io_uring.c revision 2c5d763c
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/* only define 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	(PAGE_SHIFT - 3)
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
1529/*
1530 * This should only get called when at least one event has been posted.
1531 * Some applications rely on the eventfd notification count only changing
1532 * IFF a new CQE has been added to the CQ ring. There's no depedency on
1533 * 1:1 relationship between how many times this function is called (and
1534 * hence the eventfd count) and number of CQEs posted to the CQ ring.
1535 */
1536static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1537{
1538	/*
1539	 * wake_up_all() may seem excessive, but io_wake_function() and
1540	 * io_should_wake() handle the termination of the loop and only
1541	 * wake as many waiters as we need to.
1542	 */
1543	if (wq_has_sleeper(&ctx->cq_wait))
1544		wake_up_all(&ctx->cq_wait);
1545	if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1546		wake_up(&ctx->sq_data->wait);
1547	if (io_should_trigger_evfd(ctx))
1548		eventfd_signal(ctx->cq_ev_fd, 1);
1549	if (waitqueue_active(&ctx->poll_wait)) {
1550		wake_up_interruptible(&ctx->poll_wait);
1551		kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1552	}
1553}
1554
1555static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1556{
1557	if (ctx->flags & IORING_SETUP_SQPOLL) {
1558		if (wq_has_sleeper(&ctx->cq_wait))
1559			wake_up_all(&ctx->cq_wait);
1560	}
1561	if (io_should_trigger_evfd(ctx))
1562		eventfd_signal(ctx->cq_ev_fd, 1);
1563	if (waitqueue_active(&ctx->poll_wait)) {
1564		wake_up_interruptible(&ctx->poll_wait);
1565		kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1566	}
1567}
1568
1569/* Returns true if there are no backlogged entries after the flush */
1570static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1571{
1572	bool all_flushed, posted;
1573
1574	if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1575		return false;
1576
1577	posted = false;
1578	spin_lock(&ctx->completion_lock);
1579	while (!list_empty(&ctx->cq_overflow_list)) {
1580		struct io_uring_cqe *cqe = io_get_cqe(ctx);
1581		struct io_overflow_cqe *ocqe;
1582
1583		if (!cqe && !force)
1584			break;
1585		ocqe = list_first_entry(&ctx->cq_overflow_list,
1586					struct io_overflow_cqe, list);
1587		if (cqe)
1588			memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1589		else
1590			io_account_cq_overflow(ctx);
1591
1592		posted = true;
1593		list_del(&ocqe->list);
1594		kfree(ocqe);
1595	}
1596
1597	all_flushed = list_empty(&ctx->cq_overflow_list);
1598	if (all_flushed) {
1599		clear_bit(0, &ctx->check_cq_overflow);
1600		WRITE_ONCE(ctx->rings->sq_flags,
1601			   ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1602	}
1603
1604	if (posted)
1605		io_commit_cqring(ctx);
1606	spin_unlock(&ctx->completion_lock);
1607	if (posted)
1608		io_cqring_ev_posted(ctx);
1609	return all_flushed;
1610}
1611
1612static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1613{
1614	bool ret = true;
1615
1616	if (test_bit(0, &ctx->check_cq_overflow)) {
1617		/* iopoll syncs against uring_lock, not completion_lock */
1618		if (ctx->flags & IORING_SETUP_IOPOLL)
1619			mutex_lock(&ctx->uring_lock);
1620		ret = __io_cqring_overflow_flush(ctx, false);
1621		if (ctx->flags & IORING_SETUP_IOPOLL)
1622			mutex_unlock(&ctx->uring_lock);
1623	}
1624
1625	return ret;
1626}
1627
1628/* must to be called somewhat shortly after putting a request */
1629static inline void io_put_task(struct task_struct *task, int nr)
1630{
1631	struct io_uring_task *tctx = task->io_uring;
1632
1633	if (likely(task == current)) {
1634		tctx->cached_refs += nr;
1635	} else {
1636		percpu_counter_sub(&tctx->inflight, nr);
1637		if (unlikely(atomic_read(&tctx->in_idle)))
1638			wake_up(&tctx->wait);
1639		put_task_struct_many(task, nr);
1640	}
1641}
1642
1643static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1644				     long res, unsigned int cflags)
1645{
1646	struct io_overflow_cqe *ocqe;
1647
1648	ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1649	if (!ocqe) {
1650		/*
1651		 * If we're in ring overflow flush mode, or in task cancel mode,
1652		 * or cannot allocate an overflow entry, then we need to drop it
1653		 * on the floor.
1654		 */
1655		io_account_cq_overflow(ctx);
1656		return false;
1657	}
1658	if (list_empty(&ctx->cq_overflow_list)) {
1659		set_bit(0, &ctx->check_cq_overflow);
1660		WRITE_ONCE(ctx->rings->sq_flags,
1661			   ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1662
1663	}
1664	ocqe->cqe.user_data = user_data;
1665	ocqe->cqe.res = res;
1666	ocqe->cqe.flags = cflags;
1667	list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1668	return true;
1669}
1670
1671static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1672					  long res, unsigned int cflags)
1673{
1674	struct io_uring_cqe *cqe;
1675
1676	trace_io_uring_complete(ctx, user_data, res, cflags);
1677
1678	/*
1679	 * If we can't get a cq entry, userspace overflowed the
1680	 * submission (by quite a lot). Increment the overflow count in
1681	 * the ring.
1682	 */
1683	cqe = io_get_cqe(ctx);
1684	if (likely(cqe)) {
1685		WRITE_ONCE(cqe->user_data, user_data);
1686		WRITE_ONCE(cqe->res, res);
1687		WRITE_ONCE(cqe->flags, cflags);
1688		return true;
1689	}
1690	return io_cqring_event_overflow(ctx, user_data, res, cflags);
1691}
1692
1693/* not as hot to bloat with inlining */
1694static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1695					  long res, unsigned int cflags)
1696{
1697	return __io_cqring_fill_event(ctx, user_data, res, cflags);
1698}
1699
1700static void io_req_complete_post(struct io_kiocb *req, long res,
1701				 unsigned int cflags)
1702{
1703	struct io_ring_ctx *ctx = req->ctx;
1704
1705	spin_lock(&ctx->completion_lock);
1706	__io_cqring_fill_event(ctx, req->user_data, res, cflags);
1707	/*
1708	 * If we're the last reference to this request, add to our locked
1709	 * free_list cache.
1710	 */
1711	if (req_ref_put_and_test(req)) {
1712		if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1713			if (req->flags & IO_DISARM_MASK)
1714				io_disarm_next(req);
1715			if (req->link) {
1716				io_req_task_queue(req->link);
1717				req->link = NULL;
1718			}
1719		}
1720		io_dismantle_req(req);
1721		io_put_task(req->task, 1);
1722		list_add(&req->inflight_entry, &ctx->locked_free_list);
1723		ctx->locked_free_nr++;
1724	} else {
1725		if (!percpu_ref_tryget(&ctx->refs))
1726			req = NULL;
1727	}
1728	io_commit_cqring(ctx);
1729	spin_unlock(&ctx->completion_lock);
1730
1731	if (req) {
1732		io_cqring_ev_posted(ctx);
1733		percpu_ref_put(&ctx->refs);
1734	}
1735}
1736
1737static inline bool io_req_needs_clean(struct io_kiocb *req)
1738{
1739	return req->flags & IO_REQ_CLEAN_FLAGS;
1740}
1741
1742static void io_req_complete_state(struct io_kiocb *req, long res,
1743				  unsigned int cflags)
1744{
1745	if (io_req_needs_clean(req))
1746		io_clean_op(req);
1747	req->result = res;
1748	req->compl.cflags = cflags;
1749	req->flags |= REQ_F_COMPLETE_INLINE;
1750}
1751
1752static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1753				     long res, unsigned cflags)
1754{
1755	if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1756		io_req_complete_state(req, res, cflags);
1757	else
1758		io_req_complete_post(req, res, cflags);
1759}
1760
1761static inline void io_req_complete(struct io_kiocb *req, long res)
1762{
1763	__io_req_complete(req, 0, res, 0);
1764}
1765
1766static void io_req_complete_failed(struct io_kiocb *req, long res)
1767{
1768	req_set_fail(req);
1769	io_req_complete_post(req, res, 0);
1770}
1771
1772/*
1773 * Don't initialise the fields below on every allocation, but do that in
1774 * advance and keep them valid across allocations.
1775 */
1776static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1777{
1778	req->ctx = ctx;
1779	req->link = NULL;
1780	req->async_data = NULL;
1781	/* not necessary, but safer to zero */
1782	req->result = 0;
1783}
1784
1785static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1786					struct io_submit_state *state)
1787{
1788	spin_lock(&ctx->completion_lock);
1789	list_splice_init(&ctx->locked_free_list, &state->free_list);
1790	ctx->locked_free_nr = 0;
1791	spin_unlock(&ctx->completion_lock);
1792}
1793
1794/* Returns true IFF there are requests in the cache */
1795static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1796{
1797	struct io_submit_state *state = &ctx->submit_state;
1798	int nr;
1799
1800	/*
1801	 * If we have more than a batch's worth of requests in our IRQ side
1802	 * locked cache, grab the lock and move them over to our submission
1803	 * side cache.
1804	 */
1805	if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1806		io_flush_cached_locked_reqs(ctx, state);
1807
1808	nr = state->free_reqs;
1809	while (!list_empty(&state->free_list)) {
1810		struct io_kiocb *req = list_first_entry(&state->free_list,
1811					struct io_kiocb, inflight_entry);
1812
1813		list_del(&req->inflight_entry);
1814		state->reqs[nr++] = req;
1815		if (nr == ARRAY_SIZE(state->reqs))
1816			break;
1817	}
1818
1819	state->free_reqs = nr;
1820	return nr != 0;
1821}
1822
1823/*
1824 * A request might get retired back into the request caches even before opcode
1825 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1826 * Because of that, io_alloc_req() should be called only under ->uring_lock
1827 * and with extra caution to not get a request that is still worked on.
1828 */
1829static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1830	__must_hold(&ctx->uring_lock)
1831{
1832	struct io_submit_state *state = &ctx->submit_state;
1833	gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1834	int ret, i;
1835
1836	BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1837
1838	if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1839		goto got_req;
1840
1841	ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1842				    state->reqs);
1843
1844	/*
1845	 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1846	 * retry single alloc to be on the safe side.
1847	 */
1848	if (unlikely(ret <= 0)) {
1849		state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1850		if (!state->reqs[0])
1851			return NULL;
1852		ret = 1;
1853	}
1854
1855	for (i = 0; i < ret; i++)
1856		io_preinit_req(state->reqs[i], ctx);
1857	state->free_reqs = ret;
1858got_req:
1859	state->free_reqs--;
1860	return state->reqs[state->free_reqs];
1861}
1862
1863static inline void io_put_file(struct file *file)
1864{
1865	if (file)
1866		fput(file);
1867}
1868
1869static void io_dismantle_req(struct io_kiocb *req)
1870{
1871	unsigned int flags = req->flags;
1872
1873	if (io_req_needs_clean(req))
1874		io_clean_op(req);
1875	if (!(flags & REQ_F_FIXED_FILE))
1876		io_put_file(req->file);
1877	if (req->fixed_rsrc_refs)
1878		percpu_ref_put(req->fixed_rsrc_refs);
1879	if (req->async_data) {
1880		kfree(req->async_data);
1881		req->async_data = NULL;
1882	}
1883}
1884
1885static void __io_free_req(struct io_kiocb *req)
1886{
1887	struct io_ring_ctx *ctx = req->ctx;
1888
1889	io_dismantle_req(req);
1890	io_put_task(req->task, 1);
1891
1892	spin_lock(&ctx->completion_lock);
1893	list_add(&req->inflight_entry, &ctx->locked_free_list);
1894	ctx->locked_free_nr++;
1895	spin_unlock(&ctx->completion_lock);
1896
1897	percpu_ref_put(&ctx->refs);
1898}
1899
1900static inline void io_remove_next_linked(struct io_kiocb *req)
1901{
1902	struct io_kiocb *nxt = req->link;
1903
1904	req->link = nxt->link;
1905	nxt->link = NULL;
1906}
1907
1908static bool io_kill_linked_timeout(struct io_kiocb *req)
1909	__must_hold(&req->ctx->completion_lock)
1910	__must_hold(&req->ctx->timeout_lock)
1911{
1912	struct io_kiocb *link = req->link;
1913
1914	if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
1915		struct io_timeout_data *io = link->async_data;
1916
1917		io_remove_next_linked(req);
1918		link->timeout.head = NULL;
1919		if (hrtimer_try_to_cancel(&io->timer) != -1) {
1920			io_cqring_fill_event(link->ctx, link->user_data,
1921					     -ECANCELED, 0);
1922			io_put_req_deferred(link);
1923			return true;
1924		}
1925	}
1926	return false;
1927}
1928
1929static void io_fail_links(struct io_kiocb *req)
1930	__must_hold(&req->ctx->completion_lock)
1931{
1932	struct io_kiocb *nxt, *link = req->link;
1933
1934	req->link = NULL;
1935	while (link) {
1936		nxt = link->link;
1937		link->link = NULL;
1938
1939		trace_io_uring_fail_link(req, link);
1940		io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1941		io_put_req_deferred(link);
1942		link = nxt;
1943	}
1944}
1945
1946static bool io_disarm_next(struct io_kiocb *req)
1947	__must_hold(&req->ctx->completion_lock)
1948{
1949	bool posted = false;
1950
1951	if (req->flags & REQ_F_ARM_LTIMEOUT) {
1952		struct io_kiocb *link = req->link;
1953
1954		req->flags &= ~REQ_F_ARM_LTIMEOUT;
1955		if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
1956			io_remove_next_linked(req);
1957			io_cqring_fill_event(link->ctx, link->user_data,
1958					     -ECANCELED, 0);
1959			io_put_req_deferred(link);
1960			posted = true;
1961		}
1962	} else if (req->flags & REQ_F_LINK_TIMEOUT) {
1963		struct io_ring_ctx *ctx = req->ctx;
1964
1965		spin_lock_irq(&ctx->timeout_lock);
1966		posted = io_kill_linked_timeout(req);
1967		spin_unlock_irq(&ctx->timeout_lock);
1968	}
1969	if (unlikely((req->flags & REQ_F_FAIL) &&
1970		     !(req->flags & REQ_F_HARDLINK))) {
1971		posted |= (req->link != NULL);
1972		io_fail_links(req);
1973	}
1974	return posted;
1975}
1976
1977static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1978{
1979	struct io_kiocb *nxt;
1980
1981	/*
1982	 * If LINK is set, we have dependent requests in this chain. If we
1983	 * didn't fail this request, queue the first one up, moving any other
1984	 * dependencies to the next request. In case of failure, fail the rest
1985	 * of the chain.
