io_uring.c revision 99c8bc52
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
1529static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1530{
1531	/*
1532	 * wake_up_all() may seem excessive, but io_wake_function() and
1533	 * io_should_wake() handle the termination of the loop and only
1534	 * wake as many waiters as we need to.
1535	 */
1536	if (wq_has_sleeper(&ctx->cq_wait))
1537		wake_up_all(&ctx->cq_wait);
1538	if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1539		wake_up(&ctx->sq_data->wait);
1540	if (io_should_trigger_evfd(ctx))
1541		eventfd_signal(ctx->cq_ev_fd, 1);
1542	if (waitqueue_active(&ctx->poll_wait)) {
1543		wake_up_interruptible(&ctx->poll_wait);
1544		kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1545	}
1546}
1547
1548static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1549{
1550	if (ctx->flags & IORING_SETUP_SQPOLL) {
1551		if (wq_has_sleeper(&ctx->cq_wait))
1552			wake_up_all(&ctx->cq_wait);
1553	}
1554	if (io_should_trigger_evfd(ctx))
1555		eventfd_signal(ctx->cq_ev_fd, 1);
1556	if (waitqueue_active(&ctx->poll_wait)) {
1557		wake_up_interruptible(&ctx->poll_wait);
1558		kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1559	}
1560}
1561
1562/* Returns true if there are no backlogged entries after the flush */
1563static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1564{
1565	bool all_flushed, posted;
1566
1567	if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1568		return false;
1569
1570	posted = false;
1571	spin_lock(&ctx->completion_lock);
1572	while (!list_empty(&ctx->cq_overflow_list)) {
1573		struct io_uring_cqe *cqe = io_get_cqe(ctx);
1574		struct io_overflow_cqe *ocqe;
1575
1576		if (!cqe && !force)
1577			break;
1578		ocqe = list_first_entry(&ctx->cq_overflow_list,
1579					struct io_overflow_cqe, list);
1580		if (cqe)
1581			memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1582		else
1583			io_account_cq_overflow(ctx);
1584
1585		posted = true;
1586		list_del(&ocqe->list);
1587		kfree(ocqe);
1588	}
1589
1590	all_flushed = list_empty(&ctx->cq_overflow_list);
1591	if (all_flushed) {
1592		clear_bit(0, &ctx->check_cq_overflow);
1593		WRITE_ONCE(ctx->rings->sq_flags,
1594			   ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1595	}
1596
1597	if (posted)
1598		io_commit_cqring(ctx);
1599	spin_unlock(&ctx->completion_lock);
1600	if (posted)
1601		io_cqring_ev_posted(ctx);
1602	return all_flushed;
1603}
1604
1605static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1606{
1607	bool ret = true;
1608
1609	if (test_bit(0, &ctx->check_cq_overflow)) {
1610		/* iopoll syncs against uring_lock, not completion_lock */
1611		if (ctx->flags & IORING_SETUP_IOPOLL)
1612			mutex_lock(&ctx->uring_lock);
1613		ret = __io_cqring_overflow_flush(ctx, false);
1614		if (ctx->flags & IORING_SETUP_IOPOLL)
1615			mutex_unlock(&ctx->uring_lock);
1616	}
1617
1618	return ret;
1619}
1620
1621/* must to be called somewhat shortly after putting a request */
1622static inline void io_put_task(struct task_struct *task, int nr)
1623{
1624	struct io_uring_task *tctx = task->io_uring;
1625
1626	if (likely(task == current)) {
1627		tctx->cached_refs += nr;
1628	} else {
1629		percpu_counter_sub(&tctx->inflight, nr);
1630		if (unlikely(atomic_read(&tctx->in_idle)))
1631			wake_up(&tctx->wait);
1632		put_task_struct_many(task, nr);
1633	}
1634}
1635
1636static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1637				     long res, unsigned int cflags)
1638{
1639	struct io_overflow_cqe *ocqe;
1640
1641	ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1642	if (!ocqe) {
1643		/*
1644		 * If we're in ring overflow flush mode, or in task cancel mode,
1645		 * or cannot allocate an overflow entry, then we need to drop it
1646		 * on the floor.
1647		 */
1648		io_account_cq_overflow(ctx);
1649		return false;
1650	}
1651	if (list_empty(&ctx->cq_overflow_list)) {
1652		set_bit(0, &ctx->check_cq_overflow);
1653		WRITE_ONCE(ctx->rings->sq_flags,
1654			   ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1655
1656	}
1657	ocqe->cqe.user_data = user_data;
1658	ocqe->cqe.res = res;
1659	ocqe->cqe.flags = cflags;
1660	list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1661	return true;
1662}
1663
1664static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1665					  long res, unsigned int cflags)
1666{
1667	struct io_uring_cqe *cqe;
1668
1669	trace_io_uring_complete(ctx, user_data, res, cflags);
1670
1671	/*
1672	 * If we can't get a cq entry, userspace overflowed the
1673	 * submission (by quite a lot). Increment the overflow count in
1674	 * the ring.
1675	 */
1676	cqe = io_get_cqe(ctx);
1677	if (likely(cqe)) {
1678		WRITE_ONCE(cqe->user_data, user_data);
1679		WRITE_ONCE(cqe->res, res);
1680		WRITE_ONCE(cqe->flags, cflags);
1681		return true;
1682	}
1683	return io_cqring_event_overflow(ctx, user_data, res, cflags);
1684}
1685
1686/* not as hot to bloat with inlining */
1687static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1688					  long res, unsigned int cflags)
1689{
1690	return __io_cqring_fill_event(ctx, user_data, res, cflags);
1691}
1692
1693static void io_req_complete_post(struct io_kiocb *req, long res,
1694				 unsigned int cflags)
1695{
1696	struct io_ring_ctx *ctx = req->ctx;
1697
1698	spin_lock(&ctx->completion_lock);
1699	__io_cqring_fill_event(ctx, req->user_data, res, cflags);
1700	/*
1701	 * If we're the last reference to this request, add to our locked
1702	 * free_list cache.
1703	 */
1704	if (req_ref_put_and_test(req)) {
1705		if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1706			if (req->flags & IO_DISARM_MASK)
1707				io_disarm_next(req);
1708			if (req->link) {
1709				io_req_task_queue(req->link);
1710				req->link = NULL;
1711			}
1712		}
1713		io_dismantle_req(req);
1714		io_put_task(req->task, 1);
1715		list_add(&req->inflight_entry, &ctx->locked_free_list);
1716		ctx->locked_free_nr++;
1717	} else {
1718		if (!percpu_ref_tryget(&ctx->refs))
1719			req = NULL;
1720	}
1721	io_commit_cqring(ctx);
1722	spin_unlock(&ctx->completion_lock);
1723
1724	if (req) {
1725		io_cqring_ev_posted(ctx);
1726		percpu_ref_put(&ctx->refs);
1727	}
1728}
1729
1730static inline bool io_req_needs_clean(struct io_kiocb *req)
1731{
1732	return req->flags & IO_REQ_CLEAN_FLAGS;
1733}
1734
1735static void io_req_complete_state(struct io_kiocb *req, long res,
1736				  unsigned int cflags)
1737{
1738	if (io_req_needs_clean(req))
1739		io_clean_op(req);
1740	req->result = res;
1741	req->compl.cflags = cflags;
1742	req->flags |= REQ_F_COMPLETE_INLINE;
1743}
1744
1745static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1746				     long res, unsigned cflags)
1747{
1748	if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1749		io_req_complete_state(req, res, cflags);
1750	else
1751		io_req_complete_post(req, res, cflags);
1752}
1753
1754static inline void io_req_complete(struct io_kiocb *req, long res)
1755{
1756	__io_req_complete(req, 0, res, 0);
1757}
1758
1759static void io_req_complete_failed(struct io_kiocb *req, long res)
1760{
1761	req_set_fail(req);
1762	io_req_complete_post(req, res, 0);
1763}
1764
1765/*
1766 * Don't initialise the fields below on every allocation, but do that in
1767 * advance and keep them valid across allocations.
1768 */
1769static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1770{
1771	req->ctx = ctx;
1772	req->link = NULL;
1773	req->async_data = NULL;
1774	/* not necessary, but safer to zero */
1775	req->result = 0;
1776}
1777
1778static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1779					struct io_submit_state *state)
1780{
1781	spin_lock(&ctx->completion_lock);
1782	list_splice_init(&ctx->locked_free_list, &state->free_list);
1783	ctx->locked_free_nr = 0;
1784	spin_unlock(&ctx->completion_lock);
1785}
1786
1787/* Returns true IFF there are requests in the cache */
1788static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1789{
1790	struct io_submit_state *state = &ctx->submit_state;
1791	int nr;
1792
1793	/*
1794	 * If we have more than a batch's worth of requests in our IRQ side
1795	 * locked cache, grab the lock and move them over to our submission
1796	 * side cache.
1797	 */
1798	if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1799		io_flush_cached_locked_reqs(ctx, state);
1800
1801	nr = state->free_reqs;
1802	while (!list_empty(&state->free_list)) {
1803		struct io_kiocb *req = list_first_entry(&state->free_list,
1804					struct io_kiocb, inflight_entry);
1805
1806		list_del(&req->inflight_entry);
1807		state->reqs[nr++] = req;
1808		if (nr == ARRAY_SIZE(state->reqs))
1809			break;
1810	}
1811
1812	state->free_reqs = nr;
1813	return nr != 0;
1814}
1815
1816/*
1817 * A request might get retired back into the request caches even before opcode
1818 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1819 * Because of that, io_alloc_req() should be called only under ->uring_lock
1820 * and with extra caution to not get a request that is still worked on.
1821 */
1822static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1823	__must_hold(&ctx->uring_lock)
1824{
1825	struct io_submit_state *state = &ctx->submit_state;
1826	gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1827	int ret, i;
1828
1829	BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1830
1831	if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1832		goto got_req;
1833
1834	ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1835				    state->reqs);
1836
1837	/*
1838	 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1839	 * retry single alloc to be on the safe side.
1840	 */
1841	if (unlikely(ret <= 0)) {
1842		state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1843		if (!state->reqs[0])
1844			return NULL;
1845		ret = 1;
1846	}
1847
1848	for (i = 0; i < ret; i++)
1849		io_preinit_req(state->reqs[i], ctx);
1850	state->free_reqs = ret;
1851got_req:
1852	state->free_reqs--;
1853	return state->reqs[state->free_reqs];
1854}
1855
1856static inline void io_put_file(struct file *file)
1857{
1858	if (file)
1859		fput(file);
1860}
1861
1862static void io_dismantle_req(struct io_kiocb *req)
1863{
1864	unsigned int flags = req->flags;
1865
1866	if (io_req_needs_clean(req))
1867		io_clean_op(req);
1868	if (!(flags & REQ_F_FIXED_FILE))
1869		io_put_file(req->file);
1870	if (req->fixed_rsrc_refs)
1871		percpu_ref_put(req->fixed_rsrc_refs);
1872	if (req->async_data) {
1873		kfree(req->async_data);
1874		req->async_data = NULL;
1875	}
1876}
1877
1878static void __io_free_req(struct io_kiocb *req)
1879{
1880	struct io_ring_ctx *ctx = req->ctx;
1881
1882	io_dismantle_req(req);
1883	io_put_task(req->task, 1);
1884
1885	spin_lock(&ctx->completion_lock);
1886	list_add(&req->inflight_entry, &ctx->locked_free_list);
1887	ctx->locked_free_nr++;
1888	spin_unlock(&ctx->completion_lock);
1889
1890	percpu_ref_put(&ctx->refs);
1891}
1892
1893static inline void io_remove_next_linked(struct io_kiocb *req)
1894{
1895	struct io_kiocb *nxt = req->link;
1896
1897	req->link = nxt->link;
1898	nxt->link = NULL;
1899}
1900
1901static bool io_kill_linked_timeout(struct io_kiocb *req)
1902	__must_hold(&req->ctx->completion_lock)
1903	__must_hold(&req->ctx->timeout_lock)
1904{
1905	struct io_kiocb *link = req->link;
1906
1907	if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
1908		struct io_timeout_data *io = link->async_data;
1909
1910		io_remove_next_linked(req);
1911		link->timeout.head = NULL;
1912		if (hrtimer_try_to_cancel(&io->timer) != -1) {
1913			io_cqring_fill_event(link->ctx, link->user_data,
1914					     -ECANCELED, 0);
1915			io_put_req_deferred(link);
1916			return true;
1917		}
1918	}
1919	return false;
1920}
1921
1922static void io_fail_links(struct io_kiocb *req)
1923	__must_hold(&req->ctx->completion_lock)
1924{
1925	struct io_kiocb *nxt, *link = req->link;
1926
1927	req->link = NULL;
1928	while (link) {
1929		nxt = link->link;
1930		link->link = NULL;
1931
1932		trace_io_uring_fail_link(req, link);
1933		io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1934		io_put_req_deferred(link);
1935		link = nxt;
1936	}
1937}
1938
1939static bool io_disarm_next(struct io_kiocb *req)
1940	__must_hold(&req->ctx->completion_lock)
1941{
1942	bool posted = false;
1943
1944	if (req->flags & REQ_F_ARM_LTIMEOUT) {
1945		struct io_kiocb *link = req->link;
1946
1947		req->flags &= ~REQ_F_ARM_LTIMEOUT;
1948		if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
1949			io_remove_next_linked(req);
1950			io_cqring_fill_event(link->ctx, link->user_data,
1951					     -ECANCELED, 0);
1952			io_put_req_deferred(link);
1953			posted = true;
1954		}
1955	} else if (req->flags & REQ_F_LINK_TIMEOUT) {
1956		struct io_ring_ctx *ctx = req->ctx;
1957
1958		spin_lock_irq(&ctx->timeout_lock);
1959		posted = io_kill_linked_timeout(req);
1960		spin_unlock_irq(&ctx->timeout_lock);
1961	}
1962	if (unlikely((req->flags & REQ_F_FAIL) &&
1963		     !(req->flags & REQ_F_HARDLINK))) {
1964		posted |= (req->link != NULL);
1965		io_fail_links(req);
1966	}
1967	return posted;
1968}
1969
1970static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1971{
1972	struct io_kiocb *nxt;
1973
1974	/*
1975	 * If LINK is set, we have dependent requests in this chain. If we
1976	 * didn't fail this request, queue the first one up, moving any other
1977	 * dependencies to the next request. In case of failure, fail the rest
1978	 * of the chain.
1979	 */
1980	if (req->flags & IO_DISARM_MASK) {
1981		struct io_ring_ctx *ctx = req->ctx;
1982		bool posted;
1983
1984		spin_lock(&ctx->completion_lock);
1985		posted = io_disarm_next(req);
1986		if (posted)
1987			io_commit_cqring(req->ctx);
1988		spin_unlock(&ctx->completion_lock);
1989		if (posted)
1990			io_cqring_ev_posted(ctx);
1991	}
1992	nxt = req->link;
1993	req->link = NULL;
1994	return nxt;
1995}
1996
1997static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1998{
1999	if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2000		return NULL;
2001	return __io_req_find_next(req);
2002}
2003
2004static void ctx_flush_and_put(struct io_ring_ctx *ctx)
2005{
2006	if (!ctx)
2007		return;
2008	if (ctx->submit_state.compl_nr) {
2009		mutex_lock(&ctx->uring_lock);
2010		if (ctx->submit_state.compl_nr)
2011			io_submit_flush_completions(ctx);
2012		mutex_unlock(&ctx->uring_lock);
2013	}
2014	percpu_ref_put(&ctx->refs);
2015}
2016
2017static void tctx_task_work(struct callback_head *cb)
2018{
2019	struct io_ring_ctx *ctx = NULL;
2020	struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2021						  task_work);
2022
2023	while (1) {
2024		struct io_wq_work_node *node;
2025
2026		spin_lock_irq(&tctx->task_lock);
2027		node = tctx->task_list.first;
2028		INIT_WQ_LIST(&tctx->task_list);
2029		if (!node)
2030			tctx->task_running = false;
2031		spin_unlock_irq(&tctx->task_lock);
2032		if (!node)
2033			break;
2034
2035		do {
2036			struct io_wq_work_node *next = node->next;
2037			struct io_kiocb *req = container_of(node, struct io_kiocb,
2038							    io_task_work.node);
2039
2040			if (req->ctx != ctx) {
2041				ctx_flush_and_put(ctx);
2042				ctx = req->ctx;
2043				percpu_ref_get(&ctx->refs);
2044			}
2045			req->io_task_work.func(req);
2046			node = next;
2047		} while (node);
2048
2049		cond_resched();
2050	}
2051
2052	ctx_flush_and_put(ctx);
2053}
2054
2055static void io_req_task_work_add(struct io_kiocb *req)
2056{
2057	struct task_struct *tsk = req->task;
2058	struct io_uring_task *tctx = tsk->io_uring;
2059	enum task_work_notify_mode notify;
2060	struct io_wq_work_node *node;
2061	unsigned long flags;
2062	bool running;
2063
2064	WARN_ON_ONCE(!tctx);
2065
2066	spin_lock_irqsave(&tctx->task_lock, flags);
2067	wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2068	running = tctx->task_running;
2069	if (!running)
2070		tctx->task_running = true;
2071	spin_unlock_irqrestore(&tctx->task_lock, flags);
2072
2073	/* task_work already pending, we're done */
2074	if (running)
2075		return;
2076
2077	/*
2078	 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2079	 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2080	 * processing task_work. There's no reliable way to tell if TWA_RESUME
2081	 * will do the job.
