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_cqring (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 <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
73 #include <linux/highmem.h>
74 #include <linux/namei.h>
75 #include <linux/fsnotify.h>
76 #include <linux/fadvise.h>
77 #include <linux/eventpoll.h>
78 #include <linux/fs_struct.h>
79 #include <linux/splice.h>
80 #include <linux/task_work.h>
81 #include <linux/pagemap.h>
82 #include <linux/io_uring.h>
83 #include <linux/blk-cgroup.h>
84 #include <linux/audit.h>
86 #define CREATE_TRACE_POINTS
87 #include <trace/events/io_uring.h>
89 #include <uapi/linux/io_uring.h>
94 #define IORING_MAX_ENTRIES 32768
95 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
98 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
100 #define IORING_FILE_TABLE_SHIFT 9
101 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
102 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
103 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
104 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
105 IORING_REGISTER_LAST + IORING_OP_LAST)
108 u32 head ____cacheline_aligned_in_smp;
109 u32 tail ____cacheline_aligned_in_smp;
113 * This data is shared with the application through the mmap at offsets
114 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
116 * The offsets to the member fields are published through struct
117 * io_sqring_offsets when calling io_uring_setup.
121 * Head and tail offsets into the ring; the offsets need to be
122 * masked to get valid indices.
124 * The kernel controls head of the sq ring and the tail of the cq ring,
125 * and the application controls tail of the sq ring and the head of the
128 struct io_uring sq, cq;
130 * Bitmasks to apply to head and tail offsets (constant, equals
133 u32 sq_ring_mask, cq_ring_mask;
134 /* Ring sizes (constant, power of 2) */
135 u32 sq_ring_entries, cq_ring_entries;
137 * Number of invalid entries dropped by the kernel due to
138 * invalid index stored in array
140 * Written by the kernel, shouldn't be modified by the
141 * application (i.e. get number of "new events" by comparing to
144 * After a new SQ head value was read by the application this
145 * counter includes all submissions that were dropped reaching
146 * the new SQ head (and possibly more).
152 * Written by the kernel, shouldn't be modified by the
155 * The application needs a full memory barrier before checking
156 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
162 * Written by the application, shouldn't be modified by the
167 * Number of completion events lost because the queue was full;
168 * this should be avoided by the application by making sure
169 * there are not more requests pending than there is space in
170 * the completion queue.
172 * Written by the kernel, shouldn't be modified by the
173 * application (i.e. get number of "new events" by comparing to
176 * As completion events come in out of order this counter is not
177 * ordered with any other data.
181 * Ring buffer of completion events.
183 * The kernel writes completion events fresh every time they are
184 * produced, so the application is allowed to modify pending
187 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
190 struct io_mapped_ubuf {
193 struct bio_vec *bvec;
194 unsigned int nr_bvecs;
195 unsigned long acct_pages;
198 struct fixed_file_table {
202 struct fixed_file_ref_node {
203 struct percpu_ref refs;
204 struct list_head node;
205 struct list_head file_list;
206 struct fixed_file_data *file_data;
207 struct llist_node llist;
210 struct fixed_file_data {
211 struct fixed_file_table *table;
212 struct io_ring_ctx *ctx;
214 struct fixed_file_ref_node *node;
215 struct percpu_ref refs;
216 struct completion done;
217 struct list_head ref_list;
222 struct list_head list;
228 struct io_restriction {
229 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
230 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
231 u8 sqe_flags_allowed;
232 u8 sqe_flags_required;
240 /* ctx's that are using this sqd */
241 struct list_head ctx_list;
242 struct list_head ctx_new_list;
243 struct mutex ctx_lock;
245 struct task_struct *thread;
246 struct wait_queue_head wait;
251 struct percpu_ref refs;
252 } ____cacheline_aligned_in_smp;
256 unsigned int compat: 1;
257 unsigned int limit_mem: 1;
258 unsigned int cq_overflow_flushed: 1;
259 unsigned int drain_next: 1;
260 unsigned int eventfd_async: 1;
261 unsigned int restricted: 1;
264 * Ring buffer of indices into array of io_uring_sqe, which is
265 * mmapped by the application using the IORING_OFF_SQES offset.
267 * This indirection could e.g. be used to assign fixed
268 * io_uring_sqe entries to operations and only submit them to
269 * the queue when needed.
271 * The kernel modifies neither the indices array nor the entries
275 unsigned cached_sq_head;
278 unsigned sq_thread_idle;
279 unsigned cached_sq_dropped;
280 unsigned cached_cq_overflow;
281 unsigned long sq_check_overflow;
283 struct list_head defer_list;
284 struct list_head timeout_list;
285 struct list_head cq_overflow_list;
287 wait_queue_head_t inflight_wait;
288 struct io_uring_sqe *sq_sqes;
289 } ____cacheline_aligned_in_smp;
291 struct io_rings *rings;
297 * For SQPOLL usage - we hold a reference to the parent task, so we
298 * have access to the ->files
300 struct task_struct *sqo_task;
302 /* Only used for accounting purposes */
303 struct mm_struct *mm_account;
305 #ifdef CONFIG_BLK_CGROUP
306 struct cgroup_subsys_state *sqo_blkcg_css;
309 struct io_sq_data *sq_data; /* if using sq thread polling */
311 struct wait_queue_head sqo_sq_wait;
312 struct wait_queue_entry sqo_wait_entry;
313 struct list_head sqd_list;
316 * If used, fixed file set. Writers must ensure that ->refs is dead,
317 * readers must ensure that ->refs is alive as long as the file* is
318 * used. Only updated through io_uring_register(2).
320 struct fixed_file_data *file_data;
321 unsigned nr_user_files;
323 /* if used, fixed mapped user buffers */
324 unsigned nr_user_bufs;
325 struct io_mapped_ubuf *user_bufs;
327 struct user_struct *user;
329 const struct cred *creds;
333 unsigned int sessionid;
336 struct completion ref_comp;
337 struct completion sq_thread_comp;
339 /* if all else fails... */
340 struct io_kiocb *fallback_req;
342 #if defined(CONFIG_UNIX)
343 struct socket *ring_sock;
346 struct idr io_buffer_idr;
348 struct idr personality_idr;
351 unsigned cached_cq_tail;
354 atomic_t cq_timeouts;
355 unsigned long cq_check_overflow;
356 struct wait_queue_head cq_wait;
357 struct fasync_struct *cq_fasync;
358 struct eventfd_ctx *cq_ev_fd;
359 } ____cacheline_aligned_in_smp;
362 struct mutex uring_lock;
363 wait_queue_head_t wait;
364 } ____cacheline_aligned_in_smp;
367 spinlock_t completion_lock;
370 * ->iopoll_list is protected by the ctx->uring_lock for
371 * io_uring instances that don't use IORING_SETUP_SQPOLL.
372 * For SQPOLL, only the single threaded io_sq_thread() will
373 * manipulate the list, hence no extra locking is needed there.
375 struct list_head iopoll_list;
376 struct hlist_head *cancel_hash;
377 unsigned cancel_hash_bits;
378 bool poll_multi_file;
380 spinlock_t inflight_lock;
381 struct list_head inflight_list;
382 } ____cacheline_aligned_in_smp;
384 struct delayed_work file_put_work;
385 struct llist_head file_put_llist;
387 struct work_struct exit_work;
388 struct io_restriction restrictions;
392 * First field must be the file pointer in all the
393 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
395 struct io_poll_iocb {
398 struct wait_queue_head *head;
404 struct wait_queue_entry wait;
409 struct file *put_file;
413 struct io_timeout_data {
414 struct io_kiocb *req;
415 struct hrtimer timer;
416 struct timespec64 ts;
417 enum hrtimer_mode mode;
422 struct sockaddr __user *addr;
423 int __user *addr_len;
425 unsigned long nofile;
445 struct list_head list;
448 struct io_timeout_rem {
454 /* NOTE: kiocb has the file as the first member, so don't do it here */
462 struct sockaddr __user *addr;
469 struct user_msghdr __user *umsg;
475 struct io_buffer *kbuf;
481 bool ignore_nonblock;
482 struct filename *filename;
484 unsigned long nofile;
487 struct io_files_update {
513 struct epoll_event event;
517 struct file *file_out;
518 struct file *file_in;
525 struct io_provide_buf {
539 const char __user *filename;
540 struct statx __user *buffer;
543 struct io_completion {
545 struct list_head list;
549 struct io_async_connect {
550 struct sockaddr_storage address;
553 struct io_async_msghdr {
554 struct iovec fast_iov[UIO_FASTIOV];
556 struct sockaddr __user *uaddr;
558 struct sockaddr_storage addr;
562 struct iovec fast_iov[UIO_FASTIOV];
563 const struct iovec *free_iovec;
564 struct iov_iter iter;
566 struct wait_page_queue wpq;
570 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
571 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
572 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
573 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
574 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
575 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
582 REQ_F_LINK_TIMEOUT_BIT,
584 REQ_F_NEED_CLEANUP_BIT,
586 REQ_F_BUFFER_SELECTED_BIT,
587 REQ_F_NO_FILE_TABLE_BIT,
588 REQ_F_WORK_INITIALIZED_BIT,
589 REQ_F_LTIMEOUT_ACTIVE_BIT,
591 /* not a real bit, just to check we're not overflowing the space */
597 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
598 /* drain existing IO first */
599 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
601 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
602 /* doesn't sever on completion < 0 */
603 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
605 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
606 /* IOSQE_BUFFER_SELECT */
607 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
610 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
611 /* fail rest of links */
612 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
613 /* on inflight list */
614 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
615 /* read/write uses file position */
616 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
617 /* must not punt to workers */
618 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
619 /* has or had linked timeout */
620 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
622 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
624 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
625 /* already went through poll handler */
626 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
627 /* buffer already selected */
628 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
629 /* doesn't need file table for this request */
630 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
631 /* io_wq_work is initialized */
632 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
633 /* linked timeout is active, i.e. prepared by link's head */
634 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
638 struct io_poll_iocb poll;
639 struct io_poll_iocb *double_poll;
643 * NOTE! Each of the iocb union members has the file pointer
644 * as the first entry in their struct definition. So you can
645 * access the file pointer through any of the sub-structs,
646 * or directly as just 'ki_filp' in this struct.
652 struct io_poll_iocb poll;
653 struct io_accept accept;
655 struct io_cancel cancel;
656 struct io_timeout timeout;
657 struct io_timeout_rem timeout_rem;
658 struct io_connect connect;
659 struct io_sr_msg sr_msg;
661 struct io_close close;
662 struct io_files_update files_update;
663 struct io_fadvise fadvise;
664 struct io_madvise madvise;
665 struct io_epoll epoll;
666 struct io_splice splice;
667 struct io_provide_buf pbuf;
668 struct io_statx statx;
669 /* use only after cleaning per-op data, see io_clean_op() */
670 struct io_completion compl;
673 /* opcode allocated if it needs to store data for async defer */
676 /* polled IO has completed */
682 struct io_ring_ctx *ctx;
685 struct task_struct *task;
688 struct list_head link_list;
691 * 1. used with ctx->iopoll_list with reads/writes
692 * 2. to track reqs with ->files (see io_op_def::file_table)
694 struct list_head inflight_entry;
696 struct percpu_ref *fixed_file_refs;
697 struct callback_head task_work;
698 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
699 struct hlist_node hash_node;
700 struct async_poll *apoll;
701 struct io_wq_work work;
704 struct io_defer_entry {
705 struct list_head list;
706 struct io_kiocb *req;
710 #define IO_IOPOLL_BATCH 8
712 struct io_comp_state {
714 struct list_head list;
715 struct io_ring_ctx *ctx;
718 struct io_submit_state {
719 struct blk_plug plug;
722 * io_kiocb alloc cache
724 void *reqs[IO_IOPOLL_BATCH];
725 unsigned int free_reqs;
728 * Batch completion logic
730 struct io_comp_state comp;
733 * File reference cache
737 unsigned int has_refs;
738 unsigned int ios_left;
742 /* needs req->file assigned */
743 unsigned needs_file : 1;
744 /* don't fail if file grab fails */
745 unsigned needs_file_no_error : 1;
746 /* hash wq insertion if file is a regular file */
747 unsigned hash_reg_file : 1;
748 /* unbound wq insertion if file is a non-regular file */
749 unsigned unbound_nonreg_file : 1;
750 /* opcode is not supported by this kernel */
751 unsigned not_supported : 1;
752 /* set if opcode supports polled "wait" */
754 unsigned pollout : 1;
755 /* op supports buffer selection */
756 unsigned buffer_select : 1;
757 /* must always have async data allocated */
758 unsigned needs_async_data : 1;
759 /* size of async data needed, if any */
760 unsigned short async_size;
764 static const struct io_op_def io_op_defs[] = {
765 [IORING_OP_NOP] = {},
766 [IORING_OP_READV] = {
768 .unbound_nonreg_file = 1,
771 .needs_async_data = 1,
772 .async_size = sizeof(struct io_async_rw),
773 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
775 [IORING_OP_WRITEV] = {
778 .unbound_nonreg_file = 1,
780 .needs_async_data = 1,
781 .async_size = sizeof(struct io_async_rw),
782 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
785 [IORING_OP_FSYNC] = {
787 .work_flags = IO_WQ_WORK_BLKCG,
789 [IORING_OP_READ_FIXED] = {
791 .unbound_nonreg_file = 1,
793 .async_size = sizeof(struct io_async_rw),
794 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_MM,
796 [IORING_OP_WRITE_FIXED] = {
799 .unbound_nonreg_file = 1,
801 .async_size = sizeof(struct io_async_rw),
802 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE |
805 [IORING_OP_POLL_ADD] = {
807 .unbound_nonreg_file = 1,
809 [IORING_OP_POLL_REMOVE] = {},
810 [IORING_OP_SYNC_FILE_RANGE] = {
812 .work_flags = IO_WQ_WORK_BLKCG,
814 [IORING_OP_SENDMSG] = {
816 .unbound_nonreg_file = 1,
818 .needs_async_data = 1,
819 .async_size = sizeof(struct io_async_msghdr),
820 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
823 [IORING_OP_RECVMSG] = {
825 .unbound_nonreg_file = 1,
828 .needs_async_data = 1,
829 .async_size = sizeof(struct io_async_msghdr),
830 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
833 [IORING_OP_TIMEOUT] = {
834 .needs_async_data = 1,
835 .async_size = sizeof(struct io_timeout_data),
836 .work_flags = IO_WQ_WORK_MM,
838 [IORING_OP_TIMEOUT_REMOVE] = {},
839 [IORING_OP_ACCEPT] = {
841 .unbound_nonreg_file = 1,
843 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES,
845 [IORING_OP_ASYNC_CANCEL] = {},
846 [IORING_OP_LINK_TIMEOUT] = {
847 .needs_async_data = 1,
848 .async_size = sizeof(struct io_timeout_data),
849 .work_flags = IO_WQ_WORK_MM,
851 [IORING_OP_CONNECT] = {
853 .unbound_nonreg_file = 1,
855 .needs_async_data = 1,
856 .async_size = sizeof(struct io_async_connect),
857 .work_flags = IO_WQ_WORK_MM,
859 [IORING_OP_FALLOCATE] = {
861 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE,
863 [IORING_OP_OPENAT] = {
864 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG |
867 [IORING_OP_CLOSE] = {
869 .needs_file_no_error = 1,
870 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG,
872 [IORING_OP_FILES_UPDATE] = {
873 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM,
875 [IORING_OP_STATX] = {
876 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM |
877 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
881 .unbound_nonreg_file = 1,
884 .async_size = sizeof(struct io_async_rw),
885 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
887 [IORING_OP_WRITE] = {
889 .unbound_nonreg_file = 1,
891 .async_size = sizeof(struct io_async_rw),
892 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
895 [IORING_OP_FADVISE] = {
897 .work_flags = IO_WQ_WORK_BLKCG,
899 [IORING_OP_MADVISE] = {
900 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
904 .unbound_nonreg_file = 1,
906 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
910 .unbound_nonreg_file = 1,
913 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
915 [IORING_OP_OPENAT2] = {
916 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_FS |
919 [IORING_OP_EPOLL_CTL] = {
920 .unbound_nonreg_file = 1,
921 .work_flags = IO_WQ_WORK_FILES,
923 [IORING_OP_SPLICE] = {
926 .unbound_nonreg_file = 1,
927 .work_flags = IO_WQ_WORK_BLKCG,
929 [IORING_OP_PROVIDE_BUFFERS] = {},
930 [IORING_OP_REMOVE_BUFFERS] = {},
934 .unbound_nonreg_file = 1,
938 enum io_mem_account {
943 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
944 struct io_comp_state *cs);
945 static void io_cqring_fill_event(struct io_kiocb *req, long res);
946 static void io_put_req(struct io_kiocb *req);
947 static void io_put_req_deferred(struct io_kiocb *req, int nr);
948 static void io_double_put_req(struct io_kiocb *req);
949 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
950 static void __io_queue_linked_timeout(struct io_kiocb *req);
951 static void io_queue_linked_timeout(struct io_kiocb *req);
952 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
953 struct io_uring_files_update *ip,
955 static void __io_clean_op(struct io_kiocb *req);
956 static struct file *io_file_get(struct io_submit_state *state,
957 struct io_kiocb *req, int fd, bool fixed);
958 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs);
959 static void io_file_put_work(struct work_struct *work);
961 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
962 struct iovec **iovec, struct iov_iter *iter,
964 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
965 const struct iovec *fast_iov,
966 struct iov_iter *iter, bool force);
968 static struct kmem_cache *req_cachep;
970 static const struct file_operations io_uring_fops;
972 struct sock *io_uring_get_socket(struct file *file)
974 #if defined(CONFIG_UNIX)
975 if (file->f_op == &io_uring_fops) {
976 struct io_ring_ctx *ctx = file->private_data;
978 return ctx->ring_sock->sk;
983 EXPORT_SYMBOL(io_uring_get_socket);
985 static inline void io_clean_op(struct io_kiocb *req)
987 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
992 static void io_sq_thread_drop_mm(void)
994 struct mm_struct *mm = current->mm;
997 kthread_unuse_mm(mm);
1003 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
1005 struct mm_struct *mm;
1010 /* Should never happen */
1011 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL)))
1014 task_lock(ctx->sqo_task);
1015 mm = ctx->sqo_task->mm;
1016 if (unlikely(!mm || !mmget_not_zero(mm)))
1018 task_unlock(ctx->sqo_task);
1028 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
1029 struct io_kiocb *req)
1031 if (!(io_op_defs[req->opcode].work_flags & IO_WQ_WORK_MM))
1033 return __io_sq_thread_acquire_mm(ctx);
1036 static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx,
1037 struct cgroup_subsys_state **cur_css)
1040 #ifdef CONFIG_BLK_CGROUP
1041 /* puts the old one when swapping */
1042 if (*cur_css != ctx->sqo_blkcg_css) {
1043 kthread_associate_blkcg(ctx->sqo_blkcg_css);
1044 *cur_css = ctx->sqo_blkcg_css;
1049 static void io_sq_thread_unassociate_blkcg(void)
1051 #ifdef CONFIG_BLK_CGROUP
1052 kthread_associate_blkcg(NULL);
1056 static inline void req_set_fail_links(struct io_kiocb *req)
1058 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1059 req->flags |= REQ_F_FAIL_LINK;
1063 * None of these are dereferenced, they are simply used to check if any of
1064 * them have changed. If we're under current and check they are still the
1065 * same, we're fine to grab references to them for actual out-of-line use.
1067 static void io_init_identity(struct io_identity *id)
1069 id->files = current->files;
1070 id->mm = current->mm;
1071 #ifdef CONFIG_BLK_CGROUP
1073 id->blkcg_css = blkcg_css();
1076 id->creds = current_cred();
1077 id->nsproxy = current->nsproxy;
1078 id->fs = current->fs;
1079 id->fsize = rlimit(RLIMIT_FSIZE);
1081 id->loginuid = current->loginuid;
1082 id->sessionid = current->sessionid;
1084 refcount_set(&id->count, 1);
1087 static inline void __io_req_init_async(struct io_kiocb *req)
1089 memset(&req->work, 0, sizeof(req->work));
1090 req->flags |= REQ_F_WORK_INITIALIZED;
1094 * Note: must call io_req_init_async() for the first time you
1095 * touch any members of io_wq_work.
1097 static inline void io_req_init_async(struct io_kiocb *req)
1099 struct io_uring_task *tctx = current->io_uring;
1101 if (req->flags & REQ_F_WORK_INITIALIZED)
1104 __io_req_init_async(req);
1106 /* Grab a ref if this isn't our static identity */
1107 req->work.identity = tctx->identity;
1108 if (tctx->identity != &tctx->__identity)
1109 refcount_inc(&req->work.identity->count);
1112 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1114 return ctx->flags & IORING_SETUP_SQPOLL;
1117 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1119 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1121 complete(&ctx->ref_comp);
1124 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1126 return !req->timeout.off;
1129 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1131 struct io_ring_ctx *ctx;
1134 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1138 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1139 if (!ctx->fallback_req)
1143 * Use 5 bits less than the max cq entries, that should give us around
1144 * 32 entries per hash list if totally full and uniformly spread.
1146 hash_bits = ilog2(p->cq_entries);
1150 ctx->cancel_hash_bits = hash_bits;
1151 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1153 if (!ctx->cancel_hash)
1155 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1157 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1158 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1161 ctx->flags = p->flags;
1162 init_waitqueue_head(&ctx->sqo_sq_wait);
1163 INIT_LIST_HEAD(&ctx->sqd_list);
1164 init_waitqueue_head(&ctx->cq_wait);
1165 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1166 init_completion(&ctx->ref_comp);
1167 init_completion(&ctx->sq_thread_comp);
1168 idr_init(&ctx->io_buffer_idr);
1169 idr_init(&ctx->personality_idr);
1170 mutex_init(&ctx->uring_lock);
1171 init_waitqueue_head(&ctx->wait);
1172 spin_lock_init(&ctx->completion_lock);
1173 INIT_LIST_HEAD(&ctx->iopoll_list);
1174 INIT_LIST_HEAD(&ctx->defer_list);
1175 INIT_LIST_HEAD(&ctx->timeout_list);
1176 init_waitqueue_head(&ctx->inflight_wait);
1177 spin_lock_init(&ctx->inflight_lock);
1178 INIT_LIST_HEAD(&ctx->inflight_list);
1179 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1180 init_llist_head(&ctx->file_put_llist);
1183 if (ctx->fallback_req)
1184 kmem_cache_free(req_cachep, ctx->fallback_req);
1185 kfree(ctx->cancel_hash);
1190 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1192 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1193 struct io_ring_ctx *ctx = req->ctx;
1195 return seq != ctx->cached_cq_tail
1196 + READ_ONCE(ctx->cached_cq_overflow);
1202 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1204 struct io_rings *rings = ctx->rings;
1206 /* order cqe stores with ring update */
1207 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1209 if (wq_has_sleeper(&ctx->cq_wait)) {
1210 wake_up_interruptible(&ctx->cq_wait);
1211 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1215 static void io_put_identity(struct io_uring_task *tctx, struct io_kiocb *req)
1217 if (req->work.identity == &tctx->__identity)
1219 if (refcount_dec_and_test(&req->work.identity->count))
1220 kfree(req->work.identity);
1223 static void io_req_clean_work(struct io_kiocb *req)
1225 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1228 req->flags &= ~REQ_F_WORK_INITIALIZED;
1230 if (req->work.flags & IO_WQ_WORK_MM) {
1231 mmdrop(req->work.identity->mm);
1232 req->work.flags &= ~IO_WQ_WORK_MM;
1234 #ifdef CONFIG_BLK_CGROUP
1235 if (req->work.flags & IO_WQ_WORK_BLKCG) {
1236 css_put(req->work.identity->blkcg_css);
1237 req->work.flags &= ~IO_WQ_WORK_BLKCG;
1240 if (req->work.flags & IO_WQ_WORK_CREDS) {
1241 put_cred(req->work.identity->creds);
1242 req->work.flags &= ~IO_WQ_WORK_CREDS;
1244 if (req->work.flags & IO_WQ_WORK_FS) {
1245 struct fs_struct *fs = req->work.identity->fs;
1247 spin_lock(&req->work.identity->fs->lock);
1250 spin_unlock(&req->work.identity->fs->lock);
1253 req->work.flags &= ~IO_WQ_WORK_FS;
1256 io_put_identity(req->task->io_uring, req);
1260 * Create a private copy of io_identity, since some fields don't match
1261 * the current context.
1263 static bool io_identity_cow(struct io_kiocb *req)
1265 struct io_uring_task *tctx = current->io_uring;
1266 const struct cred *creds = NULL;
1267 struct io_identity *id;
1269 if (req->work.flags & IO_WQ_WORK_CREDS)
1270 creds = req->work.identity->creds;
1272 id = kmemdup(req->work.identity, sizeof(*id), GFP_KERNEL);
1273 if (unlikely(!id)) {
1274 req->work.flags |= IO_WQ_WORK_CANCEL;
1279 * We can safely just re-init the creds we copied Either the field
1280 * matches the current one, or we haven't grabbed it yet. The only
1281 * exception is ->creds, through registered personalities, so handle
1282 * that one separately.
