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>
96 #define IORING_MAX_ENTRIES 32768
97 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
99 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
100 IORING_REGISTER_LAST + IORING_OP_LAST)
102 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
103 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
105 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
106 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
108 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
109 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
112 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
115 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
117 #define IO_COMPL_BATCH 32
118 #define IO_REQ_ALLOC_BATCH 8
121 IO_CHECK_CQ_OVERFLOW_BIT,
122 IO_CHECK_CQ_DROPPED_BIT,
125 struct io_defer_entry {
126 struct list_head list;
127 struct io_kiocb *req;
131 /* requests with any of those set should undergo io_disarm_next() */
132 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
133 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
135 static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
136 struct task_struct *task,
139 static void io_dismantle_req(struct io_kiocb *req);
140 static void io_clean_op(struct io_kiocb *req);
141 static void io_queue_sqe(struct io_kiocb *req);
143 static void __io_submit_flush_completions(struct io_ring_ctx *ctx);
145 static struct kmem_cache *req_cachep;
147 struct sock *io_uring_get_socket(struct file *file)
149 #if defined(CONFIG_UNIX)
150 if (io_is_uring_fops(file)) {
151 struct io_ring_ctx *ctx = file->private_data;
153 return ctx->ring_sock->sk;
158 EXPORT_SYMBOL(io_uring_get_socket);
160 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
162 if (!wq_list_empty(&ctx->submit_state.compl_reqs))
163 __io_submit_flush_completions(ctx);
166 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
168 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
171 static bool io_match_linked(struct io_kiocb *head)
173 struct io_kiocb *req;
175 io_for_each_link(req, head) {
176 if (req->flags & REQ_F_INFLIGHT)
183 * As io_match_task() but protected against racing with linked timeouts.
184 * User must not hold timeout_lock.
186 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
191 if (task && head->task != task)
196 if (head->flags & REQ_F_LINK_TIMEOUT) {
197 struct io_ring_ctx *ctx = head->ctx;
199 /* protect against races with linked timeouts */
200 spin_lock_irq(&ctx->timeout_lock);
201 matched = io_match_linked(head);
202 spin_unlock_irq(&ctx->timeout_lock);
204 matched = io_match_linked(head);
209 static inline void req_fail_link_node(struct io_kiocb *req, int res)
212 io_req_set_res(req, res, 0);
215 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
217 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
220 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
222 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
224 complete(&ctx->ref_comp);
227 static __cold void io_fallback_req_func(struct work_struct *work)
229 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
231 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
232 struct io_kiocb *req, *tmp;
235 percpu_ref_get(&ctx->refs);
236 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
237 req->io_task_work.func(req, &locked);
240 io_submit_flush_completions(ctx);
241 mutex_unlock(&ctx->uring_lock);
243 percpu_ref_put(&ctx->refs);
246 static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
248 unsigned hash_buckets = 1U << bits;
249 size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
251 table->hbs = kmalloc(hash_size, GFP_KERNEL);
255 table->hash_bits = bits;
256 init_hash_table(table, hash_buckets);
260 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
262 struct io_ring_ctx *ctx;
265 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
269 xa_init(&ctx->io_bl_xa);
272 * Use 5 bits less than the max cq entries, that should give us around
273 * 32 entries per hash list if totally full and uniformly spread, but
274 * don't keep too many buckets to not overconsume memory.
276 hash_bits = ilog2(p->cq_entries) - 5;
277 hash_bits = clamp(hash_bits, 1, 8);
278 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
280 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
283 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
284 if (!ctx->dummy_ubuf)
286 /* set invalid range, so io_import_fixed() fails meeting it */
287 ctx->dummy_ubuf->ubuf = -1UL;
289 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
290 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
293 ctx->flags = p->flags;
294 init_waitqueue_head(&ctx->sqo_sq_wait);
295 INIT_LIST_HEAD(&ctx->sqd_list);
296 INIT_LIST_HEAD(&ctx->cq_overflow_list);
297 INIT_LIST_HEAD(&ctx->io_buffers_cache);
298 INIT_LIST_HEAD(&ctx->apoll_cache);
299 init_completion(&ctx->ref_comp);
300 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
301 mutex_init(&ctx->uring_lock);
302 init_waitqueue_head(&ctx->cq_wait);
303 spin_lock_init(&ctx->completion_lock);
304 spin_lock_init(&ctx->timeout_lock);
305 INIT_WQ_LIST(&ctx->iopoll_list);
306 INIT_LIST_HEAD(&ctx->io_buffers_pages);
307 INIT_LIST_HEAD(&ctx->io_buffers_comp);
308 INIT_LIST_HEAD(&ctx->defer_list);
309 INIT_LIST_HEAD(&ctx->timeout_list);
310 INIT_LIST_HEAD(&ctx->ltimeout_list);
311 spin_lock_init(&ctx->rsrc_ref_lock);
312 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
313 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
314 init_llist_head(&ctx->rsrc_put_llist);
315 INIT_LIST_HEAD(&ctx->tctx_list);
316 ctx->submit_state.free_list.next = NULL;
317 INIT_WQ_LIST(&ctx->locked_free_list);
318 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
319 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
322 kfree(ctx->dummy_ubuf);
323 kfree(ctx->cancel_table.hbs);
324 kfree(ctx->cancel_table_locked.hbs);
326 xa_destroy(&ctx->io_bl_xa);
331 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
333 struct io_rings *r = ctx->rings;
335 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
339 static bool req_need_defer(struct io_kiocb *req, u32 seq)
341 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
342 struct io_ring_ctx *ctx = req->ctx;
344 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
350 static inline void io_req_track_inflight(struct io_kiocb *req)
352 if (!(req->flags & REQ_F_INFLIGHT)) {
353 req->flags |= REQ_F_INFLIGHT;
354 atomic_inc(&req->task->io_uring->inflight_tracked);
358 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
360 if (WARN_ON_ONCE(!req->link))
363 req->flags &= ~REQ_F_ARM_LTIMEOUT;
364 req->flags |= REQ_F_LINK_TIMEOUT;
366 /* linked timeouts should have two refs once prep'ed */
367 io_req_set_refcount(req);
368 __io_req_set_refcount(req->link, 2);
372 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
374 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
376 return __io_prep_linked_timeout(req);
379 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
381 io_queue_linked_timeout(__io_prep_linked_timeout(req));
384 static inline void io_arm_ltimeout(struct io_kiocb *req)
386 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
387 __io_arm_ltimeout(req);
390 static void io_prep_async_work(struct io_kiocb *req)
392 const struct io_op_def *def = &io_op_defs[req->opcode];
393 struct io_ring_ctx *ctx = req->ctx;
395 if (!(req->flags & REQ_F_CREDS)) {
396 req->flags |= REQ_F_CREDS;
397 req->creds = get_current_cred();
400 req->work.list.next = NULL;
402 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
403 if (req->flags & REQ_F_FORCE_ASYNC)
404 req->work.flags |= IO_WQ_WORK_CONCURRENT;
406 if (req->flags & REQ_F_ISREG) {
407 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
408 io_wq_hash_work(&req->work, file_inode(req->file));
409 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
410 if (def->unbound_nonreg_file)
411 req->work.flags |= IO_WQ_WORK_UNBOUND;
415 static void io_prep_async_link(struct io_kiocb *req)
417 struct io_kiocb *cur;
419 if (req->flags & REQ_F_LINK_TIMEOUT) {
420 struct io_ring_ctx *ctx = req->ctx;
422 spin_lock_irq(&ctx->timeout_lock);
423 io_for_each_link(cur, req)
424 io_prep_async_work(cur);
425 spin_unlock_irq(&ctx->timeout_lock);
427 io_for_each_link(cur, req)
428 io_prep_async_work(cur);
432 void io_queue_iowq(struct io_kiocb *req, bool *dont_use)
434 struct io_kiocb *link = io_prep_linked_timeout(req);
435 struct io_uring_task *tctx = req->task->io_uring;
438 BUG_ON(!tctx->io_wq);
440 /* init ->work of the whole link before punting */
441 io_prep_async_link(req);
444 * Not expected to happen, but if we do have a bug where this _can_
445 * happen, catch it here and ensure the request is marked as
446 * canceled. That will make io-wq go through the usual work cancel
447 * procedure rather than attempt to run this request (or create a new
450 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
451 req->work.flags |= IO_WQ_WORK_CANCEL;
453 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
454 io_wq_enqueue(tctx->io_wq, &req->work);
456 io_queue_linked_timeout(link);
459 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
461 while (!list_empty(&ctx->defer_list)) {
462 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
463 struct io_defer_entry, list);
465 if (req_need_defer(de->req, de->seq))
467 list_del_init(&de->list);
468 io_req_task_queue(de->req);
473 static void io_eventfd_signal(struct io_ring_ctx *ctx)
475 struct io_ev_fd *ev_fd;
478 spin_lock(&ctx->completion_lock);
480 * Eventfd should only get triggered when at least one event has been
481 * posted. Some applications rely on the eventfd notification count only
482 * changing IFF a new CQE has been added to the CQ ring. There's no
483 * depedency on 1:1 relationship between how many times this function is
484 * called (and hence the eventfd count) and number of CQEs posted to the
487 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
488 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
489 spin_unlock(&ctx->completion_lock);
495 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
498 ev_fd = rcu_dereference(ctx->io_ev_fd);
501 * Check again if ev_fd exists incase an io_eventfd_unregister call
502 * completed between the NULL check of ctx->io_ev_fd at the start of
503 * the function and rcu_read_lock.
