io_uring.c revision 23a65db8
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	percpu_counter_sub(&tctx->inflight, nr);
1627	if (unlikely(atomic_read(&tctx->in_idle)))
1628		wake_up(&tctx->wait);
1629	put_task_struct_many(task, nr);
1630}
1631
1632static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1633				     long res, unsigned int cflags)
1634{
1635	struct io_overflow_cqe *ocqe;
1636
1637	ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1638	if (!ocqe) {
1639		/*
1640		 * If we're in ring overflow flush mode, or in task cancel mode,
1641		 * or cannot allocate an overflow entry, then we need to drop it
1642		 * on the floor.
1643		 */
1644		io_account_cq_overflow(ctx);
1645		return false;
1646	}
1647	if (list_empty(&ctx->cq_overflow_list)) {
1648		set_bit(0, &ctx->check_cq_overflow);
1649		WRITE_ONCE(ctx->rings->sq_flags,
1650			   ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1651
1652	}
1653	ocqe->cqe.user_data = user_data;
1654	ocqe->cqe.res = res;
1655	ocqe->cqe.flags = cflags;
1656	list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1657	return true;
1658}
1659
1660static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1661					  long res, unsigned int cflags)
1662{
1663	struct io_uring_cqe *cqe;
1664
1665	trace_io_uring_complete(ctx, user_data, res, cflags);
1666
1667	/*
1668	 * If we can't get a cq entry, userspace overflowed the
1669	 * submission (by quite a lot). Increment the overflow count in
1670	 * the ring.
1671	 */
1672	cqe = io_get_cqe(ctx);
1673	if (likely(cqe)) {
1674		WRITE_ONCE(cqe->user_data, user_data);
1675		WRITE_ONCE(cqe->res, res);
1676		WRITE_ONCE(cqe->flags, cflags);
1677		return true;
1678	}
1679	return io_cqring_event_overflow(ctx, user_data, res, cflags);
1680}
1681
1682/* not as hot to bloat with inlining */
1683static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1684					  long res, unsigned int cflags)
1685{
1686	return __io_cqring_fill_event(ctx, user_data, res, cflags);
1687}
1688
1689static void io_req_complete_post(struct io_kiocb *req, long res,
1690				 unsigned int cflags)
1691{
1692	struct io_ring_ctx *ctx = req->ctx;
1693
1694	spin_lock(&ctx->completion_lock);
1695	__io_cqring_fill_event(ctx, req->user_data, res, cflags);
1696	/*
1697	 * If we're the last reference to this request, add to our locked
1698	 * free_list cache.
1699	 */
1700	if (req_ref_put_and_test(req)) {
1701		if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1702			if (req->flags & IO_DISARM_MASK)
1703				io_disarm_next(req);
1704			if (req->link) {
1705				io_req_task_queue(req->link);
1706				req->link = NULL;
1707			}
1708		}
1709		io_dismantle_req(req);
1710		io_put_task(req->task, 1);
1711		list_add(&req->inflight_entry, &ctx->locked_free_list);
1712		ctx->locked_free_nr++;
1713	} else {
1714		if (!percpu_ref_tryget(&ctx->refs))
1715			req = NULL;
1716	}
1717	io_commit_cqring(ctx);
1718	spin_unlock(&ctx->completion_lock);
1719
1720	if (req) {
1721		io_cqring_ev_posted(ctx);
1722		percpu_ref_put(&ctx->refs);
1723	}
1724}
1725
1726static inline bool io_req_needs_clean(struct io_kiocb *req)
1727{
1728	return req->flags & IO_REQ_CLEAN_FLAGS;
1729}
1730
1731static void io_req_complete_state(struct io_kiocb *req, long res,
1732				  unsigned int cflags)
1733{
1734	if (io_req_needs_clean(req))
1735		io_clean_op(req);
1736	req->result = res;
1737	req->compl.cflags = cflags;
1738	req->flags |= REQ_F_COMPLETE_INLINE;
1739}
1740
1741static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1742				     long res, unsigned cflags)
1743{
1744	if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1745		io_req_complete_state(req, res, cflags);
1746	else
1747		io_req_complete_post(req, res, cflags);
1748}
1749
1750static inline void io_req_complete(struct io_kiocb *req, long res)
1751{
1752	__io_req_complete(req, 0, res, 0);
1753}
1754
1755static void io_req_complete_failed(struct io_kiocb *req, long res)
1756{
1757	req_set_fail(req);
1758	io_req_complete_post(req, res, 0);
1759}
1760
1761/*
1762 * Don't initialise the fields below on every allocation, but do that in
1763 * advance and keep them valid across allocations.
1764 */
1765static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1766{
1767	req->ctx = ctx;
1768	req->link = NULL;
1769	req->async_data = NULL;
1770	/* not necessary, but safer to zero */
1771	req->result = 0;
1772}
1773
1774static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1775					struct io_submit_state *state)
1776{
1777	spin_lock(&ctx->completion_lock);
1778	list_splice_init(&ctx->locked_free_list, &state->free_list);
1779	ctx->locked_free_nr = 0;
1780	spin_unlock(&ctx->completion_lock);
1781}
1782
1783/* Returns true IFF there are requests in the cache */
1784static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1785{
1786	struct io_submit_state *state = &ctx->submit_state;
1787	int nr;
1788
1789	/*
1790	 * If we have more than a batch's worth of requests in our IRQ side
1791	 * locked cache, grab the lock and move them over to our submission
1792	 * side cache.
1793	 */
1794	if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1795		io_flush_cached_locked_reqs(ctx, state);
1796
1797	nr = state->free_reqs;
1798	while (!list_empty(&state->free_list)) {
1799		struct io_kiocb *req = list_first_entry(&state->free_list,
1800					struct io_kiocb, inflight_entry);
1801
1802		list_del(&req->inflight_entry);
1803		state->reqs[nr++] = req;
1804		if (nr == ARRAY_SIZE(state->reqs))
1805			break;
1806	}
1807
1808	state->free_reqs = nr;
1809	return nr != 0;
1810}
1811
1812/*
1813 * A request might get retired back into the request caches even before opcode
1814 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1815 * Because of that, io_alloc_req() should be called only under ->uring_lock
1816 * and with extra caution to not get a request that is still worked on.
1817 */
1818static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1819	__must_hold(&ctx->uring_lock)
1820{
1821	struct io_submit_state *state = &ctx->submit_state;
1822	gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1823	int ret, i;
1824
1825	BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1826
1827	if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1828		goto got_req;
1829
1830	ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1831				    state->reqs);
1832
1833	/*
1834	 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1835	 * retry single alloc to be on the safe side.
1836	 */
1837	if (unlikely(ret <= 0)) {
1838		state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1839		if (!state->reqs[0])
1840			return NULL;
1841		ret = 1;
1842	}
1843
1844	for (i = 0; i < ret; i++)
1845		io_preinit_req(state->reqs[i], ctx);
1846	state->free_reqs = ret;
1847got_req:
1848	state->free_reqs--;
1849	return state->reqs[state->free_reqs];
1850}
1851
1852static inline void io_put_file(struct file *file)
1853{
1854	if (file)
1855		fput(file);
1856}
1857
1858static void io_dismantle_req(struct io_kiocb *req)
1859{
1860	unsigned int flags = req->flags;
1861
1862	if (io_req_needs_clean(req))
1863		io_clean_op(req);
1864	if (!(flags & REQ_F_FIXED_FILE))
1865		io_put_file(req->file);
1866	if (req->fixed_rsrc_refs)
1867		percpu_ref_put(req->fixed_rsrc_refs);
1868	if (req->async_data) {
1869		kfree(req->async_data);
1870		req->async_data = NULL;
1871	}
1872}
1873
1874static void __io_free_req(struct io_kiocb *req)
1875{
1876	struct io_ring_ctx *ctx = req->ctx;
1877
1878	io_dismantle_req(req);
1879	io_put_task(req->task, 1);
1880
1881	spin_lock(&ctx->completion_lock);
1882	list_add(&req->inflight_entry, &ctx->locked_free_list);
1883	ctx->locked_free_nr++;
1884	spin_unlock(&ctx->completion_lock);
1885
1886	percpu_ref_put(&ctx->refs);
1887}
1888
1889static inline void io_remove_next_linked(struct io_kiocb *req)
1890{
1891	struct io_kiocb *nxt = req->link;
1892
1893	req->link = nxt->link;
1894	nxt->link = NULL;
1895}
1896
1897static bool io_kill_linked_timeout(struct io_kiocb *req)
1898	__must_hold(&req->ctx->completion_lock)
1899	__must_hold(&req->ctx->timeout_lock)
1900{
1901	struct io_kiocb *link = req->link;
1902
1903	if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
1904		struct io_timeout_data *io = link->async_data;
1905
1906		io_remove_next_linked(req);
1907		link->timeout.head = NULL;
1908		if (hrtimer_try_to_cancel(&io->timer) != -1) {
1909			io_cqring_fill_event(link->ctx, link->user_data,
1910					     -ECANCELED, 0);
1911			io_put_req_deferred(link);
1912			return true;
1913		}
1914	}
1915	return false;
1916}
1917
1918static void io_fail_links(struct io_kiocb *req)
1919	__must_hold(&req->ctx->completion_lock)
1920{
1921	struct io_kiocb *nxt, *link = req->link;
1922
1923	req->link = NULL;
1924	while (link) {
1925		nxt = link->link;
1926		link->link = NULL;
1927
1928		trace_io_uring_fail_link(req, link);
1929		io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1930		io_put_req_deferred(link);
1931		link = nxt;
1932	}
1933}
1934
1935static bool io_disarm_next(struct io_kiocb *req)
1936	__must_hold(&req->ctx->completion_lock)
1937{
1938	bool posted = false;
1939
1940	if (req->flags & REQ_F_ARM_LTIMEOUT) {
1941		struct io_kiocb *link = req->link;
1942
1943		req->flags &= ~REQ_F_ARM_LTIMEOUT;
1944		if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
1945			io_remove_next_linked(req);
1946			io_cqring_fill_event(link->ctx, link->user_data,
1947					     -ECANCELED, 0);
1948			io_put_req_deferred(link);
1949			posted = true;
1950		}
1951	} else if (req->flags & REQ_F_LINK_TIMEOUT) {
1952		struct io_ring_ctx *ctx = req->ctx;
1953
1954		spin_lock_irq(&ctx->timeout_lock);
1955		posted = io_kill_linked_timeout(req);
1956		spin_unlock_irq(&ctx->timeout_lock);
1957	}
1958	if (unlikely((req->flags & REQ_F_FAIL) &&
1959		     !(req->flags & REQ_F_HARDLINK))) {
1960		posted |= (req->link != NULL);
1961		io_fail_links(req);
1962	}
1963	return posted;
1964}
1965
1966static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1967{
1968	struct io_kiocb *nxt;
1969
1970	/*
1971	 * If LINK is set, we have dependent requests in this chain. If we
1972	 * didn't fail this request, queue the first one up, moving any other
1973	 * dependencies to the next request. In case of failure, fail the rest
1974	 * of the chain.
1975	 */
1976	if (req->flags & IO_DISARM_MASK) {
1977		struct io_ring_ctx *ctx = req->ctx;
1978		bool posted;
1979
1980		spin_lock(&ctx->completion_lock);
1981		posted = io_disarm_next(req);
1982		if (posted)
1983			io_commit_cqring(req->ctx);
1984		spin_unlock(&ctx->completion_lock);
1985		if (posted)
1986			io_cqring_ev_posted(ctx);
1987	}
1988	nxt = req->link;
1989	req->link = NULL;
1990	return nxt;
1991}
1992
1993static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1994{
1995	if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1996		return NULL;
1997	return __io_req_find_next(req);
1998}
1999
2000static void ctx_flush_and_put(struct io_ring_ctx *ctx)
2001{
2002	if (!ctx)
2003		return;
2004	if (ctx->submit_state.compl_nr) {
2005		mutex_lock(&ctx->uring_lock);
2006		io_submit_flush_completions(ctx);
2007		mutex_unlock(&ctx->uring_lock);
2008	}
2009	percpu_ref_put(&ctx->refs);
2010}
2011
2012static void tctx_task_work(struct callback_head *cb)
2013{
2014	struct io_ring_ctx *ctx = NULL;
2015	struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2016						  task_work);
2017
2018	while (1) {
2019		struct io_wq_work_node *node;
2020
2021		spin_lock_irq(&tctx->task_lock);
2022		node = tctx->task_list.first;
2023		INIT_WQ_LIST(&tctx->task_list);
2024		if (!node)
2025			tctx->task_running = false;
2026		spin_unlock_irq(&tctx->task_lock);
2027		if (!node)
2028			break;
2029
2030		do {
2031			struct io_wq_work_node *next = node->next;
2032			struct io_kiocb *req = container_of(node, struct io_kiocb,
2033							    io_task_work.node);
2034
2035			if (req->ctx != ctx) {
2036				ctx_flush_and_put(ctx);
2037				ctx = req->ctx;
2038				percpu_ref_get(&ctx->refs);
2039			}
2040			req->io_task_work.func(req);
2041			node = next;
2042		} while (node);
2043
2044		cond_resched();
2045	}
2046
2047	ctx_flush_and_put(ctx);
2048}
2049
2050static void io_req_task_work_add(struct io_kiocb *req)
2051{
2052	struct task_struct *tsk = req->task;
2053	struct io_uring_task *tctx = tsk->io_uring;
2054	enum task_work_notify_mode notify;
2055	struct io_wq_work_node *node;
2056	unsigned long flags;
2057	bool running;
2058
2059	WARN_ON_ONCE(!tctx);
2060
2061	spin_lock_irqsave(&tctx->task_lock, flags);
2062	wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2063	running = tctx->task_running;
2064	if (!running)
2065		tctx->task_running = true;
2066	spin_unlock_irqrestore(&tctx->task_lock, flags);
2067
2068	/* task_work already pending, we're done */
2069	if (running)
2070		return;
2071
2072	/*
2073	 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2074	 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2075	 * processing task_work. There's no reliable way to tell if TWA_RESUME
2076	 * will do the job.
