1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <net/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
49 #include <linux/bits.h>
51 #include <linux/sched/signal.h>
53 #include <linux/file.h>
54 #include <linux/fdtable.h>
56 #include <linux/mman.h>
57 #include <linux/percpu.h>
58 #include <linux/slab.h>
59 #include <linux/bvec.h>
60 #include <linux/net.h>
62 #include <net/af_unix.h>
64 #include <linux/anon_inodes.h>
65 #include <linux/sched/mm.h>
66 #include <linux/uaccess.h>
67 #include <linux/nospec.h>
68 #include <linux/highmem.h>
69 #include <linux/fsnotify.h>
70 #include <linux/fadvise.h>
71 #include <linux/task_work.h>
72 #include <linux/io_uring.h>
73 #include <linux/audit.h>
74 #include <linux/security.h>
76 #define CREATE_TRACE_POINTS
77 #include <trace/events/io_uring.h>
79 #include <uapi/linux/io_uring.h>
83 #include "io_uring_types.h"
96 #define IORING_MAX_ENTRIES 32768
97 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
99 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
100 IORING_REGISTER_LAST + IORING_OP_LAST)
102 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
103 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
105 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
106 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
108 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
109 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
112 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
115 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
117 #define IO_COMPL_BATCH 32
118 #define IO_REQ_CACHE_SIZE 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;
264 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
266 if (!ctx->cancel_hash)
268 __hash_init(ctx->cancel_hash, 1U << hash_bits);
270 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
271 if (!ctx->dummy_ubuf)
273 /* set invalid range, so io_import_fixed() fails meeting it */
274 ctx->dummy_ubuf->ubuf = -1UL;
276 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
277 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
280 ctx->flags = p->flags;
281 init_waitqueue_head(&ctx->sqo_sq_wait);
282 INIT_LIST_HEAD(&ctx->sqd_list);
283 INIT_LIST_HEAD(&ctx->cq_overflow_list);
284 INIT_LIST_HEAD(&ctx->io_buffers_cache);
285 INIT_LIST_HEAD(&ctx->apoll_cache);
286 init_completion(&ctx->ref_comp);
287 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
288 mutex_init(&ctx->uring_lock);
289 init_waitqueue_head(&ctx->cq_wait);
290 spin_lock_init(&ctx->completion_lock);
291 spin_lock_init(&ctx->timeout_lock);
292 INIT_WQ_LIST(&ctx->iopoll_list);
293 INIT_LIST_HEAD(&ctx->io_buffers_pages);
294 INIT_LIST_HEAD(&ctx->io_buffers_comp);
295 INIT_LIST_HEAD(&ctx->defer_list);
296 INIT_LIST_HEAD(&ctx->timeout_list);
297 INIT_LIST_HEAD(&ctx->ltimeout_list);
298 spin_lock_init(&ctx->rsrc_ref_lock);
299 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
300 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
301 init_llist_head(&ctx->rsrc_put_llist);
302 INIT_LIST_HEAD(&ctx->tctx_list);
303 ctx->submit_state.free_list.next = NULL;
304 INIT_WQ_LIST(&ctx->locked_free_list);
305 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
306 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
309 kfree(ctx->dummy_ubuf);
310 kfree(ctx->cancel_hash);
312 xa_destroy(&ctx->io_bl_xa);
317 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
319 struct io_rings *r = ctx->rings;
321 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
325 static bool req_need_defer(struct io_kiocb *req, u32 seq)
327 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
328 struct io_ring_ctx *ctx = req->ctx;
330 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
336 static inline void io_req_track_inflight(struct io_kiocb *req)
338 if (!(req->flags & REQ_F_INFLIGHT)) {
339 req->flags |= REQ_F_INFLIGHT;
340 atomic_inc(&req->task->io_uring->inflight_tracked);
344 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
346 if (WARN_ON_ONCE(!req->link))
349 req->flags &= ~REQ_F_ARM_LTIMEOUT;
350 req->flags |= REQ_F_LINK_TIMEOUT;
352 /* linked timeouts should have two refs once prep'ed */
353 io_req_set_refcount(req);
354 __io_req_set_refcount(req->link, 2);
358 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
360 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
362 return __io_prep_linked_timeout(req);
365 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
367 io_queue_linked_timeout(__io_prep_linked_timeout(req));
370 static inline void io_arm_ltimeout(struct io_kiocb *req)
372 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
373 __io_arm_ltimeout(req);
376 static void io_prep_async_work(struct io_kiocb *req)
378 const struct io_op_def *def = &io_op_defs[req->opcode];
379 struct io_ring_ctx *ctx = req->ctx;
381 if (!(req->flags & REQ_F_CREDS)) {
382 req->flags |= REQ_F_CREDS;
383 req->creds = get_current_cred();
386 req->work.list.next = NULL;
388 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
389 if (req->flags & REQ_F_FORCE_ASYNC)
390 req->work.flags |= IO_WQ_WORK_CONCURRENT;
392 if (req->flags & REQ_F_ISREG) {
393 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
394 io_wq_hash_work(&req->work, file_inode(req->file));
395 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
396 if (def->unbound_nonreg_file)
397 req->work.flags |= IO_WQ_WORK_UNBOUND;
401 static void io_prep_async_link(struct io_kiocb *req)
403 struct io_kiocb *cur;
405 if (req->flags & REQ_F_LINK_TIMEOUT) {
406 struct io_ring_ctx *ctx = req->ctx;
408 spin_lock_irq(&ctx->timeout_lock);
409 io_for_each_link(cur, req)
410 io_prep_async_work(cur);
411 spin_unlock_irq(&ctx->timeout_lock);
413 io_for_each_link(cur, req)
414 io_prep_async_work(cur);
418 void io_queue_iowq(struct io_kiocb *req, bool *dont_use)
420 struct io_kiocb *link = io_prep_linked_timeout(req);
421 struct io_uring_task *tctx = req->task->io_uring;
424 BUG_ON(!tctx->io_wq);
426 /* init ->work of the whole link before punting */
427 io_prep_async_link(req);
430 * Not expected to happen, but if we do have a bug where this _can_
431 * happen, catch it here and ensure the request is marked as
432 * canceled. That will make io-wq go through the usual work cancel
433 * procedure rather than attempt to run this request (or create a new
436 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
437 req->work.flags |= IO_WQ_WORK_CANCEL;
439 trace_io_uring_queue_async_work(req->ctx, req, req->cqe.user_data,
440 req->opcode, req->flags, &req->work,
441 io_wq_is_hashed(&req->work));
442 io_wq_enqueue(tctx->io_wq, &req->work);
444 io_queue_linked_timeout(link);
447 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
449 while (!list_empty(&ctx->defer_list)) {
450 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
451 struct io_defer_entry, list);
453 if (req_need_defer(de->req, de->seq))
455 list_del_init(&de->list);
456 io_req_task_queue(de->req);
461 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
463 if (ctx->off_timeout_used || ctx->drain_active) {
464 spin_lock(&ctx->completion_lock);
465 if (ctx->off_timeout_used)
466 io_flush_timeouts(ctx);
467 if (ctx->drain_active)
468 io_queue_deferred(ctx);
469 io_commit_cqring(ctx);
470 spin_unlock(&ctx->completion_lock);
473 io_eventfd_signal(ctx);
476 static void io_eventfd_signal(struct io_ring_ctx *ctx)
478 struct io_ev_fd *ev_fd;
482 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
485 ev_fd = rcu_dereference(ctx->io_ev_fd);
488 * Check again if ev_fd exists incase an io_eventfd_unregister call
489 * completed between the NULL check of ctx->io_ev_fd at the start of
490 * the function and rcu_read_lock.
492 if (unlikely(!ev_fd))
494 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
497 if (!ev_fd->eventfd_async || io_wq_current_is_worker())
498 eventfd_signal(ev_fd->cq_ev_fd, 1);
504 * This should only get called when at least one event has been posted.
505 * Some applications rely on the eventfd notification count only changing
506 * IFF a new CQE has been added to the CQ ring. There's no depedency on
507 * 1:1 relationship between how many times this function is called (and
508 * hence the eventfd count) and number of CQEs posted to the CQ ring.
