1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <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/blk-mq.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>
81 #include <linux/tracehook.h>
82 #include <linux/audit.h>
83 #include <linux/security.h>
85 #define CREATE_TRACE_POINTS
86 #include <trace/events/io_uring.h>
88 #include <uapi/linux/io_uring.h>
93 #define IORING_MAX_ENTRIES 32768
94 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
95 #define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
98 #define IORING_MAX_FIXED_FILES (1U << 15)
99 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
100 IORING_REGISTER_LAST + IORING_OP_LAST)
102 #define IO_RSRC_TAG_TABLE_SHIFT (PAGE_SHIFT - 3)
103 #define IO_RSRC_TAG_TABLE_MAX (1U << IO_RSRC_TAG_TABLE_SHIFT)
104 #define IO_RSRC_TAG_TABLE_MASK (IO_RSRC_TAG_TABLE_MAX - 1)
106 #define IORING_MAX_REG_BUFFERS (1U << 14)
108 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
109 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
111 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
112 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
114 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
115 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
118 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
121 u32 head ____cacheline_aligned_in_smp;
122 u32 tail ____cacheline_aligned_in_smp;
126 * This data is shared with the application through the mmap at offsets
127 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
129 * The offsets to the member fields are published through struct
130 * io_sqring_offsets when calling io_uring_setup.
134 * Head and tail offsets into the ring; the offsets need to be
135 * masked to get valid indices.
137 * The kernel controls head of the sq ring and the tail of the cq ring,
138 * and the application controls tail of the sq ring and the head of the
141 struct io_uring sq, cq;
143 * Bitmasks to apply to head and tail offsets (constant, equals
146 u32 sq_ring_mask, cq_ring_mask;
147 /* Ring sizes (constant, power of 2) */
148 u32 sq_ring_entries, cq_ring_entries;
150 * Number of invalid entries dropped by the kernel due to
151 * invalid index stored in array
153 * Written by the kernel, shouldn't be modified by the
154 * application (i.e. get number of "new events" by comparing to
157 * After a new SQ head value was read by the application this
158 * counter includes all submissions that were dropped reaching
159 * the new SQ head (and possibly more).
165 * Written by the kernel, shouldn't be modified by the
168 * The application needs a full memory barrier before checking
169 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
175 * Written by the application, shouldn't be modified by the
180 * Number of completion events lost because the queue was full;
181 * this should be avoided by the application by making sure
182 * there are not more requests pending than there is space in
183 * the completion queue.
185 * Written by the kernel, shouldn't be modified by the
186 * application (i.e. get number of "new events" by comparing to
189 * As completion events come in out of order this counter is not
190 * ordered with any other data.
194 * Ring buffer of completion events.
196 * The kernel writes completion events fresh every time they are
197 * produced, so the application is allowed to modify pending
200 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
203 enum io_uring_cmd_flags {
204 IO_URING_F_COMPLETE_DEFER = 1,
205 IO_URING_F_UNLOCKED = 2,
206 /* int's last bit, sign checks are usually faster than a bit test */
207 IO_URING_F_NONBLOCK = INT_MIN,
210 struct io_mapped_ubuf {
213 unsigned int nr_bvecs;
214 unsigned long acct_pages;
215 struct bio_vec bvec[];
220 struct io_overflow_cqe {
221 struct io_uring_cqe cqe;
222 struct list_head list;
225 struct io_fixed_file {
226 /* file * with additional FFS_* flags */
227 unsigned long file_ptr;
231 struct list_head list;
236 struct io_mapped_ubuf *buf;
240 struct io_file_table {
241 struct io_fixed_file *files;
244 struct io_rsrc_node {
245 struct percpu_ref refs;
246 struct list_head node;
247 struct list_head rsrc_list;
248 struct io_rsrc_data *rsrc_data;
249 struct llist_node llist;
253 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
255 struct io_rsrc_data {
256 struct io_ring_ctx *ctx;
262 struct completion done;
267 struct list_head list;
273 struct io_restriction {
274 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
275 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
276 u8 sqe_flags_allowed;
277 u8 sqe_flags_required;
282 IO_SQ_THREAD_SHOULD_STOP = 0,
283 IO_SQ_THREAD_SHOULD_PARK,
288 atomic_t park_pending;
291 /* ctx's that are using this sqd */
292 struct list_head ctx_list;
294 struct task_struct *thread;
295 struct wait_queue_head wait;
297 unsigned sq_thread_idle;
303 struct completion exited;
306 #define IO_COMPL_BATCH 32
307 #define IO_REQ_CACHE_SIZE 32
308 #define IO_REQ_ALLOC_BATCH 8
310 struct io_submit_link {
311 struct io_kiocb *head;
312 struct io_kiocb *last;
315 struct io_submit_state {
316 /* inline/task_work completion list, under ->uring_lock */
317 struct io_wq_work_node free_list;
318 /* batch completion logic */
319 struct io_wq_work_list compl_reqs;
320 struct io_submit_link link;
325 unsigned short submit_nr;
326 struct blk_plug plug;
330 /* const or read-mostly hot data */
332 struct percpu_ref refs;
334 struct io_rings *rings;
336 unsigned int compat: 1;
337 unsigned int drain_next: 1;
338 unsigned int eventfd_async: 1;
339 unsigned int restricted: 1;
340 unsigned int off_timeout_used: 1;
341 unsigned int drain_active: 1;
342 unsigned int drain_disabled: 1;
343 } ____cacheline_aligned_in_smp;
345 /* submission data */
347 struct mutex uring_lock;
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 struct io_uring_sqe *sq_sqes;
362 unsigned cached_sq_head;
364 struct list_head defer_list;
367 * Fixed resources fast path, should be accessed only under
368 * uring_lock, and updated through io_uring_register(2)
370 struct io_rsrc_node *rsrc_node;
371 int rsrc_cached_refs;
372 struct io_file_table file_table;
373 unsigned nr_user_files;
374 unsigned nr_user_bufs;
375 struct io_mapped_ubuf **user_bufs;
377 struct io_submit_state submit_state;
378 struct list_head timeout_list;
379 struct list_head ltimeout_list;
380 struct list_head cq_overflow_list;
381 struct xarray io_buffers;
382 struct xarray personalities;
384 unsigned sq_thread_idle;
385 } ____cacheline_aligned_in_smp;
387 /* IRQ completion list, under ->completion_lock */
388 struct io_wq_work_list locked_free_list;
389 unsigned int locked_free_nr;
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;
397 unsigned long check_cq_overflow;
400 unsigned cached_cq_tail;
402 struct eventfd_ctx *cq_ev_fd;
403 struct wait_queue_head cq_wait;
405 atomic_t cq_timeouts;
406 unsigned cq_last_tm_flush;
407 } ____cacheline_aligned_in_smp;
410 spinlock_t completion_lock;
412 spinlock_t timeout_lock;
415 * ->iopoll_list is protected by the ctx->uring_lock for
416 * io_uring instances that don't use IORING_SETUP_SQPOLL.
417 * For SQPOLL, only the single threaded io_sq_thread() will
418 * manipulate the list, hence no extra locking is needed there.
420 struct io_wq_work_list iopoll_list;
421 struct hlist_head *cancel_hash;
422 unsigned cancel_hash_bits;
423 bool poll_multi_queue;
424 } ____cacheline_aligned_in_smp;
426 struct io_restriction restrictions;
428 /* slow path rsrc auxilary data, used by update/register */
430 struct io_rsrc_node *rsrc_backup_node;
431 struct io_mapped_ubuf *dummy_ubuf;
432 struct io_rsrc_data *file_data;
433 struct io_rsrc_data *buf_data;
435 struct delayed_work rsrc_put_work;
436 struct llist_head rsrc_put_llist;
437 struct list_head rsrc_ref_list;
438 spinlock_t rsrc_ref_lock;
441 /* Keep this last, we don't need it for the fast path */
443 #if defined(CONFIG_UNIX)
444 struct socket *ring_sock;
446 /* hashed buffered write serialization */
447 struct io_wq_hash *hash_map;
449 /* Only used for accounting purposes */
450 struct user_struct *user;
451 struct mm_struct *mm_account;
453 /* ctx exit and cancelation */
454 struct llist_head fallback_llist;
455 struct delayed_work fallback_work;
456 struct work_struct exit_work;
457 struct list_head tctx_list;
458 struct completion ref_comp;
460 bool iowq_limits_set;
464 struct io_uring_task {
465 /* submission side */
468 struct wait_queue_head wait;
469 const struct io_ring_ctx *last;
471 struct percpu_counter inflight;
472 atomic_t inflight_tracked;
475 spinlock_t task_lock;
476 struct io_wq_work_list task_list;
477 struct io_wq_work_list prior_task_list;
478 struct callback_head task_work;
483 * First field must be the file pointer in all the
484 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
486 struct io_poll_iocb {
488 struct wait_queue_head *head;
490 struct wait_queue_entry wait;
493 struct io_poll_update {
499 bool update_user_data;
508 struct io_timeout_data {
509 struct io_kiocb *req;
510 struct hrtimer timer;
511 struct timespec64 ts;
512 enum hrtimer_mode mode;
518 struct sockaddr __user *addr;
519 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;
545 /* for linked completions */
546 struct io_kiocb *prev;
549 struct io_timeout_rem {
554 struct timespec64 ts;
560 /* NOTE: kiocb has the file as the first member, so don't do it here */
568 struct sockaddr __user *addr;
575 struct compat_msghdr __user *umsg_compat;
576 struct user_msghdr __user *umsg;
588 struct filename *filename;
590 unsigned long nofile;
593 struct io_rsrc_update {
619 struct epoll_event event;
623 struct file *file_out;
624 struct file *file_in;
631 struct io_provide_buf {
645 const char __user *filename;
646 struct statx __user *buffer;
658 struct filename *oldpath;
659 struct filename *newpath;
667 struct filename *filename;
674 struct filename *filename;
680 struct filename *oldpath;
681 struct filename *newpath;
688 struct filename *oldpath;
689 struct filename *newpath;
693 struct io_async_connect {
694 struct sockaddr_storage address;
697 struct io_async_msghdr {
698 struct iovec fast_iov[UIO_FASTIOV];
699 /* points to an allocated iov, if NULL we use fast_iov instead */
700 struct iovec *free_iov;
701 struct sockaddr __user *uaddr;
703 struct sockaddr_storage addr;
707 struct iov_iter iter;
708 struct iov_iter_state iter_state;
709 struct iovec fast_iov[UIO_FASTIOV];
713 struct io_rw_state s;
714 const struct iovec *free_iovec;
716 struct wait_page_queue wpq;
720 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
721 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
722 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
723 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
724 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
725 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
726 REQ_F_CQE_SKIP_BIT = IOSQE_CQE_SKIP_SUCCESS_BIT,
728 /* first byte is taken by user flags, shift it to not overlap */
733 REQ_F_LINK_TIMEOUT_BIT,
734 REQ_F_NEED_CLEANUP_BIT,
736 REQ_F_BUFFER_SELECTED_BIT,
737 REQ_F_COMPLETE_INLINE_BIT,
741 REQ_F_ARM_LTIMEOUT_BIT,
742 REQ_F_ASYNC_DATA_BIT,
743 REQ_F_SKIP_LINK_CQES_BIT,
744 /* keep async read/write and isreg together and in order */
745 REQ_F_SUPPORT_NOWAIT_BIT,
748 /* not a real bit, just to check we're not overflowing the space */
754 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
755 /* drain existing IO first */
756 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
758 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
759 /* doesn't sever on completion < 0 */
760 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
762 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
763 /* IOSQE_BUFFER_SELECT */
764 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
765 /* IOSQE_CQE_SKIP_SUCCESS */
766 REQ_F_CQE_SKIP = BIT(REQ_F_CQE_SKIP_BIT),
768 /* fail rest of links */
769 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
770 /* on inflight list, should be cancelled and waited on exit reliably */
771 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
772 /* read/write uses file position */
773 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
774 /* must not punt to workers */
775 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
776 /* has or had linked timeout */
777 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
779 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
780 /* already went through poll handler */
781 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
782 /* buffer already selected */
783 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
784 /* completion is deferred through io_comp_state */
785 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
786 /* caller should reissue async */
787 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
788 /* supports async reads/writes */
789 REQ_F_SUPPORT_NOWAIT = BIT(REQ_F_SUPPORT_NOWAIT_BIT),
791 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
792 /* has creds assigned */
793 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
794 /* skip refcounting if not set */
795 REQ_F_REFCOUNT = BIT(REQ_F_REFCOUNT_BIT),
796 /* there is a linked timeout that has to be armed */
797 REQ_F_ARM_LTIMEOUT = BIT(REQ_F_ARM_LTIMEOUT_BIT),
798 /* ->async_data allocated */
799 REQ_F_ASYNC_DATA = BIT(REQ_F_ASYNC_DATA_BIT),
800 /* don't post CQEs while failing linked requests */
801 REQ_F_SKIP_LINK_CQES = BIT(REQ_F_SKIP_LINK_CQES_BIT),
805 struct io_poll_iocb poll;
806 struct io_poll_iocb *double_poll;
809 typedef void (*io_req_tw_func_t)(struct io_kiocb *req, bool *locked);
811 struct io_task_work {
813 struct io_wq_work_node node;
814 struct llist_node fallback_node;
816 io_req_tw_func_t func;
820 IORING_RSRC_FILE = 0,
821 IORING_RSRC_BUFFER = 1,
825 * NOTE! Each of the iocb union members has the file pointer
826 * as the first entry in their struct definition. So you can
827 * access the file pointer through any of the sub-structs,
828 * or directly as just 'ki_filp' in this struct.
834 struct io_poll_iocb poll;
835 struct io_poll_update poll_update;
836 struct io_accept accept;
838 struct io_cancel cancel;
839 struct io_timeout timeout;
840 struct io_timeout_rem timeout_rem;
841 struct io_connect connect;
842 struct io_sr_msg sr_msg;
844 struct io_close close;
845 struct io_rsrc_update rsrc_update;
846 struct io_fadvise fadvise;
847 struct io_madvise madvise;
848 struct io_epoll epoll;
849 struct io_splice splice;
850 struct io_provide_buf pbuf;
851 struct io_statx statx;
852 struct io_shutdown shutdown;
853 struct io_rename rename;
854 struct io_unlink unlink;
855 struct io_mkdir mkdir;
856 struct io_symlink symlink;
857 struct io_hardlink hardlink;
861 /* polled IO has completed */
870 struct io_ring_ctx *ctx;
871 struct task_struct *task;
873 struct percpu_ref *fixed_rsrc_refs;
874 /* store used ubuf, so we can prevent reloading */
875 struct io_mapped_ubuf *imu;
877 /* used by request caches, completion batching and iopoll */
878 struct io_wq_work_node comp_list;
880 struct io_kiocb *link;
881 struct io_task_work io_task_work;
882 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
883 struct hlist_node hash_node;
884 /* internal polling, see IORING_FEAT_FAST_POLL */
885 struct async_poll *apoll;
886 /* opcode allocated if it needs to store data for async defer */
888 struct io_wq_work work;
889 /* custom credentials, valid IFF REQ_F_CREDS is set */
890 const struct cred *creds;
891 /* stores selected buf, valid IFF REQ_F_BUFFER_SELECTED is set */
892 struct io_buffer *kbuf;
896 struct io_tctx_node {
897 struct list_head ctx_node;
898 struct task_struct *task;
899 struct io_ring_ctx *ctx;
902 struct io_defer_entry {
903 struct list_head list;
904 struct io_kiocb *req;
909 /* needs req->file assigned */
910 unsigned needs_file : 1;
911 /* should block plug */
913 /* hash wq insertion if file is a regular file */
914 unsigned hash_reg_file : 1;
915 /* unbound wq insertion if file is a non-regular file */
916 unsigned unbound_nonreg_file : 1;
917 /* set if opcode supports polled "wait" */
919 unsigned pollout : 1;
920 /* op supports buffer selection */
921 unsigned buffer_select : 1;
922 /* do prep async if is going to be punted */
923 unsigned needs_async_setup : 1;
924 /* opcode is not supported by this kernel */
925 unsigned not_supported : 1;
927 unsigned audit_skip : 1;
928 /* size of async data needed, if any */
929 unsigned short async_size;
932 static const struct io_op_def io_op_defs[] = {
933 [IORING_OP_NOP] = {},
934 [IORING_OP_READV] = {
936 .unbound_nonreg_file = 1,
939 .needs_async_setup = 1,
942 .async_size = sizeof(struct io_async_rw),
944 [IORING_OP_WRITEV] = {
947 .unbound_nonreg_file = 1,
949 .needs_async_setup = 1,
952 .async_size = sizeof(struct io_async_rw),
954 [IORING_OP_FSYNC] = {
958 [IORING_OP_READ_FIXED] = {
960 .unbound_nonreg_file = 1,
964 .async_size = sizeof(struct io_async_rw),
966 [IORING_OP_WRITE_FIXED] = {
969 .unbound_nonreg_file = 1,
973 .async_size = sizeof(struct io_async_rw),
975 [IORING_OP_POLL_ADD] = {
977 .unbound_nonreg_file = 1,
980 [IORING_OP_POLL_REMOVE] = {
983 [IORING_OP_SYNC_FILE_RANGE] = {
987 [IORING_OP_SENDMSG] = {
989 .unbound_nonreg_file = 1,
991 .needs_async_setup = 1,
992 .async_size = sizeof(struct io_async_msghdr),
994 [IORING_OP_RECVMSG] = {
996 .unbound_nonreg_file = 1,
999 .needs_async_setup = 1,
1000 .async_size = sizeof(struct io_async_msghdr),
1002 [IORING_OP_TIMEOUT] = {
1004 .async_size = sizeof(struct io_timeout_data),
1006 [IORING_OP_TIMEOUT_REMOVE] = {
1007 /* used by timeout updates' prep() */
1010 [IORING_OP_ACCEPT] = {
1012 .unbound_nonreg_file = 1,
1015 [IORING_OP_ASYNC_CANCEL] = {
1018 [IORING_OP_LINK_TIMEOUT] = {
1020 .async_size = sizeof(struct io_timeout_data),
1022 [IORING_OP_CONNECT] = {
1024 .unbound_nonreg_file = 1,
1026 .needs_async_setup = 1,
1027 .async_size = sizeof(struct io_async_connect),
1029 [IORING_OP_FALLOCATE] = {
1032 [IORING_OP_OPENAT] = {},
1033 [IORING_OP_CLOSE] = {},
1034 [IORING_OP_FILES_UPDATE] = {
1037 [IORING_OP_STATX] = {
1040 [IORING_OP_READ] = {
1042 .unbound_nonreg_file = 1,
1047 .async_size = sizeof(struct io_async_rw),
1049 [IORING_OP_WRITE] = {
1052 .unbound_nonreg_file = 1,
1056 .async_size = sizeof(struct io_async_rw),
1058 [IORING_OP_FADVISE] = {
1062 [IORING_OP_MADVISE] = {},
1063 [IORING_OP_SEND] = {
1065 .unbound_nonreg_file = 1,
1069 [IORING_OP_RECV] = {
1071 .unbound_nonreg_file = 1,
1076 [IORING_OP_OPENAT2] = {
1078 [IORING_OP_EPOLL_CTL] = {
1079 .unbound_nonreg_file = 1,
1082 [IORING_OP_SPLICE] = {
1085 .unbound_nonreg_file = 1,
1088 [IORING_OP_PROVIDE_BUFFERS] = {
1091 [IORING_OP_REMOVE_BUFFERS] = {
1097 .unbound_nonreg_file = 1,
1100 [IORING_OP_SHUTDOWN] = {
1103 [IORING_OP_RENAMEAT] = {},
1104 [IORING_OP_UNLINKAT] = {},
1105 [IORING_OP_MKDIRAT] = {},
1106 [IORING_OP_SYMLINKAT] = {},
1107 [IORING_OP_LINKAT] = {},
1110 /* requests with any of those set should undergo io_disarm_next() */
1111 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
1113 static bool io_disarm_next(struct io_kiocb *req);
1114 static void io_uring_del_tctx_node(unsigned long index);
1115 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1116 struct task_struct *task,
1118 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1120 static void io_fill_cqe_req(struct io_kiocb *req, s32 res, u32 cflags);
1122 static void io_put_req(struct io_kiocb *req);
1123 static void io_put_req_deferred(struct io_kiocb *req);
1124 static void io_dismantle_req(struct io_kiocb *req);
1125 static void io_queue_linked_timeout(struct io_kiocb *req);
1126 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1127 struct io_uring_rsrc_update2 *up,
1129 static void io_clean_op(struct io_kiocb *req);
1130 static struct file *io_file_get(struct io_ring_ctx *ctx,
1131 struct io_kiocb *req, int fd, bool fixed);
1132 static void __io_queue_sqe(struct io_kiocb *req);
1133 static void io_rsrc_put_work(struct work_struct *work);
1135 static void io_req_task_queue(struct io_kiocb *req);
1136 static void __io_submit_flush_completions(struct io_ring_ctx *ctx);
1137 static int io_req_prep_async(struct io_kiocb *req);
1139 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
1140 unsigned int issue_flags, u32 slot_index);
1141 static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags);
1143 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer);
1145 static struct kmem_cache *req_cachep;
1147 static const struct file_operations io_uring_fops;
1149 struct sock *io_uring_get_socket(struct file *file)
1151 #if defined(CONFIG_UNIX)
1152 if (file->f_op == &io_uring_fops) {
1153 struct io_ring_ctx *ctx = file->private_data;
1155 return ctx->ring_sock->sk;
1160 EXPORT_SYMBOL(io_uring_get_socket);
1162 static inline void io_tw_lock(struct io_ring_ctx *ctx, bool *locked)
1165 mutex_lock(&ctx->uring_lock);
1170 #define io_for_each_link(pos, head) \
1171 for (pos = (head); pos; pos = pos->link)
1174 * Shamelessly stolen from the mm implementation of page reference checking,
1175 * see commit f958d7b528b1 for details.
1177 #define req_ref_zero_or_close_to_overflow(req) \
1178 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1180 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1182 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1183 return atomic_inc_not_zero(&req->refs);
1186 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1188 if (likely(!(req->flags & REQ_F_REFCOUNT)))
1191 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1192 return atomic_dec_and_test(&req->refs);
1195 static inline void req_ref_get(struct io_kiocb *req)
1197 WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1198 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1199 atomic_inc(&req->refs);
1202 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
1204 if (!wq_list_empty(&ctx->submit_state.compl_reqs))
1205 __io_submit_flush_completions(ctx);
1208 static inline void __io_req_set_refcount(struct io_kiocb *req, int nr)
1210 if (!(req->flags & REQ_F_REFCOUNT)) {
1211 req->flags |= REQ_F_REFCOUNT;
1212 atomic_set(&req->refs, nr);
1216 static inline void io_req_set_refcount(struct io_kiocb *req)
1218 __io_req_set_refcount(req, 1);
1221 #define IO_RSRC_REF_BATCH 100
1223 static inline void io_req_put_rsrc_locked(struct io_kiocb *req,
1224 struct io_ring_ctx *ctx)
1225 __must_hold(&ctx->uring_lock)
1227 struct percpu_ref *ref = req->fixed_rsrc_refs;
1230 if (ref == &ctx->rsrc_node->refs)
1231 ctx->rsrc_cached_refs++;
1233 percpu_ref_put(ref);
1237 static inline void io_req_put_rsrc(struct io_kiocb *req, struct io_ring_ctx *ctx)
1239 if (req->fixed_rsrc_refs)
1240 percpu_ref_put(req->fixed_rsrc_refs);
1243 static __cold void io_rsrc_refs_drop(struct io_ring_ctx *ctx)
1244 __must_hold(&ctx->uring_lock)
1246 if (ctx->rsrc_cached_refs) {
1247 percpu_ref_put_many(&ctx->rsrc_node->refs, ctx->rsrc_cached_refs);
1248 ctx->rsrc_cached_refs = 0;
1252 static void io_rsrc_refs_refill(struct io_ring_ctx *ctx)
1253 __must_hold(&ctx->uring_lock)
1255 ctx->rsrc_cached_refs += IO_RSRC_REF_BATCH;
1256 percpu_ref_get_many(&ctx->rsrc_node->refs, IO_RSRC_REF_BATCH);
1259 static inline void io_req_set_rsrc_node(struct io_kiocb *req,
1260 struct io_ring_ctx *ctx)
1262 if (!req->fixed_rsrc_refs) {
1263 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1264 ctx->rsrc_cached_refs--;
1265 if (unlikely(ctx->rsrc_cached_refs < 0))
1266 io_rsrc_refs_refill(ctx);
1270 static unsigned int __io_put_kbuf(struct io_kiocb *req)
1272 struct io_buffer *kbuf = req->kbuf;
1273 unsigned int cflags;
1275 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1276 cflags |= IORING_CQE_F_BUFFER;
1277 req->flags &= ~REQ_F_BUFFER_SELECTED;
1283 static inline unsigned int io_put_kbuf(struct io_kiocb *req)
1285 if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
1287 return __io_put_kbuf(req);
1290 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1292 bool got = percpu_ref_tryget(ref);
1294 /* already at zero, wait for ->release() */
1296 wait_for_completion(compl);
1297 percpu_ref_resurrect(ref);
1299 percpu_ref_put(ref);
1302 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1304 __must_hold(&req->ctx->timeout_lock)
1306 struct io_kiocb *req;
1308 if (task && head->task != task)
1313 io_for_each_link(req, head) {
1314 if (req->flags & REQ_F_INFLIGHT)
1320 static bool io_match_linked(struct io_kiocb *head)
1322 struct io_kiocb *req;
1324 io_for_each_link(req, head) {
1325 if (req->flags & REQ_F_INFLIGHT)
1332 * As io_match_task() but protected against racing with linked timeouts.
1333 * User must not hold timeout_lock.
1335 static bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
1340 if (task && head->task != task)
1345 if (head->flags & REQ_F_LINK_TIMEOUT) {
1346 struct io_ring_ctx *ctx = head->ctx;
1348 /* protect against races with linked timeouts */
1349 spin_lock_irq(&ctx->timeout_lock);
1350 matched = io_match_linked(head);
1351 spin_unlock_irq(&ctx->timeout_lock);
1353 matched = io_match_linked(head);
1358 static inline bool req_has_async_data(struct io_kiocb *req)
1360 return req->flags & REQ_F_ASYNC_DATA;
1363 static inline void req_set_fail(struct io_kiocb *req)
1365 req->flags |= REQ_F_FAIL;
1366 if (req->flags & REQ_F_CQE_SKIP) {
1367 req->flags &= ~REQ_F_CQE_SKIP;
1368 req->flags |= REQ_F_SKIP_LINK_CQES;
1372 static inline void req_fail_link_node(struct io_kiocb *req, int res)
1378 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
1380 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1382 complete(&ctx->ref_comp);
1385 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1387 return !req->timeout.off;
1390 static __cold void io_fallback_req_func(struct work_struct *work)
1392 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
1393 fallback_work.work);
1394 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
1395 struct io_kiocb *req, *tmp;
1396 bool locked = false;
1398 percpu_ref_get(&ctx->refs);
1399 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
1400 req->io_task_work.func(req, &locked);
1403 io_submit_flush_completions(ctx);
1404 mutex_unlock(&ctx->uring_lock);
1406 percpu_ref_put(&ctx->refs);
1409 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1411 struct io_ring_ctx *ctx;
1414 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1419 * Use 5 bits less than the max cq entries, that should give us around
1420 * 32 entries per hash list if totally full and uniformly spread.
1422 hash_bits = ilog2(p->cq_entries);
1426 ctx->cancel_hash_bits = hash_bits;
1427 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1429 if (!ctx->cancel_hash)
1431 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1433 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1434 if (!ctx->dummy_ubuf)
1436 /* set invalid range, so io_import_fixed() fails meeting it */
1437 ctx->dummy_ubuf->ubuf = -1UL;
1439 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1440 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1443 ctx->flags = p->flags;
1444 init_waitqueue_head(&ctx->sqo_sq_wait);
1445 INIT_LIST_HEAD(&ctx->sqd_list);
1446 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1447 init_completion(&ctx->ref_comp);
1448 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1449 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1450 mutex_init(&ctx->uring_lock);
1451 init_waitqueue_head(&ctx->cq_wait);
1452 spin_lock_init(&ctx->completion_lock);
1453 spin_lock_init(&ctx->timeout_lock);
1454 INIT_WQ_LIST(&ctx->iopoll_list);
1455 INIT_LIST_HEAD(&ctx->defer_list);
1456 INIT_LIST_HEAD(&ctx->timeout_list);
1457 INIT_LIST_HEAD(&ctx->ltimeout_list);
1458 spin_lock_init(&ctx->rsrc_ref_lock);
1459 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1460 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1461 init_llist_head(&ctx->rsrc_put_llist);
1462 INIT_LIST_HEAD(&ctx->tctx_list);
1463 ctx->submit_state.free_list.next = NULL;
1464 INIT_WQ_LIST(&ctx->locked_free_list);
1465 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1466 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
1469 kfree(ctx->dummy_ubuf);
1470 kfree(ctx->cancel_hash);
1475 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1477 struct io_rings *r = ctx->rings;
1479 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1483 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1485 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1486 struct io_ring_ctx *ctx = req->ctx;
1488 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1494 #define FFS_NOWAIT 0x1UL
1495 #define FFS_ISREG 0x2UL
1496 #define FFS_MASK ~(FFS_NOWAIT|FFS_ISREG)
1498 static inline bool io_req_ffs_set(struct io_kiocb *req)
1500 return req->flags & REQ_F_FIXED_FILE;
1503 static inline void io_req_track_inflight(struct io_kiocb *req)
1505 if (!(req->flags & REQ_F_INFLIGHT)) {
1506 req->flags |= REQ_F_INFLIGHT;
1507 atomic_inc(¤t->io_uring->inflight_tracked);
1511 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
1513 if (WARN_ON_ONCE(!req->link))
1516 req->flags &= ~REQ_F_ARM_LTIMEOUT;
1517 req->flags |= REQ_F_LINK_TIMEOUT;
1519 /* linked timeouts should have two refs once prep'ed */
1520 io_req_set_refcount(req);
1521 __io_req_set_refcount(req->link, 2);
1525 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
1527 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
1529 return __io_prep_linked_timeout(req);
1532 static void io_prep_async_work(struct io_kiocb *req)
1534 const struct io_op_def *def = &io_op_defs[req->opcode];
1535 struct io_ring_ctx *ctx = req->ctx;
1537 if (!(req->flags & REQ_F_CREDS)) {
1538 req->flags |= REQ_F_CREDS;
1539 req->creds = get_current_cred();
1542 req->work.list.next = NULL;
1543 req->work.flags = 0;
1544 if (req->flags & REQ_F_FORCE_ASYNC)
1545 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1547 if (req->flags & REQ_F_ISREG) {
1548 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1549 io_wq_hash_work(&req->work, file_inode(req->file));
1550 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1551 if (def->unbound_nonreg_file)
1552 req->work.flags |= IO_WQ_WORK_UNBOUND;
1555 switch (req->opcode) {
1556 case IORING_OP_SPLICE:
1558 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1559 req->work.flags |= IO_WQ_WORK_UNBOUND;
1564 static void io_prep_async_link(struct io_kiocb *req)
1566 struct io_kiocb *cur;
1568 if (req->flags & REQ_F_LINK_TIMEOUT) {
1569 struct io_ring_ctx *ctx = req->ctx;
1571 spin_lock_irq(&ctx->timeout_lock);
1572 io_for_each_link(cur, req)
1573 io_prep_async_work(cur);
1574 spin_unlock_irq(&ctx->timeout_lock);
1576 io_for_each_link(cur, req)
1577 io_prep_async_work(cur);
1581 static inline void io_req_add_compl_list(struct io_kiocb *req)
1583 struct io_ring_ctx *ctx = req->ctx;
1584 struct io_submit_state *state = &ctx->submit_state;
1586 if (!(req->flags & REQ_F_CQE_SKIP))
1587 ctx->submit_state.flush_cqes = true;
1588 wq_list_add_tail(&req->comp_list, &state->compl_reqs);
1591 static void io_queue_async_work(struct io_kiocb *req, bool *dont_use)
1593 struct io_ring_ctx *ctx = req->ctx;
1594 struct io_kiocb *link = io_prep_linked_timeout(req);
1595 struct io_uring_task *tctx = req->task->io_uring;
1598 BUG_ON(!tctx->io_wq);
1600 /* init ->work of the whole link before punting */
1601 io_prep_async_link(req);
1604 * Not expected to happen, but if we do have a bug where this _can_
1605 * happen, catch it here and ensure the request is marked as
1606 * canceled. That will make io-wq go through the usual work cancel
1607 * procedure rather than attempt to run this request (or create a new
1610 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1611 req->work.flags |= IO_WQ_WORK_CANCEL;
1613 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1614 &req->work, req->flags);
1615 io_wq_enqueue(tctx->io_wq, &req->work);
1617 io_queue_linked_timeout(link);
1620 static void io_kill_timeout(struct io_kiocb *req, int status)
1621 __must_hold(&req->ctx->completion_lock)
1622 __must_hold(&req->ctx->timeout_lock)
1624 struct io_timeout_data *io = req->async_data;
1626 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1629 atomic_set(&req->ctx->cq_timeouts,
1630 atomic_read(&req->ctx->cq_timeouts) + 1);
1631 list_del_init(&req->timeout.list);
1632 io_fill_cqe_req(req, status, 0);
1633 io_put_req_deferred(req);
1637 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
1639 while (!list_empty(&ctx->defer_list)) {
1640 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1641 struct io_defer_entry, list);
1643 if (req_need_defer(de->req, de->seq))
1645 list_del_init(&de->list);
1646 io_req_task_queue(de->req);
1651 static __cold void io_flush_timeouts(struct io_ring_ctx *ctx)
1652 __must_hold(&ctx->completion_lock)
1654 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1656 spin_lock_irq(&ctx->timeout_lock);
1657 while (!list_empty(&ctx->timeout_list)) {
1658 u32 events_needed, events_got;
1659 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1660 struct io_kiocb, timeout.list);
1662 if (io_is_timeout_noseq(req))
1666 * Since seq can easily wrap around over time, subtract
1667 * the last seq at which timeouts were flushed before comparing.
