1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <net/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
49 #include <linux/bits.h>
51 #include <linux/sched/signal.h>
53 #include <linux/file.h>
54 #include <linux/fdtable.h>
56 #include <linux/mman.h>
57 #include <linux/percpu.h>
58 #include <linux/slab.h>
59 #include <linux/bvec.h>
60 #include <linux/net.h>
62 #include <net/af_unix.h>
64 #include <linux/anon_inodes.h>
65 #include <linux/sched/mm.h>
66 #include <linux/uaccess.h>
67 #include <linux/nospec.h>
68 #include <linux/highmem.h>
69 #include <linux/fsnotify.h>
70 #include <linux/fadvise.h>
71 #include <linux/task_work.h>
72 #include <linux/io_uring.h>
73 #include <linux/audit.h>
74 #include <linux/security.h>
76 #define CREATE_TRACE_POINTS
77 #include <trace/events/io_uring.h>
79 #include <uapi/linux/io_uring.h>
83 #include "io_uring_types.h"
97 #define IORING_MAX_ENTRIES 32768
98 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
100 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
101 IORING_REGISTER_LAST + IORING_OP_LAST)
103 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
104 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
106 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
107 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
109 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
110 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
113 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
116 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
118 #define IO_COMPL_BATCH 32
119 #define IO_REQ_ALLOC_BATCH 8
122 IO_CHECK_CQ_OVERFLOW_BIT,
123 IO_CHECK_CQ_DROPPED_BIT,
126 struct io_defer_entry {
127 struct list_head list;
128 struct io_kiocb *req;
132 /* requests with any of those set should undergo io_disarm_next() */
133 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
134 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
136 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
137 struct task_struct *task,
140 static void io_dismantle_req(struct io_kiocb *req);
141 static void io_clean_op(struct io_kiocb *req);
142 static void io_queue_sqe(struct io_kiocb *req);
144 static void __io_submit_flush_completions(struct io_ring_ctx *ctx);
146 static void io_eventfd_signal(struct io_ring_ctx *ctx);
148 static struct kmem_cache *req_cachep;
150 struct sock *io_uring_get_socket(struct file *file)
152 #if defined(CONFIG_UNIX)
153 if (io_is_uring_fops(file)) {
154 struct io_ring_ctx *ctx = file->private_data;
156 return ctx->ring_sock->sk;
161 EXPORT_SYMBOL(io_uring_get_socket);
163 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
165 if (!wq_list_empty(&ctx->submit_state.compl_reqs))
166 __io_submit_flush_completions(ctx);
169 static bool io_match_linked(struct io_kiocb *head)
171 struct io_kiocb *req;
173 io_for_each_link(req, head) {
174 if (req->flags & REQ_F_INFLIGHT)
181 * As io_match_task() but protected against racing with linked timeouts.
182 * User must not hold timeout_lock.
184 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
189 if (task && head->task != task)
194 if (head->flags & REQ_F_LINK_TIMEOUT) {
195 struct io_ring_ctx *ctx = head->ctx;
197 /* protect against races with linked timeouts */
198 spin_lock_irq(&ctx->timeout_lock);
199 matched = io_match_linked(head);
200 spin_unlock_irq(&ctx->timeout_lock);
202 matched = io_match_linked(head);
207 static inline void req_fail_link_node(struct io_kiocb *req, int res)
210 io_req_set_res(req, res, 0);
213 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
215 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
218 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
220 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
222 complete(&ctx->ref_comp);
225 static __cold void io_fallback_req_func(struct work_struct *work)
227 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
229 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
230 struct io_kiocb *req, *tmp;
233 percpu_ref_get(&ctx->refs);
234 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
235 req->io_task_work.func(req, &locked);
238 io_submit_flush_completions(ctx);
239 mutex_unlock(&ctx->uring_lock);
241 percpu_ref_put(&ctx->refs);
244 static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
246 unsigned hash_buckets = 1U << bits;
247 size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
249 table->hbs = kmalloc(hash_size, GFP_KERNEL);
253 table->hash_bits = bits;
254 init_hash_table(table, hash_buckets);
258 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
260 struct io_ring_ctx *ctx;
263 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
267 xa_init(&ctx->io_bl_xa);
270 * Use 5 bits less than the max cq entries, that should give us around
271 * 32 entries per hash list if totally full and uniformly spread, but
272 * don't keep too many buckets to not overconsume memory.
274 hash_bits = ilog2(p->cq_entries) - 5;
275 hash_bits = clamp(hash_bits, 1, 8);
276 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
278 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
281 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
282 if (!ctx->dummy_ubuf)
284 /* set invalid range, so io_import_fixed() fails meeting it */
285 ctx->dummy_ubuf->ubuf = -1UL;
287 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
288 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
291 ctx->flags = p->flags;
292 init_waitqueue_head(&ctx->sqo_sq_wait);
293 INIT_LIST_HEAD(&ctx->sqd_list);
294 INIT_LIST_HEAD(&ctx->cq_overflow_list);
295 INIT_LIST_HEAD(&ctx->io_buffers_cache);
296 INIT_LIST_HEAD(&ctx->apoll_cache);
297 init_completion(&ctx->ref_comp);
298 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
299 mutex_init(&ctx->uring_lock);
300 init_waitqueue_head(&ctx->cq_wait);
301 spin_lock_init(&ctx->completion_lock);
302 spin_lock_init(&ctx->timeout_lock);
303 INIT_WQ_LIST(&ctx->iopoll_list);
304 INIT_LIST_HEAD(&ctx->io_buffers_pages);
305 INIT_LIST_HEAD(&ctx->io_buffers_comp);
306 INIT_LIST_HEAD(&ctx->defer_list);
307 INIT_LIST_HEAD(&ctx->timeout_list);
308 INIT_LIST_HEAD(&ctx->ltimeout_list);
309 spin_lock_init(&ctx->rsrc_ref_lock);
310 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
311 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
312 init_llist_head(&ctx->rsrc_put_llist);
313 INIT_LIST_HEAD(&ctx->tctx_list);
314 ctx->submit_state.free_list.next = NULL;
315 INIT_WQ_LIST(&ctx->locked_free_list);
316 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
317 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
320 kfree(ctx->dummy_ubuf);
321 kfree(ctx->cancel_table.hbs);
322 kfree(ctx->cancel_table_locked.hbs);
324 xa_destroy(&ctx->io_bl_xa);
329 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
331 struct io_rings *r = ctx->rings;
333 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
337 static bool req_need_defer(struct io_kiocb *req, u32 seq)
339 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
340 struct io_ring_ctx *ctx = req->ctx;
342 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
348 static inline void io_req_track_inflight(struct io_kiocb *req)
350 if (!(req->flags & REQ_F_INFLIGHT)) {
351 req->flags |= REQ_F_INFLIGHT;
352 atomic_inc(&req->task->io_uring->inflight_tracked);
356 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
358 if (WARN_ON_ONCE(!req->link))
361 req->flags &= ~REQ_F_ARM_LTIMEOUT;
362 req->flags |= REQ_F_LINK_TIMEOUT;
364 /* linked timeouts should have two refs once prep'ed */
365 io_req_set_refcount(req);
366 __io_req_set_refcount(req->link, 2);
370 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
372 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
374 return __io_prep_linked_timeout(req);
377 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
379 io_queue_linked_timeout(__io_prep_linked_timeout(req));
382 static inline void io_arm_ltimeout(struct io_kiocb *req)
384 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
385 __io_arm_ltimeout(req);
388 static void io_prep_async_work(struct io_kiocb *req)
390 const struct io_op_def *def = &io_op_defs[req->opcode];
391 struct io_ring_ctx *ctx = req->ctx;
393 if (!(req->flags & REQ_F_CREDS)) {
394 req->flags |= REQ_F_CREDS;
395 req->creds = get_current_cred();
398 req->work.list.next = NULL;
400 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
401 if (req->flags & REQ_F_FORCE_ASYNC)
402 req->work.flags |= IO_WQ_WORK_CONCURRENT;
404 if (req->flags & REQ_F_ISREG) {
405 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
406 io_wq_hash_work(&req->work, file_inode(req->file));
407 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
408 if (def->unbound_nonreg_file)
409 req->work.flags |= IO_WQ_WORK_UNBOUND;
413 static void io_prep_async_link(struct io_kiocb *req)
415 struct io_kiocb *cur;
417 if (req->flags & REQ_F_LINK_TIMEOUT) {
418 struct io_ring_ctx *ctx = req->ctx;
420 spin_lock_irq(&ctx->timeout_lock);
421 io_for_each_link(cur, req)
422 io_prep_async_work(cur);
423 spin_unlock_irq(&ctx->timeout_lock);
425 io_for_each_link(cur, req)
426 io_prep_async_work(cur);
430 void io_queue_iowq(struct io_kiocb *req, bool *dont_use)
432 struct io_kiocb *link = io_prep_linked_timeout(req);
433 struct io_uring_task *tctx = req->task->io_uring;
436 BUG_ON(!tctx->io_wq);
438 /* init ->work of the whole link before punting */
439 io_prep_async_link(req);
442 * Not expected to happen, but if we do have a bug where this _can_
443 * happen, catch it here and ensure the request is marked as
444 * canceled. That will make io-wq go through the usual work cancel
445 * procedure rather than attempt to run this request (or create a new
448 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
449 req->work.flags |= IO_WQ_WORK_CANCEL;
451 trace_io_uring_queue_async_work(req->ctx, req, req->cqe.user_data,
452 req->opcode, req->flags, &req->work,
453 io_wq_is_hashed(&req->work));
454 io_wq_enqueue(tctx->io_wq, &req->work);
456 io_queue_linked_timeout(link);
459 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
461 while (!list_empty(&ctx->defer_list)) {
462 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
463 struct io_defer_entry, list);
465 if (req_need_defer(de->req, de->seq))
467 list_del_init(&de->list);
468 io_req_task_queue(de->req);
473 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
475 if (ctx->off_timeout_used || ctx->drain_active) {
476 spin_lock(&ctx->completion_lock);
477 if (ctx->off_timeout_used)
478 io_flush_timeouts(ctx);
479 if (ctx->drain_active)
480 io_queue_deferred(ctx);
481 io_commit_cqring(ctx);
482 spin_unlock(&ctx->completion_lock);
485 io_eventfd_signal(ctx);
488 static void io_eventfd_signal(struct io_ring_ctx *ctx)
490 struct io_ev_fd *ev_fd;
494 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
497 ev_fd = rcu_dereference(ctx->io_ev_fd);
500 * Check again if ev_fd exists incase an io_eventfd_unregister call
501 * completed between the NULL check of ctx->io_ev_fd at the start of
502 * the function and rcu_read_lock.
504 if (unlikely(!ev_fd))
506 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
509 if (!ev_fd->eventfd_async || io_wq_current_is_worker())
510 eventfd_signal(ev_fd->cq_ev_fd, 1);
516 * This should only get called when at least one event has been posted.
517 * Some applications rely on the eventfd notification count only changing
518 * IFF a new CQE has been added to the CQ ring. There's no depedency on
519 * 1:1 relationship between how many times this function is called (and
520 * hence the eventfd count) and number of CQEs posted to the CQ ring.