510 void io_cqring_ev_posted(struct io_ring_ctx *ctx)
512 if (unlikely(ctx->off_timeout_used || ctx->drain_active ||
514 __io_commit_cqring_flush(ctx);
519 /* Returns true if there are no backlogged entries after the flush */
520 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
522 bool all_flushed, posted;
523 size_t cqe_size = sizeof(struct io_uring_cqe);
525 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
528 if (ctx->flags & IORING_SETUP_CQE32)
532 spin_lock(&ctx->completion_lock);
533 while (!list_empty(&ctx->cq_overflow_list)) {
534 struct io_uring_cqe *cqe = io_get_cqe(ctx);
535 struct io_overflow_cqe *ocqe;
539 ocqe = list_first_entry(&ctx->cq_overflow_list,
540 struct io_overflow_cqe, list);
542 memcpy(cqe, &ocqe->cqe, cqe_size);
544 io_account_cq_overflow(ctx);
547 list_del(&ocqe->list);
551 all_flushed = list_empty(&ctx->cq_overflow_list);
553 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
554 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
557 io_commit_cqring(ctx);
558 spin_unlock(&ctx->completion_lock);
560 io_cqring_ev_posted(ctx);
564 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
568 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
569 /* iopoll syncs against uring_lock, not completion_lock */
570 if (ctx->flags & IORING_SETUP_IOPOLL)
571 mutex_lock(&ctx->uring_lock);
572 ret = __io_cqring_overflow_flush(ctx, false);
573 if (ctx->flags & IORING_SETUP_IOPOLL)
574 mutex_unlock(&ctx->uring_lock);
580 static void __io_put_task(struct task_struct *task, int nr)
582 struct io_uring_task *tctx = task->io_uring;
584 percpu_counter_sub(&tctx->inflight, nr);
585 if (unlikely(atomic_read(&tctx->in_idle)))
586 wake_up(&tctx->wait);
587 put_task_struct_many(task, nr);
590 /* must to be called somewhat shortly after putting a request */
591 static inline void io_put_task(struct task_struct *task, int nr)
593 if (likely(task == current))
594 task->io_uring->cached_refs += nr;
596 __io_put_task(task, nr);
599 static void io_task_refs_refill(struct io_uring_task *tctx)
601 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
603 percpu_counter_add(&tctx->inflight, refill);
604 refcount_add(refill, ¤t->usage);
605 tctx->cached_refs += refill;
608 static inline void io_get_task_refs(int nr)
610 struct io_uring_task *tctx = current->io_uring;
612 tctx->cached_refs -= nr;
613 if (unlikely(tctx->cached_refs < 0))
614 io_task_refs_refill(tctx);
617 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
619 struct io_uring_task *tctx = task->io_uring;
620 unsigned int refs = tctx->cached_refs;
623 tctx->cached_refs = 0;
624 percpu_counter_sub(&tctx->inflight, refs);
625 put_task_struct_many(task, refs);
629 bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data, s32 res,
630 u32 cflags, u64 extra1, u64 extra2)
632 struct io_overflow_cqe *ocqe;
633 size_t ocq_size = sizeof(struct io_overflow_cqe);
634 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
637 ocq_size += sizeof(struct io_uring_cqe);
639 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
640 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
643 * If we're in ring overflow flush mode, or in task cancel mode,
644 * or cannot allocate an overflow entry, then we need to drop it
647 io_account_cq_overflow(ctx);
648 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
651 if (list_empty(&ctx->cq_overflow_list)) {
652 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
653 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
656 ocqe->cqe.user_data = user_data;
658 ocqe->cqe.flags = cflags;
660 ocqe->cqe.big_cqe[0] = extra1;
661 ocqe->cqe.big_cqe[1] = extra2;
663 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
667 bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
670 struct io_uring_cqe *cqe;
673 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
676 * If we can't get a cq entry, userspace overflowed the
677 * submission (by quite a lot). Increment the overflow count in
680 cqe = io_get_cqe(ctx);
682 WRITE_ONCE(cqe->user_data, user_data);
683 WRITE_ONCE(cqe->res, res);
684 WRITE_ONCE(cqe->flags, cflags);
686 if (ctx->flags & IORING_SETUP_CQE32) {
687 WRITE_ONCE(cqe->big_cqe[0], 0);
688 WRITE_ONCE(cqe->big_cqe[1], 0);
692 return io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
695 static void __io_req_complete_put(struct io_kiocb *req)
698 * If we're the last reference to this request, add to our locked
701 if (req_ref_put_and_test(req)) {
702 struct io_ring_ctx *ctx = req->ctx;
704 if (req->flags & IO_REQ_LINK_FLAGS) {
705 if (req->flags & IO_DISARM_MASK)
708 io_req_task_queue(req->link);
712 io_req_put_rsrc(req);
714 * Selected buffer deallocation in io_clean_op() assumes that
715 * we don't hold ->completion_lock. Clean them here to avoid
718 io_put_kbuf_comp(req);
719 io_dismantle_req(req);
720 io_put_task(req->task, 1);
721 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
722 ctx->locked_free_nr++;
726 void __io_req_complete_post(struct io_kiocb *req)
728 if (!(req->flags & REQ_F_CQE_SKIP))
729 __io_fill_cqe_req(req->ctx, req);
730 __io_req_complete_put(req);
733 void io_req_complete_post(struct io_kiocb *req)
735 struct io_ring_ctx *ctx = req->ctx;
737 spin_lock(&ctx->completion_lock);
738 __io_req_complete_post(req);
739 io_commit_cqring(ctx);
740 spin_unlock(&ctx->completion_lock);
741 io_cqring_ev_posted(ctx);
744 inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags)
746 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
747 req->flags |= REQ_F_COMPLETE_INLINE;
749 io_req_complete_post(req);
752 void io_req_complete_failed(struct io_kiocb *req, s32 res)
755 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
756 io_req_complete_post(req);
760 * Don't initialise the fields below on every allocation, but do that in
761 * advance and keep them valid across allocations.
763 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
767 req->async_data = NULL;
768 /* not necessary, but safer to zero */
772 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
773 struct io_submit_state *state)
775 spin_lock(&ctx->completion_lock);
776 wq_list_splice(&ctx->locked_free_list, &state->free_list);
777 ctx->locked_free_nr = 0;
778 spin_unlock(&ctx->completion_lock);
781 static inline bool io_req_cache_empty(struct io_ring_ctx *ctx)
783 return !ctx->submit_state.free_list.next;
787 * A request might get retired back into the request caches even before opcode
788 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
789 * Because of that, io_alloc_req() should be called only under ->uring_lock
790 * and with extra caution to not get a request that is still worked on.
792 static __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
793 __must_hold(&ctx->uring_lock)
795 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
796 void *reqs[IO_REQ_ALLOC_BATCH];
800 * If we have more than a batch's worth of requests in our IRQ side
801 * locked cache, grab the lock and move them over to our submission
804 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
805 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
806 if (!io_req_cache_empty(ctx))
810 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
813 * Bulk alloc is all-or-nothing. If we fail to get a batch,
814 * retry single alloc to be on the safe side.
816 if (unlikely(ret <= 0)) {
817 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
823 percpu_ref_get_many(&ctx->refs, ret);
824 for (i = 0; i < ret; i++) {
825 struct io_kiocb *req = reqs[i];
827 io_preinit_req(req, ctx);
828 io_req_add_to_cache(req, ctx);
833 static inline bool io_alloc_req_refill(struct io_ring_ctx *ctx)
835 if (unlikely(io_req_cache_empty(ctx)))
836 return __io_alloc_req_refill(ctx);
840 static inline struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
842 struct io_wq_work_node *node;
844 node = wq_stack_extract(&ctx->submit_state.free_list);
845 return container_of(node, struct io_kiocb, comp_list);
848 static inline void io_dismantle_req(struct io_kiocb *req)
850 unsigned int flags = req->flags;
852 if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
854 if (!(flags & REQ_F_FIXED_FILE))
855 io_put_file(req->file);
858 __cold void io_free_req(struct io_kiocb *req)
860 struct io_ring_ctx *ctx = req->ctx;
862 io_req_put_rsrc(req);
863 io_dismantle_req(req);
864 io_put_task(req->task, 1);
866 spin_lock(&ctx->completion_lock);
867 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
868 ctx->locked_free_nr++;
869 spin_unlock(&ctx->completion_lock);
872 static void __io_req_find_next_prep(struct io_kiocb *req)
874 struct io_ring_ctx *ctx = req->ctx;
877 spin_lock(&ctx->completion_lock);
878 posted = io_disarm_next(req);
879 io_commit_cqring(ctx);
880 spin_unlock(&ctx->completion_lock);
882 io_cqring_ev_posted(ctx);
885 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
887 struct io_kiocb *nxt;
890 * If LINK is set, we have dependent requests in this chain. If we
891 * didn't fail this request, queue the first one up, moving any other
892 * dependencies to the next request. In case of failure, fail the rest
895 if (unlikely(req->flags & IO_DISARM_MASK))
896 __io_req_find_next_prep(req);
902 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
906 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
907 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
909 io_submit_flush_completions(ctx);
910 mutex_unlock(&ctx->uring_lock);
913 percpu_ref_put(&ctx->refs);
916 static inline void ctx_commit_and_unlock(struct io_ring_ctx *ctx)
918 io_commit_cqring(ctx);
919 spin_unlock(&ctx->completion_lock);
920 io_cqring_ev_posted(ctx);
923 static void handle_prev_tw_list(struct io_wq_work_node *node,
924 struct io_ring_ctx **ctx, bool *uring_locked)
926 if (*ctx && !*uring_locked)
927 spin_lock(&(*ctx)->completion_lock);
930 struct io_wq_work_node *next = node->next;
931 struct io_kiocb *req = container_of(node, struct io_kiocb,
934 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
936 if (req->ctx != *ctx) {
937 if (unlikely(!*uring_locked && *ctx))
938 ctx_commit_and_unlock(*ctx);
940 ctx_flush_and_put(*ctx, uring_locked);
942 /* if not contended, grab and improve batching */
943 *uring_locked = mutex_trylock(&(*ctx)->uring_lock);
944 percpu_ref_get(&(*ctx)->refs);
945 if (unlikely(!*uring_locked))
946 spin_lock(&(*ctx)->completion_lock);
948 if (likely(*uring_locked)) {
949 req->io_task_work.func(req, uring_locked);
951 req->cqe.flags = io_put_kbuf_comp(req);
952 __io_req_complete_post(req);