2077	 */
2078	notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2079	if (!task_work_add(tsk, &tctx->task_work, notify)) {
2080		wake_up_process(tsk);
2081		return;
2082	}
2083
2084	spin_lock_irqsave(&tctx->task_lock, flags);
2085	tctx->task_running = false;
2086	node = tctx->task_list.first;
2087	INIT_WQ_LIST(&tctx->task_list);
2088	spin_unlock_irqrestore(&tctx->task_lock, flags);
2089
2090	while (node) {
2091		req = container_of(node, struct io_kiocb, io_task_work.node);
2092		node = node->next;
2093		if (llist_add(&req->io_task_work.fallback_node,
2094			      &req->ctx->fallback_llist))
2095			schedule_delayed_work(&req->ctx->fallback_work, 1);
2096	}
2097}
2098
2099static void io_req_task_cancel(struct io_kiocb *req)
2100{
2101	struct io_ring_ctx *ctx = req->ctx;
2102
2103	/* ctx is guaranteed to stay alive while we hold uring_lock */
2104	mutex_lock(&ctx->uring_lock);
2105	io_req_complete_failed(req, req->result);
2106	mutex_unlock(&ctx->uring_lock);
2107}
2108
2109static void io_req_task_submit(struct io_kiocb *req)
2110{
2111	struct io_ring_ctx *ctx = req->ctx;
2112
2113	/* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2114	mutex_lock(&ctx->uring_lock);
2115	if (likely(!(req->task->flags & PF_EXITING)))
2116		__io_queue_sqe(req);
2117	else
2118		io_req_complete_failed(req, -EFAULT);
2119	mutex_unlock(&ctx->uring_lock);
2120}
2121
2122static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2123{
2124	req->result = ret;
2125	req->io_task_work.func = io_req_task_cancel;
2126	io_req_task_work_add(req);
2127}
2128
2129static void io_req_task_queue(struct io_kiocb *req)
2130{
2131	req->io_task_work.func = io_req_task_submit;
2132	io_req_task_work_add(req);
2133}
2134
2135static void io_req_task_queue_reissue(struct io_kiocb *req)
2136{
2137	req->io_task_work.func = io_queue_async_work;
2138	io_req_task_work_add(req);
2139}
2140
2141static inline void io_queue_next(struct io_kiocb *req)
2142{
2143	struct io_kiocb *nxt = io_req_find_next(req);
2144
2145	if (nxt)
2146		io_req_task_queue(nxt);
2147}
2148
2149static void io_free_req(struct io_kiocb *req)
2150{
2151	io_queue_next(req);
2152	__io_free_req(req);
2153}
2154
2155struct req_batch {
2156	struct task_struct	*task;
2157	int			task_refs;
2158	int			ctx_refs;
2159};
2160
2161static inline void io_init_req_batch(struct req_batch *rb)
2162{
2163	rb->task_refs = 0;
2164	rb->ctx_refs = 0;
2165	rb->task = NULL;
2166}
2167
2168static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2169				     struct req_batch *rb)
2170{
2171	if (rb->ctx_refs)
2172		percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2173	if (rb->task == current)
2174		current->io_uring->cached_refs += rb->task_refs;
2175	else if (rb->task)
2176		io_put_task(rb->task, rb->task_refs);
2177}
2178
2179static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2180			      struct io_submit_state *state)
2181{
2182	io_queue_next(req);
2183	io_dismantle_req(req);
2184
2185	if (req->task != rb->task) {
2186		if (rb->task)
2187			io_put_task(rb->task, rb->task_refs);
2188		rb->task = req->task;
2189		rb->task_refs = 0;
2190	}
2191	rb->task_refs++;
2192	rb->ctx_refs++;
2193
2194	if (state->free_reqs != ARRAY_SIZE(state->reqs))
2195		state->reqs[state->free_reqs++] = req;
2196	else
2197		list_add(&req->inflight_entry, &state->free_list);
2198}
2199
2200static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2201	__must_hold(&ctx->uring_lock)
2202{
2203	struct io_submit_state *state = &ctx->submit_state;
2204	int i, nr = state->compl_nr;
2205	struct req_batch rb;
2206
2207	spin_lock(&ctx->completion_lock);
2208	for (i = 0; i < nr; i++) {
2209		struct io_kiocb *req = state->compl_reqs[i];
2210
2211		__io_cqring_fill_event(ctx, req->user_data, req->result,
2212					req->compl.cflags);
2213	}
2214	io_commit_cqring(ctx);
2215	spin_unlock(&ctx->completion_lock);
2216	io_cqring_ev_posted(ctx);
2217
2218	io_init_req_batch(&rb);
2219	for (i = 0; i < nr; i++) {
2220		struct io_kiocb *req = state->compl_reqs[i];
2221
2222		if (req_ref_put_and_test(req))
2223			io_req_free_batch(&rb, req, &ctx->submit_state);
2224	}
2225
2226	io_req_free_batch_finish(ctx, &rb);
2227	state->compl_nr = 0;
2228}
2229
2230/*
2231 * Drop reference to request, return next in chain (if there is one) if this
2232 * was the last reference to this request.
2233 */
2234static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2235{
2236	struct io_kiocb *nxt = NULL;
2237
2238	if (req_ref_put_and_test(req)) {
2239		nxt = io_req_find_next(req);
2240		__io_free_req(req);
2241	}
2242	return nxt;
2243}
2244
2245static inline void io_put_req(struct io_kiocb *req)
2246{
2247	if (req_ref_put_and_test(req))
2248		io_free_req(req);
2249}
2250
2251static inline void io_put_req_deferred(struct io_kiocb *req)
2252{
2253	if (req_ref_put_and_test(req)) {
2254		req->io_task_work.func = io_free_req;
2255		io_req_task_work_add(req);
2256	}
2257}
2258
2259static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2260{
2261	/* See comment at the top of this file */
2262	smp_rmb();
2263	return __io_cqring_events(ctx);
2264}
2265
2266static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2267{
2268	struct io_rings *rings = ctx->rings;
2269
2270	/* make sure SQ entry isn't read before tail */
2271	return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2272}
2273
2274static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2275{
2276	unsigned int cflags;
2277
2278	cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2279	cflags |= IORING_CQE_F_BUFFER;
2280	req->flags &= ~REQ_F_BUFFER_SELECTED;
2281	kfree(kbuf);
2282	return cflags;
2283}
2284
2285static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2286{
2287	struct io_buffer *kbuf;
2288
2289	if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
2290		return 0;
2291	kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2292	return io_put_kbuf(req, kbuf);
2293}
2294
2295static inline bool io_run_task_work(void)
2296{
2297	if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2298		__set_current_state(TASK_RUNNING);
2299		tracehook_notify_signal();
2300		return true;
2301	}
2302
2303	return false;
2304}
2305
2306/*
2307 * Find and free completed poll iocbs
2308 */
2309static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2310			       struct list_head *done)
2311{
2312	struct req_batch rb;
2313	struct io_kiocb *req;
2314
2315	/* order with ->result store in io_complete_rw_iopoll() */
2316	smp_rmb();
2317
2318	io_init_req_batch(&rb);
2319	while (!list_empty(done)) {
2320		req = list_first_entry(done, struct io_kiocb, inflight_entry);
2321		list_del(&req->inflight_entry);
2322
2323		if (READ_ONCE(req->result) == -EAGAIN &&
2324		    !(req->flags & REQ_F_DONT_REISSUE)) {
2325			req->iopoll_completed = 0;
2326			io_req_task_queue_reissue(req);
2327			continue;
2328		}
2329
2330		__io_cqring_fill_event(ctx, req->user_data, req->result,
2331					io_put_rw_kbuf(req));
2332		(*nr_events)++;
2333
2334		if (req_ref_put_and_test(req))
2335			io_req_free_batch(&rb, req, &ctx->submit_state);
2336	}
2337
2338	io_commit_cqring(ctx);
2339	io_cqring_ev_posted_iopoll(ctx);
2340	io_req_free_batch_finish(ctx, &rb);
2341}
2342
2343static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2344			long min)
2345{
2346	struct io_kiocb *req, *tmp;
2347	LIST_HEAD(done);
2348	bool spin;
2349
2350	/*
2351	 * Only spin for completions if we don't have multiple devices hanging
2352	 * off our complete list, and we're under the requested amount.
2353	 */
2354	spin = !ctx->poll_multi_queue && *nr_events < min;
2355
2356	list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2357		struct kiocb *kiocb = &req->rw.kiocb;
2358		int ret;
2359
2360		/*
2361		 * Move completed and retryable entries to our local lists.
2362		 * If we find a request that requires polling, break out
2363		 * and complete those lists first, if we have entries there.
2364		 */
2365		if (READ_ONCE(req->iopoll_completed)) {
2366			list_move_tail(&req->inflight_entry, &done);
2367			continue;
2368		}
2369		if (!list_empty(&done))
2370			break;
2371
2372		ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2373		if (unlikely(ret < 0))
2374			return ret;
2375		else if (ret)
2376			spin = false;
2377
2378		/* iopoll may have completed current req */
2379		if (READ_ONCE(req->iopoll_completed))
2380			list_move_tail(&req->inflight_entry, &done);
2381	}
2382
2383	if (!list_empty(&done))
2384		io_iopoll_complete(ctx, nr_events, &done);
2385
2386	return 0;
2387}
2388
2389/*
2390 * We can't just wait for polled events to come to us, we have to actively
2391 * find and complete them.
2392 */
2393static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2394{
2395	if (!(ctx->flags & IORING_SETUP_IOPOLL))
2396		return;
2397
2398	mutex_lock(&ctx->uring_lock);
2399	while (!list_empty(&ctx->iopoll_list)) {
2400		unsigned int nr_events = 0;
2401
2402		io_do_iopoll(ctx, &nr_events, 0);
2403
2404		/* let it sleep and repeat later if can't complete a request */
2405		if (nr_events == 0)
2406			break;
2407		/*
2408		 * Ensure we allow local-to-the-cpu processing to take place,
2409		 * in this case we need to ensure that we reap all events.
2410		 * Also let task_work, etc. to progress by releasing the mutex
2411		 */
2412		if (need_resched()) {
2413			mutex_unlock(&ctx->uring_lock);
2414			cond_resched();
2415			mutex_lock(&ctx->uring_lock);
2416		}
2417	}
2418	mutex_unlock(&ctx->uring_lock);
2419}
2420
2421static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2422{
2423	unsigned int nr_events = 0;
2424	int ret = 0;
2425
2426	/*
2427	 * We disallow the app entering submit/complete with polling, but we
2428	 * still need to lock the ring to prevent racing with polled issue
2429	 * that got punted to a workqueue.
2430	 */
2431	mutex_lock(&ctx->uring_lock);
2432	/*
2433	 * Don't enter poll loop if we already have events pending.
2434	 * If we do, we can potentially be spinning for commands that
2435	 * already triggered a CQE (eg in error).
2436	 */
2437	if (test_bit(0, &ctx->check_cq_overflow))
2438		__io_cqring_overflow_flush(ctx, false);
2439	if (io_cqring_events(ctx))
2440		goto out;
2441	do {
2442		/*
2443		 * If a submit got punted to a workqueue, we can have the
2444		 * application entering polling for a command before it gets
2445		 * issued. That app will hold the uring_lock for the duration
2446		 * of the poll right here, so we need to take a breather every
2447		 * now and then to ensure that the issue has a chance to add
2448		 * the poll to the issued list. Otherwise we can spin here
2449		 * forever, while the workqueue is stuck trying to acquire the
2450		 * very same mutex.
2451		 */
2452		if (list_empty(&ctx->iopoll_list)) {
2453			u32 tail = ctx->cached_cq_tail;
2454
2455			mutex_unlock(&ctx->uring_lock);
2456			io_run_task_work();
2457			mutex_lock(&ctx->uring_lock);
2458
2459			/* some requests don't go through iopoll_list */
2460			if (tail != ctx->cached_cq_tail ||
2461			    list_empty(&ctx->iopoll_list))
2462				break;
2463		}
2464		ret = io_do_iopoll(ctx, &nr_events, min);
2465	} while (!ret && nr_events < min && !need_resched());
2466out:
2467	mutex_unlock(&ctx->uring_lock);
2468	return ret;
2469}
2470
2471static void kiocb_end_write(struct io_kiocb *req)
2472{
2473	/*
2474	 * Tell lockdep we inherited freeze protection from submission
2475	 * thread.
2476	 */
2477	if (req->flags & REQ_F_ISREG) {
2478		struct super_block *sb = file_inode(req->file)->i_sb;
2479
2480		__sb_writers_acquired(sb, SB_FREEZE_WRITE);
2481		sb_end_write(sb);
2482	}
2483}
2484
2485#ifdef CONFIG_BLOCK
2486static bool io_resubmit_prep(struct io_kiocb *req)
2487{
2488	struct io_async_rw *rw = req->async_data;
2489
2490	if (!rw)
2491		return !io_req_prep_async(req);
2492	/* may have left rw->iter inconsistent on -EIOCBQUEUED */
2493	iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2494	return true;
2495}
2496
2497static bool io_rw_should_reissue(struct io_kiocb *req)
2498{
2499	umode_t mode = file_inode(req->file)->i_mode;
2500	struct io_ring_ctx *ctx = req->ctx;
2501
2502	if (!S_ISBLK(mode) && !S_ISREG(mode))
2503		return false;
2504	if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2505	    !(ctx->flags & IORING_SETUP_IOPOLL)))
2506		return false;
2507	/*
2508	 * If ref is dying, we might be running poll reap from the exit work.
2509	 * Don't attempt to reissue from that path, just let it fail with
2510	 * -EAGAIN.
2511	 */
2512	if (percpu_ref_is_dying(&ctx->refs))
2513		return false;
2514	/*
2515	 * Play it safe and assume not safe to re-import and reissue if we're
2516	 * not in the original thread group (or in task context).
2517	 */
2518	if (!same_thread_group(req->task, current) || !in_task())
2519		return false;
2520	return true;
2521}
2522#else
2523static bool io_resubmit_prep(struct io_kiocb *req)
2524{
2525	return false;
2526}
2527static bool io_rw_should_reissue(struct io_kiocb *req)
2528{
2529	return false;
2530}
2531#endif
2532
2533static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2534{
2535	if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2536		kiocb_end_write(req);
2537	if (res != req->result) {
2538		if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2539		    io_rw_should_reissue(req)) {
2540			req->flags |= REQ_F_REISSUE;
2541			return true;
2542		}
2543		req_set_fail(req);
2544		req->result = res;
2545	}
2546	return false;
2547}
2548
2549static void io_req_task_complete(struct io_kiocb *req)
2550{
2551	__io_req_complete(req, 0, req->result, io_put_rw_kbuf(req));
2552}
2553
2554static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2555			     unsigned int issue_flags)
2556{
2557	if (__io_complete_rw_common(req, res))
2558		return;
2559	io_req_task_complete(req);
2560}
2561
2562static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2563{
2564	struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2565
2566	if (__io_complete_rw_common(req, res))
2567		return;
2568	req->result = res;
2569	req->io_task_work.func = io_req_task_complete;
2570	io_req_task_work_add(req);
2571}
2572
2573static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2574{
2575	struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2576
2577	if (kiocb->ki_flags & IOCB_WRITE)
2578		kiocb_end_write(req);
2579	if (unlikely(res != req->result)) {
2580		if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2581		    io_resubmit_prep(req))) {
2582			req_set_fail(req);
2583			req->flags |= REQ_F_DONT_REISSUE;
2584		}
2585	}
2586
2587	WRITE_ONCE(req->result, res);
2588	/* order with io_iopoll_complete() checking ->result */
2589	smp_wmb();
2590	WRITE_ONCE(req->iopoll_completed, 1);
2591}
2592
2593/*
2594 * After the iocb has been issued, it's safe to be found on the poll list.
2595 * Adding the kiocb to the list AFTER submission ensures that we don't
2596 * find it from a io_do_iopoll() thread before the issuer is done
2597 * accessing the kiocb cookie.
2598 */
2599static void io_iopoll_req_issued(struct io_kiocb *req)
2600{
2601	struct io_ring_ctx *ctx = req->ctx;
2602	const bool in_async = io_wq_current_is_worker();
2603
2604	/* workqueue context doesn't hold uring_lock, grab it now */
2605	if (unlikely(in_async))
2606		mutex_lock(&ctx->uring_lock);
2607
2608	/*
2609	 * Track whether we have multiple files in our lists. This will impact
2610	 * how we do polling eventually, not spinning if we're on potentially
2611	 * different devices.
2612	 */
2613	if (list_empty(&ctx->iopoll_list)) {
2614		ctx->poll_multi_queue = false;
2615	} else if (!ctx->poll_multi_queue) {
2616		struct io_kiocb *list_req;
2617		unsigned int queue_num0, queue_num1;
2618
2619		list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2620						inflight_entry);
2621
2622		if (list_req->file != req->file) {
2623			ctx->poll_multi_queue = true;
2624		} else {
2625			queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2626			queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2627			if (queue_num0 != queue_num1)
2628				ctx->poll_multi_queue = true;
2629		}
2630	}
2631
2632	/*
2633	 * For fast devices, IO may have already completed. If it has, add
2634	 * it to the front so we find it first.
