1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <net/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
49 #include <linux/bits.h>
51 #include <linux/sched/signal.h>
53 #include <linux/file.h>
54 #include <linux/fdtable.h>
56 #include <linux/mman.h>
57 #include <linux/percpu.h>
58 #include <linux/slab.h>
59 #include <linux/bvec.h>
60 #include <linux/net.h>
62 #include <net/af_unix.h>
64 #include <linux/anon_inodes.h>
65 #include <linux/sched/mm.h>
66 #include <linux/uaccess.h>
67 #include <linux/nospec.h>
68 #include <linux/highmem.h>
69 #include <linux/fsnotify.h>
70 #include <linux/fadvise.h>
71 #include <linux/task_work.h>
72 #include <linux/io_uring.h>
73 #include <linux/audit.h>
74 #include <linux/security.h>
76 #define CREATE_TRACE_POINTS
77 #include <trace/events/io_uring.h>
79 #include <uapi/linux/io_uring.h>
83 #include "io_uring_types.h"
97 #define IORING_MAX_ENTRIES 32768
98 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
100 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
101 IORING_REGISTER_LAST + IORING_OP_LAST)
103 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
104 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
106 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
107 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
109 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
110 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
113 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
116 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
118 #define IO_COMPL_BATCH 32
119 #define IO_REQ_ALLOC_BATCH 8
122 IO_CHECK_CQ_OVERFLOW_BIT,
123 IO_CHECK_CQ_DROPPED_BIT,
126 struct io_defer_entry {
127 struct list_head list;
128 struct io_kiocb *req;
132 /* requests with any of those set should undergo io_disarm_next() */
133 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
134 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
136 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
137 struct task_struct *task,
140 static void io_dismantle_req(struct io_kiocb *req);
141 static void io_clean_op(struct io_kiocb *req);
142 static void io_queue_sqe(struct io_kiocb *req);
144 static void __io_submit_flush_completions(struct io_ring_ctx *ctx);
146 static void io_eventfd_signal(struct io_ring_ctx *ctx);
148 static struct kmem_cache *req_cachep;
150 struct sock *io_uring_get_socket(struct file *file)
152 #if defined(CONFIG_UNIX)
153 if (io_is_uring_fops(file)) {
154 struct io_ring_ctx *ctx = file->private_data;
156 return ctx->ring_sock->sk;
161 EXPORT_SYMBOL(io_uring_get_socket);
163 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
165 if (!wq_list_empty(&ctx->submit_state.compl_reqs))
166 __io_submit_flush_completions(ctx);
169 static bool io_match_linked(struct io_kiocb *head)
171 struct io_kiocb *req;
173 io_for_each_link(req, head) {
174 if (req->flags & REQ_F_INFLIGHT)
181 * As io_match_task() but protected against racing with linked timeouts.
182 * User must not hold timeout_lock.
184 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
189 if (task && head->task != task)
194 if (head->flags & REQ_F_LINK_TIMEOUT) {
195 struct io_ring_ctx *ctx = head->ctx;
197 /* protect against races with linked timeouts */
198 spin_lock_irq(&ctx->timeout_lock);
199 matched = io_match_linked(head);
200 spin_unlock_irq(&ctx->timeout_lock);
202 matched = io_match_linked(head);
207 static inline void req_fail_link_node(struct io_kiocb *req, int res)
210 io_req_set_res(req, res, 0);
213 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
215 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
218 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
220 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
222 complete(&ctx->ref_comp);
225 static __cold void io_fallback_req_func(struct work_struct *work)
227 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
229 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
230 struct io_kiocb *req, *tmp;
233 percpu_ref_get(&ctx->refs);
234 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
235 req->io_task_work.func(req, &locked);
238 io_submit_flush_completions(ctx);
239 mutex_unlock(&ctx->uring_lock);
241 percpu_ref_put(&ctx->refs);
244 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
246 struct io_ring_ctx *ctx;
249 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
253 xa_init(&ctx->io_bl_xa);
256 * Use 5 bits less than the max cq entries, that should give us around
257 * 32 entries per hash list if totally full and uniformly spread.
259 hash_bits = ilog2(p->cq_entries);
263 ctx->cancel_hash_bits = hash_bits;
265 kmalloc((1U << hash_bits) * sizeof(struct io_hash_bucket),
267 if (!ctx->cancel_hash)
270 init_hash_table(ctx->cancel_hash, 1U << hash_bits);
272 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
273 if (!ctx->dummy_ubuf)
275 /* set invalid range, so io_import_fixed() fails meeting it */
276 ctx->dummy_ubuf->ubuf = -1UL;
278 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
279 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
282 ctx->flags = p->flags;
283 init_waitqueue_head(&ctx->sqo_sq_wait);
284 INIT_LIST_HEAD(&ctx->sqd_list);
285 INIT_LIST_HEAD(&ctx->cq_overflow_list);
286 INIT_LIST_HEAD(&ctx->io_buffers_cache);
287 INIT_LIST_HEAD(&ctx->apoll_cache);
288 init_completion(&ctx->ref_comp);
289 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
290 mutex_init(&ctx->uring_lock);
291 init_waitqueue_head(&ctx->cq_wait);
292 spin_lock_init(&ctx->completion_lock);
293 spin_lock_init(&ctx->timeout_lock);
294 INIT_WQ_LIST(&ctx->iopoll_list);
295 INIT_LIST_HEAD(&ctx->io_buffers_pages);
296 INIT_LIST_HEAD(&ctx->io_buffers_comp);
297 INIT_LIST_HEAD(&ctx->defer_list);
298 INIT_LIST_HEAD(&ctx->timeout_list);
299 INIT_LIST_HEAD(&ctx->ltimeout_list);
300 spin_lock_init(&ctx->rsrc_ref_lock);
301 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
302 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
303 init_llist_head(&ctx->rsrc_put_llist);
304 INIT_LIST_HEAD(&ctx->tctx_list);
305 ctx->submit_state.free_list.next = NULL;
306 INIT_WQ_LIST(&ctx->locked_free_list);
307 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
308 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
311 kfree(ctx->dummy_ubuf);
312 kfree(ctx->cancel_hash);
314 xa_destroy(&ctx->io_bl_xa);
319 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
321 struct io_rings *r = ctx->rings;
323 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
327 static bool req_need_defer(struct io_kiocb *req, u32 seq)
329 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
330 struct io_ring_ctx *ctx = req->ctx;
332 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
338 static inline void io_req_track_inflight(struct io_kiocb *req)
340 if (!(req->flags & REQ_F_INFLIGHT)) {
341 req->flags |= REQ_F_INFLIGHT;
342 atomic_inc(&req->task->io_uring->inflight_tracked);
346 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
348 if (WARN_ON_ONCE(!req->link))
351 req->flags &= ~REQ_F_ARM_LTIMEOUT;
352 req->flags |= REQ_F_LINK_TIMEOUT;
354 /* linked timeouts should have two refs once prep'ed */
355 io_req_set_refcount(req);
356 __io_req_set_refcount(req->link, 2);
360 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
362 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
364 return __io_prep_linked_timeout(req);
367 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
369 io_queue_linked_timeout(__io_prep_linked_timeout(req));
372 static inline void io_arm_ltimeout(struct io_kiocb *req)
374 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
375 __io_arm_ltimeout(req);
378 static void io_prep_async_work(struct io_kiocb *req)
380 const struct io_op_def *def = &io_op_defs[req->opcode];
381 struct io_ring_ctx *ctx = req->ctx;
383 if (!(req->flags & REQ_F_CREDS)) {
384 req->flags |= REQ_F_CREDS;
385 req->creds = get_current_cred();
388 req->work.list.next = NULL;
390 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
391 if (req->flags & REQ_F_FORCE_ASYNC)
392 req->work.flags |= IO_WQ_WORK_CONCURRENT;
394 if (req->flags & REQ_F_ISREG) {
395 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
396 io_wq_hash_work(&req->work, file_inode(req->file));
397 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
398 if (def->unbound_nonreg_file)
399 req->work.flags |= IO_WQ_WORK_UNBOUND;
403 static void io_prep_async_link(struct io_kiocb *req)
405 struct io_kiocb *cur;
407 if (req->flags & REQ_F_LINK_TIMEOUT) {
408 struct io_ring_ctx *ctx = req->ctx;
410 spin_lock_irq(&ctx->timeout_lock);
411 io_for_each_link(cur, req)
412 io_prep_async_work(cur);
413 spin_unlock_irq(&ctx->timeout_lock);
415 io_for_each_link(cur, req)
416 io_prep_async_work(cur);
420 void io_queue_iowq(struct io_kiocb *req, bool *dont_use)
422 struct io_kiocb *link = io_prep_linked_timeout(req);
423 struct io_uring_task *tctx = req->task->io_uring;
426 BUG_ON(!tctx->io_wq);
428 /* init ->work of the whole link before punting */
429 io_prep_async_link(req);
432 * Not expected to happen, but if we do have a bug where this _can_
433 * happen, catch it here and ensure the request is marked as
434 * canceled. That will make io-wq go through the usual work cancel
435 * procedure rather than attempt to run this request (or create a new
438 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
439 req->work.flags |= IO_WQ_WORK_CANCEL;
441 trace_io_uring_queue_async_work(req->ctx, req, req->cqe.user_data,
442 req->opcode, req->flags, &req->work,
443 io_wq_is_hashed(&req->work));
444 io_wq_enqueue(tctx->io_wq, &req->work);
446 io_queue_linked_timeout(link);
449 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
451 while (!list_empty(&ctx->defer_list)) {
452 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
453 struct io_defer_entry, list);
455 if (req_need_defer(de->req, de->seq))
457 list_del_init(&de->list);
458 io_req_task_queue(de->req);
463 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
465 if (ctx->off_timeout_used || ctx->drain_active) {
466 spin_lock(&ctx->completion_lock);
467 if (ctx->off_timeout_used)
468 io_flush_timeouts(ctx);
469 if (ctx->drain_active)
470 io_queue_deferred(ctx);
471 io_commit_cqring(ctx);
472 spin_unlock(&ctx->completion_lock);
475 io_eventfd_signal(ctx);
478 static void io_eventfd_signal(struct io_ring_ctx *ctx)
480 struct io_ev_fd *ev_fd;
484 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
487 ev_fd = rcu_dereference(ctx->io_ev_fd);
490 * Check again if ev_fd exists incase an io_eventfd_unregister call
491 * completed between the NULL check of ctx->io_ev_fd at the start of
492 * the function and rcu_read_lock.
494 if (unlikely(!ev_fd))
496 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
499 if (!ev_fd->eventfd_async || io_wq_current_is_worker())
500 eventfd_signal(ev_fd->cq_ev_fd, 1);
506 * This should only get called when at least one event has been posted.
507 * Some applications rely on the eventfd notification count only changing
508 * IFF a new CQE has been added to the CQ ring. There's no depedency on
509 * 1:1 relationship between how many times this function is called (and
510 * hence the eventfd count) and number of CQEs posted to the CQ ring.
