io_uring.c revision 4d13d1a4
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_submit_link {
303	struct io_kiocb		*head;
304	struct io_kiocb		*last;
305};
306
307struct io_submit_state {
308	struct blk_plug		plug;
309	struct io_submit_link	link;
310
311	/*
312	 * io_kiocb alloc cache
313	 */
314	void			*reqs[IO_REQ_CACHE_SIZE];
315	unsigned int		free_reqs;
316
317	bool			plug_started;
318
319	/*
320	 * Batch completion logic
321	 */
322	struct io_kiocb		*compl_reqs[IO_COMPL_BATCH];
323	unsigned int		compl_nr;
324	/* inline/task_work completion list, under ->uring_lock */
325	struct list_head	free_list;
326
327	unsigned int		ios_left;
328};
329
330struct io_ring_ctx {
331	/* const or read-mostly hot data */
332	struct {
333		struct percpu_ref	refs;
334
335		struct io_rings		*rings;
336		unsigned int		flags;
337		unsigned int		compat: 1;
338		unsigned int		drain_next: 1;
339		unsigned int		eventfd_async: 1;
340		unsigned int		restricted: 1;
341		unsigned int		off_timeout_used: 1;
342		unsigned int		drain_active: 1;
343	} ____cacheline_aligned_in_smp;
344
345	/* submission data */
346	struct {
347		struct mutex		uring_lock;
348
349		/*
350		 * Ring buffer of indices into array of io_uring_sqe, which is
351		 * mmapped by the application using the IORING_OFF_SQES offset.
352		 *
353		 * This indirection could e.g. be used to assign fixed
354		 * io_uring_sqe entries to operations and only submit them to
355		 * the queue when needed.
356		 *
357		 * The kernel modifies neither the indices array nor the entries
358		 * array.
359		 */
360		u32			*sq_array;
361		struct io_uring_sqe	*sq_sqes;
362		unsigned		cached_sq_head;
363		unsigned		sq_entries;
364		struct list_head	defer_list;
365
366		/*
367		 * Fixed resources fast path, should be accessed only under
368		 * uring_lock, and updated through io_uring_register(2)
369		 */
370		struct io_rsrc_node	*rsrc_node;
371		struct io_file_table	file_table;
372		unsigned		nr_user_files;
373		unsigned		nr_user_bufs;
374		struct io_mapped_ubuf	**user_bufs;
375
376		struct io_submit_state	submit_state;
377		struct list_head	timeout_list;
378		struct list_head	cq_overflow_list;
379		struct xarray		io_buffers;
380		struct xarray		personalities;
381		u32			pers_next;
382		unsigned		sq_thread_idle;
383	} ____cacheline_aligned_in_smp;
384
385	/* IRQ completion list, under ->completion_lock */
386	struct list_head	locked_free_list;
387	unsigned int		locked_free_nr;
388
389	const struct cred	*sq_creds;	/* cred used for __io_sq_thread() */
390	struct io_sq_data	*sq_data;	/* if using sq thread polling */
391
392	struct wait_queue_head	sqo_sq_wait;
393	struct list_head	sqd_list;
394
395	unsigned long		check_cq_overflow;
396
397	struct {
398		unsigned		cached_cq_tail;
399		unsigned		cq_entries;
400		struct eventfd_ctx	*cq_ev_fd;
401		struct wait_queue_head	poll_wait;
402		struct wait_queue_head	cq_wait;
403		unsigned		cq_extra;
404		atomic_t		cq_timeouts;
405		struct fasync_struct	*cq_fasync;
406		unsigned		cq_last_tm_flush;
407	} ____cacheline_aligned_in_smp;
408
409	struct {
410		spinlock_t		completion_lock;
411
412		spinlock_t		timeout_lock;
413
414		/*
415		 * ->iopoll_list is protected by the ctx->uring_lock for
416		 * io_uring instances that don't use IORING_SETUP_SQPOLL.
417		 * For SQPOLL, only the single threaded io_sq_thread() will
418		 * manipulate the list, hence no extra locking is needed there.
419		 */
420		struct list_head	iopoll_list;
421		struct hlist_head	*cancel_hash;
422		unsigned		cancel_hash_bits;
423		bool			poll_multi_queue;
424	} ____cacheline_aligned_in_smp;
425
426	struct io_restriction		restrictions;
427
428	/* slow path rsrc auxilary data, used by update/register */
429	struct {
430		struct io_rsrc_node		*rsrc_backup_node;
431		struct io_mapped_ubuf		*dummy_ubuf;
432		struct io_rsrc_data		*file_data;
433		struct io_rsrc_data		*buf_data;
434
435		struct delayed_work		rsrc_put_work;
436		struct llist_head		rsrc_put_llist;
437		struct list_head		rsrc_ref_list;
438		spinlock_t			rsrc_ref_lock;
439	};
440
441	/* Keep this last, we don't need it for the fast path */
442	struct {
443		#if defined(CONFIG_UNIX)
444			struct socket		*ring_sock;
445		#endif
446		/* hashed buffered write serialization */
447		struct io_wq_hash		*hash_map;
448
449		/* Only used for accounting purposes */
450		struct user_struct		*user;
451		struct mm_struct		*mm_account;
452
453		/* ctx exit and cancelation */
454		struct llist_head		fallback_llist;
455		struct delayed_work		fallback_work;
456		struct work_struct		exit_work;
457		struct list_head		tctx_list;
458		struct completion		ref_comp;
459	};
460};
461
462struct io_uring_task {
463	/* submission side */
464	int			cached_refs;
465	struct xarray		xa;
466	struct wait_queue_head	wait;
467	const struct io_ring_ctx *last;
468	struct io_wq		*io_wq;
469	struct percpu_counter	inflight;
470	atomic_t		inflight_tracked;
471	atomic_t		in_idle;
472
473	spinlock_t		task_lock;
474	struct io_wq_work_list	task_list;
475	struct callback_head	task_work;
476	bool			task_running;
477};
478
479/*
480 * First field must be the file pointer in all the
481 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
482 */
483struct io_poll_iocb {
484	struct file			*file;
485	struct wait_queue_head		*head;
486	__poll_t			events;
487	bool				done;
488	bool				canceled;
489	struct wait_queue_entry		wait;
490};
491
492struct io_poll_update {
493	struct file			*file;
494	u64				old_user_data;
495	u64				new_user_data;
496	__poll_t			events;
497	bool				update_events;
498	bool				update_user_data;
499};
500
501struct io_close {
502	struct file			*file;
503	int				fd;
504};
505
506struct io_timeout_data {
507	struct io_kiocb			*req;
508	struct hrtimer			timer;
509	struct timespec64		ts;
510	enum hrtimer_mode		mode;
511};
512
513struct io_accept {
514	struct file			*file;
515	struct sockaddr __user		*addr;
516	int __user			*addr_len;
517	int				flags;
518	unsigned long			nofile;
519};
520
521struct io_sync {
522	struct file			*file;
523	loff_t				len;
524	loff_t				off;
525	int				flags;
526	int				mode;
527};
528
529struct io_cancel {
530	struct file			*file;
531	u64				addr;
532};
533
534struct io_timeout {
535	struct file			*file;
536	u32				off;
537	u32				target_seq;
538	struct list_head		list;
539	/* head of the link, used by linked timeouts only */
540	struct io_kiocb			*head;
541	/* for linked completions */
542	struct io_kiocb			*prev;
543};
544
545struct io_timeout_rem {
546	struct file			*file;
547	u64				addr;
548
549	/* timeout update */
550	struct timespec64		ts;
551	u32				flags;
552};
553
554struct io_rw {
555	/* NOTE: kiocb has the file as the first member, so don't do it here */
556	struct kiocb			kiocb;
557	u64				addr;
558	u64				len;
559};
560
561struct io_connect {
562	struct file			*file;
563	struct sockaddr __user		*addr;
564	int				addr_len;
565};
566
567struct io_sr_msg {
568	struct file			*file;
569	union {
570		struct compat_msghdr __user	*umsg_compat;
571		struct user_msghdr __user	*umsg;
572		void __user			*buf;
573	};
574	int				msg_flags;
575	int				bgid;
576	size_t				len;
577	struct io_buffer		*kbuf;
578};
579
580struct io_open {
581	struct file			*file;
582	int				dfd;
583	struct filename			*filename;
584	struct open_how			how;
585	unsigned long			nofile;
586};
587
588struct io_rsrc_update {
589	struct file			*file;
590	u64				arg;
591	u32				nr_args;
592	u32				offset;
593};
594
595struct io_fadvise {
596	struct file			*file;
597	u64				offset;
598	u32				len;
599	u32				advice;
600};
601
602struct io_madvise {
603	struct file			*file;
604	u64				addr;
605	u32				len;
606	u32				advice;
607};
608
609struct io_epoll {
610	struct file			*file;
611	int				epfd;
612	int				op;
613	int				fd;
614	struct epoll_event		event;
615};
616
617struct io_splice {
618	struct file			*file_out;
619	struct file			*file_in;
620	loff_t				off_out;
621	loff_t				off_in;
622	u64				len;
623	unsigned int			flags;
624};
625
626struct io_provide_buf {
627	struct file			*file;
628	__u64				addr;
629	__u32				len;
630	__u32				bgid;
631	__u16				nbufs;
632	__u16				bid;
633};
634
635struct io_statx {
636	struct file			*file;
637	int				dfd;
638	unsigned int			mask;
639	unsigned int			flags;
640	const char __user		*filename;
641	struct statx __user		*buffer;
642};
643
644struct io_shutdown {
645	struct file			*file;
646	int				how;
647};
648
649struct io_rename {
650	struct file			*file;
651	int				old_dfd;
652	int				new_dfd;
653	struct filename			*oldpath;
654	struct filename			*newpath;
655	int				flags;
656};
657
658struct io_unlink {
659	struct file			*file;
660	int				dfd;
661	int				flags;
662	struct filename			*filename;
663};
664
665struct io_completion {
666	struct file			*file;
667	u32				cflags;
668};
669
670struct io_async_connect {
671	struct sockaddr_storage		address;
672};
673
674struct io_async_msghdr {
675	struct iovec			fast_iov[UIO_FASTIOV];
676	/* points to an allocated iov, if NULL we use fast_iov instead */
677	struct iovec			*free_iov;
678	struct sockaddr __user		*uaddr;
679	struct msghdr			msg;
680	struct sockaddr_storage		addr;
681};
682
683struct io_async_rw {
684	struct iovec			fast_iov[UIO_FASTIOV];
685	const struct iovec		*free_iovec;
686	struct iov_iter			iter;
687	size_t				bytes_done;
688	struct wait_page_queue		wpq;
689};
690
691enum {
692	REQ_F_FIXED_FILE_BIT	= IOSQE_FIXED_FILE_BIT,
693	REQ_F_IO_DRAIN_BIT	= IOSQE_IO_DRAIN_BIT,
694	REQ_F_LINK_BIT		= IOSQE_IO_LINK_BIT,
695	REQ_F_HARDLINK_BIT	= IOSQE_IO_HARDLINK_BIT,
696	REQ_F_FORCE_ASYNC_BIT	= IOSQE_ASYNC_BIT,
697	REQ_F_BUFFER_SELECT_BIT	= IOSQE_BUFFER_SELECT_BIT,
698
699	/* first byte is taken by user flags, shift it to not overlap */
700	REQ_F_FAIL_BIT		= 8,
701	REQ_F_INFLIGHT_BIT,
702	REQ_F_CUR_POS_BIT,
703	REQ_F_NOWAIT_BIT,
704	REQ_F_LINK_TIMEOUT_BIT,
705	REQ_F_NEED_CLEANUP_BIT,
706	REQ_F_POLLED_BIT,
707	REQ_F_BUFFER_SELECTED_BIT,
708	REQ_F_COMPLETE_INLINE_BIT,
709	REQ_F_REISSUE_BIT,
710	REQ_F_DONT_REISSUE_BIT,
711	REQ_F_CREDS_BIT,
712	REQ_F_REFCOUNT_BIT,
713	REQ_F_ARM_LTIMEOUT_BIT,
714	/* keep async read/write and isreg together and in order */
715	REQ_F_NOWAIT_READ_BIT,
716	REQ_F_NOWAIT_WRITE_BIT,
717	REQ_F_ISREG_BIT,
718
719	/* not a real bit, just to check we're not overflowing the space */
720	__REQ_F_LAST_BIT,
721};
722
723enum {
724	/* ctx owns file */
725	REQ_F_FIXED_FILE	= BIT(REQ_F_FIXED_FILE_BIT),
726	/* drain existing IO first */
727	REQ_F_IO_DRAIN		= BIT(REQ_F_IO_DRAIN_BIT),
728	/* linked sqes */
729	REQ_F_LINK		= BIT(REQ_F_LINK_BIT),
730	/* doesn't sever on completion < 0 */
731	REQ_F_HARDLINK		= BIT(REQ_F_HARDLINK_BIT),
732	/* IOSQE_ASYNC */
733	REQ_F_FORCE_ASYNC	= BIT(REQ_F_FORCE_ASYNC_BIT),
734	/* IOSQE_BUFFER_SELECT */
735	REQ_F_BUFFER_SELECT	= BIT(REQ_F_BUFFER_SELECT_BIT),
736
737	/* fail rest of links */
738	REQ_F_FAIL		= BIT(REQ_F_FAIL_BIT),
739	/* on inflight list, should be cancelled and waited on exit reliably */
740	REQ_F_INFLIGHT		= BIT(REQ_F_INFLIGHT_BIT),
741	/* read/write uses file position */
742	REQ_F_CUR_POS		= BIT(REQ_F_CUR_POS_BIT),
743	/* must not punt to workers */
744	REQ_F_NOWAIT		= BIT(REQ_F_NOWAIT_BIT),
745	/* has or had linked timeout */
746	REQ_F_LINK_TIMEOUT	= BIT(REQ_F_LINK_TIMEOUT_BIT),
747	/* needs cleanup */
748	REQ_F_NEED_CLEANUP	= BIT(REQ_F_NEED_CLEANUP_BIT),
749	/* already went through poll handler */
750	REQ_F_POLLED		= BIT(REQ_F_POLLED_BIT),
751	/* buffer already selected */
752	REQ_F_BUFFER_SELECTED	= BIT(REQ_F_BUFFER_SELECTED_BIT),
753	/* completion is deferred through io_comp_state */
754	REQ_F_COMPLETE_INLINE	= BIT(REQ_F_COMPLETE_INLINE_BIT),
755	/* caller should reissue async */
756	REQ_F_REISSUE		= BIT(REQ_F_REISSUE_BIT),
757	/* don't attempt request reissue, see io_rw_reissue() */
758	REQ_F_DONT_REISSUE	= BIT(REQ_F_DONT_REISSUE_BIT),
759	/* supports async reads */
760	REQ_F_NOWAIT_READ	= BIT(REQ_F_NOWAIT_READ_BIT),
761	/* supports async writes */
762	REQ_F_NOWAIT_WRITE	= BIT(REQ_F_NOWAIT_WRITE_BIT),
763	/* regular file */
764	REQ_F_ISREG		= BIT(REQ_F_ISREG_BIT),
765	/* has creds assigned */
766	REQ_F_CREDS		= BIT(REQ_F_CREDS_BIT),
767	/* skip refcounting if not set */
768	REQ_F_REFCOUNT		= BIT(REQ_F_REFCOUNT_BIT),
769	/* there is a linked timeout that has to be armed */
770	REQ_F_ARM_LTIMEOUT	= BIT(REQ_F_ARM_LTIMEOUT_BIT),
771};
772
773struct async_poll {
774	struct io_poll_iocb	poll;
775	struct io_poll_iocb	*double_poll;
776};
777
778typedef void (*io_req_tw_func_t)(struct io_kiocb *req);
779
780struct io_task_work {
781	union {
782		struct io_wq_work_node	node;
783		struct llist_node	fallback_node;
784	};
785	io_req_tw_func_t		func;
786};
787
788enum {
789	IORING_RSRC_FILE		= 0,
790	IORING_RSRC_BUFFER		= 1,
791};
792
793/*
794 * NOTE! Each of the iocb union members has the file pointer
795 * as the first entry in their struct definition. So you can
796 * access the file pointer through any of the sub-structs,
797 * or directly as just 'ki_filp' in this struct.