2082	 */
2083	notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2084	if (!task_work_add(tsk, &tctx->task_work, notify)) {
2085		wake_up_process(tsk);
2086		return;
2087	}
2088
2089	spin_lock_irqsave(&tctx->task_lock, flags);
2090	tctx->task_running = false;
2091	node = tctx->task_list.first;
2092	INIT_WQ_LIST(&tctx->task_list);
2093	spin_unlock_irqrestore(&tctx->task_lock, flags);
2094
2095	while (node) {
2096		req = container_of(node, struct io_kiocb, io_task_work.node);
2097		node = node->next;
2098		if (llist_add(&req->io_task_work.fallback_node,
2099			      &req->ctx->fallback_llist))
2100			schedule_delayed_work(&req->ctx->fallback_work, 1);
2101	}
2102}
2103
2104static void io_req_task_cancel(struct io_kiocb *req)
2105{
2106	struct io_ring_ctx *ctx = req->ctx;
2107
2108	/* ctx is guaranteed to stay alive while we hold uring_lock */
2109	mutex_lock(&ctx->uring_lock);
2110	io_req_complete_failed(req, req->result);
2111	mutex_unlock(&ctx->uring_lock);
2112}
2113
2114static void io_req_task_submit(struct io_kiocb *req)
2115{
2116	struct io_ring_ctx *ctx = req->ctx;
2117
2118	/* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2119	mutex_lock(&ctx->uring_lock);
2120	/* req->task == current here, checking PF_EXITING is safe */
2121	if (likely(!(req->task->flags & PF_EXITING)))
2122		__io_queue_sqe(req);
2123	else
2124		io_req_complete_failed(req, -EFAULT);
2125	mutex_unlock(&ctx->uring_lock);
2126}
2127
2128static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2129{
2130	req->result = ret;
2131	req->io_task_work.func = io_req_task_cancel;
2132	io_req_task_work_add(req);
2133}
2134
2135static void io_req_task_queue(struct io_kiocb *req)
2136{
2137	req->io_task_work.func = io_req_task_submit;
2138	io_req_task_work_add(req);
2139}
2140
2141static void io_req_task_queue_reissue(struct io_kiocb *req)
2142{
2143	req->io_task_work.func = io_queue_async_work;
2144	io_req_task_work_add(req);
2145}
2146
2147static inline void io_queue_next(struct io_kiocb *req)
2148{
2149	struct io_kiocb *nxt = io_req_find_next(req);
2150
2151	if (nxt)
2152		io_req_task_queue(nxt);
2153}
2154
2155static void io_free_req(struct io_kiocb *req)
2156{
2157	io_queue_next(req);
2158	__io_free_req(req);
2159}
2160
2161struct req_batch {
2162	struct task_struct	*task;
2163	int			task_refs;
2164	int			ctx_refs;
2165};
2166
2167static inline void io_init_req_batch(struct req_batch *rb)
2168{
2169	rb->task_refs = 0;
2170	rb->ctx_refs = 0;
2171	rb->task = NULL;
2172}
2173
2174static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2175				     struct req_batch *rb)
2176{
2177	if (rb->ctx_refs)
2178		percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2179	if (rb->task)
2180		io_put_task(rb->task, rb->task_refs);
2181}
2182
2183static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2184			      struct io_submit_state *state)
2185{
2186	io_queue_next(req);
2187	io_dismantle_req(req);
2188
2189	if (req->task != rb->task) {
2190		if (rb->task)
2191			io_put_task(rb->task, rb->task_refs);
2192		rb->task = req->task;
2193		rb->task_refs = 0;
2194	}
2195	rb->task_refs++;
2196	rb->ctx_refs++;
2197
2198	if (state->free_reqs != ARRAY_SIZE(state->reqs))
2199		state->reqs[state->free_reqs++] = req;
2200	else
2201		list_add(&req->inflight_entry, &state->free_list);
2202}
2203
2204static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2205	__must_hold(&ctx->uring_lock)
2206{
2207	struct io_submit_state *state = &ctx->submit_state;
2208	int i, nr = state->compl_nr;
2209	struct req_batch rb;
2210
2211	spin_lock(&ctx->completion_lock);
2212	for (i = 0; i < nr; i++) {
2213		struct io_kiocb *req = state->compl_reqs[i];
2214
2215		__io_cqring_fill_event(ctx, req->user_data, req->result,
2216					req->compl.cflags);
2217	}
2218	io_commit_cqring(ctx);
2219	spin_unlock(&ctx->completion_lock);
2220	io_cqring_ev_posted(ctx);
2221
2222	io_init_req_batch(&rb);
2223	for (i = 0; i < nr; i++) {
2224		struct io_kiocb *req = state->compl_reqs[i];
2225
2226		if (req_ref_put_and_test(req))
2227			io_req_free_batch(&rb, req, &ctx->submit_state);
2228	}
2229
2230	io_req_free_batch_finish(ctx, &rb);
2231	state->compl_nr = 0;
2232}
2233
2234/*
2235 * Drop reference to request, return next in chain (if there is one) if this
2236 * was the last reference to this request.
2237 */
2238static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2239{
2240	struct io_kiocb *nxt = NULL;
2241
2242	if (req_ref_put_and_test(req)) {
2243		nxt = io_req_find_next(req);
2244		__io_free_req(req);
2245	}
2246	return nxt;
2247}
2248
2249static inline void io_put_req(struct io_kiocb *req)
2250{
2251	if (req_ref_put_and_test(req))
2252		io_free_req(req);
2253}
2254
2255static inline void io_put_req_deferred(struct io_kiocb *req)
2256{
2257	if (req_ref_put_and_test(req)) {
2258		req->io_task_work.func = io_free_req;
2259		io_req_task_work_add(req);
2260	}
2261}
2262
2263static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2264{
2265	/* See comment at the top of this file */
2266	smp_rmb();
2267	return __io_cqring_events(ctx);
2268}
2269
2270static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2271{
2272	struct io_rings *rings = ctx->rings;
2273
2274	/* make sure SQ entry isn't read before tail */
2275	return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2276}
2277
2278static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2279{
2280	unsigned int cflags;
2281
2282	cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2283	cflags |= IORING_CQE_F_BUFFER;
2284	req->flags &= ~REQ_F_BUFFER_SELECTED;
2285	kfree(kbuf);
2286	return cflags;
2287}
2288
2289static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2290{
2291	struct io_buffer *kbuf;
2292
2293	if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
2294		return 0;
2295	kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2296	return io_put_kbuf(req, kbuf);
2297}
2298
2299static inline bool io_run_task_work(void)
2300{
2301	if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2302		__set_current_state(TASK_RUNNING);
2303		tracehook_notify_signal();
2304		return true;
2305	}
2306
2307	return false;
2308}
2309
2310/*
2311 * Find and free completed poll iocbs
2312 */
2313static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2314			       struct list_head *done)
2315{
2316	struct req_batch rb;
2317	struct io_kiocb *req;
2318
2319	/* order with ->result store in io_complete_rw_iopoll() */
2320	smp_rmb();
2321
2322	io_init_req_batch(&rb);
2323	while (!list_empty(done)) {
2324		req = list_first_entry(done, struct io_kiocb, inflight_entry);
2325		list_del(&req->inflight_entry);
2326
2327		if (READ_ONCE(req->result) == -EAGAIN &&
2328		    !(req->flags & REQ_F_DONT_REISSUE)) {
2329			req->iopoll_completed = 0;
2330			io_req_task_queue_reissue(req);
2331			continue;
2332		}
2333
2334		__io_cqring_fill_event(ctx, req->user_data, req->result,
2335					io_put_rw_kbuf(req));
2336		(*nr_events)++;
2337
2338		if (req_ref_put_and_test(req))
2339			io_req_free_batch(&rb, req, &ctx->submit_state);
2340	}
2341
2342	io_commit_cqring(ctx);
2343	io_cqring_ev_posted_iopoll(ctx);
2344	io_req_free_batch_finish(ctx, &rb);
2345}
2346
2347static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2348			long min)
2349{
2350	struct io_kiocb *req, *tmp;
2351	LIST_HEAD(done);
2352	bool spin;
2353
2354	/*
2355	 * Only spin for completions if we don't have multiple devices hanging
2356	 * off our complete list, and we're under the requested amount.
2357	 */
2358	spin = !ctx->poll_multi_queue && *nr_events < min;
2359
2360	list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2361		struct kiocb *kiocb = &req->rw.kiocb;
2362		int ret;
2363
2364		/*
2365		 * Move completed and retryable entries to our local lists.
2366		 * If we find a request that requires polling, break out
2367		 * and complete those lists first, if we have entries there.
2368		 */
2369		if (READ_ONCE(req->iopoll_completed)) {
2370			list_move_tail(&req->inflight_entry, &done);
2371			continue;
2372		}
2373		if (!list_empty(&done))
2374			break;
2375
2376		ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2377		if (unlikely(ret < 0))
2378			return ret;
2379		else if (ret)
2380			spin = false;
2381
2382		/* iopoll may have completed current req */
2383		if (READ_ONCE(req->iopoll_completed))
2384			list_move_tail(&req->inflight_entry, &done);
2385	}
2386
2387	if (!list_empty(&done))
2388		io_iopoll_complete(ctx, nr_events, &done);
2389
2390	return 0;
2391}
2392
2393/*
2394 * We can't just wait for polled events to come to us, we have to actively
2395 * find and complete them.
2396 */
2397static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2398{
2399	if (!(ctx->flags & IORING_SETUP_IOPOLL))
2400		return;
2401
2402	mutex_lock(&ctx->uring_lock);
2403	while (!list_empty(&ctx->iopoll_list)) {
2404		unsigned int nr_events = 0;
2405
2406		io_do_iopoll(ctx, &nr_events, 0);
2407
2408		/* let it sleep and repeat later if can't complete a request */
2409		if (nr_events == 0)
2410			break;
2411		/*
2412		 * Ensure we allow local-to-the-cpu processing to take place,
2413		 * in this case we need to ensure that we reap all events.
2414		 * Also let task_work, etc. to progress by releasing the mutex
2415		 */
2416		if (need_resched()) {
2417			mutex_unlock(&ctx->uring_lock);
2418			cond_resched();
2419			mutex_lock(&ctx->uring_lock);
2420		}
2421	}
2422	mutex_unlock(&ctx->uring_lock);
2423}
2424
2425static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2426{
2427	unsigned int nr_events = 0;
2428	int ret = 0;
2429
2430	/*
2431	 * We disallow the app entering submit/complete with polling, but we
2432	 * still need to lock the ring to prevent racing with polled issue
2433	 * that got punted to a workqueue.
2434	 */
2435	mutex_lock(&ctx->uring_lock);
2436	/*
2437	 * Don't enter poll loop if we already have events pending.
2438	 * If we do, we can potentially be spinning for commands that
2439	 * already triggered a CQE (eg in error).
2440	 */
2441	if (test_bit(0, &ctx->check_cq_overflow))
2442		__io_cqring_overflow_flush(ctx, false);
2443	if (io_cqring_events(ctx))
2444		goto out;
2445	do {
2446		/*
2447		 * If a submit got punted to a workqueue, we can have the
2448		 * application entering polling for a command before it gets
2449		 * issued. That app will hold the uring_lock for the duration
2450		 * of the poll right here, so we need to take a breather every
2451		 * now and then to ensure that the issue has a chance to add
2452		 * the poll to the issued list. Otherwise we can spin here
2453		 * forever, while the workqueue is stuck trying to acquire the
2454		 * very same mutex.
2455		 */
2456		if (list_empty(&ctx->iopoll_list)) {
2457			u32 tail = ctx->cached_cq_tail;
2458
2459			mutex_unlock(&ctx->uring_lock);
2460			io_run_task_work();
2461			mutex_lock(&ctx->uring_lock);
2462
2463			/* some requests don't go through iopoll_list */
2464			if (tail != ctx->cached_cq_tail ||
2465			    list_empty(&ctx->iopoll_list))
2466				break;
2467		}
2468		ret = io_do_iopoll(ctx, &nr_events, min);
2469	} while (!ret && nr_events < min && !need_resched());
2470out:
2471	mutex_unlock(&ctx->uring_lock);
2472	return ret;
2473}
2474
2475static void kiocb_end_write(struct io_kiocb *req)
2476{
2477	/*
2478	 * Tell lockdep we inherited freeze protection from submission
2479	 * thread.
2480	 */
2481	if (req->flags & REQ_F_ISREG) {
2482		struct super_block *sb = file_inode(req->file)->i_sb;
2483
2484		__sb_writers_acquired(sb, SB_FREEZE_WRITE);
2485		sb_end_write(sb);
2486	}
2487}
2488
2489#ifdef CONFIG_BLOCK
2490static bool io_resubmit_prep(struct io_kiocb *req)
2491{
2492	struct io_async_rw *rw = req->async_data;
2493
2494	if (!rw)
2495		return !io_req_prep_async(req);
2496	/* may have left rw->iter inconsistent on -EIOCBQUEUED */
2497	iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2498	return true;
2499}
2500
2501static bool io_rw_should_reissue(struct io_kiocb *req)
2502{
2503	umode_t mode = file_inode(req->file)->i_mode;
2504	struct io_ring_ctx *ctx = req->ctx;
2505
2506	if (!S_ISBLK(mode) && !S_ISREG(mode))
2507		return false;
2508	if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2509	    !(ctx->flags & IORING_SETUP_IOPOLL)))
2510		return false;
2511	/*
2512	 * If ref is dying, we might be running poll reap from the exit work.
2513	 * Don't attempt to reissue from that path, just let it fail with
2514	 * -EAGAIN.
2515	 */
2516	if (percpu_ref_is_dying(&ctx->refs))
2517		return false;
2518	/*
2519	 * Play it safe and assume not safe to re-import and reissue if we're
2520	 * not in the original thread group (or in task context).
