io_uring.c revision f56165e6
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_COMPL_BATCH			32
299#define IO_REQ_CACHE_SIZE		32
300#define IO_REQ_ALLOC_BATCH		8
301
302struct io_comp_state {
303	struct io_kiocb		*reqs[IO_COMPL_BATCH];
304	unsigned int		nr;
305	/* inline/task_work completion list, under ->uring_lock */
306	struct list_head	free_list;
307};
308
309struct io_submit_link {
310	struct io_kiocb		*head;
311	struct io_kiocb		*last;
312};
313
314struct io_submit_state {
315	struct blk_plug		plug;
316	struct io_submit_link	link;
317
318	/*
319	 * io_kiocb alloc cache
320	 */
321	void			*reqs[IO_REQ_CACHE_SIZE];
322	unsigned int		free_reqs;
323
324	bool			plug_started;
325
326	/*
327	 * Batch completion logic
328	 */
329	struct io_comp_state	comp;
330
331	/*
332	 * File reference cache
333	 */
334	struct file		*file;
335	unsigned int		fd;
336	unsigned int		file_refs;
337	unsigned int		ios_left;
338};
339
340struct io_ring_ctx {
341	/* const or read-mostly hot data */
342	struct {
343		struct percpu_ref	refs;
344
345		struct io_rings		*rings;
346		unsigned int		flags;
347		unsigned int		compat: 1;
348		unsigned int		drain_next: 1;
349		unsigned int		eventfd_async: 1;
350		unsigned int		restricted: 1;
351		unsigned int		off_timeout_used: 1;
352		unsigned int		drain_active: 1;
353	} ____cacheline_aligned_in_smp;
354
355	/* submission data */
356	struct {
357		struct mutex		uring_lock;
358
359		/*
360		 * Ring buffer of indices into array of io_uring_sqe, which is
361		 * mmapped by the application using the IORING_OFF_SQES offset.
362		 *
363		 * This indirection could e.g. be used to assign fixed
364		 * io_uring_sqe entries to operations and only submit them to
365		 * the queue when needed.
366		 *
367		 * The kernel modifies neither the indices array nor the entries
368		 * array.
369		 */
370		u32			*sq_array;
371		struct io_uring_sqe	*sq_sqes;
372		unsigned		cached_sq_head;
373		unsigned		sq_entries;
374		struct list_head	defer_list;
375
376		/*
377		 * Fixed resources fast path, should be accessed only under
378		 * uring_lock, and updated through io_uring_register(2)
379		 */
380		struct io_rsrc_node	*rsrc_node;
381		struct io_file_table	file_table;
382		unsigned		nr_user_files;
383		unsigned		nr_user_bufs;
384		struct io_mapped_ubuf	**user_bufs;
385
386		struct io_submit_state	submit_state;
387		struct list_head	timeout_list;
388		struct list_head	cq_overflow_list;
389		struct xarray		io_buffers;
390		struct xarray		personalities;
391		u32			pers_next;
392		unsigned		sq_thread_idle;
393	} ____cacheline_aligned_in_smp;
394
395	/* IRQ completion list, under ->completion_lock */
396	struct list_head	locked_free_list;
397	unsigned int		locked_free_nr;
398
399	const struct cred	*sq_creds;	/* cred used for __io_sq_thread() */
400	struct io_sq_data	*sq_data;	/* if using sq thread polling */
401
402	struct wait_queue_head	sqo_sq_wait;
403	struct list_head	sqd_list;
404
405	unsigned long		check_cq_overflow;
406
407	struct {
408		unsigned		cached_cq_tail;
409		unsigned		cq_entries;
410		struct eventfd_ctx	*cq_ev_fd;
411		struct wait_queue_head	poll_wait;
412		struct wait_queue_head	cq_wait;
413		unsigned		cq_extra;
414		atomic_t		cq_timeouts;
415		struct fasync_struct	*cq_fasync;
416		unsigned		cq_last_tm_flush;
417	} ____cacheline_aligned_in_smp;
418
419	struct {
420		spinlock_t		completion_lock;
421
422		/*
423		 * ->iopoll_list is protected by the ctx->uring_lock for
424		 * io_uring instances that don't use IORING_SETUP_SQPOLL.
425		 * For SQPOLL, only the single threaded io_sq_thread() will
426		 * manipulate the list, hence no extra locking is needed there.
427		 */
428		struct list_head	iopoll_list;
429		struct hlist_head	*cancel_hash;
430		unsigned		cancel_hash_bits;
431		bool			poll_multi_queue;
432	} ____cacheline_aligned_in_smp;
433
434	struct io_restriction		restrictions;
435
436	/* slow path rsrc auxilary data, used by update/register */
437	struct {
438		struct io_rsrc_node		*rsrc_backup_node;
439		struct io_mapped_ubuf		*dummy_ubuf;
440		struct io_rsrc_data		*file_data;
441		struct io_rsrc_data		*buf_data;
442
443		struct delayed_work		rsrc_put_work;
444		struct llist_head		rsrc_put_llist;
445		struct list_head		rsrc_ref_list;
446		spinlock_t			rsrc_ref_lock;
447	};
448
449	/* Keep this last, we don't need it for the fast path */
450	struct {
451		#if defined(CONFIG_UNIX)
452			struct socket		*ring_sock;
453		#endif
454		/* hashed buffered write serialization */
455		struct io_wq_hash		*hash_map;
456
457		/* Only used for accounting purposes */
458		struct user_struct		*user;
459		struct mm_struct		*mm_account;
460
461		/* ctx exit and cancelation */
462		struct llist_head		fallback_llist;
463		struct delayed_work		fallback_work;
464		struct work_struct		exit_work;
465		struct list_head		tctx_list;
466		struct completion		ref_comp;
467	};
468};
469
470struct io_uring_task {
471	/* submission side */
472	int			cached_refs;
473	struct xarray		xa;
474	struct wait_queue_head	wait;
475	const struct io_ring_ctx *last;
476	struct io_wq		*io_wq;
477	struct percpu_counter	inflight;
478	atomic_t		inflight_tracked;
479	atomic_t		in_idle;
480
481	spinlock_t		task_lock;
482	struct io_wq_work_list	task_list;
483	unsigned long		task_state;
484	struct callback_head	task_work;
485};
486
487/*
488 * First field must be the file pointer in all the
489 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
490 */
491struct io_poll_iocb {
492	struct file			*file;
493	struct wait_queue_head		*head;
494	__poll_t			events;
495	bool				done;
496	bool				canceled;
497	struct wait_queue_entry		wait;
498};
499
500struct io_poll_update {
501	struct file			*file;
502	u64				old_user_data;
503	u64				new_user_data;
504	__poll_t			events;
505	bool				update_events;
506	bool				update_user_data;
507};
508
509struct io_close {
510	struct file			*file;
511	int				fd;
512};
513
514struct io_timeout_data {
515	struct io_kiocb			*req;
516	struct hrtimer			timer;
517	struct timespec64		ts;
518	enum hrtimer_mode		mode;
519};
520
521struct io_accept {
522	struct file			*file;
523	struct sockaddr __user		*addr;
524	int __user			*addr_len;
525	int				flags;
526	unsigned long			nofile;
527};
528
529struct io_sync {
530	struct file			*file;
531	loff_t				len;
532	loff_t				off;
533	int				flags;
534	int				mode;
535};
536
537struct io_cancel {
538	struct file			*file;
539	u64				addr;
540};
541
542struct io_timeout {
543	struct file			*file;
544	u32				off;
545	u32				target_seq;
546	struct list_head		list;
547	/* head of the link, used by linked timeouts only */
548	struct io_kiocb			*head;
549};
550
551struct io_timeout_rem {
552	struct file			*file;
553	u64				addr;
554
555	/* timeout update */
556	struct timespec64		ts;
557	u32				flags;
558};
559
560struct io_rw {
561	/* NOTE: kiocb has the file as the first member, so don't do it here */
562	struct kiocb			kiocb;
563	u64				addr;
564	u64				len;
565};
566
567struct io_connect {
568	struct file			*file;
569	struct sockaddr __user		*addr;
570	int				addr_len;
571};
572
573struct io_sr_msg {
574	struct file			*file;
575	union {
576		struct compat_msghdr __user	*umsg_compat;
577		struct user_msghdr __user	*umsg;
578		void __user			*buf;
579	};
580	int				msg_flags;
581	int				bgid;
582	size_t				len;
583	struct io_buffer		*kbuf;
584};
585
586struct io_open {
587	struct file			*file;
588	int				dfd;
589	struct filename			*filename;
590	struct open_how			how;
591	unsigned long			nofile;
592};
593
594struct io_rsrc_update {
595	struct file			*file;
596	u64				arg;
597	u32				nr_args;
598	u32				offset;
599};
600
601struct io_fadvise {
602	struct file			*file;
603	u64				offset;
604	u32				len;
605	u32				advice;
606};
607
608struct io_madvise {
609	struct file			*file;
610	u64				addr;
611	u32				len;
612	u32				advice;
613};
614
615struct io_epoll {
616	struct file			*file;
617	int				epfd;
618	int				op;
619	int				fd;
620	struct epoll_event		event;
621};
622
623struct io_splice {
624	struct file			*file_out;
625	struct file			*file_in;
626	loff_t				off_out;
627	loff_t				off_in;
628	u64				len;
629	unsigned int			flags;
630};
631
632struct io_provide_buf {
633	struct file			*file;
634	__u64				addr;
635	__u32				len;
636	__u32				bgid;
637	__u16				nbufs;
638	__u16				bid;
639};
640
641struct io_statx {
642	struct file			*file;
643	int				dfd;
644	unsigned int			mask;
645	unsigned int			flags;
646	const char __user		*filename;
647	struct statx __user		*buffer;
648};
649
650struct io_shutdown {
651	struct file			*file;
652	int				how;
653};
654
655struct io_rename {
656	struct file			*file;
657	int				old_dfd;
658	int				new_dfd;
659	struct filename			*oldpath;
660	struct filename			*newpath;
661	int				flags;
662};
663
664struct io_unlink {
665	struct file			*file;
666	int				dfd;
667	int				flags;
668	struct filename			*filename;
669};
670
671struct io_completion {
672	struct file			*file;
673	struct list_head		list;
674	u32				cflags;
675};
676
677struct io_async_connect {
678	struct sockaddr_storage		address;
679};
680
681struct io_async_msghdr {
682	struct iovec			fast_iov[UIO_FASTIOV];
683	/* points to an allocated iov, if NULL we use fast_iov instead */
684	struct iovec			*free_iov;
685	struct sockaddr __user		*uaddr;
686	struct msghdr			msg;
687	struct sockaddr_storage		addr;
688};
689
690struct io_async_rw {
691	struct iovec			fast_iov[UIO_FASTIOV];
692	const struct iovec		*free_iovec;
693	struct iov_iter			iter;
694	size_t				bytes_done;
695	struct wait_page_queue		wpq;
696};
697
698enum {
699	REQ_F_FIXED_FILE_BIT	= IOSQE_FIXED_FILE_BIT,
700	REQ_F_IO_DRAIN_BIT	= IOSQE_IO_DRAIN_BIT,
701	REQ_F_LINK_BIT		= IOSQE_IO_LINK_BIT,
702	REQ_F_HARDLINK_BIT	= IOSQE_IO_HARDLINK_BIT,
703	REQ_F_FORCE_ASYNC_BIT	= IOSQE_ASYNC_BIT,
704	REQ_F_BUFFER_SELECT_BIT	= IOSQE_BUFFER_SELECT_BIT,
705
706	/* first byte is taken by user flags, shift it to not overlap */
707	REQ_F_FAIL_BIT		= 8,
708	REQ_F_INFLIGHT_BIT,
709	REQ_F_CUR_POS_BIT,
710	REQ_F_NOWAIT_BIT,
711	REQ_F_LINK_TIMEOUT_BIT,
712	REQ_F_NEED_CLEANUP_BIT,
713	REQ_F_POLLED_BIT,
714	REQ_F_BUFFER_SELECTED_BIT,
715	REQ_F_LTIMEOUT_ACTIVE_BIT,
716	REQ_F_COMPLETE_INLINE_BIT,
717	REQ_F_REISSUE_BIT,
718	REQ_F_DONT_REISSUE_BIT,
719	REQ_F_CREDS_BIT,
720	/* keep async read/write and isreg together and in order */
721	REQ_F_NOWAIT_READ_BIT,
722	REQ_F_NOWAIT_WRITE_BIT,
723	REQ_F_ISREG_BIT,
724
725	/* not a real bit, just to check we're not overflowing the space */
726	__REQ_F_LAST_BIT,
727};
728
729enum {
730	/* ctx owns file */
731	REQ_F_FIXED_FILE	= BIT(REQ_F_FIXED_FILE_BIT),
732	/* drain existing IO first */
733	REQ_F_IO_DRAIN		= BIT(REQ_F_IO_DRAIN_BIT),
734	/* linked sqes */
735	REQ_F_LINK		= BIT(REQ_F_LINK_BIT),
736	/* doesn't sever on completion < 0 */
737	REQ_F_HARDLINK		= BIT(REQ_F_HARDLINK_BIT),
738	/* IOSQE_ASYNC */
739	REQ_F_FORCE_ASYNC	= BIT(REQ_F_FORCE_ASYNC_BIT),
740	/* IOSQE_BUFFER_SELECT */
741	REQ_F_BUFFER_SELECT	= BIT(REQ_F_BUFFER_SELECT_BIT),
742
743	/* fail rest of links */
744	REQ_F_FAIL		= BIT(REQ_F_FAIL_BIT),
745	/* on inflight list, should be cancelled and waited on exit reliably */
746	REQ_F_INFLIGHT		= BIT(REQ_F_INFLIGHT_BIT),
747	/* read/write uses file position */
748	REQ_F_CUR_POS		= BIT(REQ_F_CUR_POS_BIT),
749	/* must not punt to workers */
750	REQ_F_NOWAIT		= BIT(REQ_F_NOWAIT_BIT),
751	/* has or had linked timeout */
752	REQ_F_LINK_TIMEOUT	= BIT(REQ_F_LINK_TIMEOUT_BIT),
753	/* needs cleanup */
754	REQ_F_NEED_CLEANUP	= BIT(REQ_F_NEED_CLEANUP_BIT),
755	/* already went through poll handler */
756	REQ_F_POLLED		= BIT(REQ_F_POLLED_BIT),
757	/* buffer already selected */
758	REQ_F_BUFFER_SELECTED	= BIT(REQ_F_BUFFER_SELECTED_BIT),
759	/* linked timeout is active, i.e. prepared by link's head */
760	REQ_F_LTIMEOUT_ACTIVE	= BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
761	/* completion is deferred through io_comp_state */
762	REQ_F_COMPLETE_INLINE	= BIT(REQ_F_COMPLETE_INLINE_BIT),
763	/* caller should reissue async */
764	REQ_F_REISSUE		= BIT(REQ_F_REISSUE_BIT),
765	/* don't attempt request reissue, see io_rw_reissue() */
766	REQ_F_DONT_REISSUE	= BIT(REQ_F_DONT_REISSUE_BIT),
767	/* supports async reads */
768	REQ_F_NOWAIT_READ	= BIT(REQ_F_NOWAIT_READ_BIT),
769	/* supports async writes */
770	REQ_F_NOWAIT_WRITE	= BIT(REQ_F_NOWAIT_WRITE_BIT),
771	/* regular file */
772	REQ_F_ISREG		= BIT(REQ_F_ISREG_BIT),
773	/* has creds assigned */
774	REQ_F_CREDS		= BIT(REQ_F_CREDS_BIT),
775};
776
777struct async_poll {
778	struct io_poll_iocb	poll;
779	struct io_poll_iocb	*double_poll;
780};
781
782typedef void (*io_req_tw_func_t)(struct io_kiocb *req);
783
784struct io_task_work {
785	union {
786		struct io_wq_work_node	node;
787		struct llist_node	fallback_node;
788	};
789	io_req_tw_func_t		func;
790};
791
792enum {
793	IORING_RSRC_FILE		= 0,
794	IORING_RSRC_BUFFER		= 1,
795};
796
797/*
798 * NOTE! Each of the iocb union members has the file pointer
799 * as the first entry in their struct definition. So you can
800 * access the file pointer through any of the sub-structs,
801 * or directly as just 'ki_filp' in this struct.