798 */
799struct io_kiocb {
800	union {
801		struct file		*file;
802		struct io_rw		rw;
803		struct io_poll_iocb	poll;
804		struct io_poll_update	poll_update;
805		struct io_accept	accept;
806		struct io_sync		sync;
807		struct io_cancel	cancel;
808		struct io_timeout	timeout;
809		struct io_timeout_rem	timeout_rem;
810		struct io_connect	connect;
811		struct io_sr_msg	sr_msg;
812		struct io_open		open;
813		struct io_close		close;
814		struct io_rsrc_update	rsrc_update;
815		struct io_fadvise	fadvise;
816		struct io_madvise	madvise;
817		struct io_epoll		epoll;
818		struct io_splice	splice;
819		struct io_provide_buf	pbuf;
820		struct io_statx		statx;
821		struct io_shutdown	shutdown;
822		struct io_rename	rename;
823		struct io_unlink	unlink;
824		/* use only after cleaning per-op data, see io_clean_op() */
825		struct io_completion	compl;
826	};
827
828	/* opcode allocated if it needs to store data for async defer */
829	void				*async_data;
830	u8				opcode;
831	/* polled IO has completed */
832	u8				iopoll_completed;
833
834	u16				buf_index;
835	u32				result;
836
837	struct io_ring_ctx		*ctx;
838	unsigned int			flags;
839	atomic_t			refs;
840	struct task_struct		*task;
841	u64				user_data;
842
843	struct io_kiocb			*link;
844	struct percpu_ref		*fixed_rsrc_refs;
845
846	/* used with ctx->iopoll_list with reads/writes */
847	struct list_head		inflight_entry;
848	struct io_task_work		io_task_work;
849	/* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
850	struct hlist_node		hash_node;
851	struct async_poll		*apoll;
852	struct io_wq_work		work;
853	const struct cred		*creds;
854
855	/* store used ubuf, so we can prevent reloading */
856	struct io_mapped_ubuf		*imu;
857};
858
859struct io_tctx_node {
860	struct list_head	ctx_node;
861	struct task_struct	*task;
862	struct io_ring_ctx	*ctx;
863};
864
865struct io_defer_entry {
866	struct list_head	list;
867	struct io_kiocb		*req;
868	u32			seq;
869};
870
871struct io_op_def {
872	/* needs req->file assigned */
873	unsigned		needs_file : 1;
874	/* hash wq insertion if file is a regular file */
875	unsigned		hash_reg_file : 1;
876	/* unbound wq insertion if file is a non-regular file */
877	unsigned		unbound_nonreg_file : 1;
878	/* opcode is not supported by this kernel */
879	unsigned		not_supported : 1;
880	/* set if opcode supports polled "wait" */
881	unsigned		pollin : 1;
882	unsigned		pollout : 1;
883	/* op supports buffer selection */
884	unsigned		buffer_select : 1;
885	/* do prep async if is going to be punted */
886	unsigned		needs_async_setup : 1;
887	/* should block plug */
888	unsigned		plug : 1;
889	/* size of async data needed, if any */
890	unsigned short		async_size;
891};
892
893static const struct io_op_def io_op_defs[] = {
894	[IORING_OP_NOP] = {},
895	[IORING_OP_READV] = {
896		.needs_file		= 1,
897		.unbound_nonreg_file	= 1,
898		.pollin			= 1,
899		.buffer_select		= 1,
900		.needs_async_setup	= 1,
901		.plug			= 1,
902		.async_size		= sizeof(struct io_async_rw),
903	},
904	[IORING_OP_WRITEV] = {
905		.needs_file		= 1,
906		.hash_reg_file		= 1,
907		.unbound_nonreg_file	= 1,
908		.pollout		= 1,
909		.needs_async_setup	= 1,
910		.plug			= 1,
911		.async_size		= sizeof(struct io_async_rw),
912	},
913	[IORING_OP_FSYNC] = {
914		.needs_file		= 1,
915	},
916	[IORING_OP_READ_FIXED] = {
917		.needs_file		= 1,
918		.unbound_nonreg_file	= 1,
919		.pollin			= 1,
920		.plug			= 1,
921		.async_size		= sizeof(struct io_async_rw),
922	},
923	[IORING_OP_WRITE_FIXED] = {
924		.needs_file		= 1,
925		.hash_reg_file		= 1,
926		.unbound_nonreg_file	= 1,
927		.pollout		= 1,
928		.plug			= 1,
929		.async_size		= sizeof(struct io_async_rw),
930	},
931	[IORING_OP_POLL_ADD] = {
932		.needs_file		= 1,
933		.unbound_nonreg_file	= 1,
934	},
935	[IORING_OP_POLL_REMOVE] = {},
936	[IORING_OP_SYNC_FILE_RANGE] = {
937		.needs_file		= 1,
938	},
939	[IORING_OP_SENDMSG] = {
940		.needs_file		= 1,
941		.unbound_nonreg_file	= 1,
942		.pollout		= 1,
943		.needs_async_setup	= 1,
944		.async_size		= sizeof(struct io_async_msghdr),
945	},
946	[IORING_OP_RECVMSG] = {
947		.needs_file		= 1,
948		.unbound_nonreg_file	= 1,
949		.pollin			= 1,
950		.buffer_select		= 1,
951		.needs_async_setup	= 1,
952		.async_size		= sizeof(struct io_async_msghdr),
953	},
954	[IORING_OP_TIMEOUT] = {
955		.async_size		= sizeof(struct io_timeout_data),
956	},
957	[IORING_OP_TIMEOUT_REMOVE] = {
958		/* used by timeout updates' prep() */
959	},
960	[IORING_OP_ACCEPT] = {
961		.needs_file		= 1,
962		.unbound_nonreg_file	= 1,
963		.pollin			= 1,
964	},
965	[IORING_OP_ASYNC_CANCEL] = {},
966	[IORING_OP_LINK_TIMEOUT] = {
967		.async_size		= sizeof(struct io_timeout_data),
968	},
969	[IORING_OP_CONNECT] = {
970		.needs_file		= 1,
971		.unbound_nonreg_file	= 1,
972		.pollout		= 1,
973		.needs_async_setup	= 1,
974		.async_size		= sizeof(struct io_async_connect),
975	},
976	[IORING_OP_FALLOCATE] = {
977		.needs_file		= 1,
978	},
979	[IORING_OP_OPENAT] = {},
980	[IORING_OP_CLOSE] = {},
981	[IORING_OP_FILES_UPDATE] = {},
982	[IORING_OP_STATX] = {},
983	[IORING_OP_READ] = {
984		.needs_file		= 1,
985		.unbound_nonreg_file	= 1,
986		.pollin			= 1,
987		.buffer_select		= 1,
988		.plug			= 1,
989		.async_size		= sizeof(struct io_async_rw),
990	},
991	[IORING_OP_WRITE] = {
992		.needs_file		= 1,
993		.unbound_nonreg_file	= 1,
994		.pollout		= 1,
995		.plug			= 1,
996		.async_size		= sizeof(struct io_async_rw),
997	},
998	[IORING_OP_FADVISE] = {
999		.needs_file		= 1,
1000	},
1001	[IORING_OP_MADVISE] = {},
1002	[IORING_OP_SEND] = {
1003		.needs_file		= 1,
1004		.unbound_nonreg_file	= 1,
1005		.pollout		= 1,
1006	},
1007	[IORING_OP_RECV] = {
1008		.needs_file		= 1,
1009		.unbound_nonreg_file	= 1,
1010		.pollin			= 1,
1011		.buffer_select		= 1,
1012	},
1013	[IORING_OP_OPENAT2] = {
1014	},
1015	[IORING_OP_EPOLL_CTL] = {
1016		.unbound_nonreg_file	= 1,
1017	},
1018	[IORING_OP_SPLICE] = {
1019		.needs_file		= 1,
1020		.hash_reg_file		= 1,
1021		.unbound_nonreg_file	= 1,
1022	},
1023	[IORING_OP_PROVIDE_BUFFERS] = {},
1024	[IORING_OP_REMOVE_BUFFERS] = {},
1025	[IORING_OP_TEE] = {
1026		.needs_file		= 1,
1027		.hash_reg_file		= 1,
1028		.unbound_nonreg_file	= 1,
1029	},
1030	[IORING_OP_SHUTDOWN] = {
1031		.needs_file		= 1,
1032	},
1033	[IORING_OP_RENAMEAT] = {},
1034	[IORING_OP_UNLINKAT] = {},
1035};
1036
1037static bool io_disarm_next(struct io_kiocb *req);
1038static void io_uring_del_tctx_node(unsigned long index);
1039static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1040					 struct task_struct *task,
1041					 bool cancel_all);
1042static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1043
1044static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1045				 long res, unsigned int cflags);
1046static void io_put_req(struct io_kiocb *req);
1047static void io_put_req_deferred(struct io_kiocb *req);
1048static void io_dismantle_req(struct io_kiocb *req);
1049static void io_queue_linked_timeout(struct io_kiocb *req);
1050static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1051				     struct io_uring_rsrc_update2 *up,
1052				     unsigned nr_args);
1053static void io_clean_op(struct io_kiocb *req);
1054static struct file *io_file_get(struct io_ring_ctx *ctx,
1055				struct io_kiocb *req, int fd, bool fixed);
1056static void __io_queue_sqe(struct io_kiocb *req);
1057static void io_rsrc_put_work(struct work_struct *work);
1058
1059static void io_req_task_queue(struct io_kiocb *req);
1060static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1061static int io_req_prep_async(struct io_kiocb *req);
1062
1063static struct kmem_cache *req_cachep;
1064
1065static const struct file_operations io_uring_fops;
1066
1067struct sock *io_uring_get_socket(struct file *file)
1068{
1069#if defined(CONFIG_UNIX)
1070	if (file->f_op == &io_uring_fops) {
1071		struct io_ring_ctx *ctx = file->private_data;
1072
1073		return ctx->ring_sock->sk;
1074	}
1075#endif
1076	return NULL;
1077}
1078EXPORT_SYMBOL(io_uring_get_socket);
1079
1080#define io_for_each_link(pos, head) \
1081	for (pos = (head); pos; pos = pos->link)
1082
1083/*
1084 * Shamelessly stolen from the mm implementation of page reference checking,
1085 * see commit f958d7b528b1 for details.
1086 */
1087#define req_ref_zero_or_close_to_overflow(req)	\
1088	((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1089
1090static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1091{
1092	WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1093	return atomic_inc_not_zero(&req->refs);
1094}
1095
1096static inline bool req_ref_put_and_test(struct io_kiocb *req)
1097{
1098	if (likely(!(req->flags & REQ_F_REFCOUNT)))
1099		return true;
1100
1101	WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1102	return atomic_dec_and_test(&req->refs);
1103}
1104
1105static inline void req_ref_put(struct io_kiocb *req)
1106{
1107	WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1108	WARN_ON_ONCE(req_ref_put_and_test(req));
1109}
1110
1111static inline void req_ref_get(struct io_kiocb *req)
1112{
1113	WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1114	WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1115	atomic_inc(&req->refs);
1116}
1117
1118static inline void __io_req_set_refcount(struct io_kiocb *req, int nr)
1119{
1120	if (!(req->flags & REQ_F_REFCOUNT)) {
1121		req->flags |= REQ_F_REFCOUNT;
1122		atomic_set(&req->refs, nr);
1123	}
1124}
1125
1126static inline void io_req_set_refcount(struct io_kiocb *req)
1127{
1128	__io_req_set_refcount(req, 1);
1129}
1130
1131static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1132{
1133	struct io_ring_ctx *ctx = req->ctx;
1134
1135	if (!req->fixed_rsrc_refs) {
1136		req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1137		percpu_ref_get(req->fixed_rsrc_refs);
1138	}
1139}
1140
1141static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1142{
1143	bool got = percpu_ref_tryget(ref);
1144
1145	/* already at zero, wait for ->release() */
1146	if (!got)
1147		wait_for_completion(compl);
1148	percpu_ref_resurrect(ref);
1149	if (got)
1150		percpu_ref_put(ref);
1151}
1152
1153static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1154			  bool cancel_all)
1155{
1156	struct io_kiocb *req;
1157
1158	if (task && head->task != task)
1159		return false;
1160	if (cancel_all)
1161		return true;
1162
1163	io_for_each_link(req, head) {
1164		if (req->flags & REQ_F_INFLIGHT)
1165			return true;
1166	}
1167	return false;
1168}
1169
1170static inline void req_set_fail(struct io_kiocb *req)
1171{
1172	req->flags |= REQ_F_FAIL;
1173}
1174
1175static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1176{
1177	struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1178
1179	complete(&ctx->ref_comp);
1180}
1181
1182static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1183{
1184	return !req->timeout.off;
1185}
1186
1187static void io_fallback_req_func(struct work_struct *work)
1188{
1189	struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
1190						fallback_work.work);
1191	struct llist_node *node = llist_del_all(&ctx->fallback_llist);
1192	struct io_kiocb *req, *tmp;
1193
1194	percpu_ref_get(&ctx->refs);
1195	llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
1196		req->io_task_work.func(req);
1197	percpu_ref_put(&ctx->refs);
1198}
1199
1200static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1201{
1202	struct io_ring_ctx *ctx;
1203	int hash_bits;
1204
1205	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1206	if (!ctx)
1207		return NULL;
1208
1209	/*
1210	 * Use 5 bits less than the max cq entries, that should give us around
1211	 * 32 entries per hash list if totally full and uniformly spread.
1212	 */
1213	hash_bits = ilog2(p->cq_entries);
1214	hash_bits -= 5;
1215	if (hash_bits <= 0)
1216		hash_bits = 1;
1217	ctx->cancel_hash_bits = hash_bits;
1218	ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1219					GFP_KERNEL);
1220	if (!ctx->cancel_hash)
1221		goto err;
1222	__hash_init(ctx->cancel_hash, 1U << hash_bits);
1223
1224	ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1225	if (!ctx->dummy_ubuf)
1226		goto err;
1227	/* set invalid range, so io_import_fixed() fails meeting it */
1228	ctx->dummy_ubuf->ubuf = -1UL;
1229
1230	if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1231			    PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1232		goto err;
1233
1234	ctx->flags = p->flags;
1235	init_waitqueue_head(&ctx->sqo_sq_wait);
1236	INIT_LIST_HEAD(&ctx->sqd_list);
1237	init_waitqueue_head(&ctx->poll_wait);
1238	INIT_LIST_HEAD(&ctx->cq_overflow_list);
1239	init_completion(&ctx->ref_comp);
1240	xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1241	xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1242	mutex_init(&ctx->uring_lock);
1243	init_waitqueue_head(&ctx->cq_wait);
1244	spin_lock_init(&ctx->completion_lock);
1245	spin_lock_init(&ctx->timeout_lock);
1246	INIT_LIST_HEAD(&ctx->iopoll_list);
1247	INIT_LIST_HEAD(&ctx->defer_list);
1248	INIT_LIST_HEAD(&ctx->timeout_list);
1249	spin_lock_init(&ctx->rsrc_ref_lock);
1250	INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1251	INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1252	init_llist_head(&ctx->rsrc_put_llist);
1253	INIT_LIST_HEAD(&ctx->tctx_list);
1254	INIT_LIST_HEAD(&ctx->submit_state.free_list);
1255	INIT_LIST_HEAD(&ctx->locked_free_list);
1256	INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1257	return ctx;
1258err:
1259	kfree(ctx->dummy_ubuf);
1260	kfree(ctx->cancel_hash);
1261	kfree(ctx);
1262	return NULL;
1263}
1264
1265static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1266{
1267	struct io_rings *r = ctx->rings;
1268
1269	WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1270	ctx->cq_extra--;
1271}
1272
1273static bool req_need_defer(struct io_kiocb *req, u32 seq)
1274{
1275	if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1276		struct io_ring_ctx *ctx = req->ctx;
1277
1278		return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1279	}
1280
1281	return false;
1282}
1283
1284#define FFS_ASYNC_READ		0x1UL
1285#define FFS_ASYNC_WRITE		0x2UL
1286#ifdef CONFIG_64BIT
1287#define FFS_ISREG		0x4UL
1288#else
1289#define FFS_ISREG		0x0UL
1290#endif
1291#define FFS_MASK		~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1292
1293static inline bool io_req_ffs_set(struct io_kiocb *req)
1294{
1295	return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE);
1296}
1297
1298static void io_req_track_inflight(struct io_kiocb *req)
1299{
1300	if (!(req->flags & REQ_F_INFLIGHT)) {
1301		req->flags |= REQ_F_INFLIGHT;
1302		atomic_inc(&current->io_uring->inflight_tracked);
1303	}
1304}
1305
1306static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
1307{
1308	req->flags &= ~REQ_F_ARM_LTIMEOUT;
1309	req->flags |= REQ_F_LINK_TIMEOUT;
1310
1311	/* linked timeouts should have two refs once prep'ed */
1312	io_req_set_refcount(req);
1313	__io_req_set_refcount(req->link, 2);
1314	return req->link;
1315}
1316
1317static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
1318{
1319	if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
1320		return NULL;
1321	return __io_prep_linked_timeout(req);
1322}
1323
1324static void io_prep_async_work(struct io_kiocb *req)
1325{
1326	const struct io_op_def *def = &io_op_defs[req->opcode];
1327	struct io_ring_ctx *ctx = req->ctx;
1328
1329	if (!(req->flags & REQ_F_CREDS)) {
1330		req->flags |= REQ_F_CREDS;
1331		req->creds = get_current_cred();
1332	}
1333
1334	req->work.list.next = NULL;
1335	req->work.flags = 0;
1336	if (req->flags & REQ_F_FORCE_ASYNC)
1337		req->work.flags |= IO_WQ_WORK_CONCURRENT;
1338
1339	if (req->flags & REQ_F_ISREG) {
1340		if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1341			io_wq_hash_work(&req->work, file_inode(req->file));
1342	} else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1343		if (def->unbound_nonreg_file)
1344			req->work.flags |= IO_WQ_WORK_UNBOUND;
1345	}
1346
1347	switch (req->opcode) {
1348	case IORING_OP_SPLICE:
1349	case IORING_OP_TEE:
1350		if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1351			req->work.flags |= IO_WQ_WORK_UNBOUND;
1352		break;
1353	}
1354}
1355
1356static void io_prep_async_link(struct io_kiocb *req)
1357{
1358	struct io_kiocb *cur;
1359
1360	if (req->flags & REQ_F_LINK_TIMEOUT) {
1361		struct io_ring_ctx *ctx = req->ctx;
1362
1363		spin_lock(&ctx->completion_lock);
1364		io_for_each_link(cur, req)
1365			io_prep_async_work(cur);
1366		spin_unlock(&ctx->completion_lock);
1367	} else {
1368		io_for_each_link(cur, req)
1369			io_prep_async_work(cur);
1370	}
1371}
1372
1373static void io_queue_async_work(struct io_kiocb *req)
1374{
1375	struct io_ring_ctx *ctx = req->ctx;
1376	struct io_kiocb *link = io_prep_linked_timeout(req);
1377	struct io_uring_task *tctx = req->task->io_uring;
1378
1379	BUG_ON(!tctx);
1380	BUG_ON(!tctx->io_wq);
1381
1382	/* init ->work of the whole link before punting */
1383	io_prep_async_link(req);
1384
1385	/*
1386	 * Not expected to happen, but if we do have a bug where this _can_
1387	 * happen, catch it here and ensure the request is marked as
1388	 * canceled. That will make io-wq go through the usual work cancel
1389	 * procedure rather than attempt to run this request (or create a new
1390	 * worker for it).
1391	 */
1392	if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1393		req->work.flags |= IO_WQ_WORK_CANCEL;
1394
1395	trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1396					&req->work, req->flags);
1397	io_wq_enqueue(tctx->io_wq, &req->work);
1398	if (link)
1399		io_queue_linked_timeout(link);
1400}
1401
1402static void io_kill_timeout(struct io_kiocb *req, int status)
1403	__must_hold(&req->ctx->completion_lock)
1404	__must_hold(&req->ctx->timeout_lock)
1405{
1406	struct io_timeout_data *io = req->async_data;
1407
1408	if (hrtimer_try_to_cancel(&io->timer) != -1) {
1409		atomic_set(&req->ctx->cq_timeouts,
1410			atomic_read(&req->ctx->cq_timeouts) + 1);
1411		list_del_init(&req->timeout.list);
1412		io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1413		io_put_req_deferred(req);
1414	}
1415}
1416
1417static void io_queue_deferred(struct io_ring_ctx *ctx)
1418{
1419	while (!list_empty(&ctx->defer_list)) {
1420		struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1421						struct io_defer_entry, list);
1422
1423		if (req_need_defer(de->req, de->seq))
1424			break;
1425		list_del_init(&de->list);
1426		io_req_task_queue(de->req);
1427		kfree(de);
1428	}
1429}
1430
1431static void io_flush_timeouts(struct io_ring_ctx *ctx)
1432	__must_hold(&ctx->completion_lock)
1433{
1434	u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1435
1436	spin_lock_irq(&ctx->timeout_lock);
1437	while (!list_empty(&ctx->timeout_list)) {
1438		u32 events_needed, events_got;
1439		struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1440						struct io_kiocb, timeout.list);
1441
1442		if (io_is_timeout_noseq(req))
1443			break;
1444
1445		/*
1446		 * Since seq can easily wrap around over time, subtract
1447		 * the last seq at which timeouts were flushed before comparing.
1448		 * Assuming not more than 2^31-1 events have happened since,
1449		 * these subtractions won't have wrapped, so we can check if
1450		 * target is in [last_seq, current_seq] by comparing the two.
1451		 */
1452		events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1453		events_got = seq - ctx->cq_last_tm_flush;
1454		if (events_got < events_needed)
1455			break;
1456
1457		list_del_init(&req->timeout.list);
1458		io_kill_timeout(req, 0);
1459	}
1460	ctx->cq_last_tm_flush = seq;
1461	spin_unlock_irq(&ctx->timeout_lock);
1462}
1463
1464static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1465{
1466	if (ctx->off_timeout_used)
1467		io_flush_timeouts(ctx);
1468	if (ctx->drain_active)
1469		io_queue_deferred(ctx);
1470}
1471
1472static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1473{
1474	if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1475		__io_commit_cqring_flush(ctx);
1476	/* order cqe stores with ring update */
1477	smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1478}
1479
1480static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1481{
1482	struct io_rings *r = ctx->rings;
1483
1484	return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1485}
1486
1487static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1488{
1489	return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1490}
1491
1492static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1493{
1494	struct io_rings *rings = ctx->rings;
1495	unsigned tail, mask = ctx->cq_entries - 1;
1496
1497	/*
1498	 * writes to the cq entry need to come after reading head; the
1499	 * control dependency is enough as we're using WRITE_ONCE to
1500	 * fill the cq entry
1501	 */
1502	if (__io_cqring_events(ctx) == ctx->cq_entries)
1503		return NULL;
1504
1505	tail = ctx->cached_cq_tail++;
1506	return &rings->cqes[tail & mask];
1507}
1508
1509static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1510{
1511	if (likely(!ctx->cq_ev_fd))
1512		return false;
1513	if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1514		return false;
1515	return !ctx->eventfd_async || io_wq_current_is_worker();
1516}
1517
1518static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1519{
1520	/*
1521	 * wake_up_all() may seem excessive, but io_wake_function() and
1522	 * io_should_wake() handle the termination of the loop and only
1523	 * wake as many waiters as we need to.
1524	 */
1525	if (wq_has_sleeper(&ctx->cq_wait))
1526		wake_up_all(&ctx->cq_wait);
1527	if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1528		wake_up(&ctx->sq_data->wait);
1529	if (io_should_trigger_evfd(ctx))
1530		eventfd_signal(ctx->cq_ev_fd, 1);
1531	if (waitqueue_active(&ctx->poll_wait)) {
1532		wake_up_interruptible(&ctx->poll_wait);
1533		kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1534	}
1535}
1536
1537static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1538{
1539	if (ctx->flags & IORING_SETUP_SQPOLL) {
1540		if (wq_has_sleeper(&ctx->cq_wait))
1541			wake_up_all(&ctx->cq_wait);
1542	}
1543	if (io_should_trigger_evfd(ctx))
1544		eventfd_signal(ctx->cq_ev_fd, 1);
1545	if (waitqueue_active(&ctx->poll_wait)) {
1546		wake_up_interruptible(&ctx->poll_wait);
1547		kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1548	}
1549}
1550
1551/* Returns true if there are no backlogged entries after the flush */
1552static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1553{
1554	bool all_flushed, posted;
1555
1556	if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1557		return false;
1558
1559	posted = false;
1560	spin_lock(&ctx->completion_lock);
1561	while (!list_empty(&ctx->cq_overflow_list)) {
1562		struct io_uring_cqe *cqe = io_get_cqe(ctx);
1563		struct io_overflow_cqe *ocqe;
1564
1565		if (!cqe && !force)
1566			break;
1567		ocqe = list_first_entry(&ctx->cq_overflow_list,
1568					struct io_overflow_cqe, list);
1569		if (cqe)
1570			memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1571		else
1572			io_account_cq_overflow(ctx);
1573
1574		posted = true;
1575		list_del(&ocqe->list);
1576		kfree(ocqe);
1577	}
1578
1579	all_flushed = list_empty(&ctx->cq_overflow_list);
1580	if (all_flushed) {
1581		clear_bit(0, &ctx->check_cq_overflow);
1582		WRITE_ONCE(ctx->rings->sq_flags,
1583			   ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1584	}
1585
1586	if (posted)
1587		io_commit_cqring(ctx);
1588	spin_unlock(&ctx->completion_lock);
1589	if (posted)
1590		io_cqring_ev_posted(ctx);
1591	return all_flushed;
1592}
1593
1594static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1595{
1596	bool ret = true;
1597
1598	if (test_bit(0, &ctx->check_cq_overflow)) {
1599		/* iopoll syncs against uring_lock, not completion_lock */
1600		if (ctx->flags & IORING_SETUP_IOPOLL)
1601			mutex_lock(&ctx->uring_lock);
1602		ret = __io_cqring_overflow_flush(ctx, false);
1603		if (ctx->flags & IORING_SETUP_IOPOLL)
1604			mutex_unlock(&ctx->uring_lock);
1605	}
1606
1607	return ret;
1608}
1609
1610/* must to be called somewhat shortly after putting a request */
1611static inline void io_put_task(struct task_struct *task, int nr)
1612{
1613	struct io_uring_task *tctx = task->io_uring;
1614
1615	percpu_counter_sub(&tctx->inflight, nr);
1616	if (unlikely(atomic_read(&tctx->in_idle)))
1617		wake_up(&tctx->wait);
1618	put_task_struct_many(task, nr);
1619}
1620
1621static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1622				     long res, unsigned int cflags)
1623{
1624	struct io_overflow_cqe *ocqe;
1625
1626	ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1627	if (!ocqe) {
1628		/*
1629		 * If we're in ring overflow flush mode, or in task cancel mode,
1630		 * or cannot allocate an overflow entry, then we need to drop it
1631		 * on the floor.
