Commit | Line | Data |
---|---|---|
2b188cc1 JA |
1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* | |
3 | * Shared application/kernel submission and completion ring pairs, for | |
4 | * supporting fast/efficient IO. | |
5 | * | |
6 | * A note on the read/write ordering memory barriers that are matched between | |
1e84b97b SB |
7 | * the application and kernel side. |
8 | * | |
9 | * After the application reads the CQ ring tail, it must use an | |
10 | * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses | |
11 | * before writing the tail (using smp_load_acquire to read the tail will | |
12 | * do). It also needs a smp_mb() before updating CQ head (ordering the | |
13 | * entry load(s) with the head store), pairing with an implicit barrier | |
14 | * through a control-dependency in io_get_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. | |
2b188cc1 JA |
29 | * |
30 | * Also see the examples in the liburing library: | |
31 | * | |
32 | * git://git.kernel.dk/liburing | |
33 | * | |
34 | * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens | |
35 | * from data shared between the kernel and application. This is done both | |
36 | * for ordering purposes, but also to ensure that once a value is loaded from | |
37 | * data that the application could potentially modify, it remains stable. | |
38 | * | |
39 | * Copyright (C) 2018-2019 Jens Axboe | |
c992fe29 | 40 | * Copyright (c) 2018-2019 Christoph Hellwig |
2b188cc1 JA |
41 | */ |
42 | #include <linux/kernel.h> | |
43 | #include <linux/init.h> | |
44 | #include <linux/errno.h> | |
45 | #include <linux/syscalls.h> | |
46 | #include <linux/compat.h> | |
47 | #include <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> | |
6c271ce2 | 59 | #include <linux/kthread.h> |
2b188cc1 | 60 | #include <linux/blkdev.h> |
edafccee | 61 | #include <linux/bvec.h> |
2b188cc1 JA |
62 | #include <linux/net.h> |
63 | #include <net/sock.h> | |
64 | #include <net/af_unix.h> | |
6b06314c | 65 | #include <net/scm.h> |
2b188cc1 JA |
66 | #include <linux/anon_inodes.h> |
67 | #include <linux/sched/mm.h> | |
68 | #include <linux/uaccess.h> | |
69 | #include <linux/nospec.h> | |
edafccee JA |
70 | #include <linux/sizes.h> |
71 | #include <linux/hugetlb.h> | |
2b188cc1 | 72 | |
c826bd7a DD |
73 | #define CREATE_TRACE_POINTS |
74 | #include <trace/events/io_uring.h> | |
75 | ||
2b188cc1 JA |
76 | #include <uapi/linux/io_uring.h> |
77 | ||
78 | #include "internal.h" | |
561fb04a | 79 | #include "io-wq.h" |
2b188cc1 | 80 | |
5277deaa | 81 | #define IORING_MAX_ENTRIES 32768 |
33a107f0 | 82 | #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES) |
65e19f54 JA |
83 | |
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) | |
2b188cc1 JA |
91 | |
92 | struct io_uring { | |
93 | u32 head ____cacheline_aligned_in_smp; | |
94 | u32 tail ____cacheline_aligned_in_smp; | |
95 | }; | |
96 | ||
1e84b97b | 97 | /* |
75b28aff HV |
98 | * This data is shared with the application through the mmap at offsets |
99 | * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING. | |
1e84b97b SB |
100 | * |
101 | * The offsets to the member fields are published through struct | |
102 | * io_sqring_offsets when calling io_uring_setup. | |
103 | */ | |
75b28aff | 104 | struct io_rings { |
1e84b97b SB |
105 | /* |
106 | * Head and tail offsets into the ring; the offsets need to be | |
107 | * masked to get valid indices. | |
108 | * | |
75b28aff HV |
109 | * 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. | |
1e84b97b | 112 | */ |
75b28aff | 113 | struct io_uring sq, cq; |
1e84b97b | 114 | /* |
75b28aff | 115 | * Bitmasks to apply to head and tail offsets (constant, equals |
1e84b97b SB |
116 | * ring_entries - 1) |
117 | */ | |
75b28aff HV |
118 | u32 sq_ring_mask, cq_ring_mask; |
119 | /* Ring sizes (constant, power of 2) */ | |
120 | u32 sq_ring_entries, cq_ring_entries; | |
1e84b97b SB |
121 | /* |
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 | */ | |
75b28aff | 133 | u32 sq_dropped; |
1e84b97b SB |
134 | /* |
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 | */ | |
75b28aff | 143 | u32 sq_flags; |
1e84b97b SB |
144 | /* |
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 | */ | |
75b28aff | 157 | u32 cq_overflow; |
1e84b97b SB |
158 | /* |
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 | */ | |
75b28aff | 165 | struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp; |
2b188cc1 JA |
166 | }; |
167 | ||
edafccee JA |
168 | struct io_mapped_ubuf { |
169 | u64 ubuf; | |
170 | size_t len; | |
171 | struct bio_vec *bvec; | |
172 | unsigned int nr_bvecs; | |
173 | }; | |
174 | ||
65e19f54 JA |
175 | struct fixed_file_table { |
176 | struct file **files; | |
177 | }; | |
178 | ||
2b188cc1 JA |
179 | struct 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; | |
1d7bb1d5 | 188 | bool cq_overflow_flushed; |
2b188cc1 | 189 | |
75b28aff HV |
190 | /* |
191 | * Ring buffer of indices into array of io_uring_sqe, which is | |
192 | * mmapped by the application using the IORING_OFF_SQES offset. | |
193 | * | |
194 | * This indirection could e.g. be used to assign fixed | |
195 | * io_uring_sqe entries to operations and only submit them to | |
196 | * the queue when needed. | |
197 | * | |
198 | * The kernel modifies neither the indices array nor the entries | |
199 | * array. | |
200 | */ | |
201 | u32 *sq_array; | |
2b188cc1 JA |
202 | unsigned cached_sq_head; |
203 | unsigned sq_entries; | |
204 | unsigned sq_mask; | |
6c271ce2 | 205 | unsigned sq_thread_idle; |
498ccd9e | 206 | unsigned cached_sq_dropped; |
206aefde | 207 | atomic_t cached_cq_overflow; |
2b188cc1 | 208 | struct io_uring_sqe *sq_sqes; |
de0617e4 JA |
209 | |
210 | struct list_head defer_list; | |
5262f567 | 211 | struct list_head timeout_list; |
1d7bb1d5 | 212 | struct list_head cq_overflow_list; |
fcb323cc JA |
213 | |
214 | wait_queue_head_t inflight_wait; | |
2b188cc1 JA |
215 | } ____cacheline_aligned_in_smp; |
216 | ||
206aefde JA |
217 | struct io_rings *rings; |
218 | ||
2b188cc1 | 219 | /* IO offload */ |
561fb04a | 220 | struct io_wq *io_wq; |
6c271ce2 | 221 | struct task_struct *sqo_thread; /* if using sq thread polling */ |
2b188cc1 | 222 | struct mm_struct *sqo_mm; |
6c271ce2 | 223 | wait_queue_head_t sqo_wait; |
75b28aff | 224 | |
6b06314c JA |
225 | /* |
226 | * If used, fixed file set. Writers must ensure that ->refs is dead, | |
227 | * readers must ensure that ->refs is alive as long as the file* is | |
228 | * used. Only updated through io_uring_register(2). | |
229 | */ | |
65e19f54 | 230 | struct fixed_file_table *file_table; |
6b06314c JA |
231 | unsigned nr_user_files; |
232 | ||
edafccee JA |
233 | /* if used, fixed mapped user buffers */ |
234 | unsigned nr_user_bufs; | |
235 | struct io_mapped_ubuf *user_bufs; | |
236 | ||
2b188cc1 JA |
237 | struct user_struct *user; |
238 | ||
206aefde JA |
239 | /* 0 is for ctx quiesce/reinit/free, 1 is for sqo_thread started */ |
240 | struct completion *completions; | |
241 | ||
0ddf92e8 JA |
242 | /* if all else fails... */ |
243 | struct io_kiocb *fallback_req; | |
244 | ||
206aefde JA |
245 | #if defined(CONFIG_UNIX) |
246 | struct socket *ring_sock; | |
247 | #endif | |
248 | ||
249 | struct { | |
250 | unsigned cached_cq_tail; | |
251 | unsigned cq_entries; | |
252 | unsigned cq_mask; | |
253 | atomic_t cq_timeouts; | |
254 | struct wait_queue_head cq_wait; | |
255 | struct fasync_struct *cq_fasync; | |
256 | struct eventfd_ctx *cq_ev_fd; | |
257 | } ____cacheline_aligned_in_smp; | |
2b188cc1 JA |
258 | |
259 | struct { | |
260 | struct mutex uring_lock; | |
261 | wait_queue_head_t wait; | |
262 | } ____cacheline_aligned_in_smp; | |
263 | ||
264 | struct { | |
265 | spinlock_t completion_lock; | |
def596e9 JA |
266 | bool poll_multi_file; |
267 | /* | |
268 | * ->poll_list is protected by the ctx->uring_lock for | |
269 | * io_uring instances that don't use IORING_SETUP_SQPOLL. | |
270 | * For SQPOLL, only the single threaded io_sq_thread() will | |
271 | * manipulate the list, hence no extra locking is needed there. | |
272 | */ | |
273 | struct list_head poll_list; | |
221c5eb2 | 274 | struct list_head cancel_list; |
fcb323cc JA |
275 | |
276 | spinlock_t inflight_lock; | |
277 | struct list_head inflight_list; | |
2b188cc1 | 278 | } ____cacheline_aligned_in_smp; |
2b188cc1 JA |
279 | }; |
280 | ||
281 | struct sqe_submit { | |
282 | const struct io_uring_sqe *sqe; | |
fcb323cc JA |
283 | struct file *ring_file; |
284 | int ring_fd; | |
8776f3fa | 285 | u32 sequence; |
2b188cc1 | 286 | bool has_user; |
ba5290cc | 287 | bool in_async; |
6c271ce2 | 288 | bool needs_fixed_file; |
2b188cc1 JA |
289 | }; |
290 | ||
09bb8394 JA |
291 | /* |
292 | * First field must be the file pointer in all the | |
293 | * iocb unions! See also 'struct kiocb' in <linux/fs.h> | |
294 | */ | |
221c5eb2 JA |
295 | struct io_poll_iocb { |
296 | struct file *file; | |
297 | struct wait_queue_head *head; | |
298 | __poll_t events; | |
8c838788 | 299 | bool done; |
221c5eb2 JA |
300 | bool canceled; |
301 | struct wait_queue_entry wait; | |
302 | }; | |
303 | ||
5262f567 JA |
304 | struct io_timeout { |
305 | struct file *file; | |
306 | struct hrtimer timer; | |
307 | }; | |
308 | ||
09bb8394 JA |
309 | /* |
310 | * NOTE! Each of the iocb union members has the file pointer | |
311 | * as the first entry in their struct definition. So you can | |
312 | * access the file pointer through any of the sub-structs, | |
313 | * or directly as just 'ki_filp' in this struct. | |
314 | */ | |
2b188cc1 | 315 | struct io_kiocb { |
221c5eb2 | 316 | union { |
09bb8394 | 317 | struct file *file; |
221c5eb2 JA |
318 | struct kiocb rw; |
319 | struct io_poll_iocb poll; | |
5262f567 | 320 | struct io_timeout timeout; |
221c5eb2 | 321 | }; |
2b188cc1 JA |
322 | |
323 | struct sqe_submit submit; | |
324 | ||
325 | struct io_ring_ctx *ctx; | |
326 | struct list_head list; | |
9e645e11 | 327 | struct list_head link_list; |
2b188cc1 | 328 | unsigned int flags; |
c16361c1 | 329 | refcount_t refs; |
8449eeda | 330 | #define REQ_F_NOWAIT 1 /* must not punt to workers */ |
def596e9 | 331 | #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */ |
6b06314c | 332 | #define REQ_F_FIXED_FILE 4 /* ctx owns file */ |
31b51510 | 333 | #define REQ_F_SEQ_PREV 8 /* sequential with previous */ |
e2033e33 SB |
334 | #define REQ_F_IO_DRAIN 16 /* drain existing IO first */ |
335 | #define REQ_F_IO_DRAINED 32 /* drain done */ | |
9e645e11 | 336 | #define REQ_F_LINK 64 /* linked sqes */ |
2665abfd | 337 | #define REQ_F_LINK_TIMEOUT 128 /* has linked timeout */ |
f7b76ac9 | 338 | #define REQ_F_FAIL_LINK 256 /* fail rest of links */ |
4fe2c963 | 339 | #define REQ_F_SHADOW_DRAIN 512 /* link-drain shadow req */ |
5262f567 | 340 | #define REQ_F_TIMEOUT 1024 /* timeout request */ |
491381ce JA |
341 | #define REQ_F_ISREG 2048 /* regular file */ |
342 | #define REQ_F_MUST_PUNT 4096 /* must be punted even for NONBLOCK */ | |
fcb323cc | 343 | #define REQ_F_INFLIGHT 8192 /* on inflight list */ |
2b188cc1 | 344 | u64 user_data; |
9e645e11 | 345 | u32 result; |
de0617e4 | 346 | u32 sequence; |
2b188cc1 | 347 | |
fcb323cc JA |
348 | struct list_head inflight_entry; |
349 | ||
561fb04a | 350 | struct io_wq_work work; |
2b188cc1 JA |
351 | }; |
352 | ||
353 | #define IO_PLUG_THRESHOLD 2 | |
def596e9 | 354 | #define IO_IOPOLL_BATCH 8 |
2b188cc1 | 355 | |
9a56a232 JA |
356 | struct io_submit_state { |
357 | struct blk_plug plug; | |
358 | ||
2579f913 JA |
359 | /* |
360 | * io_kiocb alloc cache | |
361 | */ | |
362 | void *reqs[IO_IOPOLL_BATCH]; | |
363 | unsigned int free_reqs; | |
364 | unsigned int cur_req; | |
365 | ||
9a56a232 JA |
366 | /* |
367 | * File reference cache | |
368 | */ | |
369 | struct file *file; | |
370 | unsigned int fd; | |
371 | unsigned int has_refs; | |
372 | unsigned int used_refs; | |
373 | unsigned int ios_left; | |
374 | }; | |
375 | ||
561fb04a | 376 | static void io_wq_submit_work(struct io_wq_work **workptr); |
78e19bbe | 377 | static void io_cqring_fill_event(struct io_kiocb *req, long res); |
4fe2c963 | 378 | static void __io_free_req(struct io_kiocb *req); |
ec9c02ad | 379 | static void io_put_req(struct io_kiocb *req); |
78e19bbe | 380 | static void io_double_put_req(struct io_kiocb *req); |
de0617e4 | 381 | |
2b188cc1 JA |
382 | static struct kmem_cache *req_cachep; |
383 | ||
384 | static const struct file_operations io_uring_fops; | |
385 | ||
386 | struct sock *io_uring_get_socket(struct file *file) | |
387 | { | |
388 | #if defined(CONFIG_UNIX) | |
389 | if (file->f_op == &io_uring_fops) { | |
390 | struct io_ring_ctx *ctx = file->private_data; | |
391 | ||
392 | return ctx->ring_sock->sk; | |
393 | } | |
394 | #endif | |
395 | return NULL; | |
396 | } | |
397 | EXPORT_SYMBOL(io_uring_get_socket); | |
398 | ||
399 | static void io_ring_ctx_ref_free(struct percpu_ref *ref) | |
400 | { | |
401 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
402 | ||
206aefde | 403 | complete(&ctx->completions[0]); |
2b188cc1 JA |
404 | } |
405 | ||
406 | static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) | |
407 | { | |
408 | struct io_ring_ctx *ctx; | |
409 | ||
410 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
411 | if (!ctx) | |
412 | return NULL; | |
413 | ||
0ddf92e8 JA |
414 | ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL); |
415 | if (!ctx->fallback_req) | |
416 | goto err; | |
417 | ||
206aefde JA |
418 | ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL); |
419 | if (!ctx->completions) | |
420 | goto err; | |
421 | ||
21482896 | 422 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
206aefde JA |
423 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) |
424 | goto err; | |
2b188cc1 JA |
425 | |
426 | ctx->flags = p->flags; | |
427 | init_waitqueue_head(&ctx->cq_wait); | |
1d7bb1d5 | 428 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
206aefde JA |
429 | init_completion(&ctx->completions[0]); |
430 | init_completion(&ctx->completions[1]); | |
2b188cc1 JA |
431 | mutex_init(&ctx->uring_lock); |
432 | init_waitqueue_head(&ctx->wait); | |
433 | spin_lock_init(&ctx->completion_lock); | |
def596e9 | 434 | INIT_LIST_HEAD(&ctx->poll_list); |
221c5eb2 | 435 | INIT_LIST_HEAD(&ctx->cancel_list); |
de0617e4 | 436 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 437 | INIT_LIST_HEAD(&ctx->timeout_list); |
fcb323cc JA |
438 | init_waitqueue_head(&ctx->inflight_wait); |
439 | spin_lock_init(&ctx->inflight_lock); | |
440 | INIT_LIST_HEAD(&ctx->inflight_list); | |
2b188cc1 | 441 | return ctx; |
206aefde | 442 | err: |
0ddf92e8 JA |
443 | if (ctx->fallback_req) |
444 | kmem_cache_free(req_cachep, ctx->fallback_req); | |
206aefde JA |
445 | kfree(ctx->completions); |
446 | kfree(ctx); | |
447 | return NULL; | |
2b188cc1 JA |
448 | } |
449 | ||
a197f664 | 450 | static inline bool __io_sequence_defer(struct io_kiocb *req) |
7adf4eaf | 451 | { |
a197f664 JL |
452 | struct io_ring_ctx *ctx = req->ctx; |
453 | ||
498ccd9e JA |
454 | return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped |
455 | + atomic_read(&ctx->cached_cq_overflow); | |
7adf4eaf JA |
456 | } |
457 | ||
a197f664 | 458 | static inline bool io_sequence_defer(struct io_kiocb *req) |
de0617e4 | 459 | { |
7adf4eaf | 460 | if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) != REQ_F_IO_DRAIN) |
de0617e4 JA |
461 | return false; |
462 | ||
a197f664 | 463 | return __io_sequence_defer(req); |
de0617e4 JA |
464 | } |
465 | ||
7adf4eaf | 466 | static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx) |
de0617e4 JA |
467 | { |
468 | struct io_kiocb *req; | |
469 | ||
7adf4eaf | 470 | req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list); |
a197f664 | 471 | if (req && !io_sequence_defer(req)) { |
de0617e4 JA |
472 | list_del_init(&req->list); |
473 | return req; | |
474 | } | |
475 | ||
476 | return NULL; | |
477 | } | |
478 | ||
5262f567 JA |
479 | static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx) |
480 | { | |
7adf4eaf JA |
481 | struct io_kiocb *req; |
482 | ||
483 | req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list); | |
a197f664 | 484 | if (req && !__io_sequence_defer(req)) { |
7adf4eaf JA |
485 | list_del_init(&req->list); |
486 | return req; | |
487 | } | |
488 | ||
489 | return NULL; | |
5262f567 JA |
490 | } |
491 | ||
de0617e4 | 492 | static void __io_commit_cqring(struct io_ring_ctx *ctx) |
2b188cc1 | 493 | { |
75b28aff | 494 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 495 | |
75b28aff | 496 | if (ctx->cached_cq_tail != READ_ONCE(rings->cq.tail)) { |
2b188cc1 | 497 | /* order cqe stores with ring update */ |
75b28aff | 498 | smp_store_release(&rings->cq.tail, ctx->cached_cq_tail); |
2b188cc1 | 499 | |
2b188cc1 JA |
500 | if (wq_has_sleeper(&ctx->cq_wait)) { |
501 | wake_up_interruptible(&ctx->cq_wait); | |
502 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); | |
503 | } | |
504 | } | |
505 | } | |
506 | ||
561fb04a | 507 | static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe) |
18d9be1a | 508 | { |
561fb04a JA |
509 | u8 opcode = READ_ONCE(sqe->opcode); |
510 | ||
511 | return !(opcode == IORING_OP_READ_FIXED || | |
512 | opcode == IORING_OP_WRITE_FIXED); | |
513 | } | |
514 | ||
515 | static inline bool io_prep_async_work(struct io_kiocb *req) | |
516 | { | |
517 | bool do_hashed = false; | |
54a91f3b | 518 | |
6cc47d1d JA |
519 | if (req->submit.sqe) { |
520 | switch (req->submit.sqe->opcode) { | |
521 | case IORING_OP_WRITEV: | |
522 | case IORING_OP_WRITE_FIXED: | |
561fb04a | 523 | do_hashed = true; |
5f8fd2d3 JA |
524 | /* fall-through */ |
525 | case IORING_OP_READV: | |
526 | case IORING_OP_READ_FIXED: | |
527 | case IORING_OP_SENDMSG: | |
528 | case IORING_OP_RECVMSG: | |
529 | case IORING_OP_ACCEPT: | |
530 | case IORING_OP_POLL_ADD: | |
531 | /* | |
532 | * We know REQ_F_ISREG is not set on some of these | |
533 | * opcodes, but this enables us to keep the check in | |
534 | * just one place. | |
535 | */ | |
536 | if (!(req->flags & REQ_F_ISREG)) | |
537 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
6cc47d1d JA |
538 | break; |
539 | } | |
561fb04a JA |
540 | if (io_sqe_needs_user(req->submit.sqe)) |
541 | req->work.flags |= IO_WQ_WORK_NEEDS_USER; | |
54a91f3b JA |
542 | } |
543 | ||
561fb04a JA |
544 | return do_hashed; |
545 | } | |
546 | ||
a197f664 | 547 | static inline void io_queue_async_work(struct io_kiocb *req) |
561fb04a JA |
548 | { |
549 | bool do_hashed = io_prep_async_work(req); | |
a197f664 | 550 | struct io_ring_ctx *ctx = req->ctx; |
561fb04a JA |
551 | |
552 | trace_io_uring_queue_async_work(ctx, do_hashed, req, &req->work, | |
553 | req->flags); | |
554 | if (!do_hashed) { | |
555 | io_wq_enqueue(ctx->io_wq, &req->work); | |
556 | } else { | |
557 | io_wq_enqueue_hashed(ctx->io_wq, &req->work, | |
558 | file_inode(req->file)); | |
559 | } | |
18d9be1a JA |
560 | } |
561 | ||
5262f567 JA |
562 | static void io_kill_timeout(struct io_kiocb *req) |
563 | { | |
564 | int ret; | |
565 | ||
566 | ret = hrtimer_try_to_cancel(&req->timeout.timer); | |
567 | if (ret != -1) { | |
568 | atomic_inc(&req->ctx->cq_timeouts); | |
842f9612 | 569 | list_del_init(&req->list); |
78e19bbe | 570 | io_cqring_fill_event(req, 0); |
ec9c02ad | 571 | io_put_req(req); |
5262f567 JA |
572 | } |
573 | } | |
574 | ||
575 | static void io_kill_timeouts(struct io_ring_ctx *ctx) | |
576 | { | |
577 | struct io_kiocb *req, *tmp; | |
578 | ||
579 | spin_lock_irq(&ctx->completion_lock); | |
580 | list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list) | |
581 | io_kill_timeout(req); | |
582 | spin_unlock_irq(&ctx->completion_lock); | |
583 | } | |
584 | ||
de0617e4 JA |
585 | static void io_commit_cqring(struct io_ring_ctx *ctx) |
586 | { | |
587 | struct io_kiocb *req; | |
588 | ||
5262f567 JA |
589 | while ((req = io_get_timeout_req(ctx)) != NULL) |
590 | io_kill_timeout(req); | |
591 | ||
de0617e4 JA |
592 | __io_commit_cqring(ctx); |
593 | ||
594 | while ((req = io_get_deferred_req(ctx)) != NULL) { | |
4fe2c963 JL |
595 | if (req->flags & REQ_F_SHADOW_DRAIN) { |
596 | /* Just for drain, free it. */ | |
597 | __io_free_req(req); | |
598 | continue; | |
599 | } | |
de0617e4 | 600 | req->flags |= REQ_F_IO_DRAINED; |
a197f664 | 601 | io_queue_async_work(req); |
de0617e4 JA |
602 | } |
603 | } | |
604 | ||
2b188cc1 JA |
605 | static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx) |
606 | { | |
75b28aff | 607 | struct io_rings *rings = ctx->rings; |
2b188cc1 JA |
608 | unsigned tail; |
609 | ||
610 | tail = ctx->cached_cq_tail; | |
115e12e5 SB |
611 | /* |
612 | * writes to the cq entry need to come after reading head; the | |
613 | * control dependency is enough as we're using WRITE_ONCE to | |
614 | * fill the cq entry | |
615 | */ | |
75b28aff | 616 | if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries) |
2b188cc1 JA |
617 | return NULL; |
618 | ||
619 | ctx->cached_cq_tail++; | |