1986	 */
1987	if (req->flags & IO_DISARM_MASK) {
1988		struct io_ring_ctx *ctx = req->ctx;
1989		bool posted;
1990
1991		spin_lock(&ctx->completion_lock);
1992		posted = io_disarm_next(req);
1993		if (posted)
1994			io_commit_cqring(req->ctx);
1995		spin_unlock(&ctx->completion_lock);
1996		if (posted)
1997			io_cqring_ev_posted(ctx);
1998	}
1999	nxt = req->link;
2000	req->link = NULL;
2001	return nxt;
2002}
2003
2004static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2005{
2006	if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2007		return NULL;
2008	return __io_req_find_next(req);
2009}
2010
2011static void ctx_flush_and_put(struct io_ring_ctx *ctx)
2012{
2013	if (!ctx)
2014		return;
2015	if (ctx->submit_state.compl_nr) {
2016		mutex_lock(&ctx->uring_lock);
2017		if (ctx->submit_state.compl_nr)
2018			io_submit_flush_completions(ctx);
2019		mutex_unlock(&ctx->uring_lock);
2020	}
2021	percpu_ref_put(&ctx->refs);
2022}
2023
2024static void tctx_task_work(struct callback_head *cb)
2025{
2026	struct io_ring_ctx *ctx = NULL;
2027	struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2028						  task_work);
2029
2030	while (1) {
2031		struct io_wq_work_node *node;
2032
2033		spin_lock_irq(&tctx->task_lock);
2034		node = tctx->task_list.first;
2035		INIT_WQ_LIST(&tctx->task_list);
2036		if (!node)
2037			tctx->task_running = false;
2038		spin_unlock_irq(&tctx->task_lock);
2039		if (!node)
2040			break;
2041
2042		do {
2043			struct io_wq_work_node *next = node->next;
2044			struct io_kiocb *req = container_of(node, struct io_kiocb,
2045							    io_task_work.node);
2046
2047			if (req->ctx != ctx) {
2048				ctx_flush_and_put(ctx);
2049				ctx = req->ctx;
2050				percpu_ref_get(&ctx->refs);
2051			}
2052			req->io_task_work.func(req);
2053			node = next;
2054		} while (node);
2055
2056		cond_resched();
2057	}
2058
2059	ctx_flush_and_put(ctx);
2060}
2061
2062static void io_req_task_work_add(struct io_kiocb *req)
2063{
2064	struct task_struct *tsk = req->task;
2065	struct io_uring_task *tctx = tsk->io_uring;
2066	enum task_work_notify_mode notify;
2067	struct io_wq_work_node *node;
2068	unsigned long flags;
2069	bool running;
2070
2071	WARN_ON_ONCE(!tctx);
2072
2073	spin_lock_irqsave(&tctx->task_lock, flags);
2074	wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2075	running = tctx->task_running;
2076	if (!running)
2077		tctx->task_running = true;
2078	spin_unlock_irqrestore(&tctx->task_lock, flags);
2079
2080	/* task_work already pending, we're done */
2081	if (running)
2082		return;
2083
2084	/*
2085	 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2086	 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2087	 * processing task_work. There's no reliable way to tell if TWA_RESUME
2088	 * will do the job.
2089	 */
2090	notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2091	if (!task_work_add(tsk, &tctx->task_work, notify)) {
2092		wake_up_process(tsk);
2093		return;
2094	}
2095
2096	spin_lock_irqsave(&tctx->task_lock, flags);
2097	tctx->task_running = false;
2098	node = tctx->task_list.first;
2099	INIT_WQ_LIST(&tctx->task_list);
2100	spin_unlock_irqrestore(&tctx->task_lock, flags);
2101
2102	while (node) {
2103		req = container_of(node, struct io_kiocb, io_task_work.node);
2104		node = node->next;
2105		if (llist_add(&req->io_task_work.fallback_node,
2106			      &req->ctx->fallback_llist))
2107			schedule_delayed_work(&req->ctx->fallback_work, 1);
2108	}
2109}
2110
2111static void io_req_task_cancel(struct io_kiocb *req)
2112{
2113	struct io_ring_ctx *ctx = req->ctx;
2114
2115	/* ctx is guaranteed to stay alive while we hold uring_lock */
2116	mutex_lock(&ctx->uring_lock);
2117	io_req_complete_failed(req, req->result);
2118	mutex_unlock(&ctx->uring_lock);
2119}
2120
2121static void io_req_task_submit(struct io_kiocb *req)
2122{
2123	struct io_ring_ctx *ctx = req->ctx;
2124
2125	/* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2126	mutex_lock(&ctx->uring_lock);
2127	/* req->task == current here, checking PF_EXITING is safe */
2128	if (likely(!(req->task->flags & PF_EXITING)))
2129		__io_queue_sqe(req);
2130	else
2131		io_req_complete_failed(req, -EFAULT);
2132	mutex_unlock(&ctx->uring_lock);
2133}
2134
2135static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2136{
2137	req->result = ret;
2138	req->io_task_work.func = io_req_task_cancel;
2139	io_req_task_work_add(req);
2140}
2141
2142static void io_req_task_queue(struct io_kiocb *req)
2143{
2144	req->io_task_work.func = io_req_task_submit;
2145	io_req_task_work_add(req);
2146}
2147
2148static void io_req_task_queue_reissue(struct io_kiocb *req)
2149{
2150	req->io_task_work.func = io_queue_async_work;
2151	io_req_task_work_add(req);
2152}
2153
2154static inline void io_queue_next(struct io_kiocb *req)
2155{
2156	struct io_kiocb *nxt = io_req_find_next(req);
2157
2158	if (nxt)
2159		io_req_task_queue(nxt);
2160}
2161
2162static void io_free_req(struct io_kiocb *req)
2163{
2164	io_queue_next(req);
2165	__io_free_req(req);
2166}
2167
2168struct req_batch {
2169	struct task_struct	*task;
2170	int			task_refs;
2171	int			ctx_refs;
2172};
2173
2174static inline void io_init_req_batch(struct req_batch *rb)
2175{
2176	rb->task_refs = 0;
2177	rb->ctx_refs = 0;
2178	rb->task = NULL;
2179}
2180
2181static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2182				     struct req_batch *rb)
2183{
2184	if (rb->ctx_refs)
2185		percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2186	if (rb->task)
2187		io_put_task(rb->task, rb->task_refs);
2188}
2189
2190static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2191			      struct io_submit_state *state)
2192{
2193	io_queue_next(req);
2194	io_dismantle_req(req);
2195
2196	if (req->task != rb->task) {
2197		if (rb->task)
2198			io_put_task(rb->task, rb->task_refs);
2199		rb->task = req->task;
2200		rb->task_refs = 0;
2201	}
2202	rb->task_refs++;
2203	rb->ctx_refs++;
2204
2205	if (state->free_reqs != ARRAY_SIZE(state->reqs))
2206		state->reqs[state->free_reqs++] = req;
2207	else
2208		list_add(&req->inflight_entry, &state->free_list);
2209}
2210
2211static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2212	__must_hold(&ctx->uring_lock)
2213{
2214	struct io_submit_state *state = &ctx->submit_state;
2215	int i, nr = state->compl_nr;
2216	struct req_batch rb;
2217
2218	spin_lock(&ctx->completion_lock);
2219	for (i = 0; i < nr; i++) {
2220		struct io_kiocb *req = state->compl_reqs[i];
2221
2222		__io_cqring_fill_event(ctx, req->user_data, req->result,
2223					req->compl.cflags);
2224	}
2225	io_commit_cqring(ctx);
2226	spin_unlock(&ctx->completion_lock);
2227	io_cqring_ev_posted(ctx);
2228
2229	io_init_req_batch(&rb);
2230	for (i = 0; i < nr; i++) {
2231		struct io_kiocb *req = state->compl_reqs[i];
2232
2233		if (req_ref_put_and_test(req))
2234			io_req_free_batch(&rb, req, &ctx->submit_state);
2235	}
2236
2237	io_req_free_batch_finish(ctx, &rb);
2238	state->compl_nr = 0;
2239}
2240
2241/*
2242 * Drop reference to request, return next in chain (if there is one) if this
2243 * was the last reference to this request.
2244 */
2245static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2246{
2247	struct io_kiocb *nxt = NULL;
2248
2249	if (req_ref_put_and_test(req)) {
2250		nxt = io_req_find_next(req);
2251		__io_free_req(req);
2252	}
2253	return nxt;
2254}
2255
2256static inline void io_put_req(struct io_kiocb *req)
2257{
2258	if (req_ref_put_and_test(req))
2259		io_free_req(req);
2260}
2261
2262static inline void io_put_req_deferred(struct io_kiocb *req)
2263{
2264	if (req_ref_put_and_test(req)) {
2265		req->io_task_work.func = io_free_req;
2266		io_req_task_work_add(req);
2267	}
2268}
2269
2270static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2271{
2272	/* See comment at the top of this file */
2273	smp_rmb();
2274	return __io_cqring_events(ctx);
2275}
2276
2277static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2278{
2279	struct io_rings *rings = ctx->rings;
2280
2281	/* make sure SQ entry isn't read before tail */
2282	return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2283}
2284
2285static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2286{
2287	unsigned int cflags;
2288
2289	cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2290	cflags |= IORING_CQE_F_BUFFER;
2291	req->flags &= ~REQ_F_BUFFER_SELECTED;
2292	kfree(kbuf);
2293	return cflags;
2294}
2295
2296static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2297{
2298	struct io_buffer *kbuf;
2299
2300	if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
2301		return 0;
2302	kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2303	return io_put_kbuf(req, kbuf);
2304}
2305
2306static inline bool io_run_task_work(void)
2307{
2308	if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2309		__set_current_state(TASK_RUNNING);
2310		tracehook_notify_signal();
2311		return true;
2312	}
2313
2314	return false;
2315}
2316
2317/*
2318 * Find and free completed poll iocbs
2319 */
2320static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2321			       struct list_head *done)
2322{
2323	struct req_batch rb;
2324	struct io_kiocb *req;
2325
2326	/* order with ->result store in io_complete_rw_iopoll() */
2327	smp_rmb();
2328
2329	io_init_req_batch(&rb);
2330	while (!list_empty(done)) {
2331		req = list_first_entry(done, struct io_kiocb, inflight_entry);
2332		list_del(&req->inflight_entry);
2333
2334		if (READ_ONCE(req->result) == -EAGAIN &&
2335		    !(req->flags & REQ_F_DONT_REISSUE)) {
2336			req->iopoll_completed = 0;
2337			io_req_task_queue_reissue(req);
2338			continue;
2339		}
2340
2341		__io_cqring_fill_event(ctx, req->user_data, req->result,
2342					io_put_rw_kbuf(req));
2343		(*nr_events)++;
2344
2345		if (req_ref_put_and_test(req))
2346			io_req_free_batch(&rb, req, &ctx->submit_state);
2347	}
2348
2349	io_commit_cqring(ctx);
2350	io_cqring_ev_posted_iopoll(ctx);
2351	io_req_free_batch_finish(ctx, &rb);
2352}
2353
2354static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2355			long min)
2356{
2357	struct io_kiocb *req, *tmp;
2358	LIST_HEAD(done);
2359	bool spin;
2360
2361	/*
2362	 * Only spin for completions if we don't have multiple devices hanging
2363	 * off our complete list, and we're under the requested amount.
2364	 */
2365	spin = !ctx->poll_multi_queue && *nr_events < min;
2366
2367	list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2368		struct kiocb *kiocb = &req->rw.kiocb;
2369		int ret;
2370
2371		/*
2372		 * Move completed and retryable entries to our local lists.
2373		 * If we find a request that requires polling, break out
2374		 * and complete those lists first, if we have entries there.
2375		 */
2376		if (READ_ONCE(req->iopoll_completed)) {
2377			list_move_tail(&req->inflight_entry, &done);
2378			continue;
2379		}
2380		if (!list_empty(&done))
2381			break;
2382
2383		ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2384		if (unlikely(ret < 0))
2385			return ret;
2386		else if (ret)
2387			spin = false;
2388
2389		/* iopoll may have completed current req */
2390		if (READ_ONCE(req->iopoll_completed))
2391			list_move_tail(&req->inflight_entry, &done);
2392	}
2393
2394	if (!list_empty(&done))
2395		io_iopoll_complete(ctx, nr_events, &done);
2396
2397	return 0;
2398}
2399
2400/*
2401 * We can't just wait for polled events to come to us, we have to actively
2402 * find and complete them.
2403 */
2404static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2405{
2406	if (!(ctx->flags & IORING_SETUP_IOPOLL))
2407		return;
2408
2409	mutex_lock(&ctx->uring_lock);
2410	while (!list_empty(&ctx->iopoll_list)) {
2411		unsigned int nr_events = 0;
2412
2413		io_do_iopoll(ctx, &nr_events, 0);
2414
2415		/* let it sleep and repeat later if can't complete a request */
2416		if (nr_events == 0)
2417			break;
2418		/*
2419		 * Ensure we allow local-to-the-cpu processing to take place,
2420		 * in this case we need to ensure that we reap all events.
2421		 * Also let task_work, etc. to progress by releasing the mutex
2422		 */
2423		if (need_resched()) {
2424			mutex_unlock(&ctx->uring_lock);
2425			cond_resched();
2426			mutex_lock(&ctx->uring_lock);
2427		}
2428	}
2429	mutex_unlock(&ctx->uring_lock);
2430}
2431
2432static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2433{
2434	unsigned int nr_events = 0;
2435	int ret = 0;
2436
2437	/*
2438	 * We disallow the app entering submit/complete with polling, but we
2439	 * still need to lock the ring to prevent racing with polled issue
2440	 * that got punted to a workqueue.
2441	 */
2442	mutex_lock(&ctx->uring_lock);
2443	/*
2444	 * Don't enter poll loop if we already have events pending.
2445	 * If we do, we can potentially be spinning for commands that
2446	 * already triggered a CQE (eg in error).
2447	 */
2448	if (test_bit(0, &ctx->check_cq_overflow))
2449		__io_cqring_overflow_flush(ctx, false);
2450	if (io_cqring_events(ctx))
2451		goto out;
2452	do {
2453		/*
2454		 * If a submit got punted to a workqueue, we can have the
2455		 * application entering polling for a command before it gets
2456		 * issued. That app will hold the uring_lock for the duration
2457		 * of the poll right here, so we need to take a breather every
2458		 * now and then to ensure that the issue has a chance to add
2459		 * the poll to the issued list. Otherwise we can spin here
2460		 * forever, while the workqueue is stuck trying to acquire the
2461		 * very same mutex.
2462		 */
2463		if (list_empty(&ctx->iopoll_list)) {
2464			u32 tail = ctx->cached_cq_tail;
2465
2466			mutex_unlock(&ctx->uring_lock);
2467			io_run_task_work();
2468			mutex_lock(&ctx->uring_lock);
2469
2470			/* some requests don't go through iopoll_list */
2471			if (tail != ctx->cached_cq_tail ||
2472			    list_empty(&ctx->iopoll_list))
2473				break;
2474		}
2475		ret = io_do_iopoll(ctx, &nr_events, min);
2476	} while (!ret && nr_events < min && !need_resched());
2477out:
2478	mutex_unlock(&ctx->uring_lock);
2479	return ret;
2480}
2481
2482static void kiocb_end_write(struct io_kiocb *req)
2483{
2484	/*
2485	 * Tell lockdep we inherited freeze protection from submission
2486	 * thread.
2487	 */
2488	if (req->flags & REQ_F_ISREG) {
2489		struct super_block *sb = file_inode(req->file)->i_sb;
2490
2491		__sb_writers_acquired(sb, SB_FREEZE_WRITE);
2492		sb_end_write(sb);
2493	}
2494}
2495
2496#ifdef CONFIG_BLOCK
2497static bool io_resubmit_prep(struct io_kiocb *req)
2498{
2499	struct io_async_rw *rw = req->async_data;
2500
2501	if (!rw)
2502		return !io_req_prep_async(req);
2503	/* may have left rw->iter inconsistent on -EIOCBQUEUED */
2504	iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2505	return true;
2506}
2507
2508static bool io_rw_should_reissue(struct io_kiocb *req)
2509{
2510	umode_t mode = file_inode(req->file)->i_mode;
2511	struct io_ring_ctx *ctx = req->ctx;
2512
2513	if (!S_ISBLK(mode) && !S_ISREG(mode))
2514		return false;
2515	if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2516	    !(ctx->flags & IORING_SETUP_IOPOLL)))
2517		return false;
2518	/*
2519	 * If ref is dying, we might be running poll reap from the exit work.
2520	 * Don't attempt to reissue from that path, just let it fail with
2521	 * -EAGAIN.
2522	 */
2523	if (percpu_ref_is_dying(&ctx->refs))
2524		return false;
2525	/*
2526	 * Play it safe and assume not safe to re-import and reissue if we're
2527	 * not in the original thread group (or in task context).
