blob: 6524898831e05cf163ef8d620716c42a3b806b8b [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 Axboe2b188cc2019-01-07 10:46:33 -070072
Dmitrii Dolgovc826bd72019-10-15 19:02:01 +020073#define CREATE_TRACE_POINTS
74#include <trace/events/io_uring.h>
75
Jens Axboe2b188cc2019-01-07 10:46:33 -070076#include <uapi/linux/io_uring.h>
77
78#include "internal.h"
Jens Axboe561fb042019-10-24 07:25:42 -060079#include "io-wq.h"
Jens Axboe2b188cc2019-01-07 10:46:33 -070080
Daniel Xu5277dea2019-09-14 14:23:45 -070081#define IORING_MAX_ENTRIES 32768
Jens Axboe33a107f2019-10-04 12:10:03 -060082#define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
Jens Axboe65e19f52019-10-26 07:20:21 -060083
84/*
85 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
86 */
87#define IORING_FILE_TABLE_SHIFT 9
88#define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
89#define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
90#define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
Jens Axboe2b188cc2019-01-07 10:46:33 -070091
92struct io_uring {
93 u32 head ____cacheline_aligned_in_smp;
94 u32 tail ____cacheline_aligned_in_smp;
95};
96
Stefan Bühler1e84b972019-04-24 23:54:16 +020097/*
Hristo Venev75b28af2019-08-26 17:23:46 +000098 * This data is shared with the application through the mmap at offsets
99 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
Stefan Bühler1e84b972019-04-24 23:54:16 +0200100 *
101 * The offsets to the member fields are published through struct
102 * io_sqring_offsets when calling io_uring_setup.
103 */
Hristo Venev75b28af2019-08-26 17:23:46 +0000104struct io_rings {
Stefan Bühler1e84b972019-04-24 23:54:16 +0200105 /*
106 * Head and tail offsets into the ring; the offsets need to be
107 * masked to get valid indices.
108 *
Hristo Venev75b28af2019-08-26 17:23:46 +0000109 * The kernel controls head of the sq ring and the tail of the cq ring,
110 * and the application controls tail of the sq ring and the head of the
111 * cq ring.
Stefan Bühler1e84b972019-04-24 23:54:16 +0200112 */
Hristo Venev75b28af2019-08-26 17:23:46 +0000113 struct io_uring sq, cq;
Stefan Bühler1e84b972019-04-24 23:54:16 +0200114 /*
Hristo Venev75b28af2019-08-26 17:23:46 +0000115 * Bitmasks to apply to head and tail offsets (constant, equals
Stefan Bühler1e84b972019-04-24 23:54:16 +0200116 * ring_entries - 1)
117 */
Hristo Venev75b28af2019-08-26 17:23:46 +0000118 u32 sq_ring_mask, cq_ring_mask;
119 /* Ring sizes (constant, power of 2) */
120 u32 sq_ring_entries, cq_ring_entries;
Stefan Bühler1e84b972019-04-24 23:54:16 +0200121 /*
122 * Number of invalid entries dropped by the kernel due to
123 * invalid index stored in array
124 *
125 * Written by the kernel, shouldn't be modified by the
126 * application (i.e. get number of "new events" by comparing to
127 * cached value).
128 *
129 * After a new SQ head value was read by the application this
130 * counter includes all submissions that were dropped reaching
131 * the new SQ head (and possibly more).
132 */
Hristo Venev75b28af2019-08-26 17:23:46 +0000133 u32 sq_dropped;
Stefan Bühler1e84b972019-04-24 23:54:16 +0200134 /*
135 * Runtime flags
136 *
137 * Written by the kernel, shouldn't be modified by the
138 * application.
139 *
140 * The application needs a full memory barrier before checking
141 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
142 */
Hristo Venev75b28af2019-08-26 17:23:46 +0000143 u32 sq_flags;
Stefan Bühler1e84b972019-04-24 23:54:16 +0200144 /*
145 * Number of completion events lost because the queue was full;
146 * this should be avoided by the application by making sure
147 * there are not more requests pending thatn there is space in
148 * the completion queue.
149 *
150 * Written by the kernel, shouldn't be modified by the
151 * application (i.e. get number of "new events" by comparing to
152 * cached value).
153 *
154 * As completion events come in out of order this counter is not
155 * ordered with any other data.
156 */
Hristo Venev75b28af2019-08-26 17:23:46 +0000157 u32 cq_overflow;
Stefan Bühler1e84b972019-04-24 23:54:16 +0200158 /*
159 * Ring buffer of completion events.
160 *
161 * The kernel writes completion events fresh every time they are
162 * produced, so the application is allowed to modify pending
163 * entries.
164 */
Hristo Venev75b28af2019-08-26 17:23:46 +0000165 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700166};
167
Jens Axboeedafcce2019-01-09 09:16:05 -0700168struct io_mapped_ubuf {
169 u64 ubuf;
170 size_t len;
171 struct bio_vec *bvec;
172 unsigned int nr_bvecs;
173};
174
Jens Axboe65e19f52019-10-26 07:20:21 -0600175struct fixed_file_table {
176 struct file **files;
177};
178
Jens Axboe2b188cc2019-01-07 10:46:33 -0700179struct io_ring_ctx {
180 struct {
181 struct percpu_ref refs;
182 } ____cacheline_aligned_in_smp;
183
184 struct {
185 unsigned int flags;
186 bool compat;
187 bool account_mem;
188
Hristo Venev75b28af2019-08-26 17:23:46 +0000189 /*
190 * Ring buffer of indices into array of io_uring_sqe, which is
191 * mmapped by the application using the IORING_OFF_SQES offset.
192 *
193 * This indirection could e.g. be used to assign fixed
194 * io_uring_sqe entries to operations and only submit them to
195 * the queue when needed.
196 *
197 * The kernel modifies neither the indices array nor the entries
198 * array.
199 */
200 u32 *sq_array;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700201 unsigned cached_sq_head;
202 unsigned sq_entries;
203 unsigned sq_mask;
Jens Axboe6c271ce2019-01-10 11:22:30 -0700204 unsigned sq_thread_idle;
Jens Axboe498ccd92019-10-25 10:04:25 -0600205 unsigned cached_sq_dropped;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700206 struct io_uring_sqe *sq_sqes;
Jens Axboede0617e2019-04-06 21:51:27 -0600207
208 struct list_head defer_list;
Jens Axboe5262f562019-09-17 12:26:57 -0600209 struct list_head timeout_list;
Jens Axboefcb323c2019-10-24 12:39:47 -0600210
211 wait_queue_head_t inflight_wait;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700212 } ____cacheline_aligned_in_smp;
213
214 /* IO offload */
Jens Axboe561fb042019-10-24 07:25:42 -0600215 struct io_wq *io_wq;
Jens Axboe6c271ce2019-01-10 11:22:30 -0700216 struct task_struct *sqo_thread; /* if using sq thread polling */
Jens Axboe2b188cc2019-01-07 10:46:33 -0700217 struct mm_struct *sqo_mm;
Jens Axboe6c271ce2019-01-10 11:22:30 -0700218 wait_queue_head_t sqo_wait;
Jackie Liua4c0b3d2019-07-08 13:41:12 +0800219 struct completion sqo_thread_started;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700220
221 struct {
Jens Axboe2b188cc2019-01-07 10:46:33 -0700222 unsigned cached_cq_tail;
Jens Axboe498ccd92019-10-25 10:04:25 -0600223 atomic_t cached_cq_overflow;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700224 unsigned cq_entries;
225 unsigned cq_mask;
226 struct wait_queue_head cq_wait;
227 struct fasync_struct *cq_fasync;
Jens Axboe9b402842019-04-11 11:45:41 -0600228 struct eventfd_ctx *cq_ev_fd;
Jens Axboe5262f562019-09-17 12:26:57 -0600229 atomic_t cq_timeouts;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700230 } ____cacheline_aligned_in_smp;
231
Hristo Venev75b28af2019-08-26 17:23:46 +0000232 struct io_rings *rings;
233
Jens Axboe6b063142019-01-10 22:13:58 -0700234 /*
235 * If used, fixed file set. Writers must ensure that ->refs is dead,
236 * readers must ensure that ->refs is alive as long as the file* is
237 * used. Only updated through io_uring_register(2).
238 */
Jens Axboe65e19f52019-10-26 07:20:21 -0600239 struct fixed_file_table *file_table;
Jens Axboe6b063142019-01-10 22:13:58 -0700240 unsigned nr_user_files;
241
Jens Axboeedafcce2019-01-09 09:16:05 -0700242 /* if used, fixed mapped user buffers */
243 unsigned nr_user_bufs;
244 struct io_mapped_ubuf *user_bufs;
245
Jens Axboe2b188cc2019-01-07 10:46:33 -0700246 struct user_struct *user;
247
248 struct completion ctx_done;
249
250 struct {
251 struct mutex uring_lock;
252 wait_queue_head_t wait;
253 } ____cacheline_aligned_in_smp;
254
255 struct {
256 spinlock_t completion_lock;
Jens Axboedef596e2019-01-09 08:59:42 -0700257 bool poll_multi_file;
258 /*
259 * ->poll_list is protected by the ctx->uring_lock for
260 * io_uring instances that don't use IORING_SETUP_SQPOLL.
261 * For SQPOLL, only the single threaded io_sq_thread() will
262 * manipulate the list, hence no extra locking is needed there.
263 */
264 struct list_head poll_list;
Jens Axboe221c5eb2019-01-17 09:41:58 -0700265 struct list_head cancel_list;
Jens Axboefcb323c2019-10-24 12:39:47 -0600266
267 spinlock_t inflight_lock;
268 struct list_head inflight_list;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700269 } ____cacheline_aligned_in_smp;
270
271#if defined(CONFIG_UNIX)
272 struct socket *ring_sock;
273#endif
274};
275
276struct sqe_submit {
277 const struct io_uring_sqe *sqe;
Jens Axboefcb323c2019-10-24 12:39:47 -0600278 struct file *ring_file;
279 int ring_fd;
Jackie Liu8776f3f2019-09-09 20:50:39 +0800280 u32 sequence;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700281 bool has_user;
Jackie Liuba5290c2019-10-09 09:19:59 +0800282 bool in_async;
Jens Axboe6c271ce2019-01-10 11:22:30 -0700283 bool needs_fixed_file;
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;
296 struct wait_queue_entry wait;
297};
298
Jens Axboe5262f562019-09-17 12:26:57 -0600299struct io_timeout {
300 struct file *file;
301 struct hrtimer timer;
302};
303
Jens Axboe09bb8392019-03-13 12:39:28 -0600304/*
305 * NOTE! Each of the iocb union members has the file pointer
306 * as the first entry in their struct definition. So you can
307 * access the file pointer through any of the sub-structs,
308 * or directly as just 'ki_filp' in this struct.
309 */
Jens Axboe2b188cc2019-01-07 10:46:33 -0700310struct io_kiocb {
Jens Axboe221c5eb2019-01-17 09:41:58 -0700311 union {
Jens Axboe09bb8392019-03-13 12:39:28 -0600312 struct file *file;
Jens Axboe221c5eb2019-01-17 09:41:58 -0700313 struct kiocb rw;
314 struct io_poll_iocb poll;
Jens Axboe5262f562019-09-17 12:26:57 -0600315 struct io_timeout timeout;
Jens Axboe221c5eb2019-01-17 09:41:58 -0700316 };
Jens Axboe2b188cc2019-01-07 10:46:33 -0700317
318 struct sqe_submit submit;
319
320 struct io_ring_ctx *ctx;
321 struct list_head list;
Jens Axboe9e645e112019-05-10 16:07:28 -0600322 struct list_head link_list;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700323 unsigned int flags;
Jens Axboec16361c2019-01-17 08:39:48 -0700324 refcount_t refs;
Stefan Bühler8449eed2019-04-27 20:34:19 +0200325#define REQ_F_NOWAIT 1 /* must not punt to workers */
Jens Axboedef596e2019-01-09 08:59:42 -0700326#define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */
Jens Axboe6b063142019-01-10 22:13:58 -0700327#define REQ_F_FIXED_FILE 4 /* ctx owns file */
Jens Axboe31b51512019-01-18 22:56:34 -0700328#define REQ_F_SEQ_PREV 8 /* sequential with previous */
Stefan Bühlere2033e32019-05-11 19:08:01 +0200329#define REQ_F_IO_DRAIN 16 /* drain existing IO first */
330#define REQ_F_IO_DRAINED 32 /* drain done */
Jens Axboe9e645e112019-05-10 16:07:28 -0600331#define REQ_F_LINK 64 /* linked sqes */
Zhengyuan Liuf7b76ac2019-07-16 23:26:14 +0800332#define REQ_F_FAIL_LINK 256 /* fail rest of links */
Jackie Liu4fe2c962019-09-09 20:50:40 +0800333#define REQ_F_SHADOW_DRAIN 512 /* link-drain shadow req */
Jens Axboe5262f562019-09-17 12:26:57 -0600334#define REQ_F_TIMEOUT 1024 /* timeout request */
Jens Axboe491381ce2019-10-17 09:20:46 -0600335#define REQ_F_ISREG 2048 /* regular file */
336#define REQ_F_MUST_PUNT 4096 /* must be punted even for NONBLOCK */
Jens Axboefcb323c2019-10-24 12:39:47 -0600337#define REQ_F_INFLIGHT 8192 /* on inflight list */
Jens Axboe2b188cc2019-01-07 10:46:33 -0700338 u64 user_data;
Jens Axboe9e645e112019-05-10 16:07:28 -0600339 u32 result;
Jens Axboede0617e2019-04-06 21:51:27 -0600340 u32 sequence;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700341
Jens Axboefcb323c2019-10-24 12:39:47 -0600342 struct list_head inflight_entry;
343
Jens Axboe561fb042019-10-24 07:25:42 -0600344 struct io_wq_work work;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700345};
346
347#define IO_PLUG_THRESHOLD 2
Jens Axboedef596e2019-01-09 08:59:42 -0700348#define IO_IOPOLL_BATCH 8
Jens Axboe2b188cc2019-01-07 10:46:33 -0700349
Jens Axboe9a56a232019-01-09 09:06:50 -0700350struct io_submit_state {
351 struct blk_plug plug;
352
353 /*
Jens Axboe2579f912019-01-09 09:10:43 -0700354 * io_kiocb alloc cache
355 */
356 void *reqs[IO_IOPOLL_BATCH];
357 unsigned int free_reqs;
358 unsigned int cur_req;
359
360 /*
Jens Axboe9a56a232019-01-09 09:06:50 -0700361 * File reference cache
362 */
363 struct file *file;
364 unsigned int fd;
365 unsigned int has_refs;
366 unsigned int used_refs;
367 unsigned int ios_left;
368};
369
Jens Axboe561fb042019-10-24 07:25:42 -0600370static void io_wq_submit_work(struct io_wq_work **workptr);
Jens Axboe5262f562019-09-17 12:26:57 -0600371static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data,
372 long res);
Jackie Liu4fe2c962019-09-09 20:50:40 +0800373static void __io_free_req(struct io_kiocb *req);
Jens Axboede0617e2019-04-06 21:51:27 -0600374
Jens Axboe2b188cc2019-01-07 10:46:33 -0700375static struct kmem_cache *req_cachep;
376
377static const struct file_operations io_uring_fops;
378
379struct sock *io_uring_get_socket(struct file *file)
380{
381#if defined(CONFIG_UNIX)
382 if (file->f_op == &io_uring_fops) {
383 struct io_ring_ctx *ctx = file->private_data;
384
385 return ctx->ring_sock->sk;
386 }
387#endif
388 return NULL;
389}
390EXPORT_SYMBOL(io_uring_get_socket);
391
392static void io_ring_ctx_ref_free(struct percpu_ref *ref)
393{
394 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
395
396 complete(&ctx->ctx_done);
397}
398
399static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
400{
401 struct io_ring_ctx *ctx;
402
403 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
404 if (!ctx)
405 return NULL;
406
Roman Gushchin21482892019-05-07 10:01:48 -0700407 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
408 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
Jens Axboe2b188cc2019-01-07 10:46:33 -0700409 kfree(ctx);
410 return NULL;
411 }
412
413 ctx->flags = p->flags;
414 init_waitqueue_head(&ctx->cq_wait);
415 init_completion(&ctx->ctx_done);
Jackie Liua4c0b3d2019-07-08 13:41:12 +0800416 init_completion(&ctx->sqo_thread_started);
Jens Axboe2b188cc2019-01-07 10:46:33 -0700417 mutex_init(&ctx->uring_lock);
418 init_waitqueue_head(&ctx->wait);
419 spin_lock_init(&ctx->completion_lock);
Jens Axboedef596e2019-01-09 08:59:42 -0700420 INIT_LIST_HEAD(&ctx->poll_list);
Jens Axboe221c5eb2019-01-17 09:41:58 -0700421 INIT_LIST_HEAD(&ctx->cancel_list);
Jens Axboede0617e2019-04-06 21:51:27 -0600422 INIT_LIST_HEAD(&ctx->defer_list);
Jens Axboe5262f562019-09-17 12:26:57 -0600423 INIT_LIST_HEAD(&ctx->timeout_list);
Jens Axboefcb323c2019-10-24 12:39:47 -0600424 init_waitqueue_head(&ctx->inflight_wait);
425 spin_lock_init(&ctx->inflight_lock);
426 INIT_LIST_HEAD(&ctx->inflight_list);
Jens Axboe2b188cc2019-01-07 10:46:33 -0700427 return ctx;
428}
429
Jens Axboe7adf4ea2019-10-10 21:42:58 -0600430static inline bool __io_sequence_defer(struct io_ring_ctx *ctx,
431 struct io_kiocb *req)
Jens Axboede0617e2019-04-06 21:51:27 -0600432{
Jens Axboe498ccd92019-10-25 10:04:25 -0600433 return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped
434 + atomic_read(&ctx->cached_cq_overflow);
Jens Axboede0617e2019-04-06 21:51:27 -0600435}
436
Jens Axboe7adf4ea2019-10-10 21:42:58 -0600437static inline bool io_sequence_defer(struct io_ring_ctx *ctx,
438 struct io_kiocb *req)
439{
440 if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) != REQ_F_IO_DRAIN)
441 return false;
442
443 return __io_sequence_defer(ctx, req);
444}
445
446static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
Jens Axboede0617e2019-04-06 21:51:27 -0600447{
448 struct io_kiocb *req;
449
Jens Axboe7adf4ea2019-10-10 21:42:58 -0600450 req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list);
451 if (req && !io_sequence_defer(ctx, req)) {
Jens Axboede0617e2019-04-06 21:51:27 -0600452 list_del_init(&req->list);
453 return req;
454 }
455
456 return NULL;
457}
458
Jens Axboe5262f562019-09-17 12:26:57 -0600459static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx)
460{
Jens Axboe7adf4ea2019-10-10 21:42:58 -0600461 struct io_kiocb *req;
462
463 req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list);
464 if (req && !__io_sequence_defer(ctx, req)) {
465 list_del_init(&req->list);
466 return req;
467 }
468
469 return NULL;
Jens Axboe5262f562019-09-17 12:26:57 -0600470}
471
Jens Axboede0617e2019-04-06 21:51:27 -0600472static void __io_commit_cqring(struct io_ring_ctx *ctx)
Jens Axboe2b188cc2019-01-07 10:46:33 -0700473{
Hristo Venev75b28af2019-08-26 17:23:46 +0000474 struct io_rings *rings = ctx->rings;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700475
Hristo Venev75b28af2019-08-26 17:23:46 +0000476 if (ctx->cached_cq_tail != READ_ONCE(rings->cq.tail)) {
Jens Axboe2b188cc2019-01-07 10:46:33 -0700477 /* order cqe stores with ring update */
Hristo Venev75b28af2019-08-26 17:23:46 +0000478 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
Jens Axboe2b188cc2019-01-07 10:46:33 -0700479
Jens Axboe2b188cc2019-01-07 10:46:33 -0700480 if (wq_has_sleeper(&ctx->cq_wait)) {
481 wake_up_interruptible(&ctx->cq_wait);
482 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
483 }
484 }
485}
486
Jens Axboe561fb042019-10-24 07:25:42 -0600487static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe)
Jens Axboe18d9be12019-09-10 09:13:05 -0600488{
Jens Axboe561fb042019-10-24 07:25:42 -0600489 u8 opcode = READ_ONCE(sqe->opcode);
490
491 return !(opcode == IORING_OP_READ_FIXED ||
492 opcode == IORING_OP_WRITE_FIXED);
493}
494
495static inline bool io_prep_async_work(struct io_kiocb *req)
496{
497 bool do_hashed = false;
Jens Axboe54a91f32019-09-10 09:15:04 -0600498
Jens Axboe6cc47d12019-09-18 11:18:23 -0600499 if (req->submit.sqe) {
500 switch (req->submit.sqe->opcode) {
501 case IORING_OP_WRITEV:
502 case IORING_OP_WRITE_FIXED:
Jens Axboe561fb042019-10-24 07:25:42 -0600503 do_hashed = true;
Jens Axboe6cc47d12019-09-18 11:18:23 -0600504 break;
505 }
Jens Axboe561fb042019-10-24 07:25:42 -0600506 if (io_sqe_needs_user(req->submit.sqe))
507 req->work.flags |= IO_WQ_WORK_NEEDS_USER;
Jens Axboe54a91f32019-09-10 09:15:04 -0600508 }
509
Jens Axboe561fb042019-10-24 07:25:42 -0600510 return do_hashed;
511}
512
513static inline void io_queue_async_work(struct io_ring_ctx *ctx,
514 struct io_kiocb *req)
515{
516 bool do_hashed = io_prep_async_work(req);
517
518 trace_io_uring_queue_async_work(ctx, do_hashed, req, &req->work,
519 req->flags);
520 if (!do_hashed) {
521 io_wq_enqueue(ctx->io_wq, &req->work);
522 } else {
523 io_wq_enqueue_hashed(ctx->io_wq, &req->work,
524 file_inode(req->file));
525 }
Jens Axboe18d9be12019-09-10 09:13:05 -0600526}
527
Jens Axboe5262f562019-09-17 12:26:57 -0600528static void io_kill_timeout(struct io_kiocb *req)
529{
530 int ret;
531
532 ret = hrtimer_try_to_cancel(&req->timeout.timer);
533 if (ret != -1) {
534 atomic_inc(&req->ctx->cq_timeouts);
Jens Axboe842f9612019-10-29 12:34:10 -0600535 list_del_init(&req->list);
Jens Axboe5262f562019-09-17 12:26:57 -0600536 io_cqring_fill_event(req->ctx, req->user_data, 0);
537 __io_free_req(req);
538 }
539}
540
541static void io_kill_timeouts(struct io_ring_ctx *ctx)
542{
543 struct io_kiocb *req, *tmp;
544
545 spin_lock_irq(&ctx->completion_lock);
546 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
547 io_kill_timeout(req);
548 spin_unlock_irq(&ctx->completion_lock);
549}
550
Jens Axboede0617e2019-04-06 21:51:27 -0600551static void io_commit_cqring(struct io_ring_ctx *ctx)
552{
553 struct io_kiocb *req;
554
Jens Axboe5262f562019-09-17 12:26:57 -0600555 while ((req = io_get_timeout_req(ctx)) != NULL)
556 io_kill_timeout(req);
557
Jens Axboede0617e2019-04-06 21:51:27 -0600558 __io_commit_cqring(ctx);
559
560 while ((req = io_get_deferred_req(ctx)) != NULL) {
Jackie Liu4fe2c962019-09-09 20:50:40 +0800561 if (req->flags & REQ_F_SHADOW_DRAIN) {
562 /* Just for drain, free it. */
563 __io_free_req(req);
564 continue;
565 }
Jens Axboede0617e2019-04-06 21:51:27 -0600566 req->flags |= REQ_F_IO_DRAINED;
Jens Axboe18d9be12019-09-10 09:13:05 -0600567 io_queue_async_work(ctx, req);
Jens Axboede0617e2019-04-06 21:51:27 -0600568 }
569}
570
Jens Axboe2b188cc2019-01-07 10:46:33 -0700571static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
572{
Hristo Venev75b28af2019-08-26 17:23:46 +0000573 struct io_rings *rings = ctx->rings;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700574 unsigned tail;
575
576 tail = ctx->cached_cq_tail;
Stefan Bühler115e12e2019-04-24 23:54:18 +0200577 /*
578 * writes to the cq entry need to come after reading head; the
579 * control dependency is enough as we're using WRITE_ONCE to
580 * fill the cq entry
581 */
Hristo Venev75b28af2019-08-26 17:23:46 +0000582 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
Jens Axboe2b188cc2019-01-07 10:46:33 -0700583 return NULL;
584
585 ctx->cached_cq_tail++;
Hristo Venev75b28af2019-08-26 17:23:46 +0000586 return &rings->cqes[tail & ctx->cq_mask];
Jens Axboe2b188cc2019-01-07 10:46:33 -0700587}
588
589static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data,
Jens Axboec71ffb62019-05-13 20:58:29 -0600590 long res)
Jens Axboe2b188cc2019-01-07 10:46:33 -0700591{
592 struct io_uring_cqe *cqe;
593
Jens Axboe51c3ff62019-11-03 06:52:50 -0700594 trace_io_uring_complete(ctx, ki_user_data, res);
595
Jens Axboe2b188cc2019-01-07 10:46:33 -0700596 /*
597 * If we can't get a cq entry, userspace overflowed the
598 * submission (by quite a lot). Increment the overflow count in
599 * the ring.
