blob: 040e6c11ff37516ad69a0e32d759040d8cbeed97 [file] [log] [blame]
Jens Axboe2b188cc2019-01-07 10:46:33 -07001// 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
Stefan Bühler1e84b972019-04-24 23:54:16 +02007 * 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_cqring (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.
Jens Axboe2b188cc2019-01-07 10:46:33 -070029 *
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
Christoph Hellwigc992fe22019-01-11 09:43:02 -070040 * Copyright (c) 2018-2019 Christoph Hellwig
Jens Axboe2b188cc2019-01-07 10:46:33 -070041 */
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 <linux/refcount.h>
48#include <linux/uio.h>
49
50#include <linux/sched/signal.h>
51#include <linux/fs.h>
52#include <linux/file.h>
53#include <linux/fdtable.h>
54#include <linux/mm.h>
55#include <linux/mman.h>
56#include <linux/mmu_context.h>
57#include <linux/percpu.h>
58#include <linux/slab.h>
Jens Axboe6c271ce2019-01-10 11:22:30 -070059#include <linux/kthread.h>
Jens Axboe2b188cc2019-01-07 10:46:33 -070060#include <linux/blkdev.h>
Jens Axboeedafcce2019-01-09 09:16:05 -070061#include <linux/bvec.h>
Jens Axboe2b188cc2019-01-07 10:46:33 -070062#include <linux/net.h>
63#include <net/sock.h>
64#include <net/af_unix.h>
Jens Axboe6b063142019-01-10 22:13:58 -070065#include <net/scm.h>
Jens Axboe2b188cc2019-01-07 10:46:33 -070066#include <linux/anon_inodes.h>
67#include <linux/sched/mm.h>
68#include <linux/uaccess.h>
69#include <linux/nospec.h>
Jens Axboeedafcce2019-01-09 09:16:05 -070070#include <linux/sizes.h>
71#include <linux/hugetlb.h>
Jens Axboeaa4c3962019-11-29 10:14:00 -070072#include <linux/highmem.h>
Jens Axboe2b188cc2019-01-07 10:46:33 -070073
Dmitrii Dolgovc826bd72019-10-15 19:02:01 +020074#define CREATE_TRACE_POINTS
75#include <trace/events/io_uring.h>
76
Jens Axboe2b188cc2019-01-07 10:46:33 -070077#include <uapi/linux/io_uring.h>
78
79#include "internal.h"
Jens Axboe561fb042019-10-24 07:25:42 -060080#include "io-wq.h"
Jens Axboe2b188cc2019-01-07 10:46:33 -070081
Daniel Xu5277dea2019-09-14 14:23:45 -070082#define IORING_MAX_ENTRIES 32768
Jens Axboe33a107f2019-10-04 12:10:03 -060083#define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
Jens Axboe65e19f52019-10-26 07:20:21 -060084
85/*
86 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
87 */
88#define IORING_FILE_TABLE_SHIFT 9
89#define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
90#define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
91#define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
Jens Axboe2b188cc2019-01-07 10:46:33 -070092
93struct io_uring {
94 u32 head ____cacheline_aligned_in_smp;
95 u32 tail ____cacheline_aligned_in_smp;
96};
97
Stefan Bühler1e84b972019-04-24 23:54:16 +020098/*
Hristo Venev75b28af2019-08-26 17:23:46 +000099 * This data is shared with the application through the mmap at offsets
100 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
Stefan Bühler1e84b972019-04-24 23:54:16 +0200101 *
102 * The offsets to the member fields are published through struct
103 * io_sqring_offsets when calling io_uring_setup.
104 */
Hristo Venev75b28af2019-08-26 17:23:46 +0000105struct io_rings {
Stefan Bühler1e84b972019-04-24 23:54:16 +0200106 /*
107 * Head and tail offsets into the ring; the offsets need to be
108 * masked to get valid indices.
109 *
Hristo Venev75b28af2019-08-26 17:23:46 +0000110 * The kernel controls head of the sq ring and the tail of the cq ring,
111 * and the application controls tail of the sq ring and the head of the
112 * cq ring.
Stefan Bühler1e84b972019-04-24 23:54:16 +0200113 */
Hristo Venev75b28af2019-08-26 17:23:46 +0000114 struct io_uring sq, cq;
Stefan Bühler1e84b972019-04-24 23:54:16 +0200115 /*
Hristo Venev75b28af2019-08-26 17:23:46 +0000116 * Bitmasks to apply to head and tail offsets (constant, equals
Stefan Bühler1e84b972019-04-24 23:54:16 +0200117 * ring_entries - 1)
118 */
Hristo Venev75b28af2019-08-26 17:23:46 +0000119 u32 sq_ring_mask, cq_ring_mask;
120 /* Ring sizes (constant, power of 2) */
121 u32 sq_ring_entries, cq_ring_entries;
Stefan Bühler1e84b972019-04-24 23:54:16 +0200122 /*
123 * Number of invalid entries dropped by the kernel due to
124 * invalid index stored in array
125 *
126 * Written by the kernel, shouldn't be modified by the
127 * application (i.e. get number of "new events" by comparing to
128 * cached value).
129 *
130 * After a new SQ head value was read by the application this
131 * counter includes all submissions that were dropped reaching
132 * the new SQ head (and possibly more).
133 */
Hristo Venev75b28af2019-08-26 17:23:46 +0000134 u32 sq_dropped;
Stefan Bühler1e84b972019-04-24 23:54:16 +0200135 /*
136 * Runtime flags
137 *
138 * Written by the kernel, shouldn't be modified by the
139 * application.
140 *
141 * The application needs a full memory barrier before checking
142 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
143 */
Hristo Venev75b28af2019-08-26 17:23:46 +0000144 u32 sq_flags;
Stefan Bühler1e84b972019-04-24 23:54:16 +0200145 /*
146 * Number of completion events lost because the queue was full;
147 * this should be avoided by the application by making sure
LimingWu0b4295b2019-12-05 20:18:18 +0800148 * there are not more requests pending than there is space in
Stefan Bühler1e84b972019-04-24 23:54:16 +0200149 * the completion queue.
150 *
151 * Written by the kernel, shouldn't be modified by the
152 * application (i.e. get number of "new events" by comparing to
153 * cached value).
154 *
155 * As completion events come in out of order this counter is not
156 * ordered with any other data.
157 */
Hristo Venev75b28af2019-08-26 17:23:46 +0000158 u32 cq_overflow;
Stefan Bühler1e84b972019-04-24 23:54:16 +0200159 /*
160 * Ring buffer of completion events.
161 *
162 * The kernel writes completion events fresh every time they are
163 * produced, so the application is allowed to modify pending
164 * entries.
165 */
Hristo Venev75b28af2019-08-26 17:23:46 +0000166 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700167};
168
Jens Axboeedafcce2019-01-09 09:16:05 -0700169struct io_mapped_ubuf {
170 u64 ubuf;
171 size_t len;
172 struct bio_vec *bvec;
173 unsigned int nr_bvecs;
174};
175
Jens Axboe65e19f52019-10-26 07:20:21 -0600176struct fixed_file_table {
177 struct file **files;
Jens Axboe31b51512019-01-18 22:56:34 -0700178};
179
Jens Axboe2b188cc2019-01-07 10:46:33 -0700180struct io_ring_ctx {
181 struct {
182 struct percpu_ref refs;
183 } ____cacheline_aligned_in_smp;
184
185 struct {
186 unsigned int flags;
187 bool compat;
188 bool account_mem;
Jens Axboe1d7bb1d2019-11-06 11:31:17 -0700189 bool cq_overflow_flushed;
Pavel Begunkov1b4a51b2019-11-21 11:54:28 +0300190 bool drain_next;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700191
Hristo Venev75b28af2019-08-26 17:23:46 +0000192 /*
193 * Ring buffer of indices into array of io_uring_sqe, which is
194 * mmapped by the application using the IORING_OFF_SQES offset.
195 *
196 * This indirection could e.g. be used to assign fixed
197 * io_uring_sqe entries to operations and only submit them to
198 * the queue when needed.
199 *
200 * The kernel modifies neither the indices array nor the entries
201 * array.
202 */
203 u32 *sq_array;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700204 unsigned cached_sq_head;
205 unsigned sq_entries;
206 unsigned sq_mask;
Jens Axboe6c271ce2019-01-10 11:22:30 -0700207 unsigned sq_thread_idle;
Jens Axboe498ccd92019-10-25 10:04:25 -0600208 unsigned cached_sq_dropped;
Jens Axboe206aefd2019-11-07 18:27:42 -0700209 atomic_t cached_cq_overflow;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700210 struct io_uring_sqe *sq_sqes;
Jens Axboede0617e2019-04-06 21:51:27 -0600211
212 struct list_head defer_list;
Jens Axboe5262f562019-09-17 12:26:57 -0600213 struct list_head timeout_list;
Jens Axboe1d7bb1d2019-11-06 11:31:17 -0700214 struct list_head cq_overflow_list;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700215
Jens Axboefcb323c2019-10-24 12:39:47 -0600216 wait_queue_head_t inflight_wait;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700217 } ____cacheline_aligned_in_smp;
218
Hristo Venev75b28af2019-08-26 17:23:46 +0000219 struct io_rings *rings;
220
Jens Axboe2b188cc2019-01-07 10:46:33 -0700221 /* IO offload */
Jens Axboe561fb042019-10-24 07:25:42 -0600222 struct io_wq *io_wq;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700223 struct task_struct *sqo_thread; /* if using sq thread polling */
224 struct mm_struct *sqo_mm;
225 wait_queue_head_t sqo_wait;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700226
Jens Axboe6b063142019-01-10 22:13:58 -0700227 /*
228 * If used, fixed file set. Writers must ensure that ->refs is dead,
229 * readers must ensure that ->refs is alive as long as the file* is
230 * used. Only updated through io_uring_register(2).
231 */
Jens Axboe65e19f52019-10-26 07:20:21 -0600232 struct fixed_file_table *file_table;
Jens Axboe6b063142019-01-10 22:13:58 -0700233 unsigned nr_user_files;
234
Jens Axboeedafcce2019-01-09 09:16:05 -0700235 /* if used, fixed mapped user buffers */
236 unsigned nr_user_bufs;
237 struct io_mapped_ubuf *user_bufs;
238
Jens Axboe2b188cc2019-01-07 10:46:33 -0700239 struct user_struct *user;
240
Jens Axboe0b8c0ec2019-12-02 08:50:00 -0700241 const struct cred *creds;
Jens Axboe181e4482019-11-25 08:52:30 -0700242
Jens Axboe206aefd2019-11-07 18:27:42 -0700243 /* 0 is for ctx quiesce/reinit/free, 1 is for sqo_thread started */
244 struct completion *completions;
245
Jens Axboe0ddf92e2019-11-08 08:52:53 -0700246 /* if all else fails... */
247 struct io_kiocb *fallback_req;
248
Jens Axboe206aefd2019-11-07 18:27:42 -0700249#if defined(CONFIG_UNIX)
250 struct socket *ring_sock;
251#endif
252
253 struct {
254 unsigned cached_cq_tail;
255 unsigned cq_entries;
256 unsigned cq_mask;
257 atomic_t cq_timeouts;
258 struct wait_queue_head cq_wait;
259 struct fasync_struct *cq_fasync;
260 struct eventfd_ctx *cq_ev_fd;
261 } ____cacheline_aligned_in_smp;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700262
263 struct {
264 struct mutex uring_lock;
265 wait_queue_head_t wait;
266 } ____cacheline_aligned_in_smp;
267
268 struct {
269 spinlock_t completion_lock;
Jens Axboedef596e2019-01-09 08:59:42 -0700270 bool poll_multi_file;
271 /*
272 * ->poll_list is protected by the ctx->uring_lock for
273 * io_uring instances that don't use IORING_SETUP_SQPOLL.
274 * For SQPOLL, only the single threaded io_sq_thread() will
275 * manipulate the list, hence no extra locking is needed there.
276 */
277 struct list_head poll_list;
Jens Axboe78076bb2019-12-04 19:56:40 -0700278 struct hlist_head *cancel_hash;
279 unsigned cancel_hash_bits;
Jens Axboefcb323c2019-10-24 12:39:47 -0600280
281 spinlock_t inflight_lock;
282 struct list_head inflight_list;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700283 } ____cacheline_aligned_in_smp;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700284};
285
Jens Axboe09bb8392019-03-13 12:39:28 -0600286/*
287 * First field must be the file pointer in all the
288 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
289 */
Jens Axboe221c5eb2019-01-17 09:41:58 -0700290struct io_poll_iocb {
291 struct file *file;
292 struct wait_queue_head *head;
293 __poll_t events;
Jens Axboe8c838782019-03-12 15:48:16 -0600294 bool done;
Jens Axboe221c5eb2019-01-17 09:41:58 -0700295 bool canceled;
Jens Axboee9444752019-11-26 15:02:04 -0700296 struct wait_queue_entry *wait;
Jens Axboe221c5eb2019-01-17 09:41:58 -0700297};
298
Jens Axboead8a48a2019-11-15 08:49:11 -0700299struct io_timeout_data {
300 struct io_kiocb *req;
301 struct hrtimer timer;
302 struct timespec64 ts;
303 enum hrtimer_mode mode;
Pavel Begunkovcc42e0a2019-11-25 23:14:38 +0300304 u32 seq_offset;
Jens Axboead8a48a2019-11-15 08:49:11 -0700305};
306
Jens Axboef499a022019-12-02 16:28:46 -0700307struct io_async_connect {
308 struct sockaddr_storage address;
309};
310
Jens Axboe03b12302019-12-02 18:50:25 -0700311struct io_async_msghdr {
312 struct iovec fast_iov[UIO_FASTIOV];
313 struct iovec *iov;
314 struct sockaddr __user *uaddr;
315 struct msghdr msg;
316};
317
Jens Axboef67676d2019-12-02 11:03:47 -0700318struct io_async_rw {
319 struct iovec fast_iov[UIO_FASTIOV];
320 struct iovec *iov;
321 ssize_t nr_segs;
322 ssize_t size;
323};
324
Jens Axboe1a6b74f2019-12-02 10:33:15 -0700325struct io_async_ctx {
326 struct io_uring_sqe sqe;
Jens Axboef67676d2019-12-02 11:03:47 -0700327 union {
328 struct io_async_rw rw;
Jens Axboe03b12302019-12-02 18:50:25 -0700329 struct io_async_msghdr msg;
Jens Axboef499a022019-12-02 16:28:46 -0700330 struct io_async_connect connect;
Jens Axboe2d283902019-12-04 11:08:05 -0700331 struct io_timeout_data timeout;
Jens Axboef67676d2019-12-02 11:03:47 -0700332 };
Jens Axboe1a6b74f2019-12-02 10:33:15 -0700333};
334
Jens Axboe09bb8392019-03-13 12:39:28 -0600335/*
336 * NOTE! Each of the iocb union members has the file pointer
337 * as the first entry in their struct definition. So you can
338 * access the file pointer through any of the sub-structs,
339 * or directly as just 'ki_filp' in this struct.
340 */
Jens Axboe2b188cc2019-01-07 10:46:33 -0700341struct io_kiocb {
Jens Axboe221c5eb2019-01-17 09:41:58 -0700342 union {
Jens Axboe09bb8392019-03-13 12:39:28 -0600343 struct file *file;
Jens Axboe221c5eb2019-01-17 09:41:58 -0700344 struct kiocb rw;
345 struct io_poll_iocb poll;
346 };
Jens Axboe2b188cc2019-01-07 10:46:33 -0700347
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +0300348 const struct io_uring_sqe *sqe;
Jens Axboe1a6b74f2019-12-02 10:33:15 -0700349 struct io_async_ctx *io;
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +0300350 struct file *ring_file;
351 int ring_fd;
352 bool has_user;
353 bool in_async;
354 bool needs_fixed_file;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700355
356 struct io_ring_ctx *ctx;
Jens Axboeeac406c2019-11-14 12:09:58 -0700357 union {
358 struct list_head list;
Jens Axboe78076bb2019-12-04 19:56:40 -0700359 struct hlist_node hash_node;
Jens Axboeeac406c2019-11-14 12:09:58 -0700360 };
Jens Axboe9e645e112019-05-10 16:07:28 -0600361 struct list_head link_list;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700362 unsigned int flags;
Jens Axboec16361c2019-01-17 08:39:48 -0700363 refcount_t refs;
Stefan Bühler8449eed2019-04-27 20:34:19 +0200364#define REQ_F_NOWAIT 1 /* must not punt to workers */
Jens Axboedef596e2019-01-09 08:59:42 -0700365#define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */
Jens Axboe6b063142019-01-10 22:13:58 -0700366#define REQ_F_FIXED_FILE 4 /* ctx owns file */
Jens Axboe4d7dd462019-11-20 13:03:52 -0700367#define REQ_F_LINK_NEXT 8 /* already grabbed next link */
Stefan Bühlere2033e32019-05-11 19:08:01 +0200368#define REQ_F_IO_DRAIN 16 /* drain existing IO first */
369#define REQ_F_IO_DRAINED 32 /* drain done */
Jens Axboe9e645e112019-05-10 16:07:28 -0600370#define REQ_F_LINK 64 /* linked sqes */
Jens Axboe2665abf2019-11-05 12:40:47 -0700371#define REQ_F_LINK_TIMEOUT 128 /* has linked timeout */
Zhengyuan Liuf7b76ac2019-07-16 23:26:14 +0800372#define REQ_F_FAIL_LINK 256 /* fail rest of links */
Pavel Begunkov1b4a51b2019-11-21 11:54:28 +0300373#define REQ_F_DRAIN_LINK 512 /* link should be fully drained */
Jens Axboe5262f562019-09-17 12:26:57 -0600374#define REQ_F_TIMEOUT 1024 /* timeout request */
Jens Axboe491381ce2019-10-17 09:20:46 -0600375#define REQ_F_ISREG 2048 /* regular file */
376#define REQ_F_MUST_PUNT 4096 /* must be punted even for NONBLOCK */
Jens Axboe93bd25b2019-11-11 23:34:31 -0700377#define REQ_F_TIMEOUT_NOSEQ 8192 /* no timeout sequence */
Linus Torvaldsfb4b3d32019-11-25 10:40:27 -0800378#define REQ_F_INFLIGHT 16384 /* on inflight list */
379#define REQ_F_COMP_LOCKED 32768 /* completion under lock */
Jens Axboe4e88d6e2019-12-07 20:59:47 -0700380#define REQ_F_HARDLINK 65536 /* doesn't sever on completion < 0 */
Jens Axboe2b188cc2019-01-07 10:46:33 -0700381 u64 user_data;
Jens Axboe9e645e112019-05-10 16:07:28 -0600382 u32 result;
Jens Axboede0617e2019-04-06 21:51:27 -0600383 u32 sequence;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700384
Jens Axboefcb323c2019-10-24 12:39:47 -0600385 struct list_head inflight_entry;
386
Jens Axboe561fb042019-10-24 07:25:42 -0600387 struct io_wq_work work;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700388};
389
390#define IO_PLUG_THRESHOLD 2
Jens Axboedef596e2019-01-09 08:59:42 -0700391#define IO_IOPOLL_BATCH 8
Jens Axboe2b188cc2019-01-07 10:46:33 -0700392
Jens Axboe9a56a232019-01-09 09:06:50 -0700393struct io_submit_state {
394 struct blk_plug plug;
395
396 /*
Jens Axboe2579f912019-01-09 09:10:43 -0700397 * io_kiocb alloc cache
398 */
399 void *reqs[IO_IOPOLL_BATCH];
400 unsigned int free_reqs;
401 unsigned int cur_req;
402
403 /*
Jens Axboe9a56a232019-01-09 09:06:50 -0700404 * File reference cache
405 */
406 struct file *file;
407 unsigned int fd;
408 unsigned int has_refs;
409 unsigned int used_refs;
410 unsigned int ios_left;
411};
412
Jens Axboe561fb042019-10-24 07:25:42 -0600413static void io_wq_submit_work(struct io_wq_work **workptr);
Jens Axboe78e19bb2019-11-06 15:21:34 -0700414static void io_cqring_fill_event(struct io_kiocb *req, long res);
Jackie Liu4fe2c962019-09-09 20:50:40 +0800415static void __io_free_req(struct io_kiocb *req);
Jackie Liuec9c02a2019-11-08 23:50:36 +0800416static void io_put_req(struct io_kiocb *req);
Jens Axboe78e19bb2019-11-06 15:21:34 -0700417static void io_double_put_req(struct io_kiocb *req);
Jens Axboe978db572019-11-14 22:39:04 -0700418static void __io_double_put_req(struct io_kiocb *req);
Jens Axboe94ae5e72019-11-14 19:39:52 -0700419static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
420static void io_queue_linked_timeout(struct io_kiocb *req);
Jens Axboede0617e2019-04-06 21:51:27 -0600421
Jens Axboe2b188cc2019-01-07 10:46:33 -0700422static struct kmem_cache *req_cachep;
423
424static const struct file_operations io_uring_fops;
425
426struct sock *io_uring_get_socket(struct file *file)
427{
428#if defined(CONFIG_UNIX)
429 if (file->f_op == &io_uring_fops) {
430 struct io_ring_ctx *ctx = file->private_data;
431
432 return ctx->ring_sock->sk;
433 }
434#endif
435 return NULL;
436}
437EXPORT_SYMBOL(io_uring_get_socket);
438
439static void io_ring_ctx_ref_free(struct percpu_ref *ref)
440{
441 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
442
Jens Axboe206aefd2019-11-07 18:27:42 -0700443 complete(&ctx->completions[0]);
Jens Axboe2b188cc2019-01-07 10:46:33 -0700444}
445
446static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
447{
448 struct io_ring_ctx *ctx;
Jens Axboe78076bb2019-12-04 19:56:40 -0700449 int hash_bits;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700450
451 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
452 if (!ctx)
453 return NULL;
454
Jens Axboe0ddf92e2019-11-08 08:52:53 -0700455 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
456 if (!ctx->fallback_req)
457 goto err;
458
Jens Axboe206aefd2019-11-07 18:27:42 -0700459 ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL);
460 if (!ctx->completions)
461 goto err;
462
Jens Axboe78076bb2019-12-04 19:56:40 -0700463 /*
464 * Use 5 bits less than the max cq entries, that should give us around
465 * 32 entries per hash list if totally full and uniformly spread.
466 */
467 hash_bits = ilog2(p->cq_entries);
468 hash_bits -= 5;
469 if (hash_bits <= 0)
470 hash_bits = 1;
471 ctx->cancel_hash_bits = hash_bits;
472 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
473 GFP_KERNEL);
474 if (!ctx->cancel_hash)
475 goto err;
476 __hash_init(ctx->cancel_hash, 1U << hash_bits);
477
Roman Gushchin21482892019-05-07 10:01:48 -0700478 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
Jens Axboe206aefd2019-11-07 18:27:42 -0700479 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
480 goto err;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700481
482 ctx->flags = p->flags;
483 init_waitqueue_head(&ctx->cq_wait);
Jens Axboe1d7bb1d2019-11-06 11:31:17 -0700484 INIT_LIST_HEAD(&ctx->cq_overflow_list);
Jens Axboe206aefd2019-11-07 18:27:42 -0700485 init_completion(&ctx->completions[0]);
486 init_completion(&ctx->completions[1]);
Jens Axboe2b188cc2019-01-07 10:46:33 -0700487 mutex_init(&ctx->uring_lock);
488 init_waitqueue_head(&ctx->wait);
489 spin_lock_init(&ctx->completion_lock);
Jens Axboedef596e2019-01-09 08:59:42 -0700490 INIT_LIST_HEAD(&ctx->poll_list);
Jens Axboede0617e2019-04-06 21:51:27 -0600491 INIT_LIST_HEAD(&ctx->defer_list);
Jens Axboe5262f562019-09-17 12:26:57 -0600492 INIT_LIST_HEAD(&ctx->timeout_list);
Jens Axboefcb323c2019-10-24 12:39:47 -0600493 init_waitqueue_head(&ctx->inflight_wait);
494 spin_lock_init(&ctx->inflight_lock);
495 INIT_LIST_HEAD(&ctx->inflight_list);
Jens Axboe2b188cc2019-01-07 10:46:33 -0700496 return ctx;
Jens Axboe206aefd2019-11-07 18:27:42 -0700497err:
Jens Axboe0ddf92e2019-11-08 08:52:53 -0700498 if (ctx->fallback_req)
499 kmem_cache_free(req_cachep, ctx->fallback_req);
Jens Axboe206aefd2019-11-07 18:27:42 -0700500 kfree(ctx->completions);
Jens Axboe78076bb2019-12-04 19:56:40 -0700501 kfree(ctx->cancel_hash);
Jens Axboe206aefd2019-11-07 18:27:42 -0700502 kfree(ctx);
503 return NULL;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700504}
505
Bob Liu9d858b22019-11-13 18:06:25 +0800506static inline bool __req_need_defer(struct io_kiocb *req)
Jens Axboede0617e2019-04-06 21:51:27 -0600507{
Jackie Liua197f662019-11-08 08:09:12 -0700508 struct io_ring_ctx *ctx = req->ctx;
509
Jens Axboe498ccd92019-10-25 10:04:25 -0600510 return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped
511 + atomic_read(&ctx->cached_cq_overflow);
Jens Axboede0617e2019-04-06 21:51:27 -0600512}
513
Bob Liu9d858b22019-11-13 18:06:25 +0800514static inline bool req_need_defer(struct io_kiocb *req)
Jens Axboe7adf4ea2019-10-10 21:42:58 -0600515{
Bob Liu9d858b22019-11-13 18:06:25 +0800516 if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) == REQ_F_IO_DRAIN)
517 return __req_need_defer(req);
Jens Axboe7adf4ea2019-10-10 21:42:58 -0600518
Bob Liu9d858b22019-11-13 18:06:25 +0800519 return false;
Jens Axboe7adf4ea2019-10-10 21:42:58 -0600520}
521
522static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
Jens Axboede0617e2019-04-06 21:51:27 -0600523{
524 struct io_kiocb *req;
525
Jens Axboe7adf4ea2019-10-10 21:42:58 -0600526 req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list);
Bob Liu9d858b22019-11-13 18:06:25 +0800527 if (req && !req_need_defer(req)) {
Jens Axboede0617e2019-04-06 21:51:27 -0600528 list_del_init(&req->list);
529 return req;
530 }
531
532 return NULL;
533}
534
Jens Axboe5262f562019-09-17 12:26:57 -0600535static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx)
536{
Jens Axboe7adf4ea2019-10-10 21:42:58 -0600537 struct io_kiocb *req;
538
539 req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list);
Jens Axboe93bd25b2019-11-11 23:34:31 -0700540 if (req) {
541 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
542 return NULL;
Linus Torvaldsfb4b3d32019-11-25 10:40:27 -0800543 if (!__req_need_defer(req)) {
Jens Axboe93bd25b2019-11-11 23:34:31 -0700544 list_del_init(&req->list);
545 return req;
546 }
Jens Axboe7adf4ea2019-10-10 21:42:58 -0600547 }
548
549 return NULL;
Jens Axboe5262f562019-09-17 12:26:57 -0600550}
551
Jens Axboede0617e2019-04-06 21:51:27 -0600552static void __io_commit_cqring(struct io_ring_ctx *ctx)
Jens Axboe2b188cc2019-01-07 10:46:33 -0700553{
Hristo Venev75b28af2019-08-26 17:23:46 +0000554 struct io_rings *rings = ctx->rings;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700555
Hristo Venev75b28af2019-08-26 17:23:46 +0000556 if (ctx->cached_cq_tail != READ_ONCE(rings->cq.tail)) {
Jens Axboe2b188cc2019-01-07 10:46:33 -0700557 /* order cqe stores with ring update */
Hristo Venev75b28af2019-08-26 17:23:46 +0000558 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
Jens Axboe2b188cc2019-01-07 10:46:33 -0700559
Jens Axboe2b188cc2019-01-07 10:46:33 -0700560 if (wq_has_sleeper(&ctx->cq_wait)) {
561 wake_up_interruptible(&ctx->cq_wait);
562 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
563 }
564 }
565}
566
Jens Axboe561fb042019-10-24 07:25:42 -0600567static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe)
Jens Axboe18d9be12019-09-10 09:13:05 -0600568{
Jens Axboe561fb042019-10-24 07:25:42 -0600569 u8 opcode = READ_ONCE(sqe->opcode);
570
571 return !(opcode == IORING_OP_READ_FIXED ||
572 opcode == IORING_OP_WRITE_FIXED);
573}
574
Jens Axboe94ae5e72019-11-14 19:39:52 -0700575static inline bool io_prep_async_work(struct io_kiocb *req,
576 struct io_kiocb **link)
Jens Axboe561fb042019-10-24 07:25:42 -0600577{
578 bool do_hashed = false;
Jens Axboe54a91f32019-09-10 09:15:04 -0600579
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +0300580 if (req->sqe) {
581 switch (req->sqe->opcode) {
Jens Axboe6cc47d12019-09-18 11:18:23 -0600582 case IORING_OP_WRITEV:
583 case IORING_OP_WRITE_FIXED:
Jens Axboe561fb042019-10-24 07:25:42 -0600584 do_hashed = true;
Jens Axboe5f8fd2d32019-11-07 10:57:36 -0700585 /* fall-through */
586 case IORING_OP_READV:
587 case IORING_OP_READ_FIXED:
588 case IORING_OP_SENDMSG:
589 case IORING_OP_RECVMSG:
590 case IORING_OP_ACCEPT:
591 case IORING_OP_POLL_ADD:
Jens Axboef8e85cf2019-11-23 14:24:24 -0700592 case IORING_OP_CONNECT:
Jens Axboe5f8fd2d32019-11-07 10:57:36 -0700593 /*
594 * We know REQ_F_ISREG is not set on some of these
595 * opcodes, but this enables us to keep the check in
596 * just one place.
