]>
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 */ |
7c9e7f0f | 344 | #define REQ_F_COMP_LOCKED 16384 /* completion under lock */ |
2b188cc1 | 345 | u64 user_data; |
9e645e11 | 346 | u32 result; |
de0617e4 | 347 | u32 sequence; |
2b188cc1 | 348 | |
fcb323cc JA |
349 | struct list_head inflight_entry; |
350 | ||
561fb04a | 351 | struct io_wq_work work; |
2b188cc1 JA |
352 | }; |
353 | ||
354 | #define IO_PLUG_THRESHOLD 2 | |
def596e9 | 355 | #define IO_IOPOLL_BATCH 8 |
2b188cc1 | 356 | |
9a56a232 JA |
357 | struct io_submit_state { |
358 | struct blk_plug plug; | |
359 | ||
2579f913 JA |
360 | /* |
361 | * io_kiocb alloc cache | |
362 | */ | |
363 | void *reqs[IO_IOPOLL_BATCH]; | |
364 | unsigned int free_reqs; | |
365 | unsigned int cur_req; | |
366 | ||
9a56a232 JA |
367 | /* |
368 | * File reference cache | |
369 | */ | |
370 | struct file *file; | |
371 | unsigned int fd; | |
372 | unsigned int has_refs; | |
373 | unsigned int used_refs; | |
374 | unsigned int ios_left; | |
375 | }; | |
376 | ||
561fb04a | 377 | static void io_wq_submit_work(struct io_wq_work **workptr); |
78e19bbe | 378 | static void io_cqring_fill_event(struct io_kiocb *req, long res); |
4fe2c963 | 379 | static void __io_free_req(struct io_kiocb *req); |
ec9c02ad | 380 | static void io_put_req(struct io_kiocb *req); |
78e19bbe | 381 | static void io_double_put_req(struct io_kiocb *req); |
de0617e4 | 382 | |
2b188cc1 JA |
383 | static struct kmem_cache *req_cachep; |
384 | ||
385 | static const struct file_operations io_uring_fops; | |
386 | ||
387 | struct sock *io_uring_get_socket(struct file *file) | |
388 | { | |
389 | #if defined(CONFIG_UNIX) | |
390 | if (file->f_op == &io_uring_fops) { | |
391 | struct io_ring_ctx *ctx = file->private_data; | |
392 | ||
393 | return ctx->ring_sock->sk; | |
394 | } | |
395 | #endif | |
396 | return NULL; | |
397 | } | |
398 | EXPORT_SYMBOL(io_uring_get_socket); | |
399 | ||
400 | static void io_ring_ctx_ref_free(struct percpu_ref *ref) | |
401 | { | |
402 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
403 | ||
206aefde | 404 | complete(&ctx->completions[0]); |
2b188cc1 JA |
405 | } |
406 | ||
407 | static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) | |
408 | { | |
409 | struct io_ring_ctx *ctx; | |
410 | ||
411 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
412 | if (!ctx) | |
413 | return NULL; | |
414 | ||
0ddf92e8 JA |
415 | ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL); |
416 | if (!ctx->fallback_req) | |
417 | goto err; | |
418 | ||
206aefde JA |
419 | ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL); |
420 | if (!ctx->completions) | |
421 | goto err; | |
422 | ||
21482896 | 423 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
206aefde JA |
424 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) |
425 | goto err; | |
2b188cc1 JA |
426 | |
427 | ctx->flags = p->flags; | |
428 | init_waitqueue_head(&ctx->cq_wait); | |
1d7bb1d5 | 429 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
206aefde JA |
430 | init_completion(&ctx->completions[0]); |
431 | init_completion(&ctx->completions[1]); | |
2b188cc1 JA |
432 | mutex_init(&ctx->uring_lock); |
433 | init_waitqueue_head(&ctx->wait); | |
434 | spin_lock_init(&ctx->completion_lock); | |
def596e9 | 435 | INIT_LIST_HEAD(&ctx->poll_list); |
221c5eb2 | 436 | INIT_LIST_HEAD(&ctx->cancel_list); |
de0617e4 | 437 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 438 | INIT_LIST_HEAD(&ctx->timeout_list); |
fcb323cc JA |
439 | init_waitqueue_head(&ctx->inflight_wait); |
440 | spin_lock_init(&ctx->inflight_lock); | |
441 | INIT_LIST_HEAD(&ctx->inflight_list); | |
2b188cc1 | 442 | return ctx; |
206aefde | 443 | err: |
0ddf92e8 JA |
444 | if (ctx->fallback_req) |
445 | kmem_cache_free(req_cachep, ctx->fallback_req); | |
206aefde JA |
446 | kfree(ctx->completions); |
447 | kfree(ctx); | |
448 | return NULL; | |
2b188cc1 JA |
449 | } |
450 | ||
9d858b21 | 451 | static inline bool __req_need_defer(struct io_kiocb *req) |
7adf4eaf | 452 | { |
a197f664 JL |
453 | struct io_ring_ctx *ctx = req->ctx; |
454 | ||
498ccd9e JA |
455 | return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped |
456 | + atomic_read(&ctx->cached_cq_overflow); | |
7adf4eaf JA |
457 | } |
458 | ||
9d858b21 | 459 | static inline bool req_need_defer(struct io_kiocb *req) |
de0617e4 | 460 | { |
9d858b21 BL |
461 | if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) == REQ_F_IO_DRAIN) |
