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"
96 #define IORING_MAX_ENTRIES 32768
97 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
99 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
100 IORING_REGISTER_LAST + IORING_OP_LAST)
102 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
103 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
105 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
106 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
108 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
109 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
112 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
115 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
117 #define IO_COMPL_BATCH 32
118 #define IO_REQ_ALLOC_BATCH 8
121 IO_CHECK_CQ_OVERFLOW_BIT,
122 IO_CHECK_CQ_DROPPED_BIT,
125 struct io_defer_entry {
126 struct list_head list;
127 struct io_kiocb *req;
131 /* requests with any of those set should undergo io_disarm_next() */
132 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
133 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
135 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
136 struct task_struct *task,
139 static void io_dismantle_req(struct io_kiocb *req);
140 static void io_clean_op(struct io_kiocb *req);
141 static void io_queue_sqe(struct io_kiocb *req);
143 static void __io_submit_flush_completions(struct io_ring_ctx *ctx);
145 static void io_eventfd_signal(struct io_ring_ctx *ctx);
147 static struct kmem_cache *req_cachep;
149 struct sock *io_uring_get_socket(struct file *file)
151 #if defined(CONFIG_UNIX)
152 if (io_is_uring_fops(file)) {
153 struct io_ring_ctx *ctx = file->private_data;
155 return ctx->ring_sock->sk;
160 EXPORT_SYMBOL(io_uring_get_socket);
162 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
164 if (!wq_list_empty(&ctx->submit_state.compl_reqs))
165 __io_submit_flush_completions(ctx);
168 static bool io_match_linked(struct io_kiocb *head)
170 struct io_kiocb *req;
172 io_for_each_link(req, head) {
173 if (req->flags & REQ_F_INFLIGHT)
180 * As io_match_task() but protected against racing with linked timeouts.
181 * User must not hold timeout_lock.
183 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
188 if (task && head->task != task)
193 if (head->flags & REQ_F_LINK_TIMEOUT) {
194 struct io_ring_ctx *ctx = head->ctx;
196 /* protect against races with linked timeouts */
197 spin_lock_irq(&ctx->timeout_lock);
198 matched = io_match_linked(head);
199 spin_unlock_irq(&ctx->timeout_lock);
201 matched = io_match_linked(head);
206 static inline void req_fail_link_node(struct io_kiocb *req, int res)
209 io_req_set_res(req, res, 0);
212 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
214 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
217 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
219 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
221 complete(&ctx->ref_comp);
224 static __cold void io_fallback_req_func(struct work_struct *work)
226 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
228 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
229 struct io_kiocb *req, *tmp;
232 percpu_ref_get(&ctx->refs);
233 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
234 req->io_task_work.func(req, &locked);
237 io_submit_flush_completions(ctx);
238 mutex_unlock(&ctx->uring_lock);
240 percpu_ref_put(&ctx->refs);
243 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
245 struct io_ring_ctx *ctx;
248 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
252 xa_init(&ctx->io_bl_xa);
255 * Use 5 bits less than the max cq entries, that should give us around
256 * 32 entries per hash list if totally full and uniformly spread.
258 hash_bits = ilog2(p->cq_entries);
262 ctx->cancel_hash_bits = hash_bits;
263 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
265 if (!ctx->cancel_hash)
267 __hash_init(ctx->cancel_hash, 1U << hash_bits);
269 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
270 if (!ctx->dummy_ubuf)
272 /* set invalid range, so io_import_fixed() fails meeting it */
273 ctx->dummy_ubuf->ubuf = -1UL;
275 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
276 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
279 ctx->flags = p->flags;
280 init_waitqueue_head(&ctx->sqo_sq_wait);
281 INIT_LIST_HEAD(&ctx->sqd_list);
282 INIT_LIST_HEAD(&ctx->cq_overflow_list);
283 INIT_LIST_HEAD(&ctx->io_buffers_cache);
284 INIT_LIST_HEAD(&ctx->apoll_cache);
285 init_completion(&ctx->ref_comp);
286 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
287 mutex_init(&ctx->uring_lock);
288 init_waitqueue_head(&ctx->cq_wait);
289 spin_lock_init(&ctx->completion_lock);
290 spin_lock_init(&ctx->timeout_lock);
291 INIT_WQ_LIST(&ctx->iopoll_list);
292 INIT_LIST_HEAD(&ctx->io_buffers_pages);
293 INIT_LIST_HEAD(&ctx->io_buffers_comp);
294 INIT_LIST_HEAD(&ctx->defer_list);
295 INIT_LIST_HEAD(&ctx->timeout_list);
296 INIT_LIST_HEAD(&ctx->ltimeout_list);
297 spin_lock_init(&ctx->rsrc_ref_lock);
298 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
299 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
300 init_llist_head(&ctx->rsrc_put_llist);
301 INIT_LIST_HEAD(&ctx->tctx_list);
302 ctx->submit_state.free_list.next = NULL;
303 INIT_WQ_LIST(&ctx->locked_free_list);
304 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
305 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
308 kfree(ctx->dummy_ubuf);
309 kfree(ctx->cancel_hash);
311 xa_destroy(&ctx->io_bl_xa);
316 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
318 struct io_rings *r = ctx->rings;
320 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
324 static bool req_need_defer(struct io_kiocb *req, u32 seq)
326 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
327 struct io_ring_ctx *ctx = req->ctx;
329 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
335 static inline void io_req_track_inflight(struct io_kiocb *req)
337 if (!(req->flags & REQ_F_INFLIGHT)) {
338 req->flags |= REQ_F_INFLIGHT;
339 atomic_inc(&req->task->io_uring->inflight_tracked);
343 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
345 if (WARN_ON_ONCE(!req->link))
348 req->flags &= ~REQ_F_ARM_LTIMEOUT;
349 req->flags |= REQ_F_LINK_TIMEOUT;
351 /* linked timeouts should have two refs once prep'ed */
352 io_req_set_refcount(req);
353 __io_req_set_refcount(req->link, 2);
357 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
359 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
361 return __io_prep_linked_timeout(req);
364 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
366 io_queue_linked_timeout(__io_prep_linked_timeout(req));
369 static inline void io_arm_ltimeout(struct io_kiocb *req)
371 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
372 __io_arm_ltimeout(req);
375 static void io_prep_async_work(struct io_kiocb *req)
377 const struct io_op_def *def = &io_op_defs[req->opcode];
378 struct io_ring_ctx *ctx = req->ctx;
380 if (!(req->flags & REQ_F_CREDS)) {
381 req->flags |= REQ_F_CREDS;
382 req->creds = get_current_cred();
385 req->work.list.next = NULL;
387 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
388 if (req->flags & REQ_F_FORCE_ASYNC)
389 req->work.flags |= IO_WQ_WORK_CONCURRENT;
391 if (req->flags & REQ_F_ISREG) {
392 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
393 io_wq_hash_work(&req->work, file_inode(req->file));
394 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
395 if (def->unbound_nonreg_file)
396 req->work.flags |= IO_WQ_WORK_UNBOUND;
400 static void io_prep_async_link(struct io_kiocb *req)
402 struct io_kiocb *cur;
404 if (req->flags & REQ_F_LINK_TIMEOUT) {
405 struct io_ring_ctx *ctx = req->ctx;
407 spin_lock_irq(&ctx->timeout_lock);
408 io_for_each_link(cur, req)
409 io_prep_async_work(cur);
410 spin_unlock_irq(&ctx->timeout_lock);
412 io_for_each_link(cur, req)
413 io_prep_async_work(cur);
417 void io_queue_iowq(struct io_kiocb *req, bool *dont_use)
419 struct io_kiocb *link = io_prep_linked_timeout(req);
420 struct io_uring_task *tctx = req->task->io_uring;
423 BUG_ON(!tctx->io_wq);
425 /* init ->work of the whole link before punting */
426 io_prep_async_link(req);
429 * Not expected to happen, but if we do have a bug where this _can_
430 * happen, catch it here and ensure the request is marked as
431 * canceled. That will make io-wq go through the usual work cancel
432 * procedure rather than attempt to run this request (or create a new
435 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
436 req->work.flags |= IO_WQ_WORK_CANCEL;
438 trace_io_uring_queue_async_work(req->ctx, req, req->cqe.user_data,
439 req->opcode, req->flags, &req->work,
440 io_wq_is_hashed(&req->work));
441 io_wq_enqueue(tctx->io_wq, &req->work);
443 io_queue_linked_timeout(link);
446 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
448 while (!list_empty(&ctx->defer_list)) {
449 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
450 struct io_defer_entry, list);
452 if (req_need_defer(de->req, de->seq))
454 list_del_init(&de->list);
455 io_req_task_queue(de->req);
460 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
462 if (ctx->off_timeout_used || ctx->drain_active) {
463 spin_lock(&ctx->completion_lock);
464 if (ctx->off_timeout_used)
465 io_flush_timeouts(ctx);
466 if (ctx->drain_active)
467 io_queue_deferred(ctx);
468 io_commit_cqring(ctx);
469 spin_unlock(&ctx->completion_lock);
472 io_eventfd_signal(ctx);
475 static void io_eventfd_signal(struct io_ring_ctx *ctx)
477 struct io_ev_fd *ev_fd;
481 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
484 ev_fd = rcu_dereference(ctx->io_ev_fd);
487 * Check again if ev_fd exists incase an io_eventfd_unregister call
488 * completed between the NULL check of ctx->io_ev_fd at the start of
489 * the function and rcu_read_lock.
491 if (unlikely(!ev_fd))
493 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
496 if (!ev_fd->eventfd_async || io_wq_current_is_worker())
497 eventfd_signal(ev_fd->cq_ev_fd, 1);
503 * This should only get called when at least one event has been posted.
504 * Some applications rely on the eventfd notification count only changing
505 * IFF a new CQE has been added to the CQ ring. There's no depedency on
506 * 1:1 relationship between how many times this function is called (and
507 * hence the eventfd count) and number of CQEs posted to the CQ ring.