522 void io_cqring_ev_posted(struct io_ring_ctx *ctx)
524 if (unlikely(ctx->off_timeout_used || ctx->drain_active ||
526 __io_commit_cqring_flush(ctx);
531 /* Returns true if there are no backlogged entries after the flush */
532 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
534 bool all_flushed, posted;
535 size_t cqe_size = sizeof(struct io_uring_cqe);
537 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
540 if (ctx->flags & IORING_SETUP_CQE32)
544 spin_lock(&ctx->completion_lock);
545 while (!list_empty(&ctx->cq_overflow_list)) {
546 struct io_uring_cqe *cqe = io_get_cqe(ctx);
547 struct io_overflow_cqe *ocqe;
551 ocqe = list_first_entry(&ctx->cq_overflow_list,
552 struct io_overflow_cqe, list);
554 memcpy(cqe, &ocqe->cqe, cqe_size);
556 io_account_cq_overflow(ctx);
559 list_del(&ocqe->list);
563 all_flushed = list_empty(&ctx->cq_overflow_list);
565 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
566 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
569 io_commit_cqring(ctx);
570 spin_unlock(&ctx->completion_lock);
572 io_cqring_ev_posted(ctx);
576 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
580 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
581 /* iopoll syncs against uring_lock, not completion_lock */
582 if (ctx->flags & IORING_SETUP_IOPOLL)
583 mutex_lock(&ctx->uring_lock);
584 ret = __io_cqring_overflow_flush(ctx, false);
585 if (ctx->flags & IORING_SETUP_IOPOLL)
586 mutex_unlock(&ctx->uring_lock);
592 static void __io_put_task(struct task_struct *task, int nr)
594 struct io_uring_task *tctx = task->io_uring;
596 percpu_counter_sub(&tctx->inflight, nr);
597 if (unlikely(atomic_read(&tctx->in_idle)))
598 wake_up(&tctx->wait);
599 put_task_struct_many(task, nr);
602 /* must to be called somewhat shortly after putting a request */
603 static inline void io_put_task(struct task_struct *task, int nr)
605 if (likely(task == current))
606 task->io_uring->cached_refs += nr;
608 __io_put_task(task, nr);
611 static void io_task_refs_refill(struct io_uring_task *tctx)
613 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
615 percpu_counter_add(&tctx->inflight, refill);
616 refcount_add(refill, ¤t->usage);
617 tctx->cached_refs += refill;
620 static inline void io_get_task_refs(int nr)
622 struct io_uring_task *tctx = current->io_uring;
624 tctx->cached_refs -= nr;
625 if (unlikely(tctx->cached_refs < 0))
626 io_task_refs_refill(tctx);
629 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
631 struct io_uring_task *tctx = task->io_uring;
632 unsigned int refs = tctx->cached_refs;
635 tctx->cached_refs = 0;
636 percpu_counter_sub(&tctx->inflight, refs);
637 put_task_struct_many(task, refs);
641 bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data, s32 res,
642 u32 cflags, u64 extra1, u64 extra2)
644 struct io_overflow_cqe *ocqe;
645 size_t ocq_size = sizeof(struct io_overflow_cqe);
646 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
649 ocq_size += sizeof(struct io_uring_cqe);
651 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
652 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
655 * If we're in ring overflow flush mode, or in task cancel mode,
656 * or cannot allocate an overflow entry, then we need to drop it
659 io_account_cq_overflow(ctx);
660 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
663 if (list_empty(&ctx->cq_overflow_list)) {
664 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
665 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
668 ocqe->cqe.user_data = user_data;
670 ocqe->cqe.flags = cflags;
672 ocqe->cqe.big_cqe[0] = extra1;
673 ocqe->cqe.big_cqe[1] = extra2;
675 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
679 static bool io_fill_cqe_aux(struct io_ring_ctx *ctx,
680 u64 user_data, s32 res, u32 cflags)
682 struct io_uring_cqe *cqe;
685 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
688 * If we can't get a cq entry, userspace overflowed the
689 * submission (by quite a lot). Increment the overflow count in
692 cqe = io_get_cqe(ctx);
694 WRITE_ONCE(cqe->user_data, user_data);
695 WRITE_ONCE(cqe->res, res);
696 WRITE_ONCE(cqe->flags, cflags);
698 if (ctx->flags & IORING_SETUP_CQE32) {
699 WRITE_ONCE(cqe->big_cqe[0], 0);
700 WRITE_ONCE(cqe->big_cqe[1], 0);
704 return io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
707 bool io_post_aux_cqe(struct io_ring_ctx *ctx,
708 u64 user_data, s32 res, u32 cflags)
712 spin_lock(&ctx->completion_lock);
713 filled = io_fill_cqe_aux(ctx, user_data, res, cflags);
714 io_commit_cqring(ctx);
715 spin_unlock(&ctx->completion_lock);
717 io_cqring_ev_posted(ctx);
721 static void __io_req_complete_put(struct io_kiocb *req)
724 * If we're the last reference to this request, add to our locked
727 if (req_ref_put_and_test(req)) {
728 struct io_ring_ctx *ctx = req->ctx;
730 if (req->flags & IO_REQ_LINK_FLAGS) {
731 if (req->flags & IO_DISARM_MASK)
734 io_req_task_queue(req->link);
738 io_req_put_rsrc(req);
740 * Selected buffer deallocation in io_clean_op() assumes that
741 * we don't hold ->completion_lock. Clean them here to avoid
744 io_put_kbuf_comp(req);
745 io_dismantle_req(req);
746 io_put_task(req->task, 1);
747 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
748 ctx->locked_free_nr++;
752 void __io_req_complete_post(struct io_kiocb *req)
754 if (!(req->flags & REQ_F_CQE_SKIP))
755 __io_fill_cqe_req(req->ctx, req);
756 __io_req_complete_put(req);
759 void io_req_complete_post(struct io_kiocb *req)
761 struct io_ring_ctx *ctx = req->ctx;
763 spin_lock(&ctx->completion_lock);
764 __io_req_complete_post(req);
765 io_commit_cqring(ctx);
766 spin_unlock(&ctx->completion_lock);
767 io_cqring_ev_posted(ctx);
770 inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags)
772 io_req_complete_post(req);
775 void io_req_complete_failed(struct io_kiocb *req, s32 res)
778 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
779 io_req_complete_post(req);
783 * Don't initialise the fields below on every allocation, but do that in
784 * advance and keep them valid across allocations.
786 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
790 req->async_data = NULL;
791 /* not necessary, but safer to zero */
795 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
796 struct io_submit_state *state)
798 spin_lock(&ctx->completion_lock);
799 wq_list_splice(&ctx->locked_free_list, &state->free_list);
800 ctx->locked_free_nr = 0;
801 spin_unlock(&ctx->completion_lock);
804 static inline bool io_req_cache_empty(struct io_ring_ctx *ctx)
806 return !ctx->submit_state.free_list.next;
810 * A request might get retired back into the request caches even before opcode
811 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
812 * Because of that, io_alloc_req() should be called only under ->uring_lock
813 * and with extra caution to not get a request that is still worked on.
815 static __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
816 __must_hold(&ctx->uring_lock)
818 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
819 void *reqs[IO_REQ_ALLOC_BATCH];
823 * If we have more than a batch's worth of requests in our IRQ side
824 * locked cache, grab the lock and move them over to our submission
827 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
828 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
829 if (!io_req_cache_empty(ctx))
833 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
836 * Bulk alloc is all-or-nothing. If we fail to get a batch,
837 * retry single alloc to be on the safe side.
839 if (unlikely(ret <= 0)) {
840 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
846 percpu_ref_get_many(&ctx->refs, ret);
847 for (i = 0; i < ret; i++) {
848 struct io_kiocb *req = reqs[i];
850 io_preinit_req(req, ctx);
851 io_req_add_to_cache(req, ctx);
856 static inline bool io_alloc_req_refill(struct io_ring_ctx *ctx)
858 if (unlikely(io_req_cache_empty(ctx)))
859 return __io_alloc_req_refill(ctx);
863 static inline struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
865 struct io_wq_work_node *node;
867 node = wq_stack_extract(&ctx->submit_state.free_list);
868 return container_of(node, struct io_kiocb, comp_list);
871 static inline void io_dismantle_req(struct io_kiocb *req)
873 unsigned int flags = req->flags;
875 if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
877 if (!(flags & REQ_F_FIXED_FILE))
878 io_put_file(req->file);
881 __cold void io_free_req(struct io_kiocb *req)
883 struct io_ring_ctx *ctx = req->ctx;
885 io_req_put_rsrc(req);
886 io_dismantle_req(req);
887 io_put_task(req->task, 1);
889 spin_lock(&ctx->completion_lock);
890 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
891 ctx->locked_free_nr++;
892 spin_unlock(&ctx->completion_lock);
895 static void __io_req_find_next_prep(struct io_kiocb *req)
897 struct io_ring_ctx *ctx = req->ctx;
900 spin_lock(&ctx->completion_lock);
901 posted = io_disarm_next(req);
902 io_commit_cqring(ctx);
903 spin_unlock(&ctx->completion_lock);
905 io_cqring_ev_posted(ctx);
908 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
910 struct io_kiocb *nxt;
913 * If LINK is set, we have dependent requests in this chain. If we
914 * didn't fail this request, queue the first one up, moving any other
915 * dependencies to the next request. In case of failure, fail the rest
918 if (unlikely(req->flags & IO_DISARM_MASK))
919 __io_req_find_next_prep(req);
925 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
929 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
930 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
932 io_submit_flush_completions(ctx);
933 mutex_unlock(&ctx->uring_lock);
936 percpu_ref_put(&ctx->refs);
939 static inline void ctx_commit_and_unlock(struct io_ring_ctx *ctx)
941 io_commit_cqring(ctx);
942 spin_unlock(&ctx->completion_lock);
943 io_cqring_ev_posted(ctx);
946 static void handle_prev_tw_list(struct io_wq_work_node *node,
947 struct io_ring_ctx **ctx, bool *uring_locked)
949 if (*ctx && !*uring_locked)
950 spin_lock(&(*ctx)->completion_lock);
953 struct io_wq_work_node *next = node->next;
954 struct io_kiocb *req = container_of(node, struct io_kiocb,
957 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
959 if (req->ctx != *ctx) {
960 if (unlikely(!*uring_locked && *ctx))
961 ctx_commit_and_unlock(*ctx);
963 ctx_flush_and_put(*ctx, uring_locked);
965 /* if not contended, grab and improve batching */
966 *uring_locked = mutex_trylock(&(*ctx)->uring_lock);
967 percpu_ref_get(&(*ctx)->refs);
968 if (unlikely(!*uring_locked))
969 spin_lock(&(*ctx)->completion_lock);
971 if (likely(*uring_locked)) {
972 req->io_task_work.func(req, uring_locked);