462 | return __req_need_defer(req); | |
de0617e4 | 463 | |
9d858b21 | 464 | return false; |
de0617e4 JA |
465 | } |
466 | ||
7adf4eaf | 467 | static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx) |
de0617e4 JA |
468 | { |
469 | struct io_kiocb *req; | |
470 | ||
7adf4eaf | 471 | req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list); |
9d858b21 | 472 | if (req && !req_need_defer(req)) { |
de0617e4 JA |
473 | list_del_init(&req->list); |
474 | return req; | |
475 | } | |
476 | ||
477 | return NULL; | |
478 | } | |
479 | ||
5262f567 JA |
480 | static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx) |
481 | { | |
7adf4eaf JA |
482 | struct io_kiocb *req; |
483 | ||
484 | req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list); | |
9d858b21 | 485 | if (req && !__req_need_defer(req)) { |
7adf4eaf JA |
486 | list_del_init(&req->list); |
487 | return req; | |
488 | } | |
489 | ||
490 | return NULL; | |
5262f567 JA |
491 | } |
492 | ||
de0617e4 | 493 | static void __io_commit_cqring(struct io_ring_ctx *ctx) |
2b188cc1 | 494 | { |
75b28aff | 495 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 496 | |
75b28aff | 497 | if (ctx->cached_cq_tail != READ_ONCE(rings->cq.tail)) { |
2b188cc1 | 498 | /* order cqe stores with ring update */ |
75b28aff | 499 | smp_store_release(&rings->cq.tail, ctx->cached_cq_tail); |
2b188cc1 | 500 | |
2b188cc1 JA |
501 | if (wq_has_sleeper(&ctx->cq_wait)) { |
502 | wake_up_interruptible(&ctx->cq_wait); | |
503 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); | |
504 | } | |
505 | } | |
506 | } | |
507 | ||
561fb04a | 508 | static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe) |
18d9be1a | 509 | { |
561fb04a JA |
510 | u8 opcode = READ_ONCE(sqe->opcode); |
511 | ||
512 | return !(opcode == IORING_OP_READ_FIXED || | |
513 | opcode == IORING_OP_WRITE_FIXED); | |
514 | } | |
515 | ||
516 | static inline bool io_prep_async_work(struct io_kiocb *req) | |
517 | { | |
518 | bool do_hashed = false; | |
54a91f3b | 519 | |
6cc47d1d JA |
520 | if (req->submit.sqe) { |
521 | switch (req->submit.sqe->opcode) { | |
522 | case IORING_OP_WRITEV: | |
523 | case IORING_OP_WRITE_FIXED: | |
561fb04a | 524 | do_hashed = true; |
5f8fd2d3 JA |
525 | /* fall-through */ |
526 | case IORING_OP_READV: | |
527 | case IORING_OP_READ_FIXED: | |
528 | case IORING_OP_SENDMSG: | |
529 | case IORING_OP_RECVMSG: | |
530 | case IORING_OP_ACCEPT: | |
531 | case IORING_OP_POLL_ADD: | |
532 | /* | |
533 | * We know REQ_F_ISREG is not set on some of these | |
534 | * opcodes, but this enables us to keep the check in | |
535 | * just one place. | |
536 | */ | |
537 | if (!(req->flags & REQ_F_ISREG)) | |
538 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
6cc47d1d JA |
539 | break; |
540 | } | |
561fb04a JA |
541 | if (io_sqe_needs_user(req->submit.sqe)) |
542 | req->work.flags |= IO_WQ_WORK_NEEDS_USER; | |
54a91f3b JA |
543 | } |
544 | ||
561fb04a JA |
545 | return do_hashed; |
546 | } | |
547 | ||
a197f664 | 548 | static inline void io_queue_async_work(struct io_kiocb *req) |
561fb04a JA |
549 | { |
550 | bool do_hashed = io_prep_async_work(req); | |
a197f664 | 551 | struct io_ring_ctx *ctx = req->ctx; |
561fb04a JA |
552 | |
553 | trace_io_uring_queue_async_work(ctx, do_hashed, req, &req->work, | |
554 | req->flags); | |
555 | if (!do_hashed) { | |
556 | io_wq_enqueue(ctx->io_wq, &req->work); | |
557 | } else { | |
558 | io_wq_enqueue_hashed(ctx->io_wq, &req->work, | |
559 | file_inode(req->file)); | |
560 | } | |
18d9be1a JA |
561 | } |
562 | ||
5262f567 JA |
563 | static void io_kill_timeout(struct io_kiocb *req) |
564 | { | |
565 | int ret; | |
566 | ||
567 | ret = hrtimer_try_to_cancel(&req->timeout.timer); | |
568 | if (ret != -1) { | |
569 | atomic_inc(&req->ctx->cq_timeouts); | |
842f9612 | 570 | list_del_init(&req->list); |
78e19bbe | 571 | io_cqring_fill_event(req, 0); |
ec9c02ad | 572 | io_put_req(req); |
5262f567 JA |
573 | } |
574 | } | |
575 | ||
576 | static void io_kill_timeouts(struct io_ring_ctx *ctx) | |
577 | { | |
578 | struct io_kiocb *req, *tmp; | |
579 | ||
580 | spin_lock_irq(&ctx->completion_lock); | |
581 | list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list) | |
582 | io_kill_timeout(req); | |
583 | spin_unlock_irq(&ctx->completion_lock); | |
584 | } | |
585 | ||
de0617e4 JA |
586 | static void io_commit_cqring(struct io_ring_ctx *ctx) |