512 void io_cqring_ev_posted(struct io_ring_ctx *ctx)
514 if (unlikely(ctx->off_timeout_used || ctx->drain_active ||
516 __io_commit_cqring_flush(ctx);
521 /* Returns true if there are no backlogged entries after the flush */
522 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
524 bool all_flushed, posted;
525 size_t cqe_size = sizeof(struct io_uring_cqe);
527 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
530 if (ctx->flags & IORING_SETUP_CQE32)
534 spin_lock(&ctx->completion_lock);
535 while (!list_empty(&ctx->cq_overflow_list)) {
536 struct io_uring_cqe *cqe = io_get_cqe(ctx);
537 struct io_overflow_cqe *ocqe;
541 ocqe = list_first_entry(&ctx->cq_overflow_list,
542 struct io_overflow_cqe, list);
544 memcpy(cqe, &ocqe->cqe, cqe_size);
546 io_account_cq_overflow(ctx);
549 list_del(&ocqe->list);
553 all_flushed = list_empty(&ctx->cq_overflow_list);
555 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
556 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
559 io_commit_cqring(ctx);
560 spin_unlock(&ctx->completion_lock);
562 io_cqring_ev_posted(ctx);
566 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
570 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
571 /* iopoll syncs against uring_lock, not completion_lock */
572 if (ctx->flags & IORING_SETUP_IOPOLL)
573 mutex_lock(&ctx->uring_lock);
574 ret = __io_cqring_overflow_flush(ctx, false);
575 if (ctx->flags & IORING_SETUP_IOPOLL)
576 mutex_unlock(&ctx->uring_lock);
582 static void __io_put_task(struct task_struct *task, int nr)
584 struct io_uring_task *tctx = task->io_uring;
586 percpu_counter_sub(&tctx->inflight, nr);
587 if (unlikely(atomic_read(&tctx->in_idle)))
588 wake_up(&tctx->wait);
589 put_task_struct_many(task, nr);
592 /* must to be called somewhat shortly after putting a request */
593 static inline void io_put_task(struct task_struct *task, int nr)
595 if (likely(task == current))
596 task->io_uring->cached_refs += nr;
598 __io_put_task(task, nr);
601 static void io_task_refs_refill(struct io_uring_task *tctx)
603 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
605 percpu_counter_add(&tctx->inflight, refill);
606 refcount_add(refill, ¤t->usage);
607 tctx->cached_refs += refill;
610 static inline void io_get_task_refs(int nr)
612 struct io_uring_task *tctx = current->io_uring;
614 tctx->cached_refs -= nr;
615 if (unlikely(tctx->cached_refs < 0))
616 io_task_refs_refill(tctx);
619 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
621 struct io_uring_task *tctx = task->io_uring;
622 unsigned int refs = tctx->cached_refs;
625 tctx->cached_refs = 0;
626 percpu_counter_sub(&tctx->inflight, refs);
627 put_task_struct_many(task, refs);
631 bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data, s32 res,
632 u32 cflags, u64 extra1, u64 extra2)
634 struct io_overflow_cqe *ocqe;
635 size_t ocq_size = sizeof(struct io_overflow_cqe);
636 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
639 ocq_size += sizeof(struct io_uring_cqe);
641 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
642 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
645 * If we're in ring overflow flush mode, or in task cancel mode,
646 * or cannot allocate an overflow entry, then we need to drop it
649 io_account_cq_overflow(ctx);
650 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
653 if (list_empty(&ctx->cq_overflow_list)) {
654 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
655 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
658 ocqe->cqe.user_data = user_data;
660 ocqe->cqe.flags = cflags;
662 ocqe->cqe.big_cqe[0] = extra1;
663 ocqe->cqe.big_cqe[1] = extra2;
665 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
669 bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
672 struct io_uring_cqe *cqe;
675 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
678 * If we can't get a cq entry, userspace overflowed the
679 * submission (by quite a lot). Increment the overflow count in
682 cqe = io_get_cqe(ctx);
684 WRITE_ONCE(cqe->user_data, user_data);
685 WRITE_ONCE(cqe->res, res);
686 WRITE_ONCE(cqe->flags, cflags);
688 if (ctx->flags & IORING_SETUP_CQE32) {
689 WRITE_ONCE(cqe->big_cqe[0], 0);
690 WRITE_ONCE(cqe->big_cqe[1], 0);
694 return io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
697 static void __io_req_complete_put(struct io_kiocb *req)
700 * If we're the last reference to this request, add to our locked
703 if (req_ref_put_and_test(req)) {
704 struct io_ring_ctx *ctx = req->ctx;
706 if (req->flags & IO_REQ_LINK_FLAGS) {
707 if (req->flags & IO_DISARM_MASK)
710 io_req_task_queue(req->link);
714 io_req_put_rsrc(req);
716 * Selected buffer deallocation in io_clean_op() assumes that
717 * we don't hold ->completion_lock. Clean them here to avoid
720 io_put_kbuf_comp(req);
721 io_dismantle_req(req);
722 io_put_task(req->task, 1);
723 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
724 ctx->locked_free_nr++;
728 void __io_req_complete_post(struct io_kiocb *req)
730 if (!(req->flags & REQ_F_CQE_SKIP))
731 __io_fill_cqe_req(req->ctx, req);
732 __io_req_complete_put(req);
735 void io_req_complete_post(struct io_kiocb *req)
737 struct io_ring_ctx *ctx = req->ctx;
739 spin_lock(&ctx->completion_lock);
740 __io_req_complete_post(req);
741 io_commit_cqring(ctx);
742 spin_unlock(&ctx->completion_lock);
743 io_cqring_ev_posted(ctx);
746 inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags)
748 io_req_complete_post(req);
751 void io_req_complete_failed(struct io_kiocb *req, s32 res)
754 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
755 io_req_complete_post(req);
759 * Don't initialise the fields below on every allocation, but do that in
760 * advance and keep them valid across allocations.
762 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
766 req->async_data = NULL;
767 /* not necessary, but safer to zero */
771 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
772 struct io_submit_state *state)
774 spin_lock(&ctx->completion_lock);
775 wq_list_splice(&ctx->locked_free_list, &state->free_list);
776 ctx->locked_free_nr = 0;
777 spin_unlock(&ctx->completion_lock);
780 static inline bool io_req_cache_empty(struct io_ring_ctx *ctx)
782 return !ctx->submit_state.free_list.next;
786 * A request might get retired back into the request caches even before opcode
787 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
788 * Because of that, io_alloc_req() should be called only under ->uring_lock
789 * and with extra caution to not get a request that is still worked on.
791 static __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
792 __must_hold(&ctx->uring_lock)
794 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
795 void *reqs[IO_REQ_ALLOC_BATCH];
799 * If we have more than a batch's worth of requests in our IRQ side
800 * locked cache, grab the lock and move them over to our submission
803 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
804 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
805 if (!io_req_cache_empty(ctx))
809 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
812 * Bulk alloc is all-or-nothing. If we fail to get a batch,
813 * retry single alloc to be on the safe side.
815 if (unlikely(ret <= 0)) {
816 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
822 percpu_ref_get_many(&ctx->refs, ret);
823 for (i = 0; i < ret; i++) {
824 struct io_kiocb *req = reqs[i];
826 io_preinit_req(req, ctx);
827 io_req_add_to_cache(req, ctx);
832 static inline bool io_alloc_req_refill(struct io_ring_ctx *ctx)
834 if (unlikely(io_req_cache_empty(ctx)))
835 return __io_alloc_req_refill(ctx);
839 static inline struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
841 struct io_wq_work_node *node;
843 node = wq_stack_extract(&ctx->submit_state.free_list);
844 return container_of(node, struct io_kiocb, comp_list);
847 static inline void io_dismantle_req(struct io_kiocb *req)
849 unsigned int flags = req->flags;
851 if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
853 if (!(flags & REQ_F_FIXED_FILE))
854 io_put_file(req->file);
857 __cold void io_free_req(struct io_kiocb *req)
859 struct io_ring_ctx *ctx = req->ctx;
861 io_req_put_rsrc(req);
862 io_dismantle_req(req);
863 io_put_task(req->task, 1);
865 spin_lock(&ctx->completion_lock);
866 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
867 ctx->locked_free_nr++;
868 spin_unlock(&ctx->completion_lock);
871 static void __io_req_find_next_prep(struct io_kiocb *req)
873 struct io_ring_ctx *ctx = req->ctx;
876 spin_lock(&ctx->completion_lock);
877 posted = io_disarm_next(req);
878 io_commit_cqring(ctx);
879 spin_unlock(&ctx->completion_lock);
881 io_cqring_ev_posted(ctx);
884 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
886 struct io_kiocb *nxt;
889 * If LINK is set, we have dependent requests in this chain. If we
890 * didn't fail this request, queue the first one up, moving any other
891 * dependencies to the next request. In case of failure, fail the rest
894 if (unlikely(req->flags & IO_DISARM_MASK))
895 __io_req_find_next_prep(req);
901 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
905 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
906 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
908 io_submit_flush_completions(ctx);
909 mutex_unlock(&ctx->uring_lock);
912 percpu_ref_put(&ctx->refs);
915 static inline void ctx_commit_and_unlock(struct io_ring_ctx *ctx)
917 io_commit_cqring(ctx);
918 spin_unlock(&ctx->completion_lock);
919 io_cqring_ev_posted(ctx);
922 static void handle_prev_tw_list(struct io_wq_work_node *node,
923 struct io_ring_ctx **ctx, bool *uring_locked)
925 if (*ctx && !*uring_locked)
926 spin_lock(&(*ctx)->completion_lock);
929 struct io_wq_work_node *next = node->next;
930 struct io_kiocb *req = container_of(node, struct io_kiocb,
933 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
935 if (req->ctx != *ctx) {
936 if (unlikely(!*uring_locked && *ctx))
937 ctx_commit_and_unlock(*ctx);
939 ctx_flush_and_put(*ctx, uring_locked);
941 /* if not contended, grab and improve batching */
942 *uring_locked = mutex_trylock(&(*ctx)->uring_lock);
943 percpu_ref_get(&(*ctx)->refs);
944 if (unlikely(!*uring_locked))
945 spin_lock(&(*ctx)->completion_lock);
947 if (likely(*uring_locked)) {
948 req->io_task_work.func(req, uring_locked);
950 req->cqe.flags = io_put_kbuf_comp(req);
951 __io_req_complete_post(req);
956 if (unlikely(!*uring_locked))
957 ctx_commit_and_unlock(*ctx);