600 */
601 cqe = io_get_cqring(ctx);
602 if (cqe) {
603 WRITE_ONCE(cqe->user_data, ki_user_data);
604 WRITE_ONCE(cqe->res, res);
Jens Axboec71ffb62019-05-13 20:58:29 -0600605 WRITE_ONCE(cqe->flags, 0);
Jens Axboe2b188cc2019-01-07 10:46:33 -0700606 } else {
Jens Axboe498ccd92019-10-25 10:04:25 -0600607 WRITE_ONCE(ctx->rings->cq_overflow,
608 atomic_inc_return(&ctx->cached_cq_overflow));
Jens Axboe2b188cc2019-01-07 10:46:33 -0700609 }
610}
611
Jens Axboe8c838782019-03-12 15:48:16 -0600612static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
613{
614 if (waitqueue_active(&ctx->wait))
615 wake_up(&ctx->wait);
616 if (waitqueue_active(&ctx->sqo_wait))
617 wake_up(&ctx->sqo_wait);
Jens Axboe9b402842019-04-11 11:45:41 -0600618 if (ctx->cq_ev_fd)
619 eventfd_signal(ctx->cq_ev_fd, 1);
Jens Axboe8c838782019-03-12 15:48:16 -0600620}
621
622static void io_cqring_add_event(struct io_ring_ctx *ctx, u64 user_data,
Jens Axboec71ffb62019-05-13 20:58:29 -0600623 long res)
Jens Axboe2b188cc2019-01-07 10:46:33 -0700624{
625 unsigned long flags;
626
627 spin_lock_irqsave(&ctx->completion_lock, flags);
Jens Axboec71ffb62019-05-13 20:58:29 -0600628 io_cqring_fill_event(ctx, user_data, res);
Jens Axboe2b188cc2019-01-07 10:46:33 -0700629 io_commit_cqring(ctx);
630 spin_unlock_irqrestore(&ctx->completion_lock, flags);
631
Jens Axboe8c838782019-03-12 15:48:16 -0600632 io_cqring_ev_posted(ctx);
Jens Axboe2b188cc2019-01-07 10:46:33 -0700633}
634
Jens Axboe2579f912019-01-09 09:10:43 -0700635static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
636 struct io_submit_state *state)
Jens Axboe2b188cc2019-01-07 10:46:33 -0700637{
Jens Axboefd6fab22019-03-14 16:30:06 -0600638 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700639 struct io_kiocb *req;
640
641 if (!percpu_ref_tryget(&ctx->refs))
642 return NULL;
643
Jens Axboe2579f912019-01-09 09:10:43 -0700644 if (!state) {
Jens Axboefd6fab22019-03-14 16:30:06 -0600645 req = kmem_cache_alloc(req_cachep, gfp);
Jens Axboe2579f912019-01-09 09:10:43 -0700646 if (unlikely(!req))
647 goto out;
648 } else if (!state->free_reqs) {
649 size_t sz;
650 int ret;
651
652 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
Jens Axboefd6fab22019-03-14 16:30:06 -0600653 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
654
655 /*
656 * Bulk alloc is all-or-nothing. If we fail to get a batch,
657 * retry single alloc to be on the safe side.
658 */
659 if (unlikely(ret <= 0)) {
660 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
661 if (!state->reqs[0])
662 goto out;
663 ret = 1;
664 }
Jens Axboe2579f912019-01-09 09:10:43 -0700665 state->free_reqs = ret - 1;
666 state->cur_req = 1;
667 req = state->reqs[0];
668 } else {
669 req = state->reqs[state->cur_req];
670 state->free_reqs--;
671 state->cur_req++;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700672 }
673
Jens Axboe60c112b2019-06-21 10:20:18 -0600674 req->file = NULL;
Jens Axboe2579f912019-01-09 09:10:43 -0700675 req->ctx = ctx;
676 req->flags = 0;
Jens Axboee65ef562019-03-12 10:16:44 -0600677 /* one is dropped after submission, the other at completion */
678 refcount_set(&req->refs, 2);
Jens Axboe9e645e112019-05-10 16:07:28 -0600679 req->result = 0;
Jens Axboe561fb042019-10-24 07:25:42 -0600680 INIT_IO_WORK(&req->work, io_wq_submit_work);
Jens Axboe2579f912019-01-09 09:10:43 -0700681 return req;
682out:
Pavel Begunkov6805b322019-10-08 02:18:42 +0300683 percpu_ref_put(&ctx->refs);
Jens Axboe2b188cc2019-01-07 10:46:33 -0700684 return NULL;
685}
686
Jens Axboedef596e2019-01-09 08:59:42 -0700687static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr)
688{
689 if (*nr) {
690 kmem_cache_free_bulk(req_cachep, *nr, reqs);
Pavel Begunkov6805b322019-10-08 02:18:42 +0300691 percpu_ref_put_many(&ctx->refs, *nr);
Jens Axboedef596e2019-01-09 08:59:42 -0700692 *nr = 0;
693 }
694}
695
Jens Axboe9e645e112019-05-10 16:07:28 -0600696static void __io_free_req(struct io_kiocb *req)
Jens Axboe2b188cc2019-01-07 10:46:33 -0700697{
Jens Axboefcb323c2019-10-24 12:39:47 -0600698 struct io_ring_ctx *ctx = req->ctx;
699
Jens Axboe09bb8392019-03-13 12:39:28 -0600700 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
701 fput(req->file);
Jens Axboefcb323c2019-10-24 12:39:47 -0600702 if (req->flags & REQ_F_INFLIGHT) {
703 unsigned long flags;
704
705 spin_lock_irqsave(&ctx->inflight_lock, flags);
706 list_del(&req->inflight_entry);
707 if (waitqueue_active(&ctx->inflight_wait))
708 wake_up(&ctx->inflight_wait);
709 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
710 }
711 percpu_ref_put(&ctx->refs);
Jens Axboee65ef562019-03-12 10:16:44 -0600712 kmem_cache_free(req_cachep, req);
713}
714
Jens Axboeba816ad2019-09-28 11:36:45 -0600715static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
Jens Axboe9e645e112019-05-10 16:07:28 -0600716{
717 struct io_kiocb *nxt;
718
719 /*
720 * The list should never be empty when we are called here. But could
721 * potentially happen if the chain is messed up, check to be on the
722 * safe side.
723 */
724 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list);
725 if (nxt) {
726 list_del(&nxt->list);
727 if (!list_empty(&req->link_list)) {
728 INIT_LIST_HEAD(&nxt->link_list);
729 list_splice(&req->link_list, &nxt->link_list);
730 nxt->flags |= REQ_F_LINK;
731 }
732
Jens Axboeba816ad2019-09-28 11:36:45 -0600733 /*
734 * If we're in async work, we can continue processing the chain
735 * in this context instead of having to queue up new async work.
736 */
Jens Axboe561fb042019-10-24 07:25:42 -0600737 if (nxtptr && current_work())
Jens Axboeba816ad2019-09-28 11:36:45 -0600738 *nxtptr = nxt;
Jens Axboe561fb042019-10-24 07:25:42 -0600739 else
Jens Axboeba816ad2019-09-28 11:36:45 -0600740 io_queue_async_work(req->ctx, nxt);
Jens Axboe9e645e112019-05-10 16:07:28 -0600741 }
742}
743
744/*
745 * Called if REQ_F_LINK is set, and we fail the head request
746 */
747static void io_fail_links(struct io_kiocb *req)
748{
749 struct io_kiocb *link;
750
751 while (!list_empty(&req->link_list)) {
752 link = list_first_entry(&req->link_list, struct io_kiocb, list);
753 list_del(&link->list);
754
Dmitrii Dolgovc826bd72019-10-15 19:02:01 +0200755 trace_io_uring_fail_link(req, link);
Jens Axboe9e645e112019-05-10 16:07:28 -0600756 io_cqring_add_event(req->ctx, link->user_data, -ECANCELED);
757 __io_free_req(link);
758 }
759}
760
Jens Axboeba816ad2019-09-28 11:36:45 -0600761static void io_free_req(struct io_kiocb *req, struct io_kiocb **nxt)
Jens Axboe9e645e112019-05-10 16:07:28 -0600762{
763 /*
764 * If LINK is set, we have dependent requests in this chain. If we
765 * didn't fail this request, queue the first one up, moving any other
766 * dependencies to the next request. In case of failure, fail the rest
767 * of the chain.
768 */
769 if (req->flags & REQ_F_LINK) {
770 if (req->flags & REQ_F_FAIL_LINK)
771 io_fail_links(req);
772 else
Jens Axboeba816ad2019-09-28 11:36:45 -0600773 io_req_link_next(req, nxt);
Jens Axboe9e645e112019-05-10 16:07:28 -0600774 }
775
776 __io_free_req(req);
777}
778
Jens Axboeba816ad2019-09-28 11:36:45 -0600779/*
780 * Drop reference to request, return next in chain (if there is one) if this
781 * was the last reference to this request.
782 */
783static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
Jens Axboee65ef562019-03-12 10:16:44 -0600784{
Jens Axboeba816ad2019-09-28 11:36:45 -0600785 struct io_kiocb *nxt = NULL;
786
Jens Axboee65ef562019-03-12 10:16:44 -0600787 if (refcount_dec_and_test(&req->refs))
Jens Axboeba816ad2019-09-28 11:36:45 -0600788 io_free_req(req, &nxt);
789
790 return nxt;
791}
792
793static void io_put_req(struct io_kiocb *req, struct io_kiocb **nxtptr)
794{
795 struct io_kiocb *nxt;
796
797 nxt = io_put_req_find_next(req);
798 if (nxt) {
Jens Axboe561fb042019-10-24 07:25:42 -0600799 if (nxtptr)
Jens Axboeba816ad2019-09-28 11:36:45 -0600800 *nxtptr = nxt;
Jens Axboe561fb042019-10-24 07:25:42 -0600801 else
Jens Axboeba816ad2019-09-28 11:36:45 -0600802 io_queue_async_work(nxt->ctx, nxt);
Jens Axboeba816ad2019-09-28 11:36:45 -0600803 }
Jens Axboe2b188cc2019-01-07 10:46:33 -0700804}
805
Hristo Venev75b28af2019-08-26 17:23:46 +0000806static unsigned io_cqring_events(struct io_rings *rings)
Jens Axboea3a0e432019-08-20 11:03:11 -0600807{
808 /* See comment at the top of this file */
809 smp_rmb();
Hristo Venev75b28af2019-08-26 17:23:46 +0000810 return READ_ONCE(rings->cq.tail) - READ_ONCE(rings->cq.head);
Jens Axboea3a0e432019-08-20 11:03:11 -0600811}
812
Pavel Begunkovfb5ccc92019-10-25 12:31:30 +0300813static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
814{
815 struct io_rings *rings = ctx->rings;
816
817 /* make sure SQ entry isn't read before tail */
818 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
819}
820
Jens Axboedef596e2019-01-09 08:59:42 -0700821/*
822 * Find and free completed poll iocbs
823 */
824static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
825 struct list_head *done)
826{
827 void *reqs[IO_IOPOLL_BATCH];
828 struct io_kiocb *req;
Jens Axboe09bb8392019-03-13 12:39:28 -0600829 int to_free;
Jens Axboedef596e2019-01-09 08:59:42 -0700830
Jens Axboe09bb8392019-03-13 12:39:28 -0600831 to_free = 0;
Jens Axboedef596e2019-01-09 08:59:42 -0700832 while (!list_empty(done)) {
833 req = list_first_entry(done, struct io_kiocb, list);
834 list_del(&req->list);
835
Jens Axboe9e645e112019-05-10 16:07:28 -0600836 io_cqring_fill_event(ctx, req->user_data, req->result);
Jens Axboedef596e2019-01-09 08:59:42 -0700837 (*nr_events)++;
838
Jens Axboe09bb8392019-03-13 12:39:28 -0600839 if (refcount_dec_and_test(&req->refs)) {
840 /* If we're not using fixed files, we have to pair the
841 * completion part with the file put. Use regular
842 * completions for those, only batch free for fixed
Jens Axboe9e645e112019-05-10 16:07:28 -0600843 * file and non-linked commands.
Jens Axboe09bb8392019-03-13 12:39:28 -0600844 */
Jens Axboe9e645e112019-05-10 16:07:28 -0600845 if ((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) ==
846 REQ_F_FIXED_FILE) {
Jens Axboe09bb8392019-03-13 12:39:28 -0600847 reqs[to_free++] = req;
848 if (to_free == ARRAY_SIZE(reqs))
849 io_free_req_many(ctx, reqs, &to_free);
Jens Axboe6b063142019-01-10 22:13:58 -0700850 } else {
Jens Axboeba816ad2019-09-28 11:36:45 -0600851 io_free_req(req, NULL);
Jens Axboe6b063142019-01-10 22:13:58 -0700852 }
Jens Axboe9a56a232019-01-09 09:06:50 -0700853 }
Jens Axboedef596e2019-01-09 08:59:42 -0700854 }
Jens Axboedef596e2019-01-09 08:59:42 -0700855
Jens Axboe09bb8392019-03-13 12:39:28 -0600856 io_commit_cqring(ctx);
Jens Axboedef596e2019-01-09 08:59:42 -0700857 io_free_req_many(ctx, reqs, &to_free);
858}
859
860static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
861 long min)
862{
863 struct io_kiocb *req, *tmp;
864 LIST_HEAD(done);
865 bool spin;
866 int ret;
867
868 /*
869 * Only spin for completions if we don't have multiple devices hanging
870 * off our complete list, and we're under the requested amount.
871 */
872 spin = !ctx->poll_multi_file && *nr_events < min;
873
874 ret = 0;
875 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
876 struct kiocb *kiocb = &req->rw;
877
878 /*
879 * Move completed entries to our local list. If we find a
880 * request that requires polling, break out and complete
881 * the done list first, if we have entries there.
882 */
883 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
884 list_move_tail(&req->list, &done);
885 continue;
886 }
887 if (!list_empty(&done))
888 break;
889
890 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
891 if (ret < 0)
892 break;
893
894 if (ret && spin)
895 spin = false;
896 ret = 0;
897 }
898
899 if (!list_empty(&done))
900 io_iopoll_complete(ctx, nr_events, &done);
901
902 return ret;
903}
904
905/*
906 * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a
907 * non-spinning poll check - we'll still enter the driver poll loop, but only
908 * as a non-spinning completion check.
909 */
910static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
911 long min)
912{
Jens Axboe08f54392019-08-21 22:19:11 -0600913 while (!list_empty(&ctx->poll_list) && !need_resched()) {
Jens Axboedef596e2019-01-09 08:59:42 -0700914 int ret;
915
916 ret = io_do_iopoll(ctx, nr_events, min);
917 if (ret < 0)
918 return ret;
919 if (!min || *nr_events >= min)
920 return 0;
921 }
922
923 return 1;
924}
925
926/*
927 * We can't just wait for polled events to come to us, we have to actively
928 * find and complete them.
929 */
930static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
931{
932 if (!(ctx->flags & IORING_SETUP_IOPOLL))
933 return;
934
935 mutex_lock(&ctx->uring_lock);
936 while (!list_empty(&ctx->poll_list)) {
937 unsigned int nr_events = 0;
938
939 io_iopoll_getevents(ctx, &nr_events, 1);
Jens Axboe08f54392019-08-21 22:19:11 -0600940
941 /*
942 * Ensure we allow local-to-the-cpu processing to take place,
943 * in this case we need to ensure that we reap all events.
944 */
945 cond_resched();
Jens Axboedef596e2019-01-09 08:59:42 -0700946 }
947 mutex_unlock(&ctx->uring_lock);
948}
949
Jens Axboe2b2ed972019-10-25 10:06:15 -0600950static int __io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
951 long min)
Jens Axboedef596e2019-01-09 08:59:42 -0700952{
Jens Axboe2b2ed972019-10-25 10:06:15 -0600953 int iters = 0, ret = 0;
Jens Axboedef596e2019-01-09 08:59:42 -0700954
955 do {
956 int tmin = 0;
957
Jens Axboe500f9fb2019-08-19 12:15:59 -0600958 /*
Jens Axboea3a0e432019-08-20 11:03:11 -0600959 * Don't enter poll loop if we already have events pending.
960 * If we do, we can potentially be spinning for commands that
961 * already triggered a CQE (eg in error).
962 */
Hristo Venev75b28af2019-08-26 17:23:46 +0000963 if (io_cqring_events(ctx->rings))
Jens Axboea3a0e432019-08-20 11:03:11 -0600964 break;
965
966 /*
Jens Axboe500f9fb2019-08-19 12:15:59 -0600967 * If a submit got punted to a workqueue, we can have the
968 * application entering polling for a command before it gets
969 * issued. That app will hold the uring_lock for the duration
970 * of the poll right here, so we need to take a breather every
971 * now and then to ensure that the issue has a chance to add
972 * the poll to the issued list. Otherwise we can spin here
973 * forever, while the workqueue is stuck trying to acquire the
974 * very same mutex.
975 */
976 if (!(++iters & 7)) {
977 mutex_unlock(&ctx->uring_lock);
978 mutex_lock(&ctx->uring_lock);
979 }
980
Jens Axboedef596e2019-01-09 08:59:42 -0700981 if (*nr_events < min)
982 tmin = min - *nr_events;
983
984 ret = io_iopoll_getevents(ctx, nr_events, tmin);
985 if (ret <= 0)
986 break;
987 ret = 0;
988 } while (min && !*nr_events && !need_resched());
989
Jens Axboe2b2ed972019-10-25 10:06:15 -0600990 return ret;
991}
992
993static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
994 long min)
995{
996 int ret;
997
998 /*
999 * We disallow the app entering submit/complete with polling, but we
1000 * still need to lock the ring to prevent racing with polled issue
1001 * that got punted to a workqueue.
1002 */
1003 mutex_lock(&ctx->uring_lock);
1004 ret = __io_iopoll_check(ctx, nr_events, min);
Jens Axboe500f9fb2019-08-19 12:15:59 -06001005 mutex_unlock(&ctx->uring_lock);
Jens Axboedef596e2019-01-09 08:59:42 -07001006 return ret;
1007}
1008
Jens Axboe491381ce2019-10-17 09:20:46 -06001009static void kiocb_end_write(struct io_kiocb *req)
Jens Axboe2b188cc2019-01-07 10:46:33 -07001010{
Jens Axboe491381ce2019-10-17 09:20:46 -06001011 /*
1012 * Tell lockdep we inherited freeze protection from submission
1013 * thread.
1014 */
1015 if (req->flags & REQ_F_ISREG) {
1016 struct inode *inode = file_inode(req->file);
Jens Axboe2b188cc2019-01-07 10:46:33 -07001017
Jens Axboe491381ce2019-10-17 09:20:46 -06001018 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
Jens Axboe2b188cc2019-01-07 10:46:33 -07001019 }
Jens Axboe491381ce2019-10-17 09:20:46 -06001020 file_end_write(req->file);
Jens Axboe2b188cc2019-01-07 10:46:33 -07001021}
1022
Jens Axboeba816ad2019-09-28 11:36:45 -06001023static void io_complete_rw_common(struct kiocb *kiocb, long res)
Jens Axboe2b188cc2019-01-07 10:46:33 -07001024{
1025 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1026
Jens Axboe491381ce2019-10-17 09:20:46 -06001027 if (kiocb->ki_flags & IOCB_WRITE)
1028 kiocb_end_write(req);
Jens Axboe2b188cc2019-01-07 10:46:33 -07001029
Jens Axboe9e645e112019-05-10 16:07:28 -06001030 if ((req->flags & REQ_F_LINK) && res != req->result)
1031 req->flags |= REQ_F_FAIL_LINK;
Jens Axboec71ffb62019-05-13 20:58:29 -06001032 io_cqring_add_event(req->ctx, req->user_data, res);
Jens Axboeba816ad2019-09-28 11:36:45 -06001033}
1034
1035static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1036{
1037 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1038
1039 io_complete_rw_common(kiocb, res);
1040 io_put_req(req, NULL);
1041}
1042
1043static struct io_kiocb *__io_complete_rw(struct kiocb *kiocb, long res)
1044{
1045 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1046
1047 io_complete_rw_common(kiocb, res);
1048 return io_put_req_find_next(req);
Jens Axboe2b188cc2019-01-07 10:46:33 -07001049}
1050
Jens Axboedef596e2019-01-09 08:59:42 -07001051static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1052{
1053 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1054
Jens Axboe491381ce2019-10-17 09:20:46 -06001055 if (kiocb->ki_flags & IOCB_WRITE)
1056 kiocb_end_write(req);
Jens Axboedef596e2019-01-09 08:59:42 -07001057
Jens Axboe9e645e112019-05-10 16:07:28 -06001058 if ((req->flags & REQ_F_LINK) && res != req->result)
1059 req->flags |= REQ_F_FAIL_LINK;
1060 req->result = res;
Jens Axboedef596e2019-01-09 08:59:42 -07001061 if (res != -EAGAIN)
1062 req->flags |= REQ_F_IOPOLL_COMPLETED;
1063}
1064
1065/*
1066 * After the iocb has been issued, it's safe to be found on the poll list.
1067 * Adding the kiocb to the list AFTER submission ensures that we don't
1068 * find it from a io_iopoll_getevents() thread before the issuer is done
1069 * accessing the kiocb cookie.
1070 */
1071static void io_iopoll_req_issued(struct io_kiocb *req)
1072{
1073 struct io_ring_ctx *ctx = req->ctx;
1074
1075 /*
1076 * Track whether we have multiple files in our lists. This will impact
1077 * how we do polling eventually, not spinning if we're on potentially
1078 * different devices.
1079 */
1080 if (list_empty(&ctx->poll_list)) {
1081 ctx->poll_multi_file = false;
1082 } else if (!ctx->poll_multi_file) {
1083 struct io_kiocb *list_req;
1084
1085 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
1086 list);
1087 if (list_req->rw.ki_filp != req->rw.ki_filp)
1088 ctx->poll_multi_file = true;
1089 }
1090
1091 /*
1092 * For fast devices, IO may have already completed. If it has, add
1093 * it to the front so we find it first.
1094 */
1095 if (req->flags & REQ_F_IOPOLL_COMPLETED)
1096 list_add(&req->list, &ctx->poll_list);
1097 else
1098 list_add_tail(&req->list, &ctx->poll_list);
1099}
1100
Jens Axboe3d6770f2019-04-13 11:50:54 -06001101static void io_file_put(struct io_submit_state *state)
Jens Axboe9a56a232019-01-09 09:06:50 -07001102{
Jens Axboe3d6770f2019-04-13 11:50:54 -06001103 if (state->file) {
Jens Axboe9a56a232019-01-09 09:06:50 -07001104 int diff = state->has_refs - state->used_refs;
1105
1106 if (diff)
1107 fput_many(state->file, diff);
1108 state->file = NULL;
1109 }
1110}
1111
1112/*
1113 * Get as many references to a file as we have IOs left in this submission,
1114 * assuming most submissions are for one file, or at least that each file
1115 * has more than one submission.
1116 */
1117static struct file *io_file_get(struct io_submit_state *state, int fd)
1118{
1119 if (!state)
1120 return fget(fd);
1121
1122 if (state->file) {
1123 if (state->fd == fd) {
1124 state->used_refs++;
1125 state->ios_left--;
1126 return state->file;
1127 }
Jens Axboe3d6770f2019-04-13 11:50:54 -06001128 io_file_put(state);
Jens Axboe9a56a232019-01-09 09:06:50 -07001129 }
1130 state->file = fget_many(fd, state->ios_left);
1131 if (!state->file)
1132 return NULL;
1133
1134 state->fd = fd;
1135 state->has_refs = state->ios_left;
1136 state->used_refs = 1;
1137 state->ios_left--;
1138 return state->file;
1139}
1140
Jens Axboe2b188cc2019-01-07 10:46:33 -07001141/*
1142 * If we tracked the file through the SCM inflight mechanism, we could support
1143 * any file. For now, just ensure that anything potentially problematic is done
1144 * inline.