2521	 */
2522	if (!same_thread_group(req->task, current) || !in_task())
2523		return false;
2524	return true;
2525}
2526#else
2527static bool io_resubmit_prep(struct io_kiocb *req)
2528{
2529	return false;
2530}
2531static bool io_rw_should_reissue(struct io_kiocb *req)
2532{
2533	return false;
2534}
2535#endif
2536
2537static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2538{
2539	if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2540		kiocb_end_write(req);
2541	if (res != req->result) {
2542		if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2543		    io_rw_should_reissue(req)) {
2544			req->flags |= REQ_F_REISSUE;
2545			return true;
2546		}
2547		req_set_fail(req);
2548		req->result = res;
2549	}
2550	return false;
2551}
2552
2553static void io_req_task_complete(struct io_kiocb *req)
2554{
2555	__io_req_complete(req, 0, req->result, io_put_rw_kbuf(req));
2556}
2557
2558static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2559			     unsigned int issue_flags)
2560{
2561	if (__io_complete_rw_common(req, res))
2562		return;
2563	io_req_task_complete(req);
2564}
2565
2566static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2567{
2568	struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2569
2570	if (__io_complete_rw_common(req, res))
2571		return;
2572	req->result = res;
2573	req->io_task_work.func = io_req_task_complete;
2574	io_req_task_work_add(req);
2575}
2576
2577static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2578{
2579	struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2580
2581	if (kiocb->ki_flags & IOCB_WRITE)
2582		kiocb_end_write(req);
2583	if (unlikely(res != req->result)) {
2584		if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2585		    io_resubmit_prep(req))) {
2586			req_set_fail(req);
2587			req->flags |= REQ_F_DONT_REISSUE;
2588		}
2589	}
2590
2591	WRITE_ONCE(req->result, res);
2592	/* order with io_iopoll_complete() checking ->result */
2593	smp_wmb();
2594	WRITE_ONCE(req->iopoll_completed, 1);
2595}
2596
2597/*
2598 * After the iocb has been issued, it's safe to be found on the poll list.
2599 * Adding the kiocb to the list AFTER submission ensures that we don't
2600 * find it from a io_do_iopoll() thread before the issuer is done
2601 * accessing the kiocb cookie.
2602 */
2603static void io_iopoll_req_issued(struct io_kiocb *req)
2604{
2605	struct io_ring_ctx *ctx = req->ctx;
2606	const bool in_async = io_wq_current_is_worker();
2607
2608	/* workqueue context doesn't hold uring_lock, grab it now */
2609	if (unlikely(in_async))
2610		mutex_lock(&ctx->uring_lock);
2611
2612	/*
2613	 * Track whether we have multiple files in our lists. This will impact
2614	 * how we do polling eventually, not spinning if we're on potentially
2615	 * different devices.
2616	 */
2617	if (list_empty(&ctx->iopoll_list)) {
2618		ctx->poll_multi_queue = false;
2619	} else if (!ctx->poll_multi_queue) {
2620		struct io_kiocb *list_req;
2621		unsigned int queue_num0, queue_num1;
2622
2623		list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2624						inflight_entry);
2625
2626		if (list_req->file != req->file) {
2627			ctx->poll_multi_queue = true;
2628		} else {
2629			queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2630			queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2631			if (queue_num0 != queue_num1)
2632				ctx->poll_multi_queue = true;
2633		}
2634	}
2635
2636	/*
2637	 * For fast devices, IO may have already completed. If it has, add
2638	 * it to the front so we find it first.
2639	 */
2640	if (READ_ONCE(req->iopoll_completed))
2641		list_add(&req->inflight_entry, &ctx->iopoll_list);
2642	else
2643		list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2644
2645	if (unlikely(in_async)) {
2646		/*
2647		 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2648		 * in sq thread task context or in io worker task context. If
2649		 * current task context is sq thread, we don't need to check
2650		 * whether should wake up sq thread.
2651		 */
2652		if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2653		    wq_has_sleeper(&ctx->sq_data->wait))
2654			wake_up(&ctx->sq_data->wait);
2655
2656		mutex_unlock(&ctx->uring_lock);
2657	}
2658}
2659
2660static bool io_bdev_nowait(struct block_device *bdev)
2661{
2662	return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2663}
2664
2665/*
2666 * If we tracked the file through the SCM inflight mechanism, we could support
2667 * any file. For now, just ensure that anything potentially problematic is done
2668 * inline.
2669 */
2670static bool __io_file_supports_nowait(struct file *file, int rw)
2671{
2672	umode_t mode = file_inode(file)->i_mode;
2673
2674	if (S_ISBLK(mode)) {
2675		if (IS_ENABLED(CONFIG_BLOCK) &&
2676		    io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2677			return true;
2678		return false;
2679	}
2680	if (S_ISSOCK(mode))
2681		return true;
2682	if (S_ISREG(mode)) {
2683		if (IS_ENABLED(CONFIG_BLOCK) &&
2684		    io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2685		    file->f_op != &io_uring_fops)
2686			return true;
2687		return false;
2688	}
2689
2690	/* any ->read/write should understand O_NONBLOCK */
2691	if (file->f_flags & O_NONBLOCK)
2692		return true;
2693
2694	if (!(file->f_mode & FMODE_NOWAIT))
2695		return false;
2696
2697	if (rw == READ)
2698		return file->f_op->read_iter != NULL;
2699
2700	return file->f_op->write_iter != NULL;
2701}
2702
2703static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2704{
2705	if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2706		return true;
2707	else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2708		return true;
2709
2710	return __io_file_supports_nowait(req->file, rw);
2711}
2712
2713static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2714{
2715	struct io_ring_ctx *ctx = req->ctx;
2716	struct kiocb *kiocb = &req->rw.kiocb;
2717	struct file *file = req->file;
2718	unsigned ioprio;
2719	int ret;
2720
2721	if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2722		req->flags |= REQ_F_ISREG;
2723
2724	kiocb->ki_pos = READ_ONCE(sqe->off);
2725	if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2726		req->flags |= REQ_F_CUR_POS;
2727		kiocb->ki_pos = file->f_pos;
2728	}
2729	kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2730	kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2731	ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2732	if (unlikely(ret))
2733		return ret;
2734
2735	/* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2736	if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2737		req->flags |= REQ_F_NOWAIT;
2738
2739	ioprio = READ_ONCE(sqe->ioprio);
2740	if (ioprio) {
2741		ret = ioprio_check_cap(ioprio);
2742		if (ret)
2743			return ret;
2744
2745		kiocb->ki_ioprio = ioprio;
2746	} else
2747		kiocb->ki_ioprio = get_current_ioprio();
2748
2749	if (ctx->flags & IORING_SETUP_IOPOLL) {
2750		if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2751		    !kiocb->ki_filp->f_op->iopoll)
2752			return -EOPNOTSUPP;
2753
2754		kiocb->ki_flags |= IOCB_HIPRI;
2755		kiocb->ki_complete = io_complete_rw_iopoll;
2756		req->iopoll_completed = 0;
2757	} else {
2758		if (kiocb->ki_flags & IOCB_HIPRI)
2759			return -EINVAL;
2760		kiocb->ki_complete = io_complete_rw;
2761	}
2762
2763	if (req->opcode == IORING_OP_READ_FIXED ||
2764	    req->opcode == IORING_OP_WRITE_FIXED) {
2765		req->imu = NULL;
2766		io_req_set_rsrc_node(req);
2767	}
2768
2769	req->rw.addr = READ_ONCE(sqe->addr);
2770	req->rw.len = READ_ONCE(sqe->len);
2771	req->buf_index = READ_ONCE(sqe->buf_index);
2772	return 0;
2773}
2774
2775static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2776{
2777	switch (ret) {
2778	case -EIOCBQUEUED:
2779		break;
2780	case -ERESTARTSYS:
2781	case -ERESTARTNOINTR:
2782	case -ERESTARTNOHAND:
2783	case -ERESTART_RESTARTBLOCK:
2784		/*
2785		 * We can't just restart the syscall, since previously
2786		 * submitted sqes may already be in progress. Just fail this
2787		 * IO with EINTR.
2788		 */
2789		ret = -EINTR;
2790		fallthrough;
2791	default:
2792		kiocb->ki_complete(kiocb, ret, 0);
2793	}
2794}
2795
2796static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2797		       unsigned int issue_flags)
2798{
2799	struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2800	struct io_async_rw *io = req->async_data;
2801	bool check_reissue = kiocb->ki_complete == io_complete_rw;
2802
2803	/* add previously done IO, if any */
2804	if (io && io->bytes_done > 0) {
2805		if (ret < 0)
2806			ret = io->bytes_done;
2807		else
2808			ret += io->bytes_done;
2809	}
2810
2811	if (req->flags & REQ_F_CUR_POS)
2812		req->file->f_pos = kiocb->ki_pos;
2813	if (ret >= 0 && check_reissue)
2814		__io_complete_rw(req, ret, 0, issue_flags);
2815	else
2816		io_rw_done(kiocb, ret);
2817
2818	if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2819		req->flags &= ~REQ_F_REISSUE;
2820		if (io_resubmit_prep(req)) {
2821			io_req_task_queue_reissue(req);
2822		} else {
2823			req_set_fail(req);
2824			__io_req_complete(req, issue_flags, ret,
2825					  io_put_rw_kbuf(req));
2826		}
2827	}
2828}
2829
2830static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2831			     struct io_mapped_ubuf *imu)
2832{
2833	size_t len = req->rw.len;
2834	u64 buf_end, buf_addr = req->rw.addr;
2835	size_t offset;
2836
2837	if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2838		return -EFAULT;
2839	/* not inside the mapped region */
2840	if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2841		return -EFAULT;
2842
2843	/*
2844	 * May not be a start of buffer, set size appropriately
2845	 * and advance us to the beginning.
2846	 */
2847	offset = buf_addr - imu->ubuf;
2848	iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2849
2850	if (offset) {
2851		/*
2852		 * Don't use iov_iter_advance() here, as it's really slow for
2853		 * using the latter parts of a big fixed buffer - it iterates
2854		 * over each segment manually. We can cheat a bit here, because
2855		 * we know that:
2856		 *
2857		 * 1) it's a BVEC iter, we set it up
2858		 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2859		 *    first and last bvec
2860		 *
2861		 * So just find our index, and adjust the iterator afterwards.
2862		 * If the offset is within the first bvec (or the whole first
2863		 * bvec, just use iov_iter_advance(). This makes it easier
2864		 * since we can just skip the first segment, which may not
2865		 * be PAGE_SIZE aligned.
2866		 */
2867		const struct bio_vec *bvec = imu->bvec;
2868
2869		if (offset <= bvec->bv_len) {
2870			iov_iter_advance(iter, offset);
2871		} else {
2872			unsigned long seg_skip;
2873
2874			/* skip first vec */
2875			offset -= bvec->bv_len;
2876			seg_skip = 1 + (offset >> PAGE_SHIFT);
2877
2878			iter->bvec = bvec + seg_skip;
2879			iter->nr_segs -= seg_skip;
2880			iter->count -= bvec->bv_len + offset;
2881			iter->iov_offset = offset & ~PAGE_MASK;
2882		}
2883	}
2884
2885	return 0;
2886}
2887
2888static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2889{
2890	struct io_ring_ctx *ctx = req->ctx;
2891	struct io_mapped_ubuf *imu = req->imu;
2892	u16 index, buf_index = req->buf_index;
2893
2894	if (likely(!imu)) {
2895		if (unlikely(buf_index >= ctx->nr_user_bufs))
2896			return -EFAULT;
2897		index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2898		imu = READ_ONCE(ctx->user_bufs[index]);
2899		req->imu = imu;
2900	}
2901	return __io_import_fixed(req, rw, iter, imu);
2902}
2903
2904static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2905{
2906	if (needs_lock)
2907		mutex_unlock(&ctx->uring_lock);
2908}
2909
2910static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2911{
2912	/*
2913	 * "Normal" inline submissions always hold the uring_lock, since we
2914	 * grab it from the system call. Same is true for the SQPOLL offload.
2915	 * The only exception is when we've detached the request and issue it
2916	 * from an async worker thread, grab the lock for that case.
2917	 */
2918	if (needs_lock)
2919		mutex_lock(&ctx->uring_lock);
2920}
2921
2922static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2923					  int bgid, struct io_buffer *kbuf,
2924					  bool needs_lock)
2925{
2926	struct io_buffer *head;
2927
2928	if (req->flags & REQ_F_BUFFER_SELECTED)
2929		return kbuf;
2930
2931	io_ring_submit_lock(req->ctx, needs_lock);
2932
2933	lockdep_assert_held(&req->ctx->uring_lock);
2934
2935	head = xa_load(&req->ctx->io_buffers, bgid);
2936	if (head) {
2937		if (!list_empty(&head->list)) {
2938			kbuf = list_last_entry(&head->list, struct io_buffer,
2939							list);
2940			list_del(&kbuf->list);
2941		} else {
2942			kbuf = head;
2943			xa_erase(&req->ctx->io_buffers, bgid);
2944		}
2945		if (*len > kbuf->len)
2946			*len = kbuf->len;
2947	} else {
2948		kbuf = ERR_PTR(-ENOBUFS);
2949	}
2950
2951	io_ring_submit_unlock(req->ctx, needs_lock);
2952
2953	return kbuf;
2954}
2955
2956static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2957					bool needs_lock)
2958{
2959	struct io_buffer *kbuf;
2960	u16 bgid;
2961
2962	kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2963	bgid = req->buf_index;
2964	kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2965	if (IS_ERR(kbuf))
2966		return kbuf;
2967	req->rw.addr = (u64) (unsigned long) kbuf;
2968	req->flags |= REQ_F_BUFFER_SELECTED;
2969	return u64_to_user_ptr(kbuf->addr);
2970}
2971
2972#ifdef CONFIG_COMPAT
2973static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2974				bool needs_lock)
2975{
2976	struct compat_iovec __user *uiov;
2977	compat_ssize_t clen;
2978	void __user *buf;
2979	ssize_t len;
2980
2981	uiov = u64_to_user_ptr(req->rw.addr);
2982	if (!access_ok(uiov, sizeof(*uiov)))
2983		return -EFAULT;
2984	if (__get_user(clen, &uiov->iov_len))
2985		return -EFAULT;
2986	if (clen < 0)
2987		return -EINVAL;
2988
2989	len = clen;
2990	buf = io_rw_buffer_select(req, &len, needs_lock);
2991	if (IS_ERR(buf))
2992		return PTR_ERR(buf);
2993	iov[0].iov_base = buf;
2994	iov[0].iov_len = (compat_size_t) len;
2995	return 0;
2996}
2997#endif
2998
2999static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3000				      bool needs_lock)
3001{
3002	struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3003	void __user *buf;
3004	ssize_t len;
3005
3006	if (copy_from_user(iov, uiov, sizeof(*uiov)))
3007		return -EFAULT;
3008
3009	len = iov[0].iov_len;
3010	if (len < 0)
3011		return -EINVAL;
3012	buf = io_rw_buffer_select(req, &len, needs_lock);
3013	if (IS_ERR(buf))
3014		return PTR_ERR(buf);
3015	iov[0].iov_base = buf;
3016	iov[0].iov_len = len;
3017	return 0;
3018}
3019
3020static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3021				    bool needs_lock)
3022{
3023	if (req->flags & REQ_F_BUFFER_SELECTED) {
3024		struct io_buffer *kbuf;
3025
3026		kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3027		iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3028		iov[0].iov_len = kbuf->len;
3029		return 0;
3030	}
3031	if (req->rw.len != 1)
3032		return -EINVAL;
3033
3034#ifdef CONFIG_COMPAT
3035	if (req->ctx->compat)
3036		return io_compat_import(req, iov, needs_lock);
3037#endif
3038
3039	return __io_iov_buffer_select(req, iov, needs_lock);
3040}
3041
3042static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3043			   struct iov_iter *iter, bool needs_lock)
3044{
3045	void __user *buf = u64_to_user_ptr(req->rw.addr);
3046	size_t sqe_len = req->rw.len;
3047	u8 opcode = req->opcode;
3048	ssize_t ret;
3049
3050	if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3051		*iovec = NULL;
3052		return io_import_fixed(req, rw, iter);
3053	}
3054
3055	/* buffer index only valid with fixed read/write, or buffer select  */
3056	if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3057		return -EINVAL;
3058
3059	if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3060		if (req->flags & REQ_F_BUFFER_SELECT) {
3061			buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3062			if (IS_ERR(buf))
3063				return PTR_ERR(buf);
3064			req->rw.len = sqe_len;
3065		}
3066
3067		ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3068		*iovec = NULL;
3069		return ret;
3070	}
3071
3072	if (req->flags & REQ_F_BUFFER_SELECT) {
3073		ret = io_iov_buffer_select(req, *iovec, needs_lock);
3074		if (!ret)
3075			iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3076		*iovec = NULL;
3077		return ret;
3078	}
3079
3080	return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3081			      req->ctx->compat);
3082}
3083
3084static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3085{
3086	return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3087}
3088
3089/*
3090 * For files that don't have ->read_iter() and ->write_iter(), handle them
3091 * by looping over ->read() or ->write() manually.
