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>
96 #define IORING_MAX_ENTRIES 32768
97 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
99 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
100 IORING_REGISTER_LAST + IORING_OP_LAST)
102 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
103 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
105 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
106 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
108 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
109 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
112 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
115 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
117 #define IO_COMPL_BATCH 32
118 #define IO_REQ_ALLOC_BATCH 8
121 IO_CHECK_CQ_OVERFLOW_BIT,
122 IO_CHECK_CQ_DROPPED_BIT,
125 struct io_defer_entry {
126 struct list_head list;
127 struct io_kiocb *req;
131 /* requests with any of those set should undergo io_disarm_next() */
132 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
133 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
135 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
136 struct task_struct *task,
139 static void io_dismantle_req(struct io_kiocb *req);
140 static void io_clean_op(struct io_kiocb *req);
141 static void io_queue_sqe(struct io_kiocb *req);
143 static void __io_submit_flush_completions(struct io_ring_ctx *ctx);
145 static struct kmem_cache *req_cachep;
147 struct sock *io_uring_get_socket(struct file *file)
149 #if defined(CONFIG_UNIX)
150 if (io_is_uring_fops(file)) {
151 struct io_ring_ctx *ctx = file->private_data;
153 return ctx->ring_sock->sk;
158 EXPORT_SYMBOL(io_uring_get_socket);
160 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
162 if (!wq_list_empty(&ctx->submit_state.compl_reqs))
163 __io_submit_flush_completions(ctx);
166 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
168 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
171 static bool io_match_linked(struct io_kiocb *head)
173 struct io_kiocb *req;
175 io_for_each_link(req, head) {
176 if (req->flags & REQ_F_INFLIGHT)
183 * As io_match_task() but protected against racing with linked timeouts.
184 * User must not hold timeout_lock.
186 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
191 if (task && head->task != task)
196 if (head->flags & REQ_F_LINK_TIMEOUT) {
197 struct io_ring_ctx *ctx = head->ctx;
199 /* protect against races with linked timeouts */
200 spin_lock_irq(&ctx->timeout_lock);
201 matched = io_match_linked(head);
202 spin_unlock_irq(&ctx->timeout_lock);
204 matched = io_match_linked(head);
209 static inline void req_fail_link_node(struct io_kiocb *req, int res)
212 io_req_set_res(req, res, 0);
215 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
217 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
220 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
222 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
224 complete(&ctx->ref_comp);
227 static __cold void io_fallback_req_func(struct work_struct *work)
229 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
231 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
232 struct io_kiocb *req, *tmp;
235 percpu_ref_get(&ctx->refs);
236 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
237 req->io_task_work.func(req, &locked);
240 io_submit_flush_completions(ctx);
241 mutex_unlock(&ctx->uring_lock);
243 percpu_ref_put(&ctx->refs);
246 static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
248 unsigned hash_buckets = 1U << bits;
249 size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
251 table->hbs = kmalloc(hash_size, GFP_KERNEL);
255 table->hash_bits = bits;
256 init_hash_table(table, hash_buckets);
260 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
262 struct io_ring_ctx *ctx;
265 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
269 xa_init(&ctx->io_bl_xa);
272 * Use 5 bits less than the max cq entries, that should give us around
273 * 32 entries per hash list if totally full and uniformly spread, but
274 * don't keep too many buckets to not overconsume memory.
276 hash_bits = ilog2(p->cq_entries) - 5;
277 hash_bits = clamp(hash_bits, 1, 8);
278 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
280 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
283 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
284 if (!ctx->dummy_ubuf)
286 /* set invalid range, so io_import_fixed() fails meeting it */
287 ctx->dummy_ubuf->ubuf = -1UL;
289 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
290 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
293 ctx->flags = p->flags;
294 init_waitqueue_head(&ctx->sqo_sq_wait);
295 INIT_LIST_HEAD(&ctx->sqd_list);
296 INIT_LIST_HEAD(&ctx->cq_overflow_list);
297 INIT_LIST_HEAD(&ctx->io_buffers_cache);
298 INIT_LIST_HEAD(&ctx->apoll_cache);
299 init_completion(&ctx->ref_comp);
300 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
301 mutex_init(&ctx->uring_lock);
302 init_waitqueue_head(&ctx->cq_wait);
303 spin_lock_init(&ctx->completion_lock);
304 spin_lock_init(&ctx->timeout_lock);
305 INIT_WQ_LIST(&ctx->iopoll_list);
306 INIT_LIST_HEAD(&ctx->io_buffers_pages);
307 INIT_LIST_HEAD(&ctx->io_buffers_comp);
308 INIT_LIST_HEAD(&ctx->defer_list);
309 INIT_LIST_HEAD(&ctx->timeout_list);
310 INIT_LIST_HEAD(&ctx->ltimeout_list);
311 spin_lock_init(&ctx->rsrc_ref_lock);
312 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
313 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
314 init_llist_head(&ctx->rsrc_put_llist);
315 INIT_LIST_HEAD(&ctx->tctx_list);
316 ctx->submit_state.free_list.next = NULL;
317 INIT_WQ_LIST(&ctx->locked_free_list);
318 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
319 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
322 kfree(ctx->dummy_ubuf);
323 kfree(ctx->cancel_table.hbs);
324 kfree(ctx->cancel_table_locked.hbs);
326 xa_destroy(&ctx->io_bl_xa);
331 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
333 struct io_rings *r = ctx->rings;
335 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
339 static bool req_need_defer(struct io_kiocb *req, u32 seq)
341 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
342 struct io_ring_ctx *ctx = req->ctx;
344 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
350 static inline void io_req_track_inflight(struct io_kiocb *req)
352 if (!(req->flags & REQ_F_INFLIGHT)) {
353 req->flags |= REQ_F_INFLIGHT;
354 atomic_inc(&req->task->io_uring->inflight_tracked);
358 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
360 if (WARN_ON_ONCE(!req->link))
363 req->flags &= ~REQ_F_ARM_LTIMEOUT;
364 req->flags |= REQ_F_LINK_TIMEOUT;
366 /* linked timeouts should have two refs once prep'ed */
367 io_req_set_refcount(req);
368 __io_req_set_refcount(req->link, 2);
372 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
374 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
376 return __io_prep_linked_timeout(req);
379 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
381 io_queue_linked_timeout(__io_prep_linked_timeout(req));
384 static inline void io_arm_ltimeout(struct io_kiocb *req)
386 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
387 __io_arm_ltimeout(req);
390 static void io_prep_async_work(struct io_kiocb *req)
392 const struct io_op_def *def = &io_op_defs[req->opcode];
393 struct io_ring_ctx *ctx = req->ctx;
395 if (!(req->flags & REQ_F_CREDS)) {
396 req->flags |= REQ_F_CREDS;
397 req->creds = get_current_cred();
400 req->work.list.next = NULL;
402 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
403 if (req->flags & REQ_F_FORCE_ASYNC)
404 req->work.flags |= IO_WQ_WORK_CONCURRENT;
406 if (req->flags & REQ_F_ISREG) {
407 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
408 io_wq_hash_work(&req->work, file_inode(req->file));
409 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
410 if (def->unbound_nonreg_file)
411 req->work.flags |= IO_WQ_WORK_UNBOUND;
415 static void io_prep_async_link(struct io_kiocb *req)
417 struct io_kiocb *cur;
419 if (req->flags & REQ_F_LINK_TIMEOUT) {
420 struct io_ring_ctx *ctx = req->ctx;
422 spin_lock_irq(&ctx->timeout_lock);
423 io_for_each_link(cur, req)
424 io_prep_async_work(cur);
425 spin_unlock_irq(&ctx->timeout_lock);
427 io_for_each_link(cur, req)
428 io_prep_async_work(cur);
432 void io_queue_iowq(struct io_kiocb *req, bool *dont_use)
434 struct io_kiocb *link = io_prep_linked_timeout(req);
435 struct io_uring_task *tctx = req->task->io_uring;
438 BUG_ON(!tctx->io_wq);
440 /* init ->work of the whole link before punting */
441 io_prep_async_link(req);
444 * Not expected to happen, but if we do have a bug where this _can_
445 * happen, catch it here and ensure the request is marked as
446 * canceled. That will make io-wq go through the usual work cancel
447 * procedure rather than attempt to run this request (or create a new
450 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
451 req->work.flags |= IO_WQ_WORK_CANCEL;
453 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
454 io_wq_enqueue(tctx->io_wq, &req->work);
456 io_queue_linked_timeout(link);
459 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
461 while (!list_empty(&ctx->defer_list)) {
462 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
463 struct io_defer_entry, list);
465 if (req_need_defer(de->req, de->seq))
467 list_del_init(&de->list);
468 io_req_task_queue(de->req);
473 static void io_eventfd_signal(struct io_ring_ctx *ctx)
475 struct io_ev_fd *ev_fd;
479 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
482 ev_fd = rcu_dereference(ctx->io_ev_fd);
485 * Check again if ev_fd exists incase an io_eventfd_unregister call
486 * completed between the NULL check of ctx->io_ev_fd at the start of
487 * the function and rcu_read_lock.
489 if (unlikely(!ev_fd))
491 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
494 if (!ev_fd->eventfd_async || io_wq_current_is_worker())
495 eventfd_signal(ev_fd->cq_ev_fd, 1);
500 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
502 if (ctx->off_timeout_used || ctx->drain_active) {
503 spin_lock(&ctx->completion_lock);
504 if (ctx->off_timeout_used)
505 io_flush_timeouts(ctx);
506 if (ctx->drain_active)
507 io_queue_deferred(ctx);
508 spin_unlock(&ctx->completion_lock);
511 io_eventfd_signal(ctx);
515 * This should only get called when at least one event has been posted.
516 * Some applications rely on the eventfd notification count only changing
517 * IFF a new CQE has been added to the CQ ring. There's no depedency on
518 * 1:1 relationship between how many times this function is called (and
519 * hence the eventfd count) and number of CQEs posted to the CQ ring.
