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>
83 #include "io_uring_types.h"
97 #define IORING_MAX_ENTRIES 32768
98 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
100 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
101 IORING_REGISTER_LAST + IORING_OP_LAST)
103 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
104 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
106 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
107 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
109 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
110 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
113 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
116 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
118 #define IO_COMPL_BATCH 32
119 #define IO_REQ_ALLOC_BATCH 8
122 IO_CHECK_CQ_OVERFLOW_BIT,
123 IO_CHECK_CQ_DROPPED_BIT,
126 struct io_defer_entry {
127 struct list_head list;
128 struct io_kiocb *req;
132 /* requests with any of those set should undergo io_disarm_next() */
133 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
134 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
136 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
137 struct task_struct *task,
140 static void io_dismantle_req(struct io_kiocb *req);
141 static void io_clean_op(struct io_kiocb *req);
142 static void io_queue_sqe(struct io_kiocb *req);
144 static void __io_submit_flush_completions(struct io_ring_ctx *ctx);
146 static void io_eventfd_signal(struct io_ring_ctx *ctx);
148 static struct kmem_cache *req_cachep;
150 struct sock *io_uring_get_socket(struct file *file)
152 #if defined(CONFIG_UNIX)
153 if (io_is_uring_fops(file)) {
154 struct io_ring_ctx *ctx = file->private_data;
156 return ctx->ring_sock->sk;
161 EXPORT_SYMBOL(io_uring_get_socket);
163 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
165 if (!wq_list_empty(&ctx->submit_state.compl_reqs))
166 __io_submit_flush_completions(ctx);
169 static bool io_match_linked(struct io_kiocb *head)
171 struct io_kiocb *req;
173 io_for_each_link(req, head) {
174 if (req->flags & REQ_F_INFLIGHT)
181 * As io_match_task() but protected against racing with linked timeouts.
182 * User must not hold timeout_lock.
184 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
189 if (task && head->task != task)
194 if (head->flags & REQ_F_LINK_TIMEOUT) {
195 struct io_ring_ctx *ctx = head->ctx;
197 /* protect against races with linked timeouts */
198 spin_lock_irq(&ctx->timeout_lock);
199 matched = io_match_linked(head);
200 spin_unlock_irq(&ctx->timeout_lock);
202 matched = io_match_linked(head);
207 static inline void req_fail_link_node(struct io_kiocb *req, int res)
210 io_req_set_res(req, res, 0);
213 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
215 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
218 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
220 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
222 complete(&ctx->ref_comp);
225 static __cold void io_fallback_req_func(struct work_struct *work)
227 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
229 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
230 struct io_kiocb *req, *tmp;
233 percpu_ref_get(&ctx->refs);
234 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
235 req->io_task_work.func(req, &locked);
238 io_submit_flush_completions(ctx);
239 mutex_unlock(&ctx->uring_lock);
241 percpu_ref_put(&ctx->refs);
244 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
246 struct io_ring_ctx *ctx;
247 unsigned hash_buckets;
251 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
255 xa_init(&ctx->io_bl_xa);
258 * Use 5 bits less than the max cq entries, that should give us around
259 * 32 entries per hash list if totally full and uniformly spread, but
260 * don't keep too many buckets to not overconsume memory.
262 hash_bits = ilog2(p->cq_entries) - 5;
263 hash_bits = clamp(hash_bits, 1, 8);
264 hash_buckets = 1U << hash_bits;
265 hash_size = hash_buckets * sizeof(struct io_hash_bucket);
267 ctx->cancel_hash_bits = hash_bits;
268 ctx->cancel_hash = kmalloc(hash_size, GFP_KERNEL);
269 if (!ctx->cancel_hash)
272 init_hash_table(ctx->cancel_hash, hash_buckets);
274 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
275 if (!ctx->dummy_ubuf)
277 /* set invalid range, so io_import_fixed() fails meeting it */
278 ctx->dummy_ubuf->ubuf = -1UL;
280 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
281 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
284 ctx->flags = p->flags;
285 init_waitqueue_head(&ctx->sqo_sq_wait);
286 INIT_LIST_HEAD(&ctx->sqd_list);
287 INIT_LIST_HEAD(&ctx->cq_overflow_list);
288 INIT_LIST_HEAD(&ctx->io_buffers_cache);
289 INIT_LIST_HEAD(&ctx->apoll_cache);
290 init_completion(&ctx->ref_comp);
291 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
292 mutex_init(&ctx->uring_lock);
293 init_waitqueue_head(&ctx->cq_wait);
294 spin_lock_init(&ctx->completion_lock);
295 spin_lock_init(&ctx->timeout_lock);
296 INIT_WQ_LIST(&ctx->iopoll_list);
297 INIT_LIST_HEAD(&ctx->io_buffers_pages);
298 INIT_LIST_HEAD(&ctx->io_buffers_comp);
299 INIT_LIST_HEAD(&ctx->defer_list);
300 INIT_LIST_HEAD(&ctx->timeout_list);
301 INIT_LIST_HEAD(&ctx->ltimeout_list);
302 spin_lock_init(&ctx->rsrc_ref_lock);
303 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
304 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
305 init_llist_head(&ctx->rsrc_put_llist);
306 INIT_LIST_HEAD(&ctx->tctx_list);
307 ctx->submit_state.free_list.next = NULL;
308 INIT_WQ_LIST(&ctx->locked_free_list);
309 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
310 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
313 kfree(ctx->dummy_ubuf);
314 kfree(ctx->cancel_hash);
316 xa_destroy(&ctx->io_bl_xa);
321 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
323 struct io_rings *r = ctx->rings;
325 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
329 static bool req_need_defer(struct io_kiocb *req, u32 seq)
331 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
332 struct io_ring_ctx *ctx = req->ctx;
334 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
340 static inline void io_req_track_inflight(struct io_kiocb *req)
342 if (!(req->flags & REQ_F_INFLIGHT)) {
343 req->flags |= REQ_F_INFLIGHT;
344 atomic_inc(&req->task->io_uring->inflight_tracked);
348 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
350 if (WARN_ON_ONCE(!req->link))
353 req->flags &= ~REQ_F_ARM_LTIMEOUT;
354 req->flags |= REQ_F_LINK_TIMEOUT;
356 /* linked timeouts should have two refs once prep'ed */
357 io_req_set_refcount(req);
358 __io_req_set_refcount(req->link, 2);
362 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
364 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
366 return __io_prep_linked_timeout(req);
369 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
371 io_queue_linked_timeout(__io_prep_linked_timeout(req));
374 static inline void io_arm_ltimeout(struct io_kiocb *req)
376 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
377 __io_arm_ltimeout(req);
380 static void io_prep_async_work(struct io_kiocb *req)
382 const struct io_op_def *def = &io_op_defs[req->opcode];
383 struct io_ring_ctx *ctx = req->ctx;
385 if (!(req->flags & REQ_F_CREDS)) {
386 req->flags |= REQ_F_CREDS;
387 req->creds = get_current_cred();
390 req->work.list.next = NULL;
392 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
393 if (req->flags & REQ_F_FORCE_ASYNC)
394 req->work.flags |= IO_WQ_WORK_CONCURRENT;
396 if (req->flags & REQ_F_ISREG) {
397 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
398 io_wq_hash_work(&req->work, file_inode(req->file));
399 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
400 if (def->unbound_nonreg_file)
401 req->work.flags |= IO_WQ_WORK_UNBOUND;
405 static void io_prep_async_link(struct io_kiocb *req)
407 struct io_kiocb *cur;
409 if (req->flags & REQ_F_LINK_TIMEOUT) {
410 struct io_ring_ctx *ctx = req->ctx;
412 spin_lock_irq(&ctx->timeout_lock);
413 io_for_each_link(cur, req)
414 io_prep_async_work(cur);
415 spin_unlock_irq(&ctx->timeout_lock);
417 io_for_each_link(cur, req)
418 io_prep_async_work(cur);
422 void io_queue_iowq(struct io_kiocb *req, bool *dont_use)
424 struct io_kiocb *link = io_prep_linked_timeout(req);
425 struct io_uring_task *tctx = req->task->io_uring;
428 BUG_ON(!tctx->io_wq);
430 /* init ->work of the whole link before punting */
431 io_prep_async_link(req);
434 * Not expected to happen, but if we do have a bug where this _can_
435 * happen, catch it here and ensure the request is marked as
436 * canceled. That will make io-wq go through the usual work cancel
437 * procedure rather than attempt to run this request (or create a new
440 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
441 req->work.flags |= IO_WQ_WORK_CANCEL;
443 trace_io_uring_queue_async_work(req->ctx, req, req->cqe.user_data,
444 req->opcode, req->flags, &req->work,
445 io_wq_is_hashed(&req->work));
446 io_wq_enqueue(tctx->io_wq, &req->work);
448 io_queue_linked_timeout(link);
451 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
453 while (!list_empty(&ctx->defer_list)) {
454 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
455 struct io_defer_entry, list);
457 if (req_need_defer(de->req, de->seq))
459 list_del_init(&de->list);
460 io_req_task_queue(de->req);
465 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
467 if (ctx->off_timeout_used || ctx->drain_active) {
468 spin_lock(&ctx->completion_lock);
469 if (ctx->off_timeout_used)
470 io_flush_timeouts(ctx);
471 if (ctx->drain_active)
472 io_queue_deferred(ctx);
473 io_commit_cqring(ctx);
474 spin_unlock(&ctx->completion_lock);
477 io_eventfd_signal(ctx);
480 static void io_eventfd_signal(struct io_ring_ctx *ctx)
482 struct io_ev_fd *ev_fd;
486 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
489 ev_fd = rcu_dereference(ctx->io_ev_fd);
492 * Check again if ev_fd exists incase an io_eventfd_unregister call
493 * completed between the NULL check of ctx->io_ev_fd at the start of
494 * the function and rcu_read_lock.
496 if (unlikely(!ev_fd))
498 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
501 if (!ev_fd->eventfd_async || io_wq_current_is_worker())
502 eventfd_signal(ev_fd->cq_ev_fd, 1);
508 * This should only get called when at least one event has been posted.
509 * Some applications rely on the eventfd notification count only changing
510 * IFF a new CQE has been added to the CQ ring. There's no depedency on
511 * 1:1 relationship between how many times this function is called (and
512 * hence the eventfd count) and number of CQEs posted to the CQ ring.
