io_uring.c revision 41a5169c
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_LTIMEOUT_ACTIVE_BIT,
709	REQ_F_COMPLETE_INLINE_BIT,
710	REQ_F_REISSUE_BIT,
711	REQ_F_DONT_REISSUE_BIT,
712	REQ_F_CREDS_BIT,
713	REQ_F_REFCOUNT_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	/* linked timeout is active, i.e. prepared by link's head */
754	REQ_F_LTIMEOUT_ACTIVE	= BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
755	/* completion is deferred through io_comp_state */
756	REQ_F_COMPLETE_INLINE	= BIT(REQ_F_COMPLETE_INLINE_BIT),
757	/* caller should reissue async */
758	REQ_F_REISSUE		= BIT(REQ_F_REISSUE_BIT),
759	/* don't attempt request reissue, see io_rw_reissue() */
760	REQ_F_DONT_REISSUE	= BIT(REQ_F_DONT_REISSUE_BIT),
761	/* supports async reads */
762	REQ_F_NOWAIT_READ	= BIT(REQ_F_NOWAIT_READ_BIT),
763	/* supports async writes */
764	REQ_F_NOWAIT_WRITE	= BIT(REQ_F_NOWAIT_WRITE_BIT),
765	/* regular file */
766	REQ_F_ISREG		= BIT(REQ_F_ISREG_BIT),
767	/* has creds assigned */
768	REQ_F_CREDS		= BIT(REQ_F_CREDS_BIT),
769	/* skip refcounting if not set */
770	REQ_F_REFCOUNT		= BIT(REQ_F_REFCOUNT_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_refcount(struct io_kiocb *req)
1119{
1120	if (!(req->flags & REQ_F_REFCOUNT)) {
1121		req->flags |= REQ_F_REFCOUNT;
1122		atomic_set(&req->refs, 1);
1123	}
1124}
1125
1126static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1127{
1128	struct io_ring_ctx *ctx = req->ctx;
1129
1130	if (!req->fixed_rsrc_refs) {
1131		req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1132		percpu_ref_get(req->fixed_rsrc_refs);
1133	}
1134}
1135
1136static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1137{
1138	bool got = percpu_ref_tryget(ref);
1139
1140	/* already at zero, wait for ->release() */
1141	if (!got)
1142		wait_for_completion(compl);
1143	percpu_ref_resurrect(ref);
1144	if (got)
1145		percpu_ref_put(ref);
1146}
1147
1148static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1149			  bool cancel_all)
1150{
1151	struct io_kiocb *req;
1152
1153	if (task && head->task != task)
1154		return false;
1155	if (cancel_all)
1156		return true;
1157
1158	io_for_each_link(req, head) {
1159		if (req->flags & REQ_F_INFLIGHT)
1160			return true;
1161	}
1162	return false;
1163}
1164
1165static inline void req_set_fail(struct io_kiocb *req)
1166{
1167	req->flags |= REQ_F_FAIL;
1168}
1169
1170static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1171{
1172	struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1173
1174	complete(&ctx->ref_comp);
1175}
1176
1177static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1178{
1179	return !req->timeout.off;
1180}
1181
1182static void io_fallback_req_func(struct work_struct *work)
1183{
1184	struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
1185						fallback_work.work);
1186	struct llist_node *node = llist_del_all(&ctx->fallback_llist);
1187	struct io_kiocb *req, *tmp;
1188
1189	percpu_ref_get(&ctx->refs);
1190	llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
1191		req->io_task_work.func(req);
1192	percpu_ref_put(&ctx->refs);
1193}
1194
1195static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1196{
1197	struct io_ring_ctx *ctx;
1198	int hash_bits;
1199
1200	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1201	if (!ctx)
1202		return NULL;
1203
1204	/*
1205	 * Use 5 bits less than the max cq entries, that should give us around
1206	 * 32 entries per hash list if totally full and uniformly spread.
1207	 */
1208	hash_bits = ilog2(p->cq_entries);
1209	hash_bits -= 5;
1210	if (hash_bits <= 0)
1211		hash_bits = 1;
1212	ctx->cancel_hash_bits = hash_bits;
1213	ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1214					GFP_KERNEL);
1215	if (!ctx->cancel_hash)
1216		goto err;
1217	__hash_init(ctx->cancel_hash, 1U << hash_bits);
1218
1219	ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1220	if (!ctx->dummy_ubuf)
1221		goto err;
1222	/* set invalid range, so io_import_fixed() fails meeting it */
1223	ctx->dummy_ubuf->ubuf = -1UL;
1224
1225	if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1226			    PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1227		goto err;
1228
1229	ctx->flags = p->flags;
1230	init_waitqueue_head(&ctx->sqo_sq_wait);
1231	INIT_LIST_HEAD(&ctx->sqd_list);
1232	init_waitqueue_head(&ctx->poll_wait);
1233	INIT_LIST_HEAD(&ctx->cq_overflow_list);
1234	init_completion(&ctx->ref_comp);
1235	xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1236	xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1237	mutex_init(&ctx->uring_lock);
1238	init_waitqueue_head(&ctx->cq_wait);
1239	spin_lock_init(&ctx->completion_lock);
1240	spin_lock_init(&ctx->timeout_lock);
1241	INIT_LIST_HEAD(&ctx->iopoll_list);
1242	INIT_LIST_HEAD(&ctx->defer_list);
1243	INIT_LIST_HEAD(&ctx->timeout_list);
1244	spin_lock_init(&ctx->rsrc_ref_lock);
1245	INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1246	INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1247	init_llist_head(&ctx->rsrc_put_llist);
1248	INIT_LIST_HEAD(&ctx->tctx_list);
1249	INIT_LIST_HEAD(&ctx->submit_state.free_list);
1250	INIT_LIST_HEAD(&ctx->locked_free_list);
1251	INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1252	return ctx;
1253err:
1254	kfree(ctx->dummy_ubuf);
1255	kfree(ctx->cancel_hash);
1256	kfree(ctx);
1257	return NULL;
1258}
1259
1260static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1261{
1262	struct io_rings *r = ctx->rings;
1263
1264	WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1265	ctx->cq_extra--;
1266}
1267
1268static bool req_need_defer(struct io_kiocb *req, u32 seq)
1269{
1270	if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1271		struct io_ring_ctx *ctx = req->ctx;
1272
1273		return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1274	}
1275
1276	return false;
1277}
1278
1279#define FFS_ASYNC_READ		0x1UL
1280#define FFS_ASYNC_WRITE		0x2UL
1281#ifdef CONFIG_64BIT
1282#define FFS_ISREG		0x4UL
1283#else
1284#define FFS_ISREG		0x0UL
1285#endif
1286#define FFS_MASK		~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1287
1288static inline bool io_req_ffs_set(struct io_kiocb *req)
1289{
1290	return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE);
1291}
1292
1293static void io_req_track_inflight(struct io_kiocb *req)
1294{
1295	if (!(req->flags & REQ_F_INFLIGHT)) {
1296		req->flags |= REQ_F_INFLIGHT;
1297		atomic_inc(&current->io_uring->inflight_tracked);
1298	}
1299}
1300
1301static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
1302{
1303	struct io_kiocb *nxt = req->link;
1304
1305	if (req->flags & REQ_F_LINK_TIMEOUT)
1306		return NULL;
1307
1308	/* linked timeouts should have two refs once prep'ed */
1309	io_req_refcount(req);
1310	io_req_refcount(nxt);
1311	req_ref_get(nxt);
1312
1313	nxt->timeout.head = req;
1314	nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
1315	req->flags |= REQ_F_LINK_TIMEOUT;
1316	return nxt;
1317}
1318
1319static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
1320{
1321	if (likely(!req->link || req->link->opcode != IORING_OP_LINK_TIMEOUT))
1322		return NULL;
1323	return __io_prep_linked_timeout(req);
1324}
1325
1326static void io_prep_async_work(struct io_kiocb *req)
1327{
1328	const struct io_op_def *def = &io_op_defs[req->opcode];
1329	struct io_ring_ctx *ctx = req->ctx;
1330
1331	if (!(req->flags & REQ_F_CREDS)) {
1332		req->flags |= REQ_F_CREDS;
1333		req->creds = get_current_cred();
1334	}
1335
1336	req->work.list.next = NULL;
1337	req->work.flags = 0;
1338	if (req->flags & REQ_F_FORCE_ASYNC)
1339		req->work.flags |= IO_WQ_WORK_CONCURRENT;
1340
1341	if (req->flags & REQ_F_ISREG) {
1342		if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1343			io_wq_hash_work(&req->work, file_inode(req->file));
1344	} else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1345		if (def->unbound_nonreg_file)
1346			req->work.flags |= IO_WQ_WORK_UNBOUND;
1347	}
1348
1349	switch (req->opcode) {
1350	case IORING_OP_SPLICE:
1351	case IORING_OP_TEE:
1352		if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1353			req->work.flags |= IO_WQ_WORK_UNBOUND;
1354		break;
1355	}
1356}
1357
1358static void io_prep_async_link(struct io_kiocb *req)
1359{
1360	struct io_kiocb *cur;
1361
1362	if (req->flags & REQ_F_LINK_TIMEOUT) {
1363		struct io_ring_ctx *ctx = req->ctx;
1364
1365		spin_lock(&ctx->completion_lock);
1366		io_for_each_link(cur, req)
1367			io_prep_async_work(cur);
1368		spin_unlock(&ctx->completion_lock);
1369	} else {
1370		io_for_each_link(cur, req)
1371			io_prep_async_work(cur);
1372	}
1373}
1374
1375static void io_queue_async_work(struct io_kiocb *req)
1376{
1377	struct io_ring_ctx *ctx = req->ctx;
1378	struct io_kiocb *link = io_prep_linked_timeout(req);
1379	struct io_uring_task *tctx = req->task->io_uring;
1380
1381	BUG_ON(!tctx);
1382	BUG_ON(!tctx->io_wq);
1383
1384	/* init ->work of the whole link before punting */
1385	io_prep_async_link(req);
1386
1387	/*
1388	 * Not expected to happen, but if we do have a bug where this _can_
1389	 * happen, catch it here and ensure the request is marked as
1390	 * canceled. That will make io-wq go through the usual work cancel
1391	 * procedure rather than attempt to run this request (or create a new
1392	 * worker for it).
1393	 */
1394	if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1395		req->work.flags |= IO_WQ_WORK_CANCEL;
1396
1397	trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1398					&req->work, req->flags);
1399	io_wq_enqueue(tctx->io_wq, &req->work);
1400	if (link)
1401		io_queue_linked_timeout(link);
1402}
1403
1404static void io_kill_timeout(struct io_kiocb *req, int status)
1405	__must_hold(&req->ctx->completion_lock)
1406	__must_hold(&req->ctx->timeout_lock)
1407{
1408	struct io_timeout_data *io = req->async_data;
1409
1410	if (hrtimer_try_to_cancel(&io->timer) != -1) {
1411		atomic_set(&req->ctx->cq_timeouts,
1412			atomic_read(&req->ctx->cq_timeouts) + 1);
1413		list_del_init(&req->timeout.list);
1414		io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1415		io_put_req_deferred(req);
1416	}
1417}
1418
1419static void io_queue_deferred(struct io_ring_ctx *ctx)
1420{
1421	while (!list_empty(&ctx->defer_list)) {
1422		struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1423						struct io_defer_entry, list);
1424
1425		if (req_need_defer(de->req, de->seq))
1426			break;
1427		list_del_init(&de->list);
1428		io_req_task_queue(de->req);
1429		kfree(de);
1430	}
1431}
1432
1433static void io_flush_timeouts(struct io_ring_ctx *ctx)
1434	__must_hold(&ctx->completion_lock)
1435{
1436	u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1437
1438	spin_lock_irq(&ctx->timeout_lock);
1439	while (!list_empty(&ctx->timeout_list)) {
1440		u32 events_needed, events_got;
1441		struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1442						struct io_kiocb, timeout.list);
1443
1444		if (io_is_timeout_noseq(req))
1445			break;
1446
1447		/*
1448		 * Since seq can easily wrap around over time, subtract
1449		 * the last seq at which timeouts were flushed before comparing.
1450		 * Assuming not more than 2^31-1 events have happened since,
1451		 * these subtractions won't have wrapped, so we can check if
1452		 * target is in [last_seq, current_seq] by comparing the two.
1453		 */
1454		events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1455		events_got = seq - ctx->cq_last_tm_flush;
1456		if (events_got < events_needed)
1457			break;
1458
1459		list_del_init(&req->timeout.list);
1460		io_kill_timeout(req, 0);
1461	}
1462	ctx->cq_last_tm_flush = seq;
1463	spin_unlock_irq(&ctx->timeout_lock);
1464}
1465
1466static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1467{
1468	if (ctx->off_timeout_used)
1469		io_flush_timeouts(ctx);
1470	if (ctx->drain_active)
1471		io_queue_deferred(ctx);
1472}
1473
1474static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1475{
1476	if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1477		__io_commit_cqring_flush(ctx);
1478	/* order cqe stores with ring update */
1479	smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1480}
1481
1482static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1483{
1484	struct io_rings *r = ctx->rings;
1485
1486	return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1487}
1488
1489static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1490{
1491	return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1492}
1493
1494static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1495{
1496	struct io_rings *rings = ctx->rings;
1497	unsigned tail, mask = ctx->cq_entries - 1;
1498
1499	/*
1500	 * writes to the cq entry need to come after reading head; the
1501	 * control dependency is enough as we're using WRITE_ONCE to
1502	 * fill the cq entry
1503	 */
1504	if (__io_cqring_events(ctx) == ctx->cq_entries)
1505		return NULL;
1506
1507	tail = ctx->cached_cq_tail++;
1508	return &rings->cqes[tail & mask];
1509}
1510
1511static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1512{
1513	if (likely(!ctx->cq_ev_fd))
1514		return false;
1515	if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1516		return false;
1517	return !ctx->eventfd_async || io_wq_current_is_worker();
1518}
1519
1520static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1521{
1522	/*
1523	 * wake_up_all() may seem excessive, but io_wake_function() and
1524	 * io_should_wake() handle the termination of the loop and only
1525	 * wake as many waiters as we need to.
1526	 */
1527	if (wq_has_sleeper(&ctx->cq_wait))
1528		wake_up_all(&ctx->cq_wait);
1529	if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1530		wake_up(&ctx->sq_data->wait);
1531	if (io_should_trigger_evfd(ctx))
1532		eventfd_signal(ctx->cq_ev_fd, 1);
1533	if (waitqueue_active(&ctx->poll_wait)) {
1534		wake_up_interruptible(&ctx->poll_wait);
1535		kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1536	}
1537}
1538
1539static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1540{
1541	if (ctx->flags & IORING_SETUP_SQPOLL) {
1542		if (wq_has_sleeper(&ctx->cq_wait))
1543			wake_up_all(&ctx->cq_wait);
1544	}
1545	if (io_should_trigger_evfd(ctx))
1546		eventfd_signal(ctx->cq_ev_fd, 1);
1547	if (waitqueue_active(&ctx->poll_wait)) {
1548		wake_up_interruptible(&ctx->poll_wait);
1549		kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1550	}
1551}
1552
1553/* Returns true if there are no backlogged entries after the flush */
1554static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1555{
1556	bool all_flushed, posted;
1557
1558	if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1559		return false;
1560
1561	posted = false;
1562	spin_lock(&ctx->completion_lock);
1563	while (!list_empty(&ctx->cq_overflow_list)) {
1564		struct io_uring_cqe *cqe = io_get_cqe(ctx);
1565		struct io_overflow_cqe *ocqe;
1566
1567		if (!cqe && !force)
1568			break;
1569		ocqe = list_first_entry(&ctx->cq_overflow_list,
1570					struct io_overflow_cqe, list);
1571		if (cqe)
1572			memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1573		else
1574			io_account_cq_overflow(ctx);
1575
1576		posted = true;
1577		list_del(&ocqe->list);
1578		kfree(ocqe);
1579	}
1580
1581	all_flushed = list_empty(&ctx->cq_overflow_list);
1582	if (all_flushed) {
1583		clear_bit(0, &ctx->check_cq_overflow);
1584		WRITE_ONCE(ctx->rings->sq_flags,
1585			   ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1586	}
1587
1588	if (posted)
1589		io_commit_cqring(ctx);
1590	spin_unlock(&ctx->completion_lock);
1591	if (posted)
1592		io_cqring_ev_posted(ctx);
1593	return all_flushed;
1594}
1595
1596static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1597{
1598	bool ret = true;
1599
1600	if (test_bit(0, &ctx->check_cq_overflow)) {
1601		/* iopoll syncs against uring_lock, not completion_lock */
1602		if (ctx->flags & IORING_SETUP_IOPOLL)
1603			mutex_lock(&ctx->uring_lock);
1604		ret = __io_cqring_overflow_flush(ctx, false);
1605		if (ctx->flags & IORING_SETUP_IOPOLL)
1606			mutex_unlock(&ctx->uring_lock);
1607	}
1608
1609	return ret;
1610}
1611
1612/* must to be called somewhat shortly after putting a request */
1613static inline void io_put_task(struct task_struct *task, int nr)
1614{
1615	struct io_uring_task *tctx = task->io_uring;
1616
1617	percpu_counter_sub(&tctx->inflight, nr);
1618	if (unlikely(atomic_read(&tctx->in_idle)))
1619		wake_up(&tctx->wait);
1620	put_task_struct_many(task, nr);
1621}
1622
1623static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1624				     long res, unsigned int cflags)
1625{
1626	struct io_overflow_cqe *ocqe;
1627
1628	ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1629	if (!ocqe) {
1630		/*
1631		 * If we're in ring overflow flush mode, or in task cancel mode,
1632		 * or cannot allocate an overflow entry, then we need to drop it
1633		 * on the floor.