960 static void handle_tw_list(struct io_wq_work_node *node,
961 struct io_ring_ctx **ctx, bool *locked)
964 struct io_wq_work_node *next = node->next;
965 struct io_kiocb *req = container_of(node, struct io_kiocb,
968 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
970 if (req->ctx != *ctx) {
971 ctx_flush_and_put(*ctx, locked);
973 /* if not contended, grab and improve batching */
974 *locked = mutex_trylock(&(*ctx)->uring_lock);
975 percpu_ref_get(&(*ctx)->refs);
977 req->io_task_work.func(req, locked);
982 void tctx_task_work(struct callback_head *cb)
984 bool uring_locked = false;
985 struct io_ring_ctx *ctx = NULL;
986 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
990 struct io_wq_work_node *node1, *node2;
992 spin_lock_irq(&tctx->task_lock);
993 node1 = tctx->prio_task_list.first;
994 node2 = tctx->task_list.first;
995 INIT_WQ_LIST(&tctx->task_list);
996 INIT_WQ_LIST(&tctx->prio_task_list);
997 if (!node2 && !node1)
998 tctx->task_running = false;
999 spin_unlock_irq(&tctx->task_lock);
1000 if (!node2 && !node1)
1004 handle_prev_tw_list(node1, &ctx, &uring_locked);
1006 handle_tw_list(node2, &ctx, &uring_locked);
1009 if (data_race(!tctx->task_list.first) &&
1010 data_race(!tctx->prio_task_list.first) && uring_locked)
1011 io_submit_flush_completions(ctx);
1014 ctx_flush_and_put(ctx, &uring_locked);
1016 /* relaxed read is enough as only the task itself sets ->in_idle */
1017 if (unlikely(atomic_read(&tctx->in_idle)))
1018 io_uring_drop_tctx_refs(current);
1021 static void __io_req_task_work_add(struct io_kiocb *req,
1022 struct io_uring_task *tctx,
1023 struct io_wq_work_list *list)
1025 struct io_ring_ctx *ctx = req->ctx;
1026 struct io_wq_work_node *node;
1027 unsigned long flags;
1030 spin_lock_irqsave(&tctx->task_lock, flags);
1031 wq_list_add_tail(&req->io_task_work.node, list);
1032 running = tctx->task_running;
1034 tctx->task_running = true;
1035 spin_unlock_irqrestore(&tctx->task_lock, flags);
1037 /* task_work already pending, we're done */
1041 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1042 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1044 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1047 spin_lock_irqsave(&tctx->task_lock, flags);
1048 tctx->task_running = false;
1049 node = wq_list_merge(&tctx->prio_task_list, &tctx->task_list);
1050 spin_unlock_irqrestore(&tctx->task_lock, flags);
1053 req = container_of(node, struct io_kiocb, io_task_work.node);
1055 if (llist_add(&req->io_task_work.fallback_node,
1056 &req->ctx->fallback_llist))
1057 schedule_delayed_work(&req->ctx->fallback_work, 1);
1061 void io_req_task_work_add(struct io_kiocb *req)
1063 struct io_uring_task *tctx = req->task->io_uring;
1065 __io_req_task_work_add(req, tctx, &tctx->task_list);
1068 void io_req_task_prio_work_add(struct io_kiocb *req)
1070 struct io_uring_task *tctx = req->task->io_uring;
1072 if (req->ctx->flags & IORING_SETUP_SQPOLL)
1073 __io_req_task_work_add(req, tctx, &tctx->prio_task_list);
1075 __io_req_task_work_add(req, tctx, &tctx->task_list);
1078 static void io_req_tw_post(struct io_kiocb *req, bool *locked)
1080 io_req_complete_post(req);
1083 void io_req_tw_post_queue(struct io_kiocb *req, s32 res, u32 cflags)
1085 io_req_set_res(req, res, cflags);
1086 req->io_task_work.func = io_req_tw_post;
1087 io_req_task_work_add(req);
1090 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
1092 /* not needed for normal modes, but SQPOLL depends on it */
1093 io_tw_lock(req->ctx, locked);
1094 io_req_complete_failed(req, req->cqe.res);
1097 void io_req_task_submit(struct io_kiocb *req, bool *locked)
1099 io_tw_lock(req->ctx, locked);
1100 /* req->task == current here, checking PF_EXITING is safe */
1101 if (likely(!(req->task->flags & PF_EXITING)))
1104 io_req_complete_failed(req, -EFAULT);
1107 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1109 io_req_set_res(req, ret, 0);
1110 req->io_task_work.func = io_req_task_cancel;
1111 io_req_task_work_add(req);
1114 void io_req_task_queue(struct io_kiocb *req)
1116 req->io_task_work.func = io_req_task_submit;
1117 io_req_task_work_add(req);
1120 void io_queue_next(struct io_kiocb *req)
1122 struct io_kiocb *nxt = io_req_find_next(req);
1125 io_req_task_queue(nxt);
1128 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1129 __must_hold(&ctx->uring_lock)
1131 struct task_struct *task = NULL;
1135 struct io_kiocb *req = container_of(node, struct io_kiocb,
1138 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1139 if (req->flags & REQ_F_REFCOUNT) {
1140 node = req->comp_list.next;
1141 if (!req_ref_put_and_test(req))
1144 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1145 struct async_poll *apoll = req->apoll;
1147 if (apoll->double_poll)
1148 kfree(apoll->double_poll);
1149 list_add(&apoll->poll.wait.entry,
1151 req->flags &= ~REQ_F_POLLED;
1153 if (req->flags & IO_REQ_LINK_FLAGS)
1155 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1158 if (!(req->flags & REQ_F_FIXED_FILE))
1159 io_put_file(req->file);
1161 io_req_put_rsrc_locked(req, ctx);
1163 if (req->task != task) {
1165 io_put_task(task, task_refs);
1170 node = req->comp_list.next;
1171 io_req_add_to_cache(req, ctx);
1175 io_put_task(task, task_refs);
1178 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1179 __must_hold(&ctx->uring_lock)
1181 struct io_wq_work_node *node, *prev;
1182 struct io_submit_state *state = &ctx->submit_state;
1184 if (state->flush_cqes) {
1185 spin_lock(&ctx->completion_lock);
1186 wq_list_for_each(node, prev, &state->compl_reqs) {
1187 struct io_kiocb *req = container_of(node, struct io_kiocb,
1190 if (!(req->flags & REQ_F_CQE_SKIP))
1191 __io_fill_cqe_req(ctx, req);
1194 io_commit_cqring(ctx);
1195 spin_unlock(&ctx->completion_lock);
1196 io_cqring_ev_posted(ctx);
1197 state->flush_cqes = false;
1200 io_free_batch_list(ctx, state->compl_reqs.first);
1201 INIT_WQ_LIST(&state->compl_reqs);
1205 * Drop reference to request, return next in chain (if there is one) if this
1206 * was the last reference to this request.
1208 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1210 struct io_kiocb *nxt = NULL;
1212 if (req_ref_put_and_test(req)) {
1213 if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
1214 nxt = io_req_find_next(req);
1220 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1222 /* See comment at the top of this file */
1224 return __io_cqring_events(ctx);
1228 * We can't just wait for polled events to come to us, we have to actively
1229 * find and complete them.
1231 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1233 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1236 mutex_lock(&ctx->uring_lock);
1237 while (!wq_list_empty(&ctx->iopoll_list)) {
1238 /* let it sleep and repeat later if can't complete a request */
1239 if (io_do_iopoll(ctx, true) == 0)
1242 * Ensure we allow local-to-the-cpu processing to take place,
1243 * in this case we need to ensure that we reap all events.
1244 * Also let task_work, etc. to progress by releasing the mutex
1246 if (need_resched()) {
1247 mutex_unlock(&ctx->uring_lock);
1249 mutex_lock(&ctx->uring_lock);
1252 mutex_unlock(&ctx->uring_lock);
1255 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1257 unsigned int nr_events = 0;
1259 unsigned long check_cq;
1261 check_cq = READ_ONCE(ctx->check_cq);
1262 if (unlikely(check_cq)) {
1263 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1264 __io_cqring_overflow_flush(ctx, false);
1266 * Similarly do not spin if we have not informed the user of any
1269 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1273 * Don't enter poll loop if we already have events pending.
1274 * If we do, we can potentially be spinning for commands that
1275 * already triggered a CQE (eg in error).
1277 if (io_cqring_events(ctx))
1282 * If a submit got punted to a workqueue, we can have the
1283 * application entering polling for a command before it gets
1284 * issued. That app will hold the uring_lock for the duration
1285 * of the poll right here, so we need to take a breather every
1286 * now and then to ensure that the issue has a chance to add
1287 * the poll to the issued list. Otherwise we can spin here
1288 * forever, while the workqueue is stuck trying to acquire the
1291 if (wq_list_empty(&ctx->iopoll_list)) {
1292 u32 tail = ctx->cached_cq_tail;
1294 mutex_unlock(&ctx->uring_lock);
1296 mutex_lock(&ctx->uring_lock);
1298 /* some requests don't go through iopoll_list */
1299 if (tail != ctx->cached_cq_tail ||
1300 wq_list_empty(&ctx->iopoll_list))
1303 ret = io_do_iopoll(ctx, !min);
1308 } while (nr_events < min && !need_resched());
1313 void io_req_task_complete(struct io_kiocb *req, bool *locked)
1315 if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) {
1316 unsigned issue_flags = *locked ? 0 : IO_URING_F_UNLOCKED;
1318 req->cqe.flags |= io_put_kbuf(req, issue_flags);
1322 io_req_add_compl_list(req);
1324 io_req_complete_post(req);
1328 * After the iocb has been issued, it's safe to be found on the poll list.
1329 * Adding the kiocb to the list AFTER submission ensures that we don't
1330 * find it from a io_do_iopoll() thread before the issuer is done
1331 * accessing the kiocb cookie.
1333 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1335 struct io_ring_ctx *ctx = req->ctx;
1336 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1338 /* workqueue context doesn't hold uring_lock, grab it now */
1339 if (unlikely(needs_lock))
1340 mutex_lock(&ctx->uring_lock);
1343 * Track whether we have multiple files in our lists. This will impact
1344 * how we do polling eventually, not spinning if we're on potentially
1345 * different devices.
1347 if (wq_list_empty(&ctx->iopoll_list)) {
1348 ctx->poll_multi_queue = false;
1349 } else if (!ctx->poll_multi_queue) {
1350 struct io_kiocb *list_req;
1352 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1354 if (list_req->file != req->file)
1355 ctx->poll_multi_queue = true;
1359 * For fast devices, IO may have already completed. If it has, add
1360 * it to the front so we find it first.
1362 if (READ_ONCE(req->iopoll_completed))
1363 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1365 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1367 if (unlikely(needs_lock)) {
1369 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1370 * in sq thread task context or in io worker task context. If
1371 * current task context is sq thread, we don't need to check
1372 * whether should wake up sq thread.