974 req->cqe.flags = io_put_kbuf_comp(req);
975 __io_req_complete_post(req);
980 if (unlikely(!*uring_locked))
981 ctx_commit_and_unlock(*ctx);
984 static void handle_tw_list(struct io_wq_work_node *node,
985 struct io_ring_ctx **ctx, bool *locked)
988 struct io_wq_work_node *next = node->next;
989 struct io_kiocb *req = container_of(node, struct io_kiocb,
992 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
994 if (req->ctx != *ctx) {
995 ctx_flush_and_put(*ctx, locked);
997 /* if not contended, grab and improve batching */
998 *locked = mutex_trylock(&(*ctx)->uring_lock);
999 percpu_ref_get(&(*ctx)->refs);
1001 req->io_task_work.func(req, locked);
1006 void tctx_task_work(struct callback_head *cb)
1008 bool uring_locked = false;
1009 struct io_ring_ctx *ctx = NULL;
1010 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1014 struct io_wq_work_node *node1, *node2;
1016 spin_lock_irq(&tctx->task_lock);
1017 node1 = tctx->prio_task_list.first;
1018 node2 = tctx->task_list.first;
1019 INIT_WQ_LIST(&tctx->task_list);
1020 INIT_WQ_LIST(&tctx->prio_task_list);
1021 if (!node2 && !node1)
1022 tctx->task_running = false;
1023 spin_unlock_irq(&tctx->task_lock);
1024 if (!node2 && !node1)
1028 handle_prev_tw_list(node1, &ctx, &uring_locked);
1030 handle_tw_list(node2, &ctx, &uring_locked);
1033 if (data_race(!tctx->task_list.first) &&
1034 data_race(!tctx->prio_task_list.first) && uring_locked)
1035 io_submit_flush_completions(ctx);
1038 ctx_flush_and_put(ctx, &uring_locked);
1040 /* relaxed read is enough as only the task itself sets ->in_idle */
1041 if (unlikely(atomic_read(&tctx->in_idle)))
1042 io_uring_drop_tctx_refs(current);
1045 static void __io_req_task_work_add(struct io_kiocb *req,
1046 struct io_uring_task *tctx,
1047 struct io_wq_work_list *list)
1049 struct io_ring_ctx *ctx = req->ctx;
1050 struct io_wq_work_node *node;
1051 unsigned long flags;
1054 spin_lock_irqsave(&tctx->task_lock, flags);
1055 wq_list_add_tail(&req->io_task_work.node, list);
1056 running = tctx->task_running;
1058 tctx->task_running = true;
1059 spin_unlock_irqrestore(&tctx->task_lock, flags);
1061 /* task_work already pending, we're done */
1065 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1066 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1068 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1071 spin_lock_irqsave(&tctx->task_lock, flags);
1072 tctx->task_running = false;
1073 node = wq_list_merge(&tctx->prio_task_list, &tctx->task_list);
1074 spin_unlock_irqrestore(&tctx->task_lock, flags);
1077 req = container_of(node, struct io_kiocb, io_task_work.node);
1079 if (llist_add(&req->io_task_work.fallback_node,
1080 &req->ctx->fallback_llist))
1081 schedule_delayed_work(&req->ctx->fallback_work, 1);
1085 void io_req_task_work_add(struct io_kiocb *req)
1087 struct io_uring_task *tctx = req->task->io_uring;
1089 __io_req_task_work_add(req, tctx, &tctx->task_list);
1092 void io_req_task_prio_work_add(struct io_kiocb *req)
1094 struct io_uring_task *tctx = req->task->io_uring;
1096 if (req->ctx->flags & IORING_SETUP_SQPOLL)
1097 __io_req_task_work_add(req, tctx, &tctx->prio_task_list);
1099 __io_req_task_work_add(req, tctx, &tctx->task_list);
1102 static void io_req_tw_post(struct io_kiocb *req, bool *locked)
1104 io_req_complete_post(req);
1107 void io_req_tw_post_queue(struct io_kiocb *req, s32 res, u32 cflags)
1109 io_req_set_res(req, res, cflags);
1110 req->io_task_work.func = io_req_tw_post;
1111 io_req_task_work_add(req);
1114 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
1116 /* not needed for normal modes, but SQPOLL depends on it */
1117 io_tw_lock(req->ctx, locked);
1118 io_req_complete_failed(req, req->cqe.res);
1121 void io_req_task_submit(struct io_kiocb *req, bool *locked)
1123 io_tw_lock(req->ctx, locked);
1124 /* req->task == current here, checking PF_EXITING is safe */
1125 if (likely(!(req->task->flags & PF_EXITING)))
1128 io_req_complete_failed(req, -EFAULT);
1131 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1133 io_req_set_res(req, ret, 0);
1134 req->io_task_work.func = io_req_task_cancel;
1135 io_req_task_work_add(req);
1138 void io_req_task_queue(struct io_kiocb *req)
1140 req->io_task_work.func = io_req_task_submit;
1141 io_req_task_work_add(req);
1144 void io_queue_next(struct io_kiocb *req)
1146 struct io_kiocb *nxt = io_req_find_next(req);
1149 io_req_task_queue(nxt);
1152 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1153 __must_hold(&ctx->uring_lock)
1155 struct task_struct *task = NULL;
1159 struct io_kiocb *req = container_of(node, struct io_kiocb,
1162 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1163 if (req->flags & REQ_F_REFCOUNT) {
1164 node = req->comp_list.next;
1165 if (!req_ref_put_and_test(req))
1168 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1169 struct async_poll *apoll = req->apoll;
1171 if (apoll->double_poll)
1172 kfree(apoll->double_poll);
1173 list_add(&apoll->poll.wait.entry,
1175 req->flags &= ~REQ_F_POLLED;
1177 if (req->flags & IO_REQ_LINK_FLAGS)
1179 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1182 if (!(req->flags & REQ_F_FIXED_FILE))
1183 io_put_file(req->file);
1185 io_req_put_rsrc_locked(req, ctx);
1187 if (req->task != task) {
1189 io_put_task(task, task_refs);
1194 node = req->comp_list.next;
1195 io_req_add_to_cache(req, ctx);
1199 io_put_task(task, task_refs);
1202 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1203 __must_hold(&ctx->uring_lock)
1205 struct io_wq_work_node *node, *prev;
1206 struct io_submit_state *state = &ctx->submit_state;
1208 if (state->flush_cqes) {
1209 spin_lock(&ctx->completion_lock);
1210 wq_list_for_each(node, prev, &state->compl_reqs) {
1211 struct io_kiocb *req = container_of(node, struct io_kiocb,
1214 if (!(req->flags & REQ_F_CQE_SKIP))
1215 __io_fill_cqe_req(ctx, req);
1218 io_commit_cqring(ctx);
1219 spin_unlock(&ctx->completion_lock);
1220 io_cqring_ev_posted(ctx);
1221 state->flush_cqes = false;
1224 io_free_batch_list(ctx, state->compl_reqs.first);
1225 INIT_WQ_LIST(&state->compl_reqs);
1229 * Drop reference to request, return next in chain (if there is one) if this
1230 * was the last reference to this request.
1232 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1234 struct io_kiocb *nxt = NULL;
1236 if (req_ref_put_and_test(req)) {
1237 if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
1238 nxt = io_req_find_next(req);
1244 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1246 /* See comment at the top of this file */
1248 return __io_cqring_events(ctx);
1252 * We can't just wait for polled events to come to us, we have to actively
1253 * find and complete them.
1255 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1257 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1260 mutex_lock(&ctx->uring_lock);
1261 while (!wq_list_empty(&ctx->iopoll_list)) {
1262 /* let it sleep and repeat later if can't complete a request */
1263 if (io_do_iopoll(ctx, true) == 0)
1266 * Ensure we allow local-to-the-cpu processing to take place,
1267 * in this case we need to ensure that we reap all events.
1268 * Also let task_work, etc. to progress by releasing the mutex
1270 if (need_resched()) {
1271 mutex_unlock(&ctx->uring_lock);
1273 mutex_lock(&ctx->uring_lock);
1276 mutex_unlock(&ctx->uring_lock);
1279 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1281 unsigned int nr_events = 0;
1283 unsigned long check_cq;
1285 check_cq = READ_ONCE(ctx->check_cq);
1286 if (unlikely(check_cq)) {
1287 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1288 __io_cqring_overflow_flush(ctx, false);
1290 * Similarly do not spin if we have not informed the user of any
1293 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1297 * Don't enter poll loop if we already have events pending.
1298 * If we do, we can potentially be spinning for commands that
1299 * already triggered a CQE (eg in error).
1301 if (io_cqring_events(ctx))
1306 * If a submit got punted to a workqueue, we can have the
1307 * application entering polling for a command before it gets
1308 * issued. That app will hold the uring_lock for the duration
1309 * of the poll right here, so we need to take a breather every
1310 * now and then to ensure that the issue has a chance to add
1311 * the poll to the issued list. Otherwise we can spin here
1312 * forever, while the workqueue is stuck trying to acquire the
1315 if (wq_list_empty(&ctx->iopoll_list)) {
1316 u32 tail = ctx->cached_cq_tail;
1318 mutex_unlock(&ctx->uring_lock);
1320 mutex_lock(&ctx->uring_lock);
1322 /* some requests don't go through iopoll_list */
1323 if (tail != ctx->cached_cq_tail ||
1324 wq_list_empty(&ctx->iopoll_list))
1327 ret = io_do_iopoll(ctx, !min);
1332 } while (nr_events < min && !need_resched());
1337 void io_req_task_complete(struct io_kiocb *req, bool *locked)
1339 if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) {
1340 unsigned issue_flags = *locked ? 0 : IO_URING_F_UNLOCKED;
1342 req->cqe.flags |= io_put_kbuf(req, issue_flags);
1346 io_req_add_compl_list(req);
1348 io_req_complete_post(req);
1352 * After the iocb has been issued, it's safe to be found on the poll list.
1353 * Adding the kiocb to the list AFTER submission ensures that we don't
1354 * find it from a io_do_iopoll() thread before the issuer is done
1355 * accessing the kiocb cookie.
1357 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1359 struct io_ring_ctx *ctx = req->ctx;
1360 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1362 /* workqueue context doesn't hold uring_lock, grab it now */
1363 if (unlikely(needs_lock))
1364 mutex_lock(&ctx->uring_lock);
1367 * Track whether we have multiple files in our lists. This will impact
1368 * how we do polling eventually, not spinning if we're on potentially
1369 * different devices.
1371 if (wq_list_empty(&ctx->iopoll_list)) {
1372 ctx->poll_multi_queue = false;
1373 } else if (!ctx->poll_multi_queue) {
1374 struct io_kiocb *list_req;
1376 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1378 if (list_req->file != req->file)
1379 ctx->poll_multi_queue = true;
1383 * For fast devices, IO may have already completed. If it has, add
1384 * it to the front so we find it first.
1386 if (READ_ONCE(req->iopoll_completed))
1387 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1389 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1391 if (unlikely(needs_lock)) {
1393 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1394 * in sq thread task context or in io worker task context. If
1395 * current task context is sq thread, we don't need to check
1396 * whether should wake up sq thread.