597 */
598 if (!(req->flags & REQ_F_ISREG))
599 req->work.flags |= IO_WQ_WORK_UNBOUND;
Jens Axboe6cc47d12019-09-18 11:18:23 -0600600 break;
601 }
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +0300602 if (io_sqe_needs_user(req->sqe))
Jens Axboe561fb042019-10-24 07:25:42 -0600603 req->work.flags |= IO_WQ_WORK_NEEDS_USER;
Jens Axboe54a91f32019-09-10 09:15:04 -0600604 }
605
Jens Axboe94ae5e72019-11-14 19:39:52 -0700606 *link = io_prep_linked_timeout(req);
Jens Axboe561fb042019-10-24 07:25:42 -0600607 return do_hashed;
608}
609
Jackie Liua197f662019-11-08 08:09:12 -0700610static inline void io_queue_async_work(struct io_kiocb *req)
Jens Axboe561fb042019-10-24 07:25:42 -0600611{
Jackie Liua197f662019-11-08 08:09:12 -0700612 struct io_ring_ctx *ctx = req->ctx;
Jens Axboe94ae5e72019-11-14 19:39:52 -0700613 struct io_kiocb *link;
614 bool do_hashed;
615
616 do_hashed = io_prep_async_work(req, &link);
Jens Axboe561fb042019-10-24 07:25:42 -0600617
618 trace_io_uring_queue_async_work(ctx, do_hashed, req, &req->work,
619 req->flags);
620 if (!do_hashed) {
621 io_wq_enqueue(ctx->io_wq, &req->work);
622 } else {
623 io_wq_enqueue_hashed(ctx->io_wq, &req->work,
624 file_inode(req->file));
625 }
Jens Axboe94ae5e72019-11-14 19:39:52 -0700626
627 if (link)
628 io_queue_linked_timeout(link);
Jens Axboe18d9be12019-09-10 09:13:05 -0600629}
630
Jens Axboe5262f562019-09-17 12:26:57 -0600631static void io_kill_timeout(struct io_kiocb *req)
632{
633 int ret;
634
Jens Axboe2d283902019-12-04 11:08:05 -0700635 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
Jens Axboe5262f562019-09-17 12:26:57 -0600636 if (ret != -1) {
637 atomic_inc(&req->ctx->cq_timeouts);
Jens Axboe842f9612019-10-29 12:34:10 -0600638 list_del_init(&req->list);
Jens Axboe78e19bb2019-11-06 15:21:34 -0700639 io_cqring_fill_event(req, 0);
Jackie Liuec9c02a2019-11-08 23:50:36 +0800640 io_put_req(req);
Jens Axboe5262f562019-09-17 12:26:57 -0600641 }
642}
643
644static void io_kill_timeouts(struct io_ring_ctx *ctx)
645{
646 struct io_kiocb *req, *tmp;
647
648 spin_lock_irq(&ctx->completion_lock);
649 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
650 io_kill_timeout(req);
651 spin_unlock_irq(&ctx->completion_lock);
652}
653
Jens Axboede0617e2019-04-06 21:51:27 -0600654static void io_commit_cqring(struct io_ring_ctx *ctx)
655{
656 struct io_kiocb *req;
657
Jens Axboe5262f562019-09-17 12:26:57 -0600658 while ((req = io_get_timeout_req(ctx)) != NULL)
659 io_kill_timeout(req);
660
Jens Axboede0617e2019-04-06 21:51:27 -0600661 __io_commit_cqring(ctx);
662
663 while ((req = io_get_deferred_req(ctx)) != NULL) {
664 req->flags |= REQ_F_IO_DRAINED;
Jackie Liua197f662019-11-08 08:09:12 -0700665 io_queue_async_work(req);
Jens Axboede0617e2019-04-06 21:51:27 -0600666 }
667}
668
Jens Axboe2b188cc2019-01-07 10:46:33 -0700669static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
670{
Hristo Venev75b28af2019-08-26 17:23:46 +0000671 struct io_rings *rings = ctx->rings;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700672 unsigned tail;
673
674 tail = ctx->cached_cq_tail;
Stefan Bühler115e12e2019-04-24 23:54:18 +0200675 /*
676 * writes to the cq entry need to come after reading head; the
677 * control dependency is enough as we're using WRITE_ONCE to
678 * fill the cq entry
679 */
Hristo Venev75b28af2019-08-26 17:23:46 +0000680 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
Jens Axboe2b188cc2019-01-07 10:46:33 -0700681 return NULL;
682
683 ctx->cached_cq_tail++;
Hristo Venev75b28af2019-08-26 17:23:46 +0000684 return &rings->cqes[tail & ctx->cq_mask];
Jens Axboe2b188cc2019-01-07 10:46:33 -0700685}
686
Jens Axboe8c838782019-03-12 15:48:16 -0600687static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
688{
689 if (waitqueue_active(&ctx->wait))
690 wake_up(&ctx->wait);
691 if (waitqueue_active(&ctx->sqo_wait))
692 wake_up(&ctx->sqo_wait);
Jens Axboe9b402842019-04-11 11:45:41 -0600693 if (ctx->cq_ev_fd)
694 eventfd_signal(ctx->cq_ev_fd, 1);
Jens Axboe8c838782019-03-12 15:48:16 -0600695}
696
Jens Axboec4a2ed72019-11-21 21:01:26 -0700697/* Returns true if there are no backlogged entries after the flush */
698static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
Jens Axboe2b188cc2019-01-07 10:46:33 -0700699{
Jens Axboe1d7bb1d2019-11-06 11:31:17 -0700700 struct io_rings *rings = ctx->rings;
701 struct io_uring_cqe *cqe;
702 struct io_kiocb *req;
703 unsigned long flags;
704 LIST_HEAD(list);
705
706 if (!force) {
707 if (list_empty_careful(&ctx->cq_overflow_list))
Jens Axboec4a2ed72019-11-21 21:01:26 -0700708 return true;
Jens Axboe1d7bb1d2019-11-06 11:31:17 -0700709 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
710 rings->cq_ring_entries))
Jens Axboec4a2ed72019-11-21 21:01:26 -0700711 return false;
Jens Axboe1d7bb1d2019-11-06 11:31:17 -0700712 }
713
714 spin_lock_irqsave(&ctx->completion_lock, flags);
715
716 /* if force is set, the ring is going away. always drop after that */
717 if (force)
718 ctx->cq_overflow_flushed = true;
719
Jens Axboec4a2ed72019-11-21 21:01:26 -0700720 cqe = NULL;
Jens Axboe1d7bb1d2019-11-06 11:31:17 -0700721 while (!list_empty(&ctx->cq_overflow_list)) {
722 cqe = io_get_cqring(ctx);
723 if (!cqe && !force)
724 break;
725
726 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
727 list);
728 list_move(&req->list, &list);
729 if (cqe) {
730 WRITE_ONCE(cqe->user_data, req->user_data);
731 WRITE_ONCE(cqe->res, req->result);
732 WRITE_ONCE(cqe->flags, 0);
733 } else {
734 WRITE_ONCE(ctx->rings->cq_overflow,
735 atomic_inc_return(&ctx->cached_cq_overflow));
736 }
737 }
738
739 io_commit_cqring(ctx);
740 spin_unlock_irqrestore(&ctx->completion_lock, flags);
741 io_cqring_ev_posted(ctx);
742
743 while (!list_empty(&list)) {
744 req = list_first_entry(&list, struct io_kiocb, list);
745 list_del(&req->list);
Jackie Liuec9c02a2019-11-08 23:50:36 +0800746 io_put_req(req);
Jens Axboe1d7bb1d2019-11-06 11:31:17 -0700747 }
Jens Axboec4a2ed72019-11-21 21:01:26 -0700748
749 return cqe != NULL;
Jens Axboe1d7bb1d2019-11-06 11:31:17 -0700750}
751
Jens Axboe78e19bb2019-11-06 15:21:34 -0700752static void io_cqring_fill_event(struct io_kiocb *req, long res)
Jens Axboe2b188cc2019-01-07 10:46:33 -0700753{
Jens Axboe78e19bb2019-11-06 15:21:34 -0700754 struct io_ring_ctx *ctx = req->ctx;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700755 struct io_uring_cqe *cqe;
756
Jens Axboe78e19bb2019-11-06 15:21:34 -0700757 trace_io_uring_complete(ctx, req->user_data, res);
Jens Axboe51c3ff62019-11-03 06:52:50 -0700758
Jens Axboe2b188cc2019-01-07 10:46:33 -0700759 /*
760 * If we can't get a cq entry, userspace overflowed the
761 * submission (by quite a lot). Increment the overflow count in
762 * the ring.
763 */
764 cqe = io_get_cqring(ctx);
Jens Axboe1d7bb1d2019-11-06 11:31:17 -0700765 if (likely(cqe)) {
Jens Axboe78e19bb2019-11-06 15:21:34 -0700766 WRITE_ONCE(cqe->user_data, req->user_data);
Jens Axboe2b188cc2019-01-07 10:46:33 -0700767 WRITE_ONCE(cqe->res, res);
768 WRITE_ONCE(cqe->flags, 0);
Jens Axboe1d7bb1d2019-11-06 11:31:17 -0700769 } else if (ctx->cq_overflow_flushed) {
Jens Axboe2b188cc2019-01-07 10:46:33 -0700770 WRITE_ONCE(ctx->rings->cq_overflow,
771 atomic_inc_return(&ctx->cached_cq_overflow));
Jens Axboe1d7bb1d2019-11-06 11:31:17 -0700772 } else {
773 refcount_inc(&req->refs);
774 req->result = res;
775 list_add_tail(&req->list, &ctx->cq_overflow_list);
Jens Axboe2b188cc2019-01-07 10:46:33 -0700776 }
777}
778
Jens Axboe78e19bb2019-11-06 15:21:34 -0700779static void io_cqring_add_event(struct io_kiocb *req, long res)
Jens Axboe2b188cc2019-01-07 10:46:33 -0700780{
Jens Axboe78e19bb2019-11-06 15:21:34 -0700781 struct io_ring_ctx *ctx = req->ctx;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700782 unsigned long flags;
783
784 spin_lock_irqsave(&ctx->completion_lock, flags);
Jens Axboe78e19bb2019-11-06 15:21:34 -0700785 io_cqring_fill_event(req, res);
Jens Axboe2b188cc2019-01-07 10:46:33 -0700786 io_commit_cqring(ctx);
787 spin_unlock_irqrestore(&ctx->completion_lock, flags);
788
Jens Axboe8c838782019-03-12 15:48:16 -0600789 io_cqring_ev_posted(ctx);
Jens Axboe2b188cc2019-01-07 10:46:33 -0700790}
791
Jens Axboe0ddf92e2019-11-08 08:52:53 -0700792static inline bool io_is_fallback_req(struct io_kiocb *req)
793{
794 return req == (struct io_kiocb *)
795 ((unsigned long) req->ctx->fallback_req & ~1UL);
796}
797
798static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
799{
800 struct io_kiocb *req;
801
802 req = ctx->fallback_req;
803 if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req))
804 return req;
805
806 return NULL;
807}
808
Jens Axboe2579f912019-01-09 09:10:43 -0700809static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
810 struct io_submit_state *state)
Jens Axboe2b188cc2019-01-07 10:46:33 -0700811{
Jens Axboefd6fab22019-03-14 16:30:06 -0600812 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700813 struct io_kiocb *req;
814
815 if (!percpu_ref_tryget(&ctx->refs))
816 return NULL;
817
Jens Axboe2579f912019-01-09 09:10:43 -0700818 if (!state) {
Jens Axboefd6fab22019-03-14 16:30:06 -0600819 req = kmem_cache_alloc(req_cachep, gfp);
Jens Axboe2579f912019-01-09 09:10:43 -0700820 if (unlikely(!req))
Jens Axboe0ddf92e2019-11-08 08:52:53 -0700821 goto fallback;
Jens Axboe2579f912019-01-09 09:10:43 -0700822 } else if (!state->free_reqs) {
823 size_t sz;
824 int ret;
825
826 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
Jens Axboefd6fab22019-03-14 16:30:06 -0600827 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
828
829 /*
830 * Bulk alloc is all-or-nothing. If we fail to get a batch,
831 * retry single alloc to be on the safe side.
832 */
833 if (unlikely(ret <= 0)) {
834 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
835 if (!state->reqs[0])
Jens Axboe0ddf92e2019-11-08 08:52:53 -0700836 goto fallback;
Jens Axboefd6fab22019-03-14 16:30:06 -0600837 ret = 1;
838 }
Jens Axboe2579f912019-01-09 09:10:43 -0700839 state->free_reqs = ret - 1;
840 state->cur_req = 1;
841 req = state->reqs[0];
842 } else {
843 req = state->reqs[state->cur_req];
844 state->free_reqs--;
845 state->cur_req++;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700846 }
847
Jens Axboe0ddf92e2019-11-08 08:52:53 -0700848got_it:
Jens Axboe1a6b74f2019-12-02 10:33:15 -0700849 req->io = NULL;
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +0300850 req->ring_file = NULL;
Jens Axboe60c112b2019-06-21 10:20:18 -0600851 req->file = NULL;
Jens Axboe2579f912019-01-09 09:10:43 -0700852 req->ctx = ctx;
853 req->flags = 0;
Jens Axboee65ef562019-03-12 10:16:44 -0600854 /* one is dropped after submission, the other at completion */
855 refcount_set(&req->refs, 2);
Jens Axboe9e645e112019-05-10 16:07:28 -0600856 req->result = 0;
Jens Axboe561fb042019-10-24 07:25:42 -0600857 INIT_IO_WORK(&req->work, io_wq_submit_work);
Jens Axboe2579f912019-01-09 09:10:43 -0700858 return req;
Jens Axboe0ddf92e2019-11-08 08:52:53 -0700859fallback:
860 req = io_get_fallback_req(ctx);
861 if (req)
862 goto got_it;
Pavel Begunkov6805b322019-10-08 02:18:42 +0300863 percpu_ref_put(&ctx->refs);
Jens Axboe2b188cc2019-01-07 10:46:33 -0700864 return NULL;
865}
866
Jens Axboedef596e2019-01-09 08:59:42 -0700867static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr)
868{
869 if (*nr) {
870 kmem_cache_free_bulk(req_cachep, *nr, reqs);
Pavel Begunkov6805b322019-10-08 02:18:42 +0300871 percpu_ref_put_many(&ctx->refs, *nr);
Jens Axboedef596e2019-01-09 08:59:42 -0700872 *nr = 0;
873 }
874}
875
Jens Axboe9e645e112019-05-10 16:07:28 -0600876static void __io_free_req(struct io_kiocb *req)
Jens Axboe2b188cc2019-01-07 10:46:33 -0700877{
Jens Axboefcb323c2019-10-24 12:39:47 -0600878 struct io_ring_ctx *ctx = req->ctx;
879
Jens Axboe1a6b74f2019-12-02 10:33:15 -0700880 if (req->io)
881 kfree(req->io);
Jens Axboe09bb8392019-03-13 12:39:28 -0600882 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
883 fput(req->file);
Jens Axboefcb323c2019-10-24 12:39:47 -0600884 if (req->flags & REQ_F_INFLIGHT) {
885 unsigned long flags;
886
887 spin_lock_irqsave(&ctx->inflight_lock, flags);
888 list_del(&req->inflight_entry);
889 if (waitqueue_active(&ctx->inflight_wait))
890 wake_up(&ctx->inflight_wait);
891 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
892 }
893 percpu_ref_put(&ctx->refs);
Jens Axboe0ddf92e2019-11-08 08:52:53 -0700894 if (likely(!io_is_fallback_req(req)))
895 kmem_cache_free(req_cachep, req);
896 else
897 clear_bit_unlock(0, (unsigned long *) ctx->fallback_req);
Jens Axboee65ef562019-03-12 10:16:44 -0600898}
899
Jackie Liua197f662019-11-08 08:09:12 -0700900static bool io_link_cancel_timeout(struct io_kiocb *req)
Jens Axboe9e645e112019-05-10 16:07:28 -0600901{
Jackie Liua197f662019-11-08 08:09:12 -0700902 struct io_ring_ctx *ctx = req->ctx;
Jens Axboe2665abf2019-11-05 12:40:47 -0700903 int ret;
904
Jens Axboe2d283902019-12-04 11:08:05 -0700905 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
Jens Axboe2665abf2019-11-05 12:40:47 -0700906 if (ret != -1) {
Jens Axboe78e19bb2019-11-06 15:21:34 -0700907 io_cqring_fill_event(req, -ECANCELED);
Jens Axboe2665abf2019-11-05 12:40:47 -0700908 io_commit_cqring(ctx);
909 req->flags &= ~REQ_F_LINK;
Jackie Liuec9c02a2019-11-08 23:50:36 +0800910 io_put_req(req);
Jens Axboe2665abf2019-11-05 12:40:47 -0700911 return true;
912 }
913
914 return false;
915}
916
Jens Axboeba816ad2019-09-28 11:36:45 -0600917static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
Jens Axboe9e645e112019-05-10 16:07:28 -0600918{
Jens Axboe2665abf2019-11-05 12:40:47 -0700919 struct io_ring_ctx *ctx = req->ctx;
Jens Axboe2665abf2019-11-05 12:40:47 -0700920 bool wake_ev = false;
Jens Axboe9e645e112019-05-10 16:07:28 -0600921
Jens Axboe4d7dd462019-11-20 13:03:52 -0700922 /* Already got next link */
923 if (req->flags & REQ_F_LINK_NEXT)
924 return;
925
Jens Axboe9e645e112019-05-10 16:07:28 -0600926 /*
927 * The list should never be empty when we are called here. But could
928 * potentially happen if the chain is messed up, check to be on the
929 * safe side.
930 */
Pavel Begunkov44932332019-12-05 16:16:35 +0300931 while (!list_empty(&req->link_list)) {
932 struct io_kiocb *nxt = list_first_entry(&req->link_list,
933 struct io_kiocb, link_list);
Jens Axboe94ae5e72019-11-14 19:39:52 -0700934
Pavel Begunkov44932332019-12-05 16:16:35 +0300935 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
936 (nxt->flags & REQ_F_TIMEOUT))) {
937 list_del_init(&nxt->link_list);
Jens Axboe94ae5e72019-11-14 19:39:52 -0700938 wake_ev |= io_link_cancel_timeout(nxt);
Jens Axboe94ae5e72019-11-14 19:39:52 -0700939 req->flags &= ~REQ_F_LINK_TIMEOUT;
940 continue;
941 }
Jens Axboe9e645e112019-05-10 16:07:28 -0600942
Pavel Begunkov44932332019-12-05 16:16:35 +0300943 list_del_init(&req->link_list);
944 if (!list_empty(&nxt->link_list))
945 nxt->flags |= REQ_F_LINK;
Pavel Begunkovb18fdf72019-11-21 23:21:02 +0300946 *nxtptr = nxt;
Jens Axboe94ae5e72019-11-14 19:39:52 -0700947 break;
Jens Axboe9e645e112019-05-10 16:07:28 -0600948 }
Jens Axboe2665abf2019-11-05 12:40:47 -0700949
Jens Axboe4d7dd462019-11-20 13:03:52 -0700950 req->flags |= REQ_F_LINK_NEXT;
Jens Axboe2665abf2019-11-05 12:40:47 -0700951 if (wake_ev)
952 io_cqring_ev_posted(ctx);
Jens Axboe9e645e112019-05-10 16:07:28 -0600953}
954
955/*
956 * Called if REQ_F_LINK is set, and we fail the head request
957 */
958static void io_fail_links(struct io_kiocb *req)
959{
Jens Axboe2665abf2019-11-05 12:40:47 -0700960 struct io_ring_ctx *ctx = req->ctx;
Jens Axboe2665abf2019-11-05 12:40:47 -0700961 unsigned long flags;
962
963 spin_lock_irqsave(&ctx->completion_lock, flags);
Jens Axboe9e645e112019-05-10 16:07:28 -0600964
965 while (!list_empty(&req->link_list)) {
Pavel Begunkov44932332019-12-05 16:16:35 +0300966 struct io_kiocb *link = list_first_entry(&req->link_list,
967 struct io_kiocb, link_list);
Jens Axboe9e645e112019-05-10 16:07:28 -0600968
Pavel Begunkov44932332019-12-05 16:16:35 +0300969 list_del_init(&link->link_list);
Dmitrii Dolgovc826bd72019-10-15 19:02:01 +0200970 trace_io_uring_fail_link(req, link);
Jens Axboe2665abf2019-11-05 12:40:47 -0700971
972 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +0300973 link->sqe->opcode == IORING_OP_LINK_TIMEOUT) {
Jackie Liua197f662019-11-08 08:09:12 -0700974 io_link_cancel_timeout(link);
Jens Axboe2665abf2019-11-05 12:40:47 -0700975 } else {
Jens Axboe78e19bb2019-11-06 15:21:34 -0700976 io_cqring_fill_event(link, -ECANCELED);
Jens Axboe978db572019-11-14 22:39:04 -0700977 __io_double_put_req(link);
Jens Axboe2665abf2019-11-05 12:40:47 -0700978 }
Jens Axboe5d960722019-11-19 15:31:28 -0700979 req->flags &= ~REQ_F_LINK_TIMEOUT;
Jens Axboe9e645e112019-05-10 16:07:28 -0600980 }
Jens Axboe2665abf2019-11-05 12:40:47 -0700981
982 io_commit_cqring(ctx);
983 spin_unlock_irqrestore(&ctx->completion_lock, flags);
984 io_cqring_ev_posted(ctx);
Jens Axboe9e645e112019-05-10 16:07:28 -0600985}
986
Jens Axboe4d7dd462019-11-20 13:03:52 -0700987static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
Jens Axboe9e645e112019-05-10 16:07:28 -0600988{
Jens Axboe4d7dd462019-11-20 13:03:52 -0700989 if (likely(!(req->flags & REQ_F_LINK)))
Jens Axboe2665abf2019-11-05 12:40:47 -0700990 return;
Jens Axboe2665abf2019-11-05 12:40:47 -0700991
Jens Axboe9e645e112019-05-10 16:07:28 -0600992 /*
993 * If LINK is set, we have dependent requests in this chain. If we
994 * didn't fail this request, queue the first one up, moving any other
995 * dependencies to the next request. In case of failure, fail the rest
996 * of the chain.
997 */
Jens Axboe2665abf2019-11-05 12:40:47 -0700998 if (req->flags & REQ_F_FAIL_LINK) {
999 io_fail_links(req);
Jens Axboe7c9e7f02019-11-12 08:15:53 -07001000 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1001 REQ_F_LINK_TIMEOUT) {
Jens Axboe2665abf2019-11-05 12:40:47 -07001002 struct io_ring_ctx *ctx = req->ctx;
1003 unsigned long flags;
1004
1005 /*
1006 * If this is a timeout link, we could be racing with the
1007 * timeout timer. Grab the completion lock for this case to
Jens Axboe7c9e7f02019-11-12 08:15:53 -07001008 * protect against that.
Jens Axboe2665abf2019-11-05 12:40:47 -07001009 */
1010 spin_lock_irqsave(&ctx->completion_lock, flags);
1011 io_req_link_next(req, nxt);
1012 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1013 } else {
1014 io_req_link_next(req, nxt);
Jens Axboe9e645e112019-05-10 16:07:28 -06001015 }
Jens Axboe4d7dd462019-11-20 13:03:52 -07001016}
Jens Axboe9e645e112019-05-10 16:07:28 -06001017
Jackie Liuc69f8db2019-11-09 11:00:08 +08001018static void io_free_req(struct io_kiocb *req)
1019{
Pavel Begunkov944e58b2019-11-21 23:21:01 +03001020 struct io_kiocb *nxt = NULL;
1021
1022 io_req_find_next(req, &nxt);
Pavel Begunkov70cf9f32019-11-21 23:21:00 +03001023 __io_free_req(req);
Pavel Begunkov944e58b2019-11-21 23:21:01 +03001024
1025 if (nxt)
1026 io_queue_async_work(nxt);
Jackie Liuc69f8db2019-11-09 11:00:08 +08001027}
1028
Jens Axboeba816ad2019-09-28 11:36:45 -06001029/*
1030 * Drop reference to request, return next in chain (if there is one) if this
1031 * was the last reference to this request.
1032 */
Pavel Begunkovf9bd67f2019-11-21 23:21:03 +03001033__attribute__((nonnull))
Jackie Liuec9c02a2019-11-08 23:50:36 +08001034static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
Jens Axboee65ef562019-03-12 10:16:44 -06001035{
Pavel Begunkovf9bd67f2019-11-21 23:21:03 +03001036 io_req_find_next(req, nxtptr);
Jens Axboe4d7dd462019-11-20 13:03:52 -07001037
Jens Axboee65ef562019-03-12 10:16:44 -06001038 if (refcount_dec_and_test(&req->refs))
Jens Axboe4d7dd462019-11-20 13:03:52 -07001039 __io_free_req(req);
Jens Axboe2b188cc2019-01-07 10:46:33 -07001040}
1041
Jens Axboe2b188cc2019-01-07 10:46:33 -07001042static void io_put_req(struct io_kiocb *req)
1043{
Jens Axboedef596e2019-01-09 08:59:42 -07001044 if (refcount_dec_and_test(&req->refs))
1045 io_free_req(req);
1046}
1047
Jens Axboe978db572019-11-14 22:39:04 -07001048/*
1049 * Must only be used if we don't need to care about links, usually from
1050 * within the completion handling itself.
1051 */
1052static void __io_double_put_req(struct io_kiocb *req)
Jens Axboea3a0e432019-08-20 11:03:11 -06001053{
Jens Axboe78e19bb2019-11-06 15:21:34 -07001054 /* drop both submit and complete references */
1055 if (refcount_sub_and_test(2, &req->refs))
1056 __io_free_req(req);
1057}
1058
Jens Axboe978db572019-11-14 22:39:04 -07001059static void io_double_put_req(struct io_kiocb *req)
1060{
1061 /* drop both submit and complete references */
1062 if (refcount_sub_and_test(2, &req->refs))
1063 io_free_req(req);
1064}
1065
Jens Axboe1d7bb1d2019-11-06 11:31:17 -07001066static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
Jens Axboea3a0e432019-08-20 11:03:11 -06001067{
Jens Axboe84f97dc2019-11-06 11:27:53 -07001068 struct io_rings *rings = ctx->rings;
1069
Jens Axboe1d7bb1d2019-11-06 11:31:17 -07001070 /*
1071 * noflush == true is from the waitqueue handler, just ensure we wake
1072 * up the task, and the next invocation will flush the entries. We
1073 * cannot safely to it from here.
1074 */
1075 if (noflush && !list_empty(&ctx->cq_overflow_list))
1076 return -1U;
1077
1078 io_cqring_overflow_flush(ctx, false);
1079
Jens Axboea3a0e432019-08-20 11:03:11 -06001080 /* See comment at the top of this file */
1081 smp_rmb();
Hristo Venev75b28af2019-08-26 17:23:46 +00001082 return READ_ONCE(rings->cq.tail) - READ_ONCE(rings->cq.head);
Jens Axboea3a0e432019-08-20 11:03:11 -06001083}
1084
Pavel Begunkovfb5ccc92019-10-25 12:31:30 +03001085static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1086{
1087 struct io_rings *rings = ctx->rings;
1088
1089 /* make sure SQ entry isn't read before tail */
1090 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1091}
1092
Jens Axboedef596e2019-01-09 08:59:42 -07001093/*
1094 * Find and free completed poll iocbs
1095 */
1096static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1097 struct list_head *done)
1098{
1099 void *reqs[IO_IOPOLL_BATCH];
1100 struct io_kiocb *req;
Jens Axboe09bb8392019-03-13 12:39:28 -06001101 int to_free;
Jens Axboedef596e2019-01-09 08:59:42 -07001102
Jens Axboe09bb8392019-03-13 12:39:28 -06001103 to_free = 0;
Jens Axboedef596e2019-01-09 08:59:42 -07001104 while (!list_empty(done)) {
1105 req = list_first_entry(done, struct io_kiocb, list);
1106 list_del(&req->list);
1107
Jens Axboe78e19bb2019-11-06 15:21:34 -07001108 io_cqring_fill_event(req, req->result);
Jens Axboedef596e2019-01-09 08:59:42 -07001109 (*nr_events)++;
1110
Jens Axboe09bb8392019-03-13 12:39:28 -06001111 if (refcount_dec_and_test(&req->refs)) {
1112 /* If we're not using fixed files, we have to pair the
1113 * completion part with the file put. Use regular
1114 * completions for those, only batch free for fixed
Jens Axboe9e645e112019-05-10 16:07:28 -06001115 * file and non-linked commands.
Jens Axboe09bb8392019-03-13 12:39:28 -06001116 */
Jens Axboe1a6b74f2019-12-02 10:33:15 -07001117 if (((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) ==
1118 REQ_F_FIXED_FILE) && !io_is_fallback_req(req) &&
1119 !req->io) {
Jens Axboe09bb8392019-03-13 12:39:28 -06001120 reqs[to_free++] = req;
1121 if (to_free == ARRAY_SIZE(reqs))
1122 io_free_req_many(ctx, reqs, &to_free);
Jens Axboe6b063142019-01-10 22:13:58 -07001123 } else {
Jens Axboe09bb8392019-03-13 12:39:28 -06001124 io_free_req(req);
Jens Axboe6b063142019-01-10 22:13:58 -07001125 }
Jens Axboe9a56a232019-01-09 09:06:50 -07001126 }
Jens Axboedef596e2019-01-09 08:59:42 -07001127 }
Jens Axboedef596e2019-01-09 08:59:42 -07001128
Jens Axboe09bb8392019-03-13 12:39:28 -06001129 io_commit_cqring(ctx);
Jens Axboedef596e2019-01-09 08:59:42 -07001130 io_free_req_many(ctx, reqs, &to_free);
1131}
1132
1133static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1134 long min)
1135{
1136 struct io_kiocb *req, *tmp;
1137 LIST_HEAD(done);
1138 bool spin;
1139 int ret;
1140
1141 /*
1142 * Only spin for completions if we don't have multiple devices hanging
1143 * off our complete list, and we're under the requested amount.
1144 */
1145 spin = !ctx->poll_multi_file && *nr_events < min;
1146
1147 ret = 0;
1148 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1149 struct kiocb *kiocb = &req->rw;
1150
1151 /*
1152 * Move completed entries to our local list. If we find a
1153 * request that requires polling, break out and complete
1154 * the done list first, if we have entries there.
1155 */
1156 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
1157 list_move_tail(&req->list, &done);
1158 continue;
1159 }
1160 if (!list_empty(&done))
1161 break;
1162
1163 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1164 if (ret < 0)
1165 break;
1166
1167 if (ret && spin)
1168 spin = false;
1169 ret = 0;
1170 }
1171
1172 if (!list_empty(&done))
1173 io_iopoll_complete(ctx, nr_events, &done);
1174
1175 return ret;
1176}
1177
1178/*
1179 * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a
1180 * non-spinning poll check - we'll still enter the driver poll loop, but only
1181 * as a non-spinning completion check.
1182 */
1183static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1184 long min)
1185{
Jens Axboe08f54392019-08-21 22:19:11 -06001186 while (!list_empty(&ctx->poll_list) && !need_resched()) {
Jens Axboedef596e2019-01-09 08:59:42 -07001187 int ret;
1188
1189 ret = io_do_iopoll(ctx, nr_events, min);
1190 if (ret < 0)
1191 return ret;
1192 if (!min || *nr_events >= min)
1193 return 0;
1194 }
1195
1196 return 1;
1197}
1198
1199/*
1200 * We can't just wait for polled events to come to us, we have to actively
1201 * find and complete them.
1202 */
1203static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1204{
1205 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1206 return;
1207
1208 mutex_lock(&ctx->uring_lock);
1209 while (!list_empty(&ctx->poll_list)) {
1210 unsigned int nr_events = 0;
1211
1212 io_iopoll_getevents(ctx, &nr_events, 1);
Jens Axboe08f54392019-08-21 22:19:11 -06001213
1214 /*
1215 * Ensure we allow local-to-the-cpu processing to take place,
1216 * in this case we need to ensure that we reap all events.
1217 */
1218 cond_resched();
Jens Axboedef596e2019-01-09 08:59:42 -07001219 }
1220 mutex_unlock(&ctx->uring_lock);
1221}
1222
Jens Axboe2b2ed972019-10-25 10:06:15 -06001223static int __io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1224 long min)
Jens Axboedef596e2019-01-09 08:59:42 -07001225{
Jens Axboe2b2ed972019-10-25 10:06:15 -06001226 int iters = 0, ret = 0;
Jens Axboedef596e2019-01-09 08:59:42 -07001227
1228 do {
1229 int tmin = 0;
1230
Jens Axboe500f9fb2019-08-19 12:15:59 -06001231 /*
Jens Axboea3a0e432019-08-20 11:03:11 -06001232 * Don't enter poll loop if we already have events pending.
1233 * If we do, we can potentially be spinning for commands that
1234 * already triggered a CQE (eg in error).
1235 */
Jens Axboe1d7bb1d2019-11-06 11:31:17 -07001236 if (io_cqring_events(ctx, false))
Jens Axboea3a0e432019-08-20 11:03:11 -06001237 break;
1238
1239 /*
Jens Axboe500f9fb2019-08-19 12:15:59 -06001240 * If a submit got punted to a workqueue, we can have the
1241 * application entering polling for a command before it gets
1242 * issued. That app will hold the uring_lock for the duration
1243 * of the poll right here, so we need to take a breather every
1244 * now and then to ensure that the issue has a chance to add
1245 * the poll to the issued list. Otherwise we can spin here
1246 * forever, while the workqueue is stuck trying to acquire the
1247 * very same mutex.
1248 */
1249 if (!(++iters & 7)) {
1250 mutex_unlock(&ctx->uring_lock);
1251 mutex_lock(&ctx->uring_lock);
1252 }
1253
Jens Axboedef596e2019-01-09 08:59:42 -07001254 if (*nr_events < min)
1255 tmin = min - *nr_events;
1256
1257 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1258 if (ret <= 0)
1259 break;
1260 ret = 0;
1261 } while (min && !*nr_events && !need_resched());
1262
Jens Axboe2b2ed972019-10-25 10:06:15 -06001263 return ret;
1264}
1265
1266static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1267 long min)
1268{
1269 int ret;
1270
1271 /*
1272 * We disallow the app entering submit/complete with polling, but we
1273 * still need to lock the ring to prevent racing with polled issue
1274 * that got punted to a workqueue.
1275 */
1276 mutex_lock(&ctx->uring_lock);
1277 ret = __io_iopoll_check(ctx, nr_events, min);
Jens Axboe500f9fb2019-08-19 12:15:59 -06001278 mutex_unlock(&ctx->uring_lock);
Jens Axboedef596e2019-01-09 08:59:42 -07001279 return ret;
1280}
1281
Jens Axboe491381ce2019-10-17 09:20:46 -06001282static void kiocb_end_write(struct io_kiocb *req)
Jens Axboe2b188cc2019-01-07 10:46:33 -07001283{
Jens Axboe491381ce2019-10-17 09:20:46 -06001284 /*
1285 * Tell lockdep we inherited freeze protection from submission
1286 * thread.