505 if (unlikely(!ev_fd))
507 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
510 if (!ev_fd->eventfd_async || io_wq_current_is_worker())
511 eventfd_signal(ev_fd->cq_ev_fd, 1);
516 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
518 if (ctx->off_timeout_used || ctx->drain_active) {
519 spin_lock(&ctx->completion_lock);
520 if (ctx->off_timeout_used)
521 io_flush_timeouts(ctx);
522 if (ctx->drain_active)
523 io_queue_deferred(ctx);
524 spin_unlock(&ctx->completion_lock);
527 io_eventfd_signal(ctx);
530 static inline void io_cqring_ev_posted(struct io_ring_ctx *ctx)
532 io_commit_cqring_flush(ctx);
536 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
537 __releases(ctx->completion_lock)
539 io_commit_cqring(ctx);
540 spin_unlock(&ctx->completion_lock);
541 io_cqring_ev_posted(ctx);
544 void io_cq_unlock_post(struct io_ring_ctx *ctx)
546 __io_cq_unlock_post(ctx);
549 /* Returns true if there are no backlogged entries after the flush */
550 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
553 size_t cqe_size = sizeof(struct io_uring_cqe);
555 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
558 if (ctx->flags & IORING_SETUP_CQE32)
562 while (!list_empty(&ctx->cq_overflow_list)) {
563 struct io_uring_cqe *cqe = io_get_cqe(ctx);
564 struct io_overflow_cqe *ocqe;
568 ocqe = list_first_entry(&ctx->cq_overflow_list,
569 struct io_overflow_cqe, list);
571 memcpy(cqe, &ocqe->cqe, cqe_size);
573 io_account_cq_overflow(ctx);
575 list_del(&ocqe->list);
579 all_flushed = list_empty(&ctx->cq_overflow_list);
581 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
582 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
585 io_cq_unlock_post(ctx);
589 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
593 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
594 /* iopoll syncs against uring_lock, not completion_lock */
595 if (ctx->flags & IORING_SETUP_IOPOLL)
596 mutex_lock(&ctx->uring_lock);
597 ret = __io_cqring_overflow_flush(ctx, false);
598 if (ctx->flags & IORING_SETUP_IOPOLL)
599 mutex_unlock(&ctx->uring_lock);
605 static void __io_put_task(struct task_struct *task, int nr)
607 struct io_uring_task *tctx = task->io_uring;
609 percpu_counter_sub(&tctx->inflight, nr);
610 if (unlikely(atomic_read(&tctx->in_idle)))
611 wake_up(&tctx->wait);
612 put_task_struct_many(task, nr);
615 /* must to be called somewhat shortly after putting a request */
616 static inline void io_put_task(struct task_struct *task, int nr)
618 if (likely(task == current))
619 task->io_uring->cached_refs += nr;
621 __io_put_task(task, nr);
624 static void io_task_refs_refill(struct io_uring_task *tctx)
626 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
628 percpu_counter_add(&tctx->inflight, refill);
629 refcount_add(refill, ¤t->usage);
630 tctx->cached_refs += refill;
633 static inline void io_get_task_refs(int nr)
635 struct io_uring_task *tctx = current->io_uring;
637 tctx->cached_refs -= nr;
638 if (unlikely(tctx->cached_refs < 0))
639 io_task_refs_refill(tctx);
642 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
644 struct io_uring_task *tctx = task->io_uring;
645 unsigned int refs = tctx->cached_refs;
648 tctx->cached_refs = 0;
649 percpu_counter_sub(&tctx->inflight, refs);
650 put_task_struct_many(task, refs);
654 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
655 s32 res, u32 cflags, u64 extra1, u64 extra2)
657 struct io_overflow_cqe *ocqe;
658 size_t ocq_size = sizeof(struct io_overflow_cqe);
659 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
662 ocq_size += sizeof(struct io_uring_cqe);
664 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
665 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
668 * If we're in ring overflow flush mode, or in task cancel mode,
669 * or cannot allocate an overflow entry, then we need to drop it
672 io_account_cq_overflow(ctx);
673 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
676 if (list_empty(&ctx->cq_overflow_list)) {
677 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
678 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
681 ocqe->cqe.user_data = user_data;
683 ocqe->cqe.flags = cflags;
685 ocqe->cqe.big_cqe[0] = extra1;
686 ocqe->cqe.big_cqe[1] = extra2;
688 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
692 bool io_req_cqe_overflow(struct io_kiocb *req)
694 if (!(req->flags & REQ_F_CQE32_INIT)) {
698 return io_cqring_event_overflow(req->ctx, req->cqe.user_data,
699 req->cqe.res, req->cqe.flags,
700 req->extra1, req->extra2);
704 * writes to the cq entry need to come after reading head; the
705 * control dependency is enough as we're using WRITE_ONCE to
708 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx)
710 struct io_rings *rings = ctx->rings;
711 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
712 unsigned int free, queued, len;
715 /* userspace may cheat modifying the tail, be safe and do min */
716 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
717 free = ctx->cq_entries - queued;
718 /* we need a contiguous range, limit based on the current array offset */
719 len = min(free, ctx->cq_entries - off);
723 if (ctx->flags & IORING_SETUP_CQE32) {
728 ctx->cqe_cached = &rings->cqes[off];
729 ctx->cqe_sentinel = ctx->cqe_cached + len;
731 ctx->cached_cq_tail++;
733 if (ctx->flags & IORING_SETUP_CQE32)
735 return &rings->cqes[off];
738 static bool io_fill_cqe_aux(struct io_ring_ctx *ctx,
739 u64 user_data, s32 res, u32 cflags)
741 struct io_uring_cqe *cqe;
744 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
747 * If we can't get a cq entry, userspace overflowed the
748 * submission (by quite a lot). Increment the overflow count in
751 cqe = io_get_cqe(ctx);
753 WRITE_ONCE(cqe->user_data, user_data);
754 WRITE_ONCE(cqe->res, res);
755 WRITE_ONCE(cqe->flags, cflags);
757 if (ctx->flags & IORING_SETUP_CQE32) {
758 WRITE_ONCE(cqe->big_cqe[0], 0);
759 WRITE_ONCE(cqe->big_cqe[1], 0);
763 return io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
766 bool io_post_aux_cqe(struct io_ring_ctx *ctx,
767 u64 user_data, s32 res, u32 cflags)
772 filled = io_fill_cqe_aux(ctx, user_data, res, cflags);
773 io_cq_unlock_post(ctx);
777 static void __io_req_complete_put(struct io_kiocb *req)
780 * If we're the last reference to this request, add to our locked
783 if (req_ref_put_and_test(req)) {
784 struct io_ring_ctx *ctx = req->ctx;
786 if (req->flags & IO_REQ_LINK_FLAGS) {
787 if (req->flags & IO_DISARM_MASK)
790 io_req_task_queue(req->link);
794 io_req_put_rsrc(req);
796 * Selected buffer deallocation in io_clean_op() assumes that
797 * we don't hold ->completion_lock. Clean them here to avoid
800 io_put_kbuf_comp(req);
801 io_dismantle_req(req);
802 io_put_task(req->task, 1);
803 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
804 ctx->locked_free_nr++;
808 void __io_req_complete_post(struct io_kiocb *req)
810 if (!(req->flags & REQ_F_CQE_SKIP))
811 __io_fill_cqe_req(req->ctx, req);
812 __io_req_complete_put(req);
815 void io_req_complete_post(struct io_kiocb *req)
817 struct io_ring_ctx *ctx = req->ctx;
820 __io_req_complete_post(req);
821 io_cq_unlock_post(ctx);
824 inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags)
826 io_req_complete_post(req);
829 void io_req_complete_failed(struct io_kiocb *req, s32 res)
832 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
833 io_req_complete_post(req);
837 * Don't initialise the fields below on every allocation, but do that in
838 * advance and keep them valid across allocations.
840 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
844 req->async_data = NULL;
845 /* not necessary, but safer to zero */
849 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
850 struct io_submit_state *state)
852 spin_lock(&ctx->completion_lock);
853 wq_list_splice(&ctx->locked_free_list, &state->free_list);
854 ctx->locked_free_nr = 0;
855 spin_unlock(&ctx->completion_lock);
858 static inline bool io_req_cache_empty(struct io_ring_ctx *ctx)
860 return !ctx->submit_state.free_list.next;
864 * A request might get retired back into the request caches even before opcode
865 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
866 * Because of that, io_alloc_req() should be called only under ->uring_lock
867 * and with extra caution to not get a request that is still worked on.
869 static __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
870 __must_hold(&ctx->uring_lock)
872 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
873 void *reqs[IO_REQ_ALLOC_BATCH];
877 * If we have more than a batch's worth of requests in our IRQ side
878 * locked cache, grab the lock and move them over to our submission
881 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
882 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
883 if (!io_req_cache_empty(ctx))
887 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
890 * Bulk alloc is all-or-nothing. If we fail to get a batch,
891 * retry single alloc to be on the safe side.
893 if (unlikely(ret <= 0)) {
894 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
900 percpu_ref_get_many(&ctx->refs, ret);
901 for (i = 0; i < ret; i++) {
902 struct io_kiocb *req = reqs[i];
904 io_preinit_req(req, ctx);
905 io_req_add_to_cache(req, ctx);
910 static inline bool io_alloc_req_refill(struct io_ring_ctx *ctx)
912 if (unlikely(io_req_cache_empty(ctx)))
913 return __io_alloc_req_refill(ctx);
917 static inline struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
919 struct io_wq_work_node *node;
921 node = wq_stack_extract(&ctx->submit_state.free_list);
922 return container_of(node, struct io_kiocb, comp_list);
925 static inline void io_dismantle_req(struct io_kiocb *req)
927 unsigned int flags = req->flags;
929 if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
931 if (!(flags & REQ_F_FIXED_FILE))
932 io_put_file(req->file);
935 __cold void io_free_req(struct io_kiocb *req)
937 struct io_ring_ctx *ctx = req->ctx;
939 io_req_put_rsrc(req);
940 io_dismantle_req(req);
941 io_put_task(req->task, 1);
943 spin_lock(&ctx->completion_lock);
944 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
945 ctx->locked_free_nr++;
946 spin_unlock(&ctx->completion_lock);
949 static void __io_req_find_next_prep(struct io_kiocb *req)
951 struct io_ring_ctx *ctx = req->ctx;
955 io_cq_unlock_post(ctx);
958 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
960 struct io_kiocb *nxt;
963 * If LINK is set, we have dependent requests in this chain. If we
964 * didn't fail this request, queue the first one up, moving any other
965 * dependencies to the next request. In case of failure, fail the rest
968 if (unlikely(req->flags & IO_DISARM_MASK))
969 __io_req_find_next_prep(req);
975 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
979 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
980 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
982 io_submit_flush_completions(ctx);
983 mutex_unlock(&ctx->uring_lock);
986 percpu_ref_put(&ctx->refs);
989 static unsigned int handle_tw_list(struct llist_node *node,
990 struct io_ring_ctx **ctx, bool *locked,
991 struct llist_node *last)
993 unsigned int count = 0;
995 while (node != last) {
996 struct llist_node *next = node->next;
997 struct io_kiocb *req = container_of(node, struct io_kiocb,
1000 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1002 if (req->ctx != *ctx) {
1003 ctx_flush_and_put(*ctx, locked);
1005 /* if not contended, grab and improve batching */
1006 *locked = mutex_trylock(&(*ctx)->uring_lock);
1007 percpu_ref_get(&(*ctx)->refs);
1009 req->io_task_work.func(req, locked);
1018 * io_llist_xchg - swap all entries in a lock-less list
1019 * @head: the head of lock-less list to delete all entries
1020 * @new: new entry as the head of the list
1022 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1023 * The order of entries returned is from the newest to the oldest added one.
1025 static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1026 struct llist_node *new)
1028 return xchg(&head->first, new);
1032 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1033 * @head: the head of lock-less list to delete all entries
1034 * @old: expected old value of the first entry of the list
1035 * @new: new entry as the head of the list
1037 * perform a cmpxchg on the first entry of the list.
