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/blkdev.h>
61 #include <linux/bvec.h>
62 #include <linux/net.h>
64 #include <net/af_unix.h>
66 #include <linux/anon_inodes.h>
67 #include <linux/sched/mm.h>
68 #include <linux/uaccess.h>
69 #include <linux/nospec.h>
70 #include <linux/sizes.h>
71 #include <linux/hugetlb.h>
72 #include <linux/highmem.h>
73 #include <linux/namei.h>
74 #include <linux/fsnotify.h>
75 #include <linux/fadvise.h>
76 #include <linux/eventpoll.h>
77 #include <linux/splice.h>
78 #include <linux/task_work.h>
79 #include <linux/pagemap.h>
80 #include <linux/io_uring.h>
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/io_uring.h>
85 #include <uapi/linux/io_uring.h>
90 #define IORING_MAX_ENTRIES 32768
91 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
94 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
96 #define IORING_FILE_TABLE_SHIFT 9
97 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
98 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
99 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
100 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
101 IORING_REGISTER_LAST + IORING_OP_LAST)
103 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
104 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
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 enum io_uring_cmd_flags {
191 IO_URING_F_NONBLOCK = 1,
192 IO_URING_F_COMPLETE_DEFER = 2,
195 struct io_mapped_ubuf {
198 unsigned int nr_bvecs;
199 unsigned long acct_pages;
200 struct bio_vec bvec[];
205 struct io_overflow_cqe {
206 struct io_uring_cqe cqe;
207 struct list_head list;
210 struct io_fixed_file {
211 /* file * with additional FFS_* flags */
212 unsigned long file_ptr;
216 struct list_head list;
224 struct io_file_table {
225 /* two level table */
226 struct io_fixed_file **files;
229 struct io_rsrc_node {
230 struct percpu_ref refs;
231 struct list_head node;
232 struct list_head rsrc_list;
233 struct io_rsrc_data *rsrc_data;
234 struct llist_node llist;
238 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
240 struct io_rsrc_data {
241 struct io_ring_ctx *ctx;
246 struct completion done;
251 struct list_head list;
257 struct io_restriction {
258 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
259 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
260 u8 sqe_flags_allowed;
261 u8 sqe_flags_required;
266 IO_SQ_THREAD_SHOULD_STOP = 0,
267 IO_SQ_THREAD_SHOULD_PARK,
272 atomic_t park_pending;
275 /* ctx's that are using this sqd */
276 struct list_head ctx_list;
278 struct task_struct *thread;
279 struct wait_queue_head wait;
281 unsigned sq_thread_idle;
287 struct completion exited;
288 struct callback_head *park_task_work;
291 #define IO_IOPOLL_BATCH 8
292 #define IO_COMPL_BATCH 32
293 #define IO_REQ_CACHE_SIZE 32
294 #define IO_REQ_ALLOC_BATCH 8
296 struct io_comp_state {
297 struct io_kiocb *reqs[IO_COMPL_BATCH];
299 unsigned int locked_free_nr;
300 /* inline/task_work completion list, under ->uring_lock */
301 struct list_head free_list;
302 /* IRQ completion list, under ->completion_lock */
303 struct list_head locked_free_list;
306 struct io_submit_link {
307 struct io_kiocb *head;
308 struct io_kiocb *last;
311 struct io_submit_state {
312 struct blk_plug plug;
313 struct io_submit_link link;
316 * io_kiocb alloc cache
318 void *reqs[IO_REQ_CACHE_SIZE];
319 unsigned int free_reqs;
324 * Batch completion logic
326 struct io_comp_state comp;
329 * File reference cache
333 unsigned int file_refs;
334 unsigned int ios_left;
339 struct percpu_ref refs;
340 } ____cacheline_aligned_in_smp;
344 unsigned int compat: 1;
345 unsigned int drain_next: 1;
346 unsigned int eventfd_async: 1;
347 unsigned int restricted: 1;
350 * Ring buffer of indices into array of io_uring_sqe, which is
351 * mmapped by the application using the IORING_OFF_SQES offset.
353 * This indirection could e.g. be used to assign fixed
354 * io_uring_sqe entries to operations and only submit them to
355 * the queue when needed.
357 * The kernel modifies neither the indices array nor the entries
361 unsigned cached_sq_head;
364 unsigned sq_thread_idle;
365 unsigned cached_sq_dropped;
366 unsigned cached_cq_overflow;
367 unsigned long sq_check_overflow;
369 /* hashed buffered write serialization */
370 struct io_wq_hash *hash_map;
372 struct list_head defer_list;
373 struct list_head timeout_list;
374 struct list_head cq_overflow_list;
376 struct io_uring_sqe *sq_sqes;
377 } ____cacheline_aligned_in_smp;
380 struct mutex uring_lock;
381 wait_queue_head_t wait;
382 } ____cacheline_aligned_in_smp;
384 struct io_submit_state submit_state;
386 struct io_rings *rings;
388 /* Only used for accounting purposes */
389 struct mm_struct *mm_account;
391 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
392 struct io_sq_data *sq_data; /* if using sq thread polling */
394 struct wait_queue_head sqo_sq_wait;
395 struct list_head sqd_list;
398 * If used, fixed file set. Writers must ensure that ->refs is dead,
399 * readers must ensure that ->refs is alive as long as the file* is
400 * used. Only updated through io_uring_register(2).
402 struct io_rsrc_data *file_data;
403 struct io_file_table file_table;
404 unsigned nr_user_files;
406 /* if used, fixed mapped user buffers */
407 unsigned nr_user_bufs;
408 struct io_mapped_ubuf **user_bufs;
410 struct user_struct *user;
412 struct completion ref_comp;
414 #if defined(CONFIG_UNIX)
415 struct socket *ring_sock;
418 struct xarray io_buffers;
420 struct xarray personalities;
424 unsigned cached_cq_tail;
427 atomic_t cq_timeouts;
428 unsigned cq_last_tm_flush;
429 unsigned long cq_check_overflow;
430 struct wait_queue_head cq_wait;
431 struct fasync_struct *cq_fasync;
432 struct eventfd_ctx *cq_ev_fd;
433 } ____cacheline_aligned_in_smp;
436 spinlock_t completion_lock;
439 * ->iopoll_list is protected by the ctx->uring_lock for
440 * io_uring instances that don't use IORING_SETUP_SQPOLL.
441 * For SQPOLL, only the single threaded io_sq_thread() will
442 * manipulate the list, hence no extra locking is needed there.
444 struct list_head iopoll_list;
445 struct hlist_head *cancel_hash;
446 unsigned cancel_hash_bits;
447 bool poll_multi_file;
448 } ____cacheline_aligned_in_smp;
450 struct delayed_work rsrc_put_work;
451 struct llist_head rsrc_put_llist;
452 struct list_head rsrc_ref_list;
453 spinlock_t rsrc_ref_lock;
454 struct io_rsrc_node *rsrc_node;
455 struct io_rsrc_node *rsrc_backup_node;
457 struct io_restriction restrictions;
460 struct callback_head *exit_task_work;
462 /* Keep this last, we don't need it for the fast path */
463 struct work_struct exit_work;
464 struct list_head tctx_list;
467 struct io_uring_task {
468 /* submission side */
470 struct wait_queue_head wait;
471 const struct io_ring_ctx *last;
473 struct percpu_counter inflight;
474 atomic_t inflight_tracked;
477 spinlock_t task_lock;
478 struct io_wq_work_list task_list;
479 unsigned long task_state;
480 struct callback_head task_work;
484 * First field must be the file pointer in all the
485 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
487 struct io_poll_iocb {
489 struct wait_queue_head *head;
493 struct wait_queue_entry wait;
496 struct io_poll_update {
502 bool update_user_data;
510 struct io_timeout_data {
511 struct io_kiocb *req;
512 struct hrtimer timer;
513 struct timespec64 ts;
514 enum hrtimer_mode mode;
519 struct sockaddr __user *addr;
520 int __user *addr_len;
522 unsigned long nofile;
542 struct list_head list;
543 /* head of the link, used by linked timeouts only */
544 struct io_kiocb *head;
547 struct io_timeout_rem {
552 struct timespec64 ts;
557 /* NOTE: kiocb has the file as the first member, so don't do it here */
565 struct sockaddr __user *addr;
572 struct compat_msghdr __user *umsg_compat;
573 struct user_msghdr __user *umsg;
579 struct io_buffer *kbuf;
585 struct filename *filename;
587 unsigned long nofile;
590 struct io_rsrc_update {
616 struct epoll_event event;
620 struct file *file_out;
621 struct file *file_in;
628 struct io_provide_buf {
642 const char __user *filename;
643 struct statx __user *buffer;
655 struct filename *oldpath;
656 struct filename *newpath;
664 struct filename *filename;
667 struct io_completion {
669 struct list_head list;
673 struct io_async_connect {
674 struct sockaddr_storage address;
677 struct io_async_msghdr {
678 struct iovec fast_iov[UIO_FASTIOV];
679 /* points to an allocated iov, if NULL we use fast_iov instead */
680 struct iovec *free_iov;
681 struct sockaddr __user *uaddr;
683 struct sockaddr_storage addr;
687 struct iovec fast_iov[UIO_FASTIOV];
688 const struct iovec *free_iovec;
689 struct iov_iter iter;
691 struct wait_page_queue wpq;
695 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
696 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
697 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
698 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
699 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
700 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
706 REQ_F_LINK_TIMEOUT_BIT,
707 REQ_F_NEED_CLEANUP_BIT,
709 REQ_F_BUFFER_SELECTED_BIT,
710 REQ_F_LTIMEOUT_ACTIVE_BIT,
711 REQ_F_COMPLETE_INLINE_BIT,
713 REQ_F_DONT_REISSUE_BIT,
714 /* keep async read/write and isreg together and in order */
715 REQ_F_ASYNC_READ_BIT,
716 REQ_F_ASYNC_WRITE_BIT,
719 /* not a real bit, just to check we're not overflowing the space */
725 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
726 /* drain existing IO first */
727 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
729 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
730 /* doesn't sever on completion < 0 */
731 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
733 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
734 /* IOSQE_BUFFER_SELECT */
735 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
737 /* fail rest of links */
738 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
739 /* on inflight list, should be cancelled and waited on exit reliably */
740 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
741 /* read/write uses file position */
742 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
743 /* must not punt to workers */
744 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
745 /* has or had linked timeout */
746 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
748 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
749 /* already went through poll handler */
750 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
751 /* buffer already selected */
752 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
753 /* linked timeout is active, i.e. prepared by link's head */
754 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
755 /* completion is deferred through io_comp_state */
756 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
757 /* caller should reissue async */
758 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
759 /* don't attempt request reissue, see io_rw_reissue() */
760 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
761 /* supports async reads */
762 REQ_F_ASYNC_READ = BIT(REQ_F_ASYNC_READ_BIT),
763 /* supports async writes */
764 REQ_F_ASYNC_WRITE = BIT(REQ_F_ASYNC_WRITE_BIT),
766 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
770 struct io_poll_iocb poll;
771 struct io_poll_iocb *double_poll;
774 struct io_task_work {
775 struct io_wq_work_node node;
776 task_work_func_t func;
780 * NOTE! Each of the iocb union members has the file pointer
781 * as the first entry in their struct definition. So you can
782 * access the file pointer through any of the sub-structs,
783 * or directly as just 'ki_filp' in this struct.
789 struct io_poll_iocb poll;
790 struct io_poll_update poll_update;
791 struct io_accept accept;
793 struct io_cancel cancel;
794 struct io_timeout timeout;
795 struct io_timeout_rem timeout_rem;
796 struct io_connect connect;
797 struct io_sr_msg sr_msg;
799 struct io_close close;
800 struct io_rsrc_update rsrc_update;
801 struct io_fadvise fadvise;
802 struct io_madvise madvise;
803 struct io_epoll epoll;
804 struct io_splice splice;
805 struct io_provide_buf pbuf;
806 struct io_statx statx;
807 struct io_shutdown shutdown;
808 struct io_rename rename;
809 struct io_unlink unlink;
810 /* use only after cleaning per-op data, see io_clean_op() */
811 struct io_completion compl;
814 /* opcode allocated if it needs to store data for async defer */
817 /* polled IO has completed */
823 struct io_ring_ctx *ctx;
826 struct task_struct *task;
829 struct io_kiocb *link;
830 struct percpu_ref *fixed_rsrc_refs;
832 /* used with ctx->iopoll_list with reads/writes */
833 struct list_head inflight_entry;
835 struct io_task_work io_task_work;
836 struct callback_head task_work;
838 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
839 struct hlist_node hash_node;
840 struct async_poll *apoll;
841 struct io_wq_work work;
842 /* store used ubuf, so we can prevent reloading */
843 struct io_mapped_ubuf *imu;
846 struct io_tctx_node {
847 struct list_head ctx_node;
848 struct task_struct *task;
849 struct io_ring_ctx *ctx;
852 struct io_defer_entry {
853 struct list_head list;
854 struct io_kiocb *req;
859 /* needs req->file assigned */
860 unsigned needs_file : 1;
861 /* hash wq insertion if file is a regular file */
862 unsigned hash_reg_file : 1;
863 /* unbound wq insertion if file is a non-regular file */
864 unsigned unbound_nonreg_file : 1;
865 /* opcode is not supported by this kernel */
866 unsigned not_supported : 1;
867 /* set if opcode supports polled "wait" */
869 unsigned pollout : 1;
870 /* op supports buffer selection */
871 unsigned buffer_select : 1;
872 /* do prep async if is going to be punted */
873 unsigned needs_async_setup : 1;
874 /* should block plug */
876 /* size of async data needed, if any */
877 unsigned short async_size;
880 static const struct io_op_def io_op_defs[] = {
881 [IORING_OP_NOP] = {},
882 [IORING_OP_READV] = {
884 .unbound_nonreg_file = 1,
887 .needs_async_setup = 1,
889 .async_size = sizeof(struct io_async_rw),
891 [IORING_OP_WRITEV] = {
894 .unbound_nonreg_file = 1,
896 .needs_async_setup = 1,
898 .async_size = sizeof(struct io_async_rw),
900 [IORING_OP_FSYNC] = {
903 [IORING_OP_READ_FIXED] = {
905 .unbound_nonreg_file = 1,
908 .async_size = sizeof(struct io_async_rw),
910 [IORING_OP_WRITE_FIXED] = {
913 .unbound_nonreg_file = 1,
916 .async_size = sizeof(struct io_async_rw),
918 [IORING_OP_POLL_ADD] = {
920 .unbound_nonreg_file = 1,
922 [IORING_OP_POLL_REMOVE] = {},
923 [IORING_OP_SYNC_FILE_RANGE] = {
926 [IORING_OP_SENDMSG] = {
928 .unbound_nonreg_file = 1,
930 .needs_async_setup = 1,
931 .async_size = sizeof(struct io_async_msghdr),
933 [IORING_OP_RECVMSG] = {
935 .unbound_nonreg_file = 1,
938 .needs_async_setup = 1,
939 .async_size = sizeof(struct io_async_msghdr),
941 [IORING_OP_TIMEOUT] = {
942 .async_size = sizeof(struct io_timeout_data),
944 [IORING_OP_TIMEOUT_REMOVE] = {
945 /* used by timeout updates' prep() */
947 [IORING_OP_ACCEPT] = {
949 .unbound_nonreg_file = 1,
952 [IORING_OP_ASYNC_CANCEL] = {},
953 [IORING_OP_LINK_TIMEOUT] = {
954 .async_size = sizeof(struct io_timeout_data),
956 [IORING_OP_CONNECT] = {
958 .unbound_nonreg_file = 1,
960 .needs_async_setup = 1,
961 .async_size = sizeof(struct io_async_connect),
963 [IORING_OP_FALLOCATE] = {
966 [IORING_OP_OPENAT] = {},
967 [IORING_OP_CLOSE] = {},
968 [IORING_OP_FILES_UPDATE] = {},
969 [IORING_OP_STATX] = {},
972 .unbound_nonreg_file = 1,
976 .async_size = sizeof(struct io_async_rw),
978 [IORING_OP_WRITE] = {
980 .unbound_nonreg_file = 1,
983 .async_size = sizeof(struct io_async_rw),
985 [IORING_OP_FADVISE] = {
988 [IORING_OP_MADVISE] = {},
991 .unbound_nonreg_file = 1,
996 .unbound_nonreg_file = 1,
1000 [IORING_OP_OPENAT2] = {
1002 [IORING_OP_EPOLL_CTL] = {
1003 .unbound_nonreg_file = 1,
1005 [IORING_OP_SPLICE] = {
1008 .unbound_nonreg_file = 1,
1010 [IORING_OP_PROVIDE_BUFFERS] = {},
1011 [IORING_OP_REMOVE_BUFFERS] = {},
1015 .unbound_nonreg_file = 1,
1017 [IORING_OP_SHUTDOWN] = {
1020 [IORING_OP_RENAMEAT] = {},
1021 [IORING_OP_UNLINKAT] = {},
1024 static bool io_disarm_next(struct io_kiocb *req);
1025 static void io_uring_del_task_file(unsigned long index);
1026 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1027 struct task_struct *task,
1028 struct files_struct *files);
1029 static void io_uring_cancel_sqpoll(struct io_sq_data *sqd);
1030 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx);
1032 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1033 long res, unsigned int cflags);
1034 static void io_put_req(struct io_kiocb *req);
1035 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1036 static void io_dismantle_req(struct io_kiocb *req);
1037 static void io_put_task(struct task_struct *task, int nr);
1038 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1039 static void io_queue_linked_timeout(struct io_kiocb *req);
1040 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1041 struct io_uring_rsrc_update2 *up,
1043 static void io_clean_op(struct io_kiocb *req);
1044 static struct file *io_file_get(struct io_submit_state *state,
1045 struct io_kiocb *req, int fd, bool fixed);
1046 static void __io_queue_sqe(struct io_kiocb *req);
1047 static void io_rsrc_put_work(struct work_struct *work);
1049 static void io_req_task_queue(struct io_kiocb *req);
1050 static void io_submit_flush_completions(struct io_comp_state *cs,
1051 struct io_ring_ctx *ctx);
1052 static bool io_poll_remove_waitqs(struct io_kiocb *req);
1053 static int io_req_prep_async(struct io_kiocb *req);
1055 static struct kmem_cache *req_cachep;
1057 static const struct file_operations io_uring_fops;
1059 struct sock *io_uring_get_socket(struct file *file)
1061 #if defined(CONFIG_UNIX)
1062 if (file->f_op == &io_uring_fops) {
1063 struct io_ring_ctx *ctx = file->private_data;
1065 return ctx->ring_sock->sk;
1070 EXPORT_SYMBOL(io_uring_get_socket);
1072 #define io_for_each_link(pos, head) \
1073 for (pos = (head); pos; pos = pos->link)
1075 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1077 struct io_ring_ctx *ctx = req->ctx;
1079 if (!req->fixed_rsrc_refs) {
1080 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1081 percpu_ref_get(req->fixed_rsrc_refs);
1085 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1087 bool got = percpu_ref_tryget(ref);
1089 /* already at zero, wait for ->release() */
1091 wait_for_completion(compl);
1092 percpu_ref_resurrect(ref);
1094 percpu_ref_put(ref);
1097 static bool io_match_task(struct io_kiocb *head,
1098 struct task_struct *task,
1099 struct files_struct *files)
1101 struct io_kiocb *req;
1103 if (task && head->task != task)
1108 io_for_each_link(req, head) {
1109 if (req->flags & REQ_F_INFLIGHT)
1115 static inline void req_set_fail_links(struct io_kiocb *req)
1117 if (req->flags & REQ_F_LINK)
1118 req->flags |= REQ_F_FAIL_LINK;
1121 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1123 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1125 complete(&ctx->ref_comp);
1128 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1130 return !req->timeout.off;
1133 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1135 struct io_ring_ctx *ctx;
1138 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
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 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1168 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1169 mutex_init(&ctx->uring_lock);
1170 init_waitqueue_head(&ctx->wait);
1171 spin_lock_init(&ctx->completion_lock);
1172 INIT_LIST_HEAD(&ctx->iopoll_list);
1173 INIT_LIST_HEAD(&ctx->defer_list);
1174 INIT_LIST_HEAD(&ctx->timeout_list);
1175 spin_lock_init(&ctx->rsrc_ref_lock);
1176 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1177 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1178 init_llist_head(&ctx->rsrc_put_llist);
1179 INIT_LIST_HEAD(&ctx->tctx_list);
1180 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1181 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1184 kfree(ctx->cancel_hash);
1189 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1191 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1192 struct io_ring_ctx *ctx = req->ctx;
1194 return seq != ctx->cached_cq_tail
1195 + READ_ONCE(ctx->cached_cq_overflow);
1201 static void io_req_track_inflight(struct io_kiocb *req)
1203 if (!(req->flags & REQ_F_INFLIGHT)) {
1204 req->flags |= REQ_F_INFLIGHT;
1205 atomic_inc(¤t->io_uring->inflight_tracked);
1209 static void io_prep_async_work(struct io_kiocb *req)
1211 const struct io_op_def *def = &io_op_defs[req->opcode];
1212 struct io_ring_ctx *ctx = req->ctx;
1214 if (!req->work.creds)
1215 req->work.creds = get_current_cred();
1217 req->work.list.next = NULL;
1218 req->work.flags = 0;
1219 if (req->flags & REQ_F_FORCE_ASYNC)
1220 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1222 if (req->flags & REQ_F_ISREG) {
1223 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1224 io_wq_hash_work(&req->work, file_inode(req->file));
1225 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1226 if (def->unbound_nonreg_file)
1227 req->work.flags |= IO_WQ_WORK_UNBOUND;
1230 switch (req->opcode) {
1231 case IORING_OP_SPLICE:
1233 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1234 req->work.flags |= IO_WQ_WORK_UNBOUND;
1239 static void io_prep_async_link(struct io_kiocb *req)
1241 struct io_kiocb *cur;
1243 io_for_each_link(cur, req)
1244 io_prep_async_work(cur);
1247 static void io_queue_async_work(struct io_kiocb *req)
1249 struct io_ring_ctx *ctx = req->ctx;
1250 struct io_kiocb *link = io_prep_linked_timeout(req);
1251 struct io_uring_task *tctx = req->task->io_uring;
1254 BUG_ON(!tctx->io_wq);
1256 /* init ->work of the whole link before punting */
1257 io_prep_async_link(req);
1258 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1259 &req->work, req->flags);
1260 io_wq_enqueue(tctx->io_wq, &req->work);
1262 io_queue_linked_timeout(link);
1265 static void io_kill_timeout(struct io_kiocb *req, int status)
1266 __must_hold(&req->ctx->completion_lock)
1268 struct io_timeout_data *io = req->async_data;
1270 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1271 atomic_set(&req->ctx->cq_timeouts,
1272 atomic_read(&req->ctx->cq_timeouts) + 1);
1273 list_del_init(&req->timeout.list);
1274 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1275 io_put_req_deferred(req, 1);
1279 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1282 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1283 struct io_defer_entry, list);
1285 if (req_need_defer(de->req, de->seq))
1287 list_del_init(&de->list);
1288 io_req_task_queue(de->req);
1290 } while (!list_empty(&ctx->defer_list));
1293 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1297 if (list_empty(&ctx->timeout_list))
1300 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1303 u32 events_needed, events_got;
1304 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1305 struct io_kiocb, timeout.list);
1307 if (io_is_timeout_noseq(req))
1311 * Since seq can easily wrap around over time, subtract
1312 * the last seq at which timeouts were flushed before comparing.
1313 * Assuming not more than 2^31-1 events have happened since,
1314 * these subtractions won't have wrapped, so we can check if
1315 * target is in [last_seq, current_seq] by comparing the two.
1317 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1318 events_got = seq - ctx->cq_last_tm_flush;
1319 if (events_got < events_needed)
1322 list_del_init(&req->timeout.list);
1323 io_kill_timeout(req, 0);
1324 } while (!list_empty(&ctx->timeout_list));
1326 ctx->cq_last_tm_flush = seq;
1329 static void io_commit_cqring(struct io_ring_ctx *ctx)
1331 io_flush_timeouts(ctx);
1333 /* order cqe stores with ring update */
1334 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1336 if (unlikely(!list_empty(&ctx->defer_list)))
1337 __io_queue_deferred(ctx);
1340 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1342 struct io_rings *r = ctx->rings;
1344 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1347 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1349 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1352 static inline struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1354 struct io_rings *rings = ctx->rings;
1358 * writes to the cq entry need to come after reading head; the
1359 * control dependency is enough as we're using WRITE_ONCE to
1362 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1365 tail = ctx->cached_cq_tail++;
1366 return &rings->cqes[tail & ctx->cq_mask];
1369 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1371 if (likely(!ctx->cq_ev_fd))
1373 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1375 return !ctx->eventfd_async || io_wq_current_is_worker();
1378 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1380 /* see waitqueue_active() comment */
1383 if (waitqueue_active(&ctx->wait))
1384 wake_up(&ctx->wait);
1385 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1386 wake_up(&ctx->sq_data->wait);
1387 if (io_should_trigger_evfd(ctx))
1388 eventfd_signal(ctx->cq_ev_fd, 1);
1389 if (waitqueue_active(&ctx->cq_wait)) {
1390 wake_up_interruptible(&ctx->cq_wait);
1391 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1395 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1397 /* see waitqueue_active() comment */
1400 if (ctx->flags & IORING_SETUP_SQPOLL) {
1401 if (waitqueue_active(&ctx->wait))
1402 wake_up(&ctx->wait);
1404 if (io_should_trigger_evfd(ctx))
1405 eventfd_signal(ctx->cq_ev_fd, 1);
1406 if (waitqueue_active(&ctx->cq_wait)) {
1407 wake_up_interruptible(&ctx->cq_wait);
1408 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1412 /* Returns true if there are no backlogged entries after the flush */
1413 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1415 struct io_rings *rings = ctx->rings;
1416 unsigned long flags;
1417 bool all_flushed, posted;
1419 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1423 spin_lock_irqsave(&ctx->completion_lock, flags);
1424 while (!list_empty(&ctx->cq_overflow_list)) {
1425 struct io_uring_cqe *cqe = io_get_cqring(ctx);
1426 struct io_overflow_cqe *ocqe;
1430 ocqe = list_first_entry(&ctx->cq_overflow_list,
1431 struct io_overflow_cqe, list);
1433 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1435 WRITE_ONCE(ctx->rings->cq_overflow,
1436 ++ctx->cached_cq_overflow);
1438 list_del(&ocqe->list);
1442 all_flushed = list_empty(&ctx->cq_overflow_list);
1444 clear_bit(0, &ctx->sq_check_overflow);
1445 clear_bit(0, &ctx->cq_check_overflow);
1446 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1450 io_commit_cqring(ctx);
1451 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1453 io_cqring_ev_posted(ctx);
1457 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1461 if (test_bit(0, &ctx->cq_check_overflow)) {
1462 /* iopoll syncs against uring_lock, not completion_lock */
1463 if (ctx->flags & IORING_SETUP_IOPOLL)
1464 mutex_lock(&ctx->uring_lock);
1465 ret = __io_cqring_overflow_flush(ctx, force);
1466 if (ctx->flags & IORING_SETUP_IOPOLL)
1467 mutex_unlock(&ctx->uring_lock);
1474 * Shamelessly stolen from the mm implementation of page reference checking,
1475 * see commit f958d7b528b1 for details.