802 */
803struct io_kiocb {
804	union {
805		struct file		*file;
806		struct io_rw		rw;
807		struct io_poll_iocb	poll;
808		struct io_poll_update	poll_update;
809		struct io_accept	accept;
810		struct io_sync		sync;
811		struct io_cancel	cancel;
812		struct io_timeout	timeout;
813		struct io_timeout_rem	timeout_rem;
814		struct io_connect	connect;
815		struct io_sr_msg	sr_msg;
816		struct io_open		open;
817		struct io_close		close;
818		struct io_rsrc_update	rsrc_update;
819		struct io_fadvise	fadvise;
820		struct io_madvise	madvise;
821		struct io_epoll		epoll;
822		struct io_splice	splice;
823		struct io_provide_buf	pbuf;
824		struct io_statx		statx;
825		struct io_shutdown	shutdown;
826		struct io_rename	rename;
827		struct io_unlink	unlink;
828		/* use only after cleaning per-op data, see io_clean_op() */
829		struct io_completion	compl;
830	};
831
832	/* opcode allocated if it needs to store data for async defer */
833	void				*async_data;
834	u8				opcode;
835	/* polled IO has completed */
836	u8				iopoll_completed;
837
838	u16				buf_index;
839	u32				result;
840
841	struct io_ring_ctx		*ctx;
842	unsigned int			flags;
843	atomic_t			refs;
844	struct task_struct		*task;
845	u64				user_data;
846
847	struct io_kiocb			*link;
848	struct percpu_ref		*fixed_rsrc_refs;
849
850	/* used with ctx->iopoll_list with reads/writes */
851	struct list_head		inflight_entry;
852	struct io_task_work		io_task_work;
853	/* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
854	struct hlist_node		hash_node;
855	struct async_poll		*apoll;
856	struct io_wq_work		work;
857	const struct cred		*creds;
858
859	/* store used ubuf, so we can prevent reloading */
860	struct io_mapped_ubuf		*imu;
861};
862
863struct io_tctx_node {
864	struct list_head	ctx_node;
865	struct task_struct	*task;
866	struct io_ring_ctx	*ctx;
867};
868
869struct io_defer_entry {
870	struct list_head	list;
871	struct io_kiocb		*req;
872	u32			seq;
873};
874
875struct io_op_def {
876	/* needs req->file assigned */
877	unsigned		needs_file : 1;
878	/* hash wq insertion if file is a regular file */
879	unsigned		hash_reg_file : 1;
880	/* unbound wq insertion if file is a non-regular file */
881	unsigned		unbound_nonreg_file : 1;
882	/* opcode is not supported by this kernel */
883	unsigned		not_supported : 1;
884	/* set if opcode supports polled "wait" */
885	unsigned		pollin : 1;
886	unsigned		pollout : 1;
887	/* op supports buffer selection */
888	unsigned		buffer_select : 1;
889	/* do prep async if is going to be punted */
890	unsigned		needs_async_setup : 1;
891	/* should block plug */
892	unsigned		plug : 1;
893	/* size of async data needed, if any */
894	unsigned short		async_size;
895};
896
897static const struct io_op_def io_op_defs[] = {
898	[IORING_OP_NOP] = {},
899	[IORING_OP_READV] = {
900		.needs_file		= 1,
901		.unbound_nonreg_file	= 1,
902		.pollin			= 1,
903		.buffer_select		= 1,
904		.needs_async_setup	= 1,
905		.plug			= 1,
906		.async_size		= sizeof(struct io_async_rw),
907	},
908	[IORING_OP_WRITEV] = {
909		.needs_file		= 1,
910		.hash_reg_file		= 1,
911		.unbound_nonreg_file	= 1,
912		.pollout		= 1,
913		.needs_async_setup	= 1,
914		.plug			= 1,
915		.async_size		= sizeof(struct io_async_rw),
916	},
917	[IORING_OP_FSYNC] = {
918		.needs_file		= 1,
919	},
920	[IORING_OP_READ_FIXED] = {
921		.needs_file		= 1,
922		.unbound_nonreg_file	= 1,
923		.pollin			= 1,
924		.plug			= 1,
925		.async_size		= sizeof(struct io_async_rw),
926	},
927	[IORING_OP_WRITE_FIXED] = {
928		.needs_file		= 1,
929		.hash_reg_file		= 1,
930		.unbound_nonreg_file	= 1,
931		.pollout		= 1,
932		.plug			= 1,
933		.async_size		= sizeof(struct io_async_rw),
934	},
935	[IORING_OP_POLL_ADD] = {
936		.needs_file		= 1,
937		.unbound_nonreg_file	= 1,
938	},
939	[IORING_OP_POLL_REMOVE] = {},
940	[IORING_OP_SYNC_FILE_RANGE] = {
941		.needs_file		= 1,
942	},
943	[IORING_OP_SENDMSG] = {
944		.needs_file		= 1,
945		.unbound_nonreg_file	= 1,
946		.pollout		= 1,
947		.needs_async_setup	= 1,
948		.async_size		= sizeof(struct io_async_msghdr),
949	},
950	[IORING_OP_RECVMSG] = {
951		.needs_file		= 1,
952		.unbound_nonreg_file	= 1,
953		.pollin			= 1,
954		.buffer_select		= 1,
955		.needs_async_setup	= 1,
956		.async_size		= sizeof(struct io_async_msghdr),
957	},
958	[IORING_OP_TIMEOUT] = {
959		.async_size		= sizeof(struct io_timeout_data),
960	},
961	[IORING_OP_TIMEOUT_REMOVE] = {
962		/* used by timeout updates' prep() */
963	},
964	[IORING_OP_ACCEPT] = {
965		.needs_file		= 1,
966		.unbound_nonreg_file	= 1,
967		.pollin			= 1,
968	},
969	[IORING_OP_ASYNC_CANCEL] = {},
970	[IORING_OP_LINK_TIMEOUT] = {
971		.async_size		= sizeof(struct io_timeout_data),
972	},
973	[IORING_OP_CONNECT] = {
974		.needs_file		= 1,
975		.unbound_nonreg_file	= 1,
976		.pollout		= 1,
977		.needs_async_setup	= 1,
978		.async_size		= sizeof(struct io_async_connect),
979	},
980	[IORING_OP_FALLOCATE] = {
981		.needs_file		= 1,
982	},
983	[IORING_OP_OPENAT] = {},
984	[IORING_OP_CLOSE] = {},
985	[IORING_OP_FILES_UPDATE] = {},
986	[IORING_OP_STATX] = {},
987	[IORING_OP_READ] = {
988		.needs_file		= 1,
989		.unbound_nonreg_file	= 1,
990		.pollin			= 1,
991		.buffer_select		= 1,
992		.plug			= 1,
993		.async_size		= sizeof(struct io_async_rw),
994	},
995	[IORING_OP_WRITE] = {
996		.needs_file		= 1,
997		.unbound_nonreg_file	= 1,
998		.pollout		= 1,
999		.plug			= 1,
1000		.async_size		= sizeof(struct io_async_rw),
1001	},
1002	[IORING_OP_FADVISE] = {
1003		.needs_file		= 1,
1004	},
1005	[IORING_OP_MADVISE] = {},
1006	[IORING_OP_SEND] = {
1007		.needs_file		= 1,
1008		.unbound_nonreg_file	= 1,
1009		.pollout		= 1,
1010	},
1011	[IORING_OP_RECV] = {
1012		.needs_file		= 1,
1013		.unbound_nonreg_file	= 1,
1014		.pollin			= 1,
1015		.buffer_select		= 1,
1016	},
1017	[IORING_OP_OPENAT2] = {
1018	},
1019	[IORING_OP_EPOLL_CTL] = {
1020		.unbound_nonreg_file	= 1,
1021	},
1022	[IORING_OP_SPLICE] = {
1023		.needs_file		= 1,
1024		.hash_reg_file		= 1,
1025		.unbound_nonreg_file	= 1,
1026	},
1027	[IORING_OP_PROVIDE_BUFFERS] = {},
1028	[IORING_OP_REMOVE_BUFFERS] = {},
1029	[IORING_OP_TEE] = {
1030		.needs_file		= 1,
1031		.hash_reg_file		= 1,
1032		.unbound_nonreg_file	= 1,
1033	},
1034	[IORING_OP_SHUTDOWN] = {
1035		.needs_file		= 1,
1036	},
1037	[IORING_OP_RENAMEAT] = {},
1038	[IORING_OP_UNLINKAT] = {},
1039};
1040
1041static bool io_disarm_next(struct io_kiocb *req);
1042static void io_uring_del_tctx_node(unsigned long index);
1043static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1044					 struct task_struct *task,
1045					 bool cancel_all);
1046static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1047
1048static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1049				 long res, unsigned int cflags);
1050static void io_put_req(struct io_kiocb *req);
1051static void io_put_req_deferred(struct io_kiocb *req, int nr);
1052static void io_dismantle_req(struct io_kiocb *req);
1053static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1054static void io_queue_linked_timeout(struct io_kiocb *req);
1055static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1056				     struct io_uring_rsrc_update2 *up,
1057				     unsigned nr_args);
1058static void io_clean_op(struct io_kiocb *req);
1059static struct file *io_file_get(struct io_ring_ctx *ctx,
1060				struct io_submit_state *state,
1061				struct io_kiocb *req, int fd, bool fixed);
1062static void __io_queue_sqe(struct io_kiocb *req);
1063static void io_rsrc_put_work(struct work_struct *work);
1064
1065static void io_req_task_queue(struct io_kiocb *req);
1066static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1067static bool io_poll_remove_waitqs(struct io_kiocb *req);
1068static int io_req_prep_async(struct io_kiocb *req);
1069
1070static struct kmem_cache *req_cachep;
1071
1072static const struct file_operations io_uring_fops;
1073
1074struct sock *io_uring_get_socket(struct file *file)
1075{
1076#if defined(CONFIG_UNIX)
1077	if (file->f_op == &io_uring_fops) {
1078		struct io_ring_ctx *ctx = file->private_data;
1079
1080		return ctx->ring_sock->sk;
1081	}
1082#endif
1083	return NULL;
1084}
1085EXPORT_SYMBOL(io_uring_get_socket);
1086
1087#define io_for_each_link(pos, head) \
1088	for (pos = (head); pos; pos = pos->link)
1089
1090static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1091{
1092	struct io_ring_ctx *ctx = req->ctx;
1093
1094	if (!req->fixed_rsrc_refs) {
1095		req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1096		percpu_ref_get(req->fixed_rsrc_refs);
1097	}
1098}
1099
1100static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1101{
1102	bool got = percpu_ref_tryget(ref);
1103
1104	/* already at zero, wait for ->release() */
1105	if (!got)
1106		wait_for_completion(compl);
1107	percpu_ref_resurrect(ref);
1108	if (got)
1109		percpu_ref_put(ref);
1110}
1111
1112static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1113			  bool cancel_all)
1114{
1115	struct io_kiocb *req;
1116
1117	if (task && head->task != task)
1118		return false;
1119	if (cancel_all)
1120		return true;
1121
1122	io_for_each_link(req, head) {
1123		if (req->flags & REQ_F_INFLIGHT)
1124			return true;
1125	}
1126	return false;
1127}
1128
1129static inline void req_set_fail(struct io_kiocb *req)
1130{
1131	req->flags |= REQ_F_FAIL;
1132}
1133
1134static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1135{
1136	struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1137
1138	complete(&ctx->ref_comp);
1139}
1140
1141static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1142{
1143	return !req->timeout.off;
1144}
1145
1146static void io_fallback_req_func(struct work_struct *work)
1147{
1148	struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
1149						fallback_work.work);
1150	struct llist_node *node = llist_del_all(&ctx->fallback_llist);
1151	struct io_kiocb *req, *tmp;
1152
1153	percpu_ref_get(&ctx->refs);
1154	llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
1155		req->io_task_work.func(req);
1156	percpu_ref_put(&ctx->refs);
1157}
1158
1159static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1160{
1161	struct io_ring_ctx *ctx;
1162	int hash_bits;
1163
1164	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1165	if (!ctx)
1166		return NULL;
1167
1168	/*
1169	 * Use 5 bits less than the max cq entries, that should give us around
1170	 * 32 entries per hash list if totally full and uniformly spread.
1171	 */
1172	hash_bits = ilog2(p->cq_entries);
1173	hash_bits -= 5;
1174	if (hash_bits <= 0)
1175		hash_bits = 1;
1176	ctx->cancel_hash_bits = hash_bits;
1177	ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1178					GFP_KERNEL);
1179	if (!ctx->cancel_hash)
1180		goto err;
1181	__hash_init(ctx->cancel_hash, 1U << hash_bits);
1182
1183	ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1184	if (!ctx->dummy_ubuf)
1185		goto err;
1186	/* set invalid range, so io_import_fixed() fails meeting it */
1187	ctx->dummy_ubuf->ubuf = -1UL;
1188
1189	if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1190			    PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1191		goto err;
1192
1193	ctx->flags = p->flags;
1194	init_waitqueue_head(&ctx->sqo_sq_wait);
1195	INIT_LIST_HEAD(&ctx->sqd_list);
1196	init_waitqueue_head(&ctx->poll_wait);
1197	INIT_LIST_HEAD(&ctx->cq_overflow_list);
1198	init_completion(&ctx->ref_comp);
1199	xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1200	xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1201	mutex_init(&ctx->uring_lock);
1202	init_waitqueue_head(&ctx->cq_wait);
1203	spin_lock_init(&ctx->completion_lock);
1204	INIT_LIST_HEAD(&ctx->iopoll_list);
1205	INIT_LIST_HEAD(&ctx->defer_list);
1206	INIT_LIST_HEAD(&ctx->timeout_list);
1207	spin_lock_init(&ctx->rsrc_ref_lock);
1208	INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1209	INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1210	init_llist_head(&ctx->rsrc_put_llist);
1211	INIT_LIST_HEAD(&ctx->tctx_list);
1212	INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1213	INIT_LIST_HEAD(&ctx->locked_free_list);
1214	INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1215	return ctx;
1216err:
1217	kfree(ctx->dummy_ubuf);
1218	kfree(ctx->cancel_hash);
1219	kfree(ctx);
1220	return NULL;
1221}
1222
1223static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1224{
1225	struct io_rings *r = ctx->rings;
1226
1227	WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1228	ctx->cq_extra--;
1229}
1230
1231static bool req_need_defer(struct io_kiocb *req, u32 seq)
1232{
1233	if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1234		struct io_ring_ctx *ctx = req->ctx;
1235
1236		return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1237	}
1238
1239	return false;
1240}
1241
1242#define FFS_ASYNC_READ		0x1UL
1243#define FFS_ASYNC_WRITE		0x2UL
1244#ifdef CONFIG_64BIT
1245#define FFS_ISREG		0x4UL
1246#else
1247#define FFS_ISREG		0x0UL
1248#endif
1249#define FFS_MASK		~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1250
1251static inline bool io_req_ffs_set(struct io_kiocb *req)
1252{
1253	return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE);
1254}
1255
1256static void io_req_track_inflight(struct io_kiocb *req)
1257{
1258	if (!(req->flags & REQ_F_INFLIGHT)) {
1259		req->flags |= REQ_F_INFLIGHT;
1260		atomic_inc(&current->io_uring->inflight_tracked);
1261	}
1262}
1263
1264static void io_prep_async_work(struct io_kiocb *req)
1265{
1266	const struct io_op_def *def = &io_op_defs[req->opcode];
1267	struct io_ring_ctx *ctx = req->ctx;
1268
1269	if (!(req->flags & REQ_F_CREDS)) {
1270		req->flags |= REQ_F_CREDS;
1271		req->creds = get_current_cred();
1272	}
1273
1274	req->work.list.next = NULL;
1275	req->work.flags = 0;
1276	if (req->flags & REQ_F_FORCE_ASYNC)
1277		req->work.flags |= IO_WQ_WORK_CONCURRENT;
1278
1279	if (req->flags & REQ_F_ISREG) {
1280		if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1281			io_wq_hash_work(&req->work, file_inode(req->file));
1282	} else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1283		if (def->unbound_nonreg_file)
1284			req->work.flags |= IO_WQ_WORK_UNBOUND;
1285	}
1286
1287	switch (req->opcode) {
1288	case IORING_OP_SPLICE:
1289	case IORING_OP_TEE:
1290		if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1291			req->work.flags |= IO_WQ_WORK_UNBOUND;
1292		break;
1293	}
1294}
1295
1296static void io_prep_async_link(struct io_kiocb *req)
1297{
1298	struct io_kiocb *cur;
1299
1300	if (req->flags & REQ_F_LINK_TIMEOUT) {
1301		struct io_ring_ctx *ctx = req->ctx;
1302
1303		spin_lock_irq(&ctx->completion_lock);
1304		io_for_each_link(cur, req)
1305			io_prep_async_work(cur);
1306		spin_unlock_irq(&ctx->completion_lock);
1307	} else {
1308		io_for_each_link(cur, req)
1309			io_prep_async_work(cur);
1310	}
1311}
1312
1313static void io_queue_async_work(struct io_kiocb *req)
1314{
1315	struct io_ring_ctx *ctx = req->ctx;
1316	struct io_kiocb *link = io_prep_linked_timeout(req);
1317	struct io_uring_task *tctx = req->task->io_uring;
1318
1319	BUG_ON(!tctx);
1320	BUG_ON(!tctx->io_wq);
1321
1322	/* init ->work of the whole link before punting */
1323	io_prep_async_link(req);
1324
1325	/*
1326	 * Not expected to happen, but if we do have a bug where this _can_
1327	 * happen, catch it here and ensure the request is marked as
1328	 * canceled. That will make io-wq go through the usual work cancel
1329	 * procedure rather than attempt to run this request (or create a new
1330	 * worker for it).
1331	 */
1332	if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1333		req->work.flags |= IO_WQ_WORK_CANCEL;
1334
1335	trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1336					&req->work, req->flags);
1337	io_wq_enqueue(tctx->io_wq, &req->work);
1338	if (link)
1339		io_queue_linked_timeout(link);
1340}
1341
1342static void io_kill_timeout(struct io_kiocb *req, int status)
1343	__must_hold(&req->ctx->completion_lock)
1344{
1345	struct io_timeout_data *io = req->async_data;
1346
1347	if (hrtimer_try_to_cancel(&io->timer) != -1) {
1348		atomic_set(&req->ctx->cq_timeouts,
1349			atomic_read(&req->ctx->cq_timeouts) + 1);
1350		list_del_init(&req->timeout.list);
1351		io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1352		io_put_req_deferred(req, 1);
1353	}
1354}
1355
1356static void io_queue_deferred(struct io_ring_ctx *ctx)
1357{
1358	while (!list_empty(&ctx->defer_list)) {
1359		struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1360						struct io_defer_entry, list);
1361
1362		if (req_need_defer(de->req, de->seq))
1363			break;
1364		list_del_init(&de->list);
1365		io_req_task_queue(de->req);
1366		kfree(de);
1367	}
1368}
1369
1370static void io_flush_timeouts(struct io_ring_ctx *ctx)
1371{
1372	u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1373
1374	while (!list_empty(&ctx->timeout_list)) {
1375		u32 events_needed, events_got;
1376		struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1377						struct io_kiocb, timeout.list);
1378
1379		if (io_is_timeout_noseq(req))
1380			break;
1381
1382		/*
1383		 * Since seq can easily wrap around over time, subtract
1384		 * the last seq at which timeouts were flushed before comparing.
1385		 * Assuming not more than 2^31-1 events have happened since,
1386		 * these subtractions won't have wrapped, so we can check if
1387		 * target is in [last_seq, current_seq] by comparing the two.
1388		 */
1389		events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1390		events_got = seq - ctx->cq_last_tm_flush;
1391		if (events_got < events_needed)
1392			break;
1393
1394		list_del_init(&req->timeout.list);
1395		io_kill_timeout(req, 0);
1396	}
1397	ctx->cq_last_tm_flush = seq;
1398}
1399
1400static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1401{
1402	if (ctx->off_timeout_used)
1403		io_flush_timeouts(ctx);
1404	if (ctx->drain_active)
1405		io_queue_deferred(ctx);
1406}
1407
1408static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1409{
1410	if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1411		__io_commit_cqring_flush(ctx);
1412	/* order cqe stores with ring update */
1413	smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1414}
1415
1416static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1417{
1418	struct io_rings *r = ctx->rings;
1419
1420	return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1421}
1422
1423static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1424{
1425	return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1426}
1427
1428static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1429{
1430	struct io_rings *rings = ctx->rings;
1431	unsigned tail, mask = ctx->cq_entries - 1;
1432
1433	/*
1434	 * writes to the cq entry need to come after reading head; the
1435	 * control dependency is enough as we're using WRITE_ONCE to
1436	 * fill the cq entry
1437	 */
1438	if (__io_cqring_events(ctx) == ctx->cq_entries)
1439		return NULL;
1440
1441	tail = ctx->cached_cq_tail++;
1442	return &rings->cqes[tail & mask];
1443}
1444
1445static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1446{
1447	if (likely(!ctx->cq_ev_fd))
1448		return false;
1449	if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1450		return false;
1451	return !ctx->eventfd_async || io_wq_current_is_worker();
1452}
1453
1454static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1455{
1456	/*
1457	 * wake_up_all() may seem excessive, but io_wake_function() and
1458	 * io_should_wake() handle the termination of the loop and only
1459	 * wake as many waiters as we need to.
1460	 */
1461	if (wq_has_sleeper(&ctx->cq_wait))
1462		wake_up_all(&ctx->cq_wait);
1463	if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1464		wake_up(&ctx->sq_data->wait);
1465	if (io_should_trigger_evfd(ctx))
1466		eventfd_signal(ctx->cq_ev_fd, 1);
1467	if (waitqueue_active(&ctx->poll_wait)) {
1468		wake_up_interruptible(&ctx->poll_wait);
1469		kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1470	}
1471}
1472
1473static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1474{
1475	if (ctx->flags & IORING_SETUP_SQPOLL) {
1476		if (wq_has_sleeper(&ctx->cq_wait))
1477			wake_up_all(&ctx->cq_wait);
1478	}
1479	if (io_should_trigger_evfd(ctx))
1480		eventfd_signal(ctx->cq_ev_fd, 1);
1481	if (waitqueue_active(&ctx->poll_wait)) {
1482		wake_up_interruptible(&ctx->poll_wait);
1483		kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1484	}
1485}
1486
1487/* Returns true if there are no backlogged entries after the flush */
1488static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1489{
1490	unsigned long flags;
1491	bool all_flushed, posted;
1492
1493	if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1494		return false;
1495
1496	posted = false;
1497	spin_lock_irqsave(&ctx->completion_lock, flags);
1498	while (!list_empty(&ctx->cq_overflow_list)) {
1499		struct io_uring_cqe *cqe = io_get_cqe(ctx);
1500		struct io_overflow_cqe *ocqe;
1501
1502		if (!cqe && !force)
1503			break;
1504		ocqe = list_first_entry(&ctx->cq_overflow_list,
1505					struct io_overflow_cqe, list);
1506		if (cqe)
1507			memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1508		else
1509			io_account_cq_overflow(ctx);
1510
1511		posted = true;
1512		list_del(&ocqe->list);
1513		kfree(ocqe);
1514	}
1515
1516	all_flushed = list_empty(&ctx->cq_overflow_list);
1517	if (all_flushed) {
1518		clear_bit(0, &ctx->check_cq_overflow);
1519		WRITE_ONCE(ctx->rings->sq_flags,
1520			   ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1521	}
1522
1523	if (posted)
1524		io_commit_cqring(ctx);
1525	spin_unlock_irqrestore(&ctx->completion_lock, flags);
1526	if (posted)
1527		io_cqring_ev_posted(ctx);
1528	return all_flushed;
1529}
1530
1531static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1532{
1533	bool ret = true;
1534
1535	if (test_bit(0, &ctx->check_cq_overflow)) {
1536		/* iopoll syncs against uring_lock, not completion_lock */
1537		if (ctx->flags & IORING_SETUP_IOPOLL)
1538			mutex_lock(&ctx->uring_lock);
1539		ret = __io_cqring_overflow_flush(ctx, force);
1540		if (ctx->flags & IORING_SETUP_IOPOLL)
1541			mutex_unlock(&ctx->uring_lock);
1542	}
1543
1544	return ret;
1545}
1546
1547/*
1548 * Shamelessly stolen from the mm implementation of page reference checking,
1549 * see commit f958d7b528b1 for details.
1550 */
1551#define req_ref_zero_or_close_to_overflow(req)	\
1552	((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1553
1554static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1555{
1556	return atomic_inc_not_zero(&req->refs);
1557}
1558
1559static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1560{
1561	WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1562	return atomic_sub_and_test(refs, &req->refs);
1563}
1564
1565static inline bool req_ref_put_and_test(struct io_kiocb *req)
1566{
1567	WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1568	return atomic_dec_and_test(&req->refs);
1569}
1570
1571static inline void req_ref_put(struct io_kiocb *req)
1572{
1573	WARN_ON_ONCE(req_ref_put_and_test(req));
1574}
1575
1576static inline void req_ref_get(struct io_kiocb *req)
1577{
1578	WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1579	atomic_inc(&req->refs);
1580}
1581
1582/* must to be called somewhat shortly after putting a request */
1583static inline void io_put_task(struct task_struct *task, int nr)
1584{
1585	struct io_uring_task *tctx = task->io_uring;
1586
1587	percpu_counter_sub(&tctx->inflight, nr);
1588	if (unlikely(atomic_read(&tctx->in_idle)))
1589		wake_up(&tctx->wait);
1590	put_task_struct_many(task, nr);
1591}
1592
1593static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1594				     long res, unsigned int cflags)
1595{
1596	struct io_overflow_cqe *ocqe;
1597
1598	ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1599	if (!ocqe) {
1600		/*
1601		 * If we're in ring overflow flush mode, or in task cancel mode,
1602		 * or cannot allocate an overflow entry, then we need to drop it
1603		 * on the floor.