521 void io_cqring_ev_posted(struct io_ring_ctx *ctx)
523 if (unlikely(ctx->off_timeout_used || ctx->drain_active ||
525 __io_commit_cqring_flush(ctx);
530 /* Returns true if there are no backlogged entries after the flush */
531 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
533 bool all_flushed, posted;
534 size_t cqe_size = sizeof(struct io_uring_cqe);
536 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
539 if (ctx->flags & IORING_SETUP_CQE32)
543 spin_lock(&ctx->completion_lock);
544 while (!list_empty(&ctx->cq_overflow_list)) {
545 struct io_uring_cqe *cqe = io_get_cqe(ctx);
546 struct io_overflow_cqe *ocqe;
550 ocqe = list_first_entry(&ctx->cq_overflow_list,
551 struct io_overflow_cqe, list);
553 memcpy(cqe, &ocqe->cqe, cqe_size);
555 io_account_cq_overflow(ctx);
558 list_del(&ocqe->list);
562 all_flushed = list_empty(&ctx->cq_overflow_list);
564 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
565 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
568 io_commit_cqring(ctx);
569 spin_unlock(&ctx->completion_lock);
571 io_cqring_ev_posted(ctx);
575 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
579 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
580 /* iopoll syncs against uring_lock, not completion_lock */
581 if (ctx->flags & IORING_SETUP_IOPOLL)
582 mutex_lock(&ctx->uring_lock);
583 ret = __io_cqring_overflow_flush(ctx, false);
584 if (ctx->flags & IORING_SETUP_IOPOLL)
585 mutex_unlock(&ctx->uring_lock);
591 static void __io_put_task(struct task_struct *task, int nr)
593 struct io_uring_task *tctx = task->io_uring;
595 percpu_counter_sub(&tctx->inflight, nr);
596 if (unlikely(atomic_read(&tctx->in_idle)))
597 wake_up(&tctx->wait);
598 put_task_struct_many(task, nr);
601 /* must to be called somewhat shortly after putting a request */
602 static inline void io_put_task(struct task_struct *task, int nr)
604 if (likely(task == current))
605 task->io_uring->cached_refs += nr;
607 __io_put_task(task, nr);
610 static void io_task_refs_refill(struct io_uring_task *tctx)
612 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
614 percpu_counter_add(&tctx->inflight, refill);
615 refcount_add(refill, ¤t->usage);
616 tctx->cached_refs += refill;
619 static inline void io_get_task_refs(int nr)
621 struct io_uring_task *tctx = current->io_uring;
623 tctx->cached_refs -= nr;
624 if (unlikely(tctx->cached_refs < 0))
625 io_task_refs_refill(tctx);
628 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
630 struct io_uring_task *tctx = task->io_uring;
631 unsigned int refs = tctx->cached_refs;
634 tctx->cached_refs = 0;
635 percpu_counter_sub(&tctx->inflight, refs);
636 put_task_struct_many(task, refs);
640 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
641 s32 res, u32 cflags, u64 extra1, u64 extra2)
643 struct io_overflow_cqe *ocqe;
644 size_t ocq_size = sizeof(struct io_overflow_cqe);
645 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
648 ocq_size += sizeof(struct io_uring_cqe);
650 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
651 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
654 * If we're in ring overflow flush mode, or in task cancel mode,
655 * or cannot allocate an overflow entry, then we need to drop it
658 io_account_cq_overflow(ctx);
659 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
662 if (list_empty(&ctx->cq_overflow_list)) {
663 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
664 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
667 ocqe->cqe.user_data = user_data;
669 ocqe->cqe.flags = cflags;
671 ocqe->cqe.big_cqe[0] = extra1;
672 ocqe->cqe.big_cqe[1] = extra2;
674 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
678 bool io_req_cqe_overflow(struct io_kiocb *req)
680 if (!(req->flags & REQ_F_CQE32_INIT)) {
684 return io_cqring_event_overflow(req->ctx, req->cqe.user_data,
685 req->cqe.res, req->cqe.flags,
686 req->extra1, req->extra2);
690 * writes to the cq entry need to come after reading head; the
691 * control dependency is enough as we're using WRITE_ONCE to
694 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx)
696 struct io_rings *rings = ctx->rings;
697 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
698 unsigned int free, queued, len;
701 /* userspace may cheat modifying the tail, be safe and do min */
702 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
703 free = ctx->cq_entries - queued;
704 /* we need a contiguous range, limit based on the current array offset */
705 len = min(free, ctx->cq_entries - off);
709 if (ctx->flags & IORING_SETUP_CQE32) {
714 ctx->cqe_cached = &rings->cqes[off];
715 ctx->cqe_sentinel = ctx->cqe_cached + len;
717 ctx->cached_cq_tail++;
719 if (ctx->flags & IORING_SETUP_CQE32)
721 return &rings->cqes[off];
724 static bool io_fill_cqe_aux(struct io_ring_ctx *ctx,
725 u64 user_data, s32 res, u32 cflags)
727 struct io_uring_cqe *cqe;
730 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
733 * If we can't get a cq entry, userspace overflowed the
734 * submission (by quite a lot). Increment the overflow count in
737 cqe = io_get_cqe(ctx);
739 WRITE_ONCE(cqe->user_data, user_data);
740 WRITE_ONCE(cqe->res, res);
741 WRITE_ONCE(cqe->flags, cflags);
743 if (ctx->flags & IORING_SETUP_CQE32) {
744 WRITE_ONCE(cqe->big_cqe[0], 0);
745 WRITE_ONCE(cqe->big_cqe[1], 0);
749 return io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
752 bool io_post_aux_cqe(struct io_ring_ctx *ctx,
753 u64 user_data, s32 res, u32 cflags)
757 spin_lock(&ctx->completion_lock);
758 filled = io_fill_cqe_aux(ctx, user_data, res, cflags);
759 io_commit_cqring(ctx);
760 spin_unlock(&ctx->completion_lock);
762 io_cqring_ev_posted(ctx);
766 static void __io_req_complete_put(struct io_kiocb *req)
769 * If we're the last reference to this request, add to our locked
772 if (req_ref_put_and_test(req)) {
773 struct io_ring_ctx *ctx = req->ctx;
775 if (req->flags & IO_REQ_LINK_FLAGS) {
776 if (req->flags & IO_DISARM_MASK)
779 io_req_task_queue(req->link);
783 io_req_put_rsrc(req);
785 * Selected buffer deallocation in io_clean_op() assumes that
786 * we don't hold ->completion_lock. Clean them here to avoid
789 io_put_kbuf_comp(req);
790 io_dismantle_req(req);
791 io_put_task(req->task, 1);
792 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
793 ctx->locked_free_nr++;
797 void __io_req_complete_post(struct io_kiocb *req)
799 if (!(req->flags & REQ_F_CQE_SKIP))
800 __io_fill_cqe_req(req->ctx, req);
801 __io_req_complete_put(req);
804 void io_req_complete_post(struct io_kiocb *req)
806 struct io_ring_ctx *ctx = req->ctx;
808 spin_lock(&ctx->completion_lock);
809 __io_req_complete_post(req);
810 io_commit_cqring(ctx);
811 spin_unlock(&ctx->completion_lock);
812 io_cqring_ev_posted(ctx);
815 inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags)
817 io_req_complete_post(req);
820 void io_req_complete_failed(struct io_kiocb *req, s32 res)
823 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
824 io_req_complete_post(req);
828 * Don't initialise the fields below on every allocation, but do that in
829 * advance and keep them valid across allocations.
831 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
835 req->async_data = NULL;
836 /* not necessary, but safer to zero */
840 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
841 struct io_submit_state *state)
843 spin_lock(&ctx->completion_lock);
844 wq_list_splice(&ctx->locked_free_list, &state->free_list);
845 ctx->locked_free_nr = 0;
846 spin_unlock(&ctx->completion_lock);
849 static inline bool io_req_cache_empty(struct io_ring_ctx *ctx)
851 return !ctx->submit_state.free_list.next;
855 * A request might get retired back into the request caches even before opcode
856 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
857 * Because of that, io_alloc_req() should be called only under ->uring_lock
858 * and with extra caution to not get a request that is still worked on.
860 static __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
861 __must_hold(&ctx->uring_lock)
863 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
864 void *reqs[IO_REQ_ALLOC_BATCH];
868 * If we have more than a batch's worth of requests in our IRQ side
869 * locked cache, grab the lock and move them over to our submission
872 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
873 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
874 if (!io_req_cache_empty(ctx))
878 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
881 * Bulk alloc is all-or-nothing. If we fail to get a batch,
882 * retry single alloc to be on the safe side.
884 if (unlikely(ret <= 0)) {
885 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
891 percpu_ref_get_many(&ctx->refs, ret);
892 for (i = 0; i < ret; i++) {
893 struct io_kiocb *req = reqs[i];
895 io_preinit_req(req, ctx);
896 io_req_add_to_cache(req, ctx);
901 static inline bool io_alloc_req_refill(struct io_ring_ctx *ctx)
903 if (unlikely(io_req_cache_empty(ctx)))
904 return __io_alloc_req_refill(ctx);
908 static inline struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
910 struct io_wq_work_node *node;
912 node = wq_stack_extract(&ctx->submit_state.free_list);
913 return container_of(node, struct io_kiocb, comp_list);
916 static inline void io_dismantle_req(struct io_kiocb *req)
918 unsigned int flags = req->flags;
920 if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
922 if (!(flags & REQ_F_FIXED_FILE))
923 io_put_file(req->file);
926 __cold void io_free_req(struct io_kiocb *req)
928 struct io_ring_ctx *ctx = req->ctx;
930 io_req_put_rsrc(req);
931 io_dismantle_req(req);
932 io_put_task(req->task, 1);
934 spin_lock(&ctx->completion_lock);
935 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
936 ctx->locked_free_nr++;
937 spin_unlock(&ctx->completion_lock);
940 static void __io_req_find_next_prep(struct io_kiocb *req)
942 struct io_ring_ctx *ctx = req->ctx;
945 spin_lock(&ctx->completion_lock);
946 posted = io_disarm_next(req);
947 io_commit_cqring(ctx);
948 spin_unlock(&ctx->completion_lock);
950 io_cqring_ev_posted(ctx);
953 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
955 struct io_kiocb *nxt;
958 * If LINK is set, we have dependent requests in this chain. If we
959 * didn't fail this request, queue the first one up, moving any other
960 * dependencies to the next request. In case of failure, fail the rest
963 if (unlikely(req->flags & IO_DISARM_MASK))
964 __io_req_find_next_prep(req);
970 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
974 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
975 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
977 io_submit_flush_completions(ctx);
978 mutex_unlock(&ctx->uring_lock);
981 percpu_ref_put(&ctx->refs);
984 static inline void ctx_commit_and_unlock(struct io_ring_ctx *ctx)
986 io_commit_cqring(ctx);
987 spin_unlock(&ctx->completion_lock);
988 io_cqring_ev_posted(ctx);
991 static void handle_prev_tw_list(struct io_wq_work_node *node,
992 struct io_ring_ctx **ctx, bool *uring_locked)
994 if (*ctx && !*uring_locked)
995 spin_lock(&(*ctx)->completion_lock);
