_kiocb *req)
817 * If we're the last reference to this request, add to our locked
820 if (req_ref_put_and_test(req)) {
821 struct io_ring_ctx *ctx = req->ctx;
823 if (req->flags & IO_REQ_LINK_FLAGS) {
824 if (req->flags & IO_DISARM_MASK)
827 io_req_task_queue(req->link);
831 io_req_put_rsrc(req);
833 * Selected buffer deallocation in io_clean_op() assumes that
834 * we don't hold ->completion_lock. Clean them here to avoid
837 io_put_kbuf_comp(req);
838 io_dismantle_req(req);
839 io_put_task(req->task, 1);
840 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
841 ctx->locked_free_nr++;
845 void __io_req_complete_post(struct io_kiocb *req)
847 if (!(req->flags & REQ_F_CQE_SKIP))
848 __io_fill_cqe_req(req->ctx, req);
849 __io_req_complete_put(req);
852 void io_req_complete_post(struct io_kiocb *req)
854 struct io_ring_ctx *ctx = req->ctx;
857 __io_req_complete_post(req);
858 io_cq_unlock_post(ctx);
861 inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags)
863 io_req_complete_post(req);
866 void io_req_complete_failed(struct io_kiocb *req, s32 res)
869 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
870 io_req_complete_post(req);
874 * Don't initialise the fields below on every allocation, but do that in
875 * advance and keep them valid across allocations.
877 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
881 req->async_data = NULL;
882 /* not necessary, but safer to zero */
886 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
887 struct io_submit_state *state)
889 spin_lock(&ctx->completion_lock);
890 wq_list_splice(&ctx->locked_free_list, &state->free_list);
891 ctx->locked_free_nr = 0;
892 spin_unlock(&ctx->completion_lock);
896 * A request might get retired back into the request caches even before opcode
897 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
898 * Because of that, io_alloc_req() should be called only under ->uring_lock
899 * and with extra caution to not get a request that is still worked on.
901 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
902 __must_hold(&ctx->uring_lock)
904 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
905 void *reqs[IO_REQ_ALLOC_BATCH];
909 * If we have more than a batch's worth of requests in our IRQ side
910 * locked cache, grab the lock and move them over to our submission
913 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
914 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
915 if (!io_req_cache_empty(ctx))
919 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
922 * Bulk alloc is all-or-nothing. If we fail to get a batch,
923 * retry single alloc to be on the safe side.
925 if (unlikely(ret <= 0)) {
926 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
932 percpu_ref_get_many(&ctx->refs, ret);
933 for (i = 0; i < ret; i++) {
934 struct io_kiocb *req = reqs[i];
936 io_preinit_req(req, ctx);
937 io_req_add_to_cache(req, ctx);
942 static inline void io_dismantle_req(struct io_kiocb *req)
944 unsigned int flags = req->flags;
946 if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
948 if (!(flags & REQ_F_FIXED_FILE))
949 io_put_file(req->file);
952 __cold void io_free_req(struct io_kiocb *req)
954 struct io_ring_ctx *ctx = req->ctx;
956 io_req_put_rsrc(req);
957 io_dismantle_req(req);
958 io_put_task(req->task, 1);
960 spin_lock(&ctx->completion_lock);
961 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
962 ctx->locked_free_nr++;
963 spin_unlock(&ctx->completion_lock);
966 static void __io_req_find_next_prep(struct io_kiocb *req)
968 struct io_ring_ctx *ctx = req->ctx;
972 io_cq_unlock_post(ctx);
975 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
977 struct io_kiocb *nxt;
980 * If LINK is set, we have dependent requests in this chain. If we
981 * didn't fail this request, queue the first one up, moving any other
982 * dependencies to the next request. In case of failure, fail the rest
985 if (unlikely(req->flags & IO_DISARM_MASK))
986 __io_req_find_next_prep(req);
992 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
996 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
997 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
999 io_submit_flush_completions(ctx);
1000 mutex_unlock(&ctx->uring_lock);
1003 percpu_ref_put(&ctx->refs);
1006 static unsigned int handle_tw_list(struct llist_node *node,
1007 struct io_ring_ctx **ctx, bool *locked,
1008 struct llist_node *last)
1010 unsigned int count = 0;
1012 while (node != last) {
1013 struct llist_node *next = node->next;
1014 struct io_kiocb *req = container_of(node, struct io_kiocb,
1017 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1019 if (req->ctx != *ctx) {
1020 ctx_flush_and_put(*ctx, locked);
1022 /* if not contended, grab and improve batching */
1023 *locked = mutex_trylock(&(*ctx)->uring_lock);
1024 percpu_ref_get(&(*ctx)->refs);
1026 req->io_task_work.func(req, locked);
1035 * io_llist_xchg - swap all entries in a lock-less list
1036 * @head: the head of lock-less list to delete all entries
1037 * @new: new entry as the head of the list
1039 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1040 * The order of entries returned is from the newest to the oldest added one.
1042 static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1043 struct llist_node *new)
1045 return xchg(&head->first, new);
1049 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1050 * @head: the head of lock-less list to delete all entries
1051 * @old: expected old value of the first entry of the list
1052 * @new: new entry as the head of the list
1054 * perform a cmpxchg on the first entry of the list.
1057 static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1058 struct llist_node *old,
1059 struct llist_node *new)
1061 return cmpxchg(&head->first, old, new);
1064 void tctx_task_work(struct callback_head *cb)
1066 bool uring_locked = false;
1067 struct io_ring_ctx *ctx = NULL;
1068 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1070 struct llist_node fake = {};
1071 struct llist_node *node = io_llist_xchg(&tctx->task_list, &fake);
1072 unsigned int loops = 1;
1073 unsigned int count = handle_tw_list(node, &ctx, &uring_locked, NULL);
1075 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1076 while (node != &fake) {
1078 node = io_llist_xchg(&tctx->task_list, &fake);
1079 count += handle_tw_list(node, &ctx, &uring_locked, &fake);
1080 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1083 ctx_flush_and_put(ctx, &uring_locked);
1085 /* relaxed read is enough as only the task itself sets ->in_idle */
1086 if (unlikely(atomic_read(&tctx->in_idle)))
1087 io_uring_drop_tctx_refs(current);
1089 trace_io_uring_task_work_run(tctx, count, loops);
1092 static void io_req_local_work_add(struct io_kiocb *req)
1094 struct io_ring_ctx *ctx = req->ctx;
1096 if (!llist_add(&req->io_task_work.node, &ctx->work_llist))
1099 if (unlikely(atomic_read(&req->task->io_uring->in_idle))) {
1100 io_move_task_work_from_local(ctx);
1104 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1105 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1108 io_eventfd_signal(ctx);
1109 io_cqring_wake(ctx);
1113 static inline void __io_req_task_work_add(struct io_kiocb *req, bool allow_local)
1115 struct io_uring_task *tctx = req->task->io_uring;
1116 struct io_ring_ctx *ctx = req->ctx;
1117 struct llist_node *node;
1119 if (allow_local && ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
1120 io_req_local_work_add(req);
1124 /* task_work already pending, we're done */
1125 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1128 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1129 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1131 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1134 node = llist_del_all(&tctx->task_list);
1137 req = container_of(node, struct io_kiocb, io_task_work.node);
1139 if (llist_add(&req->io_task_work.node,
1140 &req->ctx->fallback_llist))
1141 schedule_delayed_work(&req->ctx->fallback_work, 1);
1145 void io_req_task_work_add(struct io_kiocb *req)
1147 __io_req_task_work_add(req, true);
1150 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1152 struct llist_node *node;
1154 node = llist_del_all(&ctx->work_llist);
1156 struct io_kiocb *req = container_of(node, struct io_kiocb,
1160 __io_req_task_work_add(req, false);
1164 int __io_run_local_work(struct io_ring_ctx *ctx, bool locked)
1166 struct llist_node *node;
1167 struct llist_node fake;
1168 struct llist_node *current_final = NULL;
1170 unsigned int loops = 1;