1477 #define req_ref_zero_or_close_to_overflow(req) \
1478 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1480 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1482 return atomic_inc_not_zero(&req->refs);
1485 static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1487 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1488 return atomic_sub_and_test(refs, &req->refs);
1491 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1493 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1494 return atomic_dec_and_test(&req->refs);
1497 static inline void req_ref_put(struct io_kiocb *req)
1499 WARN_ON_ONCE(req_ref_put_and_test(req));
1502 static inline void req_ref_get(struct io_kiocb *req)
1504 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1505 atomic_inc(&req->refs);
1508 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1509 long res, unsigned int cflags)
1511 struct io_overflow_cqe *ocqe;
1513 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1516 * If we're in ring overflow flush mode, or in task cancel mode,
1517 * or cannot allocate an overflow entry, then we need to drop it
1520 WRITE_ONCE(ctx->rings->cq_overflow, ++ctx->cached_cq_overflow);
1523 if (list_empty(&ctx->cq_overflow_list)) {
1524 set_bit(0, &ctx->sq_check_overflow);
1525 set_bit(0, &ctx->cq_check_overflow);
1526 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1528 ocqe->cqe.user_data = user_data;
1529 ocqe->cqe.res = res;
1530 ocqe->cqe.flags = cflags;
1531 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1535 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1536 long res, unsigned int cflags)
1538 struct io_uring_cqe *cqe;
1540 trace_io_uring_complete(ctx, user_data, res, cflags);
1543 * If we can't get a cq entry, userspace overflowed the
1544 * submission (by quite a lot). Increment the overflow count in
1547 cqe = io_get_cqring(ctx);
1549 WRITE_ONCE(cqe->user_data, user_data);
1550 WRITE_ONCE(cqe->res, res);
1551 WRITE_ONCE(cqe->flags, cflags);
1554 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1557 /* not as hot to bloat with inlining */
1558 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1559 long res, unsigned int cflags)
1561 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1564 static void io_req_complete_post(struct io_kiocb *req, long res,
1565 unsigned int cflags)
1567 struct io_ring_ctx *ctx = req->ctx;
1568 unsigned long flags;
1570 spin_lock_irqsave(&ctx->completion_lock, flags);
1571 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1573 * If we're the last reference to this request, add to our locked
1576 if (req_ref_put_and_test(req)) {
1577 struct io_comp_state *cs = &ctx->submit_state.comp;
1579 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1580 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK))
1581 io_disarm_next(req);
1583 io_req_task_queue(req->link);
1587 io_dismantle_req(req);
1588 io_put_task(req->task, 1);
1589 list_add(&req->compl.list, &cs->locked_free_list);
1590 cs->locked_free_nr++;
1592 if (!percpu_ref_tryget(&ctx->refs))
1595 io_commit_cqring(ctx);
1596 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1599 io_cqring_ev_posted(ctx);
1600 percpu_ref_put(&ctx->refs);
1604 static inline bool io_req_needs_clean(struct io_kiocb *req)
1606 return req->flags & (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP |
1607 REQ_F_POLLED | REQ_F_INFLIGHT);
1610 static void io_req_complete_state(struct io_kiocb *req, long res,
1611 unsigned int cflags)
1613 if (io_req_needs_clean(req))
1616 req->compl.cflags = cflags;
1617 req->flags |= REQ_F_COMPLETE_INLINE;
1620 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1621 long res, unsigned cflags)
1623 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1624 io_req_complete_state(req, res, cflags);
1626 io_req_complete_post(req, res, cflags);
1629 static inline void io_req_complete(struct io_kiocb *req, long res)
1631 __io_req_complete(req, 0, res, 0);
1634 static void io_req_complete_failed(struct io_kiocb *req, long res)
1636 req_set_fail_links(req);
1638 io_req_complete_post(req, res, 0);
1641 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1642 struct io_comp_state *cs)
1644 spin_lock_irq(&ctx->completion_lock);
1645 list_splice_init(&cs->locked_free_list, &cs->free_list);
1646 cs->locked_free_nr = 0;
1647 spin_unlock_irq(&ctx->completion_lock);
1650 /* Returns true IFF there are requests in the cache */
1651 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1653 struct io_submit_state *state = &ctx->submit_state;
1654 struct io_comp_state *cs = &state->comp;
1658 * If we have more than a batch's worth of requests in our IRQ side
1659 * locked cache, grab the lock and move them over to our submission
1662 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH)
1663 io_flush_cached_locked_reqs(ctx, cs);
1665 nr = state->free_reqs;
1666 while (!list_empty(&cs->free_list)) {
1667 struct io_kiocb *req = list_first_entry(&cs->free_list,
1668 struct io_kiocb, compl.list);
1670 list_del(&req->compl.list);
1671 state->reqs[nr++] = req;
1672 if (nr == ARRAY_SIZE(state->reqs))
1676 state->free_reqs = nr;
1680 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1682 struct io_submit_state *state = &ctx->submit_state;
1684 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1686 if (!state->free_reqs) {
1687 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1690 if (io_flush_cached_reqs(ctx))
1693 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1697 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1698 * retry single alloc to be on the safe side.
1700 if (unlikely(ret <= 0)) {
1701 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1702 if (!state->reqs[0])
1706 state->free_reqs = ret;
1710 return state->reqs[state->free_reqs];
1713 static inline void io_put_file(struct file *file)
1719 static void io_dismantle_req(struct io_kiocb *req)
1721 unsigned int flags = req->flags;
1723 if (io_req_needs_clean(req))
1725 if (!(flags & REQ_F_FIXED_FILE))
1726 io_put_file(req->file);
1727 if (req->fixed_rsrc_refs)
1728 percpu_ref_put(req->fixed_rsrc_refs);
1729 if (req->async_data)
1730 kfree(req->async_data);
1731 if (req->work.creds) {
1732 put_cred(req->work.creds);
1733 req->work.creds = NULL;
1737 /* must to be called somewhat shortly after putting a request */
1738 static inline void io_put_task(struct task_struct *task, int nr)
1740 struct io_uring_task *tctx = task->io_uring;
1742 percpu_counter_sub(&tctx->inflight, nr);
1743 if (unlikely(atomic_read(&tctx->in_idle)))
1744 wake_up(&tctx->wait);
1745 put_task_struct_many(task, nr);
1748 static void __io_free_req(struct io_kiocb *req)
1750 struct io_ring_ctx *ctx = req->ctx;
1752 io_dismantle_req(req);
1753 io_put_task(req->task, 1);
1755 kmem_cache_free(req_cachep, req);
1756 percpu_ref_put(&ctx->refs);
1759 static inline void io_remove_next_linked(struct io_kiocb *req)
1761 struct io_kiocb *nxt = req->link;
1763 req->link = nxt->link;
1767 static bool io_kill_linked_timeout(struct io_kiocb *req)
1768 __must_hold(&req->ctx->completion_lock)
1770 struct io_kiocb *link = req->link;
1773 * Can happen if a linked timeout fired and link had been like
1774 * req -> link t-out -> link t-out [-> ...]
1776 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1777 struct io_timeout_data *io = link->async_data;
1779 io_remove_next_linked(req);
1780 link->timeout.head = NULL;
1781 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1782 io_cqring_fill_event(link->ctx, link->user_data,
1784 io_put_req_deferred(link, 1);
1791 static void io_fail_links(struct io_kiocb *req)
1792 __must_hold(&req->ctx->completion_lock)
1794 struct io_kiocb *nxt, *link = req->link;
1801 trace_io_uring_fail_link(req, link);
1802 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1803 io_put_req_deferred(link, 2);
1808 static bool io_disarm_next(struct io_kiocb *req)
1809 __must_hold(&req->ctx->completion_lock)
1811 bool posted = false;
1813 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1814 posted = io_kill_linked_timeout(req);
1815 if (unlikely((req->flags & REQ_F_FAIL_LINK) &&
1816 !(req->flags & REQ_F_HARDLINK))) {
1817 posted |= (req->link != NULL);
1823 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1825 struct io_kiocb *nxt;
1828 * If LINK is set, we have dependent requests in this chain. If we
1829 * didn't fail this request, queue the first one up, moving any other
1830 * dependencies to the next request. In case of failure, fail the rest
1833 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK)) {
1834 struct io_ring_ctx *ctx = req->ctx;
1835 unsigned long flags;
1838 spin_lock_irqsave(&ctx->completion_lock, flags);
1839 posted = io_disarm_next(req);
1841 io_commit_cqring(req->ctx);
1842 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1844 io_cqring_ev_posted(ctx);
1851 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1853 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1855 return __io_req_find_next(req);
1858 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1862 if (ctx->submit_state.comp.nr) {
1863 mutex_lock(&ctx->uring_lock);
1864 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1865 mutex_unlock(&ctx->uring_lock);
1867 percpu_ref_put(&ctx->refs);
1870 static bool __tctx_task_work(struct io_uring_task *tctx)
1872 struct io_ring_ctx *ctx = NULL;
1873 struct io_wq_work_list list;
1874 struct io_wq_work_node *node;
1876 if (wq_list_empty(&tctx->task_list))
1879 spin_lock_irq(&tctx->task_lock);
1880 list = tctx->task_list;
1881 INIT_WQ_LIST(&tctx->task_list);
1882 spin_unlock_irq(&tctx->task_lock);
1886 struct io_wq_work_node *next = node->next;
1887 struct io_kiocb *req;
1889 req = container_of(node, struct io_kiocb, io_task_work.node);
1890 if (req->ctx != ctx) {
1891 ctx_flush_and_put(ctx);
1893 percpu_ref_get(&ctx->refs);
1896 req->task_work.func(&req->task_work);
1900 ctx_flush_and_put(ctx);
1901 return list.first != NULL;
1904 static void tctx_task_work(struct callback_head *cb)
1906 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1908 clear_bit(0, &tctx->task_state);
1910 while (__tctx_task_work(tctx))
1914 static int io_req_task_work_add(struct io_kiocb *req)
1916 struct task_struct *tsk = req->task;
1917 struct io_uring_task *tctx = tsk->io_uring;
1918 enum task_work_notify_mode notify;
1919 struct io_wq_work_node *node, *prev;
1920 unsigned long flags;
1923 if (unlikely(tsk->flags & PF_EXITING))
1926 WARN_ON_ONCE(!tctx);
1928 spin_lock_irqsave(&tctx->task_lock, flags);
1929 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1930 spin_unlock_irqrestore(&tctx->task_lock, flags);
1932 /* task_work already pending, we're done */
1933 if (test_bit(0, &tctx->task_state) ||
1934 test_and_set_bit(0, &tctx->task_state))
1938 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1939 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1940 * processing task_work. There's no reliable way to tell if TWA_RESUME
1943 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
1945 if (!task_work_add(tsk, &tctx->task_work, notify)) {
1946 wake_up_process(tsk);
1951 * Slow path - we failed, find and delete work. if the work is not
1952 * in the list, it got run and we're fine.
1954 spin_lock_irqsave(&tctx->task_lock, flags);
1955 wq_list_for_each(node, prev, &tctx->task_list) {
1956 if (&req->io_task_work.node == node) {
1957 wq_list_del(&tctx->task_list, node, prev);
1962 spin_unlock_irqrestore(&tctx->task_lock, flags);
1963 clear_bit(0, &tctx->task_state);
1967 static bool io_run_task_work_head(struct callback_head **work_head)
1969 struct callback_head *work, *next;
1970 bool executed = false;
1973 work = xchg(work_head, NULL);
1989 static void io_task_work_add_head(struct callback_head **work_head,
1990 struct callback_head *task_work)
1992 struct callback_head *head;
1995 head = READ_ONCE(*work_head);
1996 task_work->next = head;
1997 } while (cmpxchg(work_head, head, task_work) != head);
2000 static void io_req_task_work_add_fallback(struct io_kiocb *req,
2001 task_work_func_t cb)
2003 init_task_work(&req->task_work, cb);
2004 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
2007 static void io_req_task_cancel(struct callback_head *cb)
2009 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2010 struct io_ring_ctx *ctx = req->ctx;
2012 /* ctx is guaranteed to stay alive while we hold uring_lock */
2013 mutex_lock(&ctx->uring_lock);
2014 io_req_complete_failed(req, req->result);
2015 mutex_unlock(&ctx->uring_lock);
2018 static void __io_req_task_submit(struct io_kiocb *req)
2020 struct io_ring_ctx *ctx = req->ctx;
2022 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2023 mutex_lock(&ctx->uring_lock);
2024 if (!(current->flags & PF_EXITING) && !current->in_execve)
2025 __io_queue_sqe(req);
2027 io_req_complete_failed(req, -EFAULT);
2028 mutex_unlock(&ctx->uring_lock);
2031 static void io_req_task_submit(struct callback_head *cb)
2033 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2035 __io_req_task_submit(req);
2038 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2041 req->task_work.func = io_req_task_cancel;
2043 if (unlikely(io_req_task_work_add(req)))
2044 io_req_task_work_add_fallback(req, io_req_task_cancel);
2047 static void io_req_task_queue(struct io_kiocb *req)
2049 req->task_work.func = io_req_task_submit;
2051 if (unlikely(io_req_task_work_add(req)))
2052 io_req_task_queue_fail(req, -ECANCELED);
2055 static inline void io_queue_next(struct io_kiocb *req)
2057 struct io_kiocb *nxt = io_req_find_next(req);
2060 io_req_task_queue(nxt);
2063 static void io_free_req(struct io_kiocb *req)
2070 struct task_struct *task;
2075 static inline void io_init_req_batch(struct req_batch *rb)
2082 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2083 struct req_batch *rb)
2086 io_put_task(rb->task, rb->task_refs);
2088 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2091 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2092 struct io_submit_state *state)
2095 io_dismantle_req(req);
2097 if (req->task != rb->task) {
2099 io_put_task(rb->task, rb->task_refs);
2100 rb->task = req->task;
2106 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2107 state->reqs[state->free_reqs++] = req;
2109 list_add(&req->compl.list, &state->comp.free_list);
2112 static void io_submit_flush_completions(struct io_comp_state *cs,
2113 struct io_ring_ctx *ctx)
2116 struct io_kiocb *req;
2117 struct req_batch rb;
2119 io_init_req_batch(&rb);
2120 spin_lock_irq(&ctx->completion_lock);
2121 for (i = 0; i < nr; i++) {
2123 __io_cqring_fill_event(ctx, req->user_data, req->result,
2126 io_commit_cqring(ctx);
2127 spin_unlock_irq(&ctx->completion_lock);
2129 io_cqring_ev_posted(ctx);
2130 for (i = 0; i < nr; i++) {
2133 /* submission and completion refs */
2134 if (req_ref_sub_and_test(req, 2))
2135 io_req_free_batch(&rb, req, &ctx->submit_state);
2138 io_req_free_batch_finish(ctx, &rb);
2143 * Drop reference to request, return next in chain (if there is one) if this
2144 * was the last reference to this request.
2146 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2148 struct io_kiocb *nxt = NULL;
2150 if (req_ref_put_and_test(req)) {
2151 nxt = io_req_find_next(req);
2157 static inline void io_put_req(struct io_kiocb *req)
2159 if (req_ref_put_and_test(req))
2163 static void io_put_req_deferred_cb(struct callback_head *cb)
2165 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2170 static void io_free_req_deferred(struct io_kiocb *req)
2172 req->task_work.func = io_put_req_deferred_cb;
2173 if (unlikely(io_req_task_work_add(req)))
2174 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2177 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2179 if (req_ref_sub_and_test(req, refs))
2180 io_free_req_deferred(req);
2183 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2185 /* See comment at the top of this file */
2187 return __io_cqring_events(ctx);
2190 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2192 struct io_rings *rings = ctx->rings;
2194 /* make sure SQ entry isn't read before tail */
2195 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2198 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2200 unsigned int cflags;
2202 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2203 cflags |= IORING_CQE_F_BUFFER;
2204 req->flags &= ~REQ_F_BUFFER_SELECTED;
2209 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2211 struct io_buffer *kbuf;
2213 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2214 return io_put_kbuf(req, kbuf);
2217 static inline bool io_run_task_work(void)
2220 * Not safe to run on exiting task, and the task_work handling will
2221 * not add work to such a task.
2223 if (unlikely(current->flags & PF_EXITING))
2225 if (current->task_works) {
2226 __set_current_state(TASK_RUNNING);
2235 * Find and free completed poll iocbs
2237 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2238 struct list_head *done)
2240 struct req_batch rb;
2241 struct io_kiocb *req;
2243 /* order with ->result store in io_complete_rw_iopoll() */
2246 io_init_req_batch(&rb);
2247 while (!list_empty(done)) {
2250 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2251 list_del(&req->inflight_entry);
2253 if (READ_ONCE(req->result) == -EAGAIN &&
2254 !(req->flags & REQ_F_DONT_REISSUE)) {
2255 req->iopoll_completed = 0;
2257 io_queue_async_work(req);
2261 if (req->flags & REQ_F_BUFFER_SELECTED)
2262 cflags = io_put_rw_kbuf(req);
2264 __io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2267 if (req_ref_put_and_test(req))
2268 io_req_free_batch(&rb, req, &ctx->submit_state);
2271 io_commit_cqring(ctx);
2272 io_cqring_ev_posted_iopoll(ctx);
2273 io_req_free_batch_finish(ctx, &rb);
2276 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2279 struct io_kiocb *req, *tmp;
2285 * Only spin for completions if we don't have multiple devices hanging
2286 * off our complete list, and we're under the requested amount.
2288 spin = !ctx->poll_multi_file && *nr_events < min;
2291 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2292 struct kiocb *kiocb = &req->rw.kiocb;
2295 * Move completed and retryable entries to our local lists.
2296 * If we find a request that requires polling, break out
2297 * and complete those lists first, if we have entries there.
2299 if (READ_ONCE(req->iopoll_completed)) {
2300 list_move_tail(&req->inflight_entry, &done);
2303 if (!list_empty(&done))
2306 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2310 /* iopoll may have completed current req */
2311 if (READ_ONCE(req->iopoll_completed))
2312 list_move_tail(&req->inflight_entry, &done);
2319 if (!list_empty(&done))
2320 io_iopoll_complete(ctx, nr_events, &done);
2326 * We can't just wait for polled events to come to us, we have to actively
2327 * find and complete them.
2329 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2331 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2334 mutex_lock(&ctx->uring_lock);
2335 while (!list_empty(&ctx->iopoll_list)) {
2336 unsigned int nr_events = 0;
2338 io_do_iopoll(ctx, &nr_events, 0);
2340 /* let it sleep and repeat later if can't complete a request */
2344 * Ensure we allow local-to-the-cpu processing to take place,
2345 * in this case we need to ensure that we reap all events.
2346 * Also let task_work, etc. to progress by releasing the mutex
2348 if (need_resched()) {
2349 mutex_unlock(&ctx->uring_lock);
2351 mutex_lock(&ctx->uring_lock);
2354 mutex_unlock(&ctx->uring_lock);
2357 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2359 unsigned int nr_events = 0;
2363 * We disallow the app entering submit/complete with polling, but we
2364 * still need to lock the ring to prevent racing with polled issue
2365 * that got punted to a workqueue.
2367 mutex_lock(&ctx->uring_lock);
2369 * Don't enter poll loop if we already have events pending.
2370 * If we do, we can potentially be spinning for commands that
2371 * already triggered a CQE (eg in error).
2373 if (test_bit(0, &ctx->cq_check_overflow))
2374 __io_cqring_overflow_flush(ctx, false);
2375 if (io_cqring_events(ctx))
2379 * If a submit got punted to a workqueue, we can have the
2380 * application entering polling for a command before it gets
2381 * issued. That app will hold the uring_lock for the duration
2382 * of the poll right here, so we need to take a breather every
2383 * now and then to ensure that the issue has a chance to add
2384 * the poll to the issued list. Otherwise we can spin here
2385 * forever, while the workqueue is stuck trying to acquire the
2388 if (list_empty(&ctx->iopoll_list)) {
2389 mutex_unlock(&ctx->uring_lock);
2391 mutex_lock(&ctx->uring_lock);
2393 if (list_empty(&ctx->iopoll_list))
2396 ret = io_do_iopoll(ctx, &nr_events, min);
2397 } while (!ret && nr_events < min && !need_resched());
2399 mutex_unlock(&ctx->uring_lock);
2403 static void kiocb_end_write(struct io_kiocb *req)
2406 * Tell lockdep we inherited freeze protection from submission
2409 if (req->flags & REQ_F_ISREG) {
2410 struct super_block *sb = file_inode(req->file)->i_sb;
2412 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2418 static bool io_resubmit_prep(struct io_kiocb *req)
2420 struct io_async_rw *rw = req->async_data;
2423 return !io_req_prep_async(req);
2424 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2425 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2429 static bool io_rw_should_reissue(struct io_kiocb *req)
2431 umode_t mode = file_inode(req->file)->i_mode;
2432 struct io_ring_ctx *ctx = req->ctx;
2434 if (!S_ISBLK(mode) && !S_ISREG(mode))
2436 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2437 !(ctx->flags & IORING_SETUP_IOPOLL)))
2440 * If ref is dying, we might be running poll reap from the exit work.
2441 * Don't attempt to reissue from that path, just let it fail with
2444 if (percpu_ref_is_dying(&ctx->refs))
2449 static bool io_resubmit_prep(struct io_kiocb *req)
2453 static bool io_rw_should_reissue(struct io_kiocb *req)
2459 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2460 unsigned int issue_flags)
2464 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2465 kiocb_end_write(req);
2466 if (res != req->result) {
2467 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2468 io_rw_should_reissue(req)) {
2469 req->flags |= REQ_F_REISSUE;
2472 req_set_fail_links(req);
2474 if (req->flags & REQ_F_BUFFER_SELECTED)
2475 cflags = io_put_rw_kbuf(req);
2476 __io_req_complete(req, issue_flags, res, cflags);
2479 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2481 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2483 __io_complete_rw(req, res, res2, 0);
2486 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2488 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2490 if (kiocb->ki_flags & IOCB_WRITE)
2491 kiocb_end_write(req);
2492 if (unlikely(res != req->result)) {
2493 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2494 io_resubmit_prep(req))) {
2495 req_set_fail_links(req);
2496 req->flags |= REQ_F_DONT_REISSUE;
2500 WRITE_ONCE(req->result, res);
2501 /* order with io_iopoll_complete() checking ->result */
2503 WRITE_ONCE(req->iopoll_completed, 1);
2507 * After the iocb has been issued, it's safe to be found on the poll list.
2508 * Adding the kiocb to the list AFTER submission ensures that we don't
2509 * find it from a io_do_iopoll() thread before the issuer is done
2510 * accessing the kiocb cookie.
2512 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2514 struct io_ring_ctx *ctx = req->ctx;
2517 * Track whether we have multiple files in our lists. This will impact
2518 * how we do polling eventually, not spinning if we're on potentially
2519 * different devices.
2521 if (list_empty(&ctx->iopoll_list)) {
2522 ctx->poll_multi_file = false;
2523 } else if (!ctx->poll_multi_file) {
2524 struct io_kiocb *list_req;
2526 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2528 if (list_req->file != req->file)
2529 ctx->poll_multi_file = true;
2533 * For fast devices, IO may have already completed. If it has, add
2534 * it to the front so we find it first.
2536 if (READ_ONCE(req->iopoll_completed))
2537 list_add(&req->inflight_entry, &ctx->iopoll_list);
2539 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2542 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2543 * task context or in io worker task context. If current task context is
2544 * sq thread, we don't need to check whether should wake up sq thread.
2546 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2547 wq_has_sleeper(&ctx->sq_data->wait))
2548 wake_up(&ctx->sq_data->wait);
2551 static inline void io_state_file_put(struct io_submit_state *state)
2553 if (state->file_refs) {
2554 fput_many(state->file, state->file_refs);
2555 state->file_refs = 0;
2560 * Get as many references to a file as we have IOs left in this submission,
2561 * assuming most submissions are for one file, or at least that each file
2562 * has more than one submission.