998 struct io_wq_work_node *next = node->next;
999 struct io_kiocb *req = container_of(node, struct io_kiocb,
1002 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1004 if (req->ctx != *ctx) {
1005 if (unlikely(!*uring_locked && *ctx))
1006 ctx_commit_and_unlock(*ctx);
1008 ctx_flush_and_put(*ctx, uring_locked);
1010 /* if not contended, grab and improve batching */
1011 *uring_locked = mutex_trylock(&(*ctx)->uring_lock);
1012 percpu_ref_get(&(*ctx)->refs);
1013 if (unlikely(!*uring_locked))
1014 spin_lock(&(*ctx)->completion_lock);
1016 if (likely(*uring_locked)) {
1017 req->io_task_work.func(req, uring_locked);
1019 req->cqe.flags = io_put_kbuf_comp(req);
1020 __io_req_complete_post(req);
1025 if (unlikely(!*uring_locked))
1026 ctx_commit_and_unlock(*ctx);
1029 static void handle_tw_list(struct io_wq_work_node *node,
1030 struct io_ring_ctx **ctx, bool *locked)
1033 struct io_wq_work_node *next = node->next;
1034 struct io_kiocb *req = container_of(node, struct io_kiocb,
1037 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1039 if (req->ctx != *ctx) {
1040 ctx_flush_and_put(*ctx, locked);
1042 /* if not contended, grab and improve batching */
1043 *locked = mutex_trylock(&(*ctx)->uring_lock);
1044 percpu_ref_get(&(*ctx)->refs);
1046 req->io_task_work.func(req, locked);
1051 void tctx_task_work(struct callback_head *cb)
1053 bool uring_locked = false;
1054 struct io_ring_ctx *ctx = NULL;
1055 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1059 struct io_wq_work_node *node1, *node2;
1061 spin_lock_irq(&tctx->task_lock);
1062 node1 = tctx->prio_task_list.first;
1063 node2 = tctx->task_list.first;
1064 INIT_WQ_LIST(&tctx->task_list);
1065 INIT_WQ_LIST(&tctx->prio_task_list);
1066 if (!node2 && !node1)
1067 tctx->task_running = false;
1068 spin_unlock_irq(&tctx->task_lock);
1069 if (!node2 && !node1)
1073 handle_prev_tw_list(node1, &ctx, &uring_locked);
1075 handle_tw_list(node2, &ctx, &uring_locked);
1078 if (data_race(!tctx->task_list.first) &&
1079 data_race(!tctx->prio_task_list.first) && uring_locked)
1080 io_submit_flush_completions(ctx);
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);
1090 static void __io_req_task_work_add(struct io_kiocb *req,
1091 struct io_uring_task *tctx,
1092 struct io_wq_work_list *list)
1094 struct io_ring_ctx *ctx = req->ctx;
1095 struct io_wq_work_node *node;
1096 unsigned long flags;
1099 spin_lock_irqsave(&tctx->task_lock, flags);
1100 wq_list_add_tail(&req->io_task_work.node, list);
1101 running = tctx->task_running;
1103 tctx->task_running = true;
1104 spin_unlock_irqrestore(&tctx->task_lock, flags);
1106 /* task_work already pending, we're done */
1110 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1111 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1113 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1116 spin_lock_irqsave(&tctx->task_lock, flags);
1117 tctx->task_running = false;
1118 node = wq_list_merge(&tctx->prio_task_list, &tctx->task_list);
1119 spin_unlock_irqrestore(&tctx->task_lock, flags);
1122 req = container_of(node, struct io_kiocb, io_task_work.node);
1124 if (llist_add(&req->io_task_work.fallback_node,
1125 &req->ctx->fallback_llist))
1126 schedule_delayed_work(&req->ctx->fallback_work, 1);
1130 void io_req_task_work_add(struct io_kiocb *req)
1132 struct io_uring_task *tctx = req->task->io_uring;
1134 __io_req_task_work_add(req, tctx, &tctx->task_list);
1137 void io_req_task_prio_work_add(struct io_kiocb *req)
1139 struct io_uring_task *tctx = req->task->io_uring;
1141 if (req->ctx->flags & IORING_SETUP_SQPOLL)
1142 __io_req_task_work_add(req, tctx, &tctx->prio_task_list);
1144 __io_req_task_work_add(req, tctx, &tctx->task_list);
1147 static void io_req_tw_post(struct io_kiocb *req, bool *locked)
1149 io_req_complete_post(req);
1152 void io_req_tw_post_queue(struct io_kiocb *req, s32 res, u32 cflags)
1154 io_req_set_res(req, res, cflags);
1155 req->io_task_work.func = io_req_tw_post;
1156 io_req_task_work_add(req);
1159 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
1161 /* not needed for normal modes, but SQPOLL depends on it */
1162 io_tw_lock(req->ctx, locked);
1163 io_req_complete_failed(req, req->cqe.res);
1166 void io_req_task_submit(struct io_kiocb *req, bool *locked)
1168 io_tw_lock(req->ctx, locked);
1169 /* req->task == current here, checking PF_EXITING is safe */
1170 if (likely(!(req->task->flags & PF_EXITING)))
1173 io_req_complete_failed(req, -EFAULT);
1176 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1178 io_req_set_res(req, ret, 0);
1179 req->io_task_work.func = io_req_task_cancel;
1180 io_req_task_work_add(req);
1183 void io_req_task_queue(struct io_kiocb *req)
1185 req->io_task_work.func = io_req_task_submit;
1186 io_req_task_work_add(req);
1189 void io_queue_next(struct io_kiocb *req)
1191 struct io_kiocb *nxt = io_req_find_next(req);
1194 io_req_task_queue(nxt);
1197 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1198 __must_hold(&ctx->uring_lock)
1200 struct task_struct *task = NULL;
1204 struct io_kiocb *req = container_of(node, struct io_kiocb,
1207 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1208 if (req->flags & REQ_F_REFCOUNT) {
1209 node = req->comp_list.next;
1210 if (!req_ref_put_and_test(req))
1213 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1214 struct async_poll *apoll = req->apoll;
1216 if (apoll->double_poll)
1217 kfree(apoll->double_poll);
1218 list_add(&apoll->poll.wait.entry,
1220 req->flags &= ~REQ_F_POLLED;
1222 if (req->flags & IO_REQ_LINK_FLAGS)
1224 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1227 if (!(req->flags & REQ_F_FIXED_FILE))
1228 io_put_file(req->file);
1230 io_req_put_rsrc_locked(req, ctx);
1232 if (req->task != task) {
1234 io_put_task(task, task_refs);
1239 node = req->comp_list.next;
1240 io_req_add_to_cache(req, ctx);
1244 io_put_task(task, task_refs);
1247 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1248 __must_hold(&ctx->uring_lock)
1250 struct io_wq_work_node *node, *prev;
1251 struct io_submit_state *state = &ctx->submit_state;
1253 spin_lock(&ctx->completion_lock);
1254 wq_list_for_each(node, prev, &state->compl_reqs) {
1255 struct io_kiocb *req = container_of(node, struct io_kiocb,
1258 if (!(req->flags & REQ_F_CQE_SKIP))
1259 __io_fill_cqe_req(ctx, req);
1262 io_commit_cqring(ctx);
1263 spin_unlock(&ctx->completion_lock);
1264 io_cqring_ev_posted(ctx);
1266 io_free_batch_list(ctx, state->compl_reqs.first);
1267 INIT_WQ_LIST(&state->compl_reqs);
1271 * Drop reference to request, return next in chain (if there is one) if this
1272 * was the last reference to this request.
1274 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1276 struct io_kiocb *nxt = NULL;
1278 if (req_ref_put_and_test(req)) {
1279 if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
1280 nxt = io_req_find_next(req);
1286 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1288 /* See comment at the top of this file */
1290 return __io_cqring_events(ctx);
1294 * We can't just wait for polled events to come to us, we have to actively
1295 * find and complete them.
1297 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1299 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1302 mutex_lock(&ctx->uring_lock);
1303 while (!wq_list_empty(&ctx->iopoll_list)) {
1304 /* let it sleep and repeat later if can't complete a request */
1305 if (io_do_iopoll(ctx, true) == 0)
1308 * Ensure we allow local-to-the-cpu processing to take place,
1309 * in this case we need to ensure that we reap all events.
1310 * Also let task_work, etc. to progress by releasing the mutex
1312 if (need_resched()) {
1313 mutex_unlock(&ctx->uring_lock);
1315 mutex_lock(&ctx->uring_lock);
1318 mutex_unlock(&ctx->uring_lock);
1321 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1323 unsigned int nr_events = 0;
1325 unsigned long check_cq;
1327 check_cq = READ_ONCE(ctx->check_cq);
1328 if (unlikely(check_cq)) {
1329 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1330 __io_cqring_overflow_flush(ctx, false);
1332 * Similarly do not spin if we have not informed the user of any
1335 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1339 * Don't enter poll loop if we already have events pending.
1340 * If we do, we can potentially be spinning for commands that
1341 * already triggered a CQE (eg in error).
1343 if (io_cqring_events(ctx))
1348 * If a submit got punted to a workqueue, we can have the
1349 * application entering polling for a command before it gets
1350 * issued. That app will hold the uring_lock for the duration
1351 * of the poll right here, so we need to take a breather every
1352 * now and then to ensure that the issue has a chance to add
1353 * the poll to the issued list. Otherwise we can spin here
1354 * forever, while the workqueue is stuck trying to acquire the
1357 if (wq_list_empty(&ctx->iopoll_list)) {
1358 u32 tail = ctx->cached_cq_tail;
1360 mutex_unlock(&ctx->uring_lock);
1362 mutex_lock(&ctx->uring_lock);
1364 /* some requests don't go through iopoll_list */
1365 if (tail != ctx->cached_cq_tail ||
1366 wq_list_empty(&ctx->iopoll_list))
1369 ret = io_do_iopoll(ctx, !min);
1374 } while (nr_events < min && !need_resched());
1379 void io_req_task_complete(struct io_kiocb *req, bool *locked)
1381 if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) {
1382 unsigned issue_flags = *locked ? 0 : IO_URING_F_UNLOCKED;
1384 req->cqe.flags |= io_put_kbuf(req, issue_flags);
1388 io_req_add_compl_list(req);
1390 io_req_complete_post(req);
1394 * After the iocb has been issued, it's safe to be found on the poll list.
1395 * Adding the kiocb to the list AFTER submission ensures that we don't
1396 * find it from a io_do_iopoll() thread before the issuer is done
1397 * accessing the kiocb cookie.
1399 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1401 struct io_ring_ctx *ctx = req->ctx;
1402 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1404 /* workqueue context doesn't hold uring_lock, grab it now */
1405 if (unlikely(needs_lock))
1406 mutex_lock(&ctx->uring_lock);
1409 * Track whether we have multiple files in our lists. This will impact
1410 * how we do polling eventually, not spinning if we're on potentially
1411 * different devices.
1413 if (wq_list_empty(&ctx->iopoll_list)) {
1414 ctx->poll_multi_queue = false;
1415 } else if (!ctx->poll_multi_queue) {
1416 struct io_kiocb *list_req;
1418 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1420 if (list_req->file != req->file)
1421 ctx->poll_multi_queue = true;
1425 * For fast devices, IO may have already completed. If it has, add
1426 * it to the front so we find it first.
1428 if (READ_ONCE(req->iopoll_completed))
1429 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1431 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1433 if (unlikely(needs_lock)) {
1435 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1436 * in sq thread task context or in io worker task context. If
1437 * current task context is sq thread, we don't need to check
1438 * whether should wake up sq thread.