1040 static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1041 struct llist_node *old,
1042 struct llist_node *new)
1044 return cmpxchg(&head->first, old, new);
1047 void tctx_task_work(struct callback_head *cb)
1049 bool uring_locked = false;
1050 struct io_ring_ctx *ctx = NULL;
1051 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1053 struct llist_node fake = {};
1054 struct llist_node *node = io_llist_xchg(&tctx->task_list, &fake);
1055 unsigned int loops = 1;
1056 unsigned int count = handle_tw_list(node, &ctx, &uring_locked, NULL);
1058 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1059 while (node != &fake) {
1061 node = io_llist_xchg(&tctx->task_list, &fake);
1062 count += handle_tw_list(node, &ctx, &uring_locked, &fake);
1063 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1066 ctx_flush_and_put(ctx, &uring_locked);
1068 /* relaxed read is enough as only the task itself sets ->in_idle */
1069 if (unlikely(atomic_read(&tctx->in_idle)))
1070 io_uring_drop_tctx_refs(current);
1072 trace_io_uring_task_work_run(tctx, count, loops);
1075 void io_req_task_work_add(struct io_kiocb *req)
1077 struct io_uring_task *tctx = req->task->io_uring;
1078 struct io_ring_ctx *ctx = req->ctx;
1079 struct llist_node *node;
1082 running = !llist_add(&req->io_task_work.node, &tctx->task_list);
1084 /* task_work already pending, we're done */
1088 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1089 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1091 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1094 node = llist_del_all(&tctx->task_list);
1097 req = container_of(node, struct io_kiocb, io_task_work.node);
1099 if (llist_add(&req->io_task_work.node,
1100 &req->ctx->fallback_llist))
1101 schedule_delayed_work(&req->ctx->fallback_work, 1);
1105 static void io_req_tw_post(struct io_kiocb *req, bool *locked)
1107 io_req_complete_post(req);
1110 void io_req_tw_post_queue(struct io_kiocb *req, s32 res, u32 cflags)
1112 io_req_set_res(req, res, cflags);
1113 req->io_task_work.func = io_req_tw_post;
1114 io_req_task_work_add(req);
1117 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
1119 /* not needed for normal modes, but SQPOLL depends on it */
1120 io_tw_lock(req->ctx, locked);
1121 io_req_complete_failed(req, req->cqe.res);
1124 void io_req_task_submit(struct io_kiocb *req, bool *locked)
1126 io_tw_lock(req->ctx, locked);
1127 /* req->task == current here, checking PF_EXITING is safe */
1128 if (likely(!(req->task->flags & PF_EXITING)))
1131 io_req_complete_failed(req, -EFAULT);
1134 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1136 io_req_set_res(req, ret, 0);
1137 req->io_task_work.func = io_req_task_cancel;
1138 io_req_task_work_add(req);
1141 void io_req_task_queue(struct io_kiocb *req)
1143 req->io_task_work.func = io_req_task_submit;
1144 io_req_task_work_add(req);
1147 void io_queue_next(struct io_kiocb *req)
1149 struct io_kiocb *nxt = io_req_find_next(req);
1152 io_req_task_queue(nxt);
1155 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1156 __must_hold(&ctx->uring_lock)
1158 struct task_struct *task = NULL;
1162 struct io_kiocb *req = container_of(node, struct io_kiocb,
1165 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1166 if (req->flags & REQ_F_REFCOUNT) {
1167 node = req->comp_list.next;
1168 if (!req_ref_put_and_test(req))
1171 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1172 struct async_poll *apoll = req->apoll;
1174 if (apoll->double_poll)
1175 kfree(apoll->double_poll);
1176 list_add(&apoll->poll.wait.entry,
1178 req->flags &= ~REQ_F_POLLED;
1180 if (req->flags & IO_REQ_LINK_FLAGS)
1182 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1185 if (!(req->flags & REQ_F_FIXED_FILE))
1186 io_put_file(req->file);
1188 io_req_put_rsrc_locked(req, ctx);
1190 if (req->task != task) {
1192 io_put_task(task, task_refs);
1197 node = req->comp_list.next;
1198 io_req_add_to_cache(req, ctx);
1202 io_put_task(task, task_refs);
1205 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1206 __must_hold(&ctx->uring_lock)
1208 struct io_wq_work_node *node, *prev;
1209 struct io_submit_state *state = &ctx->submit_state;
1211 spin_lock(&ctx->completion_lock);
1212 wq_list_for_each(node, prev, &state->compl_reqs) {
1213 struct io_kiocb *req = container_of(node, struct io_kiocb,
1216 if (!(req->flags & REQ_F_CQE_SKIP))
1217 __io_fill_cqe_req(ctx, req);
1219 __io_cq_unlock_post(ctx);
1221 io_free_batch_list(ctx, state->compl_reqs.first);
1222 INIT_WQ_LIST(&state->compl_reqs);
1226 * Drop reference to request, return next in chain (if there is one) if this
1227 * was the last reference to this request.
1229 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1231 struct io_kiocb *nxt = NULL;
1233 if (req_ref_put_and_test(req)) {
1234 if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
1235 nxt = io_req_find_next(req);
1241 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1243 /* See comment at the top of this file */
1245 return __io_cqring_events(ctx);
1249 * We can't just wait for polled events to come to us, we have to actively
1250 * find and complete them.
1252 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1254 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1257 mutex_lock(&ctx->uring_lock);
1258 while (!wq_list_empty(&ctx->iopoll_list)) {
1259 /* let it sleep and repeat later if can't complete a request */
1260 if (io_do_iopoll(ctx, true) == 0)
1263 * Ensure we allow local-to-the-cpu processing to take place,
1264 * in this case we need to ensure that we reap all events.
1265 * Also let task_work, etc. to progress by releasing the mutex
1267 if (need_resched()) {
1268 mutex_unlock(&ctx->uring_lock);
1270 mutex_lock(&ctx->uring_lock);
1273 mutex_unlock(&ctx->uring_lock);
1276 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1278 unsigned int nr_events = 0;
1280 unsigned long check_cq;
1282 check_cq = READ_ONCE(ctx->check_cq);
1283 if (unlikely(check_cq)) {
1284 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1285 __io_cqring_overflow_flush(ctx, false);
1287 * Similarly do not spin if we have not informed the user of any
1290 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1294 * Don't enter poll loop if we already have events pending.
1295 * If we do, we can potentially be spinning for commands that
1296 * already triggered a CQE (eg in error).
1298 if (io_cqring_events(ctx))
1303 * If a submit got punted to a workqueue, we can have the
1304 * application entering polling for a command before it gets
1305 * issued. That app will hold the uring_lock for the duration
1306 * of the poll right here, so we need to take a breather every
1307 * now and then to ensure that the issue has a chance to add
1308 * the poll to the issued list. Otherwise we can spin here
1309 * forever, while the workqueue is stuck trying to acquire the
1312 if (wq_list_empty(&ctx->iopoll_list)) {
1313 u32 tail = ctx->cached_cq_tail;
1315 mutex_unlock(&ctx->uring_lock);
1317 mutex_lock(&ctx->uring_lock);
1319 /* some requests don't go through iopoll_list */
1320 if (tail != ctx->cached_cq_tail ||
1321 wq_list_empty(&ctx->iopoll_list))
1324 ret = io_do_iopoll(ctx, !min);
1329 } while (nr_events < min && !need_resched());
1334 void io_req_task_complete(struct io_kiocb *req, bool *locked)
1336 if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) {
1337 unsigned issue_flags = *locked ? 0 : IO_URING_F_UNLOCKED;
1339 req->cqe.flags |= io_put_kbuf(req, issue_flags);
1343 io_req_complete_defer(req);
1345 io_req_complete_post(req);
1349 * After the iocb has been issued, it's safe to be found on the poll list.
1350 * Adding the kiocb to the list AFTER submission ensures that we don't
1351 * find it from a io_do_iopoll() thread before the issuer is done
1352 * accessing the kiocb cookie.
1354 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1356 struct io_ring_ctx *ctx = req->ctx;
1357 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1359 /* workqueue context doesn't hold uring_lock, grab it now */
1360 if (unlikely(needs_lock))
1361 mutex_lock(&ctx->uring_lock);
1364 * Track whether we have multiple files in our lists. This will impact
1365 * how we do polling eventually, not spinning if we're on potentially
1366 * different devices.
1368 if (wq_list_empty(&ctx->iopoll_list)) {
1369 ctx->poll_multi_queue = false;
1370 } else if (!ctx->poll_multi_queue) {
1371 struct io_kiocb *list_req;
1373 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1375 if (list_req->file != req->file)
1376 ctx->poll_multi_queue = true;
1380 * For fast devices, IO may have already completed. If it has, add
1381 * it to the front so we find it first.
1383 if (READ_ONCE(req->iopoll_completed))
1384 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1386 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1388 if (unlikely(needs_lock)) {
1390 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1391 * in sq thread task context or in io worker task context. If
1392 * current task context is sq thread, we don't need to check
1393 * whether should wake up sq thread.