2528	 */
2529	if (!same_thread_group(req->task, current) || !in_task())
2530		return false;
2531	return true;
2532}
2533#else
2534static bool io_resubmit_prep(struct io_kiocb *req)
2535{
2536	return false;
2537}
2538static bool io_rw_should_reissue(struct io_kiocb *req)
2539{
2540	return false;
2541}
2542#endif
2543
2544static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2545{
2546	if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2547		kiocb_end_write(req);
2548	if (res != req->result) {
2549		if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2550		    io_rw_should_reissue(req)) {
2551			req->flags |= REQ_F_REISSUE;
2552			return true;
2553		}
2554		req_set_fail(req);
2555		req->result = res;
2556	}
2557	return false;
2558}
2559
2560static void io_req_task_complete(struct io_kiocb *req)
2561{
2562	__io_req_complete(req, 0, req->result, io_put_rw_kbuf(req));
2563}
2564
2565static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2566			     unsigned int issue_flags)
2567{
2568	if (__io_complete_rw_common(req, res))
2569		return;
2570	io_req_task_complete(req);
2571}
2572
2573static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2574{
2575	struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2576
2577	if (__io_complete_rw_common(req, res))
2578		return;
2579	req->result = res;
2580	req->io_task_work.func = io_req_task_complete;
2581	io_req_task_work_add(req);
2582}
2583
2584static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2585{
2586	struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2587
2588	if (kiocb->ki_flags & IOCB_WRITE)
2589		kiocb_end_write(req);
2590	if (unlikely(res != req->result)) {
2591		if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2592		    io_resubmit_prep(req))) {
2593			req_set_fail(req);
2594			req->flags |= REQ_F_DONT_REISSUE;
2595		}
2596	}
2597
2598	WRITE_ONCE(req->result, res);
2599	/* order with io_iopoll_complete() checking ->result */
2600	smp_wmb();
2601	WRITE_ONCE(req->iopoll_completed, 1);
2602}
2603
2604/*
2605 * After the iocb has been issued, it's safe to be found on the poll list.
2606 * Adding the kiocb to the list AFTER submission ensures that we don't
2607 * find it from a io_do_iopoll() thread before the issuer is done
2608 * accessing the kiocb cookie.
2609 */
2610static void io_iopoll_req_issued(struct io_kiocb *req)
2611{
2612	struct io_ring_ctx *ctx = req->ctx;
2613	const bool in_async = io_wq_current_is_worker();
2614
2615	/* workqueue context doesn't hold uring_lock, grab it now */
2616	if (unlikely(in_async))
2617		mutex_lock(&ctx->uring_lock);
2618
2619	/*
2620	 * Track whether we have multiple files in our lists. This will impact
2621	 * how we do polling eventually, not spinning if we're on potentially
2622	 * different devices.
2623	 */
2624	if (list_empty(&ctx->iopoll_list)) {
2625		ctx->poll_multi_queue = false;
2626	} else if (!ctx->poll_multi_queue) {
2627		struct io_kiocb *list_req;
2628		unsigned int queue_num0, queue_num1;
2629
2630		list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2631						inflight_entry);
2632
2633		if (list_req->file != req->file) {
2634			ctx->poll_multi_queue = true;
2635		} else {
2636			queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2637			queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2638			if (queue_num0 != queue_num1)
2639				ctx->poll_multi_queue = true;
2640		}
2641	}
2642
2643	/*
2644	 * For fast devices, IO may have already completed. If it has, add
2645	 * it to the front so we find it first.
2646	 */
2647	if (READ_ONCE(req->iopoll_completed))
2648		list_add(&req->inflight_entry, &ctx->iopoll_list);
2649	else
2650		list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2651
2652	if (unlikely(in_async)) {
2653		/*
2654		 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2655		 * in sq thread task context or in io worker task context. If
2656		 * current task context is sq thread, we don't need to check
2657		 * whether should wake up sq thread.
2658		 */
2659		if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2660		    wq_has_sleeper(&ctx->sq_data->wait))
2661			wake_up(&ctx->sq_data->wait);
2662
2663		mutex_unlock(&ctx->uring_lock);
2664	}
2665}
2666
2667static bool io_bdev_nowait(struct block_device *bdev)
2668{
2669	return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2670}
2671
2672/*
2673 * If we tracked the file through the SCM inflight mechanism, we could support
2674 * any file. For now, just ensure that anything potentially problematic is done
2675 * inline.
2676 */
2677static bool __io_file_supports_nowait(struct file *file, int rw)
2678{
2679	umode_t mode = file_inode(file)->i_mode;
2680
2681	if (S_ISBLK(mode)) {
2682		if (IS_ENABLED(CONFIG_BLOCK) &&
2683		    io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2684			return true;
2685		return false;
2686	}
2687	if (S_ISSOCK(mode))
2688		return true;
2689	if (S_ISREG(mode)) {
2690		if (IS_ENABLED(CONFIG_BLOCK) &&
2691		    io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2692		    file->f_op != &io_uring_fops)
2693			return true;
2694		return false;
2695	}
2696
2697	/* any ->read/write should understand O_NONBLOCK */
2698	if (file->f_flags & O_NONBLOCK)
2699		return true;
2700
2701	if (!(file->f_mode & FMODE_NOWAIT))
2702		return false;
2703
2704	if (rw == READ)
2705		return file->f_op->read_iter != NULL;
2706
2707	return file->f_op->write_iter != NULL;
2708}
2709
2710static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2711{
2712	if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2713		return true;
2714	else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2715		return true;
2716
2717	return __io_file_supports_nowait(req->file, rw);
2718}
2719
2720static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2721{
2722	struct io_ring_ctx *ctx = req->ctx;
2723	struct kiocb *kiocb = &req->rw.kiocb;
2724	struct file *file = req->file;
2725	unsigned ioprio;
2726	int ret;
2727
2728	if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2729		req->flags |= REQ_F_ISREG;
2730
2731	kiocb->ki_pos = READ_ONCE(sqe->off);
2732	if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2733		req->flags |= REQ_F_CUR_POS;
2734		kiocb->ki_pos = file->f_pos;
2735	}
2736	kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2737	kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2738	ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2739	if (unlikely(ret))
2740		return ret;
2741
2742	/* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2743	if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2744		req->flags |= REQ_F_NOWAIT;
2745
2746	ioprio = READ_ONCE(sqe->ioprio);
2747	if (ioprio) {
2748		ret = ioprio_check_cap(ioprio);
2749		if (ret)
2750			return ret;
2751
2752		kiocb->ki_ioprio = ioprio;
2753	} else
2754		kiocb->ki_ioprio = get_current_ioprio();
2755
2756	if (ctx->flags & IORING_SETUP_IOPOLL) {
2757		if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2758		    !kiocb->ki_filp->f_op->iopoll)
2759			return -EOPNOTSUPP;
2760
2761		kiocb->ki_flags |= IOCB_HIPRI;
2762		kiocb->ki_complete = io_complete_rw_iopoll;
2763		req->iopoll_completed = 0;
2764	} else {
2765		if (kiocb->ki_flags & IOCB_HIPRI)
2766			return -EINVAL;
2767		kiocb->ki_complete = io_complete_rw;
2768	}
2769
2770	if (req->opcode == IORING_OP_READ_FIXED ||
2771	    req->opcode == IORING_OP_WRITE_FIXED) {
2772		req->imu = NULL;
2773		io_req_set_rsrc_node(req);
2774	}
2775
2776	req->rw.addr = READ_ONCE(sqe->addr);
2777	req->rw.len = READ_ONCE(sqe->len);
2778	req->buf_index = READ_ONCE(sqe->buf_index);
2779	return 0;
2780}
2781
2782static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2783{
2784	switch (ret) {
2785	case -EIOCBQUEUED:
2786		break;
2787	case -ERESTARTSYS:
2788	case -ERESTARTNOINTR:
2789	case -ERESTARTNOHAND:
2790	case -ERESTART_RESTARTBLOCK:
2791		/*
2792		 * We can't just restart the syscall, since previously
2793		 * submitted sqes may already be in progress. Just fail this
2794		 * IO with EINTR.
2795		 */
2796		ret = -EINTR;
2797		fallthrough;
2798	default:
2799		kiocb->ki_complete(kiocb, ret, 0);
2800	}
2801}
2802
2803static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2804		       unsigned int issue_flags)
2805{
2806	struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2807	struct io_async_rw *io = req->async_data;
2808	bool check_reissue = kiocb->ki_complete == io_complete_rw;
2809
2810	/* add previously done IO, if any */
2811	if (io && io->bytes_done > 0) {
2812		if (ret < 0)
2813			ret = io->bytes_done;
2814		else
2815			ret += io->bytes_done;
2816	}
2817
2818	if (req->flags & REQ_F_CUR_POS)
2819		req->file->f_pos = kiocb->ki_pos;
2820	if (ret >= 0 && check_reissue)
2821		__io_complete_rw(req, ret, 0, issue_flags);
2822	else
2823		io_rw_done(kiocb, ret);
2824
2825	if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2826		req->flags &= ~REQ_F_REISSUE;
2827		if (io_resubmit_prep(req)) {
2828			io_req_task_queue_reissue(req);
2829		} else {
2830			req_set_fail(req);
2831			__io_req_complete(req, issue_flags, ret,
2832					  io_put_rw_kbuf(req));
2833		}
2834	}
2835}
2836
2837static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2838			     struct io_mapped_ubuf *imu)
2839{
2840	size_t len = req->rw.len;
2841	u64 buf_end, buf_addr = req->rw.addr;
2842	size_t offset;
2843
2844	if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2845		return -EFAULT;
2846	/* not inside the mapped region */
2847	if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2848		return -EFAULT;
2849
2850	/*
2851	 * May not be a start of buffer, set size appropriately
2852	 * and advance us to the beginning.
2853	 */
2854	offset = buf_addr - imu->ubuf;
2855	iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2856
2857	if (offset) {
2858		/*
2859		 * Don't use iov_iter_advance() here, as it's really slow for
2860		 * using the latter parts of a big fixed buffer - it iterates
2861		 * over each segment manually. We can cheat a bit here, because
2862		 * we know that:
2863		 *
2864		 * 1) it's a BVEC iter, we set it up
2865		 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2866		 *    first and last bvec
2867		 *
2868		 * So just find our index, and adjust the iterator afterwards.
2869		 * If the offset is within the first bvec (or the whole first
2870		 * bvec, just use iov_iter_advance(). This makes it easier
2871		 * since we can just skip the first segment, which may not
2872		 * be PAGE_SIZE aligned.
2873		 */
2874		const struct bio_vec *bvec = imu->bvec;
2875
2876		if (offset <= bvec->bv_len) {
2877			iov_iter_advance(iter, offset);
2878		} else {
2879			unsigned long seg_skip;
2880
2881			/* skip first vec */
2882			offset -= bvec->bv_len;
2883			seg_skip = 1 + (offset >> PAGE_SHIFT);
2884
2885			iter->bvec = bvec + seg_skip;
2886			iter->nr_segs -= seg_skip;
2887			iter->count -= bvec->bv_len + offset;
2888			iter->iov_offset = offset & ~PAGE_MASK;
2889		}
2890	}
2891
2892	return 0;
2893}
2894
2895static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2896{
2897	struct io_ring_ctx *ctx = req->ctx;
2898	struct io_mapped_ubuf *imu = req->imu;
2899	u16 index, buf_index = req->buf_index;
2900
2901	if (likely(!imu)) {
2902		if (unlikely(buf_index >= ctx->nr_user_bufs))
2903			return -EFAULT;
2904		index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2905		imu = READ_ONCE(ctx->user_bufs[index]);
2906		req->imu = imu;
2907	}
2908	return __io_import_fixed(req, rw, iter, imu);
2909}
2910
2911static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2912{
2913	if (needs_lock)
2914		mutex_unlock(&ctx->uring_lock);
2915}
2916
2917static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2918{
2919	/*
2920	 * "Normal" inline submissions always hold the uring_lock, since we
2921	 * grab it from the system call. Same is true for the SQPOLL offload.
2922	 * The only exception is when we've detached the request and issue it
2923	 * from an async worker thread, grab the lock for that case.
2924	 */
2925	if (needs_lock)
2926		mutex_lock(&ctx->uring_lock);
2927}
2928
2929static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2930					  int bgid, struct io_buffer *kbuf,
2931					  bool needs_lock)
2932{
2933	struct io_buffer *head;
2934
2935	if (req->flags & REQ_F_BUFFER_SELECTED)
2936		return kbuf;
2937
2938	io_ring_submit_lock(req->ctx, needs_lock);
2939
2940	lockdep_assert_held(&req->ctx->uring_lock);
2941
2942	head = xa_load(&req->ctx->io_buffers, bgid);
2943	if (head) {
2944		if (!list_empty(&head->list)) {
2945			kbuf = list_last_entry(&head->list, struct io_buffer,
2946							list);
2947			list_del(&kbuf->list);
2948		} else {
2949			kbuf = head;
2950			xa_erase(&req->ctx->io_buffers, bgid);
2951		}
2952		if (*len > kbuf->len)
2953			*len = kbuf->len;
2954	} else {
2955		kbuf = ERR_PTR(-ENOBUFS);
2956	}
2957
2958	io_ring_submit_unlock(req->ctx, needs_lock);
2959
2960	return kbuf;
2961}
2962
2963static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2964					bool needs_lock)
2965{
2966	struct io_buffer *kbuf;
2967	u16 bgid;
2968
2969	kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2970	bgid = req->buf_index;
2971	kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2972	if (IS_ERR(kbuf))
2973		return kbuf;
2974	req->rw.addr = (u64) (unsigned long) kbuf;
2975	req->flags |= REQ_F_BUFFER_SELECTED;
2976	return u64_to_user_ptr(kbuf->addr);
2977}
2978
2979#ifdef CONFIG_COMPAT
2980static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2981				bool needs_lock)
2982{
2983	struct compat_iovec __user *uiov;
2984	compat_ssize_t clen;
2985	void __user *buf;
2986	ssize_t len;
2987
2988	uiov = u64_to_user_ptr(req->rw.addr);
2989	if (!access_ok(uiov, sizeof(*uiov)))
2990		return -EFAULT;
2991	if (__get_user(clen, &uiov->iov_len))
2992		return -EFAULT;
2993	if (clen < 0)
2994		return -EINVAL;
2995
2996	len = clen;
2997	buf = io_rw_buffer_select(req, &len, needs_lock);
2998	if (IS_ERR(buf))
2999		return PTR_ERR(buf);
3000	iov[0].iov_base = buf;
3001	iov[0].iov_len = (compat_size_t) len;
3002	return 0;
3003}
3004#endif
3005
3006static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3007				      bool needs_lock)
3008{
3009	struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3010	void __user *buf;
3011	ssize_t len;
3012
3013	if (copy_from_user(iov, uiov, sizeof(*uiov)))
3014		return -EFAULT;
3015
3016	len = iov[0].iov_len;
3017	if (len < 0)
3018		return -EINVAL;
3019	buf = io_rw_buffer_select(req, &len, needs_lock);
3020	if (IS_ERR(buf))
3021		return PTR_ERR(buf);
3022	iov[0].iov_base = buf;
3023	iov[0].iov_len = len;
3024	return 0;
3025}
3026
3027static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3028				    bool needs_lock)
3029{
3030	if (req->flags & REQ_F_BUFFER_SELECTED) {
3031		struct io_buffer *kbuf;
3032
3033		kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3034		iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3035		iov[0].iov_len = kbuf->len;
3036		return 0;
3037	}
3038	if (req->rw.len != 1)
3039		return -EINVAL;
3040
3041#ifdef CONFIG_COMPAT
3042	if (req->ctx->compat)
3043		return io_compat_import(req, iov, needs_lock);
3044#endif
3045
3046	return __io_iov_buffer_select(req, iov, needs_lock);
3047}
3048
3049static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3050			   struct iov_iter *iter, bool needs_lock)
3051{
3052	void __user *buf = u64_to_user_ptr(req->rw.addr);
3053	size_t sqe_len = req->rw.len;
3054	u8 opcode = req->opcode;
3055	ssize_t ret;
3056
3057	if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3058		*iovec = NULL;
3059		return io_import_fixed(req, rw, iter);
3060	}
3061
3062	/* buffer index only valid with fixed read/write, or buffer select  */
3063	if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3064		return -EINVAL;
3065
3066	if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3067		if (req->flags & REQ_F_BUFFER_SELECT) {
3068			buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3069			if (IS_ERR(buf))
3070				return PTR_ERR(buf);
3071			req->rw.len = sqe_len;
3072		}
3073
3074		ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3075		*iovec = NULL;
3076		return ret;
3077	}
3078
3079	if (req->flags & REQ_F_BUFFER_SELECT) {
3080		ret = io_iov_buffer_select(req, *iovec, needs_lock);
3081		if (!ret)
3082			iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3083		*iovec = NULL;
3084		return ret;
3085	}
3086
3087	return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3088			      req->ctx->compat);
3089}
3090
3091static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3092{
3093	return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3094}
3095
3096/*
3097 * For files that don't have ->read_iter() and ->write_iter(), handle them
3098 * by looping over ->read() or ->write() manually.