1374 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1375 wq_has_sleeper(&ctx->sq_data->wait))
1376 wake_up(&ctx->sq_data->wait);
1378 mutex_unlock(&ctx->uring_lock);
1382 static bool io_bdev_nowait(struct block_device *bdev)
1384 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
1388 * If we tracked the file through the SCM inflight mechanism, we could support
1389 * any file. For now, just ensure that anything potentially problematic is done
1392 static bool __io_file_supports_nowait(struct file *file, umode_t mode)
1394 if (S_ISBLK(mode)) {
1395 if (IS_ENABLED(CONFIG_BLOCK) &&
1396 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
1402 if (S_ISREG(mode)) {
1403 if (IS_ENABLED(CONFIG_BLOCK) &&
1404 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
1405 !io_is_uring_fops(file))
1410 /* any ->read/write should understand O_NONBLOCK */
1411 if (file->f_flags & O_NONBLOCK)
1413 return file->f_mode & FMODE_NOWAIT;
1417 * If we tracked the file through the SCM inflight mechanism, we could support
1418 * any file. For now, just ensure that anything potentially problematic is done
1421 unsigned int io_file_get_flags(struct file *file)
1423 umode_t mode = file_inode(file)->i_mode;
1424 unsigned int res = 0;
1428 if (__io_file_supports_nowait(file, mode))
1430 if (io_file_need_scm(file))
1435 bool io_alloc_async_data(struct io_kiocb *req)
1437 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
1438 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
1439 if (req->async_data) {
1440 req->flags |= REQ_F_ASYNC_DATA;
1446 int io_req_prep_async(struct io_kiocb *req)
1448 const struct io_op_def *def = &io_op_defs[req->opcode];
1450 /* assign early for deferred execution for non-fixed file */
1451 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE))
1452 req->file = io_file_get_normal(req, req->cqe.fd);
1453 if (!def->prep_async)
1455 if (WARN_ON_ONCE(req_has_async_data(req)))
1457 if (io_alloc_async_data(req))
1460 return def->prep_async(req);
1463 static u32 io_get_sequence(struct io_kiocb *req)
1465 u32 seq = req->ctx->cached_sq_head;
1466 struct io_kiocb *cur;
1468 /* need original cached_sq_head, but it was increased for each req */
1469 io_for_each_link(cur, req)
1474 static __cold void io_drain_req(struct io_kiocb *req)
1476 struct io_ring_ctx *ctx = req->ctx;
1477 struct io_defer_entry *de;
1479 u32 seq = io_get_sequence(req);
1481 /* Still need defer if there is pending req in defer list. */
1482 spin_lock(&ctx->completion_lock);
1483 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1484 spin_unlock(&ctx->completion_lock);
1486 ctx->drain_active = false;
1487 io_req_task_queue(req);
1490 spin_unlock(&ctx->completion_lock);
1492 ret = io_req_prep_async(req);
1495 io_req_complete_failed(req, ret);
1498 io_prep_async_link(req);
1499 de = kmalloc(sizeof(*de), GFP_KERNEL);
1505 spin_lock(&ctx->completion_lock);
1506 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1507 spin_unlock(&ctx->completion_lock);
1512 trace_io_uring_defer(ctx, req, req->cqe.user_data, req->opcode);
1515 list_add_tail(&de->list, &ctx->defer_list);
1516 spin_unlock(&ctx->completion_lock);
1519 static void io_clean_op(struct io_kiocb *req)
1521 if (req->flags & REQ_F_BUFFER_SELECTED) {
1522 spin_lock(&req->ctx->completion_lock);
1523 io_put_kbuf_comp(req);
1524 spin_unlock(&req->ctx->completion_lock);
1527 if (req->flags & REQ_F_NEED_CLEANUP) {
1528 const struct io_op_def *def = &io_op_defs[req->opcode];
1533 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1534 kfree(req->apoll->double_poll);
1538 if (req->flags & REQ_F_INFLIGHT) {
1539 struct io_uring_task *tctx = req->task->io_uring;
1541 atomic_dec(&tctx->inflight_tracked);
1543 if (req->flags & REQ_F_CREDS)
1544 put_cred(req->creds);
1545 if (req->flags & REQ_F_ASYNC_DATA) {
1546 kfree(req->async_data);
1547 req->async_data = NULL;
1549 req->flags &= ~IO_REQ_CLEAN_FLAGS;
1552 static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags)
1554 if (req->file || !io_op_defs[req->opcode].needs_file)
1557 if (req->flags & REQ_F_FIXED_FILE)
1558 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1560 req->file = io_file_get_normal(req, req->cqe.fd);
1565 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1567 const struct io_op_def *def = &io_op_defs[req->opcode];
1568 const struct cred *creds = NULL;
1571 if (unlikely(!io_assign_file(req, issue_flags)))
1574 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1575 creds = override_creds(req->creds);
1577 if (!def->audit_skip)
1578 audit_uring_entry(req->opcode);
1580 ret = def->issue(req, issue_flags);
1582 if (!def->audit_skip)
1583 audit_uring_exit(!ret, ret);
1586 revert_creds(creds);
1588 if (ret == IOU_OK) {
1589 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1590 io_req_add_compl_list(req);
1592 io_req_complete_post(req);
1593 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1596 /* If the op doesn't have a file, we're not polling for it */
1597 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && req->file)
1598 io_iopoll_req_issued(req, issue_flags);
1603 int io_poll_issue(struct io_kiocb *req, bool *locked)
1605 io_tw_lock(req->ctx, locked);
1606 if (unlikely(req->task->flags & PF_EXITING))
1608 return io_issue_sqe(req, IO_URING_F_NONBLOCK);
1611 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1613 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1615 req = io_put_req_find_next(req);
1616 return req ? &req->work : NULL;
1619 void io_wq_submit_work(struct io_wq_work *work)
1621 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1622 const struct io_op_def *def = &io_op_defs[req->opcode];
1623 unsigned int issue_flags = IO_URING_F_UNLOCKED;
1624 bool needs_poll = false;
1625 int ret = 0, err = -ECANCELED;
1627 /* one will be dropped by ->io_free_work() after returning to io-wq */
1628 if (!(req->flags & REQ_F_REFCOUNT))
1629 __io_req_set_refcount(req, 2);
1633 io_arm_ltimeout(req);
1635 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1636 if (work->flags & IO_WQ_WORK_CANCEL) {
1638 io_req_task_queue_fail(req, err);
1641 if (!io_assign_file(req, issue_flags)) {
1643 work->flags |= IO_WQ_WORK_CANCEL;
1647 if (req->flags & REQ_F_FORCE_ASYNC) {
1648 bool opcode_poll = def->pollin || def->pollout;
1650 if (opcode_poll && file_can_poll(req->file)) {
1652 issue_flags |= IO_URING_F_NONBLOCK;
1657 ret = io_issue_sqe(req, issue_flags);
1661 * We can get EAGAIN for iopolled IO even though we're
1662 * forcing a sync submission from here, since we can't
1663 * wait for request slots on the block side.
1666 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1672 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1674 /* aborted or ready, in either case retry blocking */
1676 issue_flags &= ~IO_URING_F_NONBLOCK;
1679 /* avoid locking problems by failing it from a clean context */
1681 io_req_task_queue_fail(req, ret);
1684 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1685 unsigned int issue_flags)
1687 struct io_ring_ctx *ctx = req->ctx;
1688 struct file *file = NULL;
1689 unsigned long file_ptr;
1691 io_ring_submit_lock(ctx, issue_flags);
1693 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
1695 fd = array_index_nospec(fd, ctx->nr_user_files);
1696 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
1697 file = (struct file *) (file_ptr & FFS_MASK);
1698 file_ptr &= ~FFS_MASK;
1699 /* mask in overlapping REQ_F and FFS bits */
1700 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
1701 io_req_set_rsrc_node(req, ctx, 0);
1702 WARN_ON_ONCE(file && !test_bit(fd, ctx->file_table.bitmap));
1704 io_ring_submit_unlock(ctx, issue_flags);
1708 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
1710 struct file *file = fget(fd);
1712 trace_io_uring_file_get(req->ctx, req, req->cqe.user_data, fd);
1714 /* we don't allow fixed io_uring files */
1715 if (file && io_is_uring_fops(file))
1716 io_req_track_inflight(req);
1720 static void io_queue_async(struct io_kiocb *req, int ret)
1721 __must_hold(&req->ctx->uring_lock)
1723 struct io_kiocb *linked_timeout;
1725 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
1726 io_req_complete_failed(req, ret);
1730 linked_timeout = io_prep_linked_timeout(req);
1732 switch (io_arm_poll_handler(req, 0)) {
1733 case IO_APOLL_READY:
1734 io_req_task_queue(req);
1736 case IO_APOLL_ABORTED:
1738 * Queued up for async execution, worker will release
1739 * submit reference when the iocb is actually submitted.
1741 io_kbuf_recycle(req, 0);
1742 io_queue_iowq(req, NULL);
1749 io_queue_linked_timeout(linked_timeout);
1752 static inline void io_queue_sqe(struct io_kiocb *req)
1753 __must_hold(&req->ctx->uring_lock)
1757 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
1760 * We async punt it if the file wasn't marked NOWAIT, or if the file
1761 * doesn't support non-blocking read/write attempts
1764 io_arm_ltimeout(req);
1766 io_queue_async(req, ret);
1769 static void io_queue_sqe_fallback(struct io_kiocb *req)
1770 __must_hold(&req->ctx->uring_lock)
1772 if (unlikely(req->flags & REQ_F_FAIL)) {
1774 * We don't submit, fail them all, for that replace hardlinks
1775 * with normal links. Extra REQ_F_LINK is tolerated.
1777 req->flags &= ~REQ_F_HARDLINK;
1778 req->flags |= REQ_F_LINK;
1779 io_req_complete_failed(req, req->cqe.res);
1780 } else if (unlikely(req->ctx->drain_active)) {
1783 int ret = io_req_prep_async(req);
1786 io_req_complete_failed(req, ret);
1788 io_queue_iowq(req, NULL);
1793 * Check SQE restrictions (opcode and flags).
1795 * Returns 'true' if SQE is allowed, 'false' otherwise.
1797 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
1798 struct io_kiocb *req,
1799 unsigned int sqe_flags)
1801 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
1804 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
1805 ctx->restrictions.sqe_flags_required)
1808 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
1809 ctx->restrictions.sqe_flags_required))
1815 static void io_init_req_drain(struct io_kiocb *req)
1817 struct io_ring_ctx *ctx = req->ctx;
1818 struct io_kiocb *head = ctx->submit_state.link.head;
1820 ctx->drain_active = true;
1823 * If we need to drain a request in the middle of a link, drain
1824 * the head request and the next request/link after the current
1825 * link. Considering sequential execution of links,
1826 * REQ_F_IO_DRAIN will be maintained for every request of our
1829 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
1830 ctx->drain_next = true;
1834 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
1835 const struct io_uring_sqe *sqe)
1836 __must_hold(&ctx->uring_lock)
1838 const struct io_op_def *def;
1839 unsigned int sqe_flags;
1843 /* req is partially pre-initialised, see io_preinit_req() */
1844 req->opcode = opcode = READ_ONCE(sqe->opcode);
1845 /* same numerical values with corresponding REQ_F_*, safe to copy */
1846 req->flags = sqe_flags = READ_ONCE(sqe->flags);
1847 req->cqe.user_data = READ_ONCE(sqe->user_data);
1849 req->rsrc_node = NULL;
1850 req->task = current;
1852 if (unlikely(opcode >= IORING_OP_LAST)) {
1856 def = &io_op_defs[opcode];
1857 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
1858 /* enforce forwards compatibility on users */
1859 if (sqe_flags & ~SQE_VALID_FLAGS)
1861 if (sqe_flags & IOSQE_BUFFER_SELECT) {
1862 if (!def->buffer_select)
1864 req->buf_index = READ_ONCE(sqe->buf_group);
1866 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
1867 ctx->drain_disabled = true;
1868 if (sqe_flags & IOSQE_IO_DRAIN) {
1869 if (ctx->drain_disabled)
1871 io_init_req_drain(req);
1874 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
1875 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
1877 /* knock it to the slow queue path, will be drained there */
1878 if (ctx->drain_active)
1879 req->flags |= REQ_F_FORCE_ASYNC;
1880 /* if there is no link, we're at "next" request and need to drain */
1881 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
1882 ctx->drain_next = false;
1883 ctx->drain_active = true;
1884 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
1888 if (!def->ioprio && sqe->ioprio)
1890 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
1893 if (def->needs_file) {
1894 struct io_submit_state *state = &ctx->submit_state;
1896 req->cqe.fd = READ_ONCE(sqe->fd);
1899 * Plug now if we have more than 2 IO left after this, and the
1900 * target is potentially a read/write to block based storage.