587 | { | |
588 | struct io_kiocb *req; | |
589 | ||
5262f567 JA |
590 | while ((req = io_get_timeout_req(ctx)) != NULL) |
591 | io_kill_timeout(req); | |
592 | ||
de0617e4 JA |
593 | __io_commit_cqring(ctx); |
594 | ||
595 | while ((req = io_get_deferred_req(ctx)) != NULL) { | |
4fe2c963 JL |
596 | if (req->flags & REQ_F_SHADOW_DRAIN) { |
597 | /* Just for drain, free it. */ | |
598 | __io_free_req(req); | |
599 | continue; | |
600 | } | |
de0617e4 | 601 | req->flags |= REQ_F_IO_DRAINED; |
a197f664 | 602 | io_queue_async_work(req); |
de0617e4 JA |
603 | } |
604 | } | |
605 | ||
2b188cc1 JA |
606 | static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx) |
607 | { | |
75b28aff | 608 | struct io_rings *rings = ctx->rings; |
2b188cc1 JA |
609 | unsigned tail; |
610 | ||
611 | tail = ctx->cached_cq_tail; | |
115e12e5 SB |
612 | /* |
613 | * writes to the cq entry need to come after reading head; the | |
614 | * control dependency is enough as we're using WRITE_ONCE to | |
615 | * fill the cq entry | |
616 | */ | |
75b28aff | 617 | if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries) |
2b188cc1 JA |
618 | return NULL; |
619 | ||
620 | ctx->cached_cq_tail++; | |
75b28aff | 621 | return &rings->cqes[tail & ctx->cq_mask]; |
2b188cc1 JA |
622 | } |
623 | ||
1d7bb1d5 JA |
624 | static void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
625 | { | |
626 | if (waitqueue_active(&ctx->wait)) | |
627 | wake_up(&ctx->wait); | |
628 | if (waitqueue_active(&ctx->sqo_wait)) | |
629 | wake_up(&ctx->sqo_wait); | |
630 | if (ctx->cq_ev_fd) | |
631 | eventfd_signal(ctx->cq_ev_fd, 1); | |
632 | } | |
633 | ||
634 | static void io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force) | |
635 | { | |
636 | struct io_rings *rings = ctx->rings; | |
637 | struct io_uring_cqe *cqe; | |
638 | struct io_kiocb *req; | |
639 | unsigned long flags; | |
640 | LIST_HEAD(list); | |
641 | ||
642 | if (!force) { | |
643 | if (list_empty_careful(&ctx->cq_overflow_list)) | |
644 | return; | |
645 | if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) == | |
646 | rings->cq_ring_entries)) | |
647 | return; | |
648 | } | |
649 | ||
650 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
651 | ||
652 | /* if force is set, the ring is going away. always drop after that */ | |
653 | if (force) | |
654 | ctx->cq_overflow_flushed = true; | |
655 | ||
656 | while (!list_empty(&ctx->cq_overflow_list)) { | |
657 | cqe = io_get_cqring(ctx); | |
658 | if (!cqe && !force) | |
659 | break; | |
660 | ||
661 | req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb, | |
662 | list); | |
663 | list_move(&req->list, &list); | |
664 | if (cqe) { | |
665 | WRITE_ONCE(cqe->user_data, req->user_data); | |
666 | WRITE_ONCE(cqe->res, req->result); | |
667 | WRITE_ONCE(cqe->flags, 0); | |
668 | } else { | |
669 | WRITE_ONCE(ctx->rings->cq_overflow, | |
670 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
671 | } | |
672 | } | |
673 | ||
674 | io_commit_cqring(ctx); | |
675 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
676 | io_cqring_ev_posted(ctx); | |
677 | ||
678 | while (!list_empty(&list)) { | |
679 | req = list_first_entry(&list, struct io_kiocb, list); | |
680 | list_del(&req->list); | |
ec9c02ad | 681 | io_put_req(req); |
1d7bb1d5 JA |
682 | } |
683 | } | |
684 | ||
78e19bbe | 685 | static void io_cqring_fill_event(struct io_kiocb *req, long res) |
2b188cc1 | 686 | { |
78e19bbe | 687 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
688 | struct io_uring_cqe *cqe; |
689 | ||
78e19bbe | 690 | trace_io_uring_complete(ctx, req->user_data, res); |
51c3ff62 | 691 | |
2b188cc1 JA |
692 | /* |
693 | * If we can't get a cq entry, userspace overflowed the | |
694 | * submission (by quite a lot). Increment the overflow count in | |
695 | * the ring. | |
696 | */ | |
697 | cqe = io_get_cqring(ctx); | |
1d7bb1d5 | 698 | if (likely(cqe)) { |
78e19bbe | 699 | WRITE_ONCE(cqe->user_data, req->user_data); |
2b188cc1 | 700 | WRITE_ONCE(cqe->res, res); |
c71ffb67 | 701 | WRITE_ONCE(cqe->flags, 0); |
1d7bb1d5 | 702 | } else if (ctx->cq_overflow_flushed) { |
498ccd9e JA |
703 | WRITE_ONCE(ctx->rings->cq_overflow, |
704 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
1d7bb1d5 JA |
705 | } else { |
706 | refcount_inc(&req->refs); | |
707 | req->result = res; | |
708 | list_add_tail(&req->list, &ctx->cq_overflow_list); | |
2b188cc1 JA |
709 | } |
710 | } | |