509 void io_cqring_ev_posted(struct io_ring_ctx *ctx)
511 if (unlikely(ctx->off_timeout_used || ctx->drain_active ||
513 __io_commit_cqring_flush(ctx);
518 /* Returns true if there are no backlogged entries after the flush */
519 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
521 bool all_flushed, posted;
522 size_t cqe_size = sizeof(struct io_uring_cqe);
524 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
527 if (ctx->flags & IORING_SETUP_CQE32)
531 spin_lock(&ctx->completion_lock);
532 while (!list_empty(&ctx->cq_overflow_list)) {
533 struct io_uring_cqe *cqe = io_get_cqe(ctx);
534 struct io_overflow_cqe *ocqe;
538 ocqe = list_first_entry(&ctx->cq_overflow_list,
539 struct io_overflow_cqe, list);
541 memcpy(cqe, &ocqe->cqe, cqe_size);
543 io_account_cq_overflow(ctx);
546 list_del(&ocqe->list);
550 all_flushed = list_empty(&ctx->cq_overflow_list);
552 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
553 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
556 io_commit_cqring(ctx);
557 spin_unlock(&ctx->completion_lock);
559 io_cqring_ev_posted(ctx);
563 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
567 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
568 /* iopoll syncs against uring_lock, not completion_lock */
569 if (ctx->flags & IORING_SETUP_IOPOLL)
570 mutex_lock(&ctx->uring_lock);
571 ret = __io_cqring_overflow_flush(ctx, false);
572 if (ctx->flags & IORING_SETUP_IOPOLL)
573 mutex_unlock(&ctx->uring_lock);
579 static void __io_put_task(struct task_struct *task, int nr)
581 struct io_uring_task *tctx = task->io_uring;
583 percpu_counter_sub(&tctx->inflight, nr);
584 if (unlikely(atomic_read(&tctx->in_idle)))
585 wake_up(&tctx->wait);
586 put_task_struct_many(task, nr);
589 /* must to be called somewhat shortly after putting a request */
590 static inline void io_put_task(struct task_struct *task, int nr)
592 if (likely(task == current))
593 task->io_uring->cached_refs += nr;
595 __io_put_task(task, nr);
598 static void io_task_refs_refill(struct io_uring_task *tctx)
600 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
602 percpu_counter_add(&tctx->inflight, refill);
603 refcount_add(refill, ¤t->usage);
604 tctx->cached_refs += refill;
607 static inline void io_get_task_refs(int nr)
609 struct io_uring_task *tctx = current->io_uring;
611 tctx->cached_refs -= nr;
612 if (unlikely(tctx->cached_refs < 0))
613 io_task_refs_refill(tctx);
616 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
618 struct io_uring_task *tctx = task->io_uring;
619 unsigned int refs = tctx->cached_refs;
622 tctx->cached_refs = 0;
623 percpu_counter_sub(&tctx->inflight, refs);
624 put_task_struct_many(task, refs);
628 bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data, s32 res,
629 u32 cflags, u64 extra1, u64 extra2)
631 struct io_overflow_cqe *ocqe;
632 size_t ocq_size = sizeof(struct io_overflow_cqe);
633 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
636 ocq_size += sizeof(struct io_uring_cqe);
638 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
639 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
642 * If we're in ring overflow flush mode, or in task cancel mode,
643 * or cannot allocate an overflow entry, then we need to drop it
646 io_account_cq_overflow(ctx);
647 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
650 if (list_empty(&ctx->cq_overflow_list)) {
651 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
652 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
655 ocqe->cqe.user_data = user_data;
657 ocqe->cqe.flags = cflags;
659 ocqe->cqe.big_cqe[0] = extra1;
660 ocqe->cqe.big_cqe[1] = extra2;
662 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
666 bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
669 struct io_uring_cqe *cqe;
672 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
675 * If we can't get a cq entry, userspace overflowed the
676 * submission (by quite a lot). Increment the overflow count in
679 cqe = io_get_cqe(ctx);
681 WRITE_ONCE(cqe->user_data, user_data);
682 WRITE_ONCE(cqe->res, res);
683 WRITE_ONCE(cqe->flags, cflags);
685 if (ctx->flags & IORING_SETUP_CQE32) {
686 WRITE_ONCE(cqe->big_cqe[0], 0);
687 WRITE_ONCE(cqe->big_cqe[1], 0);
691 return io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
694 static void __io_req_complete_put(struct io_kiocb *req)
697 * If we're the last reference to this request, add to our locked
700 if (req_ref_put_and_test(req)) {
701 struct io_ring_ctx *ctx = req->ctx;
703 if (req->flags & IO_REQ_LINK_FLAGS) {
704 if (req->flags & IO_DISARM_MASK)
707 io_req_task_queue(req->link);
711 io_req_put_rsrc(req);
713 * Selected buffer deallocation in io_clean_op() assumes that
714 * we don't hold ->completion_lock. Clean them here to avoid
717 io_put_kbuf_comp(req);
718 io_dismantle_req(req);
719 io_put_task(req->task, 1);
720 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
721 ctx->locked_free_nr++;
725 void __io_req_complete_post(struct io_kiocb *req)
727 if (!(req->flags & REQ_F_CQE_SKIP))
728 __io_fill_cqe_req(req->ctx, req);
729 __io_req_complete_put(req);
732 void io_req_complete_post(struct io_kiocb *req)
734 struct io_ring_ctx *ctx = req->ctx;
736 spin_lock(&ctx->completion_lock);
737 __io_req_complete_post(req);
738 io_commit_cqring(ctx);
739 spin_unlock(&ctx->completion_lock);
740 io_cqring_ev_posted(ctx);
743 inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags)
745 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
746 req->flags |= REQ_F_COMPLETE_INLINE;
748 io_req_complete_post(req);
751 void io_req_complete_failed(struct io_kiocb *req, s32 res)
754 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
755 io_req_complete_post(req);
759 * Don't initialise the fields below on every allocation, but do that in
760 * advance and keep them valid across allocations.
762 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
766 req->async_data = NULL;
767 /* not necessary, but safer to zero */
771 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
772 struct io_submit_state *state)
774 spin_lock(&ctx->completion_lock);
775 wq_list_splice(&ctx->locked_free_list, &state->free_list);
776 ctx->locked_free_nr = 0;
777 spin_unlock(&ctx->completion_lock);
780 static inline bool io_req_cache_empty(struct io_ring_ctx *ctx)
782 return !ctx->submit_state.free_list.next;
786 * A request might get retired back into the request caches even before opcode
787 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
788 * Because of that, io_alloc_req() should be called only under ->uring_lock
789 * and with extra caution to not get a request that is still worked on.
791 static __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
792 __must_hold(&ctx->uring_lock)
794 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
795 void *reqs[IO_REQ_ALLOC_BATCH];
799 * If we have more than a batch's worth of requests in our IRQ side
800 * locked cache, grab the lock and move them over to our submission
803 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
804 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
805 if (!io_req_cache_empty(ctx))
809 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
812 * Bulk alloc is all-or-nothing. If we fail to get a batch,
813 * retry single alloc to be on the safe side.
815 if (unlikely(ret <= 0)) {
816 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
822 percpu_ref_get_many(&ctx->refs, ret);
823 for (i = 0; i < ret; i++) {
824 struct io_kiocb *req = reqs[i];
826 io_preinit_req(req, ctx);
827 io_req_add_to_cache(req, ctx);
832 static inline bool io_alloc_req_refill(struct io_ring_ctx *ctx)
834 if (unlikely(io_req_cache_empty(ctx)))
835 return __io_alloc_req_refill(ctx);
839 static inline struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
841 struct io_wq_work_node *node;
843 node = wq_stack_extract(&ctx->submit_state.free_list);
844 return container_of(node, struct io_kiocb, comp_list);
847 static inline void io_dismantle_req(struct io_kiocb *req)
849 unsigned int flags = req->flags;
851 if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
853 if (!(flags & REQ_F_FIXED_FILE))
854 io_put_file(req->file);
857 __cold void io_free_req(struct io_kiocb *req)
859 struct io_ring_ctx *ctx = req->ctx;
861 io_req_put_rsrc(req);
862 io_dismantle_req(req);
863 io_put_task(req->task, 1);
865 spin_lock(&ctx->completion_lock);
866 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
867 ctx->locked_free_nr++;
868 spin_unlock(&ctx->completion_lock);
871 static void __io_req_find_next_prep(struct io_kiocb *req)
873 struct io_ring_ctx *ctx = req->ctx;
876 spin_lock(&ctx->completion_lock);
877 posted = io_disarm_next(req);
878 io_commit_cqring(ctx);
879 spin_unlock(&ctx->completion_lock);
881 io_cqring_ev_posted(ctx);
884 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
886 struct io_kiocb *nxt;
889 * If LINK is set, we have dependent requests in this chain. If we
890 * didn't fail this request, queue the first one up, moving any other
891 * dependencies to the next request. In case of failure, fail the rest
894 if (unlikely(req->flags & IO_DISARM_MASK))
895 __io_req_find_next_prep(req);
901 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
905 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
906 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
908 io_submit_flush_completions(ctx);
909 mutex_unlock(&ctx->uring_lock);
912 percpu_ref_put(&ctx->refs);
915 static inline void ctx_commit_and_unlock(struct io_ring_ctx *ctx)
917 io_commit_cqring(ctx);
918 spin_unlock(&ctx->completion_lock);
919 io_cqring_ev_posted(ctx);
922 static void handle_prev_tw_list(struct io_wq_work_node *node,
923 struct io_ring_ctx **ctx, bool *uring_locked)
925 if (*ctx && !*uring_locked)
926 spin_lock(&(*ctx)->completion_lock);
929 struct io_wq_work_node *next = node->next;
930 struct io_kiocb *req = container_of(node, struct io_kiocb,
933 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
935 if (req->ctx != *ctx) {
936 if (unlikely(!*uring_locked && *ctx))
937 ctx_commit_and_unlock(*ctx);
939 ctx_flush_and_put(*ctx, uring_locked);
941 /* if not contended, grab and improve batching */
942 *uring_locked = mutex_trylock(&(*ctx)->uring_lock);
943 percpu_ref_get(&(*ctx)->refs);
944 if (unlikely(!*uring_locked))
945 spin_lock(&(*ctx)->completion_lock);
947 if (likely(*uring_locked)) {
948 req->io_task_work.func(req, uring_locked);
950 req->cqe.flags = io_put_kbuf_comp(req);
951 __io_req_complete_post(req);