1287 */
1288 if (req->flags & REQ_F_ISREG) {
1289 struct inode *inode = file_inode(req->file);
Jens Axboe2b188cc2019-01-07 10:46:33 -07001290
Jens Axboe491381ce2019-10-17 09:20:46 -06001291 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
Jens Axboe2b188cc2019-01-07 10:46:33 -07001292 }
Jens Axboe491381ce2019-10-17 09:20:46 -06001293 file_end_write(req->file);
Jens Axboe2b188cc2019-01-07 10:46:33 -07001294}
1295
Jens Axboe4e88d6e2019-12-07 20:59:47 -07001296static inline void req_set_fail_links(struct io_kiocb *req)
1297{
1298 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1299 req->flags |= REQ_F_FAIL_LINK;
1300}
1301
Jens Axboeba816ad2019-09-28 11:36:45 -06001302static void io_complete_rw_common(struct kiocb *kiocb, long res)
Jens Axboe2b188cc2019-01-07 10:46:33 -07001303{
1304 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1305
Jens Axboe491381ce2019-10-17 09:20:46 -06001306 if (kiocb->ki_flags & IOCB_WRITE)
1307 kiocb_end_write(req);
Jens Axboe2b188cc2019-01-07 10:46:33 -07001308
Jens Axboe4e88d6e2019-12-07 20:59:47 -07001309 if (res != req->result)
1310 req_set_fail_links(req);
Jens Axboe78e19bb2019-11-06 15:21:34 -07001311 io_cqring_add_event(req, res);
Jens Axboeba816ad2019-09-28 11:36:45 -06001312}
1313
1314static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1315{
1316 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1317
1318 io_complete_rw_common(kiocb, res);
Jens Axboee65ef562019-03-12 10:16:44 -06001319 io_put_req(req);
Jens Axboe2b188cc2019-01-07 10:46:33 -07001320}
1321
Jens Axboeba816ad2019-09-28 11:36:45 -06001322static struct io_kiocb *__io_complete_rw(struct kiocb *kiocb, long res)
1323{
1324 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
Jackie Liuec9c02a2019-11-08 23:50:36 +08001325 struct io_kiocb *nxt = NULL;
Jens Axboeba816ad2019-09-28 11:36:45 -06001326
1327 io_complete_rw_common(kiocb, res);
Jackie Liuec9c02a2019-11-08 23:50:36 +08001328 io_put_req_find_next(req, &nxt);
1329
1330 return nxt;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001331}
1332
Jens Axboedef596e2019-01-09 08:59:42 -07001333static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1334{
1335 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1336
Jens Axboe491381ce2019-10-17 09:20:46 -06001337 if (kiocb->ki_flags & IOCB_WRITE)
1338 kiocb_end_write(req);
Jens Axboedef596e2019-01-09 08:59:42 -07001339
Jens Axboe4e88d6e2019-12-07 20:59:47 -07001340 if (res != req->result)
1341 req_set_fail_links(req);
Jens Axboe9e645e112019-05-10 16:07:28 -06001342 req->result = res;
Jens Axboedef596e2019-01-09 08:59:42 -07001343 if (res != -EAGAIN)
1344 req->flags |= REQ_F_IOPOLL_COMPLETED;
1345}
1346
1347/*
1348 * After the iocb has been issued, it's safe to be found on the poll list.
1349 * Adding the kiocb to the list AFTER submission ensures that we don't
1350 * find it from a io_iopoll_getevents() thread before the issuer is done
1351 * accessing the kiocb cookie.
1352 */
1353static void io_iopoll_req_issued(struct io_kiocb *req)
1354{
1355 struct io_ring_ctx *ctx = req->ctx;
1356
1357 /*
1358 * Track whether we have multiple files in our lists. This will impact
1359 * how we do polling eventually, not spinning if we're on potentially
1360 * different devices.
1361 */
1362 if (list_empty(&ctx->poll_list)) {
1363 ctx->poll_multi_file = false;
1364 } else if (!ctx->poll_multi_file) {
1365 struct io_kiocb *list_req;
1366
1367 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
1368 list);
1369 if (list_req->rw.ki_filp != req->rw.ki_filp)
1370 ctx->poll_multi_file = true;
1371 }
1372
1373 /*
1374 * For fast devices, IO may have already completed. If it has, add
1375 * it to the front so we find it first.
1376 */
1377 if (req->flags & REQ_F_IOPOLL_COMPLETED)
1378 list_add(&req->list, &ctx->poll_list);
1379 else
1380 list_add_tail(&req->list, &ctx->poll_list);
1381}
1382
Jens Axboe3d6770f2019-04-13 11:50:54 -06001383static void io_file_put(struct io_submit_state *state)
Jens Axboe9a56a232019-01-09 09:06:50 -07001384{
Jens Axboe3d6770f2019-04-13 11:50:54 -06001385 if (state->file) {
Jens Axboe9a56a232019-01-09 09:06:50 -07001386 int diff = state->has_refs - state->used_refs;
1387
1388 if (diff)
1389 fput_many(state->file, diff);
1390 state->file = NULL;
1391 }
1392}
1393
1394/*
1395 * Get as many references to a file as we have IOs left in this submission,
1396 * assuming most submissions are for one file, or at least that each file
1397 * has more than one submission.
1398 */
1399static struct file *io_file_get(struct io_submit_state *state, int fd)
1400{
1401 if (!state)
1402 return fget(fd);
1403
1404 if (state->file) {
1405 if (state->fd == fd) {
1406 state->used_refs++;
1407 state->ios_left--;
1408 return state->file;
1409 }
Jens Axboe3d6770f2019-04-13 11:50:54 -06001410 io_file_put(state);
Jens Axboe9a56a232019-01-09 09:06:50 -07001411 }
1412 state->file = fget_many(fd, state->ios_left);
1413 if (!state->file)
1414 return NULL;
1415
1416 state->fd = fd;
1417 state->has_refs = state->ios_left;
1418 state->used_refs = 1;
1419 state->ios_left--;
1420 return state->file;
1421}
1422
Jens Axboe2b188cc2019-01-07 10:46:33 -07001423/*
1424 * If we tracked the file through the SCM inflight mechanism, we could support
1425 * any file. For now, just ensure that anything potentially problematic is done
1426 * inline.
1427 */
1428static bool io_file_supports_async(struct file *file)
1429{
1430 umode_t mode = file_inode(file)->i_mode;
1431
1432 if (S_ISBLK(mode) || S_ISCHR(mode))
1433 return true;
1434 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
1435 return true;
1436
1437 return false;
1438}
1439
Pavel Begunkov267bc902019-11-07 01:41:08 +03001440static int io_prep_rw(struct io_kiocb *req, bool force_nonblock)
Jens Axboe2b188cc2019-01-07 10:46:33 -07001441{
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03001442 const struct io_uring_sqe *sqe = req->sqe;
Jens Axboedef596e2019-01-09 08:59:42 -07001443 struct io_ring_ctx *ctx = req->ctx;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001444 struct kiocb *kiocb = &req->rw;
Jens Axboe09bb8392019-03-13 12:39:28 -06001445 unsigned ioprio;
1446 int ret;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001447
Jens Axboe09bb8392019-03-13 12:39:28 -06001448 if (!req->file)
1449 return -EBADF;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001450
Jens Axboe491381ce2019-10-17 09:20:46 -06001451 if (S_ISREG(file_inode(req->file)->i_mode))
1452 req->flags |= REQ_F_ISREG;
1453
Jens Axboe2b188cc2019-01-07 10:46:33 -07001454 kiocb->ki_pos = READ_ONCE(sqe->off);
1455 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
1456 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
1457
1458 ioprio = READ_ONCE(sqe->ioprio);
1459 if (ioprio) {
1460 ret = ioprio_check_cap(ioprio);
1461 if (ret)
Jens Axboe09bb8392019-03-13 12:39:28 -06001462 return ret;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001463
1464 kiocb->ki_ioprio = ioprio;
1465 } else
1466 kiocb->ki_ioprio = get_current_ioprio();
1467
1468 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
1469 if (unlikely(ret))
Jens Axboe09bb8392019-03-13 12:39:28 -06001470 return ret;
Stefan Bühler8449eed2019-04-27 20:34:19 +02001471
1472 /* don't allow async punt if RWF_NOWAIT was requested */
Jens Axboe491381ce2019-10-17 09:20:46 -06001473 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
1474 (req->file->f_flags & O_NONBLOCK))
Stefan Bühler8449eed2019-04-27 20:34:19 +02001475 req->flags |= REQ_F_NOWAIT;
1476
1477 if (force_nonblock)
Jens Axboe2b188cc2019-01-07 10:46:33 -07001478 kiocb->ki_flags |= IOCB_NOWAIT;
Stefan Bühler8449eed2019-04-27 20:34:19 +02001479
Jens Axboedef596e2019-01-09 08:59:42 -07001480 if (ctx->flags & IORING_SETUP_IOPOLL) {
Jens Axboedef596e2019-01-09 08:59:42 -07001481 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
1482 !kiocb->ki_filp->f_op->iopoll)
Jens Axboe09bb8392019-03-13 12:39:28 -06001483 return -EOPNOTSUPP;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001484
Jens Axboedef596e2019-01-09 08:59:42 -07001485 kiocb->ki_flags |= IOCB_HIPRI;
1486 kiocb->ki_complete = io_complete_rw_iopoll;
Jens Axboe6873e0b2019-10-30 13:53:09 -06001487 req->result = 0;
Jens Axboedef596e2019-01-09 08:59:42 -07001488 } else {
Jens Axboe09bb8392019-03-13 12:39:28 -06001489 if (kiocb->ki_flags & IOCB_HIPRI)
1490 return -EINVAL;
Jens Axboedef596e2019-01-09 08:59:42 -07001491 kiocb->ki_complete = io_complete_rw;
1492 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07001493 return 0;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001494}
1495
1496static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
1497{
1498 switch (ret) {
1499 case -EIOCBQUEUED:
1500 break;
1501 case -ERESTARTSYS:
1502 case -ERESTARTNOINTR:
1503 case -ERESTARTNOHAND:
1504 case -ERESTART_RESTARTBLOCK:
1505 /*
1506 * We can't just restart the syscall, since previously
1507 * submitted sqes may already be in progress. Just fail this
1508 * IO with EINTR.
1509 */
1510 ret = -EINTR;
1511 /* fall through */
1512 default:
1513 kiocb->ki_complete(kiocb, ret, 0);
1514 }
1515}
1516
Jens Axboeba816ad2019-09-28 11:36:45 -06001517static void kiocb_done(struct kiocb *kiocb, ssize_t ret, struct io_kiocb **nxt,
1518 bool in_async)
1519{
Pavel Begunkovf9bd67f2019-11-21 23:21:03 +03001520 if (in_async && ret >= 0 && kiocb->ki_complete == io_complete_rw)
Jens Axboeba816ad2019-09-28 11:36:45 -06001521 *nxt = __io_complete_rw(kiocb, ret);
1522 else
1523 io_rw_done(kiocb, ret);
1524}
1525
Pavel Begunkov7d009162019-11-25 23:14:40 +03001526static ssize_t io_import_fixed(struct io_ring_ctx *ctx, int rw,
1527 const struct io_uring_sqe *sqe,
1528 struct iov_iter *iter)
Jens Axboeedafcce2019-01-09 09:16:05 -07001529{
1530 size_t len = READ_ONCE(sqe->len);
1531 struct io_mapped_ubuf *imu;
1532 unsigned index, buf_index;
1533 size_t offset;
1534 u64 buf_addr;
1535
1536 /* attempt to use fixed buffers without having provided iovecs */
1537 if (unlikely(!ctx->user_bufs))
1538 return -EFAULT;
1539
1540 buf_index = READ_ONCE(sqe->buf_index);
1541 if (unlikely(buf_index >= ctx->nr_user_bufs))
1542 return -EFAULT;
1543
1544 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
1545 imu = &ctx->user_bufs[index];
1546 buf_addr = READ_ONCE(sqe->addr);
1547
1548 /* overflow */
1549 if (buf_addr + len < buf_addr)
1550 return -EFAULT;
1551 /* not inside the mapped region */
1552 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
1553 return -EFAULT;
1554
1555 /*
1556 * May not be a start of buffer, set size appropriately
1557 * and advance us to the beginning.
1558 */
1559 offset = buf_addr - imu->ubuf;
1560 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
Jens Axboebd11b3a2019-07-20 08:37:31 -06001561
1562 if (offset) {
1563 /*
1564 * Don't use iov_iter_advance() here, as it's really slow for
1565 * using the latter parts of a big fixed buffer - it iterates
1566 * over each segment manually. We can cheat a bit here, because
1567 * we know that:
1568 *
1569 * 1) it's a BVEC iter, we set it up
1570 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1571 * first and last bvec
1572 *
1573 * So just find our index, and adjust the iterator afterwards.
1574 * If the offset is within the first bvec (or the whole first
1575 * bvec, just use iov_iter_advance(). This makes it easier
1576 * since we can just skip the first segment, which may not
1577 * be PAGE_SIZE aligned.
1578 */
1579 const struct bio_vec *bvec = imu->bvec;
1580
1581 if (offset <= bvec->bv_len) {
1582 iov_iter_advance(iter, offset);
1583 } else {
1584 unsigned long seg_skip;
1585
1586 /* skip first vec */
1587 offset -= bvec->bv_len;
1588 seg_skip = 1 + (offset >> PAGE_SHIFT);
1589
1590 iter->bvec = bvec + seg_skip;
1591 iter->nr_segs -= seg_skip;
Aleix Roca Nonell99c79f62019-08-15 14:03:22 +02001592 iter->count -= bvec->bv_len + offset;
Jens Axboebd11b3a2019-07-20 08:37:31 -06001593 iter->iov_offset = offset & ~PAGE_MASK;
Jens Axboebd11b3a2019-07-20 08:37:31 -06001594 }
1595 }
1596
Jens Axboe5e559562019-11-13 16:12:46 -07001597 return len;
Jens Axboeedafcce2019-01-09 09:16:05 -07001598}
1599
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03001600static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
1601 struct iovec **iovec, struct iov_iter *iter)
Jens Axboe2b188cc2019-01-07 10:46:33 -07001602{
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03001603 const struct io_uring_sqe *sqe = req->sqe;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001604 void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
1605 size_t sqe_len = READ_ONCE(sqe->len);
Jens Axboeedafcce2019-01-09 09:16:05 -07001606 u8 opcode;
1607
1608 /*
1609 * We're reading ->opcode for the second time, but the first read
1610 * doesn't care whether it's _FIXED or not, so it doesn't matter
1611 * whether ->opcode changes concurrently. The first read does care
1612 * about whether it is a READ or a WRITE, so we don't trust this read
1613 * for that purpose and instead let the caller pass in the read/write
1614 * flag.
1615 */
1616 opcode = READ_ONCE(sqe->opcode);
Pavel Begunkov7d009162019-11-25 23:14:40 +03001617 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
Jens Axboeedafcce2019-01-09 09:16:05 -07001618 *iovec = NULL;
Pavel Begunkov7d009162019-11-25 23:14:40 +03001619 return io_import_fixed(req->ctx, rw, sqe, iter);
Jens Axboeedafcce2019-01-09 09:16:05 -07001620 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07001621
Jens Axboef67676d2019-12-02 11:03:47 -07001622 if (req->io) {
1623 struct io_async_rw *iorw = &req->io->rw;
1624
1625 *iovec = iorw->iov;
1626 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
1627 if (iorw->iov == iorw->fast_iov)
1628 *iovec = NULL;
1629 return iorw->size;
1630 }
1631
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03001632 if (!req->has_user)
Jens Axboe2b188cc2019-01-07 10:46:33 -07001633 return -EFAULT;
1634
1635#ifdef CONFIG_COMPAT
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03001636 if (req->ctx->compat)
Jens Axboe2b188cc2019-01-07 10:46:33 -07001637 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
1638 iovec, iter);
1639#endif
1640
1641 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
1642}
1643
Jens Axboe32960612019-09-23 11:05:34 -06001644/*
1645 * For files that don't have ->read_iter() and ->write_iter(), handle them
1646 * by looping over ->read() or ->write() manually.
1647 */
1648static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
1649 struct iov_iter *iter)
1650{
1651 ssize_t ret = 0;
1652
1653 /*
1654 * Don't support polled IO through this interface, and we can't
1655 * support non-blocking either. For the latter, this just causes
1656 * the kiocb to be handled from an async context.
1657 */
1658 if (kiocb->ki_flags & IOCB_HIPRI)
1659 return -EOPNOTSUPP;
1660 if (kiocb->ki_flags & IOCB_NOWAIT)
1661 return -EAGAIN;
1662
1663 while (iov_iter_count(iter)) {
Pavel Begunkov311ae9e2019-11-24 11:58:24 +03001664 struct iovec iovec;
Jens Axboe32960612019-09-23 11:05:34 -06001665 ssize_t nr;
1666
Pavel Begunkov311ae9e2019-11-24 11:58:24 +03001667 if (!iov_iter_is_bvec(iter)) {
1668 iovec = iov_iter_iovec(iter);
1669 } else {
1670 /* fixed buffers import bvec */
1671 iovec.iov_base = kmap(iter->bvec->bv_page)
1672 + iter->iov_offset;
1673 iovec.iov_len = min(iter->count,
1674 iter->bvec->bv_len - iter->iov_offset);
1675 }
1676
Jens Axboe32960612019-09-23 11:05:34 -06001677 if (rw == READ) {
1678 nr = file->f_op->read(file, iovec.iov_base,
1679 iovec.iov_len, &kiocb->ki_pos);
1680 } else {
1681 nr = file->f_op->write(file, iovec.iov_base,
1682 iovec.iov_len, &kiocb->ki_pos);
1683 }
1684
Pavel Begunkov311ae9e2019-11-24 11:58:24 +03001685 if (iov_iter_is_bvec(iter))
1686 kunmap(iter->bvec->bv_page);
1687
Jens Axboe32960612019-09-23 11:05:34 -06001688 if (nr < 0) {
1689 if (!ret)
1690 ret = nr;
1691 break;
1692 }
1693 ret += nr;
1694 if (nr != iovec.iov_len)
1695 break;
1696 iov_iter_advance(iter, nr);
1697 }
1698
1699 return ret;
1700}
1701
Jens Axboef67676d2019-12-02 11:03:47 -07001702static void io_req_map_io(struct io_kiocb *req, ssize_t io_size,
1703 struct iovec *iovec, struct iovec *fast_iov,
1704 struct iov_iter *iter)
1705{
1706 req->io->rw.nr_segs = iter->nr_segs;
1707 req->io->rw.size = io_size;
1708 req->io->rw.iov = iovec;
1709 if (!req->io->rw.iov) {
1710 req->io->rw.iov = req->io->rw.fast_iov;
1711 memcpy(req->io->rw.iov, fast_iov,
1712 sizeof(struct iovec) * iter->nr_segs);
1713 }
1714}
1715
1716static int io_setup_async_io(struct io_kiocb *req, ssize_t io_size,
1717 struct iovec *iovec, struct iovec *fast_iov,
1718 struct iov_iter *iter)
1719{
1720 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
1721 if (req->io) {
1722 io_req_map_io(req, io_size, iovec, fast_iov, iter);
1723 memcpy(&req->io->sqe, req->sqe, sizeof(req->io->sqe));
1724 req->sqe = &req->io->sqe;
1725 return 0;
1726 }
1727
1728 return -ENOMEM;
1729}
1730
1731static int io_read_prep(struct io_kiocb *req, struct iovec **iovec,
1732 struct iov_iter *iter, bool force_nonblock)
1733{
1734 ssize_t ret;
1735
1736 ret = io_prep_rw(req, force_nonblock);
1737 if (ret)
1738 return ret;
1739
1740 if (unlikely(!(req->file->f_mode & FMODE_READ)))
1741 return -EBADF;
1742
1743 return io_import_iovec(READ, req, iovec, iter);
1744}
1745
Pavel Begunkov267bc902019-11-07 01:41:08 +03001746static int io_read(struct io_kiocb *req, struct io_kiocb **nxt,
Jens Axboe8358e3a2019-04-23 08:17:58 -06001747 bool force_nonblock)
Jens Axboe2b188cc2019-01-07 10:46:33 -07001748{
1749 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1750 struct kiocb *kiocb = &req->rw;
1751 struct iov_iter iter;
1752 struct file *file;
Jens Axboe31b51512019-01-18 22:56:34 -07001753 size_t iov_count;
Jens Axboef67676d2019-12-02 11:03:47 -07001754 ssize_t io_size, ret;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001755
Jens Axboef67676d2019-12-02 11:03:47 -07001756 if (!req->io) {
1757 ret = io_read_prep(req, &iovec, &iter, force_nonblock);
1758 if (ret < 0)
1759 return ret;
1760 } else {
1761 ret = io_import_iovec(READ, req, &iovec, &iter);
1762 if (ret < 0)
1763 return ret;
1764 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07001765
Jens Axboef67676d2019-12-02 11:03:47 -07001766 file = req->file;
1767 io_size = ret;
Jens Axboe9e645e112019-05-10 16:07:28 -06001768 if (req->flags & REQ_F_LINK)
Jens Axboef67676d2019-12-02 11:03:47 -07001769 req->result = io_size;
1770
1771 /*
1772 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
1773 * we know to async punt it even if it was opened O_NONBLOCK
1774 */
1775 if (force_nonblock && !io_file_supports_async(file)) {
1776 req->flags |= REQ_F_MUST_PUNT;
1777 goto copy_iov;
1778 }
Jens Axboe9e645e112019-05-10 16:07:28 -06001779
Jens Axboe31b51512019-01-18 22:56:34 -07001780 iov_count = iov_iter_count(&iter);
1781 ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count);
Jens Axboe2b188cc2019-01-07 10:46:33 -07001782 if (!ret) {
1783 ssize_t ret2;
1784
Jens Axboe32960612019-09-23 11:05:34 -06001785 if (file->f_op->read_iter)
1786 ret2 = call_read_iter(file, kiocb, &iter);
1787 else
1788 ret2 = loop_rw_iter(READ, file, kiocb, &iter);
1789
Jens Axboe9d93a3f2019-05-15 13:53:07 -06001790 /*
1791 * In case of a short read, punt to async. This can happen
1792 * if we have data partially cached. Alternatively we can
1793 * return the short read, in which case the application will
1794 * need to issue another SQE and wait for it. That SQE will
1795 * need async punt anyway, so it's more efficient to do it
1796 * here.
1797 */
Jens Axboe491381ce2019-10-17 09:20:46 -06001798 if (force_nonblock && !(req->flags & REQ_F_NOWAIT) &&
1799 (req->flags & REQ_F_ISREG) &&
Jens Axboef67676d2019-12-02 11:03:47 -07001800 ret2 > 0 && ret2 < io_size)
Jens Axboe9d93a3f2019-05-15 13:53:07 -06001801 ret2 = -EAGAIN;
1802 /* Catch -EAGAIN return for forced non-blocking submission */
Jens Axboef67676d2019-12-02 11:03:47 -07001803 if (!force_nonblock || ret2 != -EAGAIN) {
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03001804 kiocb_done(kiocb, ret2, nxt, req->in_async);
Jens Axboef67676d2019-12-02 11:03:47 -07001805 } else {
1806copy_iov:
1807 ret = io_setup_async_io(req, io_size, iovec,
1808 inline_vecs, &iter);
1809 if (ret)
1810 goto out_free;
1811 return -EAGAIN;
1812 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07001813 }
Jens Axboef67676d2019-12-02 11:03:47 -07001814out_free:
Jens Axboe2b188cc2019-01-07 10:46:33 -07001815 kfree(iovec);
Jens Axboe2b188cc2019-01-07 10:46:33 -07001816 return ret;
1817}
1818
Jens Axboef67676d2019-12-02 11:03:47 -07001819static int io_write_prep(struct io_kiocb *req, struct iovec **iovec,
1820 struct iov_iter *iter, bool force_nonblock)
1821{
1822 ssize_t ret;
1823
1824 ret = io_prep_rw(req, force_nonblock);
1825 if (ret)
1826 return ret;
1827
1828 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
1829 return -EBADF;
1830
1831 return io_import_iovec(WRITE, req, iovec, iter);
1832}
1833
Pavel Begunkov267bc902019-11-07 01:41:08 +03001834static int io_write(struct io_kiocb *req, struct io_kiocb **nxt,
Jens Axboe8358e3a2019-04-23 08:17:58 -06001835 bool force_nonblock)
Jens Axboe2b188cc2019-01-07 10:46:33 -07001836{
1837 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1838 struct kiocb *kiocb = &req->rw;
1839 struct iov_iter iter;
1840 struct file *file;
Jens Axboe31b51512019-01-18 22:56:34 -07001841 size_t iov_count;
Jens Axboef67676d2019-12-02 11:03:47 -07001842 ssize_t ret, io_size;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001843
Jens Axboef67676d2019-12-02 11:03:47 -07001844 if (!req->io) {
1845 ret = io_write_prep(req, &iovec, &iter, force_nonblock);
1846 if (ret < 0)
1847 return ret;
1848 } else {
1849 ret = io_import_iovec(WRITE, req, &iovec, &iter);
1850 if (ret < 0)
1851 return ret;
1852 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07001853
Jens Axboe2b188cc2019-01-07 10:46:33 -07001854 file = kiocb->ki_filp;
Jens Axboef67676d2019-12-02 11:03:47 -07001855 io_size = ret;
Jens Axboe9e645e112019-05-10 16:07:28 -06001856 if (req->flags & REQ_F_LINK)
Jens Axboef67676d2019-12-02 11:03:47 -07001857 req->result = io_size;
1858
1859 /*
1860 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
1861 * we know to async punt it even if it was opened O_NONBLOCK
1862 */
1863 if (force_nonblock && !io_file_supports_async(req->file)) {
1864 req->flags |= REQ_F_MUST_PUNT;
1865 goto copy_iov;
1866 }
1867
1868 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT))
1869 goto copy_iov;
Jens Axboe9e645e112019-05-10 16:07:28 -06001870
Jens Axboe31b51512019-01-18 22:56:34 -07001871 iov_count = iov_iter_count(&iter);
Jens Axboe31b51512019-01-18 22:56:34 -07001872 ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count);
Jens Axboe2b188cc2019-01-07 10:46:33 -07001873 if (!ret) {
Roman Penyaev9bf79332019-03-25 20:09:24 +01001874 ssize_t ret2;
1875
Jens Axboe2b188cc2019-01-07 10:46:33 -07001876 /*
1877 * Open-code file_start_write here to grab freeze protection,
1878 * which will be released by another thread in
1879 * io_complete_rw(). Fool lockdep by telling it the lock got
1880 * released so that it doesn't complain about the held lock when
1881 * we return to userspace.
1882 */
Jens Axboe491381ce2019-10-17 09:20:46 -06001883 if (req->flags & REQ_F_ISREG) {
Jens Axboe2b188cc2019-01-07 10:46:33 -07001884 __sb_start_write(file_inode(file)->i_sb,
1885 SB_FREEZE_WRITE, true);
1886 __sb_writers_release(file_inode(file)->i_sb,
1887 SB_FREEZE_WRITE);
1888 }
1889 kiocb->ki_flags |= IOCB_WRITE;
Roman Penyaev9bf79332019-03-25 20:09:24 +01001890
Jens Axboe32960612019-09-23 11:05:34 -06001891 if (file->f_op->write_iter)
1892 ret2 = call_write_iter(file, kiocb, &iter);
1893 else
1894 ret2 = loop_rw_iter(WRITE, file, kiocb, &iter);
Jens Axboef67676d2019-12-02 11:03:47 -07001895 if (!force_nonblock || ret2 != -EAGAIN) {
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03001896 kiocb_done(kiocb, ret2, nxt, req->in_async);
Jens Axboef67676d2019-12-02 11:03:47 -07001897 } else {
1898copy_iov:
1899 ret = io_setup_async_io(req, io_size, iovec,
1900 inline_vecs, &iter);
1901 if (ret)
1902 goto out_free;
1903 return -EAGAIN;
1904 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07001905 }
Jens Axboe31b51512019-01-18 22:56:34 -07001906out_free:
Jens Axboe2b188cc2019-01-07 10:46:33 -07001907 kfree(iovec);
Jens Axboe2b188cc2019-01-07 10:46:33 -07001908 return ret;
1909}
1910
1911/*
1912 * IORING_OP_NOP just posts a completion event, nothing else.