3099 */
3100static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3101{
3102	struct kiocb *kiocb = &req->rw.kiocb;
3103	struct file *file = req->file;
3104	ssize_t ret = 0;
3105
3106	/*
3107	 * Don't support polled IO through this interface, and we can't
3108	 * support non-blocking either. For the latter, this just causes
3109	 * the kiocb to be handled from an async context.
3110	 */
3111	if (kiocb->ki_flags & IOCB_HIPRI)
3112		return -EOPNOTSUPP;
3113	if (kiocb->ki_flags & IOCB_NOWAIT)
3114		return -EAGAIN;
3115
3116	while (iov_iter_count(iter)) {
3117		struct iovec iovec;
3118		ssize_t nr;
3119
3120		if (!iov_iter_is_bvec(iter)) {
3121			iovec = iov_iter_iovec(iter);
3122		} else {
3123			iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3124			iovec.iov_len = req->rw.len;
3125		}
3126
3127		if (rw == READ) {
3128			nr = file->f_op->read(file, iovec.iov_base,
3129					      iovec.iov_len, io_kiocb_ppos(kiocb));
3130		} else {
3131			nr = file->f_op->write(file, iovec.iov_base,
3132					       iovec.iov_len, io_kiocb_ppos(kiocb));
3133		}
3134
3135		if (nr < 0) {
3136			if (!ret)
3137				ret = nr;
3138			break;
3139		}
3140		ret += nr;
3141		if (nr != iovec.iov_len)
3142			break;
3143		req->rw.len -= nr;
3144		req->rw.addr += nr;
3145		iov_iter_advance(iter, nr);
3146	}
3147
3148	return ret;
3149}
3150
3151static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3152			  const struct iovec *fast_iov, struct iov_iter *iter)
3153{
3154	struct io_async_rw *rw = req->async_data;
3155
3156	memcpy(&rw->iter, iter, sizeof(*iter));
3157	rw->free_iovec = iovec;
3158	rw->bytes_done = 0;
3159	/* can only be fixed buffers, no need to do anything */
3160	if (iov_iter_is_bvec(iter))
3161		return;
3162	if (!iovec) {
3163		unsigned iov_off = 0;
3164
3165		rw->iter.iov = rw->fast_iov;
3166		if (iter->iov != fast_iov) {
3167			iov_off = iter->iov - fast_iov;
3168			rw->iter.iov += iov_off;
3169		}
3170		if (rw->fast_iov != fast_iov)
3171			memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3172			       sizeof(struct iovec) * iter->nr_segs);
3173	} else {
3174		req->flags |= REQ_F_NEED_CLEANUP;
3175	}
3176}
3177
3178static inline int io_alloc_async_data(struct io_kiocb *req)
3179{
3180	WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3181	req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3182	return req->async_data == NULL;
3183}
3184
3185static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3186			     const struct iovec *fast_iov,
3187			     struct iov_iter *iter, bool force)
3188{
3189	if (!force && !io_op_defs[req->opcode].needs_async_setup)
3190		return 0;
3191	if (!req->async_data) {
3192		if (io_alloc_async_data(req)) {
3193			kfree(iovec);
3194			return -ENOMEM;
3195		}
3196
3197		io_req_map_rw(req, iovec, fast_iov, iter);
3198	}
3199	return 0;
3200}
3201
3202static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3203{
3204	struct io_async_rw *iorw = req->async_data;
3205	struct iovec *iov = iorw->fast_iov;
3206	int ret;
3207
3208	ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3209	if (unlikely(ret < 0))
3210		return ret;
3211
3212	iorw->bytes_done = 0;
3213	iorw->free_iovec = iov;
3214	if (iov)
3215		req->flags |= REQ_F_NEED_CLEANUP;
3216	return 0;
3217}
3218
3219static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3220{
3221	if (unlikely(!(req->file->f_mode & FMODE_READ)))
3222		return -EBADF;
3223	return io_prep_rw(req, sqe);
3224}
3225
3226/*
3227 * This is our waitqueue callback handler, registered through lock_page_async()
3228 * when we initially tried to do the IO with the iocb armed our waitqueue.
3229 * This gets called when the page is unlocked, and we generally expect that to
3230 * happen when the page IO is completed and the page is now uptodate. This will
3231 * queue a task_work based retry of the operation, attempting to copy the data
3232 * again. If the latter fails because the page was NOT uptodate, then we will
3233 * do a thread based blocking retry of the operation. That's the unexpected
3234 * slow path.
3235 */
3236static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3237			     int sync, void *arg)
3238{
3239	struct wait_page_queue *wpq;
3240	struct io_kiocb *req = wait->private;
3241	struct wait_page_key *key = arg;
3242
3243	wpq = container_of(wait, struct wait_page_queue, wait);
3244
3245	if (!wake_page_match(wpq, key))
3246		return 0;
3247
3248	req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3249	list_del_init(&wait->entry);
3250	io_req_task_queue(req);
3251	return 1;
3252}
3253
3254/*
3255 * This controls whether a given IO request should be armed for async page
3256 * based retry. If we return false here, the request is handed to the async
3257 * worker threads for retry. If we're doing buffered reads on a regular file,
3258 * we prepare a private wait_page_queue entry and retry the operation. This
3259 * will either succeed because the page is now uptodate and unlocked, or it
3260 * will register a callback when the page is unlocked at IO completion. Through
3261 * that callback, io_uring uses task_work to setup a retry of the operation.
3262 * That retry will attempt the buffered read again. The retry will generally
3263 * succeed, or in rare cases where it fails, we then fall back to using the
3264 * async worker threads for a blocking retry.
3265 */
3266static bool io_rw_should_retry(struct io_kiocb *req)
3267{
3268	struct io_async_rw *rw = req->async_data;
3269	struct wait_page_queue *wait = &rw->wpq;
3270	struct kiocb *kiocb = &req->rw.kiocb;
3271
3272	/* never retry for NOWAIT, we just complete with -EAGAIN */
3273	if (req->flags & REQ_F_NOWAIT)
3274		return false;
3275
3276	/* Only for buffered IO */
3277	if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3278		return false;
3279
3280	/*
3281	 * just use poll if we can, and don't attempt if the fs doesn't
3282	 * support callback based unlocks
3283	 */
3284	if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3285		return false;
3286
3287	wait->wait.func = io_async_buf_func;
3288	wait->wait.private = req;
3289	wait->wait.flags = 0;
3290	INIT_LIST_HEAD(&wait->wait.entry);
3291	kiocb->ki_flags |= IOCB_WAITQ;
3292	kiocb->ki_flags &= ~IOCB_NOWAIT;
3293	kiocb->ki_waitq = wait;
3294	return true;
3295}
3296
3297static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3298{
3299	if (req->file->f_op->read_iter)
3300		return call_read_iter(req->file, &req->rw.kiocb, iter);
3301	else if (req->file->f_op->read)
3302		return loop_rw_iter(READ, req, iter);
3303	else
3304		return -EINVAL;
3305}
3306
3307static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3308{
3309	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3310	struct kiocb *kiocb = &req->rw.kiocb;
3311	struct iov_iter __iter, *iter = &__iter;
3312	struct io_async_rw *rw = req->async_data;
3313	ssize_t io_size, ret, ret2;
3314	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3315
3316	if (rw) {
3317		iter = &rw->iter;
3318		iovec = NULL;
3319	} else {
3320		ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3321		if (ret < 0)
3322			return ret;
3323	}
3324	io_size = iov_iter_count(iter);
3325	req->result = io_size;
3326
3327	/* Ensure we clear previously set non-block flag */
3328	if (!force_nonblock)
3329		kiocb->ki_flags &= ~IOCB_NOWAIT;
3330	else
3331		kiocb->ki_flags |= IOCB_NOWAIT;
3332
3333	/* If the file doesn't support async, just async punt */
3334	if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3335		ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3336		return ret ?: -EAGAIN;
3337	}
3338
3339	ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3340	if (unlikely(ret)) {
3341		kfree(iovec);
3342		return ret;
3343	}
3344
3345	ret = io_iter_do_read(req, iter);
3346
3347	if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3348		req->flags &= ~REQ_F_REISSUE;
3349		/* IOPOLL retry should happen for io-wq threads */
3350		if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3351			goto done;
3352		/* no retry on NONBLOCK nor RWF_NOWAIT */
3353		if (req->flags & REQ_F_NOWAIT)
3354			goto done;
3355		/* some cases will consume bytes even on error returns */
3356		iov_iter_revert(iter, io_size - iov_iter_count(iter));
3357		ret = 0;
3358	} else if (ret == -EIOCBQUEUED) {
3359		goto out_free;
3360	} else if (ret <= 0 || ret == io_size || !force_nonblock ||
3361		   (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3362		/* read all, failed, already did sync or don't want to retry */
3363		goto done;
3364	}
3365
3366	ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3367	if (ret2)
3368		return ret2;
3369
3370	iovec = NULL;
3371	rw = req->async_data;
3372	/* now use our persistent iterator, if we aren't already */
3373	iter = &rw->iter;
3374
3375	do {
3376		io_size -= ret;
3377		rw->bytes_done += ret;
3378		/* if we can retry, do so with the callbacks armed */
3379		if (!io_rw_should_retry(req)) {
3380			kiocb->ki_flags &= ~IOCB_WAITQ;
3381			return -EAGAIN;
3382		}
3383
3384		/*
3385		 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3386		 * we get -EIOCBQUEUED, then we'll get a notification when the
3387		 * desired page gets unlocked. We can also get a partial read
3388		 * here, and if we do, then just retry at the new offset.
3389		 */
3390		ret = io_iter_do_read(req, iter);
3391		if (ret == -EIOCBQUEUED)
3392			return 0;
3393		/* we got some bytes, but not all. retry. */
3394		kiocb->ki_flags &= ~IOCB_WAITQ;
3395	} while (ret > 0 && ret < io_size);
3396done:
3397	kiocb_done(kiocb, ret, issue_flags);
3398out_free:
3399	/* it's faster to check here then delegate to kfree */
3400	if (iovec)
3401		kfree(iovec);
3402	return 0;
3403}
3404
3405static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3406{
3407	if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3408		return -EBADF;
3409	return io_prep_rw(req, sqe);
3410}
3411
3412static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3413{
3414	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3415	struct kiocb *kiocb = &req->rw.kiocb;
3416	struct iov_iter __iter, *iter = &__iter;
3417	struct io_async_rw *rw = req->async_data;
3418	ssize_t ret, ret2, io_size;
3419	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3420
3421	if (rw) {
3422		iter = &rw->iter;
3423		iovec = NULL;
3424	} else {
3425		ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3426		if (ret < 0)
3427			return ret;
3428	}
3429	io_size = iov_iter_count(iter);
3430	req->result = io_size;
3431
3432	/* Ensure we clear previously set non-block flag */
3433	if (!force_nonblock)
3434		kiocb->ki_flags &= ~IOCB_NOWAIT;
3435	else
3436		kiocb->ki_flags |= IOCB_NOWAIT;
3437
3438	/* If the file doesn't support async, just async punt */
3439	if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3440		goto copy_iov;
3441
3442	/* file path doesn't support NOWAIT for non-direct_IO */
3443	if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3444	    (req->flags & REQ_F_ISREG))
3445		goto copy_iov;
3446
3447	ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3448	if (unlikely(ret))
3449		goto out_free;
3450
3451	/*
3452	 * Open-code file_start_write here to grab freeze protection,
3453	 * which will be released by another thread in
3454	 * io_complete_rw().  Fool lockdep by telling it the lock got
3455	 * released so that it doesn't complain about the held lock when
3456	 * we return to userspace.
3457	 */
3458	if (req->flags & REQ_F_ISREG) {
3459		sb_start_write(file_inode(req->file)->i_sb);
3460		__sb_writers_release(file_inode(req->file)->i_sb,
3461					SB_FREEZE_WRITE);
3462	}
3463	kiocb->ki_flags |= IOCB_WRITE;
3464
3465	if (req->file->f_op->write_iter)
3466		ret2 = call_write_iter(req->file, kiocb, iter);
3467	else if (req->file->f_op->write)
3468		ret2 = loop_rw_iter(WRITE, req, iter);
3469	else
3470		ret2 = -EINVAL;
3471
3472	if (req->flags & REQ_F_REISSUE) {
3473		req->flags &= ~REQ_F_REISSUE;
3474		ret2 = -EAGAIN;
3475	}
3476
3477	/*
3478	 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3479	 * retry them without IOCB_NOWAIT.
