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>
97 #include "alloc_cache.h"
99 #define IORING_MAX_ENTRIES 32768
100 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
102 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
103 IORING_REGISTER_LAST + IORING_OP_LAST)
105 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
106 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
108 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
109 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
111 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
112 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
115 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
118 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
120 #define IO_COMPL_BATCH 32
121 #define IO_REQ_ALLOC_BATCH 8
124 IO_CHECK_CQ_OVERFLOW_BIT,
125 IO_CHECK_CQ_DROPPED_BIT,
128 struct io_defer_entry {
129 struct list_head list;
130 struct io_kiocb *req;
134 /* requests with any of those set should undergo io_disarm_next() */
135 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
136 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
138 static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
139 struct task_struct *task,
142 static void io_dismantle_req(struct io_kiocb *req);
143 static void io_clean_op(struct io_kiocb *req);
144 static void io_queue_sqe(struct io_kiocb *req);
146 static void __io_submit_flush_completions(struct io_ring_ctx *ctx);
148 static struct kmem_cache *req_cachep;
150 struct sock *io_uring_get_socket(struct file *file)
152 #if defined(CONFIG_UNIX)
153 if (io_is_uring_fops(file)) {
154 struct io_ring_ctx *ctx = file->private_data;
156 return ctx->ring_sock->sk;
161 EXPORT_SYMBOL(io_uring_get_socket);
163 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
165 if (!wq_list_empty(&ctx->submit_state.compl_reqs))
166 __io_submit_flush_completions(ctx);
169 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
171 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
174 static bool io_match_linked(struct io_kiocb *head)
176 struct io_kiocb *req;
178 io_for_each_link(req, head) {
179 if (req->flags & REQ_F_INFLIGHT)
186 * As io_match_task() but protected against racing with linked timeouts.
187 * User must not hold timeout_lock.
189 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
194 if (task && head->task != task)
199 if (head->flags & REQ_F_LINK_TIMEOUT) {
200 struct io_ring_ctx *ctx = head->ctx;
202 /* protect against races with linked timeouts */
203 spin_lock_irq(&ctx->timeout_lock);
204 matched = io_match_linked(head);
205 spin_unlock_irq(&ctx->timeout_lock);
207 matched = io_match_linked(head);
212 static inline void req_fail_link_node(struct io_kiocb *req, int res)
215 io_req_set_res(req, res, 0);
218 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
220 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
223 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
225 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
227 complete(&ctx->ref_comp);
230 static __cold void io_fallback_req_func(struct work_struct *work)
232 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
234 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
235 struct io_kiocb *req, *tmp;
238 percpu_ref_get(&ctx->refs);
239 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
240 req->io_task_work.func(req, &locked);
243 io_submit_flush_completions(ctx);
244 mutex_unlock(&ctx->uring_lock);
246 percpu_ref_put(&ctx->refs);
249 static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
251 unsigned hash_buckets = 1U << bits;
252 size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
254 table->hbs = kmalloc(hash_size, GFP_KERNEL);
258 table->hash_bits = bits;
259 init_hash_table(table, hash_buckets);
263 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
265 struct io_ring_ctx *ctx;
268 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
272 xa_init(&ctx->io_bl_xa);
275 * Use 5 bits less than the max cq entries, that should give us around
276 * 32 entries per hash list if totally full and uniformly spread, but
277 * don't keep too many buckets to not overconsume memory.
279 hash_bits = ilog2(p->cq_entries) - 5;
280 hash_bits = clamp(hash_bits, 1, 8);
281 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
283 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
286 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
287 if (!ctx->dummy_ubuf)
289 /* set invalid range, so io_import_fixed() fails meeting it */
290 ctx->dummy_ubuf->ubuf = -1UL;
292 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
296 ctx->flags = p->flags;
297 init_waitqueue_head(&ctx->sqo_sq_wait);
298 INIT_LIST_HEAD(&ctx->sqd_list);
299 INIT_LIST_HEAD(&ctx->cq_overflow_list);
300 INIT_LIST_HEAD(&ctx->io_buffers_cache);
301 io_alloc_cache_init(&ctx->apoll_cache);
302 io_alloc_cache_init(&ctx->netmsg_cache);
303 init_completion(&ctx->ref_comp);
304 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
305 mutex_init(&ctx->uring_lock);
306 init_waitqueue_head(&ctx->cq_wait);
307 spin_lock_init(&ctx->completion_lock);
308 spin_lock_init(&ctx->timeout_lock);
309 INIT_WQ_LIST(&ctx->iopoll_list);
310 INIT_LIST_HEAD(&ctx->io_buffers_pages);
311 INIT_LIST_HEAD(&ctx->io_buffers_comp);
312 INIT_LIST_HEAD(&ctx->defer_list);
313 INIT_LIST_HEAD(&ctx->timeout_list);
314 INIT_LIST_HEAD(&ctx->ltimeout_list);
315 spin_lock_init(&ctx->rsrc_ref_lock);
316 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
317 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
318 init_llist_head(&ctx->rsrc_put_llist);
319 INIT_LIST_HEAD(&ctx->tctx_list);
320 ctx->submit_state.free_list.next = NULL;
321 INIT_WQ_LIST(&ctx->locked_free_list);
322 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
323 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
326 kfree(ctx->dummy_ubuf);
327 kfree(ctx->cancel_table.hbs);
328 kfree(ctx->cancel_table_locked.hbs);
330 xa_destroy(&ctx->io_bl_xa);
335 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
337 struct io_rings *r = ctx->rings;
339 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
343 static bool req_need_defer(struct io_kiocb *req, u32 seq)
345 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
346 struct io_ring_ctx *ctx = req->ctx;
348 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
354 static inline void io_req_track_inflight(struct io_kiocb *req)
356 if (!(req->flags & REQ_F_INFLIGHT)) {
357 req->flags |= REQ_F_INFLIGHT;
358 atomic_inc(&req->task->io_uring->inflight_tracked);
362 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
364 if (WARN_ON_ONCE(!req->link))
367 req->flags &= ~REQ_F_ARM_LTIMEOUT;
368 req->flags |= REQ_F_LINK_TIMEOUT;
370 /* linked timeouts should have two refs once prep'ed */
371 io_req_set_refcount(req);
372 __io_req_set_refcount(req->link, 2);
376 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
378 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
380 return __io_prep_linked_timeout(req);
383 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
385 io_queue_linked_timeout(__io_prep_linked_timeout(req));
388 static inline void io_arm_ltimeout(struct io_kiocb *req)
390 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
391 __io_arm_ltimeout(req);
394 static void io_prep_async_work(struct io_kiocb *req)
396 const struct io_op_def *def = &io_op_defs[req->opcode];
397 struct io_ring_ctx *ctx = req->ctx;
399 if (!(req->flags & REQ_F_CREDS)) {
400 req->flags |= REQ_F_CREDS;
401 req->creds = get_current_cred();
404 req->work.list.next = NULL;
406 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
407 if (req->flags & REQ_F_FORCE_ASYNC)
408 req->work.flags |= IO_WQ_WORK_CONCURRENT;
410 if (req->file && !io_req_ffs_set(req))
411 req->flags |= io_file_get_flags(req->file) << REQ_F_SUPPORT_NOWAIT_BIT;
413 if (req->flags & REQ_F_ISREG) {
414 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
415 io_wq_hash_work(&req->work, file_inode(req->file));
416 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
417 if (def->unbound_nonreg_file)
418 req->work.flags |= IO_WQ_WORK_UNBOUND;
422 static void io_prep_async_link(struct io_kiocb *req)
424 struct io_kiocb *cur;
426 if (req->flags & REQ_F_LINK_TIMEOUT) {
427 struct io_ring_ctx *ctx = req->ctx;
429 spin_lock_irq(&ctx->timeout_lock);
430 io_for_each_link(cur, req)
431 io_prep_async_work(cur);
432 spin_unlock_irq(&ctx->timeout_lock);
434 io_for_each_link(cur, req)
435 io_prep_async_work(cur);
439 void io_queue_iowq(struct io_kiocb *req, bool *dont_use)
441 struct io_kiocb *link = io_prep_linked_timeout(req);
442 struct io_uring_task *tctx = req->task->io_uring;
445 BUG_ON(!tctx->io_wq);
447 /* init ->work of the whole link before punting */
448 io_prep_async_link(req);
451 * Not expected to happen, but if we do have a bug where this _can_
452 * happen, catch it here and ensure the request is marked as
453 * canceled. That will make io-wq go through the usual work cancel
454 * procedure rather than attempt to run this request (or create a new
457 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
458 req->work.flags |= IO_WQ_WORK_CANCEL;
460 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
461 io_wq_enqueue(tctx->io_wq, &req->work);
463 io_queue_linked_timeout(link);
466 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
468 while (!list_empty(&ctx->defer_list)) {
469 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
470 struct io_defer_entry, list);
472 if (req_need_defer(de->req, de->seq))
474 list_del_init(&de->list);
475 io_req_task_queue(de->req);
480 static void io_eventfd_signal(struct io_ring_ctx *ctx)
482 struct io_ev_fd *ev_fd;
485 spin_lock(&ctx->completion_lock);
487 * Eventfd should only get triggered when at least one event has been
488 * posted. Some applications rely on the eventfd notification count only
489 * changing IFF a new CQE has been added to the CQ ring. There's no
490 * depedency on 1:1 relationship between how many times this function is
491 * called (and hence the eventfd count) and number of CQEs posted to the
494 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
495 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
496 spin_unlock(&ctx->completion_lock);
502 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
505 ev_fd = rcu_dereference(ctx->io_ev_fd);
508 * Check again if ev_fd exists incase an io_eventfd_unregister call
509 * completed between the NULL check of ctx->io_ev_fd at the start of
510 * the function and rcu_read_lock.