956 if (unlikely(!*uring_locked))
957 ctx_commit_and_unlock(*ctx);
960 static void handle_tw_list(struct io_wq_work_node *node,
961 struct io_ring_ctx **ctx, bool *locked)
964 struct io_wq_work_node *next = node->next;
965 struct io_kiocb *req = container_of(node, struct io_kiocb,
968 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
970 if (req->ctx != *ctx) {
971 ctx_flush_and_put(*ctx, locked);
973 /* if not contended, grab and improve batching */
974 *locked = mutex_trylock(&(*ctx)->uring_lock);
975 percpu_ref_get(&(*ctx)->refs);
977 req->io_task_work.func(req, locked);
982 void tctx_task_work(struct callback_head *cb)
984 bool uring_locked = false;
985 struct io_ring_ctx *ctx = NULL;
986 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
990 struct io_wq_work_node *node1, *node2;
992 spin_lock_irq(&tctx->task_lock);
993 node1 = tctx->prio_task_list.first;
994 node2 = tctx->task_list.first;
995 INIT_WQ_LIST(&tctx->task_list);
996 INIT_WQ_LIST(&tctx->prio_task_list);
997 if (!node2 && !node1)
998 tctx->task_running = false;
999 spin_unlock_irq(&tctx->task_lock);
1000 if (!node2 && !node1)
1004 handle_prev_tw_list(node1, &ctx, &uring_locked);
1006 handle_tw_list(node2, &ctx, &uring_locked);
1009 if (data_race(!tctx->task_list.first) &&
1010 data_race(!tctx->prio_task_list.first) && uring_locked)
1011 io_submit_flush_completions(ctx);
1014 ctx_flush_and_put(ctx, &uring_locked);
1016 /* relaxed read is enough as only the task itself sets ->in_idle */
1017 if (unlikely(atomic_read(&tctx->in_idle)))
1018 io_uring_drop_tctx_refs(current);
1021 static void __io_req_task_work_add(struct io_kiocb *req,
1022 struct io_uring_task *tctx,
1023 struct io_wq_work_list *list)
1025 struct io_ring_ctx *ctx = req->ctx;
1026 struct io_wq_work_node *node;
1027 unsigned long flags;
1030 spin_lock_irqsave(&tctx->task_lock, flags);
1031 wq_list_add_tail(&req->io_task_work.node, list);
1032 running = tctx->task_running;
1034 tctx->task_running = true;
1035 spin_unlock_irqrestore(&tctx->task_lock, flags);
1037 /* task_work already pending, we're done */
1041 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1042 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1044 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1047 spin_lock_irqsave(&tctx->task_lock, flags);
1048 tctx->task_running = false;
1049 node = wq_list_merge(&tctx->prio_task_list, &tctx->task_list);
1050 spin_unlock_irqrestore(&tctx->task_lock, flags);
1053 req = container_of(node, struct io_kiocb, io_task_work.node);
1055 if (llist_add(&req->io_task_work.fallback_node,
1056 &req->ctx->fallback_llist))
1057 schedule_delayed_work(&req->ctx->fallback_work, 1);
1061 void io_req_task_work_add(struct io_kiocb *req)
1063 struct io_uring_task *tctx = req->task->io_uring;
1065 __io_req_task_work_add(req, tctx, &tctx->task_list);
1068 void io_req_task_prio_work_add(struct io_kiocb *req)
1070 struct io_uring_task *tctx = req->task->io_uring;
1072 if (req->ctx->flags & IORING_SETUP_SQPOLL)
1073 __io_req_task_work_add(req, tctx, &tctx->prio_task_list);
1075 __io_req_task_work_add(req, tctx, &tctx->task_list);
1078 static void io_req_tw_post(struct io_kiocb *req, bool *locked)
1080 io_req_complete_post(req);
1083 void io_req_tw_post_queue(struct io_kiocb *req, s32 res, u32 cflags)
1085 io_req_set_res(req, res, cflags);
1086 req->io_task_work.func = io_req_tw_post;
1087 io_req_task_work_add(req);
1090 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
1092 /* not needed for normal modes, but SQPOLL depends on it */
1093 io_tw_lock(req->ctx, locked);
1094 io_req_complete_failed(req, req->cqe.res);
1097 void io_req_task_submit(struct io_kiocb *req, bool *locked)
1099 io_tw_lock(req->ctx, locked);
1100 /* req->task == current here, checking PF_EXITING is safe */
1101 if (likely(!(req->task->flags & PF_EXITING)))
1104 io_req_complete_failed(req, -EFAULT);
1107 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1109 io_req_set_res(req, ret, 0);
1110 req->io_task_work.func = io_req_task_cancel;
1111 io_req_task_work_add(req);
1114 void io_req_task_queue(struct io_kiocb *req)
1116 req->io_task_work.func = io_req_task_submit;
1117 io_req_task_work_add(req);
1120 void io_queue_next(struct io_kiocb *req)
1122 struct io_kiocb *nxt = io_req_find_next(req);
1125 io_req_task_queue(nxt);
1128 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1129 __must_hold(&ctx->uring_lock)
1131 struct task_struct *task = NULL;
1135 struct io_kiocb *req = container_of(node, struct io_kiocb,
1138 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1139 if (req->flags & REQ_F_REFCOUNT) {
1140 node = req->comp_list.next;
1141 if (!req_ref_put_and_test(req))
1144 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1145 struct async_poll *apoll = req->apoll;
1147 if (apoll->double_poll)
1148 kfree(apoll->double_poll);
1149 list_add(&apoll->poll.wait.entry,
1151 req->flags &= ~REQ_F_POLLED;
1153 if (req->flags & IO_REQ_LINK_FLAGS)
1155 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1158 if (!(req->flags & REQ_F_FIXED_FILE))
1159 io_put_file(req->file);
1161 io_req_put_rsrc_locked(req, ctx);
1163 if (req->task != task) {
1165 io_put_task(task, task_refs);
1170 node = req->comp_list.next;
1171 io_req_add_to_cache(req, ctx);
1175 io_put_task(task, task_refs);
1178 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1179 __must_hold(&ctx->uring_lock)
1181 struct io_wq_work_node *node, *prev;
1182 struct io_submit_state *state = &ctx->submit_state;
1184 if (state->flush_cqes) {
1185 spin_lock(&ctx->completion_lock);
1186 wq_list_for_each(node, prev, &state->compl_reqs) {
1187 struct io_kiocb *req = container_of(node, struct io_kiocb,
1190 if (!(req->flags & REQ_F_CQE_SKIP))
1191 __io_fill_cqe_req(ctx, req);
1194 io_commit_cqring(ctx);
1195 spin_unlock(&ctx->completion_lock);
1196 io_cqring_ev_posted(ctx);
1197 state->flush_cqes = false;
1200 io_free_batch_list(ctx, state->compl_reqs.first);
1201 INIT_WQ_LIST(&state->compl_reqs);
1205 * Drop reference to request, return next in chain (if there is one) if this
1206 * was the last reference to this request.
1208 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1210 struct io_kiocb *nxt = NULL;
1212 if (req_ref_put_and_test(req)) {
1213 if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
1214 nxt = io_req_find_next(req);
1220 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1222 /* See comment at the top of this file */
1224 return __io_cqring_events(ctx);
1228 * We can't just wait for polled events to come to us, we have to actively
1229 * find and complete them.
1231 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1233 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1236 mutex_lock(&ctx->uring_lock);
1237 while (!wq_list_empty(&ctx->iopoll_list)) {
1238 /* let it sleep and repeat later if can't complete a request */
1239 if (io_do_iopoll(ctx, true) == 0)
1242 * Ensure we allow local-to-the-cpu processing to take place,
1243 * in this case we need to ensure that we reap all events.
1244 * Also let task_work, etc. to progress by releasing the mutex
1246 if (need_resched()) {
1247 mutex_unlock(&ctx->uring_lock);
1249 mutex_lock(&ctx->uring_lock);
1252 mutex_unlock(&ctx->uring_lock);
1255 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1257 unsigned int nr_events = 0;
1259 unsigned long check_cq;
1261 check_cq = READ_ONCE(ctx->check_cq);
1262 if (unlikely(check_cq)) {
1263 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1264 __io_cqring_overflow_flush(ctx, false);
1266 * Similarly do not spin if we have not informed the user of any
1269 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1273 * Don't enter poll loop if we already have events pending.
1274 * If we do, we can potentially be spinning for commands that
1275 * already triggered a CQE (eg in error).
1277 if (io_cqring_events(ctx))
1282 * If a submit got punted to a workqueue, we can have the
1283 * application entering polling for a command before it gets
1284 * issued. That app will hold the uring_lock for the duration
1285 * of the poll right here, so we need to take a breather every
1286 * now and then to ensure that the issue has a chance to add
1287 * the poll to the issued list. Otherwise we can spin here
1288 * forever, while the workqueue is stuck trying to acquire the
1291 if (wq_list_empty(&ctx->iopoll_list)) {
1292 u32 tail = ctx->cached_cq_tail;
1294 mutex_unlock(&ctx->uring_lock);
1296 mutex_lock(&ctx->uring_lock);
1298 /* some requests don't go through iopoll_list */
1299 if (tail != ctx->cached_cq_tail ||
1300 wq_list_empty(&ctx->iopoll_list))
1303 ret = io_do_iopoll(ctx, !min);
1308 } while (nr_events < min && !need_resched());
1312 inline void io_req_task_complete(struct io_kiocb *req, bool *locked)
1315 req->cqe.flags |= io_put_kbuf(req, 0);
1316 io_req_add_compl_list(req);
1318 req->cqe.flags |= io_put_kbuf(req, IO_URING_F_UNLOCKED);
1319 io_req_complete_post(req);
1324 * After the iocb has been issued, it's safe to be found on the poll list.
1325 * Adding the kiocb to the list AFTER submission ensures that we don't
1326 * find it from a io_do_iopoll() thread before the issuer is done
1327 * accessing the kiocb cookie.
1329 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1331 struct io_ring_ctx *ctx = req->ctx;
1332 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1334 /* workqueue context doesn't hold uring_lock, grab it now */
1335 if (unlikely(needs_lock))
1336 mutex_lock(&ctx->uring_lock);
1339 * Track whether we have multiple files in our lists. This will impact
1340 * how we do polling eventually, not spinning if we're on potentially
1341 * different devices.
1343 if (wq_list_empty(&ctx->iopoll_list)) {
1344 ctx->poll_multi_queue = false;
1345 } else if (!ctx->poll_multi_queue) {
1346 struct io_kiocb *list_req;
1348 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1350 if (list_req->file != req->file)
1351 ctx->poll_multi_queue = true;
1355 * For fast devices, IO may have already completed. If it has, add
1356 * it to the front so we find it first.
1358 if (READ_ONCE(req->iopoll_completed))
1359 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1361 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1363 if (unlikely(needs_lock)) {
1365 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1366 * in sq thread task context or in io worker task context. If
1367 * current task context is sq thread, we don't need to check
1368 * whether should wake up sq thread.