3092 */
3093static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3094{
3095	struct kiocb *kiocb = &req->rw.kiocb;
3096	struct file *file = req->file;
3097	ssize_t ret = 0;
3098
3099	/*
3100	 * Don't support polled IO through this interface, and we can't
3101	 * support non-blocking either. For the latter, this just causes
3102	 * the kiocb to be handled from an async context.
3103	 */
3104	if (kiocb->ki_flags & IOCB_HIPRI)
3105		return -EOPNOTSUPP;
3106	if (kiocb->ki_flags & IOCB_NOWAIT)
3107		return -EAGAIN;
3108
3109	while (iov_iter_count(iter)) {
3110		struct iovec iovec;
3111		ssize_t nr;
3112
3113		if (!iov_iter_is_bvec(iter)) {
3114			iovec = iov_iter_iovec(iter);
3115		} else {
3116			iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3117			iovec.iov_len = req->rw.len;
3118		}
3119
3120		if (rw == READ) {
3121			nr = file->f_op->read(file, iovec.iov_base,
3122					      iovec.iov_len, io_kiocb_ppos(kiocb));
3123		} else {
3124			nr = file->f_op->write(file, iovec.iov_base,
3125					       iovec.iov_len, io_kiocb_ppos(kiocb));
3126		}
3127
3128		if (nr < 0) {
3129			if (!ret)
3130				ret = nr;
3131			break;
3132		}
3133		ret += nr;
3134		if (nr != iovec.iov_len)
3135			break;
3136		req->rw.len -= nr;
3137		req->rw.addr += nr;
3138		iov_iter_advance(iter, nr);
3139	}
3140
3141	return ret;
3142}
3143
3144static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3145			  const struct iovec *fast_iov, struct iov_iter *iter)
3146{
3147	struct io_async_rw *rw = req->async_data;
3148
3149	memcpy(&rw->iter, iter, sizeof(*iter));
3150	rw->free_iovec = iovec;
3151	rw->bytes_done = 0;
3152	/* can only be fixed buffers, no need to do anything */
3153	if (iov_iter_is_bvec(iter))
3154		return;
3155	if (!iovec) {
3156		unsigned iov_off = 0;
3157
3158		rw->iter.iov = rw->fast_iov;
3159		if (iter->iov != fast_iov) {
3160			iov_off = iter->iov - fast_iov;
3161			rw->iter.iov += iov_off;
3162		}
3163		if (rw->fast_iov != fast_iov)
3164			memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3165			       sizeof(struct iovec) * iter->nr_segs);
3166	} else {
3167		req->flags |= REQ_F_NEED_CLEANUP;
3168	}
3169}
3170
3171static inline int io_alloc_async_data(struct io_kiocb *req)
3172{
3173	WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3174	req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3175	return req->async_data == NULL;
3176}
3177
3178static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3179			     const struct iovec *fast_iov,
3180			     struct iov_iter *iter, bool force)
3181{
3182	if (!force && !io_op_defs[req->opcode].needs_async_setup)
3183		return 0;
3184	if (!req->async_data) {
3185		if (io_alloc_async_data(req)) {
3186			kfree(iovec);
3187			return -ENOMEM;
3188		}
3189
3190		io_req_map_rw(req, iovec, fast_iov, iter);
3191	}
3192	return 0;
3193}
3194
3195static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3196{
3197	struct io_async_rw *iorw = req->async_data;
3198	struct iovec *iov = iorw->fast_iov;
3199	int ret;
3200
3201	ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3202	if (unlikely(ret < 0))
3203		return ret;
3204
3205	iorw->bytes_done = 0;
3206	iorw->free_iovec = iov;
3207	if (iov)
3208		req->flags |= REQ_F_NEED_CLEANUP;
3209	return 0;
3210}
3211
3212static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3213{
3214	if (unlikely(!(req->file->f_mode & FMODE_READ)))
3215		return -EBADF;
3216	return io_prep_rw(req, sqe);
3217}
3218
3219/*
3220 * This is our waitqueue callback handler, registered through lock_page_async()
3221 * when we initially tried to do the IO with the iocb armed our waitqueue.
3222 * This gets called when the page is unlocked, and we generally expect that to
3223 * happen when the page IO is completed and the page is now uptodate. This will
3224 * queue a task_work based retry of the operation, attempting to copy the data
3225 * again. If the latter fails because the page was NOT uptodate, then we will
3226 * do a thread based blocking retry of the operation. That's the unexpected
3227 * slow path.
3228 */
3229static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3230			     int sync, void *arg)
3231{
3232	struct wait_page_queue *wpq;
3233	struct io_kiocb *req = wait->private;
3234	struct wait_page_key *key = arg;
3235
3236	wpq = container_of(wait, struct wait_page_queue, wait);
3237
3238	if (!wake_page_match(wpq, key))
3239		return 0;
3240
3241	req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3242	list_del_init(&wait->entry);
3243	io_req_task_queue(req);
3244	return 1;
3245}
3246
3247/*
3248 * This controls whether a given IO request should be armed for async page
3249 * based retry. If we return false here, the request is handed to the async
3250 * worker threads for retry. If we're doing buffered reads on a regular file,
3251 * we prepare a private wait_page_queue entry and retry the operation. This
3252 * will either succeed because the page is now uptodate and unlocked, or it
3253 * will register a callback when the page is unlocked at IO completion. Through
3254 * that callback, io_uring uses task_work to setup a retry of the operation.
3255 * That retry will attempt the buffered read again. The retry will generally
3256 * succeed, or in rare cases where it fails, we then fall back to using the
3257 * async worker threads for a blocking retry.
3258 */
3259static bool io_rw_should_retry(struct io_kiocb *req)
3260{
3261	struct io_async_rw *rw = req->async_data;
3262	struct wait_page_queue *wait = &rw->wpq;
3263	struct kiocb *kiocb = &req->rw.kiocb;
3264
3265	/* never retry for NOWAIT, we just complete with -EAGAIN */
3266	if (req->flags & REQ_F_NOWAIT)
3267		return false;
3268
3269	/* Only for buffered IO */
3270	if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3271		return false;
3272
3273	/*
3274	 * just use poll if we can, and don't attempt if the fs doesn't
3275	 * support callback based unlocks
3276	 */
3277	if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3278		return false;
3279
3280	wait->wait.func = io_async_buf_func;
3281	wait->wait.private = req;
3282	wait->wait.flags = 0;
3283	INIT_LIST_HEAD(&wait->wait.entry);
3284	kiocb->ki_flags |= IOCB_WAITQ;
3285	kiocb->ki_flags &= ~IOCB_NOWAIT;
3286	kiocb->ki_waitq = wait;
3287	return true;
3288}
3289
3290static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3291{
3292	if (req->file->f_op->read_iter)
3293		return call_read_iter(req->file, &req->rw.kiocb, iter);
3294	else if (req->file->f_op->read)
3295		return loop_rw_iter(READ, req, iter);
3296	else
3297		return -EINVAL;
3298}
3299
3300static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3301{
3302	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3303	struct kiocb *kiocb = &req->rw.kiocb;
3304	struct iov_iter __iter, *iter = &__iter;
3305	struct io_async_rw *rw = req->async_data;
3306	ssize_t io_size, ret, ret2;
3307	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3308
3309	if (rw) {
3310		iter = &rw->iter;
3311		iovec = NULL;
3312	} else {
3313		ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3314		if (ret < 0)
3315			return ret;
3316	}
3317	io_size = iov_iter_count(iter);
3318	req->result = io_size;
3319
3320	/* Ensure we clear previously set non-block flag */
3321	if (!force_nonblock)
3322		kiocb->ki_flags &= ~IOCB_NOWAIT;
3323	else
3324		kiocb->ki_flags |= IOCB_NOWAIT;
3325
3326	/* If the file doesn't support async, just async punt */
3327	if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3328		ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3329		return ret ?: -EAGAIN;
3330	}
3331
3332	ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3333	if (unlikely(ret)) {
3334		kfree(iovec);
3335		return ret;
3336	}
3337
3338	ret = io_iter_do_read(req, iter);
3339
3340	if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3341		req->flags &= ~REQ_F_REISSUE;
3342		/* IOPOLL retry should happen for io-wq threads */
3343		if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3344			goto done;
3345		/* no retry on NONBLOCK nor RWF_NOWAIT */
3346		if (req->flags & REQ_F_NOWAIT)
3347			goto done;
3348		/* some cases will consume bytes even on error returns */
3349		iov_iter_revert(iter, io_size - iov_iter_count(iter));
3350		ret = 0;
3351	} else if (ret == -EIOCBQUEUED) {
3352		goto out_free;
3353	} else if (ret <= 0 || ret == io_size || !force_nonblock ||
3354		   (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3355		/* read all, failed, already did sync or don't want to retry */
3356		goto done;
3357	}
3358
3359	ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3360	if (ret2)
3361		return ret2;
3362
3363	iovec = NULL;
3364	rw = req->async_data;
3365	/* now use our persistent iterator, if we aren't already */
3366	iter = &rw->iter;
3367
3368	do {
3369		io_size -= ret;
3370		rw->bytes_done += ret;
3371		/* if we can retry, do so with the callbacks armed */
3372		if (!io_rw_should_retry(req)) {
3373			kiocb->ki_flags &= ~IOCB_WAITQ;
3374			return -EAGAIN;
3375		}
3376
3377		/*
3378		 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3379		 * we get -EIOCBQUEUED, then we'll get a notification when the
3380		 * desired page gets unlocked. We can also get a partial read
3381		 * here, and if we do, then just retry at the new offset.
3382		 */
3383		ret = io_iter_do_read(req, iter);
3384		if (ret == -EIOCBQUEUED)
3385			return 0;
3386		/* we got some bytes, but not all. retry. */
3387		kiocb->ki_flags &= ~IOCB_WAITQ;
3388	} while (ret > 0 && ret < io_size);
3389done:
3390	kiocb_done(kiocb, ret, issue_flags);
3391out_free:
3392	/* it's faster to check here then delegate to kfree */
3393	if (iovec)
3394		kfree(iovec);
3395	return 0;
3396}
3397
3398static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3399{
3400	if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3401		return -EBADF;
3402	return io_prep_rw(req, sqe);
3403}
3404
3405static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3406{
3407	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3408	struct kiocb *kiocb = &req->rw.kiocb;
3409	struct iov_iter __iter, *iter = &__iter;
3410	struct io_async_rw *rw = req->async_data;
3411	ssize_t ret, ret2, io_size;
3412	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3413
3414	if (rw) {
3415		iter = &rw->iter;
3416		iovec = NULL;
3417	} else {
3418		ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3419		if (ret < 0)
3420			return ret;
3421	}
3422	io_size = iov_iter_count(iter);
3423	req->result = io_size;
3424
3425	/* Ensure we clear previously set non-block flag */
3426	if (!force_nonblock)
3427		kiocb->ki_flags &= ~IOCB_NOWAIT;
3428	else
3429		kiocb->ki_flags |= IOCB_NOWAIT;
3430
3431	/* If the file doesn't support async, just async punt */
3432	if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3433		goto copy_iov;
3434
3435	/* file path doesn't support NOWAIT for non-direct_IO */
3436	if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3437	    (req->flags & REQ_F_ISREG))
3438		goto copy_iov;
3439
3440	ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3441	if (unlikely(ret))
3442		goto out_free;
3443
3444	/*
3445	 * Open-code file_start_write here to grab freeze protection,
3446	 * which will be released by another thread in
3447	 * io_complete_rw().  Fool lockdep by telling it the lock got
3448	 * released so that it doesn't complain about the held lock when
3449	 * we return to userspace.
3450	 */
3451	if (req->flags & REQ_F_ISREG) {
3452		sb_start_write(file_inode(req->file)->i_sb);
3453		__sb_writers_release(file_inode(req->file)->i_sb,
3454					SB_FREEZE_WRITE);
3455	}
3456	kiocb->ki_flags |= IOCB_WRITE;
3457
3458	if (req->file->f_op->write_iter)
3459		ret2 = call_write_iter(req->file, kiocb, iter);
3460	else if (req->file->f_op->write)
3461		ret2 = loop_rw_iter(WRITE, req, iter);
3462	else
3463		ret2 = -EINVAL;
3464
3465	if (req->flags & REQ_F_REISSUE) {
3466		req->flags &= ~REQ_F_REISSUE;
3467		ret2 = -EAGAIN;
3468	}
3469
3470	/*
3471	 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3472	 * retry them without IOCB_NOWAIT.