1634		 */
1635		io_account_cq_overflow(ctx);
1636		return false;
1637	}
1638	if (list_empty(&ctx->cq_overflow_list)) {
1639		set_bit(0, &ctx->check_cq_overflow);
1640		WRITE_ONCE(ctx->rings->sq_flags,
1641			   ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1642
1643	}
1644	ocqe->cqe.user_data = user_data;
1645	ocqe->cqe.res = res;
1646	ocqe->cqe.flags = cflags;
1647	list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1648	return true;
1649}
1650
1651static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1652					  long res, unsigned int cflags)
1653{
1654	struct io_uring_cqe *cqe;
1655
1656	trace_io_uring_complete(ctx, user_data, res, cflags);
1657
1658	/*
1659	 * If we can't get a cq entry, userspace overflowed the
1660	 * submission (by quite a lot). Increment the overflow count in
1661	 * the ring.
1662	 */
1663	cqe = io_get_cqe(ctx);
1664	if (likely(cqe)) {
1665		WRITE_ONCE(cqe->user_data, user_data);
1666		WRITE_ONCE(cqe->res, res);
1667		WRITE_ONCE(cqe->flags, cflags);
1668		return true;
1669	}
1670	return io_cqring_event_overflow(ctx, user_data, res, cflags);
1671}
1672
1673/* not as hot to bloat with inlining */
1674static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1675					  long res, unsigned int cflags)
1676{
1677	return __io_cqring_fill_event(ctx, user_data, res, cflags);
1678}
1679
1680static void io_req_complete_post(struct io_kiocb *req, long res,
1681				 unsigned int cflags)
1682{
1683	struct io_ring_ctx *ctx = req->ctx;
1684
1685	spin_lock(&ctx->completion_lock);
1686	__io_cqring_fill_event(ctx, req->user_data, res, cflags);
1687	/*
1688	 * If we're the last reference to this request, add to our locked
1689	 * free_list cache.
1690	 */
1691	if (req_ref_put_and_test(req)) {
1692		if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1693			if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL))
1694				io_disarm_next(req);
1695			if (req->link) {
1696				io_req_task_queue(req->link);
1697				req->link = NULL;
1698			}
1699		}
1700		io_dismantle_req(req);
1701		io_put_task(req->task, 1);
1702		list_add(&req->inflight_entry, &ctx->locked_free_list);
1703		ctx->locked_free_nr++;
1704	} else {
1705		if (!percpu_ref_tryget(&ctx->refs))
1706			req = NULL;
1707	}
1708	io_commit_cqring(ctx);
1709	spin_unlock(&ctx->completion_lock);
1710
1711	if (req) {
1712		io_cqring_ev_posted(ctx);
1713		percpu_ref_put(&ctx->refs);
1714	}
1715}
1716
1717static inline bool io_req_needs_clean(struct io_kiocb *req)
1718{
1719	return req->flags & IO_REQ_CLEAN_FLAGS;
1720}
1721
1722static void io_req_complete_state(struct io_kiocb *req, long res,
1723				  unsigned int cflags)
1724{
1725	if (io_req_needs_clean(req))
1726		io_clean_op(req);
1727	req->result = res;
1728	req->compl.cflags = cflags;
1729	req->flags |= REQ_F_COMPLETE_INLINE;
1730}
1731
1732static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1733				     long res, unsigned cflags)
1734{
1735	if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1736		io_req_complete_state(req, res, cflags);
1737	else
1738		io_req_complete_post(req, res, cflags);
1739}
1740
1741static inline void io_req_complete(struct io_kiocb *req, long res)
1742{
1743	__io_req_complete(req, 0, res, 0);
1744}
1745
1746static void io_req_complete_failed(struct io_kiocb *req, long res)
1747{
1748	req_set_fail(req);
1749	io_req_complete_post(req, res, 0);
1750}
1751
1752/*
1753 * Don't initialise the fields below on every allocation, but do that in
1754 * advance and keep them valid across allocations.
1755 */
1756static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1757{
1758	req->ctx = ctx;
1759	req->link = NULL;
1760	req->async_data = NULL;
1761	/* not necessary, but safer to zero */
1762	req->result = 0;
1763}
1764
1765static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1766					struct io_submit_state *state)
1767{
1768	spin_lock(&ctx->completion_lock);
1769	list_splice_init(&ctx->locked_free_list, &state->free_list);
1770	ctx->locked_free_nr = 0;
1771	spin_unlock(&ctx->completion_lock);
1772}
1773
1774/* Returns true IFF there are requests in the cache */
1775static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1776{
1777	struct io_submit_state *state = &ctx->submit_state;
1778	int nr;
1779
1780	/*
1781	 * If we have more than a batch's worth of requests in our IRQ side
1782	 * locked cache, grab the lock and move them over to our submission
1783	 * side cache.
1784	 */
1785	if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1786		io_flush_cached_locked_reqs(ctx, state);
1787
1788	nr = state->free_reqs;
1789	while (!list_empty(&state->free_list)) {
1790		struct io_kiocb *req = list_first_entry(&state->free_list,
1791					struct io_kiocb, inflight_entry);
1792
1793		list_del(&req->inflight_entry);
1794		state->reqs[nr++] = req;
1795		if (nr == ARRAY_SIZE(state->reqs))
1796			break;
1797	}
1798
1799	state->free_reqs = nr;
1800	return nr != 0;
1801}
1802
1803/*
1804 * A request might get retired back into the request caches even before opcode
1805 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1806 * Because of that, io_alloc_req() should be called only under ->uring_lock
1807 * and with extra caution to not get a request that is still worked on.
1808 */
1809static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1810	__must_hold(&ctx->uring_lock)
1811{
1812	struct io_submit_state *state = &ctx->submit_state;
1813	gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1814	int ret, i;
1815
1816	BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1817
1818	if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1819		goto got_req;
1820
1821	ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1822				    state->reqs);
1823
1824	/*
1825	 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1826	 * retry single alloc to be on the safe side.
1827	 */
1828	if (unlikely(ret <= 0)) {
1829		state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1830		if (!state->reqs[0])
1831			return NULL;
1832		ret = 1;
1833	}
1834
1835	for (i = 0; i < ret; i++)
1836		io_preinit_req(state->reqs[i], ctx);
1837	state->free_reqs = ret;
1838got_req:
1839	state->free_reqs--;
1840	return state->reqs[state->free_reqs];
1841}
1842
1843static inline void io_put_file(struct file *file)
1844{
1845	if (file)
1846		fput(file);
1847}
1848
1849static void io_dismantle_req(struct io_kiocb *req)
1850{
1851	unsigned int flags = req->flags;
1852
1853	if (io_req_needs_clean(req))
1854		io_clean_op(req);
1855	if (!(flags & REQ_F_FIXED_FILE))
1856		io_put_file(req->file);
1857	if (req->fixed_rsrc_refs)
1858		percpu_ref_put(req->fixed_rsrc_refs);
1859	if (req->async_data) {
1860		kfree(req->async_data);
1861		req->async_data = NULL;
1862	}
1863}
1864
1865static void __io_free_req(struct io_kiocb *req)
1866{
1867	struct io_ring_ctx *ctx = req->ctx;
1868
1869	io_dismantle_req(req);
1870	io_put_task(req->task, 1);
1871
1872	spin_lock(&ctx->completion_lock);
1873	list_add(&req->inflight_entry, &ctx->locked_free_list);
1874	ctx->locked_free_nr++;
1875	spin_unlock(&ctx->completion_lock);
1876
1877	percpu_ref_put(&ctx->refs);
1878}
1879
1880static inline void io_remove_next_linked(struct io_kiocb *req)
1881{
1882	struct io_kiocb *nxt = req->link;
1883
1884	req->link = nxt->link;
1885	nxt->link = NULL;
1886}
1887
1888static bool io_kill_linked_timeout(struct io_kiocb *req)
1889	__must_hold(&req->ctx->completion_lock)
1890	__must_hold(&req->ctx->timeout_lock)
1891{
1892	struct io_kiocb *link = req->link;
1893
1894	/*
1895	 * Can happen if a linked timeout fired and link had been like
1896	 * req -> link t-out -> link t-out [-> ...]
1897	 */
1898	if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1899		struct io_timeout_data *io = link->async_data;
1900
1901		io_remove_next_linked(req);
1902		link->timeout.head = NULL;
1903		if (hrtimer_try_to_cancel(&io->timer) != -1) {
1904			io_cqring_fill_event(link->ctx, link->user_data,
1905					     -ECANCELED, 0);
1906			io_put_req_deferred(link);
1907			return true;
1908		}
1909	}
1910	return false;
1911}
1912
1913static void io_fail_links(struct io_kiocb *req)
1914	__must_hold(&req->ctx->completion_lock)
1915{
1916	struct io_kiocb *nxt, *link = req->link;
1917
1918	req->link = NULL;
1919	while (link) {
1920		nxt = link->link;
1921		link->link = NULL;
1922
1923		trace_io_uring_fail_link(req, link);
1924		io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1925		io_put_req_deferred(link);
1926		link = nxt;
1927	}
1928}
1929
1930static bool io_disarm_next(struct io_kiocb *req)
1931	__must_hold(&req->ctx->completion_lock)
1932{
1933	bool posted = false;
1934
1935	if (likely(req->flags & REQ_F_LINK_TIMEOUT)) {
1936		struct io_ring_ctx *ctx = req->ctx;
1937
1938		spin_lock_irq(&ctx->timeout_lock);
1939		posted = io_kill_linked_timeout(req);
1940		spin_unlock_irq(&ctx->timeout_lock);
1941	}
1942	if (unlikely((req->flags & REQ_F_FAIL) &&
1943		     !(req->flags & REQ_F_HARDLINK))) {
1944		posted |= (req->link != NULL);
1945		io_fail_links(req);
1946	}
1947	return posted;
1948}
1949
1950static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1951{
1952	struct io_kiocb *nxt;
1953
1954	/*
1955	 * If LINK is set, we have dependent requests in this chain. If we
1956	 * didn't fail this request, queue the first one up, moving any other
1957	 * dependencies to the next request. In case of failure, fail the rest
1958	 * of the chain.
1959	 */
1960	if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) {
1961		struct io_ring_ctx *ctx = req->ctx;
1962		bool posted;
1963
1964		spin_lock(&ctx->completion_lock);
1965		posted = io_disarm_next(req);
1966		if (posted)
1967			io_commit_cqring(req->ctx);
1968		spin_unlock(&ctx->completion_lock);
1969		if (posted)
1970			io_cqring_ev_posted(ctx);
1971	}
1972	nxt = req->link;
1973	req->link = NULL;
1974	return nxt;
1975}
1976
1977static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1978{
1979	if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1980		return NULL;
1981	return __io_req_find_next(req);
1982}
1983
1984static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1985{
1986	if (!ctx)
1987		return;
1988	if (ctx->submit_state.compl_nr) {
1989		mutex_lock(&ctx->uring_lock);
1990		io_submit_flush_completions(ctx);
1991		mutex_unlock(&ctx->uring_lock);
1992	}
1993	percpu_ref_put(&ctx->refs);
1994}
1995
1996static void tctx_task_work(struct callback_head *cb)
1997{
1998	struct io_ring_ctx *ctx = NULL;
1999	struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2000						  task_work);
2001
2002	while (1) {
2003		struct io_wq_work_node *node;
2004
2005		spin_lock_irq(&tctx->task_lock);
2006		node = tctx->task_list.first;
2007		INIT_WQ_LIST(&tctx->task_list);
2008		if (!node)
2009			tctx->task_running = false;
2010		spin_unlock_irq(&tctx->task_lock);
2011		if (!node)
2012			break;
2013
2014		do {
2015			struct io_wq_work_node *next = node->next;
2016			struct io_kiocb *req = container_of(node, struct io_kiocb,
2017							    io_task_work.node);
2018
2019			if (req->ctx != ctx) {
2020				ctx_flush_and_put(ctx);
2021				ctx = req->ctx;
2022				percpu_ref_get(&ctx->refs);
2023			}
2024			req->io_task_work.func(req);
2025			node = next;
2026		} while (node);
2027
2028		cond_resched();
2029	}
2030
2031	ctx_flush_and_put(ctx);
2032}
2033
2034static void io_req_task_work_add(struct io_kiocb *req)
2035{
2036	struct task_struct *tsk = req->task;
2037	struct io_uring_task *tctx = tsk->io_uring;
2038	enum task_work_notify_mode notify;
2039	struct io_wq_work_node *node;
2040	unsigned long flags;
2041	bool running;
2042
2043	WARN_ON_ONCE(!tctx);
2044
2045	spin_lock_irqsave(&tctx->task_lock, flags);
2046	wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2047	running = tctx->task_running;
2048	if (!running)
2049		tctx->task_running = true;
2050	spin_unlock_irqrestore(&tctx->task_lock, flags);
2051
2052	/* task_work already pending, we're done */
2053	if (running)
2054		return;
2055
2056	/*
2057	 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2058	 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2059	 * processing task_work. There's no reliable way to tell if TWA_RESUME
2060	 * will do the job.
2061	 */
2062	notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2063	if (!task_work_add(tsk, &tctx->task_work, notify)) {
2064		wake_up_process(tsk);
2065		return;
2066	}
2067
2068	spin_lock_irqsave(&tctx->task_lock, flags);
2069	tctx->task_running = false;
2070	node = tctx->task_list.first;
2071	INIT_WQ_LIST(&tctx->task_list);
2072	spin_unlock_irqrestore(&tctx->task_lock, flags);
2073
2074	while (node) {
2075		req = container_of(node, struct io_kiocb, io_task_work.node);
2076		node = node->next;
2077		if (llist_add(&req->io_task_work.fallback_node,
2078			      &req->ctx->fallback_llist))
2079			schedule_delayed_work(&req->ctx->fallback_work, 1);
2080	}
2081}
2082
2083static void io_req_task_cancel(struct io_kiocb *req)
2084{
2085	struct io_ring_ctx *ctx = req->ctx;
2086
2087	/* ctx is guaranteed to stay alive while we hold uring_lock */
2088	mutex_lock(&ctx->uring_lock);
2089	io_req_complete_failed(req, req->result);
2090	mutex_unlock(&ctx->uring_lock);
2091}
2092
2093static void io_req_task_submit(struct io_kiocb *req)
2094{
2095	struct io_ring_ctx *ctx = req->ctx;
2096
2097	/* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2098	mutex_lock(&ctx->uring_lock);
2099	if (likely(!(req->task->flags & PF_EXITING)))
2100		__io_queue_sqe(req);
2101	else
2102		io_req_complete_failed(req, -EFAULT);
2103	mutex_unlock(&ctx->uring_lock);
2104}
2105
2106static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2107{
2108	req->result = ret;
2109	req->io_task_work.func = io_req_task_cancel;
2110	io_req_task_work_add(req);
2111}
2112
2113static void io_req_task_queue(struct io_kiocb *req)
2114{
2115	req->io_task_work.func = io_req_task_submit;
2116	io_req_task_work_add(req);
2117}
2118
2119static void io_req_task_queue_reissue(struct io_kiocb *req)
2120{
2121	req->io_task_work.func = io_queue_async_work;
2122	io_req_task_work_add(req);
2123}
2124
2125static inline void io_queue_next(struct io_kiocb *req)
2126{
2127	struct io_kiocb *nxt = io_req_find_next(req);
2128
2129	if (nxt)
2130		io_req_task_queue(nxt);
2131}
2132
2133static void io_free_req(struct io_kiocb *req)
2134{
2135	io_queue_next(req);
2136	__io_free_req(req);
2137}
2138
2139struct req_batch {
2140	struct task_struct	*task;
2141	int			task_refs;
2142	int			ctx_refs;
2143};
2144
2145static inline void io_init_req_batch(struct req_batch *rb)
2146{
2147	rb->task_refs = 0;
2148	rb->ctx_refs = 0;
2149	rb->task = NULL;
2150}
2151
2152static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2153				     struct req_batch *rb)
2154{
2155	if (rb->ctx_refs)
2156		percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2157	if (rb->task == current)
2158		current->io_uring->cached_refs += rb->task_refs;
2159	else if (rb->task)
2160		io_put_task(rb->task, rb->task_refs);
2161}
2162
2163static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2164			      struct io_submit_state *state)
2165{
2166	io_queue_next(req);
2167	io_dismantle_req(req);
2168
2169	if (req->task != rb->task) {
2170		if (rb->task)
2171			io_put_task(rb->task, rb->task_refs);
2172		rb->task = req->task;
2173		rb->task_refs = 0;
2174	}
2175	rb->task_refs++;
2176	rb->ctx_refs++;
2177
2178	if (state->free_reqs != ARRAY_SIZE(state->reqs))
2179		state->reqs[state->free_reqs++] = req;
2180	else
2181		list_add(&req->inflight_entry, &state->free_list);
2182}
2183
2184static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2185	__must_hold(&req->ctx->uring_lock)
2186{
2187	struct io_submit_state *state = &ctx->submit_state;
2188	int i, nr = state->compl_nr;
2189	struct req_batch rb;
2190
2191	spin_lock(&ctx->completion_lock);
2192	for (i = 0; i < nr; i++) {
2193		struct io_kiocb *req = state->compl_reqs[i];
2194
2195		__io_cqring_fill_event(ctx, req->user_data, req->result,
2196					req->compl.cflags);
2197	}
2198	io_commit_cqring(ctx);
2199	spin_unlock(&ctx->completion_lock);
2200	io_cqring_ev_posted(ctx);
2201
2202	io_init_req_batch(&rb);
2203	for (i = 0; i < nr; i++) {
2204		struct io_kiocb *req = state->compl_reqs[i];
2205
2206		if (req_ref_put_and_test(req))
2207			io_req_free_batch(&rb, req, &ctx->submit_state);
2208	}
2209
2210	io_req_free_batch_finish(ctx, &rb);
2211	state->compl_nr = 0;
2212}
2213
2214/*
2215 * Drop reference to request, return next in chain (if there is one) if this
2216 * was the last reference to this request.