957 if (unlikely(!*uring_locked))
958 ctx_commit_and_unlock(*ctx);
961 static void handle_tw_list(struct io_wq_work_node *node,
962 struct io_ring_ctx **ctx, bool *locked)
965 struct io_wq_work_node *next = node->next;
966 struct io_kiocb *req = container_of(node, struct io_kiocb,
969 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
971 if (req->ctx != *ctx) {
972 ctx_flush_and_put(*ctx, locked);
974 /* if not contended, grab and improve batching */
975 *locked = mutex_trylock(&(*ctx)->uring_lock);
976 percpu_ref_get(&(*ctx)->refs);
978 req->io_task_work.func(req, locked);
983 void tctx_task_work(struct callback_head *cb)
985 bool uring_locked = false;
986 struct io_ring_ctx *ctx = NULL;
987 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
991 struct io_wq_work_node *node1, *node2;
993 spin_lock_irq(&tctx->task_lock);
994 node1 = tctx->prio_task_list.first;
995 node2 = tctx->task_list.first;
996 INIT_WQ_LIST(&tctx->task_list);
997 INIT_WQ_LIST(&tctx->prio_task_list);
998 if (!node2 && !node1)
999 tctx->task_running = false;
1000 spin_unlock_irq(&tctx->task_lock);
1001 if (!node2 && !node1)
1005 handle_prev_tw_list(node1, &ctx, &uring_locked);
1007 handle_tw_list(node2, &ctx, &uring_locked);
1010 if (data_race(!tctx->task_list.first) &&
1011 data_race(!tctx->prio_task_list.first) && uring_locked)
1012 io_submit_flush_completions(ctx);
1015 ctx_flush_and_put(ctx, &uring_locked);
1017 /* relaxed read is enough as only the task itself sets ->in_idle */
1018 if (unlikely(atomic_read(&tctx->in_idle)))
1019 io_uring_drop_tctx_refs(current);
1022 static void __io_req_task_work_add(struct io_kiocb *req,
1023 struct io_uring_task *tctx,
1024 struct io_wq_work_list *list)
1026 struct io_ring_ctx *ctx = req->ctx;
1027 struct io_wq_work_node *node;
1028 unsigned long flags;
1031 spin_lock_irqsave(&tctx->task_lock, flags);
1032 wq_list_add_tail(&req->io_task_work.node, list);
1033 running = tctx->task_running;
1035 tctx->task_running = true;
1036 spin_unlock_irqrestore(&tctx->task_lock, flags);
1038 /* task_work already pending, we're done */
1042 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1043 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1045 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1048 spin_lock_irqsave(&tctx->task_lock, flags);
1049 tctx->task_running = false;
1050 node = wq_list_merge(&tctx->prio_task_list, &tctx->task_list);
1051 spin_unlock_irqrestore(&tctx->task_lock, flags);
1054 req = container_of(node, struct io_kiocb, io_task_work.node);
1056 if (llist_add(&req->io_task_work.fallback_node,
1057 &req->ctx->fallback_llist))
1058 schedule_delayed_work(&req->ctx->fallback_work, 1);
1062 void io_req_task_work_add(struct io_kiocb *req)
1064 struct io_uring_task *tctx = req->task->io_uring;
1066 __io_req_task_work_add(req, tctx, &tctx->task_list);
1069 void io_req_task_prio_work_add(struct io_kiocb *req)
1071 struct io_uring_task *tctx = req->task->io_uring;
1073 if (req->ctx->flags & IORING_SETUP_SQPOLL)
1074 __io_req_task_work_add(req, tctx, &tctx->prio_task_list);
1076 __io_req_task_work_add(req, tctx, &tctx->task_list);
1079 static void io_req_tw_post(struct io_kiocb *req, bool *locked)
1081 io_req_complete_post(req);
1084 void io_req_tw_post_queue(struct io_kiocb *req, s32 res, u32 cflags)
1086 io_req_set_res(req, res, cflags);
1087 req->io_task_work.func = io_req_tw_post;
1088 io_req_task_work_add(req);
1091 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
1093 /* not needed for normal modes, but SQPOLL depends on it */
1094 io_tw_lock(req->ctx, locked);
1095 io_req_complete_failed(req, req->cqe.res);
1098 void io_req_task_submit(struct io_kiocb *req, bool *locked)
1100 io_tw_lock(req->ctx, locked);
1101 /* req->task == current here, checking PF_EXITING is safe */
1102 if (likely(!(req->task->flags & PF_EXITING)))
1105 io_req_complete_failed(req, -EFAULT);
1108 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1110 io_req_set_res(req, ret, 0);
1111 req->io_task_work.func = io_req_task_cancel;
1112 io_req_task_work_add(req);
1115 void io_req_task_queue(struct io_kiocb *req)
1117 req->io_task_work.func = io_req_task_submit;
1118 io_req_task_work_add(req);
1121 void io_queue_next(struct io_kiocb *req)
1123 struct io_kiocb *nxt = io_req_find_next(req);
1126 io_req_task_queue(nxt);
1129 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1130 __must_hold(&ctx->uring_lock)
1132 struct task_struct *task = NULL;
1136 struct io_kiocb *req = container_of(node, struct io_kiocb,
1139 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1140 if (req->flags & REQ_F_REFCOUNT) {
1141 node = req->comp_list.next;
1142 if (!req_ref_put_and_test(req))
1145 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1146 struct async_poll *apoll = req->apoll;
1148 if (apoll->double_poll)
1149 kfree(apoll->double_poll);
1150 list_add(&apoll->poll.wait.entry,
1152 req->flags &= ~REQ_F_POLLED;
1154 if (req->flags & IO_REQ_LINK_FLAGS)
1156 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1159 if (!(req->flags & REQ_F_FIXED_FILE))
1160 io_put_file(req->file);
1162 io_req_put_rsrc_locked(req, ctx);
1164 if (req->task != task) {
1166 io_put_task(task, task_refs);
1171 node = req->comp_list.next;
1172 io_req_add_to_cache(req, ctx);
1176 io_put_task(task, task_refs);
1179 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1180 __must_hold(&ctx->uring_lock)
1182 struct io_wq_work_node *node, *prev;
1183 struct io_submit_state *state = &ctx->submit_state;
1185 if (state->flush_cqes) {
1186 spin_lock(&ctx->completion_lock);
1187 wq_list_for_each(node, prev, &state->compl_reqs) {
1188 struct io_kiocb *req = container_of(node, struct io_kiocb,
1191 if (!(req->flags & REQ_F_CQE_SKIP))
1192 __io_fill_cqe_req(ctx, req);
1195 io_commit_cqring(ctx);
1196 spin_unlock(&ctx->completion_lock);
1197 io_cqring_ev_posted(ctx);
1198 state->flush_cqes = false;
1201 io_free_batch_list(ctx, state->compl_reqs.first);
1202 INIT_WQ_LIST(&state->compl_reqs);
1206 * Drop reference to request, return next in chain (if there is one) if this
1207 * was the last reference to this request.
1209 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1211 struct io_kiocb *nxt = NULL;
1213 if (req_ref_put_and_test(req)) {
1214 if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
1215 nxt = io_req_find_next(req);
1221 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1223 /* See comment at the top of this file */
1225 return __io_cqring_events(ctx);
1229 * We can't just wait for polled events to come to us, we have to actively
1230 * find and complete them.
1232 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1234 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1237 mutex_lock(&ctx->uring_lock);
1238 while (!wq_list_empty(&ctx->iopoll_list)) {
1239 /* let it sleep and repeat later if can't complete a request */
1240 if (io_do_iopoll(ctx, true) == 0)
1243 * Ensure we allow local-to-the-cpu processing to take place,
1244 * in this case we need to ensure that we reap all events.
1245 * Also let task_work, etc. to progress by releasing the mutex
1247 if (need_resched()) {
1248 mutex_unlock(&ctx->uring_lock);
1250 mutex_lock(&ctx->uring_lock);
1253 mutex_unlock(&ctx->uring_lock);
1256 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1258 unsigned int nr_events = 0;
1260 unsigned long check_cq;
1262 check_cq = READ_ONCE(ctx->check_cq);
1263 if (unlikely(check_cq)) {
1264 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1265 __io_cqring_overflow_flush(ctx, false);
1267 * Similarly do not spin if we have not informed the user of any
1270 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1274 * Don't enter poll loop if we already have events pending.
1275 * If we do, we can potentially be spinning for commands that
1276 * already triggered a CQE (eg in error).
1278 if (io_cqring_events(ctx))
1283 * If a submit got punted to a workqueue, we can have the
1284 * application entering polling for a command before it gets
1285 * issued. That app will hold the uring_lock for the duration
1286 * of the poll right here, so we need to take a breather every
1287 * now and then to ensure that the issue has a chance to add
1288 * the poll to the issued list. Otherwise we can spin here
1289 * forever, while the workqueue is stuck trying to acquire the
1292 if (wq_list_empty(&ctx->iopoll_list)) {
1293 u32 tail = ctx->cached_cq_tail;
1295 mutex_unlock(&ctx->uring_lock);
1297 mutex_lock(&ctx->uring_lock);
1299 /* some requests don't go through iopoll_list */
1300 if (tail != ctx->cached_cq_tail ||
1301 wq_list_empty(&ctx->iopoll_list))
1304 ret = io_do_iopoll(ctx, !min);
1309 } while (nr_events < min && !need_resched());
1313 inline void io_req_task_complete(struct io_kiocb *req, bool *locked)
1316 req->cqe.flags |= io_put_kbuf(req, 0);
1317 io_req_add_compl_list(req);
1319 req->cqe.flags |= io_put_kbuf(req, IO_URING_F_UNLOCKED);
1320 io_req_complete_post(req);
1325 * After the iocb has been issued, it's safe to be found on the poll list.
1326 * Adding the kiocb to the list AFTER submission ensures that we don't
1327 * find it from a io_do_iopoll() thread before the issuer is done
1328 * accessing the kiocb cookie.
1330 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1332 struct io_ring_ctx *ctx = req->ctx;
1333 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1335 /* workqueue context doesn't hold uring_lock, grab it now */
1336 if (unlikely(needs_lock))
1337 mutex_lock(&ctx->uring_lock);
1340 * Track whether we have multiple files in our lists. This will impact
1341 * how we do polling eventually, not spinning if we're on potentially
1342 * different devices.
1344 if (wq_list_empty(&ctx->iopoll_list)) {
1345 ctx->poll_multi_queue = false;
1346 } else if (!ctx->poll_multi_queue) {
1347 struct io_kiocb *list_req;
1349 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1351 if (list_req->file != req->file)
1352 ctx->poll_multi_queue = true;
1356 * For fast devices, IO may have already completed. If it has, add
1357 * it to the front so we find it first.
1359 if (READ_ONCE(req->iopoll_completed))
1360 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1362 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1364 if (unlikely(needs_lock)) {
1366 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1367 * in sq thread task context or in io worker task context. If
1368 * current task context is sq thread, we don't need to check
1369 * whether should wake up sq thread.