1145 */
1146static bool io_file_supports_async(struct file *file)
1147{
1148 umode_t mode = file_inode(file)->i_mode;
1149
1150 if (S_ISBLK(mode) || S_ISCHR(mode))
1151 return true;
1152 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
1153 return true;
1154
1155 return false;
1156}
1157
Jens Axboe6c271ce2019-01-10 11:22:30 -07001158static int io_prep_rw(struct io_kiocb *req, const struct sqe_submit *s,
Jens Axboe8358e3a2019-04-23 08:17:58 -06001159 bool force_nonblock)
Jens Axboe2b188cc2019-01-07 10:46:33 -07001160{
Jens Axboe6c271ce2019-01-10 11:22:30 -07001161 const struct io_uring_sqe *sqe = s->sqe;
Jens Axboedef596e2019-01-09 08:59:42 -07001162 struct io_ring_ctx *ctx = req->ctx;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001163 struct kiocb *kiocb = &req->rw;
Jens Axboe09bb8392019-03-13 12:39:28 -06001164 unsigned ioprio;
1165 int ret;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001166
Jens Axboe09bb8392019-03-13 12:39:28 -06001167 if (!req->file)
1168 return -EBADF;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001169
Jens Axboe491381ce2019-10-17 09:20:46 -06001170 if (S_ISREG(file_inode(req->file)->i_mode))
1171 req->flags |= REQ_F_ISREG;
1172
1173 /*
1174 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
1175 * we know to async punt it even if it was opened O_NONBLOCK
1176 */
1177 if (force_nonblock && !io_file_supports_async(req->file)) {
1178 req->flags |= REQ_F_MUST_PUNT;
1179 return -EAGAIN;
1180 }
Jens Axboe6b063142019-01-10 22:13:58 -07001181
Jens Axboe2b188cc2019-01-07 10:46:33 -07001182 kiocb->ki_pos = READ_ONCE(sqe->off);
1183 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
1184 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
1185
1186 ioprio = READ_ONCE(sqe->ioprio);
1187 if (ioprio) {
1188 ret = ioprio_check_cap(ioprio);
1189 if (ret)
Jens Axboe09bb8392019-03-13 12:39:28 -06001190 return ret;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001191
1192 kiocb->ki_ioprio = ioprio;
1193 } else
1194 kiocb->ki_ioprio = get_current_ioprio();
1195
1196 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
1197 if (unlikely(ret))
Jens Axboe09bb8392019-03-13 12:39:28 -06001198 return ret;
Stefan Bühler8449eed2019-04-27 20:34:19 +02001199
1200 /* don't allow async punt if RWF_NOWAIT was requested */
Jens Axboe491381ce2019-10-17 09:20:46 -06001201 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
1202 (req->file->f_flags & O_NONBLOCK))
Stefan Bühler8449eed2019-04-27 20:34:19 +02001203 req->flags |= REQ_F_NOWAIT;
1204
1205 if (force_nonblock)
Jens Axboe2b188cc2019-01-07 10:46:33 -07001206 kiocb->ki_flags |= IOCB_NOWAIT;
Stefan Bühler8449eed2019-04-27 20:34:19 +02001207
Jens Axboedef596e2019-01-09 08:59:42 -07001208 if (ctx->flags & IORING_SETUP_IOPOLL) {
Jens Axboedef596e2019-01-09 08:59:42 -07001209 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
1210 !kiocb->ki_filp->f_op->iopoll)
Jens Axboe09bb8392019-03-13 12:39:28 -06001211 return -EOPNOTSUPP;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001212
Jens Axboedef596e2019-01-09 08:59:42 -07001213 kiocb->ki_flags |= IOCB_HIPRI;
1214 kiocb->ki_complete = io_complete_rw_iopoll;
1215 } else {
Jens Axboe09bb8392019-03-13 12:39:28 -06001216 if (kiocb->ki_flags & IOCB_HIPRI)
1217 return -EINVAL;
Jens Axboedef596e2019-01-09 08:59:42 -07001218 kiocb->ki_complete = io_complete_rw;
1219 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07001220 return 0;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001221}
1222
1223static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
1224{
1225 switch (ret) {
1226 case -EIOCBQUEUED:
1227 break;
1228 case -ERESTARTSYS:
1229 case -ERESTARTNOINTR:
1230 case -ERESTARTNOHAND:
1231 case -ERESTART_RESTARTBLOCK:
1232 /*
1233 * We can't just restart the syscall, since previously
1234 * submitted sqes may already be in progress. Just fail this
1235 * IO with EINTR.
1236 */
1237 ret = -EINTR;
1238 /* fall through */
1239 default:
1240 kiocb->ki_complete(kiocb, ret, 0);
1241 }
1242}
1243
Jens Axboeba816ad2019-09-28 11:36:45 -06001244static void kiocb_done(struct kiocb *kiocb, ssize_t ret, struct io_kiocb **nxt,
1245 bool in_async)
1246{
1247 if (in_async && ret >= 0 && nxt && kiocb->ki_complete == io_complete_rw)
1248 *nxt = __io_complete_rw(kiocb, ret);
1249 else
1250 io_rw_done(kiocb, ret);
1251}
1252
Jens Axboeedafcce2019-01-09 09:16:05 -07001253static int io_import_fixed(struct io_ring_ctx *ctx, int rw,
1254 const struct io_uring_sqe *sqe,
1255 struct iov_iter *iter)
1256{
1257 size_t len = READ_ONCE(sqe->len);
1258 struct io_mapped_ubuf *imu;
1259 unsigned index, buf_index;
1260 size_t offset;
1261 u64 buf_addr;
1262
1263 /* attempt to use fixed buffers without having provided iovecs */
1264 if (unlikely(!ctx->user_bufs))
1265 return -EFAULT;
1266
1267 buf_index = READ_ONCE(sqe->buf_index);
1268 if (unlikely(buf_index >= ctx->nr_user_bufs))
1269 return -EFAULT;
1270
1271 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
1272 imu = &ctx->user_bufs[index];
1273 buf_addr = READ_ONCE(sqe->addr);
1274
1275 /* overflow */
1276 if (buf_addr + len < buf_addr)
1277 return -EFAULT;
1278 /* not inside the mapped region */
1279 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
1280 return -EFAULT;
1281
1282 /*
1283 * May not be a start of buffer, set size appropriately
1284 * and advance us to the beginning.
1285 */
1286 offset = buf_addr - imu->ubuf;
1287 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
Jens Axboebd11b3a2019-07-20 08:37:31 -06001288
1289 if (offset) {
1290 /*
1291 * Don't use iov_iter_advance() here, as it's really slow for
1292 * using the latter parts of a big fixed buffer - it iterates
1293 * over each segment manually. We can cheat a bit here, because
1294 * we know that:
1295 *
1296 * 1) it's a BVEC iter, we set it up
1297 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1298 * first and last bvec
1299 *
1300 * So just find our index, and adjust the iterator afterwards.
1301 * If the offset is within the first bvec (or the whole first
1302 * bvec, just use iov_iter_advance(). This makes it easier
1303 * since we can just skip the first segment, which may not
1304 * be PAGE_SIZE aligned.
1305 */
1306 const struct bio_vec *bvec = imu->bvec;
1307
1308 if (offset <= bvec->bv_len) {
1309 iov_iter_advance(iter, offset);
1310 } else {
1311 unsigned long seg_skip;
1312
1313 /* skip first vec */
1314 offset -= bvec->bv_len;
1315 seg_skip = 1 + (offset >> PAGE_SHIFT);
1316
1317 iter->bvec = bvec + seg_skip;
1318 iter->nr_segs -= seg_skip;
Aleix Roca Nonell99c79f62019-08-15 14:03:22 +02001319 iter->count -= bvec->bv_len + offset;
Jens Axboebd11b3a2019-07-20 08:37:31 -06001320 iter->iov_offset = offset & ~PAGE_MASK;
Jens Axboebd11b3a2019-07-20 08:37:31 -06001321 }
1322 }
1323
Jens Axboeedafcce2019-01-09 09:16:05 -07001324 return 0;
1325}
1326
Jens Axboe87e5e6d2019-05-14 16:02:22 -06001327static ssize_t io_import_iovec(struct io_ring_ctx *ctx, int rw,
1328 const struct sqe_submit *s, struct iovec **iovec,
1329 struct iov_iter *iter)
Jens Axboe2b188cc2019-01-07 10:46:33 -07001330{
1331 const struct io_uring_sqe *sqe = s->sqe;
1332 void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
1333 size_t sqe_len = READ_ONCE(sqe->len);
Jens Axboeedafcce2019-01-09 09:16:05 -07001334 u8 opcode;
1335
1336 /*
1337 * We're reading ->opcode for the second time, but the first read
1338 * doesn't care whether it's _FIXED or not, so it doesn't matter
1339 * whether ->opcode changes concurrently. The first read does care
1340 * about whether it is a READ or a WRITE, so we don't trust this read
1341 * for that purpose and instead let the caller pass in the read/write
1342 * flag.
1343 */
1344 opcode = READ_ONCE(sqe->opcode);
1345 if (opcode == IORING_OP_READ_FIXED ||
1346 opcode == IORING_OP_WRITE_FIXED) {
Jens Axboe87e5e6d2019-05-14 16:02:22 -06001347 ssize_t ret = io_import_fixed(ctx, rw, sqe, iter);
Jens Axboeedafcce2019-01-09 09:16:05 -07001348 *iovec = NULL;
1349 return ret;
1350 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07001351
1352 if (!s->has_user)
1353 return -EFAULT;
1354
1355#ifdef CONFIG_COMPAT
1356 if (ctx->compat)
1357 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
1358 iovec, iter);
1359#endif
1360
1361 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
1362}
1363
Jens Axboe32960612019-09-23 11:05:34 -06001364/*
1365 * For files that don't have ->read_iter() and ->write_iter(), handle them
1366 * by looping over ->read() or ->write() manually.
1367 */
1368static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
1369 struct iov_iter *iter)
1370{
1371 ssize_t ret = 0;
1372
1373 /*
1374 * Don't support polled IO through this interface, and we can't
1375 * support non-blocking either. For the latter, this just causes
1376 * the kiocb to be handled from an async context.
1377 */
1378 if (kiocb->ki_flags & IOCB_HIPRI)
1379 return -EOPNOTSUPP;
1380 if (kiocb->ki_flags & IOCB_NOWAIT)
1381 return -EAGAIN;
1382
1383 while (iov_iter_count(iter)) {
1384 struct iovec iovec = iov_iter_iovec(iter);
1385 ssize_t nr;
1386
1387 if (rw == READ) {
1388 nr = file->f_op->read(file, iovec.iov_base,
1389 iovec.iov_len, &kiocb->ki_pos);
1390 } else {
1391 nr = file->f_op->write(file, iovec.iov_base,
1392 iovec.iov_len, &kiocb->ki_pos);
1393 }
1394
1395 if (nr < 0) {
1396 if (!ret)
1397 ret = nr;
1398 break;
1399 }
1400 ret += nr;
1401 if (nr != iovec.iov_len)
1402 break;
1403 iov_iter_advance(iter, nr);
1404 }
1405
1406 return ret;
1407}
1408
Jens Axboee0c5c572019-03-12 10:18:47 -06001409static int io_read(struct io_kiocb *req, const struct sqe_submit *s,
Jens Axboeba816ad2019-09-28 11:36:45 -06001410 struct io_kiocb **nxt, bool force_nonblock)
Jens Axboe2b188cc2019-01-07 10:46:33 -07001411{
1412 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1413 struct kiocb *kiocb = &req->rw;
1414 struct iov_iter iter;
1415 struct file *file;
Jens Axboe31b51512019-01-18 22:56:34 -07001416 size_t iov_count;
Jens Axboe9d93a3f2019-05-15 13:53:07 -06001417 ssize_t read_size, ret;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001418
Jens Axboe8358e3a2019-04-23 08:17:58 -06001419 ret = io_prep_rw(req, s, force_nonblock);
Jens Axboe2b188cc2019-01-07 10:46:33 -07001420 if (ret)
1421 return ret;
1422 file = kiocb->ki_filp;
1423
Jens Axboe2b188cc2019-01-07 10:46:33 -07001424 if (unlikely(!(file->f_mode & FMODE_READ)))
Jens Axboe09bb8392019-03-13 12:39:28 -06001425 return -EBADF;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001426
1427 ret = io_import_iovec(req->ctx, READ, s, &iovec, &iter);
Jens Axboe87e5e6d2019-05-14 16:02:22 -06001428 if (ret < 0)
Jens Axboe09bb8392019-03-13 12:39:28 -06001429 return ret;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001430
Jens Axboe9d93a3f2019-05-15 13:53:07 -06001431 read_size = ret;
Jens Axboe9e645e112019-05-10 16:07:28 -06001432 if (req->flags & REQ_F_LINK)
1433 req->result = read_size;
1434
Jens Axboe31b51512019-01-18 22:56:34 -07001435 iov_count = iov_iter_count(&iter);
1436 ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count);
Jens Axboe2b188cc2019-01-07 10:46:33 -07001437 if (!ret) {
1438 ssize_t ret2;
1439
Jens Axboe32960612019-09-23 11:05:34 -06001440 if (file->f_op->read_iter)
1441 ret2 = call_read_iter(file, kiocb, &iter);
1442 else
1443 ret2 = loop_rw_iter(READ, file, kiocb, &iter);
1444
Jens Axboe9d93a3f2019-05-15 13:53:07 -06001445 /*
1446 * In case of a short read, punt to async. This can happen
1447 * if we have data partially cached. Alternatively we can
1448 * return the short read, in which case the application will
1449 * need to issue another SQE and wait for it. That SQE will
1450 * need async punt anyway, so it's more efficient to do it
1451 * here.
1452 */
Jens Axboe491381ce2019-10-17 09:20:46 -06001453 if (force_nonblock && !(req->flags & REQ_F_NOWAIT) &&
1454 (req->flags & REQ_F_ISREG) &&
1455 ret2 > 0 && ret2 < read_size)
Jens Axboe9d93a3f2019-05-15 13:53:07 -06001456 ret2 = -EAGAIN;
1457 /* Catch -EAGAIN return for forced non-blocking submission */
Jens Axboe561fb042019-10-24 07:25:42 -06001458 if (!force_nonblock || ret2 != -EAGAIN)
Jackie Liuba5290c2019-10-09 09:19:59 +08001459 kiocb_done(kiocb, ret2, nxt, s->in_async);
Jens Axboe561fb042019-10-24 07:25:42 -06001460 else
Jens Axboe2b188cc2019-01-07 10:46:33 -07001461 ret = -EAGAIN;
1462 }
1463 kfree(iovec);
Jens Axboe2b188cc2019-01-07 10:46:33 -07001464 return ret;
1465}
1466
Jens Axboee0c5c572019-03-12 10:18:47 -06001467static int io_write(struct io_kiocb *req, const struct sqe_submit *s,
Jens Axboeba816ad2019-09-28 11:36:45 -06001468 struct io_kiocb **nxt, bool force_nonblock)
Jens Axboe2b188cc2019-01-07 10:46:33 -07001469{
1470 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1471 struct kiocb *kiocb = &req->rw;
1472 struct iov_iter iter;
1473 struct file *file;
Jens Axboe31b51512019-01-18 22:56:34 -07001474 size_t iov_count;
Jens Axboe87e5e6d2019-05-14 16:02:22 -06001475 ssize_t ret;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001476
Jens Axboe8358e3a2019-04-23 08:17:58 -06001477 ret = io_prep_rw(req, s, force_nonblock);
Jens Axboe2b188cc2019-01-07 10:46:33 -07001478 if (ret)
1479 return ret;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001480
Jens Axboe2b188cc2019-01-07 10:46:33 -07001481 file = kiocb->ki_filp;
1482 if (unlikely(!(file->f_mode & FMODE_WRITE)))
Jens Axboe09bb8392019-03-13 12:39:28 -06001483 return -EBADF;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001484
1485 ret = io_import_iovec(req->ctx, WRITE, s, &iovec, &iter);
Jens Axboe87e5e6d2019-05-14 16:02:22 -06001486 if (ret < 0)
Jens Axboe09bb8392019-03-13 12:39:28 -06001487 return ret;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001488
Jens Axboe9e645e112019-05-10 16:07:28 -06001489 if (req->flags & REQ_F_LINK)
1490 req->result = ret;
1491
Jens Axboe31b51512019-01-18 22:56:34 -07001492 iov_count = iov_iter_count(&iter);
1493
1494 ret = -EAGAIN;
Jens Axboe561fb042019-10-24 07:25:42 -06001495 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT))
Jens Axboe31b51512019-01-18 22:56:34 -07001496 goto out_free;
Jens Axboe31b51512019-01-18 22:56:34 -07001497
1498 ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count);
Jens Axboe2b188cc2019-01-07 10:46:33 -07001499 if (!ret) {
Roman Penyaev9bf79332019-03-25 20:09:24 +01001500 ssize_t ret2;
1501
Jens Axboe2b188cc2019-01-07 10:46:33 -07001502 /*
1503 * Open-code file_start_write here to grab freeze protection,
1504 * which will be released by another thread in
1505 * io_complete_rw(). Fool lockdep by telling it the lock got
1506 * released so that it doesn't complain about the held lock when
1507 * we return to userspace.
1508 */
Jens Axboe491381ce2019-10-17 09:20:46 -06001509 if (req->flags & REQ_F_ISREG) {
Jens Axboe2b188cc2019-01-07 10:46:33 -07001510 __sb_start_write(file_inode(file)->i_sb,
1511 SB_FREEZE_WRITE, true);
1512 __sb_writers_release(file_inode(file)->i_sb,
1513 SB_FREEZE_WRITE);
1514 }
1515 kiocb->ki_flags |= IOCB_WRITE;
Roman Penyaev9bf79332019-03-25 20:09:24 +01001516
Jens Axboe32960612019-09-23 11:05:34 -06001517 if (file->f_op->write_iter)
1518 ret2 = call_write_iter(file, kiocb, &iter);
1519 else
1520 ret2 = loop_rw_iter(WRITE, file, kiocb, &iter);
Jens Axboe561fb042019-10-24 07:25:42 -06001521 if (!force_nonblock || ret2 != -EAGAIN)
Jackie Liuba5290c2019-10-09 09:19:59 +08001522 kiocb_done(kiocb, ret2, nxt, s->in_async);
Jens Axboe561fb042019-10-24 07:25:42 -06001523 else
Roman Penyaev9bf79332019-03-25 20:09:24 +01001524 ret = -EAGAIN;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001525 }
Jens Axboe31b51512019-01-18 22:56:34 -07001526out_free:
Jens Axboe2b188cc2019-01-07 10:46:33 -07001527 kfree(iovec);
Jens Axboe2b188cc2019-01-07 10:46:33 -07001528 return ret;
1529}
1530
1531/*
1532 * IORING_OP_NOP just posts a completion event, nothing else.
1533 */
1534static int io_nop(struct io_kiocb *req, u64 user_data)
1535{
1536 struct io_ring_ctx *ctx = req->ctx;
1537 long err = 0;
1538
Jens Axboedef596e2019-01-09 08:59:42 -07001539 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1540 return -EINVAL;
1541
Jens Axboec71ffb62019-05-13 20:58:29 -06001542 io_cqring_add_event(ctx, user_data, err);
Jens Axboeba816ad2019-09-28 11:36:45 -06001543 io_put_req(req, NULL);
Jens Axboe2b188cc2019-01-07 10:46:33 -07001544 return 0;
1545}
1546
Christoph Hellwigc992fe22019-01-11 09:43:02 -07001547static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1548{
Jens Axboe6b063142019-01-10 22:13:58 -07001549 struct io_ring_ctx *ctx = req->ctx;
Christoph Hellwigc992fe22019-01-11 09:43:02 -07001550
Jens Axboe09bb8392019-03-13 12:39:28 -06001551 if (!req->file)
1552 return -EBADF;
Christoph Hellwigc992fe22019-01-11 09:43:02 -07001553
Jens Axboe6b063142019-01-10 22:13:58 -07001554 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
Jens Axboedef596e2019-01-09 08:59:42 -07001555 return -EINVAL;
Jens Axboeedafcce2019-01-09 09:16:05 -07001556 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
Christoph Hellwigc992fe22019-01-11 09:43:02 -07001557 return -EINVAL;
1558
Christoph Hellwigc992fe22019-01-11 09:43:02 -07001559 return 0;
1560}
1561
1562static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe,
Jens Axboeba816ad2019-09-28 11:36:45 -06001563 struct io_kiocb **nxt, bool force_nonblock)
Christoph Hellwigc992fe22019-01-11 09:43:02 -07001564{
1565 loff_t sqe_off = READ_ONCE(sqe->off);
1566 loff_t sqe_len = READ_ONCE(sqe->len);
1567 loff_t end = sqe_off + sqe_len;
1568 unsigned fsync_flags;
1569 int ret;
1570
1571 fsync_flags = READ_ONCE(sqe->fsync_flags);
1572 if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC))
1573 return -EINVAL;
1574
1575 ret = io_prep_fsync(req, sqe);
1576 if (ret)
1577 return ret;
1578
1579 /* fsync always requires a blocking context */
1580 if (force_nonblock)
1581 return -EAGAIN;
1582
1583 ret = vfs_fsync_range(req->rw.ki_filp, sqe_off,
1584 end > 0 ? end : LLONG_MAX,
1585 fsync_flags & IORING_FSYNC_DATASYNC);
1586
Jens Axboe9e645e112019-05-10 16:07:28 -06001587 if (ret < 0 && (req->flags & REQ_F_LINK))
1588 req->flags |= REQ_F_FAIL_LINK;
Jens Axboec71ffb62019-05-13 20:58:29 -06001589 io_cqring_add_event(req->ctx, sqe->user_data, ret);
Jens Axboeba816ad2019-09-28 11:36:45 -06001590 io_put_req(req, nxt);
Christoph Hellwigc992fe22019-01-11 09:43:02 -07001591 return 0;
1592}
1593
Jens Axboe5d17b4a2019-04-09 14:56:44 -06001594static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1595{
1596 struct io_ring_ctx *ctx = req->ctx;
1597 int ret = 0;
1598
1599 if (!req->file)
1600 return -EBADF;
Jens Axboe5d17b4a2019-04-09 14:56:44 -06001601
1602 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1603 return -EINVAL;
1604 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1605 return -EINVAL;
1606
Jens Axboe5d17b4a2019-04-09 14:56:44 -06001607 return ret;
1608}
1609
1610static int io_sync_file_range(struct io_kiocb *req,
1611 const struct io_uring_sqe *sqe,
Jens Axboeba816ad2019-09-28 11:36:45 -06001612 struct io_kiocb **nxt,
Jens Axboe5d17b4a2019-04-09 14:56:44 -06001613 bool force_nonblock)
1614{
1615 loff_t sqe_off;
1616 loff_t sqe_len;
1617 unsigned flags;
1618 int ret;
1619
1620 ret = io_prep_sfr(req, sqe);
1621 if (ret)
1622 return ret;
1623
1624 /* sync_file_range always requires a blocking context */
1625 if (force_nonblock)
1626 return -EAGAIN;
1627
1628 sqe_off = READ_ONCE(sqe->off);
1629 sqe_len = READ_ONCE(sqe->len);
1630 flags = READ_ONCE(sqe->sync_range_flags);
1631
1632 ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags);
1633
Jens Axboe9e645e112019-05-10 16:07:28 -06001634 if (ret < 0 && (req->flags & REQ_F_LINK))
1635 req->flags |= REQ_F_FAIL_LINK;
Jens Axboec71ffb62019-05-13 20:58:29 -06001636 io_cqring_add_event(req->ctx, sqe->user_data, ret);
Jens Axboeba816ad2019-09-28 11:36:45 -06001637 io_put_req(req, nxt);
Jens Axboe5d17b4a2019-04-09 14:56:44 -06001638 return 0;
1639}
1640
Jens Axboe0fa03c62019-04-19 13:34:07 -06001641#if defined(CONFIG_NET)
Jens Axboeaa1fa282019-04-19 13:38:09 -06001642static int io_send_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
Jens Axboeba816ad2019-09-28 11:36:45 -06001643 struct io_kiocb **nxt, bool force_nonblock,
Jens Axboeaa1fa282019-04-19 13:38:09 -06001644 long (*fn)(struct socket *, struct user_msghdr __user *,
1645 unsigned int))
1646{
Jens Axboe0fa03c62019-04-19 13:34:07 -06001647 struct socket *sock;
1648 int ret;
1649
1650 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1651 return -EINVAL;
1652
1653 sock = sock_from_file(req->file, &ret);
1654 if (sock) {
1655 struct user_msghdr __user *msg;
1656 unsigned flags;
1657
1658 flags = READ_ONCE(sqe->msg_flags);
1659 if (flags & MSG_DONTWAIT)
1660 req->flags |= REQ_F_NOWAIT;
1661 else if (force_nonblock)
1662 flags |= MSG_DONTWAIT;
1663
1664 msg = (struct user_msghdr __user *) (unsigned long)
1665 READ_ONCE(sqe->addr);
1666
Jens Axboeaa1fa282019-04-19 13:38:09 -06001667 ret = fn(sock, msg, flags);
Jens Axboe0fa03c62019-04-19 13:34:07 -06001668 if (force_nonblock && ret == -EAGAIN)
1669 return ret;
1670 }
1671
1672 io_cqring_add_event(req->ctx, sqe->user_data, ret);
Jens Axboef1f40852019-11-05 20:33:16 -07001673 if (ret < 0 && (req->flags & REQ_F_LINK))
1674 req->flags |= REQ_F_FAIL_LINK;
Jens Axboeba816ad2019-09-28 11:36:45 -06001675 io_put_req(req, nxt);
Jens Axboe0fa03c62019-04-19 13:34:07 -06001676 return 0;
Jens Axboeaa1fa282019-04-19 13:38:09 -06001677}
1678#endif
1679
1680static int io_sendmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
Jens Axboeba816ad2019-09-28 11:36:45 -06001681 struct io_kiocb **nxt, bool force_nonblock)
Jens Axboeaa1fa282019-04-19 13:38:09 -06001682{
1683#if defined(CONFIG_NET)
Jens Axboeba816ad2019-09-28 11:36:45 -06001684 return io_send_recvmsg(req, sqe, nxt, force_nonblock,
1685 __sys_sendmsg_sock);
Jens Axboeaa1fa282019-04-19 13:38:09 -06001686#else
1687 return -EOPNOTSUPP;
1688#endif
1689}
1690
1691static int io_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
Jens Axboeba816ad2019-09-28 11:36:45 -06001692 struct io_kiocb **nxt, bool force_nonblock)
Jens Axboeaa1fa282019-04-19 13:38:09 -06001693{
1694#if defined(CONFIG_NET)
Jens Axboeba816ad2019-09-28 11:36:45 -06001695 return io_send_recvmsg(req, sqe, nxt, force_nonblock,
1696 __sys_recvmsg_sock);
Jens Axboe0fa03c62019-04-19 13:34:07 -06001697#else
1698 return -EOPNOTSUPP;
1699#endif
1700}
1701
Jens Axboe17f2fe32019-10-17 14:42:58 -06001702static int io_accept(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1703 struct io_kiocb **nxt, bool force_nonblock)
1704{
1705#if defined(CONFIG_NET)
1706 struct sockaddr __user *addr;
1707 int __user *addr_len;
1708 unsigned file_flags;
1709 int flags, ret;
1710
1711 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
1712 return -EINVAL;
1713 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
1714 return -EINVAL;
1715
1716 addr = (struct sockaddr __user *) (unsigned long) READ_ONCE(sqe->addr);
1717 addr_len = (int __user *) (unsigned long) READ_ONCE(sqe->addr2);
1718 flags = READ_ONCE(sqe->accept_flags);
1719 file_flags = force_nonblock ? O_NONBLOCK : 0;
1720
1721 ret = __sys_accept4_file(req->file, file_flags, addr, addr_len, flags);
1722 if (ret == -EAGAIN && force_nonblock) {
1723 req->work.flags |= IO_WQ_WORK_NEEDS_FILES;
1724 return -EAGAIN;
1725 }
1726 if (ret < 0 && (req->flags & REQ_F_LINK))
1727 req->flags |= REQ_F_FAIL_LINK;
1728 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1729 io_put_req(req, nxt);
1730 return 0;
1731#else
1732 return -EOPNOTSUPP;
1733#endif
1734}
1735
Jens Axboe221c5eb2019-01-17 09:41:58 -07001736static void io_poll_remove_one(struct io_kiocb *req)
1737{
1738 struct io_poll_iocb *poll = &req->poll;
1739
1740 spin_lock(&poll->head->lock);
1741 WRITE_ONCE(poll->canceled, true);
1742 if (!list_empty(&poll->wait.entry)) {
1743 list_del_init(&poll->wait.entry);
Jens Axboe18d9be12019-09-10 09:13:05 -06001744 io_queue_async_work(req->ctx, req);
Jens Axboe221c5eb2019-01-17 09:41:58 -07001745 }
1746 spin_unlock(&poll->head->lock);
1747
1748 list_del_init(&req->list);
1749}
1750
1751static void io_poll_remove_all(struct io_ring_ctx *ctx)
1752{
1753 struct io_kiocb *req;
1754
1755 spin_lock_irq(&ctx->completion_lock);
1756 while (!list_empty(&ctx->cancel_list)) {
1757 req = list_first_entry(&ctx->cancel_list, struct io_kiocb,list);
1758 io_poll_remove_one(req);
1759 }
1760 spin_unlock_irq(&ctx->completion_lock);
1761}
1762
1763/*
1764 * Find a running poll command that matches one specified in sqe->addr,
1765 * and remove it if found.