1668 * Assuming not more than 2^31-1 events have happened since,
1669 * these subtractions won't have wrapped, so we can check if
1670 * target is in [last_seq, current_seq] by comparing the two.
1672 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1673 events_got = seq - ctx->cq_last_tm_flush;
1674 if (events_got < events_needed)
1677 list_del_init(&req->timeout.list);
1678 io_kill_timeout(req, 0);
1680 ctx->cq_last_tm_flush = seq;
1681 spin_unlock_irq(&ctx->timeout_lock);
1684 static __cold void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1686 if (ctx->off_timeout_used)
1687 io_flush_timeouts(ctx);
1688 if (ctx->drain_active)
1689 io_queue_deferred(ctx);
1692 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1694 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1695 __io_commit_cqring_flush(ctx);
1696 /* order cqe stores with ring update */
1697 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1700 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1702 struct io_rings *r = ctx->rings;
1704 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1707 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1709 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1712 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1714 struct io_rings *rings = ctx->rings;
1715 unsigned tail, mask = ctx->cq_entries - 1;
1718 * writes to the cq entry need to come after reading head; the
1719 * control dependency is enough as we're using WRITE_ONCE to
1722 if (__io_cqring_events(ctx) == ctx->cq_entries)
1725 tail = ctx->cached_cq_tail++;
1726 return &rings->cqes[tail & mask];
1729 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1731 if (likely(!ctx->cq_ev_fd))
1733 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1735 return !ctx->eventfd_async || io_wq_current_is_worker();
1739 * This should only get called when at least one event has been posted.
1740 * Some applications rely on the eventfd notification count only changing
1741 * IFF a new CQE has been added to the CQ ring. There's no depedency on
1742 * 1:1 relationship between how many times this function is called (and
1743 * hence the eventfd count) and number of CQEs posted to the CQ ring.
1745 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1748 * wake_up_all() may seem excessive, but io_wake_function() and
1749 * io_should_wake() handle the termination of the loop and only
1750 * wake as many waiters as we need to.
1752 if (wq_has_sleeper(&ctx->cq_wait))
1753 wake_up_all(&ctx->cq_wait);
1754 if (io_should_trigger_evfd(ctx))
1755 eventfd_signal(ctx->cq_ev_fd, 1);
1758 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1760 /* see waitqueue_active() comment */
1763 if (ctx->flags & IORING_SETUP_SQPOLL) {
1764 if (waitqueue_active(&ctx->cq_wait))
1765 wake_up_all(&ctx->cq_wait);
1767 if (io_should_trigger_evfd(ctx))
1768 eventfd_signal(ctx->cq_ev_fd, 1);
1771 /* Returns true if there are no backlogged entries after the flush */
1772 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1774 bool all_flushed, posted;
1776 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1780 spin_lock(&ctx->completion_lock);
1781 while (!list_empty(&ctx->cq_overflow_list)) {
1782 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1783 struct io_overflow_cqe *ocqe;
1787 ocqe = list_first_entry(&ctx->cq_overflow_list,
1788 struct io_overflow_cqe, list);
1790 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1792 io_account_cq_overflow(ctx);
1795 list_del(&ocqe->list);
1799 all_flushed = list_empty(&ctx->cq_overflow_list);
1801 clear_bit(0, &ctx->check_cq_overflow);
1802 WRITE_ONCE(ctx->rings->sq_flags,
1803 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1807 io_commit_cqring(ctx);
1808 spin_unlock(&ctx->completion_lock);
1810 io_cqring_ev_posted(ctx);
1814 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1818 if (test_bit(0, &ctx->check_cq_overflow)) {
1819 /* iopoll syncs against uring_lock, not completion_lock */
1820 if (ctx->flags & IORING_SETUP_IOPOLL)
1821 mutex_lock(&ctx->uring_lock);
1822 ret = __io_cqring_overflow_flush(ctx, false);
1823 if (ctx->flags & IORING_SETUP_IOPOLL)
1824 mutex_unlock(&ctx->uring_lock);
1830 /* must to be called somewhat shortly after putting a request */
1831 static inline void io_put_task(struct task_struct *task, int nr)
1833 struct io_uring_task *tctx = task->io_uring;
1835 if (likely(task == current)) {
1836 tctx->cached_refs += nr;
1838 percpu_counter_sub(&tctx->inflight, nr);
1839 if (unlikely(atomic_read(&tctx->in_idle)))
1840 wake_up(&tctx->wait);
1841 put_task_struct_many(task, nr);
1845 static void io_task_refs_refill(struct io_uring_task *tctx)
1847 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
1849 percpu_counter_add(&tctx->inflight, refill);
1850 refcount_add(refill, ¤t->usage);
1851 tctx->cached_refs += refill;
1854 static inline void io_get_task_refs(int nr)
1856 struct io_uring_task *tctx = current->io_uring;
1858 tctx->cached_refs -= nr;
1859 if (unlikely(tctx->cached_refs < 0))
1860 io_task_refs_refill(tctx);
1863 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
1865 struct io_uring_task *tctx = task->io_uring;
1866 unsigned int refs = tctx->cached_refs;
1869 tctx->cached_refs = 0;
1870 percpu_counter_sub(&tctx->inflight, refs);
1871 put_task_struct_many(task, refs);
1875 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1876 s32 res, u32 cflags)
1878 struct io_overflow_cqe *ocqe;
1880 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1883 * If we're in ring overflow flush mode, or in task cancel mode,
1884 * or cannot allocate an overflow entry, then we need to drop it
1887 io_account_cq_overflow(ctx);
1890 if (list_empty(&ctx->cq_overflow_list)) {
1891 set_bit(0, &ctx->check_cq_overflow);
1892 WRITE_ONCE(ctx->rings->sq_flags,
1893 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1896 ocqe->cqe.user_data = user_data;
1897 ocqe->cqe.res = res;
1898 ocqe->cqe.flags = cflags;
1899 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1903 static inline bool __io_fill_cqe(struct io_ring_ctx *ctx, u64 user_data,
1904 s32 res, u32 cflags)
1906 struct io_uring_cqe *cqe;
1908 trace_io_uring_complete(ctx, user_data, res, cflags);
1911 * If we can't get a cq entry, userspace overflowed the
1912 * submission (by quite a lot). Increment the overflow count in
1915 cqe = io_get_cqe(ctx);
1917 WRITE_ONCE(cqe->user_data, user_data);
1918 WRITE_ONCE(cqe->res, res);
1919 WRITE_ONCE(cqe->flags, cflags);
1922 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1925 static noinline void io_fill_cqe_req(struct io_kiocb *req, s32 res, u32 cflags)
1927 if (!(req->flags & REQ_F_CQE_SKIP))
1928 __io_fill_cqe(req->ctx, req->user_data, res, cflags);
1931 static noinline bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data,
1932 s32 res, u32 cflags)
1935 return __io_fill_cqe(ctx, user_data, res, cflags);
1938 static void __io_req_complete_post(struct io_kiocb *req, s32 res,
1941 struct io_ring_ctx *ctx = req->ctx;
1943 if (!(req->flags & REQ_F_CQE_SKIP))
1944 __io_fill_cqe(ctx, req->user_data, res, cflags);
1946 * If we're the last reference to this request, add to our locked
1949 if (req_ref_put_and_test(req)) {
1950 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1951 if (req->flags & IO_DISARM_MASK)
1952 io_disarm_next(req);
1954 io_req_task_queue(req->link);
1958 io_req_put_rsrc(req, ctx);
1959 io_dismantle_req(req);
1960 io_put_task(req->task, 1);
1961 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
1962 ctx->locked_free_nr++;
1966 static void io_req_complete_post(struct io_kiocb *req, s32 res,
1969 struct io_ring_ctx *ctx = req->ctx;
1971 spin_lock(&ctx->completion_lock);
1972 __io_req_complete_post(req, res, cflags);
1973 io_commit_cqring(ctx);
1974 spin_unlock(&ctx->completion_lock);
1975 io_cqring_ev_posted(ctx);
1978 static inline void io_req_complete_state(struct io_kiocb *req, s32 res,
1982 req->cflags = cflags;
1983 req->flags |= REQ_F_COMPLETE_INLINE;
1986 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1987 s32 res, u32 cflags)
1989 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1990 io_req_complete_state(req, res, cflags);
1992 io_req_complete_post(req, res, cflags);
1995 static inline void io_req_complete(struct io_kiocb *req, s32 res)
1997 __io_req_complete(req, 0, res, 0);
2000 static void io_req_complete_failed(struct io_kiocb *req, s32 res)
2003 io_req_complete_post(req, res, 0);
2006 static void io_req_complete_fail_submit(struct io_kiocb *req)
2009 * We don't submit, fail them all, for that replace hardlinks with
2010 * normal links. Extra REQ_F_LINK is tolerated.
2012 req->flags &= ~REQ_F_HARDLINK;
2013 req->flags |= REQ_F_LINK;
2014 io_req_complete_failed(req, req->result);
2018 * Don't initialise the fields below on every allocation, but do that in
2019 * advance and keep them valid across allocations.
2021 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
2025 req->async_data = NULL;
2026 /* not necessary, but safer to zero */
2030 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
2031 struct io_submit_state *state)
2033 spin_lock(&ctx->completion_lock);
2034 wq_list_splice(&ctx->locked_free_list, &state->free_list);
2035 ctx->locked_free_nr = 0;
2036 spin_unlock(&ctx->completion_lock);
2039 /* Returns true IFF there are requests in the cache */
2040 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
2042 struct io_submit_state *state = &ctx->submit_state;
2045 * If we have more than a batch's worth of requests in our IRQ side
2046 * locked cache, grab the lock and move them over to our submission
2049 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
2050 io_flush_cached_locked_reqs(ctx, state);
2051 return !!state->free_list.next;
2055 * A request might get retired back into the request caches even before opcode
2056 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
2057 * Because of that, io_alloc_req() should be called only under ->uring_lock
2058 * and with extra caution to not get a request that is still worked on.
2060 static __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
2061 __must_hold(&ctx->uring_lock)
2063 struct io_submit_state *state = &ctx->submit_state;
2064 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
2065 void *reqs[IO_REQ_ALLOC_BATCH];
2066 struct io_kiocb *req;
2069 if (likely(state->free_list.next || io_flush_cached_reqs(ctx)))
2072 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
2075 * Bulk alloc is all-or-nothing. If we fail to get a batch,
2076 * retry single alloc to be on the safe side.
2078 if (unlikely(ret <= 0)) {
2079 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
2085 percpu_ref_get_many(&ctx->refs, ret);
2086 for (i = 0; i < ret; i++) {
2089 io_preinit_req(req, ctx);
2090 wq_stack_add_head(&req->comp_list, &state->free_list);
2095 static inline bool io_alloc_req_refill(struct io_ring_ctx *ctx)
2097 if (unlikely(!ctx->submit_state.free_list.next))
2098 return __io_alloc_req_refill(ctx);
2102 static inline struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
2104 struct io_wq_work_node *node;
2106 node = wq_stack_extract(&ctx->submit_state.free_list);
2107 return container_of(node, struct io_kiocb, comp_list);
2110 static inline void io_put_file(struct file *file)
2116 static inline void io_dismantle_req(struct io_kiocb *req)
2118 unsigned int flags = req->flags;
2120 if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
2122 if (!(flags & REQ_F_FIXED_FILE))
2123 io_put_file(req->file);
2126 static __cold void __io_free_req(struct io_kiocb *req)
2128 struct io_ring_ctx *ctx = req->ctx;
2130 io_req_put_rsrc(req, ctx);
2131 io_dismantle_req(req);
2132 io_put_task(req->task, 1);
2134 spin_lock(&ctx->completion_lock);
2135 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
2136 ctx->locked_free_nr++;
2137 spin_unlock(&ctx->completion_lock);
2140 static inline void io_remove_next_linked(struct io_kiocb *req)
2142 struct io_kiocb *nxt = req->link;
2144 req->link = nxt->link;
2148 static bool io_kill_linked_timeout(struct io_kiocb *req)
2149 __must_hold(&req->ctx->completion_lock)
2150 __must_hold(&req->ctx->timeout_lock)
2152 struct io_kiocb *link = req->link;
2154 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2155 struct io_timeout_data *io = link->async_data;
2157 io_remove_next_linked(req);
2158 link->timeout.head = NULL;
2159 if (hrtimer_try_to_cancel(&io->timer) != -1) {
2160 list_del(&link->timeout.list);
2161 /* leave REQ_F_CQE_SKIP to io_fill_cqe_req */
2162 io_fill_cqe_req(link, -ECANCELED, 0);
2163 io_put_req_deferred(link);
2170 static void io_fail_links(struct io_kiocb *req)
2171 __must_hold(&req->ctx->completion_lock)
2173 struct io_kiocb *nxt, *link = req->link;
2174 bool ignore_cqes = req->flags & REQ_F_SKIP_LINK_CQES;
2178 long res = -ECANCELED;
2180 if (link->flags & REQ_F_FAIL)
2186 trace_io_uring_fail_link(req, link);
2188 link->flags &= ~REQ_F_CQE_SKIP;
2189 io_fill_cqe_req(link, res, 0);
2191 io_put_req_deferred(link);
2196 static bool io_disarm_next(struct io_kiocb *req)
2197 __must_hold(&req->ctx->completion_lock)
2199 bool posted = false;
2201 if (req->flags & REQ_F_ARM_LTIMEOUT) {
2202 struct io_kiocb *link = req->link;
2204 req->flags &= ~REQ_F_ARM_LTIMEOUT;
2205 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2206 io_remove_next_linked(req);
2207 /* leave REQ_F_CQE_SKIP to io_fill_cqe_req */
2208 io_fill_cqe_req(link, -ECANCELED, 0);
2209 io_put_req_deferred(link);
2212 } else if (req->flags & REQ_F_LINK_TIMEOUT) {
2213 struct io_ring_ctx *ctx = req->ctx;
2215 spin_lock_irq(&ctx->timeout_lock);
2216 posted = io_kill_linked_timeout(req);
2217 spin_unlock_irq(&ctx->timeout_lock);
2219 if (unlikely((req->flags & REQ_F_FAIL) &&
2220 !(req->flags & REQ_F_HARDLINK))) {
2221 posted |= (req->link != NULL);
2227 static void __io_req_find_next_prep(struct io_kiocb *req)
2229 struct io_ring_ctx *ctx = req->ctx;
2232 spin_lock(&ctx->completion_lock);
2233 posted = io_disarm_next(req);
2235 io_commit_cqring(ctx);
2236 spin_unlock(&ctx->completion_lock);
2238 io_cqring_ev_posted(ctx);
2241 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2243 struct io_kiocb *nxt;
2245 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2248 * If LINK is set, we have dependent requests in this chain. If we
2249 * didn't fail this request, queue the first one up, moving any other
2250 * dependencies to the next request. In case of failure, fail the rest
2253 if (unlikely(req->flags & IO_DISARM_MASK))
2254 __io_req_find_next_prep(req);
2260 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
2265 io_submit_flush_completions(ctx);
2266 mutex_unlock(&ctx->uring_lock);
2269 percpu_ref_put(&ctx->refs);
2272 static inline void ctx_commit_and_unlock(struct io_ring_ctx *ctx)
2274 io_commit_cqring(ctx);
2275 spin_unlock(&ctx->completion_lock);
2276 io_cqring_ev_posted(ctx);
2279 static void handle_prev_tw_list(struct io_wq_work_node *node,
2280 struct io_ring_ctx **ctx, bool *uring_locked)
2282 if (*ctx && !*uring_locked)
2283 spin_lock(&(*ctx)->completion_lock);
2286 struct io_wq_work_node *next = node->next;
2287 struct io_kiocb *req = container_of(node, struct io_kiocb,
2290 if (req->ctx != *ctx) {
2291 if (unlikely(!*uring_locked && *ctx))
2292 ctx_commit_and_unlock(*ctx);
2294 ctx_flush_and_put(*ctx, uring_locked);
2296 /* if not contended, grab and improve batching */
2297 *uring_locked = mutex_trylock(&(*ctx)->uring_lock);
2298 percpu_ref_get(&(*ctx)->refs);
2299 if (unlikely(!*uring_locked))
2300 spin_lock(&(*ctx)->completion_lock);
2302 if (likely(*uring_locked))
2303 req->io_task_work.func(req, uring_locked);
2305 __io_req_complete_post(req, req->result, io_put_kbuf(req));
2309 if (unlikely(!*uring_locked))
2310 ctx_commit_and_unlock(*ctx);
2313 static void handle_tw_list(struct io_wq_work_node *node,
2314 struct io_ring_ctx **ctx, bool *locked)
2317 struct io_wq_work_node *next = node->next;
2318 struct io_kiocb *req = container_of(node, struct io_kiocb,
2321 if (req->ctx != *ctx) {
2322 ctx_flush_and_put(*ctx, locked);
2324 /* if not contended, grab and improve batching */
2325 *locked = mutex_trylock(&(*ctx)->uring_lock);
2326 percpu_ref_get(&(*ctx)->refs);
2328 req->io_task_work.func(req, locked);
2333 static void tctx_task_work(struct callback_head *cb)
2335 bool uring_locked = false;
2336 struct io_ring_ctx *ctx = NULL;
2337 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2341 struct io_wq_work_node *node1, *node2;
2343 if (!tctx->task_list.first &&
2344 !tctx->prior_task_list.first && uring_locked)
2345 io_submit_flush_completions(ctx);
2347 spin_lock_irq(&tctx->task_lock);
2348 node1 = tctx->prior_task_list.first;
2349 node2 = tctx->task_list.first;
2350 INIT_WQ_LIST(&tctx->task_list);
2351 INIT_WQ_LIST(&tctx->prior_task_list);
2352 if (!node2 && !node1)
2353 tctx->task_running = false;
2354 spin_unlock_irq(&tctx->task_lock);
2355 if (!node2 && !node1)
2359 handle_prev_tw_list(node1, &ctx, &uring_locked);
2362 handle_tw_list(node2, &ctx, &uring_locked);
2366 ctx_flush_and_put(ctx, &uring_locked);
2368 /* relaxed read is enough as only the task itself sets ->in_idle */
2369 if (unlikely(atomic_read(&tctx->in_idle)))
2370 io_uring_drop_tctx_refs(current);
2373 static void io_req_task_work_add(struct io_kiocb *req, bool priority)
2375 struct task_struct *tsk = req->task;
2376 struct io_uring_task *tctx = tsk->io_uring;
2377 enum task_work_notify_mode notify;
2378 struct io_wq_work_node *node;
2379 unsigned long flags;
2382 WARN_ON_ONCE(!tctx);
2384 spin_lock_irqsave(&tctx->task_lock, flags);
2386 wq_list_add_tail(&req->io_task_work.node, &tctx->prior_task_list);
2388 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2389 running = tctx->task_running;
2391 tctx->task_running = true;
2392 spin_unlock_irqrestore(&tctx->task_lock, flags);
2394 /* task_work already pending, we're done */
2399 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2400 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2401 * processing task_work. There's no reliable way to tell if TWA_RESUME
2404 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2405 if (likely(!task_work_add(tsk, &tctx->task_work, notify))) {
2406 if (notify == TWA_NONE)
2407 wake_up_process(tsk);
2411 spin_lock_irqsave(&tctx->task_lock, flags);
2412 tctx->task_running = false;
2413 node = wq_list_merge(&tctx->prior_task_list, &tctx->task_list);
2414 spin_unlock_irqrestore(&tctx->task_lock, flags);
2417 req = container_of(node, struct io_kiocb, io_task_work.node);
2419 if (llist_add(&req->io_task_work.fallback_node,
2420 &req->ctx->fallback_llist))
2421 schedule_delayed_work(&req->ctx->fallback_work, 1);
2425 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
2427 struct io_ring_ctx *ctx = req->ctx;
2429 /* not needed for normal modes, but SQPOLL depends on it */
2430 io_tw_lock(ctx, locked);
2431 io_req_complete_failed(req, req->result);
2434 static void io_req_task_submit(struct io_kiocb *req, bool *locked)
2436 struct io_ring_ctx *ctx = req->ctx;
2438 io_tw_lock(ctx, locked);
2439 /* req->task == current here, checking PF_EXITING is safe */
2440 if (likely(!(req->task->flags & PF_EXITING)))
2441 __io_queue_sqe(req);
2443 io_req_complete_failed(req, -EFAULT);
2446 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2449 req->io_task_work.func = io_req_task_cancel;
2450 io_req_task_work_add(req, false);
2453 static void io_req_task_queue(struct io_kiocb *req)
2455 req->io_task_work.func = io_req_task_submit;
2456 io_req_task_work_add(req, false);
2459 static void io_req_task_queue_reissue(struct io_kiocb *req)
2461 req->io_task_work.func = io_queue_async_work;
2462 io_req_task_work_add(req, false);
2465 static inline void io_queue_next(struct io_kiocb *req)
2467 struct io_kiocb *nxt = io_req_find_next(req);
2470 io_req_task_queue(nxt);
2473 static void io_free_req(struct io_kiocb *req)
2479 static void io_free_req_work(struct io_kiocb *req, bool *locked)
2484 static void io_free_batch_list(struct io_ring_ctx *ctx,
2485 struct io_wq_work_node *node)
2486 __must_hold(&ctx->uring_lock)
2488 struct task_struct *task = NULL;
2492 struct io_kiocb *req = container_of(node, struct io_kiocb,
2495 if (unlikely(req->flags & REQ_F_REFCOUNT)) {
2496 node = req->comp_list.next;
2497 if (!req_ref_put_and_test(req))
2501 io_req_put_rsrc_locked(req, ctx);
2503 io_dismantle_req(req);
2505 if (req->task != task) {
2507 io_put_task(task, task_refs);
2512 node = req->comp_list.next;
2513 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
2517 io_put_task(task, task_refs);
2520 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
2521 __must_hold(&ctx->uring_lock)
2523 struct io_wq_work_node *node, *prev;
2524 struct io_submit_state *state = &ctx->submit_state;
2526 if (state->flush_cqes) {
2527 spin_lock(&ctx->completion_lock);
2528 wq_list_for_each(node, prev, &state->compl_reqs) {
2529 struct io_kiocb *req = container_of(node, struct io_kiocb,
2532 if (!(req->flags & REQ_F_CQE_SKIP))
2533 __io_fill_cqe(ctx, req->user_data, req->result,
2537 io_commit_cqring(ctx);
2538 spin_unlock(&ctx->completion_lock);
2539 io_cqring_ev_posted(ctx);
2540 state->flush_cqes = false;
2543 io_free_batch_list(ctx, state->compl_reqs.first);
2544 INIT_WQ_LIST(&state->compl_reqs);
2548 * Drop reference to request, return next in chain (if there is one) if this
2549 * was the last reference to this request.
2551 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2553 struct io_kiocb *nxt = NULL;
2555 if (req_ref_put_and_test(req)) {
2556 nxt = io_req_find_next(req);
2562 static inline void io_put_req(struct io_kiocb *req)
2564 if (req_ref_put_and_test(req))
2568 static inline void io_put_req_deferred(struct io_kiocb *req)
2570 if (req_ref_put_and_test(req)) {
2571 req->io_task_work.func = io_free_req_work;
2572 io_req_task_work_add(req, false);
2576 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2578 /* See comment at the top of this file */
2580 return __io_cqring_events(ctx);
2583 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2585 struct io_rings *rings = ctx->rings;
2587 /* make sure SQ entry isn't read before tail */
2588 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2591 static inline bool io_run_task_work(void)
2593 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2594 __set_current_state(TASK_RUNNING);
2595 tracehook_notify_signal();
2602 static int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin)
2604 struct io_wq_work_node *pos, *start, *prev;
2605 unsigned int poll_flags = BLK_POLL_NOSLEEP;
2606 DEFINE_IO_COMP_BATCH(iob);
2610 * Only spin for completions if we don't have multiple devices hanging
2611 * off our complete list.
2613 if (ctx->poll_multi_queue || force_nonspin)
2614 poll_flags |= BLK_POLL_ONESHOT;
2616 wq_list_for_each(pos, start, &ctx->iopoll_list) {
2617 struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
2618 struct kiocb *kiocb = &req->rw.kiocb;
2622 * Move completed and retryable entries to our local lists.
2623 * If we find a request that requires polling, break out
2624 * and complete those lists first, if we have entries there.
2626 if (READ_ONCE(req->iopoll_completed))
2629 ret = kiocb->ki_filp->f_op->iopoll(kiocb, &iob, poll_flags);
2630 if (unlikely(ret < 0))
2633 poll_flags |= BLK_POLL_ONESHOT;
2635 /* iopoll may have completed current req */
2636 if (!rq_list_empty(iob.req_list) ||
2637 READ_ONCE(req->iopoll_completed))
2641 if (!rq_list_empty(iob.req_list))
2647 wq_list_for_each_resume(pos, prev) {
2648 struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
2650 /* order with io_complete_rw_iopoll(), e.g. ->result updates */
2651 if (!smp_load_acquire(&req->iopoll_completed))
2653 if (unlikely(req->flags & REQ_F_CQE_SKIP))
2656 __io_fill_cqe(ctx, req->user_data, req->result, io_put_kbuf(req));
2660 if (unlikely(!nr_events))
2663 io_commit_cqring(ctx);
2664 io_cqring_ev_posted_iopoll(ctx);
2665 pos = start ? start->next : ctx->iopoll_list.first;
2666 wq_list_cut(&ctx->iopoll_list, prev, start);
2667 io_free_batch_list(ctx, pos);
2672 * We can't just wait for polled events to come to us, we have to actively
2673 * find and complete them.
2675 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2677 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2680 mutex_lock(&ctx->uring_lock);
2681 while (!wq_list_empty(&ctx->iopoll_list)) {
2682 /* let it sleep and repeat later if can't complete a request */
2683 if (io_do_iopoll(ctx, true) == 0)
2686 * Ensure we allow local-to-the-cpu processing to take place,
2687 * in this case we need to ensure that we reap all events.
2688 * Also let task_work, etc. to progress by releasing the mutex
2690 if (need_resched()) {
2691 mutex_unlock(&ctx->uring_lock);
2693 mutex_lock(&ctx->uring_lock);
2696 mutex_unlock(&ctx->uring_lock);
2699 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2701 unsigned int nr_events = 0;
2705 * We disallow the app entering submit/complete with polling, but we
2706 * still need to lock the ring to prevent racing with polled issue
2707 * that got punted to a workqueue.
2709 mutex_lock(&ctx->uring_lock);
2711 * Don't enter poll loop if we already have events pending.
2712 * If we do, we can potentially be spinning for commands that
2713 * already triggered a CQE (eg in error).
2715 if (test_bit(0, &ctx->check_cq_overflow))
2716 __io_cqring_overflow_flush(ctx, false);
2717 if (io_cqring_events(ctx))
2721 * If a submit got punted to a workqueue, we can have the
2722 * application entering polling for a command before it gets
2723 * issued. That app will hold the uring_lock for the duration
2724 * of the poll right here, so we need to take a breather every
2725 * now and then to ensure that the issue has a chance to add
2726 * the poll to the issued list. Otherwise we can spin here
2727 * forever, while the workqueue is stuck trying to acquire the
2730 if (wq_list_empty(&ctx->iopoll_list)) {
2731 u32 tail = ctx->cached_cq_tail;
2733 mutex_unlock(&ctx->uring_lock);
2735 mutex_lock(&ctx->uring_lock);
2737 /* some requests don't go through iopoll_list */
2738 if (tail != ctx->cached_cq_tail ||
2739 wq_list_empty(&ctx->iopoll_list))
2742 ret = io_do_iopoll(ctx, !min);
2747 } while (nr_events < min && !need_resched());
2749 mutex_unlock(&ctx->uring_lock);
2753 static void kiocb_end_write(struct io_kiocb *req)
2756 * Tell lockdep we inherited freeze protection from submission
2759 if (req->flags & REQ_F_ISREG) {
2760 struct super_block *sb = file_inode(req->file)->i_sb;
2762 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2768 static bool io_resubmit_prep(struct io_kiocb *req)
2770 struct io_async_rw *rw = req->async_data;
2772 if (!req_has_async_data(req))
2773 return !io_req_prep_async(req);
2774 iov_iter_restore(&rw->s.iter, &rw->s.iter_state);
2778 static bool io_rw_should_reissue(struct io_kiocb *req)
2780 umode_t mode = file_inode(req->file)->i_mode;
2781 struct io_ring_ctx *ctx = req->ctx;
2783 if (!S_ISBLK(mode) && !S_ISREG(mode))
2785 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2786 !(ctx->flags & IORING_SETUP_IOPOLL)))
2789 * If ref is dying, we might be running poll reap from the exit work.
2790 * Don't attempt to reissue from that path, just let it fail with
2793 if (percpu_ref_is_dying(&ctx->refs))
2796 * Play it safe and assume not safe to re-import and reissue if we're
2797 * not in the original thread group (or in task context).
