io_uring.c revision 90291099
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
5 *
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
8 *
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
16 * CQ entries.
17 *
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
23 * head will do).
24 *
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
28 * between.
29 *
30 * Also see the examples in the liburing library:
31 *
32 *	git://git.kernel.dk/liburing
33 *
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.
38 *
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
41 */
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>
51
52#include <linux/sched/signal.h>
53#include <linux/fs.h>
54#include <linux/file.h>
55#include <linux/fdtable.h>
56#include <linux/mm.h>
57#include <linux/mman.h>
58#include <linux/percpu.h>
59#include <linux/slab.h>
60#include <linux/blkdev.h>
61#include <linux/bvec.h>
62#include <linux/net.h>
63#include <net/sock.h>
64#include <net/af_unix.h>
65#include <net/scm.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
83#define CREATE_TRACE_POINTS
84#include <trace/events/io_uring.h>
85
86#include <uapi/linux/io_uring.h>
87
88#include "internal.h"
89#include "io-wq.h"
90
91#define IORING_MAX_ENTRIES	32768
92#define IORING_MAX_CQ_ENTRIES	(2 * IORING_MAX_ENTRIES)
93#define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
94
95/* 512 entries per page on 64-bit archs, 64 pages max */
96#define IORING_MAX_FIXED_FILES	(1U << 15)
97#define IORING_MAX_RESTRICTIONS	(IORING_RESTRICTION_LAST + \
98				 IORING_REGISTER_LAST + IORING_OP_LAST)
99
100#define IO_RSRC_TAG_TABLE_SHIFT	9
101#define IO_RSRC_TAG_TABLE_MAX	(1U << IO_RSRC_TAG_TABLE_SHIFT)
102#define IO_RSRC_TAG_TABLE_MASK	(IO_RSRC_TAG_TABLE_MAX - 1)
103
104#define IORING_MAX_REG_BUFFERS	(1U << 14)
105
106#define SQE_VALID_FLAGS	(IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK|	\
107				IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
108				IOSQE_BUFFER_SELECT)
109#define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
110				REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS)
111
112#define IO_TCTX_REFS_CACHE_NR	(1U << 10)
113
114struct io_uring {
115	u32 head ____cacheline_aligned_in_smp;
116	u32 tail ____cacheline_aligned_in_smp;
117};
118
119/*
120 * This data is shared with the application through the mmap at offsets
121 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
122 *
123 * The offsets to the member fields are published through struct
124 * io_sqring_offsets when calling io_uring_setup.
125 */
126struct io_rings {
127	/*
128	 * Head and tail offsets into the ring; the offsets need to be
129	 * masked to get valid indices.
130	 *
131	 * The kernel controls head of the sq ring and the tail of the cq ring,
132	 * and the application controls tail of the sq ring and the head of the
133	 * cq ring.
134	 */
135	struct io_uring		sq, cq;
136	/*
137	 * Bitmasks to apply to head and tail offsets (constant, equals
138	 * ring_entries - 1)
139	 */
140	u32			sq_ring_mask, cq_ring_mask;
141	/* Ring sizes (constant, power of 2) */
142	u32			sq_ring_entries, cq_ring_entries;
143	/*
144	 * Number of invalid entries dropped by the kernel due to
145	 * invalid index stored in array
146	 *
147	 * Written by the kernel, shouldn't be modified by the
148	 * application (i.e. get number of "new events" by comparing to
149	 * cached value).
150	 *
151	 * After a new SQ head value was read by the application this
152	 * counter includes all submissions that were dropped reaching
153	 * the new SQ head (and possibly more).
154	 */
155	u32			sq_dropped;
156	/*
157	 * Runtime SQ flags
158	 *
159	 * Written by the kernel, shouldn't be modified by the
160	 * application.
161	 *
162	 * The application needs a full memory barrier before checking
163	 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
164	 */
165	u32			sq_flags;
166	/*
167	 * Runtime CQ flags
168	 *
169	 * Written by the application, shouldn't be modified by the
170	 * kernel.
171	 */
172	u32			cq_flags;
173	/*
174	 * Number of completion events lost because the queue was full;
175	 * this should be avoided by the application by making sure
176	 * there are not more requests pending than there is space in
177	 * the completion queue.
178	 *
179	 * Written by the kernel, shouldn't be modified by the
180	 * application (i.e. get number of "new events" by comparing to
181	 * cached value).
182	 *
183	 * As completion events come in out of order this counter is not
184	 * ordered with any other data.
185	 */
186	u32			cq_overflow;
187	/*
188	 * Ring buffer of completion events.
189	 *
190	 * The kernel writes completion events fresh every time they are
191	 * produced, so the application is allowed to modify pending
192	 * entries.
193	 */
194	struct io_uring_cqe	cqes[] ____cacheline_aligned_in_smp;
195};
196
197enum io_uring_cmd_flags {
198	IO_URING_F_NONBLOCK		= 1,
199	IO_URING_F_COMPLETE_DEFER	= 2,
200};
201
202struct io_mapped_ubuf {
203	u64		ubuf;
204	u64		ubuf_end;
205	unsigned int	nr_bvecs;
206	unsigned long	acct_pages;
207	struct bio_vec	bvec[];
208};
209
210struct io_ring_ctx;
211
212struct io_overflow_cqe {
213	struct io_uring_cqe cqe;
214	struct list_head list;
215};
216
217struct io_fixed_file {
218	/* file * with additional FFS_* flags */
219	unsigned long file_ptr;
220};
221
222struct io_rsrc_put {
223	struct list_head list;
224	u64 tag;
225	union {
226		void *rsrc;
227		struct file *file;
228		struct io_mapped_ubuf *buf;
229	};
230};
231
232struct io_file_table {
233	struct io_fixed_file *files;
234};
235
236struct io_rsrc_node {
237	struct percpu_ref		refs;
238	struct list_head		node;
239	struct list_head		rsrc_list;
240	struct io_rsrc_data		*rsrc_data;
241	struct llist_node		llist;
242	bool				done;
243};
244
245typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
246
247struct io_rsrc_data {
248	struct io_ring_ctx		*ctx;
249
250	u64				**tags;
251	unsigned int			nr;
252	rsrc_put_fn			*do_put;
253	atomic_t			refs;
254	struct completion		done;
255	bool				quiesce;
256};
257
258struct io_buffer {
259	struct list_head list;
260	__u64 addr;
261	__u32 len;
262	__u16 bid;
263};
264
265struct io_restriction {
266	DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
267	DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
268	u8 sqe_flags_allowed;
269	u8 sqe_flags_required;
270	bool registered;
271};
272
273enum {
274	IO_SQ_THREAD_SHOULD_STOP = 0,
275	IO_SQ_THREAD_SHOULD_PARK,
276};
277
278struct io_sq_data {
279	refcount_t		refs;
280	atomic_t		park_pending;
281	struct mutex		lock;
282
283	/* ctx's that are using this sqd */
284	struct list_head	ctx_list;
285
286	struct task_struct	*thread;
287	struct wait_queue_head	wait;
288
289	unsigned		sq_thread_idle;
290	int			sq_cpu;
291	pid_t			task_pid;
292	pid_t			task_tgid;
293
294	unsigned long		state;
295	struct completion	exited;
296};
297
298#define IO_IOPOLL_BATCH			8
299#define IO_COMPL_BATCH			32
300#define IO_REQ_CACHE_SIZE		32
301#define IO_REQ_ALLOC_BATCH		8
302
303struct io_comp_state {
304	struct io_kiocb		*reqs[IO_COMPL_BATCH];
305	unsigned int		nr;
306	/* inline/task_work completion list, under ->uring_lock */
307	struct list_head	free_list;
308};
309
310struct io_submit_link {
311	struct io_kiocb		*head;
312	struct io_kiocb		*last;
313};
314
315struct io_submit_state {
316	struct blk_plug		plug;
317	struct io_submit_link	link;
318
319	/*
320	 * io_kiocb alloc cache
321	 */
322	void			*reqs[IO_REQ_CACHE_SIZE];
323	unsigned int		free_reqs;
324
325	bool			plug_started;
326
327	/*
328	 * Batch completion logic
329	 */
330	struct io_comp_state	comp;
331
332	/*
333	 * File reference cache
334	 */
335	struct file		*file;
336	unsigned int		fd;
337	unsigned int		file_refs;
338	unsigned int		ios_left;
339};
340
341struct io_ring_ctx {
342	/* const or read-mostly hot data */
343	struct {
344		struct percpu_ref	refs;
345
346		struct io_rings		*rings;
347		unsigned int		flags;
348		unsigned int		compat: 1;
349		unsigned int		drain_next: 1;
350		unsigned int		eventfd_async: 1;
351		unsigned int		restricted: 1;
352		unsigned int		off_timeout_used: 1;
353		unsigned int		drain_active: 1;
354	} ____cacheline_aligned_in_smp;
355
356	/* submission data */
357	struct {
358		struct mutex		uring_lock;
359
360		/*
361		 * Ring buffer of indices into array of io_uring_sqe, which is
362		 * mmapped by the application using the IORING_OFF_SQES offset.
363		 *
364		 * This indirection could e.g. be used to assign fixed
365		 * io_uring_sqe entries to operations and only submit them to
366		 * the queue when needed.
367		 *
368		 * The kernel modifies neither the indices array nor the entries
369		 * array.
370		 */
371		u32			*sq_array;
372		struct io_uring_sqe	*sq_sqes;
373		unsigned		cached_sq_head;
374		unsigned		sq_entries;
375		struct list_head	defer_list;
376
377		/*
378		 * Fixed resources fast path, should be accessed only under
379		 * uring_lock, and updated through io_uring_register(2)
380		 */
381		struct io_rsrc_node	*rsrc_node;
382		struct io_file_table	file_table;
383		unsigned		nr_user_files;
384		unsigned		nr_user_bufs;
385		struct io_mapped_ubuf	**user_bufs;
386
387		struct io_submit_state	submit_state;
388		struct list_head	timeout_list;
389		struct list_head	cq_overflow_list;
390		struct xarray		io_buffers;
391		struct xarray		personalities;
392		u32			pers_next;
393		unsigned		sq_thread_idle;
394	} ____cacheline_aligned_in_smp;
395
396	/* IRQ completion list, under ->completion_lock */
397	struct list_head	locked_free_list;
398	unsigned int		locked_free_nr;
399
400	const struct cred	*sq_creds;	/* cred used for __io_sq_thread() */
401	struct io_sq_data	*sq_data;	/* if using sq thread polling */
402
403	struct wait_queue_head	sqo_sq_wait;
404	struct list_head	sqd_list;
405
406	unsigned long		check_cq_overflow;
407
408	struct {
409		unsigned		cached_cq_tail;
410		unsigned		cq_entries;
411		struct eventfd_ctx	*cq_ev_fd;
412		struct wait_queue_head	poll_wait;
413		struct wait_queue_head	cq_wait;
414		unsigned		cq_extra;
415		atomic_t		cq_timeouts;
416		struct fasync_struct	*cq_fasync;
417		unsigned		cq_last_tm_flush;
418	} ____cacheline_aligned_in_smp;
419
420	struct {
421		spinlock_t		completion_lock;
422
423		/*
424		 * ->iopoll_list is protected by the ctx->uring_lock for
425		 * io_uring instances that don't use IORING_SETUP_SQPOLL.
426		 * For SQPOLL, only the single threaded io_sq_thread() will
427		 * manipulate the list, hence no extra locking is needed there.
428		 */
429		struct list_head	iopoll_list;
430		struct hlist_head	*cancel_hash;
431		unsigned		cancel_hash_bits;
432		bool			poll_multi_queue;
433	} ____cacheline_aligned_in_smp;
434
435	struct io_restriction		restrictions;
436
437	/* slow path rsrc auxilary data, used by update/register */
438	struct {
439		struct io_rsrc_node		*rsrc_backup_node;
440		struct io_mapped_ubuf		*dummy_ubuf;
441		struct io_rsrc_data		*file_data;
442		struct io_rsrc_data		*buf_data;
443
444		struct delayed_work		rsrc_put_work;
445		struct llist_head		rsrc_put_llist;
446		struct list_head		rsrc_ref_list;
447		spinlock_t			rsrc_ref_lock;
448	};
449
450	/* Keep this last, we don't need it for the fast path */
451	struct {
452		#if defined(CONFIG_UNIX)
453			struct socket		*ring_sock;
454		#endif
455		/* hashed buffered write serialization */
456		struct io_wq_hash		*hash_map;
457
458		/* Only used for accounting purposes */
459		struct user_struct		*user;
460		struct mm_struct		*mm_account;
461
462		/* ctx exit and cancelation */
463		struct llist_head		fallback_llist;
464		struct delayed_work		fallback_work;
465		struct work_struct		exit_work;
466		struct list_head		tctx_list;
467		struct completion		ref_comp;
468	};
469};
470
471struct io_uring_task {
472	/* submission side */
473	int			cached_refs;
474	struct xarray		xa;
475	struct wait_queue_head	wait;
476	const struct io_ring_ctx *last;
477	struct io_wq		*io_wq;
478	struct percpu_counter	inflight;
479	atomic_t		inflight_tracked;
480	atomic_t		in_idle;
481
482	spinlock_t		task_lock;
483	struct io_wq_work_list	task_list;
484	unsigned long		task_state;
485	struct callback_head	task_work;
486};
487
488/*
489 * First field must be the file pointer in all the
490 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
491 */
492struct io_poll_iocb {
493	struct file			*file;
494	struct wait_queue_head		*head;
495	__poll_t			events;
496	bool				done;
497	bool				canceled;
498	struct wait_queue_entry		wait;
499};
500
501struct io_poll_update {
502	struct file			*file;
503	u64				old_user_data;
504	u64				new_user_data;
505	__poll_t			events;
506	bool				update_events;
507	bool				update_user_data;
508};
509
510struct io_close {
511	struct file			*file;
512	int				fd;
513};
514
515struct io_timeout_data {
516	struct io_kiocb			*req;
517	struct hrtimer			timer;
518	struct timespec64		ts;
519	enum hrtimer_mode		mode;
520};
521
522struct io_accept {
523	struct file			*file;
524	struct sockaddr __user		*addr;
525	int __user			*addr_len;
526	int				flags;
527	unsigned long			nofile;
528};
529
530struct io_sync {
531	struct file			*file;
532	loff_t				len;
533	loff_t				off;
534	int				flags;
535	int				mode;
536};
537
538struct io_cancel {
539	struct file			*file;
540	u64				addr;
541};
542
543struct io_timeout {
544	struct file			*file;
545	u32				off;
546	u32				target_seq;
547	struct list_head		list;
548	/* head of the link, used by linked timeouts only */
549	struct io_kiocb			*head;
550};
551
552struct io_timeout_rem {
553	struct file			*file;
554	u64				addr;
555
556	/* timeout update */
557	struct timespec64		ts;
558	u32				flags;
559};
560
561struct io_rw {
562	/* NOTE: kiocb has the file as the first member, so don't do it here */
563	struct kiocb			kiocb;
564	u64				addr;
565	u64				len;
566};
567
568struct io_connect {
569	struct file			*file;
570	struct sockaddr __user		*addr;
571	int				addr_len;
572};
573
574struct io_sr_msg {
575	struct file			*file;
576	union {
577		struct compat_msghdr __user	*umsg_compat;
578		struct user_msghdr __user	*umsg;
579		void __user			*buf;
580	};
581	int				msg_flags;
582	int				bgid;
583	size_t				len;
584	struct io_buffer		*kbuf;
585};
586
587struct io_open {
588	struct file			*file;
589	int				dfd;
590	struct filename			*filename;
591	struct open_how			how;
592	unsigned long			nofile;
593};
594
595struct io_rsrc_update {
596	struct file			*file;
597	u64				arg;
598	u32				nr_args;
599	u32				offset;
600};
601
602struct io_fadvise {
603	struct file			*file;
604	u64				offset;
605	u32				len;
606	u32				advice;
607};
608
609struct io_madvise {
610	struct file			*file;
611	u64				addr;
612	u32				len;
613	u32				advice;
614};
615
616struct io_epoll {
617	struct file			*file;
618	int				epfd;
619	int				op;
620	int				fd;
621	struct epoll_event		event;
622};
623
624struct io_splice {
625	struct file			*file_out;
626	struct file			*file_in;
627	loff_t				off_out;
628	loff_t				off_in;
629	u64				len;
630	unsigned int			flags;
631};
632
633struct io_provide_buf {
634	struct file			*file;
635	__u64				addr;
636	__u32				len;
637	__u32				bgid;
638	__u16				nbufs;
639	__u16				bid;
640};
641
642struct io_statx {
643	struct file			*file;
644	int				dfd;
645	unsigned int			mask;
646	unsigned int			flags;
647	const char __user		*filename;
648	struct statx __user		*buffer;
649};
650
651struct io_shutdown {
652	struct file			*file;
653	int				how;
654};
655
656struct io_rename {
657	struct file			*file;
658	int				old_dfd;
659	int				new_dfd;
660	struct filename			*oldpath;
661	struct filename			*newpath;
662	int				flags;
663};
664
665struct io_unlink {
666	struct file			*file;
667	int				dfd;
668	int				flags;
669	struct filename			*filename;
670};
671
672struct io_completion {
673	struct file			*file;
674	struct list_head		list;
675	u32				cflags;
676};
677
678struct io_async_connect {
679	struct sockaddr_storage		address;
680};
681
682struct io_async_msghdr {
683	struct iovec			fast_iov[UIO_FASTIOV];
684	/* points to an allocated iov, if NULL we use fast_iov instead */
685	struct iovec			*free_iov;
686	struct sockaddr __user		*uaddr;
687	struct msghdr			msg;
688	struct sockaddr_storage		addr;
689};
690
691struct io_async_rw {
692	struct iovec			fast_iov[UIO_FASTIOV];
693	const struct iovec		*free_iovec;
694	struct iov_iter			iter;
695	size_t				bytes_done;
696	struct wait_page_queue		wpq;
697};
698
699enum {
700	REQ_F_FIXED_FILE_BIT	= IOSQE_FIXED_FILE_BIT,
701	REQ_F_IO_DRAIN_BIT	= IOSQE_IO_DRAIN_BIT,
702	REQ_F_LINK_BIT		= IOSQE_IO_LINK_BIT,
703	REQ_F_HARDLINK_BIT	= IOSQE_IO_HARDLINK_BIT,
704	REQ_F_FORCE_ASYNC_BIT	= IOSQE_ASYNC_BIT,
705	REQ_F_BUFFER_SELECT_BIT	= IOSQE_BUFFER_SELECT_BIT,
706
707	/* first byte is taken by user flags, shift it to not overlap */
708	REQ_F_FAIL_BIT		= 8,
709	REQ_F_INFLIGHT_BIT,
710	REQ_F_CUR_POS_BIT,
711	REQ_F_NOWAIT_BIT,
712	REQ_F_LINK_TIMEOUT_BIT,
713	REQ_F_NEED_CLEANUP_BIT,
714	REQ_F_POLLED_BIT,
715	REQ_F_BUFFER_SELECTED_BIT,
716	REQ_F_LTIMEOUT_ACTIVE_BIT,
717	REQ_F_COMPLETE_INLINE_BIT,
718	REQ_F_REISSUE_BIT,
719	REQ_F_DONT_REISSUE_BIT,
720	REQ_F_CREDS_BIT,
721	/* keep async read/write and isreg together and in order */
722	REQ_F_NOWAIT_READ_BIT,
723	REQ_F_NOWAIT_WRITE_BIT,
724	REQ_F_ISREG_BIT,
725
726	/* not a real bit, just to check we're not overflowing the space */
727	__REQ_F_LAST_BIT,
728};
729
730enum {
731	/* ctx owns file */
732	REQ_F_FIXED_FILE	= BIT(REQ_F_FIXED_FILE_BIT),
733	/* drain existing IO first */
734	REQ_F_IO_DRAIN		= BIT(REQ_F_IO_DRAIN_BIT),
735	/* linked sqes */
736	REQ_F_LINK		= BIT(REQ_F_LINK_BIT),
737	/* doesn't sever on completion < 0 */
738	REQ_F_HARDLINK		= BIT(REQ_F_HARDLINK_BIT),
739	/* IOSQE_ASYNC */
740	REQ_F_FORCE_ASYNC	= BIT(REQ_F_FORCE_ASYNC_BIT),
741	/* IOSQE_BUFFER_SELECT */
742	REQ_F_BUFFER_SELECT	= BIT(REQ_F_BUFFER_SELECT_BIT),
743
744	/* fail rest of links */
745	REQ_F_FAIL		= BIT(REQ_F_FAIL_BIT),
746	/* on inflight list, should be cancelled and waited on exit reliably */
747	REQ_F_INFLIGHT		= BIT(REQ_F_INFLIGHT_BIT),
748	/* read/write uses file position */
749	REQ_F_CUR_POS		= BIT(REQ_F_CUR_POS_BIT),
750	/* must not punt to workers */
751	REQ_F_NOWAIT		= BIT(REQ_F_NOWAIT_BIT),
752	/* has or had linked timeout */
753	REQ_F_LINK_TIMEOUT	= BIT(REQ_F_LINK_TIMEOUT_BIT),
754	/* needs cleanup */
755	REQ_F_NEED_CLEANUP	= BIT(REQ_F_NEED_CLEANUP_BIT),
756	/* already went through poll handler */
757	REQ_F_POLLED		= BIT(REQ_F_POLLED_BIT),
758	/* buffer already selected */
759	REQ_F_BUFFER_SELECTED	= BIT(REQ_F_BUFFER_SELECTED_BIT),
760	/* linked timeout is active, i.e. prepared by link's head */
761	REQ_F_LTIMEOUT_ACTIVE	= BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
762	/* completion is deferred through io_comp_state */
763	REQ_F_COMPLETE_INLINE	= BIT(REQ_F_COMPLETE_INLINE_BIT),
764	/* caller should reissue async */
765	REQ_F_REISSUE		= BIT(REQ_F_REISSUE_BIT),
766	/* don't attempt request reissue, see io_rw_reissue() */
767	REQ_F_DONT_REISSUE	= BIT(REQ_F_DONT_REISSUE_BIT),
768	/* supports async reads */
769	REQ_F_NOWAIT_READ	= BIT(REQ_F_NOWAIT_READ_BIT),
770	/* supports async writes */
771	REQ_F_NOWAIT_WRITE	= BIT(REQ_F_NOWAIT_WRITE_BIT),
772	/* regular file */
773	REQ_F_ISREG		= BIT(REQ_F_ISREG_BIT),
774	/* has creds assigned */
775	REQ_F_CREDS		= BIT(REQ_F_CREDS_BIT),
776};
777
778struct async_poll {
779	struct io_poll_iocb	poll;
780	struct io_poll_iocb	*double_poll;
781};
782
783typedef void (*io_req_tw_func_t)(struct io_kiocb *req);
784
785struct io_task_work {
786	union {
787		struct io_wq_work_node	node;
788		struct llist_node	fallback_node;
789	};
790	io_req_tw_func_t		func;
791};
792
793enum {
794	IORING_RSRC_FILE		= 0,
795	IORING_RSRC_BUFFER		= 1,
796};
797
798/*
799 * NOTE! Each of the iocb union members has the file pointer
800 * as the first entry in their struct definition. So you can
801 * access the file pointer through any of the sub-structs,
802 * or directly as just 'ki_filp' in this struct.