512 if (unlikely(!ev_fd))
514 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
517 if (!ev_fd->eventfd_async || io_wq_current_is_worker())
518 eventfd_signal(ev_fd->cq_ev_fd, 1);
523 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
525 if (ctx->off_timeout_used || ctx->drain_active) {
526 spin_lock(&ctx->completion_lock);
527 if (ctx->off_timeout_used)
528 io_flush_timeouts(ctx);
529 if (ctx->drain_active)
530 io_queue_deferred(ctx);
531 spin_unlock(&ctx->completion_lock);
534 io_eventfd_signal(ctx);
537 static inline void io_cqring_ev_posted(struct io_ring_ctx *ctx)
539 io_commit_cqring_flush(ctx);
543 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
544 __releases(ctx->completion_lock)
546 io_commit_cqring(ctx);
547 spin_unlock(&ctx->completion_lock);
548 io_cqring_ev_posted(ctx);
551 void io_cq_unlock_post(struct io_ring_ctx *ctx)
553 __io_cq_unlock_post(ctx);
556 /* Returns true if there are no backlogged entries after the flush */
557 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
560 size_t cqe_size = sizeof(struct io_uring_cqe);
562 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
565 if (ctx->flags & IORING_SETUP_CQE32)
569 while (!list_empty(&ctx->cq_overflow_list)) {
570 struct io_uring_cqe *cqe = io_get_cqe(ctx);
571 struct io_overflow_cqe *ocqe;
575 ocqe = list_first_entry(&ctx->cq_overflow_list,
576 struct io_overflow_cqe, list);
578 memcpy(cqe, &ocqe->cqe, cqe_size);
580 io_account_cq_overflow(ctx);
582 list_del(&ocqe->list);
586 all_flushed = list_empty(&ctx->cq_overflow_list);
588 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
589 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
592 io_cq_unlock_post(ctx);
596 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
600 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
601 /* iopoll syncs against uring_lock, not completion_lock */
602 if (ctx->flags & IORING_SETUP_IOPOLL)
603 mutex_lock(&ctx->uring_lock);
604 ret = __io_cqring_overflow_flush(ctx, false);
605 if (ctx->flags & IORING_SETUP_IOPOLL)
606 mutex_unlock(&ctx->uring_lock);
612 void __io_put_task(struct task_struct *task, int nr)
614 struct io_uring_task *tctx = task->io_uring;
616 percpu_counter_sub(&tctx->inflight, nr);
617 if (unlikely(atomic_read(&tctx->in_idle)))
618 wake_up(&tctx->wait);
619 put_task_struct_many(task, nr);
622 void io_task_refs_refill(struct io_uring_task *tctx)
624 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
626 percpu_counter_add(&tctx->inflight, refill);
627 refcount_add(refill, ¤t->usage);
628 tctx->cached_refs += refill;
631 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
633 struct io_uring_task *tctx = task->io_uring;
634 unsigned int refs = tctx->cached_refs;
637 tctx->cached_refs = 0;
638 percpu_counter_sub(&tctx->inflight, refs);
639 put_task_struct_many(task, refs);
643 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
644 s32 res, u32 cflags, u64 extra1, u64 extra2)
646 struct io_overflow_cqe *ocqe;
647 size_t ocq_size = sizeof(struct io_overflow_cqe);
648 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
651 ocq_size += sizeof(struct io_uring_cqe);
653 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
654 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
657 * If we're in ring overflow flush mode, or in task cancel mode,
658 * or cannot allocate an overflow entry, then we need to drop it
661 io_account_cq_overflow(ctx);
662 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
665 if (list_empty(&ctx->cq_overflow_list)) {
666 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
667 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
670 ocqe->cqe.user_data = user_data;
672 ocqe->cqe.flags = cflags;
674 ocqe->cqe.big_cqe[0] = extra1;
675 ocqe->cqe.big_cqe[1] = extra2;
677 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
681 bool io_req_cqe_overflow(struct io_kiocb *req)
683 if (!(req->flags & REQ_F_CQE32_INIT)) {
687 return io_cqring_event_overflow(req->ctx, req->cqe.user_data,
688 req->cqe.res, req->cqe.flags,
689 req->extra1, req->extra2);
693 * writes to the cq entry need to come after reading head; the
694 * control dependency is enough as we're using WRITE_ONCE to
697 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx)
699 struct io_rings *rings = ctx->rings;
700 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
701 unsigned int free, queued, len;
704 /* userspace may cheat modifying the tail, be safe and do min */
705 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
706 free = ctx->cq_entries - queued;
707 /* we need a contiguous range, limit based on the current array offset */
708 len = min(free, ctx->cq_entries - off);
712 if (ctx->flags & IORING_SETUP_CQE32) {
717 ctx->cqe_cached = &rings->cqes[off];
718 ctx->cqe_sentinel = ctx->cqe_cached + len;
720 ctx->cached_cq_tail++;
722 if (ctx->flags & IORING_SETUP_CQE32)
724 return &rings->cqes[off];
727 bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
730 struct io_uring_cqe *cqe;
735 * If we can't get a cq entry, userspace overflowed the
736 * submission (by quite a lot). Increment the overflow count in
739 cqe = io_get_cqe(ctx);
741 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
743 WRITE_ONCE(cqe->user_data, user_data);
744 WRITE_ONCE(cqe->res, res);
745 WRITE_ONCE(cqe->flags, cflags);
747 if (ctx->flags & IORING_SETUP_CQE32) {
748 WRITE_ONCE(cqe->big_cqe[0], 0);
749 WRITE_ONCE(cqe->big_cqe[1], 0);
755 return io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
760 bool io_post_aux_cqe(struct io_ring_ctx *ctx,
761 u64 user_data, s32 res, u32 cflags,
767 filled = io_fill_cqe_aux(ctx, user_data, res, cflags, allow_overflow);
768 io_cq_unlock_post(ctx);
772 static void __io_req_complete_put(struct io_kiocb *req)
775 * If we're the last reference to this request, add to our locked
778 if (req_ref_put_and_test(req)) {
779 struct io_ring_ctx *ctx = req->ctx;
781 if (req->flags & IO_REQ_LINK_FLAGS) {
782 if (req->flags & IO_DISARM_MASK)
785 io_req_task_queue(req->link);
789 io_req_put_rsrc(req);
791 * Selected buffer deallocation in io_clean_op() assumes that
792 * we don't hold ->completion_lock. Clean them here to avoid
795 io_put_kbuf_comp(req);
796 io_dismantle_req(req);
797 io_put_task(req->task, 1);
798 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
799 ctx->locked_free_nr++;
803 void __io_req_complete_post(struct io_kiocb *req)
805 if (!(req->flags & REQ_F_CQE_SKIP))
806 __io_fill_cqe_req(req->ctx, req);
807 __io_req_complete_put(req);
810 void io_req_complete_post(struct io_kiocb *req)
812 struct io_ring_ctx *ctx = req->ctx;
815 __io_req_complete_post(req);
816 io_cq_unlock_post(ctx);
819 inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags)
821 io_req_complete_post(req);
824 void io_req_complete_failed(struct io_kiocb *req, s32 res)
827 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
828 io_req_complete_post(req);
832 * Don't initialise the fields below on every allocation, but do that in
833 * advance and keep them valid across allocations.
835 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
839 req->async_data = NULL;
840 /* not necessary, but safer to zero */
844 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
845 struct io_submit_state *state)
847 spin_lock(&ctx->completion_lock);
848 wq_list_splice(&ctx->locked_free_list, &state->free_list);
849 ctx->locked_free_nr = 0;
850 spin_unlock(&ctx->completion_lock);
854 * A request might get retired back into the request caches even before opcode
855 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
856 * Because of that, io_alloc_req() should be called only under ->uring_lock
857 * and with extra caution to not get a request that is still worked on.
859 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
860 __must_hold(&ctx->uring_lock)
862 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
863 void *reqs[IO_REQ_ALLOC_BATCH];
867 * If we have more than a batch's worth of requests in our IRQ side
868 * locked cache, grab the lock and move them over to our submission
871 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
872 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
873 if (!io_req_cache_empty(ctx))
877 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
880 * Bulk alloc is all-or-nothing. If we fail to get a batch,
881 * retry single alloc to be on the safe side.
883 if (unlikely(ret <= 0)) {
884 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
890 percpu_ref_get_many(&ctx->refs, ret);
891 for (i = 0; i < ret; i++) {
892 struct io_kiocb *req = reqs[i];
894 io_preinit_req(req, ctx);
895 io_req_add_to_cache(req, ctx);
900 static inline void io_dismantle_req(struct io_kiocb *req)
902 unsigned int flags = req->flags;
904 if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
906 if (!(flags & REQ_F_FIXED_FILE))
907 io_put_file(req->file);
910 __cold void io_free_req(struct io_kiocb *req)
912 struct io_ring_ctx *ctx = req->ctx;
914 io_req_put_rsrc(req);
915 io_dismantle_req(req);
916 io_put_task(req->task, 1);
918 spin_lock(&ctx->completion_lock);
919 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
920 ctx->locked_free_nr++;
921 spin_unlock(&ctx->completion_lock);
924 static void __io_req_find_next_prep(struct io_kiocb *req)
926 struct io_ring_ctx *ctx = req->ctx;
930 io_cq_unlock_post(ctx);
933 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
935 struct io_kiocb *nxt;
938 * If LINK is set, we have dependent requests in this chain. If we
939 * didn't fail this request, queue the first one up, moving any other
940 * dependencies to the next request. In case of failure, fail the rest
943 if (unlikely(req->flags & IO_DISARM_MASK))
944 __io_req_find_next_prep(req);
950 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
954 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
955 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
957 io_submit_flush_completions(ctx);
958 mutex_unlock(&ctx->uring_lock);
961 percpu_ref_put(&ctx->refs);
964 static unsigned int handle_tw_list(struct llist_node *node,
965 struct io_ring_ctx **ctx, bool *locked,
966 struct llist_node *last)
968 unsigned int count = 0;
970 while (node != last) {
971 struct llist_node *next = node->next;
972 struct io_kiocb *req = container_of(node, struct io_kiocb,
975 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
977 if (req->ctx != *ctx) {
978 ctx_flush_and_put(*ctx, locked);
980 /* if not contended, grab and improve batching */
981 *locked = mutex_trylock(&(*ctx)->uring_lock);
982 percpu_ref_get(&(*ctx)->refs);
984 req->io_task_work.func(req, locked);
993 * io_llist_xchg - swap all entries in a lock-less list
994 * @head: the head of lock-less list to delete all entries
995 * @new: new entry as the head of the list
997 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
998 * The order of entries returned is from the newest to the oldest added one.
1000 static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1001 struct llist_node *new)
1003 return xchg(&head->first, new);
1007 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1008 * @head: the head of lock-less list to delete all entries
1009 * @old: expected old value of the first entry of the list
1010 * @new: new entry as the head of the list
1012 * perform a cmpxchg on the first entry of the list.