2799 if (!same_thread_group(req->task, current) || !in_task())
2804 static bool io_resubmit_prep(struct io_kiocb *req)
2808 static bool io_rw_should_reissue(struct io_kiocb *req)
2814 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2816 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2817 kiocb_end_write(req);
2818 if (unlikely(res != req->result)) {
2819 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2820 io_rw_should_reissue(req)) {
2821 req->flags |= REQ_F_REISSUE;
2830 static inline void io_req_task_complete(struct io_kiocb *req, bool *locked)
2832 unsigned int cflags = io_put_kbuf(req);
2833 int res = req->result;
2836 io_req_complete_state(req, res, cflags);
2837 io_req_add_compl_list(req);
2839 io_req_complete_post(req, res, cflags);
2843 static void __io_complete_rw(struct io_kiocb *req, long res,
2844 unsigned int issue_flags)
2846 if (__io_complete_rw_common(req, res))
2848 __io_req_complete(req, issue_flags, req->result, io_put_kbuf(req));
2851 static void io_complete_rw(struct kiocb *kiocb, long res)
2853 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2855 if (__io_complete_rw_common(req, res))
2858 req->io_task_work.func = io_req_task_complete;
2859 io_req_task_work_add(req, !!(req->ctx->flags & IORING_SETUP_SQPOLL));
2862 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res)
2864 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2866 if (kiocb->ki_flags & IOCB_WRITE)
2867 kiocb_end_write(req);
2868 if (unlikely(res != req->result)) {
2869 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2870 req->flags |= REQ_F_REISSUE;
2876 /* order with io_iopoll_complete() checking ->iopoll_completed */
2877 smp_store_release(&req->iopoll_completed, 1);
2881 * After the iocb has been issued, it's safe to be found on the poll list.
2882 * Adding the kiocb to the list AFTER submission ensures that we don't
2883 * find it from a io_do_iopoll() thread before the issuer is done
2884 * accessing the kiocb cookie.
2886 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
2888 struct io_ring_ctx *ctx = req->ctx;
2889 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
2891 /* workqueue context doesn't hold uring_lock, grab it now */
2892 if (unlikely(needs_lock))
2893 mutex_lock(&ctx->uring_lock);
2896 * Track whether we have multiple files in our lists. This will impact
2897 * how we do polling eventually, not spinning if we're on potentially
2898 * different devices.
2900 if (wq_list_empty(&ctx->iopoll_list)) {
2901 ctx->poll_multi_queue = false;
2902 } else if (!ctx->poll_multi_queue) {
2903 struct io_kiocb *list_req;
2905 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
2907 if (list_req->file != req->file)
2908 ctx->poll_multi_queue = true;
2912 * For fast devices, IO may have already completed. If it has, add
2913 * it to the front so we find it first.
2915 if (READ_ONCE(req->iopoll_completed))
2916 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
2918 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
2920 if (unlikely(needs_lock)) {
2922 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2923 * in sq thread task context or in io worker task context. If
2924 * current task context is sq thread, we don't need to check
2925 * whether should wake up sq thread.
2927 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2928 wq_has_sleeper(&ctx->sq_data->wait))
2929 wake_up(&ctx->sq_data->wait);
2931 mutex_unlock(&ctx->uring_lock);
2935 static bool io_bdev_nowait(struct block_device *bdev)
2937 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2941 * If we tracked the file through the SCM inflight mechanism, we could support
2942 * any file. For now, just ensure that anything potentially problematic is done
2945 static bool __io_file_supports_nowait(struct file *file, umode_t mode)
2947 if (S_ISBLK(mode)) {
2948 if (IS_ENABLED(CONFIG_BLOCK) &&
2949 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2955 if (S_ISREG(mode)) {
2956 if (IS_ENABLED(CONFIG_BLOCK) &&
2957 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2958 file->f_op != &io_uring_fops)
2963 /* any ->read/write should understand O_NONBLOCK */
2964 if (file->f_flags & O_NONBLOCK)
2966 return file->f_mode & FMODE_NOWAIT;
2970 * If we tracked the file through the SCM inflight mechanism, we could support
2971 * any file. For now, just ensure that anything potentially problematic is done
2974 static unsigned int io_file_get_flags(struct file *file)
2976 umode_t mode = file_inode(file)->i_mode;
2977 unsigned int res = 0;
2981 if (__io_file_supports_nowait(file, mode))
2986 static inline bool io_file_supports_nowait(struct io_kiocb *req)
2988 return req->flags & REQ_F_SUPPORT_NOWAIT;
2991 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2993 struct io_ring_ctx *ctx = req->ctx;
2994 struct kiocb *kiocb = &req->rw.kiocb;
2995 struct file *file = req->file;
2999 if (!io_req_ffs_set(req))
3000 req->flags |= io_file_get_flags(file) << REQ_F_SUPPORT_NOWAIT_BIT;
3002 kiocb->ki_pos = READ_ONCE(sqe->off);
3003 if (kiocb->ki_pos == -1) {
3004 if (!(file->f_mode & FMODE_STREAM)) {
3005 req->flags |= REQ_F_CUR_POS;
3006 kiocb->ki_pos = file->f_pos;
3011 kiocb->ki_flags = iocb_flags(file);
3012 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
3017 * If the file is marked O_NONBLOCK, still allow retry for it if it
3018 * supports async. Otherwise it's impossible to use O_NONBLOCK files
3019 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
3021 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
3022 ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req)))
3023 req->flags |= REQ_F_NOWAIT;
3025 if (ctx->flags & IORING_SETUP_IOPOLL) {
3026 if (!(kiocb->ki_flags & IOCB_DIRECT) || !file->f_op->iopoll)
3029 kiocb->ki_flags |= IOCB_HIPRI | IOCB_ALLOC_CACHE;
3030 kiocb->ki_complete = io_complete_rw_iopoll;
3031 req->iopoll_completed = 0;
3033 if (kiocb->ki_flags & IOCB_HIPRI)
3035 kiocb->ki_complete = io_complete_rw;
3038 ioprio = READ_ONCE(sqe->ioprio);
3040 ret = ioprio_check_cap(ioprio);
3044 kiocb->ki_ioprio = ioprio;
3046 kiocb->ki_ioprio = get_current_ioprio();
3050 req->rw.addr = READ_ONCE(sqe->addr);
3051 req->rw.len = READ_ONCE(sqe->len);
3052 req->buf_index = READ_ONCE(sqe->buf_index);
3056 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
3062 case -ERESTARTNOINTR:
3063 case -ERESTARTNOHAND:
3064 case -ERESTART_RESTARTBLOCK:
3066 * We can't just restart the syscall, since previously
3067 * submitted sqes may already be in progress. Just fail this
3073 kiocb->ki_complete(kiocb, ret);
3077 static void kiocb_done(struct io_kiocb *req, ssize_t ret,
3078 unsigned int issue_flags)
3080 struct io_async_rw *io = req->async_data;
3082 /* add previously done IO, if any */
3083 if (req_has_async_data(req) && io->bytes_done > 0) {
3085 ret = io->bytes_done;
3087 ret += io->bytes_done;
3090 if (req->flags & REQ_F_CUR_POS)
3091 req->file->f_pos = req->rw.kiocb.ki_pos;
3092 if (ret >= 0 && (req->rw.kiocb.ki_complete == io_complete_rw))
3093 __io_complete_rw(req, ret, issue_flags);
3095 io_rw_done(&req->rw.kiocb, ret);
3097 if (req->flags & REQ_F_REISSUE) {
3098 req->flags &= ~REQ_F_REISSUE;
3099 if (io_resubmit_prep(req)) {
3100 io_req_task_queue_reissue(req);
3104 req->io_task_work.func = io_req_task_complete;
3105 io_req_task_work_add(req, false);
3110 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
3111 struct io_mapped_ubuf *imu)
3113 size_t len = req->rw.len;
3114 u64 buf_end, buf_addr = req->rw.addr;
3117 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
3119 /* not inside the mapped region */
3120 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
3124 * May not be a start of buffer, set size appropriately
3125 * and advance us to the beginning.
3127 offset = buf_addr - imu->ubuf;
3128 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
3132 * Don't use iov_iter_advance() here, as it's really slow for
3133 * using the latter parts of a big fixed buffer - it iterates
3134 * over each segment manually. We can cheat a bit here, because
3137 * 1) it's a BVEC iter, we set it up
3138 * 2) all bvecs are PAGE_SIZE in size, except potentially the
3139 * first and last bvec
3141 * So just find our index, and adjust the iterator afterwards.
3142 * If the offset is within the first bvec (or the whole first
3143 * bvec, just use iov_iter_advance(). This makes it easier
3144 * since we can just skip the first segment, which may not
3145 * be PAGE_SIZE aligned.
3147 const struct bio_vec *bvec = imu->bvec;
3149 if (offset <= bvec->bv_len) {
3150 iov_iter_advance(iter, offset);
3152 unsigned long seg_skip;
3154 /* skip first vec */
3155 offset -= bvec->bv_len;
3156 seg_skip = 1 + (offset >> PAGE_SHIFT);
3158 iter->bvec = bvec + seg_skip;
3159 iter->nr_segs -= seg_skip;
3160 iter->count -= bvec->bv_len + offset;
3161 iter->iov_offset = offset & ~PAGE_MASK;
3168 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
3170 struct io_mapped_ubuf *imu = req->imu;
3171 u16 index, buf_index = req->buf_index;
3174 struct io_ring_ctx *ctx = req->ctx;
3176 if (unlikely(buf_index >= ctx->nr_user_bufs))
3178 io_req_set_rsrc_node(req, ctx);
3179 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
3180 imu = READ_ONCE(ctx->user_bufs[index]);
3183 return __io_import_fixed(req, rw, iter, imu);
3186 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3189 mutex_unlock(&ctx->uring_lock);
3192 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3195 * "Normal" inline submissions always hold the uring_lock, since we
3196 * grab it from the system call. Same is true for the SQPOLL offload.
3197 * The only exception is when we've detached the request and issue it
3198 * from an async worker thread, grab the lock for that case.
3201 mutex_lock(&ctx->uring_lock);
3204 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3205 int bgid, unsigned int issue_flags)
3207 struct io_buffer *kbuf = req->kbuf;
3208 struct io_buffer *head;
3209 bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
3211 if (req->flags & REQ_F_BUFFER_SELECTED)
3214 io_ring_submit_lock(req->ctx, needs_lock);
3216 lockdep_assert_held(&req->ctx->uring_lock);
3218 head = xa_load(&req->ctx->io_buffers, bgid);
3220 if (!list_empty(&head->list)) {
3221 kbuf = list_last_entry(&head->list, struct io_buffer,
3223 list_del(&kbuf->list);
3226 xa_erase(&req->ctx->io_buffers, bgid);
3228 if (*len > kbuf->len)
3230 req->flags |= REQ_F_BUFFER_SELECTED;
3233 kbuf = ERR_PTR(-ENOBUFS);
3236 io_ring_submit_unlock(req->ctx, needs_lock);
3240 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3241 unsigned int issue_flags)
3243 struct io_buffer *kbuf;
3246 bgid = req->buf_index;
3247 kbuf = io_buffer_select(req, len, bgid, issue_flags);
3250 return u64_to_user_ptr(kbuf->addr);
3253 #ifdef CONFIG_COMPAT
3254 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3255 unsigned int issue_flags)
3257 struct compat_iovec __user *uiov;
3258 compat_ssize_t clen;
3262 uiov = u64_to_user_ptr(req->rw.addr);
3263 if (!access_ok(uiov, sizeof(*uiov)))
3265 if (__get_user(clen, &uiov->iov_len))
3271 buf = io_rw_buffer_select(req, &len, issue_flags);
3273 return PTR_ERR(buf);
3274 iov[0].iov_base = buf;
3275 iov[0].iov_len = (compat_size_t) len;
3280 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3281 unsigned int issue_flags)
3283 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3287 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3290 len = iov[0].iov_len;
3293 buf = io_rw_buffer_select(req, &len, issue_flags);
3295 return PTR_ERR(buf);
3296 iov[0].iov_base = buf;
3297 iov[0].iov_len = len;
3301 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3302 unsigned int issue_flags)
3304 if (req->flags & REQ_F_BUFFER_SELECTED) {
3305 struct io_buffer *kbuf = req->kbuf;
3307 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3308 iov[0].iov_len = kbuf->len;
3311 if (req->rw.len != 1)
3314 #ifdef CONFIG_COMPAT
3315 if (req->ctx->compat)
3316 return io_compat_import(req, iov, issue_flags);
3319 return __io_iov_buffer_select(req, iov, issue_flags);
3322 static struct iovec *__io_import_iovec(int rw, struct io_kiocb *req,
3323 struct io_rw_state *s,
3324 unsigned int issue_flags)
3326 struct iov_iter *iter = &s->iter;
3327 u8 opcode = req->opcode;
3328 struct iovec *iovec;
3333 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3334 ret = io_import_fixed(req, rw, iter);
3336 return ERR_PTR(ret);
3340 /* buffer index only valid with fixed read/write, or buffer select */
3341 if (unlikely(req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT)))
3342 return ERR_PTR(-EINVAL);
3344 buf = u64_to_user_ptr(req->rw.addr);
3345 sqe_len = req->rw.len;
3347 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3348 if (req->flags & REQ_F_BUFFER_SELECT) {
3349 buf = io_rw_buffer_select(req, &sqe_len, issue_flags);
3351 return ERR_CAST(buf);
3352 req->rw.len = sqe_len;
3355 ret = import_single_range(rw, buf, sqe_len, s->fast_iov, iter);
3357 return ERR_PTR(ret);
3361 iovec = s->fast_iov;
3362 if (req->flags & REQ_F_BUFFER_SELECT) {
3363 ret = io_iov_buffer_select(req, iovec, issue_flags);
3365 return ERR_PTR(ret);
3366 iov_iter_init(iter, rw, iovec, 1, iovec->iov_len);
3370 ret = __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, &iovec, iter,
3372 if (unlikely(ret < 0))
3373 return ERR_PTR(ret);
3377 static inline int io_import_iovec(int rw, struct io_kiocb *req,
3378 struct iovec **iovec, struct io_rw_state *s,
3379 unsigned int issue_flags)
3381 *iovec = __io_import_iovec(rw, req, s, issue_flags);
3382 if (unlikely(IS_ERR(*iovec)))
3383 return PTR_ERR(*iovec);
3385 iov_iter_save_state(&s->iter, &s->iter_state);
3389 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3391 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3395 * For files that don't have ->read_iter() and ->write_iter(), handle them
3396 * by looping over ->read() or ->write() manually.
3398 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3400 struct kiocb *kiocb = &req->rw.kiocb;
3401 struct file *file = req->file;
3405 * Don't support polled IO through this interface, and we can't
3406 * support non-blocking either. For the latter, this just causes
3407 * the kiocb to be handled from an async context.
3409 if (kiocb->ki_flags & IOCB_HIPRI)
3411 if ((kiocb->ki_flags & IOCB_NOWAIT) &&
3412 !(kiocb->ki_filp->f_flags & O_NONBLOCK))
3415 while (iov_iter_count(iter)) {
3419 if (!iov_iter_is_bvec(iter)) {
3420 iovec = iov_iter_iovec(iter);
3422 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3423 iovec.iov_len = req->rw.len;
3427 nr = file->f_op->read(file, iovec.iov_base,
3428 iovec.iov_len, io_kiocb_ppos(kiocb));
3430 nr = file->f_op->write(file, iovec.iov_base,
3431 iovec.iov_len, io_kiocb_ppos(kiocb));
3439 if (!iov_iter_is_bvec(iter)) {
3440 iov_iter_advance(iter, nr);
3446 if (nr != iovec.iov_len)
3453 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3454 const struct iovec *fast_iov, struct iov_iter *iter)
3456 struct io_async_rw *rw = req->async_data;
3458 memcpy(&rw->s.iter, iter, sizeof(*iter));
3459 rw->free_iovec = iovec;
3461 /* can only be fixed buffers, no need to do anything */
3462 if (iov_iter_is_bvec(iter))
3465 unsigned iov_off = 0;
3467 rw->s.iter.iov = rw->s.fast_iov;
3468 if (iter->iov != fast_iov) {
3469 iov_off = iter->iov - fast_iov;
3470 rw->s.iter.iov += iov_off;
3472 if (rw->s.fast_iov != fast_iov)
3473 memcpy(rw->s.fast_iov + iov_off, fast_iov + iov_off,
3474 sizeof(struct iovec) * iter->nr_segs);
3476 req->flags |= REQ_F_NEED_CLEANUP;
3480 static inline bool io_alloc_async_data(struct io_kiocb *req)
3482 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3483 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3484 if (req->async_data) {
3485 req->flags |= REQ_F_ASYNC_DATA;
3491 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3492 struct io_rw_state *s, bool force)
3494 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3496 if (!req_has_async_data(req)) {
3497 struct io_async_rw *iorw;
3499 if (io_alloc_async_data(req)) {
3504 io_req_map_rw(req, iovec, s->fast_iov, &s->iter);
3505 iorw = req->async_data;
3506 /* we've copied and mapped the iter, ensure state is saved */
3507 iov_iter_save_state(&iorw->s.iter, &iorw->s.iter_state);
3512 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3514 struct io_async_rw *iorw = req->async_data;
3518 /* submission path, ->uring_lock should already be taken */
3519 ret = io_import_iovec(rw, req, &iov, &iorw->s, 0);
3520 if (unlikely(ret < 0))
3523 iorw->bytes_done = 0;
3524 iorw->free_iovec = iov;
3526 req->flags |= REQ_F_NEED_CLEANUP;
3530 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3532 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3534 return io_prep_rw(req, sqe);
3538 * This is our waitqueue callback handler, registered through __folio_lock_async()
3539 * when we initially tried to do the IO with the iocb armed our waitqueue.
3540 * This gets called when the page is unlocked, and we generally expect that to
3541 * happen when the page IO is completed and the page is now uptodate. This will
3542 * queue a task_work based retry of the operation, attempting to copy the data
3543 * again. If the latter fails because the page was NOT uptodate, then we will
3544 * do a thread based blocking retry of the operation. That's the unexpected
3547 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3548 int sync, void *arg)
3550 struct wait_page_queue *wpq;
3551 struct io_kiocb *req = wait->private;
3552 struct wait_page_key *key = arg;
3554 wpq = container_of(wait, struct wait_page_queue, wait);
3556 if (!wake_page_match(wpq, key))
3559 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3560 list_del_init(&wait->entry);
3561 io_req_task_queue(req);
3566 * This controls whether a given IO request should be armed for async page
3567 * based retry. If we return false here, the request is handed to the async
3568 * worker threads for retry. If we're doing buffered reads on a regular file,
3569 * we prepare a private wait_page_queue entry and retry the operation. This
3570 * will either succeed because the page is now uptodate and unlocked, or it
3571 * will register a callback when the page is unlocked at IO completion. Through
3572 * that callback, io_uring uses task_work to setup a retry of the operation.
3573 * That retry will attempt the buffered read again. The retry will generally
3574 * succeed, or in rare cases where it fails, we then fall back to using the
3575 * async worker threads for a blocking retry.
3577 static bool io_rw_should_retry(struct io_kiocb *req)
3579 struct io_async_rw *rw = req->async_data;
3580 struct wait_page_queue *wait = &rw->wpq;
3581 struct kiocb *kiocb = &req->rw.kiocb;
3583 /* never retry for NOWAIT, we just complete with -EAGAIN */
3584 if (req->flags & REQ_F_NOWAIT)
3587 /* Only for buffered IO */
3588 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3592 * just use poll if we can, and don't attempt if the fs doesn't
3593 * support callback based unlocks
3595 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3598 wait->wait.func = io_async_buf_func;
3599 wait->wait.private = req;
3600 wait->wait.flags = 0;
3601 INIT_LIST_HEAD(&wait->wait.entry);
3602 kiocb->ki_flags |= IOCB_WAITQ;
3603 kiocb->ki_flags &= ~IOCB_NOWAIT;
3604 kiocb->ki_waitq = wait;
3608 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3610 if (likely(req->file->f_op->read_iter))
3611 return call_read_iter(req->file, &req->rw.kiocb, iter);
3612 else if (req->file->f_op->read)
3613 return loop_rw_iter(READ, req, iter);
3618 static bool need_read_all(struct io_kiocb *req)
3620 return req->flags & REQ_F_ISREG ||
3621 S_ISBLK(file_inode(req->file)->i_mode);
3624 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3626 struct io_rw_state __s, *s = &__s;
3627 struct iovec *iovec;
3628 struct kiocb *kiocb = &req->rw.kiocb;
3629 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3630 struct io_async_rw *rw;
3633 if (!req_has_async_data(req)) {
3634 ret = io_import_iovec(READ, req, &iovec, s, issue_flags);
3635 if (unlikely(ret < 0))
3638 rw = req->async_data;
3641 * We come here from an earlier attempt, restore our state to
3642 * match in case it doesn't. It's cheap enough that we don't
3643 * need to make this conditional.
3645 iov_iter_restore(&s->iter, &s->iter_state);
3648 req->result = iov_iter_count(&s->iter);
3650 if (force_nonblock) {
3651 /* If the file doesn't support async, just async punt */
3652 if (unlikely(!io_file_supports_nowait(req))) {
3653 ret = io_setup_async_rw(req, iovec, s, true);
3654 return ret ?: -EAGAIN;
3656 kiocb->ki_flags |= IOCB_NOWAIT;
3658 /* Ensure we clear previously set non-block flag */
3659 kiocb->ki_flags &= ~IOCB_NOWAIT;
3662 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), req->result);
3663 if (unlikely(ret)) {
3668 ret = io_iter_do_read(req, &s->iter);
3670 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3671 req->flags &= ~REQ_F_REISSUE;
3672 /* IOPOLL retry should happen for io-wq threads */
3673 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3675 /* no retry on NONBLOCK nor RWF_NOWAIT */
3676 if (req->flags & REQ_F_NOWAIT)
3679 } else if (ret == -EIOCBQUEUED) {
3681 } else if (ret == req->result || ret <= 0 || !force_nonblock ||
3682 (req->flags & REQ_F_NOWAIT) || !need_read_all(req)) {
3683 /* read all, failed, already did sync or don't want to retry */
3688 * Don't depend on the iter state matching what was consumed, or being
3689 * untouched in case of error. Restore it and we'll advance it
3690 * manually if we need to.
3692 iov_iter_restore(&s->iter, &s->iter_state);
3694 ret2 = io_setup_async_rw(req, iovec, s, true);
3699 rw = req->async_data;
3702 * Now use our persistent iterator and state, if we aren't already.
3703 * We've restored and mapped the iter to match.
3708 * We end up here because of a partial read, either from
3709 * above or inside this loop. Advance the iter by the bytes
3710 * that were consumed.
3712 iov_iter_advance(&s->iter, ret);
3713 if (!iov_iter_count(&s->iter))
3715 rw->bytes_done += ret;
3716 iov_iter_save_state(&s->iter, &s->iter_state);
3718 /* if we can retry, do so with the callbacks armed */
3719 if (!io_rw_should_retry(req)) {
3720 kiocb->ki_flags &= ~IOCB_WAITQ;
3725 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3726 * we get -EIOCBQUEUED, then we'll get a notification when the
3727 * desired page gets unlocked. We can also get a partial read
3728 * here, and if we do, then just retry at the new offset.
3730 ret = io_iter_do_read(req, &s->iter);
3731 if (ret == -EIOCBQUEUED)
3733 /* we got some bytes, but not all. retry. */
3734 kiocb->ki_flags &= ~IOCB_WAITQ;
3735 iov_iter_restore(&s->iter, &s->iter_state);
3738 kiocb_done(req, ret, issue_flags);
3740 /* it's faster to check here then delegate to kfree */
3746 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3748 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3750 req->rw.kiocb.ki_hint = ki_hint_validate(file_write_hint(req->file));
3751 return io_prep_rw(req, sqe);
3754 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3756 struct io_rw_state __s, *s = &__s;
3757 struct iovec *iovec;
3758 struct kiocb *kiocb = &req->rw.kiocb;
3759 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3762 if (!req_has_async_data(req)) {
3763 ret = io_import_iovec(WRITE, req, &iovec, s, issue_flags);
3764 if (unlikely(ret < 0))
3767 struct io_async_rw *rw = req->async_data;
3770 iov_iter_restore(&s->iter, &s->iter_state);
3773 req->result = iov_iter_count(&s->iter);
3775 if (force_nonblock) {
3776 /* If the file doesn't support async, just async punt */
3777 if (unlikely(!io_file_supports_nowait(req)))
3780 /* file path doesn't support NOWAIT for non-direct_IO */
3781 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3782 (req->flags & REQ_F_ISREG))
3785 kiocb->ki_flags |= IOCB_NOWAIT;
3787 /* Ensure we clear previously set non-block flag */
3788 kiocb->ki_flags &= ~IOCB_NOWAIT;
3791 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), req->result);
3796 * Open-code file_start_write here to grab freeze protection,
3797 * which will be released by another thread in
3798 * io_complete_rw(). Fool lockdep by telling it the lock got
3799 * released so that it doesn't complain about the held lock when
3800 * we return to userspace.
3802 if (req->flags & REQ_F_ISREG) {
3803 sb_start_write(file_inode(req->file)->i_sb);
3804 __sb_writers_release(file_inode(req->file)->i_sb,
3807 kiocb->ki_flags |= IOCB_WRITE;
3809 if (likely(req->file->f_op->write_iter))
3810 ret2 = call_write_iter(req->file, kiocb, &s->iter);
3811 else if (req->file->f_op->write)
3812 ret2 = loop_rw_iter(WRITE, req, &s->iter);
3816 if (req->flags & REQ_F_REISSUE) {
3817 req->flags &= ~REQ_F_REISSUE;
3822 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3823 * retry them without IOCB_NOWAIT.
3825 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3827 /* no retry on NONBLOCK nor RWF_NOWAIT */
3828 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3830 if (!force_nonblock || ret2 != -EAGAIN) {
3831 /* IOPOLL retry should happen for io-wq threads */
3832 if (ret2 == -EAGAIN && (req->ctx->flags & IORING_SETUP_IOPOLL))
3835 kiocb_done(req, ret2, issue_flags);
3838 iov_iter_restore(&s->iter, &s->iter_state);
3839 ret = io_setup_async_rw(req, iovec, s, false);
3840 return ret ?: -EAGAIN;
3843 /* it's reportedly faster than delegating the null check to kfree() */
3849 static int io_renameat_prep(struct io_kiocb *req,
3850 const struct io_uring_sqe *sqe)
3852 struct io_rename *ren = &req->rename;
3853 const char __user *oldf, *newf;
3855 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3857 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
3859 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3862 ren->old_dfd = READ_ONCE(sqe->fd);
3863 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3864 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3865 ren->new_dfd = READ_ONCE(sqe->len);
3866 ren->flags = READ_ONCE(sqe->rename_flags);
3868 ren->oldpath = getname(oldf);
3869 if (IS_ERR(ren->oldpath))
3870 return PTR_ERR(ren->oldpath);
3872 ren->newpath = getname(newf);
3873 if (IS_ERR(ren->newpath)) {
3874 putname(ren->oldpath);
3875 return PTR_ERR(ren->newpath);
3878 req->flags |= REQ_F_NEED_CLEANUP;
3882 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3884 struct io_rename *ren = &req->rename;
3887 if (issue_flags & IO_URING_F_NONBLOCK)
3890 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3891 ren->newpath, ren->flags);
3893 req->flags &= ~REQ_F_NEED_CLEANUP;
3896 io_req_complete(req, ret);
3900 static int io_unlinkat_prep(struct io_kiocb *req,
3901 const struct io_uring_sqe *sqe)
3903 struct io_unlink *un = &req->unlink;
3904 const char __user *fname;
3906 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3908 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3911 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3914 un->dfd = READ_ONCE(sqe->fd);
3916 un->flags = READ_ONCE(sqe->unlink_flags);
3917 if (un->flags & ~AT_REMOVEDIR)
3920 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3921 un->filename = getname(fname);
3922 if (IS_ERR(un->filename))
3923 return PTR_ERR(un->filename);
3925 req->flags |= REQ_F_NEED_CLEANUP;
3929 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3931 struct io_unlink *un = &req->unlink;
3934 if (issue_flags & IO_URING_F_NONBLOCK)
3937 if (un->flags & AT_REMOVEDIR)
3938 ret = do_rmdir(un->dfd, un->filename);
3940 ret = do_unlinkat(un->dfd, un->filename);
3942 req->flags &= ~REQ_F_NEED_CLEANUP;
3945 io_req_complete(req, ret);
3949 static int io_mkdirat_prep(struct io_kiocb *req,
3950 const struct io_uring_sqe *sqe)
3952 struct io_mkdir *mkd = &req->mkdir;
3953 const char __user *fname;
3955 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3957 if (sqe->ioprio || sqe->off || sqe->rw_flags || sqe->buf_index ||
3960 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3963 mkd->dfd = READ_ONCE(sqe->fd);
3964 mkd->mode = READ_ONCE(sqe->len);
3966 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3967 mkd->filename = getname(fname);
3968 if (IS_ERR(mkd->filename))
3969 return PTR_ERR(mkd->filename);
3971 req->flags |= REQ_F_NEED_CLEANUP;
3975 static int io_mkdirat(struct io_kiocb *req, unsigned int issue_flags)
3977 struct io_mkdir *mkd = &req->mkdir;
3980 if (issue_flags & IO_URING_F_NONBLOCK)
3983 ret = do_mkdirat(mkd->dfd, mkd->filename, mkd->mode);
3985 req->flags &= ~REQ_F_NEED_CLEANUP;
3988 io_req_complete(req, ret);
3992 static int io_symlinkat_prep(struct io_kiocb *req,
3993 const struct io_uring_sqe *sqe)
3995 struct io_symlink *sl = &req->symlink;
3996 const char __user *oldpath, *newpath;
3998 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4000 if (sqe->ioprio || sqe->len || sqe->rw_flags || sqe->buf_index ||
4003 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4006 sl->new_dfd = READ_ONCE(sqe->fd);
4007 oldpath = u64_to_user_ptr(READ_ONCE(sqe->addr));
4008 newpath = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4010 sl->oldpath = getname(oldpath);
4011 if (IS_ERR(sl->oldpath))
4012 return PTR_ERR(sl->oldpath);
4014 sl->newpath = getname(newpath);
4015 if (IS_ERR(sl->newpath)) {
4016 putname(sl->oldpath);
4017 return PTR_ERR(sl->newpath);
4020 req->flags |= REQ_F_NEED_CLEANUP;
4024 static int io_symlinkat(struct io_kiocb *req, unsigned int issue_flags)
4026 struct io_symlink *sl = &req->symlink;
4029 if (issue_flags & IO_URING_F_NONBLOCK)
4032 ret = do_symlinkat(sl->oldpath, sl->new_dfd, sl->newpath);
4034 req->flags &= ~REQ_F_NEED_CLEANUP;
4037 io_req_complete(req, ret);
4041 static int io_linkat_prep(struct io_kiocb *req,
4042 const struct io_uring_sqe *sqe)
4044 struct io_hardlink *lnk = &req->hardlink;
4045 const char __user *oldf, *newf;
4047 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4049 if (sqe->ioprio || sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
4051 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4054 lnk->old_dfd = READ_ONCE(sqe->fd);
4055 lnk->new_dfd = READ_ONCE(sqe->len);
4056 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
4057 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4058 lnk->flags = READ_ONCE(sqe->hardlink_flags);
4060 lnk->oldpath = getname(oldf);
4061 if (IS_ERR(lnk->oldpath))
4062 return PTR_ERR(lnk->oldpath);
4064 lnk->newpath = getname(newf);
4065 if (IS_ERR(lnk->newpath)) {
4066 putname(lnk->oldpath);
4067 return PTR_ERR(lnk->newpath);
4070 req->flags |= REQ_F_NEED_CLEANUP;
4074 static int io_linkat(struct io_kiocb *req, unsigned int issue_flags)
4076 struct io_hardlink *lnk = &req->hardlink;
4079 if (issue_flags & IO_URING_F_NONBLOCK)
4082 ret = do_linkat(lnk->old_dfd, lnk->oldpath, lnk->new_dfd,
4083 lnk->newpath, lnk->flags);
4085 req->flags &= ~REQ_F_NEED_CLEANUP;
4088 io_req_complete(req, ret);
4092 static int io_shutdown_prep(struct io_kiocb *req,
4093 const struct io_uring_sqe *sqe)
4095 #if defined(CONFIG_NET)
4096 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4098 if (unlikely(sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
4099 sqe->buf_index || sqe->splice_fd_in))
4102 req->shutdown.how = READ_ONCE(sqe->len);
4109 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
4111 #if defined(CONFIG_NET)
4112 struct socket *sock;
4115 if (issue_flags & IO_URING_F_NONBLOCK)
4118 sock = sock_from_file(req->file);
4119 if (unlikely(!sock))
4122 ret = __sys_shutdown_sock(sock, req->shutdown.how);
4125 io_req_complete(req, ret);
4132 static int __io_splice_prep(struct io_kiocb *req,
4133 const struct io_uring_sqe *sqe)
4135 struct io_splice *sp = &req->splice;
4136 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
4138 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4142 sp->len = READ_ONCE(sqe->len);
4143 sp->flags = READ_ONCE(sqe->splice_flags);
4145 if (unlikely(sp->flags & ~valid_flags))
4148 sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
4149 (sp->flags & SPLICE_F_FD_IN_FIXED));
4152 req->flags |= REQ_F_NEED_CLEANUP;
4156 static int io_tee_prep(struct io_kiocb *req,
4157 const struct io_uring_sqe *sqe)
4159 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
4161 return __io_splice_prep(req, sqe);
4164 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
4166 struct io_splice *sp = &req->splice;
4167 struct file *in = sp->file_in;
4168 struct file *out = sp->file_out;
4169 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4172 if (issue_flags & IO_URING_F_NONBLOCK)
4175 ret = do_tee(in, out, sp->len, flags);
4177 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4179 req->flags &= ~REQ_F_NEED_CLEANUP;
4183 io_req_complete(req, ret);
4187 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4189 struct io_splice *sp = &req->splice;
4191 sp->off_in = READ_ONCE(sqe->splice_off_in);
4192 sp->off_out = READ_ONCE(sqe->off);
4193 return __io_splice_prep(req, sqe);
4196 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
4198 struct io_splice *sp = &req->splice;
4199 struct file *in = sp->file_in;
4200 struct file *out = sp->file_out;
4201 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4202 loff_t *poff_in, *poff_out;
4205 if (issue_flags & IO_URING_F_NONBLOCK)
4208 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
4209 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
4212 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
4214 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4216 req->flags &= ~REQ_F_NEED_CLEANUP;
4220 io_req_complete(req, ret);
4225 * IORING_OP_NOP just posts a completion event, nothing else.