1370 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1371 wq_has_sleeper(&ctx->sq_data->wait))
1372 wake_up(&ctx->sq_data->wait);
1374 mutex_unlock(&ctx->uring_lock);
1378 static bool io_bdev_nowait(struct block_device *bdev)
1380 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
1384 * If we tracked the file through the SCM inflight mechanism, we could support
1385 * any file. For now, just ensure that anything potentially problematic is done
1388 static bool __io_file_supports_nowait(struct file *file, umode_t mode)
1390 if (S_ISBLK(mode)) {
1391 if (IS_ENABLED(CONFIG_BLOCK) &&
1392 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
1398 if (S_ISREG(mode)) {
1399 if (IS_ENABLED(CONFIG_BLOCK) &&
1400 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
1401 !io_is_uring_fops(file))
1406 /* any ->read/write should understand O_NONBLOCK */
1407 if (file->f_flags & O_NONBLOCK)
1409 return file->f_mode & FMODE_NOWAIT;
1413 * If we tracked the file through the SCM inflight mechanism, we could support
1414 * any file. For now, just ensure that anything potentially problematic is done
1417 unsigned int io_file_get_flags(struct file *file)
1419 umode_t mode = file_inode(file)->i_mode;
1420 unsigned int res = 0;
1424 if (__io_file_supports_nowait(file, mode))
1426 if (io_file_need_scm(file))
1431 bool io_alloc_async_data(struct io_kiocb *req)
1433 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
1434 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
1435 if (req->async_data) {
1436 req->flags |= REQ_F_ASYNC_DATA;
1442 int io_req_prep_async(struct io_kiocb *req)
1444 const struct io_op_def *def = &io_op_defs[req->opcode];
1446 /* assign early for deferred execution for non-fixed file */
1447 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE))
1448 req->file = io_file_get_normal(req, req->cqe.fd);
1449 if (!def->prep_async)
1451 if (WARN_ON_ONCE(req_has_async_data(req)))
1453 if (io_alloc_async_data(req))
1456 return def->prep_async(req);
1459 static u32 io_get_sequence(struct io_kiocb *req)
1461 u32 seq = req->ctx->cached_sq_head;
1462 struct io_kiocb *cur;
1464 /* need original cached_sq_head, but it was increased for each req */
1465 io_for_each_link(cur, req)
1470 static __cold void io_drain_req(struct io_kiocb *req)
1472 struct io_ring_ctx *ctx = req->ctx;
1473 struct io_defer_entry *de;
1475 u32 seq = io_get_sequence(req);
1477 /* Still need defer if there is pending req in defer list. */
1478 spin_lock(&ctx->completion_lock);
1479 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1480 spin_unlock(&ctx->completion_lock);
1482 ctx->drain_active = false;
1483 io_req_task_queue(req);
1486 spin_unlock(&ctx->completion_lock);
1488 ret = io_req_prep_async(req);
1491 io_req_complete_failed(req, ret);
1494 io_prep_async_link(req);
1495 de = kmalloc(sizeof(*de), GFP_KERNEL);
1501 spin_lock(&ctx->completion_lock);
1502 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1503 spin_unlock(&ctx->completion_lock);
1508 trace_io_uring_defer(ctx, req, req->cqe.user_data, req->opcode);
1511 list_add_tail(&de->list, &ctx->defer_list);
1512 spin_unlock(&ctx->completion_lock);
1515 static void io_clean_op(struct io_kiocb *req)
1517 if (req->flags & REQ_F_BUFFER_SELECTED) {
1518 spin_lock(&req->ctx->completion_lock);
1519 io_put_kbuf_comp(req);
1520 spin_unlock(&req->ctx->completion_lock);
1523 if (req->flags & REQ_F_NEED_CLEANUP) {
1524 const struct io_op_def *def = &io_op_defs[req->opcode];
1529 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1530 kfree(req->apoll->double_poll);
1534 if (req->flags & REQ_F_INFLIGHT) {
1535 struct io_uring_task *tctx = req->task->io_uring;
1537 atomic_dec(&tctx->inflight_tracked);
1539 if (req->flags & REQ_F_CREDS)
1540 put_cred(req->creds);
1541 if (req->flags & REQ_F_ASYNC_DATA) {
1542 kfree(req->async_data);
1543 req->async_data = NULL;
1545 req->flags &= ~IO_REQ_CLEAN_FLAGS;
1548 static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags)
1550 if (req->file || !io_op_defs[req->opcode].needs_file)
1553 if (req->flags & REQ_F_FIXED_FILE)
1554 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1556 req->file = io_file_get_normal(req, req->cqe.fd);
1561 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1563 const struct io_op_def *def = &io_op_defs[req->opcode];
1564 const struct cred *creds = NULL;
1567 if (unlikely(!io_assign_file(req, issue_flags)))
1570 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1571 creds = override_creds(req->creds);
1573 if (!def->audit_skip)
1574 audit_uring_entry(req->opcode);
1576 ret = def->issue(req, issue_flags);
1578 if (!def->audit_skip)
1579 audit_uring_exit(!ret, ret);
1582 revert_creds(creds);
1585 __io_req_complete(req, issue_flags);
1586 else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1589 /* If the op doesn't have a file, we're not polling for it */
1590 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && req->file)
1591 io_iopoll_req_issued(req, issue_flags);
1596 int io_poll_issue(struct io_kiocb *req, bool *locked)
1598 io_tw_lock(req->ctx, locked);
1599 if (unlikely(req->task->flags & PF_EXITING))
1601 return io_issue_sqe(req, IO_URING_F_NONBLOCK);
1604 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1606 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1608 req = io_put_req_find_next(req);
1609 return req ? &req->work : NULL;
1612 void io_wq_submit_work(struct io_wq_work *work)
1614 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1615 const struct io_op_def *def = &io_op_defs[req->opcode];
1616 unsigned int issue_flags = IO_URING_F_UNLOCKED;
1617 bool needs_poll = false;
1618 int ret = 0, err = -ECANCELED;
1620 /* one will be dropped by ->io_free_work() after returning to io-wq */
1621 if (!(req->flags & REQ_F_REFCOUNT))
1622 __io_req_set_refcount(req, 2);
1626 io_arm_ltimeout(req);
1628 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1629 if (work->flags & IO_WQ_WORK_CANCEL) {
1631 io_req_task_queue_fail(req, err);
1634 if (!io_assign_file(req, issue_flags)) {
1636 work->flags |= IO_WQ_WORK_CANCEL;
1640 if (req->flags & REQ_F_FORCE_ASYNC) {
1641 bool opcode_poll = def->pollin || def->pollout;
1643 if (opcode_poll && file_can_poll(req->file)) {
1645 issue_flags |= IO_URING_F_NONBLOCK;
1650 ret = io_issue_sqe(req, issue_flags);
1654 * We can get EAGAIN for iopolled IO even though we're
1655 * forcing a sync submission from here, since we can't
1656 * wait for request slots on the block side.
1659 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1665 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1667 /* aborted or ready, in either case retry blocking */
1669 issue_flags &= ~IO_URING_F_NONBLOCK;
1672 /* avoid locking problems by failing it from a clean context */
1674 io_req_task_queue_fail(req, ret);
1677 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1678 unsigned int issue_flags)
1680 struct io_ring_ctx *ctx = req->ctx;
1681 struct file *file = NULL;
1682 unsigned long file_ptr;
1684 io_ring_submit_lock(ctx, issue_flags);
1686 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
1688 fd = array_index_nospec(fd, ctx->nr_user_files);
1689 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
1690 file = (struct file *) (file_ptr & FFS_MASK);
1691 file_ptr &= ~FFS_MASK;
1692 /* mask in overlapping REQ_F and FFS bits */
1693 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
1694 io_req_set_rsrc_node(req, ctx, 0);
1695 WARN_ON_ONCE(file && !test_bit(fd, ctx->file_table.bitmap));
1697 io_ring_submit_unlock(ctx, issue_flags);
1701 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
1703 struct file *file = fget(fd);
1705 trace_io_uring_file_get(req->ctx, req, req->cqe.user_data, fd);
1707 /* we don't allow fixed io_uring files */
1708 if (file && io_is_uring_fops(file))
1709 io_req_track_inflight(req);
1713 static void io_queue_async(struct io_kiocb *req, int ret)
1714 __must_hold(&req->ctx->uring_lock)
1716 struct io_kiocb *linked_timeout;
1718 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
1719 io_req_complete_failed(req, ret);
1723 linked_timeout = io_prep_linked_timeout(req);
1725 switch (io_arm_poll_handler(req, 0)) {
1726 case IO_APOLL_READY:
1727 io_req_task_queue(req);
1729 case IO_APOLL_ABORTED:
1731 * Queued up for async execution, worker will release
1732 * submit reference when the iocb is actually submitted.
1734 io_kbuf_recycle(req, 0);
1735 io_queue_iowq(req, NULL);
1742 io_queue_linked_timeout(linked_timeout);
1745 static inline void io_queue_sqe(struct io_kiocb *req)
1746 __must_hold(&req->ctx->uring_lock)
1750 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
1752 if (req->flags & REQ_F_COMPLETE_INLINE) {
1753 io_req_add_compl_list(req);
1757 * We async punt it if the file wasn't marked NOWAIT, or if the file
1758 * doesn't support non-blocking read/write attempts
1761 io_arm_ltimeout(req);
1763 io_queue_async(req, ret);
1766 static void io_queue_sqe_fallback(struct io_kiocb *req)
1767 __must_hold(&req->ctx->uring_lock)
1769 if (unlikely(req->flags & REQ_F_FAIL)) {
1771 * We don't submit, fail them all, for that replace hardlinks
1772 * with normal links. Extra REQ_F_LINK is tolerated.
1774 req->flags &= ~REQ_F_HARDLINK;
1775 req->flags |= REQ_F_LINK;
1776 io_req_complete_failed(req, req->cqe.res);
1777 } else if (unlikely(req->ctx->drain_active)) {
1780 int ret = io_req_prep_async(req);
1783 io_req_complete_failed(req, ret);
1785 io_queue_iowq(req, NULL);
1790 * Check SQE restrictions (opcode and flags).
1792 * Returns 'true' if SQE is allowed, 'false' otherwise.
1794 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
1795 struct io_kiocb *req,
1796 unsigned int sqe_flags)
1798 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
1801 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
1802 ctx->restrictions.sqe_flags_required)
1805 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
1806 ctx->restrictions.sqe_flags_required))
1812 static void io_init_req_drain(struct io_kiocb *req)
1814 struct io_ring_ctx *ctx = req->ctx;
1815 struct io_kiocb *head = ctx->submit_state.link.head;
1817 ctx->drain_active = true;
1820 * If we need to drain a request in the middle of a link, drain
1821 * the head request and the next request/link after the current
1822 * link. Considering sequential execution of links,
1823 * REQ_F_IO_DRAIN will be maintained for every request of our
1826 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
1827 ctx->drain_next = true;
1831 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
1832 const struct io_uring_sqe *sqe)
1833 __must_hold(&ctx->uring_lock)
1835 const struct io_op_def *def;
1836 unsigned int sqe_flags;
1840 /* req is partially pre-initialised, see io_preinit_req() */
1841 req->opcode = opcode = READ_ONCE(sqe->opcode);
1842 /* same numerical values with corresponding REQ_F_*, safe to copy */
1843 req->flags = sqe_flags = READ_ONCE(sqe->flags);
1844 req->cqe.user_data = READ_ONCE(sqe->user_data);
1846 req->rsrc_node = NULL;
1847 req->task = current;
1849 if (unlikely(opcode >= IORING_OP_LAST)) {
1853 def = &io_op_defs[opcode];
1854 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
1855 /* enforce forwards compatibility on users */
1856 if (sqe_flags & ~SQE_VALID_FLAGS)
1858 if (sqe_flags & IOSQE_BUFFER_SELECT) {
1859 if (!def->buffer_select)
1861 req->buf_index = READ_ONCE(sqe->buf_group);
1863 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
1864 ctx->drain_disabled = true;
1865 if (sqe_flags & IOSQE_IO_DRAIN) {
1866 if (ctx->drain_disabled)
1868 io_init_req_drain(req);
1871 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
1872 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
1874 /* knock it to the slow queue path, will be drained there */
1875 if (ctx->drain_active)
1876 req->flags |= REQ_F_FORCE_ASYNC;
1877 /* if there is no link, we're at "next" request and need to drain */
1878 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
1879 ctx->drain_next = false;
1880 ctx->drain_active = true;
1881 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
1885 if (!def->ioprio && sqe->ioprio)
1887 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
1890 if (def->needs_file) {
1891 struct io_submit_state *state = &ctx->submit_state;
1893 req->cqe.fd = READ_ONCE(sqe->fd);
1896 * Plug now if we have more than 2 IO left after this, and the
1897 * target is potentially a read/write to block based storage.