2564 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2569 if (state->file_refs) {
2570 if (state->fd == fd) {
2574 io_state_file_put(state);
2576 state->file = fget_many(fd, state->ios_left);
2577 if (unlikely(!state->file))
2581 state->file_refs = state->ios_left - 1;
2585 static bool io_bdev_nowait(struct block_device *bdev)
2587 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2591 * If we tracked the file through the SCM inflight mechanism, we could support
2592 * any file. For now, just ensure that anything potentially problematic is done
2595 static bool __io_file_supports_async(struct file *file, int rw)
2597 umode_t mode = file_inode(file)->i_mode;
2599 if (S_ISBLK(mode)) {
2600 if (IS_ENABLED(CONFIG_BLOCK) &&
2601 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2605 if (S_ISCHR(mode) || S_ISSOCK(mode))
2607 if (S_ISREG(mode)) {
2608 if (IS_ENABLED(CONFIG_BLOCK) &&
2609 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2610 file->f_op != &io_uring_fops)
2615 /* any ->read/write should understand O_NONBLOCK */
2616 if (file->f_flags & O_NONBLOCK)
2619 if (!(file->f_mode & FMODE_NOWAIT))
2623 return file->f_op->read_iter != NULL;
2625 return file->f_op->write_iter != NULL;
2628 static bool io_file_supports_async(struct io_kiocb *req, int rw)
2630 if (rw == READ && (req->flags & REQ_F_ASYNC_READ))
2632 else if (rw == WRITE && (req->flags & REQ_F_ASYNC_WRITE))
2635 return __io_file_supports_async(req->file, rw);
2638 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2640 struct io_ring_ctx *ctx = req->ctx;
2641 struct kiocb *kiocb = &req->rw.kiocb;
2642 struct file *file = req->file;
2646 if (!(req->flags & REQ_F_ISREG) && S_ISREG(file_inode(file)->i_mode))
2647 req->flags |= REQ_F_ISREG;
2649 kiocb->ki_pos = READ_ONCE(sqe->off);
2650 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2651 req->flags |= REQ_F_CUR_POS;
2652 kiocb->ki_pos = file->f_pos;
2654 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2655 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2656 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2660 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2661 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2662 req->flags |= REQ_F_NOWAIT;
2664 ioprio = READ_ONCE(sqe->ioprio);
2666 ret = ioprio_check_cap(ioprio);
2670 kiocb->ki_ioprio = ioprio;
2672 kiocb->ki_ioprio = get_current_ioprio();
2674 if (ctx->flags & IORING_SETUP_IOPOLL) {
2675 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2676 !kiocb->ki_filp->f_op->iopoll)
2679 kiocb->ki_flags |= IOCB_HIPRI;
2680 kiocb->ki_complete = io_complete_rw_iopoll;
2681 req->iopoll_completed = 0;
2683 if (kiocb->ki_flags & IOCB_HIPRI)
2685 kiocb->ki_complete = io_complete_rw;
2688 if (req->opcode == IORING_OP_READ_FIXED ||
2689 req->opcode == IORING_OP_WRITE_FIXED) {
2691 io_req_set_rsrc_node(req);
2694 req->rw.addr = READ_ONCE(sqe->addr);
2695 req->rw.len = READ_ONCE(sqe->len);
2696 req->buf_index = READ_ONCE(sqe->buf_index);
2700 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2706 case -ERESTARTNOINTR:
2707 case -ERESTARTNOHAND:
2708 case -ERESTART_RESTARTBLOCK:
2710 * We can't just restart the syscall, since previously
2711 * submitted sqes may already be in progress. Just fail this
2717 kiocb->ki_complete(kiocb, ret, 0);
2721 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2722 unsigned int issue_flags)
2724 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2725 struct io_async_rw *io = req->async_data;
2726 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2728 /* add previously done IO, if any */
2729 if (io && io->bytes_done > 0) {
2731 ret = io->bytes_done;
2733 ret += io->bytes_done;
2736 if (req->flags & REQ_F_CUR_POS)
2737 req->file->f_pos = kiocb->ki_pos;
2738 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2739 __io_complete_rw(req, ret, 0, issue_flags);
2741 io_rw_done(kiocb, ret);
2743 if (check_reissue && req->flags & REQ_F_REISSUE) {
2744 req->flags &= ~REQ_F_REISSUE;
2745 if (io_resubmit_prep(req)) {
2747 io_queue_async_work(req);
2751 req_set_fail_links(req);
2752 if (req->flags & REQ_F_BUFFER_SELECTED)
2753 cflags = io_put_rw_kbuf(req);
2754 __io_req_complete(req, issue_flags, ret, cflags);
2759 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2760 struct io_mapped_ubuf *imu)
2762 size_t len = req->rw.len;
2763 u64 buf_end, buf_addr = req->rw.addr;
2766 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2768 /* not inside the mapped region */
2769 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2773 * May not be a start of buffer, set size appropriately
2774 * and advance us to the beginning.
2776 offset = buf_addr - imu->ubuf;
2777 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2781 * Don't use iov_iter_advance() here, as it's really slow for
2782 * using the latter parts of a big fixed buffer - it iterates
2783 * over each segment manually. We can cheat a bit here, because
2786 * 1) it's a BVEC iter, we set it up
2787 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2788 * first and last bvec
2790 * So just find our index, and adjust the iterator afterwards.
2791 * If the offset is within the first bvec (or the whole first
2792 * bvec, just use iov_iter_advance(). This makes it easier
2793 * since we can just skip the first segment, which may not
2794 * be PAGE_SIZE aligned.
2796 const struct bio_vec *bvec = imu->bvec;
2798 if (offset <= bvec->bv_len) {
2799 iov_iter_advance(iter, offset);
2801 unsigned long seg_skip;
2803 /* skip first vec */
2804 offset -= bvec->bv_len;
2805 seg_skip = 1 + (offset >> PAGE_SHIFT);
2807 iter->bvec = bvec + seg_skip;
2808 iter->nr_segs -= seg_skip;
2809 iter->count -= bvec->bv_len + offset;
2810 iter->iov_offset = offset & ~PAGE_MASK;
2817 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2819 struct io_ring_ctx *ctx = req->ctx;
2820 struct io_mapped_ubuf *imu = req->imu;
2821 u16 index, buf_index = req->buf_index;
2824 if (unlikely(buf_index >= ctx->nr_user_bufs))
2826 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2827 imu = READ_ONCE(ctx->user_bufs[index]);
2830 return __io_import_fixed(req, rw, iter, imu);
2833 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2836 mutex_unlock(&ctx->uring_lock);
2839 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2842 * "Normal" inline submissions always hold the uring_lock, since we
2843 * grab it from the system call. Same is true for the SQPOLL offload.
2844 * The only exception is when we've detached the request and issue it
2845 * from an async worker thread, grab the lock for that case.
2848 mutex_lock(&ctx->uring_lock);
2851 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2852 int bgid, struct io_buffer *kbuf,
2855 struct io_buffer *head;
2857 if (req->flags & REQ_F_BUFFER_SELECTED)
2860 io_ring_submit_lock(req->ctx, needs_lock);
2862 lockdep_assert_held(&req->ctx->uring_lock);
2864 head = xa_load(&req->ctx->io_buffers, bgid);
2866 if (!list_empty(&head->list)) {
2867 kbuf = list_last_entry(&head->list, struct io_buffer,
2869 list_del(&kbuf->list);
2872 xa_erase(&req->ctx->io_buffers, bgid);
2874 if (*len > kbuf->len)
2877 kbuf = ERR_PTR(-ENOBUFS);
2880 io_ring_submit_unlock(req->ctx, needs_lock);
2885 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2888 struct io_buffer *kbuf;
2891 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2892 bgid = req->buf_index;
2893 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2896 req->rw.addr = (u64) (unsigned long) kbuf;
2897 req->flags |= REQ_F_BUFFER_SELECTED;
2898 return u64_to_user_ptr(kbuf->addr);
2901 #ifdef CONFIG_COMPAT
2902 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2905 struct compat_iovec __user *uiov;
2906 compat_ssize_t clen;
2910 uiov = u64_to_user_ptr(req->rw.addr);
2911 if (!access_ok(uiov, sizeof(*uiov)))
2913 if (__get_user(clen, &uiov->iov_len))
2919 buf = io_rw_buffer_select(req, &len, needs_lock);
2921 return PTR_ERR(buf);
2922 iov[0].iov_base = buf;
2923 iov[0].iov_len = (compat_size_t) len;
2928 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2931 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2935 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2938 len = iov[0].iov_len;
2941 buf = io_rw_buffer_select(req, &len, needs_lock);
2943 return PTR_ERR(buf);
2944 iov[0].iov_base = buf;
2945 iov[0].iov_len = len;
2949 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2952 if (req->flags & REQ_F_BUFFER_SELECTED) {
2953 struct io_buffer *kbuf;
2955 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2956 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2957 iov[0].iov_len = kbuf->len;
2960 if (req->rw.len != 1)
2963 #ifdef CONFIG_COMPAT
2964 if (req->ctx->compat)
2965 return io_compat_import(req, iov, needs_lock);
2968 return __io_iov_buffer_select(req, iov, needs_lock);
2971 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2972 struct iov_iter *iter, bool needs_lock)
2974 void __user *buf = u64_to_user_ptr(req->rw.addr);
2975 size_t sqe_len = req->rw.len;
2976 u8 opcode = req->opcode;
2979 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2981 return io_import_fixed(req, rw, iter);
2984 /* buffer index only valid with fixed read/write, or buffer select */
2985 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2988 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2989 if (req->flags & REQ_F_BUFFER_SELECT) {
2990 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2992 return PTR_ERR(buf);
2993 req->rw.len = sqe_len;
2996 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3001 if (req->flags & REQ_F_BUFFER_SELECT) {
3002 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3004 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3009 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3013 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3015 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3019 * For files that don't have ->read_iter() and ->write_iter(), handle them
3020 * by looping over ->read() or ->write() manually.
3022 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3024 struct kiocb *kiocb = &req->rw.kiocb;
3025 struct file *file = req->file;
3029 * Don't support polled IO through this interface, and we can't
3030 * support non-blocking either. For the latter, this just causes
3031 * the kiocb to be handled from an async context.
3033 if (kiocb->ki_flags & IOCB_HIPRI)
3035 if (kiocb->ki_flags & IOCB_NOWAIT)
3038 while (iov_iter_count(iter)) {
3042 if (!iov_iter_is_bvec(iter)) {
3043 iovec = iov_iter_iovec(iter);
3045 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3046 iovec.iov_len = req->rw.len;
3050 nr = file->f_op->read(file, iovec.iov_base,
3051 iovec.iov_len, io_kiocb_ppos(kiocb));
3053 nr = file->f_op->write(file, iovec.iov_base,
3054 iovec.iov_len, io_kiocb_ppos(kiocb));
3063 if (nr != iovec.iov_len)
3067 iov_iter_advance(iter, nr);
3073 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3074 const struct iovec *fast_iov, struct iov_iter *iter)
3076 struct io_async_rw *rw = req->async_data;
3078 memcpy(&rw->iter, iter, sizeof(*iter));
3079 rw->free_iovec = iovec;
3081 /* can only be fixed buffers, no need to do anything */
3082 if (iov_iter_is_bvec(iter))
3085 unsigned iov_off = 0;
3087 rw->iter.iov = rw->fast_iov;
3088 if (iter->iov != fast_iov) {
3089 iov_off = iter->iov - fast_iov;
3090 rw->iter.iov += iov_off;
3092 if (rw->fast_iov != fast_iov)
3093 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3094 sizeof(struct iovec) * iter->nr_segs);
3096 req->flags |= REQ_F_NEED_CLEANUP;
3100 static inline int io_alloc_async_data(struct io_kiocb *req)
3102 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3103 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3104 return req->async_data == NULL;
3107 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3108 const struct iovec *fast_iov,
3109 struct iov_iter *iter, bool force)
3111 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3113 if (!req->async_data) {
3114 if (io_alloc_async_data(req)) {
3119 io_req_map_rw(req, iovec, fast_iov, iter);
3124 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3126 struct io_async_rw *iorw = req->async_data;
3127 struct iovec *iov = iorw->fast_iov;
3130 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3131 if (unlikely(ret < 0))
3134 iorw->bytes_done = 0;
3135 iorw->free_iovec = iov;
3137 req->flags |= REQ_F_NEED_CLEANUP;
3141 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3143 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3145 return io_prep_rw(req, sqe);
3149 * This is our waitqueue callback handler, registered through lock_page_async()
3150 * when we initially tried to do the IO with the iocb armed our waitqueue.
3151 * This gets called when the page is unlocked, and we generally expect that to
3152 * happen when the page IO is completed and the page is now uptodate. This will
3153 * queue a task_work based retry of the operation, attempting to copy the data
3154 * again. If the latter fails because the page was NOT uptodate, then we will
3155 * do a thread based blocking retry of the operation. That's the unexpected
3158 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3159 int sync, void *arg)
3161 struct wait_page_queue *wpq;
3162 struct io_kiocb *req = wait->private;
3163 struct wait_page_key *key = arg;
3165 wpq = container_of(wait, struct wait_page_queue, wait);
3167 if (!wake_page_match(wpq, key))
3170 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3171 list_del_init(&wait->entry);
3173 /* submit ref gets dropped, acquire a new one */
3175 io_req_task_queue(req);
3180 * This controls whether a given IO request should be armed for async page
3181 * based retry. If we return false here, the request is handed to the async
3182 * worker threads for retry. If we're doing buffered reads on a regular file,
3183 * we prepare a private wait_page_queue entry and retry the operation. This
3184 * will either succeed because the page is now uptodate and unlocked, or it
3185 * will register a callback when the page is unlocked at IO completion. Through
3186 * that callback, io_uring uses task_work to setup a retry of the operation.
3187 * That retry will attempt the buffered read again. The retry will generally
3188 * succeed, or in rare cases where it fails, we then fall back to using the
3189 * async worker threads for a blocking retry.
3191 static bool io_rw_should_retry(struct io_kiocb *req)
3193 struct io_async_rw *rw = req->async_data;
3194 struct wait_page_queue *wait = &rw->wpq;
3195 struct kiocb *kiocb = &req->rw.kiocb;
3197 /* never retry for NOWAIT, we just complete with -EAGAIN */
3198 if (req->flags & REQ_F_NOWAIT)
3201 /* Only for buffered IO */
3202 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3206 * just use poll if we can, and don't attempt if the fs doesn't
3207 * support callback based unlocks
3209 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3212 wait->wait.func = io_async_buf_func;
3213 wait->wait.private = req;
3214 wait->wait.flags = 0;
3215 INIT_LIST_HEAD(&wait->wait.entry);
3216 kiocb->ki_flags |= IOCB_WAITQ;
3217 kiocb->ki_flags &= ~IOCB_NOWAIT;
3218 kiocb->ki_waitq = wait;
3222 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3224 if (req->file->f_op->read_iter)
3225 return call_read_iter(req->file, &req->rw.kiocb, iter);
3226 else if (req->file->f_op->read)
3227 return loop_rw_iter(READ, req, iter);
3232 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3234 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3235 struct kiocb *kiocb = &req->rw.kiocb;
3236 struct iov_iter __iter, *iter = &__iter;
3237 struct io_async_rw *rw = req->async_data;
3238 ssize_t io_size, ret, ret2;
3239 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3245 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3249 io_size = iov_iter_count(iter);
3250 req->result = io_size;
3252 /* Ensure we clear previously set non-block flag */
3253 if (!force_nonblock)
3254 kiocb->ki_flags &= ~IOCB_NOWAIT;
3256 kiocb->ki_flags |= IOCB_NOWAIT;
3258 /* If the file doesn't support async, just async punt */
3259 if (force_nonblock && !io_file_supports_async(req, READ)) {
3260 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3261 return ret ?: -EAGAIN;
3264 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3265 if (unlikely(ret)) {
3270 ret = io_iter_do_read(req, iter);
3272 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3273 req->flags &= ~REQ_F_REISSUE;
3274 /* IOPOLL retry should happen for io-wq threads */
3275 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3277 /* no retry on NONBLOCK nor RWF_NOWAIT */
3278 if (req->flags & REQ_F_NOWAIT)
3280 /* some cases will consume bytes even on error returns */
3281 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3283 } else if (ret == -EIOCBQUEUED) {
3285 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3286 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3287 /* read all, failed, already did sync or don't want to retry */
3291 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3296 rw = req->async_data;
3297 /* now use our persistent iterator, if we aren't already */
3302 rw->bytes_done += ret;
3303 /* if we can retry, do so with the callbacks armed */
3304 if (!io_rw_should_retry(req)) {
3305 kiocb->ki_flags &= ~IOCB_WAITQ;
3310 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3311 * we get -EIOCBQUEUED, then we'll get a notification when the
3312 * desired page gets unlocked. We can also get a partial read
3313 * here, and if we do, then just retry at the new offset.
3315 ret = io_iter_do_read(req, iter);
3316 if (ret == -EIOCBQUEUED)
3318 /* we got some bytes, but not all. retry. */
3319 kiocb->ki_flags &= ~IOCB_WAITQ;
3320 } while (ret > 0 && ret < io_size);
3322 kiocb_done(kiocb, ret, issue_flags);
3324 /* it's faster to check here then delegate to kfree */
3330 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3332 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3334 return io_prep_rw(req, sqe);
3337 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3339 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3340 struct kiocb *kiocb = &req->rw.kiocb;
3341 struct iov_iter __iter, *iter = &__iter;
3342 struct io_async_rw *rw = req->async_data;
3343 ssize_t ret, ret2, io_size;
3344 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3350 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3354 io_size = iov_iter_count(iter);
3355 req->result = io_size;
3357 /* Ensure we clear previously set non-block flag */
3358 if (!force_nonblock)
3359 kiocb->ki_flags &= ~IOCB_NOWAIT;
3361 kiocb->ki_flags |= IOCB_NOWAIT;
3363 /* If the file doesn't support async, just async punt */
3364 if (force_nonblock && !io_file_supports_async(req, WRITE))
3367 /* file path doesn't support NOWAIT for non-direct_IO */
3368 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3369 (req->flags & REQ_F_ISREG))
3372 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3377 * Open-code file_start_write here to grab freeze protection,
3378 * which will be released by another thread in
3379 * io_complete_rw(). Fool lockdep by telling it the lock got
3380 * released so that it doesn't complain about the held lock when
3381 * we return to userspace.
3383 if (req->flags & REQ_F_ISREG) {
3384 sb_start_write(file_inode(req->file)->i_sb);
3385 __sb_writers_release(file_inode(req->file)->i_sb,
3388 kiocb->ki_flags |= IOCB_WRITE;
3390 if (req->file->f_op->write_iter)
3391 ret2 = call_write_iter(req->file, kiocb, iter);
3392 else if (req->file->f_op->write)
3393 ret2 = loop_rw_iter(WRITE, req, iter);
3397 if (req->flags & REQ_F_REISSUE) {
3398 req->flags &= ~REQ_F_REISSUE;
3403 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3404 * retry them without IOCB_NOWAIT.
3406 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3408 /* no retry on NONBLOCK nor RWF_NOWAIT */
3409 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3411 if (!force_nonblock || ret2 != -EAGAIN) {
3412 /* IOPOLL retry should happen for io-wq threads */
3413 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3416 kiocb_done(kiocb, ret2, issue_flags);
3419 /* some cases will consume bytes even on error returns */
3420 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3421 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3422 return ret ?: -EAGAIN;
3425 /* it's reportedly faster than delegating the null check to kfree() */
3431 static int io_renameat_prep(struct io_kiocb *req,
3432 const struct io_uring_sqe *sqe)
3434 struct io_rename *ren = &req->rename;
3435 const char __user *oldf, *newf;
3437 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3440 ren->old_dfd = READ_ONCE(sqe->fd);
3441 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3442 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3443 ren->new_dfd = READ_ONCE(sqe->len);
3444 ren->flags = READ_ONCE(sqe->rename_flags);
3446 ren->oldpath = getname(oldf);
3447 if (IS_ERR(ren->oldpath))
3448 return PTR_ERR(ren->oldpath);
3450 ren->newpath = getname(newf);
3451 if (IS_ERR(ren->newpath)) {
3452 putname(ren->oldpath);
3453 return PTR_ERR(ren->newpath);
3456 req->flags |= REQ_F_NEED_CLEANUP;
3460 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3462 struct io_rename *ren = &req->rename;
3465 if (issue_flags & IO_URING_F_NONBLOCK)
3468 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3469 ren->newpath, ren->flags);
3471 req->flags &= ~REQ_F_NEED_CLEANUP;
3473 req_set_fail_links(req);
3474 io_req_complete(req, ret);
3478 static int io_unlinkat_prep(struct io_kiocb *req,
3479 const struct io_uring_sqe *sqe)
3481 struct io_unlink *un = &req->unlink;
3482 const char __user *fname;
3484 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3487 un->dfd = READ_ONCE(sqe->fd);
3489 un->flags = READ_ONCE(sqe->unlink_flags);
3490 if (un->flags & ~AT_REMOVEDIR)
3493 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3494 un->filename = getname(fname);
3495 if (IS_ERR(un->filename))
3496 return PTR_ERR(un->filename);
3498 req->flags |= REQ_F_NEED_CLEANUP;
3502 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3504 struct io_unlink *un = &req->unlink;
3507 if (issue_flags & IO_URING_F_NONBLOCK)
3510 if (un->flags & AT_REMOVEDIR)
3511 ret = do_rmdir(un->dfd, un->filename);
3513 ret = do_unlinkat(un->dfd, un->filename);
3515 req->flags &= ~REQ_F_NEED_CLEANUP;
3517 req_set_fail_links(req);
3518 io_req_complete(req, ret);
3522 static int io_shutdown_prep(struct io_kiocb *req,
3523 const struct io_uring_sqe *sqe)
3525 #if defined(CONFIG_NET)
3526 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3528 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3532 req->shutdown.how = READ_ONCE(sqe->len);
3539 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3541 #if defined(CONFIG_NET)
3542 struct socket *sock;
3545 if (issue_flags & IO_URING_F_NONBLOCK)
3548 sock = sock_from_file(req->file);
3549 if (unlikely(!sock))
3552 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3554 req_set_fail_links(req);
3555 io_req_complete(req, ret);
3562 static int __io_splice_prep(struct io_kiocb *req,
3563 const struct io_uring_sqe *sqe)
3565 struct io_splice* sp = &req->splice;
3566 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3568 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3572 sp->len = READ_ONCE(sqe->len);
3573 sp->flags = READ_ONCE(sqe->splice_flags);
3575 if (unlikely(sp->flags & ~valid_flags))
3578 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3579 (sp->flags & SPLICE_F_FD_IN_FIXED));
3582 req->flags |= REQ_F_NEED_CLEANUP;
3586 static int io_tee_prep(struct io_kiocb *req,
3587 const struct io_uring_sqe *sqe)
3589 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3591 return __io_splice_prep(req, sqe);
3594 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3596 struct io_splice *sp = &req->splice;
3597 struct file *in = sp->file_in;
3598 struct file *out = sp->file_out;
3599 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3602 if (issue_flags & IO_URING_F_NONBLOCK)
3605 ret = do_tee(in, out, sp->len, flags);
3607 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3609 req->flags &= ~REQ_F_NEED_CLEANUP;
3612 req_set_fail_links(req);
3613 io_req_complete(req, ret);
3617 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3619 struct io_splice* sp = &req->splice;
3621 sp->off_in = READ_ONCE(sqe->splice_off_in);
3622 sp->off_out = READ_ONCE(sqe->off);
3623 return __io_splice_prep(req, sqe);
3626 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3628 struct io_splice *sp = &req->splice;
3629 struct file *in = sp->file_in;
3630 struct file *out = sp->file_out;
3631 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3632 loff_t *poff_in, *poff_out;
3635 if (issue_flags & IO_URING_F_NONBLOCK)
3638 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3639 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3642 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3644 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3646 req->flags &= ~REQ_F_NEED_CLEANUP;
3649 req_set_fail_links(req);
3650 io_req_complete(req, ret);
3655 * IORING_OP_NOP just posts a completion event, nothing else.
3657 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3659 struct io_ring_ctx *ctx = req->ctx;
3661 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3664 __io_req_complete(req, issue_flags, 0, 0);
3668 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3670 struct io_ring_ctx *ctx = req->ctx;
3675 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3677 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3680 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3681 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3684 req->sync.off = READ_ONCE(sqe->off);
3685 req->sync.len = READ_ONCE(sqe->len);
3689 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3691 loff_t end = req->sync.off + req->sync.len;
3694 /* fsync always requires a blocking context */
3695 if (issue_flags & IO_URING_F_NONBLOCK)
3698 ret = vfs_fsync_range(req->file, req->sync.off,
3699 end > 0 ? end : LLONG_MAX,
3700 req->sync.flags & IORING_FSYNC_DATASYNC);
3702 req_set_fail_links(req);
3703 io_req_complete(req, ret);
3707 static int io_fallocate_prep(struct io_kiocb *req,
3708 const struct io_uring_sqe *sqe)
3710 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3712 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3715 req->sync.off = READ_ONCE(sqe->off);
3716 req->sync.len = READ_ONCE(sqe->addr);
3717 req->sync.mode = READ_ONCE(sqe->len);
3721 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3725 /* fallocate always requiring blocking context */
3726 if (issue_flags & IO_URING_F_NONBLOCK)
3728 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3731 req_set_fail_links(req);
3732 io_req_complete(req, ret);
3736 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3738 const char __user *fname;
3741 if (unlikely(sqe->ioprio || sqe->buf_index))
3743 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3746 /* open.how should be already initialised */
3747 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3748 req->open.how.flags |= O_LARGEFILE;
3750 req->open.dfd = READ_ONCE(sqe->fd);
3751 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3752 req->open.filename = getname(fname);
3753 if (IS_ERR(req->open.filename)) {
3754 ret = PTR_ERR(req->open.filename);
3755 req->open.filename = NULL;
3758 req->open.nofile = rlimit(RLIMIT_NOFILE);
3759 req->flags |= REQ_F_NEED_CLEANUP;
3763 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3767 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3769 mode = READ_ONCE(sqe->len);
3770 flags = READ_ONCE(sqe->open_flags);
3771 req->open.how = build_open_how(flags, mode);
3772 return __io_openat_prep(req, sqe);
3775 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3777 struct open_how __user *how;
3781 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3783 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3784 len = READ_ONCE(sqe->len);
3785 if (len < OPEN_HOW_SIZE_VER0)
3788 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3793 return __io_openat_prep(req, sqe);
3796 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3798 struct open_flags op;
3801 bool resolve_nonblock;
3804 ret = build_open_flags(&req->open.how, &op);
3807 nonblock_set = op.open_flag & O_NONBLOCK;
3808 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3809 if (issue_flags & IO_URING_F_NONBLOCK) {
3811 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3812 * it'll always -EAGAIN
3814 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3816 op.lookup_flags |= LOOKUP_CACHED;
3817 op.open_flag |= O_NONBLOCK;
3820 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3824 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3825 /* only retry if RESOLVE_CACHED wasn't already set by application */
3826 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3827 file == ERR_PTR(-EAGAIN)) {
3829 * We could hang on to this 'fd', but seems like marginal
3830 * gain for something that is now known to be a slower path.
3831 * So just put it, and we'll get a new one when we retry.