4227 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
4229 struct io_ring_ctx *ctx = req->ctx;
4231 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4234 __io_req_complete(req, issue_flags, 0, 0);
4238 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4240 struct io_ring_ctx *ctx = req->ctx;
4245 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4247 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4251 req->sync.flags = READ_ONCE(sqe->fsync_flags);
4252 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
4255 req->sync.off = READ_ONCE(sqe->off);
4256 req->sync.len = READ_ONCE(sqe->len);
4260 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
4262 loff_t end = req->sync.off + req->sync.len;
4265 /* fsync always requires a blocking context */
4266 if (issue_flags & IO_URING_F_NONBLOCK)
4269 ret = vfs_fsync_range(req->file, req->sync.off,
4270 end > 0 ? end : LLONG_MAX,
4271 req->sync.flags & IORING_FSYNC_DATASYNC);
4274 io_req_complete(req, ret);
4278 static int io_fallocate_prep(struct io_kiocb *req,
4279 const struct io_uring_sqe *sqe)
4281 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags ||
4284 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4287 req->sync.off = READ_ONCE(sqe->off);
4288 req->sync.len = READ_ONCE(sqe->addr);
4289 req->sync.mode = READ_ONCE(sqe->len);
4293 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
4297 /* fallocate always requiring blocking context */
4298 if (issue_flags & IO_URING_F_NONBLOCK)
4300 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
4304 io_req_complete(req, ret);
4308 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4310 const char __user *fname;
4313 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4315 if (unlikely(sqe->ioprio || sqe->buf_index))
4317 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4320 /* open.how should be already initialised */
4321 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4322 req->open.how.flags |= O_LARGEFILE;
4324 req->open.dfd = READ_ONCE(sqe->fd);
4325 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4326 req->open.filename = getname(fname);
4327 if (IS_ERR(req->open.filename)) {
4328 ret = PTR_ERR(req->open.filename);
4329 req->open.filename = NULL;
4333 req->open.file_slot = READ_ONCE(sqe->file_index);
4334 if (req->open.file_slot && (req->open.how.flags & O_CLOEXEC))
4337 req->open.nofile = rlimit(RLIMIT_NOFILE);
4338 req->flags |= REQ_F_NEED_CLEANUP;
4342 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4344 u64 mode = READ_ONCE(sqe->len);
4345 u64 flags = READ_ONCE(sqe->open_flags);
4347 req->open.how = build_open_how(flags, mode);
4348 return __io_openat_prep(req, sqe);
4351 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4353 struct open_how __user *how;
4357 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4358 len = READ_ONCE(sqe->len);
4359 if (len < OPEN_HOW_SIZE_VER0)
4362 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4367 return __io_openat_prep(req, sqe);
4370 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
4372 struct open_flags op;
4374 bool resolve_nonblock, nonblock_set;
4375 bool fixed = !!req->open.file_slot;
4378 ret = build_open_flags(&req->open.how, &op);
4381 nonblock_set = op.open_flag & O_NONBLOCK;
4382 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
4383 if (issue_flags & IO_URING_F_NONBLOCK) {
4385 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4386 * it'll always -EAGAIN
4388 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
4390 op.lookup_flags |= LOOKUP_CACHED;
4391 op.open_flag |= O_NONBLOCK;
4395 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4400 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4403 * We could hang on to this 'fd' on retrying, but seems like
4404 * marginal gain for something that is now known to be a slower
4405 * path. So just put it, and we'll get a new one when we retry.
4410 ret = PTR_ERR(file);
4411 /* only retry if RESOLVE_CACHED wasn't already set by application */
4412 if (ret == -EAGAIN &&
4413 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
4418 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
4419 file->f_flags &= ~O_NONBLOCK;
4420 fsnotify_open(file);
4423 fd_install(ret, file);
4425 ret = io_install_fixed_file(req, file, issue_flags,
4426 req->open.file_slot - 1);
4428 putname(req->open.filename);
4429 req->flags &= ~REQ_F_NEED_CLEANUP;
4432 __io_req_complete(req, issue_flags, ret, 0);
4436 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
4438 return io_openat2(req, issue_flags);
4441 static int io_remove_buffers_prep(struct io_kiocb *req,
4442 const struct io_uring_sqe *sqe)
4444 struct io_provide_buf *p = &req->pbuf;
4447 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
4451 tmp = READ_ONCE(sqe->fd);
4452 if (!tmp || tmp > USHRT_MAX)
4455 memset(p, 0, sizeof(*p));
4457 p->bgid = READ_ONCE(sqe->buf_group);
4461 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4462 int bgid, unsigned nbufs)
4466 /* shouldn't happen */
4470 /* the head kbuf is the list itself */
4471 while (!list_empty(&buf->list)) {
4472 struct io_buffer *nxt;
4474 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4475 list_del(&nxt->list);
4483 xa_erase(&ctx->io_buffers, bgid);
4488 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
4490 struct io_provide_buf *p = &req->pbuf;
4491 struct io_ring_ctx *ctx = req->ctx;
4492 struct io_buffer *head;
4494 bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
4496 io_ring_submit_lock(ctx, needs_lock);
4498 lockdep_assert_held(&ctx->uring_lock);
4501 head = xa_load(&ctx->io_buffers, p->bgid);
4503 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4507 /* complete before unlock, IOPOLL may need the lock */
4508 __io_req_complete(req, issue_flags, ret, 0);
4509 io_ring_submit_unlock(ctx, needs_lock);
4513 static int io_provide_buffers_prep(struct io_kiocb *req,
4514 const struct io_uring_sqe *sqe)
4516 unsigned long size, tmp_check;
4517 struct io_provide_buf *p = &req->pbuf;
4520 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
4523 tmp = READ_ONCE(sqe->fd);
4524 if (!tmp || tmp > USHRT_MAX)
4527 p->addr = READ_ONCE(sqe->addr);
4528 p->len = READ_ONCE(sqe->len);
4530 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4533 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4536 size = (unsigned long)p->len * p->nbufs;
4537 if (!access_ok(u64_to_user_ptr(p->addr), size))
4540 p->bgid = READ_ONCE(sqe->buf_group);
4541 tmp = READ_ONCE(sqe->off);
4542 if (tmp > USHRT_MAX)
4548 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4550 struct io_buffer *buf;
4551 u64 addr = pbuf->addr;
4552 int i, bid = pbuf->bid;
4554 for (i = 0; i < pbuf->nbufs; i++) {
4555 buf = kmalloc(sizeof(*buf), GFP_KERNEL_ACCOUNT);
4560 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4565 INIT_LIST_HEAD(&buf->list);
4568 list_add_tail(&buf->list, &(*head)->list);
4572 return i ? i : -ENOMEM;
4575 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4577 struct io_provide_buf *p = &req->pbuf;
4578 struct io_ring_ctx *ctx = req->ctx;
4579 struct io_buffer *head, *list;
4581 bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
4583 io_ring_submit_lock(ctx, needs_lock);
4585 lockdep_assert_held(&ctx->uring_lock);
4587 list = head = xa_load(&ctx->io_buffers, p->bgid);
4589 ret = io_add_buffers(p, &head);
4590 if (ret >= 0 && !list) {
4591 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4593 __io_remove_buffers(ctx, head, p->bgid, -1U);
4597 /* complete before unlock, IOPOLL may need the lock */
4598 __io_req_complete(req, issue_flags, ret, 0);
4599 io_ring_submit_unlock(ctx, needs_lock);
4603 static int io_epoll_ctl_prep(struct io_kiocb *req,
4604 const struct io_uring_sqe *sqe)
4606 #if defined(CONFIG_EPOLL)
4607 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4609 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4612 req->epoll.epfd = READ_ONCE(sqe->fd);
4613 req->epoll.op = READ_ONCE(sqe->len);
4614 req->epoll.fd = READ_ONCE(sqe->off);
4616 if (ep_op_has_event(req->epoll.op)) {
4617 struct epoll_event __user *ev;
4619 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4620 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4630 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4632 #if defined(CONFIG_EPOLL)
4633 struct io_epoll *ie = &req->epoll;
4635 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4637 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4638 if (force_nonblock && ret == -EAGAIN)
4643 __io_req_complete(req, issue_flags, ret, 0);
4650 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4652 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4653 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->splice_fd_in)
4655 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4658 req->madvise.addr = READ_ONCE(sqe->addr);
4659 req->madvise.len = READ_ONCE(sqe->len);
4660 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4667 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4669 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4670 struct io_madvise *ma = &req->madvise;
4673 if (issue_flags & IO_URING_F_NONBLOCK)
4676 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4679 io_req_complete(req, ret);
4686 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4688 if (sqe->ioprio || sqe->buf_index || sqe->addr || sqe->splice_fd_in)
4690 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4693 req->fadvise.offset = READ_ONCE(sqe->off);
4694 req->fadvise.len = READ_ONCE(sqe->len);
4695 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4699 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4701 struct io_fadvise *fa = &req->fadvise;
4704 if (issue_flags & IO_URING_F_NONBLOCK) {
4705 switch (fa->advice) {
4706 case POSIX_FADV_NORMAL:
4707 case POSIX_FADV_RANDOM:
4708 case POSIX_FADV_SEQUENTIAL:
4715 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4718 __io_req_complete(req, issue_flags, ret, 0);
4722 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4724 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4726 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4728 if (req->flags & REQ_F_FIXED_FILE)
4731 req->statx.dfd = READ_ONCE(sqe->fd);
4732 req->statx.mask = READ_ONCE(sqe->len);
4733 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4734 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4735 req->statx.flags = READ_ONCE(sqe->statx_flags);
4740 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4742 struct io_statx *ctx = &req->statx;
4745 if (issue_flags & IO_URING_F_NONBLOCK)
4748 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4753 io_req_complete(req, ret);
4757 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4759 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4761 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4762 sqe->rw_flags || sqe->buf_index)
4764 if (req->flags & REQ_F_FIXED_FILE)
4767 req->close.fd = READ_ONCE(sqe->fd);
4768 req->close.file_slot = READ_ONCE(sqe->file_index);
4769 if (req->close.file_slot && req->close.fd)
4775 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4777 struct files_struct *files = current->files;
4778 struct io_close *close = &req->close;
4779 struct fdtable *fdt;
4780 struct file *file = NULL;
4783 if (req->close.file_slot) {
4784 ret = io_close_fixed(req, issue_flags);
4788 spin_lock(&files->file_lock);
4789 fdt = files_fdtable(files);
4790 if (close->fd >= fdt->max_fds) {
4791 spin_unlock(&files->file_lock);
4794 file = fdt->fd[close->fd];
4795 if (!file || file->f_op == &io_uring_fops) {
4796 spin_unlock(&files->file_lock);
4801 /* if the file has a flush method, be safe and punt to async */
4802 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4803 spin_unlock(&files->file_lock);
4807 ret = __close_fd_get_file(close->fd, &file);
4808 spin_unlock(&files->file_lock);
4815 /* No ->flush() or already async, safely close from here */
4816 ret = filp_close(file, current->files);
4822 __io_req_complete(req, issue_flags, ret, 0);
4826 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4828 struct io_ring_ctx *ctx = req->ctx;
4830 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4832 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4836 req->sync.off = READ_ONCE(sqe->off);
4837 req->sync.len = READ_ONCE(sqe->len);
4838 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4842 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4846 /* sync_file_range always requires a blocking context */
4847 if (issue_flags & IO_URING_F_NONBLOCK)
4850 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4854 io_req_complete(req, ret);
4858 #if defined(CONFIG_NET)
4859 static int io_setup_async_msg(struct io_kiocb *req,
4860 struct io_async_msghdr *kmsg)
4862 struct io_async_msghdr *async_msg = req->async_data;
4866 if (io_alloc_async_data(req)) {
4867 kfree(kmsg->free_iov);
4870 async_msg = req->async_data;
4871 req->flags |= REQ_F_NEED_CLEANUP;
4872 memcpy(async_msg, kmsg, sizeof(*kmsg));
4873 async_msg->msg.msg_name = &async_msg->addr;
4874 /* if were using fast_iov, set it to the new one */
4875 if (!async_msg->free_iov)
4876 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4881 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4882 struct io_async_msghdr *iomsg)
4884 iomsg->msg.msg_name = &iomsg->addr;
4885 iomsg->free_iov = iomsg->fast_iov;
4886 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4887 req->sr_msg.msg_flags, &iomsg->free_iov);
4890 static int io_sendmsg_prep_async(struct io_kiocb *req)
4894 ret = io_sendmsg_copy_hdr(req, req->async_data);
4896 req->flags |= REQ_F_NEED_CLEANUP;
4900 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4902 struct io_sr_msg *sr = &req->sr_msg;
4904 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4907 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4908 sr->len = READ_ONCE(sqe->len);
4909 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4910 if (sr->msg_flags & MSG_DONTWAIT)
4911 req->flags |= REQ_F_NOWAIT;
4913 #ifdef CONFIG_COMPAT
4914 if (req->ctx->compat)
4915 sr->msg_flags |= MSG_CMSG_COMPAT;
4920 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4922 struct io_async_msghdr iomsg, *kmsg;
4923 struct socket *sock;
4928 sock = sock_from_file(req->file);
4929 if (unlikely(!sock))
4932 if (req_has_async_data(req)) {
4933 kmsg = req->async_data;
4935 ret = io_sendmsg_copy_hdr(req, &iomsg);
4941 flags = req->sr_msg.msg_flags;
4942 if (issue_flags & IO_URING_F_NONBLOCK)
4943 flags |= MSG_DONTWAIT;
4944 if (flags & MSG_WAITALL)
4945 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4947 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4949 if (ret < min_ret) {
4950 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
4951 return io_setup_async_msg(req, kmsg);
4952 if (ret == -ERESTARTSYS)
4956 /* fast path, check for non-NULL to avoid function call */
4958 kfree(kmsg->free_iov);
4959 req->flags &= ~REQ_F_NEED_CLEANUP;
4960 __io_req_complete(req, issue_flags, ret, 0);
4964 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4966 struct io_sr_msg *sr = &req->sr_msg;
4969 struct socket *sock;
4974 sock = sock_from_file(req->file);
4975 if (unlikely(!sock))
4978 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4982 msg.msg_name = NULL;
4983 msg.msg_control = NULL;
4984 msg.msg_controllen = 0;
4985 msg.msg_namelen = 0;
4987 flags = req->sr_msg.msg_flags;
4988 if (issue_flags & IO_URING_F_NONBLOCK)
4989 flags |= MSG_DONTWAIT;
4990 if (flags & MSG_WAITALL)
4991 min_ret = iov_iter_count(&msg.msg_iter);
4993 msg.msg_flags = flags;
4994 ret = sock_sendmsg(sock, &msg);
4995 if (ret < min_ret) {
4996 if (ret == -EAGAIN && (issue_flags & IO_URING_F_NONBLOCK))
4998 if (ret == -ERESTARTSYS)
5002 __io_req_complete(req, issue_flags, ret, 0);
5006 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
5007 struct io_async_msghdr *iomsg)
5009 struct io_sr_msg *sr = &req->sr_msg;
5010 struct iovec __user *uiov;
5014 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
5015 &iomsg->uaddr, &uiov, &iov_len);
5019 if (req->flags & REQ_F_BUFFER_SELECT) {
5022 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
5024 sr->len = iomsg->fast_iov[0].iov_len;
5025 iomsg->free_iov = NULL;
5027 iomsg->free_iov = iomsg->fast_iov;
5028 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
5029 &iomsg->free_iov, &iomsg->msg.msg_iter,
5038 #ifdef CONFIG_COMPAT
5039 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
5040 struct io_async_msghdr *iomsg)
5042 struct io_sr_msg *sr = &req->sr_msg;
5043 struct compat_iovec __user *uiov;
5048 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
5053 uiov = compat_ptr(ptr);
5054 if (req->flags & REQ_F_BUFFER_SELECT) {
5055 compat_ssize_t clen;
5059 if (!access_ok(uiov, sizeof(*uiov)))
5061 if (__get_user(clen, &uiov->iov_len))
5066 iomsg->free_iov = NULL;
5068 iomsg->free_iov = iomsg->fast_iov;
5069 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
5070 UIO_FASTIOV, &iomsg->free_iov,
5071 &iomsg->msg.msg_iter, true);
5080 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
5081 struct io_async_msghdr *iomsg)
5083 iomsg->msg.msg_name = &iomsg->addr;
5085 #ifdef CONFIG_COMPAT
5086 if (req->ctx->compat)
5087 return __io_compat_recvmsg_copy_hdr(req, iomsg);
5090 return __io_recvmsg_copy_hdr(req, iomsg);
5093 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
5094 unsigned int issue_flags)
5096 struct io_sr_msg *sr = &req->sr_msg;
5098 return io_buffer_select(req, &sr->len, sr->bgid, issue_flags);
5101 static int io_recvmsg_prep_async(struct io_kiocb *req)
5105 ret = io_recvmsg_copy_hdr(req, req->async_data);
5107 req->flags |= REQ_F_NEED_CLEANUP;
5111 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5113 struct io_sr_msg *sr = &req->sr_msg;
5115 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5118 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
5119 sr->len = READ_ONCE(sqe->len);
5120 sr->bgid = READ_ONCE(sqe->buf_group);
5121 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
5122 if (sr->msg_flags & MSG_DONTWAIT)
5123 req->flags |= REQ_F_NOWAIT;
5125 #ifdef CONFIG_COMPAT
5126 if (req->ctx->compat)
5127 sr->msg_flags |= MSG_CMSG_COMPAT;
5132 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
5134 struct io_async_msghdr iomsg, *kmsg;
5135 struct socket *sock;
5136 struct io_buffer *kbuf;
5138 int ret, min_ret = 0;
5139 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5141 sock = sock_from_file(req->file);
5142 if (unlikely(!sock))
5145 if (req_has_async_data(req)) {
5146 kmsg = req->async_data;
5148 ret = io_recvmsg_copy_hdr(req, &iomsg);
5154 if (req->flags & REQ_F_BUFFER_SELECT) {
5155 kbuf = io_recv_buffer_select(req, issue_flags);
5157 return PTR_ERR(kbuf);
5158 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
5159 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
5160 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
5161 1, req->sr_msg.len);
5164 flags = req->sr_msg.msg_flags;
5166 flags |= MSG_DONTWAIT;
5167 if (flags & MSG_WAITALL)
5168 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
5170 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
5171 kmsg->uaddr, flags);
5172 if (ret < min_ret) {
5173 if (ret == -EAGAIN && force_nonblock)
5174 return io_setup_async_msg(req, kmsg);
5175 if (ret == -ERESTARTSYS)
5178 } else if ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))) {
5182 /* fast path, check for non-NULL to avoid function call */
5184 kfree(kmsg->free_iov);
5185 req->flags &= ~REQ_F_NEED_CLEANUP;
5186 __io_req_complete(req, issue_flags, ret, io_put_kbuf(req));
5190 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
5192 struct io_buffer *kbuf;
5193 struct io_sr_msg *sr = &req->sr_msg;
5195 void __user *buf = sr->buf;
5196 struct socket *sock;
5199 int ret, min_ret = 0;
5200 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5202 sock = sock_from_file(req->file);
5203 if (unlikely(!sock))
5206 if (req->flags & REQ_F_BUFFER_SELECT) {
5207 kbuf = io_recv_buffer_select(req, issue_flags);
5209 return PTR_ERR(kbuf);
5210 buf = u64_to_user_ptr(kbuf->addr);
5213 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
5217 msg.msg_name = NULL;
5218 msg.msg_control = NULL;
5219 msg.msg_controllen = 0;
5220 msg.msg_namelen = 0;
5221 msg.msg_iocb = NULL;
5224 flags = req->sr_msg.msg_flags;
5226 flags |= MSG_DONTWAIT;
5227 if (flags & MSG_WAITALL)
5228 min_ret = iov_iter_count(&msg.msg_iter);
5230 ret = sock_recvmsg(sock, &msg, flags);
5231 if (ret < min_ret) {
5232 if (ret == -EAGAIN && force_nonblock)
5234 if (ret == -ERESTARTSYS)
5237 } else if ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))) {
5241 __io_req_complete(req, issue_flags, ret, io_put_kbuf(req));
5245 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5247 struct io_accept *accept = &req->accept;
5249 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5251 if (sqe->ioprio || sqe->len || sqe->buf_index)
5254 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5255 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
5256 accept->flags = READ_ONCE(sqe->accept_flags);
5257 accept->nofile = rlimit(RLIMIT_NOFILE);
5259 accept->file_slot = READ_ONCE(sqe->file_index);
5260 if (accept->file_slot && ((req->open.how.flags & O_CLOEXEC) ||
5261 (accept->flags & SOCK_CLOEXEC)))
5263 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
5265 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
5266 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
5270 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5272 struct io_accept *accept = &req->accept;
5273 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5274 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
5275 bool fixed = !!accept->file_slot;
5279 if (req->file->f_flags & O_NONBLOCK)
5280 req->flags |= REQ_F_NOWAIT;
5283 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
5284 if (unlikely(fd < 0))
5287 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
5292 ret = PTR_ERR(file);
5293 if (ret == -EAGAIN && force_nonblock)
5295 if (ret == -ERESTARTSYS)
5298 } else if (!fixed) {
5299 fd_install(fd, file);
5302 ret = io_install_fixed_file(req, file, issue_flags,
5303 accept->file_slot - 1);
5305 __io_req_complete(req, issue_flags, ret, 0);
5309 static int io_connect_prep_async(struct io_kiocb *req)
5311 struct io_async_connect *io = req->async_data;
5312 struct io_connect *conn = &req->connect;
5314 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
5317 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5319 struct io_connect *conn = &req->connect;
5321 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5323 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
5327 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5328 conn->addr_len = READ_ONCE(sqe->addr2);
5332 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5334 struct io_async_connect __io, *io;
5335 unsigned file_flags;
5337 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5339 if (req_has_async_data(req)) {
5340 io = req->async_data;
5342 ret = move_addr_to_kernel(req->connect.addr,
5343 req->connect.addr_len,
5350 file_flags = force_nonblock ? O_NONBLOCK : 0;
5352 ret = __sys_connect_file(req->file, &io->address,
5353 req->connect.addr_len, file_flags);
5354 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5355 if (req_has_async_data(req))
5357 if (io_alloc_async_data(req)) {
5361 memcpy(req->async_data, &__io, sizeof(__io));
5364 if (ret == -ERESTARTSYS)
5369 __io_req_complete(req, issue_flags, ret, 0);
5372 #else /* !CONFIG_NET */
5373 #define IO_NETOP_FN(op) \
5374 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5376 return -EOPNOTSUPP; \
5379 #define IO_NETOP_PREP(op) \
5381 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5383 return -EOPNOTSUPP; \
5386 #define IO_NETOP_PREP_ASYNC(op) \
5388 static int io_##op##_prep_async(struct io_kiocb *req) \
5390 return -EOPNOTSUPP; \
5393 IO_NETOP_PREP_ASYNC(sendmsg);
5394 IO_NETOP_PREP_ASYNC(recvmsg);
5395 IO_NETOP_PREP_ASYNC(connect);
5396 IO_NETOP_PREP(accept);
5399 #endif /* CONFIG_NET */
5401 struct io_poll_table {
5402 struct poll_table_struct pt;
5403 struct io_kiocb *req;
5408 #define IO_POLL_CANCEL_FLAG BIT(31)
5409 #define IO_POLL_REF_MASK ((1u << 20)-1)
5412 * If refs part of ->poll_refs (see IO_POLL_REF_MASK) is 0, it's free. We can
5413 * bump it and acquire ownership. It's disallowed to modify requests while not
5414 * owning it, that prevents from races for enqueueing task_work's and b/w
5415 * arming poll and wakeups.
5417 static inline bool io_poll_get_ownership(struct io_kiocb *req)
5419 return !(atomic_fetch_inc(&req->poll_refs) & IO_POLL_REF_MASK);
5422 static void io_poll_mark_cancelled(struct io_kiocb *req)
5424 atomic_or(IO_POLL_CANCEL_FLAG, &req->poll_refs);
5427 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5429 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5430 if (req->opcode == IORING_OP_POLL_ADD)
5431 return req->async_data;
5432 return req->apoll->double_poll;
5435 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5437 if (req->opcode == IORING_OP_POLL_ADD)
5439 return &req->apoll->poll;
5442 static void io_poll_req_insert(struct io_kiocb *req)
5444 struct io_ring_ctx *ctx = req->ctx;
5445 struct hlist_head *list;
5447 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5448 hlist_add_head(&req->hash_node, list);
5451 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5452 wait_queue_func_t wake_func)
5455 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5456 /* mask in events that we always want/need */
5457 poll->events = events | IO_POLL_UNMASK;
5458 INIT_LIST_HEAD(&poll->wait.entry);
5459 init_waitqueue_func_entry(&poll->wait, wake_func);
5462 static inline void io_poll_remove_entry(struct io_poll_iocb *poll)
5464 struct wait_queue_head *head = smp_load_acquire(&poll->head);
5467 spin_lock_irq(&head->lock);
5468 list_del_init(&poll->wait.entry);
5470 spin_unlock_irq(&head->lock);
5474 static void io_poll_remove_entries(struct io_kiocb *req)
5476 struct io_poll_iocb *poll = io_poll_get_single(req);
5477 struct io_poll_iocb *poll_double = io_poll_get_double(req);
5480 * While we hold the waitqueue lock and the waitqueue is nonempty,
5481 * wake_up_pollfree() will wait for us. However, taking the waitqueue
5482 * lock in the first place can race with the waitqueue being freed.
5484 * We solve this as eventpoll does: by taking advantage of the fact that
5485 * all users of wake_up_pollfree() will RCU-delay the actual free. If
5486 * we enter rcu_read_lock() and see that the pointer to the queue is
5487 * non-NULL, we can then lock it without the memory being freed out from
5490 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
5491 * case the caller deletes the entry from the queue, leaving it empty.
5492 * In that case, only RCU prevents the queue memory from being freed.
5495 io_poll_remove_entry(poll);
5497 io_poll_remove_entry(poll_double);
5502 * All poll tw should go through this. Checks for poll events, manages
5503 * references, does rewait, etc.
5505 * Returns a negative error on failure. >0 when no action require, which is
5506 * either spurious wakeup or multishot CQE is served. 0 when it's done with
5507 * the request, then the mask is stored in req->result.
5509 static int io_poll_check_events(struct io_kiocb *req)
5511 struct io_ring_ctx *ctx = req->ctx;
5512 struct io_poll_iocb *poll = io_poll_get_single(req);
5515 /* req->task == current here, checking PF_EXITING is safe */
5516 if (unlikely(req->task->flags & PF_EXITING))
5517 io_poll_mark_cancelled(req);
5520 v = atomic_read(&req->poll_refs);
5522 /* tw handler should be the owner, and so have some references */
5523 if (WARN_ON_ONCE(!(v & IO_POLL_REF_MASK)))
5525 if (v & IO_POLL_CANCEL_FLAG)
5529 struct poll_table_struct pt = { ._key = poll->events };
5531 req->result = vfs_poll(req->file, &pt) & poll->events;
5534 /* multishot, just fill an CQE and proceed */
5535 if (req->result && !(poll->events & EPOLLONESHOT)) {
5536 __poll_t mask = mangle_poll(req->result & poll->events);
5539 spin_lock(&ctx->completion_lock);
5540 filled = io_fill_cqe_aux(ctx, req->user_data, mask,
5542 io_commit_cqring(ctx);
5543 spin_unlock(&ctx->completion_lock);
5544 if (unlikely(!filled))
5546 io_cqring_ev_posted(ctx);
5547 } else if (req->result) {
5552 * Release all references, retry if someone tried to restart
5553 * task_work while we were executing it.
5555 } while (atomic_sub_return(v & IO_POLL_REF_MASK, &req->poll_refs));
5560 static void io_poll_task_func(struct io_kiocb *req, bool *locked)
5562 struct io_ring_ctx *ctx = req->ctx;
5565 ret = io_poll_check_events(req);
5570 req->result = mangle_poll(req->result & req->poll.events);
5576 io_poll_remove_entries(req);
5577 spin_lock(&ctx->completion_lock);
5578 hash_del(&req->hash_node);
5579 __io_req_complete_post(req, req->result, 0);
5580 io_commit_cqring(ctx);
5581 spin_unlock(&ctx->completion_lock);
5582 io_cqring_ev_posted(ctx);
5585 static void io_apoll_task_func(struct io_kiocb *req, bool *locked)
5587 struct io_ring_ctx *ctx = req->ctx;
5590 ret = io_poll_check_events(req);
5594 io_poll_remove_entries(req);
5595 spin_lock(&ctx->completion_lock);
5596 hash_del(&req->hash_node);
5597 spin_unlock(&ctx->completion_lock);
5600 io_req_task_submit(req, locked);
5602 io_req_complete_failed(req, ret);
5605 static void __io_poll_execute(struct io_kiocb *req, int mask)
5608 if (req->opcode == IORING_OP_POLL_ADD)
5609 req->io_task_work.func = io_poll_task_func;
5611 req->io_task_work.func = io_apoll_task_func;
5613 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5614 io_req_task_work_add(req, false);
5617 static inline void io_poll_execute(struct io_kiocb *req, int res)
5619 if (io_poll_get_ownership(req))
5620 __io_poll_execute(req, res);
5623 static void io_poll_cancel_req(struct io_kiocb *req)
5625 io_poll_mark_cancelled(req);
5626 /* kick tw, which should complete the request */
5627 io_poll_execute(req, 0);
5630 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5633 struct io_kiocb *req = wait->private;
5634 struct io_poll_iocb *poll = container_of(wait, struct io_poll_iocb,
5636 __poll_t mask = key_to_poll(key);
5638 if (unlikely(mask & POLLFREE)) {
5639 io_poll_mark_cancelled(req);
5640 /* we have to kick tw in case it's not already */
5641 io_poll_execute(req, 0);
5644 * If the waitqueue is being freed early but someone is already
5645 * holds ownership over it, we have to tear down the request as
5646 * best we can. That means immediately removing the request from
5647 * its waitqueue and preventing all further accesses to the
5648 * waitqueue via the request.