803 */
804struct io_kiocb {
805	union {
806		struct file		*file;
807		struct io_rw		rw;
808		struct io_poll_iocb	poll;
809		struct io_poll_update	poll_update;
810		struct io_accept	accept;
811		struct io_sync		sync;
812		struct io_cancel	cancel;
813		struct io_timeout	timeout;
814		struct io_timeout_rem	timeout_rem;
815		struct io_connect	connect;
816		struct io_sr_msg	sr_msg;
817		struct io_open		open;
818		struct io_close		close;
819		struct io_rsrc_update	rsrc_update;
820		struct io_fadvise	fadvise;
821		struct io_madvise	madvise;
822		struct io_epoll		epoll;
823		struct io_splice	splice;
824		struct io_provide_buf	pbuf;
825		struct io_statx		statx;
826		struct io_shutdown	shutdown;
827		struct io_rename	rename;
828		struct io_unlink	unlink;
829		/* use only after cleaning per-op data, see io_clean_op() */
830		struct io_completion	compl;
831	};
832
833	/* opcode allocated if it needs to store data for async defer */
834	void				*async_data;
835	u8				opcode;
836	/* polled IO has completed */
837	u8				iopoll_completed;
838
839	u16				buf_index;
840	u32				result;
841
842	struct io_ring_ctx		*ctx;
843	unsigned int			flags;
844	atomic_t			refs;
845	struct task_struct		*task;
846	u64				user_data;
847
848	struct io_kiocb			*link;
849	struct percpu_ref		*fixed_rsrc_refs;
850
851	/* used with ctx->iopoll_list with reads/writes */
852	struct list_head		inflight_entry;
853	struct io_task_work		io_task_work;
854	/* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
855	struct hlist_node		hash_node;
856	struct async_poll		*apoll;
857	struct io_wq_work		work;
858	const struct cred		*creds;
859
860	/* store used ubuf, so we can prevent reloading */
861	struct io_mapped_ubuf		*imu;
862};
863
864struct io_tctx_node {
865	struct list_head	ctx_node;
866	struct task_struct	*task;
867	struct io_ring_ctx	*ctx;
868};
869
870struct io_defer_entry {
871	struct list_head	list;
872	struct io_kiocb		*req;
873	u32			seq;
874};
875
876struct io_op_def {
877	/* needs req->file assigned */
878	unsigned		needs_file : 1;
879	/* hash wq insertion if file is a regular file */
880	unsigned		hash_reg_file : 1;
881	/* unbound wq insertion if file is a non-regular file */
882	unsigned		unbound_nonreg_file : 1;
883	/* opcode is not supported by this kernel */
884	unsigned		not_supported : 1;
885	/* set if opcode supports polled "wait" */
886	unsigned		pollin : 1;
887	unsigned		pollout : 1;
888	/* op supports buffer selection */
889	unsigned		buffer_select : 1;
890	/* do prep async if is going to be punted */
891	unsigned		needs_async_setup : 1;
892	/* should block plug */
893	unsigned		plug : 1;
894	/* size of async data needed, if any */
895	unsigned short		async_size;
896};
897
898static const struct io_op_def io_op_defs[] = {
899	[IORING_OP_NOP] = {},
900	[IORING_OP_READV] = {
901		.needs_file		= 1,
902		.unbound_nonreg_file	= 1,
903		.pollin			= 1,
904		.buffer_select		= 1,
905		.needs_async_setup	= 1,
906		.plug			= 1,
907		.async_size		= sizeof(struct io_async_rw),
908	},
909	[IORING_OP_WRITEV] = {
910		.needs_file		= 1,
911		.hash_reg_file		= 1,
912		.unbound_nonreg_file	= 1,
913		.pollout		= 1,
914		.needs_async_setup	= 1,
915		.plug			= 1,
916		.async_size		= sizeof(struct io_async_rw),
917	},
918	[IORING_OP_FSYNC] = {
919		.needs_file		= 1,
920	},
921	[IORING_OP_READ_FIXED] = {
922		.needs_file		= 1,
923		.unbound_nonreg_file	= 1,
924		.pollin			= 1,
925		.plug			= 1,
926		.async_size		= sizeof(struct io_async_rw),
927	},
928	[IORING_OP_WRITE_FIXED] = {
929		.needs_file		= 1,
930		.hash_reg_file		= 1,
931		.unbound_nonreg_file	= 1,
932		.pollout		= 1,
933		.plug			= 1,
934		.async_size		= sizeof(struct io_async_rw),
935	},
936	[IORING_OP_POLL_ADD] = {
937		.needs_file		= 1,
938		.unbound_nonreg_file	= 1,
939	},
940	[IORING_OP_POLL_REMOVE] = {},
941	[IORING_OP_SYNC_FILE_RANGE] = {
942		.needs_file		= 1,
943	},
944	[IORING_OP_SENDMSG] = {
945		.needs_file		= 1,
946		.unbound_nonreg_file	= 1,
947		.pollout		= 1,
948		.needs_async_setup	= 1,
949		.async_size		= sizeof(struct io_async_msghdr),
950	},
951	[IORING_OP_RECVMSG] = {
952		.needs_file		= 1,
953		.unbound_nonreg_file	= 1,
954		.pollin			= 1,
955		.buffer_select		= 1,
956		.needs_async_setup	= 1,
957		.async_size		= sizeof(struct io_async_msghdr),
958	},
959	[IORING_OP_TIMEOUT] = {
960		.async_size		= sizeof(struct io_timeout_data),
961	},
962	[IORING_OP_TIMEOUT_REMOVE] = {
963		/* used by timeout updates' prep() */
964	},
965	[IORING_OP_ACCEPT] = {
966		.needs_file		= 1,
967		.unbound_nonreg_file	= 1,
968		.pollin			= 1,
969	},
970	[IORING_OP_ASYNC_CANCEL] = {},
971	[IORING_OP_LINK_TIMEOUT] = {
972		.async_size		= sizeof(struct io_timeout_data),
973	},
974	[IORING_OP_CONNECT] = {
975		.needs_file		= 1,
976		.unbound_nonreg_file	= 1,
977		.pollout		= 1,
978		.needs_async_setup	= 1,
979		.async_size		= sizeof(struct io_async_connect),
980	},
981	[IORING_OP_FALLOCATE] = {
982		.needs_file		= 1,
983	},
984	[IORING_OP_OPENAT] = {},
985	[IORING_OP_CLOSE] = {},
986	[IORING_OP_FILES_UPDATE] = {},
987	[IORING_OP_STATX] = {},
988	[IORING_OP_READ] = {
989		.needs_file		= 1,
990		.unbound_nonreg_file	= 1,
991		.pollin			= 1,
992		.buffer_select		= 1,
993		.plug			= 1,
994		.async_size		= sizeof(struct io_async_rw),
995	},
996	[IORING_OP_WRITE] = {
997		.needs_file		= 1,
998		.unbound_nonreg_file	= 1,
999		.pollout		= 1,
1000		.plug			= 1,
1001		.async_size		= sizeof(struct io_async_rw),
1002	},
1003	[IORING_OP_FADVISE] = {
1004		.needs_file		= 1,
1005	},
1006	[IORING_OP_MADVISE] = {},
1007	[IORING_OP_SEND] = {
1008		.needs_file		= 1,
1009		.unbound_nonreg_file	= 1,
1010		.pollout		= 1,
1011	},
1012	[IORING_OP_RECV] = {
1013		.needs_file		= 1,
1014		.unbound_nonreg_file	= 1,
1015		.pollin			= 1,
1016		.buffer_select		= 1,
1017	},
1018	[IORING_OP_OPENAT2] = {
1019	},
1020	[IORING_OP_EPOLL_CTL] = {
1021		.unbound_nonreg_file	= 1,
1022	},
1023	[IORING_OP_SPLICE] = {
1024		.needs_file		= 1,
1025		.hash_reg_file		= 1,
1026		.unbound_nonreg_file	= 1,
1027	},
1028	[IORING_OP_PROVIDE_BUFFERS] = {},
1029	[IORING_OP_REMOVE_BUFFERS] = {},
1030	[IORING_OP_TEE] = {
1031		.needs_file		= 1,
1032		.hash_reg_file		= 1,
1033		.unbound_nonreg_file	= 1,
1034	},
1035	[IORING_OP_SHUTDOWN] = {
1036		.needs_file		= 1,
1037	},
1038	[IORING_OP_RENAMEAT] = {},
1039	[IORING_OP_UNLINKAT] = {},
1040};
1041
1042static bool io_disarm_next(struct io_kiocb *req);
1043static void io_uring_del_tctx_node(unsigned long index);
1044static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1045					 struct task_struct *task,
1046					 bool cancel_all);
1047static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1048static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx);
1049
1050static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1051				 long res, unsigned int cflags);
1052static void io_put_req(struct io_kiocb *req);
1053static void io_put_req_deferred(struct io_kiocb *req, int nr);
1054static void io_dismantle_req(struct io_kiocb *req);
1055static void io_put_task(struct task_struct *task, int nr);
1056static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1057static void io_queue_linked_timeout(struct io_kiocb *req);
1058static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1059				     struct io_uring_rsrc_update2 *up,
1060				     unsigned nr_args);
1061static void io_clean_op(struct io_kiocb *req);
1062static struct file *io_file_get(struct io_ring_ctx *ctx,
1063				struct io_submit_state *state,
1064				struct io_kiocb *req, int fd, bool fixed);
1065static void __io_queue_sqe(struct io_kiocb *req);
1066static void io_rsrc_put_work(struct work_struct *work);
1067
1068static void io_req_task_queue(struct io_kiocb *req);
1069static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1070static bool io_poll_remove_waitqs(struct io_kiocb *req);
1071static int io_req_prep_async(struct io_kiocb *req);
1072
1073static void io_fallback_req_func(struct work_struct *unused);
1074
1075static struct kmem_cache *req_cachep;
1076
1077static const struct file_operations io_uring_fops;
1078
1079struct sock *io_uring_get_socket(struct file *file)
1080{
1081#if defined(CONFIG_UNIX)
1082	if (file->f_op == &io_uring_fops) {
1083		struct io_ring_ctx *ctx = file->private_data;
1084
1085		return ctx->ring_sock->sk;
1086	}
1087#endif
1088	return NULL;
1089}
1090EXPORT_SYMBOL(io_uring_get_socket);
1091
1092#define io_for_each_link(pos, head) \
1093	for (pos = (head); pos; pos = pos->link)
1094
1095static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1096{
1097	struct io_ring_ctx *ctx = req->ctx;
1098
1099	if (!req->fixed_rsrc_refs) {
1100		req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1101		percpu_ref_get(req->fixed_rsrc_refs);
1102	}
1103}
1104
1105static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1106{
1107	bool got = percpu_ref_tryget(ref);
1108
1109	/* already at zero, wait for ->release() */
1110	if (!got)
1111		wait_for_completion(compl);
1112	percpu_ref_resurrect(ref);
1113	if (got)
1114		percpu_ref_put(ref);
1115}
1116
1117static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1118			  bool cancel_all)
1119{
1120	struct io_kiocb *req;
1121
1122	if (task && head->task != task)
1123		return false;
1124	if (cancel_all)
1125		return true;
1126
1127	io_for_each_link(req, head) {
1128		if (req->flags & REQ_F_INFLIGHT)
1129			return true;
1130	}
1131	return false;
1132}
1133
1134static inline void req_set_fail(struct io_kiocb *req)
1135{
1136	req->flags |= REQ_F_FAIL;
1137}
1138
1139static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1140{
1141	struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1142
1143	complete(&ctx->ref_comp);
1144}
1145
1146static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1147{
1148	return !req->timeout.off;
1149}
1150
1151static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1152{
1153	struct io_ring_ctx *ctx;
1154	int hash_bits;
1155
1156	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1157	if (!ctx)
1158		return NULL;
1159
1160	/*
1161	 * Use 5 bits less than the max cq entries, that should give us around
1162	 * 32 entries per hash list if totally full and uniformly spread.
1163	 */
1164	hash_bits = ilog2(p->cq_entries);
1165	hash_bits -= 5;
1166	if (hash_bits <= 0)
1167		hash_bits = 1;
1168	ctx->cancel_hash_bits = hash_bits;
1169	ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1170					GFP_KERNEL);
1171	if (!ctx->cancel_hash)
1172		goto err;
1173	__hash_init(ctx->cancel_hash, 1U << hash_bits);
1174
1175	ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1176	if (!ctx->dummy_ubuf)
1177		goto err;
1178	/* set invalid range, so io_import_fixed() fails meeting it */
1179	ctx->dummy_ubuf->ubuf = -1UL;
1180
1181	if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1182			    PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1183		goto err;
1184
1185	ctx->flags = p->flags;
1186	init_waitqueue_head(&ctx->sqo_sq_wait);
1187	INIT_LIST_HEAD(&ctx->sqd_list);
1188	init_waitqueue_head(&ctx->poll_wait);
1189	INIT_LIST_HEAD(&ctx->cq_overflow_list);
1190	init_completion(&ctx->ref_comp);
1191	xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1192	xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1193	mutex_init(&ctx->uring_lock);
1194	init_waitqueue_head(&ctx->cq_wait);
1195	spin_lock_init(&ctx->completion_lock);
1196	INIT_LIST_HEAD(&ctx->iopoll_list);
1197	INIT_LIST_HEAD(&ctx->defer_list);
1198	INIT_LIST_HEAD(&ctx->timeout_list);
1199	spin_lock_init(&ctx->rsrc_ref_lock);
1200	INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1201	INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1202	init_llist_head(&ctx->rsrc_put_llist);
1203	INIT_LIST_HEAD(&ctx->tctx_list);
1204	INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1205	INIT_LIST_HEAD(&ctx->locked_free_list);
1206	INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1207	return ctx;
1208err:
1209	kfree(ctx->dummy_ubuf);
1210	kfree(ctx->cancel_hash);
1211	kfree(ctx);
1212	return NULL;
1213}
1214
1215static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1216{
1217	struct io_rings *r = ctx->rings;
1218
1219	WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1220	ctx->cq_extra--;
1221}
1222
1223static bool req_need_defer(struct io_kiocb *req, u32 seq)
1224{
1225	if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1226		struct io_ring_ctx *ctx = req->ctx;
1227
1228		return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1229	}
1230
1231	return false;
1232}
1233
1234#define FFS_ASYNC_READ		0x1UL
1235#define FFS_ASYNC_WRITE		0x2UL
1236#ifdef CONFIG_64BIT
1237#define FFS_ISREG		0x4UL
1238#else
1239#define FFS_ISREG		0x0UL
1240#endif
1241#define FFS_MASK		~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1242
1243static inline bool io_req_ffs_set(struct io_kiocb *req)
1244{
1245	return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE);
1246}
1247
1248static void io_req_track_inflight(struct io_kiocb *req)
1249{
1250	if (!(req->flags & REQ_F_INFLIGHT)) {
1251		req->flags |= REQ_F_INFLIGHT;
1252		atomic_inc(&current->io_uring->inflight_tracked);
1253	}
1254}
1255
1256static void io_prep_async_work(struct io_kiocb *req)
1257{
1258	const struct io_op_def *def = &io_op_defs[req->opcode];
1259	struct io_ring_ctx *ctx = req->ctx;
1260
1261	if (!(req->flags & REQ_F_CREDS)) {
1262		req->flags |= REQ_F_CREDS;
1263		req->creds = get_current_cred();
1264	}
1265
1266	req->work.list.next = NULL;
1267	req->work.flags = 0;
1268	if (req->flags & REQ_F_FORCE_ASYNC)
1269		req->work.flags |= IO_WQ_WORK_CONCURRENT;
1270
1271	if (req->flags & REQ_F_ISREG) {
1272		if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1273			io_wq_hash_work(&req->work, file_inode(req->file));
1274	} else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1275		if (def->unbound_nonreg_file)
1276			req->work.flags |= IO_WQ_WORK_UNBOUND;
1277	}
1278
1279	switch (req->opcode) {
1280	case IORING_OP_SPLICE:
1281	case IORING_OP_TEE:
1282		if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1283			req->work.flags |= IO_WQ_WORK_UNBOUND;
1284		break;
1285	}
1286}
1287
1288static void io_prep_async_link(struct io_kiocb *req)
1289{
1290	struct io_kiocb *cur;
1291
1292	if (req->flags & REQ_F_LINK_TIMEOUT) {
1293		struct io_ring_ctx *ctx = req->ctx;
1294
1295		spin_lock_irq(&ctx->completion_lock);
1296		io_for_each_link(cur, req)
1297			io_prep_async_work(cur);
1298		spin_unlock_irq(&ctx->completion_lock);
1299	} else {
1300		io_for_each_link(cur, req)
1301			io_prep_async_work(cur);
1302	}
1303}
1304
1305static void io_queue_async_work(struct io_kiocb *req)
1306{
1307	struct io_ring_ctx *ctx = req->ctx;
1308	struct io_kiocb *link = io_prep_linked_timeout(req);
1309	struct io_uring_task *tctx = req->task->io_uring;
1310
1311	BUG_ON(!tctx);
1312	BUG_ON(!tctx->io_wq);
1313
1314	/* init ->work of the whole link before punting */
1315	io_prep_async_link(req);
1316
1317	/*
1318	 * Not expected to happen, but if we do have a bug where this _can_
1319	 * happen, catch it here and ensure the request is marked as
1320	 * canceled. That will make io-wq go through the usual work cancel
1321	 * procedure rather than attempt to run this request (or create a new
1322	 * worker for it).
1323	 */
1324	if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1325		req->work.flags |= IO_WQ_WORK_CANCEL;
1326
1327	trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1328					&req->work, req->flags);
1329	io_wq_enqueue(tctx->io_wq, &req->work);
1330	if (link)
1331		io_queue_linked_timeout(link);
1332}
1333
1334static void io_kill_timeout(struct io_kiocb *req, int status)
1335	__must_hold(&req->ctx->completion_lock)
1336{
1337	struct io_timeout_data *io = req->async_data;
1338
1339	if (hrtimer_try_to_cancel(&io->timer) != -1) {
1340		atomic_set(&req->ctx->cq_timeouts,
1341			atomic_read(&req->ctx->cq_timeouts) + 1);
1342		list_del_init(&req->timeout.list);
1343		io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1344		io_put_req_deferred(req, 1);
1345	}
1346}
1347
1348static void io_queue_deferred(struct io_ring_ctx *ctx)
1349{
1350	while (!list_empty(&ctx->defer_list)) {
1351		struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1352						struct io_defer_entry, list);
1353
1354		if (req_need_defer(de->req, de->seq))
1355			break;
1356		list_del_init(&de->list);
1357		io_req_task_queue(de->req);
1358		kfree(de);
1359	}
1360}
1361
1362static void io_flush_timeouts(struct io_ring_ctx *ctx)
1363{
1364	u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1365
1366	while (!list_empty(&ctx->timeout_list)) {
1367		u32 events_needed, events_got;
1368		struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1369						struct io_kiocb, timeout.list);
1370
1371		if (io_is_timeout_noseq(req))
1372			break;
1373
1374		/*
1375		 * Since seq can easily wrap around over time, subtract
1376		 * the last seq at which timeouts were flushed before comparing.
1377		 * Assuming not more than 2^31-1 events have happened since,
1378		 * these subtractions won't have wrapped, so we can check if
1379		 * target is in [last_seq, current_seq] by comparing the two.