2635	 */
2636	if (READ_ONCE(req->iopoll_completed))
2637		list_add(&req->inflight_entry, &ctx->iopoll_list);
2638	else
2639		list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2640
2641	if (unlikely(in_async)) {
2642		/*
2643		 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2644		 * in sq thread task context or in io worker task context. If
2645		 * current task context is sq thread, we don't need to check
2646		 * whether should wake up sq thread.
2647		 */
2648		if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2649		    wq_has_sleeper(&ctx->sq_data->wait))
2650			wake_up(&ctx->sq_data->wait);
2651
2652		mutex_unlock(&ctx->uring_lock);
2653	}
2654}
2655
2656static bool io_bdev_nowait(struct block_device *bdev)
2657{
2658	return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2659}
2660
2661/*
2662 * If we tracked the file through the SCM inflight mechanism, we could support
2663 * any file. For now, just ensure that anything potentially problematic is done
2664 * inline.
2665 */
2666static bool __io_file_supports_nowait(struct file *file, int rw)
2667{
2668	umode_t mode = file_inode(file)->i_mode;
2669
2670	if (S_ISBLK(mode)) {
2671		if (IS_ENABLED(CONFIG_BLOCK) &&
2672		    io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2673			return true;
2674		return false;
2675	}
2676	if (S_ISSOCK(mode))
2677		return true;
2678	if (S_ISREG(mode)) {
2679		if (IS_ENABLED(CONFIG_BLOCK) &&
2680		    io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2681		    file->f_op != &io_uring_fops)
2682			return true;
2683		return false;
2684	}
2685
2686	/* any ->read/write should understand O_NONBLOCK */
2687	if (file->f_flags & O_NONBLOCK)
2688		return true;
2689
2690	if (!(file->f_mode & FMODE_NOWAIT))
2691		return false;
2692
2693	if (rw == READ)
2694		return file->f_op->read_iter != NULL;
2695
2696	return file->f_op->write_iter != NULL;
2697}
2698
2699static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2700{
2701	if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2702		return true;
2703	else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2704		return true;
2705
2706	return __io_file_supports_nowait(req->file, rw);
2707}
2708
2709static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2710{
2711	struct io_ring_ctx *ctx = req->ctx;
2712	struct kiocb *kiocb = &req->rw.kiocb;
2713	struct file *file = req->file;
2714	unsigned ioprio;
2715	int ret;
2716
2717	if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2718		req->flags |= REQ_F_ISREG;
2719
2720	kiocb->ki_pos = READ_ONCE(sqe->off);
2721	if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2722		req->flags |= REQ_F_CUR_POS;
2723		kiocb->ki_pos = file->f_pos;
2724	}
2725	kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2726	kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2727	ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2728	if (unlikely(ret))
2729		return ret;
2730
2731	/* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2732	if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2733		req->flags |= REQ_F_NOWAIT;
2734
2735	ioprio = READ_ONCE(sqe->ioprio);
2736	if (ioprio) {
2737		ret = ioprio_check_cap(ioprio);
2738		if (ret)
2739			return ret;
2740
2741		kiocb->ki_ioprio = ioprio;
2742	} else
2743		kiocb->ki_ioprio = get_current_ioprio();
2744
2745	if (ctx->flags & IORING_SETUP_IOPOLL) {
2746		if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2747		    !kiocb->ki_filp->f_op->iopoll)
2748			return -EOPNOTSUPP;
2749
2750		kiocb->ki_flags |= IOCB_HIPRI;
2751		kiocb->ki_complete = io_complete_rw_iopoll;
2752		req->iopoll_completed = 0;
2753	} else {
2754		if (kiocb->ki_flags & IOCB_HIPRI)
2755			return -EINVAL;
2756		kiocb->ki_complete = io_complete_rw;
2757	}
2758
2759	if (req->opcode == IORING_OP_READ_FIXED ||
2760	    req->opcode == IORING_OP_WRITE_FIXED) {
2761		req->imu = NULL;
2762		io_req_set_rsrc_node(req);
2763	}
2764
2765	req->rw.addr = READ_ONCE(sqe->addr);
2766	req->rw.len = READ_ONCE(sqe->len);
2767	req->buf_index = READ_ONCE(sqe->buf_index);
2768	return 0;
2769}
2770
2771static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2772{
2773	switch (ret) {
2774	case -EIOCBQUEUED:
2775		break;
2776	case -ERESTARTSYS:
2777	case -ERESTARTNOINTR:
2778	case -ERESTARTNOHAND:
2779	case -ERESTART_RESTARTBLOCK:
2780		/*
2781		 * We can't just restart the syscall, since previously
2782		 * submitted sqes may already be in progress. Just fail this
2783		 * IO with EINTR.
2784		 */
2785		ret = -EINTR;
2786		fallthrough;
2787	default:
2788		kiocb->ki_complete(kiocb, ret, 0);
2789	}
2790}
2791
2792static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2793		       unsigned int issue_flags)
2794{
2795	struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2796	struct io_async_rw *io = req->async_data;
2797	bool check_reissue = kiocb->ki_complete == io_complete_rw;
2798
2799	/* add previously done IO, if any */
2800	if (io && io->bytes_done > 0) {
2801		if (ret < 0)
2802			ret = io->bytes_done;
2803		else
2804			ret += io->bytes_done;
2805	}
2806
2807	if (req->flags & REQ_F_CUR_POS)
2808		req->file->f_pos = kiocb->ki_pos;
2809	if (ret >= 0 && check_reissue)
2810		__io_complete_rw(req, ret, 0, issue_flags);
2811	else
2812		io_rw_done(kiocb, ret);
2813
2814	if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2815		req->flags &= ~REQ_F_REISSUE;
2816		if (io_resubmit_prep(req)) {
2817			io_req_task_queue_reissue(req);
2818		} else {
2819			req_set_fail(req);
2820			__io_req_complete(req, issue_flags, ret,
2821					  io_put_rw_kbuf(req));
2822		}
2823	}
2824}
2825
2826static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2827			     struct io_mapped_ubuf *imu)
2828{
2829	size_t len = req->rw.len;
2830	u64 buf_end, buf_addr = req->rw.addr;
2831	size_t offset;
2832
2833	if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2834		return -EFAULT;
2835	/* not inside the mapped region */
2836	if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2837		return -EFAULT;
2838
2839	/*
2840	 * May not be a start of buffer, set size appropriately
2841	 * and advance us to the beginning.
2842	 */
2843	offset = buf_addr - imu->ubuf;
2844	iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2845
2846	if (offset) {
2847		/*
2848		 * Don't use iov_iter_advance() here, as it's really slow for
2849		 * using the latter parts of a big fixed buffer - it iterates
2850		 * over each segment manually. We can cheat a bit here, because
2851		 * we know that:
2852		 *
2853		 * 1) it's a BVEC iter, we set it up
2854		 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2855		 *    first and last bvec
2856		 *
2857		 * So just find our index, and adjust the iterator afterwards.
2858		 * If the offset is within the first bvec (or the whole first
2859		 * bvec, just use iov_iter_advance(). This makes it easier
2860		 * since we can just skip the first segment, which may not
2861		 * be PAGE_SIZE aligned.
2862		 */
2863		const struct bio_vec *bvec = imu->bvec;
2864
2865		if (offset <= bvec->bv_len) {
2866			iov_iter_advance(iter, offset);
2867		} else {
2868			unsigned long seg_skip;
2869
2870			/* skip first vec */
2871			offset -= bvec->bv_len;
2872			seg_skip = 1 + (offset >> PAGE_SHIFT);
2873
2874			iter->bvec = bvec + seg_skip;
2875			iter->nr_segs -= seg_skip;
2876			iter->count -= bvec->bv_len + offset;
2877			iter->iov_offset = offset & ~PAGE_MASK;
2878		}
2879	}
2880
2881	return 0;
2882}
2883
2884static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2885{
2886	struct io_ring_ctx *ctx = req->ctx;
2887	struct io_mapped_ubuf *imu = req->imu;
2888	u16 index, buf_index = req->buf_index;
2889
2890	if (likely(!imu)) {
2891		if (unlikely(buf_index >= ctx->nr_user_bufs))
2892			return -EFAULT;
2893		index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2894		imu = READ_ONCE(ctx->user_bufs[index]);
2895		req->imu = imu;
2896	}
2897	return __io_import_fixed(req, rw, iter, imu);
2898}
2899
2900static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2901{
2902	if (needs_lock)
2903		mutex_unlock(&ctx->uring_lock);
2904}
2905
2906static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2907{
2908	/*
2909	 * "Normal" inline submissions always hold the uring_lock, since we
2910	 * grab it from the system call. Same is true for the SQPOLL offload.
2911	 * The only exception is when we've detached the request and issue it
2912	 * from an async worker thread, grab the lock for that case.
2913	 */
2914	if (needs_lock)
2915		mutex_lock(&ctx->uring_lock);
2916}
2917
2918static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2919					  int bgid, struct io_buffer *kbuf,
2920					  bool needs_lock)
2921{
2922	struct io_buffer *head;
2923
2924	if (req->flags & REQ_F_BUFFER_SELECTED)
2925		return kbuf;
2926
2927	io_ring_submit_lock(req->ctx, needs_lock);
2928
2929	lockdep_assert_held(&req->ctx->uring_lock);
2930
2931	head = xa_load(&req->ctx->io_buffers, bgid);
2932	if (head) {
2933		if (!list_empty(&head->list)) {
2934			kbuf = list_last_entry(&head->list, struct io_buffer,
2935							list);
2936			list_del(&kbuf->list);
2937		} else {
2938			kbuf = head;
2939			xa_erase(&req->ctx->io_buffers, bgid);
2940		}
2941		if (*len > kbuf->len)
2942			*len = kbuf->len;
2943	} else {
2944		kbuf = ERR_PTR(-ENOBUFS);
2945	}
2946
2947	io_ring_submit_unlock(req->ctx, needs_lock);
2948
2949	return kbuf;
2950}
2951
2952static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2953					bool needs_lock)
2954{
2955	struct io_buffer *kbuf;
2956	u16 bgid;
2957
2958	kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2959	bgid = req->buf_index;
2960	kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2961	if (IS_ERR(kbuf))
2962		return kbuf;
2963	req->rw.addr = (u64) (unsigned long) kbuf;
2964	req->flags |= REQ_F_BUFFER_SELECTED;
2965	return u64_to_user_ptr(kbuf->addr);
2966}
2967
2968#ifdef CONFIG_COMPAT
2969static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2970				bool needs_lock)
2971{
2972	struct compat_iovec __user *uiov;
2973	compat_ssize_t clen;
2974	void __user *buf;
2975	ssize_t len;
2976
2977	uiov = u64_to_user_ptr(req->rw.addr);
2978	if (!access_ok(uiov, sizeof(*uiov)))
2979		return -EFAULT;
2980	if (__get_user(clen, &uiov->iov_len))
2981		return -EFAULT;
2982	if (clen < 0)
2983		return -EINVAL;
2984
2985	len = clen;
2986	buf = io_rw_buffer_select(req, &len, needs_lock);
2987	if (IS_ERR(buf))
2988		return PTR_ERR(buf);
2989	iov[0].iov_base = buf;
2990	iov[0].iov_len = (compat_size_t) len;
2991	return 0;
2992}
2993#endif
2994
2995static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2996				      bool needs_lock)
2997{
2998	struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2999	void __user *buf;
3000	ssize_t len;
3001
3002	if (copy_from_user(iov, uiov, sizeof(*uiov)))
3003		return -EFAULT;
3004
3005	len = iov[0].iov_len;
3006	if (len < 0)
3007		return -EINVAL;
3008	buf = io_rw_buffer_select(req, &len, needs_lock);
3009	if (IS_ERR(buf))
3010		return PTR_ERR(buf);
3011	iov[0].iov_base = buf;
3012	iov[0].iov_len = len;
3013	return 0;
3014}
3015
3016static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3017				    bool needs_lock)
3018{
3019	if (req->flags & REQ_F_BUFFER_SELECTED) {
3020		struct io_buffer *kbuf;
3021
3022		kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3023		iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3024		iov[0].iov_len = kbuf->len;
3025		return 0;
3026	}
3027	if (req->rw.len != 1)
3028		return -EINVAL;
3029
3030#ifdef CONFIG_COMPAT
3031	if (req->ctx->compat)
3032		return io_compat_import(req, iov, needs_lock);
3033#endif
3034
3035	return __io_iov_buffer_select(req, iov, needs_lock);
3036}
3037
3038static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3039			   struct iov_iter *iter, bool needs_lock)
3040{
3041	void __user *buf = u64_to_user_ptr(req->rw.addr);
3042	size_t sqe_len = req->rw.len;
3043	u8 opcode = req->opcode;
3044	ssize_t ret;
3045
3046	if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3047		*iovec = NULL;
3048		return io_import_fixed(req, rw, iter);
3049	}
3050
3051	/* buffer index only valid with fixed read/write, or buffer select  */
3052	if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3053		return -EINVAL;
3054
3055	if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3056		if (req->flags & REQ_F_BUFFER_SELECT) {
3057			buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3058			if (IS_ERR(buf))
3059				return PTR_ERR(buf);
3060			req->rw.len = sqe_len;
3061		}
3062
3063		ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3064		*iovec = NULL;
3065		return ret;
3066	}
3067
3068	if (req->flags & REQ_F_BUFFER_SELECT) {
3069		ret = io_iov_buffer_select(req, *iovec, needs_lock);
3070		if (!ret)
3071			iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3072		*iovec = NULL;
3073		return ret;
3074	}
3075
3076	return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3077			      req->ctx->compat);
3078}
3079
3080static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3081{
3082	return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3083}
3084
3085/*
3086 * For files that don't have ->read_iter() and ->write_iter(), handle them
3087 * by looping over ->read() or ->write() manually.
3088 */
3089static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3090{
3091	struct kiocb *kiocb = &req->rw.kiocb;
3092	struct file *file = req->file;
3093	ssize_t ret = 0;
3094
3095	/*
3096	 * Don't support polled IO through this interface, and we can't
3097	 * support non-blocking either. For the latter, this just causes
3098	 * the kiocb to be handled from an async context.