514 void io_cqring_ev_posted(struct io_ring_ctx *ctx)
516 if (unlikely(ctx->off_timeout_used || ctx->drain_active ||
518 __io_commit_cqring_flush(ctx);
523 /* Returns true if there are no backlogged entries after the flush */
524 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
526 bool all_flushed, posted;
527 size_t cqe_size = sizeof(struct io_uring_cqe);
529 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
532 if (ctx->flags & IORING_SETUP_CQE32)
536 spin_lock(&ctx->completion_lock);
537 while (!list_empty(&ctx->cq_overflow_list)) {
538 struct io_uring_cqe *cqe = io_get_cqe(ctx);
539 struct io_overflow_cqe *ocqe;
543 ocqe = list_first_entry(&ctx->cq_overflow_list,
544 struct io_overflow_cqe, list);
546 memcpy(cqe, &ocqe->cqe, cqe_size);
548 io_account_cq_overflow(ctx);
551 list_del(&ocqe->list);
555 all_flushed = list_empty(&ctx->cq_overflow_list);
557 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
558 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
561 io_commit_cqring(ctx);
562 spin_unlock(&ctx->completion_lock);
564 io_cqring_ev_posted(ctx);
568 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
572 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
573 /* iopoll syncs against uring_lock, not completion_lock */
574 if (ctx->flags & IORING_SETUP_IOPOLL)
575 mutex_lock(&ctx->uring_lock);
576 ret = __io_cqring_overflow_flush(ctx, false);
577 if (ctx->flags & IORING_SETUP_IOPOLL)
578 mutex_unlock(&ctx->uring_lock);
584 static void __io_put_task(struct task_struct *task, int nr)
586 struct io_uring_task *tctx = task->io_uring;
588 percpu_counter_sub(&tctx->inflight, nr);
589 if (unlikely(atomic_read(&tctx->in_idle)))
590 wake_up(&tctx->wait);
591 put_task_struct_many(task, nr);
594 /* must to be called somewhat shortly after putting a request */
595 static inline void io_put_task(struct task_struct *task, int nr)
597 if (likely(task == current))
598 task->io_uring->cached_refs += nr;
600 __io_put_task(task, nr);
603 static void io_task_refs_refill(struct io_uring_task *tctx)
605 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
607 percpu_counter_add(&tctx->inflight, refill);
608 refcount_add(refill, ¤t->usage);
609 tctx->cached_refs += refill;
612 static inline void io_get_task_refs(int nr)
614 struct io_uring_task *tctx = current->io_uring;
616 tctx->cached_refs -= nr;
617 if (unlikely(tctx->cached_refs < 0))
618 io_task_refs_refill(tctx);
621 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
623 struct io_uring_task *tctx = task->io_uring;
624 unsigned int refs = tctx->cached_refs;
627 tctx->cached_refs = 0;
628 percpu_counter_sub(&tctx->inflight, refs);
629 put_task_struct_many(task, refs);
633 bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data, s32 res,
634 u32 cflags, u64 extra1, u64 extra2)
636 struct io_overflow_cqe *ocqe;
637 size_t ocq_size = sizeof(struct io_overflow_cqe);
638 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
641 ocq_size += sizeof(struct io_uring_cqe);
643 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
644 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
647 * If we're in ring overflow flush mode, or in task cancel mode,
648 * or cannot allocate an overflow entry, then we need to drop it
651 io_account_cq_overflow(ctx);
652 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
655 if (list_empty(&ctx->cq_overflow_list)) {
656 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
657 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
660 ocqe->cqe.user_data = user_data;
662 ocqe->cqe.flags = cflags;
664 ocqe->cqe.big_cqe[0] = extra1;
665 ocqe->cqe.big_cqe[1] = extra2;
667 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
671 bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
674 struct io_uring_cqe *cqe;
677 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
680 * If we can't get a cq entry, userspace overflowed the
681 * submission (by quite a lot). Increment the overflow count in
684 cqe = io_get_cqe(ctx);
686 WRITE_ONCE(cqe->user_data, user_data);
687 WRITE_ONCE(cqe->res, res);
688 WRITE_ONCE(cqe->flags, cflags);
690 if (ctx->flags & IORING_SETUP_CQE32) {
691 WRITE_ONCE(cqe->big_cqe[0], 0);
692 WRITE_ONCE(cqe->big_cqe[1], 0);
696 return io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
699 static void __io_req_complete_put(struct io_kiocb *req)
702 * If we're the last reference to this request, add to our locked
705 if (req_ref_put_and_test(req)) {
706 struct io_ring_ctx *ctx = req->ctx;
708 if (req->flags & IO_REQ_LINK_FLAGS) {
709 if (req->flags & IO_DISARM_MASK)
712 io_req_task_queue(req->link);
716 io_req_put_rsrc(req);
718 * Selected buffer deallocation in io_clean_op() assumes that
719 * we don't hold ->completion_lock. Clean them here to avoid
722 io_put_kbuf_comp(req);
723 io_dismantle_req(req);
724 io_put_task(req->task, 1);
725 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
726 ctx->locked_free_nr++;
730 void __io_req_complete_post(struct io_kiocb *req)
732 if (!(req->flags & REQ_F_CQE_SKIP))
733 __io_fill_cqe_req(req->ctx, req);
734 __io_req_complete_put(req);
737 void io_req_complete_post(struct io_kiocb *req)
739 struct io_ring_ctx *ctx = req->ctx;
741 spin_lock(&ctx->completion_lock);
742 __io_req_complete_post(req);
743 io_commit_cqring(ctx);
744 spin_unlock(&ctx->completion_lock);
745 io_cqring_ev_posted(ctx);
748 inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags)
750 io_req_complete_post(req);
753 void io_req_complete_failed(struct io_kiocb *req, s32 res)
756 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
757 io_req_complete_post(req);
761 * Don't initialise the fields below on every allocation, but do that in
762 * advance and keep them valid across allocations.
764 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
768 req->async_data = NULL;
769 /* not necessary, but safer to zero */
773 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
774 struct io_submit_state *state)
776 spin_lock(&ctx->completion_lock);
777 wq_list_splice(&ctx->locked_free_list, &state->free_list);
778 ctx->locked_free_nr = 0;
779 spin_unlock(&ctx->completion_lock);
782 static inline bool io_req_cache_empty(struct io_ring_ctx *ctx)
784 return !ctx->submit_state.free_list.next;
788 * A request might get retired back into the request caches even before opcode
789 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
790 * Because of that, io_alloc_req() should be called only under ->uring_lock
791 * and with extra caution to not get a request that is still worked on.
793 static __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
794 __must_hold(&ctx->uring_lock)
796 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
797 void *reqs[IO_REQ_ALLOC_BATCH];
801 * If we have more than a batch's worth of requests in our IRQ side
802 * locked cache, grab the lock and move them over to our submission
805 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
806 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
807 if (!io_req_cache_empty(ctx))
811 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
814 * Bulk alloc is all-or-nothing. If we fail to get a batch,
815 * retry single alloc to be on the safe side.
817 if (unlikely(ret <= 0)) {
818 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
824 percpu_ref_get_many(&ctx->refs, ret);
825 for (i = 0; i < ret; i++) {
826 struct io_kiocb *req = reqs[i];
828 io_preinit_req(req, ctx);
829 io_req_add_to_cache(req, ctx);
834 static inline bool io_alloc_req_refill(struct io_ring_ctx *ctx)
836 if (unlikely(io_req_cache_empty(ctx)))
837 return __io_alloc_req_refill(ctx);
841 static inline struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
843 struct io_wq_work_node *node;
845 node = wq_stack_extract(&ctx->submit_state.free_list);
846 return container_of(node, struct io_kiocb, comp_list);
849 static inline void io_dismantle_req(struct io_kiocb *req)
851 unsigned int flags = req->flags;
853 if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
855 if (!(flags & REQ_F_FIXED_FILE))
856 io_put_file(req->file);
859 __cold void io_free_req(struct io_kiocb *req)
861 struct io_ring_ctx *ctx = req->ctx;
863 io_req_put_rsrc(req);
864 io_dismantle_req(req);
865 io_put_task(req->task, 1);
867 spin_lock(&ctx->completion_lock);
868 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
869 ctx->locked_free_nr++;
870 spin_unlock(&ctx->completion_lock);
873 static void __io_req_find_next_prep(struct io_kiocb *req)
875 struct io_ring_ctx *ctx = req->ctx;
878 spin_lock(&ctx->completion_lock);
879 posted = io_disarm_next(req);
880 io_commit_cqring(ctx);
881 spin_unlock(&ctx->completion_lock);
883 io_cqring_ev_posted(ctx);
886 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
888 struct io_kiocb *nxt;
891 * If LINK is set, we have dependent requests in this chain. If we
892 * didn't fail this request, queue the first one up, moving any other
893 * dependencies to the next request. In case of failure, fail the rest
896 if (unlikely(req->flags & IO_DISARM_MASK))
897 __io_req_find_next_prep(req);
903 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
907 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
908 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
910 io_submit_flush_completions(ctx);
911 mutex_unlock(&ctx->uring_lock);
914 percpu_ref_put(&ctx->refs);
917 static inline void ctx_commit_and_unlock(struct io_ring_ctx *ctx)
919 io_commit_cqring(ctx);
920 spin_unlock(&ctx->completion_lock);
921 io_cqring_ev_posted(ctx);
924 static void handle_prev_tw_list(struct io_wq_work_node *node,
925 struct io_ring_ctx **ctx, bool *uring_locked)
927 if (*ctx && !*uring_locked)
928 spin_lock(&(*ctx)->completion_lock);
931 struct io_wq_work_node *next = node->next;
932 struct io_kiocb *req = container_of(node, struct io_kiocb,
935 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
937 if (req->ctx != *ctx) {
938 if (unlikely(!*uring_locked && *ctx))
939 ctx_commit_and_unlock(*ctx);
941 ctx_flush_and_put(*ctx, uring_locked);
943 /* if not contended, grab and improve batching */
944 *uring_locked = mutex_trylock(&(*ctx)->uring_lock);
945 percpu_ref_get(&(*ctx)->refs);
946 if (unlikely(!*uring_locked))
947 spin_lock(&(*ctx)->completion_lock);
949 if (likely(*uring_locked)) {
950 req->io_task_work.func(req, uring_locked);
952 req->cqe.flags = io_put_kbuf_comp(req);
953 __io_req_complete_post(req);
958 if (unlikely(!*uring_locked))
959 ctx_commit_and_unlock(*ctx);