1604		 */
1605		io_account_cq_overflow(ctx);
1606		return false;
1607	}
1608	if (list_empty(&ctx->cq_overflow_list)) {
1609		set_bit(0, &ctx->check_cq_overflow);
1610		WRITE_ONCE(ctx->rings->sq_flags,
1611			   ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1612
1613	}
1614	ocqe->cqe.user_data = user_data;
1615	ocqe->cqe.res = res;
1616	ocqe->cqe.flags = cflags;
1617	list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1618	return true;
1619}
1620
1621static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1622					  long res, unsigned int cflags)
1623{
1624	struct io_uring_cqe *cqe;
1625
1626	trace_io_uring_complete(ctx, user_data, res, cflags);
1627
1628	/*
1629	 * If we can't get a cq entry, userspace overflowed the
1630	 * submission (by quite a lot). Increment the overflow count in
1631	 * the ring.
1632	 */
1633	cqe = io_get_cqe(ctx);
1634	if (likely(cqe)) {
1635		WRITE_ONCE(cqe->user_data, user_data);
1636		WRITE_ONCE(cqe->res, res);
1637		WRITE_ONCE(cqe->flags, cflags);
1638		return true;
1639	}
1640	return io_cqring_event_overflow(ctx, user_data, res, cflags);
1641}
1642
1643/* not as hot to bloat with inlining */
1644static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1645					  long res, unsigned int cflags)
1646{
1647	return __io_cqring_fill_event(ctx, user_data, res, cflags);
1648}
1649
1650static void io_req_complete_post(struct io_kiocb *req, long res,
1651				 unsigned int cflags)
1652{
1653	struct io_ring_ctx *ctx = req->ctx;
1654	unsigned long flags;
1655
1656	spin_lock_irqsave(&ctx->completion_lock, flags);
1657	__io_cqring_fill_event(ctx, req->user_data, res, cflags);
1658	/*
1659	 * If we're the last reference to this request, add to our locked
1660	 * free_list cache.
1661	 */
1662	if (req_ref_put_and_test(req)) {
1663		if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1664			if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL))
1665				io_disarm_next(req);
1666			if (req->link) {
1667				io_req_task_queue(req->link);
1668				req->link = NULL;
1669			}
1670		}
1671		io_dismantle_req(req);
1672		io_put_task(req->task, 1);
1673		list_add(&req->compl.list, &ctx->locked_free_list);
1674		ctx->locked_free_nr++;
1675	} else {
1676		if (!percpu_ref_tryget(&ctx->refs))
1677			req = NULL;
1678	}
1679	io_commit_cqring(ctx);
1680	spin_unlock_irqrestore(&ctx->completion_lock, flags);
1681
1682	if (req) {
1683		io_cqring_ev_posted(ctx);
1684		percpu_ref_put(&ctx->refs);
1685	}
1686}
1687
1688static inline bool io_req_needs_clean(struct io_kiocb *req)
1689{
1690	return req->flags & IO_REQ_CLEAN_FLAGS;
1691}
1692
1693static void io_req_complete_state(struct io_kiocb *req, long res,
1694				  unsigned int cflags)
1695{
1696	if (io_req_needs_clean(req))
1697		io_clean_op(req);
1698	req->result = res;
1699	req->compl.cflags = cflags;
1700	req->flags |= REQ_F_COMPLETE_INLINE;
1701}
1702
1703static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1704				     long res, unsigned cflags)
1705{
1706	if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1707		io_req_complete_state(req, res, cflags);
1708	else
1709		io_req_complete_post(req, res, cflags);
1710}
1711
1712static inline void io_req_complete(struct io_kiocb *req, long res)
1713{
1714	__io_req_complete(req, 0, res, 0);
1715}
1716
1717static void io_req_complete_failed(struct io_kiocb *req, long res)
1718{
1719	req_set_fail(req);
1720	io_put_req(req);
1721	io_req_complete_post(req, res, 0);
1722}
1723
1724/*
1725 * Don't initialise the fields below on every allocation, but do that in
1726 * advance and keep them valid across allocations.
1727 */
1728static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1729{
1730	req->ctx = ctx;
1731	req->link = NULL;
1732	req->async_data = NULL;
1733	/* not necessary, but safer to zero */
1734	req->result = 0;
1735}
1736
1737static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1738					struct io_comp_state *cs)
1739{
1740	spin_lock_irq(&ctx->completion_lock);
1741	list_splice_init(&ctx->locked_free_list, &cs->free_list);
1742	ctx->locked_free_nr = 0;
1743	spin_unlock_irq(&ctx->completion_lock);
1744}
1745
1746/* Returns true IFF there are requests in the cache */
1747static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1748{
1749	struct io_submit_state *state = &ctx->submit_state;
1750	struct io_comp_state *cs = &state->comp;
1751	int nr;
1752
1753	/*
1754	 * If we have more than a batch's worth of requests in our IRQ side
1755	 * locked cache, grab the lock and move them over to our submission
1756	 * side cache.
1757	 */
1758	if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1759		io_flush_cached_locked_reqs(ctx, cs);
1760
1761	nr = state->free_reqs;
1762	while (!list_empty(&cs->free_list)) {
1763		struct io_kiocb *req = list_first_entry(&cs->free_list,
1764						struct io_kiocb, compl.list);
1765
1766		list_del(&req->compl.list);
1767		state->reqs[nr++] = req;
1768		if (nr == ARRAY_SIZE(state->reqs))
1769			break;
1770	}
1771
1772	state->free_reqs = nr;
1773	return nr != 0;
1774}
1775
1776static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1777{
1778	struct io_submit_state *state = &ctx->submit_state;
1779	gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1780	int ret, i;
1781
1782	BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1783
1784	if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1785		goto got_req;
1786
1787	ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1788				    state->reqs);
1789
1790	/*
1791	 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1792	 * retry single alloc to be on the safe side.
1793	 */
1794	if (unlikely(ret <= 0)) {
1795		state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1796		if (!state->reqs[0])
1797			return NULL;
1798		ret = 1;
1799	}
1800
1801	for (i = 0; i < ret; i++)
1802		io_preinit_req(state->reqs[i], ctx);
1803	state->free_reqs = ret;
1804got_req:
1805	state->free_reqs--;
1806	return state->reqs[state->free_reqs];
1807}
1808
1809static inline void io_put_file(struct file *file)
1810{
1811	if (file)
1812		fput(file);
1813}
1814
1815static void io_dismantle_req(struct io_kiocb *req)
1816{
1817	unsigned int flags = req->flags;
1818
1819	if (io_req_needs_clean(req))
1820		io_clean_op(req);
1821	if (!(flags & REQ_F_FIXED_FILE))
1822		io_put_file(req->file);
1823	if (req->fixed_rsrc_refs)
1824		percpu_ref_put(req->fixed_rsrc_refs);
1825	if (req->async_data) {
1826		kfree(req->async_data);
1827		req->async_data = NULL;
1828	}
1829}
1830
1831static void __io_free_req(struct io_kiocb *req)
1832{
1833	struct io_ring_ctx *ctx = req->ctx;
1834
1835	io_dismantle_req(req);
1836	io_put_task(req->task, 1);
1837
1838	kmem_cache_free(req_cachep, req);
1839	percpu_ref_put(&ctx->refs);
1840}
1841
1842static inline void io_remove_next_linked(struct io_kiocb *req)
1843{
1844	struct io_kiocb *nxt = req->link;
1845
1846	req->link = nxt->link;
1847	nxt->link = NULL;
1848}
1849
1850static bool io_kill_linked_timeout(struct io_kiocb *req)
1851	__must_hold(&req->ctx->completion_lock)
1852{
1853	struct io_kiocb *link = req->link;
1854
1855	/*
1856	 * Can happen if a linked timeout fired and link had been like
1857	 * req -> link t-out -> link t-out [-> ...]
1858	 */
1859	if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1860		struct io_timeout_data *io = link->async_data;
1861
1862		io_remove_next_linked(req);
1863		link->timeout.head = NULL;
1864		if (hrtimer_try_to_cancel(&io->timer) != -1) {
1865			io_cqring_fill_event(link->ctx, link->user_data,
1866					     -ECANCELED, 0);
1867			io_put_req_deferred(link, 1);
1868			return true;
1869		}
1870	}
1871	return false;
1872}
1873
1874static void io_fail_links(struct io_kiocb *req)
1875	__must_hold(&req->ctx->completion_lock)
1876{
1877	struct io_kiocb *nxt, *link = req->link;
1878
1879	req->link = NULL;
1880	while (link) {
1881		nxt = link->link;
1882		link->link = NULL;
1883
1884		trace_io_uring_fail_link(req, link);
1885		io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1886		io_put_req_deferred(link, 2);
1887		link = nxt;
1888	}
1889}
1890
1891static bool io_disarm_next(struct io_kiocb *req)
1892	__must_hold(&req->ctx->completion_lock)
1893{
1894	bool posted = false;
1895
1896	if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1897		posted = io_kill_linked_timeout(req);
1898	if (unlikely((req->flags & REQ_F_FAIL) &&
1899		     !(req->flags & REQ_F_HARDLINK))) {
1900		posted |= (req->link != NULL);
1901		io_fail_links(req);
1902	}
1903	return posted;
1904}
1905
1906static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1907{
1908	struct io_kiocb *nxt;
1909
1910	/*
1911	 * If LINK is set, we have dependent requests in this chain. If we
1912	 * didn't fail this request, queue the first one up, moving any other
1913	 * dependencies to the next request. In case of failure, fail the rest
1914	 * of the chain.
1915	 */
1916	if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) {
1917		struct io_ring_ctx *ctx = req->ctx;
1918		unsigned long flags;
1919		bool posted;
1920
1921		spin_lock_irqsave(&ctx->completion_lock, flags);
1922		posted = io_disarm_next(req);
1923		if (posted)
1924			io_commit_cqring(req->ctx);
1925		spin_unlock_irqrestore(&ctx->completion_lock, flags);
1926		if (posted)
1927			io_cqring_ev_posted(ctx);
1928	}
1929	nxt = req->link;
1930	req->link = NULL;
1931	return nxt;
1932}
1933
1934static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1935{
1936	if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1937		return NULL;
1938	return __io_req_find_next(req);
1939}
1940
1941static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1942{
1943	if (!ctx)
1944		return;
1945	if (ctx->submit_state.comp.nr) {
1946		mutex_lock(&ctx->uring_lock);
1947		io_submit_flush_completions(ctx);
1948		mutex_unlock(&ctx->uring_lock);
1949	}
1950	percpu_ref_put(&ctx->refs);
1951}
1952
1953static void tctx_task_work(struct callback_head *cb)
1954{
1955	struct io_ring_ctx *ctx = NULL;
1956	struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1957						  task_work);
1958
1959	while (1) {
1960		struct io_wq_work_node *node;
1961
1962		spin_lock_irq(&tctx->task_lock);
1963		node = tctx->task_list.first;
1964		INIT_WQ_LIST(&tctx->task_list);
1965		spin_unlock_irq(&tctx->task_lock);
1966
1967		while (node) {
1968			struct io_wq_work_node *next = node->next;
1969			struct io_kiocb *req = container_of(node, struct io_kiocb,
1970							    io_task_work.node);
1971
1972			if (req->ctx != ctx) {
1973				ctx_flush_and_put(ctx);
1974				ctx = req->ctx;
1975				percpu_ref_get(&ctx->refs);
1976			}
1977			req->io_task_work.func(req);
1978			node = next;
1979		}
1980		if (wq_list_empty(&tctx->task_list)) {
1981			spin_lock_irq(&tctx->task_lock);
1982			clear_bit(0, &tctx->task_state);
1983			if (wq_list_empty(&tctx->task_list)) {
1984				spin_unlock_irq(&tctx->task_lock);
1985				break;
1986			}
1987			spin_unlock_irq(&tctx->task_lock);
1988			/* another tctx_task_work() is enqueued, yield */
1989			if (test_and_set_bit(0, &tctx->task_state))
1990				break;
1991		}
1992		cond_resched();
1993	}
1994
1995	ctx_flush_and_put(ctx);
1996}
1997
1998static void io_req_task_work_add(struct io_kiocb *req)
1999{
2000	struct task_struct *tsk = req->task;
2001	struct io_uring_task *tctx = tsk->io_uring;
2002	enum task_work_notify_mode notify;
2003	struct io_wq_work_node *node;
2004	unsigned long flags;
2005
2006	WARN_ON_ONCE(!tctx);
2007
2008	spin_lock_irqsave(&tctx->task_lock, flags);
2009	wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2010	spin_unlock_irqrestore(&tctx->task_lock, flags);
2011
2012	/* task_work already pending, we're done */
2013	if (test_bit(0, &tctx->task_state) ||
2014	    test_and_set_bit(0, &tctx->task_state))
2015		return;
2016
2017	/*
2018	 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2019	 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2020	 * processing task_work. There's no reliable way to tell if TWA_RESUME
2021	 * will do the job.
2022	 */
2023	notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2024	if (!task_work_add(tsk, &tctx->task_work, notify)) {
2025		wake_up_process(tsk);
2026		return;
2027	}
2028
2029	clear_bit(0, &tctx->task_state);
2030	spin_lock_irqsave(&tctx->task_lock, flags);
2031	node = tctx->task_list.first;
2032	INIT_WQ_LIST(&tctx->task_list);
2033	spin_unlock_irqrestore(&tctx->task_lock, flags);
2034
2035	while (node) {
2036		req = container_of(node, struct io_kiocb, io_task_work.node);
2037		node = node->next;
2038		if (llist_add(&req->io_task_work.fallback_node,
2039			      &req->ctx->fallback_llist))
2040			schedule_delayed_work(&req->ctx->fallback_work, 1);
2041	}
2042}
2043
2044static void io_req_task_cancel(struct io_kiocb *req)
2045{
2046	struct io_ring_ctx *ctx = req->ctx;
2047
2048	/* ctx is guaranteed to stay alive while we hold uring_lock */
2049	mutex_lock(&ctx->uring_lock);
2050	io_req_complete_failed(req, req->result);
2051	mutex_unlock(&ctx->uring_lock);
2052}
2053
2054static void io_req_task_submit(struct io_kiocb *req)
2055{
2056	struct io_ring_ctx *ctx = req->ctx;
2057
2058	/* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2059	mutex_lock(&ctx->uring_lock);
2060	if (likely(!(req->task->flags & PF_EXITING)))
2061		__io_queue_sqe(req);
2062	else
2063		io_req_complete_failed(req, -EFAULT);
2064	mutex_unlock(&ctx->uring_lock);
2065}
2066
2067static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2068{
2069	req->result = ret;
2070	req->io_task_work.func = io_req_task_cancel;
2071	io_req_task_work_add(req);
2072}
2073
2074static void io_req_task_queue(struct io_kiocb *req)
2075{
2076	req->io_task_work.func = io_req_task_submit;
2077	io_req_task_work_add(req);
2078}
2079
2080static void io_req_task_queue_reissue(struct io_kiocb *req)
2081{
2082	req->io_task_work.func = io_queue_async_work;
2083	io_req_task_work_add(req);
2084}
2085
2086static inline void io_queue_next(struct io_kiocb *req)
2087{
2088	struct io_kiocb *nxt = io_req_find_next(req);
2089
2090	if (nxt)
2091		io_req_task_queue(nxt);
2092}
2093
2094static void io_free_req(struct io_kiocb *req)
2095{
2096	io_queue_next(req);
2097	__io_free_req(req);
2098}
2099
2100struct req_batch {
2101	struct task_struct	*task;
2102	int			task_refs;
2103	int			ctx_refs;
2104};
2105
2106static inline void io_init_req_batch(struct req_batch *rb)
2107{
2108	rb->task_refs = 0;
2109	rb->ctx_refs = 0;
2110	rb->task = NULL;
2111}
2112
2113static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2114				     struct req_batch *rb)
2115{
2116	if (rb->ctx_refs)
2117		percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2118	if (rb->task == current)
2119		current->io_uring->cached_refs += rb->task_refs;
2120	else if (rb->task)
2121		io_put_task(rb->task, rb->task_refs);
2122}
2123
2124static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2125			      struct io_submit_state *state)
2126{
2127	io_queue_next(req);
2128	io_dismantle_req(req);
2129
2130	if (req->task != rb->task) {
2131		if (rb->task)
2132			io_put_task(rb->task, rb->task_refs);
2133		rb->task = req->task;
2134		rb->task_refs = 0;
2135	}
2136	rb->task_refs++;
2137	rb->ctx_refs++;
2138
2139	if (state->free_reqs != ARRAY_SIZE(state->reqs))
2140		state->reqs[state->free_reqs++] = req;
2141	else
2142		list_add(&req->compl.list, &state->comp.free_list);
2143}
2144
2145static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2146	__must_hold(&req->ctx->uring_lock)
2147{
2148	struct io_comp_state *cs = &ctx->submit_state.comp;
2149	int i, nr = cs->nr;
2150	struct req_batch rb;
2151
2152	spin_lock_irq(&ctx->completion_lock);
2153	for (i = 0; i < nr; i++) {
2154		struct io_kiocb *req = cs->reqs[i];
2155
2156		__io_cqring_fill_event(ctx, req->user_data, req->result,
2157					req->compl.cflags);
2158	}
2159	io_commit_cqring(ctx);
2160	spin_unlock_irq(&ctx->completion_lock);
2161	io_cqring_ev_posted(ctx);
2162
2163	io_init_req_batch(&rb);
2164	for (i = 0; i < nr; i++) {
2165		struct io_kiocb *req = cs->reqs[i];
2166
2167		/* submission and completion refs */
2168		if (req_ref_sub_and_test(req, 2))
2169			io_req_free_batch(&rb, req, &ctx->submit_state);
2170	}
2171
2172	io_req_free_batch_finish(ctx, &rb);
2173	cs->nr = 0;
2174}
2175
2176/*
2177 * Drop reference to request, return next in chain (if there is one) if this
2178 * was the last reference to this request.
2179 */
2180static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2181{
2182	struct io_kiocb *nxt = NULL;
2183
2184	if (req_ref_put_and_test(req)) {
2185		nxt = io_req_find_next(req);
2186		__io_free_req(req);
2187	}
2188	return nxt;
2189}
2190
2191static inline void io_put_req(struct io_kiocb *req)
2192{
2193	if (req_ref_put_and_test(req))
2194		io_free_req(req);
2195}
2196
2197static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2198{
2199	if (req_ref_sub_and_test(req, refs)) {
2200		req->io_task_work.func = io_free_req;
2201		io_req_task_work_add(req);
2202	}
2203}
2204
2205static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2206{
2207	/* See comment at the top of this file */
2208	smp_rmb();
2209	return __io_cqring_events(ctx);
2210}
2211
2212static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2213{
2214	struct io_rings *rings = ctx->rings;
2215
2216	/* make sure SQ entry isn't read before tail */
2217	return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2218}
2219
2220static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2221{
2222	unsigned int cflags;
2223
2224	cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2225	cflags |= IORING_CQE_F_BUFFER;
2226	req->flags &= ~REQ_F_BUFFER_SELECTED;
2227	kfree(kbuf);
2228	return cflags;
2229}
2230
2231static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2232{
2233	struct io_buffer *kbuf;
2234
2235	kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2236	return io_put_kbuf(req, kbuf);
2237}
2238
2239static inline bool io_run_task_work(void)
2240{
2241	if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2242		__set_current_state(TASK_RUNNING);
2243		tracehook_notify_signal();
2244		return true;
2245	}
2246
2247	return false;
2248}
2249
2250/*
2251 * Find and free completed poll iocbs
2252 */
2253static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2254			       struct list_head *done, bool resubmit)
2255{
2256	struct req_batch rb;
2257	struct io_kiocb *req;
2258
2259	/* order with ->result store in io_complete_rw_iopoll() */
2260	smp_rmb();
2261
2262	io_init_req_batch(&rb);
2263	while (!list_empty(done)) {
2264		int cflags = 0;
2265
2266		req = list_first_entry(done, struct io_kiocb, inflight_entry);
2267		list_del(&req->inflight_entry);
2268
2269		if (READ_ONCE(req->result) == -EAGAIN && resubmit &&
2270		    !(req->flags & REQ_F_DONT_REISSUE)) {
2271			req->iopoll_completed = 0;
2272			req_ref_get(req);
2273			io_req_task_queue_reissue(req);
2274			continue;
2275		}
2276
2277		if (req->flags & REQ_F_BUFFER_SELECTED)
2278			cflags = io_put_rw_kbuf(req);
2279
2280		__io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2281		(*nr_events)++;
2282
2283		if (req_ref_put_and_test(req))
2284			io_req_free_batch(&rb, req, &ctx->submit_state);
2285	}
2286
2287	io_commit_cqring(ctx);
2288	io_cqring_ev_posted_iopoll(ctx);
2289	io_req_free_batch_finish(ctx, &rb);
2290}
2291
2292static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2293			long min, bool resubmit)
2294{
2295	struct io_kiocb *req, *tmp;
2296	LIST_HEAD(done);
2297	bool spin;
2298
2299	/*
2300	 * Only spin for completions if we don't have multiple devices hanging
2301	 * off our complete list, and we're under the requested amount.