5650 list_del_init(&poll->wait.entry);
5653 * Careful: this *must* be the last step, since as soon
5654 * as req->head is NULL'ed out, the request can be
5655 * completed and freed, since aio_poll_complete_work()
5656 * will no longer need to take the waitqueue lock.
5658 smp_store_release(&poll->head, NULL);
5662 /* for instances that support it check for an event match first */
5663 if (mask && !(mask & poll->events))
5666 if (io_poll_get_ownership(req)) {
5667 /* optional, saves extra locking for removal in tw handler */
5668 if (mask && poll->events & EPOLLONESHOT) {
5669 list_del_init(&poll->wait.entry);
5672 __io_poll_execute(req, mask);
5677 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5678 struct wait_queue_head *head,
5679 struct io_poll_iocb **poll_ptr)
5681 struct io_kiocb *req = pt->req;
5684 * The file being polled uses multiple waitqueues for poll handling
5685 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5688 if (unlikely(pt->nr_entries)) {
5689 struct io_poll_iocb *first = poll;
5691 /* double add on the same waitqueue head, ignore */
5692 if (first->head == head)
5694 /* already have a 2nd entry, fail a third attempt */
5696 if ((*poll_ptr)->head == head)
5698 pt->error = -EINVAL;
5702 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5704 pt->error = -ENOMEM;
5707 io_init_poll_iocb(poll, first->events, first->wait.func);
5709 if (req->opcode == IORING_OP_POLL_ADD)
5710 req->flags |= REQ_F_ASYNC_DATA;
5715 poll->wait.private = req;
5717 if (poll->events & EPOLLEXCLUSIVE)
5718 add_wait_queue_exclusive(head, &poll->wait);
5720 add_wait_queue(head, &poll->wait);
5723 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5724 struct poll_table_struct *p)
5726 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5728 __io_queue_proc(&pt->req->poll, pt, head,
5729 (struct io_poll_iocb **) &pt->req->async_data);
5732 static int __io_arm_poll_handler(struct io_kiocb *req,
5733 struct io_poll_iocb *poll,
5734 struct io_poll_table *ipt, __poll_t mask)
5736 struct io_ring_ctx *ctx = req->ctx;
5739 INIT_HLIST_NODE(&req->hash_node);
5740 io_init_poll_iocb(poll, mask, io_poll_wake);
5741 poll->file = req->file;
5742 poll->wait.private = req;
5744 ipt->pt._key = mask;
5747 ipt->nr_entries = 0;
5750 * Take the ownership to delay any tw execution up until we're done
5751 * with poll arming. see io_poll_get_ownership().
5753 atomic_set(&req->poll_refs, 1);
5754 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5756 if (mask && (poll->events & EPOLLONESHOT)) {
5757 io_poll_remove_entries(req);
5758 /* no one else has access to the req, forget about the ref */
5761 if (!mask && unlikely(ipt->error || !ipt->nr_entries)) {
5762 io_poll_remove_entries(req);
5764 ipt->error = -EINVAL;
5768 spin_lock(&ctx->completion_lock);
5769 io_poll_req_insert(req);
5770 spin_unlock(&ctx->completion_lock);
5773 /* can't multishot if failed, just queue the event we've got */
5774 if (unlikely(ipt->error || !ipt->nr_entries))
5775 poll->events |= EPOLLONESHOT;
5776 __io_poll_execute(req, mask);
5781 * Release ownership. If someone tried to queue a tw while it was
5782 * locked, kick it off for them.
5784 v = atomic_dec_return(&req->poll_refs);
5785 if (unlikely(v & IO_POLL_REF_MASK))
5786 __io_poll_execute(req, 0);
5790 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5791 struct poll_table_struct *p)
5793 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5794 struct async_poll *apoll = pt->req->apoll;
5796 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5805 static int io_arm_poll_handler(struct io_kiocb *req)
5807 const struct io_op_def *def = &io_op_defs[req->opcode];
5808 struct io_ring_ctx *ctx = req->ctx;
5809 struct async_poll *apoll;
5810 struct io_poll_table ipt;
5811 __poll_t mask = EPOLLONESHOT | POLLERR | POLLPRI;
5814 if (!def->pollin && !def->pollout)
5815 return IO_APOLL_ABORTED;
5816 if (!file_can_poll(req->file) || (req->flags & REQ_F_POLLED))
5817 return IO_APOLL_ABORTED;
5820 mask |= POLLIN | POLLRDNORM;
5822 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5823 if ((req->opcode == IORING_OP_RECVMSG) &&
5824 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5827 mask |= POLLOUT | POLLWRNORM;
5830 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5831 if (unlikely(!apoll))
5832 return IO_APOLL_ABORTED;
5833 apoll->double_poll = NULL;
5835 req->flags |= REQ_F_POLLED;
5836 ipt.pt._qproc = io_async_queue_proc;
5838 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask);
5839 if (ret || ipt.error)
5840 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5842 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5843 mask, apoll->poll.events);
5848 * Returns true if we found and killed one or more poll requests
5850 static __cold bool io_poll_remove_all(struct io_ring_ctx *ctx,
5851 struct task_struct *tsk, bool cancel_all)
5853 struct hlist_node *tmp;
5854 struct io_kiocb *req;
5858 spin_lock(&ctx->completion_lock);
5859 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5860 struct hlist_head *list;
5862 list = &ctx->cancel_hash[i];
5863 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5864 if (io_match_task_safe(req, tsk, cancel_all)) {
5865 io_poll_cancel_req(req);
5870 spin_unlock(&ctx->completion_lock);
5874 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5876 __must_hold(&ctx->completion_lock)
5878 struct hlist_head *list;
5879 struct io_kiocb *req;
5881 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5882 hlist_for_each_entry(req, list, hash_node) {
5883 if (sqe_addr != req->user_data)
5885 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5892 static bool io_poll_disarm(struct io_kiocb *req)
5893 __must_hold(&ctx->completion_lock)
5895 if (!io_poll_get_ownership(req))
5897 io_poll_remove_entries(req);
5898 hash_del(&req->hash_node);
5902 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5904 __must_hold(&ctx->completion_lock)
5906 struct io_kiocb *req = io_poll_find(ctx, sqe_addr, poll_only);
5910 io_poll_cancel_req(req);
5914 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5919 events = READ_ONCE(sqe->poll32_events);
5921 events = swahw32(events);
5923 if (!(flags & IORING_POLL_ADD_MULTI))
5924 events |= EPOLLONESHOT;
5925 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5928 static int io_poll_update_prep(struct io_kiocb *req,
5929 const struct io_uring_sqe *sqe)
5931 struct io_poll_update *upd = &req->poll_update;
5934 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5936 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
5938 flags = READ_ONCE(sqe->len);
5939 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5940 IORING_POLL_ADD_MULTI))
5942 /* meaningless without update */
5943 if (flags == IORING_POLL_ADD_MULTI)
5946 upd->old_user_data = READ_ONCE(sqe->addr);
5947 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5948 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5950 upd->new_user_data = READ_ONCE(sqe->off);
5951 if (!upd->update_user_data && upd->new_user_data)
5953 if (upd->update_events)
5954 upd->events = io_poll_parse_events(sqe, flags);
5955 else if (sqe->poll32_events)
5961 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5963 struct io_poll_iocb *poll = &req->poll;
5966 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5968 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5970 flags = READ_ONCE(sqe->len);
5971 if (flags & ~IORING_POLL_ADD_MULTI)
5973 if ((flags & IORING_POLL_ADD_MULTI) && (req->flags & REQ_F_CQE_SKIP))
5976 io_req_set_refcount(req);
5977 poll->events = io_poll_parse_events(sqe, flags);
5981 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5983 struct io_poll_iocb *poll = &req->poll;
5984 struct io_poll_table ipt;
5987 ipt.pt._qproc = io_poll_queue_proc;
5989 ret = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events);
5990 ret = ret ?: ipt.error;
5992 __io_req_complete(req, issue_flags, ret, 0);
5996 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5998 struct io_ring_ctx *ctx = req->ctx;
5999 struct io_kiocb *preq;
6003 spin_lock(&ctx->completion_lock);
6004 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
6005 if (!preq || !io_poll_disarm(preq)) {
6006 spin_unlock(&ctx->completion_lock);
6007 ret = preq ? -EALREADY : -ENOENT;
6010 spin_unlock(&ctx->completion_lock);
6012 if (req->poll_update.update_events || req->poll_update.update_user_data) {
6013 /* only mask one event flags, keep behavior flags */
6014 if (req->poll_update.update_events) {
6015 preq->poll.events &= ~0xffff;
6016 preq->poll.events |= req->poll_update.events & 0xffff;
6017 preq->poll.events |= IO_POLL_UNMASK;
6019 if (req->poll_update.update_user_data)
6020 preq->user_data = req->poll_update.new_user_data;
6022 ret2 = io_poll_add(preq, issue_flags);
6023 /* successfully updated, don't complete poll request */
6029 preq->result = -ECANCELED;
6030 locked = !(issue_flags & IO_URING_F_UNLOCKED);
6031 io_req_task_complete(preq, &locked);
6035 /* complete update request, we're done with it */
6036 __io_req_complete(req, issue_flags, ret, 0);
6040 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
6042 struct io_timeout_data *data = container_of(timer,
6043 struct io_timeout_data, timer);
6044 struct io_kiocb *req = data->req;
6045 struct io_ring_ctx *ctx = req->ctx;
6046 unsigned long flags;
6048 spin_lock_irqsave(&ctx->timeout_lock, flags);
6049 list_del_init(&req->timeout.list);
6050 atomic_set(&req->ctx->cq_timeouts,
6051 atomic_read(&req->ctx->cq_timeouts) + 1);
6052 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6054 if (!(data->flags & IORING_TIMEOUT_ETIME_SUCCESS))
6057 req->result = -ETIME;
6058 req->io_task_work.func = io_req_task_complete;
6059 io_req_task_work_add(req, false);
6060 return HRTIMER_NORESTART;
6063 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
6065 __must_hold(&ctx->timeout_lock)
6067 struct io_timeout_data *io;
6068 struct io_kiocb *req;
6071 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
6072 found = user_data == req->user_data;
6077 return ERR_PTR(-ENOENT);
6079 io = req->async_data;
6080 if (hrtimer_try_to_cancel(&io->timer) == -1)
6081 return ERR_PTR(-EALREADY);
6082 list_del_init(&req->timeout.list);
6086 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
6087 __must_hold(&ctx->completion_lock)
6088 __must_hold(&ctx->timeout_lock)
6090 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
6093 return PTR_ERR(req);
6096 io_fill_cqe_req(req, -ECANCELED, 0);
6097 io_put_req_deferred(req);
6101 static clockid_t io_timeout_get_clock(struct io_timeout_data *data)
6103 switch (data->flags & IORING_TIMEOUT_CLOCK_MASK) {
6104 case IORING_TIMEOUT_BOOTTIME:
6105 return CLOCK_BOOTTIME;
6106 case IORING_TIMEOUT_REALTIME:
6107 return CLOCK_REALTIME;
6109 /* can't happen, vetted at prep time */
6113 return CLOCK_MONOTONIC;
6117 static int io_linked_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6118 struct timespec64 *ts, enum hrtimer_mode mode)
6119 __must_hold(&ctx->timeout_lock)
6121 struct io_timeout_data *io;
6122 struct io_kiocb *req;
6125 list_for_each_entry(req, &ctx->ltimeout_list, timeout.list) {
6126 found = user_data == req->user_data;
6133 io = req->async_data;
6134 if (hrtimer_try_to_cancel(&io->timer) == -1)
6136 hrtimer_init(&io->timer, io_timeout_get_clock(io), mode);
6137 io->timer.function = io_link_timeout_fn;
6138 hrtimer_start(&io->timer, timespec64_to_ktime(*ts), mode);
6142 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6143 struct timespec64 *ts, enum hrtimer_mode mode)
6144 __must_hold(&ctx->timeout_lock)
6146 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
6147 struct io_timeout_data *data;
6150 return PTR_ERR(req);
6152 req->timeout.off = 0; /* noseq */
6153 data = req->async_data;
6154 list_add_tail(&req->timeout.list, &ctx->timeout_list);
6155 hrtimer_init(&data->timer, io_timeout_get_clock(data), mode);
6156 data->timer.function = io_timeout_fn;
6157 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
6161 static int io_timeout_remove_prep(struct io_kiocb *req,
6162 const struct io_uring_sqe *sqe)
6164 struct io_timeout_rem *tr = &req->timeout_rem;
6166 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6168 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6170 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->splice_fd_in)
6173 tr->ltimeout = false;
6174 tr->addr = READ_ONCE(sqe->addr);
6175 tr->flags = READ_ONCE(sqe->timeout_flags);
6176 if (tr->flags & IORING_TIMEOUT_UPDATE_MASK) {
6177 if (hweight32(tr->flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6179 if (tr->flags & IORING_LINK_TIMEOUT_UPDATE)
6180 tr->ltimeout = true;
6181 if (tr->flags & ~(IORING_TIMEOUT_UPDATE_MASK|IORING_TIMEOUT_ABS))
6183 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
6185 if (tr->ts.tv_sec < 0 || tr->ts.tv_nsec < 0)
6187 } else if (tr->flags) {
6188 /* timeout removal doesn't support flags */
6195 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
6197 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
6202 * Remove or update an existing timeout command
6204 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
6206 struct io_timeout_rem *tr = &req->timeout_rem;
6207 struct io_ring_ctx *ctx = req->ctx;
6210 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
6211 spin_lock(&ctx->completion_lock);
6212 spin_lock_irq(&ctx->timeout_lock);
6213 ret = io_timeout_cancel(ctx, tr->addr);
6214 spin_unlock_irq(&ctx->timeout_lock);
6215 spin_unlock(&ctx->completion_lock);
6217 enum hrtimer_mode mode = io_translate_timeout_mode(tr->flags);
6219 spin_lock_irq(&ctx->timeout_lock);
6221 ret = io_linked_timeout_update(ctx, tr->addr, &tr->ts, mode);
6223 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
6224 spin_unlock_irq(&ctx->timeout_lock);
6229 io_req_complete_post(req, ret, 0);
6233 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6234 bool is_timeout_link)
6236 struct io_timeout_data *data;
6238 u32 off = READ_ONCE(sqe->off);
6240 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6242 if (sqe->ioprio || sqe->buf_index || sqe->len != 1 ||
6245 if (off && is_timeout_link)
6247 flags = READ_ONCE(sqe->timeout_flags);
6248 if (flags & ~(IORING_TIMEOUT_ABS | IORING_TIMEOUT_CLOCK_MASK |
6249 IORING_TIMEOUT_ETIME_SUCCESS))
6251 /* more than one clock specified is invalid, obviously */
6252 if (hweight32(flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6255 INIT_LIST_HEAD(&req->timeout.list);
6256 req->timeout.off = off;
6257 if (unlikely(off && !req->ctx->off_timeout_used))
6258 req->ctx->off_timeout_used = true;
6260 if (WARN_ON_ONCE(req_has_async_data(req)))
6262 if (io_alloc_async_data(req))
6265 data = req->async_data;
6267 data->flags = flags;
6269 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
6272 if (data->ts.tv_sec < 0 || data->ts.tv_nsec < 0)
6275 data->mode = io_translate_timeout_mode(flags);
6276 hrtimer_init(&data->timer, io_timeout_get_clock(data), data->mode);
6278 if (is_timeout_link) {
6279 struct io_submit_link *link = &req->ctx->submit_state.link;
6283 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
6285 req->timeout.head = link->last;
6286 link->last->flags |= REQ_F_ARM_LTIMEOUT;
6291 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
6293 struct io_ring_ctx *ctx = req->ctx;
6294 struct io_timeout_data *data = req->async_data;
6295 struct list_head *entry;
6296 u32 tail, off = req->timeout.off;
6298 spin_lock_irq(&ctx->timeout_lock);
6301 * sqe->off holds how many events that need to occur for this
6302 * timeout event to be satisfied. If it isn't set, then this is
6303 * a pure timeout request, sequence isn't used.
6305 if (io_is_timeout_noseq(req)) {
6306 entry = ctx->timeout_list.prev;
6310 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
6311 req->timeout.target_seq = tail + off;
6313 /* Update the last seq here in case io_flush_timeouts() hasn't.
6314 * This is safe because ->completion_lock is held, and submissions
6315 * and completions are never mixed in the same ->completion_lock section.
6317 ctx->cq_last_tm_flush = tail;
6320 * Insertion sort, ensuring the first entry in the list is always
6321 * the one we need first.
6323 list_for_each_prev(entry, &ctx->timeout_list) {
6324 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
6327 if (io_is_timeout_noseq(nxt))
6329 /* nxt.seq is behind @tail, otherwise would've been completed */
6330 if (off >= nxt->timeout.target_seq - tail)
6334 list_add(&req->timeout.list, entry);
6335 data->timer.function = io_timeout_fn;
6336 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
6337 spin_unlock_irq(&ctx->timeout_lock);
6341 struct io_cancel_data {
6342 struct io_ring_ctx *ctx;
6346 static bool io_cancel_cb(struct io_wq_work *work, void *data)
6348 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6349 struct io_cancel_data *cd = data;
6351 return req->ctx == cd->ctx && req->user_data == cd->user_data;
6354 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
6355 struct io_ring_ctx *ctx)
6357 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
6358 enum io_wq_cancel cancel_ret;
6361 if (!tctx || !tctx->io_wq)
6364 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
6365 switch (cancel_ret) {
6366 case IO_WQ_CANCEL_OK:
6369 case IO_WQ_CANCEL_RUNNING:
6372 case IO_WQ_CANCEL_NOTFOUND:
6380 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
6382 struct io_ring_ctx *ctx = req->ctx;
6385 WARN_ON_ONCE(!io_wq_current_is_worker() && req->task != current);
6387 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
6389 * Fall-through even for -EALREADY, as we may have poll armed
6390 * that need unarming.
6395 spin_lock(&ctx->completion_lock);
6396 ret = io_poll_cancel(ctx, sqe_addr, false);
6400 spin_lock_irq(&ctx->timeout_lock);
6401 ret = io_timeout_cancel(ctx, sqe_addr);
6402 spin_unlock_irq(&ctx->timeout_lock);
6404 spin_unlock(&ctx->completion_lock);
6408 static int io_async_cancel_prep(struct io_kiocb *req,
6409 const struct io_uring_sqe *sqe)
6411 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6413 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6415 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
6419 req->cancel.addr = READ_ONCE(sqe->addr);
6423 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
6425 struct io_ring_ctx *ctx = req->ctx;
6426 u64 sqe_addr = req->cancel.addr;
6427 bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
6428 struct io_tctx_node *node;
6431 ret = io_try_cancel_userdata(req, sqe_addr);
6435 /* slow path, try all io-wq's */
6436 io_ring_submit_lock(ctx, needs_lock);
6438 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
6439 struct io_uring_task *tctx = node->task->io_uring;
6441 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
6445 io_ring_submit_unlock(ctx, needs_lock);
6449 io_req_complete_post(req, ret, 0);
6453 static int io_rsrc_update_prep(struct io_kiocb *req,
6454 const struct io_uring_sqe *sqe)
6456 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6458 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
6461 req->rsrc_update.offset = READ_ONCE(sqe->off);
6462 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
6463 if (!req->rsrc_update.nr_args)
6465 req->rsrc_update.arg = READ_ONCE(sqe->addr);
6469 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
6471 struct io_ring_ctx *ctx = req->ctx;
6472 bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
6473 struct io_uring_rsrc_update2 up;
6476 up.offset = req->rsrc_update.offset;
6477 up.data = req->rsrc_update.arg;
6482 io_ring_submit_lock(ctx, needs_lock);
6483 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
6484 &up, req->rsrc_update.nr_args);
6485 io_ring_submit_unlock(ctx, needs_lock);
6489 __io_req_complete(req, issue_flags, ret, 0);
6493 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6495 switch (req->opcode) {
6498 case IORING_OP_READV:
6499 case IORING_OP_READ_FIXED:
6500 case IORING_OP_READ:
6501 return io_read_prep(req, sqe);
6502 case IORING_OP_WRITEV:
6503 case IORING_OP_WRITE_FIXED:
6504 case IORING_OP_WRITE:
6505 return io_write_prep(req, sqe);
6506 case IORING_OP_POLL_ADD:
6507 return io_poll_add_prep(req, sqe);
6508 case IORING_OP_POLL_REMOVE:
6509 return io_poll_update_prep(req, sqe);
6510 case IORING_OP_FSYNC:
6511 return io_fsync_prep(req, sqe);
6512 case IORING_OP_SYNC_FILE_RANGE:
6513 return io_sfr_prep(req, sqe);
6514 case IORING_OP_SENDMSG:
6515 case IORING_OP_SEND:
6516 return io_sendmsg_prep(req, sqe);
6517 case IORING_OP_RECVMSG:
6518 case IORING_OP_RECV:
6519 return io_recvmsg_prep(req, sqe);
6520 case IORING_OP_CONNECT:
6521 return io_connect_prep(req, sqe);
6522 case IORING_OP_TIMEOUT:
6523 return io_timeout_prep(req, sqe, false);
6524 case IORING_OP_TIMEOUT_REMOVE:
6525 return io_timeout_remove_prep(req, sqe);
6526 case IORING_OP_ASYNC_CANCEL:
6527 return io_async_cancel_prep(req, sqe);
6528 case IORING_OP_LINK_TIMEOUT:
6529 return io_timeout_prep(req, sqe, true);
6530 case IORING_OP_ACCEPT:
6531 return io_accept_prep(req, sqe);
6532 case IORING_OP_FALLOCATE:
6533 return io_fallocate_prep(req, sqe);
6534 case IORING_OP_OPENAT:
6535 return io_openat_prep(req, sqe);
6536 case IORING_OP_CLOSE:
6537 return io_close_prep(req, sqe);
6538 case IORING_OP_FILES_UPDATE:
6539 return io_rsrc_update_prep(req, sqe);
6540 case IORING_OP_STATX:
6541 return io_statx_prep(req, sqe);
6542 case IORING_OP_FADVISE:
6543 return io_fadvise_prep(req, sqe);
6544 case IORING_OP_MADVISE:
6545 return io_madvise_prep(req, sqe);
6546 case IORING_OP_OPENAT2:
6547 return io_openat2_prep(req, sqe);
6548 case IORING_OP_EPOLL_CTL:
6549 return io_epoll_ctl_prep(req, sqe);
6550 case IORING_OP_SPLICE:
6551 return io_splice_prep(req, sqe);
6552 case IORING_OP_PROVIDE_BUFFERS:
6553 return io_provide_buffers_prep(req, sqe);
6554 case IORING_OP_REMOVE_BUFFERS:
6555 return io_remove_buffers_prep(req, sqe);
6557 return io_tee_prep(req, sqe);
6558 case IORING_OP_SHUTDOWN:
6559 return io_shutdown_prep(req, sqe);
6560 case IORING_OP_RENAMEAT:
6561 return io_renameat_prep(req, sqe);
6562 case IORING_OP_UNLINKAT:
6563 return io_unlinkat_prep(req, sqe);
6564 case IORING_OP_MKDIRAT:
6565 return io_mkdirat_prep(req, sqe);
6566 case IORING_OP_SYMLINKAT:
6567 return io_symlinkat_prep(req, sqe);
6568 case IORING_OP_LINKAT:
6569 return io_linkat_prep(req, sqe);
6572 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6577 static int io_req_prep_async(struct io_kiocb *req)
6579 if (!io_op_defs[req->opcode].needs_async_setup)
6581 if (WARN_ON_ONCE(req_has_async_data(req)))
6583 if (io_alloc_async_data(req))
6586 switch (req->opcode) {
6587 case IORING_OP_READV:
6588 return io_rw_prep_async(req, READ);
6589 case IORING_OP_WRITEV:
6590 return io_rw_prep_async(req, WRITE);
6591 case IORING_OP_SENDMSG:
6592 return io_sendmsg_prep_async(req);
6593 case IORING_OP_RECVMSG:
6594 return io_recvmsg_prep_async(req);
6595 case IORING_OP_CONNECT:
6596 return io_connect_prep_async(req);
6598 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6603 static u32 io_get_sequence(struct io_kiocb *req)
6605 u32 seq = req->ctx->cached_sq_head;
6607 /* need original cached_sq_head, but it was increased for each req */
6608 io_for_each_link(req, req)
6613 static __cold void io_drain_req(struct io_kiocb *req)
6615 struct io_ring_ctx *ctx = req->ctx;
6616 struct io_defer_entry *de;
6618 u32 seq = io_get_sequence(req);
6620 /* Still need defer if there is pending req in defer list. */
6621 spin_lock(&ctx->completion_lock);
6622 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
6623 spin_unlock(&ctx->completion_lock);
6625 ctx->drain_active = false;
6626 io_req_task_queue(req);
6629 spin_unlock(&ctx->completion_lock);
6631 ret = io_req_prep_async(req);
6634 io_req_complete_failed(req, ret);
6637 io_prep_async_link(req);
6638 de = kmalloc(sizeof(*de), GFP_KERNEL);
6644 spin_lock(&ctx->completion_lock);
6645 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6646 spin_unlock(&ctx->completion_lock);
6651 trace_io_uring_defer(ctx, req, req->user_data);
6654 list_add_tail(&de->list, &ctx->defer_list);
6655 spin_unlock(&ctx->completion_lock);
6658 static void io_clean_op(struct io_kiocb *req)
6660 if (req->flags & REQ_F_BUFFER_SELECTED)
6663 if (req->flags & REQ_F_NEED_CLEANUP) {
6664 switch (req->opcode) {
6665 case IORING_OP_READV:
6666 case IORING_OP_READ_FIXED:
6667 case IORING_OP_READ:
6668 case IORING_OP_WRITEV:
6669 case IORING_OP_WRITE_FIXED:
6670 case IORING_OP_WRITE: {
6671 struct io_async_rw *io = req->async_data;
6673 kfree(io->free_iovec);
6676 case IORING_OP_RECVMSG:
6677 case IORING_OP_SENDMSG: {
6678 struct io_async_msghdr *io = req->async_data;
6680 kfree(io->free_iov);
6683 case IORING_OP_SPLICE:
6685 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6686 io_put_file(req->splice.file_in);
6688 case IORING_OP_OPENAT:
6689 case IORING_OP_OPENAT2:
6690 if (req->open.filename)
6691 putname(req->open.filename);
6693 case IORING_OP_RENAMEAT:
6694 putname(req->rename.oldpath);
6695 putname(req->rename.newpath);
6697 case IORING_OP_UNLINKAT:
6698 putname(req->unlink.filename);
6700 case IORING_OP_MKDIRAT:
6701 putname(req->mkdir.filename);
6703 case IORING_OP_SYMLINKAT:
6704 putname(req->symlink.oldpath);
6705 putname(req->symlink.newpath);
6707 case IORING_OP_LINKAT:
6708 putname(req->hardlink.oldpath);
6709 putname(req->hardlink.newpath);
6713 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6714 kfree(req->apoll->double_poll);
6718 if (req->flags & REQ_F_INFLIGHT) {
6719 struct io_uring_task *tctx = req->task->io_uring;
6721 atomic_dec(&tctx->inflight_tracked);
6723 if (req->flags & REQ_F_CREDS)
6724 put_cred(req->creds);
6725 if (req->flags & REQ_F_ASYNC_DATA) {
6726 kfree(req->async_data);
6727 req->async_data = NULL;
6729 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6732 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6734 const struct cred *creds = NULL;
6737 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
6738 creds = override_creds(req->creds);
6740 if (!io_op_defs[req->opcode].audit_skip)
6741 audit_uring_entry(req->opcode);
6743 switch (req->opcode) {
6745 ret = io_nop(req, issue_flags);
6747 case IORING_OP_READV:
6748 case IORING_OP_READ_FIXED:
6749 case IORING_OP_READ:
6750 ret = io_read(req, issue_flags);
6752 case IORING_OP_WRITEV:
6753 case IORING_OP_WRITE_FIXED:
6754 case IORING_OP_WRITE:
6755 ret = io_write(req, issue_flags);
6757 case IORING_OP_FSYNC:
6758 ret = io_fsync(req, issue_flags);
6760 case IORING_OP_POLL_ADD:
6761 ret = io_poll_add(req, issue_flags);
6763 case IORING_OP_POLL_REMOVE:
6764 ret = io_poll_update(req, issue_flags);
6766 case IORING_OP_SYNC_FILE_RANGE:
6767 ret = io_sync_file_range(req, issue_flags);
6769 case IORING_OP_SENDMSG:
6770 ret = io_sendmsg(req, issue_flags);
6772 case IORING_OP_SEND:
6773 ret = io_send(req, issue_flags);
6775 case IORING_OP_RECVMSG:
6776 ret = io_recvmsg(req, issue_flags);
6778 case IORING_OP_RECV:
6779 ret = io_recv(req, issue_flags);
6781 case IORING_OP_TIMEOUT:
6782 ret = io_timeout(req, issue_flags);
6784 case IORING_OP_TIMEOUT_REMOVE:
6785 ret = io_timeout_remove(req, issue_flags);
6787 case IORING_OP_ACCEPT:
6788 ret = io_accept(req, issue_flags);
6790 case IORING_OP_CONNECT:
6791 ret = io_connect(req, issue_flags);
6793 case IORING_OP_ASYNC_CANCEL:
6794 ret = io_async_cancel(req, issue_flags);
6796 case IORING_OP_FALLOCATE:
6797 ret = io_fallocate(req, issue_flags);
6799 case IORING_OP_OPENAT:
6800 ret = io_openat(req, issue_flags);
6802 case IORING_OP_CLOSE:
6803 ret = io_close(req, issue_flags);
6805 case IORING_OP_FILES_UPDATE:
6806 ret = io_files_update(req, issue_flags);
6808 case IORING_OP_STATX:
6809 ret = io_statx(req, issue_flags);
6811 case IORING_OP_FADVISE:
6812 ret = io_fadvise(req, issue_flags);
6814 case IORING_OP_MADVISE:
6815 ret = io_madvise(req, issue_flags);
6817 case IORING_OP_OPENAT2:
6818 ret = io_openat2(req, issue_flags);
6820 case IORING_OP_EPOLL_CTL:
6821 ret = io_epoll_ctl(req, issue_flags);
6823 case IORING_OP_SPLICE:
6824 ret = io_splice(req, issue_flags);
6826 case IORING_OP_PROVIDE_BUFFERS:
6827 ret = io_provide_buffers(req, issue_flags);
6829 case IORING_OP_REMOVE_BUFFERS:
6830 ret = io_remove_buffers(req, issue_flags);
6833 ret = io_tee(req, issue_flags);
6835 case IORING_OP_SHUTDOWN:
6836 ret = io_shutdown(req, issue_flags);
6838 case IORING_OP_RENAMEAT:
6839 ret = io_renameat(req, issue_flags);
6841 case IORING_OP_UNLINKAT:
6842 ret = io_unlinkat(req, issue_flags);
6844 case IORING_OP_MKDIRAT:
6845 ret = io_mkdirat(req, issue_flags);
6847 case IORING_OP_SYMLINKAT:
6848 ret = io_symlinkat(req, issue_flags);
6850 case IORING_OP_LINKAT:
6851 ret = io_linkat(req, issue_flags);
6858 if (!io_op_defs[req->opcode].audit_skip)
6859 audit_uring_exit(!ret, ret);
6862 revert_creds(creds);
6865 /* If the op doesn't have a file, we're not polling for it */
6866 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6867 io_iopoll_req_issued(req, issue_flags);
6872 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6874 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6876 req = io_put_req_find_next(req);
6877 return req ? &req->work : NULL;
6880 static void io_wq_submit_work(struct io_wq_work *work)
6882 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6883 unsigned int issue_flags = IO_URING_F_UNLOCKED;
6884 bool needs_poll = false;
6885 struct io_kiocb *timeout;
6888 /* one will be dropped by ->io_free_work() after returning to io-wq */
6889 if (!(req->flags & REQ_F_REFCOUNT))
6890 __io_req_set_refcount(req, 2);
6894 timeout = io_prep_linked_timeout(req);
6896 io_queue_linked_timeout(timeout);
6898 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
6899 if (work->flags & IO_WQ_WORK_CANCEL) {
6900 io_req_task_queue_fail(req, -ECANCELED);
6904 if (req->flags & REQ_F_FORCE_ASYNC) {
6905 const struct io_op_def *def = &io_op_defs[req->opcode];
6906 bool opcode_poll = def->pollin || def->pollout;
6908 if (opcode_poll && file_can_poll(req->file)) {
6910 issue_flags |= IO_URING_F_NONBLOCK;
6915 ret = io_issue_sqe(req, issue_flags);
6919 * We can get EAGAIN for iopolled IO even though we're
6920 * forcing a sync submission from here, since we can't
6921 * wait for request slots on the block side.