1380		 */
1381		events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1382		events_got = seq - ctx->cq_last_tm_flush;
1383		if (events_got < events_needed)
1384			break;
1385
1386		list_del_init(&req->timeout.list);
1387		io_kill_timeout(req, 0);
1388	}
1389	ctx->cq_last_tm_flush = seq;
1390}
1391
1392static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1393{
1394	if (ctx->off_timeout_used)
1395		io_flush_timeouts(ctx);
1396	if (ctx->drain_active)
1397		io_queue_deferred(ctx);
1398}
1399
1400static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1401{
1402	if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1403		__io_commit_cqring_flush(ctx);
1404	/* order cqe stores with ring update */
1405	smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1406}
1407
1408static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1409{
1410	struct io_rings *r = ctx->rings;
1411
1412	return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1413}
1414
1415static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1416{
1417	return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1418}
1419
1420static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1421{
1422	struct io_rings *rings = ctx->rings;
1423	unsigned tail, mask = ctx->cq_entries - 1;
1424
1425	/*
1426	 * writes to the cq entry need to come after reading head; the
1427	 * control dependency is enough as we're using WRITE_ONCE to
1428	 * fill the cq entry
1429	 */
1430	if (__io_cqring_events(ctx) == ctx->cq_entries)
1431		return NULL;
1432
1433	tail = ctx->cached_cq_tail++;
1434	return &rings->cqes[tail & mask];
1435}
1436
1437static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1438{
1439	if (likely(!ctx->cq_ev_fd))
1440		return false;
1441	if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1442		return false;
1443	return !ctx->eventfd_async || io_wq_current_is_worker();
1444}
1445
1446static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1447{
1448	/*
1449	 * wake_up_all() may seem excessive, but io_wake_function() and
1450	 * io_should_wake() handle the termination of the loop and only
1451	 * wake as many waiters as we need to.
1452	 */
1453	if (wq_has_sleeper(&ctx->cq_wait))
1454		wake_up_all(&ctx->cq_wait);
1455	if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1456		wake_up(&ctx->sq_data->wait);
1457	if (io_should_trigger_evfd(ctx))
1458		eventfd_signal(ctx->cq_ev_fd, 1);
1459	if (waitqueue_active(&ctx->poll_wait)) {
1460		wake_up_interruptible(&ctx->poll_wait);
1461		kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1462	}
1463}
1464
1465static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1466{
1467	if (ctx->flags & IORING_SETUP_SQPOLL) {
1468		if (wq_has_sleeper(&ctx->cq_wait))
1469			wake_up_all(&ctx->cq_wait);
1470	}
1471	if (io_should_trigger_evfd(ctx))
1472		eventfd_signal(ctx->cq_ev_fd, 1);
1473	if (waitqueue_active(&ctx->poll_wait)) {
1474		wake_up_interruptible(&ctx->poll_wait);
1475		kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1476	}
1477}
1478
1479/* Returns true if there are no backlogged entries after the flush */
1480static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1481{
1482	unsigned long flags;
1483	bool all_flushed, posted;
1484
1485	if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1486		return false;
1487
1488	posted = false;
1489	spin_lock_irqsave(&ctx->completion_lock, flags);
1490	while (!list_empty(&ctx->cq_overflow_list)) {
1491		struct io_uring_cqe *cqe = io_get_cqe(ctx);
1492		struct io_overflow_cqe *ocqe;
1493
1494		if (!cqe && !force)
1495			break;
1496		ocqe = list_first_entry(&ctx->cq_overflow_list,
1497					struct io_overflow_cqe, list);
1498		if (cqe)
1499			memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1500		else
1501			io_account_cq_overflow(ctx);
1502
1503		posted = true;
1504		list_del(&ocqe->list);
1505		kfree(ocqe);
1506	}
1507
1508	all_flushed = list_empty(&ctx->cq_overflow_list);
1509	if (all_flushed) {
1510		clear_bit(0, &ctx->check_cq_overflow);
1511		WRITE_ONCE(ctx->rings->sq_flags,
1512			   ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1513	}
1514
1515	if (posted)
1516		io_commit_cqring(ctx);
1517	spin_unlock_irqrestore(&ctx->completion_lock, flags);
1518	if (posted)
1519		io_cqring_ev_posted(ctx);
1520	return all_flushed;
1521}
1522
1523static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1524{
1525	bool ret = true;
1526
1527	if (test_bit(0, &ctx->check_cq_overflow)) {
1528		/* iopoll syncs against uring_lock, not completion_lock */
1529		if (ctx->flags & IORING_SETUP_IOPOLL)
1530			mutex_lock(&ctx->uring_lock);
1531		ret = __io_cqring_overflow_flush(ctx, force);
1532		if (ctx->flags & IORING_SETUP_IOPOLL)
1533			mutex_unlock(&ctx->uring_lock);
1534	}
1535
1536	return ret;
1537}
1538
1539/*
1540 * Shamelessly stolen from the mm implementation of page reference checking,
1541 * see commit f958d7b528b1 for details.
1542 */
1543#define req_ref_zero_or_close_to_overflow(req)	\
1544	((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1545
1546static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1547{
1548	return atomic_inc_not_zero(&req->refs);
1549}
1550
1551static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1552{
1553	WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1554	return atomic_sub_and_test(refs, &req->refs);
1555}
1556
1557static inline bool req_ref_put_and_test(struct io_kiocb *req)
1558{
1559	WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1560	return atomic_dec_and_test(&req->refs);
1561}
1562
1563static inline void req_ref_put(struct io_kiocb *req)
1564{
1565	WARN_ON_ONCE(req_ref_put_and_test(req));
1566}
1567
1568static inline void req_ref_get(struct io_kiocb *req)
1569{
1570	WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1571	atomic_inc(&req->refs);
1572}
1573
1574static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1575				     long res, unsigned int cflags)
1576{
1577	struct io_overflow_cqe *ocqe;
1578
1579	ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1580	if (!ocqe) {
1581		/*
1582		 * If we're in ring overflow flush mode, or in task cancel mode,
1583		 * or cannot allocate an overflow entry, then we need to drop it
1584		 * on the floor.
1585		 */
1586		io_account_cq_overflow(ctx);
1587		return false;
1588	}
1589	if (list_empty(&ctx->cq_overflow_list)) {
1590		set_bit(0, &ctx->check_cq_overflow);
1591		WRITE_ONCE(ctx->rings->sq_flags,
1592			   ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1593
1594	}
1595	ocqe->cqe.user_data = user_data;
1596	ocqe->cqe.res = res;
1597	ocqe->cqe.flags = cflags;
1598	list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1599	return true;
1600}
1601
1602static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1603					  long res, unsigned int cflags)
1604{
1605	struct io_uring_cqe *cqe;
1606
1607	trace_io_uring_complete(ctx, user_data, res, cflags);
1608
1609	/*
1610	 * If we can't get a cq entry, userspace overflowed the
1611	 * submission (by quite a lot). Increment the overflow count in
1612	 * the ring.
1613	 */
1614	cqe = io_get_cqe(ctx);
1615	if (likely(cqe)) {
1616		WRITE_ONCE(cqe->user_data, user_data);
1617		WRITE_ONCE(cqe->res, res);
1618		WRITE_ONCE(cqe->flags, cflags);
1619		return true;
1620	}
1621	return io_cqring_event_overflow(ctx, user_data, res, cflags);
1622}
1623
1624/* not as hot to bloat with inlining */
1625static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1626					  long res, unsigned int cflags)
1627{
1628	return __io_cqring_fill_event(ctx, user_data, res, cflags);
1629}
1630
1631static void io_req_complete_post(struct io_kiocb *req, long res,
1632				 unsigned int cflags)
1633{
1634	struct io_ring_ctx *ctx = req->ctx;
1635	unsigned long flags;
1636
1637	spin_lock_irqsave(&ctx->completion_lock, flags);
1638	__io_cqring_fill_event(ctx, req->user_data, res, cflags);
1639	/*
1640	 * If we're the last reference to this request, add to our locked
1641	 * free_list cache.
1642	 */
1643	if (req_ref_put_and_test(req)) {
1644		if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1645			if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL))
1646				io_disarm_next(req);
1647			if (req->link) {
1648				io_req_task_queue(req->link);
1649				req->link = NULL;
1650			}
1651		}
1652		io_dismantle_req(req);
1653		io_put_task(req->task, 1);
1654		list_add(&req->compl.list, &ctx->locked_free_list);
1655		ctx->locked_free_nr++;
1656	} else {
1657		if (!percpu_ref_tryget(&ctx->refs))
1658			req = NULL;
1659	}
1660	io_commit_cqring(ctx);
1661	spin_unlock_irqrestore(&ctx->completion_lock, flags);
1662
1663	if (req) {
1664		io_cqring_ev_posted(ctx);
1665		percpu_ref_put(&ctx->refs);
1666	}
1667}
1668
1669static inline bool io_req_needs_clean(struct io_kiocb *req)
1670{
1671	return req->flags & IO_REQ_CLEAN_FLAGS;
1672}
1673
1674static void io_req_complete_state(struct io_kiocb *req, long res,
1675				  unsigned int cflags)
1676{
1677	if (io_req_needs_clean(req))
1678		io_clean_op(req);
1679	req->result = res;
1680	req->compl.cflags = cflags;
1681	req->flags |= REQ_F_COMPLETE_INLINE;
1682}
1683
1684static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1685				     long res, unsigned cflags)
1686{
1687	if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1688		io_req_complete_state(req, res, cflags);
1689	else
1690		io_req_complete_post(req, res, cflags);
1691}
1692
1693static inline void io_req_complete(struct io_kiocb *req, long res)
1694{
1695	__io_req_complete(req, 0, res, 0);
1696}
1697
1698static void io_req_complete_failed(struct io_kiocb *req, long res)
1699{
1700	req_set_fail(req);
1701	io_put_req(req);
1702	io_req_complete_post(req, res, 0);
1703}
1704
1705/*
1706 * Don't initialise the fields below on every allocation, but do that in
1707 * advance and keep them valid across allocations.
1708 */
1709static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1710{
1711	req->ctx = ctx;
1712	req->link = NULL;
1713	req->async_data = NULL;
1714	/* not necessary, but safer to zero */
1715	req->result = 0;
1716}
1717
1718static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1719					struct io_comp_state *cs)
1720{
1721	spin_lock_irq(&ctx->completion_lock);
1722	list_splice_init(&ctx->locked_free_list, &cs->free_list);
1723	ctx->locked_free_nr = 0;
1724	spin_unlock_irq(&ctx->completion_lock);
1725}
1726
1727/* Returns true IFF there are requests in the cache */
1728static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1729{
1730	struct io_submit_state *state = &ctx->submit_state;
1731	struct io_comp_state *cs = &state->comp;
1732	int nr;
1733
1734	/*
1735	 * If we have more than a batch's worth of requests in our IRQ side
1736	 * locked cache, grab the lock and move them over to our submission
1737	 * side cache.
1738	 */
1739	if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1740		io_flush_cached_locked_reqs(ctx, cs);
1741
1742	nr = state->free_reqs;
1743	while (!list_empty(&cs->free_list)) {
1744		struct io_kiocb *req = list_first_entry(&cs->free_list,
1745						struct io_kiocb, compl.list);
1746
1747		list_del(&req->compl.list);
1748		state->reqs[nr++] = req;
1749		if (nr == ARRAY_SIZE(state->reqs))
1750			break;
1751	}
1752
1753	state->free_reqs = nr;
1754	return nr != 0;
1755}
1756
1757static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1758{
1759	struct io_submit_state *state = &ctx->submit_state;
1760	gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1761	int ret, i;
1762
1763	BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1764
1765	if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1766		goto got_req;
1767
1768	ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1769				    state->reqs);
1770
1771	/*
1772	 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1773	 * retry single alloc to be on the safe side.
1774	 */
1775	if (unlikely(ret <= 0)) {
1776		state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1777		if (!state->reqs[0])
1778			return NULL;
1779		ret = 1;
1780	}
1781
1782	for (i = 0; i < ret; i++)
1783		io_preinit_req(state->reqs[i], ctx);
1784	state->free_reqs = ret;
1785got_req:
1786	state->free_reqs--;
1787	return state->reqs[state->free_reqs];
1788}
1789
1790static inline void io_put_file(struct file *file)
1791{
1792	if (file)
1793		fput(file);
1794}
1795
1796static void io_dismantle_req(struct io_kiocb *req)
1797{
1798	unsigned int flags = req->flags;
1799
1800	if (io_req_needs_clean(req))
1801		io_clean_op(req);
1802	if (!(flags & REQ_F_FIXED_FILE))
1803		io_put_file(req->file);
1804	if (req->fixed_rsrc_refs)
1805		percpu_ref_put(req->fixed_rsrc_refs);
1806	if (req->async_data) {
1807		kfree(req->async_data);
1808		req->async_data = NULL;
1809	}
1810}
1811
1812/* must to be called somewhat shortly after putting a request */
1813static inline void io_put_task(struct task_struct *task, int nr)
1814{
1815	struct io_uring_task *tctx = task->io_uring;
1816
1817	percpu_counter_sub(&tctx->inflight, nr);
1818	if (unlikely(atomic_read(&tctx->in_idle)))
1819		wake_up(&tctx->wait);
1820	put_task_struct_many(task, nr);
1821}
1822
1823static void __io_free_req(struct io_kiocb *req)
1824{
1825	struct io_ring_ctx *ctx = req->ctx;
1826
1827	io_dismantle_req(req);
1828	io_put_task(req->task, 1);
1829
1830	kmem_cache_free(req_cachep, req);
1831	percpu_ref_put(&ctx->refs);
1832}
1833
1834static inline void io_remove_next_linked(struct io_kiocb *req)
1835{
1836	struct io_kiocb *nxt = req->link;
1837
1838	req->link = nxt->link;
1839	nxt->link = NULL;
1840}
1841
1842static bool io_kill_linked_timeout(struct io_kiocb *req)
1843	__must_hold(&req->ctx->completion_lock)
1844{
1845	struct io_kiocb *link = req->link;
1846
1847	/*
1848	 * Can happen if a linked timeout fired and link had been like
1849	 * req -> link t-out -> link t-out [-> ...]
1850	 */
1851	if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1852		struct io_timeout_data *io = link->async_data;
1853
1854		io_remove_next_linked(req);
1855		link->timeout.head = NULL;
1856		if (hrtimer_try_to_cancel(&io->timer) != -1) {
1857			io_cqring_fill_event(link->ctx, link->user_data,
1858					     -ECANCELED, 0);
1859			io_put_req_deferred(link, 1);
1860			return true;
1861		}
1862	}
1863	return false;
1864}
1865
1866static void io_fail_links(struct io_kiocb *req)
1867	__must_hold(&req->ctx->completion_lock)
1868{
1869	struct io_kiocb *nxt, *link = req->link;
1870
1871	req->link = NULL;
1872	while (link) {
1873		nxt = link->link;
1874		link->link = NULL;
1875
1876		trace_io_uring_fail_link(req, link);
1877		io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1878		io_put_req_deferred(link, 2);
1879		link = nxt;
1880	}
1881}
1882
1883static bool io_disarm_next(struct io_kiocb *req)
1884	__must_hold(&req->ctx->completion_lock)
1885{
1886	bool posted = false;
1887
1888	if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1889		posted = io_kill_linked_timeout(req);
1890	if (unlikely((req->flags & REQ_F_FAIL) &&
1891		     !(req->flags & REQ_F_HARDLINK))) {
1892		posted |= (req->link != NULL);
1893		io_fail_links(req);
1894	}
1895	return posted;
1896}
1897
1898static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1899{
1900	struct io_kiocb *nxt;
1901
1902	/*
1903	 * If LINK is set, we have dependent requests in this chain. If we
1904	 * didn't fail this request, queue the first one up, moving any other
1905	 * dependencies to the next request. In case of failure, fail the rest
1906	 * of the chain.
1907	 */
1908	if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) {
1909		struct io_ring_ctx *ctx = req->ctx;
1910		unsigned long flags;
1911		bool posted;
1912
1913		spin_lock_irqsave(&ctx->completion_lock, flags);
1914		posted = io_disarm_next(req);
1915		if (posted)
1916			io_commit_cqring(req->ctx);
1917		spin_unlock_irqrestore(&ctx->completion_lock, flags);
1918		if (posted)
1919			io_cqring_ev_posted(ctx);
1920	}
1921	nxt = req->link;
1922	req->link = NULL;
1923	return nxt;
1924}
1925
1926static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1927{
1928	if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1929		return NULL;
1930	return __io_req_find_next(req);
1931}
1932
1933static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1934{
1935	if (!ctx)
1936		return;
1937	if (ctx->submit_state.comp.nr) {
1938		mutex_lock(&ctx->uring_lock);
1939		io_submit_flush_completions(ctx);
1940		mutex_unlock(&ctx->uring_lock);
1941	}
1942	percpu_ref_put(&ctx->refs);
1943}
1944
1945static void tctx_task_work(struct callback_head *cb)
1946{
1947	struct io_ring_ctx *ctx = NULL;
1948	struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1949						  task_work);
1950
1951	while (1) {
1952		struct io_wq_work_node *node;
1953
1954		spin_lock_irq(&tctx->task_lock);
1955		node = tctx->task_list.first;
1956		INIT_WQ_LIST(&tctx->task_list);
1957		spin_unlock_irq(&tctx->task_lock);
1958
1959		while (node) {
1960			struct io_wq_work_node *next = node->next;
1961			struct io_kiocb *req = container_of(node, struct io_kiocb,
1962							    io_task_work.node);
1963
1964			if (req->ctx != ctx) {
1965				ctx_flush_and_put(ctx);
1966				ctx = req->ctx;
1967				percpu_ref_get(&ctx->refs);
1968			}
1969			req->io_task_work.func(req);
1970			node = next;
1971		}
1972		if (wq_list_empty(&tctx->task_list)) {
1973			spin_lock_irq(&tctx->task_lock);
1974			clear_bit(0, &tctx->task_state);
1975			if (wq_list_empty(&tctx->task_list)) {
1976				spin_unlock_irq(&tctx->task_lock);
1977				break;
1978			}
1979			spin_unlock_irq(&tctx->task_lock);
1980			/* another tctx_task_work() is enqueued, yield */
1981			if (test_and_set_bit(0, &tctx->task_state))
1982				break;
1983		}
1984		cond_resched();
1985	}
1986
1987	ctx_flush_and_put(ctx);
1988}
1989
1990static void io_req_task_work_add(struct io_kiocb *req)
1991{
1992	struct task_struct *tsk = req->task;
1993	struct io_uring_task *tctx = tsk->io_uring;
1994	enum task_work_notify_mode notify;
1995	struct io_wq_work_node *node;
1996	unsigned long flags;
1997
1998	WARN_ON_ONCE(!tctx);
1999
2000	spin_lock_irqsave(&tctx->task_lock, flags);
2001	wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2002	spin_unlock_irqrestore(&tctx->task_lock, flags);
2003
2004	/* task_work already pending, we're done */
2005	if (test_bit(0, &tctx->task_state) ||
2006	    test_and_set_bit(0, &tctx->task_state))
2007		return;
2008
2009	/*
2010	 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2011	 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2012	 * processing task_work. There's no reliable way to tell if TWA_RESUME
2013	 * will do the job.
2014	 */
2015	notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2016	if (!task_work_add(tsk, &tctx->task_work, notify)) {
2017		wake_up_process(tsk);
2018		return;
2019	}
2020
2021	clear_bit(0, &tctx->task_state);
2022	spin_lock_irqsave(&tctx->task_lock, flags);
2023	node = tctx->task_list.first;
2024	INIT_WQ_LIST(&tctx->task_list);
2025	spin_unlock_irqrestore(&tctx->task_lock, flags);
2026
2027	while (node) {
2028		req = container_of(node, struct io_kiocb, io_task_work.node);
2029		node = node->next;
2030		if (llist_add(&req->io_task_work.fallback_node,
2031			      &req->ctx->fallback_llist))
2032			schedule_delayed_work(&req->ctx->fallback_work, 1);
2033	}
2034}
2035
2036static void io_req_task_cancel(struct io_kiocb *req)
2037{
2038	struct io_ring_ctx *ctx = req->ctx;
2039
2040	/* ctx is guaranteed to stay alive while we hold uring_lock */
2041	mutex_lock(&ctx->uring_lock);
2042	io_req_complete_failed(req, req->result);
2043	mutex_unlock(&ctx->uring_lock);
2044}
2045
2046static void io_req_task_submit(struct io_kiocb *req)
2047{
2048	struct io_ring_ctx *ctx = req->ctx;
2049
2050	/* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2051	mutex_lock(&ctx->uring_lock);
2052	if (!(req->task->flags & PF_EXITING) && !req->task->in_execve)
2053		__io_queue_sqe(req);
2054	else
2055		io_req_complete_failed(req, -EFAULT);
2056	mutex_unlock(&ctx->uring_lock);
2057}
2058
2059static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2060{
2061	req->result = ret;
2062	req->io_task_work.func = io_req_task_cancel;
2063	io_req_task_work_add(req);
2064}
2065
2066static void io_req_task_queue(struct io_kiocb *req)
2067{
2068	req->io_task_work.func = io_req_task_submit;
2069	io_req_task_work_add(req);
2070}
2071
2072static void io_req_task_queue_reissue(struct io_kiocb *req)
2073{
2074	req->io_task_work.func = io_queue_async_work;
2075	io_req_task_work_add(req);
2076}
2077
2078static inline void io_queue_next(struct io_kiocb *req)
2079{
2080	struct io_kiocb *nxt = io_req_find_next(req);
2081
2082	if (nxt)
2083		io_req_task_queue(nxt);
2084}
2085
2086static void io_free_req(struct io_kiocb *req)
2087{
2088	io_queue_next(req);
2089	__io_free_req(req);
2090}
2091
2092struct req_batch {
2093	struct task_struct	*task;
2094	int			task_refs;
2095	int			ctx_refs;
2096};
2097
2098static inline void io_init_req_batch(struct req_batch *rb)
2099{
2100	rb->task_refs = 0;
2101	rb->ctx_refs = 0;
2102	rb->task = NULL;
2103}
2104
2105static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2106				     struct req_batch *rb)
2107{
2108	if (rb->task)
2109		io_put_task(rb->task, rb->task_refs);
2110	if (rb->ctx_refs)
2111		percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2112}
2113
2114static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2115			      struct io_submit_state *state)
2116{
2117	io_queue_next(req);
2118	io_dismantle_req(req);
2119
2120	if (req->task != rb->task) {
2121		if (rb->task)
2122			io_put_task(rb->task, rb->task_refs);
2123		rb->task = req->task;
2124		rb->task_refs = 0;
2125	}
2126	rb->task_refs++;
2127	rb->ctx_refs++;
2128
2129	if (state->free_reqs != ARRAY_SIZE(state->reqs))
2130		state->reqs[state->free_reqs++] = req;
2131	else
2132		list_add(&req->compl.list, &state->comp.free_list);
2133}
2134
2135static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2136	__must_hold(&req->ctx->uring_lock)
2137{
2138	struct io_comp_state *cs = &ctx->submit_state.comp;
2139	int i, nr = cs->nr;
2140	struct req_batch rb;
2141
2142	spin_lock_irq(&ctx->completion_lock);
2143	for (i = 0; i < nr; i++) {
2144		struct io_kiocb *req = cs->reqs[i];
2145
2146		__io_cqring_fill_event(ctx, req->user_data, req->result,
2147					req->compl.cflags);
2148	}
2149	io_commit_cqring(ctx);
2150	spin_unlock_irq(&ctx->completion_lock);
2151	io_cqring_ev_posted(ctx);
2152
2153	io_init_req_batch(&rb);
2154	for (i = 0; i < nr; i++) {
2155		struct io_kiocb *req = cs->reqs[i];
2156
2157		/* submission and completion refs */
2158		if (req_ref_sub_and_test(req, 2))
2159			io_req_free_batch(&rb, req, &ctx->submit_state);
2160	}
2161
2162	io_req_free_batch_finish(ctx, &rb);
2163	cs->nr = 0;
2164}
2165
2166/*
2167 * Drop reference to request, return next in chain (if there is one) if this
2168 * was the last reference to this request.