1172 if (unlikely(ctx->submitter_task != current)) {
1173 /* maybe this is before any submissions */
1174 if (!ctx->submitter_task)
1180 node = io_llist_xchg(&ctx->work_llist, &fake);
1183 while (node != current_final) {
1184 struct llist_node *next = node->next;
1185 struct io_kiocb *req = container_of(node, struct io_kiocb,
1187 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1188 req->io_task_work.func(req, &locked);
1193 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1194 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1196 node = io_llist_cmpxchg(&ctx->work_llist, &fake, NULL);
1197 if (node != &fake) {
1199 current_final = &fake;
1200 node = io_llist_xchg(&ctx->work_llist, &fake);
1205 io_submit_flush_completions(ctx);
1206 trace_io_uring_local_work_run(ctx, ret, loops);
1211 int io_run_local_work(struct io_ring_ctx *ctx)
1216 locked = mutex_trylock(&ctx->uring_lock);
1217 ret = __io_run_local_work(ctx, locked);
1219 mutex_unlock(&ctx->uring_lock);
1224 static void io_req_tw_post(struct io_kiocb *req, bool *locked)
1226 io_req_complete_post(req);
1229 void io_req_tw_post_queue(struct io_kiocb *req, s32 res, u32 cflags)
1231 io_req_set_res(req, res, cflags);
1232 req->io_task_work.func = io_req_tw_post;
1233 io_req_task_work_add(req);
1236 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
1238 /* not needed for normal modes, but SQPOLL depends on it */
1239 io_tw_lock(req->ctx, locked);
1240 io_req_complete_failed(req, req->cqe.res);
1243 void io_req_task_submit(struct io_kiocb *req, bool *locked)
1245 io_tw_lock(req->ctx, locked);
1246 /* req->task == current here, checking PF_EXITING is safe */
1247 if (likely(!(req->task->flags & PF_EXITING)))
1250 io_req_complete_failed(req, -EFAULT);
1253 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1255 io_req_set_res(req, ret, 0);
1256 req->io_task_work.func = io_req_task_cancel;
1257 io_req_task_work_add(req);
1260 void io_req_task_queue(struct io_kiocb *req)
1262 req->io_task_work.func = io_req_task_submit;
1263 io_req_task_work_add(req);
1266 void io_queue_next(struct io_kiocb *req)
1268 struct io_kiocb *nxt = io_req_find_next(req);
1271 io_req_task_queue(nxt);
1274 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1275 __must_hold(&ctx->uring_lock)
1277 struct task_struct *task = NULL;
1281 struct io_kiocb *req = container_of(node, struct io_kiocb,
1284 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1285 if (req->flags & REQ_F_REFCOUNT) {
1286 node = req->comp_list.next;
1287 if (!req_ref_put_and_test(req))
1290 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1291 struct async_poll *apoll = req->apoll;
1293 if (apoll->double_poll)
1294 kfree(apoll->double_poll);
1295 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1297 req->flags &= ~REQ_F_POLLED;
1299 if (req->flags & IO_REQ_LINK_FLAGS)
1301 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1304 if (!(req->flags & REQ_F_FIXED_FILE))
1305 io_put_file(req->file);
1307 io_req_put_rsrc_locked(req, ctx);
1309 if (req->task != task) {
1311 io_put_task(task, task_refs);
1316 node = req->comp_list.next;
1317 io_req_add_to_cache(req, ctx);
1321 io_put_task(task, task_refs);
1324 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1325 __must_hold(&ctx->uring_lock)
1327 struct io_wq_work_node *node, *prev;
1328 struct io_submit_state *state = &ctx->submit_state;
1331 wq_list_for_each(node, prev, &state->compl_reqs) {
1332 struct io_kiocb *req = container_of(node, struct io_kiocb,
1335 if (!(req->flags & REQ_F_CQE_SKIP))
1336 __io_fill_cqe_req(ctx, req);
1338 __io_cq_unlock_post(ctx);
1340 io_free_batch_list(ctx, state->compl_reqs.first);
1341 INIT_WQ_LIST(&state->compl_reqs);
1345 * Drop reference to request, return next in chain (if there is one) if this
1346 * was the last reference to this request.
1348 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1350 struct io_kiocb *nxt = NULL;
1352 if (req_ref_put_and_test(req)) {
1353 if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
1354 nxt = io_req_find_next(req);
1360 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1362 /* See comment at the top of this file */
1364 return __io_cqring_events(ctx);
1368 * We can't just wait for polled events to come to us, we have to actively
1369 * find and complete them.
1371 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1373 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1376 mutex_lock(&ctx->uring_lock);
1377 while (!wq_list_empty(&ctx->iopoll_list)) {
1378 /* let it sleep and repeat later if can't complete a request */
1379 if (io_do_iopoll(ctx, true) == 0)
1382 * Ensure we allow local-to-the-cpu processing to take place,
1383 * in this case we need to ensure that we reap all events.
1384 * Also let task_work, etc. to progress by releasing the mutex
1386 if (need_resched()) {
1387 mutex_unlock(&ctx->uring_lock);
1389 mutex_lock(&ctx->uring_lock);
1392 mutex_unlock(&ctx->uring_lock);
1395 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1397 unsigned int nr_events = 0;
1399 unsigned long check_cq;
1401 check_cq = READ_ONCE(ctx->check_cq);
1402 if (unlikely(check_cq)) {
1403 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1404 __io_cqring_overflow_flush(ctx, false);
1406 * Similarly do not spin if we have not informed the user of any
1409 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1413 * Don't enter poll loop if we already have events pending.
1414 * If we do, we can potentially be spinning for commands that
1415 * already triggered a CQE (eg in error).
1417 if (io_cqring_events(ctx))
1422 * If a submit got punted to a workqueue, we can have the
1423 * application entering polling for a command before it gets
1424 * issued. That app will hold the uring_lock for the duration
1425 * of the poll right here, so we need to take a breather every
1426 * now and then to ensure that the issue has a chance to add
1427 * the poll to the issued list. Otherwise we can spin here
1428 * forever, while the workqueue is stuck trying to acquire the
1431 if (wq_list_empty(&ctx->iopoll_list) ||
1432 io_task_work_pending(ctx)) {
1433 if (!llist_empty(&ctx->work_llist))
1434 __io_run_local_work(ctx, true);
1435 if (task_work_pending(current) ||
1436 wq_list_empty(&ctx->iopoll_list)) {
1437 u32 tail = ctx->cached_cq_tail;
1439 mutex_unlock(&ctx->uring_lock);
1440 ret = io_run_task_work();
1441 mutex_lock(&ctx->uring_lock);
1446 /* some requests don't go through iopoll_list */
1447 if (tail != ctx->cached_cq_tail ||
1448 wq_list_empty(&ctx->iopoll_list))
1452 ret = io_do_iopoll(ctx, !min);
1457 } while (nr_events < min && !need_resched());
1462 void io_req_task_complete(struct io_kiocb *req, bool *locked)
1464 if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) {
1465 unsigned issue_flags = *locked ? 0 : IO_URING_F_UNLOCKED;
1467 req->cqe.flags |= io_put_kbuf(req, issue_flags);
1471 io_req_complete_defer(req);
1473 io_req_complete_post(req);
1477 * After the iocb has been issued, it's safe to be found on the poll list.
1478 * Adding the kiocb to the list AFTER submission ensures that we don't
1479 * find it from a io_do_iopoll() thread before the issuer is done
1480 * accessing the kiocb cookie.
1482 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1484 struct io_ring_ctx *ctx = req->ctx;
1485 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1487 /* workqueue context doesn't hold uring_lock, grab it now */
1488 if (unlikely(needs_lock))
1489 mutex_lock(&ctx->uring_lock);
1492 * Track whether we have multiple files in our lists. This will impact
1493 * how we do polling eventually, not spinning if we're on potentially
1494 * different devices.
1496 if (wq_list_empty(&ctx->iopoll_list)) {
1497 ctx->poll_multi_queue = false;
1498 } else if (!ctx->poll_multi_queue) {
1499 struct io_kiocb *list_req;
1501 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1503 if (list_req->file != req->file)
1504 ctx->poll_multi_queue = true;
1508 * For fast devices, IO may have already completed. If it has, add
1509 * it to the front so we find it first.
1511 if (READ_ONCE(req->iopoll_completed))
1512 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1514 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1516 if (unlikely(needs_lock)) {
1518 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1519 * in sq thread task context or in io worker task context. If
1520 * current task context is sq thread, we don't need to check
1521 * whether should wake up sq thread.
