io_uring.c revision e98e49b2
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
5 *
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
8 *
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
16 * CQ entries.
17 *
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
23 * head will do).
24 *
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
28 * between.
29 *
30 * Also see the examples in the liburing library:
31 *
32 *	git://git.kernel.dk/liburing
33 *
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
38 *
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
41 */
42#include <linux/kernel.h>
43#include <linux/init.h>
44#include <linux/errno.h>
45#include <linux/syscalls.h>
46#include <linux/compat.h>
47#include <net/compat.h>
48#include <linux/refcount.h>
49#include <linux/uio.h>
50#include <linux/bits.h>
51
52#include <linux/sched/signal.h>
53#include <linux/fs.h>
54#include <linux/file.h>
55#include <linux/fdtable.h>
56#include <linux/mm.h>
57#include <linux/mman.h>
58#include <linux/percpu.h>
59#include <linux/slab.h>
60#include <linux/blkdev.h>
61#include <linux/bvec.h>
62#include <linux/net.h>
63#include <net/sock.h>
64#include <net/af_unix.h>
65#include <net/scm.h>
66#include <linux/anon_inodes.h>
67#include <linux/sched/mm.h>
68#include <linux/uaccess.h>
69#include <linux/nospec.h>
70#include <linux/sizes.h>
71#include <linux/hugetlb.h>
72#include <linux/highmem.h>
73#include <linux/namei.h>
74#include <linux/fsnotify.h>
75#include <linux/fadvise.h>
76#include <linux/eventpoll.h>
77#include <linux/splice.h>
78#include <linux/task_work.h>
79#include <linux/pagemap.h>
80#include <linux/io_uring.h>
81#include <linux/tracehook.h>
82
83#define CREATE_TRACE_POINTS
84#include <trace/events/io_uring.h>
85
86#include <uapi/linux/io_uring.h>
87
88#include "internal.h"
89#include "io-wq.h"
90
91#define IORING_MAX_ENTRIES	32768
92#define IORING_MAX_CQ_ENTRIES	(2 * IORING_MAX_ENTRIES)
93#define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
94
95/* 512 entries per page on 64-bit archs, 64 pages max */
96#define IORING_MAX_FIXED_FILES	(1U << 15)
97#define IORING_MAX_RESTRICTIONS	(IORING_RESTRICTION_LAST + \
98				 IORING_REGISTER_LAST + IORING_OP_LAST)
99
100#define IO_RSRC_TAG_TABLE_SHIFT	9
101#define IO_RSRC_TAG_TABLE_MAX	(1U << IO_RSRC_TAG_TABLE_SHIFT)
102#define IO_RSRC_TAG_TABLE_MASK	(IO_RSRC_TAG_TABLE_MAX - 1)
103
104#define IORING_MAX_REG_BUFFERS	(1U << 14)
105
106#define SQE_VALID_FLAGS	(IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK|	\
107				IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
108				IOSQE_BUFFER_SELECT)
109#define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
110				REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS)
111
112#define IO_TCTX_REFS_CACHE_NR	(1U << 10)
113
114struct io_uring {
115	u32 head ____cacheline_aligned_in_smp;
116	u32 tail ____cacheline_aligned_in_smp;
117};
118
119/*
120 * This data is shared with the application through the mmap at offsets
121 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
122 *
123 * The offsets to the member fields are published through struct
124 * io_sqring_offsets when calling io_uring_setup.
125 */
126struct io_rings {
127	/*
128	 * Head and tail offsets into the ring; the offsets need to be
129	 * masked to get valid indices.
130	 *
131	 * The kernel controls head of the sq ring and the tail of the cq ring,
132	 * and the application controls tail of the sq ring and the head of the
133	 * cq ring.
134	 */
135	struct io_uring		sq, cq;
136	/*
137	 * Bitmasks to apply to head and tail offsets (constant, equals
138	 * ring_entries - 1)
139	 */
140	u32			sq_ring_mask, cq_ring_mask;
141	/* Ring sizes (constant, power of 2) */
142	u32			sq_ring_entries, cq_ring_entries;
143	/*
144	 * Number of invalid entries dropped by the kernel due to
145	 * invalid index stored in array
146	 *
147	 * Written by the kernel, shouldn't be modified by the
148	 * application (i.e. get number of "new events" by comparing to
149	 * cached value).
150	 *
151	 * After a new SQ head value was read by the application this
152	 * counter includes all submissions that were dropped reaching
153	 * the new SQ head (and possibly more).
154	 */
155	u32			sq_dropped;
156	/*
157	 * Runtime SQ flags
158	 *
159	 * Written by the kernel, shouldn't be modified by the
160	 * application.
161	 *
162	 * The application needs a full memory barrier before checking
163	 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
164	 */
165	u32			sq_flags;
166	/*
167	 * Runtime CQ flags
168	 *
169	 * Written by the application, shouldn't be modified by the
170	 * kernel.
171	 */
172	u32			cq_flags;
173	/*
174	 * Number of completion events lost because the queue was full;
175	 * this should be avoided by the application by making sure
176	 * there are not more requests pending than there is space in
177	 * the completion queue.
178	 *
179	 * Written by the kernel, shouldn't be modified by the
180	 * application (i.e. get number of "new events" by comparing to
181	 * cached value).
182	 *
183	 * As completion events come in out of order this counter is not
184	 * ordered with any other data.
185	 */
186	u32			cq_overflow;
187	/*
188	 * Ring buffer of completion events.
189	 *
190	 * The kernel writes completion events fresh every time they are
191	 * produced, so the application is allowed to modify pending
192	 * entries.
193	 */
194	struct io_uring_cqe	cqes[] ____cacheline_aligned_in_smp;
195};
196
197enum io_uring_cmd_flags {
198	IO_URING_F_NONBLOCK		= 1,
199	IO_URING_F_COMPLETE_DEFER	= 2,
200};
201
202struct io_mapped_ubuf {
203	u64		ubuf;
204	u64		ubuf_end;
205	unsigned int	nr_bvecs;
206	unsigned long	acct_pages;
207	struct bio_vec	bvec[];
208};
209
210struct io_ring_ctx;
211
212struct io_overflow_cqe {
213	struct io_uring_cqe cqe;
214	struct list_head list;
215};
216
217struct io_fixed_file {
218	/* file * with additional FFS_* flags */
219	unsigned long file_ptr;
220};
221
222struct io_rsrc_put {
223	struct list_head list;
224	u64 tag;
225	union {
226		void *rsrc;
227		struct file *file;
228		struct io_mapped_ubuf *buf;
229	};
230};
231
232struct io_file_table {
233	struct io_fixed_file *files;
234};
235
236struct io_rsrc_node {
237	struct percpu_ref		refs;
238	struct list_head		node;
239	struct list_head		rsrc_list;
240	struct io_rsrc_data		*rsrc_data;
241	struct llist_node		llist;
242	bool				done;
243};
244
245typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
246
247struct io_rsrc_data {
248	struct io_ring_ctx		*ctx;
249
250	u64				**tags;
251	unsigned int			nr;
252	rsrc_put_fn			*do_put;
253	atomic_t			refs;
254	struct completion		done;
255	bool				quiesce;
256};
257
258struct io_buffer {
259	struct list_head list;
260	__u64 addr;
261	__u32 len;
262	__u16 bid;
263};
264
265struct io_restriction {
266	DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
267	DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
268	u8 sqe_flags_allowed;
269	u8 sqe_flags_required;
270	bool registered;
271};
272
273enum {
274	IO_SQ_THREAD_SHOULD_STOP = 0,
275	IO_SQ_THREAD_SHOULD_PARK,
276};
277
278struct io_sq_data {
279	refcount_t		refs;
280	atomic_t		park_pending;
281	struct mutex		lock;
282
283	/* ctx's that are using this sqd */
284	struct list_head	ctx_list;
285
286	struct task_struct	*thread;
287	struct wait_queue_head	wait;
288
289	unsigned		sq_thread_idle;
290	int			sq_cpu;
291	pid_t			task_pid;
292	pid_t			task_tgid;
293
294	unsigned long		state;
295	struct completion	exited;
296};
297
298#define IO_COMPL_BATCH			32
299#define IO_REQ_CACHE_SIZE		32
300#define IO_REQ_ALLOC_BATCH		8
301
302struct io_submit_link {
303	struct io_kiocb		*head;
304	struct io_kiocb		*last;
305};
306
307struct io_submit_state {
308	struct blk_plug		plug;
309	struct io_submit_link	link;
310
311	/*
312	 * io_kiocb alloc cache
313	 */
314	void			*reqs[IO_REQ_CACHE_SIZE];
315	unsigned int		free_reqs;
316
317	bool			plug_started;
318
319	/*
320	 * Batch completion logic
321	 */
322	struct io_kiocb		*compl_reqs[IO_COMPL_BATCH];
323	unsigned int		compl_nr;
324	/* inline/task_work completion list, under ->uring_lock */
325	struct list_head	free_list;
326
327	unsigned int		ios_left;
328};
329
330struct io_ring_ctx {
331	/* const or read-mostly hot data */
332	struct {
333		struct percpu_ref	refs;
334
335		struct io_rings		*rings;
336		unsigned int		flags;
337		unsigned int		compat: 1;
338		unsigned int		drain_next: 1;
339		unsigned int		eventfd_async: 1;
340		unsigned int		restricted: 1;
341		unsigned int		off_timeout_used: 1;
342		unsigned int		drain_active: 1;
343	} ____cacheline_aligned_in_smp;
344
345	/* submission data */
346	struct {
347		struct mutex		uring_lock;
348
349		/*
350		 * Ring buffer of indices into array of io_uring_sqe, which is
351		 * mmapped by the application using the IORING_OFF_SQES offset.
352		 *
353		 * This indirection could e.g. be used to assign fixed
354		 * io_uring_sqe entries to operations and only submit them to
355		 * the queue when needed.
356		 *
357		 * The kernel modifies neither the indices array nor the entries
358		 * array.
359		 */
360		u32			*sq_array;
361		struct io_uring_sqe	*sq_sqes;
362		unsigned		cached_sq_head;
363		unsigned		sq_entries;
364		struct list_head	defer_list;
365
366		/*
367		 * Fixed resources fast path, should be accessed only under
368		 * uring_lock, and updated through io_uring_register(2)
369		 */
370		struct io_rsrc_node	*rsrc_node;
371		struct io_file_table	file_table;
372		unsigned		nr_user_files;
373		unsigned		nr_user_bufs;
374		struct io_mapped_ubuf	**user_bufs;
375
376		struct io_submit_state	submit_state;
377		struct list_head	timeout_list;
378		struct list_head	cq_overflow_list;
379		struct xarray		io_buffers;
380		struct xarray		personalities;
381		u32			pers_next;
382		unsigned		sq_thread_idle;
383	} ____cacheline_aligned_in_smp;
384
385	/* IRQ completion list, under ->completion_lock */
386	struct list_head	locked_free_list;
387	unsigned int		locked_free_nr;
388
389	const struct cred	*sq_creds;	/* cred used for __io_sq_thread() */
390	struct io_sq_data	*sq_data;	/* if using sq thread polling */
391
392	struct wait_queue_head	sqo_sq_wait;
393	struct list_head	sqd_list;
394
395	unsigned long		check_cq_overflow;
396
397	struct {
398		unsigned		cached_cq_tail;
399		unsigned		cq_entries;
400		struct eventfd_ctx	*cq_ev_fd;
401		struct wait_queue_head	poll_wait;
402		struct wait_queue_head	cq_wait;
403		unsigned		cq_extra;
404		atomic_t		cq_timeouts;
405		struct fasync_struct	*cq_fasync;
406		unsigned		cq_last_tm_flush;
407	} ____cacheline_aligned_in_smp;
408
409	struct {
410		spinlock_t		completion_lock;
411
412		spinlock_t		timeout_lock;
413
414		/*
415		 * ->iopoll_list is protected by the ctx->uring_lock for
416		 * io_uring instances that don't use IORING_SETUP_SQPOLL.
417		 * For SQPOLL, only the single threaded io_sq_thread() will
418		 * manipulate the list, hence no extra locking is needed there.
419		 */
420		struct list_head	iopoll_list;
421		struct hlist_head	*cancel_hash;
422		unsigned		cancel_hash_bits;
423		bool			poll_multi_queue;
424	} ____cacheline_aligned_in_smp;
425
426	struct io_restriction		restrictions;
427
428	/* slow path rsrc auxilary data, used by update/register */
429	struct {
430		struct io_rsrc_node		*rsrc_backup_node;
431		struct io_mapped_ubuf		*dummy_ubuf;
432		struct io_rsrc_data		*file_data;
433		struct io_rsrc_data		*buf_data;
434
435		struct delayed_work		rsrc_put_work;
436		struct llist_head		rsrc_put_llist;
437		struct list_head		rsrc_ref_list;
438		spinlock_t			rsrc_ref_lock;
439	};
440
441	/* Keep this last, we don't need it for the fast path */
442	struct {
443		#if defined(CONFIG_UNIX)
444			struct socket		*ring_sock;
445		#endif
446		/* hashed buffered write serialization */
447		struct io_wq_hash		*hash_map;
448
449		/* Only used for accounting purposes */
450		struct user_struct		*user;
451		struct mm_struct		*mm_account;
452
453		/* ctx exit and cancelation */
454		struct llist_head		fallback_llist;
455		struct delayed_work		fallback_work;
456		struct work_struct		exit_work;
457		struct list_head		tctx_list;
458		struct completion		ref_comp;
459	};
460};
461
462struct io_uring_task {
463	/* submission side */
464	int			cached_refs;
465	struct xarray		xa;
466	struct wait_queue_head	wait;
467	const struct io_ring_ctx *last;
468	struct io_wq		*io_wq;
469	struct percpu_counter	inflight;
470	atomic_t		inflight_tracked;
471	atomic_t		in_idle;
472
473	spinlock_t		task_lock;
474	struct io_wq_work_list	task_list;
475	struct callback_head	task_work;
476	bool			task_running;
477};
478
479/*
480 * First field must be the file pointer in all the
481 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
482 */
483struct io_poll_iocb {
484	struct file			*file;
485	struct wait_queue_head		*head;
486	__poll_t			events;
487	bool				done;
488	bool				canceled;
489	struct wait_queue_entry		wait;
490};
491
492struct io_poll_update {
493	struct file			*file;
494	u64				old_user_data;
495	u64				new_user_data;
496	__poll_t			events;
497	bool				update_events;
498	bool				update_user_data;
499};
500
501struct io_close {
502	struct file			*file;
503	int				fd;
504};
505
506struct io_timeout_data {
507	struct io_kiocb			*req;
508	struct hrtimer			timer;
509	struct timespec64		ts;
510	enum hrtimer_mode		mode;
511};
512
513struct io_accept {
514	struct file			*file;
515	struct sockaddr __user		*addr;
516	int __user			*addr_len;
517	int				flags;
518	unsigned long			nofile;
519};
520
521struct io_sync {
522	struct file			*file;
523	loff_t				len;
524	loff_t				off;
525	int				flags;
526	int				mode;
527};
528
529struct io_cancel {
530	struct file			*file;
531	u64				addr;
532};
533
534struct io_timeout {
535	struct file			*file;
536	u32				off;
537	u32				target_seq;
538	struct list_head		list;
539	/* head of the link, used by linked timeouts only */
540	struct io_kiocb			*head;
541	/* for linked completions */
542	struct io_kiocb			*prev;
543};
544
545struct io_timeout_rem {
546	struct file			*file;
547	u64				addr;
548
549	/* timeout update */
550	struct timespec64		ts;
551	u32				flags;
552};
553
554struct io_rw {
555	/* NOTE: kiocb has the file as the first member, so don't do it here */
556	struct kiocb			kiocb;
557	u64				addr;
558	u64				len;
559};
560
561struct io_connect {
562	struct file			*file;
563	struct sockaddr __user		*addr;
564	int				addr_len;
565};
566
567struct io_sr_msg {
568	struct file			*file;
569	union {
570		struct compat_msghdr __user	*umsg_compat;
571		struct user_msghdr __user	*umsg;
572		void __user			*buf;
573	};
574	int				msg_flags;
575	int				bgid;
576	size_t				len;
577	struct io_buffer		*kbuf;
578};
579
580struct io_open {
581	struct file			*file;
582	int				dfd;
583	struct filename			*filename;
584	struct open_how			how;
585	unsigned long			nofile;
586};
587
588struct io_rsrc_update {
589	struct file			*file;
590	u64				arg;
591	u32				nr_args;
592	u32				offset;
593};
594
595struct io_fadvise {
596	struct file			*file;
597	u64				offset;
598	u32				len;
599	u32				advice;
600};
601
602struct io_madvise {
603	struct file			*file;
604	u64				addr;
605	u32				len;
606	u32				advice;
607};
608
609struct io_epoll {
610	struct file			*file;
611	int				epfd;
612	int				op;
613	int				fd;
614	struct epoll_event		event;
615};
616
617struct io_splice {
618	struct file			*file_out;
619	struct file			*file_in;
620	loff_t				off_out;
621	loff_t				off_in;
622	u64				len;
623	unsigned int			flags;
624};
625
626struct io_provide_buf {
627	struct file			*file;
628	__u64				addr;
629	__u32				len;
630	__u32				bgid;
631	__u16				nbufs;
632	__u16				bid;
633};
634
635struct io_statx {
636	struct file			*file;
637	int				dfd;
638	unsigned int			mask;
639	unsigned int			flags;
640	const char __user		*filename;
641	struct statx __user		*buffer;
642};
643
644struct io_shutdown {
645	struct file			*file;
646	int				how;
647};
648
649struct io_rename {
650	struct file			*file;
651	int				old_dfd;
652	int				new_dfd;
653	struct filename			*oldpath;
654	struct filename			*newpath;
655	int				flags;
656};
657
658struct io_unlink {
659	struct file			*file;
660	int				dfd;
661	int				flags;
662	struct filename			*filename;
663};
664
665struct io_completion {
666	struct file			*file;
667	u32				cflags;
668};
669
670struct io_async_connect {
671	struct sockaddr_storage		address;
672};
673
674struct io_async_msghdr {
675	struct iovec			fast_iov[UIO_FASTIOV];
676	/* points to an allocated iov, if NULL we use fast_iov instead */
677	struct iovec			*free_iov;
678	struct sockaddr __user		*uaddr;
679	struct msghdr			msg;
680	struct sockaddr_storage		addr;
681};
682
683struct io_async_rw {
684	struct iovec			fast_iov[UIO_FASTIOV];
685	const struct iovec		*free_iovec;
686	struct iov_iter			iter;
687	size_t				bytes_done;
688	struct wait_page_queue		wpq;
689};
690
691enum {
692	REQ_F_FIXED_FILE_BIT	= IOSQE_FIXED_FILE_BIT,
693	REQ_F_IO_DRAIN_BIT	= IOSQE_IO_DRAIN_BIT,
694	REQ_F_LINK_BIT		= IOSQE_IO_LINK_BIT,
695	REQ_F_HARDLINK_BIT	= IOSQE_IO_HARDLINK_BIT,
696	REQ_F_FORCE_ASYNC_BIT	= IOSQE_ASYNC_BIT,
697	REQ_F_BUFFER_SELECT_BIT	= IOSQE_BUFFER_SELECT_BIT,
698
699	/* first byte is taken by user flags, shift it to not overlap */
700	REQ_F_FAIL_BIT		= 8,
701	REQ_F_INFLIGHT_BIT,
702	REQ_F_CUR_POS_BIT,
703	REQ_F_NOWAIT_BIT,
704	REQ_F_LINK_TIMEOUT_BIT,
705	REQ_F_NEED_CLEANUP_BIT,
706	REQ_F_POLLED_BIT,
707	REQ_F_BUFFER_SELECTED_BIT,
708	REQ_F_COMPLETE_INLINE_BIT,
709	REQ_F_REISSUE_BIT,
710	REQ_F_DONT_REISSUE_BIT,
711	REQ_F_CREDS_BIT,
712	REQ_F_REFCOUNT_BIT,
713	REQ_F_ARM_LTIMEOUT_BIT,
714	/* keep async read/write and isreg together and in order */
715	REQ_F_NOWAIT_READ_BIT,
716	REQ_F_NOWAIT_WRITE_BIT,
717	REQ_F_ISREG_BIT,
718
719	/* not a real bit, just to check we're not overflowing the space */
720	__REQ_F_LAST_BIT,
721};
722
723enum {
724	/* ctx owns file */
725	REQ_F_FIXED_FILE	= BIT(REQ_F_FIXED_FILE_BIT),
726	/* drain existing IO first */
727	REQ_F_IO_DRAIN		= BIT(REQ_F_IO_DRAIN_BIT),
728	/* linked sqes */
729	REQ_F_LINK		= BIT(REQ_F_LINK_BIT),
730	/* doesn't sever on completion < 0 */
731	REQ_F_HARDLINK		= BIT(REQ_F_HARDLINK_BIT),
732	/* IOSQE_ASYNC */
733	REQ_F_FORCE_ASYNC	= BIT(REQ_F_FORCE_ASYNC_BIT),
734	/* IOSQE_BUFFER_SELECT */
735	REQ_F_BUFFER_SELECT	= BIT(REQ_F_BUFFER_SELECT_BIT),
736
737	/* fail rest of links */
738	REQ_F_FAIL		= BIT(REQ_F_FAIL_BIT),
739	/* on inflight list, should be cancelled and waited on exit reliably */
740	REQ_F_INFLIGHT		= BIT(REQ_F_INFLIGHT_BIT),
741	/* read/write uses file position */
742	REQ_F_CUR_POS		= BIT(REQ_F_CUR_POS_BIT),
743	/* must not punt to workers */
744	REQ_F_NOWAIT		= BIT(REQ_F_NOWAIT_BIT),
745	/* has or had linked timeout */
746	REQ_F_LINK_TIMEOUT	= BIT(REQ_F_LINK_TIMEOUT_BIT),
747	/* needs cleanup */
748	REQ_F_NEED_CLEANUP	= BIT(REQ_F_NEED_CLEANUP_BIT),
749	/* already went through poll handler */
750	REQ_F_POLLED		= BIT(REQ_F_POLLED_BIT),
751	/* buffer already selected */
752	REQ_F_BUFFER_SELECTED	= BIT(REQ_F_BUFFER_SELECTED_BIT),
753	/* completion is deferred through io_comp_state */
754	REQ_F_COMPLETE_INLINE	= BIT(REQ_F_COMPLETE_INLINE_BIT),
755	/* caller should reissue async */
756	REQ_F_REISSUE		= BIT(REQ_F_REISSUE_BIT),
757	/* don't attempt request reissue, see io_rw_reissue() */
758	REQ_F_DONT_REISSUE	= BIT(REQ_F_DONT_REISSUE_BIT),
759	/* supports async reads */
760	REQ_F_NOWAIT_READ	= BIT(REQ_F_NOWAIT_READ_BIT),
761	/* supports async writes */
762	REQ_F_NOWAIT_WRITE	= BIT(REQ_F_NOWAIT_WRITE_BIT),
763	/* regular file */
764	REQ_F_ISREG		= BIT(REQ_F_ISREG_BIT),
765	/* has creds assigned */
766	REQ_F_CREDS		= BIT(REQ_F_CREDS_BIT),
767	/* skip refcounting if not set */
768	REQ_F_REFCOUNT		= BIT(REQ_F_REFCOUNT_BIT),
769	/* there is a linked timeout that has to be armed */
770	REQ_F_ARM_LTIMEOUT	= BIT(REQ_F_ARM_LTIMEOUT_BIT),
771};
772
773struct async_poll {
774	struct io_poll_iocb	poll;
775	struct io_poll_iocb	*double_poll;
776};
777
778typedef void (*io_req_tw_func_t)(struct io_kiocb *req);
779
780struct io_task_work {
781	union {
782		struct io_wq_work_node	node;
783		struct llist_node	fallback_node;
784	};
785	io_req_tw_func_t		func;
786};
787
788enum {
789	IORING_RSRC_FILE		= 0,
790	IORING_RSRC_BUFFER		= 1,
791};
792
793/*
794 * NOTE! Each of the iocb union members has the file pointer
795 * as the first entry in their struct definition. So you can
796 * access the file pointer through any of the sub-structs,
797 * or directly as just 'ki_filp' in this struct.