1371 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1372 wq_has_sleeper(&ctx->sq_data->wait))
1373 wake_up(&ctx->sq_data->wait);
1375 mutex_unlock(&ctx->uring_lock);
1379 static bool io_bdev_nowait(struct block_device *bdev)
1381 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
1385 * If we tracked the file through the SCM inflight mechanism, we could support
1386 * any file. For now, just ensure that anything potentially problematic is done
1389 static bool __io_file_supports_nowait(struct file *file, umode_t mode)
1391 if (S_ISBLK(mode)) {
1392 if (IS_ENABLED(CONFIG_BLOCK) &&
1393 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
1399 if (S_ISREG(mode)) {
1400 if (IS_ENABLED(CONFIG_BLOCK) &&
1401 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
1402 !io_is_uring_fops(file))
1407 /* any ->read/write should understand O_NONBLOCK */
1408 if (file->f_flags & O_NONBLOCK)
1410 return file->f_mode & FMODE_NOWAIT;
1414 * If we tracked the file through the SCM inflight mechanism, we could support
1415 * any file. For now, just ensure that anything potentially problematic is done
1418 unsigned int io_file_get_flags(struct file *file)
1420 umode_t mode = file_inode(file)->i_mode;
1421 unsigned int res = 0;
1425 if (__io_file_supports_nowait(file, mode))
1427 if (io_file_need_scm(file))
1432 bool io_alloc_async_data(struct io_kiocb *req)
1434 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
1435 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
1436 if (req->async_data) {
1437 req->flags |= REQ_F_ASYNC_DATA;
1443 int io_req_prep_async(struct io_kiocb *req)
1445 const struct io_op_def *def = &io_op_defs[req->opcode];
1447 /* assign early for deferred execution for non-fixed file */
1448 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE))
1449 req->file = io_file_get_normal(req, req->cqe.fd);
1450 if (!def->prep_async)
1452 if (WARN_ON_ONCE(req_has_async_data(req)))
1454 if (io_alloc_async_data(req))
1457 return def->prep_async(req);
1460 static u32 io_get_sequence(struct io_kiocb *req)
1462 u32 seq = req->ctx->cached_sq_head;
1463 struct io_kiocb *cur;
1465 /* need original cached_sq_head, but it was increased for each req */
1466 io_for_each_link(cur, req)
1471 static __cold void io_drain_req(struct io_kiocb *req)
1473 struct io_ring_ctx *ctx = req->ctx;
1474 struct io_defer_entry *de;
1476 u32 seq = io_get_sequence(req);
1478 /* Still need defer if there is pending req in defer list. */
1479 spin_lock(&ctx->completion_lock);
1480 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1481 spin_unlock(&ctx->completion_lock);
1483 ctx->drain_active = false;
1484 io_req_task_queue(req);
1487 spin_unlock(&ctx->completion_lock);
1489 ret = io_req_prep_async(req);
1492 io_req_complete_failed(req, ret);
1495 io_prep_async_link(req);
1496 de = kmalloc(sizeof(*de), GFP_KERNEL);
1502 spin_lock(&ctx->completion_lock);
1503 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1504 spin_unlock(&ctx->completion_lock);
1509 trace_io_uring_defer(ctx, req, req->cqe.user_data, req->opcode);
1512 list_add_tail(&de->list, &ctx->defer_list);
1513 spin_unlock(&ctx->completion_lock);
1516 static void io_clean_op(struct io_kiocb *req)
1518 if (req->flags & REQ_F_BUFFER_SELECTED) {
1519 spin_lock(&req->ctx->completion_lock);
1520 io_put_kbuf_comp(req);
1521 spin_unlock(&req->ctx->completion_lock);
1524 if (req->flags & REQ_F_NEED_CLEANUP) {
1525 const struct io_op_def *def = &io_op_defs[req->opcode];
1530 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1531 kfree(req->apoll->double_poll);
1535 if (req->flags & REQ_F_INFLIGHT) {
1536 struct io_uring_task *tctx = req->task->io_uring;
1538 atomic_dec(&tctx->inflight_tracked);
1540 if (req->flags & REQ_F_CREDS)
1541 put_cred(req->creds);
1542 if (req->flags & REQ_F_ASYNC_DATA) {
1543 kfree(req->async_data);
1544 req->async_data = NULL;
1546 req->flags &= ~IO_REQ_CLEAN_FLAGS;
1549 static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags)
1551 if (req->file || !io_op_defs[req->opcode].needs_file)
1554 if (req->flags & REQ_F_FIXED_FILE)
1555 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1557 req->file = io_file_get_normal(req, req->cqe.fd);
1562 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1564 const struct io_op_def *def = &io_op_defs[req->opcode];
1565 const struct cred *creds = NULL;
1568 if (unlikely(!io_assign_file(req, issue_flags)))
1571 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1572 creds = override_creds(req->creds);
1574 if (!def->audit_skip)
1575 audit_uring_entry(req->opcode);
1577 ret = def->issue(req, issue_flags);
1579 if (!def->audit_skip)
1580 audit_uring_exit(!ret, ret);
1583 revert_creds(creds);
1586 __io_req_complete(req, issue_flags);
1587 else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1590 /* If the op doesn't have a file, we're not polling for it */
1591 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && req->file)
1592 io_iopoll_req_issued(req, issue_flags);
1597 int io_poll_issue(struct io_kiocb *req, bool *locked)
1599 io_tw_lock(req->ctx, locked);
1600 if (unlikely(req->task->flags & PF_EXITING))
1602 return io_issue_sqe(req, IO_URING_F_NONBLOCK);
1605 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1607 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1609 req = io_put_req_find_next(req);
1610 return req ? &req->work : NULL;
1613 void io_wq_submit_work(struct io_wq_work *work)
1615 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1616 const struct io_op_def *def = &io_op_defs[req->opcode];
1617 unsigned int issue_flags = IO_URING_F_UNLOCKED;
1618 bool needs_poll = false;
1619 int ret = 0, err = -ECANCELED;
1621 /* one will be dropped by ->io_free_work() after returning to io-wq */
1622 if (!(req->flags & REQ_F_REFCOUNT))
1623 __io_req_set_refcount(req, 2);
1627 io_arm_ltimeout(req);
1629 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1630 if (work->flags & IO_WQ_WORK_CANCEL) {
1632 io_req_task_queue_fail(req, err);
1635 if (!io_assign_file(req, issue_flags)) {
1637 work->flags |= IO_WQ_WORK_CANCEL;
1641 if (req->flags & REQ_F_FORCE_ASYNC) {
1642 bool opcode_poll = def->pollin || def->pollout;
1644 if (opcode_poll && file_can_poll(req->file)) {
1646 issue_flags |= IO_URING_F_NONBLOCK;
1651 ret = io_issue_sqe(req, issue_flags);
1655 * We can get EAGAIN for iopolled IO even though we're
1656 * forcing a sync submission from here, since we can't
1657 * wait for request slots on the block side.
1660 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1666 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1668 /* aborted or ready, in either case retry blocking */
1670 issue_flags &= ~IO_URING_F_NONBLOCK;
1673 /* avoid locking problems by failing it from a clean context */
1675 io_req_task_queue_fail(req, ret);
1678 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1679 unsigned int issue_flags)
1681 struct io_ring_ctx *ctx = req->ctx;
1682 struct file *file = NULL;
1683 unsigned long file_ptr;
1685 io_ring_submit_lock(ctx, issue_flags);
1687 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
1689 fd = array_index_nospec(fd, ctx->nr_user_files);
1690 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
1691 file = (struct file *) (file_ptr & FFS_MASK);
1692 file_ptr &= ~FFS_MASK;
1693 /* mask in overlapping REQ_F and FFS bits */
1694 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
1695 io_req_set_rsrc_node(req, ctx, 0);
1696 WARN_ON_ONCE(file && !test_bit(fd, ctx->file_table.bitmap));
1698 io_ring_submit_unlock(ctx, issue_flags);
1702 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
1704 struct file *file = fget(fd);
1706 trace_io_uring_file_get(req->ctx, req, req->cqe.user_data, fd);
1708 /* we don't allow fixed io_uring files */
1709 if (file && io_is_uring_fops(file))
1710 io_req_track_inflight(req);
1714 static void io_queue_async(struct io_kiocb *req, int ret)
1715 __must_hold(&req->ctx->uring_lock)
1717 struct io_kiocb *linked_timeout;
1719 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
1720 io_req_complete_failed(req, ret);
1724 linked_timeout = io_prep_linked_timeout(req);
1726 switch (io_arm_poll_handler(req, 0)) {
1727 case IO_APOLL_READY:
1728 io_req_task_queue(req);
1730 case IO_APOLL_ABORTED:
1732 * Queued up for async execution, worker will release
1733 * submit reference when the iocb is actually submitted.
1735 io_kbuf_recycle(req, 0);
1736 io_queue_iowq(req, NULL);
1743 io_queue_linked_timeout(linked_timeout);
1746 static inline void io_queue_sqe(struct io_kiocb *req)
1747 __must_hold(&req->ctx->uring_lock)
1751 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
1753 if (req->flags & REQ_F_COMPLETE_INLINE) {
1754 io_req_add_compl_list(req);
1758 * We async punt it if the file wasn't marked NOWAIT, or if the file
1759 * doesn't support non-blocking read/write attempts
1762 io_arm_ltimeout(req);
1764 io_queue_async(req, ret);
1767 static void io_queue_sqe_fallback(struct io_kiocb *req)
1768 __must_hold(&req->ctx->uring_lock)
1770 if (unlikely(req->flags & REQ_F_FAIL)) {
1772 * We don't submit, fail them all, for that replace hardlinks
1773 * with normal links. Extra REQ_F_LINK is tolerated.
1775 req->flags &= ~REQ_F_HARDLINK;
1776 req->flags |= REQ_F_LINK;
1777 io_req_complete_failed(req, req->cqe.res);
1778 } else if (unlikely(req->ctx->drain_active)) {
1781 int ret = io_req_prep_async(req);
1784 io_req_complete_failed(req, ret);
1786 io_queue_iowq(req, NULL);
1791 * Check SQE restrictions (opcode and flags).
1793 * Returns 'true' if SQE is allowed, 'false' otherwise.
1795 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
1796 struct io_kiocb *req,
1797 unsigned int sqe_flags)
1799 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
1802 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
1803 ctx->restrictions.sqe_flags_required)
1806 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
1807 ctx->restrictions.sqe_flags_required))
1813 static void io_init_req_drain(struct io_kiocb *req)
1815 struct io_ring_ctx *ctx = req->ctx;
1816 struct io_kiocb *head = ctx->submit_state.link.head;
1818 ctx->drain_active = true;
1821 * If we need to drain a request in the middle of a link, drain
1822 * the head request and the next request/link after the current
1823 * link. Considering sequential execution of links,
1824 * REQ_F_IO_DRAIN will be maintained for every request of our
1827 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
1828 ctx->drain_next = true;
1832 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
1833 const struct io_uring_sqe *sqe)
1834 __must_hold(&ctx->uring_lock)
1836 const struct io_op_def *def;
1837 unsigned int sqe_flags;
1841 /* req is partially pre-initialised, see io_preinit_req() */
1842 req->opcode = opcode = READ_ONCE(sqe->opcode);
1843 /* same numerical values with corresponding REQ_F_*, safe to copy */
1844 req->flags = sqe_flags = READ_ONCE(sqe->flags);
1845 req->cqe.user_data = READ_ONCE(sqe->user_data);
1847 req->rsrc_node = NULL;
1848 req->task = current;
1850 if (unlikely(opcode >= IORING_OP_LAST)) {
1854 def = &io_op_defs[opcode];
1855 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
1856 /* enforce forwards compatibility on users */
1857 if (sqe_flags & ~SQE_VALID_FLAGS)
1859 if (sqe_flags & IOSQE_BUFFER_SELECT) {
1860 if (!def->buffer_select)
1862 req->buf_index = READ_ONCE(sqe->buf_group);
1864 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
1865 ctx->drain_disabled = true;
1866 if (sqe_flags & IOSQE_IO_DRAIN) {
1867 if (ctx->drain_disabled)
1869 io_init_req_drain(req);
1872 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
1873 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
1875 /* knock it to the slow queue path, will be drained there */
1876 if (ctx->drain_active)
1877 req->flags |= REQ_F_FORCE_ASYNC;
1878 /* if there is no link, we're at "next" request and need to drain */
1879 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
1880 ctx->drain_next = false;
1881 ctx->drain_active = true;
1882 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
1886 if (!def->ioprio && sqe->ioprio)
1888 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
1891 if (def->needs_file) {
1892 struct io_submit_state *state = &ctx->submit_state;
1894 req->cqe.fd = READ_ONCE(sqe->fd);
1897 * Plug now if we have more than 2 IO left after this, and the
1898 * target is potentially a read/write to block based storage.
1900 if (state->need_plug && def->plug) {
1901 state->plug_started = true;
1902 state->need_plug = false;
1903 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
1907 personality = READ_ONCE(sqe->personality);
1911 req->creds = xa_load(&ctx->personalities, personality);
1914 get_cred(req->creds);
1915 ret = security_uring_override_creds(req->creds);
1917 put_cred(req->creds);
1920 req->flags |= REQ_F_CREDS;
1923 return def->prep(req, sqe);
1926 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
1927 struct io_kiocb *req, int ret)
1929 struct io_ring_ctx *ctx = req->ctx;
1930 struct io_submit_link *link = &ctx->submit_state.link;
1931 struct io_kiocb *head = link->head;
1933 trace_io_uring_req_failed(sqe, ctx, req, ret);
1936 * Avoid breaking links in the middle as it renders links with SQPOLL
1937 * unusable. Instead of failing eagerly, continue assembling the link if
1938 * applicable and mark the head with REQ_F_FAIL. The link flushing code
1939 * should find the flag and handle the rest.