1899 if (state->need_plug && def->plug) {
1900 state->plug_started = true;
1901 state->need_plug = false;
1902 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
1906 personality = READ_ONCE(sqe->personality);
1910 req->creds = xa_load(&ctx->personalities, personality);
1913 get_cred(req->creds);
1914 ret = security_uring_override_creds(req->creds);
1916 put_cred(req->creds);
1919 req->flags |= REQ_F_CREDS;
1922 return def->prep(req, sqe);
1925 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
1926 struct io_kiocb *req, int ret)
1928 struct io_ring_ctx *ctx = req->ctx;
1929 struct io_submit_link *link = &ctx->submit_state.link;
1930 struct io_kiocb *head = link->head;
1932 trace_io_uring_req_failed(sqe, ctx, req, ret);
1935 * Avoid breaking links in the middle as it renders links with SQPOLL
1936 * unusable. Instead of failing eagerly, continue assembling the link if
1937 * applicable and mark the head with REQ_F_FAIL. The link flushing code
1938 * should find the flag and handle the rest.
1940 req_fail_link_node(req, ret);
1941 if (head && !(head->flags & REQ_F_FAIL))
1942 req_fail_link_node(head, -ECANCELED);
1944 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
1946 link->last->link = req;
1950 io_queue_sqe_fallback(req);
1955 link->last->link = req;
1962 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
1963 const struct io_uring_sqe *sqe)
1964 __must_hold(&ctx->uring_lock)
1966 struct io_submit_link *link = &ctx->submit_state.link;
1969 ret = io_init_req(ctx, req, sqe);
1971 return io_submit_fail_init(sqe, req, ret);
1973 /* don't need @sqe from now on */
1974 trace_io_uring_submit_sqe(ctx, req, req->cqe.user_data, req->opcode,
1976 ctx->flags & IORING_SETUP_SQPOLL);
1979 * If we already have a head request, queue this one for async
1980 * submittal once the head completes. If we don't have a head but
1981 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
1982 * submitted sync once the chain is complete. If none of those
1983 * conditions are true (normal request), then just queue it.
1985 if (unlikely(link->head)) {
1986 ret = io_req_prep_async(req);
1988 return io_submit_fail_init(sqe, req, ret);
1990 trace_io_uring_link(ctx, req, link->head);
1991 link->last->link = req;
1994 if (req->flags & IO_REQ_LINK_FLAGS)
1996 /* last request of the link, flush it */
1999 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2002 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2003 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2004 if (req->flags & IO_REQ_LINK_FLAGS) {
2009 io_queue_sqe_fallback(req);
2019 * Batched submission is done, ensure local IO is flushed out.
2021 static void io_submit_state_end(struct io_ring_ctx *ctx)
2023 struct io_submit_state *state = &ctx->submit_state;
2025 if (unlikely(state->link.head))
2026 io_queue_sqe_fallback(state->link.head);
2027 /* flush only after queuing links as they can generate completions */
2028 io_submit_flush_completions(ctx);
2029 if (state->plug_started)
2030 blk_finish_plug(&state->plug);
2034 * Start submission side cache.
2036 static void io_submit_state_start(struct io_submit_state *state,
2037 unsigned int max_ios)
2039 state->plug_started = false;
2040 state->need_plug = max_ios > 2;
2041 state->submit_nr = max_ios;
2042 /* set only head, no need to init link_last in advance */
2043 state->link.head = NULL;
2046 static void io_commit_sqring(struct io_ring_ctx *ctx)
2048 struct io_rings *rings = ctx->rings;
2051 * Ensure any loads from the SQEs are done at this point,
2052 * since once we write the new head, the application could
2053 * write new data to them.
2055 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2059 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2060 * that is mapped by userspace. This means that care needs to be taken to
2061 * ensure that reads are stable, as we cannot rely on userspace always
2062 * being a good citizen. If members of the sqe are validated and then later
2063 * used, it's important that those reads are done through READ_ONCE() to
2064 * prevent a re-load down the line.
2066 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
2068 unsigned head, mask = ctx->sq_entries - 1;
2069 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2072 * The cached sq head (or cq tail) serves two purposes:
2074 * 1) allows us to batch the cost of updating the user visible
2076 * 2) allows the kernel side to track the head on its own, even
2077 * though the application is the one updating it.
2079 head = READ_ONCE(ctx->sq_array[sq_idx]);
2080 if (likely(head < ctx->sq_entries)) {
2081 /* double index for 128-byte SQEs, twice as long */
2082 if (ctx->flags & IORING_SETUP_SQE128)
2084 return &ctx->sq_sqes[head];
2087 /* drop invalid entries */
2089 WRITE_ONCE(ctx->rings->sq_dropped,
2090 READ_ONCE(ctx->rings->sq_dropped) + 1);
2094 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2095 __must_hold(&ctx->uring_lock)
2097 unsigned int entries = io_sqring_entries(ctx);
2101 if (unlikely(!entries))
2103 /* make sure SQ entry isn't read before tail */
2104 ret = left = min3(nr, ctx->sq_entries, entries);
2105 io_get_task_refs(left);
2106 io_submit_state_start(&ctx->submit_state, left);
2109 const struct io_uring_sqe *sqe;
2110 struct io_kiocb *req;
2112 if (unlikely(!io_alloc_req_refill(ctx)))
2114 req = io_alloc_req(ctx);
2115 sqe = io_get_sqe(ctx);
2116 if (unlikely(!sqe)) {
2117 io_req_add_to_cache(req, ctx);
2122 * Continue submitting even for sqe failure if the
2123 * ring was setup with IORING_SETUP_SUBMIT_ALL
2125 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2126 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2132 if (unlikely(left)) {
2134 /* try again if it submitted nothing and can't allocate a req */
2135 if (!ret && io_req_cache_empty(ctx))
2137 current->io_uring->cached_refs += left;
2140 io_submit_state_end(ctx);
2141 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2142 io_commit_sqring(ctx);
2146 struct io_wait_queue {
2147 struct wait_queue_entry wq;
2148 struct io_ring_ctx *ctx;
2150 unsigned nr_timeouts;
2153 static inline bool io_should_wake(struct io_wait_queue *iowq)
2155 struct io_ring_ctx *ctx = iowq->ctx;
2156 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
2159 * Wake up if we have enough events, or if a timeout occurred since we
2160 * started waiting. For timeouts, we always want to return to userspace,
2161 * regardless of event count.
2163 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2166 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2167 int wake_flags, void *key)
2169 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
2173 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2174 * the task, and the next invocation will do it.
2176 if (io_should_wake(iowq) ||
2177 test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &iowq->ctx->check_cq))
2178 return autoremove_wake_function(curr, mode, wake_flags, key);
2182 int io_run_task_work_sig(void)
2184 if (io_run_task_work())
2186 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
2187 return -ERESTARTSYS;
2188 if (task_sigpending(current))
2193 /* when returns >0, the caller should retry */
2194 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2195 struct io_wait_queue *iowq,
2199 unsigned long check_cq;
2201 /* make sure we run task_work before checking for signals */
2202 ret = io_run_task_work_sig();
2203 if (ret || io_should_wake(iowq))
2206 check_cq = READ_ONCE(ctx->check_cq);
2207 if (unlikely(check_cq)) {
2208 /* let the caller flush overflows, retry */
2209 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2211 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
2214 if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS))
2220 * Wait until events become available, if we don't already have some. The
2221 * application must reap them itself, as they reside on the shared cq ring.
2223 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2224 const sigset_t __user *sig, size_t sigsz,
2225 struct __kernel_timespec __user *uts)
2227 struct io_wait_queue iowq;
2228 struct io_rings *rings = ctx->rings;
2229 ktime_t timeout = KTIME_MAX;
2233 io_cqring_overflow_flush(ctx);
2234 if (io_cqring_events(ctx) >= min_events)
2236 if (!io_run_task_work())
2241 #ifdef CONFIG_COMPAT
2242 if (in_compat_syscall())
2243 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2247 ret = set_user_sigmask(sig, sigsz);
2254 struct timespec64 ts;
2256 if (get_timespec64(&ts, uts))
2258 timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2261 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2262 iowq.wq.private = current;
2263 INIT_LIST_HEAD(&iowq.wq.entry);
2265 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2266 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2268 trace_io_uring_cqring_wait(ctx, min_events);
2270 /* if we can't even flush overflow, don't wait for more */
2271 if (!io_cqring_overflow_flush(ctx)) {
2275 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2276 TASK_INTERRUPTIBLE);
2277 ret = io_cqring_wait_schedule(ctx, &iowq, timeout);
2281 finish_wait(&ctx->cq_wait, &iowq.wq);
2282 restore_saved_sigmask_unless(ret == -EINTR);
2284 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2287 static void io_mem_free(void *ptr)
2294 page = virt_to_head_page(ptr);
2295 if (put_page_testzero(page))
2296 free_compound_page(page);
2299 static void *io_mem_alloc(size_t size)
2301 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2303 return (void *) __get_free_pages(gfp, get_order(size));
2306 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2307 unsigned int cq_entries, size_t *sq_offset)
2309 struct io_rings *rings;
2310 size_t off, sq_array_size;
2312 off = struct_size(rings, cqes, cq_entries);
2313 if (off == SIZE_MAX)
2315 if (ctx->flags & IORING_SETUP_CQE32) {
2316 if (check_shl_overflow(off, 1, &off))
2321 off = ALIGN(off, SMP_CACHE_BYTES);
2329 sq_array_size = array_size(sizeof(u32), sq_entries);
2330 if (sq_array_size == SIZE_MAX)
2333 if (check_add_overflow(off, sq_array_size, &off))
2339 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2340 unsigned int eventfd_async)
2342 struct io_ev_fd *ev_fd;
2343 __s32 __user *fds = arg;
2346 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2347 lockdep_is_held(&ctx->uring_lock));
2351 if (copy_from_user(&fd, fds, sizeof(*fds)))
2354 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2358 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2359 if (IS_ERR(ev_fd->cq_ev_fd)) {
2360 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2364 ev_fd->eventfd_async = eventfd_async;
2365 ctx->has_evfd = true;
2366 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2370 static void io_eventfd_put(struct rcu_head *rcu)
2372 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
2374 eventfd_ctx_put(ev_fd->cq_ev_fd);
2378 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2380 struct io_ev_fd *ev_fd;
2382 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2383 lockdep_is_held(&ctx->uring_lock));
2385 ctx->has_evfd = false;
2386 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2387 call_rcu(&ev_fd->rcu, io_eventfd_put);
2394 static void io_req_caches_free(struct io_ring_ctx *ctx)
2396 struct io_submit_state *state = &ctx->submit_state;
2399 mutex_lock(&ctx->uring_lock);
2400 io_flush_cached_locked_reqs(ctx, state);
2402 while (!io_req_cache_empty(ctx)) {
2403 struct io_wq_work_node *node;
2404 struct io_kiocb *req;
2406 node = wq_stack_extract(&state->free_list);
2407 req = container_of(node, struct io_kiocb, comp_list);
2408 kmem_cache_free(req_cachep, req);
2412 percpu_ref_put_many(&ctx->refs, nr);
2413 mutex_unlock(&ctx->uring_lock);
2416 static void io_flush_apoll_cache(struct io_ring_ctx *ctx)
2418 struct async_poll *apoll;
2420 while (!list_empty(&ctx->apoll_cache)) {
2421 apoll = list_first_entry(&ctx->apoll_cache, struct async_poll,
2423 list_del(&apoll->poll.wait.entry);
2428 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2430 io_sq_thread_finish(ctx);
2432 if (ctx->mm_account) {
2433 mmdrop(ctx->mm_account);
2434 ctx->mm_account = NULL;
2437 io_rsrc_refs_drop(ctx);
2438 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2439 io_wait_rsrc_data(ctx->buf_data);
2440 io_wait_rsrc_data(ctx->file_data);
2442 mutex_lock(&ctx->uring_lock);
2444 __io_sqe_buffers_unregister(ctx);
2446 __io_sqe_files_unregister(ctx);
2448 __io_cqring_overflow_flush(ctx, true);
2449 io_eventfd_unregister(ctx);
2450 io_flush_apoll_cache(ctx);
2451 mutex_unlock(&ctx->uring_lock);
2452 io_destroy_buffers(ctx);
2454 put_cred(ctx->sq_creds);
2456 /* there are no registered resources left, nobody uses it */
2458 io_rsrc_node_destroy(ctx->rsrc_node);
2459 if (ctx->rsrc_backup_node)
2460 io_rsrc_node_destroy(ctx->rsrc_backup_node);
2461 flush_delayed_work(&ctx->rsrc_put_work);
2462 flush_delayed_work(&ctx->fallback_work);
2464 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2465 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
2467 #if defined(CONFIG_UNIX)
2468 if (ctx->ring_sock) {
2469 ctx->ring_sock->file = NULL; /* so that iput() is called */
2470 sock_release(ctx->ring_sock);
2473 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2475 io_mem_free(ctx->rings);
2476 io_mem_free(ctx->sq_sqes);
2478 percpu_ref_exit(&ctx->refs);
2479 free_uid(ctx->user);
2480 io_req_caches_free(ctx);
2482 io_wq_put_hash(ctx->hash_map);
2483 kfree(ctx->cancel_hash);
2484 kfree(ctx->dummy_ubuf);
2486 xa_destroy(&ctx->io_bl_xa);
2490 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2492 struct io_ring_ctx *ctx = file->private_data;
2495 poll_wait(file, &ctx->cq_wait, wait);
2497 * synchronizes with barrier from wq_has_sleeper call in
2501 if (!io_sqring_full(ctx))
2502 mask |= EPOLLOUT | EPOLLWRNORM;
2505 * Don't flush cqring overflow list here, just do a simple check.