798 */
799struct io_kiocb {
800	union {
801		struct file		*file;
802		struct io_rw		rw;
803		struct io_poll_iocb	poll;
804		struct io_poll_update	poll_update;
805		struct io_accept	accept;
806		struct io_sync		sync;
807		struct io_cancel	cancel;
808		struct io_timeout	timeout;
809		struct io_timeout_rem	timeout_rem;
810		struct io_connect	connect;
811		struct io_sr_msg	sr_msg;
812		struct io_open		open;
813		struct io_close		close;
814		struct io_rsrc_update	rsrc_update;
815		struct io_fadvise	fadvise;
816		struct io_madvise	madvise;
817		struct io_epoll		epoll;
818		struct io_splice	splice;
819		struct io_provide_buf	pbuf;
820		struct io_statx		statx;
821		struct io_shutdown	shutdown;
822		struct io_rename	rename;
823		struct io_unlink	unlink;
824		/* use only after cleaning per-op data, see io_clean_op() */
825		struct io_completion	compl;
826	};
827
828	/* opcode allocated if it needs to store data for async defer */
829	void				*async_data;
830	u8				opcode;
831	/* polled IO has completed */
832	u8				iopoll_completed;
833
834	u16				buf_index;
835	u32				result;
836
837	struct io_ring_ctx		*ctx;
838	unsigned int			flags;
839	atomic_t			refs;
840	struct task_struct		*task;
841	u64				user_data;
842
843	struct io_kiocb			*link;
844	struct percpu_ref		*fixed_rsrc_refs;
845
846	/* used with ctx->iopoll_list with reads/writes */
847	struct list_head		inflight_entry;
848	struct io_task_work		io_task_work;
849	/* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
850	struct hlist_node		hash_node;
851	struct async_poll		*apoll;
852	struct io_wq_work		work;
853	const struct cred		*creds;
854
855	/* store used ubuf, so we can prevent reloading */
856	struct io_mapped_ubuf		*imu;
857};
858
859struct io_tctx_node {
860	struct list_head	ctx_node;
861	struct task_struct	*task;
862	struct io_ring_ctx	*ctx;
863};
864
865struct io_defer_entry {
866	struct list_head	list;
867	struct io_kiocb		*req;
868	u32			seq;
869};
870
871struct io_op_def {
872	/* needs req->file assigned */
873	unsigned		needs_file : 1;
874	/* hash wq insertion if file is a regular file */
875	unsigned		hash_reg_file : 1;
876	/* unbound wq insertion if file is a non-regular file */
877	unsigned		unbound_nonreg_file : 1;
878	/* opcode is not supported by this kernel */
879	unsigned		not_supported : 1;
880	/* set if opcode supports polled "wait" */
881	unsigned		pollin : 1;
882	unsigned		pollout : 1;
883	/* op supports buffer selection */
884	unsigned		buffer_select : 1;
885	/* do prep async if is going to be punted */
886	unsigned		needs_async_setup : 1;
887	/* should block plug */
888	unsigned		plug : 1;
889	/* size of async data needed, if any */
890	unsigned short		async_size;
891};
892
893static const struct io_op_def io_op_defs[] = {
894	[IORING_OP_NOP] = {},
895	[IORING_OP_READV] = {
896		.needs_file		= 1,
897		.unbound_nonreg_file	= 1,
898		.pollin			= 1,
899		.buffer_select		= 1,
900		.needs_async_setup	= 1,
901		.plug			= 1,
902		.async_size		= sizeof(struct io_async_rw),
903	},
904	[IORING_OP_WRITEV] = {
905		.needs_file		= 1,
906		.hash_reg_file		= 1,
907		.unbound_nonreg_file	= 1,
908		.pollout		= 1,
909		.needs_async_setup	= 1,
910		.plug			= 1,
911		.async_size		= sizeof(struct io_async_rw),
912	},
913	[IORING_OP_FSYNC] = {
914		.needs_file		= 1,
915	},
916	[IORING_OP_READ_FIXED] = {
917		.needs_file		= 1,
918		.unbound_nonreg_file	= 1,
919		.pollin			= 1,
920		.plug			= 1,
921		.async_size		= sizeof(struct io_async_rw),
922	},
923	[IORING_OP_WRITE_FIXED] = {
924		.needs_file		= 1,
925		.hash_reg_file		= 1,
926		.unbound_nonreg_file	= 1,
927		.pollout		= 1,
928		.plug			= 1,
929		.async_size		= sizeof(struct io_async_rw),
930	},
931	[IORING_OP_POLL_ADD] = {
932		.needs_file		= 1,
933		.unbound_nonreg_file	= 1,
934	},
935	[IORING_OP_POLL_REMOVE] = {},
936	[IORING_OP_SYNC_FILE_RANGE] = {
937		.needs_file		= 1,
938	},
939	[IORING_OP_SENDMSG] = {
940		.needs_file		= 1,
941		.unbound_nonreg_file	= 1,
942		.pollout		= 1,
943		.needs_async_setup	= 1,
944		.async_size		= sizeof(struct io_async_msghdr),
945	},
946	[IORING_OP_RECVMSG] = {
947		.needs_file		= 1,
948		.unbound_nonreg_file	= 1,
949		.pollin			= 1,
950		.buffer_select		= 1,
951		.needs_async_setup	= 1,
952		.async_size		= sizeof(struct io_async_msghdr),
953	},
954	[IORING_OP_TIMEOUT] = {
955		.async_size		= sizeof(struct io_timeout_data),
956	},
957	[IORING_OP_TIMEOUT_REMOVE] = {
958		/* used by timeout updates' prep() */
959	},
960	[IORING_OP_ACCEPT] = {
961		.needs_file		= 1,
962		.unbound_nonreg_file	= 1,
963		.pollin			= 1,
964	},
965	[IORING_OP_ASYNC_CANCEL] = {},
966	[IORING_OP_LINK_TIMEOUT] = {
967		.async_size		= sizeof(struct io_timeout_data),
968	},
969	[IORING_OP_CONNECT] = {
970		.needs_file		= 1,
971		.unbound_nonreg_file	= 1,
972		.pollout		= 1,
973		.needs_async_setup	= 1,
974		.async_size		= sizeof(struct io_async_connect),
975	},
976	[IORING_OP_FALLOCATE] = {
977		.needs_file		= 1,
978	},
979	[IORING_OP_OPENAT] = {},
980	[IORING_OP_CLOSE] = {},
981	[IORING_OP_FILES_UPDATE] = {},
982	[IORING_OP_STATX] = {},
983	[IORING_OP_READ] = {
984		.needs_file		= 1,
985		.unbound_nonreg_file	= 1,
986		.pollin			= 1,
987		.buffer_select		= 1,
988		.plug			= 1,
989		.async_size		= sizeof(struct io_async_rw),
990	},
991	[IORING_OP_WRITE] = {
992		.needs_file		= 1,
993		.unbound_nonreg_file	= 1,
994		.pollout		= 1,
995		.plug			= 1,
996		.async_size		= sizeof(struct io_async_rw),
997	},
998	[IORING_OP_FADVISE] = {
999		.needs_file		= 1,
1000	},
1001	[IORING_OP_MADVISE] = {},
1002	[IORING_OP_SEND] = {
1003		.needs_file		= 1,
1004		.unbound_nonreg_file	= 1,
1005		.pollout		= 1,
1006	},
1007	[IORING_OP_RECV] = {
1008		.needs_file		= 1,
1009		.unbound_nonreg_file	= 1,
1010		.pollin			= 1,
1011		.buffer_select		= 1,
1012	},
1013	[IORING_OP_OPENAT2] = {
1014	},
1015	[IORING_OP_EPOLL_CTL] = {
1016		.unbound_nonreg_file	= 1,
1017	},
1018	[IORING_OP_SPLICE] = {
1019		.needs_file		= 1,
1020		.hash_reg_file		= 1,
1021		.unbound_nonreg_file	= 1,
1022	},
1023	[IORING_OP_PROVIDE_BUFFERS] = {},
1024	[IORING_OP_REMOVE_BUFFERS] = {},
1025	[IORING_OP_TEE] = {
1026		.needs_file		= 1,
1027		.hash_reg_file		= 1,
1028		.unbound_nonreg_file	= 1,
1029	},
1030	[IORING_OP_SHUTDOWN] = {
1031		.needs_file		= 1,
1032	},
1033	[IORING_OP_RENAMEAT] = {},
1034	[IORING_OP_UNLINKAT] = {},
1035};
1036
1037/* requests with any of those set should undergo io_disarm_next() */
1038#define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
1039
1040static bool io_disarm_next(struct io_kiocb *req);
1041static void io_uring_del_tctx_node(unsigned long index);
1042static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1043					 struct task_struct *task,
1044					 bool cancel_all);
1045static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1046
1047static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1048				 long res, unsigned int cflags);
1049static void io_put_req(struct io_kiocb *req);
1050static void io_put_req_deferred(struct io_kiocb *req);
1051static void io_dismantle_req(struct io_kiocb *req);
1052static void io_queue_linked_timeout(struct io_kiocb *req);
1053static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1054				     struct io_uring_rsrc_update2 *up,
1055				     unsigned nr_args);
1056static void io_clean_op(struct io_kiocb *req);
1057static struct file *io_file_get(struct io_ring_ctx *ctx,
1058				struct io_kiocb *req, int fd, bool fixed);
1059static void __io_queue_sqe(struct io_kiocb *req);
1060static void io_rsrc_put_work(struct work_struct *work);
1061
1062static void io_req_task_queue(struct io_kiocb *req);
1063static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1064static int io_req_prep_async(struct io_kiocb *req);
1065
1066static struct kmem_cache *req_cachep;
1067
1068static const struct file_operations io_uring_fops;
1069
1070struct sock *io_uring_get_socket(struct file *file)
1071{
1072#if defined(CONFIG_UNIX)
1073	if (file->f_op == &io_uring_fops) {
1074		struct io_ring_ctx *ctx = file->private_data;
1075
1076		return ctx->ring_sock->sk;
1077	}
1078#endif
1079	return NULL;
1080}
1081EXPORT_SYMBOL(io_uring_get_socket);
1082
1083#define io_for_each_link(pos, head) \
1084	for (pos = (head); pos; pos = pos->link)
1085
1086/*
1087 * Shamelessly stolen from the mm implementation of page reference checking,
1088 * see commit f958d7b528b1 for details.
1089 */
1090#define req_ref_zero_or_close_to_overflow(req)	\
1091	((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1092
1093static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1094{
1095	WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1096	return atomic_inc_not_zero(&req->refs);
1097}
1098
1099static inline bool req_ref_put_and_test(struct io_kiocb *req)
1100{
1101	if (likely(!(req->flags & REQ_F_REFCOUNT)))
1102		return true;
1103
1104	WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1105	return atomic_dec_and_test(&req->refs);
1106}
1107
1108static inline void req_ref_put(struct io_kiocb *req)
1109{
1110	WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1111	WARN_ON_ONCE(req_ref_put_and_test(req));
1112}
1113
1114static inline void req_ref_get(struct io_kiocb *req)
1115{
1116	WARN_ON_ONCE(!(req->flags & REQ_F_REFCOUNT));
1117	WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1118	atomic_inc(&req->refs);
1119}
1120
1121static inline void __io_req_set_refcount(struct io_kiocb *req, int nr)
1122{
1123	if (!(req->flags & REQ_F_REFCOUNT)) {
1124		req->flags |= REQ_F_REFCOUNT;
1125		atomic_set(&req->refs, nr);
1126	}
1127}
1128
1129static inline void io_req_set_refcount(struct io_kiocb *req)
1130{
1131	__io_req_set_refcount(req, 1);
1132}
1133
1134static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1135{
1136	struct io_ring_ctx *ctx = req->ctx;
1137
1138	if (!req->fixed_rsrc_refs) {
1139		req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1140		percpu_ref_get(req->fixed_rsrc_refs);
1141	}
1142}
1143
1144static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1145{
1146	bool got = percpu_ref_tryget(ref);
1147
1148	/* already at zero, wait for ->release() */
1149	if (!got)
1150		wait_for_completion(compl);
1151	percpu_ref_resurrect(ref);
1152	if (got)
1153		percpu_ref_put(ref);
1154}
1155
1156static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1157			  bool cancel_all)
1158{
1159	struct io_kiocb *req;
1160
1161	if (task && head->task != task)
1162		return false;
1163	if (cancel_all)
1164		return true;
1165
1166	io_for_each_link(req, head) {
1167		if (req->flags & REQ_F_INFLIGHT)
1168			return true;
1169	}
1170	return false;
1171}
1172
1173static inline void req_set_fail(struct io_kiocb *req)
1174{
1175	req->flags |= REQ_F_FAIL;
1176}
1177
1178static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1179{
1180	struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1181
1182	complete(&ctx->ref_comp);
1183}
1184
1185static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1186{
1187	return !req->timeout.off;
1188}
1189
1190static void io_fallback_req_func(struct work_struct *work)
1191{
1192	struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
1193						fallback_work.work);
1194	struct llist_node *node = llist_del_all(&ctx->fallback_llist);
1195	struct io_kiocb *req, *tmp;
1196
1197	percpu_ref_get(&ctx->refs);
1198	llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
1199		req->io_task_work.func(req);
1200	percpu_ref_put(&ctx->refs);
1201}
1202
1203static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1204{
1205	struct io_ring_ctx *ctx;
1206	int hash_bits;
1207
1208	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1209	if (!ctx)
1210		return NULL;
1211
1212	/*
1213	 * Use 5 bits less than the max cq entries, that should give us around
1214	 * 32 entries per hash list if totally full and uniformly spread.
1215	 */
1216	hash_bits = ilog2(p->cq_entries);
1217	hash_bits -= 5;
1218	if (hash_bits <= 0)
1219		hash_bits = 1;
1220	ctx->cancel_hash_bits = hash_bits;
1221	ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1222					GFP_KERNEL);
1223	if (!ctx->cancel_hash)
1224		goto err;
1225	__hash_init(ctx->cancel_hash, 1U << hash_bits);
1226
1227	ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1228	if (!ctx->dummy_ubuf)
1229		goto err;
1230	/* set invalid range, so io_import_fixed() fails meeting it */
1231	ctx->dummy_ubuf->ubuf = -1UL;
1232
1233	if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1234			    PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1235		goto err;
1236
1237	ctx->flags = p->flags;
1238	init_waitqueue_head(&ctx->sqo_sq_wait);
1239	INIT_LIST_HEAD(&ctx->sqd_list);
1240	init_waitqueue_head(&ctx->poll_wait);
1241	INIT_LIST_HEAD(&ctx->cq_overflow_list);
1242	init_completion(&ctx->ref_comp);
1243	xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1244	xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1245	mutex_init(&ctx->uring_lock);
1246	init_waitqueue_head(&ctx->cq_wait);
1247	spin_lock_init(&ctx->completion_lock);
1248	spin_lock_init(&ctx->timeout_lock);
1249	INIT_LIST_HEAD(&ctx->iopoll_list);
1250	INIT_LIST_HEAD(&ctx->defer_list);
1251	INIT_LIST_HEAD(&ctx->timeout_list);
1252	spin_lock_init(&ctx->rsrc_ref_lock);
1253	INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1254	INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1255	init_llist_head(&ctx->rsrc_put_llist);
1256	INIT_LIST_HEAD(&ctx->tctx_list);
1257	INIT_LIST_HEAD(&ctx->submit_state.free_list);
1258	INIT_LIST_HEAD(&ctx->locked_free_list);
1259	INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1260	return ctx;
1261err:
1262	kfree(ctx->dummy_ubuf);
1263	kfree(ctx->cancel_hash);
1264	kfree(ctx);
1265	return NULL;
1266}
1267
1268static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1269{
1270	struct io_rings *r = ctx->rings;
1271
1272	WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1273	ctx->cq_extra--;
1274}
1275
1276static bool req_need_defer(struct io_kiocb *req, u32 seq)
1277{
1278	if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1279		struct io_ring_ctx *ctx = req->ctx;
1280
1281		return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1282	}
1283
1284	return false;
1285}
1286
1287#define FFS_ASYNC_READ		0x1UL
1288#define FFS_ASYNC_WRITE		0x2UL
1289#ifdef CONFIG_64BIT
1290#define FFS_ISREG		0x4UL
1291#else
1292#define FFS_ISREG		0x0UL
1293#endif
1294#define FFS_MASK		~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1295
1296static inline bool io_req_ffs_set(struct io_kiocb *req)
1297{
1298	return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE);
1299}
1300
1301static void io_req_track_inflight(struct io_kiocb *req)
1302{
1303	if (!(req->flags & REQ_F_INFLIGHT)) {
1304		req->flags |= REQ_F_INFLIGHT;
1305		atomic_inc(&current->io_uring->inflight_tracked);
1306	}
1307}
1308
1309static inline void io_unprep_linked_timeout(struct io_kiocb *req)
1310{
1311	req->flags &= ~REQ_F_LINK_TIMEOUT;
1312}
1313
1314static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
1315{
1316	if (WARN_ON_ONCE(!req->link))
1317		return NULL;
1318
1319	req->flags &= ~REQ_F_ARM_LTIMEOUT;
1320	req->flags |= REQ_F_LINK_TIMEOUT;
1321
1322	/* linked timeouts should have two refs once prep'ed */
1323	io_req_set_refcount(req);
1324	__io_req_set_refcount(req->link, 2);
1325	return req->link;
1326}
1327
1328static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
1329{
1330	if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
1331		return NULL;
1332	return __io_prep_linked_timeout(req);
1333}
1334
1335static void io_prep_async_work(struct io_kiocb *req)
1336{
1337	const struct io_op_def *def = &io_op_defs[req->opcode];
1338	struct io_ring_ctx *ctx = req->ctx;
1339
1340	if (!(req->flags & REQ_F_CREDS)) {
1341		req->flags |= REQ_F_CREDS;
1342		req->creds = get_current_cred();
1343	}
1344
1345	req->work.list.next = NULL;
1346	req->work.flags = 0;
1347	if (req->flags & REQ_F_FORCE_ASYNC)
1348		req->work.flags |= IO_WQ_WORK_CONCURRENT;
1349
1350	if (req->flags & REQ_F_ISREG) {
1351		if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1352			io_wq_hash_work(&req->work, file_inode(req->file));
1353	} else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1354		if (def->unbound_nonreg_file)
1355			req->work.flags |= IO_WQ_WORK_UNBOUND;
1356	}
1357
1358	switch (req->opcode) {
1359	case IORING_OP_SPLICE:
1360	case IORING_OP_TEE:
1361		if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1362			req->work.flags |= IO_WQ_WORK_UNBOUND;
1363		break;
1364	}
1365}
1366
1367static void io_prep_async_link(struct io_kiocb *req)
1368{
1369	struct io_kiocb *cur;
1370
1371	if (req->flags & REQ_F_LINK_TIMEOUT) {
1372		struct io_ring_ctx *ctx = req->ctx;
1373
1374		spin_lock(&ctx->completion_lock);
1375		io_for_each_link(cur, req)
1376			io_prep_async_work(cur);
1377		spin_unlock(&ctx->completion_lock);
1378	} else {
1379		io_for_each_link(cur, req)
1380			io_prep_async_work(cur);
1381	}
1382}
1383
1384static void io_queue_async_work(struct io_kiocb *req)
1385{
1386	struct io_ring_ctx *ctx = req->ctx;
1387	struct io_kiocb *link = io_prep_linked_timeout(req);
1388	struct io_uring_task *tctx = req->task->io_uring;
1389
1390	BUG_ON(!tctx);
1391	BUG_ON(!tctx->io_wq);
1392
1393	/* init ->work of the whole link before punting */
1394	io_prep_async_link(req);
1395
1396	/*
1397	 * Not expected to happen, but if we do have a bug where this _can_
1398	 * happen, catch it here and ensure the request is marked as
1399	 * canceled. That will make io-wq go through the usual work cancel
1400	 * procedure rather than attempt to run this request (or create a new
1401	 * worker for it).