75b28aff | 620 | return &rings->cqes[tail & ctx->cq_mask]; |
2b188cc1 JA |
621 | } |
622 | ||
1d7bb1d5 JA |
623 | static void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
624 | { | |
625 | if (waitqueue_active(&ctx->wait)) | |
626 | wake_up(&ctx->wait); | |
627 | if (waitqueue_active(&ctx->sqo_wait)) | |
628 | wake_up(&ctx->sqo_wait); | |
629 | if (ctx->cq_ev_fd) | |
630 | eventfd_signal(ctx->cq_ev_fd, 1); | |
631 | } | |
632 | ||
633 | static void io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force) | |
634 | { | |
635 | struct io_rings *rings = ctx->rings; | |
636 | struct io_uring_cqe *cqe; | |
637 | struct io_kiocb *req; | |
638 | unsigned long flags; | |
639 | LIST_HEAD(list); | |
640 | ||
641 | if (!force) { | |
642 | if (list_empty_careful(&ctx->cq_overflow_list)) | |
643 | return; | |
644 | if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) == | |
645 | rings->cq_ring_entries)) | |
646 | return; | |
647 | } | |
648 | ||
649 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
650 | ||
651 | /* if force is set, the ring is going away. always drop after that */ | |
652 | if (force) | |
653 | ctx->cq_overflow_flushed = true; | |
654 | ||
655 | while (!list_empty(&ctx->cq_overflow_list)) { | |
656 | cqe = io_get_cqring(ctx); | |
657 | if (!cqe && !force) | |
658 | break; | |
659 | ||
660 | req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb, | |
661 | list); | |
662 | list_move(&req->list, &list); | |
663 | if (cqe) { | |
664 | WRITE_ONCE(cqe->user_data, req->user_data); | |
665 | WRITE_ONCE(cqe->res, req->result); | |
666 | WRITE_ONCE(cqe->flags, 0); | |
667 | } else { | |
668 | WRITE_ONCE(ctx->rings->cq_overflow, | |
669 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
670 | } | |
671 | } | |
672 | ||
673 | io_commit_cqring(ctx); | |
674 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
675 | io_cqring_ev_posted(ctx); | |
676 | ||
677 | while (!list_empty(&list)) { | |
678 | req = list_first_entry(&list, struct io_kiocb, list); | |
679 | list_del(&req->list); | |
ec9c02ad | 680 | io_put_req(req); |
1d7bb1d5 JA |
681 | } |
682 | } | |
683 | ||
78e19bbe | 684 | static void io_cqring_fill_event(struct io_kiocb *req, long res) |
2b188cc1 | 685 | { |
78e19bbe | 686 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
687 | struct io_uring_cqe *cqe; |
688 | ||
78e19bbe | 689 | trace_io_uring_complete(ctx, req->user_data, res); |
51c3ff62 | 690 | |
2b188cc1 JA |
691 | /* |
692 | * If we can't get a cq entry, userspace overflowed the | |
693 | * submission (by quite a lot). Increment the overflow count in | |
694 | * the ring. | |
695 | */ | |
696 | cqe = io_get_cqring(ctx); | |
1d7bb1d5 | 697 | if (likely(cqe)) { |
78e19bbe | 698 | WRITE_ONCE(cqe->user_data, req->user_data); |
2b188cc1 | 699 | WRITE_ONCE(cqe->res, res); |
c71ffb67 | 700 | WRITE_ONCE(cqe->flags, 0); |
1d7bb1d5 | 701 | } else if (ctx->cq_overflow_flushed) { |
498ccd9e JA |
702 | WRITE_ONCE(ctx->rings->cq_overflow, |
703 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
1d7bb1d5 JA |
704 | } else { |
705 | refcount_inc(&req->refs); | |
706 | req->result = res; | |
707 | list_add_tail(&req->list, &ctx->cq_overflow_list); | |
2b188cc1 JA |
708 | } |
709 | } | |
710 | ||
78e19bbe | 711 | static void io_cqring_add_event(struct io_kiocb *req, long res) |
2b188cc1 | 712 | { |
78e19bbe | 713 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
714 | unsigned long flags; |
715 | ||
716 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
78e19bbe | 717 | io_cqring_fill_event(req, res); |
2b188cc1 JA |
718 | io_commit_cqring(ctx); |
719 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
720 | ||
8c838788 | 721 | io_cqring_ev_posted(ctx); |
2b188cc1 JA |
722 | } |
723 | ||
0ddf92e8 JA |
724 | static inline bool io_is_fallback_req(struct io_kiocb *req) |
725 | { | |
726 | return req == (struct io_kiocb *) | |
727 | ((unsigned long) req->ctx->fallback_req & ~1UL); | |
728 | } | |
729 | ||
730 | static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx) | |
731 | { | |
732 | struct io_kiocb *req; | |
733 | ||
734 | req = ctx->fallback_req; | |
735 | if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req)) | |
736 | return req; | |
737 | ||
738 | return NULL; | |
739 | } | |
740 | ||
2579f913 JA |
741 | static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx, |
742 | struct io_submit_state *state) | |
2b188cc1 | 743 | { |
fd6fab2c | 744 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
2b188cc1 JA |
745 | struct io_kiocb *req; |
746 | ||
747 | if (!percpu_ref_tryget(&ctx->refs)) | |
748 | return NULL; | |
749 | ||
2579f913 | 750 | if (!state) { |
fd6fab2c | 751 | req = kmem_cache_alloc(req_cachep, gfp); |
2579f913 | 752 | if (unlikely(!req)) |
0ddf92e8 | 753 | goto fallback; |
2579f913 JA |
754 | } else if (!state->free_reqs) { |
755 | size_t sz; | |
756 | int ret; | |
757 | ||
758 | sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs)); | |
fd6fab2c JA |
759 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs); |
760 | ||
761 | /* | |
762 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
763 | * retry single alloc to be on the safe side. | |
764 | */ | |
765 | if (unlikely(ret <= 0)) { | |
766 | state->reqs[0] = kmem_cache_alloc(req_cachep, gfp); | |
767 | if (!state->reqs[0]) | |
0ddf92e8 | 768 | goto fallback; |
fd6fab2c JA |
769 | ret = 1; |
770 | } | |
2579f913 JA |
771 | state->free_reqs = ret - 1; |
772 | state->cur_req = 1; | |
773 | req = state->reqs[0]; | |
774 | } else { | |
775 | req = state->reqs[state->cur_req]; | |
776 | state->free_reqs--; | |
777 | state->cur_req++; | |
2b188cc1 JA |
778 | } |
779 | ||
0ddf92e8 | 780 | got_it: |
60c112b0 | 781 | req->file = NULL; |
2579f913 JA |
782 | req->ctx = ctx; |
783 | req->flags = 0; | |
e65ef56d JA |
784 | /* one is dropped after submission, the other at completion */ |
785 | refcount_set(&req->refs, 2); | |
9e645e11 | 786 | req->result = 0; |
561fb04a | 787 | INIT_IO_WORK(&req->work, io_wq_submit_work); |
2579f913 | 788 | return req; |
0ddf92e8 JA |
789 | fallback: |
790 | req = io_get_fallback_req(ctx); | |
791 | if (req) | |
792 | goto got_it; | |
6805b32e | 793 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
794 | return NULL; |
795 | } | |
796 | ||
def596e9 JA |
797 | static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr) |
798 | { | |
799 | if (*nr) { | |
800 | kmem_cache_free_bulk(req_cachep, *nr, reqs); | |
6805b32e | 801 | percpu_ref_put_many(&ctx->refs, *nr); |
def596e9 JA |
802 | *nr = 0; |
803 | } | |
804 | } | |
805 | ||
9e645e11 | 806 | static void __io_free_req(struct io_kiocb *req) |
2b188cc1 | 807 | { |
fcb323cc JA |
808 | struct io_ring_ctx *ctx = req->ctx; |
809 | ||
09bb8394 JA |
810 | if (req->file && !(req->flags & REQ_F_FIXED_FILE)) |
811 | fput(req->file); | |
fcb323cc JA |
812 | if (req->flags & REQ_F_INFLIGHT) { |
813 | unsigned long flags; | |
814 | ||
815 | spin_lock_irqsave(&ctx->inflight_lock, flags); | |
816 | list_del(&req->inflight_entry); | |
817 | if (waitqueue_active(&ctx->inflight_wait)) | |
818 | wake_up(&ctx->inflight_wait); | |
819 | spin_unlock_irqrestore(&ctx->inflight_lock, flags); | |
820 | } | |
821 | percpu_ref_put(&ctx->refs); | |
0ddf92e8 JA |
822 | if (likely(!io_is_fallback_req(req))) |
823 | kmem_cache_free(req_cachep, req); | |
824 | else | |
825 | clear_bit_unlock(0, (unsigned long *) ctx->fallback_req); | |
e65ef56d JA |
826 | } |
827 | ||
a197f664 | 828 | static bool io_link_cancel_timeout(struct io_kiocb *req) |
2665abfd | 829 | { |
a197f664 | 830 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd JA |
831 | int ret; |
832 | ||
833 | ret = hrtimer_try_to_cancel(&req->timeout.timer); | |
834 | if (ret != -1) { | |
78e19bbe | 835 | io_cqring_fill_event(req, -ECANCELED); |
2665abfd JA |
836 | io_commit_cqring(ctx); |
837 | req->flags &= ~REQ_F_LINK; | |
ec9c02ad | 838 | io_put_req(req); |
2665abfd JA |
839 | return true; |
840 | } | |
841 | ||
842 | return false; | |
843 | } | |
844 | ||
ba816ad6 | 845 | static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr) |
9e645e11 | 846 | { |
2665abfd | 847 | struct io_ring_ctx *ctx = req->ctx; |
9e645e11 | 848 | struct io_kiocb *nxt; |
2665abfd | 849 | bool wake_ev = false; |
9e645e11 JA |
850 | |
851 | /* | |
852 | * The list should never be empty when we are called here. But could | |
853 | * potentially happen if the chain is messed up, check to be on the | |
854 | * safe side. | |
855 | */ | |
856 | nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list); | |
2665abfd | 857 | while (nxt) { |
9e645e11 JA |
858 | list_del(&nxt->list); |
859 | if (!list_empty(&req->link_list)) { | |
860 | INIT_LIST_HEAD(&nxt->link_list); | |
861 | list_splice(&req->link_list, &nxt->link_list); | |
862 | nxt->flags |= REQ_F_LINK; | |
863 | } | |
864 | ||
ba816ad6 JA |
865 | /* |
866 | * If we're in async work, we can continue processing the chain | |
867 | * in this context instead of having to queue up new async work. | |
868 | */ | |
2665abfd | 869 | if (req->flags & REQ_F_LINK_TIMEOUT) { |
a197f664 | 870 | wake_ev = io_link_cancel_timeout(nxt); |
2665abfd JA |
871 | |
872 | /* we dropped this link, get next */ | |
873 | nxt = list_first_entry_or_null(&req->link_list, | |
874 | struct io_kiocb, list); | |
875 | } else if (nxtptr && current_work()) { | |
ba816ad6 | 876 | *nxtptr = nxt; |
2665abfd JA |
877 | break; |
878 | } else { | |
a197f664 | 879 | io_queue_async_work(nxt); |
2665abfd JA |
880 | break; |
881 | } | |
9e645e11 | 882 | } |
2665abfd JA |
883 | |
884 | if (wake_ev) | |
885 | io_cqring_ev_posted(ctx); | |
9e645e11 JA |
886 | } |
887 | ||
888 | /* | |
889 | * Called if REQ_F_LINK is set, and we fail the head request | |
890 | */ | |
891 | static void io_fail_links(struct io_kiocb *req) | |
892 | { | |
2665abfd | 893 | struct io_ring_ctx *ctx = req->ctx; |
9e645e11 | 894 | struct io_kiocb *link; |
2665abfd JA |
895 | unsigned long flags; |
896 | ||
897 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
9e645e11 JA |
898 | |
899 | while (!list_empty(&req->link_list)) { | |
900 | link = list_first_entry(&req->link_list, struct io_kiocb, list); | |
2665abfd | 901 | list_del_init(&link->list); |
9e645e11 | 902 | |
c826bd7a | 903 | trace_io_uring_fail_link(req, link); |
2665abfd JA |
904 | |
905 | if ((req->flags & REQ_F_LINK_TIMEOUT) && | |
906 | link->submit.sqe->opcode == IORING_OP_LINK_TIMEOUT) { | |
a197f664 | 907 | io_link_cancel_timeout(link); |
2665abfd | 908 | } else { |
78e19bbe JA |
909 | io_cqring_fill_event(link, -ECANCELED); |
910 | io_double_put_req(link); | |
2665abfd | 911 | } |
9e645e11 | 912 | } |
2665abfd JA |
913 | |
914 | io_commit_cqring(ctx); | |
915 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
916 | io_cqring_ev_posted(ctx); | |
9e645e11 JA |
917 | } |
918 | ||
c69f8dbe | 919 | static void io_free_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt) |
9e645e11 | 920 | { |
2665abfd JA |
921 | if (likely(!(req->flags & REQ_F_LINK))) { |
922 | __io_free_req(req); | |
923 | return; | |
924 | } | |
925 | ||
9e645e11 JA |
926 | /* |
927 | * If LINK is set, we have dependent requests in this chain. If we | |
928 | * didn't fail this request, queue the first one up, moving any other | |
929 | * dependencies to the next request. In case of failure, fail the rest | |
930 | * of the chain. | |
931 | */ | |
2665abfd JA |
932 | if (req->flags & REQ_F_FAIL_LINK) { |
933 | io_fail_links(req); | |
934 | } else if (req->flags & REQ_F_LINK_TIMEOUT) { | |
935 | struct io_ring_ctx *ctx = req->ctx; | |
936 | unsigned long flags; | |
937 | ||
938 | /* | |
939 | * If this is a timeout link, we could be racing with the | |
940 | * timeout timer. Grab the completion lock for this case to | |
941 | * protection against that. | |
942 | */ | |
943 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
944 | io_req_link_next(req, nxt); | |
945 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
946 | } else { | |
947 | io_req_link_next(req, nxt); | |
9e645e11 JA |
948 | } |
949 | ||
950 | __io_free_req(req); | |
951 | } | |
952 | ||
c69f8dbe JL |
953 | static void io_free_req(struct io_kiocb *req) |
954 | { | |
955 | io_free_req_find_next(req, NULL); | |
956 | } | |
957 | ||
ba816ad6 JA |
958 | /* |
959 | * Drop reference to request, return next in chain (if there is one) if this | |
960 | * was the last reference to this request. | |
961 | */ | |
ec9c02ad | 962 | static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr) |
e65ef56d | 963 | { |
ba816ad6 JA |
964 | struct io_kiocb *nxt = NULL; |
965 | ||
e65ef56d | 966 | if (refcount_dec_and_test(&req->refs)) |
c69f8dbe | 967 | io_free_req_find_next(req, &nxt); |
ba816ad6 | 968 | |
ba816ad6 | 969 | if (nxt) { |
561fb04a | 970 | if (nxtptr) |
ba816ad6 | 971 | *nxtptr = nxt; |
561fb04a | 972 | else |
a197f664 | 973 | io_queue_async_work(nxt); |
ba816ad6 | 974 | } |
2b188cc1 JA |
975 | } |
976 | ||
ec9c02ad JL |
977 | static void io_put_req(struct io_kiocb *req) |
978 | { | |
979 | if (refcount_dec_and_test(&req->refs)) | |
c69f8dbe | 980 | io_free_req(req); |
ec9c02ad JL |
981 | } |
982 | ||
78e19bbe JA |
983 | static void io_double_put_req(struct io_kiocb *req) |
984 | { | |
985 | /* drop both submit and complete references */ | |
986 | if (refcount_sub_and_test(2, &req->refs)) | |
987 | __io_free_req(req); | |
988 | } | |
989 | ||
1d7bb1d5 | 990 | static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush) |
a3a0e43f | 991 | { |
84f97dc2 JA |
992 | struct io_rings *rings = ctx->rings; |
993 | ||
1d7bb1d5 JA |
994 | /* |
995 | * noflush == true is from the waitqueue handler, just ensure we wake | |
996 | * up the task, and the next invocation will flush the entries. We | |
997 | * cannot safely to it from here. | |
998 | */ | |
999 | if (noflush && !list_empty(&ctx->cq_overflow_list)) | |
1000 | return -1U; | |
1001 | ||
1002 | io_cqring_overflow_flush(ctx, false); | |
1003 | ||
a3a0e43f JA |
1004 | /* See comment at the top of this file */ |
1005 | smp_rmb(); | |
75b28aff | 1006 | return READ_ONCE(rings->cq.tail) - READ_ONCE(rings->cq.head); |
a3a0e43f JA |
1007 | } |
1008 | ||
fb5ccc98 PB |
1009 | static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx) |
1010 | { | |
1011 | struct io_rings *rings = ctx->rings; | |
1012 | ||
1013 | /* make sure SQ entry isn't read before tail */ | |
1014 | return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head; | |
1015 | } | |
1016 | ||
def596e9 JA |
1017 | /* |
1018 | * Find and free completed poll iocbs | |
1019 | */ | |
1020 | static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1021 | struct list_head *done) | |
1022 | { | |
1023 | void *reqs[IO_IOPOLL_BATCH]; | |
1024 | struct io_kiocb *req; | |
09bb8394 | 1025 | int to_free; |
def596e9 | 1026 | |
09bb8394 | 1027 | to_free = 0; |
def596e9 JA |
1028 | while (!list_empty(done)) { |
1029 | req = list_first_entry(done, struct io_kiocb, list); | |
1030 | list_del(&req->list); | |
1031 | ||
78e19bbe | 1032 | io_cqring_fill_event(req, req->result); |
def596e9 JA |
1033 | (*nr_events)++; |
1034 | ||
09bb8394 JA |
1035 | if (refcount_dec_and_test(&req->refs)) { |
1036 | /* If we're not using fixed files, we have to pair the | |
1037 | * completion part with the file put. Use regular | |
1038 | * completions for those, only batch free for fixed | |
9e645e11 | 1039 | * file and non-linked commands. |
09bb8394 | 1040 | */ |
0ddf92e8 JA |
1041 | if (((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) == |
1042 | REQ_F_FIXED_FILE) && !io_is_fallback_req(req)) { | |
09bb8394 JA |
1043 | reqs[to_free++] = req; |
1044 | if (to_free == ARRAY_SIZE(reqs)) | |
1045 | io_free_req_many(ctx, reqs, &to_free); | |
6b06314c | 1046 | } else { |
c69f8dbe | 1047 | io_free_req(req); |
6b06314c | 1048 | } |
9a56a232 | 1049 | } |
def596e9 | 1050 | } |
def596e9 | 1051 | |
09bb8394 | 1052 | io_commit_cqring(ctx); |
def596e9 JA |
1053 | io_free_req_many(ctx, reqs, &to_free); |
1054 | } | |
1055 | ||
1056 | static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1057 | long min) | |
1058 | { | |
1059 | struct io_kiocb *req, *tmp; | |
1060 | LIST_HEAD(done); | |
1061 | bool spin; | |
1062 | int ret; | |
1063 | ||
1064 | /* | |
1065 | * Only spin for completions if we don't have multiple devices hanging | |
1066 | * off our complete list, and we're under the requested amount. | |
1067 | */ | |
1068 | spin = !ctx->poll_multi_file && *nr_events < min; | |
1069 | ||
1070 | ret = 0; | |
1071 | list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) { | |
1072 | struct kiocb *kiocb = &req->rw; | |
1073 | ||
1074 | /* | |
1075 | * Move completed entries to our local list. If we find a | |
1076 | * request that requires polling, break out and complete | |
1077 | * the done list first, if we have entries there. | |
1078 | */ | |
1079 | if (req->flags & REQ_F_IOPOLL_COMPLETED) { | |
1080 | list_move_tail(&req->list, &done); | |
1081 | continue; | |
1082 | } | |
1083 | if (!list_empty(&done)) | |
1084 | break; | |
1085 | ||
1086 | ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin); | |
1087 | if (ret < 0) | |
1088 | break; | |
1089 | ||
1090 | if (ret && spin) | |
1091 | spin = false; | |
1092 | ret = 0; | |
1093 | } | |
1094 | ||
1095 | if (!list_empty(&done)) | |
1096 | io_iopoll_complete(ctx, nr_events, &done); | |
1097 | ||
1098 | return ret; | |
1099 | } | |
1100 | ||
1101 | /* | |
1102 | * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a | |
1103 | * non-spinning poll check - we'll still enter the driver poll loop, but only | |
1104 | * as a non-spinning completion check. | |
1105 | */ | |
1106 | static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1107 | long min) | |
1108 | { | |
08f5439f | 1109 | while (!list_empty(&ctx->poll_list) && !need_resched()) { |
def596e9 JA |
1110 | int ret; |
1111 | ||
1112 | ret = io_do_iopoll(ctx, nr_events, min); | |
1113 | if (ret < 0) | |
1114 | return ret; | |
1115 | if (!min || *nr_events >= min) | |
1116 | return 0; | |
1117 | } | |
1118 | ||
1119 | return 1; | |
1120 | } | |
1121 | ||
1122 | /* | |
1123 | * We can't just wait for polled events to come to us, we have to actively | |
1124 | * find and complete them. | |
1125 | */ | |
1126 | static void io_iopoll_reap_events(struct io_ring_ctx *ctx) | |
1127 | { | |
1128 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
1129 | return; | |
1130 | ||
1131 | mutex_lock(&ctx->uring_lock); | |
1132 | while (!list_empty(&ctx->poll_list)) { | |
1133 | unsigned int nr_events = 0; | |
1134 | ||
1135 | io_iopoll_getevents(ctx, &nr_events, 1); | |
08f5439f JA |
1136 | |
1137 | /* | |
1138 | * Ensure we allow local-to-the-cpu processing to take place, | |
1139 | * in this case we need to ensure that we reap all events. | |
1140 | */ | |
1141 | cond_resched(); | |
def596e9 JA |
1142 | } |
1143 | mutex_unlock(&ctx->uring_lock); | |
1144 | } | |
1145 | ||
2b2ed975 JA |
1146 | static int __io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, |
1147 | long min) | |
def596e9 | 1148 | { |
2b2ed975 | 1149 | int iters = 0, ret = 0; |
500f9fba | 1150 | |
def596e9 JA |
1151 | do { |
1152 | int tmin = 0; | |
1153 | ||
a3a0e43f JA |
1154 | /* |
1155 | * Don't enter poll loop if we already have events pending. | |
1156 | * If we do, we can potentially be spinning for commands that | |
1157 | * already triggered a CQE (eg in error). | |
1158 | */ | |
1d7bb1d5 | 1159 | if (io_cqring_events(ctx, false)) |
a3a0e43f JA |
1160 | break; |
1161 | ||
500f9fba JA |
1162 | /* |
1163 | * If a submit got punted to a workqueue, we can have the | |
1164 | * application entering polling for a command before it gets | |
1165 | * issued. That app will hold the uring_lock for the duration | |
1166 | * of the poll right here, so we need to take a breather every | |
1167 | * now and then to ensure that the issue has a chance to add | |
1168 | * the poll to the issued list. Otherwise we can spin here | |
1169 | * forever, while the workqueue is stuck trying to acquire the | |
1170 | * very same mutex. | |
1171 | */ | |
1172 | if (!(++iters & 7)) { | |
1173 | mutex_unlock(&ctx->uring_lock); | |
1174 | mutex_lock(&ctx->uring_lock); | |
1175 | } | |
1176 | ||
def596e9 JA |
1177 | if (*nr_events < min) |
1178 | tmin = min - *nr_events; | |
1179 | ||
1180 | ret = io_iopoll_getevents(ctx, nr_events, tmin); | |
1181 | if (ret <= 0) | |
1182 | break; | |
1183 | ret = 0; | |
1184 | } while (min && !*nr_events && !need_resched()); | |
1185 | ||
2b2ed975 JA |
1186 | return ret; |
1187 | } | |
1188 | ||
1189 | static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, | |
1190 | long min) | |
1191 | { | |
1192 | int ret; | |
1193 | ||
1194 | /* | |
1195 | * We disallow the app entering submit/complete with polling, but we | |
1196 | * still need to lock the ring to prevent racing with polled issue | |
1197 | * that got punted to a workqueue. | |
1198 | */ | |
1199 | mutex_lock(&ctx->uring_lock); | |
1200 | ret = __io_iopoll_check(ctx, nr_events, min); | |