2217 */
2218static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2219{
2220	struct io_kiocb *nxt = NULL;
2221
2222	if (req_ref_put_and_test(req)) {
2223		nxt = io_req_find_next(req);
2224		__io_free_req(req);
2225	}
2226	return nxt;
2227}
2228
2229static inline void io_put_req(struct io_kiocb *req)
2230{
2231	if (req_ref_put_and_test(req))
2232		io_free_req(req);
2233}
2234
2235static inline void io_put_req_deferred(struct io_kiocb *req)
2236{
2237	if (req_ref_put_and_test(req)) {
2238		req->io_task_work.func = io_free_req;
2239		io_req_task_work_add(req);
2240	}
2241}
2242
2243static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2244{
2245	/* See comment at the top of this file */
2246	smp_rmb();
2247	return __io_cqring_events(ctx);
2248}
2249
2250static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2251{
2252	struct io_rings *rings = ctx->rings;
2253
2254	/* make sure SQ entry isn't read before tail */
2255	return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2256}
2257
2258static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2259{
2260	unsigned int cflags;
2261
2262	cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2263	cflags |= IORING_CQE_F_BUFFER;
2264	req->flags &= ~REQ_F_BUFFER_SELECTED;
2265	kfree(kbuf);
2266	return cflags;
2267}
2268
2269static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2270{
2271	struct io_buffer *kbuf;
2272
2273	kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2274	return io_put_kbuf(req, kbuf);
2275}
2276
2277static inline bool io_run_task_work(void)
2278{
2279	if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2280		__set_current_state(TASK_RUNNING);
2281		tracehook_notify_signal();
2282		return true;
2283	}
2284
2285	return false;
2286}
2287
2288/*
2289 * Find and free completed poll iocbs
2290 */
2291static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2292			       struct list_head *done, bool resubmit)
2293{
2294	struct req_batch rb;
2295	struct io_kiocb *req;
2296
2297	/* order with ->result store in io_complete_rw_iopoll() */
2298	smp_rmb();
2299
2300	io_init_req_batch(&rb);
2301	while (!list_empty(done)) {
2302		int cflags = 0;
2303
2304		req = list_first_entry(done, struct io_kiocb, inflight_entry);
2305		list_del(&req->inflight_entry);
2306
2307		if (READ_ONCE(req->result) == -EAGAIN && resubmit &&
2308		    !(req->flags & REQ_F_DONT_REISSUE)) {
2309			req->iopoll_completed = 0;
2310			io_req_task_queue_reissue(req);
2311			continue;
2312		}
2313
2314		if (req->flags & REQ_F_BUFFER_SELECTED)
2315			cflags = io_put_rw_kbuf(req);
2316
2317		__io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2318		(*nr_events)++;
2319
2320		if (req_ref_put_and_test(req))
2321			io_req_free_batch(&rb, req, &ctx->submit_state);
2322	}
2323
2324	io_commit_cqring(ctx);
2325	io_cqring_ev_posted_iopoll(ctx);
2326	io_req_free_batch_finish(ctx, &rb);
2327}
2328
2329static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2330			long min, bool resubmit)
2331{
2332	struct io_kiocb *req, *tmp;
2333	LIST_HEAD(done);
2334	bool spin;
2335
2336	/*
2337	 * Only spin for completions if we don't have multiple devices hanging
2338	 * off our complete list, and we're under the requested amount.
2339	 */
2340	spin = !ctx->poll_multi_queue && *nr_events < min;
2341
2342	list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2343		struct kiocb *kiocb = &req->rw.kiocb;
2344		int ret;
2345
2346		/*
2347		 * Move completed and retryable entries to our local lists.
2348		 * If we find a request that requires polling, break out
2349		 * and complete those lists first, if we have entries there.
2350		 */
2351		if (READ_ONCE(req->iopoll_completed)) {
2352			list_move_tail(&req->inflight_entry, &done);
2353			continue;
2354		}
2355		if (!list_empty(&done))
2356			break;
2357
2358		ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2359		if (unlikely(ret < 0))
2360			return ret;
2361		else if (ret)
2362			spin = false;
2363
2364		/* iopoll may have completed current req */
2365		if (READ_ONCE(req->iopoll_completed))
2366			list_move_tail(&req->inflight_entry, &done);
2367	}
2368
2369	if (!list_empty(&done))
2370		io_iopoll_complete(ctx, nr_events, &done, resubmit);
2371
2372	return 0;
2373}
2374
2375/*
2376 * We can't just wait for polled events to come to us, we have to actively
2377 * find and complete them.
2378 */
2379static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2380{
2381	if (!(ctx->flags & IORING_SETUP_IOPOLL))
2382		return;
2383
2384	mutex_lock(&ctx->uring_lock);
2385	while (!list_empty(&ctx->iopoll_list)) {
2386		unsigned int nr_events = 0;
2387
2388		io_do_iopoll(ctx, &nr_events, 0, false);
2389
2390		/* let it sleep and repeat later if can't complete a request */
2391		if (nr_events == 0)
2392			break;
2393		/*
2394		 * Ensure we allow local-to-the-cpu processing to take place,
2395		 * in this case we need to ensure that we reap all events.
2396		 * Also let task_work, etc. to progress by releasing the mutex
2397		 */
2398		if (need_resched()) {
2399			mutex_unlock(&ctx->uring_lock);
2400			cond_resched();
2401			mutex_lock(&ctx->uring_lock);
2402		}
2403	}
2404	mutex_unlock(&ctx->uring_lock);
2405}
2406
2407static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2408{
2409	unsigned int nr_events = 0;
2410	int ret = 0;
2411
2412	/*
2413	 * We disallow the app entering submit/complete with polling, but we
2414	 * still need to lock the ring to prevent racing with polled issue
2415	 * that got punted to a workqueue.
2416	 */
2417	mutex_lock(&ctx->uring_lock);
2418	/*
2419	 * Don't enter poll loop if we already have events pending.
2420	 * If we do, we can potentially be spinning for commands that
2421	 * already triggered a CQE (eg in error).
2422	 */
2423	if (test_bit(0, &ctx->check_cq_overflow))
2424		__io_cqring_overflow_flush(ctx, false);
2425	if (io_cqring_events(ctx))
2426		goto out;
2427	do {
2428		/*
2429		 * If a submit got punted to a workqueue, we can have the
2430		 * application entering polling for a command before it gets
2431		 * issued. That app will hold the uring_lock for the duration
2432		 * of the poll right here, so we need to take a breather every
2433		 * now and then to ensure that the issue has a chance to add
2434		 * the poll to the issued list. Otherwise we can spin here
2435		 * forever, while the workqueue is stuck trying to acquire the
2436		 * very same mutex.
2437		 */
2438		if (list_empty(&ctx->iopoll_list)) {
2439			u32 tail = ctx->cached_cq_tail;
2440
2441			mutex_unlock(&ctx->uring_lock);
2442			io_run_task_work();
2443			mutex_lock(&ctx->uring_lock);
2444
2445			/* some requests don't go through iopoll_list */
2446			if (tail != ctx->cached_cq_tail ||
2447			    list_empty(&ctx->iopoll_list))
2448				break;
2449		}
2450		ret = io_do_iopoll(ctx, &nr_events, min, true);
2451	} while (!ret && nr_events < min && !need_resched());
2452out:
2453	mutex_unlock(&ctx->uring_lock);
2454	return ret;
2455}
2456
2457static void kiocb_end_write(struct io_kiocb *req)
2458{
2459	/*
2460	 * Tell lockdep we inherited freeze protection from submission
2461	 * thread.
2462	 */
2463	if (req->flags & REQ_F_ISREG) {
2464		struct super_block *sb = file_inode(req->file)->i_sb;
2465
2466		__sb_writers_acquired(sb, SB_FREEZE_WRITE);
2467		sb_end_write(sb);
2468	}
2469}
2470
2471#ifdef CONFIG_BLOCK
2472static bool io_resubmit_prep(struct io_kiocb *req)
2473{
2474	struct io_async_rw *rw = req->async_data;
2475
2476	if (!rw)
2477		return !io_req_prep_async(req);
2478	/* may have left rw->iter inconsistent on -EIOCBQUEUED */
2479	iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2480	return true;
2481}
2482
2483static bool io_rw_should_reissue(struct io_kiocb *req)
2484{
2485	umode_t mode = file_inode(req->file)->i_mode;
2486	struct io_ring_ctx *ctx = req->ctx;
2487
2488	if (!S_ISBLK(mode) && !S_ISREG(mode))
2489		return false;
2490	if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2491	    !(ctx->flags & IORING_SETUP_IOPOLL)))
2492		return false;
2493	/*
2494	 * If ref is dying, we might be running poll reap from the exit work.
2495	 * Don't attempt to reissue from that path, just let it fail with
2496	 * -EAGAIN.
2497	 */
2498	if (percpu_ref_is_dying(&ctx->refs))
2499		return false;
2500	/*
2501	 * Play it safe and assume not safe to re-import and reissue if we're
2502	 * not in the original thread group (or in task context).
2503	 */
2504	if (!same_thread_group(req->task, current) || !in_task())
2505		return false;
2506	return true;
2507}
2508#else
2509static bool io_resubmit_prep(struct io_kiocb *req)
2510{
2511	return false;
2512}
2513static bool io_rw_should_reissue(struct io_kiocb *req)
2514{
2515	return false;
2516}
2517#endif
2518
2519static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2520{
2521	if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2522		kiocb_end_write(req);
2523	if (res != req->result) {
2524		if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2525		    io_rw_should_reissue(req)) {
2526			req->flags |= REQ_F_REISSUE;
2527			return true;
2528		}
2529		req_set_fail(req);
2530		req->result = res;
2531	}
2532	return false;
2533}
2534
2535static void io_req_task_complete(struct io_kiocb *req)
2536{
2537	int cflags = 0;
2538
2539	if (req->flags & REQ_F_BUFFER_SELECTED)
2540		cflags = io_put_rw_kbuf(req);
2541	__io_req_complete(req, 0, req->result, cflags);
2542}
2543
2544static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2545			     unsigned int issue_flags)
2546{
2547	if (__io_complete_rw_common(req, res))
2548		return;
2549	io_req_task_complete(req);
2550}
2551
2552static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2553{
2554	struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2555
2556	if (__io_complete_rw_common(req, res))
2557		return;
2558	req->result = res;
2559	req->io_task_work.func = io_req_task_complete;
2560	io_req_task_work_add(req);
2561}
2562
2563static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2564{
2565	struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2566
2567	if (kiocb->ki_flags & IOCB_WRITE)
2568		kiocb_end_write(req);
2569	if (unlikely(res != req->result)) {
2570		if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2571		    io_resubmit_prep(req))) {
2572			req_set_fail(req);
2573			req->flags |= REQ_F_DONT_REISSUE;
2574		}
2575	}
2576
2577	WRITE_ONCE(req->result, res);
2578	/* order with io_iopoll_complete() checking ->result */
2579	smp_wmb();
2580	WRITE_ONCE(req->iopoll_completed, 1);
2581}
2582
2583/*
2584 * After the iocb has been issued, it's safe to be found on the poll list.
2585 * Adding the kiocb to the list AFTER submission ensures that we don't
2586 * find it from a io_do_iopoll() thread before the issuer is done
2587 * accessing the kiocb cookie.
2588 */
2589static void io_iopoll_req_issued(struct io_kiocb *req)
2590{
2591	struct io_ring_ctx *ctx = req->ctx;
2592	const bool in_async = io_wq_current_is_worker();
2593
2594	/* workqueue context doesn't hold uring_lock, grab it now */
2595	if (unlikely(in_async))
2596		mutex_lock(&ctx->uring_lock);
2597
2598	/*
2599	 * Track whether we have multiple files in our lists. This will impact
2600	 * how we do polling eventually, not spinning if we're on potentially
2601	 * different devices.
2602	 */
2603	if (list_empty(&ctx->iopoll_list)) {
2604		ctx->poll_multi_queue = false;
2605	} else if (!ctx->poll_multi_queue) {
2606		struct io_kiocb *list_req;
2607		unsigned int queue_num0, queue_num1;
2608
2609		list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2610						inflight_entry);
2611
2612		if (list_req->file != req->file) {
2613			ctx->poll_multi_queue = true;
2614		} else {
2615			queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2616			queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2617			if (queue_num0 != queue_num1)
2618				ctx->poll_multi_queue = true;
2619		}
2620	}
2621
2622	/*
2623	 * For fast devices, IO may have already completed. If it has, add
2624	 * it to the front so we find it first.
2625	 */
2626	if (READ_ONCE(req->iopoll_completed))
2627		list_add(&req->inflight_entry, &ctx->iopoll_list);
2628	else
2629		list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2630
2631	if (unlikely(in_async)) {
2632		/*
2633		 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2634		 * in sq thread task context or in io worker task context. If
2635		 * current task context is sq thread, we don't need to check
2636		 * whether should wake up sq thread.
2637		 */
2638		if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2639		    wq_has_sleeper(&ctx->sq_data->wait))
2640			wake_up(&ctx->sq_data->wait);
2641
2642		mutex_unlock(&ctx->uring_lock);
2643	}
2644}
2645
2646static bool io_bdev_nowait(struct block_device *bdev)
2647{
2648	return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2649}
2650
2651/*
2652 * If we tracked the file through the SCM inflight mechanism, we could support
2653 * any file. For now, just ensure that anything potentially problematic is done
2654 * inline.
2655 */
2656static bool __io_file_supports_nowait(struct file *file, int rw)
2657{
2658	umode_t mode = file_inode(file)->i_mode;
2659
2660	if (S_ISBLK(mode)) {
2661		if (IS_ENABLED(CONFIG_BLOCK) &&
2662		    io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2663			return true;
2664		return false;
2665	}
2666	if (S_ISSOCK(mode))
2667		return true;
2668	if (S_ISREG(mode)) {
2669		if (IS_ENABLED(CONFIG_BLOCK) &&
2670		    io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2671		    file->f_op != &io_uring_fops)
2672			return true;
2673		return false;
2674	}
2675
2676	/* any ->read/write should understand O_NONBLOCK */
2677	if (file->f_flags & O_NONBLOCK)
2678		return true;
2679
2680	if (!(file->f_mode & FMODE_NOWAIT))
2681		return false;
2682
2683	if (rw == READ)
2684		return file->f_op->read_iter != NULL;
2685
2686	return file->f_op->write_iter != NULL;
2687}
2688
2689static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2690{
2691	if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2692		return true;
2693	else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2694		return true;
2695
2696	return __io_file_supports_nowait(req->file, rw);
2697}
2698
2699static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2700{
2701	struct io_ring_ctx *ctx = req->ctx;
2702	struct kiocb *kiocb = &req->rw.kiocb;
2703	struct file *file = req->file;
2704	unsigned ioprio;
2705	int ret;
2706
2707	if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2708		req->flags |= REQ_F_ISREG;
2709
2710	kiocb->ki_pos = READ_ONCE(sqe->off);
2711	if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2712		req->flags |= REQ_F_CUR_POS;
2713		kiocb->ki_pos = file->f_pos;
2714	}
2715	kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2716	kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2717	ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2718	if (unlikely(ret))
2719		return ret;
2720
2721	/* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2722	if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2723		req->flags |= REQ_F_NOWAIT;
2724
2725	ioprio = READ_ONCE(sqe->ioprio);
2726	if (ioprio) {
2727		ret = ioprio_check_cap(ioprio);
2728		if (ret)
2729			return ret;
2730
2731		kiocb->ki_ioprio = ioprio;
2732	} else
2733		kiocb->ki_ioprio = get_current_ioprio();
2734
2735	if (ctx->flags & IORING_SETUP_IOPOLL) {
2736		if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2737		    !kiocb->ki_filp->f_op->iopoll)
2738			return -EOPNOTSUPP;
2739
2740		kiocb->ki_flags |= IOCB_HIPRI;
2741		kiocb->ki_complete = io_complete_rw_iopoll;
2742		req->iopoll_completed = 0;
2743	} else {
2744		if (kiocb->ki_flags & IOCB_HIPRI)
2745			return -EINVAL;
2746		kiocb->ki_complete = io_complete_rw;
2747	}
2748
2749	if (req->opcode == IORING_OP_READ_FIXED ||
2750	    req->opcode == IORING_OP_WRITE_FIXED) {
2751		req->imu = NULL;
2752		io_req_set_rsrc_node(req);
2753	}
2754
2755	req->rw.addr = READ_ONCE(sqe->addr);
2756	req->rw.len = READ_ONCE(sqe->len);
2757	req->buf_index = READ_ONCE(sqe->buf_index);
2758	return 0;
2759}
2760
2761static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2762{
2763	switch (ret) {
2764	case -EIOCBQUEUED:
2765		break;
2766	case -ERESTARTSYS:
2767	case -ERESTARTNOINTR:
2768	case -ERESTARTNOHAND:
2769	case -ERESTART_RESTARTBLOCK:
2770		/*
2771		 * We can't just restart the syscall, since previously
2772		 * submitted sqes may already be in progress. Just fail this
2773		 * IO with EINTR.