3480	 */
3481	if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3482		ret2 = -EAGAIN;
3483	/* no retry on NONBLOCK nor RWF_NOWAIT */
3484	if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3485		goto done;
3486	if (!force_nonblock || ret2 != -EAGAIN) {
3487		/* IOPOLL retry should happen for io-wq threads */
3488		if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3489			goto copy_iov;
3490done:
3491		kiocb_done(kiocb, ret2, issue_flags);
3492	} else {
3493copy_iov:
3494		/* some cases will consume bytes even on error returns */
3495		iov_iter_revert(iter, io_size - iov_iter_count(iter));
3496		ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3497		return ret ?: -EAGAIN;
3498	}
3499out_free:
3500	/* it's reportedly faster than delegating the null check to kfree() */
3501	if (iovec)
3502		kfree(iovec);
3503	return ret;
3504}
3505
3506static int io_renameat_prep(struct io_kiocb *req,
3507			    const struct io_uring_sqe *sqe)
3508{
3509	struct io_rename *ren = &req->rename;
3510	const char __user *oldf, *newf;
3511
3512	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3513		return -EINVAL;
3514	if (sqe->ioprio || sqe->buf_index)
3515		return -EINVAL;
3516	if (unlikely(req->flags & REQ_F_FIXED_FILE))
3517		return -EBADF;
3518
3519	ren->old_dfd = READ_ONCE(sqe->fd);
3520	oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3521	newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3522	ren->new_dfd = READ_ONCE(sqe->len);
3523	ren->flags = READ_ONCE(sqe->rename_flags);
3524
3525	ren->oldpath = getname(oldf);
3526	if (IS_ERR(ren->oldpath))
3527		return PTR_ERR(ren->oldpath);
3528
3529	ren->newpath = getname(newf);
3530	if (IS_ERR(ren->newpath)) {
3531		putname(ren->oldpath);
3532		return PTR_ERR(ren->newpath);
3533	}
3534
3535	req->flags |= REQ_F_NEED_CLEANUP;
3536	return 0;
3537}
3538
3539static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3540{
3541	struct io_rename *ren = &req->rename;
3542	int ret;
3543
3544	if (issue_flags & IO_URING_F_NONBLOCK)
3545		return -EAGAIN;
3546
3547	ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3548				ren->newpath, ren->flags);
3549
3550	req->flags &= ~REQ_F_NEED_CLEANUP;
3551	if (ret < 0)
3552		req_set_fail(req);
3553	io_req_complete(req, ret);
3554	return 0;
3555}
3556
3557static int io_unlinkat_prep(struct io_kiocb *req,
3558			    const struct io_uring_sqe *sqe)
3559{
3560	struct io_unlink *un = &req->unlink;
3561	const char __user *fname;
3562
3563	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3564		return -EINVAL;
3565	if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3566		return -EINVAL;
3567	if (unlikely(req->flags & REQ_F_FIXED_FILE))
3568		return -EBADF;
3569
3570	un->dfd = READ_ONCE(sqe->fd);
3571
3572	un->flags = READ_ONCE(sqe->unlink_flags);
3573	if (un->flags & ~AT_REMOVEDIR)
3574		return -EINVAL;
3575
3576	fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3577	un->filename = getname(fname);
3578	if (IS_ERR(un->filename))
3579		return PTR_ERR(un->filename);
3580
3581	req->flags |= REQ_F_NEED_CLEANUP;
3582	return 0;
3583}
3584
3585static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3586{
3587	struct io_unlink *un = &req->unlink;
3588	int ret;
3589
3590	if (issue_flags & IO_URING_F_NONBLOCK)
3591		return -EAGAIN;
3592
3593	if (un->flags & AT_REMOVEDIR)
3594		ret = do_rmdir(un->dfd, un->filename);
3595	else
3596		ret = do_unlinkat(un->dfd, un->filename);
3597
3598	req->flags &= ~REQ_F_NEED_CLEANUP;
3599	if (ret < 0)
3600		req_set_fail(req);
3601	io_req_complete(req, ret);
3602	return 0;
3603}
3604
3605static int io_shutdown_prep(struct io_kiocb *req,
3606			    const struct io_uring_sqe *sqe)
3607{
3608#if defined(CONFIG_NET)
3609	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3610		return -EINVAL;
3611	if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3612	    sqe->buf_index)
3613		return -EINVAL;
3614
3615	req->shutdown.how = READ_ONCE(sqe->len);
3616	return 0;
3617#else
3618	return -EOPNOTSUPP;
3619#endif
3620}
3621
3622static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3623{
3624#if defined(CONFIG_NET)
3625	struct socket *sock;
3626	int ret;
3627
3628	if (issue_flags & IO_URING_F_NONBLOCK)
3629		return -EAGAIN;
3630
3631	sock = sock_from_file(req->file);
3632	if (unlikely(!sock))
3633		return -ENOTSOCK;
3634
3635	ret = __sys_shutdown_sock(sock, req->shutdown.how);
3636	if (ret < 0)
3637		req_set_fail(req);
3638	io_req_complete(req, ret);
3639	return 0;
3640#else
3641	return -EOPNOTSUPP;
3642#endif
3643}
3644
3645static int __io_splice_prep(struct io_kiocb *req,
3646			    const struct io_uring_sqe *sqe)
3647{
3648	struct io_splice *sp = &req->splice;
3649	unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3650
3651	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3652		return -EINVAL;
3653
3654	sp->file_in = NULL;
3655	sp->len = READ_ONCE(sqe->len);
3656	sp->flags = READ_ONCE(sqe->splice_flags);
3657
3658	if (unlikely(sp->flags & ~valid_flags))
3659		return -EINVAL;
3660
3661	sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3662				  (sp->flags & SPLICE_F_FD_IN_FIXED));
3663	if (!sp->file_in)
3664		return -EBADF;
3665	req->flags |= REQ_F_NEED_CLEANUP;
3666	return 0;
3667}
3668
3669static int io_tee_prep(struct io_kiocb *req,
3670		       const struct io_uring_sqe *sqe)
3671{
3672	if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3673		return -EINVAL;
3674	return __io_splice_prep(req, sqe);
3675}
3676
3677static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3678{
3679	struct io_splice *sp = &req->splice;
3680	struct file *in = sp->file_in;
3681	struct file *out = sp->file_out;
3682	unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3683	long ret = 0;
3684
3685	if (issue_flags & IO_URING_F_NONBLOCK)
3686		return -EAGAIN;
3687	if (sp->len)
3688		ret = do_tee(in, out, sp->len, flags);
3689
3690	if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3691		io_put_file(in);
3692	req->flags &= ~REQ_F_NEED_CLEANUP;
3693
3694	if (ret != sp->len)
3695		req_set_fail(req);
3696	io_req_complete(req, ret);
3697	return 0;
3698}
3699
3700static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3701{
3702	struct io_splice *sp = &req->splice;
3703
3704	sp->off_in = READ_ONCE(sqe->splice_off_in);
3705	sp->off_out = READ_ONCE(sqe->off);
3706	return __io_splice_prep(req, sqe);
3707}
3708
3709static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3710{
3711	struct io_splice *sp = &req->splice;
3712	struct file *in = sp->file_in;
3713	struct file *out = sp->file_out;
3714	unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3715	loff_t *poff_in, *poff_out;
3716	long ret = 0;
3717
3718	if (issue_flags & IO_URING_F_NONBLOCK)
3719		return -EAGAIN;
3720
3721	poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3722	poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3723
3724	if (sp->len)
3725		ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3726
3727	if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3728		io_put_file(in);
3729	req->flags &= ~REQ_F_NEED_CLEANUP;
3730
3731	if (ret != sp->len)
3732		req_set_fail(req);
3733	io_req_complete(req, ret);
3734	return 0;
3735}
3736
3737/*
3738 * IORING_OP_NOP just posts a completion event, nothing else.
3739 */
3740static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3741{
3742	struct io_ring_ctx *ctx = req->ctx;
3743
3744	if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3745		return -EINVAL;
3746
3747	__io_req_complete(req, issue_flags, 0, 0);
3748	return 0;
3749}
3750
3751static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3752{
3753	struct io_ring_ctx *ctx = req->ctx;
3754
3755	if (!req->file)
3756		return -EBADF;
3757
3758	if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3759		return -EINVAL;
3760	if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3761		return -EINVAL;
3762
3763	req->sync.flags = READ_ONCE(sqe->fsync_flags);
3764	if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3765		return -EINVAL;
3766
3767	req->sync.off = READ_ONCE(sqe->off);
3768	req->sync.len = READ_ONCE(sqe->len);
3769	return 0;
3770}
3771
3772static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3773{
3774	loff_t end = req->sync.off + req->sync.len;
3775	int ret;
3776
3777	/* fsync always requires a blocking context */
3778	if (issue_flags & IO_URING_F_NONBLOCK)
3779		return -EAGAIN;
3780
3781	ret = vfs_fsync_range(req->file, req->sync.off,
3782				end > 0 ? end : LLONG_MAX,
3783				req->sync.flags & IORING_FSYNC_DATASYNC);
3784	if (ret < 0)
3785		req_set_fail(req);
3786	io_req_complete(req, ret);
3787	return 0;
3788}
3789
3790static int io_fallocate_prep(struct io_kiocb *req,
3791			     const struct io_uring_sqe *sqe)
3792{
3793	if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3794		return -EINVAL;
3795	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3796		return -EINVAL;
3797
3798	req->sync.off = READ_ONCE(sqe->off);
3799	req->sync.len = READ_ONCE(sqe->addr);
3800	req->sync.mode = READ_ONCE(sqe->len);
3801	return 0;
3802}
3803
3804static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3805{
3806	int ret;
3807
3808	/* fallocate always requiring blocking context */
3809	if (issue_flags & IO_URING_F_NONBLOCK)
3810		return -EAGAIN;
3811	ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3812				req->sync.len);
3813	if (ret < 0)
3814		req_set_fail(req);
3815	io_req_complete(req, ret);
3816	return 0;
3817}
3818
3819static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3820{
3821	const char __user *fname;
3822	int ret;
3823
3824	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3825		return -EINVAL;
3826	if (unlikely(sqe->ioprio || sqe->buf_index))
3827		return -EINVAL;
3828	if (unlikely(req->flags & REQ_F_FIXED_FILE))
3829		return -EBADF;
3830
3831	/* open.how should be already initialised */
3832	if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3833		req->open.how.flags |= O_LARGEFILE;
3834
3835	req->open.dfd = READ_ONCE(sqe->fd);
3836	fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3837	req->open.filename = getname(fname);
3838	if (IS_ERR(req->open.filename)) {
3839		ret = PTR_ERR(req->open.filename);
3840		req->open.filename = NULL;
3841		return ret;
3842	}
3843	req->open.nofile = rlimit(RLIMIT_NOFILE);
3844	req->flags |= REQ_F_NEED_CLEANUP;
3845	return 0;
3846}
3847
3848static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3849{
3850	u64 mode = READ_ONCE(sqe->len);
3851	u64 flags = READ_ONCE(sqe->open_flags);
3852
3853	req->open.how = build_open_how(flags, mode);
3854	return __io_openat_prep(req, sqe);
3855}
3856
3857static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3858{
3859	struct open_how __user *how;
3860	size_t len;
3861	int ret;
3862
3863	how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3864	len = READ_ONCE(sqe->len);
3865	if (len < OPEN_HOW_SIZE_VER0)
3866		return -EINVAL;
3867
3868	ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3869					len);
3870	if (ret)
3871		return ret;
3872
3873	return __io_openat_prep(req, sqe);
3874}
3875
3876static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3877{
3878	struct open_flags op;
3879	struct file *file;
3880	bool nonblock_set;
3881	bool resolve_nonblock;
3882	int ret;
3883
3884	ret = build_open_flags(&req->open.how, &op);
3885	if (ret)
3886		goto err;
3887	nonblock_set = op.open_flag & O_NONBLOCK;
3888	resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3889	if (issue_flags & IO_URING_F_NONBLOCK) {
3890		/*
3891		 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3892		 * it'll always -EAGAIN
3893		 */
3894		if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3895			return -EAGAIN;
3896		op.lookup_flags |= LOOKUP_CACHED;
3897		op.open_flag |= O_NONBLOCK;
3898	}
3899
3900	ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3901	if (ret < 0)
3902		goto err;
3903
3904	file = do_filp_open(req->open.dfd, req->open.filename, &op);
3905	if (IS_ERR(file)) {
3906		/*
3907		 * We could hang on to this 'fd' on retrying, but seems like
3908		 * marginal gain for something that is now known to be a slower
3909		 * path. So just put it, and we'll get a new one when we retry.
3910		 */
3911		put_unused_fd(ret);
3912
3913		ret = PTR_ERR(file);
3914		/* only retry if RESOLVE_CACHED wasn't already set by application */
3915		if (ret == -EAGAIN &&
3916		    (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3917			return -EAGAIN;
3918		goto err;
3919	}
3920
3921	if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3922		file->f_flags &= ~O_NONBLOCK;
3923	fsnotify_open(file);
3924	fd_install(ret, file);
3925err:
3926	putname(req->open.filename);
3927	req->flags &= ~REQ_F_NEED_CLEANUP;
3928	if (ret < 0)
3929		req_set_fail(req);
3930	__io_req_complete(req, issue_flags, ret, 0);
3931	return 0;
3932}
3933
3934static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3935{
3936	return io_openat2(req, issue_flags);
3937}
3938
3939static int io_remove_buffers_prep(struct io_kiocb *req,
3940				  const struct io_uring_sqe *sqe)
3941{
3942	struct io_provide_buf *p = &req->pbuf;
3943	u64 tmp;
3944
3945	if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3946		return -EINVAL;
3947
3948	tmp = READ_ONCE(sqe->fd);
3949	if (!tmp || tmp > USHRT_MAX)
3950		return -EINVAL;
3951
3952	memset(p, 0, sizeof(*p));
3953	p->nbufs = tmp;
3954	p->bgid = READ_ONCE(sqe->buf_group);
3955	return 0;
3956}
3957
3958static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3959			       int bgid, unsigned nbufs)
3960{
3961	unsigned i = 0;
3962
3963	/* shouldn't happen */
3964	if (!nbufs)
3965		return 0;
3966
3967	/* the head kbuf is the list itself */
3968	while (!list_empty(&buf->list)) {
3969		struct io_buffer *nxt;
3970
3971		nxt = list_first_entry(&buf->list, struct io_buffer, list);
3972		list_del(&nxt->list);
3973		kfree(nxt);
3974		if (++i == nbufs)
3975			return i;
3976	}
3977	i++;
3978	kfree(buf);
3979	xa_erase(&ctx->io_buffers, bgid);
3980
3981	return i;
3982}
3983
3984static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3985{
3986	struct io_provide_buf *p = &req->pbuf;
3987	struct io_ring_ctx *ctx = req->ctx;
3988	struct io_buffer *head;
3989	int ret = 0;
3990	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3991
3992	io_ring_submit_lock(ctx, !force_nonblock);
3993
3994	lockdep_assert_held(&ctx->uring_lock);
3995
3996	ret = -ENOENT;
3997	head = xa_load(&ctx->io_buffers, p->bgid);
3998	if (head)
3999		ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4000	if (ret < 0)
4001		req_set_fail(req);
4002
4003	/* complete before unlock, IOPOLL may need the lock */
4004	__io_req_complete(req, issue_flags, ret, 0);
4005	io_ring_submit_unlock(ctx, !force_nonblock);
4006	return 0;
4007}
4008
4009static int io_provide_buffers_prep(struct io_kiocb *req,
4010				   const struct io_uring_sqe *sqe)
4011{
4012	unsigned long size, tmp_check;
4013	struct io_provide_buf *p = &req->pbuf;
4014	u64 tmp;
4015
4016	if (sqe->ioprio || sqe->rw_flags)
4017		return -EINVAL;
4018