500f9fba | 1201 | mutex_unlock(&ctx->uring_lock); |
def596e9 JA |
1202 | return ret; |
1203 | } | |
1204 | ||
491381ce | 1205 | static void kiocb_end_write(struct io_kiocb *req) |
2b188cc1 | 1206 | { |
491381ce JA |
1207 | /* |
1208 | * Tell lockdep we inherited freeze protection from submission | |
1209 | * thread. | |
1210 | */ | |
1211 | if (req->flags & REQ_F_ISREG) { | |
1212 | struct inode *inode = file_inode(req->file); | |
2b188cc1 | 1213 | |
491381ce | 1214 | __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE); |
2b188cc1 | 1215 | } |
491381ce | 1216 | file_end_write(req->file); |
2b188cc1 JA |
1217 | } |
1218 | ||
ba816ad6 | 1219 | static void io_complete_rw_common(struct kiocb *kiocb, long res) |
2b188cc1 JA |
1220 | { |
1221 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
1222 | ||
491381ce JA |
1223 | if (kiocb->ki_flags & IOCB_WRITE) |
1224 | kiocb_end_write(req); | |
2b188cc1 | 1225 | |
9e645e11 JA |
1226 | if ((req->flags & REQ_F_LINK) && res != req->result) |
1227 | req->flags |= REQ_F_FAIL_LINK; | |
78e19bbe | 1228 | io_cqring_add_event(req, res); |
ba816ad6 JA |
1229 | } |
1230 | ||
1231 | static void io_complete_rw(struct kiocb *kiocb, long res, long res2) | |
1232 | { | |
1233 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
1234 | ||
1235 | io_complete_rw_common(kiocb, res); | |
ec9c02ad | 1236 | io_put_req(req); |
ba816ad6 JA |
1237 | } |
1238 | ||
1239 | static struct io_kiocb *__io_complete_rw(struct kiocb *kiocb, long res) | |
1240 | { | |
1241 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
ec9c02ad | 1242 | struct io_kiocb *nxt = NULL; |
ba816ad6 JA |
1243 | |
1244 | io_complete_rw_common(kiocb, res); | |
ec9c02ad JL |
1245 | io_put_req_find_next(req, &nxt); |
1246 | ||
1247 | return nxt; | |
2b188cc1 JA |
1248 | } |
1249 | ||
def596e9 JA |
1250 | static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2) |
1251 | { | |
1252 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
1253 | ||
491381ce JA |
1254 | if (kiocb->ki_flags & IOCB_WRITE) |
1255 | kiocb_end_write(req); | |
def596e9 | 1256 | |
9e645e11 JA |
1257 | if ((req->flags & REQ_F_LINK) && res != req->result) |
1258 | req->flags |= REQ_F_FAIL_LINK; | |
1259 | req->result = res; | |
def596e9 JA |
1260 | if (res != -EAGAIN) |
1261 | req->flags |= REQ_F_IOPOLL_COMPLETED; | |
1262 | } | |
1263 | ||
1264 | /* | |
1265 | * After the iocb has been issued, it's safe to be found on the poll list. | |
1266 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
1267 | * find it from a io_iopoll_getevents() thread before the issuer is done | |
1268 | * accessing the kiocb cookie. | |
1269 | */ | |
1270 | static void io_iopoll_req_issued(struct io_kiocb *req) | |
1271 | { | |
1272 | struct io_ring_ctx *ctx = req->ctx; | |
1273 | ||
1274 | /* | |
1275 | * Track whether we have multiple files in our lists. This will impact | |
1276 | * how we do polling eventually, not spinning if we're on potentially | |
1277 | * different devices. | |
1278 | */ | |
1279 | if (list_empty(&ctx->poll_list)) { | |
1280 | ctx->poll_multi_file = false; | |
1281 | } else if (!ctx->poll_multi_file) { | |
1282 | struct io_kiocb *list_req; | |
1283 | ||
1284 | list_req = list_first_entry(&ctx->poll_list, struct io_kiocb, | |
1285 | list); | |
1286 | if (list_req->rw.ki_filp != req->rw.ki_filp) | |
1287 | ctx->poll_multi_file = true; | |
1288 | } | |
1289 | ||
1290 | /* | |
1291 | * For fast devices, IO may have already completed. If it has, add | |
1292 | * it to the front so we find it first. | |
1293 | */ | |
1294 | if (req->flags & REQ_F_IOPOLL_COMPLETED) | |
1295 | list_add(&req->list, &ctx->poll_list); | |
1296 | else | |
1297 | list_add_tail(&req->list, &ctx->poll_list); | |
1298 | } | |
1299 | ||
3d6770fb | 1300 | static void io_file_put(struct io_submit_state *state) |
9a56a232 | 1301 | { |
3d6770fb | 1302 | if (state->file) { |
9a56a232 JA |
1303 | int diff = state->has_refs - state->used_refs; |
1304 | ||
1305 | if (diff) | |
1306 | fput_many(state->file, diff); | |
1307 | state->file = NULL; | |
1308 | } | |
1309 | } | |
1310 | ||
1311 | /* | |
1312 | * Get as many references to a file as we have IOs left in this submission, | |
1313 | * assuming most submissions are for one file, or at least that each file | |
1314 | * has more than one submission. | |
1315 | */ | |
1316 | static struct file *io_file_get(struct io_submit_state *state, int fd) | |
1317 | { | |
1318 | if (!state) | |
1319 | return fget(fd); | |
1320 | ||
1321 | if (state->file) { | |
1322 | if (state->fd == fd) { | |
1323 | state->used_refs++; | |
1324 | state->ios_left--; | |
1325 | return state->file; | |
1326 | } | |
3d6770fb | 1327 | io_file_put(state); |
9a56a232 JA |
1328 | } |
1329 | state->file = fget_many(fd, state->ios_left); | |
1330 | if (!state->file) | |
1331 | return NULL; | |
1332 | ||
1333 | state->fd = fd; | |
1334 | state->has_refs = state->ios_left; | |
1335 | state->used_refs = 1; | |
1336 | state->ios_left--; | |
1337 | return state->file; | |
1338 | } | |
1339 | ||
2b188cc1 JA |
1340 | /* |
1341 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1342 | * any file. For now, just ensure that anything potentially problematic is done | |
1343 | * inline. | |
1344 | */ | |
1345 | static bool io_file_supports_async(struct file *file) | |
1346 | { | |
1347 | umode_t mode = file_inode(file)->i_mode; | |
1348 | ||
1349 | if (S_ISBLK(mode) || S_ISCHR(mode)) | |
1350 | return true; | |
1351 | if (S_ISREG(mode) && file->f_op != &io_uring_fops) | |
1352 | return true; | |
1353 | ||
1354 | return false; | |
1355 | } | |
1356 | ||
267bc904 | 1357 | static int io_prep_rw(struct io_kiocb *req, bool force_nonblock) |
2b188cc1 | 1358 | { |
267bc904 | 1359 | const struct io_uring_sqe *sqe = req->submit.sqe; |
def596e9 | 1360 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 | 1361 | struct kiocb *kiocb = &req->rw; |
09bb8394 JA |
1362 | unsigned ioprio; |
1363 | int ret; | |
2b188cc1 | 1364 | |
09bb8394 JA |
1365 | if (!req->file) |
1366 | return -EBADF; | |
2b188cc1 | 1367 | |
491381ce JA |
1368 | if (S_ISREG(file_inode(req->file)->i_mode)) |
1369 | req->flags |= REQ_F_ISREG; | |
1370 | ||
1371 | /* | |
1372 | * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so | |
1373 | * we know to async punt it even if it was opened O_NONBLOCK | |
1374 | */ | |
1375 | if (force_nonblock && !io_file_supports_async(req->file)) { | |
1376 | req->flags |= REQ_F_MUST_PUNT; | |
1377 | return -EAGAIN; | |
1378 | } | |
6b06314c | 1379 | |
2b188cc1 JA |
1380 | kiocb->ki_pos = READ_ONCE(sqe->off); |
1381 | kiocb->ki_flags = iocb_flags(kiocb->ki_filp); | |
1382 | kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp)); | |
1383 | ||
1384 | ioprio = READ_ONCE(sqe->ioprio); | |
1385 | if (ioprio) { | |
1386 | ret = ioprio_check_cap(ioprio); | |
1387 | if (ret) | |
09bb8394 | 1388 | return ret; |
2b188cc1 JA |
1389 | |
1390 | kiocb->ki_ioprio = ioprio; | |
1391 | } else | |
1392 | kiocb->ki_ioprio = get_current_ioprio(); | |
1393 | ||
1394 | ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags)); | |
1395 | if (unlikely(ret)) | |
09bb8394 | 1396 | return ret; |
8449eeda SB |
1397 | |
1398 | /* don't allow async punt if RWF_NOWAIT was requested */ | |
491381ce JA |
1399 | if ((kiocb->ki_flags & IOCB_NOWAIT) || |
1400 | (req->file->f_flags & O_NONBLOCK)) | |
8449eeda SB |
1401 | req->flags |= REQ_F_NOWAIT; |
1402 | ||
1403 | if (force_nonblock) | |
2b188cc1 | 1404 | kiocb->ki_flags |= IOCB_NOWAIT; |
8449eeda | 1405 | |
def596e9 | 1406 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 JA |
1407 | if (!(kiocb->ki_flags & IOCB_DIRECT) || |
1408 | !kiocb->ki_filp->f_op->iopoll) | |
09bb8394 | 1409 | return -EOPNOTSUPP; |
2b188cc1 | 1410 | |
def596e9 JA |
1411 | kiocb->ki_flags |= IOCB_HIPRI; |
1412 | kiocb->ki_complete = io_complete_rw_iopoll; | |
1413 | } else { | |
09bb8394 JA |
1414 | if (kiocb->ki_flags & IOCB_HIPRI) |
1415 | return -EINVAL; | |
def596e9 JA |
1416 | kiocb->ki_complete = io_complete_rw; |
1417 | } | |
2b188cc1 | 1418 | return 0; |
2b188cc1 JA |
1419 | } |
1420 | ||
1421 | static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) | |
1422 | { | |
1423 | switch (ret) { | |
1424 | case -EIOCBQUEUED: | |
1425 | break; | |
1426 | case -ERESTARTSYS: | |
1427 | case -ERESTARTNOINTR: | |
1428 | case -ERESTARTNOHAND: | |
1429 | case -ERESTART_RESTARTBLOCK: | |
1430 | /* | |
1431 | * We can't just restart the syscall, since previously | |
1432 | * submitted sqes may already be in progress. Just fail this | |
1433 | * IO with EINTR. | |
1434 | */ | |
1435 | ret = -EINTR; | |
1436 | /* fall through */ | |
1437 | default: | |
1438 | kiocb->ki_complete(kiocb, ret, 0); | |
1439 | } | |
1440 | } | |
1441 | ||
ba816ad6 JA |
1442 | static void kiocb_done(struct kiocb *kiocb, ssize_t ret, struct io_kiocb **nxt, |
1443 | bool in_async) | |
1444 | { | |
1445 | if (in_async && ret >= 0 && nxt && kiocb->ki_complete == io_complete_rw) | |
1446 | *nxt = __io_complete_rw(kiocb, ret); | |
1447 | else | |
1448 | io_rw_done(kiocb, ret); | |
1449 | } | |
1450 | ||
edafccee JA |
1451 | static int io_import_fixed(struct io_ring_ctx *ctx, int rw, |
1452 | const struct io_uring_sqe *sqe, | |
1453 | struct iov_iter *iter) | |
1454 | { | |
1455 | size_t len = READ_ONCE(sqe->len); | |
1456 | struct io_mapped_ubuf *imu; | |
1457 | unsigned index, buf_index; | |
1458 | size_t offset; | |
1459 | u64 buf_addr; | |
1460 | ||
1461 | /* attempt to use fixed buffers without having provided iovecs */ | |
1462 | if (unlikely(!ctx->user_bufs)) | |
1463 | return -EFAULT; | |
1464 | ||
1465 | buf_index = READ_ONCE(sqe->buf_index); | |
1466 | if (unlikely(buf_index >= ctx->nr_user_bufs)) | |
1467 | return -EFAULT; | |
1468 | ||
1469 | index = array_index_nospec(buf_index, ctx->nr_user_bufs); | |
1470 | imu = &ctx->user_bufs[index]; | |
1471 | buf_addr = READ_ONCE(sqe->addr); | |
1472 | ||
1473 | /* overflow */ | |
1474 | if (buf_addr + len < buf_addr) | |
1475 | return -EFAULT; | |
1476 | /* not inside the mapped region */ | |
1477 | if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len) | |
1478 | return -EFAULT; | |
1479 | ||
1480 | /* | |
1481 | * May not be a start of buffer, set size appropriately | |
1482 | * and advance us to the beginning. | |
1483 | */ | |
1484 | offset = buf_addr - imu->ubuf; | |
1485 | iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len); | |
bd11b3a3 JA |
1486 | |
1487 | if (offset) { | |
1488 | /* | |
1489 | * Don't use iov_iter_advance() here, as it's really slow for | |
1490 | * using the latter parts of a big fixed buffer - it iterates | |
1491 | * over each segment manually. We can cheat a bit here, because | |
1492 | * we know that: | |
1493 | * | |
1494 | * 1) it's a BVEC iter, we set it up | |
1495 | * 2) all bvecs are PAGE_SIZE in size, except potentially the | |
1496 | * first and last bvec | |
1497 | * | |
1498 | * So just find our index, and adjust the iterator afterwards. | |
1499 | * If the offset is within the first bvec (or the whole first | |
1500 | * bvec, just use iov_iter_advance(). This makes it easier | |
1501 | * since we can just skip the first segment, which may not | |
1502 | * be PAGE_SIZE aligned. | |
1503 | */ | |
1504 | const struct bio_vec *bvec = imu->bvec; | |
1505 | ||
1506 | if (offset <= bvec->bv_len) { | |
1507 | iov_iter_advance(iter, offset); | |
1508 | } else { | |
1509 | unsigned long seg_skip; | |
1510 | ||
1511 | /* skip first vec */ | |
1512 | offset -= bvec->bv_len; | |
1513 | seg_skip = 1 + (offset >> PAGE_SHIFT); | |
1514 | ||
1515 | iter->bvec = bvec + seg_skip; | |
1516 | iter->nr_segs -= seg_skip; | |
99c79f66 | 1517 | iter->count -= bvec->bv_len + offset; |
bd11b3a3 | 1518 | iter->iov_offset = offset & ~PAGE_MASK; |
bd11b3a3 JA |
1519 | } |
1520 | } | |
1521 | ||
edafccee JA |
1522 | return 0; |
1523 | } | |
1524 | ||
87e5e6da JA |
1525 | static ssize_t io_import_iovec(struct io_ring_ctx *ctx, int rw, |
1526 | const struct sqe_submit *s, struct iovec **iovec, | |
1527 | struct iov_iter *iter) | |
2b188cc1 JA |
1528 | { |
1529 | const struct io_uring_sqe *sqe = s->sqe; | |
1530 | void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
1531 | size_t sqe_len = READ_ONCE(sqe->len); | |
edafccee JA |
1532 | u8 opcode; |
1533 | ||
1534 | /* | |
1535 | * We're reading ->opcode for the second time, but the first read | |
1536 | * doesn't care whether it's _FIXED or not, so it doesn't matter | |
1537 | * whether ->opcode changes concurrently. The first read does care | |
1538 | * about whether it is a READ or a WRITE, so we don't trust this read | |
1539 | * for that purpose and instead let the caller pass in the read/write | |
1540 | * flag. | |
1541 | */ | |
1542 | opcode = READ_ONCE(sqe->opcode); | |
1543 | if (opcode == IORING_OP_READ_FIXED || | |
1544 | opcode == IORING_OP_WRITE_FIXED) { | |
87e5e6da | 1545 | ssize_t ret = io_import_fixed(ctx, rw, sqe, iter); |
edafccee JA |
1546 | *iovec = NULL; |
1547 | return ret; | |
1548 | } | |
2b188cc1 JA |
1549 | |
1550 | if (!s->has_user) | |
1551 | return -EFAULT; | |
1552 | ||
1553 | #ifdef CONFIG_COMPAT | |
1554 | if (ctx->compat) | |
1555 | return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV, | |
1556 | iovec, iter); | |
1557 | #endif | |
1558 | ||
1559 | return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter); | |
1560 | } | |
1561 | ||
32960613 JA |
1562 | /* |
1563 | * For files that don't have ->read_iter() and ->write_iter(), handle them | |
1564 | * by looping over ->read() or ->write() manually. | |
1565 | */ | |
1566 | static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb, | |
1567 | struct iov_iter *iter) | |
1568 | { | |
1569 | ssize_t ret = 0; | |
1570 | ||
1571 | /* | |
1572 | * Don't support polled IO through this interface, and we can't | |
1573 | * support non-blocking either. For the latter, this just causes | |
1574 | * the kiocb to be handled from an async context. | |
1575 | */ | |
1576 | if (kiocb->ki_flags & IOCB_HIPRI) | |
1577 | return -EOPNOTSUPP; | |
1578 | if (kiocb->ki_flags & IOCB_NOWAIT) | |
1579 | return -EAGAIN; | |
1580 | ||
1581 | while (iov_iter_count(iter)) { | |
1582 | struct iovec iovec = iov_iter_iovec(iter); | |
1583 | ssize_t nr; | |
1584 | ||
1585 | if (rw == READ) { | |
1586 | nr = file->f_op->read(file, iovec.iov_base, | |
1587 | iovec.iov_len, &kiocb->ki_pos); | |
1588 | } else { | |
1589 | nr = file->f_op->write(file, iovec.iov_base, | |
1590 | iovec.iov_len, &kiocb->ki_pos); | |
1591 | } | |
1592 | ||
1593 | if (nr < 0) { | |
1594 | if (!ret) | |
1595 | ret = nr; | |
1596 | break; | |
1597 | } | |
1598 | ret += nr; | |
1599 | if (nr != iovec.iov_len) | |
1600 | break; | |
1601 | iov_iter_advance(iter, nr); | |
1602 | } | |
1603 | ||
1604 | return ret; | |
1605 | } | |
1606 | ||
267bc904 PB |
1607 | static int io_read(struct io_kiocb *req, struct io_kiocb **nxt, |
1608 | bool force_nonblock) | |
2b188cc1 JA |
1609 | { |
1610 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1611 | struct kiocb *kiocb = &req->rw; | |
1612 | struct iov_iter iter; | |
1613 | struct file *file; | |
31b51510 | 1614 | size_t iov_count; |
9d93a3f5 | 1615 | ssize_t read_size, ret; |
2b188cc1 | 1616 | |
267bc904 | 1617 | ret = io_prep_rw(req, force_nonblock); |
2b188cc1 JA |
1618 | if (ret) |
1619 | return ret; | |
1620 | file = kiocb->ki_filp; | |
1621 | ||
2b188cc1 | 1622 | if (unlikely(!(file->f_mode & FMODE_READ))) |
09bb8394 | 1623 | return -EBADF; |
2b188cc1 | 1624 | |
267bc904 | 1625 | ret = io_import_iovec(req->ctx, READ, &req->submit, &iovec, &iter); |
87e5e6da | 1626 | if (ret < 0) |
09bb8394 | 1627 | return ret; |
2b188cc1 | 1628 | |
9d93a3f5 | 1629 | read_size = ret; |
9e645e11 JA |
1630 | if (req->flags & REQ_F_LINK) |
1631 | req->result = read_size; | |
1632 | ||
31b51510 JA |
1633 | iov_count = iov_iter_count(&iter); |
1634 | ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count); | |
2b188cc1 JA |
1635 | if (!ret) { |
1636 | ssize_t ret2; | |
1637 | ||
32960613 JA |
1638 | if (file->f_op->read_iter) |
1639 | ret2 = call_read_iter(file, kiocb, &iter); | |
1640 | else | |
1641 | ret2 = loop_rw_iter(READ, file, kiocb, &iter); | |
1642 | ||
9d93a3f5 JA |
1643 | /* |
1644 | * In case of a short read, punt to async. This can happen | |
1645 | * if we have data partially cached. Alternatively we can | |
1646 | * return the short read, in which case the application will | |
1647 | * need to issue another SQE and wait for it. That SQE will | |
1648 | * need async punt anyway, so it's more efficient to do it | |
1649 | * here. | |
1650 | */ | |
491381ce JA |
1651 | if (force_nonblock && !(req->flags & REQ_F_NOWAIT) && |
1652 | (req->flags & REQ_F_ISREG) && | |
1653 | ret2 > 0 && ret2 < read_size) | |
9d93a3f5 JA |
1654 | ret2 = -EAGAIN; |
1655 | /* Catch -EAGAIN return for forced non-blocking submission */ | |
561fb04a | 1656 | if (!force_nonblock || ret2 != -EAGAIN) |
267bc904 | 1657 | kiocb_done(kiocb, ret2, nxt, req->submit.in_async); |
561fb04a | 1658 | else |
2b188cc1 JA |
1659 | ret = -EAGAIN; |
1660 | } | |
1661 | kfree(iovec); | |
2b188cc1 JA |
1662 | return ret; |
1663 | } | |
1664 | ||
267bc904 PB |
1665 | static int io_write(struct io_kiocb *req, struct io_kiocb **nxt, |
1666 | bool force_nonblock) | |
2b188cc1 JA |
1667 | { |
1668 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1669 | struct kiocb *kiocb = &req->rw; | |
1670 | struct iov_iter iter; | |
1671 | struct file *file; | |
31b51510 | 1672 | size_t iov_count; |
87e5e6da | 1673 | ssize_t ret; |
2b188cc1 | 1674 | |
267bc904 | 1675 | ret = io_prep_rw(req, force_nonblock); |
2b188cc1 JA |
1676 | if (ret) |
1677 | return ret; | |
2b188cc1 | 1678 | |
2b188cc1 JA |
1679 | file = kiocb->ki_filp; |
1680 | if (unlikely(!(file->f_mode & FMODE_WRITE))) | |
09bb8394 | 1681 | return -EBADF; |
2b188cc1 | 1682 | |
267bc904 | 1683 | ret = io_import_iovec(req->ctx, WRITE, &req->submit, &iovec, &iter); |
87e5e6da | 1684 | if (ret < 0) |
09bb8394 | 1685 | return ret; |
2b188cc1 | 1686 | |
9e645e11 JA |
1687 | if (req->flags & REQ_F_LINK) |
1688 | req->result = ret; | |
1689 | ||
31b51510 JA |
1690 | iov_count = iov_iter_count(&iter); |
1691 | ||
1692 | ret = -EAGAIN; | |
561fb04a | 1693 | if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT)) |
31b51510 | 1694 | goto out_free; |
31b51510 JA |
1695 | |
1696 | ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count); | |
2b188cc1 | 1697 | if (!ret) { |
9bf7933f RP |
1698 | ssize_t ret2; |
1699 | ||
2b188cc1 JA |
1700 | /* |
1701 | * Open-code file_start_write here to grab freeze protection, | |
1702 | * which will be released by another thread in | |
1703 | * io_complete_rw(). Fool lockdep by telling it the lock got | |
1704 | * released so that it doesn't complain about the held lock when | |
1705 | * we return to userspace. | |
1706 | */ | |
491381ce | 1707 | if (req->flags & REQ_F_ISREG) { |
2b188cc1 JA |
1708 | __sb_start_write(file_inode(file)->i_sb, |
1709 | SB_FREEZE_WRITE, true); | |
1710 | __sb_writers_release(file_inode(file)->i_sb, | |
1711 | SB_FREEZE_WRITE); | |
1712 | } | |
1713 | kiocb->ki_flags |= IOCB_WRITE; | |
9bf7933f | 1714 | |
32960613 JA |
1715 | if (file->f_op->write_iter) |
1716 | ret2 = call_write_iter(file, kiocb, &iter); | |
1717 | else | |
1718 | ret2 = loop_rw_iter(WRITE, file, kiocb, &iter); | |
561fb04a | 1719 | if (!force_nonblock || ret2 != -EAGAIN) |
267bc904 | 1720 | kiocb_done(kiocb, ret2, nxt, req->submit.in_async); |
561fb04a | 1721 | else |
9bf7933f | 1722 | ret = -EAGAIN; |
2b188cc1 | 1723 | } |
31b51510 | 1724 | out_free: |
2b188cc1 | 1725 | kfree(iovec); |
2b188cc1 JA |
1726 | return ret; |
1727 | } | |
1728 | ||
1729 | /* | |
1730 | * IORING_OP_NOP just posts a completion event, nothing else. | |
1731 | */ | |
78e19bbe | 1732 | static int io_nop(struct io_kiocb *req) |
2b188cc1 JA |
1733 | { |
1734 | struct io_ring_ctx *ctx = req->ctx; | |
2b188cc1 | 1735 | |
def596e9 JA |
1736 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
1737 | return -EINVAL; | |
1738 | ||
78e19bbe | 1739 | io_cqring_add_event(req, 0); |
ec9c02ad | 1740 | io_put_req(req); |
2b188cc1 JA |
1741 | return 0; |
1742 | } | |
1743 | ||
c992fe29 CH |
1744 | static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
1745 | { | |
6b06314c | 1746 | struct io_ring_ctx *ctx = req->ctx; |
c992fe29 | 1747 | |
09bb8394 JA |
1748 | if (!req->file) |
1749 | return -EBADF; | |
c992fe29 | 1750 | |
6b06314c | 1751 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
def596e9 | 1752 | return -EINVAL; |
edafccee | 1753 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) |
c992fe29 CH |
1754 | return -EINVAL; |
1755 | ||
c992fe29 CH |
1756 | return 0; |
1757 | } | |
1758 | ||
1759 | static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
ba816ad6 | 1760 | struct io_kiocb **nxt, bool force_nonblock) |
c992fe29 CH |
1761 | { |
1762 | loff_t sqe_off = READ_ONCE(sqe->off); | |
1763 | loff_t sqe_len = READ_ONCE(sqe->len); | |