1015 static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1016 struct llist_node *old,
1017 struct llist_node *new)
1019 return cmpxchg(&head->first, old, new);
1022 void tctx_task_work(struct callback_head *cb)
1024 bool uring_locked = false;
1025 struct io_ring_ctx *ctx = NULL;
1026 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1028 struct llist_node fake = {};
1029 struct llist_node *node = io_llist_xchg(&tctx->task_list, &fake);
1030 unsigned int loops = 1;
1031 unsigned int count = handle_tw_list(node, &ctx, &uring_locked, NULL);
1033 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1034 while (node != &fake) {
1036 node = io_llist_xchg(&tctx->task_list, &fake);
1037 count += handle_tw_list(node, &ctx, &uring_locked, &fake);
1038 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1041 ctx_flush_and_put(ctx, &uring_locked);
1043 /* relaxed read is enough as only the task itself sets ->in_idle */
1044 if (unlikely(atomic_read(&tctx->in_idle)))
1045 io_uring_drop_tctx_refs(current);
1047 trace_io_uring_task_work_run(tctx, count, loops);
1050 void io_req_task_work_add(struct io_kiocb *req)
1052 struct io_uring_task *tctx = req->task->io_uring;
1053 struct io_ring_ctx *ctx = req->ctx;
1054 struct llist_node *node;
1057 running = !llist_add(&req->io_task_work.node, &tctx->task_list);
1059 /* task_work already pending, we're done */
1063 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1064 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1066 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1069 node = llist_del_all(&tctx->task_list);
1072 req = container_of(node, struct io_kiocb, io_task_work.node);
1074 if (llist_add(&req->io_task_work.node,
1075 &req->ctx->fallback_llist))
1076 schedule_delayed_work(&req->ctx->fallback_work, 1);
1080 static void io_req_tw_post(struct io_kiocb *req, bool *locked)
1082 io_req_complete_post(req);
1085 void io_req_tw_post_queue(struct io_kiocb *req, s32 res, u32 cflags)
1087 io_req_set_res(req, res, cflags);
1088 req->io_task_work.func = io_req_tw_post;
1089 io_req_task_work_add(req);
1092 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
1094 /* not needed for normal modes, but SQPOLL depends on it */
1095 io_tw_lock(req->ctx, locked);
1096 io_req_complete_failed(req, req->cqe.res);
1099 void io_req_task_submit(struct io_kiocb *req, bool *locked)
1101 io_tw_lock(req->ctx, locked);
1102 /* req->task == current here, checking PF_EXITING is safe */
1103 if (likely(!(req->task->flags & PF_EXITING)))
1106 io_req_complete_failed(req, -EFAULT);
1109 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1111 io_req_set_res(req, ret, 0);
1112 req->io_task_work.func = io_req_task_cancel;
1113 io_req_task_work_add(req);
1116 void io_req_task_queue(struct io_kiocb *req)
1118 req->io_task_work.func = io_req_task_submit;
1119 io_req_task_work_add(req);
1122 void io_queue_next(struct io_kiocb *req)
1124 struct io_kiocb *nxt = io_req_find_next(req);
1127 io_req_task_queue(nxt);
1130 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1131 __must_hold(&ctx->uring_lock)
1133 struct task_struct *task = NULL;
1137 struct io_kiocb *req = container_of(node, struct io_kiocb,
1140 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1141 if (req->flags & REQ_F_REFCOUNT) {
1142 node = req->comp_list.next;
1143 if (!req_ref_put_and_test(req))
1146 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1147 struct async_poll *apoll = req->apoll;
1149 if (apoll->double_poll)
1150 kfree(apoll->double_poll);
1151 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1153 req->flags &= ~REQ_F_POLLED;
1155 if (req->flags & IO_REQ_LINK_FLAGS)
1157 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1160 if (!(req->flags & REQ_F_FIXED_FILE))
1161 io_put_file(req->file);
1163 io_req_put_rsrc_locked(req, ctx);
1165 if (req->task != task) {
1167 io_put_task(task, task_refs);
1172 node = req->comp_list.next;
1173 io_req_add_to_cache(req, ctx);
1177 io_put_task(task, task_refs);
1180 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1181 __must_hold(&ctx->uring_lock)
1183 struct io_wq_work_node *node, *prev;
1184 struct io_submit_state *state = &ctx->submit_state;
1186 spin_lock(&ctx->completion_lock);
1187 wq_list_for_each(node, prev, &state->compl_reqs) {
1188 struct io_kiocb *req = container_of(node, struct io_kiocb,
1191 if (!(req->flags & REQ_F_CQE_SKIP))
1192 __io_fill_cqe_req(ctx, req);
1194 __io_cq_unlock_post(ctx);
1196 io_free_batch_list(ctx, state->compl_reqs.first);
1197 INIT_WQ_LIST(&state->compl_reqs);
1201 * Drop reference to request, return next in chain (if there is one) if this
1202 * was the last reference to this request.
1204 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1206 struct io_kiocb *nxt = NULL;
1208 if (req_ref_put_and_test(req)) {
1209 if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
1210 nxt = io_req_find_next(req);
1216 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1218 /* See comment at the top of this file */
1220 return __io_cqring_events(ctx);
1224 * We can't just wait for polled events to come to us, we have to actively
1225 * find and complete them.
1227 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1229 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1232 mutex_lock(&ctx->uring_lock);
1233 while (!wq_list_empty(&ctx->iopoll_list)) {
1234 /* let it sleep and repeat later if can't complete a request */
1235 if (io_do_iopoll(ctx, true) == 0)
1238 * Ensure we allow local-to-the-cpu processing to take place,
1239 * in this case we need to ensure that we reap all events.
1240 * Also let task_work, etc. to progress by releasing the mutex
1242 if (need_resched()) {
1243 mutex_unlock(&ctx->uring_lock);
1245 mutex_lock(&ctx->uring_lock);
1248 mutex_unlock(&ctx->uring_lock);
1251 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1253 unsigned int nr_events = 0;
1255 unsigned long check_cq;
1257 check_cq = READ_ONCE(ctx->check_cq);
1258 if (unlikely(check_cq)) {
1259 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1260 __io_cqring_overflow_flush(ctx, false);
1262 * Similarly do not spin if we have not informed the user of any
1265 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1269 * Don't enter poll loop if we already have events pending.
1270 * If we do, we can potentially be spinning for commands that
1271 * already triggered a CQE (eg in error).
1273 if (io_cqring_events(ctx))
1278 * If a submit got punted to a workqueue, we can have the
1279 * application entering polling for a command before it gets
1280 * issued. That app will hold the uring_lock for the duration
1281 * of the poll right here, so we need to take a breather every
1282 * now and then to ensure that the issue has a chance to add
1283 * the poll to the issued list. Otherwise we can spin here
1284 * forever, while the workqueue is stuck trying to acquire the
1287 if (wq_list_empty(&ctx->iopoll_list)) {
1288 u32 tail = ctx->cached_cq_tail;
1290 mutex_unlock(&ctx->uring_lock);
1292 mutex_lock(&ctx->uring_lock);
1294 /* some requests don't go through iopoll_list */
1295 if (tail != ctx->cached_cq_tail ||
1296 wq_list_empty(&ctx->iopoll_list))
1299 ret = io_do_iopoll(ctx, !min);
1304 } while (nr_events < min && !need_resched());
1309 void io_req_task_complete(struct io_kiocb *req, bool *locked)
1311 if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) {
1312 unsigned issue_flags = *locked ? 0 : IO_URING_F_UNLOCKED;
1314 req->cqe.flags |= io_put_kbuf(req, issue_flags);
1318 io_req_complete_defer(req);
1320 io_req_complete_post(req);
1324 * After the iocb has been issued, it's safe to be found on the poll list.
1325 * Adding the kiocb to the list AFTER submission ensures that we don't
1326 * find it from a io_do_iopoll() thread before the issuer is done
1327 * accessing the kiocb cookie.
1329 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1331 struct io_ring_ctx *ctx = req->ctx;
1332 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1334 /* workqueue context doesn't hold uring_lock, grab it now */
1335 if (unlikely(needs_lock))
1336 mutex_lock(&ctx->uring_lock);
1339 * Track whether we have multiple files in our lists. This will impact
1340 * how we do polling eventually, not spinning if we're on potentially
1341 * different devices.
1343 if (wq_list_empty(&ctx->iopoll_list)) {
1344 ctx->poll_multi_queue = false;
1345 } else if (!ctx->poll_multi_queue) {
1346 struct io_kiocb *list_req;
1348 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1350 if (list_req->file != req->file)
1351 ctx->poll_multi_queue = true;
1355 * For fast devices, IO may have already completed. If it has, add
1356 * it to the front so we find it first.
1358 if (READ_ONCE(req->iopoll_completed))
1359 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1361 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1363 if (unlikely(needs_lock)) {
1365 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1366 * in sq thread task context or in io worker task context. If
1367 * current task context is sq thread, we don't need to check
1368 * whether should wake up sq thread.