1632		 */
1633		io_account_cq_overflow(ctx);
1634		return false;
1635	}
1636	if (list_empty(&ctx->cq_overflow_list)) {
1637		set_bit(0, &ctx->check_cq_overflow);
1638		WRITE_ONCE(ctx->rings->sq_flags,
1639			   ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1640
1641	}
1642	ocqe->cqe.user_data = user_data;
1643	ocqe->cqe.res = res;
1644	ocqe->cqe.flags = cflags;
1645	list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1646	return true;
1647}
1648
1649static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1650					  long res, unsigned int cflags)
1651{
1652	struct io_uring_cqe *cqe;
1653
1654	trace_io_uring_complete(ctx, user_data, res, cflags);
1655
1656	/*
1657	 * If we can't get a cq entry, userspace overflowed the
1658	 * submission (by quite a lot). Increment the overflow count in
1659	 * the ring.
1660	 */
1661	cqe = io_get_cqe(ctx);
1662	if (likely(cqe)) {
1663		WRITE_ONCE(cqe->user_data, user_data);
1664		WRITE_ONCE(cqe->res, res);
1665		WRITE_ONCE(cqe->flags, cflags);
1666		return true;
1667	}
1668	return io_cqring_event_overflow(ctx, user_data, res, cflags);
1669}
1670
1671/* not as hot to bloat with inlining */
1672static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1673					  long res, unsigned int cflags)
1674{
1675	return __io_cqring_fill_event(ctx, user_data, res, cflags);
1676}
1677
1678static void io_req_complete_post(struct io_kiocb *req, long res,
1679				 unsigned int cflags)
1680{
1681	struct io_ring_ctx *ctx = req->ctx;
1682
1683	spin_lock(&ctx->completion_lock);
1684	__io_cqring_fill_event(ctx, req->user_data, res, cflags);
1685	/*
1686	 * If we're the last reference to this request, add to our locked
1687	 * free_list cache.
1688	 */
1689	if (req_ref_put_and_test(req)) {
1690		if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1691			if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL))
1692				io_disarm_next(req);
1693			if (req->link) {
1694				io_req_task_queue(req->link);
1695				req->link = NULL;
1696			}
1697		}
1698		io_dismantle_req(req);
1699		io_put_task(req->task, 1);
1700		list_add(&req->inflight_entry, &ctx->locked_free_list);
1701		ctx->locked_free_nr++;
1702	} else {
1703		if (!percpu_ref_tryget(&ctx->refs))
1704			req = NULL;
1705	}
1706	io_commit_cqring(ctx);
1707	spin_unlock(&ctx->completion_lock);
1708
1709	if (req) {
1710		io_cqring_ev_posted(ctx);
1711		percpu_ref_put(&ctx->refs);
1712	}
1713}
1714
1715static inline bool io_req_needs_clean(struct io_kiocb *req)
1716{
1717	return req->flags & IO_REQ_CLEAN_FLAGS;
1718}
1719
1720static void io_req_complete_state(struct io_kiocb *req, long res,
1721				  unsigned int cflags)
1722{
1723	if (io_req_needs_clean(req))
1724		io_clean_op(req);
1725	req->result = res;
1726	req->compl.cflags = cflags;
1727	req->flags |= REQ_F_COMPLETE_INLINE;
1728}
1729
1730static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1731				     long res, unsigned cflags)
1732{
1733	if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1734		io_req_complete_state(req, res, cflags);
1735	else
1736		io_req_complete_post(req, res, cflags);
1737}
1738
1739static inline void io_req_complete(struct io_kiocb *req, long res)
1740{
1741	__io_req_complete(req, 0, res, 0);
1742}
1743
1744static void io_req_complete_failed(struct io_kiocb *req, long res)
1745{
1746	req_set_fail(req);
1747	io_req_complete_post(req, res, 0);
1748}
1749
1750/*
1751 * Don't initialise the fields below on every allocation, but do that in
1752 * advance and keep them valid across allocations.
1753 */
1754static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1755{
1756	req->ctx = ctx;
1757	req->link = NULL;
1758	req->async_data = NULL;
1759	/* not necessary, but safer to zero */
1760	req->result = 0;
1761}
1762
1763static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1764					struct io_submit_state *state)
1765{
1766	spin_lock(&ctx->completion_lock);
1767	list_splice_init(&ctx->locked_free_list, &state->free_list);
1768	ctx->locked_free_nr = 0;
1769	spin_unlock(&ctx->completion_lock);
1770}
1771
1772/* Returns true IFF there are requests in the cache */
1773static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1774{
1775	struct io_submit_state *state = &ctx->submit_state;
1776	int nr;
1777
1778	/*
1779	 * If we have more than a batch's worth of requests in our IRQ side
1780	 * locked cache, grab the lock and move them over to our submission
1781	 * side cache.
1782	 */
1783	if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1784		io_flush_cached_locked_reqs(ctx, state);
1785
1786	nr = state->free_reqs;
1787	while (!list_empty(&state->free_list)) {
1788		struct io_kiocb *req = list_first_entry(&state->free_list,
1789					struct io_kiocb, inflight_entry);
1790
1791		list_del(&req->inflight_entry);
1792		state->reqs[nr++] = req;
1793		if (nr == ARRAY_SIZE(state->reqs))
1794			break;
1795	}
1796
1797	state->free_reqs = nr;
1798	return nr != 0;
1799}
1800
1801/*
1802 * A request might get retired back into the request caches even before opcode
1803 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1804 * Because of that, io_alloc_req() should be called only under ->uring_lock
1805 * and with extra caution to not get a request that is still worked on.
1806 */
1807static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1808	__must_hold(&ctx->uring_lock)
1809{
1810	struct io_submit_state *state = &ctx->submit_state;
1811	gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1812	int ret, i;
1813
1814	BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1815
1816	if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1817		goto got_req;
1818
1819	ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1820				    state->reqs);
1821
1822	/*
1823	 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1824	 * retry single alloc to be on the safe side.
1825	 */
1826	if (unlikely(ret <= 0)) {
1827		state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1828		if (!state->reqs[0])
1829			return NULL;
1830		ret = 1;
1831	}
1832
1833	for (i = 0; i < ret; i++)
1834		io_preinit_req(state->reqs[i], ctx);
1835	state->free_reqs = ret;
1836got_req:
1837	state->free_reqs--;
1838	return state->reqs[state->free_reqs];
1839}
1840
1841static inline void io_put_file(struct file *file)
1842{
1843	if (file)
1844		fput(file);
1845}
1846
1847static void io_dismantle_req(struct io_kiocb *req)
1848{
1849	unsigned int flags = req->flags;
1850
1851	if (io_req_needs_clean(req))
1852		io_clean_op(req);
1853	if (!(flags & REQ_F_FIXED_FILE))
1854		io_put_file(req->file);
1855	if (req->fixed_rsrc_refs)
1856		percpu_ref_put(req->fixed_rsrc_refs);
1857	if (req->async_data) {
1858		kfree(req->async_data);
1859		req->async_data = NULL;
1860	}
1861}
1862
1863static void __io_free_req(struct io_kiocb *req)
1864{
1865	struct io_ring_ctx *ctx = req->ctx;
1866
1867	io_dismantle_req(req);
1868	io_put_task(req->task, 1);
1869
1870	spin_lock(&ctx->completion_lock);
1871	list_add(&req->inflight_entry, &ctx->locked_free_list);
1872	ctx->locked_free_nr++;
1873	spin_unlock(&ctx->completion_lock);
1874
1875	percpu_ref_put(&ctx->refs);
1876}
1877
1878static inline void io_remove_next_linked(struct io_kiocb *req)
1879{
1880	struct io_kiocb *nxt = req->link;
1881
1882	req->link = nxt->link;
1883	nxt->link = NULL;
1884}
1885
1886static bool io_kill_linked_timeout(struct io_kiocb *req)
1887	__must_hold(&req->ctx->completion_lock)
1888	__must_hold(&req->ctx->timeout_lock)
1889{
1890	struct io_kiocb *link = req->link;
1891
1892	if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
1893		struct io_timeout_data *io = link->async_data;
1894
1895		io_remove_next_linked(req);
1896		link->timeout.head = NULL;
1897		if (hrtimer_try_to_cancel(&io->timer) != -1) {
1898			io_cqring_fill_event(link->ctx, link->user_data,
1899					     -ECANCELED, 0);
1900			io_put_req_deferred(link);
1901			return true;
1902		}
1903	}
1904	return false;
1905}
1906
1907static void io_fail_links(struct io_kiocb *req)
1908	__must_hold(&req->ctx->completion_lock)
1909{
1910	struct io_kiocb *nxt, *link = req->link;
1911
1912	req->link = NULL;
1913	while (link) {
1914		nxt = link->link;
1915		link->link = NULL;
1916
1917		trace_io_uring_fail_link(req, link);
1918		io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1919		io_put_req_deferred(link);
1920		link = nxt;
1921	}
1922}
1923
1924static bool io_disarm_next(struct io_kiocb *req)
1925	__must_hold(&req->ctx->completion_lock)
1926{
1927	bool posted = false;
1928
1929	if (likely(req->flags & REQ_F_LINK_TIMEOUT)) {
1930		struct io_ring_ctx *ctx = req->ctx;
1931
1932		spin_lock_irq(&ctx->timeout_lock);
1933		posted = io_kill_linked_timeout(req);
1934		spin_unlock_irq(&ctx->timeout_lock);
1935	}
1936	if (unlikely((req->flags & REQ_F_FAIL) &&
1937		     !(req->flags & REQ_F_HARDLINK))) {
1938		posted |= (req->link != NULL);
1939		io_fail_links(req);
1940	}
1941	return posted;
1942}
1943
1944static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1945{
1946	struct io_kiocb *nxt;
1947
1948	/*
1949	 * If LINK is set, we have dependent requests in this chain. If we
1950	 * didn't fail this request, queue the first one up, moving any other
1951	 * dependencies to the next request. In case of failure, fail the rest
1952	 * of the chain.
1953	 */
1954	if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) {
1955		struct io_ring_ctx *ctx = req->ctx;
1956		bool posted;
1957
1958		spin_lock(&ctx->completion_lock);
1959		posted = io_disarm_next(req);
1960		if (posted)
1961			io_commit_cqring(req->ctx);
1962		spin_unlock(&ctx->completion_lock);
1963		if (posted)
1964			io_cqring_ev_posted(ctx);
1965	}
1966	nxt = req->link;
1967	req->link = NULL;
1968	return nxt;
1969}
1970
1971static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1972{
1973	if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1974		return NULL;
1975	return __io_req_find_next(req);
1976}
1977
1978static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1979{
1980	if (!ctx)
1981		return;
1982	if (ctx->submit_state.compl_nr) {
1983		mutex_lock(&ctx->uring_lock);
1984		io_submit_flush_completions(ctx);
1985		mutex_unlock(&ctx->uring_lock);
1986	}
1987	percpu_ref_put(&ctx->refs);
1988}
1989
1990static void tctx_task_work(struct callback_head *cb)
1991{
1992	struct io_ring_ctx *ctx = NULL;
1993	struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1994						  task_work);
1995
1996	while (1) {
1997		struct io_wq_work_node *node;
1998
1999		spin_lock_irq(&tctx->task_lock);
2000		node = tctx->task_list.first;
2001		INIT_WQ_LIST(&tctx->task_list);
2002		if (!node)
2003			tctx->task_running = false;
2004		spin_unlock_irq(&tctx->task_lock);
2005		if (!node)
2006			break;
2007
2008		do {
2009			struct io_wq_work_node *next = node->next;
2010			struct io_kiocb *req = container_of(node, struct io_kiocb,
2011							    io_task_work.node);
2012
2013			if (req->ctx != ctx) {
2014				ctx_flush_and_put(ctx);
2015				ctx = req->ctx;
2016				percpu_ref_get(&ctx->refs);
2017			}
2018			req->io_task_work.func(req);
2019			node = next;
2020		} while (node);
2021
2022		cond_resched();
2023	}
2024
2025	ctx_flush_and_put(ctx);
2026}
2027
2028static void io_req_task_work_add(struct io_kiocb *req)
2029{
2030	struct task_struct *tsk = req->task;
2031	struct io_uring_task *tctx = tsk->io_uring;
2032	enum task_work_notify_mode notify;
2033	struct io_wq_work_node *node;
2034	unsigned long flags;
2035	bool running;
2036
2037	WARN_ON_ONCE(!tctx);
2038
2039	spin_lock_irqsave(&tctx->task_lock, flags);
2040	wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2041	running = tctx->task_running;
2042	if (!running)
2043		tctx->task_running = true;
2044	spin_unlock_irqrestore(&tctx->task_lock, flags);
2045
2046	/* task_work already pending, we're done */
2047	if (running)
2048		return;
2049
2050	/*
2051	 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2052	 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2053	 * processing task_work. There's no reliable way to tell if TWA_RESUME
2054	 * will do the job.
2055	 */
2056	notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2057	if (!task_work_add(tsk, &tctx->task_work, notify)) {
2058		wake_up_process(tsk);
2059		return;
2060	}
2061
2062	spin_lock_irqsave(&tctx->task_lock, flags);
2063	tctx->task_running = false;
2064	node = tctx->task_list.first;
2065	INIT_WQ_LIST(&tctx->task_list);
2066	spin_unlock_irqrestore(&tctx->task_lock, flags);
2067
2068	while (node) {
2069		req = container_of(node, struct io_kiocb, io_task_work.node);
2070		node = node->next;
2071		if (llist_add(&req->io_task_work.fallback_node,
2072			      &req->ctx->fallback_llist))
2073			schedule_delayed_work(&req->ctx->fallback_work, 1);
2074	}
2075}
2076
2077static void io_req_task_cancel(struct io_kiocb *req)
2078{
2079	struct io_ring_ctx *ctx = req->ctx;
2080
2081	/* ctx is guaranteed to stay alive while we hold uring_lock */
2082	mutex_lock(&ctx->uring_lock);
2083	io_req_complete_failed(req, req->result);
2084	mutex_unlock(&ctx->uring_lock);
2085}
2086
2087static void io_req_task_submit(struct io_kiocb *req)
2088{
2089	struct io_ring_ctx *ctx = req->ctx;
2090
2091	/* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2092	mutex_lock(&ctx->uring_lock);
2093	if (likely(!(req->task->flags & PF_EXITING)))
2094		__io_queue_sqe(req);
2095	else
2096		io_req_complete_failed(req, -EFAULT);
2097	mutex_unlock(&ctx->uring_lock);
2098}
2099
2100static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2101{
2102	req->result = ret;
2103	req->io_task_work.func = io_req_task_cancel;
2104	io_req_task_work_add(req);
2105}
2106
2107static void io_req_task_queue(struct io_kiocb *req)
2108{
2109	req->io_task_work.func = io_req_task_submit;
2110	io_req_task_work_add(req);
2111}
2112
2113static void io_req_task_queue_reissue(struct io_kiocb *req)
2114{
2115	req->io_task_work.func = io_queue_async_work;
2116	io_req_task_work_add(req);
2117}
2118
2119static inline void io_queue_next(struct io_kiocb *req)
2120{
2121	struct io_kiocb *nxt = io_req_find_next(req);
2122
2123	if (nxt)
2124		io_req_task_queue(nxt);
2125}
2126
2127static void io_free_req(struct io_kiocb *req)
2128{
2129	io_queue_next(req);
2130	__io_free_req(req);
2131}
2132
2133struct req_batch {
2134	struct task_struct	*task;
2135	int			task_refs;
2136	int			ctx_refs;
2137};
2138
2139static inline void io_init_req_batch(struct req_batch *rb)
2140{
2141	rb->task_refs = 0;
2142	rb->ctx_refs = 0;
2143	rb->task = NULL;
2144}
2145
2146static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2147				     struct req_batch *rb)
2148{
2149	if (rb->ctx_refs)
2150		percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2151	if (rb->task == current)
2152		current->io_uring->cached_refs += rb->task_refs;
2153	else if (rb->task)
2154		io_put_task(rb->task, rb->task_refs);
2155}
2156
2157static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2158			      struct io_submit_state *state)
2159{
2160	io_queue_next(req);
2161	io_dismantle_req(req);
2162
2163	if (req->task != rb->task) {
2164		if (rb->task)
2165			io_put_task(rb->task, rb->task_refs);
2166		rb->task = req->task;
2167		rb->task_refs = 0;
2168	}
2169	rb->task_refs++;
2170	rb->ctx_refs++;
2171
2172	if (state->free_reqs != ARRAY_SIZE(state->reqs))
2173		state->reqs[state->free_reqs++] = req;
2174	else
2175		list_add(&req->inflight_entry, &state->free_list);
2176}
2177
2178static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2179	__must_hold(&ctx->uring_lock)
2180{
2181	struct io_submit_state *state = &ctx->submit_state;
2182	int i, nr = state->compl_nr;
2183	struct req_batch rb;
2184
2185	spin_lock(&ctx->completion_lock);
2186	for (i = 0; i < nr; i++) {
2187		struct io_kiocb *req = state->compl_reqs[i];
2188
2189		__io_cqring_fill_event(ctx, req->user_data, req->result,
2190					req->compl.cflags);
2191	}
2192	io_commit_cqring(ctx);
2193	spin_unlock(&ctx->completion_lock);
2194	io_cqring_ev_posted(ctx);
2195
2196	io_init_req_batch(&rb);
2197	for (i = 0; i < nr; i++) {
2198		struct io_kiocb *req = state->compl_reqs[i];
2199
2200		if (req_ref_put_and_test(req))
2201			io_req_free_batch(&rb, req, &ctx->submit_state);
2202	}
2203
2204	io_req_free_batch_finish(ctx, &rb);
2205	state->compl_nr = 0;
2206}
2207
2208/*
2209 * Drop reference to request, return next in chain (if there is one) if this
2210 * was the last reference to this request.