4019	tmp = READ_ONCE(sqe->fd);
4020	if (!tmp || tmp > USHRT_MAX)
4021		return -E2BIG;
4022	p->nbufs = tmp;
4023	p->addr = READ_ONCE(sqe->addr);
4024	p->len = READ_ONCE(sqe->len);
4025
4026	if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4027				&size))
4028		return -EOVERFLOW;
4029	if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4030		return -EOVERFLOW;
4031
4032	size = (unsigned long)p->len * p->nbufs;
4033	if (!access_ok(u64_to_user_ptr(p->addr), size))
4034		return -EFAULT;
4035
4036	p->bgid = READ_ONCE(sqe->buf_group);
4037	tmp = READ_ONCE(sqe->off);
4038	if (tmp > USHRT_MAX)
4039		return -E2BIG;
4040	p->bid = tmp;
4041	return 0;
4042}
4043
4044static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4045{
4046	struct io_buffer *buf;
4047	u64 addr = pbuf->addr;
4048	int i, bid = pbuf->bid;
4049
4050	for (i = 0; i < pbuf->nbufs; i++) {
4051		buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4052		if (!buf)
4053			break;
4054
4055		buf->addr = addr;
4056		buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4057		buf->bid = bid;
4058		addr += pbuf->len;
4059		bid++;
4060		if (!*head) {
4061			INIT_LIST_HEAD(&buf->list);
4062			*head = buf;
4063		} else {
4064			list_add_tail(&buf->list, &(*head)->list);
4065		}
4066	}
4067
4068	return i ? i : -ENOMEM;
4069}
4070
4071static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4072{
4073	struct io_provide_buf *p = &req->pbuf;
4074	struct io_ring_ctx *ctx = req->ctx;
4075	struct io_buffer *head, *list;
4076	int ret = 0;
4077	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4078
4079	io_ring_submit_lock(ctx, !force_nonblock);
4080
4081	lockdep_assert_held(&ctx->uring_lock);
4082
4083	list = head = xa_load(&ctx->io_buffers, p->bgid);
4084
4085	ret = io_add_buffers(p, &head);
4086	if (ret >= 0 && !list) {
4087		ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4088		if (ret < 0)
4089			__io_remove_buffers(ctx, head, p->bgid, -1U);
4090	}
4091	if (ret < 0)
4092		req_set_fail(req);
4093	/* complete before unlock, IOPOLL may need the lock */
4094	__io_req_complete(req, issue_flags, ret, 0);
4095	io_ring_submit_unlock(ctx, !force_nonblock);
4096	return 0;
4097}
4098
4099static int io_epoll_ctl_prep(struct io_kiocb *req,
4100			     const struct io_uring_sqe *sqe)
4101{
4102#if defined(CONFIG_EPOLL)
4103	if (sqe->ioprio || sqe->buf_index)
4104		return -EINVAL;
4105	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4106		return -EINVAL;
4107
4108	req->epoll.epfd = READ_ONCE(sqe->fd);
4109	req->epoll.op = READ_ONCE(sqe->len);
4110	req->epoll.fd = READ_ONCE(sqe->off);
4111
4112	if (ep_op_has_event(req->epoll.op)) {
4113		struct epoll_event __user *ev;
4114
4115		ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4116		if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4117			return -EFAULT;
4118	}
4119
4120	return 0;
4121#else
4122	return -EOPNOTSUPP;
4123#endif
4124}
4125
4126static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4127{
4128#if defined(CONFIG_EPOLL)
4129	struct io_epoll *ie = &req->epoll;
4130	int ret;
4131	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4132
4133	ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4134	if (force_nonblock && ret == -EAGAIN)
4135		return -EAGAIN;
4136
4137	if (ret < 0)
4138		req_set_fail(req);
4139	__io_req_complete(req, issue_flags, ret, 0);
4140	return 0;
4141#else
4142	return -EOPNOTSUPP;
4143#endif
4144}
4145
4146static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4147{
4148#if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4149	if (sqe->ioprio || sqe->buf_index || sqe->off)
4150		return -EINVAL;
4151	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4152		return -EINVAL;
4153
4154	req->madvise.addr = READ_ONCE(sqe->addr);
4155	req->madvise.len = READ_ONCE(sqe->len);
4156	req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4157	return 0;
4158#else
4159	return -EOPNOTSUPP;
4160#endif
4161}
4162
4163static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4164{
4165#if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4166	struct io_madvise *ma = &req->madvise;
4167	int ret;
4168
4169	if (issue_flags & IO_URING_F_NONBLOCK)
4170		return -EAGAIN;
4171
4172	ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4173	if (ret < 0)
4174		req_set_fail(req);
4175	io_req_complete(req, ret);
4176	return 0;
4177#else
4178	return -EOPNOTSUPP;
4179#endif
4180}
4181
4182static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4183{
4184	if (sqe->ioprio || sqe->buf_index || sqe->addr)
4185		return -EINVAL;
4186	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4187		return -EINVAL;
4188
4189	req->fadvise.offset = READ_ONCE(sqe->off);
4190	req->fadvise.len = READ_ONCE(sqe->len);
4191	req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4192	return 0;
4193}
4194
4195static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4196{
4197	struct io_fadvise *fa = &req->fadvise;
4198	int ret;
4199
4200	if (issue_flags & IO_URING_F_NONBLOCK) {
4201		switch (fa->advice) {
4202		case POSIX_FADV_NORMAL:
4203		case POSIX_FADV_RANDOM:
4204		case POSIX_FADV_SEQUENTIAL:
4205			break;
4206		default:
4207			return -EAGAIN;
4208		}
4209	}
4210
4211	ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4212	if (ret < 0)
4213		req_set_fail(req);
4214	__io_req_complete(req, issue_flags, ret, 0);
4215	return 0;
4216}
4217
4218static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4219{
4220	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4221		return -EINVAL;
4222	if (sqe->ioprio || sqe->buf_index)
4223		return -EINVAL;
4224	if (req->flags & REQ_F_FIXED_FILE)
4225		return -EBADF;
4226
4227	req->statx.dfd = READ_ONCE(sqe->fd);
4228	req->statx.mask = READ_ONCE(sqe->len);
4229	req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4230	req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4231	req->statx.flags = READ_ONCE(sqe->statx_flags);
4232
4233	return 0;
4234}
4235
4236static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4237{
4238	struct io_statx *ctx = &req->statx;
4239	int ret;
4240
4241	if (issue_flags & IO_URING_F_NONBLOCK)
4242		return -EAGAIN;
4243
4244	ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4245		       ctx->buffer);
4246
4247	if (ret < 0)
4248		req_set_fail(req);
4249	io_req_complete(req, ret);
4250	return 0;
4251}
4252
4253static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4254{
4255	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4256		return -EINVAL;
4257	if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4258	    sqe->rw_flags || sqe->buf_index)
4259		return -EINVAL;
4260	if (req->flags & REQ_F_FIXED_FILE)
4261		return -EBADF;
4262
4263	req->close.fd = READ_ONCE(sqe->fd);
4264	return 0;
4265}
4266
4267static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4268{
4269	struct files_struct *files = current->files;
4270	struct io_close *close = &req->close;
4271	struct fdtable *fdt;
4272	struct file *file = NULL;
4273	int ret = -EBADF;
4274
4275	spin_lock(&files->file_lock);
4276	fdt = files_fdtable(files);
4277	if (close->fd >= fdt->max_fds) {
4278		spin_unlock(&files->file_lock);
4279		goto err;
4280	}
4281	file = fdt->fd[close->fd];
4282	if (!file || file->f_op == &io_uring_fops) {
4283		spin_unlock(&files->file_lock);
4284		file = NULL;
4285		goto err;
4286	}
4287
4288	/* if the file has a flush method, be safe and punt to async */
4289	if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4290		spin_unlock(&files->file_lock);
4291		return -EAGAIN;
4292	}
4293
4294	ret = __close_fd_get_file(close->fd, &file);
4295	spin_unlock(&files->file_lock);
4296	if (ret < 0) {
4297		if (ret == -ENOENT)
4298			ret = -EBADF;
4299		goto err;
4300	}
4301
4302	/* No ->flush() or already async, safely close from here */
4303	ret = filp_close(file, current->files);
4304err:
4305	if (ret < 0)
4306		req_set_fail(req);
4307	if (file)
4308		fput(file);
4309	__io_req_complete(req, issue_flags, ret, 0);
4310	return 0;
4311}
4312
4313static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4314{
4315	struct io_ring_ctx *ctx = req->ctx;
4316
4317	if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4318		return -EINVAL;
4319	if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4320		return -EINVAL;
4321
4322	req->sync.off = READ_ONCE(sqe->off);
4323	req->sync.len = READ_ONCE(sqe->len);
4324	req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4325	return 0;
4326}
4327
4328static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4329{
4330	int ret;
4331
4332	/* sync_file_range always requires a blocking context */
4333	if (issue_flags & IO_URING_F_NONBLOCK)
4334		return -EAGAIN;
4335
4336	ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4337				req->sync.flags);
4338	if (ret < 0)
4339		req_set_fail(req);
4340	io_req_complete(req, ret);
4341	return 0;
4342}
4343
4344#if defined(CONFIG_NET)
4345static int io_setup_async_msg(struct io_kiocb *req,
4346			      struct io_async_msghdr *kmsg)
4347{
4348	struct io_async_msghdr *async_msg = req->async_data;
4349
4350	if (async_msg)
4351		return -EAGAIN;
4352	if (io_alloc_async_data(req)) {
4353		kfree(kmsg->free_iov);
4354		return -ENOMEM;
4355	}
4356	async_msg = req->async_data;
4357	req->flags |= REQ_F_NEED_CLEANUP;
4358	memcpy(async_msg, kmsg, sizeof(*kmsg));
4359	async_msg->msg.msg_name = &async_msg->addr;
4360	/* if were using fast_iov, set it to the new one */
4361	if (!async_msg->free_iov)
4362		async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4363
4364	return -EAGAIN;
4365}
4366
4367static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4368			       struct io_async_msghdr *iomsg)
4369{
4370	iomsg->msg.msg_name = &iomsg->addr;
4371	iomsg->free_iov = iomsg->fast_iov;
4372	return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4373				   req->sr_msg.msg_flags, &iomsg->free_iov);
4374}
4375
4376static int io_sendmsg_prep_async(struct io_kiocb *req)
4377{
4378	int ret;
4379
4380	ret = io_sendmsg_copy_hdr(req, req->async_data);
4381	if (!ret)
4382		req->flags |= REQ_F_NEED_CLEANUP;
4383	return ret;
4384}
4385
4386static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4387{
4388	struct io_sr_msg *sr = &req->sr_msg;
4389
4390	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4391		return -EINVAL;
4392
4393	sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4394	sr->len = READ_ONCE(sqe->len);
4395	sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4396	if (sr->msg_flags & MSG_DONTWAIT)
4397		req->flags |= REQ_F_NOWAIT;
4398
4399#ifdef CONFIG_COMPAT
4400	if (req->ctx->compat)
4401		sr->msg_flags |= MSG_CMSG_COMPAT;
4402#endif
4403	return 0;
4404}
4405
4406static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4407{
4408	struct io_async_msghdr iomsg, *kmsg;
4409	struct socket *sock;
4410	unsigned flags;
4411	int min_ret = 0;
4412	int ret;
4413
4414	sock = sock_from_file(req->file);
4415	if (unlikely(!sock))
4416		return -ENOTSOCK;
4417
4418	kmsg = req->async_data;
4419	if (!kmsg) {
4420		ret = io_sendmsg_copy_hdr(req, &iomsg);
4421		if (ret)
4422			return ret;
4423		kmsg = &iomsg;
4424	}
4425
4426	flags = req->sr_msg.msg_flags;
4427	if (issue_flags & IO_URING_F_NONBLOCK)
4428		flags |= MSG_DONTWAIT;
4429	if (flags & MSG_WAITALL)
4430		min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4431
4432	ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4433	if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4434		return io_setup_async_msg(req, kmsg);
4435	if (ret == -ERESTARTSYS)
4436		ret = -EINTR;
4437
4438	/* fast path, check for non-NULL to avoid function call */
4439	if (kmsg->free_iov)
4440		kfree(kmsg->free_iov);
4441	req->flags &= ~REQ_F_NEED_CLEANUP;
4442	if (ret < min_ret)
4443		req_set_fail(req);
4444	__io_req_complete(req, issue_flags, ret, 0);
4445	return 0;
4446}
4447
4448static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4449{
4450	struct io_sr_msg *sr = &req->sr_msg;
4451	struct msghdr msg;
4452	struct iovec iov;
4453	struct socket *sock;
4454	unsigned flags;
4455	int min_ret = 0;
4456	int ret;
4457
4458	sock = sock_from_file(req->file);
4459	if (unlikely(!sock))
4460		return -ENOTSOCK;
4461
4462	ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4463	if (unlikely(ret))
4464		return ret;
4465
4466	msg.msg_name = NULL;
4467	msg.msg_control = NULL;
4468	msg.msg_controllen = 0;
4469	msg.msg_namelen = 0;
4470
4471	flags = req->sr_msg.msg_flags;
4472	if (issue_flags & IO_URING_F_NONBLOCK)
4473		flags |= MSG_DONTWAIT;
4474	if (flags & MSG_WAITALL)
4475		min_ret = iov_iter_count(&msg.msg_iter);
4476
4477	msg.msg_flags = flags;
4478	ret = sock_sendmsg(sock, &msg);
4479	if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4480		return -EAGAIN;
4481	if (ret == -ERESTARTSYS)
4482		ret = -EINTR;
4483
4484	if (ret < min_ret)
4485		req_set_fail(req);
4486	__io_req_complete(req, issue_flags, ret, 0);
4487	return 0;
4488}
4489
4490static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4491				 struct io_async_msghdr *iomsg)
4492{
4493	struct io_sr_msg *sr = &req->sr_msg;
4494	struct iovec __user *uiov;
4495	size_t iov_len;
4496	int ret;
4497
4498	ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4499					&iomsg->uaddr, &uiov, &iov_len);
4500	if (ret)
4501		return ret;
4502
4503	if (req->flags & REQ_F_BUFFER_SELECT) {
4504		if (iov_len > 1)
4505			return -EINVAL;
4506		if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4507			return -EFAULT;
4508		sr->len = iomsg->fast_iov[0].iov_len;
4509		iomsg->free_iov = NULL;
4510	} else {
4511		iomsg->free_iov = iomsg->fast_iov;
4512		ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4513				     &iomsg->free_iov, &iomsg->msg.msg_iter,
4514				     false);
4515		if (ret > 0)
4516			ret = 0;
4517	}
4518
4519	return ret;
4520}
4521
4522#ifdef CONFIG_COMPAT
4523static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4524					struct io_async_msghdr *iomsg)
4525{
4526	struct io_sr_msg *sr = &req->sr_msg;
4527	struct compat_iovec __user *uiov;
4528	compat_uptr_t ptr;
4529	compat_size_t len;
4530	int ret;
4531
4532	ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4533				  &ptr, &len);
4534	if (ret)
4535		return ret;
4536
4537	uiov = compat_ptr(ptr);
4538	if (req->flags & REQ_F_BUFFER_SELECT) {
4539		compat_ssize_t clen;
4540
4541		if (len > 1)
4542			return -EINVAL;
4543		if (!access_ok(uiov, sizeof(*uiov)))
4544			return -EFAULT;
4545		if (__get_user(clen, &uiov->iov_len))
4546			return -EFAULT;
4547		if (clen < 0)
4548			return -EINVAL;
4549		sr->len = clen;
4550		iomsg->free_iov = NULL;
4551	} else {
4552		iomsg->free_iov = iomsg->fast_iov;
4553		ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4554				   UIO_FASTIOV, &iomsg->free_iov,
4555				   &iomsg->msg.msg_iter, true);
4556		if (ret < 0)
4557			return ret;
4558	}
4559
4560	return 0;
4561}
4562#endif
4563
4564static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4565			       struct io_async_msghdr *iomsg)
4566{
4567	iomsg->msg.msg_name = &iomsg->addr;
4568
4569#ifdef CONFIG_COMPAT