3473	 */
3474	if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3475		ret2 = -EAGAIN;
3476	/* no retry on NONBLOCK nor RWF_NOWAIT */
3477	if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3478		goto done;
3479	if (!force_nonblock || ret2 != -EAGAIN) {
3480		/* IOPOLL retry should happen for io-wq threads */
3481		if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3482			goto copy_iov;
3483done:
3484		kiocb_done(kiocb, ret2, issue_flags);
3485	} else {
3486copy_iov:
3487		/* some cases will consume bytes even on error returns */
3488		iov_iter_revert(iter, io_size - iov_iter_count(iter));
3489		ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3490		return ret ?: -EAGAIN;
3491	}
3492out_free:
3493	/* it's reportedly faster than delegating the null check to kfree() */
3494	if (iovec)
3495		kfree(iovec);
3496	return ret;
3497}
3498
3499static int io_renameat_prep(struct io_kiocb *req,
3500			    const struct io_uring_sqe *sqe)
3501{
3502	struct io_rename *ren = &req->rename;
3503	const char __user *oldf, *newf;
3504
3505	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3506		return -EINVAL;
3507	if (sqe->ioprio || sqe->buf_index)
3508		return -EINVAL;
3509	if (unlikely(req->flags & REQ_F_FIXED_FILE))
3510		return -EBADF;
3511
3512	ren->old_dfd = READ_ONCE(sqe->fd);
3513	oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3514	newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3515	ren->new_dfd = READ_ONCE(sqe->len);
3516	ren->flags = READ_ONCE(sqe->rename_flags);
3517
3518	ren->oldpath = getname(oldf);
3519	if (IS_ERR(ren->oldpath))
3520		return PTR_ERR(ren->oldpath);
3521
3522	ren->newpath = getname(newf);
3523	if (IS_ERR(ren->newpath)) {
3524		putname(ren->oldpath);
3525		return PTR_ERR(ren->newpath);
3526	}
3527
3528	req->flags |= REQ_F_NEED_CLEANUP;
3529	return 0;
3530}
3531
3532static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3533{
3534	struct io_rename *ren = &req->rename;
3535	int ret;
3536
3537	if (issue_flags & IO_URING_F_NONBLOCK)
3538		return -EAGAIN;
3539
3540	ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3541				ren->newpath, ren->flags);
3542
3543	req->flags &= ~REQ_F_NEED_CLEANUP;
3544	if (ret < 0)
3545		req_set_fail(req);
3546	io_req_complete(req, ret);
3547	return 0;
3548}
3549
3550static int io_unlinkat_prep(struct io_kiocb *req,
3551			    const struct io_uring_sqe *sqe)
3552{
3553	struct io_unlink *un = &req->unlink;
3554	const char __user *fname;
3555
3556	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3557		return -EINVAL;
3558	if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3559		return -EINVAL;
3560	if (unlikely(req->flags & REQ_F_FIXED_FILE))
3561		return -EBADF;
3562
3563	un->dfd = READ_ONCE(sqe->fd);
3564
3565	un->flags = READ_ONCE(sqe->unlink_flags);
3566	if (un->flags & ~AT_REMOVEDIR)
3567		return -EINVAL;
3568
3569	fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3570	un->filename = getname(fname);
3571	if (IS_ERR(un->filename))
3572		return PTR_ERR(un->filename);
3573
3574	req->flags |= REQ_F_NEED_CLEANUP;
3575	return 0;
3576}
3577
3578static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3579{
3580	struct io_unlink *un = &req->unlink;
3581	int ret;
3582
3583	if (issue_flags & IO_URING_F_NONBLOCK)
3584		return -EAGAIN;
3585
3586	if (un->flags & AT_REMOVEDIR)
3587		ret = do_rmdir(un->dfd, un->filename);
3588	else
3589		ret = do_unlinkat(un->dfd, un->filename);
3590
3591	req->flags &= ~REQ_F_NEED_CLEANUP;
3592	if (ret < 0)
3593		req_set_fail(req);
3594	io_req_complete(req, ret);
3595	return 0;
3596}
3597
3598static int io_shutdown_prep(struct io_kiocb *req,
3599			    const struct io_uring_sqe *sqe)
3600{
3601#if defined(CONFIG_NET)
3602	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3603		return -EINVAL;
3604	if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3605	    sqe->buf_index)
3606		return -EINVAL;
3607
3608	req->shutdown.how = READ_ONCE(sqe->len);
3609	return 0;
3610#else
3611	return -EOPNOTSUPP;
3612#endif
3613}
3614
3615static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3616{
3617#if defined(CONFIG_NET)
3618	struct socket *sock;
3619	int ret;
3620
3621	if (issue_flags & IO_URING_F_NONBLOCK)
3622		return -EAGAIN;
3623
3624	sock = sock_from_file(req->file);
3625	if (unlikely(!sock))
3626		return -ENOTSOCK;
3627
3628	ret = __sys_shutdown_sock(sock, req->shutdown.how);
3629	if (ret < 0)
3630		req_set_fail(req);
3631	io_req_complete(req, ret);
3632	return 0;
3633#else
3634	return -EOPNOTSUPP;
3635#endif
3636}
3637
3638static int __io_splice_prep(struct io_kiocb *req,
3639			    const struct io_uring_sqe *sqe)
3640{
3641	struct io_splice *sp = &req->splice;
3642	unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3643
3644	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3645		return -EINVAL;
3646
3647	sp->file_in = NULL;
3648	sp->len = READ_ONCE(sqe->len);
3649	sp->flags = READ_ONCE(sqe->splice_flags);
3650
3651	if (unlikely(sp->flags & ~valid_flags))
3652		return -EINVAL;
3653
3654	sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3655				  (sp->flags & SPLICE_F_FD_IN_FIXED));
3656	if (!sp->file_in)
3657		return -EBADF;
3658	req->flags |= REQ_F_NEED_CLEANUP;
3659	return 0;
3660}
3661
3662static int io_tee_prep(struct io_kiocb *req,
3663		       const struct io_uring_sqe *sqe)
3664{
3665	if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3666		return -EINVAL;
3667	return __io_splice_prep(req, sqe);
3668}
3669
3670static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3671{
3672	struct io_splice *sp = &req->splice;
3673	struct file *in = sp->file_in;
3674	struct file *out = sp->file_out;
3675	unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3676	long ret = 0;
3677
3678	if (issue_flags & IO_URING_F_NONBLOCK)
3679		return -EAGAIN;
3680	if (sp->len)
3681		ret = do_tee(in, out, sp->len, flags);
3682
3683	if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3684		io_put_file(in);
3685	req->flags &= ~REQ_F_NEED_CLEANUP;
3686
3687	if (ret != sp->len)
3688		req_set_fail(req);
3689	io_req_complete(req, ret);
3690	return 0;
3691}
3692
3693static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3694{
3695	struct io_splice *sp = &req->splice;
3696
3697	sp->off_in = READ_ONCE(sqe->splice_off_in);
3698	sp->off_out = READ_ONCE(sqe->off);
3699	return __io_splice_prep(req, sqe);
3700}
3701
3702static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3703{
3704	struct io_splice *sp = &req->splice;
3705	struct file *in = sp->file_in;
3706	struct file *out = sp->file_out;
3707	unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3708	loff_t *poff_in, *poff_out;
3709	long ret = 0;
3710
3711	if (issue_flags & IO_URING_F_NONBLOCK)
3712		return -EAGAIN;
3713
3714	poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3715	poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3716
3717	if (sp->len)
3718		ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3719
3720	if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3721		io_put_file(in);
3722	req->flags &= ~REQ_F_NEED_CLEANUP;
3723
3724	if (ret != sp->len)
3725		req_set_fail(req);
3726	io_req_complete(req, ret);
3727	return 0;
3728}
3729
3730/*
3731 * IORING_OP_NOP just posts a completion event, nothing else.
3732 */
3733static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3734{
3735	struct io_ring_ctx *ctx = req->ctx;
3736
3737	if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3738		return -EINVAL;
3739
3740	__io_req_complete(req, issue_flags, 0, 0);
3741	return 0;
3742}
3743
3744static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3745{
3746	struct io_ring_ctx *ctx = req->ctx;
3747
3748	if (!req->file)
3749		return -EBADF;
3750
3751	if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3752		return -EINVAL;
3753	if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3754		return -EINVAL;
3755
3756	req->sync.flags = READ_ONCE(sqe->fsync_flags);
3757	if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3758		return -EINVAL;
3759
3760	req->sync.off = READ_ONCE(sqe->off);
3761	req->sync.len = READ_ONCE(sqe->len);
3762	return 0;
3763}
3764
3765static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3766{
3767	loff_t end = req->sync.off + req->sync.len;
3768	int ret;
3769
3770	/* fsync always requires a blocking context */
3771	if (issue_flags & IO_URING_F_NONBLOCK)
3772		return -EAGAIN;
3773
3774	ret = vfs_fsync_range(req->file, req->sync.off,
3775				end > 0 ? end : LLONG_MAX,
3776				req->sync.flags & IORING_FSYNC_DATASYNC);
3777	if (ret < 0)
3778		req_set_fail(req);
3779	io_req_complete(req, ret);
3780	return 0;
3781}
3782
3783static int io_fallocate_prep(struct io_kiocb *req,
3784			     const struct io_uring_sqe *sqe)
3785{
3786	if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3787		return -EINVAL;
3788	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3789		return -EINVAL;
3790
3791	req->sync.off = READ_ONCE(sqe->off);
3792	req->sync.len = READ_ONCE(sqe->addr);
3793	req->sync.mode = READ_ONCE(sqe->len);
3794	return 0;
3795}
3796
3797static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3798{
3799	int ret;
3800
3801	/* fallocate always requiring blocking context */
3802	if (issue_flags & IO_URING_F_NONBLOCK)
3803		return -EAGAIN;
3804	ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3805				req->sync.len);
3806	if (ret < 0)
3807		req_set_fail(req);
3808	io_req_complete(req, ret);
3809	return 0;
3810}
3811
3812static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3813{
3814	const char __user *fname;
3815	int ret;
3816
3817	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3818		return -EINVAL;
3819	if (unlikely(sqe->ioprio || sqe->buf_index))
3820		return -EINVAL;
3821	if (unlikely(req->flags & REQ_F_FIXED_FILE))
3822		return -EBADF;
3823
3824	/* open.how should be already initialised */
3825	if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3826		req->open.how.flags |= O_LARGEFILE;
3827
3828	req->open.dfd = READ_ONCE(sqe->fd);
3829	fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3830	req->open.filename = getname(fname);
3831	if (IS_ERR(req->open.filename)) {
3832		ret = PTR_ERR(req->open.filename);
3833		req->open.filename = NULL;
3834		return ret;
3835	}
3836	req->open.nofile = rlimit(RLIMIT_NOFILE);
3837	req->flags |= REQ_F_NEED_CLEANUP;
3838	return 0;
3839}
3840
3841static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3842{
3843	u64 mode = READ_ONCE(sqe->len);
3844	u64 flags = READ_ONCE(sqe->open_flags);
3845
3846	req->open.how = build_open_how(flags, mode);
3847	return __io_openat_prep(req, sqe);
3848}
3849
3850static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3851{
3852	struct open_how __user *how;
3853	size_t len;
3854	int ret;
3855
3856	how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3857	len = READ_ONCE(sqe->len);
3858	if (len < OPEN_HOW_SIZE_VER0)
3859		return -EINVAL;
3860
3861	ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3862					len);
3863	if (ret)
3864		return ret;
3865
3866	return __io_openat_prep(req, sqe);
3867}
3868
3869static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3870{
3871	struct open_flags op;
3872	struct file *file;
3873	bool nonblock_set;
3874	bool resolve_nonblock;
3875	int ret;
3876
3877	ret = build_open_flags(&req->open.how, &op);
3878	if (ret)
3879		goto err;
3880	nonblock_set = op.open_flag & O_NONBLOCK;
3881	resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3882	if (issue_flags & IO_URING_F_NONBLOCK) {
3883		/*
3884		 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3885		 * it'll always -EAGAIN
3886		 */
3887		if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3888			return -EAGAIN;
3889		op.lookup_flags |= LOOKUP_CACHED;
3890		op.open_flag |= O_NONBLOCK;
3891	}
3892
3893	ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3894	if (ret < 0)
3895		goto err;
3896
3897	file = do_filp_open(req->open.dfd, req->open.filename, &op);
3898	if (IS_ERR(file)) {
3899		/*
3900		 * We could hang on to this 'fd' on retrying, but seems like
3901		 * marginal gain for something that is now known to be a slower
3902		 * path. So just put it, and we'll get a new one when we retry.
3903		 */
3904		put_unused_fd(ret);
3905
3906		ret = PTR_ERR(file);
3907		/* only retry if RESOLVE_CACHED wasn't already set by application */
3908		if (ret == -EAGAIN &&
3909		    (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3910			return -EAGAIN;
3911		goto err;
3912	}
3913
3914	if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3915		file->f_flags &= ~O_NONBLOCK;
3916	fsnotify_open(file);
3917	fd_install(ret, file);
3918err:
3919	putname(req->open.filename);
3920	req->flags &= ~REQ_F_NEED_CLEANUP;
3921	if (ret < 0)
3922		req_set_fail(req);
3923	__io_req_complete(req, issue_flags, ret, 0);
3924	return 0;
3925}
3926
3927static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3928{
3929	return io_openat2(req, issue_flags);
3930}
3931
3932static int io_remove_buffers_prep(struct io_kiocb *req,
3933				  const struct io_uring_sqe *sqe)
3934{
3935	struct io_provide_buf *p = &req->pbuf;
3936	u64 tmp;
3937
3938	if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3939		return -EINVAL;
3940
3941	tmp = READ_ONCE(sqe->fd);
3942	if (!tmp || tmp > USHRT_MAX)
3943		return -EINVAL;
3944
3945	memset(p, 0, sizeof(*p));
3946	p->nbufs = tmp;
3947	p->bgid = READ_ONCE(sqe->buf_group);
3948	return 0;
3949}
3950
3951static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3952			       int bgid, unsigned nbufs)
3953{
3954	unsigned i = 0;
3955
3956	/* shouldn't happen */
3957	if (!nbufs)
3958		return 0;
3959
3960	/* the head kbuf is the list itself */
3961	while (!list_empty(&buf->list)) {
3962		struct io_buffer *nxt;
3963
3964		nxt = list_first_entry(&buf->list, struct io_buffer, list);
3965		list_del(&nxt->list);
3966		kfree(nxt);
3967		if (++i == nbufs)
3968			return i;
3969	}
3970	i++;
3971	kfree(buf);
3972	xa_erase(&ctx->io_buffers, bgid);
3973
3974	return i;
3975}
3976
3977static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3978{
3979	struct io_provide_buf *p = &req->pbuf;
3980	struct io_ring_ctx *ctx = req->ctx;
3981	struct io_buffer *head;
3982	int ret = 0;
3983	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3984
3985	io_ring_submit_lock(ctx, !force_nonblock);
3986
3987	lockdep_assert_held(&ctx->uring_lock);
3988
3989	ret = -ENOENT;
3990	head = xa_load(&ctx->io_buffers, p->bgid);
3991	if (head)
3992		ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3993	if (ret < 0)
3994		req_set_fail(req);
3995
3996	/* complete before unlock, IOPOLL may need the lock */
3997	__io_req_complete(req, issue_flags, ret, 0);
3998	io_ring_submit_unlock(ctx, !force_nonblock);
3999	return 0;
4000}
4001
4002static int io_provide_buffers_prep(struct io_kiocb *req,
4003				   const struct io_uring_sqe *sqe)
4004{
4005	unsigned long size, tmp_check;
4006	struct io_provide_buf *p = &req->pbuf;
4007	u64 tmp;
4008
4009	if (sqe->ioprio || sqe->rw_flags)
4010		return -EINVAL;
4011