1913 */
Jens Axboe78e19bb2019-11-06 15:21:34 -07001914static int io_nop(struct io_kiocb *req)
Jens Axboe2b188cc2019-01-07 10:46:33 -07001915{
1916 struct io_ring_ctx *ctx = req->ctx;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001917
Jens Axboedef596e2019-01-09 08:59:42 -07001918 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1919 return -EINVAL;
1920
Jens Axboe78e19bb2019-11-06 15:21:34 -07001921 io_cqring_add_event(req, 0);
Jens Axboee65ef562019-03-12 10:16:44 -06001922 io_put_req(req);
Jens Axboe2b188cc2019-01-07 10:46:33 -07001923 return 0;
1924}
1925
Christoph Hellwigc992fe22019-01-11 09:43:02 -07001926static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1927{
Jens Axboe6b063142019-01-10 22:13:58 -07001928 struct io_ring_ctx *ctx = req->ctx;
Christoph Hellwigc992fe22019-01-11 09:43:02 -07001929
Jens Axboe09bb8392019-03-13 12:39:28 -06001930 if (!req->file)
1931 return -EBADF;
Christoph Hellwigc992fe22019-01-11 09:43:02 -07001932
Jens Axboe6b063142019-01-10 22:13:58 -07001933 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
Jens Axboedef596e2019-01-09 08:59:42 -07001934 return -EINVAL;
Jens Axboeedafcce2019-01-09 09:16:05 -07001935 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
Christoph Hellwigc992fe22019-01-11 09:43:02 -07001936 return -EINVAL;
1937
Christoph Hellwigc992fe22019-01-11 09:43:02 -07001938 return 0;
1939}
1940
1941static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe,
Jens Axboeba816ad2019-09-28 11:36:45 -06001942 struct io_kiocb **nxt, bool force_nonblock)
Christoph Hellwigc992fe22019-01-11 09:43:02 -07001943{
1944 loff_t sqe_off = READ_ONCE(sqe->off);
1945 loff_t sqe_len = READ_ONCE(sqe->len);
1946 loff_t end = sqe_off + sqe_len;
1947 unsigned fsync_flags;
1948 int ret;
1949
1950 fsync_flags = READ_ONCE(sqe->fsync_flags);
1951 if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC))
1952 return -EINVAL;
1953
1954 ret = io_prep_fsync(req, sqe);
1955 if (ret)
1956 return ret;
1957
1958 /* fsync always requires a blocking context */
1959 if (force_nonblock)
1960 return -EAGAIN;
1961
1962 ret = vfs_fsync_range(req->rw.ki_filp, sqe_off,
1963 end > 0 ? end : LLONG_MAX,
1964 fsync_flags & IORING_FSYNC_DATASYNC);
1965
Jens Axboe4e88d6e2019-12-07 20:59:47 -07001966 if (ret < 0)
1967 req_set_fail_links(req);
Jens Axboe78e19bb2019-11-06 15:21:34 -07001968 io_cqring_add_event(req, ret);
Jackie Liuec9c02a2019-11-08 23:50:36 +08001969 io_put_req_find_next(req, nxt);
Christoph Hellwigc992fe22019-01-11 09:43:02 -07001970 return 0;
1971}
1972
Jens Axboe5d17b4a2019-04-09 14:56:44 -06001973static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1974{
1975 struct io_ring_ctx *ctx = req->ctx;
1976 int ret = 0;
1977
1978 if (!req->file)
1979 return -EBADF;
Jens Axboe5d17b4a2019-04-09 14:56:44 -06001980
1981 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1982 return -EINVAL;
1983 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1984 return -EINVAL;
1985
Jens Axboe5d17b4a2019-04-09 14:56:44 -06001986 return ret;
1987}
1988
1989static int io_sync_file_range(struct io_kiocb *req,
1990 const struct io_uring_sqe *sqe,
Jens Axboeba816ad2019-09-28 11:36:45 -06001991 struct io_kiocb **nxt,
Jens Axboe5d17b4a2019-04-09 14:56:44 -06001992 bool force_nonblock)
1993{
1994 loff_t sqe_off;
1995 loff_t sqe_len;
1996 unsigned flags;
1997 int ret;
1998
1999 ret = io_prep_sfr(req, sqe);
2000 if (ret)
2001 return ret;
2002
2003 /* sync_file_range always requires a blocking context */
2004 if (force_nonblock)
2005 return -EAGAIN;
2006
2007 sqe_off = READ_ONCE(sqe->off);
2008 sqe_len = READ_ONCE(sqe->len);
2009 flags = READ_ONCE(sqe->sync_range_flags);
2010
2011 ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags);
2012
Jens Axboe4e88d6e2019-12-07 20:59:47 -07002013 if (ret < 0)
2014 req_set_fail_links(req);
Jens Axboe78e19bb2019-11-06 15:21:34 -07002015 io_cqring_add_event(req, ret);
Jackie Liuec9c02a2019-11-08 23:50:36 +08002016 io_put_req_find_next(req, nxt);
Jens Axboe5d17b4a2019-04-09 14:56:44 -06002017 return 0;
2018}
2019
Jens Axboe03b12302019-12-02 18:50:25 -07002020static int io_sendmsg_prep(struct io_kiocb *req, struct io_async_ctx *io)
Jens Axboeaa1fa282019-04-19 13:38:09 -06002021{
Jens Axboe03b12302019-12-02 18:50:25 -07002022#if defined(CONFIG_NET)
2023 const struct io_uring_sqe *sqe = req->sqe;
2024 struct user_msghdr __user *msg;
2025 unsigned flags;
2026
2027 flags = READ_ONCE(sqe->msg_flags);
2028 msg = (struct user_msghdr __user *)(unsigned long) READ_ONCE(sqe->addr);
2029 return sendmsg_copy_msghdr(&io->msg.msg, msg, flags, &io->msg.iov);
2030#else
2031 return 0;
2032#endif
2033}
2034
2035static int io_sendmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2036 struct io_kiocb **nxt, bool force_nonblock)
2037{
2038#if defined(CONFIG_NET)
2039 struct socket *sock;
2040 int ret;
2041
2042 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2043 return -EINVAL;
2044
2045 sock = sock_from_file(req->file, &ret);
2046 if (sock) {
2047 struct io_async_ctx io, *copy;
2048 struct sockaddr_storage addr;
2049 struct msghdr *kmsg;
2050 unsigned flags;
2051
2052 flags = READ_ONCE(sqe->msg_flags);
2053 if (flags & MSG_DONTWAIT)
2054 req->flags |= REQ_F_NOWAIT;
2055 else if (force_nonblock)
2056 flags |= MSG_DONTWAIT;
2057
2058 if (req->io) {
2059 kmsg = &req->io->msg.msg;
2060 kmsg->msg_name = &addr;
2061 } else {
2062 kmsg = &io.msg.msg;
2063 kmsg->msg_name = &addr;
2064 io.msg.iov = io.msg.fast_iov;
2065 ret = io_sendmsg_prep(req, &io);
2066 if (ret)
2067 goto out;
2068 }
2069
2070 ret = __sys_sendmsg_sock(sock, kmsg, flags);
2071 if (force_nonblock && ret == -EAGAIN) {
2072 copy = kmalloc(sizeof(*copy), GFP_KERNEL);
2073 if (!copy) {
2074 ret = -ENOMEM;
2075 goto out;
2076 }
2077 memcpy(&copy->msg, &io.msg, sizeof(copy->msg));
2078 req->io = copy;
2079 memcpy(&req->io->sqe, req->sqe, sizeof(*req->sqe));
2080 req->sqe = &req->io->sqe;
2081 return ret;
2082 }
2083 if (ret == -ERESTARTSYS)
2084 ret = -EINTR;
2085 }
2086
2087out:
2088 io_cqring_add_event(req, ret);
Jens Axboe4e88d6e2019-12-07 20:59:47 -07002089 if (ret < 0)
2090 req_set_fail_links(req);
Jens Axboe03b12302019-12-02 18:50:25 -07002091 io_put_req_find_next(req, nxt);
2092 return 0;
2093#else
2094 return -EOPNOTSUPP;
2095#endif
2096}
2097
2098static int io_recvmsg_prep(struct io_kiocb *req, struct io_async_ctx *io)
2099{
2100#if defined(CONFIG_NET)
2101 const struct io_uring_sqe *sqe = req->sqe;
2102 struct user_msghdr __user *msg;
2103 unsigned flags;
2104
2105 flags = READ_ONCE(sqe->msg_flags);
2106 msg = (struct user_msghdr __user *)(unsigned long) READ_ONCE(sqe->addr);
2107 return recvmsg_copy_msghdr(&io->msg.msg, msg, flags, &io->msg.uaddr,
2108 &io->msg.iov);
2109#else
2110 return 0;
2111#endif
2112}
2113
2114static int io_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2115 struct io_kiocb **nxt, bool force_nonblock)
2116{
2117#if defined(CONFIG_NET)
Jens Axboe0fa03c62019-04-19 13:34:07 -06002118 struct socket *sock;
2119 int ret;
2120
2121 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2122 return -EINVAL;
2123
2124 sock = sock_from_file(req->file, &ret);
2125 if (sock) {
2126 struct user_msghdr __user *msg;
Jens Axboe03b12302019-12-02 18:50:25 -07002127 struct io_async_ctx io, *copy;
2128 struct sockaddr_storage addr;
2129 struct msghdr *kmsg;
Jens Axboe0fa03c62019-04-19 13:34:07 -06002130 unsigned flags;
2131
2132 flags = READ_ONCE(sqe->msg_flags);
2133 if (flags & MSG_DONTWAIT)
2134 req->flags |= REQ_F_NOWAIT;
2135 else if (force_nonblock)
2136 flags |= MSG_DONTWAIT;
2137
2138 msg = (struct user_msghdr __user *) (unsigned long)
2139 READ_ONCE(sqe->addr);
Jens Axboe03b12302019-12-02 18:50:25 -07002140 if (req->io) {
2141 kmsg = &req->io->msg.msg;
2142 kmsg->msg_name = &addr;
2143 } else {
2144 kmsg = &io.msg.msg;
2145 kmsg->msg_name = &addr;
2146 io.msg.iov = io.msg.fast_iov;
2147 ret = io_recvmsg_prep(req, &io);
2148 if (ret)
2149 goto out;
2150 }
Jens Axboe0fa03c62019-04-19 13:34:07 -06002151
Jens Axboe03b12302019-12-02 18:50:25 -07002152 ret = __sys_recvmsg_sock(sock, kmsg, msg, io.msg.uaddr, flags);
2153 if (force_nonblock && ret == -EAGAIN) {
2154 copy = kmalloc(sizeof(*copy), GFP_KERNEL);
2155 if (!copy) {
2156 ret = -ENOMEM;
2157 goto out;
2158 }
2159 memcpy(copy, &io, sizeof(*copy));
2160 req->io = copy;
2161 memcpy(&req->io->sqe, req->sqe, sizeof(*req->sqe));
2162 req->sqe = &req->io->sqe;
Jens Axboe0fa03c62019-04-19 13:34:07 -06002163 return ret;
Jens Axboe03b12302019-12-02 18:50:25 -07002164 }
Jens Axboe441cdbd2019-12-02 18:49:10 -07002165 if (ret == -ERESTARTSYS)
2166 ret = -EINTR;
Jens Axboe0fa03c62019-04-19 13:34:07 -06002167 }
2168
Jens Axboe03b12302019-12-02 18:50:25 -07002169out:
Jens Axboe78e19bb2019-11-06 15:21:34 -07002170 io_cqring_add_event(req, ret);
Jens Axboe4e88d6e2019-12-07 20:59:47 -07002171 if (ret < 0)
2172 req_set_fail_links(req);
Jackie Liuec9c02a2019-11-08 23:50:36 +08002173 io_put_req_find_next(req, nxt);
Jens Axboe0fa03c62019-04-19 13:34:07 -06002174 return 0;
2175#else
2176 return -EOPNOTSUPP;
2177#endif
2178}
2179
Jens Axboe17f2fe32019-10-17 14:42:58 -06002180static int io_accept(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2181 struct io_kiocb **nxt, bool force_nonblock)
2182{
2183#if defined(CONFIG_NET)
2184 struct sockaddr __user *addr;
2185 int __user *addr_len;
2186 unsigned file_flags;
2187 int flags, ret;
2188
2189 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
2190 return -EINVAL;
Hrvoje Zeba8042d6c2019-11-25 14:40:22 -05002191 if (sqe->ioprio || sqe->len || sqe->buf_index)
Jens Axboe17f2fe32019-10-17 14:42:58 -06002192 return -EINVAL;
2193
2194 addr = (struct sockaddr __user *) (unsigned long) READ_ONCE(sqe->addr);
2195 addr_len = (int __user *) (unsigned long) READ_ONCE(sqe->addr2);
2196 flags = READ_ONCE(sqe->accept_flags);
2197 file_flags = force_nonblock ? O_NONBLOCK : 0;
2198
2199 ret = __sys_accept4_file(req->file, file_flags, addr, addr_len, flags);
2200 if (ret == -EAGAIN && force_nonblock) {
2201 req->work.flags |= IO_WQ_WORK_NEEDS_FILES;
2202 return -EAGAIN;
2203 }
Jens Axboe8e3cca12019-11-09 19:52:33 -07002204 if (ret == -ERESTARTSYS)
2205 ret = -EINTR;
Jens Axboe4e88d6e2019-12-07 20:59:47 -07002206 if (ret < 0)
2207 req_set_fail_links(req);
Jens Axboe78e19bb2019-11-06 15:21:34 -07002208 io_cqring_add_event(req, ret);
Jackie Liuec9c02a2019-11-08 23:50:36 +08002209 io_put_req_find_next(req, nxt);
Jens Axboe17f2fe32019-10-17 14:42:58 -06002210 return 0;
2211#else
2212 return -EOPNOTSUPP;
2213#endif
2214}
2215
Jens Axboef499a022019-12-02 16:28:46 -07002216static int io_connect_prep(struct io_kiocb *req, struct io_async_ctx *io)
2217{
2218#if defined(CONFIG_NET)
2219 const struct io_uring_sqe *sqe = req->sqe;
2220 struct sockaddr __user *addr;
2221 int addr_len;
2222
2223 addr = (struct sockaddr __user *) (unsigned long) READ_ONCE(sqe->addr);
2224 addr_len = READ_ONCE(sqe->addr2);
2225 return move_addr_to_kernel(addr, addr_len, &io->connect.address);
2226#else
2227 return 0;
2228#endif
2229}
2230
Jens Axboef8e85cf2019-11-23 14:24:24 -07002231static int io_connect(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2232 struct io_kiocb **nxt, bool force_nonblock)
2233{
2234#if defined(CONFIG_NET)
Jens Axboef499a022019-12-02 16:28:46 -07002235 struct io_async_ctx __io, *io;
Jens Axboef8e85cf2019-11-23 14:24:24 -07002236 unsigned file_flags;
2237 int addr_len, ret;
2238
2239 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
2240 return -EINVAL;
2241 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
2242 return -EINVAL;
2243
Jens Axboef8e85cf2019-11-23 14:24:24 -07002244 addr_len = READ_ONCE(sqe->addr2);
2245 file_flags = force_nonblock ? O_NONBLOCK : 0;
2246
Jens Axboef499a022019-12-02 16:28:46 -07002247 if (req->io) {
2248 io = req->io;
2249 } else {
2250 ret = io_connect_prep(req, &__io);
2251 if (ret)
2252 goto out;
2253 io = &__io;
2254 }
2255
2256 ret = __sys_connect_file(req->file, &io->connect.address, addr_len,
2257 file_flags);
Jens Axboe87f80d62019-12-03 11:23:54 -07002258 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
Jens Axboef499a022019-12-02 16:28:46 -07002259 io = kmalloc(sizeof(*io), GFP_KERNEL);
2260 if (!io) {
2261 ret = -ENOMEM;
2262 goto out;
2263 }
2264 memcpy(&io->connect, &__io.connect, sizeof(io->connect));
2265 req->io = io;
2266 memcpy(&io->sqe, req->sqe, sizeof(*req->sqe));
2267 req->sqe = &io->sqe;
Jens Axboef8e85cf2019-11-23 14:24:24 -07002268 return -EAGAIN;
Jens Axboef499a022019-12-02 16:28:46 -07002269 }
Jens Axboef8e85cf2019-11-23 14:24:24 -07002270 if (ret == -ERESTARTSYS)
2271 ret = -EINTR;
Jens Axboef499a022019-12-02 16:28:46 -07002272out:
Jens Axboe4e88d6e2019-12-07 20:59:47 -07002273 if (ret < 0)
2274 req_set_fail_links(req);
Jens Axboef8e85cf2019-11-23 14:24:24 -07002275 io_cqring_add_event(req, ret);
2276 io_put_req_find_next(req, nxt);
2277 return 0;
2278#else
2279 return -EOPNOTSUPP;
2280#endif
2281}
2282
Jens Axboe221c5eb2019-01-17 09:41:58 -07002283static void io_poll_remove_one(struct io_kiocb *req)
2284{
2285 struct io_poll_iocb *poll = &req->poll;
2286
2287 spin_lock(&poll->head->lock);
2288 WRITE_ONCE(poll->canceled, true);
Jens Axboee9444752019-11-26 15:02:04 -07002289 if (!list_empty(&poll->wait->entry)) {
2290 list_del_init(&poll->wait->entry);
Jackie Liua197f662019-11-08 08:09:12 -07002291 io_queue_async_work(req);
Jens Axboe221c5eb2019-01-17 09:41:58 -07002292 }
2293 spin_unlock(&poll->head->lock);
Jens Axboe78076bb2019-12-04 19:56:40 -07002294 hash_del(&req->hash_node);
Jens Axboe221c5eb2019-01-17 09:41:58 -07002295}
2296
2297static void io_poll_remove_all(struct io_ring_ctx *ctx)
2298{
Jens Axboe78076bb2019-12-04 19:56:40 -07002299 struct hlist_node *tmp;
Jens Axboe221c5eb2019-01-17 09:41:58 -07002300 struct io_kiocb *req;
Jens Axboe78076bb2019-12-04 19:56:40 -07002301 int i;
Jens Axboe221c5eb2019-01-17 09:41:58 -07002302
2303 spin_lock_irq(&ctx->completion_lock);
Jens Axboe78076bb2019-12-04 19:56:40 -07002304 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
2305 struct hlist_head *list;
2306
2307 list = &ctx->cancel_hash[i];
2308 hlist_for_each_entry_safe(req, tmp, list, hash_node)
2309 io_poll_remove_one(req);
Jens Axboe221c5eb2019-01-17 09:41:58 -07002310 }
2311 spin_unlock_irq(&ctx->completion_lock);
2312}
2313
Jens Axboe47f46762019-11-09 17:43:02 -07002314static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
2315{
Jens Axboe78076bb2019-12-04 19:56:40 -07002316 struct hlist_head *list;
Jens Axboe47f46762019-11-09 17:43:02 -07002317 struct io_kiocb *req;
2318
Jens Axboe78076bb2019-12-04 19:56:40 -07002319 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
2320 hlist_for_each_entry(req, list, hash_node) {
2321 if (sqe_addr == req->user_data) {
Jens Axboeeac406c2019-11-14 12:09:58 -07002322 io_poll_remove_one(req);
2323 return 0;
2324 }
Jens Axboe47f46762019-11-09 17:43:02 -07002325 }
2326
2327 return -ENOENT;
2328}
2329
Jens Axboe221c5eb2019-01-17 09:41:58 -07002330/*
2331 * Find a running poll command that matches one specified in sqe->addr,
2332 * and remove it if found.
2333 */
2334static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2335{
2336 struct io_ring_ctx *ctx = req->ctx;
Jens Axboe47f46762019-11-09 17:43:02 -07002337 int ret;
Jens Axboe221c5eb2019-01-17 09:41:58 -07002338
2339 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2340 return -EINVAL;
2341 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
2342 sqe->poll_events)
2343 return -EINVAL;
2344
2345 spin_lock_irq(&ctx->completion_lock);
Jens Axboe47f46762019-11-09 17:43:02 -07002346 ret = io_poll_cancel(ctx, READ_ONCE(sqe->addr));
Jens Axboe221c5eb2019-01-17 09:41:58 -07002347 spin_unlock_irq(&ctx->completion_lock);
2348
Jens Axboe78e19bb2019-11-06 15:21:34 -07002349 io_cqring_add_event(req, ret);
Jens Axboe4e88d6e2019-12-07 20:59:47 -07002350 if (ret < 0)
2351 req_set_fail_links(req);
Jens Axboee65ef562019-03-12 10:16:44 -06002352 io_put_req(req);
Jens Axboe221c5eb2019-01-17 09:41:58 -07002353 return 0;
2354}
2355
Jens Axboeb0dd8a42019-11-18 12:14:54 -07002356static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
Jens Axboe221c5eb2019-01-17 09:41:58 -07002357{
Jackie Liua197f662019-11-08 08:09:12 -07002358 struct io_ring_ctx *ctx = req->ctx;
2359
Jens Axboe8c838782019-03-12 15:48:16 -06002360 req->poll.done = true;
Jens Axboee9444752019-11-26 15:02:04 -07002361 kfree(req->poll.wait);
Jens Axboeb0dd8a42019-11-18 12:14:54 -07002362 if (error)
2363 io_cqring_fill_event(req, error);
2364 else
2365 io_cqring_fill_event(req, mangle_poll(mask));
Jens Axboe8c838782019-03-12 15:48:16 -06002366 io_commit_cqring(ctx);
Jens Axboe221c5eb2019-01-17 09:41:58 -07002367}
2368
Jens Axboe561fb042019-10-24 07:25:42 -06002369static void io_poll_complete_work(struct io_wq_work **workptr)
Jens Axboe221c5eb2019-01-17 09:41:58 -07002370{
Jens Axboe561fb042019-10-24 07:25:42 -06002371 struct io_wq_work *work = *workptr;
Jens Axboe221c5eb2019-01-17 09:41:58 -07002372 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2373 struct io_poll_iocb *poll = &req->poll;
2374 struct poll_table_struct pt = { ._key = poll->events };
2375 struct io_ring_ctx *ctx = req->ctx;
Jens Axboe89723d02019-11-05 15:32:58 -07002376 struct io_kiocb *nxt = NULL;
Jens Axboe221c5eb2019-01-17 09:41:58 -07002377 __poll_t mask = 0;
Jens Axboeb0dd8a42019-11-18 12:14:54 -07002378 int ret = 0;
Jens Axboe221c5eb2019-01-17 09:41:58 -07002379
Jens Axboeb0dd8a42019-11-18 12:14:54 -07002380 if (work->flags & IO_WQ_WORK_CANCEL) {
Jens Axboe561fb042019-10-24 07:25:42 -06002381 WRITE_ONCE(poll->canceled, true);
Jens Axboeb0dd8a42019-11-18 12:14:54 -07002382 ret = -ECANCELED;
2383 } else if (READ_ONCE(poll->canceled)) {
2384 ret = -ECANCELED;
2385 }
Jens Axboe561fb042019-10-24 07:25:42 -06002386
Jens Axboeb0dd8a42019-11-18 12:14:54 -07002387 if (ret != -ECANCELED)
Jens Axboe221c5eb2019-01-17 09:41:58 -07002388 mask = vfs_poll(poll->file, &pt) & poll->events;
2389
2390 /*
2391 * Note that ->ki_cancel callers also delete iocb from active_reqs after
2392 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
2393 * synchronize with them. In the cancellation case the list_del_init
2394 * itself is not actually needed, but harmless so we keep it in to
2395 * avoid further branches in the fast path.
2396 */
2397 spin_lock_irq(&ctx->completion_lock);
Jens Axboeb0dd8a42019-11-18 12:14:54 -07002398 if (!mask && ret != -ECANCELED) {
Jens Axboee9444752019-11-26 15:02:04 -07002399 add_wait_queue(poll->head, poll->wait);
Jens Axboe221c5eb2019-01-17 09:41:58 -07002400 spin_unlock_irq(&ctx->completion_lock);
2401 return;
2402 }
Jens Axboe78076bb2019-12-04 19:56:40 -07002403 hash_del(&req->hash_node);
Jens Axboeb0dd8a42019-11-18 12:14:54 -07002404 io_poll_complete(req, mask, ret);
Jens Axboe221c5eb2019-01-17 09:41:58 -07002405 spin_unlock_irq(&ctx->completion_lock);
2406
Jens Axboe8c838782019-03-12 15:48:16 -06002407 io_cqring_ev_posted(ctx);
Jens Axboe89723d02019-11-05 15:32:58 -07002408
Jens Axboe4e88d6e2019-12-07 20:59:47 -07002409 if (ret < 0)
2410 req_set_fail_links(req);
Jackie Liuec9c02a2019-11-08 23:50:36 +08002411 io_put_req_find_next(req, &nxt);
Jens Axboe89723d02019-11-05 15:32:58 -07002412 if (nxt)
2413 *workptr = &nxt->work;
Jens Axboe221c5eb2019-01-17 09:41:58 -07002414}
2415
2416static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
2417 void *key)
2418{
Jens Axboee9444752019-11-26 15:02:04 -07002419 struct io_poll_iocb *poll = wait->private;
Jens Axboe221c5eb2019-01-17 09:41:58 -07002420 struct io_kiocb *req = container_of(poll, struct io_kiocb, poll);
2421 struct io_ring_ctx *ctx = req->ctx;
2422 __poll_t mask = key_to_poll(key);
Jens Axboe8c838782019-03-12 15:48:16 -06002423 unsigned long flags;
Jens Axboe221c5eb2019-01-17 09:41:58 -07002424
2425 /* for instances that support it check for an event match first: */
Jens Axboe8c838782019-03-12 15:48:16 -06002426 if (mask && !(mask & poll->events))
2427 return 0;
Jens Axboe221c5eb2019-01-17 09:41:58 -07002428
Jens Axboee9444752019-11-26 15:02:04 -07002429 list_del_init(&poll->wait->entry);
Jens Axboe8c838782019-03-12 15:48:16 -06002430
Jens Axboe7c9e7f02019-11-12 08:15:53 -07002431 /*
2432 * Run completion inline if we can. We're using trylock here because
2433 * we are violating the completion_lock -> poll wq lock ordering.
2434 * If we have a link timeout we're going to need the completion_lock
2435 * for finalizing the request, mark us as having grabbed that already.
2436 */
Jens Axboe8c838782019-03-12 15:48:16 -06002437 if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) {
Jens Axboe78076bb2019-12-04 19:56:40 -07002438 hash_del(&req->hash_node);
Jens Axboeb0dd8a42019-11-18 12:14:54 -07002439 io_poll_complete(req, mask, 0);
Jens Axboe7c9e7f02019-11-12 08:15:53 -07002440 req->flags |= REQ_F_COMP_LOCKED;
2441 io_put_req(req);
Jens Axboe8c838782019-03-12 15:48:16 -06002442 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2443
2444 io_cqring_ev_posted(ctx);
Jens Axboe8c838782019-03-12 15:48:16 -06002445 } else {
Jackie Liua197f662019-11-08 08:09:12 -07002446 io_queue_async_work(req);
Jens Axboe8c838782019-03-12 15:48:16 -06002447 }
2448
Jens Axboe221c5eb2019-01-17 09:41:58 -07002449 return 1;
2450}
2451
2452struct io_poll_table {
2453 struct poll_table_struct pt;
2454 struct io_kiocb *req;
2455 int error;
2456};
2457
2458static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
2459 struct poll_table_struct *p)
2460{
2461 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
2462
2463 if (unlikely(pt->req->poll.head)) {
2464 pt->error = -EINVAL;
2465 return;
2466 }
2467
2468 pt->error = 0;
2469 pt->req->poll.head = head;
Jens Axboee9444752019-11-26 15:02:04 -07002470 add_wait_queue(head, pt->req->poll.wait);
Jens Axboe221c5eb2019-01-17 09:41:58 -07002471}
2472
Jens Axboeeac406c2019-11-14 12:09:58 -07002473static void io_poll_req_insert(struct io_kiocb *req)
2474{
2475 struct io_ring_ctx *ctx = req->ctx;
Jens Axboe78076bb2019-12-04 19:56:40 -07002476 struct hlist_head *list;
Jens Axboeeac406c2019-11-14 12:09:58 -07002477
Jens Axboe78076bb2019-12-04 19:56:40 -07002478 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
2479 hlist_add_head(&req->hash_node, list);
Jens Axboeeac406c2019-11-14 12:09:58 -07002480}
2481
Jens Axboe89723d02019-11-05 15:32:58 -07002482static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2483 struct io_kiocb **nxt)
Jens Axboe221c5eb2019-01-17 09:41:58 -07002484{
2485 struct io_poll_iocb *poll = &req->poll;
2486 struct io_ring_ctx *ctx = req->ctx;
2487 struct io_poll_table ipt;
Jens Axboe8c838782019-03-12 15:48:16 -06002488 bool cancel = false;
Jens Axboe221c5eb2019-01-17 09:41:58 -07002489 __poll_t mask;
2490 u16 events;
Jens Axboe221c5eb2019-01-17 09:41:58 -07002491
2492 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2493 return -EINVAL;
2494 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
2495 return -EINVAL;
Jens Axboe09bb8392019-03-13 12:39:28 -06002496 if (!poll->file)
2497 return -EBADF;
Jens Axboe221c5eb2019-01-17 09:41:58 -07002498
Jens Axboee9444752019-11-26 15:02:04 -07002499 poll->wait = kmalloc(sizeof(*poll->wait), GFP_KERNEL);
2500 if (!poll->wait)
2501 return -ENOMEM;
2502
Jens Axboe1a6b74f2019-12-02 10:33:15 -07002503 req->io = NULL;
Jens Axboe561fb042019-10-24 07:25:42 -06002504 INIT_IO_WORK(&req->work, io_poll_complete_work);
Jens Axboe221c5eb2019-01-17 09:41:58 -07002505 events = READ_ONCE(sqe->poll_events);
2506 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
Jens Axboe78076bb2019-12-04 19:56:40 -07002507 INIT_HLIST_NODE(&req->hash_node);
Jens Axboe221c5eb2019-01-17 09:41:58 -07002508
Jens Axboe221c5eb2019-01-17 09:41:58 -07002509 poll->head = NULL;
Jens Axboe8c838782019-03-12 15:48:16 -06002510 poll->done = false;
Jens Axboe221c5eb2019-01-17 09:41:58 -07002511 poll->canceled = false;
2512
2513 ipt.pt._qproc = io_poll_queue_proc;
2514 ipt.pt._key = poll->events;
2515 ipt.req = req;
2516 ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
2517
2518 /* initialized the list so that we can do list_empty checks */
Jens Axboee9444752019-11-26 15:02:04 -07002519 INIT_LIST_HEAD(&poll->wait->entry);
2520 init_waitqueue_func_entry(poll->wait, io_poll_wake);
2521 poll->wait->private = poll;
Jens Axboe221c5eb2019-01-17 09:41:58 -07002522
Jens Axboe36703242019-07-25 10:20:18 -06002523 INIT_LIST_HEAD(&req->list);
2524
Jens Axboe221c5eb2019-01-17 09:41:58 -07002525 mask = vfs_poll(poll->file, &ipt.pt) & poll->events;
Jens Axboe221c5eb2019-01-17 09:41:58 -07002526
2527 spin_lock_irq(&ctx->completion_lock);
Jens Axboe8c838782019-03-12 15:48:16 -06002528 if (likely(poll->head)) {
2529 spin_lock(&poll->head->lock);
Jens Axboee9444752019-11-26 15:02:04 -07002530 if (unlikely(list_empty(&poll->wait->entry))) {
Jens Axboe8c838782019-03-12 15:48:16 -06002531 if (ipt.error)
2532 cancel = true;
2533 ipt.error = 0;
2534 mask = 0;
2535 }
2536 if (mask || ipt.error)
Jens Axboee9444752019-11-26 15:02:04 -07002537 list_del_init(&poll->wait->entry);
Jens Axboe8c838782019-03-12 15:48:16 -06002538 else if (cancel)
2539 WRITE_ONCE(poll->canceled, true);
2540 else if (!poll->done) /* actually waiting for an event */
Jens Axboeeac406c2019-11-14 12:09:58 -07002541 io_poll_req_insert(req);
Jens Axboe8c838782019-03-12 15:48:16 -06002542 spin_unlock(&poll->head->lock);
Jens Axboe221c5eb2019-01-17 09:41:58 -07002543 }
Jens Axboe8c838782019-03-12 15:48:16 -06002544 if (mask) { /* no async, we'd stolen it */
Jens Axboe8c838782019-03-12 15:48:16 -06002545 ipt.error = 0;
Jens Axboeb0dd8a42019-11-18 12:14:54 -07002546 io_poll_complete(req, mask, 0);
Jens Axboe8c838782019-03-12 15:48:16 -06002547 }
Jens Axboe221c5eb2019-01-17 09:41:58 -07002548 spin_unlock_irq(&ctx->completion_lock);
2549
Jens Axboe8c838782019-03-12 15:48:16 -06002550 if (mask) {
2551 io_cqring_ev_posted(ctx);
Jackie Liuec9c02a2019-11-08 23:50:36 +08002552 io_put_req_find_next(req, nxt);
Jens Axboe221c5eb2019-01-17 09:41:58 -07002553 }
Jens Axboe8c838782019-03-12 15:48:16 -06002554 return ipt.error;
Jens Axboe221c5eb2019-01-17 09:41:58 -07002555}
2556
Jens Axboe5262f562019-09-17 12:26:57 -06002557static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
2558{
Jens Axboead8a48a2019-11-15 08:49:11 -07002559 struct io_timeout_data *data = container_of(timer,
2560 struct io_timeout_data, timer);
2561 struct io_kiocb *req = data->req;
2562 struct io_ring_ctx *ctx = req->ctx;
Jens Axboe5262f562019-09-17 12:26:57 -06002563 unsigned long flags;
2564
Jens Axboe5262f562019-09-17 12:26:57 -06002565 atomic_inc(&ctx->cq_timeouts);
2566
2567 spin_lock_irqsave(&ctx->completion_lock, flags);
zhangyi (F)ef036812019-10-23 15:10:08 +08002568 /*
Jens Axboe11365042019-10-16 09:08:32 -06002569 * We could be racing with timeout deletion. If the list is empty,
2570 * then timeout lookup already found it and will be handling it.
zhangyi (F)ef036812019-10-23 15:10:08 +08002571 */
Jens Axboe842f9612019-10-29 12:34:10 -06002572 if (!list_empty(&req->list)) {
Jens Axboe11365042019-10-16 09:08:32 -06002573 struct io_kiocb *prev;
Jens Axboe5262f562019-09-17 12:26:57 -06002574
Jens Axboe11365042019-10-16 09:08:32 -06002575 /*
2576 * Adjust the reqs sequence before the current one because it
2577 * will consume a slot in the cq_ring and the the cq_tail
2578 * pointer will be increased, otherwise other timeout reqs may
2579 * return in advance without waiting for enough wait_nr.