3099	 */
3100	if (kiocb->ki_flags & IOCB_HIPRI)
3101		return -EOPNOTSUPP;
3102	if (kiocb->ki_flags & IOCB_NOWAIT)
3103		return -EAGAIN;
3104
3105	while (iov_iter_count(iter)) {
3106		struct iovec iovec;
3107		ssize_t nr;
3108
3109		if (!iov_iter_is_bvec(iter)) {
3110			iovec = iov_iter_iovec(iter);
3111		} else {
3112			iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3113			iovec.iov_len = req->rw.len;
3114		}
3115
3116		if (rw == READ) {
3117			nr = file->f_op->read(file, iovec.iov_base,
3118					      iovec.iov_len, io_kiocb_ppos(kiocb));
3119		} else {
3120			nr = file->f_op->write(file, iovec.iov_base,
3121					       iovec.iov_len, io_kiocb_ppos(kiocb));
3122		}
3123
3124		if (nr < 0) {
3125			if (!ret)
3126				ret = nr;
3127			break;
3128		}
3129		ret += nr;
3130		if (nr != iovec.iov_len)
3131			break;
3132		req->rw.len -= nr;
3133		req->rw.addr += nr;
3134		iov_iter_advance(iter, nr);
3135	}
3136
3137	return ret;
3138}
3139
3140static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3141			  const struct iovec *fast_iov, struct iov_iter *iter)
3142{
3143	struct io_async_rw *rw = req->async_data;
3144
3145	memcpy(&rw->iter, iter, sizeof(*iter));
3146	rw->free_iovec = iovec;
3147	rw->bytes_done = 0;
3148	/* can only be fixed buffers, no need to do anything */
3149	if (iov_iter_is_bvec(iter))
3150		return;
3151	if (!iovec) {
3152		unsigned iov_off = 0;
3153
3154		rw->iter.iov = rw->fast_iov;
3155		if (iter->iov != fast_iov) {
3156			iov_off = iter->iov - fast_iov;
3157			rw->iter.iov += iov_off;
3158		}
3159		if (rw->fast_iov != fast_iov)
3160			memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3161			       sizeof(struct iovec) * iter->nr_segs);
3162	} else {
3163		req->flags |= REQ_F_NEED_CLEANUP;
3164	}
3165}
3166
3167static inline int io_alloc_async_data(struct io_kiocb *req)
3168{
3169	WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3170	req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3171	return req->async_data == NULL;
3172}
3173
3174static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3175			     const struct iovec *fast_iov,
3176			     struct iov_iter *iter, bool force)
3177{
3178	if (!force && !io_op_defs[req->opcode].needs_async_setup)
3179		return 0;
3180	if (!req->async_data) {
3181		if (io_alloc_async_data(req)) {
3182			kfree(iovec);
3183			return -ENOMEM;
3184		}
3185
3186		io_req_map_rw(req, iovec, fast_iov, iter);
3187	}
3188	return 0;
3189}
3190
3191static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3192{
3193	struct io_async_rw *iorw = req->async_data;
3194	struct iovec *iov = iorw->fast_iov;
3195	int ret;
3196
3197	ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3198	if (unlikely(ret < 0))
3199		return ret;
3200
3201	iorw->bytes_done = 0;
3202	iorw->free_iovec = iov;
3203	if (iov)
3204		req->flags |= REQ_F_NEED_CLEANUP;
3205	return 0;
3206}
3207
3208static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3209{
3210	if (unlikely(!(req->file->f_mode & FMODE_READ)))
3211		return -EBADF;
3212	return io_prep_rw(req, sqe);
3213}
3214
3215/*
3216 * This is our waitqueue callback handler, registered through lock_page_async()
3217 * when we initially tried to do the IO with the iocb armed our waitqueue.
3218 * This gets called when the page is unlocked, and we generally expect that to
3219 * happen when the page IO is completed and the page is now uptodate. This will
3220 * queue a task_work based retry of the operation, attempting to copy the data
3221 * again. If the latter fails because the page was NOT uptodate, then we will
3222 * do a thread based blocking retry of the operation. That's the unexpected
3223 * slow path.
3224 */
3225static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3226			     int sync, void *arg)
3227{
3228	struct wait_page_queue *wpq;
3229	struct io_kiocb *req = wait->private;
3230	struct wait_page_key *key = arg;
3231
3232	wpq = container_of(wait, struct wait_page_queue, wait);
3233
3234	if (!wake_page_match(wpq, key))
3235		return 0;
3236
3237	req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3238	list_del_init(&wait->entry);
3239	io_req_task_queue(req);
3240	return 1;
3241}
3242
3243/*
3244 * This controls whether a given IO request should be armed for async page
3245 * based retry. If we return false here, the request is handed to the async
3246 * worker threads for retry. If we're doing buffered reads on a regular file,
3247 * we prepare a private wait_page_queue entry and retry the operation. This
3248 * will either succeed because the page is now uptodate and unlocked, or it
3249 * will register a callback when the page is unlocked at IO completion. Through
3250 * that callback, io_uring uses task_work to setup a retry of the operation.
3251 * That retry will attempt the buffered read again. The retry will generally
3252 * succeed, or in rare cases where it fails, we then fall back to using the
3253 * async worker threads for a blocking retry.
3254 */
3255static bool io_rw_should_retry(struct io_kiocb *req)
3256{
3257	struct io_async_rw *rw = req->async_data;
3258	struct wait_page_queue *wait = &rw->wpq;
3259	struct kiocb *kiocb = &req->rw.kiocb;
3260
3261	/* never retry for NOWAIT, we just complete with -EAGAIN */
3262	if (req->flags & REQ_F_NOWAIT)
3263		return false;
3264
3265	/* Only for buffered IO */
3266	if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3267		return false;
3268
3269	/*
3270	 * just use poll if we can, and don't attempt if the fs doesn't
3271	 * support callback based unlocks
3272	 */
3273	if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3274		return false;
3275
3276	wait->wait.func = io_async_buf_func;
3277	wait->wait.private = req;
3278	wait->wait.flags = 0;
3279	INIT_LIST_HEAD(&wait->wait.entry);
3280	kiocb->ki_flags |= IOCB_WAITQ;
3281	kiocb->ki_flags &= ~IOCB_NOWAIT;
3282	kiocb->ki_waitq = wait;
3283	return true;
3284}
3285
3286static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3287{
3288	if (req->file->f_op->read_iter)
3289		return call_read_iter(req->file, &req->rw.kiocb, iter);
3290	else if (req->file->f_op->read)
3291		return loop_rw_iter(READ, req, iter);
3292	else
3293		return -EINVAL;
3294}
3295
3296static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3297{
3298	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3299	struct kiocb *kiocb = &req->rw.kiocb;
3300	struct iov_iter __iter, *iter = &__iter;
3301	struct io_async_rw *rw = req->async_data;
3302	ssize_t io_size, ret, ret2;
3303	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3304
3305	if (rw) {
3306		iter = &rw->iter;
3307		iovec = NULL;
3308	} else {
3309		ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3310		if (ret < 0)
3311			return ret;
3312	}
3313	io_size = iov_iter_count(iter);
3314	req->result = io_size;
3315
3316	/* Ensure we clear previously set non-block flag */
3317	if (!force_nonblock)
3318		kiocb->ki_flags &= ~IOCB_NOWAIT;
3319	else
3320		kiocb->ki_flags |= IOCB_NOWAIT;
3321
3322	/* If the file doesn't support async, just async punt */
3323	if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3324		ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3325		return ret ?: -EAGAIN;
3326	}
3327
3328	ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3329	if (unlikely(ret)) {
3330		kfree(iovec);
3331		return ret;
3332	}
3333
3334	ret = io_iter_do_read(req, iter);
3335
3336	if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3337		req->flags &= ~REQ_F_REISSUE;
3338		/* IOPOLL retry should happen for io-wq threads */
3339		if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3340			goto done;
3341		/* no retry on NONBLOCK nor RWF_NOWAIT */
3342		if (req->flags & REQ_F_NOWAIT)
3343			goto done;
3344		/* some cases will consume bytes even on error returns */
3345		iov_iter_revert(iter, io_size - iov_iter_count(iter));
3346		ret = 0;
3347	} else if (ret == -EIOCBQUEUED) {
3348		goto out_free;
3349	} else if (ret <= 0 || ret == io_size || !force_nonblock ||
3350		   (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3351		/* read all, failed, already did sync or don't want to retry */
3352		goto done;
3353	}
3354
3355	ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3356	if (ret2)
3357		return ret2;
3358
3359	iovec = NULL;
3360	rw = req->async_data;
3361	/* now use our persistent iterator, if we aren't already */
3362	iter = &rw->iter;
3363
3364	do {
3365		io_size -= ret;
3366		rw->bytes_done += ret;
3367		/* if we can retry, do so with the callbacks armed */
3368		if (!io_rw_should_retry(req)) {
3369			kiocb->ki_flags &= ~IOCB_WAITQ;
3370			return -EAGAIN;
3371		}
3372
3373		/*
3374		 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3375		 * we get -EIOCBQUEUED, then we'll get a notification when the
3376		 * desired page gets unlocked. We can also get a partial read
3377		 * here, and if we do, then just retry at the new offset.
3378		 */
3379		ret = io_iter_do_read(req, iter);
3380		if (ret == -EIOCBQUEUED)
3381			return 0;
3382		/* we got some bytes, but not all. retry. */
3383		kiocb->ki_flags &= ~IOCB_WAITQ;
3384	} while (ret > 0 && ret < io_size);
3385done:
3386	kiocb_done(kiocb, ret, issue_flags);
3387out_free:
3388	/* it's faster to check here then delegate to kfree */
3389	if (iovec)
3390		kfree(iovec);
3391	return 0;
3392}
3393
3394static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3395{
3396	if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3397		return -EBADF;
3398	return io_prep_rw(req, sqe);
3399}
3400
3401static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3402{
3403	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3404	struct kiocb *kiocb = &req->rw.kiocb;
3405	struct iov_iter __iter, *iter = &__iter;
3406	struct io_async_rw *rw = req->async_data;
3407	ssize_t ret, ret2, io_size;
3408	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3409
3410	if (rw) {
3411		iter = &rw->iter;
3412		iovec = NULL;
3413	} else {
3414		ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3415		if (ret < 0)
3416			return ret;
3417	}
3418	io_size = iov_iter_count(iter);
3419	req->result = io_size;
3420
3421	/* Ensure we clear previously set non-block flag */
3422	if (!force_nonblock)
3423		kiocb->ki_flags &= ~IOCB_NOWAIT;
3424	else
3425		kiocb->ki_flags |= IOCB_NOWAIT;
3426
3427	/* If the file doesn't support async, just async punt */
3428	if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3429		goto copy_iov;
3430
3431	/* file path doesn't support NOWAIT for non-direct_IO */
3432	if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3433	    (req->flags & REQ_F_ISREG))
3434		goto copy_iov;
3435
3436	ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3437	if (unlikely(ret))
3438		goto out_free;
3439
3440	/*
3441	 * Open-code file_start_write here to grab freeze protection,
3442	 * which will be released by another thread in
3443	 * io_complete_rw().  Fool lockdep by telling it the lock got
3444	 * released so that it doesn't complain about the held lock when
3445	 * we return to userspace.
3446	 */
3447	if (req->flags & REQ_F_ISREG) {
3448		sb_start_write(file_inode(req->file)->i_sb);
3449		__sb_writers_release(file_inode(req->file)->i_sb,
3450					SB_FREEZE_WRITE);
3451	}
3452	kiocb->ki_flags |= IOCB_WRITE;
3453
3454	if (req->file->f_op->write_iter)
3455		ret2 = call_write_iter(req->file, kiocb, iter);
3456	else if (req->file->f_op->write)
3457		ret2 = loop_rw_iter(WRITE, req, iter);
3458	else
3459		ret2 = -EINVAL;
3460
3461	if (req->flags & REQ_F_REISSUE) {
3462		req->flags &= ~REQ_F_REISSUE;
3463		ret2 = -EAGAIN;
3464	}
3465
3466	/*
3467	 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3468	 * retry them without IOCB_NOWAIT.
3469	 */
3470	if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3471		ret2 = -EAGAIN;
3472	/* no retry on NONBLOCK nor RWF_NOWAIT */
3473	if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3474		goto done;
3475	if (!force_nonblock || ret2 != -EAGAIN) {
3476		/* IOPOLL retry should happen for io-wq threads */
3477		if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3478			goto copy_iov;
3479done:
3480		kiocb_done(kiocb, ret2, issue_flags);
3481	} else {
3482copy_iov:
3483		/* some cases will consume bytes even on error returns */
3484		iov_iter_revert(iter, io_size - iov_iter_count(iter));
3485		ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3486		return ret ?: -EAGAIN;
3487	}
3488out_free:
3489	/* it's reportedly faster than delegating the null check to kfree() */
3490	if (iovec)
3491		kfree(iovec);
3492	return ret;
3493}
3494
3495static int io_renameat_prep(struct io_kiocb *req,
3496			    const struct io_uring_sqe *sqe)
3497{
3498	struct io_rename *ren = &req->rename;
3499	const char __user *oldf, *newf;
3500
3501	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3502		return -EINVAL;
3503	if (sqe->ioprio || sqe->buf_index)
3504		return -EINVAL;
3505	if (unlikely(req->flags & REQ_F_FIXED_FILE))
3506		return -EBADF;
3507
3508	ren->old_dfd = READ_ONCE(sqe->fd);
3509	oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3510	newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3511	ren->new_dfd = READ_ONCE(sqe->len);
3512	ren->flags = READ_ONCE(sqe->rename_flags);
3513
3514	ren->oldpath = getname(oldf);
3515	if (IS_ERR(ren->oldpath))
3516		return PTR_ERR(ren->oldpath);
3517
3518	ren->newpath = getname(newf);
3519	if (IS_ERR(ren->newpath)) {
3520		putname(ren->oldpath);
3521		return PTR_ERR(ren->newpath);
3522	}
3523
3524	req->flags |= REQ_F_NEED_CLEANUP;
3525	return 0;
3526}
3527
3528static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3529{
3530	struct io_rename *ren = &req->rename;
3531	int ret;
3532
3533	if (issue_flags & IO_URING_F_NONBLOCK)
3534		return -EAGAIN;
3535
3536	ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3537				ren->newpath, ren->flags);
3538
3539	req->flags &= ~REQ_F_NEED_CLEANUP;
3540	if (ret < 0)
3541		req_set_fail(req);
3542	io_req_complete(req, ret);
3543	return 0;
3544}
3545
3546static int io_unlinkat_prep(struct io_kiocb *req,
3547			    const struct io_uring_sqe *sqe)
3548{
3549	struct io_unlink *un = &req->unlink;
3550	const char __user *fname;
3551
3552	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3553		return -EINVAL;
3554	if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3555		return -EINVAL;
3556	if (unlikely(req->flags & REQ_F_FIXED_FILE))
3557		return -EBADF;
3558
3559	un->dfd = READ_ONCE(sqe->fd);
3560
3561	un->flags = READ_ONCE(sqe->unlink_flags);
3562	if (un->flags & ~AT_REMOVEDIR)
3563		return -EINVAL;
3564
3565	fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3566	un->filename = getname(fname);
3567	if (IS_ERR(un->filename))
3568		return PTR_ERR(un->filename);
3569
3570	req->flags |= REQ_F_NEED_CLEANUP;
3571	return 0;
3572}
3573
3574static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3575{
3576	struct io_unlink *un = &req->unlink;
3577	int ret;
3578
3579	if (issue_flags & IO_URING_F_NONBLOCK)
3580		return -EAGAIN;
3581
3582	if (un->flags & AT_REMOVEDIR)
3583		ret = do_rmdir(un->dfd, un->filename);
3584	else
3585		ret = do_unlinkat(un->dfd, un->filename);
3586
3587	req->flags &= ~REQ_F_NEED_CLEANUP;
3588	if (ret < 0)
3589		req_set_fail(req);
3590	io_req_complete(req, ret);
3591	return 0;
3592}
3593
3594static int io_shutdown_prep(struct io_kiocb *req,
3595			    const struct io_uring_sqe *sqe)
3596{
3597#if defined(CONFIG_NET)
3598	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3599		return -EINVAL;
3600	if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3601	    sqe->buf_index)
3602		return -EINVAL;
3603
3604	req->shutdown.how = READ_ONCE(sqe->len);
3605	return 0;
3606#else
3607	return -EOPNOTSUPP;
3608#endif
3609}
3610
3611static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3612{
3613#if defined(CONFIG_NET)
3614	struct socket *sock;
3615	int ret;
3616
3617	if (issue_flags & IO_URING_F_NONBLOCK)
3618		return -EAGAIN;
3619
3620	sock = sock_from_file(req->file);
3621	if (unlikely(!sock))
3622		return -ENOTSOCK;
3623
3624	ret = __sys_shutdown_sock(sock, req->shutdown.how);
3625	if (ret < 0)
3626		req_set_fail(req);
3627	io_req_complete(req, ret);
3628	return 0;
3629#else
3630	return -EOPNOTSUPP;
3631#endif
3632}
3633
3634static int __io_splice_prep(struct io_kiocb *req,
3635			    const struct io_uring_sqe *sqe)
3636{
3637	struct io_splice *sp = &req->splice;
3638	unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3639
3640	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3641		return -EINVAL;
3642
3643	sp->file_in = NULL;
3644	sp->len = READ_ONCE(sqe->len);
3645	sp->flags = READ_ONCE(sqe->splice_flags);
3646
3647	if (unlikely(sp->flags & ~valid_flags))
3648		return -EINVAL;
3649
3650	sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3651				  (sp->flags & SPLICE_F_FD_IN_FIXED));
3652	if (!sp->file_in)
3653		return -EBADF;
3654	req->flags |= REQ_F_NEED_CLEANUP;
3655	return 0;
3656}
3657
3658static int io_tee_prep(struct io_kiocb *req,
3659		       const struct io_uring_sqe *sqe)
3660{
3661	if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3662		return -EINVAL;
3663	return __io_splice_prep(req, sqe);
3664}
3665
3666static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3667{
3668	struct io_splice *sp = &req->splice;
3669	struct file *in = sp->file_in;
3670	struct file *out = sp->file_out;
3671	unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3672	long ret = 0;
3673
3674	if (issue_flags & IO_URING_F_NONBLOCK)
3675		return -EAGAIN;
3676	if (sp->len)
3677		ret = do_tee(in, out, sp->len, flags);
3678
3679	if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3680		io_put_file(in);
3681	req->flags &= ~REQ_F_NEED_CLEANUP;
3682
3683	if (ret != sp->len)
3684		req_set_fail(req);
3685	io_req_complete(req, ret);
3686	return 0;
3687}
3688
3689static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3690{
3691	struct io_splice *sp = &req->splice;
3692
3693	sp->off_in = READ_ONCE(sqe->splice_off_in);
3694	sp->off_out = READ_ONCE(sqe->off);
3695	return __io_splice_prep(req, sqe);
3696}
3697
3698static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3699{
3700	struct io_splice *sp = &req->splice;
3701	struct file *in = sp->file_in;
3702	struct file *out = sp->file_out;
3703	unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3704	loff_t *poff_in, *poff_out;
3705	long ret = 0;
3706
3707	if (issue_flags & IO_URING_F_NONBLOCK)
3708		return -EAGAIN;
3709
3710	poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3711	poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3712
3713	if (sp->len)
3714		ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3715
3716	if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3717		io_put_file(in);
3718	req->flags &= ~REQ_F_NEED_CLEANUP;
3719
3720	if (ret != sp->len)
3721		req_set_fail(req);
3722	io_req_complete(req, ret);
3723	return 0;
3724}
3725
3726/*
3727 * IORING_OP_NOP just posts a completion event, nothing else.