1402	 */
1403	if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1404		req->work.flags |= IO_WQ_WORK_CANCEL;
1405
1406	trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1407					&req->work, req->flags);
1408	io_wq_enqueue(tctx->io_wq, &req->work);
1409	if (link)
1410		io_queue_linked_timeout(link);
1411}
1412
1413static void io_kill_timeout(struct io_kiocb *req, int status)
1414	__must_hold(&req->ctx->completion_lock)
1415	__must_hold(&req->ctx->timeout_lock)
1416{
1417	struct io_timeout_data *io = req->async_data;
1418
1419	if (hrtimer_try_to_cancel(&io->timer) != -1) {
1420		atomic_set(&req->ctx->cq_timeouts,
1421			atomic_read(&req->ctx->cq_timeouts) + 1);
1422		list_del_init(&req->timeout.list);
1423		io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1424		io_put_req_deferred(req);
1425	}
1426}
1427
1428static void io_queue_deferred(struct io_ring_ctx *ctx)
1429{
1430	while (!list_empty(&ctx->defer_list)) {
1431		struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1432						struct io_defer_entry, list);
1433
1434		if (req_need_defer(de->req, de->seq))
1435			break;
1436		list_del_init(&de->list);
1437		io_req_task_queue(de->req);
1438		kfree(de);
1439	}
1440}
1441
1442static void io_flush_timeouts(struct io_ring_ctx *ctx)
1443	__must_hold(&ctx->completion_lock)
1444{
1445	u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1446
1447	spin_lock_irq(&ctx->timeout_lock);
1448	while (!list_empty(&ctx->timeout_list)) {
1449		u32 events_needed, events_got;
1450		struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1451						struct io_kiocb, timeout.list);
1452
1453		if (io_is_timeout_noseq(req))
1454			break;
1455
1456		/*
1457		 * Since seq can easily wrap around over time, subtract
1458		 * the last seq at which timeouts were flushed before comparing.
1459		 * Assuming not more than 2^31-1 events have happened since,
1460		 * these subtractions won't have wrapped, so we can check if
1461		 * target is in [last_seq, current_seq] by comparing the two.
1462		 */
1463		events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1464		events_got = seq - ctx->cq_last_tm_flush;
1465		if (events_got < events_needed)
1466			break;
1467
1468		list_del_init(&req->timeout.list);
1469		io_kill_timeout(req, 0);
1470	}
1471	ctx->cq_last_tm_flush = seq;
1472	spin_unlock_irq(&ctx->timeout_lock);
1473}
1474
1475static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1476{
1477	if (ctx->off_timeout_used)
1478		io_flush_timeouts(ctx);
1479	if (ctx->drain_active)
1480		io_queue_deferred(ctx);
1481}
1482
1483static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1484{
1485	if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1486		__io_commit_cqring_flush(ctx);
1487	/* order cqe stores with ring update */
1488	smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1489}
1490
1491static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1492{
1493	struct io_rings *r = ctx->rings;
1494
1495	return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1496}
1497
1498static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1499{
1500	return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1501}
1502
1503static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1504{
1505	struct io_rings *rings = ctx->rings;
1506	unsigned tail, mask = ctx->cq_entries - 1;
1507
1508	/*
1509	 * writes to the cq entry need to come after reading head; the
1510	 * control dependency is enough as we're using WRITE_ONCE to
1511	 * fill the cq entry
1512	 */
1513	if (__io_cqring_events(ctx) == ctx->cq_entries)
1514		return NULL;
1515
1516	tail = ctx->cached_cq_tail++;
1517	return &rings->cqes[tail & mask];
1518}
1519
1520static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1521{
1522	if (likely(!ctx->cq_ev_fd))
1523		return false;
1524	if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1525		return false;
1526	return !ctx->eventfd_async || io_wq_current_is_worker();
1527}
1528
1529static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1530{
1531	/*
1532	 * wake_up_all() may seem excessive, but io_wake_function() and
1533	 * io_should_wake() handle the termination of the loop and only
1534	 * wake as many waiters as we need to.
1535	 */
1536	if (wq_has_sleeper(&ctx->cq_wait))
1537		wake_up_all(&ctx->cq_wait);
1538	if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1539		wake_up(&ctx->sq_data->wait);
1540	if (io_should_trigger_evfd(ctx))
1541		eventfd_signal(ctx->cq_ev_fd, 1);
1542	if (waitqueue_active(&ctx->poll_wait)) {
1543		wake_up_interruptible(&ctx->poll_wait);
1544		kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1545	}
1546}
1547
1548static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1549{
1550	if (ctx->flags & IORING_SETUP_SQPOLL) {
1551		if (wq_has_sleeper(&ctx->cq_wait))
1552			wake_up_all(&ctx->cq_wait);
1553	}
1554	if (io_should_trigger_evfd(ctx))
1555		eventfd_signal(ctx->cq_ev_fd, 1);
1556	if (waitqueue_active(&ctx->poll_wait)) {
1557		wake_up_interruptible(&ctx->poll_wait);
1558		kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1559	}
1560}
1561
1562/* Returns true if there are no backlogged entries after the flush */
1563static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1564{
1565	bool all_flushed, posted;
1566
1567	if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1568		return false;
1569
1570	posted = false;
1571	spin_lock(&ctx->completion_lock);
1572	while (!list_empty(&ctx->cq_overflow_list)) {
1573		struct io_uring_cqe *cqe = io_get_cqe(ctx);
1574		struct io_overflow_cqe *ocqe;
1575
1576		if (!cqe && !force)
1577			break;
1578		ocqe = list_first_entry(&ctx->cq_overflow_list,
1579					struct io_overflow_cqe, list);
1580		if (cqe)
1581			memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1582		else
1583			io_account_cq_overflow(ctx);
1584
1585		posted = true;
1586		list_del(&ocqe->list);
1587		kfree(ocqe);
1588	}
1589
1590	all_flushed = list_empty(&ctx->cq_overflow_list);
1591	if (all_flushed) {
1592		clear_bit(0, &ctx->check_cq_overflow);
1593		WRITE_ONCE(ctx->rings->sq_flags,
1594			   ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1595	}
1596
1597	if (posted)
1598		io_commit_cqring(ctx);
1599	spin_unlock(&ctx->completion_lock);
1600	if (posted)
1601		io_cqring_ev_posted(ctx);
1602	return all_flushed;
1603}
1604
1605static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1606{
1607	bool ret = true;
1608
1609	if (test_bit(0, &ctx->check_cq_overflow)) {
1610		/* iopoll syncs against uring_lock, not completion_lock */
1611		if (ctx->flags & IORING_SETUP_IOPOLL)
1612			mutex_lock(&ctx->uring_lock);
1613		ret = __io_cqring_overflow_flush(ctx, false);
1614		if (ctx->flags & IORING_SETUP_IOPOLL)
1615			mutex_unlock(&ctx->uring_lock);
1616	}
1617
1618	return ret;
1619}
1620
1621/* must to be called somewhat shortly after putting a request */
1622static inline void io_put_task(struct task_struct *task, int nr)
1623{
1624	struct io_uring_task *tctx = task->io_uring;
1625
1626	if (likely(task == current)) {
1627		tctx->cached_refs += nr;
1628	} else {
1629		percpu_counter_sub(&tctx->inflight, nr);
1630		if (unlikely(atomic_read(&tctx->in_idle)))
1631			wake_up(&tctx->wait);
1632		put_task_struct_many(task, nr);
1633	}
1634}
1635
1636static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1637				     long res, unsigned int cflags)
1638{
1639	struct io_overflow_cqe *ocqe;
1640
1641	ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1642	if (!ocqe) {
1643		/*
1644		 * If we're in ring overflow flush mode, or in task cancel mode,
1645		 * or cannot allocate an overflow entry, then we need to drop it
1646		 * on the floor.
1647		 */
1648		io_account_cq_overflow(ctx);
1649		return false;
1650	}
1651	if (list_empty(&ctx->cq_overflow_list)) {
1652		set_bit(0, &ctx->check_cq_overflow);
1653		WRITE_ONCE(ctx->rings->sq_flags,
1654			   ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1655
1656	}
1657	ocqe->cqe.user_data = user_data;
1658	ocqe->cqe.res = res;
1659	ocqe->cqe.flags = cflags;
1660	list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1661	return true;
1662}
1663
1664static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1665					  long res, unsigned int cflags)
1666{
1667	struct io_uring_cqe *cqe;
1668
1669	trace_io_uring_complete(ctx, user_data, res, cflags);
1670
1671	/*
1672	 * If we can't get a cq entry, userspace overflowed the
1673	 * submission (by quite a lot). Increment the overflow count in
1674	 * the ring.
1675	 */
1676	cqe = io_get_cqe(ctx);
1677	if (likely(cqe)) {
1678		WRITE_ONCE(cqe->user_data, user_data);
1679		WRITE_ONCE(cqe->res, res);
1680		WRITE_ONCE(cqe->flags, cflags);
1681		return true;
1682	}
1683	return io_cqring_event_overflow(ctx, user_data, res, cflags);
1684}
1685
1686/* not as hot to bloat with inlining */
1687static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1688					  long res, unsigned int cflags)
1689{
1690	return __io_cqring_fill_event(ctx, user_data, res, cflags);
1691}
1692
1693static void io_req_complete_post(struct io_kiocb *req, long res,
1694				 unsigned int cflags)
1695{
1696	struct io_ring_ctx *ctx = req->ctx;
1697
1698	spin_lock(&ctx->completion_lock);
1699	__io_cqring_fill_event(ctx, req->user_data, res, cflags);
1700	/*
1701	 * If we're the last reference to this request, add to our locked
1702	 * free_list cache.
1703	 */
1704	if (req_ref_put_and_test(req)) {
1705		if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1706			if (req->flags & IO_DISARM_MASK)
1707				io_disarm_next(req);
1708			if (req->link) {
1709				io_req_task_queue(req->link);
1710				req->link = NULL;
1711			}
1712		}
1713		io_dismantle_req(req);
1714		io_put_task(req->task, 1);
1715		list_add(&req->inflight_entry, &ctx->locked_free_list);
1716		ctx->locked_free_nr++;
1717	} else {
1718		if (!percpu_ref_tryget(&ctx->refs))
1719			req = NULL;
1720	}
1721	io_commit_cqring(ctx);
1722	spin_unlock(&ctx->completion_lock);
1723
1724	if (req) {
1725		io_cqring_ev_posted(ctx);
1726		percpu_ref_put(&ctx->refs);
1727	}
1728}
1729
1730static inline bool io_req_needs_clean(struct io_kiocb *req)
1731{
1732	return req->flags & IO_REQ_CLEAN_FLAGS;
1733}
1734
1735static void io_req_complete_state(struct io_kiocb *req, long res,
1736				  unsigned int cflags)
1737{
1738	if (io_req_needs_clean(req))
1739		io_clean_op(req);
1740	req->result = res;
1741	req->compl.cflags = cflags;
1742	req->flags |= REQ_F_COMPLETE_INLINE;
1743}
1744
1745static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1746				     long res, unsigned cflags)
1747{
1748	if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1749		io_req_complete_state(req, res, cflags);
1750	else
1751		io_req_complete_post(req, res, cflags);
1752}
1753
1754static inline void io_req_complete(struct io_kiocb *req, long res)
1755{
1756	__io_req_complete(req, 0, res, 0);
1757}
1758
1759static void io_req_complete_failed(struct io_kiocb *req, long res)
1760{
1761	req_set_fail(req);
1762	io_req_complete_post(req, res, 0);
1763}
1764
1765/*
1766 * Don't initialise the fields below on every allocation, but do that in
1767 * advance and keep them valid across allocations.
1768 */
1769static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1770{
1771	req->ctx = ctx;
1772	req->link = NULL;
1773	req->async_data = NULL;
1774	/* not necessary, but safer to zero */
1775	req->result = 0;
1776}
1777
1778static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1779					struct io_submit_state *state)
1780{
1781	spin_lock(&ctx->completion_lock);
1782	list_splice_init(&ctx->locked_free_list, &state->free_list);
1783	ctx->locked_free_nr = 0;
1784	spin_unlock(&ctx->completion_lock);
1785}
1786
1787/* Returns true IFF there are requests in the cache */
1788static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1789{
1790	struct io_submit_state *state = &ctx->submit_state;
1791	int nr;
1792
1793	/*
1794	 * If we have more than a batch's worth of requests in our IRQ side
1795	 * locked cache, grab the lock and move them over to our submission
1796	 * side cache.
1797	 */
1798	if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1799		io_flush_cached_locked_reqs(ctx, state);
1800
1801	nr = state->free_reqs;
1802	while (!list_empty(&state->free_list)) {
1803		struct io_kiocb *req = list_first_entry(&state->free_list,
1804					struct io_kiocb, inflight_entry);
1805
1806		list_del(&req->inflight_entry);
1807		state->reqs[nr++] = req;
1808		if (nr == ARRAY_SIZE(state->reqs))
1809			break;
1810	}
1811
1812	state->free_reqs = nr;
1813	return nr != 0;
1814}
1815
1816/*
1817 * A request might get retired back into the request caches even before opcode
1818 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1819 * Because of that, io_alloc_req() should be called only under ->uring_lock
1820 * and with extra caution to not get a request that is still worked on.
1821 */
1822static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1823	__must_hold(&ctx->uring_lock)
1824{
1825	struct io_submit_state *state = &ctx->submit_state;
1826	gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1827	int ret, i;
1828
1829	BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1830
1831	if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1832		goto got_req;
1833
1834	ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1835				    state->reqs);
1836
1837	/*
1838	 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1839	 * retry single alloc to be on the safe side.
1840	 */
1841	if (unlikely(ret <= 0)) {
1842		state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1843		if (!state->reqs[0])
1844			return NULL;
1845		ret = 1;
1846	}
1847
1848	for (i = 0; i < ret; i++)
1849		io_preinit_req(state->reqs[i], ctx);
1850	state->free_reqs = ret;
1851got_req:
1852	state->free_reqs--;
1853	return state->reqs[state->free_reqs];
1854}
1855
1856static inline void io_put_file(struct file *file)
1857{
1858	if (file)
1859		fput(file);
1860}
1861
1862static void io_dismantle_req(struct io_kiocb *req)
1863{
1864	unsigned int flags = req->flags;
1865
1866	if (io_req_needs_clean(req))
1867		io_clean_op(req);
1868	if (!(flags & REQ_F_FIXED_FILE))
1869		io_put_file(req->file);
1870	if (req->fixed_rsrc_refs)
1871		percpu_ref_put(req->fixed_rsrc_refs);
1872	if (req->async_data) {
1873		kfree(req->async_data);
1874		req->async_data = NULL;
1875	}
1876}
1877
1878static void __io_free_req(struct io_kiocb *req)
1879{
1880	struct io_ring_ctx *ctx = req->ctx;
1881
1882	io_dismantle_req(req);
1883	io_put_task(req->task, 1);
1884
1885	spin_lock(&ctx->completion_lock);
1886	list_add(&req->inflight_entry, &ctx->locked_free_list);
1887	ctx->locked_free_nr++;
1888	spin_unlock(&ctx->completion_lock);
1889
1890	percpu_ref_put(&ctx->refs);
1891}
1892
1893static inline void io_remove_next_linked(struct io_kiocb *req)
1894{
1895	struct io_kiocb *nxt = req->link;
1896
1897	req->link = nxt->link;
1898	nxt->link = NULL;
1899}
1900
1901static bool io_kill_linked_timeout(struct io_kiocb *req)
1902	__must_hold(&req->ctx->completion_lock)
1903	__must_hold(&req->ctx->timeout_lock)
1904{
1905	struct io_kiocb *link = req->link;
1906
1907	if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
1908		struct io_timeout_data *io = link->async_data;
1909
1910		io_remove_next_linked(req);
1911		link->timeout.head = NULL;
1912		if (hrtimer_try_to_cancel(&io->timer) != -1) {
1913			io_cqring_fill_event(link->ctx, link->user_data,
1914					     -ECANCELED, 0);
1915			io_put_req_deferred(link);
1916			return true;
1917		}
1918	}
1919	return false;
1920}
1921
1922static void io_fail_links(struct io_kiocb *req)
1923	__must_hold(&req->ctx->completion_lock)
1924{
1925	struct io_kiocb *nxt, *link = req->link;
1926
1927	req->link = NULL;
1928	while (link) {
1929		nxt = link->link;
1930		link->link = NULL;
1931
1932		trace_io_uring_fail_link(req, link);
1933		io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1934		io_put_req_deferred(link);
1935		link = nxt;
1936	}
1937}
1938
1939static bool io_disarm_next(struct io_kiocb *req)
1940	__must_hold(&req->ctx->completion_lock)
1941{
1942	bool posted = false;
1943
1944	if (req->flags & REQ_F_ARM_LTIMEOUT) {
1945		struct io_kiocb *link = req->link;
1946
1947		req->flags &= ~REQ_F_ARM_LTIMEOUT;
1948		if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
1949			io_remove_next_linked(req);
1950			io_cqring_fill_event(link->ctx, link->user_data,
1951					     -ECANCELED, 0);
1952			io_put_req_deferred(link);
1953			posted = true;
1954		}
1955	} else if (req->flags & REQ_F_LINK_TIMEOUT) {
1956		struct io_ring_ctx *ctx = req->ctx;
1957
1958		spin_lock_irq(&ctx->timeout_lock);
1959		posted = io_kill_linked_timeout(req);
1960		spin_unlock_irq(&ctx->timeout_lock);
1961	}
1962	if (unlikely((req->flags & REQ_F_FAIL) &&
1963		     !(req->flags & REQ_F_HARDLINK))) {
1964		posted |= (req->link != NULL);
1965		io_fail_links(req);
1966	}
1967	return posted;
1968}
1969
1970static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1971{
1972	struct io_kiocb *nxt;
1973
1974	/*
1975	 * If LINK is set, we have dependent requests in this chain. If we
1976	 * didn't fail this request, queue the first one up, moving any other
1977	 * dependencies to the next request. In case of failure, fail the rest
1978	 * of the chain.
1979	 */
1980	if (req->flags & IO_DISARM_MASK) {
1981		struct io_ring_ctx *ctx = req->ctx;
1982		bool posted;
1983
1984		spin_lock(&ctx->completion_lock);
1985		posted = io_disarm_next(req);
1986		if (posted)
1987			io_commit_cqring(req->ctx);
1988		spin_unlock(&ctx->completion_lock);
1989		if (posted)
1990			io_cqring_ev_posted(ctx);
1991	}
1992	nxt = req->link;
1993	req->link = NULL;
1994	return nxt;
1995}
1996
1997static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1998{
1999	if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2000		return NULL;
2001	return __io_req_find_next(req);
2002}
2003
2004static void ctx_flush_and_put(struct io_ring_ctx *ctx)
2005{
2006	if (!ctx)
2007		return;
2008	if (ctx->submit_state.compl_nr) {
2009		mutex_lock(&ctx->uring_lock);
2010		io_submit_flush_completions(ctx);
2011		mutex_unlock(&ctx->uring_lock);
2012	}
2013	percpu_ref_put(&ctx->refs);
2014}
2015
2016static void tctx_task_work(struct callback_head *cb)
2017{
2018	struct io_ring_ctx *ctx = NULL;
2019	struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2020						  task_work);
2021
2022	while (1) {
2023		struct io_wq_work_node *node;
2024
2025		spin_lock_irq(&tctx->task_lock);
2026		node = tctx->task_list.first;
2027		INIT_WQ_LIST(&tctx->task_list);
2028		if (!node)
2029			tctx->task_running = false;
2030		spin_unlock_irq(&tctx->task_lock);
2031		if (!node)
2032			break;
2033
2034		do {
2035			struct io_wq_work_node *next = node->next;
2036			struct io_kiocb *req = container_of(node, struct io_kiocb,
2037							    io_task_work.node);
2038
2039			if (req->ctx != ctx) {
2040				ctx_flush_and_put(ctx);
2041				ctx = req->ctx;
2042				percpu_ref_get(&ctx->refs);
2043			}
2044			req->io_task_work.func(req);
2045			node = next;
2046		} while (node);
2047
2048		cond_resched();
2049	}
2050
2051	ctx_flush_and_put(ctx);
2052}
2053
2054static void io_req_task_work_add(struct io_kiocb *req)
2055{
2056	struct task_struct *tsk = req->task;
2057	struct io_uring_task *tctx = tsk->io_uring;
2058	enum task_work_notify_mode notify;
2059	struct io_wq_work_node *node;
2060	unsigned long flags;
2061	bool running;
2062
2063	WARN_ON_ONCE(!tctx);
2064
2065	spin_lock_irqsave(&tctx->task_lock, flags);
2066	wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2067	running = tctx->task_running;
2068	if (!running)
2069		tctx->task_running = true;
2070	spin_unlock_irqrestore(&tctx->task_lock, flags);
2071
2072	/* task_work already pending, we're done */
2073	if (running)
2074		return;
2075
2076	/*
2077	 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2078	 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2079	 * processing task_work. There's no reliable way to tell if TWA_RESUME
2080	 * will do the job.