1766 */
1767static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1768{
1769 struct io_ring_ctx *ctx = req->ctx;
1770 struct io_kiocb *poll_req, *next;
1771 int ret = -ENOENT;
1772
1773 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1774 return -EINVAL;
1775 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
1776 sqe->poll_events)
1777 return -EINVAL;
1778
1779 spin_lock_irq(&ctx->completion_lock);
1780 list_for_each_entry_safe(poll_req, next, &ctx->cancel_list, list) {
1781 if (READ_ONCE(sqe->addr) == poll_req->user_data) {
1782 io_poll_remove_one(poll_req);
1783 ret = 0;
1784 break;
1785 }
1786 }
1787 spin_unlock_irq(&ctx->completion_lock);
1788
Jens Axboec71ffb62019-05-13 20:58:29 -06001789 io_cqring_add_event(req->ctx, sqe->user_data, ret);
Jens Axboef1f40852019-11-05 20:33:16 -07001790 if (ret < 0 && (req->flags & REQ_F_LINK))
1791 req->flags |= REQ_F_FAIL_LINK;
Jens Axboeba816ad2019-09-28 11:36:45 -06001792 io_put_req(req, NULL);
Jens Axboe221c5eb2019-01-17 09:41:58 -07001793 return 0;
1794}
1795
Jens Axboe8c838782019-03-12 15:48:16 -06001796static void io_poll_complete(struct io_ring_ctx *ctx, struct io_kiocb *req,
1797 __poll_t mask)
Jens Axboe221c5eb2019-01-17 09:41:58 -07001798{
Jens Axboe8c838782019-03-12 15:48:16 -06001799 req->poll.done = true;
Jens Axboec71ffb62019-05-13 20:58:29 -06001800 io_cqring_fill_event(ctx, req->user_data, mangle_poll(mask));
Jens Axboe8c838782019-03-12 15:48:16 -06001801 io_commit_cqring(ctx);
Jens Axboe221c5eb2019-01-17 09:41:58 -07001802}
1803
Jens Axboe561fb042019-10-24 07:25:42 -06001804static void io_poll_complete_work(struct io_wq_work **workptr)
Jens Axboe221c5eb2019-01-17 09:41:58 -07001805{
Jens Axboe561fb042019-10-24 07:25:42 -06001806 struct io_wq_work *work = *workptr;
Jens Axboe221c5eb2019-01-17 09:41:58 -07001807 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1808 struct io_poll_iocb *poll = &req->poll;
1809 struct poll_table_struct pt = { ._key = poll->events };
1810 struct io_ring_ctx *ctx = req->ctx;
Jens Axboe89723d02019-11-05 15:32:58 -07001811 struct io_kiocb *nxt = NULL;
Jens Axboe221c5eb2019-01-17 09:41:58 -07001812 __poll_t mask = 0;
1813
Jens Axboe561fb042019-10-24 07:25:42 -06001814 if (work->flags & IO_WQ_WORK_CANCEL)
1815 WRITE_ONCE(poll->canceled, true);
1816
Jens Axboe221c5eb2019-01-17 09:41:58 -07001817 if (!READ_ONCE(poll->canceled))
1818 mask = vfs_poll(poll->file, &pt) & poll->events;
1819
1820 /*
1821 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1822 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1823 * synchronize with them. In the cancellation case the list_del_init
1824 * itself is not actually needed, but harmless so we keep it in to
1825 * avoid further branches in the fast path.
1826 */
1827 spin_lock_irq(&ctx->completion_lock);
1828 if (!mask && !READ_ONCE(poll->canceled)) {
1829 add_wait_queue(poll->head, &poll->wait);
1830 spin_unlock_irq(&ctx->completion_lock);
1831 return;
1832 }
1833 list_del_init(&req->list);
Jens Axboe8c838782019-03-12 15:48:16 -06001834 io_poll_complete(ctx, req, mask);
Jens Axboe221c5eb2019-01-17 09:41:58 -07001835 spin_unlock_irq(&ctx->completion_lock);
1836
Jens Axboe8c838782019-03-12 15:48:16 -06001837 io_cqring_ev_posted(ctx);
Jens Axboe89723d02019-11-05 15:32:58 -07001838
1839 io_put_req(req, &nxt);
1840 if (nxt)
1841 *workptr = &nxt->work;
Jens Axboe221c5eb2019-01-17 09:41:58 -07001842}
1843
1844static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1845 void *key)
1846{
1847 struct io_poll_iocb *poll = container_of(wait, struct io_poll_iocb,
1848 wait);
1849 struct io_kiocb *req = container_of(poll, struct io_kiocb, poll);
1850 struct io_ring_ctx *ctx = req->ctx;
1851 __poll_t mask = key_to_poll(key);
Jens Axboe8c838782019-03-12 15:48:16 -06001852 unsigned long flags;
Jens Axboe221c5eb2019-01-17 09:41:58 -07001853
1854 /* for instances that support it check for an event match first: */
Jens Axboe8c838782019-03-12 15:48:16 -06001855 if (mask && !(mask & poll->events))
1856 return 0;
Jens Axboe221c5eb2019-01-17 09:41:58 -07001857
1858 list_del_init(&poll->wait.entry);
Jens Axboe8c838782019-03-12 15:48:16 -06001859
1860 if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) {
1861 list_del(&req->list);
1862 io_poll_complete(ctx, req, mask);
1863 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1864
1865 io_cqring_ev_posted(ctx);
Jens Axboeba816ad2019-09-28 11:36:45 -06001866 io_put_req(req, NULL);
Jens Axboe8c838782019-03-12 15:48:16 -06001867 } else {
Jens Axboe18d9be12019-09-10 09:13:05 -06001868 io_queue_async_work(ctx, req);
Jens Axboe8c838782019-03-12 15:48:16 -06001869 }
1870
Jens Axboe221c5eb2019-01-17 09:41:58 -07001871 return 1;
1872}
1873
1874struct io_poll_table {
1875 struct poll_table_struct pt;
1876 struct io_kiocb *req;
1877 int error;
1878};
1879
1880static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1881 struct poll_table_struct *p)
1882{
1883 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
1884
1885 if (unlikely(pt->req->poll.head)) {
1886 pt->error = -EINVAL;
1887 return;
1888 }
1889
1890 pt->error = 0;
1891 pt->req->poll.head = head;
1892 add_wait_queue(head, &pt->req->poll.wait);
1893}
1894
Jens Axboe89723d02019-11-05 15:32:58 -07001895static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1896 struct io_kiocb **nxt)
Jens Axboe221c5eb2019-01-17 09:41:58 -07001897{
1898 struct io_poll_iocb *poll = &req->poll;
1899 struct io_ring_ctx *ctx = req->ctx;
1900 struct io_poll_table ipt;
Jens Axboe8c838782019-03-12 15:48:16 -06001901 bool cancel = false;
Jens Axboe221c5eb2019-01-17 09:41:58 -07001902 __poll_t mask;
1903 u16 events;
Jens Axboe221c5eb2019-01-17 09:41:58 -07001904
1905 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1906 return -EINVAL;
1907 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
1908 return -EINVAL;
Jens Axboe09bb8392019-03-13 12:39:28 -06001909 if (!poll->file)
1910 return -EBADF;
Jens Axboe221c5eb2019-01-17 09:41:58 -07001911
Jens Axboe6cc47d12019-09-18 11:18:23 -06001912 req->submit.sqe = NULL;
Jens Axboe561fb042019-10-24 07:25:42 -06001913 INIT_IO_WORK(&req->work, io_poll_complete_work);
Jens Axboe221c5eb2019-01-17 09:41:58 -07001914 events = READ_ONCE(sqe->poll_events);
1915 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
1916
Jens Axboe221c5eb2019-01-17 09:41:58 -07001917 poll->head = NULL;
Jens Axboe8c838782019-03-12 15:48:16 -06001918 poll->done = false;
Jens Axboe221c5eb2019-01-17 09:41:58 -07001919 poll->canceled = false;
1920
1921 ipt.pt._qproc = io_poll_queue_proc;
1922 ipt.pt._key = poll->events;
1923 ipt.req = req;
1924 ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1925
1926 /* initialized the list so that we can do list_empty checks */
1927 INIT_LIST_HEAD(&poll->wait.entry);
1928 init_waitqueue_func_entry(&poll->wait, io_poll_wake);
1929
Jens Axboe36703242019-07-25 10:20:18 -06001930 INIT_LIST_HEAD(&req->list);
1931
Jens Axboe221c5eb2019-01-17 09:41:58 -07001932 mask = vfs_poll(poll->file, &ipt.pt) & poll->events;
Jens Axboe221c5eb2019-01-17 09:41:58 -07001933
1934 spin_lock_irq(&ctx->completion_lock);
Jens Axboe8c838782019-03-12 15:48:16 -06001935 if (likely(poll->head)) {
1936 spin_lock(&poll->head->lock);
1937 if (unlikely(list_empty(&poll->wait.entry))) {
1938 if (ipt.error)
1939 cancel = true;
1940 ipt.error = 0;
1941 mask = 0;
1942 }
1943 if (mask || ipt.error)
1944 list_del_init(&poll->wait.entry);
1945 else if (cancel)
1946 WRITE_ONCE(poll->canceled, true);
1947 else if (!poll->done) /* actually waiting for an event */
1948 list_add_tail(&req->list, &ctx->cancel_list);
1949 spin_unlock(&poll->head->lock);
Jens Axboe221c5eb2019-01-17 09:41:58 -07001950 }
Jens Axboe8c838782019-03-12 15:48:16 -06001951 if (mask) { /* no async, we'd stolen it */
Jens Axboe8c838782019-03-12 15:48:16 -06001952 ipt.error = 0;
1953 io_poll_complete(ctx, req, mask);
1954 }
Jens Axboe221c5eb2019-01-17 09:41:58 -07001955 spin_unlock_irq(&ctx->completion_lock);
1956
Jens Axboe8c838782019-03-12 15:48:16 -06001957 if (mask) {
1958 io_cqring_ev_posted(ctx);
Jens Axboe89723d02019-11-05 15:32:58 -07001959 io_put_req(req, nxt);
Jens Axboe221c5eb2019-01-17 09:41:58 -07001960 }
Jens Axboe8c838782019-03-12 15:48:16 -06001961 return ipt.error;
Jens Axboe221c5eb2019-01-17 09:41:58 -07001962}
1963
Jens Axboe5262f562019-09-17 12:26:57 -06001964static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
1965{
1966 struct io_ring_ctx *ctx;
Jens Axboe11365042019-10-16 09:08:32 -06001967 struct io_kiocb *req;
Jens Axboe5262f562019-09-17 12:26:57 -06001968 unsigned long flags;
1969
1970 req = container_of(timer, struct io_kiocb, timeout.timer);
1971 ctx = req->ctx;
1972 atomic_inc(&ctx->cq_timeouts);
1973
1974 spin_lock_irqsave(&ctx->completion_lock, flags);
zhangyi (F)ef036812019-10-23 15:10:08 +08001975 /*
Jens Axboe11365042019-10-16 09:08:32 -06001976 * We could be racing with timeout deletion. If the list is empty,
1977 * then timeout lookup already found it and will be handling it.
zhangyi (F)ef036812019-10-23 15:10:08 +08001978 */
Jens Axboe842f9612019-10-29 12:34:10 -06001979 if (!list_empty(&req->list)) {
Jens Axboe11365042019-10-16 09:08:32 -06001980 struct io_kiocb *prev;
Jens Axboe5262f562019-09-17 12:26:57 -06001981
Jens Axboe11365042019-10-16 09:08:32 -06001982 /*
1983 * Adjust the reqs sequence before the current one because it
1984 * will consume a slot in the cq_ring and the the cq_tail
1985 * pointer will be increased, otherwise other timeout reqs may
1986 * return in advance without waiting for enough wait_nr.
1987 */
1988 prev = req;
1989 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
1990 prev->sequence++;
Jens Axboe11365042019-10-16 09:08:32 -06001991 list_del_init(&req->list);
Jens Axboe11365042019-10-16 09:08:32 -06001992 }
Jens Axboe842f9612019-10-29 12:34:10 -06001993
1994 io_cqring_fill_event(ctx, req->user_data, -ETIME);
1995 io_commit_cqring(ctx);
Jens Axboe5262f562019-09-17 12:26:57 -06001996 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1997
Jens Axboe842f9612019-10-29 12:34:10 -06001998 io_cqring_ev_posted(ctx);
Jens Axboef1f40852019-11-05 20:33:16 -07001999 if (req->flags & REQ_F_LINK)
2000 req->flags |= REQ_F_FAIL_LINK;
Jens Axboe842f9612019-10-29 12:34:10 -06002001 io_put_req(req, NULL);
Jens Axboe11365042019-10-16 09:08:32 -06002002 return HRTIMER_NORESTART;
2003}
2004
2005/*
2006 * Remove or update an existing timeout command
2007 */
2008static int io_timeout_remove(struct io_kiocb *req,
2009 const struct io_uring_sqe *sqe)
2010{
2011 struct io_ring_ctx *ctx = req->ctx;
2012 struct io_kiocb *treq;
2013 int ret = -ENOENT;
2014 __u64 user_data;
2015 unsigned flags;
2016
2017 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2018 return -EINVAL;
2019 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
2020 return -EINVAL;
2021 flags = READ_ONCE(sqe->timeout_flags);
2022 if (flags)
2023 return -EINVAL;
2024
2025 user_data = READ_ONCE(sqe->addr);
2026 spin_lock_irq(&ctx->completion_lock);
2027 list_for_each_entry(treq, &ctx->timeout_list, list) {
2028 if (user_data == treq->user_data) {
2029 list_del_init(&treq->list);
2030 ret = 0;
2031 break;
2032 }
2033 }
2034
2035 /* didn't find timeout */
2036 if (ret) {
2037fill_ev:
2038 io_cqring_fill_event(ctx, req->user_data, ret);
2039 io_commit_cqring(ctx);
2040 spin_unlock_irq(&ctx->completion_lock);
2041 io_cqring_ev_posted(ctx);
Jens Axboef1f40852019-11-05 20:33:16 -07002042 if (req->flags & REQ_F_LINK)
2043 req->flags |= REQ_F_FAIL_LINK;
Jens Axboe11365042019-10-16 09:08:32 -06002044 io_put_req(req, NULL);
2045 return 0;
2046 }
2047
2048 ret = hrtimer_try_to_cancel(&treq->timeout.timer);
2049 if (ret == -1) {
2050 ret = -EBUSY;
2051 goto fill_ev;
2052 }
2053
2054 io_cqring_fill_event(ctx, req->user_data, 0);
2055 io_cqring_fill_event(ctx, treq->user_data, -ECANCELED);
2056 io_commit_cqring(ctx);
2057 spin_unlock_irq(&ctx->completion_lock);
Jens Axboe5262f562019-09-17 12:26:57 -06002058 io_cqring_ev_posted(ctx);
2059
Jens Axboe11365042019-10-16 09:08:32 -06002060 io_put_req(treq, NULL);
Jens Axboeba816ad2019-09-28 11:36:45 -06002061 io_put_req(req, NULL);
Jens Axboe11365042019-10-16 09:08:32 -06002062 return 0;
Jens Axboe5262f562019-09-17 12:26:57 -06002063}
2064
2065static int io_timeout(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2066{
yangerkun5da0fb12019-10-15 21:59:29 +08002067 unsigned count;
Jens Axboe5262f562019-09-17 12:26:57 -06002068 struct io_ring_ctx *ctx = req->ctx;
2069 struct list_head *entry;
Jens Axboea41525a2019-10-15 16:48:15 -06002070 enum hrtimer_mode mode;
Arnd Bergmannbdf20072019-10-01 09:53:29 -06002071 struct timespec64 ts;
zhangyi (F)a1f58ba2019-10-23 15:10:09 +08002072 unsigned span = 0;
Jens Axboea41525a2019-10-15 16:48:15 -06002073 unsigned flags;
Jens Axboe5262f562019-09-17 12:26:57 -06002074
2075 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2076 return -EINVAL;
Jens Axboea41525a2019-10-15 16:48:15 -06002077 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len != 1)
2078 return -EINVAL;
2079 flags = READ_ONCE(sqe->timeout_flags);
2080 if (flags & ~IORING_TIMEOUT_ABS)
Jens Axboe5262f562019-09-17 12:26:57 -06002081 return -EINVAL;
Arnd Bergmannbdf20072019-10-01 09:53:29 -06002082
2083 if (get_timespec64(&ts, u64_to_user_ptr(sqe->addr)))
Jens Axboe5262f562019-09-17 12:26:57 -06002084 return -EFAULT;
2085
Jens Axboe11365042019-10-16 09:08:32 -06002086 if (flags & IORING_TIMEOUT_ABS)
2087 mode = HRTIMER_MODE_ABS;
2088 else
2089 mode = HRTIMER_MODE_REL;
2090
2091 hrtimer_init(&req->timeout.timer, CLOCK_MONOTONIC, mode);
2092
Jens Axboe5262f562019-09-17 12:26:57 -06002093 /*
2094 * sqe->off holds how many events that need to occur for this
2095 * timeout event to be satisfied.
2096 */
2097 count = READ_ONCE(sqe->off);
2098 if (!count)
2099 count = 1;
2100
2101 req->sequence = ctx->cached_sq_head + count - 1;
yangerkun5da0fb12019-10-15 21:59:29 +08002102 /* reuse it to store the count */
2103 req->submit.sequence = count;
Jens Axboe5262f562019-09-17 12:26:57 -06002104 req->flags |= REQ_F_TIMEOUT;
2105
2106 /*
2107 * Insertion sort, ensuring the first entry in the list is always
2108 * the one we need first.
2109 */
Jens Axboe5262f562019-09-17 12:26:57 -06002110 spin_lock_irq(&ctx->completion_lock);
2111 list_for_each_prev(entry, &ctx->timeout_list) {
2112 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
yangerkun5da0fb12019-10-15 21:59:29 +08002113 unsigned nxt_sq_head;
2114 long long tmp, tmp_nxt;
Jens Axboe5262f562019-09-17 12:26:57 -06002115
yangerkun5da0fb12019-10-15 21:59:29 +08002116 /*
2117 * Since cached_sq_head + count - 1 can overflow, use type long
2118 * long to store it.
2119 */
2120 tmp = (long long)ctx->cached_sq_head + count - 1;
2121 nxt_sq_head = nxt->sequence - nxt->submit.sequence + 1;
2122 tmp_nxt = (long long)nxt_sq_head + nxt->submit.sequence - 1;
2123
2124 /*
2125 * cached_sq_head may overflow, and it will never overflow twice
2126 * once there is some timeout req still be valid.
2127 */
2128 if (ctx->cached_sq_head < nxt_sq_head)
yangerkun8b07a652019-10-17 12:12:35 +08002129 tmp += UINT_MAX;
yangerkun5da0fb12019-10-15 21:59:29 +08002130
zhangyi (F)a1f58ba2019-10-23 15:10:09 +08002131 if (tmp > tmp_nxt)
Jens Axboe5262f562019-09-17 12:26:57 -06002132 break;
zhangyi (F)a1f58ba2019-10-23 15:10:09 +08002133
2134 /*
2135 * Sequence of reqs after the insert one and itself should
2136 * be adjusted because each timeout req consumes a slot.