1523 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1524 wq_has_sleeper(&ctx->sq_data->wait))
1525 wake_up(&ctx->sq_data->wait);
1527 mutex_unlock(&ctx->uring_lock);
1531 static bool io_bdev_nowait(struct block_device *bdev)
1533 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
1537 * If we tracked the file through the SCM inflight mechanism, we could support
1538 * any file. For now, just ensure that anything potentially problematic is done
1541 static bool __io_file_supports_nowait(struct file *file, umode_t mode)
1543 if (S_ISBLK(mode)) {
1544 if (IS_ENABLED(CONFIG_BLOCK) &&
1545 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
1551 if (S_ISREG(mode)) {
1552 if (IS_ENABLED(CONFIG_BLOCK) &&
1553 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
1554 !io_is_uring_fops(file))
1559 /* any ->read/write should understand O_NONBLOCK */
1560 if (file->f_flags & O_NONBLOCK)
1562 return file->f_mode & FMODE_NOWAIT;
1566 * If we tracked the file through the SCM inflight mechanism, we could support
1567 * any file. For now, just ensure that anything potentially problematic is done
1570 unsigned int io_file_get_flags(struct file *file)
1572 umode_t mode = file_inode(file)->i_mode;
1573 unsigned int res = 0;
1577 if (__io_file_supports_nowait(file, mode))
1579 if (io_file_need_scm(file))
1584 bool io_alloc_async_data(struct io_kiocb *req)
1586 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
1587 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
1588 if (req->async_data) {
1589 req->flags |= REQ_F_ASYNC_DATA;
1595 int io_req_prep_async(struct io_kiocb *req)
1597 const struct io_op_def *def = &io_op_defs[req->opcode];
1599 /* assign early for deferred execution for non-fixed file */
1600 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE))
1601 req->file = io_file_get_normal(req, req->cqe.fd);
1602 if (!def->prep_async)
1604 if (WARN_ON_ONCE(req_has_async_data(req)))
1606 if (!io_op_defs[req->opcode].manual_alloc) {
1607 if (io_alloc_async_data(req))
1610 return def->prep_async(req);
1613 static u32 io_get_sequence(struct io_kiocb *req)
1615 u32 seq = req->ctx->cached_sq_head;
1616 struct io_kiocb *cur;
1618 /* need original cached_sq_head, but it was increased for each req */
1619 io_for_each_link(cur, req)
1624 static __cold void io_drain_req(struct io_kiocb *req)
1626 struct io_ring_ctx *ctx = req->ctx;
1627 struct io_defer_entry *de;
1629 u32 seq = io_get_sequence(req);
1631 /* Still need defer if there is pending req in defer list. */
1632 spin_lock(&ctx->completion_lock);
1633 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1634 spin_unlock(&ctx->completion_lock);
1636 ctx->drain_active = false;
1637 io_req_task_queue(req);
1640 spin_unlock(&ctx->completion_lock);
1642 ret = io_req_prep_async(req);
1645 io_req_complete_failed(req, ret);
1648 io_prep_async_link(req);
1649 de = kmalloc(sizeof(*de), GFP_KERNEL);
1655 spin_lock(&ctx->completion_lock);
1656 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1657 spin_unlock(&ctx->completion_lock);
1662 trace_io_uring_defer(req);
1665 list_add_tail(&de->list, &ctx->defer_list);
1666 spin_unlock(&ctx->completion_lock);
1669 static void io_clean_op(struct io_kiocb *req)
1671 if (req->flags & REQ_F_BUFFER_SELECTED) {
1672 spin_lock(&req->ctx->completion_lock);
1673 io_put_kbuf_comp(req);
1674 spin_unlock(&req->ctx->completion_lock);
1677 if (req->flags & REQ_F_NEED_CLEANUP) {
1678 const struct io_op_def *def = &io_op_defs[req->opcode];
1683 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1684 kfree(req->apoll->double_poll);
1688 if (req->flags & REQ_F_INFLIGHT) {
1689 struct io_uring_task *tctx = req->task->io_uring;
1691 atomic_dec(&tctx->inflight_tracked);
1693 if (req->flags & REQ_F_CREDS)
1694 put_cred(req->creds);
1695 if (req->flags & REQ_F_ASYNC_DATA) {
1696 kfree(req->async_data);
1697 req->async_data = NULL;
1699 req->flags &= ~IO_REQ_CLEAN_FLAGS;
1702 static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags)
1704 if (req->file || !io_op_defs[req->opcode].needs_file)
1707 if (req->flags & REQ_F_FIXED_FILE)
1708 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1710 req->file = io_file_get_normal(req, req->cqe.fd);
1715 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1717 const struct io_op_def *def = &io_op_defs[req->opcode];
1718 const struct cred *creds = NULL;
1721 if (unlikely(!io_assign_file(req, issue_flags)))
1724 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1725 creds = override_creds(req->creds);
1727 if (!def->audit_skip)
1728 audit_uring_entry(req->opcode);
1730 ret = def->issue(req, issue_flags);
1732 if (!def->audit_skip)
1733 audit_uring_exit(!ret, ret);
1736 revert_creds(creds);
1738 if (ret == IOU_OK) {
1739 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1740 io_req_complete_defer(req);
1742 io_req_complete_post(req);
1743 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1746 /* If the op doesn't have a file, we're not polling for it */
1747 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && req->file)
1748 io_iopoll_req_issued(req, issue_flags);
1753 int io_poll_issue(struct io_kiocb *req, bool *locked)
1755 io_tw_lock(req->ctx, locked);
1756 if (unlikely(req->task->flags & PF_EXITING))
1758 return io_issue_sqe(req, IO_URING_F_NONBLOCK);
1761 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1763 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1765 req = io_put_req_find_next(req);
1766 return req ? &req->work : NULL;
1769 void io_wq_submit_work(struct io_wq_work *work)
1771 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1772 const struct io_op_def *def = &io_op_defs[req->opcode];
1773 unsigned int issue_flags = IO_URING_F_UNLOCKED;
1774 bool needs_poll = false;
1775 int ret = 0, err = -ECANCELED;
1777 /* one will be dropped by ->io_free_work() after returning to io-wq */
1778 if (!(req->flags & REQ_F_REFCOUNT))
1779 __io_req_set_refcount(req, 2);
1783 io_arm_ltimeout(req);
1785 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1786 if (work->flags & IO_WQ_WORK_CANCEL) {
1788 io_req_task_queue_fail(req, err);
1791 if (!io_assign_file(req, issue_flags)) {
1793 work->flags |= IO_WQ_WORK_CANCEL;
1797 if (req->flags & REQ_F_FORCE_ASYNC) {
1798 bool opcode_poll = def->pollin || def->pollout;
1800 if (opcode_poll && file_can_poll(req->file)) {
1802 issue_flags |= IO_URING_F_NONBLOCK;
1807 ret = io_issue_sqe(req, issue_flags);
1811 * We can get EAGAIN for iopolled IO even though we're
1812 * forcing a sync submission from here, since we can't
1813 * wait for request slots on the block side.
1816 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1822 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1824 /* aborted or ready, in either case retry blocking */
1826 issue_flags &= ~IO_URING_F_NONBLOCK;
1829 /* avoid locking problems by failing it from a clean context */
1831 io_req_task_queue_fail(req, ret);
1834 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1835 unsigned int issue_flags)
1837 struct io_ring_ctx *ctx = req->ctx;
1838 struct file *file = NULL;
1839 unsigned long file_ptr;
1841 io_ring_submit_lock(ctx, issue_flags);
1843 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
1845 fd = array_index_nospec(fd, ctx->nr_user_files);
1846 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
1847 file = (struct file *) (file_ptr & FFS_MASK);
1848 file_ptr &= ~FFS_MASK;
1849 /* mask in overlapping REQ_F and FFS bits */
1850 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
1851 io_req_set_rsrc_node(req, ctx, 0);
1852 WARN_ON_ONCE(file && !test_bit(fd, ctx->file_table.bitmap));
1854 io_ring_submit_unlock(ctx, issue_flags);
1858 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
1860 struct file *file = fget(fd);
1862 trace_io_uring_file_get(req, fd);
1864 /* we don't allow fixed io_uring files */
1865 if (file && io_is_uring_fops(file))
1866 io_req_track_inflight(req);
1870 static void io_queue_async(struct io_kiocb *req, int ret)
1871 __must_hold(&req->ctx->uring_lock)
1873 struct io_kiocb *linked_timeout;
1875 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
1876 io_req_complete_failed(req, ret);
1880 linked_timeout = io_prep_linked_timeout(req);
1882 switch (io_arm_poll_handler(req, 0)) {
1883 case IO_APOLL_READY:
1884 io_kbuf_recycle(req, 0);
1885 io_req_task_queue(req);
1887 case IO_APOLL_ABORTED:
1888 io_kbuf_recycle(req, 0);
1889 io_queue_iowq(req, NULL);
1896 io_queue_linked_timeout(linked_timeout);
1899 static inline void io_queue_sqe(struct io_kiocb *req)
1900 __must_hold(&req->ctx->uring_lock)
1904 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
1907 * We async punt it if the file wasn't marked NOWAIT, or if the file
1908 * doesn't support non-blocking read/write attempts
1911 io_arm_ltimeout(req);
1913 io_queue_async(req, ret);
1916 static void io_queue_sqe_fallback(struct io_kiocb *req)
1917 __must_hold(&req->ctx->uring_lock)
1919 if (unlikely(req->flags & REQ_F_FAIL)) {
1921 * We don't submit, fail them all, for that replace hardlinks
1922 * with normal links. Extra REQ_F_LINK is tolerated.
1924 req->flags &= ~REQ_F_HARDLINK;
1925 req->flags |= REQ_F_LINK;
1926 io_req_complete_failed(req, req->cqe.res);
1927 } else if (unlikely(req->ctx->drain_active)) {
1930 int ret = io_req_prep_async(req);
1933 io_req_complete_failed(req, ret);
1935 io_queue_iowq(req, NULL);
1940 * Check SQE restrictions (opcode and flags).
1942 * Returns 'true' if SQE is allowed, 'false' otherwise.