2302	 */
2303	spin = !ctx->poll_multi_queue && *nr_events < min;
2304
2305	list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2306		struct kiocb *kiocb = &req->rw.kiocb;
2307		int ret;
2308
2309		/*
2310		 * Move completed and retryable entries to our local lists.
2311		 * If we find a request that requires polling, break out
2312		 * and complete those lists first, if we have entries there.
2313		 */
2314		if (READ_ONCE(req->iopoll_completed)) {
2315			list_move_tail(&req->inflight_entry, &done);
2316			continue;
2317		}
2318		if (!list_empty(&done))
2319			break;
2320
2321		ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2322		if (unlikely(ret < 0))
2323			return ret;
2324		else if (ret)
2325			spin = false;
2326
2327		/* iopoll may have completed current req */
2328		if (READ_ONCE(req->iopoll_completed))
2329			list_move_tail(&req->inflight_entry, &done);
2330	}
2331
2332	if (!list_empty(&done))
2333		io_iopoll_complete(ctx, nr_events, &done, resubmit);
2334
2335	return 0;
2336}
2337
2338/*
2339 * We can't just wait for polled events to come to us, we have to actively
2340 * find and complete them.
2341 */
2342static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2343{
2344	if (!(ctx->flags & IORING_SETUP_IOPOLL))
2345		return;
2346
2347	mutex_lock(&ctx->uring_lock);
2348	while (!list_empty(&ctx->iopoll_list)) {
2349		unsigned int nr_events = 0;
2350
2351		io_do_iopoll(ctx, &nr_events, 0, false);
2352
2353		/* let it sleep and repeat later if can't complete a request */
2354		if (nr_events == 0)
2355			break;
2356		/*
2357		 * Ensure we allow local-to-the-cpu processing to take place,
2358		 * in this case we need to ensure that we reap all events.
2359		 * Also let task_work, etc. to progress by releasing the mutex
2360		 */
2361		if (need_resched()) {
2362			mutex_unlock(&ctx->uring_lock);
2363			cond_resched();
2364			mutex_lock(&ctx->uring_lock);
2365		}
2366	}
2367	mutex_unlock(&ctx->uring_lock);
2368}
2369
2370static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2371{
2372	unsigned int nr_events = 0;
2373	int ret = 0;
2374
2375	/*
2376	 * We disallow the app entering submit/complete with polling, but we
2377	 * still need to lock the ring to prevent racing with polled issue
2378	 * that got punted to a workqueue.
2379	 */
2380	mutex_lock(&ctx->uring_lock);
2381	/*
2382	 * Don't enter poll loop if we already have events pending.
2383	 * If we do, we can potentially be spinning for commands that
2384	 * already triggered a CQE (eg in error).
2385	 */
2386	if (test_bit(0, &ctx->check_cq_overflow))
2387		__io_cqring_overflow_flush(ctx, false);
2388	if (io_cqring_events(ctx))
2389		goto out;
2390	do {
2391		/*
2392		 * If a submit got punted to a workqueue, we can have the
2393		 * application entering polling for a command before it gets
2394		 * issued. That app will hold the uring_lock for the duration
2395		 * of the poll right here, so we need to take a breather every
2396		 * now and then to ensure that the issue has a chance to add
2397		 * the poll to the issued list. Otherwise we can spin here
2398		 * forever, while the workqueue is stuck trying to acquire the
2399		 * very same mutex.
2400		 */
2401		if (list_empty(&ctx->iopoll_list)) {
2402			u32 tail = ctx->cached_cq_tail;
2403
2404			mutex_unlock(&ctx->uring_lock);
2405			io_run_task_work();
2406			mutex_lock(&ctx->uring_lock);
2407
2408			/* some requests don't go through iopoll_list */
2409			if (tail != ctx->cached_cq_tail ||
2410			    list_empty(&ctx->iopoll_list))
2411				break;
2412		}
2413		ret = io_do_iopoll(ctx, &nr_events, min, true);
2414	} while (!ret && nr_events < min && !need_resched());
2415out:
2416	mutex_unlock(&ctx->uring_lock);
2417	return ret;
2418}
2419
2420static void kiocb_end_write(struct io_kiocb *req)
2421{
2422	/*
2423	 * Tell lockdep we inherited freeze protection from submission
2424	 * thread.
2425	 */
2426	if (req->flags & REQ_F_ISREG) {
2427		struct super_block *sb = file_inode(req->file)->i_sb;
2428
2429		__sb_writers_acquired(sb, SB_FREEZE_WRITE);
2430		sb_end_write(sb);
2431	}
2432}
2433
2434#ifdef CONFIG_BLOCK
2435static bool io_resubmit_prep(struct io_kiocb *req)
2436{
2437	struct io_async_rw *rw = req->async_data;
2438
2439	if (!rw)
2440		return !io_req_prep_async(req);
2441	/* may have left rw->iter inconsistent on -EIOCBQUEUED */
2442	iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2443	return true;
2444}
2445
2446static bool io_rw_should_reissue(struct io_kiocb *req)
2447{
2448	umode_t mode = file_inode(req->file)->i_mode;
2449	struct io_ring_ctx *ctx = req->ctx;
2450
2451	if (!S_ISBLK(mode) && !S_ISREG(mode))
2452		return false;
2453	if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2454	    !(ctx->flags & IORING_SETUP_IOPOLL)))
2455		return false;
2456	/*
2457	 * If ref is dying, we might be running poll reap from the exit work.
2458	 * Don't attempt to reissue from that path, just let it fail with
2459	 * -EAGAIN.
2460	 */
2461	if (percpu_ref_is_dying(&ctx->refs))
2462		return false;
2463	/*
2464	 * Play it safe and assume not safe to re-import and reissue if we're
2465	 * not in the original thread group (or in task context).
2466	 */
2467	if (!same_thread_group(req->task, current) || !in_task())
2468		return false;
2469	return true;
2470}
2471#else
2472static bool io_resubmit_prep(struct io_kiocb *req)
2473{
2474	return false;
2475}
2476static bool io_rw_should_reissue(struct io_kiocb *req)
2477{
2478	return false;
2479}
2480#endif
2481
2482static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2483			     unsigned int issue_flags)
2484{
2485	int cflags = 0;
2486
2487	if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2488		kiocb_end_write(req);
2489	if (res != req->result) {
2490		if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2491		    io_rw_should_reissue(req)) {
2492			req->flags |= REQ_F_REISSUE;
2493			return;
2494		}
2495		req_set_fail(req);
2496	}
2497	if (req->flags & REQ_F_BUFFER_SELECTED)
2498		cflags = io_put_rw_kbuf(req);
2499	__io_req_complete(req, issue_flags, res, cflags);
2500}
2501
2502static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2503{
2504	struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2505
2506	__io_complete_rw(req, res, res2, 0);
2507}
2508
2509static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2510{
2511	struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2512
2513	if (kiocb->ki_flags & IOCB_WRITE)
2514		kiocb_end_write(req);
2515	if (unlikely(res != req->result)) {
2516		if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2517		    io_resubmit_prep(req))) {
2518			req_set_fail(req);
2519			req->flags |= REQ_F_DONT_REISSUE;
2520		}
2521	}
2522
2523	WRITE_ONCE(req->result, res);
2524	/* order with io_iopoll_complete() checking ->result */
2525	smp_wmb();
2526	WRITE_ONCE(req->iopoll_completed, 1);
2527}
2528
2529/*
2530 * After the iocb has been issued, it's safe to be found on the poll list.
2531 * Adding the kiocb to the list AFTER submission ensures that we don't
2532 * find it from a io_do_iopoll() thread before the issuer is done
2533 * accessing the kiocb cookie.
2534 */
2535static void io_iopoll_req_issued(struct io_kiocb *req)
2536{
2537	struct io_ring_ctx *ctx = req->ctx;
2538	const bool in_async = io_wq_current_is_worker();
2539
2540	/* workqueue context doesn't hold uring_lock, grab it now */
2541	if (unlikely(in_async))
2542		mutex_lock(&ctx->uring_lock);
2543
2544	/*
2545	 * Track whether we have multiple files in our lists. This will impact
2546	 * how we do polling eventually, not spinning if we're on potentially
2547	 * different devices.
2548	 */
2549	if (list_empty(&ctx->iopoll_list)) {
2550		ctx->poll_multi_queue = false;
2551	} else if (!ctx->poll_multi_queue) {
2552		struct io_kiocb *list_req;
2553		unsigned int queue_num0, queue_num1;
2554
2555		list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2556						inflight_entry);
2557
2558		if (list_req->file != req->file) {
2559			ctx->poll_multi_queue = true;
2560		} else {
2561			queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2562			queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2563			if (queue_num0 != queue_num1)
2564				ctx->poll_multi_queue = true;
2565		}
2566	}
2567
2568	/*
2569	 * For fast devices, IO may have already completed. If it has, add
2570	 * it to the front so we find it first.
2571	 */
2572	if (READ_ONCE(req->iopoll_completed))
2573		list_add(&req->inflight_entry, &ctx->iopoll_list);
2574	else
2575		list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2576
2577	if (unlikely(in_async)) {
2578		/*
2579		 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2580		 * in sq thread task context or in io worker task context. If
2581		 * current task context is sq thread, we don't need to check
2582		 * whether should wake up sq thread.
2583		 */
2584		if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2585		    wq_has_sleeper(&ctx->sq_data->wait))
2586			wake_up(&ctx->sq_data->wait);
2587
2588		mutex_unlock(&ctx->uring_lock);
2589	}
2590}
2591
2592static inline void io_state_file_put(struct io_submit_state *state)
2593{
2594	if (state->file_refs) {
2595		fput_many(state->file, state->file_refs);
2596		state->file_refs = 0;
2597	}
2598}
2599
2600/*
2601 * Get as many references to a file as we have IOs left in this submission,
2602 * assuming most submissions are for one file, or at least that each file
2603 * has more than one submission.
2604 */
2605static struct file *__io_file_get(struct io_submit_state *state, int fd)
2606{
2607	if (!state)
2608		return fget(fd);
2609
2610	if (state->file_refs) {
2611		if (state->fd == fd) {
2612			state->file_refs--;
2613			return state->file;
2614		}
2615		io_state_file_put(state);
2616	}
2617	state->file = fget_many(fd, state->ios_left);
2618	if (unlikely(!state->file))
2619		return NULL;
2620
2621	state->fd = fd;
2622	state->file_refs = state->ios_left - 1;
2623	return state->file;
2624}
2625
2626static bool io_bdev_nowait(struct block_device *bdev)
2627{
2628	return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2629}
2630
2631/*
2632 * If we tracked the file through the SCM inflight mechanism, we could support
2633 * any file. For now, just ensure that anything potentially problematic is done
2634 * inline.
2635 */
2636static bool __io_file_supports_nowait(struct file *file, int rw)
2637{
2638	umode_t mode = file_inode(file)->i_mode;
2639
2640	if (S_ISBLK(mode)) {
2641		if (IS_ENABLED(CONFIG_BLOCK) &&
2642		    io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2643			return true;
2644		return false;
2645	}
2646	if (S_ISSOCK(mode))
2647		return true;
2648	if (S_ISREG(mode)) {
2649		if (IS_ENABLED(CONFIG_BLOCK) &&
2650		    io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2651		    file->f_op != &io_uring_fops)
2652			return true;
2653		return false;
2654	}
2655
2656	/* any ->read/write should understand O_NONBLOCK */
2657	if (file->f_flags & O_NONBLOCK)
2658		return true;
2659
2660	if (!(file->f_mode & FMODE_NOWAIT))
2661		return false;
2662
2663	if (rw == READ)
2664		return file->f_op->read_iter != NULL;
2665
2666	return file->f_op->write_iter != NULL;
2667}
2668
2669static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2670{
2671	if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2672		return true;
2673	else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2674		return true;
2675
2676	return __io_file_supports_nowait(req->file, rw);
2677}
2678
2679static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2680{
2681	struct io_ring_ctx *ctx = req->ctx;
2682	struct kiocb *kiocb = &req->rw.kiocb;
2683	struct file *file = req->file;
2684	unsigned ioprio;
2685	int ret;
2686
2687	if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2688		req->flags |= REQ_F_ISREG;
2689
2690	kiocb->ki_pos = READ_ONCE(sqe->off);
2691	if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2692		req->flags |= REQ_F_CUR_POS;
2693		kiocb->ki_pos = file->f_pos;
2694	}
2695	kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2696	kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2697	ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2698	if (unlikely(ret))
2699		return ret;
2700
2701	/* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2702	if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2703		req->flags |= REQ_F_NOWAIT;
2704
2705	ioprio = READ_ONCE(sqe->ioprio);
2706	if (ioprio) {
2707		ret = ioprio_check_cap(ioprio);
2708		if (ret)
2709			return ret;
2710
2711		kiocb->ki_ioprio = ioprio;
2712	} else
2713		kiocb->ki_ioprio = get_current_ioprio();
2714
2715	if (ctx->flags & IORING_SETUP_IOPOLL) {
2716		if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2717		    !kiocb->ki_filp->f_op->iopoll)
2718			return -EOPNOTSUPP;
2719
2720		kiocb->ki_flags |= IOCB_HIPRI;
2721		kiocb->ki_complete = io_complete_rw_iopoll;
2722		req->iopoll_completed = 0;
2723	} else {
2724		if (kiocb->ki_flags & IOCB_HIPRI)
2725			return -EINVAL;
2726		kiocb->ki_complete = io_complete_rw;
2727	}
2728
2729	if (req->opcode == IORING_OP_READ_FIXED ||
2730	    req->opcode == IORING_OP_WRITE_FIXED) {
2731		req->imu = NULL;
2732		io_req_set_rsrc_node(req);
2733	}
2734
2735	req->rw.addr = READ_ONCE(sqe->addr);
2736	req->rw.len = READ_ONCE(sqe->len);
2737	req->buf_index = READ_ONCE(sqe->buf_index);
2738	return 0;
2739}
2740
2741static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2742{
2743	switch (ret) {
2744	case -EIOCBQUEUED:
2745		break;
2746	case -ERESTARTSYS:
2747	case -ERESTARTNOINTR:
2748	case -ERESTARTNOHAND:
2749	case -ERESTART_RESTARTBLOCK:
2750		/*
2751		 * We can't just restart the syscall, since previously
2752		 * submitted sqes may already be in progress. Just fail this
2753		 * IO with EINTR.
2754		 */
2755		ret = -EINTR;
2756		fallthrough;
2757	default:
2758		kiocb->ki_complete(kiocb, ret, 0);
2759	}
2760}
2761
2762static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2763		       unsigned int issue_flags)
2764{
2765	struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2766	struct io_async_rw *io = req->async_data;
2767	bool check_reissue = kiocb->ki_complete == io_complete_rw;
2768
2769	/* add previously done IO, if any */
2770	if (io && io->bytes_done > 0) {
2771		if (ret < 0)
2772			ret = io->bytes_done;
2773		else
2774			ret += io->bytes_done;
2775	}
2776
2777	if (req->flags & REQ_F_CUR_POS)
2778		req->file->f_pos = kiocb->ki_pos;
2779	if (ret >= 0 && check_reissue)
2780		__io_complete_rw(req, ret, 0, issue_flags);
2781	else
2782		io_rw_done(kiocb, ret);
2783
2784	if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2785		req->flags &= ~REQ_F_REISSUE;
2786		if (io_resubmit_prep(req)) {
2787			req_ref_get(req);
2788			io_req_task_queue_reissue(req);
2789		} else {
2790			int cflags = 0;
2791
2792			req_set_fail(req);
2793			if (req->flags & REQ_F_BUFFER_SELECTED)
2794				cflags = io_put_rw_kbuf(req);
2795			__io_req_complete(req, issue_flags, ret, cflags);
2796		}
2797	}
2798}
2799
2800static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2801			     struct io_mapped_ubuf *imu)
2802{
2803	size_t len = req->rw.len;
2804	u64 buf_end, buf_addr = req->rw.addr;
2805	size_t offset;
2806
2807	if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2808		return -EFAULT;
2809	/* not inside the mapped region */
2810	if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2811		return -EFAULT;
2812
2813	/*
2814	 * May not be a start of buffer, set size appropriately
2815	 * and advance us to the beginning.
2816	 */
2817	offset = buf_addr - imu->ubuf;
2818	iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2819
2820	if (offset) {
2821		/*
2822		 * Don't use iov_iter_advance() here, as it's really slow for
2823		 * using the latter parts of a big fixed buffer - it iterates
2824		 * over each segment manually. We can cheat a bit here, because
2825		 * we know that:
2826		 *
2827		 * 1) it's a BVEC iter, we set it up
2828		 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2829		 *    first and last bvec
2830		 *
2831		 * So just find our index, and adjust the iterator afterwards.
2832		 * If the offset is within the first bvec (or the whole first
2833		 * bvec, just use iov_iter_advance(). This makes it easier
2834		 * since we can just skip the first segment, which may not
2835		 * be PAGE_SIZE aligned.
2836		 */
2837		const struct bio_vec *bvec = imu->bvec;
2838
2839		if (offset <= bvec->bv_len) {
2840			iov_iter_advance(iter, offset);
2841		} else {
2842			unsigned long seg_skip;
2843
2844			/* skip first vec */
2845			offset -= bvec->bv_len;
2846			seg_skip = 1 + (offset >> PAGE_SHIFT);
2847
2848			iter->bvec = bvec + seg_skip;
2849			iter->nr_segs -= seg_skip;
2850			iter->count -= bvec->bv_len + offset;
2851			iter->iov_offset = offset & ~PAGE_MASK;
2852		}
2853	}
2854
2855	return 0;
2856}
2857
2858static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2859{
2860	struct io_ring_ctx *ctx = req->ctx;
2861	struct io_mapped_ubuf *imu = req->imu;
2862	u16 index, buf_index = req->buf_index;
2863
2864	if (likely(!imu)) {
2865		if (unlikely(buf_index >= ctx->nr_user_bufs))
2866			return -EFAULT;
2867		index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2868		imu = READ_ONCE(ctx->user_bufs[index]);
2869		req->imu = imu;
2870	}
2871	return __io_import_fixed(req, rw, iter, imu);
2872}
2873
2874static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2875{
2876	if (needs_lock)
2877		mutex_unlock(&ctx->uring_lock);
2878}
2879
2880static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2881{
2882	/*
2883	 * "Normal" inline submissions always hold the uring_lock, since we
2884	 * grab it from the system call. Same is true for the SQPOLL offload.
2885	 * The only exception is when we've detached the request and issue it
2886	 * from an async worker thread, grab the lock for that case.