1395 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1396 wq_has_sleeper(&ctx->sq_data->wait))
1397 wake_up(&ctx->sq_data->wait);
1399 mutex_unlock(&ctx->uring_lock);
1403 static bool io_bdev_nowait(struct block_device *bdev)
1405 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
1409 * If we tracked the file through the SCM inflight mechanism, we could support
1410 * any file. For now, just ensure that anything potentially problematic is done
1413 static bool __io_file_supports_nowait(struct file *file, umode_t mode)
1415 if (S_ISBLK(mode)) {
1416 if (IS_ENABLED(CONFIG_BLOCK) &&
1417 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
1423 if (S_ISREG(mode)) {
1424 if (IS_ENABLED(CONFIG_BLOCK) &&
1425 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
1426 !io_is_uring_fops(file))
1431 /* any ->read/write should understand O_NONBLOCK */
1432 if (file->f_flags & O_NONBLOCK)
1434 return file->f_mode & FMODE_NOWAIT;
1438 * If we tracked the file through the SCM inflight mechanism, we could support
1439 * any file. For now, just ensure that anything potentially problematic is done
1442 unsigned int io_file_get_flags(struct file *file)
1444 umode_t mode = file_inode(file)->i_mode;
1445 unsigned int res = 0;
1449 if (__io_file_supports_nowait(file, mode))
1451 if (io_file_need_scm(file))
1456 bool io_alloc_async_data(struct io_kiocb *req)
1458 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
1459 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
1460 if (req->async_data) {
1461 req->flags |= REQ_F_ASYNC_DATA;
1467 int io_req_prep_async(struct io_kiocb *req)
1469 const struct io_op_def *def = &io_op_defs[req->opcode];
1471 /* assign early for deferred execution for non-fixed file */
1472 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE))
1473 req->file = io_file_get_normal(req, req->cqe.fd);
1474 if (!def->prep_async)
1476 if (WARN_ON_ONCE(req_has_async_data(req)))
1478 if (io_alloc_async_data(req))
1481 return def->prep_async(req);
1484 static u32 io_get_sequence(struct io_kiocb *req)
1486 u32 seq = req->ctx->cached_sq_head;
1487 struct io_kiocb *cur;
1489 /* need original cached_sq_head, but it was increased for each req */
1490 io_for_each_link(cur, req)
1495 static __cold void io_drain_req(struct io_kiocb *req)
1497 struct io_ring_ctx *ctx = req->ctx;
1498 struct io_defer_entry *de;
1500 u32 seq = io_get_sequence(req);
1502 /* Still need defer if there is pending req in defer list. */
1503 spin_lock(&ctx->completion_lock);
1504 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1505 spin_unlock(&ctx->completion_lock);
1507 ctx->drain_active = false;
1508 io_req_task_queue(req);
1511 spin_unlock(&ctx->completion_lock);
1513 ret = io_req_prep_async(req);
1516 io_req_complete_failed(req, ret);
1519 io_prep_async_link(req);
1520 de = kmalloc(sizeof(*de), GFP_KERNEL);
1526 spin_lock(&ctx->completion_lock);
1527 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1528 spin_unlock(&ctx->completion_lock);
1533 trace_io_uring_defer(req);
1536 list_add_tail(&de->list, &ctx->defer_list);
1537 spin_unlock(&ctx->completion_lock);
1540 static void io_clean_op(struct io_kiocb *req)
1542 if (req->flags & REQ_F_BUFFER_SELECTED) {
1543 spin_lock(&req->ctx->completion_lock);
1544 io_put_kbuf_comp(req);
1545 spin_unlock(&req->ctx->completion_lock);
1548 if (req->flags & REQ_F_NEED_CLEANUP) {
1549 const struct io_op_def *def = &io_op_defs[req->opcode];
1554 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1555 kfree(req->apoll->double_poll);
1559 if (req->flags & REQ_F_INFLIGHT) {
1560 struct io_uring_task *tctx = req->task->io_uring;
1562 atomic_dec(&tctx->inflight_tracked);
1564 if (req->flags & REQ_F_CREDS)
1565 put_cred(req->creds);
1566 if (req->flags & REQ_F_ASYNC_DATA) {
1567 kfree(req->async_data);
1568 req->async_data = NULL;
1570 req->flags &= ~IO_REQ_CLEAN_FLAGS;
1573 static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags)
1575 if (req->file || !io_op_defs[req->opcode].needs_file)
1578 if (req->flags & REQ_F_FIXED_FILE)
1579 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1581 req->file = io_file_get_normal(req, req->cqe.fd);
1586 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1588 const struct io_op_def *def = &io_op_defs[req->opcode];
1589 const struct cred *creds = NULL;
1592 if (unlikely(!io_assign_file(req, issue_flags)))
1595 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1596 creds = override_creds(req->creds);
1598 if (!def->audit_skip)
1599 audit_uring_entry(req->opcode);
1601 ret = def->issue(req, issue_flags);
1603 if (!def->audit_skip)
1604 audit_uring_exit(!ret, ret);
1607 revert_creds(creds);
1609 if (ret == IOU_OK) {
1610 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1611 io_req_complete_defer(req);
1613 io_req_complete_post(req);
1614 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1617 /* If the op doesn't have a file, we're not polling for it */
1618 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && req->file)
1619 io_iopoll_req_issued(req, issue_flags);
1624 int io_poll_issue(struct io_kiocb *req, bool *locked)
1626 io_tw_lock(req->ctx, locked);
1627 if (unlikely(req->task->flags & PF_EXITING))
1629 return io_issue_sqe(req, IO_URING_F_NONBLOCK);
1632 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1634 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1636 req = io_put_req_find_next(req);
1637 return req ? &req->work : NULL;
1640 void io_wq_submit_work(struct io_wq_work *work)
1642 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1643 const struct io_op_def *def = &io_op_defs[req->opcode];
1644 unsigned int issue_flags = IO_URING_F_UNLOCKED;
1645 bool needs_poll = false;
1646 int ret = 0, err = -ECANCELED;
1648 /* one will be dropped by ->io_free_work() after returning to io-wq */
1649 if (!(req->flags & REQ_F_REFCOUNT))
1650 __io_req_set_refcount(req, 2);
1654 io_arm_ltimeout(req);
1656 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1657 if (work->flags & IO_WQ_WORK_CANCEL) {
1659 io_req_task_queue_fail(req, err);
1662 if (!io_assign_file(req, issue_flags)) {
1664 work->flags |= IO_WQ_WORK_CANCEL;
1668 if (req->flags & REQ_F_FORCE_ASYNC) {
1669 bool opcode_poll = def->pollin || def->pollout;
1671 if (opcode_poll && file_can_poll(req->file)) {
1673 issue_flags |= IO_URING_F_NONBLOCK;
1678 ret = io_issue_sqe(req, issue_flags);
1682 * We can get EAGAIN for iopolled IO even though we're
1683 * forcing a sync submission from here, since we can't
1684 * wait for request slots on the block side.
1687 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1693 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1695 /* aborted or ready, in either case retry blocking */
1697 issue_flags &= ~IO_URING_F_NONBLOCK;
1700 /* avoid locking problems by failing it from a clean context */
1702 io_req_task_queue_fail(req, ret);
1705 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1706 unsigned int issue_flags)
1708 struct io_ring_ctx *ctx = req->ctx;
1709 struct file *file = NULL;
1710 unsigned long file_ptr;
1712 io_ring_submit_lock(ctx, issue_flags);
1714 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
1716 fd = array_index_nospec(fd, ctx->nr_user_files);
1717 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
1718 file = (struct file *) (file_ptr & FFS_MASK);
1719 file_ptr &= ~FFS_MASK;
1720 /* mask in overlapping REQ_F and FFS bits */
1721 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
1722 io_req_set_rsrc_node(req, ctx, 0);
1723 WARN_ON_ONCE(file && !test_bit(fd, ctx->file_table.bitmap));
1725 io_ring_submit_unlock(ctx, issue_flags);
1729 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
1731 struct file *file = fget(fd);
1733 trace_io_uring_file_get(req, fd);
1735 /* we don't allow fixed io_uring files */
1736 if (file && io_is_uring_fops(file))
1737 io_req_track_inflight(req);
1741 static void io_queue_async(struct io_kiocb *req, int ret)
1742 __must_hold(&req->ctx->uring_lock)
1744 struct io_kiocb *linked_timeout;
1746 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
1747 io_req_complete_failed(req, ret);
1751 linked_timeout = io_prep_linked_timeout(req);
1753 switch (io_arm_poll_handler(req, 0)) {
1754 case IO_APOLL_READY:
1755 io_req_task_queue(req);
1757 case IO_APOLL_ABORTED:
1759 * Queued up for async execution, worker will release
1760 * submit reference when the iocb is actually submitted.
1762 io_kbuf_recycle(req, 0);
1763 io_queue_iowq(req, NULL);
1770 io_queue_linked_timeout(linked_timeout);
1773 static inline void io_queue_sqe(struct io_kiocb *req)
1774 __must_hold(&req->ctx->uring_lock)
1778 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
1781 * We async punt it if the file wasn't marked NOWAIT, or if the file
1782 * doesn't support non-blocking read/write attempts
1785 io_arm_ltimeout(req);
1787 io_queue_async(req, ret);
1790 static void io_queue_sqe_fallback(struct io_kiocb *req)
1791 __must_hold(&req->ctx->uring_lock)
1793 if (unlikely(req->flags & REQ_F_FAIL)) {
1795 * We don't submit, fail them all, for that replace hardlinks
1796 * with normal links. Extra REQ_F_LINK is tolerated.
1798 req->flags &= ~REQ_F_HARDLINK;
1799 req->flags |= REQ_F_LINK;
1800 io_req_complete_failed(req, req->cqe.res);
1801 } else if (unlikely(req->ctx->drain_active)) {
1804 int ret = io_req_prep_async(req);
1807 io_req_complete_failed(req, ret);
1809 io_queue_iowq(req, NULL);
1814 * Check SQE restrictions (opcode and flags).
1816 * Returns 'true' if SQE is allowed, 'false' otherwise.
1818 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
1819 struct io_kiocb *req,
1820 unsigned int sqe_flags)
1822 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
1825 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
1826 ctx->restrictions.sqe_flags_required)
1829 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
1830 ctx->restrictions.sqe_flags_required))
1836 static void io_init_req_drain(struct io_kiocb *req)
1838 struct io_ring_ctx *ctx = req->ctx;
1839 struct io_kiocb *head = ctx->submit_state.link.head;
1841 ctx->drain_active = true;
1844 * If we need to drain a request in the middle of a link, drain
1845 * the head request and the next request/link after the current
1846 * link. Considering sequential execution of links,
1847 * REQ_F_IO_DRAIN will be maintained for every request of our
1850 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
1851 ctx->drain_next = true;
1855 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
1856 const struct io_uring_sqe *sqe)
1857 __must_hold(&ctx->uring_lock)
1859 const struct io_op_def *def;
1860 unsigned int sqe_flags;
1864 /* req is partially pre-initialised, see io_preinit_req() */
1865 req->opcode = opcode = READ_ONCE(sqe->opcode);
1866 /* same numerical values with corresponding REQ_F_*, safe to copy */
1867 req->flags = sqe_flags = READ_ONCE(sqe->flags);
1868 req->cqe.user_data = READ_ONCE(sqe->user_data);
1870 req->rsrc_node = NULL;
1871 req->task = current;
1873 if (unlikely(opcode >= IORING_OP_LAST)) {
1877 def = &io_op_defs[opcode];
1878 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
1879 /* enforce forwards compatibility on users */
1880 if (sqe_flags & ~SQE_VALID_FLAGS)
1882 if (sqe_flags & IOSQE_BUFFER_SELECT) {
1883 if (!def->buffer_select)
1885 req->buf_index = READ_ONCE(sqe->buf_group);
1887 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
1888 ctx->drain_disabled = true;
1889 if (sqe_flags & IOSQE_IO_DRAIN) {
1890 if (ctx->drain_disabled)
1892 io_init_req_drain(req);
1895 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
1896 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
1898 /* knock it to the slow queue path, will be drained there */
1899 if (ctx->drain_active)
1900 req->flags |= REQ_F_FORCE_ASYNC;
1901 /* if there is no link, we're at "next" request and need to drain */
1902 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
1903 ctx->drain_next = false;
1904 ctx->drain_active = true;
1905 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
1909 if (!def->ioprio && sqe->ioprio)
1911 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
1914 if (def->needs_file) {
1915 struct io_submit_state *state = &ctx->submit_state;
1917 req->cqe.fd = READ_ONCE(sqe->fd);
1920 * Plug now if we have more than 2 IO left after this, and the
1921 * target is potentially a read/write to block based storage.