1284 io_init_identity(id);
1286 req->work.identity->creds = creds;
1288 /* add one for this request */
1289 refcount_inc(&id->count);
1291 /* drop tctx and req identity references, if needed */
1292 if (tctx->identity != &tctx->__identity &&
1293 refcount_dec_and_test(&tctx->identity->count))
1294 kfree(tctx->identity);
1295 if (req->work.identity != &tctx->__identity &&
1296 refcount_dec_and_test(&req->work.identity->count))
1297 kfree(req->work.identity);
1299 req->work.identity = id;
1300 tctx->identity = id;
1304 static bool io_grab_identity(struct io_kiocb *req)
1306 const struct io_op_def *def = &io_op_defs[req->opcode];
1307 struct io_identity *id = req->work.identity;
1308 struct io_ring_ctx *ctx = req->ctx;
1310 if (def->work_flags & IO_WQ_WORK_FSIZE) {
1311 if (id->fsize != rlimit(RLIMIT_FSIZE))
1313 req->work.flags |= IO_WQ_WORK_FSIZE;
1316 if (!(req->work.flags & IO_WQ_WORK_FILES) &&
1317 (def->work_flags & IO_WQ_WORK_FILES) &&
1318 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1319 if (id->files != current->files ||
1320 id->nsproxy != current->nsproxy)
1322 atomic_inc(&id->files->count);
1323 get_nsproxy(id->nsproxy);
1324 req->flags |= REQ_F_INFLIGHT;
1326 spin_lock_irq(&ctx->inflight_lock);
1327 list_add(&req->inflight_entry, &ctx->inflight_list);
1328 spin_unlock_irq(&ctx->inflight_lock);
1329 req->work.flags |= IO_WQ_WORK_FILES;
1331 #ifdef CONFIG_BLK_CGROUP
1332 if (!(req->work.flags & IO_WQ_WORK_BLKCG) &&
1333 (def->work_flags & IO_WQ_WORK_BLKCG)) {
1335 if (id->blkcg_css != blkcg_css()) {
1340 * This should be rare, either the cgroup is dying or the task
1341 * is moving cgroups. Just punt to root for the handful of ios.
1343 if (css_tryget_online(id->blkcg_css))
1344 req->work.flags |= IO_WQ_WORK_BLKCG;
1348 if (!(req->work.flags & IO_WQ_WORK_CREDS)) {
1349 if (id->creds != current_cred())
1351 get_cred(id->creds);
1352 req->work.flags |= IO_WQ_WORK_CREDS;
1355 if (!uid_eq(current->loginuid, id->loginuid) ||
1356 current->sessionid != id->sessionid)
1359 if (!(req->work.flags & IO_WQ_WORK_FS) &&
1360 (def->work_flags & IO_WQ_WORK_FS)) {
1361 if (current->fs != id->fs)
1363 spin_lock(&id->fs->lock);
1364 if (!id->fs->in_exec) {
1366 req->work.flags |= IO_WQ_WORK_FS;
1368 req->work.flags |= IO_WQ_WORK_CANCEL;
1370 spin_unlock(¤t->fs->lock);
1376 static void io_prep_async_work(struct io_kiocb *req)
1378 const struct io_op_def *def = &io_op_defs[req->opcode];
1379 struct io_ring_ctx *ctx = req->ctx;
1380 struct io_identity *id;
1382 io_req_init_async(req);
1383 id = req->work.identity;
1385 if (req->flags & REQ_F_FORCE_ASYNC)
1386 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1388 if (req->flags & REQ_F_ISREG) {
1389 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1390 io_wq_hash_work(&req->work, file_inode(req->file));
1392 if (def->unbound_nonreg_file)
1393 req->work.flags |= IO_WQ_WORK_UNBOUND;
1396 /* ->mm can never change on us */
1397 if (!(req->work.flags & IO_WQ_WORK_MM) &&
1398 (def->work_flags & IO_WQ_WORK_MM)) {
1400 req->work.flags |= IO_WQ_WORK_MM;
1403 /* if we fail grabbing identity, we must COW, regrab, and retry */
1404 if (io_grab_identity(req))
1407 if (!io_identity_cow(req))
1410 /* can't fail at this point */
1411 if (!io_grab_identity(req))
1415 static void io_prep_async_link(struct io_kiocb *req)
1417 struct io_kiocb *cur;
1419 io_prep_async_work(req);
1420 if (req->flags & REQ_F_LINK_HEAD)
1421 list_for_each_entry(cur, &req->link_list, link_list)
1422 io_prep_async_work(cur);
1425 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1427 struct io_ring_ctx *ctx = req->ctx;
1428 struct io_kiocb *link = io_prep_linked_timeout(req);
1430 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1431 &req->work, req->flags);
1432 io_wq_enqueue(ctx->io_wq, &req->work);
1436 static void io_queue_async_work(struct io_kiocb *req)
1438 struct io_kiocb *link;
1440 /* init ->work of the whole link before punting */
1441 io_prep_async_link(req);
1442 link = __io_queue_async_work(req);
1445 io_queue_linked_timeout(link);
1448 static void io_kill_timeout(struct io_kiocb *req)
1450 struct io_timeout_data *io = req->async_data;
1453 ret = hrtimer_try_to_cancel(&io->timer);
1455 atomic_set(&req->ctx->cq_timeouts,
1456 atomic_read(&req->ctx->cq_timeouts) + 1);
1457 list_del_init(&req->timeout.list);
1458 io_cqring_fill_event(req, 0);
1459 io_put_req_deferred(req, 1);
1463 static bool io_task_match(struct io_kiocb *req, struct task_struct *tsk)
1465 struct io_ring_ctx *ctx = req->ctx;
1467 if (!tsk || req->task == tsk)
1469 if (ctx->flags & IORING_SETUP_SQPOLL) {
1470 if (ctx->sq_data && req->task == ctx->sq_data->thread)
1477 * Returns true if we found and killed one or more timeouts
1479 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk)
1481 struct io_kiocb *req, *tmp;
1484 spin_lock_irq(&ctx->completion_lock);
1485 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1486 if (io_task_match(req, tsk)) {
1487 io_kill_timeout(req);
1491 spin_unlock_irq(&ctx->completion_lock);
1492 return canceled != 0;
1495 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1498 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1499 struct io_defer_entry, list);
1500 struct io_kiocb *link;
1502 if (req_need_defer(de->req, de->seq))
1504 list_del_init(&de->list);
1505 /* punt-init is done before queueing for defer */
1506 link = __io_queue_async_work(de->req);
1508 __io_queue_linked_timeout(link);
1509 /* drop submission reference */
1510 io_put_req_deferred(link, 1);
1513 } while (!list_empty(&ctx->defer_list));
1516 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1518 while (!list_empty(&ctx->timeout_list)) {
1519 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1520 struct io_kiocb, timeout.list);
1522 if (io_is_timeout_noseq(req))
1524 if (req->timeout.target_seq != ctx->cached_cq_tail
1525 - atomic_read(&ctx->cq_timeouts))
1528 list_del_init(&req->timeout.list);
1529 io_kill_timeout(req);
1533 static void io_commit_cqring(struct io_ring_ctx *ctx)
1535 io_flush_timeouts(ctx);
1536 __io_commit_cqring(ctx);
1538 if (unlikely(!list_empty(&ctx->defer_list)))
1539 __io_queue_deferred(ctx);
1542 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1544 struct io_rings *r = ctx->rings;
1546 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1549 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1551 struct io_rings *rings = ctx->rings;
1554 tail = ctx->cached_cq_tail;
1556 * writes to the cq entry need to come after reading head; the
1557 * control dependency is enough as we're using WRITE_ONCE to
1560 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1563 ctx->cached_cq_tail++;
1564 return &rings->cqes[tail & ctx->cq_mask];
1567 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1571 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1573 if (!ctx->eventfd_async)
1575 return io_wq_current_is_worker();
1578 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1580 if (waitqueue_active(&ctx->wait))
1581 wake_up(&ctx->wait);
1582 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1583 wake_up(&ctx->sq_data->wait);
1584 if (io_should_trigger_evfd(ctx))
1585 eventfd_signal(ctx->cq_ev_fd, 1);
1588 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1590 if (list_empty(&ctx->cq_overflow_list)) {
1591 clear_bit(0, &ctx->sq_check_overflow);
1592 clear_bit(0, &ctx->cq_check_overflow);
1593 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1597 static inline bool __io_match_files(struct io_kiocb *req,
1598 struct files_struct *files)
1600 return ((req->flags & REQ_F_WORK_INITIALIZED) &&
1601 (req->work.flags & IO_WQ_WORK_FILES)) &&
1602 req->work.identity->files == files;
1605 static bool io_match_files(struct io_kiocb *req,
1606 struct files_struct *files)
1608 struct io_kiocb *link;
1612 if (__io_match_files(req, files))
1614 if (req->flags & REQ_F_LINK_HEAD) {
1615 list_for_each_entry(link, &req->link_list, link_list) {
1616 if (__io_match_files(link, files))
1623 /* Returns true if there are no backlogged entries after the flush */
1624 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1625 struct task_struct *tsk,
1626 struct files_struct *files)
1628 struct io_rings *rings = ctx->rings;
1629 struct io_kiocb *req, *tmp;
1630 struct io_uring_cqe *cqe;
1631 unsigned long flags;
1635 if (list_empty_careful(&ctx->cq_overflow_list))
1637 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1638 rings->cq_ring_entries))
1642 spin_lock_irqsave(&ctx->completion_lock, flags);
1644 /* if force is set, the ring is going away. always drop after that */
1646 ctx->cq_overflow_flushed = 1;
1649 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1650 if (tsk && req->task != tsk)
1652 if (!io_match_files(req, files))
1655 cqe = io_get_cqring(ctx);
1659 list_move(&req->compl.list, &list);
1661 WRITE_ONCE(cqe->user_data, req->user_data);
1662 WRITE_ONCE(cqe->res, req->result);
1663 WRITE_ONCE(cqe->flags, req->compl.cflags);
1665 ctx->cached_cq_overflow++;
1666 WRITE_ONCE(ctx->rings->cq_overflow,
1667 ctx->cached_cq_overflow);
1671 io_commit_cqring(ctx);
1672 io_cqring_mark_overflow(ctx);
1674 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1675 io_cqring_ev_posted(ctx);
1677 while (!list_empty(&list)) {
1678 req = list_first_entry(&list, struct io_kiocb, compl.list);
1679 list_del(&req->compl.list);
1686 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1688 struct io_ring_ctx *ctx = req->ctx;
1689 struct io_uring_cqe *cqe;
1691 trace_io_uring_complete(ctx, req->user_data, res);
1694 * If we can't get a cq entry, userspace overflowed the
1695 * submission (by quite a lot). Increment the overflow count in
1698 cqe = io_get_cqring(ctx);
1700 WRITE_ONCE(cqe->user_data, req->user_data);
1701 WRITE_ONCE(cqe->res, res);
1702 WRITE_ONCE(cqe->flags, cflags);
1703 } else if (ctx->cq_overflow_flushed ||
1704 atomic_read(&req->task->io_uring->in_idle)) {
1706 * If we're in ring overflow flush mode, or in task cancel mode,
1707 * then we cannot store the request for later flushing, we need
1708 * to drop it on the floor.
1710 ctx->cached_cq_overflow++;
1711 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1713 if (list_empty(&ctx->cq_overflow_list)) {
1714 set_bit(0, &ctx->sq_check_overflow);
1715 set_bit(0, &ctx->cq_check_overflow);
1716 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1720 req->compl.cflags = cflags;
1721 refcount_inc(&req->refs);
1722 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1726 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1728 __io_cqring_fill_event(req, res, 0);
1731 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1733 struct io_ring_ctx *ctx = req->ctx;
1734 unsigned long flags;
1736 spin_lock_irqsave(&ctx->completion_lock, flags);
1737 __io_cqring_fill_event(req, res, cflags);
1738 io_commit_cqring(ctx);
1739 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1741 io_cqring_ev_posted(ctx);
1744 static void io_submit_flush_completions(struct io_comp_state *cs)
1746 struct io_ring_ctx *ctx = cs->ctx;
1748 spin_lock_irq(&ctx->completion_lock);
1749 while (!list_empty(&cs->list)) {
1750 struct io_kiocb *req;
1752 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1753 list_del(&req->compl.list);
1754 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1757 * io_free_req() doesn't care about completion_lock unless one
1758 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1759 * because of a potential deadlock with req->work.fs->lock
1761 if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
1762 |REQ_F_WORK_INITIALIZED)) {
1763 spin_unlock_irq(&ctx->completion_lock);
1765 spin_lock_irq(&ctx->completion_lock);
1770 io_commit_cqring(ctx);
1771 spin_unlock_irq(&ctx->completion_lock);
1773 io_cqring_ev_posted(ctx);
1777 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1778 struct io_comp_state *cs)
1781 io_cqring_add_event(req, res, cflags);
1786 req->compl.cflags = cflags;
1787 list_add_tail(&req->compl.list, &cs->list);
1789 io_submit_flush_completions(cs);
1793 static void io_req_complete(struct io_kiocb *req, long res)
1795 __io_req_complete(req, res, 0, NULL);
1798 static inline bool io_is_fallback_req(struct io_kiocb *req)
1800 return req == (struct io_kiocb *)
1801 ((unsigned long) req->ctx->fallback_req & ~1UL);
1804 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1806 struct io_kiocb *req;
1808 req = ctx->fallback_req;
1809 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1815 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1816 struct io_submit_state *state)
1818 if (!state->free_reqs) {
1819 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1823 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1824 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1827 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1828 * retry single alloc to be on the safe side.
1830 if (unlikely(ret <= 0)) {
1831 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1832 if (!state->reqs[0])
1836 state->free_reqs = ret;
1840 return state->reqs[state->free_reqs];
1842 return io_get_fallback_req(ctx);
1845 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1849 percpu_ref_put(req->fixed_file_refs);
1854 static void io_dismantle_req(struct io_kiocb *req)
1858 if (req->async_data)
1859 kfree(req->async_data);
1861 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1863 io_req_clean_work(req);
1866 static void __io_free_req(struct io_kiocb *req)
1868 struct io_uring_task *tctx = req->task->io_uring;
1869 struct io_ring_ctx *ctx = req->ctx;
1871 io_dismantle_req(req);
1873 percpu_counter_dec(&tctx->inflight);
1874 if (atomic_read(&tctx->in_idle))
1875 wake_up(&tctx->wait);
1876 put_task_struct(req->task);
1878 if (likely(!io_is_fallback_req(req)))
1879 kmem_cache_free(req_cachep, req);
1881 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1882 percpu_ref_put(&ctx->refs);
1885 static void io_kill_linked_timeout(struct io_kiocb *req)
1887 struct io_ring_ctx *ctx = req->ctx;
1888 struct io_kiocb *link;
1889 bool cancelled = false;
1890 unsigned long flags;
1892 spin_lock_irqsave(&ctx->completion_lock, flags);
1893 link = list_first_entry_or_null(&req->link_list, struct io_kiocb,
1896 * Can happen if a linked timeout fired and link had been like
1897 * req -> link t-out -> link t-out [-> ...]
1899 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1900 struct io_timeout_data *io = link->async_data;
1903 list_del_init(&link->link_list);
1904 ret = hrtimer_try_to_cancel(&io->timer);
1906 io_cqring_fill_event(link, -ECANCELED);
1907 io_commit_cqring(ctx);
1911 req->flags &= ~REQ_F_LINK_TIMEOUT;
1912 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1915 io_cqring_ev_posted(ctx);
1920 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1922 struct io_kiocb *nxt;
1925 * The list should never be empty when we are called here. But could
1926 * potentially happen if the chain is messed up, check to be on the
1929 if (unlikely(list_empty(&req->link_list)))
1932 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1933 list_del_init(&req->link_list);
1934 if (!list_empty(&nxt->link_list))
1935 nxt->flags |= REQ_F_LINK_HEAD;
1940 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1942 static void io_fail_links(struct io_kiocb *req)
1944 struct io_ring_ctx *ctx = req->ctx;
1945 unsigned long flags;
1947 spin_lock_irqsave(&ctx->completion_lock, flags);
1948 while (!list_empty(&req->link_list)) {
1949 struct io_kiocb *link = list_first_entry(&req->link_list,
1950 struct io_kiocb, link_list);
1952 list_del_init(&link->link_list);
1953 trace_io_uring_fail_link(req, link);
1955 io_cqring_fill_event(link, -ECANCELED);
1958 * It's ok to free under spinlock as they're not linked anymore,
1959 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
1962 if (link->flags & REQ_F_WORK_INITIALIZED)
1963 io_put_req_deferred(link, 2);
1965 io_double_put_req(link);
1968 io_commit_cqring(ctx);
1969 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1971 io_cqring_ev_posted(ctx);
1974 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1976 req->flags &= ~REQ_F_LINK_HEAD;
1977 if (req->flags & REQ_F_LINK_TIMEOUT)
1978 io_kill_linked_timeout(req);
1981 * If LINK is set, we have dependent requests in this chain. If we
1982 * didn't fail this request, queue the first one up, moving any other
1983 * dependencies to the next request. In case of failure, fail the rest
1986 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1987 return io_req_link_next(req);
1992 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1994 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1996 return __io_req_find_next(req);
1999 static int io_req_task_work_add(struct io_kiocb *req, bool twa_signal_ok)
2001 struct task_struct *tsk = req->task;
2002 struct io_ring_ctx *ctx = req->ctx;
2003 enum task_work_notify_mode notify;
2006 if (tsk->flags & PF_EXITING)
2010 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2011 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2012 * processing task_work. There's no reliable way to tell if TWA_RESUME
2016 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
2017 notify = TWA_SIGNAL;
2019 ret = task_work_add(tsk, &req->task_work, notify);
2021 wake_up_process(tsk);
2026 static void __io_req_task_cancel(struct io_kiocb *req, int error)
2028 struct io_ring_ctx *ctx = req->ctx;
2030 spin_lock_irq(&ctx->completion_lock);
2031 io_cqring_fill_event(req, error);
2032 io_commit_cqring(ctx);
2033 spin_unlock_irq(&ctx->completion_lock);
2035 io_cqring_ev_posted(ctx);
2036 req_set_fail_links(req);
2037 io_double_put_req(req);
2040 static void io_req_task_cancel(struct callback_head *cb)
2042 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2043 struct io_ring_ctx *ctx = req->ctx;
2045 __io_req_task_cancel(req, -ECANCELED);
2046 percpu_ref_put(&ctx->refs);
2049 static void __io_req_task_submit(struct io_kiocb *req)
2051 struct io_ring_ctx *ctx = req->ctx;
2053 if (!__io_sq_thread_acquire_mm(ctx)) {
2054 mutex_lock(&ctx->uring_lock);
2055 __io_queue_sqe(req, NULL);
2056 mutex_unlock(&ctx->uring_lock);
2058 __io_req_task_cancel(req, -EFAULT);
2062 static void io_req_task_submit(struct callback_head *cb)
2064 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2065 struct io_ring_ctx *ctx = req->ctx;
2067 __io_req_task_submit(req);
2068 percpu_ref_put(&ctx->refs);
2071 static void io_req_task_queue(struct io_kiocb *req)
2075 init_task_work(&req->task_work, io_req_task_submit);
2076 percpu_ref_get(&req->ctx->refs);
2078 ret = io_req_task_work_add(req, true);
2079 if (unlikely(ret)) {
2080 struct task_struct *tsk;
2082 init_task_work(&req->task_work, io_req_task_cancel);
2083 tsk = io_wq_get_task(req->ctx->io_wq);
2084 task_work_add(tsk, &req->task_work, TWA_NONE);
2085 wake_up_process(tsk);
2089 static void io_queue_next(struct io_kiocb *req)
2091 struct io_kiocb *nxt = io_req_find_next(req);
2094 io_req_task_queue(nxt);
2097 static void io_free_req(struct io_kiocb *req)
2104 void *reqs[IO_IOPOLL_BATCH];
2107 struct task_struct *task;
2111 static inline void io_init_req_batch(struct req_batch *rb)
2118 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2119 struct req_batch *rb)
2121 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2122 percpu_ref_put_many(&ctx->refs, rb->to_free);
2126 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2127 struct req_batch *rb)
2130 __io_req_free_batch_flush(ctx, rb);
2132 struct io_uring_task *tctx = rb->task->io_uring;
2134 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2135 put_task_struct_many(rb->task, rb->task_refs);
2140 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2142 if (unlikely(io_is_fallback_req(req))) {
2146 if (req->flags & REQ_F_LINK_HEAD)
2149 if (req->task != rb->task) {
2151 struct io_uring_task *tctx = rb->task->io_uring;
2153 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2154 put_task_struct_many(rb->task, rb->task_refs);
2156 rb->task = req->task;
2161 io_dismantle_req(req);
2162 rb->reqs[rb->to_free++] = req;
2163 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2164 __io_req_free_batch_flush(req->ctx, rb);
2168 * Drop reference to request, return next in chain (if there is one) if this
2169 * was the last reference to this request.
2171 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2173 struct io_kiocb *nxt = NULL;
2175 if (refcount_dec_and_test(&req->refs)) {
2176 nxt = io_req_find_next(req);
2182 static void io_put_req(struct io_kiocb *req)
2184 if (refcount_dec_and_test(&req->refs))
2188 static void io_put_req_deferred_cb(struct callback_head *cb)
2190 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2195 static void io_free_req_deferred(struct io_kiocb *req)
2199 init_task_work(&req->task_work, io_put_req_deferred_cb);
2200 ret = io_req_task_work_add(req, true);
2201 if (unlikely(ret)) {
2202 struct task_struct *tsk;
2204 tsk = io_wq_get_task(req->ctx->io_wq);
2205 task_work_add(tsk, &req->task_work, TWA_NONE);
2206 wake_up_process(tsk);
2210 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2212 if (refcount_sub_and_test(refs, &req->refs))
2213 io_free_req_deferred(req);
2216 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2218 struct io_kiocb *nxt;
2221 * A ref is owned by io-wq in which context we're. So, if that's the
2222 * last one, it's safe to steal next work. False negatives are Ok,
2223 * it just will be re-punted async in io_put_work()
2225 if (refcount_read(&req->refs) != 1)
2228 nxt = io_req_find_next(req);
2229 return nxt ? &nxt->work : NULL;
2232 static void io_double_put_req(struct io_kiocb *req)
2234 /* drop both submit and complete references */
2235 if (refcount_sub_and_test(2, &req->refs))
2239 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2241 struct io_rings *rings = ctx->rings;
2243 if (test_bit(0, &ctx->cq_check_overflow)) {
2245 * noflush == true is from the waitqueue handler, just ensure
2246 * we wake up the task, and the next invocation will flush the
2247 * entries. We cannot safely to it from here.
2249 if (noflush && !list_empty(&ctx->cq_overflow_list))
2252 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2255 /* See comment at the top of this file */
2257 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2260 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2262 struct io_rings *rings = ctx->rings;
2264 /* make sure SQ entry isn't read before tail */
2265 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2268 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2270 unsigned int cflags;
2272 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2273 cflags |= IORING_CQE_F_BUFFER;
2274 req->flags &= ~REQ_F_BUFFER_SELECTED;
2279 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2281 struct io_buffer *kbuf;
2283 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2284 return io_put_kbuf(req, kbuf);
2287 static inline bool io_run_task_work(void)
2290 * Not safe to run on exiting task, and the task_work handling will
2291 * not add work to such a task.
2293 if (unlikely(current->flags & PF_EXITING))
2295 if (current->task_works) {
2296 __set_current_state(TASK_RUNNING);
2304 static void io_iopoll_queue(struct list_head *again)
2306 struct io_kiocb *req;
2309 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2310 list_del(&req->inflight_entry);
2311 __io_complete_rw(req, -EAGAIN, 0, NULL);
2312 } while (!list_empty(again));
2316 * Find and free completed poll iocbs
2318 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2319 struct list_head *done)
2321 struct req_batch rb;
2322 struct io_kiocb *req;
2325 /* order with ->result store in io_complete_rw_iopoll() */
2328 io_init_req_batch(&rb);
2329 while (!list_empty(done)) {
2332 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2333 if (READ_ONCE(req->result) == -EAGAIN) {
2335 req->iopoll_completed = 0;
2336 list_move_tail(&req->inflight_entry, &again);
2339 list_del(&req->inflight_entry);
2341 if (req->flags & REQ_F_BUFFER_SELECTED)
2342 cflags = io_put_rw_kbuf(req);
2344 __io_cqring_fill_event(req, req->result, cflags);
2347 if (refcount_dec_and_test(&req->refs))
2348 io_req_free_batch(&rb, req);
2351 io_commit_cqring(ctx);
2352 if (ctx->flags & IORING_SETUP_SQPOLL)
2353 io_cqring_ev_posted(ctx);
2354 io_req_free_batch_finish(ctx, &rb);
2356 if (!list_empty(&again))
2357 io_iopoll_queue(&again);
2360 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2363 struct io_kiocb *req, *tmp;
2369 * Only spin for completions if we don't have multiple devices hanging
2370 * off our complete list, and we're under the requested amount.
2372 spin = !ctx->poll_multi_file && *nr_events < min;
2375 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2376 struct kiocb *kiocb = &req->rw.kiocb;
2379 * Move completed and retryable entries to our local lists.
2380 * If we find a request that requires polling, break out
2381 * and complete those lists first, if we have entries there.
2383 if (READ_ONCE(req->iopoll_completed)) {
2384 list_move_tail(&req->inflight_entry, &done);
2387 if (!list_empty(&done))
2390 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2394 /* iopoll may have completed current req */
2395 if (READ_ONCE(req->iopoll_completed))
2396 list_move_tail(&req->inflight_entry, &done);
2403 if (!list_empty(&done))
2404 io_iopoll_complete(ctx, nr_events, &done);
2410 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2411 * non-spinning poll check - we'll still enter the driver poll loop, but only
2412 * as a non-spinning completion check.
2414 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2417 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2420 ret = io_do_iopoll(ctx, nr_events, min);
2423 if (*nr_events >= min)
2431 * We can't just wait for polled events to come to us, we have to actively
2432 * find and complete them.