2887	 */
2888	if (needs_lock)
2889		mutex_lock(&ctx->uring_lock);
2890}
2891
2892static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2893					  int bgid, struct io_buffer *kbuf,
2894					  bool needs_lock)
2895{
2896	struct io_buffer *head;
2897
2898	if (req->flags & REQ_F_BUFFER_SELECTED)
2899		return kbuf;
2900
2901	io_ring_submit_lock(req->ctx, needs_lock);
2902
2903	lockdep_assert_held(&req->ctx->uring_lock);
2904
2905	head = xa_load(&req->ctx->io_buffers, bgid);
2906	if (head) {
2907		if (!list_empty(&head->list)) {
2908			kbuf = list_last_entry(&head->list, struct io_buffer,
2909							list);
2910			list_del(&kbuf->list);
2911		} else {
2912			kbuf = head;
2913			xa_erase(&req->ctx->io_buffers, bgid);
2914		}
2915		if (*len > kbuf->len)
2916			*len = kbuf->len;
2917	} else {
2918		kbuf = ERR_PTR(-ENOBUFS);
2919	}
2920
2921	io_ring_submit_unlock(req->ctx, needs_lock);
2922
2923	return kbuf;
2924}
2925
2926static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2927					bool needs_lock)
2928{
2929	struct io_buffer *kbuf;
2930	u16 bgid;
2931
2932	kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2933	bgid = req->buf_index;
2934	kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2935	if (IS_ERR(kbuf))
2936		return kbuf;
2937	req->rw.addr = (u64) (unsigned long) kbuf;
2938	req->flags |= REQ_F_BUFFER_SELECTED;
2939	return u64_to_user_ptr(kbuf->addr);
2940}
2941
2942#ifdef CONFIG_COMPAT
2943static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2944				bool needs_lock)
2945{
2946	struct compat_iovec __user *uiov;
2947	compat_ssize_t clen;
2948	void __user *buf;
2949	ssize_t len;
2950
2951	uiov = u64_to_user_ptr(req->rw.addr);
2952	if (!access_ok(uiov, sizeof(*uiov)))
2953		return -EFAULT;
2954	if (__get_user(clen, &uiov->iov_len))
2955		return -EFAULT;
2956	if (clen < 0)
2957		return -EINVAL;
2958
2959	len = clen;
2960	buf = io_rw_buffer_select(req, &len, needs_lock);
2961	if (IS_ERR(buf))
2962		return PTR_ERR(buf);
2963	iov[0].iov_base = buf;
2964	iov[0].iov_len = (compat_size_t) len;
2965	return 0;
2966}
2967#endif
2968
2969static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2970				      bool needs_lock)
2971{
2972	struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2973	void __user *buf;
2974	ssize_t len;
2975
2976	if (copy_from_user(iov, uiov, sizeof(*uiov)))
2977		return -EFAULT;
2978
2979	len = iov[0].iov_len;
2980	if (len < 0)
2981		return -EINVAL;
2982	buf = io_rw_buffer_select(req, &len, needs_lock);
2983	if (IS_ERR(buf))
2984		return PTR_ERR(buf);
2985	iov[0].iov_base = buf;
2986	iov[0].iov_len = len;
2987	return 0;
2988}
2989
2990static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2991				    bool needs_lock)
2992{
2993	if (req->flags & REQ_F_BUFFER_SELECTED) {
2994		struct io_buffer *kbuf;
2995
2996		kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2997		iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2998		iov[0].iov_len = kbuf->len;
2999		return 0;
3000	}
3001	if (req->rw.len != 1)
3002		return -EINVAL;
3003
3004#ifdef CONFIG_COMPAT
3005	if (req->ctx->compat)
3006		return io_compat_import(req, iov, needs_lock);
3007#endif
3008
3009	return __io_iov_buffer_select(req, iov, needs_lock);
3010}
3011
3012static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3013			   struct iov_iter *iter, bool needs_lock)
3014{
3015	void __user *buf = u64_to_user_ptr(req->rw.addr);
3016	size_t sqe_len = req->rw.len;
3017	u8 opcode = req->opcode;
3018	ssize_t ret;
3019
3020	if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3021		*iovec = NULL;
3022		return io_import_fixed(req, rw, iter);
3023	}
3024
3025	/* buffer index only valid with fixed read/write, or buffer select  */
3026	if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3027		return -EINVAL;
3028
3029	if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3030		if (req->flags & REQ_F_BUFFER_SELECT) {
3031			buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3032			if (IS_ERR(buf))
3033				return PTR_ERR(buf);
3034			req->rw.len = sqe_len;
3035		}
3036
3037		ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3038		*iovec = NULL;
3039		return ret;
3040	}
3041
3042	if (req->flags & REQ_F_BUFFER_SELECT) {
3043		ret = io_iov_buffer_select(req, *iovec, needs_lock);
3044		if (!ret)
3045			iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3046		*iovec = NULL;
3047		return ret;
3048	}
3049
3050	return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3051			      req->ctx->compat);
3052}
3053
3054static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3055{
3056	return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3057}
3058
3059/*
3060 * For files that don't have ->read_iter() and ->write_iter(), handle them
3061 * by looping over ->read() or ->write() manually.
3062 */
3063static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3064{
3065	struct kiocb *kiocb = &req->rw.kiocb;
3066	struct file *file = req->file;
3067	ssize_t ret = 0;
3068
3069	/*
3070	 * Don't support polled IO through this interface, and we can't
3071	 * support non-blocking either. For the latter, this just causes
3072	 * the kiocb to be handled from an async context.
3073	 */
3074	if (kiocb->ki_flags & IOCB_HIPRI)
3075		return -EOPNOTSUPP;
3076	if (kiocb->ki_flags & IOCB_NOWAIT)
3077		return -EAGAIN;
3078
3079	while (iov_iter_count(iter)) {
3080		struct iovec iovec;
3081		ssize_t nr;
3082
3083		if (!iov_iter_is_bvec(iter)) {
3084			iovec = iov_iter_iovec(iter);
3085		} else {
3086			iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3087			iovec.iov_len = req->rw.len;
3088		}
3089
3090		if (rw == READ) {
3091			nr = file->f_op->read(file, iovec.iov_base,
3092					      iovec.iov_len, io_kiocb_ppos(kiocb));
3093		} else {
3094			nr = file->f_op->write(file, iovec.iov_base,
3095					       iovec.iov_len, io_kiocb_ppos(kiocb));
3096		}
3097
3098		if (nr < 0) {
3099			if (!ret)
3100				ret = nr;
3101			break;
3102		}
3103		ret += nr;
3104		if (nr != iovec.iov_len)
3105			break;
3106		req->rw.len -= nr;
3107		req->rw.addr += nr;
3108		iov_iter_advance(iter, nr);
3109	}
3110
3111	return ret;
3112}
3113
3114static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3115			  const struct iovec *fast_iov, struct iov_iter *iter)
3116{
3117	struct io_async_rw *rw = req->async_data;
3118
3119	memcpy(&rw->iter, iter, sizeof(*iter));
3120	rw->free_iovec = iovec;
3121	rw->bytes_done = 0;
3122	/* can only be fixed buffers, no need to do anything */
3123	if (iov_iter_is_bvec(iter))
3124		return;
3125	if (!iovec) {
3126		unsigned iov_off = 0;
3127
3128		rw->iter.iov = rw->fast_iov;
3129		if (iter->iov != fast_iov) {
3130			iov_off = iter->iov - fast_iov;
3131			rw->iter.iov += iov_off;
3132		}
3133		if (rw->fast_iov != fast_iov)
3134			memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3135			       sizeof(struct iovec) * iter->nr_segs);
3136	} else {
3137		req->flags |= REQ_F_NEED_CLEANUP;
3138	}
3139}
3140
3141static inline int io_alloc_async_data(struct io_kiocb *req)
3142{
3143	WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3144	req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3145	return req->async_data == NULL;
3146}
3147
3148static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3149			     const struct iovec *fast_iov,
3150			     struct iov_iter *iter, bool force)
3151{
3152	if (!force && !io_op_defs[req->opcode].needs_async_setup)
3153		return 0;
3154	if (!req->async_data) {
3155		if (io_alloc_async_data(req)) {
3156			kfree(iovec);
3157			return -ENOMEM;
3158		}
3159
3160		io_req_map_rw(req, iovec, fast_iov, iter);
3161	}
3162	return 0;
3163}
3164
3165static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3166{
3167	struct io_async_rw *iorw = req->async_data;
3168	struct iovec *iov = iorw->fast_iov;
3169	int ret;
3170
3171	ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3172	if (unlikely(ret < 0))
3173		return ret;
3174
3175	iorw->bytes_done = 0;
3176	iorw->free_iovec = iov;
3177	if (iov)
3178		req->flags |= REQ_F_NEED_CLEANUP;
3179	return 0;
3180}
3181
3182static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3183{
3184	if (unlikely(!(req->file->f_mode & FMODE_READ)))
3185		return -EBADF;
3186	return io_prep_rw(req, sqe);
3187}
3188
3189/*
3190 * This is our waitqueue callback handler, registered through lock_page_async()
3191 * when we initially tried to do the IO with the iocb armed our waitqueue.
3192 * This gets called when the page is unlocked, and we generally expect that to
3193 * happen when the page IO is completed and the page is now uptodate. This will
3194 * queue a task_work based retry of the operation, attempting to copy the data
3195 * again. If the latter fails because the page was NOT uptodate, then we will
3196 * do a thread based blocking retry of the operation. That's the unexpected
3197 * slow path.
3198 */
3199static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3200			     int sync, void *arg)
3201{
3202	struct wait_page_queue *wpq;
3203	struct io_kiocb *req = wait->private;
3204	struct wait_page_key *key = arg;
3205
3206	wpq = container_of(wait, struct wait_page_queue, wait);
3207
3208	if (!wake_page_match(wpq, key))
3209		return 0;
3210
3211	req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3212	list_del_init(&wait->entry);
3213
3214	/* submit ref gets dropped, acquire a new one */
3215	req_ref_get(req);
3216	io_req_task_queue(req);
3217	return 1;
3218}
3219
3220/*
3221 * This controls whether a given IO request should be armed for async page
3222 * based retry. If we return false here, the request is handed to the async
3223 * worker threads for retry. If we're doing buffered reads on a regular file,
3224 * we prepare a private wait_page_queue entry and retry the operation. This
3225 * will either succeed because the page is now uptodate and unlocked, or it
3226 * will register a callback when the page is unlocked at IO completion. Through
3227 * that callback, io_uring uses task_work to setup a retry of the operation.
3228 * That retry will attempt the buffered read again. The retry will generally
3229 * succeed, or in rare cases where it fails, we then fall back to using the
3230 * async worker threads for a blocking retry.
3231 */
3232static bool io_rw_should_retry(struct io_kiocb *req)
3233{
3234	struct io_async_rw *rw = req->async_data;
3235	struct wait_page_queue *wait = &rw->wpq;
3236	struct kiocb *kiocb = &req->rw.kiocb;
3237
3238	/* never retry for NOWAIT, we just complete with -EAGAIN */
3239	if (req->flags & REQ_F_NOWAIT)
3240		return false;
3241
3242	/* Only for buffered IO */
3243	if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3244		return false;
3245
3246	/*
3247	 * just use poll if we can, and don't attempt if the fs doesn't
3248	 * support callback based unlocks
3249	 */
3250	if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3251		return false;
3252
3253	wait->wait.func = io_async_buf_func;
3254	wait->wait.private = req;
3255	wait->wait.flags = 0;
3256	INIT_LIST_HEAD(&wait->wait.entry);
3257	kiocb->ki_flags |= IOCB_WAITQ;
3258	kiocb->ki_flags &= ~IOCB_NOWAIT;
3259	kiocb->ki_waitq = wait;
3260	return true;
3261}
3262
3263static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3264{
3265	if (req->file->f_op->read_iter)
3266		return call_read_iter(req->file, &req->rw.kiocb, iter);
3267	else if (req->file->f_op->read)
3268		return loop_rw_iter(READ, req, iter);
3269	else
3270		return -EINVAL;
3271}
3272
3273static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3274{
3275	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3276	struct kiocb *kiocb = &req->rw.kiocb;
3277	struct iov_iter __iter, *iter = &__iter;
3278	struct io_async_rw *rw = req->async_data;
3279	ssize_t io_size, ret, ret2;
3280	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3281
3282	if (rw) {
3283		iter = &rw->iter;
3284		iovec = NULL;
3285	} else {
3286		ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3287		if (ret < 0)
3288			return ret;
3289	}
3290	io_size = iov_iter_count(iter);
3291	req->result = io_size;
3292
3293	/* Ensure we clear previously set non-block flag */
3294	if (!force_nonblock)
3295		kiocb->ki_flags &= ~IOCB_NOWAIT;
3296	else
3297		kiocb->ki_flags |= IOCB_NOWAIT;
3298
3299	/* If the file doesn't support async, just async punt */
3300	if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3301		ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3302		return ret ?: -EAGAIN;
3303	}
3304
3305	ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3306	if (unlikely(ret)) {
3307		kfree(iovec);
3308		return ret;
3309	}
3310
3311	ret = io_iter_do_read(req, iter);
3312
3313	if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3314		req->flags &= ~REQ_F_REISSUE;
3315		/* IOPOLL retry should happen for io-wq threads */
3316		if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3317			goto done;
3318		/* no retry on NONBLOCK nor RWF_NOWAIT */
3319		if (req->flags & REQ_F_NOWAIT)
3320			goto done;
3321		/* some cases will consume bytes even on error returns */
3322		iov_iter_revert(iter, io_size - iov_iter_count(iter));
3323		ret = 0;
3324	} else if (ret == -EIOCBQUEUED) {
3325		goto out_free;
3326	} else if (ret <= 0 || ret == io_size || !force_nonblock ||
3327		   (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3328		/* read all, failed, already did sync or don't want to retry */
3329		goto done;
3330	}
3331
3332	ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3333	if (ret2)
3334		return ret2;
3335
3336	iovec = NULL;
3337	rw = req->async_data;
3338	/* now use our persistent iterator, if we aren't already */
3339	iter = &rw->iter;
3340
3341	do {
3342		io_size -= ret;
3343		rw->bytes_done += ret;
3344		/* if we can retry, do so with the callbacks armed */
3345		if (!io_rw_should_retry(req)) {
3346			kiocb->ki_flags &= ~IOCB_WAITQ;
3347			return -EAGAIN;
3348		}
3349
3350		/*
3351		 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3352		 * we get -EIOCBQUEUED, then we'll get a notification when the
3353		 * desired page gets unlocked. We can also get a partial read
3354		 * here, and if we do, then just retry at the new offset.
3355		 */
3356		ret = io_iter_do_read(req, iter);
3357		if (ret == -EIOCBQUEUED)
3358			return 0;
3359		/* we got some bytes, but not all. retry. */
3360		kiocb->ki_flags &= ~IOCB_WAITQ;
3361	} while (ret > 0 && ret < io_size);
3362done:
3363	kiocb_done(kiocb, ret, issue_flags);
3364out_free:
3365	/* it's faster to check here then delegate to kfree */
3366	if (iovec)
3367		kfree(iovec);
3368	return 0;
3369}
3370
3371static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3372{
3373	if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3374		return -EBADF;
3375	return io_prep_rw(req, sqe);
3376}
3377
3378static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3379{
3380	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3381	struct kiocb *kiocb = &req->rw.kiocb;
3382	struct iov_iter __iter, *iter = &__iter;
3383	struct io_async_rw *rw = req->async_data;
3384	ssize_t ret, ret2, io_size;
3385	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3386
3387	if (rw) {
3388		iter = &rw->iter;
3389		iovec = NULL;
3390	} else {
3391		ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3392		if (ret < 0)
3393			return ret;
3394	}
3395	io_size = iov_iter_count(iter);
3396	req->result = io_size;
3397
3398	/* Ensure we clear previously set non-block flag */
3399	if (!force_nonblock)
3400		kiocb->ki_flags &= ~IOCB_NOWAIT;
3401	else
3402		kiocb->ki_flags |= IOCB_NOWAIT;
3403
3404	/* If the file doesn't support async, just async punt */
3405	if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3406		goto copy_iov;
3407
3408	/* file path doesn't support NOWAIT for non-direct_IO */
3409	if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3410	    (req->flags & REQ_F_ISREG))
3411		goto copy_iov;
3412
3413	ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3414	if (unlikely(ret))
3415		goto out_free;
3416
3417	/*
3418	 * Open-code file_start_write here to grab freeze protection,
3419	 * which will be released by another thread in
3420	 * io_complete_rw().  Fool lockdep by telling it the lock got
3421	 * released so that it doesn't complain about the held lock when
3422	 * we return to userspace.
3423	 */
3424	if (req->flags & REQ_F_ISREG) {
3425		sb_start_write(file_inode(req->file)->i_sb);
3426		__sb_writers_release(file_inode(req->file)->i_sb,
3427					SB_FREEZE_WRITE);
3428	}
3429	kiocb->ki_flags |= IOCB_WRITE;
3430
3431	if (req->file->f_op->write_iter)
3432		ret2 = call_write_iter(req->file, kiocb, iter);
3433	else if (req->file->f_op->write)
3434		ret2 = loop_rw_iter(WRITE, req, iter);
3435	else
3436		ret2 = -EINVAL;
3437
3438	if (req->flags & REQ_F_REISSUE) {
3439		req->flags &= ~REQ_F_REISSUE;
3440		ret2 = -EAGAIN;
3441	}
3442
3443	/*
3444	 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3445	 * retry them without IOCB_NOWAIT.
3446	 */
3447	if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3448		ret2 = -EAGAIN;
3449	/* no retry on NONBLOCK nor RWF_NOWAIT */
3450	if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3451		goto done;
3452	if (!force_nonblock || ret2 != -EAGAIN) {
3453		/* IOPOLL retry should happen for io-wq threads */
3454		if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3455			goto copy_iov;
3456done:
3457		kiocb_done(kiocb, ret2, issue_flags);
3458	} else {
3459copy_iov:
3460		/* some cases will consume bytes even on error returns */
3461		iov_iter_revert(iter, io_size - iov_iter_count(iter));
3462		ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3463		return ret ?: -EAGAIN;
3464	}
3465out_free:
3466	/* it's reportedly faster than delegating the null check to kfree() */
3467	if (iovec)
3468		kfree(iovec);
3469	return ret;
3470}
3471
3472static int io_renameat_prep(struct io_kiocb *req,
3473			    const struct io_uring_sqe *sqe)
3474{
3475	struct io_rename *ren = &req->rename;
3476	const char __user *oldf, *newf;
3477
3478	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3479		return -EINVAL;
3480	if (sqe->ioprio || sqe->buf_index)
3481		return -EINVAL;
3482	if (unlikely(req->flags & REQ_F_FIXED_FILE))
3483		return -EBADF;
3484
3485	ren->old_dfd = READ_ONCE(sqe->fd);
3486	oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3487	newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3488	ren->new_dfd = READ_ONCE(sqe->len);
3489	ren->flags = READ_ONCE(sqe->rename_flags);
3490
3491	ren->oldpath = getname(oldf);
3492	if (IS_ERR(ren->oldpath))
3493		return PTR_ERR(ren->oldpath);
3494
3495	ren->newpath = getname(newf);
3496	if (IS_ERR(ren->newpath)) {
3497		putname(ren->oldpath);
3498		return PTR_ERR(ren->newpath);
3499	}
3500
3501	req->flags |= REQ_F_NEED_CLEANUP;
3502	return 0;
3503}
3504
3505static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3506{
3507	struct io_rename *ren = &req->rename;
3508	int ret;
3509
3510	if (issue_flags & IO_URING_F_NONBLOCK)
3511		return -EAGAIN;
3512
3513	ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3514				ren->newpath, ren->flags);
3515
3516	req->flags &= ~REQ_F_NEED_CLEANUP;
3517	if (ret < 0)
3518		req_set_fail(req);
3519	io_req_complete(req, ret);
3520	return 0;
3521}
3522
3523static int io_unlinkat_prep(struct io_kiocb *req,
3524			    const struct io_uring_sqe *sqe)
3525{
3526	struct io_unlink *un = &req->unlink;
3527	const char __user *fname;
3528
3529	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3530		return -EINVAL;
3531	if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3532		return -EINVAL;
3533	if (unlikely(req->flags & REQ_F_FIXED_FILE))
3534		return -EBADF;
3535
3536	un->dfd = READ_ONCE(sqe->fd);
3537
3538	un->flags = READ_ONCE(sqe->unlink_flags);
3539	if (un->flags & ~AT_REMOVEDIR)
3540		return -EINVAL;
3541
3542	fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3543	un->filename = getname(fname);
3544	if (IS_ERR(un->filename))
3545		return PTR_ERR(un->filename);
3546
3547	req->flags |= REQ_F_NEED_CLEANUP;
3548	return 0;
3549}
3550
3551static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3552{
3553	struct io_unlink *un = &req->unlink;
3554	int ret;
3555
3556	if (issue_flags & IO_URING_F_NONBLOCK)
3557		return -EAGAIN;
3558
3559	if (un->flags & AT_REMOVEDIR)
3560		ret = do_rmdir(un->dfd, un->filename);
3561	else
3562		ret = do_unlinkat(un->dfd, un->filename);
3563
3564	req->flags &= ~REQ_F_NEED_CLEANUP;
3565	if (ret < 0)
3566		req_set_fail(req);
3567	io_req_complete(req, ret);
3568	return 0;
3569}
3570
3571static int io_shutdown_prep(struct io_kiocb *req,
3572			    const struct io_uring_sqe *sqe)
3573{
3574#if defined(CONFIG_NET)
3575	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3576		return -EINVAL;
3577	if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3578	    sqe->buf_index)
3579		return -EINVAL;
3580
3581	req->shutdown.how = READ_ONCE(sqe->len);
3582	return 0;
3583#else
3584	return -EOPNOTSUPP;
3585#endif
3586}
3587
3588static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3589{
3590#if defined(CONFIG_NET)
3591	struct socket *sock;
3592	int ret;
3593
3594	if (issue_flags & IO_URING_F_NONBLOCK)
3595		return -EAGAIN;
3596
3597	sock = sock_from_file(req->file);
3598	if (unlikely(!sock))
3599		return -ENOTSOCK;
3600
3601	ret = __sys_shutdown_sock(sock, req->shutdown.how);
3602	if (ret < 0)
3603		req_set_fail(req);
3604	io_req_complete(req, ret);
3605	return 0;
3606#else
3607	return -EOPNOTSUPP;
3608#endif
3609}
3610
3611static int __io_splice_prep(struct io_kiocb *req,
3612			    const struct io_uring_sqe *sqe)
3613{
3614	struct io_splice *sp = &req->splice;
3615	unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3616
3617	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3618		return -EINVAL;
3619
3620	sp->file_in = NULL;
3621	sp->len = READ_ONCE(sqe->len);
3622	sp->flags = READ_ONCE(sqe->splice_flags);
3623
3624	if (unlikely(sp->flags & ~valid_flags))
3625		return -EINVAL;
3626
3627	sp->file_in = io_file_get(req->ctx, NULL, req,
3628				  READ_ONCE(sqe->splice_fd_in),
3629				  (sp->flags & SPLICE_F_FD_IN_FIXED));
3630	if (!sp->file_in)
3631		return -EBADF;
3632	req->flags |= REQ_F_NEED_CLEANUP;
3633	return 0;
3634}
3635
3636static int io_tee_prep(struct io_kiocb *req,
3637		       const struct io_uring_sqe *sqe)
3638{
3639	if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3640		return -EINVAL;
3641	return __io_splice_prep(req, sqe);
3642}
3643
3644static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3645{
3646	struct io_splice *sp = &req->splice;
3647	struct file *in = sp->file_in;
3648	struct file *out = sp->file_out;
3649	unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3650	long ret = 0;
3651
3652	if (issue_flags & IO_URING_F_NONBLOCK)
3653		return -EAGAIN;
3654	if (sp->len)
3655		ret = do_tee(in, out, sp->len, flags);
3656
3657	if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3658		io_put_file(in);
3659	req->flags &= ~REQ_F_NEED_CLEANUP;
3660
3661	if (ret != sp->len)
3662		req_set_fail(req);
3663	io_req_complete(req, ret);
3664	return 0;
3665}
3666
3667static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3668{
3669	struct io_splice *sp = &req->splice;
3670
3671	sp->off_in = READ_ONCE(sqe->splice_off_in);
3672	sp->off_out = READ_ONCE(sqe->off);
3673	return __io_splice_prep(req, sqe);
3674}
3675
3676static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3677{
3678	struct io_splice *sp = &req->splice;
3679	struct file *in = sp->file_in;
3680	struct file *out = sp->file_out;
3681	unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3682	loff_t *poff_in, *poff_out;
3683	long ret = 0;
3684
3685	if (issue_flags & IO_URING_F_NONBLOCK)
3686		return -EAGAIN;
3687
3688	poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3689	poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3690
3691	if (sp->len)
3692		ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3693
3694	if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3695		io_put_file(in);
3696	req->flags &= ~REQ_F_NEED_CLEANUP;
3697
3698	if (ret != sp->len)
3699		req_set_fail(req);
3700	io_req_complete(req, ret);
3701	return 0;
3702}
3703
3704/*
3705 * IORING_OP_NOP just posts a completion event, nothing else.