2137 */
2138 span++;
2139 nxt->sequence++;
Jens Axboe5262f562019-09-17 12:26:57 -06002140 }
zhangyi (F)a1f58ba2019-10-23 15:10:09 +08002141 req->sequence -= span;
Jens Axboe5262f562019-09-17 12:26:57 -06002142 list_add(&req->list, entry);
Jens Axboe5262f562019-09-17 12:26:57 -06002143 req->timeout.timer.function = io_timeout_fn;
Jens Axboea41525a2019-10-15 16:48:15 -06002144 hrtimer_start(&req->timeout.timer, timespec64_to_ktime(ts), mode);
Jens Axboe842f9612019-10-29 12:34:10 -06002145 spin_unlock_irq(&ctx->completion_lock);
Jens Axboe5262f562019-09-17 12:26:57 -06002146 return 0;
2147}
2148
Jens Axboe62755e32019-10-28 21:49:21 -06002149static bool io_cancel_cb(struct io_wq_work *work, void *data)
2150{
2151 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2152
2153 return req->user_data == (unsigned long) data;
2154}
2155
2156static int io_async_cancel(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2157 struct io_kiocb **nxt)
2158{
2159 struct io_ring_ctx *ctx = req->ctx;
2160 enum io_wq_cancel cancel_ret;
2161 void *sqe_addr;
2162 int ret = 0;
2163
2164 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2165 return -EINVAL;
2166 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
2167 sqe->cancel_flags)
2168 return -EINVAL;
2169
2170 sqe_addr = (void *) (unsigned long) READ_ONCE(sqe->addr);
2171 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
2172 switch (cancel_ret) {
2173 case IO_WQ_CANCEL_OK:
2174 ret = 0;
2175 break;
2176 case IO_WQ_CANCEL_RUNNING:
2177 ret = -EALREADY;
2178 break;
2179 case IO_WQ_CANCEL_NOTFOUND:
2180 ret = -ENOENT;
2181 break;
2182 }
2183
2184 if (ret < 0 && (req->flags & REQ_F_LINK))
2185 req->flags |= REQ_F_FAIL_LINK;
2186 io_cqring_add_event(req->ctx, sqe->user_data, ret);
2187 io_put_req(req, nxt);
2188 return 0;
2189}
2190
Jens Axboede0617e2019-04-06 21:51:27 -06002191static int io_req_defer(struct io_ring_ctx *ctx, struct io_kiocb *req,
2192 const struct io_uring_sqe *sqe)
2193{
2194 struct io_uring_sqe *sqe_copy;
2195
2196 if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list))
2197 return 0;
2198
2199 sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL);
2200 if (!sqe_copy)
2201 return -EAGAIN;
2202
2203 spin_lock_irq(&ctx->completion_lock);
2204 if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list)) {
2205 spin_unlock_irq(&ctx->completion_lock);
2206 kfree(sqe_copy);
2207 return 0;
2208 }
2209
2210 memcpy(sqe_copy, sqe, sizeof(*sqe_copy));
2211 req->submit.sqe = sqe_copy;
2212
Dmitrii Dolgovc826bd72019-10-15 19:02:01 +02002213 trace_io_uring_defer(ctx, req, false);
Jens Axboede0617e2019-04-06 21:51:27 -06002214 list_add_tail(&req->list, &ctx->defer_list);
2215 spin_unlock_irq(&ctx->completion_lock);
2216 return -EIOCBQUEUED;
2217}
2218
Jens Axboe2b188cc2019-01-07 10:46:33 -07002219static int __io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
Jens Axboeba816ad2019-09-28 11:36:45 -06002220 const struct sqe_submit *s, struct io_kiocb **nxt,
2221 bool force_nonblock)
Jens Axboe2b188cc2019-01-07 10:46:33 -07002222{
Jens Axboee0c5c572019-03-12 10:18:47 -06002223 int ret, opcode;
Jens Axboe2b188cc2019-01-07 10:46:33 -07002224
Jens Axboe9e645e112019-05-10 16:07:28 -06002225 req->user_data = READ_ONCE(s->sqe->user_data);
2226
Jens Axboe2b188cc2019-01-07 10:46:33 -07002227 opcode = READ_ONCE(s->sqe->opcode);
2228 switch (opcode) {
2229 case IORING_OP_NOP:
2230 ret = io_nop(req, req->user_data);
2231 break;
2232 case IORING_OP_READV:
Jens Axboeedafcce2019-01-09 09:16:05 -07002233 if (unlikely(s->sqe->buf_index))
2234 return -EINVAL;
Jens Axboeba816ad2019-09-28 11:36:45 -06002235 ret = io_read(req, s, nxt, force_nonblock);
Jens Axboe2b188cc2019-01-07 10:46:33 -07002236 break;
2237 case IORING_OP_WRITEV:
Jens Axboeedafcce2019-01-09 09:16:05 -07002238 if (unlikely(s->sqe->buf_index))
2239 return -EINVAL;
Jens Axboeba816ad2019-09-28 11:36:45 -06002240 ret = io_write(req, s, nxt, force_nonblock);
Jens Axboeedafcce2019-01-09 09:16:05 -07002241 break;
2242 case IORING_OP_READ_FIXED:
Jens Axboeba816ad2019-09-28 11:36:45 -06002243 ret = io_read(req, s, nxt, force_nonblock);
Jens Axboeedafcce2019-01-09 09:16:05 -07002244 break;
2245 case IORING_OP_WRITE_FIXED:
Jens Axboeba816ad2019-09-28 11:36:45 -06002246 ret = io_write(req, s, nxt, force_nonblock);
Jens Axboe2b188cc2019-01-07 10:46:33 -07002247 break;
Christoph Hellwigc992fe22019-01-11 09:43:02 -07002248 case IORING_OP_FSYNC:
Jens Axboeba816ad2019-09-28 11:36:45 -06002249 ret = io_fsync(req, s->sqe, nxt, force_nonblock);
Christoph Hellwigc992fe22019-01-11 09:43:02 -07002250 break;
Jens Axboe221c5eb2019-01-17 09:41:58 -07002251 case IORING_OP_POLL_ADD:
Jens Axboe89723d02019-11-05 15:32:58 -07002252 ret = io_poll_add(req, s->sqe, nxt);
Jens Axboe221c5eb2019-01-17 09:41:58 -07002253 break;
2254 case IORING_OP_POLL_REMOVE:
2255 ret = io_poll_remove(req, s->sqe);
2256 break;
Jens Axboe5d17b4a2019-04-09 14:56:44 -06002257 case IORING_OP_SYNC_FILE_RANGE:
Jens Axboeba816ad2019-09-28 11:36:45 -06002258 ret = io_sync_file_range(req, s->sqe, nxt, force_nonblock);
Jens Axboe5d17b4a2019-04-09 14:56:44 -06002259 break;
Jens Axboe0fa03c62019-04-19 13:34:07 -06002260 case IORING_OP_SENDMSG:
Jens Axboeba816ad2019-09-28 11:36:45 -06002261 ret = io_sendmsg(req, s->sqe, nxt, force_nonblock);
Jens Axboe0fa03c62019-04-19 13:34:07 -06002262 break;
Jens Axboeaa1fa282019-04-19 13:38:09 -06002263 case IORING_OP_RECVMSG:
Jens Axboeba816ad2019-09-28 11:36:45 -06002264 ret = io_recvmsg(req, s->sqe, nxt, force_nonblock);
Jens Axboeaa1fa282019-04-19 13:38:09 -06002265 break;
Jens Axboe5262f562019-09-17 12:26:57 -06002266 case IORING_OP_TIMEOUT:
2267 ret = io_timeout(req, s->sqe);
2268 break;
Jens Axboe11365042019-10-16 09:08:32 -06002269 case IORING_OP_TIMEOUT_REMOVE:
2270 ret = io_timeout_remove(req, s->sqe);
2271 break;
Jens Axboe17f2fe32019-10-17 14:42:58 -06002272 case IORING_OP_ACCEPT:
2273 ret = io_accept(req, s->sqe, nxt, force_nonblock);
2274 break;
Jens Axboe62755e32019-10-28 21:49:21 -06002275 case IORING_OP_ASYNC_CANCEL:
2276 ret = io_async_cancel(req, s->sqe, nxt);
2277 break;
Jens Axboe2b188cc2019-01-07 10:46:33 -07002278 default:
2279 ret = -EINVAL;
2280 break;
2281 }
2282
Jens Axboedef596e2019-01-09 08:59:42 -07002283 if (ret)
2284 return ret;
2285
2286 if (ctx->flags & IORING_SETUP_IOPOLL) {
Jens Axboe9e645e112019-05-10 16:07:28 -06002287 if (req->result == -EAGAIN)
Jens Axboedef596e2019-01-09 08:59:42 -07002288 return -EAGAIN;
2289
2290 /* workqueue context doesn't hold uring_lock, grab it now */
Jackie Liuba5290c2019-10-09 09:19:59 +08002291 if (s->in_async)
Jens Axboedef596e2019-01-09 08:59:42 -07002292 mutex_lock(&ctx->uring_lock);
2293 io_iopoll_req_issued(req);
Jackie Liuba5290c2019-10-09 09:19:59 +08002294 if (s->in_async)
Jens Axboedef596e2019-01-09 08:59:42 -07002295 mutex_unlock(&ctx->uring_lock);
2296 }
2297
2298 return 0;
Jens Axboe2b188cc2019-01-07 10:46:33 -07002299}
2300
Jens Axboe561fb042019-10-24 07:25:42 -06002301static void io_wq_submit_work(struct io_wq_work **workptr)
Jens Axboe31b51512019-01-18 22:56:34 -07002302{
Jens Axboe561fb042019-10-24 07:25:42 -06002303 struct io_wq_work *work = *workptr;
Jens Axboe2b188cc2019-01-07 10:46:33 -07002304 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
Jens Axboe2b188cc2019-01-07 10:46:33 -07002305 struct io_ring_ctx *ctx = req->ctx;
Jens Axboe561fb042019-10-24 07:25:42 -06002306 struct sqe_submit *s = &req->submit;
2307 const struct io_uring_sqe *sqe = s->sqe;
2308 struct io_kiocb *nxt = NULL;
2309 int ret = 0;
Jens Axboe2b188cc2019-01-07 10:46:33 -07002310
Jens Axboe561fb042019-10-24 07:25:42 -06002311 /* Ensure we clear previously set non-block flag */
2312 req->rw.ki_flags &= ~IOCB_NOWAIT;
Jens Axboe2b188cc2019-01-07 10:46:33 -07002313
Jens Axboe561fb042019-10-24 07:25:42 -06002314 if (work->flags & IO_WQ_WORK_CANCEL)
2315 ret = -ECANCELED;
Jens Axboe31b51512019-01-18 22:56:34 -07002316
Jens Axboe561fb042019-10-24 07:25:42 -06002317 if (!ret) {
2318 s->has_user = (work->flags & IO_WQ_WORK_HAS_MM) != 0;
2319 s->in_async = true;
2320 do {
2321 ret = __io_submit_sqe(ctx, req, s, &nxt, false);
2322 /*
2323 * We can get EAGAIN for polled IO even though we're
2324 * forcing a sync submission from here, since we can't
2325 * wait for request slots on the block side.
2326 */
2327 if (ret != -EAGAIN)
2328 break;
2329 cond_resched();
2330 } while (1);
2331 }
Jens Axboe31b51512019-01-18 22:56:34 -07002332
Jens Axboe561fb042019-10-24 07:25:42 -06002333 /* drop submission reference */
2334 io_put_req(req, NULL);
Jens Axboe817869d2019-04-30 14:44:05 -06002335
Jens Axboe561fb042019-10-24 07:25:42 -06002336 if (ret) {
Jens Axboef1f40852019-11-05 20:33:16 -07002337 if (req->flags & REQ_F_LINK)
2338 req->flags |= REQ_F_FAIL_LINK;
Jens Axboe561fb042019-10-24 07:25:42 -06002339 io_cqring_add_event(ctx, sqe->user_data, ret);
Jens Axboeba816ad2019-09-28 11:36:45 -06002340 io_put_req(req, NULL);
Jens Axboeedafcce2019-01-09 09:16:05 -07002341 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07002342
Jens Axboe561fb042019-10-24 07:25:42 -06002343 /* async context always use a copy of the sqe */
2344 kfree(sqe);
2345
2346 /* if a dependent link is ready, pass it back */
2347 if (!ret && nxt) {
2348 io_prep_async_work(nxt);
2349 *workptr = &nxt->work;
Jens Axboeedafcce2019-01-09 09:16:05 -07002350 }
Jens Axboe31b51512019-01-18 22:56:34 -07002351}
Jens Axboe2b188cc2019-01-07 10:46:33 -07002352
Jens Axboe09bb8392019-03-13 12:39:28 -06002353static bool io_op_needs_file(const struct io_uring_sqe *sqe)
2354{
2355 int op = READ_ONCE(sqe->opcode);
2356
2357 switch (op) {
2358 case IORING_OP_NOP:
2359 case IORING_OP_POLL_REMOVE:
2360 return false;
2361 default:
2362 return true;
2363 }
2364}
2365
Jens Axboe65e19f52019-10-26 07:20:21 -06002366static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
2367 int index)
2368{
2369 struct fixed_file_table *table;
2370
2371 table = &ctx->file_table[index >> IORING_FILE_TABLE_SHIFT];
2372 return table->files[index & IORING_FILE_TABLE_MASK];
2373}
2374
Jens Axboe09bb8392019-03-13 12:39:28 -06002375static int io_req_set_file(struct io_ring_ctx *ctx, const struct sqe_submit *s,
2376 struct io_submit_state *state, struct io_kiocb *req)
2377{
2378 unsigned flags;
2379 int fd;
2380
2381 flags = READ_ONCE(s->sqe->flags);
2382 fd = READ_ONCE(s->sqe->fd);
2383
Jackie Liu4fe2c962019-09-09 20:50:40 +08002384 if (flags & IOSQE_IO_DRAIN)
Jens Axboede0617e2019-04-06 21:51:27 -06002385 req->flags |= REQ_F_IO_DRAIN;
Jackie Liu4fe2c962019-09-09 20:50:40 +08002386 /*
2387 * All io need record the previous position, if LINK vs DARIN,
2388 * it can be used to mark the position of the first IO in the
2389 * link list.
2390 */
2391 req->sequence = s->sequence;
Jens Axboede0617e2019-04-06 21:51:27 -06002392
Jens Axboe60c112b2019-06-21 10:20:18 -06002393 if (!io_op_needs_file(s->sqe))
Jens Axboe09bb8392019-03-13 12:39:28 -06002394 return 0;
Jens Axboe09bb8392019-03-13 12:39:28 -06002395
2396 if (flags & IOSQE_FIXED_FILE) {
Jens Axboe65e19f52019-10-26 07:20:21 -06002397 if (unlikely(!ctx->file_table ||
Jens Axboe09bb8392019-03-13 12:39:28 -06002398 (unsigned) fd >= ctx->nr_user_files))
2399 return -EBADF;
Jens Axboeb7620122019-10-26 07:22:55 -06002400 fd = array_index_nospec(fd, ctx->nr_user_files);
Jens Axboe65e19f52019-10-26 07:20:21 -06002401 req->file = io_file_from_index(ctx, fd);
2402 if (!req->file)
Jens Axboe08a45172019-10-03 08:11:03 -06002403 return -EBADF;
Jens Axboe09bb8392019-03-13 12:39:28 -06002404 req->flags |= REQ_F_FIXED_FILE;
2405 } else {
2406 if (s->needs_fixed_file)
2407 return -EBADF;
Dmitrii Dolgovc826bd72019-10-15 19:02:01 +02002408 trace_io_uring_file_get(ctx, fd);
Jens Axboe09bb8392019-03-13 12:39:28 -06002409 req->file = io_file_get(state, fd);
2410 if (unlikely(!req->file))
2411 return -EBADF;
2412 }
2413
2414 return 0;
2415}
2416
Jens Axboefcb323c2019-10-24 12:39:47 -06002417static int io_grab_files(struct io_ring_ctx *ctx, struct io_kiocb *req)
2418{
2419 int ret = -EBADF;
2420
2421 rcu_read_lock();
2422 spin_lock_irq(&ctx->inflight_lock);
2423 /*
2424 * We use the f_ops->flush() handler to ensure that we can flush
2425 * out work accessing these files if the fd is closed. Check if
2426 * the fd has changed since we started down this path, and disallow
2427 * this operation if it has.
2428 */
2429 if (fcheck(req->submit.ring_fd) == req->submit.ring_file) {
2430 list_add(&req->inflight_entry, &ctx->inflight_list);
2431 req->flags |= REQ_F_INFLIGHT;
2432 req->work.files = current->files;
2433 ret = 0;
2434 }
2435 spin_unlock_irq(&ctx->inflight_lock);
2436 rcu_read_unlock();
2437
2438 return ret;
2439}
2440
Jackie Liu4fe2c962019-09-09 20:50:40 +08002441static int __io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
Jens Axboebc808bc2019-10-22 13:14:37 -06002442 struct sqe_submit *s)
Jens Axboe2b188cc2019-01-07 10:46:33 -07002443{
Jens Axboee0c5c572019-03-12 10:18:47 -06002444 int ret;
Jens Axboe2b188cc2019-01-07 10:46:33 -07002445
Jens Axboeba816ad2019-09-28 11:36:45 -06002446 ret = __io_submit_sqe(ctx, req, s, NULL, true);
Jens Axboe491381ce2019-10-17 09:20:46 -06002447
2448 /*
2449 * We async punt it if the file wasn't marked NOWAIT, or if the file
2450 * doesn't support non-blocking read/write attempts
2451 */
2452 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
2453 (req->flags & REQ_F_MUST_PUNT))) {
Jens Axboe2b188cc2019-01-07 10:46:33 -07002454 struct io_uring_sqe *sqe_copy;
2455
Jackie Liu954dab12019-09-18 10:37:52 +08002456 sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL);
Jens Axboe2b188cc2019-01-07 10:46:33 -07002457 if (sqe_copy) {
Jens Axboe2b188cc2019-01-07 10:46:33 -07002458 s->sqe = sqe_copy;
Jens Axboe2b188cc2019-01-07 10:46:33 -07002459 memcpy(&req->submit, s, sizeof(*s));
Jens Axboefcb323c2019-10-24 12:39:47 -06002460 if (req->work.flags & IO_WQ_WORK_NEEDS_FILES) {
2461 ret = io_grab_files(ctx, req);
2462 if (ret) {
2463 kfree(sqe_copy);
2464 goto err;
2465 }
2466 }
Jens Axboee65ef562019-03-12 10:16:44 -06002467
2468 /*
2469 * Queued up for async execution, worker will release
Jens Axboe9e645e112019-05-10 16:07:28 -06002470 * submit reference when the iocb is actually submitted.
Jens Axboee65ef562019-03-12 10:16:44 -06002471 */
Jens Axboefcb323c2019-10-24 12:39:47 -06002472 io_queue_async_work(ctx, req);
Jens Axboee65ef562019-03-12 10:16:44 -06002473 return 0;
Jens Axboe2b188cc2019-01-07 10:46:33 -07002474 }
2475 }
Jens Axboee65ef562019-03-12 10:16:44 -06002476
2477 /* drop submission reference */
Jens Axboefcb323c2019-10-24 12:39:47 -06002478err:
Jens Axboeba816ad2019-09-28 11:36:45 -06002479 io_put_req(req, NULL);
Jens Axboee65ef562019-03-12 10:16:44 -06002480
2481 /* and drop final reference, if we failed */
Jens Axboe9e645e112019-05-10 16:07:28 -06002482 if (ret) {
2483 io_cqring_add_event(ctx, req->user_data, ret);
2484 if (req->flags & REQ_F_LINK)
2485 req->flags |= REQ_F_FAIL_LINK;
Jens Axboeba816ad2019-09-28 11:36:45 -06002486 io_put_req(req, NULL);
Jens Axboe9e645e112019-05-10 16:07:28 -06002487 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07002488
2489 return ret;
2490}
2491
Jackie Liu4fe2c962019-09-09 20:50:40 +08002492static int io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
Jens Axboebc808bc2019-10-22 13:14:37 -06002493 struct sqe_submit *s)
Jackie Liu4fe2c962019-09-09 20:50:40 +08002494{
2495 int ret;
2496
2497 ret = io_req_defer(ctx, req, s->sqe);
2498 if (ret) {
2499 if (ret != -EIOCBQUEUED) {
Jens Axboeba816ad2019-09-28 11:36:45 -06002500 io_free_req(req, NULL);
Jackie Liu4fe2c962019-09-09 20:50:40 +08002501 io_cqring_add_event(ctx, s->sqe->user_data, ret);
2502 }
2503 return 0;
2504 }
2505
Jens Axboebc808bc2019-10-22 13:14:37 -06002506 return __io_queue_sqe(ctx, req, s);
Jackie Liu4fe2c962019-09-09 20:50:40 +08002507}
2508
2509static int io_queue_link_head(struct io_ring_ctx *ctx, struct io_kiocb *req,
Jens Axboebc808bc2019-10-22 13:14:37 -06002510 struct sqe_submit *s, struct io_kiocb *shadow)
Jackie Liu4fe2c962019-09-09 20:50:40 +08002511{
2512 int ret;
2513 int need_submit = false;
2514
2515 if (!shadow)
Jens Axboebc808bc2019-10-22 13:14:37 -06002516 return io_queue_sqe(ctx, req, s);
Jackie Liu4fe2c962019-09-09 20:50:40 +08002517
2518 /*
2519 * Mark the first IO in link list as DRAIN, let all the following
2520 * IOs enter the defer list. all IO needs to be completed before link
2521 * list.
2522 */
2523 req->flags |= REQ_F_IO_DRAIN;
2524 ret = io_req_defer(ctx, req, s->sqe);
2525 if (ret) {
2526 if (ret != -EIOCBQUEUED) {
Jens Axboeba816ad2019-09-28 11:36:45 -06002527 io_free_req(req, NULL);
Pavel Begunkov7b202382019-10-27 22:10:36 +03002528 __io_free_req(shadow);
Jackie Liu4fe2c962019-09-09 20:50:40 +08002529 io_cqring_add_event(ctx, s->sqe->user_data, ret);
2530 return 0;
2531 }
2532 } else {
2533 /*
2534 * If ret == 0 means that all IOs in front of link io are
2535 * running done. let's queue link head.
2536 */
2537 need_submit = true;
2538 }
2539
2540 /* Insert shadow req to defer_list, blocking next IOs */
2541 spin_lock_irq(&ctx->completion_lock);
Dmitrii Dolgovc826bd72019-10-15 19:02:01 +02002542 trace_io_uring_defer(ctx, shadow, true);
Jackie Liu4fe2c962019-09-09 20:50:40 +08002543 list_add_tail(&shadow->list, &ctx->defer_list);
2544 spin_unlock_irq(&ctx->completion_lock);
2545
2546 if (need_submit)
Jens Axboebc808bc2019-10-22 13:14:37 -06002547 return __io_queue_sqe(ctx, req, s);
Jackie Liu4fe2c962019-09-09 20:50:40 +08002548
2549 return 0;
2550}
2551
Jens Axboe9e645e112019-05-10 16:07:28 -06002552#define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK)
2553
2554static void io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s,
Jens Axboebc808bc2019-10-22 13:14:37 -06002555 struct io_submit_state *state, struct io_kiocb **link)
Jens Axboe9e645e112019-05-10 16:07:28 -06002556{
2557 struct io_uring_sqe *sqe_copy;
2558 struct io_kiocb *req;
2559 int ret;
2560
2561 /* enforce forwards compatibility on users */
2562 if (unlikely(s->sqe->flags & ~SQE_VALID_FLAGS)) {
2563 ret = -EINVAL;
2564 goto err;
2565 }
2566
2567 req = io_get_req(ctx, state);
2568 if (unlikely(!req)) {
2569 ret = -EAGAIN;
2570 goto err;
2571 }
2572
2573 ret = io_req_set_file(ctx, s, state, req);
2574 if (unlikely(ret)) {
2575err_req:
Jens Axboeba816ad2019-09-28 11:36:45 -06002576 io_free_req(req, NULL);
Jens Axboe9e645e112019-05-10 16:07:28 -06002577err:
2578 io_cqring_add_event(ctx, s->sqe->user_data, ret);
2579 return;
2580 }
2581
Pavel Begunkov84d55dc2019-10-25 12:31:29 +03002582 req->user_data = s->sqe->user_data;
2583
Jens Axboe9e645e112019-05-10 16:07:28 -06002584 /*
2585 * If we already have a head request, queue this one for async
2586 * submittal once the head completes. If we don't have a head but
2587 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2588 * submitted sync once the chain is complete. If none of those
2589 * conditions are true (normal request), then just queue it.
2590 */
2591 if (*link) {
2592 struct io_kiocb *prev = *link;
2593
2594 sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL);
2595 if (!sqe_copy) {
2596 ret = -EAGAIN;
2597 goto err_req;
2598 }
2599
2600 s->sqe = sqe_copy;
2601 memcpy(&req->submit, s, sizeof(*s));
Dmitrii Dolgovc826bd72019-10-15 19:02:01 +02002602 trace_io_uring_link(ctx, req, prev);
Jens Axboe9e645e112019-05-10 16:07:28 -06002603 list_add_tail(&req->list, &prev->link_list);
2604 } else if (s->sqe->flags & IOSQE_IO_LINK) {
2605 req->flags |= REQ_F_LINK;
2606
2607 memcpy(&req->submit, s, sizeof(*s));
2608 INIT_LIST_HEAD(&req->link_list);
2609 *link = req;
2610 } else {
Jens Axboebc808bc2019-10-22 13:14:37 -06002611 io_queue_sqe(ctx, req, s);
Jens Axboe9e645e112019-05-10 16:07:28 -06002612 }
2613}
2614
Jens Axboe9a56a232019-01-09 09:06:50 -07002615/*
2616 * Batched submission is done, ensure local IO is flushed out.
2617 */
2618static void io_submit_state_end(struct io_submit_state *state)
2619{
2620 blk_finish_plug(&state->plug);
Jens Axboe3d6770f2019-04-13 11:50:54 -06002621 io_file_put(state);
Jens Axboe2579f912019-01-09 09:10:43 -07002622 if (state->free_reqs)
2623 kmem_cache_free_bulk(req_cachep, state->free_reqs,
2624 &state->reqs[state->cur_req]);
Jens Axboe9a56a232019-01-09 09:06:50 -07002625}
2626
2627/*
2628 * Start submission side cache.