2169 */
2170static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2171{
2172	struct io_kiocb *nxt = NULL;
2173
2174	if (req_ref_put_and_test(req)) {
2175		nxt = io_req_find_next(req);
2176		__io_free_req(req);
2177	}
2178	return nxt;
2179}
2180
2181static inline void io_put_req(struct io_kiocb *req)
2182{
2183	if (req_ref_put_and_test(req))
2184		io_free_req(req);
2185}
2186
2187static void io_free_req_deferred(struct io_kiocb *req)
2188{
2189	req->io_task_work.func = io_free_req;
2190	io_req_task_work_add(req);
2191}
2192
2193static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2194{
2195	if (req_ref_sub_and_test(req, refs))
2196		io_free_req_deferred(req);
2197}
2198
2199static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2200{
2201	/* See comment at the top of this file */
2202	smp_rmb();
2203	return __io_cqring_events(ctx);
2204}
2205
2206static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2207{
2208	struct io_rings *rings = ctx->rings;
2209
2210	/* make sure SQ entry isn't read before tail */
2211	return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2212}
2213
2214static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2215{
2216	unsigned int cflags;
2217
2218	cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2219	cflags |= IORING_CQE_F_BUFFER;
2220	req->flags &= ~REQ_F_BUFFER_SELECTED;
2221	kfree(kbuf);
2222	return cflags;
2223}
2224
2225static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2226{
2227	struct io_buffer *kbuf;
2228
2229	kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2230	return io_put_kbuf(req, kbuf);
2231}
2232
2233static inline bool io_run_task_work(void)
2234{
2235	if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2236		__set_current_state(TASK_RUNNING);
2237		tracehook_notify_signal();
2238		return true;
2239	}
2240
2241	return false;
2242}
2243
2244/*
2245 * Find and free completed poll iocbs
2246 */
2247static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2248			       struct list_head *done, bool resubmit)
2249{
2250	struct req_batch rb;
2251	struct io_kiocb *req;
2252
2253	/* order with ->result store in io_complete_rw_iopoll() */
2254	smp_rmb();
2255
2256	io_init_req_batch(&rb);
2257	while (!list_empty(done)) {
2258		int cflags = 0;
2259
2260		req = list_first_entry(done, struct io_kiocb, inflight_entry);
2261		list_del(&req->inflight_entry);
2262
2263		if (READ_ONCE(req->result) == -EAGAIN && resubmit &&
2264		    !(req->flags & REQ_F_DONT_REISSUE)) {
2265			req->iopoll_completed = 0;
2266			req_ref_get(req);
2267			io_req_task_queue_reissue(req);
2268			continue;
2269		}
2270
2271		if (req->flags & REQ_F_BUFFER_SELECTED)
2272			cflags = io_put_rw_kbuf(req);
2273
2274		__io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2275		(*nr_events)++;
2276
2277		if (req_ref_put_and_test(req))
2278			io_req_free_batch(&rb, req, &ctx->submit_state);
2279	}
2280
2281	io_commit_cqring(ctx);
2282	io_cqring_ev_posted_iopoll(ctx);
2283	io_req_free_batch_finish(ctx, &rb);
2284}
2285
2286static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2287			long min, bool resubmit)
2288{
2289	struct io_kiocb *req, *tmp;
2290	LIST_HEAD(done);
2291	bool spin;
2292
2293	/*
2294	 * Only spin for completions if we don't have multiple devices hanging
2295	 * off our complete list, and we're under the requested amount.
2296	 */
2297	spin = !ctx->poll_multi_queue && *nr_events < min;
2298
2299	list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2300		struct kiocb *kiocb = &req->rw.kiocb;
2301		int ret;
2302
2303		/*
2304		 * Move completed and retryable entries to our local lists.
2305		 * If we find a request that requires polling, break out
2306		 * and complete those lists first, if we have entries there.
2307		 */
2308		if (READ_ONCE(req->iopoll_completed)) {
2309			list_move_tail(&req->inflight_entry, &done);
2310			continue;
2311		}
2312		if (!list_empty(&done))
2313			break;
2314
2315		ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2316		if (unlikely(ret < 0))
2317			return ret;
2318		else if (ret)
2319			spin = false;
2320
2321		/* iopoll may have completed current req */
2322		if (READ_ONCE(req->iopoll_completed))
2323			list_move_tail(&req->inflight_entry, &done);
2324	}
2325
2326	if (!list_empty(&done))
2327		io_iopoll_complete(ctx, nr_events, &done, resubmit);
2328
2329	return 0;
2330}
2331
2332/*
2333 * We can't just wait for polled events to come to us, we have to actively
2334 * find and complete them.
2335 */
2336static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2337{
2338	if (!(ctx->flags & IORING_SETUP_IOPOLL))
2339		return;
2340
2341	mutex_lock(&ctx->uring_lock);
2342	while (!list_empty(&ctx->iopoll_list)) {
2343		unsigned int nr_events = 0;
2344
2345		io_do_iopoll(ctx, &nr_events, 0, false);
2346
2347		/* let it sleep and repeat later if can't complete a request */
2348		if (nr_events == 0)
2349			break;
2350		/*
2351		 * Ensure we allow local-to-the-cpu processing to take place,
2352		 * in this case we need to ensure that we reap all events.
2353		 * Also let task_work, etc. to progress by releasing the mutex
2354		 */
2355		if (need_resched()) {
2356			mutex_unlock(&ctx->uring_lock);
2357			cond_resched();
2358			mutex_lock(&ctx->uring_lock);
2359		}
2360	}
2361	mutex_unlock(&ctx->uring_lock);
2362}
2363
2364static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2365{
2366	unsigned int nr_events = 0;
2367	int ret = 0;
2368
2369	/*
2370	 * We disallow the app entering submit/complete with polling, but we
2371	 * still need to lock the ring to prevent racing with polled issue
2372	 * that got punted to a workqueue.
2373	 */
2374	mutex_lock(&ctx->uring_lock);
2375	/*
2376	 * Don't enter poll loop if we already have events pending.
2377	 * If we do, we can potentially be spinning for commands that
2378	 * already triggered a CQE (eg in error).
2379	 */
2380	if (test_bit(0, &ctx->check_cq_overflow))
2381		__io_cqring_overflow_flush(ctx, false);
2382	if (io_cqring_events(ctx))
2383		goto out;
2384	do {
2385		/*
2386		 * If a submit got punted to a workqueue, we can have the
2387		 * application entering polling for a command before it gets
2388		 * issued. That app will hold the uring_lock for the duration
2389		 * of the poll right here, so we need to take a breather every
2390		 * now and then to ensure that the issue has a chance to add
2391		 * the poll to the issued list. Otherwise we can spin here
2392		 * forever, while the workqueue is stuck trying to acquire the
2393		 * very same mutex.
2394		 */
2395		if (list_empty(&ctx->iopoll_list)) {
2396			u32 tail = ctx->cached_cq_tail;
2397
2398			mutex_unlock(&ctx->uring_lock);
2399			io_run_task_work();
2400			mutex_lock(&ctx->uring_lock);
2401
2402			/* some requests don't go through iopoll_list */
2403			if (tail != ctx->cached_cq_tail ||
2404			    list_empty(&ctx->iopoll_list))
2405				break;
2406		}
2407		ret = io_do_iopoll(ctx, &nr_events, min, true);
2408	} while (!ret && nr_events < min && !need_resched());
2409out:
2410	mutex_unlock(&ctx->uring_lock);
2411	return ret;
2412}
2413
2414static void kiocb_end_write(struct io_kiocb *req)
2415{
2416	/*
2417	 * Tell lockdep we inherited freeze protection from submission
2418	 * thread.
2419	 */
2420	if (req->flags & REQ_F_ISREG) {
2421		struct super_block *sb = file_inode(req->file)->i_sb;
2422
2423		__sb_writers_acquired(sb, SB_FREEZE_WRITE);
2424		sb_end_write(sb);
2425	}
2426}
2427
2428#ifdef CONFIG_BLOCK
2429static bool io_resubmit_prep(struct io_kiocb *req)
2430{
2431	struct io_async_rw *rw = req->async_data;
2432
2433	if (!rw)
2434		return !io_req_prep_async(req);
2435	/* may have left rw->iter inconsistent on -EIOCBQUEUED */
2436	iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2437	return true;
2438}
2439
2440static bool io_rw_should_reissue(struct io_kiocb *req)
2441{
2442	umode_t mode = file_inode(req->file)->i_mode;
2443	struct io_ring_ctx *ctx = req->ctx;
2444
2445	if (!S_ISBLK(mode) && !S_ISREG(mode))
2446		return false;
2447	if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2448	    !(ctx->flags & IORING_SETUP_IOPOLL)))
2449		return false;
2450	/*
2451	 * If ref is dying, we might be running poll reap from the exit work.
2452	 * Don't attempt to reissue from that path, just let it fail with
2453	 * -EAGAIN.
2454	 */
2455	if (percpu_ref_is_dying(&ctx->refs))
2456		return false;
2457	/*
2458	 * Play it safe and assume not safe to re-import and reissue if we're
2459	 * not in the original thread group (or in task context).
2460	 */
2461	if (!same_thread_group(req->task, current) || !in_task())
2462		return false;
2463	return true;
2464}
2465#else
2466static bool io_resubmit_prep(struct io_kiocb *req)
2467{
2468	return false;
2469}
2470static bool io_rw_should_reissue(struct io_kiocb *req)
2471{
2472	return false;
2473}
2474#endif
2475
2476static void io_fallback_req_func(struct work_struct *work)
2477{
2478	struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
2479						fallback_work.work);
2480	struct llist_node *node = llist_del_all(&ctx->fallback_llist);
2481	struct io_kiocb *req, *tmp;
2482
2483	percpu_ref_get(&ctx->refs);
2484	llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
2485		req->io_task_work.func(req);
2486	percpu_ref_put(&ctx->refs);
2487}
2488
2489static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2490			     unsigned int issue_flags)
2491{
2492	int cflags = 0;
2493
2494	if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2495		kiocb_end_write(req);
2496	if (res != req->result) {
2497		if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2498		    io_rw_should_reissue(req)) {
2499			req->flags |= REQ_F_REISSUE;
2500			return;
2501		}
2502		req_set_fail(req);
2503	}
2504	if (req->flags & REQ_F_BUFFER_SELECTED)
2505		cflags = io_put_rw_kbuf(req);
2506	__io_req_complete(req, issue_flags, res, cflags);
2507}
2508
2509static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2510{
2511	struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2512
2513	__io_complete_rw(req, res, res2, 0);
2514}
2515
2516static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2517{
2518	struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2519
2520	if (kiocb->ki_flags & IOCB_WRITE)
2521		kiocb_end_write(req);
2522	if (unlikely(res != req->result)) {
2523		if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2524		    io_resubmit_prep(req))) {
2525			req_set_fail(req);
2526			req->flags |= REQ_F_DONT_REISSUE;
2527		}
2528	}
2529
2530	WRITE_ONCE(req->result, res);
2531	/* order with io_iopoll_complete() checking ->result */
2532	smp_wmb();
2533	WRITE_ONCE(req->iopoll_completed, 1);
2534}
2535
2536/*
2537 * After the iocb has been issued, it's safe to be found on the poll list.
2538 * Adding the kiocb to the list AFTER submission ensures that we don't
2539 * find it from a io_do_iopoll() thread before the issuer is done
2540 * accessing the kiocb cookie.
2541 */
2542static void io_iopoll_req_issued(struct io_kiocb *req)
2543{
2544	struct io_ring_ctx *ctx = req->ctx;
2545	const bool in_async = io_wq_current_is_worker();
2546
2547	/* workqueue context doesn't hold uring_lock, grab it now */
2548	if (unlikely(in_async))
2549		mutex_lock(&ctx->uring_lock);
2550
2551	/*
2552	 * Track whether we have multiple files in our lists. This will impact
2553	 * how we do polling eventually, not spinning if we're on potentially
2554	 * different devices.
2555	 */
2556	if (list_empty(&ctx->iopoll_list)) {
2557		ctx->poll_multi_queue = false;
2558	} else if (!ctx->poll_multi_queue) {
2559		struct io_kiocb *list_req;
2560		unsigned int queue_num0, queue_num1;
2561
2562		list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2563						inflight_entry);
2564
2565		if (list_req->file != req->file) {
2566			ctx->poll_multi_queue = true;
2567		} else {
2568			queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2569			queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2570			if (queue_num0 != queue_num1)
2571				ctx->poll_multi_queue = true;
2572		}
2573	}
2574
2575	/*
2576	 * For fast devices, IO may have already completed. If it has, add
2577	 * it to the front so we find it first.
2578	 */
2579	if (READ_ONCE(req->iopoll_completed))
2580		list_add(&req->inflight_entry, &ctx->iopoll_list);
2581	else
2582		list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2583
2584	if (unlikely(in_async)) {
2585		/*
2586		 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2587		 * in sq thread task context or in io worker task context. If
2588		 * current task context is sq thread, we don't need to check
2589		 * whether should wake up sq thread.
2590		 */
2591		if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2592		    wq_has_sleeper(&ctx->sq_data->wait))
2593			wake_up(&ctx->sq_data->wait);
2594
2595		mutex_unlock(&ctx->uring_lock);
2596	}
2597}
2598
2599static inline void io_state_file_put(struct io_submit_state *state)
2600{
2601	if (state->file_refs) {
2602		fput_many(state->file, state->file_refs);
2603		state->file_refs = 0;
2604	}
2605}
2606
2607/*
2608 * Get as many references to a file as we have IOs left in this submission,
2609 * assuming most submissions are for one file, or at least that each file
2610 * has more than one submission.
2611 */
2612static struct file *__io_file_get(struct io_submit_state *state, int fd)
2613{
2614	if (!state)
2615		return fget(fd);
2616
2617	if (state->file_refs) {
2618		if (state->fd == fd) {
2619			state->file_refs--;
2620			return state->file;
2621		}
2622		io_state_file_put(state);
2623	}
2624	state->file = fget_many(fd, state->ios_left);
2625	if (unlikely(!state->file))
2626		return NULL;
2627
2628	state->fd = fd;
2629	state->file_refs = state->ios_left - 1;
2630	return state->file;
2631}
2632
2633static bool io_bdev_nowait(struct block_device *bdev)
2634{
2635	return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2636}
2637
2638/*
2639 * If we tracked the file through the SCM inflight mechanism, we could support
2640 * any file. For now, just ensure that anything potentially problematic is done
2641 * inline.
2642 */
2643static bool __io_file_supports_nowait(struct file *file, int rw)
2644{
2645	umode_t mode = file_inode(file)->i_mode;
2646
2647	if (S_ISBLK(mode)) {
2648		if (IS_ENABLED(CONFIG_BLOCK) &&
2649		    io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2650			return true;
2651		return false;
2652	}
2653	if (S_ISSOCK(mode))
2654		return true;
2655	if (S_ISREG(mode)) {
2656		if (IS_ENABLED(CONFIG_BLOCK) &&
2657		    io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2658		    file->f_op != &io_uring_fops)
2659			return true;
2660		return false;
2661	}
2662
2663	/* any ->read/write should understand O_NONBLOCK */
2664	if (file->f_flags & O_NONBLOCK)
2665		return true;
2666
2667	if (!(file->f_mode & FMODE_NOWAIT))
2668		return false;
2669
2670	if (rw == READ)
2671		return file->f_op->read_iter != NULL;
2672
2673	return file->f_op->write_iter != NULL;
2674}
2675
2676static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2677{
2678	if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2679		return true;
2680	else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2681		return true;
2682
2683	return __io_file_supports_nowait(req->file, rw);
2684}
2685
2686static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2687{
2688	struct io_ring_ctx *ctx = req->ctx;
2689	struct kiocb *kiocb = &req->rw.kiocb;
2690	struct file *file = req->file;
2691	unsigned ioprio;
2692	int ret;
2693
2694	if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2695		req->flags |= REQ_F_ISREG;
2696
2697	kiocb->ki_pos = READ_ONCE(sqe->off);
2698	if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2699		req->flags |= REQ_F_CUR_POS;
2700		kiocb->ki_pos = file->f_pos;
2701	}
2702	kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2703	kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2704	ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2705	if (unlikely(ret))
2706		return ret;
2707
2708	/* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2709	if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2710		req->flags |= REQ_F_NOWAIT;
2711
2712	ioprio = READ_ONCE(sqe->ioprio);
2713	if (ioprio) {
2714		ret = ioprio_check_cap(ioprio);
2715		if (ret)
2716			return ret;
2717
2718		kiocb->ki_ioprio = ioprio;
2719	} else
2720		kiocb->ki_ioprio = get_current_ioprio();
2721
2722	if (ctx->flags & IORING_SETUP_IOPOLL) {
2723		if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2724		    !kiocb->ki_filp->f_op->iopoll)
2725			return -EOPNOTSUPP;
2726
2727		kiocb->ki_flags |= IOCB_HIPRI;
2728		kiocb->ki_complete = io_complete_rw_iopoll;
2729		req->iopoll_completed = 0;
2730	} else {
2731		if (kiocb->ki_flags & IOCB_HIPRI)
2732			return -EINVAL;
2733		kiocb->ki_complete = io_complete_rw;
2734	}
2735
2736	if (req->opcode == IORING_OP_READ_FIXED ||
2737	    req->opcode == IORING_OP_WRITE_FIXED) {
2738		req->imu = NULL;
2739		io_req_set_rsrc_node(req);
2740	}
2741
2742	req->rw.addr = READ_ONCE(sqe->addr);
2743	req->rw.len = READ_ONCE(sqe->len);
2744	req->buf_index = READ_ONCE(sqe->buf_index);
2745	return 0;
2746}
2747
2748static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2749{
2750	switch (ret) {
2751	case -EIOCBQUEUED:
2752		break;
2753	case -ERESTARTSYS:
2754	case -ERESTARTNOINTR:
2755	case -ERESTARTNOHAND:
2756	case -ERESTART_RESTARTBLOCK:
2757		/*
2758		 * We can't just restart the syscall, since previously
2759		 * submitted sqes may already be in progress. Just fail this
2760		 * IO with EINTR.
2761		 */
2762		ret = -EINTR;
2763		fallthrough;
2764	default:
2765		kiocb->ki_complete(kiocb, ret, 0);
2766	}
2767}
2768
2769static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2770		       unsigned int issue_flags)
2771{
2772	struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2773	struct io_async_rw *io = req->async_data;
2774	bool check_reissue = kiocb->ki_complete == io_complete_rw;
2775
2776	/* add previously done IO, if any */
2777	if (io && io->bytes_done > 0) {
2778		if (ret < 0)
2779			ret = io->bytes_done;
2780		else
2781			ret += io->bytes_done;
2782	}
2783
2784	if (req->flags & REQ_F_CUR_POS)
2785		req->file->f_pos = kiocb->ki_pos;
2786	if (ret >= 0 && check_reissue)
2787		__io_complete_rw(req, ret, 0, issue_flags);
2788	else
2789		io_rw_done(kiocb, ret);
2790
2791	if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2792		req->flags &= ~REQ_F_REISSUE;
2793		if (io_resubmit_prep(req)) {
2794			req_ref_get(req);
2795			io_req_task_queue_reissue(req);
2796		} else {
2797			int cflags = 0;
2798
2799			req_set_fail(req);
2800			if (req->flags & REQ_F_BUFFER_SELECTED)
2801				cflags = io_put_rw_kbuf(req);
2802			__io_req_complete(req, issue_flags, ret, cflags);
2803		}
2804	}
2805}
2806
2807static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2808			     struct io_mapped_ubuf *imu)
2809{
2810	size_t len = req->rw.len;
2811	u64 buf_end, buf_addr = req->rw.addr;
2812	size_t offset;
2813
2814	if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2815		return -EFAULT;
2816	/* not inside the mapped region */
2817	if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2818		return -EFAULT;
2819
2820	/*
2821	 * May not be a start of buffer, set size appropriately
2822	 * and advance us to the beginning.
2823	 */
2824	offset = buf_addr - imu->ubuf;
2825	iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2826
2827	if (offset) {
2828		/*
2829		 * Don't use iov_iter_advance() here, as it's really slow for
2830		 * using the latter parts of a big fixed buffer - it iterates
2831		 * over each segment manually. We can cheat a bit here, because
2832		 * we know that:
2833		 *
2834		 * 1) it's a BVEC iter, we set it up
2835		 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2836		 *    first and last bvec
2837		 *
2838		 * So just find our index, and adjust the iterator afterwards.