1440 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1441 wq_has_sleeper(&ctx->sq_data->wait))
1442 wake_up(&ctx->sq_data->wait);
1444 mutex_unlock(&ctx->uring_lock);
1448 static bool io_bdev_nowait(struct block_device *bdev)
1450 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
1454 * If we tracked the file through the SCM inflight mechanism, we could support
1455 * any file. For now, just ensure that anything potentially problematic is done
1458 static bool __io_file_supports_nowait(struct file *file, umode_t mode)
1460 if (S_ISBLK(mode)) {
1461 if (IS_ENABLED(CONFIG_BLOCK) &&
1462 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
1468 if (S_ISREG(mode)) {
1469 if (IS_ENABLED(CONFIG_BLOCK) &&
1470 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
1471 !io_is_uring_fops(file))
1476 /* any ->read/write should understand O_NONBLOCK */
1477 if (file->f_flags & O_NONBLOCK)
1479 return file->f_mode & FMODE_NOWAIT;
1483 * If we tracked the file through the SCM inflight mechanism, we could support
1484 * any file. For now, just ensure that anything potentially problematic is done
1487 unsigned int io_file_get_flags(struct file *file)
1489 umode_t mode = file_inode(file)->i_mode;
1490 unsigned int res = 0;
1494 if (__io_file_supports_nowait(file, mode))
1496 if (io_file_need_scm(file))
1501 bool io_alloc_async_data(struct io_kiocb *req)
1503 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
1504 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
1505 if (req->async_data) {
1506 req->flags |= REQ_F_ASYNC_DATA;
1512 int io_req_prep_async(struct io_kiocb *req)
1514 const struct io_op_def *def = &io_op_defs[req->opcode];
1516 /* assign early for deferred execution for non-fixed file */
1517 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE))
1518 req->file = io_file_get_normal(req, req->cqe.fd);
1519 if (!def->prep_async)
1521 if (WARN_ON_ONCE(req_has_async_data(req)))
1523 if (io_alloc_async_data(req))
1526 return def->prep_async(req);
1529 static u32 io_get_sequence(struct io_kiocb *req)
1531 u32 seq = req->ctx->cached_sq_head;
1532 struct io_kiocb *cur;
1534 /* need original cached_sq_head, but it was increased for each req */
1535 io_for_each_link(cur, req)
1540 static __cold void io_drain_req(struct io_kiocb *req)
1542 struct io_ring_ctx *ctx = req->ctx;
1543 struct io_defer_entry *de;
1545 u32 seq = io_get_sequence(req);
1547 /* Still need defer if there is pending req in defer list. */
1548 spin_lock(&ctx->completion_lock);
1549 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1550 spin_unlock(&ctx->completion_lock);
1552 ctx->drain_active = false;
1553 io_req_task_queue(req);
1556 spin_unlock(&ctx->completion_lock);
1558 ret = io_req_prep_async(req);
1561 io_req_complete_failed(req, ret);
1564 io_prep_async_link(req);
1565 de = kmalloc(sizeof(*de), GFP_KERNEL);
1571 spin_lock(&ctx->completion_lock);
1572 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1573 spin_unlock(&ctx->completion_lock);
1578 trace_io_uring_defer(req);
1581 list_add_tail(&de->list, &ctx->defer_list);
1582 spin_unlock(&ctx->completion_lock);
1585 static void io_clean_op(struct io_kiocb *req)
1587 if (req->flags & REQ_F_BUFFER_SELECTED) {
1588 spin_lock(&req->ctx->completion_lock);
1589 io_put_kbuf_comp(req);
1590 spin_unlock(&req->ctx->completion_lock);
1593 if (req->flags & REQ_F_NEED_CLEANUP) {
1594 const struct io_op_def *def = &io_op_defs[req->opcode];
1599 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1600 kfree(req->apoll->double_poll);
1604 if (req->flags & REQ_F_INFLIGHT) {
1605 struct io_uring_task *tctx = req->task->io_uring;
1607 atomic_dec(&tctx->inflight_tracked);
1609 if (req->flags & REQ_F_CREDS)
1610 put_cred(req->creds);
1611 if (req->flags & REQ_F_ASYNC_DATA) {
1612 kfree(req->async_data);
1613 req->async_data = NULL;
1615 req->flags &= ~IO_REQ_CLEAN_FLAGS;
1618 static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags)
1620 if (req->file || !io_op_defs[req->opcode].needs_file)
1623 if (req->flags & REQ_F_FIXED_FILE)
1624 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1626 req->file = io_file_get_normal(req, req->cqe.fd);
1631 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1633 const struct io_op_def *def = &io_op_defs[req->opcode];
1634 const struct cred *creds = NULL;
1637 if (unlikely(!io_assign_file(req, issue_flags)))
1640 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1641 creds = override_creds(req->creds);
1643 if (!def->audit_skip)
1644 audit_uring_entry(req->opcode);
1646 ret = def->issue(req, issue_flags);
1648 if (!def->audit_skip)
1649 audit_uring_exit(!ret, ret);
1652 revert_creds(creds);
1654 if (ret == IOU_OK) {
1655 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1656 io_req_add_compl_list(req);
1658 io_req_complete_post(req);
1659 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1662 /* If the op doesn't have a file, we're not polling for it */
1663 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && req->file)
1664 io_iopoll_req_issued(req, issue_flags);
1669 int io_poll_issue(struct io_kiocb *req, bool *locked)
1671 io_tw_lock(req->ctx, locked);
1672 if (unlikely(req->task->flags & PF_EXITING))
1674 return io_issue_sqe(req, IO_URING_F_NONBLOCK);
1677 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1679 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1681 req = io_put_req_find_next(req);
1682 return req ? &req->work : NULL;
1685 void io_wq_submit_work(struct io_wq_work *work)
1687 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1688 const struct io_op_def *def = &io_op_defs[req->opcode];
1689 unsigned int issue_flags = IO_URING_F_UNLOCKED;
1690 bool needs_poll = false;
1691 int ret = 0, err = -ECANCELED;
1693 /* one will be dropped by ->io_free_work() after returning to io-wq */
1694 if (!(req->flags & REQ_F_REFCOUNT))
1695 __io_req_set_refcount(req, 2);
1699 io_arm_ltimeout(req);
1701 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1702 if (work->flags & IO_WQ_WORK_CANCEL) {
1704 io_req_task_queue_fail(req, err);
1707 if (!io_assign_file(req, issue_flags)) {
1709 work->flags |= IO_WQ_WORK_CANCEL;
1713 if (req->flags & REQ_F_FORCE_ASYNC) {
1714 bool opcode_poll = def->pollin || def->pollout;
1716 if (opcode_poll && file_can_poll(req->file)) {
1718 issue_flags |= IO_URING_F_NONBLOCK;
1723 ret = io_issue_sqe(req, issue_flags);
1727 * We can get EAGAIN for iopolled IO even though we're
1728 * forcing a sync submission from here, since we can't
1729 * wait for request slots on the block side.
1732 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1738 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1740 /* aborted or ready, in either case retry blocking */
1742 issue_flags &= ~IO_URING_F_NONBLOCK;
1745 /* avoid locking problems by failing it from a clean context */
1747 io_req_task_queue_fail(req, ret);
1750 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1751 unsigned int issue_flags)
1753 struct io_ring_ctx *ctx = req->ctx;
1754 struct file *file = NULL;
1755 unsigned long file_ptr;
1757 io_ring_submit_lock(ctx, issue_flags);
1759 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
1761 fd = array_index_nospec(fd, ctx->nr_user_files);
1762 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
1763 file = (struct file *) (file_ptr & FFS_MASK);
1764 file_ptr &= ~FFS_MASK;
1765 /* mask in overlapping REQ_F and FFS bits */
1766 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
1767 io_req_set_rsrc_node(req, ctx, 0);
1768 WARN_ON_ONCE(file && !test_bit(fd, ctx->file_table.bitmap));
1770 io_ring_submit_unlock(ctx, issue_flags);
1774 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
1776 struct file *file = fget(fd);
1778 trace_io_uring_file_get(req, fd);
1780 /* we don't allow fixed io_uring files */
1781 if (file && io_is_uring_fops(file))
1782 io_req_track_inflight(req);
1786 static void io_queue_async(struct io_kiocb *req, int ret)
1787 __must_hold(&req->ctx->uring_lock)
1789 struct io_kiocb *linked_timeout;
1791 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
1792 io_req_complete_failed(req, ret);
1796 linked_timeout = io_prep_linked_timeout(req);
1798 switch (io_arm_poll_handler(req, 0)) {
1799 case IO_APOLL_READY:
1800 io_req_task_queue(req);
1802 case IO_APOLL_ABORTED:
1804 * Queued up for async execution, worker will release
1805 * submit reference when the iocb is actually submitted.
1807 io_kbuf_recycle(req, 0);
1808 io_queue_iowq(req, NULL);
1815 io_queue_linked_timeout(linked_timeout);
1818 static inline void io_queue_sqe(struct io_kiocb *req)
1819 __must_hold(&req->ctx->uring_lock)
1823 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
1826 * We async punt it if the file wasn't marked NOWAIT, or if the file
1827 * doesn't support non-blocking read/write attempts
1830 io_arm_ltimeout(req);
1832 io_queue_async(req, ret);
1835 static void io_queue_sqe_fallback(struct io_kiocb *req)
1836 __must_hold(&req->ctx->uring_lock)
1838 if (unlikely(req->flags & REQ_F_FAIL)) {
1840 * We don't submit, fail them all, for that replace hardlinks
1841 * with normal links. Extra REQ_F_LINK is tolerated.
1843 req->flags &= ~REQ_F_HARDLINK;
1844 req->flags |= REQ_F_LINK;
1845 io_req_complete_failed(req, req->cqe.res);
1846 } else if (unlikely(req->ctx->drain_active)) {
1849 int ret = io_req_prep_async(req);
1852 io_req_complete_failed(req, ret);
1854 io_queue_iowq(req, NULL);
1859 * Check SQE restrictions (opcode and flags).
1861 * Returns 'true' if SQE is allowed, 'false' otherwise.
1863 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
1864 struct io_kiocb *req,
1865 unsigned int sqe_flags)
1867 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
1870 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
1871 ctx->restrictions.sqe_flags_required)
1874 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
1875 ctx->restrictions.sqe_flags_required))
1881 static void io_init_req_drain(struct io_kiocb *req)
1883 struct io_ring_ctx *ctx = req->ctx;
1884 struct io_kiocb *head = ctx->submit_state.link.head;
1886 ctx->drain_active = true;
1889 * If we need to drain a request in the middle of a link, drain
1890 * the head request and the next request/link after the current
1891 * link. Considering sequential execution of links,
1892 * REQ_F_IO_DRAIN will be maintained for every request of our
1895 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
1896 ctx->drain_next = true;
1900 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
1901 const struct io_uring_sqe *sqe)
1902 __must_hold(&ctx->uring_lock)
1904 const struct io_op_def *def;
1905 unsigned int sqe_flags;
1909 /* req is partially pre-initialised, see io_preinit_req() */
1910 req->opcode = opcode = READ_ONCE(sqe->opcode);
1911 /* same numerical values with corresponding REQ_F_*, safe to copy */
1912 req->flags = sqe_flags = READ_ONCE(sqe->flags);
1913 req->cqe.user_data = READ_ONCE(sqe->user_data);
1915 req->rsrc_node = NULL;
1916 req->task = current;
1918 if (unlikely(opcode >= IORING_OP_LAST)) {
1922 def = &io_op_defs[opcode];
1923 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
1924 /* enforce forwards compatibility on users */
1925 if (sqe_flags & ~SQE_VALID_FLAGS)
1927 if (sqe_flags & IOSQE_BUFFER_SELECT) {
1928 if (!def->buffer_select)
1930 req->buf_index = READ_ONCE(sqe->buf_group);
1932 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
1933 ctx->drain_disabled = true;
1934 if (sqe_flags & IOSQE_IO_DRAIN) {
1935 if (ctx->drain_disabled)
1937 io_init_req_drain(req);
1940 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
1941 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
1943 /* knock it to the slow queue path, will be drained there */
1944 if (ctx->drain_active)
1945 req->flags |= REQ_F_FORCE_ASYNC;
1946 /* if there is no link, we're at "next" request and need to drain */
1947 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
1948 ctx->drain_next = false;
1949 ctx->drain_active = true;
1950 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
1954 if (!def->ioprio && sqe->ioprio)
1956 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
1959 if (def->needs_file) {
1960 struct io_submit_state *state = &ctx->submit_state;
1962 req->cqe.fd = READ_ONCE(sqe->fd);
1965 * Plug now if we have more than 2 IO left after this, and the
1966 * target is potentially a read/write to block based storage.