2211 */
2212static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2213{
2214	struct io_kiocb *nxt = NULL;
2215
2216	if (req_ref_put_and_test(req)) {
2217		nxt = io_req_find_next(req);
2218		__io_free_req(req);
2219	}
2220	return nxt;
2221}
2222
2223static inline void io_put_req(struct io_kiocb *req)
2224{
2225	if (req_ref_put_and_test(req))
2226		io_free_req(req);
2227}
2228
2229static inline void io_put_req_deferred(struct io_kiocb *req)
2230{
2231	if (req_ref_put_and_test(req)) {
2232		req->io_task_work.func = io_free_req;
2233		io_req_task_work_add(req);
2234	}
2235}
2236
2237static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2238{
2239	/* See comment at the top of this file */
2240	smp_rmb();
2241	return __io_cqring_events(ctx);
2242}
2243
2244static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2245{
2246	struct io_rings *rings = ctx->rings;
2247
2248	/* make sure SQ entry isn't read before tail */
2249	return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2250}
2251
2252static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2253{
2254	unsigned int cflags;
2255
2256	cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2257	cflags |= IORING_CQE_F_BUFFER;
2258	req->flags &= ~REQ_F_BUFFER_SELECTED;
2259	kfree(kbuf);
2260	return cflags;
2261}
2262
2263static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2264{
2265	struct io_buffer *kbuf;
2266
2267	kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2268	return io_put_kbuf(req, kbuf);
2269}
2270
2271static inline bool io_run_task_work(void)
2272{
2273	if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2274		__set_current_state(TASK_RUNNING);
2275		tracehook_notify_signal();
2276		return true;
2277	}
2278
2279	return false;
2280}
2281
2282/*
2283 * Find and free completed poll iocbs
2284 */
2285static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2286			       struct list_head *done)
2287{
2288	struct req_batch rb;
2289	struct io_kiocb *req;
2290
2291	/* order with ->result store in io_complete_rw_iopoll() */
2292	smp_rmb();
2293
2294	io_init_req_batch(&rb);
2295	while (!list_empty(done)) {
2296		int cflags = 0;
2297
2298		req = list_first_entry(done, struct io_kiocb, inflight_entry);
2299		list_del(&req->inflight_entry);
2300
2301		if (READ_ONCE(req->result) == -EAGAIN &&
2302		    !(req->flags & REQ_F_DONT_REISSUE)) {
2303			req->iopoll_completed = 0;
2304			io_req_task_queue_reissue(req);
2305			continue;
2306		}
2307
2308		if (req->flags & REQ_F_BUFFER_SELECTED)
2309			cflags = io_put_rw_kbuf(req);
2310
2311		__io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2312		(*nr_events)++;
2313
2314		if (req_ref_put_and_test(req))
2315			io_req_free_batch(&rb, req, &ctx->submit_state);
2316	}
2317
2318	io_commit_cqring(ctx);
2319	io_cqring_ev_posted_iopoll(ctx);
2320	io_req_free_batch_finish(ctx, &rb);
2321}
2322
2323static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2324			long min)
2325{
2326	struct io_kiocb *req, *tmp;
2327	LIST_HEAD(done);
2328	bool spin;
2329
2330	/*
2331	 * Only spin for completions if we don't have multiple devices hanging
2332	 * off our complete list, and we're under the requested amount.
2333	 */
2334	spin = !ctx->poll_multi_queue && *nr_events < min;
2335
2336	list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2337		struct kiocb *kiocb = &req->rw.kiocb;
2338		int ret;
2339
2340		/*
2341		 * Move completed and retryable entries to our local lists.
2342		 * If we find a request that requires polling, break out
2343		 * and complete those lists first, if we have entries there.
2344		 */
2345		if (READ_ONCE(req->iopoll_completed)) {
2346			list_move_tail(&req->inflight_entry, &done);
2347			continue;
2348		}
2349		if (!list_empty(&done))
2350			break;
2351
2352		ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2353		if (unlikely(ret < 0))
2354			return ret;
2355		else if (ret)
2356			spin = false;
2357
2358		/* iopoll may have completed current req */
2359		if (READ_ONCE(req->iopoll_completed))
2360			list_move_tail(&req->inflight_entry, &done);
2361	}
2362
2363	if (!list_empty(&done))
2364		io_iopoll_complete(ctx, nr_events, &done);
2365
2366	return 0;
2367}
2368
2369/*
2370 * We can't just wait for polled events to come to us, we have to actively
2371 * find and complete them.
2372 */
2373static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2374{
2375	if (!(ctx->flags & IORING_SETUP_IOPOLL))
2376		return;
2377
2378	mutex_lock(&ctx->uring_lock);
2379	while (!list_empty(&ctx->iopoll_list)) {
2380		unsigned int nr_events = 0;
2381
2382		io_do_iopoll(ctx, &nr_events, 0);
2383
2384		/* let it sleep and repeat later if can't complete a request */
2385		if (nr_events == 0)
2386			break;
2387		/*
2388		 * Ensure we allow local-to-the-cpu processing to take place,
2389		 * in this case we need to ensure that we reap all events.
2390		 * Also let task_work, etc. to progress by releasing the mutex
2391		 */
2392		if (need_resched()) {
2393			mutex_unlock(&ctx->uring_lock);
2394			cond_resched();
2395			mutex_lock(&ctx->uring_lock);
2396		}
2397	}
2398	mutex_unlock(&ctx->uring_lock);
2399}
2400
2401static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2402{
2403	unsigned int nr_events = 0;
2404	int ret = 0;
2405
2406	/*
2407	 * We disallow the app entering submit/complete with polling, but we
2408	 * still need to lock the ring to prevent racing with polled issue
2409	 * that got punted to a workqueue.
2410	 */
2411	mutex_lock(&ctx->uring_lock);
2412	/*
2413	 * Don't enter poll loop if we already have events pending.
2414	 * If we do, we can potentially be spinning for commands that
2415	 * already triggered a CQE (eg in error).
2416	 */
2417	if (test_bit(0, &ctx->check_cq_overflow))
2418		__io_cqring_overflow_flush(ctx, false);
2419	if (io_cqring_events(ctx))
2420		goto out;
2421	do {
2422		/*
2423		 * If a submit got punted to a workqueue, we can have the
2424		 * application entering polling for a command before it gets
2425		 * issued. That app will hold the uring_lock for the duration
2426		 * of the poll right here, so we need to take a breather every
2427		 * now and then to ensure that the issue has a chance to add
2428		 * the poll to the issued list. Otherwise we can spin here
2429		 * forever, while the workqueue is stuck trying to acquire the
2430		 * very same mutex.
2431		 */
2432		if (list_empty(&ctx->iopoll_list)) {
2433			u32 tail = ctx->cached_cq_tail;
2434
2435			mutex_unlock(&ctx->uring_lock);
2436			io_run_task_work();
2437			mutex_lock(&ctx->uring_lock);
2438
2439			/* some requests don't go through iopoll_list */
2440			if (tail != ctx->cached_cq_tail ||
2441			    list_empty(&ctx->iopoll_list))
2442				break;
2443		}
2444		ret = io_do_iopoll(ctx, &nr_events, min);
2445	} while (!ret && nr_events < min && !need_resched());
2446out:
2447	mutex_unlock(&ctx->uring_lock);
2448	return ret;
2449}
2450
2451static void kiocb_end_write(struct io_kiocb *req)
2452{
2453	/*
2454	 * Tell lockdep we inherited freeze protection from submission
2455	 * thread.
2456	 */
2457	if (req->flags & REQ_F_ISREG) {
2458		struct super_block *sb = file_inode(req->file)->i_sb;
2459
2460		__sb_writers_acquired(sb, SB_FREEZE_WRITE);
2461		sb_end_write(sb);
2462	}
2463}
2464
2465#ifdef CONFIG_BLOCK
2466static bool io_resubmit_prep(struct io_kiocb *req)
2467{
2468	struct io_async_rw *rw = req->async_data;
2469
2470	if (!rw)
2471		return !io_req_prep_async(req);
2472	/* may have left rw->iter inconsistent on -EIOCBQUEUED */
2473	iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2474	return true;
2475}
2476
2477static bool io_rw_should_reissue(struct io_kiocb *req)
2478{
2479	umode_t mode = file_inode(req->file)->i_mode;
2480	struct io_ring_ctx *ctx = req->ctx;
2481
2482	if (!S_ISBLK(mode) && !S_ISREG(mode))
2483		return false;
2484	if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2485	    !(ctx->flags & IORING_SETUP_IOPOLL)))
2486		return false;
2487	/*
2488	 * If ref is dying, we might be running poll reap from the exit work.
2489	 * Don't attempt to reissue from that path, just let it fail with
2490	 * -EAGAIN.
2491	 */
2492	if (percpu_ref_is_dying(&ctx->refs))
2493		return false;
2494	/*
2495	 * Play it safe and assume not safe to re-import and reissue if we're
2496	 * not in the original thread group (or in task context).
2497	 */
2498	if (!same_thread_group(req->task, current) || !in_task())
2499		return false;
2500	return true;
2501}
2502#else
2503static bool io_resubmit_prep(struct io_kiocb *req)
2504{
2505	return false;
2506}
2507static bool io_rw_should_reissue(struct io_kiocb *req)
2508{
2509	return false;
2510}
2511#endif
2512
2513static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2514{
2515	if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2516		kiocb_end_write(req);
2517	if (res != req->result) {
2518		if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2519		    io_rw_should_reissue(req)) {
2520			req->flags |= REQ_F_REISSUE;
2521			return true;
2522		}
2523		req_set_fail(req);
2524		req->result = res;
2525	}
2526	return false;
2527}
2528
2529static void io_req_task_complete(struct io_kiocb *req)
2530{
2531	int cflags = 0;
2532
2533	if (req->flags & REQ_F_BUFFER_SELECTED)
2534		cflags = io_put_rw_kbuf(req);
2535	__io_req_complete(req, 0, req->result, cflags);
2536}
2537
2538static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2539			     unsigned int issue_flags)
2540{
2541	if (__io_complete_rw_common(req, res))
2542		return;
2543	io_req_task_complete(req);
2544}
2545
2546static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2547{
2548	struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2549
2550	if (__io_complete_rw_common(req, res))
2551		return;
2552	req->result = res;
2553	req->io_task_work.func = io_req_task_complete;
2554	io_req_task_work_add(req);
2555}
2556
2557static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2558{
2559	struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2560
2561	if (kiocb->ki_flags & IOCB_WRITE)
2562		kiocb_end_write(req);
2563	if (unlikely(res != req->result)) {
2564		if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2565		    io_resubmit_prep(req))) {
2566			req_set_fail(req);
2567			req->flags |= REQ_F_DONT_REISSUE;
2568		}
2569	}
2570
2571	WRITE_ONCE(req->result, res);
2572	/* order with io_iopoll_complete() checking ->result */
2573	smp_wmb();
2574	WRITE_ONCE(req->iopoll_completed, 1);
2575}
2576
2577/*
2578 * After the iocb has been issued, it's safe to be found on the poll list.
2579 * Adding the kiocb to the list AFTER submission ensures that we don't
2580 * find it from a io_do_iopoll() thread before the issuer is done
2581 * accessing the kiocb cookie.
2582 */
2583static void io_iopoll_req_issued(struct io_kiocb *req)
2584{
2585	struct io_ring_ctx *ctx = req->ctx;
2586	const bool in_async = io_wq_current_is_worker();
2587
2588	/* workqueue context doesn't hold uring_lock, grab it now */
2589	if (unlikely(in_async))
2590		mutex_lock(&ctx->uring_lock);
2591
2592	/*
2593	 * Track whether we have multiple files in our lists. This will impact
2594	 * how we do polling eventually, not spinning if we're on potentially
2595	 * different devices.
2596	 */
2597	if (list_empty(&ctx->iopoll_list)) {
2598		ctx->poll_multi_queue = false;
2599	} else if (!ctx->poll_multi_queue) {
2600		struct io_kiocb *list_req;
2601		unsigned int queue_num0, queue_num1;
2602
2603		list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2604						inflight_entry);
2605
2606		if (list_req->file != req->file) {
2607			ctx->poll_multi_queue = true;
2608		} else {
2609			queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2610			queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2611			if (queue_num0 != queue_num1)
2612				ctx->poll_multi_queue = true;
2613		}
2614	}
2615
2616	/*
2617	 * For fast devices, IO may have already completed. If it has, add
2618	 * it to the front so we find it first.
2619	 */
2620	if (READ_ONCE(req->iopoll_completed))
2621		list_add(&req->inflight_entry, &ctx->iopoll_list);
2622	else
2623		list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2624
2625	if (unlikely(in_async)) {
2626		/*
2627		 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2628		 * in sq thread task context or in io worker task context. If
2629		 * current task context is sq thread, we don't need to check
2630		 * whether should wake up sq thread.
2631		 */
2632		if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2633		    wq_has_sleeper(&ctx->sq_data->wait))
2634			wake_up(&ctx->sq_data->wait);
2635
2636		mutex_unlock(&ctx->uring_lock);
2637	}
2638}
2639
2640static bool io_bdev_nowait(struct block_device *bdev)
2641{
2642	return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2643}
2644
2645/*
2646 * If we tracked the file through the SCM inflight mechanism, we could support
2647 * any file. For now, just ensure that anything potentially problematic is done
2648 * inline.
2649 */
2650static bool __io_file_supports_nowait(struct file *file, int rw)
2651{
2652	umode_t mode = file_inode(file)->i_mode;
2653
2654	if (S_ISBLK(mode)) {
2655		if (IS_ENABLED(CONFIG_BLOCK) &&
2656		    io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2657			return true;
2658		return false;
2659	}
2660	if (S_ISSOCK(mode))
2661		return true;
2662	if (S_ISREG(mode)) {
2663		if (IS_ENABLED(CONFIG_BLOCK) &&
2664		    io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2665		    file->f_op != &io_uring_fops)
2666			return true;
2667		return false;
2668	}
2669
2670	/* any ->read/write should understand O_NONBLOCK */
2671	if (file->f_flags & O_NONBLOCK)
2672		return true;
2673
2674	if (!(file->f_mode & FMODE_NOWAIT))
2675		return false;
2676
2677	if (rw == READ)
2678		return file->f_op->read_iter != NULL;
2679
2680	return file->f_op->write_iter != NULL;
2681}
2682
2683static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2684{
2685	if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2686		return true;
2687	else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2688		return true;
2689
2690	return __io_file_supports_nowait(req->file, rw);
2691}
2692
2693static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2694{
2695	struct io_ring_ctx *ctx = req->ctx;
2696	struct kiocb *kiocb = &req->rw.kiocb;
2697	struct file *file = req->file;
2698	unsigned ioprio;
2699	int ret;
2700
2701	if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2702		req->flags |= REQ_F_ISREG;
2703
2704	kiocb->ki_pos = READ_ONCE(sqe->off);
2705	if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2706		req->flags |= REQ_F_CUR_POS;
2707		kiocb->ki_pos = file->f_pos;
2708	}
2709	kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2710	kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2711	ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2712	if (unlikely(ret))
2713		return ret;
2714
2715	/* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2716	if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2717		req->flags |= REQ_F_NOWAIT;
2718
2719	ioprio = READ_ONCE(sqe->ioprio);
2720	if (ioprio) {
2721		ret = ioprio_check_cap(ioprio);
2722		if (ret)
2723			return ret;
2724
2725		kiocb->ki_ioprio = ioprio;
2726	} else
2727		kiocb->ki_ioprio = get_current_ioprio();
2728
2729	if (ctx->flags & IORING_SETUP_IOPOLL) {
2730		if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2731		    !kiocb->ki_filp->f_op->iopoll)
2732			return -EOPNOTSUPP;
2733
2734		kiocb->ki_flags |= IOCB_HIPRI;
2735		kiocb->ki_complete = io_complete_rw_iopoll;
2736		req->iopoll_completed = 0;
2737	} else {
2738		if (kiocb->ki_flags & IOCB_HIPRI)
2739			return -EINVAL;
2740		kiocb->ki_complete = io_complete_rw;
2741	}
2742
2743	if (req->opcode == IORING_OP_READ_FIXED ||
2744	    req->opcode == IORING_OP_WRITE_FIXED) {
2745		req->imu = NULL;
2746		io_req_set_rsrc_node(req);
2747	}
2748
2749	req->rw.addr = READ_ONCE(sqe->addr);
2750	req->rw.len = READ_ONCE(sqe->len);
2751	req->buf_index = READ_ONCE(sqe->buf_index);
2752	return 0;
2753}
2754
2755static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2756{
2757	switch (ret) {
2758	case -EIOCBQUEUED:
2759		break;
2760	case -ERESTARTSYS:
2761	case -ERESTARTNOINTR:
2762	case -ERESTARTNOHAND:
2763	case -ERESTART_RESTARTBLOCK:
2764		/*
2765		 * We can't just restart the syscall, since previously
2766		 * submitted sqes may already be in progress. Just fail this
2767		 * IO with EINTR.
2768		 */
2769		ret = -EINTR;
2770		fallthrough;
2771	default:
2772		kiocb->ki_complete(kiocb, ret, 0);
2773	}
2774}
2775
2776static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2777		       unsigned int issue_flags)
2778{
2779	struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2780	struct io_async_rw *io = req->async_data;
2781	bool check_reissue = kiocb->ki_complete == io_complete_rw;
2782
2783	/* add previously done IO, if any */
2784	if (io && io->bytes_done > 0) {
2785		if (ret < 0)
2786			ret = io->bytes_done;
2787		else
2788			ret += io->bytes_done;
2789	}
2790
2791	if (req->flags & REQ_F_CUR_POS)
2792		req->file->f_pos = kiocb->ki_pos;
2793	if (ret >= 0 && check_reissue)
2794		__io_complete_rw(req, ret, 0, issue_flags);
2795	else
2796		io_rw_done(kiocb, ret);
2797
2798	if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2799		req->flags &= ~REQ_F_REISSUE;
2800		if (io_resubmit_prep(req)) {
2801			io_req_task_queue_reissue(req);
2802		} else {
2803			int cflags = 0;
2804
2805			req_set_fail(req);
2806			if (req->flags & REQ_F_BUFFER_SELECTED)
2807				cflags = io_put_rw_kbuf(req);
2808			__io_req_complete(req, issue_flags, ret, cflags);
2809		}
2810	}
2811}
2812
2813static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2814			     struct io_mapped_ubuf *imu)
2815{
2816	size_t len = req->rw.len;
2817	u64 buf_end, buf_addr = req->rw.addr;
2818	size_t offset;
2819
2820	if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2821		return -EFAULT;
2822	/* not inside the mapped region */
2823	if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2824		return -EFAULT;
2825
2826	/*
2827	 * May not be a start of buffer, set size appropriately
2828	 * and advance us to the beginning.
2829	 */
2830	offset = buf_addr - imu->ubuf;
2831	iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2832
2833	if (offset) {
2834		/*
2835		 * Don't use iov_iter_advance() here, as it's really slow for
2836		 * using the latter parts of a big fixed buffer - it iterates
2837		 * over each segment manually. We can cheat a bit here, because
2838		 * we know that:
2839		 *
2840		 * 1) it's a BVEC iter, we set it up
2841		 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2842		 *    first and last bvec
2843		 *
2844		 * So just find our index, and adjust the iterator afterwards.
2845		 * If the offset is within the first bvec (or the whole first
2846		 * bvec, just use iov_iter_advance(). This makes it easier
2847		 * since we can just skip the first segment, which may not
2848		 * be PAGE_SIZE aligned.
2849		 */
2850		const struct bio_vec *bvec = imu->bvec;
2851
2852		if (offset <= bvec->bv_len) {
2853			iov_iter_advance(iter, offset);
2854		} else {
2855			unsigned long seg_skip;
2856
2857			/* skip first vec */
2858			offset -= bvec->bv_len;
2859			seg_skip = 1 + (offset >> PAGE_SHIFT);
2860
2861			iter->bvec = bvec + seg_skip;
2862			iter->nr_segs -= seg_skip;
2863			iter->count -= bvec->bv_len + offset;
2864			iter->iov_offset = offset & ~PAGE_MASK;
2865		}
2866	}
2867
2868	return 0;
2869}
2870
2871static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2872{
2873	struct io_ring_ctx *ctx = req->ctx;
2874	struct io_mapped_ubuf *imu = req->imu;
2875	u16 index, buf_index = req->buf_index;
2876
2877	if (likely(!imu)) {
2878		if (unlikely(buf_index >= ctx->nr_user_bufs))
2879			return -EFAULT;
2880		index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2881		imu = READ_ONCE(ctx->user_bufs[index]);
2882		req->imu = imu;
2883	}
2884	return __io_import_fixed(req, rw, iter, imu);
2885}
2886
2887static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2888{
2889	if (needs_lock)
2890		mutex_unlock(&ctx->uring_lock);
2891}
2892
2893static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2894{
2895	/*
2896	 * "Normal" inline submissions always hold the uring_lock, since we
2897	 * grab it from the system call. Same is true for the SQPOLL offload.
2898	 * The only exception is when we've detached the request and issue it
2899	 * from an async worker thread, grab the lock for that case.