1370 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1371 wq_has_sleeper(&ctx->sq_data->wait))
1372 wake_up(&ctx->sq_data->wait);
1374 mutex_unlock(&ctx->uring_lock);
1378 static bool io_bdev_nowait(struct block_device *bdev)
1380 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
1384 * If we tracked the file through the SCM inflight mechanism, we could support
1385 * any file. For now, just ensure that anything potentially problematic is done
1388 static bool __io_file_supports_nowait(struct file *file, umode_t mode)
1390 if (S_ISBLK(mode)) {
1391 if (IS_ENABLED(CONFIG_BLOCK) &&
1392 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
1398 if (S_ISREG(mode)) {
1399 if (IS_ENABLED(CONFIG_BLOCK) &&
1400 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
1401 !io_is_uring_fops(file))
1406 /* any ->read/write should understand O_NONBLOCK */
1407 if (file->f_flags & O_NONBLOCK)
1409 return file->f_mode & FMODE_NOWAIT;
1413 * If we tracked the file through the SCM inflight mechanism, we could support
1414 * any file. For now, just ensure that anything potentially problematic is done
1417 unsigned int io_file_get_flags(struct file *file)
1419 umode_t mode = file_inode(file)->i_mode;
1420 unsigned int res = 0;
1424 if (__io_file_supports_nowait(file, mode))
1426 if (io_file_need_scm(file))
1431 bool io_alloc_async_data(struct io_kiocb *req)
1433 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
1434 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
1435 if (req->async_data) {
1436 req->flags |= REQ_F_ASYNC_DATA;
1442 int io_req_prep_async(struct io_kiocb *req)
1444 const struct io_op_def *def = &io_op_defs[req->opcode];
1446 /* assign early for deferred execution for non-fixed file */
1447 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE))
1448 req->file = io_file_get_normal(req, req->cqe.fd);
1449 if (!def->prep_async)
1451 if (WARN_ON_ONCE(req_has_async_data(req)))
1453 if (!io_op_defs[req->opcode].manual_alloc) {
1454 if (io_alloc_async_data(req))
1457 return def->prep_async(req);
1460 static u32 io_get_sequence(struct io_kiocb *req)
1462 u32 seq = req->ctx->cached_sq_head;
1463 struct io_kiocb *cur;
1465 /* need original cached_sq_head, but it was increased for each req */
1466 io_for_each_link(cur, req)
1471 static __cold void io_drain_req(struct io_kiocb *req)
1473 struct io_ring_ctx *ctx = req->ctx;
1474 struct io_defer_entry *de;
1476 u32 seq = io_get_sequence(req);
1478 /* Still need defer if there is pending req in defer list. */
1479 spin_lock(&ctx->completion_lock);
1480 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1481 spin_unlock(&ctx->completion_lock);
1483 ctx->drain_active = false;
1484 io_req_task_queue(req);
1487 spin_unlock(&ctx->completion_lock);
1489 ret = io_req_prep_async(req);
1492 io_req_complete_failed(req, ret);
1495 io_prep_async_link(req);
1496 de = kmalloc(sizeof(*de), GFP_KERNEL);
1502 spin_lock(&ctx->completion_lock);
1503 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1504 spin_unlock(&ctx->completion_lock);
1509 trace_io_uring_defer(req);
1512 list_add_tail(&de->list, &ctx->defer_list);
1513 spin_unlock(&ctx->completion_lock);
1516 static void io_clean_op(struct io_kiocb *req)
1518 if (req->flags & REQ_F_BUFFER_SELECTED) {
1519 spin_lock(&req->ctx->completion_lock);
1520 io_put_kbuf_comp(req);
1521 spin_unlock(&req->ctx->completion_lock);
1524 if (req->flags & REQ_F_NEED_CLEANUP) {
1525 const struct io_op_def *def = &io_op_defs[req->opcode];
1530 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1531 kfree(req->apoll->double_poll);
1535 if (req->flags & REQ_F_INFLIGHT) {
1536 struct io_uring_task *tctx = req->task->io_uring;
1538 atomic_dec(&tctx->inflight_tracked);
1540 if (req->flags & REQ_F_CREDS)
1541 put_cred(req->creds);
1542 if (req->flags & REQ_F_ASYNC_DATA) {
1543 kfree(req->async_data);
1544 req->async_data = NULL;
1546 req->flags &= ~IO_REQ_CLEAN_FLAGS;
1549 static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags)
1551 if (req->file || !io_op_defs[req->opcode].needs_file)
1554 if (req->flags & REQ_F_FIXED_FILE)
1555 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1557 req->file = io_file_get_normal(req, req->cqe.fd);
1562 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1564 const struct io_op_def *def = &io_op_defs[req->opcode];
1565 const struct cred *creds = NULL;
1568 if (unlikely(!io_assign_file(req, issue_flags)))
1571 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1572 creds = override_creds(req->creds);
1574 if (!def->audit_skip)
1575 audit_uring_entry(req->opcode);
1577 ret = def->issue(req, issue_flags);
1579 if (!def->audit_skip)
1580 audit_uring_exit(!ret, ret);
1583 revert_creds(creds);
1585 if (ret == IOU_OK) {
1586 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1587 io_req_complete_defer(req);
1589 io_req_complete_post(req);
1590 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1593 /* If the op doesn't have a file, we're not polling for it */
1594 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && req->file)
1595 io_iopoll_req_issued(req, issue_flags);
1600 int io_poll_issue(struct io_kiocb *req, bool *locked)
1602 io_tw_lock(req->ctx, locked);
1603 if (unlikely(req->task->flags & PF_EXITING))
1605 return io_issue_sqe(req, IO_URING_F_NONBLOCK);
1608 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1610 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1612 req = io_put_req_find_next(req);
1613 return req ? &req->work : NULL;
1616 void io_wq_submit_work(struct io_wq_work *work)
1618 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1619 const struct io_op_def *def = &io_op_defs[req->opcode];
1620 unsigned int issue_flags = IO_URING_F_UNLOCKED;
1621 bool needs_poll = false;
1622 int ret = 0, err = -ECANCELED;
1624 /* one will be dropped by ->io_free_work() after returning to io-wq */
1625 if (!(req->flags & REQ_F_REFCOUNT))
1626 __io_req_set_refcount(req, 2);
1630 io_arm_ltimeout(req);
1632 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1633 if (work->flags & IO_WQ_WORK_CANCEL) {
1635 io_req_task_queue_fail(req, err);
1638 if (!io_assign_file(req, issue_flags)) {
1640 work->flags |= IO_WQ_WORK_CANCEL;
1644 if (req->flags & REQ_F_FORCE_ASYNC) {
1645 bool opcode_poll = def->pollin || def->pollout;
1647 if (opcode_poll && file_can_poll(req->file)) {
1649 issue_flags |= IO_URING_F_NONBLOCK;
1654 ret = io_issue_sqe(req, issue_flags);
1658 * We can get EAGAIN for iopolled IO even though we're
1659 * forcing a sync submission from here, since we can't
1660 * wait for request slots on the block side.
1663 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1669 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1671 /* aborted or ready, in either case retry blocking */
1673 issue_flags &= ~IO_URING_F_NONBLOCK;
1676 /* avoid locking problems by failing it from a clean context */
1678 io_req_task_queue_fail(req, ret);
1681 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1682 unsigned int issue_flags)
1684 struct io_ring_ctx *ctx = req->ctx;
1685 struct file *file = NULL;
1686 unsigned long file_ptr;
1688 io_ring_submit_lock(ctx, issue_flags);
1690 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
1692 fd = array_index_nospec(fd, ctx->nr_user_files);
1693 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
1694 file = (struct file *) (file_ptr & FFS_MASK);
1695 file_ptr &= ~FFS_MASK;
1696 /* mask in overlapping REQ_F and FFS bits */
1697 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
1698 io_req_set_rsrc_node(req, ctx, 0);
1699 WARN_ON_ONCE(file && !test_bit(fd, ctx->file_table.bitmap));
1701 io_ring_submit_unlock(ctx, issue_flags);
1705 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
1707 struct file *file = fget(fd);
1709 trace_io_uring_file_get(req, fd);
1711 /* we don't allow fixed io_uring files */
1712 if (file && io_is_uring_fops(file))
1713 io_req_track_inflight(req);
1717 static void io_queue_async(struct io_kiocb *req, int ret)
1718 __must_hold(&req->ctx->uring_lock)
1720 struct io_kiocb *linked_timeout;
1722 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
1723 io_req_complete_failed(req, ret);
1727 linked_timeout = io_prep_linked_timeout(req);
1729 switch (io_arm_poll_handler(req, 0)) {
1730 case IO_APOLL_READY:
1731 io_kbuf_recycle(req, 0);
1732 io_req_task_queue(req);
1734 case IO_APOLL_ABORTED:
1736 * Queued up for async execution, worker will release
1737 * submit reference when the iocb is actually submitted.
1739 io_kbuf_recycle(req, 0);
1740 io_queue_iowq(req, NULL);
1747 io_queue_linked_timeout(linked_timeout);
1750 static inline void io_queue_sqe(struct io_kiocb *req)
1751 __must_hold(&req->ctx->uring_lock)
1755 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
1758 * We async punt it if the file wasn't marked NOWAIT, or if the file
1759 * doesn't support non-blocking read/write attempts
1762 io_arm_ltimeout(req);
1764 io_queue_async(req, ret);
1767 static void io_queue_sqe_fallback(struct io_kiocb *req)
1768 __must_hold(&req->ctx->uring_lock)
1770 if (unlikely(req->flags & REQ_F_FAIL)) {
1772 * We don't submit, fail them all, for that replace hardlinks
1773 * with normal links. Extra REQ_F_LINK is tolerated.
1775 req->flags &= ~REQ_F_HARDLINK;
1776 req->flags |= REQ_F_LINK;
1777 io_req_complete_failed(req, req->cqe.res);
1778 } else if (unlikely(req->ctx->drain_active)) {
1781 int ret = io_req_prep_async(req);
1784 io_req_complete_failed(req, ret);
1786 io_queue_iowq(req, NULL);
1791 * Check SQE restrictions (opcode and flags).
1793 * Returns 'true' if SQE is allowed, 'false' otherwise.
1795 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
1796 struct io_kiocb *req,
1797 unsigned int sqe_flags)
1799 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
1802 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
1803 ctx->restrictions.sqe_flags_required)
1806 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
1807 ctx->restrictions.sqe_flags_required))
1813 static void io_init_req_drain(struct io_kiocb *req)
1815 struct io_ring_ctx *ctx = req->ctx;
1816 struct io_kiocb *head = ctx->submit_state.link.head;
1818 ctx->drain_active = true;
1821 * If we need to drain a request in the middle of a link, drain
1822 * the head request and the next request/link after the current
1823 * link. Considering sequential execution of links,
1824 * REQ_F_IO_DRAIN will be maintained for every request of our
1827 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
1828 ctx->drain_next = true;
1832 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
1833 const struct io_uring_sqe *sqe)
1834 __must_hold(&ctx->uring_lock)
1836 const struct io_op_def *def;
1837 unsigned int sqe_flags;
1841 /* req is partially pre-initialised, see io_preinit_req() */
1842 req->opcode = opcode = READ_ONCE(sqe->opcode);
1843 /* same numerical values with corresponding REQ_F_*, safe to copy */
1844 req->flags = sqe_flags = READ_ONCE(sqe->flags);
1845 req->cqe.user_data = READ_ONCE(sqe->user_data);
1847 req->rsrc_node = NULL;
1848 req->task = current;
1850 if (unlikely(opcode >= IORING_OP_LAST)) {
1854 def = &io_op_defs[opcode];
1855 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
1856 /* enforce forwards compatibility on users */
1857 if (sqe_flags & ~SQE_VALID_FLAGS)
1859 if (sqe_flags & IOSQE_BUFFER_SELECT) {
1860 if (!def->buffer_select)
1862 req->buf_index = READ_ONCE(sqe->buf_group);
1864 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
1865 ctx->drain_disabled = true;
1866 if (sqe_flags & IOSQE_IO_DRAIN) {
1867 if (ctx->drain_disabled)
1869 io_init_req_drain(req);
1872 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
1873 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
1875 /* knock it to the slow queue path, will be drained there */
1876 if (ctx->drain_active)
1877 req->flags |= REQ_F_FORCE_ASYNC;
1878 /* if there is no link, we're at "next" request and need to drain */
1879 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
1880 ctx->drain_next = false;
1881 ctx->drain_active = true;
1882 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
1886 if (!def->ioprio && sqe->ioprio)
1888 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
1891 if (def->needs_file) {
1892 struct io_submit_state *state = &ctx->submit_state;
1894 req->cqe.fd = READ_ONCE(sqe->fd);
1897 * Plug now if we have more than 2 IO left after this, and the
1898 * target is potentially a read/write to block based storage.
1900 if (state->need_plug && def->plug) {
1901 state->plug_started = true;
1902 state->need_plug = false;
1903 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
1907 personality = READ_ONCE(sqe->personality);
1911 req->creds = xa_load(&ctx->personalities, personality);
1914 get_cred(req->creds);
1915 ret = security_uring_override_creds(req->creds);
1917 put_cred(req->creds);
1920 req->flags |= REQ_F_CREDS;
1923 return def->prep(req, sqe);
1926 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
1927 struct io_kiocb *req, int ret)
1929 struct io_ring_ctx *ctx = req->ctx;
1930 struct io_submit_link *link = &ctx->submit_state.link;
1931 struct io_kiocb *head = link->head;
1933 trace_io_uring_req_failed(sqe, req, ret);
1936 * Avoid breaking links in the middle as it renders links with SQPOLL
1937 * unusable. Instead of failing eagerly, continue assembling the link if
1938 * applicable and mark the head with REQ_F_FAIL. The link flushing code
1939 * should find the flag and handle the rest.