3728 */
3729static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3730{
3731	struct io_ring_ctx *ctx = req->ctx;
3732
3733	if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3734		return -EINVAL;
3735
3736	__io_req_complete(req, issue_flags, 0, 0);
3737	return 0;
3738}
3739
3740static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3741{
3742	struct io_ring_ctx *ctx = req->ctx;
3743
3744	if (!req->file)
3745		return -EBADF;
3746
3747	if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3748		return -EINVAL;
3749	if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3750		return -EINVAL;
3751
3752	req->sync.flags = READ_ONCE(sqe->fsync_flags);
3753	if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3754		return -EINVAL;
3755
3756	req->sync.off = READ_ONCE(sqe->off);
3757	req->sync.len = READ_ONCE(sqe->len);
3758	return 0;
3759}
3760
3761static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3762{
3763	loff_t end = req->sync.off + req->sync.len;
3764	int ret;
3765
3766	/* fsync always requires a blocking context */
3767	if (issue_flags & IO_URING_F_NONBLOCK)
3768		return -EAGAIN;
3769
3770	ret = vfs_fsync_range(req->file, req->sync.off,
3771				end > 0 ? end : LLONG_MAX,
3772				req->sync.flags & IORING_FSYNC_DATASYNC);
3773	if (ret < 0)
3774		req_set_fail(req);
3775	io_req_complete(req, ret);
3776	return 0;
3777}
3778
3779static int io_fallocate_prep(struct io_kiocb *req,
3780			     const struct io_uring_sqe *sqe)
3781{
3782	if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3783		return -EINVAL;
3784	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3785		return -EINVAL;
3786
3787	req->sync.off = READ_ONCE(sqe->off);
3788	req->sync.len = READ_ONCE(sqe->addr);
3789	req->sync.mode = READ_ONCE(sqe->len);
3790	return 0;
3791}
3792
3793static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3794{
3795	int ret;
3796
3797	/* fallocate always requiring blocking context */
3798	if (issue_flags & IO_URING_F_NONBLOCK)
3799		return -EAGAIN;
3800	ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3801				req->sync.len);
3802	if (ret < 0)
3803		req_set_fail(req);
3804	io_req_complete(req, ret);
3805	return 0;
3806}
3807
3808static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3809{
3810	const char __user *fname;
3811	int ret;
3812
3813	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3814		return -EINVAL;
3815	if (unlikely(sqe->ioprio || sqe->buf_index))
3816		return -EINVAL;
3817	if (unlikely(req->flags & REQ_F_FIXED_FILE))
3818		return -EBADF;
3819
3820	/* open.how should be already initialised */
3821	if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3822		req->open.how.flags |= O_LARGEFILE;
3823
3824	req->open.dfd = READ_ONCE(sqe->fd);
3825	fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3826	req->open.filename = getname(fname);
3827	if (IS_ERR(req->open.filename)) {
3828		ret = PTR_ERR(req->open.filename);
3829		req->open.filename = NULL;
3830		return ret;
3831	}
3832	req->open.nofile = rlimit(RLIMIT_NOFILE);
3833	req->flags |= REQ_F_NEED_CLEANUP;
3834	return 0;
3835}
3836
3837static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3838{
3839	u64 mode = READ_ONCE(sqe->len);
3840	u64 flags = READ_ONCE(sqe->open_flags);
3841
3842	req->open.how = build_open_how(flags, mode);
3843	return __io_openat_prep(req, sqe);
3844}
3845
3846static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3847{
3848	struct open_how __user *how;
3849	size_t len;
3850	int ret;
3851
3852	how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3853	len = READ_ONCE(sqe->len);
3854	if (len < OPEN_HOW_SIZE_VER0)
3855		return -EINVAL;
3856
3857	ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3858					len);
3859	if (ret)
3860		return ret;
3861
3862	return __io_openat_prep(req, sqe);
3863}
3864
3865static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3866{
3867	struct open_flags op;
3868	struct file *file;
3869	bool nonblock_set;
3870	bool resolve_nonblock;
3871	int ret;
3872
3873	ret = build_open_flags(&req->open.how, &op);
3874	if (ret)
3875		goto err;
3876	nonblock_set = op.open_flag & O_NONBLOCK;
3877	resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3878	if (issue_flags & IO_URING_F_NONBLOCK) {
3879		/*
3880		 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3881		 * it'll always -EAGAIN
3882		 */
3883		if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3884			return -EAGAIN;
3885		op.lookup_flags |= LOOKUP_CACHED;
3886		op.open_flag |= O_NONBLOCK;
3887	}
3888
3889	ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3890	if (ret < 0)
3891		goto err;
3892
3893	file = do_filp_open(req->open.dfd, req->open.filename, &op);
3894	if (IS_ERR(file)) {
3895		/*
3896		 * We could hang on to this 'fd' on retrying, but seems like
3897		 * marginal gain for something that is now known to be a slower
3898		 * path. So just put it, and we'll get a new one when we retry.
3899		 */
3900		put_unused_fd(ret);
3901
3902		ret = PTR_ERR(file);
3903		/* only retry if RESOLVE_CACHED wasn't already set by application */
3904		if (ret == -EAGAIN &&
3905		    (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3906			return -EAGAIN;
3907		goto err;
3908	}
3909
3910	if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3911		file->f_flags &= ~O_NONBLOCK;
3912	fsnotify_open(file);
3913	fd_install(ret, file);
3914err:
3915	putname(req->open.filename);
3916	req->flags &= ~REQ_F_NEED_CLEANUP;
3917	if (ret < 0)
3918		req_set_fail(req);
3919	__io_req_complete(req, issue_flags, ret, 0);
3920	return 0;
3921}
3922
3923static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3924{
3925	return io_openat2(req, issue_flags);
3926}
3927
3928static int io_remove_buffers_prep(struct io_kiocb *req,
3929				  const struct io_uring_sqe *sqe)
3930{
3931	struct io_provide_buf *p = &req->pbuf;
3932	u64 tmp;
3933
3934	if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3935		return -EINVAL;
3936
3937	tmp = READ_ONCE(sqe->fd);
3938	if (!tmp || tmp > USHRT_MAX)
3939		return -EINVAL;
3940
3941	memset(p, 0, sizeof(*p));
3942	p->nbufs = tmp;
3943	p->bgid = READ_ONCE(sqe->buf_group);
3944	return 0;
3945}
3946
3947static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3948			       int bgid, unsigned nbufs)
3949{
3950	unsigned i = 0;
3951
3952	/* shouldn't happen */
3953	if (!nbufs)
3954		return 0;
3955
3956	/* the head kbuf is the list itself */
3957	while (!list_empty(&buf->list)) {
3958		struct io_buffer *nxt;
3959
3960		nxt = list_first_entry(&buf->list, struct io_buffer, list);
3961		list_del(&nxt->list);
3962		kfree(nxt);
3963		if (++i == nbufs)
3964			return i;
3965	}
3966	i++;
3967	kfree(buf);
3968	xa_erase(&ctx->io_buffers, bgid);
3969
3970	return i;
3971}
3972
3973static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3974{
3975	struct io_provide_buf *p = &req->pbuf;
3976	struct io_ring_ctx *ctx = req->ctx;
3977	struct io_buffer *head;
3978	int ret = 0;
3979	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3980
3981	io_ring_submit_lock(ctx, !force_nonblock);
3982
3983	lockdep_assert_held(&ctx->uring_lock);
3984
3985	ret = -ENOENT;
3986	head = xa_load(&ctx->io_buffers, p->bgid);
3987	if (head)
3988		ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3989	if (ret < 0)
3990		req_set_fail(req);
3991
3992	/* complete before unlock, IOPOLL may need the lock */
3993	__io_req_complete(req, issue_flags, ret, 0);
3994	io_ring_submit_unlock(ctx, !force_nonblock);
3995	return 0;
3996}
3997
3998static int io_provide_buffers_prep(struct io_kiocb *req,
3999				   const struct io_uring_sqe *sqe)
4000{
4001	unsigned long size, tmp_check;
4002	struct io_provide_buf *p = &req->pbuf;
4003	u64 tmp;
4004
4005	if (sqe->ioprio || sqe->rw_flags)
4006		return -EINVAL;
4007
4008	tmp = READ_ONCE(sqe->fd);
4009	if (!tmp || tmp > USHRT_MAX)
4010		return -E2BIG;
4011	p->nbufs = tmp;
4012	p->addr = READ_ONCE(sqe->addr);
4013	p->len = READ_ONCE(sqe->len);
4014
4015	if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4016				&size))
4017		return -EOVERFLOW;
4018	if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4019		return -EOVERFLOW;
4020
4021	size = (unsigned long)p->len * p->nbufs;
4022	if (!access_ok(u64_to_user_ptr(p->addr), size))
4023		return -EFAULT;
4024
4025	p->bgid = READ_ONCE(sqe->buf_group);
4026	tmp = READ_ONCE(sqe->off);
4027	if (tmp > USHRT_MAX)
4028		return -E2BIG;
4029	p->bid = tmp;
4030	return 0;
4031}
4032
4033static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4034{
4035	struct io_buffer *buf;
4036	u64 addr = pbuf->addr;
4037	int i, bid = pbuf->bid;
4038
4039	for (i = 0; i < pbuf->nbufs; i++) {
4040		buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4041		if (!buf)
4042			break;
4043
4044		buf->addr = addr;
4045		buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4046		buf->bid = bid;
4047		addr += pbuf->len;
4048		bid++;
4049		if (!*head) {
4050			INIT_LIST_HEAD(&buf->list);
4051			*head = buf;
4052		} else {
4053			list_add_tail(&buf->list, &(*head)->list);
4054		}
4055	}
4056
4057	return i ? i : -ENOMEM;
4058}
4059
4060static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4061{
4062	struct io_provide_buf *p = &req->pbuf;
4063	struct io_ring_ctx *ctx = req->ctx;
4064	struct io_buffer *head, *list;
4065	int ret = 0;
4066	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4067
4068	io_ring_submit_lock(ctx, !force_nonblock);
4069
4070	lockdep_assert_held(&ctx->uring_lock);
4071
4072	list = head = xa_load(&ctx->io_buffers, p->bgid);
4073
4074	ret = io_add_buffers(p, &head);
4075	if (ret >= 0 && !list) {
4076		ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4077		if (ret < 0)
4078			__io_remove_buffers(ctx, head, p->bgid, -1U);
4079	}
4080	if (ret < 0)
4081		req_set_fail(req);
4082	/* complete before unlock, IOPOLL may need the lock */
4083	__io_req_complete(req, issue_flags, ret, 0);
4084	io_ring_submit_unlock(ctx, !force_nonblock);
4085	return 0;
4086}
4087
4088static int io_epoll_ctl_prep(struct io_kiocb *req,
4089			     const struct io_uring_sqe *sqe)
4090{
4091#if defined(CONFIG_EPOLL)
4092	if (sqe->ioprio || sqe->buf_index)
4093		return -EINVAL;
4094	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4095		return -EINVAL;
4096
4097	req->epoll.epfd = READ_ONCE(sqe->fd);
4098	req->epoll.op = READ_ONCE(sqe->len);
4099	req->epoll.fd = READ_ONCE(sqe->off);
4100
4101	if (ep_op_has_event(req->epoll.op)) {
4102		struct epoll_event __user *ev;
4103
4104		ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4105		if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4106			return -EFAULT;
4107	}
4108
4109	return 0;
4110#else
4111	return -EOPNOTSUPP;
4112#endif
4113}
4114
4115static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4116{
4117#if defined(CONFIG_EPOLL)
4118	struct io_epoll *ie = &req->epoll;
4119	int ret;
4120	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4121
4122	ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4123	if (force_nonblock && ret == -EAGAIN)
4124		return -EAGAIN;
4125
4126	if (ret < 0)
4127		req_set_fail(req);
4128	__io_req_complete(req, issue_flags, ret, 0);
4129	return 0;
4130#else
4131	return -EOPNOTSUPP;
4132#endif
4133}
4134
4135static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4136{
4137#if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4138	if (sqe->ioprio || sqe->buf_index || sqe->off)
4139		return -EINVAL;
4140	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4141		return -EINVAL;
4142
4143	req->madvise.addr = READ_ONCE(sqe->addr);
4144	req->madvise.len = READ_ONCE(sqe->len);
4145	req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4146	return 0;
4147#else
4148	return -EOPNOTSUPP;
4149#endif
4150}
4151
4152static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4153{
4154#if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4155	struct io_madvise *ma = &req->madvise;
4156	int ret;
4157
4158	if (issue_flags & IO_URING_F_NONBLOCK)
4159		return -EAGAIN;
4160
4161	ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4162	if (ret < 0)
4163		req_set_fail(req);
4164	io_req_complete(req, ret);
4165	return 0;
4166#else
4167	return -EOPNOTSUPP;
4168#endif
4169}
4170
4171static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4172{
4173	if (sqe->ioprio || sqe->buf_index || sqe->addr)
4174		return -EINVAL;
4175	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4176		return -EINVAL;
4177
4178	req->fadvise.offset = READ_ONCE(sqe->off);
4179	req->fadvise.len = READ_ONCE(sqe->len);
4180	req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4181	return 0;
4182}