6928 if (io_arm_poll_handler(req) == IO_APOLL_OK)
6930 /* aborted or ready, in either case retry blocking */
6932 issue_flags &= ~IO_URING_F_NONBLOCK;
6935 /* avoid locking problems by failing it from a clean context */
6937 io_req_task_queue_fail(req, ret);
6940 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6943 return &table->files[i];
6946 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6949 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6951 return (struct file *) (slot->file_ptr & FFS_MASK);
6954 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6956 unsigned long file_ptr = (unsigned long) file;
6958 file_ptr |= io_file_get_flags(file);
6959 file_slot->file_ptr = file_ptr;
6962 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6963 struct io_kiocb *req, int fd)
6966 unsigned long file_ptr;
6968 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6970 fd = array_index_nospec(fd, ctx->nr_user_files);
6971 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6972 file = (struct file *) (file_ptr & FFS_MASK);
6973 file_ptr &= ~FFS_MASK;
6974 /* mask in overlapping REQ_F and FFS bits */
6975 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
6976 io_req_set_rsrc_node(req, ctx);
6980 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6981 struct io_kiocb *req, int fd)
6983 struct file *file = fget(fd);
6985 trace_io_uring_file_get(ctx, fd);
6987 /* we don't allow fixed io_uring files */
6988 if (file && unlikely(file->f_op == &io_uring_fops))
6989 io_req_track_inflight(req);
6993 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6994 struct io_kiocb *req, int fd, bool fixed)
6997 return io_file_get_fixed(ctx, req, fd);
6999 return io_file_get_normal(ctx, req, fd);
7002 static void io_req_task_link_timeout(struct io_kiocb *req, bool *locked)
7004 struct io_kiocb *prev = req->timeout.prev;
7008 if (!(req->task->flags & PF_EXITING))
7009 ret = io_try_cancel_userdata(req, prev->user_data);
7010 io_req_complete_post(req, ret ?: -ETIME, 0);
7013 io_req_complete_post(req, -ETIME, 0);
7017 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
7019 struct io_timeout_data *data = container_of(timer,
7020 struct io_timeout_data, timer);
7021 struct io_kiocb *prev, *req = data->req;
7022 struct io_ring_ctx *ctx = req->ctx;
7023 unsigned long flags;
7025 spin_lock_irqsave(&ctx->timeout_lock, flags);
7026 prev = req->timeout.head;
7027 req->timeout.head = NULL;
7030 * We don't expect the list to be empty, that will only happen if we
7031 * race with the completion of the linked work.
7034 io_remove_next_linked(prev);
7035 if (!req_ref_inc_not_zero(prev))
7038 list_del(&req->timeout.list);
7039 req->timeout.prev = prev;
7040 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
7042 req->io_task_work.func = io_req_task_link_timeout;
7043 io_req_task_work_add(req, false);
7044 return HRTIMER_NORESTART;
7047 static void io_queue_linked_timeout(struct io_kiocb *req)
7049 struct io_ring_ctx *ctx = req->ctx;
7051 spin_lock_irq(&ctx->timeout_lock);
7053 * If the back reference is NULL, then our linked request finished
7054 * before we got a chance to setup the timer
7056 if (req->timeout.head) {
7057 struct io_timeout_data *data = req->async_data;
7059 data->timer.function = io_link_timeout_fn;
7060 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
7062 list_add_tail(&req->timeout.list, &ctx->ltimeout_list);
7064 spin_unlock_irq(&ctx->timeout_lock);
7065 /* drop submission reference */
7069 static void io_queue_sqe_arm_apoll(struct io_kiocb *req)
7070 __must_hold(&req->ctx->uring_lock)
7072 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
7074 switch (io_arm_poll_handler(req)) {
7075 case IO_APOLL_READY:
7076 io_req_task_queue(req);
7078 case IO_APOLL_ABORTED:
7080 * Queued up for async execution, worker will release
7081 * submit reference when the iocb is actually submitted.
7083 io_queue_async_work(req, NULL);
7088 io_queue_linked_timeout(linked_timeout);
7091 static inline void __io_queue_sqe(struct io_kiocb *req)
7092 __must_hold(&req->ctx->uring_lock)
7094 struct io_kiocb *linked_timeout;
7097 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
7099 if (req->flags & REQ_F_COMPLETE_INLINE) {
7100 io_req_add_compl_list(req);
7104 * We async punt it if the file wasn't marked NOWAIT, or if the file
7105 * doesn't support non-blocking read/write attempts
7108 linked_timeout = io_prep_linked_timeout(req);
7110 io_queue_linked_timeout(linked_timeout);
7111 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
7112 io_queue_sqe_arm_apoll(req);
7114 io_req_complete_failed(req, ret);
7118 static void io_queue_sqe_fallback(struct io_kiocb *req)
7119 __must_hold(&req->ctx->uring_lock)
7121 if (req->flags & REQ_F_FAIL) {
7122 io_req_complete_fail_submit(req);
7123 } else if (unlikely(req->ctx->drain_active)) {
7126 int ret = io_req_prep_async(req);
7129 io_req_complete_failed(req, ret);
7131 io_queue_async_work(req, NULL);
7135 static inline void io_queue_sqe(struct io_kiocb *req)
7136 __must_hold(&req->ctx->uring_lock)
7138 if (likely(!(req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))))
7139 __io_queue_sqe(req);
7141 io_queue_sqe_fallback(req);
7145 * Check SQE restrictions (opcode and flags).
7147 * Returns 'true' if SQE is allowed, 'false' otherwise.
7149 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
7150 struct io_kiocb *req,
7151 unsigned int sqe_flags)
7153 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
7156 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
7157 ctx->restrictions.sqe_flags_required)
7160 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
7161 ctx->restrictions.sqe_flags_required))
7167 static void io_init_req_drain(struct io_kiocb *req)
7169 struct io_ring_ctx *ctx = req->ctx;
7170 struct io_kiocb *head = ctx->submit_state.link.head;
7172 ctx->drain_active = true;
7175 * If we need to drain a request in the middle of a link, drain
7176 * the head request and the next request/link after the current
7177 * link. Considering sequential execution of links,
7178 * REQ_F_IO_DRAIN will be maintained for every request of our
7181 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
7182 ctx->drain_next = true;
7186 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
7187 const struct io_uring_sqe *sqe)
7188 __must_hold(&ctx->uring_lock)
7190 unsigned int sqe_flags;
7194 /* req is partially pre-initialised, see io_preinit_req() */
7195 req->opcode = opcode = READ_ONCE(sqe->opcode);
7196 /* same numerical values with corresponding REQ_F_*, safe to copy */
7197 req->flags = sqe_flags = READ_ONCE(sqe->flags);
7198 req->user_data = READ_ONCE(sqe->user_data);
7200 req->fixed_rsrc_refs = NULL;
7201 req->task = current;
7203 if (unlikely(opcode >= IORING_OP_LAST)) {
7207 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
7208 /* enforce forwards compatibility on users */
7209 if (sqe_flags & ~SQE_VALID_FLAGS)
7211 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
7212 !io_op_defs[opcode].buffer_select)
7214 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
7215 ctx->drain_disabled = true;
7216 if (sqe_flags & IOSQE_IO_DRAIN) {
7217 if (ctx->drain_disabled)
7219 io_init_req_drain(req);
7222 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
7223 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
7225 /* knock it to the slow queue path, will be drained there */
7226 if (ctx->drain_active)
7227 req->flags |= REQ_F_FORCE_ASYNC;
7228 /* if there is no link, we're at "next" request and need to drain */
7229 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
7230 ctx->drain_next = false;
7231 ctx->drain_active = true;
7232 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
7236 if (io_op_defs[opcode].needs_file) {
7237 struct io_submit_state *state = &ctx->submit_state;
7240 * Plug now if we have more than 2 IO left after this, and the
7241 * target is potentially a read/write to block based storage.
7243 if (state->need_plug && io_op_defs[opcode].plug) {
7244 state->plug_started = true;
7245 state->need_plug = false;
7246 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
7249 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
7250 (sqe_flags & IOSQE_FIXED_FILE));
7251 if (unlikely(!req->file))
7255 personality = READ_ONCE(sqe->personality);
7259 req->creds = xa_load(&ctx->personalities, personality);
7262 get_cred(req->creds);
7263 ret = security_uring_override_creds(req->creds);
7265 put_cred(req->creds);
7268 req->flags |= REQ_F_CREDS;
7271 return io_req_prep(req, sqe);
7274 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
7275 const struct io_uring_sqe *sqe)
7276 __must_hold(&ctx->uring_lock)
7278 struct io_submit_link *link = &ctx->submit_state.link;
7281 ret = io_init_req(ctx, req, sqe);
7282 if (unlikely(ret)) {
7283 trace_io_uring_req_failed(sqe, ret);
7285 /* fail even hard links since we don't submit */
7288 * we can judge a link req is failed or cancelled by if
7289 * REQ_F_FAIL is set, but the head is an exception since
7290 * it may be set REQ_F_FAIL because of other req's failure
7291 * so let's leverage req->result to distinguish if a head
7292 * is set REQ_F_FAIL because of its failure or other req's
7293 * failure so that we can set the correct ret code for it.
7294 * init result here to avoid affecting the normal path.
7296 if (!(link->head->flags & REQ_F_FAIL))
7297 req_fail_link_node(link->head, -ECANCELED);
7298 } else if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7300 * the current req is a normal req, we should return
7301 * error and thus break the submittion loop.
7303 io_req_complete_failed(req, ret);
7306 req_fail_link_node(req, ret);
7309 /* don't need @sqe from now on */
7310 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
7312 ctx->flags & IORING_SETUP_SQPOLL);
7315 * If we already have a head request, queue this one for async
7316 * submittal once the head completes. If we don't have a head but
7317 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
7318 * submitted sync once the chain is complete. If none of those
7319 * conditions are true (normal request), then just queue it.
7322 struct io_kiocb *head = link->head;
7324 if (!(req->flags & REQ_F_FAIL)) {
7325 ret = io_req_prep_async(req);
7326 if (unlikely(ret)) {
7327 req_fail_link_node(req, ret);
7328 if (!(head->flags & REQ_F_FAIL))
7329 req_fail_link_node(head, -ECANCELED);
7332 trace_io_uring_link(ctx, req, head);
7333 link->last->link = req;
7336 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK))
7338 /* last request of a link, enqueue the link */
7341 } else if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
7352 * Batched submission is done, ensure local IO is flushed out.
7354 static void io_submit_state_end(struct io_ring_ctx *ctx)
7356 struct io_submit_state *state = &ctx->submit_state;
7358 if (state->link.head)
7359 io_queue_sqe(state->link.head);
7360 /* flush only after queuing links as they can generate completions */
7361 io_submit_flush_completions(ctx);
7362 if (state->plug_started)
7363 blk_finish_plug(&state->plug);
7367 * Start submission side cache.
7369 static void io_submit_state_start(struct io_submit_state *state,
7370 unsigned int max_ios)
7372 state->plug_started = false;
7373 state->need_plug = max_ios > 2;
7374 state->submit_nr = max_ios;
7375 /* set only head, no need to init link_last in advance */
7376 state->link.head = NULL;
7379 static void io_commit_sqring(struct io_ring_ctx *ctx)
7381 struct io_rings *rings = ctx->rings;
7384 * Ensure any loads from the SQEs are done at this point,
7385 * since once we write the new head, the application could
7386 * write new data to them.
7388 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
7392 * Fetch an sqe, if one is available. Note this returns a pointer to memory
7393 * that is mapped by userspace. This means that care needs to be taken to
7394 * ensure that reads are stable, as we cannot rely on userspace always
7395 * being a good citizen. If members of the sqe are validated and then later
7396 * used, it's important that those reads are done through READ_ONCE() to
7397 * prevent a re-load down the line.
7399 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
7401 unsigned head, mask = ctx->sq_entries - 1;
7402 unsigned sq_idx = ctx->cached_sq_head++ & mask;
7405 * The cached sq head (or cq tail) serves two purposes:
7407 * 1) allows us to batch the cost of updating the user visible
7409 * 2) allows the kernel side to track the head on its own, even
7410 * though the application is the one updating it.
7412 head = READ_ONCE(ctx->sq_array[sq_idx]);
7413 if (likely(head < ctx->sq_entries))
7414 return &ctx->sq_sqes[head];
7416 /* drop invalid entries */
7418 WRITE_ONCE(ctx->rings->sq_dropped,
7419 READ_ONCE(ctx->rings->sq_dropped) + 1);
7423 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
7424 __must_hold(&ctx->uring_lock)
7426 unsigned int entries = io_sqring_entries(ctx);
7429 if (unlikely(!entries))
7431 /* make sure SQ entry isn't read before tail */
7432 nr = min3(nr, ctx->sq_entries, entries);
7433 io_get_task_refs(nr);
7435 io_submit_state_start(&ctx->submit_state, nr);
7437 const struct io_uring_sqe *sqe;
7438 struct io_kiocb *req;
7440 if (unlikely(!io_alloc_req_refill(ctx))) {
7442 submitted = -EAGAIN;
7445 req = io_alloc_req(ctx);
7446 sqe = io_get_sqe(ctx);
7447 if (unlikely(!sqe)) {
7448 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
7451 /* will complete beyond this point, count as submitted */
7453 if (io_submit_sqe(ctx, req, sqe))
7455 } while (submitted < nr);
7457 if (unlikely(submitted != nr)) {
7458 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
7459 int unused = nr - ref_used;
7461 current->io_uring->cached_refs += unused;
7464 io_submit_state_end(ctx);
7465 /* Commit SQ ring head once we've consumed and submitted all SQEs */
7466 io_commit_sqring(ctx);
7471 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
7473 return READ_ONCE(sqd->state);
7476 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
7478 /* Tell userspace we may need a wakeup call */
7479 spin_lock(&ctx->completion_lock);
7480 WRITE_ONCE(ctx->rings->sq_flags,
7481 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
7482 spin_unlock(&ctx->completion_lock);
7485 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
7487 spin_lock(&ctx->completion_lock);
7488 WRITE_ONCE(ctx->rings->sq_flags,
7489 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
7490 spin_unlock(&ctx->completion_lock);
7493 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
7495 unsigned int to_submit;
7498 to_submit = io_sqring_entries(ctx);
7499 /* if we're handling multiple rings, cap submit size for fairness */
7500 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
7501 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
7503 if (!wq_list_empty(&ctx->iopoll_list) || to_submit) {
7504 const struct cred *creds = NULL;
7506 if (ctx->sq_creds != current_cred())
7507 creds = override_creds(ctx->sq_creds);
7509 mutex_lock(&ctx->uring_lock);
7510 if (!wq_list_empty(&ctx->iopoll_list))
7511 io_do_iopoll(ctx, true);
7514 * Don't submit if refs are dying, good for io_uring_register(),
7515 * but also it is relied upon by io_ring_exit_work()
7517 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
7518 !(ctx->flags & IORING_SETUP_R_DISABLED))
7519 ret = io_submit_sqes(ctx, to_submit);
7520 mutex_unlock(&ctx->uring_lock);
7522 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
7523 wake_up(&ctx->sqo_sq_wait);
7525 revert_creds(creds);
7531 static __cold void io_sqd_update_thread_idle(struct io_sq_data *sqd)
7533 struct io_ring_ctx *ctx;
7534 unsigned sq_thread_idle = 0;
7536 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7537 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
7538 sqd->sq_thread_idle = sq_thread_idle;
7541 static bool io_sqd_handle_event(struct io_sq_data *sqd)
7543 bool did_sig = false;
7544 struct ksignal ksig;
7546 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
7547 signal_pending(current)) {
7548 mutex_unlock(&sqd->lock);
7549 if (signal_pending(current))
7550 did_sig = get_signal(&ksig);
7552 mutex_lock(&sqd->lock);
7554 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7557 static int io_sq_thread(void *data)
7559 struct io_sq_data *sqd = data;
7560 struct io_ring_ctx *ctx;
7561 unsigned long timeout = 0;
7562 char buf[TASK_COMM_LEN];
7565 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
7566 set_task_comm(current, buf);
7568 if (sqd->sq_cpu != -1)
7569 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
7571 set_cpus_allowed_ptr(current, cpu_online_mask);
7572 current->flags |= PF_NO_SETAFFINITY;
7574 audit_alloc_kernel(current);
7576 mutex_lock(&sqd->lock);
7578 bool cap_entries, sqt_spin = false;
7580 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
7581 if (io_sqd_handle_event(sqd))
7583 timeout = jiffies + sqd->sq_thread_idle;
7586 cap_entries = !list_is_singular(&sqd->ctx_list);
7587 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7588 int ret = __io_sq_thread(ctx, cap_entries);
7590 if (!sqt_spin && (ret > 0 || !wq_list_empty(&ctx->iopoll_list)))
7593 if (io_run_task_work())
7596 if (sqt_spin || !time_after(jiffies, timeout)) {
7599 timeout = jiffies + sqd->sq_thread_idle;
7603 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7604 if (!io_sqd_events_pending(sqd) && !current->task_works) {
7605 bool needs_sched = true;
7607 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7608 io_ring_set_wakeup_flag(ctx);
7610 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7611 !wq_list_empty(&ctx->iopoll_list)) {
7612 needs_sched = false;
7615 if (io_sqring_entries(ctx)) {
7616 needs_sched = false;
7622 mutex_unlock(&sqd->lock);
7624 mutex_lock(&sqd->lock);
7626 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7627 io_ring_clear_wakeup_flag(ctx);
7630 finish_wait(&sqd->wait, &wait);
7631 timeout = jiffies + sqd->sq_thread_idle;
7634 io_uring_cancel_generic(true, sqd);
7636 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7637 io_ring_set_wakeup_flag(ctx);
7639 mutex_unlock(&sqd->lock);
7641 audit_free(current);
7643 complete(&sqd->exited);
7647 struct io_wait_queue {
7648 struct wait_queue_entry wq;
7649 struct io_ring_ctx *ctx;
7651 unsigned nr_timeouts;
7654 static inline bool io_should_wake(struct io_wait_queue *iowq)
7656 struct io_ring_ctx *ctx = iowq->ctx;
7657 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7660 * Wake up if we have enough events, or if a timeout occurred since we
7661 * started waiting. For timeouts, we always want to return to userspace,
7662 * regardless of event count.
7664 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7667 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7668 int wake_flags, void *key)
7670 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7674 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7675 * the task, and the next invocation will do it.
7677 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7678 return autoremove_wake_function(curr, mode, wake_flags, key);
7682 static int io_run_task_work_sig(void)
7684 if (io_run_task_work())
7686 if (!signal_pending(current))
7688 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7689 return -ERESTARTSYS;
7693 /* when returns >0, the caller should retry */
7694 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7695 struct io_wait_queue *iowq,
7696 signed long *timeout)
7700 /* make sure we run task_work before checking for signals */
7701 ret = io_run_task_work_sig();
7702 if (ret || io_should_wake(iowq))
7704 /* let the caller flush overflows, retry */
7705 if (test_bit(0, &ctx->check_cq_overflow))
7708 *timeout = schedule_timeout(*timeout);
7709 return !*timeout ? -ETIME : 1;
7713 * Wait until events become available, if we don't already have some. The
7714 * application must reap them itself, as they reside on the shared cq ring.
7716 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7717 const sigset_t __user *sig, size_t sigsz,
7718 struct __kernel_timespec __user *uts)
7720 struct io_wait_queue iowq;
7721 struct io_rings *rings = ctx->rings;
7722 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7726 io_cqring_overflow_flush(ctx);
7727 if (io_cqring_events(ctx) >= min_events)
7729 if (!io_run_task_work())
7734 struct timespec64 ts;
7736 if (get_timespec64(&ts, uts))
7738 timeout = timespec64_to_jiffies(&ts);
7742 #ifdef CONFIG_COMPAT
7743 if (in_compat_syscall())
7744 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7748 ret = set_user_sigmask(sig, sigsz);
7754 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7755 iowq.wq.private = current;
7756 INIT_LIST_HEAD(&iowq.wq.entry);
7758 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7759 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7761 trace_io_uring_cqring_wait(ctx, min_events);
7763 /* if we can't even flush overflow, don't wait for more */
7764 if (!io_cqring_overflow_flush(ctx)) {
7768 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7769 TASK_INTERRUPTIBLE);
7770 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7771 finish_wait(&ctx->cq_wait, &iowq.wq);
7775 restore_saved_sigmask_unless(ret == -EINTR);
7777 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7780 static void io_free_page_table(void **table, size_t size)
7782 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7784 for (i = 0; i < nr_tables; i++)
7789 static __cold void **io_alloc_page_table(size_t size)
7791 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7792 size_t init_size = size;
7795 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
7799 for (i = 0; i < nr_tables; i++) {
7800 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7802 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
7804 io_free_page_table(table, init_size);
7812 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7814 percpu_ref_exit(&ref_node->refs);
7818 static __cold void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7820 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7821 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7822 unsigned long flags;
7823 bool first_add = false;
7824 unsigned long delay = HZ;
7826 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7829 /* if we are mid-quiesce then do not delay */
7830 if (node->rsrc_data->quiesce)
7833 while (!list_empty(&ctx->rsrc_ref_list)) {
7834 node = list_first_entry(&ctx->rsrc_ref_list,
7835 struct io_rsrc_node, node);
7836 /* recycle ref nodes in order */
7839 list_del(&node->node);
7840 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7842 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7845 mod_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7848 static struct io_rsrc_node *io_rsrc_node_alloc(void)
7850 struct io_rsrc_node *ref_node;
7852 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7856 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7861 INIT_LIST_HEAD(&ref_node->node);
7862 INIT_LIST_HEAD(&ref_node->rsrc_list);
7863 ref_node->done = false;
7867 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7868 struct io_rsrc_data *data_to_kill)
7869 __must_hold(&ctx->uring_lock)
7871 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7872 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7874 io_rsrc_refs_drop(ctx);
7877 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7879 rsrc_node->rsrc_data = data_to_kill;
7880 spin_lock_irq(&ctx->rsrc_ref_lock);
7881 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7882 spin_unlock_irq(&ctx->rsrc_ref_lock);
7884 atomic_inc(&data_to_kill->refs);
7885 percpu_ref_kill(&rsrc_node->refs);
7886 ctx->rsrc_node = NULL;
7889 if (!ctx->rsrc_node) {
7890 ctx->rsrc_node = ctx->rsrc_backup_node;
7891 ctx->rsrc_backup_node = NULL;
7895 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7897 if (ctx->rsrc_backup_node)
7899 ctx->rsrc_backup_node = io_rsrc_node_alloc();
7900 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7903 static __cold int io_rsrc_ref_quiesce(struct io_rsrc_data *data,
7904 struct io_ring_ctx *ctx)
7908 /* As we may drop ->uring_lock, other task may have started quiesce */
7912 data->quiesce = true;
7914 ret = io_rsrc_node_switch_start(ctx);
7917 io_rsrc_node_switch(ctx, data);
7919 /* kill initial ref, already quiesced if zero */
7920 if (atomic_dec_and_test(&data->refs))
7922 mutex_unlock(&ctx->uring_lock);
7923 flush_delayed_work(&ctx->rsrc_put_work);
7924 ret = wait_for_completion_interruptible(&data->done);
7926 mutex_lock(&ctx->uring_lock);
7930 atomic_inc(&data->refs);
7931 /* wait for all works potentially completing data->done */
7932 flush_delayed_work(&ctx->rsrc_put_work);
7933 reinit_completion(&data->done);
7935 ret = io_run_task_work_sig();
7936 mutex_lock(&ctx->uring_lock);
7938 data->quiesce = false;
7943 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7945 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7946 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7948 return &data->tags[table_idx][off];
7951 static void io_rsrc_data_free(struct io_rsrc_data *data)
7953 size_t size = data->nr * sizeof(data->tags[0][0]);
7956 io_free_page_table((void **)data->tags, size);
7960 static __cold int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7961 u64 __user *utags, unsigned nr,
7962 struct io_rsrc_data **pdata)
7964 struct io_rsrc_data *data;
7968 data = kzalloc(sizeof(*data), GFP_KERNEL);
7971 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7979 data->do_put = do_put;
7982 for (i = 0; i < nr; i++) {
7983 u64 *tag_slot = io_get_tag_slot(data, i);
7985 if (copy_from_user(tag_slot, &utags[i],
7991 atomic_set(&data->refs, 1);
7992 init_completion(&data->done);
7996 io_rsrc_data_free(data);
8000 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
8002 table->files = kvcalloc(nr_files, sizeof(table->files[0]),
8003 GFP_KERNEL_ACCOUNT);
8004 return !!table->files;
8007 static void io_free_file_tables(struct io_file_table *table)
8009 kvfree(table->files);
8010 table->files = NULL;
8013 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
8015 #if defined(CONFIG_UNIX)
8016 if (ctx->ring_sock) {
8017 struct sock *sock = ctx->ring_sock->sk;
8018 struct sk_buff *skb;
8020 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
8026 for (i = 0; i < ctx->nr_user_files; i++) {
8029 file = io_file_from_index(ctx, i);
8034 io_free_file_tables(&ctx->file_table);
8035 io_rsrc_data_free(ctx->file_data);
8036 ctx->file_data = NULL;
8037 ctx->nr_user_files = 0;
8040 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
8044 if (!ctx->file_data)
8046 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
8048 __io_sqe_files_unregister(ctx);
8052 static void io_sq_thread_unpark(struct io_sq_data *sqd)
8053 __releases(&sqd->lock)
8055 WARN_ON_ONCE(sqd->thread == current);
8058 * Do the dance but not conditional clear_bit() because it'd race with
8059 * other threads incrementing park_pending and setting the bit.
8061 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
8062 if (atomic_dec_return(&sqd->park_pending))
8063 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
8064 mutex_unlock(&sqd->lock);
8067 static void io_sq_thread_park(struct io_sq_data *sqd)
8068 __acquires(&sqd->lock)
8070 WARN_ON_ONCE(sqd->thread == current);
8072 atomic_inc(&sqd->park_pending);
8073 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
8074 mutex_lock(&sqd->lock);
8076 wake_up_process(sqd->thread);
8079 static void io_sq_thread_stop(struct io_sq_data *sqd)
8081 WARN_ON_ONCE(sqd->thread == current);
8082 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
8084 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
8085 mutex_lock(&sqd->lock);
8087 wake_up_process(sqd->thread);
8088 mutex_unlock(&sqd->lock);
8089 wait_for_completion(&sqd->exited);
8092 static void io_put_sq_data(struct io_sq_data *sqd)
8094 if (refcount_dec_and_test(&sqd->refs)) {
8095 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
8097 io_sq_thread_stop(sqd);
8102 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
8104 struct io_sq_data *sqd = ctx->sq_data;
8107 io_sq_thread_park(sqd);
8108 list_del_init(&ctx->sqd_list);
8109 io_sqd_update_thread_idle(sqd);
8110 io_sq_thread_unpark(sqd);
8112 io_put_sq_data(sqd);
8113 ctx->sq_data = NULL;
8117 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
8119 struct io_ring_ctx *ctx_attach;
8120 struct io_sq_data *sqd;
8123 f = fdget(p->wq_fd);
8125 return ERR_PTR(-ENXIO);
8126 if (f.file->f_op != &io_uring_fops) {
8128 return ERR_PTR(-EINVAL);
8131 ctx_attach = f.file->private_data;
8132 sqd = ctx_attach->sq_data;
8135 return ERR_PTR(-EINVAL);
8137 if (sqd->task_tgid != current->tgid) {
8139 return ERR_PTR(-EPERM);
8142 refcount_inc(&sqd->refs);
8147 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
8150 struct io_sq_data *sqd;
8153 if (p->flags & IORING_SETUP_ATTACH_WQ) {
8154 sqd = io_attach_sq_data(p);
8159 /* fall through for EPERM case, setup new sqd/task */
8160 if (PTR_ERR(sqd) != -EPERM)
8164 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
8166 return ERR_PTR(-ENOMEM);
8168 atomic_set(&sqd->park_pending, 0);
8169 refcount_set(&sqd->refs, 1);
8170 INIT_LIST_HEAD(&sqd->ctx_list);
8171 mutex_init(&sqd->lock);
8172 init_waitqueue_head(&sqd->wait);
8173 init_completion(&sqd->exited);
8177 #if defined(CONFIG_UNIX)
8179 * Ensure the UNIX gc is aware of our file set, so we are certain that
8180 * the io_uring can be safely unregistered on process exit, even if we have
8181 * loops in the file referencing.
8183 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
8185 struct sock *sk = ctx->ring_sock->sk;
8186 struct scm_fp_list *fpl;
8187 struct sk_buff *skb;
8190 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
8194 skb = alloc_skb(0, GFP_KERNEL);
8203 fpl->user = get_uid(current_user());
8204 for (i = 0; i < nr; i++) {
8205 struct file *file = io_file_from_index(ctx, i + offset);
8209 fpl->fp[nr_files] = get_file(file);
8210 unix_inflight(fpl->user, fpl->fp[nr_files]);
8215 fpl->max = SCM_MAX_FD;
8216 fpl->count = nr_files;
8217 UNIXCB(skb).fp = fpl;
8218 skb->destructor = unix_destruct_scm;
8219 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
8220 skb_queue_head(&sk->sk_receive_queue, skb);
8222 for (i = 0; i < nr_files; i++)
8233 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
8234 * causes regular reference counting to break down. We rely on the UNIX
8235 * garbage collection to take care of this problem for us.
8237 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8239 unsigned left, total;
8243 left = ctx->nr_user_files;
8245 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
8247 ret = __io_sqe_files_scm(ctx, this_files, total);
8251 total += this_files;
8257 while (total < ctx->nr_user_files) {
8258 struct file *file = io_file_from_index(ctx, total);
8268 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8274 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8276 struct file *file = prsrc->file;
8277 #if defined(CONFIG_UNIX)
8278 struct sock *sock = ctx->ring_sock->sk;
8279 struct sk_buff_head list, *head = &sock->sk_receive_queue;
8280 struct sk_buff *skb;
8283 __skb_queue_head_init(&list);
8286 * Find the skb that holds this file in its SCM_RIGHTS. When found,
8287 * remove this entry and rearrange the file array.