1764 | loff_t end = sqe_off + sqe_len; | |
1765 | unsigned fsync_flags; | |
1766 | int ret; | |
1767 | ||
1768 | fsync_flags = READ_ONCE(sqe->fsync_flags); | |
1769 | if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC)) | |
1770 | return -EINVAL; | |
1771 | ||
1772 | ret = io_prep_fsync(req, sqe); | |
1773 | if (ret) | |
1774 | return ret; | |
1775 | ||
1776 | /* fsync always requires a blocking context */ | |
1777 | if (force_nonblock) | |
1778 | return -EAGAIN; | |
1779 | ||
1780 | ret = vfs_fsync_range(req->rw.ki_filp, sqe_off, | |
1781 | end > 0 ? end : LLONG_MAX, | |
1782 | fsync_flags & IORING_FSYNC_DATASYNC); | |
1783 | ||
9e645e11 JA |
1784 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1785 | req->flags |= REQ_F_FAIL_LINK; | |
78e19bbe | 1786 | io_cqring_add_event(req, ret); |
ec9c02ad | 1787 | io_put_req_find_next(req, nxt); |
c992fe29 CH |
1788 | return 0; |
1789 | } | |
1790 | ||
5d17b4a4 JA |
1791 | static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
1792 | { | |
1793 | struct io_ring_ctx *ctx = req->ctx; | |
1794 | int ret = 0; | |
1795 | ||
1796 | if (!req->file) | |
1797 | return -EBADF; | |
5d17b4a4 JA |
1798 | |
1799 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
1800 | return -EINVAL; | |
1801 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) | |
1802 | return -EINVAL; | |
1803 | ||
5d17b4a4 JA |
1804 | return ret; |
1805 | } | |
1806 | ||
1807 | static int io_sync_file_range(struct io_kiocb *req, | |
1808 | const struct io_uring_sqe *sqe, | |
ba816ad6 | 1809 | struct io_kiocb **nxt, |
5d17b4a4 JA |
1810 | bool force_nonblock) |
1811 | { | |
1812 | loff_t sqe_off; | |
1813 | loff_t sqe_len; | |
1814 | unsigned flags; | |
1815 | int ret; | |
1816 | ||
1817 | ret = io_prep_sfr(req, sqe); | |
1818 | if (ret) | |
1819 | return ret; | |
1820 | ||
1821 | /* sync_file_range always requires a blocking context */ | |
1822 | if (force_nonblock) | |
1823 | return -EAGAIN; | |
1824 | ||
1825 | sqe_off = READ_ONCE(sqe->off); | |
1826 | sqe_len = READ_ONCE(sqe->len); | |
1827 | flags = READ_ONCE(sqe->sync_range_flags); | |
1828 | ||
1829 | ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags); | |
1830 | ||
9e645e11 JA |
1831 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1832 | req->flags |= REQ_F_FAIL_LINK; | |
78e19bbe | 1833 | io_cqring_add_event(req, ret); |
ec9c02ad | 1834 | io_put_req_find_next(req, nxt); |
5d17b4a4 JA |
1835 | return 0; |
1836 | } | |
1837 | ||
0fa03c62 | 1838 | #if defined(CONFIG_NET) |
aa1fa28f | 1839 | static int io_send_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
ba816ad6 | 1840 | struct io_kiocb **nxt, bool force_nonblock, |
aa1fa28f JA |
1841 | long (*fn)(struct socket *, struct user_msghdr __user *, |
1842 | unsigned int)) | |
1843 | { | |
0fa03c62 JA |
1844 | struct socket *sock; |
1845 | int ret; | |
1846 | ||
1847 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1848 | return -EINVAL; | |
1849 | ||
1850 | sock = sock_from_file(req->file, &ret); | |
1851 | if (sock) { | |
1852 | struct user_msghdr __user *msg; | |
1853 | unsigned flags; | |
1854 | ||
1855 | flags = READ_ONCE(sqe->msg_flags); | |
1856 | if (flags & MSG_DONTWAIT) | |
1857 | req->flags |= REQ_F_NOWAIT; | |
1858 | else if (force_nonblock) | |
1859 | flags |= MSG_DONTWAIT; | |
1860 | ||
1861 | msg = (struct user_msghdr __user *) (unsigned long) | |
1862 | READ_ONCE(sqe->addr); | |
1863 | ||
aa1fa28f | 1864 | ret = fn(sock, msg, flags); |
0fa03c62 JA |
1865 | if (force_nonblock && ret == -EAGAIN) |
1866 | return ret; | |
1867 | } | |
1868 | ||
78e19bbe | 1869 | io_cqring_add_event(req, ret); |
f1f40853 JA |
1870 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1871 | req->flags |= REQ_F_FAIL_LINK; | |
ec9c02ad | 1872 | io_put_req_find_next(req, nxt); |
5d17b4a4 JA |
1873 | return 0; |
1874 | } | |
aa1fa28f JA |
1875 | #endif |
1876 | ||
1877 | static int io_sendmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
ba816ad6 | 1878 | struct io_kiocb **nxt, bool force_nonblock) |
aa1fa28f JA |
1879 | { |
1880 | #if defined(CONFIG_NET) | |
ba816ad6 JA |
1881 | return io_send_recvmsg(req, sqe, nxt, force_nonblock, |
1882 | __sys_sendmsg_sock); | |
aa1fa28f JA |
1883 | #else |
1884 | return -EOPNOTSUPP; | |
1885 | #endif | |
1886 | } | |
1887 | ||
1888 | static int io_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
ba816ad6 | 1889 | struct io_kiocb **nxt, bool force_nonblock) |
aa1fa28f JA |
1890 | { |
1891 | #if defined(CONFIG_NET) | |
ba816ad6 JA |
1892 | return io_send_recvmsg(req, sqe, nxt, force_nonblock, |
1893 | __sys_recvmsg_sock); | |
0fa03c62 JA |
1894 | #else |
1895 | return -EOPNOTSUPP; | |
1896 | #endif | |
1897 | } | |
5d17b4a4 | 1898 | |
17f2fe35 JA |
1899 | static int io_accept(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
1900 | struct io_kiocb **nxt, bool force_nonblock) | |
1901 | { | |
1902 | #if defined(CONFIG_NET) | |
1903 | struct sockaddr __user *addr; | |
1904 | int __user *addr_len; | |
1905 | unsigned file_flags; | |
1906 | int flags, ret; | |
1907 | ||
1908 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL))) | |
1909 | return -EINVAL; | |
1910 | if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index) | |
1911 | return -EINVAL; | |
1912 | ||
1913 | addr = (struct sockaddr __user *) (unsigned long) READ_ONCE(sqe->addr); | |
1914 | addr_len = (int __user *) (unsigned long) READ_ONCE(sqe->addr2); | |
1915 | flags = READ_ONCE(sqe->accept_flags); | |
1916 | file_flags = force_nonblock ? O_NONBLOCK : 0; | |
1917 | ||
1918 | ret = __sys_accept4_file(req->file, file_flags, addr, addr_len, flags); | |
1919 | if (ret == -EAGAIN && force_nonblock) { | |
1920 | req->work.flags |= IO_WQ_WORK_NEEDS_FILES; | |
1921 | return -EAGAIN; | |
1922 | } | |
8e3cca12 JA |
1923 | if (ret == -ERESTARTSYS) |
1924 | ret = -EINTR; | |
17f2fe35 JA |
1925 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1926 | req->flags |= REQ_F_FAIL_LINK; | |
78e19bbe | 1927 | io_cqring_add_event(req, ret); |
ec9c02ad | 1928 | io_put_req_find_next(req, nxt); |
17f2fe35 JA |
1929 | return 0; |
1930 | #else | |
1931 | return -EOPNOTSUPP; | |
1932 | #endif | |
1933 | } | |
1934 | ||
221c5eb2 JA |
1935 | static void io_poll_remove_one(struct io_kiocb *req) |
1936 | { | |
1937 | struct io_poll_iocb *poll = &req->poll; | |
1938 | ||
1939 | spin_lock(&poll->head->lock); | |
1940 | WRITE_ONCE(poll->canceled, true); | |
1941 | if (!list_empty(&poll->wait.entry)) { | |
1942 | list_del_init(&poll->wait.entry); | |
a197f664 | 1943 | io_queue_async_work(req); |
221c5eb2 JA |
1944 | } |
1945 | spin_unlock(&poll->head->lock); | |
1946 | ||
1947 | list_del_init(&req->list); | |
1948 | } | |
1949 | ||
1950 | static void io_poll_remove_all(struct io_ring_ctx *ctx) | |
1951 | { | |
1952 | struct io_kiocb *req; | |
1953 | ||
1954 | spin_lock_irq(&ctx->completion_lock); | |
1955 | while (!list_empty(&ctx->cancel_list)) { | |
1956 | req = list_first_entry(&ctx->cancel_list, struct io_kiocb,list); | |
1957 | io_poll_remove_one(req); | |
1958 | } | |
1959 | spin_unlock_irq(&ctx->completion_lock); | |
1960 | } | |
1961 | ||
47f46768 JA |
1962 | static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr) |
1963 | { | |
1964 | struct io_kiocb *req; | |
1965 | ||
1966 | list_for_each_entry(req, &ctx->cancel_list, list) { | |
1967 | if (req->user_data != sqe_addr) | |
1968 | continue; | |
1969 | io_poll_remove_one(req); | |
1970 | return 0; | |
1971 | } | |
1972 | ||
1973 | return -ENOENT; | |
1974 | } | |
1975 | ||
221c5eb2 JA |
1976 | /* |
1977 | * Find a running poll command that matches one specified in sqe->addr, | |
1978 | * and remove it if found. | |
1979 | */ | |
1980 | static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
1981 | { | |
1982 | struct io_ring_ctx *ctx = req->ctx; | |
47f46768 | 1983 | int ret; |
221c5eb2 JA |
1984 | |
1985 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1986 | return -EINVAL; | |
1987 | if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index || | |
1988 | sqe->poll_events) | |
1989 | return -EINVAL; | |
1990 | ||
1991 | spin_lock_irq(&ctx->completion_lock); | |
47f46768 | 1992 | ret = io_poll_cancel(ctx, READ_ONCE(sqe->addr)); |
221c5eb2 JA |
1993 | spin_unlock_irq(&ctx->completion_lock); |
1994 | ||
78e19bbe | 1995 | io_cqring_add_event(req, ret); |
f1f40853 JA |
1996 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1997 | req->flags |= REQ_F_FAIL_LINK; | |
ec9c02ad | 1998 | io_put_req(req); |
221c5eb2 JA |
1999 | return 0; |
2000 | } | |
2001 | ||
a197f664 | 2002 | static void io_poll_complete(struct io_kiocb *req, __poll_t mask) |
221c5eb2 | 2003 | { |
a197f664 JL |
2004 | struct io_ring_ctx *ctx = req->ctx; |
2005 | ||
8c838788 | 2006 | req->poll.done = true; |
78e19bbe | 2007 | io_cqring_fill_event(req, mangle_poll(mask)); |
8c838788 | 2008 | io_commit_cqring(ctx); |
221c5eb2 JA |
2009 | } |
2010 | ||
561fb04a | 2011 | static void io_poll_complete_work(struct io_wq_work **workptr) |
221c5eb2 | 2012 | { |
561fb04a | 2013 | struct io_wq_work *work = *workptr; |
221c5eb2 JA |
2014 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
2015 | struct io_poll_iocb *poll = &req->poll; | |
2016 | struct poll_table_struct pt = { ._key = poll->events }; | |
2017 | struct io_ring_ctx *ctx = req->ctx; | |
89723d0b | 2018 | struct io_kiocb *nxt = NULL; |
221c5eb2 JA |
2019 | __poll_t mask = 0; |
2020 | ||
561fb04a JA |
2021 | if (work->flags & IO_WQ_WORK_CANCEL) |
2022 | WRITE_ONCE(poll->canceled, true); | |
2023 | ||
221c5eb2 JA |
2024 | if (!READ_ONCE(poll->canceled)) |
2025 | mask = vfs_poll(poll->file, &pt) & poll->events; | |
2026 | ||
2027 | /* | |
2028 | * Note that ->ki_cancel callers also delete iocb from active_reqs after | |
2029 | * calling ->ki_cancel. We need the ctx_lock roundtrip here to | |
2030 | * synchronize with them. In the cancellation case the list_del_init | |
2031 | * itself is not actually needed, but harmless so we keep it in to | |
2032 | * avoid further branches in the fast path. | |
2033 | */ | |
2034 | spin_lock_irq(&ctx->completion_lock); | |
2035 | if (!mask && !READ_ONCE(poll->canceled)) { | |
2036 | add_wait_queue(poll->head, &poll->wait); | |
2037 | spin_unlock_irq(&ctx->completion_lock); | |
2038 | return; | |
2039 | } | |
2040 | list_del_init(&req->list); | |
a197f664 | 2041 | io_poll_complete(req, mask); |
221c5eb2 JA |
2042 | spin_unlock_irq(&ctx->completion_lock); |
2043 | ||
8c838788 | 2044 | io_cqring_ev_posted(ctx); |
89723d0b | 2045 | |
ec9c02ad | 2046 | io_put_req_find_next(req, &nxt); |
89723d0b JA |
2047 | if (nxt) |
2048 | *workptr = &nxt->work; | |
221c5eb2 JA |
2049 | } |
2050 | ||
2051 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, | |
2052 | void *key) | |
2053 | { | |
2054 | struct io_poll_iocb *poll = container_of(wait, struct io_poll_iocb, | |
2055 | wait); | |
2056 | struct io_kiocb *req = container_of(poll, struct io_kiocb, poll); | |
2057 | struct io_ring_ctx *ctx = req->ctx; | |
2058 | __poll_t mask = key_to_poll(key); | |
8c838788 | 2059 | unsigned long flags; |
221c5eb2 JA |
2060 | |
2061 | /* for instances that support it check for an event match first: */ | |
8c838788 JA |
2062 | if (mask && !(mask & poll->events)) |
2063 | return 0; | |
221c5eb2 | 2064 | |
8c838788 | 2065 | list_del_init(&poll->wait.entry); |
221c5eb2 | 2066 | |
8c838788 JA |
2067 | if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) { |
2068 | list_del(&req->list); | |
a197f664 | 2069 | io_poll_complete(req, mask); |
8c838788 | 2070 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
221c5eb2 | 2071 | |
8c838788 | 2072 | io_cqring_ev_posted(ctx); |
ec9c02ad | 2073 | io_put_req(req); |
8c838788 | 2074 | } else { |
a197f664 | 2075 | io_queue_async_work(req); |
221c5eb2 JA |
2076 | } |
2077 | ||
221c5eb2 JA |
2078 | return 1; |
2079 | } | |
2080 | ||
2081 | struct io_poll_table { | |
2082 | struct poll_table_struct pt; | |
2083 | struct io_kiocb *req; | |
2084 | int error; | |
2085 | }; | |
2086 | ||
2087 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, | |
2088 | struct poll_table_struct *p) | |
2089 | { | |
2090 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
2091 | ||
2092 | if (unlikely(pt->req->poll.head)) { | |
2093 | pt->error = -EINVAL; | |
2094 | return; | |
2095 | } | |
2096 | ||
2097 | pt->error = 0; | |
2098 | pt->req->poll.head = head; | |
2099 | add_wait_queue(head, &pt->req->poll.wait); | |
2100 | } | |
2101 | ||
89723d0b JA |
2102 | static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2103 | struct io_kiocb **nxt) | |
221c5eb2 JA |
2104 | { |
2105 | struct io_poll_iocb *poll = &req->poll; | |
2106 | struct io_ring_ctx *ctx = req->ctx; | |
2107 | struct io_poll_table ipt; | |
8c838788 | 2108 | bool cancel = false; |
221c5eb2 JA |
2109 | __poll_t mask; |
2110 | u16 events; | |
221c5eb2 JA |
2111 | |
2112 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
2113 | return -EINVAL; | |
2114 | if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index) | |
2115 | return -EINVAL; | |
09bb8394 JA |
2116 | if (!poll->file) |
2117 | return -EBADF; | |
221c5eb2 | 2118 | |
6cc47d1d | 2119 | req->submit.sqe = NULL; |
561fb04a | 2120 | INIT_IO_WORK(&req->work, io_poll_complete_work); |
221c5eb2 JA |
2121 | events = READ_ONCE(sqe->poll_events); |
2122 | poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP; | |
2123 | ||
221c5eb2 | 2124 | poll->head = NULL; |
8c838788 | 2125 | poll->done = false; |
221c5eb2 JA |
2126 | poll->canceled = false; |
2127 | ||
2128 | ipt.pt._qproc = io_poll_queue_proc; | |
2129 | ipt.pt._key = poll->events; | |
2130 | ipt.req = req; | |
2131 | ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */ | |
2132 | ||
2133 | /* initialized the list so that we can do list_empty checks */ | |
2134 | INIT_LIST_HEAD(&poll->wait.entry); | |
2135 | init_waitqueue_func_entry(&poll->wait, io_poll_wake); | |
2136 | ||
36703247 JA |
2137 | INIT_LIST_HEAD(&req->list); |
2138 | ||
221c5eb2 | 2139 | mask = vfs_poll(poll->file, &ipt.pt) & poll->events; |
221c5eb2 JA |
2140 | |
2141 | spin_lock_irq(&ctx->completion_lock); | |
8c838788 JA |
2142 | if (likely(poll->head)) { |
2143 | spin_lock(&poll->head->lock); | |
2144 | if (unlikely(list_empty(&poll->wait.entry))) { | |
2145 | if (ipt.error) | |
2146 | cancel = true; | |
2147 | ipt.error = 0; | |
2148 | mask = 0; | |
2149 | } | |
2150 | if (mask || ipt.error) | |
2151 | list_del_init(&poll->wait.entry); | |
2152 | else if (cancel) | |
2153 | WRITE_ONCE(poll->canceled, true); | |
2154 | else if (!poll->done) /* actually waiting for an event */ | |
2155 | list_add_tail(&req->list, &ctx->cancel_list); | |
2156 | spin_unlock(&poll->head->lock); | |
2157 | } | |
2158 | if (mask) { /* no async, we'd stolen it */ | |
221c5eb2 | 2159 | ipt.error = 0; |
a197f664 | 2160 | io_poll_complete(req, mask); |
221c5eb2 | 2161 | } |
221c5eb2 JA |
2162 | spin_unlock_irq(&ctx->completion_lock); |
2163 | ||
8c838788 JA |
2164 | if (mask) { |
2165 | io_cqring_ev_posted(ctx); | |
ec9c02ad | 2166 | io_put_req_find_next(req, nxt); |
221c5eb2 | 2167 | } |
8c838788 | 2168 | return ipt.error; |
221c5eb2 JA |
2169 | } |
2170 | ||
5262f567 JA |
2171 | static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer) |
2172 | { | |
2173 | struct io_ring_ctx *ctx; | |
11365043 | 2174 | struct io_kiocb *req; |
5262f567 JA |
2175 | unsigned long flags; |
2176 | ||
2177 | req = container_of(timer, struct io_kiocb, timeout.timer); | |
2178 | ctx = req->ctx; | |
2179 | atomic_inc(&ctx->cq_timeouts); | |
2180 | ||
2181 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
ef03681a | 2182 | /* |
11365043 JA |
2183 | * We could be racing with timeout deletion. If the list is empty, |
2184 | * then timeout lookup already found it and will be handling it. | |
ef03681a | 2185 | */ |
842f9612 | 2186 | if (!list_empty(&req->list)) { |
11365043 | 2187 | struct io_kiocb *prev; |
5262f567 | 2188 | |
11365043 JA |
2189 | /* |
2190 | * Adjust the reqs sequence before the current one because it | |
2191 | * will consume a slot in the cq_ring and the the cq_tail | |
2192 | * pointer will be increased, otherwise other timeout reqs may | |
2193 | * return in advance without waiting for enough wait_nr. | |
2194 | */ | |
2195 | prev = req; | |
2196 | list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list) | |
2197 | prev->sequence++; | |
11365043 | 2198 | list_del_init(&req->list); |
11365043 | 2199 | } |
842f9612 | 2200 | |
78e19bbe | 2201 | io_cqring_fill_event(req, -ETIME); |
842f9612 | 2202 | io_commit_cqring(ctx); |
5262f567 JA |
2203 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
2204 | ||
842f9612 | 2205 | io_cqring_ev_posted(ctx); |
f1f40853 JA |
2206 | if (req->flags & REQ_F_LINK) |
2207 | req->flags |= REQ_F_FAIL_LINK; | |
ec9c02ad | 2208 | io_put_req(req); |
11365043 JA |
2209 | return HRTIMER_NORESTART; |
2210 | } | |
2211 | ||
47f46768 JA |
2212 | static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data) |
2213 | { | |
2214 | struct io_kiocb *req; | |
2215 | int ret = -ENOENT; | |
2216 | ||
2217 | list_for_each_entry(req, &ctx->timeout_list, list) { | |
2218 | if (user_data == req->user_data) { | |
2219 | list_del_init(&req->list); | |
2220 | ret = 0; | |
2221 | break; | |
2222 | } | |
2223 | } | |
2224 | ||
2225 | if (ret == -ENOENT) | |
2226 | return ret; | |
2227 | ||
2228 | ret = hrtimer_try_to_cancel(&req->timeout.timer); | |
2229 | if (ret == -1) | |
2230 | return -EALREADY; | |
2231 | ||
2232 | io_cqring_fill_event(req, -ECANCELED); | |
2233 | io_put_req(req); | |
2234 | return 0; | |
2235 | } | |
2236 | ||
11365043 JA |
2237 | /* |
2238 | * Remove or update an existing timeout command | |
2239 | */ | |
2240 | static int io_timeout_remove(struct io_kiocb *req, | |
2241 | const struct io_uring_sqe *sqe) | |
2242 | { | |
2243 | struct io_ring_ctx *ctx = req->ctx; | |
11365043 | 2244 | unsigned flags; |
47f46768 | 2245 | int ret; |
11365043 JA |
2246 | |
2247 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
2248 | return -EINVAL; | |
2249 | if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len) | |
2250 | return -EINVAL; | |
2251 | flags = READ_ONCE(sqe->timeout_flags); | |
2252 | if (flags) | |
2253 | return -EINVAL; | |
2254 | ||
11365043 | 2255 | spin_lock_irq(&ctx->completion_lock); |
47f46768 | 2256 | ret = io_timeout_cancel(ctx, READ_ONCE(sqe->addr)); |
11365043 | 2257 | |
47f46768 | 2258 | io_cqring_fill_event(req, ret); |
11365043 JA |
2259 | io_commit_cqring(ctx); |
2260 | spin_unlock_irq(&ctx->completion_lock); | |
5262f567 | 2261 | io_cqring_ev_posted(ctx); |
47f46768 JA |
2262 | if (ret < 0 && req->flags & REQ_F_LINK) |
2263 | req->flags |= REQ_F_FAIL_LINK; | |
ec9c02ad | 2264 | io_put_req(req); |
11365043 | 2265 | return 0; |
5262f567 JA |
2266 | } |
2267 | ||
2268 | static int io_timeout(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
2269 | { | |
5da0fb1a | 2270 | unsigned count; |
5262f567 JA |
2271 | struct io_ring_ctx *ctx = req->ctx; |
2272 | struct list_head *entry; | |
a41525ab | 2273 | enum hrtimer_mode mode; |
bdf20073 | 2274 | struct timespec64 ts; |
a1f58ba4 | 2275 | unsigned span = 0; |
a41525ab | 2276 | unsigned flags; |
5262f567 JA |
2277 | |
2278 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
2279 | return -EINVAL; | |
a41525ab JA |
2280 | if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len != 1) |
2281 | return -EINVAL; | |
2282 | flags = READ_ONCE(sqe->timeout_flags); | |
2283 | if (flags & ~IORING_TIMEOUT_ABS) | |
5262f567 | 2284 | return -EINVAL; |
bdf20073 AB |
2285 | |
2286 | if (get_timespec64(&ts, u64_to_user_ptr(sqe->addr))) | |
5262f567 JA |
2287 | return -EFAULT; |
2288 | ||
11365043 JA |
2289 | if (flags & IORING_TIMEOUT_ABS) |
2290 | mode = HRTIMER_MODE_ABS; | |
2291 | else | |
2292 | mode = HRTIMER_MODE_REL; | |
2293 | ||
2294 | hrtimer_init(&req->timeout.timer, CLOCK_MONOTONIC, mode); | |
2295 | ||
5262f567 JA |
2296 | /* |
2297 | * sqe->off holds how many events that need to occur for this | |
2298 | * timeout event to be satisfied. | |
2299 | */ | |
2300 | count = READ_ONCE(sqe->off); | |
2301 | if (!count) | |
2302 | count = 1; | |
2303 | ||
2304 | req->sequence = ctx->cached_sq_head + count - 1; | |
5da0fb1a | 2305 | /* reuse it to store the count */ |
2306 | req->submit.sequence = count; | |
5262f567 JA |
2307 | req->flags |= REQ_F_TIMEOUT; |
2308 | ||
2309 | /* | |
2310 | * Insertion sort, ensuring the first entry in the list is always | |
2311 | * the one we need first. | |
2312 | */ | |
5262f567 JA |
2313 | spin_lock_irq(&ctx->completion_lock); |
2314 | list_for_each_prev(entry, &ctx->timeout_list) { | |
2315 | struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list); | |
5da0fb1a | 2316 | unsigned nxt_sq_head; |
2317 | long long tmp, tmp_nxt; | |
5262f567 | 2318 | |
5da0fb1a | 2319 | /* |
2320 | * Since cached_sq_head + count - 1 can overflow, use type long | |
2321 | * long to store it. | |
2322 | */ | |
2323 | tmp = (long long)ctx->cached_sq_head + count - 1; | |