4012	tmp = READ_ONCE(sqe->fd);
4013	if (!tmp || tmp > USHRT_MAX)
4014		return -E2BIG;
4015	p->nbufs = tmp;
4016	p->addr = READ_ONCE(sqe->addr);
4017	p->len = READ_ONCE(sqe->len);
4018
4019	if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4020				&size))
4021		return -EOVERFLOW;
4022	if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4023		return -EOVERFLOW;
4024
4025	size = (unsigned long)p->len * p->nbufs;
4026	if (!access_ok(u64_to_user_ptr(p->addr), size))
4027		return -EFAULT;
4028
4029	p->bgid = READ_ONCE(sqe->buf_group);
4030	tmp = READ_ONCE(sqe->off);
4031	if (tmp > USHRT_MAX)
4032		return -E2BIG;
4033	p->bid = tmp;
4034	return 0;
4035}
4036
4037static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4038{
4039	struct io_buffer *buf;
4040	u64 addr = pbuf->addr;
4041	int i, bid = pbuf->bid;
4042
4043	for (i = 0; i < pbuf->nbufs; i++) {
4044		buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4045		if (!buf)
4046			break;
4047
4048		buf->addr = addr;
4049		buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4050		buf->bid = bid;
4051		addr += pbuf->len;
4052		bid++;
4053		if (!*head) {
4054			INIT_LIST_HEAD(&buf->list);
4055			*head = buf;
4056		} else {
4057			list_add_tail(&buf->list, &(*head)->list);
4058		}
4059	}
4060
4061	return i ? i : -ENOMEM;
4062}
4063
4064static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4065{
4066	struct io_provide_buf *p = &req->pbuf;
4067	struct io_ring_ctx *ctx = req->ctx;
4068	struct io_buffer *head, *list;
4069	int ret = 0;
4070	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4071
4072	io_ring_submit_lock(ctx, !force_nonblock);
4073
4074	lockdep_assert_held(&ctx->uring_lock);
4075
4076	list = head = xa_load(&ctx->io_buffers, p->bgid);
4077
4078	ret = io_add_buffers(p, &head);
4079	if (ret >= 0 && !list) {
4080		ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4081		if (ret < 0)
4082			__io_remove_buffers(ctx, head, p->bgid, -1U);
4083	}
4084	if (ret < 0)
4085		req_set_fail(req);
4086	/* complete before unlock, IOPOLL may need the lock */
4087	__io_req_complete(req, issue_flags, ret, 0);
4088	io_ring_submit_unlock(ctx, !force_nonblock);
4089	return 0;
4090}
4091
4092static int io_epoll_ctl_prep(struct io_kiocb *req,
4093			     const struct io_uring_sqe *sqe)
4094{
4095#if defined(CONFIG_EPOLL)
4096	if (sqe->ioprio || sqe->buf_index)
4097		return -EINVAL;
4098	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4099		return -EINVAL;
4100
4101	req->epoll.epfd = READ_ONCE(sqe->fd);
4102	req->epoll.op = READ_ONCE(sqe->len);
4103	req->epoll.fd = READ_ONCE(sqe->off);
4104
4105	if (ep_op_has_event(req->epoll.op)) {
4106		struct epoll_event __user *ev;
4107
4108		ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4109		if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4110			return -EFAULT;
4111	}
4112
4113	return 0;
4114#else
4115	return -EOPNOTSUPP;
4116#endif
4117}
4118
4119static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4120{
4121#if defined(CONFIG_EPOLL)
4122	struct io_epoll *ie = &req->epoll;
4123	int ret;
4124	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4125
4126	ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4127	if (force_nonblock && ret == -EAGAIN)
4128		return -EAGAIN;
4129
4130	if (ret < 0)
4131		req_set_fail(req);
4132	__io_req_complete(req, issue_flags, ret, 0);
4133	return 0;
4134#else
4135	return -EOPNOTSUPP;
4136#endif
4137}
4138
4139static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4140{
4141#if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4142	if (sqe->ioprio || sqe->buf_index || sqe->off)
4143		return -EINVAL;
4144	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4145		return -EINVAL;
4146
4147	req->madvise.addr = READ_ONCE(sqe->addr);
4148	req->madvise.len = READ_ONCE(sqe->len);
4149	req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4150	return 0;
4151#else
4152	return -EOPNOTSUPP;
4153#endif
4154}
4155
4156static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4157{
4158#if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4159	struct io_madvise *ma = &req->madvise;
4160	int ret;
4161
4162	if (issue_flags & IO_URING_F_NONBLOCK)
4163		return -EAGAIN;
4164
4165	ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4166	if (ret < 0)
4167		req_set_fail(req);
4168	io_req_complete(req, ret);
4169	return 0;
4170#else
4171	return -EOPNOTSUPP;
4172#endif
4173}
4174
4175static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4176{
4177	if (sqe->ioprio || sqe->buf_index || sqe->addr)
4178		return -EINVAL;
4179	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4180		return -EINVAL;
4181
4182	req->fadvise.offset = READ_ONCE(sqe->off);
4183	req->fadvise.len = READ_ONCE(sqe->len);
4184	req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4185	return 0;
4186}
4187
4188static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4189{
4190	struct io_fadvise *fa = &req->fadvise;
4191	int ret;
4192
4193	if (issue_flags & IO_URING_F_NONBLOCK) {
4194		switch (fa->advice) {
4195		case POSIX_FADV_NORMAL:
4196		case POSIX_FADV_RANDOM:
4197		case POSIX_FADV_SEQUENTIAL:
4198			break;
4199		default:
4200			return -EAGAIN;
4201		}
4202	}
4203
4204	ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4205	if (ret < 0)
4206		req_set_fail(req);
4207	__io_req_complete(req, issue_flags, ret, 0);
4208	return 0;
4209}
4210
4211static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4212{
4213	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4214		return -EINVAL;
4215	if (sqe->ioprio || sqe->buf_index)
4216		return -EINVAL;
4217	if (req->flags & REQ_F_FIXED_FILE)
4218		return -EBADF;
4219
4220	req->statx.dfd = READ_ONCE(sqe->fd);
4221	req->statx.mask = READ_ONCE(sqe->len);
4222	req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4223	req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4224	req->statx.flags = READ_ONCE(sqe->statx_flags);
4225
4226	return 0;
4227}
4228
4229static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4230{
4231	struct io_statx *ctx = &req->statx;
4232	int ret;
4233
4234	if (issue_flags & IO_URING_F_NONBLOCK)
4235		return -EAGAIN;
4236
4237	ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4238		       ctx->buffer);
4239
4240	if (ret < 0)
4241		req_set_fail(req);
4242	io_req_complete(req, ret);
4243	return 0;
4244}
4245
4246static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4247{
4248	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4249		return -EINVAL;
4250	if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4251	    sqe->rw_flags || sqe->buf_index)
4252		return -EINVAL;
4253	if (req->flags & REQ_F_FIXED_FILE)
4254		return -EBADF;
4255
4256	req->close.fd = READ_ONCE(sqe->fd);
4257	return 0;
4258}
4259
4260static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4261{
4262	struct files_struct *files = current->files;
4263	struct io_close *close = &req->close;
4264	struct fdtable *fdt;
4265	struct file *file = NULL;
4266	int ret = -EBADF;
4267
4268	spin_lock(&files->file_lock);
4269	fdt = files_fdtable(files);
4270	if (close->fd >= fdt->max_fds) {
4271		spin_unlock(&files->file_lock);
4272		goto err;
4273	}
4274	file = fdt->fd[close->fd];
4275	if (!file || file->f_op == &io_uring_fops) {
4276		spin_unlock(&files->file_lock);
4277		file = NULL;
4278		goto err;
4279	}
4280
4281	/* if the file has a flush method, be safe and punt to async */
4282	if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4283		spin_unlock(&files->file_lock);
4284		return -EAGAIN;
4285	}
4286
4287	ret = __close_fd_get_file(close->fd, &file);
4288	spin_unlock(&files->file_lock);
4289	if (ret < 0) {
4290		if (ret == -ENOENT)
4291			ret = -EBADF;
4292		goto err;
4293	}
4294
4295	/* No ->flush() or already async, safely close from here */
4296	ret = filp_close(file, current->files);
4297err:
4298	if (ret < 0)
4299		req_set_fail(req);
4300	if (file)
4301		fput(file);
4302	__io_req_complete(req, issue_flags, ret, 0);
4303	return 0;
4304}
4305
4306static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4307{
4308	struct io_ring_ctx *ctx = req->ctx;
4309
4310	if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4311		return -EINVAL;
4312	if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4313		return -EINVAL;
4314
4315	req->sync.off = READ_ONCE(sqe->off);
4316	req->sync.len = READ_ONCE(sqe->len);
4317	req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4318	return 0;
4319}
4320
4321static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4322{
4323	int ret;
4324
4325	/* sync_file_range always requires a blocking context */
4326	if (issue_flags & IO_URING_F_NONBLOCK)
4327		return -EAGAIN;
4328
4329	ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4330				req->sync.flags);
4331	if (ret < 0)
4332		req_set_fail(req);
4333	io_req_complete(req, ret);
4334	return 0;
4335}
4336
4337#if defined(CONFIG_NET)
4338static int io_setup_async_msg(struct io_kiocb *req,
4339			      struct io_async_msghdr *kmsg)
4340{
4341	struct io_async_msghdr *async_msg = req->async_data;
4342
4343	if (async_msg)
4344		return -EAGAIN;
4345	if (io_alloc_async_data(req)) {
4346		kfree(kmsg->free_iov);
4347		return -ENOMEM;
4348	}
4349	async_msg = req->async_data;
4350	req->flags |= REQ_F_NEED_CLEANUP;
4351	memcpy(async_msg, kmsg, sizeof(*kmsg));
4352	async_msg->msg.msg_name = &async_msg->addr;
4353	/* if were using fast_iov, set it to the new one */
4354	if (!async_msg->free_iov)
4355		async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4356
4357	return -EAGAIN;
4358}
4359
4360static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4361			       struct io_async_msghdr *iomsg)
4362{
4363	iomsg->msg.msg_name = &iomsg->addr;
4364	iomsg->free_iov = iomsg->fast_iov;
4365	return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4366				   req->sr_msg.msg_flags, &iomsg->free_iov);
4367}
4368
4369static int io_sendmsg_prep_async(struct io_kiocb *req)
4370{
4371	int ret;
4372
4373	ret = io_sendmsg_copy_hdr(req, req->async_data);
4374	if (!ret)
4375		req->flags |= REQ_F_NEED_CLEANUP;
4376	return ret;
4377}
4378
4379static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4380{
4381	struct io_sr_msg *sr = &req->sr_msg;
4382
4383	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4384		return -EINVAL;
4385
4386	sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4387	sr->len = READ_ONCE(sqe->len);
4388	sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4389	if (sr->msg_flags & MSG_DONTWAIT)
4390		req->flags |= REQ_F_NOWAIT;
4391
4392#ifdef CONFIG_COMPAT
4393	if (req->ctx->compat)
4394		sr->msg_flags |= MSG_CMSG_COMPAT;
4395#endif
4396	return 0;
4397}
4398
4399static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4400{
4401	struct io_async_msghdr iomsg, *kmsg;
4402	struct socket *sock;
4403	unsigned flags;
4404	int min_ret = 0;
4405	int ret;
4406
4407	sock = sock_from_file(req->file);
4408	if (unlikely(!sock))
4409		return -ENOTSOCK;
4410
4411	kmsg = req->async_data;
4412	if (!kmsg) {
4413		ret = io_sendmsg_copy_hdr(req, &iomsg);
4414		if (ret)
4415			return ret;
4416		kmsg = &iomsg;
4417	}
4418
4419	flags = req->sr_msg.msg_flags;
4420	if (issue_flags & IO_URING_F_NONBLOCK)
4421		flags |= MSG_DONTWAIT;
4422	if (flags & MSG_WAITALL)
4423		min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4424
4425	ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4426	if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4427		return io_setup_async_msg(req, kmsg);
4428	if (ret == -ERESTARTSYS)
4429		ret = -EINTR;
4430
4431	/* fast path, check for non-NULL to avoid function call */
4432	if (kmsg->free_iov)
4433		kfree(kmsg->free_iov);
4434	req->flags &= ~REQ_F_NEED_CLEANUP;
4435	if (ret < min_ret)
4436		req_set_fail(req);
4437	__io_req_complete(req, issue_flags, ret, 0);
4438	return 0;
4439}
4440
4441static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4442{
4443	struct io_sr_msg *sr = &req->sr_msg;
4444	struct msghdr msg;
4445	struct iovec iov;
4446	struct socket *sock;
4447	unsigned flags;
4448	int min_ret = 0;
4449	int ret;
4450
4451	sock = sock_from_file(req->file);
4452	if (unlikely(!sock))
4453		return -ENOTSOCK;
4454
4455	ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4456	if (unlikely(ret))
4457		return ret;
4458
4459	msg.msg_name = NULL;
4460	msg.msg_control = NULL;
4461	msg.msg_controllen = 0;
4462	msg.msg_namelen = 0;
4463
4464	flags = req->sr_msg.msg_flags;
4465	if (issue_flags & IO_URING_F_NONBLOCK)
4466		flags |= MSG_DONTWAIT;
4467	if (flags & MSG_WAITALL)
4468		min_ret = iov_iter_count(&msg.msg_iter);
4469
4470	msg.msg_flags = flags;
4471	ret = sock_sendmsg(sock, &msg);
4472	if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4473		return -EAGAIN;
4474	if (ret == -ERESTARTSYS)
4475		ret = -EINTR;
4476
4477	if (ret < min_ret)
4478		req_set_fail(req);
4479	__io_req_complete(req, issue_flags, ret, 0);
4480	return 0;
4481}
4482
4483static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4484				 struct io_async_msghdr *iomsg)
4485{
4486	struct io_sr_msg *sr = &req->sr_msg;
4487	struct iovec __user *uiov;
4488	size_t iov_len;
4489	int ret;
4490
4491	ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4492					&iomsg->uaddr, &uiov, &iov_len);
4493	if (ret)
4494		return ret;
4495
4496	if (req->flags & REQ_F_BUFFER_SELECT) {
4497		if (iov_len > 1)
4498			return -EINVAL;
4499		if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4500			return -EFAULT;
4501		sr->len = iomsg->fast_iov[0].iov_len;
4502		iomsg->free_iov = NULL;
4503	} else {
4504		iomsg->free_iov = iomsg->fast_iov;
4505		ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4506				     &iomsg->free_iov, &iomsg->msg.msg_iter,
4507				     false);
4508		if (ret > 0)
4509			ret = 0;
4510	}
4511
4512	return ret;
4513}
4514
4515#ifdef CONFIG_COMPAT
4516static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4517					struct io_async_msghdr *iomsg)
4518{
4519	struct io_sr_msg *sr = &req->sr_msg;
4520	struct compat_iovec __user *uiov;
4521	compat_uptr_t ptr;
4522	compat_size_t len;
4523	int ret;
4524
4525	ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4526				  &ptr, &len);
4527	if (ret)
4528		return ret;
4529
4530	uiov = compat_ptr(ptr);
4531	if (req->flags & REQ_F_BUFFER_SELECT) {
4532		compat_ssize_t clen;
4533
4534		if (len > 1)
4535			return -EINVAL;
4536		if (!access_ok(uiov, sizeof(*uiov)))
4537			return -EFAULT;
4538		if (__get_user(clen, &uiov->iov_len))
4539			return -EFAULT;
4540		if (clen < 0)
4541			return -EINVAL;
4542		sr->len = clen;
4543		iomsg->free_iov = NULL;
4544	} else {
4545		iomsg->free_iov = iomsg->fast_iov;
4546		ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4547				   UIO_FASTIOV, &iomsg->free_iov,
4548				   &iomsg->msg.msg_iter, true);
4549		if (ret < 0)
4550			return ret;
4551	}
4552
4553	return 0;
4554}
4555#endif
4556
4557static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4558			       struct io_async_msghdr *iomsg)
4559{
4560	iomsg->msg.msg_name = &iomsg->addr;
4561
4562#ifdef CONFIG_COMPAT