1923 if (state->need_plug && def->plug) {
1924 state->plug_started = true;
1925 state->need_plug = false;
1926 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
1930 personality = READ_ONCE(sqe->personality);
1934 req->creds = xa_load(&ctx->personalities, personality);
1937 get_cred(req->creds);
1938 ret = security_uring_override_creds(req->creds);
1940 put_cred(req->creds);
1943 req->flags |= REQ_F_CREDS;
1946 return def->prep(req, sqe);
1949 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
1950 struct io_kiocb *req, int ret)
1952 struct io_ring_ctx *ctx = req->ctx;
1953 struct io_submit_link *link = &ctx->submit_state.link;
1954 struct io_kiocb *head = link->head;
1956 trace_io_uring_req_failed(sqe, req, ret);
1959 * Avoid breaking links in the middle as it renders links with SQPOLL
1960 * unusable. Instead of failing eagerly, continue assembling the link if
1961 * applicable and mark the head with REQ_F_FAIL. The link flushing code
1962 * should find the flag and handle the rest.
1964 req_fail_link_node(req, ret);
1965 if (head && !(head->flags & REQ_F_FAIL))
1966 req_fail_link_node(head, -ECANCELED);
1968 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
1970 link->last->link = req;
1974 io_queue_sqe_fallback(req);
1979 link->last->link = req;
1986 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
1987 const struct io_uring_sqe *sqe)
1988 __must_hold(&ctx->uring_lock)
1990 struct io_submit_link *link = &ctx->submit_state.link;
1993 ret = io_init_req(ctx, req, sqe);
1995 return io_submit_fail_init(sqe, req, ret);
1997 /* don't need @sqe from now on */
1998 trace_io_uring_submit_sqe(req, true);
2001 * If we already have a head request, queue this one for async
2002 * submittal once the head completes. If we don't have a head but
2003 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2004 * submitted sync once the chain is complete. If none of those
2005 * conditions are true (normal request), then just queue it.
2007 if (unlikely(link->head)) {
2008 ret = io_req_prep_async(req);
2010 return io_submit_fail_init(sqe, req, ret);
2012 trace_io_uring_link(req, link->head);
2013 link->last->link = req;
2016 if (req->flags & IO_REQ_LINK_FLAGS)
2018 /* last request of the link, flush it */
2021 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2024 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2025 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2026 if (req->flags & IO_REQ_LINK_FLAGS) {
2031 io_queue_sqe_fallback(req);
2041 * Batched submission is done, ensure local IO is flushed out.
2043 static void io_submit_state_end(struct io_ring_ctx *ctx)
2045 struct io_submit_state *state = &ctx->submit_state;
2047 if (unlikely(state->link.head))
2048 io_queue_sqe_fallback(state->link.head);
2049 /* flush only after queuing links as they can generate completions */
2050 io_submit_flush_completions(ctx);
2051 if (state->plug_started)
2052 blk_finish_plug(&state->plug);
2056 * Start submission side cache.
2058 static void io_submit_state_start(struct io_submit_state *state,
2059 unsigned int max_ios)
2061 state->plug_started = false;
2062 state->need_plug = max_ios > 2;
2063 state->submit_nr = max_ios;
2064 /* set only head, no need to init link_last in advance */
2065 state->link.head = NULL;
2068 static void io_commit_sqring(struct io_ring_ctx *ctx)
2070 struct io_rings *rings = ctx->rings;
2073 * Ensure any loads from the SQEs are done at this point,
2074 * since once we write the new head, the application could
2075 * write new data to them.
2077 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2081 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2082 * that is mapped by userspace. This means that care needs to be taken to
2083 * ensure that reads are stable, as we cannot rely on userspace always
2084 * being a good citizen. If members of the sqe are validated and then later
2085 * used, it's important that those reads are done through READ_ONCE() to
2086 * prevent a re-load down the line.
2088 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
2090 unsigned head, mask = ctx->sq_entries - 1;
2091 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2094 * The cached sq head (or cq tail) serves two purposes:
2096 * 1) allows us to batch the cost of updating the user visible
2098 * 2) allows the kernel side to track the head on its own, even
2099 * though the application is the one updating it.
2101 head = READ_ONCE(ctx->sq_array[sq_idx]);
2102 if (likely(head < ctx->sq_entries)) {
2103 /* double index for 128-byte SQEs, twice as long */
2104 if (ctx->flags & IORING_SETUP_SQE128)
2106 return &ctx->sq_sqes[head];
2109 /* drop invalid entries */
2111 WRITE_ONCE(ctx->rings->sq_dropped,
2112 READ_ONCE(ctx->rings->sq_dropped) + 1);
2116 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2117 __must_hold(&ctx->uring_lock)
2119 unsigned int entries = io_sqring_entries(ctx);
2123 if (unlikely(!entries))
2125 /* make sure SQ entry isn't read before tail */
2126 ret = left = min3(nr, ctx->sq_entries, entries);
2127 io_get_task_refs(left);
2128 io_submit_state_start(&ctx->submit_state, left);
2131 const struct io_uring_sqe *sqe;
2132 struct io_kiocb *req;
2134 if (unlikely(!io_alloc_req_refill(ctx)))
2136 req = io_alloc_req(ctx);
2137 sqe = io_get_sqe(ctx);
2138 if (unlikely(!sqe)) {
2139 io_req_add_to_cache(req, ctx);
2144 * Continue submitting even for sqe failure if the
2145 * ring was setup with IORING_SETUP_SUBMIT_ALL
2147 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2148 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2154 if (unlikely(left)) {
2156 /* try again if it submitted nothing and can't allocate a req */
2157 if (!ret && io_req_cache_empty(ctx))
2159 current->io_uring->cached_refs += left;
2162 io_submit_state_end(ctx);
2163 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2164 io_commit_sqring(ctx);
2168 struct io_wait_queue {
2169 struct wait_queue_entry wq;
2170 struct io_ring_ctx *ctx;
2172 unsigned nr_timeouts;
2175 static inline bool io_should_wake(struct io_wait_queue *iowq)
2177 struct io_ring_ctx *ctx = iowq->ctx;
2178 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
2181 * Wake up if we have enough events, or if a timeout occurred since we
2182 * started waiting. For timeouts, we always want to return to userspace,
2183 * regardless of event count.
2185 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2188 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2189 int wake_flags, void *key)
2191 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
2195 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2196 * the task, and the next invocation will do it.
2198 if (io_should_wake(iowq) ||
2199 test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &iowq->ctx->check_cq))
2200 return autoremove_wake_function(curr, mode, wake_flags, key);
2204 int io_run_task_work_sig(void)
2206 if (io_run_task_work())
2208 if (task_sigpending(current))
2213 /* when returns >0, the caller should retry */
2214 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2215 struct io_wait_queue *iowq,
2219 unsigned long check_cq;
2221 /* make sure we run task_work before checking for signals */
2222 ret = io_run_task_work_sig();
2223 if (ret || io_should_wake(iowq))
2226 check_cq = READ_ONCE(ctx->check_cq);
2227 if (unlikely(check_cq)) {
2228 /* let the caller flush overflows, retry */
2229 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2231 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
2234 if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS))
2240 * Wait until events become available, if we don't already have some. The
2241 * application must reap them itself, as they reside on the shared cq ring.
2243 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2244 const sigset_t __user *sig, size_t sigsz,
2245 struct __kernel_timespec __user *uts)
2247 struct io_wait_queue iowq;
2248 struct io_rings *rings = ctx->rings;
2249 ktime_t timeout = KTIME_MAX;
2253 io_cqring_overflow_flush(ctx);
2254 if (io_cqring_events(ctx) >= min_events)
2256 if (!io_run_task_work())
2261 #ifdef CONFIG_COMPAT
2262 if (in_compat_syscall())
2263 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2267 ret = set_user_sigmask(sig, sigsz);
2274 struct timespec64 ts;
2276 if (get_timespec64(&ts, uts))
2278 timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2281 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2282 iowq.wq.private = current;
2283 INIT_LIST_HEAD(&iowq.wq.entry);
2285 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2286 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2288 trace_io_uring_cqring_wait(ctx, min_events);
2290 /* if we can't even flush overflow, don't wait for more */
2291 if (!io_cqring_overflow_flush(ctx)) {
2295 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2296 TASK_INTERRUPTIBLE);
2297 ret = io_cqring_wait_schedule(ctx, &iowq, timeout);
2301 finish_wait(&ctx->cq_wait, &iowq.wq);
2302 restore_saved_sigmask_unless(ret == -EINTR);
2304 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2307 static void io_mem_free(void *ptr)
2314 page = virt_to_head_page(ptr);
2315 if (put_page_testzero(page))
2316 free_compound_page(page);
2319 static void *io_mem_alloc(size_t size)
2321 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2323 return (void *) __get_free_pages(gfp, get_order(size));
2326 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2327 unsigned int cq_entries, size_t *sq_offset)
2329 struct io_rings *rings;
2330 size_t off, sq_array_size;
2332 off = struct_size(rings, cqes, cq_entries);
2333 if (off == SIZE_MAX)
2335 if (ctx->flags & IORING_SETUP_CQE32) {
2336 if (check_shl_overflow(off, 1, &off))
2341 off = ALIGN(off, SMP_CACHE_BYTES);
2349 sq_array_size = array_size(sizeof(u32), sq_entries);
2350 if (sq_array_size == SIZE_MAX)
2353 if (check_add_overflow(off, sq_array_size, &off))
2359 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2360 unsigned int eventfd_async)
2362 struct io_ev_fd *ev_fd;
2363 __s32 __user *fds = arg;
2366 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2367 lockdep_is_held(&ctx->uring_lock));
2371 if (copy_from_user(&fd, fds, sizeof(*fds)))
2374 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2378 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2379 if (IS_ERR(ev_fd->cq_ev_fd)) {
2380 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2385 spin_lock(&ctx->completion_lock);
2386 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2387 spin_unlock(&ctx->completion_lock);
2389 ev_fd->eventfd_async = eventfd_async;
2390 ctx->has_evfd = true;
2391 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2395 static void io_eventfd_put(struct rcu_head *rcu)
2397 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
2399 eventfd_ctx_put(ev_fd->cq_ev_fd);
2403 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2405 struct io_ev_fd *ev_fd;
2407 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2408 lockdep_is_held(&ctx->uring_lock));
2410 ctx->has_evfd = false;
2411 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2412 call_rcu(&ev_fd->rcu, io_eventfd_put);
2419 static void io_req_caches_free(struct io_ring_ctx *ctx)
2421 struct io_submit_state *state = &ctx->submit_state;
2424 mutex_lock(&ctx->uring_lock);
2425 io_flush_cached_locked_reqs(ctx, state);
2427 while (!io_req_cache_empty(ctx)) {
2428 struct io_wq_work_node *node;
2429 struct io_kiocb *req;
2431 node = wq_stack_extract(&state->free_list);
2432 req = container_of(node, struct io_kiocb, comp_list);
2433 kmem_cache_free(req_cachep, req);
2437 percpu_ref_put_many(&ctx->refs, nr);
2438 mutex_unlock(&ctx->uring_lock);
2441 static void io_flush_apoll_cache(struct io_ring_ctx *ctx)
2443 struct async_poll *apoll;
2445 while (!list_empty(&ctx->apoll_cache)) {
2446 apoll = list_first_entry(&ctx->apoll_cache, struct async_poll,
2448 list_del(&apoll->poll.wait.entry);
2453 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2455 io_sq_thread_finish(ctx);
2457 if (ctx->mm_account) {
2458 mmdrop(ctx->mm_account);
2459 ctx->mm_account = NULL;
2462 io_rsrc_refs_drop(ctx);
2463 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2464 io_wait_rsrc_data(ctx->buf_data);
2465 io_wait_rsrc_data(ctx->file_data);
2467 mutex_lock(&ctx->uring_lock);
2469 __io_sqe_buffers_unregister(ctx);
2471 __io_sqe_files_unregister(ctx);
2473 __io_cqring_overflow_flush(ctx, true);
2474 io_eventfd_unregister(ctx);
2475 io_flush_apoll_cache(ctx);
2476 mutex_unlock(&ctx->uring_lock);
2477 io_destroy_buffers(ctx);
2479 put_cred(ctx->sq_creds);
2480 if (ctx->submitter_task)
2481 put_task_struct(ctx->submitter_task);
2483 /* there are no registered resources left, nobody uses it */
2485 io_rsrc_node_destroy(ctx->rsrc_node);
2486 if (ctx->rsrc_backup_node)
2487 io_rsrc_node_destroy(ctx->rsrc_backup_node);
2488 flush_delayed_work(&ctx->rsrc_put_work);
2489 flush_delayed_work(&ctx->fallback_work);
2491 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2492 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
2494 #if defined(CONFIG_UNIX)
2495 if (ctx->ring_sock) {
2496 ctx->ring_sock->file = NULL; /* so that iput() is called */
2497 sock_release(ctx->ring_sock);
2500 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2502 io_mem_free(ctx->rings);
2503 io_mem_free(ctx->sq_sqes);
2505 percpu_ref_exit(&ctx->refs);
2506 free_uid(ctx->user);
2507 io_req_caches_free(ctx);
2509 io_wq_put_hash(ctx->hash_map);
2510 kfree(ctx->cancel_table.hbs);
2511 kfree(ctx->cancel_table_locked.hbs);
2512 kfree(ctx->dummy_ubuf);
2514 xa_destroy(&ctx->io_bl_xa);
2518 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2520 struct io_ring_ctx *ctx = file->private_data;
2523 poll_wait(file, &ctx->cq_wait, wait);
2525 * synchronizes with barrier from wq_has_sleeper call in
2529 if (!io_sqring_full(ctx))
2530 mask |= EPOLLOUT | EPOLLWRNORM;
2533 * Don't flush cqring overflow list here, just do a simple check.