1968 if (state->need_plug && def->plug) {
1969 state->plug_started = true;
1970 state->need_plug = false;
1971 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
1975 personality = READ_ONCE(sqe->personality);
1979 req->creds = xa_load(&ctx->personalities, personality);
1982 get_cred(req->creds);
1983 ret = security_uring_override_creds(req->creds);
1985 put_cred(req->creds);
1988 req->flags |= REQ_F_CREDS;
1991 return def->prep(req, sqe);
1994 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
1995 struct io_kiocb *req, int ret)
1997 struct io_ring_ctx *ctx = req->ctx;
1998 struct io_submit_link *link = &ctx->submit_state.link;
1999 struct io_kiocb *head = link->head;
2001 trace_io_uring_req_failed(sqe, req, ret);
2004 * Avoid breaking links in the middle as it renders links with SQPOLL
2005 * unusable. Instead of failing eagerly, continue assembling the link if
2006 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2007 * should find the flag and handle the rest.
2009 req_fail_link_node(req, ret);
2010 if (head && !(head->flags & REQ_F_FAIL))
2011 req_fail_link_node(head, -ECANCELED);
2013 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2015 link->last->link = req;
2019 io_queue_sqe_fallback(req);
2024 link->last->link = req;
2031 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2032 const struct io_uring_sqe *sqe)
2033 __must_hold(&ctx->uring_lock)
2035 struct io_submit_link *link = &ctx->submit_state.link;
2038 ret = io_init_req(ctx, req, sqe);
2040 return io_submit_fail_init(sqe, req, ret);
2042 /* don't need @sqe from now on */
2043 trace_io_uring_submit_sqe(req, true);
2046 * If we already have a head request, queue this one for async
2047 * submittal once the head completes. If we don't have a head but
2048 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2049 * submitted sync once the chain is complete. If none of those
2050 * conditions are true (normal request), then just queue it.
2052 if (unlikely(link->head)) {
2053 ret = io_req_prep_async(req);
2055 return io_submit_fail_init(sqe, req, ret);
2057 trace_io_uring_link(req, link->head);
2058 link->last->link = req;
2061 if (req->flags & IO_REQ_LINK_FLAGS)
2063 /* last request of the link, flush it */
2066 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2069 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2070 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2071 if (req->flags & IO_REQ_LINK_FLAGS) {
2076 io_queue_sqe_fallback(req);
2086 * Batched submission is done, ensure local IO is flushed out.
2088 static void io_submit_state_end(struct io_ring_ctx *ctx)
2090 struct io_submit_state *state = &ctx->submit_state;
2092 if (unlikely(state->link.head))
2093 io_queue_sqe_fallback(state->link.head);
2094 /* flush only after queuing links as they can generate completions */
2095 io_submit_flush_completions(ctx);
2096 if (state->plug_started)
2097 blk_finish_plug(&state->plug);
2101 * Start submission side cache.
2103 static void io_submit_state_start(struct io_submit_state *state,
2104 unsigned int max_ios)
2106 state->plug_started = false;
2107 state->need_plug = max_ios > 2;
2108 state->submit_nr = max_ios;
2109 /* set only head, no need to init link_last in advance */
2110 state->link.head = NULL;
2113 static void io_commit_sqring(struct io_ring_ctx *ctx)
2115 struct io_rings *rings = ctx->rings;
2118 * Ensure any loads from the SQEs are done at this point,
2119 * since once we write the new head, the application could
2120 * write new data to them.
2122 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2126 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2127 * that is mapped by userspace. This means that care needs to be taken to
2128 * ensure that reads are stable, as we cannot rely on userspace always
2129 * being a good citizen. If members of the sqe are validated and then later
2130 * used, it's important that those reads are done through READ_ONCE() to
2131 * prevent a re-load down the line.
2133 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
2135 unsigned head, mask = ctx->sq_entries - 1;
2136 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2139 * The cached sq head (or cq tail) serves two purposes:
2141 * 1) allows us to batch the cost of updating the user visible
2143 * 2) allows the kernel side to track the head on its own, even
2144 * though the application is the one updating it.
2146 head = READ_ONCE(ctx->sq_array[sq_idx]);
2147 if (likely(head < ctx->sq_entries)) {
2148 /* double index for 128-byte SQEs, twice as long */
2149 if (ctx->flags & IORING_SETUP_SQE128)
2151 return &ctx->sq_sqes[head];
2154 /* drop invalid entries */
2156 WRITE_ONCE(ctx->rings->sq_dropped,
2157 READ_ONCE(ctx->rings->sq_dropped) + 1);
2161 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2162 __must_hold(&ctx->uring_lock)
2164 unsigned int entries = io_sqring_entries(ctx);
2168 if (unlikely(!entries))
2170 /* make sure SQ entry isn't read before tail */
2171 ret = left = min3(nr, ctx->sq_entries, entries);
2172 io_get_task_refs(left);
2173 io_submit_state_start(&ctx->submit_state, left);
2176 const struct io_uring_sqe *sqe;
2177 struct io_kiocb *req;
2179 if (unlikely(!io_alloc_req_refill(ctx)))
2181 req = io_alloc_req(ctx);
2182 sqe = io_get_sqe(ctx);
2183 if (unlikely(!sqe)) {
2184 io_req_add_to_cache(req, ctx);
2189 * Continue submitting even for sqe failure if the
2190 * ring was setup with IORING_SETUP_SUBMIT_ALL
2192 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2193 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2199 if (unlikely(left)) {
2201 /* try again if it submitted nothing and can't allocate a req */
2202 if (!ret && io_req_cache_empty(ctx))
2204 current->io_uring->cached_refs += left;
2207 io_submit_state_end(ctx);
2208 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2209 io_commit_sqring(ctx);
2213 struct io_wait_queue {
2214 struct wait_queue_entry wq;
2215 struct io_ring_ctx *ctx;
2217 unsigned nr_timeouts;
2220 static inline bool io_should_wake(struct io_wait_queue *iowq)
2222 struct io_ring_ctx *ctx = iowq->ctx;
2223 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
2226 * Wake up if we have enough events, or if a timeout occurred since we
2227 * started waiting. For timeouts, we always want to return to userspace,
2228 * regardless of event count.
2230 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2233 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2234 int wake_flags, void *key)
2236 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
2240 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2241 * the task, and the next invocation will do it.
2243 if (io_should_wake(iowq) ||
2244 test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &iowq->ctx->check_cq))
2245 return autoremove_wake_function(curr, mode, wake_flags, key);
2249 int io_run_task_work_sig(void)
2251 if (io_run_task_work())
2253 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
2254 return -ERESTARTSYS;
2255 if (task_sigpending(current))
2260 /* when returns >0, the caller should retry */
2261 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2262 struct io_wait_queue *iowq,
2266 unsigned long check_cq;
2268 /* make sure we run task_work before checking for signals */
2269 ret = io_run_task_work_sig();
2270 if (ret || io_should_wake(iowq))
2273 check_cq = READ_ONCE(ctx->check_cq);
2274 if (unlikely(check_cq)) {
2275 /* let the caller flush overflows, retry */
2276 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2278 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
2281 if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS))
2287 * Wait until events become available, if we don't already have some. The
2288 * application must reap them itself, as they reside on the shared cq ring.