2434 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2436 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2439 mutex_lock(&ctx->uring_lock);
2440 while (!list_empty(&ctx->iopoll_list)) {
2441 unsigned int nr_events = 0;
2443 io_do_iopoll(ctx, &nr_events, 0);
2445 /* let it sleep and repeat later if can't complete a request */
2449 * Ensure we allow local-to-the-cpu processing to take place,
2450 * in this case we need to ensure that we reap all events.
2451 * Also let task_work, etc. to progress by releasing the mutex
2453 if (need_resched()) {
2454 mutex_unlock(&ctx->uring_lock);
2456 mutex_lock(&ctx->uring_lock);
2459 mutex_unlock(&ctx->uring_lock);
2462 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2464 unsigned int nr_events = 0;
2465 int iters = 0, ret = 0;
2468 * We disallow the app entering submit/complete with polling, but we
2469 * still need to lock the ring to prevent racing with polled issue
2470 * that got punted to a workqueue.
2472 mutex_lock(&ctx->uring_lock);
2475 * Don't enter poll loop if we already have events pending.
2476 * If we do, we can potentially be spinning for commands that
2477 * already triggered a CQE (eg in error).
2479 if (io_cqring_events(ctx, false))
2483 * If a submit got punted to a workqueue, we can have the
2484 * application entering polling for a command before it gets
2485 * issued. That app will hold the uring_lock for the duration
2486 * of the poll right here, so we need to take a breather every
2487 * now and then to ensure that the issue has a chance to add
2488 * the poll to the issued list. Otherwise we can spin here
2489 * forever, while the workqueue is stuck trying to acquire the
2492 if (!(++iters & 7)) {
2493 mutex_unlock(&ctx->uring_lock);
2495 mutex_lock(&ctx->uring_lock);
2498 ret = io_iopoll_getevents(ctx, &nr_events, min);
2502 } while (min && !nr_events && !need_resched());
2504 mutex_unlock(&ctx->uring_lock);
2508 static void kiocb_end_write(struct io_kiocb *req)
2511 * Tell lockdep we inherited freeze protection from submission
2514 if (req->flags & REQ_F_ISREG) {
2515 struct inode *inode = file_inode(req->file);
2517 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2519 file_end_write(req->file);
2522 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2523 struct io_comp_state *cs)
2525 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2528 if (kiocb->ki_flags & IOCB_WRITE)
2529 kiocb_end_write(req);
2531 if (res != req->result)
2532 req_set_fail_links(req);
2533 if (req->flags & REQ_F_BUFFER_SELECTED)
2534 cflags = io_put_rw_kbuf(req);
2535 __io_req_complete(req, res, cflags, cs);
2539 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2541 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2542 ssize_t ret = -ECANCELED;
2543 struct iov_iter iter;
2551 switch (req->opcode) {
2552 case IORING_OP_READV:
2553 case IORING_OP_READ_FIXED:
2554 case IORING_OP_READ:
2557 case IORING_OP_WRITEV:
2558 case IORING_OP_WRITE_FIXED:
2559 case IORING_OP_WRITE:
2563 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2568 if (!req->async_data) {
2569 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2572 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2580 req_set_fail_links(req);
2585 static bool io_rw_reissue(struct io_kiocb *req, long res)
2588 umode_t mode = file_inode(req->file)->i_mode;
2591 if (!S_ISBLK(mode) && !S_ISREG(mode))
2593 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2596 ret = io_sq_thread_acquire_mm(req->ctx, req);
2598 if (io_resubmit_prep(req, ret)) {
2599 refcount_inc(&req->refs);
2600 io_queue_async_work(req);
2608 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2609 struct io_comp_state *cs)
2611 if (!io_rw_reissue(req, res))
2612 io_complete_rw_common(&req->rw.kiocb, res, cs);
2615 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2617 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2619 __io_complete_rw(req, res, res2, NULL);
2622 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2624 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2626 if (kiocb->ki_flags & IOCB_WRITE)
2627 kiocb_end_write(req);
2629 if (res != -EAGAIN && res != req->result)
2630 req_set_fail_links(req);
2632 WRITE_ONCE(req->result, res);
2633 /* order with io_poll_complete() checking ->result */
2635 WRITE_ONCE(req->iopoll_completed, 1);
2639 * After the iocb has been issued, it's safe to be found on the poll list.
2640 * Adding the kiocb to the list AFTER submission ensures that we don't
2641 * find it from a io_iopoll_getevents() thread before the issuer is done
2642 * accessing the kiocb cookie.
2644 static void io_iopoll_req_issued(struct io_kiocb *req)
2646 struct io_ring_ctx *ctx = req->ctx;
2649 * Track whether we have multiple files in our lists. This will impact
2650 * how we do polling eventually, not spinning if we're on potentially
2651 * different devices.
2653 if (list_empty(&ctx->iopoll_list)) {
2654 ctx->poll_multi_file = false;
2655 } else if (!ctx->poll_multi_file) {
2656 struct io_kiocb *list_req;
2658 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2660 if (list_req->file != req->file)
2661 ctx->poll_multi_file = true;
2665 * For fast devices, IO may have already completed. If it has, add
2666 * it to the front so we find it first.
2668 if (READ_ONCE(req->iopoll_completed))
2669 list_add(&req->inflight_entry, &ctx->iopoll_list);
2671 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2673 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2674 wq_has_sleeper(&ctx->sq_data->wait))
2675 wake_up(&ctx->sq_data->wait);
2678 static void __io_state_file_put(struct io_submit_state *state)
2680 if (state->has_refs)
2681 fput_many(state->file, state->has_refs);
2685 static inline void io_state_file_put(struct io_submit_state *state)
2688 __io_state_file_put(state);
2692 * Get as many references to a file as we have IOs left in this submission,
2693 * assuming most submissions are for one file, or at least that each file
2694 * has more than one submission.
2696 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2702 if (state->fd == fd) {
2706 __io_state_file_put(state);
2708 state->file = fget_many(fd, state->ios_left);
2713 state->has_refs = state->ios_left - 1;
2717 static bool io_bdev_nowait(struct block_device *bdev)
2720 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2727 * If we tracked the file through the SCM inflight mechanism, we could support
2728 * any file. For now, just ensure that anything potentially problematic is done
2731 static bool io_file_supports_async(struct file *file, int rw)
2733 umode_t mode = file_inode(file)->i_mode;
2735 if (S_ISBLK(mode)) {
2736 if (io_bdev_nowait(file->f_inode->i_bdev))
2740 if (S_ISCHR(mode) || S_ISSOCK(mode))
2742 if (S_ISREG(mode)) {
2743 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2744 file->f_op != &io_uring_fops)
2749 /* any ->read/write should understand O_NONBLOCK */
2750 if (file->f_flags & O_NONBLOCK)
2753 if (!(file->f_mode & FMODE_NOWAIT))
2757 return file->f_op->read_iter != NULL;
2759 return file->f_op->write_iter != NULL;
2762 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2764 struct io_ring_ctx *ctx = req->ctx;
2765 struct kiocb *kiocb = &req->rw.kiocb;
2769 if (S_ISREG(file_inode(req->file)->i_mode))
2770 req->flags |= REQ_F_ISREG;
2772 kiocb->ki_pos = READ_ONCE(sqe->off);
2773 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2774 req->flags |= REQ_F_CUR_POS;
2775 kiocb->ki_pos = req->file->f_pos;
2777 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2778 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2779 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2783 ioprio = READ_ONCE(sqe->ioprio);
2785 ret = ioprio_check_cap(ioprio);
2789 kiocb->ki_ioprio = ioprio;
2791 kiocb->ki_ioprio = get_current_ioprio();
2793 /* don't allow async punt if RWF_NOWAIT was requested */
2794 if (kiocb->ki_flags & IOCB_NOWAIT)
2795 req->flags |= REQ_F_NOWAIT;
2797 if (ctx->flags & IORING_SETUP_IOPOLL) {
2798 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2799 !kiocb->ki_filp->f_op->iopoll)
2802 kiocb->ki_flags |= IOCB_HIPRI;
2803 kiocb->ki_complete = io_complete_rw_iopoll;
2804 req->iopoll_completed = 0;
2806 if (kiocb->ki_flags & IOCB_HIPRI)
2808 kiocb->ki_complete = io_complete_rw;
2811 req->rw.addr = READ_ONCE(sqe->addr);
2812 req->rw.len = READ_ONCE(sqe->len);
2813 req->buf_index = READ_ONCE(sqe->buf_index);
2817 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2823 case -ERESTARTNOINTR:
2824 case -ERESTARTNOHAND:
2825 case -ERESTART_RESTARTBLOCK:
2827 * We can't just restart the syscall, since previously
2828 * submitted sqes may already be in progress. Just fail this
2834 kiocb->ki_complete(kiocb, ret, 0);
2838 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2839 struct io_comp_state *cs)
2841 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2842 struct io_async_rw *io = req->async_data;
2844 /* add previously done IO, if any */
2845 if (io && io->bytes_done > 0) {
2847 ret = io->bytes_done;
2849 ret += io->bytes_done;
2852 if (req->flags & REQ_F_CUR_POS)
2853 req->file->f_pos = kiocb->ki_pos;
2854 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2855 __io_complete_rw(req, ret, 0, cs);
2857 io_rw_done(kiocb, ret);
2860 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2861 struct iov_iter *iter)
2863 struct io_ring_ctx *ctx = req->ctx;
2864 size_t len = req->rw.len;
2865 struct io_mapped_ubuf *imu;
2866 u16 index, buf_index = req->buf_index;
2870 if (unlikely(buf_index >= ctx->nr_user_bufs))
2872 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2873 imu = &ctx->user_bufs[index];
2874 buf_addr = req->rw.addr;
2877 if (buf_addr + len < buf_addr)
2879 /* not inside the mapped region */
2880 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2884 * May not be a start of buffer, set size appropriately
2885 * and advance us to the beginning.
2887 offset = buf_addr - imu->ubuf;
2888 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2892 * Don't use iov_iter_advance() here, as it's really slow for
2893 * using the latter parts of a big fixed buffer - it iterates
2894 * over each segment manually. We can cheat a bit here, because
2897 * 1) it's a BVEC iter, we set it up
2898 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2899 * first and last bvec
2901 * So just find our index, and adjust the iterator afterwards.
2902 * If the offset is within the first bvec (or the whole first
2903 * bvec, just use iov_iter_advance(). This makes it easier
2904 * since we can just skip the first segment, which may not
2905 * be PAGE_SIZE aligned.
2907 const struct bio_vec *bvec = imu->bvec;
2909 if (offset <= bvec->bv_len) {
2910 iov_iter_advance(iter, offset);
2912 unsigned long seg_skip;
2914 /* skip first vec */
2915 offset -= bvec->bv_len;
2916 seg_skip = 1 + (offset >> PAGE_SHIFT);
2918 iter->bvec = bvec + seg_skip;
2919 iter->nr_segs -= seg_skip;
2920 iter->count -= bvec->bv_len + offset;
2921 iter->iov_offset = offset & ~PAGE_MASK;
2928 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2931 mutex_unlock(&ctx->uring_lock);
2934 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2937 * "Normal" inline submissions always hold the uring_lock, since we
2938 * grab it from the system call. Same is true for the SQPOLL offload.
2939 * The only exception is when we've detached the request and issue it
2940 * from an async worker thread, grab the lock for that case.
2943 mutex_lock(&ctx->uring_lock);
2946 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2947 int bgid, struct io_buffer *kbuf,
2950 struct io_buffer *head;
2952 if (req->flags & REQ_F_BUFFER_SELECTED)
2955 io_ring_submit_lock(req->ctx, needs_lock);
2957 lockdep_assert_held(&req->ctx->uring_lock);
2959 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2961 if (!list_empty(&head->list)) {
2962 kbuf = list_last_entry(&head->list, struct io_buffer,
2964 list_del(&kbuf->list);
2967 idr_remove(&req->ctx->io_buffer_idr, bgid);
2969 if (*len > kbuf->len)
2972 kbuf = ERR_PTR(-ENOBUFS);
2975 io_ring_submit_unlock(req->ctx, needs_lock);
2980 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2983 struct io_buffer *kbuf;
2986 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2987 bgid = req->buf_index;
2988 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2991 req->rw.addr = (u64) (unsigned long) kbuf;
2992 req->flags |= REQ_F_BUFFER_SELECTED;
2993 return u64_to_user_ptr(kbuf->addr);
2996 #ifdef CONFIG_COMPAT
2997 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3000 struct compat_iovec __user *uiov;
3001 compat_ssize_t clen;
3005 uiov = u64_to_user_ptr(req->rw.addr);
3006 if (!access_ok(uiov, sizeof(*uiov)))
3008 if (__get_user(clen, &uiov->iov_len))
3014 buf = io_rw_buffer_select(req, &len, needs_lock);
3016 return PTR_ERR(buf);
3017 iov[0].iov_base = buf;
3018 iov[0].iov_len = (compat_size_t) len;
3023 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3026 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3030 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3033 len = iov[0].iov_len;
3036 buf = io_rw_buffer_select(req, &len, needs_lock);
3038 return PTR_ERR(buf);
3039 iov[0].iov_base = buf;
3040 iov[0].iov_len = len;
3044 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3047 if (req->flags & REQ_F_BUFFER_SELECTED) {
3048 struct io_buffer *kbuf;
3050 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3051 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3052 iov[0].iov_len = kbuf->len;
3057 else if (req->rw.len > 1)
3060 #ifdef CONFIG_COMPAT
3061 if (req->ctx->compat)
3062 return io_compat_import(req, iov, needs_lock);
3065 return __io_iov_buffer_select(req, iov, needs_lock);
3068 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
3069 struct iovec **iovec, struct iov_iter *iter,
3072 void __user *buf = u64_to_user_ptr(req->rw.addr);
3073 size_t sqe_len = req->rw.len;
3077 opcode = req->opcode;
3078 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3080 return io_import_fixed(req, rw, iter);
3083 /* buffer index only valid with fixed read/write, or buffer select */
3084 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3087 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3088 if (req->flags & REQ_F_BUFFER_SELECT) {
3089 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3091 return PTR_ERR(buf);
3092 req->rw.len = sqe_len;
3095 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3097 return ret < 0 ? ret : sqe_len;
3100 if (req->flags & REQ_F_BUFFER_SELECT) {
3101 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3103 ret = (*iovec)->iov_len;
3104 iov_iter_init(iter, rw, *iovec, 1, ret);
3110 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3114 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3115 struct iovec **iovec, struct iov_iter *iter,
3118 struct io_async_rw *iorw = req->async_data;
3121 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
3123 return iov_iter_count(&iorw->iter);
3126 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3128 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3132 * For files that don't have ->read_iter() and ->write_iter(), handle them
3133 * by looping over ->read() or ->write() manually.
3135 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3137 struct kiocb *kiocb = &req->rw.kiocb;
3138 struct file *file = req->file;
3142 * Don't support polled IO through this interface, and we can't
3143 * support non-blocking either. For the latter, this just causes
3144 * the kiocb to be handled from an async context.
3146 if (kiocb->ki_flags & IOCB_HIPRI)
3148 if (kiocb->ki_flags & IOCB_NOWAIT)
3151 while (iov_iter_count(iter)) {
3155 if (!iov_iter_is_bvec(iter)) {
3156 iovec = iov_iter_iovec(iter);
3158 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3159 iovec.iov_len = req->rw.len;
3163 nr = file->f_op->read(file, iovec.iov_base,
3164 iovec.iov_len, io_kiocb_ppos(kiocb));
3166 nr = file->f_op->write(file, iovec.iov_base,
3167 iovec.iov_len, io_kiocb_ppos(kiocb));
3176 if (nr != iovec.iov_len)
3180 iov_iter_advance(iter, nr);
3186 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3187 const struct iovec *fast_iov, struct iov_iter *iter)
3189 struct io_async_rw *rw = req->async_data;
3191 memcpy(&rw->iter, iter, sizeof(*iter));
3192 rw->free_iovec = iovec;
3194 /* can only be fixed buffers, no need to do anything */
3195 if (iter->type == ITER_BVEC)
3198 unsigned iov_off = 0;
3200 rw->iter.iov = rw->fast_iov;
3201 if (iter->iov != fast_iov) {
3202 iov_off = iter->iov - fast_iov;
3203 rw->iter.iov += iov_off;
3205 if (rw->fast_iov != fast_iov)
3206 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3207 sizeof(struct iovec) * iter->nr_segs);
3209 req->flags |= REQ_F_NEED_CLEANUP;
3213 static inline int __io_alloc_async_data(struct io_kiocb *req)
3215 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3216 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3217 return req->async_data == NULL;
3220 static int io_alloc_async_data(struct io_kiocb *req)
3222 if (!io_op_defs[req->opcode].needs_async_data)
3225 return __io_alloc_async_data(req);
3228 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3229 const struct iovec *fast_iov,
3230 struct iov_iter *iter, bool force)
3232 if (!force && !io_op_defs[req->opcode].needs_async_data)
3234 if (!req->async_data) {
3235 if (__io_alloc_async_data(req))
3238 io_req_map_rw(req, iovec, fast_iov, iter);
3243 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3245 struct io_async_rw *iorw = req->async_data;
3246 struct iovec *iov = iorw->fast_iov;
3249 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, false);
3250 if (unlikely(ret < 0))
3253 iorw->bytes_done = 0;
3254 iorw->free_iovec = iov;
3256 req->flags |= REQ_F_NEED_CLEANUP;
3260 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3264 ret = io_prep_rw(req, sqe);
3268 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3271 /* either don't need iovec imported or already have it */
3272 if (!req->async_data)
3274 return io_rw_prep_async(req, READ);
3278 * This is our waitqueue callback handler, registered through lock_page_async()
3279 * when we initially tried to do the IO with the iocb armed our waitqueue.
3280 * This gets called when the page is unlocked, and we generally expect that to
3281 * happen when the page IO is completed and the page is now uptodate. This will
3282 * queue a task_work based retry of the operation, attempting to copy the data
3283 * again. If the latter fails because the page was NOT uptodate, then we will
3284 * do a thread based blocking retry of the operation. That's the unexpected
3287 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3288 int sync, void *arg)
3290 struct wait_page_queue *wpq;
3291 struct io_kiocb *req = wait->private;
3292 struct wait_page_key *key = arg;
3295 wpq = container_of(wait, struct wait_page_queue, wait);
3297 if (!wake_page_match(wpq, key))
3300 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3301 list_del_init(&wait->entry);
3303 init_task_work(&req->task_work, io_req_task_submit);
3304 percpu_ref_get(&req->ctx->refs);
3306 /* submit ref gets dropped, acquire a new one */
3307 refcount_inc(&req->refs);
3308 ret = io_req_task_work_add(req, true);
3309 if (unlikely(ret)) {
3310 struct task_struct *tsk;
3312 /* queue just for cancelation */
3313 init_task_work(&req->task_work, io_req_task_cancel);
3314 tsk = io_wq_get_task(req->ctx->io_wq);
3315 task_work_add(tsk, &req->task_work, TWA_NONE);
3316 wake_up_process(tsk);
3322 * This controls whether a given IO request should be armed for async page
3323 * based retry. If we return false here, the request is handed to the async
3324 * worker threads for retry. If we're doing buffered reads on a regular file,
3325 * we prepare a private wait_page_queue entry and retry the operation. This
3326 * will either succeed because the page is now uptodate and unlocked, or it
3327 * will register a callback when the page is unlocked at IO completion. Through
3328 * that callback, io_uring uses task_work to setup a retry of the operation.
3329 * That retry will attempt the buffered read again. The retry will generally
3330 * succeed, or in rare cases where it fails, we then fall back to using the
3331 * async worker threads for a blocking retry.
3333 static bool io_rw_should_retry(struct io_kiocb *req)
3335 struct io_async_rw *rw = req->async_data;
3336 struct wait_page_queue *wait = &rw->wpq;
3337 struct kiocb *kiocb = &req->rw.kiocb;
3339 /* never retry for NOWAIT, we just complete with -EAGAIN */
3340 if (req->flags & REQ_F_NOWAIT)
3343 /* Only for buffered IO */
3344 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3348 * just use poll if we can, and don't attempt if the fs doesn't
3349 * support callback based unlocks
3351 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3354 wait->wait.func = io_async_buf_func;
3355 wait->wait.private = req;
3356 wait->wait.flags = 0;
3357 INIT_LIST_HEAD(&wait->wait.entry);
3358 kiocb->ki_flags |= IOCB_WAITQ;
3359 kiocb->ki_flags &= ~IOCB_NOWAIT;
3360 kiocb->ki_waitq = wait;
3364 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3366 if (req->file->f_op->read_iter)
3367 return call_read_iter(req->file, &req->rw.kiocb, iter);
3368 else if (req->file->f_op->read)
3369 return loop_rw_iter(READ, req, iter);
3374 static int io_read(struct io_kiocb *req, bool force_nonblock,
3375 struct io_comp_state *cs)
3377 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3378 struct kiocb *kiocb = &req->rw.kiocb;
3379 struct iov_iter __iter, *iter = &__iter;
3380 struct io_async_rw *rw = req->async_data;
3381 ssize_t io_size, ret, ret2;
3388 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3391 iov_count = iov_iter_count(iter);
3393 req->result = io_size;
3396 /* Ensure we clear previously set non-block flag */
3397 if (!force_nonblock)
3398 kiocb->ki_flags &= ~IOCB_NOWAIT;
3400 kiocb->ki_flags |= IOCB_NOWAIT;
3403 /* If the file doesn't support async, just async punt */
3404 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3408 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3412 ret = io_iter_do_read(req, iter);
3416 } else if (ret == -EIOCBQUEUED) {
3419 } else if (ret == -EAGAIN) {
3420 /* IOPOLL retry should happen for io-wq threads */
3421 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3423 /* no retry on NONBLOCK marked file */
3424 if (req->file->f_flags & O_NONBLOCK)
3426 /* some cases will consume bytes even on error returns */
3427 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3430 } else if (ret < 0) {
3431 /* make sure -ERESTARTSYS -> -EINTR is done */
3435 /* read it all, or we did blocking attempt. no retry. */
3436 if (!iov_iter_count(iter) || !force_nonblock ||
3437 (req->file->f_flags & O_NONBLOCK))
3442 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3449 rw = req->async_data;
3450 /* it's copied and will be cleaned with ->io */
3452 /* now use our persistent iterator, if we aren't already */
3455 rw->bytes_done += ret;
3456 /* if we can retry, do so with the callbacks armed */
3457 if (!io_rw_should_retry(req)) {
3458 kiocb->ki_flags &= ~IOCB_WAITQ;
3463 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3464 * get -EIOCBQUEUED, then we'll get a notification when the desired
3465 * page gets unlocked. We can also get a partial read here, and if we
3466 * do, then just retry at the new offset.
3468 ret = io_iter_do_read(req, iter);
3469 if (ret == -EIOCBQUEUED) {
3472 } else if (ret > 0 && ret < io_size) {
3473 /* we got some bytes, but not all. retry. */
3477 kiocb_done(kiocb, ret, cs);
3480 /* it's reportedly faster than delegating the null check to kfree() */
3486 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3490 ret = io_prep_rw(req, sqe);
3494 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3497 /* either don't need iovec imported or already have it */
3498 if (!req->async_data)
3500 return io_rw_prep_async(req, WRITE);
3503 static int io_write(struct io_kiocb *req, bool force_nonblock,
3504 struct io_comp_state *cs)
3506 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3507 struct kiocb *kiocb = &req->rw.kiocb;
3508 struct iov_iter __iter, *iter = &__iter;
3509 struct io_async_rw *rw = req->async_data;
3511 ssize_t ret, ret2, io_size;
3516 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3519 iov_count = iov_iter_count(iter);
3521 req->result = io_size;
3523 /* Ensure we clear previously set non-block flag */
3524 if (!force_nonblock)
3525 kiocb->ki_flags &= ~IOCB_NOWAIT;
3527 kiocb->ki_flags |= IOCB_NOWAIT;
3529 /* If the file doesn't support async, just async punt */
3530 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3533 /* file path doesn't support NOWAIT for non-direct_IO */
3534 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3535 (req->flags & REQ_F_ISREG))
3538 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3543 * Open-code file_start_write here to grab freeze protection,
3544 * which will be released by another thread in
3545 * io_complete_rw(). Fool lockdep by telling it the lock got
3546 * released so that it doesn't complain about the held lock when
3547 * we return to userspace.
3549 if (req->flags & REQ_F_ISREG) {
3550 __sb_start_write(file_inode(req->file)->i_sb,
3551 SB_FREEZE_WRITE, true);
3552 __sb_writers_release(file_inode(req->file)->i_sb,
3555 kiocb->ki_flags |= IOCB_WRITE;
3557 if (req->file->f_op->write_iter)
3558 ret2 = call_write_iter(req->file, kiocb, iter);
3559 else if (req->file->f_op->write)
3560 ret2 = loop_rw_iter(WRITE, req, iter);
3565 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3566 * retry them without IOCB_NOWAIT.