1941 req_fail_link_node(req, ret);
1942 if (head && !(head->flags & REQ_F_FAIL))
1943 req_fail_link_node(head, -ECANCELED);
1945 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
1947 link->last->link = req;
1951 io_queue_sqe_fallback(req);
1956 link->last->link = req;
1963 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
1964 const struct io_uring_sqe *sqe)
1965 __must_hold(&ctx->uring_lock)
1967 struct io_submit_link *link = &ctx->submit_state.link;
1970 ret = io_init_req(ctx, req, sqe);
1972 return io_submit_fail_init(sqe, req, ret);
1974 /* don't need @sqe from now on */
1975 trace_io_uring_submit_sqe(req, true);
1978 * If we already have a head request, queue this one for async
1979 * submittal once the head completes. If we don't have a head but
1980 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
1981 * submitted sync once the chain is complete. If none of those
1982 * conditions are true (normal request), then just queue it.
1984 if (unlikely(link->head)) {
1985 ret = io_req_prep_async(req);
1987 return io_submit_fail_init(sqe, req, ret);
1989 trace_io_uring_link(req, link->head);
1990 link->last->link = req;
1993 if (req->flags & IO_REQ_LINK_FLAGS)
1995 /* last request of the link, flush it */
1998 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2001 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2002 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2003 if (req->flags & IO_REQ_LINK_FLAGS) {
2008 io_queue_sqe_fallback(req);
2018 * Batched submission is done, ensure local IO is flushed out.
2020 static void io_submit_state_end(struct io_ring_ctx *ctx)
2022 struct io_submit_state *state = &ctx->submit_state;
2024 if (unlikely(state->link.head))
2025 io_queue_sqe_fallback(state->link.head);
2026 /* flush only after queuing links as they can generate completions */
2027 io_submit_flush_completions(ctx);
2028 if (state->plug_started)
2029 blk_finish_plug(&state->plug);
2033 * Start submission side cache.
2035 static void io_submit_state_start(struct io_submit_state *state,
2036 unsigned int max_ios)
2038 state->plug_started = false;
2039 state->need_plug = max_ios > 2;
2040 state->submit_nr = max_ios;
2041 /* set only head, no need to init link_last in advance */
2042 state->link.head = NULL;
2045 static void io_commit_sqring(struct io_ring_ctx *ctx)
2047 struct io_rings *rings = ctx->rings;
2050 * Ensure any loads from the SQEs are done at this point,
2051 * since once we write the new head, the application could
2052 * write new data to them.
2054 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2058 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2059 * that is mapped by userspace. This means that care needs to be taken to
2060 * ensure that reads are stable, as we cannot rely on userspace always
2061 * being a good citizen. If members of the sqe are validated and then later
2062 * used, it's important that those reads are done through READ_ONCE() to
2063 * prevent a re-load down the line.
2065 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
2067 unsigned head, mask = ctx->sq_entries - 1;
2068 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2071 * The cached sq head (or cq tail) serves two purposes:
2073 * 1) allows us to batch the cost of updating the user visible
2075 * 2) allows the kernel side to track the head on its own, even
2076 * though the application is the one updating it.
2078 head = READ_ONCE(ctx->sq_array[sq_idx]);
2079 if (likely(head < ctx->sq_entries)) {
2080 /* double index for 128-byte SQEs, twice as long */
2081 if (ctx->flags & IORING_SETUP_SQE128)
2083 return &ctx->sq_sqes[head];
2086 /* drop invalid entries */
2088 WRITE_ONCE(ctx->rings->sq_dropped,
2089 READ_ONCE(ctx->rings->sq_dropped) + 1);
2093 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2094 __must_hold(&ctx->uring_lock)
2096 unsigned int entries = io_sqring_entries(ctx);
2100 if (unlikely(!entries))
2102 /* make sure SQ entry isn't read before tail */
2103 ret = left = min3(nr, ctx->sq_entries, entries);
2104 io_get_task_refs(left);
2105 io_submit_state_start(&ctx->submit_state, left);
2108 const struct io_uring_sqe *sqe;
2109 struct io_kiocb *req;
2111 if (unlikely(!io_alloc_req_refill(ctx)))
2113 req = io_alloc_req(ctx);
2114 sqe = io_get_sqe(ctx);
2115 if (unlikely(!sqe)) {
2116 io_req_add_to_cache(req, ctx);
2121 * Continue submitting even for sqe failure if the
2122 * ring was setup with IORING_SETUP_SUBMIT_ALL
2124 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2125 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2131 if (unlikely(left)) {
2133 /* try again if it submitted nothing and can't allocate a req */
2134 if (!ret && io_req_cache_empty(ctx))
2136 current->io_uring->cached_refs += left;
2139 io_submit_state_end(ctx);
2140 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2141 io_commit_sqring(ctx);
2145 struct io_wait_queue {
2146 struct wait_queue_entry wq;
2147 struct io_ring_ctx *ctx;
2149 unsigned nr_timeouts;
2152 static inline bool io_should_wake(struct io_wait_queue *iowq)
2154 struct io_ring_ctx *ctx = iowq->ctx;
2155 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
2158 * Wake up if we have enough events, or if a timeout occurred since we
2159 * started waiting. For timeouts, we always want to return to userspace,
2160 * regardless of event count.
2162 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2165 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2166 int wake_flags, void *key)
2168 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
2172 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2173 * the task, and the next invocation will do it.
2175 if (io_should_wake(iowq) ||
2176 test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &iowq->ctx->check_cq))
2177 return autoremove_wake_function(curr, mode, wake_flags, key);
2181 int io_run_task_work_sig(void)
2183 if (io_run_task_work())
2185 if (task_sigpending(current))
2190 /* when returns >0, the caller should retry */
2191 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2192 struct io_wait_queue *iowq,
2196 unsigned long check_cq;
2198 /* make sure we run task_work before checking for signals */
2199 ret = io_run_task_work_sig();
2200 if (ret || io_should_wake(iowq))
2203 check_cq = READ_ONCE(ctx->check_cq);
2204 if (unlikely(check_cq)) {
2205 /* let the caller flush overflows, retry */
2206 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2208 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
2211 if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS))
2217 * Wait until events become available, if we don't already have some. The
2218 * application must reap them itself, as they reside on the shared cq ring.
2220 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2221 const sigset_t __user *sig, size_t sigsz,
2222 struct __kernel_timespec __user *uts)
2224 struct io_wait_queue iowq;
2225 struct io_rings *rings = ctx->rings;
2226 ktime_t timeout = KTIME_MAX;
2230 io_cqring_overflow_flush(ctx);
2231 if (io_cqring_events(ctx) >= min_events)
2233 if (!io_run_task_work())
2238 #ifdef CONFIG_COMPAT
2239 if (in_compat_syscall())
2240 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2244 ret = set_user_sigmask(sig, sigsz);
2251 struct timespec64 ts;
2253 if (get_timespec64(&ts, uts))
2255 timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2258 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2259 iowq.wq.private = current;
2260 INIT_LIST_HEAD(&iowq.wq.entry);
2262 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2263 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2265 trace_io_uring_cqring_wait(ctx, min_events);
2267 /* if we can't even flush overflow, don't wait for more */
2268 if (!io_cqring_overflow_flush(ctx)) {
2272 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2273 TASK_INTERRUPTIBLE);
2274 ret = io_cqring_wait_schedule(ctx, &iowq, timeout);
2278 finish_wait(&ctx->cq_wait, &iowq.wq);
2279 restore_saved_sigmask_unless(ret == -EINTR);
2281 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2284 static void io_mem_free(void *ptr)
2291 page = virt_to_head_page(ptr);
2292 if (put_page_testzero(page))
2293 free_compound_page(page);
2296 static void *io_mem_alloc(size_t size)
2298 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2300 return (void *) __get_free_pages(gfp, get_order(size));
2303 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2304 unsigned int cq_entries, size_t *sq_offset)
2306 struct io_rings *rings;
2307 size_t off, sq_array_size;
2309 off = struct_size(rings, cqes, cq_entries);
2310 if (off == SIZE_MAX)
2312 if (ctx->flags & IORING_SETUP_CQE32) {
2313 if (check_shl_overflow(off, 1, &off))
2318 off = ALIGN(off, SMP_CACHE_BYTES);
2326 sq_array_size = array_size(sizeof(u32), sq_entries);
2327 if (sq_array_size == SIZE_MAX)
2330 if (check_add_overflow(off, sq_array_size, &off))
2336 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2337 unsigned int eventfd_async)
2339 struct io_ev_fd *ev_fd;
2340 __s32 __user *fds = arg;
2343 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2344 lockdep_is_held(&ctx->uring_lock));
2348 if (copy_from_user(&fd, fds, sizeof(*fds)))
2351 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2355 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2356 if (IS_ERR(ev_fd->cq_ev_fd)) {
2357 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2362 spin_lock(&ctx->completion_lock);
2363 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2364 spin_unlock(&ctx->completion_lock);
2366 ev_fd->eventfd_async = eventfd_async;
2367 ctx->has_evfd = true;
2368 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2372 static void io_eventfd_put(struct rcu_head *rcu)
2374 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
2376 eventfd_ctx_put(ev_fd->cq_ev_fd);
2380 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2382 struct io_ev_fd *ev_fd;
2384 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2385 lockdep_is_held(&ctx->uring_lock));
2387 ctx->has_evfd = false;
2388 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2389 call_rcu(&ev_fd->rcu, io_eventfd_put);
2396 static void io_req_caches_free(struct io_ring_ctx *ctx)
2398 struct io_submit_state *state = &ctx->submit_state;
2401 mutex_lock(&ctx->uring_lock);
2402 io_flush_cached_locked_reqs(ctx, state);
2404 while (!io_req_cache_empty(ctx)) {
2405 struct io_wq_work_node *node;
2406 struct io_kiocb *req;
2408 node = wq_stack_extract(&state->free_list);
2409 req = container_of(node, struct io_kiocb, comp_list);
2410 kmem_cache_free(req_cachep, req);
2414 percpu_ref_put_many(&ctx->refs, nr);
2415 mutex_unlock(&ctx->uring_lock);
2418 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2420 io_sq_thread_finish(ctx);
2422 if (ctx->mm_account) {
2423 mmdrop(ctx->mm_account);
2424 ctx->mm_account = NULL;
2427 io_rsrc_refs_drop(ctx);
2428 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2429 io_wait_rsrc_data(ctx->buf_data);
2430 io_wait_rsrc_data(ctx->file_data);
2432 mutex_lock(&ctx->uring_lock);
2434 __io_sqe_buffers_unregister(ctx);
2436 __io_sqe_files_unregister(ctx);
2438 __io_cqring_overflow_flush(ctx, true);
2439 io_eventfd_unregister(ctx);
2440 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2441 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2442 mutex_unlock(&ctx->uring_lock);
2443 io_destroy_buffers(ctx);
2445 put_cred(ctx->sq_creds);
2446 if (ctx->submitter_task)
2447 put_task_struct(ctx->submitter_task);
2449 /* there are no registered resources left, nobody uses it */
2451 io_rsrc_node_destroy(ctx->rsrc_node);
2452 if (ctx->rsrc_backup_node)
2453 io_rsrc_node_destroy(ctx->rsrc_backup_node);
2454 flush_delayed_work(&ctx->rsrc_put_work);
2455 flush_delayed_work(&ctx->fallback_work);
2457 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2458 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
2460 #if defined(CONFIG_UNIX)
2461 if (ctx->ring_sock) {
2462 ctx->ring_sock->file = NULL; /* so that iput() is called */
2463 sock_release(ctx->ring_sock);
2466 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2467 WARN_ON_ONCE(ctx->notif_slots || ctx->nr_notif_slots);
2469 io_mem_free(ctx->rings);
2470 io_mem_free(ctx->sq_sqes);
2472 percpu_ref_exit(&ctx->refs);
2473 free_uid(ctx->user);
2474 io_req_caches_free(ctx);
2476 io_wq_put_hash(ctx->hash_map);
2477 kfree(ctx->cancel_table.hbs);
2478 kfree(ctx->cancel_table_locked.hbs);
2479 kfree(ctx->dummy_ubuf);
2481 xa_destroy(&ctx->io_bl_xa);
2485 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2487 struct io_ring_ctx *ctx = file->private_data;
2490 poll_wait(file, &ctx->cq_wait, wait);
2492 * synchronizes with barrier from wq_has_sleeper call in
2496 if (!io_sqring_full(ctx))
2497 mask |= EPOLLOUT | EPOLLWRNORM;
2500 * Don't flush cqring overflow list here, just do a simple check.