2290 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2291 const sigset_t __user *sig, size_t sigsz,
2292 struct __kernel_timespec __user *uts)
2294 struct io_wait_queue iowq;
2295 struct io_rings *rings = ctx->rings;
2296 ktime_t timeout = KTIME_MAX;
2300 io_cqring_overflow_flush(ctx);
2301 if (io_cqring_events(ctx) >= min_events)
2303 if (!io_run_task_work())
2308 #ifdef CONFIG_COMPAT
2309 if (in_compat_syscall())
2310 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2314 ret = set_user_sigmask(sig, sigsz);
2321 struct timespec64 ts;
2323 if (get_timespec64(&ts, uts))
2325 timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2328 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2329 iowq.wq.private = current;
2330 INIT_LIST_HEAD(&iowq.wq.entry);
2332 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2333 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2335 trace_io_uring_cqring_wait(ctx, min_events);
2337 /* if we can't even flush overflow, don't wait for more */
2338 if (!io_cqring_overflow_flush(ctx)) {
2342 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2343 TASK_INTERRUPTIBLE);
2344 ret = io_cqring_wait_schedule(ctx, &iowq, timeout);
2348 finish_wait(&ctx->cq_wait, &iowq.wq);
2349 restore_saved_sigmask_unless(ret == -EINTR);
2351 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2354 static void io_mem_free(void *ptr)
2361 page = virt_to_head_page(ptr);
2362 if (put_page_testzero(page))
2363 free_compound_page(page);
2366 static void *io_mem_alloc(size_t size)
2368 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2370 return (void *) __get_free_pages(gfp, get_order(size));
2373 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2374 unsigned int cq_entries, size_t *sq_offset)
2376 struct io_rings *rings;
2377 size_t off, sq_array_size;
2379 off = struct_size(rings, cqes, cq_entries);
2380 if (off == SIZE_MAX)
2382 if (ctx->flags & IORING_SETUP_CQE32) {
2383 if (check_shl_overflow(off, 1, &off))
2388 off = ALIGN(off, SMP_CACHE_BYTES);
2396 sq_array_size = array_size(sizeof(u32), sq_entries);
2397 if (sq_array_size == SIZE_MAX)
2400 if (check_add_overflow(off, sq_array_size, &off))
2406 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2407 unsigned int eventfd_async)
2409 struct io_ev_fd *ev_fd;
2410 __s32 __user *fds = arg;
2413 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2414 lockdep_is_held(&ctx->uring_lock));
2418 if (copy_from_user(&fd, fds, sizeof(*fds)))
2421 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2425 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2426 if (IS_ERR(ev_fd->cq_ev_fd)) {
2427 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2431 ev_fd->eventfd_async = eventfd_async;
2432 ctx->has_evfd = true;
2433 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2437 static void io_eventfd_put(struct rcu_head *rcu)
2439 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
2441 eventfd_ctx_put(ev_fd->cq_ev_fd);
2445 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2447 struct io_ev_fd *ev_fd;
2449 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2450 lockdep_is_held(&ctx->uring_lock));
2452 ctx->has_evfd = false;
2453 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2454 call_rcu(&ev_fd->rcu, io_eventfd_put);
2461 static void io_req_caches_free(struct io_ring_ctx *ctx)
2463 struct io_submit_state *state = &ctx->submit_state;
2466 mutex_lock(&ctx->uring_lock);
2467 io_flush_cached_locked_reqs(ctx, state);
2469 while (!io_req_cache_empty(ctx)) {
2470 struct io_wq_work_node *node;
2471 struct io_kiocb *req;
2473 node = wq_stack_extract(&state->free_list);
2474 req = container_of(node, struct io_kiocb, comp_list);
2475 kmem_cache_free(req_cachep, req);
2479 percpu_ref_put_many(&ctx->refs, nr);
2480 mutex_unlock(&ctx->uring_lock);
2483 static void io_flush_apoll_cache(struct io_ring_ctx *ctx)
2485 struct async_poll *apoll;
2487 while (!list_empty(&ctx->apoll_cache)) {
2488 apoll = list_first_entry(&ctx->apoll_cache, struct async_poll,
2490 list_del(&apoll->poll.wait.entry);
2495 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2497 io_sq_thread_finish(ctx);
2499 if (ctx->mm_account) {
2500 mmdrop(ctx->mm_account);
2501 ctx->mm_account = NULL;
2504 io_rsrc_refs_drop(ctx);
2505 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2506 io_wait_rsrc_data(ctx->buf_data);
2507 io_wait_rsrc_data(ctx->file_data);
2509 mutex_lock(&ctx->uring_lock);
2511 __io_sqe_buffers_unregister(ctx);
2513 __io_sqe_files_unregister(ctx);
2515 __io_cqring_overflow_flush(ctx, true);
2516 io_eventfd_unregister(ctx);
2517 io_flush_apoll_cache(ctx);
2518 mutex_unlock(&ctx->uring_lock);
2519 io_destroy_buffers(ctx);
2521 put_cred(ctx->sq_creds);
2522 if (ctx->submitter_task)
2523 put_task_struct(ctx->submitter_task);
2525 /* there are no registered resources left, nobody uses it */
2527 io_rsrc_node_destroy(ctx->rsrc_node);
2528 if (ctx->rsrc_backup_node)
2529 io_rsrc_node_destroy(ctx->rsrc_backup_node);
2530 flush_delayed_work(&ctx->rsrc_put_work);
2531 flush_delayed_work(&ctx->fallback_work);
2533 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2534 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
2536 #if defined(CONFIG_UNIX)
2537 if (ctx->ring_sock) {
2538 ctx->ring_sock->file = NULL; /* so that iput() is called */
2539 sock_release(ctx->ring_sock);
2542 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2544 io_mem_free(ctx->rings);
2545 io_mem_free(ctx->sq_sqes);
2547 percpu_ref_exit(&ctx->refs);
2548 free_uid(ctx->user);
2549 io_req_caches_free(ctx);
2551 io_wq_put_hash(ctx->hash_map);
2552 kfree(ctx->cancel_table.hbs);
2553 kfree(ctx->cancel_table_locked.hbs);
2554 kfree(ctx->dummy_ubuf);
2556 xa_destroy(&ctx->io_bl_xa);
2560 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2562 struct io_ring_ctx *ctx = file->private_data;
2565 poll_wait(file, &ctx->cq_wait, wait);
2567 * synchronizes with barrier from wq_has_sleeper call in
2571 if (!io_sqring_full(ctx))
2572 mask |= EPOLLOUT | EPOLLWRNORM;
2575 * Don't flush cqring overflow list here, just do a simple check.
2576 * Otherwise there could possible be ABBA deadlock:
2579 * lock(&ctx->uring_lock);
2581 * lock(&ctx->uring_lock);
2584 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2585 * pushs them to do the flush.
2587 if (io_cqring_events(ctx) ||
2588 test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
2589 mask |= EPOLLIN | EPOLLRDNORM;
2594 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2596 const struct cred *creds;
2598 creds = xa_erase(&ctx->personalities, id);
2607 struct io_tctx_exit {
2608 struct callback_head task_work;
2609 struct completion completion;
2610 struct io_ring_ctx *ctx;
2613 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2615 struct io_uring_task *tctx = current->io_uring;
2616 struct io_tctx_exit *work;
2618 work = container_of(cb, struct io_tctx_exit, task_work);
2620 * When @in_idle, we're in cancellation and it's racy to remove the
2621 * node. It'll be removed by the end of cancellation, just ignore it.
2623 if (!atomic_read(&tctx->in_idle))
2624 io_uring_del_tctx_node((unsigned long)work->ctx);
2625 complete(&work->completion);
2628 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2630 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2632 return req->ctx == data;
2635 static __cold void io_ring_exit_work(struct work_struct *work)
2637 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2638 unsigned long timeout = jiffies + HZ * 60 * 5;
2639 unsigned long interval = HZ / 20;
2640 struct io_tctx_exit exit;
2641 struct io_tctx_node *node;
2645 * If we're doing polled IO and end up having requests being
2646 * submitted async (out-of-line), then completions can come in while
2647 * we're waiting for refs to drop. We need to reap these manually,
2648 * as nobody else will be looking for them.
2651 io_uring_try_cancel_requests(ctx, NULL, true);
2653 struct io_sq_data *sqd = ctx->sq_data;
2654 struct task_struct *tsk;
2656 io_sq_thread_park(sqd);
2658 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
2659 io_wq_cancel_cb(tsk->io_uring->io_wq,
2660 io_cancel_ctx_cb, ctx, true);
2661 io_sq_thread_unpark(sqd);
2664 io_req_caches_free(ctx);
2666 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
2667 /* there is little hope left, don't run it too often */
2670 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
2672 init_completion(&exit.completion);
2673 init_task_work(&exit.task_work, io_tctx_exit_cb);
2676 * Some may use context even when all refs and requests have been put,
2677 * and they are free to do so while still holding uring_lock or
2678 * completion_lock, see io_req_task_submit(). Apart from other work,
2679 * this lock/unlock section also waits them to finish.
2681 mutex_lock(&ctx->uring_lock);
2682 while (!list_empty(&ctx->tctx_list)) {
2683 WARN_ON_ONCE(time_after(jiffies, timeout));
2685 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
2687 /* don't spin on a single task if cancellation failed */
2688 list_rotate_left(&ctx->tctx_list);
2689 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
2690 if (WARN_ON_ONCE(ret))
2693 mutex_unlock(&ctx->uring_lock);
2694 wait_for_completion(&exit.completion);
2695 mutex_lock(&ctx->uring_lock);
2697 mutex_unlock(&ctx->uring_lock);
2698 spin_lock(&ctx->completion_lock);
2699 spin_unlock(&ctx->completion_lock);
2701 io_ring_ctx_free(ctx);
2704 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
2706 unsigned long index;
2707 struct creds *creds;
2709 mutex_lock(&ctx->uring_lock);
2710 percpu_ref_kill(&ctx->refs);
2712 __io_cqring_overflow_flush(ctx, true);
2713 xa_for_each(&ctx->personalities, index, creds)
2714 io_unregister_personality(ctx, index);
2716 io_poll_remove_all(ctx, NULL, true);
2717 mutex_unlock(&ctx->uring_lock);
2719 /* failed during ring init, it couldn't have issued any requests */
2721 io_kill_timeouts(ctx, NULL, true);
2722 /* if we failed setting up the ctx, we might not have any rings */
2723 io_iopoll_try_reap_events(ctx);
2726 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
2728 * Use system_unbound_wq to avoid spawning tons of event kworkers
2729 * if we're exiting a ton of rings at the same time. It just adds
2730 * noise and overhead, there's no discernable change in runtime
2731 * over using system_wq.
2733 queue_work(system_unbound_wq, &ctx->exit_work);
2736 static int io_uring_release(struct inode *inode, struct file *file)
2738 struct io_ring_ctx *ctx = file->private_data;
2740 file->private_data = NULL;
2741 io_ring_ctx_wait_and_kill(ctx);
2745 struct io_task_cancel {
2746 struct task_struct *task;
2750 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
2752 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2753 struct io_task_cancel *cancel = data;
2755 return io_match_task_safe(req, cancel->task, cancel->all);
2758 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
2759 struct task_struct *task,
2762 struct io_defer_entry *de;
2765 spin_lock(&ctx->completion_lock);
2766 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
2767 if (io_match_task_safe(de->req, task, cancel_all)) {
2768 list_cut_position(&list, &ctx->defer_list, &de->list);
2772 spin_unlock(&ctx->completion_lock);
2773 if (list_empty(&list))
2776 while (!list_empty(&list)) {
2777 de = list_first_entry(&list, struct io_defer_entry, list);
2778 list_del_init(&de->list);
2779 io_req_complete_failed(de->req, -ECANCELED);
2785 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
2787 struct io_tctx_node *node;
2788 enum io_wq_cancel cret;
2791 mutex_lock(&ctx->uring_lock);
2792 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
2793 struct io_uring_task *tctx = node->task->io_uring;
2796 * io_wq will stay alive while we hold uring_lock, because it's
2797 * killed after ctx nodes, which requires to take the lock.
2799 if (!tctx || !tctx->io_wq)
2801 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
2802 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2804 mutex_unlock(&ctx->uring_lock);
2809 static __cold void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
2810 struct task_struct *task,
2813 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
2814 struct io_uring_task *tctx = task ? task->io_uring : NULL;
2816 /* failed during ring init, it couldn't have issued any requests */
2821 enum io_wq_cancel cret;
2825 ret |= io_uring_try_cancel_iowq(ctx);
2826 } else if (tctx && tctx->io_wq) {
2828 * Cancels requests of all rings, not only @ctx, but
2829 * it's fine as the task is in exit/exec.
2831 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
2833 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2836 /* SQPOLL thread does its own polling */
2837 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
2838 (ctx->sq_data && ctx->sq_data->thread == current)) {
2839 while (!wq_list_empty(&ctx->iopoll_list)) {
2840 io_iopoll_try_reap_events(ctx);
2845 ret |= io_cancel_defer_files(ctx, task, cancel_all);
2846 mutex_lock(&ctx->uring_lock);
2847 ret |= io_poll_remove_all(ctx, task, cancel_all);
2848 mutex_unlock(&ctx->uring_lock);
2849 ret |= io_kill_timeouts(ctx, task, cancel_all);
2851 ret |= io_run_task_work();
2858 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
2861 return atomic_read(&tctx->inflight_tracked);
2862 return percpu_counter_sum(&tctx->inflight);
2866 * Find any io_uring ctx that this task has registered or done IO on, and cancel
2867 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
2869 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
2871 struct io_uring_task *tctx = current->io_uring;
2872 struct io_ring_ctx *ctx;
2876 WARN_ON_ONCE(sqd && sqd->thread != current);
2878 if (!current->io_uring)
2881 io_wq_exit_start(tctx->io_wq);
2883 atomic_inc(&tctx->in_idle);
2885 io_uring_drop_tctx_refs(current);
2886 /* read completions before cancelations */
2887 inflight = tctx_inflight(tctx, !cancel_all);
2892 struct io_tctx_node *node;
2893 unsigned long index;
2895 xa_for_each(&tctx->xa, index, node) {
2896 /* sqpoll task will cancel all its requests */
2897 if (node->ctx->sq_data)
2899 io_uring_try_cancel_requests(node->ctx, current,
2903 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
2904 io_uring_try_cancel_requests(ctx, current,
2908 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
2910 io_uring_drop_tctx_refs(current);
2913 * If we've seen completions, retry without waiting. This
2914 * avoids a race where a completion comes in before we did
2915 * prepare_to_wait().