4570	if (req->ctx->compat)
4571		return __io_compat_recvmsg_copy_hdr(req, iomsg);
4572#endif
4573
4574	return __io_recvmsg_copy_hdr(req, iomsg);
4575}
4576
4577static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4578					       bool needs_lock)
4579{
4580	struct io_sr_msg *sr = &req->sr_msg;
4581	struct io_buffer *kbuf;
4582
4583	kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4584	if (IS_ERR(kbuf))
4585		return kbuf;
4586
4587	sr->kbuf = kbuf;
4588	req->flags |= REQ_F_BUFFER_SELECTED;
4589	return kbuf;
4590}
4591
4592static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4593{
4594	return io_put_kbuf(req, req->sr_msg.kbuf);
4595}
4596
4597static int io_recvmsg_prep_async(struct io_kiocb *req)
4598{
4599	int ret;
4600
4601	ret = io_recvmsg_copy_hdr(req, req->async_data);
4602	if (!ret)
4603		req->flags |= REQ_F_NEED_CLEANUP;
4604	return ret;
4605}
4606
4607static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4608{
4609	struct io_sr_msg *sr = &req->sr_msg;
4610
4611	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4612		return -EINVAL;
4613
4614	sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4615	sr->len = READ_ONCE(sqe->len);
4616	sr->bgid = READ_ONCE(sqe->buf_group);
4617	sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4618	if (sr->msg_flags & MSG_DONTWAIT)
4619		req->flags |= REQ_F_NOWAIT;
4620
4621#ifdef CONFIG_COMPAT
4622	if (req->ctx->compat)
4623		sr->msg_flags |= MSG_CMSG_COMPAT;
4624#endif
4625	return 0;
4626}
4627
4628static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4629{
4630	struct io_async_msghdr iomsg, *kmsg;
4631	struct socket *sock;
4632	struct io_buffer *kbuf;
4633	unsigned flags;
4634	int min_ret = 0;
4635	int ret, cflags = 0;
4636	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4637
4638	sock = sock_from_file(req->file);
4639	if (unlikely(!sock))
4640		return -ENOTSOCK;
4641
4642	kmsg = req->async_data;
4643	if (!kmsg) {
4644		ret = io_recvmsg_copy_hdr(req, &iomsg);
4645		if (ret)
4646			return ret;
4647		kmsg = &iomsg;
4648	}
4649
4650	if (req->flags & REQ_F_BUFFER_SELECT) {
4651		kbuf = io_recv_buffer_select(req, !force_nonblock);
4652		if (IS_ERR(kbuf))
4653			return PTR_ERR(kbuf);
4654		kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4655		kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4656		iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4657				1, req->sr_msg.len);
4658	}
4659
4660	flags = req->sr_msg.msg_flags;
4661	if (force_nonblock)
4662		flags |= MSG_DONTWAIT;
4663	if (flags & MSG_WAITALL)
4664		min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4665
4666	ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4667					kmsg->uaddr, flags);
4668	if (force_nonblock && ret == -EAGAIN)
4669		return io_setup_async_msg(req, kmsg);
4670	if (ret == -ERESTARTSYS)
4671		ret = -EINTR;
4672
4673	if (req->flags & REQ_F_BUFFER_SELECTED)
4674		cflags = io_put_recv_kbuf(req);
4675	/* fast path, check for non-NULL to avoid function call */
4676	if (kmsg->free_iov)
4677		kfree(kmsg->free_iov);
4678	req->flags &= ~REQ_F_NEED_CLEANUP;
4679	if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4680		req_set_fail(req);
4681	__io_req_complete(req, issue_flags, ret, cflags);
4682	return 0;
4683}
4684
4685static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4686{
4687	struct io_buffer *kbuf;
4688	struct io_sr_msg *sr = &req->sr_msg;
4689	struct msghdr msg;
4690	void __user *buf = sr->buf;
4691	struct socket *sock;
4692	struct iovec iov;
4693	unsigned flags;
4694	int min_ret = 0;
4695	int ret, cflags = 0;
4696	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4697
4698	sock = sock_from_file(req->file);
4699	if (unlikely(!sock))
4700		return -ENOTSOCK;
4701
4702	if (req->flags & REQ_F_BUFFER_SELECT) {
4703		kbuf = io_recv_buffer_select(req, !force_nonblock);
4704		if (IS_ERR(kbuf))
4705			return PTR_ERR(kbuf);
4706		buf = u64_to_user_ptr(kbuf->addr);
4707	}
4708
4709	ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4710	if (unlikely(ret))
4711		goto out_free;
4712
4713	msg.msg_name = NULL;
4714	msg.msg_control = NULL;
4715	msg.msg_controllen = 0;
4716	msg.msg_namelen = 0;
4717	msg.msg_iocb = NULL;
4718	msg.msg_flags = 0;
4719
4720	flags = req->sr_msg.msg_flags;
4721	if (force_nonblock)
4722		flags |= MSG_DONTWAIT;
4723	if (flags & MSG_WAITALL)
4724		min_ret = iov_iter_count(&msg.msg_iter);
4725
4726	ret = sock_recvmsg(sock, &msg, flags);
4727	if (force_nonblock && ret == -EAGAIN)
4728		return -EAGAIN;
4729	if (ret == -ERESTARTSYS)
4730		ret = -EINTR;
4731out_free:
4732	if (req->flags & REQ_F_BUFFER_SELECTED)
4733		cflags = io_put_recv_kbuf(req);
4734	if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4735		req_set_fail(req);
4736	__io_req_complete(req, issue_flags, ret, cflags);
4737	return 0;
4738}
4739
4740static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4741{
4742	struct io_accept *accept = &req->accept;
4743
4744	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4745		return -EINVAL;
4746	if (sqe->ioprio || sqe->len || sqe->buf_index)
4747		return -EINVAL;
4748
4749	accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4750	accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4751	accept->flags = READ_ONCE(sqe->accept_flags);
4752	accept->nofile = rlimit(RLIMIT_NOFILE);
4753	return 0;
4754}
4755
4756static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4757{
4758	struct io_accept *accept = &req->accept;
4759	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4760	unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4761	int ret;
4762
4763	if (req->file->f_flags & O_NONBLOCK)
4764		req->flags |= REQ_F_NOWAIT;
4765
4766	ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4767					accept->addr_len, accept->flags,
4768					accept->nofile);
4769	if (ret == -EAGAIN && force_nonblock)
4770		return -EAGAIN;
4771	if (ret < 0) {
4772		if (ret == -ERESTARTSYS)
4773			ret = -EINTR;
4774		req_set_fail(req);
4775	}
4776	__io_req_complete(req, issue_flags, ret, 0);
4777	return 0;
4778}
4779
4780static int io_connect_prep_async(struct io_kiocb *req)
4781{
4782	struct io_async_connect *io = req->async_data;
4783	struct io_connect *conn = &req->connect;
4784
4785	return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4786}
4787
4788static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4789{
4790	struct io_connect *conn = &req->connect;
4791
4792	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4793		return -EINVAL;
4794	if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4795		return -EINVAL;
4796
4797	conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4798	conn->addr_len =  READ_ONCE(sqe->addr2);
4799	return 0;
4800}
4801
4802static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4803{
4804	struct io_async_connect __io, *io;
4805	unsigned file_flags;
4806	int ret;
4807	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4808
4809	if (req->async_data) {
4810		io = req->async_data;
4811	} else {
4812		ret = move_addr_to_kernel(req->connect.addr,
4813						req->connect.addr_len,
4814						&__io.address);
4815		if (ret)
4816			goto out;
4817		io = &__io;
4818	}
4819
4820	file_flags = force_nonblock ? O_NONBLOCK : 0;
4821
4822	ret = __sys_connect_file(req->file, &io->address,
4823					req->connect.addr_len, file_flags);
4824	if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4825		if (req->async_data)
4826			return -EAGAIN;
4827		if (io_alloc_async_data(req)) {
4828			ret = -ENOMEM;
4829			goto out;
4830		}
4831		memcpy(req->async_data, &__io, sizeof(__io));
4832		return -EAGAIN;
4833	}
4834	if (ret == -ERESTARTSYS)
4835		ret = -EINTR;
4836out:
4837	if (ret < 0)
4838		req_set_fail(req);
4839	__io_req_complete(req, issue_flags, ret, 0);
4840	return 0;
4841}
4842#else /* !CONFIG_NET */
4843#define IO_NETOP_FN(op)							\
4844static int io_##op(struct io_kiocb *req, unsigned int issue_flags)	\
4845{									\
4846	return -EOPNOTSUPP;						\
4847}
4848
4849#define IO_NETOP_PREP(op)						\
4850IO_NETOP_FN(op)								\
4851static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4852{									\
4853	return -EOPNOTSUPP;						\
4854}									\
4855
4856#define IO_NETOP_PREP_ASYNC(op)						\
4857IO_NETOP_PREP(op)							\
4858static int io_##op##_prep_async(struct io_kiocb *req)			\
4859{									\
4860	return -EOPNOTSUPP;						\
4861}
4862
4863IO_NETOP_PREP_ASYNC(sendmsg);
4864IO_NETOP_PREP_ASYNC(recvmsg);
4865IO_NETOP_PREP_ASYNC(connect);
4866IO_NETOP_PREP(accept);
4867IO_NETOP_FN(send);
4868IO_NETOP_FN(recv);
4869#endif /* CONFIG_NET */
4870
4871struct io_poll_table {
4872	struct poll_table_struct pt;
4873	struct io_kiocb *req;
4874	int nr_entries;
4875	int error;
4876};
4877
4878static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4879			   __poll_t mask, io_req_tw_func_t func)
4880{
4881	/* for instances that support it check for an event match first: */
4882	if (mask && !(mask & poll->events))
4883		return 0;
4884
4885	trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4886
4887	list_del_init(&poll->wait.entry);
4888
4889	req->result = mask;
4890	req->io_task_work.func = func;
4891
4892	/*
4893	 * If this fails, then the task is exiting. When a task exits, the
4894	 * work gets canceled, so just cancel this request as well instead
4895	 * of executing it. We can't safely execute it anyway, as we may not
4896	 * have the needed state needed for it anyway.
4897	 */
4898	io_req_task_work_add(req);
4899	return 1;
4900}
4901
4902static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4903	__acquires(&req->ctx->completion_lock)
4904{
4905	struct io_ring_ctx *ctx = req->ctx;
4906
4907	/* req->task == current here, checking PF_EXITING is safe */
4908	if (unlikely(req->task->flags & PF_EXITING))
4909		WRITE_ONCE(poll->canceled, true);
4910
4911	if (!req->result && !READ_ONCE(poll->canceled)) {
4912		struct poll_table_struct pt = { ._key = poll->events };
4913
4914		req->result = vfs_poll(req->file, &pt) & poll->events;
4915	}
4916
4917	spin_lock(&ctx->completion_lock);
4918	if (!req->result && !READ_ONCE(poll->canceled)) {
4919		add_wait_queue(poll->head, &poll->wait);
4920		return true;
4921	}
4922
4923	return false;
4924}
4925
4926static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4927{
4928	/* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4929	if (req->opcode == IORING_OP_POLL_ADD)
4930		return req->async_data;
4931	return req->apoll->double_poll;
4932}
4933
4934static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4935{
4936	if (req->opcode == IORING_OP_POLL_ADD)
4937		return &req->poll;
4938	return &req->apoll->poll;
4939}
4940
4941static void io_poll_remove_double(struct io_kiocb *req)
4942	__must_hold(&req->ctx->completion_lock)
4943{
4944	struct io_poll_iocb *poll = io_poll_get_double(req);
4945
4946	lockdep_assert_held(&req->ctx->completion_lock);
4947
4948	if (poll && poll->head) {
4949		struct wait_queue_head *head = poll->head;
4950
4951		spin_lock_irq(&head->lock);
4952		list_del_init(&poll->wait.entry);
4953		if (poll->wait.private)
4954			req_ref_put(req);
4955		poll->head = NULL;
4956		spin_unlock_irq(&head->lock);
4957	}
4958}
4959
4960static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4961	__must_hold(&req->ctx->completion_lock)
4962{
4963	struct io_ring_ctx *ctx = req->ctx;
4964	unsigned flags = IORING_CQE_F_MORE;
4965	int error;
4966
4967	if (READ_ONCE(req->poll.canceled)) {
4968		error = -ECANCELED;
4969		req->poll.events |= EPOLLONESHOT;
4970	} else {
4971		error = mangle_poll(mask);
4972	}
4973	if (req->poll.events & EPOLLONESHOT)
4974		flags = 0;
4975	if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4976		req->poll.done = true;
4977		flags = 0;
4978	}
4979	if (flags & IORING_CQE_F_MORE)
4980		ctx->cq_extra++;
4981
4982	io_commit_cqring(ctx);
4983	return !(flags & IORING_CQE_F_MORE);
4984}
4985
4986static void io_poll_task_func(struct io_kiocb *req)
4987{
4988	struct io_ring_ctx *ctx = req->ctx;
4989	struct io_kiocb *nxt;
4990
4991	if (io_poll_rewait(req, &req->poll)) {
4992		spin_unlock(&ctx->completion_lock);
4993	} else {
4994		bool done;
4995
4996		done = io_poll_complete(req, req->result);
4997		if (done) {
4998			io_poll_remove_double(req);
4999			hash_del(&req->hash_node);
5000		} else {
5001			req->result = 0;
5002			add_wait_queue(req->poll.head, &req->poll.wait);
5003		}
5004		spin_unlock(&ctx->completion_lock);
5005		io_cqring_ev_posted(ctx);
5006
5007		if (done) {
5008			nxt = io_put_req_find_next(req);
5009			if (nxt)
5010				io_req_task_submit(nxt);
5011		}
5012	}
5013}
5014
5015static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5016			       int sync, void *key)
5017{
5018	struct io_kiocb *req = wait->private;
5019	struct io_poll_iocb *poll = io_poll_get_single(req);
5020	__poll_t mask = key_to_poll(key);
5021	unsigned long flags;
5022
5023	/* for instances that support it check for an event match first: */
5024	if (mask && !(mask & poll->events))
5025		return 0;
5026	if (!(poll->events & EPOLLONESHOT))
5027		return poll->wait.func(&poll->wait, mode, sync, key);
5028
5029	list_del_init(&wait->entry);
5030
5031	if (poll->head) {
5032		bool done;
5033
5034		spin_lock_irqsave(&poll->head->lock, flags);
5035		done = list_empty(&poll->wait.entry);
5036		if (!done)
5037			list_del_init(&poll->wait.entry);
5038		/* make sure double remove sees this as being gone */
5039		wait->private = NULL;
5040		spin_unlock_irqrestore(&poll->head->lock, flags);
5041		if (!done) {
5042			/* use wait func handler, so it matches the rq type */
5043			poll->wait.func(&poll->wait, mode, sync, key);
5044		}
5045	}
5046	req_ref_put(req);
5047	return 1;
5048}
5049
5050static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5051			      wait_queue_func_t wake_func)
5052{
5053	poll->head = NULL;
5054	poll->done = false;
5055	poll->canceled = false;
5056#define IO_POLL_UNMASK	(EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5057	/* mask in events that we always want/need */
5058	poll->events = events | IO_POLL_UNMASK;
5059	INIT_LIST_HEAD(&poll->wait.entry);
5060	init_waitqueue_func_entry(&poll->wait, wake_func);
5061}
5062
5063static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5064			    struct wait_queue_head *head,
5065			    struct io_poll_iocb **poll_ptr)
5066{
5067	struct io_kiocb *req = pt->req;
5068
5069	/*
5070	 * The file being polled uses multiple waitqueues for poll handling
5071	 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5072	 * if this happens.
5073	 */
5074	if (unlikely(pt->nr_entries)) {
5075		struct io_poll_iocb *poll_one = poll;
5076
5077		/* double add on the same waitqueue head, ignore */
5078		if (poll_one->head == head)
5079			return;
5080		/* already have a 2nd entry, fail a third attempt */
5081		if (*poll_ptr) {
5082			if ((*poll_ptr)->head == head)
5083				return;
5084			pt->error = -EINVAL;
5085			return;
5086		}
5087		/*
5088		 * Can't handle multishot for double wait for now, turn it
5089		 * into one-shot mode.