3839 ret = PTR_ERR(file);
3841 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3842 file->f_flags &= ~O_NONBLOCK;
3843 fsnotify_open(file);
3844 fd_install(ret, file);
3847 putname(req->open.filename);
3848 req->flags &= ~REQ_F_NEED_CLEANUP;
3850 req_set_fail_links(req);
3851 __io_req_complete(req, issue_flags, ret, 0);
3855 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3857 return io_openat2(req, issue_flags);
3860 static int io_remove_buffers_prep(struct io_kiocb *req,
3861 const struct io_uring_sqe *sqe)
3863 struct io_provide_buf *p = &req->pbuf;
3866 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3869 tmp = READ_ONCE(sqe->fd);
3870 if (!tmp || tmp > USHRT_MAX)
3873 memset(p, 0, sizeof(*p));
3875 p->bgid = READ_ONCE(sqe->buf_group);
3879 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3880 int bgid, unsigned nbufs)
3884 /* shouldn't happen */
3888 /* the head kbuf is the list itself */
3889 while (!list_empty(&buf->list)) {
3890 struct io_buffer *nxt;
3892 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3893 list_del(&nxt->list);
3900 xa_erase(&ctx->io_buffers, bgid);
3905 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3907 struct io_provide_buf *p = &req->pbuf;
3908 struct io_ring_ctx *ctx = req->ctx;
3909 struct io_buffer *head;
3911 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3913 io_ring_submit_lock(ctx, !force_nonblock);
3915 lockdep_assert_held(&ctx->uring_lock);
3918 head = xa_load(&ctx->io_buffers, p->bgid);
3920 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3922 req_set_fail_links(req);
3924 /* complete before unlock, IOPOLL may need the lock */
3925 __io_req_complete(req, issue_flags, ret, 0);
3926 io_ring_submit_unlock(ctx, !force_nonblock);
3930 static int io_provide_buffers_prep(struct io_kiocb *req,
3931 const struct io_uring_sqe *sqe)
3933 unsigned long size, tmp_check;
3934 struct io_provide_buf *p = &req->pbuf;
3937 if (sqe->ioprio || sqe->rw_flags)
3940 tmp = READ_ONCE(sqe->fd);
3941 if (!tmp || tmp > USHRT_MAX)
3944 p->addr = READ_ONCE(sqe->addr);
3945 p->len = READ_ONCE(sqe->len);
3947 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
3950 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
3953 size = (unsigned long)p->len * p->nbufs;
3954 if (!access_ok(u64_to_user_ptr(p->addr), size))
3957 p->bgid = READ_ONCE(sqe->buf_group);
3958 tmp = READ_ONCE(sqe->off);
3959 if (tmp > USHRT_MAX)
3965 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3967 struct io_buffer *buf;
3968 u64 addr = pbuf->addr;
3969 int i, bid = pbuf->bid;
3971 for (i = 0; i < pbuf->nbufs; i++) {
3972 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3977 buf->len = pbuf->len;
3982 INIT_LIST_HEAD(&buf->list);
3985 list_add_tail(&buf->list, &(*head)->list);
3989 return i ? i : -ENOMEM;
3992 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
3994 struct io_provide_buf *p = &req->pbuf;
3995 struct io_ring_ctx *ctx = req->ctx;
3996 struct io_buffer *head, *list;
3998 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4000 io_ring_submit_lock(ctx, !force_nonblock);
4002 lockdep_assert_held(&ctx->uring_lock);
4004 list = head = xa_load(&ctx->io_buffers, p->bgid);
4006 ret = io_add_buffers(p, &head);
4007 if (ret >= 0 && !list) {
4008 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4010 __io_remove_buffers(ctx, head, p->bgid, -1U);
4013 req_set_fail_links(req);
4014 /* complete before unlock, IOPOLL may need the lock */
4015 __io_req_complete(req, issue_flags, ret, 0);
4016 io_ring_submit_unlock(ctx, !force_nonblock);
4020 static int io_epoll_ctl_prep(struct io_kiocb *req,
4021 const struct io_uring_sqe *sqe)
4023 #if defined(CONFIG_EPOLL)
4024 if (sqe->ioprio || sqe->buf_index)
4026 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4029 req->epoll.epfd = READ_ONCE(sqe->fd);
4030 req->epoll.op = READ_ONCE(sqe->len);
4031 req->epoll.fd = READ_ONCE(sqe->off);
4033 if (ep_op_has_event(req->epoll.op)) {
4034 struct epoll_event __user *ev;
4036 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4037 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4047 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4049 #if defined(CONFIG_EPOLL)
4050 struct io_epoll *ie = &req->epoll;
4052 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4054 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4055 if (force_nonblock && ret == -EAGAIN)
4059 req_set_fail_links(req);
4060 __io_req_complete(req, issue_flags, ret, 0);
4067 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4069 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4070 if (sqe->ioprio || sqe->buf_index || sqe->off)
4072 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4075 req->madvise.addr = READ_ONCE(sqe->addr);
4076 req->madvise.len = READ_ONCE(sqe->len);
4077 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4084 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4086 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4087 struct io_madvise *ma = &req->madvise;
4090 if (issue_flags & IO_URING_F_NONBLOCK)
4093 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4095 req_set_fail_links(req);
4096 io_req_complete(req, ret);
4103 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4105 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4107 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4110 req->fadvise.offset = READ_ONCE(sqe->off);
4111 req->fadvise.len = READ_ONCE(sqe->len);
4112 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4116 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4118 struct io_fadvise *fa = &req->fadvise;
4121 if (issue_flags & IO_URING_F_NONBLOCK) {
4122 switch (fa->advice) {
4123 case POSIX_FADV_NORMAL:
4124 case POSIX_FADV_RANDOM:
4125 case POSIX_FADV_SEQUENTIAL:
4132 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4134 req_set_fail_links(req);
4135 __io_req_complete(req, issue_flags, ret, 0);
4139 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4141 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4143 if (sqe->ioprio || sqe->buf_index)
4145 if (req->flags & REQ_F_FIXED_FILE)
4148 req->statx.dfd = READ_ONCE(sqe->fd);
4149 req->statx.mask = READ_ONCE(sqe->len);
4150 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4151 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4152 req->statx.flags = READ_ONCE(sqe->statx_flags);
4157 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4159 struct io_statx *ctx = &req->statx;
4162 if (issue_flags & IO_URING_F_NONBLOCK)
4165 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4169 req_set_fail_links(req);
4170 io_req_complete(req, ret);
4174 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4176 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4178 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4179 sqe->rw_flags || sqe->buf_index)
4181 if (req->flags & REQ_F_FIXED_FILE)
4184 req->close.fd = READ_ONCE(sqe->fd);
4188 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4190 struct files_struct *files = current->files;
4191 struct io_close *close = &req->close;
4192 struct fdtable *fdt;
4193 struct file *file = NULL;
4196 spin_lock(&files->file_lock);
4197 fdt = files_fdtable(files);
4198 if (close->fd >= fdt->max_fds) {
4199 spin_unlock(&files->file_lock);
4202 file = fdt->fd[close->fd];
4203 if (!file || file->f_op == &io_uring_fops) {
4204 spin_unlock(&files->file_lock);
4209 /* if the file has a flush method, be safe and punt to async */
4210 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4211 spin_unlock(&files->file_lock);
4215 ret = __close_fd_get_file(close->fd, &file);
4216 spin_unlock(&files->file_lock);
4223 /* No ->flush() or already async, safely close from here */
4224 ret = filp_close(file, current->files);
4227 req_set_fail_links(req);
4230 __io_req_complete(req, issue_flags, ret, 0);
4234 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4236 struct io_ring_ctx *ctx = req->ctx;
4238 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4240 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4243 req->sync.off = READ_ONCE(sqe->off);
4244 req->sync.len = READ_ONCE(sqe->len);
4245 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4249 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4253 /* sync_file_range always requires a blocking context */
4254 if (issue_flags & IO_URING_F_NONBLOCK)
4257 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4260 req_set_fail_links(req);
4261 io_req_complete(req, ret);
4265 #if defined(CONFIG_NET)
4266 static int io_setup_async_msg(struct io_kiocb *req,
4267 struct io_async_msghdr *kmsg)
4269 struct io_async_msghdr *async_msg = req->async_data;
4273 if (io_alloc_async_data(req)) {
4274 kfree(kmsg->free_iov);
4277 async_msg = req->async_data;
4278 req->flags |= REQ_F_NEED_CLEANUP;
4279 memcpy(async_msg, kmsg, sizeof(*kmsg));
4280 async_msg->msg.msg_name = &async_msg->addr;
4281 /* if were using fast_iov, set it to the new one */
4282 if (!async_msg->free_iov)
4283 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4288 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4289 struct io_async_msghdr *iomsg)
4291 iomsg->msg.msg_name = &iomsg->addr;
4292 iomsg->free_iov = iomsg->fast_iov;
4293 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4294 req->sr_msg.msg_flags, &iomsg->free_iov);
4297 static int io_sendmsg_prep_async(struct io_kiocb *req)
4301 ret = io_sendmsg_copy_hdr(req, req->async_data);
4303 req->flags |= REQ_F_NEED_CLEANUP;
4307 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4309 struct io_sr_msg *sr = &req->sr_msg;
4311 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4314 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4315 sr->len = READ_ONCE(sqe->len);
4316 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4317 if (sr->msg_flags & MSG_DONTWAIT)
4318 req->flags |= REQ_F_NOWAIT;
4320 #ifdef CONFIG_COMPAT
4321 if (req->ctx->compat)
4322 sr->msg_flags |= MSG_CMSG_COMPAT;
4327 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4329 struct io_async_msghdr iomsg, *kmsg;
4330 struct socket *sock;
4335 sock = sock_from_file(req->file);
4336 if (unlikely(!sock))
4339 kmsg = req->async_data;
4341 ret = io_sendmsg_copy_hdr(req, &iomsg);
4347 flags = req->sr_msg.msg_flags;
4348 if (issue_flags & IO_URING_F_NONBLOCK)
4349 flags |= MSG_DONTWAIT;
4350 if (flags & MSG_WAITALL)
4351 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4353 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4354 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4355 return io_setup_async_msg(req, kmsg);
4356 if (ret == -ERESTARTSYS)
4359 /* fast path, check for non-NULL to avoid function call */
4361 kfree(kmsg->free_iov);
4362 req->flags &= ~REQ_F_NEED_CLEANUP;
4364 req_set_fail_links(req);
4365 __io_req_complete(req, issue_flags, ret, 0);
4369 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4371 struct io_sr_msg *sr = &req->sr_msg;
4374 struct socket *sock;
4379 sock = sock_from_file(req->file);
4380 if (unlikely(!sock))
4383 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4387 msg.msg_name = NULL;
4388 msg.msg_control = NULL;
4389 msg.msg_controllen = 0;
4390 msg.msg_namelen = 0;
4392 flags = req->sr_msg.msg_flags;
4393 if (issue_flags & IO_URING_F_NONBLOCK)
4394 flags |= MSG_DONTWAIT;
4395 if (flags & MSG_WAITALL)
4396 min_ret = iov_iter_count(&msg.msg_iter);
4398 msg.msg_flags = flags;
4399 ret = sock_sendmsg(sock, &msg);
4400 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4402 if (ret == -ERESTARTSYS)
4406 req_set_fail_links(req);
4407 __io_req_complete(req, issue_flags, ret, 0);
4411 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4412 struct io_async_msghdr *iomsg)
4414 struct io_sr_msg *sr = &req->sr_msg;
4415 struct iovec __user *uiov;
4419 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4420 &iomsg->uaddr, &uiov, &iov_len);
4424 if (req->flags & REQ_F_BUFFER_SELECT) {
4427 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4429 sr->len = iomsg->fast_iov[0].iov_len;
4430 iomsg->free_iov = NULL;
4432 iomsg->free_iov = iomsg->fast_iov;
4433 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4434 &iomsg->free_iov, &iomsg->msg.msg_iter,
4443 #ifdef CONFIG_COMPAT
4444 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4445 struct io_async_msghdr *iomsg)
4447 struct io_sr_msg *sr = &req->sr_msg;
4448 struct compat_iovec __user *uiov;
4453 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4458 uiov = compat_ptr(ptr);
4459 if (req->flags & REQ_F_BUFFER_SELECT) {
4460 compat_ssize_t clen;
4464 if (!access_ok(uiov, sizeof(*uiov)))
4466 if (__get_user(clen, &uiov->iov_len))
4471 iomsg->free_iov = NULL;
4473 iomsg->free_iov = iomsg->fast_iov;
4474 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4475 UIO_FASTIOV, &iomsg->free_iov,
4476 &iomsg->msg.msg_iter, true);
4485 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4486 struct io_async_msghdr *iomsg)
4488 iomsg->msg.msg_name = &iomsg->addr;
4490 #ifdef CONFIG_COMPAT
4491 if (req->ctx->compat)
4492 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4495 return __io_recvmsg_copy_hdr(req, iomsg);
4498 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4501 struct io_sr_msg *sr = &req->sr_msg;
4502 struct io_buffer *kbuf;
4504 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4509 req->flags |= REQ_F_BUFFER_SELECTED;
4513 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4515 return io_put_kbuf(req, req->sr_msg.kbuf);
4518 static int io_recvmsg_prep_async(struct io_kiocb *req)
4522 ret = io_recvmsg_copy_hdr(req, req->async_data);
4524 req->flags |= REQ_F_NEED_CLEANUP;
4528 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4530 struct io_sr_msg *sr = &req->sr_msg;
4532 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4535 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4536 sr->len = READ_ONCE(sqe->len);
4537 sr->bgid = READ_ONCE(sqe->buf_group);
4538 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4539 if (sr->msg_flags & MSG_DONTWAIT)
4540 req->flags |= REQ_F_NOWAIT;
4542 #ifdef CONFIG_COMPAT
4543 if (req->ctx->compat)
4544 sr->msg_flags |= MSG_CMSG_COMPAT;
4549 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4551 struct io_async_msghdr iomsg, *kmsg;
4552 struct socket *sock;
4553 struct io_buffer *kbuf;
4556 int ret, cflags = 0;
4557 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4559 sock = sock_from_file(req->file);
4560 if (unlikely(!sock))
4563 kmsg = req->async_data;
4565 ret = io_recvmsg_copy_hdr(req, &iomsg);
4571 if (req->flags & REQ_F_BUFFER_SELECT) {
4572 kbuf = io_recv_buffer_select(req, !force_nonblock);
4574 return PTR_ERR(kbuf);
4575 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4576 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4577 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4578 1, req->sr_msg.len);
4581 flags = req->sr_msg.msg_flags;
4583 flags |= MSG_DONTWAIT;
4584 if (flags & MSG_WAITALL)
4585 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4587 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4588 kmsg->uaddr, flags);
4589 if (force_nonblock && ret == -EAGAIN)
4590 return io_setup_async_msg(req, kmsg);
4591 if (ret == -ERESTARTSYS)
4594 if (req->flags & REQ_F_BUFFER_SELECTED)
4595 cflags = io_put_recv_kbuf(req);
4596 /* fast path, check for non-NULL to avoid function call */
4598 kfree(kmsg->free_iov);
4599 req->flags &= ~REQ_F_NEED_CLEANUP;
4600 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4601 req_set_fail_links(req);
4602 __io_req_complete(req, issue_flags, ret, cflags);
4606 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4608 struct io_buffer *kbuf;
4609 struct io_sr_msg *sr = &req->sr_msg;
4611 void __user *buf = sr->buf;
4612 struct socket *sock;
4616 int ret, cflags = 0;
4617 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4619 sock = sock_from_file(req->file);
4620 if (unlikely(!sock))
4623 if (req->flags & REQ_F_BUFFER_SELECT) {
4624 kbuf = io_recv_buffer_select(req, !force_nonblock);
4626 return PTR_ERR(kbuf);
4627 buf = u64_to_user_ptr(kbuf->addr);
4630 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4634 msg.msg_name = NULL;
4635 msg.msg_control = NULL;
4636 msg.msg_controllen = 0;
4637 msg.msg_namelen = 0;
4638 msg.msg_iocb = NULL;
4641 flags = req->sr_msg.msg_flags;
4643 flags |= MSG_DONTWAIT;
4644 if (flags & MSG_WAITALL)
4645 min_ret = iov_iter_count(&msg.msg_iter);
4647 ret = sock_recvmsg(sock, &msg, flags);
4648 if (force_nonblock && ret == -EAGAIN)
4650 if (ret == -ERESTARTSYS)
4653 if (req->flags & REQ_F_BUFFER_SELECTED)
4654 cflags = io_put_recv_kbuf(req);
4655 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4656 req_set_fail_links(req);
4657 __io_req_complete(req, issue_flags, ret, cflags);
4661 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4663 struct io_accept *accept = &req->accept;
4665 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4667 if (sqe->ioprio || sqe->len || sqe->buf_index)
4670 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4671 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4672 accept->flags = READ_ONCE(sqe->accept_flags);
4673 accept->nofile = rlimit(RLIMIT_NOFILE);
4677 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4679 struct io_accept *accept = &req->accept;
4680 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4681 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4684 if (req->file->f_flags & O_NONBLOCK)
4685 req->flags |= REQ_F_NOWAIT;
4687 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4688 accept->addr_len, accept->flags,
4690 if (ret == -EAGAIN && force_nonblock)
4693 if (ret == -ERESTARTSYS)
4695 req_set_fail_links(req);
4697 __io_req_complete(req, issue_flags, ret, 0);
4701 static int io_connect_prep_async(struct io_kiocb *req)
4703 struct io_async_connect *io = req->async_data;
4704 struct io_connect *conn = &req->connect;
4706 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4709 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4711 struct io_connect *conn = &req->connect;
4713 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4715 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4718 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4719 conn->addr_len = READ_ONCE(sqe->addr2);
4723 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4725 struct io_async_connect __io, *io;
4726 unsigned file_flags;
4728 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4730 if (req->async_data) {
4731 io = req->async_data;
4733 ret = move_addr_to_kernel(req->connect.addr,
4734 req->connect.addr_len,
4741 file_flags = force_nonblock ? O_NONBLOCK : 0;
4743 ret = __sys_connect_file(req->file, &io->address,
4744 req->connect.addr_len, file_flags);
4745 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4746 if (req->async_data)
4748 if (io_alloc_async_data(req)) {
4752 memcpy(req->async_data, &__io, sizeof(__io));
4755 if (ret == -ERESTARTSYS)
4759 req_set_fail_links(req);
4760 __io_req_complete(req, issue_flags, ret, 0);
4763 #else /* !CONFIG_NET */
4764 #define IO_NETOP_FN(op) \
4765 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4767 return -EOPNOTSUPP; \
4770 #define IO_NETOP_PREP(op) \
4772 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4774 return -EOPNOTSUPP; \
4777 #define IO_NETOP_PREP_ASYNC(op) \
4779 static int io_##op##_prep_async(struct io_kiocb *req) \
4781 return -EOPNOTSUPP; \
4784 IO_NETOP_PREP_ASYNC(sendmsg);
4785 IO_NETOP_PREP_ASYNC(recvmsg);
4786 IO_NETOP_PREP_ASYNC(connect);
4787 IO_NETOP_PREP(accept);
4790 #endif /* CONFIG_NET */
4792 struct io_poll_table {
4793 struct poll_table_struct pt;
4794 struct io_kiocb *req;
4798 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4799 __poll_t mask, task_work_func_t func)
4803 /* for instances that support it check for an event match first: */
4804 if (mask && !(mask & poll->events))
4807 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4809 list_del_init(&poll->wait.entry);
4812 req->task_work.func = func;
4815 * If this fails, then the task is exiting. When a task exits, the
4816 * work gets canceled, so just cancel this request as well instead
4817 * of executing it. We can't safely execute it anyway, as we may not
4818 * have the needed state needed for it anyway.
4820 ret = io_req_task_work_add(req);
4821 if (unlikely(ret)) {
4822 WRITE_ONCE(poll->canceled, true);
4823 io_req_task_work_add_fallback(req, func);
4828 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4829 __acquires(&req->ctx->completion_lock)
4831 struct io_ring_ctx *ctx = req->ctx;
4833 if (!req->result && !READ_ONCE(poll->canceled)) {
4834 struct poll_table_struct pt = { ._key = poll->events };
4836 req->result = vfs_poll(req->file, &pt) & poll->events;
4839 spin_lock_irq(&ctx->completion_lock);
4840 if (!req->result && !READ_ONCE(poll->canceled)) {
4841 add_wait_queue(poll->head, &poll->wait);
4848 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4850 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4851 if (req->opcode == IORING_OP_POLL_ADD)
4852 return req->async_data;
4853 return req->apoll->double_poll;
4856 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4858 if (req->opcode == IORING_OP_POLL_ADD)
4860 return &req->apoll->poll;
4863 static void io_poll_remove_double(struct io_kiocb *req)
4864 __must_hold(&req->ctx->completion_lock)
4866 struct io_poll_iocb *poll = io_poll_get_double(req);
4868 lockdep_assert_held(&req->ctx->completion_lock);
4870 if (poll && poll->head) {
4871 struct wait_queue_head *head = poll->head;
4873 spin_lock(&head->lock);
4874 list_del_init(&poll->wait.entry);
4875 if (poll->wait.private)
4878 spin_unlock(&head->lock);
4882 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4883 __must_hold(&req->ctx->completion_lock)
4885 struct io_ring_ctx *ctx = req->ctx;
4886 unsigned flags = IORING_CQE_F_MORE;
4889 if (READ_ONCE(req->poll.canceled)) {
4891 req->poll.events |= EPOLLONESHOT;
4893 error = mangle_poll(mask);
4895 if (req->poll.events & EPOLLONESHOT)
4897 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4898 io_poll_remove_waitqs(req);
4899 req->poll.done = true;
4902 io_commit_cqring(ctx);
4903 return !(flags & IORING_CQE_F_MORE);
4906 static void io_poll_task_func(struct callback_head *cb)
4908 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4909 struct io_ring_ctx *ctx = req->ctx;
4910 struct io_kiocb *nxt;
4912 if (io_poll_rewait(req, &req->poll)) {
4913 spin_unlock_irq(&ctx->completion_lock);
4917 done = io_poll_complete(req, req->result);
4919 hash_del(&req->hash_node);
4922 add_wait_queue(req->poll.head, &req->poll.wait);
4924 spin_unlock_irq(&ctx->completion_lock);
4925 io_cqring_ev_posted(ctx);
4928 nxt = io_put_req_find_next(req);
4930 __io_req_task_submit(nxt);
4935 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4936 int sync, void *key)
4938 struct io_kiocb *req = wait->private;
4939 struct io_poll_iocb *poll = io_poll_get_single(req);
4940 __poll_t mask = key_to_poll(key);
4942 /* for instances that support it check for an event match first: */
4943 if (mask && !(mask & poll->events))
4945 if (!(poll->events & EPOLLONESHOT))
4946 return poll->wait.func(&poll->wait, mode, sync, key);
4948 list_del_init(&wait->entry);
4950 if (poll && poll->head) {
4953 spin_lock(&poll->head->lock);
4954 done = list_empty(&poll->wait.entry);
4956 list_del_init(&poll->wait.entry);
4957 /* make sure double remove sees this as being gone */
4958 wait->private = NULL;
4959 spin_unlock(&poll->head->lock);
4961 /* use wait func handler, so it matches the rq type */
4962 poll->wait.func(&poll->wait, mode, sync, key);
4969 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4970 wait_queue_func_t wake_func)
4974 poll->canceled = false;
4975 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
4976 /* mask in events that we always want/need */
4977 poll->events = events | IO_POLL_UNMASK;
4978 INIT_LIST_HEAD(&poll->wait.entry);
4979 init_waitqueue_func_entry(&poll->wait, wake_func);
4982 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4983 struct wait_queue_head *head,
4984 struct io_poll_iocb **poll_ptr)
4986 struct io_kiocb *req = pt->req;
4989 * If poll->head is already set, it's because the file being polled
4990 * uses multiple waitqueues for poll handling (eg one for read, one
4991 * for write). Setup a separate io_poll_iocb if this happens.
4993 if (unlikely(poll->head)) {
4994 struct io_poll_iocb *poll_one = poll;
4996 /* already have a 2nd entry, fail a third attempt */
4998 pt->error = -EINVAL;
5002 * Can't handle multishot for double wait for now, turn it
5003 * into one-shot mode.