1902 if (state->need_plug && def->plug) {
1903 state->plug_started = true;
1904 state->need_plug = false;
1905 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
1909 personality = READ_ONCE(sqe->personality);
1913 req->creds = xa_load(&ctx->personalities, personality);
1916 get_cred(req->creds);
1917 ret = security_uring_override_creds(req->creds);
1919 put_cred(req->creds);
1922 req->flags |= REQ_F_CREDS;
1925 return def->prep(req, sqe);
1928 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
1929 struct io_kiocb *req, int ret)
1931 struct io_ring_ctx *ctx = req->ctx;
1932 struct io_submit_link *link = &ctx->submit_state.link;
1933 struct io_kiocb *head = link->head;
1935 trace_io_uring_req_failed(sqe, ctx, req, ret);
1938 * Avoid breaking links in the middle as it renders links with SQPOLL
1939 * unusable. Instead of failing eagerly, continue assembling the link if
1940 * applicable and mark the head with REQ_F_FAIL. The link flushing code
1941 * should find the flag and handle the rest.
1943 req_fail_link_node(req, ret);
1944 if (head && !(head->flags & REQ_F_FAIL))
1945 req_fail_link_node(head, -ECANCELED);
1947 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
1949 link->last->link = req;
1953 io_queue_sqe_fallback(req);
1958 link->last->link = req;
1965 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
1966 const struct io_uring_sqe *sqe)
1967 __must_hold(&ctx->uring_lock)
1969 struct io_submit_link *link = &ctx->submit_state.link;
1972 ret = io_init_req(ctx, req, sqe);
1974 return io_submit_fail_init(sqe, req, ret);
1976 /* don't need @sqe from now on */
1977 trace_io_uring_submit_sqe(ctx, req, req->cqe.user_data, req->opcode,
1979 ctx->flags & IORING_SETUP_SQPOLL);
1982 * If we already have a head request, queue this one for async
1983 * submittal once the head completes. If we don't have a head but
1984 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
1985 * submitted sync once the chain is complete. If none of those
1986 * conditions are true (normal request), then just queue it.
1988 if (unlikely(link->head)) {
1989 ret = io_req_prep_async(req);
1991 return io_submit_fail_init(sqe, req, ret);
1993 trace_io_uring_link(ctx, req, link->head);
1994 link->last->link = req;
1997 if (req->flags & IO_REQ_LINK_FLAGS)
1999 /* last request of the link, flush it */
2002 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2005 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2006 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2007 if (req->flags & IO_REQ_LINK_FLAGS) {
2012 io_queue_sqe_fallback(req);
2022 * Batched submission is done, ensure local IO is flushed out.
2024 static void io_submit_state_end(struct io_ring_ctx *ctx)
2026 struct io_submit_state *state = &ctx->submit_state;
2028 if (unlikely(state->link.head))
2029 io_queue_sqe_fallback(state->link.head);
2030 /* flush only after queuing links as they can generate completions */
2031 io_submit_flush_completions(ctx);
2032 if (state->plug_started)
2033 blk_finish_plug(&state->plug);
2037 * Start submission side cache.
2039 static void io_submit_state_start(struct io_submit_state *state,
2040 unsigned int max_ios)
2042 state->plug_started = false;
2043 state->need_plug = max_ios > 2;
2044 state->submit_nr = max_ios;
2045 /* set only head, no need to init link_last in advance */
2046 state->link.head = NULL;
2049 static void io_commit_sqring(struct io_ring_ctx *ctx)
2051 struct io_rings *rings = ctx->rings;
2054 * Ensure any loads from the SQEs are done at this point,
2055 * since once we write the new head, the application could
2056 * write new data to them.
2058 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2062 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2063 * that is mapped by userspace. This means that care needs to be taken to
2064 * ensure that reads are stable, as we cannot rely on userspace always
2065 * being a good citizen. If members of the sqe are validated and then later
2066 * used, it's important that those reads are done through READ_ONCE() to
2067 * prevent a re-load down the line.
2069 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
2071 unsigned head, mask = ctx->sq_entries - 1;
2072 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2075 * The cached sq head (or cq tail) serves two purposes:
2077 * 1) allows us to batch the cost of updating the user visible
2079 * 2) allows the kernel side to track the head on its own, even
2080 * though the application is the one updating it.
2082 head = READ_ONCE(ctx->sq_array[sq_idx]);
2083 if (likely(head < ctx->sq_entries)) {
2084 /* double index for 128-byte SQEs, twice as long */
2085 if (ctx->flags & IORING_SETUP_SQE128)
2087 return &ctx->sq_sqes[head];
2090 /* drop invalid entries */
2092 WRITE_ONCE(ctx->rings->sq_dropped,
2093 READ_ONCE(ctx->rings->sq_dropped) + 1);
2097 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2098 __must_hold(&ctx->uring_lock)
2100 unsigned int entries = io_sqring_entries(ctx);
2104 if (unlikely(!entries))
2106 /* make sure SQ entry isn't read before tail */
2107 ret = left = min3(nr, ctx->sq_entries, entries);
2108 io_get_task_refs(left);
2109 io_submit_state_start(&ctx->submit_state, left);
2112 const struct io_uring_sqe *sqe;
2113 struct io_kiocb *req;
2115 if (unlikely(!io_alloc_req_refill(ctx)))
2117 req = io_alloc_req(ctx);
2118 sqe = io_get_sqe(ctx);
2119 if (unlikely(!sqe)) {
2120 io_req_add_to_cache(req, ctx);
2125 * Continue submitting even for sqe failure if the
2126 * ring was setup with IORING_SETUP_SUBMIT_ALL
2128 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2129 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2135 if (unlikely(left)) {
2137 /* try again if it submitted nothing and can't allocate a req */
2138 if (!ret && io_req_cache_empty(ctx))
2140 current->io_uring->cached_refs += left;
2143 io_submit_state_end(ctx);
2144 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2145 io_commit_sqring(ctx);
2149 struct io_wait_queue {
2150 struct wait_queue_entry wq;
2151 struct io_ring_ctx *ctx;
2153 unsigned nr_timeouts;
2156 static inline bool io_should_wake(struct io_wait_queue *iowq)
2158 struct io_ring_ctx *ctx = iowq->ctx;
2159 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
2162 * Wake up if we have enough events, or if a timeout occurred since we
2163 * started waiting. For timeouts, we always want to return to userspace,
2164 * regardless of event count.
2166 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2169 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2170 int wake_flags, void *key)
2172 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
2176 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2177 * the task, and the next invocation will do it.
2179 if (io_should_wake(iowq) ||
2180 test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &iowq->ctx->check_cq))
2181 return autoremove_wake_function(curr, mode, wake_flags, key);
2185 int io_run_task_work_sig(void)
2187 if (io_run_task_work())
2189 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
2190 return -ERESTARTSYS;
2191 if (task_sigpending(current))
2196 /* when returns >0, the caller should retry */
2197 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2198 struct io_wait_queue *iowq,
2202 unsigned long check_cq;
2204 /* make sure we run task_work before checking for signals */
2205 ret = io_run_task_work_sig();
2206 if (ret || io_should_wake(iowq))
2209 check_cq = READ_ONCE(ctx->check_cq);
2210 if (unlikely(check_cq)) {
2211 /* let the caller flush overflows, retry */
2212 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2214 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
2217 if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS))
2223 * Wait until events become available, if we don't already have some. The
2224 * application must reap them itself, as they reside on the shared cq ring.