2506 * Otherwise there could possible be ABBA deadlock:
2509 * lock(&ctx->uring_lock);
2511 * lock(&ctx->uring_lock);
2514 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2515 * pushs them to do the flush.
2517 if (io_cqring_events(ctx) ||
2518 test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
2519 mask |= EPOLLIN | EPOLLRDNORM;
2524 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2526 const struct cred *creds;
2528 creds = xa_erase(&ctx->personalities, id);
2537 struct io_tctx_exit {
2538 struct callback_head task_work;
2539 struct completion completion;
2540 struct io_ring_ctx *ctx;
2543 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2545 struct io_uring_task *tctx = current->io_uring;
2546 struct io_tctx_exit *work;
2548 work = container_of(cb, struct io_tctx_exit, task_work);
2550 * When @in_idle, we're in cancellation and it's racy to remove the
2551 * node. It'll be removed by the end of cancellation, just ignore it.
2553 if (!atomic_read(&tctx->in_idle))
2554 io_uring_del_tctx_node((unsigned long)work->ctx);
2555 complete(&work->completion);
2558 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2560 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2562 return req->ctx == data;
2565 static __cold void io_ring_exit_work(struct work_struct *work)
2567 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2568 unsigned long timeout = jiffies + HZ * 60 * 5;
2569 unsigned long interval = HZ / 20;
2570 struct io_tctx_exit exit;
2571 struct io_tctx_node *node;
2575 * If we're doing polled IO and end up having requests being
2576 * submitted async (out-of-line), then completions can come in while
2577 * we're waiting for refs to drop. We need to reap these manually,
2578 * as nobody else will be looking for them.
2581 io_uring_try_cancel_requests(ctx, NULL, true);
2583 struct io_sq_data *sqd = ctx->sq_data;
2584 struct task_struct *tsk;
2586 io_sq_thread_park(sqd);
2588 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
2589 io_wq_cancel_cb(tsk->io_uring->io_wq,
2590 io_cancel_ctx_cb, ctx, true);
2591 io_sq_thread_unpark(sqd);
2594 io_req_caches_free(ctx);
2596 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
2597 /* there is little hope left, don't run it too often */
2600 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
2602 init_completion(&exit.completion);
2603 init_task_work(&exit.task_work, io_tctx_exit_cb);
2606 * Some may use context even when all refs and requests have been put,
2607 * and they are free to do so while still holding uring_lock or
2608 * completion_lock, see io_req_task_submit(). Apart from other work,
2609 * this lock/unlock section also waits them to finish.
2611 mutex_lock(&ctx->uring_lock);
2612 while (!list_empty(&ctx->tctx_list)) {
2613 WARN_ON_ONCE(time_after(jiffies, timeout));
2615 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
2617 /* don't spin on a single task if cancellation failed */
2618 list_rotate_left(&ctx->tctx_list);
2619 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
2620 if (WARN_ON_ONCE(ret))
2623 mutex_unlock(&ctx->uring_lock);
2624 wait_for_completion(&exit.completion);
2625 mutex_lock(&ctx->uring_lock);
2627 mutex_unlock(&ctx->uring_lock);
2628 spin_lock(&ctx->completion_lock);
2629 spin_unlock(&ctx->completion_lock);
2631 io_ring_ctx_free(ctx);
2634 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
2636 unsigned long index;
2637 struct creds *creds;
2639 mutex_lock(&ctx->uring_lock);
2640 percpu_ref_kill(&ctx->refs);
2642 __io_cqring_overflow_flush(ctx, true);
2643 xa_for_each(&ctx->personalities, index, creds)
2644 io_unregister_personality(ctx, index);
2645 mutex_unlock(&ctx->uring_lock);
2647 /* failed during ring init, it couldn't have issued any requests */
2649 io_kill_timeouts(ctx, NULL, true);
2650 io_poll_remove_all(ctx, NULL, true);
2651 /* if we failed setting up the ctx, we might not have any rings */
2652 io_iopoll_try_reap_events(ctx);
2655 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
2657 * Use system_unbound_wq to avoid spawning tons of event kworkers
2658 * if we're exiting a ton of rings at the same time. It just adds
2659 * noise and overhead, there's no discernable change in runtime
2660 * over using system_wq.
2662 queue_work(system_unbound_wq, &ctx->exit_work);
2665 static int io_uring_release(struct inode *inode, struct file *file)
2667 struct io_ring_ctx *ctx = file->private_data;
2669 file->private_data = NULL;
2670 io_ring_ctx_wait_and_kill(ctx);
2674 struct io_task_cancel {
2675 struct task_struct *task;
2679 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
2681 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2682 struct io_task_cancel *cancel = data;
2684 return io_match_task_safe(req, cancel->task, cancel->all);
2687 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
2688 struct task_struct *task,
2691 struct io_defer_entry *de;
2694 spin_lock(&ctx->completion_lock);
2695 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
2696 if (io_match_task_safe(de->req, task, cancel_all)) {
2697 list_cut_position(&list, &ctx->defer_list, &de->list);
2701 spin_unlock(&ctx->completion_lock);
2702 if (list_empty(&list))
2705 while (!list_empty(&list)) {
2706 de = list_first_entry(&list, struct io_defer_entry, list);
2707 list_del_init(&de->list);
2708 io_req_complete_failed(de->req, -ECANCELED);
2714 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
2716 struct io_tctx_node *node;
2717 enum io_wq_cancel cret;
2720 mutex_lock(&ctx->uring_lock);
2721 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
2722 struct io_uring_task *tctx = node->task->io_uring;
2725 * io_wq will stay alive while we hold uring_lock, because it's
2726 * killed after ctx nodes, which requires to take the lock.
2728 if (!tctx || !tctx->io_wq)
2730 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
2731 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2733 mutex_unlock(&ctx->uring_lock);
2738 static __cold void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
2739 struct task_struct *task,
2742 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
2743 struct io_uring_task *tctx = task ? task->io_uring : NULL;
2745 /* failed during ring init, it couldn't have issued any requests */
2750 enum io_wq_cancel cret;
2754 ret |= io_uring_try_cancel_iowq(ctx);
2755 } else if (tctx && tctx->io_wq) {
2757 * Cancels requests of all rings, not only @ctx, but
2758 * it's fine as the task is in exit/exec.
2760 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
2762 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2765 /* SQPOLL thread does its own polling */
2766 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
2767 (ctx->sq_data && ctx->sq_data->thread == current)) {
2768 while (!wq_list_empty(&ctx->iopoll_list)) {
2769 io_iopoll_try_reap_events(ctx);
2774 ret |= io_cancel_defer_files(ctx, task, cancel_all);
2775 ret |= io_poll_remove_all(ctx, task, cancel_all);
2776 ret |= io_kill_timeouts(ctx, task, cancel_all);
2778 ret |= io_run_task_work();
2785 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
2788 return atomic_read(&tctx->inflight_tracked);
2789 return percpu_counter_sum(&tctx->inflight);
2793 * Find any io_uring ctx that this task has registered or done IO on, and cancel
2794 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
2796 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
2798 struct io_uring_task *tctx = current->io_uring;
2799 struct io_ring_ctx *ctx;
2803 WARN_ON_ONCE(sqd && sqd->thread != current);
2805 if (!current->io_uring)
2808 io_wq_exit_start(tctx->io_wq);
2810 atomic_inc(&tctx->in_idle);
2812 io_uring_drop_tctx_refs(current);
2813 /* read completions before cancelations */
2814 inflight = tctx_inflight(tctx, !cancel_all);
2819 struct io_tctx_node *node;
2820 unsigned long index;
2822 xa_for_each(&tctx->xa, index, node) {
2823 /* sqpoll task will cancel all its requests */
2824 if (node->ctx->sq_data)
2826 io_uring_try_cancel_requests(node->ctx, current,
2830 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
2831 io_uring_try_cancel_requests(ctx, current,
2835 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
2837 io_uring_drop_tctx_refs(current);
2840 * If we've seen completions, retry without waiting. This
2841 * avoids a race where a completion comes in before we did
2842 * prepare_to_wait().