2580 */
2581 prev = req;
2582 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
2583 prev->sequence++;
Jens Axboe11365042019-10-16 09:08:32 -06002584 list_del_init(&req->list);
Jens Axboe11365042019-10-16 09:08:32 -06002585 }
Jens Axboe842f9612019-10-29 12:34:10 -06002586
Jens Axboe78e19bb2019-11-06 15:21:34 -07002587 io_cqring_fill_event(req, -ETIME);
Jens Axboe5262f562019-09-17 12:26:57 -06002588 io_commit_cqring(ctx);
2589 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2590
2591 io_cqring_ev_posted(ctx);
Jens Axboe4e88d6e2019-12-07 20:59:47 -07002592 req_set_fail_links(req);
Jens Axboe5262f562019-09-17 12:26:57 -06002593 io_put_req(req);
2594 return HRTIMER_NORESTART;
2595}
2596
Jens Axboe47f46762019-11-09 17:43:02 -07002597static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
2598{
2599 struct io_kiocb *req;
2600 int ret = -ENOENT;
2601
2602 list_for_each_entry(req, &ctx->timeout_list, list) {
2603 if (user_data == req->user_data) {
2604 list_del_init(&req->list);
2605 ret = 0;
2606 break;
2607 }
2608 }
2609
2610 if (ret == -ENOENT)
2611 return ret;
2612
Jens Axboe2d283902019-12-04 11:08:05 -07002613 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
Jens Axboe47f46762019-11-09 17:43:02 -07002614 if (ret == -1)
2615 return -EALREADY;
2616
Jens Axboe4e88d6e2019-12-07 20:59:47 -07002617 req_set_fail_links(req);
Jens Axboe47f46762019-11-09 17:43:02 -07002618 io_cqring_fill_event(req, -ECANCELED);
2619 io_put_req(req);
2620 return 0;
2621}
2622
Jens Axboe11365042019-10-16 09:08:32 -06002623/*
2624 * Remove or update an existing timeout command
2625 */
2626static int io_timeout_remove(struct io_kiocb *req,
2627 const struct io_uring_sqe *sqe)
2628{
2629 struct io_ring_ctx *ctx = req->ctx;
Jens Axboe11365042019-10-16 09:08:32 -06002630 unsigned flags;
Jens Axboe47f46762019-11-09 17:43:02 -07002631 int ret;
Jens Axboe11365042019-10-16 09:08:32 -06002632
2633 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2634 return -EINVAL;
2635 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
2636 return -EINVAL;
2637 flags = READ_ONCE(sqe->timeout_flags);
2638 if (flags)
2639 return -EINVAL;
2640
Jens Axboe11365042019-10-16 09:08:32 -06002641 spin_lock_irq(&ctx->completion_lock);
Jens Axboe47f46762019-11-09 17:43:02 -07002642 ret = io_timeout_cancel(ctx, READ_ONCE(sqe->addr));
Jens Axboe11365042019-10-16 09:08:32 -06002643
Jens Axboe47f46762019-11-09 17:43:02 -07002644 io_cqring_fill_event(req, ret);
Jens Axboe11365042019-10-16 09:08:32 -06002645 io_commit_cqring(ctx);
2646 spin_unlock_irq(&ctx->completion_lock);
Jens Axboe5262f562019-09-17 12:26:57 -06002647 io_cqring_ev_posted(ctx);
Jens Axboe4e88d6e2019-12-07 20:59:47 -07002648 if (ret < 0)
2649 req_set_fail_links(req);
Jackie Liuec9c02a2019-11-08 23:50:36 +08002650 io_put_req(req);
Jens Axboe11365042019-10-16 09:08:32 -06002651 return 0;
Jens Axboe5262f562019-09-17 12:26:57 -06002652}
2653
Jens Axboe2d283902019-12-04 11:08:05 -07002654static int io_timeout_prep(struct io_kiocb *req, struct io_async_ctx *io,
2655 bool is_timeout_link)
Jens Axboe5262f562019-09-17 12:26:57 -06002656{
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03002657 const struct io_uring_sqe *sqe = req->sqe;
Jens Axboead8a48a2019-11-15 08:49:11 -07002658 struct io_timeout_data *data;
Jens Axboea41525a2019-10-15 16:48:15 -06002659 unsigned flags;
Jens Axboe5262f562019-09-17 12:26:57 -06002660
Jens Axboead8a48a2019-11-15 08:49:11 -07002661 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
Jens Axboe5262f562019-09-17 12:26:57 -06002662 return -EINVAL;
Jens Axboead8a48a2019-11-15 08:49:11 -07002663 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
Jens Axboea41525a2019-10-15 16:48:15 -06002664 return -EINVAL;
Jens Axboe2d283902019-12-04 11:08:05 -07002665 if (sqe->off && is_timeout_link)
2666 return -EINVAL;
Jens Axboea41525a2019-10-15 16:48:15 -06002667 flags = READ_ONCE(sqe->timeout_flags);
2668 if (flags & ~IORING_TIMEOUT_ABS)
Jens Axboe5262f562019-09-17 12:26:57 -06002669 return -EINVAL;
Arnd Bergmannbdf20072019-10-01 09:53:29 -06002670
Jens Axboe2d283902019-12-04 11:08:05 -07002671 data = &io->timeout;
Jens Axboead8a48a2019-11-15 08:49:11 -07002672 data->req = req;
Jens Axboead8a48a2019-11-15 08:49:11 -07002673 req->flags |= REQ_F_TIMEOUT;
2674
2675 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
Jens Axboe5262f562019-09-17 12:26:57 -06002676 return -EFAULT;
2677
Jens Axboe11365042019-10-16 09:08:32 -06002678 if (flags & IORING_TIMEOUT_ABS)
Jens Axboead8a48a2019-11-15 08:49:11 -07002679 data->mode = HRTIMER_MODE_ABS;
Jens Axboe11365042019-10-16 09:08:32 -06002680 else
Jens Axboead8a48a2019-11-15 08:49:11 -07002681 data->mode = HRTIMER_MODE_REL;
Jens Axboe11365042019-10-16 09:08:32 -06002682
Jens Axboead8a48a2019-11-15 08:49:11 -07002683 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
Jens Axboe2d283902019-12-04 11:08:05 -07002684 req->io = io;
Jens Axboead8a48a2019-11-15 08:49:11 -07002685 return 0;
2686}
2687
2688static int io_timeout(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2689{
2690 unsigned count;
2691 struct io_ring_ctx *ctx = req->ctx;
2692 struct io_timeout_data *data;
Jens Axboe2d283902019-12-04 11:08:05 -07002693 struct io_async_ctx *io;
Jens Axboead8a48a2019-11-15 08:49:11 -07002694 struct list_head *entry;
2695 unsigned span = 0;
Jens Axboead8a48a2019-11-15 08:49:11 -07002696
Jens Axboe2d283902019-12-04 11:08:05 -07002697 io = req->io;
2698 if (!io) {
2699 int ret;
2700
2701 io = kmalloc(sizeof(*io), GFP_KERNEL);
2702 if (!io)
2703 return -ENOMEM;
2704 ret = io_timeout_prep(req, io, false);
2705 if (ret) {
2706 kfree(io);
2707 return ret;
2708 }
2709 }
2710 data = &req->io->timeout;
Jens Axboe93bd25b2019-11-11 23:34:31 -07002711
Jens Axboe5262f562019-09-17 12:26:57 -06002712 /*
2713 * sqe->off holds how many events that need to occur for this
Jens Axboe93bd25b2019-11-11 23:34:31 -07002714 * timeout event to be satisfied. If it isn't set, then this is
2715 * a pure timeout request, sequence isn't used.
Jens Axboe5262f562019-09-17 12:26:57 -06002716 */
2717 count = READ_ONCE(sqe->off);
Jens Axboe93bd25b2019-11-11 23:34:31 -07002718 if (!count) {
2719 req->flags |= REQ_F_TIMEOUT_NOSEQ;
2720 spin_lock_irq(&ctx->completion_lock);
2721 entry = ctx->timeout_list.prev;
2722 goto add;
2723 }
Jens Axboe5262f562019-09-17 12:26:57 -06002724
2725 req->sequence = ctx->cached_sq_head + count - 1;
Jens Axboe2d283902019-12-04 11:08:05 -07002726 data->seq_offset = count;
Jens Axboe5262f562019-09-17 12:26:57 -06002727
2728 /*
2729 * Insertion sort, ensuring the first entry in the list is always
2730 * the one we need first.
2731 */
Jens Axboe5262f562019-09-17 12:26:57 -06002732 spin_lock_irq(&ctx->completion_lock);
2733 list_for_each_prev(entry, &ctx->timeout_list) {
2734 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
yangerkun5da0fb12019-10-15 21:59:29 +08002735 unsigned nxt_sq_head;
2736 long long tmp, tmp_nxt;
Jens Axboe2d283902019-12-04 11:08:05 -07002737 u32 nxt_offset = nxt->io->timeout.seq_offset;
Jens Axboe5262f562019-09-17 12:26:57 -06002738
Jens Axboe93bd25b2019-11-11 23:34:31 -07002739 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
2740 continue;
2741
yangerkun5da0fb12019-10-15 21:59:29 +08002742 /*
2743 * Since cached_sq_head + count - 1 can overflow, use type long
2744 * long to store it.
2745 */
2746 tmp = (long long)ctx->cached_sq_head + count - 1;
Pavel Begunkovcc42e0a2019-11-25 23:14:38 +03002747 nxt_sq_head = nxt->sequence - nxt_offset + 1;
2748 tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1;
yangerkun5da0fb12019-10-15 21:59:29 +08002749
2750 /*
2751 * cached_sq_head may overflow, and it will never overflow twice
2752 * once there is some timeout req still be valid.
2753 */
2754 if (ctx->cached_sq_head < nxt_sq_head)
yangerkun8b07a652019-10-17 12:12:35 +08002755 tmp += UINT_MAX;
yangerkun5da0fb12019-10-15 21:59:29 +08002756
zhangyi (F)a1f58ba2019-10-23 15:10:09 +08002757 if (tmp > tmp_nxt)
Jens Axboe5262f562019-09-17 12:26:57 -06002758 break;
zhangyi (F)a1f58ba2019-10-23 15:10:09 +08002759
2760 /*
2761 * Sequence of reqs after the insert one and itself should
2762 * be adjusted because each timeout req consumes a slot.
2763 */
2764 span++;
2765 nxt->sequence++;
Jens Axboe5262f562019-09-17 12:26:57 -06002766 }
zhangyi (F)a1f58ba2019-10-23 15:10:09 +08002767 req->sequence -= span;
Jens Axboe93bd25b2019-11-11 23:34:31 -07002768add:
Jens Axboe5262f562019-09-17 12:26:57 -06002769 list_add(&req->list, entry);
Jens Axboead8a48a2019-11-15 08:49:11 -07002770 data->timer.function = io_timeout_fn;
2771 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
Jens Axboe842f9612019-10-29 12:34:10 -06002772 spin_unlock_irq(&ctx->completion_lock);
Jens Axboe5262f562019-09-17 12:26:57 -06002773 return 0;
2774}
2775
Jens Axboe62755e32019-10-28 21:49:21 -06002776static bool io_cancel_cb(struct io_wq_work *work, void *data)
Jens Axboede0617e2019-04-06 21:51:27 -06002777{
Jens Axboe62755e32019-10-28 21:49:21 -06002778 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
Jens Axboede0617e2019-04-06 21:51:27 -06002779
Jens Axboe62755e32019-10-28 21:49:21 -06002780 return req->user_data == (unsigned long) data;
2781}
2782
Jens Axboee977d6d2019-11-05 12:39:45 -07002783static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
Jens Axboe62755e32019-10-28 21:49:21 -06002784{
Jens Axboe62755e32019-10-28 21:49:21 -06002785 enum io_wq_cancel cancel_ret;
Jens Axboe62755e32019-10-28 21:49:21 -06002786 int ret = 0;
2787
Jens Axboe62755e32019-10-28 21:49:21 -06002788 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
2789 switch (cancel_ret) {
2790 case IO_WQ_CANCEL_OK:
2791 ret = 0;
2792 break;
2793 case IO_WQ_CANCEL_RUNNING:
2794 ret = -EALREADY;
2795 break;
2796 case IO_WQ_CANCEL_NOTFOUND:
2797 ret = -ENOENT;
2798 break;
2799 }
2800
Jens Axboee977d6d2019-11-05 12:39:45 -07002801 return ret;
2802}
2803
Jens Axboe47f46762019-11-09 17:43:02 -07002804static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
2805 struct io_kiocb *req, __u64 sqe_addr,
Jens Axboeb0dd8a42019-11-18 12:14:54 -07002806 struct io_kiocb **nxt, int success_ret)
Jens Axboe47f46762019-11-09 17:43:02 -07002807{
2808 unsigned long flags;
2809 int ret;
2810
2811 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
2812 if (ret != -ENOENT) {
2813 spin_lock_irqsave(&ctx->completion_lock, flags);
2814 goto done;
2815 }
2816
2817 spin_lock_irqsave(&ctx->completion_lock, flags);
2818 ret = io_timeout_cancel(ctx, sqe_addr);
2819 if (ret != -ENOENT)
2820 goto done;
2821 ret = io_poll_cancel(ctx, sqe_addr);
2822done:
Jens Axboeb0dd8a42019-11-18 12:14:54 -07002823 if (!ret)
2824 ret = success_ret;
Jens Axboe47f46762019-11-09 17:43:02 -07002825 io_cqring_fill_event(req, ret);
2826 io_commit_cqring(ctx);
2827 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2828 io_cqring_ev_posted(ctx);
2829
Jens Axboe4e88d6e2019-12-07 20:59:47 -07002830 if (ret < 0)
2831 req_set_fail_links(req);
Jens Axboe47f46762019-11-09 17:43:02 -07002832 io_put_req_find_next(req, nxt);
2833}
2834
Jens Axboee977d6d2019-11-05 12:39:45 -07002835static int io_async_cancel(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2836 struct io_kiocb **nxt)
2837{
2838 struct io_ring_ctx *ctx = req->ctx;
Jens Axboee977d6d2019-11-05 12:39:45 -07002839
2840 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2841 return -EINVAL;
2842 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
2843 sqe->cancel_flags)
2844 return -EINVAL;
2845
Jens Axboeb0dd8a42019-11-18 12:14:54 -07002846 io_async_find_and_cancel(ctx, req, READ_ONCE(sqe->addr), nxt, 0);
Jens Axboe62755e32019-10-28 21:49:21 -06002847 return 0;
2848}
2849
Jens Axboef67676d2019-12-02 11:03:47 -07002850static int io_req_defer_prep(struct io_kiocb *req, struct io_async_ctx *io)
2851{
2852 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2853 struct iov_iter iter;
2854 ssize_t ret;
2855
2856 memcpy(&io->sqe, req->sqe, sizeof(io->sqe));
2857 req->sqe = &io->sqe;
2858
2859 switch (io->sqe.opcode) {
2860 case IORING_OP_READV:
2861 case IORING_OP_READ_FIXED:
2862 ret = io_read_prep(req, &iovec, &iter, true);
2863 break;
2864 case IORING_OP_WRITEV:
2865 case IORING_OP_WRITE_FIXED:
2866 ret = io_write_prep(req, &iovec, &iter, true);
2867 break;
Jens Axboe03b12302019-12-02 18:50:25 -07002868 case IORING_OP_SENDMSG:
2869 ret = io_sendmsg_prep(req, io);
2870 break;
2871 case IORING_OP_RECVMSG:
2872 ret = io_recvmsg_prep(req, io);
2873 break;
Jens Axboef499a022019-12-02 16:28:46 -07002874 case IORING_OP_CONNECT:
2875 ret = io_connect_prep(req, io);
2876 break;
Jens Axboe2d283902019-12-04 11:08:05 -07002877 case IORING_OP_TIMEOUT:
2878 return io_timeout_prep(req, io, false);
2879 case IORING_OP_LINK_TIMEOUT:
2880 return io_timeout_prep(req, io, true);
Jens Axboef67676d2019-12-02 11:03:47 -07002881 default:
2882 req->io = io;
2883 return 0;
2884 }
2885
2886 if (ret < 0)
2887 return ret;
2888
2889 req->io = io;
2890 io_req_map_io(req, ret, iovec, inline_vecs, &iter);
2891 return 0;
2892}
2893
Jackie Liua197f662019-11-08 08:09:12 -07002894static int io_req_defer(struct io_kiocb *req)
Jens Axboede0617e2019-04-06 21:51:27 -06002895{
Jackie Liua197f662019-11-08 08:09:12 -07002896 struct io_ring_ctx *ctx = req->ctx;
Jens Axboe1a6b74f2019-12-02 10:33:15 -07002897 struct io_async_ctx *io;
Jens Axboef67676d2019-12-02 11:03:47 -07002898 int ret;
Jens Axboede0617e2019-04-06 21:51:27 -06002899
Bob Liu9d858b22019-11-13 18:06:25 +08002900 /* Still need defer if there is pending req in defer list. */
2901 if (!req_need_defer(req) && list_empty(&ctx->defer_list))
Jens Axboede0617e2019-04-06 21:51:27 -06002902 return 0;
2903
Jens Axboe1a6b74f2019-12-02 10:33:15 -07002904 io = kmalloc(sizeof(*io), GFP_KERNEL);
2905 if (!io)
Jens Axboede0617e2019-04-06 21:51:27 -06002906 return -EAGAIN;
2907
Jens Axboe2d283902019-12-04 11:08:05 -07002908 ret = io_req_defer_prep(req, io);
2909 if (ret < 0) {
2910 kfree(io);
2911 return ret;
2912 }
2913
Jens Axboede0617e2019-04-06 21:51:27 -06002914 spin_lock_irq(&ctx->completion_lock);
Bob Liu9d858b22019-11-13 18:06:25 +08002915 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
Jens Axboede0617e2019-04-06 21:51:27 -06002916 spin_unlock_irq(&ctx->completion_lock);
Jens Axboede0617e2019-04-06 21:51:27 -06002917 return 0;
2918 }
2919
Jens Axboe915967f2019-11-21 09:01:20 -07002920 trace_io_uring_defer(ctx, req, req->user_data);
Jens Axboede0617e2019-04-06 21:51:27 -06002921 list_add_tail(&req->list, &ctx->defer_list);
2922 spin_unlock_irq(&ctx->completion_lock);
2923 return -EIOCBQUEUED;
2924}
2925
Pavel Begunkovf9bd67f2019-11-21 23:21:03 +03002926__attribute__((nonnull))
Pavel Begunkovd7324472019-11-21 21:24:36 +03002927static int io_issue_sqe(struct io_kiocb *req, struct io_kiocb **nxt,
2928 bool force_nonblock)
Jens Axboe2b188cc2019-01-07 10:46:33 -07002929{
Jens Axboee0c5c572019-03-12 10:18:47 -06002930 int ret, opcode;
Jackie Liua197f662019-11-08 08:09:12 -07002931 struct io_ring_ctx *ctx = req->ctx;
Jens Axboe2b188cc2019-01-07 10:46:33 -07002932
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03002933 opcode = READ_ONCE(req->sqe->opcode);
Jens Axboe2b188cc2019-01-07 10:46:33 -07002934 switch (opcode) {
2935 case IORING_OP_NOP:
Jens Axboe78e19bb2019-11-06 15:21:34 -07002936 ret = io_nop(req);
Jens Axboe2b188cc2019-01-07 10:46:33 -07002937 break;
2938 case IORING_OP_READV:
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03002939 if (unlikely(req->sqe->buf_index))
Jens Axboeedafcce2019-01-09 09:16:05 -07002940 return -EINVAL;
Pavel Begunkov267bc902019-11-07 01:41:08 +03002941 ret = io_read(req, nxt, force_nonblock);
Jens Axboe2b188cc2019-01-07 10:46:33 -07002942 break;
2943 case IORING_OP_WRITEV:
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03002944 if (unlikely(req->sqe->buf_index))
Jens Axboeedafcce2019-01-09 09:16:05 -07002945 return -EINVAL;
Pavel Begunkov267bc902019-11-07 01:41:08 +03002946 ret = io_write(req, nxt, force_nonblock);
Jens Axboeedafcce2019-01-09 09:16:05 -07002947 break;
2948 case IORING_OP_READ_FIXED:
Pavel Begunkov267bc902019-11-07 01:41:08 +03002949 ret = io_read(req, nxt, force_nonblock);
Jens Axboeedafcce2019-01-09 09:16:05 -07002950 break;
2951 case IORING_OP_WRITE_FIXED:
Pavel Begunkov267bc902019-11-07 01:41:08 +03002952 ret = io_write(req, nxt, force_nonblock);
Jens Axboe2b188cc2019-01-07 10:46:33 -07002953 break;
Christoph Hellwigc992fe22019-01-11 09:43:02 -07002954 case IORING_OP_FSYNC:
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03002955 ret = io_fsync(req, req->sqe, nxt, force_nonblock);
Christoph Hellwigc992fe22019-01-11 09:43:02 -07002956 break;
Jens Axboe221c5eb2019-01-17 09:41:58 -07002957 case IORING_OP_POLL_ADD:
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03002958 ret = io_poll_add(req, req->sqe, nxt);
Jens Axboe221c5eb2019-01-17 09:41:58 -07002959 break;
2960 case IORING_OP_POLL_REMOVE:
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03002961 ret = io_poll_remove(req, req->sqe);
Jens Axboe221c5eb2019-01-17 09:41:58 -07002962 break;
Jens Axboe5d17b4a2019-04-09 14:56:44 -06002963 case IORING_OP_SYNC_FILE_RANGE:
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03002964 ret = io_sync_file_range(req, req->sqe, nxt, force_nonblock);
Jens Axboe5d17b4a2019-04-09 14:56:44 -06002965 break;
Jens Axboe0fa03c62019-04-19 13:34:07 -06002966 case IORING_OP_SENDMSG:
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03002967 ret = io_sendmsg(req, req->sqe, nxt, force_nonblock);
Jens Axboe0fa03c62019-04-19 13:34:07 -06002968 break;
Jens Axboeaa1fa282019-04-19 13:38:09 -06002969 case IORING_OP_RECVMSG:
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03002970 ret = io_recvmsg(req, req->sqe, nxt, force_nonblock);
Jens Axboeaa1fa282019-04-19 13:38:09 -06002971 break;
Jens Axboe5262f562019-09-17 12:26:57 -06002972 case IORING_OP_TIMEOUT:
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03002973 ret = io_timeout(req, req->sqe);
Jens Axboe5262f562019-09-17 12:26:57 -06002974 break;
Jens Axboe11365042019-10-16 09:08:32 -06002975 case IORING_OP_TIMEOUT_REMOVE:
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03002976 ret = io_timeout_remove(req, req->sqe);
Jens Axboe11365042019-10-16 09:08:32 -06002977 break;
Jens Axboe17f2fe32019-10-17 14:42:58 -06002978 case IORING_OP_ACCEPT:
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03002979 ret = io_accept(req, req->sqe, nxt, force_nonblock);
Jens Axboe17f2fe32019-10-17 14:42:58 -06002980 break;
Jens Axboef8e85cf2019-11-23 14:24:24 -07002981 case IORING_OP_CONNECT:
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03002982 ret = io_connect(req, req->sqe, nxt, force_nonblock);
Jens Axboef8e85cf2019-11-23 14:24:24 -07002983 break;
Jens Axboe62755e32019-10-28 21:49:21 -06002984 case IORING_OP_ASYNC_CANCEL:
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03002985 ret = io_async_cancel(req, req->sqe, nxt);
Jens Axboe62755e32019-10-28 21:49:21 -06002986 break;
Jens Axboe2b188cc2019-01-07 10:46:33 -07002987 default:
2988 ret = -EINVAL;
2989 break;
2990 }
2991
Jens Axboedef596e2019-01-09 08:59:42 -07002992 if (ret)
2993 return ret;
2994
2995 if (ctx->flags & IORING_SETUP_IOPOLL) {
Jens Axboe9e645e112019-05-10 16:07:28 -06002996 if (req->result == -EAGAIN)
Jens Axboedef596e2019-01-09 08:59:42 -07002997 return -EAGAIN;
2998
2999 /* workqueue context doesn't hold uring_lock, grab it now */
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03003000 if (req->in_async)
Jens Axboedef596e2019-01-09 08:59:42 -07003001 mutex_lock(&ctx->uring_lock);
3002 io_iopoll_req_issued(req);
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03003003 if (req->in_async)
Jens Axboedef596e2019-01-09 08:59:42 -07003004 mutex_unlock(&ctx->uring_lock);
3005 }
3006
3007 return 0;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003008}
3009
Jens Axboeb76da702019-11-20 13:05:32 -07003010static void io_link_work_cb(struct io_wq_work **workptr)
3011{
3012 struct io_wq_work *work = *workptr;
3013 struct io_kiocb *link = work->data;
3014
3015 io_queue_linked_timeout(link);
3016 work->func = io_wq_submit_work;
3017}
3018
Jens Axboe561fb042019-10-24 07:25:42 -06003019static void io_wq_submit_work(struct io_wq_work **workptr)
Jens Axboe31b51512019-01-18 22:56:34 -07003020{
Jens Axboe561fb042019-10-24 07:25:42 -06003021 struct io_wq_work *work = *workptr;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003022 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
Jens Axboe561fb042019-10-24 07:25:42 -06003023 struct io_kiocb *nxt = NULL;
3024 int ret = 0;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003025
Jens Axboe561fb042019-10-24 07:25:42 -06003026 /* Ensure we clear previously set non-block flag */
3027 req->rw.ki_flags &= ~IOCB_NOWAIT;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003028
Jens Axboe561fb042019-10-24 07:25:42 -06003029 if (work->flags & IO_WQ_WORK_CANCEL)
3030 ret = -ECANCELED;
Jens Axboe31b51512019-01-18 22:56:34 -07003031
Jens Axboe561fb042019-10-24 07:25:42 -06003032 if (!ret) {
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03003033 req->has_user = (work->flags & IO_WQ_WORK_HAS_MM) != 0;
3034 req->in_async = true;
Jens Axboe561fb042019-10-24 07:25:42 -06003035 do {
Pavel Begunkovd7324472019-11-21 21:24:36 +03003036 ret = io_issue_sqe(req, &nxt, false);
Jens Axboe561fb042019-10-24 07:25:42 -06003037 /*
3038 * We can get EAGAIN for polled IO even though we're
3039 * forcing a sync submission from here, since we can't
3040 * wait for request slots on the block side.
3041 */
3042 if (ret != -EAGAIN)
3043 break;
3044 cond_resched();
3045 } while (1);
3046 }
Jens Axboe31b51512019-01-18 22:56:34 -07003047
Jens Axboe561fb042019-10-24 07:25:42 -06003048 /* drop submission reference */
Jackie Liuec9c02a2019-11-08 23:50:36 +08003049 io_put_req(req);
Jens Axboe817869d2019-04-30 14:44:05 -06003050
Jens Axboe561fb042019-10-24 07:25:42 -06003051 if (ret) {
Jens Axboe4e88d6e2019-12-07 20:59:47 -07003052 req_set_fail_links(req);
Jens Axboe78e19bb2019-11-06 15:21:34 -07003053 io_cqring_add_event(req, ret);
Jens Axboe817869d2019-04-30 14:44:05 -06003054 io_put_req(req);
Jens Axboeedafcce2019-01-09 09:16:05 -07003055 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07003056
Jens Axboe561fb042019-10-24 07:25:42 -06003057 /* if a dependent link is ready, pass it back */
3058 if (!ret && nxt) {
Jens Axboe94ae5e72019-11-14 19:39:52 -07003059 struct io_kiocb *link;
3060
3061 io_prep_async_work(nxt, &link);
Jens Axboe561fb042019-10-24 07:25:42 -06003062 *workptr = &nxt->work;
Jens Axboeb76da702019-11-20 13:05:32 -07003063 if (link) {
3064 nxt->work.flags |= IO_WQ_WORK_CB;
3065 nxt->work.func = io_link_work_cb;
3066 nxt->work.data = link;
3067 }
Jens Axboeedafcce2019-01-09 09:16:05 -07003068 }
Jens Axboe31b51512019-01-18 22:56:34 -07003069}
Jens Axboe2b188cc2019-01-07 10:46:33 -07003070
Jens Axboe09bb8392019-03-13 12:39:28 -06003071static bool io_op_needs_file(const struct io_uring_sqe *sqe)
3072{
3073 int op = READ_ONCE(sqe->opcode);
3074
3075 switch (op) {
3076 case IORING_OP_NOP:
3077 case IORING_OP_POLL_REMOVE:
Pavel Begunkov5683e542019-11-14 00:59:19 +03003078 case IORING_OP_TIMEOUT:
Pavel Begunkova320e9f2019-11-14 00:11:01 +03003079 case IORING_OP_TIMEOUT_REMOVE:
3080 case IORING_OP_ASYNC_CANCEL:
3081 case IORING_OP_LINK_TIMEOUT:
Jens Axboe09bb8392019-03-13 12:39:28 -06003082 return false;
3083 default:
3084 return true;
3085 }
3086}
3087
Jens Axboe65e19f52019-10-26 07:20:21 -06003088static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
3089 int index)
Jens Axboe09bb8392019-03-13 12:39:28 -06003090{
Jens Axboe65e19f52019-10-26 07:20:21 -06003091 struct fixed_file_table *table;
3092
3093 table = &ctx->file_table[index >> IORING_FILE_TABLE_SHIFT];
3094 return table->files[index & IORING_FILE_TABLE_MASK];
3095}
3096
Jackie Liua197f662019-11-08 08:09:12 -07003097static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req)
Jens Axboe09bb8392019-03-13 12:39:28 -06003098{
Jackie Liua197f662019-11-08 08:09:12 -07003099 struct io_ring_ctx *ctx = req->ctx;
Jens Axboe09bb8392019-03-13 12:39:28 -06003100 unsigned flags;
3101 int fd;
3102
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03003103 flags = READ_ONCE(req->sqe->flags);
3104 fd = READ_ONCE(req->sqe->fd);
Jens Axboe09bb8392019-03-13 12:39:28 -06003105
Jackie Liu4fe2c962019-09-09 20:50:40 +08003106 if (flags & IOSQE_IO_DRAIN)
Jens Axboede0617e2019-04-06 21:51:27 -06003107 req->flags |= REQ_F_IO_DRAIN;
Jens Axboede0617e2019-04-06 21:51:27 -06003108
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03003109 if (!io_op_needs_file(req->sqe))
Jens Axboe09bb8392019-03-13 12:39:28 -06003110 return 0;
Jens Axboe09bb8392019-03-13 12:39:28 -06003111
3112 if (flags & IOSQE_FIXED_FILE) {
Jens Axboe65e19f52019-10-26 07:20:21 -06003113 if (unlikely(!ctx->file_table ||
Jens Axboe09bb8392019-03-13 12:39:28 -06003114 (unsigned) fd >= ctx->nr_user_files))
3115 return -EBADF;
Jens Axboeb7620122019-10-26 07:22:55 -06003116 fd = array_index_nospec(fd, ctx->nr_user_files);
Jens Axboe65e19f52019-10-26 07:20:21 -06003117 req->file = io_file_from_index(ctx, fd);
3118 if (!req->file)
Jens Axboe08a45172019-10-03 08:11:03 -06003119 return -EBADF;
Jens Axboe09bb8392019-03-13 12:39:28 -06003120 req->flags |= REQ_F_FIXED_FILE;
3121 } else {
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03003122 if (req->needs_fixed_file)
Jens Axboe09bb8392019-03-13 12:39:28 -06003123 return -EBADF;
Dmitrii Dolgovc826bd72019-10-15 19:02:01 +02003124 trace_io_uring_file_get(ctx, fd);
Jens Axboe09bb8392019-03-13 12:39:28 -06003125 req->file = io_file_get(state, fd);
3126 if (unlikely(!req->file))
3127 return -EBADF;
3128 }
3129
3130 return 0;
3131}
3132
Jackie Liua197f662019-11-08 08:09:12 -07003133static int io_grab_files(struct io_kiocb *req)
Jens Axboe2b188cc2019-01-07 10:46:33 -07003134{
Jens Axboefcb323c2019-10-24 12:39:47 -06003135 int ret = -EBADF;
Jackie Liua197f662019-11-08 08:09:12 -07003136 struct io_ring_ctx *ctx = req->ctx;
Jens Axboefcb323c2019-10-24 12:39:47 -06003137
3138 rcu_read_lock();
3139 spin_lock_irq(&ctx->inflight_lock);
3140 /*
3141 * We use the f_ops->flush() handler to ensure that we can flush
3142 * out work accessing these files if the fd is closed. Check if
3143 * the fd has changed since we started down this path, and disallow
3144 * this operation if it has.
3145 */
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03003146 if (fcheck(req->ring_fd) == req->ring_file) {
Jens Axboefcb323c2019-10-24 12:39:47 -06003147 list_add(&req->inflight_entry, &ctx->inflight_list);
3148 req->flags |= REQ_F_INFLIGHT;
3149 req->work.files = current->files;
3150 ret = 0;
3151 }
3152 spin_unlock_irq(&ctx->inflight_lock);
3153 rcu_read_unlock();
3154
3155 return ret;
3156}
3157
Jens Axboe2665abf2019-11-05 12:40:47 -07003158static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
3159{
Jens Axboead8a48a2019-11-15 08:49:11 -07003160 struct io_timeout_data *data = container_of(timer,
3161 struct io_timeout_data, timer);
3162 struct io_kiocb *req = data->req;
Jens Axboe2665abf2019-11-05 12:40:47 -07003163 struct io_ring_ctx *ctx = req->ctx;
3164 struct io_kiocb *prev = NULL;
3165 unsigned long flags;
Jens Axboe2665abf2019-11-05 12:40:47 -07003166
3167 spin_lock_irqsave(&ctx->completion_lock, flags);
3168
3169 /*
3170 * We don't expect the list to be empty, that will only happen if we
3171 * race with the completion of the linked work.