2534 * Otherwise there could possible be ABBA deadlock:
2537 * lock(&ctx->uring_lock);
2539 * lock(&ctx->uring_lock);
2542 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2543 * pushs them to do the flush.
2545 if (io_cqring_events(ctx) ||
2546 test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
2547 mask |= EPOLLIN | EPOLLRDNORM;
2552 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2554 const struct cred *creds;
2556 creds = xa_erase(&ctx->personalities, id);
2565 struct io_tctx_exit {
2566 struct callback_head task_work;
2567 struct completion completion;
2568 struct io_ring_ctx *ctx;
2571 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2573 struct io_uring_task *tctx = current->io_uring;
2574 struct io_tctx_exit *work;
2576 work = container_of(cb, struct io_tctx_exit, task_work);
2578 * When @in_idle, we're in cancellation and it's racy to remove the
2579 * node. It'll be removed by the end of cancellation, just ignore it.
2581 if (!atomic_read(&tctx->in_idle))
2582 io_uring_del_tctx_node((unsigned long)work->ctx);
2583 complete(&work->completion);
2586 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2588 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2590 return req->ctx == data;
2593 static __cold void io_ring_exit_work(struct work_struct *work)
2595 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2596 unsigned long timeout = jiffies + HZ * 60 * 5;
2597 unsigned long interval = HZ / 20;
2598 struct io_tctx_exit exit;
2599 struct io_tctx_node *node;
2603 * If we're doing polled IO and end up having requests being
2604 * submitted async (out-of-line), then completions can come in while
2605 * we're waiting for refs to drop. We need to reap these manually,
2606 * as nobody else will be looking for them.
2609 while (io_uring_try_cancel_requests(ctx, NULL, true))
2613 struct io_sq_data *sqd = ctx->sq_data;
2614 struct task_struct *tsk;
2616 io_sq_thread_park(sqd);
2618 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
2619 io_wq_cancel_cb(tsk->io_uring->io_wq,
2620 io_cancel_ctx_cb, ctx, true);
2621 io_sq_thread_unpark(sqd);
2624 io_req_caches_free(ctx);
2626 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
2627 /* there is little hope left, don't run it too often */
2630 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
2632 init_completion(&exit.completion);
2633 init_task_work(&exit.task_work, io_tctx_exit_cb);
2636 * Some may use context even when all refs and requests have been put,
2637 * and they are free to do so while still holding uring_lock or
2638 * completion_lock, see io_req_task_submit(). Apart from other work,
2639 * this lock/unlock section also waits them to finish.
2641 mutex_lock(&ctx->uring_lock);
2642 while (!list_empty(&ctx->tctx_list)) {
2643 WARN_ON_ONCE(time_after(jiffies, timeout));
2645 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
2647 /* don't spin on a single task if cancellation failed */
2648 list_rotate_left(&ctx->tctx_list);
2649 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
2650 if (WARN_ON_ONCE(ret))
2653 mutex_unlock(&ctx->uring_lock);
2654 wait_for_completion(&exit.completion);
2655 mutex_lock(&ctx->uring_lock);
2657 mutex_unlock(&ctx->uring_lock);
2658 spin_lock(&ctx->completion_lock);
2659 spin_unlock(&ctx->completion_lock);
2661 io_ring_ctx_free(ctx);
2664 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
2666 unsigned long index;
2667 struct creds *creds;
2669 mutex_lock(&ctx->uring_lock);
2670 percpu_ref_kill(&ctx->refs);
2672 __io_cqring_overflow_flush(ctx, true);
2673 xa_for_each(&ctx->personalities, index, creds)
2674 io_unregister_personality(ctx, index);
2676 io_poll_remove_all(ctx, NULL, true);
2677 mutex_unlock(&ctx->uring_lock);
2679 /* failed during ring init, it couldn't have issued any requests */
2681 io_kill_timeouts(ctx, NULL, true);
2682 /* if we failed setting up the ctx, we might not have any rings */
2683 io_iopoll_try_reap_events(ctx);
2686 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
2688 * Use system_unbound_wq to avoid spawning tons of event kworkers
2689 * if we're exiting a ton of rings at the same time. It just adds
2690 * noise and overhead, there's no discernable change in runtime
2691 * over using system_wq.
2693 queue_work(system_unbound_wq, &ctx->exit_work);
2696 static int io_uring_release(struct inode *inode, struct file *file)
2698 struct io_ring_ctx *ctx = file->private_data;
2700 file->private_data = NULL;
2701 io_ring_ctx_wait_and_kill(ctx);
2705 struct io_task_cancel {
2706 struct task_struct *task;
2710 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
2712 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2713 struct io_task_cancel *cancel = data;
2715 return io_match_task_safe(req, cancel->task, cancel->all);
2718 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
2719 struct task_struct *task,
2722 struct io_defer_entry *de;
2725 spin_lock(&ctx->completion_lock);
2726 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
2727 if (io_match_task_safe(de->req, task, cancel_all)) {
2728 list_cut_position(&list, &ctx->defer_list, &de->list);
2732 spin_unlock(&ctx->completion_lock);
2733 if (list_empty(&list))
2736 while (!list_empty(&list)) {
2737 de = list_first_entry(&list, struct io_defer_entry, list);
2738 list_del_init(&de->list);
2739 io_req_complete_failed(de->req, -ECANCELED);
2745 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
2747 struct io_tctx_node *node;
2748 enum io_wq_cancel cret;
2751 mutex_lock(&ctx->uring_lock);
2752 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
2753 struct io_uring_task *tctx = node->task->io_uring;
2756 * io_wq will stay alive while we hold uring_lock, because it's
2757 * killed after ctx nodes, which requires to take the lock.
2759 if (!tctx || !tctx->io_wq)
2761 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
2762 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2764 mutex_unlock(&ctx->uring_lock);
2769 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
2770 struct task_struct *task,
2773 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
2774 struct io_uring_task *tctx = task ? task->io_uring : NULL;
2775 enum io_wq_cancel cret;
2778 /* failed during ring init, it couldn't have issued any requests */
2783 ret |= io_uring_try_cancel_iowq(ctx);
2784 } else if (tctx && tctx->io_wq) {
2786 * Cancels requests of all rings, not only @ctx, but
2787 * it's fine as the task is in exit/exec.
2789 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
2791 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2794 /* SQPOLL thread does its own polling */
2795 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
2796 (ctx->sq_data && ctx->sq_data->thread == current)) {
2797 while (!wq_list_empty(&ctx->iopoll_list)) {
2798 io_iopoll_try_reap_events(ctx);
2803 ret |= io_cancel_defer_files(ctx, task, cancel_all);
2804 mutex_lock(&ctx->uring_lock);
2805 ret |= io_poll_remove_all(ctx, task, cancel_all);
2806 mutex_unlock(&ctx->uring_lock);
2807 ret |= io_kill_timeouts(ctx, task, cancel_all);
2809 ret |= io_run_task_work();
2813 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
2816 return atomic_read(&tctx->inflight_tracked);
2817 return percpu_counter_sum(&tctx->inflight);
2821 * Find any io_uring ctx that this task has registered or done IO on, and cancel
2822 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
2824 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
2826 struct io_uring_task *tctx = current->io_uring;
2827 struct io_ring_ctx *ctx;
2831 WARN_ON_ONCE(sqd && sqd->thread != current);
2833 if (!current->io_uring)
2836 io_wq_exit_start(tctx->io_wq);
2838 atomic_inc(&tctx->in_idle);
2842 io_uring_drop_tctx_refs(current);
2843 /* read completions before cancelations */
2844 inflight = tctx_inflight(tctx, !cancel_all);
2849 struct io_tctx_node *node;
2850 unsigned long index;
2852 xa_for_each(&tctx->xa, index, node) {
2853 /* sqpoll task will cancel all its requests */
2854 if (node->ctx->sq_data)
2856 loop |= io_uring_try_cancel_requests(node->ctx,
2857 current, cancel_all);
2860 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
2861 loop |= io_uring_try_cancel_requests(ctx,
2871 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
2873 io_uring_drop_tctx_refs(current);
2876 * If we've seen completions, retry without waiting. This
2877 * avoids a race where a completion comes in before we did
2878 * prepare_to_wait().
2880 if (inflight == tctx_inflight(tctx, !cancel_all))
2882 finish_wait(&tctx->wait, &wait);
2885 io_uring_clean_tctx(tctx);
2888 * We shouldn't run task_works after cancel, so just leave
2889 * ->in_idle set for normal exit.