3706 */
3707static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3708{
3709	struct io_ring_ctx *ctx = req->ctx;
3710
3711	if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3712		return -EINVAL;
3713
3714	__io_req_complete(req, issue_flags, 0, 0);
3715	return 0;
3716}
3717
3718static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3719{
3720	struct io_ring_ctx *ctx = req->ctx;
3721
3722	if (!req->file)
3723		return -EBADF;
3724
3725	if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3726		return -EINVAL;
3727	if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3728		return -EINVAL;
3729
3730	req->sync.flags = READ_ONCE(sqe->fsync_flags);
3731	if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3732		return -EINVAL;
3733
3734	req->sync.off = READ_ONCE(sqe->off);
3735	req->sync.len = READ_ONCE(sqe->len);
3736	return 0;
3737}
3738
3739static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3740{
3741	loff_t end = req->sync.off + req->sync.len;
3742	int ret;
3743
3744	/* fsync always requires a blocking context */
3745	if (issue_flags & IO_URING_F_NONBLOCK)
3746		return -EAGAIN;
3747
3748	ret = vfs_fsync_range(req->file, req->sync.off,
3749				end > 0 ? end : LLONG_MAX,
3750				req->sync.flags & IORING_FSYNC_DATASYNC);
3751	if (ret < 0)
3752		req_set_fail(req);
3753	io_req_complete(req, ret);
3754	return 0;
3755}
3756
3757static int io_fallocate_prep(struct io_kiocb *req,
3758			     const struct io_uring_sqe *sqe)
3759{
3760	if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3761		return -EINVAL;
3762	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3763		return -EINVAL;
3764
3765	req->sync.off = READ_ONCE(sqe->off);
3766	req->sync.len = READ_ONCE(sqe->addr);
3767	req->sync.mode = READ_ONCE(sqe->len);
3768	return 0;
3769}
3770
3771static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3772{
3773	int ret;
3774
3775	/* fallocate always requiring blocking context */
3776	if (issue_flags & IO_URING_F_NONBLOCK)
3777		return -EAGAIN;
3778	ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3779				req->sync.len);
3780	if (ret < 0)
3781		req_set_fail(req);
3782	io_req_complete(req, ret);
3783	return 0;
3784}
3785
3786static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3787{
3788	const char __user *fname;
3789	int ret;
3790
3791	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3792		return -EINVAL;
3793	if (unlikely(sqe->ioprio || sqe->buf_index))
3794		return -EINVAL;
3795	if (unlikely(req->flags & REQ_F_FIXED_FILE))
3796		return -EBADF;
3797
3798	/* open.how should be already initialised */
3799	if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3800		req->open.how.flags |= O_LARGEFILE;
3801
3802	req->open.dfd = READ_ONCE(sqe->fd);
3803	fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3804	req->open.filename = getname(fname);
3805	if (IS_ERR(req->open.filename)) {
3806		ret = PTR_ERR(req->open.filename);
3807		req->open.filename = NULL;
3808		return ret;
3809	}
3810	req->open.nofile = rlimit(RLIMIT_NOFILE);
3811	req->flags |= REQ_F_NEED_CLEANUP;
3812	return 0;
3813}
3814
3815static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3816{
3817	u64 mode = READ_ONCE(sqe->len);
3818	u64 flags = READ_ONCE(sqe->open_flags);
3819
3820	req->open.how = build_open_how(flags, mode);
3821	return __io_openat_prep(req, sqe);
3822}
3823
3824static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3825{
3826	struct open_how __user *how;
3827	size_t len;
3828	int ret;
3829
3830	how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3831	len = READ_ONCE(sqe->len);
3832	if (len < OPEN_HOW_SIZE_VER0)
3833		return -EINVAL;
3834
3835	ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3836					len);
3837	if (ret)
3838		return ret;
3839
3840	return __io_openat_prep(req, sqe);
3841}
3842
3843static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3844{
3845	struct open_flags op;
3846	struct file *file;
3847	bool nonblock_set;
3848	bool resolve_nonblock;
3849	int ret;
3850
3851	ret = build_open_flags(&req->open.how, &op);
3852	if (ret)
3853		goto err;
3854	nonblock_set = op.open_flag & O_NONBLOCK;
3855	resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3856	if (issue_flags & IO_URING_F_NONBLOCK) {
3857		/*
3858		 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3859		 * it'll always -EAGAIN
3860		 */
3861		if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3862			return -EAGAIN;
3863		op.lookup_flags |= LOOKUP_CACHED;
3864		op.open_flag |= O_NONBLOCK;
3865	}
3866
3867	ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3868	if (ret < 0)
3869		goto err;
3870
3871	file = do_filp_open(req->open.dfd, req->open.filename, &op);
3872	if (IS_ERR(file)) {
3873		/*
3874		 * We could hang on to this 'fd' on retrying, but seems like
3875		 * marginal gain for something that is now known to be a slower
3876		 * path. So just put it, and we'll get a new one when we retry.
3877		 */
3878		put_unused_fd(ret);
3879
3880		ret = PTR_ERR(file);
3881		/* only retry if RESOLVE_CACHED wasn't already set by application */
3882		if (ret == -EAGAIN &&
3883		    (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3884			return -EAGAIN;
3885		goto err;
3886	}
3887
3888	if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3889		file->f_flags &= ~O_NONBLOCK;
3890	fsnotify_open(file);
3891	fd_install(ret, file);
3892err:
3893	putname(req->open.filename);
3894	req->flags &= ~REQ_F_NEED_CLEANUP;
3895	if (ret < 0)
3896		req_set_fail(req);
3897	__io_req_complete(req, issue_flags, ret, 0);
3898	return 0;
3899}
3900
3901static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3902{
3903	return io_openat2(req, issue_flags);
3904}
3905
3906static int io_remove_buffers_prep(struct io_kiocb *req,
3907				  const struct io_uring_sqe *sqe)
3908{
3909	struct io_provide_buf *p = &req->pbuf;
3910	u64 tmp;
3911
3912	if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3913		return -EINVAL;
3914
3915	tmp = READ_ONCE(sqe->fd);
3916	if (!tmp || tmp > USHRT_MAX)
3917		return -EINVAL;
3918
3919	memset(p, 0, sizeof(*p));
3920	p->nbufs = tmp;
3921	p->bgid = READ_ONCE(sqe->buf_group);
3922	return 0;
3923}
3924
3925static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3926			       int bgid, unsigned nbufs)
3927{
3928	unsigned i = 0;
3929
3930	/* shouldn't happen */
3931	if (!nbufs)
3932		return 0;
3933
3934	/* the head kbuf is the list itself */
3935	while (!list_empty(&buf->list)) {
3936		struct io_buffer *nxt;
3937
3938		nxt = list_first_entry(&buf->list, struct io_buffer, list);
3939		list_del(&nxt->list);
3940		kfree(nxt);
3941		if (++i == nbufs)
3942			return i;
3943	}
3944	i++;
3945	kfree(buf);
3946	xa_erase(&ctx->io_buffers, bgid);
3947
3948	return i;
3949}
3950
3951static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3952{
3953	struct io_provide_buf *p = &req->pbuf;
3954	struct io_ring_ctx *ctx = req->ctx;
3955	struct io_buffer *head;
3956	int ret = 0;
3957	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3958
3959	io_ring_submit_lock(ctx, !force_nonblock);
3960
3961	lockdep_assert_held(&ctx->uring_lock);
3962
3963	ret = -ENOENT;
3964	head = xa_load(&ctx->io_buffers, p->bgid);
3965	if (head)
3966		ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3967	if (ret < 0)
3968		req_set_fail(req);
3969
3970	/* complete before unlock, IOPOLL may need the lock */
3971	__io_req_complete(req, issue_flags, ret, 0);
3972	io_ring_submit_unlock(ctx, !force_nonblock);
3973	return 0;
3974}
3975
3976static int io_provide_buffers_prep(struct io_kiocb *req,
3977				   const struct io_uring_sqe *sqe)
3978{
3979	unsigned long size, tmp_check;
3980	struct io_provide_buf *p = &req->pbuf;
3981	u64 tmp;
3982
3983	if (sqe->ioprio || sqe->rw_flags)
3984		return -EINVAL;
3985
3986	tmp = READ_ONCE(sqe->fd);
3987	if (!tmp || tmp > USHRT_MAX)
3988		return -E2BIG;
3989	p->nbufs = tmp;
3990	p->addr = READ_ONCE(sqe->addr);
3991	p->len = READ_ONCE(sqe->len);
3992
3993	if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
3994				&size))
3995		return -EOVERFLOW;
3996	if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
3997		return -EOVERFLOW;
3998
3999	size = (unsigned long)p->len * p->nbufs;
4000	if (!access_ok(u64_to_user_ptr(p->addr), size))
4001		return -EFAULT;
4002
4003	p->bgid = READ_ONCE(sqe->buf_group);
4004	tmp = READ_ONCE(sqe->off);
4005	if (tmp > USHRT_MAX)
4006		return -E2BIG;
4007	p->bid = tmp;
4008	return 0;
4009}
4010
4011static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4012{
4013	struct io_buffer *buf;
4014	u64 addr = pbuf->addr;
4015	int i, bid = pbuf->bid;
4016
4017	for (i = 0; i < pbuf->nbufs; i++) {
4018		buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4019		if (!buf)
4020			break;
4021
4022		buf->addr = addr;
4023		buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4024		buf->bid = bid;
4025		addr += pbuf->len;
4026		bid++;
4027		if (!*head) {
4028			INIT_LIST_HEAD(&buf->list);
4029			*head = buf;
4030		} else {
4031			list_add_tail(&buf->list, &(*head)->list);
4032		}
4033	}
4034
4035	return i ? i : -ENOMEM;
4036}
4037
4038static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4039{
4040	struct io_provide_buf *p = &req->pbuf;
4041	struct io_ring_ctx *ctx = req->ctx;
4042	struct io_buffer *head, *list;
4043	int ret = 0;
4044	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4045
4046	io_ring_submit_lock(ctx, !force_nonblock);
4047
4048	lockdep_assert_held(&ctx->uring_lock);
4049
4050	list = head = xa_load(&ctx->io_buffers, p->bgid);
4051
4052	ret = io_add_buffers(p, &head);
4053	if (ret >= 0 && !list) {
4054		ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4055		if (ret < 0)
4056			__io_remove_buffers(ctx, head, p->bgid, -1U);
4057	}
4058	if (ret < 0)
4059		req_set_fail(req);
4060	/* complete before unlock, IOPOLL may need the lock */
4061	__io_req_complete(req, issue_flags, ret, 0);
4062	io_ring_submit_unlock(ctx, !force_nonblock);
4063	return 0;
4064}
4065
4066static int io_epoll_ctl_prep(struct io_kiocb *req,
4067			     const struct io_uring_sqe *sqe)
4068{
4069#if defined(CONFIG_EPOLL)
4070	if (sqe->ioprio || sqe->buf_index)
4071		return -EINVAL;
4072	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4073		return -EINVAL;
4074
4075	req->epoll.epfd = READ_ONCE(sqe->fd);
4076	req->epoll.op = READ_ONCE(sqe->len);
4077	req->epoll.fd = READ_ONCE(sqe->off);
4078
4079	if (ep_op_has_event(req->epoll.op)) {
4080		struct epoll_event __user *ev;
4081
4082		ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4083		if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4084			return -EFAULT;
4085	}
4086
4087	return 0;
4088#else
4089	return -EOPNOTSUPP;
4090#endif
4091}
4092
4093static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4094{
4095#if defined(CONFIG_EPOLL)
4096	struct io_epoll *ie = &req->epoll;
4097	int ret;
4098	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4099
4100	ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4101	if (force_nonblock && ret == -EAGAIN)
4102		return -EAGAIN;
4103
4104	if (ret < 0)
4105		req_set_fail(req);
4106	__io_req_complete(req, issue_flags, ret, 0);
4107	return 0;
4108#else
4109	return -EOPNOTSUPP;
4110#endif
4111}
4112
4113static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4114{
4115#if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4116	if (sqe->ioprio || sqe->buf_index || sqe->off)
4117		return -EINVAL;
4118	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4119		return -EINVAL;
4120
4121	req->madvise.addr = READ_ONCE(sqe->addr);
4122	req->madvise.len = READ_ONCE(sqe->len);
4123	req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4124	return 0;
4125#else
4126	return -EOPNOTSUPP;
4127#endif
4128}
4129
4130static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4131{
4132#if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4133	struct io_madvise *ma = &req->madvise;
4134	int ret;
4135
4136	if (issue_flags & IO_URING_F_NONBLOCK)
4137		return -EAGAIN;
4138
4139	ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4140	if (ret < 0)
4141		req_set_fail(req);
4142	io_req_complete(req, ret);
4143	return 0;
4144#else
4145	return -EOPNOTSUPP;
4146#endif
4147}
4148
4149static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4150{
4151	if (sqe->ioprio || sqe->buf_index || sqe->addr)
4152		return -EINVAL;
4153	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4154		return -EINVAL;
4155
4156	req->fadvise.offset = READ_ONCE(sqe->off);
4157	req->fadvise.len = READ_ONCE(sqe->len);
4158	req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4159	return 0;
4160}
4161
4162static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4163{
4164	struct io_fadvise *fa = &req->fadvise;
4165	int ret;
4166
4167	if (issue_flags & IO_URING_F_NONBLOCK) {
4168		switch (fa->advice) {
4169		case POSIX_FADV_NORMAL:
4170		case POSIX_FADV_RANDOM:
4171		case POSIX_FADV_SEQUENTIAL:
4172			break;
4173		default:
4174			return -EAGAIN;
4175		}
4176	}
4177
4178	ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4179	if (ret < 0)
4180		req_set_fail(req);
4181	__io_req_complete(req, issue_flags, ret, 0);
4182	return 0;
4183}
4184
4185static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4186{
4187	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4188		return -EINVAL;
4189	if (sqe->ioprio || sqe->buf_index)
4190		return -EINVAL;
4191	if (req->flags & REQ_F_FIXED_FILE)
4192		return -EBADF;
4193
4194	req->statx.dfd = READ_ONCE(sqe->fd);
4195	req->statx.mask = READ_ONCE(sqe->len);
4196	req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4197	req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4198	req->statx.flags = READ_ONCE(sqe->statx_flags);
4199
4200	return 0;
4201}
4202
4203static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4204{
4205	struct io_statx *ctx = &req->statx;
4206	int ret;
4207
4208	if (issue_flags & IO_URING_F_NONBLOCK)
4209		return -EAGAIN;
4210
4211	ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4212		       ctx->buffer);
4213
4214	if (ret < 0)
4215		req_set_fail(req);
4216	io_req_complete(req, ret);
4217	return 0;
4218}
4219
4220static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4221{
4222	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4223		return -EINVAL;
4224	if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4225	    sqe->rw_flags || sqe->buf_index)
4226		return -EINVAL;
4227	if (req->flags & REQ_F_FIXED_FILE)
4228		return -EBADF;
4229
4230	req->close.fd = READ_ONCE(sqe->fd);
4231	return 0;
4232}
4233
4234static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4235{
4236	struct files_struct *files = current->files;
4237	struct io_close *close = &req->close;
4238	struct fdtable *fdt;
4239	struct file *file = NULL;
4240	int ret = -EBADF;
4241
4242	spin_lock(&files->file_lock);
4243	fdt = files_fdtable(files);
4244	if (close->fd >= fdt->max_fds) {
4245		spin_unlock(&files->file_lock);
4246		goto err;
4247	}
4248	file = fdt->fd[close->fd];
4249	if (!file || file->f_op == &io_uring_fops) {
4250		spin_unlock(&files->file_lock);
4251		file = NULL;
4252		goto err;
4253	}
4254
4255	/* if the file has a flush method, be safe and punt to async */
4256	if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4257		spin_unlock(&files->file_lock);
4258		return -EAGAIN;
4259	}
4260
4261	ret = __close_fd_get_file(close->fd, &file);
4262	spin_unlock(&files->file_lock);
4263	if (ret < 0) {
4264		if (ret == -ENOENT)
4265			ret = -EBADF;
4266		goto err;
4267	}
4268
4269	/* No ->flush() or already async, safely close from here */
4270	ret = filp_close(file, current->files);
4271err:
4272	if (ret < 0)
4273		req_set_fail(req);
4274	if (file)
4275		fput(file);
4276	__io_req_complete(req, issue_flags, ret, 0);
4277	return 0;
4278}
4279
4280static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4281{
4282	struct io_ring_ctx *ctx = req->ctx;
4283
4284	if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4285		return -EINVAL;
4286	if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4287		return -EINVAL;
4288
4289	req->sync.off = READ_ONCE(sqe->off);
4290	req->sync.len = READ_ONCE(sqe->len);
4291	req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4292	return 0;
4293}
4294
4295static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4296{
4297	int ret;
4298
4299	/* sync_file_range always requires a blocking context */
4300	if (issue_flags & IO_URING_F_NONBLOCK)
4301		return -EAGAIN;
4302
4303	ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4304				req->sync.flags);
4305	if (ret < 0)
4306		req_set_fail(req);
4307	io_req_complete(req, ret);
4308	return 0;
4309}
4310
4311#if defined(CONFIG_NET)
4312static int io_setup_async_msg(struct io_kiocb *req,
4313			      struct io_async_msghdr *kmsg)
4314{
4315	struct io_async_msghdr *async_msg = req->async_data;
4316
4317	if (async_msg)
4318		return -EAGAIN;
4319	if (io_alloc_async_data(req)) {
4320		kfree(kmsg->free_iov);
4321		return -ENOMEM;
4322	}
4323	async_msg = req->async_data;
4324	req->flags |= REQ_F_NEED_CLEANUP;
4325	memcpy(async_msg, kmsg, sizeof(*kmsg));
4326	async_msg->msg.msg_name = &async_msg->addr;
4327	/* if were using fast_iov, set it to the new one */
4328	if (!async_msg->free_iov)
4329		async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4330
4331	return -EAGAIN;
4332}
4333
4334static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4335			       struct io_async_msghdr *iomsg)
4336{
4337	iomsg->msg.msg_name = &iomsg->addr;
4338	iomsg->free_iov = iomsg->fast_iov;
4339	return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4340				   req->sr_msg.msg_flags, &iomsg->free_iov);
4341}
4342
4343static int io_sendmsg_prep_async(struct io_kiocb *req)
4344{
4345	int ret;
4346
4347	ret = io_sendmsg_copy_hdr(req, req->async_data);
4348	if (!ret)
4349		req->flags |= REQ_F_NEED_CLEANUP;
4350	return ret;
4351}
4352
4353static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4354{
4355	struct io_sr_msg *sr = &req->sr_msg;
4356
4357	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4358		return -EINVAL;
4359
4360	sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4361	sr->len = READ_ONCE(sqe->len);
4362	sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4363	if (sr->msg_flags & MSG_DONTWAIT)
4364		req->flags |= REQ_F_NOWAIT;
4365
4366#ifdef CONFIG_COMPAT
4367	if (req->ctx->compat)
4368		sr->msg_flags |= MSG_CMSG_COMPAT;
4369#endif
4370	return 0;
4371}
4372
4373static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4374{
4375	struct io_async_msghdr iomsg, *kmsg;
4376	struct socket *sock;
4377	unsigned flags;
4378	int min_ret = 0;