3172 */
Pavel Begunkov44932332019-12-05 16:16:35 +03003173 if (!list_empty(&req->link_list)) {
3174 prev = list_entry(req->link_list.prev, struct io_kiocb,
3175 link_list);
Jens Axboe5d960722019-11-19 15:31:28 -07003176 if (refcount_inc_not_zero(&prev->refs)) {
Pavel Begunkov44932332019-12-05 16:16:35 +03003177 list_del_init(&req->link_list);
Jens Axboe5d960722019-11-19 15:31:28 -07003178 prev->flags &= ~REQ_F_LINK_TIMEOUT;
3179 } else
Jens Axboe76a46e02019-11-10 23:34:16 -07003180 prev = NULL;
Jens Axboe2665abf2019-11-05 12:40:47 -07003181 }
3182
3183 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3184
3185 if (prev) {
Jens Axboe4e88d6e2019-12-07 20:59:47 -07003186 req_set_fail_links(prev);
Jens Axboeb0dd8a42019-11-18 12:14:54 -07003187 io_async_find_and_cancel(ctx, req, prev->user_data, NULL,
3188 -ETIME);
Jens Axboe76a46e02019-11-10 23:34:16 -07003189 io_put_req(prev);
Jens Axboe47f46762019-11-09 17:43:02 -07003190 } else {
3191 io_cqring_add_event(req, -ETIME);
3192 io_put_req(req);
Jens Axboe2665abf2019-11-05 12:40:47 -07003193 }
Jens Axboe2665abf2019-11-05 12:40:47 -07003194 return HRTIMER_NORESTART;
3195}
3196
Jens Axboead8a48a2019-11-15 08:49:11 -07003197static void io_queue_linked_timeout(struct io_kiocb *req)
Jens Axboe2665abf2019-11-05 12:40:47 -07003198{
Jens Axboe76a46e02019-11-10 23:34:16 -07003199 struct io_ring_ctx *ctx = req->ctx;
Jens Axboe2665abf2019-11-05 12:40:47 -07003200
Jens Axboe76a46e02019-11-10 23:34:16 -07003201 /*
3202 * If the list is now empty, then our linked request finished before
3203 * we got a chance to setup the timer
3204 */
3205 spin_lock_irq(&ctx->completion_lock);
Pavel Begunkov44932332019-12-05 16:16:35 +03003206 if (!list_empty(&req->link_list)) {
Jens Axboe2d283902019-12-04 11:08:05 -07003207 struct io_timeout_data *data = &req->io->timeout;
Jens Axboe94ae5e72019-11-14 19:39:52 -07003208
Jens Axboead8a48a2019-11-15 08:49:11 -07003209 data->timer.function = io_link_timeout_fn;
3210 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
3211 data->mode);
Jens Axboe2665abf2019-11-05 12:40:47 -07003212 }
Jens Axboe76a46e02019-11-10 23:34:16 -07003213 spin_unlock_irq(&ctx->completion_lock);
Jens Axboe2665abf2019-11-05 12:40:47 -07003214
Jens Axboe2665abf2019-11-05 12:40:47 -07003215 /* drop submission reference */
Jens Axboe76a46e02019-11-10 23:34:16 -07003216 io_put_req(req);
Jens Axboe2665abf2019-11-05 12:40:47 -07003217}
3218
Jens Axboead8a48a2019-11-15 08:49:11 -07003219static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
Jens Axboe2665abf2019-11-05 12:40:47 -07003220{
3221 struct io_kiocb *nxt;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003222
Jens Axboe2665abf2019-11-05 12:40:47 -07003223 if (!(req->flags & REQ_F_LINK))
3224 return NULL;
3225
Pavel Begunkov44932332019-12-05 16:16:35 +03003226 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
3227 link_list);
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03003228 if (!nxt || nxt->sqe->opcode != IORING_OP_LINK_TIMEOUT)
Jens Axboe76a46e02019-11-10 23:34:16 -07003229 return NULL;
Jens Axboe2665abf2019-11-05 12:40:47 -07003230
Jens Axboe76a46e02019-11-10 23:34:16 -07003231 req->flags |= REQ_F_LINK_TIMEOUT;
Jens Axboe76a46e02019-11-10 23:34:16 -07003232 return nxt;
Jens Axboe2665abf2019-11-05 12:40:47 -07003233}
3234
Jens Axboe0e0702d2019-11-14 21:42:10 -07003235static void __io_queue_sqe(struct io_kiocb *req)
Jens Axboe2b188cc2019-01-07 10:46:33 -07003236{
Pavel Begunkovf9bd67f2019-11-21 23:21:03 +03003237 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
3238 struct io_kiocb *nxt = NULL;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003239 int ret;
3240
Pavel Begunkovf9bd67f2019-11-21 23:21:03 +03003241 ret = io_issue_sqe(req, &nxt, true);
3242 if (nxt)
3243 io_queue_async_work(nxt);
Jens Axboe491381ce2019-10-17 09:20:46 -06003244
3245 /*
3246 * We async punt it if the file wasn't marked NOWAIT, or if the file
3247 * doesn't support non-blocking read/write attempts
3248 */
3249 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
3250 (req->flags & REQ_F_MUST_PUNT))) {
Pavel Begunkovbbad27b2019-11-19 23:32:47 +03003251 if (req->work.flags & IO_WQ_WORK_NEEDS_FILES) {
3252 ret = io_grab_files(req);
3253 if (ret)
3254 goto err;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003255 }
Pavel Begunkovbbad27b2019-11-19 23:32:47 +03003256
3257 /*
3258 * Queued up for async execution, worker will release
3259 * submit reference when the iocb is actually submitted.
3260 */
3261 io_queue_async_work(req);
3262 return;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003263 }
Jens Axboee65ef562019-03-12 10:16:44 -06003264
Jens Axboefcb323c2019-10-24 12:39:47 -06003265err:
Jens Axboee65ef562019-03-12 10:16:44 -06003266 /* drop submission reference */
3267 io_put_req(req);
3268
Pavel Begunkovf9bd67f2019-11-21 23:21:03 +03003269 if (linked_timeout) {
Jens Axboe76a46e02019-11-10 23:34:16 -07003270 if (!ret)
Pavel Begunkovf9bd67f2019-11-21 23:21:03 +03003271 io_queue_linked_timeout(linked_timeout);
Jens Axboe76a46e02019-11-10 23:34:16 -07003272 else
Pavel Begunkovf9bd67f2019-11-21 23:21:03 +03003273 io_put_req(linked_timeout);
Jens Axboe76a46e02019-11-10 23:34:16 -07003274 }
3275
Jens Axboee65ef562019-03-12 10:16:44 -06003276 /* and drop final reference, if we failed */
Jens Axboe9e645e112019-05-10 16:07:28 -06003277 if (ret) {
Jens Axboe78e19bb2019-11-06 15:21:34 -07003278 io_cqring_add_event(req, ret);
Jens Axboe4e88d6e2019-12-07 20:59:47 -07003279 req_set_fail_links(req);
Jens Axboee65ef562019-03-12 10:16:44 -06003280 io_put_req(req);
Jens Axboe9e645e112019-05-10 16:07:28 -06003281 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07003282}
3283
Jens Axboe0e0702d2019-11-14 21:42:10 -07003284static void io_queue_sqe(struct io_kiocb *req)
Jackie Liu4fe2c962019-09-09 20:50:40 +08003285{
3286 int ret;
3287
Pavel Begunkov1b4a51b2019-11-21 11:54:28 +03003288 if (unlikely(req->ctx->drain_next)) {
3289 req->flags |= REQ_F_IO_DRAIN;
3290 req->ctx->drain_next = false;
3291 }
3292 req->ctx->drain_next = (req->flags & REQ_F_DRAIN_LINK);
3293
Jackie Liua197f662019-11-08 08:09:12 -07003294 ret = io_req_defer(req);
Jackie Liu4fe2c962019-09-09 20:50:40 +08003295 if (ret) {
3296 if (ret != -EIOCBQUEUED) {
Jens Axboe78e19bb2019-11-06 15:21:34 -07003297 io_cqring_add_event(req, ret);
Jens Axboe4e88d6e2019-12-07 20:59:47 -07003298 req_set_fail_links(req);
Jens Axboe78e19bb2019-11-06 15:21:34 -07003299 io_double_put_req(req);
Jackie Liu4fe2c962019-09-09 20:50:40 +08003300 }
Jens Axboe0e0702d2019-11-14 21:42:10 -07003301 } else
3302 __io_queue_sqe(req);
Jackie Liu4fe2c962019-09-09 20:50:40 +08003303}
3304
Pavel Begunkov1b4a51b2019-11-21 11:54:28 +03003305static inline void io_queue_link_head(struct io_kiocb *req)
Jackie Liu4fe2c962019-09-09 20:50:40 +08003306{
Jens Axboe94ae5e72019-11-14 19:39:52 -07003307 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
Pavel Begunkov1b4a51b2019-11-21 11:54:28 +03003308 io_cqring_add_event(req, -ECANCELED);
3309 io_double_put_req(req);
3310 } else
Jens Axboe0e0702d2019-11-14 21:42:10 -07003311 io_queue_sqe(req);
Jackie Liu4fe2c962019-09-09 20:50:40 +08003312}
3313
Pavel Begunkov1b4a51b2019-11-21 11:54:28 +03003314
Jens Axboe4e88d6e2019-12-07 20:59:47 -07003315#define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
3316 IOSQE_IO_HARDLINK)
Jens Axboe9e645e112019-05-10 16:07:28 -06003317
Pavel Begunkov2e6e1fd2019-12-05 16:15:45 +03003318static bool io_submit_sqe(struct io_kiocb *req, struct io_submit_state *state,
Jackie Liua197f662019-11-08 08:09:12 -07003319 struct io_kiocb **link)
Jens Axboe9e645e112019-05-10 16:07:28 -06003320{
Jackie Liua197f662019-11-08 08:09:12 -07003321 struct io_ring_ctx *ctx = req->ctx;
Jens Axboe9e645e112019-05-10 16:07:28 -06003322 int ret;
3323
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03003324 req->user_data = req->sqe->user_data;
Jens Axboe78e19bb2019-11-06 15:21:34 -07003325
Jens Axboe9e645e112019-05-10 16:07:28 -06003326 /* enforce forwards compatibility on users */
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03003327 if (unlikely(req->sqe->flags & ~SQE_VALID_FLAGS)) {
Jens Axboe9e645e112019-05-10 16:07:28 -06003328 ret = -EINVAL;
Pavel Begunkov196be952019-11-07 01:41:06 +03003329 goto err_req;
Jens Axboe9e645e112019-05-10 16:07:28 -06003330 }
3331
Jackie Liua197f662019-11-08 08:09:12 -07003332 ret = io_req_set_file(state, req);
Jens Axboe9e645e112019-05-10 16:07:28 -06003333 if (unlikely(ret)) {
3334err_req:
Jens Axboe78e19bb2019-11-06 15:21:34 -07003335 io_cqring_add_event(req, ret);
3336 io_double_put_req(req);
Pavel Begunkov2e6e1fd2019-12-05 16:15:45 +03003337 return false;
Jens Axboe9e645e112019-05-10 16:07:28 -06003338 }
3339
Jens Axboe9e645e112019-05-10 16:07:28 -06003340 /*
3341 * If we already have a head request, queue this one for async
3342 * submittal once the head completes. If we don't have a head but
3343 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
3344 * submitted sync once the chain is complete. If none of those
3345 * conditions are true (normal request), then just queue it.
3346 */
3347 if (*link) {
3348 struct io_kiocb *prev = *link;
Jens Axboe1a6b74f2019-12-02 10:33:15 -07003349 struct io_async_ctx *io;
Jens Axboe9e645e112019-05-10 16:07:28 -06003350
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03003351 if (req->sqe->flags & IOSQE_IO_DRAIN)
Pavel Begunkov1b4a51b2019-11-21 11:54:28 +03003352 (*link)->flags |= REQ_F_DRAIN_LINK | REQ_F_IO_DRAIN;
3353
Jens Axboe4e88d6e2019-12-07 20:59:47 -07003354 if (req->sqe->flags & IOSQE_IO_HARDLINK)
3355 req->flags |= REQ_F_HARDLINK;
3356
Jens Axboe1a6b74f2019-12-02 10:33:15 -07003357 io = kmalloc(sizeof(*io), GFP_KERNEL);
3358 if (!io) {
Jens Axboe9e645e112019-05-10 16:07:28 -06003359 ret = -EAGAIN;
3360 goto err_req;
3361 }
3362
Jens Axboef67676d2019-12-02 11:03:47 -07003363 ret = io_req_defer_prep(req, io);
Jens Axboe2d283902019-12-04 11:08:05 -07003364 if (ret) {
3365 kfree(io);
Jens Axboe4e88d6e2019-12-07 20:59:47 -07003366 /* fail even hard links since we don't submit */
Jens Axboe2d283902019-12-04 11:08:05 -07003367 prev->flags |= REQ_F_FAIL_LINK;
Jens Axboef67676d2019-12-02 11:03:47 -07003368 goto err_req;
Jens Axboe2d283902019-12-04 11:08:05 -07003369 }
Dmitrii Dolgovc826bd72019-10-15 19:02:01 +02003370 trace_io_uring_link(ctx, req, prev);
Pavel Begunkov44932332019-12-05 16:16:35 +03003371 list_add_tail(&req->link_list, &prev->link_list);
Jens Axboe4e88d6e2019-12-07 20:59:47 -07003372 } else if (req->sqe->flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) {
Jens Axboe9e645e112019-05-10 16:07:28 -06003373 req->flags |= REQ_F_LINK;
Jens Axboe4e88d6e2019-12-07 20:59:47 -07003374 if (req->sqe->flags & IOSQE_IO_HARDLINK)
3375 req->flags |= REQ_F_HARDLINK;
Jens Axboe9e645e112019-05-10 16:07:28 -06003376
Jens Axboe9e645e112019-05-10 16:07:28 -06003377 INIT_LIST_HEAD(&req->link_list);
3378 *link = req;
3379 } else {
Jackie Liua197f662019-11-08 08:09:12 -07003380 io_queue_sqe(req);
Jens Axboe9e645e112019-05-10 16:07:28 -06003381 }
Pavel Begunkov2e6e1fd2019-12-05 16:15:45 +03003382
3383 return true;
Jens Axboe9e645e112019-05-10 16:07:28 -06003384}
3385
Jens Axboe9a56a232019-01-09 09:06:50 -07003386/*
3387 * Batched submission is done, ensure local IO is flushed out.
3388 */
3389static void io_submit_state_end(struct io_submit_state *state)
3390{
3391 blk_finish_plug(&state->plug);
Jens Axboe3d6770f2019-04-13 11:50:54 -06003392 io_file_put(state);
Jens Axboe2579f912019-01-09 09:10:43 -07003393 if (state->free_reqs)
3394 kmem_cache_free_bulk(req_cachep, state->free_reqs,
3395 &state->reqs[state->cur_req]);
Jens Axboe9a56a232019-01-09 09:06:50 -07003396}
3397
3398/*
3399 * Start submission side cache.
3400 */
3401static void io_submit_state_start(struct io_submit_state *state,
Jackie Liu22efde52019-12-02 17:14:52 +08003402 unsigned int max_ios)
Jens Axboe9a56a232019-01-09 09:06:50 -07003403{
3404 blk_start_plug(&state->plug);
Jens Axboe2579f912019-01-09 09:10:43 -07003405 state->free_reqs = 0;
Jens Axboe9a56a232019-01-09 09:06:50 -07003406 state->file = NULL;
3407 state->ios_left = max_ios;
3408}
3409
Jens Axboe2b188cc2019-01-07 10:46:33 -07003410static void io_commit_sqring(struct io_ring_ctx *ctx)
3411{
Hristo Venev75b28af2019-08-26 17:23:46 +00003412 struct io_rings *rings = ctx->rings;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003413
Hristo Venev75b28af2019-08-26 17:23:46 +00003414 if (ctx->cached_sq_head != READ_ONCE(rings->sq.head)) {
Jens Axboe2b188cc2019-01-07 10:46:33 -07003415 /*
3416 * Ensure any loads from the SQEs are done at this point,
3417 * since once we write the new head, the application could
3418 * write new data to them.
3419 */
Hristo Venev75b28af2019-08-26 17:23:46 +00003420 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
Jens Axboe2b188cc2019-01-07 10:46:33 -07003421 }
3422}
3423
3424/*
Jens Axboe2b188cc2019-01-07 10:46:33 -07003425 * Fetch an sqe, if one is available. Note that s->sqe will point to memory
3426 * that is mapped by userspace. This means that care needs to be taken to
3427 * ensure that reads are stable, as we cannot rely on userspace always
3428 * being a good citizen. If members of the sqe are validated and then later
3429 * used, it's important that those reads are done through READ_ONCE() to
3430 * prevent a re-load down the line.
3431 */
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03003432static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req)
Jens Axboe2b188cc2019-01-07 10:46:33 -07003433{
Hristo Venev75b28af2019-08-26 17:23:46 +00003434 struct io_rings *rings = ctx->rings;
3435 u32 *sq_array = ctx->sq_array;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003436 unsigned head;
3437
3438 /*
3439 * The cached sq head (or cq tail) serves two purposes:
3440 *
3441 * 1) allows us to batch the cost of updating the user visible
3442 * head updates.
3443 * 2) allows the kernel side to track the head on its own, even
3444 * though the application is the one updating it.
3445 */
3446 head = ctx->cached_sq_head;
Stefan Bühlere523a292019-04-19 11:57:44 +02003447 /* make sure SQ entry isn't read before tail */
Pavel Begunkov9835d6f2019-11-21 21:24:56 +03003448 if (unlikely(head == smp_load_acquire(&rings->sq.tail)))
Jens Axboe2b188cc2019-01-07 10:46:33 -07003449 return false;
3450
Hristo Venev75b28af2019-08-26 17:23:46 +00003451 head = READ_ONCE(sq_array[head & ctx->sq_mask]);
Pavel Begunkov9835d6f2019-11-21 21:24:56 +03003452 if (likely(head < ctx->sq_entries)) {
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03003453 /*
3454 * All io need record the previous position, if LINK vs DARIN,
3455 * it can be used to mark the position of the first IO in the
3456 * link list.
3457 */
3458 req->sequence = ctx->cached_sq_head;
3459 req->sqe = &ctx->sq_sqes[head];
Jens Axboe2b188cc2019-01-07 10:46:33 -07003460 ctx->cached_sq_head++;
3461 return true;
3462 }
3463
3464 /* drop invalid entries */
3465 ctx->cached_sq_head++;
Jens Axboe498ccd92019-10-25 10:04:25 -06003466 ctx->cached_sq_dropped++;
3467 WRITE_ONCE(rings->sq_dropped, ctx->cached_sq_dropped);
Jens Axboe2b188cc2019-01-07 10:46:33 -07003468 return false;
3469}
3470
Pavel Begunkovfb5ccc92019-10-25 12:31:30 +03003471static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
Pavel Begunkovae9428c2019-11-06 00:22:14 +03003472 struct file *ring_file, int ring_fd,
3473 struct mm_struct **mm, bool async)
Jens Axboe6c271ce2019-01-10 11:22:30 -07003474{
3475 struct io_submit_state state, *statep = NULL;
Jens Axboe9e645e112019-05-10 16:07:28 -06003476 struct io_kiocb *link = NULL;
Jens Axboe9e645e112019-05-10 16:07:28 -06003477 int i, submitted = 0;
Pavel Begunkov95a1b3ff2019-10-27 23:15:41 +03003478 bool mm_fault = false;
Jens Axboe6c271ce2019-01-10 11:22:30 -07003479
Jens Axboec4a2ed72019-11-21 21:01:26 -07003480 /* if we have a backlog and couldn't flush it all, return BUSY */
3481 if (!list_empty(&ctx->cq_overflow_list) &&
3482 !io_cqring_overflow_flush(ctx, false))
Jens Axboe1d7bb1d2019-11-06 11:31:17 -07003483 return -EBUSY;
Jens Axboe6c271ce2019-01-10 11:22:30 -07003484
3485 if (nr > IO_PLUG_THRESHOLD) {
Jackie Liu22efde52019-12-02 17:14:52 +08003486 io_submit_state_start(&state, nr);
Jens Axboe6c271ce2019-01-10 11:22:30 -07003487 statep = &state;
3488 }
3489
3490 for (i = 0; i < nr; i++) {
Pavel Begunkov196be952019-11-07 01:41:06 +03003491 struct io_kiocb *req;
Pavel Begunkov50585b92019-11-07 01:41:07 +03003492 unsigned int sqe_flags;
Pavel Begunkovfb5ccc92019-10-25 12:31:30 +03003493
Pavel Begunkov196be952019-11-07 01:41:06 +03003494 req = io_get_req(ctx, statep);
3495 if (unlikely(!req)) {
3496 if (!submitted)
3497 submitted = -EAGAIN;
Pavel Begunkovfb5ccc92019-10-25 12:31:30 +03003498 break;
Jens Axboe9e645e112019-05-10 16:07:28 -06003499 }
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03003500 if (!io_get_sqring(ctx, req)) {
Pavel Begunkov196be952019-11-07 01:41:06 +03003501 __io_free_req(req);
3502 break;
3503 }
Jens Axboe9e645e112019-05-10 16:07:28 -06003504
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03003505 if (io_sqe_needs_user(req->sqe) && !*mm) {
Pavel Begunkov95a1b3ff2019-10-27 23:15:41 +03003506 mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm);
3507 if (!mm_fault) {
3508 use_mm(ctx->sqo_mm);
3509 *mm = ctx->sqo_mm;
3510 }
3511 }
3512
Pavel Begunkov2e6e1fd2019-12-05 16:15:45 +03003513 submitted++;
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03003514 sqe_flags = req->sqe->flags;
Pavel Begunkov50585b92019-11-07 01:41:07 +03003515
Pavel Begunkovcf6fd4b2019-11-25 23:14:39 +03003516 req->ring_file = ring_file;
3517 req->ring_fd = ring_fd;
3518 req->has_user = *mm != NULL;
3519 req->in_async = async;
3520 req->needs_fixed_file = async;
3521 trace_io_uring_submit_sqe(ctx, req->sqe->user_data,
Pavel Begunkov50585b92019-11-07 01:41:07 +03003522 true, async);
Pavel Begunkov2e6e1fd2019-12-05 16:15:45 +03003523 if (!io_submit_sqe(req, statep, &link))
3524 break;
Pavel Begunkove5eb6362019-11-06 00:22:15 +03003525 /*
3526 * If previous wasn't linked and we have a linked command,
3527 * that's the end of the chain. Submit the previous link.
3528 */
Pavel Begunkov50585b92019-11-07 01:41:07 +03003529 if (!(sqe_flags & IOSQE_IO_LINK) && link) {
Pavel Begunkov1b4a51b2019-11-21 11:54:28 +03003530 io_queue_link_head(link);
Pavel Begunkove5eb6362019-11-06 00:22:15 +03003531 link = NULL;
Jens Axboe6c271ce2019-01-10 11:22:30 -07003532 }
Jens Axboe6c271ce2019-01-10 11:22:30 -07003533 }
3534
Jens Axboe9e645e112019-05-10 16:07:28 -06003535 if (link)
Pavel Begunkov1b4a51b2019-11-21 11:54:28 +03003536 io_queue_link_head(link);
Jens Axboe6c271ce2019-01-10 11:22:30 -07003537 if (statep)
3538 io_submit_state_end(&state);
3539
Pavel Begunkovae9428c2019-11-06 00:22:14 +03003540 /* Commit SQ ring head once we've consumed and submitted all SQEs */
3541 io_commit_sqring(ctx);
3542
Jens Axboe6c271ce2019-01-10 11:22:30 -07003543 return submitted;
3544}
3545
3546static int io_sq_thread(void *data)
3547{
Jens Axboe6c271ce2019-01-10 11:22:30 -07003548 struct io_ring_ctx *ctx = data;
3549 struct mm_struct *cur_mm = NULL;
Jens Axboe181e4482019-11-25 08:52:30 -07003550 const struct cred *old_cred;
Jens Axboe6c271ce2019-01-10 11:22:30 -07003551 mm_segment_t old_fs;
3552 DEFINE_WAIT(wait);
3553 unsigned inflight;
3554 unsigned long timeout;
Jens Axboec1edbf52019-11-10 16:56:04 -07003555 int ret;
Jens Axboe6c271ce2019-01-10 11:22:30 -07003556
Jens Axboe206aefd2019-11-07 18:27:42 -07003557 complete(&ctx->completions[1]);
Jackie Liua4c0b3d2019-07-08 13:41:12 +08003558
Jens Axboe6c271ce2019-01-10 11:22:30 -07003559 old_fs = get_fs();
3560 set_fs(USER_DS);
Jens Axboe181e4482019-11-25 08:52:30 -07003561 old_cred = override_creds(ctx->creds);
Jens Axboe6c271ce2019-01-10 11:22:30 -07003562
Jens Axboec1edbf52019-11-10 16:56:04 -07003563 ret = timeout = inflight = 0;
Roman Penyaev2bbcd6d2019-05-16 10:53:57 +02003564 while (!kthread_should_park()) {
Pavel Begunkovfb5ccc92019-10-25 12:31:30 +03003565 unsigned int to_submit;
Jens Axboe6c271ce2019-01-10 11:22:30 -07003566
3567 if (inflight) {
3568 unsigned nr_events = 0;
3569
3570 if (ctx->flags & IORING_SETUP_IOPOLL) {
Jens Axboe2b2ed972019-10-25 10:06:15 -06003571 /*
3572 * inflight is the count of the maximum possible
3573 * entries we submitted, but it can be smaller
3574 * if we dropped some of them. If we don't have
3575 * poll entries available, then we know that we
3576 * have nothing left to poll for. Reset the
3577 * inflight count to zero in that case.
3578 */
3579 mutex_lock(&ctx->uring_lock);
3580 if (!list_empty(&ctx->poll_list))
3581 __io_iopoll_check(ctx, &nr_events, 0);
3582 else
3583 inflight = 0;
3584 mutex_unlock(&ctx->uring_lock);
Jens Axboe6c271ce2019-01-10 11:22:30 -07003585 } else {
3586 /*
3587 * Normal IO, just pretend everything completed.
3588 * We don't have to poll completions for that.
3589 */
3590 nr_events = inflight;
3591 }
3592
3593 inflight -= nr_events;
3594 if (!inflight)
3595 timeout = jiffies + ctx->sq_thread_idle;
3596 }
3597
Pavel Begunkovfb5ccc92019-10-25 12:31:30 +03003598 to_submit = io_sqring_entries(ctx);
Jens Axboec1edbf52019-11-10 16:56:04 -07003599
3600 /*
3601 * If submit got -EBUSY, flag us as needing the application
3602 * to enter the kernel to reap and flush events.
3603 */
3604 if (!to_submit || ret == -EBUSY) {
Jens Axboe6c271ce2019-01-10 11:22:30 -07003605 /*
3606 * We're polling. If we're within the defined idle
3607 * period, then let us spin without work before going
Jens Axboec1edbf52019-11-10 16:56:04 -07003608 * to sleep. The exception is if we got EBUSY doing
3609 * more IO, we should wait for the application to
3610 * reap events and wake us up.
Jens Axboe6c271ce2019-01-10 11:22:30 -07003611 */
Jens Axboec1edbf52019-11-10 16:56:04 -07003612 if (inflight ||
3613 (!time_after(jiffies, timeout) && ret != -EBUSY)) {
Jens Axboe9831a902019-09-19 09:48:55 -06003614 cond_resched();
Jens Axboe6c271ce2019-01-10 11:22:30 -07003615 continue;
3616 }
3617
3618 /*
3619 * Drop cur_mm before scheduling, we can't hold it for
3620 * long periods (or over schedule()). Do this before
3621 * adding ourselves to the waitqueue, as the unuse/drop
3622 * may sleep.
3623 */
3624 if (cur_mm) {
3625 unuse_mm(cur_mm);
3626 mmput(cur_mm);
3627 cur_mm = NULL;
3628 }
3629
3630 prepare_to_wait(&ctx->sqo_wait, &wait,
3631 TASK_INTERRUPTIBLE);
3632
3633 /* Tell userspace we may need a wakeup call */
Hristo Venev75b28af2019-08-26 17:23:46 +00003634 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
Stefan Bühler0d7bae62019-04-19 11:57:45 +02003635 /* make sure to read SQ tail after writing flags */
3636 smp_mb();
Jens Axboe6c271ce2019-01-10 11:22:30 -07003637
Pavel Begunkovfb5ccc92019-10-25 12:31:30 +03003638 to_submit = io_sqring_entries(ctx);
Jens Axboec1edbf52019-11-10 16:56:04 -07003639 if (!to_submit || ret == -EBUSY) {
Roman Penyaev2bbcd6d2019-05-16 10:53:57 +02003640 if (kthread_should_park()) {
Jens Axboe6c271ce2019-01-10 11:22:30 -07003641 finish_wait(&ctx->sqo_wait, &wait);
3642 break;
3643 }
3644 if (signal_pending(current))
3645 flush_signals(current);
3646 schedule();
3647 finish_wait(&ctx->sqo_wait, &wait);
3648
Hristo Venev75b28af2019-08-26 17:23:46 +00003649 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
Jens Axboe6c271ce2019-01-10 11:22:30 -07003650 continue;
3651 }
3652 finish_wait(&ctx->sqo_wait, &wait);
3653
Hristo Venev75b28af2019-08-26 17:23:46 +00003654 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
Jens Axboe6c271ce2019-01-10 11:22:30 -07003655 }
3656
Pavel Begunkovfb5ccc92019-10-25 12:31:30 +03003657 to_submit = min(to_submit, ctx->sq_entries);
Jens Axboe1d7bb1d2019-11-06 11:31:17 -07003658 ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true);
3659 if (ret > 0)
3660 inflight += ret;
Jens Axboe6c271ce2019-01-10 11:22:30 -07003661 }
3662
3663 set_fs(old_fs);
3664 if (cur_mm) {
3665 unuse_mm(cur_mm);
3666 mmput(cur_mm);
3667 }
Jens Axboe181e4482019-11-25 08:52:30 -07003668 revert_creds(old_cred);
Jens Axboe06058632019-04-13 09:26:03 -06003669
Roman Penyaev2bbcd6d2019-05-16 10:53:57 +02003670 kthread_parkme();
Jens Axboe06058632019-04-13 09:26:03 -06003671
Jens Axboe6c271ce2019-01-10 11:22:30 -07003672 return 0;
3673}
3674
Jens Axboebda52162019-09-24 13:47:15 -06003675struct io_wait_queue {
3676 struct wait_queue_entry wq;
3677 struct io_ring_ctx *ctx;
3678 unsigned to_wait;
3679 unsigned nr_timeouts;
3680};
3681
Jens Axboe1d7bb1d2019-11-06 11:31:17 -07003682static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
Jens Axboebda52162019-09-24 13:47:15 -06003683{
3684 struct io_ring_ctx *ctx = iowq->ctx;
3685
3686 /*
3687 * Wake up if we have enough events, or if a timeout occured since we
3688 * started waiting. For timeouts, we always want to return to userspace,
3689 * regardless of event count.
3690 */
Jens Axboe1d7bb1d2019-11-06 11:31:17 -07003691 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
Jens Axboebda52162019-09-24 13:47:15 -06003692 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
3693}
3694
3695static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
3696 int wake_flags, void *key)
3697{
3698 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
3699 wq);
3700
Jens Axboe1d7bb1d2019-11-06 11:31:17 -07003701 /* use noflush == true, as we can't safely rely on locking context */
3702 if (!io_should_wake(iowq, true))
Jens Axboebda52162019-09-24 13:47:15 -06003703 return -1;
3704
3705 return autoremove_wake_function(curr, mode, wake_flags, key);
3706}
3707
Jens Axboe2b188cc2019-01-07 10:46:33 -07003708/*
3709 * Wait until events become available, if we don't already have some. The
3710 * application must reap them itself, as they reside on the shared cq ring.