1398 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1399 wq_has_sleeper(&ctx->sq_data->wait))
1400 wake_up(&ctx->sq_data->wait);
1402 mutex_unlock(&ctx->uring_lock);
1406 static bool io_bdev_nowait(struct block_device *bdev)
1408 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
1412 * If we tracked the file through the SCM inflight mechanism, we could support
1413 * any file. For now, just ensure that anything potentially problematic is done
1416 static bool __io_file_supports_nowait(struct file *file, umode_t mode)
1418 if (S_ISBLK(mode)) {
1419 if (IS_ENABLED(CONFIG_BLOCK) &&
1420 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
1426 if (S_ISREG(mode)) {
1427 if (IS_ENABLED(CONFIG_BLOCK) &&
1428 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
1429 !io_is_uring_fops(file))
1434 /* any ->read/write should understand O_NONBLOCK */
1435 if (file->f_flags & O_NONBLOCK)
1437 return file->f_mode & FMODE_NOWAIT;
1441 * If we tracked the file through the SCM inflight mechanism, we could support
1442 * any file. For now, just ensure that anything potentially problematic is done
1445 unsigned int io_file_get_flags(struct file *file)
1447 umode_t mode = file_inode(file)->i_mode;
1448 unsigned int res = 0;
1452 if (__io_file_supports_nowait(file, mode))
1454 if (io_file_need_scm(file))
1459 bool io_alloc_async_data(struct io_kiocb *req)
1461 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
1462 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
1463 if (req->async_data) {
1464 req->flags |= REQ_F_ASYNC_DATA;
1470 int io_req_prep_async(struct io_kiocb *req)
1472 const struct io_op_def *def = &io_op_defs[req->opcode];
1474 /* assign early for deferred execution for non-fixed file */
1475 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE))
1476 req->file = io_file_get_normal(req, req->cqe.fd);
1477 if (!def->prep_async)
1479 if (WARN_ON_ONCE(req_has_async_data(req)))
1481 if (io_alloc_async_data(req))
1484 return def->prep_async(req);
1487 static u32 io_get_sequence(struct io_kiocb *req)
1489 u32 seq = req->ctx->cached_sq_head;
1490 struct io_kiocb *cur;
1492 /* need original cached_sq_head, but it was increased for each req */
1493 io_for_each_link(cur, req)
1498 static __cold void io_drain_req(struct io_kiocb *req)
1500 struct io_ring_ctx *ctx = req->ctx;
1501 struct io_defer_entry *de;
1503 u32 seq = io_get_sequence(req);
1505 /* Still need defer if there is pending req in defer list. */
1506 spin_lock(&ctx->completion_lock);
1507 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1508 spin_unlock(&ctx->completion_lock);
1510 ctx->drain_active = false;
1511 io_req_task_queue(req);
1514 spin_unlock(&ctx->completion_lock);
1516 ret = io_req_prep_async(req);
1519 io_req_complete_failed(req, ret);
1522 io_prep_async_link(req);
1523 de = kmalloc(sizeof(*de), GFP_KERNEL);
1529 spin_lock(&ctx->completion_lock);
1530 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1531 spin_unlock(&ctx->completion_lock);
1536 trace_io_uring_defer(ctx, req, req->cqe.user_data, req->opcode);
1539 list_add_tail(&de->list, &ctx->defer_list);
1540 spin_unlock(&ctx->completion_lock);
1543 static void io_clean_op(struct io_kiocb *req)
1545 if (req->flags & REQ_F_BUFFER_SELECTED) {
1546 spin_lock(&req->ctx->completion_lock);
1547 io_put_kbuf_comp(req);
1548 spin_unlock(&req->ctx->completion_lock);
1551 if (req->flags & REQ_F_NEED_CLEANUP) {
1552 const struct io_op_def *def = &io_op_defs[req->opcode];
1557 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1558 kfree(req->apoll->double_poll);
1562 if (req->flags & REQ_F_INFLIGHT) {
1563 struct io_uring_task *tctx = req->task->io_uring;
1565 atomic_dec(&tctx->inflight_tracked);
1567 if (req->flags & REQ_F_CREDS)
1568 put_cred(req->creds);
1569 if (req->flags & REQ_F_ASYNC_DATA) {
1570 kfree(req->async_data);
1571 req->async_data = NULL;
1573 req->flags &= ~IO_REQ_CLEAN_FLAGS;
1576 static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags)
1578 if (req->file || !io_op_defs[req->opcode].needs_file)
1581 if (req->flags & REQ_F_FIXED_FILE)
1582 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1584 req->file = io_file_get_normal(req, req->cqe.fd);
1589 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1591 const struct io_op_def *def = &io_op_defs[req->opcode];
1592 const struct cred *creds = NULL;
1595 if (unlikely(!io_assign_file(req, issue_flags)))
1598 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1599 creds = override_creds(req->creds);
1601 if (!def->audit_skip)
1602 audit_uring_entry(req->opcode);
1604 ret = def->issue(req, issue_flags);
1606 if (!def->audit_skip)
1607 audit_uring_exit(!ret, ret);
1610 revert_creds(creds);
1612 if (ret == IOU_OK) {
1613 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1614 io_req_add_compl_list(req);
1616 io_req_complete_post(req);
1617 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1620 /* If the op doesn't have a file, we're not polling for it */
1621 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && req->file)
1622 io_iopoll_req_issued(req, issue_flags);
1627 int io_poll_issue(struct io_kiocb *req, bool *locked)
1629 io_tw_lock(req->ctx, locked);
1630 if (unlikely(req->task->flags & PF_EXITING))
1632 return io_issue_sqe(req, IO_URING_F_NONBLOCK);
1635 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1637 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1639 req = io_put_req_find_next(req);
1640 return req ? &req->work : NULL;
1643 void io_wq_submit_work(struct io_wq_work *work)
1645 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1646 const struct io_op_def *def = &io_op_defs[req->opcode];
1647 unsigned int issue_flags = IO_URING_F_UNLOCKED;
1648 bool needs_poll = false;
1649 int ret = 0, err = -ECANCELED;
1651 /* one will be dropped by ->io_free_work() after returning to io-wq */
1652 if (!(req->flags & REQ_F_REFCOUNT))
1653 __io_req_set_refcount(req, 2);
1657 io_arm_ltimeout(req);
1659 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1660 if (work->flags & IO_WQ_WORK_CANCEL) {
1662 io_req_task_queue_fail(req, err);
1665 if (!io_assign_file(req, issue_flags)) {
1667 work->flags |= IO_WQ_WORK_CANCEL;
1671 if (req->flags & REQ_F_FORCE_ASYNC) {
1672 bool opcode_poll = def->pollin || def->pollout;
1674 if (opcode_poll && file_can_poll(req->file)) {
1676 issue_flags |= IO_URING_F_NONBLOCK;
1681 ret = io_issue_sqe(req, issue_flags);
1685 * We can get EAGAIN for iopolled IO even though we're
1686 * forcing a sync submission from here, since we can't
1687 * wait for request slots on the block side.
1690 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1696 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1698 /* aborted or ready, in either case retry blocking */
1700 issue_flags &= ~IO_URING_F_NONBLOCK;
1703 /* avoid locking problems by failing it from a clean context */
1705 io_req_task_queue_fail(req, ret);
1708 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1709 unsigned int issue_flags)
1711 struct io_ring_ctx *ctx = req->ctx;
1712 struct file *file = NULL;
1713 unsigned long file_ptr;
1715 io_ring_submit_lock(ctx, issue_flags);
1717 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
1719 fd = array_index_nospec(fd, ctx->nr_user_files);
1720 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
1721 file = (struct file *) (file_ptr & FFS_MASK);
1722 file_ptr &= ~FFS_MASK;
1723 /* mask in overlapping REQ_F and FFS bits */
1724 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
1725 io_req_set_rsrc_node(req, ctx, 0);
1726 WARN_ON_ONCE(file && !test_bit(fd, ctx->file_table.bitmap));
1728 io_ring_submit_unlock(ctx, issue_flags);
1732 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
1734 struct file *file = fget(fd);
1736 trace_io_uring_file_get(req->ctx, req, req->cqe.user_data, fd);
1738 /* we don't allow fixed io_uring files */
1739 if (file && io_is_uring_fops(file))
1740 io_req_track_inflight(req);
1744 static void io_queue_async(struct io_kiocb *req, int ret)
1745 __must_hold(&req->ctx->uring_lock)
1747 struct io_kiocb *linked_timeout;
1749 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
1750 io_req_complete_failed(req, ret);
1754 linked_timeout = io_prep_linked_timeout(req);
1756 switch (io_arm_poll_handler(req, 0)) {
1757 case IO_APOLL_READY:
1758 io_req_task_queue(req);
1760 case IO_APOLL_ABORTED:
1762 * Queued up for async execution, worker will release
1763 * submit reference when the iocb is actually submitted.
1765 io_kbuf_recycle(req, 0);
1766 io_queue_iowq(req, NULL);
1773 io_queue_linked_timeout(linked_timeout);
1776 static inline void io_queue_sqe(struct io_kiocb *req)
1777 __must_hold(&req->ctx->uring_lock)
1781 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
1784 * We async punt it if the file wasn't marked NOWAIT, or if the file
1785 * doesn't support non-blocking read/write attempts
1788 io_arm_ltimeout(req);
1790 io_queue_async(req, ret);
1793 static void io_queue_sqe_fallback(struct io_kiocb *req)
1794 __must_hold(&req->ctx->uring_lock)
1796 if (unlikely(req->flags & REQ_F_FAIL)) {
1798 * We don't submit, fail them all, for that replace hardlinks
1799 * with normal links. Extra REQ_F_LINK is tolerated.
1801 req->flags &= ~REQ_F_HARDLINK;
1802 req->flags |= REQ_F_LINK;
1803 io_req_complete_failed(req, req->cqe.res);
1804 } else if (unlikely(req->ctx->drain_active)) {
1807 int ret = io_req_prep_async(req);
1810 io_req_complete_failed(req, ret);
1812 io_queue_iowq(req, NULL);
1817 * Check SQE restrictions (opcode and flags).
1819 * Returns 'true' if SQE is allowed, 'false' otherwise.
1821 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
1822 struct io_kiocb *req,
1823 unsigned int sqe_flags)
1825 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
1828 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
1829 ctx->restrictions.sqe_flags_required)
1832 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
1833 ctx->restrictions.sqe_flags_required))
1839 static void io_init_req_drain(struct io_kiocb *req)
1841 struct io_ring_ctx *ctx = req->ctx;
1842 struct io_kiocb *head = ctx->submit_state.link.head;
1844 ctx->drain_active = true;
1847 * If we need to drain a request in the middle of a link, drain
1848 * the head request and the next request/link after the current
1849 * link. Considering sequential execution of links,
1850 * REQ_F_IO_DRAIN will be maintained for every request of our
1853 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
1854 ctx->drain_next = true;
1858 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
1859 const struct io_uring_sqe *sqe)
1860 __must_hold(&ctx->uring_lock)
1862 const struct io_op_def *def;
1863 unsigned int sqe_flags;
1867 /* req is partially pre-initialised, see io_preinit_req() */
1868 req->opcode = opcode = READ_ONCE(sqe->opcode);
1869 /* same numerical values with corresponding REQ_F_*, safe to copy */
1870 req->flags = sqe_flags = READ_ONCE(sqe->flags);
1871 req->cqe.user_data = READ_ONCE(sqe->user_data);
1873 req->rsrc_node = NULL;
1874 req->task = current;
1876 if (unlikely(opcode >= IORING_OP_LAST)) {
1880 def = &io_op_defs[opcode];
1881 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
1882 /* enforce forwards compatibility on users */
1883 if (sqe_flags & ~SQE_VALID_FLAGS)
1885 if (sqe_flags & IOSQE_BUFFER_SELECT) {
1886 if (!def->buffer_select)
1888 req->buf_index = READ_ONCE(sqe->buf_group);
1890 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
1891 ctx->drain_disabled = true;
1892 if (sqe_flags & IOSQE_IO_DRAIN) {
1893 if (ctx->drain_disabled)
1895 io_init_req_drain(req);
1898 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
1899 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
1901 /* knock it to the slow queue path, will be drained there */
1902 if (ctx->drain_active)
1903 req->flags |= REQ_F_FORCE_ASYNC;
1904 /* if there is no link, we're at "next" request and need to drain */
1905 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
1906 ctx->drain_next = false;
1907 ctx->drain_active = true;
1908 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
1912 if (!def->ioprio && sqe->ioprio)
1914 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
1917 if (def->needs_file) {
1918 struct io_submit_state *state = &ctx->submit_state;
1920 req->cqe.fd = READ_ONCE(sqe->fd);
1923 * Plug now if we have more than 2 IO left after this, and the
1924 * target is potentially a read/write to block based storage.