2774		 */
2775		ret = -EINTR;
2776		fallthrough;
2777	default:
2778		kiocb->ki_complete(kiocb, ret, 0);
2779	}
2780}
2781
2782static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2783		       unsigned int issue_flags)
2784{
2785	struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2786	struct io_async_rw *io = req->async_data;
2787	bool check_reissue = kiocb->ki_complete == io_complete_rw;
2788
2789	/* add previously done IO, if any */
2790	if (io && io->bytes_done > 0) {
2791		if (ret < 0)
2792			ret = io->bytes_done;
2793		else
2794			ret += io->bytes_done;
2795	}
2796
2797	if (req->flags & REQ_F_CUR_POS)
2798		req->file->f_pos = kiocb->ki_pos;
2799	if (ret >= 0 && check_reissue)
2800		__io_complete_rw(req, ret, 0, issue_flags);
2801	else
2802		io_rw_done(kiocb, ret);
2803
2804	if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2805		req->flags &= ~REQ_F_REISSUE;
2806		if (io_resubmit_prep(req)) {
2807			io_req_task_queue_reissue(req);
2808		} else {
2809			int cflags = 0;
2810
2811			req_set_fail(req);
2812			if (req->flags & REQ_F_BUFFER_SELECTED)
2813				cflags = io_put_rw_kbuf(req);
2814			__io_req_complete(req, issue_flags, ret, cflags);
2815		}
2816	}
2817}
2818
2819static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2820			     struct io_mapped_ubuf *imu)
2821{
2822	size_t len = req->rw.len;
2823	u64 buf_end, buf_addr = req->rw.addr;
2824	size_t offset;
2825
2826	if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2827		return -EFAULT;
2828	/* not inside the mapped region */
2829	if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2830		return -EFAULT;
2831
2832	/*
2833	 * May not be a start of buffer, set size appropriately
2834	 * and advance us to the beginning.
2835	 */
2836	offset = buf_addr - imu->ubuf;
2837	iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2838
2839	if (offset) {
2840		/*
2841		 * Don't use iov_iter_advance() here, as it's really slow for
2842		 * using the latter parts of a big fixed buffer - it iterates
2843		 * over each segment manually. We can cheat a bit here, because
2844		 * we know that:
2845		 *
2846		 * 1) it's a BVEC iter, we set it up
2847		 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2848		 *    first and last bvec
2849		 *
2850		 * So just find our index, and adjust the iterator afterwards.
2851		 * If the offset is within the first bvec (or the whole first
2852		 * bvec, just use iov_iter_advance(). This makes it easier
2853		 * since we can just skip the first segment, which may not
2854		 * be PAGE_SIZE aligned.
2855		 */
2856		const struct bio_vec *bvec = imu->bvec;
2857
2858		if (offset <= bvec->bv_len) {
2859			iov_iter_advance(iter, offset);
2860		} else {
2861			unsigned long seg_skip;
2862
2863			/* skip first vec */
2864			offset -= bvec->bv_len;
2865			seg_skip = 1 + (offset >> PAGE_SHIFT);
2866
2867			iter->bvec = bvec + seg_skip;
2868			iter->nr_segs -= seg_skip;
2869			iter->count -= bvec->bv_len + offset;
2870			iter->iov_offset = offset & ~PAGE_MASK;
2871		}
2872	}
2873
2874	return 0;
2875}
2876
2877static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2878{
2879	struct io_ring_ctx *ctx = req->ctx;
2880	struct io_mapped_ubuf *imu = req->imu;
2881	u16 index, buf_index = req->buf_index;
2882
2883	if (likely(!imu)) {
2884		if (unlikely(buf_index >= ctx->nr_user_bufs))
2885			return -EFAULT;
2886		index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2887		imu = READ_ONCE(ctx->user_bufs[index]);
2888		req->imu = imu;
2889	}
2890	return __io_import_fixed(req, rw, iter, imu);
2891}
2892
2893static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2894{
2895	if (needs_lock)
2896		mutex_unlock(&ctx->uring_lock);
2897}
2898
2899static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2900{
2901	/*
2902	 * "Normal" inline submissions always hold the uring_lock, since we
2903	 * grab it from the system call. Same is true for the SQPOLL offload.
2904	 * The only exception is when we've detached the request and issue it
2905	 * from an async worker thread, grab the lock for that case.
2906	 */
2907	if (needs_lock)
2908		mutex_lock(&ctx->uring_lock);
2909}
2910
2911static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2912					  int bgid, struct io_buffer *kbuf,
2913					  bool needs_lock)
2914{
2915	struct io_buffer *head;
2916
2917	if (req->flags & REQ_F_BUFFER_SELECTED)
2918		return kbuf;
2919
2920	io_ring_submit_lock(req->ctx, needs_lock);
2921
2922	lockdep_assert_held(&req->ctx->uring_lock);
2923
2924	head = xa_load(&req->ctx->io_buffers, bgid);
2925	if (head) {
2926		if (!list_empty(&head->list)) {
2927			kbuf = list_last_entry(&head->list, struct io_buffer,
2928							list);
2929			list_del(&kbuf->list);
2930		} else {
2931			kbuf = head;
2932			xa_erase(&req->ctx->io_buffers, bgid);
2933		}
2934		if (*len > kbuf->len)
2935			*len = kbuf->len;
2936	} else {
2937		kbuf = ERR_PTR(-ENOBUFS);
2938	}
2939
2940	io_ring_submit_unlock(req->ctx, needs_lock);
2941
2942	return kbuf;
2943}
2944
2945static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2946					bool needs_lock)
2947{
2948	struct io_buffer *kbuf;
2949	u16 bgid;
2950
2951	kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2952	bgid = req->buf_index;
2953	kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2954	if (IS_ERR(kbuf))
2955		return kbuf;
2956	req->rw.addr = (u64) (unsigned long) kbuf;
2957	req->flags |= REQ_F_BUFFER_SELECTED;
2958	return u64_to_user_ptr(kbuf->addr);
2959}
2960
2961#ifdef CONFIG_COMPAT
2962static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2963				bool needs_lock)
2964{
2965	struct compat_iovec __user *uiov;
2966	compat_ssize_t clen;
2967	void __user *buf;
2968	ssize_t len;
2969
2970	uiov = u64_to_user_ptr(req->rw.addr);
2971	if (!access_ok(uiov, sizeof(*uiov)))
2972		return -EFAULT;
2973	if (__get_user(clen, &uiov->iov_len))
2974		return -EFAULT;
2975	if (clen < 0)
2976		return -EINVAL;
2977
2978	len = clen;
2979	buf = io_rw_buffer_select(req, &len, needs_lock);
2980	if (IS_ERR(buf))
2981		return PTR_ERR(buf);
2982	iov[0].iov_base = buf;
2983	iov[0].iov_len = (compat_size_t) len;
2984	return 0;
2985}
2986#endif
2987
2988static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2989				      bool needs_lock)
2990{
2991	struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2992	void __user *buf;
2993	ssize_t len;
2994
2995	if (copy_from_user(iov, uiov, sizeof(*uiov)))
2996		return -EFAULT;
2997
2998	len = iov[0].iov_len;
2999	if (len < 0)
3000		return -EINVAL;
3001	buf = io_rw_buffer_select(req, &len, needs_lock);
3002	if (IS_ERR(buf))
3003		return PTR_ERR(buf);
3004	iov[0].iov_base = buf;
3005	iov[0].iov_len = len;
3006	return 0;
3007}
3008
3009static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3010				    bool needs_lock)
3011{
3012	if (req->flags & REQ_F_BUFFER_SELECTED) {
3013		struct io_buffer *kbuf;
3014
3015		kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3016		iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3017		iov[0].iov_len = kbuf->len;
3018		return 0;
3019	}
3020	if (req->rw.len != 1)
3021		return -EINVAL;
3022
3023#ifdef CONFIG_COMPAT
3024	if (req->ctx->compat)
3025		return io_compat_import(req, iov, needs_lock);
3026#endif
3027
3028	return __io_iov_buffer_select(req, iov, needs_lock);
3029}
3030
3031static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3032			   struct iov_iter *iter, bool needs_lock)
3033{
3034	void __user *buf = u64_to_user_ptr(req->rw.addr);
3035	size_t sqe_len = req->rw.len;
3036	u8 opcode = req->opcode;
3037	ssize_t ret;
3038
3039	if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3040		*iovec = NULL;
3041		return io_import_fixed(req, rw, iter);
3042	}
3043
3044	/* buffer index only valid with fixed read/write, or buffer select  */
3045	if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3046		return -EINVAL;
3047
3048	if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3049		if (req->flags & REQ_F_BUFFER_SELECT) {
3050			buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3051			if (IS_ERR(buf))
3052				return PTR_ERR(buf);
3053			req->rw.len = sqe_len;
3054		}
3055
3056		ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3057		*iovec = NULL;
3058		return ret;
3059	}
3060
3061	if (req->flags & REQ_F_BUFFER_SELECT) {
3062		ret = io_iov_buffer_select(req, *iovec, needs_lock);
3063		if (!ret)
3064			iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3065		*iovec = NULL;
3066		return ret;
3067	}
3068
3069	return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3070			      req->ctx->compat);
3071}
3072
3073static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3074{
3075	return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3076}
3077
3078/*
3079 * For files that don't have ->read_iter() and ->write_iter(), handle them
3080 * by looping over ->read() or ->write() manually.
3081 */
3082static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3083{
3084	struct kiocb *kiocb = &req->rw.kiocb;
3085	struct file *file = req->file;
3086	ssize_t ret = 0;
3087
3088	/*
3089	 * Don't support polled IO through this interface, and we can't
3090	 * support non-blocking either. For the latter, this just causes
3091	 * the kiocb to be handled from an async context.
3092	 */
3093	if (kiocb->ki_flags & IOCB_HIPRI)
3094		return -EOPNOTSUPP;
3095	if (kiocb->ki_flags & IOCB_NOWAIT)
3096		return -EAGAIN;
3097
3098	while (iov_iter_count(iter)) {
3099		struct iovec iovec;
3100		ssize_t nr;
3101
3102		if (!iov_iter_is_bvec(iter)) {
3103			iovec = iov_iter_iovec(iter);
3104		} else {
3105			iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3106			iovec.iov_len = req->rw.len;
3107		}
3108
3109		if (rw == READ) {
3110			nr = file->f_op->read(file, iovec.iov_base,
3111					      iovec.iov_len, io_kiocb_ppos(kiocb));
3112		} else {
3113			nr = file->f_op->write(file, iovec.iov_base,
3114					       iovec.iov_len, io_kiocb_ppos(kiocb));
3115		}
3116
3117		if (nr < 0) {
3118			if (!ret)
3119				ret = nr;
3120			break;
3121		}
3122		ret += nr;
3123		if (nr != iovec.iov_len)
3124			break;
3125		req->rw.len -= nr;
3126		req->rw.addr += nr;
3127		iov_iter_advance(iter, nr);
3128	}
3129
3130	return ret;
3131}
3132
3133static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3134			  const struct iovec *fast_iov, struct iov_iter *iter)
3135{
3136	struct io_async_rw *rw = req->async_data;
3137
3138	memcpy(&rw->iter, iter, sizeof(*iter));
3139	rw->free_iovec = iovec;
3140	rw->bytes_done = 0;
3141	/* can only be fixed buffers, no need to do anything */
3142	if (iov_iter_is_bvec(iter))
3143		return;
3144	if (!iovec) {
3145		unsigned iov_off = 0;
3146
3147		rw->iter.iov = rw->fast_iov;
3148		if (iter->iov != fast_iov) {
3149			iov_off = iter->iov - fast_iov;
3150			rw->iter.iov += iov_off;
3151		}
3152		if (rw->fast_iov != fast_iov)
3153			memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3154			       sizeof(struct iovec) * iter->nr_segs);
3155	} else {
3156		req->flags |= REQ_F_NEED_CLEANUP;
3157	}
3158}
3159
3160static inline int io_alloc_async_data(struct io_kiocb *req)
3161{
3162	WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3163	req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3164	return req->async_data == NULL;
3165}
3166
3167static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3168			     const struct iovec *fast_iov,
3169			     struct iov_iter *iter, bool force)
3170{
3171	if (!force && !io_op_defs[req->opcode].needs_async_setup)
3172		return 0;
3173	if (!req->async_data) {
3174		if (io_alloc_async_data(req)) {
3175			kfree(iovec);
3176			return -ENOMEM;
3177		}
3178
3179		io_req_map_rw(req, iovec, fast_iov, iter);
3180	}
3181	return 0;
3182}
3183
3184static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3185{
3186	struct io_async_rw *iorw = req->async_data;
3187	struct iovec *iov = iorw->fast_iov;
3188	int ret;
3189
3190	ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3191	if (unlikely(ret < 0))
3192		return ret;
3193
3194	iorw->bytes_done = 0;
3195	iorw->free_iovec = iov;
3196	if (iov)
3197		req->flags |= REQ_F_NEED_CLEANUP;
3198	return 0;
3199}
3200
3201static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3202{
3203	if (unlikely(!(req->file->f_mode & FMODE_READ)))
3204		return -EBADF;
3205	return io_prep_rw(req, sqe);
3206}
3207
3208/*
3209 * This is our waitqueue callback handler, registered through lock_page_async()
3210 * when we initially tried to do the IO with the iocb armed our waitqueue.
3211 * This gets called when the page is unlocked, and we generally expect that to
3212 * happen when the page IO is completed and the page is now uptodate. This will
3213 * queue a task_work based retry of the operation, attempting to copy the data
3214 * again. If the latter fails because the page was NOT uptodate, then we will
3215 * do a thread based blocking retry of the operation. That's the unexpected
3216 * slow path.
3217 */
3218static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3219			     int sync, void *arg)
3220{
3221	struct wait_page_queue *wpq;
3222	struct io_kiocb *req = wait->private;
3223	struct wait_page_key *key = arg;
3224
3225	wpq = container_of(wait, struct wait_page_queue, wait);
3226
3227	if (!wake_page_match(wpq, key))
3228		return 0;
3229
3230	req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3231	list_del_init(&wait->entry);
3232	io_req_task_queue(req);
3233	return 1;
3234}
3235
3236/*
3237 * This controls whether a given IO request should be armed for async page
3238 * based retry. If we return false here, the request is handed to the async
3239 * worker threads for retry. If we're doing buffered reads on a regular file,
3240 * we prepare a private wait_page_queue entry and retry the operation. This
3241 * will either succeed because the page is now uptodate and unlocked, or it
3242 * will register a callback when the page is unlocked at IO completion. Through
3243 * that callback, io_uring uses task_work to setup a retry of the operation.
3244 * That retry will attempt the buffered read again. The retry will generally
3245 * succeed, or in rare cases where it fails, we then fall back to using the
3246 * async worker threads for a blocking retry.
3247 */
3248static bool io_rw_should_retry(struct io_kiocb *req)
3249{
3250	struct io_async_rw *rw = req->async_data;
3251	struct wait_page_queue *wait = &rw->wpq;
3252	struct kiocb *kiocb = &req->rw.kiocb;
3253
3254	/* never retry for NOWAIT, we just complete with -EAGAIN */
3255	if (req->flags & REQ_F_NOWAIT)
3256		return false;
3257
3258	/* Only for buffered IO */
3259	if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3260		return false;
3261
3262	/*
3263	 * just use poll if we can, and don't attempt if the fs doesn't
3264	 * support callback based unlocks
3265	 */
3266	if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3267		return false;
3268
3269	wait->wait.func = io_async_buf_func;
3270	wait->wait.private = req;
3271	wait->wait.flags = 0;
3272	INIT_LIST_HEAD(&wait->wait.entry);
3273	kiocb->ki_flags |= IOCB_WAITQ;
3274	kiocb->ki_flags &= ~IOCB_NOWAIT;
3275	kiocb->ki_waitq = wait;
3276	return true;
3277}
3278
3279static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3280{
3281	if (req->file->f_op->read_iter)
3282		return call_read_iter(req->file, &req->rw.kiocb, iter);
3283	else if (req->file->f_op->read)
3284		return loop_rw_iter(READ, req, iter);
3285	else
3286		return -EINVAL;
3287}
3288
3289static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3290{
3291	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3292	struct kiocb *kiocb = &req->rw.kiocb;
3293	struct iov_iter __iter, *iter = &__iter;
3294	struct io_async_rw *rw = req->async_data;
3295	ssize_t io_size, ret, ret2;
3296	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3297
3298	if (rw) {
3299		iter = &rw->iter;
3300		iovec = NULL;
3301	} else {
3302		ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3303		if (ret < 0)
3304			return ret;
3305	}
3306	io_size = iov_iter_count(iter);
3307	req->result = io_size;
3308
3309	/* Ensure we clear previously set non-block flag */
3310	if (!force_nonblock)
3311		kiocb->ki_flags &= ~IOCB_NOWAIT;
3312	else
3313		kiocb->ki_flags |= IOCB_NOWAIT;
3314
3315	/* If the file doesn't support async, just async punt */
3316	if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3317		ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3318		return ret ?: -EAGAIN;
3319	}
3320
3321	ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3322	if (unlikely(ret)) {
3323		kfree(iovec);
3324		return ret;
3325	}
3326
3327	ret = io_iter_do_read(req, iter);
3328
3329	if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3330		req->flags &= ~REQ_F_REISSUE;
3331		/* IOPOLL retry should happen for io-wq threads */
3332		if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3333			goto done;
3334		/* no retry on NONBLOCK nor RWF_NOWAIT */
3335		if (req->flags & REQ_F_NOWAIT)
3336			goto done;
3337		/* some cases will consume bytes even on error returns */
3338		iov_iter_revert(iter, io_size - iov_iter_count(iter));
3339		ret = 0;
3340	} else if (ret == -EIOCBQUEUED) {
3341		goto out_free;
3342	} else if (ret <= 0 || ret == io_size || !force_nonblock ||
3343		   (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3344		/* read all, failed, already did sync or don't want to retry */
3345		goto done;
3346	}
3347
3348	ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3349	if (ret2)
3350		return ret2;
3351
3352	iovec = NULL;
3353	rw = req->async_data;
3354	/* now use our persistent iterator, if we aren't already */
3355	iter = &rw->iter;
3356
3357	do {
3358		io_size -= ret;
3359		rw->bytes_done += ret;
3360		/* if we can retry, do so with the callbacks armed */
3361		if (!io_rw_should_retry(req)) {
3362			kiocb->ki_flags &= ~IOCB_WAITQ;
3363			return -EAGAIN;
3364		}
3365
3366		/*
3367		 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3368		 * we get -EIOCBQUEUED, then we'll get a notification when the
3369		 * desired page gets unlocked. We can also get a partial read
3370		 * here, and if we do, then just retry at the new offset.