4379	int ret;
4380
4381	sock = sock_from_file(req->file);
4382	if (unlikely(!sock))
4383		return -ENOTSOCK;
4384
4385	kmsg = req->async_data;
4386	if (!kmsg) {
4387		ret = io_sendmsg_copy_hdr(req, &iomsg);
4388		if (ret)
4389			return ret;
4390		kmsg = &iomsg;
4391	}
4392
4393	flags = req->sr_msg.msg_flags;
4394	if (issue_flags & IO_URING_F_NONBLOCK)
4395		flags |= MSG_DONTWAIT;
4396	if (flags & MSG_WAITALL)
4397		min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4398
4399	ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4400	if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4401		return io_setup_async_msg(req, kmsg);
4402	if (ret == -ERESTARTSYS)
4403		ret = -EINTR;
4404
4405	/* fast path, check for non-NULL to avoid function call */
4406	if (kmsg->free_iov)
4407		kfree(kmsg->free_iov);
4408	req->flags &= ~REQ_F_NEED_CLEANUP;
4409	if (ret < min_ret)
4410		req_set_fail(req);
4411	__io_req_complete(req, issue_flags, ret, 0);
4412	return 0;
4413}
4414
4415static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4416{
4417	struct io_sr_msg *sr = &req->sr_msg;
4418	struct msghdr msg;
4419	struct iovec iov;
4420	struct socket *sock;
4421	unsigned flags;
4422	int min_ret = 0;
4423	int ret;
4424
4425	sock = sock_from_file(req->file);
4426	if (unlikely(!sock))
4427		return -ENOTSOCK;
4428
4429	ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4430	if (unlikely(ret))
4431		return ret;
4432
4433	msg.msg_name = NULL;
4434	msg.msg_control = NULL;
4435	msg.msg_controllen = 0;
4436	msg.msg_namelen = 0;
4437
4438	flags = req->sr_msg.msg_flags;
4439	if (issue_flags & IO_URING_F_NONBLOCK)
4440		flags |= MSG_DONTWAIT;
4441	if (flags & MSG_WAITALL)
4442		min_ret = iov_iter_count(&msg.msg_iter);
4443
4444	msg.msg_flags = flags;
4445	ret = sock_sendmsg(sock, &msg);
4446	if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4447		return -EAGAIN;
4448	if (ret == -ERESTARTSYS)
4449		ret = -EINTR;
4450
4451	if (ret < min_ret)
4452		req_set_fail(req);
4453	__io_req_complete(req, issue_flags, ret, 0);
4454	return 0;
4455}
4456
4457static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4458				 struct io_async_msghdr *iomsg)
4459{
4460	struct io_sr_msg *sr = &req->sr_msg;
4461	struct iovec __user *uiov;
4462	size_t iov_len;
4463	int ret;
4464
4465	ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4466					&iomsg->uaddr, &uiov, &iov_len);
4467	if (ret)
4468		return ret;
4469
4470	if (req->flags & REQ_F_BUFFER_SELECT) {
4471		if (iov_len > 1)
4472			return -EINVAL;
4473		if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4474			return -EFAULT;
4475		sr->len = iomsg->fast_iov[0].iov_len;
4476		iomsg->free_iov = NULL;
4477	} else {
4478		iomsg->free_iov = iomsg->fast_iov;
4479		ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4480				     &iomsg->free_iov, &iomsg->msg.msg_iter,
4481				     false);
4482		if (ret > 0)
4483			ret = 0;
4484	}
4485
4486	return ret;
4487}
4488
4489#ifdef CONFIG_COMPAT
4490static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4491					struct io_async_msghdr *iomsg)
4492{
4493	struct io_sr_msg *sr = &req->sr_msg;
4494	struct compat_iovec __user *uiov;
4495	compat_uptr_t ptr;
4496	compat_size_t len;
4497	int ret;
4498
4499	ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4500				  &ptr, &len);
4501	if (ret)
4502		return ret;
4503
4504	uiov = compat_ptr(ptr);
4505	if (req->flags & REQ_F_BUFFER_SELECT) {
4506		compat_ssize_t clen;
4507
4508		if (len > 1)
4509			return -EINVAL;
4510		if (!access_ok(uiov, sizeof(*uiov)))
4511			return -EFAULT;
4512		if (__get_user(clen, &uiov->iov_len))
4513			return -EFAULT;
4514		if (clen < 0)
4515			return -EINVAL;
4516		sr->len = clen;
4517		iomsg->free_iov = NULL;
4518	} else {
4519		iomsg->free_iov = iomsg->fast_iov;
4520		ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4521				   UIO_FASTIOV, &iomsg->free_iov,
4522				   &iomsg->msg.msg_iter, true);
4523		if (ret < 0)
4524			return ret;
4525	}
4526
4527	return 0;
4528}
4529#endif
4530
4531static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4532			       struct io_async_msghdr *iomsg)
4533{
4534	iomsg->msg.msg_name = &iomsg->addr;
4535
4536#ifdef CONFIG_COMPAT
4537	if (req->ctx->compat)
4538		return __io_compat_recvmsg_copy_hdr(req, iomsg);
4539#endif
4540
4541	return __io_recvmsg_copy_hdr(req, iomsg);
4542}
4543
4544static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4545					       bool needs_lock)
4546{
4547	struct io_sr_msg *sr = &req->sr_msg;
4548	struct io_buffer *kbuf;
4549
4550	kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4551	if (IS_ERR(kbuf))
4552		return kbuf;
4553
4554	sr->kbuf = kbuf;
4555	req->flags |= REQ_F_BUFFER_SELECTED;
4556	return kbuf;
4557}
4558
4559static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4560{
4561	return io_put_kbuf(req, req->sr_msg.kbuf);
4562}
4563
4564static int io_recvmsg_prep_async(struct io_kiocb *req)
4565{
4566	int ret;
4567
4568	ret = io_recvmsg_copy_hdr(req, req->async_data);
4569	if (!ret)
4570		req->flags |= REQ_F_NEED_CLEANUP;
4571	return ret;
4572}
4573
4574static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4575{
4576	struct io_sr_msg *sr = &req->sr_msg;
4577
4578	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4579		return -EINVAL;
4580
4581	sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4582	sr->len = READ_ONCE(sqe->len);
4583	sr->bgid = READ_ONCE(sqe->buf_group);
4584	sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4585	if (sr->msg_flags & MSG_DONTWAIT)
4586		req->flags |= REQ_F_NOWAIT;
4587
4588#ifdef CONFIG_COMPAT
4589	if (req->ctx->compat)
4590		sr->msg_flags |= MSG_CMSG_COMPAT;
4591#endif
4592	return 0;
4593}
4594
4595static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4596{
4597	struct io_async_msghdr iomsg, *kmsg;
4598	struct socket *sock;
4599	struct io_buffer *kbuf;
4600	unsigned flags;
4601	int min_ret = 0;
4602	int ret, cflags = 0;
4603	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4604
4605	sock = sock_from_file(req->file);
4606	if (unlikely(!sock))
4607		return -ENOTSOCK;
4608
4609	kmsg = req->async_data;
4610	if (!kmsg) {
4611		ret = io_recvmsg_copy_hdr(req, &iomsg);
4612		if (ret)
4613			return ret;
4614		kmsg = &iomsg;
4615	}
4616
4617	if (req->flags & REQ_F_BUFFER_SELECT) {
4618		kbuf = io_recv_buffer_select(req, !force_nonblock);
4619		if (IS_ERR(kbuf))
4620			return PTR_ERR(kbuf);
4621		kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4622		kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4623		iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4624				1, req->sr_msg.len);
4625	}
4626
4627	flags = req->sr_msg.msg_flags;
4628	if (force_nonblock)
4629		flags |= MSG_DONTWAIT;
4630	if (flags & MSG_WAITALL)
4631		min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4632
4633	ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4634					kmsg->uaddr, flags);
4635	if (force_nonblock && ret == -EAGAIN)
4636		return io_setup_async_msg(req, kmsg);
4637	if (ret == -ERESTARTSYS)
4638		ret = -EINTR;
4639
4640	if (req->flags & REQ_F_BUFFER_SELECTED)
4641		cflags = io_put_recv_kbuf(req);
4642	/* fast path, check for non-NULL to avoid function call */
4643	if (kmsg->free_iov)
4644		kfree(kmsg->free_iov);
4645	req->flags &= ~REQ_F_NEED_CLEANUP;
4646	if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4647		req_set_fail(req);
4648	__io_req_complete(req, issue_flags, ret, cflags);
4649	return 0;
4650}
4651
4652static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4653{
4654	struct io_buffer *kbuf;
4655	struct io_sr_msg *sr = &req->sr_msg;
4656	struct msghdr msg;
4657	void __user *buf = sr->buf;
4658	struct socket *sock;
4659	struct iovec iov;
4660	unsigned flags;
4661	int min_ret = 0;
4662	int ret, cflags = 0;
4663	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4664
4665	sock = sock_from_file(req->file);
4666	if (unlikely(!sock))
4667		return -ENOTSOCK;
4668
4669	if (req->flags & REQ_F_BUFFER_SELECT) {
4670		kbuf = io_recv_buffer_select(req, !force_nonblock);
4671		if (IS_ERR(kbuf))
4672			return PTR_ERR(kbuf);
4673		buf = u64_to_user_ptr(kbuf->addr);
4674	}
4675
4676	ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4677	if (unlikely(ret))
4678		goto out_free;
4679
4680	msg.msg_name = NULL;
4681	msg.msg_control = NULL;
4682	msg.msg_controllen = 0;
4683	msg.msg_namelen = 0;
4684	msg.msg_iocb = NULL;
4685	msg.msg_flags = 0;
4686
4687	flags = req->sr_msg.msg_flags;
4688	if (force_nonblock)
4689		flags |= MSG_DONTWAIT;
4690	if (flags & MSG_WAITALL)
4691		min_ret = iov_iter_count(&msg.msg_iter);
4692
4693	ret = sock_recvmsg(sock, &msg, flags);
4694	if (force_nonblock && ret == -EAGAIN)
4695		return -EAGAIN;
4696	if (ret == -ERESTARTSYS)
4697		ret = -EINTR;
4698out_free:
4699	if (req->flags & REQ_F_BUFFER_SELECTED)
4700		cflags = io_put_recv_kbuf(req);
4701	if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4702		req_set_fail(req);
4703	__io_req_complete(req, issue_flags, ret, cflags);
4704	return 0;
4705}
4706
4707static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4708{
4709	struct io_accept *accept = &req->accept;
4710
4711	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4712		return -EINVAL;
4713	if (sqe->ioprio || sqe->len || sqe->buf_index)
4714		return -EINVAL;
4715
4716	accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4717	accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4718	accept->flags = READ_ONCE(sqe->accept_flags);
4719	accept->nofile = rlimit(RLIMIT_NOFILE);
4720	return 0;
4721}
4722
4723static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4724{
4725	struct io_accept *accept = &req->accept;
4726	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4727	unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4728	int ret;
4729
4730	if (req->file->f_flags & O_NONBLOCK)
4731		req->flags |= REQ_F_NOWAIT;
4732
4733	ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4734					accept->addr_len, accept->flags,
4735					accept->nofile);
4736	if (ret == -EAGAIN && force_nonblock)
4737		return -EAGAIN;
4738	if (ret < 0) {
4739		if (ret == -ERESTARTSYS)
4740			ret = -EINTR;
4741		req_set_fail(req);
4742	}
4743	__io_req_complete(req, issue_flags, ret, 0);
4744	return 0;
4745}
4746
4747static int io_connect_prep_async(struct io_kiocb *req)
4748{
4749	struct io_async_connect *io = req->async_data;
4750	struct io_connect *conn = &req->connect;
4751
4752	return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4753}
4754
4755static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4756{
4757	struct io_connect *conn = &req->connect;
4758
4759	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4760		return -EINVAL;
4761	if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4762		return -EINVAL;
4763
4764	conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4765	conn->addr_len =  READ_ONCE(sqe->addr2);
4766	return 0;
4767}
4768
4769static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4770{
4771	struct io_async_connect __io, *io;
4772	unsigned file_flags;
4773	int ret;
4774	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4775
4776	if (req->async_data) {
4777		io = req->async_data;
4778	} else {
4779		ret = move_addr_to_kernel(req->connect.addr,
4780						req->connect.addr_len,
4781						&__io.address);
4782		if (ret)
4783			goto out;
4784		io = &__io;
4785	}
4786
4787	file_flags = force_nonblock ? O_NONBLOCK : 0;
4788
4789	ret = __sys_connect_file(req->file, &io->address,
4790					req->connect.addr_len, file_flags);
4791	if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4792		if (req->async_data)
4793			return -EAGAIN;
4794		if (io_alloc_async_data(req)) {
4795			ret = -ENOMEM;
4796			goto out;
4797		}
4798		memcpy(req->async_data, &__io, sizeof(__io));
4799		return -EAGAIN;
4800	}
4801	if (ret == -ERESTARTSYS)
4802		ret = -EINTR;
4803out:
4804	if (ret < 0)
4805		req_set_fail(req);
4806	__io_req_complete(req, issue_flags, ret, 0);
4807	return 0;
4808}
4809#else /* !CONFIG_NET */
4810#define IO_NETOP_FN(op)							\
4811static int io_##op(struct io_kiocb *req, unsigned int issue_flags)	\
4812{									\
4813	return -EOPNOTSUPP;						\
4814}
4815
4816#define IO_NETOP_PREP(op)						\
4817IO_NETOP_FN(op)								\
4818static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4819{									\
4820	return -EOPNOTSUPP;						\
4821}									\
4822
4823#define IO_NETOP_PREP_ASYNC(op)						\
4824IO_NETOP_PREP(op)							\
4825static int io_##op##_prep_async(struct io_kiocb *req)			\
4826{									\
4827	return -EOPNOTSUPP;						\
4828}
4829
4830IO_NETOP_PREP_ASYNC(sendmsg);
4831IO_NETOP_PREP_ASYNC(recvmsg);
4832IO_NETOP_PREP_ASYNC(connect);
4833IO_NETOP_PREP(accept);
4834IO_NETOP_FN(send);
4835IO_NETOP_FN(recv);
4836#endif /* CONFIG_NET */
4837
4838struct io_poll_table {
4839	struct poll_table_struct pt;
4840	struct io_kiocb *req;
4841	int nr_entries;
4842	int error;
4843};
4844
4845static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4846			   __poll_t mask, io_req_tw_func_t func)
4847{
4848	/* for instances that support it check for an event match first: */
4849	if (mask && !(mask & poll->events))
4850		return 0;
4851
4852	trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4853
4854	list_del_init(&poll->wait.entry);
4855
4856	req->result = mask;
4857	req->io_task_work.func = func;
4858
4859	/*
4860	 * If this fails, then the task is exiting. When a task exits, the
4861	 * work gets canceled, so just cancel this request as well instead
4862	 * of executing it. We can't safely execute it anyway, as we may not
4863	 * have the needed state needed for it anyway.
4864	 */
4865	io_req_task_work_add(req);
4866	return 1;
4867}
4868
4869static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4870	__acquires(&req->ctx->completion_lock)
4871{
4872	struct io_ring_ctx *ctx = req->ctx;
4873
4874	if (unlikely(req->task->flags & PF_EXITING))
4875		WRITE_ONCE(poll->canceled, true);
4876
4877	if (!req->result && !READ_ONCE(poll->canceled)) {
4878		struct poll_table_struct pt = { ._key = poll->events };
4879
4880		req->result = vfs_poll(req->file, &pt) & poll->events;
4881	}
4882
4883	spin_lock_irq(&ctx->completion_lock);
4884	if (!req->result && !READ_ONCE(poll->canceled)) {
4885		add_wait_queue(poll->head, &poll->wait);
4886		return true;
4887	}
4888
4889	return false;
4890}
4891
4892static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4893{
4894	/* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4895	if (req->opcode == IORING_OP_POLL_ADD)
4896		return req->async_data;
4897	return req->apoll->double_poll;
4898}
4899
4900static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4901{
4902	if (req->opcode == IORING_OP_POLL_ADD)
4903		return &req->poll;
4904	return &req->apoll->poll;
4905}
4906
4907static void io_poll_remove_double(struct io_kiocb *req)
4908	__must_hold(&req->ctx->completion_lock)
4909{
4910	struct io_poll_iocb *poll = io_poll_get_double(req);
4911
4912	lockdep_assert_held(&req->ctx->completion_lock);
4913
4914	if (poll && poll->head) {
4915		struct wait_queue_head *head = poll->head;
4916
4917		spin_lock(&head->lock);
4918		list_del_init(&poll->wait.entry);
4919		if (poll->wait.private)
4920			req_ref_put(req);
4921		poll->head = NULL;
4922		spin_unlock(&head->lock);
4923	}
4924}
4925
4926static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4927	__must_hold(&req->ctx->completion_lock)
4928{
4929	struct io_ring_ctx *ctx = req->ctx;
4930	unsigned flags = IORING_CQE_F_MORE;
4931	int error;
4932
4933	if (READ_ONCE(req->poll.canceled)) {
4934		error = -ECANCELED;
4935		req->poll.events |= EPOLLONESHOT;
4936	} else {
4937		error = mangle_poll(mask);
4938	}
4939	if (req->poll.events & EPOLLONESHOT)
4940		flags = 0;
4941	if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4942		req->poll.done = true;
4943		flags = 0;
4944	}
4945	if (flags & IORING_CQE_F_MORE)
4946		ctx->cq_extra++;
4947
4948	io_commit_cqring(ctx);
4949	return !(flags & IORING_CQE_F_MORE);
4950}
4951
4952static void io_poll_task_func(struct io_kiocb *req)
4953{
4954	struct io_ring_ctx *ctx = req->ctx;
4955	struct io_kiocb *nxt;
4956
4957	if (io_poll_rewait(req, &req->poll)) {
4958		spin_unlock_irq(&ctx->completion_lock);
4959	} else {
4960		bool done;
4961
4962		done = io_poll_complete(req, req->result);
4963		if (done) {
4964			io_poll_remove_double(req);
4965			hash_del(&req->hash_node);
4966		} else {
4967			req->result = 0;
4968			add_wait_queue(req->poll.head, &req->poll.wait);
4969		}
4970		spin_unlock_irq(&ctx->completion_lock);
4971		io_cqring_ev_posted(ctx);
4972
4973		if (done) {
4974			nxt = io_put_req_find_next(req);
4975			if (nxt)
4976				io_req_task_submit(nxt);
4977		}
4978	}
4979}
4980
4981static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4982			       int sync, void *key)
4983{
4984	struct io_kiocb *req = wait->private;
4985	struct io_poll_iocb *poll = io_poll_get_single(req);
4986	__poll_t mask = key_to_poll(key);
4987
4988	/* for instances that support it check for an event match first: */
4989	if (mask && !(mask & poll->events))
4990		return 0;
4991	if (!(poll->events & EPOLLONESHOT))
4992		return poll->wait.func(&poll->wait, mode, sync, key);
4993
4994	list_del_init(&wait->entry);
4995
4996	if (poll->head) {
4997		bool done;
4998
4999		spin_lock(&poll->head->lock);
5000		done = list_empty(&poll->wait.entry);
5001		if (!done)
5002			list_del_init(&poll->wait.entry);
5003		/* make sure double remove sees this as being gone */
5004		wait->private = NULL;
5005		spin_unlock(&poll->head->lock);
5006		if (!done) {
5007			/* use wait func handler, so it matches the rq type */
5008			poll->wait.func(&poll->wait, mode, sync, key);
5009		}
5010	}
5011	req_ref_put(req);
5012	return 1;
5013}
5014
5015static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5016			      wait_queue_func_t wake_func)
5017{
5018	poll->head = NULL;
5019	poll->done = false;
5020	poll->canceled = false;
5021#define IO_POLL_UNMASK	(EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5022	/* mask in events that we always want/need */
5023	poll->events = events | IO_POLL_UNMASK;
5024	INIT_LIST_HEAD(&poll->wait.entry);
5025	init_waitqueue_func_entry(&poll->wait, wake_func);
5026}
5027
5028static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5029			    struct wait_queue_head *head,
5030			    struct io_poll_iocb **poll_ptr)
5031{
5032	struct io_kiocb *req = pt->req;
5033
5034	/*
5035	 * The file being polled uses multiple waitqueues for poll handling
5036	 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5037	 * if this happens.