2226 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2227 const sigset_t __user *sig, size_t sigsz,
2228 struct __kernel_timespec __user *uts)
2230 struct io_wait_queue iowq;
2231 struct io_rings *rings = ctx->rings;
2232 ktime_t timeout = KTIME_MAX;
2236 io_cqring_overflow_flush(ctx);
2237 if (io_cqring_events(ctx) >= min_events)
2239 if (!io_run_task_work())
2244 #ifdef CONFIG_COMPAT
2245 if (in_compat_syscall())
2246 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2250 ret = set_user_sigmask(sig, sigsz);
2257 struct timespec64 ts;
2259 if (get_timespec64(&ts, uts))
2261 timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2264 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2265 iowq.wq.private = current;
2266 INIT_LIST_HEAD(&iowq.wq.entry);
2268 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2269 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2271 trace_io_uring_cqring_wait(ctx, min_events);
2273 /* if we can't even flush overflow, don't wait for more */
2274 if (!io_cqring_overflow_flush(ctx)) {
2278 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2279 TASK_INTERRUPTIBLE);
2280 ret = io_cqring_wait_schedule(ctx, &iowq, timeout);
2284 finish_wait(&ctx->cq_wait, &iowq.wq);
2285 restore_saved_sigmask_unless(ret == -EINTR);
2287 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2290 static void io_mem_free(void *ptr)
2297 page = virt_to_head_page(ptr);
2298 if (put_page_testzero(page))
2299 free_compound_page(page);
2302 static void *io_mem_alloc(size_t size)
2304 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2306 return (void *) __get_free_pages(gfp, get_order(size));
2309 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2310 unsigned int cq_entries, size_t *sq_offset)
2312 struct io_rings *rings;
2313 size_t off, sq_array_size;
2315 off = struct_size(rings, cqes, cq_entries);
2316 if (off == SIZE_MAX)
2318 if (ctx->flags & IORING_SETUP_CQE32) {
2319 if (check_shl_overflow(off, 1, &off))
2324 off = ALIGN(off, SMP_CACHE_BYTES);
2332 sq_array_size = array_size(sizeof(u32), sq_entries);
2333 if (sq_array_size == SIZE_MAX)
2336 if (check_add_overflow(off, sq_array_size, &off))
2342 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2343 unsigned int eventfd_async)
2345 struct io_ev_fd *ev_fd;
2346 __s32 __user *fds = arg;
2349 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2350 lockdep_is_held(&ctx->uring_lock));
2354 if (copy_from_user(&fd, fds, sizeof(*fds)))
2357 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2361 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2362 if (IS_ERR(ev_fd->cq_ev_fd)) {
2363 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2367 ev_fd->eventfd_async = eventfd_async;
2368 ctx->has_evfd = true;
2369 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2373 static void io_eventfd_put(struct rcu_head *rcu)
2375 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
2377 eventfd_ctx_put(ev_fd->cq_ev_fd);
2381 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2383 struct io_ev_fd *ev_fd;
2385 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2386 lockdep_is_held(&ctx->uring_lock));
2388 ctx->has_evfd = false;
2389 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2390 call_rcu(&ev_fd->rcu, io_eventfd_put);
2397 static void io_req_caches_free(struct io_ring_ctx *ctx)
2399 struct io_submit_state *state = &ctx->submit_state;
2402 mutex_lock(&ctx->uring_lock);
2403 io_flush_cached_locked_reqs(ctx, state);
2405 while (!io_req_cache_empty(ctx)) {
2406 struct io_wq_work_node *node;
2407 struct io_kiocb *req;
2409 node = wq_stack_extract(&state->free_list);
2410 req = container_of(node, struct io_kiocb, comp_list);
2411 kmem_cache_free(req_cachep, req);
2415 percpu_ref_put_many(&ctx->refs, nr);
2416 mutex_unlock(&ctx->uring_lock);
2419 static void io_flush_apoll_cache(struct io_ring_ctx *ctx)
2421 struct async_poll *apoll;
2423 while (!list_empty(&ctx->apoll_cache)) {
2424 apoll = list_first_entry(&ctx->apoll_cache, struct async_poll,
2426 list_del(&apoll->poll.wait.entry);
2431 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2433 io_sq_thread_finish(ctx);
2435 if (ctx->mm_account) {
2436 mmdrop(ctx->mm_account);
2437 ctx->mm_account = NULL;
2440 io_rsrc_refs_drop(ctx);
2441 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2442 io_wait_rsrc_data(ctx->buf_data);
2443 io_wait_rsrc_data(ctx->file_data);
2445 mutex_lock(&ctx->uring_lock);
2447 __io_sqe_buffers_unregister(ctx);
2449 __io_sqe_files_unregister(ctx);
2451 __io_cqring_overflow_flush(ctx, true);
2452 io_eventfd_unregister(ctx);
2453 io_flush_apoll_cache(ctx);
2454 mutex_unlock(&ctx->uring_lock);
2455 io_destroy_buffers(ctx);
2457 put_cred(ctx->sq_creds);
2459 /* there are no registered resources left, nobody uses it */
2461 io_rsrc_node_destroy(ctx->rsrc_node);
2462 if (ctx->rsrc_backup_node)
2463 io_rsrc_node_destroy(ctx->rsrc_backup_node);
2464 flush_delayed_work(&ctx->rsrc_put_work);
2465 flush_delayed_work(&ctx->fallback_work);
2467 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2468 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
2470 #if defined(CONFIG_UNIX)
2471 if (ctx->ring_sock) {
2472 ctx->ring_sock->file = NULL; /* so that iput() is called */
2473 sock_release(ctx->ring_sock);
2476 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2478 io_mem_free(ctx->rings);
2479 io_mem_free(ctx->sq_sqes);
2481 percpu_ref_exit(&ctx->refs);
2482 free_uid(ctx->user);
2483 io_req_caches_free(ctx);
2485 io_wq_put_hash(ctx->hash_map);
2486 kfree(ctx->cancel_hash);
2487 kfree(ctx->dummy_ubuf);
2489 xa_destroy(&ctx->io_bl_xa);
2493 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2495 struct io_ring_ctx *ctx = file->private_data;
2498 poll_wait(file, &ctx->cq_wait, wait);
2500 * synchronizes with barrier from wq_has_sleeper call in
2504 if (!io_sqring_full(ctx))
2505 mask |= EPOLLOUT | EPOLLWRNORM;
2508 * Don't flush cqring overflow list here, just do a simple check.
2509 * Otherwise there could possible be ABBA deadlock:
2512 * lock(&ctx->uring_lock);
2514 * lock(&ctx->uring_lock);
2517 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2518 * pushs them to do the flush.
2520 if (io_cqring_events(ctx) ||
2521 test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
2522 mask |= EPOLLIN | EPOLLRDNORM;
2527 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2529 const struct cred *creds;
2531 creds = xa_erase(&ctx->personalities, id);
2540 struct io_tctx_exit {
2541 struct callback_head task_work;
2542 struct completion completion;
2543 struct io_ring_ctx *ctx;
2546 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2548 struct io_uring_task *tctx = current->io_uring;
2549 struct io_tctx_exit *work;
2551 work = container_of(cb, struct io_tctx_exit, task_work);
2553 * When @in_idle, we're in cancellation and it's racy to remove the
2554 * node. It'll be removed by the end of cancellation, just ignore it.
2556 if (!atomic_read(&tctx->in_idle))
2557 io_uring_del_tctx_node((unsigned long)work->ctx);
2558 complete(&work->completion);
2561 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2563 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2565 return req->ctx == data;
2568 static __cold void io_ring_exit_work(struct work_struct *work)
2570 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2571 unsigned long timeout = jiffies + HZ * 60 * 5;
2572 unsigned long interval = HZ / 20;
2573 struct io_tctx_exit exit;
2574 struct io_tctx_node *node;
2578 * If we're doing polled IO and end up having requests being
2579 * submitted async (out-of-line), then completions can come in while
2580 * we're waiting for refs to drop. We need to reap these manually,
2581 * as nobody else will be looking for them.
2584 io_uring_try_cancel_requests(ctx, NULL, true);
2586 struct io_sq_data *sqd = ctx->sq_data;
2587 struct task_struct *tsk;
2589 io_sq_thread_park(sqd);
2591 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
2592 io_wq_cancel_cb(tsk->io_uring->io_wq,
2593 io_cancel_ctx_cb, ctx, true);
2594 io_sq_thread_unpark(sqd);
2597 io_req_caches_free(ctx);
2599 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
2600 /* there is little hope left, don't run it too often */
2603 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
2605 init_completion(&exit.completion);
2606 init_task_work(&exit.task_work, io_tctx_exit_cb);
2609 * Some may use context even when all refs and requests have been put,
2610 * and they are free to do so while still holding uring_lock or
2611 * completion_lock, see io_req_task_submit(). Apart from other work,
2612 * this lock/unlock section also waits them to finish.
2614 mutex_lock(&ctx->uring_lock);
2615 while (!list_empty(&ctx->tctx_list)) {
2616 WARN_ON_ONCE(time_after(jiffies, timeout));
2618 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
2620 /* don't spin on a single task if cancellation failed */
2621 list_rotate_left(&ctx->tctx_list);
2622 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
2623 if (WARN_ON_ONCE(ret))
2626 mutex_unlock(&ctx->uring_lock);
2627 wait_for_completion(&exit.completion);
2628 mutex_lock(&ctx->uring_lock);
2630 mutex_unlock(&ctx->uring_lock);
2631 spin_lock(&ctx->completion_lock);
2632 spin_unlock(&ctx->completion_lock);
2634 io_ring_ctx_free(ctx);
2637 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
2639 unsigned long index;
2640 struct creds *creds;
2642 mutex_lock(&ctx->uring_lock);
2643 percpu_ref_kill(&ctx->refs);
2645 __io_cqring_overflow_flush(ctx, true);
2646 xa_for_each(&ctx->personalities, index, creds)
2647 io_unregister_personality(ctx, index);
2648 mutex_unlock(&ctx->uring_lock);
2650 /* failed during ring init, it couldn't have issued any requests */
2652 io_kill_timeouts(ctx, NULL, true);
2653 io_poll_remove_all(ctx, NULL, true);
2654 /* if we failed setting up the ctx, we might not have any rings */
2655 io_iopoll_try_reap_events(ctx);
2658 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
2660 * Use system_unbound_wq to avoid spawning tons of event kworkers
2661 * if we're exiting a ton of rings at the same time. It just adds
2662 * noise and overhead, there's no discernable change in runtime
2663 * over using system_wq.
2665 queue_work(system_unbound_wq, &ctx->exit_work);
2668 static int io_uring_release(struct inode *inode, struct file *file)
2670 struct io_ring_ctx *ctx = file->private_data;
2672 file->private_data = NULL;
2673 io_ring_ctx_wait_and_kill(ctx);
2677 struct io_task_cancel {
2678 struct task_struct *task;
2682 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
2684 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2685 struct io_task_cancel *cancel = data;
2687 return io_match_task_safe(req, cancel->task, cancel->all);
2690 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
2691 struct task_struct *task,
2694 struct io_defer_entry *de;
2697 spin_lock(&ctx->completion_lock);
2698 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
2699 if (io_match_task_safe(de->req, task, cancel_all)) {
2700 list_cut_position(&list, &ctx->defer_list, &de->list);
2704 spin_unlock(&ctx->completion_lock);
2705 if (list_empty(&list))
2708 while (!list_empty(&list)) {
2709 de = list_first_entry(&list, struct io_defer_entry, list);
2710 list_del_init(&de->list);
2711 io_req_complete_failed(de->req, -ECANCELED);
2717 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
2719 struct io_tctx_node *node;
2720 enum io_wq_cancel cret;
2723 mutex_lock(&ctx->uring_lock);
2724 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
2725 struct io_uring_task *tctx = node->task->io_uring;
2728 * io_wq will stay alive while we hold uring_lock, because it's
2729 * killed after ctx nodes, which requires to take the lock.
2731 if (!tctx || !tctx->io_wq)
2733 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
2734 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2736 mutex_unlock(&ctx->uring_lock);
2741 static __cold void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
2742 struct task_struct *task,
2745 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
2746 struct io_uring_task *tctx = task ? task->io_uring : NULL;
2748 /* failed during ring init, it couldn't have issued any requests */
2753 enum io_wq_cancel cret;
2757 ret |= io_uring_try_cancel_iowq(ctx);
2758 } else if (tctx && tctx->io_wq) {
2760 * Cancels requests of all rings, not only @ctx, but
2761 * it's fine as the task is in exit/exec.
2763 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
2765 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2768 /* SQPOLL thread does its own polling */
2769 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
2770 (ctx->sq_data && ctx->sq_data->thread == current)) {
2771 while (!wq_list_empty(&ctx->iopoll_list)) {
2772 io_iopoll_try_reap_events(ctx);
2777 ret |= io_cancel_defer_files(ctx, task, cancel_all);
2778 ret |= io_poll_remove_all(ctx, task, cancel_all);
2779 ret |= io_kill_timeouts(ctx, task, cancel_all);
2781 ret |= io_run_task_work();
2788 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
2791 return atomic_read(&tctx->inflight_tracked);
2792 return percpu_counter_sum(&tctx->inflight);
2796 * Find any io_uring ctx that this task has registered or done IO on, and cancel
2797 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
2799 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
2801 struct io_uring_task *tctx = current->io_uring;
2802 struct io_ring_ctx *ctx;
2806 WARN_ON_ONCE(sqd && sqd->thread != current);
2808 if (!current->io_uring)
2811 io_wq_exit_start(tctx->io_wq);
2813 atomic_inc(&tctx->in_idle);
2815 io_uring_drop_tctx_refs(current);
2816 /* read completions before cancelations */
2817 inflight = tctx_inflight(tctx, !cancel_all);
2822 struct io_tctx_node *node;
2823 unsigned long index;
2825 xa_for_each(&tctx->xa, index, node) {
2826 /* sqpoll task will cancel all its requests */
2827 if (node->ctx->sq_data)
2829 io_uring_try_cancel_requests(node->ctx, current,
2833 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
2834 io_uring_try_cancel_requests(ctx, current,
2838 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
2840 io_uring_drop_tctx_refs(current);
2843 * If we've seen completions, retry without waiting. This
2844 * avoids a race where a completion comes in before we did
2845 * prepare_to_wait().