3568 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3570 /* no retry on NONBLOCK marked file */
3571 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3573 if (!force_nonblock || ret2 != -EAGAIN) {
3574 /* IOPOLL retry should happen for io-wq threads */
3575 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3578 kiocb_done(kiocb, ret2, cs);
3581 /* some cases will consume bytes even on error returns */
3582 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3583 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3588 /* it's reportedly faster than delegating the null check to kfree() */
3594 static int __io_splice_prep(struct io_kiocb *req,
3595 const struct io_uring_sqe *sqe)
3597 struct io_splice* sp = &req->splice;
3598 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3600 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3604 sp->len = READ_ONCE(sqe->len);
3605 sp->flags = READ_ONCE(sqe->splice_flags);
3607 if (unlikely(sp->flags & ~valid_flags))
3610 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3611 (sp->flags & SPLICE_F_FD_IN_FIXED));
3614 req->flags |= REQ_F_NEED_CLEANUP;
3616 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3618 * Splice operation will be punted aync, and here need to
3619 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3621 io_req_init_async(req);
3622 req->work.flags |= IO_WQ_WORK_UNBOUND;
3628 static int io_tee_prep(struct io_kiocb *req,
3629 const struct io_uring_sqe *sqe)
3631 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3633 return __io_splice_prep(req, sqe);
3636 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3638 struct io_splice *sp = &req->splice;
3639 struct file *in = sp->file_in;
3640 struct file *out = sp->file_out;
3641 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3647 ret = do_tee(in, out, sp->len, flags);
3649 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3650 req->flags &= ~REQ_F_NEED_CLEANUP;
3653 req_set_fail_links(req);
3654 io_req_complete(req, ret);
3658 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3660 struct io_splice* sp = &req->splice;
3662 sp->off_in = READ_ONCE(sqe->splice_off_in);
3663 sp->off_out = READ_ONCE(sqe->off);
3664 return __io_splice_prep(req, sqe);
3667 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3669 struct io_splice *sp = &req->splice;
3670 struct file *in = sp->file_in;
3671 struct file *out = sp->file_out;
3672 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3673 loff_t *poff_in, *poff_out;
3679 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3680 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3683 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3685 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3686 req->flags &= ~REQ_F_NEED_CLEANUP;
3689 req_set_fail_links(req);
3690 io_req_complete(req, ret);
3695 * IORING_OP_NOP just posts a completion event, nothing else.
3697 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3699 struct io_ring_ctx *ctx = req->ctx;
3701 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3704 __io_req_complete(req, 0, 0, cs);
3708 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3710 struct io_ring_ctx *ctx = req->ctx;
3715 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3717 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3720 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3721 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3724 req->sync.off = READ_ONCE(sqe->off);
3725 req->sync.len = READ_ONCE(sqe->len);
3729 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3731 loff_t end = req->sync.off + req->sync.len;
3734 /* fsync always requires a blocking context */
3738 ret = vfs_fsync_range(req->file, req->sync.off,
3739 end > 0 ? end : LLONG_MAX,
3740 req->sync.flags & IORING_FSYNC_DATASYNC);
3742 req_set_fail_links(req);
3743 io_req_complete(req, ret);
3747 static int io_fallocate_prep(struct io_kiocb *req,
3748 const struct io_uring_sqe *sqe)
3750 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3752 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3755 req->sync.off = READ_ONCE(sqe->off);
3756 req->sync.len = READ_ONCE(sqe->addr);
3757 req->sync.mode = READ_ONCE(sqe->len);
3761 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3765 /* fallocate always requiring blocking context */
3768 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3771 req_set_fail_links(req);
3772 io_req_complete(req, ret);
3776 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3778 const char __user *fname;
3781 if (unlikely(sqe->ioprio || sqe->buf_index))
3783 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3786 /* open.how should be already initialised */
3787 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3788 req->open.how.flags |= O_LARGEFILE;
3790 req->open.dfd = READ_ONCE(sqe->fd);
3791 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3792 req->open.filename = getname(fname);
3793 if (IS_ERR(req->open.filename)) {
3794 ret = PTR_ERR(req->open.filename);
3795 req->open.filename = NULL;
3798 req->open.nofile = rlimit(RLIMIT_NOFILE);
3799 req->open.ignore_nonblock = false;
3800 req->flags |= REQ_F_NEED_CLEANUP;
3804 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3808 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3810 mode = READ_ONCE(sqe->len);
3811 flags = READ_ONCE(sqe->open_flags);
3812 req->open.how = build_open_how(flags, mode);
3813 return __io_openat_prep(req, sqe);
3816 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3818 struct open_how __user *how;
3822 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3824 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3825 len = READ_ONCE(sqe->len);
3826 if (len < OPEN_HOW_SIZE_VER0)
3829 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3834 return __io_openat_prep(req, sqe);
3837 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3839 struct open_flags op;
3843 if (force_nonblock && !req->open.ignore_nonblock)
3846 ret = build_open_flags(&req->open.how, &op);
3850 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3854 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3857 ret = PTR_ERR(file);
3859 * A work-around to ensure that /proc/self works that way
3860 * that it should - if we get -EOPNOTSUPP back, then assume
3861 * that proc_self_get_link() failed us because we're in async
3862 * context. We should be safe to retry this from the task
3863 * itself with force_nonblock == false set, as it should not
3864 * block on lookup. Would be nice to know this upfront and
3865 * avoid the async dance, but doesn't seem feasible.
3867 if (ret == -EOPNOTSUPP && io_wq_current_is_worker()) {
3868 req->open.ignore_nonblock = true;
3869 refcount_inc(&req->refs);
3870 io_req_task_queue(req);
3874 fsnotify_open(file);
3875 fd_install(ret, file);
3878 putname(req->open.filename);
3879 req->flags &= ~REQ_F_NEED_CLEANUP;
3881 req_set_fail_links(req);
3882 io_req_complete(req, ret);
3886 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3888 return io_openat2(req, force_nonblock);
3891 static int io_remove_buffers_prep(struct io_kiocb *req,
3892 const struct io_uring_sqe *sqe)
3894 struct io_provide_buf *p = &req->pbuf;
3897 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3900 tmp = READ_ONCE(sqe->fd);
3901 if (!tmp || tmp > USHRT_MAX)
3904 memset(p, 0, sizeof(*p));
3906 p->bgid = READ_ONCE(sqe->buf_group);
3910 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3911 int bgid, unsigned nbufs)
3915 /* shouldn't happen */
3919 /* the head kbuf is the list itself */
3920 while (!list_empty(&buf->list)) {
3921 struct io_buffer *nxt;
3923 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3924 list_del(&nxt->list);
3931 idr_remove(&ctx->io_buffer_idr, bgid);
3936 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3937 struct io_comp_state *cs)
3939 struct io_provide_buf *p = &req->pbuf;
3940 struct io_ring_ctx *ctx = req->ctx;
3941 struct io_buffer *head;
3944 io_ring_submit_lock(ctx, !force_nonblock);
3946 lockdep_assert_held(&ctx->uring_lock);
3949 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3951 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3953 io_ring_submit_lock(ctx, !force_nonblock);
3955 req_set_fail_links(req);
3956 __io_req_complete(req, ret, 0, cs);
3960 static int io_provide_buffers_prep(struct io_kiocb *req,
3961 const struct io_uring_sqe *sqe)
3963 struct io_provide_buf *p = &req->pbuf;
3966 if (sqe->ioprio || sqe->rw_flags)
3969 tmp = READ_ONCE(sqe->fd);
3970 if (!tmp || tmp > USHRT_MAX)
3973 p->addr = READ_ONCE(sqe->addr);
3974 p->len = READ_ONCE(sqe->len);
3976 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3979 p->bgid = READ_ONCE(sqe->buf_group);
3980 tmp = READ_ONCE(sqe->off);
3981 if (tmp > USHRT_MAX)
3987 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3989 struct io_buffer *buf;
3990 u64 addr = pbuf->addr;
3991 int i, bid = pbuf->bid;
3993 for (i = 0; i < pbuf->nbufs; i++) {
3994 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3999 buf->len = pbuf->len;
4004 INIT_LIST_HEAD(&buf->list);
4007 list_add_tail(&buf->list, &(*head)->list);
4011 return i ? i : -ENOMEM;
4014 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
4015 struct io_comp_state *cs)
4017 struct io_provide_buf *p = &req->pbuf;
4018 struct io_ring_ctx *ctx = req->ctx;
4019 struct io_buffer *head, *list;
4022 io_ring_submit_lock(ctx, !force_nonblock);
4024 lockdep_assert_held(&ctx->uring_lock);
4026 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
4028 ret = io_add_buffers(p, &head);
4033 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
4036 __io_remove_buffers(ctx, head, p->bgid, -1U);
4041 io_ring_submit_unlock(ctx, !force_nonblock);
4043 req_set_fail_links(req);
4044 __io_req_complete(req, ret, 0, cs);
4048 static int io_epoll_ctl_prep(struct io_kiocb *req,
4049 const struct io_uring_sqe *sqe)
4051 #if defined(CONFIG_EPOLL)
4052 if (sqe->ioprio || sqe->buf_index)
4054 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4057 req->epoll.epfd = READ_ONCE(sqe->fd);
4058 req->epoll.op = READ_ONCE(sqe->len);
4059 req->epoll.fd = READ_ONCE(sqe->off);
4061 if (ep_op_has_event(req->epoll.op)) {
4062 struct epoll_event __user *ev;
4064 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4065 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4075 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4076 struct io_comp_state *cs)
4078 #if defined(CONFIG_EPOLL)
4079 struct io_epoll *ie = &req->epoll;
4082 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4083 if (force_nonblock && ret == -EAGAIN)
4087 req_set_fail_links(req);
4088 __io_req_complete(req, ret, 0, cs);
4095 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4097 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4098 if (sqe->ioprio || sqe->buf_index || sqe->off)
4100 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4103 req->madvise.addr = READ_ONCE(sqe->addr);
4104 req->madvise.len = READ_ONCE(sqe->len);
4105 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4112 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4114 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4115 struct io_madvise *ma = &req->madvise;
4121 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4123 req_set_fail_links(req);
4124 io_req_complete(req, ret);
4131 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4133 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4135 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4138 req->fadvise.offset = READ_ONCE(sqe->off);
4139 req->fadvise.len = READ_ONCE(sqe->len);
4140 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4144 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4146 struct io_fadvise *fa = &req->fadvise;
4149 if (force_nonblock) {
4150 switch (fa->advice) {
4151 case POSIX_FADV_NORMAL:
4152 case POSIX_FADV_RANDOM:
4153 case POSIX_FADV_SEQUENTIAL:
4160 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4162 req_set_fail_links(req);
4163 io_req_complete(req, ret);
4167 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4169 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4171 if (sqe->ioprio || sqe->buf_index)
4173 if (req->flags & REQ_F_FIXED_FILE)
4176 req->statx.dfd = READ_ONCE(sqe->fd);
4177 req->statx.mask = READ_ONCE(sqe->len);
4178 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4179 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4180 req->statx.flags = READ_ONCE(sqe->statx_flags);
4185 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4187 struct io_statx *ctx = &req->statx;
4190 if (force_nonblock) {
4191 /* only need file table for an actual valid fd */
4192 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4193 req->flags |= REQ_F_NO_FILE_TABLE;
4197 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4201 req_set_fail_links(req);
4202 io_req_complete(req, ret);
4206 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4209 * If we queue this for async, it must not be cancellable. That would
4210 * leave the 'file' in an undeterminate state, and here need to modify
4211 * io_wq_work.flags, so initialize io_wq_work firstly.
4213 io_req_init_async(req);
4214 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4216 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4218 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4219 sqe->rw_flags || sqe->buf_index)
4221 if (req->flags & REQ_F_FIXED_FILE)
4224 req->close.fd = READ_ONCE(sqe->fd);
4225 if ((req->file && req->file->f_op == &io_uring_fops))
4228 req->close.put_file = NULL;
4232 static int io_close(struct io_kiocb *req, bool force_nonblock,
4233 struct io_comp_state *cs)
4235 struct io_close *close = &req->close;
4238 /* might be already done during nonblock submission */
4239 if (!close->put_file) {
4240 ret = __close_fd_get_file(close->fd, &close->put_file);
4242 return (ret == -ENOENT) ? -EBADF : ret;
4245 /* if the file has a flush method, be safe and punt to async */
4246 if (close->put_file->f_op->flush && force_nonblock) {
4247 /* was never set, but play safe */
4248 req->flags &= ~REQ_F_NOWAIT;
4249 /* avoid grabbing files - we don't need the files */
4250 req->flags |= REQ_F_NO_FILE_TABLE;
4254 /* No ->flush() or already async, safely close from here */
4255 ret = filp_close(close->put_file, req->work.identity->files);
4257 req_set_fail_links(req);
4258 fput(close->put_file);
4259 close->put_file = NULL;
4260 __io_req_complete(req, ret, 0, cs);
4264 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4266 struct io_ring_ctx *ctx = req->ctx;
4271 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4273 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4276 req->sync.off = READ_ONCE(sqe->off);
4277 req->sync.len = READ_ONCE(sqe->len);
4278 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4282 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4286 /* sync_file_range always requires a blocking context */
4290 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4293 req_set_fail_links(req);
4294 io_req_complete(req, ret);
4298 #if defined(CONFIG_NET)
4299 static int io_setup_async_msg(struct io_kiocb *req,
4300 struct io_async_msghdr *kmsg)
4302 struct io_async_msghdr *async_msg = req->async_data;
4306 if (io_alloc_async_data(req)) {
4307 if (kmsg->iov != kmsg->fast_iov)
4311 async_msg = req->async_data;
4312 req->flags |= REQ_F_NEED_CLEANUP;
4313 memcpy(async_msg, kmsg, sizeof(*kmsg));
4317 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4318 struct io_async_msghdr *iomsg)
4320 iomsg->iov = iomsg->fast_iov;
4321 iomsg->msg.msg_name = &iomsg->addr;
4322 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4323 req->sr_msg.msg_flags, &iomsg->iov);
4326 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4328 struct io_async_msghdr *async_msg = req->async_data;
4329 struct io_sr_msg *sr = &req->sr_msg;
4332 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4335 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4336 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4337 sr->len = READ_ONCE(sqe->len);
4339 #ifdef CONFIG_COMPAT
4340 if (req->ctx->compat)
4341 sr->msg_flags |= MSG_CMSG_COMPAT;
4344 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4346 ret = io_sendmsg_copy_hdr(req, async_msg);
4348 req->flags |= REQ_F_NEED_CLEANUP;
4352 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4353 struct io_comp_state *cs)
4355 struct io_async_msghdr iomsg, *kmsg;
4356 struct socket *sock;
4360 sock = sock_from_file(req->file, &ret);
4361 if (unlikely(!sock))
4364 if (req->async_data) {
4365 kmsg = req->async_data;
4366 kmsg->msg.msg_name = &kmsg->addr;
4367 /* if iov is set, it's allocated already */
4369 kmsg->iov = kmsg->fast_iov;
4370 kmsg->msg.msg_iter.iov = kmsg->iov;
4372 ret = io_sendmsg_copy_hdr(req, &iomsg);
4378 flags = req->sr_msg.msg_flags;
4379 if (flags & MSG_DONTWAIT)
4380 req->flags |= REQ_F_NOWAIT;
4381 else if (force_nonblock)
4382 flags |= MSG_DONTWAIT;
4384 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4385 if (force_nonblock && ret == -EAGAIN)
4386 return io_setup_async_msg(req, kmsg);
4387 if (ret == -ERESTARTSYS)
4390 if (kmsg->iov != kmsg->fast_iov)
4392 req->flags &= ~REQ_F_NEED_CLEANUP;
4394 req_set_fail_links(req);
4395 __io_req_complete(req, ret, 0, cs);
4399 static int io_send(struct io_kiocb *req, bool force_nonblock,
4400 struct io_comp_state *cs)
4402 struct io_sr_msg *sr = &req->sr_msg;
4405 struct socket *sock;
4409 sock = sock_from_file(req->file, &ret);
4410 if (unlikely(!sock))
4413 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4417 msg.msg_name = NULL;
4418 msg.msg_control = NULL;
4419 msg.msg_controllen = 0;
4420 msg.msg_namelen = 0;
4422 flags = req->sr_msg.msg_flags;
4423 if (flags & MSG_DONTWAIT)
4424 req->flags |= REQ_F_NOWAIT;
4425 else if (force_nonblock)
4426 flags |= MSG_DONTWAIT;
4428 msg.msg_flags = flags;
4429 ret = sock_sendmsg(sock, &msg);
4430 if (force_nonblock && ret == -EAGAIN)
4432 if (ret == -ERESTARTSYS)
4436 req_set_fail_links(req);
4437 __io_req_complete(req, ret, 0, cs);
4441 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4442 struct io_async_msghdr *iomsg)
4444 struct io_sr_msg *sr = &req->sr_msg;
4445 struct iovec __user *uiov;
4449 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4450 &iomsg->uaddr, &uiov, &iov_len);
4454 if (req->flags & REQ_F_BUFFER_SELECT) {
4457 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4459 sr->len = iomsg->iov[0].iov_len;
4460 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4464 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4465 &iomsg->iov, &iomsg->msg.msg_iter,
4474 #ifdef CONFIG_COMPAT
4475 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4476 struct io_async_msghdr *iomsg)
4478 struct compat_msghdr __user *msg_compat;
4479 struct io_sr_msg *sr = &req->sr_msg;
4480 struct compat_iovec __user *uiov;
4485 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4486 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4491 uiov = compat_ptr(ptr);
4492 if (req->flags & REQ_F_BUFFER_SELECT) {
4493 compat_ssize_t clen;
4497 if (!access_ok(uiov, sizeof(*uiov)))
4499 if (__get_user(clen, &uiov->iov_len))
4503 sr->len = iomsg->iov[0].iov_len;
4506 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4507 UIO_FASTIOV, &iomsg->iov,
4508 &iomsg->msg.msg_iter, true);
4517 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4518 struct io_async_msghdr *iomsg)
4520 iomsg->msg.msg_name = &iomsg->addr;
4521 iomsg->iov = iomsg->fast_iov;
4523 #ifdef CONFIG_COMPAT
4524 if (req->ctx->compat)
4525 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4528 return __io_recvmsg_copy_hdr(req, iomsg);
4531 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4534 struct io_sr_msg *sr = &req->sr_msg;
4535 struct io_buffer *kbuf;
4537 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4542 req->flags |= REQ_F_BUFFER_SELECTED;
4546 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4548 return io_put_kbuf(req, req->sr_msg.kbuf);
4551 static int io_recvmsg_prep(struct io_kiocb *req,
4552 const struct io_uring_sqe *sqe)
4554 struct io_async_msghdr *async_msg = req->async_data;
4555 struct io_sr_msg *sr = &req->sr_msg;
4558 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4561 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4562 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4563 sr->len = READ_ONCE(sqe->len);
4564 sr->bgid = READ_ONCE(sqe->buf_group);
4566 #ifdef CONFIG_COMPAT
4567 if (req->ctx->compat)
4568 sr->msg_flags |= MSG_CMSG_COMPAT;
4571 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4573 ret = io_recvmsg_copy_hdr(req, async_msg);
4575 req->flags |= REQ_F_NEED_CLEANUP;
4579 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4580 struct io_comp_state *cs)
4582 struct io_async_msghdr iomsg, *kmsg;
4583 struct socket *sock;
4584 struct io_buffer *kbuf;
4586 int ret, cflags = 0;
4588 sock = sock_from_file(req->file, &ret);
4589 if (unlikely(!sock))
4592 if (req->async_data) {
4593 kmsg = req->async_data;
4594 kmsg->msg.msg_name = &kmsg->addr;
4595 /* if iov is set, it's allocated already */
4597 kmsg->iov = kmsg->fast_iov;
4598 kmsg->msg.msg_iter.iov = kmsg->iov;
4600 ret = io_recvmsg_copy_hdr(req, &iomsg);
4606 if (req->flags & REQ_F_BUFFER_SELECT) {
4607 kbuf = io_recv_buffer_select(req, !force_nonblock);
4609 return PTR_ERR(kbuf);
4610 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4611 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4612 1, req->sr_msg.len);
4615 flags = req->sr_msg.msg_flags;
4616 if (flags & MSG_DONTWAIT)
4617 req->flags |= REQ_F_NOWAIT;
4618 else if (force_nonblock)
4619 flags |= MSG_DONTWAIT;
4621 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4622 kmsg->uaddr, flags);
4623 if (force_nonblock && ret == -EAGAIN)
4624 return io_setup_async_msg(req, kmsg);
4625 if (ret == -ERESTARTSYS)
4628 if (req->flags & REQ_F_BUFFER_SELECTED)
4629 cflags = io_put_recv_kbuf(req);
4630 if (kmsg->iov != kmsg->fast_iov)
4632 req->flags &= ~REQ_F_NEED_CLEANUP;
4634 req_set_fail_links(req);
4635 __io_req_complete(req, ret, cflags, cs);
4639 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4640 struct io_comp_state *cs)
4642 struct io_buffer *kbuf;
4643 struct io_sr_msg *sr = &req->sr_msg;
4645 void __user *buf = sr->buf;
4646 struct socket *sock;
4649 int ret, cflags = 0;
4651 sock = sock_from_file(req->file, &ret);
4652 if (unlikely(!sock))
4655 if (req->flags & REQ_F_BUFFER_SELECT) {
4656 kbuf = io_recv_buffer_select(req, !force_nonblock);
4658 return PTR_ERR(kbuf);
4659 buf = u64_to_user_ptr(kbuf->addr);
4662 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4666 msg.msg_name = NULL;
4667 msg.msg_control = NULL;
4668 msg.msg_controllen = 0;
4669 msg.msg_namelen = 0;
4670 msg.msg_iocb = NULL;
4673 flags = req->sr_msg.msg_flags;
4674 if (flags & MSG_DONTWAIT)
4675 req->flags |= REQ_F_NOWAIT;
4676 else if (force_nonblock)
4677 flags |= MSG_DONTWAIT;
4679 ret = sock_recvmsg(sock, &msg, flags);
4680 if (force_nonblock && ret == -EAGAIN)
4682 if (ret == -ERESTARTSYS)
4685 if (req->flags & REQ_F_BUFFER_SELECTED)
4686 cflags = io_put_recv_kbuf(req);
4688 req_set_fail_links(req);
4689 __io_req_complete(req, ret, cflags, cs);
4693 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4695 struct io_accept *accept = &req->accept;
4697 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4699 if (sqe->ioprio || sqe->len || sqe->buf_index)
4702 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4703 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4704 accept->flags = READ_ONCE(sqe->accept_flags);
4705 accept->nofile = rlimit(RLIMIT_NOFILE);
4709 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4710 struct io_comp_state *cs)
4712 struct io_accept *accept = &req->accept;
4713 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4716 if (req->file->f_flags & O_NONBLOCK)
4717 req->flags |= REQ_F_NOWAIT;
4719 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4720 accept->addr_len, accept->flags,
4722 if (ret == -EAGAIN && force_nonblock)
4725 if (ret == -ERESTARTSYS)
4727 req_set_fail_links(req);
4729 __io_req_complete(req, ret, 0, cs);
4733 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4735 struct io_connect *conn = &req->connect;
4736 struct io_async_connect *io = req->async_data;
4738 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4740 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4743 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4744 conn->addr_len = READ_ONCE(sqe->addr2);
4749 return move_addr_to_kernel(conn->addr, conn->addr_len,
4753 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4754 struct io_comp_state *cs)
4756 struct io_async_connect __io, *io;
4757 unsigned file_flags;
4760 if (req->async_data) {
4761 io = req->async_data;
4763 ret = move_addr_to_kernel(req->connect.addr,
4764 req->connect.addr_len,
4771 file_flags = force_nonblock ? O_NONBLOCK : 0;
4773 ret = __sys_connect_file(req->file, &io->address,
4774 req->connect.addr_len, file_flags);
4775 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4776 if (req->async_data)
4778 if (io_alloc_async_data(req)) {
4782 io = req->async_data;
4783 memcpy(req->async_data, &__io, sizeof(__io));
4786 if (ret == -ERESTARTSYS)
4790 req_set_fail_links(req);
4791 __io_req_complete(req, ret, 0, cs);
4794 #else /* !CONFIG_NET */
4795 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4800 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4801 struct io_comp_state *cs)
4806 static int io_send(struct io_kiocb *req, bool force_nonblock,
4807 struct io_comp_state *cs)
4812 static int io_recvmsg_prep(struct io_kiocb *req,
4813 const struct io_uring_sqe *sqe)
4818 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4819 struct io_comp_state *cs)
4824 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4825 struct io_comp_state *cs)
4830 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4835 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4836 struct io_comp_state *cs)
4841 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4846 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4847 struct io_comp_state *cs)
4851 #endif /* CONFIG_NET */
4853 struct io_poll_table {
4854 struct poll_table_struct pt;
4855 struct io_kiocb *req;
4859 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4860 __poll_t mask, task_work_func_t func)
4865 /* for instances that support it check for an event match first: */
4866 if (mask && !(mask & poll->events))
4869 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4871 list_del_init(&poll->wait.entry);
4874 init_task_work(&req->task_work, func);
4875 percpu_ref_get(&req->ctx->refs);
4878 * If we using the signalfd wait_queue_head for this wakeup, then
4879 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4880 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4881 * either, as the normal wakeup will suffice.
4883 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4886 * If this fails, then the task is exiting. When a task exits, the
4887 * work gets canceled, so just cancel this request as well instead
4888 * of executing it. We can't safely execute it anyway, as we may not
4889 * have the needed state needed for it anyway.