1944 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
1945 struct io_kiocb *req,
1946 unsigned int sqe_flags)
1948 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
1951 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
1952 ctx->restrictions.sqe_flags_required)
1955 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
1956 ctx->restrictions.sqe_flags_required))
1962 static void io_init_req_drain(struct io_kiocb *req)
1964 struct io_ring_ctx *ctx = req->ctx;
1965 struct io_kiocb *head = ctx->submit_state.link.head;
1967 ctx->drain_active = true;
1970 * If we need to drain a request in the middle of a link, drain
1971 * the head request and the next request/link after the current
1972 * link. Considering sequential execution of links,
1973 * REQ_F_IO_DRAIN will be maintained for every request of our
1976 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
1977 ctx->drain_next = true;
1981 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
1982 const struct io_uring_sqe *sqe)
1983 __must_hold(&ctx->uring_lock)
1985 const struct io_op_def *def;
1986 unsigned int sqe_flags;
1990 /* req is partially pre-initialised, see io_preinit_req() */
1991 req->opcode = opcode = READ_ONCE(sqe->opcode);
1992 /* same numerical values with corresponding REQ_F_*, safe to copy */
1993 req->flags = sqe_flags = READ_ONCE(sqe->flags);
1994 req->cqe.user_data = READ_ONCE(sqe->user_data);
1996 req->rsrc_node = NULL;
1997 req->task = current;
1999 if (unlikely(opcode >= IORING_OP_LAST)) {
2003 def = &io_op_defs[opcode];
2004 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2005 /* enforce forwards compatibility on users */
2006 if (sqe_flags & ~SQE_VALID_FLAGS)
2008 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2009 if (!def->buffer_select)
2011 req->buf_index = READ_ONCE(sqe->buf_group);
2013 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2014 ctx->drain_disabled = true;
2015 if (sqe_flags & IOSQE_IO_DRAIN) {
2016 if (ctx->drain_disabled)
2018 io_init_req_drain(req);
2021 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2022 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2024 /* knock it to the slow queue path, will be drained there */
2025 if (ctx->drain_active)
2026 req->flags |= REQ_F_FORCE_ASYNC;
2027 /* if there is no link, we're at "next" request and need to drain */
2028 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2029 ctx->drain_next = false;
2030 ctx->drain_active = true;
2031 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2035 if (!def->ioprio && sqe->ioprio)
2037 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2040 if (def->needs_file) {
2041 struct io_submit_state *state = &ctx->submit_state;
2043 req->cqe.fd = READ_ONCE(sqe->fd);
2046 * Plug now if we have more than 2 IO left after this, and the
2047 * target is potentially a read/write to block based storage.
2049 if (state->need_plug && def->plug) {
2050 state->plug_started = true;
2051 state->need_plug = false;
2052 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2056 personality = READ_ONCE(sqe->personality);
2060 req->creds = xa_load(&ctx->personalities, personality);
2063 get_cred(req->creds);
2064 ret = security_uring_override_creds(req->creds);
2066 put_cred(req->creds);
2069 req->flags |= REQ_F_CREDS;
2072 return def->prep(req, sqe);
2075 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2076 struct io_kiocb *req, int ret)
2078 struct io_ring_ctx *ctx = req->ctx;
2079 struct io_submit_link *link = &ctx->submit_state.link;
2080 struct io_kiocb *head = link->head;
2082 trace_io_uring_req_failed(sqe, req, ret);
2085 * Avoid breaking links in the middle as it renders links with SQPOLL
2086 * unusable. Instead of failing eagerly, continue assembling the link if
2087 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2088 * should find the flag and handle the rest.
2090 req_fail_link_node(req, ret);
2091 if (head && !(head->flags & REQ_F_FAIL))
2092 req_fail_link_node(head, -ECANCELED);
2094 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2096 link->last->link = req;
2100 io_queue_sqe_fallback(req);
2105 link->last->link = req;
2112 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2113 const struct io_uring_sqe *sqe)
2114 __must_hold(&ctx->uring_lock)
2116 struct io_submit_link *link = &ctx->submit_state.link;
2119 ret = io_init_req(ctx, req, sqe);
2121 return io_submit_fail_init(sqe, req, ret);
2123 /* don't need @sqe from now on */
2124 trace_io_uring_submit_sqe(req, true);
2127 * If we already have a head request, queue this one for async
2128 * submittal once the head completes. If we don't have a head but
2129 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2130 * submitted sync once the chain is complete. If none of those
2131 * conditions are true (normal request), then just queue it.
2133 if (unlikely(link->head)) {
2134 ret = io_req_prep_async(req);
2136 return io_submit_fail_init(sqe, req, ret);
2138 trace_io_uring_link(req, link->head);
2139 link->last->link = req;
2142 if (req->flags & IO_REQ_LINK_FLAGS)
2144 /* last request of the link, flush it */
2147 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2150 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2151 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2152 if (req->flags & IO_REQ_LINK_FLAGS) {
2157 io_queue_sqe_fallback(req);
2167 * Batched submission is done, ensure local IO is flushed out.
2169 static void io_submit_state_end(struct io_ring_ctx *ctx)
2171 struct io_submit_state *state = &ctx->submit_state;
2173 if (unlikely(state->link.head))
2174 io_queue_sqe_fallback(state->link.head);
2175 /* flush only after queuing links as they can generate completions */
2176 io_submit_flush_completions(ctx);
2177 if (state->plug_started)
2178 blk_finish_plug(&state->plug);
2182 * Start submission side cache.
2184 static void io_submit_state_start(struct io_submit_state *state,
2185 unsigned int max_ios)
2187 state->plug_started = false;
2188 state->need_plug = max_ios > 2;
2189 state->submit_nr = max_ios;
2190 /* set only head, no need to init link_last in advance */
2191 state->link.head = NULL;
2194 static void io_commit_sqring(struct io_ring_ctx *ctx)
2196 struct io_rings *rings = ctx->rings;
2199 * Ensure any loads from the SQEs are done at this point,
2200 * since once we write the new head, the application could
2201 * write new data to them.
2203 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2207 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2208 * that is mapped by userspace. This means that care needs to be taken to
2209 * ensure that reads are stable, as we cannot rely on userspace always
2210 * being a good citizen. If members of the sqe are validated and then later
2211 * used, it's important that those reads are done through READ_ONCE() to
2212 * prevent a re-load down the line.
2214 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
2216 unsigned head, mask = ctx->sq_entries - 1;
2217 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2220 * The cached sq head (or cq tail) serves two purposes:
2222 * 1) allows us to batch the cost of updating the user visible
2224 * 2) allows the kernel side to track the head on its own, even
2225 * though the application is the one updating it.
2227 head = READ_ONCE(ctx->sq_array[sq_idx]);
2228 if (likely(head < ctx->sq_entries)) {
2229 /* double index for 128-byte SQEs, twice as long */
2230 if (ctx->flags & IORING_SETUP_SQE128)
2232 return &ctx->sq_sqes[head];
2235 /* drop invalid entries */
2237 WRITE_ONCE(ctx->rings->sq_dropped,
2238 READ_ONCE(ctx->rings->sq_dropped) + 1);
2242 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2243 __must_hold(&ctx->uring_lock)
2245 unsigned int entries = io_sqring_entries(ctx);
2249 if (unlikely(!entries))
2251 /* make sure SQ entry isn't read before tail */
2252 ret = left = min3(nr, ctx->sq_entries, entries);
2253 io_get_task_refs(left);
2254 io_submit_state_start(&ctx->submit_state, left);
2257 const struct io_uring_sqe *sqe;
2258 struct io_kiocb *req;
2260 if (unlikely(!io_alloc_req_refill(ctx)))
2262 req = io_alloc_req(ctx);
2263 sqe = io_get_sqe(ctx);
2264 if (unlikely(!sqe)) {
2265 io_req_add_to_cache(req, ctx);
2270 * Continue submitting even for sqe failure if the
2271 * ring was setup with IORING_SETUP_SUBMIT_ALL
2273 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2274 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2280 if (unlikely(left)) {
2282 /* try again if it submitted nothing and can't allocate a req */
2283 if (!ret && io_req_cache_empty(ctx))
2285 current->io_uring->cached_refs += left;
2288 io_submit_state_end(ctx);
2289 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2290 io_commit_sqring(ctx);
2294 struct io_wait_queue {
2295 struct wait_queue_entry wq;
2296 struct io_ring_ctx *ctx;
2298 unsigned nr_timeouts;
2301 static inline bool io_has_work(struct io_ring_ctx *ctx)
2303 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
2304 ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
2305 !llist_empty(&ctx->work_llist));
2308 static inline bool io_should_wake(struct io_wait_queue *iowq)
2310 struct io_ring_ctx *ctx = iowq->ctx;
2311 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
2314 * Wake up if we have enough events, or if a timeout occurred since we
2315 * started waiting. For timeouts, we always want to return to userspace,
2316 * regardless of event count.
2318 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2321 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2322 int wake_flags, void *key)
2324 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
2326 struct io_ring_ctx *ctx = iowq->ctx;
2329 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2330 * the task, and the next invocation will do it.
2332 if (io_should_wake(iowq) || io_has_work(ctx))
2333 return autoremove_wake_function(curr, mode, wake_flags, key);
2337 int io_run_task_work_sig(struct io_ring_ctx *ctx)
2339 if (io_run_task_work_ctx(ctx) > 0)
2341 if (task_sigpending(current))
2346 /* when returns >0, the caller should retry */
2347 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2348 struct io_wait_queue *iowq,
2352 unsigned long check_cq;
2354 /* make sure we run task_work before checking for signals */
2355 ret = io_run_task_work_sig(ctx);
2356 if (ret || io_should_wake(iowq))
2359 check_cq = READ_ONCE(ctx->check_cq);
2360 if (unlikely(check_cq)) {
2361 /* let the caller flush overflows, retry */
2362 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2364 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
2367 if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS))
2373 * Wait until events become available, if we don't already have some. The
2374 * application must reap them itself, as they reside on the shared cq ring.