1926 if (state->need_plug && def->plug) {
1927 state->plug_started = true;
1928 state->need_plug = false;
1929 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
1933 personality = READ_ONCE(sqe->personality);
1937 req->creds = xa_load(&ctx->personalities, personality);
1940 get_cred(req->creds);
1941 ret = security_uring_override_creds(req->creds);
1943 put_cred(req->creds);
1946 req->flags |= REQ_F_CREDS;
1949 return def->prep(req, sqe);
1952 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
1953 struct io_kiocb *req, int ret)
1955 struct io_ring_ctx *ctx = req->ctx;
1956 struct io_submit_link *link = &ctx->submit_state.link;
1957 struct io_kiocb *head = link->head;
1959 trace_io_uring_req_failed(sqe, ctx, req, ret);
1962 * Avoid breaking links in the middle as it renders links with SQPOLL
1963 * unusable. Instead of failing eagerly, continue assembling the link if
1964 * applicable and mark the head with REQ_F_FAIL. The link flushing code
1965 * should find the flag and handle the rest.
1967 req_fail_link_node(req, ret);
1968 if (head && !(head->flags & REQ_F_FAIL))
1969 req_fail_link_node(head, -ECANCELED);
1971 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
1973 link->last->link = req;
1977 io_queue_sqe_fallback(req);
1982 link->last->link = req;
1989 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
1990 const struct io_uring_sqe *sqe)
1991 __must_hold(&ctx->uring_lock)
1993 struct io_submit_link *link = &ctx->submit_state.link;
1996 ret = io_init_req(ctx, req, sqe);
1998 return io_submit_fail_init(sqe, req, ret);
2000 /* don't need @sqe from now on */
2001 trace_io_uring_submit_sqe(ctx, req, req->cqe.user_data, req->opcode,
2003 ctx->flags & IORING_SETUP_SQPOLL);
2006 * If we already have a head request, queue this one for async
2007 * submittal once the head completes. If we don't have a head but
2008 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2009 * submitted sync once the chain is complete. If none of those
2010 * conditions are true (normal request), then just queue it.
2012 if (unlikely(link->head)) {
2013 ret = io_req_prep_async(req);
2015 return io_submit_fail_init(sqe, req, ret);
2017 trace_io_uring_link(ctx, req, link->head);
2018 link->last->link = req;
2021 if (req->flags & IO_REQ_LINK_FLAGS)
2023 /* last request of the link, flush it */
2026 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2029 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2030 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2031 if (req->flags & IO_REQ_LINK_FLAGS) {
2036 io_queue_sqe_fallback(req);
2046 * Batched submission is done, ensure local IO is flushed out.
2048 static void io_submit_state_end(struct io_ring_ctx *ctx)
2050 struct io_submit_state *state = &ctx->submit_state;
2052 if (unlikely(state->link.head))
2053 io_queue_sqe_fallback(state->link.head);
2054 /* flush only after queuing links as they can generate completions */
2055 io_submit_flush_completions(ctx);
2056 if (state->plug_started)
2057 blk_finish_plug(&state->plug);
2061 * Start submission side cache.
2063 static void io_submit_state_start(struct io_submit_state *state,
2064 unsigned int max_ios)
2066 state->plug_started = false;
2067 state->need_plug = max_ios > 2;
2068 state->submit_nr = max_ios;
2069 /* set only head, no need to init link_last in advance */
2070 state->link.head = NULL;
2073 static void io_commit_sqring(struct io_ring_ctx *ctx)
2075 struct io_rings *rings = ctx->rings;
2078 * Ensure any loads from the SQEs are done at this point,
2079 * since once we write the new head, the application could
2080 * write new data to them.
2082 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2086 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2087 * that is mapped by userspace. This means that care needs to be taken to
2088 * ensure that reads are stable, as we cannot rely on userspace always
2089 * being a good citizen. If members of the sqe are validated and then later
2090 * used, it's important that those reads are done through READ_ONCE() to
2091 * prevent a re-load down the line.
2093 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
2095 unsigned head, mask = ctx->sq_entries - 1;
2096 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2099 * The cached sq head (or cq tail) serves two purposes:
2101 * 1) allows us to batch the cost of updating the user visible
2103 * 2) allows the kernel side to track the head on its own, even
2104 * though the application is the one updating it.
2106 head = READ_ONCE(ctx->sq_array[sq_idx]);
2107 if (likely(head < ctx->sq_entries)) {
2108 /* double index for 128-byte SQEs, twice as long */
2109 if (ctx->flags & IORING_SETUP_SQE128)
2111 return &ctx->sq_sqes[head];
2114 /* drop invalid entries */
2116 WRITE_ONCE(ctx->rings->sq_dropped,
2117 READ_ONCE(ctx->rings->sq_dropped) + 1);
2121 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2122 __must_hold(&ctx->uring_lock)
2124 unsigned int entries = io_sqring_entries(ctx);
2128 if (unlikely(!entries))
2130 /* make sure SQ entry isn't read before tail */
2131 ret = left = min3(nr, ctx->sq_entries, entries);
2132 io_get_task_refs(left);
2133 io_submit_state_start(&ctx->submit_state, left);
2136 const struct io_uring_sqe *sqe;
2137 struct io_kiocb *req;
2139 if (unlikely(!io_alloc_req_refill(ctx)))
2141 req = io_alloc_req(ctx);
2142 sqe = io_get_sqe(ctx);
2143 if (unlikely(!sqe)) {
2144 io_req_add_to_cache(req, ctx);
2149 * Continue submitting even for sqe failure if the
2150 * ring was setup with IORING_SETUP_SUBMIT_ALL
2152 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2153 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2159 if (unlikely(left)) {
2161 /* try again if it submitted nothing and can't allocate a req */
2162 if (!ret && io_req_cache_empty(ctx))
2164 current->io_uring->cached_refs += left;
2167 io_submit_state_end(ctx);
2168 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2169 io_commit_sqring(ctx);
2173 struct io_wait_queue {
2174 struct wait_queue_entry wq;
2175 struct io_ring_ctx *ctx;
2177 unsigned nr_timeouts;
2180 static inline bool io_should_wake(struct io_wait_queue *iowq)
2182 struct io_ring_ctx *ctx = iowq->ctx;
2183 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
2186 * Wake up if we have enough events, or if a timeout occurred since we
2187 * started waiting. For timeouts, we always want to return to userspace,
2188 * regardless of event count.
2190 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2193 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2194 int wake_flags, void *key)
2196 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
2200 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2201 * the task, and the next invocation will do it.
2203 if (io_should_wake(iowq) ||
2204 test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &iowq->ctx->check_cq))
2205 return autoremove_wake_function(curr, mode, wake_flags, key);
2209 int io_run_task_work_sig(void)
2211 if (io_run_task_work())
2213 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
2214 return -ERESTARTSYS;
2215 if (task_sigpending(current))
2220 /* when returns >0, the caller should retry */
2221 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2222 struct io_wait_queue *iowq,
2226 unsigned long check_cq;
2228 /* make sure we run task_work before checking for signals */
2229 ret = io_run_task_work_sig();
2230 if (ret || io_should_wake(iowq))
2233 check_cq = READ_ONCE(ctx->check_cq);
2234 if (unlikely(check_cq)) {
2235 /* let the caller flush overflows, retry */
2236 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2238 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
2241 if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS))
2247 * Wait until events become available, if we don't already have some. The
2248 * application must reap them itself, as they reside on the shared cq ring.