962 static void handle_tw_list(struct io_wq_work_node *node,
963 struct io_ring_ctx **ctx, bool *locked)
966 struct io_wq_work_node *next = node->next;
967 struct io_kiocb *req = container_of(node, struct io_kiocb,
970 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
972 if (req->ctx != *ctx) {
973 ctx_flush_and_put(*ctx, locked);
975 /* if not contended, grab and improve batching */
976 *locked = mutex_trylock(&(*ctx)->uring_lock);
977 percpu_ref_get(&(*ctx)->refs);
979 req->io_task_work.func(req, locked);
984 void tctx_task_work(struct callback_head *cb)
986 bool uring_locked = false;
987 struct io_ring_ctx *ctx = NULL;
988 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
992 struct io_wq_work_node *node1, *node2;
994 spin_lock_irq(&tctx->task_lock);
995 node1 = tctx->prio_task_list.first;
996 node2 = tctx->task_list.first;
997 INIT_WQ_LIST(&tctx->task_list);
998 INIT_WQ_LIST(&tctx->prio_task_list);
999 if (!node2 && !node1)
1000 tctx->task_running = false;
1001 spin_unlock_irq(&tctx->task_lock);
1002 if (!node2 && !node1)
1006 handle_prev_tw_list(node1, &ctx, &uring_locked);
1008 handle_tw_list(node2, &ctx, &uring_locked);
1011 if (data_race(!tctx->task_list.first) &&
1012 data_race(!tctx->prio_task_list.first) && uring_locked)
1013 io_submit_flush_completions(ctx);
1016 ctx_flush_and_put(ctx, &uring_locked);
1018 /* relaxed read is enough as only the task itself sets ->in_idle */
1019 if (unlikely(atomic_read(&tctx->in_idle)))
1020 io_uring_drop_tctx_refs(current);
1023 static void __io_req_task_work_add(struct io_kiocb *req,
1024 struct io_uring_task *tctx,
1025 struct io_wq_work_list *list)
1027 struct io_ring_ctx *ctx = req->ctx;
1028 struct io_wq_work_node *node;
1029 unsigned long flags;
1032 spin_lock_irqsave(&tctx->task_lock, flags);
1033 wq_list_add_tail(&req->io_task_work.node, list);
1034 running = tctx->task_running;
1036 tctx->task_running = true;
1037 spin_unlock_irqrestore(&tctx->task_lock, flags);
1039 /* task_work already pending, we're done */
1043 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1044 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1046 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1049 spin_lock_irqsave(&tctx->task_lock, flags);
1050 tctx->task_running = false;
1051 node = wq_list_merge(&tctx->prio_task_list, &tctx->task_list);
1052 spin_unlock_irqrestore(&tctx->task_lock, flags);
1055 req = container_of(node, struct io_kiocb, io_task_work.node);
1057 if (llist_add(&req->io_task_work.fallback_node,
1058 &req->ctx->fallback_llist))
1059 schedule_delayed_work(&req->ctx->fallback_work, 1);
1063 void io_req_task_work_add(struct io_kiocb *req)
1065 struct io_uring_task *tctx = req->task->io_uring;
1067 __io_req_task_work_add(req, tctx, &tctx->task_list);
1070 void io_req_task_prio_work_add(struct io_kiocb *req)
1072 struct io_uring_task *tctx = req->task->io_uring;
1074 if (req->ctx->flags & IORING_SETUP_SQPOLL)
1075 __io_req_task_work_add(req, tctx, &tctx->prio_task_list);
1077 __io_req_task_work_add(req, tctx, &tctx->task_list);
1080 static void io_req_tw_post(struct io_kiocb *req, bool *locked)
1082 io_req_complete_post(req);
1085 void io_req_tw_post_queue(struct io_kiocb *req, s32 res, u32 cflags)
1087 io_req_set_res(req, res, cflags);
1088 req->io_task_work.func = io_req_tw_post;
1089 io_req_task_work_add(req);
1092 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
1094 /* not needed for normal modes, but SQPOLL depends on it */
1095 io_tw_lock(req->ctx, locked);
1096 io_req_complete_failed(req, req->cqe.res);
1099 void io_req_task_submit(struct io_kiocb *req, bool *locked)
1101 io_tw_lock(req->ctx, locked);
1102 /* req->task == current here, checking PF_EXITING is safe */
1103 if (likely(!(req->task->flags & PF_EXITING)))
1106 io_req_complete_failed(req, -EFAULT);
1109 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1111 io_req_set_res(req, ret, 0);
1112 req->io_task_work.func = io_req_task_cancel;
1113 io_req_task_work_add(req);
1116 void io_req_task_queue(struct io_kiocb *req)
1118 req->io_task_work.func = io_req_task_submit;
1119 io_req_task_work_add(req);
1122 void io_queue_next(struct io_kiocb *req)
1124 struct io_kiocb *nxt = io_req_find_next(req);
1127 io_req_task_queue(nxt);
1130 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1131 __must_hold(&ctx->uring_lock)
1133 struct task_struct *task = NULL;
1137 struct io_kiocb *req = container_of(node, struct io_kiocb,
1140 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1141 if (req->flags & REQ_F_REFCOUNT) {
1142 node = req->comp_list.next;
1143 if (!req_ref_put_and_test(req))
1146 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1147 struct async_poll *apoll = req->apoll;
1149 if (apoll->double_poll)
1150 kfree(apoll->double_poll);
1151 list_add(&apoll->poll.wait.entry,
1153 req->flags &= ~REQ_F_POLLED;
1155 if (req->flags & IO_REQ_LINK_FLAGS)
1157 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1160 if (!(req->flags & REQ_F_FIXED_FILE))
1161 io_put_file(req->file);
1163 io_req_put_rsrc_locked(req, ctx);
1165 if (req->task != task) {
1167 io_put_task(task, task_refs);
1172 node = req->comp_list.next;
1173 io_req_add_to_cache(req, ctx);
1177 io_put_task(task, task_refs);
1180 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1181 __must_hold(&ctx->uring_lock)
1183 struct io_wq_work_node *node, *prev;
1184 struct io_submit_state *state = &ctx->submit_state;
1186 if (state->flush_cqes) {
1187 spin_lock(&ctx->completion_lock);
1188 wq_list_for_each(node, prev, &state->compl_reqs) {
1189 struct io_kiocb *req = container_of(node, struct io_kiocb,
1192 if (!(req->flags & REQ_F_CQE_SKIP))
1193 __io_fill_cqe_req(ctx, req);
1196 io_commit_cqring(ctx);
1197 spin_unlock(&ctx->completion_lock);
1198 io_cqring_ev_posted(ctx);
1199 state->flush_cqes = false;
1202 io_free_batch_list(ctx, state->compl_reqs.first);
1203 INIT_WQ_LIST(&state->compl_reqs);
1207 * Drop reference to request, return next in chain (if there is one) if this
1208 * was the last reference to this request.
1210 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1212 struct io_kiocb *nxt = NULL;
1214 if (req_ref_put_and_test(req)) {
1215 if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
1216 nxt = io_req_find_next(req);
1222 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1224 /* See comment at the top of this file */
1226 return __io_cqring_events(ctx);
1230 * We can't just wait for polled events to come to us, we have to actively
1231 * find and complete them.
1233 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1235 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1238 mutex_lock(&ctx->uring_lock);
1239 while (!wq_list_empty(&ctx->iopoll_list)) {
1240 /* let it sleep and repeat later if can't complete a request */
1241 if (io_do_iopoll(ctx, true) == 0)
1244 * Ensure we allow local-to-the-cpu processing to take place,
1245 * in this case we need to ensure that we reap all events.
1246 * Also let task_work, etc. to progress by releasing the mutex
1248 if (need_resched()) {
1249 mutex_unlock(&ctx->uring_lock);
1251 mutex_lock(&ctx->uring_lock);
1254 mutex_unlock(&ctx->uring_lock);
1257 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1259 unsigned int nr_events = 0;
1261 unsigned long check_cq;
1263 check_cq = READ_ONCE(ctx->check_cq);
1264 if (unlikely(check_cq)) {
1265 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1266 __io_cqring_overflow_flush(ctx, false);
1268 * Similarly do not spin if we have not informed the user of any
1271 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1275 * Don't enter poll loop if we already have events pending.
1276 * If we do, we can potentially be spinning for commands that
1277 * already triggered a CQE (eg in error).
1279 if (io_cqring_events(ctx))
1284 * If a submit got punted to a workqueue, we can have the
1285 * application entering polling for a command before it gets
1286 * issued. That app will hold the uring_lock for the duration
1287 * of the poll right here, so we need to take a breather every
1288 * now and then to ensure that the issue has a chance to add
1289 * the poll to the issued list. Otherwise we can spin here
1290 * forever, while the workqueue is stuck trying to acquire the
1293 if (wq_list_empty(&ctx->iopoll_list)) {
1294 u32 tail = ctx->cached_cq_tail;
1296 mutex_unlock(&ctx->uring_lock);
1298 mutex_lock(&ctx->uring_lock);
1300 /* some requests don't go through iopoll_list */
1301 if (tail != ctx->cached_cq_tail ||
1302 wq_list_empty(&ctx->iopoll_list))
1305 ret = io_do_iopoll(ctx, !min);
1310 } while (nr_events < min && !need_resched());
1315 void io_req_task_complete(struct io_kiocb *req, bool *locked)
1317 if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) {
1318 unsigned issue_flags = *locked ? 0 : IO_URING_F_UNLOCKED;
1320 req->cqe.flags |= io_put_kbuf(req, issue_flags);
1324 io_req_add_compl_list(req);
1326 io_req_complete_post(req);
1330 * After the iocb has been issued, it's safe to be found on the poll list.
1331 * Adding the kiocb to the list AFTER submission ensures that we don't
1332 * find it from a io_do_iopoll() thread before the issuer is done
1333 * accessing the kiocb cookie.
1335 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1337 struct io_ring_ctx *ctx = req->ctx;
1338 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1340 /* workqueue context doesn't hold uring_lock, grab it now */
1341 if (unlikely(needs_lock))
1342 mutex_lock(&ctx->uring_lock);
1345 * Track whether we have multiple files in our lists. This will impact
1346 * how we do polling eventually, not spinning if we're on potentially
1347 * different devices.
1349 if (wq_list_empty(&ctx->iopoll_list)) {
1350 ctx->poll_multi_queue = false;
1351 } else if (!ctx->poll_multi_queue) {
1352 struct io_kiocb *list_req;
1354 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1356 if (list_req->file != req->file)
1357 ctx->poll_multi_queue = true;
1361 * For fast devices, IO may have already completed. If it has, add
1362 * it to the front so we find it first.
1364 if (READ_ONCE(req->iopoll_completed))
1365 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1367 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1369 if (unlikely(needs_lock)) {
1371 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1372 * in sq thread task context or in io worker task context. If
1373 * current task context is sq thread, we don't need to check
1374 * whether should wake up sq thread.