2376 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2377 const sigset_t __user *sig, size_t sigsz,
2378 struct __kernel_timespec __user *uts)
2380 struct io_wait_queue iowq;
2381 struct io_rings *rings = ctx->rings;
2382 ktime_t timeout = KTIME_MAX;
2386 /* always run at least 1 task work to process local work */
2387 ret = io_run_task_work_ctx(ctx);
2390 io_cqring_overflow_flush(ctx);
2391 if (io_cqring_events(ctx) >= min_events)
2396 #ifdef CONFIG_COMPAT
2397 if (in_compat_syscall())
2398 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2402 ret = set_user_sigmask(sig, sigsz);
2409 struct timespec64 ts;
2411 if (get_timespec64(&ts, uts))
2413 timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2416 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2417 iowq.wq.private = current;
2418 INIT_LIST_HEAD(&iowq.wq.entry);
2420 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2421 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2423 trace_io_uring_cqring_wait(ctx, min_events);
2425 /* if we can't even flush overflow, don't wait for more */
2426 if (!io_cqring_overflow_flush(ctx)) {
2430 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2431 TASK_INTERRUPTIBLE);
2432 ret = io_cqring_wait_schedule(ctx, &iowq, timeout);
2436 finish_wait(&ctx->cq_wait, &iowq.wq);
2437 restore_saved_sigmask_unless(ret == -EINTR);
2439 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2442 static void io_mem_free(void *ptr)
2449 page = virt_to_head_page(ptr);
2450 if (put_page_testzero(page))
2451 free_compound_page(page);
2454 static void *io_mem_alloc(size_t size)
2456 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2458 return (void *) __get_free_pages(gfp, get_order(size));
2461 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2462 unsigned int cq_entries, size_t *sq_offset)
2464 struct io_rings *rings;
2465 size_t off, sq_array_size;
2467 off = struct_size(rings, cqes, cq_entries);
2468 if (off == SIZE_MAX)
2470 if (ctx->flags & IORING_SETUP_CQE32) {
2471 if (check_shl_overflow(off, 1, &off))
2476 off = ALIGN(off, SMP_CACHE_BYTES);
2484 sq_array_size = array_size(sizeof(u32), sq_entries);
2485 if (sq_array_size == SIZE_MAX)
2488 if (check_add_overflow(off, sq_array_size, &off))
2494 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2495 unsigned int eventfd_async)
2497 struct io_ev_fd *ev_fd;
2498 __s32 __user *fds = arg;
2501 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2502 lockdep_is_held(&ctx->uring_lock));
2506 if (copy_from_user(&fd, fds, sizeof(*fds)))
2509 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2513 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2514 if (IS_ERR(ev_fd->cq_ev_fd)) {
2515 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2520 spin_lock(&ctx->completion_lock);
2521 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2522 spin_unlock(&ctx->completion_lock);
2524 ev_fd->eventfd_async = eventfd_async;
2525 ctx->has_evfd = true;
2526 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2527 atomic_set(&ev_fd->refs, 1);
2528 atomic_set(&ev_fd->ops, 0);
2532 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2534 struct io_ev_fd *ev_fd;
2536 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2537 lockdep_is_held(&ctx->uring_lock));
2539 ctx->has_evfd = false;
2540 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2541 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops))
2542 call_rcu(&ev_fd->rcu, io_eventfd_ops);
2549 static void io_req_caches_free(struct io_ring_ctx *ctx)
2551 struct io_submit_state *state = &ctx->submit_state;
2554 mutex_lock(&ctx->uring_lock);
2555 io_flush_cached_locked_reqs(ctx, state);
2557 while (!io_req_cache_empty(ctx)) {
2558 struct io_wq_work_node *node;
2559 struct io_kiocb *req;
2561 node = wq_stack_extract(&state->free_list);
2562 req = container_of(node, struct io_kiocb, comp_list);
2563 kmem_cache_free(req_cachep, req);
2567 percpu_ref_put_many(&ctx->refs, nr);
2568 mutex_unlock(&ctx->uring_lock);
2571 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2573 io_sq_thread_finish(ctx);
2575 if (ctx->mm_account) {
2576 mmdrop(ctx->mm_account);
2577 ctx->mm_account = NULL;
2580 io_rsrc_refs_drop(ctx);
2581 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2582 io_wait_rsrc_data(ctx->buf_data);
2583 io_wait_rsrc_data(ctx->file_data);
2585 mutex_lock(&ctx->uring_lock);
2587 __io_sqe_buffers_unregister(ctx);
2589 __io_sqe_files_unregister(ctx);
2591 __io_cqring_overflow_flush(ctx, true);
2592 io_eventfd_unregister(ctx);
2593 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2594 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2595 mutex_unlock(&ctx->uring_lock);
2596 io_destroy_buffers(ctx);
2598 put_cred(ctx->sq_creds);
2599 if (ctx->submitter_task)
2600 put_task_struct(ctx->submitter_task);
2602 /* there are no registered resources left, nobody uses it */
2604 io_rsrc_node_destroy(ctx->rsrc_node);
2605 if (ctx->rsrc_backup_node)
2606 io_rsrc_node_destroy(ctx->rsrc_backup_node);
2607 flush_delayed_work(&ctx->rsrc_put_work);
2608 flush_delayed_work(&ctx->fallback_work);
2610 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2611 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
2613 #if defined(CONFIG_UNIX)
2614 if (ctx->ring_sock) {
2615 ctx->ring_sock->file = NULL; /* so that iput() is called */
2616 sock_release(ctx->ring_sock);
2619 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2620 WARN_ON_ONCE(ctx->notif_slots || ctx->nr_notif_slots);
2622 io_mem_free(ctx->rings);
2623 io_mem_free(ctx->sq_sqes);
2625 percpu_ref_exit(&ctx->refs);
2626 free_uid(ctx->user);
2627 io_req_caches_free(ctx);
2629 io_wq_put_hash(ctx->hash_map);
2630 kfree(ctx->cancel_table.hbs);
2631 kfree(ctx->cancel_table_locked.hbs);
2632 kfree(ctx->dummy_ubuf);
2634 xa_destroy(&ctx->io_bl_xa);
2638 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2640 struct io_ring_ctx *ctx = file->private_data;
2643 poll_wait(file, &ctx->cq_wait, wait);
2645 * synchronizes with barrier from wq_has_sleeper call in
2649 if (!io_sqring_full(ctx))
2650 mask |= EPOLLOUT | EPOLLWRNORM;
2653 * Don't flush cqring overflow list here, just do a simple check.
2654 * Otherwise there could possible be ABBA deadlock:
2657 * lock(&ctx->uring_lock);
2659 * lock(&ctx->uring_lock);
2662 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2663 * pushs them to do the flush.
2666 if (io_cqring_events(ctx) || io_has_work(ctx))
2667 mask |= EPOLLIN | EPOLLRDNORM;
2672 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2674 const struct cred *creds;
2676 creds = xa_erase(&ctx->personalities, id);
2685 struct io_tctx_exit {
2686 struct callback_head task_work;
2687 struct completion completion;
2688 struct io_ring_ctx *ctx;
2691 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2693 struct io_uring_task *tctx = current->io_uring;
2694 struct io_tctx_exit *work;
2696 work = container_of(cb, struct io_tctx_exit, task_work);
2698 * When @in_idle, we're in cancellation and it's racy to remove the
2699 * node. It'll be removed by the end of cancellation, just ignore it.
2701 if (!atomic_read(&tctx->in_idle))
2702 io_uring_del_tctx_node((unsigned long)work->ctx);
2703 complete(&work->completion);
2706 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2708 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2710 return req->ctx == data;
2713 static __cold void io_ring_exit_work(struct work_struct *work)
2715 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2716 unsigned long timeout = jiffies + HZ * 60 * 5;
2717 unsigned long interval = HZ / 20;
2718 struct io_tctx_exit exit;
2719 struct io_tctx_node *node;
2723 * If we're doing polled IO and end up having requests being
2724 * submitted async (out-of-line), then completions can come in while
2725 * we're waiting for refs to drop. We need to reap these manually,
2726 * as nobody else will be looking for them.
2729 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
2730 io_move_task_work_from_local(ctx);
2732 while (io_uring_try_cancel_requests(ctx, NULL, true))
2736 struct io_sq_data *sqd = ctx->sq_data;
2737 struct task_struct *tsk;
2739 io_sq_thread_park(sqd);
2741 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
2742 io_wq_cancel_cb(tsk->io_uring->io_wq,
2743 io_cancel_ctx_cb, ctx, true);
2744 io_sq_thread_unpark(sqd);
2747 io_req_caches_free(ctx);
2749 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
2750 /* there is little hope left, don't run it too often */
2753 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
2755 init_completion(&exit.completion);
2756 init_task_work(&exit.task_work, io_tctx_exit_cb);
2759 * Some may use context even when all refs and requests have been put,
2760 * and they are free to do so while still holding uring_lock or
2761 * completion_lock, see io_req_task_submit(). Apart from other work,
2762 * this lock/unlock section also waits them to finish.