5005 if (!(req->poll.events & EPOLLONESHOT))
5006 req->poll.events |= EPOLLONESHOT;
5007 /* double add on the same waitqueue head, ignore */
5008 if (poll->head == head)
5010 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5012 pt->error = -ENOMEM;
5015 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5017 poll->wait.private = req;
5024 if (poll->events & EPOLLEXCLUSIVE)
5025 add_wait_queue_exclusive(head, &poll->wait);
5027 add_wait_queue(head, &poll->wait);
5030 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5031 struct poll_table_struct *p)
5033 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5034 struct async_poll *apoll = pt->req->apoll;
5036 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5039 static void io_async_task_func(struct callback_head *cb)
5041 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5042 struct async_poll *apoll = req->apoll;
5043 struct io_ring_ctx *ctx = req->ctx;
5045 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5047 if (io_poll_rewait(req, &apoll->poll)) {
5048 spin_unlock_irq(&ctx->completion_lock);
5052 hash_del(&req->hash_node);
5053 io_poll_remove_double(req);
5054 spin_unlock_irq(&ctx->completion_lock);
5056 if (!READ_ONCE(apoll->poll.canceled))
5057 __io_req_task_submit(req);
5059 io_req_complete_failed(req, -ECANCELED);
5062 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5065 struct io_kiocb *req = wait->private;
5066 struct io_poll_iocb *poll = &req->apoll->poll;
5068 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5071 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5074 static void io_poll_req_insert(struct io_kiocb *req)
5076 struct io_ring_ctx *ctx = req->ctx;
5077 struct hlist_head *list;
5079 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5080 hlist_add_head(&req->hash_node, list);
5083 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5084 struct io_poll_iocb *poll,
5085 struct io_poll_table *ipt, __poll_t mask,
5086 wait_queue_func_t wake_func)
5087 __acquires(&ctx->completion_lock)
5089 struct io_ring_ctx *ctx = req->ctx;
5090 bool cancel = false;
5092 INIT_HLIST_NODE(&req->hash_node);
5093 io_init_poll_iocb(poll, mask, wake_func);
5094 poll->file = req->file;
5095 poll->wait.private = req;
5097 ipt->pt._key = mask;
5099 ipt->error = -EINVAL;
5101 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5103 spin_lock_irq(&ctx->completion_lock);
5104 if (likely(poll->head)) {
5105 spin_lock(&poll->head->lock);
5106 if (unlikely(list_empty(&poll->wait.entry))) {
5112 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5113 list_del_init(&poll->wait.entry);
5115 WRITE_ONCE(poll->canceled, true);
5116 else if (!poll->done) /* actually waiting for an event */
5117 io_poll_req_insert(req);
5118 spin_unlock(&poll->head->lock);
5124 static bool io_arm_poll_handler(struct io_kiocb *req)
5126 const struct io_op_def *def = &io_op_defs[req->opcode];
5127 struct io_ring_ctx *ctx = req->ctx;
5128 struct async_poll *apoll;
5129 struct io_poll_table ipt;
5133 if (!req->file || !file_can_poll(req->file))
5135 if (req->flags & REQ_F_POLLED)
5139 else if (def->pollout)
5143 /* if we can't nonblock try, then no point in arming a poll handler */
5144 if (!io_file_supports_async(req, rw))
5147 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5148 if (unlikely(!apoll))
5150 apoll->double_poll = NULL;
5152 req->flags |= REQ_F_POLLED;
5155 mask = EPOLLONESHOT;
5157 mask |= POLLIN | POLLRDNORM;
5159 mask |= POLLOUT | POLLWRNORM;
5161 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5162 if ((req->opcode == IORING_OP_RECVMSG) &&
5163 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5166 mask |= POLLERR | POLLPRI;
5168 ipt.pt._qproc = io_async_queue_proc;
5170 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5172 if (ret || ipt.error) {
5173 io_poll_remove_double(req);
5174 spin_unlock_irq(&ctx->completion_lock);
5177 spin_unlock_irq(&ctx->completion_lock);
5178 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5179 apoll->poll.events);
5183 static bool __io_poll_remove_one(struct io_kiocb *req,
5184 struct io_poll_iocb *poll, bool do_cancel)
5185 __must_hold(&req->ctx->completion_lock)
5187 bool do_complete = false;
5191 spin_lock(&poll->head->lock);
5193 WRITE_ONCE(poll->canceled, true);
5194 if (!list_empty(&poll->wait.entry)) {
5195 list_del_init(&poll->wait.entry);
5198 spin_unlock(&poll->head->lock);
5199 hash_del(&req->hash_node);
5203 static bool io_poll_remove_waitqs(struct io_kiocb *req)
5204 __must_hold(&req->ctx->completion_lock)
5208 io_poll_remove_double(req);
5209 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5211 if (req->opcode != IORING_OP_POLL_ADD && do_complete) {
5212 /* non-poll requests have submit ref still */
5218 static bool io_poll_remove_one(struct io_kiocb *req)
5219 __must_hold(&req->ctx->completion_lock)
5223 do_complete = io_poll_remove_waitqs(req);
5225 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5226 io_commit_cqring(req->ctx);
5227 req_set_fail_links(req);
5228 io_put_req_deferred(req, 1);
5235 * Returns true if we found and killed one or more poll requests
5237 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5238 struct files_struct *files)
5240 struct hlist_node *tmp;
5241 struct io_kiocb *req;
5244 spin_lock_irq(&ctx->completion_lock);
5245 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5246 struct hlist_head *list;
5248 list = &ctx->cancel_hash[i];
5249 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5250 if (io_match_task(req, tsk, files))
5251 posted += io_poll_remove_one(req);
5254 spin_unlock_irq(&ctx->completion_lock);
5257 io_cqring_ev_posted(ctx);
5262 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5264 __must_hold(&ctx->completion_lock)
5266 struct hlist_head *list;
5267 struct io_kiocb *req;
5269 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5270 hlist_for_each_entry(req, list, hash_node) {
5271 if (sqe_addr != req->user_data)
5273 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5280 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5282 __must_hold(&ctx->completion_lock)
5284 struct io_kiocb *req;
5286 req = io_poll_find(ctx, sqe_addr, poll_only);
5289 if (io_poll_remove_one(req))
5295 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5300 events = READ_ONCE(sqe->poll32_events);
5302 events = swahw32(events);
5304 if (!(flags & IORING_POLL_ADD_MULTI))
5305 events |= EPOLLONESHOT;
5306 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5309 static int io_poll_update_prep(struct io_kiocb *req,
5310 const struct io_uring_sqe *sqe)
5312 struct io_poll_update *upd = &req->poll_update;
5315 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5317 if (sqe->ioprio || sqe->buf_index)
5319 flags = READ_ONCE(sqe->len);
5320 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5321 IORING_POLL_ADD_MULTI))
5323 /* meaningless without update */
5324 if (flags == IORING_POLL_ADD_MULTI)
5327 upd->old_user_data = READ_ONCE(sqe->addr);
5328 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5329 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5331 upd->new_user_data = READ_ONCE(sqe->off);
5332 if (!upd->update_user_data && upd->new_user_data)
5334 if (upd->update_events)
5335 upd->events = io_poll_parse_events(sqe, flags);
5336 else if (sqe->poll32_events)
5342 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5345 struct io_kiocb *req = wait->private;
5346 struct io_poll_iocb *poll = &req->poll;
5348 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5351 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5352 struct poll_table_struct *p)
5354 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5356 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5359 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5361 struct io_poll_iocb *poll = &req->poll;
5364 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5366 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5368 flags = READ_ONCE(sqe->len);
5369 if (flags & ~IORING_POLL_ADD_MULTI)
5372 poll->events = io_poll_parse_events(sqe, flags);
5376 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5378 struct io_poll_iocb *poll = &req->poll;
5379 struct io_ring_ctx *ctx = req->ctx;
5380 struct io_poll_table ipt;
5383 ipt.pt._qproc = io_poll_queue_proc;
5385 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5388 if (mask) { /* no async, we'd stolen it */
5390 io_poll_complete(req, mask);
5392 spin_unlock_irq(&ctx->completion_lock);
5395 io_cqring_ev_posted(ctx);
5396 if (poll->events & EPOLLONESHOT)
5402 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5404 struct io_ring_ctx *ctx = req->ctx;
5405 struct io_kiocb *preq;
5409 spin_lock_irq(&ctx->completion_lock);
5410 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5416 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5418 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5423 * Don't allow racy completion with singleshot, as we cannot safely
5424 * update those. For multishot, if we're racing with completion, just
5425 * let completion re-add it.
5427 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5428 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5432 /* we now have a detached poll request. reissue. */
5436 spin_unlock_irq(&ctx->completion_lock);
5437 req_set_fail_links(req);
5438 io_req_complete(req, ret);
5441 /* only mask one event flags, keep behavior flags */
5442 if (req->poll_update.update_events) {
5443 preq->poll.events &= ~0xffff;
5444 preq->poll.events |= req->poll_update.events & 0xffff;
5445 preq->poll.events |= IO_POLL_UNMASK;
5447 if (req->poll_update.update_user_data)
5448 preq->user_data = req->poll_update.new_user_data;
5449 spin_unlock_irq(&ctx->completion_lock);
5451 /* complete update request, we're done with it */
5452 io_req_complete(req, ret);
5455 ret = io_poll_add(preq, issue_flags);
5457 req_set_fail_links(preq);
5458 io_req_complete(preq, ret);
5464 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5466 struct io_timeout_data *data = container_of(timer,
5467 struct io_timeout_data, timer);
5468 struct io_kiocb *req = data->req;
5469 struct io_ring_ctx *ctx = req->ctx;
5470 unsigned long flags;
5472 spin_lock_irqsave(&ctx->completion_lock, flags);
5473 list_del_init(&req->timeout.list);
5474 atomic_set(&req->ctx->cq_timeouts,
5475 atomic_read(&req->ctx->cq_timeouts) + 1);
5477 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5478 io_commit_cqring(ctx);
5479 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5481 io_cqring_ev_posted(ctx);
5482 req_set_fail_links(req);
5484 return HRTIMER_NORESTART;
5487 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5489 __must_hold(&ctx->completion_lock)
5491 struct io_timeout_data *io;
5492 struct io_kiocb *req;
5495 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5496 found = user_data == req->user_data;
5501 return ERR_PTR(-ENOENT);
5503 io = req->async_data;
5504 if (hrtimer_try_to_cancel(&io->timer) == -1)
5505 return ERR_PTR(-EALREADY);
5506 list_del_init(&req->timeout.list);
5510 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5511 __must_hold(&ctx->completion_lock)
5513 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5516 return PTR_ERR(req);
5518 req_set_fail_links(req);
5519 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5520 io_put_req_deferred(req, 1);
5524 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5525 struct timespec64 *ts, enum hrtimer_mode mode)
5526 __must_hold(&ctx->completion_lock)
5528 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5529 struct io_timeout_data *data;
5532 return PTR_ERR(req);
5534 req->timeout.off = 0; /* noseq */
5535 data = req->async_data;
5536 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5537 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5538 data->timer.function = io_timeout_fn;
5539 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5543 static int io_timeout_remove_prep(struct io_kiocb *req,
5544 const struct io_uring_sqe *sqe)
5546 struct io_timeout_rem *tr = &req->timeout_rem;
5548 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5550 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5552 if (sqe->ioprio || sqe->buf_index || sqe->len)
5555 tr->addr = READ_ONCE(sqe->addr);
5556 tr->flags = READ_ONCE(sqe->timeout_flags);
5557 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5558 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5560 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5562 } else if (tr->flags) {
5563 /* timeout removal doesn't support flags */
5570 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5572 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5577 * Remove or update an existing timeout command
5579 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5581 struct io_timeout_rem *tr = &req->timeout_rem;
5582 struct io_ring_ctx *ctx = req->ctx;
5585 spin_lock_irq(&ctx->completion_lock);
5586 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5587 ret = io_timeout_cancel(ctx, tr->addr);
5589 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5590 io_translate_timeout_mode(tr->flags));
5592 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5593 io_commit_cqring(ctx);
5594 spin_unlock_irq(&ctx->completion_lock);
5595 io_cqring_ev_posted(ctx);
5597 req_set_fail_links(req);
5602 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5603 bool is_timeout_link)
5605 struct io_timeout_data *data;
5607 u32 off = READ_ONCE(sqe->off);
5609 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5611 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5613 if (off && is_timeout_link)
5615 flags = READ_ONCE(sqe->timeout_flags);
5616 if (flags & ~IORING_TIMEOUT_ABS)
5619 req->timeout.off = off;
5621 if (!req->async_data && io_alloc_async_data(req))
5624 data = req->async_data;
5627 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5630 data->mode = io_translate_timeout_mode(flags);
5631 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5632 if (is_timeout_link)
5633 io_req_track_inflight(req);
5637 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5639 struct io_ring_ctx *ctx = req->ctx;
5640 struct io_timeout_data *data = req->async_data;
5641 struct list_head *entry;
5642 u32 tail, off = req->timeout.off;
5644 spin_lock_irq(&ctx->completion_lock);
5647 * sqe->off holds how many events that need to occur for this
5648 * timeout event to be satisfied. If it isn't set, then this is
5649 * a pure timeout request, sequence isn't used.
5651 if (io_is_timeout_noseq(req)) {
5652 entry = ctx->timeout_list.prev;
5656 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5657 req->timeout.target_seq = tail + off;
5659 /* Update the last seq here in case io_flush_timeouts() hasn't.
5660 * This is safe because ->completion_lock is held, and submissions
5661 * and completions are never mixed in the same ->completion_lock section.
5663 ctx->cq_last_tm_flush = tail;
5666 * Insertion sort, ensuring the first entry in the list is always
5667 * the one we need first.
5669 list_for_each_prev(entry, &ctx->timeout_list) {
5670 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5673 if (io_is_timeout_noseq(nxt))
5675 /* nxt.seq is behind @tail, otherwise would've been completed */
5676 if (off >= nxt->timeout.target_seq - tail)
5680 list_add(&req->timeout.list, entry);
5681 data->timer.function = io_timeout_fn;
5682 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5683 spin_unlock_irq(&ctx->completion_lock);
5687 struct io_cancel_data {
5688 struct io_ring_ctx *ctx;
5692 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5694 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5695 struct io_cancel_data *cd = data;
5697 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5700 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5701 struct io_ring_ctx *ctx)
5703 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5704 enum io_wq_cancel cancel_ret;
5707 if (!tctx || !tctx->io_wq)
5710 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5711 switch (cancel_ret) {
5712 case IO_WQ_CANCEL_OK:
5715 case IO_WQ_CANCEL_RUNNING:
5718 case IO_WQ_CANCEL_NOTFOUND:
5726 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5727 struct io_kiocb *req, __u64 sqe_addr,
5730 unsigned long flags;
5733 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5734 spin_lock_irqsave(&ctx->completion_lock, flags);
5737 ret = io_timeout_cancel(ctx, sqe_addr);
5740 ret = io_poll_cancel(ctx, sqe_addr, false);
5744 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5745 io_commit_cqring(ctx);
5746 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5747 io_cqring_ev_posted(ctx);
5750 req_set_fail_links(req);
5753 static int io_async_cancel_prep(struct io_kiocb *req,
5754 const struct io_uring_sqe *sqe)
5756 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5758 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5760 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5763 req->cancel.addr = READ_ONCE(sqe->addr);
5767 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5769 struct io_ring_ctx *ctx = req->ctx;
5770 u64 sqe_addr = req->cancel.addr;
5771 struct io_tctx_node *node;
5774 /* tasks should wait for their io-wq threads, so safe w/o sync */
5775 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5776 spin_lock_irq(&ctx->completion_lock);
5779 ret = io_timeout_cancel(ctx, sqe_addr);
5782 ret = io_poll_cancel(ctx, sqe_addr, false);
5785 spin_unlock_irq(&ctx->completion_lock);
5787 /* slow path, try all io-wq's */
5788 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5790 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5791 struct io_uring_task *tctx = node->task->io_uring;
5793 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5797 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5799 spin_lock_irq(&ctx->completion_lock);
5801 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5802 io_commit_cqring(ctx);
5803 spin_unlock_irq(&ctx->completion_lock);
5804 io_cqring_ev_posted(ctx);
5807 req_set_fail_links(req);
5812 static int io_rsrc_update_prep(struct io_kiocb *req,
5813 const struct io_uring_sqe *sqe)
5815 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5817 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5819 if (sqe->ioprio || sqe->rw_flags)
5822 req->rsrc_update.offset = READ_ONCE(sqe->off);
5823 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5824 if (!req->rsrc_update.nr_args)
5826 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5830 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5832 struct io_ring_ctx *ctx = req->ctx;
5833 struct io_uring_rsrc_update2 up;
5836 if (issue_flags & IO_URING_F_NONBLOCK)
5839 up.offset = req->rsrc_update.offset;
5840 up.data = req->rsrc_update.arg;
5844 mutex_lock(&ctx->uring_lock);
5845 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5846 &up, req->rsrc_update.nr_args);
5847 mutex_unlock(&ctx->uring_lock);
5850 req_set_fail_links(req);
5851 __io_req_complete(req, issue_flags, ret, 0);
5855 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5857 switch (req->opcode) {
5860 case IORING_OP_READV:
5861 case IORING_OP_READ_FIXED:
5862 case IORING_OP_READ:
5863 return io_read_prep(req, sqe);
5864 case IORING_OP_WRITEV:
5865 case IORING_OP_WRITE_FIXED:
5866 case IORING_OP_WRITE:
5867 return io_write_prep(req, sqe);
5868 case IORING_OP_POLL_ADD:
5869 return io_poll_add_prep(req, sqe);
5870 case IORING_OP_POLL_REMOVE:
5871 return io_poll_update_prep(req, sqe);
5872 case IORING_OP_FSYNC:
5873 return io_fsync_prep(req, sqe);
5874 case IORING_OP_SYNC_FILE_RANGE:
5875 return io_sfr_prep(req, sqe);
5876 case IORING_OP_SENDMSG:
5877 case IORING_OP_SEND:
5878 return io_sendmsg_prep(req, sqe);
5879 case IORING_OP_RECVMSG:
5880 case IORING_OP_RECV:
5881 return io_recvmsg_prep(req, sqe);
5882 case IORING_OP_CONNECT:
5883 return io_connect_prep(req, sqe);
5884 case IORING_OP_TIMEOUT:
5885 return io_timeout_prep(req, sqe, false);
5886 case IORING_OP_TIMEOUT_REMOVE:
5887 return io_timeout_remove_prep(req, sqe);
5888 case IORING_OP_ASYNC_CANCEL:
5889 return io_async_cancel_prep(req, sqe);
5890 case IORING_OP_LINK_TIMEOUT:
5891 return io_timeout_prep(req, sqe, true);
5892 case IORING_OP_ACCEPT:
5893 return io_accept_prep(req, sqe);
5894 case IORING_OP_FALLOCATE:
5895 return io_fallocate_prep(req, sqe);
5896 case IORING_OP_OPENAT:
5897 return io_openat_prep(req, sqe);
5898 case IORING_OP_CLOSE:
5899 return io_close_prep(req, sqe);
5900 case IORING_OP_FILES_UPDATE:
5901 return io_rsrc_update_prep(req, sqe);
5902 case IORING_OP_STATX:
5903 return io_statx_prep(req, sqe);
5904 case IORING_OP_FADVISE:
5905 return io_fadvise_prep(req, sqe);
5906 case IORING_OP_MADVISE:
5907 return io_madvise_prep(req, sqe);
5908 case IORING_OP_OPENAT2:
5909 return io_openat2_prep(req, sqe);
5910 case IORING_OP_EPOLL_CTL:
5911 return io_epoll_ctl_prep(req, sqe);
5912 case IORING_OP_SPLICE:
5913 return io_splice_prep(req, sqe);
5914 case IORING_OP_PROVIDE_BUFFERS:
5915 return io_provide_buffers_prep(req, sqe);
5916 case IORING_OP_REMOVE_BUFFERS:
5917 return io_remove_buffers_prep(req, sqe);
5919 return io_tee_prep(req, sqe);
5920 case IORING_OP_SHUTDOWN:
5921 return io_shutdown_prep(req, sqe);
5922 case IORING_OP_RENAMEAT:
5923 return io_renameat_prep(req, sqe);
5924 case IORING_OP_UNLINKAT:
5925 return io_unlinkat_prep(req, sqe);
5928 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5933 static int io_req_prep_async(struct io_kiocb *req)
5935 if (!io_op_defs[req->opcode].needs_async_setup)
5937 if (WARN_ON_ONCE(req->async_data))
5939 if (io_alloc_async_data(req))
5942 switch (req->opcode) {
5943 case IORING_OP_READV:
5944 return io_rw_prep_async(req, READ);
5945 case IORING_OP_WRITEV:
5946 return io_rw_prep_async(req, WRITE);
5947 case IORING_OP_SENDMSG:
5948 return io_sendmsg_prep_async(req);
5949 case IORING_OP_RECVMSG:
5950 return io_recvmsg_prep_async(req);
5951 case IORING_OP_CONNECT:
5952 return io_connect_prep_async(req);
5954 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
5959 static u32 io_get_sequence(struct io_kiocb *req)
5961 struct io_kiocb *pos;
5962 struct io_ring_ctx *ctx = req->ctx;
5963 u32 total_submitted, nr_reqs = 0;
5965 io_for_each_link(pos, req)
5968 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5969 return total_submitted - nr_reqs;
5972 static int io_req_defer(struct io_kiocb *req)
5974 struct io_ring_ctx *ctx = req->ctx;
5975 struct io_defer_entry *de;
5979 /* Still need defer if there is pending req in defer list. */
5980 if (likely(list_empty_careful(&ctx->defer_list) &&
5981 !(req->flags & REQ_F_IO_DRAIN)))
5984 seq = io_get_sequence(req);
5985 /* Still a chance to pass the sequence check */
5986 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5989 ret = io_req_prep_async(req);
5992 io_prep_async_link(req);
5993 de = kmalloc(sizeof(*de), GFP_KERNEL);
5997 spin_lock_irq(&ctx->completion_lock);
5998 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5999 spin_unlock_irq(&ctx->completion_lock);
6001 io_queue_async_work(req);
6002 return -EIOCBQUEUED;
6005 trace_io_uring_defer(ctx, req, req->user_data);
6008 list_add_tail(&de->list, &ctx->defer_list);
6009 spin_unlock_irq(&ctx->completion_lock);
6010 return -EIOCBQUEUED;
6013 static void io_clean_op(struct io_kiocb *req)
6015 if (req->flags & REQ_F_BUFFER_SELECTED) {
6016 switch (req->opcode) {
6017 case IORING_OP_READV:
6018 case IORING_OP_READ_FIXED:
6019 case IORING_OP_READ:
6020 kfree((void *)(unsigned long)req->rw.addr);
6022 case IORING_OP_RECVMSG:
6023 case IORING_OP_RECV:
6024 kfree(req->sr_msg.kbuf);
6027 req->flags &= ~REQ_F_BUFFER_SELECTED;
6030 if (req->flags & REQ_F_NEED_CLEANUP) {
6031 switch (req->opcode) {
6032 case IORING_OP_READV:
6033 case IORING_OP_READ_FIXED:
6034 case IORING_OP_READ:
6035 case IORING_OP_WRITEV:
6036 case IORING_OP_WRITE_FIXED:
6037 case IORING_OP_WRITE: {
6038 struct io_async_rw *io = req->async_data;
6040 kfree(io->free_iovec);
6043 case IORING_OP_RECVMSG:
6044 case IORING_OP_SENDMSG: {
6045 struct io_async_msghdr *io = req->async_data;
6047 kfree(io->free_iov);
6050 case IORING_OP_SPLICE:
6052 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6053 io_put_file(req->splice.file_in);
6055 case IORING_OP_OPENAT:
6056 case IORING_OP_OPENAT2:
6057 if (req->open.filename)
6058 putname(req->open.filename);
6060 case IORING_OP_RENAMEAT:
6061 putname(req->rename.oldpath);
6062 putname(req->rename.newpath);
6064 case IORING_OP_UNLINKAT:
6065 putname(req->unlink.filename);
6068 req->flags &= ~REQ_F_NEED_CLEANUP;
6070 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6071 kfree(req->apoll->double_poll);
6075 if (req->flags & REQ_F_INFLIGHT) {
6076 struct io_uring_task *tctx = req->task->io_uring;
6078 atomic_dec(&tctx->inflight_tracked);
6079 req->flags &= ~REQ_F_INFLIGHT;
6083 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6085 struct io_ring_ctx *ctx = req->ctx;
6086 const struct cred *creds = NULL;
6089 if (req->work.creds && req->work.creds != current_cred())
6090 creds = override_creds(req->work.creds);
6092 switch (req->opcode) {
6094 ret = io_nop(req, issue_flags);
6096 case IORING_OP_READV:
6097 case IORING_OP_READ_FIXED:
6098 case IORING_OP_READ:
6099 ret = io_read(req, issue_flags);
6101 case IORING_OP_WRITEV:
6102 case IORING_OP_WRITE_FIXED:
6103 case IORING_OP_WRITE:
6104 ret = io_write(req, issue_flags);
6106 case IORING_OP_FSYNC:
6107 ret = io_fsync(req, issue_flags);
6109 case IORING_OP_POLL_ADD:
6110 ret = io_poll_add(req, issue_flags);
6112 case IORING_OP_POLL_REMOVE:
6113 ret = io_poll_update(req, issue_flags);
6115 case IORING_OP_SYNC_FILE_RANGE:
6116 ret = io_sync_file_range(req, issue_flags);
6118 case IORING_OP_SENDMSG:
6119 ret = io_sendmsg(req, issue_flags);
6121 case IORING_OP_SEND:
6122 ret = io_send(req, issue_flags);
6124 case IORING_OP_RECVMSG:
6125 ret = io_recvmsg(req, issue_flags);
6127 case IORING_OP_RECV:
6128 ret = io_recv(req, issue_flags);
6130 case IORING_OP_TIMEOUT:
6131 ret = io_timeout(req, issue_flags);
6133 case IORING_OP_TIMEOUT_REMOVE:
6134 ret = io_timeout_remove(req, issue_flags);
6136 case IORING_OP_ACCEPT:
6137 ret = io_accept(req, issue_flags);
6139 case IORING_OP_CONNECT:
6140 ret = io_connect(req, issue_flags);
6142 case IORING_OP_ASYNC_CANCEL:
6143 ret = io_async_cancel(req, issue_flags);
6145 case IORING_OP_FALLOCATE:
6146 ret = io_fallocate(req, issue_flags);
6148 case IORING_OP_OPENAT:
6149 ret = io_openat(req, issue_flags);
6151 case IORING_OP_CLOSE:
6152 ret = io_close(req, issue_flags);
6154 case IORING_OP_FILES_UPDATE:
6155 ret = io_files_update(req, issue_flags);
6157 case IORING_OP_STATX:
6158 ret = io_statx(req, issue_flags);
6160 case IORING_OP_FADVISE:
6161 ret = io_fadvise(req, issue_flags);
6163 case IORING_OP_MADVISE:
6164 ret = io_madvise(req, issue_flags);
6166 case IORING_OP_OPENAT2:
6167 ret = io_openat2(req, issue_flags);
6169 case IORING_OP_EPOLL_CTL:
6170 ret = io_epoll_ctl(req, issue_flags);
6172 case IORING_OP_SPLICE:
6173 ret = io_splice(req, issue_flags);
6175 case IORING_OP_PROVIDE_BUFFERS:
6176 ret = io_provide_buffers(req, issue_flags);
6178 case IORING_OP_REMOVE_BUFFERS:
6179 ret = io_remove_buffers(req, issue_flags);
6182 ret = io_tee(req, issue_flags);
6184 case IORING_OP_SHUTDOWN:
6185 ret = io_shutdown(req, issue_flags);
6187 case IORING_OP_RENAMEAT:
6188 ret = io_renameat(req, issue_flags);
6190 case IORING_OP_UNLINKAT:
6191 ret = io_unlinkat(req, issue_flags);
6199 revert_creds(creds);
6204 /* If the op doesn't have a file, we're not polling for it */
6205 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6206 const bool in_async = io_wq_current_is_worker();
6208 /* workqueue context doesn't hold uring_lock, grab it now */
6210 mutex_lock(&ctx->uring_lock);
6212 io_iopoll_req_issued(req, in_async);
6215 mutex_unlock(&ctx->uring_lock);
6221 static void io_wq_submit_work(struct io_wq_work *work)
6223 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6224 struct io_kiocb *timeout;
6227 timeout = io_prep_linked_timeout(req);
6229 io_queue_linked_timeout(timeout);
6231 if (work->flags & IO_WQ_WORK_CANCEL)
6236 ret = io_issue_sqe(req, 0);
6238 * We can get EAGAIN for polled IO even though we're
6239 * forcing a sync submission from here, since we can't
6240 * wait for request slots on the block side.