2250 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2251 const sigset_t __user *sig, size_t sigsz,
2252 struct __kernel_timespec __user *uts)
2254 struct io_wait_queue iowq;
2255 struct io_rings *rings = ctx->rings;
2256 ktime_t timeout = KTIME_MAX;
2260 io_cqring_overflow_flush(ctx);
2261 if (io_cqring_events(ctx) >= min_events)
2263 if (!io_run_task_work())
2268 #ifdef CONFIG_COMPAT
2269 if (in_compat_syscall())
2270 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2274 ret = set_user_sigmask(sig, sigsz);
2281 struct timespec64 ts;
2283 if (get_timespec64(&ts, uts))
2285 timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2288 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2289 iowq.wq.private = current;
2290 INIT_LIST_HEAD(&iowq.wq.entry);
2292 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2293 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2295 trace_io_uring_cqring_wait(ctx, min_events);
2297 /* if we can't even flush overflow, don't wait for more */
2298 if (!io_cqring_overflow_flush(ctx)) {
2302 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2303 TASK_INTERRUPTIBLE);
2304 ret = io_cqring_wait_schedule(ctx, &iowq, timeout);
2308 finish_wait(&ctx->cq_wait, &iowq.wq);
2309 restore_saved_sigmask_unless(ret == -EINTR);
2311 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2314 static void io_mem_free(void *ptr)
2321 page = virt_to_head_page(ptr);
2322 if (put_page_testzero(page))
2323 free_compound_page(page);
2326 static void *io_mem_alloc(size_t size)
2328 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2330 return (void *) __get_free_pages(gfp, get_order(size));
2333 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2334 unsigned int cq_entries, size_t *sq_offset)
2336 struct io_rings *rings;
2337 size_t off, sq_array_size;
2339 off = struct_size(rings, cqes, cq_entries);
2340 if (off == SIZE_MAX)
2342 if (ctx->flags & IORING_SETUP_CQE32) {
2343 if (check_shl_overflow(off, 1, &off))
2348 off = ALIGN(off, SMP_CACHE_BYTES);
2356 sq_array_size = array_size(sizeof(u32), sq_entries);
2357 if (sq_array_size == SIZE_MAX)
2360 if (check_add_overflow(off, sq_array_size, &off))
2366 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2367 unsigned int eventfd_async)
2369 struct io_ev_fd *ev_fd;
2370 __s32 __user *fds = arg;
2373 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2374 lockdep_is_held(&ctx->uring_lock));
2378 if (copy_from_user(&fd, fds, sizeof(*fds)))
2381 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2385 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2386 if (IS_ERR(ev_fd->cq_ev_fd)) {
2387 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2391 ev_fd->eventfd_async = eventfd_async;
2392 ctx->has_evfd = true;
2393 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2397 static void io_eventfd_put(struct rcu_head *rcu)
2399 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
2401 eventfd_ctx_put(ev_fd->cq_ev_fd);
2405 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2407 struct io_ev_fd *ev_fd;
2409 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2410 lockdep_is_held(&ctx->uring_lock));
2412 ctx->has_evfd = false;
2413 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2414 call_rcu(&ev_fd->rcu, io_eventfd_put);
2421 static void io_req_caches_free(struct io_ring_ctx *ctx)
2423 struct io_submit_state *state = &ctx->submit_state;
2426 mutex_lock(&ctx->uring_lock);
2427 io_flush_cached_locked_reqs(ctx, state);
2429 while (!io_req_cache_empty(ctx)) {
2430 struct io_wq_work_node *node;
2431 struct io_kiocb *req;
2433 node = wq_stack_extract(&state->free_list);
2434 req = container_of(node, struct io_kiocb, comp_list);
2435 kmem_cache_free(req_cachep, req);
2439 percpu_ref_put_many(&ctx->refs, nr);
2440 mutex_unlock(&ctx->uring_lock);
2443 static void io_flush_apoll_cache(struct io_ring_ctx *ctx)
2445 struct async_poll *apoll;
2447 while (!list_empty(&ctx->apoll_cache)) {
2448 apoll = list_first_entry(&ctx->apoll_cache, struct async_poll,
2450 list_del(&apoll->poll.wait.entry);
2455 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2457 io_sq_thread_finish(ctx);
2459 if (ctx->mm_account) {
2460 mmdrop(ctx->mm_account);
2461 ctx->mm_account = NULL;
2464 io_rsrc_refs_drop(ctx);
2465 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2466 io_wait_rsrc_data(ctx->buf_data);
2467 io_wait_rsrc_data(ctx->file_data);
2469 mutex_lock(&ctx->uring_lock);
2471 __io_sqe_buffers_unregister(ctx);
2473 __io_sqe_files_unregister(ctx);
2475 __io_cqring_overflow_flush(ctx, true);
2476 io_eventfd_unregister(ctx);
2477 io_flush_apoll_cache(ctx);
2478 mutex_unlock(&ctx->uring_lock);
2479 io_destroy_buffers(ctx);
2481 put_cred(ctx->sq_creds);
2482 if (ctx->submitter_task)
2483 put_task_struct(ctx->submitter_task);
2485 /* there are no registered resources left, nobody uses it */
2487 io_rsrc_node_destroy(ctx->rsrc_node);
2488 if (ctx->rsrc_backup_node)
2489 io_rsrc_node_destroy(ctx->rsrc_backup_node);
2490 flush_delayed_work(&ctx->rsrc_put_work);
2491 flush_delayed_work(&ctx->fallback_work);
2493 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2494 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
2496 #if defined(CONFIG_UNIX)
2497 if (ctx->ring_sock) {
2498 ctx->ring_sock->file = NULL; /* so that iput() is called */
2499 sock_release(ctx->ring_sock);
2502 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2504 io_mem_free(ctx->rings);
2505 io_mem_free(ctx->sq_sqes);
2507 percpu_ref_exit(&ctx->refs);
2508 free_uid(ctx->user);
2509 io_req_caches_free(ctx);
2511 io_wq_put_hash(ctx->hash_map);
2512 kfree(ctx->cancel_table.hbs);
2513 kfree(ctx->cancel_table_locked.hbs);
2514 kfree(ctx->dummy_ubuf);
2516 xa_destroy(&ctx->io_bl_xa);
2520 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2522 struct io_ring_ctx *ctx = file->private_data;
2525 poll_wait(file, &ctx->cq_wait, wait);
2527 * synchronizes with barrier from wq_has_sleeper call in
2531 if (!io_sqring_full(ctx))
2532 mask |= EPOLLOUT | EPOLLWRNORM;
2535 * Don't flush cqring overflow list here, just do a simple check.
2536 * Otherwise there could possible be ABBA deadlock:
2539 * lock(&ctx->uring_lock);
2541 * lock(&ctx->uring_lock);
2544 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2545 * pushs them to do the flush.
2547 if (io_cqring_events(ctx) ||
2548 test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
2549 mask |= EPOLLIN | EPOLLRDNORM;
2554 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2556 const struct cred *creds;
2558 creds = xa_erase(&ctx->personalities, id);
2567 struct io_tctx_exit {
2568 struct callback_head task_work;
2569 struct completion completion;
2570 struct io_ring_ctx *ctx;
2573 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2575 struct io_uring_task *tctx = current->io_uring;
2576 struct io_tctx_exit *work;
2578 work = container_of(cb, struct io_tctx_exit, task_work);
2580 * When @in_idle, we're in cancellation and it's racy to remove the
2581 * node. It'll be removed by the end of cancellation, just ignore it.
2583 if (!atomic_read(&tctx->in_idle))
2584 io_uring_del_tctx_node((unsigned long)work->ctx);
2585 complete(&work->completion);
2588 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2590 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2592 return req->ctx == data;
2595 static __cold void io_ring_exit_work(struct work_struct *work)
2597 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2598 unsigned long timeout = jiffies + HZ * 60 * 5;
2599 unsigned long interval = HZ / 20;
2600 struct io_tctx_exit exit;
2601 struct io_tctx_node *node;
2605 * If we're doing polled IO and end up having requests being
2606 * submitted async (out-of-line), then completions can come in while
2607 * we're waiting for refs to drop. We need to reap these manually,
2608 * as nobody else will be looking for them.
2611 io_uring_try_cancel_requests(ctx, NULL, true);
2613 struct io_sq_data *sqd = ctx->sq_data;
2614 struct task_struct *tsk;
2616 io_sq_thread_park(sqd);
2618 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
2619 io_wq_cancel_cb(tsk->io_uring->io_wq,
2620 io_cancel_ctx_cb, ctx, true);
2621 io_sq_thread_unpark(sqd);
2624 io_req_caches_free(ctx);
2626 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
2627 /* there is little hope left, don't run it too often */
2630 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
2632 init_completion(&exit.completion);
2633 init_task_work(&exit.task_work, io_tctx_exit_cb);
2636 * Some may use context even when all refs and requests have been put,
2637 * and they are free to do so while still holding uring_lock or
2638 * completion_lock, see io_req_task_submit(). Apart from other work,
2639 * this lock/unlock section also waits them to finish.
2641 mutex_lock(&ctx->uring_lock);
2642 while (!list_empty(&ctx->tctx_list)) {
2643 WARN_ON_ONCE(time_after(jiffies, timeout));
2645 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
2647 /* don't spin on a single task if cancellation failed */
2648 list_rotate_left(&ctx->tctx_list);
2649 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
2650 if (WARN_ON_ONCE(ret))
2653 mutex_unlock(&ctx->uring_lock);
2654 wait_for_completion(&exit.completion);
2655 mutex_lock(&ctx->uring_lock);
2657 mutex_unlock(&ctx->uring_lock);
2658 spin_lock(&ctx->completion_lock);
2659 spin_unlock(&ctx->completion_lock);
2661 io_ring_ctx_free(ctx);
2664 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
2666 unsigned long index;
2667 struct creds *creds;
2669 mutex_lock(&ctx->uring_lock);
2670 percpu_ref_kill(&ctx->refs);
2672 __io_cqring_overflow_flush(ctx, true);
2673 xa_for_each(&ctx->personalities, index, creds)
2674 io_unregister_personality(ctx, index);
2676 io_poll_remove_all(ctx, NULL, true);
2677 mutex_unlock(&ctx->uring_lock);
2679 /* failed during ring init, it couldn't have issued any requests */
2681 io_kill_timeouts(ctx, NULL, true);
2682 /* if we failed setting up the ctx, we might not have any rings */
2683 io_iopoll_try_reap_events(ctx);
2686 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
2688 * Use system_unbound_wq to avoid spawning tons of event kworkers
2689 * if we're exiting a ton of rings at the same time. It just adds
2690 * noise and overhead, there's no discernable change in runtime
2691 * over using system_wq.
2693 queue_work(system_unbound_wq, &ctx->exit_work);
2696 static int io_uring_release(struct inode *inode, struct file *file)
2698 struct io_ring_ctx *ctx = file->private_data;
2700 file->private_data = NULL;
2701 io_ring_ctx_wait_and_kill(ctx);
2705 struct io_task_cancel {
2706 struct task_struct *task;
2710 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
2712 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2713 struct io_task_cancel *cancel = data;
2715 return io_match_task_safe(req, cancel->task, cancel->all);
2718 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
2719 struct task_struct *task,
2722 struct io_defer_entry *de;
2725 spin_lock(&ctx->completion_lock);
2726 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
2727 if (io_match_task_safe(de->req, task, cancel_all)) {
2728 list_cut_position(&list, &ctx->defer_list, &de->list);
2732 spin_unlock(&ctx->completion_lock);
2733 if (list_empty(&list))
2736 while (!list_empty(&list)) {
2737 de = list_first_entry(&list, struct io_defer_entry, list);
2738 list_del_init(&de->list);
2739 io_req_complete_failed(de->req, -ECANCELED);
2745 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
2747 struct io_tctx_node *node;
2748 enum io_wq_cancel cret;
2751 mutex_lock(&ctx->uring_lock);
2752 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
2753 struct io_uring_task *tctx = node->task->io_uring;
2756 * io_wq will stay alive while we hold uring_lock, because it's
2757 * killed after ctx nodes, which requires to take the lock.
2759 if (!tctx || !tctx->io_wq)
2761 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
2762 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2764 mutex_unlock(&ctx->uring_lock);
2769 static __cold void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
2770 struct task_struct *task,
2773 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
2774 struct io_uring_task *tctx = task ? task->io_uring : NULL;
2776 /* failed during ring init, it couldn't have issued any requests */
2781 enum io_wq_cancel cret;
2785 ret |= io_uring_try_cancel_iowq(ctx);
2786 } else if (tctx && tctx->io_wq) {
2788 * Cancels requests of all rings, not only @ctx, but
2789 * it's fine as the task is in exit/exec.
2791 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
2793 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2796 /* SQPOLL thread does its own polling */
2797 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
2798 (ctx->sq_data && ctx->sq_data->thread == current)) {
2799 while (!wq_list_empty(&ctx->iopoll_list)) {
2800 io_iopoll_try_reap_events(ctx);
2805 ret |= io_cancel_defer_files(ctx, task, cancel_all);
2806 mutex_lock(&ctx->uring_lock);
2807 ret |= io_poll_remove_all(ctx, task, cancel_all);
2808 mutex_unlock(&ctx->uring_lock);
2809 ret |= io_kill_timeouts(ctx, task, cancel_all);
2811 ret |= io_run_task_work();
2818 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
2821 return atomic_read(&tctx->inflight_tracked);
2822 return percpu_counter_sum(&tctx->inflight);
2826 * Find any io_uring ctx that this task has registered or done IO on, and cancel
2827 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
2829 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
2831 struct io_uring_task *tctx = current->io_uring;
2832 struct io_ring_ctx *ctx;
2836 WARN_ON_ONCE(sqd && sqd->thread != current);
2838 if (!current->io_uring)
2841 io_wq_exit_start(tctx->io_wq);
2843 atomic_inc(&tctx->in_idle);
2845 io_uring_drop_tctx_refs(current);
2846 /* read completions before cancelations */
2847 inflight = tctx_inflight(tctx, !cancel_all);
2852 struct io_tctx_node *node;
2853 unsigned long index;
2855 xa_for_each(&tctx->xa, index, node) {
2856 /* sqpoll task will cancel all its requests */
2857 if (node->ctx->sq_data)
2859 io_uring_try_cancel_requests(node->ctx, current,
2863 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
2864 io_uring_try_cancel_requests(ctx, current,
2868 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
2870 io_uring_drop_tctx_refs(current);
2873 * If we've seen completions, retry without waiting. This
2874 * avoids a race where a completion comes in before we did
2875 * prepare_to_wait().