4183
4184static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4185{
4186	struct io_fadvise *fa = &req->fadvise;
4187	int ret;
4188
4189	if (issue_flags & IO_URING_F_NONBLOCK) {
4190		switch (fa->advice) {
4191		case POSIX_FADV_NORMAL:
4192		case POSIX_FADV_RANDOM:
4193		case POSIX_FADV_SEQUENTIAL:
4194			break;
4195		default:
4196			return -EAGAIN;
4197		}
4198	}
4199
4200	ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4201	if (ret < 0)
4202		req_set_fail(req);
4203	__io_req_complete(req, issue_flags, ret, 0);
4204	return 0;
4205}
4206
4207static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4208{
4209	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4210		return -EINVAL;
4211	if (sqe->ioprio || sqe->buf_index)
4212		return -EINVAL;
4213	if (req->flags & REQ_F_FIXED_FILE)
4214		return -EBADF;
4215
4216	req->statx.dfd = READ_ONCE(sqe->fd);
4217	req->statx.mask = READ_ONCE(sqe->len);
4218	req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4219	req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4220	req->statx.flags = READ_ONCE(sqe->statx_flags);
4221
4222	return 0;
4223}
4224
4225static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4226{
4227	struct io_statx *ctx = &req->statx;
4228	int ret;
4229
4230	if (issue_flags & IO_URING_F_NONBLOCK)
4231		return -EAGAIN;
4232
4233	ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4234		       ctx->buffer);
4235
4236	if (ret < 0)
4237		req_set_fail(req);
4238	io_req_complete(req, ret);
4239	return 0;
4240}
4241
4242static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4243{
4244	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4245		return -EINVAL;
4246	if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4247	    sqe->rw_flags || sqe->buf_index)
4248		return -EINVAL;
4249	if (req->flags & REQ_F_FIXED_FILE)
4250		return -EBADF;
4251
4252	req->close.fd = READ_ONCE(sqe->fd);
4253	return 0;
4254}
4255
4256static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4257{
4258	struct files_struct *files = current->files;
4259	struct io_close *close = &req->close;
4260	struct fdtable *fdt;
4261	struct file *file = NULL;
4262	int ret = -EBADF;
4263
4264	spin_lock(&files->file_lock);
4265	fdt = files_fdtable(files);
4266	if (close->fd >= fdt->max_fds) {
4267		spin_unlock(&files->file_lock);
4268		goto err;
4269	}
4270	file = fdt->fd[close->fd];
4271	if (!file || file->f_op == &io_uring_fops) {
4272		spin_unlock(&files->file_lock);
4273		file = NULL;
4274		goto err;
4275	}
4276
4277	/* if the file has a flush method, be safe and punt to async */
4278	if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4279		spin_unlock(&files->file_lock);
4280		return -EAGAIN;
4281	}
4282
4283	ret = __close_fd_get_file(close->fd, &file);
4284	spin_unlock(&files->file_lock);
4285	if (ret < 0) {
4286		if (ret == -ENOENT)
4287			ret = -EBADF;
4288		goto err;
4289	}
4290
4291	/* No ->flush() or already async, safely close from here */
4292	ret = filp_close(file, current->files);
4293err:
4294	if (ret < 0)
4295		req_set_fail(req);
4296	if (file)
4297		fput(file);
4298	__io_req_complete(req, issue_flags, ret, 0);
4299	return 0;
4300}
4301
4302static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4303{
4304	struct io_ring_ctx *ctx = req->ctx;
4305
4306	if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4307		return -EINVAL;
4308	if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4309		return -EINVAL;
4310
4311	req->sync.off = READ_ONCE(sqe->off);
4312	req->sync.len = READ_ONCE(sqe->len);
4313	req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4314	return 0;
4315}
4316
4317static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4318{
4319	int ret;
4320
4321	/* sync_file_range always requires a blocking context */
4322	if (issue_flags & IO_URING_F_NONBLOCK)
4323		return -EAGAIN;
4324
4325	ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4326				req->sync.flags);
4327	if (ret < 0)
4328		req_set_fail(req);
4329	io_req_complete(req, ret);
4330	return 0;
4331}
4332
4333#if defined(CONFIG_NET)
4334static int io_setup_async_msg(struct io_kiocb *req,
4335			      struct io_async_msghdr *kmsg)
4336{
4337	struct io_async_msghdr *async_msg = req->async_data;
4338
4339	if (async_msg)
4340		return -EAGAIN;
4341	if (io_alloc_async_data(req)) {
4342		kfree(kmsg->free_iov);
4343		return -ENOMEM;
4344	}
4345	async_msg = req->async_data;
4346	req->flags |= REQ_F_NEED_CLEANUP;
4347	memcpy(async_msg, kmsg, sizeof(*kmsg));
4348	async_msg->msg.msg_name = &async_msg->addr;
4349	/* if were using fast_iov, set it to the new one */
4350	if (!async_msg->free_iov)
4351		async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4352
4353	return -EAGAIN;
4354}
4355
4356static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4357			       struct io_async_msghdr *iomsg)
4358{
4359	iomsg->msg.msg_name = &iomsg->addr;
4360	iomsg->free_iov = iomsg->fast_iov;
4361	return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4362				   req->sr_msg.msg_flags, &iomsg->free_iov);
4363}
4364
4365static int io_sendmsg_prep_async(struct io_kiocb *req)
4366{
4367	int ret;
4368
4369	ret = io_sendmsg_copy_hdr(req, req->async_data);
4370	if (!ret)
4371		req->flags |= REQ_F_NEED_CLEANUP;
4372	return ret;
4373}
4374
4375static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4376{
4377	struct io_sr_msg *sr = &req->sr_msg;
4378
4379	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4380		return -EINVAL;
4381
4382	sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4383	sr->len = READ_ONCE(sqe->len);
4384	sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4385	if (sr->msg_flags & MSG_DONTWAIT)
4386		req->flags |= REQ_F_NOWAIT;
4387
4388#ifdef CONFIG_COMPAT
4389	if (req->ctx->compat)
4390		sr->msg_flags |= MSG_CMSG_COMPAT;
4391#endif
4392	return 0;
4393}
4394
4395static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4396{
4397	struct io_async_msghdr iomsg, *kmsg;
4398	struct socket *sock;
4399	unsigned flags;
4400	int min_ret = 0;
4401	int ret;
4402
4403	sock = sock_from_file(req->file);
4404	if (unlikely(!sock))
4405		return -ENOTSOCK;
4406
4407	kmsg = req->async_data;
4408	if (!kmsg) {
4409		ret = io_sendmsg_copy_hdr(req, &iomsg);
4410		if (ret)
4411			return ret;
4412		kmsg = &iomsg;
4413	}
4414
4415	flags = req->sr_msg.msg_flags;
4416	if (issue_flags & IO_URING_F_NONBLOCK)
4417		flags |= MSG_DONTWAIT;
4418	if (flags & MSG_WAITALL)
4419		min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4420
4421	ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4422	if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4423		return io_setup_async_msg(req, kmsg);
4424	if (ret == -ERESTARTSYS)
4425		ret = -EINTR;
4426
4427	/* fast path, check for non-NULL to avoid function call */
4428	if (kmsg->free_iov)
4429		kfree(kmsg->free_iov);
4430	req->flags &= ~REQ_F_NEED_CLEANUP;
4431	if (ret < min_ret)
4432		req_set_fail(req);
4433	__io_req_complete(req, issue_flags, ret, 0);
4434	return 0;
4435}
4436
4437static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4438{
4439	struct io_sr_msg *sr = &req->sr_msg;
4440	struct msghdr msg;
4441	struct iovec iov;
4442	struct socket *sock;
4443	unsigned flags;
4444	int min_ret = 0;
4445	int ret;
4446
4447	sock = sock_from_file(req->file);
4448	if (unlikely(!sock))
4449		return -ENOTSOCK;
4450
4451	ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4452	if (unlikely(ret))
4453		return ret;
4454
4455	msg.msg_name = NULL;
4456	msg.msg_control = NULL;
4457	msg.msg_controllen = 0;
4458	msg.msg_namelen = 0;
4459
4460	flags = req->sr_msg.msg_flags;
4461	if (issue_flags & IO_URING_F_NONBLOCK)
4462		flags |= MSG_DONTWAIT;
4463	if (flags & MSG_WAITALL)
4464		min_ret = iov_iter_count(&msg.msg_iter);
4465
4466	msg.msg_flags = flags;
4467	ret = sock_sendmsg(sock, &msg);
4468	if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4469		return -EAGAIN;
4470	if (ret == -ERESTARTSYS)
4471		ret = -EINTR;
4472
4473	if (ret < min_ret)
4474		req_set_fail(req);
4475	__io_req_complete(req, issue_flags, ret, 0);
4476	return 0;
4477}
4478
4479static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4480				 struct io_async_msghdr *iomsg)
4481{
4482	struct io_sr_msg *sr = &req->sr_msg;
4483	struct iovec __user *uiov;
4484	size_t iov_len;
4485	int ret;
4486
4487	ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4488					&iomsg->uaddr, &uiov, &iov_len);
4489	if (ret)
4490		return ret;
4491
4492	if (req->flags & REQ_F_BUFFER_SELECT) {
4493		if (iov_len > 1)
4494			return -EINVAL;
4495		if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4496			return -EFAULT;
4497		sr->len = iomsg->fast_iov[0].iov_len;
4498		iomsg->free_iov = NULL;
4499	} else {
4500		iomsg->free_iov = iomsg->fast_iov;
4501		ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4502				     &iomsg->free_iov, &iomsg->msg.msg_iter,
4503				     false);
4504		if (ret > 0)
4505			ret = 0;
4506	}
4507
4508	return ret;
4509}
4510
4511#ifdef CONFIG_COMPAT
4512static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4513					struct io_async_msghdr *iomsg)
4514{
4515	struct io_sr_msg *sr = &req->sr_msg;
4516	struct compat_iovec __user *uiov;
4517	compat_uptr_t ptr;
4518	compat_size_t len;
4519	int ret;
4520
4521	ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4522				  &ptr, &len);
4523	if (ret)
4524		return ret;
4525
4526	uiov = compat_ptr(ptr);
4527	if (req->flags & REQ_F_BUFFER_SELECT) {
4528		compat_ssize_t clen;
4529
4530		if (len > 1)
4531			return -EINVAL;
4532		if (!access_ok(uiov, sizeof(*uiov)))
4533			return -EFAULT;
4534		if (__get_user(clen, &uiov->iov_len))
4535			return -EFAULT;
4536		if (clen < 0)
4537			return -EINVAL;
4538		sr->len = clen;
4539		iomsg->free_iov = NULL;
4540	} else {
4541		iomsg->free_iov = iomsg->fast_iov;
4542		ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4543				   UIO_FASTIOV, &iomsg->free_iov,
4544				   &iomsg->msg.msg_iter, true);
4545		if (ret < 0)
4546			return ret;
4547	}
4548
4549	return 0;
4550}
4551#endif
4552
4553static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4554			       struct io_async_msghdr *iomsg)
4555{
4556	iomsg->msg.msg_name = &iomsg->addr;
4557
4558#ifdef CONFIG_COMPAT
4559	if (req->ctx->compat)
4560		return __io_compat_recvmsg_copy_hdr(req, iomsg);
4561#endif
4562
4563	return __io_recvmsg_copy_hdr(req, iomsg);
4564}
4565
4566static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4567					       bool needs_lock)
4568{
4569	struct io_sr_msg *sr = &req->sr_msg;
4570	struct io_buffer *kbuf;
4571
4572	kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4573	if (IS_ERR(kbuf))
4574		return kbuf;
4575
4576	sr->kbuf = kbuf;
4577	req->flags |= REQ_F_BUFFER_SELECTED;
4578	return kbuf;
4579}
4580
4581static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4582{
4583	return io_put_kbuf(req, req->sr_msg.kbuf);
4584}
4585
4586static int io_recvmsg_prep_async(struct io_kiocb *req)
4587{
4588	int ret;
4589
4590	ret = io_recvmsg_copy_hdr(req, req->async_data);
4591	if (!ret)
4592		req->flags |= REQ_F_NEED_CLEANUP;
4593	return ret;
4594}
4595
4596static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4597{
4598	struct io_sr_msg *sr = &req->sr_msg;
4599
4600	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4601		return -EINVAL;
4602
4603	sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4604	sr->len = READ_ONCE(sqe->len);
4605	sr->bgid = READ_ONCE(sqe->buf_group);
4606	sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4607	if (sr->msg_flags & MSG_DONTWAIT)
4608		req->flags |= REQ_F_NOWAIT;
4609
4610#ifdef CONFIG_COMPAT
4611	if (req->ctx->compat)
4612		sr->msg_flags |= MSG_CMSG_COMPAT;
4613#endif
4614	return 0;
4615}
4616
4617static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4618{
4619	struct io_async_msghdr iomsg, *kmsg;
4620	struct socket *sock;
4621	struct io_buffer *kbuf;
4622	unsigned flags;
4623	int min_ret = 0;
4624	int ret, cflags = 0;
4625	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4626
4627	sock = sock_from_file(req->file);
4628	if (unlikely(!sock))
4629		return -ENOTSOCK;
4630
4631	kmsg = req->async_data;
4632	if (!kmsg) {
4633		ret = io_recvmsg_copy_hdr(req, &iomsg);
4634		if (ret)
4635			return ret;
4636		kmsg = &iomsg;
4637	}
4638
4639	if (req->flags & REQ_F_BUFFER_SELECT) {
4640		kbuf = io_recv_buffer_select(req, !force_nonblock);
4641		if (IS_ERR(kbuf))
4642			return PTR_ERR(kbuf);
4643		kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4644		kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4645		iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4646				1, req->sr_msg.len);