6248 /* avoid locking problems by failing it from a clean context */
6250 /* io-wq is going to take one down */
6252 io_req_task_queue_fail(req, ret);
6256 #define FFS_ASYNC_READ 0x1UL
6257 #define FFS_ASYNC_WRITE 0x2UL
6259 #define FFS_ISREG 0x4UL
6261 #define FFS_ISREG 0x0UL
6263 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6265 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6268 struct io_fixed_file *table_l2;
6270 table_l2 = table->files[i >> IORING_FILE_TABLE_SHIFT];
6271 return &table_l2[i & IORING_FILE_TABLE_MASK];
6274 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6277 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6279 return (struct file *) (slot->file_ptr & FFS_MASK);
6282 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6284 unsigned long file_ptr = (unsigned long) file;
6286 if (__io_file_supports_async(file, READ))
6287 file_ptr |= FFS_ASYNC_READ;
6288 if (__io_file_supports_async(file, WRITE))
6289 file_ptr |= FFS_ASYNC_WRITE;
6290 if (S_ISREG(file_inode(file)->i_mode))
6291 file_ptr |= FFS_ISREG;
6292 file_slot->file_ptr = file_ptr;
6295 static struct file *io_file_get(struct io_submit_state *state,
6296 struct io_kiocb *req, int fd, bool fixed)
6298 struct io_ring_ctx *ctx = req->ctx;
6302 unsigned long file_ptr;
6304 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6306 fd = array_index_nospec(fd, ctx->nr_user_files);
6307 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6308 file = (struct file *) (file_ptr & FFS_MASK);
6309 file_ptr &= ~FFS_MASK;
6310 /* mask in overlapping REQ_F and FFS bits */
6311 req->flags |= (file_ptr << REQ_F_ASYNC_READ_BIT);
6312 io_req_set_rsrc_node(req);
6314 trace_io_uring_file_get(ctx, fd);
6315 file = __io_file_get(state, fd);
6317 /* we don't allow fixed io_uring files */
6318 if (file && unlikely(file->f_op == &io_uring_fops))
6319 io_req_track_inflight(req);
6325 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6327 struct io_timeout_data *data = container_of(timer,
6328 struct io_timeout_data, timer);
6329 struct io_kiocb *prev, *req = data->req;
6330 struct io_ring_ctx *ctx = req->ctx;
6331 unsigned long flags;
6333 spin_lock_irqsave(&ctx->completion_lock, flags);
6334 prev = req->timeout.head;
6335 req->timeout.head = NULL;
6338 * We don't expect the list to be empty, that will only happen if we
6339 * race with the completion of the linked work.
6341 if (prev && req_ref_inc_not_zero(prev))
6342 io_remove_next_linked(prev);
6345 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6348 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6349 io_put_req_deferred(prev, 1);
6351 io_req_complete_post(req, -ETIME, 0);
6353 io_put_req_deferred(req, 1);
6354 return HRTIMER_NORESTART;
6357 static void io_queue_linked_timeout(struct io_kiocb *req)
6359 struct io_ring_ctx *ctx = req->ctx;
6361 spin_lock_irq(&ctx->completion_lock);
6363 * If the back reference is NULL, then our linked request finished
6364 * before we got a chance to setup the timer
6366 if (req->timeout.head) {
6367 struct io_timeout_data *data = req->async_data;
6369 data->timer.function = io_link_timeout_fn;
6370 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6373 spin_unlock_irq(&ctx->completion_lock);
6374 /* drop submission reference */
6378 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6380 struct io_kiocb *nxt = req->link;
6382 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6383 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6386 nxt->timeout.head = req;
6387 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6388 req->flags |= REQ_F_LINK_TIMEOUT;
6392 static void __io_queue_sqe(struct io_kiocb *req)
6394 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6397 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6400 * We async punt it if the file wasn't marked NOWAIT, or if the file
6401 * doesn't support non-blocking read/write attempts
6404 /* drop submission reference */
6405 if (req->flags & REQ_F_COMPLETE_INLINE) {
6406 struct io_ring_ctx *ctx = req->ctx;
6407 struct io_comp_state *cs = &ctx->submit_state.comp;
6409 cs->reqs[cs->nr++] = req;
6410 if (cs->nr == ARRAY_SIZE(cs->reqs))
6411 io_submit_flush_completions(cs, ctx);
6415 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6416 if (!io_arm_poll_handler(req)) {
6418 * Queued up for async execution, worker will release
6419 * submit reference when the iocb is actually submitted.
6421 io_queue_async_work(req);
6424 io_req_complete_failed(req, ret);
6427 io_queue_linked_timeout(linked_timeout);
6430 static void io_queue_sqe(struct io_kiocb *req)
6434 ret = io_req_defer(req);
6436 if (ret != -EIOCBQUEUED) {
6438 io_req_complete_failed(req, ret);
6440 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6441 ret = io_req_prep_async(req);
6444 io_queue_async_work(req);
6446 __io_queue_sqe(req);
6451 * Check SQE restrictions (opcode and flags).
6453 * Returns 'true' if SQE is allowed, 'false' otherwise.
6455 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6456 struct io_kiocb *req,
6457 unsigned int sqe_flags)
6459 if (!ctx->restricted)
6462 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6465 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6466 ctx->restrictions.sqe_flags_required)
6469 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6470 ctx->restrictions.sqe_flags_required))
6476 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6477 const struct io_uring_sqe *sqe)
6479 struct io_submit_state *state;
6480 unsigned int sqe_flags;
6481 int personality, ret = 0;
6483 req->opcode = READ_ONCE(sqe->opcode);
6484 /* same numerical values with corresponding REQ_F_*, safe to copy */
6485 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6486 req->user_data = READ_ONCE(sqe->user_data);
6487 req->async_data = NULL;
6491 req->fixed_rsrc_refs = NULL;
6492 /* one is dropped after submission, the other at completion */
6493 atomic_set(&req->refs, 2);
6494 req->task = current;
6496 req->work.creds = NULL;
6498 /* enforce forwards compatibility on users */
6499 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6504 if (unlikely(req->opcode >= IORING_OP_LAST))
6507 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6510 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6511 !io_op_defs[req->opcode].buffer_select)
6514 personality = READ_ONCE(sqe->personality);
6516 req->work.creds = xa_load(&ctx->personalities, personality);
6517 if (!req->work.creds)
6519 get_cred(req->work.creds);
6521 state = &ctx->submit_state;
6524 * Plug now if we have more than 1 IO left after this, and the target
6525 * is potentially a read/write to block based storage.
6527 if (!state->plug_started && state->ios_left > 1 &&
6528 io_op_defs[req->opcode].plug) {
6529 blk_start_plug(&state->plug);
6530 state->plug_started = true;
6533 if (io_op_defs[req->opcode].needs_file) {
6534 bool fixed = req->flags & REQ_F_FIXED_FILE;
6536 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6537 if (unlikely(!req->file))
6545 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6546 const struct io_uring_sqe *sqe)
6548 struct io_submit_link *link = &ctx->submit_state.link;
6551 ret = io_init_req(ctx, req, sqe);
6552 if (unlikely(ret)) {
6555 /* fail even hard links since we don't submit */
6556 link->head->flags |= REQ_F_FAIL_LINK;
6557 io_req_complete_failed(link->head, -ECANCELED);
6560 io_req_complete_failed(req, ret);
6563 ret = io_req_prep(req, sqe);
6567 /* don't need @sqe from now on */
6568 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6569 true, ctx->flags & IORING_SETUP_SQPOLL);
6572 * If we already have a head request, queue this one for async
6573 * submittal once the head completes. If we don't have a head but
6574 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6575 * submitted sync once the chain is complete. If none of those
6576 * conditions are true (normal request), then just queue it.
6579 struct io_kiocb *head = link->head;
6582 * Taking sequential execution of a link, draining both sides
6583 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6584 * requests in the link. So, it drains the head and the
6585 * next after the link request. The last one is done via
6586 * drain_next flag to persist the effect across calls.
6588 if (req->flags & REQ_F_IO_DRAIN) {
6589 head->flags |= REQ_F_IO_DRAIN;
6590 ctx->drain_next = 1;
6592 ret = io_req_prep_async(req);
6595 trace_io_uring_link(ctx, req, head);
6596 link->last->link = req;
6599 /* last request of a link, enqueue the link */
6600 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6605 if (unlikely(ctx->drain_next)) {
6606 req->flags |= REQ_F_IO_DRAIN;
6607 ctx->drain_next = 0;
6609 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6621 * Batched submission is done, ensure local IO is flushed out.
6623 static void io_submit_state_end(struct io_submit_state *state,
6624 struct io_ring_ctx *ctx)
6626 if (state->link.head)
6627 io_queue_sqe(state->link.head);
6629 io_submit_flush_completions(&state->comp, ctx);
6630 if (state->plug_started)
6631 blk_finish_plug(&state->plug);
6632 io_state_file_put(state);
6636 * Start submission side cache.
6638 static void io_submit_state_start(struct io_submit_state *state,
6639 unsigned int max_ios)
6641 state->plug_started = false;
6642 state->ios_left = max_ios;
6643 /* set only head, no need to init link_last in advance */
6644 state->link.head = NULL;
6647 static void io_commit_sqring(struct io_ring_ctx *ctx)
6649 struct io_rings *rings = ctx->rings;
6652 * Ensure any loads from the SQEs are done at this point,
6653 * since once we write the new head, the application could
6654 * write new data to them.
6656 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6660 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6661 * that is mapped by userspace. This means that care needs to be taken to
6662 * ensure that reads are stable, as we cannot rely on userspace always
6663 * being a good citizen. If members of the sqe are validated and then later
6664 * used, it's important that those reads are done through READ_ONCE() to
6665 * prevent a re-load down the line.
6667 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6669 u32 *sq_array = ctx->sq_array;
6673 * The cached sq head (or cq tail) serves two purposes:
6675 * 1) allows us to batch the cost of updating the user visible
6677 * 2) allows the kernel side to track the head on its own, even
6678 * though the application is the one updating it.
6680 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6681 if (likely(head < ctx->sq_entries))
6682 return &ctx->sq_sqes[head];
6684 /* drop invalid entries */
6685 ctx->cached_sq_dropped++;
6686 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6690 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6694 /* make sure SQ entry isn't read before tail */
6695 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6697 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6700 percpu_counter_add(¤t->io_uring->inflight, nr);
6701 refcount_add(nr, ¤t->usage);
6702 io_submit_state_start(&ctx->submit_state, nr);
6704 while (submitted < nr) {
6705 const struct io_uring_sqe *sqe;
6706 struct io_kiocb *req;
6708 req = io_alloc_req(ctx);
6709 if (unlikely(!req)) {
6711 submitted = -EAGAIN;
6714 sqe = io_get_sqe(ctx);
6715 if (unlikely(!sqe)) {
6716 kmem_cache_free(req_cachep, req);
6719 /* will complete beyond this point, count as submitted */
6721 if (io_submit_sqe(ctx, req, sqe))
6725 if (unlikely(submitted != nr)) {
6726 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6727 struct io_uring_task *tctx = current->io_uring;
6728 int unused = nr - ref_used;
6730 percpu_ref_put_many(&ctx->refs, unused);
6731 percpu_counter_sub(&tctx->inflight, unused);
6732 put_task_struct_many(current, unused);
6735 io_submit_state_end(&ctx->submit_state, ctx);
6736 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6737 io_commit_sqring(ctx);
6742 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6744 /* Tell userspace we may need a wakeup call */
6745 spin_lock_irq(&ctx->completion_lock);
6746 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6747 spin_unlock_irq(&ctx->completion_lock);
6750 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6752 spin_lock_irq(&ctx->completion_lock);
6753 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6754 spin_unlock_irq(&ctx->completion_lock);
6757 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6759 unsigned int to_submit;
6762 to_submit = io_sqring_entries(ctx);
6763 /* if we're handling multiple rings, cap submit size for fairness */
6764 if (cap_entries && to_submit > 8)
6767 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6768 unsigned nr_events = 0;
6770 mutex_lock(&ctx->uring_lock);
6771 if (!list_empty(&ctx->iopoll_list))
6772 io_do_iopoll(ctx, &nr_events, 0);
6775 * Don't submit if refs are dying, good for io_uring_register(),
6776 * but also it is relied upon by io_ring_exit_work()
6778 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6779 !(ctx->flags & IORING_SETUP_R_DISABLED))
6780 ret = io_submit_sqes(ctx, to_submit);
6781 mutex_unlock(&ctx->uring_lock);
6784 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6785 wake_up(&ctx->sqo_sq_wait);
6790 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6792 struct io_ring_ctx *ctx;
6793 unsigned sq_thread_idle = 0;
6795 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6796 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6797 sqd->sq_thread_idle = sq_thread_idle;
6800 static int io_sq_thread(void *data)
6802 struct io_sq_data *sqd = data;
6803 struct io_ring_ctx *ctx;
6804 unsigned long timeout = 0;
6805 char buf[TASK_COMM_LEN];
6808 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6809 set_task_comm(current, buf);
6810 current->pf_io_worker = NULL;
6812 if (sqd->sq_cpu != -1)
6813 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6815 set_cpus_allowed_ptr(current, cpu_online_mask);
6816 current->flags |= PF_NO_SETAFFINITY;
6818 mutex_lock(&sqd->lock);
6819 while (!test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)) {
6821 bool cap_entries, sqt_spin, needs_sched;
6823 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6824 signal_pending(current)) {
6825 bool did_sig = false;
6827 mutex_unlock(&sqd->lock);
6828 if (signal_pending(current)) {
6829 struct ksignal ksig;
6831 did_sig = get_signal(&ksig);
6834 mutex_lock(&sqd->lock);
6838 io_run_task_work_head(&sqd->park_task_work);
6839 timeout = jiffies + sqd->sq_thread_idle;
6843 cap_entries = !list_is_singular(&sqd->ctx_list);
6844 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6845 const struct cred *creds = NULL;
6847 if (ctx->sq_creds != current_cred())
6848 creds = override_creds(ctx->sq_creds);
6849 ret = __io_sq_thread(ctx, cap_entries);
6851 revert_creds(creds);
6852 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6856 if (sqt_spin || !time_after(jiffies, timeout)) {
6860 timeout = jiffies + sqd->sq_thread_idle;
6864 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6865 if (!test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6866 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6867 io_ring_set_wakeup_flag(ctx);
6870 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6871 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6872 !list_empty_careful(&ctx->iopoll_list)) {
6873 needs_sched = false;
6876 if (io_sqring_entries(ctx)) {
6877 needs_sched = false;
6883 mutex_unlock(&sqd->lock);
6885 mutex_lock(&sqd->lock);
6887 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6888 io_ring_clear_wakeup_flag(ctx);
6891 finish_wait(&sqd->wait, &wait);
6892 io_run_task_work_head(&sqd->park_task_work);
6893 timeout = jiffies + sqd->sq_thread_idle;
6896 io_uring_cancel_sqpoll(sqd);
6898 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6899 io_ring_set_wakeup_flag(ctx);
6901 io_run_task_work_head(&sqd->park_task_work);
6902 mutex_unlock(&sqd->lock);
6904 complete(&sqd->exited);
6908 struct io_wait_queue {
6909 struct wait_queue_entry wq;
6910 struct io_ring_ctx *ctx;
6912 unsigned nr_timeouts;
6915 static inline bool io_should_wake(struct io_wait_queue *iowq)
6917 struct io_ring_ctx *ctx = iowq->ctx;
6920 * Wake up if we have enough events, or if a timeout occurred since we
6921 * started waiting. For timeouts, we always want to return to userspace,
6922 * regardless of event count.
6924 return io_cqring_events(ctx) >= iowq->to_wait ||
6925 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6928 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6929 int wake_flags, void *key)
6931 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6935 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6936 * the task, and the next invocation will do it.
6938 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6939 return autoremove_wake_function(curr, mode, wake_flags, key);
6943 static int io_run_task_work_sig(void)
6945 if (io_run_task_work())
6947 if (!signal_pending(current))
6949 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6950 return -ERESTARTSYS;
6954 /* when returns >0, the caller should retry */
6955 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6956 struct io_wait_queue *iowq,
6957 signed long *timeout)
6961 /* make sure we run task_work before checking for signals */
6962 ret = io_run_task_work_sig();
6963 if (ret || io_should_wake(iowq))
6965 /* let the caller flush overflows, retry */
6966 if (test_bit(0, &ctx->cq_check_overflow))
6969 *timeout = schedule_timeout(*timeout);
6970 return !*timeout ? -ETIME : 1;
6974 * Wait until events become available, if we don't already have some. The
6975 * application must reap them itself, as they reside on the shared cq ring.
6977 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6978 const sigset_t __user *sig, size_t sigsz,
6979 struct __kernel_timespec __user *uts)
6981 struct io_wait_queue iowq = {
6984 .func = io_wake_function,
6985 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6988 .to_wait = min_events,
6990 struct io_rings *rings = ctx->rings;
6991 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6995 io_cqring_overflow_flush(ctx, false);
6996 if (io_cqring_events(ctx) >= min_events)
6998 if (!io_run_task_work())
7003 #ifdef CONFIG_COMPAT
7004 if (in_compat_syscall())
7005 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7009 ret = set_user_sigmask(sig, sigsz);
7016 struct timespec64 ts;
7018 if (get_timespec64(&ts, uts))
7020 timeout = timespec64_to_jiffies(&ts);
7023 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7024 trace_io_uring_cqring_wait(ctx, min_events);
7026 /* if we can't even flush overflow, don't wait for more */
7027 if (!io_cqring_overflow_flush(ctx, false)) {
7031 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7032 TASK_INTERRUPTIBLE);
7033 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7034 finish_wait(&ctx->wait, &iowq.wq);
7038 restore_saved_sigmask_unless(ret == -EINTR);
7040 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7043 static void io_free_file_tables(struct io_file_table *table, unsigned nr_files)
7045 unsigned i, nr_tables = DIV_ROUND_UP(nr_files, IORING_MAX_FILES_TABLE);
7047 for (i = 0; i < nr_tables; i++)
7048 kfree(table->files[i]);
7049 kfree(table->files);
7050 table->files = NULL;
7053 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7055 spin_lock_bh(&ctx->rsrc_ref_lock);
7058 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7060 spin_unlock_bh(&ctx->rsrc_ref_lock);
7063 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7065 percpu_ref_exit(&ref_node->refs);
7069 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7070 struct io_rsrc_data *data_to_kill)
7072 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7073 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7076 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7078 rsrc_node->rsrc_data = data_to_kill;
7079 io_rsrc_ref_lock(ctx);
7080 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7081 io_rsrc_ref_unlock(ctx);
7083 atomic_inc(&data_to_kill->refs);
7084 percpu_ref_kill(&rsrc_node->refs);
7085 ctx->rsrc_node = NULL;
7088 if (!ctx->rsrc_node) {
7089 ctx->rsrc_node = ctx->rsrc_backup_node;
7090 ctx->rsrc_backup_node = NULL;
7094 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7096 if (ctx->rsrc_backup_node)
7098 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7099 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7102 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7106 /* As we may drop ->uring_lock, other task may have started quiesce */
7110 data->quiesce = true;
7112 ret = io_rsrc_node_switch_start(ctx);
7115 io_rsrc_node_switch(ctx, data);
7117 /* kill initial ref, already quiesced if zero */
7118 if (atomic_dec_and_test(&data->refs))
7120 flush_delayed_work(&ctx->rsrc_put_work);
7121 ret = wait_for_completion_interruptible(&data->done);
7125 atomic_inc(&data->refs);
7126 /* wait for all works potentially completing data->done */
7127 flush_delayed_work(&ctx->rsrc_put_work);
7128 reinit_completion(&data->done);
7130 mutex_unlock(&ctx->uring_lock);
7131 ret = io_run_task_work_sig();
7132 mutex_lock(&ctx->uring_lock);
7134 data->quiesce = false;
7139 static void io_rsrc_data_free(struct io_rsrc_data *data)
7145 static struct io_rsrc_data *io_rsrc_data_alloc(struct io_ring_ctx *ctx,
7146 rsrc_put_fn *do_put,
7149 struct io_rsrc_data *data;
7151 data = kzalloc(sizeof(*data), GFP_KERNEL);
7155 data->tags = kvcalloc(nr, sizeof(*data->tags), GFP_KERNEL);
7161 atomic_set(&data->refs, 1);
7163 data->do_put = do_put;
7164 init_completion(&data->done);
7168 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7170 #if defined(CONFIG_UNIX)
7171 if (ctx->ring_sock) {
7172 struct sock *sock = ctx->ring_sock->sk;
7173 struct sk_buff *skb;
7175 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7181 for (i = 0; i < ctx->nr_user_files; i++) {
7184 file = io_file_from_index(ctx, i);
7189 io_free_file_tables(&ctx->file_table, ctx->nr_user_files);
7190 io_rsrc_data_free(ctx->file_data);
7191 ctx->file_data = NULL;
7192 ctx->nr_user_files = 0;
7195 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7199 if (!ctx->file_data)
7201 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7203 __io_sqe_files_unregister(ctx);
7207 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7208 __releases(&sqd->lock)
7210 WARN_ON_ONCE(sqd->thread == current);
7213 * Do the dance but not conditional clear_bit() because it'd race with
7214 * other threads incrementing park_pending and setting the bit.
7216 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7217 if (atomic_dec_return(&sqd->park_pending))
7218 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7219 mutex_unlock(&sqd->lock);
7222 static void io_sq_thread_park(struct io_sq_data *sqd)
7223 __acquires(&sqd->lock)
7225 WARN_ON_ONCE(sqd->thread == current);
7227 atomic_inc(&sqd->park_pending);
7228 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7229 mutex_lock(&sqd->lock);
7231 wake_up_process(sqd->thread);
7234 static void io_sq_thread_stop(struct io_sq_data *sqd)
7236 WARN_ON_ONCE(sqd->thread == current);
7237 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7239 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7240 mutex_lock(&sqd->lock);
7242 wake_up_process(sqd->thread);
7243 mutex_unlock(&sqd->lock);
7244 wait_for_completion(&sqd->exited);
7247 static void io_put_sq_data(struct io_sq_data *sqd)
7249 if (refcount_dec_and_test(&sqd->refs)) {
7250 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7252 io_sq_thread_stop(sqd);
7257 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7259 struct io_sq_data *sqd = ctx->sq_data;
7262 io_sq_thread_park(sqd);
7263 list_del_init(&ctx->sqd_list);
7264 io_sqd_update_thread_idle(sqd);
7265 io_sq_thread_unpark(sqd);
7267 io_put_sq_data(sqd);
7268 ctx->sq_data = NULL;
7272 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7274 struct io_ring_ctx *ctx_attach;
7275 struct io_sq_data *sqd;
7278 f = fdget(p->wq_fd);
7280 return ERR_PTR(-ENXIO);
7281 if (f.file->f_op != &io_uring_fops) {
7283 return ERR_PTR(-EINVAL);
7286 ctx_attach = f.file->private_data;
7287 sqd = ctx_attach->sq_data;
7290 return ERR_PTR(-EINVAL);
7292 if (sqd->task_tgid != current->tgid) {
7294 return ERR_PTR(-EPERM);
7297 refcount_inc(&sqd->refs);
7302 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7305 struct io_sq_data *sqd;
7308 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7309 sqd = io_attach_sq_data(p);
7314 /* fall through for EPERM case, setup new sqd/task */
7315 if (PTR_ERR(sqd) != -EPERM)
7319 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7321 return ERR_PTR(-ENOMEM);
7323 atomic_set(&sqd->park_pending, 0);
7324 refcount_set(&sqd->refs, 1);
7325 INIT_LIST_HEAD(&sqd->ctx_list);
7326 mutex_init(&sqd->lock);
7327 init_waitqueue_head(&sqd->wait);
7328 init_completion(&sqd->exited);
7332 #if defined(CONFIG_UNIX)
7334 * Ensure the UNIX gc is aware of our file set, so we are certain that
7335 * the io_uring can be safely unregistered on process exit, even if we have
7336 * loops in the file referencing.
7338 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7340 struct sock *sk = ctx->ring_sock->sk;
7341 struct scm_fp_list *fpl;
7342 struct sk_buff *skb;
7345 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7349 skb = alloc_skb(0, GFP_KERNEL);
7358 fpl->user = get_uid(current_user());
7359 for (i = 0; i < nr; i++) {
7360 struct file *file = io_file_from_index(ctx, i + offset);
7364 fpl->fp[nr_files] = get_file(file);
7365 unix_inflight(fpl->user, fpl->fp[nr_files]);
7370 fpl->max = SCM_MAX_FD;
7371 fpl->count = nr_files;
7372 UNIXCB(skb).fp = fpl;
7373 skb->destructor = unix_destruct_scm;
7374 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7375 skb_queue_head(&sk->sk_receive_queue, skb);
7377 for (i = 0; i < nr_files; i++)
7388 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7389 * causes regular reference counting to break down. We rely on the UNIX
7390 * garbage collection to take care of this problem for us.