2891 atomic_dec(&tctx->in_idle);
2892 /* for exec all current's requests should be gone, kill tctx */
2893 __io_uring_free(current);
2897 void __io_uring_cancel(bool cancel_all)
2899 io_uring_cancel_generic(cancel_all, NULL);
2902 static void *io_uring_validate_mmap_request(struct file *file,
2903 loff_t pgoff, size_t sz)
2905 struct io_ring_ctx *ctx = file->private_data;
2906 loff_t offset = pgoff << PAGE_SHIFT;
2911 case IORING_OFF_SQ_RING:
2912 case IORING_OFF_CQ_RING:
2915 case IORING_OFF_SQES:
2919 return ERR_PTR(-EINVAL);
2922 page = virt_to_head_page(ptr);
2923 if (sz > page_size(page))
2924 return ERR_PTR(-EINVAL);
2931 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
2933 size_t sz = vma->vm_end - vma->vm_start;
2937 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
2939 return PTR_ERR(ptr);
2941 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
2942 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
2945 #else /* !CONFIG_MMU */
2947 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
2949 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
2952 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
2954 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
2957 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
2958 unsigned long addr, unsigned long len,
2959 unsigned long pgoff, unsigned long flags)
2963 ptr = io_uring_validate_mmap_request(file, pgoff, len);
2965 return PTR_ERR(ptr);
2967 return (unsigned long) ptr;
2970 #endif /* !CONFIG_MMU */
2972 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
2974 if (flags & IORING_ENTER_EXT_ARG) {
2975 struct io_uring_getevents_arg arg;
2977 if (argsz != sizeof(arg))
2979 if (copy_from_user(&arg, argp, sizeof(arg)))
2985 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
2986 struct __kernel_timespec __user **ts,
2987 const sigset_t __user **sig)
2989 struct io_uring_getevents_arg arg;
2992 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
2993 * is just a pointer to the sigset_t.
2995 if (!(flags & IORING_ENTER_EXT_ARG)) {
2996 *sig = (const sigset_t __user *) argp;
3002 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3003 * timespec and sigset_t pointers if good.
3005 if (*argsz != sizeof(arg))
3007 if (copy_from_user(&arg, argp, sizeof(arg)))
3011 *sig = u64_to_user_ptr(arg.sigmask);
3012 *argsz = arg.sigmask_sz;
3013 *ts = u64_to_user_ptr(arg.ts);
3017 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3018 u32, min_complete, u32, flags, const void __user *, argp,
3021 struct io_ring_ctx *ctx;
3027 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3028 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3029 IORING_ENTER_REGISTERED_RING)))
3033 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3034 * need only dereference our task private array to find it.
3036 if (flags & IORING_ENTER_REGISTERED_RING) {
3037 struct io_uring_task *tctx = current->io_uring;
3039 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3041 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3042 f.file = tctx->registered_rings[fd];
3044 if (unlikely(!f.file))
3048 if (unlikely(!f.file))
3051 if (unlikely(!io_is_uring_fops(f.file)))
3055 ctx = f.file->private_data;
3057 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3061 * For SQ polling, the thread will do all submissions and completions.
3062 * Just return the requested submit count, and wake the thread if
3066 if (ctx->flags & IORING_SETUP_SQPOLL) {
3067 io_cqring_overflow_flush(ctx);
3069 if (unlikely(ctx->sq_data->thread == NULL)) {
3073 if (flags & IORING_ENTER_SQ_WAKEUP)
3074 wake_up(&ctx->sq_data->wait);
3075 if (flags & IORING_ENTER_SQ_WAIT) {
3076 ret = io_sqpoll_wait_sq(ctx);
3081 } else if (to_submit) {
3082 ret = io_uring_add_tctx_node(ctx);
3086 mutex_lock(&ctx->uring_lock);
3087 ret = io_submit_sqes(ctx, to_submit);
3088 if (ret != to_submit) {
3089 mutex_unlock(&ctx->uring_lock);
3092 if ((flags & IORING_ENTER_GETEVENTS) && ctx->syscall_iopoll)
3094 mutex_unlock(&ctx->uring_lock);
3096 if (flags & IORING_ENTER_GETEVENTS) {
3098 if (ctx->syscall_iopoll) {
3100 * We disallow the app entering submit/complete with
3101 * polling, but we still need to lock the ring to
3102 * prevent racing with polled issue that got punted to
3105 mutex_lock(&ctx->uring_lock);
3107 ret2 = io_validate_ext_arg(flags, argp, argsz);
3108 if (likely(!ret2)) {
3109 min_complete = min(min_complete,
3111 ret2 = io_iopoll_check(ctx, min_complete);
3113 mutex_unlock(&ctx->uring_lock);
3115 const sigset_t __user *sig;
3116 struct __kernel_timespec __user *ts;
3118 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3119 if (likely(!ret2)) {
3120 min_complete = min(min_complete,
3122 ret2 = io_cqring_wait(ctx, min_complete, sig,
3131 * EBADR indicates that one or more CQE were dropped.
3132 * Once the user has been informed we can clear the bit
3133 * as they are obviously ok with those drops.
3135 if (unlikely(ret2 == -EBADR))
3136 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3145 static const struct file_operations io_uring_fops = {
3146 .release = io_uring_release,
3147 .mmap = io_uring_mmap,
3149 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3150 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3152 .poll = io_uring_poll,
3153 #ifdef CONFIG_PROC_FS
3154 .show_fdinfo = io_uring_show_fdinfo,
3158 bool io_is_uring_fops(struct file *file)
3160 return file->f_op == &io_uring_fops;
3163 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3164 struct io_uring_params *p)
3166 struct io_rings *rings;
3167 size_t size, sq_array_offset;
3169 /* make sure these are sane, as we already accounted them */
3170 ctx->sq_entries = p->sq_entries;
3171 ctx->cq_entries = p->cq_entries;
3173 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3174 if (size == SIZE_MAX)
3177 rings = io_mem_alloc(size);
3182 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3183 rings->sq_ring_mask = p->sq_entries - 1;
3184 rings->cq_ring_mask = p->cq_entries - 1;
3185 rings->sq_ring_entries = p->sq_entries;
3186 rings->cq_ring_entries = p->cq_entries;
3188 if (p->flags & IORING_SETUP_SQE128)
3189 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3191 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3192 if (size == SIZE_MAX) {
3193 io_mem_free(ctx->rings);
3198 ctx->sq_sqes = io_mem_alloc(size);
3199 if (!ctx->sq_sqes) {
3200 io_mem_free(ctx->rings);
3208 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3212 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3216 ret = __io_uring_add_tctx_node(ctx, false);
3221 fd_install(fd, file);
3226 * Allocate an anonymous fd, this is what constitutes the application
3227 * visible backing of an io_uring instance. The application mmaps this
3228 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3229 * we have to tie this fd to a socket for file garbage collection purposes.
3231 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3234 #if defined(CONFIG_UNIX)
3237 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3240 return ERR_PTR(ret);
3243 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3244 O_RDWR | O_CLOEXEC, NULL);
3245 #if defined(CONFIG_UNIX)
3247 sock_release(ctx->ring_sock);
3248 ctx->ring_sock = NULL;
3250 ctx->ring_sock->file = file;
3256 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3257 struct io_uring_params __user *params)
3259 struct io_ring_ctx *ctx;
3265 if (entries > IORING_MAX_ENTRIES) {
3266 if (!(p->flags & IORING_SETUP_CLAMP))
3268 entries = IORING_MAX_ENTRIES;
3272 * Use twice as many entries for the CQ ring. It's possible for the
3273 * application to drive a higher depth than the size of the SQ ring,
3274 * since the sqes are only used at submission time. This allows for
3275 * some flexibility in overcommitting a bit. If the application has
3276 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3277 * of CQ ring entries manually.
3279 p->sq_entries = roundup_pow_of_two(entries);
3280 if (p->flags & IORING_SETUP_CQSIZE) {
3282 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3283 * to a power-of-two, if it isn't already. We do NOT impose
3284 * any cq vs sq ring sizing.
3288 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3289 if (!(p->flags & IORING_SETUP_CLAMP))
3291 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3293 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3294 if (p->cq_entries < p->sq_entries)
3297 p->cq_entries = 2 * p->sq_entries;
3300 ctx = io_ring_ctx_alloc(p);
3305 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3306 * space applications don't need to do io completion events
3307 * polling again, they can rely on io_sq_thread to do polling
3308 * work, which can reduce cpu usage and uring_lock contention.
3310 if (ctx->flags & IORING_SETUP_IOPOLL &&
3311 !(ctx->flags & IORING_SETUP_SQPOLL))
3312 ctx->syscall_iopoll = 1;
3314 ctx->compat = in_compat_syscall();
3315 if (!capable(CAP_IPC_LOCK))
3316 ctx->user = get_uid(current_user());
3319 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3320 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3323 if (ctx->flags & IORING_SETUP_SQPOLL) {
3324 /* IPI related flags don't make sense with SQPOLL */
3325 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3326 IORING_SETUP_TASKRUN_FLAG))
3328 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3329 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3330 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3332 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
3334 ctx->notify_method = TWA_SIGNAL;
3338 * This is just grabbed for accounting purposes. When a process exits,
3339 * the mm is exited and dropped before the files, hence we need to hang
3340 * on to this mm purely for the purposes of being able to unaccount
3341 * memory (locked/pinned vm). It's not used for anything else.
3343 mmgrab(current->mm);
3344 ctx->mm_account = current->mm;
3346 ret = io_allocate_scq_urings(ctx, p);
3350 ret = io_sq_offload_create(ctx, p);
3353 /* always set a rsrc node */
3354 ret = io_rsrc_node_switch_start(ctx);
3357 io_rsrc_node_switch(ctx, NULL);
3359 memset(&p->sq_off, 0, sizeof(p->sq_off));
3360 p->sq_off.head = offsetof(struct io_rings, sq.head);
3361 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3362 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3363 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3364 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3365 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3366 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3368 memset(&p->cq_off, 0, sizeof(p->cq_off));
3369 p->cq_off.head = offsetof(struct io_rings, cq.head);
3370 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3371 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3372 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3373 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3374 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3375 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3377 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3378 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3379 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3380 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3381 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3382 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3383 IORING_FEAT_LINKED_FILE;
3385 if (copy_to_user(params, p, sizeof(*p))) {
3390 file = io_uring_get_file(ctx);
3392 ret = PTR_ERR(file);
3397 * Install ring fd as the very last thing, so we don't risk someone
3398 * having closed it before we finish setup
3400 ret = io_uring_install_fd(ctx, file);
3402 /* fput will clean it up */
3407 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3410 io_ring_ctx_wait_and_kill(ctx);
3415 * Sets up an aio uring context, and returns the fd. Applications asks for a
3416 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3417 * params structure passed in.