8289 skb = skb_dequeue(head);
8291 struct scm_fp_list *fp;
8293 fp = UNIXCB(skb).fp;
8294 for (i = 0; i < fp->count; i++) {
8297 if (fp->fp[i] != file)
8300 unix_notinflight(fp->user, fp->fp[i]);
8301 left = fp->count - 1 - i;
8303 memmove(&fp->fp[i], &fp->fp[i + 1],
8304 left * sizeof(struct file *));
8311 __skb_queue_tail(&list, skb);
8321 __skb_queue_tail(&list, skb);
8323 skb = skb_dequeue(head);
8326 if (skb_peek(&list)) {
8327 spin_lock_irq(&head->lock);
8328 while ((skb = __skb_dequeue(&list)) != NULL)
8329 __skb_queue_tail(head, skb);
8330 spin_unlock_irq(&head->lock);
8337 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
8339 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
8340 struct io_ring_ctx *ctx = rsrc_data->ctx;
8341 struct io_rsrc_put *prsrc, *tmp;
8343 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
8344 list_del(&prsrc->list);
8347 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
8349 io_ring_submit_lock(ctx, lock_ring);
8350 spin_lock(&ctx->completion_lock);
8351 io_fill_cqe_aux(ctx, prsrc->tag, 0, 0);
8352 io_commit_cqring(ctx);
8353 spin_unlock(&ctx->completion_lock);
8354 io_cqring_ev_posted(ctx);
8355 io_ring_submit_unlock(ctx, lock_ring);
8358 rsrc_data->do_put(ctx, prsrc);
8362 io_rsrc_node_destroy(ref_node);
8363 if (atomic_dec_and_test(&rsrc_data->refs))
8364 complete(&rsrc_data->done);
8367 static void io_rsrc_put_work(struct work_struct *work)
8369 struct io_ring_ctx *ctx;
8370 struct llist_node *node;
8372 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
8373 node = llist_del_all(&ctx->rsrc_put_llist);
8376 struct io_rsrc_node *ref_node;
8377 struct llist_node *next = node->next;
8379 ref_node = llist_entry(node, struct io_rsrc_node, llist);
8380 __io_rsrc_put_work(ref_node);
8385 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
8386 unsigned nr_args, u64 __user *tags)
8388 __s32 __user *fds = (__s32 __user *) arg;
8397 if (nr_args > IORING_MAX_FIXED_FILES)
8399 if (nr_args > rlimit(RLIMIT_NOFILE))
8401 ret = io_rsrc_node_switch_start(ctx);
8404 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
8410 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
8413 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
8414 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
8418 /* allow sparse sets */
8421 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
8428 if (unlikely(!file))
8432 * Don't allow io_uring instances to be registered. If UNIX
8433 * isn't enabled, then this causes a reference cycle and this
8434 * instance can never get freed. If UNIX is enabled we'll
8435 * handle it just fine, but there's still no point in allowing
8436 * a ring fd as it doesn't support regular read/write anyway.
8438 if (file->f_op == &io_uring_fops) {
8442 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
8445 ret = io_sqe_files_scm(ctx);
8447 __io_sqe_files_unregister(ctx);
8451 io_rsrc_node_switch(ctx, NULL);
8454 for (i = 0; i < ctx->nr_user_files; i++) {
8455 file = io_file_from_index(ctx, i);
8459 io_free_file_tables(&ctx->file_table);
8460 ctx->nr_user_files = 0;
8462 io_rsrc_data_free(ctx->file_data);
8463 ctx->file_data = NULL;
8467 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
8470 #if defined(CONFIG_UNIX)
8471 struct sock *sock = ctx->ring_sock->sk;
8472 struct sk_buff_head *head = &sock->sk_receive_queue;
8473 struct sk_buff *skb;
8476 * See if we can merge this file into an existing skb SCM_RIGHTS
8477 * file set. If there's no room, fall back to allocating a new skb
8478 * and filling it in.
8480 spin_lock_irq(&head->lock);
8481 skb = skb_peek(head);
8483 struct scm_fp_list *fpl = UNIXCB(skb).fp;
8485 if (fpl->count < SCM_MAX_FD) {
8486 __skb_unlink(skb, head);
8487 spin_unlock_irq(&head->lock);
8488 fpl->fp[fpl->count] = get_file(file);
8489 unix_inflight(fpl->user, fpl->fp[fpl->count]);
8491 spin_lock_irq(&head->lock);
8492 __skb_queue_head(head, skb);
8497 spin_unlock_irq(&head->lock);
8504 return __io_sqe_files_scm(ctx, 1, index);
8510 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
8511 struct io_rsrc_node *node, void *rsrc)
8513 struct io_rsrc_put *prsrc;
8515 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
8519 prsrc->tag = *io_get_tag_slot(data, idx);
8521 list_add(&prsrc->list, &node->rsrc_list);
8525 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
8526 unsigned int issue_flags, u32 slot_index)
8528 struct io_ring_ctx *ctx = req->ctx;
8529 bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
8530 bool needs_switch = false;
8531 struct io_fixed_file *file_slot;
8534 io_ring_submit_lock(ctx, needs_lock);
8535 if (file->f_op == &io_uring_fops)
8538 if (!ctx->file_data)
8541 if (slot_index >= ctx->nr_user_files)
8544 slot_index = array_index_nospec(slot_index, ctx->nr_user_files);
8545 file_slot = io_fixed_file_slot(&ctx->file_table, slot_index);
8547 if (file_slot->file_ptr) {
8548 struct file *old_file;
8550 ret = io_rsrc_node_switch_start(ctx);
8554 old_file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8555 ret = io_queue_rsrc_removal(ctx->file_data, slot_index,
8556 ctx->rsrc_node, old_file);
8559 file_slot->file_ptr = 0;
8560 needs_switch = true;
8563 *io_get_tag_slot(ctx->file_data, slot_index) = 0;
8564 io_fixed_file_set(file_slot, file);
8565 ret = io_sqe_file_register(ctx, file, slot_index);
8567 file_slot->file_ptr = 0;
8574 io_rsrc_node_switch(ctx, ctx->file_data);
8575 io_ring_submit_unlock(ctx, needs_lock);
8581 static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags)
8583 unsigned int offset = req->close.file_slot - 1;
8584 struct io_ring_ctx *ctx = req->ctx;
8585 bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
8586 struct io_fixed_file *file_slot;
8590 io_ring_submit_lock(ctx, needs_lock);
8592 if (unlikely(!ctx->file_data))
8595 if (offset >= ctx->nr_user_files)
8597 ret = io_rsrc_node_switch_start(ctx);
8601 i = array_index_nospec(offset, ctx->nr_user_files);
8602 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8604 if (!file_slot->file_ptr)
8607 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8608 ret = io_queue_rsrc_removal(ctx->file_data, offset, ctx->rsrc_node, file);
8612 file_slot->file_ptr = 0;
8613 io_rsrc_node_switch(ctx, ctx->file_data);
8616 io_ring_submit_unlock(ctx, needs_lock);
8620 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
8621 struct io_uring_rsrc_update2 *up,
8624 u64 __user *tags = u64_to_user_ptr(up->tags);
8625 __s32 __user *fds = u64_to_user_ptr(up->data);
8626 struct io_rsrc_data *data = ctx->file_data;
8627 struct io_fixed_file *file_slot;
8631 bool needs_switch = false;
8633 if (!ctx->file_data)
8635 if (up->offset + nr_args > ctx->nr_user_files)
8638 for (done = 0; done < nr_args; done++) {
8641 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
8642 copy_from_user(&fd, &fds[done], sizeof(fd))) {
8646 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
8650 if (fd == IORING_REGISTER_FILES_SKIP)
8653 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
8654 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8656 if (file_slot->file_ptr) {
8657 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8658 err = io_queue_rsrc_removal(data, up->offset + done,
8659 ctx->rsrc_node, file);
8662 file_slot->file_ptr = 0;
8663 needs_switch = true;
8672 * Don't allow io_uring instances to be registered. If
8673 * UNIX isn't enabled, then this causes a reference
8674 * cycle and this instance can never get freed. If UNIX
8675 * is enabled we'll handle it just fine, but there's
8676 * still no point in allowing a ring fd as it doesn't
8677 * support regular read/write anyway.
8679 if (file->f_op == &io_uring_fops) {
8684 *io_get_tag_slot(data, up->offset + done) = tag;
8685 io_fixed_file_set(file_slot, file);
8686 err = io_sqe_file_register(ctx, file, i);
8688 file_slot->file_ptr = 0;
8696 io_rsrc_node_switch(ctx, data);
8697 return done ? done : err;
8700 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
8701 struct task_struct *task)
8703 struct io_wq_hash *hash;
8704 struct io_wq_data data;
8705 unsigned int concurrency;
8707 mutex_lock(&ctx->uring_lock);
8708 hash = ctx->hash_map;
8710 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
8712 mutex_unlock(&ctx->uring_lock);
8713 return ERR_PTR(-ENOMEM);
8715 refcount_set(&hash->refs, 1);
8716 init_waitqueue_head(&hash->wait);
8717 ctx->hash_map = hash;
8719 mutex_unlock(&ctx->uring_lock);
8723 data.free_work = io_wq_free_work;
8724 data.do_work = io_wq_submit_work;
8726 /* Do QD, or 4 * CPUS, whatever is smallest */
8727 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8729 return io_wq_create(concurrency, &data);
8732 static __cold int io_uring_alloc_task_context(struct task_struct *task,
8733 struct io_ring_ctx *ctx)
8735 struct io_uring_task *tctx;
8738 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
8739 if (unlikely(!tctx))
8742 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8743 if (unlikely(ret)) {
8748 tctx->io_wq = io_init_wq_offload(ctx, task);
8749 if (IS_ERR(tctx->io_wq)) {
8750 ret = PTR_ERR(tctx->io_wq);
8751 percpu_counter_destroy(&tctx->inflight);
8757 init_waitqueue_head(&tctx->wait);
8758 atomic_set(&tctx->in_idle, 0);
8759 atomic_set(&tctx->inflight_tracked, 0);
8760 task->io_uring = tctx;
8761 spin_lock_init(&tctx->task_lock);
8762 INIT_WQ_LIST(&tctx->task_list);
8763 INIT_WQ_LIST(&tctx->prior_task_list);
8764 init_task_work(&tctx->task_work, tctx_task_work);
8768 void __io_uring_free(struct task_struct *tsk)
8770 struct io_uring_task *tctx = tsk->io_uring;
8772 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8773 WARN_ON_ONCE(tctx->io_wq);
8774 WARN_ON_ONCE(tctx->cached_refs);
8776 percpu_counter_destroy(&tctx->inflight);
8778 tsk->io_uring = NULL;
8781 static __cold int io_sq_offload_create(struct io_ring_ctx *ctx,
8782 struct io_uring_params *p)
8786 /* Retain compatibility with failing for an invalid attach attempt */
8787 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8788 IORING_SETUP_ATTACH_WQ) {
8791 f = fdget(p->wq_fd);
8794 if (f.file->f_op != &io_uring_fops) {
8800 if (ctx->flags & IORING_SETUP_SQPOLL) {
8801 struct task_struct *tsk;
8802 struct io_sq_data *sqd;
8805 ret = security_uring_sqpoll();
8809 sqd = io_get_sq_data(p, &attached);
8815 ctx->sq_creds = get_current_cred();
8817 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8818 if (!ctx->sq_thread_idle)
8819 ctx->sq_thread_idle = HZ;
8821 io_sq_thread_park(sqd);
8822 list_add(&ctx->sqd_list, &sqd->ctx_list);
8823 io_sqd_update_thread_idle(sqd);
8824 /* don't attach to a dying SQPOLL thread, would be racy */
8825 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8826 io_sq_thread_unpark(sqd);
8833 if (p->flags & IORING_SETUP_SQ_AFF) {
8834 int cpu = p->sq_thread_cpu;
8837 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8844 sqd->task_pid = current->pid;
8845 sqd->task_tgid = current->tgid;
8846 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8853 ret = io_uring_alloc_task_context(tsk, ctx);
8854 wake_up_new_task(tsk);
8857 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8858 /* Can't have SQ_AFF without SQPOLL */
8865 complete(&ctx->sq_data->exited);
8867 io_sq_thread_finish(ctx);
8871 static inline void __io_unaccount_mem(struct user_struct *user,
8872 unsigned long nr_pages)
8874 atomic_long_sub(nr_pages, &user->locked_vm);
8877 static inline int __io_account_mem(struct user_struct *user,
8878 unsigned long nr_pages)
8880 unsigned long page_limit, cur_pages, new_pages;
8882 /* Don't allow more pages than we can safely lock */
8883 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8886 cur_pages = atomic_long_read(&user->locked_vm);
8887 new_pages = cur_pages + nr_pages;
8888 if (new_pages > page_limit)
8890 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8891 new_pages) != cur_pages);
8896 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8899 __io_unaccount_mem(ctx->user, nr_pages);
8901 if (ctx->mm_account)
8902 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8905 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8910 ret = __io_account_mem(ctx->user, nr_pages);
8915 if (ctx->mm_account)
8916 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8921 static void io_mem_free(void *ptr)
8928 page = virt_to_head_page(ptr);
8929 if (put_page_testzero(page))
8930 free_compound_page(page);
8933 static void *io_mem_alloc(size_t size)
8935 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
8937 return (void *) __get_free_pages(gfp, get_order(size));
8940 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8943 struct io_rings *rings;
8944 size_t off, sq_array_size;
8946 off = struct_size(rings, cqes, cq_entries);
8947 if (off == SIZE_MAX)
8951 off = ALIGN(off, SMP_CACHE_BYTES);
8959 sq_array_size = array_size(sizeof(u32), sq_entries);
8960 if (sq_array_size == SIZE_MAX)
8963 if (check_add_overflow(off, sq_array_size, &off))
8969 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8971 struct io_mapped_ubuf *imu = *slot;
8974 if (imu != ctx->dummy_ubuf) {
8975 for (i = 0; i < imu->nr_bvecs; i++)
8976 unpin_user_page(imu->bvec[i].bv_page);
8977 if (imu->acct_pages)
8978 io_unaccount_mem(ctx, imu->acct_pages);
8984 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8986 io_buffer_unmap(ctx, &prsrc->buf);
8990 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8994 for (i = 0; i < ctx->nr_user_bufs; i++)
8995 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8996 kfree(ctx->user_bufs);
8997 io_rsrc_data_free(ctx->buf_data);
8998 ctx->user_bufs = NULL;
8999 ctx->buf_data = NULL;
9000 ctx->nr_user_bufs = 0;
9003 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
9010 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
9012 __io_sqe_buffers_unregister(ctx);
9016 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
9017 void __user *arg, unsigned index)
9019 struct iovec __user *src;
9021 #ifdef CONFIG_COMPAT
9023 struct compat_iovec __user *ciovs;
9024 struct compat_iovec ciov;
9026 ciovs = (struct compat_iovec __user *) arg;
9027 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
9030 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
9031 dst->iov_len = ciov.iov_len;
9035 src = (struct iovec __user *) arg;
9036 if (copy_from_user(dst, &src[index], sizeof(*dst)))
9042 * Not super efficient, but this is just a registration time. And we do cache
9043 * the last compound head, so generally we'll only do a full search if we don't
9046 * We check if the given compound head page has already been accounted, to
9047 * avoid double accounting it. This allows us to account the full size of the
9048 * page, not just the constituent pages of a huge page.
9050 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
9051 int nr_pages, struct page *hpage)
9055 /* check current page array */
9056 for (i = 0; i < nr_pages; i++) {
9057 if (!PageCompound(pages[i]))
9059 if (compound_head(pages[i]) == hpage)
9063 /* check previously registered pages */
9064 for (i = 0; i < ctx->nr_user_bufs; i++) {
9065 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
9067 for (j = 0; j < imu->nr_bvecs; j++) {
9068 if (!PageCompound(imu->bvec[j].bv_page))
9070 if (compound_head(imu->bvec[j].bv_page) == hpage)
9078 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
9079 int nr_pages, struct io_mapped_ubuf *imu,
9080 struct page **last_hpage)
9084 imu->acct_pages = 0;
9085 for (i = 0; i < nr_pages; i++) {
9086 if (!PageCompound(pages[i])) {
9091 hpage = compound_head(pages[i]);
9092 if (hpage == *last_hpage)
9094 *last_hpage = hpage;
9095 if (headpage_already_acct(ctx, pages, i, hpage))
9097 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
9101 if (!imu->acct_pages)
9104 ret = io_account_mem(ctx, imu->acct_pages);
9106 imu->acct_pages = 0;
9110 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
9111 struct io_mapped_ubuf **pimu,
9112 struct page **last_hpage)
9114 struct io_mapped_ubuf *imu = NULL;
9115 struct vm_area_struct **vmas = NULL;
9116 struct page **pages = NULL;
9117 unsigned long off, start, end, ubuf;
9119 int ret, pret, nr_pages, i;
9121 if (!iov->iov_base) {
9122 *pimu = ctx->dummy_ubuf;
9126 ubuf = (unsigned long) iov->iov_base;
9127 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
9128 start = ubuf >> PAGE_SHIFT;
9129 nr_pages = end - start;
9134 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
9138 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
9143 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
9148 mmap_read_lock(current->mm);
9149 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
9151 if (pret == nr_pages) {
9152 /* don't support file backed memory */
9153 for (i = 0; i < nr_pages; i++) {
9154 struct vm_area_struct *vma = vmas[i];
9156 if (vma_is_shmem(vma))
9159 !is_file_hugepages(vma->vm_file)) {
9165 ret = pret < 0 ? pret : -EFAULT;
9167 mmap_read_unlock(current->mm);
9170 * if we did partial map, or found file backed vmas,
9171 * release any pages we did get
9174 unpin_user_pages(pages, pret);
9178 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
9180 unpin_user_pages(pages, pret);
9184 off = ubuf & ~PAGE_MASK;
9185 size = iov->iov_len;
9186 for (i = 0; i < nr_pages; i++) {
9189 vec_len = min_t(size_t, size, PAGE_SIZE - off);
9190 imu->bvec[i].bv_page = pages[i];
9191 imu->bvec[i].bv_len = vec_len;
9192 imu->bvec[i].bv_offset = off;
9196 /* store original address for later verification */
9198 imu->ubuf_end = ubuf + iov->iov_len;
9199 imu->nr_bvecs = nr_pages;
9210 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
9212 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
9213 return ctx->user_bufs ? 0 : -ENOMEM;
9216 static int io_buffer_validate(struct iovec *iov)
9218 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
9221 * Don't impose further limits on the size and buffer
9222 * constraints here, we'll -EINVAL later when IO is
9223 * submitted if they are wrong.
9226 return iov->iov_len ? -EFAULT : 0;
9230 /* arbitrary limit, but we need something */
9231 if (iov->iov_len > SZ_1G)
9234 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
9240 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
9241 unsigned int nr_args, u64 __user *tags)
9243 struct page *last_hpage = NULL;
9244 struct io_rsrc_data *data;
9250 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
9252 ret = io_rsrc_node_switch_start(ctx);
9255 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
9258 ret = io_buffers_map_alloc(ctx, nr_args);
9260 io_rsrc_data_free(data);
9264 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
9265 ret = io_copy_iov(ctx, &iov, arg, i);
9268 ret = io_buffer_validate(&iov);
9271 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
9276 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
9282 WARN_ON_ONCE(ctx->buf_data);
9284 ctx->buf_data = data;
9286 __io_sqe_buffers_unregister(ctx);
9288 io_rsrc_node_switch(ctx, NULL);
9292 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
9293 struct io_uring_rsrc_update2 *up,
9294 unsigned int nr_args)
9296 u64 __user *tags = u64_to_user_ptr(up->tags);
9297 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
9298 struct page *last_hpage = NULL;
9299 bool needs_switch = false;
9305 if (up->offset + nr_args > ctx->nr_user_bufs)
9308 for (done = 0; done < nr_args; done++) {
9309 struct io_mapped_ubuf *imu;
9310 int offset = up->offset + done;
9313 err = io_copy_iov(ctx, &iov, iovs, done);
9316 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
9320 err = io_buffer_validate(&iov);
9323 if (!iov.iov_base && tag) {
9327 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
9331 i = array_index_nospec(offset, ctx->nr_user_bufs);
9332 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
9333 err = io_queue_rsrc_removal(ctx->buf_data, offset,
9334 ctx->rsrc_node, ctx->user_bufs[i]);
9335 if (unlikely(err)) {
9336 io_buffer_unmap(ctx, &imu);
9339 ctx->user_bufs[i] = NULL;
9340 needs_switch = true;
9343 ctx->user_bufs[i] = imu;
9344 *io_get_tag_slot(ctx->buf_data, offset) = tag;
9348 io_rsrc_node_switch(ctx, ctx->buf_data);
9349 return done ? done : err;
9352 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
9354 __s32 __user *fds = arg;
9360 if (copy_from_user(&fd, fds, sizeof(*fds)))
9363 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
9364 if (IS_ERR(ctx->cq_ev_fd)) {
9365 int ret = PTR_ERR(ctx->cq_ev_fd);
9367 ctx->cq_ev_fd = NULL;
9374 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
9376 if (ctx->cq_ev_fd) {
9377 eventfd_ctx_put(ctx->cq_ev_fd);
9378 ctx->cq_ev_fd = NULL;
9385 static void io_destroy_buffers(struct io_ring_ctx *ctx)
9387 struct io_buffer *buf;
9388 unsigned long index;
9390 xa_for_each(&ctx->io_buffers, index, buf)
9391 __io_remove_buffers(ctx, buf, index, -1U);
9394 static void io_req_caches_free(struct io_ring_ctx *ctx)
9396 struct io_submit_state *state = &ctx->submit_state;
9399 mutex_lock(&ctx->uring_lock);
9400 io_flush_cached_locked_reqs(ctx, state);
9402 while (state->free_list.next) {
9403 struct io_wq_work_node *node;
9404 struct io_kiocb *req;
9406 node = wq_stack_extract(&state->free_list);
9407 req = container_of(node, struct io_kiocb, comp_list);
9408 kmem_cache_free(req_cachep, req);
9412 percpu_ref_put_many(&ctx->refs, nr);
9413 mutex_unlock(&ctx->uring_lock);
9416 static void io_wait_rsrc_data(struct io_rsrc_data *data)
9418 if (data && !atomic_dec_and_test(&data->refs))
9419 wait_for_completion(&data->done);
9422 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
9424 io_sq_thread_finish(ctx);
9426 if (ctx->mm_account) {
9427 mmdrop(ctx->mm_account);
9428 ctx->mm_account = NULL;
9431 io_rsrc_refs_drop(ctx);
9432 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
9433 io_wait_rsrc_data(ctx->buf_data);
9434 io_wait_rsrc_data(ctx->file_data);
9436 mutex_lock(&ctx->uring_lock);
9438 __io_sqe_buffers_unregister(ctx);
9440 __io_sqe_files_unregister(ctx);
9442 __io_cqring_overflow_flush(ctx, true);
9443 mutex_unlock(&ctx->uring_lock);
9444 io_eventfd_unregister(ctx);
9445 io_destroy_buffers(ctx);
9447 put_cred(ctx->sq_creds);
9449 /* there are no registered resources left, nobody uses it */
9451 io_rsrc_node_destroy(ctx->rsrc_node);
9452 if (ctx->rsrc_backup_node)
9453 io_rsrc_node_destroy(ctx->rsrc_backup_node);
9454 flush_delayed_work(&ctx->rsrc_put_work);
9455 flush_delayed_work(&ctx->fallback_work);
9457 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
9458 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
9460 #if defined(CONFIG_UNIX)
9461 if (ctx->ring_sock) {
9462 ctx->ring_sock->file = NULL; /* so that iput() is called */
9463 sock_release(ctx->ring_sock);
9466 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
9468 io_mem_free(ctx->rings);
9469 io_mem_free(ctx->sq_sqes);
9471 percpu_ref_exit(&ctx->refs);
9472 free_uid(ctx->user);
9473 io_req_caches_free(ctx);
9475 io_wq_put_hash(ctx->hash_map);
9476 kfree(ctx->cancel_hash);
9477 kfree(ctx->dummy_ubuf);
9481 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
9483 struct io_ring_ctx *ctx = file->private_data;
9486 poll_wait(file, &ctx->cq_wait, wait);
9488 * synchronizes with barrier from wq_has_sleeper call in
9492 if (!io_sqring_full(ctx))
9493 mask |= EPOLLOUT | EPOLLWRNORM;
9496 * Don't flush cqring overflow list here, just do a simple check.
9497 * Otherwise there could possible be ABBA deadlock:
9500 * lock(&ctx->uring_lock);
9502 * lock(&ctx->uring_lock);
9505 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
9506 * pushs them to do the flush.
9508 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
9509 mask |= EPOLLIN | EPOLLRDNORM;
9514 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9516 const struct cred *creds;
9518 creds = xa_erase(&ctx->personalities, id);
9527 struct io_tctx_exit {
9528 struct callback_head task_work;
9529 struct completion completion;
9530 struct io_ring_ctx *ctx;
9533 static __cold void io_tctx_exit_cb(struct callback_head *cb)
9535 struct io_uring_task *tctx = current->io_uring;
9536 struct io_tctx_exit *work;
9538 work = container_of(cb, struct io_tctx_exit, task_work);
9540 * When @in_idle, we're in cancellation and it's racy to remove the
9541 * node. It'll be removed by the end of cancellation, just ignore it.
9543 if (!atomic_read(&tctx->in_idle))
9544 io_uring_del_tctx_node((unsigned long)work->ctx);
9545 complete(&work->completion);
9548 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
9550 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9552 return req->ctx == data;
9555 static __cold void io_ring_exit_work(struct work_struct *work)
9557 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
9558 unsigned long timeout = jiffies + HZ * 60 * 5;
9559 unsigned long interval = HZ / 20;
9560 struct io_tctx_exit exit;
9561 struct io_tctx_node *node;
9565 * If we're doing polled IO and end up having requests being
9566 * submitted async (out-of-line), then completions can come in while
9567 * we're waiting for refs to drop. We need to reap these manually,
9568 * as nobody else will be looking for them.
9571 io_uring_try_cancel_requests(ctx, NULL, true);
9573 struct io_sq_data *sqd = ctx->sq_data;
9574 struct task_struct *tsk;
9576 io_sq_thread_park(sqd);
9578 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
9579 io_wq_cancel_cb(tsk->io_uring->io_wq,
9580 io_cancel_ctx_cb, ctx, true);
9581 io_sq_thread_unpark(sqd);
9584 io_req_caches_free(ctx);
9586 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
9587 /* there is little hope left, don't run it too often */
9590 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
9592 init_completion(&exit.completion);
9593 init_task_work(&exit.task_work, io_tctx_exit_cb);
9596 * Some may use context even when all refs and requests have been put,
9597 * and they are free to do so while still holding uring_lock or
9598 * completion_lock, see io_req_task_submit(). Apart from other work,
9599 * this lock/unlock section also waits them to finish.
9601 mutex_lock(&ctx->uring_lock);
9602 while (!list_empty(&ctx->tctx_list)) {
9603 WARN_ON_ONCE(time_after(jiffies, timeout));
9605 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
9607 /* don't spin on a single task if cancellation failed */
9608 list_rotate_left(&ctx->tctx_list);
9609 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
9610 if (WARN_ON_ONCE(ret))
9613 mutex_unlock(&ctx->uring_lock);
9614 wait_for_completion(&exit.completion);
9615 mutex_lock(&ctx->uring_lock);
9617 mutex_unlock(&ctx->uring_lock);
9618 spin_lock(&ctx->completion_lock);
9619 spin_unlock(&ctx->completion_lock);
9621 io_ring_ctx_free(ctx);
9624 /* Returns true if we found and killed one or more timeouts */
9625 static __cold bool io_kill_timeouts(struct io_ring_ctx *ctx,
9626 struct task_struct *tsk, bool cancel_all)
9628 struct io_kiocb *req, *tmp;
9631 spin_lock(&ctx->completion_lock);
9632 spin_lock_irq(&ctx->timeout_lock);
9633 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
9634 if (io_match_task(req, tsk, cancel_all)) {
9635 io_kill_timeout(req, -ECANCELED);
9639 spin_unlock_irq(&ctx->timeout_lock);
9641 io_commit_cqring(ctx);
9642 spin_unlock(&ctx->completion_lock);
9644 io_cqring_ev_posted(ctx);
9645 return canceled != 0;
9648 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
9650 unsigned long index;
9651 struct creds *creds;
9653 mutex_lock(&ctx->uring_lock);
9654 percpu_ref_kill(&ctx->refs);
9656 __io_cqring_overflow_flush(ctx, true);
9657 xa_for_each(&ctx->personalities, index, creds)
9658 io_unregister_personality(ctx, index);
9659 mutex_unlock(&ctx->uring_lock);
9661 io_kill_timeouts(ctx, NULL, true);
9662 io_poll_remove_all(ctx, NULL, true);
9664 /* if we failed setting up the ctx, we might not have any rings */
9665 io_iopoll_try_reap_events(ctx);
9667 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
9669 * Use system_unbound_wq to avoid spawning tons of event kworkers
9670 * if we're exiting a ton of rings at the same time. It just adds
9671 * noise and overhead, there's no discernable change in runtime
9672 * over using system_wq.
9674 queue_work(system_unbound_wq, &ctx->exit_work);
9677 static int io_uring_release(struct inode *inode, struct file *file)
9679 struct io_ring_ctx *ctx = file->private_data;
9681 file->private_data = NULL;
9682 io_ring_ctx_wait_and_kill(ctx);
9686 struct io_task_cancel {
9687 struct task_struct *task;
9691 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
9693 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9694 struct io_task_cancel *cancel = data;
9696 return io_match_task_safe(req, cancel->task, cancel->all);
9699 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
9700 struct task_struct *task,
9703 struct io_defer_entry *de;
9706 spin_lock(&ctx->completion_lock);
9707 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
9708 if (io_match_task_safe(de->req, task, cancel_all)) {
9709 list_cut_position(&list, &ctx->defer_list, &de->list);
9713 spin_unlock(&ctx->completion_lock);
9714 if (list_empty(&list))
9717 while (!list_empty(&list)) {
9718 de = list_first_entry(&list, struct io_defer_entry, list);
9719 list_del_init(&de->list);
9720 io_req_complete_failed(de->req, -ECANCELED);
9726 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
9728 struct io_tctx_node *node;
9729 enum io_wq_cancel cret;
9732 mutex_lock(&ctx->uring_lock);
9733 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
9734 struct io_uring_task *tctx = node->task->io_uring;
9737 * io_wq will stay alive while we hold uring_lock, because it's
9738 * killed after ctx nodes, which requires to take the lock.
9740 if (!tctx || !tctx->io_wq)
9742 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
9743 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9745 mutex_unlock(&ctx->uring_lock);
9750 static __cold void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9751 struct task_struct *task,
9754 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9755 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9758 enum io_wq_cancel cret;
9762 ret |= io_uring_try_cancel_iowq(ctx);
9763 } else if (tctx && tctx->io_wq) {
9765 * Cancels requests of all rings, not only @ctx, but
9766 * it's fine as the task is in exit/exec.
9768 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9770 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9773 /* SQPOLL thread does its own polling */
9774 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9775 (ctx->sq_data && ctx->sq_data->thread == current)) {
9776 while (!wq_list_empty(&ctx->iopoll_list)) {
9777 io_iopoll_try_reap_events(ctx);
9782 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9783 ret |= io_poll_remove_all(ctx, task, cancel_all);
9784 ret |= io_kill_timeouts(ctx, task, cancel_all);
9786 ret |= io_run_task_work();
9793 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9795 struct io_uring_task *tctx = current->io_uring;
9796 struct io_tctx_node *node;
9799 if (unlikely(!tctx)) {
9800 ret = io_uring_alloc_task_context(current, ctx);
9804 tctx = current->io_uring;
9805 if (ctx->iowq_limits_set) {
9806 unsigned int limits[2] = { ctx->iowq_limits[0],
9807 ctx->iowq_limits[1], };
9809 ret = io_wq_max_workers(tctx->io_wq, limits);
9814 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9815 node = kmalloc(sizeof(*node), GFP_KERNEL);
9819 node->task = current;
9821 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9828 mutex_lock(&ctx->uring_lock);
9829 list_add(&node->ctx_node, &ctx->tctx_list);
9830 mutex_unlock(&ctx->uring_lock);
9837 * Note that this task has used io_uring. We use it for cancelation purposes.