2081	 */
2082	notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2083	if (!task_work_add(tsk, &tctx->task_work, notify)) {
2084		wake_up_process(tsk);
2085		return;
2086	}
2087
2088	spin_lock_irqsave(&tctx->task_lock, flags);
2089	tctx->task_running = false;
2090	node = tctx->task_list.first;
2091	INIT_WQ_LIST(&tctx->task_list);
2092	spin_unlock_irqrestore(&tctx->task_lock, flags);
2093
2094	while (node) {
2095		req = container_of(node, struct io_kiocb, io_task_work.node);
2096		node = node->next;
2097		if (llist_add(&req->io_task_work.fallback_node,
2098			      &req->ctx->fallback_llist))
2099			schedule_delayed_work(&req->ctx->fallback_work, 1);
2100	}
2101}
2102
2103static void io_req_task_cancel(struct io_kiocb *req)
2104{
2105	struct io_ring_ctx *ctx = req->ctx;
2106
2107	/* ctx is guaranteed to stay alive while we hold uring_lock */
2108	mutex_lock(&ctx->uring_lock);
2109	io_req_complete_failed(req, req->result);
2110	mutex_unlock(&ctx->uring_lock);
2111}
2112
2113static void io_req_task_submit(struct io_kiocb *req)
2114{
2115	struct io_ring_ctx *ctx = req->ctx;
2116
2117	/* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2118	mutex_lock(&ctx->uring_lock);
2119	/* req->task == current here, checking PF_EXITING is safe */
2120	if (likely(!(req->task->flags & PF_EXITING)))
2121		__io_queue_sqe(req);
2122	else
2123		io_req_complete_failed(req, -EFAULT);
2124	mutex_unlock(&ctx->uring_lock);
2125}
2126
2127static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2128{
2129	req->result = ret;
2130	req->io_task_work.func = io_req_task_cancel;
2131	io_req_task_work_add(req);
2132}
2133
2134static void io_req_task_queue(struct io_kiocb *req)
2135{
2136	req->io_task_work.func = io_req_task_submit;
2137	io_req_task_work_add(req);
2138}
2139
2140static void io_req_task_queue_reissue(struct io_kiocb *req)
2141{
2142	req->io_task_work.func = io_queue_async_work;
2143	io_req_task_work_add(req);
2144}
2145
2146static inline void io_queue_next(struct io_kiocb *req)
2147{
2148	struct io_kiocb *nxt = io_req_find_next(req);
2149
2150	if (nxt)
2151		io_req_task_queue(nxt);
2152}
2153
2154static void io_free_req(struct io_kiocb *req)
2155{
2156	io_queue_next(req);
2157	__io_free_req(req);
2158}
2159
2160struct req_batch {
2161	struct task_struct	*task;
2162	int			task_refs;
2163	int			ctx_refs;
2164};
2165
2166static inline void io_init_req_batch(struct req_batch *rb)
2167{
2168	rb->task_refs = 0;
2169	rb->ctx_refs = 0;
2170	rb->task = NULL;
2171}
2172
2173static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2174				     struct req_batch *rb)
2175{
2176	if (rb->ctx_refs)
2177		percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2178	if (rb->task)
2179		io_put_task(rb->task, rb->task_refs);
2180}
2181
2182static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2183			      struct io_submit_state *state)
2184{
2185	io_queue_next(req);
2186	io_dismantle_req(req);
2187
2188	if (req->task != rb->task) {
2189		if (rb->task)
2190			io_put_task(rb->task, rb->task_refs);
2191		rb->task = req->task;
2192		rb->task_refs = 0;
2193	}
2194	rb->task_refs++;
2195	rb->ctx_refs++;
2196
2197	if (state->free_reqs != ARRAY_SIZE(state->reqs))
2198		state->reqs[state->free_reqs++] = req;
2199	else
2200		list_add(&req->inflight_entry, &state->free_list);
2201}
2202
2203static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2204	__must_hold(&ctx->uring_lock)
2205{
2206	struct io_submit_state *state = &ctx->submit_state;
2207	int i, nr = state->compl_nr;
2208	struct req_batch rb;
2209
2210	spin_lock(&ctx->completion_lock);
2211	for (i = 0; i < nr; i++) {
2212		struct io_kiocb *req = state->compl_reqs[i];
2213
2214		__io_cqring_fill_event(ctx, req->user_data, req->result,
2215					req->compl.cflags);
2216	}
2217	io_commit_cqring(ctx);
2218	spin_unlock(&ctx->completion_lock);
2219	io_cqring_ev_posted(ctx);
2220
2221	io_init_req_batch(&rb);
2222	for (i = 0; i < nr; i++) {
2223		struct io_kiocb *req = state->compl_reqs[i];
2224
2225		if (req_ref_put_and_test(req))
2226			io_req_free_batch(&rb, req, &ctx->submit_state);
2227	}
2228
2229	io_req_free_batch_finish(ctx, &rb);
2230	state->compl_nr = 0;
2231}
2232
2233/*
2234 * Drop reference to request, return next in chain (if there is one) if this
2235 * was the last reference to this request.
2236 */
2237static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2238{
2239	struct io_kiocb *nxt = NULL;
2240
2241	if (req_ref_put_and_test(req)) {
2242		nxt = io_req_find_next(req);
2243		__io_free_req(req);
2244	}
2245	return nxt;
2246}
2247
2248static inline void io_put_req(struct io_kiocb *req)
2249{
2250	if (req_ref_put_and_test(req))
2251		io_free_req(req);
2252}
2253
2254static inline void io_put_req_deferred(struct io_kiocb *req)
2255{
2256	if (req_ref_put_and_test(req)) {
2257		req->io_task_work.func = io_free_req;
2258		io_req_task_work_add(req);
2259	}
2260}
2261
2262static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2263{
2264	/* See comment at the top of this file */
2265	smp_rmb();
2266	return __io_cqring_events(ctx);
2267}
2268
2269static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2270{
2271	struct io_rings *rings = ctx->rings;
2272
2273	/* make sure SQ entry isn't read before tail */
2274	return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2275}
2276
2277static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2278{
2279	unsigned int cflags;
2280
2281	cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2282	cflags |= IORING_CQE_F_BUFFER;
2283	req->flags &= ~REQ_F_BUFFER_SELECTED;
2284	kfree(kbuf);
2285	return cflags;
2286}
2287
2288static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2289{
2290	struct io_buffer *kbuf;
2291
2292	if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
2293		return 0;
2294	kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2295	return io_put_kbuf(req, kbuf);
2296}
2297
2298static inline bool io_run_task_work(void)
2299{
2300	if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2301		__set_current_state(TASK_RUNNING);
2302		tracehook_notify_signal();
2303		return true;
2304	}
2305
2306	return false;
2307}
2308
2309/*
2310 * Find and free completed poll iocbs
2311 */
2312static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2313			       struct list_head *done)
2314{
2315	struct req_batch rb;
2316	struct io_kiocb *req;
2317
2318	/* order with ->result store in io_complete_rw_iopoll() */
2319	smp_rmb();
2320
2321	io_init_req_batch(&rb);
2322	while (!list_empty(done)) {
2323		req = list_first_entry(done, struct io_kiocb, inflight_entry);
2324		list_del(&req->inflight_entry);
2325
2326		if (READ_ONCE(req->result) == -EAGAIN &&
2327		    !(req->flags & REQ_F_DONT_REISSUE)) {
2328			req->iopoll_completed = 0;
2329			io_req_task_queue_reissue(req);
2330			continue;
2331		}
2332
2333		__io_cqring_fill_event(ctx, req->user_data, req->result,
2334					io_put_rw_kbuf(req));
2335		(*nr_events)++;
2336
2337		if (req_ref_put_and_test(req))
2338			io_req_free_batch(&rb, req, &ctx->submit_state);
2339	}
2340
2341	io_commit_cqring(ctx);
2342	io_cqring_ev_posted_iopoll(ctx);
2343	io_req_free_batch_finish(ctx, &rb);
2344}
2345
2346static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2347			long min)
2348{
2349	struct io_kiocb *req, *tmp;
2350	LIST_HEAD(done);
2351	bool spin;
2352
2353	/*
2354	 * Only spin for completions if we don't have multiple devices hanging
2355	 * off our complete list, and we're under the requested amount.
2356	 */
2357	spin = !ctx->poll_multi_queue && *nr_events < min;
2358
2359	list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2360		struct kiocb *kiocb = &req->rw.kiocb;
2361		int ret;
2362
2363		/*
2364		 * Move completed and retryable entries to our local lists.
2365		 * If we find a request that requires polling, break out
2366		 * and complete those lists first, if we have entries there.
2367		 */
2368		if (READ_ONCE(req->iopoll_completed)) {
2369			list_move_tail(&req->inflight_entry, &done);
2370			continue;
2371		}
2372		if (!list_empty(&done))
2373			break;
2374
2375		ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2376		if (unlikely(ret < 0))
2377			return ret;
2378		else if (ret)
2379			spin = false;
2380
2381		/* iopoll may have completed current req */
2382		if (READ_ONCE(req->iopoll_completed))
2383			list_move_tail(&req->inflight_entry, &done);
2384	}
2385
2386	if (!list_empty(&done))
2387		io_iopoll_complete(ctx, nr_events, &done);
2388
2389	return 0;
2390}
2391
2392/*
2393 * We can't just wait for polled events to come to us, we have to actively
2394 * find and complete them.
2395 */
2396static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2397{
2398	if (!(ctx->flags & IORING_SETUP_IOPOLL))
2399		return;
2400
2401	mutex_lock(&ctx->uring_lock);
2402	while (!list_empty(&ctx->iopoll_list)) {
2403		unsigned int nr_events = 0;
2404
2405		io_do_iopoll(ctx, &nr_events, 0);
2406
2407		/* let it sleep and repeat later if can't complete a request */
2408		if (nr_events == 0)
2409			break;
2410		/*
2411		 * Ensure we allow local-to-the-cpu processing to take place,
2412		 * in this case we need to ensure that we reap all events.
2413		 * Also let task_work, etc. to progress by releasing the mutex
2414		 */
2415		if (need_resched()) {
2416			mutex_unlock(&ctx->uring_lock);
2417			cond_resched();
2418			mutex_lock(&ctx->uring_lock);
2419		}
2420	}
2421	mutex_unlock(&ctx->uring_lock);
2422}
2423
2424static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2425{
2426	unsigned int nr_events = 0;
2427	int ret = 0;
2428
2429	/*
2430	 * We disallow the app entering submit/complete with polling, but we
2431	 * still need to lock the ring to prevent racing with polled issue
2432	 * that got punted to a workqueue.
2433	 */
2434	mutex_lock(&ctx->uring_lock);
2435	/*
2436	 * Don't enter poll loop if we already have events pending.
2437	 * If we do, we can potentially be spinning for commands that
2438	 * already triggered a CQE (eg in error).
2439	 */
2440	if (test_bit(0, &ctx->check_cq_overflow))
2441		__io_cqring_overflow_flush(ctx, false);
2442	if (io_cqring_events(ctx))
2443		goto out;
2444	do {
2445		/*
2446		 * If a submit got punted to a workqueue, we can have the
2447		 * application entering polling for a command before it gets
2448		 * issued. That app will hold the uring_lock for the duration
2449		 * of the poll right here, so we need to take a breather every
2450		 * now and then to ensure that the issue has a chance to add
2451		 * the poll to the issued list. Otherwise we can spin here
2452		 * forever, while the workqueue is stuck trying to acquire the
2453		 * very same mutex.
2454		 */
2455		if (list_empty(&ctx->iopoll_list)) {
2456			u32 tail = ctx->cached_cq_tail;
2457
2458			mutex_unlock(&ctx->uring_lock);
2459			io_run_task_work();
2460			mutex_lock(&ctx->uring_lock);
2461
2462			/* some requests don't go through iopoll_list */
2463			if (tail != ctx->cached_cq_tail ||
2464			    list_empty(&ctx->iopoll_list))
2465				break;
2466		}
2467		ret = io_do_iopoll(ctx, &nr_events, min);
2468	} while (!ret && nr_events < min && !need_resched());
2469out:
2470	mutex_unlock(&ctx->uring_lock);
2471	return ret;
2472}
2473
2474static void kiocb_end_write(struct io_kiocb *req)
2475{
2476	/*
2477	 * Tell lockdep we inherited freeze protection from submission
2478	 * thread.
2479	 */
2480	if (req->flags & REQ_F_ISREG) {
2481		struct super_block *sb = file_inode(req->file)->i_sb;
2482
2483		__sb_writers_acquired(sb, SB_FREEZE_WRITE);
2484		sb_end_write(sb);
2485	}
2486}
2487
2488#ifdef CONFIG_BLOCK
2489static bool io_resubmit_prep(struct io_kiocb *req)
2490{
2491	struct io_async_rw *rw = req->async_data;
2492
2493	if (!rw)
2494		return !io_req_prep_async(req);
2495	/* may have left rw->iter inconsistent on -EIOCBQUEUED */
2496	iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2497	return true;
2498}
2499
2500static bool io_rw_should_reissue(struct io_kiocb *req)
2501{
2502	umode_t mode = file_inode(req->file)->i_mode;
2503	struct io_ring_ctx *ctx = req->ctx;
2504
2505	if (!S_ISBLK(mode) && !S_ISREG(mode))
2506		return false;
2507	if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2508	    !(ctx->flags & IORING_SETUP_IOPOLL)))
2509		return false;
2510	/*
2511	 * If ref is dying, we might be running poll reap from the exit work.
2512	 * Don't attempt to reissue from that path, just let it fail with
2513	 * -EAGAIN.
2514	 */
2515	if (percpu_ref_is_dying(&ctx->refs))
2516		return false;
2517	/*
2518	 * Play it safe and assume not safe to re-import and reissue if we're
2519	 * not in the original thread group (or in task context).
2520	 */
2521	if (!same_thread_group(req->task, current) || !in_task())
2522		return false;
2523	return true;
2524}
2525#else
2526static bool io_resubmit_prep(struct io_kiocb *req)
2527{
2528	return false;
2529}
2530static bool io_rw_should_reissue(struct io_kiocb *req)
2531{
2532	return false;
2533}
2534#endif
2535
2536static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2537{
2538	if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2539		kiocb_end_write(req);
2540	if (res != req->result) {
2541		if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2542		    io_rw_should_reissue(req)) {
2543			req->flags |= REQ_F_REISSUE;
2544			return true;
2545		}
2546		req_set_fail(req);
2547		req->result = res;
2548	}
2549	return false;
2550}
2551
2552static void io_req_task_complete(struct io_kiocb *req)
2553{
2554	__io_req_complete(req, 0, req->result, io_put_rw_kbuf(req));
2555}
2556
2557static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2558			     unsigned int issue_flags)
2559{
2560	if (__io_complete_rw_common(req, res))
2561		return;
2562	io_req_task_complete(req);
2563}
2564
2565static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2566{
2567	struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2568
2569	if (__io_complete_rw_common(req, res))
2570		return;
2571	req->result = res;
2572	req->io_task_work.func = io_req_task_complete;
2573	io_req_task_work_add(req);
2574}
2575
2576static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2577{
2578	struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2579
2580	if (kiocb->ki_flags & IOCB_WRITE)
2581		kiocb_end_write(req);
2582	if (unlikely(res != req->result)) {
2583		if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2584		    io_resubmit_prep(req))) {
2585			req_set_fail(req);
2586			req->flags |= REQ_F_DONT_REISSUE;
2587		}
2588	}
2589
2590	WRITE_ONCE(req->result, res);
2591	/* order with io_iopoll_complete() checking ->result */
2592	smp_wmb();
2593	WRITE_ONCE(req->iopoll_completed, 1);
2594}
2595
2596/*
2597 * After the iocb has been issued, it's safe to be found on the poll list.
2598 * Adding the kiocb to the list AFTER submission ensures that we don't
2599 * find it from a io_do_iopoll() thread before the issuer is done
2600 * accessing the kiocb cookie.
2601 */
2602static void io_iopoll_req_issued(struct io_kiocb *req)
2603{
2604	struct io_ring_ctx *ctx = req->ctx;
2605	const bool in_async = io_wq_current_is_worker();
2606
2607	/* workqueue context doesn't hold uring_lock, grab it now */
2608	if (unlikely(in_async))
2609		mutex_lock(&ctx->uring_lock);
2610
2611	/*
2612	 * Track whether we have multiple files in our lists. This will impact
2613	 * how we do polling eventually, not spinning if we're on potentially
2614	 * different devices.
2615	 */
2616	if (list_empty(&ctx->iopoll_list)) {
2617		ctx->poll_multi_queue = false;
2618	} else if (!ctx->poll_multi_queue) {
2619		struct io_kiocb *list_req;
2620		unsigned int queue_num0, queue_num1;
2621
2622		list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2623						inflight_entry);
2624
2625		if (list_req->file != req->file) {
2626			ctx->poll_multi_queue = true;
2627		} else {
2628			queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2629			queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2630			if (queue_num0 != queue_num1)
2631				ctx->poll_multi_queue = true;
2632		}
2633	}
2634
2635	/*
2636	 * For fast devices, IO may have already completed. If it has, add
2637	 * it to the front so we find it first.
2638	 */
2639	if (READ_ONCE(req->iopoll_completed))
2640		list_add(&req->inflight_entry, &ctx->iopoll_list);
2641	else
2642		list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2643
2644	if (unlikely(in_async)) {
2645		/*
2646		 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2647		 * in sq thread task context or in io worker task context. If
2648		 * current task context is sq thread, we don't need to check
2649		 * whether should wake up sq thread.
2650		 */
2651		if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2652		    wq_has_sleeper(&ctx->sq_data->wait))
2653			wake_up(&ctx->sq_data->wait);
2654
2655		mutex_unlock(&ctx->uring_lock);
2656	}
2657}
2658
2659static bool io_bdev_nowait(struct block_device *bdev)
2660{
2661	return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2662}
2663
2664/*
2665 * If we tracked the file through the SCM inflight mechanism, we could support
2666 * any file. For now, just ensure that anything potentially problematic is done
2667 * inline.
2668 */
2669static bool __io_file_supports_nowait(struct file *file, int rw)
2670{
2671	umode_t mode = file_inode(file)->i_mode;
2672
2673	if (S_ISBLK(mode)) {
2674		if (IS_ENABLED(CONFIG_BLOCK) &&
2675		    io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2676			return true;
2677		return false;
2678	}
2679	if (S_ISSOCK(mode))
2680		return true;
2681	if (S_ISREG(mode)) {
2682		if (IS_ENABLED(CONFIG_BLOCK) &&
2683		    io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2684		    file->f_op != &io_uring_fops)
2685			return true;
2686		return false;
2687	}
2688
2689	/* any ->read/write should understand O_NONBLOCK */
2690	if (file->f_flags & O_NONBLOCK)
2691		return true;
2692
2693	if (!(file->f_mode & FMODE_NOWAIT))
2694		return false;
2695
2696	if (rw == READ)
2697		return file->f_op->read_iter != NULL;
2698
2699	return file->f_op->write_iter != NULL;
2700}
2701
2702static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2703{
2704	if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2705		return true;
2706	else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2707		return true;
2708
2709	return __io_file_supports_nowait(req->file, rw);
2710}
2711
2712static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2713{
2714	struct io_ring_ctx *ctx = req->ctx;
2715	struct kiocb *kiocb = &req->rw.kiocb;
2716	struct file *file = req->file;
2717	unsigned ioprio;
2718	int ret;
2719
2720	if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2721		req->flags |= REQ_F_ISREG;
2722
2723	kiocb->ki_pos = READ_ONCE(sqe->off);
2724	if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2725		req->flags |= REQ_F_CUR_POS;
2726		kiocb->ki_pos = file->f_pos;
2727	}
2728	kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2729	kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2730	ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2731	if (unlikely(ret))
2732		return ret;
2733
2734	/* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2735	if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2736		req->flags |= REQ_F_NOWAIT;
2737
2738	ioprio = READ_ONCE(sqe->ioprio);
2739	if (ioprio) {
2740		ret = ioprio_check_cap(ioprio);
2741		if (ret)
2742			return ret;
2743
2744		kiocb->ki_ioprio = ioprio;
2745	} else
2746		kiocb->ki_ioprio = get_current_ioprio();
2747
2748	if (ctx->flags & IORING_SETUP_IOPOLL) {
2749		if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2750		    !kiocb->ki_filp->f_op->iopoll)
2751			return -EOPNOTSUPP;
2752
2753		kiocb->ki_flags |= IOCB_HIPRI;
2754		kiocb->ki_complete = io_complete_rw_iopoll;
2755		req->iopoll_completed = 0;
2756	} else {
2757		if (kiocb->ki_flags & IOCB_HIPRI)
2758			return -EINVAL;
2759		kiocb->ki_complete = io_complete_rw;
2760	}
2761
2762	if (req->opcode == IORING_OP_READ_FIXED ||
2763	    req->opcode == IORING_OP_WRITE_FIXED) {
2764		req->imu = NULL;
2765		io_req_set_rsrc_node(req);
2766	}
2767
2768	req->rw.addr = READ_ONCE(sqe->addr);
2769	req->rw.len = READ_ONCE(sqe->len);
2770	req->buf_index = READ_ONCE(sqe->buf_index);
2771	return 0;
2772}
2773
2774static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2775{
2776	switch (ret) {
2777	case -EIOCBQUEUED:
2778		break;
2779	case -ERESTARTSYS:
2780	case -ERESTARTNOINTR:
2781	case -ERESTARTNOHAND:
2782	case -ERESTART_RESTARTBLOCK:
2783		/*
2784		 * We can't just restart the syscall, since previously
2785		 * submitted sqes may already be in progress. Just fail this
2786		 * IO with EINTR.