4647	}
4648
4649	flags = req->sr_msg.msg_flags;
4650	if (force_nonblock)
4651		flags |= MSG_DONTWAIT;
4652	if (flags & MSG_WAITALL)
4653		min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4654
4655	ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4656					kmsg->uaddr, flags);
4657	if (force_nonblock && ret == -EAGAIN)
4658		return io_setup_async_msg(req, kmsg);
4659	if (ret == -ERESTARTSYS)
4660		ret = -EINTR;
4661
4662	if (req->flags & REQ_F_BUFFER_SELECTED)
4663		cflags = io_put_recv_kbuf(req);
4664	/* fast path, check for non-NULL to avoid function call */
4665	if (kmsg->free_iov)
4666		kfree(kmsg->free_iov);
4667	req->flags &= ~REQ_F_NEED_CLEANUP;
4668	if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4669		req_set_fail(req);
4670	__io_req_complete(req, issue_flags, ret, cflags);
4671	return 0;
4672}
4673
4674static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4675{
4676	struct io_buffer *kbuf;
4677	struct io_sr_msg *sr = &req->sr_msg;
4678	struct msghdr msg;
4679	void __user *buf = sr->buf;
4680	struct socket *sock;
4681	struct iovec iov;
4682	unsigned flags;
4683	int min_ret = 0;
4684	int ret, cflags = 0;
4685	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4686
4687	sock = sock_from_file(req->file);
4688	if (unlikely(!sock))
4689		return -ENOTSOCK;
4690
4691	if (req->flags & REQ_F_BUFFER_SELECT) {
4692		kbuf = io_recv_buffer_select(req, !force_nonblock);
4693		if (IS_ERR(kbuf))
4694			return PTR_ERR(kbuf);
4695		buf = u64_to_user_ptr(kbuf->addr);
4696	}
4697
4698	ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4699	if (unlikely(ret))
4700		goto out_free;
4701
4702	msg.msg_name = NULL;
4703	msg.msg_control = NULL;
4704	msg.msg_controllen = 0;
4705	msg.msg_namelen = 0;
4706	msg.msg_iocb = NULL;
4707	msg.msg_flags = 0;
4708
4709	flags = req->sr_msg.msg_flags;
4710	if (force_nonblock)
4711		flags |= MSG_DONTWAIT;
4712	if (flags & MSG_WAITALL)
4713		min_ret = iov_iter_count(&msg.msg_iter);
4714
4715	ret = sock_recvmsg(sock, &msg, flags);
4716	if (force_nonblock && ret == -EAGAIN)
4717		return -EAGAIN;
4718	if (ret == -ERESTARTSYS)
4719		ret = -EINTR;
4720out_free:
4721	if (req->flags & REQ_F_BUFFER_SELECTED)
4722		cflags = io_put_recv_kbuf(req);
4723	if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4724		req_set_fail(req);
4725	__io_req_complete(req, issue_flags, ret, cflags);
4726	return 0;
4727}
4728
4729static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4730{
4731	struct io_accept *accept = &req->accept;
4732
4733	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4734		return -EINVAL;
4735	if (sqe->ioprio || sqe->len || sqe->buf_index)
4736		return -EINVAL;
4737
4738	accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4739	accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4740	accept->flags = READ_ONCE(sqe->accept_flags);
4741	accept->nofile = rlimit(RLIMIT_NOFILE);
4742	return 0;
4743}
4744
4745static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4746{
4747	struct io_accept *accept = &req->accept;
4748	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4749	unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4750	int ret;
4751
4752	if (req->file->f_flags & O_NONBLOCK)
4753		req->flags |= REQ_F_NOWAIT;
4754
4755	ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4756					accept->addr_len, accept->flags,
4757					accept->nofile);
4758	if (ret == -EAGAIN && force_nonblock)
4759		return -EAGAIN;
4760	if (ret < 0) {
4761		if (ret == -ERESTARTSYS)
4762			ret = -EINTR;
4763		req_set_fail(req);
4764	}
4765	__io_req_complete(req, issue_flags, ret, 0);
4766	return 0;
4767}
4768
4769static int io_connect_prep_async(struct io_kiocb *req)
4770{
4771	struct io_async_connect *io = req->async_data;
4772	struct io_connect *conn = &req->connect;
4773
4774	return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4775}
4776
4777static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4778{
4779	struct io_connect *conn = &req->connect;
4780
4781	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4782		return -EINVAL;
4783	if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4784		return -EINVAL;
4785
4786	conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4787	conn->addr_len =  READ_ONCE(sqe->addr2);
4788	return 0;
4789}
4790
4791static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4792{
4793	struct io_async_connect __io, *io;
4794	unsigned file_flags;
4795	int ret;
4796	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4797
4798	if (req->async_data) {
4799		io = req->async_data;
4800	} else {
4801		ret = move_addr_to_kernel(req->connect.addr,
4802						req->connect.addr_len,
4803						&__io.address);
4804		if (ret)
4805			goto out;
4806		io = &__io;
4807	}
4808
4809	file_flags = force_nonblock ? O_NONBLOCK : 0;
4810
4811	ret = __sys_connect_file(req->file, &io->address,
4812					req->connect.addr_len, file_flags);
4813	if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4814		if (req->async_data)
4815			return -EAGAIN;
4816		if (io_alloc_async_data(req)) {
4817			ret = -ENOMEM;
4818			goto out;
4819		}
4820		memcpy(req->async_data, &__io, sizeof(__io));
4821		return -EAGAIN;
4822	}
4823	if (ret == -ERESTARTSYS)
4824		ret = -EINTR;
4825out:
4826	if (ret < 0)
4827		req_set_fail(req);
4828	__io_req_complete(req, issue_flags, ret, 0);
4829	return 0;
4830}
4831#else /* !CONFIG_NET */
4832#define IO_NETOP_FN(op)							\
4833static int io_##op(struct io_kiocb *req, unsigned int issue_flags)	\
4834{									\
4835	return -EOPNOTSUPP;						\
4836}
4837
4838#define IO_NETOP_PREP(op)						\
4839IO_NETOP_FN(op)								\
4840static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4841{									\
4842	return -EOPNOTSUPP;						\
4843}									\
4844
4845#define IO_NETOP_PREP_ASYNC(op)						\
4846IO_NETOP_PREP(op)							\
4847static int io_##op##_prep_async(struct io_kiocb *req)			\
4848{									\
4849	return -EOPNOTSUPP;						\
4850}
4851
4852IO_NETOP_PREP_ASYNC(sendmsg);
4853IO_NETOP_PREP_ASYNC(recvmsg);
4854IO_NETOP_PREP_ASYNC(connect);
4855IO_NETOP_PREP(accept);
4856IO_NETOP_FN(send);
4857IO_NETOP_FN(recv);
4858#endif /* CONFIG_NET */
4859
4860struct io_poll_table {
4861	struct poll_table_struct pt;
4862	struct io_kiocb *req;
4863	int nr_entries;
4864	int error;
4865};
4866
4867static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4868			   __poll_t mask, io_req_tw_func_t func)
4869{
4870	/* for instances that support it check for an event match first: */
4871	if (mask && !(mask & poll->events))
4872		return 0;
4873
4874	trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4875
4876	list_del_init(&poll->wait.entry);
4877
4878	req->result = mask;
4879	req->io_task_work.func = func;
4880
4881	/*
4882	 * If this fails, then the task is exiting. When a task exits, the
4883	 * work gets canceled, so just cancel this request as well instead
4884	 * of executing it. We can't safely execute it anyway, as we may not
4885	 * have the needed state needed for it anyway.
4886	 */
4887	io_req_task_work_add(req);
4888	return 1;
4889}
4890
4891static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4892	__acquires(&req->ctx->completion_lock)
4893{
4894	struct io_ring_ctx *ctx = req->ctx;
4895
4896	if (unlikely(req->task->flags & PF_EXITING))
4897		WRITE_ONCE(poll->canceled, true);
4898
4899	if (!req->result && !READ_ONCE(poll->canceled)) {
4900		struct poll_table_struct pt = { ._key = poll->events };
4901
4902		req->result = vfs_poll(req->file, &pt) & poll->events;
4903	}
4904
4905	spin_lock(&ctx->completion_lock);
4906	if (!req->result && !READ_ONCE(poll->canceled)) {
4907		add_wait_queue(poll->head, &poll->wait);
4908		return true;
4909	}
4910
4911	return false;
4912}
4913
4914static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4915{
4916	/* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4917	if (req->opcode == IORING_OP_POLL_ADD)
4918		return req->async_data;
4919	return req->apoll->double_poll;
4920}
4921
4922static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4923{
4924	if (req->opcode == IORING_OP_POLL_ADD)
4925		return &req->poll;
4926	return &req->apoll->poll;
4927}
4928
4929static void io_poll_remove_double(struct io_kiocb *req)
4930	__must_hold(&req->ctx->completion_lock)
4931{
4932	struct io_poll_iocb *poll = io_poll_get_double(req);
4933
4934	lockdep_assert_held(&req->ctx->completion_lock);
4935
4936	if (poll && poll->head) {
4937		struct wait_queue_head *head = poll->head;
4938
4939		spin_lock_irq(&head->lock);
4940		list_del_init(&poll->wait.entry);
4941		if (poll->wait.private)
4942			req_ref_put(req);
4943		poll->head = NULL;
4944		spin_unlock_irq(&head->lock);
4945	}
4946}
4947
4948static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4949	__must_hold(&req->ctx->completion_lock)
4950{
4951	struct io_ring_ctx *ctx = req->ctx;
4952	unsigned flags = IORING_CQE_F_MORE;
4953	int error;
4954
4955	if (READ_ONCE(req->poll.canceled)) {
4956		error = -ECANCELED;
4957		req->poll.events |= EPOLLONESHOT;
4958	} else {
4959		error = mangle_poll(mask);
4960	}
4961	if (req->poll.events & EPOLLONESHOT)
4962		flags = 0;
4963	if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4964		req->poll.done = true;
4965		flags = 0;
4966	}
4967	if (flags & IORING_CQE_F_MORE)
4968		ctx->cq_extra++;
4969
4970	io_commit_cqring(ctx);
4971	return !(flags & IORING_CQE_F_MORE);
4972}
4973
4974static void io_poll_task_func(struct io_kiocb *req)
4975{
4976	struct io_ring_ctx *ctx = req->ctx;
4977	struct io_kiocb *nxt;
4978
4979	if (io_poll_rewait(req, &req->poll)) {
4980		spin_unlock(&ctx->completion_lock);
4981	} else {
4982		bool done;
4983
4984		done = io_poll_complete(req, req->result);
4985		if (done) {
4986			io_poll_remove_double(req);
4987			hash_del(&req->hash_node);
4988		} else {
4989			req->result = 0;
4990			add_wait_queue(req->poll.head, &req->poll.wait);
4991		}
4992		spin_unlock(&ctx->completion_lock);
4993		io_cqring_ev_posted(ctx);
4994
4995		if (done) {
4996			nxt = io_put_req_find_next(req);
4997			if (nxt)
4998				io_req_task_submit(nxt);
4999		}
5000	}
5001}
5002
5003static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5004			       int sync, void *key)
5005{
5006	struct io_kiocb *req = wait->private;
5007	struct io_poll_iocb *poll = io_poll_get_single(req);
5008	__poll_t mask = key_to_poll(key);
5009	unsigned long flags;
5010
5011	/* for instances that support it check for an event match first: */
5012	if (mask && !(mask & poll->events))
5013		return 0;
5014	if (!(poll->events & EPOLLONESHOT))
5015		return poll->wait.func(&poll->wait, mode, sync, key);
5016
5017	list_del_init(&wait->entry);
5018
5019	if (poll->head) {
5020		bool done;
5021
5022		spin_lock_irqsave(&poll->head->lock, flags);
5023		done = list_empty(&poll->wait.entry);
5024		if (!done)
5025			list_del_init(&poll->wait.entry);
5026		/* make sure double remove sees this as being gone */
5027		wait->private = NULL;
5028		spin_unlock_irqrestore(&poll->head->lock, flags);
5029		if (!done) {
5030			/* use wait func handler, so it matches the rq type */
5031			poll->wait.func(&poll->wait, mode, sync, key);
5032		}
5033	}
5034	req_ref_put(req);
5035	return 1;
5036}
5037
5038static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5039			      wait_queue_func_t wake_func)
5040{
5041	poll->head = NULL;
5042	poll->done = false;
5043	poll->canceled = false;
5044#define IO_POLL_UNMASK	(EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5045	/* mask in events that we always want/need */
5046	poll->events = events | IO_POLL_UNMASK;
5047	INIT_LIST_HEAD(&poll->wait.entry);
5048	init_waitqueue_func_entry(&poll->wait, wake_func);
5049}
5050
5051static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5052			    struct wait_queue_head *head,
5053			    struct io_poll_iocb **poll_ptr)
5054{
5055	struct io_kiocb *req = pt->req;
5056
5057	/*
5058	 * The file being polled uses multiple waitqueues for poll handling
5059	 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5060	 * if this happens.
5061	 */
5062	if (unlikely(pt->nr_entries)) {
5063		struct io_poll_iocb *poll_one = poll;
5064
5065		/* double add on the same waitqueue head, ignore */
5066		if (poll_one->head == head)
5067			return;
5068		/* already have a 2nd entry, fail a third attempt */
5069		if (*poll_ptr) {
5070			if ((*poll_ptr)->head == head)
5071				return;
5072			pt->error = -EINVAL;
5073			return;
5074		}
5075		/*
5076		 * Can't handle multishot for double wait for now, turn it
5077		 * into one-shot mode.