3419 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3421 struct io_uring_params p;
3424 if (copy_from_user(&p, params, sizeof(p)))
3426 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3431 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3432 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3433 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3434 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3435 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3436 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3437 IORING_SETUP_SINGLE_ISSUER))
3440 return io_uring_create(entries, &p, params);
3443 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3444 struct io_uring_params __user *, params)
3446 return io_uring_setup(entries, params);
3449 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3452 struct io_uring_probe *p;
3456 size = struct_size(p, ops, nr_args);
3457 if (size == SIZE_MAX)
3459 p = kzalloc(size, GFP_KERNEL);
3464 if (copy_from_user(p, arg, size))
3467 if (memchr_inv(p, 0, size))
3470 p->last_op = IORING_OP_LAST - 1;
3471 if (nr_args > IORING_OP_LAST)
3472 nr_args = IORING_OP_LAST;
3474 for (i = 0; i < nr_args; i++) {
3476 if (!io_op_defs[i].not_supported)
3477 p->ops[i].flags = IO_URING_OP_SUPPORTED;
3482 if (copy_to_user(arg, p, size))
3489 static int io_register_personality(struct io_ring_ctx *ctx)
3491 const struct cred *creds;
3495 creds = get_current_cred();
3497 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
3498 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
3506 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
3507 void __user *arg, unsigned int nr_args)
3509 struct io_uring_restriction *res;
3513 /* Restrictions allowed only if rings started disabled */
3514 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3517 /* We allow only a single restrictions registration */
3518 if (ctx->restrictions.registered)
3521 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
3524 size = array_size(nr_args, sizeof(*res));
3525 if (size == SIZE_MAX)
3528 res = memdup_user(arg, size);
3530 return PTR_ERR(res);
3534 for (i = 0; i < nr_args; i++) {
3535 switch (res[i].opcode) {
3536 case IORING_RESTRICTION_REGISTER_OP:
3537 if (res[i].register_op >= IORING_REGISTER_LAST) {
3542 __set_bit(res[i].register_op,
3543 ctx->restrictions.register_op);
3545 case IORING_RESTRICTION_SQE_OP:
3546 if (res[i].sqe_op >= IORING_OP_LAST) {
3551 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
3553 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
3554 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
3556 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
3557 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
3566 /* Reset all restrictions if an error happened */
3568 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
3570 ctx->restrictions.registered = true;
3576 static int io_register_enable_rings(struct io_ring_ctx *ctx)
3578 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3581 if (ctx->restrictions.registered)
3582 ctx->restricted = 1;
3584 ctx->flags &= ~IORING_SETUP_R_DISABLED;
3585 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
3586 wake_up(&ctx->sq_data->wait);
3590 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
3591 void __user *arg, unsigned len)
3593 struct io_uring_task *tctx = current->io_uring;
3594 cpumask_var_t new_mask;
3597 if (!tctx || !tctx->io_wq)
3600 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3603 cpumask_clear(new_mask);
3604 if (len > cpumask_size())
3605 len = cpumask_size();
3607 if (in_compat_syscall()) {
3608 ret = compat_get_bitmap(cpumask_bits(new_mask),
3609 (const compat_ulong_t __user *)arg,
3610 len * 8 /* CHAR_BIT */);
3612 ret = copy_from_user(new_mask, arg, len);
3616 free_cpumask_var(new_mask);
3620 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
3621 free_cpumask_var(new_mask);
3625 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
3627 struct io_uring_task *tctx = current->io_uring;
3629 if (!tctx || !tctx->io_wq)
3632 return io_wq_cpu_affinity(tctx->io_wq, NULL);
3635 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
3637 __must_hold(&ctx->uring_lock)
3639 struct io_tctx_node *node;
3640 struct io_uring_task *tctx = NULL;
3641 struct io_sq_data *sqd = NULL;
3645 if (copy_from_user(new_count, arg, sizeof(new_count)))
3647 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3648 if (new_count[i] > INT_MAX)
3651 if (ctx->flags & IORING_SETUP_SQPOLL) {
3655 * Observe the correct sqd->lock -> ctx->uring_lock
3656 * ordering. Fine to drop uring_lock here, we hold
3659 refcount_inc(&sqd->refs);
3660 mutex_unlock(&ctx->uring_lock);
3661 mutex_lock(&sqd->lock);
3662 mutex_lock(&ctx->uring_lock);
3664 tctx = sqd->thread->io_uring;
3667 tctx = current->io_uring;
3670 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
3672 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3674 ctx->iowq_limits[i] = new_count[i];
3675 ctx->iowq_limits_set = true;
3677 if (tctx && tctx->io_wq) {
3678 ret = io_wq_max_workers(tctx->io_wq, new_count);
3682 memset(new_count, 0, sizeof(new_count));
3686 mutex_unlock(&sqd->lock);
3687 io_put_sq_data(sqd);
3690 if (copy_to_user(arg, new_count, sizeof(new_count)))
3693 /* that's it for SQPOLL, only the SQPOLL task creates requests */
3697 /* now propagate the restriction to all registered users */
3698 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3699 struct io_uring_task *tctx = node->task->io_uring;
3701 if (WARN_ON_ONCE(!tctx->io_wq))
3704 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3705 new_count[i] = ctx->iowq_limits[i];
3706 /* ignore errors, it always returns zero anyway */
3707 (void)io_wq_max_workers(tctx->io_wq, new_count);
3712 mutex_unlock(&sqd->lock);
3713 io_put_sq_data(sqd);
3718 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
3719 void __user *arg, unsigned nr_args)
3720 __releases(ctx->uring_lock)
3721 __acquires(ctx->uring_lock)
3726 * We don't quiesce the refs for register anymore and so it can't be
3727 * dying as we're holding a file ref here.
3729 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
3732 if (ctx->restricted) {
3733 if (opcode >= IORING_REGISTER_LAST)
3735 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
3736 if (!test_bit(opcode, ctx->restrictions.register_op))
3741 case IORING_REGISTER_BUFFERS:
3745 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
3747 case IORING_UNREGISTER_BUFFERS:
3751 ret = io_sqe_buffers_unregister(ctx);
3753 case IORING_REGISTER_FILES:
3757 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
3759 case IORING_UNREGISTER_FILES:
3763 ret = io_sqe_files_unregister(ctx);
3765 case IORING_REGISTER_FILES_UPDATE:
3766 ret = io_register_files_update(ctx, arg, nr_args);
3768 case IORING_REGISTER_EVENTFD:
3772 ret = io_eventfd_register(ctx, arg, 0);
3774 case IORING_REGISTER_EVENTFD_ASYNC:
3778 ret = io_eventfd_register(ctx, arg, 1);
3780 case IORING_UNREGISTER_EVENTFD:
3784 ret = io_eventfd_unregister(ctx);
3786 case IORING_REGISTER_PROBE:
3788 if (!arg || nr_args > 256)
3790 ret = io_probe(ctx, arg, nr_args);
3792 case IORING_REGISTER_PERSONALITY:
3796 ret = io_register_personality(ctx);
3798 case IORING_UNREGISTER_PERSONALITY:
3802 ret = io_unregister_personality(ctx, nr_args);
3804 case IORING_REGISTER_ENABLE_RINGS:
3808 ret = io_register_enable_rings(ctx);
3810 case IORING_REGISTER_RESTRICTIONS:
3811 ret = io_register_restrictions(ctx, arg, nr_args);
3813 case IORING_REGISTER_FILES2:
3814 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
3816 case IORING_REGISTER_FILES_UPDATE2:
3817 ret = io_register_rsrc_update(ctx, arg, nr_args,
3820 case IORING_REGISTER_BUFFERS2:
3821 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
3823 case IORING_REGISTER_BUFFERS_UPDATE:
3824 ret = io_register_rsrc_update(ctx, arg, nr_args,
3825 IORING_RSRC_BUFFER);
3827 case IORING_REGISTER_IOWQ_AFF:
3829 if (!arg || !nr_args)
3831 ret = io_register_iowq_aff(ctx, arg, nr_args);
3833 case IORING_UNREGISTER_IOWQ_AFF:
3837 ret = io_unregister_iowq_aff(ctx);
3839 case IORING_REGISTER_IOWQ_MAX_WORKERS:
3841 if (!arg || nr_args != 2)
3843 ret = io_register_iowq_max_workers(ctx, arg);
3845 case IORING_REGISTER_RING_FDS:
3846 ret = io_ringfd_register(ctx, arg, nr_args);
3848 case IORING_UNREGISTER_RING_FDS:
3849 ret = io_ringfd_unregister(ctx, arg, nr_args);
3851 case IORING_REGISTER_PBUF_RING:
3853 if (!arg || nr_args != 1)
3855 ret = io_register_pbuf_ring(ctx, arg);
3857 case IORING_UNREGISTER_PBUF_RING:
3859 if (!arg || nr_args != 1)
3861 ret = io_unregister_pbuf_ring(ctx, arg);
3863 case IORING_REGISTER_SYNC_CANCEL:
3865 if (!arg || nr_args != 1)
3867 ret = io_sync_cancel(ctx, arg);
3869 case IORING_REGISTER_FILE_ALLOC_RANGE:
3871 if (!arg || nr_args)
3873 ret = io_register_file_alloc_range(ctx, arg);
3883 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
3884 void __user *, arg, unsigned int, nr_args)
3886 struct io_ring_ctx *ctx;
3895 if (!io_is_uring_fops(f.file))
3898 ctx = f.file->private_data;
3902 mutex_lock(&ctx->uring_lock);
3903 ret = __io_uring_register(ctx, opcode, arg, nr_args);
3904 mutex_unlock(&ctx->uring_lock);
3905 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
3911 static int __init io_uring_init(void)
3913 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
3914 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
3915 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
3918 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
3919 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
3920 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
3921 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
3922 BUILD_BUG_SQE_ELEM(1, __u8, flags);
3923 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
3924 BUILD_BUG_SQE_ELEM(4, __s32, fd);
3925 BUILD_BUG_SQE_ELEM(8, __u64, off);
3926 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
3927 BUILD_BUG_SQE_ELEM(16, __u64, addr);
3928 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
3929 BUILD_BUG_SQE_ELEM(24, __u32, len);
3930 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
3931 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
3932 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
3933 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
3934 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
3935 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
3936 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
3937 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
3938 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
3939 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
3940 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
3941 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
3942 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
3943 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
3944 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
3945 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
3946 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
3947 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
3948 BUILD_BUG_SQE_ELEM(42, __u16, personality);
3949 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
3950 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
3951 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
3953 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
3954 sizeof(struct io_uring_rsrc_update));
3955 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
3956 sizeof(struct io_uring_rsrc_update2));
3958 /* ->buf_index is u16 */
3959 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
3960 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
3961 offsetof(struct io_uring_buf_ring, tail));
3963 /* should fit into one byte */
3964 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
3965 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
3966 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
3968 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
3970 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
3972 io_uring_optable_init();
3974 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
3978 __initcall(io_uring_init);