2501 * Otherwise there could possible be ABBA deadlock:
2504 * lock(&ctx->uring_lock);
2506 * lock(&ctx->uring_lock);
2509 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2510 * pushs them to do the flush.
2512 if (io_cqring_events(ctx) ||
2513 test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
2514 mask |= EPOLLIN | EPOLLRDNORM;
2519 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2521 const struct cred *creds;
2523 creds = xa_erase(&ctx->personalities, id);
2532 struct io_tctx_exit {
2533 struct callback_head task_work;
2534 struct completion completion;
2535 struct io_ring_ctx *ctx;
2538 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2540 struct io_uring_task *tctx = current->io_uring;
2541 struct io_tctx_exit *work;
2543 work = container_of(cb, struct io_tctx_exit, task_work);
2545 * When @in_idle, we're in cancellation and it's racy to remove the
2546 * node. It'll be removed by the end of cancellation, just ignore it.
2548 if (!atomic_read(&tctx->in_idle))
2549 io_uring_del_tctx_node((unsigned long)work->ctx);
2550 complete(&work->completion);
2553 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2555 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2557 return req->ctx == data;
2560 static __cold void io_ring_exit_work(struct work_struct *work)
2562 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2563 unsigned long timeout = jiffies + HZ * 60 * 5;
2564 unsigned long interval = HZ / 20;
2565 struct io_tctx_exit exit;
2566 struct io_tctx_node *node;
2570 * If we're doing polled IO and end up having requests being
2571 * submitted async (out-of-line), then completions can come in while
2572 * we're waiting for refs to drop. We need to reap these manually,
2573 * as nobody else will be looking for them.
2576 while (io_uring_try_cancel_requests(ctx, NULL, true))
2580 struct io_sq_data *sqd = ctx->sq_data;
2581 struct task_struct *tsk;
2583 io_sq_thread_park(sqd);
2585 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
2586 io_wq_cancel_cb(tsk->io_uring->io_wq,
2587 io_cancel_ctx_cb, ctx, true);
2588 io_sq_thread_unpark(sqd);
2591 io_req_caches_free(ctx);
2593 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
2594 /* there is little hope left, don't run it too often */
2597 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
2599 init_completion(&exit.completion);
2600 init_task_work(&exit.task_work, io_tctx_exit_cb);
2603 * Some may use context even when all refs and requests have been put,
2604 * and they are free to do so while still holding uring_lock or
2605 * completion_lock, see io_req_task_submit(). Apart from other work,
2606 * this lock/unlock section also waits them to finish.
2608 mutex_lock(&ctx->uring_lock);
2609 while (!list_empty(&ctx->tctx_list)) {
2610 WARN_ON_ONCE(time_after(jiffies, timeout));
2612 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
2614 /* don't spin on a single task if cancellation failed */
2615 list_rotate_left(&ctx->tctx_list);
2616 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
2617 if (WARN_ON_ONCE(ret))
2620 mutex_unlock(&ctx->uring_lock);
2621 wait_for_completion(&exit.completion);
2622 mutex_lock(&ctx->uring_lock);
2624 mutex_unlock(&ctx->uring_lock);
2625 spin_lock(&ctx->completion_lock);
2626 spin_unlock(&ctx->completion_lock);
2628 io_ring_ctx_free(ctx);
2631 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
2633 unsigned long index;
2634 struct creds *creds;
2636 mutex_lock(&ctx->uring_lock);
2637 percpu_ref_kill(&ctx->refs);
2639 __io_cqring_overflow_flush(ctx, true);
2640 xa_for_each(&ctx->personalities, index, creds)
2641 io_unregister_personality(ctx, index);
2643 io_poll_remove_all(ctx, NULL, true);
2644 mutex_unlock(&ctx->uring_lock);
2646 /* failed during ring init, it couldn't have issued any requests */
2648 io_kill_timeouts(ctx, NULL, true);
2649 /* if we failed setting up the ctx, we might not have any rings */
2650 io_iopoll_try_reap_events(ctx);
2651 /* drop cached put refs after potentially doing completions */
2652 if (current->io_uring)
2653 io_uring_drop_tctx_refs(current);
2656 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
2658 * Use system_unbound_wq to avoid spawning tons of event kworkers
2659 * if we're exiting a ton of rings at the same time. It just adds
2660 * noise and overhead, there's no discernable change in runtime
2661 * over using system_wq.
2663 queue_work(system_unbound_wq, &ctx->exit_work);
2666 static int io_uring_release(struct inode *inode, struct file *file)
2668 struct io_ring_ctx *ctx = file->private_data;
2670 file->private_data = NULL;
2671 io_ring_ctx_wait_and_kill(ctx);
2675 struct io_task_cancel {
2676 struct task_struct *task;
2680 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
2682 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2683 struct io_task_cancel *cancel = data;
2685 return io_match_task_safe(req, cancel->task, cancel->all);
2688 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
2689 struct task_struct *task,
2692 struct io_defer_entry *de;
2695 spin_lock(&ctx->completion_lock);
2696 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
2697 if (io_match_task_safe(de->req, task, cancel_all)) {
2698 list_cut_position(&list, &ctx->defer_list, &de->list);
2702 spin_unlock(&ctx->completion_lock);
2703 if (list_empty(&list))
2706 while (!list_empty(&list)) {
2707 de = list_first_entry(&list, struct io_defer_entry, list);
2708 list_del_init(&de->list);
2709 io_req_complete_failed(de->req, -ECANCELED);
2715 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
2717 struct io_tctx_node *node;
2718 enum io_wq_cancel cret;
2721 mutex_lock(&ctx->uring_lock);
2722 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
2723 struct io_uring_task *tctx = node->task->io_uring;
2726 * io_wq will stay alive while we hold uring_lock, because it's
2727 * killed after ctx nodes, which requires to take the lock.
2729 if (!tctx || !tctx->io_wq)
2731 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
2732 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2734 mutex_unlock(&ctx->uring_lock);
2739 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
2740 struct task_struct *task,
2743 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
2744 struct io_uring_task *tctx = task ? task->io_uring : NULL;
2745 enum io_wq_cancel cret;
2748 /* failed during ring init, it couldn't have issued any requests */
2753 ret |= io_uring_try_cancel_iowq(ctx);
2754 } else if (tctx && tctx->io_wq) {
2756 * Cancels requests of all rings, not only @ctx, but
2757 * it's fine as the task is in exit/exec.
2759 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
2761 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2764 /* SQPOLL thread does its own polling */
2765 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
2766 (ctx->sq_data && ctx->sq_data->thread == current)) {
2767 while (!wq_list_empty(&ctx->iopoll_list)) {
2768 io_iopoll_try_reap_events(ctx);
2773 ret |= io_cancel_defer_files(ctx, task, cancel_all);
2774 mutex_lock(&ctx->uring_lock);
2775 ret |= io_poll_remove_all(ctx, task, cancel_all);
2776 mutex_unlock(&ctx->uring_lock);
2777 ret |= io_kill_timeouts(ctx, task, cancel_all);
2779 ret |= io_run_task_work();
2783 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
2786 return atomic_read(&tctx->inflight_tracked);
2787 return percpu_counter_sum(&tctx->inflight);
2791 * Find any io_uring ctx that this task has registered or done IO on, and cancel
2792 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
2794 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
2796 struct io_uring_task *tctx = current->io_uring;
2797 struct io_ring_ctx *ctx;
2801 WARN_ON_ONCE(sqd && sqd->thread != current);
2803 if (!current->io_uring)
2806 io_wq_exit_start(tctx->io_wq);
2808 atomic_inc(&tctx->in_idle);
2812 io_uring_drop_tctx_refs(current);
2813 /* read completions before cancelations */
2814 inflight = tctx_inflight(tctx, !cancel_all);
2819 struct io_tctx_node *node;
2820 unsigned long index;
2822 xa_for_each(&tctx->xa, index, node) {
2823 /* sqpoll task will cancel all its requests */
2824 if (node->ctx->sq_data)
2826 loop |= io_uring_try_cancel_requests(node->ctx,
2827 current, cancel_all);
2830 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
2831 loop |= io_uring_try_cancel_requests(ctx,
2841 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
2843 io_uring_drop_tctx_refs(current);
2846 * If we've seen completions, retry without waiting. This
2847 * avoids a race where a completion comes in before we did
2848 * prepare_to_wait().
2850 if (inflight == tctx_inflight(tctx, !cancel_all))
2852 finish_wait(&tctx->wait, &wait);
2855 io_uring_clean_tctx(tctx);
2858 * We shouldn't run task_works after cancel, so just leave
2859 * ->in_idle set for normal exit.