4891 ret = io_req_task_work_add(req, twa_signal_ok);
4892 if (unlikely(ret)) {
4893 struct task_struct *tsk;
4895 WRITE_ONCE(poll->canceled, true);
4896 tsk = io_wq_get_task(req->ctx->io_wq);
4897 task_work_add(tsk, &req->task_work, TWA_NONE);
4898 wake_up_process(tsk);
4903 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4904 __acquires(&req->ctx->completion_lock)
4906 struct io_ring_ctx *ctx = req->ctx;
4908 if (!req->result && !READ_ONCE(poll->canceled)) {
4909 struct poll_table_struct pt = { ._key = poll->events };
4911 req->result = vfs_poll(req->file, &pt) & poll->events;
4914 spin_lock_irq(&ctx->completion_lock);
4915 if (!req->result && !READ_ONCE(poll->canceled)) {
4916 add_wait_queue(poll->head, &poll->wait);
4923 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4925 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4926 if (req->opcode == IORING_OP_POLL_ADD)
4927 return req->async_data;
4928 return req->apoll->double_poll;
4931 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4933 if (req->opcode == IORING_OP_POLL_ADD)
4935 return &req->apoll->poll;
4938 static void io_poll_remove_double(struct io_kiocb *req)
4940 struct io_poll_iocb *poll = io_poll_get_double(req);
4942 lockdep_assert_held(&req->ctx->completion_lock);
4944 if (poll && poll->head) {
4945 struct wait_queue_head *head = poll->head;
4947 spin_lock(&head->lock);
4948 list_del_init(&poll->wait.entry);
4949 if (poll->wait.private)
4950 refcount_dec(&req->refs);
4952 spin_unlock(&head->lock);
4956 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4958 struct io_ring_ctx *ctx = req->ctx;
4960 io_poll_remove_double(req);
4961 req->poll.done = true;
4962 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4963 io_commit_cqring(ctx);
4966 static void io_poll_task_func(struct callback_head *cb)
4968 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4969 struct io_ring_ctx *ctx = req->ctx;
4970 struct io_kiocb *nxt;
4972 if (io_poll_rewait(req, &req->poll)) {
4973 spin_unlock_irq(&ctx->completion_lock);
4975 hash_del(&req->hash_node);
4976 io_poll_complete(req, req->result, 0);
4977 spin_unlock_irq(&ctx->completion_lock);
4979 nxt = io_put_req_find_next(req);
4980 io_cqring_ev_posted(ctx);
4982 __io_req_task_submit(nxt);
4985 percpu_ref_put(&ctx->refs);
4988 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4989 int sync, void *key)
4991 struct io_kiocb *req = wait->private;
4992 struct io_poll_iocb *poll = io_poll_get_single(req);
4993 __poll_t mask = key_to_poll(key);
4995 /* for instances that support it check for an event match first: */
4996 if (mask && !(mask & poll->events))
4999 list_del_init(&wait->entry);
5001 if (poll && poll->head) {
5004 spin_lock(&poll->head->lock);
5005 done = list_empty(&poll->wait.entry);
5007 list_del_init(&poll->wait.entry);
5008 /* make sure double remove sees this as being gone */
5009 wait->private = NULL;
5010 spin_unlock(&poll->head->lock);
5012 /* use wait func handler, so it matches the rq type */
5013 poll->wait.func(&poll->wait, mode, sync, key);
5016 refcount_dec(&req->refs);
5020 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5021 wait_queue_func_t wake_func)
5025 poll->canceled = false;
5026 poll->events = events;
5027 INIT_LIST_HEAD(&poll->wait.entry);
5028 init_waitqueue_func_entry(&poll->wait, wake_func);
5031 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5032 struct wait_queue_head *head,
5033 struct io_poll_iocb **poll_ptr)
5035 struct io_kiocb *req = pt->req;
5038 * If poll->head is already set, it's because the file being polled
5039 * uses multiple waitqueues for poll handling (eg one for read, one
5040 * for write). Setup a separate io_poll_iocb if this happens.
5042 if (unlikely(poll->head)) {
5043 struct io_poll_iocb *poll_one = poll;
5045 /* already have a 2nd entry, fail a third attempt */
5047 pt->error = -EINVAL;
5050 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5052 pt->error = -ENOMEM;
5055 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5056 refcount_inc(&req->refs);
5057 poll->wait.private = req;
5064 if (poll->events & EPOLLEXCLUSIVE)
5065 add_wait_queue_exclusive(head, &poll->wait);
5067 add_wait_queue(head, &poll->wait);
5070 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5071 struct poll_table_struct *p)
5073 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5074 struct async_poll *apoll = pt->req->apoll;
5076 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5079 static void io_async_task_func(struct callback_head *cb)
5081 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5082 struct async_poll *apoll = req->apoll;
5083 struct io_ring_ctx *ctx = req->ctx;
5085 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5087 if (io_poll_rewait(req, &apoll->poll)) {
5088 spin_unlock_irq(&ctx->completion_lock);
5089 percpu_ref_put(&ctx->refs);
5093 /* If req is still hashed, it cannot have been canceled. Don't check. */
5094 if (hash_hashed(&req->hash_node))
5095 hash_del(&req->hash_node);
5097 io_poll_remove_double(req);
5098 spin_unlock_irq(&ctx->completion_lock);
5100 if (!READ_ONCE(apoll->poll.canceled))
5101 __io_req_task_submit(req);
5103 __io_req_task_cancel(req, -ECANCELED);
5105 percpu_ref_put(&ctx->refs);
5106 kfree(apoll->double_poll);
5110 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5113 struct io_kiocb *req = wait->private;
5114 struct io_poll_iocb *poll = &req->apoll->poll;
5116 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5119 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5122 static void io_poll_req_insert(struct io_kiocb *req)
5124 struct io_ring_ctx *ctx = req->ctx;
5125 struct hlist_head *list;
5127 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5128 hlist_add_head(&req->hash_node, list);
5131 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5132 struct io_poll_iocb *poll,
5133 struct io_poll_table *ipt, __poll_t mask,
5134 wait_queue_func_t wake_func)
5135 __acquires(&ctx->completion_lock)
5137 struct io_ring_ctx *ctx = req->ctx;
5138 bool cancel = false;
5140 INIT_HLIST_NODE(&req->hash_node);
5141 io_init_poll_iocb(poll, mask, wake_func);
5142 poll->file = req->file;
5143 poll->wait.private = req;
5145 ipt->pt._key = mask;
5147 ipt->error = -EINVAL;
5149 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5151 spin_lock_irq(&ctx->completion_lock);
5152 if (likely(poll->head)) {
5153 spin_lock(&poll->head->lock);
5154 if (unlikely(list_empty(&poll->wait.entry))) {
5160 if (mask || ipt->error)
5161 list_del_init(&poll->wait.entry);
5163 WRITE_ONCE(poll->canceled, true);
5164 else if (!poll->done) /* actually waiting for an event */
5165 io_poll_req_insert(req);
5166 spin_unlock(&poll->head->lock);
5172 static bool io_arm_poll_handler(struct io_kiocb *req)
5174 const struct io_op_def *def = &io_op_defs[req->opcode];
5175 struct io_ring_ctx *ctx = req->ctx;
5176 struct async_poll *apoll;
5177 struct io_poll_table ipt;
5181 if (!req->file || !file_can_poll(req->file))
5183 if (req->flags & REQ_F_POLLED)
5187 else if (def->pollout)
5191 /* if we can't nonblock try, then no point in arming a poll handler */
5192 if (!io_file_supports_async(req->file, rw))
5195 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5196 if (unlikely(!apoll))
5198 apoll->double_poll = NULL;
5200 req->flags |= REQ_F_POLLED;
5205 mask |= POLLIN | POLLRDNORM;
5207 mask |= POLLOUT | POLLWRNORM;
5209 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5210 if ((req->opcode == IORING_OP_RECVMSG) &&
5211 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5214 mask |= POLLERR | POLLPRI;
5216 ipt.pt._qproc = io_async_queue_proc;
5218 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5220 if (ret || ipt.error) {
5221 io_poll_remove_double(req);
5222 spin_unlock_irq(&ctx->completion_lock);
5223 kfree(apoll->double_poll);
5227 spin_unlock_irq(&ctx->completion_lock);
5228 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5229 apoll->poll.events);
5233 static bool __io_poll_remove_one(struct io_kiocb *req,
5234 struct io_poll_iocb *poll)
5236 bool do_complete = false;
5238 spin_lock(&poll->head->lock);
5239 WRITE_ONCE(poll->canceled, true);
5240 if (!list_empty(&poll->wait.entry)) {
5241 list_del_init(&poll->wait.entry);
5244 spin_unlock(&poll->head->lock);
5245 hash_del(&req->hash_node);
5249 static bool io_poll_remove_one(struct io_kiocb *req)
5253 io_poll_remove_double(req);
5255 if (req->opcode == IORING_OP_POLL_ADD) {
5256 do_complete = __io_poll_remove_one(req, &req->poll);
5258 struct async_poll *apoll = req->apoll;
5260 /* non-poll requests have submit ref still */
5261 do_complete = __io_poll_remove_one(req, &apoll->poll);
5264 kfree(apoll->double_poll);
5270 io_cqring_fill_event(req, -ECANCELED);
5271 io_commit_cqring(req->ctx);
5272 req_set_fail_links(req);
5273 io_put_req_deferred(req, 1);
5280 * Returns true if we found and killed one or more poll requests
5282 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk)
5284 struct hlist_node *tmp;
5285 struct io_kiocb *req;
5288 spin_lock_irq(&ctx->completion_lock);
5289 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5290 struct hlist_head *list;
5292 list = &ctx->cancel_hash[i];
5293 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5294 if (io_task_match(req, tsk))
5295 posted += io_poll_remove_one(req);
5298 spin_unlock_irq(&ctx->completion_lock);
5301 io_cqring_ev_posted(ctx);
5306 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5308 struct hlist_head *list;
5309 struct io_kiocb *req;
5311 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5312 hlist_for_each_entry(req, list, hash_node) {
5313 if (sqe_addr != req->user_data)
5315 if (io_poll_remove_one(req))
5323 static int io_poll_remove_prep(struct io_kiocb *req,
5324 const struct io_uring_sqe *sqe)
5326 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5328 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5332 req->poll.addr = READ_ONCE(sqe->addr);
5337 * Find a running poll command that matches one specified in sqe->addr,
5338 * and remove it if found.
5340 static int io_poll_remove(struct io_kiocb *req)
5342 struct io_ring_ctx *ctx = req->ctx;
5346 addr = req->poll.addr;
5347 spin_lock_irq(&ctx->completion_lock);
5348 ret = io_poll_cancel(ctx, addr);
5349 spin_unlock_irq(&ctx->completion_lock);
5352 req_set_fail_links(req);
5353 io_req_complete(req, ret);
5357 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5360 struct io_kiocb *req = wait->private;
5361 struct io_poll_iocb *poll = &req->poll;
5363 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5366 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5367 struct poll_table_struct *p)
5369 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5371 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5374 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5376 struct io_poll_iocb *poll = &req->poll;
5379 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5381 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5384 events = READ_ONCE(sqe->poll32_events);
5386 events = swahw32(events);
5388 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5389 (events & EPOLLEXCLUSIVE);
5393 static int io_poll_add(struct io_kiocb *req)
5395 struct io_poll_iocb *poll = &req->poll;
5396 struct io_ring_ctx *ctx = req->ctx;
5397 struct io_poll_table ipt;
5400 ipt.pt._qproc = io_poll_queue_proc;
5402 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5405 if (mask) { /* no async, we'd stolen it */
5407 io_poll_complete(req, mask, 0);
5409 spin_unlock_irq(&ctx->completion_lock);
5412 io_cqring_ev_posted(ctx);
5418 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5420 struct io_timeout_data *data = container_of(timer,
5421 struct io_timeout_data, timer);
5422 struct io_kiocb *req = data->req;
5423 struct io_ring_ctx *ctx = req->ctx;
5424 unsigned long flags;
5426 spin_lock_irqsave(&ctx->completion_lock, flags);
5427 list_del_init(&req->timeout.list);
5428 atomic_set(&req->ctx->cq_timeouts,
5429 atomic_read(&req->ctx->cq_timeouts) + 1);
5431 io_cqring_fill_event(req, -ETIME);
5432 io_commit_cqring(ctx);
5433 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5435 io_cqring_ev_posted(ctx);
5436 req_set_fail_links(req);
5438 return HRTIMER_NORESTART;
5441 static int __io_timeout_cancel(struct io_kiocb *req)
5443 struct io_timeout_data *io = req->async_data;
5446 ret = hrtimer_try_to_cancel(&io->timer);
5449 list_del_init(&req->timeout.list);
5451 req_set_fail_links(req);
5452 io_cqring_fill_event(req, -ECANCELED);
5453 io_put_req_deferred(req, 1);
5457 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5459 struct io_kiocb *req;
5462 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5463 if (user_data == req->user_data) {
5472 return __io_timeout_cancel(req);
5475 static int io_timeout_remove_prep(struct io_kiocb *req,
5476 const struct io_uring_sqe *sqe)
5478 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5480 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5482 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->timeout_flags)
5485 req->timeout_rem.addr = READ_ONCE(sqe->addr);
5490 * Remove or update an existing timeout command
5492 static int io_timeout_remove(struct io_kiocb *req)
5494 struct io_ring_ctx *ctx = req->ctx;
5497 spin_lock_irq(&ctx->completion_lock);
5498 ret = io_timeout_cancel(ctx, req->timeout_rem.addr);
5500 io_cqring_fill_event(req, ret);
5501 io_commit_cqring(ctx);
5502 spin_unlock_irq(&ctx->completion_lock);
5503 io_cqring_ev_posted(ctx);
5505 req_set_fail_links(req);
5510 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5511 bool is_timeout_link)
5513 struct io_timeout_data *data;
5515 u32 off = READ_ONCE(sqe->off);
5517 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5519 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5521 if (off && is_timeout_link)
5523 flags = READ_ONCE(sqe->timeout_flags);
5524 if (flags & ~IORING_TIMEOUT_ABS)
5527 req->timeout.off = off;
5529 if (!req->async_data && io_alloc_async_data(req))
5532 data = req->async_data;
5535 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5538 if (flags & IORING_TIMEOUT_ABS)
5539 data->mode = HRTIMER_MODE_ABS;
5541 data->mode = HRTIMER_MODE_REL;
5543 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5547 static int io_timeout(struct io_kiocb *req)
5549 struct io_ring_ctx *ctx = req->ctx;
5550 struct io_timeout_data *data = req->async_data;
5551 struct list_head *entry;
5552 u32 tail, off = req->timeout.off;
5554 spin_lock_irq(&ctx->completion_lock);
5557 * sqe->off holds how many events that need to occur for this
5558 * timeout event to be satisfied. If it isn't set, then this is
5559 * a pure timeout request, sequence isn't used.
5561 if (io_is_timeout_noseq(req)) {
5562 entry = ctx->timeout_list.prev;
5566 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5567 req->timeout.target_seq = tail + off;
5570 * Insertion sort, ensuring the first entry in the list is always
5571 * the one we need first.
5573 list_for_each_prev(entry, &ctx->timeout_list) {
5574 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5577 if (io_is_timeout_noseq(nxt))
5579 /* nxt.seq is behind @tail, otherwise would've been completed */
5580 if (off >= nxt->timeout.target_seq - tail)
5584 list_add(&req->timeout.list, entry);
5585 data->timer.function = io_timeout_fn;
5586 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5587 spin_unlock_irq(&ctx->completion_lock);
5591 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5593 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5595 return req->user_data == (unsigned long) data;
5598 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5600 enum io_wq_cancel cancel_ret;
5603 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5604 switch (cancel_ret) {
5605 case IO_WQ_CANCEL_OK:
5608 case IO_WQ_CANCEL_RUNNING:
5611 case IO_WQ_CANCEL_NOTFOUND:
5619 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5620 struct io_kiocb *req, __u64 sqe_addr,
5623 unsigned long flags;
5626 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5627 if (ret != -ENOENT) {
5628 spin_lock_irqsave(&ctx->completion_lock, flags);
5632 spin_lock_irqsave(&ctx->completion_lock, flags);
5633 ret = io_timeout_cancel(ctx, sqe_addr);
5636 ret = io_poll_cancel(ctx, sqe_addr);
5640 io_cqring_fill_event(req, ret);
5641 io_commit_cqring(ctx);
5642 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5643 io_cqring_ev_posted(ctx);
5646 req_set_fail_links(req);
5650 static int io_async_cancel_prep(struct io_kiocb *req,
5651 const struct io_uring_sqe *sqe)
5653 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5655 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5657 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5660 req->cancel.addr = READ_ONCE(sqe->addr);
5664 static int io_async_cancel(struct io_kiocb *req)
5666 struct io_ring_ctx *ctx = req->ctx;
5668 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5672 static int io_files_update_prep(struct io_kiocb *req,
5673 const struct io_uring_sqe *sqe)
5675 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5677 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5679 if (sqe->ioprio || sqe->rw_flags)
5682 req->files_update.offset = READ_ONCE(sqe->off);
5683 req->files_update.nr_args = READ_ONCE(sqe->len);
5684 if (!req->files_update.nr_args)
5686 req->files_update.arg = READ_ONCE(sqe->addr);
5690 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5691 struct io_comp_state *cs)
5693 struct io_ring_ctx *ctx = req->ctx;
5694 struct io_uring_files_update up;
5700 up.offset = req->files_update.offset;
5701 up.fds = req->files_update.arg;
5703 mutex_lock(&ctx->uring_lock);
5704 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5705 mutex_unlock(&ctx->uring_lock);
5708 req_set_fail_links(req);
5709 __io_req_complete(req, ret, 0, cs);
5713 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5715 switch (req->opcode) {
5718 case IORING_OP_READV:
5719 case IORING_OP_READ_FIXED:
5720 case IORING_OP_READ:
5721 return io_read_prep(req, sqe);
5722 case IORING_OP_WRITEV:
5723 case IORING_OP_WRITE_FIXED:
5724 case IORING_OP_WRITE:
5725 return io_write_prep(req, sqe);
5726 case IORING_OP_POLL_ADD:
5727 return io_poll_add_prep(req, sqe);
5728 case IORING_OP_POLL_REMOVE:
5729 return io_poll_remove_prep(req, sqe);
5730 case IORING_OP_FSYNC:
5731 return io_prep_fsync(req, sqe);
5732 case IORING_OP_SYNC_FILE_RANGE:
5733 return io_prep_sfr(req, sqe);
5734 case IORING_OP_SENDMSG:
5735 case IORING_OP_SEND:
5736 return io_sendmsg_prep(req, sqe);
5737 case IORING_OP_RECVMSG:
5738 case IORING_OP_RECV:
5739 return io_recvmsg_prep(req, sqe);
5740 case IORING_OP_CONNECT:
5741 return io_connect_prep(req, sqe);
5742 case IORING_OP_TIMEOUT:
5743 return io_timeout_prep(req, sqe, false);
5744 case IORING_OP_TIMEOUT_REMOVE:
5745 return io_timeout_remove_prep(req, sqe);
5746 case IORING_OP_ASYNC_CANCEL:
5747 return io_async_cancel_prep(req, sqe);
5748 case IORING_OP_LINK_TIMEOUT:
5749 return io_timeout_prep(req, sqe, true);
5750 case IORING_OP_ACCEPT:
5751 return io_accept_prep(req, sqe);
5752 case IORING_OP_FALLOCATE:
5753 return io_fallocate_prep(req, sqe);
5754 case IORING_OP_OPENAT:
5755 return io_openat_prep(req, sqe);
5756 case IORING_OP_CLOSE:
5757 return io_close_prep(req, sqe);
5758 case IORING_OP_FILES_UPDATE:
5759 return io_files_update_prep(req, sqe);
5760 case IORING_OP_STATX:
5761 return io_statx_prep(req, sqe);
5762 case IORING_OP_FADVISE:
5763 return io_fadvise_prep(req, sqe);
5764 case IORING_OP_MADVISE:
5765 return io_madvise_prep(req, sqe);
5766 case IORING_OP_OPENAT2:
5767 return io_openat2_prep(req, sqe);
5768 case IORING_OP_EPOLL_CTL:
5769 return io_epoll_ctl_prep(req, sqe);
5770 case IORING_OP_SPLICE:
5771 return io_splice_prep(req, sqe);
5772 case IORING_OP_PROVIDE_BUFFERS:
5773 return io_provide_buffers_prep(req, sqe);
5774 case IORING_OP_REMOVE_BUFFERS:
5775 return io_remove_buffers_prep(req, sqe);
5777 return io_tee_prep(req, sqe);
5780 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5785 static int io_req_defer_prep(struct io_kiocb *req,
5786 const struct io_uring_sqe *sqe)
5790 if (io_alloc_async_data(req))
5792 return io_req_prep(req, sqe);
5795 static u32 io_get_sequence(struct io_kiocb *req)
5797 struct io_kiocb *pos;
5798 struct io_ring_ctx *ctx = req->ctx;
5799 u32 total_submitted, nr_reqs = 1;
5801 if (req->flags & REQ_F_LINK_HEAD)
5802 list_for_each_entry(pos, &req->link_list, link_list)
5805 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5806 return total_submitted - nr_reqs;
5809 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5811 struct io_ring_ctx *ctx = req->ctx;
5812 struct io_defer_entry *de;
5816 /* Still need defer if there is pending req in defer list. */
5817 if (likely(list_empty_careful(&ctx->defer_list) &&
5818 !(req->flags & REQ_F_IO_DRAIN)))
5821 seq = io_get_sequence(req);
5822 /* Still a chance to pass the sequence check */
5823 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5826 if (!req->async_data) {
5827 ret = io_req_defer_prep(req, sqe);
5831 io_prep_async_link(req);
5832 de = kmalloc(sizeof(*de), GFP_KERNEL);
5836 spin_lock_irq(&ctx->completion_lock);
5837 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5838 spin_unlock_irq(&ctx->completion_lock);
5840 io_queue_async_work(req);
5841 return -EIOCBQUEUED;
5844 trace_io_uring_defer(ctx, req, req->user_data);
5847 list_add_tail(&de->list, &ctx->defer_list);
5848 spin_unlock_irq(&ctx->completion_lock);
5849 return -EIOCBQUEUED;
5852 static void io_req_drop_files(struct io_kiocb *req)
5854 struct io_ring_ctx *ctx = req->ctx;
5855 unsigned long flags;
5857 spin_lock_irqsave(&ctx->inflight_lock, flags);
5858 list_del(&req->inflight_entry);
5859 if (waitqueue_active(&ctx->inflight_wait))
5860 wake_up(&ctx->inflight_wait);
5861 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5862 req->flags &= ~REQ_F_INFLIGHT;
5863 put_files_struct(req->work.identity->files);
5864 put_nsproxy(req->work.identity->nsproxy);
5865 req->work.flags &= ~IO_WQ_WORK_FILES;
5868 static void __io_clean_op(struct io_kiocb *req)
5870 if (req->flags & REQ_F_BUFFER_SELECTED) {
5871 switch (req->opcode) {
5872 case IORING_OP_READV:
5873 case IORING_OP_READ_FIXED:
5874 case IORING_OP_READ:
5875 kfree((void *)(unsigned long)req->rw.addr);
5877 case IORING_OP_RECVMSG:
5878 case IORING_OP_RECV:
5879 kfree(req->sr_msg.kbuf);
5882 req->flags &= ~REQ_F_BUFFER_SELECTED;
5885 if (req->flags & REQ_F_NEED_CLEANUP) {
5886 switch (req->opcode) {
5887 case IORING_OP_READV:
5888 case IORING_OP_READ_FIXED:
5889 case IORING_OP_READ:
5890 case IORING_OP_WRITEV:
5891 case IORING_OP_WRITE_FIXED:
5892 case IORING_OP_WRITE: {
5893 struct io_async_rw *io = req->async_data;
5895 kfree(io->free_iovec);
5898 case IORING_OP_RECVMSG:
5899 case IORING_OP_SENDMSG: {
5900 struct io_async_msghdr *io = req->async_data;
5901 if (io->iov != io->fast_iov)
5905 case IORING_OP_SPLICE:
5907 io_put_file(req, req->splice.file_in,
5908 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5910 case IORING_OP_OPENAT:
5911 case IORING_OP_OPENAT2:
5912 if (req->open.filename)
5913 putname(req->open.filename);
5916 req->flags &= ~REQ_F_NEED_CLEANUP;
5919 if (req->flags & REQ_F_INFLIGHT)
5920 io_req_drop_files(req);
5923 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
5924 struct io_comp_state *cs)
5926 struct io_ring_ctx *ctx = req->ctx;
5929 switch (req->opcode) {
5931 ret = io_nop(req, cs);
5933 case IORING_OP_READV:
5934 case IORING_OP_READ_FIXED:
5935 case IORING_OP_READ:
5936 ret = io_read(req, force_nonblock, cs);
5938 case IORING_OP_WRITEV:
5939 case IORING_OP_WRITE_FIXED:
5940 case IORING_OP_WRITE:
5941 ret = io_write(req, force_nonblock, cs);
5943 case IORING_OP_FSYNC:
5944 ret = io_fsync(req, force_nonblock);
5946 case IORING_OP_POLL_ADD:
5947 ret = io_poll_add(req);
5949 case IORING_OP_POLL_REMOVE:
5950 ret = io_poll_remove(req);
5952 case IORING_OP_SYNC_FILE_RANGE:
5953 ret = io_sync_file_range(req, force_nonblock);
5955 case IORING_OP_SENDMSG:
5956 ret = io_sendmsg(req, force_nonblock, cs);
5958 case IORING_OP_SEND:
5959 ret = io_send(req, force_nonblock, cs);
5961 case IORING_OP_RECVMSG:
5962 ret = io_recvmsg(req, force_nonblock, cs);
5964 case IORING_OP_RECV:
5965 ret = io_recv(req, force_nonblock, cs);
5967 case IORING_OP_TIMEOUT:
5968 ret = io_timeout(req);
5970 case IORING_OP_TIMEOUT_REMOVE:
5971 ret = io_timeout_remove(req);
5973 case IORING_OP_ACCEPT:
5974 ret = io_accept(req, force_nonblock, cs);
5976 case IORING_OP_CONNECT:
5977 ret = io_connect(req, force_nonblock, cs);
5979 case IORING_OP_ASYNC_CANCEL:
5980 ret = io_async_cancel(req);
5982 case IORING_OP_FALLOCATE:
5983 ret = io_fallocate(req, force_nonblock);
5985 case IORING_OP_OPENAT:
5986 ret = io_openat(req, force_nonblock);
5988 case IORING_OP_CLOSE:
5989 ret = io_close(req, force_nonblock, cs);
5991 case IORING_OP_FILES_UPDATE:
5992 ret = io_files_update(req, force_nonblock, cs);
5994 case IORING_OP_STATX:
5995 ret = io_statx(req, force_nonblock);
5997 case IORING_OP_FADVISE:
5998 ret = io_fadvise(req, force_nonblock);
6000 case IORING_OP_MADVISE:
6001 ret = io_madvise(req, force_nonblock);
6003 case IORING_OP_OPENAT2:
6004 ret = io_openat2(req, force_nonblock);
6006 case IORING_OP_EPOLL_CTL:
6007 ret = io_epoll_ctl(req, force_nonblock, cs);
6009 case IORING_OP_SPLICE:
6010 ret = io_splice(req, force_nonblock);
6012 case IORING_OP_PROVIDE_BUFFERS:
6013 ret = io_provide_buffers(req, force_nonblock, cs);
6015 case IORING_OP_REMOVE_BUFFERS:
6016 ret = io_remove_buffers(req, force_nonblock, cs);
6019 ret = io_tee(req, force_nonblock);
6029 /* If the op doesn't have a file, we're not polling for it */
6030 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6031 const bool in_async = io_wq_current_is_worker();
6033 /* workqueue context doesn't hold uring_lock, grab it now */
6035 mutex_lock(&ctx->uring_lock);
6037 io_iopoll_req_issued(req);
6040 mutex_unlock(&ctx->uring_lock);
6046 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
6048 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6049 struct io_kiocb *timeout;
6052 timeout = io_prep_linked_timeout(req);
6054 io_queue_linked_timeout(timeout);
6056 /* if NO_CANCEL is set, we must still run the work */
6057 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6058 IO_WQ_WORK_CANCEL) {
6064 ret = io_issue_sqe(req, false, NULL);
6066 * We can get EAGAIN for polled IO even though we're
6067 * forcing a sync submission from here, since we can't
6068 * wait for request slots on the block side.