1941 req_fail_link_node(req, ret);
1942 if (head && !(head->flags & REQ_F_FAIL))
1943 req_fail_link_node(head, -ECANCELED);
1945 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
1947 link->last->link = req;
1951 io_queue_sqe_fallback(req);
1956 link->last->link = req;
1963 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
1964 const struct io_uring_sqe *sqe)
1965 __must_hold(&ctx->uring_lock)
1967 struct io_submit_link *link = &ctx->submit_state.link;
1970 ret = io_init_req(ctx, req, sqe);
1972 return io_submit_fail_init(sqe, req, ret);
1974 /* don't need @sqe from now on */
1975 trace_io_uring_submit_sqe(ctx, req, req->cqe.user_data, req->opcode,
1977 ctx->flags & IORING_SETUP_SQPOLL);
1980 * If we already have a head request, queue this one for async
1981 * submittal once the head completes. If we don't have a head but
1982 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
1983 * submitted sync once the chain is complete. If none of those
1984 * conditions are true (normal request), then just queue it.
1986 if (unlikely(link->head)) {
1987 ret = io_req_prep_async(req);
1989 return io_submit_fail_init(sqe, req, ret);
1991 trace_io_uring_link(ctx, req, link->head);
1992 link->last->link = req;
1995 if (req->flags & IO_REQ_LINK_FLAGS)
1997 /* last request of the link, flush it */
2000 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2003 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2004 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2005 if (req->flags & IO_REQ_LINK_FLAGS) {
2010 io_queue_sqe_fallback(req);
2020 * Batched submission is done, ensure local IO is flushed out.
2022 static void io_submit_state_end(struct io_ring_ctx *ctx)
2024 struct io_submit_state *state = &ctx->submit_state;
2026 if (unlikely(state->link.head))
2027 io_queue_sqe_fallback(state->link.head);
2028 /* flush only after queuing links as they can generate completions */
2029 io_submit_flush_completions(ctx);
2030 if (state->plug_started)
2031 blk_finish_plug(&state->plug);
2035 * Start submission side cache.
2037 static void io_submit_state_start(struct io_submit_state *state,
2038 unsigned int max_ios)
2040 state->plug_started = false;
2041 state->need_plug = max_ios > 2;
2042 state->submit_nr = max_ios;
2043 /* set only head, no need to init link_last in advance */
2044 state->link.head = NULL;
2047 static void io_commit_sqring(struct io_ring_ctx *ctx)
2049 struct io_rings *rings = ctx->rings;
2052 * Ensure any loads from the SQEs are done at this point,
2053 * since once we write the new head, the application could
2054 * write new data to them.
2056 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2060 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2061 * that is mapped by userspace. This means that care needs to be taken to
2062 * ensure that reads are stable, as we cannot rely on userspace always
2063 * being a good citizen. If members of the sqe are validated and then later
2064 * used, it's important that those reads are done through READ_ONCE() to
2065 * prevent a re-load down the line.
2067 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
2069 unsigned head, mask = ctx->sq_entries - 1;
2070 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2073 * The cached sq head (or cq tail) serves two purposes:
2075 * 1) allows us to batch the cost of updating the user visible
2077 * 2) allows the kernel side to track the head on its own, even
2078 * though the application is the one updating it.
2080 head = READ_ONCE(ctx->sq_array[sq_idx]);
2081 if (likely(head < ctx->sq_entries)) {
2082 /* double index for 128-byte SQEs, twice as long */
2083 if (ctx->flags & IORING_SETUP_SQE128)
2085 return &ctx->sq_sqes[head];
2088 /* drop invalid entries */
2090 WRITE_ONCE(ctx->rings->sq_dropped,
2091 READ_ONCE(ctx->rings->sq_dropped) + 1);
2095 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2096 __must_hold(&ctx->uring_lock)
2098 unsigned int entries = io_sqring_entries(ctx);
2102 if (unlikely(!entries))
2104 /* make sure SQ entry isn't read before tail */
2105 ret = left = min3(nr, ctx->sq_entries, entries);
2106 io_get_task_refs(left);
2107 io_submit_state_start(&ctx->submit_state, left);
2110 const struct io_uring_sqe *sqe;
2111 struct io_kiocb *req;
2113 if (unlikely(!io_alloc_req_refill(ctx)))
2115 req = io_alloc_req(ctx);
2116 sqe = io_get_sqe(ctx);
2117 if (unlikely(!sqe)) {
2118 io_req_add_to_cache(req, ctx);
2123 * Continue submitting even for sqe failure if the
2124 * ring was setup with IORING_SETUP_SUBMIT_ALL
2126 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2127 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2133 if (unlikely(left)) {
2135 /* try again if it submitted nothing and can't allocate a req */
2136 if (!ret && io_req_cache_empty(ctx))
2138 current->io_uring->cached_refs += left;
2141 io_submit_state_end(ctx);
2142 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2143 io_commit_sqring(ctx);
2147 struct io_wait_queue {
2148 struct wait_queue_entry wq;
2149 struct io_ring_ctx *ctx;
2151 unsigned nr_timeouts;
2154 static inline bool io_should_wake(struct io_wait_queue *iowq)
2156 struct io_ring_ctx *ctx = iowq->ctx;
2157 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
2160 * Wake up if we have enough events, or if a timeout occurred since we
2161 * started waiting. For timeouts, we always want to return to userspace,
2162 * regardless of event count.
2164 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2167 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2168 int wake_flags, void *key)
2170 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
2174 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2175 * the task, and the next invocation will do it.
2177 if (io_should_wake(iowq) ||
2178 test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &iowq->ctx->check_cq))
2179 return autoremove_wake_function(curr, mode, wake_flags, key);
2183 int io_run_task_work_sig(void)
2185 if (io_run_task_work())
2187 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
2188 return -ERESTARTSYS;
2189 if (task_sigpending(current))
2194 /* when returns >0, the caller should retry */
2195 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2196 struct io_wait_queue *iowq,
2200 unsigned long check_cq;
2202 /* make sure we run task_work before checking for signals */
2203 ret = io_run_task_work_sig();
2204 if (ret || io_should_wake(iowq))
2207 check_cq = READ_ONCE(ctx->check_cq);
2208 if (unlikely(check_cq)) {
2209 /* let the caller flush overflows, retry */
2210 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2212 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
2215 if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS))
2221 * Wait until events become available, if we don't already have some. The
2222 * application must reap them itself, as they reside on the shared cq ring.
2224 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2225 const sigset_t __user *sig, size_t sigsz,
2226 struct __kernel_timespec __user *uts)
2228 struct io_wait_queue iowq;
2229 struct io_rings *rings = ctx->rings;
2230 ktime_t timeout = KTIME_MAX;
2234 io_cqring_overflow_flush(ctx);
2235 if (io_cqring_events(ctx) >= min_events)
2237 if (!io_run_task_work())
2242 #ifdef CONFIG_COMPAT
2243 if (in_compat_syscall())
2244 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2248 ret = set_user_sigmask(sig, sigsz);
2255 struct timespec64 ts;
2257 if (get_timespec64(&ts, uts))
2259 timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2262 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2263 iowq.wq.private = current;
2264 INIT_LIST_HEAD(&iowq.wq.entry);
2266 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2267 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2269 trace_io_uring_cqring_wait(ctx, min_events);
2271 /* if we can't even flush overflow, don't wait for more */
2272 if (!io_cqring_overflow_flush(ctx)) {
2276 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2277 TASK_INTERRUPTIBLE);
2278 ret = io_cqring_wait_schedule(ctx, &iowq, timeout);
2282 finish_wait(&ctx->cq_wait, &iowq.wq);
2283 restore_saved_sigmask_unless(ret == -EINTR);
2285 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2288 static void io_mem_free(void *ptr)
2295 page = virt_to_head_page(ptr);
2296 if (put_page_testzero(page))
2297 free_compound_page(page);
2300 static void *io_mem_alloc(size_t size)
2302 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2304 return (void *) __get_free_pages(gfp, get_order(size));
2307 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2308 unsigned int cq_entries, size_t *sq_offset)
2310 struct io_rings *rings;
2311 size_t off, sq_array_size;
2313 off = struct_size(rings, cqes, cq_entries);
2314 if (off == SIZE_MAX)
2316 if (ctx->flags & IORING_SETUP_CQE32) {
2317 if (check_shl_overflow(off, 1, &off))
2322 off = ALIGN(off, SMP_CACHE_BYTES);
2330 sq_array_size = array_size(sizeof(u32), sq_entries);
2331 if (sq_array_size == SIZE_MAX)
2334 if (check_add_overflow(off, sq_array_size, &off))
2340 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2341 unsigned int eventfd_async)
2343 struct io_ev_fd *ev_fd;
2344 __s32 __user *fds = arg;
2347 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2348 lockdep_is_held(&ctx->uring_lock));
2352 if (copy_from_user(&fd, fds, sizeof(*fds)))
2355 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2359 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2360 if (IS_ERR(ev_fd->cq_ev_fd)) {
2361 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2365 ev_fd->eventfd_async = eventfd_async;
2366 ctx->has_evfd = true;
2367 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2371 static void io_eventfd_put(struct rcu_head *rcu)
2373 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
2375 eventfd_ctx_put(ev_fd->cq_ev_fd);
2379 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2381 struct io_ev_fd *ev_fd;
2383 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2384 lockdep_is_held(&ctx->uring_lock));
2386 ctx->has_evfd = false;
2387 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2388 call_rcu(&ev_fd->rcu, io_eventfd_put);
2395 static void io_req_caches_free(struct io_ring_ctx *ctx)
2397 struct io_submit_state *state = &ctx->submit_state;
2400 mutex_lock(&ctx->uring_lock);
2401 io_flush_cached_locked_reqs(ctx, state);
2403 while (!io_req_cache_empty(ctx)) {
2404 struct io_wq_work_node *node;
2405 struct io_kiocb *req;
2407 node = wq_stack_extract(&state->free_list);
2408 req = container_of(node, struct io_kiocb, comp_list);
2409 kmem_cache_free(req_cachep, req);
2413 percpu_ref_put_many(&ctx->refs, nr);
2414 mutex_unlock(&ctx->uring_lock);
2417 static void io_flush_apoll_cache(struct io_ring_ctx *ctx)
2419 struct async_poll *apoll;
2421 while (!list_empty(&ctx->apoll_cache)) {
2422 apoll = list_first_entry(&ctx->apoll_cache, struct async_poll,
2424 list_del(&apoll->poll.wait.entry);
2429 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2431 io_sq_thread_finish(ctx);
2433 if (ctx->mm_account) {
2434 mmdrop(ctx->mm_account);
2435 ctx->mm_account = NULL;
2438 io_rsrc_refs_drop(ctx);
2439 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2440 io_wait_rsrc_data(ctx->buf_data);
2441 io_wait_rsrc_data(ctx->file_data);
2443 mutex_lock(&ctx->uring_lock);
2445 __io_sqe_buffers_unregister(ctx);
2447 __io_sqe_files_unregister(ctx);
2449 __io_cqring_overflow_flush(ctx, true);
2450 io_eventfd_unregister(ctx);
2451 io_flush_apoll_cache(ctx);
2452 mutex_unlock(&ctx->uring_lock);
2453 io_destroy_buffers(ctx);
2455 put_cred(ctx->sq_creds);
2457 /* there are no registered resources left, nobody uses it */
2459 io_rsrc_node_destroy(ctx->rsrc_node);
2460 if (ctx->rsrc_backup_node)
2461 io_rsrc_node_destroy(ctx->rsrc_backup_node);
2462 flush_delayed_work(&ctx->rsrc_put_work);
2463 flush_delayed_work(&ctx->fallback_work);
2465 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2466 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
2468 #if defined(CONFIG_UNIX)
2469 if (ctx->ring_sock) {
2470 ctx->ring_sock->file = NULL; /* so that iput() is called */
2471 sock_release(ctx->ring_sock);
2474 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2476 io_mem_free(ctx->rings);
2477 io_mem_free(ctx->sq_sqes);
2479 percpu_ref_exit(&ctx->refs);
2480 free_uid(ctx->user);
2481 io_req_caches_free(ctx);
2483 io_wq_put_hash(ctx->hash_map);
2484 kfree(ctx->cancel_hash);
2485 kfree(ctx->dummy_ubuf);
2487 xa_destroy(&ctx->io_bl_xa);
2491 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2493 struct io_ring_ctx *ctx = file->private_data;
2496 poll_wait(file, &ctx->cq_wait, wait);
2498 * synchronizes with barrier from wq_has_sleeper call in
2502 if (!io_sqring_full(ctx))
2503 mask |= EPOLLOUT | EPOLLWRNORM;
2506 * Don't flush cqring overflow list here, just do a simple check.