711 | ||
78e19bbe | 712 | static void io_cqring_add_event(struct io_kiocb *req, long res) |
2b188cc1 | 713 | { |
78e19bbe | 714 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
715 | unsigned long flags; |
716 | ||
717 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
78e19bbe | 718 | io_cqring_fill_event(req, res); |
2b188cc1 JA |
719 | io_commit_cqring(ctx); |
720 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
721 | ||
8c838788 | 722 | io_cqring_ev_posted(ctx); |
2b188cc1 JA |
723 | } |
724 | ||
0ddf92e8 JA |
725 | static inline bool io_is_fallback_req(struct io_kiocb *req) |
726 | { | |
727 | return req == (struct io_kiocb *) | |
728 | ((unsigned long) req->ctx->fallback_req & ~1UL); | |
729 | } | |
730 | ||
731 | static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx) | |
732 | { | |
733 | struct io_kiocb *req; | |
734 | ||
735 | req = ctx->fallback_req; | |
736 | if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req)) | |
737 | return req; | |
738 | ||
739 | return NULL; | |
740 | } | |
741 | ||
2579f913 JA |
742 | static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx, |
743 | struct io_submit_state *state) | |
2b188cc1 | 744 | { |
fd6fab2c | 745 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
2b188cc1 JA |
746 | struct io_kiocb *req; |
747 | ||
748 | if (!percpu_ref_tryget(&ctx->refs)) | |
749 | return NULL; | |
750 | ||
2579f913 | 751 | if (!state) { |
fd6fab2c | 752 | req = kmem_cache_alloc(req_cachep, gfp); |
2579f913 | 753 | if (unlikely(!req)) |
0ddf92e8 | 754 | goto fallback; |
2579f913 JA |
755 | } else if (!state->free_reqs) { |
756 | size_t sz; | |
757 | int ret; | |
758 | ||
759 | sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs)); | |
fd6fab2c JA |
760 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs); |
761 | ||
762 | /* | |
763 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
764 | * retry single alloc to be on the safe side. | |
765 | */ | |
766 | if (unlikely(ret <= 0)) { | |
767 | state->reqs[0] = kmem_cache_alloc(req_cachep, gfp); | |
768 | if (!state->reqs[0]) | |
0ddf92e8 | 769 | goto fallback; |
fd6fab2c JA |
770 | ret = 1; |
771 | } | |
2579f913 JA |
772 | state->free_reqs = ret - 1; |
773 | state->cur_req = 1; | |
774 | req = state->reqs[0]; | |
775 | } else { | |
776 | req = state->reqs[state->cur_req]; | |
777 | state->free_reqs--; | |
778 | state->cur_req++; | |
2b188cc1 JA |
779 | } |
780 | ||
0ddf92e8 | 781 | got_it: |
60c112b0 | 782 | req->file = NULL; |
2579f913 JA |
783 | req->ctx = ctx; |
784 | req->flags = 0; | |
e65ef56d JA |
785 | /* one is dropped after submission, the other at completion */ |
786 | refcount_set(&req->refs, 2); | |
9e645e11 | 787 | req->result = 0; |
561fb04a | 788 | INIT_IO_WORK(&req->work, io_wq_submit_work); |
2579f913 | 789 | return req; |
0ddf92e8 JA |
790 | fallback: |
791 | req = io_get_fallback_req(ctx); | |
792 | if (req) | |
793 | goto got_it; | |
6805b32e | 794 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
795 | return NULL; |
796 | } | |
797 | ||
def596e9 JA |
798 | static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr) |
799 | { | |
800 | if (*nr) { | |
801 | kmem_cache_free_bulk(req_cachep, *nr, reqs); | |
6805b32e | 802 | percpu_ref_put_many(&ctx->refs, *nr); |
def596e9 JA |
803 | *nr = 0; |
804 | } | |
805 | } | |
806 | ||
9e645e11 | 807 | static void __io_free_req(struct io_kiocb *req) |
2b188cc1 | 808 | { |
fcb323cc JA |
809 | struct io_ring_ctx *ctx = req->ctx; |
810 | ||
09bb8394 JA |
811 | if (req->file && !(req->flags & REQ_F_FIXED_FILE)) |
812 | fput(req->file); | |
fcb323cc JA |
813 | if (req->flags & REQ_F_INFLIGHT) { |
814 | unsigned long flags; | |
815 | ||
816 | spin_lock_irqsave(&ctx->inflight_lock, flags); | |
817 | list_del(&req->inflight_entry); | |
818 | if (waitqueue_active(&ctx->inflight_wait)) | |
819 | wake_up(&ctx->inflight_wait); | |
820 | spin_unlock_irqrestore(&ctx->inflight_lock, flags); | |
821 | } | |
822 | percpu_ref_put(&ctx->refs); | |
0ddf92e8 JA |
823 | if (likely(!io_is_fallback_req(req))) |
824 | kmem_cache_free(req_cachep, req); | |
825 | else | |
826 | clear_bit_unlock(0, (unsigned long *) ctx->fallback_req); | |
e65ef56d JA |
827 | } |
828 | ||
a197f664 | 829 | static bool io_link_cancel_timeout(struct io_kiocb *req) |
2665abfd | 830 | { |
a197f664 | 831 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd JA |
832 | int ret; |
833 | ||