4563	if (req->ctx->compat)
4564		return __io_compat_recvmsg_copy_hdr(req, iomsg);
4565#endif
4566
4567	return __io_recvmsg_copy_hdr(req, iomsg);
4568}
4569
4570static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4571					       bool needs_lock)
4572{
4573	struct io_sr_msg *sr = &req->sr_msg;
4574	struct io_buffer *kbuf;
4575
4576	kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4577	if (IS_ERR(kbuf))
4578		return kbuf;
4579
4580	sr->kbuf = kbuf;
4581	req->flags |= REQ_F_BUFFER_SELECTED;
4582	return kbuf;
4583}
4584
4585static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4586{
4587	return io_put_kbuf(req, req->sr_msg.kbuf);
4588}
4589
4590static int io_recvmsg_prep_async(struct io_kiocb *req)
4591{
4592	int ret;
4593
4594	ret = io_recvmsg_copy_hdr(req, req->async_data);
4595	if (!ret)
4596		req->flags |= REQ_F_NEED_CLEANUP;
4597	return ret;
4598}
4599
4600static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4601{
4602	struct io_sr_msg *sr = &req->sr_msg;
4603
4604	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4605		return -EINVAL;
4606
4607	sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4608	sr->len = READ_ONCE(sqe->len);
4609	sr->bgid = READ_ONCE(sqe->buf_group);
4610	sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4611	if (sr->msg_flags & MSG_DONTWAIT)
4612		req->flags |= REQ_F_NOWAIT;
4613
4614#ifdef CONFIG_COMPAT
4615	if (req->ctx->compat)
4616		sr->msg_flags |= MSG_CMSG_COMPAT;
4617#endif
4618	return 0;
4619}
4620
4621static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4622{
4623	struct io_async_msghdr iomsg, *kmsg;
4624	struct socket *sock;
4625	struct io_buffer *kbuf;
4626	unsigned flags;
4627	int min_ret = 0;
4628	int ret, cflags = 0;
4629	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4630
4631	sock = sock_from_file(req->file);
4632	if (unlikely(!sock))
4633		return -ENOTSOCK;
4634
4635	kmsg = req->async_data;
4636	if (!kmsg) {
4637		ret = io_recvmsg_copy_hdr(req, &iomsg);
4638		if (ret)
4639			return ret;
4640		kmsg = &iomsg;
4641	}
4642
4643	if (req->flags & REQ_F_BUFFER_SELECT) {
4644		kbuf = io_recv_buffer_select(req, !force_nonblock);
4645		if (IS_ERR(kbuf))
4646			return PTR_ERR(kbuf);
4647		kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4648		kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4649		iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4650				1, req->sr_msg.len);
4651	}
4652
4653	flags = req->sr_msg.msg_flags;
4654	if (force_nonblock)
4655		flags |= MSG_DONTWAIT;
4656	if (flags & MSG_WAITALL)
4657		min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4658
4659	ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4660					kmsg->uaddr, flags);
4661	if (force_nonblock && ret == -EAGAIN)
4662		return io_setup_async_msg(req, kmsg);
4663	if (ret == -ERESTARTSYS)
4664		ret = -EINTR;
4665
4666	if (req->flags & REQ_F_BUFFER_SELECTED)
4667		cflags = io_put_recv_kbuf(req);
4668	/* fast path, check for non-NULL to avoid function call */
4669	if (kmsg->free_iov)
4670		kfree(kmsg->free_iov);
4671	req->flags &= ~REQ_F_NEED_CLEANUP;
4672	if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4673		req_set_fail(req);
4674	__io_req_complete(req, issue_flags, ret, cflags);
4675	return 0;
4676}
4677
4678static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4679{
4680	struct io_buffer *kbuf;
4681	struct io_sr_msg *sr = &req->sr_msg;
4682	struct msghdr msg;
4683	void __user *buf = sr->buf;
4684	struct socket *sock;
4685	struct iovec iov;
4686	unsigned flags;
4687	int min_ret = 0;
4688	int ret, cflags = 0;
4689	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4690
4691	sock = sock_from_file(req->file);
4692	if (unlikely(!sock))
4693		return -ENOTSOCK;
4694
4695	if (req->flags & REQ_F_BUFFER_SELECT) {
4696		kbuf = io_recv_buffer_select(req, !force_nonblock);
4697		if (IS_ERR(kbuf))
4698			return PTR_ERR(kbuf);
4699		buf = u64_to_user_ptr(kbuf->addr);
4700	}
4701
4702	ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4703	if (unlikely(ret))
4704		goto out_free;
4705
4706	msg.msg_name = NULL;
4707	msg.msg_control = NULL;
4708	msg.msg_controllen = 0;
4709	msg.msg_namelen = 0;
4710	msg.msg_iocb = NULL;
4711	msg.msg_flags = 0;
4712
4713	flags = req->sr_msg.msg_flags;
4714	if (force_nonblock)
4715		flags |= MSG_DONTWAIT;
4716	if (flags & MSG_WAITALL)
4717		min_ret = iov_iter_count(&msg.msg_iter);
4718
4719	ret = sock_recvmsg(sock, &msg, flags);
4720	if (force_nonblock && ret == -EAGAIN)
4721		return -EAGAIN;
4722	if (ret == -ERESTARTSYS)
4723		ret = -EINTR;
4724out_free:
4725	if (req->flags & REQ_F_BUFFER_SELECTED)
4726		cflags = io_put_recv_kbuf(req);
4727	if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4728		req_set_fail(req);
4729	__io_req_complete(req, issue_flags, ret, cflags);
4730	return 0;
4731}
4732
4733static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4734{
4735	struct io_accept *accept = &req->accept;
4736
4737	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4738		return -EINVAL;
4739	if (sqe->ioprio || sqe->len || sqe->buf_index)
4740		return -EINVAL;
4741
4742	accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4743	accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4744	accept->flags = READ_ONCE(sqe->accept_flags);
4745	accept->nofile = rlimit(RLIMIT_NOFILE);
4746	return 0;
4747}
4748
4749static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4750{
4751	struct io_accept *accept = &req->accept;
4752	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4753	unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4754	int ret;
4755
4756	if (req->file->f_flags & O_NONBLOCK)
4757		req->flags |= REQ_F_NOWAIT;
4758
4759	ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4760					accept->addr_len, accept->flags,
4761					accept->nofile);
4762	if (ret == -EAGAIN && force_nonblock)
4763		return -EAGAIN;
4764	if (ret < 0) {
4765		if (ret == -ERESTARTSYS)
4766			ret = -EINTR;
4767		req_set_fail(req);
4768	}
4769	__io_req_complete(req, issue_flags, ret, 0);
4770	return 0;
4771}
4772
4773static int io_connect_prep_async(struct io_kiocb *req)
4774{
4775	struct io_async_connect *io = req->async_data;
4776	struct io_connect *conn = &req->connect;
4777
4778	return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4779}
4780
4781static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4782{
4783	struct io_connect *conn = &req->connect;
4784
4785	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4786		return -EINVAL;
4787	if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4788		return -EINVAL;
4789
4790	conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4791	conn->addr_len =  READ_ONCE(sqe->addr2);
4792	return 0;
4793}
4794
4795static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4796{
4797	struct io_async_connect __io, *io;
4798	unsigned file_flags;
4799	int ret;
4800	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4801
4802	if (req->async_data) {
4803		io = req->async_data;
4804	} else {
4805		ret = move_addr_to_kernel(req->connect.addr,
4806						req->connect.addr_len,
4807						&__io.address);
4808		if (ret)
4809			goto out;
4810		io = &__io;
4811	}
4812
4813	file_flags = force_nonblock ? O_NONBLOCK : 0;
4814
4815	ret = __sys_connect_file(req->file, &io->address,
4816					req->connect.addr_len, file_flags);
4817	if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4818		if (req->async_data)
4819			return -EAGAIN;
4820		if (io_alloc_async_data(req)) {
4821			ret = -ENOMEM;
4822			goto out;
4823		}
4824		memcpy(req->async_data, &__io, sizeof(__io));
4825		return -EAGAIN;
4826	}
4827	if (ret == -ERESTARTSYS)
4828		ret = -EINTR;
4829out:
4830	if (ret < 0)
4831		req_set_fail(req);
4832	__io_req_complete(req, issue_flags, ret, 0);
4833	return 0;
4834}
4835#else /* !CONFIG_NET */
4836#define IO_NETOP_FN(op)							\
4837static int io_##op(struct io_kiocb *req, unsigned int issue_flags)	\
4838{									\
4839	return -EOPNOTSUPP;						\
4840}
4841
4842#define IO_NETOP_PREP(op)						\
4843IO_NETOP_FN(op)								\
4844static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4845{									\
4846	return -EOPNOTSUPP;						\
4847}									\
4848
4849#define IO_NETOP_PREP_ASYNC(op)						\
4850IO_NETOP_PREP(op)							\
4851static int io_##op##_prep_async(struct io_kiocb *req)			\
4852{									\
4853	return -EOPNOTSUPP;						\
4854}
4855
4856IO_NETOP_PREP_ASYNC(sendmsg);
4857IO_NETOP_PREP_ASYNC(recvmsg);
4858IO_NETOP_PREP_ASYNC(connect);
4859IO_NETOP_PREP(accept);
4860IO_NETOP_FN(send);
4861IO_NETOP_FN(recv);
4862#endif /* CONFIG_NET */
4863
4864struct io_poll_table {
4865	struct poll_table_struct pt;
4866	struct io_kiocb *req;
4867	int nr_entries;
4868	int error;
4869};
4870
4871static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4872			   __poll_t mask, io_req_tw_func_t func)
4873{
4874	/* for instances that support it check for an event match first: */
4875	if (mask && !(mask & poll->events))
4876		return 0;
4877
4878	trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4879
4880	list_del_init(&poll->wait.entry);
4881
4882	req->result = mask;
4883	req->io_task_work.func = func;
4884
4885	/*
4886	 * If this fails, then the task is exiting. When a task exits, the
4887	 * work gets canceled, so just cancel this request as well instead
4888	 * of executing it. We can't safely execute it anyway, as we may not
4889	 * have the needed state needed for it anyway.
4890	 */
4891	io_req_task_work_add(req);
4892	return 1;
4893}
4894
4895static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4896	__acquires(&req->ctx->completion_lock)
4897{
4898	struct io_ring_ctx *ctx = req->ctx;
4899
4900	/* req->task == current here, checking PF_EXITING is safe */
4901	if (unlikely(req->task->flags & PF_EXITING))
4902		WRITE_ONCE(poll->canceled, true);
4903
4904	if (!req->result && !READ_ONCE(poll->canceled)) {
4905		struct poll_table_struct pt = { ._key = poll->events };
4906
4907		req->result = vfs_poll(req->file, &pt) & poll->events;
4908	}
4909
4910	spin_lock(&ctx->completion_lock);
4911	if (!req->result && !READ_ONCE(poll->canceled)) {
4912		add_wait_queue(poll->head, &poll->wait);
4913		return true;
4914	}
4915
4916	return false;
4917}
4918
4919static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4920{
4921	/* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4922	if (req->opcode == IORING_OP_POLL_ADD)
4923		return req->async_data;
4924	return req->apoll->double_poll;
4925}
4926
4927static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4928{
4929	if (req->opcode == IORING_OP_POLL_ADD)
4930		return &req->poll;
4931	return &req->apoll->poll;
4932}
4933
4934static void io_poll_remove_double(struct io_kiocb *req)
4935	__must_hold(&req->ctx->completion_lock)
4936{
4937	struct io_poll_iocb *poll = io_poll_get_double(req);
4938
4939	lockdep_assert_held(&req->ctx->completion_lock);
4940
4941	if (poll && poll->head) {
4942		struct wait_queue_head *head = poll->head;
4943
4944		spin_lock_irq(&head->lock);
4945		list_del_init(&poll->wait.entry);
4946		if (poll->wait.private)
4947			req_ref_put(req);
4948		poll->head = NULL;
4949		spin_unlock_irq(&head->lock);
4950	}
4951}
4952
4953static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4954	__must_hold(&req->ctx->completion_lock)
4955{
4956	struct io_ring_ctx *ctx = req->ctx;
4957	unsigned flags = IORING_CQE_F_MORE;
4958	int error;
4959
4960	if (READ_ONCE(req->poll.canceled)) {
4961		error = -ECANCELED;
4962		req->poll.events |= EPOLLONESHOT;
4963	} else {
4964		error = mangle_poll(mask);
4965	}
4966	if (req->poll.events & EPOLLONESHOT)
4967		flags = 0;
4968	if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4969		req->poll.done = true;
4970		flags = 0;
4971	}
4972	if (flags & IORING_CQE_F_MORE)
4973		ctx->cq_extra++;
4974
4975	io_commit_cqring(ctx);
4976	return !(flags & IORING_CQE_F_MORE);
4977}
4978
4979static void io_poll_task_func(struct io_kiocb *req)
4980{
4981	struct io_ring_ctx *ctx = req->ctx;
4982	struct io_kiocb *nxt;
4983
4984	if (io_poll_rewait(req, &req->poll)) {
4985		spin_unlock(&ctx->completion_lock);
4986	} else {
4987		bool done;
4988
4989		done = io_poll_complete(req, req->result);
4990		if (done) {
4991			io_poll_remove_double(req);
4992			hash_del(&req->hash_node);
4993		} else {
4994			req->result = 0;
4995			add_wait_queue(req->poll.head, &req->poll.wait);
4996		}
4997		spin_unlock(&ctx->completion_lock);
4998		io_cqring_ev_posted(ctx);
4999
5000		if (done) {
5001			nxt = io_put_req_find_next(req);
5002			if (nxt)
5003				io_req_task_submit(nxt);
5004		}
5005	}
5006}
5007
5008static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5009			       int sync, void *key)
5010{
5011	struct io_kiocb *req = wait->private;
5012	struct io_poll_iocb *poll = io_poll_get_single(req);
5013	__poll_t mask = key_to_poll(key);
5014	unsigned long flags;
5015
5016	/* for instances that support it check for an event match first: */
5017	if (mask && !(mask & poll->events))
5018		return 0;
5019	if (!(poll->events & EPOLLONESHOT))
5020		return poll->wait.func(&poll->wait, mode, sync, key);
5021
5022	list_del_init(&wait->entry);
5023
5024	if (poll->head) {
5025		bool done;
5026
5027		spin_lock_irqsave(&poll->head->lock, flags);
5028		done = list_empty(&poll->wait.entry);
5029		if (!done)
5030			list_del_init(&poll->wait.entry);
5031		/* make sure double remove sees this as being gone */
5032		wait->private = NULL;
5033		spin_unlock_irqrestore(&poll->head->lock, flags);
5034		if (!done) {
5035			/* use wait func handler, so it matches the rq type */
5036			poll->wait.func(&poll->wait, mode, sync, key);
5037		}
5038	}
5039	req_ref_put(req);
5040	return 1;
5041}
5042
5043static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5044			      wait_queue_func_t wake_func)
5045{
5046	poll->head = NULL;
5047	poll->done = false;
5048	poll->canceled = false;
5049#define IO_POLL_UNMASK	(EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5050	/* mask in events that we always want/need */
5051	poll->events = events | IO_POLL_UNMASK;
5052	INIT_LIST_HEAD(&poll->wait.entry);
5053	init_waitqueue_func_entry(&poll->wait, wake_func);
5054}
5055
5056static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5057			    struct wait_queue_head *head,
5058			    struct io_poll_iocb **poll_ptr)
5059{
5060	struct io_kiocb *req = pt->req;
5061
5062	/*
5063	 * The file being polled uses multiple waitqueues for poll handling
5064	 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5065	 * if this happens.
5066	 */
5067	if (unlikely(pt->nr_entries)) {
5068		struct io_poll_iocb *poll_one = poll;
5069
5070		/* double add on the same waitqueue head, ignore */
5071		if (poll_one->head == head)
5072			return;
5073		/* already have a 2nd entry, fail a third attempt */
5074		if (*poll_ptr) {
5075			if ((*poll_ptr)->head == head)
5076				return;
5077			pt->error = -EINVAL;
5078			return;
5079		}
5080		/*
5081		 * Can't handle multishot for double wait for now, turn it
5082		 * into one-shot mode.