3711 */
3712static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
3713 const sigset_t __user *sig, size_t sigsz)
3714{
Jens Axboebda52162019-09-24 13:47:15 -06003715 struct io_wait_queue iowq = {
3716 .wq = {
3717 .private = current,
3718 .func = io_wake_function,
3719 .entry = LIST_HEAD_INIT(iowq.wq.entry),
3720 },
3721 .ctx = ctx,
3722 .to_wait = min_events,
3723 };
Hristo Venev75b28af2019-08-26 17:23:46 +00003724 struct io_rings *rings = ctx->rings;
Jackie Liue9ffa5c2019-10-29 11:16:42 +08003725 int ret = 0;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003726
Jens Axboe1d7bb1d2019-11-06 11:31:17 -07003727 if (io_cqring_events(ctx, false) >= min_events)
Jens Axboe2b188cc2019-01-07 10:46:33 -07003728 return 0;
3729
3730 if (sig) {
Arnd Bergmann9e75ad52019-03-25 15:34:53 +01003731#ifdef CONFIG_COMPAT
3732 if (in_compat_syscall())
3733 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
Oleg Nesterovb7724342019-07-16 16:29:53 -07003734 sigsz);
Arnd Bergmann9e75ad52019-03-25 15:34:53 +01003735 else
3736#endif
Oleg Nesterovb7724342019-07-16 16:29:53 -07003737 ret = set_user_sigmask(sig, sigsz);
Arnd Bergmann9e75ad52019-03-25 15:34:53 +01003738
Jens Axboe2b188cc2019-01-07 10:46:33 -07003739 if (ret)
3740 return ret;
3741 }
3742
Jens Axboebda52162019-09-24 13:47:15 -06003743 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
Dmitrii Dolgovc826bd72019-10-15 19:02:01 +02003744 trace_io_uring_cqring_wait(ctx, min_events);
Jens Axboebda52162019-09-24 13:47:15 -06003745 do {
3746 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
3747 TASK_INTERRUPTIBLE);
Jens Axboe1d7bb1d2019-11-06 11:31:17 -07003748 if (io_should_wake(&iowq, false))
Jens Axboebda52162019-09-24 13:47:15 -06003749 break;
3750 schedule();
3751 if (signal_pending(current)) {
Jackie Liue9ffa5c2019-10-29 11:16:42 +08003752 ret = -EINTR;
Jens Axboebda52162019-09-24 13:47:15 -06003753 break;
3754 }
3755 } while (1);
3756 finish_wait(&ctx->wait, &iowq.wq);
3757
Jackie Liue9ffa5c2019-10-29 11:16:42 +08003758 restore_saved_sigmask_unless(ret == -EINTR);
Jens Axboe2b188cc2019-01-07 10:46:33 -07003759
Hristo Venev75b28af2019-08-26 17:23:46 +00003760 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003761}
3762
Jens Axboe6b063142019-01-10 22:13:58 -07003763static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
3764{
3765#if defined(CONFIG_UNIX)
3766 if (ctx->ring_sock) {
3767 struct sock *sock = ctx->ring_sock->sk;
3768 struct sk_buff *skb;
3769
3770 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
3771 kfree_skb(skb);
3772 }
3773#else
3774 int i;
3775
Jens Axboe65e19f52019-10-26 07:20:21 -06003776 for (i = 0; i < ctx->nr_user_files; i++) {
3777 struct file *file;
3778
3779 file = io_file_from_index(ctx, i);
3780 if (file)
3781 fput(file);
3782 }
Jens Axboe6b063142019-01-10 22:13:58 -07003783#endif
3784}
3785
3786static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
3787{
Jens Axboe65e19f52019-10-26 07:20:21 -06003788 unsigned nr_tables, i;
3789
3790 if (!ctx->file_table)
Jens Axboe6b063142019-01-10 22:13:58 -07003791 return -ENXIO;
3792
3793 __io_sqe_files_unregister(ctx);
Jens Axboe65e19f52019-10-26 07:20:21 -06003794 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
3795 for (i = 0; i < nr_tables; i++)
3796 kfree(ctx->file_table[i].files);
3797 kfree(ctx->file_table);
3798 ctx->file_table = NULL;
Jens Axboe6b063142019-01-10 22:13:58 -07003799 ctx->nr_user_files = 0;
3800 return 0;
3801}
3802
Jens Axboe6c271ce2019-01-10 11:22:30 -07003803static void io_sq_thread_stop(struct io_ring_ctx *ctx)
3804{
3805 if (ctx->sqo_thread) {
Jens Axboe206aefd2019-11-07 18:27:42 -07003806 wait_for_completion(&ctx->completions[1]);
Roman Penyaev2bbcd6d2019-05-16 10:53:57 +02003807 /*
3808 * The park is a bit of a work-around, without it we get
3809 * warning spews on shutdown with SQPOLL set and affinity
3810 * set to a single CPU.
3811 */
Jens Axboe06058632019-04-13 09:26:03 -06003812 kthread_park(ctx->sqo_thread);
Jens Axboe6c271ce2019-01-10 11:22:30 -07003813 kthread_stop(ctx->sqo_thread);
3814 ctx->sqo_thread = NULL;
3815 }
3816}
3817
Jens Axboe6b063142019-01-10 22:13:58 -07003818static void io_finish_async(struct io_ring_ctx *ctx)
3819{
Jens Axboe6c271ce2019-01-10 11:22:30 -07003820 io_sq_thread_stop(ctx);
3821
Jens Axboe561fb042019-10-24 07:25:42 -06003822 if (ctx->io_wq) {
3823 io_wq_destroy(ctx->io_wq);
3824 ctx->io_wq = NULL;
Jens Axboe6b063142019-01-10 22:13:58 -07003825 }
3826}
3827
3828#if defined(CONFIG_UNIX)
3829static void io_destruct_skb(struct sk_buff *skb)
3830{
3831 struct io_ring_ctx *ctx = skb->sk->sk_user_data;
3832
Jens Axboe561fb042019-10-24 07:25:42 -06003833 if (ctx->io_wq)
3834 io_wq_flush(ctx->io_wq);
Jens Axboe8a997342019-10-09 14:40:13 -06003835
Jens Axboe6b063142019-01-10 22:13:58 -07003836 unix_destruct_scm(skb);
3837}
3838
3839/*
3840 * Ensure the UNIX gc is aware of our file set, so we are certain that
3841 * the io_uring can be safely unregistered on process exit, even if we have
3842 * loops in the file referencing.
3843 */
3844static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
3845{
3846 struct sock *sk = ctx->ring_sock->sk;
3847 struct scm_fp_list *fpl;
3848 struct sk_buff *skb;
Jens Axboe08a45172019-10-03 08:11:03 -06003849 int i, nr_files;
Jens Axboe6b063142019-01-10 22:13:58 -07003850
3851 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
3852 unsigned long inflight = ctx->user->unix_inflight + nr;
3853
3854 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
3855 return -EMFILE;
3856 }
3857
3858 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
3859 if (!fpl)
3860 return -ENOMEM;
3861
3862 skb = alloc_skb(0, GFP_KERNEL);
3863 if (!skb) {
3864 kfree(fpl);
3865 return -ENOMEM;
3866 }
3867
3868 skb->sk = sk;
Jens Axboe6b063142019-01-10 22:13:58 -07003869
Jens Axboe08a45172019-10-03 08:11:03 -06003870 nr_files = 0;
Jens Axboe6b063142019-01-10 22:13:58 -07003871 fpl->user = get_uid(ctx->user);
3872 for (i = 0; i < nr; i++) {
Jens Axboe65e19f52019-10-26 07:20:21 -06003873 struct file *file = io_file_from_index(ctx, i + offset);
3874
3875 if (!file)
Jens Axboe08a45172019-10-03 08:11:03 -06003876 continue;
Jens Axboe65e19f52019-10-26 07:20:21 -06003877 fpl->fp[nr_files] = get_file(file);
Jens Axboe08a45172019-10-03 08:11:03 -06003878 unix_inflight(fpl->user, fpl->fp[nr_files]);
3879 nr_files++;
Jens Axboe6b063142019-01-10 22:13:58 -07003880 }
3881
Jens Axboe08a45172019-10-03 08:11:03 -06003882 if (nr_files) {
3883 fpl->max = SCM_MAX_FD;
3884 fpl->count = nr_files;
3885 UNIXCB(skb).fp = fpl;
3886 skb->destructor = io_destruct_skb;
3887 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
3888 skb_queue_head(&sk->sk_receive_queue, skb);
Jens Axboe6b063142019-01-10 22:13:58 -07003889
Jens Axboe08a45172019-10-03 08:11:03 -06003890 for (i = 0; i < nr_files; i++)
3891 fput(fpl->fp[i]);
3892 } else {
3893 kfree_skb(skb);
3894 kfree(fpl);
3895 }
Jens Axboe6b063142019-01-10 22:13:58 -07003896
3897 return 0;
3898}
3899
3900/*
3901 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
3902 * causes regular reference counting to break down. We rely on the UNIX
3903 * garbage collection to take care of this problem for us.
3904 */
3905static int io_sqe_files_scm(struct io_ring_ctx *ctx)
3906{
3907 unsigned left, total;
3908 int ret = 0;
3909
3910 total = 0;
3911 left = ctx->nr_user_files;
3912 while (left) {
3913 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
Jens Axboe6b063142019-01-10 22:13:58 -07003914
3915 ret = __io_sqe_files_scm(ctx, this_files, total);
3916 if (ret)
3917 break;
3918 left -= this_files;
3919 total += this_files;
3920 }
3921
3922 if (!ret)
3923 return 0;
3924
3925 while (total < ctx->nr_user_files) {
Jens Axboe65e19f52019-10-26 07:20:21 -06003926 struct file *file = io_file_from_index(ctx, total);
3927
3928 if (file)
3929 fput(file);
Jens Axboe6b063142019-01-10 22:13:58 -07003930 total++;
3931 }
3932
3933 return ret;
3934}
3935#else
3936static int io_sqe_files_scm(struct io_ring_ctx *ctx)
3937{
3938 return 0;
3939}
3940#endif
3941
Jens Axboe65e19f52019-10-26 07:20:21 -06003942static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
3943 unsigned nr_files)
3944{
3945 int i;
3946
3947 for (i = 0; i < nr_tables; i++) {
3948 struct fixed_file_table *table = &ctx->file_table[i];
3949 unsigned this_files;
3950
3951 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
3952 table->files = kcalloc(this_files, sizeof(struct file *),
3953 GFP_KERNEL);
3954 if (!table->files)
3955 break;
3956 nr_files -= this_files;
3957 }
3958
3959 if (i == nr_tables)
3960 return 0;
3961
3962 for (i = 0; i < nr_tables; i++) {
3963 struct fixed_file_table *table = &ctx->file_table[i];
3964 kfree(table->files);
3965 }
3966 return 1;
3967}
3968
Jens Axboe6b063142019-01-10 22:13:58 -07003969static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
3970 unsigned nr_args)
3971{
3972 __s32 __user *fds = (__s32 __user *) arg;
Jens Axboe65e19f52019-10-26 07:20:21 -06003973 unsigned nr_tables;
Jens Axboe6b063142019-01-10 22:13:58 -07003974 int fd, ret = 0;
3975 unsigned i;
3976
Jens Axboe65e19f52019-10-26 07:20:21 -06003977 if (ctx->file_table)
Jens Axboe6b063142019-01-10 22:13:58 -07003978 return -EBUSY;
3979 if (!nr_args)
3980 return -EINVAL;
3981 if (nr_args > IORING_MAX_FIXED_FILES)
3982 return -EMFILE;
3983
Jens Axboe65e19f52019-10-26 07:20:21 -06003984 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
3985 ctx->file_table = kcalloc(nr_tables, sizeof(struct fixed_file_table),
3986 GFP_KERNEL);
3987 if (!ctx->file_table)
Jens Axboe6b063142019-01-10 22:13:58 -07003988 return -ENOMEM;
3989
Jens Axboe65e19f52019-10-26 07:20:21 -06003990 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
3991 kfree(ctx->file_table);
Jens Axboe46568e92019-11-10 08:40:53 -07003992 ctx->file_table = NULL;
Jens Axboe65e19f52019-10-26 07:20:21 -06003993 return -ENOMEM;
3994 }
3995
Jens Axboe08a45172019-10-03 08:11:03 -06003996 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
Jens Axboe65e19f52019-10-26 07:20:21 -06003997 struct fixed_file_table *table;
3998 unsigned index;
3999
Jens Axboe6b063142019-01-10 22:13:58 -07004000 ret = -EFAULT;
4001 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
4002 break;
Jens Axboe08a45172019-10-03 08:11:03 -06004003 /* allow sparse sets */
4004 if (fd == -1) {
4005 ret = 0;
4006 continue;
4007 }
Jens Axboe6b063142019-01-10 22:13:58 -07004008
Jens Axboe65e19f52019-10-26 07:20:21 -06004009 table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT];
4010 index = i & IORING_FILE_TABLE_MASK;
4011 table->files[index] = fget(fd);
Jens Axboe6b063142019-01-10 22:13:58 -07004012
4013 ret = -EBADF;
Jens Axboe65e19f52019-10-26 07:20:21 -06004014 if (!table->files[index])
Jens Axboe6b063142019-01-10 22:13:58 -07004015 break;
4016 /*
4017 * Don't allow io_uring instances to be registered. If UNIX
4018 * isn't enabled, then this causes a reference cycle and this
4019 * instance can never get freed. If UNIX is enabled we'll
4020 * handle it just fine, but there's still no point in allowing
4021 * a ring fd as it doesn't support regular read/write anyway.
4022 */
Jens Axboe65e19f52019-10-26 07:20:21 -06004023 if (table->files[index]->f_op == &io_uring_fops) {
4024 fput(table->files[index]);
Jens Axboe6b063142019-01-10 22:13:58 -07004025 break;
4026 }
Jens Axboe6b063142019-01-10 22:13:58 -07004027 ret = 0;
4028 }
4029
4030 if (ret) {
Jens Axboe65e19f52019-10-26 07:20:21 -06004031 for (i = 0; i < ctx->nr_user_files; i++) {
4032 struct file *file;
Jens Axboe6b063142019-01-10 22:13:58 -07004033
Jens Axboe65e19f52019-10-26 07:20:21 -06004034 file = io_file_from_index(ctx, i);
4035 if (file)
4036 fput(file);
4037 }
4038 for (i = 0; i < nr_tables; i++)
4039 kfree(ctx->file_table[i].files);
4040
4041 kfree(ctx->file_table);
4042 ctx->file_table = NULL;
Jens Axboe6b063142019-01-10 22:13:58 -07004043 ctx->nr_user_files = 0;
4044 return ret;
4045 }
4046
4047 ret = io_sqe_files_scm(ctx);
4048 if (ret)
4049 io_sqe_files_unregister(ctx);
4050
4051 return ret;
4052}
4053
Jens Axboec3a31e62019-10-03 13:59:56 -06004054static void io_sqe_file_unregister(struct io_ring_ctx *ctx, int index)
4055{
4056#if defined(CONFIG_UNIX)
Jens Axboe65e19f52019-10-26 07:20:21 -06004057 struct file *file = io_file_from_index(ctx, index);
Jens Axboec3a31e62019-10-03 13:59:56 -06004058 struct sock *sock = ctx->ring_sock->sk;
4059 struct sk_buff_head list, *head = &sock->sk_receive_queue;
4060 struct sk_buff *skb;
4061 int i;
4062
4063 __skb_queue_head_init(&list);
4064
4065 /*
4066 * Find the skb that holds this file in its SCM_RIGHTS. When found,
4067 * remove this entry and rearrange the file array.
4068 */
4069 skb = skb_dequeue(head);
4070 while (skb) {
4071 struct scm_fp_list *fp;
4072
4073 fp = UNIXCB(skb).fp;
4074 for (i = 0; i < fp->count; i++) {
4075 int left;
4076
4077 if (fp->fp[i] != file)
4078 continue;
4079
4080 unix_notinflight(fp->user, fp->fp[i]);
4081 left = fp->count - 1 - i;
4082 if (left) {
4083 memmove(&fp->fp[i], &fp->fp[i + 1],
4084 left * sizeof(struct file *));
4085 }
4086 fp->count--;
4087 if (!fp->count) {
4088 kfree_skb(skb);
4089 skb = NULL;
4090 } else {
4091 __skb_queue_tail(&list, skb);
4092 }
4093 fput(file);
4094 file = NULL;
4095 break;
4096 }
4097
4098 if (!file)
4099 break;
4100
4101 __skb_queue_tail(&list, skb);
4102
4103 skb = skb_dequeue(head);
4104 }
4105
4106 if (skb_peek(&list)) {
4107 spin_lock_irq(&head->lock);
4108 while ((skb = __skb_dequeue(&list)) != NULL)
4109 __skb_queue_tail(head, skb);
4110 spin_unlock_irq(&head->lock);
4111 }
4112#else
Jens Axboe65e19f52019-10-26 07:20:21 -06004113 fput(io_file_from_index(ctx, index));
Jens Axboec3a31e62019-10-03 13:59:56 -06004114#endif
4115}
4116
4117static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
4118 int index)
4119{
4120#if defined(CONFIG_UNIX)
4121 struct sock *sock = ctx->ring_sock->sk;
4122 struct sk_buff_head *head = &sock->sk_receive_queue;
4123 struct sk_buff *skb;
4124
4125 /*
4126 * See if we can merge this file into an existing skb SCM_RIGHTS
4127 * file set. If there's no room, fall back to allocating a new skb
4128 * and filling it in.
4129 */
4130 spin_lock_irq(&head->lock);
4131 skb = skb_peek(head);
4132 if (skb) {
4133 struct scm_fp_list *fpl = UNIXCB(skb).fp;
4134
4135 if (fpl->count < SCM_MAX_FD) {
4136 __skb_unlink(skb, head);
4137 spin_unlock_irq(&head->lock);
4138 fpl->fp[fpl->count] = get_file(file);
4139 unix_inflight(fpl->user, fpl->fp[fpl->count]);
4140 fpl->count++;
4141 spin_lock_irq(&head->lock);
4142 __skb_queue_head(head, skb);
4143 } else {
4144 skb = NULL;
4145 }
4146 }
4147 spin_unlock_irq(&head->lock);
4148
4149 if (skb) {
4150 fput(file);
4151 return 0;
4152 }
4153
4154 return __io_sqe_files_scm(ctx, 1, index);
4155#else
4156 return 0;
4157#endif
4158}
4159
4160static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
4161 unsigned nr_args)
4162{
4163 struct io_uring_files_update up;
4164 __s32 __user *fds;
4165 int fd, i, err;
4166 __u32 done;
4167
Jens Axboe65e19f52019-10-26 07:20:21 -06004168 if (!ctx->file_table)
Jens Axboec3a31e62019-10-03 13:59:56 -06004169 return -ENXIO;
4170 if (!nr_args)
4171 return -EINVAL;
4172 if (copy_from_user(&up, arg, sizeof(up)))
4173 return -EFAULT;
4174 if (check_add_overflow(up.offset, nr_args, &done))
4175 return -EOVERFLOW;
4176 if (done > ctx->nr_user_files)
4177 return -EINVAL;
4178
4179 done = 0;
4180 fds = (__s32 __user *) up.fds;
4181 while (nr_args) {
Jens Axboe65e19f52019-10-26 07:20:21 -06004182 struct fixed_file_table *table;
4183 unsigned index;
4184
Jens Axboec3a31e62019-10-03 13:59:56 -06004185 err = 0;
4186 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
4187 err = -EFAULT;
4188 break;
4189 }
4190 i = array_index_nospec(up.offset, ctx->nr_user_files);
Jens Axboe65e19f52019-10-26 07:20:21 -06004191 table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT];
4192 index = i & IORING_FILE_TABLE_MASK;
4193 if (table->files[index]) {
Jens Axboec3a31e62019-10-03 13:59:56 -06004194 io_sqe_file_unregister(ctx, i);
Jens Axboe65e19f52019-10-26 07:20:21 -06004195 table->files[index] = NULL;
Jens Axboec3a31e62019-10-03 13:59:56 -06004196 }
4197 if (fd != -1) {
4198 struct file *file;
4199
4200 file = fget(fd);
4201 if (!file) {
4202 err = -EBADF;
4203 break;
4204 }
4205 /*
4206 * Don't allow io_uring instances to be registered. If
4207 * UNIX isn't enabled, then this causes a reference
4208 * cycle and this instance can never get freed. If UNIX
4209 * is enabled we'll handle it just fine, but there's
4210 * still no point in allowing a ring fd as it doesn't
4211 * support regular read/write anyway.
4212 */
4213 if (file->f_op == &io_uring_fops) {
4214 fput(file);
4215 err = -EBADF;
4216 break;
4217 }
Jens Axboe65e19f52019-10-26 07:20:21 -06004218 table->files[index] = file;
Jens Axboec3a31e62019-10-03 13:59:56 -06004219 err = io_sqe_file_register(ctx, file, i);
4220 if (err)
4221 break;
4222 }
4223 nr_args--;
4224 done++;
4225 up.offset++;
4226 }
4227
4228 return done ? done : err;
4229}
4230
Jens Axboe7d723062019-11-12 22:31:31 -07004231static void io_put_work(struct io_wq_work *work)
4232{
4233 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4234
4235 io_put_req(req);
4236}
4237
4238static void io_get_work(struct io_wq_work *work)
4239{
4240 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4241
4242 refcount_inc(&req->refs);
4243}
4244
Jens Axboe6c271ce2019-01-10 11:22:30 -07004245static int io_sq_offload_start(struct io_ring_ctx *ctx,
4246 struct io_uring_params *p)
Jens Axboe2b188cc2019-01-07 10:46:33 -07004247{
Jens Axboe576a3472019-11-25 08:49:20 -07004248 struct io_wq_data data;
Jens Axboe561fb042019-10-24 07:25:42 -06004249 unsigned concurrency;
Jens Axboe2b188cc2019-01-07 10:46:33 -07004250 int ret;
4251
Jens Axboe6c271ce2019-01-10 11:22:30 -07004252 init_waitqueue_head(&ctx->sqo_wait);
Jens Axboe2b188cc2019-01-07 10:46:33 -07004253 mmgrab(current->mm);
4254 ctx->sqo_mm = current->mm;
4255
Jens Axboe6c271ce2019-01-10 11:22:30 -07004256 if (ctx->flags & IORING_SETUP_SQPOLL) {
Jens Axboe3ec482d2019-04-08 10:51:01 -06004257 ret = -EPERM;
4258 if (!capable(CAP_SYS_ADMIN))
4259 goto err;
4260
Jens Axboe917257d2019-04-13 09:28:55 -06004261 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
4262 if (!ctx->sq_thread_idle)
4263 ctx->sq_thread_idle = HZ;
4264
Jens Axboe6c271ce2019-01-10 11:22:30 -07004265 if (p->flags & IORING_SETUP_SQ_AFF) {
Jens Axboe44a9bd12019-05-14 20:00:30 -06004266 int cpu = p->sq_thread_cpu;
Jens Axboe6c271ce2019-01-10 11:22:30 -07004267
Jens Axboe917257d2019-04-13 09:28:55 -06004268 ret = -EINVAL;
Jens Axboe44a9bd12019-05-14 20:00:30 -06004269 if (cpu >= nr_cpu_ids)
4270 goto err;
Shenghui Wang7889f442019-05-07 16:03:19 +08004271 if (!cpu_online(cpu))
Jens Axboe917257d2019-04-13 09:28:55 -06004272 goto err;
4273
Jens Axboe6c271ce2019-01-10 11:22:30 -07004274 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
4275 ctx, cpu,
4276 "io_uring-sq");
4277 } else {
4278 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
4279 "io_uring-sq");
4280 }
4281 if (IS_ERR(ctx->sqo_thread)) {
4282 ret = PTR_ERR(ctx->sqo_thread);
4283 ctx->sqo_thread = NULL;
4284 goto err;
4285 }
4286 wake_up_process(ctx->sqo_thread);
4287 } else if (p->flags & IORING_SETUP_SQ_AFF) {
4288 /* Can't have SQ_AFF without SQPOLL */
4289 ret = -EINVAL;
4290 goto err;
4291 }
4292
Jens Axboe576a3472019-11-25 08:49:20 -07004293 data.mm = ctx->sqo_mm;
4294 data.user = ctx->user;
Jens Axboe181e4482019-11-25 08:52:30 -07004295 data.creds = ctx->creds;
Jens Axboe576a3472019-11-25 08:49:20 -07004296 data.get_work = io_get_work;
4297 data.put_work = io_put_work;
4298
Jens Axboe561fb042019-10-24 07:25:42 -06004299 /* Do QD, or 4 * CPUS, whatever is smallest */
4300 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
Jens Axboe576a3472019-11-25 08:49:20 -07004301 ctx->io_wq = io_wq_create(concurrency, &data);
Jens Axboe975c99a52019-10-30 08:42:56 -06004302 if (IS_ERR(ctx->io_wq)) {
4303 ret = PTR_ERR(ctx->io_wq);
4304 ctx->io_wq = NULL;
Jens Axboe2b188cc2019-01-07 10:46:33 -07004305 goto err;
4306 }
4307
4308 return 0;
4309err:
Jens Axboe54a91f32019-09-10 09:15:04 -06004310 io_finish_async(ctx);
Jens Axboe2b188cc2019-01-07 10:46:33 -07004311 mmdrop(ctx->sqo_mm);
4312 ctx->sqo_mm = NULL;
4313 return ret;
4314}
4315
4316static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
4317{
4318 atomic_long_sub(nr_pages, &user->locked_vm);
4319}
4320
4321static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
4322{
4323 unsigned long page_limit, cur_pages, new_pages;
4324
4325 /* Don't allow more pages than we can safely lock */
4326 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
4327
4328 do {
4329 cur_pages = atomic_long_read(&user->locked_vm);
4330 new_pages = cur_pages + nr_pages;
4331 if (new_pages > page_limit)
4332 return -ENOMEM;
4333 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
4334 new_pages) != cur_pages);
4335
4336 return 0;
4337}
4338
4339static void io_mem_free(void *ptr)
4340{
Mark Rutland52e04ef2019-04-30 17:30:21 +01004341 struct page *page;
Jens Axboe2b188cc2019-01-07 10:46:33 -07004342
Mark Rutland52e04ef2019-04-30 17:30:21 +01004343 if (!ptr)
4344 return;
4345
4346 page = virt_to_head_page(ptr);
Jens Axboe2b188cc2019-01-07 10:46:33 -07004347 if (put_page_testzero(page))
4348 free_compound_page(page);
4349}
4350
4351static void *io_mem_alloc(size_t size)
4352{
4353 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
4354 __GFP_NORETRY;
4355
4356 return (void *) __get_free_pages(gfp_flags, get_order(size));
4357}
4358
Hristo Venev75b28af2019-08-26 17:23:46 +00004359static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
4360 size_t *sq_offset)
4361{
4362 struct io_rings *rings;
4363 size_t off, sq_array_size;
4364
4365 off = struct_size(rings, cqes, cq_entries);
4366 if (off == SIZE_MAX)
4367 return SIZE_MAX;
4368
4369#ifdef CONFIG_SMP
4370 off = ALIGN(off, SMP_CACHE_BYTES);
4371 if (off == 0)
4372 return SIZE_MAX;
4373#endif
4374
4375 sq_array_size = array_size(sizeof(u32), sq_entries);
4376 if (sq_array_size == SIZE_MAX)
4377 return SIZE_MAX;
4378
4379 if (check_add_overflow(off, sq_array_size, &off))
4380 return SIZE_MAX;
4381
4382 if (sq_offset)
4383 *sq_offset = off;
4384
4385 return off;
4386}
4387
Jens Axboe2b188cc2019-01-07 10:46:33 -07004388static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
4389{
Hristo Venev75b28af2019-08-26 17:23:46 +00004390 size_t pages;
Jens Axboe2b188cc2019-01-07 10:46:33 -07004391
Hristo Venev75b28af2019-08-26 17:23:46 +00004392 pages = (size_t)1 << get_order(
4393 rings_size(sq_entries, cq_entries, NULL));
4394 pages += (size_t)1 << get_order(
4395 array_size(sizeof(struct io_uring_sqe), sq_entries));
Jens Axboe2b188cc2019-01-07 10:46:33 -07004396
Hristo Venev75b28af2019-08-26 17:23:46 +00004397 return pages;
Jens Axboe2b188cc2019-01-07 10:46:33 -07004398}
4399
Jens Axboeedafcce2019-01-09 09:16:05 -07004400static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
4401{
4402 int i, j;
4403
4404 if (!ctx->user_bufs)
4405 return -ENXIO;
4406
4407 for (i = 0; i < ctx->nr_user_bufs; i++) {
4408 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
4409
4410 for (j = 0; j < imu->nr_bvecs; j++)
John Hubbard27c4d3a2019-08-04 19:32:06 -07004411 put_user_page(imu->bvec[j].bv_page);
Jens Axboeedafcce2019-01-09 09:16:05 -07004412
4413 if (ctx->account_mem)
4414 io_unaccount_mem(ctx->user, imu->nr_bvecs);
Mark Rutlandd4ef6472019-05-01 16:59:16 +01004415 kvfree(imu->bvec);
Jens Axboeedafcce2019-01-09 09:16:05 -07004416 imu->nr_bvecs = 0;
4417 }
4418
4419 kfree(ctx->user_bufs);
4420 ctx->user_bufs = NULL;
4421 ctx->nr_user_bufs = 0;
4422 return 0;
4423}
4424
4425static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
4426 void __user *arg, unsigned index)
4427{
4428 struct iovec __user *src;
4429
4430#ifdef CONFIG_COMPAT
4431 if (ctx->compat) {
4432 struct compat_iovec __user *ciovs;
4433 struct compat_iovec ciov;
4434
4435 ciovs = (struct compat_iovec __user *) arg;
4436 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
4437 return -EFAULT;
4438
4439 dst->iov_base = (void __user *) (unsigned long) ciov.iov_base;
4440 dst->iov_len = ciov.iov_len;
4441 return 0;
4442 }
4443#endif
4444 src = (struct iovec __user *) arg;
4445 if (copy_from_user(dst, &src[index], sizeof(*dst)))
4446 return -EFAULT;
4447 return 0;
4448}
4449
4450static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
4451 unsigned nr_args)
4452{
4453 struct vm_area_struct **vmas = NULL;
4454 struct page **pages = NULL;
4455 int i, j, got_pages = 0;
4456 int ret = -EINVAL;
4457
4458 if (ctx->user_bufs)
4459 return -EBUSY;
4460 if (!nr_args || nr_args > UIO_MAXIOV)
4461 return -EINVAL;
4462
4463 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
4464 GFP_KERNEL);
4465 if (!ctx->user_bufs)
4466 return -ENOMEM;
4467
4468 for (i = 0; i < nr_args; i++) {
4469 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
4470 unsigned long off, start, end, ubuf;
4471 int pret, nr_pages;
4472 struct iovec iov;
4473 size_t size;
4474
4475 ret = io_copy_iov(ctx, &iov, arg, i);
4476 if (ret)
Pavel Begunkova2786822019-05-26 12:35:47 +03004477 goto err;
Jens Axboeedafcce2019-01-09 09:16:05 -07004478
4479 /*
4480 * Don't impose further limits on the size and buffer
4481 * constraints here, we'll -EINVAL later when IO is
4482 * submitted if they are wrong.