2839		 * If the offset is within the first bvec (or the whole first
2840		 * bvec, just use iov_iter_advance(). This makes it easier
2841		 * since we can just skip the first segment, which may not
2842		 * be PAGE_SIZE aligned.
2843		 */
2844		const struct bio_vec *bvec = imu->bvec;
2845
2846		if (offset <= bvec->bv_len) {
2847			iov_iter_advance(iter, offset);
2848		} else {
2849			unsigned long seg_skip;
2850
2851			/* skip first vec */
2852			offset -= bvec->bv_len;
2853			seg_skip = 1 + (offset >> PAGE_SHIFT);
2854
2855			iter->bvec = bvec + seg_skip;
2856			iter->nr_segs -= seg_skip;
2857			iter->count -= bvec->bv_len + offset;
2858			iter->iov_offset = offset & ~PAGE_MASK;
2859		}
2860	}
2861
2862	return 0;
2863}
2864
2865static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2866{
2867	struct io_ring_ctx *ctx = req->ctx;
2868	struct io_mapped_ubuf *imu = req->imu;
2869	u16 index, buf_index = req->buf_index;
2870
2871	if (likely(!imu)) {
2872		if (unlikely(buf_index >= ctx->nr_user_bufs))
2873			return -EFAULT;
2874		index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2875		imu = READ_ONCE(ctx->user_bufs[index]);
2876		req->imu = imu;
2877	}
2878	return __io_import_fixed(req, rw, iter, imu);
2879}
2880
2881static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2882{
2883	if (needs_lock)
2884		mutex_unlock(&ctx->uring_lock);
2885}
2886
2887static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2888{
2889	/*
2890	 * "Normal" inline submissions always hold the uring_lock, since we
2891	 * grab it from the system call. Same is true for the SQPOLL offload.
2892	 * The only exception is when we've detached the request and issue it
2893	 * from an async worker thread, grab the lock for that case.
2894	 */
2895	if (needs_lock)
2896		mutex_lock(&ctx->uring_lock);
2897}
2898
2899static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2900					  int bgid, struct io_buffer *kbuf,
2901					  bool needs_lock)
2902{
2903	struct io_buffer *head;
2904
2905	if (req->flags & REQ_F_BUFFER_SELECTED)
2906		return kbuf;
2907
2908	io_ring_submit_lock(req->ctx, needs_lock);
2909
2910	lockdep_assert_held(&req->ctx->uring_lock);
2911
2912	head = xa_load(&req->ctx->io_buffers, bgid);
2913	if (head) {
2914		if (!list_empty(&head->list)) {
2915			kbuf = list_last_entry(&head->list, struct io_buffer,
2916							list);
2917			list_del(&kbuf->list);
2918		} else {
2919			kbuf = head;
2920			xa_erase(&req->ctx->io_buffers, bgid);
2921		}
2922		if (*len > kbuf->len)
2923			*len = kbuf->len;
2924	} else {
2925		kbuf = ERR_PTR(-ENOBUFS);
2926	}
2927
2928	io_ring_submit_unlock(req->ctx, needs_lock);
2929
2930	return kbuf;
2931}
2932
2933static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2934					bool needs_lock)
2935{
2936	struct io_buffer *kbuf;
2937	u16 bgid;
2938
2939	kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2940	bgid = req->buf_index;
2941	kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2942	if (IS_ERR(kbuf))
2943		return kbuf;
2944	req->rw.addr = (u64) (unsigned long) kbuf;
2945	req->flags |= REQ_F_BUFFER_SELECTED;
2946	return u64_to_user_ptr(kbuf->addr);
2947}
2948
2949#ifdef CONFIG_COMPAT
2950static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2951				bool needs_lock)
2952{
2953	struct compat_iovec __user *uiov;
2954	compat_ssize_t clen;
2955	void __user *buf;
2956	ssize_t len;
2957
2958	uiov = u64_to_user_ptr(req->rw.addr);
2959	if (!access_ok(uiov, sizeof(*uiov)))
2960		return -EFAULT;
2961	if (__get_user(clen, &uiov->iov_len))
2962		return -EFAULT;
2963	if (clen < 0)
2964		return -EINVAL;
2965
2966	len = clen;
2967	buf = io_rw_buffer_select(req, &len, needs_lock);
2968	if (IS_ERR(buf))
2969		return PTR_ERR(buf);
2970	iov[0].iov_base = buf;
2971	iov[0].iov_len = (compat_size_t) len;
2972	return 0;
2973}
2974#endif
2975
2976static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2977				      bool needs_lock)
2978{
2979	struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2980	void __user *buf;
2981	ssize_t len;
2982
2983	if (copy_from_user(iov, uiov, sizeof(*uiov)))
2984		return -EFAULT;
2985
2986	len = iov[0].iov_len;
2987	if (len < 0)
2988		return -EINVAL;
2989	buf = io_rw_buffer_select(req, &len, needs_lock);
2990	if (IS_ERR(buf))
2991		return PTR_ERR(buf);
2992	iov[0].iov_base = buf;
2993	iov[0].iov_len = len;
2994	return 0;
2995}
2996
2997static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2998				    bool needs_lock)
2999{
3000	if (req->flags & REQ_F_BUFFER_SELECTED) {
3001		struct io_buffer *kbuf;
3002
3003		kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3004		iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3005		iov[0].iov_len = kbuf->len;
3006		return 0;
3007	}
3008	if (req->rw.len != 1)
3009		return -EINVAL;
3010
3011#ifdef CONFIG_COMPAT
3012	if (req->ctx->compat)
3013		return io_compat_import(req, iov, needs_lock);
3014#endif
3015
3016	return __io_iov_buffer_select(req, iov, needs_lock);
3017}
3018
3019static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3020			   struct iov_iter *iter, bool needs_lock)
3021{
3022	void __user *buf = u64_to_user_ptr(req->rw.addr);
3023	size_t sqe_len = req->rw.len;
3024	u8 opcode = req->opcode;
3025	ssize_t ret;
3026
3027	if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3028		*iovec = NULL;
3029		return io_import_fixed(req, rw, iter);
3030	}
3031
3032	/* buffer index only valid with fixed read/write, or buffer select  */
3033	if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3034		return -EINVAL;
3035
3036	if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3037		if (req->flags & REQ_F_BUFFER_SELECT) {
3038			buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3039			if (IS_ERR(buf))
3040				return PTR_ERR(buf);
3041			req->rw.len = sqe_len;
3042		}
3043
3044		ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3045		*iovec = NULL;
3046		return ret;
3047	}
3048
3049	if (req->flags & REQ_F_BUFFER_SELECT) {
3050		ret = io_iov_buffer_select(req, *iovec, needs_lock);
3051		if (!ret)
3052			iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3053		*iovec = NULL;
3054		return ret;
3055	}
3056
3057	return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3058			      req->ctx->compat);
3059}
3060
3061static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3062{
3063	return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3064}
3065
3066/*
3067 * For files that don't have ->read_iter() and ->write_iter(), handle them
3068 * by looping over ->read() or ->write() manually.
3069 */
3070static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3071{
3072	struct kiocb *kiocb = &req->rw.kiocb;
3073	struct file *file = req->file;
3074	ssize_t ret = 0;
3075
3076	/*
3077	 * Don't support polled IO through this interface, and we can't
3078	 * support non-blocking either. For the latter, this just causes
3079	 * the kiocb to be handled from an async context.
3080	 */
3081	if (kiocb->ki_flags & IOCB_HIPRI)
3082		return -EOPNOTSUPP;
3083	if (kiocb->ki_flags & IOCB_NOWAIT)
3084		return -EAGAIN;
3085
3086	while (iov_iter_count(iter)) {
3087		struct iovec iovec;
3088		ssize_t nr;
3089
3090		if (!iov_iter_is_bvec(iter)) {
3091			iovec = iov_iter_iovec(iter);
3092		} else {
3093			iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3094			iovec.iov_len = req->rw.len;
3095		}
3096
3097		if (rw == READ) {
3098			nr = file->f_op->read(file, iovec.iov_base,
3099					      iovec.iov_len, io_kiocb_ppos(kiocb));
3100		} else {
3101			nr = file->f_op->write(file, iovec.iov_base,
3102					       iovec.iov_len, io_kiocb_ppos(kiocb));
3103		}
3104
3105		if (nr < 0) {
3106			if (!ret)
3107				ret = nr;
3108			break;
3109		}
3110		ret += nr;
3111		if (nr != iovec.iov_len)
3112			break;
3113		req->rw.len -= nr;
3114		req->rw.addr += nr;
3115		iov_iter_advance(iter, nr);
3116	}
3117
3118	return ret;
3119}
3120
3121static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3122			  const struct iovec *fast_iov, struct iov_iter *iter)
3123{
3124	struct io_async_rw *rw = req->async_data;
3125
3126	memcpy(&rw->iter, iter, sizeof(*iter));
3127	rw->free_iovec = iovec;
3128	rw->bytes_done = 0;
3129	/* can only be fixed buffers, no need to do anything */
3130	if (iov_iter_is_bvec(iter))
3131		return;
3132	if (!iovec) {
3133		unsigned iov_off = 0;
3134
3135		rw->iter.iov = rw->fast_iov;
3136		if (iter->iov != fast_iov) {
3137			iov_off = iter->iov - fast_iov;
3138			rw->iter.iov += iov_off;
3139		}
3140		if (rw->fast_iov != fast_iov)
3141			memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3142			       sizeof(struct iovec) * iter->nr_segs);
3143	} else {
3144		req->flags |= REQ_F_NEED_CLEANUP;
3145	}
3146}
3147
3148static inline int io_alloc_async_data(struct io_kiocb *req)
3149{
3150	WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3151	req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3152	return req->async_data == NULL;
3153}
3154
3155static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3156			     const struct iovec *fast_iov,
3157			     struct iov_iter *iter, bool force)
3158{
3159	if (!force && !io_op_defs[req->opcode].needs_async_setup)
3160		return 0;
3161	if (!req->async_data) {
3162		if (io_alloc_async_data(req)) {
3163			kfree(iovec);
3164			return -ENOMEM;
3165		}
3166
3167		io_req_map_rw(req, iovec, fast_iov, iter);
3168	}
3169	return 0;
3170}
3171
3172static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3173{
3174	struct io_async_rw *iorw = req->async_data;
3175	struct iovec *iov = iorw->fast_iov;
3176	int ret;
3177
3178	ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3179	if (unlikely(ret < 0))
3180		return ret;
3181
3182	iorw->bytes_done = 0;
3183	iorw->free_iovec = iov;
3184	if (iov)
3185		req->flags |= REQ_F_NEED_CLEANUP;
3186	return 0;
3187}
3188
3189static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3190{
3191	if (unlikely(!(req->file->f_mode & FMODE_READ)))
3192		return -EBADF;
3193	return io_prep_rw(req, sqe);
3194}
3195
3196/*
3197 * This is our waitqueue callback handler, registered through lock_page_async()
3198 * when we initially tried to do the IO with the iocb armed our waitqueue.
3199 * This gets called when the page is unlocked, and we generally expect that to
3200 * happen when the page IO is completed and the page is now uptodate. This will
3201 * queue a task_work based retry of the operation, attempting to copy the data
3202 * again. If the latter fails because the page was NOT uptodate, then we will
3203 * do a thread based blocking retry of the operation. That's the unexpected
3204 * slow path.
3205 */
3206static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3207			     int sync, void *arg)
3208{
3209	struct wait_page_queue *wpq;
3210	struct io_kiocb *req = wait->private;
3211	struct wait_page_key *key = arg;
3212
3213	wpq = container_of(wait, struct wait_page_queue, wait);
3214
3215	if (!wake_page_match(wpq, key))
3216		return 0;
3217
3218	req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3219	list_del_init(&wait->entry);
3220
3221	/* submit ref gets dropped, acquire a new one */
3222	req_ref_get(req);
3223	io_req_task_queue(req);
3224	return 1;
3225}
3226
3227/*
3228 * This controls whether a given IO request should be armed for async page
3229 * based retry. If we return false here, the request is handed to the async
3230 * worker threads for retry. If we're doing buffered reads on a regular file,
3231 * we prepare a private wait_page_queue entry and retry the operation. This
3232 * will either succeed because the page is now uptodate and unlocked, or it
3233 * will register a callback when the page is unlocked at IO completion. Through
3234 * that callback, io_uring uses task_work to setup a retry of the operation.
3235 * That retry will attempt the buffered read again. The retry will generally
3236 * succeed, or in rare cases where it fails, we then fall back to using the
3237 * async worker threads for a blocking retry.
3238 */
3239static bool io_rw_should_retry(struct io_kiocb *req)
3240{
3241	struct io_async_rw *rw = req->async_data;
3242	struct wait_page_queue *wait = &rw->wpq;
3243	struct kiocb *kiocb = &req->rw.kiocb;
3244
3245	/* never retry for NOWAIT, we just complete with -EAGAIN */
3246	if (req->flags & REQ_F_NOWAIT)
3247		return false;
3248
3249	/* Only for buffered IO */
3250	if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3251		return false;
3252
3253	/*
3254	 * just use poll if we can, and don't attempt if the fs doesn't
3255	 * support callback based unlocks
3256	 */
3257	if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3258		return false;
3259
3260	wait->wait.func = io_async_buf_func;
3261	wait->wait.private = req;
3262	wait->wait.flags = 0;
3263	INIT_LIST_HEAD(&wait->wait.entry);
3264	kiocb->ki_flags |= IOCB_WAITQ;
3265	kiocb->ki_flags &= ~IOCB_NOWAIT;
3266	kiocb->ki_waitq = wait;
3267	return true;
3268}
3269
3270static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3271{
3272	if (req->file->f_op->read_iter)
3273		return call_read_iter(req->file, &req->rw.kiocb, iter);
3274	else if (req->file->f_op->read)
3275		return loop_rw_iter(READ, req, iter);
3276	else
3277		return -EINVAL;
3278}
3279
3280static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3281{
3282	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3283	struct kiocb *kiocb = &req->rw.kiocb;
3284	struct iov_iter __iter, *iter = &__iter;
3285	struct io_async_rw *rw = req->async_data;
3286	ssize_t io_size, ret, ret2;
3287	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3288
3289	if (rw) {
3290		iter = &rw->iter;
3291		iovec = NULL;
3292	} else {
3293		ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3294		if (ret < 0)
3295			return ret;
3296	}
3297	io_size = iov_iter_count(iter);
3298	req->result = io_size;
3299
3300	/* Ensure we clear previously set non-block flag */
3301	if (!force_nonblock)
3302		kiocb->ki_flags &= ~IOCB_NOWAIT;
3303	else
3304		kiocb->ki_flags |= IOCB_NOWAIT;
3305
3306	/* If the file doesn't support async, just async punt */
3307	if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3308		ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3309		return ret ?: -EAGAIN;
3310	}
3311
3312	ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3313	if (unlikely(ret)) {
3314		kfree(iovec);
3315		return ret;
3316	}
3317
3318	ret = io_iter_do_read(req, iter);
3319
3320	if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3321		req->flags &= ~REQ_F_REISSUE;
3322		/* IOPOLL retry should happen for io-wq threads */
3323		if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3324			goto done;
3325		/* no retry on NONBLOCK nor RWF_NOWAIT */
3326		if (req->flags & REQ_F_NOWAIT)
3327			goto done;
3328		/* some cases will consume bytes even on error returns */
3329		iov_iter_revert(iter, io_size - iov_iter_count(iter));
3330		ret = 0;
3331	} else if (ret == -EIOCBQUEUED) {
3332		goto out_free;
3333	} else if (ret <= 0 || ret == io_size || !force_nonblock ||
3334		   (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3335		/* read all, failed, already did sync or don't want to retry */
3336		goto done;
3337	}
3338
3339	ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3340	if (ret2)
3341		return ret2;
3342
3343	iovec = NULL;
3344	rw = req->async_data;
3345	/* now use our persistent iterator, if we aren't already */
3346	iter = &rw->iter;
3347
3348	do {
3349		io_size -= ret;
3350		rw->bytes_done += ret;
3351		/* if we can retry, do so with the callbacks armed */
3352		if (!io_rw_should_retry(req)) {
3353			kiocb->ki_flags &= ~IOCB_WAITQ;
3354			return -EAGAIN;
3355		}
3356
3357		/*
3358		 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3359		 * we get -EIOCBQUEUED, then we'll get a notification when the
3360		 * desired page gets unlocked. We can also get a partial read
3361		 * here, and if we do, then just retry at the new offset.
3362		 */
3363		ret = io_iter_do_read(req, iter);
3364		if (ret == -EIOCBQUEUED)
3365			return 0;
3366		/* we got some bytes, but not all. retry. */
3367		kiocb->ki_flags &= ~IOCB_WAITQ;
3368	} while (ret > 0 && ret < io_size);
3369done:
3370	kiocb_done(kiocb, ret, issue_flags);
3371out_free:
3372	/* it's faster to check here then delegate to kfree */
3373	if (iovec)
3374		kfree(iovec);
3375	return 0;
3376}
3377
3378static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3379{
3380	if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3381		return -EBADF;
3382	return io_prep_rw(req, sqe);
3383}
3384
3385static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3386{
3387	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3388	struct kiocb *kiocb = &req->rw.kiocb;
3389	struct iov_iter __iter, *iter = &__iter;
3390	struct io_async_rw *rw = req->async_data;
3391	ssize_t ret, ret2, io_size;
3392	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3393
3394	if (rw) {
3395		iter = &rw->iter;
3396		iovec = NULL;
3397	} else {
3398		ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3399		if (ret < 0)
3400			return ret;
3401	}
3402	io_size = iov_iter_count(iter);
3403	req->result = io_size;
3404
3405	/* Ensure we clear previously set non-block flag */
3406	if (!force_nonblock)
3407		kiocb->ki_flags &= ~IOCB_NOWAIT;
3408	else
3409		kiocb->ki_flags |= IOCB_NOWAIT;
3410
3411	/* If the file doesn't support async, just async punt */
3412	if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3413		goto copy_iov;
3414
3415	/* file path doesn't support NOWAIT for non-direct_IO */
3416	if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3417	    (req->flags & REQ_F_ISREG))
3418		goto copy_iov;
3419
3420	ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3421	if (unlikely(ret))
3422		goto out_free;
3423
3424	/*
3425	 * Open-code file_start_write here to grab freeze protection,
3426	 * which will be released by another thread in
3427	 * io_complete_rw().  Fool lockdep by telling it the lock got
3428	 * released so that it doesn't complain about the held lock when
3429	 * we return to userspace.
3430	 */
3431	if (req->flags & REQ_F_ISREG) {
3432		sb_start_write(file_inode(req->file)->i_sb);
3433		__sb_writers_release(file_inode(req->file)->i_sb,
3434					SB_FREEZE_WRITE);
3435	}
3436	kiocb->ki_flags |= IOCB_WRITE;
3437
3438	if (req->file->f_op->write_iter)
3439		ret2 = call_write_iter(req->file, kiocb, iter);
3440	else if (req->file->f_op->write)
3441		ret2 = loop_rw_iter(WRITE, req, iter);
3442	else
3443		ret2 = -EINVAL;
3444
3445	if (req->flags & REQ_F_REISSUE) {
3446		req->flags &= ~REQ_F_REISSUE;
3447		ret2 = -EAGAIN;
3448	}
3449
3450	/*
3451	 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3452	 * retry them without IOCB_NOWAIT.