834 | ret = hrtimer_try_to_cancel(&req->timeout.timer); | |
835 | if (ret != -1) { | |
78e19bbe | 836 | io_cqring_fill_event(req, -ECANCELED); |
2665abfd JA |
837 | io_commit_cqring(ctx); |
838 | req->flags &= ~REQ_F_LINK; | |
ec9c02ad | 839 | io_put_req(req); |
2665abfd JA |
840 | return true; |
841 | } | |
842 | ||
843 | return false; | |
844 | } | |
845 | ||
ba816ad6 | 846 | static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr) |
9e645e11 | 847 | { |
2665abfd | 848 | struct io_ring_ctx *ctx = req->ctx; |
9e645e11 | 849 | struct io_kiocb *nxt; |
2665abfd | 850 | bool wake_ev = false; |
9e645e11 JA |
851 | |
852 | /* | |
853 | * The list should never be empty when we are called here. But could | |
854 | * potentially happen if the chain is messed up, check to be on the | |
855 | * safe side. | |
856 | */ | |
857 | nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list); | |
2665abfd | 858 | while (nxt) { |
76a46e06 | 859 | list_del_init(&nxt->list); |
9e645e11 JA |
860 | if (!list_empty(&req->link_list)) { |
861 | INIT_LIST_HEAD(&nxt->link_list); | |
862 | list_splice(&req->link_list, &nxt->link_list); | |
863 | nxt->flags |= REQ_F_LINK; | |
864 | } | |
865 | ||
ba816ad6 JA |
866 | /* |
867 | * If we're in async work, we can continue processing the chain | |
868 | * in this context instead of having to queue up new async work. | |
869 | */ | |
2665abfd | 870 | if (req->flags & REQ_F_LINK_TIMEOUT) { |
a197f664 | 871 | wake_ev = io_link_cancel_timeout(nxt); |
2665abfd JA |
872 | |
873 | /* we dropped this link, get next */ | |
874 | nxt = list_first_entry_or_null(&req->link_list, | |
875 | struct io_kiocb, list); | |
960e432d | 876 | } else if (nxtptr && io_wq_current_is_worker()) { |
ba816ad6 | 877 | *nxtptr = nxt; |
2665abfd JA |
878 | break; |
879 | } else { | |
a197f664 | 880 | io_queue_async_work(nxt); |
2665abfd JA |
881 | break; |
882 | } | |
9e645e11 | 883 | } |
2665abfd JA |
884 | |
885 | if (wake_ev) | |
886 | io_cqring_ev_posted(ctx); | |
9e645e11 JA |
887 | } |
888 | ||
889 | /* | |
890 | * Called if REQ_F_LINK is set, and we fail the head request | |
891 | */ | |
892 | static void io_fail_links(struct io_kiocb *req) | |
893 | { | |
2665abfd | 894 | struct io_ring_ctx *ctx = req->ctx; |
9e645e11 | 895 | struct io_kiocb *link; |
2665abfd JA |
896 | unsigned long flags; |
897 | ||
898 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
9e645e11 JA |
899 | |
900 | while (!list_empty(&req->link_list)) { | |
901 | link = list_first_entry(&req->link_list, struct io_kiocb, list); | |
2665abfd | 902 | list_del_init(&link->list); |
9e645e11 | 903 | |
c826bd7a | 904 | trace_io_uring_fail_link(req, link); |
2665abfd JA |
905 | |
906 | if ((req->flags & REQ_F_LINK_TIMEOUT) && | |
907 | link->submit.sqe->opcode == IORING_OP_LINK_TIMEOUT) { | |
a197f664 | 908 | io_link_cancel_timeout(link); |
2665abfd | 909 | } else { |
78e19bbe JA |
910 | io_cqring_fill_event(link, -ECANCELED); |
911 | io_double_put_req(link); | |
2665abfd | 912 | } |
9e645e11 | 913 | } |
2665abfd JA |
914 | |
915 | io_commit_cqring(ctx); | |
916 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
917 | io_cqring_ev_posted(ctx); | |
9e645e11 JA |
918 | } |
919 | ||
c69f8dbe | 920 | static void io_free_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt) |
9e645e11 | 921 | { |
2665abfd JA |
922 | if (likely(!(req->flags & REQ_F_LINK))) { |
923 | __io_free_req(req); | |
924 | return; | |
925 | } | |
926 | ||
9e645e11 JA |
927 | /* |
928 | * If LINK is set, we have dependent requests in this chain. If we | |
929 | * didn't fail this request, queue the first one up, moving any other | |
930 | * dependencies to the next request. In case of failure, fail the rest | |
931 | * of the chain. | |
932 | */ | |
2665abfd JA |
933 | if (req->flags & REQ_F_FAIL_LINK) { |
934 | io_fail_links(req); | |
7c9e7f0f JA |
935 | } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) == |
936 | REQ_F_LINK_TIMEOUT) { | |
2665abfd JA |
937 | struct io_ring_ctx *ctx = req->ctx; |
938 | unsigned long flags; | |
939 | ||
940 | /* | |
941 | * If this is a timeout link, we could be racing with the | |
942 | * timeout timer. Grab the completion lock for this case to | |
7c9e7f0f | 943 | * protect against that. |
2665abfd JA |
944 | */ |
945 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
946 | io_req_link_next(req, nxt); | |
947 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
948 | } else { | |