2324 | nxt_sq_head = nxt->sequence - nxt->submit.sequence + 1; | |
2325 | tmp_nxt = (long long)nxt_sq_head + nxt->submit.sequence - 1; | |
2326 | ||
2327 | /* | |
2328 | * cached_sq_head may overflow, and it will never overflow twice | |
2329 | * once there is some timeout req still be valid. | |
2330 | */ | |
2331 | if (ctx->cached_sq_head < nxt_sq_head) | |
8b07a65a | 2332 | tmp += UINT_MAX; |
5da0fb1a | 2333 | |
a1f58ba4 | 2334 | if (tmp > tmp_nxt) |
5262f567 | 2335 | break; |
a1f58ba4 | 2336 | |
2337 | /* | |
2338 | * Sequence of reqs after the insert one and itself should | |
2339 | * be adjusted because each timeout req consumes a slot. | |
2340 | */ | |
2341 | span++; | |
2342 | nxt->sequence++; | |
5262f567 | 2343 | } |
a1f58ba4 | 2344 | req->sequence -= span; |
5262f567 | 2345 | list_add(&req->list, entry); |
5262f567 | 2346 | req->timeout.timer.function = io_timeout_fn; |
a41525ab | 2347 | hrtimer_start(&req->timeout.timer, timespec64_to_ktime(ts), mode); |
842f9612 | 2348 | spin_unlock_irq(&ctx->completion_lock); |
5262f567 JA |
2349 | return 0; |
2350 | } | |
2351 | ||
62755e35 JA |
2352 | static bool io_cancel_cb(struct io_wq_work *work, void *data) |
2353 | { | |
2354 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
2355 | ||
2356 | return req->user_data == (unsigned long) data; | |
2357 | } | |
2358 | ||
e977d6d3 | 2359 | static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr) |
62755e35 | 2360 | { |
62755e35 | 2361 | enum io_wq_cancel cancel_ret; |
62755e35 JA |
2362 | int ret = 0; |
2363 | ||
62755e35 JA |
2364 | cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr); |
2365 | switch (cancel_ret) { | |
2366 | case IO_WQ_CANCEL_OK: | |
2367 | ret = 0; | |
2368 | break; | |
2369 | case IO_WQ_CANCEL_RUNNING: | |
2370 | ret = -EALREADY; | |
2371 | break; | |
2372 | case IO_WQ_CANCEL_NOTFOUND: | |
2373 | ret = -ENOENT; | |
2374 | break; | |
2375 | } | |
2376 | ||
e977d6d3 JA |
2377 | return ret; |
2378 | } | |
2379 | ||
47f46768 JA |
2380 | static void io_async_find_and_cancel(struct io_ring_ctx *ctx, |
2381 | struct io_kiocb *req, __u64 sqe_addr, | |
2382 | struct io_kiocb **nxt) | |
2383 | { | |
2384 | unsigned long flags; | |
2385 | int ret; | |
2386 | ||
2387 | ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr); | |
2388 | if (ret != -ENOENT) { | |
2389 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
2390 | goto done; | |
2391 | } | |
2392 | ||
2393 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
2394 | ret = io_timeout_cancel(ctx, sqe_addr); | |
2395 | if (ret != -ENOENT) | |
2396 | goto done; | |
2397 | ret = io_poll_cancel(ctx, sqe_addr); | |
2398 | done: | |
2399 | io_cqring_fill_event(req, ret); | |
2400 | io_commit_cqring(ctx); | |
2401 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
2402 | io_cqring_ev_posted(ctx); | |
2403 | ||
2404 | if (ret < 0 && (req->flags & REQ_F_LINK)) | |
2405 | req->flags |= REQ_F_FAIL_LINK; | |
2406 | io_put_req_find_next(req, nxt); | |
2407 | } | |
2408 | ||
e977d6d3 JA |
2409 | static int io_async_cancel(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2410 | struct io_kiocb **nxt) | |
2411 | { | |
2412 | struct io_ring_ctx *ctx = req->ctx; | |
e977d6d3 JA |
2413 | |
2414 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
2415 | return -EINVAL; | |
2416 | if (sqe->flags || sqe->ioprio || sqe->off || sqe->len || | |
2417 | sqe->cancel_flags) | |
2418 | return -EINVAL; | |
2419 | ||
47f46768 | 2420 | io_async_find_and_cancel(ctx, req, READ_ONCE(sqe->addr), NULL); |
62755e35 JA |
2421 | return 0; |
2422 | } | |
2423 | ||
a197f664 | 2424 | static int io_req_defer(struct io_kiocb *req) |
de0617e4 | 2425 | { |
267bc904 | 2426 | const struct io_uring_sqe *sqe = req->submit.sqe; |
de0617e4 | 2427 | struct io_uring_sqe *sqe_copy; |
a197f664 | 2428 | struct io_ring_ctx *ctx = req->ctx; |
de0617e4 | 2429 | |
a197f664 | 2430 | if (!io_sequence_defer(req) && list_empty(&ctx->defer_list)) |
de0617e4 JA |
2431 | return 0; |
2432 | ||
2433 | sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL); | |
2434 | if (!sqe_copy) | |
2435 | return -EAGAIN; | |
2436 | ||
2437 | spin_lock_irq(&ctx->completion_lock); | |
a197f664 | 2438 | if (!io_sequence_defer(req) && list_empty(&ctx->defer_list)) { |
de0617e4 JA |
2439 | spin_unlock_irq(&ctx->completion_lock); |
2440 | kfree(sqe_copy); | |
2441 | return 0; | |
2442 | } | |
2443 | ||
2444 | memcpy(sqe_copy, sqe, sizeof(*sqe_copy)); | |
2445 | req->submit.sqe = sqe_copy; | |
2446 | ||
c826bd7a | 2447 | trace_io_uring_defer(ctx, req, false); |
de0617e4 JA |
2448 | list_add_tail(&req->list, &ctx->defer_list); |
2449 | spin_unlock_irq(&ctx->completion_lock); | |
2450 | return -EIOCBQUEUED; | |
2451 | } | |
2452 | ||
a197f664 JL |
2453 | static int __io_submit_sqe(struct io_kiocb *req, struct io_kiocb **nxt, |
2454 | bool force_nonblock) | |
2b188cc1 | 2455 | { |
e0c5c576 | 2456 | int ret, opcode; |
267bc904 | 2457 | struct sqe_submit *s = &req->submit; |
a197f664 | 2458 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 | 2459 | |
2b188cc1 JA |
2460 | opcode = READ_ONCE(s->sqe->opcode); |
2461 | switch (opcode) { | |
2462 | case IORING_OP_NOP: | |
78e19bbe | 2463 | ret = io_nop(req); |
2b188cc1 JA |
2464 | break; |
2465 | case IORING_OP_READV: | |
edafccee JA |
2466 | if (unlikely(s->sqe->buf_index)) |
2467 | return -EINVAL; | |
267bc904 | 2468 | ret = io_read(req, nxt, force_nonblock); |
2b188cc1 JA |
2469 | break; |
2470 | case IORING_OP_WRITEV: | |
edafccee JA |
2471 | if (unlikely(s->sqe->buf_index)) |
2472 | return -EINVAL; | |
267bc904 | 2473 | ret = io_write(req, nxt, force_nonblock); |
edafccee JA |
2474 | break; |
2475 | case IORING_OP_READ_FIXED: | |
267bc904 | 2476 | ret = io_read(req, nxt, force_nonblock); |
edafccee JA |
2477 | break; |
2478 | case IORING_OP_WRITE_FIXED: | |
267bc904 | 2479 | ret = io_write(req, nxt, force_nonblock); |
2b188cc1 | 2480 | break; |
c992fe29 | 2481 | case IORING_OP_FSYNC: |
ba816ad6 | 2482 | ret = io_fsync(req, s->sqe, nxt, force_nonblock); |
c992fe29 | 2483 | break; |
221c5eb2 | 2484 | case IORING_OP_POLL_ADD: |
89723d0b | 2485 | ret = io_poll_add(req, s->sqe, nxt); |
221c5eb2 JA |
2486 | break; |
2487 | case IORING_OP_POLL_REMOVE: | |
2488 | ret = io_poll_remove(req, s->sqe); | |
2489 | break; | |
5d17b4a4 | 2490 | case IORING_OP_SYNC_FILE_RANGE: |
ba816ad6 | 2491 | ret = io_sync_file_range(req, s->sqe, nxt, force_nonblock); |
5d17b4a4 | 2492 | break; |
0fa03c62 | 2493 | case IORING_OP_SENDMSG: |
ba816ad6 | 2494 | ret = io_sendmsg(req, s->sqe, nxt, force_nonblock); |
0fa03c62 | 2495 | break; |
aa1fa28f | 2496 | case IORING_OP_RECVMSG: |
ba816ad6 | 2497 | ret = io_recvmsg(req, s->sqe, nxt, force_nonblock); |
aa1fa28f | 2498 | break; |
5262f567 JA |
2499 | case IORING_OP_TIMEOUT: |
2500 | ret = io_timeout(req, s->sqe); | |
2501 | break; | |
11365043 JA |
2502 | case IORING_OP_TIMEOUT_REMOVE: |
2503 | ret = io_timeout_remove(req, s->sqe); | |
2504 | break; | |
17f2fe35 JA |
2505 | case IORING_OP_ACCEPT: |
2506 | ret = io_accept(req, s->sqe, nxt, force_nonblock); | |
2507 | break; | |
62755e35 JA |
2508 | case IORING_OP_ASYNC_CANCEL: |
2509 | ret = io_async_cancel(req, s->sqe, nxt); | |
2510 | break; | |
2b188cc1 JA |
2511 | default: |
2512 | ret = -EINVAL; | |
2513 | break; | |
2514 | } | |
2515 | ||
def596e9 JA |
2516 | if (ret) |
2517 | return ret; | |
2518 | ||
2519 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
9e645e11 | 2520 | if (req->result == -EAGAIN) |
def596e9 JA |
2521 | return -EAGAIN; |
2522 | ||
2523 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
ba5290cc | 2524 | if (s->in_async) |
def596e9 JA |
2525 | mutex_lock(&ctx->uring_lock); |
2526 | io_iopoll_req_issued(req); | |
ba5290cc | 2527 | if (s->in_async) |
def596e9 JA |
2528 | mutex_unlock(&ctx->uring_lock); |
2529 | } | |
2530 | ||
2531 | return 0; | |
2b188cc1 JA |
2532 | } |
2533 | ||
561fb04a | 2534 | static void io_wq_submit_work(struct io_wq_work **workptr) |
2b188cc1 | 2535 | { |
561fb04a | 2536 | struct io_wq_work *work = *workptr; |
2b188cc1 | 2537 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
561fb04a JA |
2538 | struct sqe_submit *s = &req->submit; |
2539 | const struct io_uring_sqe *sqe = s->sqe; | |
2540 | struct io_kiocb *nxt = NULL; | |
2541 | int ret = 0; | |
2b188cc1 | 2542 | |
561fb04a JA |
2543 | /* Ensure we clear previously set non-block flag */ |
2544 | req->rw.ki_flags &= ~IOCB_NOWAIT; | |
2b188cc1 | 2545 | |
561fb04a JA |
2546 | if (work->flags & IO_WQ_WORK_CANCEL) |
2547 | ret = -ECANCELED; | |
31b51510 | 2548 | |
561fb04a JA |
2549 | if (!ret) { |
2550 | s->has_user = (work->flags & IO_WQ_WORK_HAS_MM) != 0; | |
2551 | s->in_async = true; | |
2552 | do { | |
a197f664 | 2553 | ret = __io_submit_sqe(req, &nxt, false); |
561fb04a JA |
2554 | /* |
2555 | * We can get EAGAIN for polled IO even though we're | |
2556 | * forcing a sync submission from here, since we can't | |
2557 | * wait for request slots on the block side. | |
2558 | */ | |
2559 | if (ret != -EAGAIN) | |
2560 | break; | |
2561 | cond_resched(); | |
2562 | } while (1); | |
2563 | } | |
31b51510 | 2564 | |
561fb04a | 2565 | /* drop submission reference */ |
ec9c02ad | 2566 | io_put_req(req); |
817869d2 | 2567 | |
561fb04a | 2568 | if (ret) { |
f1f40853 JA |
2569 | if (req->flags & REQ_F_LINK) |
2570 | req->flags |= REQ_F_FAIL_LINK; | |
78e19bbe | 2571 | io_cqring_add_event(req, ret); |
ec9c02ad | 2572 | io_put_req(req); |
2b188cc1 | 2573 | } |
31b51510 | 2574 | |
561fb04a JA |
2575 | /* async context always use a copy of the sqe */ |
2576 | kfree(sqe); | |
31b51510 | 2577 | |
561fb04a JA |
2578 | /* if a dependent link is ready, pass it back */ |
2579 | if (!ret && nxt) { | |
2580 | io_prep_async_work(nxt); | |
2581 | *workptr = &nxt->work; | |
31b51510 | 2582 | } |
2b188cc1 JA |
2583 | } |
2584 | ||
09bb8394 JA |
2585 | static bool io_op_needs_file(const struct io_uring_sqe *sqe) |
2586 | { | |
2587 | int op = READ_ONCE(sqe->opcode); | |
2588 | ||
2589 | switch (op) { | |
2590 | case IORING_OP_NOP: | |
2591 | case IORING_OP_POLL_REMOVE: | |
2592 | return false; | |
2593 | default: | |
2594 | return true; | |
2595 | } | |
2596 | } | |
2597 | ||
65e19f54 JA |
2598 | static inline struct file *io_file_from_index(struct io_ring_ctx *ctx, |
2599 | int index) | |
2600 | { | |
2601 | struct fixed_file_table *table; | |
2602 | ||
2603 | table = &ctx->file_table[index >> IORING_FILE_TABLE_SHIFT]; | |
2604 | return table->files[index & IORING_FILE_TABLE_MASK]; | |
2605 | } | |
2606 | ||
a197f664 | 2607 | static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req) |
09bb8394 | 2608 | { |
267bc904 | 2609 | struct sqe_submit *s = &req->submit; |
a197f664 | 2610 | struct io_ring_ctx *ctx = req->ctx; |
09bb8394 JA |
2611 | unsigned flags; |
2612 | int fd; | |
2613 | ||
2614 | flags = READ_ONCE(s->sqe->flags); | |
2615 | fd = READ_ONCE(s->sqe->fd); | |
2616 | ||
4fe2c963 | 2617 | if (flags & IOSQE_IO_DRAIN) |
de0617e4 | 2618 | req->flags |= REQ_F_IO_DRAIN; |
4fe2c963 JL |
2619 | /* |
2620 | * All io need record the previous position, if LINK vs DARIN, | |
2621 | * it can be used to mark the position of the first IO in the | |
2622 | * link list. | |
2623 | */ | |
2624 | req->sequence = s->sequence; | |
de0617e4 | 2625 | |
60c112b0 | 2626 | if (!io_op_needs_file(s->sqe)) |
09bb8394 | 2627 | return 0; |
09bb8394 JA |
2628 | |
2629 | if (flags & IOSQE_FIXED_FILE) { | |
65e19f54 | 2630 | if (unlikely(!ctx->file_table || |
09bb8394 JA |
2631 | (unsigned) fd >= ctx->nr_user_files)) |
2632 | return -EBADF; | |
b7620121 | 2633 | fd = array_index_nospec(fd, ctx->nr_user_files); |
65e19f54 JA |
2634 | req->file = io_file_from_index(ctx, fd); |
2635 | if (!req->file) | |
08a45173 | 2636 | return -EBADF; |
09bb8394 JA |
2637 | req->flags |= REQ_F_FIXED_FILE; |
2638 | } else { | |
2639 | if (s->needs_fixed_file) | |
2640 | return -EBADF; | |
c826bd7a | 2641 | trace_io_uring_file_get(ctx, fd); |
09bb8394 JA |
2642 | req->file = io_file_get(state, fd); |
2643 | if (unlikely(!req->file)) | |
2644 | return -EBADF; | |
2645 | } | |
2646 | ||
2647 | return 0; | |
2648 | } | |
2649 | ||
a197f664 | 2650 | static int io_grab_files(struct io_kiocb *req) |
fcb323cc JA |
2651 | { |
2652 | int ret = -EBADF; | |
a197f664 | 2653 | struct io_ring_ctx *ctx = req->ctx; |
fcb323cc JA |
2654 | |
2655 | rcu_read_lock(); | |
2656 | spin_lock_irq(&ctx->inflight_lock); | |
2657 | /* | |
2658 | * We use the f_ops->flush() handler to ensure that we can flush | |
2659 | * out work accessing these files if the fd is closed. Check if | |
2660 | * the fd has changed since we started down this path, and disallow | |
2661 | * this operation if it has. | |
2662 | */ | |
2663 | if (fcheck(req->submit.ring_fd) == req->submit.ring_file) { | |
2664 | list_add(&req->inflight_entry, &ctx->inflight_list); | |
2665 | req->flags |= REQ_F_INFLIGHT; | |
2666 | req->work.files = current->files; | |
2667 | ret = 0; | |
2668 | } | |
2669 | spin_unlock_irq(&ctx->inflight_lock); | |
2670 | rcu_read_unlock(); | |
2671 | ||
2672 | return ret; | |
2673 | } | |
2674 | ||
2665abfd JA |
2675 | static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer) |
2676 | { | |
2677 | struct io_kiocb *req = container_of(timer, struct io_kiocb, | |
2678 | timeout.timer); | |
2679 | struct io_ring_ctx *ctx = req->ctx; | |
2680 | struct io_kiocb *prev = NULL; | |
2681 | unsigned long flags; | |
2665abfd JA |
2682 | |
2683 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
2684 | ||
2685 | /* | |
2686 | * We don't expect the list to be empty, that will only happen if we | |
2687 | * race with the completion of the linked work. | |
2688 | */ | |
2689 | if (!list_empty(&req->list)) { | |
2690 | prev = list_entry(req->list.prev, struct io_kiocb, link_list); | |
2691 | list_del_init(&req->list); | |
2692 | } | |
2693 | ||
2694 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
2695 | ||
2696 | if (prev) { | |
47f46768 JA |
2697 | io_async_find_and_cancel(ctx, req, prev->user_data, NULL); |
2698 | } else { | |
2699 | io_cqring_add_event(req, -ETIME); | |
2700 | io_put_req(req); | |
2665abfd | 2701 | } |
2665abfd JA |
2702 | return HRTIMER_NORESTART; |
2703 | } | |
2704 | ||
2705 | static int io_queue_linked_timeout(struct io_kiocb *req, struct io_kiocb *nxt) | |
2706 | { | |
2707 | const struct io_uring_sqe *sqe = nxt->submit.sqe; | |
2708 | enum hrtimer_mode mode; | |
2709 | struct timespec64 ts; | |
2710 | int ret = -EINVAL; | |
2711 | ||
2712 | if (sqe->ioprio || sqe->buf_index || sqe->len != 1 || sqe->off) | |
2713 | goto err; | |
2714 | if (sqe->timeout_flags & ~IORING_TIMEOUT_ABS) | |
2715 | goto err; | |
2716 | if (get_timespec64(&ts, u64_to_user_ptr(sqe->addr))) { | |
2717 | ret = -EFAULT; | |
2718 | goto err; | |
2719 | } | |
2720 | ||
2721 | req->flags |= REQ_F_LINK_TIMEOUT; | |
2722 | ||
2723 | if (sqe->timeout_flags & IORING_TIMEOUT_ABS) | |
2724 | mode = HRTIMER_MODE_ABS; | |
2725 | else | |
2726 | mode = HRTIMER_MODE_REL; | |
2727 | hrtimer_init(&nxt->timeout.timer, CLOCK_MONOTONIC, mode); | |
2728 | nxt->timeout.timer.function = io_link_timeout_fn; | |
2729 | hrtimer_start(&nxt->timeout.timer, timespec64_to_ktime(ts), mode); | |
2730 | ret = 0; | |
2731 | err: | |
2732 | /* drop submission reference */ | |
ec9c02ad | 2733 | io_put_req(nxt); |
2665abfd JA |
2734 | |
2735 | if (ret) { | |
2736 | struct io_ring_ctx *ctx = req->ctx; | |
2737 | ||
2738 | /* | |
2739 | * Break the link and fail linked timeout, parent will get | |
2740 | * failed by the regular submission path. | |
2741 | */ | |
2742 | list_del(&nxt->list); | |
78e19bbe | 2743 | io_cqring_fill_event(nxt, ret); |
2665abfd JA |
2744 | trace_io_uring_fail_link(req, nxt); |
2745 | io_commit_cqring(ctx); | |
ec9c02ad | 2746 | io_put_req(nxt); |
2665abfd JA |
2747 | ret = -ECANCELED; |
2748 | } | |
2749 | ||
2750 | return ret; | |
2751 | } | |
2752 | ||
2753 | static inline struct io_kiocb *io_get_linked_timeout(struct io_kiocb *req) | |
2754 | { | |
2755 | struct io_kiocb *nxt; | |
2756 | ||
2757 | if (!(req->flags & REQ_F_LINK)) | |
2758 | return NULL; | |
2759 | ||
2760 | nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list); | |
2761 | if (nxt && nxt->submit.sqe->opcode == IORING_OP_LINK_TIMEOUT) | |
2762 | return nxt; | |
2763 | ||
2764 | return NULL; | |
2765 | } | |
2766 | ||
a197f664 | 2767 | static int __io_queue_sqe(struct io_kiocb *req) |
2b188cc1 | 2768 | { |
2665abfd | 2769 | struct io_kiocb *nxt; |
e0c5c576 | 2770 | int ret; |
2b188cc1 | 2771 | |
2665abfd JA |
2772 | nxt = io_get_linked_timeout(req); |
2773 | if (unlikely(nxt)) { | |
2774 | ret = io_queue_linked_timeout(req, nxt); | |
2775 | if (ret) | |
2776 | goto err; | |
2777 | } | |
2778 | ||
a197f664 | 2779 | ret = __io_submit_sqe(req, NULL, true); |
491381ce JA |
2780 | |
2781 | /* | |
2782 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
2783 | * doesn't support non-blocking read/write attempts | |
2784 | */ | |
2785 | if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) || | |
2786 | (req->flags & REQ_F_MUST_PUNT))) { | |
267bc904 | 2787 | struct sqe_submit *s = &req->submit; |
2b188cc1 JA |
2788 | struct io_uring_sqe *sqe_copy; |
2789 | ||
954dab19 | 2790 | sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL); |
2b188cc1 | 2791 | if (sqe_copy) { |
2b188cc1 | 2792 | s->sqe = sqe_copy; |
fcb323cc | 2793 | if (req->work.flags & IO_WQ_WORK_NEEDS_FILES) { |
a197f664 | 2794 | ret = io_grab_files(req); |
fcb323cc JA |
2795 | if (ret) { |
2796 | kfree(sqe_copy); | |
2797 | goto err; | |
2798 | } | |
2799 | } | |
e65ef56d JA |
2800 | |
2801 | /* | |
2802 | * Queued up for async execution, worker will release | |
9e645e11 | 2803 | * submit reference when the iocb is actually submitted. |
e65ef56d | 2804 | */ |
a197f664 | 2805 | io_queue_async_work(req); |
e65ef56d | 2806 | return 0; |
2b188cc1 JA |
2807 | } |
2808 | } | |
e65ef56d JA |
2809 | |
2810 | /* drop submission reference */ | |
fcb323cc | 2811 | err: |
ec9c02ad | 2812 | io_put_req(req); |
e65ef56d JA |
2813 | |
2814 | /* and drop final reference, if we failed */ | |
9e645e11 | 2815 | if (ret) { |
78e19bbe | 2816 | io_cqring_add_event(req, ret); |
9e645e11 JA |
2817 | if (req->flags & REQ_F_LINK) |
2818 | req->flags |= REQ_F_FAIL_LINK; | |
ec9c02ad | 2819 | io_put_req(req); |
9e645e11 | 2820 | } |
2b188cc1 JA |
2821 | |
2822 | return ret; | |
2823 | } | |
2824 | ||
a197f664 | 2825 | static int io_queue_sqe(struct io_kiocb *req) |
4fe2c963 JL |
2826 | { |
2827 | int ret; | |
2828 | ||
a197f664 | 2829 | ret = io_req_defer(req); |
4fe2c963 JL |
2830 | if (ret) { |
2831 | if (ret != -EIOCBQUEUED) { | |
78e19bbe JA |
2832 | io_cqring_add_event(req, ret); |
2833 | io_double_put_req(req); | |
4fe2c963 JL |
2834 | } |
2835 | return 0; | |
2836 | } | |
2837 | ||
a197f664 | 2838 | return __io_queue_sqe(req); |
4fe2c963 JL |
2839 | } |
2840 | ||
a197f664 | 2841 | static int io_queue_link_head(struct io_kiocb *req, struct io_kiocb *shadow) |
4fe2c963 JL |
2842 | { |
2843 | int ret; | |
2844 | int need_submit = false; | |
a197f664 | 2845 | struct io_ring_ctx *ctx = req->ctx; |
4fe2c963 JL |
2846 | |
2847 | if (!shadow) | |
a197f664 | 2848 | return io_queue_sqe(req); |
4fe2c963 JL |
2849 | |
2850 | /* | |
2851 | * Mark the first IO in link list as DRAIN, let all the following | |
2852 | * IOs enter the defer list. all IO needs to be completed before link | |
2853 | * list. | |
2854 | */ | |
2855 | req->flags |= REQ_F_IO_DRAIN; | |
a197f664 | 2856 | ret = io_req_defer(req); |
4fe2c963 JL |
2857 | if (ret) { |
2858 | if (ret != -EIOCBQUEUED) { | |
78e19bbe JA |
2859 | io_cqring_add_event(req, ret); |
2860 | io_double_put_req(req); | |
7b20238d | 2861 | __io_free_req(shadow); |
4fe2c963 JL |
2862 | return 0; |
2863 | } | |
2864 | } else { | |
2865 | /* | |
2866 | * If ret == 0 means that all IOs in front of link io are | |
2867 | * running done. let's queue link head. | |
2868 | */ | |
2869 | need_submit = true; | |
2870 | } | |
2871 | ||
2872 | /* Insert shadow req to defer_list, blocking next IOs */ | |
2873 | spin_lock_irq(&ctx->completion_lock); | |
c826bd7a | 2874 | trace_io_uring_defer(ctx, shadow, true); |
4fe2c963 JL |
2875 | list_add_tail(&shadow->list, &ctx->defer_list); |
2876 | spin_unlock_irq(&ctx->completion_lock); | |
2877 | ||
2878 | if (need_submit) | |
a197f664 | 2879 | return __io_queue_sqe(req); |
4fe2c963 JL |
2880 | |
2881 | return 0; | |
2882 | } | |
2883 | ||
9e645e11 JA |
2884 | #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK) |
2885 | ||
a197f664 JL |
2886 | static void io_submit_sqe(struct io_kiocb *req, struct io_submit_state *state, |
2887 | struct io_kiocb **link) | |
9e645e11 JA |
2888 | { |
2889 | struct io_uring_sqe *sqe_copy; | |
267bc904 | 2890 | struct sqe_submit *s = &req->submit; |
a197f664 | 2891 | struct io_ring_ctx *ctx = req->ctx; |
9e645e11 JA |
2892 | int ret; |
2893 | ||
78e19bbe JA |
2894 | req->user_data = s->sqe->user_data; |
2895 | ||
9e645e11 JA |