1376 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1377 wq_has_sleeper(&ctx->sq_data->wait))
1378 wake_up(&ctx->sq_data->wait);
1380 mutex_unlock(&ctx->uring_lock);
1384 static bool io_bdev_nowait(struct block_device *bdev)
1386 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
1390 * If we tracked the file through the SCM inflight mechanism, we could support
1391 * any file. For now, just ensure that anything potentially problematic is done
1394 static bool __io_file_supports_nowait(struct file *file, umode_t mode)
1396 if (S_ISBLK(mode)) {
1397 if (IS_ENABLED(CONFIG_BLOCK) &&
1398 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
1404 if (S_ISREG(mode)) {
1405 if (IS_ENABLED(CONFIG_BLOCK) &&
1406 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
1407 !io_is_uring_fops(file))
1412 /* any ->read/write should understand O_NONBLOCK */
1413 if (file->f_flags & O_NONBLOCK)
1415 return file->f_mode & FMODE_NOWAIT;
1419 * If we tracked the file through the SCM inflight mechanism, we could support
1420 * any file. For now, just ensure that anything potentially problematic is done
1423 unsigned int io_file_get_flags(struct file *file)
1425 umode_t mode = file_inode(file)->i_mode;
1426 unsigned int res = 0;
1430 if (__io_file_supports_nowait(file, mode))
1432 if (io_file_need_scm(file))
1437 bool io_alloc_async_data(struct io_kiocb *req)
1439 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
1440 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
1441 if (req->async_data) {
1442 req->flags |= REQ_F_ASYNC_DATA;
1448 int io_req_prep_async(struct io_kiocb *req)
1450 const struct io_op_def *def = &io_op_defs[req->opcode];
1452 /* assign early for deferred execution for non-fixed file */
1453 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE))
1454 req->file = io_file_get_normal(req, req->cqe.fd);
1455 if (!def->prep_async)
1457 if (WARN_ON_ONCE(req_has_async_data(req)))
1459 if (io_alloc_async_data(req))
1462 return def->prep_async(req);
1465 static u32 io_get_sequence(struct io_kiocb *req)
1467 u32 seq = req->ctx->cached_sq_head;
1468 struct io_kiocb *cur;
1470 /* need original cached_sq_head, but it was increased for each req */
1471 io_for_each_link(cur, req)
1476 static __cold void io_drain_req(struct io_kiocb *req)
1478 struct io_ring_ctx *ctx = req->ctx;
1479 struct io_defer_entry *de;
1481 u32 seq = io_get_sequence(req);
1483 /* Still need defer if there is pending req in defer list. */
1484 spin_lock(&ctx->completion_lock);
1485 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1486 spin_unlock(&ctx->completion_lock);
1488 ctx->drain_active = false;
1489 io_req_task_queue(req);
1492 spin_unlock(&ctx->completion_lock);
1494 ret = io_req_prep_async(req);
1497 io_req_complete_failed(req, ret);
1500 io_prep_async_link(req);
1501 de = kmalloc(sizeof(*de), GFP_KERNEL);
1507 spin_lock(&ctx->completion_lock);
1508 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1509 spin_unlock(&ctx->completion_lock);
1514 trace_io_uring_defer(ctx, req, req->cqe.user_data, req->opcode);
1517 list_add_tail(&de->list, &ctx->defer_list);
1518 spin_unlock(&ctx->completion_lock);
1521 static void io_clean_op(struct io_kiocb *req)
1523 if (req->flags & REQ_F_BUFFER_SELECTED) {
1524 spin_lock(&req->ctx->completion_lock);
1525 io_put_kbuf_comp(req);
1526 spin_unlock(&req->ctx->completion_lock);
1529 if (req->flags & REQ_F_NEED_CLEANUP) {
1530 const struct io_op_def *def = &io_op_defs[req->opcode];
1535 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1536 kfree(req->apoll->double_poll);
1540 if (req->flags & REQ_F_INFLIGHT) {
1541 struct io_uring_task *tctx = req->task->io_uring;
1543 atomic_dec(&tctx->inflight_tracked);
1545 if (req->flags & REQ_F_CREDS)
1546 put_cred(req->creds);
1547 if (req->flags & REQ_F_ASYNC_DATA) {
1548 kfree(req->async_data);
1549 req->async_data = NULL;
1551 req->flags &= ~IO_REQ_CLEAN_FLAGS;
1554 static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags)
1556 if (req->file || !io_op_defs[req->opcode].needs_file)
1559 if (req->flags & REQ_F_FIXED_FILE)
1560 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1562 req->file = io_file_get_normal(req, req->cqe.fd);
1567 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1569 const struct io_op_def *def = &io_op_defs[req->opcode];
1570 const struct cred *creds = NULL;
1573 if (unlikely(!io_assign_file(req, issue_flags)))
1576 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1577 creds = override_creds(req->creds);
1579 if (!def->audit_skip)
1580 audit_uring_entry(req->opcode);
1582 ret = def->issue(req, issue_flags);
1584 if (!def->audit_skip)
1585 audit_uring_exit(!ret, ret);
1588 revert_creds(creds);
1590 if (ret == IOU_OK) {
1591 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1592 io_req_add_compl_list(req);
1594 io_req_complete_post(req);
1595 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1598 /* If the op doesn't have a file, we're not polling for it */
1599 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && req->file)
1600 io_iopoll_req_issued(req, issue_flags);
1605 int io_poll_issue(struct io_kiocb *req, bool *locked)
1607 io_tw_lock(req->ctx, locked);
1608 if (unlikely(req->task->flags & PF_EXITING))
1610 return io_issue_sqe(req, IO_URING_F_NONBLOCK);
1613 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1615 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1617 req = io_put_req_find_next(req);
1618 return req ? &req->work : NULL;
1621 void io_wq_submit_work(struct io_wq_work *work)
1623 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1624 const struct io_op_def *def = &io_op_defs[req->opcode];
1625 unsigned int issue_flags = IO_URING_F_UNLOCKED;
1626 bool needs_poll = false;
1627 int ret = 0, err = -ECANCELED;
1629 /* one will be dropped by ->io_free_work() after returning to io-wq */
1630 if (!(req->flags & REQ_F_REFCOUNT))
1631 __io_req_set_refcount(req, 2);
1635 io_arm_ltimeout(req);
1637 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1638 if (work->flags & IO_WQ_WORK_CANCEL) {
1640 io_req_task_queue_fail(req, err);
1643 if (!io_assign_file(req, issue_flags)) {
1645 work->flags |= IO_WQ_WORK_CANCEL;
1649 if (req->flags & REQ_F_FORCE_ASYNC) {
1650 bool opcode_poll = def->pollin || def->pollout;
1652 if (opcode_poll && file_can_poll(req->file)) {
1654 issue_flags |= IO_URING_F_NONBLOCK;
1659 ret = io_issue_sqe(req, issue_flags);
1663 * We can get EAGAIN for iopolled IO even though we're
1664 * forcing a sync submission from here, since we can't
1665 * wait for request slots on the block side.
1668 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1674 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1676 /* aborted or ready, in either case retry blocking */
1678 issue_flags &= ~IO_URING_F_NONBLOCK;
1681 /* avoid locking problems by failing it from a clean context */
1683 io_req_task_queue_fail(req, ret);
1686 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1687 unsigned int issue_flags)
1689 struct io_ring_ctx *ctx = req->ctx;
1690 struct file *file = NULL;
1691 unsigned long file_ptr;
1693 io_ring_submit_lock(ctx, issue_flags);
1695 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
1697 fd = array_index_nospec(fd, ctx->nr_user_files);
1698 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
1699 file = (struct file *) (file_ptr & FFS_MASK);
1700 file_ptr &= ~FFS_MASK;
1701 /* mask in overlapping REQ_F and FFS bits */
1702 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
1703 io_req_set_rsrc_node(req, ctx, 0);
1704 WARN_ON_ONCE(file && !test_bit(fd, ctx->file_table.bitmap));
1706 io_ring_submit_unlock(ctx, issue_flags);
1710 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
1712 struct file *file = fget(fd);
1714 trace_io_uring_file_get(req->ctx, req, req->cqe.user_data, fd);
1716 /* we don't allow fixed io_uring files */
1717 if (file && io_is_uring_fops(file))
1718 io_req_track_inflight(req);
1722 static void io_queue_async(struct io_kiocb *req, int ret)
1723 __must_hold(&req->ctx->uring_lock)
1725 struct io_kiocb *linked_timeout;
1727 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
1728 io_req_complete_failed(req, ret);
1732 linked_timeout = io_prep_linked_timeout(req);
1734 switch (io_arm_poll_handler(req, 0)) {
1735 case IO_APOLL_READY:
1736 io_req_task_queue(req);
1738 case IO_APOLL_ABORTED:
1740 * Queued up for async execution, worker will release
1741 * submit reference when the iocb is actually submitted.
1743 io_kbuf_recycle(req, 0);
1744 io_queue_iowq(req, NULL);
1751 io_queue_linked_timeout(linked_timeout);
1754 static inline void io_queue_sqe(struct io_kiocb *req)
1755 __must_hold(&req->ctx->uring_lock)
1759 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
1762 * We async punt it if the file wasn't marked NOWAIT, or if the file
1763 * doesn't support non-blocking read/write attempts
1766 io_arm_ltimeout(req);
1768 io_queue_async(req, ret);
1771 static void io_queue_sqe_fallback(struct io_kiocb *req)
1772 __must_hold(&req->ctx->uring_lock)
1774 if (unlikely(req->flags & REQ_F_FAIL)) {
1776 * We don't submit, fail them all, for that replace hardlinks
1777 * with normal links. Extra REQ_F_LINK is tolerated.
1779 req->flags &= ~REQ_F_HARDLINK;
1780 req->flags |= REQ_F_LINK;
1781 io_req_complete_failed(req, req->cqe.res);
1782 } else if (unlikely(req->ctx->drain_active)) {
1785 int ret = io_req_prep_async(req);
1788 io_req_complete_failed(req, ret);
1790 io_queue_iowq(req, NULL);
1795 * Check SQE restrictions (opcode and flags).
1797 * Returns 'true' if SQE is allowed, 'false' otherwise.
1799 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
1800 struct io_kiocb *req,
1801 unsigned int sqe_flags)
1803 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
1806 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
1807 ctx->restrictions.sqe_flags_required)
1810 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
1811 ctx->restrictions.sqe_flags_required))
1817 static void io_init_req_drain(struct io_kiocb *req)
1819 struct io_ring_ctx *ctx = req->ctx;
1820 struct io_kiocb *head = ctx->submit_state.link.head;
1822 ctx->drain_active = true;
1825 * If we need to drain a request in the middle of a link, drain
1826 * the head request and the next request/link after the current
1827 * link. Considering sequential execution of links,
1828 * REQ_F_IO_DRAIN will be maintained for every request of our
1831 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
1832 ctx->drain_next = true;
1836 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
1837 const struct io_uring_sqe *sqe)
1838 __must_hold(&ctx->uring_lock)
1840 const struct io_op_def *def;
1841 unsigned int sqe_flags;
1845 /* req is partially pre-initialised, see io_preinit_req() */
1846 req->opcode = opcode = READ_ONCE(sqe->opcode);
1847 /* same numerical values with corresponding REQ_F_*, safe to copy */
1848 req->flags = sqe_flags = READ_ONCE(sqe->flags);
1849 req->cqe.user_data = READ_ONCE(sqe->user_data);
1851 req->rsrc_node = NULL;
1852 req->task = current;
1854 if (unlikely(opcode >= IORING_OP_LAST)) {
1858 def = &io_op_defs[opcode];
1859 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
1860 /* enforce forwards compatibility on users */
1861 if (sqe_flags & ~SQE_VALID_FLAGS)
1863 if (sqe_flags & IOSQE_BUFFER_SELECT) {
1864 if (!def->buffer_select)
1866 req->buf_index = READ_ONCE(sqe->buf_group);
1868 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
1869 ctx->drain_disabled = true;
1870 if (sqe_flags & IOSQE_IO_DRAIN) {
1871 if (ctx->drain_disabled)
1873 io_init_req_drain(req);
1876 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
1877 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
1879 /* knock it to the slow queue path, will be drained there */
1880 if (ctx->drain_active)
1881 req->flags |= REQ_F_FORCE_ASYNC;
1882 /* if there is no link, we're at "next" request and need to drain */
1883 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
1884 ctx->drain_next = false;
1885 ctx->drain_active = true;
1886 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
1890 if (!def->ioprio && sqe->ioprio)
1892 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
1895 if (def->needs_file) {
1896 struct io_submit_state *state = &ctx->submit_state;
1898 req->cqe.fd = READ_ONCE(sqe->fd);
1901 * Plug now if we have more than 2 IO left after this, and the
1902 * target is potentially a read/write to block based storage.