5090		 */
5091		if (!(poll_one->events & EPOLLONESHOT))
5092			poll_one->events |= EPOLLONESHOT;
5093		poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5094		if (!poll) {
5095			pt->error = -ENOMEM;
5096			return;
5097		}
5098		io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5099		req_ref_get(req);
5100		poll->wait.private = req;
5101		*poll_ptr = poll;
5102	}
5103
5104	pt->nr_entries++;
5105	poll->head = head;
5106
5107	if (poll->events & EPOLLEXCLUSIVE)
5108		add_wait_queue_exclusive(head, &poll->wait);
5109	else
5110		add_wait_queue(head, &poll->wait);
5111}
5112
5113static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5114			       struct poll_table_struct *p)
5115{
5116	struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5117	struct async_poll *apoll = pt->req->apoll;
5118
5119	__io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5120}
5121
5122static void io_async_task_func(struct io_kiocb *req)
5123{
5124	struct async_poll *apoll = req->apoll;
5125	struct io_ring_ctx *ctx = req->ctx;
5126
5127	trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5128
5129	if (io_poll_rewait(req, &apoll->poll)) {
5130		spin_unlock(&ctx->completion_lock);
5131		return;
5132	}
5133
5134	hash_del(&req->hash_node);
5135	io_poll_remove_double(req);
5136	spin_unlock(&ctx->completion_lock);
5137
5138	if (!READ_ONCE(apoll->poll.canceled))
5139		io_req_task_submit(req);
5140	else
5141		io_req_complete_failed(req, -ECANCELED);
5142}
5143
5144static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5145			void *key)
5146{
5147	struct io_kiocb *req = wait->private;
5148	struct io_poll_iocb *poll = &req->apoll->poll;
5149
5150	trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5151					key_to_poll(key));
5152
5153	return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5154}
5155
5156static void io_poll_req_insert(struct io_kiocb *req)
5157{
5158	struct io_ring_ctx *ctx = req->ctx;
5159	struct hlist_head *list;
5160
5161	list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5162	hlist_add_head(&req->hash_node, list);
5163}
5164
5165static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5166				      struct io_poll_iocb *poll,
5167				      struct io_poll_table *ipt, __poll_t mask,
5168				      wait_queue_func_t wake_func)
5169	__acquires(&ctx->completion_lock)
5170{
5171	struct io_ring_ctx *ctx = req->ctx;
5172	bool cancel = false;
5173
5174	INIT_HLIST_NODE(&req->hash_node);
5175	io_init_poll_iocb(poll, mask, wake_func);
5176	poll->file = req->file;
5177	poll->wait.private = req;
5178
5179	ipt->pt._key = mask;
5180	ipt->req = req;
5181	ipt->error = 0;
5182	ipt->nr_entries = 0;
5183
5184	mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5185	if (unlikely(!ipt->nr_entries) && !ipt->error)
5186		ipt->error = -EINVAL;
5187
5188	spin_lock(&ctx->completion_lock);
5189	if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5190		io_poll_remove_double(req);
5191	if (likely(poll->head)) {
5192		spin_lock_irq(&poll->head->lock);
5193		if (unlikely(list_empty(&poll->wait.entry))) {
5194			if (ipt->error)
5195				cancel = true;
5196			ipt->error = 0;
5197			mask = 0;
5198		}
5199		if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5200			list_del_init(&poll->wait.entry);
5201		else if (cancel)
5202			WRITE_ONCE(poll->canceled, true);
5203		else if (!poll->done) /* actually waiting for an event */
5204			io_poll_req_insert(req);
5205		spin_unlock_irq(&poll->head->lock);
5206	}
5207
5208	return mask;
5209}
5210
5211enum {
5212	IO_APOLL_OK,
5213	IO_APOLL_ABORTED,
5214	IO_APOLL_READY
5215};
5216
5217static int io_arm_poll_handler(struct io_kiocb *req)
5218{
5219	const struct io_op_def *def = &io_op_defs[req->opcode];
5220	struct io_ring_ctx *ctx = req->ctx;
5221	struct async_poll *apoll;
5222	struct io_poll_table ipt;
5223	__poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5224	int rw;
5225
5226	if (!req->file || !file_can_poll(req->file))
5227		return IO_APOLL_ABORTED;
5228	if (req->flags & REQ_F_POLLED)
5229		return IO_APOLL_ABORTED;
5230	if (!def->pollin && !def->pollout)
5231		return IO_APOLL_ABORTED;
5232
5233	if (def->pollin) {
5234		rw = READ;
5235		mask |= POLLIN | POLLRDNORM;
5236
5237		/* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5238		if ((req->opcode == IORING_OP_RECVMSG) &&
5239		    (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5240			mask &= ~POLLIN;
5241	} else {
5242		rw = WRITE;
5243		mask |= POLLOUT | POLLWRNORM;
5244	}
5245
5246	/* if we can't nonblock try, then no point in arming a poll handler */
5247	if (!io_file_supports_nowait(req, rw))
5248		return IO_APOLL_ABORTED;
5249
5250	apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5251	if (unlikely(!apoll))
5252		return IO_APOLL_ABORTED;
5253	apoll->double_poll = NULL;
5254	req->apoll = apoll;
5255	req->flags |= REQ_F_POLLED;
5256	ipt.pt._qproc = io_async_queue_proc;
5257	io_req_set_refcount(req);
5258
5259	ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5260					io_async_wake);
5261	spin_unlock(&ctx->completion_lock);
5262	if (ret || ipt.error)
5263		return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5264
5265	trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5266				mask, apoll->poll.events);
5267	return IO_APOLL_OK;
5268}
5269
5270static bool __io_poll_remove_one(struct io_kiocb *req,
5271				 struct io_poll_iocb *poll, bool do_cancel)
5272	__must_hold(&req->ctx->completion_lock)
5273{
5274	bool do_complete = false;
5275
5276	if (!poll->head)
5277		return false;
5278	spin_lock_irq(&poll->head->lock);
5279	if (do_cancel)
5280		WRITE_ONCE(poll->canceled, true);
5281	if (!list_empty(&poll->wait.entry)) {
5282		list_del_init(&poll->wait.entry);
5283		do_complete = true;
5284	}
5285	spin_unlock_irq(&poll->head->lock);
5286	hash_del(&req->hash_node);
5287	return do_complete;
5288}
5289
5290static bool io_poll_remove_one(struct io_kiocb *req)
5291	__must_hold(&req->ctx->completion_lock)
5292{
5293	bool do_complete;
5294
5295	io_poll_remove_double(req);
5296	do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5297
5298	if (do_complete) {
5299		io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5300		io_commit_cqring(req->ctx);
5301		req_set_fail(req);
5302		io_put_req_deferred(req);
5303	}
5304	return do_complete;
5305}
5306
5307/*
5308 * Returns true if we found and killed one or more poll requests
5309 */
5310static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5311			       bool cancel_all)
5312{
5313	struct hlist_node *tmp;
5314	struct io_kiocb *req;
5315	int posted = 0, i;
5316
5317	spin_lock(&ctx->completion_lock);
5318	for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5319		struct hlist_head *list;
5320
5321		list = &ctx->cancel_hash[i];
5322		hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5323			if (io_match_task(req, tsk, cancel_all))
5324				posted += io_poll_remove_one(req);
5325		}
5326	}
5327	spin_unlock(&ctx->completion_lock);
5328
5329	if (posted)
5330		io_cqring_ev_posted(ctx);
5331
5332	return posted != 0;
5333}
5334
5335static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5336				     bool poll_only)
5337	__must_hold(&ctx->completion_lock)
5338{
5339	struct hlist_head *list;
5340	struct io_kiocb *req;
5341
5342	list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5343	hlist_for_each_entry(req, list, hash_node) {
5344		if (sqe_addr != req->user_data)
5345			continue;
5346		if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5347			continue;
5348		return req;
5349	}
5350	return NULL;
5351}
5352
5353static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5354			  bool poll_only)
5355	__must_hold(&ctx->completion_lock)
5356{
5357	struct io_kiocb *req;
5358
5359	req = io_poll_find(ctx, sqe_addr, poll_only);
5360	if (!req)
5361		return -ENOENT;
5362	if (io_poll_remove_one(req))
5363		return 0;
5364
5365	return -EALREADY;
5366}
5367
5368static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5369				     unsigned int flags)
5370{
5371	u32 events;
5372
5373	events = READ_ONCE(sqe->poll32_events);
5374#ifdef __BIG_ENDIAN
5375	events = swahw32(events);
5376#endif
5377	if (!(flags & IORING_POLL_ADD_MULTI))
5378		events |= EPOLLONESHOT;
5379	return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5380}
5381
5382static int io_poll_update_prep(struct io_kiocb *req,
5383			       const struct io_uring_sqe *sqe)
5384{
5385	struct io_poll_update *upd = &req->poll_update;
5386	u32 flags;
5387
5388	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5389		return -EINVAL;
5390	if (sqe->ioprio || sqe->buf_index)
5391		return -EINVAL;
5392	flags = READ_ONCE(sqe->len);
5393	if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5394		      IORING_POLL_ADD_MULTI))
5395		return -EINVAL;
5396	/* meaningless without update */
5397	if (flags == IORING_POLL_ADD_MULTI)
5398		return -EINVAL;
5399
5400	upd->old_user_data = READ_ONCE(sqe->addr);
5401	upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5402	upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5403
5404	upd->new_user_data = READ_ONCE(sqe->off);
5405	if (!upd->update_user_data && upd->new_user_data)
5406		return -EINVAL;
5407	if (upd->update_events)
5408		upd->events = io_poll_parse_events(sqe, flags);
5409	else if (sqe->poll32_events)
5410		return -EINVAL;
5411
5412	return 0;
5413}
5414
5415static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5416			void *key)
5417{
5418	struct io_kiocb *req = wait->private;
5419	struct io_poll_iocb *poll = &req->poll;
5420
5421	return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5422}
5423
5424static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5425			       struct poll_table_struct *p)
5426{
5427	struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5428
5429	__io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5430}
5431
5432static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5433{
5434	struct io_poll_iocb *poll = &req->poll;
5435	u32 flags;
5436
5437	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5438		return -EINVAL;
5439	if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5440		return -EINVAL;
5441	flags = READ_ONCE(sqe->len);
5442	if (flags & ~IORING_POLL_ADD_MULTI)
5443		return -EINVAL;
5444
5445	io_req_set_refcount(req);
5446	poll->events = io_poll_parse_events(sqe, flags);
5447	return 0;
5448}
5449
5450static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5451{
5452	struct io_poll_iocb *poll = &req->poll;
5453	struct io_ring_ctx *ctx = req->ctx;
5454	struct io_poll_table ipt;
5455	__poll_t mask;
5456
5457	ipt.pt._qproc = io_poll_queue_proc;
5458
5459	mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5460					io_poll_wake);
5461
5462	if (mask) { /* no async, we'd stolen it */
5463		ipt.error = 0;
5464		io_poll_complete(req, mask);
5465	}
5466	spin_unlock(&ctx->completion_lock);
5467
5468	if (mask) {
5469		io_cqring_ev_posted(ctx);
5470		if (poll->events & EPOLLONESHOT)
5471			io_put_req(req);
5472	}
5473	return ipt.error;
5474}
5475
5476static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5477{
5478	struct io_ring_ctx *ctx = req->ctx;
5479	struct io_kiocb *preq;
5480	bool completing;
5481	int ret;
5482
5483	spin_lock(&ctx->completion_lock);
5484	preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5485	if (!preq) {
5486		ret = -ENOENT;
5487		goto err;
5488	}
5489
5490	if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5491		completing = true;
5492		ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5493		goto err;
5494	}
5495
5496	/*
5497	 * Don't allow racy completion with singleshot, as we cannot safely
5498	 * update those. For multishot, if we're racing with completion, just
5499	 * let completion re-add it.
5500	 */
5501	completing = !__io_poll_remove_one(preq, &preq->poll, false);
5502	if (completing && (preq->poll.events & EPOLLONESHOT)) {
5503		ret = -EALREADY;
5504		goto err;
5505	}
5506	/* we now have a detached poll request. reissue. */
5507	ret = 0;
5508err:
5509	if (ret < 0) {
5510		spin_unlock(&ctx->completion_lock);
5511		req_set_fail(req);
5512		io_req_complete(req, ret);
5513		return 0;
5514	}
5515	/* only mask one event flags, keep behavior flags */
5516	if (req->poll_update.update_events) {
5517		preq->poll.events &= ~0xffff;
5518		preq->poll.events |= req->poll_update.events & 0xffff;
5519		preq->poll.events |= IO_POLL_UNMASK;
5520	}
5521	if (req->poll_update.update_user_data)
5522		preq->user_data = req->poll_update.new_user_data;
5523	spin_unlock(&ctx->completion_lock);
5524
5525	/* complete update request, we're done with it */
5526	io_req_complete(req, ret);
5527
5528	if (!completing) {
5529		ret = io_poll_add(preq, issue_flags);
5530		if (ret < 0) {
5531			req_set_fail(preq);
5532			io_req_complete(preq, ret);
5533		}
5534	}
5535	return 0;
5536}
5537
5538static void io_req_task_timeout(struct io_kiocb *req)
5539{
5540	req_set_fail(req);
5541	io_req_complete_post(req, -ETIME, 0);
5542}
5543
5544static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5545{
5546	struct io_timeout_data *data = container_of(timer,
5547						struct io_timeout_data, timer);
5548	struct io_kiocb *req = data->req;
5549	struct io_ring_ctx *ctx = req->ctx;
5550	unsigned long flags;
5551
5552	spin_lock_irqsave(&ctx->timeout_lock, flags);
5553	list_del_init(&req->timeout.list);
5554	atomic_set(&req->ctx->cq_timeouts,
5555		atomic_read(&req->ctx->cq_timeouts) + 1);
5556	spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5557
5558	req->io_task_work.func = io_req_task_timeout;
5559	io_req_task_work_add(req);
5560	return HRTIMER_NORESTART;
5561}
5562
5563static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5564					   __u64 user_data)
5565	__must_hold(&ctx->timeout_lock)
5566{
5567	struct io_timeout_data *io;
5568	struct io_kiocb *req;
5569	bool found = false;
5570
5571	list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5572		found = user_data == req->user_data;
5573		if (found)
5574			break;
5575	}
5576	if (!found)
5577		return ERR_PTR(-ENOENT);
5578
5579	io = req->async_data;
5580	if (hrtimer_try_to_cancel(&io->timer) == -1)
5581		return ERR_PTR(-EALREADY);
5582	list_del_init(&req->timeout.list);
5583	return req;
5584}
5585
5586static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5587	__must_hold(&ctx->completion_lock)
5588	__must_hold(&ctx->timeout_lock)
5589{
5590	struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5591
5592	if (IS_ERR(req))
5593		return PTR_ERR(req);
5594
5595	req_set_fail(req);
5596	io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5597	io_put_req_deferred(req);
5598	return 0;
5599}
5600
5601static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5602			     struct timespec64 *ts, enum hrtimer_mode mode)
5603	__must_hold(&ctx->timeout_lock)
5604{
5605	struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5606	struct io_timeout_data *data;
5607
5608	if (IS_ERR(req))
5609		return PTR_ERR(req);
5610
5611	req->timeout.off = 0; /* noseq */
5612	data = req->async_data;
5613	list_add_tail(&req->timeout.list, &ctx->timeout_list);
5614	hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5615	data->timer.function = io_timeout_fn;
5616	hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5617	return 0;
5618}
5619
5620static int io_timeout_remove_prep(struct io_kiocb *req,
5621				  const struct io_uring_sqe *sqe)
5622{
5623	struct io_timeout_rem *tr = &req->timeout_rem;
5624
5625	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5626		return -EINVAL;
5627	if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5628		return -EINVAL;
5629	if (sqe->ioprio || sqe->buf_index || sqe->len)
5630		return -EINVAL;
5631
5632	tr->addr = READ_ONCE(sqe->addr);
5633	tr->flags = READ_ONCE(sqe->timeout_flags);
5634	if (tr->flags & IORING_TIMEOUT_UPDATE) {
5635		if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5636			return -EINVAL;
5637		if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5638			return -EFAULT;
5639	} else if (tr->flags) {
5640		/* timeout removal doesn't support flags */
5641		return -EINVAL;
5642	}
5643
5644	return 0;
5645}
5646
5647static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5648{
5649	return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5650					    : HRTIMER_MODE_REL;
5651}
5652
5653/*
5654 * Remove or update an existing timeout command
5655 */
5656static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5657{
5658	struct io_timeout_rem *tr = &req->timeout_rem;
5659	struct io_ring_ctx *ctx = req->ctx;
5660	int ret;
5661
5662	if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
5663		spin_lock(&ctx->completion_lock);
5664		spin_lock_irq(&ctx->timeout_lock);
5665		ret = io_timeout_cancel(ctx, tr->addr);
5666		spin_unlock_irq(&ctx->timeout_lock);
5667		spin_unlock(&ctx->completion_lock);
5668	} else {
5669		spin_lock_irq(&ctx->timeout_lock);
5670		ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5671					io_translate_timeout_mode(tr->flags));
5672		spin_unlock_irq(&ctx->timeout_lock);
5673	}
5674
5675	if (ret < 0)
5676		req_set_fail(req);
5677	io_req_complete_post(req, ret, 0);
5678	return 0;
5679}
5680
5681static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5682			   bool is_timeout_link)
5683{
5684	struct io_timeout_data *data;
5685	unsigned flags;
5686	u32 off = READ_ONCE(sqe->off);
5687
5688	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5689		return -EINVAL;
5690	if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5691		return -EINVAL;
5692	if (off && is_timeout_link)
5693		return -EINVAL;
5694	flags = READ_ONCE(sqe->timeout_flags);
5695	if (flags & ~IORING_TIMEOUT_ABS)
5696		return -EINVAL;
5697
5698	req->timeout.off = off;
5699	if (unlikely(off && !req->ctx->off_timeout_used))
5700		req->ctx->off_timeout_used = true;
5701
5702	if (!req->async_data && io_alloc_async_data(req))
5703		return -ENOMEM;
5704
5705	data = req->async_data;
5706	data->req = req;
5707
5708	if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5709		return -EFAULT;
5710
5711	data->mode = io_translate_timeout_mode(flags);
5712	hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5713
5714	if (is_timeout_link) {
5715		struct io_submit_link *link = &req->ctx->submit_state.link;
5716
5717		if (!link->head)
5718			return -EINVAL;
5719		if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
5720			return -EINVAL;
5721		req->timeout.head = link->last;
5722		link->last->flags |= REQ_F_ARM_LTIMEOUT;
5723	}
5724	return 0;
5725}
5726
5727static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5728{
5729	struct io_ring_ctx *ctx = req->ctx;
5730	struct io_timeout_data *data = req->