2847 if (inflight == tctx_inflight(tctx, !cancel_all))
2849 finish_wait(&tctx->wait, &wait);
2852 io_uring_clean_tctx(tctx);
2855 * We shouldn't run task_works after cancel, so just leave
2856 * ->in_idle set for normal exit.
2858 atomic_dec(&tctx->in_idle);
2859 /* for exec all current's requests should be gone, kill tctx */
2860 __io_uring_free(current);
2864 void __io_uring_cancel(bool cancel_all)
2866 io_uring_cancel_generic(cancel_all, NULL);
2869 static void *io_uring_validate_mmap_request(struct file *file,
2870 loff_t pgoff, size_t sz)
2872 struct io_ring_ctx *ctx = file->private_data;
2873 loff_t offset = pgoff << PAGE_SHIFT;
2878 case IORING_OFF_SQ_RING:
2879 case IORING_OFF_CQ_RING:
2882 case IORING_OFF_SQES:
2886 return ERR_PTR(-EINVAL);
2889 page = virt_to_head_page(ptr);
2890 if (sz > page_size(page))
2891 return ERR_PTR(-EINVAL);
2898 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
2900 size_t sz = vma->vm_end - vma->vm_start;
2904 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
2906 return PTR_ERR(ptr);
2908 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
2909 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
2912 #else /* !CONFIG_MMU */
2914 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
2916 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
2919 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
2921 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
2924 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
2925 unsigned long addr, unsigned long len,
2926 unsigned long pgoff, unsigned long flags)
2930 ptr = io_uring_validate_mmap_request(file, pgoff, len);
2932 return PTR_ERR(ptr);
2934 return (unsigned long) ptr;
2937 #endif /* !CONFIG_MMU */
2939 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
2941 if (flags & IORING_ENTER_EXT_ARG) {
2942 struct io_uring_getevents_arg arg;
2944 if (argsz != sizeof(arg))
2946 if (copy_from_user(&arg, argp, sizeof(arg)))
2952 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
2953 struct __kernel_timespec __user **ts,
2954 const sigset_t __user **sig)
2956 struct io_uring_getevents_arg arg;
2959 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
2960 * is just a pointer to the sigset_t.
2962 if (!(flags & IORING_ENTER_EXT_ARG)) {
2963 *sig = (const sigset_t __user *) argp;
2969 * EXT_ARG is set - ensure we agree on the size of it and copy in our
2970 * timespec and sigset_t pointers if good.
2972 if (*argsz != sizeof(arg))
2974 if (copy_from_user(&arg, argp, sizeof(arg)))
2978 *sig = u64_to_user_ptr(arg.sigmask);
2979 *argsz = arg.sigmask_sz;
2980 *ts = u64_to_user_ptr(arg.ts);
2984 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
2985 u32, min_complete, u32, flags, const void __user *, argp,
2988 struct io_ring_ctx *ctx;
2994 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
2995 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
2996 IORING_ENTER_REGISTERED_RING)))
3000 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3001 * need only dereference our task private array to find it.
3003 if (flags & IORING_ENTER_REGISTERED_RING) {
3004 struct io_uring_task *tctx = current->io_uring;
3006 if (!tctx || fd >= IO_RINGFD_REG_MAX)
3008 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3009 f.file = tctx->registered_rings[fd];
3015 if (unlikely(!f.file))
3019 if (unlikely(!io_is_uring_fops(f.file)))
3023 ctx = f.file->private_data;
3024 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
3028 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3032 * For SQ polling, the thread will do all submissions and completions.
3033 * Just return the requested submit count, and wake the thread if
3037 if (ctx->flags & IORING_SETUP_SQPOLL) {
3038 io_cqring_overflow_flush(ctx);
3040 if (unlikely(ctx->sq_data->thread == NULL)) {
3044 if (flags & IORING_ENTER_SQ_WAKEUP)
3045 wake_up(&ctx->sq_data->wait);
3046 if (flags & IORING_ENTER_SQ_WAIT) {
3047 ret = io_sqpoll_wait_sq(ctx);
3052 } else if (to_submit) {
3053 ret = io_uring_add_tctx_node(ctx);
3057 mutex_lock(&ctx->uring_lock);
3058 ret = io_submit_sqes(ctx, to_submit);
3059 if (ret != to_submit) {
3060 mutex_unlock(&ctx->uring_lock);
3063 if ((flags & IORING_ENTER_GETEVENTS) && ctx->syscall_iopoll)
3065 mutex_unlock(&ctx->uring_lock);
3067 if (flags & IORING_ENTER_GETEVENTS) {
3069 if (ctx->syscall_iopoll) {
3071 * We disallow the app entering submit/complete with
3072 * polling, but we still need to lock the ring to
3073 * prevent racing with polled issue that got punted to
3076 mutex_lock(&ctx->uring_lock);
3078 ret2 = io_validate_ext_arg(flags, argp, argsz);
3079 if (likely(!ret2)) {
3080 min_complete = min(min_complete,
3082 ret2 = io_iopoll_check(ctx, min_complete);
3084 mutex_unlock(&ctx->uring_lock);
3086 const sigset_t __user *sig;
3087 struct __kernel_timespec __user *ts;
3089 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3090 if (likely(!ret2)) {
3091 min_complete = min(min_complete,
3093 ret2 = io_cqring_wait(ctx, min_complete, sig,
3102 * EBADR indicates that one or more CQE were dropped.
3103 * Once the user has been informed we can clear the bit
3104 * as they are obviously ok with those drops.
3106 if (unlikely(ret2 == -EBADR))
3107 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3113 percpu_ref_put(&ctx->refs);
3119 static const struct file_operations io_uring_fops = {
3120 .release = io_uring_release,
3121 .mmap = io_uring_mmap,
3123 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3124 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3126 .poll = io_uring_poll,
3127 #ifdef CONFIG_PROC_FS
3128 .show_fdinfo = io_uring_show_fdinfo,
3132 bool io_is_uring_fops(struct file *file)
3134 return file->f_op == &io_uring_fops;
3137 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3138 struct io_uring_params *p)
3140 struct io_rings *rings;
3141 size_t size, sq_array_offset;
3143 /* make sure these are sane, as we already accounted them */
3144 ctx->sq_entries = p->sq_entries;
3145 ctx->cq_entries = p->cq_entries;
3147 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3148 if (size == SIZE_MAX)
3151 rings = io_mem_alloc(size);
3156 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3157 rings->sq_ring_mask = p->sq_entries - 1;
3158 rings->cq_ring_mask = p->cq_entries - 1;
3159 rings->sq_ring_entries = p->sq_entries;
3160 rings->cq_ring_entries = p->cq_entries;
3162 if (p->flags & IORING_SETUP_SQE128)
3163 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3165 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3166 if (size == SIZE_MAX) {
3167 io_mem_free(ctx->rings);
3172 ctx->sq_sqes = io_mem_alloc(size);
3173 if (!ctx->sq_sqes) {
3174 io_mem_free(ctx->rings);
3182 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3186 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3190 ret = io_uring_add_tctx_node(ctx);
3195 fd_install(fd, file);
3200 * Allocate an anonymous fd, this is what constitutes the application
3201 * visible backing of an io_uring instance. The application mmaps this
3202 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3203 * we have to tie this fd to a socket for file garbage collection purposes.
3205 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3208 #if defined(CONFIG_UNIX)
3211 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3214 return ERR_PTR(ret);
3217 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3218 O_RDWR | O_CLOEXEC, NULL);
3219 #if defined(CONFIG_UNIX)
3221 sock_release(ctx->ring_sock);
3222 ctx->ring_sock = NULL;
3224 ctx->ring_sock->file = file;
3230 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3231 struct io_uring_params __user *params)
3233 struct io_ring_ctx *ctx;
3239 if (entries > IORING_MAX_ENTRIES) {
3240 if (!(p->flags & IORING_SETUP_CLAMP))
3242 entries = IORING_MAX_ENTRIES;
3246 * Use twice as many entries for the CQ ring. It's possible for the
3247 * application to drive a higher depth than the size of the SQ ring,
3248 * since the sqes are only used at submission time. This allows for
3249 * some flexibility in overcommitting a bit. If the application has
3250 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3251 * of CQ ring entries manually.
3253 p->sq_entries = roundup_pow_of_two(entries);
3254 if (p->flags & IORING_SETUP_CQSIZE) {
3256 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3257 * to a power-of-two, if it isn't already. We do NOT impose
3258 * any cq vs sq ring sizing.
3262 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3263 if (!(p->flags & IORING_SETUP_CLAMP))
3265 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3267 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3268 if (p->cq_entries < p->sq_entries)
3271 p->cq_entries = 2 * p->sq_entries;
3274 ctx = io_ring_ctx_alloc(p);
3279 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3280 * space applications don't need to do io completion events
3281 * polling again, they can rely on io_sq_thread to do polling
3282 * work, which can reduce cpu usage and uring_lock contention.
3284 if (ctx->flags & IORING_SETUP_IOPOLL &&
3285 !(ctx->flags & IORING_SETUP_SQPOLL))
3286 ctx->syscall_iopoll = 1;
3288 ctx->compat = in_compat_syscall();
3289 if (!capable(CAP_IPC_LOCK))
3290 ctx->user = get_uid(current_user());
3293 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3294 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3297 if (ctx->flags & IORING_SETUP_SQPOLL) {
3298 /* IPI related flags don't make sense with SQPOLL */
3299 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3300 IORING_SETUP_TASKRUN_FLAG))
3302 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3303 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3304 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3306 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
3308 ctx->notify_method = TWA_SIGNAL;
3312 * This is just grabbed for accounting purposes. When a process exits,
3313 * the mm is exited and dropped before the files, hence we need to hang
3314 * on to this mm purely for the purposes of being able to unaccount
3315 * memory (locked/pinned vm). It's not used for anything else.
3317 mmgrab(current->mm);
3318 ctx->mm_account = current->mm;
3320 ret = io_allocate_scq_urings(ctx, p);
3324 ret = io_sq_offload_create(ctx, p);
3327 /* always set a rsrc node */
3328 ret = io_rsrc_node_switch_start(ctx);
3331 io_rsrc_node_switch(ctx, NULL);
3333 memset(&p->sq_off, 0, sizeof(p->sq_off));
3334 p->sq_off.head = offsetof(struct io_rings, sq.head);
3335 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3336 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3337 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3338 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3339 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3340 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3342 memset(&p->cq_off, 0, sizeof(p->cq_off));
3343 p->cq_off.head = offsetof(struct io_rings, cq.head);
3344 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3345 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3346 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3347 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3348 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3349 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3351 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3352 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3353 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3354 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3355 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3356 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3357 IORING_FEAT_LINKED_FILE;
3359 if (copy_to_user(params, p, sizeof(*p))) {
3364 file = io_uring_get_file(ctx);
3366 ret = PTR_ERR(file);
3371 * Install ring fd as the very last thing, so we don't risk someone
3372 * having closed it before we finish setup
3374 ret = io_uring_install_fd(ctx, file);
3376 /* fput will clean it up */
3381 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3384 io_ring_ctx_wait_and_kill(ctx);
3389 * Sets up an aio uring context, and returns the fd. Applications asks for a
3390 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3391 * params structure passed in.