1904 if (state->need_plug && def->plug) {
1905 state->plug_started = true;
1906 state->need_plug = false;
1907 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
1911 personality = READ_ONCE(sqe->personality);
1915 req->creds = xa_load(&ctx->personalities, personality);
1918 get_cred(req->creds);
1919 ret = security_uring_override_creds(req->creds);
1921 put_cred(req->creds);
1924 req->flags |= REQ_F_CREDS;
1927 return def->prep(req, sqe);
1930 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
1931 struct io_kiocb *req, int ret)
1933 struct io_ring_ctx *ctx = req->ctx;
1934 struct io_submit_link *link = &ctx->submit_state.link;
1935 struct io_kiocb *head = link->head;
1937 trace_io_uring_req_failed(sqe, ctx, req, ret);
1940 * Avoid breaking links in the middle as it renders links with SQPOLL
1941 * unusable. Instead of failing eagerly, continue assembling the link if
1942 * applicable and mark the head with REQ_F_FAIL. The link flushing code
1943 * should find the flag and handle the rest.
1945 req_fail_link_node(req, ret);
1946 if (head && !(head->flags & REQ_F_FAIL))
1947 req_fail_link_node(head, -ECANCELED);
1949 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
1951 link->last->link = req;
1955 io_queue_sqe_fallback(req);
1960 link->last->link = req;
1967 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
1968 const struct io_uring_sqe *sqe)
1969 __must_hold(&ctx->uring_lock)
1971 struct io_submit_link *link = &ctx->submit_state.link;
1974 ret = io_init_req(ctx, req, sqe);
1976 return io_submit_fail_init(sqe, req, ret);
1978 /* don't need @sqe from now on */
1979 trace_io_uring_submit_sqe(ctx, req, req->cqe.user_data, req->opcode,
1981 ctx->flags & IORING_SETUP_SQPOLL);
1984 * If we already have a head request, queue this one for async
1985 * submittal once the head completes. If we don't have a head but
1986 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
1987 * submitted sync once the chain is complete. If none of those
1988 * conditions are true (normal request), then just queue it.
1990 if (unlikely(link->head)) {
1991 ret = io_req_prep_async(req);
1993 return io_submit_fail_init(sqe, req, ret);
1995 trace_io_uring_link(ctx, req, link->head);
1996 link->last->link = req;
1999 if (req->flags & IO_REQ_LINK_FLAGS)
2001 /* last request of the link, flush it */
2004 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2007 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2008 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2009 if (req->flags & IO_REQ_LINK_FLAGS) {
2014 io_queue_sqe_fallback(req);
2024 * Batched submission is done, ensure local IO is flushed out.
2026 static void io_submit_state_end(struct io_ring_ctx *ctx)
2028 struct io_submit_state *state = &ctx->submit_state;
2030 if (unlikely(state->link.head))
2031 io_queue_sqe_fallback(state->link.head);
2032 /* flush only after queuing links as they can generate completions */
2033 io_submit_flush_completions(ctx);
2034 if (state->plug_started)
2035 blk_finish_plug(&state->plug);
2039 * Start submission side cache.
2041 static void io_submit_state_start(struct io_submit_state *state,
2042 unsigned int max_ios)
2044 state->plug_started = false;
2045 state->need_plug = max_ios > 2;
2046 state->submit_nr = max_ios;
2047 /* set only head, no need to init link_last in advance */
2048 state->link.head = NULL;
2051 static void io_commit_sqring(struct io_ring_ctx *ctx)
2053 struct io_rings *rings = ctx->rings;
2056 * Ensure any loads from the SQEs are done at this point,
2057 * since once we write the new head, the application could
2058 * write new data to them.
2060 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2064 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2065 * that is mapped by userspace. This means that care needs to be taken to
2066 * ensure that reads are stable, as we cannot rely on userspace always
2067 * being a good citizen. If members of the sqe are validated and then later
2068 * used, it's important that those reads are done through READ_ONCE() to
2069 * prevent a re-load down the line.
2071 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
2073 unsigned head, mask = ctx->sq_entries - 1;
2074 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2077 * The cached sq head (or cq tail) serves two purposes:
2079 * 1) allows us to batch the cost of updating the user visible
2081 * 2) allows the kernel side to track the head on its own, even
2082 * though the application is the one updating it.
2084 head = READ_ONCE(ctx->sq_array[sq_idx]);
2085 if (likely(head < ctx->sq_entries)) {
2086 /* double index for 128-byte SQEs, twice as long */
2087 if (ctx->flags & IORING_SETUP_SQE128)
2089 return &ctx->sq_sqes[head];
2092 /* drop invalid entries */
2094 WRITE_ONCE(ctx->rings->sq_dropped,
2095 READ_ONCE(ctx->rings->sq_dropped) + 1);
2099 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2100 __must_hold(&ctx->uring_lock)
2102 unsigned int entries = io_sqring_entries(ctx);
2106 if (unlikely(!entries))
2108 /* make sure SQ entry isn't read before tail */
2109 ret = left = min3(nr, ctx->sq_entries, entries);
2110 io_get_task_refs(left);
2111 io_submit_state_start(&ctx->submit_state, left);
2114 const struct io_uring_sqe *sqe;
2115 struct io_kiocb *req;
2117 if (unlikely(!io_alloc_req_refill(ctx)))
2119 req = io_alloc_req(ctx);
2120 sqe = io_get_sqe(ctx);
2121 if (unlikely(!sqe)) {
2122 io_req_add_to_cache(req, ctx);
2127 * Continue submitting even for sqe failure if the
2128 * ring was setup with IORING_SETUP_SUBMIT_ALL
2130 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2131 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2137 if (unlikely(left)) {
2139 /* try again if it submitted nothing and can't allocate a req */
2140 if (!ret && io_req_cache_empty(ctx))
2142 current->io_uring->cached_refs += left;
2145 io_submit_state_end(ctx);
2146 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2147 io_commit_sqring(ctx);
2151 struct io_wait_queue {
2152 struct wait_queue_entry wq;
2153 struct io_ring_ctx *ctx;
2155 unsigned nr_timeouts;
2158 static inline bool io_should_wake(struct io_wait_queue *iowq)
2160 struct io_ring_ctx *ctx = iowq->ctx;
2161 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
2164 * Wake up if we have enough events, or if a timeout occurred since we
2165 * started waiting. For timeouts, we always want to return to userspace,
2166 * regardless of event count.
2168 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2171 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2172 int wake_flags, void *key)
2174 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
2178 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2179 * the task, and the next invocation will do it.
2181 if (io_should_wake(iowq) ||
2182 test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &iowq->ctx->check_cq))
2183 return autoremove_wake_function(curr, mode, wake_flags, key);
2187 int io_run_task_work_sig(void)
2189 if (io_run_task_work())
2191 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
2192 return -ERESTARTSYS;
2193 if (task_sigpending(current))
2198 /* when returns >0, the caller should retry */
2199 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2200 struct io_wait_queue *iowq,
2204 unsigned long check_cq;
2206 /* make sure we run task_work before checking for signals */
2207 ret = io_run_task_work_sig();
2208 if (ret || io_should_wake(iowq))
2211 check_cq = READ_ONCE(ctx->check_cq);
2212 if (unlikely(check_cq)) {
2213 /* let the caller flush overflows, retry */
2214 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2216 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
2219 if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS))
2225 * Wait until events become available, if we don't already have some. The
2226 * application must reap them itself, as they reside on the shared cq ring.
2228 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2229 const sigset_t __user *sig, size_t sigsz,
2230 struct __kernel_timespec __user *uts)
2232 struct io_wait_queue iowq;
2233 struct io_rings *rings = ctx->rings;
2234 ktime_t timeout = KTIME_MAX;
2238 io_cqring_overflow_flush(ctx);
2239 if (io_cqring_events(ctx) >= min_events)
2241 if (!io_run_task_work())
2246 #ifdef CONFIG_COMPAT
2247 if (in_compat_syscall())
2248 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2252 ret = set_user_sigmask(sig, sigsz);
2259 struct timespec64 ts;
2261 if (get_timespec64(&ts, uts))
2263 timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2266 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2267 iowq.wq.private = current;
2268 INIT_LIST_HEAD(&iowq.wq.entry);
2270 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2271 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2273 trace_io_uring_cqring_wait(ctx, min_events);
2275 /* if we can't even flush overflow, don't wait for more */
2276 if (!io_cqring_overflow_flush(ctx)) {
2280 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2281 TASK_INTERRUPTIBLE);
2282 ret = io_cqring_wait_schedule(ctx, &iowq, timeout);
2286 finish_wait(&ctx->cq_wait, &iowq.wq);
2287 restore_saved_sigmask_unless(ret == -EINTR);
2289 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2292 static void io_mem_free(void *ptr)
2299 page = virt_to_head_page(ptr);
2300 if (put_page_testzero(page))
2301 free_compound_page(page);
2304 static void *io_mem_alloc(size_t size)
2306 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2308 return (void *) __get_free_pages(gfp, get_order(size));
2311 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2312 unsigned int cq_entries, size_t *sq_offset)
2314 struct io_rings *rings;
2315 size_t off, sq_array_size;
2317 off = struct_size(rings, cqes, cq_entries);
2318 if (off == SIZE_MAX)
2320 if (ctx->flags & IORING_SETUP_CQE32) {
2321 if (check_shl_overflow(off, 1, &off))
2326 off = ALIGN(off, SMP_CACHE_BYTES);
2334 sq_array_size = array_size(sizeof(u32), sq_entries);
2335 if (sq_array_size == SIZE_MAX)
2338 if (check_add_overflow(off, sq_array_size, &off))
2344 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2345 unsigned int eventfd_async)
2347 struct io_ev_fd *ev_fd;
2348 __s32 __user *fds = arg;
2351 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2352 lockdep_is_held(&ctx->uring_lock));
2356 if (copy_from_user(&fd, fds, sizeof(*fds)))
2359 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2363 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2364 if (IS_ERR(ev_fd->cq_ev_fd)) {
2365 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2369 ev_fd->eventfd_async = eventfd_async;
2370 ctx->has_evfd = true;
2371 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2375 static void io_eventfd_put(struct rcu_head *rcu)
2377 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
2379 eventfd_ctx_put(ev_fd->cq_ev_fd);
2383 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2385 struct io_ev_fd *ev_fd;
2387 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2388 lockdep_is_held(&ctx->uring_lock));
2390 ctx->has_evfd = false;
2391 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2392 call_rcu(&ev_fd->rcu, io_eventfd_put);
2399 static void io_req_caches_free(struct io_ring_ctx *ctx)
2401 struct io_submit_state *state = &ctx->submit_state;
2404 mutex_lock(&ctx->uring_lock);
2405 io_flush_cached_locked_reqs(ctx, state);
2407 while (!io_req_cache_empty(ctx)) {
2408 struct io_wq_work_node *node;
2409 struct io_kiocb *req;
2411 node = wq_stack_extract(&state->free_list);
2412 req = container_of(node, struct io_kiocb, comp_list);
2413 kmem_cache_free(req_cachep, req);
2417 percpu_ref_put_many(&ctx->refs, nr);
2418 mutex_unlock(&ctx->uring_lock);
2421 static void io_flush_apoll_cache(struct io_ring_ctx *ctx)
2423 struct async_poll *apoll;
2425 while (!list_empty(&ctx->apoll_cache)) {
2426 apoll = list_first_entry(&ctx->apoll_cache, struct async_poll,
2428 list_del(&apoll->poll.wait.entry);
2433 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2435 io_sq_thread_finish(ctx);
2437 if (ctx->mm_account) {
2438 mmdrop(ctx->mm_account);
2439 ctx->mm_account = NULL;
2442 io_rsrc_refs_drop(ctx);
2443 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2444 io_wait_rsrc_data(ctx->buf_data);
2445 io_wait_rsrc_data(ctx->file_data);
2447 mutex_lock(&ctx->uring_lock);
2449 __io_sqe_buffers_unregister(ctx);
2451 __io_sqe_files_unregister(ctx);
2453 __io_cqring_overflow_flush(ctx, true);
2454 io_eventfd_unregister(ctx);
2455 io_flush_apoll_cache(ctx);
2456 mutex_unlock(&ctx->uring_lock);
2457 io_destroy_buffers(ctx);
2459 put_cred(ctx->sq_creds);
2461 /* there are no registered resources left, nobody uses it */
2463 io_rsrc_node_destroy(ctx->rsrc_node);
2464 if (ctx->rsrc_backup_node)
2465 io_rsrc_node_destroy(ctx->rsrc_backup_node);
2466 flush_delayed_work(&ctx->rsrc_put_work);
2467 flush_delayed_work(&ctx->fallback_work);
2469 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2470 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
2472 #if defined(CONFIG_UNIX)
2473 if (ctx->ring_sock) {
2474 ctx->ring_sock->file = NULL; /* so that iput() is called */
2475 sock_release(ctx->ring_sock);
2478 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2480 io_mem_free(ctx->rings);
2481 io_mem_free(ctx->sq_sqes);
2483 percpu_ref_exit(&ctx->refs);
2484 free_uid(ctx->user);
2485 io_req_caches_free(ctx);
2487 io_wq_put_hash(ctx->hash_map);
2488 kfree(ctx->cancel_hash);
2489 kfree(ctx->dummy_ubuf);
2491 xa_destroy(&ctx->io_bl_xa);
2495 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2497 struct io_ring_ctx *ctx = file->private_data;
2500 poll_wait(file, &ctx->cq_wait, wait);
2502 * synchronizes with barrier from wq_has_sleeper call in
2506 if (!io_sqring_full(ctx))
2507 mask |= EPOLLOUT | EPOLLWRNORM;
2510 * Don't flush cqring overflow list here, just do a simple check.