2764 mutex_lock(&ctx->uring_lock);
2765 while (!list_empty(&ctx->tctx_list)) {
2766 WARN_ON_ONCE(time_after(jiffies, timeout));
2768 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
2770 /* don't spin on a single task if cancellation failed */
2771 list_rotate_left(&ctx->tctx_list);
2772 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
2773 if (WARN_ON_ONCE(ret))
2776 mutex_unlock(&ctx->uring_lock);
2777 wait_for_completion(&exit.completion);
2778 mutex_lock(&ctx->uring_lock);
2780 mutex_unlock(&ctx->uring_lock);
2781 spin_lock(&ctx->completion_lock);
2782 spin_unlock(&ctx->completion_lock);
2784 io_ring_ctx_free(ctx);
2787 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
2789 unsigned long index;
2790 struct creds *creds;
2792 mutex_lock(&ctx->uring_lock);
2793 percpu_ref_kill(&ctx->refs);
2795 __io_cqring_overflow_flush(ctx, true);
2796 xa_for_each(&ctx->personalities, index, creds)
2797 io_unregister_personality(ctx, index);
2799 io_poll_remove_all(ctx, NULL, true);
2800 mutex_unlock(&ctx->uring_lock);
2802 /* failed during ring init, it couldn't have issued any requests */
2804 io_kill_timeouts(ctx, NULL, true);
2805 /* if we failed setting up the ctx, we might not have any rings */
2806 io_iopoll_try_reap_events(ctx);
2809 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
2811 * Use system_unbound_wq to avoid spawning tons of event kworkers
2812 * if we're exiting a ton of rings at the same time. It just adds
2813 * noise and overhead, there's no discernable change in runtime
2814 * over using system_wq.
2816 queue_work(system_unbound_wq, &ctx->exit_work);
2819 static int io_uring_release(struct inode *inode, struct file *file)
2821 struct io_ring_ctx *ctx = file->private_data;
2823 file->private_data = NULL;
2824 io_ring_ctx_wait_and_kill(ctx);
2828 struct io_task_cancel {
2829 struct task_struct *task;
2833 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
2835 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2836 struct io_task_cancel *cancel = data;
2838 return io_match_task_safe(req, cancel->task, cancel->all);
2841 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
2842 struct task_struct *task,
2845 struct io_defer_entry *de;
2848 spin_lock(&ctx->completion_lock);
2849 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
2850 if (io_match_task_safe(de->req, task, cancel_all)) {
2851 list_cut_position(&list, &ctx->defer_list, &de->list);
2855 spin_unlock(&ctx->completion_lock);
2856 if (list_empty(&list))
2859 while (!list_empty(&list)) {
2860 de = list_first_entry(&list, struct io_defer_entry, list);
2861 list_del_init(&de->list);
2862 io_req_complete_failed(de->req, -ECANCELED);
2868 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
2870 struct io_tctx_node *node;
2871 enum io_wq_cancel cret;
2874 mutex_lock(&ctx->uring_lock);
2875 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
2876 struct io_uring_task *tctx = node->task->io_uring;
2879 * io_wq will stay alive while we hold uring_lock, because it's
2880 * killed after ctx nodes, which requires to take the lock.
2882 if (!tctx || !tctx->io_wq)
2884 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
2885 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2887 mutex_unlock(&ctx->uring_lock);
2892 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
2893 struct task_struct *task,
2896 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
2897 struct io_uring_task *tctx = task ? task->io_uring : NULL;
2898 enum io_wq_cancel cret;
2901 /* failed during ring init, it couldn't have issued any requests */
2906 ret |= io_uring_try_cancel_iowq(ctx);
2907 } else if (tctx && tctx->io_wq) {
2909 * Cancels requests of all rings, not only @ctx, but
2910 * it's fine as the task is in exit/exec.
2912 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
2914 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2917 /* SQPOLL thread does its own polling */
2918 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
2919 (ctx->sq_data && ctx->sq_data->thread == current)) {
2920 while (!wq_list_empty(&ctx->iopoll_list)) {
2921 io_iopoll_try_reap_events(ctx);
2926 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
2927 ret |= io_run_local_work(ctx) > 0;
2928 ret |= io_cancel_defer_files(ctx, task, cancel_all);
2929 mutex_lock(&ctx->uring_lock);
2930 ret |= io_poll_remove_all(ctx, task, cancel_all);
2931 mutex_unlock(&ctx->uring_lock);
2932 ret |= io_kill_timeouts(ctx, task, cancel_all);
2934 ret |= io_run_task_work() > 0;
2938 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
2941 return atomic_read(&tctx->inflight_tracked);
2942 return percpu_counter_sum(&tctx->inflight);
2946 * Find any io_uring ctx that this task has registered or done IO on, and cancel
2947 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
2949 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
2951 struct io_uring_task *tctx = current->io_uring;
2952 struct io_ring_ctx *ctx;
2956 WARN_ON_ONCE(sqd && sqd->thread != current);
2958 if (!current->io_uring)
2961 io_wq_exit_start(tctx->io_wq);
2963 atomic_inc(&tctx->in_idle);
2967 io_uring_drop_tctx_refs(current);
2968 /* read completions before cancelations */
2969 inflight = tctx_inflight(tctx, !cancel_all);
2974 struct io_tctx_node *node;
2975 unsigned long index;
2977 xa_for_each(&tctx->xa, index, node) {
2978 /* sqpoll task will cancel all its requests */
2979 if (node->ctx->sq_data)
2981 loop |= io_uring_try_cancel_requests(node->ctx,
2982 current, cancel_all);
2985 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
2986 loop |= io_uring_try_cancel_requests(ctx,
2996 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
2998 io_uring_drop_tctx_refs(current);
3001 * If we've seen completions, retry without waiting. This
3002 * avoids a race where a completion comes in before we did
3003 * prepare_to_wait().
3005 if (inflight == tctx_inflight(tctx, !cancel_all))
3007 finish_wait(&tctx->wait, &wait);
3010 io_uring_clean_tctx(tctx);
3013 * We shouldn't run task_works after cancel, so just leave
3014 * ->in_idle set for normal exit.
3016 atomic_dec(&tctx->in_idle);
3017 /* for exec all current's requests should be gone, kill tctx */
3018 __io_uring_free(current);
3022 void __io_uring_cancel(bool cancel_all)
3024 io_uring_cancel_generic(cancel_all, NULL);
3027 static void *io_uring_validate_mmap_request(struct file *file,
3028 loff_t pgoff, size_t sz)
3030 struct io_ring_ctx *ctx = file->private_data;
3031 loff_t offset = pgoff << PAGE_SHIFT;
3036 case IORING_OFF_SQ_RING:
3037 case IORING_OFF_CQ_RING:
3040 case IORING_OFF_SQES:
3044 return ERR_PTR(-EINVAL);
3047 page = virt_to_head_page(ptr);
3048 if (sz > page_size(page))
3049 return ERR_PTR(-EINVAL);
3056 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3058 size_t sz = vma->vm_end - vma->vm_start;
3062 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3064 return PTR_ERR(ptr);
3066 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3067 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3070 #else /* !CONFIG_MMU */
3072 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3074 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
3077 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3079 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3082 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3083 unsigned long addr, unsigned long len,
3084 unsigned long pgoff, unsigned long flags)
3088 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3090 return PTR_ERR(ptr);
3092 return (unsigned long) ptr;
3095 #endif /* !CONFIG_MMU */
3097 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3099 if (flags & IORING_ENTER_EXT_ARG) {
3100 struct io_uring_getevents_arg arg;
3102 if (argsz != sizeof(arg))
3104 if (copy_from_user(&arg, argp, sizeof(arg)))
3110 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3111 struct __kernel_timespec __user **ts,
3112 const sigset_t __user **sig)
3114 struct io_uring_getevents_arg arg;
3117 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3118 * is just a pointer to the sigset_t.
3120 if (!(flags & IORING_ENTER_EXT_ARG)) {
3121 *sig = (const sigset_t __user *) argp;
3127 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3128 * timespec and sigset_t pointers if good.
3130 if (*argsz != sizeof(arg))
3132 if (copy_from_user(&arg, argp, sizeof(arg)))
3136 *sig = u64_to_user_ptr(arg.sigmask);
3137 *argsz = arg.sigmask_sz;
3138 *ts = u64_to_user_ptr(arg.ts);
3142 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3143 u32, min_complete, u32, flags, const void __user *, argp,
3146 struct io_ring_ctx *ctx;
3150 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3151 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3152 IORING_ENTER_REGISTERED_RING)))
3156 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3157 * need only dereference our task private array to find it.
3159 if (flags & IORING_ENTER_REGISTERED_RING) {
3160 struct io_uring_task *tctx = current->io_uring;
3162 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3164 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3165 f.file = tctx->registered_rings[fd];
3167 if (unlikely(!f.file))
3171 if (unlikely(!f.file))
3174 if (unlikely(!io_is_uring_fops(f.file)))
3178 ctx = f.file->private_data;
3180 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3184 * For SQ polling, the thread will do all submissions and completions.