2629 */
2630static void io_submit_state_start(struct io_submit_state *state,
2631 struct io_ring_ctx *ctx, unsigned max_ios)
2632{
2633 blk_start_plug(&state->plug);
Jens Axboe2579f912019-01-09 09:10:43 -07002634 state->free_reqs = 0;
Jens Axboe9a56a232019-01-09 09:06:50 -07002635 state->file = NULL;
2636 state->ios_left = max_ios;
2637}
2638
Jens Axboe2b188cc2019-01-07 10:46:33 -07002639static void io_commit_sqring(struct io_ring_ctx *ctx)
2640{
Hristo Venev75b28af2019-08-26 17:23:46 +00002641 struct io_rings *rings = ctx->rings;
Jens Axboe2b188cc2019-01-07 10:46:33 -07002642
Hristo Venev75b28af2019-08-26 17:23:46 +00002643 if (ctx->cached_sq_head != READ_ONCE(rings->sq.head)) {
Jens Axboe2b188cc2019-01-07 10:46:33 -07002644 /*
2645 * Ensure any loads from the SQEs are done at this point,
2646 * since once we write the new head, the application could
2647 * write new data to them.
2648 */
Hristo Venev75b28af2019-08-26 17:23:46 +00002649 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
Jens Axboe2b188cc2019-01-07 10:46:33 -07002650 }
2651}
2652
2653/*
Jens Axboe2b188cc2019-01-07 10:46:33 -07002654 * Fetch an sqe, if one is available. Note that s->sqe will point to memory
2655 * that is mapped by userspace. This means that care needs to be taken to
2656 * ensure that reads are stable, as we cannot rely on userspace always
2657 * being a good citizen. If members of the sqe are validated and then later
2658 * used, it's important that those reads are done through READ_ONCE() to
2659 * prevent a re-load down the line.
2660 */
2661static bool io_get_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s)
2662{
Hristo Venev75b28af2019-08-26 17:23:46 +00002663 struct io_rings *rings = ctx->rings;
2664 u32 *sq_array = ctx->sq_array;
Jens Axboe2b188cc2019-01-07 10:46:33 -07002665 unsigned head;
2666
2667 /*
2668 * The cached sq head (or cq tail) serves two purposes:
2669 *
2670 * 1) allows us to batch the cost of updating the user visible
2671 * head updates.
2672 * 2) allows the kernel side to track the head on its own, even
2673 * though the application is the one updating it.
2674 */
2675 head = ctx->cached_sq_head;
Stefan Bühlere523a292019-04-19 11:57:44 +02002676 /* make sure SQ entry isn't read before tail */
Hristo Venev75b28af2019-08-26 17:23:46 +00002677 if (head == smp_load_acquire(&rings->sq.tail))
Jens Axboe2b188cc2019-01-07 10:46:33 -07002678 return false;
2679
Hristo Venev75b28af2019-08-26 17:23:46 +00002680 head = READ_ONCE(sq_array[head & ctx->sq_mask]);
Jens Axboe2b188cc2019-01-07 10:46:33 -07002681 if (head < ctx->sq_entries) {
Jens Axboefcb323c2019-10-24 12:39:47 -06002682 s->ring_file = NULL;
Jens Axboe2b188cc2019-01-07 10:46:33 -07002683 s->sqe = &ctx->sq_sqes[head];
Jackie Liu8776f3f2019-09-09 20:50:39 +08002684 s->sequence = ctx->cached_sq_head;
Jens Axboe2b188cc2019-01-07 10:46:33 -07002685 ctx->cached_sq_head++;
2686 return true;
2687 }
2688
2689 /* drop invalid entries */
2690 ctx->cached_sq_head++;
Jens Axboe498ccd92019-10-25 10:04:25 -06002691 ctx->cached_sq_dropped++;
2692 WRITE_ONCE(rings->sq_dropped, ctx->cached_sq_dropped);
Jens Axboe2b188cc2019-01-07 10:46:33 -07002693 return false;
2694}
2695
Pavel Begunkovfb5ccc92019-10-25 12:31:30 +03002696static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
Pavel Begunkovae9428c2019-11-06 00:22:14 +03002697 struct file *ring_file, int ring_fd,
2698 struct mm_struct **mm, bool async)
Jens Axboe6c271ce2019-01-10 11:22:30 -07002699{
2700 struct io_submit_state state, *statep = NULL;
Jens Axboe9e645e112019-05-10 16:07:28 -06002701 struct io_kiocb *link = NULL;
Jackie Liu4fe2c962019-09-09 20:50:40 +08002702 struct io_kiocb *shadow_req = NULL;
Jens Axboe9e645e112019-05-10 16:07:28 -06002703 int i, submitted = 0;
Pavel Begunkov95a1b3ff2019-10-27 23:15:41 +03002704 bool mm_fault = false;
Jens Axboe6c271ce2019-01-10 11:22:30 -07002705
2706 if (nr > IO_PLUG_THRESHOLD) {
2707 io_submit_state_start(&state, ctx, nr);
2708 statep = &state;
2709 }
2710
2711 for (i = 0; i < nr; i++) {
Pavel Begunkovfb5ccc92019-10-25 12:31:30 +03002712 struct sqe_submit s;
2713
2714 if (!io_get_sqring(ctx, &s))
2715 break;
2716
Pavel Begunkov95a1b3ff2019-10-27 23:15:41 +03002717 if (io_sqe_needs_user(s.sqe) && !*mm) {
2718 mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm);
2719 if (!mm_fault) {
2720 use_mm(ctx->sqo_mm);
2721 *mm = ctx->sqo_mm;
2722 }
2723 }
2724
Pavel Begunkovfb5ccc92019-10-25 12:31:30 +03002725 if (link && (s.sqe->flags & IOSQE_IO_DRAIN)) {
Jackie Liu4fe2c962019-09-09 20:50:40 +08002726 if (!shadow_req) {
2727 shadow_req = io_get_req(ctx, NULL);
Jackie Liua1041c22019-09-18 17:25:52 +08002728 if (unlikely(!shadow_req))
2729 goto out;
Jackie Liu4fe2c962019-09-09 20:50:40 +08002730 shadow_req->flags |= (REQ_F_IO_DRAIN | REQ_F_SHADOW_DRAIN);
2731 refcount_dec(&shadow_req->refs);
2732 }
Pavel Begunkovfb5ccc92019-10-25 12:31:30 +03002733 shadow_req->sequence = s.sequence;
Jackie Liu4fe2c962019-09-09 20:50:40 +08002734 }
2735
Jackie Liua1041c22019-09-18 17:25:52 +08002736out:
Pavel Begunkovae9428c2019-11-06 00:22:14 +03002737 s.ring_file = ring_file;
2738 s.ring_fd = ring_fd;
Pavel Begunkov95a1b3ff2019-10-27 23:15:41 +03002739 s.has_user = *mm != NULL;
Pavel Begunkovae9428c2019-11-06 00:22:14 +03002740 s.in_async = async;
2741 s.needs_fixed_file = async;
2742 trace_io_uring_submit_sqe(ctx, s.sqe->user_data, true, async);
Pavel Begunkov95a1b3ff2019-10-27 23:15:41 +03002743 io_submit_sqe(ctx, &s, statep, &link);
2744 submitted++;
Pavel Begunkove5eb6362019-11-06 00:22:15 +03002745
2746 /*
2747 * If previous wasn't linked and we have a linked command,
2748 * that's the end of the chain. Submit the previous link.
2749 */
2750 if (!(s.sqe->flags & IOSQE_IO_LINK) && link) {
2751 io_queue_link_head(ctx, link, &link->submit, shadow_req);
2752 link = NULL;
2753 shadow_req = NULL;
2754 }
Jens Axboe6c271ce2019-01-10 11:22:30 -07002755 }
2756
Jens Axboe9e645e112019-05-10 16:07:28 -06002757 if (link)
Jens Axboebc808bc2019-10-22 13:14:37 -06002758 io_queue_link_head(ctx, link, &link->submit, shadow_req);
Jens Axboe6c271ce2019-01-10 11:22:30 -07002759 if (statep)
2760 io_submit_state_end(&state);
2761
Pavel Begunkovae9428c2019-11-06 00:22:14 +03002762 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2763 io_commit_sqring(ctx);
2764
Jens Axboe6c271ce2019-01-10 11:22:30 -07002765 return submitted;
2766}
2767
2768static int io_sq_thread(void *data)
2769{
Jens Axboe6c271ce2019-01-10 11:22:30 -07002770 struct io_ring_ctx *ctx = data;
2771 struct mm_struct *cur_mm = NULL;
2772 mm_segment_t old_fs;
2773 DEFINE_WAIT(wait);
2774 unsigned inflight;
2775 unsigned long timeout;
2776
Jackie Liua4c0b3d2019-07-08 13:41:12 +08002777 complete(&ctx->sqo_thread_started);
2778
Jens Axboe6c271ce2019-01-10 11:22:30 -07002779 old_fs = get_fs();
2780 set_fs(USER_DS);
2781
2782 timeout = inflight = 0;
Roman Penyaev2bbcd6d2019-05-16 10:53:57 +02002783 while (!kthread_should_park()) {
Pavel Begunkovfb5ccc92019-10-25 12:31:30 +03002784 unsigned int to_submit;
Jens Axboe6c271ce2019-01-10 11:22:30 -07002785
2786 if (inflight) {
2787 unsigned nr_events = 0;
2788
2789 if (ctx->flags & IORING_SETUP_IOPOLL) {
Jens Axboe2b2ed972019-10-25 10:06:15 -06002790 /*
2791 * inflight is the count of the maximum possible
2792 * entries we submitted, but it can be smaller
2793 * if we dropped some of them. If we don't have
2794 * poll entries available, then we know that we
2795 * have nothing left to poll for. Reset the
2796 * inflight count to zero in that case.
2797 */
2798 mutex_lock(&ctx->uring_lock);
2799 if (!list_empty(&ctx->poll_list))
2800 __io_iopoll_check(ctx, &nr_events, 0);
2801 else
2802 inflight = 0;
2803 mutex_unlock(&ctx->uring_lock);
Jens Axboe6c271ce2019-01-10 11:22:30 -07002804 } else {
2805 /*
2806 * Normal IO, just pretend everything completed.
2807 * We don't have to poll completions for that.
2808 */
2809 nr_events = inflight;
2810 }
2811
2812 inflight -= nr_events;
2813 if (!inflight)
2814 timeout = jiffies + ctx->sq_thread_idle;
2815 }
2816
Pavel Begunkovfb5ccc92019-10-25 12:31:30 +03002817 to_submit = io_sqring_entries(ctx);
2818 if (!to_submit) {
Jens Axboe6c271ce2019-01-10 11:22:30 -07002819 /*
2820 * We're polling. If we're within the defined idle
2821 * period, then let us spin without work before going
2822 * to sleep.
2823 */
2824 if (inflight || !time_after(jiffies, timeout)) {
Jens Axboe9831a902019-09-19 09:48:55 -06002825 cond_resched();
Jens Axboe6c271ce2019-01-10 11:22:30 -07002826 continue;
2827 }
2828
2829 /*
2830 * Drop cur_mm before scheduling, we can't hold it for
2831 * long periods (or over schedule()). Do this before
2832 * adding ourselves to the waitqueue, as the unuse/drop
2833 * may sleep.
2834 */
2835 if (cur_mm) {
2836 unuse_mm(cur_mm);
2837 mmput(cur_mm);
2838 cur_mm = NULL;
2839 }
2840
2841 prepare_to_wait(&ctx->sqo_wait, &wait,
2842 TASK_INTERRUPTIBLE);
2843
2844 /* Tell userspace we may need a wakeup call */
Hristo Venev75b28af2019-08-26 17:23:46 +00002845 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
Stefan Bühler0d7bae62019-04-19 11:57:45 +02002846 /* make sure to read SQ tail after writing flags */
2847 smp_mb();
Jens Axboe6c271ce2019-01-10 11:22:30 -07002848
Pavel Begunkovfb5ccc92019-10-25 12:31:30 +03002849 to_submit = io_sqring_entries(ctx);
2850 if (!to_submit) {
Roman Penyaev2bbcd6d2019-05-16 10:53:57 +02002851 if (kthread_should_park()) {
Jens Axboe6c271ce2019-01-10 11:22:30 -07002852 finish_wait(&ctx->sqo_wait, &wait);
2853 break;
2854 }
2855 if (signal_pending(current))
2856 flush_signals(current);
2857 schedule();
2858 finish_wait(&ctx->sqo_wait, &wait);
2859
Hristo Venev75b28af2019-08-26 17:23:46 +00002860 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
Jens Axboe6c271ce2019-01-10 11:22:30 -07002861 continue;
2862 }
2863 finish_wait(&ctx->sqo_wait, &wait);
2864
Hristo Venev75b28af2019-08-26 17:23:46 +00002865 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
Jens Axboe6c271ce2019-01-10 11:22:30 -07002866 }
2867
Pavel Begunkovfb5ccc92019-10-25 12:31:30 +03002868 to_submit = min(to_submit, ctx->sq_entries);
Pavel Begunkovae9428c2019-11-06 00:22:14 +03002869 inflight += io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm,
2870 true);
Jens Axboe6c271ce2019-01-10 11:22:30 -07002871 }
2872
2873 set_fs(old_fs);
2874 if (cur_mm) {
2875 unuse_mm(cur_mm);
2876 mmput(cur_mm);
2877 }
Jens Axboe06058632019-04-13 09:26:03 -06002878
Roman Penyaev2bbcd6d2019-05-16 10:53:57 +02002879 kthread_parkme();
Jens Axboe06058632019-04-13 09:26:03 -06002880
Jens Axboe6c271ce2019-01-10 11:22:30 -07002881 return 0;
2882}
2883
Jens Axboebda52162019-09-24 13:47:15 -06002884struct io_wait_queue {
2885 struct wait_queue_entry wq;
2886 struct io_ring_ctx *ctx;
2887 unsigned to_wait;
2888 unsigned nr_timeouts;
2889};
2890
2891static inline bool io_should_wake(struct io_wait_queue *iowq)
2892{
2893 struct io_ring_ctx *ctx = iowq->ctx;
2894
2895 /*
2896 * Wake up if we have enough events, or if a timeout occured since we
2897 * started waiting. For timeouts, we always want to return to userspace,
2898 * regardless of event count.
2899 */
2900 return io_cqring_events(ctx->rings) >= iowq->to_wait ||
2901 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2902}
2903
2904static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2905 int wake_flags, void *key)
2906{
2907 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
2908 wq);
2909
2910 if (!io_should_wake(iowq))
2911 return -1;
2912
2913 return autoremove_wake_function(curr, mode, wake_flags, key);
2914}
2915
Jens Axboe2b188cc2019-01-07 10:46:33 -07002916/*
2917 * Wait until events become available, if we don't already have some. The
2918 * application must reap them itself, as they reside on the shared cq ring.
2919 */
2920static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2921 const sigset_t __user *sig, size_t sigsz)
2922{
Jens Axboebda52162019-09-24 13:47:15 -06002923 struct io_wait_queue iowq = {
2924 .wq = {
2925 .private = current,
2926 .func = io_wake_function,
2927 .entry = LIST_HEAD_INIT(iowq.wq.entry),
2928 },
2929 .ctx = ctx,
2930 .to_wait = min_events,
2931 };
Hristo Venev75b28af2019-08-26 17:23:46 +00002932 struct io_rings *rings = ctx->rings;
Jackie Liue9ffa5c2019-10-29 11:16:42 +08002933 int ret = 0;
Jens Axboe2b188cc2019-01-07 10:46:33 -07002934
Hristo Venev75b28af2019-08-26 17:23:46 +00002935 if (io_cqring_events(rings) >= min_events)
Jens Axboe2b188cc2019-01-07 10:46:33 -07002936 return 0;
2937
2938 if (sig) {
Arnd Bergmann9e75ad52019-03-25 15:34:53 +01002939#ifdef CONFIG_COMPAT
2940 if (in_compat_syscall())
2941 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
Oleg Nesterovb7724342019-07-16 16:29:53 -07002942 sigsz);
Arnd Bergmann9e75ad52019-03-25 15:34:53 +01002943 else
2944#endif
Oleg Nesterovb7724342019-07-16 16:29:53 -07002945 ret = set_user_sigmask(sig, sigsz);
Arnd Bergmann9e75ad52019-03-25 15:34:53 +01002946
Jens Axboe2b188cc2019-01-07 10:46:33 -07002947 if (ret)
2948 return ret;
2949 }
2950
Jens Axboebda52162019-09-24 13:47:15 -06002951 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
Dmitrii Dolgovc826bd72019-10-15 19:02:01 +02002952 trace_io_uring_cqring_wait(ctx, min_events);
Jens Axboebda52162019-09-24 13:47:15 -06002953 do {
2954 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
2955 TASK_INTERRUPTIBLE);
2956 if (io_should_wake(&iowq))
2957 break;
2958 schedule();
2959 if (signal_pending(current)) {
Jackie Liue9ffa5c2019-10-29 11:16:42 +08002960 ret = -EINTR;
Jens Axboebda52162019-09-24 13:47:15 -06002961 break;
2962 }
2963 } while (1);
2964 finish_wait(&ctx->wait, &iowq.wq);
2965
Jackie Liue9ffa5c2019-10-29 11:16:42 +08002966 restore_saved_sigmask_unless(ret == -EINTR);
Jens Axboe2b188cc2019-01-07 10:46:33 -07002967
Hristo Venev75b28af2019-08-26 17:23:46 +00002968 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
Jens Axboe2b188cc2019-01-07 10:46:33 -07002969}
2970
Jens Axboe6b063142019-01-10 22:13:58 -07002971static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
2972{
2973#if defined(CONFIG_UNIX)
2974 if (ctx->ring_sock) {
2975 struct sock *sock = ctx->ring_sock->sk;
2976 struct sk_buff *skb;
2977
2978 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
2979 kfree_skb(skb);
2980 }
2981#else
2982 int i;
2983
Jens Axboe65e19f52019-10-26 07:20:21 -06002984 for (i = 0; i < ctx->nr_user_files; i++) {
2985 struct file *file;
2986
2987 file = io_file_from_index(ctx, i);
2988 if (file)
2989 fput(file);
2990 }
Jens Axboe6b063142019-01-10 22:13:58 -07002991#endif
2992}
2993
2994static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
2995{
Jens Axboe65e19f52019-10-26 07:20:21 -06002996 unsigned nr_tables, i;
2997
2998 if (!ctx->file_table)
Jens Axboe6b063142019-01-10 22:13:58 -07002999 return -ENXIO;
3000
3001 __io_sqe_files_unregister(ctx);
Jens Axboe65e19f52019-10-26 07:20:21 -06003002 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
3003 for (i = 0; i < nr_tables; i++)
3004 kfree(ctx->file_table[i].files);
3005 kfree(ctx->file_table);
3006 ctx->file_table = NULL;
Jens Axboe6b063142019-01-10 22:13:58 -07003007 ctx->nr_user_files = 0;
3008 return 0;
3009}
3010
Jens Axboe6c271ce2019-01-10 11:22:30 -07003011static void io_sq_thread_stop(struct io_ring_ctx *ctx)
3012{
3013 if (ctx->sqo_thread) {
Jackie Liua4c0b3d2019-07-08 13:41:12 +08003014 wait_for_completion(&ctx->sqo_thread_started);
Roman Penyaev2bbcd6d2019-05-16 10:53:57 +02003015 /*
3016 * The park is a bit of a work-around, without it we get
3017 * warning spews on shutdown with SQPOLL set and affinity
3018 * set to a single CPU.
3019 */
Jens Axboe06058632019-04-13 09:26:03 -06003020 kthread_park(ctx->sqo_thread);
Jens Axboe6c271ce2019-01-10 11:22:30 -07003021 kthread_stop(ctx->sqo_thread);
3022 ctx->sqo_thread = NULL;
3023 }
3024}
3025
Jens Axboe6b063142019-01-10 22:13:58 -07003026static void io_finish_async(struct io_ring_ctx *ctx)
3027{
Jens Axboe6c271ce2019-01-10 11:22:30 -07003028 io_sq_thread_stop(ctx);
3029
Jens Axboe561fb042019-10-24 07:25:42 -06003030 if (ctx->io_wq) {
3031 io_wq_destroy(ctx->io_wq);
3032 ctx->io_wq = NULL;
Jens Axboe6b063142019-01-10 22:13:58 -07003033 }
3034}
3035
3036#if defined(CONFIG_UNIX)
3037static void io_destruct_skb(struct sk_buff *skb)
3038{
3039 struct io_ring_ctx *ctx = skb->sk->sk_user_data;
3040
Jens Axboe561fb042019-10-24 07:25:42 -06003041 if (ctx->io_wq)
3042 io_wq_flush(ctx->io_wq);
Jens Axboe8a997342019-10-09 14:40:13 -06003043
Jens Axboe6b063142019-01-10 22:13:58 -07003044 unix_destruct_scm(skb);
3045}
3046
3047/*
3048 * Ensure the UNIX gc is aware of our file set, so we are certain that
3049 * the io_uring can be safely unregistered on process exit, even if we have
3050 * loops in the file referencing.
3051 */
3052static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
3053{
3054 struct sock *sk = ctx->ring_sock->sk;
3055 struct scm_fp_list *fpl;
3056 struct sk_buff *skb;
Jens Axboe08a45172019-10-03 08:11:03 -06003057 int i, nr_files;
Jens Axboe6b063142019-01-10 22:13:58 -07003058
3059 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
3060 unsigned long inflight = ctx->user->unix_inflight + nr;
3061
3062 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
3063 return -EMFILE;
3064 }
3065
3066 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
3067 if (!fpl)
3068 return -ENOMEM;
3069
3070 skb = alloc_skb(0, GFP_KERNEL);
3071 if (!skb) {
3072 kfree(fpl);
3073 return -ENOMEM;
3074 }
3075
3076 skb->sk = sk;
Jens Axboe6b063142019-01-10 22:13:58 -07003077
Jens Axboe08a45172019-10-03 08:11:03 -06003078 nr_files = 0;
Jens Axboe6b063142019-01-10 22:13:58 -07003079 fpl->user = get_uid(ctx->user);
3080 for (i = 0; i < nr; i++) {
Jens Axboe65e19f52019-10-26 07:20:21 -06003081 struct file *file = io_file_from_index(ctx, i + offset);
3082
3083 if (!file)
Jens Axboe08a45172019-10-03 08:11:03 -06003084 continue;
Jens Axboe65e19f52019-10-26 07:20:21 -06003085 fpl->fp[nr_files] = get_file(file);
Jens Axboe08a45172019-10-03 08:11:03 -06003086 unix_inflight(fpl->user, fpl->fp[nr_files]);
3087 nr_files++;
Jens Axboe6b063142019-01-10 22:13:58 -07003088 }
3089
Jens Axboe08a45172019-10-03 08:11:03 -06003090 if (nr_files) {
3091 fpl->max = SCM_MAX_FD;
3092 fpl->count = nr_files;
3093 UNIXCB(skb).fp = fpl;
3094 skb->destructor = io_destruct_skb;
3095 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
3096 skb_queue_head(&sk->sk_receive_queue, skb);
Jens Axboe6b063142019-01-10 22:13:58 -07003097
Jens Axboe08a45172019-10-03 08:11:03 -06003098 for (i = 0; i < nr_files; i++)
3099 fput(fpl->fp[i]);
3100 } else {
3101 kfree_skb(skb);
3102 kfree(fpl);
3103 }
Jens Axboe6b063142019-01-10 22:13:58 -07003104
3105 return 0;
3106}
3107
3108/*
3109 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
3110 * causes regular reference counting to break down. We rely on the UNIX
3111 * garbage collection to take care of this problem for us.