2900	 */
2901	if (needs_lock)
2902		mutex_lock(&ctx->uring_lock);
2903}
2904
2905static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2906					  int bgid, struct io_buffer *kbuf,
2907					  bool needs_lock)
2908{
2909	struct io_buffer *head;
2910
2911	if (req->flags & REQ_F_BUFFER_SELECTED)
2912		return kbuf;
2913
2914	io_ring_submit_lock(req->ctx, needs_lock);
2915
2916	lockdep_assert_held(&req->ctx->uring_lock);
2917
2918	head = xa_load(&req->ctx->io_buffers, bgid);
2919	if (head) {
2920		if (!list_empty(&head->list)) {
2921			kbuf = list_last_entry(&head->list, struct io_buffer,
2922							list);
2923			list_del(&kbuf->list);
2924		} else {
2925			kbuf = head;
2926			xa_erase(&req->ctx->io_buffers, bgid);
2927		}
2928		if (*len > kbuf->len)
2929			*len = kbuf->len;
2930	} else {
2931		kbuf = ERR_PTR(-ENOBUFS);
2932	}
2933
2934	io_ring_submit_unlock(req->ctx, needs_lock);
2935
2936	return kbuf;
2937}
2938
2939static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2940					bool needs_lock)
2941{
2942	struct io_buffer *kbuf;
2943	u16 bgid;
2944
2945	kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2946	bgid = req->buf_index;
2947	kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2948	if (IS_ERR(kbuf))
2949		return kbuf;
2950	req->rw.addr = (u64) (unsigned long) kbuf;
2951	req->flags |= REQ_F_BUFFER_SELECTED;
2952	return u64_to_user_ptr(kbuf->addr);
2953}
2954
2955#ifdef CONFIG_COMPAT
2956static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2957				bool needs_lock)
2958{
2959	struct compat_iovec __user *uiov;
2960	compat_ssize_t clen;
2961	void __user *buf;
2962	ssize_t len;
2963
2964	uiov = u64_to_user_ptr(req->rw.addr);
2965	if (!access_ok(uiov, sizeof(*uiov)))
2966		return -EFAULT;
2967	if (__get_user(clen, &uiov->iov_len))
2968		return -EFAULT;
2969	if (clen < 0)
2970		return -EINVAL;
2971
2972	len = clen;
2973	buf = io_rw_buffer_select(req, &len, needs_lock);
2974	if (IS_ERR(buf))
2975		return PTR_ERR(buf);
2976	iov[0].iov_base = buf;
2977	iov[0].iov_len = (compat_size_t) len;
2978	return 0;
2979}
2980#endif
2981
2982static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2983				      bool needs_lock)
2984{
2985	struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2986	void __user *buf;
2987	ssize_t len;
2988
2989	if (copy_from_user(iov, uiov, sizeof(*uiov)))
2990		return -EFAULT;
2991
2992	len = iov[0].iov_len;
2993	if (len < 0)
2994		return -EINVAL;
2995	buf = io_rw_buffer_select(req, &len, needs_lock);
2996	if (IS_ERR(buf))
2997		return PTR_ERR(buf);
2998	iov[0].iov_base = buf;
2999	iov[0].iov_len = len;
3000	return 0;
3001}
3002
3003static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3004				    bool needs_lock)
3005{
3006	if (req->flags & REQ_F_BUFFER_SELECTED) {
3007		struct io_buffer *kbuf;
3008
3009		kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3010		iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3011		iov[0].iov_len = kbuf->len;
3012		return 0;
3013	}
3014	if (req->rw.len != 1)
3015		return -EINVAL;
3016
3017#ifdef CONFIG_COMPAT
3018	if (req->ctx->compat)
3019		return io_compat_import(req, iov, needs_lock);
3020#endif
3021
3022	return __io_iov_buffer_select(req, iov, needs_lock);
3023}
3024
3025static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3026			   struct iov_iter *iter, bool needs_lock)
3027{
3028	void __user *buf = u64_to_user_ptr(req->rw.addr);
3029	size_t sqe_len = req->rw.len;
3030	u8 opcode = req->opcode;
3031	ssize_t ret;
3032
3033	if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3034		*iovec = NULL;
3035		return io_import_fixed(req, rw, iter);
3036	}
3037
3038	/* buffer index only valid with fixed read/write, or buffer select  */
3039	if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3040		return -EINVAL;
3041
3042	if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3043		if (req->flags & REQ_F_BUFFER_SELECT) {
3044			buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3045			if (IS_ERR(buf))
3046				return PTR_ERR(buf);
3047			req->rw.len = sqe_len;
3048		}
3049
3050		ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3051		*iovec = NULL;
3052		return ret;
3053	}
3054
3055	if (req->flags & REQ_F_BUFFER_SELECT) {
3056		ret = io_iov_buffer_select(req, *iovec, needs_lock);
3057		if (!ret)
3058			iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3059		*iovec = NULL;
3060		return ret;
3061	}
3062
3063	return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3064			      req->ctx->compat);
3065}
3066
3067static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3068{
3069	return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3070}
3071
3072/*
3073 * For files that don't have ->read_iter() and ->write_iter(), handle them
3074 * by looping over ->read() or ->write() manually.
3075 */
3076static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3077{
3078	struct kiocb *kiocb = &req->rw.kiocb;
3079	struct file *file = req->file;
3080	ssize_t ret = 0;
3081
3082	/*
3083	 * Don't support polled IO through this interface, and we can't
3084	 * support non-blocking either. For the latter, this just causes
3085	 * the kiocb to be handled from an async context.
3086	 */
3087	if (kiocb->ki_flags & IOCB_HIPRI)
3088		return -EOPNOTSUPP;
3089	if (kiocb->ki_flags & IOCB_NOWAIT)
3090		return -EAGAIN;
3091
3092	while (iov_iter_count(iter)) {
3093		struct iovec iovec;
3094		ssize_t nr;
3095
3096		if (!iov_iter_is_bvec(iter)) {
3097			iovec = iov_iter_iovec(iter);
3098		} else {
3099			iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3100			iovec.iov_len = req->rw.len;
3101		}
3102
3103		if (rw == READ) {
3104			nr = file->f_op->read(file, iovec.iov_base,
3105					      iovec.iov_len, io_kiocb_ppos(kiocb));
3106		} else {
3107			nr = file->f_op->write(file, iovec.iov_base,
3108					       iovec.iov_len, io_kiocb_ppos(kiocb));
3109		}
3110
3111		if (nr < 0) {
3112			if (!ret)
3113				ret = nr;
3114			break;
3115		}
3116		ret += nr;
3117		if (nr != iovec.iov_len)
3118			break;
3119		req->rw.len -= nr;
3120		req->rw.addr += nr;
3121		iov_iter_advance(iter, nr);
3122	}
3123
3124	return ret;
3125}
3126
3127static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3128			  const struct iovec *fast_iov, struct iov_iter *iter)
3129{
3130	struct io_async_rw *rw = req->async_data;
3131
3132	memcpy(&rw->iter, iter, sizeof(*iter));
3133	rw->free_iovec = iovec;
3134	rw->bytes_done = 0;
3135	/* can only be fixed buffers, no need to do anything */
3136	if (iov_iter_is_bvec(iter))
3137		return;
3138	if (!iovec) {
3139		unsigned iov_off = 0;
3140
3141		rw->iter.iov = rw->fast_iov;
3142		if (iter->iov != fast_iov) {
3143			iov_off = iter->iov - fast_iov;
3144			rw->iter.iov += iov_off;
3145		}
3146		if (rw->fast_iov != fast_iov)
3147			memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3148			       sizeof(struct iovec) * iter->nr_segs);
3149	} else {
3150		req->flags |= REQ_F_NEED_CLEANUP;
3151	}
3152}
3153
3154static inline int io_alloc_async_data(struct io_kiocb *req)
3155{
3156	WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3157	req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3158	return req->async_data == NULL;
3159}
3160
3161static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3162			     const struct iovec *fast_iov,
3163			     struct iov_iter *iter, bool force)
3164{
3165	if (!force && !io_op_defs[req->opcode].needs_async_setup)
3166		return 0;
3167	if (!req->async_data) {
3168		if (io_alloc_async_data(req)) {
3169			kfree(iovec);
3170			return -ENOMEM;
3171		}
3172
3173		io_req_map_rw(req, iovec, fast_iov, iter);
3174	}
3175	return 0;
3176}
3177
3178static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3179{
3180	struct io_async_rw *iorw = req->async_data;
3181	struct iovec *iov = iorw->fast_iov;
3182	int ret;
3183
3184	ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3185	if (unlikely(ret < 0))
3186		return ret;
3187
3188	iorw->bytes_done = 0;
3189	iorw->free_iovec = iov;
3190	if (iov)
3191		req->flags |= REQ_F_NEED_CLEANUP;
3192	return 0;
3193}
3194
3195static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3196{
3197	if (unlikely(!(req->file->f_mode & FMODE_READ)))
3198		return -EBADF;
3199	return io_prep_rw(req, sqe);
3200}
3201
3202/*
3203 * This is our waitqueue callback handler, registered through lock_page_async()
3204 * when we initially tried to do the IO with the iocb armed our waitqueue.
3205 * This gets called when the page is unlocked, and we generally expect that to
3206 * happen when the page IO is completed and the page is now uptodate. This will
3207 * queue a task_work based retry of the operation, attempting to copy the data
3208 * again. If the latter fails because the page was NOT uptodate, then we will
3209 * do a thread based blocking retry of the operation. That's the unexpected
3210 * slow path.
3211 */
3212static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3213			     int sync, void *arg)
3214{
3215	struct wait_page_queue *wpq;
3216	struct io_kiocb *req = wait->private;
3217	struct wait_page_key *key = arg;
3218
3219	wpq = container_of(wait, struct wait_page_queue, wait);
3220
3221	if (!wake_page_match(wpq, key))
3222		return 0;
3223
3224	req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3225	list_del_init(&wait->entry);
3226	io_req_task_queue(req);
3227	return 1;
3228}
3229
3230/*
3231 * This controls whether a given IO request should be armed for async page
3232 * based retry. If we return false here, the request is handed to the async
3233 * worker threads for retry. If we're doing buffered reads on a regular file,
3234 * we prepare a private wait_page_queue entry and retry the operation. This
3235 * will either succeed because the page is now uptodate and unlocked, or it
3236 * will register a callback when the page is unlocked at IO completion. Through
3237 * that callback, io_uring uses task_work to setup a retry of the operation.
3238 * That retry will attempt the buffered read again. The retry will generally
3239 * succeed, or in rare cases where it fails, we then fall back to using the
3240 * async worker threads for a blocking retry.
3241 */
3242static bool io_rw_should_retry(struct io_kiocb *req)
3243{
3244	struct io_async_rw *rw = req->async_data;
3245	struct wait_page_queue *wait = &rw->wpq;
3246	struct kiocb *kiocb = &req->rw.kiocb;
3247
3248	/* never retry for NOWAIT, we just complete with -EAGAIN */
3249	if (req->flags & REQ_F_NOWAIT)
3250		return false;
3251
3252	/* Only for buffered IO */
3253	if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3254		return false;
3255
3256	/*
3257	 * just use poll if we can, and don't attempt if the fs doesn't
3258	 * support callback based unlocks
3259	 */
3260	if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3261		return false;
3262
3263	wait->wait.func = io_async_buf_func;
3264	wait->wait.private = req;
3265	wait->wait.flags = 0;
3266	INIT_LIST_HEAD(&wait->wait.entry);
3267	kiocb->ki_flags |= IOCB_WAITQ;
3268	kiocb->ki_flags &= ~IOCB_NOWAIT;
3269	kiocb->ki_waitq = wait;
3270	return true;
3271}
3272
3273static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3274{
3275	if (req->file->f_op->read_iter)
3276		return call_read_iter(req->file, &req->rw.kiocb, iter);
3277	else if (req->file->f_op->read)
3278		return loop_rw_iter(READ, req, iter);
3279	else
3280		return -EINVAL;
3281}
3282
3283static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3284{
3285	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3286	struct kiocb *kiocb = &req->rw.kiocb;
3287	struct iov_iter __iter, *iter = &__iter;
3288	struct io_async_rw *rw = req->async_data;
3289	ssize_t io_size, ret, ret2;
3290	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3291
3292	if (rw) {
3293		iter = &rw->iter;
3294		iovec = NULL;
3295	} else {
3296		ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3297		if (ret < 0)
3298			return ret;
3299	}
3300	io_size = iov_iter_count(iter);
3301	req->result = io_size;
3302
3303	/* Ensure we clear previously set non-block flag */
3304	if (!force_nonblock)
3305		kiocb->ki_flags &= ~IOCB_NOWAIT;
3306	else
3307		kiocb->ki_flags |= IOCB_NOWAIT;
3308
3309	/* If the file doesn't support async, just async punt */
3310	if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3311		ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3312		return ret ?: -EAGAIN;
3313	}
3314
3315	ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3316	if (unlikely(ret)) {
3317		kfree(iovec);
3318		return ret;
3319	}
3320
3321	ret = io_iter_do_read(req, iter);
3322
3323	if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3324		req->flags &= ~REQ_F_REISSUE;
3325		/* IOPOLL retry should happen for io-wq threads */
3326		if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3327			goto done;
3328		/* no retry on NONBLOCK nor RWF_NOWAIT */
3329		if (req->flags & REQ_F_NOWAIT)
3330			goto done;
3331		/* some cases will consume bytes even on error returns */
3332		iov_iter_revert(iter, io_size - iov_iter_count(iter));
3333		ret = 0;
3334	} else if (ret == -EIOCBQUEUED) {
3335		goto out_free;
3336	} else if (ret <= 0 || ret == io_size || !force_nonblock ||
3337		   (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3338		/* read all, failed, already did sync or don't want to retry */
3339		goto done;
3340	}
3341
3342	ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3343	if (ret2)
3344		return ret2;
3345
3346	iovec = NULL;
3347	rw = req->async_data;
3348	/* now use our persistent iterator, if we aren't already */
3349	iter = &rw->iter;
3350
3351	do {
3352		io_size -= ret;
3353		rw->bytes_done += ret;
3354		/* if we can retry, do so with the callbacks armed */
3355		if (!io_rw_should_retry(req)) {
3356			kiocb->ki_flags &= ~IOCB_WAITQ;
3357			return -EAGAIN;
3358		}
3359
3360		/*
3361		 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3362		 * we get -EIOCBQUEUED, then we'll get a notification when the
3363		 * desired page gets unlocked. We can also get a partial read
3364		 * here, and if we do, then just retry at the new offset.
3365		 */
3366		ret = io_iter_do_read(req, iter);
3367		if (ret == -EIOCBQUEUED)
3368			return 0;
3369		/* we got some bytes, but not all. retry. */
3370		kiocb->ki_flags &= ~IOCB_WAITQ;
3371	} while (ret > 0 && ret < io_size);
3372done:
3373	kiocb_done(kiocb, ret, issue_flags);
3374out_free:
3375	/* it's faster to check here then delegate to kfree */
3376	if (iovec)
3377		kfree(iovec);
3378	return 0;
3379}
3380
3381static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3382{
3383	if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3384		return -EBADF;
3385	return io_prep_rw(req, sqe);
3386}
3387
3388static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3389{
3390	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3391	struct kiocb *kiocb = &req->rw.kiocb;
3392	struct iov_iter __iter, *iter = &__iter;
3393	struct io_async_rw *rw = req->async_data;
3394	ssize_t ret, ret2, io_size;
3395	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3396
3397	if (rw) {
3398		iter = &rw->iter;
3399		iovec = NULL;
3400	} else {
3401		ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3402		if (ret < 0)
3403			return ret;
3404	}
3405	io_size = iov_iter_count(iter);
3406	req->result = io_size;
3407
3408	/* Ensure we clear previously set non-block flag */
3409	if (!force_nonblock)
3410		kiocb->ki_flags &= ~IOCB_NOWAIT;
3411	else
3412		kiocb->ki_flags |= IOCB_NOWAIT;
3413
3414	/* If the file doesn't support async, just async punt */
3415	if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3416		goto copy_iov;
3417
3418	/* file path doesn't support NOWAIT for non-direct_IO */
3419	if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3420	    (req->flags & REQ_F_ISREG))
3421		goto copy_iov;
3422
3423	ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3424	if (unlikely(ret))
3425		goto out_free;
3426
3427	/*
3428	 * Open-code file_start_write here to grab freeze protection,
3429	 * which will be released by another thread in
3430	 * io_complete_rw().  Fool lockdep by telling it the lock got
3431	 * released so that it doesn't complain about the held lock when
3432	 * we return to userspace.
3433	 */
3434	if (req->flags & REQ_F_ISREG) {
3435		sb_start_write(file_inode(req->file)->i_sb);
3436		__sb_writers_release(file_inode(req->file)->i_sb,
3437					SB_FREEZE_WRITE);
3438	}
3439	kiocb->ki_flags |= IOCB_WRITE;
3440
3441	if (req->file->f_op->write_iter)
3442		ret2 = call_write_iter(req->file, kiocb, iter);
3443	else if (req->file->f_op->write)
3444		ret2 = loop_rw_iter(WRITE, req, iter);
3445	else
3446		ret2 = -EINVAL;
3447
3448	if (req->flags & REQ_F_REISSUE) {
3449		req->flags &= ~REQ_F_REISSUE;
3450		ret2 = -EAGAIN;
3451	}
3452
3453	/*
3454	 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3455	 * retry them without IOCB_NOWAIT.
3456	 */
3457	if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3458		ret2 = -EAGAIN;
3459	/* no retry on NONBLOCK nor RWF_NOWAIT */
3460	if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3461		goto done;
3462	if (!force_nonblock || ret2 != -EAGAIN) {
3463		/* IOPOLL retry should happen for io-wq threads */
3464		if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3465			goto copy_iov;
3466done:
3467		kiocb_done(kiocb, ret2, issue_flags);
3468	} else {
3469copy_iov:
3470		/* some cases will consume bytes even on error returns */
3471		iov_iter_revert(iter, io_size - iov_iter_count(iter));
3472		ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3473		return ret ?: -EAGAIN;
3474	}
3475out_free:
3476	/* it's reportedly faster than delegating the null check to kfree() */
3477	if (iovec)
3478		kfree(iovec);
3479	return ret;
3480}
3481
3482static int io_renameat_prep(struct io_kiocb *req,
3483			    const struct io_uring_sqe *sqe)
3484{
3485	struct io_rename *ren = &req->rename;
3486	const char __user *oldf, *newf;
3487
3488	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3489		return -EINVAL;
3490	if (sqe->ioprio || sqe->buf_index)
3491		return -EINVAL;
3492	if (unlikely(req->flags & REQ_F_FIXED_FILE))
3493		return -EBADF;
3494
3495	ren->old_dfd = READ_ONCE(sqe->fd);
3496	oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3497	newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3498	ren->new_dfd = READ_ONCE(sqe->len);
3499	ren->flags = READ_ONCE(sqe->rename_flags);
3500
3501	ren->oldpath = getname(oldf);
3502	if (IS_ERR(ren->oldpath))
3503		return PTR_ERR(ren->oldpath);
3504
3505	ren->newpath = getname(newf);
3506	if (IS_ERR(ren->newpath)) {
3507		putname(ren->oldpath);
3508		return PTR_ERR(ren->newpath);
3509	}
3510
3511	req->flags |= REQ_F_NEED_CLEANUP;
3512	return 0;
3513}
3514
3515static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3516{
3517	struct io_rename *ren = &req->rename;
3518	int ret;
3519
3520	if (issue_flags & IO_URING_F_NONBLOCK)
3521		return -EAGAIN;
3522
3523	ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3524				ren->newpath, ren->flags);
3525
3526	req->flags &= ~REQ_F_NEED_CLEANUP;
3527	if (ret < 0)
3528		req_set_fail(req);
3529	io_req_complete(req, ret);
3530	return 0;
3531}
3532
3533static int io_unlinkat_prep(struct io_kiocb *req,
3534			    const struct io_uring_sqe *sqe)
3535{
3536	struct io_unlink *un = &req->unlink;
3537	const char __user *fname;
3538
3539	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3540		return -EINVAL;
3541	if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3542		return -EINVAL;
3543	if (unlikely(req->flags & REQ_F_FIXED_FILE))
3544		return -EBADF;
3545
3546	un->dfd = READ_ONCE(sqe->fd);
3547
3548	un->flags = READ_ONCE(sqe->unlink_flags);
3549	if (un->flags & ~AT_REMOVEDIR)
3550		return -EINVAL;
3551
3552	fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3553	un->filename = getname(fname);
3554	if (IS_ERR(un->filename))
3555		return PTR_ERR(un->filename);
3556
3557	req->flags |= REQ_F_NEED_CLEANUP;
3558	return 0;
3559}
3560
3561static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3562{
3563	struct io_unlink *un = &req->unlink;
3564	int ret;
3565
3566	if (issue_flags & IO_URING_F_NONBLOCK)
3567		return -EAGAIN;
3568
3569	if (un->flags & AT_REMOVEDIR)
3570		ret = do_rmdir(un->dfd, un->filename);
3571	else
3572		ret = do_unlinkat(un->dfd, un->filename);
3573
3574	req->flags &= ~REQ_F_NEED_CLEANUP;
3575	if (ret < 0)
3576		req_set_fail(req);
3577	io_req_complete(req, ret);
3578	return 0;
3579}
3580
3581static int io_shutdown_prep(struct io_kiocb *req,
3582			    const struct io_uring_sqe *sqe)
3583{
3584#if defined(CONFIG_NET)
3585	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3586		return -EINVAL;
3587	if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3588	    sqe->buf_index)
3589		return -EINVAL;
3590
3591	req->shutdown.how = READ_ONCE(sqe->len);
3592	return 0;
3593#else
3594	return -EOPNOTSUPP;
3595#endif
3596}
3597
3598static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3599{
3600#if defined(CONFIG_NET)
3601	struct socket *sock;
3602	int ret;
3603
3604	if (issue_flags & IO_URING_F_NONBLOCK)
3605		return -EAGAIN;
3606
3607	sock = sock_from_file(req->file);
3608	if (unlikely(!sock))
3609		return -ENOTSOCK;
3610
3611	ret = __sys_shutdown_sock(sock, req->shutdown.how);
3612	if (ret < 0)
3613		req_set_fail(req);
3614	io_req_complete(req, ret);
3615	return 0;
3616#else
3617	return -EOPNOTSUPP;
3618#endif
3619}
3620
3621static int __io_splice_prep(struct io_kiocb *req,
3622			    const struct io_uring_sqe *sqe)
3623{
3624	struct io_splice *sp = &req->splice;
3625	unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3626
3627	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3628		return -EINVAL;
3629
3630	sp->file_in = NULL;
3631	sp->len = READ_ONCE(sqe->len);
3632	sp->flags = READ_ONCE(sqe->splice_flags);
3633
3634	if (unlikely(sp->flags & ~valid_flags))
3635		return -EINVAL;
3636
3637	sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3638				  (sp->flags & SPLICE_F_FD_IN_FIXED));
3639	if (!sp->file_in)
3640		return -EBADF;
3641	req->flags |= REQ_F_NEED_CLEANUP;
3642	return 0;
3643}
3644
3645static int io_tee_prep(struct io_kiocb *req,
3646		       const struct io_uring_sqe *sqe)
3647{
3648	if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3649		return -EINVAL;
3650	return __io_splice_prep(req, sqe);
3651}
3652
3653static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3654{
3655	struct io_splice *sp = &req->splice;
3656	struct file *in = sp->file_in;
3657	struct file *out = sp->file_out;
3658	unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3659	long ret = 0;
3660
3661	if (issue_flags & IO_URING_F_NONBLOCK)
3662		return -EAGAIN;
3663	if (sp->len)
3664		ret = do_tee(in, out, sp->len, flags);
3665
3666	if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3667		io_put_file(in);
3668	req->flags &= ~REQ_F_NEED_CLEANUP;
3669
3670	if (ret != sp->len)
3671		req_set_fail(req);
3672	io_req_complete(req, ret);
3673	return 0;
3674}
3675
3676static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3677{
3678	struct io_splice *sp = &req->splice;
3679
3680	sp->off_in = READ_ONCE(sqe->splice_off_in);
3681	sp->off_out = READ_ONCE(sqe->off);
3682	return __io_splice_prep(req, sqe);
3683}
3684
3685static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3686{
3687	struct io_splice *sp = &req->splice;
3688	struct file *in = sp->file_in;
3689	struct file *out = sp->file_out;
3690	unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3691	loff_t *poff_in, *poff_out;
3692	long ret = 0;
3693
3694	if (issue_flags & IO_URING_F_NONBLOCK)
3695		return -EAGAIN;
3696
3697	poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3698	poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3699
3700	if (sp->len)
3701		ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3702
3703	if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3704		io_put_file(in);
3705	req->flags &= ~REQ_F_NEED_CLEANUP;
3706
3707	if (ret != sp->len)
3708		req_set_fail(req);
3709	io_req_complete(req, ret);
3710	return 0;
3711}
3712
3713/*
3714 * IORING_OP_NOP just posts a completion event, nothing else.