2917 if (inflight == tctx_inflight(tctx, !cancel_all))
2919 finish_wait(&tctx->wait, &wait);
2922 io_uring_clean_tctx(tctx);
2925 * We shouldn't run task_works after cancel, so just leave
2926 * ->in_idle set for normal exit.
2928 atomic_dec(&tctx->in_idle);
2929 /* for exec all current's requests should be gone, kill tctx */
2930 __io_uring_free(current);
2934 void __io_uring_cancel(bool cancel_all)
2936 io_uring_cancel_generic(cancel_all, NULL);
2939 static void *io_uring_validate_mmap_request(struct file *file,
2940 loff_t pgoff, size_t sz)
2942 struct io_ring_ctx *ctx = file->private_data;
2943 loff_t offset = pgoff << PAGE_SHIFT;
2948 case IORING_OFF_SQ_RING:
2949 case IORING_OFF_CQ_RING:
2952 case IORING_OFF_SQES:
2956 return ERR_PTR(-EINVAL);
2959 page = virt_to_head_page(ptr);
2960 if (sz > page_size(page))
2961 return ERR_PTR(-EINVAL);
2968 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
2970 size_t sz = vma->vm_end - vma->vm_start;
2974 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
2976 return PTR_ERR(ptr);
2978 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
2979 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
2982 #else /* !CONFIG_MMU */
2984 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
2986 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
2989 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
2991 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
2994 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
2995 unsigned long addr, unsigned long len,
2996 unsigned long pgoff, unsigned long flags)
3000 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3002 return PTR_ERR(ptr);
3004 return (unsigned long) ptr;
3007 #endif /* !CONFIG_MMU */
3009 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3011 if (flags & IORING_ENTER_EXT_ARG) {
3012 struct io_uring_getevents_arg arg;
3014 if (argsz != sizeof(arg))
3016 if (copy_from_user(&arg, argp, sizeof(arg)))
3022 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3023 struct __kernel_timespec __user **ts,
3024 const sigset_t __user **sig)
3026 struct io_uring_getevents_arg arg;
3029 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3030 * is just a pointer to the sigset_t.
3032 if (!(flags & IORING_ENTER_EXT_ARG)) {
3033 *sig = (const sigset_t __user *) argp;
3039 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3040 * timespec and sigset_t pointers if good.
3042 if (*argsz != sizeof(arg))
3044 if (copy_from_user(&arg, argp, sizeof(arg)))
3048 *sig = u64_to_user_ptr(arg.sigmask);
3049 *argsz = arg.sigmask_sz;
3050 *ts = u64_to_user_ptr(arg.ts);
3054 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3055 u32, min_complete, u32, flags, const void __user *, argp,
3058 struct io_ring_ctx *ctx;
3064 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3065 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3066 IORING_ENTER_REGISTERED_RING)))
3070 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3071 * need only dereference our task private array to find it.
3073 if (flags & IORING_ENTER_REGISTERED_RING) {
3074 struct io_uring_task *tctx = current->io_uring;
3076 if (!tctx || fd >= IO_RINGFD_REG_MAX)
3078 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3079 f.file = tctx->registered_rings[fd];
3085 if (unlikely(!f.file))
3089 if (unlikely(!io_is_uring_fops(f.file)))
3093 ctx = f.file->private_data;
3094 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
3098 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3102 * For SQ polling, the thread will do all submissions and completions.
3103 * Just return the requested submit count, and wake the thread if
3107 if (ctx->flags & IORING_SETUP_SQPOLL) {
3108 io_cqring_overflow_flush(ctx);
3110 if (unlikely(ctx->sq_data->thread == NULL)) {
3114 if (flags & IORING_ENTER_SQ_WAKEUP)
3115 wake_up(&ctx->sq_data->wait);
3116 if (flags & IORING_ENTER_SQ_WAIT) {
3117 ret = io_sqpoll_wait_sq(ctx);
3122 } else if (to_submit) {
3123 ret = io_uring_add_tctx_node(ctx);
3127 mutex_lock(&ctx->uring_lock);
3128 ret = io_submit_sqes(ctx, to_submit);
3129 if (ret != to_submit) {
3130 mutex_unlock(&ctx->uring_lock);
3133 if ((flags & IORING_ENTER_GETEVENTS) && ctx->syscall_iopoll)
3135 mutex_unlock(&ctx->uring_lock);
3137 if (flags & IORING_ENTER_GETEVENTS) {
3139 if (ctx->syscall_iopoll) {
3141 * We disallow the app entering submit/complete with
3142 * polling, but we still need to lock the ring to
3143 * prevent racing with polled issue that got punted to
3146 mutex_lock(&ctx->uring_lock);
3148 ret2 = io_validate_ext_arg(flags, argp, argsz);
3149 if (likely(!ret2)) {
3150 min_complete = min(min_complete,
3152 ret2 = io_iopoll_check(ctx, min_complete);
3154 mutex_unlock(&ctx->uring_lock);
3156 const sigset_t __user *sig;
3157 struct __kernel_timespec __user *ts;
3159 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3160 if (likely(!ret2)) {
3161 min_complete = min(min_complete,
3163 ret2 = io_cqring_wait(ctx, min_complete, sig,
3172 * EBADR indicates that one or more CQE were dropped.
3173 * Once the user has been informed we can clear the bit
3174 * as they are obviously ok with those drops.
3176 if (unlikely(ret2 == -EBADR))
3177 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3183 percpu_ref_put(&ctx->refs);
3189 static const struct file_operations io_uring_fops = {
3190 .release = io_uring_release,
3191 .mmap = io_uring_mmap,
3193 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3194 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3196 .poll = io_uring_poll,
3197 #ifdef CONFIG_PROC_FS
3198 .show_fdinfo = io_uring_show_fdinfo,
3202 bool io_is_uring_fops(struct file *file)
3204 return file->f_op == &io_uring_fops;
3207 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3208 struct io_uring_params *p)
3210 struct io_rings *rings;
3211 size_t size, sq_array_offset;
3213 /* make sure these are sane, as we already accounted them */
3214 ctx->sq_entries = p->sq_entries;
3215 ctx->cq_entries = p->cq_entries;
3217 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3218 if (size == SIZE_MAX)
3221 rings = io_mem_alloc(size);
3226 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3227 rings->sq_ring_mask = p->sq_entries - 1;
3228 rings->cq_ring_mask = p->cq_entries - 1;
3229 rings->sq_ring_entries = p->sq_entries;
3230 rings->cq_ring_entries = p->cq_entries;
3232 if (p->flags & IORING_SETUP_SQE128)
3233 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3235 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3236 if (size == SIZE_MAX) {
3237 io_mem_free(ctx->rings);
3242 ctx->sq_sqes = io_mem_alloc(size);
3243 if (!ctx->sq_sqes) {
3244 io_mem_free(ctx->rings);
3252 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3256 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3260 ret = __io_uring_add_tctx_node(ctx, false);
3265 fd_install(fd, file);
3270 * Allocate an anonymous fd, this is what constitutes the application
3271 * visible backing of an io_uring instance. The application mmaps this
3272 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3273 * we have to tie this fd to a socket for file garbage collection purposes.
3275 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3278 #if defined(CONFIG_UNIX)
3281 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3284 return ERR_PTR(ret);
3287 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3288 O_RDWR | O_CLOEXEC, NULL);
3289 #if defined(CONFIG_UNIX)
3291 sock_release(ctx->ring_sock);
3292 ctx->ring_sock = NULL;
3294 ctx->ring_sock->file = file;
3300 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3301 struct io_uring_params __user *params)
3303 struct io_ring_ctx *ctx;
3309 if (entries > IORING_MAX_ENTRIES) {
3310 if (!(p->flags & IORING_SETUP_CLAMP))
3312 entries = IORING_MAX_ENTRIES;
3316 * Use twice as many entries for the CQ ring. It's possible for the
3317 * application to drive a higher depth than the size of the SQ ring,
3318 * since the sqes are only used at submission time. This allows for
3319 * some flexibility in overcommitting a bit. If the application has
3320 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3321 * of CQ ring entries manually.
3323 p->sq_entries = roundup_pow_of_two(entries);
3324 if (p->flags & IORING_SETUP_CQSIZE) {
3326 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3327 * to a power-of-two, if it isn't already. We do NOT impose
3328 * any cq vs sq ring sizing.
3332 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3333 if (!(p->flags & IORING_SETUP_CLAMP))
3335 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3337 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3338 if (p->cq_entries < p->sq_entries)
3341 p->cq_entries = 2 * p->sq_entries;
3344 ctx = io_ring_ctx_alloc(p);
3349 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3350 * space applications don't need to do io completion events
3351 * polling again, they can rely on io_sq_thread to do polling
3352 * work, which can reduce cpu usage and uring_lock contention.
3354 if (ctx->flags & IORING_SETUP_IOPOLL &&
3355 !(ctx->flags & IORING_SETUP_SQPOLL))
3356 ctx->syscall_iopoll = 1;
3358 ctx->compat = in_compat_syscall();
3359 if (!capable(CAP_IPC_LOCK))
3360 ctx->user = get_uid(current_user());
3363 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3364 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3367 if (ctx->flags & IORING_SETUP_SQPOLL) {
3368 /* IPI related flags don't make sense with SQPOLL */
3369 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3370 IORING_SETUP_TASKRUN_FLAG))
3372 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3373 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3374 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3376 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
3378 ctx->notify_method = TWA_SIGNAL;
3382 * This is just grabbed for accounting purposes. When a process exits,
3383 * the mm is exited and dropped before the files, hence we need to hang
3384 * on to this mm purely for the purposes of being able to unaccount
3385 * memory (locked/pinned vm). It's not used for anything else.
3387 mmgrab(current->mm);
3388 ctx->mm_account = current->mm;
3390 ret = io_allocate_scq_urings(ctx, p);
3394 ret = io_sq_offload_create(ctx, p);
3397 /* always set a rsrc node */
3398 ret = io_rsrc_node_switch_start(ctx);
3401 io_rsrc_node_switch(ctx, NULL);
3403 memset(&p->sq_off, 0, sizeof(p->sq_off));
3404 p->sq_off.head = offsetof(struct io_rings, sq.head);
3405 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3406 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3407 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3408 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3409 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3410 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3412 memset(&p->cq_off, 0, sizeof(p->cq_off));
3413 p->cq_off.head = offsetof(struct io_rings, cq.head);
3414 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3415 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3416 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3417 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3418 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3419 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3421 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3422 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3423 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3424 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3425 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3426 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3427 IORING_FEAT_LINKED_FILE;
3429 if (copy_to_user(params, p, sizeof(*p))) {
3434 file = io_uring_get_file(ctx);
3436 ret = PTR_ERR(file);
3441 * Install ring fd as the very last thing, so we don't risk someone
3442 * having closed it before we finish setup
3444 ret = io_uring_install_fd(ctx, file);
3446 /* fput will clean it up */
3451 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3454 io_ring_ctx_wait_and_kill(ctx);
3459 * Sets up an aio uring context, and returns the fd. Applications asks for a
3460 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3461 * params structure passed in.