3185 * Just return the requested submit count, and wake the thread if
3189 if (ctx->flags & IORING_SETUP_SQPOLL) {
3190 io_cqring_overflow_flush(ctx);
3192 if (unlikely(ctx->sq_data->thread == NULL)) {
3196 if (flags & IORING_ENTER_SQ_WAKEUP)
3197 wake_up(&ctx->sq_data->wait);
3198 if (flags & IORING_ENTER_SQ_WAIT) {
3199 ret = io_sqpoll_wait_sq(ctx);
3204 } else if (to_submit) {
3205 ret = io_uring_add_tctx_node(ctx);
3209 mutex_lock(&ctx->uring_lock);
3210 ret = io_submit_sqes(ctx, to_submit);
3211 if (ret != to_submit) {
3212 mutex_unlock(&ctx->uring_lock);
3215 if ((flags & IORING_ENTER_GETEVENTS) && ctx->syscall_iopoll)
3217 mutex_unlock(&ctx->uring_lock);
3220 if (flags & IORING_ENTER_GETEVENTS) {
3223 if (ctx->syscall_iopoll) {
3225 * We disallow the app entering submit/complete with
3226 * polling, but we still need to lock the ring to
3227 * prevent racing with polled issue that got punted to
3230 mutex_lock(&ctx->uring_lock);
3232 ret2 = io_validate_ext_arg(flags, argp, argsz);
3233 if (likely(!ret2)) {
3234 min_complete = min(min_complete,
3236 ret2 = io_iopoll_check(ctx, min_complete);
3238 mutex_unlock(&ctx->uring_lock);
3240 const sigset_t __user *sig;
3241 struct __kernel_timespec __user *ts;
3243 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3244 if (likely(!ret2)) {
3245 min_complete = min(min_complete,
3247 ret2 = io_cqring_wait(ctx, min_complete, sig,
3256 * EBADR indicates that one or more CQE were dropped.
3257 * Once the user has been informed we can clear the bit
3258 * as they are obviously ok with those drops.
3260 if (unlikely(ret2 == -EBADR))
3261 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3270 static const struct file_operations io_uring_fops = {
3271 .release = io_uring_release,
3272 .mmap = io_uring_mmap,
3274 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3275 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3277 .poll = io_uring_poll,
3278 #ifdef CONFIG_PROC_FS
3279 .show_fdinfo = io_uring_show_fdinfo,
3283 bool io_is_uring_fops(struct file *file)
3285 return file->f_op == &io_uring_fops;
3288 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3289 struct io_uring_params *p)
3291 struct io_rings *rings;
3292 size_t size, sq_array_offset;
3294 /* make sure these are sane, as we already accounted them */
3295 ctx->sq_entries = p->sq_entries;
3296 ctx->cq_entries = p->cq_entries;
3298 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3299 if (size == SIZE_MAX)
3302 rings = io_mem_alloc(size);
3307 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3308 rings->sq_ring_mask = p->sq_entries - 1;
3309 rings->cq_ring_mask = p->cq_entries - 1;
3310 rings->sq_ring_entries = p->sq_entries;
3311 rings->cq_ring_entries = p->cq_entries;
3313 if (p->flags & IORING_SETUP_SQE128)
3314 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3316 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3317 if (size == SIZE_MAX) {
3318 io_mem_free(ctx->rings);
3323 ctx->sq_sqes = io_mem_alloc(size);
3324 if (!ctx->sq_sqes) {
3325 io_mem_free(ctx->rings);
3333 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3337 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3341 ret = __io_uring_add_tctx_node(ctx, false);
3346 fd_install(fd, file);
3351 * Allocate an anonymous fd, this is what constitutes the application
3352 * visible backing of an io_uring instance. The application mmaps this
3353 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3354 * we have to tie this fd to a socket for file garbage collection purposes.
3356 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3359 #if defined(CONFIG_UNIX)
3362 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3365 return ERR_PTR(ret);
3368 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3369 O_RDWR | O_CLOEXEC, NULL);
3370 #if defined(CONFIG_UNIX)
3372 sock_release(ctx->ring_sock);
3373 ctx->ring_sock = NULL;
3375 ctx->ring_sock->file = file;
3381 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3382 struct io_uring_params __user *params)
3384 struct io_ring_ctx *ctx;
3390 if (entries > IORING_MAX_ENTRIES) {
3391 if (!(p->flags & IORING_SETUP_CLAMP))
3393 entries = IORING_MAX_ENTRIES;
3397 * Use twice as many entries for the CQ ring. It's possible for the
3398 * application to drive a higher depth than the size of the SQ ring,
3399 * since the sqes are only used at submission time. This allows for
3400 * some flexibility in overcommitting a bit. If the application has
3401 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3402 * of CQ ring entries manually.
3404 p->sq_entries = roundup_pow_of_two(entries);
3405 if (p->flags & IORING_SETUP_CQSIZE) {
3407 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3408 * to a power-of-two, if it isn't already. We do NOT impose
3409 * any cq vs sq ring sizing.
3413 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3414 if (!(p->flags & IORING_SETUP_CLAMP))
3416 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3418 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3419 if (p->cq_entries < p->sq_entries)
3422 p->cq_entries = 2 * p->sq_entries;
3425 ctx = io_ring_ctx_alloc(p);
3430 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3431 * space applications don't need to do io completion events
3432 * polling again, they can rely on io_sq_thread to do polling
3433 * work, which can reduce cpu usage and uring_lock contention.
3435 if (ctx->flags & IORING_SETUP_IOPOLL &&
3436 !(ctx->flags & IORING_SETUP_SQPOLL))
3437 ctx->syscall_iopoll = 1;
3439 ctx->compat = in_compat_syscall();
3440 if (!capable(CAP_IPC_LOCK))
3441 ctx->user = get_uid(current_user());
3444 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3445 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3448 if (ctx->flags & IORING_SETUP_SQPOLL) {
3449 /* IPI related flags don't make sense with SQPOLL */
3450 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3451 IORING_SETUP_TASKRUN_FLAG |
3452 IORING_SETUP_DEFER_TASKRUN))
3454 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3455 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3456 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3458 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3459 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3461 ctx->notify_method = TWA_SIGNAL;
3465 * For DEFER_TASKRUN we require the completion task to be the same as the
3466 * submission task. This implies that there is only one submitter, so enforce
3469 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3470 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3475 * This is just grabbed for accounting purposes. When a process exits,
3476 * the mm is exited and dropped before the files, hence we need to hang
3477 * on to this mm purely for the purposes of being able to unaccount
3478 * memory (locked/pinned vm). It's not used for anything else.
3480 mmgrab(current->mm);
3481 ctx->mm_account = current->mm;
3483 ret = io_allocate_scq_urings(ctx, p);
3487 ret = io_sq_offload_create(ctx, p);
3490 /* always set a rsrc node */
3491 ret = io_rsrc_node_switch_start(ctx);
3494 io_rsrc_node_switch(ctx, NULL);
3496 memset(&p->sq_off, 0, sizeof(p->sq_off));
3497 p->sq_off.head = offsetof(struct io_rings, sq.head);
3498 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3499 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3500 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3501 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3502 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3503 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3505 memset(&p->cq_off, 0, sizeof(p->cq_off));
3506 p->cq_off.head = offsetof(struct io_rings, cq.head);
3507 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3508 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3509 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3510 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3511 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3512 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3514 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3515 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3516 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3517 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3518 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3519 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3520 IORING_FEAT_LINKED_FILE;
3522 if (copy_to_user(params, p, sizeof(*p))) {
3527 file = io_uring_get_file(ctx);
3529 ret = PTR_ERR(file);
3534 * Install ring fd as the very last thing, so we don't risk someone
3535 * having closed it before we finish setup
3537 ret = io_uring_install_fd(ctx, file);
3539 /* fput will clean it up */
3544 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3547 io_ring_ctx_wait_and_kill(ctx);
3552 * Sets up an aio uring context, and returns the fd. Applications asks for a
3553 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3554 * params structure passed in.