2861 atomic_dec(&tctx->in_idle);
2862 /* for exec all current's requests should be gone, kill tctx */
2863 __io_uring_free(current);
2867 void __io_uring_cancel(bool cancel_all)
2869 io_uring_cancel_generic(cancel_all, NULL);
2872 static void *io_uring_validate_mmap_request(struct file *file,
2873 loff_t pgoff, size_t sz)
2875 struct io_ring_ctx *ctx = file->private_data;
2876 loff_t offset = pgoff << PAGE_SHIFT;
2881 case IORING_OFF_SQ_RING:
2882 case IORING_OFF_CQ_RING:
2885 case IORING_OFF_SQES:
2889 return ERR_PTR(-EINVAL);
2892 page = virt_to_head_page(ptr);
2893 if (sz > page_size(page))
2894 return ERR_PTR(-EINVAL);
2901 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
2903 size_t sz = vma->vm_end - vma->vm_start;
2907 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
2909 return PTR_ERR(ptr);
2911 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
2912 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
2915 #else /* !CONFIG_MMU */
2917 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
2919 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
2922 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
2924 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
2927 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
2928 unsigned long addr, unsigned long len,
2929 unsigned long pgoff, unsigned long flags)
2933 ptr = io_uring_validate_mmap_request(file, pgoff, len);
2935 return PTR_ERR(ptr);
2937 return (unsigned long) ptr;
2940 #endif /* !CONFIG_MMU */
2942 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
2944 if (flags & IORING_ENTER_EXT_ARG) {
2945 struct io_uring_getevents_arg arg;
2947 if (argsz != sizeof(arg))
2949 if (copy_from_user(&arg, argp, sizeof(arg)))
2955 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
2956 struct __kernel_timespec __user **ts,
2957 const sigset_t __user **sig)
2959 struct io_uring_getevents_arg arg;
2962 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
2963 * is just a pointer to the sigset_t.
2965 if (!(flags & IORING_ENTER_EXT_ARG)) {
2966 *sig = (const sigset_t __user *) argp;
2972 * EXT_ARG is set - ensure we agree on the size of it and copy in our
2973 * timespec and sigset_t pointers if good.
2975 if (*argsz != sizeof(arg))
2977 if (copy_from_user(&arg, argp, sizeof(arg)))
2981 *sig = u64_to_user_ptr(arg.sigmask);
2982 *argsz = arg.sigmask_sz;
2983 *ts = u64_to_user_ptr(arg.ts);
2987 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
2988 u32, min_complete, u32, flags, const void __user *, argp,
2991 struct io_ring_ctx *ctx;
2997 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
2998 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
2999 IORING_ENTER_REGISTERED_RING)))
3003 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3004 * need only dereference our task private array to find it.
3006 if (flags & IORING_ENTER_REGISTERED_RING) {
3007 struct io_uring_task *tctx = current->io_uring;
3009 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3011 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3012 f.file = tctx->registered_rings[fd];
3014 if (unlikely(!f.file))
3018 if (unlikely(!f.file))
3021 if (unlikely(!io_is_uring_fops(f.file)))
3025 ctx = f.file->private_data;
3027 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3031 * For SQ polling, the thread will do all submissions and completions.
3032 * Just return the requested submit count, and wake the thread if
3036 if (ctx->flags & IORING_SETUP_SQPOLL) {
3037 io_cqring_overflow_flush(ctx);
3039 if (unlikely(ctx->sq_data->thread == NULL)) {
3043 if (flags & IORING_ENTER_SQ_WAKEUP)
3044 wake_up(&ctx->sq_data->wait);
3045 if (flags & IORING_ENTER_SQ_WAIT) {
3046 ret = io_sqpoll_wait_sq(ctx);
3051 } else if (to_submit) {
3052 ret = io_uring_add_tctx_node(ctx);
3056 mutex_lock(&ctx->uring_lock);
3057 ret = io_submit_sqes(ctx, to_submit);
3058 if (ret != to_submit) {
3059 mutex_unlock(&ctx->uring_lock);
3062 if ((flags & IORING_ENTER_GETEVENTS) && ctx->syscall_iopoll)
3064 mutex_unlock(&ctx->uring_lock);
3066 if (flags & IORING_ENTER_GETEVENTS) {
3068 if (ctx->syscall_iopoll) {
3070 * We disallow the app entering submit/complete with
3071 * polling, but we still need to lock the ring to
3072 * prevent racing with polled issue that got punted to
3075 mutex_lock(&ctx->uring_lock);
3077 ret2 = io_validate_ext_arg(flags, argp, argsz);
3078 if (likely(!ret2)) {
3079 min_complete = min(min_complete,
3081 ret2 = io_iopoll_check(ctx, min_complete);
3083 mutex_unlock(&ctx->uring_lock);
3085 const sigset_t __user *sig;
3086 struct __kernel_timespec __user *ts;
3088 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3089 if (likely(!ret2)) {
3090 min_complete = min(min_complete,
3092 ret2 = io_cqring_wait(ctx, min_complete, sig,
3101 * EBADR indicates that one or more CQE were dropped.
3102 * Once the user has been informed we can clear the bit
3103 * as they are obviously ok with those drops.
3105 if (unlikely(ret2 == -EBADR))
3106 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3115 static const struct file_operations io_uring_fops = {
3116 .release = io_uring_release,
3117 .mmap = io_uring_mmap,
3119 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3120 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3122 .poll = io_uring_poll,
3123 #ifdef CONFIG_PROC_FS
3124 .show_fdinfo = io_uring_show_fdinfo,
3128 bool io_is_uring_fops(struct file *file)
3130 return file->f_op == &io_uring_fops;
3133 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3134 struct io_uring_params *p)
3136 struct io_rings *rings;
3137 size_t size, sq_array_offset;
3139 /* make sure these are sane, as we already accounted them */
3140 ctx->sq_entries = p->sq_entries;
3141 ctx->cq_entries = p->cq_entries;
3143 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3144 if (size == SIZE_MAX)
3147 rings = io_mem_alloc(size);
3152 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3153 rings->sq_ring_mask = p->sq_entries - 1;
3154 rings->cq_ring_mask = p->cq_entries - 1;
3155 rings->sq_ring_entries = p->sq_entries;
3156 rings->cq_ring_entries = p->cq_entries;
3158 if (p->flags & IORING_SETUP_SQE128)
3159 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3161 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3162 if (size == SIZE_MAX) {
3163 io_mem_free(ctx->rings);
3168 ctx->sq_sqes = io_mem_alloc(size);
3169 if (!ctx->sq_sqes) {
3170 io_mem_free(ctx->rings);
3178 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3182 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3186 ret = __io_uring_add_tctx_node(ctx, false);
3191 fd_install(fd, file);
3196 * Allocate an anonymous fd, this is what constitutes the application
3197 * visible backing of an io_uring instance. The application mmaps this
3198 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3199 * we have to tie this fd to a socket for file garbage collection purposes.
3201 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3204 #if defined(CONFIG_UNIX)
3207 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3210 return ERR_PTR(ret);
3213 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3214 O_RDWR | O_CLOEXEC, NULL);
3215 #if defined(CONFIG_UNIX)
3217 sock_release(ctx->ring_sock);
3218 ctx->ring_sock = NULL;
3220 ctx->ring_sock->file = file;
3226 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3227 struct io_uring_params __user *params)
3229 struct io_ring_ctx *ctx;
3235 if (entries > IORING_MAX_ENTRIES) {
3236 if (!(p->flags & IORING_SETUP_CLAMP))
3238 entries = IORING_MAX_ENTRIES;
3242 * Use twice as many entries for the CQ ring. It's possible for the
3243 * application to drive a higher depth than the size of the SQ ring,
3244 * since the sqes are only used at submission time. This allows for
3245 * some flexibility in overcommitting a bit. If the application has
3246 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3247 * of CQ ring entries manually.
3249 p->sq_entries = roundup_pow_of_two(entries);
3250 if (p->flags & IORING_SETUP_CQSIZE) {
3252 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3253 * to a power-of-two, if it isn't already. We do NOT impose
3254 * any cq vs sq ring sizing.
3258 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3259 if (!(p->flags & IORING_SETUP_CLAMP))
3261 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3263 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3264 if (p->cq_entries < p->sq_entries)
3267 p->cq_entries = 2 * p->sq_entries;
3270 ctx = io_ring_ctx_alloc(p);
3275 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3276 * space applications don't need to do io completion events
3277 * polling again, they can rely on io_sq_thread to do polling
3278 * work, which can reduce cpu usage and uring_lock contention.
3280 if (ctx->flags & IORING_SETUP_IOPOLL &&
3281 !(ctx->flags & IORING_SETUP_SQPOLL))
3282 ctx->syscall_iopoll = 1;
3284 ctx->compat = in_compat_syscall();
3285 if (!capable(CAP_IPC_LOCK))
3286 ctx->user = get_uid(current_user());
3289 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3290 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3293 if (ctx->flags & IORING_SETUP_SQPOLL) {
3294 /* IPI related flags don't make sense with SQPOLL */
3295 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3296 IORING_SETUP_TASKRUN_FLAG))
3298 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3299 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3300 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3302 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
3304 ctx->notify_method = TWA_SIGNAL;
3308 * This is just grabbed for accounting purposes. When a process exits,
3309 * the mm is exited and dropped before the files, hence we need to hang
3310 * on to this mm purely for the purposes of being able to unaccount
3311 * memory (locked/pinned vm). It's not used for anything else.
3313 mmgrab(current->mm);
3314 ctx->mm_account = current->mm;
3316 ret = io_allocate_scq_urings(ctx, p);
3320 ret = io_sq_offload_create(ctx, p);
3323 /* always set a rsrc node */
3324 ret = io_rsrc_node_switch_start(ctx);
3327 io_rsrc_node_switch(ctx, NULL);
3329 memset(&p->sq_off, 0, sizeof(p->sq_off));
3330 p->sq_off.head = offsetof(struct io_rings, sq.head);
3331 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3332 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3333 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3334 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3335 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3336 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3338 memset(&p->cq_off, 0, sizeof(p->cq_off));
3339 p->cq_off.head = offsetof(struct io_rings, cq.head);
3340 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3341 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3342 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3343 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3344 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3345 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3347 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3348 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3349 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3350 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3351 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3352 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3353 IORING_FEAT_LINKED_FILE;
3355 if (copy_to_user(params, p, sizeof(*p))) {
3360 file = io_uring_get_file(ctx);
3362 ret = PTR_ERR(file);
3367 * Install ring fd as the very last thing, so we don't risk someone
3368 * having closed it before we finish setup
3370 ret = io_uring_install_fd(ctx, file);
3372 /* fput will clean it up */
3377 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3380 io_ring_ctx_wait_and_kill(ctx);
3385 * Sets up an aio uring context, and returns the fd. Applications asks for a
3386 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3387 * params structure passed in.