2507 * Otherwise there could possible be ABBA deadlock:
2510 * lock(&ctx->uring_lock);
2512 * lock(&ctx->uring_lock);
2515 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2516 * pushs them to do the flush.
2518 if (io_cqring_events(ctx) ||
2519 test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
2520 mask |= EPOLLIN | EPOLLRDNORM;
2525 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2527 const struct cred *creds;
2529 creds = xa_erase(&ctx->personalities, id);
2538 struct io_tctx_exit {
2539 struct callback_head task_work;
2540 struct completion completion;
2541 struct io_ring_ctx *ctx;
2544 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2546 struct io_uring_task *tctx = current->io_uring;
2547 struct io_tctx_exit *work;
2549 work = container_of(cb, struct io_tctx_exit, task_work);
2551 * When @in_idle, we're in cancellation and it's racy to remove the
2552 * node. It'll be removed by the end of cancellation, just ignore it.
2554 if (!atomic_read(&tctx->in_idle))
2555 io_uring_del_tctx_node((unsigned long)work->ctx);
2556 complete(&work->completion);
2559 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2561 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2563 return req->ctx == data;
2566 static __cold void io_ring_exit_work(struct work_struct *work)
2568 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2569 unsigned long timeout = jiffies + HZ * 60 * 5;
2570 unsigned long interval = HZ / 20;
2571 struct io_tctx_exit exit;
2572 struct io_tctx_node *node;
2576 * If we're doing polled IO and end up having requests being
2577 * submitted async (out-of-line), then completions can come in while
2578 * we're waiting for refs to drop. We need to reap these manually,
2579 * as nobody else will be looking for them.
2582 io_uring_try_cancel_requests(ctx, NULL, true);
2584 struct io_sq_data *sqd = ctx->sq_data;
2585 struct task_struct *tsk;
2587 io_sq_thread_park(sqd);
2589 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
2590 io_wq_cancel_cb(tsk->io_uring->io_wq,
2591 io_cancel_ctx_cb, ctx, true);
2592 io_sq_thread_unpark(sqd);
2595 io_req_caches_free(ctx);
2597 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
2598 /* there is little hope left, don't run it too often */
2601 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
2603 init_completion(&exit.completion);
2604 init_task_work(&exit.task_work, io_tctx_exit_cb);
2607 * Some may use context even when all refs and requests have been put,
2608 * and they are free to do so while still holding uring_lock or
2609 * completion_lock, see io_req_task_submit(). Apart from other work,
2610 * this lock/unlock section also waits them to finish.
2612 mutex_lock(&ctx->uring_lock);
2613 while (!list_empty(&ctx->tctx_list)) {
2614 WARN_ON_ONCE(time_after(jiffies, timeout));
2616 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
2618 /* don't spin on a single task if cancellation failed */
2619 list_rotate_left(&ctx->tctx_list);
2620 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
2621 if (WARN_ON_ONCE(ret))
2624 mutex_unlock(&ctx->uring_lock);
2625 wait_for_completion(&exit.completion);
2626 mutex_lock(&ctx->uring_lock);
2628 mutex_unlock(&ctx->uring_lock);
2629 spin_lock(&ctx->completion_lock);
2630 spin_unlock(&ctx->completion_lock);
2632 io_ring_ctx_free(ctx);
2635 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
2637 unsigned long index;
2638 struct creds *creds;
2640 mutex_lock(&ctx->uring_lock);
2641 percpu_ref_kill(&ctx->refs);
2643 __io_cqring_overflow_flush(ctx, true);
2644 xa_for_each(&ctx->personalities, index, creds)
2645 io_unregister_personality(ctx, index);
2646 mutex_unlock(&ctx->uring_lock);
2648 /* failed during ring init, it couldn't have issued any requests */
2650 io_kill_timeouts(ctx, NULL, true);
2651 io_poll_remove_all(ctx, NULL, true);
2652 /* if we failed setting up the ctx, we might not have any rings */
2653 io_iopoll_try_reap_events(ctx);
2656 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
2658 * Use system_unbound_wq to avoid spawning tons of event kworkers
2659 * if we're exiting a ton of rings at the same time. It just adds
2660 * noise and overhead, there's no discernable change in runtime
2661 * over using system_wq.
2663 queue_work(system_unbound_wq, &ctx->exit_work);
2666 static int io_uring_release(struct inode *inode, struct file *file)
2668 struct io_ring_ctx *ctx = file->private_data;
2670 file->private_data = NULL;
2671 io_ring_ctx_wait_and_kill(ctx);
2675 struct io_task_cancel {
2676 struct task_struct *task;
2680 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
2682 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2683 struct io_task_cancel *cancel = data;
2685 return io_match_task_safe(req, cancel->task, cancel->all);
2688 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
2689 struct task_struct *task,
2692 struct io_defer_entry *de;
2695 spin_lock(&ctx->completion_lock);
2696 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
2697 if (io_match_task_safe(de->req, task, cancel_all)) {
2698 list_cut_position(&list, &ctx->defer_list, &de->list);
2702 spin_unlock(&ctx->completion_lock);
2703 if (list_empty(&list))
2706 while (!list_empty(&list)) {
2707 de = list_first_entry(&list, struct io_defer_entry, list);
2708 list_del_init(&de->list);
2709 io_req_complete_failed(de->req, -ECANCELED);
2715 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
2717 struct io_tctx_node *node;
2718 enum io_wq_cancel cret;
2721 mutex_lock(&ctx->uring_lock);
2722 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
2723 struct io_uring_task *tctx = node->task->io_uring;
2726 * io_wq will stay alive while we hold uring_lock, because it's
2727 * killed after ctx nodes, which requires to take the lock.
2729 if (!tctx || !tctx->io_wq)
2731 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
2732 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2734 mutex_unlock(&ctx->uring_lock);
2739 static __cold void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
2740 struct task_struct *task,
2743 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
2744 struct io_uring_task *tctx = task ? task->io_uring : NULL;
2746 /* failed during ring init, it couldn't have issued any requests */
2751 enum io_wq_cancel cret;
2755 ret |= io_uring_try_cancel_iowq(ctx);
2756 } else if (tctx && tctx->io_wq) {
2758 * Cancels requests of all rings, not only @ctx, but
2759 * it's fine as the task is in exit/exec.
2761 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
2763 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2766 /* SQPOLL thread does its own polling */
2767 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
2768 (ctx->sq_data && ctx->sq_data->thread == current)) {
2769 while (!wq_list_empty(&ctx->iopoll_list)) {
2770 io_iopoll_try_reap_events(ctx);
2775 ret |= io_cancel_defer_files(ctx, task, cancel_all);
2776 ret |= io_poll_remove_all(ctx, task, cancel_all);
2777 ret |= io_kill_timeouts(ctx, task, cancel_all);
2779 ret |= io_run_task_work();
2786 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
2789 return atomic_read(&tctx->inflight_tracked);
2790 return percpu_counter_sum(&tctx->inflight);
2794 * Find any io_uring ctx that this task has registered or done IO on, and cancel
2795 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
2797 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
2799 struct io_uring_task *tctx = current->io_uring;
2800 struct io_ring_ctx *ctx;
2804 WARN_ON_ONCE(sqd && sqd->thread != current);
2806 if (!current->io_uring)
2809 io_wq_exit_start(tctx->io_wq);
2811 atomic_inc(&tctx->in_idle);
2813 io_uring_drop_tctx_refs(current);
2814 /* read completions before cancelations */
2815 inflight = tctx_inflight(tctx, !cancel_all);
2820 struct io_tctx_node *node;
2821 unsigned long index;
2823 xa_for_each(&tctx->xa, index, node) {
2824 /* sqpoll task will cancel all its requests */
2825 if (node->ctx->sq_data)
2827 io_uring_try_cancel_requests(node->ctx, current,
2831 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
2832 io_uring_try_cancel_requests(ctx, current,
2836 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
2838 io_uring_drop_tctx_refs(current);
2841 * If we've seen completions, retry without waiting. This
2842 * avoids a race where a completion comes in before we did
2843 * prepare_to_wait().
2845 if (inflight == tctx_inflight(tctx, !cancel_all))
2847 finish_wait(&tctx->wait, &wait);
2850 io_uring_clean_tctx(tctx);
2853 * We shouldn't run task_works after cancel, so just leave
2854 * ->in_idle set for normal exit.