2511 * Otherwise there could possible be ABBA deadlock:
2514 * lock(&ctx->uring_lock);
2516 * lock(&ctx->uring_lock);
2519 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2520 * pushs them to do the flush.
2522 if (io_cqring_events(ctx) ||
2523 test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
2524 mask |= EPOLLIN | EPOLLRDNORM;
2529 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2531 const struct cred *creds;
2533 creds = xa_erase(&ctx->personalities, id);
2542 struct io_tctx_exit {
2543 struct callback_head task_work;
2544 struct completion completion;
2545 struct io_ring_ctx *ctx;
2548 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2550 struct io_uring_task *tctx = current->io_uring;
2551 struct io_tctx_exit *work;
2553 work = container_of(cb, struct io_tctx_exit, task_work);
2555 * When @in_idle, we're in cancellation and it's racy to remove the
2556 * node. It'll be removed by the end of cancellation, just ignore it.
2558 if (!atomic_read(&tctx->in_idle))
2559 io_uring_del_tctx_node((unsigned long)work->ctx);
2560 complete(&work->completion);
2563 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2565 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2567 return req->ctx == data;
2570 static __cold void io_ring_exit_work(struct work_struct *work)
2572 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2573 unsigned long timeout = jiffies + HZ * 60 * 5;
2574 unsigned long interval = HZ / 20;
2575 struct io_tctx_exit exit;
2576 struct io_tctx_node *node;
2580 * If we're doing polled IO and end up having requests being
2581 * submitted async (out-of-line), then completions can come in while
2582 * we're waiting for refs to drop. We need to reap these manually,
2583 * as nobody else will be looking for them.
2586 io_uring_try_cancel_requests(ctx, NULL, true);
2588 struct io_sq_data *sqd = ctx->sq_data;
2589 struct task_struct *tsk;
2591 io_sq_thread_park(sqd);
2593 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
2594 io_wq_cancel_cb(tsk->io_uring->io_wq,
2595 io_cancel_ctx_cb, ctx, true);
2596 io_sq_thread_unpark(sqd);
2599 io_req_caches_free(ctx);
2601 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
2602 /* there is little hope left, don't run it too often */
2605 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
2607 init_completion(&exit.completion);
2608 init_task_work(&exit.task_work, io_tctx_exit_cb);
2611 * Some may use context even when all refs and requests have been put,
2612 * and they are free to do so while still holding uring_lock or
2613 * completion_lock, see io_req_task_submit(). Apart from other work,
2614 * this lock/unlock section also waits them to finish.
2616 mutex_lock(&ctx->uring_lock);
2617 while (!list_empty(&ctx->tctx_list)) {
2618 WARN_ON_ONCE(time_after(jiffies, timeout));
2620 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
2622 /* don't spin on a single task if cancellation failed */
2623 list_rotate_left(&ctx->tctx_list);
2624 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
2625 if (WARN_ON_ONCE(ret))
2628 mutex_unlock(&ctx->uring_lock);
2629 wait_for_completion(&exit.completion);
2630 mutex_lock(&ctx->uring_lock);
2632 mutex_unlock(&ctx->uring_lock);
2633 spin_lock(&ctx->completion_lock);
2634 spin_unlock(&ctx->completion_lock);
2636 io_ring_ctx_free(ctx);
2639 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
2641 unsigned long index;
2642 struct creds *creds;
2644 mutex_lock(&ctx->uring_lock);
2645 percpu_ref_kill(&ctx->refs);
2647 __io_cqring_overflow_flush(ctx, true);
2648 xa_for_each(&ctx->personalities, index, creds)
2649 io_unregister_personality(ctx, index);
2650 mutex_unlock(&ctx->uring_lock);
2652 /* failed during ring init, it couldn't have issued any requests */
2654 io_kill_timeouts(ctx, NULL, true);
2655 io_poll_remove_all(ctx, NULL, true);
2656 /* if we failed setting up the ctx, we might not have any rings */
2657 io_iopoll_try_reap_events(ctx);
2660 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
2662 * Use system_unbound_wq to avoid spawning tons of event kworkers
2663 * if we're exiting a ton of rings at the same time. It just adds
2664 * noise and overhead, there's no discernable change in runtime
2665 * over using system_wq.
2667 queue_work(system_unbound_wq, &ctx->exit_work);
2670 static int io_uring_release(struct inode *inode, struct file *file)
2672 struct io_ring_ctx *ctx = file->private_data;
2674 file->private_data = NULL;
2675 io_ring_ctx_wait_and_kill(ctx);
2679 struct io_task_cancel {
2680 struct task_struct *task;
2684 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
2686 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2687 struct io_task_cancel *cancel = data;
2689 return io_match_task_safe(req, cancel->task, cancel->all);
2692 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
2693 struct task_struct *task,
2696 struct io_defer_entry *de;
2699 spin_lock(&ctx->completion_lock);
2700 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
2701 if (io_match_task_safe(de->req, task, cancel_all)) {
2702 list_cut_position(&list, &ctx->defer_list, &de->list);
2706 spin_unlock(&ctx->completion_lock);
2707 if (list_empty(&list))
2710 while (!list_empty(&list)) {
2711 de = list_first_entry(&list, struct io_defer_entry, list);
2712 list_del_init(&de->list);
2713 io_req_complete_failed(de->req, -ECANCELED);
2719 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
2721 struct io_tctx_node *node;
2722 enum io_wq_cancel cret;
2725 mutex_lock(&ctx->uring_lock);
2726 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
2727 struct io_uring_task *tctx = node->task->io_uring;
2730 * io_wq will stay alive while we hold uring_lock, because it's
2731 * killed after ctx nodes, which requires to take the lock.
2733 if (!tctx || !tctx->io_wq)
2735 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
2736 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2738 mutex_unlock(&ctx->uring_lock);
2743 static __cold void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
2744 struct task_struct *task,
2747 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
2748 struct io_uring_task *tctx = task ? task->io_uring : NULL;
2750 /* failed during ring init, it couldn't have issued any requests */
2755 enum io_wq_cancel cret;
2759 ret |= io_uring_try_cancel_iowq(ctx);
2760 } else if (tctx && tctx->io_wq) {
2762 * Cancels requests of all rings, not only @ctx, but
2763 * it's fine as the task is in exit/exec.
2765 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
2767 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2770 /* SQPOLL thread does its own polling */
2771 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
2772 (ctx->sq_data && ctx->sq_data->thread == current)) {
2773 while (!wq_list_empty(&ctx->iopoll_list)) {
2774 io_iopoll_try_reap_events(ctx);
2779 ret |= io_cancel_defer_files(ctx, task, cancel_all);
2780 ret |= io_poll_remove_all(ctx, task, cancel_all);
2781 ret |= io_kill_timeouts(ctx, task, cancel_all);
2783 ret |= io_run_task_work();
2790 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
2793 return atomic_read(&tctx->inflight_tracked);
2794 return percpu_counter_sum(&tctx->inflight);
2798 * Find any io_uring ctx that this task has registered or done IO on, and cancel
2799 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
2801 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
2803 struct io_uring_task *tctx = current->io_uring;
2804 struct io_ring_ctx *ctx;
2808 WARN_ON_ONCE(sqd && sqd->thread != current);
2810 if (!current->io_uring)
2813 io_wq_exit_start(tctx->io_wq);
2815 atomic_inc(&tctx->in_idle);
2817 io_uring_drop_tctx_refs(current);
2818 /* read completions before cancelations */
2819 inflight = tctx_inflight(tctx, !cancel_all);
2824 struct io_tctx_node *node;
2825 unsigned long index;
2827 xa_for_each(&tctx->xa, index, node) {
2828 /* sqpoll task will cancel all its requests */
2829 if (node->ctx->sq_data)
2831 io_uring_try_cancel_requests(node->ctx, current,
2835 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
2836 io_uring_try_cancel_requests(ctx, current,
2840 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
2842 io_uring_drop_tctx_refs(current);
2845 * If we've seen completions, retry without waiting. This
2846 * avoids a race where a completion comes in before we did
2847 * prepare_to_wait().