3371		 */
3372		ret = io_iter_do_read(req, iter);
3373		if (ret == -EIOCBQUEUED)
3374			return 0;
3375		/* we got some bytes, but not all. retry. */
3376		kiocb->ki_flags &= ~IOCB_WAITQ;
3377	} while (ret > 0 && ret < io_size);
3378done:
3379	kiocb_done(kiocb, ret, issue_flags);
3380out_free:
3381	/* it's faster to check here then delegate to kfree */
3382	if (iovec)
3383		kfree(iovec);
3384	return 0;
3385}
3386
3387static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3388{
3389	if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3390		return -EBADF;
3391	return io_prep_rw(req, sqe);
3392}
3393
3394static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3395{
3396	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3397	struct kiocb *kiocb = &req->rw.kiocb;
3398	struct iov_iter __iter, *iter = &__iter;
3399	struct io_async_rw *rw = req->async_data;
3400	ssize_t ret, ret2, io_size;
3401	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3402
3403	if (rw) {
3404		iter = &rw->iter;
3405		iovec = NULL;
3406	} else {
3407		ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3408		if (ret < 0)
3409			return ret;
3410	}
3411	io_size = iov_iter_count(iter);
3412	req->result = io_size;
3413
3414	/* Ensure we clear previously set non-block flag */
3415	if (!force_nonblock)
3416		kiocb->ki_flags &= ~IOCB_NOWAIT;
3417	else
3418		kiocb->ki_flags |= IOCB_NOWAIT;
3419
3420	/* If the file doesn't support async, just async punt */
3421	if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3422		goto copy_iov;
3423
3424	/* file path doesn't support NOWAIT for non-direct_IO */
3425	if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3426	    (req->flags & REQ_F_ISREG))
3427		goto copy_iov;
3428
3429	ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3430	if (unlikely(ret))
3431		goto out_free;
3432
3433	/*
3434	 * Open-code file_start_write here to grab freeze protection,
3435	 * which will be released by another thread in
3436	 * io_complete_rw().  Fool lockdep by telling it the lock got
3437	 * released so that it doesn't complain about the held lock when
3438	 * we return to userspace.
3439	 */
3440	if (req->flags & REQ_F_ISREG) {
3441		sb_start_write(file_inode(req->file)->i_sb);
3442		__sb_writers_release(file_inode(req->file)->i_sb,
3443					SB_FREEZE_WRITE);
3444	}
3445	kiocb->ki_flags |= IOCB_WRITE;
3446
3447	if (req->file->f_op->write_iter)
3448		ret2 = call_write_iter(req->file, kiocb, iter);
3449	else if (req->file->f_op->write)
3450		ret2 = loop_rw_iter(WRITE, req, iter);
3451	else
3452		ret2 = -EINVAL;
3453
3454	if (req->flags & REQ_F_REISSUE) {
3455		req->flags &= ~REQ_F_REISSUE;
3456		ret2 = -EAGAIN;
3457	}
3458
3459	/*
3460	 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3461	 * retry them without IOCB_NOWAIT.
3462	 */
3463	if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3464		ret2 = -EAGAIN;
3465	/* no retry on NONBLOCK nor RWF_NOWAIT */
3466	if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3467		goto done;
3468	if (!force_nonblock || ret2 != -EAGAIN) {
3469		/* IOPOLL retry should happen for io-wq threads */
3470		if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3471			goto copy_iov;
3472done:
3473		kiocb_done(kiocb, ret2, issue_flags);
3474	} else {
3475copy_iov:
3476		/* some cases will consume bytes even on error returns */
3477		iov_iter_revert(iter, io_size - iov_iter_count(iter));
3478		ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3479		return ret ?: -EAGAIN;
3480	}
3481out_free:
3482	/* it's reportedly faster than delegating the null check to kfree() */
3483	if (iovec)
3484		kfree(iovec);
3485	return ret;
3486}
3487
3488static int io_renameat_prep(struct io_kiocb *req,
3489			    const struct io_uring_sqe *sqe)
3490{
3491	struct io_rename *ren = &req->rename;
3492	const char __user *oldf, *newf;
3493
3494	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3495		return -EINVAL;
3496	if (sqe->ioprio || sqe->buf_index)
3497		return -EINVAL;
3498	if (unlikely(req->flags & REQ_F_FIXED_FILE))
3499		return -EBADF;
3500
3501	ren->old_dfd = READ_ONCE(sqe->fd);
3502	oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3503	newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3504	ren->new_dfd = READ_ONCE(sqe->len);
3505	ren->flags = READ_ONCE(sqe->rename_flags);
3506
3507	ren->oldpath = getname(oldf);
3508	if (IS_ERR(ren->oldpath))
3509		return PTR_ERR(ren->oldpath);
3510
3511	ren->newpath = getname(newf);
3512	if (IS_ERR(ren->newpath)) {
3513		putname(ren->oldpath);
3514		return PTR_ERR(ren->newpath);
3515	}
3516
3517	req->flags |= REQ_F_NEED_CLEANUP;
3518	return 0;
3519}
3520
3521static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3522{
3523	struct io_rename *ren = &req->rename;
3524	int ret;
3525
3526	if (issue_flags & IO_URING_F_NONBLOCK)
3527		return -EAGAIN;
3528
3529	ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3530				ren->newpath, ren->flags);
3531
3532	req->flags &= ~REQ_F_NEED_CLEANUP;
3533	if (ret < 0)
3534		req_set_fail(req);
3535	io_req_complete(req, ret);
3536	return 0;
3537}
3538
3539static int io_unlinkat_prep(struct io_kiocb *req,
3540			    const struct io_uring_sqe *sqe)
3541{
3542	struct io_unlink *un = &req->unlink;
3543	const char __user *fname;
3544
3545	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3546		return -EINVAL;
3547	if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3548		return -EINVAL;
3549	if (unlikely(req->flags & REQ_F_FIXED_FILE))
3550		return -EBADF;
3551
3552	un->dfd = READ_ONCE(sqe->fd);
3553
3554	un->flags = READ_ONCE(sqe->unlink_flags);
3555	if (un->flags & ~AT_REMOVEDIR)
3556		return -EINVAL;
3557
3558	fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3559	un->filename = getname(fname);
3560	if (IS_ERR(un->filename))
3561		return PTR_ERR(un->filename);
3562
3563	req->flags |= REQ_F_NEED_CLEANUP;
3564	return 0;
3565}
3566
3567static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3568{
3569	struct io_unlink *un = &req->unlink;
3570	int ret;
3571
3572	if (issue_flags & IO_URING_F_NONBLOCK)
3573		return -EAGAIN;
3574
3575	if (un->flags & AT_REMOVEDIR)
3576		ret = do_rmdir(un->dfd, un->filename);
3577	else
3578		ret = do_unlinkat(un->dfd, un->filename);
3579
3580	req->flags &= ~REQ_F_NEED_CLEANUP;
3581	if (ret < 0)
3582		req_set_fail(req);
3583	io_req_complete(req, ret);
3584	return 0;
3585}
3586
3587static int io_shutdown_prep(struct io_kiocb *req,
3588			    const struct io_uring_sqe *sqe)
3589{
3590#if defined(CONFIG_NET)
3591	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3592		return -EINVAL;
3593	if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3594	    sqe->buf_index)
3595		return -EINVAL;
3596
3597	req->shutdown.how = READ_ONCE(sqe->len);
3598	return 0;
3599#else
3600	return -EOPNOTSUPP;
3601#endif
3602}
3603
3604static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3605{
3606#if defined(CONFIG_NET)
3607	struct socket *sock;
3608	int ret;
3609
3610	if (issue_flags & IO_URING_F_NONBLOCK)
3611		return -EAGAIN;
3612
3613	sock = sock_from_file(req->file);
3614	if (unlikely(!sock))
3615		return -ENOTSOCK;
3616
3617	ret = __sys_shutdown_sock(sock, req->shutdown.how);
3618	if (ret < 0)
3619		req_set_fail(req);
3620	io_req_complete(req, ret);
3621	return 0;
3622#else
3623	return -EOPNOTSUPP;
3624#endif
3625}
3626
3627static int __io_splice_prep(struct io_kiocb *req,
3628			    const struct io_uring_sqe *sqe)
3629{
3630	struct io_splice *sp = &req->splice;
3631	unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3632
3633	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3634		return -EINVAL;
3635
3636	sp->file_in = NULL;
3637	sp->len = READ_ONCE(sqe->len);
3638	sp->flags = READ_ONCE(sqe->splice_flags);
3639
3640	if (unlikely(sp->flags & ~valid_flags))
3641		return -EINVAL;
3642
3643	sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3644				  (sp->flags & SPLICE_F_FD_IN_FIXED));
3645	if (!sp->file_in)
3646		return -EBADF;
3647	req->flags |= REQ_F_NEED_CLEANUP;
3648	return 0;
3649}
3650
3651static int io_tee_prep(struct io_kiocb *req,
3652		       const struct io_uring_sqe *sqe)
3653{
3654	if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3655		return -EINVAL;
3656	return __io_splice_prep(req, sqe);
3657}
3658
3659static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3660{
3661	struct io_splice *sp = &req->splice;
3662	struct file *in = sp->file_in;
3663	struct file *out = sp->file_out;
3664	unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3665	long ret = 0;
3666
3667	if (issue_flags & IO_URING_F_NONBLOCK)
3668		return -EAGAIN;
3669	if (sp->len)
3670		ret = do_tee(in, out, sp->len, flags);
3671
3672	if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3673		io_put_file(in);
3674	req->flags &= ~REQ_F_NEED_CLEANUP;
3675
3676	if (ret != sp->len)
3677		req_set_fail(req);
3678	io_req_complete(req, ret);
3679	return 0;
3680}
3681
3682static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3683{
3684	struct io_splice *sp = &req->splice;
3685
3686	sp->off_in = READ_ONCE(sqe->splice_off_in);
3687	sp->off_out = READ_ONCE(sqe->off);
3688	return __io_splice_prep(req, sqe);
3689}
3690
3691static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3692{
3693	struct io_splice *sp = &req->splice;
3694	struct file *in = sp->file_in;
3695	struct file *out = sp->file_out;
3696	unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3697	loff_t *poff_in, *poff_out;
3698	long ret = 0;
3699
3700	if (issue_flags & IO_URING_F_NONBLOCK)
3701		return -EAGAIN;
3702
3703	poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3704	poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3705
3706	if (sp->len)
3707		ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3708
3709	if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3710		io_put_file(in);
3711	req->flags &= ~REQ_F_NEED_CLEANUP;
3712
3713	if (ret != sp->len)
3714		req_set_fail(req);
3715	io_req_complete(req, ret);
3716	return 0;
3717}
3718
3719/*
3720 * IORING_OP_NOP just posts a completion event, nothing else.
3721 */
3722static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3723{
3724	struct io_ring_ctx *ctx = req->ctx;
3725
3726	if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3727		return -EINVAL;
3728
3729	__io_req_complete(req, issue_flags, 0, 0);
3730	return 0;
3731}
3732
3733static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3734{
3735	struct io_ring_ctx *ctx = req->ctx;
3736
3737	if (!req->file)
3738		return -EBADF;
3739
3740	if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3741		return -EINVAL;
3742	if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3743		return -EINVAL;
3744
3745	req->sync.flags = READ_ONCE(sqe->fsync_flags);
3746	if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3747		return -EINVAL;
3748
3749	req->sync.off = READ_ONCE(sqe->off);
3750	req->sync.len = READ_ONCE(sqe->len);
3751	return 0;
3752}
3753
3754static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3755{
3756	loff_t end = req->sync.off + req->sync.len;
3757	int ret;
3758
3759	/* fsync always requires a blocking context */
3760	if (issue_flags & IO_URING_F_NONBLOCK)
3761		return -EAGAIN;
3762
3763	ret = vfs_fsync_range(req->file, req->sync.off,
3764				end > 0 ? end : LLONG_MAX,
3765				req->sync.flags & IORING_FSYNC_DATASYNC);
3766	if (ret < 0)
3767		req_set_fail(req);
3768	io_req_complete(req, ret);
3769	return 0;
3770}
3771
3772static int io_fallocate_prep(struct io_kiocb *req,
3773			     const struct io_uring_sqe *sqe)
3774{
3775	if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3776		return -EINVAL;
3777	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3778		return -EINVAL;
3779
3780	req->sync.off = READ_ONCE(sqe->off);
3781	req->sync.len = READ_ONCE(sqe->addr);
3782	req->sync.mode = READ_ONCE(sqe->len);
3783	return 0;
3784}
3785
3786static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3787{
3788	int ret;
3789
3790	/* fallocate always requiring blocking context */
3791	if (issue_flags & IO_URING_F_NONBLOCK)
3792		return -EAGAIN;
3793	ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3794				req->sync.len);
3795	if (ret < 0)
3796		req_set_fail(req);
3797	io_req_complete(req, ret);
3798	return 0;
3799}
3800
3801static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3802{
3803	const char __user *fname;
3804	int ret;
3805
3806	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3807		return -EINVAL;
3808	if (unlikely(sqe->ioprio || sqe->buf_index))
3809		return -EINVAL;
3810	if (unlikely(req->flags & REQ_F_FIXED_FILE))
3811		return -EBADF;
3812
3813	/* open.how should be already initialised */
3814	if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3815		req->open.how.flags |= O_LARGEFILE;
3816
3817	req->open.dfd = READ_ONCE(sqe->fd);
3818	fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3819	req->open.filename = getname(fname);
3820	if (IS_ERR(req->open.filename)) {
3821		ret = PTR_ERR(req->open.filename);
3822		req->open.filename = NULL;
3823		return ret;
3824	}
3825	req->open.nofile = rlimit(RLIMIT_NOFILE);
3826	req->flags |= REQ_F_NEED_CLEANUP;
3827	return 0;
3828}
3829
3830static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3831{
3832	u64 mode = READ_ONCE(sqe->len);
3833	u64 flags = READ_ONCE(sqe->open_flags);
3834
3835	req->open.how = build_open_how(flags, mode);
3836	return __io_openat_prep(req, sqe);
3837}
3838
3839static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3840{
3841	struct open_how __user *how;
3842	size_t len;
3843	int ret;
3844
3845	how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3846	len = READ_ONCE(sqe->len);
3847	if (len < OPEN_HOW_SIZE_VER0)
3848		return -EINVAL;
3849
3850	ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3851					len);
3852	if (ret)
3853		return ret;
3854
3855	return __io_openat_prep(req, sqe);
3856}
3857
3858static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3859{
3860	struct open_flags op;
3861	struct file *file;
3862	bool nonblock_set;
3863	bool resolve_nonblock;
3864	int ret;
3865
3866	ret = build_open_flags(&req->open.how, &op);
3867	if (ret)
3868		goto err;
3869	nonblock_set = op.open_flag & O_NONBLOCK;
3870	resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3871	if (issue_flags & IO_URING_F_NONBLOCK) {
3872		/*
3873		 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3874		 * it'll always -EAGAIN
3875		 */
3876		if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3877			return -EAGAIN;
3878		op.lookup_flags |= LOOKUP_CACHED;
3879		op.open_flag |= O_NONBLOCK;
3880	}
3881
3882	ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3883	if (ret < 0)
3884		goto err;
3885
3886	file = do_filp_open(req->open.dfd, req->open.filename, &op);
3887	if (IS_ERR(file)) {
3888		/*
3889		 * We could hang on to this 'fd' on retrying, but seems like
3890		 * marginal gain for something that is now known to be a slower
3891		 * path. So just put it, and we'll get a new one when we retry.