9839 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9841 struct io_uring_task *tctx = current->io_uring;
9843 if (likely(tctx && tctx->last == ctx))
9845 return __io_uring_add_tctx_node(ctx);
9849 * Remove this io_uring_file -> task mapping.
9851 static __cold void io_uring_del_tctx_node(unsigned long index)
9853 struct io_uring_task *tctx = current->io_uring;
9854 struct io_tctx_node *node;
9858 node = xa_erase(&tctx->xa, index);
9862 WARN_ON_ONCE(current != node->task);
9863 WARN_ON_ONCE(list_empty(&node->ctx_node));
9865 mutex_lock(&node->ctx->uring_lock);
9866 list_del(&node->ctx_node);
9867 mutex_unlock(&node->ctx->uring_lock);
9869 if (tctx->last == node->ctx)
9874 static __cold void io_uring_clean_tctx(struct io_uring_task *tctx)
9876 struct io_wq *wq = tctx->io_wq;
9877 struct io_tctx_node *node;
9878 unsigned long index;
9880 xa_for_each(&tctx->xa, index, node) {
9881 io_uring_del_tctx_node(index);
9886 * Must be after io_uring_del_tctx_node() (removes nodes under
9887 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9889 io_wq_put_and_exit(wq);
9894 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9897 return atomic_read(&tctx->inflight_tracked);
9898 return percpu_counter_sum(&tctx->inflight);
9902 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9903 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
9905 static __cold void io_uring_cancel_generic(bool cancel_all,
9906 struct io_sq_data *sqd)
9908 struct io_uring_task *tctx = current->io_uring;
9909 struct io_ring_ctx *ctx;
9913 WARN_ON_ONCE(sqd && sqd->thread != current);
9915 if (!current->io_uring)
9918 io_wq_exit_start(tctx->io_wq);
9920 atomic_inc(&tctx->in_idle);
9922 io_uring_drop_tctx_refs(current);
9923 /* read completions before cancelations */
9924 inflight = tctx_inflight(tctx, !cancel_all);
9929 struct io_tctx_node *node;
9930 unsigned long index;
9932 xa_for_each(&tctx->xa, index, node) {
9933 /* sqpoll task will cancel all its requests */
9934 if (node->ctx->sq_data)
9936 io_uring_try_cancel_requests(node->ctx, current,
9940 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9941 io_uring_try_cancel_requests(ctx, current,
9945 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
9947 io_uring_drop_tctx_refs(current);
9950 * If we've seen completions, retry without waiting. This
9951 * avoids a race where a completion comes in before we did
9952 * prepare_to_wait().
9954 if (inflight == tctx_inflight(tctx, !cancel_all))
9956 finish_wait(&tctx->wait, &wait);
9959 io_uring_clean_tctx(tctx);
9962 * We shouldn't run task_works after cancel, so just leave
9963 * ->in_idle set for normal exit.
9965 atomic_dec(&tctx->in_idle);
9966 /* for exec all current's requests should be gone, kill tctx */
9967 __io_uring_free(current);
9971 void __io_uring_cancel(bool cancel_all)
9973 io_uring_cancel_generic(cancel_all, NULL);
9976 static void *io_uring_validate_mmap_request(struct file *file,
9977 loff_t pgoff, size_t sz)
9979 struct io_ring_ctx *ctx = file->private_data;
9980 loff_t offset = pgoff << PAGE_SHIFT;
9985 case IORING_OFF_SQ_RING:
9986 case IORING_OFF_CQ_RING:
9989 case IORING_OFF_SQES:
9993 return ERR_PTR(-EINVAL);
9996 page = virt_to_head_page(ptr);
9997 if (sz > page_size(page))
9998 return ERR_PTR(-EINVAL);
10005 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
10007 size_t sz = vma->vm_end - vma->vm_start;
10011 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
10013 return PTR_ERR(ptr);
10015 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
10016 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
10019 #else /* !CONFIG_MMU */
10021 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
10023 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
10026 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
10028 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
10031 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
10032 unsigned long addr, unsigned long len,
10033 unsigned long pgoff, unsigned long flags)
10037 ptr = io_uring_validate_mmap_request(file, pgoff, len);
10039 return PTR_ERR(ptr);
10041 return (unsigned long) ptr;
10044 #endif /* !CONFIG_MMU */
10046 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
10051 if (!io_sqring_full(ctx))
10053 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
10055 if (!io_sqring_full(ctx))
10058 } while (!signal_pending(current));
10060 finish_wait(&ctx->sqo_sq_wait, &wait);
10064 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
10065 struct __kernel_timespec __user **ts,
10066 const sigset_t __user **sig)
10068 struct io_uring_getevents_arg arg;
10071 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
10072 * is just a pointer to the sigset_t.
10074 if (!(flags & IORING_ENTER_EXT_ARG)) {
10075 *sig = (const sigset_t __user *) argp;
10081 * EXT_ARG is set - ensure we agree on the size of it and copy in our
10082 * timespec and sigset_t pointers if good.
10084 if (*argsz != sizeof(arg))
10086 if (copy_from_user(&arg, argp, sizeof(arg)))
10088 *sig = u64_to_user_ptr(arg.sigmask);
10089 *argsz = arg.sigmask_sz;
10090 *ts = u64_to_user_ptr(arg.ts);
10094 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
10095 u32, min_complete, u32, flags, const void __user *, argp,
10098 struct io_ring_ctx *ctx;
10103 io_run_task_work();
10105 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
10106 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
10110 if (unlikely(!f.file))
10114 if (unlikely(f.file->f_op != &io_uring_fops))
10118 ctx = f.file->private_data;
10119 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
10123 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
10127 * For SQ polling, the thread will do all submissions and completions.
10128 * Just return the requested submit count, and wake the thread if
10129 * we were asked to.
10132 if (ctx->flags & IORING_SETUP_SQPOLL) {
10133 io_cqring_overflow_flush(ctx);
10135 if (unlikely(ctx->sq_data->thread == NULL)) {
10139 if (flags & IORING_ENTER_SQ_WAKEUP)
10140 wake_up(&ctx->sq_data->wait);
10141 if (flags & IORING_ENTER_SQ_WAIT) {
10142 ret = io_sqpoll_wait_sq(ctx);
10146 submitted = to_submit;
10147 } else if (to_submit) {
10148 ret = io_uring_add_tctx_node(ctx);
10151 mutex_lock(&ctx->uring_lock);
10152 submitted = io_submit_sqes(ctx, to_submit);
10153 mutex_unlock(&ctx->uring_lock);
10155 if (submitted != to_submit)
10158 if (flags & IORING_ENTER_GETEVENTS) {
10159 const sigset_t __user *sig;
10160 struct __kernel_timespec __user *ts;
10162 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
10166 min_complete = min(min_complete, ctx->cq_entries);
10169 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
10170 * space applications don't need to do io completion events
10171 * polling again, they can rely on io_sq_thread to do polling
10172 * work, which can reduce cpu usage and uring_lock contention.
10174 if (ctx->flags & IORING_SETUP_IOPOLL &&
10175 !(ctx->flags & IORING_SETUP_SQPOLL)) {
10176 ret = io_iopoll_check(ctx, min_complete);
10178 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
10183 percpu_ref_put(&ctx->refs);
10186 return submitted ? submitted : ret;
10189 #ifdef CONFIG_PROC_FS
10190 static __cold int io_uring_show_cred(struct seq_file *m, unsigned int id,
10191 const struct cred *cred)
10193 struct user_namespace *uns = seq_user_ns(m);
10194 struct group_info *gi;
10199 seq_printf(m, "%5d\n", id);
10200 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
10201 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
10202 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
10203 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
10204 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
10205 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
10206 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
10207 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
10208 seq_puts(m, "\n\tGroups:\t");
10209 gi = cred->group_info;
10210 for (g = 0; g < gi->ngroups; g++) {
10211 seq_put_decimal_ull(m, g ? " " : "",
10212 from_kgid_munged(uns, gi->gid[g]));
10214 seq_puts(m, "\n\tCapEff:\t");
10215 cap = cred->cap_effective;
10216 CAP_FOR_EACH_U32(__capi)
10217 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
10222 static __cold void __io_uring_show_fdinfo(struct io_ring_ctx *ctx,
10223 struct seq_file *m)
10225 struct io_sq_data *sq = NULL;
10226 struct io_overflow_cqe *ocqe;
10227 struct io_rings *r = ctx->rings;
10228 unsigned int sq_mask = ctx->sq_entries - 1, cq_mask = ctx->cq_entries - 1;
10229 unsigned int sq_head = READ_ONCE(r->sq.head);
10230 unsigned int sq_tail = READ_ONCE(r->sq.tail);
10231 unsigned int cq_head = READ_ONCE(r->cq.head);
10232 unsigned int cq_tail = READ_ONCE(r->cq.tail);
10233 unsigned int sq_entries, cq_entries;
10238 * we may get imprecise sqe and cqe info if uring is actively running
10239 * since we get cached_sq_head and cached_cq_tail without uring_lock
10240 * and sq_tail and cq_head are changed by userspace. But it's ok since
10241 * we usually use these info when it is stuck.
10243 seq_printf(m, "SqMask:\t0x%x\n", sq_mask);
10244 seq_printf(m, "SqHead:\t%u\n", sq_head);
10245 seq_printf(m, "SqTail:\t%u\n", sq_tail);
10246 seq_printf(m, "CachedSqHead:\t%u\n", ctx->cached_sq_head);
10247 seq_printf(m, "CqMask:\t0x%x\n", cq_mask);
10248 seq_printf(m, "CqHead:\t%u\n", cq_head);
10249 seq_printf(m, "CqTail:\t%u\n", cq_tail);
10250 seq_printf(m, "CachedCqTail:\t%u\n", ctx->cached_cq_tail);
10251 seq_printf(m, "SQEs:\t%u\n", sq_tail - ctx->cached_sq_head);
10252 sq_entries = min(sq_tail - sq_head, ctx->sq_entries);
10253 for (i = 0; i < sq_entries; i++) {
10254 unsigned int entry = i + sq_head;
10255 unsigned int sq_idx = READ_ONCE(ctx->sq_array[entry & sq_mask]);
10256 struct io_uring_sqe *sqe;
10258 if (sq_idx > sq_mask)
10260 sqe = &ctx->sq_sqes[sq_idx];
10261 seq_printf(m, "%5u: opcode:%d, fd:%d, flags:%x, user_data:%llu\n",
10262 sq_idx, sqe->opcode, sqe->fd, sqe->flags,
10265 seq_printf(m, "CQEs:\t%u\n", cq_tail - cq_head);
10266 cq_entries = min(cq_tail - cq_head, ctx->cq_entries);
10267 for (i = 0; i < cq_entries; i++) {
10268 unsigned int entry = i + cq_head;
10269 struct io_uring_cqe *cqe = &r->cqes[entry & cq_mask];
10271 seq_printf(m, "%5u: user_data:%llu, res:%d, flag:%x\n",
10272 entry & cq_mask, cqe->user_data, cqe->res,
10277 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
10278 * since fdinfo case grabs it in the opposite direction of normal use
10279 * cases. If we fail to get the lock, we just don't iterate any
10280 * structures that could be going away outside the io_uring mutex.
10282 has_lock = mutex_trylock(&ctx->uring_lock);
10284 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
10290 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
10291 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
10292 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
10293 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
10294 struct file *f = io_file_from_index(ctx, i);
10297 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
10299 seq_printf(m, "%5u: <none>\n", i);
10301 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
10302 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
10303 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
10304 unsigned int len = buf->ubuf_end - buf->ubuf;
10306 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
10308 if (has_lock && !xa_empty(&ctx->personalities)) {
10309 unsigned long index;
10310 const struct cred *cred;
10312 seq_printf(m, "Personalities:\n");
10313 xa_for_each(&ctx->personalities, index, cred)
10314 io_uring_show_cred(m, index, cred);
10317 mutex_unlock(&ctx->uring_lock);
10319 seq_puts(m, "PollList:\n");
10320 spin_lock(&ctx->completion_lock);
10321 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
10322 struct hlist_head *list = &ctx->cancel_hash[i];
10323 struct io_kiocb *req;
10325 hlist_for_each_entry(req, list, hash_node)
10326 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
10327 req->task->task_works != NULL);
10330 seq_puts(m, "CqOverflowList:\n");
10331 list_for_each_entry(ocqe, &ctx->cq_overflow_list, list) {
10332 struct io_uring_cqe *cqe = &ocqe->cqe;
10334 seq_printf(m, " user_data=%llu, res=%d, flags=%x\n",
10335 cqe->user_data, cqe->res, cqe->flags);
10339 spin_unlock(&ctx->completion_lock);
10342 static __cold void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
10344 struct io_ring_ctx *ctx = f->private_data;
10346 if (percpu_ref_tryget(&ctx->refs)) {
10347 __io_uring_show_fdinfo(ctx, m);
10348 percpu_ref_put(&ctx->refs);
10353 static const struct file_operations io_uring_fops = {
10354 .release = io_uring_release,
10355 .mmap = io_uring_mmap,
10357 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
10358 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
10360 .poll = io_uring_poll,
10361 #ifdef CONFIG_PROC_FS
10362 .show_fdinfo = io_uring_show_fdinfo,
10366 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
10367 struct io_uring_params *p)
10369 struct io_rings *rings;
10370 size_t size, sq_array_offset;
10372 /* make sure these are sane, as we already accounted them */
10373 ctx->sq_entries = p->sq_entries;
10374 ctx->cq_entries = p->cq_entries;
10376 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
10377 if (size == SIZE_MAX)
10380 rings = io_mem_alloc(size);
10384 ctx->rings = rings;
10385 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
10386 rings->sq_ring_mask = p->sq_entries - 1;
10387 rings->cq_ring_mask = p->cq_entries - 1;
10388 rings->sq_ring_entries = p->sq_entries;
10389 rings->cq_ring_entries = p->cq_entries;
10391 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
10392 if (size == SIZE_MAX) {
10393 io_mem_free(ctx->rings);
10398 ctx->sq_sqes = io_mem_alloc(size);
10399 if (!ctx->sq_sqes) {
10400 io_mem_free(ctx->rings);
10408 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
10412 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
10416 ret = io_uring_add_tctx_node(ctx);
10421 fd_install(fd, file);
10426 * Allocate an anonymous fd, this is what constitutes the application
10427 * visible backing of an io_uring instance. The application mmaps this
10428 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
10429 * we have to tie this fd to a socket for file garbage collection purposes.
10431 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
10434 #if defined(CONFIG_UNIX)
10437 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
10440 return ERR_PTR(ret);
10443 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
10444 O_RDWR | O_CLOEXEC, NULL);
10445 #if defined(CONFIG_UNIX)
10446 if (IS_ERR(file)) {
10447 sock_release(ctx->ring_sock);
10448 ctx->ring_sock = NULL;
10450 ctx->ring_sock->file = file;
10456 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
10457 struct io_uring_params __user *params)
10459 struct io_ring_ctx *ctx;
10465 if (entries > IORING_MAX_ENTRIES) {
10466 if (!(p->flags & IORING_SETUP_CLAMP))
10468 entries = IORING_MAX_ENTRIES;
10472 * Use twice as many entries for the CQ ring. It's possible for the
10473 * application to drive a higher depth than the size of the SQ ring,
10474 * since the sqes are only used at submission time. This allows for
10475 * some flexibility in overcommitting a bit. If the application has
10476 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
10477 * of CQ ring entries manually.
10479 p->sq_entries = roundup_pow_of_two(entries);
10480 if (p->flags & IORING_SETUP_CQSIZE) {
10482 * If IORING_SETUP_CQSIZE is set, we do the same roundup
10483 * to a power-of-two, if it isn't already. We do NOT impose
10484 * any cq vs sq ring sizing.
10486 if (!p->cq_entries)
10488 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
10489 if (!(p->flags & IORING_SETUP_CLAMP))
10491 p->cq_entries = IORING_MAX_CQ_ENTRIES;
10493 p->cq_entries = roundup_pow_of_two(p->cq_entries);
10494 if (p->cq_entries < p->sq_entries)
10497 p->cq_entries = 2 * p->sq_entries;
10500 ctx = io_ring_ctx_alloc(p);
10503 ctx->compat = in_compat_syscall();
10504 if (!capable(CAP_IPC_LOCK))
10505 ctx->user = get_uid(current_user());
10508 * This is just grabbed for accounting purposes. When a process exits,
10509 * the mm is exited and dropped before the files, hence we need to hang
10510 * on to this mm purely for the purposes of being able to unaccount
10511 * memory (locked/pinned vm). It's not used for anything else.
10513 mmgrab(current->mm);
10514 ctx->mm_account = current->mm;
10516 ret = io_allocate_scq_urings(ctx, p);
10520 ret = io_sq_offload_create(ctx, p);
10523 /* always set a rsrc node */
10524 ret = io_rsrc_node_switch_start(ctx);
10527 io_rsrc_node_switch(ctx, NULL);
10529 memset(&p->sq_off, 0, sizeof(p->sq_off));
10530 p->sq_off.head = offsetof(struct io_rings, sq.head);
10531 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
10532 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
10533 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
10534 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
10535 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
10536 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
10538 memset(&p->cq_off, 0, sizeof(p->cq_off));
10539 p->cq_off.head = offsetof(struct io_rings, cq.head);
10540 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
10541 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
10542 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
10543 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
10544 p->cq_off.cqes = offsetof(struct io_rings, cqes);
10545 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
10547 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
10548 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
10549 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
10550 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
10551 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
10552 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP;
10554 if (copy_to_user(params, p, sizeof(*p))) {
10559 file = io_uring_get_file(ctx);
10560 if (IS_ERR(file)) {
10561 ret = PTR_ERR(file);
10566 * Install ring fd as the very last thing, so we don't risk someone
10567 * having closed it before we finish setup
10569 ret = io_uring_install_fd(ctx, file);
10571 /* fput will clean it up */
10576 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
10579 io_ring_ctx_wait_and_kill(ctx);
10584 * Sets up an aio uring context, and returns the fd. Applications asks for a
10585 * ring size, we return the actual sq/cq ring sizes (among other things) in the
10586 * params structure passed in.
10588 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
10590 struct io_uring_params p;
10593 if (copy_from_user(&p, params, sizeof(p)))
10595 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
10600 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
10601 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
10602 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
10603 IORING_SETUP_R_DISABLED))
10606 return io_uring_create(entries, &p, params);
10609 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
10610 struct io_uring_params __user *, params)
10612 return io_uring_setup(entries, params);
10615 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
10618 struct io_uring_probe *p;
10622 size = struct_size(p, ops, nr_args);
10623 if (size == SIZE_MAX)
10625 p = kzalloc(size, GFP_KERNEL);
10630 if (copy_from_user(p, arg, size))
10633 if (memchr_inv(p, 0, size))
10636 p->last_op = IORING_OP_LAST - 1;
10637 if (nr_args > IORING_OP_LAST)
10638 nr_args = IORING_OP_LAST;
10640 for (i = 0; i < nr_args; i++) {
10642 if (!io_op_defs[i].not_supported)
10643 p->ops[i].flags = IO_URING_OP_SUPPORTED;
10648 if (copy_to_user(arg, p, size))
10655 static int io_register_personality(struct io_ring_ctx *ctx)
10657 const struct cred *creds;
10661 creds = get_current_cred();
10663 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
10664 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
10672 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
10673 void __user *arg, unsigned int nr_args)
10675 struct io_uring_restriction *res;
10679 /* Restrictions allowed only if rings started disabled */
10680 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10683 /* We allow only a single restrictions registration */
10684 if (ctx->restrictions.registered)
10687 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
10690 size = array_size(nr_args, sizeof(*res));
10691 if (size == SIZE_MAX)
10694 res = memdup_user(arg, size);
10696 return PTR_ERR(res);
10700 for (i = 0; i < nr_args; i++) {
10701 switch (res[i].opcode) {
10702 case IORING_RESTRICTION_REGISTER_OP:
10703 if (res[i].register_op >= IORING_REGISTER_LAST) {
10708 __set_bit(res[i].register_op,
10709 ctx->restrictions.register_op);
10711 case IORING_RESTRICTION_SQE_OP:
10712 if (res[i].sqe_op >= IORING_OP_LAST) {
10717 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
10719 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
10720 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
10722 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
10723 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
10732 /* Reset all restrictions if an error happened */
10734 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
10736 ctx->restrictions.registered = true;
10742 static int io_register_enable_rings(struct io_ring_ctx *ctx)
10744 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10747 if (ctx->restrictions.registered)
10748 ctx->restricted = 1;
10750 ctx->flags &= ~IORING_SETUP_R_DISABLED;
10751 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
10752 wake_up(&ctx->sq_data->wait);
10756 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
10757 struct io_uring_rsrc_update2 *up,
10765 if (check_add_overflow(up->offset, nr_args, &tmp))
10767 err = io_rsrc_node_switch_start(ctx);
10772 case IORING_RSRC_FILE:
10773 return __io_sqe_files_update(ctx, up, nr_args);
10774 case IORING_RSRC_BUFFER:
10775 return __io_sqe_buffers_update(ctx, up, nr_args);
10780 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
10783 struct io_uring_rsrc_update2 up;
10787 memset(&up, 0, sizeof(up));
10788 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
10790 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
10793 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
10794 unsigned size, unsigned type)
10796 struct io_uring_rsrc_update2 up;
10798 if (size != sizeof(up))
10800 if (copy_from_user(&up, arg, sizeof(up)))
10802 if (!up.nr || up.resv)
10804 return __io_register_rsrc_update(ctx, type, &up, up.nr);
10807 static __cold int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
10808 unsigned int size, unsigned int type)
10810 struct io_uring_rsrc_register rr;
10812 /* keep it extendible */
10813 if (size != sizeof(rr))
10816 memset(&rr, 0, sizeof(rr));
10817 if (copy_from_user(&rr, arg, size))
10819 if (!rr.nr || rr.resv || rr.resv2)
10823 case IORING_RSRC_FILE:
10824 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10825 rr.nr, u64_to_user_ptr(rr.tags));
10826 case IORING_RSRC_BUFFER:
10827 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10828 rr.nr, u64_to_user_ptr(rr.tags));
10833 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
10834 void __user *arg, unsigned len)
10836 struct io_uring_task *tctx = current->io_uring;
10837 cpumask_var_t new_mask;
10840 if (!tctx || !tctx->io_wq)
10843 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10846 cpumask_clear(new_mask);
10847 if (len > cpumask_size())
10848 len = cpumask_size();
10850 if (copy_from_user(new_mask, arg, len)) {
10851 free_cpumask_var(new_mask);
10855 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10856 free_cpumask_var(new_mask);
10860 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10862 struct io_uring_task *tctx = current->io_uring;
10864 if (!tctx || !tctx->io_wq)
10867 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10870 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
10872 __must_hold(&ctx->uring_lock)
10874 struct io_tctx_node *node;
10875 struct io_uring_task *tctx = NULL;
10876 struct io_sq_data *sqd = NULL;
10877 __u32 new_count[2];
10880 if (copy_from_user(new_count, arg, sizeof(new_count)))
10882 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10883 if (new_count[i] > INT_MAX)
10886 if (ctx->flags & IORING_SETUP_SQPOLL) {
10887 sqd = ctx->sq_data;
10890 * Observe the correct sqd->lock -> ctx->uring_lock
10891 * ordering. Fine to drop uring_lock here, we hold
10892 * a ref to the ctx.
10894 refcount_inc(&sqd->refs);
10895 mutex_unlock(&ctx->uring_lock);
10896 mutex_lock(&sqd->lock);
10897 mutex_lock(&ctx->uring_lock);
10899 tctx = sqd->thread->io_uring;
10902 tctx = current->io_uring;
10905 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
10907 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10909 ctx->iowq_limits[i] = new_count[i];
10910 ctx->iowq_limits_set = true;
10912 if (tctx && tctx->io_wq) {
10913 ret = io_wq_max_workers(tctx->io_wq, new_count);
10917 memset(new_count, 0, sizeof(new_count));
10921 mutex_unlock(&sqd->lock);
10922 io_put_sq_data(sqd);
10925 if (copy_to_user(arg, new_count, sizeof(new_count)))
10928 /* that's it for SQPOLL, only the SQPOLL task creates requests */
10932 /* now propagate the restriction to all registered users */
10933 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
10934 struct io_uring_task *tctx = node->task->io_uring;
10936 if (WARN_ON_ONCE(!tctx->io_wq))
10939 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10940 new_count[i] = ctx->iowq_limits[i];
10941 /* ignore errors, it always returns zero anyway */
10942 (void)io_wq_max_workers(tctx->io_wq, new_count);
10947 mutex_unlock(&sqd->lock);
10948 io_put_sq_data(sqd);
10953 static bool io_register_op_must_quiesce(int op)
10956 case IORING_REGISTER_BUFFERS:
10957 case IORING_UNREGISTER_BUFFERS:
10958 case IORING_REGISTER_FILES:
10959 case IORING_UNREGISTER_FILES:
10960 case IORING_REGISTER_FILES_UPDATE:
10961 case IORING_REGISTER_PROBE:
10962 case IORING_REGISTER_PERSONALITY:
10963 case IORING_UNREGISTER_PERSONALITY:
10964 case IORING_REGISTER_FILES2:
10965 case IORING_REGISTER_FILES_UPDATE2:
10966 case IORING_REGISTER_BUFFERS2:
10967 case IORING_REGISTER_BUFFERS_UPDATE:
10968 case IORING_REGISTER_IOWQ_AFF:
10969 case IORING_UNREGISTER_IOWQ_AFF:
10970 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10977 static __cold int io_ctx_quiesce(struct io_ring_ctx *ctx)
10981 percpu_ref_kill(&ctx->refs);
10984 * Drop uring mutex before waiting for references to exit. If another
10985 * thread is currently inside io_uring_enter() it might need to grab the
10986 * uring_lock to make progress. If we hold it here across the drain
10987 * wait, then we can deadlock. It's safe to drop the mutex here, since
10988 * no new references will come in after we've killed the percpu ref.
10990 mutex_unlock(&ctx->uring_lock);
10992 ret = wait_for_completion_interruptible_timeout(&ctx->ref_comp, HZ);
10994 ret = min(0L, ret);
10998 ret = io_run_task_work_sig();
10999 io_req_caches_free(ctx);
11000 } while (ret >= 0);
11001 mutex_lock(&ctx->uring_lock);
11004 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
11008 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
11009 void __user *arg, unsigned nr_args)
11010 __releases(ctx->uring_lock)
11011 __acquires(ctx->uring_lock)
11016 * We're inside the ring mutex, if the ref is already dying, then
11017 * someone else killed the ctx or is already going through
11018 * io_uring_register().
11020 if (percpu_ref_is_dying(&ctx->refs))
11023 if (ctx->restricted) {
11024 if (opcode >= IORING_REGISTER_LAST)
11026 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
11027 if (!test_bit(opcode, ctx->restrictions.register_op))
11031 if (io_register_op_must_quiesce(opcode)) {
11032 ret = io_ctx_quiesce(ctx);
11038 case IORING_REGISTER_BUFFERS:
11039 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
11041 case IORING_UNREGISTER_BUFFERS:
11043 if (arg || nr_args)
11045 ret = io_sqe_buffers_unregister(ctx);
11047 case IORING_REGISTER_FILES:
11048 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
11050 case IORING_UNREGISTER_FILES:
11052 if (arg || nr_args)
11054 ret = io_sqe_files_unregister(ctx);
11056 case IORING_REGISTER_FILES_UPDATE:
11057 ret = io_register_files_update(ctx, arg, nr_args);
11059 case IORING_REGISTER_EVENTFD:
11060 case IORING_REGISTER_EVENTFD_ASYNC:
11064 ret = io_eventfd_register(ctx, arg);
11067 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
11068 ctx->eventfd_async = 1;
11070 ctx->eventfd_async = 0;
11072 case IORING_UNREGISTER_EVENTFD:
11074 if (arg || nr_args)
11076 ret = io_eventfd_unregister(ctx);
11078 case IORING_REGISTER_PROBE:
11080 if (!arg || nr_args > 256)
11082 ret = io_probe(ctx, arg, nr_args);
11084 case IORING_REGISTER_PERSONALITY:
11086 if (arg || nr_args)
11088 ret = io_register_personality(ctx);
11090 case IORING_UNREGISTER_PERSONALITY:
11094 ret = io_unregister_personality(ctx, nr_args);
11096 case IORING_REGISTER_ENABLE_RINGS:
11098 if (arg || nr_args)
11100 ret = io_register_enable_rings(ctx);
11102 case IORING_REGISTER_RESTRICTIONS:
11103 ret = io_register_restrictions(ctx, arg, nr_args);
11105 case IORING_REGISTER_FILES2:
11106 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
11108 case IORING_REGISTER_FILES_UPDATE2:
11109 ret = io_register_rsrc_update(ctx, arg, nr_args,
11112 case IORING_REGISTER_BUFFERS2:
11113 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
11115 case IORING_REGISTER_BUFFERS_UPDATE:
11116 ret = io_register_rsrc_update(ctx, arg, nr_args,
11117 IORING_RSRC_BUFFER);
11119 case IORING_REGISTER_IOWQ_AFF:
11121 if (!arg || !nr_args)
11123 ret = io_register_iowq_aff(ctx, arg, nr_args);
11125 case IORING_UNREGISTER_IOWQ_AFF:
11127 if (arg || nr_args)
11129 ret = io_unregister_iowq_aff(ctx);
11131 case IORING_REGISTER_IOWQ_MAX_WORKERS:
11133 if (!arg || nr_args != 2)
11135 ret = io_register_iowq_max_workers(ctx, arg);
11142 if (io_register_op_must_quiesce(opcode)) {
11143 /* bring the ctx back to life */
11144 percpu_ref_reinit(&ctx->refs);
11145 reinit_completion(&ctx->ref_comp);
11150 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
11151 void __user *, arg, unsigned int, nr_args)
11153 struct io_ring_ctx *ctx;
11162 if (f.file->f_op != &io_uring_fops)
11165 ctx = f.file->private_data;
11167 io_run_task_work();
11169 mutex_lock(&ctx->uring_lock);
11170 ret = __io_uring_register(ctx, opcode, arg, nr_args);
11171 mutex_unlock(&ctx->uring_lock);
11172 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
11173 ctx->cq_ev_fd != NULL, ret);
11179 static int __init io_uring_init(void)
11181 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
11182 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
11183 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
11186 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
11187 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
11188 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
11189 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
11190 BUILD_BUG_SQE_ELEM(1, __u8, flags);
11191 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
11192 BUILD_BUG_SQE_ELEM(4, __s32, fd);
11193 BUILD_BUG_SQE_ELEM(8, __u64, off);
11194 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
11195 BUILD_BUG_SQE_ELEM(16, __u64, addr);
11196 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
11197 BUILD_BUG_SQE_ELEM(24, __u32, len);
11198 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
11199 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
11200 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
11201 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
11202 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
11203 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
11204 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
11205 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
11206 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
11207 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
11208 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
11209 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
11210 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
11211 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
11212 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
11213 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
11214 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
11215 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
11216 BUILD_BUG_SQE_ELEM(42, __u16, personality);
11217 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
11218 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
11220 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
11221 sizeof(struct io_uring_rsrc_update));
11222 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
11223 sizeof(struct io_uring_rsrc_update2));
11225 /* ->buf_index is u16 */
11226 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
11228 /* should fit into one byte */
11229 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
11230 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
11231 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
11233 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
11234 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
11236 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
11240 __initcall(io_uring_init);