6077 req_set_fail_links(req);
6078 io_req_complete(req, ret);
6081 return io_steal_work(req);
6084 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6087 struct fixed_file_table *table;
6089 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6090 return table->files[index & IORING_FILE_TABLE_MASK];
6093 static struct file *io_file_get(struct io_submit_state *state,
6094 struct io_kiocb *req, int fd, bool fixed)
6096 struct io_ring_ctx *ctx = req->ctx;
6100 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6102 fd = array_index_nospec(fd, ctx->nr_user_files);
6103 file = io_file_from_index(ctx, fd);
6105 req->fixed_file_refs = &ctx->file_data->node->refs;
6106 percpu_ref_get(req->fixed_file_refs);
6109 trace_io_uring_file_get(ctx, fd);
6110 file = __io_file_get(state, fd);
6116 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6121 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6122 if (unlikely(!fixed && io_async_submit(req->ctx)))
6125 req->file = io_file_get(state, req, fd, fixed);
6126 if (req->file || io_op_defs[req->opcode].needs_file_no_error)
6131 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6133 struct io_timeout_data *data = container_of(timer,
6134 struct io_timeout_data, timer);
6135 struct io_kiocb *req = data->req;
6136 struct io_ring_ctx *ctx = req->ctx;
6137 struct io_kiocb *prev = NULL;
6138 unsigned long flags;
6140 spin_lock_irqsave(&ctx->completion_lock, flags);
6143 * We don't expect the list to be empty, that will only happen if we
6144 * race with the completion of the linked work.
6146 if (!list_empty(&req->link_list)) {
6147 prev = list_entry(req->link_list.prev, struct io_kiocb,
6149 if (refcount_inc_not_zero(&prev->refs))
6150 list_del_init(&req->link_list);
6155 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6158 req_set_fail_links(prev);
6159 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6162 io_req_complete(req, -ETIME);
6164 return HRTIMER_NORESTART;
6167 static void __io_queue_linked_timeout(struct io_kiocb *req)
6170 * If the list is now empty, then our linked request finished before
6171 * we got a chance to setup the timer
6173 if (!list_empty(&req->link_list)) {
6174 struct io_timeout_data *data = req->async_data;
6176 data->timer.function = io_link_timeout_fn;
6177 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6182 static void io_queue_linked_timeout(struct io_kiocb *req)
6184 struct io_ring_ctx *ctx = req->ctx;
6186 spin_lock_irq(&ctx->completion_lock);
6187 __io_queue_linked_timeout(req);
6188 spin_unlock_irq(&ctx->completion_lock);
6190 /* drop submission reference */
6194 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6196 struct io_kiocb *nxt;
6198 if (!(req->flags & REQ_F_LINK_HEAD))
6200 if (req->flags & REQ_F_LINK_TIMEOUT)
6203 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6205 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6208 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6209 req->flags |= REQ_F_LINK_TIMEOUT;
6213 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6215 struct io_kiocb *linked_timeout;
6216 const struct cred *old_creds = NULL;
6220 linked_timeout = io_prep_linked_timeout(req);
6222 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6223 (req->work.flags & IO_WQ_WORK_CREDS) &&
6224 req->work.identity->creds != current_cred()) {
6226 revert_creds(old_creds);
6227 if (old_creds == req->work.identity->creds)
6228 old_creds = NULL; /* restored original creds */
6230 old_creds = override_creds(req->work.identity->creds);
6233 ret = io_issue_sqe(req, true, cs);
6236 * We async punt it if the file wasn't marked NOWAIT, or if the file
6237 * doesn't support non-blocking read/write attempts
6239 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6240 if (!io_arm_poll_handler(req)) {
6242 * Queued up for async execution, worker will release
6243 * submit reference when the iocb is actually submitted.
6245 io_queue_async_work(req);
6249 io_queue_linked_timeout(linked_timeout);
6250 } else if (likely(!ret)) {
6251 /* drop submission reference */
6252 req = io_put_req_find_next(req);
6254 io_queue_linked_timeout(linked_timeout);
6257 if (!(req->flags & REQ_F_FORCE_ASYNC))
6259 io_queue_async_work(req);
6262 /* un-prep timeout, so it'll be killed as any other linked */
6263 req->flags &= ~REQ_F_LINK_TIMEOUT;
6264 req_set_fail_links(req);
6266 io_req_complete(req, ret);
6270 revert_creds(old_creds);
6273 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6274 struct io_comp_state *cs)
6278 ret = io_req_defer(req, sqe);
6280 if (ret != -EIOCBQUEUED) {
6282 req_set_fail_links(req);
6284 io_req_complete(req, ret);
6286 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6287 if (!req->async_data) {
6288 ret = io_req_defer_prep(req, sqe);
6292 io_queue_async_work(req);
6295 ret = io_req_prep(req, sqe);
6299 __io_queue_sqe(req, cs);
6303 static inline void io_queue_link_head(struct io_kiocb *req,
6304 struct io_comp_state *cs)
6306 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6308 io_req_complete(req, -ECANCELED);
6310 io_queue_sqe(req, NULL, cs);
6313 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6314 struct io_kiocb **link, struct io_comp_state *cs)
6316 struct io_ring_ctx *ctx = req->ctx;
6320 * If we already have a head request, queue this one for async
6321 * submittal once the head completes. If we don't have a head but
6322 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6323 * submitted sync once the chain is complete. If none of those
6324 * conditions are true (normal request), then just queue it.
6327 struct io_kiocb *head = *link;
6330 * Taking sequential execution of a link, draining both sides
6331 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6332 * requests in the link. So, it drains the head and the
6333 * next after the link request. The last one is done via
6334 * drain_next flag to persist the effect across calls.
6336 if (req->flags & REQ_F_IO_DRAIN) {
6337 head->flags |= REQ_F_IO_DRAIN;
6338 ctx->drain_next = 1;
6340 ret = io_req_defer_prep(req, sqe);
6341 if (unlikely(ret)) {
6342 /* fail even hard links since we don't submit */
6343 head->flags |= REQ_F_FAIL_LINK;
6346 trace_io_uring_link(ctx, req, head);
6347 list_add_tail(&req->link_list, &head->link_list);
6349 /* last request of a link, enqueue the link */
6350 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6351 io_queue_link_head(head, cs);
6355 if (unlikely(ctx->drain_next)) {
6356 req->flags |= REQ_F_IO_DRAIN;
6357 ctx->drain_next = 0;
6359 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6360 req->flags |= REQ_F_LINK_HEAD;
6361 INIT_LIST_HEAD(&req->link_list);
6363 ret = io_req_defer_prep(req, sqe);
6365 req->flags |= REQ_F_FAIL_LINK;
6368 io_queue_sqe(req, sqe, cs);
6376 * Batched submission is done, ensure local IO is flushed out.
6378 static void io_submit_state_end(struct io_submit_state *state)
6380 if (!list_empty(&state->comp.list))
6381 io_submit_flush_completions(&state->comp);
6382 blk_finish_plug(&state->plug);
6383 io_state_file_put(state);
6384 if (state->free_reqs)
6385 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6389 * Start submission side cache.
6391 static void io_submit_state_start(struct io_submit_state *state,
6392 struct io_ring_ctx *ctx, unsigned int max_ios)
6394 blk_start_plug(&state->plug);
6396 INIT_LIST_HEAD(&state->comp.list);
6397 state->comp.ctx = ctx;
6398 state->free_reqs = 0;
6400 state->ios_left = max_ios;
6403 static void io_commit_sqring(struct io_ring_ctx *ctx)
6405 struct io_rings *rings = ctx->rings;
6408 * Ensure any loads from the SQEs are done at this point,
6409 * since once we write the new head, the application could
6410 * write new data to them.
6412 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6416 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6417 * that is mapped by userspace. This means that care needs to be taken to
6418 * ensure that reads are stable, as we cannot rely on userspace always
6419 * being a good citizen. If members of the sqe are validated and then later
6420 * used, it's important that those reads are done through READ_ONCE() to
6421 * prevent a re-load down the line.
6423 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6425 u32 *sq_array = ctx->sq_array;
6429 * The cached sq head (or cq tail) serves two purposes:
6431 * 1) allows us to batch the cost of updating the user visible
6433 * 2) allows the kernel side to track the head on its own, even
6434 * though the application is the one updating it.
6436 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6437 if (likely(head < ctx->sq_entries))
6438 return &ctx->sq_sqes[head];
6440 /* drop invalid entries */
6441 ctx->cached_sq_dropped++;
6442 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6446 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6448 ctx->cached_sq_head++;
6452 * Check SQE restrictions (opcode and flags).
6454 * Returns 'true' if SQE is allowed, 'false' otherwise.
6456 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6457 struct io_kiocb *req,
6458 unsigned int sqe_flags)
6460 if (!ctx->restricted)
6463 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6466 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6467 ctx->restrictions.sqe_flags_required)
6470 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6471 ctx->restrictions.sqe_flags_required))
6477 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6478 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6479 IOSQE_BUFFER_SELECT)
6481 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6482 const struct io_uring_sqe *sqe,
6483 struct io_submit_state *state)
6485 unsigned int sqe_flags;
6488 req->opcode = READ_ONCE(sqe->opcode);
6489 req->user_data = READ_ONCE(sqe->user_data);
6490 req->async_data = NULL;
6494 /* one is dropped after submission, the other at completion */
6495 refcount_set(&req->refs, 2);
6496 req->task = current;
6499 if (unlikely(req->opcode >= IORING_OP_LAST))
6502 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6505 sqe_flags = READ_ONCE(sqe->flags);
6506 /* enforce forwards compatibility on users */
6507 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6510 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6513 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6514 !io_op_defs[req->opcode].buffer_select)
6517 id = READ_ONCE(sqe->personality);
6519 struct io_identity *iod;
6521 iod = idr_find(&ctx->personality_idr, id);
6524 refcount_inc(&iod->count);
6526 __io_req_init_async(req);
6527 get_cred(iod->creds);
6528 req->work.identity = iod;
6529 req->work.flags |= IO_WQ_WORK_CREDS;
6532 /* same numerical values with corresponding REQ_F_*, safe to copy */
6533 req->flags |= sqe_flags;
6535 if (!io_op_defs[req->opcode].needs_file)
6538 ret = io_req_set_file(state, req, READ_ONCE(sqe->fd));
6543 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6545 struct io_submit_state state;
6546 struct io_kiocb *link = NULL;
6547 int i, submitted = 0;
6549 /* if we have a backlog and couldn't flush it all, return BUSY */
6550 if (test_bit(0, &ctx->sq_check_overflow)) {
6551 if (!list_empty(&ctx->cq_overflow_list) &&
6552 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6556 /* make sure SQ entry isn't read before tail */
6557 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6559 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6562 percpu_counter_add(¤t->io_uring->inflight, nr);
6563 refcount_add(nr, ¤t->usage);
6565 io_submit_state_start(&state, ctx, nr);
6567 for (i = 0; i < nr; i++) {
6568 const struct io_uring_sqe *sqe;
6569 struct io_kiocb *req;
6572 sqe = io_get_sqe(ctx);
6573 if (unlikely(!sqe)) {
6574 io_consume_sqe(ctx);
6577 req = io_alloc_req(ctx, &state);
6578 if (unlikely(!req)) {
6580 submitted = -EAGAIN;
6583 io_consume_sqe(ctx);
6584 /* will complete beyond this point, count as submitted */
6587 err = io_init_req(ctx, req, sqe, &state);
6588 if (unlikely(err)) {
6591 io_req_complete(req, err);
6595 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6596 true, io_async_submit(ctx));
6597 err = io_submit_sqe(req, sqe, &link, &state.comp);
6602 if (unlikely(submitted != nr)) {
6603 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6604 struct io_uring_task *tctx = current->io_uring;
6605 int unused = nr - ref_used;
6607 percpu_ref_put_many(&ctx->refs, unused);
6608 percpu_counter_sub(&tctx->inflight, unused);
6609 put_task_struct_many(current, unused);
6612 io_queue_link_head(link, &state.comp);
6613 io_submit_state_end(&state);
6615 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6616 io_commit_sqring(ctx);
6621 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6623 /* Tell userspace we may need a wakeup call */
6624 spin_lock_irq(&ctx->completion_lock);
6625 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6626 spin_unlock_irq(&ctx->completion_lock);
6629 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6631 spin_lock_irq(&ctx->completion_lock);
6632 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6633 spin_unlock_irq(&ctx->completion_lock);
6636 static int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode,
6637 int sync, void *key)
6639 struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry);
6642 ret = autoremove_wake_function(wqe, mode, sync, key);
6644 unsigned long flags;
6646 spin_lock_irqsave(&ctx->completion_lock, flags);
6647 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6648 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6659 static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx,
6660 unsigned long start_jiffies, bool cap_entries)
6662 unsigned long timeout = start_jiffies + ctx->sq_thread_idle;
6663 struct io_sq_data *sqd = ctx->sq_data;
6664 unsigned int to_submit;
6668 if (!list_empty(&ctx->iopoll_list)) {
6669 unsigned nr_events = 0;
6671 mutex_lock(&ctx->uring_lock);
6672 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6673 io_do_iopoll(ctx, &nr_events, 0);
6674 mutex_unlock(&ctx->uring_lock);
6677 to_submit = io_sqring_entries(ctx);
6680 * If submit got -EBUSY, flag us as needing the application
6681 * to enter the kernel to reap and flush events.
6683 if (!to_submit || ret == -EBUSY || need_resched()) {
6685 * Drop cur_mm before scheduling, we can't hold it for
6686 * long periods (or over schedule()). Do this before
6687 * adding ourselves to the waitqueue, as the unuse/drop
6690 io_sq_thread_drop_mm();
6693 * We're polling. If we're within the defined idle
6694 * period, then let us spin without work before going
6695 * to sleep. The exception is if we got EBUSY doing
6696 * more IO, we should wait for the application to
6697 * reap events and wake us up.
6699 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6700 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6701 !percpu_ref_is_dying(&ctx->refs)))
6704 prepare_to_wait(&sqd->wait, &ctx->sqo_wait_entry,
6705 TASK_INTERRUPTIBLE);
6708 * While doing polled IO, before going to sleep, we need
6709 * to check if there are new reqs added to iopoll_list,
6710 * it is because reqs may have been punted to io worker
6711 * and will be added to iopoll_list later, hence check
6712 * the iopoll_list again.
6714 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6715 !list_empty_careful(&ctx->iopoll_list)) {
6716 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6720 to_submit = io_sqring_entries(ctx);
6721 if (!to_submit || ret == -EBUSY)
6725 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6726 io_ring_clear_wakeup_flag(ctx);
6728 /* if we're handling multiple rings, cap submit size for fairness */
6729 if (cap_entries && to_submit > 8)
6732 mutex_lock(&ctx->uring_lock);
6733 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6734 ret = io_submit_sqes(ctx, to_submit);
6735 mutex_unlock(&ctx->uring_lock);
6737 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6738 wake_up(&ctx->sqo_sq_wait);
6740 return SQT_DID_WORK;
6743 static void io_sqd_init_new(struct io_sq_data *sqd)
6745 struct io_ring_ctx *ctx;
6747 while (!list_empty(&sqd->ctx_new_list)) {
6748 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6749 init_wait(&ctx->sqo_wait_entry);
6750 ctx->sqo_wait_entry.func = io_sq_wake_function;
6751 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6752 complete(&ctx->sq_thread_comp);
6756 static int io_sq_thread(void *data)
6758 struct cgroup_subsys_state *cur_css = NULL;
6759 const struct cred *old_cred = NULL;
6760 struct io_sq_data *sqd = data;
6761 struct io_ring_ctx *ctx;
6762 unsigned long start_jiffies;
6764 start_jiffies = jiffies;
6765 while (!kthread_should_stop()) {
6766 enum sq_ret ret = 0;
6770 * Any changes to the sqd lists are synchronized through the
6771 * kthread parking. This synchronizes the thread vs users,
6772 * the users are synchronized on the sqd->ctx_lock.
6774 if (kthread_should_park())
6777 if (unlikely(!list_empty(&sqd->ctx_new_list)))
6778 io_sqd_init_new(sqd);
6780 cap_entries = !list_is_singular(&sqd->ctx_list);
6782 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6783 if (current->cred != ctx->creds) {
6785 revert_creds(old_cred);
6786 old_cred = override_creds(ctx->creds);
6788 io_sq_thread_associate_blkcg(ctx, &cur_css);
6790 current->loginuid = ctx->loginuid;
6791 current->sessionid = ctx->sessionid;
6794 ret |= __io_sq_thread(ctx, start_jiffies, cap_entries);
6796 io_sq_thread_drop_mm();
6799 if (ret & SQT_SPIN) {
6802 } else if (ret == SQT_IDLE) {
6803 if (kthread_should_park())
6805 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6806 io_ring_set_wakeup_flag(ctx);
6808 start_jiffies = jiffies;
6809 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6810 io_ring_clear_wakeup_flag(ctx);
6817 io_sq_thread_unassociate_blkcg();
6819 revert_creds(old_cred);
6826 struct io_wait_queue {
6827 struct wait_queue_entry wq;
6828 struct io_ring_ctx *ctx;
6830 unsigned nr_timeouts;
6833 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6835 struct io_ring_ctx *ctx = iowq->ctx;
6838 * Wake up if we have enough events, or if a timeout occurred since we
6839 * started waiting. For timeouts, we always want to return to userspace,
6840 * regardless of event count.
6842 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6843 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6846 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6847 int wake_flags, void *key)
6849 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6852 /* use noflush == true, as we can't safely rely on locking context */
6853 if (!io_should_wake(iowq, true))
6856 return autoremove_wake_function(curr, mode, wake_flags, key);
6859 static int io_run_task_work_sig(void)
6861 if (io_run_task_work())
6863 if (!signal_pending(current))
6865 if (current->jobctl & JOBCTL_TASK_WORK) {
6866 spin_lock_irq(¤t->sighand->siglock);
6867 current->jobctl &= ~JOBCTL_TASK_WORK;
6868 recalc_sigpending();
6869 spin_unlock_irq(¤t->sighand->siglock);
6876 * Wait until events become available, if we don't already have some. The
6877 * application must reap them itself, as they reside on the shared cq ring.
6879 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6880 const sigset_t __user *sig, size_t sigsz)
6882 struct io_wait_queue iowq = {
6885 .func = io_wake_function,
6886 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6889 .to_wait = min_events,
6891 struct io_rings *rings = ctx->rings;
6895 if (io_cqring_events(ctx, false) >= min_events)
6897 if (!io_run_task_work())
6902 #ifdef CONFIG_COMPAT
6903 if (in_compat_syscall())
6904 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6908 ret = set_user_sigmask(sig, sigsz);
6914 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6915 trace_io_uring_cqring_wait(ctx, min_events);
6917 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6918 TASK_INTERRUPTIBLE);
6919 /* make sure we run task_work before checking for signals */
6920 ret = io_run_task_work_sig();
6925 if (io_should_wake(&iowq, false))
6929 finish_wait(&ctx->wait, &iowq.wq);
6931 restore_saved_sigmask_unless(ret == -EINTR);
6933 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6936 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6938 #if defined(CONFIG_UNIX)
6939 if (ctx->ring_sock) {
6940 struct sock *sock = ctx->ring_sock->sk;
6941 struct sk_buff *skb;
6943 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6949 for (i = 0; i < ctx->nr_user_files; i++) {
6952 file = io_file_from_index(ctx, i);
6959 static void io_file_ref_kill(struct percpu_ref *ref)
6961 struct fixed_file_data *data;
6963 data = container_of(ref, struct fixed_file_data, refs);
6964 complete(&data->done);
6967 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6969 struct fixed_file_data *data = ctx->file_data;
6970 struct fixed_file_ref_node *ref_node = NULL;
6971 unsigned nr_tables, i;
6976 spin_lock(&data->lock);
6977 ref_node = data->node;
6978 spin_unlock(&data->lock);
6980 percpu_ref_kill(&ref_node->refs);
6982 percpu_ref_kill(&data->refs);
6984 /* wait for all refs nodes to complete */
6985 flush_delayed_work(&ctx->file_put_work);
6986 wait_for_completion(&data->done);
6988 __io_sqe_files_unregister(ctx);
6989 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6990 for (i = 0; i < nr_tables; i++)
6991 kfree(data->table[i].files);
6993 percpu_ref_exit(&data->refs);
6995 ctx->file_data = NULL;
6996 ctx->nr_user_files = 0;
7000 static void io_put_sq_data(struct io_sq_data *sqd)
7002 if (refcount_dec_and_test(&sqd->refs)) {
7004 * The park is a bit of a work-around, without it we get
7005 * warning spews on shutdown with SQPOLL set and affinity
7006 * set to a single CPU.
7009 kthread_park(sqd->thread);
7010 kthread_stop(sqd->thread);
7017 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7019 struct io_ring_ctx *ctx_attach;
7020 struct io_sq_data *sqd;
7023 f = fdget(p->wq_fd);
7025 return ERR_PTR(-ENXIO);
7026 if (f.file->f_op != &io_uring_fops) {
7028 return ERR_PTR(-EINVAL);
7031 ctx_attach = f.file->private_data;
7032 sqd = ctx_attach->sq_data;
7035 return ERR_PTR(-EINVAL);
7038 refcount_inc(&sqd->refs);
7043 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7045 struct io_sq_data *sqd;
7047 if (p->flags & IORING_SETUP_ATTACH_WQ)
7048 return io_attach_sq_data(p);
7050 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7052 return ERR_PTR(-ENOMEM);
7054 refcount_set(&sqd->refs, 1);
7055 INIT_LIST_HEAD(&sqd->ctx_list);
7056 INIT_LIST_HEAD(&sqd->ctx_new_list);
7057 mutex_init(&sqd->ctx_lock);
7058 mutex_init(&sqd->lock);
7059 init_waitqueue_head(&sqd->wait);
7063 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7064 __releases(&sqd->lock)
7068 kthread_unpark(sqd->thread);
7069 mutex_unlock(&sqd->lock);
7072 static void io_sq_thread_park(struct io_sq_data *sqd)
7073 __acquires(&sqd->lock)
7077 mutex_lock(&sqd->lock);
7078 kthread_park(sqd->thread);
7081 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7083 struct io_sq_data *sqd = ctx->sq_data;
7088 * We may arrive here from the error branch in
7089 * io_sq_offload_create() where the kthread is created
7090 * without being waked up, thus wake it up now to make
7091 * sure the wait will complete.
7093 wake_up_process(sqd->thread);
7094 wait_for_completion(&ctx->sq_thread_comp);
7096 io_sq_thread_park(sqd);
7099 mutex_lock(&sqd->ctx_lock);
7100 list_del(&ctx->sqd_list);
7101 mutex_unlock(&sqd->ctx_lock);
7104 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
7105 io_sq_thread_unpark(sqd);
7108 io_put_sq_data(sqd);
7109 ctx->sq_data = NULL;
7113 static void io_finish_async(struct io_ring_ctx *ctx)
7115 io_sq_thread_stop(ctx);
7118 io_wq_destroy(ctx->io_wq);
7123 #if defined(CONFIG_UNIX)
7125 * Ensure the UNIX gc is aware of our file set, so we are certain that
7126 * the io_uring can be safely unregistered on process exit, even if we have
7127 * loops in the file referencing.