3393 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3395 struct io_uring_params p;
3398 if (copy_from_user(&p, params, sizeof(p)))
3400 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3405 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3406 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3407 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3408 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3409 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3410 IORING_SETUP_SQE128 | IORING_SETUP_CQE32))
3413 return io_uring_create(entries, &p, params);
3416 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3417 struct io_uring_params __user *, params)
3419 return io_uring_setup(entries, params);
3422 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3425 struct io_uring_probe *p;
3429 size = struct_size(p, ops, nr_args);
3430 if (size == SIZE_MAX)
3432 p = kzalloc(size, GFP_KERNEL);
3437 if (copy_from_user(p, arg, size))
3440 if (memchr_inv(p, 0, size))
3443 p->last_op = IORING_OP_LAST - 1;
3444 if (nr_args > IORING_OP_LAST)
3445 nr_args = IORING_OP_LAST;
3447 for (i = 0; i < nr_args; i++) {
3449 if (!io_op_defs[i].not_supported)
3450 p->ops[i].flags = IO_URING_OP_SUPPORTED;
3455 if (copy_to_user(arg, p, size))
3462 static int io_register_personality(struct io_ring_ctx *ctx)
3464 const struct cred *creds;
3468 creds = get_current_cred();
3470 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
3471 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
3479 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
3480 void __user *arg, unsigned int nr_args)
3482 struct io_uring_restriction *res;
3486 /* Restrictions allowed only if rings started disabled */
3487 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3490 /* We allow only a single restrictions registration */
3491 if (ctx->restrictions.registered)
3494 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
3497 size = array_size(nr_args, sizeof(*res));
3498 if (size == SIZE_MAX)
3501 res = memdup_user(arg, size);
3503 return PTR_ERR(res);
3507 for (i = 0; i < nr_args; i++) {
3508 switch (res[i].opcode) {
3509 case IORING_RESTRICTION_REGISTER_OP:
3510 if (res[i].register_op >= IORING_REGISTER_LAST) {
3515 __set_bit(res[i].register_op,
3516 ctx->restrictions.register_op);
3518 case IORING_RESTRICTION_SQE_OP:
3519 if (res[i].sqe_op >= IORING_OP_LAST) {
3524 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
3526 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
3527 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
3529 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
3530 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
3539 /* Reset all restrictions if an error happened */
3541 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
3543 ctx->restrictions.registered = true;
3549 static int io_register_enable_rings(struct io_ring_ctx *ctx)
3551 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3554 if (ctx->restrictions.registered)
3555 ctx->restricted = 1;
3557 ctx->flags &= ~IORING_SETUP_R_DISABLED;
3558 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
3559 wake_up(&ctx->sq_data->wait);
3563 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
3564 void __user *arg, unsigned len)
3566 struct io_uring_task *tctx = current->io_uring;
3567 cpumask_var_t new_mask;
3570 if (!tctx || !tctx->io_wq)
3573 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3576 cpumask_clear(new_mask);
3577 if (len > cpumask_size())
3578 len = cpumask_size();
3580 if (in_compat_syscall()) {
3581 ret = compat_get_bitmap(cpumask_bits(new_mask),
3582 (const compat_ulong_t __user *)arg,
3583 len * 8 /* CHAR_BIT */);
3585 ret = copy_from_user(new_mask, arg, len);
3589 free_cpumask_var(new_mask);
3593 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
3594 free_cpumask_var(new_mask);
3598 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
3600 struct io_uring_task *tctx = current->io_uring;
3602 if (!tctx || !tctx->io_wq)
3605 return io_wq_cpu_affinity(tctx->io_wq, NULL);
3608 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
3610 __must_hold(&ctx->uring_lock)
3612 struct io_tctx_node *node;
3613 struct io_uring_task *tctx = NULL;
3614 struct io_sq_data *sqd = NULL;
3618 if (copy_from_user(new_count, arg, sizeof(new_count)))
3620 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3621 if (new_count[i] > INT_MAX)
3624 if (ctx->flags & IORING_SETUP_SQPOLL) {
3628 * Observe the correct sqd->lock -> ctx->uring_lock
3629 * ordering. Fine to drop uring_lock here, we hold
3632 refcount_inc(&sqd->refs);
3633 mutex_unlock(&ctx->uring_lock);
3634 mutex_lock(&sqd->lock);
3635 mutex_lock(&ctx->uring_lock);
3637 tctx = sqd->thread->io_uring;
3640 tctx = current->io_uring;
3643 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
3645 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3647 ctx->iowq_limits[i] = new_count[i];
3648 ctx->iowq_limits_set = true;
3650 if (tctx && tctx->io_wq) {
3651 ret = io_wq_max_workers(tctx->io_wq, new_count);
3655 memset(new_count, 0, sizeof(new_count));
3659 mutex_unlock(&sqd->lock);
3660 io_put_sq_data(sqd);
3663 if (copy_to_user(arg, new_count, sizeof(new_count)))
3666 /* that's it for SQPOLL, only the SQPOLL task creates requests */
3670 /* now propagate the restriction to all registered users */
3671 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3672 struct io_uring_task *tctx = node->task->io_uring;
3674 if (WARN_ON_ONCE(!tctx->io_wq))
3677 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3678 new_count[i] = ctx->iowq_limits[i];
3679 /* ignore errors, it always returns zero anyway */
3680 (void)io_wq_max_workers(tctx->io_wq, new_count);
3685 mutex_unlock(&sqd->lock);
3686 io_put_sq_data(sqd);
3691 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
3692 void __user *arg, unsigned nr_args)
3693 __releases(ctx->uring_lock)
3694 __acquires(ctx->uring_lock)
3699 * We're inside the ring mutex, if the ref is already dying, then
3700 * someone else killed the ctx or is already going through
3701 * io_uring_register().
3703 if (percpu_ref_is_dying(&ctx->refs))
3706 if (ctx->restricted) {
3707 if (opcode >= IORING_REGISTER_LAST)
3709 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
3710 if (!test_bit(opcode, ctx->restrictions.register_op))
3715 case IORING_REGISTER_BUFFERS:
3719 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
3721 case IORING_UNREGISTER_BUFFERS:
3725 ret = io_sqe_buffers_unregister(ctx);
3727 case IORING_REGISTER_FILES:
3731 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
3733 case IORING_UNREGISTER_FILES:
3737 ret = io_sqe_files_unregister(ctx);
3739 case IORING_REGISTER_FILES_UPDATE:
3740 ret = io_register_files_update(ctx, arg, nr_args);
3742 case IORING_REGISTER_EVENTFD:
3746 ret = io_eventfd_register(ctx, arg, 0);
3748 case IORING_REGISTER_EVENTFD_ASYNC:
3752 ret = io_eventfd_register(ctx, arg, 1);
3754 case IORING_UNREGISTER_EVENTFD:
3758 ret = io_eventfd_unregister(ctx);
3760 case IORING_REGISTER_PROBE:
3762 if (!arg || nr_args > 256)
3764 ret = io_probe(ctx, arg, nr_args);
3766 case IORING_REGISTER_PERSONALITY:
3770 ret = io_register_personality(ctx);
3772 case IORING_UNREGISTER_PERSONALITY:
3776 ret = io_unregister_personality(ctx, nr_args);
3778 case IORING_REGISTER_ENABLE_RINGS:
3782 ret = io_register_enable_rings(ctx);
3784 case IORING_REGISTER_RESTRICTIONS:
3785 ret = io_register_restrictions(ctx, arg, nr_args);
3787 case IORING_REGISTER_FILES2:
3788 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
3790 case IORING_REGISTER_FILES_UPDATE2:
3791 ret = io_register_rsrc_update(ctx, arg, nr_args,
3794 case IORING_REGISTER_BUFFERS2:
3795 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
3797 case IORING_REGISTER_BUFFERS_UPDATE:
3798 ret = io_register_rsrc_update(ctx, arg, nr_args,
3799 IORING_RSRC_BUFFER);
3801 case IORING_REGISTER_IOWQ_AFF:
3803 if (!arg || !nr_args)
3805 ret = io_register_iowq_aff(ctx, arg, nr_args);
3807 case IORING_UNREGISTER_IOWQ_AFF:
3811 ret = io_unregister_iowq_aff(ctx);
3813 case IORING_REGISTER_IOWQ_MAX_WORKERS:
3815 if (!arg || nr_args != 2)
3817 ret = io_register_iowq_max_workers(ctx, arg);
3819 case IORING_REGISTER_RING_FDS:
3820 ret = io_ringfd_register(ctx, arg, nr_args);
3822 case IORING_UNREGISTER_RING_FDS:
3823 ret = io_ringfd_unregister(ctx, arg, nr_args);
3825 case IORING_REGISTER_PBUF_RING:
3827 if (!arg || nr_args != 1)
3829 ret = io_register_pbuf_ring(ctx, arg);
3831 case IORING_UNREGISTER_PBUF_RING:
3833 if (!arg || nr_args != 1)
3835 ret = io_unregister_pbuf_ring(ctx, arg);
3845 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
3846 void __user *, arg, unsigned int, nr_args)
3848 struct io_ring_ctx *ctx;
3857 if (!io_is_uring_fops(f.file))
3860 ctx = f.file->private_data;
3864 mutex_lock(&ctx->uring_lock);
3865 ret = __io_uring_register(ctx, opcode, arg, nr_args);
3866 mutex_unlock(&ctx->uring_lock);
3867 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
3873 static int __init io_uring_init(void)
3875 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
3876 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
3877 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
3880 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
3881 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
3882 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
3883 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
3884 BUILD_BUG_SQE_ELEM(1, __u8, flags);
3885 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
3886 BUILD_BUG_SQE_ELEM(4, __s32, fd);
3887 BUILD_BUG_SQE_ELEM(8, __u64, off);
3888 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
3889 BUILD_BUG_SQE_ELEM(16, __u64, addr);
3890 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
3891 BUILD_BUG_SQE_ELEM(24, __u32, len);
3892 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
3893 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
3894 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
3895 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
3896 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
3897 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
3898 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
3899 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
3900 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
3901 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
3902 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
3903 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
3904 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
3905 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
3906 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
3907 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
3908 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
3909 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
3910 BUILD_BUG_SQE_ELEM(42, __u16, personality);
3911 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
3912 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
3913 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
3915 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
3916 sizeof(struct io_uring_rsrc_update));
3917 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
3918 sizeof(struct io_uring_rsrc_update2));
3920 /* ->buf_index is u16 */
3921 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
3922 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
3923 offsetof(struct io_uring_buf_ring, tail));
3925 /* should fit into one byte */
3926 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
3927 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
3928 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
3930 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
3932 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
3934 io_uring_optable_init();
3936 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
3940 __initcall(io_uring_init);