2896 | /* enforce forwards compatibility on users */ |
2897 | if (unlikely(s->sqe->flags & ~SQE_VALID_FLAGS)) { | |
2898 | ret = -EINVAL; | |
196be95c | 2899 | goto err_req; |
9e645e11 JA |
2900 | } |
2901 | ||
a197f664 | 2902 | ret = io_req_set_file(state, req); |
9e645e11 JA |
2903 | if (unlikely(ret)) { |
2904 | err_req: | |
78e19bbe JA |
2905 | io_cqring_add_event(req, ret); |
2906 | io_double_put_req(req); | |
9e645e11 JA |
2907 | return; |
2908 | } | |
2909 | ||
9e645e11 JA |
2910 | /* |
2911 | * If we already have a head request, queue this one for async | |
2912 | * submittal once the head completes. If we don't have a head but | |
2913 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
2914 | * submitted sync once the chain is complete. If none of those | |
2915 | * conditions are true (normal request), then just queue it. | |
2916 | */ | |
2917 | if (*link) { | |
2918 | struct io_kiocb *prev = *link; | |
2919 | ||
2920 | sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL); | |
2921 | if (!sqe_copy) { | |
2922 | ret = -EAGAIN; | |
2923 | goto err_req; | |
2924 | } | |
2925 | ||
2926 | s->sqe = sqe_copy; | |
c826bd7a | 2927 | trace_io_uring_link(ctx, req, prev); |
9e645e11 JA |
2928 | list_add_tail(&req->list, &prev->link_list); |
2929 | } else if (s->sqe->flags & IOSQE_IO_LINK) { | |
2930 | req->flags |= REQ_F_LINK; | |
2931 | ||
9e645e11 JA |
2932 | INIT_LIST_HEAD(&req->link_list); |
2933 | *link = req; | |
2665abfd JA |
2934 | } else if (READ_ONCE(s->sqe->opcode) == IORING_OP_LINK_TIMEOUT) { |
2935 | /* Only valid as a linked SQE */ | |
2936 | ret = -EINVAL; | |
2937 | goto err_req; | |
9e645e11 | 2938 | } else { |
a197f664 | 2939 | io_queue_sqe(req); |
9e645e11 JA |
2940 | } |
2941 | } | |
2942 | ||
9a56a232 JA |
2943 | /* |
2944 | * Batched submission is done, ensure local IO is flushed out. | |
2945 | */ | |
2946 | static void io_submit_state_end(struct io_submit_state *state) | |
2947 | { | |
2948 | blk_finish_plug(&state->plug); | |
3d6770fb | 2949 | io_file_put(state); |
2579f913 JA |
2950 | if (state->free_reqs) |
2951 | kmem_cache_free_bulk(req_cachep, state->free_reqs, | |
2952 | &state->reqs[state->cur_req]); | |
9a56a232 JA |
2953 | } |
2954 | ||
2955 | /* | |
2956 | * Start submission side cache. | |
2957 | */ | |
2958 | static void io_submit_state_start(struct io_submit_state *state, | |
2959 | struct io_ring_ctx *ctx, unsigned max_ios) | |
2960 | { | |
2961 | blk_start_plug(&state->plug); | |
2579f913 | 2962 | state->free_reqs = 0; |
9a56a232 JA |
2963 | state->file = NULL; |
2964 | state->ios_left = max_ios; | |
2965 | } | |
2966 | ||
2b188cc1 JA |
2967 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
2968 | { | |
75b28aff | 2969 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 2970 | |
75b28aff | 2971 | if (ctx->cached_sq_head != READ_ONCE(rings->sq.head)) { |
2b188cc1 JA |
2972 | /* |
2973 | * Ensure any loads from the SQEs are done at this point, | |
2974 | * since once we write the new head, the application could | |
2975 | * write new data to them. | |
2976 | */ | |
75b28aff | 2977 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); |
2b188cc1 JA |
2978 | } |
2979 | } | |
2980 | ||
2b188cc1 JA |
2981 | /* |
2982 | * Fetch an sqe, if one is available. Note that s->sqe will point to memory | |
2983 | * that is mapped by userspace. This means that care needs to be taken to | |
2984 | * ensure that reads are stable, as we cannot rely on userspace always | |
2985 | * being a good citizen. If members of the sqe are validated and then later | |
2986 | * used, it's important that those reads are done through READ_ONCE() to | |
2987 | * prevent a re-load down the line. | |
2988 | */ | |
2989 | static bool io_get_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s) | |
2990 | { | |
75b28aff HV |
2991 | struct io_rings *rings = ctx->rings; |
2992 | u32 *sq_array = ctx->sq_array; | |
2b188cc1 JA |
2993 | unsigned head; |
2994 | ||
2995 | /* | |
2996 | * The cached sq head (or cq tail) serves two purposes: | |
2997 | * | |
2998 | * 1) allows us to batch the cost of updating the user visible | |
2999 | * head updates. | |
3000 | * 2) allows the kernel side to track the head on its own, even | |
3001 | * though the application is the one updating it. | |
3002 | */ | |
3003 | head = ctx->cached_sq_head; | |
e523a29c | 3004 | /* make sure SQ entry isn't read before tail */ |
75b28aff | 3005 | if (head == smp_load_acquire(&rings->sq.tail)) |
2b188cc1 JA |
3006 | return false; |
3007 | ||
75b28aff | 3008 | head = READ_ONCE(sq_array[head & ctx->sq_mask]); |
2b188cc1 | 3009 | if (head < ctx->sq_entries) { |
fcb323cc | 3010 | s->ring_file = NULL; |
2b188cc1 | 3011 | s->sqe = &ctx->sq_sqes[head]; |
8776f3fa | 3012 | s->sequence = ctx->cached_sq_head; |
2b188cc1 JA |
3013 | ctx->cached_sq_head++; |
3014 | return true; | |
3015 | } | |
3016 | ||
3017 | /* drop invalid entries */ | |
3018 | ctx->cached_sq_head++; | |
498ccd9e JA |
3019 | ctx->cached_sq_dropped++; |
3020 | WRITE_ONCE(rings->sq_dropped, ctx->cached_sq_dropped); | |
2b188cc1 JA |
3021 | return false; |
3022 | } | |
3023 | ||
fb5ccc98 | 3024 | static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr, |
ae9428ca PB |
3025 | struct file *ring_file, int ring_fd, |
3026 | struct mm_struct **mm, bool async) | |
6c271ce2 JA |
3027 | { |
3028 | struct io_submit_state state, *statep = NULL; | |
9e645e11 | 3029 | struct io_kiocb *link = NULL; |
4fe2c963 | 3030 | struct io_kiocb *shadow_req = NULL; |
9e645e11 | 3031 | int i, submitted = 0; |
95a1b3ff | 3032 | bool mm_fault = false; |
6c271ce2 | 3033 | |
1d7bb1d5 JA |
3034 | if (!list_empty(&ctx->cq_overflow_list)) { |
3035 | io_cqring_overflow_flush(ctx, false); | |
3036 | return -EBUSY; | |
3037 | } | |
3038 | ||
6c271ce2 JA |
3039 | if (nr > IO_PLUG_THRESHOLD) { |
3040 | io_submit_state_start(&state, ctx, nr); | |
3041 | statep = &state; | |
3042 | } | |
3043 | ||
3044 | for (i = 0; i < nr; i++) { | |
196be95c | 3045 | struct io_kiocb *req; |
50585b9a | 3046 | unsigned int sqe_flags; |
fb5ccc98 | 3047 | |
196be95c PB |
3048 | req = io_get_req(ctx, statep); |
3049 | if (unlikely(!req)) { | |
3050 | if (!submitted) | |
3051 | submitted = -EAGAIN; | |
fb5ccc98 | 3052 | break; |
196be95c | 3053 | } |
50585b9a | 3054 | if (!io_get_sqring(ctx, &req->submit)) { |
196be95c PB |
3055 | __io_free_req(req); |
3056 | break; | |
3057 | } | |
fb5ccc98 | 3058 | |
50585b9a | 3059 | if (io_sqe_needs_user(req->submit.sqe) && !*mm) { |
95a1b3ff PB |
3060 | mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm); |
3061 | if (!mm_fault) { | |
3062 | use_mm(ctx->sqo_mm); | |
3063 | *mm = ctx->sqo_mm; | |
3064 | } | |
3065 | } | |
3066 | ||
50585b9a PB |
3067 | sqe_flags = req->submit.sqe->flags; |
3068 | ||
3069 | if (link && (sqe_flags & IOSQE_IO_DRAIN)) { | |
4fe2c963 JL |
3070 | if (!shadow_req) { |
3071 | shadow_req = io_get_req(ctx, NULL); | |
a1041c27 JL |
3072 | if (unlikely(!shadow_req)) |
3073 | goto out; | |
4fe2c963 JL |
3074 | shadow_req->flags |= (REQ_F_IO_DRAIN | REQ_F_SHADOW_DRAIN); |
3075 | refcount_dec(&shadow_req->refs); | |
3076 | } | |
50585b9a | 3077 | shadow_req->sequence = req->submit.sequence; |
4fe2c963 JL |
3078 | } |
3079 | ||
a1041c27 | 3080 | out: |
50585b9a PB |
3081 | req->submit.ring_file = ring_file; |
3082 | req->submit.ring_fd = ring_fd; | |
3083 | req->submit.has_user = *mm != NULL; | |
3084 | req->submit.in_async = async; | |
3085 | req->submit.needs_fixed_file = async; | |
3086 | trace_io_uring_submit_sqe(ctx, req->submit.sqe->user_data, | |
3087 | true, async); | |
a197f664 | 3088 | io_submit_sqe(req, statep, &link); |
95a1b3ff | 3089 | submitted++; |
e5eb6366 PB |
3090 | |
3091 | /* | |
3092 | * If previous wasn't linked and we have a linked command, | |
3093 | * that's the end of the chain. Submit the previous link. | |
3094 | */ | |
50585b9a | 3095 | if (!(sqe_flags & IOSQE_IO_LINK) && link) { |
a197f664 | 3096 | io_queue_link_head(link, shadow_req); |
e5eb6366 PB |
3097 | link = NULL; |
3098 | shadow_req = NULL; | |
3099 | } | |
6c271ce2 JA |
3100 | } |
3101 | ||
9e645e11 | 3102 | if (link) |
a197f664 | 3103 | io_queue_link_head(link, shadow_req); |
6c271ce2 JA |
3104 | if (statep) |
3105 | io_submit_state_end(&state); | |
3106 | ||
ae9428ca PB |
3107 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
3108 | io_commit_sqring(ctx); | |
3109 | ||
6c271ce2 JA |
3110 | return submitted; |
3111 | } | |
3112 | ||
3113 | static int io_sq_thread(void *data) | |
3114 | { | |
6c271ce2 JA |
3115 | struct io_ring_ctx *ctx = data; |
3116 | struct mm_struct *cur_mm = NULL; | |
3117 | mm_segment_t old_fs; | |
3118 | DEFINE_WAIT(wait); | |
3119 | unsigned inflight; | |
3120 | unsigned long timeout; | |
3121 | ||
206aefde | 3122 | complete(&ctx->completions[1]); |
a4c0b3de | 3123 | |
6c271ce2 JA |
3124 | old_fs = get_fs(); |
3125 | set_fs(USER_DS); | |
3126 | ||
3127 | timeout = inflight = 0; | |
2bbcd6d3 | 3128 | while (!kthread_should_park()) { |
fb5ccc98 | 3129 | unsigned int to_submit; |
1d7bb1d5 | 3130 | int ret; |
6c271ce2 JA |
3131 | |
3132 | if (inflight) { | |
3133 | unsigned nr_events = 0; | |
3134 | ||
3135 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
2b2ed975 JA |
3136 | /* |
3137 | * inflight is the count of the maximum possible | |
3138 | * entries we submitted, but it can be smaller | |
3139 | * if we dropped some of them. If we don't have | |
3140 | * poll entries available, then we know that we | |
3141 | * have nothing left to poll for. Reset the | |
3142 | * inflight count to zero in that case. | |
3143 | */ | |
3144 | mutex_lock(&ctx->uring_lock); | |
3145 | if (!list_empty(&ctx->poll_list)) | |
3146 | __io_iopoll_check(ctx, &nr_events, 0); | |
3147 | else | |
3148 | inflight = 0; | |
3149 | mutex_unlock(&ctx->uring_lock); | |
6c271ce2 JA |
3150 | } else { |
3151 | /* | |
3152 | * Normal IO, just pretend everything completed. | |
3153 | * We don't have to poll completions for that. | |
3154 | */ | |
3155 | nr_events = inflight; | |
3156 | } | |
3157 | ||
3158 | inflight -= nr_events; | |
3159 | if (!inflight) | |
3160 | timeout = jiffies + ctx->sq_thread_idle; | |
3161 | } | |
3162 | ||
fb5ccc98 PB |
3163 | to_submit = io_sqring_entries(ctx); |
3164 | if (!to_submit) { | |
6c271ce2 JA |
3165 | /* |
3166 | * We're polling. If we're within the defined idle | |
3167 | * period, then let us spin without work before going | |
3168 | * to sleep. | |
3169 | */ | |
3170 | if (inflight || !time_after(jiffies, timeout)) { | |
9831a90c | 3171 | cond_resched(); |
6c271ce2 JA |
3172 | continue; |
3173 | } | |
3174 | ||
3175 | /* | |
3176 | * Drop cur_mm before scheduling, we can't hold it for | |
3177 | * long periods (or over schedule()). Do this before | |
3178 | * adding ourselves to the waitqueue, as the unuse/drop | |
3179 | * may sleep. | |
3180 | */ | |
3181 | if (cur_mm) { | |
3182 | unuse_mm(cur_mm); | |
3183 | mmput(cur_mm); | |
3184 | cur_mm = NULL; | |
3185 | } | |
3186 | ||
3187 | prepare_to_wait(&ctx->sqo_wait, &wait, | |
3188 | TASK_INTERRUPTIBLE); | |
3189 | ||
3190 | /* Tell userspace we may need a wakeup call */ | |
75b28aff | 3191 | ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP; |
0d7bae69 SB |
3192 | /* make sure to read SQ tail after writing flags */ |
3193 | smp_mb(); | |
6c271ce2 | 3194 | |
fb5ccc98 PB |
3195 | to_submit = io_sqring_entries(ctx); |
3196 | if (!to_submit) { | |
2bbcd6d3 | 3197 | if (kthread_should_park()) { |
6c271ce2 JA |
3198 | finish_wait(&ctx->sqo_wait, &wait); |
3199 | break; | |
3200 | } | |
3201 | if (signal_pending(current)) | |
3202 | flush_signals(current); | |
3203 | schedule(); | |
3204 | finish_wait(&ctx->sqo_wait, &wait); | |
3205 | ||
75b28aff | 3206 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
3207 | continue; |
3208 | } | |
3209 | finish_wait(&ctx->sqo_wait, &wait); | |
3210 | ||
75b28aff | 3211 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
3212 | } |
3213 | ||
fb5ccc98 | 3214 | to_submit = min(to_submit, ctx->sq_entries); |
1d7bb1d5 JA |
3215 | ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true); |
3216 | if (ret > 0) | |
3217 | inflight += ret; | |
6c271ce2 JA |
3218 | } |
3219 | ||
3220 | set_fs(old_fs); | |
3221 | if (cur_mm) { | |
3222 | unuse_mm(cur_mm); | |
3223 | mmput(cur_mm); | |
3224 | } | |
06058632 | 3225 | |
2bbcd6d3 | 3226 | kthread_parkme(); |
06058632 | 3227 | |
6c271ce2 JA |
3228 | return 0; |
3229 | } | |
3230 | ||
bda52162 JA |
3231 | struct io_wait_queue { |
3232 | struct wait_queue_entry wq; | |
3233 | struct io_ring_ctx *ctx; | |
3234 | unsigned to_wait; | |
3235 | unsigned nr_timeouts; | |
3236 | }; | |
3237 | ||
1d7bb1d5 | 3238 | static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush) |
bda52162 JA |
3239 | { |
3240 | struct io_ring_ctx *ctx = iowq->ctx; | |
3241 | ||
3242 | /* | |
3243 | * Wake up if we have enough events, or if a timeout occured since we | |
3244 | * started waiting. For timeouts, we always want to return to userspace, | |
3245 | * regardless of event count. | |
3246 | */ | |
1d7bb1d5 | 3247 | return io_cqring_events(ctx, noflush) >= iowq->to_wait || |
bda52162 JA |
3248 | atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
3249 | } | |
3250 | ||
3251 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
3252 | int wake_flags, void *key) | |
3253 | { | |
3254 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, | |
3255 | wq); | |
3256 | ||
1d7bb1d5 JA |
3257 | /* use noflush == true, as we can't safely rely on locking context */ |
3258 | if (!io_should_wake(iowq, true)) | |
bda52162 JA |
3259 | return -1; |
3260 | ||
3261 | return autoremove_wake_function(curr, mode, wake_flags, key); | |
3262 | } | |
3263 | ||
2b188cc1 JA |
3264 | /* |
3265 | * Wait until events become available, if we don't already have some. The | |
3266 | * application must reap them itself, as they reside on the shared cq ring. | |
3267 | */ | |
3268 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
3269 | const sigset_t __user *sig, size_t sigsz) | |
3270 | { | |
bda52162 JA |
3271 | struct io_wait_queue iowq = { |
3272 | .wq = { | |
3273 | .private = current, | |
3274 | .func = io_wake_function, | |
3275 | .entry = LIST_HEAD_INIT(iowq.wq.entry), | |
3276 | }, | |
3277 | .ctx = ctx, | |
3278 | .to_wait = min_events, | |
3279 | }; | |
75b28aff | 3280 | struct io_rings *rings = ctx->rings; |
e9ffa5c2 | 3281 | int ret = 0; |
2b188cc1 | 3282 | |
1d7bb1d5 | 3283 | if (io_cqring_events(ctx, false) >= min_events) |
2b188cc1 JA |
3284 | return 0; |
3285 | ||
3286 | if (sig) { | |
9e75ad5d AB |
3287 | #ifdef CONFIG_COMPAT |
3288 | if (in_compat_syscall()) | |
3289 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 3290 | sigsz); |
9e75ad5d AB |
3291 | else |
3292 | #endif | |
b772434b | 3293 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 3294 | |
2b188cc1 JA |
3295 | if (ret) |
3296 | return ret; | |
3297 | } | |
3298 | ||
bda52162 | 3299 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
c826bd7a | 3300 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 JA |
3301 | do { |
3302 | prepare_to_wait_exclusive(&ctx->wait, &iowq.wq, | |
3303 | TASK_INTERRUPTIBLE); | |
1d7bb1d5 | 3304 | if (io_should_wake(&iowq, false)) |
bda52162 JA |
3305 | break; |
3306 | schedule(); | |
3307 | if (signal_pending(current)) { | |
e9ffa5c2 | 3308 | ret = -EINTR; |
bda52162 JA |
3309 | break; |
3310 | } | |
3311 | } while (1); | |
3312 | finish_wait(&ctx->wait, &iowq.wq); | |
3313 | ||
e9ffa5c2 | 3314 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 3315 | |
75b28aff | 3316 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
3317 | } |
3318 | ||
6b06314c JA |
3319 | static void __io_sqe_files_unregister(struct io_ring_ctx *ctx) |
3320 | { | |
3321 | #if defined(CONFIG_UNIX) | |
3322 | if (ctx->ring_sock) { | |
3323 | struct sock *sock = ctx->ring_sock->sk; | |
3324 | struct sk_buff *skb; | |
3325 | ||
3326 | while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) | |
3327 | kfree_skb(skb); | |
3328 | } | |
3329 | #else | |
3330 | int i; | |
3331 | ||
65e19f54 JA |
3332 | for (i = 0; i < ctx->nr_user_files; i++) { |
3333 | struct file *file; | |
3334 | ||
3335 | file = io_file_from_index(ctx, i); | |
3336 | if (file) | |
3337 | fput(file); | |
3338 | } | |
6b06314c JA |
3339 | #endif |
3340 | } | |
3341 | ||
3342 | static int io_sqe_files_unregister(struct io_ring_ctx *ctx) | |
3343 | { | |
65e19f54 JA |
3344 | unsigned nr_tables, i; |
3345 | ||
3346 | if (!ctx->file_table) | |
6b06314c JA |
3347 | return -ENXIO; |
3348 | ||
3349 | __io_sqe_files_unregister(ctx); | |
65e19f54 JA |
3350 | nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE); |
3351 | for (i = 0; i < nr_tables; i++) | |
3352 | kfree(ctx->file_table[i].files); | |
3353 | kfree(ctx->file_table); | |
3354 | ctx->file_table = NULL; | |
6b06314c JA |
3355 | ctx->nr_user_files = 0; |
3356 | return 0; | |
3357 | } | |
3358 | ||
6c271ce2 JA |
3359 | static void io_sq_thread_stop(struct io_ring_ctx *ctx) |
3360 | { | |
3361 | if (ctx->sqo_thread) { | |
206aefde | 3362 | wait_for_completion(&ctx->completions[1]); |
2bbcd6d3 RP |
3363 | /* |
3364 | * The park is a bit of a work-around, without it we get | |
3365 | * warning spews on shutdown with SQPOLL set and affinity | |
3366 | * set to a single CPU. | |
3367 | */ | |
06058632 | 3368 | kthread_park(ctx->sqo_thread); |
6c271ce2 JA |
3369 | kthread_stop(ctx->sqo_thread); |
3370 | ctx->sqo_thread = NULL; | |
3371 | } | |
3372 | } | |
3373 | ||
6b06314c JA |
3374 | static void io_finish_async(struct io_ring_ctx *ctx) |
3375 | { | |
6c271ce2 JA |
3376 | io_sq_thread_stop(ctx); |
3377 | ||
561fb04a JA |
3378 | if (ctx->io_wq) { |
3379 | io_wq_destroy(ctx->io_wq); | |
3380 | ctx->io_wq = NULL; | |
6b06314c JA |
3381 | } |
3382 | } | |
3383 | ||
3384 | #if defined(CONFIG_UNIX) | |
3385 | static void io_destruct_skb(struct sk_buff *skb) | |
3386 | { | |
3387 | struct io_ring_ctx *ctx = skb->sk->sk_user_data; | |
8a997340 | 3388 | |
561fb04a JA |
3389 | if (ctx->io_wq) |
3390 | io_wq_flush(ctx->io_wq); | |
6b06314c | 3391 | |
6b06314c JA |
3392 | unix_destruct_scm(skb); |
3393 | } | |
3394 | ||
3395 | /* | |
3396 | * Ensure the UNIX gc is aware of our file set, so we are certain that | |
3397 | * the io_uring can be safely unregistered on process exit, even if we have | |
3398 | * loops in the file referencing. | |
3399 | */ | |
3400 | static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset) | |
3401 | { | |
3402 | struct sock *sk = ctx->ring_sock->sk; | |
3403 | struct scm_fp_list *fpl; | |
3404 | struct sk_buff *skb; | |
08a45173 | 3405 | int i, nr_files; |
6b06314c JA |
3406 | |
3407 | if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) { | |
3408 | unsigned long inflight = ctx->user->unix_inflight + nr; | |
3409 | ||
3410 | if (inflight > task_rlimit(current, RLIMIT_NOFILE)) | |
3411 | return -EMFILE; | |
3412 | } | |
3413 | ||
3414 | fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); | |
3415 | if (!fpl) | |
3416 | return -ENOMEM; | |
3417 | ||
3418 | skb = alloc_skb(0, GFP_KERNEL); | |
3419 | if (!skb) { | |
3420 | kfree(fpl); | |
3421 | return -ENOMEM; | |
3422 | } | |
3423 | ||
3424 | skb->sk = sk; | |
6b06314c | 3425 | |
08a45173 | 3426 | nr_files = 0; |
6b06314c JA |
3427 | fpl->user = get_uid(ctx->user); |
3428 | for (i = 0; i < nr; i++) { | |
65e19f54 JA |
3429 | struct file *file = io_file_from_index(ctx, i + offset); |
3430 | ||
3431 | if (!file) | |
08a45173 | 3432 | continue; |
65e19f54 | 3433 | fpl->fp[nr_files] = get_file(file); |
08a45173 JA |
3434 | unix_inflight(fpl->user, fpl->fp[nr_files]); |
3435 | nr_files++; | |
6b06314c JA |
3436 | } |
3437 | ||
08a45173 JA |
3438 | if (nr_files) { |
3439 | fpl->max = SCM_MAX_FD; | |
3440 | fpl->count = nr_files; | |
3441 | UNIXCB(skb).fp = fpl; | |
3442 | skb->destructor = io_destruct_skb; | |
3443 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); | |
3444 | skb_queue_head(&sk->sk_receive_queue, skb); | |
6b06314c | 3445 | |
08a45173 JA |
3446 | for (i = 0; i < nr_files; i++) |
3447 | fput(fpl->fp[i]); | |
3448 | } else { | |
3449 | kfree_skb(skb); | |
3450 | kfree(fpl); | |
3451 | } | |
6b06314c JA |
3452 | |
3453 | return 0; | |
3454 | } | |
3455 | ||
3456 | /* | |
3457 | * If UNIX sockets are enabled, fd passing can cause a reference cycle which | |
3458 | * causes regular reference counting to break down. We rely on the UNIX | |
3459 | * garbage collection to take care of this problem for us. | |
3460 | */ | |
3461 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
3462 | { | |
3463 | unsigned left, total; | |
3464 | int ret = 0; | |
3465 | ||
3466 | total = 0; | |
3467 | left = ctx->nr_user_files; | |
3468 | while (left) { | |
3469 | unsigned this_files = min_t(unsigned, left, SCM_MAX_FD); | |
6b06314c JA |
3470 | |
3471 | ret = __io_sqe_files_scm(ctx, this_files, total); | |
3472 | if (ret) | |
3473 | break; | |
3474 | left -= this_files; | |
3475 | total += this_files; | |