2877 if (inflight == tctx_inflight(tctx, !cancel_all))
2879 finish_wait(&tctx->wait, &wait);
2882 io_uring_clean_tctx(tctx);
2885 * We shouldn't run task_works after cancel, so just leave
2886 * ->in_idle set for normal exit.
2888 atomic_dec(&tctx->in_idle);
2889 /* for exec all current's requests should be gone, kill tctx */
2890 __io_uring_free(current);
2894 void __io_uring_cancel(bool cancel_all)
2896 io_uring_cancel_generic(cancel_all, NULL);
2899 static void *io_uring_validate_mmap_request(struct file *file,
2900 loff_t pgoff, size_t sz)
2902 struct io_ring_ctx *ctx = file->private_data;
2903 loff_t offset = pgoff << PAGE_SHIFT;
2908 case IORING_OFF_SQ_RING:
2909 case IORING_OFF_CQ_RING:
2912 case IORING_OFF_SQES:
2916 return ERR_PTR(-EINVAL);
2919 page = virt_to_head_page(ptr);
2920 if (sz > page_size(page))
2921 return ERR_PTR(-EINVAL);
2928 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
2930 size_t sz = vma->vm_end - vma->vm_start;
2934 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
2936 return PTR_ERR(ptr);
2938 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
2939 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
2942 #else /* !CONFIG_MMU */
2944 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
2946 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
2949 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
2951 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
2954 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
2955 unsigned long addr, unsigned long len,
2956 unsigned long pgoff, unsigned long flags)
2960 ptr = io_uring_validate_mmap_request(file, pgoff, len);
2962 return PTR_ERR(ptr);
2964 return (unsigned long) ptr;
2967 #endif /* !CONFIG_MMU */
2969 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
2971 if (flags & IORING_ENTER_EXT_ARG) {
2972 struct io_uring_getevents_arg arg;
2974 if (argsz != sizeof(arg))
2976 if (copy_from_user(&arg, argp, sizeof(arg)))
2982 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
2983 struct __kernel_timespec __user **ts,
2984 const sigset_t __user **sig)
2986 struct io_uring_getevents_arg arg;
2989 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
2990 * is just a pointer to the sigset_t.
2992 if (!(flags & IORING_ENTER_EXT_ARG)) {
2993 *sig = (const sigset_t __user *) argp;
2999 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3000 * timespec and sigset_t pointers if good.
3002 if (*argsz != sizeof(arg))
3004 if (copy_from_user(&arg, argp, sizeof(arg)))
3008 *sig = u64_to_user_ptr(arg.sigmask);
3009 *argsz = arg.sigmask_sz;
3010 *ts = u64_to_user_ptr(arg.ts);
3014 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3015 u32, min_complete, u32, flags, const void __user *, argp,
3018 struct io_ring_ctx *ctx;
3024 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3025 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3026 IORING_ENTER_REGISTERED_RING)))
3030 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3031 * need only dereference our task private array to find it.
3033 if (flags & IORING_ENTER_REGISTERED_RING) {
3034 struct io_uring_task *tctx = current->io_uring;
3036 if (!tctx || fd >= IO_RINGFD_REG_MAX)
3038 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3039 f.file = tctx->registered_rings[fd];
3045 if (unlikely(!f.file))
3049 if (unlikely(!io_is_uring_fops(f.file)))
3053 ctx = f.file->private_data;
3054 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
3058 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3062 * For SQ polling, the thread will do all submissions and completions.
3063 * Just return the requested submit count, and wake the thread if
3067 if (ctx->flags & IORING_SETUP_SQPOLL) {
3068 io_cqring_overflow_flush(ctx);
3070 if (unlikely(ctx->sq_data->thread == NULL)) {
3074 if (flags & IORING_ENTER_SQ_WAKEUP)
3075 wake_up(&ctx->sq_data->wait);
3076 if (flags & IORING_ENTER_SQ_WAIT) {
3077 ret = io_sqpoll_wait_sq(ctx);
3082 } else if (to_submit) {
3083 ret = io_uring_add_tctx_node(ctx);
3087 mutex_lock(&ctx->uring_lock);
3088 ret = io_submit_sqes(ctx, to_submit);
3089 if (ret != to_submit) {
3090 mutex_unlock(&ctx->uring_lock);
3093 if ((flags & IORING_ENTER_GETEVENTS) && ctx->syscall_iopoll)
3095 mutex_unlock(&ctx->uring_lock);
3097 if (flags & IORING_ENTER_GETEVENTS) {
3099 if (ctx->syscall_iopoll) {
3101 * We disallow the app entering submit/complete with
3102 * polling, but we still need to lock the ring to
3103 * prevent racing with polled issue that got punted to
3106 mutex_lock(&ctx->uring_lock);
3108 ret2 = io_validate_ext_arg(flags, argp, argsz);
3109 if (likely(!ret2)) {
3110 min_complete = min(min_complete,
3112 ret2 = io_iopoll_check(ctx, min_complete);
3114 mutex_unlock(&ctx->uring_lock);
3116 const sigset_t __user *sig;
3117 struct __kernel_timespec __user *ts;
3119 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3120 if (likely(!ret2)) {
3121 min_complete = min(min_complete,
3123 ret2 = io_cqring_wait(ctx, min_complete, sig,
3132 * EBADR indicates that one or more CQE were dropped.
3133 * Once the user has been informed we can clear the bit
3134 * as they are obviously ok with those drops.
3136 if (unlikely(ret2 == -EBADR))
3137 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3143 percpu_ref_put(&ctx->refs);
3149 static const struct file_operations io_uring_fops = {
3150 .release = io_uring_release,
3151 .mmap = io_uring_mmap,
3153 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3154 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3156 .poll = io_uring_poll,
3157 #ifdef CONFIG_PROC_FS
3158 .show_fdinfo = io_uring_show_fdinfo,
3162 bool io_is_uring_fops(struct file *file)
3164 return file->f_op == &io_uring_fops;
3167 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3168 struct io_uring_params *p)
3170 struct io_rings *rings;
3171 size_t size, sq_array_offset;
3173 /* make sure these are sane, as we already accounted them */
3174 ctx->sq_entries = p->sq_entries;
3175 ctx->cq_entries = p->cq_entries;
3177 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3178 if (size == SIZE_MAX)
3181 rings = io_mem_alloc(size);
3186 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3187 rings->sq_ring_mask = p->sq_entries - 1;
3188 rings->cq_ring_mask = p->cq_entries - 1;
3189 rings->sq_ring_entries = p->sq_entries;
3190 rings->cq_ring_entries = p->cq_entries;
3192 if (p->flags & IORING_SETUP_SQE128)
3193 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3195 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3196 if (size == SIZE_MAX) {
3197 io_mem_free(ctx->rings);
3202 ctx->sq_sqes = io_mem_alloc(size);
3203 if (!ctx->sq_sqes) {
3204 io_mem_free(ctx->rings);
3212 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3216 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3220 ret = __io_uring_add_tctx_node(ctx, false);
3225 fd_install(fd, file);
3230 * Allocate an anonymous fd, this is what constitutes the application
3231 * visible backing of an io_uring instance. The application mmaps this
3232 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3233 * we have to tie this fd to a socket for file garbage collection purposes.
3235 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3238 #if defined(CONFIG_UNIX)
3241 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3244 return ERR_PTR(ret);
3247 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3248 O_RDWR | O_CLOEXEC, NULL);
3249 #if defined(CONFIG_UNIX)
3251 sock_release(ctx->ring_sock);
3252 ctx->ring_sock = NULL;
3254 ctx->ring_sock->file = file;
3260 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3261 struct io_uring_params __user *params)
3263 struct io_ring_ctx *ctx;
3269 if (entries > IORING_MAX_ENTRIES) {
3270 if (!(p->flags & IORING_SETUP_CLAMP))
3272 entries = IORING_MAX_ENTRIES;
3276 * Use twice as many entries for the CQ ring. It's possible for the
3277 * application to drive a higher depth than the size of the SQ ring,
3278 * since the sqes are only used at submission time. This allows for
3279 * some flexibility in overcommitting a bit. If the application has
3280 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3281 * of CQ ring entries manually.
3283 p->sq_entries = roundup_pow_of_two(entries);
3284 if (p->flags & IORING_SETUP_CQSIZE) {
3286 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3287 * to a power-of-two, if it isn't already. We do NOT impose
3288 * any cq vs sq ring sizing.
3292 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3293 if (!(p->flags & IORING_SETUP_CLAMP))
3295 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3297 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3298 if (p->cq_entries < p->sq_entries)
3301 p->cq_entries = 2 * p->sq_entries;
3304 ctx = io_ring_ctx_alloc(p);
3309 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3310 * space applications don't need to do io completion events
3311 * polling again, they can rely on io_sq_thread to do polling
3312 * work, which can reduce cpu usage and uring_lock contention.
3314 if (ctx->flags & IORING_SETUP_IOPOLL &&
3315 !(ctx->flags & IORING_SETUP_SQPOLL))
3316 ctx->syscall_iopoll = 1;
3318 ctx->compat = in_compat_syscall();
3319 if (!capable(CAP_IPC_LOCK))
3320 ctx->user = get_uid(current_user());
3323 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3324 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3327 if (ctx->flags & IORING_SETUP_SQPOLL) {
3328 /* IPI related flags don't make sense with SQPOLL */
3329 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3330 IORING_SETUP_TASKRUN_FLAG))
3332 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3333 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3334 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3336 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
3338 ctx->notify_method = TWA_SIGNAL;
3342 * This is just grabbed for accounting purposes. When a process exits,
3343 * the mm is exited and dropped before the files, hence we need to hang
3344 * on to this mm purely for the purposes of being able to unaccount
3345 * memory (locked/pinned vm). It's not used for anything else.
3347 mmgrab(current->mm);
3348 ctx->mm_account = current->mm;
3350 ret = io_allocate_scq_urings(ctx, p);
3354 ret = io_sq_offload_create(ctx, p);
3357 /* always set a rsrc node */
3358 ret = io_rsrc_node_switch_start(ctx);
3361 io_rsrc_node_switch(ctx, NULL);
3363 memset(&p->sq_off, 0, sizeof(p->sq_off));
3364 p->sq_off.head = offsetof(struct io_rings, sq.head);
3365 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3366 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3367 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3368 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3369 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3370 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3372 memset(&p->cq_off, 0, sizeof(p->cq_off));
3373 p->cq_off.head = offsetof(struct io_rings, cq.head);
3374 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3375 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3376 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3377 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3378 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3379 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3381 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3382 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3383 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3384 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3385 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3386 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3387 IORING_FEAT_LINKED_FILE;
3389 if (copy_to_user(params, p, sizeof(*p))) {
3394 file = io_uring_get_file(ctx);
3396 ret = PTR_ERR(file);
3401 * Install ring fd as the very last thing, so we don't risk someone
3402 * having closed it before we finish setup
3404 ret = io_uring_install_fd(ctx, file);
3406 /* fput will clean it up */
3411 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3414 io_ring_ctx_wait_and_kill(ctx);
3419 * Sets up an aio uring context, and returns the fd. Applications asks for a
3420 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3421 * params structure passed in.