5038	 */
5039	if (unlikely(pt->nr_entries)) {
5040		struct io_poll_iocb *poll_one = poll;
5041
5042		/* already have a 2nd entry, fail a third attempt */
5043		if (*poll_ptr) {
5044			pt->error = -EINVAL;
5045			return;
5046		}
5047		/*
5048		 * Can't handle multishot for double wait for now, turn it
5049		 * into one-shot mode.
5050		 */
5051		if (!(poll_one->events & EPOLLONESHOT))
5052			poll_one->events |= EPOLLONESHOT;
5053		/* double add on the same waitqueue head, ignore */
5054		if (poll_one->head == head)
5055			return;
5056		poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5057		if (!poll) {
5058			pt->error = -ENOMEM;
5059			return;
5060		}
5061		io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5062		req_ref_get(req);
5063		poll->wait.private = req;
5064		*poll_ptr = poll;
5065	}
5066
5067	pt->nr_entries++;
5068	poll->head = head;
5069
5070	if (poll->events & EPOLLEXCLUSIVE)
5071		add_wait_queue_exclusive(head, &poll->wait);
5072	else
5073		add_wait_queue(head, &poll->wait);
5074}
5075
5076static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5077			       struct poll_table_struct *p)
5078{
5079	struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5080	struct async_poll *apoll = pt->req->apoll;
5081
5082	__io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5083}
5084
5085static void io_async_task_func(struct io_kiocb *req)
5086{
5087	struct async_poll *apoll = req->apoll;
5088	struct io_ring_ctx *ctx = req->ctx;
5089
5090	trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5091
5092	if (io_poll_rewait(req, &apoll->poll)) {
5093		spin_unlock_irq(&ctx->completion_lock);
5094		return;
5095	}
5096
5097	hash_del(&req->hash_node);
5098	io_poll_remove_double(req);
5099	spin_unlock_irq(&ctx->completion_lock);
5100
5101	if (!READ_ONCE(apoll->poll.canceled))
5102		io_req_task_submit(req);
5103	else
5104		io_req_complete_failed(req, -ECANCELED);
5105}
5106
5107static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5108			void *key)
5109{
5110	struct io_kiocb *req = wait->private;
5111	struct io_poll_iocb *poll = &req->apoll->poll;
5112
5113	trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5114					key_to_poll(key));
5115
5116	return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5117}
5118
5119static void io_poll_req_insert(struct io_kiocb *req)
5120{
5121	struct io_ring_ctx *ctx = req->ctx;
5122	struct hlist_head *list;
5123
5124	list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5125	hlist_add_head(&req->hash_node, list);
5126}
5127
5128static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5129				      struct io_poll_iocb *poll,
5130				      struct io_poll_table *ipt, __poll_t mask,
5131				      wait_queue_func_t wake_func)
5132	__acquires(&ctx->completion_lock)
5133{
5134	struct io_ring_ctx *ctx = req->ctx;
5135	bool cancel = false;
5136
5137	INIT_HLIST_NODE(&req->hash_node);
5138	io_init_poll_iocb(poll, mask, wake_func);
5139	poll->file = req->file;
5140	poll->wait.private = req;
5141
5142	ipt->pt._key = mask;
5143	ipt->req = req;
5144	ipt->error = 0;
5145	ipt->nr_entries = 0;
5146
5147	mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5148	if (unlikely(!ipt->nr_entries) && !ipt->error)
5149		ipt->error = -EINVAL;
5150
5151	spin_lock_irq(&ctx->completion_lock);
5152	if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5153		io_poll_remove_double(req);
5154	if (likely(poll->head)) {
5155		spin_lock(&poll->head->lock);
5156		if (unlikely(list_empty(&poll->wait.entry))) {
5157			if (ipt->error)
5158				cancel = true;
5159			ipt->error = 0;
5160			mask = 0;
5161		}
5162		if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5163			list_del_init(&poll->wait.entry);
5164		else if (cancel)
5165			WRITE_ONCE(poll->canceled, true);
5166		else if (!poll->done) /* actually waiting for an event */
5167			io_poll_req_insert(req);
5168		spin_unlock(&poll->head->lock);
5169	}
5170
5171	return mask;
5172}
5173
5174enum {
5175	IO_APOLL_OK,
5176	IO_APOLL_ABORTED,
5177	IO_APOLL_READY
5178};
5179
5180static int io_arm_poll_handler(struct io_kiocb *req)
5181{
5182	const struct io_op_def *def = &io_op_defs[req->opcode];
5183	struct io_ring_ctx *ctx = req->ctx;
5184	struct async_poll *apoll;
5185	struct io_poll_table ipt;
5186	__poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5187	int rw;
5188
5189	if (!req->file || !file_can_poll(req->file))
5190		return IO_APOLL_ABORTED;
5191	if (req->flags & REQ_F_POLLED)
5192		return IO_APOLL_ABORTED;
5193	if (!def->pollin && !def->pollout)
5194		return IO_APOLL_ABORTED;
5195
5196	if (def->pollin) {
5197		rw = READ;
5198		mask |= POLLIN | POLLRDNORM;
5199
5200		/* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5201		if ((req->opcode == IORING_OP_RECVMSG) &&
5202		    (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5203			mask &= ~POLLIN;
5204	} else {
5205		rw = WRITE;
5206		mask |= POLLOUT | POLLWRNORM;
5207	}
5208
5209	/* if we can't nonblock try, then no point in arming a poll handler */
5210	if (!io_file_supports_nowait(req, rw))
5211		return IO_APOLL_ABORTED;
5212
5213	apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5214	if (unlikely(!apoll))
5215		return IO_APOLL_ABORTED;
5216	apoll->double_poll = NULL;
5217	req->apoll = apoll;
5218	req->flags |= REQ_F_POLLED;
5219	ipt.pt._qproc = io_async_queue_proc;
5220
5221	ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5222					io_async_wake);
5223	if (ret || ipt.error) {
5224		spin_unlock_irq(&ctx->completion_lock);
5225		if (ret)
5226			return IO_APOLL_READY;
5227		return IO_APOLL_ABORTED;
5228	}
5229	spin_unlock_irq(&ctx->completion_lock);
5230	trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5231				mask, apoll->poll.events);
5232	return IO_APOLL_OK;
5233}
5234
5235static bool __io_poll_remove_one(struct io_kiocb *req,
5236				 struct io_poll_iocb *poll, bool do_cancel)
5237	__must_hold(&req->ctx->completion_lock)
5238{
5239	bool do_complete = false;
5240
5241	if (!poll->head)
5242		return false;
5243	spin_lock(&poll->head->lock);
5244	if (do_cancel)
5245		WRITE_ONCE(poll->canceled, true);
5246	if (!list_empty(&poll->wait.entry)) {
5247		list_del_init(&poll->wait.entry);
5248		do_complete = true;
5249	}
5250	spin_unlock(&poll->head->lock);
5251	hash_del(&req->hash_node);
5252	return do_complete;
5253}
5254
5255static bool io_poll_remove_waitqs(struct io_kiocb *req)
5256	__must_hold(&req->ctx->completion_lock)
5257{
5258	bool do_complete;
5259
5260	io_poll_remove_double(req);
5261	do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5262
5263	if (req->opcode != IORING_OP_POLL_ADD && do_complete) {
5264		/* non-poll requests have submit ref still */
5265		req_ref_put(req);
5266	}
5267	return do_complete;
5268}
5269
5270static bool io_poll_remove_one(struct io_kiocb *req)
5271	__must_hold(&req->ctx->completion_lock)
5272{
5273	bool do_complete;
5274
5275	do_complete = io_poll_remove_waitqs(req);
5276	if (do_complete) {
5277		io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5278		io_commit_cqring(req->ctx);
5279		req_set_fail(req);
5280		io_put_req_deferred(req, 1);
5281	}
5282
5283	return do_complete;
5284}
5285
5286/*
5287 * Returns true if we found and killed one or more poll requests
5288 */
5289static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5290			       bool cancel_all)
5291{
5292	struct hlist_node *tmp;
5293	struct io_kiocb *req;
5294	int posted = 0, i;
5295
5296	spin_lock_irq(&ctx->completion_lock);
5297	for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5298		struct hlist_head *list;
5299
5300		list = &ctx->cancel_hash[i];
5301		hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5302			if (io_match_task(req, tsk, cancel_all))
5303				posted += io_poll_remove_one(req);
5304		}
5305	}
5306	spin_unlock_irq(&ctx->completion_lock);
5307
5308	if (posted)
5309		io_cqring_ev_posted(ctx);
5310
5311	return posted != 0;
5312}
5313
5314static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5315				     bool poll_only)
5316	__must_hold(&ctx->completion_lock)
5317{
5318	struct hlist_head *list;
5319	struct io_kiocb *req;
5320
5321	list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5322	hlist_for_each_entry(req, list, hash_node) {
5323		if (sqe_addr != req->user_data)
5324			continue;
5325		if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5326			continue;
5327		return req;
5328	}
5329	return NULL;
5330}
5331
5332static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5333			  bool poll_only)
5334	__must_hold(&ctx->completion_lock)
5335{
5336	struct io_kiocb *req;
5337
5338	req = io_poll_find(ctx, sqe_addr, poll_only);
5339	if (!req)
5340		return -ENOENT;
5341	if (io_poll_remove_one(req))
5342		return 0;
5343
5344	return -EALREADY;
5345}
5346
5347static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5348				     unsigned int flags)
5349{
5350	u32 events;
5351
5352	events = READ_ONCE(sqe->poll32_events);
5353#ifdef __BIG_ENDIAN
5354	events = swahw32(events);
5355#endif
5356	if (!(flags & IORING_POLL_ADD_MULTI))
5357		events |= EPOLLONESHOT;
5358	return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5359}
5360
5361static int io_poll_update_prep(struct io_kiocb *req,
5362			       const struct io_uring_sqe *sqe)
5363{
5364	struct io_poll_update *upd = &req->poll_update;
5365	u32 flags;
5366
5367	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5368		return -EINVAL;
5369	if (sqe->ioprio || sqe->buf_index)
5370		return -EINVAL;
5371	flags = READ_ONCE(sqe->len);
5372	if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5373		      IORING_POLL_ADD_MULTI))
5374		return -EINVAL;
5375	/* meaningless without update */
5376	if (flags == IORING_POLL_ADD_MULTI)
5377		return -EINVAL;
5378
5379	upd->old_user_data = READ_ONCE(sqe->addr);
5380	upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5381	upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5382
5383	upd->new_user_data = READ_ONCE(sqe->off);
5384	if (!upd->update_user_data && upd->new_user_data)
5385		return -EINVAL;
5386	if (upd->update_events)
5387		upd->events = io_poll_parse_events(sqe, flags);
5388	else if (sqe->poll32_events)
5389		return -EINVAL;
5390
5391	return 0;
5392}
5393
5394static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5395			void *key)
5396{
5397	struct io_kiocb *req = wait->private;
5398	struct io_poll_iocb *poll = &req->poll;
5399
5400	return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5401}
5402
5403static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5404			       struct poll_table_struct *p)
5405{
5406	struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5407
5408	__io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5409}
5410
5411static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5412{
5413	struct io_poll_iocb *poll = &req->poll;
5414	u32 flags;
5415
5416	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5417		return -EINVAL;
5418	if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5419		return -EINVAL;
5420	flags = READ_ONCE(sqe->len);
5421	if (flags & ~IORING_POLL_ADD_MULTI)
5422		return -EINVAL;
5423
5424	poll->events = io_poll_parse_events(sqe, flags);
5425	return 0;
5426}
5427
5428static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5429{
5430	struct io_poll_iocb *poll = &req->poll;
5431	struct io_ring_ctx *ctx = req->ctx;
5432	struct io_poll_table ipt;
5433	__poll_t mask;
5434
5435	ipt.pt._qproc = io_poll_queue_proc;
5436
5437	mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5438					io_poll_wake);
5439
5440	if (mask) { /* no async, we'd stolen it */
5441		ipt.error = 0;
5442		io_poll_complete(req, mask);
5443	}
5444	spin_unlock_irq(&ctx->completion_lock);
5445
5446	if (mask) {
5447		io_cqring_ev_posted(ctx);
5448		if (poll->events & EPOLLONESHOT)
5449			io_put_req(req);
5450	}
5451	return ipt.error;
5452}
5453
5454static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5455{
5456	struct io_ring_ctx *ctx = req->ctx;
5457	struct io_kiocb *preq;
5458	bool completing;
5459	int ret;
5460
5461	spin_lock_irq(&ctx->completion_lock);
5462	preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5463	if (!preq) {
5464		ret = -ENOENT;
5465		goto err;
5466	}
5467
5468	if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5469		completing = true;
5470		ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5471		goto err;
5472	}
5473
5474	/*
5475	 * Don't allow racy completion with singleshot, as we cannot safely
5476	 * update those. For multishot, if we're racing with completion, just
5477	 * let completion re-add it.
5478	 */
5479	completing = !__io_poll_remove_one(preq, &preq->poll, false);
5480	if (completing && (preq->poll.events & EPOLLONESHOT)) {
5481		ret = -EALREADY;
5482		goto err;
5483	}
5484	/* we now have a detached poll request. reissue. */
5485	ret = 0;
5486err:
5487	if (ret < 0) {
5488		spin_unlock_irq(&ctx->completion_lock);
5489		req_set_fail(req);
5490		io_req_complete(req, ret);
5491		return 0;
5492	}
5493	/* only mask one event flags, keep behavior flags */
5494	if (req->poll_update.update_events) {
5495		preq->poll.events &= ~0xffff;
5496		preq->poll.events |= req->poll_update.events & 0xffff;
5497		preq->poll.events |= IO_POLL_UNMASK;
5498	}
5499	if (req->poll_update.update_user_data)
5500		preq->user_data = req->poll_update.new_user_data;
5501	spin_unlock_irq(&ctx->completion_lock);
5502
5503	/* complete update request, we're done with it */
5504	io_req_complete(req, ret);
5505
5506	if (!completing) {
5507		ret = io_poll_add(preq, issue_flags);
5508		if (ret < 0) {
5509			req_set_fail(preq);
5510			io_req_complete(preq, ret);
5511		}
5512	}
5513	return 0;
5514}
5515
5516static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5517{
5518	struct io_timeout_data *data = container_of(timer,
5519						struct io_timeout_data, timer);
5520	struct io_kiocb *req = data->req;
5521	struct io_ring_ctx *ctx = req->ctx;
5522	unsigned long flags;
5523
5524	spin_lock_irqsave(&ctx->completion_lock, flags);
5525	list_del_init(&req->timeout.list);
5526	atomic_set(&req->ctx->cq_timeouts,
5527		atomic_read(&req->ctx->cq_timeouts) + 1);
5528
5529	io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5530	io_commit_cqring(ctx);
5531	spin_unlock_irqrestore(&ctx->completion_lock, flags);
5532
5533	io_cqring_ev_posted(ctx);
5534	req_set_fail(req);
5535	io_put_req(req);
5536	return HRTIMER_NORESTART;
5537}
5538
5539static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5540					   __u64 user_data)
5541	__must_hold(&ctx->completion_lock)
5542{
5543	struct io_timeout_data *io;
5544	struct io_kiocb *req;
5545	bool found = false;
5546
5547	list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5548		found = user_data == req->user_data;
5549		if (found)
5550			break;
5551	}
5552	if (!found)
5553		return ERR_PTR(-ENOENT);
5554
5555	io = req->async_data;
5556	if (hrtimer_try_to_cancel(&io->timer) == -1)
5557		return ERR_PTR(-EALREADY);
5558	list_del_init(&req->timeout.list);
5559	return req;
5560}
5561
5562static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5563	__must_hold(&ctx->completion_lock)
5564{
5565	struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5566
5567	if (IS_ERR(req))
5568		return PTR_ERR(req);
5569
5570	req_set_fail(req);
5571	io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5572	io_put_req_deferred(req, 1);
5573	return 0;
5574}
5575
5576static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5577			     struct timespec64 *ts, enum hrtimer_mode mode)
5578	__must_hold(&ctx->completion_lock)
5579{
5580	struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5581	struct io_timeout_data *data;
5582
5583	if (IS_ERR(req))
5584		return PTR_ERR(req);
5585
5586	req->timeout.off = 0; /* noseq */
5587	data = req->async_data;
5588	list_add_tail(&req->timeout.list, &ctx->timeout_list);
5589	hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5590	data->timer.function = io_timeout_fn;
5591	hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5592	return 0;
5593}
5594
5595static int io_timeout_remove_prep(struct io_kiocb *req,
5596				  const struct io_uring_sqe *sqe)
5597{
5598	struct io_timeout_rem *tr = &req->timeout_rem;
5599
5600	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5601		return -EINVAL;
5602	if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5603		return -EINVAL;
5604	if (sqe->ioprio || sqe->buf_index || sqe->len)
5605		return -EINVAL;
5606
5607	tr->addr = READ_ONCE(sqe->addr);
5608	tr->flags = READ_ONCE(sqe->timeout_flags);
5609	if (tr->flags & IORING_TIMEOUT_UPDATE) {
5610		if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5611			return -EINVAL;
5612		if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5613			return -EFAULT;
5614	} else if (tr->flags) {
5615		/* timeout removal doesn't support flags */
5616		return -EINVAL;
5617	}
5618
5619	return 0;
5620}
5621
5622static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5623{
5624	return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5625					    : HRTIMER_MODE_REL;
5626}
5627
5628/*
5629 * Remove or update an existing timeout command
5630 */
5631static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5632{
5633	struct io_timeout_rem *tr = &req->timeout_rem;
5634	struct io_ring_ctx *ctx = req->ctx;
5635	int ret;
5636
5637	spin_lock_irq(&ctx->completion_lock);
5638	if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5639		ret = io_timeout_cancel(ctx, tr->addr);
5640	else
5641		ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5642					io_translate_timeout_mode(tr->flags));
5643
5644	io_cqring_fill_event(ctx, req->user_data, ret, 0);
5645	io_commit_cqring(ctx);
5646	spin_unlock_irq(&ctx->completion_lock);
5647	io_cqring_ev_posted(ctx);
5648	if (ret < 0)
5649		req_set_fail(req);
5650	io_put_req(req);
5651	return 0;
5652}
5653
5654static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5655			   bool is_timeout_link)
5656{
5657	struct io_timeout_data *data;
5658	unsigned flags;
5659	u32 off = READ_ONCE(sqe->off);
5660
5661	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5662		return -EINVAL;
5663	if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5664		return -EINVAL;
5665	if (off && is_timeout_link)
5666		return -EINVAL;
5667	flags = READ_ONCE(sqe->timeout_flags);
5668	if (flags & ~IORING_TIMEOUT_ABS)
5669		return -EINVAL;
5670
5671	req->timeout.off = off;
5672	if (unlikely(off && !req->ctx->off_timeout_used))
5673		req->ctx->off_timeout_used = true;
5674
5675	if (!req->async_data && io_alloc_async_data(req))
5676		return -ENOMEM;
5677
5678	data = req->async_data;
5679	data->req = req;
5680
5681	if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5682		return -EFAULT;
5683
5684	data->mode = io_translate_timeout_mode(flags);
5685	hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5686	if (is_timeout_link)
5687		io_req_track_inflight(req);
5688	return 0;
5689}
5690
5691static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5692{
5693	struct io_ring_ctx *ctx = req->ctx;
5694	struct io_timeout_data *data = req->async_data;
5695	struct list_head *entry;
5696	u32 tail, off = req->timeout.off;
5697
5698	spin_lock_irq(&ctx->completion_lock);
5699
5700	/*
5701	 * sqe->off holds how many events that need to occur for this
5702	 * timeout event to be satisfied. If it isn't set, then this is
5703	 * a pure timeout request, sequence isn't used.
5704	 */
5705	if (io_is_timeout_noseq(req)) {
5706		entry = ctx->timeout_list.prev;
5707		goto add;
5708	}
5709
5710	tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5711	req->timeout.target_seq = tail + off;
5712
5713	/* Update the last seq here in case io_flush_timeouts() hasn't.
5714	 * This is safe because ->completion_lock is held, and submissions
5715	 * and completions are never mixed in the same ->completion_lock section.
5716	 */
5717	ctx->cq_last_tm_flush = tail;
5718
5719	/*
5720	 * Insertion sort, ensuring the first entry in the list is always
5721	 * the one we need first.
5722	 */
5723	list_for_each_prev(entry, &ctx->timeout_list) {
5724		struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5725						  timeout.list);
5726
5727		if (io_is_timeout_noseq(nxt))
5728			continue;
5729		/* nxt.seq is behind @tail, otherwise would've been completed */
5730		if (