3892		 */
3893		put_unused_fd(ret);
3894
3895		ret = PTR_ERR(file);
3896		/* only retry if RESOLVE_CACHED wasn't already set by application */
3897		if (ret == -EAGAIN &&
3898		    (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3899			return -EAGAIN;
3900		goto err;
3901	}
3902
3903	if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3904		file->f_flags &= ~O_NONBLOCK;
3905	fsnotify_open(file);
3906	fd_install(ret, file);
3907err:
3908	putname(req->open.filename);
3909	req->flags &= ~REQ_F_NEED_CLEANUP;
3910	if (ret < 0)
3911		req_set_fail(req);
3912	__io_req_complete(req, issue_flags, ret, 0);
3913	return 0;
3914}
3915
3916static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3917{
3918	return io_openat2(req, issue_flags);
3919}
3920
3921static int io_remove_buffers_prep(struct io_kiocb *req,
3922				  const struct io_uring_sqe *sqe)
3923{
3924	struct io_provide_buf *p = &req->pbuf;
3925	u64 tmp;
3926
3927	if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3928		return -EINVAL;
3929
3930	tmp = READ_ONCE(sqe->fd);
3931	if (!tmp || tmp > USHRT_MAX)
3932		return -EINVAL;
3933
3934	memset(p, 0, sizeof(*p));
3935	p->nbufs = tmp;
3936	p->bgid = READ_ONCE(sqe->buf_group);
3937	return 0;
3938}
3939
3940static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3941			       int bgid, unsigned nbufs)
3942{
3943	unsigned i = 0;
3944
3945	/* shouldn't happen */
3946	if (!nbufs)
3947		return 0;
3948
3949	/* the head kbuf is the list itself */
3950	while (!list_empty(&buf->list)) {
3951		struct io_buffer *nxt;
3952
3953		nxt = list_first_entry(&buf->list, struct io_buffer, list);
3954		list_del(&nxt->list);
3955		kfree(nxt);
3956		if (++i == nbufs)
3957			return i;
3958	}
3959	i++;
3960	kfree(buf);
3961	xa_erase(&ctx->io_buffers, bgid);
3962
3963	return i;
3964}
3965
3966static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3967{
3968	struct io_provide_buf *p = &req->pbuf;
3969	struct io_ring_ctx *ctx = req->ctx;
3970	struct io_buffer *head;
3971	int ret = 0;
3972	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3973
3974	io_ring_submit_lock(ctx, !force_nonblock);
3975
3976	lockdep_assert_held(&ctx->uring_lock);
3977
3978	ret = -ENOENT;
3979	head = xa_load(&ctx->io_buffers, p->bgid);
3980	if (head)
3981		ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3982	if (ret < 0)
3983		req_set_fail(req);
3984
3985	/* complete before unlock, IOPOLL may need the lock */
3986	__io_req_complete(req, issue_flags, ret, 0);
3987	io_ring_submit_unlock(ctx, !force_nonblock);
3988	return 0;
3989}
3990
3991static int io_provide_buffers_prep(struct io_kiocb *req,
3992				   const struct io_uring_sqe *sqe)
3993{
3994	unsigned long size, tmp_check;
3995	struct io_provide_buf *p = &req->pbuf;
3996	u64 tmp;
3997
3998	if (sqe->ioprio || sqe->rw_flags)
3999		return -EINVAL;
4000
4001	tmp = READ_ONCE(sqe->fd);
4002	if (!tmp || tmp > USHRT_MAX)
4003		return -E2BIG;
4004	p->nbufs = tmp;
4005	p->addr = READ_ONCE(sqe->addr);
4006	p->len = READ_ONCE(sqe->len);
4007
4008	if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4009				&size))
4010		return -EOVERFLOW;
4011	if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4012		return -EOVERFLOW;
4013
4014	size = (unsigned long)p->len * p->nbufs;
4015	if (!access_ok(u64_to_user_ptr(p->addr), size))
4016		return -EFAULT;
4017
4018	p->bgid = READ_ONCE(sqe->buf_group);
4019	tmp = READ_ONCE(sqe->off);
4020	if (tmp > USHRT_MAX)
4021		return -E2BIG;
4022	p->bid = tmp;
4023	return 0;
4024}
4025
4026static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4027{
4028	struct io_buffer *buf;
4029	u64 addr = pbuf->addr;
4030	int i, bid = pbuf->bid;
4031
4032	for (i = 0; i < pbuf->nbufs; i++) {
4033		buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4034		if (!buf)
4035			break;
4036
4037		buf->addr = addr;
4038		buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4039		buf->bid = bid;
4040		addr += pbuf->len;
4041		bid++;
4042		if (!*head) {
4043			INIT_LIST_HEAD(&buf->list);
4044			*head = buf;
4045		} else {
4046			list_add_tail(&buf->list, &(*head)->list);
4047		}
4048	}
4049
4050	return i ? i : -ENOMEM;
4051}
4052
4053static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4054{
4055	struct io_provide_buf *p = &req->pbuf;
4056	struct io_ring_ctx *ctx = req->ctx;
4057	struct io_buffer *head, *list;
4058	int ret = 0;
4059	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4060
4061	io_ring_submit_lock(ctx, !force_nonblock);
4062
4063	lockdep_assert_held(&ctx->uring_lock);
4064
4065	list = head = xa_load(&ctx->io_buffers, p->bgid);
4066
4067	ret = io_add_buffers(p, &head);
4068	if (ret >= 0 && !list) {
4069		ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4070		if (ret < 0)
4071			__io_remove_buffers(ctx, head, p->bgid, -1U);
4072	}
4073	if (ret < 0)
4074		req_set_fail(req);
4075	/* complete before unlock, IOPOLL may need the lock */
4076	__io_req_complete(req, issue_flags, ret, 0);
4077	io_ring_submit_unlock(ctx, !force_nonblock);
4078	return 0;
4079}
4080
4081static int io_epoll_ctl_prep(struct io_kiocb *req,
4082			     const struct io_uring_sqe *sqe)
4083{
4084#if defined(CONFIG_EPOLL)
4085	if (sqe->ioprio || sqe->buf_index)
4086		return -EINVAL;
4087	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4088		return -EINVAL;
4089
4090	req->epoll.epfd = READ_ONCE(sqe->fd);
4091	req->epoll.op = READ_ONCE(sqe->len);
4092	req->epoll.fd = READ_ONCE(sqe->off);
4093
4094	if (ep_op_has_event(req->epoll.op)) {
4095		struct epoll_event __user *ev;
4096
4097		ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4098		if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4099			return -EFAULT;
4100	}
4101
4102	return 0;
4103#else
4104	return -EOPNOTSUPP;
4105#endif
4106}
4107
4108static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4109{
4110#if defined(CONFIG_EPOLL)
4111	struct io_epoll *ie = &req->epoll;
4112	int ret;
4113	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4114
4115	ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4116	if (force_nonblock && ret == -EAGAIN)
4117		return -EAGAIN;
4118
4119	if (ret < 0)
4120		req_set_fail(req);
4121	__io_req_complete(req, issue_flags, ret, 0);
4122	return 0;
4123#else
4124	return -EOPNOTSUPP;
4125#endif
4126}
4127
4128static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4129{
4130#if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4131	if (sqe->ioprio || sqe->buf_index || sqe->off)
4132		return -EINVAL;
4133	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4134		return -EINVAL;
4135
4136	req->madvise.addr = READ_ONCE(sqe->addr);
4137	req->madvise.len = READ_ONCE(sqe->len);
4138	req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4139	return 0;
4140#else
4141	return -EOPNOTSUPP;
4142#endif
4143}
4144
4145static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4146{
4147#if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4148	struct io_madvise *ma = &req->madvise;
4149	int ret;
4150
4151	if (issue_flags & IO_URING_F_NONBLOCK)
4152		return -EAGAIN;
4153
4154	ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4155	if (ret < 0)
4156		req_set_fail(req);
4157	io_req_complete(req, ret);
4158	return 0;
4159#else
4160	return -EOPNOTSUPP;
4161#endif
4162}
4163
4164static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4165{
4166	if (sqe->ioprio || sqe->buf_index || sqe->addr)
4167		return -EINVAL;
4168	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4169		return -EINVAL;
4170
4171	req->fadvise.offset = READ_ONCE(sqe->off);
4172	req->fadvise.len = READ_ONCE(sqe->len);
4173	req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4174	return 0;
4175}
4176
4177static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4178{
4179	struct io_fadvise *fa = &req->fadvise;
4180	int ret;
4181
4182	if (issue_flags & IO_URING_F_NONBLOCK) {
4183		switch (fa->advice) {
4184		case POSIX_FADV_NORMAL:
4185		case POSIX_FADV_RANDOM:
4186		case POSIX_FADV_SEQUENTIAL:
4187			break;
4188		default:
4189			return -EAGAIN;
4190		}
4191	}
4192
4193	ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4194	if (ret < 0)
4195		req_set_fail(req);
4196	__io_req_complete(req, issue_flags, ret, 0);
4197	return 0;
4198}
4199
4200static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4201{
4202	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4203		return -EINVAL;
4204	if (sqe->ioprio || sqe->buf_index)
4205		return -EINVAL;
4206	if (req->flags & REQ_F_FIXED_FILE)
4207		return -EBADF;
4208
4209	req->statx.dfd = READ_ONCE(sqe->fd);
4210	req->statx.mask = READ_ONCE(sqe->len);
4211	req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4212	req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4213	req->statx.flags = READ_ONCE(sqe->statx_flags);
4214
4215	return 0;
4216}
4217
4218static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4219{
4220	struct io_statx *ctx = &req->statx;
4221	int ret;
4222
4223	if (issue_flags & IO_URING_F_NONBLOCK)
4224		return -EAGAIN;
4225
4226	ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4227		       ctx->buffer);
4228
4229	if (ret < 0)
4230		req_set_fail(req);
4231	io_req_complete(req, ret);
4232	return 0;
4233}
4234
4235static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4236{
4237	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4238		return -EINVAL;
4239	if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4240	    sqe->rw_flags || sqe->buf_index)
4241		return -EINVAL;
4242	if (req->flags & REQ_F_FIXED_FILE)
4243		return -EBADF;
4244
4245	req->close.fd = READ_ONCE(sqe->fd);
4246	return 0;
4247}
4248
4249static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4250{
4251	struct files_struct *files = current->files;
4252	struct io_close *close = &req->close;
4253	struct fdtable *fdt;
4254	struct file *file = NULL;
4255	int ret = -EBADF;
4256
4257	spin_lock(&files->file_lock);
4258	fdt = files_fdtable(files);
4259	if (close->fd >= fdt->max_fds) {
4260		spin_unlock(&files->file_lock);
4261		goto err;
4262	}
4263	file = fdt->fd[close->fd];
4264	if (!file || file->f_op == &io_uring_fops) {
4265		spin_unlock(&files->file_lock);
4266		file = NULL;
4267		goto err;
4268	}
4269
4270	/* if the file has a flush method, be safe and punt to async */
4271	if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4272		spin_unlock(&files->file_lock);
4273		return -EAGAIN;
4274	}
4275
4276	ret = __close_fd_get_file(close->fd, &file);
4277	spin_unlock(&files->file_lock);
4278	if (ret < 0) {
4279		if (ret == -ENOENT)
4280			ret = -EBADF;
4281		goto err;
4282	}
4283
4284	/* No ->flush() or already async, safely close from here */
4285	ret = filp_close(file, current->files);
4286err:
4287	if (ret < 0)
4288		req_set_fail(req);
4289	if (file)
4290		fput(file);
4291	__io_req_complete(req, issue_flags, ret, 0);
4292	return 0;
4293}
4294
4295static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4296{
4297	struct io_ring_ctx *ctx = req->ctx;
4298
4299	if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4300		return -EINVAL;
4301	if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4302		return -EINVAL;
4303
4304	req->sync.off = READ_ONCE(sqe->off);
4305	req->sync.len = READ_ONCE(sqe->len);
4306	req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4307	return 0;
4308}
4309
4310static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4311{
4312	int ret;
4313
4314	/* sync_file_range always requires a blocking context */
4315	if (issue_flags & IO_URING_F_NONBLOCK)
4316		return -EAGAIN;
4317
4318	ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4319				req->sync.flags);
4320	if (ret < 0)
4321		req_set_fail(req);
4322	io_req_complete(req, ret);
4323	return 0;
4324}
4325
4326#if defined(CONFIG_NET)
4327static int io_setup_async_msg(struct io_kiocb *req,
4328			      struct io_async_msghdr *kmsg)
4329{
4330	struct io_async_msghdr *async_msg = req->async_data;
4331
4332	if (async_msg)
4333		return -EAGAIN;
4334	if (io_alloc_async_data(req)) {
4335		kfree(kmsg->free_iov);
4336		return -ENOMEM;
4337	}
4338	async_msg = req->async_data;
4339	req->flags |= REQ_F_NEED_CLEANUP;
4340	memcpy(async_msg, kmsg, sizeof(*kmsg));
4341	async_msg->msg.msg_name = &async_msg->addr;
4342	/* if were using fast_iov, set it to the new one */
4343	if (!async_msg->free_iov)
4344		async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4345
4346	return -EAGAIN;
4347}
4348
4349static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4350			       struct io_async_msghdr *iomsg)
4351{
4352	iomsg->msg.msg_name = &iomsg->addr;
4353	iomsg->free_iov = iomsg->fast_iov;
4354	return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4355				   req->sr_msg.msg_flags, &iomsg->free_iov);
4356}
4357
4358static int io_sendmsg_prep_async(struct io_kiocb *req)
4359{
4360	int ret;
4361
4362	ret = io_sendmsg_copy_hdr(req, req->async_data);
4363	if (!ret)
4364		req->flags |= REQ_F_NEED_CLEANUP;
4365	return ret;
4366}
4367
4368static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4369{
4370	struct io_sr_msg *sr = &req->sr_msg;
4371
4372	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4373		return -EINVAL;
4374
4375	sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4376	sr->len = READ_ONCE(sqe->len);
4377	sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4378	if (sr->msg_flags & MSG_DONTWAIT)
4379		req->flags |= REQ_F_NOWAIT;
4380
4381#ifdef CONFIG_COMPAT
4382	if (req->ctx->compat)
4383		sr->msg_flags |= MSG_CMSG_COMPAT;
4384#endif
4385	return 0;
4386}
4387
4388static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4389{
4390	struct io_async_msghdr iomsg, *kmsg;
4391	struct socket *sock;
4392	unsigned flags;
4393	int min_ret = 0;
4394	int ret;
4395
4396	sock = sock_from_file(req->file);
4397	if (unlikely(!sock))
4398		return -ENOTSOCK;
4399
4400	kmsg = req->async_data;
4401	if (!kmsg) {
4402		ret = io_sendmsg_copy_hdr(req, &iomsg);
4403		if (ret)
4404			return ret;
4405		kmsg = &iomsg;
4406	}
4407
4408	flags = req->sr_msg.msg_flags;
4409	if (issue_flags & IO_URING_F_NONBLOCK)
4410		flags |= MSG_DONTWAIT;
4411	if (flags & MSG_WAITALL)
4412		min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4413
4414	ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4415	if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4416		return io_setup_async_msg(req, kmsg);
4417	if (ret == -ERESTARTSYS)
4418		ret = -EINTR;
4419
4420	/* fast path, check for non-NULL to avoid function call */
4421	if (kmsg->free_iov)
4422		kfree(kmsg->free_iov);
4423	req->flags &= ~REQ_F_NEED_CLEANUP;
4424	if (ret < min_ret)
4425		req_set_fail(req);
4426	__io_req_complete(req, issue_flags, ret, 0);
4427	return 0;
4428}
4429
4430static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4431{
4432	struct io_sr_msg *sr = &req->sr_msg;
4433	struct msghdr msg;
4434	struct iovec iov;
4435	struct socket *sock;
4436	unsigned flags;
4437	int min_ret = 0;
4438	int ret;
4439
4440	sock = sock_from_file(req->file);
4441	if (unlikely(!sock))
4442		return -ENOTSOCK;
4443
4444	ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4445	if (unlikely(ret))
4446		return ret;
4447
4448	msg.msg_name = NULL;
4449	msg.msg_control = NULL;
4450	msg.msg_controllen = 0;
4451	msg.msg_namelen = 0;
4452
4453	flags = req->sr_msg.msg_flags;
4454	if (issue_flags & IO_URING_F_NONBLOCK)
4455		flags |= MSG_DONTWAIT;
4456	if (flags & MSG_WAITALL)
4457		min_ret = iov_iter_count(&msg.msg_iter);
4458
4459	msg.msg_flags = flags;
4460	ret = sock_sendmsg(sock, &msg);
4461	if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4462		return -EAGAIN;
4463	if (ret == -ERESTARTSYS)
4464		ret = -EINTR;
4465
4466	if (ret < min_ret)
4467		req_set_fail(req);
4468	__io_req_complete(req, issue_flags, ret, 0);
4469	return 0;
4470}
4471
4472static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4473				 struct io_async_msghdr *iomsg)
4474{
4475	struct io_sr_msg *sr = &req->sr_msg;
4476	struct iovec __user *uiov;
4477	size_t iov_len;
4478	int ret;
4479
4480	ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4481					&iomsg->uaddr, &uiov, &iov_len);
4482	if (ret)
4483		return ret;
4484
4485	if (req->flags & REQ_F_BUFFER_SELECT) {
4486		if (iov_len > 1)
4487			return -EINVAL;
4488		if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4489			return -EFAULT;
4490		sr->len = iomsg->fast_iov[0].iov_len;