3463 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3465 struct io_uring_params p;
3468 if (copy_from_user(&p, params, sizeof(p)))
3470 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3475 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3476 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3477 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3478 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3479 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3480 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3481 IORING_SETUP_SINGLE_ISSUER))
3484 return io_uring_create(entries, &p, params);
3487 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3488 struct io_uring_params __user *, params)
3490 return io_uring_setup(entries, params);
3493 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3496 struct io_uring_probe *p;
3500 size = struct_size(p, ops, nr_args);
3501 if (size == SIZE_MAX)
3503 p = kzalloc(size, GFP_KERNEL);
3508 if (copy_from_user(p, arg, size))
3511 if (memchr_inv(p, 0, size))
3514 p->last_op = IORING_OP_LAST - 1;
3515 if (nr_args > IORING_OP_LAST)
3516 nr_args = IORING_OP_LAST;
3518 for (i = 0; i < nr_args; i++) {
3520 if (!io_op_defs[i].not_supported)
3521 p->ops[i].flags = IO_URING_OP_SUPPORTED;
3526 if (copy_to_user(arg, p, size))
3533 static int io_register_personality(struct io_ring_ctx *ctx)
3535 const struct cred *creds;
3539 creds = get_current_cred();
3541 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
3542 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
3550 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
3551 void __user *arg, unsigned int nr_args)
3553 struct io_uring_restriction *res;
3557 /* Restrictions allowed only if rings started disabled */
3558 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3561 /* We allow only a single restrictions registration */
3562 if (ctx->restrictions.registered)
3565 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
3568 size = array_size(nr_args, sizeof(*res));
3569 if (size == SIZE_MAX)
3572 res = memdup_user(arg, size);
3574 return PTR_ERR(res);
3578 for (i = 0; i < nr_args; i++) {
3579 switch (res[i].opcode) {
3580 case IORING_RESTRICTION_REGISTER_OP:
3581 if (res[i].register_op >= IORING_REGISTER_LAST) {
3586 __set_bit(res[i].register_op,
3587 ctx->restrictions.register_op);
3589 case IORING_RESTRICTION_SQE_OP:
3590 if (res[i].sqe_op >= IORING_OP_LAST) {
3595 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
3597 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
3598 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
3600 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
3601 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
3610 /* Reset all restrictions if an error happened */
3612 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
3614 ctx->restrictions.registered = true;
3620 static int io_register_enable_rings(struct io_ring_ctx *ctx)
3622 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3625 if (ctx->restrictions.registered)
3626 ctx->restricted = 1;
3628 ctx->flags &= ~IORING_SETUP_R_DISABLED;
3629 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
3630 wake_up(&ctx->sq_data->wait);
3634 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
3635 void __user *arg, unsigned len)
3637 struct io_uring_task *tctx = current->io_uring;
3638 cpumask_var_t new_mask;
3641 if (!tctx || !tctx->io_wq)
3644 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3647 cpumask_clear(new_mask);
3648 if (len > cpumask_size())
3649 len = cpumask_size();
3651 if (in_compat_syscall()) {
3652 ret = compat_get_bitmap(cpumask_bits(new_mask),
3653 (const compat_ulong_t __user *)arg,
3654 len * 8 /* CHAR_BIT */);
3656 ret = copy_from_user(new_mask, arg, len);
3660 free_cpumask_var(new_mask);
3664 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
3665 free_cpumask_var(new_mask);
3669 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
3671 struct io_uring_task *tctx = current->io_uring;
3673 if (!tctx || !tctx->io_wq)
3676 return io_wq_cpu_affinity(tctx->io_wq, NULL);
3679 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
3681 __must_hold(&ctx->uring_lock)
3683 struct io_tctx_node *node;
3684 struct io_uring_task *tctx = NULL;
3685 struct io_sq_data *sqd = NULL;
3689 if (copy_from_user(new_count, arg, sizeof(new_count)))
3691 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3692 if (new_count[i] > INT_MAX)
3695 if (ctx->flags & IORING_SETUP_SQPOLL) {
3699 * Observe the correct sqd->lock -> ctx->uring_lock
3700 * ordering. Fine to drop uring_lock here, we hold
3703 refcount_inc(&sqd->refs);
3704 mutex_unlock(&ctx->uring_lock);
3705 mutex_lock(&sqd->lock);
3706 mutex_lock(&ctx->uring_lock);
3708 tctx = sqd->thread->io_uring;
3711 tctx = current->io_uring;
3714 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
3716 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3718 ctx->iowq_limits[i] = new_count[i];
3719 ctx->iowq_limits_set = true;
3721 if (tctx && tctx->io_wq) {
3722 ret = io_wq_max_workers(tctx->io_wq, new_count);
3726 memset(new_count, 0, sizeof(new_count));
3730 mutex_unlock(&sqd->lock);
3731 io_put_sq_data(sqd);
3734 if (copy_to_user(arg, new_count, sizeof(new_count)))
3737 /* that's it for SQPOLL, only the SQPOLL task creates requests */
3741 /* now propagate the restriction to all registered users */
3742 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3743 struct io_uring_task *tctx = node->task->io_uring;
3745 if (WARN_ON_ONCE(!tctx->io_wq))
3748 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3749 new_count[i] = ctx->iowq_limits[i];
3750 /* ignore errors, it always returns zero anyway */
3751 (void)io_wq_max_workers(tctx->io_wq, new_count);
3756 mutex_unlock(&sqd->lock);
3757 io_put_sq_data(sqd);
3762 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
3763 void __user *arg, unsigned nr_args)
3764 __releases(ctx->uring_lock)
3765 __acquires(ctx->uring_lock)
3770 * We're inside the ring mutex, if the ref is already dying, then
3771 * someone else killed the ctx or is already going through
3772 * io_uring_register().
3774 if (percpu_ref_is_dying(&ctx->refs))
3777 if (ctx->restricted) {
3778 if (opcode >= IORING_REGISTER_LAST)
3780 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
3781 if (!test_bit(opcode, ctx->restrictions.register_op))
3786 case IORING_REGISTER_BUFFERS:
3790 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
3792 case IORING_UNREGISTER_BUFFERS:
3796 ret = io_sqe_buffers_unregister(ctx);
3798 case IORING_REGISTER_FILES:
3802 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
3804 case IORING_UNREGISTER_FILES:
3808 ret = io_sqe_files_unregister(ctx);
3810 case IORING_REGISTER_FILES_UPDATE:
3811 ret = io_register_files_update(ctx, arg, nr_args);
3813 case IORING_REGISTER_EVENTFD:
3817 ret = io_eventfd_register(ctx, arg, 0);
3819 case IORING_REGISTER_EVENTFD_ASYNC:
3823 ret = io_eventfd_register(ctx, arg, 1);
3825 case IORING_UNREGISTER_EVENTFD:
3829 ret = io_eventfd_unregister(ctx);
3831 case IORING_REGISTER_PROBE:
3833 if (!arg || nr_args > 256)
3835 ret = io_probe(ctx, arg, nr_args);
3837 case IORING_REGISTER_PERSONALITY:
3841 ret = io_register_personality(ctx);
3843 case IORING_UNREGISTER_PERSONALITY:
3847 ret = io_unregister_personality(ctx, nr_args);
3849 case IORING_REGISTER_ENABLE_RINGS:
3853 ret = io_register_enable_rings(ctx);
3855 case IORING_REGISTER_RESTRICTIONS:
3856 ret = io_register_restrictions(ctx, arg, nr_args);
3858 case IORING_REGISTER_FILES2:
3859 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
3861 case IORING_REGISTER_FILES_UPDATE2:
3862 ret = io_register_rsrc_update(ctx, arg, nr_args,
3865 case IORING_REGISTER_BUFFERS2:
3866 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
3868 case IORING_REGISTER_BUFFERS_UPDATE:
3869 ret = io_register_rsrc_update(ctx, arg, nr_args,
3870 IORING_RSRC_BUFFER);
3872 case IORING_REGISTER_IOWQ_AFF:
3874 if (!arg || !nr_args)
3876 ret = io_register_iowq_aff(ctx, arg, nr_args);
3878 case IORING_UNREGISTER_IOWQ_AFF:
3882 ret = io_unregister_iowq_aff(ctx);
3884 case IORING_REGISTER_IOWQ_MAX_WORKERS:
3886 if (!arg || nr_args != 2)
3888 ret = io_register_iowq_max_workers(ctx, arg);
3890 case IORING_REGISTER_RING_FDS:
3891 ret = io_ringfd_register(ctx, arg, nr_args);
3893 case IORING_UNREGISTER_RING_FDS:
3894 ret = io_ringfd_unregister(ctx, arg, nr_args);
3896 case IORING_REGISTER_PBUF_RING:
3898 if (!arg || nr_args != 1)
3900 ret = io_register_pbuf_ring(ctx, arg);
3902 case IORING_UNREGISTER_PBUF_RING:
3904 if (!arg || nr_args != 1)
3906 ret = io_unregister_pbuf_ring(ctx, arg);
3916 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
3917 void __user *, arg, unsigned int, nr_args)
3919 struct io_ring_ctx *ctx;
3928 if (!io_is_uring_fops(f.file))
3931 ctx = f.file->private_data;
3935 mutex_lock(&ctx->uring_lock);
3936 ret = __io_uring_register(ctx, opcode, arg, nr_args);
3937 mutex_unlock(&ctx->uring_lock);
3938 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
3944 static int __init io_uring_init(void)
3946 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
3947 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
3948 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
3951 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
3952 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
3953 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
3954 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
3955 BUILD_BUG_SQE_ELEM(1, __u8, flags);
3956 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
3957 BUILD_BUG_SQE_ELEM(4, __s32, fd);
3958 BUILD_BUG_SQE_ELEM(8, __u64, off);
3959 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
3960 BUILD_BUG_SQE_ELEM(16, __u64, addr);
3961 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
3962 BUILD_BUG_SQE_ELEM(24, __u32, len);
3963 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
3964 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
3965 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
3966 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
3967 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
3968 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
3969 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
3970 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
3971 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
3972 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
3973 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
3974 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
3975 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
3976 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
3977 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
3978 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
3979 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
3980 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
3981 BUILD_BUG_SQE_ELEM(42, __u16, personality);
3982 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
3983 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
3984 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
3986 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
3987 sizeof(struct io_uring_rsrc_update));
3988 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
3989 sizeof(struct io_uring_rsrc_update2));
3991 /* ->buf_index is u16 */
3992 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
3993 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
3994 offsetof(struct io_uring_buf_ring, tail));
3996 /* should fit into one byte */
3997 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
3998 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
3999 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4001 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
4003 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4005 io_uring_optable_init();
4007 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4011 __initcall(io_uring_init);