5083		 */
5084		if (!(poll_one->events & EPOLLONESHOT))
5085			poll_one->events |= EPOLLONESHOT;
5086		poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5087		if (!poll) {
5088			pt->error = -ENOMEM;
5089			return;
5090		}
5091		io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5092		req_ref_get(req);
5093		poll->wait.private = req;
5094		*poll_ptr = poll;
5095	}
5096
5097	pt->nr_entries++;
5098	poll->head = head;
5099
5100	if (poll->events & EPOLLEXCLUSIVE)
5101		add_wait_queue_exclusive(head, &poll->wait);
5102	else
5103		add_wait_queue(head, &poll->wait);
5104}
5105
5106static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5107			       struct poll_table_struct *p)
5108{
5109	struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5110	struct async_poll *apoll = pt->req->apoll;
5111
5112	__io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5113}
5114
5115static void io_async_task_func(struct io_kiocb *req)
5116{
5117	struct async_poll *apoll = req->apoll;
5118	struct io_ring_ctx *ctx = req->ctx;
5119
5120	trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5121
5122	if (io_poll_rewait(req, &apoll->poll)) {
5123		spin_unlock(&ctx->completion_lock);
5124		return;
5125	}
5126
5127	hash_del(&req->hash_node);
5128	io_poll_remove_double(req);
5129	spin_unlock(&ctx->completion_lock);
5130
5131	if (!READ_ONCE(apoll->poll.canceled))
5132		io_req_task_submit(req);
5133	else
5134		io_req_complete_failed(req, -ECANCELED);
5135}
5136
5137static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5138			void *key)
5139{
5140	struct io_kiocb *req = wait->private;
5141	struct io_poll_iocb *poll = &req->apoll->poll;
5142
5143	trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5144					key_to_poll(key));
5145
5146	return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5147}
5148
5149static void io_poll_req_insert(struct io_kiocb *req)
5150{
5151	struct io_ring_ctx *ctx = req->ctx;
5152	struct hlist_head *list;
5153
5154	list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5155	hlist_add_head(&req->hash_node, list);
5156}
5157
5158static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5159				      struct io_poll_iocb *poll,
5160				      struct io_poll_table *ipt, __poll_t mask,
5161				      wait_queue_func_t wake_func)
5162	__acquires(&ctx->completion_lock)
5163{
5164	struct io_ring_ctx *ctx = req->ctx;
5165	bool cancel = false;
5166
5167	INIT_HLIST_NODE(&req->hash_node);
5168	io_init_poll_iocb(poll, mask, wake_func);
5169	poll->file = req->file;
5170	poll->wait.private = req;
5171
5172	ipt->pt._key = mask;
5173	ipt->req = req;
5174	ipt->error = 0;
5175	ipt->nr_entries = 0;
5176
5177	mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5178	if (unlikely(!ipt->nr_entries) && !ipt->error)
5179		ipt->error = -EINVAL;
5180
5181	spin_lock(&ctx->completion_lock);
5182	if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5183		io_poll_remove_double(req);
5184	if (likely(poll->head)) {
5185		spin_lock_irq(&poll->head->lock);
5186		if (unlikely(list_empty(&poll->wait.entry))) {
5187			if (ipt->error)
5188				cancel = true;
5189			ipt->error = 0;
5190			mask = 0;
5191		}
5192		if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5193			list_del_init(&poll->wait.entry);
5194		else if (cancel)
5195			WRITE_ONCE(poll->canceled, true);
5196		else if (!poll->done) /* actually waiting for an event */
5197			io_poll_req_insert(req);
5198		spin_unlock_irq(&poll->head->lock);
5199	}
5200
5201	return mask;
5202}
5203
5204enum {
5205	IO_APOLL_OK,
5206	IO_APOLL_ABORTED,
5207	IO_APOLL_READY
5208};
5209
5210static int io_arm_poll_handler(struct io_kiocb *req)
5211{
5212	const struct io_op_def *def = &io_op_defs[req->opcode];
5213	struct io_ring_ctx *ctx = req->ctx;
5214	struct async_poll *apoll;
5215	struct io_poll_table ipt;
5216	__poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5217	int rw;
5218
5219	if (!req->file || !file_can_poll(req->file))
5220		return IO_APOLL_ABORTED;
5221	if (req->flags & REQ_F_POLLED)
5222		return IO_APOLL_ABORTED;
5223	if (!def->pollin && !def->pollout)
5224		return IO_APOLL_ABORTED;
5225
5226	if (def->pollin) {
5227		rw = READ;
5228		mask |= POLLIN | POLLRDNORM;
5229
5230		/* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5231		if ((req->opcode == IORING_OP_RECVMSG) &&
5232		    (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5233			mask &= ~POLLIN;
5234	} else {
5235		rw = WRITE;
5236		mask |= POLLOUT | POLLWRNORM;
5237	}
5238
5239	/* if we can't nonblock try, then no point in arming a poll handler */
5240	if (!io_file_supports_nowait(req, rw))
5241		return IO_APOLL_ABORTED;
5242
5243	apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5244	if (unlikely(!apoll))
5245		return IO_APOLL_ABORTED;
5246	apoll->double_poll = NULL;
5247	req->apoll = apoll;
5248	req->flags |= REQ_F_POLLED;
5249	ipt.pt._qproc = io_async_queue_proc;
5250	io_req_set_refcount(req);
5251
5252	ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5253					io_async_wake);
5254	spin_unlock(&ctx->completion_lock);
5255	if (ret || ipt.error)
5256		return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5257
5258	trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5259				mask, apoll->poll.events);
5260	return IO_APOLL_OK;
5261}
5262
5263static bool __io_poll_remove_one(struct io_kiocb *req,
5264				 struct io_poll_iocb *poll, bool do_cancel)
5265	__must_hold(&req->ctx->completion_lock)
5266{
5267	bool do_complete = false;
5268
5269	if (!poll->head)
5270		return false;
5271	spin_lock_irq(&poll->head->lock);
5272	if (do_cancel)
5273		WRITE_ONCE(poll->canceled, true);
5274	if (!list_empty(&poll->wait.entry)) {
5275		list_del_init(&poll->wait.entry);
5276		do_complete = true;
5277	}
5278	spin_unlock_irq(&poll->head->lock);
5279	hash_del(&req->hash_node);
5280	return do_complete;
5281}
5282
5283static bool io_poll_remove_one(struct io_kiocb *req)
5284	__must_hold(&req->ctx->completion_lock)
5285{
5286	bool do_complete;
5287
5288	io_poll_remove_double(req);
5289	do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5290
5291	if (do_complete) {
5292		io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5293		io_commit_cqring(req->ctx);
5294		req_set_fail(req);
5295		io_put_req_deferred(req);
5296	}
5297	return do_complete;
5298}
5299
5300/*
5301 * Returns true if we found and killed one or more poll requests
5302 */
5303static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5304			       bool cancel_all)
5305{
5306	struct hlist_node *tmp;
5307	struct io_kiocb *req;
5308	int posted = 0, i;
5309
5310	spin_lock(&ctx->completion_lock);
5311	for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5312		struct hlist_head *list;
5313
5314		list = &ctx->cancel_hash[i];
5315		hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5316			if (io_match_task(req, tsk, cancel_all))
5317				posted += io_poll_remove_one(req);
5318		}
5319	}
5320	spin_unlock(&ctx->completion_lock);
5321
5322	if (posted)
5323		io_cqring_ev_posted(ctx);
5324
5325	return posted != 0;
5326}
5327
5328static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5329				     bool poll_only)
5330	__must_hold(&ctx->completion_lock)
5331{
5332	struct hlist_head *list;
5333	struct io_kiocb *req;
5334
5335	list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5336	hlist_for_each_entry(req, list, hash_node) {
5337		if (sqe_addr != req->user_data)
5338			continue;
5339		if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5340			continue;
5341		return req;
5342	}
5343	return NULL;
5344}
5345
5346static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5347			  bool poll_only)
5348	__must_hold(&ctx->completion_lock)
5349{
5350	struct io_kiocb *req;
5351
5352	req = io_poll_find(ctx, sqe_addr, poll_only);
5353	if (!req)
5354		return -ENOENT;
5355	if (io_poll_remove_one(req))
5356		return 0;
5357
5358	return -EALREADY;
5359}
5360
5361static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5362				     unsigned int flags)
5363{
5364	u32 events;
5365
5366	events = READ_ONCE(sqe->poll32_events);
5367#ifdef __BIG_ENDIAN
5368	events = swahw32(events);
5369#endif
5370	if (!(flags & IORING_POLL_ADD_MULTI))
5371		events |= EPOLLONESHOT;
5372	return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5373}
5374
5375static int io_poll_update_prep(struct io_kiocb *req,
5376			       const struct io_uring_sqe *sqe)
5377{
5378	struct io_poll_update *upd = &req->poll_update;
5379	u32 flags;
5380
5381	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5382		return -EINVAL;
5383	if (sqe->ioprio || sqe->buf_index)
5384		return -EINVAL;
5385	flags = READ_ONCE(sqe->len);
5386	if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5387		      IORING_POLL_ADD_MULTI))
5388		return -EINVAL;
5389	/* meaningless without update */
5390	if (flags == IORING_POLL_ADD_MULTI)
5391		return -EINVAL;
5392
5393	upd->old_user_data = READ_ONCE(sqe->addr);
5394	upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5395	upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5396
5397	upd->new_user_data = READ_ONCE(sqe->off);
5398	if (!upd->update_user_data && upd->new_user_data)
5399		return -EINVAL;
5400	if (upd->update_events)
5401		upd->events = io_poll_parse_events(sqe, flags);
5402	else if (sqe->poll32_events)
5403		return -EINVAL;
5404
5405	return 0;
5406}
5407
5408static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5409			void *key)
5410{
5411	struct io_kiocb *req = wait->private;
5412	struct io_poll_iocb *poll = &req->poll;
5413
5414	return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5415}
5416
5417static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5418			       struct poll_table_struct *p)
5419{
5420	struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5421
5422	__io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5423}
5424
5425static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5426{
5427	struct io_poll_iocb *poll = &req->poll;
5428	u32 flags;
5429
5430	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5431		return -EINVAL;
5432	if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5433		return -EINVAL;
5434	flags = READ_ONCE(sqe->len);
5435	if (flags & ~IORING_POLL_ADD_MULTI)
5436		return -EINVAL;
5437
5438	io_req_set_refcount(req);
5439	poll->events = io_poll_parse_events(sqe, flags);
5440	return 0;
5441}
5442
5443static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5444{
5445	struct io_poll_iocb *poll = &req->poll;
5446	struct io_ring_ctx *ctx = req->ctx;
5447	struct io_poll_table ipt;
5448	__poll_t mask;
5449
5450	ipt.pt._qproc = io_poll_queue_proc;
5451
5452	mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5453					io_poll_wake);
5454
5455	if (mask) { /* no async, we'd stolen it */
5456		ipt.error = 0;
5457		io_poll_complete(req, mask);
5458	}
5459	spin_unlock(&ctx->completion_lock);
5460
5461	if (mask) {
5462		io_cqring_ev_posted(ctx);
5463		if (poll->events & EPOLLONESHOT)
5464			io_put_req(req);
5465	}
5466	return ipt.error;
5467}
5468
5469static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5470{
5471	struct io_ring_ctx *ctx = req->ctx;
5472	struct io_kiocb *preq;
5473	bool completing;
5474	int ret;
5475
5476	spin_lock(&ctx->completion_lock);
5477	preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5478	if (!preq) {
5479		ret = -ENOENT;
5480		goto err;
5481	}
5482
5483	if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5484		completing = true;
5485		ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5486		goto err;
5487	}
5488
5489	/*
5490	 * Don't allow racy completion with singleshot, as we cannot safely
5491	 * update those. For multishot, if we're racing with completion, just
5492	 * let completion re-add it.
5493	 */
5494	completing = !__io_poll_remove_one(preq, &preq->poll, false);
5495	if (completing && (preq->poll.events & EPOLLONESHOT)) {
5496		ret = -EALREADY;
5497		goto err;
5498	}
5499	/* we now have a detached poll request. reissue. */
5500	ret = 0;
5501err:
5502	if (ret < 0) {
5503		spin_unlock(&ctx->completion_lock);
5504		req_set_fail(req);
5505		io_req_complete(req, ret);
5506		return 0;
5507	}
5508	/* only mask one event flags, keep behavior flags */
5509	if (req->poll_update.update_events) {
5510		preq->poll.events &= ~0xffff;
5511		preq->poll.events |= req->poll_update.events & 0xffff;
5512		preq->poll.events |= IO_POLL_UNMASK;
5513	}
5514	if (req->poll_update.update_user_data)
5515		preq->user_data = req->poll_update.new_user_data;
5516	spin_unlock(&ctx->completion_lock);
5517
5518	/* complete update request, we're done with it */
5519	io_req_complete(req, ret);
5520
5521	if (!completing) {
5522		ret = io_poll_add(preq, issue_flags);
5523		if (ret < 0) {
5524			req_set_fail(preq);
5525			io_req_complete(preq, ret);
5526		}
5527	}
5528	return 0;
5529}
5530
5531static void io_req_task_timeout(struct io_kiocb *req)
5532{
5533	req_set_fail(req);
5534	io_req_complete_post(req, -ETIME, 0);
5535}
5536
5537static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5538{
5539	struct io_timeout_data *data = container_of(timer,
5540						struct io_timeout_data, timer);
5541	struct io_kiocb *req = data->req;
5542	struct io_ring_ctx *ctx = req->ctx;
5543	unsigned long flags;
5544
5545	spin_lock_irqsave(&ctx->timeout_lock, flags);
5546	list_del_init(&req->timeout.list);
5547	atomic_set(&req->ctx->cq_timeouts,
5548		atomic_read(&req->ctx->cq_timeouts) + 1);
5549	spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5550
5551	req->io_task_work.func = io_req_task_timeout;
5552	io_req_task_work_add(req);
5553	return HRTIMER_NORESTART;
5554}
5555
5556static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5557					   __u64 user_data)
5558	__must_hold(&ctx->timeout_lock)
5559{
5560	struct io_timeout_data *io;
5561	struct io_kiocb *req;
5562	bool found = false;
5563
5564	list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5565		found = user_data == req->user_data;
5566		if (found)
5567			break;
5568	}
5569	if (!found)
5570		return ERR_PTR(-ENOENT);
5571
5572	io = req->async_data;
5573	if (hrtimer_try_to_cancel(&io->timer) == -1)
5574		return ERR_PTR(-EALREADY);
5575	list_del_init(&req->timeout.list);
5576	return req;
5577}
5578
5579static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5580	__must_hold(&ctx->completion_lock)
5581	__must_hold(&ctx->timeout_lock)
5582{
5583	struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5584
5585	if (IS_ERR(req))
5586		return PTR_ERR(req);
5587
5588	req_set_fail(req);
5589	io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5590	io_put_req_deferred(req);
5591	return 0;
5592}
5593
5594static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5595			     struct timespec64 *ts, enum hrtimer_mode mode)
5596	__must_hold(&ctx->timeout_lock)
5597{
5598	struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5599	struct io_timeout_data *data;
5600
5601	if (IS_ERR(req))
5602		return PTR_ERR(req);
5603
5604	req->timeout.off = 0; /* noseq */
5605	data = req->async_data;
5606	list_add_tail(&req->timeout.list, &ctx->timeout_list);
5607	hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5608	data->timer.function = io_timeout_fn;
5609	hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5610	return 0;
5611}
5612
5613static int io_timeout_remove_prep(struct io_kiocb *req,
5614				  const struct io_uring_sqe *sqe)
5615{
5616	struct io_timeout_rem *tr = &req->timeout_rem;
5617
5618	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5619		return -EINVAL;
5620	if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5621		return -EINVAL;
5622	if (sqe->ioprio || sqe->buf_index || sqe->len)
5623		return -EINVAL;
5624
5625	tr->addr = READ_ONCE(sqe->addr);
5626	tr->flags = READ_ONCE(sqe->timeout_flags);
5627	if (tr->flags & IORING_TIMEOUT_UPDATE) {
5628		if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5629			return -EINVAL;
5630		if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5631			return -EFAULT;
5632	} else if (tr->flags) {
5633		/* timeout removal doesn't support flags */
5634		return -EINVAL;
5635	}
5636
5637	return 0;
5638}
5639
5640static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5641{
5642	return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5643					    : HRTIMER_MODE_REL;
5644}
5645
5646/*
5647 * Remove or update an existing timeout command
5648 */
5649static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5650{
5651	struct io_timeout_rem *tr = &req->timeout_rem;
5652	struct io_ring_ctx *ctx = req->ctx;
5653	int ret;
5654
5655	if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
5656		spin_lock(&ctx->completion_lock);
5657		spin_lock_irq(&ctx->timeout_lock);
5658		ret = io_timeout_cancel(ctx, tr->addr);
5659		spin_unlock_irq(&ctx->timeout_lock);
5660		spin_unlock(&ctx->completion_lock);
5661	} else {
5662		spin_lock_irq(&ctx->timeout_lock);
5663		ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5664					io_translate_timeout_mode(tr->flags));
5665		spin_unlock_irq(&ctx->timeout_lock);
5666	}
5667
5668	if (ret < 0)
5669		req_set_fail(req);
5670	io_req_complete_post(req, ret, 0);
5671	return 0;
5672}
5673
5674static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5675			   bool is_timeout_link)
5676{
5677	struct io_timeout_data *data;
5678	unsigned flags;
5679	u32 off = READ_ONCE(sqe->off);
5680
5681	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5682		return -EINVAL;
5683	if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5684		return -EINVAL;
5685	if (off && is_timeout_link)
5686		return -EINVAL;
5687	flags = READ_ONCE(sqe->timeout_flags);
5688	if (flags & ~IORING_TIMEOUT_ABS)
5689		return -EINVAL;
5690
5691	req->timeout.off = off;
5692	if (unlikely(off && !req->ctx->off_timeout_used))
5693		req->ctx->off_timeout_used = true;
5694
5695	if (!req->async_data && io_alloc_async_data(req))
5696		return -ENOMEM;
5697
5698	data = req->async_data;
5699	data->req = req;
5700
5701	if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5702		return -EFAULT;
5703
5704	data->mode = io_translate_timeout_mode(flags);
5705	hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5706
5707	if (is_timeout_link) {
5708		struct io_submit_link *link = &req->ctx->submit_state.link;
5709
5710		if (!link->head)
5711			return -EINVAL;
5712		if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
5713			return -EINVAL;
5714		req->timeout.head = link->last;
5715		link->last->flags |= REQ_F_ARM_LTIMEOUT;
5716	}
5717	return 0;
5718}
5719
5720static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5721{
5722	struct io_ring_ctx *ctx = req->ctx;
5723	struct io_timeout_data *data = req->async_data;
5724	struct list_head *entry;
5725	u32 tail, off = req->timeout.off;