2844 if (inflight == tctx_inflight(tctx, !cancel_all))
2846 finish_wait(&tctx->wait, &wait);
2849 io_uring_clean_tctx(tctx);
2852 * We shouldn't run task_works after cancel, so just leave
2853 * ->in_idle set for normal exit.
2855 atomic_dec(&tctx->in_idle);
2856 /* for exec all current's requests should be gone, kill tctx */
2857 __io_uring_free(current);
2861 void __io_uring_cancel(bool cancel_all)
2863 io_uring_cancel_generic(cancel_all, NULL);
2866 static void *io_uring_validate_mmap_request(struct file *file,
2867 loff_t pgoff, size_t sz)
2869 struct io_ring_ctx *ctx = file->private_data;
2870 loff_t offset = pgoff << PAGE_SHIFT;
2875 case IORING_OFF_SQ_RING:
2876 case IORING_OFF_CQ_RING:
2879 case IORING_OFF_SQES:
2883 return ERR_PTR(-EINVAL);
2886 page = virt_to_head_page(ptr);
2887 if (sz > page_size(page))
2888 return ERR_PTR(-EINVAL);
2895 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
2897 size_t sz = vma->vm_end - vma->vm_start;
2901 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
2903 return PTR_ERR(ptr);
2905 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
2906 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
2909 #else /* !CONFIG_MMU */
2911 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
2913 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
2916 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
2918 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
2921 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
2922 unsigned long addr, unsigned long len,
2923 unsigned long pgoff, unsigned long flags)
2927 ptr = io_uring_validate_mmap_request(file, pgoff, len);
2929 return PTR_ERR(ptr);
2931 return (unsigned long) ptr;
2934 #endif /* !CONFIG_MMU */
2936 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
2938 if (flags & IORING_ENTER_EXT_ARG) {
2939 struct io_uring_getevents_arg arg;
2941 if (argsz != sizeof(arg))
2943 if (copy_from_user(&arg, argp, sizeof(arg)))
2949 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
2950 struct __kernel_timespec __user **ts,
2951 const sigset_t __user **sig)
2953 struct io_uring_getevents_arg arg;
2956 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
2957 * is just a pointer to the sigset_t.
2959 if (!(flags & IORING_ENTER_EXT_ARG)) {
2960 *sig = (const sigset_t __user *) argp;
2966 * EXT_ARG is set - ensure we agree on the size of it and copy in our
2967 * timespec and sigset_t pointers if good.
2969 if (*argsz != sizeof(arg))
2971 if (copy_from_user(&arg, argp, sizeof(arg)))
2975 *sig = u64_to_user_ptr(arg.sigmask);
2976 *argsz = arg.sigmask_sz;
2977 *ts = u64_to_user_ptr(arg.ts);
2981 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
2982 u32, min_complete, u32, flags, const void __user *, argp,
2985 struct io_ring_ctx *ctx;
2991 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
2992 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
2993 IORING_ENTER_REGISTERED_RING)))
2997 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
2998 * need only dereference our task private array to find it.
3000 if (flags & IORING_ENTER_REGISTERED_RING) {
3001 struct io_uring_task *tctx = current->io_uring;
3003 if (!tctx || fd >= IO_RINGFD_REG_MAX)
3005 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3006 f.file = tctx->registered_rings[fd];
3012 if (unlikely(!f.file))
3016 if (unlikely(!io_is_uring_fops(f.file)))
3020 ctx = f.file->private_data;
3021 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
3025 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3029 * For SQ polling, the thread will do all submissions and completions.
3030 * Just return the requested submit count, and wake the thread if
3034 if (ctx->flags & IORING_SETUP_SQPOLL) {
3035 io_cqring_overflow_flush(ctx);
3037 if (unlikely(ctx->sq_data->thread == NULL)) {
3041 if (flags & IORING_ENTER_SQ_WAKEUP)
3042 wake_up(&ctx->sq_data->wait);
3043 if (flags & IORING_ENTER_SQ_WAIT) {
3044 ret = io_sqpoll_wait_sq(ctx);
3049 } else if (to_submit) {
3050 ret = io_uring_add_tctx_node(ctx);
3054 mutex_lock(&ctx->uring_lock);
3055 ret = io_submit_sqes(ctx, to_submit);
3056 if (ret != to_submit) {
3057 mutex_unlock(&ctx->uring_lock);
3060 if ((flags & IORING_ENTER_GETEVENTS) && ctx->syscall_iopoll)
3062 mutex_unlock(&ctx->uring_lock);
3064 if (flags & IORING_ENTER_GETEVENTS) {
3066 if (ctx->syscall_iopoll) {
3068 * We disallow the app entering submit/complete with
3069 * polling, but we still need to lock the ring to
3070 * prevent racing with polled issue that got punted to
3073 mutex_lock(&ctx->uring_lock);
3075 ret2 = io_validate_ext_arg(flags, argp, argsz);
3076 if (likely(!ret2)) {
3077 min_complete = min(min_complete,
3079 ret2 = io_iopoll_check(ctx, min_complete);
3081 mutex_unlock(&ctx->uring_lock);
3083 const sigset_t __user *sig;
3084 struct __kernel_timespec __user *ts;
3086 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3087 if (likely(!ret2)) {
3088 min_complete = min(min_complete,
3090 ret2 = io_cqring_wait(ctx, min_complete, sig,
3099 * EBADR indicates that one or more CQE were dropped.
3100 * Once the user has been informed we can clear the bit
3101 * as they are obviously ok with those drops.
3103 if (unlikely(ret2 == -EBADR))
3104 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3110 percpu_ref_put(&ctx->refs);
3116 static const struct file_operations io_uring_fops = {
3117 .release = io_uring_release,
3118 .mmap = io_uring_mmap,
3120 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3121 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3123 .poll = io_uring_poll,
3124 #ifdef CONFIG_PROC_FS
3125 .show_fdinfo = io_uring_show_fdinfo,
3129 bool io_is_uring_fops(struct file *file)
3131 return file->f_op == &io_uring_fops;
3134 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3135 struct io_uring_params *p)
3137 struct io_rings *rings;
3138 size_t size, sq_array_offset;
3140 /* make sure these are sane, as we already accounted them */
3141 ctx->sq_entries = p->sq_entries;
3142 ctx->cq_entries = p->cq_entries;
3144 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3145 if (size == SIZE_MAX)
3148 rings = io_mem_alloc(size);
3153 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3154 rings->sq_ring_mask = p->sq_entries - 1;
3155 rings->cq_ring_mask = p->cq_entries - 1;
3156 rings->sq_ring_entries = p->sq_entries;
3157 rings->cq_ring_entries = p->cq_entries;
3159 if (p->flags & IORING_SETUP_SQE128)
3160 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3162 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3163 if (size == SIZE_MAX) {
3164 io_mem_free(ctx->rings);
3169 ctx->sq_sqes = io_mem_alloc(size);
3170 if (!ctx->sq_sqes) {
3171 io_mem_free(ctx->rings);
3179 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3183 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3187 ret = io_uring_add_tctx_node(ctx);
3192 fd_install(fd, file);
3197 * Allocate an anonymous fd, this is what constitutes the application
3198 * visible backing of an io_uring instance. The application mmaps this
3199 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3200 * we have to tie this fd to a socket for file garbage collection purposes.
3202 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3205 #if defined(CONFIG_UNIX)
3208 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3211 return ERR_PTR(ret);
3214 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3215 O_RDWR | O_CLOEXEC, NULL);
3216 #if defined(CONFIG_UNIX)
3218 sock_release(ctx->ring_sock);
3219 ctx->ring_sock = NULL;
3221 ctx->ring_sock->file = file;
3227 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3228 struct io_uring_params __user *params)
3230 struct io_ring_ctx *ctx;
3236 if (entries > IORING_MAX_ENTRIES) {
3237 if (!(p->flags & IORING_SETUP_CLAMP))
3239 entries = IORING_MAX_ENTRIES;
3243 * Use twice as many entries for the CQ ring. It's possible for the
3244 * application to drive a higher depth than the size of the SQ ring,
3245 * since the sqes are only used at submission time. This allows for
3246 * some flexibility in overcommitting a bit. If the application has
3247 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3248 * of CQ ring entries manually.
3250 p->sq_entries = roundup_pow_of_two(entries);
3251 if (p->flags & IORING_SETUP_CQSIZE) {
3253 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3254 * to a power-of-two, if it isn't already. We do NOT impose
3255 * any cq vs sq ring sizing.
3259 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3260 if (!(p->flags & IORING_SETUP_CLAMP))
3262 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3264 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3265 if (p->cq_entries < p->sq_entries)
3268 p->cq_entries = 2 * p->sq_entries;
3271 ctx = io_ring_ctx_alloc(p);
3276 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3277 * space applications don't need to do io completion events
3278 * polling again, they can rely on io_sq_thread to do polling
3279 * work, which can reduce cpu usage and uring_lock contention.
3281 if (ctx->flags & IORING_SETUP_IOPOLL &&
3282 !(ctx->flags & IORING_SETUP_SQPOLL))
3283 ctx->syscall_iopoll = 1;
3285 ctx->compat = in_compat_syscall();
3286 if (!capable(CAP_IPC_LOCK))
3287 ctx->user = get_uid(current_user());
3290 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3291 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3294 if (ctx->flags & IORING_SETUP_SQPOLL) {
3295 /* IPI related flags don't make sense with SQPOLL */
3296 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3297 IORING_SETUP_TASKRUN_FLAG))
3299 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3300 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3301 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3303 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
3305 ctx->notify_method = TWA_SIGNAL;
3309 * This is just grabbed for accounting purposes. When a process exits,
3310 * the mm is exited and dropped before the files, hence we need to hang
3311 * on to this mm purely for the purposes of being able to unaccount
3312 * memory (locked/pinned vm). It's not used for anything else.
3314 mmgrab(current->mm);
3315 ctx->mm_account = current->mm;
3317 ret = io_allocate_scq_urings(ctx, p);
3321 ret = io_sq_offload_create(ctx, p);
3324 /* always set a rsrc node */
3325 ret = io_rsrc_node_switch_start(ctx);
3328 io_rsrc_node_switch(ctx, NULL);
3330 memset(&p->sq_off, 0, sizeof(p->sq_off));
3331 p->sq_off.head = offsetof(struct io_rings, sq.head);
3332 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3333 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3334 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3335 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3336 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3337 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3339 memset(&p->cq_off, 0, sizeof(p->cq_off));
3340 p->cq_off.head = offsetof(struct io_rings, cq.head);
3341 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3342 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3343 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3344 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3345 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3346 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3348 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3349 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3350 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3351 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3352 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3353 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3354 IORING_FEAT_LINKED_FILE;
3356 if (copy_to_user(params, p, sizeof(*p))) {
3361 file = io_uring_get_file(ctx);
3363 ret = PTR_ERR(file);
3368 * Install ring fd as the very last thing, so we don't risk someone
3369 * having closed it before we finish setup
3371 ret = io_uring_install_fd(ctx, file);
3373 /* fput will clean it up */
3378 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3381 io_ring_ctx_wait_and_kill(ctx);
3386 * Sets up an aio uring context, and returns the fd. Applications asks for a
3387 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3388 * params structure passed in.