4483 */
4484 ret = -EFAULT;
4485 if (!iov.iov_base || !iov.iov_len)
4486 goto err;
4487
4488 /* arbitrary limit, but we need something */
4489 if (iov.iov_len > SZ_1G)
4490 goto err;
4491
4492 ubuf = (unsigned long) iov.iov_base;
4493 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
4494 start = ubuf >> PAGE_SHIFT;
4495 nr_pages = end - start;
4496
4497 if (ctx->account_mem) {
4498 ret = io_account_mem(ctx->user, nr_pages);
4499 if (ret)
4500 goto err;
4501 }
4502
4503 ret = 0;
4504 if (!pages || nr_pages > got_pages) {
4505 kfree(vmas);
4506 kfree(pages);
Mark Rutlandd4ef6472019-05-01 16:59:16 +01004507 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
Jens Axboeedafcce2019-01-09 09:16:05 -07004508 GFP_KERNEL);
Mark Rutlandd4ef6472019-05-01 16:59:16 +01004509 vmas = kvmalloc_array(nr_pages,
Jens Axboeedafcce2019-01-09 09:16:05 -07004510 sizeof(struct vm_area_struct *),
4511 GFP_KERNEL);
4512 if (!pages || !vmas) {
4513 ret = -ENOMEM;
4514 if (ctx->account_mem)
4515 io_unaccount_mem(ctx->user, nr_pages);
4516 goto err;
4517 }
4518 got_pages = nr_pages;
4519 }
4520
Mark Rutlandd4ef6472019-05-01 16:59:16 +01004521 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
Jens Axboeedafcce2019-01-09 09:16:05 -07004522 GFP_KERNEL);
4523 ret = -ENOMEM;
4524 if (!imu->bvec) {
4525 if (ctx->account_mem)
4526 io_unaccount_mem(ctx->user, nr_pages);
4527 goto err;
4528 }
4529
4530 ret = 0;
4531 down_read(&current->mm->mmap_sem);
Ira Weiny932f4a62019-05-13 17:17:03 -07004532 pret = get_user_pages(ubuf, nr_pages,
4533 FOLL_WRITE | FOLL_LONGTERM,
4534 pages, vmas);
Jens Axboeedafcce2019-01-09 09:16:05 -07004535 if (pret == nr_pages) {
4536 /* don't support file backed memory */
4537 for (j = 0; j < nr_pages; j++) {
4538 struct vm_area_struct *vma = vmas[j];
4539
4540 if (vma->vm_file &&
4541 !is_file_hugepages(vma->vm_file)) {
4542 ret = -EOPNOTSUPP;
4543 break;
4544 }
4545 }
4546 } else {
4547 ret = pret < 0 ? pret : -EFAULT;
4548 }
4549 up_read(&current->mm->mmap_sem);
4550 if (ret) {
4551 /*
4552 * if we did partial map, or found file backed vmas,
4553 * release any pages we did get
4554 */
John Hubbard27c4d3a2019-08-04 19:32:06 -07004555 if (pret > 0)
4556 put_user_pages(pages, pret);
Jens Axboeedafcce2019-01-09 09:16:05 -07004557 if (ctx->account_mem)
4558 io_unaccount_mem(ctx->user, nr_pages);
Mark Rutlandd4ef6472019-05-01 16:59:16 +01004559 kvfree(imu->bvec);
Jens Axboeedafcce2019-01-09 09:16:05 -07004560 goto err;
4561 }
4562
4563 off = ubuf & ~PAGE_MASK;
4564 size = iov.iov_len;
4565 for (j = 0; j < nr_pages; j++) {
4566 size_t vec_len;
4567
4568 vec_len = min_t(size_t, size, PAGE_SIZE - off);
4569 imu->bvec[j].bv_page = pages[j];
4570 imu->bvec[j].bv_len = vec_len;
4571 imu->bvec[j].bv_offset = off;
4572 off = 0;
4573 size -= vec_len;
4574 }
4575 /* store original address for later verification */
4576 imu->ubuf = ubuf;
4577 imu->len = iov.iov_len;
4578 imu->nr_bvecs = nr_pages;
4579
4580 ctx->nr_user_bufs++;
4581 }
Mark Rutlandd4ef6472019-05-01 16:59:16 +01004582 kvfree(pages);
4583 kvfree(vmas);
Jens Axboeedafcce2019-01-09 09:16:05 -07004584 return 0;
4585err:
Mark Rutlandd4ef6472019-05-01 16:59:16 +01004586 kvfree(pages);
4587 kvfree(vmas);
Jens Axboeedafcce2019-01-09 09:16:05 -07004588 io_sqe_buffer_unregister(ctx);
4589 return ret;
4590}
4591
Jens Axboe9b402842019-04-11 11:45:41 -06004592static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
4593{
4594 __s32 __user *fds = arg;
4595 int fd;
4596
4597 if (ctx->cq_ev_fd)
4598 return -EBUSY;
4599
4600 if (copy_from_user(&fd, fds, sizeof(*fds)))
4601 return -EFAULT;
4602
4603 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
4604 if (IS_ERR(ctx->cq_ev_fd)) {
4605 int ret = PTR_ERR(ctx->cq_ev_fd);
4606 ctx->cq_ev_fd = NULL;
4607 return ret;
4608 }
4609
4610 return 0;
4611}
4612
4613static int io_eventfd_unregister(struct io_ring_ctx *ctx)
4614{
4615 if (ctx->cq_ev_fd) {
4616 eventfd_ctx_put(ctx->cq_ev_fd);
4617 ctx->cq_ev_fd = NULL;
4618 return 0;
4619 }
4620
4621 return -ENXIO;
4622}
4623
Jens Axboe2b188cc2019-01-07 10:46:33 -07004624static void io_ring_ctx_free(struct io_ring_ctx *ctx)
4625{
Jens Axboe6b063142019-01-10 22:13:58 -07004626 io_finish_async(ctx);
Jens Axboe2b188cc2019-01-07 10:46:33 -07004627 if (ctx->sqo_mm)
4628 mmdrop(ctx->sqo_mm);
Jens Axboedef596e2019-01-09 08:59:42 -07004629
4630 io_iopoll_reap_events(ctx);
Jens Axboeedafcce2019-01-09 09:16:05 -07004631 io_sqe_buffer_unregister(ctx);
Jens Axboe6b063142019-01-10 22:13:58 -07004632 io_sqe_files_unregister(ctx);
Jens Axboe9b402842019-04-11 11:45:41 -06004633 io_eventfd_unregister(ctx);
Jens Axboedef596e2019-01-09 08:59:42 -07004634
Jens Axboe2b188cc2019-01-07 10:46:33 -07004635#if defined(CONFIG_UNIX)
Eric Biggers355e8d22019-06-12 14:58:43 -07004636 if (ctx->ring_sock) {
4637 ctx->ring_sock->file = NULL; /* so that iput() is called */
Jens Axboe2b188cc2019-01-07 10:46:33 -07004638 sock_release(ctx->ring_sock);
Eric Biggers355e8d22019-06-12 14:58:43 -07004639 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07004640#endif
4641
Hristo Venev75b28af2019-08-26 17:23:46 +00004642 io_mem_free(ctx->rings);
Jens Axboe2b188cc2019-01-07 10:46:33 -07004643 io_mem_free(ctx->sq_sqes);
Jens Axboe2b188cc2019-01-07 10:46:33 -07004644
4645 percpu_ref_exit(&ctx->refs);
4646 if (ctx->account_mem)
4647 io_unaccount_mem(ctx->user,
4648 ring_pages(ctx->sq_entries, ctx->cq_entries));
4649 free_uid(ctx->user);
Jens Axboe181e4482019-11-25 08:52:30 -07004650 put_cred(ctx->creds);
Jens Axboe206aefd2019-11-07 18:27:42 -07004651 kfree(ctx->completions);
Jens Axboe78076bb2019-12-04 19:56:40 -07004652 kfree(ctx->cancel_hash);
Jens Axboe0ddf92e2019-11-08 08:52:53 -07004653 kmem_cache_free(req_cachep, ctx->fallback_req);
Jens Axboe2b188cc2019-01-07 10:46:33 -07004654 kfree(ctx);
4655}
4656
4657static __poll_t io_uring_poll(struct file *file, poll_table *wait)
4658{
4659 struct io_ring_ctx *ctx = file->private_data;
4660 __poll_t mask = 0;
4661
4662 poll_wait(file, &ctx->cq_wait, wait);
Stefan Bühler4f7067c2019-04-24 23:54:17 +02004663 /*
4664 * synchronizes with barrier from wq_has_sleeper call in
4665 * io_commit_cqring
4666 */
Jens Axboe2b188cc2019-01-07 10:46:33 -07004667 smp_rmb();
Hristo Venev75b28af2019-08-26 17:23:46 +00004668 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
4669 ctx->rings->sq_ring_entries)
Jens Axboe2b188cc2019-01-07 10:46:33 -07004670 mask |= EPOLLOUT | EPOLLWRNORM;
yangerkundaa5de52019-09-24 20:53:34 +08004671 if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail)
Jens Axboe2b188cc2019-01-07 10:46:33 -07004672 mask |= EPOLLIN | EPOLLRDNORM;
4673
4674 return mask;
4675}
4676
4677static int io_uring_fasync(int fd, struct file *file, int on)
4678{
4679 struct io_ring_ctx *ctx = file->private_data;
4680
4681 return fasync_helper(fd, file, on, &ctx->cq_fasync);
4682}
4683
4684static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
4685{
4686 mutex_lock(&ctx->uring_lock);
4687 percpu_ref_kill(&ctx->refs);
4688 mutex_unlock(&ctx->uring_lock);
4689
Jens Axboe5262f562019-09-17 12:26:57 -06004690 io_kill_timeouts(ctx);
Jens Axboe221c5eb2019-01-17 09:41:58 -07004691 io_poll_remove_all(ctx);
Jens Axboe561fb042019-10-24 07:25:42 -06004692
4693 if (ctx->io_wq)
4694 io_wq_cancel_all(ctx->io_wq);
4695
Jens Axboedef596e2019-01-09 08:59:42 -07004696 io_iopoll_reap_events(ctx);
Jens Axboe15dff282019-11-13 09:09:23 -07004697 /* if we failed setting up the ctx, we might not have any rings */
4698 if (ctx->rings)
4699 io_cqring_overflow_flush(ctx, true);
Jens Axboe206aefd2019-11-07 18:27:42 -07004700 wait_for_completion(&ctx->completions[0]);
Jens Axboe2b188cc2019-01-07 10:46:33 -07004701 io_ring_ctx_free(ctx);
4702}
4703
4704static int io_uring_release(struct inode *inode, struct file *file)
4705{
4706 struct io_ring_ctx *ctx = file->private_data;
4707
4708 file->private_data = NULL;
4709 io_ring_ctx_wait_and_kill(ctx);
4710 return 0;
4711}
4712
Jens Axboefcb323c2019-10-24 12:39:47 -06004713static void io_uring_cancel_files(struct io_ring_ctx *ctx,
4714 struct files_struct *files)
4715{
4716 struct io_kiocb *req;
4717 DEFINE_WAIT(wait);
4718
4719 while (!list_empty_careful(&ctx->inflight_list)) {
Jens Axboe768134d2019-11-10 20:30:53 -07004720 struct io_kiocb *cancel_req = NULL;
Jens Axboefcb323c2019-10-24 12:39:47 -06004721
4722 spin_lock_irq(&ctx->inflight_lock);
4723 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
Jens Axboe768134d2019-11-10 20:30:53 -07004724 if (req->work.files != files)
4725 continue;
4726 /* req is being completed, ignore */
4727 if (!refcount_inc_not_zero(&req->refs))
4728 continue;
4729 cancel_req = req;
4730 break;
Jens Axboefcb323c2019-10-24 12:39:47 -06004731 }
Jens Axboe768134d2019-11-10 20:30:53 -07004732 if (cancel_req)
Jens Axboefcb323c2019-10-24 12:39:47 -06004733 prepare_to_wait(&ctx->inflight_wait, &wait,
Jens Axboe768134d2019-11-10 20:30:53 -07004734 TASK_UNINTERRUPTIBLE);
Jens Axboefcb323c2019-10-24 12:39:47 -06004735 spin_unlock_irq(&ctx->inflight_lock);
4736
Jens Axboe768134d2019-11-10 20:30:53 -07004737 /* We need to keep going until we don't find a matching req */
4738 if (!cancel_req)
Jens Axboefcb323c2019-10-24 12:39:47 -06004739 break;
Bob Liu2f6d9b92019-11-13 18:06:24 +08004740
4741 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
4742 io_put_req(cancel_req);
Jens Axboefcb323c2019-10-24 12:39:47 -06004743 schedule();
4744 }
Jens Axboe768134d2019-11-10 20:30:53 -07004745 finish_wait(&ctx->inflight_wait, &wait);
Jens Axboefcb323c2019-10-24 12:39:47 -06004746}
4747
4748static int io_uring_flush(struct file *file, void *data)
4749{
4750 struct io_ring_ctx *ctx = file->private_data;
4751
4752 io_uring_cancel_files(ctx, data);
Jens Axboe1d7bb1d2019-11-06 11:31:17 -07004753 if (fatal_signal_pending(current) || (current->flags & PF_EXITING)) {
4754 io_cqring_overflow_flush(ctx, true);
Jens Axboefcb323c2019-10-24 12:39:47 -06004755 io_wq_cancel_all(ctx->io_wq);
Jens Axboe1d7bb1d2019-11-06 11:31:17 -07004756 }
Jens Axboefcb323c2019-10-24 12:39:47 -06004757 return 0;
4758}
4759
Roman Penyaev6c5c2402019-11-28 12:53:22 +01004760static void *io_uring_validate_mmap_request(struct file *file,
4761 loff_t pgoff, size_t sz)
Jens Axboe2b188cc2019-01-07 10:46:33 -07004762{
Jens Axboe2b188cc2019-01-07 10:46:33 -07004763 struct io_ring_ctx *ctx = file->private_data;
Roman Penyaev6c5c2402019-11-28 12:53:22 +01004764 loff_t offset = pgoff << PAGE_SHIFT;
Jens Axboe2b188cc2019-01-07 10:46:33 -07004765 struct page *page;
4766 void *ptr;
4767
4768 switch (offset) {
4769 case IORING_OFF_SQ_RING:
Hristo Venev75b28af2019-08-26 17:23:46 +00004770 case IORING_OFF_CQ_RING:
4771 ptr = ctx->rings;
Jens Axboe2b188cc2019-01-07 10:46:33 -07004772 break;
4773 case IORING_OFF_SQES:
4774 ptr = ctx->sq_sqes;
4775 break;
Jens Axboe2b188cc2019-01-07 10:46:33 -07004776 default:
Roman Penyaev6c5c2402019-11-28 12:53:22 +01004777 return ERR_PTR(-EINVAL);
Jens Axboe2b188cc2019-01-07 10:46:33 -07004778 }
4779
4780 page = virt_to_head_page(ptr);
Matthew Wilcox (Oracle)a50b8542019-09-23 15:34:25 -07004781 if (sz > page_size(page))
Roman Penyaev6c5c2402019-11-28 12:53:22 +01004782 return ERR_PTR(-EINVAL);
4783
4784 return ptr;
4785}
4786
4787#ifdef CONFIG_MMU
4788
4789static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
4790{
4791 size_t sz = vma->vm_end - vma->vm_start;
4792 unsigned long pfn;
4793 void *ptr;
4794
4795 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
4796 if (IS_ERR(ptr))
4797 return PTR_ERR(ptr);
Jens Axboe2b188cc2019-01-07 10:46:33 -07004798
4799 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
4800 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
4801}
4802
Roman Penyaev6c5c2402019-11-28 12:53:22 +01004803#else /* !CONFIG_MMU */
4804
4805static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
4806{
4807 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
4808}
4809
4810static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
4811{
4812 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
4813}
4814
4815static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
4816 unsigned long addr, unsigned long len,
4817 unsigned long pgoff, unsigned long flags)
4818{
4819 void *ptr;
4820
4821 ptr = io_uring_validate_mmap_request(file, pgoff, len);
4822 if (IS_ERR(ptr))
4823 return PTR_ERR(ptr);
4824
4825 return (unsigned long) ptr;
4826}
4827
4828#endif /* !CONFIG_MMU */
4829
Jens Axboe2b188cc2019-01-07 10:46:33 -07004830SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
4831 u32, min_complete, u32, flags, const sigset_t __user *, sig,
4832 size_t, sigsz)
4833{
4834 struct io_ring_ctx *ctx;
4835 long ret = -EBADF;
4836 int submitted = 0;
4837 struct fd f;
4838
Jens Axboe6c271ce2019-01-10 11:22:30 -07004839 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
Jens Axboe2b188cc2019-01-07 10:46:33 -07004840 return -EINVAL;
4841
4842 f = fdget(fd);
4843 if (!f.file)
4844 return -EBADF;
4845
4846 ret = -EOPNOTSUPP;
4847 if (f.file->f_op != &io_uring_fops)
4848 goto out_fput;
4849
4850 ret = -ENXIO;
4851 ctx = f.file->private_data;
4852 if (!percpu_ref_tryget(&ctx->refs))
4853 goto out_fput;
4854
Jens Axboe6c271ce2019-01-10 11:22:30 -07004855 /*
4856 * For SQ polling, the thread will do all submissions and completions.
4857 * Just return the requested submit count, and wake the thread if
4858 * we were asked to.
4859 */
Jens Axboeb2a9ead2019-09-12 14:19:16 -06004860 ret = 0;
Jens Axboe6c271ce2019-01-10 11:22:30 -07004861 if (ctx->flags & IORING_SETUP_SQPOLL) {
Jens Axboec1edbf52019-11-10 16:56:04 -07004862 if (!list_empty_careful(&ctx->cq_overflow_list))
4863 io_cqring_overflow_flush(ctx, false);
Jens Axboe6c271ce2019-01-10 11:22:30 -07004864 if (flags & IORING_ENTER_SQ_WAKEUP)
4865 wake_up(&ctx->sqo_wait);
4866 submitted = to_submit;
Jens Axboeb2a9ead2019-09-12 14:19:16 -06004867 } else if (to_submit) {
Pavel Begunkovae9428c2019-11-06 00:22:14 +03004868 struct mm_struct *cur_mm;
Jens Axboe2b188cc2019-01-07 10:46:33 -07004869
Pavel Begunkovae9428c2019-11-06 00:22:14 +03004870 to_submit = min(to_submit, ctx->sq_entries);
Jens Axboe2b188cc2019-01-07 10:46:33 -07004871 mutex_lock(&ctx->uring_lock);
Pavel Begunkovae9428c2019-11-06 00:22:14 +03004872 /* already have mm, so io_submit_sqes() won't try to grab it */
4873 cur_mm = ctx->sqo_mm;
4874 submitted = io_submit_sqes(ctx, to_submit, f.file, fd,
4875 &cur_mm, false);
Jens Axboe2b188cc2019-01-07 10:46:33 -07004876 mutex_unlock(&ctx->uring_lock);
Jens Axboe2b188cc2019-01-07 10:46:33 -07004877 }
4878 if (flags & IORING_ENTER_GETEVENTS) {
Jens Axboedef596e2019-01-09 08:59:42 -07004879 unsigned nr_events = 0;
4880
Jens Axboe2b188cc2019-01-07 10:46:33 -07004881 min_complete = min(min_complete, ctx->cq_entries);
4882
Jens Axboedef596e2019-01-09 08:59:42 -07004883 if (ctx->flags & IORING_SETUP_IOPOLL) {
Jens Axboedef596e2019-01-09 08:59:42 -07004884 ret = io_iopoll_check(ctx, &nr_events, min_complete);
Jens Axboedef596e2019-01-09 08:59:42 -07004885 } else {
4886 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
4887 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07004888 }
4889
Pavel Begunkov6805b322019-10-08 02:18:42 +03004890 percpu_ref_put(&ctx->refs);
Jens Axboe2b188cc2019-01-07 10:46:33 -07004891out_fput:
4892 fdput(f);
4893 return submitted ? submitted : ret;
4894}
4895
4896static const struct file_operations io_uring_fops = {
4897 .release = io_uring_release,
Jens Axboefcb323c2019-10-24 12:39:47 -06004898 .flush = io_uring_flush,
Jens Axboe2b188cc2019-01-07 10:46:33 -07004899 .mmap = io_uring_mmap,
Roman Penyaev6c5c2402019-11-28 12:53:22 +01004900#ifndef CONFIG_MMU
4901 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
4902 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
4903#endif
Jens Axboe2b188cc2019-01-07 10:46:33 -07004904 .poll = io_uring_poll,
4905 .fasync = io_uring_fasync,
4906};
4907
4908static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
4909 struct io_uring_params *p)
4910{
Hristo Venev75b28af2019-08-26 17:23:46 +00004911 struct io_rings *rings;
4912 size_t size, sq_array_offset;
Jens Axboe2b188cc2019-01-07 10:46:33 -07004913
Hristo Venev75b28af2019-08-26 17:23:46 +00004914 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
4915 if (size == SIZE_MAX)
4916 return -EOVERFLOW;
4917
4918 rings = io_mem_alloc(size);
4919 if (!rings)
Jens Axboe2b188cc2019-01-07 10:46:33 -07004920 return -ENOMEM;
4921
Hristo Venev75b28af2019-08-26 17:23:46 +00004922 ctx->rings = rings;
4923 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
4924 rings->sq_ring_mask = p->sq_entries - 1;
4925 rings->cq_ring_mask = p->cq_entries - 1;
4926 rings->sq_ring_entries = p->sq_entries;
4927 rings->cq_ring_entries = p->cq_entries;
4928 ctx->sq_mask = rings->sq_ring_mask;
4929 ctx->cq_mask = rings->cq_ring_mask;
4930 ctx->sq_entries = rings->sq_ring_entries;
4931 ctx->cq_entries = rings->cq_ring_entries;
Jens Axboe2b188cc2019-01-07 10:46:33 -07004932
4933 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
Jens Axboeeb065d32019-11-20 09:26:29 -07004934 if (size == SIZE_MAX) {
4935 io_mem_free(ctx->rings);
4936 ctx->rings = NULL;
Jens Axboe2b188cc2019-01-07 10:46:33 -07004937 return -EOVERFLOW;
Jens Axboeeb065d32019-11-20 09:26:29 -07004938 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07004939
4940 ctx->sq_sqes = io_mem_alloc(size);
Jens Axboeeb065d32019-11-20 09:26:29 -07004941 if (!ctx->sq_sqes) {
4942 io_mem_free(ctx->rings);
4943 ctx->rings = NULL;
Jens Axboe2b188cc2019-01-07 10:46:33 -07004944 return -ENOMEM;
Jens Axboeeb065d32019-11-20 09:26:29 -07004945 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07004946
Jens Axboe2b188cc2019-01-07 10:46:33 -07004947 return 0;
4948}
4949
4950/*
4951 * Allocate an anonymous fd, this is what constitutes the application
4952 * visible backing of an io_uring instance. The application mmaps this
4953 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
4954 * we have to tie this fd to a socket for file garbage collection purposes.
4955 */
4956static int io_uring_get_fd(struct io_ring_ctx *ctx)
4957{
4958 struct file *file;
4959 int ret;
4960
4961#if defined(CONFIG_UNIX)
4962 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
4963 &ctx->ring_sock);
4964 if (ret)
4965 return ret;
4966#endif
4967
4968 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
4969 if (ret < 0)
4970 goto err;
4971
4972 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
4973 O_RDWR | O_CLOEXEC);
4974 if (IS_ERR(file)) {
4975 put_unused_fd(ret);
4976 ret = PTR_ERR(file);
4977 goto err;
4978 }
4979
4980#if defined(CONFIG_UNIX)
4981 ctx->ring_sock->file = file;
Jens Axboe6b063142019-01-10 22:13:58 -07004982 ctx->ring_sock->sk->sk_user_data = ctx;
Jens Axboe2b188cc2019-01-07 10:46:33 -07004983#endif
4984 fd_install(ret, file);
4985 return ret;
4986err:
4987#if defined(CONFIG_UNIX)
4988 sock_release(ctx->ring_sock);
4989 ctx->ring_sock = NULL;
4990#endif
4991 return ret;
4992}
4993
4994static int io_uring_create(unsigned entries, struct io_uring_params *p)
4995{
4996 struct user_struct *user = NULL;
4997 struct io_ring_ctx *ctx;
4998 bool account_mem;
4999 int ret;
5000
5001 if (!entries || entries > IORING_MAX_ENTRIES)
5002 return -EINVAL;
5003
5004 /*
5005 * Use twice as many entries for the CQ ring. It's possible for the
5006 * application to drive a higher depth than the size of the SQ ring,
5007 * since the sqes are only used at submission time. This allows for
Jens Axboe33a107f2019-10-04 12:10:03 -06005008 * some flexibility in overcommitting a bit. If the application has
5009 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
5010 * of CQ ring entries manually.
Jens Axboe2b188cc2019-01-07 10:46:33 -07005011 */
5012 p->sq_entries = roundup_pow_of_two(entries);
Jens Axboe33a107f2019-10-04 12:10:03 -06005013 if (p->flags & IORING_SETUP_CQSIZE) {
5014 /*
5015 * If IORING_SETUP_CQSIZE is set, we do the same roundup
5016 * to a power-of-two, if it isn't already. We do NOT impose
5017 * any cq vs sq ring sizing.
5018 */
5019 if (p->cq_entries < p->sq_entries || p->cq_entries > IORING_MAX_CQ_ENTRIES)
5020 return -EINVAL;
5021 p->cq_entries = roundup_pow_of_two(p->cq_entries);
5022 } else {
5023 p->cq_entries = 2 * p->sq_entries;
5024 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07005025
5026 user = get_uid(current_user());
5027 account_mem = !capable(CAP_IPC_LOCK);
5028
5029 if (account_mem) {
5030 ret = io_account_mem(user,
5031 ring_pages(p->sq_entries, p->cq_entries));
5032 if (ret) {
5033 free_uid(user);
5034 return ret;
5035 }
5036 }
5037
5038 ctx = io_ring_ctx_alloc(p);
5039 if (!ctx) {
5040 if (account_mem)
5041 io_unaccount_mem(user, ring_pages(p->sq_entries,
5042 p->cq_entries));
5043 free_uid(user);
5044 return -ENOMEM;
5045 }
5046 ctx->compat = in_compat_syscall();
5047 ctx->account_mem = account_mem;
5048 ctx->user = user;
Jens Axboe0b8c0ec2019-12-02 08:50:00 -07005049 ctx->creds = get_current_cred();
Jens Axboe2b188cc2019-01-07 10:46:33 -07005050
5051 ret = io_allocate_scq_urings(ctx, p);
5052 if (ret)
5053 goto err;
5054
Jens Axboe6c271ce2019-01-10 11:22:30 -07005055 ret = io_sq_offload_start(ctx, p);
Jens Axboe2b188cc2019-01-07 10:46:33 -07005056 if (ret)
5057 goto err;
5058
Jens Axboe2b188cc2019-01-07 10:46:33 -07005059 memset(&p->sq_off, 0, sizeof(p->sq_off));
Hristo Venev75b28af2019-08-26 17:23:46 +00005060 p->sq_off.head = offsetof(struct io_rings, sq.head);
5061 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
5062 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
5063 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
5064 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
5065 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
5066 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
Jens Axboe2b188cc2019-01-07 10:46:33 -07005067
5068 memset(&p->cq_off, 0, sizeof(p->cq_off));
Hristo Venev75b28af2019-08-26 17:23:46 +00005069 p->cq_off.head = offsetof(struct io_rings, cq.head);
5070 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
5071 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
5072 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
5073 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
5074 p->cq_off.cqes = offsetof(struct io_rings, cqes);
Jens Axboeac90f242019-09-06 10:26:21 -06005075
Jens Axboe044c1ab2019-10-28 09:15:33 -06005076 /*
5077 * Install ring fd as the very last thing, so we don't risk someone
5078 * having closed it before we finish setup
5079 */
5080 ret = io_uring_get_fd(ctx);
5081 if (ret < 0)
5082 goto err;
5083
Jens Axboeda8c9692019-12-02 18:51:26 -07005084 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
5085 IORING_FEAT_SUBMIT_STABLE;
Dmitrii Dolgovc826bd72019-10-15 19:02:01 +02005086 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
Jens Axboe2b188cc2019-01-07 10:46:33 -07005087 return ret;
5088err:
5089 io_ring_ctx_wait_and_kill(ctx);
5090 return ret;
5091}
5092
5093/*
5094 * Sets up an aio uring context, and returns the fd. Applications asks for a
5095 * ring size, we return the actual sq/cq ring sizes (among other things) in the
5096 * params structure passed in.
5097 */
5098static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
5099{
5100 struct io_uring_params p;
5101 long ret;
5102 int i;
5103
5104 if (copy_from_user(&p, params, sizeof(p)))
5105 return -EFAULT;
5106 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
5107 if (p.resv[i])
5108 return -EINVAL;
5109 }
5110
Jens Axboe6c271ce2019-01-10 11:22:30 -07005111 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
Jens Axboe33a107f2019-10-04 12:10:03 -06005112 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE))
Jens Axboe2b188cc2019-01-07 10:46:33 -07005113 return -EINVAL;
5114
5115 ret = io_uring_create(entries, &p);
5116 if (ret < 0)
5117 return ret;
5118
5119 if (copy_to_user(params, &p, sizeof(p)))
5120 return -EFAULT;
5121
5122 return ret;
5123}
5124
5125SYSCALL_DEFINE2(io_uring_setup, u32, entries,
5126 struct io_uring_params __user *, params)
5127{
5128 return io_uring_setup(entries, params);
5129}
5130
Jens Axboeedafcce2019-01-09 09:16:05 -07005131static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
5132 void __user *arg, unsigned nr_args)
Jens Axboeb19062a2019-04-15 10:49:38 -06005133 __releases(ctx->uring_lock)
5134 __acquires(ctx->uring_lock)
Jens Axboeedafcce2019-01-09 09:16:05 -07005135{
5136 int ret;
5137
Jens Axboe35fa71a2019-04-22 10:23:23 -06005138 /*
5139 * We're inside the ring mutex, if the ref is already dying, then
5140 * someone else killed the ctx or is already going through
5141 * io_uring_register().
5142 */
5143 if (percpu_ref_is_dying(&ctx->refs))
5144 return -ENXIO;
5145
Jens Axboeedafcce2019-01-09 09:16:05 -07005146 percpu_ref_kill(&ctx->refs);
Jens Axboeb19062a2019-04-15 10:49:38 -06005147
5148 /*
5149 * Drop uring mutex before waiting for references to exit. If another
5150 * thread is currently inside io_uring_enter() it might need to grab
5151 * the uring_lock to make progress. If we hold it here across the drain
5152 * wait, then we can deadlock. It's safe to drop the mutex here, since
5153 * no new references will come in after we've killed the percpu ref.
5154 */
5155 mutex_unlock(&ctx->uring_lock);
Jens Axboe206aefd2019-11-07 18:27:42 -07005156 wait_for_completion(&ctx->completions[0]);
Jens Axboeb19062a2019-04-15 10:49:38 -06005157 mutex_lock(&ctx->uring_lock);
Jens Axboeedafcce2019-01-09 09:16:05 -07005158
5159 switch (opcode) {
5160 case IORING_REGISTER_BUFFERS:
5161 ret = io_sqe_buffer_register(ctx, arg, nr_args);
5162 break;
5163 case IORING_UNREGISTER_BUFFERS:
5164 ret = -EINVAL;
5165 if (arg || nr_args)
5166 break;
5167 ret = io_sqe_buffer_unregister(ctx);
5168 break;
Jens Axboe6b063142019-01-10 22:13:58 -07005169 case IORING_REGISTER_FILES:
5170 ret = io_sqe_files_register(ctx, arg, nr_args);
5171 break;
5172 case IORING_UNREGISTER_FILES:
5173 ret = -EINVAL;
5174 if (arg || nr_args)
5175 break;
5176 ret = io_sqe_files_unregister(ctx);
5177 break;
Jens Axboec3a31e62019-10-03 13:59:56 -06005178 case IORING_REGISTER_FILES_UPDATE:
5179 ret = io_sqe_files_update(ctx, arg, nr_args);
5180 break;
Jens Axboe9b402842019-04-11 11:45:41 -06005181 case IORING_REGISTER_EVENTFD:
5182 ret = -EINVAL;
5183 if (nr_args != 1)
5184 break;
5185 ret = io_eventfd_register(ctx, arg);
5186 break;
5187 case IORING_UNREGISTER_EVENTFD:
5188 ret = -EINVAL;
5189 if (arg || nr_args)
5190 break;
5191 ret = io_eventfd_unregister(ctx);
5192 break;
Jens Axboeedafcce2019-01-09 09:16:05 -07005193 default:
5194 ret = -EINVAL;
5195 break;
5196 }
5197
5198 /* bring the ctx back to life */
Jens Axboe206aefd2019-11-07 18:27:42 -07005199 reinit_completion(&ctx->completions[0]);
Jens Axboeedafcce2019-01-09 09:16:05 -07005200 percpu_ref_reinit(&ctx->refs);
5201 return ret;
5202}
5203
5204SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
5205 void __user *, arg, unsigned int, nr_args)
5206{
5207 struct io_ring_ctx *ctx;
5208 long ret = -EBADF;
5209 struct fd f;
5210
5211 f = fdget(fd);
5212 if (!f.file)
5213 return -EBADF;
5214
5215 ret = -EOPNOTSUPP;
5216 if (f.file->f_op != &io_uring_fops)
5217 goto out_fput;
5218
5219 ctx = f.file->private_data;
5220
5221 mutex_lock(&ctx->uring_lock);
5222 ret = __io_uring_register(ctx, opcode, arg, nr_args);
5223 mutex_unlock(&ctx->uring_lock);
Dmitrii Dolgovc826bd72019-10-15 19:02:01 +02005224 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
5225 ctx->cq_ev_fd != NULL, ret);
Jens Axboeedafcce2019-01-09 09:16:05 -07005226out_fput:
5227 fdput(f);
5228 return ret;
5229}
5230
Jens Axboe2b188cc2019-01-07 10:46:33 -07005231static int __init io_uring_init(void)
5232{
5233 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
5234 return 0;
5235};
5236__initcall(io_uring_init);