2849 if (inflight == tctx_inflight(tctx, !cancel_all))
2851 finish_wait(&tctx->wait, &wait);
2854 io_uring_clean_tctx(tctx);
2857 * We shouldn't run task_works after cancel, so just leave
2858 * ->in_idle set for normal exit.
2860 atomic_dec(&tctx->in_idle);
2861 /* for exec all current's requests should be gone, kill tctx */
2862 __io_uring_free(current);
2866 void __io_uring_cancel(bool cancel_all)
2868 io_uring_cancel_generic(cancel_all, NULL);
2871 static void *io_uring_validate_mmap_request(struct file *file,
2872 loff_t pgoff, size_t sz)
2874 struct io_ring_ctx *ctx = file->private_data;
2875 loff_t offset = pgoff << PAGE_SHIFT;
2880 case IORING_OFF_SQ_RING:
2881 case IORING_OFF_CQ_RING:
2884 case IORING_OFF_SQES:
2888 return ERR_PTR(-EINVAL);
2891 page = virt_to_head_page(ptr);
2892 if (sz > page_size(page))
2893 return ERR_PTR(-EINVAL);
2900 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
2902 size_t sz = vma->vm_end - vma->vm_start;
2906 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
2908 return PTR_ERR(ptr);
2910 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
2911 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
2914 #else /* !CONFIG_MMU */
2916 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
2918 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
2921 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
2923 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
2926 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
2927 unsigned long addr, unsigned long len,
2928 unsigned long pgoff, unsigned long flags)
2932 ptr = io_uring_validate_mmap_request(file, pgoff, len);
2934 return PTR_ERR(ptr);
2936 return (unsigned long) ptr;
2939 #endif /* !CONFIG_MMU */
2941 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
2943 if (flags & IORING_ENTER_EXT_ARG) {
2944 struct io_uring_getevents_arg arg;
2946 if (argsz != sizeof(arg))
2948 if (copy_from_user(&arg, argp, sizeof(arg)))
2954 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
2955 struct __kernel_timespec __user **ts,
2956 const sigset_t __user **sig)
2958 struct io_uring_getevents_arg arg;
2961 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
2962 * is just a pointer to the sigset_t.
2964 if (!(flags & IORING_ENTER_EXT_ARG)) {
2965 *sig = (const sigset_t __user *) argp;
2971 * EXT_ARG is set - ensure we agree on the size of it and copy in our
2972 * timespec and sigset_t pointers if good.
2974 if (*argsz != sizeof(arg))
2976 if (copy_from_user(&arg, argp, sizeof(arg)))
2980 *sig = u64_to_user_ptr(arg.sigmask);
2981 *argsz = arg.sigmask_sz;
2982 *ts = u64_to_user_ptr(arg.ts);
2986 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
2987 u32, min_complete, u32, flags, const void __user *, argp,
2990 struct io_ring_ctx *ctx;
2996 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
2997 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
2998 IORING_ENTER_REGISTERED_RING)))
3002 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3003 * need only dereference our task private array to find it.
3005 if (flags & IORING_ENTER_REGISTERED_RING) {
3006 struct io_uring_task *tctx = current->io_uring;
3008 if (!tctx || fd >= IO_RINGFD_REG_MAX)
3010 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3011 f.file = tctx->registered_rings[fd];
3017 if (unlikely(!f.file))
3021 if (unlikely(!io_is_uring_fops(f.file)))
3025 ctx = f.file->private_data;
3026 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
3030 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3034 * For SQ polling, the thread will do all submissions and completions.
3035 * Just return the requested submit count, and wake the thread if
3039 if (ctx->flags & IORING_SETUP_SQPOLL) {
3040 io_cqring_overflow_flush(ctx);
3042 if (unlikely(ctx->sq_data->thread == NULL)) {
3046 if (flags & IORING_ENTER_SQ_WAKEUP)
3047 wake_up(&ctx->sq_data->wait);
3048 if (flags & IORING_ENTER_SQ_WAIT) {
3049 ret = io_sqpoll_wait_sq(ctx);
3054 } else if (to_submit) {
3055 ret = io_uring_add_tctx_node(ctx);
3059 mutex_lock(&ctx->uring_lock);
3060 ret = io_submit_sqes(ctx, to_submit);
3061 if (ret != to_submit) {
3062 mutex_unlock(&ctx->uring_lock);
3065 if ((flags & IORING_ENTER_GETEVENTS) && ctx->syscall_iopoll)
3067 mutex_unlock(&ctx->uring_lock);
3069 if (flags & IORING_ENTER_GETEVENTS) {
3071 if (ctx->syscall_iopoll) {
3073 * We disallow the app entering submit/complete with
3074 * polling, but we still need to lock the ring to
3075 * prevent racing with polled issue that got punted to
3078 mutex_lock(&ctx->uring_lock);
3080 ret2 = io_validate_ext_arg(flags, argp, argsz);
3081 if (likely(!ret2)) {
3082 min_complete = min(min_complete,
3084 ret2 = io_iopoll_check(ctx, min_complete);
3086 mutex_unlock(&ctx->uring_lock);
3088 const sigset_t __user *sig;
3089 struct __kernel_timespec __user *ts;
3091 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3092 if (likely(!ret2)) {
3093 min_complete = min(min_complete,
3095 ret2 = io_cqring_wait(ctx, min_complete, sig,
3104 * EBADR indicates that one or more CQE were dropped.
3105 * Once the user has been informed we can clear the bit
3106 * as they are obviously ok with those drops.
3108 if (unlikely(ret2 == -EBADR))
3109 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3115 percpu_ref_put(&ctx->refs);
3121 static const struct file_operations io_uring_fops = {
3122 .release = io_uring_release,
3123 .mmap = io_uring_mmap,
3125 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3126 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3128 .poll = io_uring_poll,
3129 #ifdef CONFIG_PROC_FS
3130 .show_fdinfo = io_uring_show_fdinfo,
3134 bool io_is_uring_fops(struct file *file)
3136 return file->f_op == &io_uring_fops;
3139 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3140 struct io_uring_params *p)
3142 struct io_rings *rings;
3143 size_t size, sq_array_offset;
3145 /* make sure these are sane, as we already accounted them */
3146 ctx->sq_entries = p->sq_entries;
3147 ctx->cq_entries = p->cq_entries;
3149 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3150 if (size == SIZE_MAX)
3153 rings = io_mem_alloc(size);
3158 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3159 rings->sq_ring_mask = p->sq_entries - 1;
3160 rings->cq_ring_mask = p->cq_entries - 1;
3161 rings->sq_ring_entries = p->sq_entries;
3162 rings->cq_ring_entries = p->cq_entries;
3164 if (p->flags & IORING_SETUP_SQE128)
3165 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3167 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3168 if (size == SIZE_MAX) {
3169 io_mem_free(ctx->rings);
3174 ctx->sq_sqes = io_mem_alloc(size);
3175 if (!ctx->sq_sqes) {
3176 io_mem_free(ctx->rings);
3184 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3188 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3192 ret = io_uring_add_tctx_node(ctx);
3197 fd_install(fd, file);
3202 * Allocate an anonymous fd, this is what constitutes the application
3203 * visible backing of an io_uring instance. The application mmaps this
3204 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3205 * we have to tie this fd to a socket for file garbage collection purposes.
3207 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3210 #if defined(CONFIG_UNIX)
3213 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3216 return ERR_PTR(ret);
3219 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3220 O_RDWR | O_CLOEXEC, NULL);
3221 #if defined(CONFIG_UNIX)
3223 sock_release(ctx->ring_sock);
3224 ctx->ring_sock = NULL;
3226 ctx->ring_sock->file = file;
3232 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3233 struct io_uring_params __user *params)
3235 struct io_ring_ctx *ctx;
3241 if (entries > IORING_MAX_ENTRIES) {
3242 if (!(p->flags & IORING_SETUP_CLAMP))
3244 entries = IORING_MAX_ENTRIES;
3248 * Use twice as many entries for the CQ ring. It's possible for the
3249 * application to drive a higher depth than the size of the SQ ring,
3250 * since the sqes are only used at submission time. This allows for
3251 * some flexibility in overcommitting a bit. If the application has
3252 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3253 * of CQ ring entries manually.
3255 p->sq_entries = roundup_pow_of_two(entries);
3256 if (p->flags & IORING_SETUP_CQSIZE) {
3258 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3259 * to a power-of-two, if it isn't already. We do NOT impose
3260 * any cq vs sq ring sizing.
3264 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3265 if (!(p->flags & IORING_SETUP_CLAMP))
3267 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3269 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3270 if (p->cq_entries < p->sq_entries)
3273 p->cq_entries = 2 * p->sq_entries;
3276 ctx = io_ring_ctx_alloc(p);
3281 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3282 * space applications don't need to do io completion events
3283 * polling again, they can rely on io_sq_thread to do polling
3284 * work, which can reduce cpu usage and uring_lock contention.
3286 if (ctx->flags & IORING_SETUP_IOPOLL &&
3287 !(ctx->flags & IORING_SETUP_SQPOLL))
3288 ctx->syscall_iopoll = 1;
3290 ctx->compat = in_compat_syscall();
3291 if (!capable(CAP_IPC_LOCK))
3292 ctx->user = get_uid(current_user());
3295 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3296 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3299 if (ctx->flags & IORING_SETUP_SQPOLL) {
3300 /* IPI related flags don't make sense with SQPOLL */
3301 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3302 IORING_SETUP_TASKRUN_FLAG))
3304 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3305 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3306 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3308 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
3310 ctx->notify_method = TWA_SIGNAL;
3314 * This is just grabbed for accounting purposes. When a process exits,
3315 * the mm is exited and dropped before the files, hence we need to hang
3316 * on to this mm purely for the purposes of being able to unaccount
3317 * memory (locked/pinned vm). It's not used for anything else.
3319 mmgrab(current->mm);
3320 ctx->mm_account = current->mm;
3322 ret = io_allocate_scq_urings(ctx, p);
3326 ret = io_sq_offload_create(ctx, p);
3329 /* always set a rsrc node */
3330 ret = io_rsrc_node_switch_start(ctx);
3333 io_rsrc_node_switch(ctx, NULL);
3335 memset(&p->sq_off, 0, sizeof(p->sq_off));
3336 p->sq_off.head = offsetof(struct io_rings, sq.head);
3337 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3338 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3339 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3340 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3341 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3342 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3344 memset(&p->cq_off, 0, sizeof(p->cq_off));
3345 p->cq_off.head = offsetof(struct io_rings, cq.head);
3346 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3347 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3348 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3349 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3350 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3351 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3353 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3354 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3355 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3356 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3357 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3358 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3359 IORING_FEAT_LINKED_FILE;
3361 if (copy_to_user(params, p, sizeof(*p))) {
3366 file = io_uring_get_file(ctx);
3368 ret = PTR_ERR(file);
3373 * Install ring fd as the very last thing, so we don't risk someone
3374 * having closed it before we finish setup
3376 ret = io_uring_install_fd(ctx, file);
3378 /* fput will clean it up */
3383 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3386 io_ring_ctx_wait_and_kill(ctx);
3391 * Sets up an aio uring context, and returns the fd. Applications asks for a
3392 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3393 * params structure passed in.