949 | io_req_link_next(req, nxt); | |
9e645e11 JA |
950 | } |
951 | ||
952 | __io_free_req(req); | |
953 | } | |
954 | ||
c69f8dbe JL |
955 | static void io_free_req(struct io_kiocb *req) |
956 | { | |
957 | io_free_req_find_next(req, NULL); | |
958 | } | |
959 | ||
ba816ad6 JA |
960 | /* |
961 | * Drop reference to request, return next in chain (if there is one) if this | |
962 | * was the last reference to this request. | |
963 | */ | |
ec9c02ad | 964 | static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr) |
e65ef56d | 965 | { |
ba816ad6 JA |
966 | struct io_kiocb *nxt = NULL; |
967 | ||
e65ef56d | 968 | if (refcount_dec_and_test(&req->refs)) |
c69f8dbe | 969 | io_free_req_find_next(req, &nxt); |
ba816ad6 | 970 | |
ba816ad6 | 971 | if (nxt) { |
561fb04a | 972 | if (nxtptr) |
ba816ad6 | 973 | *nxtptr = nxt; |
561fb04a | 974 | else |
a197f664 | 975 | io_queue_async_work(nxt); |
ba816ad6 | 976 | } |
2b188cc1 JA |
977 | } |
978 | ||
ec9c02ad JL |
979 | static void io_put_req(struct io_kiocb *req) |
980 | { | |
981 | if (refcount_dec_and_test(&req->refs)) | |
c69f8dbe | 982 | io_free_req(req); |
ec9c02ad JL |
983 | } |
984 | ||
78e19bbe JA |
985 | static void io_double_put_req(struct io_kiocb *req) |
986 | { | |
987 | /* drop both submit and complete references */ | |
988 | if (refcount_sub_and_test(2, &req->refs)) | |
989 | __io_free_req(req); | |
990 | } | |
991 | ||
1d7bb1d5 | 992 | static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush) |
a3a0e43f | 993 | { |
84f97dc2 JA |
994 | struct io_rings *rings = ctx->rings; |
995 | ||
1d7bb1d5 JA |
996 | /* |
997 | * noflush == true is from the waitqueue handler, just ensure we wake | |
998 | * up the task, and the next invocation will flush the entries. We | |
999 | * cannot safely to it from here. | |
1000 | */ | |
1001 | if (noflush && !list_empty(&ctx->cq_overflow_list)) | |
1002 | return -1U; | |
1003 | ||
1004 | io_cqring_overflow_flush(ctx, false); | |
1005 | ||
a3a0e43f JA |
1006 | /* See comment at the top of this file */ |
1007 | smp_rmb(); | |
75b28aff | 1008 | return READ_ONCE(rings->cq.tail) - READ_ONCE(rings->cq.head); |
a3a0e43f JA |
1009 | } |
1010 | ||
fb5ccc98 PB |
1011 | static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx) |
1012 | { | |
1013 | struct io_rings *rings = ctx->rings; | |
1014 | ||
1015 | /* make sure SQ entry isn't read before tail */ | |
1016 | return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head; | |
1017 | } | |
1018 | ||
def596e9 JA |
1019 | /* |
1020 | * Find and free completed poll iocbs | |
1021 | */ | |
1022 | static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1023 | struct list_head *done) | |
1024 | { | |
1025 | void *reqs[IO_IOPOLL_BATCH]; | |
1026 | struct io_kiocb *req; | |
09bb8394 | 1027 | int to_free; |
def596e9 | 1028 | |
09bb8394 | 1029 | to_free = 0; |
def596e9 JA |
1030 | while (!list_empty(done)) { |
1031 | req = list_first_entry(done, struct io_kiocb, list); | |
1032 | list_del(&req->list); | |
1033 | ||
78e19bbe | 1034 | io_cqring_fill_event(req, req->result); |
def596e9 JA |
1035 | (*nr_events)++; |
1036 | ||
09bb8394 JA |
1037 | if (refcount_dec_and_test(&req->refs)) { |
1038 | /* If we're not using fixed files, we have to pair the | |
1039 | * completion part with the file put. Use regular | |
1040 | * completions for those, only batch free for fixed | |
9e645e11 | 1041 | * file and non-linked commands. |
09bb8394 | 1042 | */ |
0ddf92e8 JA |
1043 | if (((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) == |
1044 | REQ_F_FIXED_FILE) && !io_is_fallback_req(req)) { | |
09bb8394 JA |
1045 | reqs[to_free++] = req; |
1046 | if (to_free == ARRAY_SIZE(reqs)) | |
1047 | io_free_req_many(ctx, reqs, &to_free); | |
6b06314c | 1048 | } else { |
c69f8dbe | 1049 | io_free_req(req); |
6b06314c | 1050 | } |
9a56a232 | 1051 | } |
def596e9 | 1052 | } |
def596e9 | 1053 | |
09bb8394 | 1054 | io_commit_cqring(ctx); |
def596e9 JA |
1055 | io_free_req_many(ctx, reqs, &to_free); |
1056 | } | |
1057 | ||
1058 | static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1059 | long min) | |
1060 | { | |
1061 | struct io_kiocb *req, *tmp; | |
1062 | LIST_HEAD(done); | |
1063 | bool spin; | |
1064 | int ret; | |
1065 | ||
1066 | /* | |
1067 | * Only spin for completions if we don't have multiple devices hanging | |
1068 | * off our complete list, and we're under the requested amount. | |
1069 | */ | |