3476 | } | |
3477 | ||
3478 | if (!ret) | |
3479 | return 0; | |
3480 | ||
3481 | while (total < ctx->nr_user_files) { | |
65e19f54 JA |
3482 | struct file *file = io_file_from_index(ctx, total); |
3483 | ||
3484 | if (file) | |
3485 | fput(file); | |
6b06314c JA |
3486 | total++; |
3487 | } | |
3488 | ||
3489 | return ret; | |
3490 | } | |
3491 | #else | |
3492 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
3493 | { | |
3494 | return 0; | |
3495 | } | |
3496 | #endif | |
3497 | ||
65e19f54 JA |
3498 | static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables, |
3499 | unsigned nr_files) | |
3500 | { | |
3501 | int i; | |
3502 | ||
3503 | for (i = 0; i < nr_tables; i++) { | |
3504 | struct fixed_file_table *table = &ctx->file_table[i]; | |
3505 | unsigned this_files; | |
3506 | ||
3507 | this_files = min(nr_files, IORING_MAX_FILES_TABLE); | |
3508 | table->files = kcalloc(this_files, sizeof(struct file *), | |
3509 | GFP_KERNEL); | |
3510 | if (!table->files) | |
3511 | break; | |
3512 | nr_files -= this_files; | |
3513 | } | |
3514 | ||
3515 | if (i == nr_tables) | |
3516 | return 0; | |
3517 | ||
3518 | for (i = 0; i < nr_tables; i++) { | |
3519 | struct fixed_file_table *table = &ctx->file_table[i]; | |
3520 | kfree(table->files); | |
3521 | } | |
3522 | return 1; | |
3523 | } | |
3524 | ||
6b06314c JA |
3525 | static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, |
3526 | unsigned nr_args) | |
3527 | { | |
3528 | __s32 __user *fds = (__s32 __user *) arg; | |
65e19f54 | 3529 | unsigned nr_tables; |
6b06314c JA |
3530 | int fd, ret = 0; |
3531 | unsigned i; | |
3532 | ||
65e19f54 | 3533 | if (ctx->file_table) |
6b06314c JA |
3534 | return -EBUSY; |
3535 | if (!nr_args) | |
3536 | return -EINVAL; | |
3537 | if (nr_args > IORING_MAX_FIXED_FILES) | |
3538 | return -EMFILE; | |
3539 | ||
65e19f54 JA |
3540 | nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE); |
3541 | ctx->file_table = kcalloc(nr_tables, sizeof(struct fixed_file_table), | |
3542 | GFP_KERNEL); | |
3543 | if (!ctx->file_table) | |
6b06314c JA |
3544 | return -ENOMEM; |
3545 | ||
65e19f54 JA |
3546 | if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) { |
3547 | kfree(ctx->file_table); | |
46568e9b | 3548 | ctx->file_table = NULL; |
65e19f54 JA |
3549 | return -ENOMEM; |
3550 | } | |
3551 | ||
08a45173 | 3552 | for (i = 0; i < nr_args; i++, ctx->nr_user_files++) { |
65e19f54 JA |
3553 | struct fixed_file_table *table; |
3554 | unsigned index; | |
3555 | ||
6b06314c JA |
3556 | ret = -EFAULT; |
3557 | if (copy_from_user(&fd, &fds[i], sizeof(fd))) | |
3558 | break; | |
08a45173 JA |
3559 | /* allow sparse sets */ |
3560 | if (fd == -1) { | |
3561 | ret = 0; | |
3562 | continue; | |
3563 | } | |
6b06314c | 3564 | |
65e19f54 JA |
3565 | table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT]; |
3566 | index = i & IORING_FILE_TABLE_MASK; | |
3567 | table->files[index] = fget(fd); | |
6b06314c JA |
3568 | |
3569 | ret = -EBADF; | |
65e19f54 | 3570 | if (!table->files[index]) |
6b06314c JA |
3571 | break; |
3572 | /* | |
3573 | * Don't allow io_uring instances to be registered. If UNIX | |
3574 | * isn't enabled, then this causes a reference cycle and this | |
3575 | * instance can never get freed. If UNIX is enabled we'll | |
3576 | * handle it just fine, but there's still no point in allowing | |
3577 | * a ring fd as it doesn't support regular read/write anyway. | |
3578 | */ | |
65e19f54 JA |
3579 | if (table->files[index]->f_op == &io_uring_fops) { |
3580 | fput(table->files[index]); | |
6b06314c JA |
3581 | break; |
3582 | } | |
6b06314c JA |
3583 | ret = 0; |
3584 | } | |
3585 | ||
3586 | if (ret) { | |
65e19f54 JA |
3587 | for (i = 0; i < ctx->nr_user_files; i++) { |
3588 | struct file *file; | |
6b06314c | 3589 | |
65e19f54 JA |
3590 | file = io_file_from_index(ctx, i); |
3591 | if (file) | |
3592 | fput(file); | |
3593 | } | |
3594 | for (i = 0; i < nr_tables; i++) | |
3595 | kfree(ctx->file_table[i].files); | |
3596 | ||
3597 | kfree(ctx->file_table); | |
3598 | ctx->file_table = NULL; | |
6b06314c JA |
3599 | ctx->nr_user_files = 0; |
3600 | return ret; | |
3601 | } | |
3602 | ||
3603 | ret = io_sqe_files_scm(ctx); | |
3604 | if (ret) | |
3605 | io_sqe_files_unregister(ctx); | |
3606 | ||
3607 | return ret; | |
3608 | } | |
3609 | ||
c3a31e60 JA |
3610 | static void io_sqe_file_unregister(struct io_ring_ctx *ctx, int index) |
3611 | { | |
3612 | #if defined(CONFIG_UNIX) | |
65e19f54 | 3613 | struct file *file = io_file_from_index(ctx, index); |
c3a31e60 JA |
3614 | struct sock *sock = ctx->ring_sock->sk; |
3615 | struct sk_buff_head list, *head = &sock->sk_receive_queue; | |
3616 | struct sk_buff *skb; | |
3617 | int i; | |
3618 | ||
3619 | __skb_queue_head_init(&list); | |
3620 | ||
3621 | /* | |
3622 | * Find the skb that holds this file in its SCM_RIGHTS. When found, | |
3623 | * remove this entry and rearrange the file array. | |
3624 | */ | |
3625 | skb = skb_dequeue(head); | |
3626 | while (skb) { | |
3627 | struct scm_fp_list *fp; | |
3628 | ||
3629 | fp = UNIXCB(skb).fp; | |
3630 | for (i = 0; i < fp->count; i++) { | |
3631 | int left; | |
3632 | ||
3633 | if (fp->fp[i] != file) | |
3634 | continue; | |
3635 | ||
3636 | unix_notinflight(fp->user, fp->fp[i]); | |
3637 | left = fp->count - 1 - i; | |
3638 | if (left) { | |
3639 | memmove(&fp->fp[i], &fp->fp[i + 1], | |
3640 | left * sizeof(struct file *)); | |
3641 | } | |
3642 | fp->count--; | |
3643 | if (!fp->count) { | |
3644 | kfree_skb(skb); | |
3645 | skb = NULL; | |
3646 | } else { | |
3647 | __skb_queue_tail(&list, skb); | |
3648 | } | |
3649 | fput(file); | |
3650 | file = NULL; | |
3651 | break; | |
3652 | } | |
3653 | ||
3654 | if (!file) | |
3655 | break; | |
3656 | ||
3657 | __skb_queue_tail(&list, skb); | |
3658 | ||
3659 | skb = skb_dequeue(head); | |
3660 | } | |
3661 | ||
3662 | if (skb_peek(&list)) { | |
3663 | spin_lock_irq(&head->lock); | |
3664 | while ((skb = __skb_dequeue(&list)) != NULL) | |
3665 | __skb_queue_tail(head, skb); | |
3666 | spin_unlock_irq(&head->lock); | |
3667 | } | |
3668 | #else | |
65e19f54 | 3669 | fput(io_file_from_index(ctx, index)); |
c3a31e60 JA |
3670 | #endif |
3671 | } | |
3672 | ||
3673 | static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file, | |
3674 | int index) | |
3675 | { | |
3676 | #if defined(CONFIG_UNIX) | |
3677 | struct sock *sock = ctx->ring_sock->sk; | |
3678 | struct sk_buff_head *head = &sock->sk_receive_queue; | |
3679 | struct sk_buff *skb; | |
3680 | ||
3681 | /* | |
3682 | * See if we can merge this file into an existing skb SCM_RIGHTS | |
3683 | * file set. If there's no room, fall back to allocating a new skb | |
3684 | * and filling it in. | |
3685 | */ | |
3686 | spin_lock_irq(&head->lock); | |
3687 | skb = skb_peek(head); | |
3688 | if (skb) { | |
3689 | struct scm_fp_list *fpl = UNIXCB(skb).fp; | |
3690 | ||
3691 | if (fpl->count < SCM_MAX_FD) { | |
3692 | __skb_unlink(skb, head); | |
3693 | spin_unlock_irq(&head->lock); | |
3694 | fpl->fp[fpl->count] = get_file(file); | |
3695 | unix_inflight(fpl->user, fpl->fp[fpl->count]); | |
3696 | fpl->count++; | |
3697 | spin_lock_irq(&head->lock); | |
3698 | __skb_queue_head(head, skb); | |
3699 | } else { | |
3700 | skb = NULL; | |
3701 | } | |
3702 | } | |
3703 | spin_unlock_irq(&head->lock); | |
3704 | ||
3705 | if (skb) { | |
3706 | fput(file); | |
3707 | return 0; | |
3708 | } | |
3709 | ||
3710 | return __io_sqe_files_scm(ctx, 1, index); | |
3711 | #else | |
3712 | return 0; | |
3713 | #endif | |
3714 | } | |
3715 | ||
3716 | static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg, | |
3717 | unsigned nr_args) | |
3718 | { | |
3719 | struct io_uring_files_update up; | |
3720 | __s32 __user *fds; | |
3721 | int fd, i, err; | |
3722 | __u32 done; | |
3723 | ||
65e19f54 | 3724 | if (!ctx->file_table) |
c3a31e60 JA |
3725 | return -ENXIO; |
3726 | if (!nr_args) | |
3727 | return -EINVAL; | |
3728 | if (copy_from_user(&up, arg, sizeof(up))) | |
3729 | return -EFAULT; | |
3730 | if (check_add_overflow(up.offset, nr_args, &done)) | |
3731 | return -EOVERFLOW; | |
3732 | if (done > ctx->nr_user_files) | |
3733 | return -EINVAL; | |
3734 | ||
3735 | done = 0; | |
3736 | fds = (__s32 __user *) up.fds; | |
3737 | while (nr_args) { | |
65e19f54 JA |
3738 | struct fixed_file_table *table; |
3739 | unsigned index; | |
3740 | ||
c3a31e60 JA |
3741 | err = 0; |
3742 | if (copy_from_user(&fd, &fds[done], sizeof(fd))) { | |
3743 | err = -EFAULT; | |
3744 | break; | |
3745 | } | |
3746 | i = array_index_nospec(up.offset, ctx->nr_user_files); | |
65e19f54 JA |
3747 | table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT]; |
3748 | index = i & IORING_FILE_TABLE_MASK; | |
3749 | if (table->files[index]) { | |
c3a31e60 | 3750 | io_sqe_file_unregister(ctx, i); |
65e19f54 | 3751 | table->files[index] = NULL; |
c3a31e60 JA |
3752 | } |
3753 | if (fd != -1) { | |
3754 | struct file *file; | |
3755 | ||
3756 | file = fget(fd); | |
3757 | if (!file) { | |
3758 | err = -EBADF; | |
3759 | break; | |
3760 | } | |
3761 | /* | |
3762 | * Don't allow io_uring instances to be registered. If | |
3763 | * UNIX isn't enabled, then this causes a reference | |
3764 | * cycle and this instance can never get freed. If UNIX | |
3765 | * is enabled we'll handle it just fine, but there's | |
3766 | * still no point in allowing a ring fd as it doesn't | |
3767 | * support regular read/write anyway. | |
3768 | */ | |
3769 | if (file->f_op == &io_uring_fops) { | |
3770 | fput(file); | |
3771 | err = -EBADF; | |
3772 | break; | |
3773 | } | |
65e19f54 | 3774 | table->files[index] = file; |
c3a31e60 JA |
3775 | err = io_sqe_file_register(ctx, file, i); |
3776 | if (err) | |
3777 | break; | |
3778 | } | |
3779 | nr_args--; | |
3780 | done++; | |
3781 | up.offset++; | |
3782 | } | |
3783 | ||
3784 | return done ? done : err; | |
3785 | } | |
3786 | ||
6c271ce2 JA |
3787 | static int io_sq_offload_start(struct io_ring_ctx *ctx, |
3788 | struct io_uring_params *p) | |
2b188cc1 | 3789 | { |
561fb04a | 3790 | unsigned concurrency; |
2b188cc1 JA |
3791 | int ret; |
3792 | ||
6c271ce2 | 3793 | init_waitqueue_head(&ctx->sqo_wait); |
2b188cc1 JA |
3794 | mmgrab(current->mm); |
3795 | ctx->sqo_mm = current->mm; | |
3796 | ||
6c271ce2 | 3797 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
3ec482d1 JA |
3798 | ret = -EPERM; |
3799 | if (!capable(CAP_SYS_ADMIN)) | |
3800 | goto err; | |
3801 | ||
917257da JA |
3802 | ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle); |
3803 | if (!ctx->sq_thread_idle) | |
3804 | ctx->sq_thread_idle = HZ; | |
3805 | ||
6c271ce2 | 3806 | if (p->flags & IORING_SETUP_SQ_AFF) { |
44a9bd18 | 3807 | int cpu = p->sq_thread_cpu; |
6c271ce2 | 3808 | |
917257da | 3809 | ret = -EINVAL; |
44a9bd18 JA |
3810 | if (cpu >= nr_cpu_ids) |
3811 | goto err; | |
7889f44d | 3812 | if (!cpu_online(cpu)) |
917257da JA |
3813 | goto err; |
3814 | ||
6c271ce2 JA |
3815 | ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread, |
3816 | ctx, cpu, | |
3817 | "io_uring-sq"); | |
3818 | } else { | |
3819 | ctx->sqo_thread = kthread_create(io_sq_thread, ctx, | |
3820 | "io_uring-sq"); | |
3821 | } | |
3822 | if (IS_ERR(ctx->sqo_thread)) { | |
3823 | ret = PTR_ERR(ctx->sqo_thread); | |
3824 | ctx->sqo_thread = NULL; | |
3825 | goto err; | |
3826 | } | |
3827 | wake_up_process(ctx->sqo_thread); | |
3828 | } else if (p->flags & IORING_SETUP_SQ_AFF) { | |
3829 | /* Can't have SQ_AFF without SQPOLL */ | |
3830 | ret = -EINVAL; | |
3831 | goto err; | |
3832 | } | |
3833 | ||
561fb04a JA |
3834 | /* Do QD, or 4 * CPUS, whatever is smallest */ |
3835 | concurrency = min(ctx->sq_entries, 4 * num_online_cpus()); | |
5f8fd2d3 | 3836 | ctx->io_wq = io_wq_create(concurrency, ctx->sqo_mm, ctx->user); |
975c99a5 JA |
3837 | if (IS_ERR(ctx->io_wq)) { |
3838 | ret = PTR_ERR(ctx->io_wq); | |
3839 | ctx->io_wq = NULL; | |
2b188cc1 JA |
3840 | goto err; |
3841 | } | |
3842 | ||
3843 | return 0; | |
3844 | err: | |
54a91f3b | 3845 | io_finish_async(ctx); |
2b188cc1 JA |
3846 | mmdrop(ctx->sqo_mm); |
3847 | ctx->sqo_mm = NULL; | |
3848 | return ret; | |
3849 | } | |
3850 | ||
3851 | static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages) | |
3852 | { | |
3853 | atomic_long_sub(nr_pages, &user->locked_vm); | |
3854 | } | |
3855 | ||
3856 | static int io_account_mem(struct user_struct *user, unsigned long nr_pages) | |
3857 | { | |
3858 | unsigned long page_limit, cur_pages, new_pages; | |
3859 | ||
3860 | /* Don't allow more pages than we can safely lock */ | |
3861 | page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
3862 | ||
3863 | do { | |
3864 | cur_pages = atomic_long_read(&user->locked_vm); | |
3865 | new_pages = cur_pages + nr_pages; | |
3866 | if (new_pages > page_limit) | |
3867 | return -ENOMEM; | |
3868 | } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages, | |
3869 | new_pages) != cur_pages); | |
3870 | ||
3871 | return 0; | |
3872 | } | |
3873 | ||
3874 | static void io_mem_free(void *ptr) | |
3875 | { | |
52e04ef4 MR |
3876 | struct page *page; |
3877 | ||
3878 | if (!ptr) | |
3879 | return; | |
2b188cc1 | 3880 | |
52e04ef4 | 3881 | page = virt_to_head_page(ptr); |
2b188cc1 JA |
3882 | if (put_page_testzero(page)) |
3883 | free_compound_page(page); | |
3884 | } | |
3885 | ||
3886 | static void *io_mem_alloc(size_t size) | |
3887 | { | |
3888 | gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP | | |
3889 | __GFP_NORETRY; | |
3890 | ||
3891 | return (void *) __get_free_pages(gfp_flags, get_order(size)); | |
3892 | } | |
3893 | ||
75b28aff HV |
3894 | static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries, |
3895 | size_t *sq_offset) | |
3896 | { | |
3897 | struct io_rings *rings; | |
3898 | size_t off, sq_array_size; | |
3899 | ||
3900 | off = struct_size(rings, cqes, cq_entries); | |
3901 | if (off == SIZE_MAX) | |
3902 | return SIZE_MAX; | |
3903 | ||
3904 | #ifdef CONFIG_SMP | |
3905 | off = ALIGN(off, SMP_CACHE_BYTES); | |
3906 | if (off == 0) | |
3907 | return SIZE_MAX; | |
3908 | #endif | |
3909 | ||
3910 | sq_array_size = array_size(sizeof(u32), sq_entries); | |
3911 | if (sq_array_size == SIZE_MAX) | |
3912 | return SIZE_MAX; | |
3913 | ||
3914 | if (check_add_overflow(off, sq_array_size, &off)) | |
3915 | return SIZE_MAX; | |
3916 | ||
3917 | if (sq_offset) | |
3918 | *sq_offset = off; | |
3919 | ||
3920 | return off; | |
3921 | } | |
3922 | ||
2b188cc1 JA |
3923 | static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries) |
3924 | { | |
75b28aff | 3925 | size_t pages; |
2b188cc1 | 3926 | |
75b28aff HV |
3927 | pages = (size_t)1 << get_order( |
3928 | rings_size(sq_entries, cq_entries, NULL)); | |
3929 | pages += (size_t)1 << get_order( | |
3930 | array_size(sizeof(struct io_uring_sqe), sq_entries)); | |
2b188cc1 | 3931 | |
75b28aff | 3932 | return pages; |
2b188cc1 JA |
3933 | } |
3934 | ||
edafccee JA |
3935 | static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx) |
3936 | { | |
3937 | int i, j; | |
3938 | ||
3939 | if (!ctx->user_bufs) | |
3940 | return -ENXIO; | |
3941 | ||
3942 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
3943 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
3944 | ||
3945 | for (j = 0; j < imu->nr_bvecs; j++) | |
27c4d3a3 | 3946 | put_user_page(imu->bvec[j].bv_page); |
edafccee JA |
3947 | |
3948 | if (ctx->account_mem) | |
3949 | io_unaccount_mem(ctx->user, imu->nr_bvecs); | |
d4ef6475 | 3950 | kvfree(imu->bvec); |
edafccee JA |
3951 | imu->nr_bvecs = 0; |
3952 | } | |
3953 | ||
3954 | kfree(ctx->user_bufs); | |
3955 | ctx->user_bufs = NULL; | |
3956 | ctx->nr_user_bufs = 0; | |
3957 | return 0; | |
3958 | } | |
3959 | ||
3960 | static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, | |
3961 | void __user *arg, unsigned index) | |
3962 | { | |
3963 | struct iovec __user *src; | |
3964 | ||
3965 | #ifdef CONFIG_COMPAT | |
3966 | if (ctx->compat) { | |
3967 | struct compat_iovec __user *ciovs; | |
3968 | struct compat_iovec ciov; | |
3969 | ||
3970 | ciovs = (struct compat_iovec __user *) arg; | |
3971 | if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) | |
3972 | return -EFAULT; | |
3973 | ||
3974 | dst->iov_base = (void __user *) (unsigned long) ciov.iov_base; | |
3975 | dst->iov_len = ciov.iov_len; | |
3976 | return 0; | |
3977 | } | |
3978 | #endif | |
3979 | src = (struct iovec __user *) arg; | |
3980 | if (copy_from_user(dst, &src[index], sizeof(*dst))) | |
3981 | return -EFAULT; | |
3982 | return 0; | |
3983 | } | |
3984 | ||
3985 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg, | |
3986 | unsigned nr_args) | |
3987 | { | |
3988 | struct vm_area_struct **vmas = NULL; | |
3989 | struct page **pages = NULL; | |
3990 | int i, j, got_pages = 0; | |
3991 | int ret = -EINVAL; | |
3992 | ||
3993 | if (ctx->user_bufs) | |
3994 | return -EBUSY; | |
3995 | if (!nr_args || nr_args > UIO_MAXIOV) | |
3996 | return -EINVAL; | |
3997 | ||
3998 | ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf), | |
3999 | GFP_KERNEL); | |
4000 | if (!ctx->user_bufs) | |
4001 | return -ENOMEM; | |
4002 | ||
4003 | for (i = 0; i < nr_args; i++) { | |
4004 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
4005 | unsigned long off, start, end, ubuf; | |
4006 | int pret, nr_pages; | |
4007 | struct iovec iov; | |
4008 | size_t size; | |
4009 | ||
4010 | ret = io_copy_iov(ctx, &iov, arg, i); | |
4011 | if (ret) | |
a278682d | 4012 | goto err; |
edafccee JA |
4013 | |
4014 | /* | |
4015 | * Don't impose further limits on the size and buffer | |
4016 | * constraints here, we'll -EINVAL later when IO is | |
4017 | * submitted if they are wrong. | |
4018 | */ | |
4019 | ret = -EFAULT; | |
4020 | if (!iov.iov_base || !iov.iov_len) | |
4021 | goto err; | |
4022 | ||
4023 | /* arbitrary limit, but we need something */ | |
4024 | if (iov.iov_len > SZ_1G) | |
4025 | goto err; | |
4026 | ||
4027 | ubuf = (unsigned long) iov.iov_base; | |
4028 | end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
4029 | start = ubuf >> PAGE_SHIFT; | |
4030 | nr_pages = end - start; | |
4031 | ||
4032 | if (ctx->account_mem) { | |
4033 | ret = io_account_mem(ctx->user, nr_pages); | |
4034 | if (ret) | |
4035 | goto err; | |
4036 | } | |
4037 | ||
4038 | ret = 0; | |
4039 | if (!pages || nr_pages > got_pages) { | |
4040 | kfree(vmas); | |
4041 | kfree(pages); | |
d4ef6475 | 4042 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), |
edafccee | 4043 | GFP_KERNEL); |
d4ef6475 | 4044 | vmas = kvmalloc_array(nr_pages, |
edafccee JA |
4045 | sizeof(struct vm_area_struct *), |
4046 | GFP_KERNEL); | |
4047 | if (!pages || !vmas) { | |
4048 | ret = -ENOMEM; | |
4049 | if (ctx->account_mem) | |
4050 | io_unaccount_mem(ctx->user, nr_pages); | |
4051 | goto err; | |
4052 | } | |
4053 | got_pages = nr_pages; | |
4054 | } | |
4055 | ||
d4ef6475 | 4056 | imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec), |
edafccee JA |
4057 | GFP_KERNEL); |
4058 | ret = -ENOMEM; | |
4059 | if (!imu->bvec) { | |
4060 | if (ctx->account_mem) | |
4061 | io_unaccount_mem(ctx->user, nr_pages); | |
4062 | goto err; | |
4063 | } | |
4064 | ||
4065 | ret = 0; | |
4066 | down_read(¤t->mm->mmap_sem); | |
932f4a63 IW |
4067 | pret = get_user_pages(ubuf, nr_pages, |
4068 | FOLL_WRITE | FOLL_LONGTERM, | |
4069 | pages, vmas); | |
edafccee JA |
4070 | if (pret == nr_pages) { |
4071 | /* don't support file backed memory */ | |
4072 | for (j = 0; j < nr_pages; j++) { | |
4073 | struct vm_area_struct *vma = vmas[j]; | |
4074 | ||
4075 | if (vma->vm_file && | |
4076 | !is_file_hugepages(vma->vm_file)) { | |
4077 | ret = -EOPNOTSUPP; | |
4078 | break; | |
4079 | } | |
4080 | } | |
4081 | } else { | |
4082 | ret = pret < 0 ? pret : -EFAULT; | |
4083 | } | |
4084 | up_read(¤t->mm->mmap_sem); | |
4085 | if (ret) { | |
4086 | /* | |
4087 | * if we did partial map, or found file backed vmas, | |
4088 | * release any pages we did get | |
4089 | */ | |
27c4d3a3 JH |
4090 | if (pret > 0) |
4091 | put_user_pages(pages, pret); | |
edafccee JA |
4092 | if (ctx->account_mem) |
4093 | io_unaccount_mem(ctx->user, nr_pages); | |
d4ef6475 | 4094 | kvfree(imu->bvec); |
edafccee JA |
4095 | goto err; |
4096 | } | |
4097 | ||
4098 | off = ubuf & ~PAGE_MASK; | |
4099 | size = iov.iov_len; | |
4100 | for (j = 0; j < nr_pages; j++) { | |
4101 | size_t vec_len; | |
4102 | ||
4103 | vec_len = min_t(size_t, size, PAGE_SIZE - off); | |
4104 | imu->bvec[j].bv_page = pages[j]; | |
4105 | imu->bvec[j].bv_len = vec_len; | |
4106 | imu->bvec[j].bv_offset = off; | |
4107 | off = 0; | |
4108 | size -= vec_len; | |
4109 | } | |
4110 | /* store original address for later verification */ | |
4111 | imu->ubuf = ubuf; | |
4112 | imu->len = iov.iov_len; | |
4113 | imu->nr_bvecs = nr_pages; | |
4114 | ||
4115 | ctx->nr_user_bufs++; | |
4116 | } | |
d4ef6475 MR |
4117 | kvfree(pages); |
4118 | kvfree(vmas); | |
edafccee JA |
4119 | return 0; |
4120 | err: | |
d4ef6475 MR |
4121 | kvfree(pages); |