3423 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3425 struct io_uring_params p;
3428 if (copy_from_user(&p, params, sizeof(p)))
3430 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3435 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3436 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3437 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3438 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3439 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3440 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3441 IORING_SETUP_SINGLE_ISSUER))
3444 return io_uring_create(entries, &p, params);
3447 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3448 struct io_uring_params __user *, params)
3450 return io_uring_setup(entries, params);
3453 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3456 struct io_uring_probe *p;
3460 size = struct_size(p, ops, nr_args);
3461 if (size == SIZE_MAX)
3463 p = kzalloc(size, GFP_KERNEL);
3468 if (copy_from_user(p, arg, size))
3471 if (memchr_inv(p, 0, size))
3474 p->last_op = IORING_OP_LAST - 1;
3475 if (nr_args > IORING_OP_LAST)
3476 nr_args = IORING_OP_LAST;
3478 for (i = 0; i < nr_args; i++) {
3480 if (!io_op_defs[i].not_supported)
3481 p->ops[i].flags = IO_URING_OP_SUPPORTED;
3486 if (copy_to_user(arg, p, size))
3493 static int io_register_personality(struct io_ring_ctx *ctx)
3495 const struct cred *creds;
3499 creds = get_current_cred();
3501 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
3502 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
3510 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
3511 void __user *arg, unsigned int nr_args)
3513 struct io_uring_restriction *res;
3517 /* Restrictions allowed only if rings started disabled */
3518 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3521 /* We allow only a single restrictions registration */
3522 if (ctx->restrictions.registered)
3525 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
3528 size = array_size(nr_args, sizeof(*res));
3529 if (size == SIZE_MAX)
3532 res = memdup_user(arg, size);
3534 return PTR_ERR(res);
3538 for (i = 0; i < nr_args; i++) {
3539 switch (res[i].opcode) {
3540 case IORING_RESTRICTION_REGISTER_OP:
3541 if (res[i].register_op >= IORING_REGISTER_LAST) {
3546 __set_bit(res[i].register_op,
3547 ctx->restrictions.register_op);
3549 case IORING_RESTRICTION_SQE_OP:
3550 if (res[i].sqe_op >= IORING_OP_LAST) {
3555 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
3557 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
3558 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
3560 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
3561 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
3570 /* Reset all restrictions if an error happened */
3572 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
3574 ctx->restrictions.registered = true;
3580 static int io_register_enable_rings(struct io_ring_ctx *ctx)
3582 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3585 if (ctx->restrictions.registered)
3586 ctx->restricted = 1;
3588 ctx->flags &= ~IORING_SETUP_R_DISABLED;
3589 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
3590 wake_up(&ctx->sq_data->wait);
3594 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
3595 void __user *arg, unsigned len)
3597 struct io_uring_task *tctx = current->io_uring;
3598 cpumask_var_t new_mask;
3601 if (!tctx || !tctx->io_wq)
3604 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3607 cpumask_clear(new_mask);
3608 if (len > cpumask_size())
3609 len = cpumask_size();
3611 if (in_compat_syscall()) {
3612 ret = compat_get_bitmap(cpumask_bits(new_mask),
3613 (const compat_ulong_t __user *)arg,
3614 len * 8 /* CHAR_BIT */);
3616 ret = copy_from_user(new_mask, arg, len);
3620 free_cpumask_var(new_mask);
3624 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
3625 free_cpumask_var(new_mask);
3629 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
3631 struct io_uring_task *tctx = current->io_uring;
3633 if (!tctx || !tctx->io_wq)
3636 return io_wq_cpu_affinity(tctx->io_wq, NULL);
3639 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
3641 __must_hold(&ctx->uring_lock)
3643 struct io_tctx_node *node;
3644 struct io_uring_task *tctx = NULL;
3645 struct io_sq_data *sqd = NULL;
3649 if (copy_from_user(new_count, arg, sizeof(new_count)))
3651 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3652 if (new_count[i] > INT_MAX)
3655 if (ctx->flags & IORING_SETUP_SQPOLL) {
3659 * Observe the correct sqd->lock -> ctx->uring_lock
3660 * ordering. Fine to drop uring_lock here, we hold
3663 refcount_inc(&sqd->refs);
3664 mutex_unlock(&ctx->uring_lock);
3665 mutex_lock(&sqd->lock);
3666 mutex_lock(&ctx->uring_lock);
3668 tctx = sqd->thread->io_uring;
3671 tctx = current->io_uring;
3674 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
3676 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3678 ctx->iowq_limits[i] = new_count[i];
3679 ctx->iowq_limits_set = true;
3681 if (tctx && tctx->io_wq) {
3682 ret = io_wq_max_workers(tctx->io_wq, new_count);
3686 memset(new_count, 0, sizeof(new_count));
3690 mutex_unlock(&sqd->lock);
3691 io_put_sq_data(sqd);
3694 if (copy_to_user(arg, new_count, sizeof(new_count)))
3697 /* that's it for SQPOLL, only the SQPOLL task creates requests */
3701 /* now propagate the restriction to all registered users */
3702 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3703 struct io_uring_task *tctx = node->task->io_uring;
3705 if (WARN_ON_ONCE(!tctx->io_wq))
3708 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3709 new_count[i] = ctx->iowq_limits[i];
3710 /* ignore errors, it always returns zero anyway */
3711 (void)io_wq_max_workers(tctx->io_wq, new_count);
3716 mutex_unlock(&sqd->lock);
3717 io_put_sq_data(sqd);
3722 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
3723 void __user *arg, unsigned nr_args)
3724 __releases(ctx->uring_lock)
3725 __acquires(ctx->uring_lock)
3730 * We're inside the ring mutex, if the ref is already dying, then
3731 * someone else killed the ctx or is already going through
3732 * io_uring_register().
3734 if (percpu_ref_is_dying(&ctx->refs))
3737 if (ctx->restricted) {
3738 if (opcode >= IORING_REGISTER_LAST)
3740 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
3741 if (!test_bit(opcode, ctx->restrictions.register_op))
3746 case IORING_REGISTER_BUFFERS:
3750 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
3752 case IORING_UNREGISTER_BUFFERS:
3756 ret = io_sqe_buffers_unregister(ctx);
3758 case IORING_REGISTER_FILES:
3762 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
3764 case IORING_UNREGISTER_FILES:
3768 ret = io_sqe_files_unregister(ctx);
3770 case IORING_REGISTER_FILES_UPDATE:
3771 ret = io_register_files_update(ctx, arg, nr_args);
3773 case IORING_REGISTER_EVENTFD:
3777 ret = io_eventfd_register(ctx, arg, 0);
3779 case IORING_REGISTER_EVENTFD_ASYNC:
3783 ret = io_eventfd_register(ctx, arg, 1);
3785 case IORING_UNREGISTER_EVENTFD:
3789 ret = io_eventfd_unregister(ctx);
3791 case IORING_REGISTER_PROBE:
3793 if (!arg || nr_args > 256)
3795 ret = io_probe(ctx, arg, nr_args);
3797 case IORING_REGISTER_PERSONALITY:
3801 ret = io_register_personality(ctx);
3803 case IORING_UNREGISTER_PERSONALITY:
3807 ret = io_unregister_personality(ctx, nr_args);
3809 case IORING_REGISTER_ENABLE_RINGS:
3813 ret = io_register_enable_rings(ctx);
3815 case IORING_REGISTER_RESTRICTIONS:
3816 ret = io_register_restrictions(ctx, arg, nr_args);
3818 case IORING_REGISTER_FILES2:
3819 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
3821 case IORING_REGISTER_FILES_UPDATE2:
3822 ret = io_register_rsrc_update(ctx, arg, nr_args,
3825 case IORING_REGISTER_BUFFERS2:
3826 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
3828 case IORING_REGISTER_BUFFERS_UPDATE:
3829 ret = io_register_rsrc_update(ctx, arg, nr_args,
3830 IORING_RSRC_BUFFER);
3832 case IORING_REGISTER_IOWQ_AFF:
3834 if (!arg || !nr_args)
3836 ret = io_register_iowq_aff(ctx, arg, nr_args);
3838 case IORING_UNREGISTER_IOWQ_AFF:
3842 ret = io_unregister_iowq_aff(ctx);
3844 case IORING_REGISTER_IOWQ_MAX_WORKERS:
3846 if (!arg || nr_args != 2)
3848 ret = io_register_iowq_max_workers(ctx, arg);
3850 case IORING_REGISTER_RING_FDS:
3851 ret = io_ringfd_register(ctx, arg, nr_args);
3853 case IORING_UNREGISTER_RING_FDS:
3854 ret = io_ringfd_unregister(ctx, arg, nr_args);
3856 case IORING_REGISTER_PBUF_RING:
3858 if (!arg || nr_args != 1)
3860 ret = io_register_pbuf_ring(ctx, arg);
3862 case IORING_UNREGISTER_PBUF_RING:
3864 if (!arg || nr_args != 1)
3866 ret = io_unregister_pbuf_ring(ctx, arg);
3876 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
3877 void __user *, arg, unsigned int, nr_args)
3879 struct io_ring_ctx *ctx;
3888 if (!io_is_uring_fops(f.file))
3891 ctx = f.file->private_data;
3895 mutex_lock(&ctx->uring_lock);
3896 ret = __io_uring_register(ctx, opcode, arg, nr_args);
3897 mutex_unlock(&ctx->uring_lock);
3898 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
3904 static int __init io_uring_init(void)
3906 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
3907 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
3908 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
3911 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
3912 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
3913 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
3914 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
3915 BUILD_BUG_SQE_ELEM(1, __u8, flags);
3916 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
3917 BUILD_BUG_SQE_ELEM(4, __s32, fd);
3918 BUILD_BUG_SQE_ELEM(8, __u64, off);
3919 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
3920 BUILD_BUG_SQE_ELEM(16, __u64, addr);
3921 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
3922 BUILD_BUG_SQE_ELEM(24, __u32, len);
3923 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
3924 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
3925 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
3926 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
3927 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
3928 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
3929 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
3930 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
3931 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
3932 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
3933 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
3934 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
3935 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
3936 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
3937 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
3938 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
3939 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
3940 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
3941 BUILD_BUG_SQE_ELEM(42, __u16, personality);
3942 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
3943 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
3944 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
3946 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
3947 sizeof(struct io_uring_rsrc_update));
3948 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
3949 sizeof(struct io_uring_rsrc_update2));
3951 /* ->buf_index is u16 */
3952 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
3953 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
3954 offsetof(struct io_uring_buf_ring, tail));
3956 /* should fit into one byte */
3957 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
3958 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
3959 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
3961 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
3963 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
3965 io_uring_optable_init();
3967 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
3971 __initcall(io_uring_init);