1070 | spin = !ctx->poll_multi_file && *nr_events < min; | |
1071 | ||
1072 | ret = 0; | |
1073 | list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) { | |
1074 | struct kiocb *kiocb = &req->rw; | |
1075 | ||
1076 | /* | |
1077 | * Move completed entries to our local list. If we find a | |
1078 | * request that requires polling, break out and complete | |
1079 | * the done list first, if we have entries there. | |
1080 | */ | |
1081 | if (req->flags & REQ_F_IOPOLL_COMPLETED) { | |
1082 | list_move_tail(&req->list, &done); | |
1083 | continue; | |
1084 | } | |
1085 | if (!list_empty(&done)) | |
1086 | break; | |
1087 | ||
1088 | ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin); | |
1089 | if (ret < 0) | |
1090 | break; | |
1091 | ||
1092 | if (ret && spin) | |
1093 | spin = false; | |
1094 | ret = 0; | |
1095 | } | |
1096 | ||
1097 | if (!list_empty(&done)) | |
1098 | io_iopoll_complete(ctx, nr_events, &done); | |
1099 | ||
1100 | return ret; | |
1101 | } | |
1102 | ||
1103 | /* | |
1104 | * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a | |
1105 | * non-spinning poll check - we'll still enter the driver poll loop, but only | |
1106 | * as a non-spinning completion check. | |
1107 | */ | |
1108 | static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1109 | long min) | |
1110 | { | |
08f5439f | 1111 | while (!list_empty(&ctx->poll_list) && !need_resched()) { |
def596e9 JA |
1112 | int ret; |
1113 | ||
1114 | ret = io_do_iopoll(ctx, nr_events, min); | |
1115 | if (ret < 0) | |
1116 | return ret; | |
1117 | if (!min || *nr_events >= min) | |
1118 | return 0; | |
1119 | } | |
1120 | ||
1121 | return 1; | |
1122 | } | |
1123 | ||
1124 | /* | |
1125 | * We can't just wait for polled events to come to us, we have to actively | |
1126 | * find and complete them. | |
1127 | */ | |
1128 | static void io_iopoll_reap_events(struct io_ring_ctx *ctx) | |
1129 | { | |
1130 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
1131 | return; | |
1132 | ||
1133 | mutex_lock(&ctx->uring_lock); | |
1134 | while (!list_empty(&ctx->poll_list)) { | |
1135 | unsigned int nr_events = 0; | |
1136 | ||
1137 | io_iopoll_getevents(ctx, &nr_events, 1); | |
08f5439f JA |
1138 | |
1139 | /* | |
1140 | * Ensure we allow local-to-the-cpu processing to take place, | |
1141 | * in this case we need to ensure that we reap all events. | |
1142 | */ | |
1143 | cond_resched(); | |
def596e9 JA |
1144 | } |
1145 | mutex_unlock(&ctx->uring_lock); | |
1146 | } | |
1147 | ||
2b2ed975 JA |
1148 | static int __io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, |
1149 | long min) | |
def596e9 | 1150 | { |
2b2ed975 | 1151 | int iters = 0, ret = 0; |
500f9fba | 1152 | |
def596e9 JA |
1153 | do { |
1154 | int tmin = 0; | |
1155 | ||
a3a0e43f JA |
1156 | /* |
1157 | * Don't enter poll loop if we already have events pending. | |
1158 | * If we do, we can potentially be spinning for commands that | |
1159 | * already triggered a CQE (eg in error). | |
1160 | */ | |
1d7bb1d5 | 1161 | if (io_cqring_events(ctx, false)) |
a3a0e43f JA |
1162 | break; |
1163 | ||
500f9fba JA |
1164 | /* |
1165 | * If a submit got punted to a workqueue, we can have the | |
1166 | * application entering polling for a command before it gets | |
1167 | * issued. That app will hold the uring_lock for the duration | |
1168 | * of the poll right here, so we need to take a breather every | |
1169 | * now and then to ensure that the issue has a chance to add | |
1170 | * the poll to the issued list. Otherwise we can spin here | |
1171 | * forever, while the workqueue is stuck trying to acquire the | |
1172 | * very same mutex. | |
1173 | */ | |
1174 | if (!(++iters & 7)) { | |
1175 | mutex_unlock(&ctx->uring_lock); | |
1176 | mutex_lock(&ctx->uring_lock); | |
1177 | } | |
1178 | ||
def596e9 JA |
1179 | if (*nr_events < min) |
1180 | tmin = min - *nr_events; | |
1181 | ||
1182 | ret = io_iopoll_getevents(ctx, nr_events, tmin); | |
1183 | if (ret <= 0) | |
1184 | break; | |
1185 | ret = 0; | |
1186 | } while (min && !*nr_events && !need_resched()); | |
1187 | ||
2b2ed975 JA |
1188 | return ret; |
1189 | } | |
1190 | ||
1191 | static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, | |
1192 | long min) | |
1193 | { | |
1194 | int ret; | |
1195 | ||
1196 | /* | |
1197 | * We disallow the app entering submit/complete with polling, but we | |
1198 | * still need to lock the ring to prevent racing with polled issue | |