7392 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7394 unsigned left, total;
7398 left = ctx->nr_user_files;
7400 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7402 ret = __io_sqe_files_scm(ctx, this_files, total);
7406 total += this_files;
7412 while (total < ctx->nr_user_files) {
7413 struct file *file = io_file_from_index(ctx, total);
7423 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7429 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7431 unsigned i, nr_tables = DIV_ROUND_UP(nr_files, IORING_MAX_FILES_TABLE);
7433 table->files = kcalloc(nr_tables, sizeof(*table->files), GFP_KERNEL);
7437 for (i = 0; i < nr_tables; i++) {
7438 unsigned int this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7440 table->files[i] = kcalloc(this_files, sizeof(*table->files[i]),
7442 if (!table->files[i])
7444 nr_files -= this_files;
7450 io_free_file_tables(table, nr_tables * IORING_MAX_FILES_TABLE);
7454 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7456 struct file *file = prsrc->file;
7457 #if defined(CONFIG_UNIX)
7458 struct sock *sock = ctx->ring_sock->sk;
7459 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7460 struct sk_buff *skb;
7463 __skb_queue_head_init(&list);
7466 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7467 * remove this entry and rearrange the file array.
7469 skb = skb_dequeue(head);
7471 struct scm_fp_list *fp;
7473 fp = UNIXCB(skb).fp;
7474 for (i = 0; i < fp->count; i++) {
7477 if (fp->fp[i] != file)
7480 unix_notinflight(fp->user, fp->fp[i]);
7481 left = fp->count - 1 - i;
7483 memmove(&fp->fp[i], &fp->fp[i + 1],
7484 left * sizeof(struct file *));
7491 __skb_queue_tail(&list, skb);
7501 __skb_queue_tail(&list, skb);
7503 skb = skb_dequeue(head);
7506 if (skb_peek(&list)) {
7507 spin_lock_irq(&head->lock);
7508 while ((skb = __skb_dequeue(&list)) != NULL)
7509 __skb_queue_tail(head, skb);
7510 spin_unlock_irq(&head->lock);
7517 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7519 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7520 struct io_ring_ctx *ctx = rsrc_data->ctx;
7521 struct io_rsrc_put *prsrc, *tmp;
7523 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7524 list_del(&prsrc->list);
7527 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7528 unsigned long flags;
7530 io_ring_submit_lock(ctx, lock_ring);
7531 spin_lock_irqsave(&ctx->completion_lock, flags);
7532 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7533 io_commit_cqring(ctx);
7534 spin_unlock_irqrestore(&ctx->completion_lock, flags);
7535 io_cqring_ev_posted(ctx);
7536 io_ring_submit_unlock(ctx, lock_ring);
7539 rsrc_data->do_put(ctx, prsrc);
7543 io_rsrc_node_destroy(ref_node);
7544 if (atomic_dec_and_test(&rsrc_data->refs))
7545 complete(&rsrc_data->done);
7548 static void io_rsrc_put_work(struct work_struct *work)
7550 struct io_ring_ctx *ctx;
7551 struct llist_node *node;
7553 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7554 node = llist_del_all(&ctx->rsrc_put_llist);
7557 struct io_rsrc_node *ref_node;
7558 struct llist_node *next = node->next;
7560 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7561 __io_rsrc_put_work(ref_node);
7566 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7568 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7569 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7570 bool first_add = false;
7572 io_rsrc_ref_lock(ctx);
7575 while (!list_empty(&ctx->rsrc_ref_list)) {
7576 node = list_first_entry(&ctx->rsrc_ref_list,
7577 struct io_rsrc_node, node);
7578 /* recycle ref nodes in order */
7581 list_del(&node->node);
7582 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7584 io_rsrc_ref_unlock(ctx);
7587 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7590 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7592 struct io_rsrc_node *ref_node;
7594 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7598 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7603 INIT_LIST_HEAD(&ref_node->node);
7604 INIT_LIST_HEAD(&ref_node->rsrc_list);
7605 ref_node->done = false;
7609 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7610 unsigned nr_args, u64 __user *tags)
7612 __s32 __user *fds = (__s32 __user *) arg;
7616 struct io_rsrc_data *file_data;
7622 if (nr_args > IORING_MAX_FIXED_FILES)
7624 ret = io_rsrc_node_switch_start(ctx);
7628 file_data = io_rsrc_data_alloc(ctx, io_rsrc_file_put, nr_args);
7631 ctx->file_data = file_data;
7633 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7636 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7639 if ((tags && copy_from_user(&tag, &tags[i], sizeof(tag))) ||
7640 copy_from_user(&fd, &fds[i], sizeof(fd))) {
7644 /* allow sparse sets */
7654 if (unlikely(!file))
7658 * Don't allow io_uring instances to be registered. If UNIX
7659 * isn't enabled, then this causes a reference cycle and this
7660 * instance can never get freed. If UNIX is enabled we'll
7661 * handle it just fine, but there's still no point in allowing
7662 * a ring fd as it doesn't support regular read/write anyway.
7664 if (file->f_op == &io_uring_fops) {
7668 ctx->file_data->tags[i] = tag;
7669 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7672 ret = io_sqe_files_scm(ctx);
7674 __io_sqe_files_unregister(ctx);
7678 io_rsrc_node_switch(ctx, NULL);
7681 for (i = 0; i < ctx->nr_user_files; i++) {
7682 file = io_file_from_index(ctx, i);
7686 io_free_file_tables(&ctx->file_table, nr_args);
7687 ctx->nr_user_files = 0;
7689 io_rsrc_data_free(ctx->file_data);
7690 ctx->file_data = NULL;
7694 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7697 #if defined(CONFIG_UNIX)
7698 struct sock *sock = ctx->ring_sock->sk;
7699 struct sk_buff_head *head = &sock->sk_receive_queue;
7700 struct sk_buff *skb;
7703 * See if we can merge this file into an existing skb SCM_RIGHTS
7704 * file set. If there's no room, fall back to allocating a new skb
7705 * and filling it in.
7707 spin_lock_irq(&head->lock);
7708 skb = skb_peek(head);
7710 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7712 if (fpl->count < SCM_MAX_FD) {
7713 __skb_unlink(skb, head);
7714 spin_unlock_irq(&head->lock);
7715 fpl->fp[fpl->count] = get_file(file);
7716 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7718 spin_lock_irq(&head->lock);
7719 __skb_queue_head(head, skb);
7724 spin_unlock_irq(&head->lock);
7731 return __io_sqe_files_scm(ctx, 1, index);
7737 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7738 struct io_rsrc_node *node, void *rsrc)
7740 struct io_rsrc_put *prsrc;
7742 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7746 prsrc->tag = data->tags[idx];
7748 list_add(&prsrc->list, &node->rsrc_list);
7752 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7753 struct io_uring_rsrc_update2 *up,
7756 u64 __user *tags = u64_to_user_ptr(up->tags);
7757 __s32 __user *fds = u64_to_user_ptr(up->data);
7758 struct io_rsrc_data *data = ctx->file_data;
7759 struct io_fixed_file *file_slot;
7763 bool needs_switch = false;
7765 if (!ctx->file_data)
7767 if (up->offset + nr_args > ctx->nr_user_files)
7770 for (done = 0; done < nr_args; done++) {
7773 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7774 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7778 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7782 if (fd == IORING_REGISTER_FILES_SKIP)
7785 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7786 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7788 if (file_slot->file_ptr) {
7789 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7790 err = io_queue_rsrc_removal(data, up->offset + done,
7791 ctx->rsrc_node, file);
7794 file_slot->file_ptr = 0;
7795 needs_switch = true;
7804 * Don't allow io_uring instances to be registered. If
7805 * UNIX isn't enabled, then this causes a reference
7806 * cycle and this instance can never get freed. If UNIX
7807 * is enabled we'll handle it just fine, but there's
7808 * still no point in allowing a ring fd as it doesn't
7809 * support regular read/write anyway.
7811 if (file->f_op == &io_uring_fops) {
7816 data->tags[up->offset + done] = tag;
7817 io_fixed_file_set(file_slot, file);
7818 err = io_sqe_file_register(ctx, file, i);
7820 file_slot->file_ptr = 0;
7828 io_rsrc_node_switch(ctx, data);
7829 return done ? done : err;
7832 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7834 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7836 req = io_put_req_find_next(req);
7837 return req ? &req->work : NULL;
7840 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7841 struct task_struct *task)
7843 struct io_wq_hash *hash;
7844 struct io_wq_data data;
7845 unsigned int concurrency;
7847 hash = ctx->hash_map;
7849 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7851 return ERR_PTR(-ENOMEM);
7852 refcount_set(&hash->refs, 1);
7853 init_waitqueue_head(&hash->wait);
7854 ctx->hash_map = hash;
7859 data.free_work = io_free_work;
7860 data.do_work = io_wq_submit_work;
7862 /* Do QD, or 4 * CPUS, whatever is smallest */
7863 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7865 return io_wq_create(concurrency, &data);
7868 static int io_uring_alloc_task_context(struct task_struct *task,
7869 struct io_ring_ctx *ctx)
7871 struct io_uring_task *tctx;
7874 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7875 if (unlikely(!tctx))
7878 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7879 if (unlikely(ret)) {
7884 tctx->io_wq = io_init_wq_offload(ctx, task);
7885 if (IS_ERR(tctx->io_wq)) {
7886 ret = PTR_ERR(tctx->io_wq);
7887 percpu_counter_destroy(&tctx->inflight);
7893 init_waitqueue_head(&tctx->wait);
7895 atomic_set(&tctx->in_idle, 0);
7896 atomic_set(&tctx->inflight_tracked, 0);
7897 task->io_uring = tctx;
7898 spin_lock_init(&tctx->task_lock);
7899 INIT_WQ_LIST(&tctx->task_list);
7900 tctx->task_state = 0;
7901 init_task_work(&tctx->task_work, tctx_task_work);
7905 void __io_uring_free(struct task_struct *tsk)
7907 struct io_uring_task *tctx = tsk->io_uring;
7909 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7910 WARN_ON_ONCE(tctx->io_wq);
7912 percpu_counter_destroy(&tctx->inflight);
7914 tsk->io_uring = NULL;
7917 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7918 struct io_uring_params *p)
7922 /* Retain compatibility with failing for an invalid attach attempt */
7923 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7924 IORING_SETUP_ATTACH_WQ) {
7927 f = fdget(p->wq_fd);
7931 if (f.file->f_op != &io_uring_fops)
7934 if (ctx->flags & IORING_SETUP_SQPOLL) {
7935 struct task_struct *tsk;
7936 struct io_sq_data *sqd;
7939 sqd = io_get_sq_data(p, &attached);
7945 ctx->sq_creds = get_current_cred();
7947 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7948 if (!ctx->sq_thread_idle)
7949 ctx->sq_thread_idle = HZ;
7951 io_sq_thread_park(sqd);
7952 list_add(&ctx->sqd_list, &sqd->ctx_list);
7953 io_sqd_update_thread_idle(sqd);
7954 /* don't attach to a dying SQPOLL thread, would be racy */
7955 ret = (attached && !sqd->thread) ? -ENXIO : 0;
7956 io_sq_thread_unpark(sqd);
7963 if (p->flags & IORING_SETUP_SQ_AFF) {
7964 int cpu = p->sq_thread_cpu;
7967 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
7974 sqd->task_pid = current->pid;
7975 sqd->task_tgid = current->tgid;
7976 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7983 ret = io_uring_alloc_task_context(tsk, ctx);
7984 wake_up_new_task(tsk);
7987 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7988 /* Can't have SQ_AFF without SQPOLL */
7995 complete(&ctx->sq_data->exited);
7997 io_sq_thread_finish(ctx);
8001 static inline void __io_unaccount_mem(struct user_struct *user,
8002 unsigned long nr_pages)
8004 atomic_long_sub(nr_pages, &user->locked_vm);
8007 static inline int __io_account_mem(struct user_struct *user,
8008 unsigned long nr_pages)
8010 unsigned long page_limit, cur_pages, new_pages;
8012 /* Don't allow more pages than we can safely lock */
8013 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8016 cur_pages = atomic_long_read(&user->locked_vm);
8017 new_pages = cur_pages + nr_pages;
8018 if (new_pages > page_limit)
8020 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8021 new_pages) != cur_pages);
8026 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8029 __io_unaccount_mem(ctx->user, nr_pages);
8031 if (ctx->mm_account)
8032 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8035 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8040 ret = __io_account_mem(ctx->user, nr_pages);
8045 if (ctx->mm_account)
8046 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8051 static void io_mem_free(void *ptr)
8058 page = virt_to_head_page(ptr);
8059 if (put_page_testzero(page))
8060 free_compound_page(page);
8063 static void *io_mem_alloc(size_t size)
8065 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8066 __GFP_NORETRY | __GFP_ACCOUNT;
8068 return (void *) __get_free_pages(gfp_flags, get_order(size));
8071 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8074 struct io_rings *rings;
8075 size_t off, sq_array_size;
8077 off = struct_size(rings, cqes, cq_entries);
8078 if (off == SIZE_MAX)
8082 off = ALIGN(off, SMP_CACHE_BYTES);
8090 sq_array_size = array_size(sizeof(u32), sq_entries);
8091 if (sq_array_size == SIZE_MAX)
8094 if (check_add_overflow(off, sq_array_size, &off))
8100 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8102 struct io_mapped_ubuf *imu = *slot;
8105 for (i = 0; i < imu->nr_bvecs; i++)
8106 unpin_user_page(imu->bvec[i].bv_page);
8107 if (imu->acct_pages)
8108 io_unaccount_mem(ctx, imu->acct_pages);
8113 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8117 if (!ctx->user_bufs)
8120 for (i = 0; i < ctx->nr_user_bufs; i++)
8121 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8122 kfree(ctx->user_bufs);
8123 ctx->user_bufs = NULL;
8124 ctx->nr_user_bufs = 0;
8128 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8129 void __user *arg, unsigned index)
8131 struct iovec __user *src;
8133 #ifdef CONFIG_COMPAT
8135 struct compat_iovec __user *ciovs;
8136 struct compat_iovec ciov;
8138 ciovs = (struct compat_iovec __user *) arg;
8139 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8142 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8143 dst->iov_len = ciov.iov_len;
8147 src = (struct iovec __user *) arg;
8148 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8154 * Not super efficient, but this is just a registration time. And we do cache
8155 * the last compound head, so generally we'll only do a full search if we don't
8158 * We check if the given compound head page has already been accounted, to
8159 * avoid double accounting it. This allows us to account the full size of the
8160 * page, not just the constituent pages of a huge page.
8162 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8163 int nr_pages, struct page *hpage)
8167 /* check current page array */
8168 for (i = 0; i < nr_pages; i++) {
8169 if (!PageCompound(pages[i]))
8171 if (compound_head(pages[i]) == hpage)
8175 /* check previously registered pages */
8176 for (i = 0; i < ctx->nr_user_bufs; i++) {
8177 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8179 for (j = 0; j < imu->nr_bvecs; j++) {
8180 if (!PageCompound(imu->bvec[j].bv_page))
8182 if (compound_head(imu->bvec[j].bv_page) == hpage)
8190 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8191 int nr_pages, struct io_mapped_ubuf *imu,
8192 struct page **last_hpage)
8196 for (i = 0; i < nr_pages; i++) {
8197 if (!PageCompound(pages[i])) {
8202 hpage = compound_head(pages[i]);
8203 if (hpage == *last_hpage)
8205 *last_hpage = hpage;
8206 if (headpage_already_acct(ctx, pages, i, hpage))
8208 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8212 if (!imu->acct_pages)
8215 ret = io_account_mem(ctx, imu->acct_pages);
8217 imu->acct_pages = 0;
8221 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8222 struct io_mapped_ubuf **pimu,
8223 struct page **last_hpage)
8225 struct io_mapped_ubuf *imu = NULL;
8226 struct vm_area_struct **vmas = NULL;
8227 struct page **pages = NULL;
8228 unsigned long off, start, end, ubuf;
8230 int ret, pret, nr_pages, i;
8232 ubuf = (unsigned long) iov->iov_base;
8233 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8234 start = ubuf >> PAGE_SHIFT;
8235 nr_pages = end - start;
8240 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8244 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8249 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8254 mmap_read_lock(current->mm);
8255 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8257 if (pret == nr_pages) {
8258 /* don't support file backed memory */
8259 for (i = 0; i < nr_pages; i++) {
8260 struct vm_area_struct *vma = vmas[i];
8263 !is_file_hugepages(vma->vm_file)) {
8269 ret = pret < 0 ? pret : -EFAULT;
8271 mmap_read_unlock(current->mm);
8274 * if we did partial map, or found file backed vmas,
8275 * release any pages we did get
8278 unpin_user_pages(pages, pret);
8282 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8284 unpin_user_pages(pages, pret);
8288 off = ubuf & ~PAGE_MASK;
8289 size = iov->iov_len;
8290 for (i = 0; i < nr_pages; i++) {
8293 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8294 imu->bvec[i].bv_page = pages[i];
8295 imu->bvec[i].bv_len = vec_len;
8296 imu->bvec[i].bv_offset = off;
8300 /* store original address for later verification */
8302 imu->ubuf_end = ubuf + iov->iov_len;
8303 imu->nr_bvecs = nr_pages;
8314 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8316 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8317 return ctx->user_bufs ? 0 : -ENOMEM;
8320 static int io_buffer_validate(struct iovec *iov)
8322 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8325 * Don't impose further limits on the size and buffer
8326 * constraints here, we'll -EINVAL later when IO is
8327 * submitted if they are wrong.
8329 if (!iov->iov_base || !iov->iov_len)
8332 /* arbitrary limit, but we need something */
8333 if (iov->iov_len > SZ_1G)
8336 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8342 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8343 unsigned int nr_args)
8347 struct page *last_hpage = NULL;
8351 if (!nr_args || nr_args > UIO_MAXIOV)
8353 ret = io_buffers_map_alloc(ctx, nr_args);
8357 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8358 ret = io_copy_iov(ctx, &iov, arg, i);
8361 ret = io_buffer_validate(&iov);
8365 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8372 io_sqe_buffers_unregister(ctx);
8377 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8379 __s32 __user *fds = arg;
8385 if (copy_from_user(&fd, fds, sizeof(*fds)))
8388 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8389 if (IS_ERR(ctx->cq_ev_fd)) {
8390 int ret = PTR_ERR(ctx->cq_ev_fd);
8391 ctx->cq_ev_fd = NULL;
8398 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8400 if (ctx->cq_ev_fd) {
8401 eventfd_ctx_put(ctx->cq_ev_fd);
8402 ctx->cq_ev_fd = NULL;
8409 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8411 struct io_buffer *buf;
8412 unsigned long index;
8414 xa_for_each(&ctx->io_buffers, index, buf)
8415 __io_remove_buffers(ctx, buf, index, -1U);
8418 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8420 struct io_kiocb *req, *nxt;
8422 list_for_each_entry_safe(req, nxt, list, compl.list) {
8423 if (tsk && req->task != tsk)
8425 list_del(&req->compl.list);
8426 kmem_cache_free(req_cachep, req);
8430 static void io_req_caches_free(struct io_ring_ctx *ctx)
8432 struct io_submit_state *submit_state = &ctx->submit_state;
8433 struct io_comp_state *cs = &ctx->submit_state.comp;
8435 mutex_lock(&ctx->uring_lock);
8437 if (submit_state->free_reqs) {
8438 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8439 submit_state->reqs);
8440 submit_state->free_reqs = 0;
8443 io_flush_cached_locked_reqs(ctx, cs);
8444 io_req_cache_free(&cs->free_list, NULL);
8445 mutex_unlock(&ctx->uring_lock);
8448 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8450 io_sq_thread_finish(ctx);
8451 io_sqe_buffers_unregister(ctx);
8453 if (ctx->mm_account) {
8454 mmdrop(ctx->mm_account);
8455 ctx->mm_account = NULL;
8458 mutex_lock(&ctx->uring_lock);
8459 if (ctx->file_data) {
8460 if (!atomic_dec_and_test(&ctx->file_data->refs))
8461 wait_for_completion(&ctx->file_data->done);
8462 __io_sqe_files_unregister(ctx);
8465 __io_cqring_overflow_flush(ctx, true);
8466 mutex_unlock(&ctx->uring_lock);
8467 io_eventfd_unregister(ctx);
8468 io_destroy_buffers(ctx);
8470 put_cred(ctx->sq_creds);
8472 /* there are no registered resources left, nobody uses it */
8474 io_rsrc_node_destroy(ctx->rsrc_node);
8475 if (ctx->rsrc_backup_node)
8476 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8477 flush_delayed_work(&ctx->rsrc_put_work);
8479 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8480 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8482 #if defined(CONFIG_UNIX)
8483 if (ctx->ring_sock) {
8484 ctx->ring_sock->file = NULL; /* so that iput() is called */
8485 sock_release(ctx->ring_sock);
8489 io_mem_free(ctx->rings);
8490 io_mem_free(ctx->sq_sqes);
8492 percpu_ref_exit(&ctx->refs);
8493 free_uid(ctx->user);
8494 io_req_caches_free(ctx);
8496 io_wq_put_hash(ctx->hash_map);
8497 kfree(ctx->cancel_hash);
8501 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8503 struct io_ring_ctx *ctx = file->private_data;
8506 poll_wait(file, &ctx->cq_wait, wait);
8508 * synchronizes with barrier from wq_has_sleeper call in
8512 if (!io_sqring_full(ctx))
8513 mask |= EPOLLOUT | EPOLLWRNORM;
8516 * Don't flush cqring overflow list here, just do a simple check.
8517 * Otherwise there could possible be ABBA deadlock:
8520 * lock(&ctx->uring_lock);
8522 * lock(&ctx->uring_lock);
8525 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8526 * pushs them to do the flush.
8528 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8529 mask |= EPOLLIN | EPOLLRDNORM;
8534 static int io_uring_fasync(int fd, struct file *file, int on)
8536 struct io_ring_ctx *ctx = file->private_data;
8538 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8541 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8543 const struct cred *creds;
8545 creds = xa_erase(&ctx->personalities, id);
8554 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8556 return io_run_task_work_head(&ctx->exit_task_work);
8559 struct io_tctx_exit {
8560 struct callback_head task_work;
8561 struct completion completion;
8562 struct io_ring_ctx *ctx;
8565 static void io_tctx_exit_cb(struct callback_head *cb)
8567 struct io_uring_task *tctx = current->io_uring;
8568 struct io_tctx_exit *work;
8570 work = container_of(cb, struct io_tctx_exit, task_work);
8572 * When @in_idle, we're in cancellation and it's racy to remove the
8573 * node. It'll be removed by the end of cancellation, just ignore it.
8575 if (!atomic_read(&tctx->in_idle))
8576 io_uring_del_task_file((unsigned long)work->ctx);
8577 complete(&work->completion);
8580 static void io_ring_exit_work(struct work_struct *work)
8582 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8583 unsigned long timeout = jiffies + HZ * 60 * 5;
8584 struct io_tctx_exit exit;
8585 struct io_tctx_node *node;
8589 * If we're doing polled IO and end up having requests being
8590 * submitted async (out-of-line), then completions can come in while
8591 * we're waiting for refs to drop. We need to reap these manually,
8592 * as nobody else will be looking for them.
8595 io_uring_try_cancel_requests(ctx, NULL, NULL);
8597 WARN_ON_ONCE(time_after(jiffies, timeout));
8598 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8600 init_completion(&exit.completion);
8601 init_task_work(&exit.task_work, io_tctx_exit_cb);
8604 * Some may use context even when all refs and requests have been put,
8605 * and they are free to do so while still holding uring_lock or
8606 * completion_lock, see __io_req_task_submit(). Apart from other work,
8607 * this lock/unlock section also waits them to finish.
8609 mutex_lock(&ctx->uring_lock);
8610 while (!list_empty(&ctx->tctx_list)) {
8611 WARN_ON_ONCE(time_after(jiffies, timeout));
8613 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8615 /* don't spin on a single task if cancellation failed */
8616 list_rotate_left(&ctx->tctx_list);
8617 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8618 if (WARN_ON_ONCE(ret))
8620 wake_up_process(node->task);
8622 mutex_unlock(&ctx->uring_lock);
8623 wait_for_completion(&exit.completion);
8624 mutex_lock(&ctx->uring_lock);
8626 mutex_unlock(&ctx->uring_lock);
8627 spin_lock_irq(&ctx->completion_lock);
8628 spin_unlock_irq(&ctx->completion_lock);
8630 io_ring_ctx_free(ctx);
8633 /* Returns true if we found and killed one or more timeouts */
8634 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8635 struct files_struct *files)
8637 struct io_kiocb *req, *tmp;
8640 spin_lock_irq(&ctx->completion_lock);
8641 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8642 if (io_match_task(req, tsk, files)) {
8643 io_kill_timeout(req, -ECANCELED);
8648 io_commit_cqring(ctx);
8649 spin_unlock_irq(&ctx->completion_lock);
8651 io_cqring_ev_posted(ctx);
8652 return canceled != 0;
8655 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8657 unsigned long index;
8658 struct creds *creds;
8660 mutex_lock(&ctx->uring_lock);
8661 percpu_ref_kill(&ctx->refs);
8663 __io_cqring_overflow_flush(ctx, true);
8664 xa_for_each(&ctx->personalities, index, creds)
8665 io_unregister_personality(ctx, index);
8666 mutex_unlock(&ctx->uring_lock);
8668 io_kill_timeouts(ctx, NULL, NULL);
8669 io_poll_remove_all(ctx, NULL, NULL);
8671 /* if we failed setting up the ctx, we might not have any rings */
8672 io_iopoll_try_reap_events(ctx);
8674 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8676 * Use system_unbound_wq to avoid spawning tons of event kworkers
8677 * if we're exiting a ton of rings at the same time. It just adds
8678 * noise and overhead, there's no discernable change in runtime
8679 * over using system_wq.