2856 atomic_dec(&tctx->in_idle);
2857 /* for exec all current's requests should be gone, kill tctx */
2858 __io_uring_free(current);
2862 void __io_uring_cancel(bool cancel_all)
2864 io_uring_cancel_generic(cancel_all, NULL);
2867 static void *io_uring_validate_mmap_request(struct file *file,
2868 loff_t pgoff, size_t sz)
2870 struct io_ring_ctx *ctx = file->private_data;
2871 loff_t offset = pgoff << PAGE_SHIFT;
2876 case IORING_OFF_SQ_RING:
2877 case IORING_OFF_CQ_RING:
2880 case IORING_OFF_SQES:
2884 return ERR_PTR(-EINVAL);
2887 page = virt_to_head_page(ptr);
2888 if (sz > page_size(page))
2889 return ERR_PTR(-EINVAL);
2896 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
2898 size_t sz = vma->vm_end - vma->vm_start;
2902 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
2904 return PTR_ERR(ptr);
2906 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
2907 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
2910 #else /* !CONFIG_MMU */
2912 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
2914 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
2917 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
2919 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
2922 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
2923 unsigned long addr, unsigned long len,
2924 unsigned long pgoff, unsigned long flags)
2928 ptr = io_uring_validate_mmap_request(file, pgoff, len);
2930 return PTR_ERR(ptr);
2932 return (unsigned long) ptr;
2935 #endif /* !CONFIG_MMU */
2937 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
2939 if (flags & IORING_ENTER_EXT_ARG) {
2940 struct io_uring_getevents_arg arg;
2942 if (argsz != sizeof(arg))
2944 if (copy_from_user(&arg, argp, sizeof(arg)))
2950 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
2951 struct __kernel_timespec __user **ts,
2952 const sigset_t __user **sig)
2954 struct io_uring_getevents_arg arg;
2957 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
2958 * is just a pointer to the sigset_t.
2960 if (!(flags & IORING_ENTER_EXT_ARG)) {
2961 *sig = (const sigset_t __user *) argp;
2967 * EXT_ARG is set - ensure we agree on the size of it and copy in our
2968 * timespec and sigset_t pointers if good.
2970 if (*argsz != sizeof(arg))
2972 if (copy_from_user(&arg, argp, sizeof(arg)))
2976 *sig = u64_to_user_ptr(arg.sigmask);
2977 *argsz = arg.sigmask_sz;
2978 *ts = u64_to_user_ptr(arg.ts);
2982 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
2983 u32, min_complete, u32, flags, const void __user *, argp,
2986 struct io_ring_ctx *ctx;
2992 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
2993 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
2994 IORING_ENTER_REGISTERED_RING)))
2998 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
2999 * need only dereference our task private array to find it.
3001 if (flags & IORING_ENTER_REGISTERED_RING) {
3002 struct io_uring_task *tctx = current->io_uring;
3004 if (!tctx || fd >= IO_RINGFD_REG_MAX)
3006 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3007 f.file = tctx->registered_rings[fd];
3013 if (unlikely(!f.file))
3017 if (unlikely(!io_is_uring_fops(f.file)))
3021 ctx = f.file->private_data;
3022 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
3026 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3030 * For SQ polling, the thread will do all submissions and completions.
3031 * Just return the requested submit count, and wake the thread if
3035 if (ctx->flags & IORING_SETUP_SQPOLL) {
3036 io_cqring_overflow_flush(ctx);
3038 if (unlikely(ctx->sq_data->thread == NULL)) {
3042 if (flags & IORING_ENTER_SQ_WAKEUP)
3043 wake_up(&ctx->sq_data->wait);
3044 if (flags & IORING_ENTER_SQ_WAIT) {
3045 ret = io_sqpoll_wait_sq(ctx);
3050 } else if (to_submit) {
3051 ret = io_uring_add_tctx_node(ctx);
3055 mutex_lock(&ctx->uring_lock);
3056 ret = io_submit_sqes(ctx, to_submit);
3057 if (ret != to_submit) {
3058 mutex_unlock(&ctx->uring_lock);
3061 if ((flags & IORING_ENTER_GETEVENTS) && ctx->syscall_iopoll)
3063 mutex_unlock(&ctx->uring_lock);
3065 if (flags & IORING_ENTER_GETEVENTS) {
3067 if (ctx->syscall_iopoll) {
3069 * We disallow the app entering submit/complete with
3070 * polling, but we still need to lock the ring to
3071 * prevent racing with polled issue that got punted to
3074 mutex_lock(&ctx->uring_lock);
3076 ret2 = io_validate_ext_arg(flags, argp, argsz);
3077 if (likely(!ret2)) {
3078 min_complete = min(min_complete,
3080 ret2 = io_iopoll_check(ctx, min_complete);
3082 mutex_unlock(&ctx->uring_lock);
3084 const sigset_t __user *sig;
3085 struct __kernel_timespec __user *ts;
3087 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3088 if (likely(!ret2)) {
3089 min_complete = min(min_complete,
3091 ret2 = io_cqring_wait(ctx, min_complete, sig,
3100 * EBADR indicates that one or more CQE were dropped.
3101 * Once the user has been informed we can clear the bit
3102 * as they are obviously ok with those drops.
3104 if (unlikely(ret2 == -EBADR))
3105 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3111 percpu_ref_put(&ctx->refs);
3117 static const struct file_operations io_uring_fops = {
3118 .release = io_uring_release,
3119 .mmap = io_uring_mmap,
3121 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3122 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3124 .poll = io_uring_poll,
3125 #ifdef CONFIG_PROC_FS
3126 .show_fdinfo = io_uring_show_fdinfo,
3130 bool io_is_uring_fops(struct file *file)
3132 return file->f_op == &io_uring_fops;
3135 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3136 struct io_uring_params *p)
3138 struct io_rings *rings;
3139 size_t size, sq_array_offset;
3141 /* make sure these are sane, as we already accounted them */
3142 ctx->sq_entries = p->sq_entries;
3143 ctx->cq_entries = p->cq_entries;
3145 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3146 if (size == SIZE_MAX)
3149 rings = io_mem_alloc(size);
3154 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3155 rings->sq_ring_mask = p->sq_entries - 1;
3156 rings->cq_ring_mask = p->cq_entries - 1;
3157 rings->sq_ring_entries = p->sq_entries;
3158 rings->cq_ring_entries = p->cq_entries;
3160 if (p->flags & IORING_SETUP_SQE128)
3161 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3163 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3164 if (size == SIZE_MAX) {
3165 io_mem_free(ctx->rings);
3170 ctx->sq_sqes = io_mem_alloc(size);
3171 if (!ctx->sq_sqes) {
3172 io_mem_free(ctx->rings);
3180 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3184 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3188 ret = io_uring_add_tctx_node(ctx);
3193 fd_install(fd, file);
3198 * Allocate an anonymous fd, this is what constitutes the application
3199 * visible backing of an io_uring instance. The application mmaps this
3200 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3201 * we have to tie this fd to a socket for file garbage collection purposes.
3203 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3206 #if defined(CONFIG_UNIX)
3209 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3212 return ERR_PTR(ret);
3215 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3216 O_RDWR | O_CLOEXEC, NULL);
3217 #if defined(CONFIG_UNIX)
3219 sock_release(ctx->ring_sock);
3220 ctx->ring_sock = NULL;
3222 ctx->ring_sock->file = file;
3228 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3229 struct io_uring_params __user *params)
3231 struct io_ring_ctx *ctx;
3237 if (entries > IORING_MAX_ENTRIES) {
3238 if (!(p->flags & IORING_SETUP_CLAMP))
3240 entries = IORING_MAX_ENTRIES;
3244 * Use twice as many entries for the CQ ring. It's possible for the
3245 * application to drive a higher depth than the size of the SQ ring,
3246 * since the sqes are only used at submission time. This allows for
3247 * some flexibility in overcommitting a bit. If the application has
3248 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3249 * of CQ ring entries manually.
3251 p->sq_entries = roundup_pow_of_two(entries);
3252 if (p->flags & IORING_SETUP_CQSIZE) {
3254 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3255 * to a power-of-two, if it isn't already. We do NOT impose
3256 * any cq vs sq ring sizing.
3260 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3261 if (!(p->flags & IORING_SETUP_CLAMP))
3263 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3265 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3266 if (p->cq_entries < p->sq_entries)
3269 p->cq_entries = 2 * p->sq_entries;
3272 ctx = io_ring_ctx_alloc(p);
3277 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3278 * space applications don't need to do io completion events
3279 * polling again, they can rely on io_sq_thread to do polling
3280 * work, which can reduce cpu usage and uring_lock contention.
3282 if (ctx->flags & IORING_SETUP_IOPOLL &&
3283 !(ctx->flags & IORING_SETUP_SQPOLL))
3284 ctx->syscall_iopoll = 1;
3286 ctx->compat = in_compat_syscall();
3287 if (!capable(CAP_IPC_LOCK))
3288 ctx->user = get_uid(current_user());
3291 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3292 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3295 if (ctx->flags & IORING_SETUP_SQPOLL) {
3296 /* IPI related flags don't make sense with SQPOLL */
3297 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3298 IORING_SETUP_TASKRUN_FLAG))
3300 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3301 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3302 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3304 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
3306 ctx->notify_method = TWA_SIGNAL;
3310 * This is just grabbed for accounting purposes. When a process exits,
3311 * the mm is exited and dropped before the files, hence we need to hang
3312 * on to this mm purely for the purposes of being able to unaccount
3313 * memory (locked/pinned vm). It's not used for anything else.
3315 mmgrab(current->mm);
3316 ctx->mm_account = current->mm;
3318 ret = io_allocate_scq_urings(ctx, p);
3322 ret = io_sq_offload_create(ctx, p);
3325 /* always set a rsrc node */
3326 ret = io_rsrc_node_switch_start(ctx);
3329 io_rsrc_node_switch(ctx, NULL);
3331 memset(&p->sq_off, 0, sizeof(p->sq_off));
3332 p->sq_off.head = offsetof(struct io_rings, sq.head);
3333 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3334 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3335 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3336 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3337 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3338 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3340 memset(&p->cq_off, 0, sizeof(p->cq_off));
3341 p->cq_off.head = offsetof(struct io_rings, cq.head);
3342 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3343 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3344 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3345 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3346 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3347 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3349 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3350 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3351 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3352 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3353 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3354 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3355 IORING_FEAT_LINKED_FILE;
3357 if (copy_to_user(params, p, sizeof(*p))) {
3362 file = io_uring_get_file(ctx);
3364 ret = PTR_ERR(file);
3369 * Install ring fd as the very last thing, so we don't risk someone
3370 * having closed it before we finish setup
3372 ret = io_uring_install_fd(ctx, file);
3374 /* fput will clean it up */
3379 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3382 io_ring_ctx_wait_and_kill(ctx);
3387 * Sets up an aio uring context, and returns the fd. Applications asks for a
3388 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3389 * params structure passed in.