2787		 */
2788		ret = -EINTR;
2789		fallthrough;
2790	default:
2791		kiocb->ki_complete(kiocb, ret, 0);
2792	}
2793}
2794
2795static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2796		       unsigned int issue_flags)
2797{
2798	struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2799	struct io_async_rw *io = req->async_data;
2800	bool check_reissue = kiocb->ki_complete == io_complete_rw;
2801
2802	/* add previously done IO, if any */
2803	if (io && io->bytes_done > 0) {
2804		if (ret < 0)
2805			ret = io->bytes_done;
2806		else
2807			ret += io->bytes_done;
2808	}
2809
2810	if (req->flags & REQ_F_CUR_POS)
2811		req->file->f_pos = kiocb->ki_pos;
2812	if (ret >= 0 && check_reissue)
2813		__io_complete_rw(req, ret, 0, issue_flags);
2814	else
2815		io_rw_done(kiocb, ret);
2816
2817	if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2818		req->flags &= ~REQ_F_REISSUE;
2819		if (io_resubmit_prep(req)) {
2820			io_req_task_queue_reissue(req);
2821		} else {
2822			req_set_fail(req);
2823			__io_req_complete(req, issue_flags, ret,
2824					  io_put_rw_kbuf(req));
2825		}
2826	}
2827}
2828
2829static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2830			     struct io_mapped_ubuf *imu)
2831{
2832	size_t len = req->rw.len;
2833	u64 buf_end, buf_addr = req->rw.addr;
2834	size_t offset;
2835
2836	if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2837		return -EFAULT;
2838	/* not inside the mapped region */
2839	if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2840		return -EFAULT;
2841
2842	/*
2843	 * May not be a start of buffer, set size appropriately
2844	 * and advance us to the beginning.
2845	 */
2846	offset = buf_addr - imu->ubuf;
2847	iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2848
2849	if (offset) {
2850		/*
2851		 * Don't use iov_iter_advance() here, as it's really slow for
2852		 * using the latter parts of a big fixed buffer - it iterates
2853		 * over each segment manually. We can cheat a bit here, because
2854		 * we know that:
2855		 *
2856		 * 1) it's a BVEC iter, we set it up
2857		 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2858		 *    first and last bvec
2859		 *
2860		 * So just find our index, and adjust the iterator afterwards.
2861		 * If the offset is within the first bvec (or the whole first
2862		 * bvec, just use iov_iter_advance(). This makes it easier
2863		 * since we can just skip the first segment, which may not
2864		 * be PAGE_SIZE aligned.
2865		 */
2866		const struct bio_vec *bvec = imu->bvec;
2867
2868		if (offset <= bvec->bv_len) {
2869			iov_iter_advance(iter, offset);
2870		} else {
2871			unsigned long seg_skip;
2872
2873			/* skip first vec */
2874			offset -= bvec->bv_len;
2875			seg_skip = 1 + (offset >> PAGE_SHIFT);
2876
2877			iter->bvec = bvec + seg_skip;
2878			iter->nr_segs -= seg_skip;
2879			iter->count -= bvec->bv_len + offset;
2880			iter->iov_offset = offset & ~PAGE_MASK;
2881		}
2882	}
2883
2884	return 0;
2885}
2886
2887static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2888{
2889	struct io_ring_ctx *ctx = req->ctx;
2890	struct io_mapped_ubuf *imu = req->imu;
2891	u16 index, buf_index = req->buf_index;
2892
2893	if (likely(!imu)) {
2894		if (unlikely(buf_index >= ctx->nr_user_bufs))
2895			return -EFAULT;
2896		index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2897		imu = READ_ONCE(ctx->user_bufs[index]);
2898		req->imu = imu;
2899	}
2900	return __io_import_fixed(req, rw, iter, imu);
2901}
2902
2903static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2904{
2905	if (needs_lock)
2906		mutex_unlock(&ctx->uring_lock);
2907}
2908
2909static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2910{
2911	/*
2912	 * "Normal" inline submissions always hold the uring_lock, since we
2913	 * grab it from the system call. Same is true for the SQPOLL offload.
2914	 * The only exception is when we've detached the request and issue it
2915	 * from an async worker thread, grab the lock for that case.
2916	 */
2917	if (needs_lock)
2918		mutex_lock(&ctx->uring_lock);
2919}
2920
2921static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2922					  int bgid, struct io_buffer *kbuf,
2923					  bool needs_lock)
2924{
2925	struct io_buffer *head;
2926
2927	if (req->flags & REQ_F_BUFFER_SELECTED)
2928		return kbuf;
2929
2930	io_ring_submit_lock(req->ctx, needs_lock);
2931
2932	lockdep_assert_held(&req->ctx->uring_lock);
2933
2934	head = xa_load(&req->ctx->io_buffers, bgid);
2935	if (head) {
2936		if (!list_empty(&head->list)) {
2937			kbuf = list_last_entry(&head->list, struct io_buffer,
2938							list);
2939			list_del(&kbuf->list);
2940		} else {
2941			kbuf = head;
2942			xa_erase(&req->ctx->io_buffers, bgid);
2943		}
2944		if (*len > kbuf->len)
2945			*len = kbuf->len;
2946	} else {
2947		kbuf = ERR_PTR(-ENOBUFS);
2948	}
2949
2950	io_ring_submit_unlock(req->ctx, needs_lock);
2951
2952	return kbuf;
2953}
2954
2955static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2956					bool needs_lock)
2957{
2958	struct io_buffer *kbuf;
2959	u16 bgid;
2960
2961	kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2962	bgid = req->buf_index;
2963	kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2964	if (IS_ERR(kbuf))
2965		return kbuf;
2966	req->rw.addr = (u64) (unsigned long) kbuf;
2967	req->flags |= REQ_F_BUFFER_SELECTED;
2968	return u64_to_user_ptr(kbuf->addr);
2969}
2970
2971#ifdef CONFIG_COMPAT
2972static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2973				bool needs_lock)
2974{
2975	struct compat_iovec __user *uiov;
2976	compat_ssize_t clen;
2977	void __user *buf;
2978	ssize_t len;
2979
2980	uiov = u64_to_user_ptr(req->rw.addr);
2981	if (!access_ok(uiov, sizeof(*uiov)))
2982		return -EFAULT;
2983	if (__get_user(clen, &uiov->iov_len))
2984		return -EFAULT;
2985	if (clen < 0)
2986		return -EINVAL;
2987
2988	len = clen;
2989	buf = io_rw_buffer_select(req, &len, needs_lock);
2990	if (IS_ERR(buf))
2991		return PTR_ERR(buf);
2992	iov[0].iov_base = buf;
2993	iov[0].iov_len = (compat_size_t) len;
2994	return 0;
2995}
2996#endif
2997
2998static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2999				      bool needs_lock)
3000{
3001	struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3002	void __user *buf;
3003	ssize_t len;
3004
3005	if (copy_from_user(iov, uiov, sizeof(*uiov)))
3006		return -EFAULT;
3007
3008	len = iov[0].iov_len;
3009	if (len < 0)
3010		return -EINVAL;
3011	buf = io_rw_buffer_select(req, &len, needs_lock);
3012	if (IS_ERR(buf))
3013		return PTR_ERR(buf);
3014	iov[0].iov_base = buf;
3015	iov[0].iov_len = len;
3016	return 0;
3017}
3018
3019static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3020				    bool needs_lock)
3021{
3022	if (req->flags & REQ_F_BUFFER_SELECTED) {
3023		struct io_buffer *kbuf;
3024
3025		kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3026		iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3027		iov[0].iov_len = kbuf->len;
3028		return 0;
3029	}
3030	if (req->rw.len != 1)
3031		return -EINVAL;
3032
3033#ifdef CONFIG_COMPAT
3034	if (req->ctx->compat)
3035		return io_compat_import(req, iov, needs_lock);
3036#endif
3037
3038	return __io_iov_buffer_select(req, iov, needs_lock);
3039}
3040
3041static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3042			   struct iov_iter *iter, bool needs_lock)
3043{
3044	void __user *buf = u64_to_user_ptr(req->rw.addr);
3045	size_t sqe_len = req->rw.len;
3046	u8 opcode = req->opcode;
3047	ssize_t ret;
3048
3049	if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3050		*iovec = NULL;
3051		return io_import_fixed(req, rw, iter);
3052	}
3053
3054	/* buffer index only valid with fixed read/write, or buffer select  */
3055	if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3056		return -EINVAL;
3057
3058	if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3059		if (req->flags & REQ_F_BUFFER_SELECT) {
3060			buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3061			if (IS_ERR(buf))
3062				return PTR_ERR(buf);
3063			req->rw.len = sqe_len;
3064		}
3065
3066		ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3067		*iovec = NULL;
3068		return ret;
3069	}
3070
3071	if (req->flags & REQ_F_BUFFER_SELECT) {
3072		ret = io_iov_buffer_select(req, *iovec, needs_lock);
3073		if (!ret)
3074			iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3075		*iovec = NULL;
3076		return ret;
3077	}
3078
3079	return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3080			      req->ctx->compat);
3081}
3082
3083static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3084{
3085	return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3086}
3087
3088/*
3089 * For files that don't have ->read_iter() and ->write_iter(), handle them
3090 * by looping over ->read() or ->write() manually.
3091 */
3092static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3093{
3094	struct kiocb *kiocb = &req->rw.kiocb;
3095	struct file *file = req->file;
3096	ssize_t ret = 0;
3097
3098	/*
3099	 * Don't support polled IO through this interface, and we can't
3100	 * support non-blocking either. For the latter, this just causes
3101	 * the kiocb to be handled from an async context.
3102	 */
3103	if (kiocb->ki_flags & IOCB_HIPRI)
3104		return -EOPNOTSUPP;
3105	if (kiocb->ki_flags & IOCB_NOWAIT)
3106		return -EAGAIN;
3107
3108	while (iov_iter_count(iter)) {
3109		struct iovec iovec;
3110		ssize_t nr;
3111
3112		if (!iov_iter_is_bvec(iter)) {
3113			iovec = iov_iter_iovec(iter);
3114		} else {
3115			iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3116			iovec.iov_len = req->rw.len;
3117		}
3118
3119		if (rw == READ) {
3120			nr = file->f_op->read(file, iovec.iov_base,
3121					      iovec.iov_len, io_kiocb_ppos(kiocb));
3122		} else {
3123			nr = file->f_op->write(file, iovec.iov_base,
3124					       iovec.iov_len, io_kiocb_ppos(kiocb));
3125		}
3126
3127		if (nr < 0) {
3128			if (!ret)
3129				ret = nr;
3130			break;
3131		}
3132		ret += nr;
3133		if (nr != iovec.iov_len)
3134			break;
3135		req->rw.len -= nr;
3136		req->rw.addr += nr;
3137		iov_iter_advance(iter, nr);
3138	}
3139
3140	return ret;
3141}
3142
3143static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3144			  const struct iovec *fast_iov, struct iov_iter *iter)
3145{
3146	struct io_async_rw *rw = req->async_data;
3147
3148	memcpy(&rw->iter, iter, sizeof(*iter));
3149	rw->free_iovec = iovec;
3150	rw->bytes_done = 0;
3151	/* can only be fixed buffers, no need to do anything */
3152	if (iov_iter_is_bvec(iter))
3153		return;
3154	if (!iovec) {
3155		unsigned iov_off = 0;
3156
3157		rw->iter.iov = rw->fast_iov;
3158		if (iter->iov != fast_iov) {
3159			iov_off = iter->iov - fast_iov;
3160			rw->iter.iov += iov_off;
3161		}
3162		if (rw->fast_iov != fast_iov)
3163			memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3164			       sizeof(struct iovec) * iter->nr_segs);
3165	} else {
3166		req->flags |= REQ_F_NEED_CLEANUP;
3167	}
3168}
3169
3170static inline int io_alloc_async_data(struct io_kiocb *req)
3171{
3172	WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3173	req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3174	return req->async_data == NULL;
3175}
3176
3177static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3178			     const struct iovec *fast_iov,
3179			     struct iov_iter *iter, bool force)
3180{
3181	if (!force && !io_op_defs[req->opcode].needs_async_setup)
3182		return 0;
3183	if (!req->async_data) {
3184		if (io_alloc_async_data(req)) {
3185			kfree(iovec);
3186			return -ENOMEM;
3187		}
3188
3189		io_req_map_rw(req, iovec, fast_iov, iter);
3190	}
3191	return 0;
3192}
3193
3194static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3195{
3196	struct io_async_rw *iorw = req->async_data;
3197	struct iovec *iov = iorw->fast_iov;
3198	int ret;
3199
3200	ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3201	if (unlikely(ret < 0))
3202		return ret;
3203
3204	iorw->bytes_done = 0;
3205	iorw->free_iovec = iov;
3206	if (iov)
3207		req->flags |= REQ_F_NEED_CLEANUP;
3208	return 0;
3209}
3210
3211static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3212{
3213	if (unlikely(!(req->file->f_mode & FMODE_READ)))
3214		return -EBADF;
3215	return io_prep_rw(req, sqe);
3216}
3217
3218/*
3219 * This is our waitqueue callback handler, registered through lock_page_async()
3220 * when we initially tried to do the IO with the iocb armed our waitqueue.
3221 * This gets called when the page is unlocked, and we generally expect that to
3222 * happen when the page IO is completed and the page is now uptodate. This will
3223 * queue a task_work based retry of the operation, attempting to copy the data
3224 * again. If the latter fails because the page was NOT uptodate, then we will
3225 * do a thread based blocking retry of the operation. That's the unexpected
3226 * slow path.
3227 */
3228static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3229			     int sync, void *arg)
3230{
3231	struct wait_page_queue *wpq;
3232	struct io_kiocb *req = wait->private;
3233	struct wait_page_key *key = arg;
3234
3235	wpq = container_of(wait, struct wait_page_queue, wait);
3236
3237	if (!wake_page_match(wpq, key))
3238		return 0;
3239
3240	req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3241	list_del_init(&wait->entry);
3242	io_req_task_queue(req);
3243	return 1;
3244}
3245
3246/*
3247 * This controls whether a given IO request should be armed for async page
3248 * based retry. If we return false here, the request is handed to the async
3249 * worker threads for retry. If we're doing buffered reads on a regular file,
3250 * we prepare a private wait_page_queue entry and retry the operation. This
3251 * will either succeed because the page is now uptodate and unlocked, or it
3252 * will register a callback when the page is unlocked at IO completion. Through
3253 * that callback, io_uring uses task_work to setup a retry of the operation.
3254 * That retry will attempt the buffered read again. The retry will generally
3255 * succeed, or in rare cases where it fails, we then fall back to using the
3256 * async worker threads for a blocking retry.
3257 */
3258static bool io_rw_should_retry(struct io_kiocb *req)
3259{
3260	struct io_async_rw *rw = req->async_data;
3261	struct wait_page_queue *wait = &rw->wpq;
3262	struct kiocb *kiocb = &req->rw.kiocb;
3263
3264	/* never retry for NOWAIT, we just complete with -EAGAIN */
3265	if (req->flags & REQ_F_NOWAIT)
3266		return false;
3267
3268	/* Only for buffered IO */
3269	if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3270		return false;
3271
3272	/*
3273	 * just use poll if we can, and don't attempt if the fs doesn't
3274	 * support callback based unlocks
3275	 */
3276	if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3277		return false;
3278
3279	wait->wait.func = io_async_buf_func;
3280	wait->wait.private = req;
3281	wait->wait.flags = 0;
3282	INIT_LIST_HEAD(&wait->wait.entry);
3283	kiocb->ki_flags |= IOCB_WAITQ;
3284	kiocb->ki_flags &= ~IOCB_NOWAIT;
3285	kiocb->ki_waitq = wait;
3286	return true;
3287}
3288
3289static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3290{
3291	if (req->file->f_op->read_iter)
3292		return call_read_iter(req->file, &req->rw.kiocb, iter);
3293	else if (req->file->f_op->read)
3294		return loop_rw_iter(READ, req, iter);
3295	else
3296		return -EINVAL;
3297}
3298
3299static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3300{
3301	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3302	struct kiocb *kiocb = &req->rw.kiocb;
3303	struct iov_iter __iter, *iter = &__iter;
3304	struct io_async_rw *rw = req->async_data;
3305	ssize_t io_size, ret, ret2;
3306	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3307
3308	if (rw) {
3309		iter = &rw->iter;
3310		iovec = NULL;
3311	} else {
3312		ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3313		if (ret < 0)
3314			return ret;
3315	}
3316	io_size = iov_iter_count(iter);
3317	req->result = io_size;
3318
3319	/* Ensure we clear previously set non-block flag */
3320	if (!force_nonblock)
3321		kiocb->ki_flags &= ~IOCB_NOWAIT;
3322	else
3323		kiocb->ki_flags |= IOCB_NOWAIT;
3324
3325	/* If the file doesn't support async, just async punt */
3326	if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3327		ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3328		return ret ?: -EAGAIN;
3329	}
3330
3331	ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3332	if (unlikely(ret)) {
3333		kfree(iovec);
3334		return ret;
3335	}
3336
3337	ret = io_iter_do_read(req, iter);
3338
3339	if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3340		req->flags &= ~REQ_F_REISSUE;
3341		/* IOPOLL retry should happen for io-wq threads */
3342		if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3343			goto done;
3344		/* no retry on NONBLOCK nor RWF_NOWAIT */
3345		if (req->flags & REQ_F_NOWAIT)
3346			goto done;
3347		/* some cases will consume bytes even on error returns */
3348		iov_iter_revert(iter, io_size - iov_iter_count(iter));
3349		ret = 0;
3350	} else if (ret == -EIOCBQUEUED) {
3351		goto out_free;
3352	} else if (ret <= 0 || ret == io_size || !force_nonblock ||
3353		   (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3354		/* read all, failed, already did sync or don't want to retry */
3355		goto done;
3356	}
3357
3358	ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3359	if (ret2)
3360		return ret2;
3361
3362	iovec = NULL;
3363	rw = req->async_data;
3364	/* now use our persistent iterator, if we aren't already */
3365	iter = &rw->iter;
3366
3367	do {
3368		io_size -= ret;
3369		rw->bytes_done += ret;
3370		/* if we can retry, do so with the callbacks armed */
3371		if (!io_rw_should_retry(req)) {
3372			kiocb->ki_flags &= ~IOCB_WAITQ;
3373			return -EAGAIN;
3374		}
3375
3376		/*
3377		 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3378		 * we get -EIOCBQUEUED, then we'll get a notification when the
3379		 * desired page gets unlocked. We can also get a partial read
3380		 * here, and if we do, then just retry at the new offset.
3381		 */
3382		ret = io_iter_do_read(req, iter);
3383		if (ret == -EIOCBQUEUED)
3384			return 0;
3385		/* we got some bytes, but not all. retry. */
3386		kiocb->ki_flags &= ~IOCB_WAITQ;
3387	} while (ret > 0 && ret < io_size);
3388done:
3389	kiocb_done(kiocb, ret, issue_flags);
3390out_free:
3391	/* it's faster to check here then delegate to kfree */
3392	if (iovec)
3393		kfree(iovec);
3394	return 0;
3395}
3396
3397static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3398{
3399	if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3400		return -EBADF;
3401	return io_prep_rw(req, sqe);
3402}
3403
3404static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3405{
3406	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3407	struct kiocb *kiocb = &req->rw.kiocb;
3408	struct iov_iter __iter, *iter = &__iter;
3409	struct io_async_rw *rw = req->async_data;
3410	ssize_t ret, ret2, io_size;
3411	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3412
3413	if (rw) {
3414		iter = &rw->iter;
3415		iovec = NULL;
3416	} else {
3417		ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3418		if (ret < 0)
3419			return ret;
3420	}
3421	io_size = iov_iter_count(iter);
3422	req->result = io_size;
3423
3424	/* Ensure we clear previously set non-block flag */
3425	if (!force_nonblock)
3426		kiocb->ki_flags &= ~IOCB_NOWAIT;
3427	else
3428		kiocb->ki_flags |= IOCB_NOWAIT;
3429
3430	/* If the file doesn't support async, just async punt */
3431	if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3432		goto copy_iov;
3433
3434	/* file path doesn't support NOWAIT for non-direct_IO */
3435	if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3436	    (req->flags & REQ_F_ISREG))
3437		goto copy_iov;
3438
3439	ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3440	if (unlikely(ret))
3441		goto out_free;
3442
3443	/*
3444	 * Open-code file_start_write here to grab freeze protection,
3445	 * which will be released by another thread in
3446	 * io_complete_rw().  Fool lockdep by telling it the lock got
3447	 * released so that it doesn't complain about the held lock when
3448	 * we return to userspace.
3449	 */
3450	if (req->flags & REQ_F_ISREG) {
3451		sb_start_write(file_inode(req->file)->i_sb);
3452		__sb_writers_release(file_inode(req->file)->i_sb,
3453					SB_FREEZE_WRITE);
3454	}
3455	kiocb->ki_flags |= IOCB_WRITE;
3456
3457	if (req->file->f_op->write_iter)
3458		ret2 = call_write_iter(req->file, kiocb, iter);
3459	else if (req->file->f_op->write)
3460		ret2 = loop_rw_iter(WRITE, req, iter);
3461	else
3462		ret2 = -EINVAL;
3463
3464	if (req->flags & REQ_F_REISSUE) {
3465		req->flags &= ~REQ_F_REISSUE;
3466		ret2 = -EAGAIN;
3467	}
3468
3469	/*
3470	 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3471	 * retry them without IOCB_NOWAIT.