8681 queue_work(system_unbound_wq, &ctx->exit_work);
8684 static int io_uring_release(struct inode *inode, struct file *file)
8686 struct io_ring_ctx *ctx = file->private_data;
8688 file->private_data = NULL;
8689 io_ring_ctx_wait_and_kill(ctx);
8693 struct io_task_cancel {
8694 struct task_struct *task;
8695 struct files_struct *files;
8698 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8700 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8701 struct io_task_cancel *cancel = data;
8704 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8705 unsigned long flags;
8706 struct io_ring_ctx *ctx = req->ctx;
8708 /* protect against races with linked timeouts */
8709 spin_lock_irqsave(&ctx->completion_lock, flags);
8710 ret = io_match_task(req, cancel->task, cancel->files);
8711 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8713 ret = io_match_task(req, cancel->task, cancel->files);
8718 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8719 struct task_struct *task,
8720 struct files_struct *files)
8722 struct io_defer_entry *de;
8725 spin_lock_irq(&ctx->completion_lock);
8726 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8727 if (io_match_task(de->req, task, files)) {
8728 list_cut_position(&list, &ctx->defer_list, &de->list);
8732 spin_unlock_irq(&ctx->completion_lock);
8733 if (list_empty(&list))
8736 while (!list_empty(&list)) {
8737 de = list_first_entry(&list, struct io_defer_entry, list);
8738 list_del_init(&de->list);
8739 io_req_complete_failed(de->req, -ECANCELED);
8745 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8747 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8749 return req->ctx == data;
8752 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8754 struct io_tctx_node *node;
8755 enum io_wq_cancel cret;
8758 mutex_lock(&ctx->uring_lock);
8759 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8760 struct io_uring_task *tctx = node->task->io_uring;
8763 * io_wq will stay alive while we hold uring_lock, because it's
8764 * killed after ctx nodes, which requires to take the lock.
8766 if (!tctx || !tctx->io_wq)
8768 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8769 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8771 mutex_unlock(&ctx->uring_lock);
8776 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8777 struct task_struct *task,
8778 struct files_struct *files)
8780 struct io_task_cancel cancel = { .task = task, .files = files, };
8781 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8784 enum io_wq_cancel cret;
8788 ret |= io_uring_try_cancel_iowq(ctx);
8789 } else if (tctx && tctx->io_wq) {
8791 * Cancels requests of all rings, not only @ctx, but
8792 * it's fine as the task is in exit/exec.
8794 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8796 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8799 /* SQPOLL thread does its own polling */
8800 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && !files) ||
8801 (ctx->sq_data && ctx->sq_data->thread == current)) {
8802 while (!list_empty_careful(&ctx->iopoll_list)) {
8803 io_iopoll_try_reap_events(ctx);
8808 ret |= io_cancel_defer_files(ctx, task, files);
8809 ret |= io_poll_remove_all(ctx, task, files);
8810 ret |= io_kill_timeouts(ctx, task, files);
8811 ret |= io_run_task_work();
8812 ret |= io_run_ctx_fallback(ctx);
8819 static int __io_uring_add_task_file(struct io_ring_ctx *ctx)
8821 struct io_uring_task *tctx = current->io_uring;
8822 struct io_tctx_node *node;
8825 if (unlikely(!tctx)) {
8826 ret = io_uring_alloc_task_context(current, ctx);
8829 tctx = current->io_uring;
8831 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
8832 node = kmalloc(sizeof(*node), GFP_KERNEL);
8836 node->task = current;
8838 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
8845 mutex_lock(&ctx->uring_lock);
8846 list_add(&node->ctx_node, &ctx->tctx_list);
8847 mutex_unlock(&ctx->uring_lock);
8854 * Note that this task has used io_uring. We use it for cancelation purposes.
8856 static inline int io_uring_add_task_file(struct io_ring_ctx *ctx)
8858 struct io_uring_task *tctx = current->io_uring;
8860 if (likely(tctx && tctx->last == ctx))
8862 return __io_uring_add_task_file(ctx);
8866 * Remove this io_uring_file -> task mapping.
8868 static void io_uring_del_task_file(unsigned long index)
8870 struct io_uring_task *tctx = current->io_uring;
8871 struct io_tctx_node *node;
8875 node = xa_erase(&tctx->xa, index);
8879 WARN_ON_ONCE(current != node->task);
8880 WARN_ON_ONCE(list_empty(&node->ctx_node));
8882 mutex_lock(&node->ctx->uring_lock);
8883 list_del(&node->ctx_node);
8884 mutex_unlock(&node->ctx->uring_lock);
8886 if (tctx->last == node->ctx)
8891 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8893 struct io_tctx_node *node;
8894 unsigned long index;
8896 xa_for_each(&tctx->xa, index, node)
8897 io_uring_del_task_file(index);
8899 io_wq_put_and_exit(tctx->io_wq);
8904 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
8907 return atomic_read(&tctx->inflight_tracked);
8908 return percpu_counter_sum(&tctx->inflight);
8911 static void io_sqpoll_cancel_cb(struct callback_head *cb)
8913 struct io_tctx_exit *work = container_of(cb, struct io_tctx_exit, task_work);
8914 struct io_sq_data *sqd = work->ctx->sq_data;
8917 io_uring_cancel_sqpoll(sqd);
8918 list_del_init(&work->ctx->sqd_list);
8919 complete(&work->completion);
8922 static void io_sqpoll_cancel_sync(struct io_ring_ctx *ctx)
8924 struct io_sq_data *sqd = ctx->sq_data;
8925 struct io_tctx_exit work = { .ctx = ctx, };
8926 struct task_struct *task;
8928 io_sq_thread_park(sqd);
8929 io_sqd_update_thread_idle(sqd);
8932 init_completion(&work.completion);
8933 init_task_work(&work.task_work, io_sqpoll_cancel_cb);
8934 io_task_work_add_head(&sqd->park_task_work, &work.task_work);
8935 wake_up_process(task);
8937 list_del_init(&ctx->sqd_list);
8939 io_sq_thread_unpark(sqd);
8942 wait_for_completion(&work.completion);
8945 static void io_uring_try_cancel(struct files_struct *files)
8947 struct io_uring_task *tctx = current->io_uring;
8948 struct io_tctx_node *node;
8949 unsigned long index;
8951 xa_for_each(&tctx->xa, index, node) {
8952 struct io_ring_ctx *ctx = node->ctx;
8955 io_sqpoll_cancel_sync(ctx);
8958 io_uring_try_cancel_requests(ctx, current, files);
8962 /* should only be called by SQPOLL task */
8963 static void io_uring_cancel_sqpoll(struct io_sq_data *sqd)
8965 struct io_uring_task *tctx = current->io_uring;
8966 struct io_ring_ctx *ctx;
8970 WARN_ON_ONCE(!sqd || sqd->thread != current);
8972 atomic_inc(&tctx->in_idle);
8974 /* read completions before cancelations */
8975 inflight = tctx_inflight(tctx, false);
8978 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
8979 io_uring_try_cancel_requests(ctx, current, NULL);
8981 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8983 * If we've seen completions, retry without waiting. This
8984 * avoids a race where a completion comes in before we did
8985 * prepare_to_wait().
8987 if (inflight == tctx_inflight(tctx, false))
8989 finish_wait(&tctx->wait, &wait);
8991 atomic_dec(&tctx->in_idle);
8995 * Find any io_uring fd that this task has registered or done IO on, and cancel
8998 void __io_uring_cancel(struct files_struct *files)
9000 struct io_uring_task *tctx = current->io_uring;
9004 /* make sure overflow events are dropped */
9005 atomic_inc(&tctx->in_idle);
9006 io_uring_try_cancel(files);
9009 /* read completions before cancelations */
9010 inflight = tctx_inflight(tctx, !!files);
9013 io_uring_try_cancel(files);
9014 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9017 * If we've seen completions, retry without waiting. This
9018 * avoids a race where a completion comes in before we did
9019 * prepare_to_wait().
9021 if (inflight == tctx_inflight(tctx, !!files))
9023 finish_wait(&tctx->wait, &wait);
9025 atomic_dec(&tctx->in_idle);
9027 io_uring_clean_tctx(tctx);
9029 /* for exec all current's requests should be gone, kill tctx */
9030 __io_uring_free(current);
9034 static void *io_uring_validate_mmap_request(struct file *file,
9035 loff_t pgoff, size_t sz)
9037 struct io_ring_ctx *ctx = file->private_data;
9038 loff_t offset = pgoff << PAGE_SHIFT;
9043 case IORING_OFF_SQ_RING:
9044 case IORING_OFF_CQ_RING:
9047 case IORING_OFF_SQES:
9051 return ERR_PTR(-EINVAL);
9054 page = virt_to_head_page(ptr);
9055 if (sz > page_size(page))
9056 return ERR_PTR(-EINVAL);
9063 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9065 size_t sz = vma->vm_end - vma->vm_start;
9069 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9071 return PTR_ERR(ptr);
9073 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9074 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9077 #else /* !CONFIG_MMU */
9079 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9081 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9084 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9086 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9089 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9090 unsigned long addr, unsigned long len,
9091 unsigned long pgoff, unsigned long flags)
9095 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9097 return PTR_ERR(ptr);
9099 return (unsigned long) ptr;
9102 #endif /* !CONFIG_MMU */
9104 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9109 if (!io_sqring_full(ctx))
9111 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9113 if (!io_sqring_full(ctx))
9116 } while (!signal_pending(current));
9118 finish_wait(&ctx->sqo_sq_wait, &wait);
9122 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9123 struct __kernel_timespec __user **ts,
9124 const sigset_t __user **sig)
9126 struct io_uring_getevents_arg arg;
9129 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9130 * is just a pointer to the sigset_t.
9132 if (!(flags & IORING_ENTER_EXT_ARG)) {
9133 *sig = (const sigset_t __user *) argp;
9139 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9140 * timespec and sigset_t pointers if good.
9142 if (*argsz != sizeof(arg))
9144 if (copy_from_user(&arg, argp, sizeof(arg)))
9146 *sig = u64_to_user_ptr(arg.sigmask);
9147 *argsz = arg.sigmask_sz;
9148 *ts = u64_to_user_ptr(arg.ts);
9152 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9153 u32, min_complete, u32, flags, const void __user *, argp,
9156 struct io_ring_ctx *ctx;
9163 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9164 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9168 if (unlikely(!f.file))
9172 if (unlikely(f.file->f_op != &io_uring_fops))
9176 ctx = f.file->private_data;
9177 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9181 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9185 * For SQ polling, the thread will do all submissions and completions.
9186 * Just return the requested submit count, and wake the thread if
9190 if (ctx->flags & IORING_SETUP_SQPOLL) {
9191 io_cqring_overflow_flush(ctx, false);
9194 if (unlikely(ctx->sq_data->thread == NULL)) {
9197 if (flags & IORING_ENTER_SQ_WAKEUP)
9198 wake_up(&ctx->sq_data->wait);
9199 if (flags & IORING_ENTER_SQ_WAIT) {
9200 ret = io_sqpoll_wait_sq(ctx);
9204 submitted = to_submit;
9205 } else if (to_submit) {
9206 ret = io_uring_add_task_file(ctx);
9209 mutex_lock(&ctx->uring_lock);
9210 submitted = io_submit_sqes(ctx, to_submit);
9211 mutex_unlock(&ctx->uring_lock);
9213 if (submitted != to_submit)
9216 if (flags & IORING_ENTER_GETEVENTS) {
9217 const sigset_t __user *sig;
9218 struct __kernel_timespec __user *ts;
9220 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9224 min_complete = min(min_complete, ctx->cq_entries);
9227 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9228 * space applications don't need to do io completion events
9229 * polling again, they can rely on io_sq_thread to do polling
9230 * work, which can reduce cpu usage and uring_lock contention.
9232 if (ctx->flags & IORING_SETUP_IOPOLL &&
9233 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9234 ret = io_iopoll_check(ctx, min_complete);
9236 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9241 percpu_ref_put(&ctx->refs);
9244 return submitted ? submitted : ret;
9247 #ifdef CONFIG_PROC_FS
9248 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9249 const struct cred *cred)
9251 struct user_namespace *uns = seq_user_ns(m);
9252 struct group_info *gi;
9257 seq_printf(m, "%5d\n", id);
9258 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9259 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9260 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9261 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9262 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9263 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9264 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9265 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9266 seq_puts(m, "\n\tGroups:\t");
9267 gi = cred->group_info;
9268 for (g = 0; g < gi->ngroups; g++) {
9269 seq_put_decimal_ull(m, g ? " " : "",
9270 from_kgid_munged(uns, gi->gid[g]));
9272 seq_puts(m, "\n\tCapEff:\t");
9273 cap = cred->cap_effective;
9274 CAP_FOR_EACH_U32(__capi)
9275 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9280 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9282 struct io_sq_data *sq = NULL;
9287 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9288 * since fdinfo case grabs it in the opposite direction of normal use
9289 * cases. If we fail to get the lock, we just don't iterate any
9290 * structures that could be going away outside the io_uring mutex.
9292 has_lock = mutex_trylock(&ctx->uring_lock);
9294 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9300 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9301 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9302 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9303 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9304 struct file *f = io_file_from_index(ctx, i);
9307 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9309 seq_printf(m, "%5u: <none>\n", i);
9311 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9312 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9313 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9314 unsigned int len = buf->ubuf_end - buf->ubuf;
9316 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9318 if (has_lock && !xa_empty(&ctx->personalities)) {
9319 unsigned long index;
9320 const struct cred *cred;
9322 seq_printf(m, "Personalities:\n");
9323 xa_for_each(&ctx->personalities, index, cred)
9324 io_uring_show_cred(m, index, cred);
9326 seq_printf(m, "PollList:\n");
9327 spin_lock_irq(&ctx->completion_lock);
9328 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9329 struct hlist_head *list = &ctx->cancel_hash[i];
9330 struct io_kiocb *req;
9332 hlist_for_each_entry(req, list, hash_node)
9333 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9334 req->task->task_works != NULL);
9336 spin_unlock_irq(&ctx->completion_lock);
9338 mutex_unlock(&ctx->uring_lock);
9341 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9343 struct io_ring_ctx *ctx = f->private_data;
9345 if (percpu_ref_tryget(&ctx->refs)) {
9346 __io_uring_show_fdinfo(ctx, m);
9347 percpu_ref_put(&ctx->refs);
9352 static const struct file_operations io_uring_fops = {
9353 .release = io_uring_release,
9354 .mmap = io_uring_mmap,
9356 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9357 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9359 .poll = io_uring_poll,
9360 .fasync = io_uring_fasync,
9361 #ifdef CONFIG_PROC_FS
9362 .show_fdinfo = io_uring_show_fdinfo,
9366 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9367 struct io_uring_params *p)
9369 struct io_rings *rings;
9370 size_t size, sq_array_offset;
9372 /* make sure these are sane, as we already accounted them */
9373 ctx->sq_entries = p->sq_entries;
9374 ctx->cq_entries = p->cq_entries;
9376 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9377 if (size == SIZE_MAX)
9380 rings = io_mem_alloc(size);
9385 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9386 rings->sq_ring_mask = p->sq_entries - 1;
9387 rings->cq_ring_mask = p->cq_entries - 1;
9388 rings->sq_ring_entries = p->sq_entries;
9389 rings->cq_ring_entries = p->cq_entries;
9390 ctx->sq_mask = rings->sq_ring_mask;
9391 ctx->cq_mask = rings->cq_ring_mask;
9393 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9394 if (size == SIZE_MAX) {
9395 io_mem_free(ctx->rings);
9400 ctx->sq_sqes = io_mem_alloc(size);
9401 if (!ctx->sq_sqes) {
9402 io_mem_free(ctx->rings);
9410 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9414 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9418 ret = io_uring_add_task_file(ctx);
9423 fd_install(fd, file);
9428 * Allocate an anonymous fd, this is what constitutes the application
9429 * visible backing of an io_uring instance. The application mmaps this
9430 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9431 * we have to tie this fd to a socket for file garbage collection purposes.
9433 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9436 #if defined(CONFIG_UNIX)
9439 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9442 return ERR_PTR(ret);
9445 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9446 O_RDWR | O_CLOEXEC);
9447 #if defined(CONFIG_UNIX)
9449 sock_release(ctx->ring_sock);
9450 ctx->ring_sock = NULL;
9452 ctx->ring_sock->file = file;
9458 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9459 struct io_uring_params __user *params)
9461 struct io_ring_ctx *ctx;
9467 if (entries > IORING_MAX_ENTRIES) {
9468 if (!(p->flags & IORING_SETUP_CLAMP))
9470 entries = IORING_MAX_ENTRIES;
9474 * Use twice as many entries for the CQ ring. It's possible for the
9475 * application to drive a higher depth than the size of the SQ ring,
9476 * since the sqes are only used at submission time. This allows for
9477 * some flexibility in overcommitting a bit. If the application has
9478 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9479 * of CQ ring entries manually.
9481 p->sq_entries = roundup_pow_of_two(entries);
9482 if (p->flags & IORING_SETUP_CQSIZE) {
9484 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9485 * to a power-of-two, if it isn't already. We do NOT impose
9486 * any cq vs sq ring sizing.
9490 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9491 if (!(p->flags & IORING_SETUP_CLAMP))
9493 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9495 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9496 if (p->cq_entries < p->sq_entries)
9499 p->cq_entries = 2 * p->sq_entries;
9502 ctx = io_ring_ctx_alloc(p);
9505 ctx->compat = in_compat_syscall();
9506 if (!capable(CAP_IPC_LOCK))
9507 ctx->user = get_uid(current_user());
9510 * This is just grabbed for accounting purposes. When a process exits,
9511 * the mm is exited and dropped before the files, hence we need to hang
9512 * on to this mm purely for the purposes of being able to unaccount
9513 * memory (locked/pinned vm). It's not used for anything else.
9515 mmgrab(current->mm);
9516 ctx->mm_account = current->mm;
9518 ret = io_allocate_scq_urings(ctx, p);
9522 ret = io_sq_offload_create(ctx, p);
9525 /* always set a rsrc node */
9526 io_rsrc_node_switch_start(ctx);
9527 io_rsrc_node_switch(ctx, NULL);
9529 memset(&p->sq_off, 0, sizeof(p->sq_off));
9530 p->sq_off.head = offsetof(struct io_rings, sq.head);
9531 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9532 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9533 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9534 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9535 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9536 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9538 memset(&p->cq_off, 0, sizeof(p->cq_off));
9539 p->cq_off.head = offsetof(struct io_rings, cq.head);
9540 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9541 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9542 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9543 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9544 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9545 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9547 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9548 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9549 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9550 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9551 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9553 if (copy_to_user(params, p, sizeof(*p))) {
9558 file = io_uring_get_file(ctx);
9560 ret = PTR_ERR(file);
9565 * Install ring fd as the very last thing, so we don't risk someone
9566 * having closed it before we finish setup
9568 ret = io_uring_install_fd(ctx, file);
9570 /* fput will clean it up */
9575 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9578 io_ring_ctx_wait_and_kill(ctx);
9583 * Sets up an aio uring context, and returns the fd. Applications asks for a
9584 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9585 * params structure passed in.
9587 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9589 struct io_uring_params p;
9592 if (copy_from_user(&p, params, sizeof(p)))
9594 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9599 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9600 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9601 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9602 IORING_SETUP_R_DISABLED))
9605 return io_uring_create(entries, &p, params);
9608 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9609 struct io_uring_params __user *, params)
9611 return io_uring_setup(entries, params);
9614 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9616 struct io_uring_probe *p;
9620 size = struct_size(p, ops, nr_args);
9621 if (size == SIZE_MAX)
9623 p = kzalloc(size, GFP_KERNEL);
9628 if (copy_from_user(p, arg, size))
9631 if (memchr_inv(p, 0, size))
9634 p->last_op = IORING_OP_LAST - 1;
9635 if (nr_args > IORING_OP_LAST)
9636 nr_args = IORING_OP_LAST;
9638 for (i = 0; i < nr_args; i++) {
9640 if (!io_op_defs[i].not_supported)
9641 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9646 if (copy_to_user(arg, p, size))
9653 static int io_register_personality(struct io_ring_ctx *ctx)
9655 const struct cred *creds;
9659 creds = get_current_cred();
9661 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9662 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9669 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9670 unsigned int nr_args)
9672 struct io_uring_restriction *res;
9676 /* Restrictions allowed only if rings started disabled */
9677 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9680 /* We allow only a single restrictions registration */
9681 if (ctx->restrictions.registered)
9684 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9687 size = array_size(nr_args, sizeof(*res));
9688 if (size == SIZE_MAX)
9691 res = memdup_user(arg, size);
9693 return PTR_ERR(res);
9697 for (i = 0; i < nr_args; i++) {
9698 switch (res[i].opcode) {
9699 case IORING_RESTRICTION_REGISTER_OP:
9700 if (res[i].register_op >= IORING_REGISTER_LAST) {
9705 __set_bit(res[i].register_op,
9706 ctx->restrictions.register_op);
9708 case IORING_RESTRICTION_SQE_OP:
9709 if (res[i].sqe_op >= IORING_OP_LAST) {
9714 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9716 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9717 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9719 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9720 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9729 /* Reset all restrictions if an error happened */
9731 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9733 ctx->restrictions.registered = true;
9739 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9741 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9744 if (ctx->restrictions.registered)
9745 ctx->restricted = 1;
9747 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9748 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9749 wake_up(&ctx->sq_data->wait);
9753 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9754 struct io_uring_rsrc_update2 *up,
9762 if (check_add_overflow(up->offset, nr_args, &tmp))
9764 err = io_rsrc_node_switch_start(ctx);
9769 case IORING_RSRC_FILE:
9770 return __io_sqe_files_update(ctx, up, nr_args);
9775 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9778 struct io_uring_rsrc_update2 up;
9782 memset(&up, 0, sizeof(up));
9783 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9785 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9788 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9791 struct io_uring_rsrc_update2 up;
9793 if (size != sizeof(up))
9795 if (copy_from_user(&up, arg, sizeof(up)))
9799 return __io_register_rsrc_update(ctx, up.type, &up, up.nr);
9802 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9805 struct io_uring_rsrc_register rr;
9807 /* keep it extendible */
9808 if (size != sizeof(rr))
9811 memset(&rr, 0, sizeof(rr));
9812 if (copy_from_user(&rr, arg, size))
9818 case IORING_RSRC_FILE:
9819 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
9820 rr.nr, u64_to_user_ptr(rr.tags));
9825 static bool io_register_op_must_quiesce(int op)
9828 case IORING_REGISTER_FILES:
9829 case IORING_UNREGISTER_FILES:
9830 case IORING_REGISTER_FILES_UPDATE:
9831 case IORING_REGISTER_PROBE:
9832 case IORING_REGISTER_PERSONALITY:
9833 case IORING_UNREGISTER_PERSONALITY:
9834 case IORING_REGISTER_RSRC:
9835 case IORING_REGISTER_RSRC_UPDATE:
9842 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9843 void __user *arg, unsigned nr_args)
9844 __releases(ctx->uring_lock)
9845 __acquires(ctx->uring_lock)
9850 * We're inside the ring mutex, if the ref is already dying, then
9851 * someone else killed the ctx or is already going through
9852 * io_uring_register().
9854 if (percpu_ref_is_dying(&ctx->refs))
9857 if (ctx->restricted) {
9858 if (opcode >= IORING_REGISTER_LAST)
9860 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
9861 if (!test_bit(opcode, ctx->restrictions.register_op))
9865 if (io_register_op_must_quiesce(opcode)) {
9866 percpu_ref_kill(&ctx->refs);
9869 * Drop uring mutex before waiting for references to exit. If
9870 * another thread is currently inside io_uring_enter() it might
9871 * need to grab the uring_lock to make progress. If we hold it
9872 * here across the drain wait, then we can deadlock. It's safe
9873 * to drop the mutex here, since no new references will come in
9874 * after we've killed the percpu ref.
9876 mutex_unlock(&ctx->uring_lock);
9878 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9881 ret = io_run_task_work_sig();
9885 mutex_lock(&ctx->uring_lock);
9888 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
9894 case IORING_REGISTER_BUFFERS:
9895 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9897 case IORING_UNREGISTER_BUFFERS:
9901 ret = io_sqe_buffers_unregister(ctx);
9903 case IORING_REGISTER_FILES:
9904 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
9906 case IORING_UNREGISTER_FILES:
9910 ret = io_sqe_files_unregister(ctx);
9912 case IORING_REGISTER_FILES_UPDATE:
9913 ret = io_register_files_update(ctx, arg, nr_args);
9915 case IORING_REGISTER_EVENTFD:
9916 case IORING_REGISTER_EVENTFD_ASYNC:
9920 ret = io_eventfd_register(ctx, arg);
9923 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9924 ctx->eventfd_async = 1;
9926 ctx->eventfd_async = 0;
9928 case IORING_UNREGISTER_EVENTFD:
9932 ret = io_eventfd_unregister(ctx);
9934 case IORING_REGISTER_PROBE:
9936 if (!arg || nr_args > 256)
9938 ret = io_probe(ctx, arg, nr_args);
9940 case IORING_REGISTER_PERSONALITY:
9944 ret = io_register_personality(ctx);
9946 case IORING_UNREGISTER_PERSONALITY:
9950 ret = io_unregister_personality(ctx, nr_args);
9952 case IORING_REGISTER_ENABLE_RINGS:
9956 ret = io_register_enable_rings(ctx);
9958 case IORING_REGISTER_RESTRICTIONS:
9959 ret = io_register_restrictions(ctx, arg, nr_args);
9961 case IORING_REGISTER_RSRC:
9962 ret = io_register_rsrc(ctx, arg, nr_args);
9964 case IORING_REGISTER_RSRC_UPDATE:
9965 ret = io_register_rsrc_update(ctx, arg, nr_args);
9972 if (io_register_op_must_quiesce(opcode)) {
9973 /* bring the ctx back to life */
9974 percpu_ref_reinit(&ctx->refs);
9975 reinit_completion(&ctx->ref_comp);
9980 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9981 void __user *, arg, unsigned int, nr_args)
9983 struct io_ring_ctx *ctx;
9992 if (f.file->f_op != &io_uring_fops)
9995 ctx = f.file->private_data;
9999 mutex_lock(&ctx->uring_lock);
10000 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10001 mutex_unlock(&ctx->uring_lock);
10002 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10003 ctx->cq_ev_fd != NULL, ret);
10009 static int __init io_uring_init(void)
10011 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10012 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10013 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10016 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10017 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10018 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10019 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10020 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10021 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10022 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10023 BUILD_BUG_SQE_ELEM(8, __u64, off);
10024 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10025 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10026 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10027 BUILD_BUG_SQE_ELEM(24, __u32, len);
10028 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10029 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10030 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10031 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10032 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10033 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10034 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10035 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10036 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10037 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10038 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10039 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10040 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10041 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10042 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10043 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10044 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10045 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10046 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10048 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10049 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10050 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10054 __initcall(io_uring_init);