4491		iomsg->free_iov = NULL;
4492	} else {
4493		iomsg->free_iov = iomsg->fast_iov;
4494		ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4495				     &iomsg->free_iov, &iomsg->msg.msg_iter,
4496				     false);
4497		if (ret > 0)
4498			ret = 0;
4499	}
4500
4501	return ret;
4502}
4503
4504#ifdef CONFIG_COMPAT
4505static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4506					struct io_async_msghdr *iomsg)
4507{
4508	struct io_sr_msg *sr = &req->sr_msg;
4509	struct compat_iovec __user *uiov;
4510	compat_uptr_t ptr;
4511	compat_size_t len;
4512	int ret;
4513
4514	ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4515				  &ptr, &len);
4516	if (ret)
4517		return ret;
4518
4519	uiov = compat_ptr(ptr);
4520	if (req->flags & REQ_F_BUFFER_SELECT) {
4521		compat_ssize_t clen;
4522
4523		if (len > 1)
4524			return -EINVAL;
4525		if (!access_ok(uiov, sizeof(*uiov)))
4526			return -EFAULT;
4527		if (__get_user(clen, &uiov->iov_len))
4528			return -EFAULT;
4529		if (clen < 0)
4530			return -EINVAL;
4531		sr->len = clen;
4532		iomsg->free_iov = NULL;
4533	} else {
4534		iomsg->free_iov = iomsg->fast_iov;
4535		ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4536				   UIO_FASTIOV, &iomsg->free_iov,
4537				   &iomsg->msg.msg_iter, true);
4538		if (ret < 0)
4539			return ret;
4540	}
4541
4542	return 0;
4543}
4544#endif
4545
4546static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4547			       struct io_async_msghdr *iomsg)
4548{
4549	iomsg->msg.msg_name = &iomsg->addr;
4550
4551#ifdef CONFIG_COMPAT
4552	if (req->ctx->compat)
4553		return __io_compat_recvmsg_copy_hdr(req, iomsg);
4554#endif
4555
4556	return __io_recvmsg_copy_hdr(req, iomsg);
4557}
4558
4559static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4560					       bool needs_lock)
4561{
4562	struct io_sr_msg *sr = &req->sr_msg;
4563	struct io_buffer *kbuf;
4564
4565	kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4566	if (IS_ERR(kbuf))
4567		return kbuf;
4568
4569	sr->kbuf = kbuf;
4570	req->flags |= REQ_F_BUFFER_SELECTED;
4571	return kbuf;
4572}
4573
4574static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4575{
4576	return io_put_kbuf(req, req->sr_msg.kbuf);
4577}
4578
4579static int io_recvmsg_prep_async(struct io_kiocb *req)
4580{
4581	int ret;
4582
4583	ret = io_recvmsg_copy_hdr(req, req->async_data);
4584	if (!ret)
4585		req->flags |= REQ_F_NEED_CLEANUP;
4586	return ret;
4587}
4588
4589static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4590{
4591	struct io_sr_msg *sr = &req->sr_msg;
4592
4593	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4594		return -EINVAL;
4595
4596	sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4597	sr->len = READ_ONCE(sqe->len);
4598	sr->bgid = READ_ONCE(sqe->buf_group);
4599	sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4600	if (sr->msg_flags & MSG_DONTWAIT)
4601		req->flags |= REQ_F_NOWAIT;
4602
4603#ifdef CONFIG_COMPAT
4604	if (req->ctx->compat)
4605		sr->msg_flags |= MSG_CMSG_COMPAT;
4606#endif
4607	return 0;
4608}
4609
4610static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4611{
4612	struct io_async_msghdr iomsg, *kmsg;
4613	struct socket *sock;
4614	struct io_buffer *kbuf;
4615	unsigned flags;
4616	int min_ret = 0;
4617	int ret, cflags = 0;
4618	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4619
4620	sock = sock_from_file(req->file);
4621	if (unlikely(!sock))
4622		return -ENOTSOCK;
4623
4624	kmsg = req->async_data;
4625	if (!kmsg) {
4626		ret = io_recvmsg_copy_hdr(req, &iomsg);
4627		if (ret)
4628			return ret;
4629		kmsg = &iomsg;
4630	}
4631
4632	if (req->flags & REQ_F_BUFFER_SELECT) {
4633		kbuf = io_recv_buffer_select(req, !force_nonblock);
4634		if (IS_ERR(kbuf))
4635			return PTR_ERR(kbuf);
4636		kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4637		kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4638		iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4639				1, req->sr_msg.len);
4640	}
4641
4642	flags = req->sr_msg.msg_flags;
4643	if (force_nonblock)
4644		flags |= MSG_DONTWAIT;
4645	if (flags & MSG_WAITALL)
4646		min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4647
4648	ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4649					kmsg->uaddr, flags);
4650	if (force_nonblock && ret == -EAGAIN)
4651		return io_setup_async_msg(req, kmsg);
4652	if (ret == -ERESTARTSYS)
4653		ret = -EINTR;
4654
4655	if (req->flags & REQ_F_BUFFER_SELECTED)
4656		cflags = io_put_recv_kbuf(req);
4657	/* fast path, check for non-NULL to avoid function call */
4658	if (kmsg->free_iov)
4659		kfree(kmsg->free_iov);
4660	req->flags &= ~REQ_F_NEED_CLEANUP;
4661	if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4662		req_set_fail(req);
4663	__io_req_complete(req, issue_flags, ret, cflags);
4664	return 0;
4665}
4666
4667static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4668{
4669	struct io_buffer *kbuf;
4670	struct io_sr_msg *sr = &req->sr_msg;
4671	struct msghdr msg;
4672	void __user *buf = sr->buf;
4673	struct socket *sock;
4674	struct iovec iov;
4675	unsigned flags;
4676	int min_ret = 0;
4677	int ret, cflags = 0;
4678	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4679
4680	sock = sock_from_file(req->file);
4681	if (unlikely(!sock))
4682		return -ENOTSOCK;
4683
4684	if (req->flags & REQ_F_BUFFER_SELECT) {
4685		kbuf = io_recv_buffer_select(req, !force_nonblock);
4686		if (IS_ERR(kbuf))
4687			return PTR_ERR(kbuf);
4688		buf = u64_to_user_ptr(kbuf->addr);
4689	}
4690
4691	ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4692	if (unlikely(ret))
4693		goto out_free;
4694
4695	msg.msg_name = NULL;
4696	msg.msg_control = NULL;
4697	msg.msg_controllen = 0;
4698	msg.msg_namelen = 0;
4699	msg.msg_iocb = NULL;
4700	msg.msg_flags = 0;
4701
4702	flags = req->sr_msg.msg_flags;
4703	if (force_nonblock)
4704		flags |= MSG_DONTWAIT;
4705	if (flags & MSG_WAITALL)
4706		min_ret = iov_iter_count(&msg.msg_iter);
4707
4708	ret = sock_recvmsg(sock, &msg, flags);
4709	if (force_nonblock && ret == -EAGAIN)
4710		return -EAGAIN;
4711	if (ret == -ERESTARTSYS)
4712		ret = -EINTR;
4713out_free:
4714	if (req->flags & REQ_F_BUFFER_SELECTED)
4715		cflags = io_put_recv_kbuf(req);
4716	if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4717		req_set_fail(req);
4718	__io_req_complete(req, issue_flags, ret, cflags);
4719	return 0;
4720}
4721
4722static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4723{
4724	struct io_accept *accept = &req->accept;
4725
4726	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4727		return -EINVAL;
4728	if (sqe->ioprio || sqe->len || sqe->buf_index)
4729		return -EINVAL;
4730
4731	accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4732	accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4733	accept->flags = READ_ONCE(sqe->accept_flags);
4734	accept->nofile = rlimit(RLIMIT_NOFILE);
4735	return 0;
4736}
4737
4738static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4739{
4740	struct io_accept *accept = &req->accept;
4741	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4742	unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4743	int ret;
4744
4745	if (req->file->f_flags & O_NONBLOCK)
4746		req->flags |= REQ_F_NOWAIT;
4747
4748	ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4749					accept->addr_len, accept->flags,
4750					accept->nofile);
4751	if (ret == -EAGAIN && force_nonblock)
4752		return -EAGAIN;
4753	if (ret < 0) {
4754		if (ret == -ERESTARTSYS)
4755			ret = -EINTR;
4756		req_set_fail(req);
4757	}
4758	__io_req_complete(req, issue_flags, ret, 0);
4759	return 0;
4760}
4761
4762static int io_connect_prep_async(struct io_kiocb *req)
4763{
4764	struct io_async_connect *io = req->async_data;
4765	struct io_connect *conn = &req->connect;
4766
4767	return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4768}
4769
4770static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4771{
4772	struct io_connect *conn = &req->connect;
4773
4774	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4775		return -EINVAL;
4776	if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4777		return -EINVAL;
4778
4779	conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4780	conn->addr_len =  READ_ONCE(sqe->addr2);
4781	return 0;
4782}
4783
4784static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4785{
4786	struct io_async_connect __io, *io;
4787	unsigned file_flags;
4788	int ret;
4789	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4790
4791	if (req->async_data) {
4792		io = req->async_data;
4793	} else {
4794		ret = move_addr_to_kernel(req->connect.addr,
4795						req->connect.addr_len,
4796						&__io.address);
4797		if (ret)
4798			goto out;
4799		io = &__io;
4800	}
4801
4802	file_flags = force_nonblock ? O_NONBLOCK : 0;
4803
4804	ret = __sys_connect_file(req->file, &io->address,
4805					req->connect.addr_len, file_flags);
4806	if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4807		if (req->async_data)
4808			return -EAGAIN;
4809		if (io_alloc_async_data(req)) {
4810			ret = -ENOMEM;
4811			goto out;
4812		}
4813		memcpy(req->async_data, &__io, sizeof(__io));
4814		return -EAGAIN;
4815	}
4816	if (ret == -ERESTARTSYS)
4817		ret = -EINTR;
4818out:
4819	if (ret < 0)
4820		req_set_fail(req);
4821	__io_req_complete(req, issue_flags, ret, 0);
4822	return 0;
4823}
4824#else /* !CONFIG_NET */
4825#define IO_NETOP_FN(op)							\
4826static int io_##op(struct io_kiocb *req, unsigned int issue_flags)	\
4827{									\
4828	return -EOPNOTSUPP;						\
4829}
4830
4831#define IO_NETOP_PREP(op)						\
4832IO_NETOP_FN(op)								\
4833static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4834{									\
4835	return -EOPNOTSUPP;						\
4836}									\
4837
4838#define IO_NETOP_PREP_ASYNC(op)						\
4839IO_NETOP_PREP(op)							\
4840static int io_##op##_prep_async(struct io_kiocb *req)			\
4841{									\
4842	return -EOPNOTSUPP;						\
4843}
4844
4845IO_NETOP_PREP_ASYNC(sendmsg);
4846IO_NETOP_PREP_ASYNC(recvmsg);
4847IO_NETOP_PREP_ASYNC(connect);
4848IO_NETOP_PREP(accept);
4849IO_NETOP_FN(send);
4850IO_NETOP_FN(recv);
4851#endif /* CONFIG_NET */
4852
4853struct io_poll_table {
4854	struct poll_table_struct pt;
4855	struct io_kiocb *req;
4856	int nr_entries;
4857	int error;
4858};
4859
4860static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4861			   __poll_t mask, io_req_tw_func_t func)
4862{
4863	/* for instances that support it check for an event match first: */
4864	if (mask && !(mask & poll->events))
4865		return 0;
4866
4867	trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4868
4869	list_del_init(&poll->wait.entry);
4870
4871	req->result = mask;
4872	req->io_task_work.func = func;
4873
4874	/*
4875	 * If this fails, then the task is exiting. When a task exits, the
4876	 * work gets canceled, so just cancel this request as well instead
4877	 * of executing it. We can't safely execute it anyway, as we may not
4878	 * have the needed state needed for it anyway.
4879	 */
4880	io_req_task_work_add(req);
4881	return 1;
4882}
4883
4884static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4885	__acquires(&req->ctx->completion_lock)
4886{
4887	struct io_ring_ctx *ctx = req->ctx;
4888
4889	if (unlikely(req->task->flags & PF_EXITING))
4890		WRITE_ONCE(poll->canceled, true);
4891
4892	if (!req->result && !READ_ONCE(poll->canceled)) {
4893		struct poll_table_struct pt = { ._key = poll->events };
4894
4895		req->result = vfs_poll(req->file, &pt) & poll->events;
4896	}
4897
4898	spin_lock(&ctx->completion_lock);
4899	if (!req->result && !READ_ONCE(poll->canceled)) {
4900		add_wait_queue(poll->head, &poll->wait);
4901		return true;
4902	}
4903
4904	return false;
4905}
4906
4907static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4908{
4909	/* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4910	if (req->opcode == IORING_OP_POLL_ADD)
4911		return req->async_data;
4912	return req->apoll->double_poll;
4913}
4914
4915static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4916{
4917	if (req->opcode == IORING_OP_POLL_ADD)
4918		return &req->poll;
4919	return &req->apoll->poll;
4920}
4921
4922static void io_poll_remove_double(struct io_kiocb *req)
4923	__must_hold(&req->ctx->completion_lock)
4924{
4925	struct io_poll_iocb *poll = io_poll_get_double(req);
4926
4927	lockdep_assert_held(&req->ctx->completion_lock);
4928
4929	if (poll && poll->head) {
4930		struct wait_queue_head *head = poll->head;
4931
4932		spin_lock_irq(&head->lock);
4933		list_del_init(&poll->wait.entry);
4934		if (poll->wait.private)
4935			req_ref_put(req);
4936		poll->head = NULL;
4937		spin_unlock_irq(&head->lock);
4938	}
4939}
4940
4941static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4942	__must_hold(&req->ctx->completion_lock)
4943{
4944	struct io_ring_ctx *ctx = req->ctx;
4945	unsigned flags = IORING_CQE_F_MORE;
4946	int error;
4947
4948	if (READ_ONCE(req->poll.canceled)) {
4949		error = -ECANCELED;
4950		req->poll.events |= EPOLLONESHOT;
4951	} else {
4952		error = mangle_poll(mask);
4953	}
4954	if (req->poll.events & EPOLLONESHOT)
4955		flags = 0;
4956	if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4957		req->poll.done = true;
4958		flags = 0;
4959	}
4960	if (flags & IORING_CQE_F_MORE)
4961		ctx->cq_extra++;
4962
4963	io_commit_cqring(ctx);
4964	return !(flags & IORING_CQE_F_MORE);
4965}
4966
4967static void io_poll_task_func(struct io_kiocb *req)
4968{
4969	struct io_ring_ctx *ctx = req->ctx;
4970	struct io_kiocb *nxt;
4971
4972	if (io_poll_rewait(req, &req->poll)) {
4973		spin_unlock(&ctx->completion_lock);
4974	} else {
4975		bool done;
4976
4977		done = io_poll_complete(req, req->result);
4978		if (done) {
4979			io_poll_remove_double(req);
4980			hash_del(&req->hash_node);
4981		} else {
4982			req->result = 0;
4983			add_wait_queue(req->poll.head, &req->poll.wait);
4984		}
4985		spin_unlock(&ctx->completion_lock);
4986		io_cqring_ev_posted(ctx);
4987
4988		if (done) {
4989			nxt = io_put_req_find_next(req);
4990			if (nxt)
4991				io_req_task_submit(nxt);
4992		}
4993	}
4994}
4995
4996static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4997			       int sync, void *key)
4998{
4999	struct io_kiocb *req = wait->private;
5000	struct io_poll_iocb *poll = io_poll_get_single(req);
5001	__poll_t mask = key_to_poll(key);
5002	unsigned long flags;
5003
5004	/* for instances that support it check for an event match first: */
5005	if (mask && !(mask & poll->events))
5006		return 0;
5007	if (!(poll->events & EPOLLONESHOT))
5008		return poll->wait.func(&poll->wait, mode, sync, key);
5009
5010	list_del_init(&wait->entry);
5011
5012	if (poll->head) {
5013		bool done;
5014
5015		spin_lock_irqsave(&poll->head->lock, flags);
5016		done = list_empty(&poll->wait.entry);
5017		if (!done)
5018			list_del_init(&poll->wait.entry);
5019		/* make sure double remove sees this as being gone */
5020		wait->private = NULL;
5021		spin_unlock_irqrestore(&poll->head->lock, flags);
5022		if (!done) {
5023			/* use wait func handler, so it matches the rq type */
5024			poll->wait.func(&poll->wait, mode, sync, key);
5025		}
5026	}
5027	req_ref_put(req);
5028	return 1;
5029}
5030
5031static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5032			      wait_queue_func_t wake_func)
5033{
5034	poll->head = NULL;
5035	poll->done = false;
5036	poll->canceled = false;
5037#define IO_POLL_UNMASK	(EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5038	/* mask in events that we always want/need */
5039	poll->events = events | IO_POLL_UNMASK;
5040	INIT_LIST_HEAD(&poll->wait.entry);
5041	init_waitqueue_func_entry(&poll->wait, wake_func);
5042}
5043
5044static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5045			    struct wait_queue_head *head,
5046			    struct io_poll_iocb **poll_ptr)
5047{
5048	struct io_kiocb *req = pt->req;
5049
5050	/*
5051	 * The file being polled uses multiple waitqueues for poll handling
5052	 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5053	 * if this happens.
5054	 */
5055	if (unlikely(pt->nr_entries)) {
5056		struct io_poll_iocb *poll_one = poll;
5057
5058		/* already have a 2nd entry, fail a third attempt */
5059		if (*poll_ptr) {
5060			pt->error = -EINVAL;
5061			return;
5062		}
5063		/*
5064		 * Can't handle multishot for double wait for now, turn it
5065		 * into one-shot mode.