3112 */
3113static int io_sqe_files_scm(struct io_ring_ctx *ctx)
3114{
3115 unsigned left, total;
3116 int ret = 0;
3117
3118 total = 0;
3119 left = ctx->nr_user_files;
3120 while (left) {
3121 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
Jens Axboe6b063142019-01-10 22:13:58 -07003122
3123 ret = __io_sqe_files_scm(ctx, this_files, total);
3124 if (ret)
3125 break;
3126 left -= this_files;
3127 total += this_files;
3128 }
3129
3130 if (!ret)
3131 return 0;
3132
3133 while (total < ctx->nr_user_files) {
Jens Axboe65e19f52019-10-26 07:20:21 -06003134 struct file *file = io_file_from_index(ctx, total);
3135
3136 if (file)
3137 fput(file);
Jens Axboe6b063142019-01-10 22:13:58 -07003138 total++;
3139 }
3140
3141 return ret;
3142}
3143#else
3144static int io_sqe_files_scm(struct io_ring_ctx *ctx)
3145{
3146 return 0;
3147}
3148#endif
3149
Jens Axboe65e19f52019-10-26 07:20:21 -06003150static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
3151 unsigned nr_files)
3152{
3153 int i;
3154
3155 for (i = 0; i < nr_tables; i++) {
3156 struct fixed_file_table *table = &ctx->file_table[i];
3157 unsigned this_files;
3158
3159 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
3160 table->files = kcalloc(this_files, sizeof(struct file *),
3161 GFP_KERNEL);
3162 if (!table->files)
3163 break;
3164 nr_files -= this_files;
3165 }
3166
3167 if (i == nr_tables)
3168 return 0;
3169
3170 for (i = 0; i < nr_tables; i++) {
3171 struct fixed_file_table *table = &ctx->file_table[i];
3172 kfree(table->files);
3173 }
3174 return 1;
3175}
3176
Jens Axboe6b063142019-01-10 22:13:58 -07003177static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
3178 unsigned nr_args)
3179{
3180 __s32 __user *fds = (__s32 __user *) arg;
Jens Axboe65e19f52019-10-26 07:20:21 -06003181 unsigned nr_tables;
Jens Axboe6b063142019-01-10 22:13:58 -07003182 int fd, ret = 0;
3183 unsigned i;
3184
Jens Axboe65e19f52019-10-26 07:20:21 -06003185 if (ctx->file_table)
Jens Axboe6b063142019-01-10 22:13:58 -07003186 return -EBUSY;
3187 if (!nr_args)
3188 return -EINVAL;
3189 if (nr_args > IORING_MAX_FIXED_FILES)
3190 return -EMFILE;
3191
Jens Axboe65e19f52019-10-26 07:20:21 -06003192 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
3193 ctx->file_table = kcalloc(nr_tables, sizeof(struct fixed_file_table),
3194 GFP_KERNEL);
3195 if (!ctx->file_table)
Jens Axboe6b063142019-01-10 22:13:58 -07003196 return -ENOMEM;
3197
Jens Axboe65e19f52019-10-26 07:20:21 -06003198 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
3199 kfree(ctx->file_table);
3200 return -ENOMEM;
3201 }
3202
Jens Axboe08a45172019-10-03 08:11:03 -06003203 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
Jens Axboe65e19f52019-10-26 07:20:21 -06003204 struct fixed_file_table *table;
3205 unsigned index;
3206
Jens Axboe6b063142019-01-10 22:13:58 -07003207 ret = -EFAULT;
3208 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
3209 break;
Jens Axboe08a45172019-10-03 08:11:03 -06003210 /* allow sparse sets */
3211 if (fd == -1) {
3212 ret = 0;
3213 continue;
3214 }
Jens Axboe6b063142019-01-10 22:13:58 -07003215
Jens Axboe65e19f52019-10-26 07:20:21 -06003216 table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT];
3217 index = i & IORING_FILE_TABLE_MASK;
3218 table->files[index] = fget(fd);
Jens Axboe6b063142019-01-10 22:13:58 -07003219
3220 ret = -EBADF;
Jens Axboe65e19f52019-10-26 07:20:21 -06003221 if (!table->files[index])
Jens Axboe6b063142019-01-10 22:13:58 -07003222 break;
3223 /*
3224 * Don't allow io_uring instances to be registered. If UNIX
3225 * isn't enabled, then this causes a reference cycle and this
3226 * instance can never get freed. If UNIX is enabled we'll
3227 * handle it just fine, but there's still no point in allowing
3228 * a ring fd as it doesn't support regular read/write anyway.
3229 */
Jens Axboe65e19f52019-10-26 07:20:21 -06003230 if (table->files[index]->f_op == &io_uring_fops) {
3231 fput(table->files[index]);
Jens Axboe6b063142019-01-10 22:13:58 -07003232 break;
3233 }
Jens Axboe6b063142019-01-10 22:13:58 -07003234 ret = 0;
3235 }
3236
3237 if (ret) {
Jens Axboe65e19f52019-10-26 07:20:21 -06003238 for (i = 0; i < ctx->nr_user_files; i++) {
3239 struct file *file;
Jens Axboe6b063142019-01-10 22:13:58 -07003240
Jens Axboe65e19f52019-10-26 07:20:21 -06003241 file = io_file_from_index(ctx, i);
3242 if (file)
3243 fput(file);
3244 }
3245 for (i = 0; i < nr_tables; i++)
3246 kfree(ctx->file_table[i].files);
3247
3248 kfree(ctx->file_table);
3249 ctx->file_table = NULL;
Jens Axboe6b063142019-01-10 22:13:58 -07003250 ctx->nr_user_files = 0;
3251 return ret;
3252 }
3253
3254 ret = io_sqe_files_scm(ctx);
3255 if (ret)
3256 io_sqe_files_unregister(ctx);
3257
3258 return ret;
3259}
3260
Jens Axboec3a31e62019-10-03 13:59:56 -06003261static void io_sqe_file_unregister(struct io_ring_ctx *ctx, int index)
3262{
3263#if defined(CONFIG_UNIX)
Jens Axboe65e19f52019-10-26 07:20:21 -06003264 struct file *file = io_file_from_index(ctx, index);
Jens Axboec3a31e62019-10-03 13:59:56 -06003265 struct sock *sock = ctx->ring_sock->sk;
3266 struct sk_buff_head list, *head = &sock->sk_receive_queue;
3267 struct sk_buff *skb;
3268 int i;
3269
3270 __skb_queue_head_init(&list);
3271
3272 /*
3273 * Find the skb that holds this file in its SCM_RIGHTS. When found,
3274 * remove this entry and rearrange the file array.
3275 */
3276 skb = skb_dequeue(head);
3277 while (skb) {
3278 struct scm_fp_list *fp;
3279
3280 fp = UNIXCB(skb).fp;
3281 for (i = 0; i < fp->count; i++) {
3282 int left;
3283
3284 if (fp->fp[i] != file)
3285 continue;
3286
3287 unix_notinflight(fp->user, fp->fp[i]);
3288 left = fp->count - 1 - i;
3289 if (left) {
3290 memmove(&fp->fp[i], &fp->fp[i + 1],
3291 left * sizeof(struct file *));
3292 }
3293 fp->count--;
3294 if (!fp->count) {
3295 kfree_skb(skb);
3296 skb = NULL;
3297 } else {
3298 __skb_queue_tail(&list, skb);
3299 }
3300 fput(file);
3301 file = NULL;
3302 break;
3303 }
3304
3305 if (!file)
3306 break;
3307
3308 __skb_queue_tail(&list, skb);
3309
3310 skb = skb_dequeue(head);
3311 }
3312
3313 if (skb_peek(&list)) {
3314 spin_lock_irq(&head->lock);
3315 while ((skb = __skb_dequeue(&list)) != NULL)
3316 __skb_queue_tail(head, skb);
3317 spin_unlock_irq(&head->lock);
3318 }
3319#else
Jens Axboe65e19f52019-10-26 07:20:21 -06003320 fput(io_file_from_index(ctx, index));
Jens Axboec3a31e62019-10-03 13:59:56 -06003321#endif
3322}
3323
3324static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
3325 int index)
3326{
3327#if defined(CONFIG_UNIX)
3328 struct sock *sock = ctx->ring_sock->sk;
3329 struct sk_buff_head *head = &sock->sk_receive_queue;
3330 struct sk_buff *skb;
3331
3332 /*
3333 * See if we can merge this file into an existing skb SCM_RIGHTS
3334 * file set. If there's no room, fall back to allocating a new skb
3335 * and filling it in.
3336 */
3337 spin_lock_irq(&head->lock);
3338 skb = skb_peek(head);
3339 if (skb) {
3340 struct scm_fp_list *fpl = UNIXCB(skb).fp;
3341
3342 if (fpl->count < SCM_MAX_FD) {
3343 __skb_unlink(skb, head);
3344 spin_unlock_irq(&head->lock);
3345 fpl->fp[fpl->count] = get_file(file);
3346 unix_inflight(fpl->user, fpl->fp[fpl->count]);
3347 fpl->count++;
3348 spin_lock_irq(&head->lock);
3349 __skb_queue_head(head, skb);
3350 } else {
3351 skb = NULL;
3352 }
3353 }
3354 spin_unlock_irq(&head->lock);
3355
3356 if (skb) {
3357 fput(file);
3358 return 0;
3359 }
3360
3361 return __io_sqe_files_scm(ctx, 1, index);
3362#else
3363 return 0;
3364#endif
3365}
3366
3367static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
3368 unsigned nr_args)
3369{
3370 struct io_uring_files_update up;
3371 __s32 __user *fds;
3372 int fd, i, err;
3373 __u32 done;
3374
Jens Axboe65e19f52019-10-26 07:20:21 -06003375 if (!ctx->file_table)
Jens Axboec3a31e62019-10-03 13:59:56 -06003376 return -ENXIO;
3377 if (!nr_args)
3378 return -EINVAL;
3379 if (copy_from_user(&up, arg, sizeof(up)))
3380 return -EFAULT;
3381 if (check_add_overflow(up.offset, nr_args, &done))
3382 return -EOVERFLOW;
3383 if (done > ctx->nr_user_files)
3384 return -EINVAL;
3385
3386 done = 0;
3387 fds = (__s32 __user *) up.fds;
3388 while (nr_args) {
Jens Axboe65e19f52019-10-26 07:20:21 -06003389 struct fixed_file_table *table;
3390 unsigned index;
3391
Jens Axboec3a31e62019-10-03 13:59:56 -06003392 err = 0;
3393 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
3394 err = -EFAULT;
3395 break;
3396 }
3397 i = array_index_nospec(up.offset, ctx->nr_user_files);
Jens Axboe65e19f52019-10-26 07:20:21 -06003398 table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT];
3399 index = i & IORING_FILE_TABLE_MASK;
3400 if (table->files[index]) {
Jens Axboec3a31e62019-10-03 13:59:56 -06003401 io_sqe_file_unregister(ctx, i);
Jens Axboe65e19f52019-10-26 07:20:21 -06003402 table->files[index] = NULL;
Jens Axboec3a31e62019-10-03 13:59:56 -06003403 }
3404 if (fd != -1) {
3405 struct file *file;
3406
3407 file = fget(fd);
3408 if (!file) {
3409 err = -EBADF;
3410 break;
3411 }
3412 /*
3413 * Don't allow io_uring instances to be registered. If
3414 * UNIX isn't enabled, then this causes a reference
3415 * cycle and this instance can never get freed. If UNIX
3416 * is enabled we'll handle it just fine, but there's
3417 * still no point in allowing a ring fd as it doesn't
3418 * support regular read/write anyway.
3419 */
3420 if (file->f_op == &io_uring_fops) {
3421 fput(file);
3422 err = -EBADF;
3423 break;
3424 }
Jens Axboe65e19f52019-10-26 07:20:21 -06003425 table->files[index] = file;
Jens Axboec3a31e62019-10-03 13:59:56 -06003426 err = io_sqe_file_register(ctx, file, i);
3427 if (err)
3428 break;
3429 }
3430 nr_args--;
3431 done++;
3432 up.offset++;
3433 }
3434
3435 return done ? done : err;
3436}
3437
Jens Axboe6c271ce2019-01-10 11:22:30 -07003438static int io_sq_offload_start(struct io_ring_ctx *ctx,
3439 struct io_uring_params *p)
Jens Axboe2b188cc2019-01-07 10:46:33 -07003440{
Jens Axboe561fb042019-10-24 07:25:42 -06003441 unsigned concurrency;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003442 int ret;
3443
Jens Axboe6c271ce2019-01-10 11:22:30 -07003444 init_waitqueue_head(&ctx->sqo_wait);
Jens Axboe2b188cc2019-01-07 10:46:33 -07003445 mmgrab(current->mm);
3446 ctx->sqo_mm = current->mm;
3447
Jens Axboe6c271ce2019-01-10 11:22:30 -07003448 if (ctx->flags & IORING_SETUP_SQPOLL) {
Jens Axboe3ec482d2019-04-08 10:51:01 -06003449 ret = -EPERM;
3450 if (!capable(CAP_SYS_ADMIN))
3451 goto err;
3452
Jens Axboe917257d2019-04-13 09:28:55 -06003453 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
3454 if (!ctx->sq_thread_idle)
3455 ctx->sq_thread_idle = HZ;
3456
Jens Axboe6c271ce2019-01-10 11:22:30 -07003457 if (p->flags & IORING_SETUP_SQ_AFF) {
Jens Axboe44a9bd12019-05-14 20:00:30 -06003458 int cpu = p->sq_thread_cpu;
Jens Axboe6c271ce2019-01-10 11:22:30 -07003459
Jens Axboe917257d2019-04-13 09:28:55 -06003460 ret = -EINVAL;
Jens Axboe44a9bd12019-05-14 20:00:30 -06003461 if (cpu >= nr_cpu_ids)
3462 goto err;
Shenghui Wang7889f442019-05-07 16:03:19 +08003463 if (!cpu_online(cpu))
Jens Axboe917257d2019-04-13 09:28:55 -06003464 goto err;
3465
Jens Axboe6c271ce2019-01-10 11:22:30 -07003466 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
3467 ctx, cpu,
3468 "io_uring-sq");
3469 } else {
3470 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
3471 "io_uring-sq");
3472 }
3473 if (IS_ERR(ctx->sqo_thread)) {
3474 ret = PTR_ERR(ctx->sqo_thread);
3475 ctx->sqo_thread = NULL;
3476 goto err;
3477 }
3478 wake_up_process(ctx->sqo_thread);
3479 } else if (p->flags & IORING_SETUP_SQ_AFF) {
3480 /* Can't have SQ_AFF without SQPOLL */
3481 ret = -EINVAL;
3482 goto err;
3483 }
3484
Jens Axboe561fb042019-10-24 07:25:42 -06003485 /* Do QD, or 4 * CPUS, whatever is smallest */
3486 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
3487 ctx->io_wq = io_wq_create(concurrency, ctx->sqo_mm);
Jens Axboe975c99a52019-10-30 08:42:56 -06003488 if (IS_ERR(ctx->io_wq)) {
3489 ret = PTR_ERR(ctx->io_wq);
3490 ctx->io_wq = NULL;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003491 goto err;
3492 }
3493
3494 return 0;
3495err:
Jens Axboe54a91f32019-09-10 09:15:04 -06003496 io_finish_async(ctx);
Jens Axboe2b188cc2019-01-07 10:46:33 -07003497 mmdrop(ctx->sqo_mm);
3498 ctx->sqo_mm = NULL;
3499 return ret;
3500}
3501
3502static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
3503{
3504 atomic_long_sub(nr_pages, &user->locked_vm);
3505}
3506
3507static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
3508{
3509 unsigned long page_limit, cur_pages, new_pages;
3510
3511 /* Don't allow more pages than we can safely lock */
3512 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
3513
3514 do {
3515 cur_pages = atomic_long_read(&user->locked_vm);
3516 new_pages = cur_pages + nr_pages;
3517 if (new_pages > page_limit)
3518 return -ENOMEM;
3519 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
3520 new_pages) != cur_pages);
3521
3522 return 0;
3523}
3524
3525static void io_mem_free(void *ptr)
3526{
Mark Rutland52e04ef2019-04-30 17:30:21 +01003527 struct page *page;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003528
Mark Rutland52e04ef2019-04-30 17:30:21 +01003529 if (!ptr)
3530 return;
3531
3532 page = virt_to_head_page(ptr);
Jens Axboe2b188cc2019-01-07 10:46:33 -07003533 if (put_page_testzero(page))
3534 free_compound_page(page);
3535}
3536
3537static void *io_mem_alloc(size_t size)
3538{
3539 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
3540 __GFP_NORETRY;
3541
3542 return (void *) __get_free_pages(gfp_flags, get_order(size));
3543}
3544
Hristo Venev75b28af2019-08-26 17:23:46 +00003545static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
3546 size_t *sq_offset)
3547{
3548 struct io_rings *rings;
3549 size_t off, sq_array_size;
3550
3551 off = struct_size(rings, cqes, cq_entries);
3552 if (off == SIZE_MAX)
3553 return SIZE_MAX;
3554
3555#ifdef CONFIG_SMP
3556 off = ALIGN(off, SMP_CACHE_BYTES);
3557 if (off == 0)
3558 return SIZE_MAX;
3559#endif
3560
3561 sq_array_size = array_size(sizeof(u32), sq_entries);
3562 if (sq_array_size == SIZE_MAX)
3563 return SIZE_MAX;
3564
3565 if (check_add_overflow(off, sq_array_size, &off))
3566 return SIZE_MAX;
3567
3568 if (sq_offset)
3569 *sq_offset = off;
3570
3571 return off;
3572}
3573
Jens Axboe2b188cc2019-01-07 10:46:33 -07003574static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
3575{
Hristo Venev75b28af2019-08-26 17:23:46 +00003576 size_t pages;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003577
Hristo Venev75b28af2019-08-26 17:23:46 +00003578 pages = (size_t)1 << get_order(
3579 rings_size(sq_entries, cq_entries, NULL));
3580 pages += (size_t)1 << get_order(
3581 array_size(sizeof(struct io_uring_sqe), sq_entries));
Jens Axboe2b188cc2019-01-07 10:46:33 -07003582
Hristo Venev75b28af2019-08-26 17:23:46 +00003583 return pages;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003584}
3585
Jens Axboeedafcce2019-01-09 09:16:05 -07003586static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
3587{
3588 int i, j;
3589
3590 if (!ctx->user_bufs)
3591 return -ENXIO;
3592
3593 for (i = 0; i < ctx->nr_user_bufs; i++) {
3594 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
3595
3596 for (j = 0; j < imu->nr_bvecs; j++)
John Hubbard27c4d3a2019-08-04 19:32:06 -07003597 put_user_page(imu->bvec[j].bv_page);
Jens Axboeedafcce2019-01-09 09:16:05 -07003598
3599 if (ctx->account_mem)
3600 io_unaccount_mem(ctx->user, imu->nr_bvecs);
Mark Rutlandd4ef6472019-05-01 16:59:16 +01003601 kvfree(imu->bvec);
Jens Axboeedafcce2019-01-09 09:16:05 -07003602 imu->nr_bvecs = 0;
3603 }
3604
3605 kfree(ctx->user_bufs);
3606 ctx->user_bufs = NULL;
3607 ctx->nr_user_bufs = 0;
3608 return 0;
3609}
3610
3611static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
3612 void __user *arg, unsigned index)
3613{
3614 struct iovec __user *src;
3615
3616#ifdef CONFIG_COMPAT
3617 if (ctx->compat) {
3618 struct compat_iovec __user *ciovs;
3619 struct compat_iovec ciov;
3620
3621 ciovs = (struct compat_iovec __user *) arg;
3622 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
3623 return -EFAULT;
3624
3625 dst->iov_base = (void __user *) (unsigned long) ciov.iov_base;
3626 dst->iov_len = ciov.iov_len;
3627 return 0;
3628 }
3629#endif
3630 src = (struct iovec __user *) arg;
3631 if (copy_from_user(dst, &src[index], sizeof(*dst)))
3632 return -EFAULT;
3633 return 0;
3634}
3635
3636static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
3637 unsigned nr_args)
3638{
3639 struct vm_area_struct **vmas = NULL;
3640 struct page **pages = NULL;
3641 int i, j, got_pages = 0;
3642 int ret = -EINVAL;
3643
3644 if (ctx->user_bufs)
3645 return -EBUSY;
3646 if (!nr_args || nr_args > UIO_MAXIOV)
3647 return -EINVAL;
3648
3649 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
3650 GFP_KERNEL);
3651 if (!ctx->user_bufs)
3652 return -ENOMEM;
3653
3654 for (i = 0; i < nr_args; i++) {
3655 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
3656 unsigned long off, start, end, ubuf;
3657 int pret, nr_pages;
3658 struct iovec iov;
3659 size_t size;
3660
3661 ret = io_copy_iov(ctx, &iov, arg, i);
3662 if (ret)
Pavel Begunkova2786822019-05-26 12:35:47 +03003663 goto err;
Jens Axboeedafcce2019-01-09 09:16:05 -07003664
3665 /*
3666 * Don't impose further limits on the size and buffer
3667 * constraints here, we'll -EINVAL later when IO is
3668 * submitted if they are wrong.