7129 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7131 struct sock *sk = ctx->ring_sock->sk;
7132 struct scm_fp_list *fpl;
7133 struct sk_buff *skb;
7136 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7140 skb = alloc_skb(0, GFP_KERNEL);
7149 fpl->user = get_uid(ctx->user);
7150 for (i = 0; i < nr; i++) {
7151 struct file *file = io_file_from_index(ctx, i + offset);
7155 fpl->fp[nr_files] = get_file(file);
7156 unix_inflight(fpl->user, fpl->fp[nr_files]);
7161 fpl->max = SCM_MAX_FD;
7162 fpl->count = nr_files;
7163 UNIXCB(skb).fp = fpl;
7164 skb->destructor = unix_destruct_scm;
7165 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7166 skb_queue_head(&sk->sk_receive_queue, skb);
7168 for (i = 0; i < nr_files; i++)
7179 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7180 * causes regular reference counting to break down. We rely on the UNIX
7181 * garbage collection to take care of this problem for us.
7183 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7185 unsigned left, total;
7189 left = ctx->nr_user_files;
7191 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7193 ret = __io_sqe_files_scm(ctx, this_files, total);
7197 total += this_files;
7203 while (total < ctx->nr_user_files) {
7204 struct file *file = io_file_from_index(ctx, total);
7214 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7220 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7221 unsigned nr_tables, unsigned nr_files)
7225 for (i = 0; i < nr_tables; i++) {
7226 struct fixed_file_table *table = &file_data->table[i];
7227 unsigned this_files;
7229 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7230 table->files = kcalloc(this_files, sizeof(struct file *),
7234 nr_files -= this_files;
7240 for (i = 0; i < nr_tables; i++) {
7241 struct fixed_file_table *table = &file_data->table[i];
7242 kfree(table->files);
7247 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7249 #if defined(CONFIG_UNIX)
7250 struct sock *sock = ctx->ring_sock->sk;
7251 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7252 struct sk_buff *skb;
7255 __skb_queue_head_init(&list);
7258 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7259 * remove this entry and rearrange the file array.
7261 skb = skb_dequeue(head);
7263 struct scm_fp_list *fp;
7265 fp = UNIXCB(skb).fp;
7266 for (i = 0; i < fp->count; i++) {
7269 if (fp->fp[i] != file)
7272 unix_notinflight(fp->user, fp->fp[i]);
7273 left = fp->count - 1 - i;
7275 memmove(&fp->fp[i], &fp->fp[i + 1],
7276 left * sizeof(struct file *));
7283 __skb_queue_tail(&list, skb);
7293 __skb_queue_tail(&list, skb);
7295 skb = skb_dequeue(head);
7298 if (skb_peek(&list)) {
7299 spin_lock_irq(&head->lock);
7300 while ((skb = __skb_dequeue(&list)) != NULL)
7301 __skb_queue_tail(head, skb);
7302 spin_unlock_irq(&head->lock);
7309 struct io_file_put {
7310 struct list_head list;
7314 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7316 struct fixed_file_data *file_data = ref_node->file_data;
7317 struct io_ring_ctx *ctx = file_data->ctx;
7318 struct io_file_put *pfile, *tmp;
7320 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7321 list_del(&pfile->list);
7322 io_ring_file_put(ctx, pfile->file);
7326 spin_lock(&file_data->lock);
7327 list_del(&ref_node->node);
7328 spin_unlock(&file_data->lock);
7330 percpu_ref_exit(&ref_node->refs);
7332 percpu_ref_put(&file_data->refs);
7335 static void io_file_put_work(struct work_struct *work)
7337 struct io_ring_ctx *ctx;
7338 struct llist_node *node;
7340 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7341 node = llist_del_all(&ctx->file_put_llist);
7344 struct fixed_file_ref_node *ref_node;
7345 struct llist_node *next = node->next;
7347 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7348 __io_file_put_work(ref_node);
7353 static void io_file_data_ref_zero(struct percpu_ref *ref)
7355 struct fixed_file_ref_node *ref_node;
7356 struct io_ring_ctx *ctx;
7360 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7361 ctx = ref_node->file_data->ctx;
7363 if (percpu_ref_is_dying(&ctx->file_data->refs))
7366 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7368 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7370 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7373 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7374 struct io_ring_ctx *ctx)
7376 struct fixed_file_ref_node *ref_node;
7378 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7380 return ERR_PTR(-ENOMEM);
7382 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7385 return ERR_PTR(-ENOMEM);
7387 INIT_LIST_HEAD(&ref_node->node);
7388 INIT_LIST_HEAD(&ref_node->file_list);
7389 ref_node->file_data = ctx->file_data;
7393 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7395 percpu_ref_exit(&ref_node->refs);
7399 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7402 __s32 __user *fds = (__s32 __user *) arg;
7403 unsigned nr_tables, i;
7405 int fd, ret = -ENOMEM;
7406 struct fixed_file_ref_node *ref_node;
7407 struct fixed_file_data *file_data;
7413 if (nr_args > IORING_MAX_FIXED_FILES)
7416 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7419 file_data->ctx = ctx;
7420 init_completion(&file_data->done);
7421 INIT_LIST_HEAD(&file_data->ref_list);
7422 spin_lock_init(&file_data->lock);
7424 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7425 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7427 if (!file_data->table)
7430 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7431 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7434 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7436 ctx->file_data = file_data;
7438 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7439 struct fixed_file_table *table;
7442 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7446 /* allow sparse sets */
7456 * Don't allow io_uring instances to be registered. If UNIX
7457 * isn't enabled, then this causes a reference cycle and this
7458 * instance can never get freed. If UNIX is enabled we'll
7459 * handle it just fine, but there's still no point in allowing
7460 * a ring fd as it doesn't support regular read/write anyway.
7462 if (file->f_op == &io_uring_fops) {
7466 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7467 index = i & IORING_FILE_TABLE_MASK;
7468 table->files[index] = file;
7471 ret = io_sqe_files_scm(ctx);
7473 io_sqe_files_unregister(ctx);
7477 ref_node = alloc_fixed_file_ref_node(ctx);
7478 if (IS_ERR(ref_node)) {
7479 io_sqe_files_unregister(ctx);
7480 return PTR_ERR(ref_node);
7483 file_data->node = ref_node;
7484 spin_lock(&file_data->lock);
7485 list_add(&ref_node->node, &file_data->ref_list);
7486 spin_unlock(&file_data->lock);
7487 percpu_ref_get(&file_data->refs);
7490 for (i = 0; i < ctx->nr_user_files; i++) {
7491 file = io_file_from_index(ctx, i);
7495 for (i = 0; i < nr_tables; i++)
7496 kfree(file_data->table[i].files);
7497 ctx->nr_user_files = 0;
7499 percpu_ref_exit(&file_data->refs);
7501 kfree(file_data->table);
7503 ctx->file_data = NULL;
7507 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7510 #if defined(CONFIG_UNIX)
7511 struct sock *sock = ctx->ring_sock->sk;
7512 struct sk_buff_head *head = &sock->sk_receive_queue;
7513 struct sk_buff *skb;
7516 * See if we can merge this file into an existing skb SCM_RIGHTS
7517 * file set. If there's no room, fall back to allocating a new skb
7518 * and filling it in.
7520 spin_lock_irq(&head->lock);
7521 skb = skb_peek(head);
7523 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7525 if (fpl->count < SCM_MAX_FD) {
7526 __skb_unlink(skb, head);
7527 spin_unlock_irq(&head->lock);
7528 fpl->fp[fpl->count] = get_file(file);
7529 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7531 spin_lock_irq(&head->lock);
7532 __skb_queue_head(head, skb);
7537 spin_unlock_irq(&head->lock);
7544 return __io_sqe_files_scm(ctx, 1, index);
7550 static int io_queue_file_removal(struct fixed_file_data *data,
7553 struct io_file_put *pfile;
7554 struct fixed_file_ref_node *ref_node = data->node;
7556 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7561 list_add(&pfile->list, &ref_node->file_list);
7566 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7567 struct io_uring_files_update *up,
7570 struct fixed_file_data *data = ctx->file_data;
7571 struct fixed_file_ref_node *ref_node;
7576 bool needs_switch = false;
7578 if (check_add_overflow(up->offset, nr_args, &done))
7580 if (done > ctx->nr_user_files)
7583 ref_node = alloc_fixed_file_ref_node(ctx);
7584 if (IS_ERR(ref_node))
7585 return PTR_ERR(ref_node);
7588 fds = u64_to_user_ptr(up->fds);
7590 struct fixed_file_table *table;
7594 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7598 i = array_index_nospec(up->offset, ctx->nr_user_files);
7599 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7600 index = i & IORING_FILE_TABLE_MASK;
7601 if (table->files[index]) {
7602 file = table->files[index];
7603 err = io_queue_file_removal(data, file);
7606 table->files[index] = NULL;
7607 needs_switch = true;
7616 * Don't allow io_uring instances to be registered. If
7617 * UNIX isn't enabled, then this causes a reference
7618 * cycle and this instance can never get freed. If UNIX
7619 * is enabled we'll handle it just fine, but there's
7620 * still no point in allowing a ring fd as it doesn't
7621 * support regular read/write anyway.
7623 if (file->f_op == &io_uring_fops) {
7628 table->files[index] = file;
7629 err = io_sqe_file_register(ctx, file, i);
7631 table->files[index] = NULL;
7642 percpu_ref_kill(&data->node->refs);
7643 spin_lock(&data->lock);
7644 list_add(&ref_node->node, &data->ref_list);
7645 data->node = ref_node;
7646 spin_unlock(&data->lock);
7647 percpu_ref_get(&ctx->file_data->refs);
7649 destroy_fixed_file_ref_node(ref_node);
7651 return done ? done : err;
7654 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7657 struct io_uring_files_update up;
7659 if (!ctx->file_data)
7663 if (copy_from_user(&up, arg, sizeof(up)))
7668 return __io_sqe_files_update(ctx, &up, nr_args);
7671 static void io_free_work(struct io_wq_work *work)
7673 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7675 /* Consider that io_steal_work() relies on this ref */
7679 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7680 struct io_uring_params *p)
7682 struct io_wq_data data;
7684 struct io_ring_ctx *ctx_attach;
7685 unsigned int concurrency;
7688 data.user = ctx->user;
7689 data.free_work = io_free_work;
7690 data.do_work = io_wq_submit_work;
7692 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7693 /* Do QD, or 4 * CPUS, whatever is smallest */
7694 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7696 ctx->io_wq = io_wq_create(concurrency, &data);
7697 if (IS_ERR(ctx->io_wq)) {
7698 ret = PTR_ERR(ctx->io_wq);
7704 f = fdget(p->wq_fd);
7708 if (f.file->f_op != &io_uring_fops) {
7713 ctx_attach = f.file->private_data;
7714 /* @io_wq is protected by holding the fd */
7715 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7720 ctx->io_wq = ctx_attach->io_wq;
7726 static int io_uring_alloc_task_context(struct task_struct *task)
7728 struct io_uring_task *tctx;
7731 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7732 if (unlikely(!tctx))
7735 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7736 if (unlikely(ret)) {
7742 init_waitqueue_head(&tctx->wait);
7744 atomic_set(&tctx->in_idle, 0);
7745 tctx->sqpoll = false;
7746 io_init_identity(&tctx->__identity);
7747 tctx->identity = &tctx->__identity;
7748 task->io_uring = tctx;
7752 void __io_uring_free(struct task_struct *tsk)
7754 struct io_uring_task *tctx = tsk->io_uring;
7756 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7757 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
7758 if (tctx->identity != &tctx->__identity)
7759 kfree(tctx->identity);
7760 percpu_counter_destroy(&tctx->inflight);
7762 tsk->io_uring = NULL;
7765 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7766 struct io_uring_params *p)
7770 if (ctx->flags & IORING_SETUP_SQPOLL) {
7771 struct io_sq_data *sqd;
7774 if (!capable(CAP_SYS_ADMIN))
7777 sqd = io_get_sq_data(p);
7784 io_sq_thread_park(sqd);
7785 mutex_lock(&sqd->ctx_lock);
7786 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7787 mutex_unlock(&sqd->ctx_lock);
7788 io_sq_thread_unpark(sqd);
7790 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7791 if (!ctx->sq_thread_idle)
7792 ctx->sq_thread_idle = HZ;
7797 if (p->flags & IORING_SETUP_SQ_AFF) {
7798 int cpu = p->sq_thread_cpu;
7801 if (cpu >= nr_cpu_ids)
7803 if (!cpu_online(cpu))
7806 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
7807 cpu, "io_uring-sq");
7809 sqd->thread = kthread_create(io_sq_thread, sqd,
7812 if (IS_ERR(sqd->thread)) {
7813 ret = PTR_ERR(sqd->thread);
7817 ret = io_uring_alloc_task_context(sqd->thread);
7820 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7821 /* Can't have SQ_AFF without SQPOLL */
7827 ret = io_init_wq_offload(ctx, p);
7833 io_finish_async(ctx);
7837 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7839 struct io_sq_data *sqd = ctx->sq_data;
7841 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
7842 wake_up_process(sqd->thread);
7845 static inline void __io_unaccount_mem(struct user_struct *user,
7846 unsigned long nr_pages)
7848 atomic_long_sub(nr_pages, &user->locked_vm);
7851 static inline int __io_account_mem(struct user_struct *user,
7852 unsigned long nr_pages)
7854 unsigned long page_limit, cur_pages, new_pages;
7856 /* Don't allow more pages than we can safely lock */
7857 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7860 cur_pages = atomic_long_read(&user->locked_vm);
7861 new_pages = cur_pages + nr_pages;
7862 if (new_pages > page_limit)
7864 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7865 new_pages) != cur_pages);
7870 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7871 enum io_mem_account acct)
7874 __io_unaccount_mem(ctx->user, nr_pages);
7876 if (ctx->mm_account) {
7877 if (acct == ACCT_LOCKED)
7878 ctx->mm_account->locked_vm -= nr_pages;
7879 else if (acct == ACCT_PINNED)
7880 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7884 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7885 enum io_mem_account acct)
7889 if (ctx->limit_mem) {
7890 ret = __io_account_mem(ctx->user, nr_pages);
7895 if (ctx->mm_account) {
7896 if (acct == ACCT_LOCKED)
7897 ctx->mm_account->locked_vm += nr_pages;
7898 else if (acct == ACCT_PINNED)
7899 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7905 static void io_mem_free(void *ptr)
7912 page = virt_to_head_page(ptr);
7913 if (put_page_testzero(page))
7914 free_compound_page(page);
7917 static void *io_mem_alloc(size_t size)
7919 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7922 return (void *) __get_free_pages(gfp_flags, get_order(size));
7925 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7928 struct io_rings *rings;
7929 size_t off, sq_array_size;
7931 off = struct_size(rings, cqes, cq_entries);
7932 if (off == SIZE_MAX)
7936 off = ALIGN(off, SMP_CACHE_BYTES);
7944 sq_array_size = array_size(sizeof(u32), sq_entries);
7945 if (sq_array_size == SIZE_MAX)
7948 if (check_add_overflow(off, sq_array_size, &off))
7954 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7958 pages = (size_t)1 << get_order(
7959 rings_size(sq_entries, cq_entries, NULL));
7960 pages += (size_t)1 << get_order(
7961 array_size(sizeof(struct io_uring_sqe), sq_entries));
7966 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7970 if (!ctx->user_bufs)
7973 for (i = 0; i < ctx->nr_user_bufs; i++) {
7974 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7976 for (j = 0; j < imu->nr_bvecs; j++)
7977 unpin_user_page(imu->bvec[j].bv_page);
7979 if (imu->acct_pages)
7980 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
7985 kfree(ctx->user_bufs);
7986 ctx->user_bufs = NULL;
7987 ctx->nr_user_bufs = 0;
7991 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7992 void __user *arg, unsigned index)
7994 struct iovec __user *src;
7996 #ifdef CONFIG_COMPAT
7998 struct compat_iovec __user *ciovs;
7999 struct compat_iovec ciov;
8001 ciovs = (struct compat_iovec __user *) arg;
8002 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8005 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8006 dst->iov_len = ciov.iov_len;
8010 src = (struct iovec __user *) arg;
8011 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8017 * Not super efficient, but this is just a registration time. And we do cache
8018 * the last compound head, so generally we'll only do a full search if we don't
8021 * We check if the given compound head page has already been accounted, to
8022 * avoid double accounting it. This allows us to account the full size of the
8023 * page, not just the constituent pages of a huge page.
8025 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8026 int nr_pages, struct page *hpage)
8030 /* check current page array */
8031 for (i = 0; i < nr_pages; i++) {
8032 if (!PageCompound(pages[i]))
8034 if (compound_head(pages[i]) == hpage)
8038 /* check previously registered pages */
8039 for (i = 0; i < ctx->nr_user_bufs; i++) {
8040 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8042 for (j = 0; j < imu->nr_bvecs; j++) {
8043 if (!PageCompound(imu->bvec[j].bv_page))
8045 if (compound_head(imu->bvec[j].bv_page) == hpage)
8053 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8054 int nr_pages, struct io_mapped_ubuf *imu,
8055 struct page **last_hpage)
8059 for (i = 0; i < nr_pages; i++) {
8060 if (!PageCompound(pages[i])) {
8065 hpage = compound_head(pages[i]);
8066 if (hpage == *last_hpage)
8068 *last_hpage = hpage;
8069 if (headpage_already_acct(ctx, pages, i, hpage))
8071 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8075 if (!imu->acct_pages)
8078 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8080 imu->acct_pages = 0;
8084 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8087 struct vm_area_struct **vmas = NULL;
8088 struct page **pages = NULL;
8089 struct page *last_hpage = NULL;
8090 int i, j, got_pages = 0;
8095 if (!nr_args || nr_args > UIO_MAXIOV)
8098 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8100 if (!ctx->user_bufs)
8103 for (i = 0; i < nr_args; i++) {
8104 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8105 unsigned long off, start, end, ubuf;
8110 ret = io_copy_iov(ctx, &iov, arg, i);
8115 * Don't impose further limits on the size and buffer
8116 * constraints here, we'll -EINVAL later when IO is
8117 * submitted if they are wrong.
8120 if (!iov.iov_base || !iov.iov_len)
8123 /* arbitrary limit, but we need something */
8124 if (iov.iov_len > SZ_1G)
8127 ubuf = (unsigned long) iov.iov_base;
8128 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8129 start = ubuf >> PAGE_SHIFT;
8130 nr_pages = end - start;
8133 if (!pages || nr_pages > got_pages) {
8136 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8138 vmas = kvmalloc_array(nr_pages,
8139 sizeof(struct vm_area_struct *),
8141 if (!pages || !vmas) {
8145 got_pages = nr_pages;
8148 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8155 mmap_read_lock(current->mm);
8156 pret = pin_user_pages(ubuf, nr_pages,
8157 FOLL_WRITE | FOLL_LONGTERM,
8159 if (pret == nr_pages) {
8160 /* don't support file backed memory */
8161 for (j = 0; j < nr_pages; j++) {
8162 struct vm_area_struct *vma = vmas[j];
8165 !is_file_hugepages(vma->vm_file)) {
8171 ret = pret < 0 ? pret : -EFAULT;
8173 mmap_read_unlock(current->mm);
8176 * if we did partial map, or found file backed vmas,
8177 * release any pages we did get
8180 unpin_user_pages(pages, pret);
8185 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8187 unpin_user_pages(pages, pret);
8192 off = ubuf & ~PAGE_MASK;
8194 for (j = 0; j < nr_pages; j++) {
8197 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8198 imu->bvec[j].bv_page = pages[j];
8199 imu->bvec[j].bv_len = vec_len;
8200 imu->bvec[j].bv_offset = off;
8204 /* store original address for later verification */
8206 imu->len = iov.iov_len;
8207 imu->nr_bvecs = nr_pages;
8209 ctx->nr_user_bufs++;
8217 io_sqe_buffer_unregister(ctx);
8221 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8223 __s32 __user *fds = arg;
8229 if (copy_from_user(&fd, fds, sizeof(*fds)))
8232 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8233 if (IS_ERR(ctx->cq_ev_fd)) {
8234 int ret = PTR_ERR(ctx->cq_ev_fd);
8235 ctx->cq_ev_fd = NULL;
8242 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8244 if (ctx->cq_ev_fd) {
8245 eventfd_ctx_put(ctx->cq_ev_fd);
8246 ctx->cq_ev_fd = NULL;
8253 static int __io_destroy_buffers(int id, void *p, void *data)
8255 struct io_ring_ctx *ctx = data;
8256 struct io_buffer *buf = p;
8258 __io_remove_buffers(ctx, buf, id, -1U);
8262 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8264 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8265 idr_destroy(&ctx->io_buffer_idr);
8268 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8270 io_finish_async(ctx);
8271 io_sqe_buffer_unregister(ctx);
8273 if (ctx->sqo_task) {
8274 put_task_struct(ctx->sqo_task);
8275 ctx->sqo_task = NULL;
8276 mmdrop(ctx->mm_account);
8277 ctx->mm_account = NULL;
8280 #ifdef CONFIG_BLK_CGROUP
8281 if (ctx->sqo_blkcg_css)
8282 css_put(ctx->sqo_blkcg_css);
8285 io_sqe_files_unregister(ctx);
8286 io_eventfd_unregister(ctx);
8287 io_destroy_buffers(ctx);
8288 idr_destroy(&ctx->personality_idr);
8290 #if defined(CONFIG_UNIX)
8291 if (ctx->ring_sock) {
8292 ctx->ring_sock->file = NULL; /* so that iput() is called */
8293 sock_release(ctx->ring_sock);
8297 io_mem_free(ctx->rings);
8298 io_mem_free(ctx->sq_sqes);
8300 percpu_ref_exit(&ctx->refs);
8301 free_uid(ctx->user);
8302 put_cred(ctx->creds);
8303 kfree(ctx->cancel_hash);
8304 kmem_cache_free(req_cachep, ctx->fallback_req);
8308 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8310 struct io_ring_ctx *ctx = file->private_data;
8313 poll_wait(file, &ctx->cq_wait, wait);
8315 * synchronizes with barrier from wq_has_sleeper call in
8319 if (!io_sqring_full(ctx))
8320 mask |= EPOLLOUT | EPOLLWRNORM;
8321 if (io_cqring_events(ctx, false))
8322 mask |= EPOLLIN | EPOLLRDNORM;
8327 static int io_uring_fasync(int fd, struct file *file, int on)
8329 struct io_ring_ctx *ctx = file->private_data;
8331 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8334 static int io_remove_personalities(int id, void *p, void *data)
8336 struct io_ring_ctx *ctx = data;
8337 struct io_identity *iod;
8339 iod = idr_remove(&ctx->personality_idr, id);
8341 put_cred(iod->creds);
8342 if (refcount_dec_and_test(&iod->count))
8348 static void io_ring_exit_work(struct work_struct *work)
8350 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8354 * If we're doing polled IO and end up having requests being
8355 * submitted async (out-of-line), then completions can come in while
8356 * we're waiting for refs to drop. We need to reap these manually,
8357 * as nobody else will be looking for them.
8361 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8362 io_iopoll_try_reap_events(ctx);
8363 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8364 io_ring_ctx_free(ctx);
8367 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8369 mutex_lock(&ctx->uring_lock);
8370 percpu_ref_kill(&ctx->refs);
8371 mutex_unlock(&ctx->uring_lock);
8373 io_kill_timeouts(ctx, NULL);
8374 io_poll_remove_all(ctx, NULL);
8377 io_wq_cancel_all(ctx->io_wq);
8379 /* if we failed setting up the ctx, we might not have any rings */
8381 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8382 io_iopoll_try_reap_events(ctx);
8383 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8386 * Do this upfront, so we won't have a grace period where the ring
8387 * is closed but resources aren't reaped yet. This can cause
8388 * spurious failure in setting up a new ring.
8390 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8393 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8395 * Use system_unbound_wq to avoid spawning tons of event kworkers
8396 * if we're exiting a ton of rings at the same time. It just adds
8397 * noise and overhead, there's no discernable change in runtime
8398 * over using system_wq.
8400 queue_work(system_unbound_wq, &ctx->exit_work);
8403 static int io_uring_release(struct inode *inode, struct file *file)
8405 struct io_ring_ctx *ctx = file->private_data;
8407 file->private_data = NULL;
8408 io_ring_ctx_wait_and_kill(ctx);
8412 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8414 struct files_struct *files = data;
8416 return !files || ((work->flags & IO_WQ_WORK_FILES) &&
8417 work->identity->files == files);
8421 * Returns true if 'preq' is the link parent of 'req'
8423 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8425 struct io_kiocb *link;
8427 if (!(preq->flags & REQ_F_LINK_HEAD))
8430 list_for_each_entry(link, &preq->link_list, link_list) {
8439 * We're looking to cancel 'req' because it's holding on to our files, but
8440 * 'req' could be a link to another request. See if it is, and cancel that
8441 * parent request if so.