3391 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3393 struct io_uring_params p;
3396 if (copy_from_user(&p, params, sizeof(p)))
3398 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3403 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3404 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3405 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3406 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3407 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3408 IORING_SETUP_SQE128 | IORING_SETUP_CQE32))
3411 return io_uring_create(entries, &p, params);
3414 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3415 struct io_uring_params __user *, params)
3417 return io_uring_setup(entries, params);
3420 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3423 struct io_uring_probe *p;
3427 size = struct_size(p, ops, nr_args);
3428 if (size == SIZE_MAX)
3430 p = kzalloc(size, GFP_KERNEL);
3435 if (copy_from_user(p, arg, size))
3438 if (memchr_inv(p, 0, size))
3441 p->last_op = IORING_OP_LAST - 1;
3442 if (nr_args > IORING_OP_LAST)
3443 nr_args = IORING_OP_LAST;
3445 for (i = 0; i < nr_args; i++) {
3447 if (!io_op_defs[i].not_supported)
3448 p->ops[i].flags = IO_URING_OP_SUPPORTED;
3453 if (copy_to_user(arg, p, size))
3460 static int io_register_personality(struct io_ring_ctx *ctx)
3462 const struct cred *creds;
3466 creds = get_current_cred();
3468 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
3469 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
3477 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
3478 void __user *arg, unsigned int nr_args)
3480 struct io_uring_restriction *res;
3484 /* Restrictions allowed only if rings started disabled */
3485 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3488 /* We allow only a single restrictions registration */
3489 if (ctx->restrictions.registered)
3492 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
3495 size = array_size(nr_args, sizeof(*res));
3496 if (size == SIZE_MAX)
3499 res = memdup_user(arg, size);
3501 return PTR_ERR(res);
3505 for (i = 0; i < nr_args; i++) {
3506 switch (res[i].opcode) {
3507 case IORING_RESTRICTION_REGISTER_OP:
3508 if (res[i].register_op >= IORING_REGISTER_LAST) {
3513 __set_bit(res[i].register_op,
3514 ctx->restrictions.register_op);
3516 case IORING_RESTRICTION_SQE_OP:
3517 if (res[i].sqe_op >= IORING_OP_LAST) {
3522 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
3524 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
3525 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
3527 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
3528 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
3537 /* Reset all restrictions if an error happened */
3539 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
3541 ctx->restrictions.registered = true;
3547 static int io_register_enable_rings(struct io_ring_ctx *ctx)
3549 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3552 if (ctx->restrictions.registered)
3553 ctx->restricted = 1;
3555 ctx->flags &= ~IORING_SETUP_R_DISABLED;
3556 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
3557 wake_up(&ctx->sq_data->wait);
3561 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
3562 void __user *arg, unsigned len)
3564 struct io_uring_task *tctx = current->io_uring;
3565 cpumask_var_t new_mask;
3568 if (!tctx || !tctx->io_wq)
3571 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3574 cpumask_clear(new_mask);
3575 if (len > cpumask_size())
3576 len = cpumask_size();
3578 if (in_compat_syscall()) {
3579 ret = compat_get_bitmap(cpumask_bits(new_mask),
3580 (const compat_ulong_t __user *)arg,
3581 len * 8 /* CHAR_BIT */);
3583 ret = copy_from_user(new_mask, arg, len);
3587 free_cpumask_var(new_mask);
3591 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
3592 free_cpumask_var(new_mask);
3596 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
3598 struct io_uring_task *tctx = current->io_uring;
3600 if (!tctx || !tctx->io_wq)
3603 return io_wq_cpu_affinity(tctx->io_wq, NULL);
3606 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
3608 __must_hold(&ctx->uring_lock)
3610 struct io_tctx_node *node;
3611 struct io_uring_task *tctx = NULL;
3612 struct io_sq_data *sqd = NULL;
3616 if (copy_from_user(new_count, arg, sizeof(new_count)))
3618 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3619 if (new_count[i] > INT_MAX)
3622 if (ctx->flags & IORING_SETUP_SQPOLL) {
3626 * Observe the correct sqd->lock -> ctx->uring_lock
3627 * ordering. Fine to drop uring_lock here, we hold
3630 refcount_inc(&sqd->refs);
3631 mutex_unlock(&ctx->uring_lock);
3632 mutex_lock(&sqd->lock);
3633 mutex_lock(&ctx->uring_lock);
3635 tctx = sqd->thread->io_uring;
3638 tctx = current->io_uring;
3641 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
3643 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3645 ctx->iowq_limits[i] = new_count[i];
3646 ctx->iowq_limits_set = true;
3648 if (tctx && tctx->io_wq) {
3649 ret = io_wq_max_workers(tctx->io_wq, new_count);
3653 memset(new_count, 0, sizeof(new_count));
3657 mutex_unlock(&sqd->lock);
3658 io_put_sq_data(sqd);
3661 if (copy_to_user(arg, new_count, sizeof(new_count)))
3664 /* that's it for SQPOLL, only the SQPOLL task creates requests */
3668 /* now propagate the restriction to all registered users */
3669 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3670 struct io_uring_task *tctx = node->task->io_uring;
3672 if (WARN_ON_ONCE(!tctx->io_wq))
3675 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3676 new_count[i] = ctx->iowq_limits[i];
3677 /* ignore errors, it always returns zero anyway */
3678 (void)io_wq_max_workers(tctx->io_wq, new_count);
3683 mutex_unlock(&sqd->lock);
3684 io_put_sq_data(sqd);
3689 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
3690 void __user *arg, unsigned nr_args)
3691 __releases(ctx->uring_lock)
3692 __acquires(ctx->uring_lock)
3697 * We're inside the ring mutex, if the ref is already dying, then
3698 * someone else killed the ctx or is already going through
3699 * io_uring_register().
3701 if (percpu_ref_is_dying(&ctx->refs))
3704 if (ctx->restricted) {
3705 if (opcode >= IORING_REGISTER_LAST)
3707 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
3708 if (!test_bit(opcode, ctx->restrictions.register_op))
3713 case IORING_REGISTER_BUFFERS:
3717 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
3719 case IORING_UNREGISTER_BUFFERS:
3723 ret = io_sqe_buffers_unregister(ctx);
3725 case IORING_REGISTER_FILES:
3729 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
3731 case IORING_UNREGISTER_FILES:
3735 ret = io_sqe_files_unregister(ctx);
3737 case IORING_REGISTER_FILES_UPDATE:
3738 ret = io_register_files_update(ctx, arg, nr_args);
3740 case IORING_REGISTER_EVENTFD:
3744 ret = io_eventfd_register(ctx, arg, 0);
3746 case IORING_REGISTER_EVENTFD_ASYNC:
3750 ret = io_eventfd_register(ctx, arg, 1);
3752 case IORING_UNREGISTER_EVENTFD:
3756 ret = io_eventfd_unregister(ctx);
3758 case IORING_REGISTER_PROBE:
3760 if (!arg || nr_args > 256)
3762 ret = io_probe(ctx, arg, nr_args);
3764 case IORING_REGISTER_PERSONALITY:
3768 ret = io_register_personality(ctx);
3770 case IORING_UNREGISTER_PERSONALITY:
3774 ret = io_unregister_personality(ctx, nr_args);
3776 case IORING_REGISTER_ENABLE_RINGS:
3780 ret = io_register_enable_rings(ctx);
3782 case IORING_REGISTER_RESTRICTIONS:
3783 ret = io_register_restrictions(ctx, arg, nr_args);
3785 case IORING_REGISTER_FILES2:
3786 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
3788 case IORING_REGISTER_FILES_UPDATE2:
3789 ret = io_register_rsrc_update(ctx, arg, nr_args,
3792 case IORING_REGISTER_BUFFERS2:
3793 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
3795 case IORING_REGISTER_BUFFERS_UPDATE:
3796 ret = io_register_rsrc_update(ctx, arg, nr_args,
3797 IORING_RSRC_BUFFER);
3799 case IORING_REGISTER_IOWQ_AFF:
3801 if (!arg || !nr_args)
3803 ret = io_register_iowq_aff(ctx, arg, nr_args);
3805 case IORING_UNREGISTER_IOWQ_AFF:
3809 ret = io_unregister_iowq_aff(ctx);
3811 case IORING_REGISTER_IOWQ_MAX_WORKERS:
3813 if (!arg || nr_args != 2)
3815 ret = io_register_iowq_max_workers(ctx, arg);
3817 case IORING_REGISTER_RING_FDS:
3818 ret = io_ringfd_register(ctx, arg, nr_args);
3820 case IORING_UNREGISTER_RING_FDS:
3821 ret = io_ringfd_unregister(ctx, arg, nr_args);
3823 case IORING_REGISTER_PBUF_RING:
3825 if (!arg || nr_args != 1)
3827 ret = io_register_pbuf_ring(ctx, arg);
3829 case IORING_UNREGISTER_PBUF_RING:
3831 if (!arg || nr_args != 1)
3833 ret = io_unregister_pbuf_ring(ctx, arg);
3843 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
3844 void __user *, arg, unsigned int, nr_args)
3846 struct io_ring_ctx *ctx;
3855 if (!io_is_uring_fops(f.file))
3858 ctx = f.file->private_data;
3862 mutex_lock(&ctx->uring_lock);
3863 ret = __io_uring_register(ctx, opcode, arg, nr_args);
3864 mutex_unlock(&ctx->uring_lock);
3865 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
3871 static int __init io_uring_init(void)
3873 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
3874 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
3875 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
3878 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
3879 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
3880 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
3881 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
3882 BUILD_BUG_SQE_ELEM(1, __u8, flags);
3883 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
3884 BUILD_BUG_SQE_ELEM(4, __s32, fd);
3885 BUILD_BUG_SQE_ELEM(8, __u64, off);
3886 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
3887 BUILD_BUG_SQE_ELEM(16, __u64, addr);
3888 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
3889 BUILD_BUG_SQE_ELEM(24, __u32, len);
3890 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
3891 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
3892 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
3893 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
3894 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
3895 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
3896 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
3897 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
3898 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
3899 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
3900 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
3901 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
3902 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
3903 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
3904 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
3905 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
3906 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
3907 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
3908 BUILD_BUG_SQE_ELEM(42, __u16, personality);
3909 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
3910 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
3911 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
3913 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
3914 sizeof(struct io_uring_rsrc_update));
3915 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
3916 sizeof(struct io_uring_rsrc_update2));
3918 /* ->buf_index is u16 */
3919 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
3920 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
3921 offsetof(struct io_uring_buf_ring, tail));
3923 /* should fit into one byte */
3924 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
3925 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
3926 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
3928 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
3930 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
3932 io_uring_optable_init();
3934 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
3938 __initcall(io_uring_init);