3453	 */
3454	if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3455		ret2 = -EAGAIN;
3456	/* no retry on NONBLOCK nor RWF_NOWAIT */
3457	if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3458		goto done;
3459	if (!force_nonblock || ret2 != -EAGAIN) {
3460		/* IOPOLL retry should happen for io-wq threads */
3461		if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3462			goto copy_iov;
3463done:
3464		kiocb_done(kiocb, ret2, issue_flags);
3465	} else {
3466copy_iov:
3467		/* some cases will consume bytes even on error returns */
3468		iov_iter_revert(iter, io_size - iov_iter_count(iter));
3469		ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3470		return ret ?: -EAGAIN;
3471	}
3472out_free:
3473	/* it's reportedly faster than delegating the null check to kfree() */
3474	if (iovec)
3475		kfree(iovec);
3476	return ret;
3477}
3478
3479static int io_renameat_prep(struct io_kiocb *req,
3480			    const struct io_uring_sqe *sqe)
3481{
3482	struct io_rename *ren = &req->rename;
3483	const char __user *oldf, *newf;
3484
3485	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3486		return -EINVAL;
3487	if (sqe->ioprio || sqe->buf_index)
3488		return -EINVAL;
3489	if (unlikely(req->flags & REQ_F_FIXED_FILE))
3490		return -EBADF;
3491
3492	ren->old_dfd = READ_ONCE(sqe->fd);
3493	oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3494	newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3495	ren->new_dfd = READ_ONCE(sqe->len);
3496	ren->flags = READ_ONCE(sqe->rename_flags);
3497
3498	ren->oldpath = getname(oldf);
3499	if (IS_ERR(ren->oldpath))
3500		return PTR_ERR(ren->oldpath);
3501
3502	ren->newpath = getname(newf);
3503	if (IS_ERR(ren->newpath)) {
3504		putname(ren->oldpath);
3505		return PTR_ERR(ren->newpath);
3506	}
3507
3508	req->flags |= REQ_F_NEED_CLEANUP;
3509	return 0;
3510}
3511
3512static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3513{
3514	struct io_rename *ren = &req->rename;
3515	int ret;
3516
3517	if (issue_flags & IO_URING_F_NONBLOCK)
3518		return -EAGAIN;
3519
3520	ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3521				ren->newpath, ren->flags);
3522
3523	req->flags &= ~REQ_F_NEED_CLEANUP;
3524	if (ret < 0)
3525		req_set_fail(req);
3526	io_req_complete(req, ret);
3527	return 0;
3528}
3529
3530static int io_unlinkat_prep(struct io_kiocb *req,
3531			    const struct io_uring_sqe *sqe)
3532{
3533	struct io_unlink *un = &req->unlink;
3534	const char __user *fname;
3535
3536	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3537		return -EINVAL;
3538	if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3539		return -EINVAL;
3540	if (unlikely(req->flags & REQ_F_FIXED_FILE))
3541		return -EBADF;
3542
3543	un->dfd = READ_ONCE(sqe->fd);
3544
3545	un->flags = READ_ONCE(sqe->unlink_flags);
3546	if (un->flags & ~AT_REMOVEDIR)
3547		return -EINVAL;
3548
3549	fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3550	un->filename = getname(fname);
3551	if (IS_ERR(un->filename))
3552		return PTR_ERR(un->filename);
3553
3554	req->flags |= REQ_F_NEED_CLEANUP;
3555	return 0;
3556}
3557
3558static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3559{
3560	struct io_unlink *un = &req->unlink;
3561	int ret;
3562
3563	if (issue_flags & IO_URING_F_NONBLOCK)
3564		return -EAGAIN;
3565
3566	if (un->flags & AT_REMOVEDIR)
3567		ret = do_rmdir(un->dfd, un->filename);
3568	else
3569		ret = do_unlinkat(un->dfd, un->filename);
3570
3571	req->flags &= ~REQ_F_NEED_CLEANUP;
3572	if (ret < 0)
3573		req_set_fail(req);
3574	io_req_complete(req, ret);
3575	return 0;
3576}
3577
3578static int io_shutdown_prep(struct io_kiocb *req,
3579			    const struct io_uring_sqe *sqe)
3580{
3581#if defined(CONFIG_NET)
3582	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3583		return -EINVAL;
3584	if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3585	    sqe->buf_index)
3586		return -EINVAL;
3587
3588	req->shutdown.how = READ_ONCE(sqe->len);
3589	return 0;
3590#else
3591	return -EOPNOTSUPP;
3592#endif
3593}
3594
3595static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3596{
3597#if defined(CONFIG_NET)
3598	struct socket *sock;
3599	int ret;
3600
3601	if (issue_flags & IO_URING_F_NONBLOCK)
3602		return -EAGAIN;
3603
3604	sock = sock_from_file(req->file);
3605	if (unlikely(!sock))
3606		return -ENOTSOCK;
3607
3608	ret = __sys_shutdown_sock(sock, req->shutdown.how);
3609	if (ret < 0)
3610		req_set_fail(req);
3611	io_req_complete(req, ret);
3612	return 0;
3613#else
3614	return -EOPNOTSUPP;
3615#endif
3616}
3617
3618static int __io_splice_prep(struct io_kiocb *req,
3619			    const struct io_uring_sqe *sqe)
3620{
3621	struct io_splice *sp = &req->splice;
3622	unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3623
3624	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3625		return -EINVAL;
3626
3627	sp->file_in = NULL;
3628	sp->len = READ_ONCE(sqe->len);
3629	sp->flags = READ_ONCE(sqe->splice_flags);
3630
3631	if (unlikely(sp->flags & ~valid_flags))
3632		return -EINVAL;
3633
3634	sp->file_in = io_file_get(req->ctx, NULL, req,
3635				  READ_ONCE(sqe->splice_fd_in),
3636				  (sp->flags & SPLICE_F_FD_IN_FIXED));
3637	if (!sp->file_in)
3638		return -EBADF;
3639	req->flags |= REQ_F_NEED_CLEANUP;
3640	return 0;
3641}
3642
3643static int io_tee_prep(struct io_kiocb *req,
3644		       const struct io_uring_sqe *sqe)
3645{
3646	if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3647		return -EINVAL;
3648	return __io_splice_prep(req, sqe);
3649}
3650
3651static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3652{
3653	struct io_splice *sp = &req->splice;
3654	struct file *in = sp->file_in;
3655	struct file *out = sp->file_out;
3656	unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3657	long ret = 0;
3658
3659	if (issue_flags & IO_URING_F_NONBLOCK)
3660		return -EAGAIN;
3661	if (sp->len)
3662		ret = do_tee(in, out, sp->len, flags);
3663
3664	if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3665		io_put_file(in);
3666	req->flags &= ~REQ_F_NEED_CLEANUP;
3667
3668	if (ret != sp->len)
3669		req_set_fail(req);
3670	io_req_complete(req, ret);
3671	return 0;
3672}
3673
3674static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3675{
3676	struct io_splice *sp = &req->splice;
3677
3678	sp->off_in = READ_ONCE(sqe->splice_off_in);
3679	sp->off_out = READ_ONCE(sqe->off);
3680	return __io_splice_prep(req, sqe);
3681}
3682
3683static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3684{
3685	struct io_splice *sp = &req->splice;
3686	struct file *in = sp->file_in;
3687	struct file *out = sp->file_out;
3688	unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3689	loff_t *poff_in, *poff_out;
3690	long ret = 0;
3691
3692	if (issue_flags & IO_URING_F_NONBLOCK)
3693		return -EAGAIN;
3694
3695	poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3696	poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3697
3698	if (sp->len)
3699		ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3700
3701	if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3702		io_put_file(in);
3703	req->flags &= ~REQ_F_NEED_CLEANUP;
3704
3705	if (ret != sp->len)
3706		req_set_fail(req);
3707	io_req_complete(req, ret);
3708	return 0;
3709}
3710
3711/*
3712 * IORING_OP_NOP just posts a completion event, nothing else.
3713 */
3714static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3715{
3716	struct io_ring_ctx *ctx = req->ctx;
3717
3718	if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3719		return -EINVAL;
3720
3721	__io_req_complete(req, issue_flags, 0, 0);
3722	return 0;
3723}
3724
3725static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3726{
3727	struct io_ring_ctx *ctx = req->ctx;
3728
3729	if (!req->file)
3730		return -EBADF;
3731
3732	if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3733		return -EINVAL;
3734	if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3735		return -EINVAL;
3736
3737	req->sync.flags = READ_ONCE(sqe->fsync_flags);
3738	if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3739		return -EINVAL;
3740
3741	req->sync.off = READ_ONCE(sqe->off);
3742	req->sync.len = READ_ONCE(sqe->len);
3743	return 0;
3744}
3745
3746static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3747{
3748	loff_t end = req->sync.off + req->sync.len;
3749	int ret;
3750
3751	/* fsync always requires a blocking context */
3752	if (issue_flags & IO_URING_F_NONBLOCK)
3753		return -EAGAIN;
3754
3755	ret = vfs_fsync_range(req->file, req->sync.off,
3756				end > 0 ? end : LLONG_MAX,
3757				req->sync.flags & IORING_FSYNC_DATASYNC);
3758	if (ret < 0)
3759		req_set_fail(req);
3760	io_req_complete(req, ret);
3761	return 0;
3762}
3763
3764static int io_fallocate_prep(struct io_kiocb *req,
3765			     const struct io_uring_sqe *sqe)
3766{
3767	if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3768		return -EINVAL;
3769	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3770		return -EINVAL;
3771
3772	req->sync.off = READ_ONCE(sqe->off);
3773	req->sync.len = READ_ONCE(sqe->addr);
3774	req->sync.mode = READ_ONCE(sqe->len);
3775	return 0;
3776}
3777
3778static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3779{
3780	int ret;
3781
3782	/* fallocate always requiring blocking context */
3783	if (issue_flags & IO_URING_F_NONBLOCK)
3784		return -EAGAIN;
3785	ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3786				req->sync.len);
3787	if (ret < 0)
3788		req_set_fail(req);
3789	io_req_complete(req, ret);
3790	return 0;
3791}
3792
3793static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3794{
3795	const char __user *fname;
3796	int ret;
3797
3798	if (unlikely(sqe->ioprio || sqe->buf_index))
3799		return -EINVAL;
3800	if (unlikely(req->flags & REQ_F_FIXED_FILE))
3801		return -EBADF;
3802
3803	/* open.how should be already initialised */
3804	if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3805		req->open.how.flags |= O_LARGEFILE;
3806
3807	req->open.dfd = READ_ONCE(sqe->fd);
3808	fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3809	req->open.filename = getname(fname);
3810	if (IS_ERR(req->open.filename)) {
3811		ret = PTR_ERR(req->open.filename);
3812		req->open.filename = NULL;
3813		return ret;
3814	}
3815	req->open.nofile = rlimit(RLIMIT_NOFILE);
3816	req->flags |= REQ_F_NEED_CLEANUP;
3817	return 0;
3818}
3819
3820static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3821{
3822	u64 flags, mode;
3823
3824	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3825		return -EINVAL;
3826	mode = READ_ONCE(sqe->len);
3827	flags = READ_ONCE(sqe->open_flags);
3828	req->open.how = build_open_how(flags, mode);
3829	return __io_openat_prep(req, sqe);
3830}
3831
3832static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3833{
3834	struct open_how __user *how;
3835	size_t len;
3836	int ret;
3837
3838	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3839		return -EINVAL;
3840	how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3841	len = READ_ONCE(sqe->len);
3842	if (len < OPEN_HOW_SIZE_VER0)
3843		return -EINVAL;
3844
3845	ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3846					len);
3847	if (ret)
3848		return ret;
3849
3850	return __io_openat_prep(req, sqe);
3851}
3852
3853static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3854{
3855	struct open_flags op;
3856	struct file *file;
3857	bool nonblock_set;
3858	bool resolve_nonblock;
3859	int ret;
3860
3861	ret = build_open_flags(&req->open.how, &op);
3862	if (ret)
3863		goto err;
3864	nonblock_set = op.open_flag & O_NONBLOCK;
3865	resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3866	if (issue_flags & IO_URING_F_NONBLOCK) {
3867		/*
3868		 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3869		 * it'll always -EAGAIN
3870		 */
3871		if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3872			return -EAGAIN;
3873		op.lookup_flags |= LOOKUP_CACHED;
3874		op.open_flag |= O_NONBLOCK;
3875	}
3876
3877	ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3878	if (ret < 0)
3879		goto err;
3880
3881	file = do_filp_open(req->open.dfd, req->open.filename, &op);
3882	if (IS_ERR(file)) {
3883		/*
3884		 * We could hang on to this 'fd' on retrying, but seems like
3885		 * marginal gain for something that is now known to be a slower
3886		 * path. So just put it, and we'll get a new one when we retry.
3887		 */
3888		put_unused_fd(ret);
3889
3890		ret = PTR_ERR(file);
3891		/* only retry if RESOLVE_CACHED wasn't already set by application */
3892		if (ret == -EAGAIN &&
3893		    (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3894			return -EAGAIN;
3895		goto err;
3896	}
3897
3898	if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3899		file->f_flags &= ~O_NONBLOCK;
3900	fsnotify_open(file);
3901	fd_install(ret, file);
3902err:
3903	putname(req->open.filename);
3904	req->flags &= ~REQ_F_NEED_CLEANUP;
3905	if (ret < 0)
3906		req_set_fail(req);
3907	__io_req_complete(req, issue_flags, ret, 0);
3908	return 0;
3909}
3910
3911static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3912{
3913	return io_openat2(req, issue_flags);
3914}
3915
3916static int io_remove_buffers_prep(struct io_kiocb *req,
3917				  const struct io_uring_sqe *sqe)
3918{
3919	struct io_provide_buf *p = &req->pbuf;
3920	u64 tmp;
3921
3922	if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3923		return -EINVAL;
3924
3925	tmp = READ_ONCE(sqe->fd);
3926	if (!tmp || tmp > USHRT_MAX)
3927		return -EINVAL;
3928
3929	memset(p, 0, sizeof(*p));
3930	p->nbufs = tmp;
3931	p->bgid = READ_ONCE(sqe->buf_group);
3932	return 0;
3933}
3934
3935static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3936			       int bgid, unsigned nbufs)
3937{
3938	unsigned i = 0;
3939
3940	/* shouldn't happen */
3941	if (!nbufs)
3942		return 0;
3943
3944	/* the head kbuf is the list itself */
3945	while (!list_empty(&buf->list)) {
3946		struct io_buffer *nxt;
3947
3948		nxt = list_first_entry(&buf->list, struct io_buffer, list);
3949		list_del(&nxt->list);
3950		kfree(nxt);
3951		if (++i == nbufs)
3952			return i;
3953	}
3954	i++;
3955	kfree(buf);
3956	xa_erase(&ctx->io_buffers, bgid);
3957
3958	return i;
3959}
3960
3961static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3962{
3963	struct io_provide_buf *p = &req->pbuf;
3964	struct io_ring_ctx *ctx = req->ctx;
3965	struct io_buffer *head;
3966	int ret = 0;
3967	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3968
3969	io_ring_submit_lock(ctx, !force_nonblock);
3970
3971	lockdep_assert_held(&ctx->uring_lock);
3972
3973	ret = -ENOENT;
3974	head = xa_load(&ctx->io_buffers, p->bgid);
3975	if (head)
3976		ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3977	if (ret < 0)
3978		req_set_fail(req);
3979
3980	/* complete before unlock, IOPOLL may need the lock */
3981	__io_req_complete(req, issue_flags, ret, 0);
3982	io_ring_submit_unlock(ctx, !force_nonblock);
3983	return 0;
3984}
3985
3986static int io_provide_buffers_prep(struct io_kiocb *req,
3987				   const struct io_uring_sqe *sqe)
3988{
3989	unsigned long size, tmp_check;
3990	struct io_provide_buf *p = &req->pbuf;
3991	u64 tmp;
3992
3993	if (sqe->ioprio || sqe->rw_flags)
3994		return -EINVAL;
3995
3996	tmp = READ_ONCE(sqe->fd);
3997	if (!tmp || tmp > USHRT_MAX)
3998		return -E2BIG;
3999	p->nbufs = tmp;
4000	p->addr = READ_ONCE(sqe->addr);
4001	p->len = READ_ONCE(sqe->len);
4002
4003	if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4004				&size))
4005		return -EOVERFLOW;
4006	if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4007		return -EOVERFLOW;
4008
4009	size = (unsigned long)p->len * p->nbufs;
4010	if (!access_ok(u64_to_user_ptr(p->addr), size))
4011		return -EFAULT;
4012
4013	p->bgid = READ_ONCE(sqe->buf_group);
4014	tmp = READ_ONCE(sqe->off);
4015	if (tmp > USHRT_MAX)
4016		return -E2BIG;
4017	p->bid = tmp;
4018	return 0;
4019}
4020
4021static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4022{
4023	struct io_buffer *buf;
4024	u64 addr = pbuf->addr;
4025	int i, bid = pbuf->bid;
4026
4027	for (i = 0; i < pbuf->nbufs; i++) {
4028		buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4029		if (!buf)
4030			break;
4031
4032		buf->addr = addr;
4033		buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4034		buf->bid = bid;
4035		addr += pbuf->len;
4036		bid++;
4037		if (!*head) {
4038			INIT_LIST_HEAD(&buf->list);
4039			*head = buf;
4040		} else {
4041			list_add_tail(&buf->list, &(*head)->list);
4042		}
4043	}
4044
4045	return i ? i : -ENOMEM;
4046}
4047
4048static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4049{
4050	struct io_provide_buf *p = &req->pbuf;
4051	struct io_ring_ctx *ctx = req->ctx;
4052	struct io_buffer *head, *list;
4053	int ret = 0;
4054	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4055
4056	io_ring_submit_lock(ctx, !force_nonblock);
4057
4058	lockdep_assert_held(&ctx->uring_lock);
4059
4060	list = head = xa_load(&ctx->io_buffers, p->bgid);
4061
4062	ret = io_add_buffers(p, &head);
4063	if (ret >= 0 && !list) {
4064		ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4065		if (ret < 0)
4066			__io_remove_buffers(ctx, head, p->bgid, -1U);
4067	}
4068	if (ret < 0)
4069		req_set_fail(req);
4070	/* complete before unlock, IOPOLL may need the lock */
4071	__io_req_complete(req, issue_flags, ret, 0);
4072	io_ring_submit_unlock(ctx, !force_nonblock);
4073	return 0;
4074}
4075
4076static int io_epoll_ctl_prep(struct io_kiocb *req,
4077			     const struct io_uring_sqe *sqe)
4078{
4079#if defined(CONFIG_EPOLL)
4080	if (sqe->ioprio || sqe->buf_index)
4081		return -EINVAL;
4082	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4083		return -EINVAL;
4084
4085	req->epoll.epfd = READ_ONCE(sqe->fd);
4086	req->epoll.op = READ_ONCE(sqe->len);
4087	req->epoll.fd = READ_ONCE(sqe->off);
4088
4089	if (ep_op_has_event(req->epoll.op)) {
4090		struct epoll_event __user *ev;
4091
4092		ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4093		if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4094			return -EFAULT;
4095	}
4096
4097	return 0;
4098#else
4099	return -EOPNOTSUPP;
4100#endif
4101}
4102
4103static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4104{
4105#if defined(CONFIG_EPOLL)
4106	struct io_epoll *ie = &req->epoll;
4107	int ret;
4108	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4109
4110	ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4111	if (force_nonblock && ret == -EAGAIN)
4112		return -EAGAIN;
4113
4114	if (ret < 0)
4115		req_set_fail(req);
4116	__io_req_complete(req, issue_flags, ret, 0);
4117	return 0;
4118#else
4119	return -EOPNOTSUPP;
4120#endif
4121}
4122
4123static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4124{
4125#if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4126	if (sqe->ioprio || sqe->buf_index || sqe->off)
4127		return -EINVAL;
4128	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4129		return -EINVAL;
4130
4131	req->madvise.addr = READ_ONCE(sqe->addr);
4132	req->madvise.len = READ_ONCE(sqe->len);
4133	req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4134	return 0;
4135#else
4136	return -EOPNOTSUPP;
4137#endif
4138}
4139
4140static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4141{
4142#if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4143	struct io_madvise *ma = &req->madvise;
4144	int ret;
4145
4146	if (issue_flags & IO_URING_F_NONBLOCK)
4147		return -EAGAIN;
4148
4149	ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4150	if (ret < 0)
4151		req_set_fail(req);
4152	io_req_complete(req, ret);
4153	return 0;
4154#else
4155	return -EOPNOTSUPP;
4156#endif
4157}
4158
4159static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4160{
4161	if (sqe->ioprio || sqe->buf_index || sqe->addr)
4162		return -EINVAL;
4163	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4164		return -EINVAL;
4165
4166	req->fadvise.offset = READ_ONCE(sqe->off);
4167	req->fadvise.len = READ_ONCE(sqe->len);
4168	req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4169	return 0;
4170}
4171
4172static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4173{
4174	struct io_fadvise *fa = &req->fadvise;
4175	int ret;
4176
4177	if (issue_flags & IO_URING_F_NONBLOCK) {
4178		switch (fa->advice) {
4179		case POSIX_FADV_NORMAL:
4180		case POSIX_FADV_RANDOM:
4181		case POSIX_FADV_SEQUENTIAL:
4182			break;
4183		default:
4184			return -EAGAIN;
4185		}
4186	}
4187
4188	ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4189	if (ret < 0)
4190		req_set_fail(req);
4191	__io_req_complete(req, issue_flags, ret, 0);
4192	return 0;
4193}
4194
4195static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4196{
4197	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4198		return -EINVAL;
4199	if (sqe->ioprio || sqe->buf_index)
4200		return -EINVAL;
4201	if (req->flags & REQ_F_FIXED_FILE)
4202		return -EBADF;
4203
4204	req->statx.dfd = READ_ONCE(sqe->fd);
4205	req->statx.mask = READ_ONCE(sqe->len);
4206	req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4207	req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4208	req->statx.flags = READ_ONCE(sqe->statx_flags);
4209
4210	return 0;
4211}
4212
4213static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4214{
4215	struct io_statx *ctx = &req->statx;
4216	int ret;
4217
4218	if (issue_flags & IO_URING_F_NONBLOCK)
4219		return -EAGAIN;
4220
4221	ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4222		       ctx->buffer);
4223
4224	if (ret < 0)
4225		req_set_fail(req);
4226	io_req_complete(req, ret);
4227	return 0;
4228}
4229
4230static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4231{
4232	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4233		return -EINVAL;
4234	if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4235	    sqe->rw_flags || sqe->buf_index)
4236		return -EINVAL;
4237	if (req->flags & REQ_F_FIXED_FILE)
4238		return -EBADF;
4239
4240	req->close.fd = READ_ONCE(sqe->fd);
4241	return 0;
4242}
4243
4244static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4245{
4246	struct files_struct *files = current->files;
4247	struct io_close *close = &req->close;
4248	struct fdtable *fdt;
4249	struct file *file = NULL;
4250	int ret = -EBADF;
4251
4252	spin_lock(&files->file_lock);
4253	fdt = files_fdtable(files);
4254	if (close->fd >= fdt->max_fds) {
4255		spin_unlock(&files->file_lock);