3715 */
3716static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3717{
3718	struct io_ring_ctx *ctx = req->ctx;
3719
3720	if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3721		return -EINVAL;
3722
3723	__io_req_complete(req, issue_flags, 0, 0);
3724	return 0;
3725}
3726
3727static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3728{
3729	struct io_ring_ctx *ctx = req->ctx;
3730
3731	if (!req->file)
3732		return -EBADF;
3733
3734	if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3735		return -EINVAL;
3736	if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3737		return -EINVAL;
3738
3739	req->sync.flags = READ_ONCE(sqe->fsync_flags);
3740	if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3741		return -EINVAL;
3742
3743	req->sync.off = READ_ONCE(sqe->off);
3744	req->sync.len = READ_ONCE(sqe->len);
3745	return 0;
3746}
3747
3748static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3749{
3750	loff_t end = req->sync.off + req->sync.len;
3751	int ret;
3752
3753	/* fsync always requires a blocking context */
3754	if (issue_flags & IO_URING_F_NONBLOCK)
3755		return -EAGAIN;
3756
3757	ret = vfs_fsync_range(req->file, req->sync.off,
3758				end > 0 ? end : LLONG_MAX,
3759				req->sync.flags & IORING_FSYNC_DATASYNC);
3760	if (ret < 0)
3761		req_set_fail(req);
3762	io_req_complete(req, ret);
3763	return 0;
3764}
3765
3766static int io_fallocate_prep(struct io_kiocb *req,
3767			     const struct io_uring_sqe *sqe)
3768{
3769	if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3770		return -EINVAL;
3771	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3772		return -EINVAL;
3773
3774	req->sync.off = READ_ONCE(sqe->off);
3775	req->sync.len = READ_ONCE(sqe->addr);
3776	req->sync.mode = READ_ONCE(sqe->len);
3777	return 0;
3778}
3779
3780static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3781{
3782	int ret;
3783
3784	/* fallocate always requiring blocking context */
3785	if (issue_flags & IO_URING_F_NONBLOCK)
3786		return -EAGAIN;
3787	ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3788				req->sync.len);
3789	if (ret < 0)
3790		req_set_fail(req);
3791	io_req_complete(req, ret);
3792	return 0;
3793}
3794
3795static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3796{
3797	const char __user *fname;
3798	int ret;
3799
3800	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3801		return -EINVAL;
3802	if (unlikely(sqe->ioprio || sqe->buf_index))
3803		return -EINVAL;
3804	if (unlikely(req->flags & REQ_F_FIXED_FILE))
3805		return -EBADF;
3806
3807	/* open.how should be already initialised */
3808	if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3809		req->open.how.flags |= O_LARGEFILE;
3810
3811	req->open.dfd = READ_ONCE(sqe->fd);
3812	fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3813	req->open.filename = getname(fname);
3814	if (IS_ERR(req->open.filename)) {
3815		ret = PTR_ERR(req->open.filename);
3816		req->open.filename = NULL;
3817		return ret;
3818	}
3819	req->open.nofile = rlimit(RLIMIT_NOFILE);
3820	req->flags |= REQ_F_NEED_CLEANUP;
3821	return 0;
3822}
3823
3824static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3825{
3826	u64 mode = READ_ONCE(sqe->len);
3827	u64 flags = READ_ONCE(sqe->open_flags);
3828
3829	req->open.how = build_open_how(flags, mode);
3830	return __io_openat_prep(req, sqe);
3831}
3832
3833static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3834{
3835	struct open_how __user *how;
3836	size_t len;
3837	int ret;
3838
3839	how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3840	len = READ_ONCE(sqe->len);
3841	if (len < OPEN_HOW_SIZE_VER0)
3842		return -EINVAL;
3843
3844	ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3845					len);
3846	if (ret)
3847		return ret;
3848
3849	return __io_openat_prep(req, sqe);
3850}
3851
3852static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3853{
3854	struct open_flags op;
3855	struct file *file;
3856	bool nonblock_set;
3857	bool resolve_nonblock;
3858	int ret;
3859
3860	ret = build_open_flags(&req->open.how, &op);
3861	if (ret)
3862		goto err;
3863	nonblock_set = op.open_flag & O_NONBLOCK;
3864	resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3865	if (issue_flags & IO_URING_F_NONBLOCK) {
3866		/*
3867		 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3868		 * it'll always -EAGAIN
3869		 */
3870		if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3871			return -EAGAIN;
3872		op.lookup_flags |= LOOKUP_CACHED;
3873		op.open_flag |= O_NONBLOCK;
3874	}
3875
3876	ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3877	if (ret < 0)
3878		goto err;
3879
3880	file = do_filp_open(req->open.dfd, req->open.filename, &op);
3881	if (IS_ERR(file)) {
3882		/*
3883		 * We could hang on to this 'fd' on retrying, but seems like
3884		 * marginal gain for something that is now known to be a slower
3885		 * path. So just put it, and we'll get a new one when we retry.
3886		 */
3887		put_unused_fd(ret);
3888
3889		ret = PTR_ERR(file);
3890		/* only retry if RESOLVE_CACHED wasn't already set by application */
3891		if (ret == -EAGAIN &&
3892		    (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3893			return -EAGAIN;
3894		goto err;
3895	}
3896
3897	if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3898		file->f_flags &= ~O_NONBLOCK;
3899	fsnotify_open(file);
3900	fd_install(ret, file);
3901err:
3902	putname(req->open.filename);
3903	req->flags &= ~REQ_F_NEED_CLEANUP;
3904	if (ret < 0)
3905		req_set_fail(req);
3906	__io_req_complete(req, issue_flags, ret, 0);
3907	return 0;
3908}
3909
3910static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3911{
3912	return io_openat2(req, issue_flags);
3913}
3914
3915static int io_remove_buffers_prep(struct io_kiocb *req,
3916				  const struct io_uring_sqe *sqe)
3917{
3918	struct io_provide_buf *p = &req->pbuf;
3919	u64 tmp;
3920
3921	if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3922		return -EINVAL;
3923
3924	tmp = READ_ONCE(sqe->fd);
3925	if (!tmp || tmp > USHRT_MAX)
3926		return -EINVAL;
3927
3928	memset(p, 0, sizeof(*p));
3929	p->nbufs = tmp;
3930	p->bgid = READ_ONCE(sqe->buf_group);
3931	return 0;
3932}
3933
3934static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3935			       int bgid, unsigned nbufs)
3936{
3937	unsigned i = 0;
3938
3939	/* shouldn't happen */
3940	if (!nbufs)
3941		return 0;
3942
3943	/* the head kbuf is the list itself */
3944	while (!list_empty(&buf->list)) {
3945		struct io_buffer *nxt;
3946
3947		nxt = list_first_entry(&buf->list, struct io_buffer, list);
3948		list_del(&nxt->list);
3949		kfree(nxt);
3950		if (++i == nbufs)
3951			return i;
3952	}
3953	i++;
3954	kfree(buf);
3955	xa_erase(&ctx->io_buffers, bgid);
3956
3957	return i;
3958}
3959
3960static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3961{
3962	struct io_provide_buf *p = &req->pbuf;
3963	struct io_ring_ctx *ctx = req->ctx;
3964	struct io_buffer *head;
3965	int ret = 0;
3966	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3967
3968	io_ring_submit_lock(ctx, !force_nonblock);
3969
3970	lockdep_assert_held(&ctx->uring_lock);
3971
3972	ret = -ENOENT;
3973	head = xa_load(&ctx->io_buffers, p->bgid);
3974	if (head)
3975		ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3976	if (ret < 0)
3977		req_set_fail(req);
3978
3979	/* complete before unlock, IOPOLL may need the lock */
3980	__io_req_complete(req, issue_flags, ret, 0);
3981	io_ring_submit_unlock(ctx, !force_nonblock);
3982	return 0;
3983}
3984
3985static int io_provide_buffers_prep(struct io_kiocb *req,
3986				   const struct io_uring_sqe *sqe)
3987{
3988	unsigned long size, tmp_check;
3989	struct io_provide_buf *p = &req->pbuf;
3990	u64 tmp;
3991
3992	if (sqe->ioprio || sqe->rw_flags)
3993		return -EINVAL;
3994
3995	tmp = READ_ONCE(sqe->fd);
3996	if (!tmp || tmp > USHRT_MAX)
3997		return -E2BIG;
3998	p->nbufs = tmp;
3999	p->addr = READ_ONCE(sqe->addr);
4000	p->len = READ_ONCE(sqe->len);
4001
4002	if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4003				&size))
4004		return -EOVERFLOW;
4005	if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4006		return -EOVERFLOW;
4007
4008	size = (unsigned long)p->len * p->nbufs;
4009	if (!access_ok(u64_to_user_ptr(p->addr), size))
4010		return -EFAULT;
4011
4012	p->bgid = READ_ONCE(sqe->buf_group);
4013	tmp = READ_ONCE(sqe->off);
4014	if (tmp > USHRT_MAX)
4015		return -E2BIG;
4016	p->bid = tmp;
4017	return 0;
4018}
4019
4020static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4021{
4022	struct io_buffer *buf;
4023	u64 addr = pbuf->addr;
4024	int i, bid = pbuf->bid;
4025
4026	for (i = 0; i < pbuf->nbufs; i++) {
4027		buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4028		if (!buf)
4029			break;
4030
4031		buf->addr = addr;
4032		buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4033		buf->bid = bid;
4034		addr += pbuf->len;
4035		bid++;
4036		if (!*head) {
4037			INIT_LIST_HEAD(&buf->list);
4038			*head = buf;
4039		} else {
4040			list_add_tail(&buf->list, &(*head)->list);
4041		}
4042	}
4043
4044	return i ? i : -ENOMEM;
4045}
4046
4047static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4048{
4049	struct io_provide_buf *p = &req->pbuf;
4050	struct io_ring_ctx *ctx = req->ctx;
4051	struct io_buffer *head, *list;
4052	int ret = 0;
4053	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4054
4055	io_ring_submit_lock(ctx, !force_nonblock);
4056
4057	lockdep_assert_held(&ctx->uring_lock);
4058
4059	list = head = xa_load(&ctx->io_buffers, p->bgid);
4060
4061	ret = io_add_buffers(p, &head);
4062	if (ret >= 0 && !list) {
4063		ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4064		if (ret < 0)
4065			__io_remove_buffers(ctx, head, p->bgid, -1U);
4066	}
4067	if (ret < 0)
4068		req_set_fail(req);
4069	/* complete before unlock, IOPOLL may need the lock */
4070	__io_req_complete(req, issue_flags, ret, 0);
4071	io_ring_submit_unlock(ctx, !force_nonblock);
4072	return 0;
4073}
4074
4075static int io_epoll_ctl_prep(struct io_kiocb *req,
4076			     const struct io_uring_sqe *sqe)
4077{
4078#if defined(CONFIG_EPOLL)
4079	if (sqe->ioprio || sqe->buf_index)
4080		return -EINVAL;
4081	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4082		return -EINVAL;
4083
4084	req->epoll.epfd = READ_ONCE(sqe->fd);
4085	req->epoll.op = READ_ONCE(sqe->len);
4086	req->epoll.fd = READ_ONCE(sqe->off);
4087
4088	if (ep_op_has_event(req->epoll.op)) {
4089		struct epoll_event __user *ev;
4090
4091		ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4092		if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4093			return -EFAULT;
4094	}
4095
4096	return 0;
4097#else
4098	return -EOPNOTSUPP;
4099#endif
4100}
4101
4102static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4103{
4104#if defined(CONFIG_EPOLL)
4105	struct io_epoll *ie = &req->epoll;
4106	int ret;
4107	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4108
4109	ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4110	if (force_nonblock && ret == -EAGAIN)
4111		return -EAGAIN;
4112
4113	if (ret < 0)
4114		req_set_fail(req);
4115	__io_req_complete(req, issue_flags, ret, 0);
4116	return 0;
4117#else
4118	return -EOPNOTSUPP;
4119#endif
4120}
4121
4122static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4123{
4124#if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4125	if (sqe->ioprio || sqe->buf_index || sqe->off)
4126		return -EINVAL;
4127	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4128		return -EINVAL;
4129
4130	req->madvise.addr = READ_ONCE(sqe->addr);
4131	req->madvise.len = READ_ONCE(sqe->len);
4132	req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4133	return 0;
4134#else
4135	return -EOPNOTSUPP;
4136#endif
4137}
4138
4139static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4140{
4141#if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4142	struct io_madvise *ma = &req->madvise;
4143	int ret;
4144
4145	if (issue_flags & IO_URING_F_NONBLOCK)
4146		return -EAGAIN;
4147
4148	ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4149	if (ret < 0)
4150		req_set_fail(req);
4151	io_req_complete(req, ret);
4152	return 0;
4153#else
4154	return -EOPNOTSUPP;
4155#endif
4156}
4157
4158static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4159{
4160	if (sqe->ioprio || sqe->buf_index || sqe->addr)
4161		return -EINVAL;
4162	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4163		return -EINVAL;
4164
4165	req->fadvise.offset = READ_ONCE(sqe->off);
4166	req->fadvise.len = READ_ONCE(sqe->len);
4167	req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4168	return 0;
4169}
4170
4171static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4172{
4173	struct io_fadvise *fa = &req->fadvise;
4174	int ret;
4175
4176	if (issue_flags & IO_URING_F_NONBLOCK) {
4177		switch (fa->advice) {
4178		case POSIX_FADV_NORMAL:
4179		case POSIX_FADV_RANDOM:
4180		case POSIX_FADV_SEQUENTIAL:
4181			break;
4182		default:
4183			return -EAGAIN;
4184		}
4185	}
4186
4187	ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4188	if (ret < 0)
4189		req_set_fail(req);
4190	__io_req_complete(req, issue_flags, ret, 0);
4191	return 0;
4192}
4193
4194static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4195{
4196	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4197		return -EINVAL;
4198	if (sqe->ioprio || sqe->buf_index)
4199		return -EINVAL;
4200	if (req->flags & REQ_F_FIXED_FILE)
4201		return -EBADF;
4202
4203	req->statx.dfd = READ_ONCE(sqe->fd);
4204	req->statx.mask = READ_ONCE(sqe->len);
4205	req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4206	req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4207	req->statx.flags = READ_ONCE(sqe->statx_flags);
4208
4209	return 0;
4210}
4211
4212static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4213{
4214	struct io_statx *ctx = &req->statx;
4215	int ret;
4216
4217	if (issue_flags & IO_URING_F_NONBLOCK)
4218		return -EAGAIN;
4219
4220	ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4221		       ctx->buffer);
4222
4223	if (ret < 0)
4224		req_set_fail(req);
4225	io_req_complete(req, ret);
4226	return 0;
4227}
4228
4229static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4230{
4231	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4232		return -EINVAL;
4233	if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4234	    sqe->rw_flags || sqe->buf_index)
4235		return -EINVAL;
4236	if (req->flags & REQ_F_FIXED_FILE)
4237		return -EBADF;
4238
4239	req->close.fd = READ_ONCE(sqe->fd);
4240	return 0;
4241}
4242
4243static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4244{
4245	struct files_struct *files = current->files;
4246	struct io_close *close = &req->close;
4247	struct fdtable *fdt;
4248	struct file *file = NULL;
4249	int ret = -EBADF;
4250
4251	spin_lock(&files->file_lock);
4252	fdt = files_fdtable(files);
4253	if (close->fd >= fdt->max_fds) {
4254		spin_unlock(&files->file_lock);
4255		goto err;
4256	}
4257	file = fdt->fd[close->fd];
4258	if (!file || file->f_op == &io_uring_fops) {
4259		spin_unlock(&files->file_lock);
4260		file = NULL;
4261		goto err;
4262	}
4263
4264	/* if the file has a flush method, be safe and punt to async */
4265	if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4266		spin_unlock(&files->file_lock);
4267		return -EAGAIN;
4268	}
4269
4270	ret = __close_fd_get_file(close->fd, &file);
4271	spin_unlock(&files->file_lock);
4272	if (ret < 0) {
4273		if (ret == -ENOENT)
4274			ret = -EBADF;
4275		goto err;
4276	}
4277
4278	/* No ->flush() or already async, safely close from here */
4279	ret = filp_close(file, current->files);
4280err:
4281	if (ret < 0)
4282		req_set_fail(req);
4283	if (file)
4284		fput(file);
4285	__io_req_complete(req, issue_flags, ret, 0);
4286	return 0;
4287}
4288
4289static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4290{
4291	struct io_ring_ctx *ctx = req->ctx;
4292
4293	if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4294		return -EINVAL;
4295	if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4296		return -EINVAL;
4297
4298	req->sync.off = READ_ONCE(sqe->off);
4299	req->sync.len = READ_ONCE(sqe->len);
4300	req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4301	return 0;
4302}
4303
4304static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4305{
4306	int ret;
4307
4308	/* sync_file_range always requires a blocking context */
4309	if (issue_flags & IO_URING_F_NONBLOCK)
4310		return -EAGAIN;
4311
4312	ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4313				req->sync.flags);
4314	if (ret < 0)
4315		req_set_fail(req);
4316	io_req_complete(req, ret);
4317	return 0;
4318}
4319
4320#if defined(CONFIG_NET)
4321static int io_setup_async_msg(struct io_kiocb *req,
4322			      struct io_async_msghdr *kmsg)
4323{
4324	struct io_async_msghdr *async_msg = req->async_data;
4325
4326	if (async_msg)
4327		return -EAGAIN;
4328	if (io_alloc_async_data(req)) {
4329		kfree(kmsg->free_iov);
4330		return -ENOMEM;
4331	}
4332	async_msg = req->async_data;
4333	req->flags |= REQ_F_NEED_CLEANUP;
4334	memcpy(async_msg, kmsg, sizeof(*kmsg));
4335	async_msg->msg.msg_name = &async_msg->addr;
4336	/* if were using fast_iov, set it to the new one */
4337	if (!async_msg->free_iov)
4338		async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4339
4340	return -EAGAIN;
4341}
4342
4343static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4344			       struct io_async_msghdr *iomsg)
4345{
4346	iomsg->msg.msg_name = &iomsg->addr;
4347	iomsg->free_iov = iomsg->fast_iov;
4348	return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4349				   req->sr_msg.msg_flags, &iomsg->free_iov);
4350}
4351
4352static int io_sendmsg_prep_async(struct io_kiocb *req)
4353{
4354	int ret;
4355
4356	ret = io_sendmsg_copy_hdr(req, req->async_data);
4357	if (!ret)
4358		req->flags |= REQ_F_NEED_CLEANUP;
4359	return ret;
4360}
4361
4362static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4363{
4364	struct io_sr_msg *sr = &req->sr_msg;
4365
4366	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4367		return -EINVAL;
4368
4369	sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4370	sr->len = READ_ONCE(sqe->len);
4371	sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4372	if (sr->msg_flags & MSG_DONTWAIT)
4373		req->flags |= REQ_F_NOWAIT;
4374
4375#ifdef CONFIG_COMPAT
4376	if (req->ctx->compat)
4377		sr->msg_flags |= MSG_CMSG_COMPAT;
4378#endif
4379	return 0;
4380}
4381
4382static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4383{
4384	struct io_async_msghdr iomsg, *kmsg;
4385	struct socket *sock;
4386	unsigned flags;
4387	int min_ret = 0;
4388	int ret;
4389