3390 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3392 struct io_uring_params p;
3395 if (copy_from_user(&p, params, sizeof(p)))
3397 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3402 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3403 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3404 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3405 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3406 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3407 IORING_SETUP_SQE128 | IORING_SETUP_CQE32))
3410 return io_uring_create(entries, &p, params);
3413 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3414 struct io_uring_params __user *, params)
3416 return io_uring_setup(entries, params);
3419 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3422 struct io_uring_probe *p;
3426 size = struct_size(p, ops, nr_args);
3427 if (size == SIZE_MAX)
3429 p = kzalloc(size, GFP_KERNEL);
3434 if (copy_from_user(p, arg, size))
3437 if (memchr_inv(p, 0, size))
3440 p->last_op = IORING_OP_LAST - 1;
3441 if (nr_args > IORING_OP_LAST)
3442 nr_args = IORING_OP_LAST;
3444 for (i = 0; i < nr_args; i++) {
3446 if (!io_op_defs[i].not_supported)
3447 p->ops[i].flags = IO_URING_OP_SUPPORTED;
3452 if (copy_to_user(arg, p, size))
3459 static int io_register_personality(struct io_ring_ctx *ctx)
3461 const struct cred *creds;
3465 creds = get_current_cred();
3467 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
3468 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
3476 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
3477 void __user *arg, unsigned int nr_args)
3479 struct io_uring_restriction *res;
3483 /* Restrictions allowed only if rings started disabled */
3484 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3487 /* We allow only a single restrictions registration */
3488 if (ctx->restrictions.registered)
3491 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
3494 size = array_size(nr_args, sizeof(*res));
3495 if (size == SIZE_MAX)
3498 res = memdup_user(arg, size);
3500 return PTR_ERR(res);
3504 for (i = 0; i < nr_args; i++) {
3505 switch (res[i].opcode) {
3506 case IORING_RESTRICTION_REGISTER_OP:
3507 if (res[i].register_op >= IORING_REGISTER_LAST) {
3512 __set_bit(res[i].register_op,
3513 ctx->restrictions.register_op);
3515 case IORING_RESTRICTION_SQE_OP:
3516 if (res[i].sqe_op >= IORING_OP_LAST) {
3521 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
3523 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
3524 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
3526 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
3527 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
3536 /* Reset all restrictions if an error happened */
3538 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
3540 ctx->restrictions.registered = true;
3546 static int io_register_enable_rings(struct io_ring_ctx *ctx)
3548 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3551 if (ctx->restrictions.registered)
3552 ctx->restricted = 1;
3554 ctx->flags &= ~IORING_SETUP_R_DISABLED;
3555 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
3556 wake_up(&ctx->sq_data->wait);
3560 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
3561 void __user *arg, unsigned len)
3563 struct io_uring_task *tctx = current->io_uring;
3564 cpumask_var_t new_mask;
3567 if (!tctx || !tctx->io_wq)
3570 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3573 cpumask_clear(new_mask);
3574 if (len > cpumask_size())
3575 len = cpumask_size();
3577 if (in_compat_syscall()) {
3578 ret = compat_get_bitmap(cpumask_bits(new_mask),
3579 (const compat_ulong_t __user *)arg,
3580 len * 8 /* CHAR_BIT */);
3582 ret = copy_from_user(new_mask, arg, len);
3586 free_cpumask_var(new_mask);
3590 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
3591 free_cpumask_var(new_mask);
3595 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
3597 struct io_uring_task *tctx = current->io_uring;
3599 if (!tctx || !tctx->io_wq)
3602 return io_wq_cpu_affinity(tctx->io_wq, NULL);
3605 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
3607 __must_hold(&ctx->uring_lock)
3609 struct io_tctx_node *node;
3610 struct io_uring_task *tctx = NULL;
3611 struct io_sq_data *sqd = NULL;
3615 if (copy_from_user(new_count, arg, sizeof(new_count)))
3617 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3618 if (new_count[i] > INT_MAX)
3621 if (ctx->flags & IORING_SETUP_SQPOLL) {
3625 * Observe the correct sqd->lock -> ctx->uring_lock
3626 * ordering. Fine to drop uring_lock here, we hold
3629 refcount_inc(&sqd->refs);
3630 mutex_unlock(&ctx->uring_lock);
3631 mutex_lock(&sqd->lock);
3632 mutex_lock(&ctx->uring_lock);
3634 tctx = sqd->thread->io_uring;
3637 tctx = current->io_uring;
3640 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
3642 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3644 ctx->iowq_limits[i] = new_count[i];
3645 ctx->iowq_limits_set = true;
3647 if (tctx && tctx->io_wq) {
3648 ret = io_wq_max_workers(tctx->io_wq, new_count);
3652 memset(new_count, 0, sizeof(new_count));
3656 mutex_unlock(&sqd->lock);
3657 io_put_sq_data(sqd);
3660 if (copy_to_user(arg, new_count, sizeof(new_count)))
3663 /* that's it for SQPOLL, only the SQPOLL task creates requests */
3667 /* now propagate the restriction to all registered users */
3668 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3669 struct io_uring_task *tctx = node->task->io_uring;
3671 if (WARN_ON_ONCE(!tctx->io_wq))
3674 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3675 new_count[i] = ctx->iowq_limits[i];
3676 /* ignore errors, it always returns zero anyway */
3677 (void)io_wq_max_workers(tctx->io_wq, new_count);
3682 mutex_unlock(&sqd->lock);
3683 io_put_sq_data(sqd);
3688 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
3689 void __user *arg, unsigned nr_args)
3690 __releases(ctx->uring_lock)
3691 __acquires(ctx->uring_lock)
3696 * We're inside the ring mutex, if the ref is already dying, then
3697 * someone else killed the ctx or is already going through
3698 * io_uring_register().
3700 if (percpu_ref_is_dying(&ctx->refs))
3703 if (ctx->restricted) {
3704 if (opcode >= IORING_REGISTER_LAST)
3706 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
3707 if (!test_bit(opcode, ctx->restrictions.register_op))
3712 case IORING_REGISTER_BUFFERS:
3716 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
3718 case IORING_UNREGISTER_BUFFERS:
3722 ret = io_sqe_buffers_unregister(ctx);
3724 case IORING_REGISTER_FILES:
3728 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
3730 case IORING_UNREGISTER_FILES:
3734 ret = io_sqe_files_unregister(ctx);
3736 case IORING_REGISTER_FILES_UPDATE:
3737 ret = io_register_files_update(ctx, arg, nr_args);
3739 case IORING_REGISTER_EVENTFD:
3743 ret = io_eventfd_register(ctx, arg, 0);
3745 case IORING_REGISTER_EVENTFD_ASYNC:
3749 ret = io_eventfd_register(ctx, arg, 1);
3751 case IORING_UNREGISTER_EVENTFD:
3755 ret = io_eventfd_unregister(ctx);
3757 case IORING_REGISTER_PROBE:
3759 if (!arg || nr_args > 256)
3761 ret = io_probe(ctx, arg, nr_args);
3763 case IORING_REGISTER_PERSONALITY:
3767 ret = io_register_personality(ctx);
3769 case IORING_UNREGISTER_PERSONALITY:
3773 ret = io_unregister_personality(ctx, nr_args);
3775 case IORING_REGISTER_ENABLE_RINGS:
3779 ret = io_register_enable_rings(ctx);
3781 case IORING_REGISTER_RESTRICTIONS:
3782 ret = io_register_restrictions(ctx, arg, nr_args);
3784 case IORING_REGISTER_FILES2:
3785 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
3787 case IORING_REGISTER_FILES_UPDATE2:
3788 ret = io_register_rsrc_update(ctx, arg, nr_args,
3791 case IORING_REGISTER_BUFFERS2:
3792 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
3794 case IORING_REGISTER_BUFFERS_UPDATE:
3795 ret = io_register_rsrc_update(ctx, arg, nr_args,
3796 IORING_RSRC_BUFFER);
3798 case IORING_REGISTER_IOWQ_AFF:
3800 if (!arg || !nr_args)
3802 ret = io_register_iowq_aff(ctx, arg, nr_args);
3804 case IORING_UNREGISTER_IOWQ_AFF:
3808 ret = io_unregister_iowq_aff(ctx);
3810 case IORING_REGISTER_IOWQ_MAX_WORKERS:
3812 if (!arg || nr_args != 2)
3814 ret = io_register_iowq_max_workers(ctx, arg);
3816 case IORING_REGISTER_RING_FDS:
3817 ret = io_ringfd_register(ctx, arg, nr_args);
3819 case IORING_UNREGISTER_RING_FDS:
3820 ret = io_ringfd_unregister(ctx, arg, nr_args);
3822 case IORING_REGISTER_PBUF_RING:
3824 if (!arg || nr_args != 1)
3826 ret = io_register_pbuf_ring(ctx, arg);
3828 case IORING_UNREGISTER_PBUF_RING:
3830 if (!arg || nr_args != 1)
3832 ret = io_unregister_pbuf_ring(ctx, arg);
3842 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
3843 void __user *, arg, unsigned int, nr_args)
3845 struct io_ring_ctx *ctx;
3854 if (!io_is_uring_fops(f.file))
3857 ctx = f.file->private_data;
3861 mutex_lock(&ctx->uring_lock);
3862 ret = __io_uring_register(ctx, opcode, arg, nr_args);
3863 mutex_unlock(&ctx->uring_lock);
3864 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
3870 static int __init io_uring_init(void)
3872 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
3873 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
3874 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
3877 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
3878 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
3879 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
3880 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
3881 BUILD_BUG_SQE_ELEM(1, __u8, flags);
3882 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
3883 BUILD_BUG_SQE_ELEM(4, __s32, fd);
3884 BUILD_BUG_SQE_ELEM(8, __u64, off);
3885 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
3886 BUILD_BUG_SQE_ELEM(16, __u64, addr);
3887 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
3888 BUILD_BUG_SQE_ELEM(24, __u32, len);
3889 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
3890 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
3891 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
3892 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
3893 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
3894 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
3895 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
3896 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
3897 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
3898 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
3899 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
3900 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
3901 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
3902 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
3903 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
3904 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
3905 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
3906 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
3907 BUILD_BUG_SQE_ELEM(42, __u16, personality);
3908 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
3909 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
3910 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
3912 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
3913 sizeof(struct io_uring_rsrc_update));
3914 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
3915 sizeof(struct io_uring_rsrc_update2));
3917 /* ->buf_index is u16 */
3918 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
3919 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
3920 offsetof(struct io_uring_buf_ring, tail));
3922 /* should fit into one byte */
3923 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
3924 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
3925 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
3927 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
3929 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
3931 io_uring_optable_init();
3933 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
3937 __initcall(io_uring_init);