3472	 */
3473	if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3474		ret2 = -EAGAIN;
3475	/* no retry on NONBLOCK nor RWF_NOWAIT */
3476	if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3477		goto done;
3478	if (!force_nonblock || ret2 != -EAGAIN) {
3479		/* IOPOLL retry should happen for io-wq threads */
3480		if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3481			goto copy_iov;
3482done:
3483		kiocb_done(kiocb, ret2, issue_flags);
3484	} else {
3485copy_iov:
3486		/* some cases will consume bytes even on error returns */
3487		iov_iter_revert(iter, io_size - iov_iter_count(iter));
3488		ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3489		return ret ?: -EAGAIN;
3490	}
3491out_free:
3492	/* it's reportedly faster than delegating the null check to kfree() */
3493	if (iovec)
3494		kfree(iovec);
3495	return ret;
3496}
3497
3498static int io_renameat_prep(struct io_kiocb *req,
3499			    const struct io_uring_sqe *sqe)
3500{
3501	struct io_rename *ren = &req->rename;
3502	const char __user *oldf, *newf;
3503
3504	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3505		return -EINVAL;
3506	if (sqe->ioprio || sqe->buf_index)
3507		return -EINVAL;
3508	if (unlikely(req->flags & REQ_F_FIXED_FILE))
3509		return -EBADF;
3510
3511	ren->old_dfd = READ_ONCE(sqe->fd);
3512	oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3513	newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3514	ren->new_dfd = READ_ONCE(sqe->len);
3515	ren->flags = READ_ONCE(sqe->rename_flags);
3516
3517	ren->oldpath = getname(oldf);
3518	if (IS_ERR(ren->oldpath))
3519		return PTR_ERR(ren->oldpath);
3520
3521	ren->newpath = getname(newf);
3522	if (IS_ERR(ren->newpath)) {
3523		putname(ren->oldpath);
3524		return PTR_ERR(ren->newpath);
3525	}
3526
3527	req->flags |= REQ_F_NEED_CLEANUP;
3528	return 0;
3529}
3530
3531static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3532{
3533	struct io_rename *ren = &req->rename;
3534	int ret;
3535
3536	if (issue_flags & IO_URING_F_NONBLOCK)
3537		return -EAGAIN;
3538
3539	ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3540				ren->newpath, ren->flags);
3541
3542	req->flags &= ~REQ_F_NEED_CLEANUP;
3543	if (ret < 0)
3544		req_set_fail(req);
3545	io_req_complete(req, ret);
3546	return 0;
3547}
3548
3549static int io_unlinkat_prep(struct io_kiocb *req,
3550			    const struct io_uring_sqe *sqe)
3551{
3552	struct io_unlink *un = &req->unlink;
3553	const char __user *fname;
3554
3555	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3556		return -EINVAL;
3557	if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3558		return -EINVAL;
3559	if (unlikely(req->flags & REQ_F_FIXED_FILE))
3560		return -EBADF;
3561
3562	un->dfd = READ_ONCE(sqe->fd);
3563
3564	un->flags = READ_ONCE(sqe->unlink_flags);
3565	if (un->flags & ~AT_REMOVEDIR)
3566		return -EINVAL;
3567
3568	fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3569	un->filename = getname(fname);
3570	if (IS_ERR(un->filename))
3571		return PTR_ERR(un->filename);
3572
3573	req->flags |= REQ_F_NEED_CLEANUP;
3574	return 0;
3575}
3576
3577static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3578{
3579	struct io_unlink *un = &req->unlink;
3580	int ret;
3581
3582	if (issue_flags & IO_URING_F_NONBLOCK)
3583		return -EAGAIN;
3584
3585	if (un->flags & AT_REMOVEDIR)
3586		ret = do_rmdir(un->dfd, un->filename);
3587	else
3588		ret = do_unlinkat(un->dfd, un->filename);
3589
3590	req->flags &= ~REQ_F_NEED_CLEANUP;
3591	if (ret < 0)
3592		req_set_fail(req);
3593	io_req_complete(req, ret);
3594	return 0;
3595}
3596
3597static int io_shutdown_prep(struct io_kiocb *req,
3598			    const struct io_uring_sqe *sqe)
3599{
3600#if defined(CONFIG_NET)
3601	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3602		return -EINVAL;
3603	if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3604	    sqe->buf_index)
3605		return -EINVAL;
3606
3607	req->shutdown.how = READ_ONCE(sqe->len);
3608	return 0;
3609#else
3610	return -EOPNOTSUPP;
3611#endif
3612}
3613
3614static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3615{
3616#if defined(CONFIG_NET)
3617	struct socket *sock;
3618	int ret;
3619
3620	if (issue_flags & IO_URING_F_NONBLOCK)
3621		return -EAGAIN;
3622
3623	sock = sock_from_file(req->file);
3624	if (unlikely(!sock))
3625		return -ENOTSOCK;
3626
3627	ret = __sys_shutdown_sock(sock, req->shutdown.how);
3628	if (ret < 0)
3629		req_set_fail(req);
3630	io_req_complete(req, ret);
3631	return 0;
3632#else
3633	return -EOPNOTSUPP;
3634#endif
3635}
3636
3637static int __io_splice_prep(struct io_kiocb *req,
3638			    const struct io_uring_sqe *sqe)
3639{
3640	struct io_splice *sp = &req->splice;
3641	unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3642
3643	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3644		return -EINVAL;
3645
3646	sp->file_in = NULL;
3647	sp->len = READ_ONCE(sqe->len);
3648	sp->flags = READ_ONCE(sqe->splice_flags);
3649
3650	if (unlikely(sp->flags & ~valid_flags))
3651		return -EINVAL;
3652
3653	sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3654				  (sp->flags & SPLICE_F_FD_IN_FIXED));
3655	if (!sp->file_in)
3656		return -EBADF;
3657	req->flags |= REQ_F_NEED_CLEANUP;
3658	return 0;
3659}
3660
3661static int io_tee_prep(struct io_kiocb *req,
3662		       const struct io_uring_sqe *sqe)
3663{
3664	if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3665		return -EINVAL;
3666	return __io_splice_prep(req, sqe);
3667}
3668
3669static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3670{
3671	struct io_splice *sp = &req->splice;
3672	struct file *in = sp->file_in;
3673	struct file *out = sp->file_out;
3674	unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3675	long ret = 0;
3676
3677	if (issue_flags & IO_URING_F_NONBLOCK)
3678		return -EAGAIN;
3679	if (sp->len)
3680		ret = do_tee(in, out, sp->len, flags);
3681
3682	if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3683		io_put_file(in);
3684	req->flags &= ~REQ_F_NEED_CLEANUP;
3685
3686	if (ret != sp->len)
3687		req_set_fail(req);
3688	io_req_complete(req, ret);
3689	return 0;
3690}
3691
3692static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3693{
3694	struct io_splice *sp = &req->splice;
3695
3696	sp->off_in = READ_ONCE(sqe->splice_off_in);
3697	sp->off_out = READ_ONCE(sqe->off);
3698	return __io_splice_prep(req, sqe);
3699}
3700
3701static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3702{
3703	struct io_splice *sp = &req->splice;
3704	struct file *in = sp->file_in;
3705	struct file *out = sp->file_out;
3706	unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3707	loff_t *poff_in, *poff_out;
3708	long ret = 0;
3709
3710	if (issue_flags & IO_URING_F_NONBLOCK)
3711		return -EAGAIN;
3712
3713	poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3714	poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3715
3716	if (sp->len)
3717		ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3718
3719	if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3720		io_put_file(in);
3721	req->flags &= ~REQ_F_NEED_CLEANUP;
3722
3723	if (ret != sp->len)
3724		req_set_fail(req);
3725	io_req_complete(req, ret);
3726	return 0;
3727}
3728
3729/*
3730 * IORING_OP_NOP just posts a completion event, nothing else.
3731 */
3732static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3733{
3734	struct io_ring_ctx *ctx = req->ctx;
3735
3736	if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3737		return -EINVAL;
3738
3739	__io_req_complete(req, issue_flags, 0, 0);
3740	return 0;
3741}
3742
3743static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3744{
3745	struct io_ring_ctx *ctx = req->ctx;
3746
3747	if (!req->file)
3748		return -EBADF;
3749
3750	if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3751		return -EINVAL;
3752	if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3753		return -EINVAL;
3754
3755	req->sync.flags = READ_ONCE(sqe->fsync_flags);
3756	if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3757		return -EINVAL;
3758
3759	req->sync.off = READ_ONCE(sqe->off);
3760	req->sync.len = READ_ONCE(sqe->len);
3761	return 0;
3762}
3763
3764static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3765{
3766	loff_t end = req->sync.off + req->sync.len;
3767	int ret;
3768
3769	/* fsync always requires a blocking context */
3770	if (issue_flags & IO_URING_F_NONBLOCK)
3771		return -EAGAIN;
3772
3773	ret = vfs_fsync_range(req->file, req->sync.off,
3774				end > 0 ? end : LLONG_MAX,
3775				req->sync.flags & IORING_FSYNC_DATASYNC);
3776	if (ret < 0)
3777		req_set_fail(req);
3778	io_req_complete(req, ret);
3779	return 0;
3780}
3781
3782static int io_fallocate_prep(struct io_kiocb *req,
3783			     const struct io_uring_sqe *sqe)
3784{
3785	if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3786		return -EINVAL;
3787	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3788		return -EINVAL;
3789
3790	req->sync.off = READ_ONCE(sqe->off);
3791	req->sync.len = READ_ONCE(sqe->addr);
3792	req->sync.mode = READ_ONCE(sqe->len);
3793	return 0;
3794}
3795
3796static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3797{
3798	int ret;
3799
3800	/* fallocate always requiring blocking context */
3801	if (issue_flags & IO_URING_F_NONBLOCK)
3802		return -EAGAIN;
3803	ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3804				req->sync.len);
3805	if (ret < 0)
3806		req_set_fail(req);
3807	io_req_complete(req, ret);
3808	return 0;
3809}
3810
3811static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3812{
3813	const char __user *fname;
3814	int ret;
3815
3816	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3817		return -EINVAL;
3818	if (unlikely(sqe->ioprio || sqe->buf_index))
3819		return -EINVAL;
3820	if (unlikely(req->flags & REQ_F_FIXED_FILE))
3821		return -EBADF;
3822
3823	/* open.how should be already initialised */
3824	if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3825		req->open.how.flags |= O_LARGEFILE;
3826
3827	req->open.dfd = READ_ONCE(sqe->fd);
3828	fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3829	req->open.filename = getname(fname);
3830	if (IS_ERR(req->open.filename)) {
3831		ret = PTR_ERR(req->open.filename);
3832		req->open.filename = NULL;
3833		return ret;
3834	}
3835	req->open.nofile = rlimit(RLIMIT_NOFILE);
3836	req->flags |= REQ_F_NEED_CLEANUP;
3837	return 0;
3838}
3839
3840static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3841{
3842	u64 mode = READ_ONCE(sqe->len);
3843	u64 flags = READ_ONCE(sqe->open_flags);
3844
3845	req->open.how = build_open_how(flags, mode);
3846	return __io_openat_prep(req, sqe);
3847}
3848
3849static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3850{
3851	struct open_how __user *how;
3852	size_t len;
3853	int ret;
3854
3855	how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3856	len = READ_ONCE(sqe->len);
3857	if (len < OPEN_HOW_SIZE_VER0)
3858		return -EINVAL;
3859
3860	ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3861					len);
3862	if (ret)
3863		return ret;
3864
3865	return __io_openat_prep(req, sqe);
3866}
3867
3868static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3869{
3870	struct open_flags op;
3871	struct file *file;
3872	bool nonblock_set;
3873	bool resolve_nonblock;
3874	int ret;
3875
3876	ret = build_open_flags(&req->open.how, &op);
3877	if (ret)
3878		goto err;
3879	nonblock_set = op.open_flag & O_NONBLOCK;
3880	resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3881	if (issue_flags & IO_URING_F_NONBLOCK) {
3882		/*
3883		 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3884		 * it'll always -EAGAIN
3885		 */
3886		if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3887			return -EAGAIN;
3888		op.lookup_flags |= LOOKUP_CACHED;
3889		op.open_flag |= O_NONBLOCK;
3890	}
3891
3892	ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3893	if (ret < 0)
3894		goto err;
3895
3896	file = do_filp_open(req->open.dfd, req->open.filename, &op);
3897	if (IS_ERR(file)) {
3898		/*
3899		 * We could hang on to this 'fd' on retrying, but seems like
3900		 * marginal gain for something that is now known to be a slower
3901		 * path. So just put it, and we'll get a new one when we retry.
3902		 */
3903		put_unused_fd(ret);
3904
3905		ret = PTR_ERR(file);
3906		/* only retry if RESOLVE_CACHED wasn't already set by application */
3907		if (ret == -EAGAIN &&
3908		    (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3909			return -EAGAIN;
3910		goto err;
3911	}
3912
3913	if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3914		file->f_flags &= ~O_NONBLOCK;
3915	fsnotify_open(file);
3916	fd_install(ret, file);
3917err:
3918	putname(req->open.filename);
3919	req->flags &= ~REQ_F_NEED_CLEANUP;
3920	if (ret < 0)
3921		req_set_fail(req);
3922	__io_req_complete(req, issue_flags, ret, 0);
3923	return 0;
3924}
3925
3926static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3927{
3928	return io_openat2(req, issue_flags);
3929}
3930
3931static int io_remove_buffers_prep(struct io_kiocb *req,
3932				  const struct io_uring_sqe *sqe)
3933{
3934	struct io_provide_buf *p = &req->pbuf;
3935	u64 tmp;
3936
3937	if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3938		return -EINVAL;
3939
3940	tmp = READ_ONCE(sqe->fd);
3941	if (!tmp || tmp > USHRT_MAX)
3942		return -EINVAL;
3943
3944	memset(p, 0, sizeof(*p));
3945	p->nbufs = tmp;
3946	p->bgid = READ_ONCE(sqe->buf_group);
3947	return 0;
3948}
3949
3950static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3951			       int bgid, unsigned nbufs)
3952{
3953	unsigned i = 0;
3954
3955	/* shouldn't happen */
3956	if (!nbufs)
3957		return 0;
3958
3959	/* the head kbuf is the list itself */
3960	while (!list_empty(&buf->list)) {
3961		struct io_buffer *nxt;
3962
3963		nxt = list_first_entry(&buf->list, struct io_buffer, list);
3964		list_del(&nxt->list);
3965		kfree(nxt);
3966		if (++i == nbufs)
3967			return i;
3968	}
3969	i++;
3970	kfree(buf);
3971	xa_erase(&ctx->io_buffers, bgid);
3972
3973	return i;
3974}
3975
3976static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3977{
3978	struct io_provide_buf *p = &req->pbuf;
3979	struct io_ring_ctx *ctx = req->ctx;
3980	struct io_buffer *head;
3981	int ret = 0;
3982	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3983
3984	io_ring_submit_lock(ctx, !force_nonblock);
3985
3986	lockdep_assert_held(&ctx->uring_lock);
3987
3988	ret = -ENOENT;
3989	head = xa_load(&ctx->io_buffers, p->bgid);
3990	if (head)
3991		ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3992	if (ret < 0)
3993		req_set_fail(req);
3994
3995	/* complete before unlock, IOPOLL may need the lock */
3996	__io_req_complete(req, issue_flags, ret, 0);
3997	io_ring_submit_unlock(ctx, !force_nonblock);
3998	return 0;
3999}
4000
4001static int io_provide_buffers_prep(struct io_kiocb *req,
4002				   const struct io_uring_sqe *sqe)
4003{
4004	unsigned long size, tmp_check;
4005	struct io_provide_buf *p = &req->pbuf;
4006	u64 tmp;
4007
4008	if (sqe->ioprio || sqe->rw_flags)
4009		return -EINVAL;
4010
4011	tmp = READ_ONCE(sqe->fd);
4012	if (!tmp || tmp > USHRT_MAX)
4013		return -E2BIG;
4014	p->nbufs = tmp;
4015	p->addr = READ_ONCE(sqe->addr);
4016	p->len = READ_ONCE(sqe->len);
4017
4018	if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4019				&size))
4020		return -EOVERFLOW;
4021	if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4022		return -EOVERFLOW;
4023
4024	size = (unsigned long)p->len * p->nbufs;
4025	if (!access_ok(u64_to_user_ptr(p->addr), size))
4026		return -EFAULT;
4027
4028	p->bgid = READ_ONCE(sqe->buf_group);
4029	tmp = READ_ONCE(sqe->off);
4030	if (tmp > USHRT_MAX)
4031		return -E2BIG;
4032	p->bid = tmp;
4033	return 0;
4034}
4035
4036static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4037{
4038	struct io_buffer *buf;
4039	u64 addr = pbuf->addr;
4040	int i, bid = pbuf->bid;
4041
4042	for (i = 0; i < pbuf->nbufs; i++) {
4043		buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4044		if (!buf)
4045			break;
4046
4047		buf->addr = addr;
4048		buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4049		buf->bid = bid;
4050		addr += pbuf->len;
4051		bid++;
4052		if (!*head) {
4053			INIT_LIST_HEAD(&buf->list);
4054			*head = buf;
4055		} else {
4056			list_add_tail(&buf->list, &(*head)->list);
4057		}
4058	}
4059
4060	return i ? i : -ENOMEM;
4061}
4062
4063static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4064{
4065	struct io_provide_buf *p = &req->pbuf;
4066	struct io_ring_ctx *ctx = req->ctx;
4067	struct io_buffer *head, *list;
4068	int ret = 0;
4069	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4070
4071	io_ring_submit_lock(ctx, !force_nonblock);
4072
4073	lockdep_assert_held(&ctx->uring_lock);
4074
4075	list = head = xa_load(&ctx->io_buffers, p->bgid);
4076
4077	ret = io_add_buffers(p, &head);
4078	if (ret >= 0 && !list) {
4079		ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4080		if (ret < 0)
4081			__io_remove_buffers(ctx, head, p->bgid, -1U);
4082	}
4083	if (ret < 0)
4084		req_set_fail(req);
4085	/* complete before unlock, IOPOLL may need the lock */
4086	__io_req_complete(req, issue_flags, ret, 0);
4087	io_ring_submit_unlock(ctx, !force_nonblock);
4088	return 0;
4089}
4090
4091static int io_epoll_ctl_prep(struct io_kiocb *req,
4092			     const struct io_uring_sqe *sqe)
4093{
4094#if defined(CONFIG_EPOLL)
4095	if (sqe->ioprio || sqe->buf_index)
4096		return -EINVAL;
4097	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4098		return -EINVAL;
4099
4100	req->epoll.epfd = READ_ONCE(sqe->fd);
4101	req->epoll.op = READ_ONCE(sqe->len);
4102	req->epoll.fd = READ_ONCE(sqe->off);
4103
4104	if (ep_op_has_event(req->epoll.op)) {
4105		struct epoll_event __user *ev;
4106
4107		ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4108		if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4109			return -EFAULT;
4110	}
4111
4112	return 0;
4113#else
4114	return -EOPNOTSUPP;
4115#endif
4116}
4117
4118static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4119{
4120#if defined(CONFIG_EPOLL)
4121	struct io_epoll *ie = &req->epoll;
4122	int ret;
4123	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4124
4125	ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4126	if (force_nonblock && ret == -EAGAIN)
4127		return -EAGAIN;
4128
4129	if (ret < 0)
4130		req_set_fail(req);
4131	__io_req_complete(req, issue_flags, ret, 0);
4132	return 0;
4133#else
4134	return -EOPNOTSUPP;
4135#endif
4136}
4137
4138static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4139{
4140#if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4141	if (sqe->ioprio || sqe->buf_index || sqe->off)
4142		return -EINVAL;
4143	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4144		return -EINVAL;
4145
4146	req->madvise.addr = READ_ONCE(sqe->addr);
4147	req->madvise.len = READ_ONCE(sqe->len);
4148	req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4149	return 0;
4150#else
4151	return -EOPNOTSUPP;
4152#endif
4153}
4154
4155static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4156{
4157#if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4158	struct io_madvise *ma = &req->madvise;
4159	int ret;
4160
4161	if (issue_flags & IO_URING_F_NONBLOCK)
4162		return -EAGAIN;
4163
4164	ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4165	if (ret < 0)
4166		req_set_fail(req);
4167	io_req_complete(req, ret);
4168	return 0;
4169#else
4170	return -EOPNOTSUPP;
4171#endif
4172}
4173
4174static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4175{
4176	if (sqe->ioprio || sqe->buf_index || sqe->addr)
4177		return -EINVAL;
4178	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4179		return -EINVAL;
4180
4181	req->fadvise.offset = READ_ONCE(sqe->off);
4182	req->fadvise.len = READ_ONCE(sqe->len);
4183	req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4184	return 0;
4185}
4186
4187static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4188{
4189	struct io_fadvise *fa = &req->fadvise;
4190	int ret;
4191
4192	if (issue_flags & IO_URING_F_NONBLOCK) {
4193		switch (fa->advice) {
4194		case POSIX_FADV_NORMAL:
4195		case POSIX_FADV_RANDOM:
4196		case POSIX_FADV_SEQUENTIAL:
4197			break;
4198		default:
4199			return -EAGAIN;
4200		}
4201	}
4202
4203	ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4204	if (ret < 0)
4205		req_set_fail(req);
4206	__io_req_complete(req, issue_flags, ret, 0);
4207	return 0;
4208}
4209
4210static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4211{
4212	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4213		return -EINVAL;
4214	if (sqe->ioprio || sqe->buf_index)
4215		return -EINVAL;
4216	if (req->flags & REQ_F_FIXED_FILE)
4217		return -EBADF;
4218
4219	req->statx.dfd = READ_ONCE(sqe->fd);
4220	req->statx.mask = READ_ONCE(sqe->len);
4221	req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4222	req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4223	req->statx.flags = READ_ONCE(sqe->statx_flags);
4224
4225	return 0;
4226}
4227
4228static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4229{
4230	struct io_statx *ctx = &req->statx;
4231	int ret;
4232
4233	if (issue_flags & IO_URING_F_NONBLOCK)
4234		return -EAGAIN;
4235
4236	ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4237		       ctx->buffer);
4238
4239	if (ret < 0)
4240		req_set_fail(req);
4241	io_req_complete(req, ret);
4242	return 0;
4243}
4244
4245static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4246{
4247	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4248		return -EINVAL;
4249	if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4250	    sqe->rw_flags || sqe->buf_index)
4251		return -EINVAL;
4252	if (req->flags & REQ_F_FIXED_FILE)
4253		return -EBADF;
4254
4255	req->close.fd = READ_ONCE(sqe->fd);
4256	return 0;
4257}
4258
4259static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4260{
4261	struct files_struct *files = current->files;
4262	struct io_close *close = &req->close;
4263	struct fdtable *fdt;
4264	struct file *file = NULL;
4265	int ret = -EBADF;
4266
4267	spin_lock(&files->file_lock);
4268	fdt = files_fdtable(files);
4269	if (close->fd >= fdt->max_fds) {
4270		spin_unlock(&files->file_lock);
4271		goto err;
4272	}
4273	file = fdt->fd[close->fd];
4274	if (!file || file->f_op == &io_uring_fops) {