3389 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3391 struct io_uring_params p;
3394 if (copy_from_user(&p, params, sizeof(p)))
3396 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3401 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3402 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3403 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3404 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3405 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3406 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3407 IORING_SETUP_SINGLE_ISSUER))
3410 return io_uring_create(entries, &p, params);
3413 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3414 struct io_uring_params __user *, params)
3416 return io_uring_setup(entries, params);
3419 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3422 struct io_uring_probe *p;
3426 size = struct_size(p, ops, nr_args);
3427 if (size == SIZE_MAX)
3429 p = kzalloc(size, GFP_KERNEL);
3434 if (copy_from_user(p, arg, size))
3437 if (memchr_inv(p, 0, size))
3440 p->last_op = IORING_OP_LAST - 1;
3441 if (nr_args > IORING_OP_LAST)
3442 nr_args = IORING_OP_LAST;
3444 for (i = 0; i < nr_args; i++) {
3446 if (!io_op_defs[i].not_supported)
3447 p->ops[i].flags = IO_URING_OP_SUPPORTED;
3452 if (copy_to_user(arg, p, size))
3459 static int io_register_personality(struct io_ring_ctx *ctx)
3461 const struct cred *creds;
3465 creds = get_current_cred();
3467 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
3468 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
3476 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
3477 void __user *arg, unsigned int nr_args)
3479 struct io_uring_restriction *res;
3483 /* Restrictions allowed only if rings started disabled */
3484 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3487 /* We allow only a single restrictions registration */
3488 if (ctx->restrictions.registered)
3491 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
3494 size = array_size(nr_args, sizeof(*res));
3495 if (size == SIZE_MAX)
3498 res = memdup_user(arg, size);
3500 return PTR_ERR(res);
3504 for (i = 0; i < nr_args; i++) {
3505 switch (res[i].opcode) {
3506 case IORING_RESTRICTION_REGISTER_OP:
3507 if (res[i].register_op >= IORING_REGISTER_LAST) {
3512 __set_bit(res[i].register_op,
3513 ctx->restrictions.register_op);
3515 case IORING_RESTRICTION_SQE_OP:
3516 if (res[i].sqe_op >= IORING_OP_LAST) {
3521 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
3523 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
3524 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
3526 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
3527 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
3536 /* Reset all restrictions if an error happened */
3538 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
3540 ctx->restrictions.registered = true;
3546 static int io_register_enable_rings(struct io_ring_ctx *ctx)
3548 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3551 if (ctx->restrictions.registered)
3552 ctx->restricted = 1;
3554 ctx->flags &= ~IORING_SETUP_R_DISABLED;
3555 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
3556 wake_up(&ctx->sq_data->wait);
3560 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
3561 void __user *arg, unsigned len)
3563 struct io_uring_task *tctx = current->io_uring;
3564 cpumask_var_t new_mask;
3567 if (!tctx || !tctx->io_wq)
3570 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3573 cpumask_clear(new_mask);
3574 if (len > cpumask_size())
3575 len = cpumask_size();
3577 if (in_compat_syscall()) {
3578 ret = compat_get_bitmap(cpumask_bits(new_mask),
3579 (const compat_ulong_t __user *)arg,
3580 len * 8 /* CHAR_BIT */);
3582 ret = copy_from_user(new_mask, arg, len);
3586 free_cpumask_var(new_mask);
3590 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
3591 free_cpumask_var(new_mask);
3595 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
3597 struct io_uring_task *tctx = current->io_uring;
3599 if (!tctx || !tctx->io_wq)
3602 return io_wq_cpu_affinity(tctx->io_wq, NULL);
3605 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
3607 __must_hold(&ctx->uring_lock)
3609 struct io_tctx_node *node;
3610 struct io_uring_task *tctx = NULL;
3611 struct io_sq_data *sqd = NULL;
3615 if (copy_from_user(new_count, arg, sizeof(new_count)))
3617 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3618 if (new_count[i] > INT_MAX)
3621 if (ctx->flags & IORING_SETUP_SQPOLL) {
3625 * Observe the correct sqd->lock -> ctx->uring_lock
3626 * ordering. Fine to drop uring_lock here, we hold
3629 refcount_inc(&sqd->refs);
3630 mutex_unlock(&ctx->uring_lock);
3631 mutex_lock(&sqd->lock);
3632 mutex_lock(&ctx->uring_lock);
3634 tctx = sqd->thread->io_uring;
3637 tctx = current->io_uring;
3640 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
3642 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3644 ctx->iowq_limits[i] = new_count[i];
3645 ctx->iowq_limits_set = true;
3647 if (tctx && tctx->io_wq) {
3648 ret = io_wq_max_workers(tctx->io_wq, new_count);
3652 memset(new_count, 0, sizeof(new_count));
3656 mutex_unlock(&sqd->lock);
3657 io_put_sq_data(sqd);
3660 if (copy_to_user(arg, new_count, sizeof(new_count)))
3663 /* that's it for SQPOLL, only the SQPOLL task creates requests */
3667 /* now propagate the restriction to all registered users */
3668 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3669 struct io_uring_task *tctx = node->task->io_uring;
3671 if (WARN_ON_ONCE(!tctx->io_wq))
3674 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3675 new_count[i] = ctx->iowq_limits[i];
3676 /* ignore errors, it always returns zero anyway */
3677 (void)io_wq_max_workers(tctx->io_wq, new_count);
3682 mutex_unlock(&sqd->lock);
3683 io_put_sq_data(sqd);
3688 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
3689 void __user *arg, unsigned nr_args)
3690 __releases(ctx->uring_lock)
3691 __acquires(ctx->uring_lock)
3696 * We don't quiesce the refs for register anymore and so it can't be
3697 * dying as we're holding a file ref here.
3699 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
3702 if (ctx->restricted) {
3703 if (opcode >= IORING_REGISTER_LAST)
3705 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
3706 if (!test_bit(opcode, ctx->restrictions.register_op))
3711 case IORING_REGISTER_BUFFERS:
3715 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
3717 case IORING_UNREGISTER_BUFFERS:
3721 ret = io_sqe_buffers_unregister(ctx);
3723 case IORING_REGISTER_FILES:
3727 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
3729 case IORING_UNREGISTER_FILES:
3733 ret = io_sqe_files_unregister(ctx);
3735 case IORING_REGISTER_FILES_UPDATE:
3736 ret = io_register_files_update(ctx, arg, nr_args);
3738 case IORING_REGISTER_EVENTFD:
3742 ret = io_eventfd_register(ctx, arg, 0);
3744 case IORING_REGISTER_EVENTFD_ASYNC:
3748 ret = io_eventfd_register(ctx, arg, 1);
3750 case IORING_UNREGISTER_EVENTFD:
3754 ret = io_eventfd_unregister(ctx);
3756 case IORING_REGISTER_PROBE:
3758 if (!arg || nr_args > 256)
3760 ret = io_probe(ctx, arg, nr_args);
3762 case IORING_REGISTER_PERSONALITY:
3766 ret = io_register_personality(ctx);
3768 case IORING_UNREGISTER_PERSONALITY:
3772 ret = io_unregister_personality(ctx, nr_args);
3774 case IORING_REGISTER_ENABLE_RINGS:
3778 ret = io_register_enable_rings(ctx);
3780 case IORING_REGISTER_RESTRICTIONS:
3781 ret = io_register_restrictions(ctx, arg, nr_args);
3783 case IORING_REGISTER_FILES2:
3784 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
3786 case IORING_REGISTER_FILES_UPDATE2:
3787 ret = io_register_rsrc_update(ctx, arg, nr_args,
3790 case IORING_REGISTER_BUFFERS2:
3791 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
3793 case IORING_REGISTER_BUFFERS_UPDATE:
3794 ret = io_register_rsrc_update(ctx, arg, nr_args,
3795 IORING_RSRC_BUFFER);
3797 case IORING_REGISTER_IOWQ_AFF:
3799 if (!arg || !nr_args)
3801 ret = io_register_iowq_aff(ctx, arg, nr_args);
3803 case IORING_UNREGISTER_IOWQ_AFF:
3807 ret = io_unregister_iowq_aff(ctx);
3809 case IORING_REGISTER_IOWQ_MAX_WORKERS:
3811 if (!arg || nr_args != 2)
3813 ret = io_register_iowq_max_workers(ctx, arg);
3815 case IORING_REGISTER_RING_FDS:
3816 ret = io_ringfd_register(ctx, arg, nr_args);
3818 case IORING_UNREGISTER_RING_FDS:
3819 ret = io_ringfd_unregister(ctx, arg, nr_args);
3821 case IORING_REGISTER_PBUF_RING:
3823 if (!arg || nr_args != 1)
3825 ret = io_register_pbuf_ring(ctx, arg);
3827 case IORING_UNREGISTER_PBUF_RING:
3829 if (!arg || nr_args != 1)
3831 ret = io_unregister_pbuf_ring(ctx, arg);
3833 case IORING_REGISTER_SYNC_CANCEL:
3835 if (!arg || nr_args != 1)
3837 ret = io_sync_cancel(ctx, arg);
3839 case IORING_REGISTER_FILE_ALLOC_RANGE:
3841 if (!arg || nr_args)
3843 ret = io_register_file_alloc_range(ctx, arg);
3853 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
3854 void __user *, arg, unsigned int, nr_args)
3856 struct io_ring_ctx *ctx;
3865 if (!io_is_uring_fops(f.file))
3868 ctx = f.file->private_data;
3872 mutex_lock(&ctx->uring_lock);
3873 ret = __io_uring_register(ctx, opcode, arg, nr_args);
3874 mutex_unlock(&ctx->uring_lock);
3875 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
3881 static int __init io_uring_init(void)
3883 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
3884 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
3885 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
3888 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
3889 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
3890 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
3891 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
3892 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
3893 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
3894 BUILD_BUG_SQE_ELEM(1, __u8, flags);
3895 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
3896 BUILD_BUG_SQE_ELEM(4, __s32, fd);
3897 BUILD_BUG_SQE_ELEM(8, __u64, off);
3898 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
3899 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
3900 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
3901 BUILD_BUG_SQE_ELEM(16, __u64, addr);
3902 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
3903 BUILD_BUG_SQE_ELEM(24, __u32, len);
3904 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
3905 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
3906 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
3907 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
3908 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
3909 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
3910 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
3911 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
3912 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
3913 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
3914 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
3915 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
3916 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
3917 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
3918 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
3919 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
3920 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
3921 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
3922 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
3923 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
3924 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
3925 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
3926 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
3927 BUILD_BUG_SQE_ELEM(42, __u16, personality);
3928 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
3929 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
3930 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
3931 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
3932 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
3933 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
3934 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
3936 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
3937 sizeof(struct io_uring_rsrc_update));
3938 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
3939 sizeof(struct io_uring_rsrc_update2));
3941 /* ->buf_index is u16 */
3942 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
3943 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
3944 offsetof(struct io_uring_buf_ring, tail));
3946 /* should fit into one byte */
3947 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
3948 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
3949 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
3951 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
3953 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
3955 io_uring_optable_init();
3957 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
3961 __initcall(io_uring_init);