4390	sock = sock_from_file(req->file);
4391	if (unlikely(!sock))
4392		return -ENOTSOCK;
4393
4394	kmsg = req->async_data;
4395	if (!kmsg) {
4396		ret = io_sendmsg_copy_hdr(req, &iomsg);
4397		if (ret)
4398			return ret;
4399		kmsg = &iomsg;
4400	}
4401
4402	flags = req->sr_msg.msg_flags;
4403	if (issue_flags & IO_URING_F_NONBLOCK)
4404		flags |= MSG_DONTWAIT;
4405	if (flags & MSG_WAITALL)
4406		min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4407
4408	ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4409	if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4410		return io_setup_async_msg(req, kmsg);
4411	if (ret == -ERESTARTSYS)
4412		ret = -EINTR;
4413
4414	/* fast path, check for non-NULL to avoid function call */
4415	if (kmsg->free_iov)
4416		kfree(kmsg->free_iov);
4417	req->flags &= ~REQ_F_NEED_CLEANUP;
4418	if (ret < min_ret)
4419		req_set_fail(req);
4420	__io_req_complete(req, issue_flags, ret, 0);
4421	return 0;
4422}
4423
4424static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4425{
4426	struct io_sr_msg *sr = &req->sr_msg;
4427	struct msghdr msg;
4428	struct iovec iov;
4429	struct socket *sock;
4430	unsigned flags;
4431	int min_ret = 0;
4432	int ret;
4433
4434	sock = sock_from_file(req->file);
4435	if (unlikely(!sock))
4436		return -ENOTSOCK;
4437
4438	ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4439	if (unlikely(ret))
4440		return ret;
4441
4442	msg.msg_name = NULL;
4443	msg.msg_control = NULL;
4444	msg.msg_controllen = 0;
4445	msg.msg_namelen = 0;
4446
4447	flags = req->sr_msg.msg_flags;
4448	if (issue_flags & IO_URING_F_NONBLOCK)
4449		flags |= MSG_DONTWAIT;
4450	if (flags & MSG_WAITALL)
4451		min_ret = iov_iter_count(&msg.msg_iter);
4452
4453	msg.msg_flags = flags;
4454	ret = sock_sendmsg(sock, &msg);
4455	if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4456		return -EAGAIN;
4457	if (ret == -ERESTARTSYS)
4458		ret = -EINTR;
4459
4460	if (ret < min_ret)
4461		req_set_fail(req);
4462	__io_req_complete(req, issue_flags, ret, 0);
4463	return 0;
4464}
4465
4466static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4467				 struct io_async_msghdr *iomsg)
4468{
4469	struct io_sr_msg *sr = &req->sr_msg;
4470	struct iovec __user *uiov;
4471	size_t iov_len;
4472	int ret;
4473
4474	ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4475					&iomsg->uaddr, &uiov, &iov_len);
4476	if (ret)
4477		return ret;
4478
4479	if (req->flags & REQ_F_BUFFER_SELECT) {
4480		if (iov_len > 1)
4481			return -EINVAL;
4482		if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4483			return -EFAULT;
4484		sr->len = iomsg->fast_iov[0].iov_len;
4485		iomsg->free_iov = NULL;
4486	} else {
4487		iomsg->free_iov = iomsg->fast_iov;
4488		ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4489				     &iomsg->free_iov, &iomsg->msg.msg_iter,
4490				     false);
4491		if (ret > 0)
4492			ret = 0;
4493	}
4494
4495	return ret;
4496}
4497
4498#ifdef CONFIG_COMPAT
4499static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4500					struct io_async_msghdr *iomsg)
4501{
4502	struct io_sr_msg *sr = &req->sr_msg;
4503	struct compat_iovec __user *uiov;
4504	compat_uptr_t ptr;
4505	compat_size_t len;
4506	int ret;
4507
4508	ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4509				  &ptr, &len);
4510	if (ret)
4511		return ret;
4512
4513	uiov = compat_ptr(ptr);
4514	if (req->flags & REQ_F_BUFFER_SELECT) {
4515		compat_ssize_t clen;
4516
4517		if (len > 1)
4518			return -EINVAL;
4519		if (!access_ok(uiov, sizeof(*uiov)))
4520			return -EFAULT;
4521		if (__get_user(clen, &uiov->iov_len))
4522			return -EFAULT;
4523		if (clen < 0)
4524			return -EINVAL;
4525		sr->len = clen;
4526		iomsg->free_iov = NULL;
4527	} else {
4528		iomsg->free_iov = iomsg->fast_iov;
4529		ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4530				   UIO_FASTIOV, &iomsg->free_iov,
4531				   &iomsg->msg.msg_iter, true);
4532		if (ret < 0)
4533			return ret;
4534	}
4535
4536	return 0;
4537}
4538#endif
4539
4540static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4541			       struct io_async_msghdr *iomsg)
4542{
4543	iomsg->msg.msg_name = &iomsg->addr;
4544
4545#ifdef CONFIG_COMPAT
4546	if (req->ctx->compat)
4547		return __io_compat_recvmsg_copy_hdr(req, iomsg);
4548#endif
4549
4550	return __io_recvmsg_copy_hdr(req, iomsg);
4551}
4552
4553static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4554					       bool needs_lock)
4555{
4556	struct io_sr_msg *sr = &req->sr_msg;
4557	struct io_buffer *kbuf;
4558
4559	kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4560	if (IS_ERR(kbuf))
4561		return kbuf;
4562
4563	sr->kbuf = kbuf;
4564	req->flags |= REQ_F_BUFFER_SELECTED;
4565	return kbuf;
4566}
4567
4568static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4569{
4570	return io_put_kbuf(req, req->sr_msg.kbuf);
4571}
4572
4573static int io_recvmsg_prep_async(struct io_kiocb *req)
4574{
4575	int ret;
4576
4577	ret = io_recvmsg_copy_hdr(req, req->async_data);
4578	if (!ret)
4579		req->flags |= REQ_F_NEED_CLEANUP;
4580	return ret;
4581}
4582
4583static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4584{
4585	struct io_sr_msg *sr = &req->sr_msg;
4586
4587	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4588		return -EINVAL;
4589
4590	sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4591	sr->len = READ_ONCE(sqe->len);
4592	sr->bgid = READ_ONCE(sqe->buf_group);
4593	sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4594	if (sr->msg_flags & MSG_DONTWAIT)
4595		req->flags |= REQ_F_NOWAIT;
4596
4597#ifdef CONFIG_COMPAT
4598	if (req->ctx->compat)
4599		sr->msg_flags |= MSG_CMSG_COMPAT;
4600#endif
4601	return 0;
4602}
4603
4604static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4605{
4606	struct io_async_msghdr iomsg, *kmsg;
4607	struct socket *sock;
4608	struct io_buffer *kbuf;
4609	unsigned flags;
4610	int min_ret = 0;
4611	int ret, cflags = 0;
4612	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4613
4614	sock = sock_from_file(req->file);
4615	if (unlikely(!sock))
4616		return -ENOTSOCK;
4617
4618	kmsg = req->async_data;
4619	if (!kmsg) {
4620		ret = io_recvmsg_copy_hdr(req, &iomsg);
4621		if (ret)
4622			return ret;
4623		kmsg = &iomsg;
4624	}
4625
4626	if (req->flags & REQ_F_BUFFER_SELECT) {
4627		kbuf = io_recv_buffer_select(req, !force_nonblock);
4628		if (IS_ERR(kbuf))
4629			return PTR_ERR(kbuf);
4630		kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4631		kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4632		iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4633				1, req->sr_msg.len);
4634	}
4635
4636	flags = req->sr_msg.msg_flags;
4637	if (force_nonblock)
4638		flags |= MSG_DONTWAIT;
4639	if (flags & MSG_WAITALL)
4640		min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4641
4642	ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4643					kmsg->uaddr, flags);
4644	if (force_nonblock && ret == -EAGAIN)
4645		return io_setup_async_msg(req, kmsg);
4646	if (ret == -ERESTARTSYS)
4647		ret = -EINTR;
4648
4649	if (req->flags & REQ_F_BUFFER_SELECTED)
4650		cflags = io_put_recv_kbuf(req);
4651	/* fast path, check for non-NULL to avoid function call */
4652	if (kmsg->free_iov)
4653		kfree(kmsg->free_iov);
4654	req->flags &= ~REQ_F_NEED_CLEANUP;
4655	if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4656		req_set_fail(req);
4657	__io_req_complete(req, issue_flags, ret, cflags);
4658	return 0;
4659}
4660
4661static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4662{
4663	struct io_buffer *kbuf;
4664	struct io_sr_msg *sr = &req->sr_msg;
4665	struct msghdr msg;
4666	void __user *buf = sr->buf;
4667	struct socket *sock;
4668	struct iovec iov;
4669	unsigned flags;
4670	int min_ret = 0;
4671	int ret, cflags = 0;
4672	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4673
4674	sock = sock_from_file(req->file);
4675	if (unlikely(!sock))
4676		return -ENOTSOCK;
4677
4678	if (req->flags & REQ_F_BUFFER_SELECT) {
4679		kbuf = io_recv_buffer_select(req, !force_nonblock);
4680		if (IS_ERR(kbuf))
4681			return PTR_ERR(kbuf);
4682		buf = u64_to_user_ptr(kbuf->addr);
4683	}
4684
4685	ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4686	if (unlikely(ret))
4687		goto out_free;
4688
4689	msg.msg_name = NULL;
4690	msg.msg_control = NULL;
4691	msg.msg_controllen = 0;
4692	msg.msg_namelen = 0;
4693	msg.msg_iocb = NULL;
4694	msg.msg_flags = 0;
4695
4696	flags = req->sr_msg.msg_flags;
4697	if (force_nonblock)
4698		flags |= MSG_DONTWAIT;
4699	if (flags & MSG_WAITALL)
4700		min_ret = iov_iter_count(&msg.msg_iter);
4701
4702	ret = sock_recvmsg(sock, &msg, flags);
4703	if (force_nonblock && ret == -EAGAIN)
4704		return -EAGAIN;
4705	if (ret == -ERESTARTSYS)
4706		ret = -EINTR;
4707out_free:
4708	if (req->flags & REQ_F_BUFFER_SELECTED)
4709		cflags = io_put_recv_kbuf(req);
4710	if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4711		req_set_fail(req);
4712	__io_req_complete(req, issue_flags, ret, cflags);
4713	return 0;
4714}
4715
4716static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4717{
4718	struct io_accept *accept = &req->accept;
4719
4720	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4721		return -EINVAL;
4722	if (sqe->ioprio || sqe->len || sqe->buf_index)
4723		return -EINVAL;
4724
4725	accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4726	accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4727	accept->flags = READ_ONCE(sqe->accept_flags);
4728	accept->nofile = rlimit(RLIMIT_NOFILE);
4729	return 0;
4730}
4731
4732static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4733{
4734	struct io_accept *accept = &req->accept;
4735	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4736	unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4737	int ret;
4738
4739	if (req->file->f_flags & O_NONBLOCK)
4740		req->flags |= REQ_F_NOWAIT;
4741
4742	ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4743					accept->addr_len, accept->flags,
4744					accept->nofile);
4745	if (ret == -EAGAIN && force_nonblock)
4746		return -EAGAIN;
4747	if (ret < 0) {
4748		if (ret == -ERESTARTSYS)
4749			ret = -EINTR;
4750		req_set_fail(req);
4751	}
4752	__io_req_complete(req, issue_flags, ret, 0);
4753	return 0;
4754}
4755
4756static int io_connect_prep_async(struct io_kiocb *req)
4757{
4758	struct io_async_connect *io = req->async_data;
4759	struct io_connect *conn = &req->connect;
4760
4761	return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4762}
4763
4764static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4765{
4766	struct io_connect *conn = &req->connect;
4767
4768	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4769		return -EINVAL;
4770	if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4771		return -EINVAL;
4772
4773	conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4774	conn->addr_len =  READ_ONCE(sqe->addr2);
4775	return 0;
4776}
4777
4778static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4779{
4780	struct io_async_connect __io, *io;
4781	unsigned file_flags;
4782	int ret;
4783	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4784
4785	if (req->async_data) {
4786		io = req->async_data;
4787	} else {
4788		ret = move_addr_to_kernel(req->connect.addr,
4789						req->connect.addr_len,
4790						&__io.address);
4791		if (ret)
4792			goto out;
4793		io = &__io;
4794	}
4795
4796	file_flags = force_nonblock ? O_NONBLOCK : 0;
4797
4798	ret = __sys_connect_file(req->file, &io->address,
4799					req->connect.addr_len, file_flags);
4800	if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4801		if (req->async_data)
4802			return -EAGAIN;
4803		if (io_alloc_async_data(req)) {
4804			ret = -ENOMEM;
4805			goto out;
4806		}
4807		memcpy(req->async_data, &__io, sizeof(__io));
4808		return -EAGAIN;
4809	}
4810	if (ret == -ERESTARTSYS)
4811		ret = -EINTR;
4812out:
4813	if (ret < 0)
4814		req_set_fail(req);
4815	__io_req_complete(req, issue_flags, ret, 0);
4816	return 0;
4817}
4818#else /* !CONFIG_NET */
4819#define IO_NETOP_FN(op)							\
4820static int io_##op(struct io_kiocb *req, unsigned int issue_flags)	\
4821{									\
4822	return -EOPNOTSUPP;						\
4823}
4824
4825#define IO_NETOP_PREP(op)						\
4826IO_NETOP_FN(op)								\
4827static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4828{									\
4829	return -EOPNOTSUPP;						\
4830}									\
4831
4832#define IO_NETOP_PREP_ASYNC(op)						\
4833IO_NETOP_PREP(op)							\
4834static int io_##op##_prep_async(struct io_kiocb *req)			\
4835{									\
4836	return -EOPNOTSUPP;						\
4837}
4838
4839IO_NETOP_PREP_ASYNC(sendmsg);
4840IO_NETOP_PREP_ASYNC(recvmsg);
4841IO_NETOP_PREP_ASYNC(connect);
4842IO_NETOP_PREP(accept);
4843IO_NETOP_FN(send);
4844IO_NETOP_FN(recv);
4845#endif /* CONFIG_NET */
4846
4847struct io_poll_table {
4848	struct poll_table_struct pt;
4849	struct io_kiocb *req;
4850	int nr_entries;
4851	int error;
4852};
4853
4854static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4855			   __poll_t mask, io_req_tw_func_t func)
4856{
4857	/* for instances that support it check for an event match first: */
4858	if (mask && !(mask & poll->events))
4859		return 0;
4860
4861	trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4862
4863	list_del_init(&poll->wait.entry);
4864
4865	req->result = mask;
4866	req->io_task_work.func = func;
4867
4868	/*
4869	 * If this fails, then the task is exiting. When a task exits, the
4870	 * work gets canceled, so just cancel this request as well instead
4871	 * of executing it. We can't safely execute it anyway, as we may not
4872	 * have the needed state needed for it anyway.
4873	 */
4874	io_req_task_work_add(req);
4875	return 1;
4876}
4877
4878static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4879	__acquires(&req->ctx->completion_lock)
4880{
4881	struct io_ring_ctx *ctx = req->ctx;
4882
4883	if (unlikely(req->task->flags & PF_EXITING))
4884		WRITE_ONCE(poll->canceled, true);
4885
4886	if (!req->result && !READ_ONCE(poll->canceled)) {
4887		struct poll_table_struct pt = { ._key = poll->events };
4888
4889		req->result = vfs_poll(req->file, &pt) & poll->events;
4890	}
4891
4892	spin_lock(&ctx->completion_lock);
4893	if (!req->result && !READ_ONCE(poll->canceled)) {
4894		add_wait_queue(poll->head, &poll->wait);
4895		return true;
4896	}
4897
4898	return false;
4899}
4900
4901static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4902{
4903	/* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4904	if (req->opcode == IORING_OP_POLL_ADD)
4905		return req->async_data;
4906	return req->apoll->double_poll;
4907}
4908
4909static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4910{
4911	if (req->opcode == IORING_OP_POLL_ADD)
4912		return &req->poll;
4913	return &req->apoll->poll;
4914}
4915
4916static void io_poll_remove_double(struct io_kiocb *req)
4917	__must_hold(&req->ctx->completion_lock)
4918{
4919	struct io_poll_iocb *poll = io_poll_get_double(req);
4920
4921	lockdep_assert_held(&req->ctx->completion_lock);
4922
4923	if (poll && poll->head) {
4924		struct wait_queue_head *head = poll->head;
4925
4926		spin_lock_irq(&head->lock);
4927		list_del_init(&poll->wait.entry);
4928		if (poll->wait.private)
4929			req_ref_put(req);
4930		poll->head = NULL;
4931		spin_unlock_irq(&head->lock);
4932	}
4933}
4934
4935static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4936	__must_hold(&req->ctx->completion_lock)
4937{
4938	struct io_ring_ctx *ctx = req->ctx;
4939	unsigned flags = IORING_CQE_F_MORE;
4940	int error;
4941
4942	if (READ_ONCE(req->poll.canceled)) {
4943		error = -ECANCELED;
4944		req->poll.events |= EPOLLONESHOT;
4945	} else {
4946		error = mangle_poll(mask);
4947	}
4948	if (req->poll.events & EPOLLONESHOT)
4949		flags = 0;
4950	if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4951		req->poll.done = true;
4952		flags = 0;
4953	}
4954	if (flags & IORING_CQE_F_MORE)
4955		ctx->cq_extra++;
4956
4957	io_commit_cqring(ctx);
4958	return !(flags & IORING_CQE_F_MORE);
4959}
4960
4961static void io_poll_task_func(struct io_kiocb *req)
4962{
4963	struct io_ring_ctx *ctx = req->ctx;
4964	struct io_kiocb *nxt;
4965
4966	if (io_poll_rewait(req, &req->poll)) {
4967		spin_unlock(&ctx->completion_lock);
4968	} else {
4969		bool done;
4970
4971		done = io_poll_complete(req, req->result);
4972		if (done) {
4973			io_poll_remove_double(req);
4974			hash_del(&req->hash_node);
4975		} else {
4976			req->result = 0;
4977			add_wait_queue(req->poll.head, &req->poll.wait);
4978		}
4979		spin_unlock(&ctx->completion_lock);
4980		io_cqring_ev_posted(ctx);
4981
4982		if (done) {
4983			nxt = io_put_req_find_next(req);
4984			if (nxt)
4985				io_req_task_submit(nxt);
4986		}
4987	}
4988}
4989
4990static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4991			       int sync, void *key)
4992{
4993	struct io_kiocb *req = wait->private;
4994	struct io_poll_iocb *poll = io_poll_get_single(req);
4995	__poll_t mask = key_to_poll(key);
4996	unsigned long flags;
4997
4998	/* for instances that support it check for an event match first: */
4999	if (mask && !(mask & poll->events))
5000		return 0;
5001	if (!(poll->events & EPOLLONESHOT))
5002		return poll->wait.func(&poll->wait, mode, sync, key);
5003
5004	list_del_init(&wait->entry);
5005
5006	if (poll->head) {
5007		bool done;
5008
5009		spin_lock_irqsave(&poll->head->lock, flags);
5010		done = list_empty(&poll->wait.entry);
5011		if (!done)
5012			list_del_init(&poll->wait.entry);
5013		/* make sure double remove sees this as being gone */
5014		wait->private = NULL;
5015		spin_unlock_irqrestore(&poll->head->lock, flags);
5016		if (!done) {
5017			/* use wait func handler, so it matches the rq type */
5018			poll->wait.func(&poll->wait, mode, sync, key);
5019		}
5020	}
5021	req_ref_put(req);
5022	return 1;
5023}
5024
5025static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5026			      wait_queue_func_t wake_func)
5027{
5028	poll->head = NULL;
5029	poll->done = false;
5030	poll->canceled = false;
5031#define IO_POLL_UNMASK	(EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5032	/* mask in events that we always want/need */
5033	poll->events = events | IO_POLL_UNMASK;
5034	INIT_LIST_HEAD(&poll->wait.entry);
5035	init_waitqueue_func_entry(&poll->wait, wake_func);
5036}
5037
5038static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5039			    struct wait_queue_head *head,
5040			    struct io_poll_iocb **poll_ptr)
5041{
5042	struct io_kiocb *req = pt->req;
5043
5044	/*
5045	 * The file being polled uses multiple waitqueues for poll handling
5046	 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5047	 * if this happens.