5066		 */
5067		if (!(poll_one->events & EPOLLONESHOT))
5068			poll_one->events |= EPOLLONESHOT;
5069		/* double add on the same waitqueue head, ignore */
5070		if (poll_one->head == head)
5071			return;
5072		poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5073		if (!poll) {
5074			pt->error = -ENOMEM;
5075			return;
5076		}
5077		io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5078		req_ref_get(req);
5079		poll->wait.private = req;
5080		*poll_ptr = poll;
5081	}
5082
5083	pt->nr_entries++;
5084	poll->head = head;
5085
5086	if (poll->events & EPOLLEXCLUSIVE)
5087		add_wait_queue_exclusive(head, &poll->wait);
5088	else
5089		add_wait_queue(head, &poll->wait);
5090}
5091
5092static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5093			       struct poll_table_struct *p)
5094{
5095	struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5096	struct async_poll *apoll = pt->req->apoll;
5097
5098	__io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5099}
5100
5101static void io_async_task_func(struct io_kiocb *req)
5102{
5103	struct async_poll *apoll = req->apoll;
5104	struct io_ring_ctx *ctx = req->ctx;
5105
5106	trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5107
5108	if (io_poll_rewait(req, &apoll->poll)) {
5109		spin_unlock(&ctx->completion_lock);
5110		return;
5111	}
5112
5113	hash_del(&req->hash_node);
5114	io_poll_remove_double(req);
5115	spin_unlock(&ctx->completion_lock);
5116
5117	if (!READ_ONCE(apoll->poll.canceled))
5118		io_req_task_submit(req);
5119	else
5120		io_req_complete_failed(req, -ECANCELED);
5121}
5122
5123static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5124			void *key)
5125{
5126	struct io_kiocb *req = wait->private;
5127	struct io_poll_iocb *poll = &req->apoll->poll;
5128
5129	trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5130					key_to_poll(key));
5131
5132	return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5133}
5134
5135static void io_poll_req_insert(struct io_kiocb *req)
5136{
5137	struct io_ring_ctx *ctx = req->ctx;
5138	struct hlist_head *list;
5139
5140	list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5141	hlist_add_head(&req->hash_node, list);
5142}
5143
5144static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5145				      struct io_poll_iocb *poll,
5146				      struct io_poll_table *ipt, __poll_t mask,
5147				      wait_queue_func_t wake_func)
5148	__acquires(&ctx->completion_lock)
5149{
5150	struct io_ring_ctx *ctx = req->ctx;
5151	bool cancel = false;
5152
5153	INIT_HLIST_NODE(&req->hash_node);
5154	io_init_poll_iocb(poll, mask, wake_func);
5155	poll->file = req->file;
5156	poll->wait.private = req;
5157
5158	ipt->pt._key = mask;
5159	ipt->req = req;
5160	ipt->error = 0;
5161	ipt->nr_entries = 0;
5162
5163	mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5164	if (unlikely(!ipt->nr_entries) && !ipt->error)
5165		ipt->error = -EINVAL;
5166
5167	spin_lock(&ctx->completion_lock);
5168	if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5169		io_poll_remove_double(req);
5170	if (likely(poll->head)) {
5171		spin_lock_irq(&poll->head->lock);
5172		if (unlikely(list_empty(&poll->wait.entry))) {
5173			if (ipt->error)
5174				cancel = true;
5175			ipt->error = 0;
5176			mask = 0;
5177		}
5178		if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5179			list_del_init(&poll->wait.entry);
5180		else if (cancel)
5181			WRITE_ONCE(poll->canceled, true);
5182		else if (!poll->done) /* actually waiting for an event */
5183			io_poll_req_insert(req);
5184		spin_unlock_irq(&poll->head->lock);
5185	}
5186
5187	return mask;
5188}
5189
5190enum {
5191	IO_APOLL_OK,
5192	IO_APOLL_ABORTED,
5193	IO_APOLL_READY
5194};
5195
5196static int io_arm_poll_handler(struct io_kiocb *req)
5197{
5198	const struct io_op_def *def = &io_op_defs[req->opcode];
5199	struct io_ring_ctx *ctx = req->ctx;
5200	struct async_poll *apoll;
5201	struct io_poll_table ipt;
5202	__poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5203	int rw;
5204
5205	if (!req->file || !file_can_poll(req->file))
5206		return IO_APOLL_ABORTED;
5207	if (req->flags & REQ_F_POLLED)
5208		return IO_APOLL_ABORTED;
5209	if (!def->pollin && !def->pollout)
5210		return IO_APOLL_ABORTED;
5211
5212	if (def->pollin) {
5213		rw = READ;
5214		mask |= POLLIN | POLLRDNORM;
5215
5216		/* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5217		if ((req->opcode == IORING_OP_RECVMSG) &&
5218		    (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5219			mask &= ~POLLIN;
5220	} else {
5221		rw = WRITE;
5222		mask |= POLLOUT | POLLWRNORM;
5223	}
5224
5225	/* if we can't nonblock try, then no point in arming a poll handler */
5226	if (!io_file_supports_nowait(req, rw))
5227		return IO_APOLL_ABORTED;
5228
5229	apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5230	if (unlikely(!apoll))
5231		return IO_APOLL_ABORTED;
5232	apoll->double_poll = NULL;
5233	req->apoll = apoll;
5234	req->flags |= REQ_F_POLLED;
5235	ipt.pt._qproc = io_async_queue_proc;
5236	io_req_refcount(req);
5237
5238	ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5239					io_async_wake);
5240	spin_unlock(&ctx->completion_lock);
5241	if (ret || ipt.error)
5242		return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5243
5244	trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5245				mask, apoll->poll.events);
5246	return IO_APOLL_OK;
5247}
5248
5249static bool __io_poll_remove_one(struct io_kiocb *req,
5250				 struct io_poll_iocb *poll, bool do_cancel)
5251	__must_hold(&req->ctx->completion_lock)
5252{
5253	bool do_complete = false;
5254
5255	if (!poll->head)
5256		return false;
5257	spin_lock_irq(&poll->head->lock);
5258	if (do_cancel)
5259		WRITE_ONCE(poll->canceled, true);
5260	if (!list_empty(&poll->wait.entry)) {
5261		list_del_init(&poll->wait.entry);
5262		do_complete = true;
5263	}
5264	spin_unlock_irq(&poll->head->lock);
5265	hash_del(&req->hash_node);
5266	return do_complete;
5267}
5268
5269static bool io_poll_remove_one(struct io_kiocb *req)
5270	__must_hold(&req->ctx->completion_lock)
5271{
5272	bool do_complete;
5273
5274	io_poll_remove_double(req);
5275	do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5276
5277	if (do_complete) {
5278		io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5279		io_commit_cqring(req->ctx);
5280		req_set_fail(req);
5281		io_put_req_deferred(req);
5282	}
5283	return do_complete;
5284}
5285
5286/*
5287 * Returns true if we found and killed one or more poll requests
5288 */
5289static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5290			       bool cancel_all)
5291{
5292	struct hlist_node *tmp;
5293	struct io_kiocb *req;
5294	int posted = 0, i;
5295
5296	spin_lock(&ctx->completion_lock);
5297	for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5298		struct hlist_head *list;
5299
5300		list = &ctx->cancel_hash[i];
5301		hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5302			if (io_match_task(req, tsk, cancel_all))
5303				posted += io_poll_remove_one(req);
5304		}
5305	}
5306	spin_unlock(&ctx->completion_lock);
5307
5308	if (posted)
5309		io_cqring_ev_posted(ctx);
5310
5311	return posted != 0;
5312}
5313
5314static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5315				     bool poll_only)
5316	__must_hold(&ctx->completion_lock)
5317{
5318	struct hlist_head *list;
5319	struct io_kiocb *req;
5320
5321	list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5322	hlist_for_each_entry(req, list, hash_node) {
5323		if (sqe_addr != req->user_data)
5324			continue;
5325		if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5326			continue;
5327		return req;
5328	}
5329	return NULL;
5330}
5331
5332static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5333			  bool poll_only)
5334	__must_hold(&ctx->completion_lock)
5335{
5336	struct io_kiocb *req;
5337
5338	req = io_poll_find(ctx, sqe_addr, poll_only);
5339	if (!req)
5340		return -ENOENT;
5341	if (io_poll_remove_one(req))
5342		return 0;
5343
5344	return -EALREADY;
5345}
5346
5347static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5348				     unsigned int flags)
5349{
5350	u32 events;
5351
5352	events = READ_ONCE(sqe->poll32_events);
5353#ifdef __BIG_ENDIAN
5354	events = swahw32(events);
5355#endif
5356	if (!(flags & IORING_POLL_ADD_MULTI))
5357		events |= EPOLLONESHOT;
5358	return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5359}
5360
5361static int io_poll_update_prep(struct io_kiocb *req,
5362			       const struct io_uring_sqe *sqe)
5363{
5364	struct io_poll_update *upd = &req->poll_update;
5365	u32 flags;
5366
5367	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5368		return -EINVAL;
5369	if (sqe->ioprio || sqe->buf_index)
5370		return -EINVAL;
5371	flags = READ_ONCE(sqe->len);
5372	if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5373		      IORING_POLL_ADD_MULTI))
5374		return -EINVAL;
5375	/* meaningless without update */
5376	if (flags == IORING_POLL_ADD_MULTI)
5377		return -EINVAL;
5378
5379	upd->old_user_data = READ_ONCE(sqe->addr);
5380	upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5381	upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5382
5383	upd->new_user_data = READ_ONCE(sqe->off);
5384	if (!upd->update_user_data && upd->new_user_data)
5385		return -EINVAL;
5386	if (upd->update_events)
5387		upd->events = io_poll_parse_events(sqe, flags);
5388	else if (sqe->poll32_events)
5389		return -EINVAL;
5390
5391	return 0;
5392}
5393
5394static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5395			void *key)
5396{
5397	struct io_kiocb *req = wait->private;
5398	struct io_poll_iocb *poll = &req->poll;
5399
5400	return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5401}
5402
5403static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5404			       struct poll_table_struct *p)
5405{
5406	struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5407
5408	__io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5409}
5410
5411static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5412{
5413	struct io_poll_iocb *poll = &req->poll;
5414	u32 flags;
5415
5416	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5417		return -EINVAL;
5418	if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5419		return -EINVAL;
5420	flags = READ_ONCE(sqe->len);
5421	if (flags & ~IORING_POLL_ADD_MULTI)
5422		return -EINVAL;
5423
5424	io_req_refcount(req);
5425	poll->events = io_poll_parse_events(sqe, flags);
5426	return 0;
5427}
5428
5429static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5430{
5431	struct io_poll_iocb *poll = &req->poll;
5432	struct io_ring_ctx *ctx = req->ctx;
5433	struct io_poll_table ipt;
5434	__poll_t mask;
5435
5436	ipt.pt._qproc = io_poll_queue_proc;
5437
5438	mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5439					io_poll_wake);
5440
5441	if (mask) { /* no async, we'd stolen it */
5442		ipt.error = 0;
5443		io_poll_complete(req, mask);
5444	}
5445	spin_unlock(&ctx->completion_lock);
5446
5447	if (mask) {
5448		io_cqring_ev_posted(ctx);
5449		if (poll->events & EPOLLONESHOT)
5450			io_put_req(req);
5451	}
5452	return ipt.error;
5453}
5454
5455static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5456{
5457	struct io_ring_ctx *ctx = req->ctx;
5458	struct io_kiocb *preq;
5459	bool completing;
5460	int ret;
5461
5462	spin_lock(&ctx->completion_lock);
5463	preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5464	if (!preq) {
5465		ret = -ENOENT;
5466		goto err;
5467	}
5468
5469	if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5470		completing = true;
5471		ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5472		goto err;
5473	}
5474
5475	/*
5476	 * Don't allow racy completion with singleshot, as we cannot safely
5477	 * update those. For multishot, if we're racing with completion, just
5478	 * let completion re-add it.
5479	 */
5480	completing = !__io_poll_remove_one(preq, &preq->poll, false);
5481	if (completing && (preq->poll.events & EPOLLONESHOT)) {
5482		ret = -EALREADY;
5483		goto err;
5484	}
5485	/* we now have a detached poll request. reissue. */
5486	ret = 0;
5487err:
5488	if (ret < 0) {
5489		spin_unlock(&ctx->completion_lock);
5490		req_set_fail(req);
5491		io_req_complete(req, ret);
5492		return 0;
5493	}
5494	/* only mask one event flags, keep behavior flags */
5495	if (req->poll_update.update_events) {
5496		preq->poll.events &= ~0xffff;
5497		preq->poll.events |= req->poll_update.events & 0xffff;
5498		preq->poll.events |= IO_POLL_UNMASK;
5499	}
5500	if (req->poll_update.update_user_data)
5501		preq->user_data = req->poll_update.new_user_data;
5502	spin_unlock(&ctx->completion_lock);
5503
5504	/* complete update request, we're done with it */
5505	io_req_complete(req, ret);
5506
5507	if (!completing) {
5508		ret = io_poll_add(preq, issue_flags);
5509		if (ret < 0) {
5510			req_set_fail(preq);
5511			io_req_complete(preq, ret);
5512		}
5513	}
5514	return 0;
5515}
5516
5517static void io_req_task_timeout(struct io_kiocb *req)
5518{
5519	struct io_ring_ctx *ctx = req->ctx;
5520
5521	spin_lock(&ctx->completion_lock);
5522	io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5523	io_commit_cqring(ctx);
5524	spin_unlock(&ctx->completion_lock);
5525
5526	io_cqring_ev_posted(ctx);
5527	req_set_fail(req);
5528	io_put_req(req);
5529}
5530
5531static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5532{
5533	struct io_timeout_data *data = container_of(timer,
5534						struct io_timeout_data, timer);
5535	struct io_kiocb *req = data->req;
5536	struct io_ring_ctx *ctx = req->ctx;
5537	unsigned long flags;
5538
5539	spin_lock_irqsave(&ctx->timeout_lock, flags);
5540	list_del_init(&req->timeout.list);
5541	atomic_set(&req->ctx->cq_timeouts,
5542		atomic_read(&req->ctx->cq_timeouts) + 1);
5543	spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5544
5545	req->io_task_work.func = io_req_task_timeout;
5546	io_req_task_work_add(req);
5547	return HRTIMER_NORESTART;
5548}
5549
5550static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5551					   __u64 user_data)
5552	__must_hold(&ctx->timeout_lock)
5553{
5554	struct io_timeout_data *io;
5555	struct io_kiocb *req;
5556	bool found = false;
5557
5558	list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5559		found = user_data == req->user_data;
5560		if (found)
5561			break;
5562	}
5563	if (!found)
5564		return ERR_PTR(-ENOENT);
5565
5566	io = req->async_data;
5567	if (hrtimer_try_to_cancel(&io->timer) == -1)
5568		return ERR_PTR(-EALREADY);
5569	list_del_init(&req->timeout.list);
5570	return req;
5571}
5572
5573static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5574	__must_hold(&ctx->timeout_lock)
5575{
5576	struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5577
5578	if (IS_ERR(req))
5579		return PTR_ERR(req);
5580
5581	req_set_fail(req);
5582	io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5583	io_put_req_deferred(req);
5584	return 0;
5585}
5586
5587static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5588			     struct timespec64 *ts, enum hrtimer_mode mode)
5589	__must_hold(&ctx->timeout_lock)
5590{
5591	struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5592	struct io_timeout_data *data;
5593
5594	if (IS_ERR(req))
5595		return PTR_ERR(req);
5596
5597	req->timeout.off = 0; /* noseq */
5598	data = req->async_data;
5599	list_add_tail(&req->timeout.list, &ctx->timeout_list);
5600	hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5601	data->timer.function = io_timeout_fn;
5602	hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5603	return 0;
5604}
5605
5606static int io_timeout_remove_prep(struct io_kiocb *req,
5607				  const struct io_uring_sqe *sqe)
5608{
5609	struct io_timeout_rem *tr = &req->timeout_rem;
5610
5611	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5612		return -EINVAL;
5613	if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5614		return -EINVAL;
5615	if (sqe->ioprio || sqe->buf_index || sqe->len)
5616		return -EINVAL;
5617
5618	tr->addr = READ_ONCE(sqe->addr);
5619	tr->flags = READ_ONCE(sqe->timeout_flags);
5620	if (tr->flags & IORING_TIMEOUT_UPDATE) {
5621		if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5622			return -EINVAL;
5623		if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5624			return -EFAULT;
5625	} else if (tr->flags) {
5626		/* timeout removal doesn't support flags */
5627		return -EINVAL;
5628	}
5629
5630	return 0;
5631}
5632
5633static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5634{
5635	return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5636					    : HRTIMER_MODE_REL;
5637}
5638
5639/*
5640 * Remove or update an existing timeout command
5641 */
5642static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5643{
5644	struct io_timeout_rem *tr = &req->timeout_rem;
5645	struct io_ring_ctx *ctx = req->ctx;
5646	int ret;
5647
5648	spin_lock_irq(&ctx->timeout_lock);
5649	if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5650		ret = io_timeout_cancel(ctx, tr->addr);
5651	else
5652		ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5653					io_translate_timeout_mode(tr->flags));
5654	spin_unlock_irq(&ctx->timeout_lock);
5655
5656	spin_lock(&ctx->completion_lock);
5657	io_cqring_fill_event(ctx, req->user_data, ret, 0);
5658	io_commit_cqring(ctx);
5659	spin_unlock(&ctx->completion_lock);
5660	io_cqring_ev_posted(ctx);
5661	if (ret < 0)
5662		req_set_fail(req);
5663	io_put_req(req);
5664	return 0;
5665}
5666
5667static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5668			   bool is_timeout_link)
5669{
5670	struct io_timeout_data *data;
5671	unsigned flags;
5672	u32 off = READ_ONCE(sqe->off);
5673
5674	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5675		return -EINVAL;
5676	if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5677		return -EINVAL;
5678	if (off && is_timeout_link)
5679		return -EINVAL;
5680	flags = READ_ONCE(sqe->timeout_flags);
5681	if (flags & ~IORING_TIMEOUT_ABS)
5682		return -EINVAL;
5683
5684	req->timeout.off = off;
5685	if (unlikely(off && !req->ctx->off_timeout_used))
5686		req->ctx->off_timeout_used = true;
5687
5688	if (!req->async_data && io_alloc_async_data(req))
5689		return -ENOMEM;
5690
5691	data = req->async_data;
5692	data->req = req;
5693
5694	if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5695		return -EFAULT;
5696
5697	data->mode = io_translate_timeout_mode(flags);
5698	hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5699	if (is_timeout_link)
5700		io_req_track_inflight(req);
5701	return 0;
5702}
5703
5704static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5705{
5706	struct io_ring_ctx *ctx = req->ctx;
5707	struct io_timeout_data *data = req->async_data;
5708	struct list_head *entry;
5709	u32 tail, off = req->timeout.off;
5710
5711	spin_lock_irq(&ctx->timeout_lock);
5712
5713	/*
5714	 * sqe->off holds how many events that need to occur for this
5715	 * timeout event to be satisfied. If it isn't set, then this is
5716	 * a pure timeout request, sequence isn't used.
5717	 */
5718	if (io_is_timeout_noseq(req)) {
5719		entry = ctx->timeout_list.prev;
5720		goto add;
5721	}
5722
5723	tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5724	req->timeout.target_seq = tail + off;
5725
5726	/* Update the las