3556 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3558 struct io_uring_params p;
3561 if (copy_from_user(&p, params, sizeof(p)))
3563 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3568 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3569 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3570 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3571 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3572 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3573 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3574 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN))
3577 return io_uring_create(entries, &p, params);
3580 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3581 struct io_uring_params __user *, params)
3583 return io_uring_setup(entries, params);
3586 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3589 struct io_uring_probe *p;
3593 size = struct_size(p, ops, nr_args);
3594 if (size == SIZE_MAX)
3596 p = kzalloc(size, GFP_KERNEL);
3601 if (copy_from_user(p, arg, size))
3604 if (memchr_inv(p, 0, size))
3607 p->last_op = IORING_OP_LAST - 1;
3608 if (nr_args > IORING_OP_LAST)
3609 nr_args = IORING_OP_LAST;
3611 for (i = 0; i < nr_args; i++) {
3613 if (!io_op_defs[i].not_supported)
3614 p->ops[i].flags = IO_URING_OP_SUPPORTED;
3619 if (copy_to_user(arg, p, size))
3626 static int io_register_personality(struct io_ring_ctx *ctx)
3628 const struct cred *creds;
3632 creds = get_current_cred();
3634 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
3635 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
3643 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
3644 void __user *arg, unsigned int nr_args)
3646 struct io_uring_restriction *res;
3650 /* Restrictions allowed only if rings started disabled */
3651 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3654 /* We allow only a single restrictions registration */
3655 if (ctx->restrictions.registered)
3658 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
3661 size = array_size(nr_args, sizeof(*res));
3662 if (size == SIZE_MAX)
3665 res = memdup_user(arg, size);
3667 return PTR_ERR(res);
3671 for (i = 0; i < nr_args; i++) {
3672 switch (res[i].opcode) {
3673 case IORING_RESTRICTION_REGISTER_OP:
3674 if (res[i].register_op >= IORING_REGISTER_LAST) {
3679 __set_bit(res[i].register_op,
3680 ctx->restrictions.register_op);
3682 case IORING_RESTRICTION_SQE_OP:
3683 if (res[i].sqe_op >= IORING_OP_LAST) {
3688 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
3690 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
3691 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
3693 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
3694 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
3703 /* Reset all restrictions if an error happened */
3705 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
3707 ctx->restrictions.registered = true;
3713 static int io_register_enable_rings(struct io_ring_ctx *ctx)
3715 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3718 if (ctx->restrictions.registered)
3719 ctx->restricted = 1;
3721 ctx->flags &= ~IORING_SETUP_R_DISABLED;
3722 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
3723 wake_up(&ctx->sq_data->wait);
3727 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
3728 void __user *arg, unsigned len)
3730 struct io_uring_task *tctx = current->io_uring;
3731 cpumask_var_t new_mask;
3734 if (!tctx || !tctx->io_wq)
3737 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3740 cpumask_clear(new_mask);
3741 if (len > cpumask_size())
3742 len = cpumask_size();
3744 if (in_compat_syscall()) {
3745 ret = compat_get_bitmap(cpumask_bits(new_mask),
3746 (const compat_ulong_t __user *)arg,
3747 len * 8 /* CHAR_BIT */);
3749 ret = copy_from_user(new_mask, arg, len);
3753 free_cpumask_var(new_mask);
3757 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
3758 free_cpumask_var(new_mask);
3762 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
3764 struct io_uring_task *tctx = current->io_uring;
3766 if (!tctx || !tctx->io_wq)
3769 return io_wq_cpu_affinity(tctx->io_wq, NULL);
3772 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
3774 __must_hold(&ctx->uring_lock)
3776 struct io_tctx_node *node;
3777 struct io_uring_task *tctx = NULL;
3778 struct io_sq_data *sqd = NULL;
3782 if (copy_from_user(new_count, arg, sizeof(new_count)))
3784 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3785 if (new_count[i] > INT_MAX)
3788 if (ctx->flags & IORING_SETUP_SQPOLL) {
3792 * Observe the correct sqd->lock -> ctx->uring_lock
3793 * ordering. Fine to drop uring_lock here, we hold
3796 refcount_inc(&sqd->refs);
3797 mutex_unlock(&ctx->uring_lock);
3798 mutex_lock(&sqd->lock);
3799 mutex_lock(&ctx->uring_lock);
3801 tctx = sqd->thread->io_uring;
3804 tctx = current->io_uring;
3807 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
3809 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3811 ctx->iowq_limits[i] = new_count[i];
3812 ctx->iowq_limits_set = true;
3814 if (tctx && tctx->io_wq) {
3815 ret = io_wq_max_workers(tctx->io_wq, new_count);
3819 memset(new_count, 0, sizeof(new_count));
3823 mutex_unlock(&sqd->lock);
3824 io_put_sq_data(sqd);
3827 if (copy_to_user(arg, new_count, sizeof(new_count)))
3830 /* that's it for SQPOLL, only the SQPOLL task creates requests */
3834 /* now propagate the restriction to all registered users */
3835 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3836 struct io_uring_task *tctx = node->task->io_uring;
3838 if (WARN_ON_ONCE(!tctx->io_wq))
3841 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3842 new_count[i] = ctx->iowq_limits[i];
3843 /* ignore errors, it always returns zero anyway */
3844 (void)io_wq_max_workers(tctx->io_wq, new_count);
3849 mutex_unlock(&sqd->lock);
3850 io_put_sq_data(sqd);
3855 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
3856 void __user *arg, unsigned nr_args)
3857 __releases(ctx->uring_lock)
3858 __acquires(ctx->uring_lock)
3863 * We don't quiesce the refs for register anymore and so it can't be
3864 * dying as we're holding a file ref here.
3866 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
3869 if (ctx->restricted) {
3870 if (opcode >= IORING_REGISTER_LAST)
3872 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
3873 if (!test_bit(opcode, ctx->restrictions.register_op))
3878 case IORING_REGISTER_BUFFERS:
3882 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
3884 case IORING_UNREGISTER_BUFFERS:
3888 ret = io_sqe_buffers_unregister(ctx);
3890 case IORING_REGISTER_FILES:
3894 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
3896 case IORING_UNREGISTER_FILES:
3900 ret = io_sqe_files_unregister(ctx);
3902 case IORING_REGISTER_FILES_UPDATE:
3903 ret = io_register_files_update(ctx, arg, nr_args);
3905 case IORING_REGISTER_EVENTFD:
3909 ret = io_eventfd_register(ctx, arg, 0);
3911 case IORING_REGISTER_EVENTFD_ASYNC:
3915 ret = io_eventfd_register(ctx, arg, 1);
3917 case IORING_UNREGISTER_EVENTFD:
3921 ret = io_eventfd_unregister(ctx);
3923 case IORING_REGISTER_PROBE:
3925 if (!arg || nr_args > 256)
3927 ret = io_probe(ctx, arg, nr_args);
3929 case IORING_REGISTER_PERSONALITY:
3933 ret = io_register_personality(ctx);
3935 case IORING_UNREGISTER_PERSONALITY:
3939 ret = io_unregister_personality(ctx, nr_args);
3941 case IORING_REGISTER_ENABLE_RINGS:
3945 ret = io_register_enable_rings(ctx);
3947 case IORING_REGISTER_RESTRICTIONS:
3948 ret = io_register_restrictions(ctx, arg, nr_args);
3950 case IORING_REGISTER_FILES2:
3951 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
3953 case IORING_REGISTER_FILES_UPDATE2:
3954 ret = io_register_rsrc_update(ctx, arg, nr_args,
3957 case IORING_REGISTER_BUFFERS2:
3958 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
3960 case IORING_REGISTER_BUFFERS_UPDATE:
3961 ret = io_register_rsrc_update(ctx, arg, nr_args,
3962 IORING_RSRC_BUFFER);
3964 case IORING_REGISTER_IOWQ_AFF:
3966 if (!arg || !nr_args)
3968 ret = io_register_iowq_aff(ctx, arg, nr_args);
3970 case IORING_UNREGISTER_IOWQ_AFF:
3974 ret = io_unregister_iowq_aff(ctx);
3976 case IORING_REGISTER_IOWQ_MAX_WORKERS:
3978 if (!arg || nr_args != 2)
3980 ret = io_register_iowq_max_workers(ctx, arg);
3982 case IORING_REGISTER_RING_FDS:
3983 ret = io_ringfd_register(ctx, arg, nr_args);
3985 case IORING_UNREGISTER_RING_FDS:
3986 ret = io_ringfd_unregister(ctx, arg, nr_args);
3988 case IORING_REGISTER_PBUF_RING:
3990 if (!arg || nr_args != 1)
3992 ret = io_register_pbuf_ring(ctx, arg);
3994 case IORING_UNREGISTER_PBUF_RING:
3996 if (!arg || nr_args != 1)
3998 ret = io_unregister_pbuf_ring(ctx, arg);
4000 case IORING_REGISTER_SYNC_CANCEL:
4002 if (!arg || nr_args != 1)
4004 ret = io_sync_cancel(ctx, arg);
4006 case IORING_REGISTER_FILE_ALLOC_RANGE:
4008 if (!arg || nr_args)
4010 ret = io_register_file_alloc_range(ctx, arg);
4020 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
4021 void __user *, arg, unsigned int, nr_args)
4023 struct io_ring_ctx *ctx;
4032 if (!io_is_uring_fops(f.file))
4035 ctx = f.file->private_data;
4037 io_run_task_work_ctx(ctx);
4039 mutex_lock(&ctx->uring_lock);
4040 ret = __io_uring_register(ctx, opcode, arg, nr_args);
4041 mutex_unlock(&ctx->uring_lock);
4042 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
4048 static int __init io_uring_init(void)
4050 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4051 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4052 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4055 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4056 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4057 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4058 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4059 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4060 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
4061 BUILD_BUG_SQE_ELEM(1, __u8, flags);
4062 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
4063 BUILD_BUG_SQE_ELEM(4, __s32, fd);
4064 BUILD_BUG_SQE_ELEM(8, __u64, off);
4065 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
4066 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
4067 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
4068 BUILD_BUG_SQE_ELEM(16, __u64, addr);
4069 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
4070 BUILD_BUG_SQE_ELEM(24, __u32, len);
4071 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
4072 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
4073 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4074 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
4075 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
4076 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
4077 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
4078 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
4079 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
4080 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
4081 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
4082 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
4083 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
4084 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
4085 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
4086 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
4087 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
4088 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
4089 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
4090 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
4091 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
4092 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
4093 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
4094 BUILD_BUG_SQE_ELEM(42, __u16, personality);
4095 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
4096 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
4097 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
4098 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
4099 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
4100 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4101 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
4103 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4104 sizeof(struct io_uring_rsrc_update));
4105 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4106 sizeof(struct io_uring_rsrc_update2));
4108 /* ->buf_index is u16 */
4109 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4110 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4111 offsetof(struct io_uring_buf_ring, tail));
4113 /* should fit into one byte */
4114 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
4115 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4116 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4118 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
4120 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4122 io_uring_optable_init();
4124 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4128 __initcall(io_uring_init);