4275		spin_unlock(&files->file_lock);
4276		file = NULL;
4277		goto err;
4278	}
4279
4280	/* if the file has a flush method, be safe and punt to async */
4281	if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4282		spin_unlock(&files->file_lock);
4283		return -EAGAIN;
4284	}
4285
4286	ret = __close_fd_get_file(close->fd, &file);
4287	spin_unlock(&files->file_lock);
4288	if (ret < 0) {
4289		if (ret == -ENOENT)
4290			ret = -EBADF;
4291		goto err;
4292	}
4293
4294	/* No ->flush() or already async, safely close from here */
4295	ret = filp_close(file, current->files);
4296err:
4297	if (ret < 0)
4298		req_set_fail(req);
4299	if (file)
4300		fput(file);
4301	__io_req_complete(req, issue_flags, ret, 0);
4302	return 0;
4303}
4304
4305static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4306{
4307	struct io_ring_ctx *ctx = req->ctx;
4308
4309	if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4310		return -EINVAL;
4311	if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4312		return -EINVAL;
4313
4314	req->sync.off = READ_ONCE(sqe->off);
4315	req->sync.len = READ_ONCE(sqe->len);
4316	req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4317	return 0;
4318}
4319
4320static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4321{
4322	int ret;
4323
4324	/* sync_file_range always requires a blocking context */
4325	if (issue_flags & IO_URING_F_NONBLOCK)
4326		return -EAGAIN;
4327
4328	ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4329				req->sync.flags);
4330	if (ret < 0)
4331		req_set_fail(req);
4332	io_req_complete(req, ret);
4333	return 0;
4334}
4335
4336#if defined(CONFIG_NET)
4337static int io_setup_async_msg(struct io_kiocb *req,
4338			      struct io_async_msghdr *kmsg)
4339{
4340	struct io_async_msghdr *async_msg = req->async_data;
4341
4342	if (async_msg)
4343		return -EAGAIN;
4344	if (io_alloc_async_data(req)) {
4345		kfree(kmsg->free_iov);
4346		return -ENOMEM;
4347	}
4348	async_msg = req->async_data;
4349	req->flags |= REQ_F_NEED_CLEANUP;
4350	memcpy(async_msg, kmsg, sizeof(*kmsg));
4351	async_msg->msg.msg_name = &async_msg->addr;
4352	/* if were using fast_iov, set it to the new one */
4353	if (!async_msg->free_iov)
4354		async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4355
4356	return -EAGAIN;
4357}
4358
4359static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4360			       struct io_async_msghdr *iomsg)
4361{
4362	iomsg->msg.msg_name = &iomsg->addr;
4363	iomsg->free_iov = iomsg->fast_iov;
4364	return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4365				   req->sr_msg.msg_flags, &iomsg->free_iov);
4366}
4367
4368static int io_sendmsg_prep_async(struct io_kiocb *req)
4369{
4370	int ret;
4371
4372	ret = io_sendmsg_copy_hdr(req, req->async_data);
4373	if (!ret)
4374		req->flags |= REQ_F_NEED_CLEANUP;
4375	return ret;
4376}
4377
4378static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4379{
4380	struct io_sr_msg *sr = &req->sr_msg;
4381
4382	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4383		return -EINVAL;
4384
4385	sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4386	sr->len = READ_ONCE(sqe->len);
4387	sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4388	if (sr->msg_flags & MSG_DONTWAIT)
4389		req->flags |= REQ_F_NOWAIT;
4390
4391#ifdef CONFIG_COMPAT
4392	if (req->ctx->compat)
4393		sr->msg_flags |= MSG_CMSG_COMPAT;
4394#endif
4395	return 0;
4396}
4397
4398static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4399{
4400	struct io_async_msghdr iomsg, *kmsg;
4401	struct socket *sock;
4402	unsigned flags;
4403	int min_ret = 0;
4404	int ret;
4405
4406	sock = sock_from_file(req->file);
4407	if (unlikely(!sock))
4408		return -ENOTSOCK;
4409
4410	kmsg = req->async_data;
4411	if (!kmsg) {
4412		ret = io_sendmsg_copy_hdr(req, &iomsg);
4413		if (ret)
4414			return ret;
4415		kmsg = &iomsg;
4416	}
4417
4418	flags = req->sr_msg.msg_flags;
4419	if (issue_flags & IO_URING_F_NONBLOCK)
4420		flags |= MSG_DONTWAIT;
4421	if (flags & MSG_WAITALL)
4422		min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4423
4424	ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4425	if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4426		return io_setup_async_msg(req, kmsg);
4427	if (ret == -ERESTARTSYS)
4428		ret = -EINTR;
4429
4430	/* fast path, check for non-NULL to avoid function call */
4431	if (kmsg->free_iov)
4432		kfree(kmsg->free_iov);
4433	req->flags &= ~REQ_F_NEED_CLEANUP;
4434	if (ret < min_ret)
4435		req_set_fail(req);
4436	__io_req_complete(req, issue_flags, ret, 0);
4437	return 0;
4438}
4439
4440static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4441{
4442	struct io_sr_msg *sr = &req->sr_msg;
4443	struct msghdr msg;
4444	struct iovec iov;
4445	struct socket *sock;
4446	unsigned flags;
4447	int min_ret = 0;
4448	int ret;
4449
4450	sock = sock_from_file(req->file);
4451	if (unlikely(!sock))
4452		return -ENOTSOCK;
4453
4454	ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4455	if (unlikely(ret))
4456		return ret;
4457
4458	msg.msg_name = NULL;
4459	msg.msg_control = NULL;
4460	msg.msg_controllen = 0;
4461	msg.msg_namelen = 0;
4462
4463	flags = req->sr_msg.msg_flags;
4464	if (issue_flags & IO_URING_F_NONBLOCK)
4465		flags |= MSG_DONTWAIT;
4466	if (flags & MSG_WAITALL)
4467		min_ret = iov_iter_count(&msg.msg_iter);
4468
4469	msg.msg_flags = flags;
4470	ret = sock_sendmsg(sock, &msg);
4471	if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4472		return -EAGAIN;
4473	if (ret == -ERESTARTSYS)
4474		ret = -EINTR;
4475
4476	if (ret < min_ret)
4477		req_set_fail(req);
4478	__io_req_complete(req, issue_flags, ret, 0);
4479	return 0;
4480}
4481
4482static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4483				 struct io_async_msghdr *iomsg)
4484{
4485	struct io_sr_msg *sr = &req->sr_msg;
4486	struct iovec __user *uiov;
4487	size_t iov_len;
4488	int ret;
4489
4490	ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4491					&iomsg->uaddr, &uiov, &iov_len);
4492	if (ret)
4493		return ret;
4494
4495	if (req->flags & REQ_F_BUFFER_SELECT) {
4496		if (iov_len > 1)
4497			return -EINVAL;
4498		if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4499			return -EFAULT;
4500		sr->len = iomsg->fast_iov[0].iov_len;
4501		iomsg->free_iov = NULL;
4502	} else {
4503		iomsg->free_iov = iomsg->fast_iov;
4504		ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4505				     &iomsg->free_iov, &iomsg->msg.msg_iter,
4506				     false);
4507		if (ret > 0)
4508			ret = 0;
4509	}
4510
4511	return ret;
4512}
4513
4514#ifdef CONFIG_COMPAT
4515static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4516					struct io_async_msghdr *iomsg)
4517{
4518	struct io_sr_msg *sr = &req->sr_msg;
4519	struct compat_iovec __user *uiov;
4520	compat_uptr_t ptr;
4521	compat_size_t len;
4522	int ret;
4523
4524	ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4525				  &ptr, &len);
4526	if (ret)
4527		return ret;
4528
4529	uiov = compat_ptr(ptr);
4530	if (req->flags & REQ_F_BUFFER_SELECT) {
4531		compat_ssize_t clen;
4532
4533		if (len > 1)
4534			return -EINVAL;
4535		if (!access_ok(uiov, sizeof(*uiov)))
4536			return -EFAULT;
4537		if (__get_user(clen, &uiov->iov_len))
4538			return -EFAULT;
4539		if (clen < 0)
4540			return -EINVAL;
4541		sr->len = clen;
4542		iomsg->free_iov = NULL;
4543	} else {
4544		iomsg->free_iov = iomsg->fast_iov;
4545		ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4546				   UIO_FASTIOV, &iomsg->free_iov,
4547				   &iomsg->msg.msg_iter, true);
4548		if (ret < 0)
4549			return ret;
4550	}
4551
4552	return 0;
4553}
4554#endif
4555
4556static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4557			       struct io_async_msghdr *iomsg)
4558{
4559	iomsg->msg.msg_name = &iomsg->addr;
4560
4561#ifdef CONFIG_COMPAT
4562	if (req->ctx->compat)
4563		return __io_compat_recvmsg_copy_hdr(req, iomsg);
4564#endif
4565
4566	return __io_recvmsg_copy_hdr(req, iomsg);
4567}
4568
4569static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4570					       bool needs_lock)
4571{
4572	struct io_sr_msg *sr = &req->sr_msg;
4573	struct io_buffer *kbuf;
4574
4575	kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4576	if (IS_ERR(kbuf))
4577		return kbuf;
4578
4579	sr->kbuf = kbuf;
4580	req->flags |= REQ_F_BUFFER_SELECTED;
4581	return kbuf;
4582}
4583
4584static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4585{
4586	return io_put_kbuf(req, req->sr_msg.kbuf);
4587}
4588
4589static int io_recvmsg_prep_async(struct io_kiocb *req)
4590{
4591	int ret;
4592
4593	ret = io_recvmsg_copy_hdr(req, req->async_data);
4594	if (!ret)
4595		req->flags |= REQ_F_NEED_CLEANUP;
4596	return ret;
4597}
4598
4599static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4600{
4601	struct io_sr_msg *sr = &req->sr_msg;
4602
4603	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4604		return -EINVAL;
4605
4606	sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4607	sr->len = READ_ONCE(sqe->len);
4608	sr->bgid = READ_ONCE(sqe->buf_group);
4609	sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4610	if (sr->msg_flags & MSG_DONTWAIT)
4611		req->flags |= REQ_F_NOWAIT;
4612
4613#ifdef CONFIG_COMPAT
4614	if (req->ctx->compat)
4615		sr->msg_flags |= MSG_CMSG_COMPAT;
4616#endif
4617	return 0;
4618}
4619
4620static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4621{
4622	struct io_async_msghdr iomsg, *kmsg;
4623	struct socket *sock;
4624	struct io_buffer *kbuf;
4625	unsigned flags;
4626	int min_ret = 0;
4627	int ret, cflags = 0;
4628	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4629
4630	sock = sock_from_file(req->file);
4631	if (unlikely(!sock))
4632		return -ENOTSOCK;
4633
4634	kmsg = req->async_data;
4635	if (!kmsg) {
4636		ret = io_recvmsg_copy_hdr(req, &iomsg);
4637		if (ret)
4638			return ret;
4639		kmsg = &iomsg;
4640	}
4641
4642	if (req->flags & REQ_F_BUFFER_SELECT) {
4643		kbuf = io_recv_buffer_select(req, !force_nonblock);
4644		if (IS_ERR(kbuf))
4645			return PTR_ERR(kbuf);
4646		kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4647		kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4648		iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4649				1, req->sr_msg.len);
4650	}
4651
4652	flags = req->sr_msg.msg_flags;
4653	if (force_nonblock)
4654		flags |= MSG_DONTWAIT;
4655	if (flags & MSG_WAITALL)
4656		min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4657
4658	ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4659					kmsg->uaddr, flags);
4660	if (force_nonblock && ret == -EAGAIN)
4661		return io_setup_async_msg(req, kmsg);
4662	if (ret == -ERESTARTSYS)
4663		ret = -EINTR;
4664
4665	if (req->flags & REQ_F_BUFFER_SELECTED)
4666		cflags = io_put_recv_kbuf(req);
4667	/* fast path, check for non-NULL to avoid function call */
4668	if (kmsg->free_iov)
4669		kfree(kmsg->free_iov);
4670	req->flags &= ~REQ_F_NEED_CLEANUP;
4671	if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4672		req_set_fail(req);
4673	__io_req_complete(req, issue_flags, ret, cflags);
4674	return 0;
4675}
4676
4677static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4678{
4679	struct io_buffer *kbuf;
4680	struct io_sr_msg *sr = &req->sr_msg;
4681	struct msghdr msg;
4682	void __user *buf = sr->buf;
4683	struct socket *sock;
4684	struct iovec iov;
4685	unsigned flags;
4686	int min_ret = 0;
4687	int ret, cflags = 0;
4688	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4689
4690	sock = sock_from_file(req->file);
4691	if (unlikely(!sock))
4692		return -ENOTSOCK;
4693
4694	if (req->flags & REQ_F_BUFFER_SELECT) {
4695		kbuf = io_recv_buffer_select(req, !force_nonblock);
4696		if (IS_ERR(kbuf))
4697			return PTR_ERR(kbuf);
4698		buf = u64_to_user_ptr(kbuf->addr);
4699	}
4700
4701	ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4702	if (unlikely(ret))
4703		goto out_free;
4704
4705	msg.msg_name = NULL;
4706	msg.msg_control = NULL;
4707	msg.msg_controllen = 0;
4708	msg.msg_namelen = 0;
4709	msg.msg_iocb = NULL;
4710	msg.msg_flags = 0;
4711
4712	flags = req->sr_msg.msg_flags;
4713	if (force_nonblock)
4714		flags |= MSG_DONTWAIT;
4715	if (flags & MSG_WAITALL)
4716		min_ret = iov_iter_count(&msg.msg_iter);
4717
4718	ret = sock_recvmsg(sock, &msg, flags);
4719	if (force_nonblock && ret == -EAGAIN)
4720		return -EAGAIN;
4721	if (ret == -ERESTARTSYS)
4722		ret = -EINTR;
4723out_free:
4724	if (req->flags & REQ_F_BUFFER_SELECTED)
4725		cflags = io_put_recv_kbuf(req);
4726	if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4727		req_set_fail(req);
4728	__io_req_complete(req, issue_flags, ret, cflags);
4729	return 0;
4730}
4731
4732static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4733{
4734	struct io_accept *accept = &req->accept;
4735
4736	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4737		return -EINVAL;
4738	if (sqe->ioprio || sqe->len || sqe->buf_index)
4739		return -EINVAL;
4740
4741	accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4742	accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4743	accept->flags = READ_ONCE(sqe->accept_flags);
4744	accept->nofile = rlimit(RLIMIT_NOFILE);
4745	return 0;
4746}
4747
4748static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4749{
4750	struct io_accept *accept = &req->accept;
4751	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4752	unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4753	int ret;
4754
4755	if (req->file->f_flags & O_NONBLOCK)
4756		req->flags |= REQ_F_NOWAIT;
4757
4758	ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4759					accept->addr_len, accept->flags,
4760					accept->nofile);
4761	if (ret == -EAGAIN && force_nonblock)
4762		return -EAGAIN;
4763	if (ret < 0) {
4764		if (ret == -ERESTARTSYS)
4765			ret = -EINTR;
4766		req_set_fail(req);
4767	}
4768	__io_req_complete(req, issue_flags, ret, 0);
4769	return 0;
4770}
4771
4772static int io_connect_prep_async(struct io_kiocb *req)
4773{
4774	struct io_async_connect *io = req->async_data;
4775	struct io_connect *conn = &req->connect;
4776
4777	return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4778}
4779
4780static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4781{
4782	struct io_connect *conn = &req->connect;
4783
4784	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4785		return -EINVAL;
4786	if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4787		return -EINVAL;
4788
4789	conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4790	conn->addr_len =  READ_ONCE(sqe->addr2);
4791	return 0;
4792}
4793
4794static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4795{
4796	struct io_async_connect __io, *io;
4797	unsigned file_flags;
4798	int ret;
4799	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4800
4801	if (req->async_data) {
4802		io = req->async_data;
4803	} else {
4804		ret = move_addr_to_kernel(req->connect.addr,
4805						req->connect.addr_len,
4806						&__io.address);
4807		if (ret)
4808			goto out;
4809		io = &__io;
4810	}
4811
4812	file_flags = force_nonblock ? O_NONBLOCK : 0;
4813
4814	ret = __sys_connect_file(req->file, &io->address,
4815					req->connect.addr_len, file_flags);
4816	if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4817		if (req->async_data)
4818			return -EAGAIN;
4819		if (io_alloc_async_data(req)) {
4820			ret = -ENOMEM;
4821			goto out;
4822		}
4823		memcpy(req->async_data, &__io, sizeof(__io));
4824		return -EAGAIN;
4825	}
4826	if (ret == -ERESTARTSYS)
4827		ret = -EINTR;
4828out:
4829	if (ret < 0)
4830		req_set_fail(req);
4831	__io_req_complete(req, issue_flags, ret, 0);
4832	return 0;
4833}
4834#else /* !CONFIG_NET */
4835#define IO_NETOP_FN(op)							\
4836static int io_##op(struct io_kiocb *req, unsigned int issue_flags)	\
4837{									\
4838	return -EOPNOTSUPP;						\
4839}
4840
4841#define IO_NETOP_PREP(op)						\
4842IO_NETOP_FN(op)								\
4843static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4844{									\
4845	return -EOPNOTSUPP;						\
4846}									\
4847
4848#define IO_NETOP_PREP_ASYNC(op)						\
4849IO_NETOP_PREP(op)							\
4850static int io_##op##_prep_async(struct io_kiocb *req)			\
4851{									\
4852	return -EOPNOTSUPP;						\
4853}
4854
4855IO_NETOP_PREP_ASYNC(sendmsg);
4856IO_NETOP_PREP_ASYNC(recvmsg);
4857IO_NETOP_PREP_ASYNC(connect);
4858IO_NETOP_PREP(accept);
4859IO_NETOP_FN(send);
4860IO_NETOP_FN(recv);
4861#endif /* CONFIG_NET */
4862
4863struct io_poll_table {
4864	struct poll_table_struct pt;
4865	struct io_kiocb *req;
4866	int nr_entries;
4867	int error;
4868};
4869
4870static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4871			   __poll_t mask, io_req_tw_func_t func)
4872{
4873	/* for instances that support it check for an event match first: */
4874	if (mask && !(mask & poll->events))
4875		return 0;
4876
4877	trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4878
4879	list_del_init(&poll->wait.entry);
4880
4881	req->result = mask;
4882	req->io_task_work.func = func;
4883
4884	/*
4885	 * If this fails, then the task is exiting. When a task exits, the
4886	 * work gets canceled, so just cancel this request as well instead
4887	 * of executing it. We can't safely execute it anyway, as we may not
4888	 * have the needed state needed for it anyway.
4889	 */
4890	io_req_task_work_add(req);
4891	return 1;
4892}
4893
4894static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4895	__acquires(&req->ctx->completion_lock)
4896{
4897	struct io_ring_ctx *ctx = req->ctx;
4898
4899	/* req->task == current here, checking PF_EXITING is safe */
4900	if (unlikely(req->task->flags & PF_EXITING))
4901		WRITE_ONCE(poll->canceled, true);
4902
4903	if (!req->result && !READ_ONCE(poll->canceled)) {
4904		struct poll_table_struct pt = { ._key = poll->events };
4905
4906		req->result = vfs_poll(req->file, &pt) & poll->events;
4907	}
4908
4909	spin_lock(&ctx->completion_lock);
4910	if (!req->result && !READ_ONCE(poll->canceled)) {
4911		add_wait_queue(poll->head, &poll->wait);
4912		return true;
4913	}
4914
4915	return false;
4916}
4917
4918static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4919{
4920	/* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4921	if (req->opcode == IORING_OP_POLL_ADD)
4922		return req->async_data;
4923	return req->apoll->double_poll;
4924}
4925
4926static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4927{
4928	if (req->opcode == IORING_OP_POLL_ADD)
4929		return &req->poll;
4930	return &req->apoll->poll;
4931}
4932
4933static void io_poll_remove_double(struct io_kiocb *req)
4934	__must_hold(&req->ctx->completion_lock)
4935{
4936	struct io_poll_iocb *poll = io_poll_get_double(req);
4937
4938	lockdep_assert_held(&req->ctx->completion_lock);
4939
4940	if (poll && poll->head) {
4941		struct wait_queue_head *head = poll->head;
4942
4943		spin_lock_irq(&head->lock);
4944		list_del_init(&poll->wait.entry);
4945		if (poll->wait.private)
4946			req_ref_put(req);
4947		poll->head = NULL;
4948		spin_unlock_irq(&head->lock);
4949	}
4950}
4951
4952static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4953	__must_hold(&req->ctx->completion_lock)
4954{
4955	struct io_ring_ctx *ctx = req->ctx;
4956	unsigned flags = IORING_CQE_F_MORE;
4957	int error;
4958
4959	if (READ_ONCE(req->poll.canceled)) {
4960		error = -ECANCELED;
4961		req->poll.events |= EPOLLONESHOT;
4962	} else {
4963		error = mangle_poll(mask);
4964	}
4965	if (req->poll.events & EPOLLONESHOT)
4966		flags = 0;
4967	if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4968		req->poll.done = true;
4969		flags = 0;
4970	}
4971	if (flags & IORING_CQE_F_MORE)
4972		ctx->cq_extra++;
4973
4974	io_commit_cqring(ctx);
4975	return !(flags & IORING_CQE_F_MORE);
4976}
4977
4978static void io_poll_task_func(struct io_kiocb *req)
4979{
4980	struct io_ring_ctx *ctx = req->ctx;
4981	struct io_kiocb *nxt;
4982
4983	if (io_poll_rewait(req, &req->poll)) {
4984		spin_unlock(&ctx->completion_lock);
4985	} else {
4986		bool done;
4987
4988		done = io_poll_complete(req, req->result);
4989		if (done) {
4990			io_poll_remove_double(req);
4991			hash_del(&req->hash_node);
4992		} else {
4993			req->result = 0;
4994			add_wait_queue(req->poll.head, &req->poll.wait);
4995		}
4996		spin_unlock(&ctx->completion_lock);
4997		io_cqring_ev_posted(ctx);
4998
4999		if (done) {
5000			nxt = io_put_req_find_next(req);
5001			if (nxt)
5002				io_req_task_submit(nxt);
5003		}
5004	}
5005}
5006
5007static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5008			       int sync, void *key)
5009{
5010	struct io_kiocb *req = wait->private;
5011	struct io_poll_iocb *poll = io_poll_get_single(req);
5012	__poll_t mask = key_to_poll(key);
5013	unsigned long flags;
5014
5015	/* for instances that support it check for an event match first: */
5016	if (mask && !(mask & poll->events))
5017		return 0;
5018	if (!(poll->events & EPOLLONESHOT))
5019		return poll->wait.func(&poll->wait, mode, sync, key);
5020
5021	list_del_init(&wait->entry);
5022
5023	if (poll->head) {
5024		bool done;
5025
5026		spin_lock_irqsave(&poll->head->lock, flags);
5027		done = list_empty(&poll->wait.entry);
5028		if (!done)
5029			list_del_init(&poll->wait.entry);
5030		/* make sure double remove sees this as being gone */
5031		wait->private = NULL;
5032		spin_unlock_irqrestore(&poll->head->lock, flags);
5033		if (!done) {
5034			/* use wait func handler, so it matches the rq type */
5035			poll->wait.func(&poll->wait, mode, sync, key);
5036		}
5037	}
5038	req_ref_put(req);
5039	return 1;
5040}
5041
5042static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5043			      wait_queue_func_t wake_func)
5044{
5045	poll->head = NULL;
5046	poll->done = false;
5047	poll->canceled = false;
5048#define IO_POLL_UNMASK	(EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5049	/* mask in events that we always want/need */
5050	poll->events = events | IO_POLL_UNMASK;
5051	INIT_LIST_HEAD(&poll->wait.entry);
5052	init_waitqueue_func_entry(&poll->wait, wake_func);
5053}
5054
5055static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5056			    struct wait_queue_head *head,
5057			    struct io_poll_iocb **poll_ptr)
5058{
5059	struct io_kiocb *req = pt->req;
5060
5061	/*
5062	 * The file being polled uses multiple waitqueues for poll handling
5063	 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5064	 * if this happens.