3669 */
3670 ret = -EFAULT;
3671 if (!iov.iov_base || !iov.iov_len)
3672 goto err;
3673
3674 /* arbitrary limit, but we need something */
3675 if (iov.iov_len > SZ_1G)
3676 goto err;
3677
3678 ubuf = (unsigned long) iov.iov_base;
3679 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
3680 start = ubuf >> PAGE_SHIFT;
3681 nr_pages = end - start;
3682
3683 if (ctx->account_mem) {
3684 ret = io_account_mem(ctx->user, nr_pages);
3685 if (ret)
3686 goto err;
3687 }
3688
3689 ret = 0;
3690 if (!pages || nr_pages > got_pages) {
3691 kfree(vmas);
3692 kfree(pages);
Mark Rutlandd4ef6472019-05-01 16:59:16 +01003693 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
Jens Axboeedafcce2019-01-09 09:16:05 -07003694 GFP_KERNEL);
Mark Rutlandd4ef6472019-05-01 16:59:16 +01003695 vmas = kvmalloc_array(nr_pages,
Jens Axboeedafcce2019-01-09 09:16:05 -07003696 sizeof(struct vm_area_struct *),
3697 GFP_KERNEL);
3698 if (!pages || !vmas) {
3699 ret = -ENOMEM;
3700 if (ctx->account_mem)
3701 io_unaccount_mem(ctx->user, nr_pages);
3702 goto err;
3703 }
3704 got_pages = nr_pages;
3705 }
3706
Mark Rutlandd4ef6472019-05-01 16:59:16 +01003707 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
Jens Axboeedafcce2019-01-09 09:16:05 -07003708 GFP_KERNEL);
3709 ret = -ENOMEM;
3710 if (!imu->bvec) {
3711 if (ctx->account_mem)
3712 io_unaccount_mem(ctx->user, nr_pages);
3713 goto err;
3714 }
3715
3716 ret = 0;
3717 down_read(&current->mm->mmap_sem);
Ira Weiny932f4a62019-05-13 17:17:03 -07003718 pret = get_user_pages(ubuf, nr_pages,
3719 FOLL_WRITE | FOLL_LONGTERM,
3720 pages, vmas);
Jens Axboeedafcce2019-01-09 09:16:05 -07003721 if (pret == nr_pages) {
3722 /* don't support file backed memory */
3723 for (j = 0; j < nr_pages; j++) {
3724 struct vm_area_struct *vma = vmas[j];
3725
3726 if (vma->vm_file &&
3727 !is_file_hugepages(vma->vm_file)) {
3728 ret = -EOPNOTSUPP;
3729 break;
3730 }
3731 }
3732 } else {
3733 ret = pret < 0 ? pret : -EFAULT;
3734 }
3735 up_read(&current->mm->mmap_sem);
3736 if (ret) {
3737 /*
3738 * if we did partial map, or found file backed vmas,
3739 * release any pages we did get
3740 */
John Hubbard27c4d3a2019-08-04 19:32:06 -07003741 if (pret > 0)
3742 put_user_pages(pages, pret);
Jens Axboeedafcce2019-01-09 09:16:05 -07003743 if (ctx->account_mem)
3744 io_unaccount_mem(ctx->user, nr_pages);
Mark Rutlandd4ef6472019-05-01 16:59:16 +01003745 kvfree(imu->bvec);
Jens Axboeedafcce2019-01-09 09:16:05 -07003746 goto err;
3747 }
3748
3749 off = ubuf & ~PAGE_MASK;
3750 size = iov.iov_len;
3751 for (j = 0; j < nr_pages; j++) {
3752 size_t vec_len;
3753
3754 vec_len = min_t(size_t, size, PAGE_SIZE - off);
3755 imu->bvec[j].bv_page = pages[j];
3756 imu->bvec[j].bv_len = vec_len;
3757 imu->bvec[j].bv_offset = off;
3758 off = 0;
3759 size -= vec_len;
3760 }
3761 /* store original address for later verification */
3762 imu->ubuf = ubuf;
3763 imu->len = iov.iov_len;
3764 imu->nr_bvecs = nr_pages;
3765
3766 ctx->nr_user_bufs++;
3767 }
Mark Rutlandd4ef6472019-05-01 16:59:16 +01003768 kvfree(pages);
3769 kvfree(vmas);
Jens Axboeedafcce2019-01-09 09:16:05 -07003770 return 0;
3771err:
Mark Rutlandd4ef6472019-05-01 16:59:16 +01003772 kvfree(pages);
3773 kvfree(vmas);
Jens Axboeedafcce2019-01-09 09:16:05 -07003774 io_sqe_buffer_unregister(ctx);
3775 return ret;
3776}
3777
Jens Axboe9b402842019-04-11 11:45:41 -06003778static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
3779{
3780 __s32 __user *fds = arg;
3781 int fd;
3782
3783 if (ctx->cq_ev_fd)
3784 return -EBUSY;
3785
3786 if (copy_from_user(&fd, fds, sizeof(*fds)))
3787 return -EFAULT;
3788
3789 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
3790 if (IS_ERR(ctx->cq_ev_fd)) {
3791 int ret = PTR_ERR(ctx->cq_ev_fd);
3792 ctx->cq_ev_fd = NULL;
3793 return ret;
3794 }
3795
3796 return 0;
3797}
3798
3799static int io_eventfd_unregister(struct io_ring_ctx *ctx)
3800{
3801 if (ctx->cq_ev_fd) {
3802 eventfd_ctx_put(ctx->cq_ev_fd);
3803 ctx->cq_ev_fd = NULL;
3804 return 0;
3805 }
3806
3807 return -ENXIO;
3808}
3809
Jens Axboe2b188cc2019-01-07 10:46:33 -07003810static void io_ring_ctx_free(struct io_ring_ctx *ctx)
3811{
Jens Axboe6b063142019-01-10 22:13:58 -07003812 io_finish_async(ctx);
Jens Axboe2b188cc2019-01-07 10:46:33 -07003813 if (ctx->sqo_mm)
3814 mmdrop(ctx->sqo_mm);
Jens Axboedef596e2019-01-09 08:59:42 -07003815
3816 io_iopoll_reap_events(ctx);
Jens Axboeedafcce2019-01-09 09:16:05 -07003817 io_sqe_buffer_unregister(ctx);
Jens Axboe6b063142019-01-10 22:13:58 -07003818 io_sqe_files_unregister(ctx);
Jens Axboe9b402842019-04-11 11:45:41 -06003819 io_eventfd_unregister(ctx);
Jens Axboedef596e2019-01-09 08:59:42 -07003820
Jens Axboe2b188cc2019-01-07 10:46:33 -07003821#if defined(CONFIG_UNIX)
Eric Biggers355e8d22019-06-12 14:58:43 -07003822 if (ctx->ring_sock) {
3823 ctx->ring_sock->file = NULL; /* so that iput() is called */
Jens Axboe2b188cc2019-01-07 10:46:33 -07003824 sock_release(ctx->ring_sock);
Eric Biggers355e8d22019-06-12 14:58:43 -07003825 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07003826#endif
3827
Hristo Venev75b28af2019-08-26 17:23:46 +00003828 io_mem_free(ctx->rings);
Jens Axboe2b188cc2019-01-07 10:46:33 -07003829 io_mem_free(ctx->sq_sqes);
Jens Axboe2b188cc2019-01-07 10:46:33 -07003830
3831 percpu_ref_exit(&ctx->refs);
3832 if (ctx->account_mem)
3833 io_unaccount_mem(ctx->user,
3834 ring_pages(ctx->sq_entries, ctx->cq_entries));
3835 free_uid(ctx->user);
3836 kfree(ctx);
3837}
3838
3839static __poll_t io_uring_poll(struct file *file, poll_table *wait)
3840{
3841 struct io_ring_ctx *ctx = file->private_data;
3842 __poll_t mask = 0;
3843
3844 poll_wait(file, &ctx->cq_wait, wait);
Stefan Bühler4f7067c2019-04-24 23:54:17 +02003845 /*
3846 * synchronizes with barrier from wq_has_sleeper call in
3847 * io_commit_cqring
3848 */
Jens Axboe2b188cc2019-01-07 10:46:33 -07003849 smp_rmb();
Hristo Venev75b28af2019-08-26 17:23:46 +00003850 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
3851 ctx->rings->sq_ring_entries)
Jens Axboe2b188cc2019-01-07 10:46:33 -07003852 mask |= EPOLLOUT | EPOLLWRNORM;
yangerkundaa5de52019-09-24 20:53:34 +08003853 if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail)
Jens Axboe2b188cc2019-01-07 10:46:33 -07003854 mask |= EPOLLIN | EPOLLRDNORM;
3855
3856 return mask;
3857}
3858
3859static int io_uring_fasync(int fd, struct file *file, int on)
3860{
3861 struct io_ring_ctx *ctx = file->private_data;
3862
3863 return fasync_helper(fd, file, on, &ctx->cq_fasync);
3864}
3865
3866static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
3867{
3868 mutex_lock(&ctx->uring_lock);
3869 percpu_ref_kill(&ctx->refs);
3870 mutex_unlock(&ctx->uring_lock);
3871
Jens Axboe5262f562019-09-17 12:26:57 -06003872 io_kill_timeouts(ctx);
Jens Axboe221c5eb2019-01-17 09:41:58 -07003873 io_poll_remove_all(ctx);
Jens Axboe561fb042019-10-24 07:25:42 -06003874
3875 if (ctx->io_wq)
3876 io_wq_cancel_all(ctx->io_wq);
3877
Jens Axboedef596e2019-01-09 08:59:42 -07003878 io_iopoll_reap_events(ctx);
Jens Axboe2b188cc2019-01-07 10:46:33 -07003879 wait_for_completion(&ctx->ctx_done);
3880 io_ring_ctx_free(ctx);
3881}
3882
3883static int io_uring_release(struct inode *inode, struct file *file)
3884{
3885 struct io_ring_ctx *ctx = file->private_data;
3886
3887 file->private_data = NULL;
3888 io_ring_ctx_wait_and_kill(ctx);
3889 return 0;
3890}
3891
Jens Axboefcb323c2019-10-24 12:39:47 -06003892static void io_uring_cancel_files(struct io_ring_ctx *ctx,
3893 struct files_struct *files)
3894{
3895 struct io_kiocb *req;
3896 DEFINE_WAIT(wait);
3897
3898 while (!list_empty_careful(&ctx->inflight_list)) {
3899 enum io_wq_cancel ret = IO_WQ_CANCEL_NOTFOUND;
3900
3901 spin_lock_irq(&ctx->inflight_lock);
3902 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
3903 if (req->work.files == files) {
3904 ret = io_wq_cancel_work(ctx->io_wq, &req->work);
3905 break;
3906 }
3907 }
3908 if (ret == IO_WQ_CANCEL_RUNNING)
3909 prepare_to_wait(&ctx->inflight_wait, &wait,
3910 TASK_UNINTERRUPTIBLE);
3911
3912 spin_unlock_irq(&ctx->inflight_lock);
3913
3914 /*
3915 * We need to keep going until we get NOTFOUND. We only cancel
3916 * one work at the time.
3917 *
3918 * If we get CANCEL_RUNNING, then wait for a work to complete
3919 * before continuing.
3920 */
3921 if (ret == IO_WQ_CANCEL_OK)
3922 continue;
3923 else if (ret != IO_WQ_CANCEL_RUNNING)
3924 break;
3925 schedule();
3926 }
3927}
3928
3929static int io_uring_flush(struct file *file, void *data)
3930{
3931 struct io_ring_ctx *ctx = file->private_data;
3932
3933 io_uring_cancel_files(ctx, data);
3934 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
3935 io_wq_cancel_all(ctx->io_wq);
3936 return 0;
3937}
3938
Jens Axboe2b188cc2019-01-07 10:46:33 -07003939static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3940{
3941 loff_t offset = (loff_t) vma->vm_pgoff << PAGE_SHIFT;
3942 unsigned long sz = vma->vm_end - vma->vm_start;
3943 struct io_ring_ctx *ctx = file->private_data;
3944 unsigned long pfn;
3945 struct page *page;
3946 void *ptr;
3947
3948 switch (offset) {
3949 case IORING_OFF_SQ_RING:
Hristo Venev75b28af2019-08-26 17:23:46 +00003950 case IORING_OFF_CQ_RING:
3951 ptr = ctx->rings;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003952 break;
3953 case IORING_OFF_SQES:
3954 ptr = ctx->sq_sqes;
3955 break;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003956 default:
3957 return -EINVAL;
3958 }
3959
3960 page = virt_to_head_page(ptr);
Matthew Wilcox (Oracle)a50b8542019-09-23 15:34:25 -07003961 if (sz > page_size(page))
Jens Axboe2b188cc2019-01-07 10:46:33 -07003962 return -EINVAL;
3963
3964 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3965 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3966}
3967
3968SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3969 u32, min_complete, u32, flags, const sigset_t __user *, sig,
3970 size_t, sigsz)
3971{
3972 struct io_ring_ctx *ctx;
3973 long ret = -EBADF;
3974 int submitted = 0;
3975 struct fd f;
3976
Jens Axboe6c271ce2019-01-10 11:22:30 -07003977 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
Jens Axboe2b188cc2019-01-07 10:46:33 -07003978 return -EINVAL;
3979
3980 f = fdget(fd);
3981 if (!f.file)
3982 return -EBADF;
3983
3984 ret = -EOPNOTSUPP;
3985 if (f.file->f_op != &io_uring_fops)
3986 goto out_fput;
3987
3988 ret = -ENXIO;
3989 ctx = f.file->private_data;
3990 if (!percpu_ref_tryget(&ctx->refs))
3991 goto out_fput;
3992
Jens Axboe6c271ce2019-01-10 11:22:30 -07003993 /*
3994 * For SQ polling, the thread will do all submissions and completions.
3995 * Just return the requested submit count, and wake the thread if
3996 * we were asked to.
3997 */
Jens Axboeb2a9ead2019-09-12 14:19:16 -06003998 ret = 0;
Jens Axboe6c271ce2019-01-10 11:22:30 -07003999 if (ctx->flags & IORING_SETUP_SQPOLL) {
4000 if (flags & IORING_ENTER_SQ_WAKEUP)
4001 wake_up(&ctx->sqo_wait);
4002 submitted = to_submit;
Jens Axboeb2a9ead2019-09-12 14:19:16 -06004003 } else if (to_submit) {
Pavel Begunkovae9428c2019-11-06 00:22:14 +03004004 struct mm_struct *cur_mm;
Jens Axboe2b188cc2019-01-07 10:46:33 -07004005
Pavel Begunkovae9428c2019-11-06 00:22:14 +03004006 to_submit = min(to_submit, ctx->sq_entries);
Jens Axboe2b188cc2019-01-07 10:46:33 -07004007 mutex_lock(&ctx->uring_lock);
Pavel Begunkovae9428c2019-11-06 00:22:14 +03004008 /* already have mm, so io_submit_sqes() won't try to grab it */
4009 cur_mm = ctx->sqo_mm;
4010 submitted = io_submit_sqes(ctx, to_submit, f.file, fd,
4011 &cur_mm, false);
Jens Axboe2b188cc2019-01-07 10:46:33 -07004012 mutex_unlock(&ctx->uring_lock);
Jens Axboe2b188cc2019-01-07 10:46:33 -07004013 }
4014 if (flags & IORING_ENTER_GETEVENTS) {
Jens Axboedef596e2019-01-09 08:59:42 -07004015 unsigned nr_events = 0;
4016
Jens Axboe2b188cc2019-01-07 10:46:33 -07004017 min_complete = min(min_complete, ctx->cq_entries);
4018
Jens Axboedef596e2019-01-09 08:59:42 -07004019 if (ctx->flags & IORING_SETUP_IOPOLL) {
Jens Axboedef596e2019-01-09 08:59:42 -07004020 ret = io_iopoll_check(ctx, &nr_events, min_complete);
Jens Axboedef596e2019-01-09 08:59:42 -07004021 } else {
4022 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
4023 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07004024 }
4025
Pavel Begunkov6805b322019-10-08 02:18:42 +03004026 percpu_ref_put(&ctx->refs);
Jens Axboe2b188cc2019-01-07 10:46:33 -07004027out_fput:
4028 fdput(f);
4029 return submitted ? submitted : ret;
4030}
4031
4032static const struct file_operations io_uring_fops = {
4033 .release = io_uring_release,
Jens Axboefcb323c2019-10-24 12:39:47 -06004034 .flush = io_uring_flush,
Jens Axboe2b188cc2019-01-07 10:46:33 -07004035 .mmap = io_uring_mmap,
4036 .poll = io_uring_poll,
4037 .fasync = io_uring_fasync,
4038};
4039
4040static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
4041 struct io_uring_params *p)
4042{
Hristo Venev75b28af2019-08-26 17:23:46 +00004043 struct io_rings *rings;
4044 size_t size, sq_array_offset;
Jens Axboe2b188cc2019-01-07 10:46:33 -07004045
Hristo Venev75b28af2019-08-26 17:23:46 +00004046 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
4047 if (size == SIZE_MAX)
4048 return -EOVERFLOW;
4049
4050 rings = io_mem_alloc(size);
4051 if (!rings)
Jens Axboe2b188cc2019-01-07 10:46:33 -07004052 return -ENOMEM;
4053
Hristo Venev75b28af2019-08-26 17:23:46 +00004054 ctx->rings = rings;
4055 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
4056 rings->sq_ring_mask = p->sq_entries - 1;
4057 rings->cq_ring_mask = p->cq_entries - 1;
4058 rings->sq_ring_entries = p->sq_entries;
4059 rings->cq_ring_entries = p->cq_entries;
4060 ctx->sq_mask = rings->sq_ring_mask;
4061 ctx->cq_mask = rings->cq_ring_mask;
4062 ctx->sq_entries = rings->sq_ring_entries;
4063 ctx->cq_entries = rings->cq_ring_entries;
Jens Axboe2b188cc2019-01-07 10:46:33 -07004064
4065 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
4066 if (size == SIZE_MAX)
4067 return -EOVERFLOW;
4068
4069 ctx->sq_sqes = io_mem_alloc(size);
Mark Rutland52e04ef2019-04-30 17:30:21 +01004070 if (!ctx->sq_sqes)
Jens Axboe2b188cc2019-01-07 10:46:33 -07004071 return -ENOMEM;
Jens Axboe2b188cc2019-01-07 10:46:33 -07004072
Jens Axboe2b188cc2019-01-07 10:46:33 -07004073 return 0;
4074}
4075
4076/*
4077 * Allocate an anonymous fd, this is what constitutes the application
4078 * visible backing of an io_uring instance. The application mmaps this
4079 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
4080 * we have to tie this fd to a socket for file garbage collection purposes.
4081 */
4082static int io_uring_get_fd(struct io_ring_ctx *ctx)
4083{
4084 struct file *file;
4085 int ret;
4086
4087#if defined(CONFIG_UNIX)
4088 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
4089 &ctx->ring_sock);
4090 if (ret)
4091 return ret;
4092#endif
4093
4094 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
4095 if (ret < 0)
4096 goto err;
4097
4098 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
4099 O_RDWR | O_CLOEXEC);
4100 if (IS_ERR(file)) {
4101 put_unused_fd(ret);
4102 ret = PTR_ERR(file);
4103 goto err;
4104 }
4105
4106#if defined(CONFIG_UNIX)
4107 ctx->ring_sock->file = file;
Jens Axboe6b063142019-01-10 22:13:58 -07004108 ctx->ring_sock->sk->sk_user_data = ctx;
Jens Axboe2b188cc2019-01-07 10:46:33 -07004109#endif
4110 fd_install(ret, file);
4111 return ret;
4112err:
4113#if defined(CONFIG_UNIX)
4114 sock_release(ctx->ring_sock);
4115 ctx->ring_sock = NULL;
4116#endif
4117 return ret;
4118}
4119
4120static int io_uring_create(unsigned entries, struct io_uring_params *p)
4121{
4122 struct user_struct *user = NULL;
4123 struct io_ring_ctx *ctx;
4124 bool account_mem;
4125 int ret;
4126
4127 if (!entries || entries > IORING_MAX_ENTRIES)
4128 return -EINVAL;
4129
4130 /*
4131 * Use twice as many entries for the CQ ring. It's possible for the
4132 * application to drive a higher depth than the size of the SQ ring,
4133 * since the sqes are only used at submission time. This allows for
Jens Axboe33a107f2019-10-04 12:10:03 -06004134 * some flexibility in overcommitting a bit. If the application has
4135 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
4136 * of CQ ring entries manually.
Jens Axboe2b188cc2019-01-07 10:46:33 -07004137 */
4138 p->sq_entries = roundup_pow_of_two(entries);
Jens Axboe33a107f2019-10-04 12:10:03 -06004139 if (p->flags & IORING_SETUP_CQSIZE) {
4140 /*
4141 * If IORING_SETUP_CQSIZE is set, we do the same roundup
4142 * to a power-of-two, if it isn't already. We do NOT impose
4143 * any cq vs sq ring sizing.
4144 */
4145 if (p->cq_entries < p->sq_entries || p->cq_entries > IORING_MAX_CQ_ENTRIES)
4146 return -EINVAL;
4147 p->cq_entries = roundup_pow_of_two(p->cq_entries);
4148 } else {
4149 p->cq_entries = 2 * p->sq_entries;
4150 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07004151
4152 user = get_uid(current_user());
4153 account_mem = !capable(CAP_IPC_LOCK);
4154
4155 if (account_mem) {
4156 ret = io_account_mem(user,
4157 ring_pages(p->sq_entries, p->cq_entries));
4158 if (ret) {
4159 free_uid(user);
4160 return ret;
4161 }
4162 }
4163
4164 ctx = io_ring_ctx_alloc(p);
4165 if (!ctx) {
4166 if (account_mem)
4167 io_unaccount_mem(user, ring_pages(p->sq_entries,
4168 p->cq_entries));
4169 free_uid(user);
4170 return -ENOMEM;
4171 }
4172 ctx->compat = in_compat_syscall();
4173 ctx->account_mem = account_mem;
4174 ctx->user = user;
4175
4176 ret = io_allocate_scq_urings(ctx, p);
4177 if (ret)
4178 goto err;
4179
Jens Axboe6c271ce2019-01-10 11:22:30 -07004180 ret = io_sq_offload_start(ctx, p);
Jens Axboe2b188cc2019-01-07 10:46:33 -07004181 if (ret)
4182 goto err;
4183
Jens Axboe2b188cc2019-01-07 10:46:33 -07004184 memset(&p->sq_off, 0, sizeof(p->sq_off));
Hristo Venev75b28af2019-08-26 17:23:46 +00004185 p->sq_off.head = offsetof(struct io_rings, sq.head);
4186 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
4187 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
4188 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
4189 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
4190 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
4191 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
Jens Axboe2b188cc2019-01-07 10:46:33 -07004192
4193 memset(&p->cq_off, 0, sizeof(p->cq_off));
Hristo Venev75b28af2019-08-26 17:23:46 +00004194 p->cq_off.head = offsetof(struct io_rings, cq.head);
4195 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
4196 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
4197 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
4198 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
4199 p->cq_off.cqes = offsetof(struct io_rings, cqes);
Jens Axboeac90f242019-09-06 10:26:21 -06004200
Jens Axboe044c1ab2019-10-28 09:15:33 -06004201 /*
4202 * Install ring fd as the very last thing, so we don't risk someone
4203 * having closed it before we finish setup
4204 */
4205 ret = io_uring_get_fd(ctx);
4206 if (ret < 0)
4207 goto err;
4208
Jens Axboeac90f242019-09-06 10:26:21 -06004209 p->features = IORING_FEAT_SINGLE_MMAP;
Dmitrii Dolgovc826bd72019-10-15 19:02:01 +02004210 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
Jens Axboe2b188cc2019-01-07 10:46:33 -07004211 return ret;
4212err:
4213 io_ring_ctx_wait_and_kill(ctx);
4214 return ret;
4215}
4216
4217/*
4218 * Sets up an aio uring context, and returns the fd. Applications asks for a
4219 * ring size, we return the actual sq/cq ring sizes (among other things) in the
4220 * params structure passed in.
4221 */
4222static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
4223{
4224 struct io_uring_params p;
4225 long ret;
4226 int i;
4227
4228 if (copy_from_user(&p, params, sizeof(p)))
4229 return -EFAULT;
4230 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
4231 if (p.resv[i])
4232 return -EINVAL;
4233 }
4234
Jens Axboe6c271ce2019-01-10 11:22:30 -07004235 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
Jens Axboe33a107f2019-10-04 12:10:03 -06004236 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE))
Jens Axboe2b188cc2019-01-07 10:46:33 -07004237 return -EINVAL;
4238
4239 ret = io_uring_create(entries, &p);
4240 if (ret < 0)
4241 return ret;
4242
4243 if (copy_to_user(params, &p, sizeof(p)))
4244 return -EFAULT;
4245
4246 return ret;
4247}
4248
4249SYSCALL_DEFINE2(io_uring_setup, u32, entries,
4250 struct io_uring_params __user *, params)
4251{
4252 return io_uring_setup(entries, params);
4253}
4254
Jens Axboeedafcce2019-01-09 09:16:05 -07004255static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
4256 void __user *arg, unsigned nr_args)
Jens Axboeb19062a2019-04-15 10:49:38 -06004257 __releases(ctx->uring_lock)
4258 __acquires(ctx->uring_lock)
Jens Axboeedafcce2019-01-09 09:16:05 -07004259{
4260 int ret;
4261
Jens Axboe35fa71a2019-04-22 10:23:23 -06004262 /*
4263 * We're inside the ring mutex, if the ref is already dying, then
4264 * someone else killed the ctx or is already going through
4265 * io_uring_register().
4266 */
4267 if (percpu_ref_is_dying(&ctx->refs))
4268 return -ENXIO;
4269
Jens Axboeedafcce2019-01-09 09:16:05 -07004270 percpu_ref_kill(&ctx->refs);
Jens Axboeb19062a2019-04-15 10:49:38 -06004271
4272 /*
4273 * Drop uring mutex before waiting for references to exit. If another
4274 * thread is currently inside io_uring_enter() it might need to grab
4275 * the uring_lock to make progress. If we hold it here across the drain
4276 * wait, then we can deadlock. It's safe to drop the mutex here, since
4277 * no new references will come in after we've killed the percpu ref.
4278 */
4279 mutex_unlock(&ctx->uring_lock);
Jens Axboeedafcce2019-01-09 09:16:05 -07004280 wait_for_completion(&ctx->ctx_done);
Jens Axboeb19062a2019-04-15 10:49:38 -06004281 mutex_lock(&ctx->uring_lock);
Jens Axboeedafcce2019-01-09 09:16:05 -07004282
4283 switch (opcode) {
4284 case IORING_REGISTER_BUFFERS:
4285 ret = io_sqe_buffer_register(ctx, arg, nr_args);
4286 break;
4287 case IORING_UNREGISTER_BUFFERS:
4288 ret = -EINVAL;
4289 if (arg || nr_args)
4290 break;
4291 ret = io_sqe_buffer_unregister(ctx);
4292 break;
Jens Axboe6b063142019-01-10 22:13:58 -07004293 case IORING_REGISTER_FILES:
4294 ret = io_sqe_files_register(ctx, arg, nr_args);
4295 break;
4296 case IORING_UNREGISTER_FILES:
4297 ret = -EINVAL;
4298 if (arg || nr_args)
4299 break;
4300 ret = io_sqe_files_unregister(ctx);
4301 break;
Jens Axboec3a31e62019-10-03 13:59:56 -06004302 case IORING_REGISTER_FILES_UPDATE:
4303 ret = io_sqe_files_update(ctx, arg, nr_args);
4304 break;
Jens Axboe9b402842019-04-11 11:45:41 -06004305 case IORING_REGISTER_EVENTFD:
4306 ret = -EINVAL;
4307 if (nr_args != 1)
4308 break;
4309 ret = io_eventfd_register(ctx, arg);
4310 break;
4311 case IORING_UNREGISTER_EVENTFD:
4312 ret = -EINVAL;
4313 if (arg || nr_args)
4314 break;
4315 ret = io_eventfd_unregister(ctx);
4316 break;
Jens Axboeedafcce2019-01-09 09:16:05 -07004317 default:
4318 ret = -EINVAL;
4319 break;
4320 }
4321
4322 /* bring the ctx back to life */
4323 reinit_completion(&ctx->ctx_done);
4324 percpu_ref_reinit(&ctx->refs);
4325 return ret;
4326}
4327
4328SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
4329 void __user *, arg, unsigned int, nr_args)
4330{
4331 struct io_ring_ctx *ctx;
4332 long ret = -EBADF;
4333 struct fd f;
4334
4335 f = fdget(fd);
4336 if (!f.file)
4337 return -EBADF;
4338
4339 ret = -EOPNOTSUPP;
4340 if (f.file->f_op != &io_uring_fops)
4341 goto out_fput;
4342
4343 ctx = f.file->private_data;
4344
4345 mutex_lock(&ctx->uring_lock);
4346 ret = __io_uring_register(ctx, opcode, arg, nr_args);
4347 mutex_unlock(&ctx->uring_lock);
Dmitrii Dolgovc826bd72019-10-15 19:02:01 +02004348 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
4349 ctx->cq_ev_fd != NULL, ret);
Jens Axboeedafcce2019-01-09 09:16:05 -07004350out_fput:
4351 fdput(f);
4352 return ret;
4353}
4354
Jens Axboe2b188cc2019-01-07 10:46:33 -07004355static int __init io_uring_init(void)
4356{
4357 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
4358 return 0;
4359};
4360__initcall(io_uring_init);