3395 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3397 struct io_uring_params p;
3400 if (copy_from_user(&p, params, sizeof(p)))
3402 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3407 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3408 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3409 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3410 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3411 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3412 IORING_SETUP_SQE128 | IORING_SETUP_CQE32))
3415 return io_uring_create(entries, &p, params);
3418 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3419 struct io_uring_params __user *, params)
3421 return io_uring_setup(entries, params);
3424 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3427 struct io_uring_probe *p;
3431 size = struct_size(p, ops, nr_args);
3432 if (size == SIZE_MAX)
3434 p = kzalloc(size, GFP_KERNEL);
3439 if (copy_from_user(p, arg, size))
3442 if (memchr_inv(p, 0, size))
3445 p->last_op = IORING_OP_LAST - 1;
3446 if (nr_args > IORING_OP_LAST)
3447 nr_args = IORING_OP_LAST;
3449 for (i = 0; i < nr_args; i++) {
3451 if (!io_op_defs[i].not_supported)
3452 p->ops[i].flags = IO_URING_OP_SUPPORTED;
3457 if (copy_to_user(arg, p, size))
3464 static int io_register_personality(struct io_ring_ctx *ctx)
3466 const struct cred *creds;
3470 creds = get_current_cred();
3472 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
3473 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
3481 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
3482 void __user *arg, unsigned int nr_args)
3484 struct io_uring_restriction *res;
3488 /* Restrictions allowed only if rings started disabled */
3489 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3492 /* We allow only a single restrictions registration */
3493 if (ctx->restrictions.registered)
3496 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
3499 size = array_size(nr_args, sizeof(*res));
3500 if (size == SIZE_MAX)
3503 res = memdup_user(arg, size);
3505 return PTR_ERR(res);
3509 for (i = 0; i < nr_args; i++) {
3510 switch (res[i].opcode) {
3511 case IORING_RESTRICTION_REGISTER_OP:
3512 if (res[i].register_op >= IORING_REGISTER_LAST) {
3517 __set_bit(res[i].register_op,
3518 ctx->restrictions.register_op);
3520 case IORING_RESTRICTION_SQE_OP:
3521 if (res[i].sqe_op >= IORING_OP_LAST) {
3526 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
3528 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
3529 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
3531 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
3532 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
3541 /* Reset all restrictions if an error happened */
3543 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
3545 ctx->restrictions.registered = true;
3551 static int io_register_enable_rings(struct io_ring_ctx *ctx)
3553 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3556 if (ctx->restrictions.registered)
3557 ctx->restricted = 1;
3559 ctx->flags &= ~IORING_SETUP_R_DISABLED;
3560 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
3561 wake_up(&ctx->sq_data->wait);
3565 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
3566 void __user *arg, unsigned len)
3568 struct io_uring_task *tctx = current->io_uring;
3569 cpumask_var_t new_mask;
3572 if (!tctx || !tctx->io_wq)
3575 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3578 cpumask_clear(new_mask);
3579 if (len > cpumask_size())
3580 len = cpumask_size();
3582 if (in_compat_syscall()) {
3583 ret = compat_get_bitmap(cpumask_bits(new_mask),
3584 (const compat_ulong_t __user *)arg,
3585 len * 8 /* CHAR_BIT */);
3587 ret = copy_from_user(new_mask, arg, len);
3591 free_cpumask_var(new_mask);
3595 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
3596 free_cpumask_var(new_mask);
3600 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
3602 struct io_uring_task *tctx = current->io_uring;
3604 if (!tctx || !tctx->io_wq)
3607 return io_wq_cpu_affinity(tctx->io_wq, NULL);
3610 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
3612 __must_hold(&ctx->uring_lock)
3614 struct io_tctx_node *node;
3615 struct io_uring_task *tctx = NULL;
3616 struct io_sq_data *sqd = NULL;
3620 if (copy_from_user(new_count, arg, sizeof(new_count)))
3622 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3623 if (new_count[i] > INT_MAX)
3626 if (ctx->flags & IORING_SETUP_SQPOLL) {
3630 * Observe the correct sqd->lock -> ctx->uring_lock
3631 * ordering. Fine to drop uring_lock here, we hold
3634 refcount_inc(&sqd->refs);
3635 mutex_unlock(&ctx->uring_lock);
3636 mutex_lock(&sqd->lock);
3637 mutex_lock(&ctx->uring_lock);
3639 tctx = sqd->thread->io_uring;
3642 tctx = current->io_uring;
3645 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
3647 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3649 ctx->iowq_limits[i] = new_count[i];
3650 ctx->iowq_limits_set = true;
3652 if (tctx && tctx->io_wq) {
3653 ret = io_wq_max_workers(tctx->io_wq, new_count);
3657 memset(new_count, 0, sizeof(new_count));
3661 mutex_unlock(&sqd->lock);
3662 io_put_sq_data(sqd);
3665 if (copy_to_user(arg, new_count, sizeof(new_count)))
3668 /* that's it for SQPOLL, only the SQPOLL task creates requests */
3672 /* now propagate the restriction to all registered users */
3673 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3674 struct io_uring_task *tctx = node->task->io_uring;
3676 if (WARN_ON_ONCE(!tctx->io_wq))
3679 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3680 new_count[i] = ctx->iowq_limits[i];
3681 /* ignore errors, it always returns zero anyway */
3682 (void)io_wq_max_workers(tctx->io_wq, new_count);
3687 mutex_unlock(&sqd->lock);
3688 io_put_sq_data(sqd);
3693 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
3694 void __user *arg, unsigned nr_args)
3695 __releases(ctx->uring_lock)
3696 __acquires(ctx->uring_lock)
3701 * We're inside the ring mutex, if the ref is already dying, then
3702 * someone else killed the ctx or is already going through
3703 * io_uring_register().
3705 if (percpu_ref_is_dying(&ctx->refs))
3708 if (ctx->restricted) {
3709 if (opcode >= IORING_REGISTER_LAST)
3711 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
3712 if (!test_bit(opcode, ctx->restrictions.register_op))
3717 case IORING_REGISTER_BUFFERS:
3721 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
3723 case IORING_UNREGISTER_BUFFERS:
3727 ret = io_sqe_buffers_unregister(ctx);
3729 case IORING_REGISTER_FILES:
3733 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
3735 case IORING_UNREGISTER_FILES:
3739 ret = io_sqe_files_unregister(ctx);
3741 case IORING_REGISTER_FILES_UPDATE:
3742 ret = io_register_files_update(ctx, arg, nr_args);
3744 case IORING_REGISTER_EVENTFD:
3748 ret = io_eventfd_register(ctx, arg, 0);
3750 case IORING_REGISTER_EVENTFD_ASYNC:
3754 ret = io_eventfd_register(ctx, arg, 1);
3756 case IORING_UNREGISTER_EVENTFD:
3760 ret = io_eventfd_unregister(ctx);
3762 case IORING_REGISTER_PROBE:
3764 if (!arg || nr_args > 256)
3766 ret = io_probe(ctx, arg, nr_args);
3768 case IORING_REGISTER_PERSONALITY:
3772 ret = io_register_personality(ctx);
3774 case IORING_UNREGISTER_PERSONALITY:
3778 ret = io_unregister_personality(ctx, nr_args);
3780 case IORING_REGISTER_ENABLE_RINGS:
3784 ret = io_register_enable_rings(ctx);
3786 case IORING_REGISTER_RESTRICTIONS:
3787 ret = io_register_restrictions(ctx, arg, nr_args);
3789 case IORING_REGISTER_FILES2:
3790 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
3792 case IORING_REGISTER_FILES_UPDATE2:
3793 ret = io_register_rsrc_update(ctx, arg, nr_args,
3796 case IORING_REGISTER_BUFFERS2:
3797 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
3799 case IORING_REGISTER_BUFFERS_UPDATE:
3800 ret = io_register_rsrc_update(ctx, arg, nr_args,
3801 IORING_RSRC_BUFFER);
3803 case IORING_REGISTER_IOWQ_AFF:
3805 if (!arg || !nr_args)
3807 ret = io_register_iowq_aff(ctx, arg, nr_args);
3809 case IORING_UNREGISTER_IOWQ_AFF:
3813 ret = io_unregister_iowq_aff(ctx);
3815 case IORING_REGISTER_IOWQ_MAX_WORKERS:
3817 if (!arg || nr_args != 2)
3819 ret = io_register_iowq_max_workers(ctx, arg);
3821 case IORING_REGISTER_RING_FDS:
3822 ret = io_ringfd_register(ctx, arg, nr_args);
3824 case IORING_UNREGISTER_RING_FDS:
3825 ret = io_ringfd_unregister(ctx, arg, nr_args);
3827 case IORING_REGISTER_PBUF_RING:
3829 if (!arg || nr_args != 1)
3831 ret = io_register_pbuf_ring(ctx, arg);
3833 case IORING_UNREGISTER_PBUF_RING:
3835 if (!arg || nr_args != 1)
3837 ret = io_unregister_pbuf_ring(ctx, arg);
3847 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
3848 void __user *, arg, unsigned int, nr_args)
3850 struct io_ring_ctx *ctx;
3859 if (!io_is_uring_fops(f.file))
3862 ctx = f.file->private_data;
3866 mutex_lock(&ctx->uring_lock);
3867 ret = __io_uring_register(ctx, opcode, arg, nr_args);
3868 mutex_unlock(&ctx->uring_lock);
3869 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
3875 static int __init io_uring_init(void)
3877 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
3878 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
3879 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
3882 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
3883 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
3884 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
3885 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
3886 BUILD_BUG_SQE_ELEM(1, __u8, flags);
3887 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
3888 BUILD_BUG_SQE_ELEM(4, __s32, fd);
3889 BUILD_BUG_SQE_ELEM(8, __u64, off);
3890 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
3891 BUILD_BUG_SQE_ELEM(16, __u64, addr);
3892 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
3893 BUILD_BUG_SQE_ELEM(24, __u32, len);
3894 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
3895 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
3896 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
3897 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
3898 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
3899 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
3900 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
3901 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
3902 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
3903 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
3904 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
3905 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
3906 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
3907 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
3908 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
3909 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
3910 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
3911 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
3912 BUILD_BUG_SQE_ELEM(42, __u16, personality);
3913 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
3914 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
3915 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
3917 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
3918 sizeof(struct io_uring_rsrc_update));
3919 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
3920 sizeof(struct io_uring_rsrc_update2));
3922 /* ->buf_index is u16 */
3923 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
3924 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
3925 offsetof(struct io_uring_buf_ring, tail));
3927 /* should fit into one byte */
3928 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
3929 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
3930 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
3932 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
3934 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
3936 io_uring_optable_init();
3938 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
3942 __initcall(io_uring_init);