5065	 */
5066	if (unlikely(pt->nr_entries)) {
5067		struct io_poll_iocb *poll_one = poll;
5068
5069		/* double add on the same waitqueue head, ignore */
5070		if (poll_one->head == head)
5071			return;
5072		/* already have a 2nd entry, fail a third attempt */
5073		if (*poll_ptr) {
5074			if ((*poll_ptr)->head == head)
5075				return;
5076			pt->error = -EINVAL;
5077			return;
5078		}
5079		/*
5080		 * Can't handle multishot for double wait for now, turn it
5081		 * into one-shot mode.
5082		 */
5083		if (!(poll_one->events & EPOLLONESHOT))
5084			poll_one->events |= EPOLLONESHOT;
5085		poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5086		if (!poll) {
5087			pt->error = -ENOMEM;
5088			return;
5089		}
5090		io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5091		req_ref_get(req);
5092		poll->wait.private = req;
5093		*poll_ptr = poll;
5094	}
5095
5096	pt->nr_entries++;
5097	poll->head = head;
5098
5099	if (poll->events & EPOLLEXCLUSIVE)
5100		add_wait_queue_exclusive(head, &poll->wait);
5101	else
5102		add_wait_queue(head, &poll->wait);
5103}
5104
5105static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5106			       struct poll_table_struct *p)
5107{
5108	struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5109	struct async_poll *apoll = pt->req->apoll;
5110
5111	__io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5112}
5113
5114static void io_async_task_func(struct io_kiocb *req)
5115{
5116	struct async_poll *apoll = req->apoll;
5117	struct io_ring_ctx *ctx = req->ctx;
5118
5119	trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5120
5121	if (io_poll_rewait(req, &apoll->poll)) {
5122		spin_unlock(&ctx->completion_lock);
5123		return;
5124	}
5125
5126	hash_del(&req->hash_node);
5127	io_poll_remove_double(req);
5128	spin_unlock(&ctx->completion_lock);
5129
5130	if (!READ_ONCE(apoll->poll.canceled))
5131		io_req_task_submit(req);
5132	else
5133		io_req_complete_failed(req, -ECANCELED);
5134}
5135
5136static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5137			void *key)
5138{
5139	struct io_kiocb *req = wait->private;
5140	struct io_poll_iocb *poll = &req->apoll->poll;
5141
5142	trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5143					key_to_poll(key));
5144
5145	return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5146}
5147
5148static void io_poll_req_insert(struct io_kiocb *req)
5149{
5150	struct io_ring_ctx *ctx = req->ctx;
5151	struct hlist_head *list;
5152
5153	list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5154	hlist_add_head(&req->hash_node, list);
5155}
5156
5157static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5158				      struct io_poll_iocb *poll,
5159				      struct io_poll_table *ipt, __poll_t mask,
5160				      wait_queue_func_t wake_func)
5161	__acquires(&ctx->completion_lock)
5162{
5163	struct io_ring_ctx *ctx = req->ctx;
5164	bool cancel = false;
5165
5166	INIT_HLIST_NODE(&req->hash_node);
5167	io_init_poll_iocb(poll, mask, wake_func);
5168	poll->file = req->file;
5169	poll->wait.private = req;
5170
5171	ipt->pt._key = mask;
5172	ipt->req = req;
5173	ipt->error = 0;
5174	ipt->nr_entries = 0;
5175
5176	mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5177	if (unlikely(!ipt->nr_entries) && !ipt->error)
5178		ipt->error = -EINVAL;
5179
5180	spin_lock(&ctx->completion_lock);
5181	if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5182		io_poll_remove_double(req);
5183	if (likely(poll->head)) {
5184		spin_lock_irq(&poll->head->lock);
5185		if (unlikely(list_empty(&poll->wait.entry))) {
5186			if (ipt->error)
5187				cancel = true;
5188			ipt->error = 0;
5189			mask = 0;
5190		}
5191		if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5192			list_del_init(&poll->wait.entry);
5193		else if (cancel)
5194			WRITE_ONCE(poll->canceled, true);
5195		else if (!poll->done) /* actually waiting for an event */
5196			io_poll_req_insert(req);
5197		spin_unlock_irq(&poll->head->lock);
5198	}
5199
5200	return mask;
5201}
5202
5203enum {
5204	IO_APOLL_OK,
5205	IO_APOLL_ABORTED,
5206	IO_APOLL_READY
5207};
5208
5209static int io_arm_poll_handler(struct io_kiocb *req)
5210{
5211	const struct io_op_def *def = &io_op_defs[req->opcode];
5212	struct io_ring_ctx *ctx = req->ctx;
5213	struct async_poll *apoll;
5214	struct io_poll_table ipt;
5215	__poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5216	int rw;
5217
5218	if (!req->file || !file_can_poll(req->file))
5219		return IO_APOLL_ABORTED;
5220	if (req->flags & REQ_F_POLLED)
5221		return IO_APOLL_ABORTED;
5222	if (!def->pollin && !def->pollout)
5223		return IO_APOLL_ABORTED;
5224
5225	if (def->pollin) {
5226		rw = READ;
5227		mask |= POLLIN | POLLRDNORM;
5228
5229		/* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5230		if ((req->opcode == IORING_OP_RECVMSG) &&
5231		    (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5232			mask &= ~POLLIN;
5233	} else {
5234		rw = WRITE;
5235		mask |= POLLOUT | POLLWRNORM;
5236	}
5237
5238	/* if we can't nonblock try, then no point in arming a poll handler */
5239	if (!io_file_supports_nowait(req, rw))
5240		return IO_APOLL_ABORTED;
5241
5242	apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5243	if (unlikely(!apoll))
5244		return IO_APOLL_ABORTED;
5245	apoll->double_poll = NULL;
5246	req->apoll = apoll;
5247	req->flags |= REQ_F_POLLED;
5248	ipt.pt._qproc = io_async_queue_proc;
5249	io_req_set_refcount(req);
5250
5251	ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5252					io_async_wake);
5253	spin_unlock(&ctx->completion_lock);
5254	if (ret || ipt.error)
5255		return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5256
5257	trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5258				mask, apoll->poll.events);
5259	return IO_APOLL_OK;
5260}
5261
5262static bool __io_poll_remove_one(struct io_kiocb *req,
5263				 struct io_poll_iocb *poll, bool do_cancel)
5264	__must_hold(&req->ctx->completion_lock)
5265{
5266	bool do_complete = false;
5267
5268	if (!poll->head)
5269		return false;
5270	spin_lock_irq(&poll->head->lock);
5271	if (do_cancel)
5272		WRITE_ONCE(poll->canceled, true);
5273	if (!list_empty(&poll->wait.entry)) {
5274		list_del_init(&poll->wait.entry);
5275		do_complete = true;
5276	}
5277	spin_unlock_irq(&poll->head->lock);
5278	hash_del(&req->hash_node);
5279	return do_complete;
5280}
5281
5282static bool io_poll_remove_one(struct io_kiocb *req)
5283	__must_hold(&req->ctx->completion_lock)
5284{
5285	bool do_complete;
5286
5287	io_poll_remove_double(req);
5288	do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5289
5290	if (do_complete) {
5291		io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5292		io_commit_cqring(req->ctx);
5293		req_set_fail(req);
5294		io_put_req_deferred(req);
5295	}
5296	return do_complete;
5297}
5298
5299/*
5300 * Returns true if we found and killed one or more poll requests
5301 */
5302static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5303			       bool cancel_all)
5304{
5305	struct hlist_node *tmp;
5306	struct io_kiocb *req;
5307	int posted = 0, i;
5308
5309	spin_lock(&ctx->completion_lock);
5310	for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5311		struct hlist_head *list;
5312
5313		list = &ctx->cancel_hash[i];
5314		hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5315			if (io_match_task(req, tsk, cancel_all))
5316				posted += io_poll_remove_one(req);
5317		}
5318	}
5319	spin_unlock(&ctx->completion_lock);
5320
5321	if (posted)
5322		io_cqring_ev_posted(ctx);
5323
5324	return posted != 0;
5325}
5326
5327static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5328				     bool poll_only)
5329	__must_hold(&ctx->completion_lock)
5330{
5331	struct hlist_head *list;
5332	struct io_kiocb *req;
5333
5334	list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5335	hlist_for_each_entry(req, list, hash_node) {
5336		if (sqe_addr != req->user_data)
5337			continue;
5338		if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5339			continue;
5340		return req;
5341	}
5342	return NULL;
5343}
5344
5345static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5346			  bool poll_only)
5347	__must_hold(&ctx->completion_lock)
5348{
5349	struct io_kiocb *req;
5350
5351	req = io_poll_find(ctx, sqe_addr, poll_only);
5352	if (!req)
5353		return -ENOENT;
5354	if (io_poll_remove_one(req))
5355		return 0;
5356
5357	return -EALREADY;
5358}
5359
5360static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5361				     unsigned int flags)
5362{
5363	u32 events;
5364
5365	events = READ_ONCE(sqe->poll32_events);
5366#ifdef __BIG_ENDIAN
5367	events = swahw32(events);
5368#endif
5369	if (!(flags & IORING_POLL_ADD_MULTI))
5370		events |= EPOLLONESHOT;
5371	return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5372}
5373
5374static int io_poll_update_prep(struct io_kiocb *req,
5375			       const struct io_uring_sqe *sqe)
5376{
5377	struct io_poll_update *upd = &req->poll_update;
5378	u32 flags;
5379
5380	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5381		return -EINVAL;
5382	if (sqe->ioprio || sqe->buf_index)
5383		return -EINVAL;
5384	flags = READ_ONCE(sqe->len);
5385	if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5386		      IORING_POLL_ADD_MULTI))
5387		return -EINVAL;
5388	/* meaningless without update */
5389	if (flags == IORING_POLL_ADD_MULTI)
5390		return -EINVAL;
5391
5392	upd->old_user_data = READ_ONCE(sqe->addr);
5393	upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5394	upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5395
5396	upd->new_user_data = READ_ONCE(sqe->off);
5397	if (!upd->update_user_data && upd->new_user_data)
5398		return -EINVAL;
5399	if (upd->update_events)
5400		upd->events = io_poll_parse_events(sqe, flags);
5401	else if (sqe->poll32_events)
5402		return -EINVAL;
5403
5404	return 0;
5405}
5406
5407static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5408			void *key)
5409{
5410	struct io_kiocb *req = wait->private;
5411	struct io_poll_iocb *poll = &req->poll;
5412
5413	return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5414}
5415
5416static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5417			       struct poll_table_struct *p)
5418{
5419	struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5420
5421	__io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5422}
5423
5424static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5425{
5426	struct io_poll_iocb *poll = &req->poll;
5427	u32 flags;
5428
5429	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5430		return -EINVAL;
5431	if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5432		return -EINVAL;
5433	flags = READ_ONCE(sqe->len);
5434	if (flags & ~IORING_POLL_ADD_MULTI)
5435		return -EINVAL;
5436
5437	io_req_set_refcount(req);
5438	poll->events = io_poll_parse_events(sqe, flags);
5439	return 0;
5440}
5441
5442static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5443{
5444	struct io_poll_iocb *poll = &req->poll;
5445	struct io_ring_ctx *ctx = req->ctx;
5446	struct io_poll_table ipt;
5447	__poll_t mask;
5448
5449	ipt.pt._qproc = io_poll_queue_proc;
5450
5451	mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5452					io_poll_wake);
5453
5454	if (mask) { /* no async, we'd stolen it */
5455		ipt.error = 0;
5456		io_poll_complete(req, mask);
5457	}
5458	spin_unlock(&ctx->completion_lock);
5459
5460	if (mask) {
5461		io_cqring_ev_posted(ctx);
5462		if (poll->events & EPOLLONESHOT)
5463			io_put_req(req);
5464	}
5465	return ipt.error;
5466}
5467
5468static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5469{
5470	struct io_ring_ctx *ctx = req->ctx;
5471	struct io_kiocb *preq;
5472	bool completing;
5473	int ret;
5474
5475	spin_lock(&ctx->completion_lock);
5476	preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5477	if (!preq) {
5478		ret = -ENOENT;
5479		goto err;
5480	}
5481
5482	if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5483		completing = true;
5484		ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5485		goto err;
5486	}
5487
5488	/*
5489	 * Don't allow racy completion with singleshot, as we cannot safely
5490	 * update those. For multishot, if we're racing with completion, just
5491	 * let completion re-add it.
5492	 */
5493	completing = !__io_poll_remove_one(preq, &preq->poll, false);
5494	if (completing && (preq->poll.events & EPOLLONESHOT)) {
5495		ret = -EALREADY;
5496		goto err;
5497	}
5498	/* we now have a detached poll request. reissue. */
5499	ret = 0;
5500err:
5501	if (ret < 0) {
5502		spin_unlock(&ctx->completion_lock);
5503		req_set_fail(req);
5504		io_req_complete(req, ret);
5505		return 0;
5506	}
5507	/* only mask one event flags, keep behavior flags */
5508	if (req->poll_update.update_events) {
5509		preq->poll.events &= ~0xffff;
5510		preq->poll.events |= req->poll_update.events & 0xffff;
5511		preq->poll.events |= IO_POLL_UNMASK;
5512	}
5513	if (req->poll_update.update_user_data)
5514		preq->user_data = req->poll_update.new_user_data;
5515	spin_unlock(&ctx->completion_lock);
5516
5517	/* complete update request, we're done with it */
5518	io_req_complete(req, ret);
5519
5520	if (!completing) {
5521		ret = io_poll_add(preq, issue_flags);
5522		if (ret < 0) {
5523			req_set_fail(preq);
5524			io_req_complete(preq, ret);
5525		}
5526	}
5527	return 0;
5528}
5529
5530static void io_req_task_timeout(struct io_kiocb *req)
5531{
5532	req_set_fail(req);
5533	io_req_complete_post(req, -ETIME, 0);
5534}
5535
5536static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5537{
5538	struct io_timeout_data *data = container_of(timer,
5539						struct io_timeout_data, timer);
5540	struct io_kiocb *req = data->req;
5541	struct io_ring_ctx *ctx = req->ctx;
5542	unsigned long flags;
5543
5544	spin_lock_irqsave(&ctx->timeout_lock, flags);
5545	list_del_init(&req->timeout.list);
5546	atomic_set(&req->ctx->cq_timeouts,
5547		atomic_read(&req->ctx->cq_timeouts) + 1);
5548	spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5549
5550	req->io_task_work.func = io_req_task_timeout;
5551	io_req_task_work_add(req);
5552	return HRTIMER_NORESTART;
5553}
5554
5555static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5556					   __u64 user_data)
5557	__must_hold(&ctx->timeout_lock)
5558{
5559	struct io_timeout_data *io;
5560	struct io_kiocb *req;
5561	bool found = false;
5562
5563	list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5564		found = user_data == req->user_data;
5565		if (found)
5566			break;
5567	}
5568	if (!found)
5569		return ERR_PTR(-ENOENT);
5570
5571	io = req->async_data;
5572	if (hrtimer_try_to_cancel(&io->timer) == -1)
5573		return ERR_PTR(-EALREADY);
5574	list_del_init(&req->timeout.list);
5575	return req;
5576}
5577
5578static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5579	__must_hold(&ctx->completion_lock)
5580	__must_hold(&ctx->timeout_lock)
5581{
5582	struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5583
5584	if (IS_ERR(req))
5585		return PTR_ERR(req);
5586
5587	req_set_fail(req);
5588	io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5589	io_put_req_deferred(req);
5590	return 0;
5591}
5592
5593static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5594			     struct timespec64 *ts, enum hrtimer_mode mode)
5595	__must_hold(&ctx->timeout_lock)
5596{
5597	struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5598	struct io_timeout_data *data;
5599
5600	if (IS_ERR(req))
5601		return PTR_ERR(req);
5602
5603	req->timeout.off = 0; /* noseq */
5604	data = req->async_data;
5605	list_add_tail(&req->timeout.list, &ctx->timeout_list);
5606	hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5607	data->timer.function = io_timeout_fn;
5608	hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5609	return 0;
5610}
5611
5612static int io_timeout_remove_prep(struct io_kiocb *req,
5613				  const struct io_uring_sqe *sqe)
5614{
5615	struct io_timeout_rem *tr = &req->timeout_rem;
5616
5617	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5618		return -EINVAL;
5619	if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5620		return -EINVAL;
5621	if (sqe->ioprio || sqe->buf_index || sqe->len)
5622		return -EINVAL;
5623
5624	tr->addr = READ_ONCE(sqe->addr);
5625	tr->flags = READ_ONCE(sqe->timeout_flags);
5626	if (tr->flags & IORING_TIMEOUT_UPDATE) {
5627		if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5628			return -EINVAL;
5629		if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5630			return -EFAULT;
5631	} else if (tr->flags) {
5632		/* timeout removal doesn't support flags */
5633		return -EINVAL;
5634	}
5635
5636	return 0;
5637}
5638
5639static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5640{
5641	return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5642					    : HRTIMER_MODE_REL;
5643}
5644
5645/*
5646 * Remove or update an existing timeout command
5647 */
5648static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5649{
5650	struct io_timeout_rem *tr = &req->timeout_rem;
5651	struct io_ring_ctx *ctx = req->ctx;
5652	int ret;
5653
5654	if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
5655		spin_lock(&ctx->completion_lock);
5656		spin_lock_irq(&ctx->timeout_lock);
5657		ret = io_timeout_cancel(ctx, tr->addr);
5658		spin_unlock_irq(&ctx->timeout_lock);
5659		spin_unlock(&ctx->completion_lock);
5660	} else {
5661		spin_lock_irq(&ctx->timeout_lock);
5662		ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5663					io_translate_timeout_mode(tr->flags));
5664		spin_unlock_irq(&ctx->timeout_lock);
5665	}
5666
5667	if (ret < 0)
5668		req_set_fail(req);
5669	io_req_complete_post(req, ret, 0);
5670	return 0;
5671}
5672
5673static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5674			   bool is_timeout_link)
5675{
5676	struct io_timeout_data *data;
5677	unsigned flags;
5678	u32 off = READ_ONCE(sqe->off);
5679
5680	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5681		return -EINVAL;
5682	if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5683		return -EINVAL;
5684	if (off && is_timeout_link)
5685		return -EINVAL;
5686	flags = READ_ONCE(sqe->timeout_flags);
5687	if (flags & ~IORING_TIMEOUT_ABS)
5688		return -EINVAL;
5689
5690	req->timeout.off = off;
5691	if (unlikely(off && !req->ctx->off_timeout_used))
5692		req->ctx->off_timeout_used = true;
5693
5694	if (!req->async_data && io_alloc_async_data(req))
5695		return -ENOMEM;
5696
5697	data = req->async_data;
5698	data->req = req;
5699
5700	if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5701		return -EFAULT;
5702
5703	data->mode = io_translate_timeout_mode(flags);
5704	hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5705
5706	if (is_timeout_link) {
5707		struct io_submit_link *link = &req->ctx->submit_state.link;
5708
5709		if (!link->head)
5710			return -EINVAL;
5711		if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
5712			return -EINVAL;
5713		req->timeout.head = link->last;
5714		link->last->flags |= REQ_F_ARM_LTIMEOUT;
5715	}
5716	return 0;
5717}
5718
5719static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5720{
5721	struct io_ring_ctx *ctx = req->ctx;
5722	struct io_timeout_data *data = req->async_data;
5723	struct list_head *entry;
5724	u32 tail, off = req->timeout.off;
5725
5726	spin_lock_irq(&ctx->timeout_lock);