5078		 */
5079		if (!(poll_one->events & EPOLLONESHOT))
5080			poll_one->events |= EPOLLONESHOT;
5081		poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5082		if (!poll) {
5083			pt->error = -ENOMEM;
5084			return;
5085		}
5086		io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5087		req_ref_get(req);
5088		poll->wait.private = req;
5089		*poll_ptr = poll;
5090	}
5091
5092	pt->nr_entries++;
5093	poll->head = head;
5094
5095	if (poll->events & EPOLLEXCLUSIVE)
5096		add_wait_queue_exclusive(head, &poll->wait);
5097	else
5098		add_wait_queue(head, &poll->wait);
5099}
5100
5101static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5102			       struct poll_table_struct *p)
5103{
5104	struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5105	struct async_poll *apoll = pt->req->apoll;
5106
5107	__io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5108}
5109
5110static void io_async_task_func(struct io_kiocb *req)
5111{
5112	struct async_poll *apoll = req->apoll;
5113	struct io_ring_ctx *ctx = req->ctx;
5114
5115	trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5116
5117	if (io_poll_rewait(req, &apoll->poll)) {
5118		spin_unlock(&ctx->completion_lock);
5119		return;
5120	}
5121
5122	hash_del(&req->hash_node);
5123	io_poll_remove_double(req);
5124	spin_unlock(&ctx->completion_lock);
5125
5126	if (!READ_ONCE(apoll->poll.canceled))
5127		io_req_task_submit(req);
5128	else
5129		io_req_complete_failed(req, -ECANCELED);
5130}
5131
5132static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5133			void *key)
5134{
5135	struct io_kiocb *req = wait->private;
5136	struct io_poll_iocb *poll = &req->apoll->poll;
5137
5138	trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5139					key_to_poll(key));
5140
5141	return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5142}
5143
5144static void io_poll_req_insert(struct io_kiocb *req)
5145{
5146	struct io_ring_ctx *ctx = req->ctx;
5147	struct hlist_head *list;
5148
5149	list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5150	hlist_add_head(&req->hash_node, list);
5151}
5152
5153static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5154				      struct io_poll_iocb *poll,
5155				      struct io_poll_table *ipt, __poll_t mask,
5156				      wait_queue_func_t wake_func)
5157	__acquires(&ctx->completion_lock)
5158{
5159	struct io_ring_ctx *ctx = req->ctx;
5160	bool cancel = false;
5161
5162	INIT_HLIST_NODE(&req->hash_node);
5163	io_init_poll_iocb(poll, mask, wake_func);
5164	poll->file = req->file;
5165	poll->wait.private = req;
5166
5167	ipt->pt._key = mask;
5168	ipt->req = req;
5169	ipt->error = 0;
5170	ipt->nr_entries = 0;
5171
5172	mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5173	if (unlikely(!ipt->nr_entries) && !ipt->error)
5174		ipt->error = -EINVAL;
5175
5176	spin_lock(&ctx->completion_lock);
5177	if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5178		io_poll_remove_double(req);
5179	if (likely(poll->head)) {
5180		spin_lock_irq(&poll->head->lock);
5181		if (unlikely(list_empty(&poll->wait.entry))) {
5182			if (ipt->error)
5183				cancel = true;
5184			ipt->error = 0;
5185			mask = 0;
5186		}
5187		if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5188			list_del_init(&poll->wait.entry);
5189		else if (cancel)
5190			WRITE_ONCE(poll->canceled, true);
5191		else if (!poll->done) /* actually waiting for an event */
5192			io_poll_req_insert(req);
5193		spin_unlock_irq(&poll->head->lock);
5194	}
5195
5196	return mask;
5197}
5198
5199enum {
5200	IO_APOLL_OK,
5201	IO_APOLL_ABORTED,
5202	IO_APOLL_READY
5203};
5204
5205static int io_arm_poll_handler(struct io_kiocb *req)
5206{
5207	const struct io_op_def *def = &io_op_defs[req->opcode];
5208	struct io_ring_ctx *ctx = req->ctx;
5209	struct async_poll *apoll;
5210	struct io_poll_table ipt;
5211	__poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5212	int rw;
5213
5214	if (!req->file || !file_can_poll(req->file))
5215		return IO_APOLL_ABORTED;
5216	if (req->flags & REQ_F_POLLED)
5217		return IO_APOLL_ABORTED;
5218	if (!def->pollin && !def->pollout)
5219		return IO_APOLL_ABORTED;
5220
5221	if (def->pollin) {
5222		rw = READ;
5223		mask |= POLLIN | POLLRDNORM;
5224
5225		/* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5226		if ((req->opcode == IORING_OP_RECVMSG) &&
5227		    (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5228			mask &= ~POLLIN;
5229	} else {
5230		rw = WRITE;
5231		mask |= POLLOUT | POLLWRNORM;
5232	}
5233
5234	/* if we can't nonblock try, then no point in arming a poll handler */
5235	if (!io_file_supports_nowait(req, rw))
5236		return IO_APOLL_ABORTED;
5237
5238	apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5239	if (unlikely(!apoll))
5240		return IO_APOLL_ABORTED;
5241	apoll->double_poll = NULL;
5242	req->apoll = apoll;
5243	req->flags |= REQ_F_POLLED;
5244	ipt.pt._qproc = io_async_queue_proc;
5245	io_req_set_refcount(req);
5246
5247	ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5248					io_async_wake);
5249	spin_unlock(&ctx->completion_lock);
5250	if (ret || ipt.error)
5251		return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5252
5253	trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5254				mask, apoll->poll.events);
5255	return IO_APOLL_OK;
5256}
5257
5258static bool __io_poll_remove_one(struct io_kiocb *req,
5259				 struct io_poll_iocb *poll, bool do_cancel)
5260	__must_hold(&req->ctx->completion_lock)
5261{
5262	bool do_complete = false;
5263
5264	if (!poll->head)
5265		return false;
5266	spin_lock_irq(&poll->head->lock);
5267	if (do_cancel)
5268		WRITE_ONCE(poll->canceled, true);
5269	if (!list_empty(&poll->wait.entry)) {
5270		list_del_init(&poll->wait.entry);
5271		do_complete = true;
5272	}
5273	spin_unlock_irq(&poll->head->lock);
5274	hash_del(&req->hash_node);
5275	return do_complete;
5276}
5277
5278static bool io_poll_remove_one(struct io_kiocb *req)
5279	__must_hold(&req->ctx->completion_lock)
5280{
5281	bool do_complete;
5282
5283	io_poll_remove_double(req);
5284	do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5285
5286	if (do_complete) {
5287		io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5288		io_commit_cqring(req->ctx);
5289		req_set_fail(req);
5290		io_put_req_deferred(req);
5291	}
5292	return do_complete;
5293}
5294
5295/*
5296 * Returns true if we found and killed one or more poll requests
5297 */
5298static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5299			       bool cancel_all)
5300{
5301	struct hlist_node *tmp;
5302	struct io_kiocb *req;
5303	int posted = 0, i;
5304
5305	spin_lock(&ctx->completion_lock);
5306	for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5307		struct hlist_head *list;
5308
5309		list = &ctx->cancel_hash[i];
5310		hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5311			if (io_match_task(req, tsk, cancel_all))
5312				posted += io_poll_remove_one(req);
5313		}
5314	}
5315	spin_unlock(&ctx->completion_lock);
5316
5317	if (posted)
5318		io_cqring_ev_posted(ctx);
5319
5320	return posted != 0;
5321}
5322
5323static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5324				     bool poll_only)
5325	__must_hold(&ctx->completion_lock)
5326{
5327	struct hlist_head *list;
5328	struct io_kiocb *req;
5329
5330	list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5331	hlist_for_each_entry(req, list, hash_node) {
5332		if (sqe_addr != req->user_data)
5333			continue;
5334		if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5335			continue;
5336		return req;
5337	}
5338	return NULL;
5339}
5340
5341static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5342			  bool poll_only)
5343	__must_hold(&ctx->completion_lock)
5344{
5345	struct io_kiocb *req;
5346
5347	req = io_poll_find(ctx, sqe_addr, poll_only);
5348	if (!req)
5349		return -ENOENT;
5350	if (io_poll_remove_one(req))
5351		return 0;
5352
5353	return -EALREADY;
5354}
5355
5356static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5357				     unsigned int flags)
5358{
5359	u32 events;
5360
5361	events = READ_ONCE(sqe->poll32_events);
5362#ifdef __BIG_ENDIAN
5363	events = swahw32(events);
5364#endif
5365	if (!(flags & IORING_POLL_ADD_MULTI))
5366		events |= EPOLLONESHOT;
5367	return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5368}
5369
5370static int io_poll_update_prep(struct io_kiocb *req,
5371			       const struct io_uring_sqe *sqe)
5372{
5373	struct io_poll_update *upd = &req->poll_update;
5374	u32 flags;
5375
5376	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5377		return -EINVAL;
5378	if (sqe->ioprio || sqe->buf_index)
5379		return -EINVAL;
5380	flags = READ_ONCE(sqe->len);
5381	if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5382		      IORING_POLL_ADD_MULTI))
5383		return -EINVAL;
5384	/* meaningless without update */
5385	if (flags == IORING_POLL_ADD_MULTI)
5386		return -EINVAL;
5387
5388	upd->old_user_data = READ_ONCE(sqe->addr);
5389	upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5390	upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5391
5392	upd->new_user_data = READ_ONCE(sqe->off);
5393	if (!upd->update_user_data && upd->new_user_data)
5394		return -EINVAL;
5395	if (upd->update_events)
5396		upd->events = io_poll_parse_events(sqe, flags);
5397	else if (sqe->poll32_events)
5398		return -EINVAL;
5399
5400	return 0;
5401}
5402
5403static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5404			void *key)
5405{
5406	struct io_kiocb *req = wait->private;
5407	struct io_poll_iocb *poll = &req->poll;
5408
5409	return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5410}
5411
5412static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5413			       struct poll_table_struct *p)
5414{
5415	struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5416
5417	__io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5418}
5419
5420static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5421{
5422	struct io_poll_iocb *poll = &req->poll;
5423	u32 flags;
5424
5425	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5426		return -EINVAL;
5427	if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5428		return -EINVAL;
5429	flags = READ_ONCE(sqe->len);
5430	if (flags & ~IORING_POLL_ADD_MULTI)
5431		return -EINVAL;
5432
5433	io_req_set_refcount(req);
5434	poll->events = io_poll_parse_events(sqe, flags);
5435	return 0;
5436}
5437
5438static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5439{
5440	struct io_poll_iocb *poll = &req->poll;
5441	struct io_ring_ctx *ctx = req->ctx;
5442	struct io_poll_table ipt;
5443	__poll_t mask;
5444
5445	ipt.pt._qproc = io_poll_queue_proc;
5446
5447	mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5448					io_poll_wake);
5449
5450	if (mask) { /* no async, we'd stolen it */
5451		ipt.error = 0;
5452		io_poll_complete(req, mask);
5453	}
5454	spin_unlock(&ctx->completion_lock);
5455
5456	if (mask) {
5457		io_cqring_ev_posted(ctx);
5458		if (poll->events & EPOLLONESHOT)
5459			io_put_req(req);
5460	}
5461	return ipt.error;
5462}
5463
5464static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5465{
5466	struct io_ring_ctx *ctx = req->ctx;
5467	struct io_kiocb *preq;
5468	bool completing;
5469	int ret;
5470
5471	spin_lock(&ctx->completion_lock);
5472	preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5473	if (!preq) {
5474		ret = -ENOENT;
5475		goto err;
5476	}
5477
5478	if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5479		completing = true;
5480		ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5481		goto err;
5482	}
5483
5484	/*
5485	 * Don't allow racy completion with singleshot, as we cannot safely
5486	 * update those. For multishot, if we're racing with completion, just
5487	 * let completion re-add it.
5488	 */
5489	completing = !__io_poll_remove_one(preq, &preq->poll, false);
5490	if (completing && (preq->poll.events & EPOLLONESHOT)) {
5491		ret = -EALREADY;
5492		goto err;
5493	}
5494	/* we now have a detached poll request. reissue. */
5495	ret = 0;
5496err:
5497	if (ret < 0) {
5498		spin_unlock(&ctx->completion_lock);
5499		req_set_fail(req);
5500		io_req_complete(req, ret);
5501		return 0;
5502	}
5503	/* only mask one event flags, keep behavior flags */
5504	if (req->poll_update.update_events) {
5505		preq->poll.events &= ~0xffff;
5506		preq->poll.events |= req->poll_update.events & 0xffff;
5507		preq->poll.events |= IO_POLL_UNMASK;
5508	}
5509	if (req->poll_update.update_user_data)
5510		preq->user_data = req->poll_update.new_user_data;
5511	spin_unlock(&ctx->completion_lock);
5512
5513	/* complete update request, we're done with it */
5514	io_req_complete(req, ret);
5515
5516	if (!completing) {
5517		ret = io_poll_add(preq, issue_flags);
5518		if (ret < 0) {
5519			req_set_fail(preq);
5520			io_req_complete(preq, ret);
5521		}
5522	}
5523	return 0;
5524}
5525
5526static void io_req_task_timeout(struct io_kiocb *req)
5527{
5528	req_set_fail(req);
5529	io_req_complete_post(req, -ETIME, 0);
5530}
5531
5532static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5533{
5534	struct io_timeout_data *data = container_of(timer,
5535						struct io_timeout_data, timer);
5536	struct io_kiocb *req = data->req;
5537	struct io_ring_ctx *ctx = req->ctx;
5538	unsigned long flags;
5539
5540	spin_lock_irqsave(&ctx->timeout_lock, flags);
5541	list_del_init(&req->timeout.list);
5542	atomic_set(&req->ctx->cq_timeouts,
5543		atomic_read(&req->ctx->cq_timeouts) + 1);
5544	spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5545
5546	req->io_task_work.func = io_req_task_timeout;
5547	io_req_task_work_add(req);
5548	return HRTIMER_NORESTART;
5549}
5550
5551static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5552					   __u64 user_data)
5553	__must_hold(&ctx->timeout_lock)
5554{
5555	struct io_timeout_data *io;
5556	struct io_kiocb *req;
5557	bool found = false;
5558
5559	list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5560		found = user_data == req->user_data;
5561		if (found)
5562			break;
5563	}
5564	if (!found)
5565		return ERR_PTR(-ENOENT);
5566
5567	io = req->async_data;
5568	if (hrtimer_try_to_cancel(&io->timer) == -1)
5569		return ERR_PTR(-EALREADY);
5570	list_del_init(&req->timeout.list);
5571	return req;
5572}
5573
5574static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5575	__must_hold(&ctx->timeout_lock)
5576{
5577	struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5578
5579	if (IS_ERR(req))
5580		return PTR_ERR(req);
5581
5582	req_set_fail(req);
5583	io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5584	io_put_req_deferred(req);
5585	return 0;
5586}
5587
5588static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5589			     struct timespec64 *ts, enum hrtimer_mode mode)
5590	__must_hold(&ctx->timeout_lock)
5591{
5592	struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5593	struct io_timeout_data *data;
5594
5595	if (IS_ERR(req))
5596		return PTR_ERR(req);
5597
5598	req->timeout.off = 0; /* noseq */
5599	data = req->async_data;
5600	list_add_tail(&req->timeout.list, &ctx->timeout_list);
5601	hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5602	data->timer.function = io_timeout_fn;
5603	hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5604	return 0;
5605}
5606
5607static int io_timeout_remove_prep(struct io_kiocb *req,
5608				  const struct io_uring_sqe *sqe)
5609{
5610	struct io_timeout_rem *tr = &req->timeout_rem;
5611
5612	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5613		return -EINVAL;
5614	if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5615		return -EINVAL;
5616	if (sqe->ioprio || sqe->buf_index || sqe->len)
5617		return -EINVAL;
5618
5619	tr->addr = READ_ONCE(sqe->addr);
5620	tr->flags = READ_ONCE(sqe->timeout_flags);
5621	if (tr->flags & IORING_TIMEOUT_UPDATE) {
5622		if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5623			return -EINVAL;
5624		if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5625			return -EFAULT;
5626	} else if (tr->flags) {
5627		/* timeout removal doesn't support flags */
5628		return -EINVAL;
5629	}
5630
5631	return 0;
5632}
5633
5634static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5635{
5636	return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5637					    : HRTIMER_MODE_REL;
5638}
5639
5640/*
5641 * Remove or update an existing timeout command
5642 */
5643static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5644{
5645	struct io_timeout_rem *tr = &req->timeout_rem;
5646	struct io_ring_ctx *ctx = req->ctx;
5647	int ret;
5648
5649	spin_lock_irq(&ctx->timeout_lock);
5650	if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5651		ret = io_timeout_cancel(ctx, tr->addr);
5652	else
5653		ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5654					io_translate_timeout_mode(tr->flags));
5655	spin_unlock_irq(&ctx->timeout_lock);
5656
5657	if (ret < 0)
5658		req_set_fail(req);
5659	io_req_complete_post(req, ret, 0);
5660	return 0;
5661}
5662
5663static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5664			   bool is_timeout_link)
5665{
5666	struct io_timeout_data *data;
5667	unsigned flags;
5668	u32 off = READ_ONCE(sqe->off);
5669
5670	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5671		return -EINVAL;
5672	if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5673		return -EINVAL;
5674	if (off && is_timeout_link)
5675		return -EINVAL;
5676	flags = READ_ONCE(sqe->timeout_flags);
5677	if (flags & ~IORING_TIMEOUT_ABS)
5678		return -EINVAL;
5679
5680	req->timeout.off = off;
5681	if (unlikely(off && !req->ctx->off_timeout_used))
5682		req->ctx->off_timeout_used = true;
5683
5684	if (!req->async_data && io_alloc_async_data(req))
5685		return -ENOMEM;
5686
5687	data = req->async_data;
5688	data->req = req;
5689
5690	if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5691		return -EFAULT;
5692
5693	data->mode = io_translate_timeout_mode(flags);
5694	hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5695
5696	if (is_timeout_link) {
5697		struct io_submit_link *link = &req->ctx->submit_state.link;
5698
5699		if (!link->head)
5700			return -EINVAL;
5701		if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
5702			return -EINVAL;
5703		req->timeout.head = link->last;
5704		link->last->flags |= REQ_F_ARM_LTIMEOUT;
5705	}
5706	return 0;
5707}
5708
5709static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5710{
5711	struct io_ring_ctx *ctx = req->ctx;
5712	struct io_timeout_data *data = req->async_data;
5713	struct list_head *entry;
5714	u32 tail, off = req->timeout.off;
5715
5716	spin_lock_irq(&ctx->timeout_lock);
5717
5718	/*
5719	 * sqe->off holds how many events that need to occur for this
5720	 * timeout event to be satisfied. If it isn't set, then this is
5721	 * a pure timeout request, sequence isn't used.
5722	 */
5723	if (io_is_timeout_noseq(req)) {
5724		entry = ctx->timeout_list.prev;
5725		goto add;
5726	}
5727
5728	tail = ctx->cached_cq_tail - atomic_read(&ctx->