4256		goto err;
4257	}
4258	file = fdt->fd[close->fd];
4259	if (!file || file->f_op == &io_uring_fops) {
4260		spin_unlock(&files->file_lock);
4261		file = NULL;
4262		goto err;
4263	}
4264
4265	/* if the file has a flush method, be safe and punt to async */
4266	if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4267		spin_unlock(&files->file_lock);
4268		return -EAGAIN;
4269	}
4270
4271	ret = __close_fd_get_file(close->fd, &file);
4272	spin_unlock(&files->file_lock);
4273	if (ret < 0) {
4274		if (ret == -ENOENT)
4275			ret = -EBADF;
4276		goto err;
4277	}
4278
4279	/* No ->flush() or already async, safely close from here */
4280	ret = filp_close(file, current->files);
4281err:
4282	if (ret < 0)
4283		req_set_fail(req);
4284	if (file)
4285		fput(file);
4286	__io_req_complete(req, issue_flags, ret, 0);
4287	return 0;
4288}
4289
4290static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4291{
4292	struct io_ring_ctx *ctx = req->ctx;
4293
4294	if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4295		return -EINVAL;
4296	if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4297		return -EINVAL;
4298
4299	req->sync.off = READ_ONCE(sqe->off);
4300	req->sync.len = READ_ONCE(sqe->len);
4301	req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4302	return 0;
4303}
4304
4305static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4306{
4307	int ret;
4308
4309	/* sync_file_range always requires a blocking context */
4310	if (issue_flags & IO_URING_F_NONBLOCK)
4311		return -EAGAIN;
4312
4313	ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4314				req->sync.flags);
4315	if (ret < 0)
4316		req_set_fail(req);
4317	io_req_complete(req, ret);
4318	return 0;
4319}
4320
4321#if defined(CONFIG_NET)
4322static int io_setup_async_msg(struct io_kiocb *req,
4323			      struct io_async_msghdr *kmsg)
4324{
4325	struct io_async_msghdr *async_msg = req->async_data;
4326
4327	if (async_msg)
4328		return -EAGAIN;
4329	if (io_alloc_async_data(req)) {
4330		kfree(kmsg->free_iov);
4331		return -ENOMEM;
4332	}
4333	async_msg = req->async_data;
4334	req->flags |= REQ_F_NEED_CLEANUP;
4335	memcpy(async_msg, kmsg, sizeof(*kmsg));
4336	async_msg->msg.msg_name = &async_msg->addr;
4337	/* if were using fast_iov, set it to the new one */
4338	if (!async_msg->free_iov)
4339		async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4340
4341	return -EAGAIN;
4342}
4343
4344static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4345			       struct io_async_msghdr *iomsg)
4346{
4347	iomsg->msg.msg_name = &iomsg->addr;
4348	iomsg->free_iov = iomsg->fast_iov;
4349	return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4350				   req->sr_msg.msg_flags, &iomsg->free_iov);
4351}
4352
4353static int io_sendmsg_prep_async(struct io_kiocb *req)
4354{
4355	int ret;
4356
4357	ret = io_sendmsg_copy_hdr(req, req->async_data);
4358	if (!ret)
4359		req->flags |= REQ_F_NEED_CLEANUP;
4360	return ret;
4361}
4362
4363static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4364{
4365	struct io_sr_msg *sr = &req->sr_msg;
4366
4367	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4368		return -EINVAL;
4369
4370	sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4371	sr->len = READ_ONCE(sqe->len);
4372	sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4373	if (sr->msg_flags & MSG_DONTWAIT)
4374		req->flags |= REQ_F_NOWAIT;
4375
4376#ifdef CONFIG_COMPAT
4377	if (req->ctx->compat)
4378		sr->msg_flags |= MSG_CMSG_COMPAT;
4379#endif
4380	return 0;
4381}
4382
4383static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4384{
4385	struct io_async_msghdr iomsg, *kmsg;
4386	struct socket *sock;
4387	unsigned flags;
4388	int min_ret = 0;
4389	int ret;
4390
4391	sock = sock_from_file(req->file);
4392	if (unlikely(!sock))
4393		return -ENOTSOCK;
4394
4395	kmsg = req->async_data;
4396	if (!kmsg) {
4397		ret = io_sendmsg_copy_hdr(req, &iomsg);
4398		if (ret)
4399			return ret;
4400		kmsg = &iomsg;
4401	}
4402
4403	flags = req->sr_msg.msg_flags;
4404	if (issue_flags & IO_URING_F_NONBLOCK)
4405		flags |= MSG_DONTWAIT;
4406	if (flags & MSG_WAITALL)
4407		min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4408
4409	ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4410	if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4411		return io_setup_async_msg(req, kmsg);
4412	if (ret == -ERESTARTSYS)
4413		ret = -EINTR;
4414
4415	/* fast path, check for non-NULL to avoid function call */
4416	if (kmsg->free_iov)
4417		kfree(kmsg->free_iov);
4418	req->flags &= ~REQ_F_NEED_CLEANUP;
4419	if (ret < min_ret)
4420		req_set_fail(req);
4421	__io_req_complete(req, issue_flags, ret, 0);
4422	return 0;
4423}
4424
4425static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4426{
4427	struct io_sr_msg *sr = &req->sr_msg;
4428	struct msghdr msg;
4429	struct iovec iov;
4430	struct socket *sock;
4431	unsigned flags;
4432	int min_ret = 0;
4433	int ret;
4434
4435	sock = sock_from_file(req->file);
4436	if (unlikely(!sock))
4437		return -ENOTSOCK;
4438
4439	ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4440	if (unlikely(ret))
4441		return ret;
4442
4443	msg.msg_name = NULL;
4444	msg.msg_control = NULL;
4445	msg.msg_controllen = 0;
4446	msg.msg_namelen = 0;
4447
4448	flags = req->sr_msg.msg_flags;
4449	if (issue_flags & IO_URING_F_NONBLOCK)
4450		flags |= MSG_DONTWAIT;
4451	if (flags & MSG_WAITALL)
4452		min_ret = iov_iter_count(&msg.msg_iter);
4453
4454	msg.msg_flags = flags;
4455	ret = sock_sendmsg(sock, &msg);
4456	if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4457		return -EAGAIN;
4458	if (ret == -ERESTARTSYS)
4459		ret = -EINTR;
4460
4461	if (ret < min_ret)
4462		req_set_fail(req);
4463	__io_req_complete(req, issue_flags, ret, 0);
4464	return 0;
4465}
4466
4467static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4468				 struct io_async_msghdr *iomsg)
4469{
4470	struct io_sr_msg *sr = &req->sr_msg;
4471	struct iovec __user *uiov;
4472	size_t iov_len;
4473	int ret;
4474
4475	ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4476					&iomsg->uaddr, &uiov, &iov_len);
4477	if (ret)
4478		return ret;
4479
4480	if (req->flags & REQ_F_BUFFER_SELECT) {
4481		if (iov_len > 1)
4482			return -EINVAL;
4483		if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4484			return -EFAULT;
4485		sr->len = iomsg->fast_iov[0].iov_len;
4486		iomsg->free_iov = NULL;
4487	} else {
4488		iomsg->free_iov = iomsg->fast_iov;
4489		ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4490				     &iomsg->free_iov, &iomsg->msg.msg_iter,
4491				     false);
4492		if (ret > 0)
4493			ret = 0;
4494	}
4495
4496	return ret;
4497}
4498
4499#ifdef CONFIG_COMPAT
4500static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4501					struct io_async_msghdr *iomsg)
4502{
4503	struct io_sr_msg *sr = &req->sr_msg;
4504	struct compat_iovec __user *uiov;
4505	compat_uptr_t ptr;
4506	compat_size_t len;
4507	int ret;
4508
4509	ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4510				  &ptr, &len);
4511	if (ret)
4512		return ret;
4513
4514	uiov = compat_ptr(ptr);
4515	if (req->flags & REQ_F_BUFFER_SELECT) {
4516		compat_ssize_t clen;
4517
4518		if (len > 1)
4519			return -EINVAL;
4520		if (!access_ok(uiov, sizeof(*uiov)))
4521			return -EFAULT;
4522		if (__get_user(clen, &uiov->iov_len))
4523			return -EFAULT;
4524		if (clen < 0)
4525			return -EINVAL;
4526		sr->len = clen;
4527		iomsg->free_iov = NULL;
4528	} else {
4529		iomsg->free_iov = iomsg->fast_iov;
4530		ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4531				   UIO_FASTIOV, &iomsg->free_iov,
4532				   &iomsg->msg.msg_iter, true);
4533		if (ret < 0)
4534			return ret;
4535	}
4536
4537	return 0;
4538}
4539#endif
4540
4541static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4542			       struct io_async_msghdr *iomsg)
4543{
4544	iomsg->msg.msg_name = &iomsg->addr;
4545
4546#ifdef CONFIG_COMPAT
4547	if (req->ctx->compat)
4548		return __io_compat_recvmsg_copy_hdr(req, iomsg);
4549#endif
4550
4551	return __io_recvmsg_copy_hdr(req, iomsg);
4552}
4553
4554static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4555					       bool needs_lock)
4556{
4557	struct io_sr_msg *sr = &req->sr_msg;
4558	struct io_buffer *kbuf;
4559
4560	kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4561	if (IS_ERR(kbuf))
4562		return kbuf;
4563
4564	sr->kbuf = kbuf;
4565	req->flags |= REQ_F_BUFFER_SELECTED;
4566	return kbuf;
4567}
4568
4569static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4570{
4571	return io_put_kbuf(req, req->sr_msg.kbuf);
4572}
4573
4574static int io_recvmsg_prep_async(struct io_kiocb *req)
4575{
4576	int ret;
4577
4578	ret = io_recvmsg_copy_hdr(req, req->async_data);
4579	if (!ret)
4580		req->flags |= REQ_F_NEED_CLEANUP;
4581	return ret;
4582}
4583
4584static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4585{
4586	struct io_sr_msg *sr = &req->sr_msg;
4587
4588	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4589		return -EINVAL;
4590
4591	sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4592	sr->len = READ_ONCE(sqe->len);
4593	sr->bgid = READ_ONCE(sqe->buf_group);
4594	sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4595	if (sr->msg_flags & MSG_DONTWAIT)
4596		req->flags |= REQ_F_NOWAIT;
4597
4598#ifdef CONFIG_COMPAT
4599	if (req->ctx->compat)
4600		sr->msg_flags |= MSG_CMSG_COMPAT;
4601#endif
4602	return 0;
4603}
4604
4605static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4606{
4607	struct io_async_msghdr iomsg, *kmsg;
4608	struct socket *sock;
4609	struct io_buffer *kbuf;
4610	unsigned flags;
4611	int min_ret = 0;
4612	int ret, cflags = 0;
4613	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4614
4615	sock = sock_from_file(req->file);
4616	if (unlikely(!sock))
4617		return -ENOTSOCK;
4618
4619	kmsg = req->async_data;
4620	if (!kmsg) {
4621		ret = io_recvmsg_copy_hdr(req, &iomsg);
4622		if (ret)
4623			return ret;
4624		kmsg = &iomsg;
4625	}
4626
4627	if (req->flags & REQ_F_BUFFER_SELECT) {
4628		kbuf = io_recv_buffer_select(req, !force_nonblock);
4629		if (IS_ERR(kbuf))
4630			return PTR_ERR(kbuf);
4631		kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4632		kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4633		iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4634				1, req->sr_msg.len);
4635	}
4636
4637	flags = req->sr_msg.msg_flags;
4638	if (force_nonblock)
4639		flags |= MSG_DONTWAIT;
4640	if (flags & MSG_WAITALL)
4641		min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4642
4643	ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4644					kmsg->uaddr, flags);
4645	if (force_nonblock && ret == -EAGAIN)
4646		return io_setup_async_msg(req, kmsg);
4647	if (ret == -ERESTARTSYS)
4648		ret = -EINTR;
4649
4650	if (req->flags & REQ_F_BUFFER_SELECTED)
4651		cflags = io_put_recv_kbuf(req);
4652	/* fast path, check for non-NULL to avoid function call */
4653	if (kmsg->free_iov)
4654		kfree(kmsg->free_iov);
4655	req->flags &= ~REQ_F_NEED_CLEANUP;
4656	if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4657		req_set_fail(req);
4658	__io_req_complete(req, issue_flags, ret, cflags);
4659	return 0;
4660}
4661
4662static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4663{
4664	struct io_buffer *kbuf;
4665	struct io_sr_msg *sr = &req->sr_msg;
4666	struct msghdr msg;
4667	void __user *buf = sr->buf;
4668	struct socket *sock;
4669	struct iovec iov;
4670	unsigned flags;
4671	int min_ret = 0;
4672	int ret, cflags = 0;
4673	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4674
4675	sock = sock_from_file(req->file);
4676	if (unlikely(!sock))
4677		return -ENOTSOCK;
4678
4679	if (req->flags & REQ_F_BUFFER_SELECT) {
4680		kbuf = io_recv_buffer_select(req, !force_nonblock);
4681		if (IS_ERR(kbuf))
4682			return PTR_ERR(kbuf);
4683		buf = u64_to_user_ptr(kbuf->addr);
4684	}
4685
4686	ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4687	if (unlikely(ret))
4688		goto out_free;
4689
4690	msg.msg_name = NULL;
4691	msg.msg_control = NULL;
4692	msg.msg_controllen = 0;
4693	msg.msg_namelen = 0;
4694	msg.msg_iocb = NULL;
4695	msg.msg_flags = 0;
4696
4697	flags = req->sr_msg.msg_flags;
4698	if (force_nonblock)
4699		flags |= MSG_DONTWAIT;
4700	if (flags & MSG_WAITALL)
4701		min_ret = iov_iter_count(&msg.msg_iter);
4702
4703	ret = sock_recvmsg(sock, &msg, flags);
4704	if (force_nonblock && ret == -EAGAIN)
4705		return -EAGAIN;
4706	if (ret == -ERESTARTSYS)
4707		ret = -EINTR;
4708out_free:
4709	if (req->flags & REQ_F_BUFFER_SELECTED)
4710		cflags = io_put_recv_kbuf(req);
4711	if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4712		req_set_fail(req);
4713	__io_req_complete(req, issue_flags, ret, cflags);
4714	return 0;
4715}
4716
4717static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4718{
4719	struct io_accept *accept = &req->accept;
4720
4721	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4722		return -EINVAL;
4723	if (sqe->ioprio || sqe->len || sqe->buf_index)
4724		return -EINVAL;
4725
4726	accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4727	accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4728	accept->flags = READ_ONCE(sqe->accept_flags);
4729	accept->nofile = rlimit(RLIMIT_NOFILE);
4730	return 0;
4731}
4732
4733static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4734{
4735	struct io_accept *accept = &req->accept;
4736	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4737	unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4738	int ret;
4739
4740	if (req->file->f_flags & O_NONBLOCK)
4741		req->flags |= REQ_F_NOWAIT;
4742
4743	ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4744					accept->addr_len, accept->flags,
4745					accept->nofile);
4746	if (ret == -EAGAIN && force_nonblock)
4747		return -EAGAIN;
4748	if (ret < 0) {
4749		if (ret == -ERESTARTSYS)
4750			ret = -EINTR;
4751		req_set_fail(req);
4752	}
4753	__io_req_complete(req, issue_flags, ret, 0);
4754	return 0;
4755}
4756
4757static int io_connect_prep_async(struct io_kiocb *req)
4758{
4759	struct io_async_connect *io = req->async_data;
4760	struct io_connect *conn = &req->connect;
4761
4762	return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4763}
4764
4765static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4766{
4767	struct io_connect *conn = &req->connect;
4768
4769	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4770		return -EINVAL;
4771	if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4772		return -EINVAL;
4773
4774	conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4775	conn->addr_len =  READ_ONCE(sqe->addr2);
4776	return 0;
4777}
4778
4779static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4780{
4781	struct io_async_connect __io, *io;
4782	unsigned file_flags;
4783	int ret;
4784	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4785
4786	if (req->async_data) {
4787		io = req->async_data;
4788	} else {
4789		ret = move_addr_to_kernel(req->connect.addr,
4790						req->connect.addr_len,
4791						&__io.address);
4792		if (ret)
4793			goto out;
4794		io = &__io;
4795	}
4796
4797	file_flags = force_nonblock ? O_NONBLOCK : 0;
4798
4799	ret = __sys_connect_file(req->file, &io->address,
4800					req->connect.addr_len, file_flags);
4801	if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4802		if (req->async_data)
4803			return -EAGAIN;
4804		if (io_alloc_async_data(req)) {
4805			ret = -ENOMEM;
4806			goto out;
4807		}
4808		memcpy(req->async_data, &__io, sizeof(__io));
4809		return -EAGAIN;
4810	}
4811	if (ret == -ERESTARTSYS)
4812		ret = -EINTR;
4813out:
4814	if (ret < 0)
4815		req_set_fail(req);
4816	__io_req_complete(req, issue_flags, ret, 0);
4817	return 0;
4818}
4819#else /* !CONFIG_NET */
4820#define IO_NETOP_FN(op)							\
4821static int io_##op(struct io_kiocb *req, unsigned int issue_flags)	\
4822{									\
4823	return -EOPNOTSUPP;						\
4824}
4825
4826#define IO_NETOP_PREP(op)						\
4827IO_NETOP_FN(op)								\
4828static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4829{									\
4830	return -EOPNOTSUPP;						\
4831}									\
4832
4833#define IO_NETOP_PREP_ASYNC(op)						\
4834IO_NETOP_PREP(op)							\
4835static int io_##op##_prep_async(struct io_kiocb *req)			\
4836{									\
4837	return -EOPNOTSUPP;						\
4838}
4839
4840IO_NETOP_PREP_ASYNC(sendmsg);
4841IO_NETOP_PREP_ASYNC(recvmsg);
4842IO_NETOP_PREP_ASYNC(connect);
4843IO_NETOP_PREP(accept);
4844IO_NETOP_FN(send);
4845IO_NETOP_FN(recv);
4846#endif /* CONFIG_NET */
4847
4848struct io_poll_table {
4849	struct poll_table_struct pt;
4850	struct io_kiocb *req;
4851	int nr_entries;
4852	int error;
4853};
4854
4855static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4856			   __poll_t mask, io_req_tw_func_t func)
4857{
4858	/* for instances that support it check for an event match first: */
4859	if (mask && !(mask & poll->events))
4860		return 0;
4861
4862	trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4863
4864	list_del_init(&poll->wait.entry);
4865
4866	req->result = mask;
4867	req->io_task_work.func = func;
4868
4869	/*
4870	 * If this fails, then the task is exiting. When a task exits, the
4871	 * work gets canceled, so just cancel this request as well instead
4872	 * of executing it. We can't safely execute it anyway, as we may not
4873	 * have the needed state needed for it anyway.
4874	 */
4875	io_req_task_work_add(req);
4876	return 1;
4877}
4878
4879static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4880	__acquires(&req->ctx->completion_lock)
4881{
4882	struct io_ring_ctx *ctx = req->ctx;
4883
4884	if (unlikely(req->task->flags & PF_EXITING))
4885		WRITE_ONCE(poll->canceled, true);
4886
4887	if (!req->result && !READ_ONCE(poll->canceled)) {
4888		struct poll_table_struct pt = { ._key = poll->events };
4889
4890		req->result = vfs_poll(req->file, &pt) & poll->events;
4891	}
4892
4893	spin_lock_irq(&ctx->completion_lock);
4894	if (!req->result && !READ_ONCE(poll->canceled)) {
4895		add_wait_queue(poll->head, &poll->wait);
4896		return true;
4897	}
4898
4899	return false;
4900}
4901
4902static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4903{
4904	/* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4905	if (req->opcode == IORING_OP_POLL_ADD)
4906		return req->async_data;
4907	return req->apoll->double_poll;
4908}
4909
4910static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4911{
4912	if (req->opcode == IORING_OP_POLL_ADD)
4913		return &req->poll;
4914	return &req->apoll->poll;
4915}
4916
4917static void io_poll_remove_double(struct io_kiocb *req)
4918	__must_hold(&req->ctx->completion_lock)
4919{
4920	struct io_poll_iocb *poll = io_poll_get_double(req);
4921
4922	lockdep_assert_held(&req->ctx->completion_lock);
4923
4924	if (poll && poll->head) {
4925		struct wait_queue_head *head = poll->head;
4926
4927		spin_lock(&head->lock);
4928		list_del_init(&poll->wait.entry);
4929		if (poll->wait.private)
4930			req_ref_put(req);
4931		poll->head = NULL;
4932		spin_unlock(&head->lock);
4933	}
4934}
4935
4936static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4937	__must_hold(&req->ctx->completion_lock)
4938{
4939	struct io_ring_ctx *ctx = req->ctx;
4940	unsigned flags = IORING_CQE_F_MORE;
4941	int error;
4942
4943	if (READ_ONCE(req->poll.canceled)) {
4944		error = -ECANCELED;
4945		req->poll.events |= EPOLLONESHOT;
4946	} else {
4947		error = mangle_poll(mask);
4948	}
4949	if (req->poll.events & EPOLLONESHOT)
4950		flags = 0;
4951	if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4952		req->poll.done = true;
4953		flags = 0;
4954	}
4955	if (flags & IORING_CQE_F_MORE)
4956		ctx->cq_extra++;
4957
4958	io_commit_cqring(ctx);
4959	return !(flags & IORING_CQE_F_MORE);
4960}
4961
4962static void io_poll_task_func(struct io_kiocb *req)
4963{
4964	struct io_ring_ctx *ctx = req->ctx;
4965	struct io_kiocb *nxt;
4966
4967	if (io_poll_rewait(req, &req->poll)) {
4968		spin_unlock_irq(&ctx->completion_lock);
4969	} else {
4970		bool done;
4971
4972		done = io_poll_complete(req, req->result);
4973		if (done) {
4974			io_poll_remove_double(req);
4975			hash_del(&req->hash_node);
4976		} else {
4977			req->result = 0;
4978			add_wait_queue(req->poll.head, &req->poll.wait);
4979		}
4980		spin_unlock_irq(&ctx->completion_lock);
4981		io_cqring_ev_posted(ctx);
4982
4983		if (done) {
4984			nxt = io_put_req_find_next(req);
4985			if (nxt)
4986				io_req_task_submit(nxt);
4987		}
4988	}
4989}
4990
4991static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4992			       int sync, void *key)
4993{
4994	struct io_kiocb *req = wait->private;
4995	struct io_poll_iocb *poll = io_poll_get_single(req);
4996	__poll_t mask = key_to_poll(key);
4997
4998	/* for instances that support it check for an event match first: */
4999	if (mask && !(mask & poll->events))
5000		return 0;
5001	if (!(poll->events & EPOLLONESHOT))
5002		return poll->wait.func(&poll->wait, mode, sync, key);
5003
5004	list_del_init(&wait->entry);
5005
5006	if (poll->head) {
5007		bool done;
5008
5009		spin_lock(&poll->head->lock);
5010		done = list_empty(&poll->wait.entry);
5011		if (!done)
5012			list_del_init(&poll->wait.entry);
5013		/* make sure double remove sees this as being gone */
5014		wait->private = NULL;
5015		spin_unlock(&poll->head->lock);
5016		if (!done) {
5017			/* use wait func handler, so it matches the rq type */
5018			poll->wait.func(&poll->wait, mode, sync, key);
5019		}
5020	}
5021	req_ref_put(req);
5022	return 1;
5023}
5024
5025static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5026			      wait_queue_func_t wake_func)
5027{
5028	poll->head = NULL;
5029	poll->done = false;
5030	poll->canceled = false;
5031#define IO_POLL_UNMASK	(EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5032	/* mask in events that we always want/need */
5033	poll->events = events | IO_POLL_UNMASK;
5034	INIT_LIST_HEAD(&poll->wait.entry);
5035	init_waitqueue_func_entry(&poll->wait, wake_func);
5036}
5037
5038static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5039			    struct wait_queue_head *head,
5040			    struct io_poll_iocb **poll_ptr)
5041{
5042	struct io_kiocb *req = pt->req;
5043
5044	/*
5045	 * The file being polled uses multiple waitqueues for poll handling
5046	 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5047	 * if this happens.