4122 | kvfree(vmas); | |
edafccee JA |
4123 | io_sqe_buffer_unregister(ctx); |
4124 | return ret; | |
4125 | } | |
4126 | ||
9b402849 JA |
4127 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg) |
4128 | { | |
4129 | __s32 __user *fds = arg; | |
4130 | int fd; | |
4131 | ||
4132 | if (ctx->cq_ev_fd) | |
4133 | return -EBUSY; | |
4134 | ||
4135 | if (copy_from_user(&fd, fds, sizeof(*fds))) | |
4136 | return -EFAULT; | |
4137 | ||
4138 | ctx->cq_ev_fd = eventfd_ctx_fdget(fd); | |
4139 | if (IS_ERR(ctx->cq_ev_fd)) { | |
4140 | int ret = PTR_ERR(ctx->cq_ev_fd); | |
4141 | ctx->cq_ev_fd = NULL; | |
4142 | return ret; | |
4143 | } | |
4144 | ||
4145 | return 0; | |
4146 | } | |
4147 | ||
4148 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) | |
4149 | { | |
4150 | if (ctx->cq_ev_fd) { | |
4151 | eventfd_ctx_put(ctx->cq_ev_fd); | |
4152 | ctx->cq_ev_fd = NULL; | |
4153 | return 0; | |
4154 | } | |
4155 | ||
4156 | return -ENXIO; | |
4157 | } | |
4158 | ||
2b188cc1 JA |
4159 | static void io_ring_ctx_free(struct io_ring_ctx *ctx) |
4160 | { | |
6b06314c | 4161 | io_finish_async(ctx); |
2b188cc1 JA |
4162 | if (ctx->sqo_mm) |
4163 | mmdrop(ctx->sqo_mm); | |
def596e9 JA |
4164 | |
4165 | io_iopoll_reap_events(ctx); | |
edafccee | 4166 | io_sqe_buffer_unregister(ctx); |
6b06314c | 4167 | io_sqe_files_unregister(ctx); |
9b402849 | 4168 | io_eventfd_unregister(ctx); |
def596e9 | 4169 | |
2b188cc1 | 4170 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
4171 | if (ctx->ring_sock) { |
4172 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 4173 | sock_release(ctx->ring_sock); |
355e8d26 | 4174 | } |
2b188cc1 JA |
4175 | #endif |
4176 | ||
75b28aff | 4177 | io_mem_free(ctx->rings); |
2b188cc1 | 4178 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
4179 | |
4180 | percpu_ref_exit(&ctx->refs); | |
4181 | if (ctx->account_mem) | |
4182 | io_unaccount_mem(ctx->user, | |
4183 | ring_pages(ctx->sq_entries, ctx->cq_entries)); | |
4184 | free_uid(ctx->user); | |
206aefde | 4185 | kfree(ctx->completions); |
0ddf92e8 | 4186 | kmem_cache_free(req_cachep, ctx->fallback_req); |
2b188cc1 JA |
4187 | kfree(ctx); |
4188 | } | |
4189 | ||
4190 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
4191 | { | |
4192 | struct io_ring_ctx *ctx = file->private_data; | |
4193 | __poll_t mask = 0; | |
4194 | ||
4195 | poll_wait(file, &ctx->cq_wait, wait); | |
4f7067c3 SB |
4196 | /* |
4197 | * synchronizes with barrier from wq_has_sleeper call in | |
4198 | * io_commit_cqring | |
4199 | */ | |
2b188cc1 | 4200 | smp_rmb(); |
75b28aff HV |
4201 | if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head != |
4202 | ctx->rings->sq_ring_entries) | |
2b188cc1 | 4203 | mask |= EPOLLOUT | EPOLLWRNORM; |
daa5de54 | 4204 | if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail) |
2b188cc1 JA |
4205 | mask |= EPOLLIN | EPOLLRDNORM; |
4206 | ||
4207 | return mask; | |
4208 | } | |
4209 | ||
4210 | static int io_uring_fasync(int fd, struct file *file, int on) | |
4211 | { | |
4212 | struct io_ring_ctx *ctx = file->private_data; | |
4213 | ||
4214 | return fasync_helper(fd, file, on, &ctx->cq_fasync); | |
4215 | } | |
4216 | ||
4217 | static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) | |
4218 | { | |
4219 | mutex_lock(&ctx->uring_lock); | |
4220 | percpu_ref_kill(&ctx->refs); | |
4221 | mutex_unlock(&ctx->uring_lock); | |
4222 | ||
5262f567 | 4223 | io_kill_timeouts(ctx); |
221c5eb2 | 4224 | io_poll_remove_all(ctx); |
561fb04a JA |
4225 | |
4226 | if (ctx->io_wq) | |
4227 | io_wq_cancel_all(ctx->io_wq); | |
4228 | ||
def596e9 | 4229 | io_iopoll_reap_events(ctx); |
1d7bb1d5 | 4230 | io_cqring_overflow_flush(ctx, true); |
206aefde | 4231 | wait_for_completion(&ctx->completions[0]); |
2b188cc1 JA |
4232 | io_ring_ctx_free(ctx); |
4233 | } | |
4234 | ||
4235 | static int io_uring_release(struct inode *inode, struct file *file) | |
4236 | { | |
4237 | struct io_ring_ctx *ctx = file->private_data; | |
4238 | ||
4239 | file->private_data = NULL; | |
4240 | io_ring_ctx_wait_and_kill(ctx); | |
4241 | return 0; | |
4242 | } | |
4243 | ||
fcb323cc JA |
4244 | static void io_uring_cancel_files(struct io_ring_ctx *ctx, |
4245 | struct files_struct *files) | |
4246 | { | |
4247 | struct io_kiocb *req; | |
4248 | DEFINE_WAIT(wait); | |
4249 | ||
4250 | while (!list_empty_careful(&ctx->inflight_list)) { | |
4251 | enum io_wq_cancel ret = IO_WQ_CANCEL_NOTFOUND; | |
4252 | ||
4253 | spin_lock_irq(&ctx->inflight_lock); | |
4254 | list_for_each_entry(req, &ctx->inflight_list, inflight_entry) { | |
4255 | if (req->work.files == files) { | |
4256 | ret = io_wq_cancel_work(ctx->io_wq, &req->work); | |
4257 | break; | |
4258 | } | |
4259 | } | |
4260 | if (ret == IO_WQ_CANCEL_RUNNING) | |
4261 | prepare_to_wait(&ctx->inflight_wait, &wait, | |
4262 | TASK_UNINTERRUPTIBLE); | |
4263 | ||
4264 | spin_unlock_irq(&ctx->inflight_lock); | |
4265 | ||
4266 | /* | |
4267 | * We need to keep going until we get NOTFOUND. We only cancel | |
4268 | * one work at the time. | |
4269 | * | |
4270 | * If we get CANCEL_RUNNING, then wait for a work to complete | |
4271 | * before continuing. | |
4272 | */ | |
4273 | if (ret == IO_WQ_CANCEL_OK) | |
4274 | continue; | |
4275 | else if (ret != IO_WQ_CANCEL_RUNNING) | |
4276 | break; | |
4277 | schedule(); | |
4278 | } | |
4279 | } | |
4280 | ||
4281 | static int io_uring_flush(struct file *file, void *data) | |
4282 | { | |
4283 | struct io_ring_ctx *ctx = file->private_data; | |
4284 | ||
4285 | io_uring_cancel_files(ctx, data); | |
1d7bb1d5 JA |
4286 | if (fatal_signal_pending(current) || (current->flags & PF_EXITING)) { |
4287 | io_cqring_overflow_flush(ctx, true); | |
fcb323cc | 4288 | io_wq_cancel_all(ctx->io_wq); |
1d7bb1d5 | 4289 | } |
fcb323cc JA |
4290 | return 0; |
4291 | } | |
4292 | ||
2b188cc1 JA |
4293 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) |
4294 | { | |
4295 | loff_t offset = (loff_t) vma->vm_pgoff << PAGE_SHIFT; | |
4296 | unsigned long sz = vma->vm_end - vma->vm_start; | |
4297 | struct io_ring_ctx *ctx = file->private_data; | |
4298 | unsigned long pfn; | |
4299 | struct page *page; | |
4300 | void *ptr; | |
4301 | ||
4302 | switch (offset) { | |
4303 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
4304 | case IORING_OFF_CQ_RING: |
4305 | ptr = ctx->rings; | |
2b188cc1 JA |
4306 | break; |
4307 | case IORING_OFF_SQES: | |
4308 | ptr = ctx->sq_sqes; | |
4309 | break; | |
2b188cc1 JA |
4310 | default: |
4311 | return -EINVAL; | |
4312 | } | |
4313 | ||
4314 | page = virt_to_head_page(ptr); | |
a50b854e | 4315 | if (sz > page_size(page)) |
2b188cc1 JA |
4316 | return -EINVAL; |
4317 | ||
4318 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
4319 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
4320 | } | |
4321 | ||
4322 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, | |
4323 | u32, min_complete, u32, flags, const sigset_t __user *, sig, | |
4324 | size_t, sigsz) | |
4325 | { | |
4326 | struct io_ring_ctx *ctx; | |
4327 | long ret = -EBADF; | |
4328 | int submitted = 0; | |
4329 | struct fd f; | |
4330 | ||
6c271ce2 | 4331 | if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP)) |
2b188cc1 JA |
4332 | return -EINVAL; |
4333 | ||
4334 | f = fdget(fd); | |
4335 | if (!f.file) | |
4336 | return -EBADF; | |
4337 | ||
4338 | ret = -EOPNOTSUPP; | |
4339 | if (f.file->f_op != &io_uring_fops) | |
4340 | goto out_fput; | |
4341 | ||
4342 | ret = -ENXIO; | |
4343 | ctx = f.file->private_data; | |
4344 | if (!percpu_ref_tryget(&ctx->refs)) | |
4345 | goto out_fput; | |
4346 | ||
6c271ce2 JA |
4347 | /* |
4348 | * For SQ polling, the thread will do all submissions and completions. | |
4349 | * Just return the requested submit count, and wake the thread if | |
4350 | * we were asked to. | |
4351 | */ | |
b2a9eada | 4352 | ret = 0; |
6c271ce2 JA |
4353 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
4354 | if (flags & IORING_ENTER_SQ_WAKEUP) | |
4355 | wake_up(&ctx->sqo_wait); | |
4356 | submitted = to_submit; | |
b2a9eada | 4357 | } else if (to_submit) { |
ae9428ca | 4358 | struct mm_struct *cur_mm; |
2b188cc1 | 4359 | |
ae9428ca | 4360 | to_submit = min(to_submit, ctx->sq_entries); |
2b188cc1 | 4361 | mutex_lock(&ctx->uring_lock); |
ae9428ca PB |
4362 | /* already have mm, so io_submit_sqes() won't try to grab it */ |
4363 | cur_mm = ctx->sqo_mm; | |
4364 | submitted = io_submit_sqes(ctx, to_submit, f.file, fd, | |
4365 | &cur_mm, false); | |
2b188cc1 | 4366 | mutex_unlock(&ctx->uring_lock); |
2b188cc1 JA |
4367 | } |
4368 | if (flags & IORING_ENTER_GETEVENTS) { | |
def596e9 JA |
4369 | unsigned nr_events = 0; |
4370 | ||
2b188cc1 JA |
4371 | min_complete = min(min_complete, ctx->cq_entries); |
4372 | ||
def596e9 | 4373 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 | 4374 | ret = io_iopoll_check(ctx, &nr_events, min_complete); |
def596e9 JA |
4375 | } else { |
4376 | ret = io_cqring_wait(ctx, min_complete, sig, sigsz); | |
4377 | } | |
2b188cc1 JA |
4378 | } |
4379 | ||
6805b32e | 4380 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
4381 | out_fput: |
4382 | fdput(f); | |
4383 | return submitted ? submitted : ret; | |
4384 | } | |
4385 | ||
4386 | static const struct file_operations io_uring_fops = { | |
4387 | .release = io_uring_release, | |
fcb323cc | 4388 | .flush = io_uring_flush, |
2b188cc1 JA |
4389 | .mmap = io_uring_mmap, |
4390 | .poll = io_uring_poll, | |
4391 | .fasync = io_uring_fasync, | |
4392 | }; | |
4393 | ||
4394 | static int io_allocate_scq_urings(struct io_ring_ctx *ctx, | |
4395 | struct io_uring_params *p) | |
4396 | { | |
75b28aff HV |
4397 | struct io_rings *rings; |
4398 | size_t size, sq_array_offset; | |
2b188cc1 | 4399 | |
75b28aff HV |
4400 | size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset); |
4401 | if (size == SIZE_MAX) | |
4402 | return -EOVERFLOW; | |
4403 | ||
4404 | rings = io_mem_alloc(size); | |
4405 | if (!rings) | |
2b188cc1 JA |
4406 | return -ENOMEM; |
4407 | ||
75b28aff HV |
4408 | ctx->rings = rings; |
4409 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
4410 | rings->sq_ring_mask = p->sq_entries - 1; | |
4411 | rings->cq_ring_mask = p->cq_entries - 1; | |
4412 | rings->sq_ring_entries = p->sq_entries; | |
4413 | rings->cq_ring_entries = p->cq_entries; | |
4414 | ctx->sq_mask = rings->sq_ring_mask; | |
4415 | ctx->cq_mask = rings->cq_ring_mask; | |
4416 | ctx->sq_entries = rings->sq_ring_entries; | |
4417 | ctx->cq_entries = rings->cq_ring_entries; | |
2b188cc1 JA |
4418 | |
4419 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
4420 | if (size == SIZE_MAX) | |
4421 | return -EOVERFLOW; | |
4422 | ||
4423 | ctx->sq_sqes = io_mem_alloc(size); | |
52e04ef4 | 4424 | if (!ctx->sq_sqes) |
2b188cc1 | 4425 | return -ENOMEM; |
2b188cc1 | 4426 | |
2b188cc1 JA |
4427 | return 0; |
4428 | } | |
4429 | ||
4430 | /* | |
4431 | * Allocate an anonymous fd, this is what constitutes the application | |
4432 | * visible backing of an io_uring instance. The application mmaps this | |
4433 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
4434 | * we have to tie this fd to a socket for file garbage collection purposes. | |
4435 | */ | |
4436 | static int io_uring_get_fd(struct io_ring_ctx *ctx) | |
4437 | { | |
4438 | struct file *file; | |
4439 | int ret; | |
4440 | ||
4441 | #if defined(CONFIG_UNIX) | |
4442 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, | |
4443 | &ctx->ring_sock); | |
4444 | if (ret) | |
4445 | return ret; | |
4446 | #endif | |
4447 | ||
4448 | ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
4449 | if (ret < 0) | |
4450 | goto err; | |
4451 | ||
4452 | file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx, | |
4453 | O_RDWR | O_CLOEXEC); | |
4454 | if (IS_ERR(file)) { | |
4455 | put_unused_fd(ret); | |
4456 | ret = PTR_ERR(file); | |
4457 | goto err; | |
4458 | } | |
4459 | ||
4460 | #if defined(CONFIG_UNIX) | |
4461 | ctx->ring_sock->file = file; | |
6b06314c | 4462 | ctx->ring_sock->sk->sk_user_data = ctx; |
2b188cc1 JA |
4463 | #endif |
4464 | fd_install(ret, file); | |
4465 | return ret; | |
4466 | err: | |
4467 | #if defined(CONFIG_UNIX) | |
4468 | sock_release(ctx->ring_sock); | |
4469 | ctx->ring_sock = NULL; | |
4470 | #endif | |
4471 | return ret; | |
4472 | } | |
4473 | ||
4474 | static int io_uring_create(unsigned entries, struct io_uring_params *p) | |
4475 | { | |
4476 | struct user_struct *user = NULL; | |
4477 | struct io_ring_ctx *ctx; | |
4478 | bool account_mem; | |
4479 | int ret; | |
4480 | ||
4481 | if (!entries || entries > IORING_MAX_ENTRIES) | |
4482 | return -EINVAL; | |
4483 | ||
4484 | /* | |
4485 | * Use twice as many entries for the CQ ring. It's possible for the | |
4486 | * application to drive a higher depth than the size of the SQ ring, | |
4487 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
4488 | * some flexibility in overcommitting a bit. If the application has |
4489 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
4490 | * of CQ ring entries manually. | |
2b188cc1 JA |
4491 | */ |
4492 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
4493 | if (p->flags & IORING_SETUP_CQSIZE) { |
4494 | /* | |
4495 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
4496 | * to a power-of-two, if it isn't already. We do NOT impose | |
4497 | * any cq vs sq ring sizing. | |
4498 | */ | |
4499 | if (p->cq_entries < p->sq_entries || p->cq_entries > IORING_MAX_CQ_ENTRIES) | |
4500 | return -EINVAL; | |
4501 | p->cq_entries = roundup_pow_of_two(p->cq_entries); | |
4502 | } else { | |
4503 | p->cq_entries = 2 * p->sq_entries; | |
4504 | } | |
2b188cc1 JA |
4505 | |
4506 | user = get_uid(current_user()); | |
4507 | account_mem = !capable(CAP_IPC_LOCK); | |
4508 | ||
4509 | if (account_mem) { | |
4510 | ret = io_account_mem(user, | |
4511 | ring_pages(p->sq_entries, p->cq_entries)); | |
4512 | if (ret) { | |
4513 | free_uid(user); | |
4514 | return ret; | |
4515 | } | |
4516 | } | |
4517 | ||
4518 | ctx = io_ring_ctx_alloc(p); | |
4519 | if (!ctx) { | |
4520 | if (account_mem) | |
4521 | io_unaccount_mem(user, ring_pages(p->sq_entries, | |
4522 | p->cq_entries)); | |
4523 | free_uid(user); | |
4524 | return -ENOMEM; | |
4525 | } | |
4526 | ctx->compat = in_compat_syscall(); | |
4527 | ctx->account_mem = account_mem; | |
4528 | ctx->user = user; | |
4529 | ||
4530 | ret = io_allocate_scq_urings(ctx, p); | |
4531 | if (ret) | |
4532 | goto err; | |
4533 | ||
6c271ce2 | 4534 | ret = io_sq_offload_start(ctx, p); |
2b188cc1 JA |
4535 | if (ret) |
4536 | goto err; | |
4537 | ||
2b188cc1 | 4538 | memset(&p->sq_off, 0, sizeof(p->sq_off)); |
75b28aff HV |
4539 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
4540 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
4541 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
4542 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
4543 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
4544 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
4545 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
4546 | |
4547 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
4548 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
4549 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
4550 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
4551 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
4552 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
4553 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
ac90f249 | 4554 | |
044c1ab3 JA |
4555 | /* |
4556 | * Install ring fd as the very last thing, so we don't risk someone | |
4557 | * having closed it before we finish setup | |
4558 | */ | |
4559 | ret = io_uring_get_fd(ctx); | |
4560 | if (ret < 0) | |
4561 | goto err; | |
4562 | ||
1d7bb1d5 | 4563 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP; |
c826bd7a | 4564 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
4565 | return ret; |
4566 | err: | |
4567 | io_ring_ctx_wait_and_kill(ctx); | |
4568 | return ret; | |
4569 | } | |
4570 | ||
4571 | /* | |
4572 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
4573 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
4574 | * params structure passed in. | |
4575 | */ | |
4576 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
4577 | { | |
4578 | struct io_uring_params p; | |
4579 | long ret; | |
4580 | int i; | |
4581 | ||
4582 | if (copy_from_user(&p, params, sizeof(p))) | |
4583 | return -EFAULT; | |
4584 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
4585 | if (p.resv[i]) | |
4586 | return -EINVAL; | |
4587 | } | |
4588 | ||
6c271ce2 | 4589 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
33a107f0 | 4590 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE)) |
2b188cc1 JA |
4591 | return -EINVAL; |
4592 | ||
4593 | ret = io_uring_create(entries, &p); | |
4594 | if (ret < 0) | |
4595 | return ret; | |
4596 | ||
4597 | if (copy_to_user(params, &p, sizeof(p))) | |
4598 | return -EFAULT; | |
4599 | ||
4600 | return ret; | |
4601 | } | |
4602 | ||
4603 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
4604 | struct io_uring_params __user *, params) | |
4605 | { | |
4606 | return io_uring_setup(entries, params); | |
4607 | } | |
4608 | ||
edafccee JA |
4609 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
4610 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
4611 | __releases(ctx->uring_lock) |
4612 | __acquires(ctx->uring_lock) | |
edafccee JA |
4613 | { |
4614 | int ret; | |
4615 | ||
35fa71a0 JA |
4616 | /* |
4617 | * We're inside the ring mutex, if the ref is already dying, then | |
4618 | * someone else killed the ctx or is already going through | |
4619 | * io_uring_register(). | |
4620 | */ | |
4621 | if (percpu_ref_is_dying(&ctx->refs)) | |
4622 | return -ENXIO; | |
4623 | ||
edafccee | 4624 | percpu_ref_kill(&ctx->refs); |
b19062a5 JA |
4625 | |
4626 | /* | |
4627 | * Drop uring mutex before waiting for references to exit. If another | |
4628 | * thread is currently inside io_uring_enter() it might need to grab | |
4629 | * the uring_lock to make progress. If we hold it here across the drain | |
4630 | * wait, then we can deadlock. It's safe to drop the mutex here, since | |
4631 | * no new references will come in after we've killed the percpu ref. | |
4632 | */ | |
4633 | mutex_unlock(&ctx->uring_lock); | |
206aefde | 4634 | wait_for_completion(&ctx->completions[0]); |
b19062a5 | 4635 | mutex_lock(&ctx->uring_lock); |
edafccee JA |
4636 | |
4637 | switch (opcode) { | |
4638 | case IORING_REGISTER_BUFFERS: | |
4639 | ret = io_sqe_buffer_register(ctx, arg, nr_args); | |
4640 | break; | |
4641 | case IORING_UNREGISTER_BUFFERS: | |
4642 | ret = -EINVAL; | |
4643 | if (arg || nr_args) | |
4644 | break; | |
4645 | ret = io_sqe_buffer_unregister(ctx); | |
4646 | break; | |
6b06314c JA |
4647 | case IORING_REGISTER_FILES: |
4648 | ret = io_sqe_files_register(ctx, arg, nr_args); | |
4649 | break; | |
4650 | case IORING_UNREGISTER_FILES: | |
4651 | ret = -EINVAL; | |
4652 | if (arg || nr_args) | |
4653 | break; | |
4654 | ret = io_sqe_files_unregister(ctx); | |
4655 | break; | |
c3a31e60 JA |
4656 | case IORING_REGISTER_FILES_UPDATE: |
4657 | ret = io_sqe_files_update(ctx, arg, nr_args); | |
4658 | break; | |
9b402849 JA |
4659 | case IORING_REGISTER_EVENTFD: |
4660 | ret = -EINVAL; | |
4661 | if (nr_args != 1) | |
4662 | break; | |
4663 | ret = io_eventfd_register(ctx, arg); | |
4664 | break; | |
4665 | case IORING_UNREGISTER_EVENTFD: | |
4666 | ret = -EINVAL; | |
4667 | if (arg || nr_args) | |
4668 | break; | |
4669 | ret = io_eventfd_unregister(ctx); | |
4670 | break; | |
edafccee JA |
4671 | default: |
4672 | ret = -EINVAL; | |
4673 | break; | |
4674 | } | |
4675 | ||
4676 | /* bring the ctx back to life */ | |
206aefde | 4677 | reinit_completion(&ctx->completions[0]); |
edafccee JA |
4678 | percpu_ref_reinit(&ctx->refs); |
4679 | return ret; | |
4680 | } | |
4681 | ||
4682 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
4683 | void __user *, arg, unsigned int, nr_args) | |
4684 | { | |
4685 | struct io_ring_ctx *ctx; | |
4686 | long ret = -EBADF; | |
4687 | struct fd f; | |
4688 | ||
4689 | f = fdget(fd); | |
4690 | if (!f.file) | |
4691 | return -EBADF; | |
4692 | ||
4693 | ret = -EOPNOTSUPP; | |
4694 | if (f.file->f_op != &io_uring_fops) | |
4695 | goto out_fput; | |
4696 | ||
4697 | ctx = f.file->private_data; | |
4698 | ||
4699 | mutex_lock(&ctx->uring_lock); | |
4700 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
4701 | mutex_unlock(&ctx->uring_lock); | |
c826bd7a DD |
4702 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, |
4703 | ctx->cq_ev_fd != NULL, ret); | |
edafccee JA |
4704 | out_fput: |
4705 | fdput(f); | |
4706 | return ret; | |
4707 | } | |
4708 | ||
2b188cc1 JA |
4709 | static int __init io_uring_init(void) |
4710 | { | |
4711 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC); | |
4712 | return 0; | |
4713 | }; | |
4714 | __initcall(io_uring_init); |