5048	 */
5049	if (unlikely(pt->nr_entries)) {
5050		struct io_poll_iocb *poll_one = poll;
5051
5052		/* already have a 2nd entry, fail a third attempt */
5053		if (*poll_ptr) {
5054			pt->error = -EINVAL;
5055			return;
5056		}
5057		/*
5058		 * Can't handle multishot for double wait for now, turn it
5059		 * into one-shot mode.
5060		 */
5061		if (!(poll_one->events & EPOLLONESHOT))
5062			poll_one->events |= EPOLLONESHOT;
5063		/* double add on the same waitqueue head, ignore */
5064		if (poll_one->head == head)
5065			return;
5066		poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5067		if (!poll) {
5068			pt->error = -ENOMEM;
5069			return;
5070		}
5071		io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5072		req_ref_get(req);
5073		poll->wait.private = req;
5074		*poll_ptr = poll;
5075	}
5076
5077	pt->nr_entries++;
5078	poll->head = head;
5079
5080	if (poll->events & EPOLLEXCLUSIVE)
5081		add_wait_queue_exclusive(head, &poll->wait);
5082	else
5083		add_wait_queue(head, &poll->wait);
5084}
5085
5086static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5087			       struct poll_table_struct *p)
5088{
5089	struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5090	struct async_poll *apoll = pt->req->apoll;
5091
5092	__io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5093}
5094
5095static void io_async_task_func(struct io_kiocb *req)
5096{
5097	struct async_poll *apoll = req->apoll;
5098	struct io_ring_ctx *ctx = req->ctx;
5099
5100	trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5101
5102	if (io_poll_rewait(req, &apoll->poll)) {
5103		spin_unlock(&ctx->completion_lock);
5104		return;
5105	}
5106
5107	hash_del(&req->hash_node);
5108	io_poll_remove_double(req);
5109	spin_unlock(&ctx->completion_lock);
5110
5111	if (!READ_ONCE(apoll->poll.canceled))
5112		io_req_task_submit(req);
5113	else
5114		io_req_complete_failed(req, -ECANCELED);
5115}
5116
5117static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5118			void *key)
5119{
5120	struct io_kiocb *req = wait->private;
5121	struct io_poll_iocb *poll = &req->apoll->poll;
5122
5123	trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5124					key_to_poll(key));
5125
5126	return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5127}
5128
5129static void io_poll_req_insert(struct io_kiocb *req)
5130{
5131	struct io_ring_ctx *ctx = req->ctx;
5132	struct hlist_head *list;
5133
5134	list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5135	hlist_add_head(&req->hash_node, list);
5136}
5137
5138static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5139				      struct io_poll_iocb *poll,
5140				      struct io_poll_table *ipt, __poll_t mask,
5141				      wait_queue_func_t wake_func)
5142	__acquires(&ctx->completion_lock)
5143{
5144	struct io_ring_ctx *ctx = req->ctx;
5145	bool cancel = false;
5146
5147	INIT_HLIST_NODE(&req->hash_node);
5148	io_init_poll_iocb(poll, mask, wake_func);
5149	poll->file = req->file;
5150	poll->wait.private = req;
5151
5152	ipt->pt._key = mask;
5153	ipt->req = req;
5154	ipt->error = 0;
5155	ipt->nr_entries = 0;
5156
5157	mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5158	if (unlikely(!ipt->nr_entries) && !ipt->error)
5159		ipt->error = -EINVAL;
5160
5161	spin_lock(&ctx->completion_lock);
5162	if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5163		io_poll_remove_double(req);
5164	if (likely(poll->head)) {
5165		spin_lock_irq(&poll->head->lock);
5166		if (unlikely(list_empty(&poll->wait.entry))) {
5167			if (ipt->error)
5168				cancel = true;
5169			ipt->error = 0;
5170			mask = 0;
5171		}
5172		if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5173			list_del_init(&poll->wait.entry);
5174		else if (cancel)
5175			WRITE_ONCE(poll->canceled, true);
5176		else if (!poll->done) /* actually waiting for an event */
5177			io_poll_req_insert(req);
5178		spin_unlock_irq(&poll->head->lock);
5179	}
5180
5181	return mask;
5182}
5183
5184enum {
5185	IO_APOLL_OK,
5186	IO_APOLL_ABORTED,
5187	IO_APOLL_READY
5188};
5189
5190static int io_arm_poll_handler(struct io_kiocb *req)
5191{
5192	const struct io_op_def *def = &io_op_defs[req->opcode];
5193	struct io_ring_ctx *ctx = req->ctx;
5194	struct async_poll *apoll;
5195	struct io_poll_table ipt;
5196	__poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5197	int rw;
5198
5199	if (!req->file || !file_can_poll(req->file))
5200		return IO_APOLL_ABORTED;
5201	if (req->flags & REQ_F_POLLED)
5202		return IO_APOLL_ABORTED;
5203	if (!def->pollin && !def->pollout)
5204		return IO_APOLL_ABORTED;
5205
5206	if (def->pollin) {
5207		rw = READ;
5208		mask |= POLLIN | POLLRDNORM;
5209
5210		/* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5211		if ((req->opcode == IORING_OP_RECVMSG) &&
5212		    (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5213			mask &= ~POLLIN;
5214	} else {
5215		rw = WRITE;
5216		mask |= POLLOUT | POLLWRNORM;
5217	}
5218
5219	/* if we can't nonblock try, then no point in arming a poll handler */
5220	if (!io_file_supports_nowait(req, rw))
5221		return IO_APOLL_ABORTED;
5222
5223	apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5224	if (unlikely(!apoll))
5225		return IO_APOLL_ABORTED;
5226	apoll->double_poll = NULL;
5227	req->apoll = apoll;
5228	req->flags |= REQ_F_POLLED;
5229	ipt.pt._qproc = io_async_queue_proc;
5230	io_req_set_refcount(req);
5231
5232	ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5233					io_async_wake);
5234	spin_unlock(&ctx->completion_lock);
5235	if (ret || ipt.error)
5236		return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5237
5238	trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5239				mask, apoll->poll.events);
5240	return IO_APOLL_OK;
5241}
5242
5243static bool __io_poll_remove_one(struct io_kiocb *req,
5244				 struct io_poll_iocb *poll, bool do_cancel)
5245	__must_hold(&req->ctx->completion_lock)
5246{
5247	bool do_complete = false;
5248
5249	if (!poll->head)
5250		return false;
5251	spin_lock_irq(&poll->head->lock);
5252	if (do_cancel)
5253		WRITE_ONCE(poll->canceled, true);
5254	if (!list_empty(&poll->wait.entry)) {
5255		list_del_init(&poll->wait.entry);
5256		do_complete = true;
5257	}
5258	spin_unlock_irq(&poll->head->lock);
5259	hash_del(&req->hash_node);
5260	return do_complete;
5261}
5262
5263static bool io_poll_remove_one(struct io_kiocb *req)
5264	__must_hold(&req->ctx->completion_lock)
5265{
5266	bool do_complete;
5267
5268	io_poll_remove_double(req);
5269	do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5270
5271	if (do_complete) {
5272		io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5273		io_commit_cqring(req->ctx);
5274		req_set_fail(req);
5275		io_put_req_deferred(req);
5276	}
5277	return do_complete;
5278}
5279
5280/*
5281 * Returns true if we found and killed one or more poll requests
5282 */
5283static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5284			       bool cancel_all)
5285{
5286	struct hlist_node *tmp;
5287	struct io_kiocb *req;
5288	int posted = 0, i;
5289
5290	spin_lock(&ctx->completion_lock);
5291	for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5292		struct hlist_head *list;
5293
5294		list = &ctx->cancel_hash[i];
5295		hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5296			if (io_match_task(req, tsk, cancel_all))
5297				posted += io_poll_remove_one(req);
5298		}
5299	}
5300	spin_unlock(&ctx->completion_lock);
5301
5302	if (posted)
5303		io_cqring_ev_posted(ctx);
5304
5305	return posted != 0;
5306}
5307
5308static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5309				     bool poll_only)
5310	__must_hold(&ctx->completion_lock)
5311{
5312	struct hlist_head *list;
5313	struct io_kiocb *req;
5314
5315	list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5316	hlist_for_each_entry(req, list, hash_node) {
5317		if (sqe_addr != req->user_data)
5318			continue;
5319		if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5320			continue;
5321		return req;
5322	}
5323	return NULL;
5324}
5325
5326static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5327			  bool poll_only)
5328	__must_hold(&ctx->completion_lock)
5329{
5330	struct io_kiocb *req;
5331
5332	req = io_poll_find(ctx, sqe_addr, poll_only);
5333	if (!req)
5334		return -ENOENT;
5335	if (io_poll_remove_one(req))
5336		return 0;
5337
5338	return -EALREADY;
5339}
5340
5341static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5342				     unsigned int flags)
5343{
5344	u32 events;
5345
5346	events = READ_ONCE(sqe->poll32_events);
5347#ifdef __BIG_ENDIAN
5348	events = swahw32(events);
5349#endif
5350	if (!(flags & IORING_POLL_ADD_MULTI))
5351		events |= EPOLLONESHOT;
5352	return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5353}
5354
5355static int io_poll_update_prep(struct io_kiocb *req,
5356			       const struct io_uring_sqe *sqe)
5357{
5358	struct io_poll_update *upd = &req->poll_update;
5359	u32 flags;
5360
5361	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5362		return -EINVAL;
5363	if (sqe->ioprio || sqe->buf_index)
5364		return -EINVAL;
5365	flags = READ_ONCE(sqe->len);
5366	if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5367		      IORING_POLL_ADD_MULTI))
5368		return -EINVAL;
5369	/* meaningless without update */
5370	if (flags == IORING_POLL_ADD_MULTI)
5371		return -EINVAL;
5372
5373	upd->old_user_data = READ_ONCE(sqe->addr);
5374	upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5375	upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5376
5377	upd->new_user_data = READ_ONCE(sqe->off);
5378	if (!upd->update_user_data && upd->new_user_data)
5379		return -EINVAL;
5380	if (upd->update_events)
5381		upd->events = io_poll_parse_events(sqe, flags);
5382	else if (sqe->poll32_events)
5383		return -EINVAL;
5384
5385	return 0;
5386}
5387
5388static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5389			void *key)
5390{
5391	struct io_kiocb *req = wait->private;
5392	struct io_poll_iocb *poll = &req->poll;
5393
5394	return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5395}
5396
5397static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5398			       struct poll_table_struct *p)
5399{
5400	struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5401
5402	__io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5403}
5404
5405static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5406{
5407	struct io_poll_iocb *poll = &req->poll;
5408	u32 flags;
5409
5410	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5411		return -EINVAL;
5412	if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5413		return -EINVAL;
5414	flags = READ_ONCE(sqe->len);
5415	if (flags & ~IORING_POLL_ADD_MULTI)
5416		return -EINVAL;
5417
5418	io_req_set_refcount(req);
5419	poll->events = io_poll_parse_events(sqe, flags);
5420	return 0;
5421}
5422
5423static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5424{
5425	struct io_poll_iocb *poll = &req->poll;
5426	struct io_ring_ctx *ctx = req->ctx;
5427	struct io_poll_table ipt;
5428	__poll_t mask;
5429
5430	ipt.pt._qproc = io_poll_queue_proc;
5431
5432	mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5433					io_poll_wake);
5434
5435	if (mask) { /* no async, we'd stolen it */
5436		ipt.error = 0;
5437		io_poll_complete(req, mask);
5438	}
5439	spin_unlock(&ctx->completion_lock);
5440
5441	if (mask) {
5442		io_cqring_ev_posted(ctx);
5443		if (poll->events & EPOLLONESHOT)
5444			io_put_req(req);
5445	}
5446	return ipt.error;
5447}
5448
5449static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5450{
5451	struct io_ring_ctx *ctx = req->ctx;
5452	struct io_kiocb *preq;
5453	bool completing;
5454	int ret;
5455
5456	spin_lock(&ctx->completion_lock);
5457	preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5458	if (!preq) {
5459		ret = -ENOENT;
5460		goto err;
5461	}
5462
5463	if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5464		completing = true;
5465		ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5466		goto err;
5467	}
5468
5469	/*
5470	 * Don't allow racy completion with singleshot, as we cannot safely
5471	 * update those. For multishot, if we're racing with completion, just
5472	 * let completion re-add it.
5473	 */
5474	completing = !__io_poll_remove_one(preq, &preq->poll, false);
5475	if (completing && (preq->poll.events & EPOLLONESHOT)) {
5476		ret = -EALREADY;
5477		goto err;
5478	}
5479	/* we now have a detached poll request. reissue. */
5480	ret = 0;
5481err:
5482	if (ret < 0) {
5483		spin_unlock(&ctx->completion_lock);
5484		req_set_fail(req);
5485		io_req_complete(req, ret);
5486		return 0;
5487	}
5488	/* only mask one event flags, keep behavior flags */
5489	if (req->poll_update.update_events) {
5490		preq->poll.events &= ~0xffff;
5491		preq->poll.events |= req->poll_update.events & 0xffff;
5492		preq->poll.events |= IO_POLL_UNMASK;
5493	}
5494	if (req->poll_update.update_user_data)
5495		preq->user_data = req->poll_update.new_user_data;
5496	spin_unlock(&ctx->completion_lock);
5497
5498	/* complete update request, we're done with it */
5499	io_req_complete(req, ret);
5500
5501	if (!completing) {
5502		ret = io_poll_add(preq, issue_flags);
5503		if (ret < 0) {
5504			req_set_fail(preq);
5505			io_req_complete(preq, ret);
5506		}
5507	}
5508	return 0;
5509}
5510
5511static void io_req_task_timeout(struct io_kiocb *req)
5512{
5513	struct io_ring_ctx *ctx = req->ctx;
5514
5515	spin_lock(&ctx->completion_lock);
5516	io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5517	io_commit_cqring(ctx);
5518	spin_unlock(&ctx->completion_lock);
5519
5520	io_cqring_ev_posted(ctx);
5521	req_set_fail(req);
5522	io_put_req(req);
5523}
5524
5525static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5526{
5527	struct io_timeout_data *data = container_of(timer,
5528						struct io_timeout_data, timer);
5529	struct io_kiocb *req = data->req;
5530	struct io_ring_ctx *ctx = req->ctx;
5531	unsigned long flags;
5532
5533	spin_lock_irqsave(&ctx->timeout_lock, flags);
5534	list_del_init(&req->timeout.list);
5535	atomic_set(&req->ctx->cq_timeouts,
5536		atomic_read(&req->ctx->cq_timeouts) + 1);
5537	spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5538
5539	req->io_task_work.func = io_req_task_timeout;
5540	io_req_task_work_add(req);
5541	return HRTIMER_NORESTART;
5542}
5543
5544static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5545					   __u64 user_data)
5546	__must_hold(&ctx->timeout_lock)
5547{
5548	struct io_timeout_data *io;
5549	struct io_kiocb *req;
5550	bool found = false;
5551
5552	list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5553		found = user_data == req->user_data;
5554		if (found)
5555			break;
5556	}
5557	if (!found)
5558		return ERR_PTR(-ENOENT);
5559
5560	io = req->async_data;
5561	if (hrtimer_try_to_cancel(&io->timer) == -1)
5562		return ERR_PTR(-EALREADY);
5563	list_del_init(&req->timeout.list);
5564	return req;
5565}
5566
5567static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5568	__must_hold(&ctx->timeout_lock)
5569{
5570	struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5571
5572	if (IS_ERR(req))
5573		return PTR_ERR(req);
5574
5575	req_set_fail(req);
5576	io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5577	io_put_req_deferred(req);
5578	return 0;
5579}
5580
5581static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5582			     struct timespec64 *ts, enum hrtimer_mode mode)
5583	__must_hold(&ctx->timeout_lock)
5584{
5585	struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5586	struct io_timeout_data *data;
5587
5588	if (IS_ERR(req))
5589		return PTR_ERR(req);
5590
5591	req->timeout.off = 0; /* noseq */
5592	data = req->async_data;
5593	list_add_tail(&req->timeout.list, &ctx->timeout_list);
5594	hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5595	data->timer.function = io_timeout_fn;
5596	hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5597	return 0;
5598}
5599
5600static int io_timeout_remove_prep(struct io_kiocb *req,
5601				  const struct io_uring_sqe *sqe)
5602{
5603	struct io_timeout_rem *tr = &req->timeout_rem;
5604
5605	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5606		return -EINVAL;
5607	if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5608		return -EINVAL;
5609	if (sqe->ioprio || sqe->buf_index || sqe->len)
5610		return -EINVAL;
5611
5612	tr->addr = READ_ONCE(sqe->addr);
5613	tr->flags = READ_ONCE(sqe->timeout_flags);
5614	if (tr->flags & IORING_TIMEOUT_UPDATE) {
5615		if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5616			return -EINVAL;
5617		if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5618			return -EFAULT;
5619	} else if (tr->flags) {
5620		/* timeout removal doesn't support flags */
5621		return -EINVAL;
5622	}
5623
5624	return 0;
5625}
5626
5627static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5628{
5629	return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5630					    : HRTIMER_MODE_REL;
5631}
5632
5633/*
5634 * Remove or update an existing timeout command
5635 */
5636static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5637{
5638	struct io_timeout_rem *tr = &req->timeout_rem;
5639	struct io_ring_ctx *ctx = req->ctx;
5640	int ret;
5641
5642	spin_lock_irq(&ctx->timeout_lock);
5643	if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5644		ret = io_timeout_cancel(ctx, tr->addr);
5645	else
5646		ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5647					io_translate_timeout_mode(tr->flags));
5648	spin_unlock_irq(&ctx->timeout_lock);
5649
5650	spin_lock(&ctx->completion_lock);
5651	io_cqring_fill_event(ctx, req->user_data, ret, 0);
5652	io_commit_cqring(ctx);
5653	spin_unlock(&ctx->completion_lock);
5654	io_cqring_ev_posted(ctx);
5655	if (ret < 0)
5656		req_set_fail(req);
5657	io_put_req(req);
5658	return 0;
5659}
5660
5661static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5662			   bool is_timeout_link)
5663{
5664	struct io_timeout_data *data;
5665	unsigned flags;
5666	u32 off = READ_ONCE(sqe->off);
5667
5668	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5669		return -EINVAL;
5670	if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5671		return -EINVAL;
5672	if (off && is_timeout_link)
5673		return -EINVAL;
5674	flags = READ_ONCE(sqe->timeout_flags);
5675	if (flags & ~IORING_TIMEOUT_ABS)
5676		return -EINVAL;
5677
5678	req->timeout.off = off;
5679	if (unlikely(off && !req->ctx->off_timeout_used))
5680		req->ctx->off_timeout_used = true;
5681
5682	if (!req->async_data && io_alloc_async_data(req))
5683		return -ENOMEM;
5684
5685	data = req->async_data;
5686	data->req = req;
5687
5688	if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5689		return -EFAULT;
5690
5691	data->mode = io_translate_timeout_mode(flags);
5692	hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5693
5694	if (is_timeout_link) {
5695		struct io_submit_link *link = &req->ctx->submit_state.link;
5696
5697		if (!link->head)
5698			return -EINVAL;
5699		if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
5700			return -EINVAL;
5701		req->timeout.head = link->last;
5702		link->last->flags |= REQ_F_ARM_LTIMEOUT;
5703	}
5704	return 0;
5705}
5706
5707static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5708{
5709	struct io_ring_ctx *ctx = req->ctx;
5710	struct io_timeout_data *data = req->async_data;
5711	struct list_head *entry;
5712	u32 tail, off = req->timeout.off;
5713
5714	spin_lock_irq(&ctx->timeout_lock);
5715
5716	/*
5717	 * sqe->off holds how many events that need to occur for this
5718	 * timeout event to be satisfied. If it isn't set, then this is
5719	 * a pure timeout request, sequence isn't used.
5720	 */
5721	if (io_is_timeout_noseq(req)) {
5722		entry = ctx->timeout_list.prev;
5723		goto add;
5724	}
5725
5726	tail = ctx->cached_cq_tail - a