5048	 */
5049	if (unlikely(pt->nr_entries)) {
5050		struct io_poll_iocb *poll_one = poll;
5051
5052		/* already have a 2nd entry, fail a third attempt */
5053		if (*poll_ptr) {
5054			pt->error = -EINVAL;
5055			return;
5056		}
5057		/*
5058		 * Can't handle multishot for double wait for now, turn it
5059		 * into one-shot mode.
5060		 */
5061		if (!(poll_one->events & EPOLLONESHOT))
5062			poll_one->events |= EPOLLONESHOT;
5063		/* double add on the same waitqueue head, ignore */
5064		if (poll_one->head == head)
5065			return;
5066		poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5067		if (!poll) {
5068			pt->error = -ENOMEM;
5069			return;
5070		}
5071		io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5072		req_ref_get(req);
5073		poll->wait.private = req;
5074		*poll_ptr = poll;
5075	}
5076
5077	pt->nr_entries++;
5078	poll->head = head;
5079
5080	if (poll->events & EPOLLEXCLUSIVE)
5081		add_wait_queue_exclusive(head, &poll->wait);
5082	else
5083		add_wait_queue(head, &poll->wait);
5084}
5085
5086static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5087			       struct poll_table_struct *p)
5088{
5089	struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5090	struct async_poll *apoll = pt->req->apoll;
5091
5092	__io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5093}
5094
5095static void io_async_task_func(struct io_kiocb *req)
5096{
5097	struct async_poll *apoll = req->apoll;
5098	struct io_ring_ctx *ctx = req->ctx;
5099
5100	trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5101
5102	if (io_poll_rewait(req, &apoll->poll)) {
5103		spin_unlock_irq(&ctx->completion_lock);
5104		return;
5105	}
5106
5107	hash_del(&req->hash_node);
5108	io_poll_remove_double(req);
5109	spin_unlock_irq(&ctx->completion_lock);
5110
5111	if (!READ_ONCE(apoll->poll.canceled))
5112		io_req_task_submit(req);
5113	else
5114		io_req_complete_failed(req, -ECANCELED);
5115}
5116
5117static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5118			void *key)
5119{
5120	struct io_kiocb *req = wait->private;
5121	struct io_poll_iocb *poll = &req->apoll->poll;
5122
5123	trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5124					key_to_poll(key));
5125
5126	return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5127}
5128
5129static void io_poll_req_insert(struct io_kiocb *req)
5130{
5131	struct io_ring_ctx *ctx = req->ctx;
5132	struct hlist_head *list;
5133
5134	list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5135	hlist_add_head(&req->hash_node, list);
5136}
5137
5138static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5139				      struct io_poll_iocb *poll,
5140				      struct io_poll_table *ipt, __poll_t mask,
5141				      wait_queue_func_t wake_func)
5142	__acquires(&ctx->completion_lock)
5143{
5144	struct io_ring_ctx *ctx = req->ctx;
5145	bool cancel = false;
5146
5147	INIT_HLIST_NODE(&req->hash_node);
5148	io_init_poll_iocb(poll, mask, wake_func);
5149	poll->file = req->file;
5150	poll->wait.private = req;
5151
5152	ipt->pt._key = mask;
5153	ipt->req = req;
5154	ipt->error = 0;
5155	ipt->nr_entries = 0;
5156
5157	mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5158	if (unlikely(!ipt->nr_entries) && !ipt->error)
5159		ipt->error = -EINVAL;
5160
5161	spin_lock_irq(&ctx->completion_lock);
5162	if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5163		io_poll_remove_double(req);
5164	if (likely(poll->head)) {
5165		spin_lock(&poll->head->lock);
5166		if (unlikely(list_empty(&poll->wait.entry))) {
5167			if (ipt->error)
5168				cancel = true;
5169			ipt->error = 0;
5170			mask = 0;
5171		}
5172		if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5173			list_del_init(&poll->wait.entry);
5174		else if (cancel)
5175			WRITE_ONCE(poll->canceled, true);
5176		else if (!poll->done) /* actually waiting for an event */
5177			io_poll_req_insert(req);
5178		spin_unlock(&poll->head->lock);
5179	}
5180
5181	return mask;
5182}
5183
5184enum {
5185	IO_APOLL_OK,
5186	IO_APOLL_ABORTED,
5187	IO_APOLL_READY
5188};
5189
5190static int io_arm_poll_handler(struct io_kiocb *req)
5191{
5192	const struct io_op_def *def = &io_op_defs[req->opcode];
5193	struct io_ring_ctx *ctx = req->ctx;
5194	struct async_poll *apoll;
5195	struct io_poll_table ipt;
5196	__poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5197	int rw;
5198
5199	if (!req->file || !file_can_poll(req->file))
5200		return IO_APOLL_ABORTED;
5201	if (req->flags & REQ_F_POLLED)
5202		return IO_APOLL_ABORTED;
5203	if (!def->pollin && !def->pollout)
5204		return IO_APOLL_ABORTED;
5205
5206	if (def->pollin) {
5207		rw = READ;
5208		mask |= POLLIN | POLLRDNORM;
5209
5210		/* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5211		if ((req->opcode == IORING_OP_RECVMSG) &&
5212		    (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5213			mask &= ~POLLIN;
5214	} else {
5215		rw = WRITE;
5216		mask |= POLLOUT | POLLWRNORM;
5217	}
5218
5219	/* if we can't nonblock try, then no point in arming a poll handler */
5220	if (!io_file_supports_nowait(req, rw))
5221		return IO_APOLL_ABORTED;
5222
5223	apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5224	if (unlikely(!apoll))
5225		return IO_APOLL_ABORTED;
5226	apoll->double_poll = NULL;
5227	req->apoll = apoll;
5228	req->flags |= REQ_F_POLLED;
5229	ipt.pt._qproc = io_async_queue_proc;
5230
5231	ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5232					io_async_wake);
5233	if (ret || ipt.error) {
5234		spin_unlock_irq(&ctx->completion_lock);
5235		if (ret)
5236			return IO_APOLL_READY;
5237		return IO_APOLL_ABORTED;
5238	}
5239	spin_unlock_irq(&ctx->completion_lock);
5240	trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5241				mask, apoll->poll.events);
5242	return IO_APOLL_OK;
5243}
5244
5245static bool __io_poll_remove_one(struct io_kiocb *req,
5246				 struct io_poll_iocb *poll, bool do_cancel)
5247	__must_hold(&req->ctx->completion_lock)
5248{
5249	bool do_complete = false;
5250
5251	if (!poll->head)
5252		return false;
5253	spin_lock(&poll->head->lock);
5254	if (do_cancel)
5255		WRITE_ONCE(poll->canceled, true);
5256	if (!list_empty(&poll->wait.entry)) {
5257		list_del_init(&poll->wait.entry);
5258		do_complete = true;
5259	}
5260	spin_unlock(&poll->head->lock);
5261	hash_del(&req->hash_node);
5262	return do_complete;
5263}
5264
5265static bool io_poll_remove_waitqs(struct io_kiocb *req)
5266	__must_hold(&req->ctx->completion_lock)
5267{
5268	bool do_complete;
5269
5270	io_poll_remove_double(req);
5271	do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5272
5273	if (req->opcode != IORING_OP_POLL_ADD && do_complete) {
5274		/* non-poll requests have submit ref still */
5275		req_ref_put(req);
5276	}
5277	return do_complete;
5278}
5279
5280static bool io_poll_remove_one(struct io_kiocb *req)
5281	__must_hold(&req->ctx->completion_lock)
5282{
5283	bool do_complete;
5284
5285	do_complete = io_poll_remove_waitqs(req);
5286	if (do_complete) {
5287		io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5288		io_commit_cqring(req->ctx);
5289		req_set_fail(req);
5290		io_put_req_deferred(req, 1);
5291	}
5292
5293	return do_complete;
5294}
5295
5296/*
5297 * Returns true if we found and killed one or more poll requests
5298 */
5299static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5300			       bool cancel_all)
5301{
5302	struct hlist_node *tmp;
5303	struct io_kiocb *req;
5304	int posted = 0, i;
5305
5306	spin_lock_irq(&ctx->completion_lock);
5307	for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5308		struct hlist_head *list;
5309
5310		list = &ctx->cancel_hash[i];
5311		hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5312			if (io_match_task(req, tsk, cancel_all))
5313				posted += io_poll_remove_one(req);
5314		}
5315	}
5316	spin_unlock_irq(&ctx->completion_lock);
5317
5318	if (posted)
5319		io_cqring_ev_posted(ctx);
5320
5321	return posted != 0;
5322}
5323
5324static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5325				     bool poll_only)
5326	__must_hold(&ctx->completion_lock)
5327{
5328	struct hlist_head *list;
5329	struct io_kiocb *req;
5330
5331	list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5332	hlist_for_each_entry(req, list, hash_node) {
5333		if (sqe_addr != req->user_data)
5334			continue;
5335		if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5336			continue;
5337		return req;
5338	}
5339	return NULL;
5340}
5341
5342static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5343			  bool poll_only)
5344	__must_hold(&ctx->completion_lock)
5345{
5346	struct io_kiocb *req;
5347
5348	req = io_poll_find(ctx, sqe_addr, poll_only);
5349	if (!req)
5350		return -ENOENT;
5351	if (io_poll_remove_one(req))
5352		return 0;
5353
5354	return -EALREADY;
5355}
5356
5357static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5358				     unsigned int flags)
5359{
5360	u32 events;
5361
5362	events = READ_ONCE(sqe->poll32_events);
5363#ifdef __BIG_ENDIAN
5364	events = swahw32(events);
5365#endif
5366	if (!(flags & IORING_POLL_ADD_MULTI))
5367		events |= EPOLLONESHOT;
5368	return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5369}
5370
5371static int io_poll_update_prep(struct io_kiocb *req,
5372			       const struct io_uring_sqe *sqe)
5373{
5374	struct io_poll_update *upd = &req->poll_update;
5375	u32 flags;
5376
5377	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5378		return -EINVAL;
5379	if (sqe->ioprio || sqe->buf_index)
5380		return -EINVAL;
5381	flags = READ_ONCE(sqe->len);
5382	if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5383		      IORING_POLL_ADD_MULTI))
5384		return -EINVAL;
5385	/* meaningless without update */
5386	if (flags == IORING_POLL_ADD_MULTI)
5387		return -EINVAL;
5388
5389	upd->old_user_data = READ_ONCE(sqe->addr);
5390	upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5391	upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5392
5393	upd->new_user_data = READ_ONCE(sqe->off);
5394	if (!upd->update_user_data && upd->new_user_data)
5395		return -EINVAL;
5396	if (upd->update_events)
5397		upd->events = io_poll_parse_events(sqe, flags);
5398	else if (sqe->poll32_events)
5399		return -EINVAL;
5400
5401	return 0;
5402}
5403
5404static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5405			void *key)
5406{
5407	struct io_kiocb *req = wait->private;
5408	struct io_poll_iocb *poll = &req->poll;
5409
5410	return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5411}
5412
5413static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5414			       struct poll_table_struct *p)
5415{
5416	struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5417
5418	__io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5419}
5420
5421static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5422{
5423	struct io_poll_iocb *poll = &req->poll;
5424	u32 flags;
5425
5426	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5427		return -EINVAL;
5428	if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5429		return -EINVAL;
5430	flags = READ_ONCE(sqe->len);
5431	if (flags & ~IORING_POLL_ADD_MULTI)
5432		return -EINVAL;
5433
5434	poll->events = io_poll_parse_events(sqe, flags);
5435	return 0;
5436}
5437
5438static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5439{
5440	struct io_poll_iocb *poll = &req->poll;
5441	struct io_ring_ctx *ctx = req->ctx;
5442	struct io_poll_table ipt;
5443	__poll_t mask;
5444
5445	ipt.pt._qproc = io_poll_queue_proc;
5446
5447	mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5448					io_poll_wake);
5449
5450	if (mask) { /* no async, we'd stolen it */
5451		ipt.error = 0;
5452		io_poll_complete(req, mask);
5453	}
5454	spin_unlock_irq(&ctx->completion_lock);
5455
5456	if (mask) {
5457		io_cqring_ev_posted(ctx);
5458		if (poll->events & EPOLLONESHOT)
5459			io_put_req(req);
5460	}
5461	return ipt.error;
5462}
5463
5464static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5465{
5466	struct io_ring_ctx *ctx = req->ctx;
5467	struct io_kiocb *preq;
5468	bool completing;
5469	int ret;
5470
5471	spin_lock_irq(&ctx->completion_lock);
5472	preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5473	if (!preq) {
5474		ret = -ENOENT;
5475		goto err;
5476	}
5477
5478	if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5479		completing = true;
5480		ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5481		goto err;
5482	}
5483
5484	/*
5485	 * Don't allow racy completion with singleshot, as we cannot safely
5486	 * update those. For multishot, if we're racing with completion, just
5487	 * let completion re-add it.
5488	 */
5489	completing = !__io_poll_remove_one(preq, &preq->poll, false);
5490	if (completing && (preq->poll.events & EPOLLONESHOT)) {
5491		ret = -EALREADY;
5492		goto err;
5493	}
5494	/* we now have a detached poll request. reissue. */
5495	ret = 0;
5496err:
5497	if (ret < 0) {
5498		spin_unlock_irq(&ctx->completion_lock);
5499		req_set_fail(req);
5500		io_req_complete(req, ret);
5501		return 0;
5502	}
5503	/* only mask one event flags, keep behavior flags */
5504	if (req->poll_update.update_events) {
5505		preq->poll.events &= ~0xffff;
5506		preq->poll.events |= req->poll_update.events & 0xffff;
5507		preq->poll.events |= IO_POLL_UNMASK;
5508	}
5509	if (req->poll_update.update_user_data)
5510		preq->user_data = req->poll_update.new_user_data;
5511	spin_unlock_irq(&ctx->completion_lock);
5512
5513	/* complete update request, we're done with it */
5514	io_req_complete(req, ret);
5515
5516	if (!completing) {
5517		ret = io_poll_add(preq, issue_flags);
5518		if (ret < 0) {
5519			req_set_fail(preq);
5520			io_req_complete(preq, ret);
5521		}
5522	}
5523	return 0;
5524}
5525
5526static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5527{
5528	struct io_timeout_data *data = container_of(timer,
5529						struct io_timeout_data, timer);
5530	struct io_kiocb *req = data->req;
5531	struct io_ring_ctx *ctx = req->ctx;
5532	unsigned long flags;
5533
5534	spin_lock_irqsave(&ctx->completion_lock, flags);
5535	list_del_init(&req->timeout.list);
5536	atomic_set(&req->ctx->cq_timeouts,
5537		atomic_read(&req->ctx->cq_timeouts) + 1);
5538
5539	io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5540	io_commit_cqring(ctx);
5541	spin_unlock_irqrestore(&ctx->completion_lock, flags);
5542
5543	io_cqring_ev_posted(ctx);
5544	req_set_fail(req);
5545	io_put_req(req);
5546	return HRTIMER_NORESTART;
5547}
5548
5549static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5550					   __u64 user_data)
5551	__must_hold(&ctx->completion_lock)
5552{
5553	struct io_timeout_data *io;
5554	struct io_kiocb *req;
5555	bool found = false;
5556
5557	list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5558		found = user_data == req->user_data;
5559		if (found)
5560			break;
5561	}
5562	if (!found)
5563		return ERR_PTR(-ENOENT);
5564
5565	io = req->async_data;
5566	if (hrtimer_try_to_cancel(&io->timer) == -1)
5567		return ERR_PTR(-EALREADY);
5568	list_del_init(&req->timeout.list);
5569	return req;
5570}
5571
5572static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5573	__must_hold(&ctx->completion_lock)
5574{
5575	struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5576
5577	if (IS_ERR(req))
5578		return PTR_ERR(req);
5579
5580	req_set_fail(req);
5581	io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5582	io_put_req_deferred(req, 1);
5583	return 0;
5584}
5585
5586static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5587			     struct timespec64 *ts, enum hrtimer_mode mode)
5588	__must_hold(&ctx->completion_lock)
5589{
5590	struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5591	struct io_timeout_data *data;
5592
5593	if (IS_ERR(req))
5594		return PTR_ERR(req);
5595
5596	req->timeout.off = 0; /* noseq */
5597	data = req->async_data;
5598	list_add_tail(&req->timeout.list, &ctx->timeout_list);
5599	hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5600	data->timer.function = io_timeout_fn;
5601	hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5602	return 0;
5603}
5604
5605static int io_timeout_remove_prep(struct io_kiocb *req,
5606				  const struct io_uring_sqe *sqe)
5607{
5608	struct io_timeout_rem *tr = &req->timeout_rem;
5609
5610	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5611		return -EINVAL;
5612	if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5613		return -EINVAL;
5614	if (sqe->ioprio || sqe->buf_index || sqe->len)
5615		return -EINVAL;
5616
5617	tr->addr = READ_ONCE(sqe->addr);
5618	tr->flags = READ_ONCE(sqe->timeout_flags);
5619	if (tr->flags & IORING_TIMEOUT_UPDATE) {
5620		if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5621			return -EINVAL;
5622		if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5623			return -EFAULT;
5624	} else if (tr->flags) {
5625		/* timeout removal doesn't support flags */
5626		return -EINVAL;
5627	}
5628
5629	return 0;
5630}
5631
5632static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5633{
5634	return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5635					    : HRTIMER_MODE_REL;
5636}
5637
5638/*
5639 * Remove or update an existing timeout command
5640 */
5641static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5642{
5643	struct io_timeout_rem *tr = &req->timeout_rem;
5644	struct io_ring_ctx *ctx = req->ctx;
5645	int ret;
5646
5647	spin_lock_irq(&ctx->completion_lock);
5648	if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5649		ret = io_timeout_cancel(ctx, tr->addr);
5650	else
5651		ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5652					io_translate_timeout_mode(tr->flags));
5653
5654	io_cqring_fill_event(ctx, req->user_data, ret, 0);
5655	io_commit_cqring(ctx);
5656	spin_unlock_irq(&ctx->completion_lock);
5657	io_cqring_ev_posted(ctx);
5658	if (ret < 0)
5659		req_set_fail(req);
5660	io_put_req(req);
5661	return 0;
5662}
5663
5664static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5665			   bool is_timeout_link)
5666{
5667	struct io_timeout_data *data;
5668	unsigned flags;
5669	u32 off = READ_ONCE(sqe->off);
5670
5671	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5672		return -EINVAL;
5673	if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5674		return -EINVAL;
5675	if (off && is_timeout_link)
5676		return -EINVAL;
5677	flags = READ_ONCE(sqe->timeout_flags);
5678	if (flags & ~IORING_TIMEOUT_ABS)
5679		return -EINVAL;
5680
5681	req->timeout.off = off;
5682	if (unlikely(off && !req->ctx->off_timeout_used))
5683		req->ctx->off_timeout_used = true;
5684
5685	if (!req->async_data && io_alloc_async_data(req))
5686		return -ENOMEM;
5687
5688	data = req->async_data;
5689	data->req = req;
5690
5691	if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5692		return -EFAULT;
5693
5694	data->mode = io_translate_timeout_mode(flags);
5695	hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5696	if (is_timeout_link)
5697		io_req_track_inflight(req);
5698	return 0;
5699}
5700
5701static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5702{
5703	struct io_ring_ctx *ctx = req->ctx;
5704	struct io_timeout_data *data = req->async_data;
5705	struct list_head *entry;
5706	u32 tail, off = req->timeout.off;
5707
5708	spin_lock_irq(&ctx->completion_lock);
5709
5710	/*
5711	 * sqe->off holds how many events that need to occur for this
5712	 * timeout event to be satisfied. If it isn't set, then this is
5713	 * a pure timeout request, sequence isn't used.
5714	 */
5715	if (io_is_timeout_noseq(req)) {
5716		entry = ctx->timeout_list.prev;
5717		goto add;
5718	}
5719
5720	tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5721	req->timeout.target_seq = tail + off;
5722
5723	/* Update the last seq here in case io_flush_timeouts() hasn't.
5724	 * This is safe because ->completion_lock is held, and submissions
5725	 * and completions are never mixed in the same ->completion_lock section.
5726	 */
5727	ctx->