8443 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8445 struct hlist_node *tmp;
8446 struct io_kiocb *preq;
8450 spin_lock_irq(&ctx->completion_lock);
8451 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8452 struct hlist_head *list;
8454 list = &ctx->cancel_hash[i];
8455 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8456 found = io_match_link(preq, req);
8458 io_poll_remove_one(preq);
8463 spin_unlock_irq(&ctx->completion_lock);
8467 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8468 struct io_kiocb *req)
8470 struct io_kiocb *preq;
8473 spin_lock_irq(&ctx->completion_lock);
8474 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8475 found = io_match_link(preq, req);
8477 __io_timeout_cancel(preq);
8481 spin_unlock_irq(&ctx->completion_lock);
8485 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8487 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8490 if (req->flags & REQ_F_LINK_TIMEOUT) {
8491 unsigned long flags;
8492 struct io_ring_ctx *ctx = req->ctx;
8494 /* protect against races with linked timeouts */
8495 spin_lock_irqsave(&ctx->completion_lock, flags);
8496 ret = io_match_link(req, data);
8497 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8499 ret = io_match_link(req, data);
8504 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8506 enum io_wq_cancel cret;
8508 /* cancel this particular work, if it's running */
8509 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8510 if (cret != IO_WQ_CANCEL_NOTFOUND)
8513 /* find links that hold this pending, cancel those */
8514 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8515 if (cret != IO_WQ_CANCEL_NOTFOUND)
8518 /* if we have a poll link holding this pending, cancel that */
8519 if (io_poll_remove_link(ctx, req))
8522 /* final option, timeout link is holding this req pending */
8523 io_timeout_remove_link(ctx, req);
8526 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8527 struct task_struct *task,
8528 struct files_struct *files)
8530 struct io_defer_entry *de = NULL;
8533 spin_lock_irq(&ctx->completion_lock);
8534 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8535 if (io_task_match(de->req, task) &&
8536 io_match_files(de->req, files)) {
8537 list_cut_position(&list, &ctx->defer_list, &de->list);
8541 spin_unlock_irq(&ctx->completion_lock);
8543 while (!list_empty(&list)) {
8544 de = list_first_entry(&list, struct io_defer_entry, list);
8545 list_del_init(&de->list);
8546 req_set_fail_links(de->req);
8547 io_put_req(de->req);
8548 io_req_complete(de->req, -ECANCELED);
8554 * Returns true if we found and killed one or more files pinning requests
8556 static bool io_uring_cancel_files(struct io_ring_ctx *ctx,
8557 struct files_struct *files)
8559 if (list_empty_careful(&ctx->inflight_list))
8562 /* cancel all at once, should be faster than doing it one by one*/
8563 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8565 while (!list_empty_careful(&ctx->inflight_list)) {
8566 struct io_kiocb *cancel_req = NULL, *req;
8569 spin_lock_irq(&ctx->inflight_lock);
8570 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8571 if (files && (req->work.flags & IO_WQ_WORK_FILES) &&
8572 req->work.identity->files != files)
8574 /* req is being completed, ignore */
8575 if (!refcount_inc_not_zero(&req->refs))
8581 prepare_to_wait(&ctx->inflight_wait, &wait,
8582 TASK_UNINTERRUPTIBLE);
8583 spin_unlock_irq(&ctx->inflight_lock);
8585 /* We need to keep going until we don't find a matching req */
8588 /* cancel this request, or head link requests */
8589 io_attempt_cancel(ctx, cancel_req);
8590 io_put_req(cancel_req);
8591 /* cancellations _may_ trigger task work */
8594 finish_wait(&ctx->inflight_wait, &wait);
8600 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8602 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8603 struct task_struct *task = data;
8605 return io_task_match(req, task);
8608 static bool __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8609 struct task_struct *task,
8610 struct files_struct *files)
8614 ret = io_uring_cancel_files(ctx, files);
8616 enum io_wq_cancel cret;
8618 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, task, true);
8619 if (cret != IO_WQ_CANCEL_NOTFOUND)
8622 /* SQPOLL thread does its own polling */
8623 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8624 while (!list_empty_careful(&ctx->iopoll_list)) {
8625 io_iopoll_try_reap_events(ctx);
8630 ret |= io_poll_remove_all(ctx, task);
8631 ret |= io_kill_timeouts(ctx, task);
8638 * We need to iteratively cancel requests, in case a request has dependent
8639 * hard links. These persist even for failure of cancelations, hence keep
8640 * looping until none are found.
8642 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8643 struct files_struct *files)
8645 struct task_struct *task = current;
8647 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8648 task = ctx->sq_data->thread;
8649 atomic_inc(&task->io_uring->in_idle);
8650 io_sq_thread_park(ctx->sq_data);
8654 io_cancel_defer_files(ctx, NULL, files);
8656 io_cancel_defer_files(ctx, task, NULL);
8658 io_cqring_overflow_flush(ctx, true, task, files);
8660 while (__io_uring_cancel_task_requests(ctx, task, files)) {
8665 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8666 atomic_dec(&task->io_uring->in_idle);
8668 * If the files that are going away are the ones in the thread
8669 * identity, clear them out.
8671 if (task->io_uring->identity->files == files)
8672 task->io_uring->identity->files = NULL;
8673 io_sq_thread_unpark(ctx->sq_data);
8678 * Note that this task has used io_uring. We use it for cancelation purposes.
8680 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
8682 struct io_uring_task *tctx = current->io_uring;
8684 if (unlikely(!tctx)) {
8687 ret = io_uring_alloc_task_context(current);
8690 tctx = current->io_uring;
8692 if (tctx->last != file) {
8693 void *old = xa_load(&tctx->xa, (unsigned long)file);
8697 xa_store(&tctx->xa, (unsigned long)file, file, GFP_KERNEL);
8703 * This is race safe in that the task itself is doing this, hence it
8704 * cannot be going through the exit/cancel paths at the same time.
8705 * This cannot be modified while exit/cancel is running.
8707 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8708 tctx->sqpoll = true;
8714 * Remove this io_uring_file -> task mapping.
8716 static void io_uring_del_task_file(struct file *file)
8718 struct io_uring_task *tctx = current->io_uring;
8720 if (tctx->last == file)
8722 file = xa_erase(&tctx->xa, (unsigned long)file);
8728 * Drop task note for this file if we're the only ones that hold it after
8731 static void io_uring_attempt_task_drop(struct file *file)
8733 if (!current->io_uring)
8736 * fput() is pending, will be 2 if the only other ref is our potential
8737 * task file note. If the task is exiting, drop regardless of count.
8739 if (fatal_signal_pending(current) || (current->flags & PF_EXITING) ||
8740 atomic_long_read(&file->f_count) == 2)
8741 io_uring_del_task_file(file);
8744 void __io_uring_files_cancel(struct files_struct *files)
8746 struct io_uring_task *tctx = current->io_uring;
8748 unsigned long index;
8750 /* make sure overflow events are dropped */
8751 atomic_inc(&tctx->in_idle);
8753 xa_for_each(&tctx->xa, index, file) {
8754 struct io_ring_ctx *ctx = file->private_data;
8756 io_uring_cancel_task_requests(ctx, files);
8758 io_uring_del_task_file(file);
8761 atomic_dec(&tctx->in_idle);
8764 static s64 tctx_inflight(struct io_uring_task *tctx)
8766 unsigned long index;
8770 inflight = percpu_counter_sum(&tctx->inflight);
8775 * If we have SQPOLL rings, then we need to iterate and find them, and
8776 * add the pending count for those.
8778 xa_for_each(&tctx->xa, index, file) {
8779 struct io_ring_ctx *ctx = file->private_data;
8781 if (ctx->flags & IORING_SETUP_SQPOLL) {
8782 struct io_uring_task *__tctx = ctx->sqo_task->io_uring;
8784 inflight += percpu_counter_sum(&__tctx->inflight);
8792 * Find any io_uring fd that this task has registered or done IO on, and cancel
8795 void __io_uring_task_cancel(void)
8797 struct io_uring_task *tctx = current->io_uring;
8801 /* make sure overflow events are dropped */
8802 atomic_inc(&tctx->in_idle);
8805 /* read completions before cancelations */
8806 inflight = tctx_inflight(tctx);
8809 __io_uring_files_cancel(NULL);
8811 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8814 * If we've seen completions, retry. This avoids a race where
8815 * a completion comes in before we did prepare_to_wait().
8817 if (inflight != tctx_inflight(tctx))
8822 finish_wait(&tctx->wait, &wait);
8823 atomic_dec(&tctx->in_idle);
8826 static int io_uring_flush(struct file *file, void *data)
8828 io_uring_attempt_task_drop(file);
8832 static void *io_uring_validate_mmap_request(struct file *file,
8833 loff_t pgoff, size_t sz)
8835 struct io_ring_ctx *ctx = file->private_data;
8836 loff_t offset = pgoff << PAGE_SHIFT;
8841 case IORING_OFF_SQ_RING:
8842 case IORING_OFF_CQ_RING:
8845 case IORING_OFF_SQES:
8849 return ERR_PTR(-EINVAL);
8852 page = virt_to_head_page(ptr);
8853 if (sz > page_size(page))
8854 return ERR_PTR(-EINVAL);
8861 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8863 size_t sz = vma->vm_end - vma->vm_start;
8867 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8869 return PTR_ERR(ptr);
8871 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8872 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8875 #else /* !CONFIG_MMU */
8877 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8879 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8882 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8884 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8887 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8888 unsigned long addr, unsigned long len,
8889 unsigned long pgoff, unsigned long flags)
8893 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8895 return PTR_ERR(ptr);
8897 return (unsigned long) ptr;
8900 #endif /* !CONFIG_MMU */
8902 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8907 if (!io_sqring_full(ctx))
8910 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8912 if (!io_sqring_full(ctx))
8916 } while (!signal_pending(current));
8918 finish_wait(&ctx->sqo_sq_wait, &wait);
8921 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8922 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8925 struct io_ring_ctx *ctx;
8932 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
8933 IORING_ENTER_SQ_WAIT))
8941 if (f.file->f_op != &io_uring_fops)
8945 ctx = f.file->private_data;
8946 if (!percpu_ref_tryget(&ctx->refs))
8950 if (ctx->flags & IORING_SETUP_R_DISABLED)
8954 * For SQ polling, the thread will do all submissions and completions.
8955 * Just return the requested submit count, and wake the thread if
8959 if (ctx->flags & IORING_SETUP_SQPOLL) {
8960 if (!list_empty_careful(&ctx->cq_overflow_list))
8961 io_cqring_overflow_flush(ctx, false, NULL, NULL);
8962 if (flags & IORING_ENTER_SQ_WAKEUP)
8963 wake_up(&ctx->sq_data->wait);
8964 if (flags & IORING_ENTER_SQ_WAIT)
8965 io_sqpoll_wait_sq(ctx);
8966 submitted = to_submit;
8967 } else if (to_submit) {
8968 ret = io_uring_add_task_file(ctx, f.file);
8971 mutex_lock(&ctx->uring_lock);
8972 submitted = io_submit_sqes(ctx, to_submit);
8973 mutex_unlock(&ctx->uring_lock);
8975 if (submitted != to_submit)
8978 if (flags & IORING_ENTER_GETEVENTS) {
8979 min_complete = min(min_complete, ctx->cq_entries);
8982 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8983 * space applications don't need to do io completion events
8984 * polling again, they can rely on io_sq_thread to do polling
8985 * work, which can reduce cpu usage and uring_lock contention.
8987 if (ctx->flags & IORING_SETUP_IOPOLL &&
8988 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8989 ret = io_iopoll_check(ctx, min_complete);
8991 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8996 percpu_ref_put(&ctx->refs);
8999 return submitted ? submitted : ret;
9002 #ifdef CONFIG_PROC_FS
9003 static int io_uring_show_cred(int id, void *p, void *data)
9005 struct io_identity *iod = p;
9006 const struct cred *cred = iod->creds;
9007 struct seq_file *m = data;
9008 struct user_namespace *uns = seq_user_ns(m);
9009 struct group_info *gi;
9014 seq_printf(m, "%5d\n", id);
9015 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9016 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9017 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9018 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9019 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9020 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9021 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9022 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9023 seq_puts(m, "\n\tGroups:\t");
9024 gi = cred->group_info;
9025 for (g = 0; g < gi->ngroups; g++) {
9026 seq_put_decimal_ull(m, g ? " " : "",
9027 from_kgid_munged(uns, gi->gid[g]));
9029 seq_puts(m, "\n\tCapEff:\t");
9030 cap = cred->cap_effective;
9031 CAP_FOR_EACH_U32(__capi)
9032 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9037 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9039 struct io_sq_data *sq = NULL;
9044 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9045 * since fdinfo case grabs it in the opposite direction of normal use
9046 * cases. If we fail to get the lock, we just don't iterate any
9047 * structures that could be going away outside the io_uring mutex.
9049 has_lock = mutex_trylock(&ctx->uring_lock);
9051 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
9054 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9055 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9056 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9057 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9058 struct fixed_file_table *table;
9061 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
9062 f = table->files[i & IORING_FILE_TABLE_MASK];
9064 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9066 seq_printf(m, "%5u: <none>\n", i);
9068 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9069 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9070 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9072 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9073 (unsigned int) buf->len);
9075 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9076 seq_printf(m, "Personalities:\n");
9077 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9079 seq_printf(m, "PollList:\n");
9080 spin_lock_irq(&ctx->completion_lock);
9081 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9082 struct hlist_head *list = &ctx->cancel_hash[i];
9083 struct io_kiocb *req;
9085 hlist_for_each_entry(req, list, hash_node)
9086 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9087 req->task->task_works != NULL);
9089 spin_unlock_irq(&ctx->completion_lock);
9091 mutex_unlock(&ctx->uring_lock);
9094 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9096 struct io_ring_ctx *ctx = f->private_data;
9098 if (percpu_ref_tryget(&ctx->refs)) {
9099 __io_uring_show_fdinfo(ctx, m);
9100 percpu_ref_put(&ctx->refs);
9105 static const struct file_operations io_uring_fops = {
9106 .release = io_uring_release,
9107 .flush = io_uring_flush,
9108 .mmap = io_uring_mmap,
9110 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9111 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9113 .poll = io_uring_poll,
9114 .fasync = io_uring_fasync,
9115 #ifdef CONFIG_PROC_FS
9116 .show_fdinfo = io_uring_show_fdinfo,
9120 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9121 struct io_uring_params *p)
9123 struct io_rings *rings;
9124 size_t size, sq_array_offset;
9126 /* make sure these are sane, as we already accounted them */
9127 ctx->sq_entries = p->sq_entries;
9128 ctx->cq_entries = p->cq_entries;
9130 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9131 if (size == SIZE_MAX)
9134 rings = io_mem_alloc(size);
9139 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9140 rings->sq_ring_mask = p->sq_entries - 1;
9141 rings->cq_ring_mask = p->cq_entries - 1;
9142 rings->sq_ring_entries = p->sq_entries;
9143 rings->cq_ring_entries = p->cq_entries;
9144 ctx->sq_mask = rings->sq_ring_mask;
9145 ctx->cq_mask = rings->cq_ring_mask;
9147 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9148 if (size == SIZE_MAX) {
9149 io_mem_free(ctx->rings);
9154 ctx->sq_sqes = io_mem_alloc(size);
9155 if (!ctx->sq_sqes) {
9156 io_mem_free(ctx->rings);
9165 * Allocate an anonymous fd, this is what constitutes the application
9166 * visible backing of an io_uring instance. The application mmaps this
9167 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9168 * we have to tie this fd to a socket for file garbage collection purposes.
9170 static int io_uring_get_fd(struct io_ring_ctx *ctx)
9175 #if defined(CONFIG_UNIX)
9176 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9182 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9186 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9187 O_RDWR | O_CLOEXEC);
9191 ret = PTR_ERR(file);
9195 #if defined(CONFIG_UNIX)
9196 ctx->ring_sock->file = file;
9198 if (unlikely(io_uring_add_task_file(ctx, file))) {
9199 file = ERR_PTR(-ENOMEM);
9202 fd_install(ret, file);
9205 #if defined(CONFIG_UNIX)
9206 sock_release(ctx->ring_sock);
9207 ctx->ring_sock = NULL;
9212 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9213 struct io_uring_params __user *params)
9215 struct user_struct *user = NULL;
9216 struct io_ring_ctx *ctx;
9222 if (entries > IORING_MAX_ENTRIES) {
9223 if (!(p->flags & IORING_SETUP_CLAMP))
9225 entries = IORING_MAX_ENTRIES;
9229 * Use twice as many entries for the CQ ring. It's possible for the
9230 * application to drive a higher depth than the size of the SQ ring,
9231 * since the sqes are only used at submission time. This allows for
9232 * some flexibility in overcommitting a bit. If the application has
9233 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9234 * of CQ ring entries manually.
9236 p->sq_entries = roundup_pow_of_two(entries);
9237 if (p->flags & IORING_SETUP_CQSIZE) {
9239 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9240 * to a power-of-two, if it isn't already. We do NOT impose
9241 * any cq vs sq ring sizing.
9243 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9244 if (p->cq_entries < p->sq_entries)
9246 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9247 if (!(p->flags & IORING_SETUP_CLAMP))
9249 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9252 p->cq_entries = 2 * p->sq_entries;
9255 user = get_uid(current_user());
9256 limit_mem = !capable(CAP_IPC_LOCK);
9259 ret = __io_account_mem(user,
9260 ring_pages(p->sq_entries, p->cq_entries));
9267 ctx = io_ring_ctx_alloc(p);
9270 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9275 ctx->compat = in_compat_syscall();
9277 ctx->creds = get_current_cred();
9279 ctx->loginuid = current->loginuid;
9280 ctx->sessionid = current->sessionid;
9282 ctx->sqo_task = get_task_struct(current);
9285 * This is just grabbed for accounting purposes. When a process exits,
9286 * the mm is exited and dropped before the files, hence we need to hang
9287 * on to this mm purely for the purposes of being able to unaccount
9288 * memory (locked/pinned vm). It's not used for anything else.
9290 mmgrab(current->mm);
9291 ctx->mm_account = current->mm;
9293 #ifdef CONFIG_BLK_CGROUP
9295 * The sq thread will belong to the original cgroup it was inited in.
9296 * If the cgroup goes offline (e.g. disabling the io controller), then
9297 * issued bios will be associated with the closest cgroup later in the
9301 ctx->sqo_blkcg_css = blkcg_css();
9302 ret = css_tryget_online(ctx->sqo_blkcg_css);
9305 /* don't init against a dying cgroup, have the user try again */
9306 ctx->sqo_blkcg_css = NULL;
9313 * Account memory _before_ installing the file descriptor. Once
9314 * the descriptor is installed, it can get closed at any time. Also
9315 * do this before hitting the general error path, as ring freeing
9316 * will un-account as well.
9318 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9320 ctx->limit_mem = limit_mem;
9322 ret = io_allocate_scq_urings(ctx, p);
9326 ret = io_sq_offload_create(ctx, p);
9330 if (!(p->flags & IORING_SETUP_R_DISABLED))
9331 io_sq_offload_start(ctx);
9333 memset(&p->sq_off, 0, sizeof(p->sq_off));
9334 p->sq_off.head = offsetof(struct io_rings, sq.head);
9335 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9336 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9337 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9338 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9339 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9340 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9342 memset(&p->cq_off, 0, sizeof(p->cq_off));
9343 p->cq_off.head = offsetof(struct io_rings, cq.head);
9344 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9345 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9346 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9347 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9348 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9349 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9351 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9352 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9353 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9354 IORING_FEAT_POLL_32BITS;
9356 if (copy_to_user(params, p, sizeof(*p))) {
9362 * Install ring fd as the very last thing, so we don't risk someone
9363 * having closed it before we finish setup
9365 ret = io_uring_get_fd(ctx);
9369 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9372 io_ring_ctx_wait_and_kill(ctx);
9377 * Sets up an aio uring context, and returns the fd. Applications asks for a
9378 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9379 * params structure passed in.
9381 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9383 struct io_uring_params p;
9386 if (copy_from_user(&p, params, sizeof(p)))
9388 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9393 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9394 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9395 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9396 IORING_SETUP_R_DISABLED))
9399 return io_uring_create(entries, &p, params);
9402 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9403 struct io_uring_params __user *, params)
9405 return io_uring_setup(entries, params);
9408 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9410 struct io_uring_probe *p;
9414 size = struct_size(p, ops, nr_args);
9415 if (size == SIZE_MAX)
9417 p = kzalloc(size, GFP_KERNEL);
9422 if (copy_from_user(p, arg, size))
9425 if (memchr_inv(p, 0, size))
9428 p->last_op = IORING_OP_LAST - 1;
9429 if (nr_args > IORING_OP_LAST)
9430 nr_args = IORING_OP_LAST;
9432 for (i = 0; i < nr_args; i++) {
9434 if (!io_op_defs[i].not_supported)
9435 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9440 if (copy_to_user(arg, p, size))
9447 static int io_register_personality(struct io_ring_ctx *ctx)
9449 struct io_identity *id;
9452 id = kmalloc(sizeof(*id), GFP_KERNEL);
9456 io_init_identity(id);
9457 id->creds = get_current_cred();
9459 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9461 put_cred(id->creds);
9467 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9469 struct io_identity *iod;
9471 iod = idr_remove(&ctx->personality_idr, id);
9473 put_cred(iod->creds);
9474 if (refcount_dec_and_test(&iod->count))
9482 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9483 unsigned int nr_args)
9485 struct io_uring_restriction *res;
9489 /* Restrictions allowed only if rings started disabled */
9490 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9493 /* We allow only a single restrictions registration */
9494 if (ctx->restrictions.registered)
9497 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9500 size = array_size(nr_args, sizeof(*res));
9501 if (size == SIZE_MAX)
9504 res = memdup_user(arg, size);
9506 return PTR_ERR(res);
9510 for (i = 0; i < nr_args; i++) {
9511 switch (res[i].opcode) {
9512 case IORING_RESTRICTION_REGISTER_OP:
9513 if (res[i].register_op >= IORING_REGISTER_LAST) {
9518 __set_bit(res[i].register_op,
9519 ctx->restrictions.register_op);
9521 case IORING_RESTRICTION_SQE_OP:
9522 if (res[i].sqe_op >= IORING_OP_LAST) {
9527 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9529 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9530 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9532 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9533 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9542 /* Reset all restrictions if an error happened */
9544 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9546 ctx->restrictions.registered = true;
9552 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9554 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9557 if (ctx->restrictions.registered)
9558 ctx->restricted = 1;
9560 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9562 io_sq_offload_start(ctx);
9567 static bool io_register_op_must_quiesce(int op)
9570 case IORING_UNREGISTER_FILES:
9571 case IORING_REGISTER_FILES_UPDATE:
9572 case IORING_REGISTER_PROBE:
9573 case IORING_REGISTER_PERSONALITY:
9574 case IORING_UNREGISTER_PERSONALITY:
9581 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9582 void __user *arg, unsigned nr_args)
9583 __releases(ctx->uring_lock)
9584 __acquires(ctx->uring_lock)
9589 * We're inside the ring mutex, if the ref is already dying, then
9590 * someone else killed the ctx or is already going through
9591 * io_uring_register().
9593 if (percpu_ref_is_dying(&ctx->refs))
9596 if (io_register_op_must_quiesce(opcode)) {
9597 percpu_ref_kill(&ctx->refs);
9600 * Drop uring mutex before waiting for references to exit. If
9601 * another thread is currently inside io_uring_enter() it might
9602 * need to grab the uring_lock to make progress. If we hold it
9603 * here across the drain wait, then we can deadlock. It's safe
9604 * to drop the mutex here, since no new references will come in
9605 * after we've killed the percpu ref.
9607 mutex_unlock(&ctx->uring_lock);
9609 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9612 ret = io_run_task_work_sig();
9617 mutex_lock(&ctx->uring_lock);
9620 percpu_ref_resurrect(&ctx->refs);
9625 if (ctx->restricted) {
9626 if (opcode >= IORING_REGISTER_LAST) {
9631 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9638 case IORING_REGISTER_BUFFERS:
9639 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9641 case IORING_UNREGISTER_BUFFERS:
9645 ret = io_sqe_buffer_unregister(ctx);
9647 case IORING_REGISTER_FILES:
9648 ret = io_sqe_files_register(ctx, arg, nr_args);
9650 case IORING_UNREGISTER_FILES:
9654 ret = io_sqe_files_unregister(ctx);
9656 case IORING_REGISTER_FILES_UPDATE:
9657 ret = io_sqe_files_update(ctx, arg, nr_args);
9659 case IORING_REGISTER_EVENTFD:
9660 case IORING_REGISTER_EVENTFD_ASYNC:
9664 ret = io_eventfd_register(ctx, arg);
9667 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9668 ctx->eventfd_async = 1;
9670 ctx->eventfd_async = 0;
9672 case IORING_UNREGISTER_EVENTFD:
9676 ret = io_eventfd_unregister(ctx);
9678 case IORING_REGISTER_PROBE:
9680 if (!arg || nr_args > 256)
9682 ret = io_probe(ctx, arg, nr_args);
9684 case IORING_REGISTER_PERSONALITY:
9688 ret = io_register_personality(ctx);
9690 case IORING_UNREGISTER_PERSONALITY:
9694 ret = io_unregister_personality(ctx, nr_args);
9696 case IORING_REGISTER_ENABLE_RINGS:
9700 ret = io_register_enable_rings(ctx);
9702 case IORING_REGISTER_RESTRICTIONS:
9703 ret = io_register_restrictions(ctx, arg, nr_args);
9711 if (io_register_op_must_quiesce(opcode)) {
9712 /* bring the ctx back to life */
9713 percpu_ref_reinit(&ctx->refs);
9715 reinit_completion(&ctx->ref_comp);
9720 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9721 void __user *, arg, unsigned int, nr_args)
9723 struct io_ring_ctx *ctx;
9732 if (f.file->f_op != &io_uring_fops)
9735 ctx = f.file->private_data;
9737 mutex_lock(&ctx->uring_lock);
9738 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9739 mutex_unlock(&ctx->uring_lock);
9740 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9741 ctx->cq_ev_fd != NULL, ret);
9747 static int __init io_uring_init(void)
9749 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9750 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9751 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9754 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9755 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9756 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9757 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9758 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9759 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9760 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9761 BUILD_BUG_SQE_ELEM(8, __u64, off);
9762 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9763 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9764 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9765 BUILD_BUG_SQE_ELEM(24, __u32, len);
9766 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9767 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9768 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9769 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9770 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9771 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9772 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9773 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9774 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9775 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9776 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9777 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9778 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9779 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9780 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9781 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9782 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9783 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9784 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9786 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9787 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9788 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
9791 __initcall(io_uring_init);