1199 | * that got punted to a workqueue. | |
1200 | */ | |
1201 | mutex_lock(&ctx->uring_lock); | |
1202 | ret = __io_iopoll_check(ctx, nr_events, min); | |
500f9fba | 1203 | mutex_unlock(&ctx->uring_lock); |
def596e9 JA |
1204 | return ret; |
1205 | } | |
1206 | ||
491381ce | 1207 | static void kiocb_end_write(struct io_kiocb *req) |
2b188cc1 | 1208 | { |
491381ce JA |
1209 | /* |
1210 | * Tell lockdep we inherited freeze protection from submission | |
1211 | * thread. | |
1212 | */ | |
1213 | if (req->flags & REQ_F_ISREG) { | |
1214 | struct inode *inode = file_inode(req->file); | |
2b188cc1 | 1215 | |
491381ce | 1216 | __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE); |
2b188cc1 | 1217 | } |
491381ce | 1218 | file_end_write(req->file); |
2b188cc1 JA |
1219 | } |
1220 | ||
ba816ad6 | 1221 | static void io_complete_rw_common(struct kiocb *kiocb, long res) |
2b188cc1 JA |
1222 | { |
1223 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
1224 | ||
491381ce JA |
1225 | if (kiocb->ki_flags & IOCB_WRITE) |
1226 | kiocb_end_write(req); | |
2b188cc1 | 1227 | |
9e645e11 JA |
1228 | if ((req->flags & REQ_F_LINK) && res != req->result) |
1229 | req->flags |= REQ_F_FAIL_LINK; | |
78e19bbe | 1230 | io_cqring_add_event(req, res); |
ba816ad6 JA |
1231 | } |
1232 | ||
1233 | static void io_complete_rw(struct kiocb *kiocb, long res, long res2) | |
1234 | { | |
1235 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
1236 | ||
1237 | io_complete_rw_common(kiocb, res); | |
ec9c02ad | 1238 | io_put_req(req); |
ba816ad6 JA |
1239 | } |
1240 | ||
1241 | static struct io_kiocb *__io_complete_rw(struct kiocb *kiocb, long res) | |
1242 | { | |
1243 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
ec9c02ad | 1244 | struct io_kiocb *nxt = NULL; |
ba816ad6 JA |
1245 | |
1246 | io_complete_rw_common(kiocb, res); | |
ec9c02ad JL |
1247 | io_put_req_find_next(req, &nxt); |
1248 | ||
1249 | return nxt; | |
2b188cc1 JA |
1250 | } |
1251 | ||
def596e9 JA |
1252 | static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2) |
1253 | { | |
1254 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
1255 | ||
491381ce JA |
1256 | if (kiocb->ki_flags & IOCB_WRITE) |
1257 | kiocb_end_write(req); | |
def596e9 | 1258 | |
9e645e11 JA |
1259 | if ((req->flags & REQ_F_LINK) && res != req->result) |
1260 | req->flags |= REQ_F_FAIL_LINK; | |
1261 | req->result = res; | |
def596e9 JA |
1262 | if (res != -EAGAIN) |
1263 | req->flags |= REQ_F_IOPOLL_COMPLETED; | |
1264 | } | |
1265 | ||
1266 | /* | |
1267 | * After the iocb has been issued, it's safe to be found on the poll list. | |
1268 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
1269 | * find it from a io_iopoll_getevents() thread before the issuer is done | |
1270 | * accessing the kiocb cookie. | |
1271 | */ | |
1272 | static void io_iopoll_req_issued(struct io_kiocb *req) | |
1273 | { | |
1274 | struct io_ring_ctx *ctx = req->ctx; | |
1275 | ||
1276 | /* | |
1277 | * Track whether we have multiple files in our lists. This will impact | |
1278 | * how we do polling eventually, not spinning if we're on potentially | |
1279 | * different devices. | |
1280 | */ | |
1281 | if (list_empty(&ctx->poll_list)) { | |
1282 | ctx->poll_multi_file = false; | |
1283 | } else if (!ctx->poll_multi_file) { | |
1284 | struct io_kiocb *list_req; | |
1285 | ||
1286 | list_req = list_first_entry(&ctx->poll_list, struct io_kiocb, | |
1287 | list); | |
1288 | if (list_req->rw.ki_filp != req->rw.ki_filp) | |
1289 | ctx->poll_multi_file = true; | |
1290 | } | |
1291 | ||
1292 | /* | |
1293 | * For fast devices, IO may have already completed. If it has, add | |
1294 | * it to the front so we find it first. | |
1295 | */ | |
1296 | if (req->flags & REQ_F_IOPOLL_COMPLETED) | |
1297 | list_add(&req->list, &ctx->poll_list); | |
1298 | else | |
1299 | list_add_tail(&req->list, &ctx->poll_list); | |
1300 | } | |
1301 | ||
3d6770fb | 1302 | static void io_file_put(struct io_submit_state *state) |
9a56a232 | 1303 | { |
3d6770fb | 1304 | if (state->file) { |
9a56a232 JA |
1305 | int diff = state->has_refs - state->used_refs; |
1306 | ||
1307 | if (diff) | |
1308 | fput_many(state->file, diff); | |
1309 | state->file = NULL; | |
1310 | } | |
1311 | } | |
1312 | ||
1313 | /* | |
1314 | * Get as many references to a file as we have IOs left in this submission, | |
1315 | * assuming most submissions are for one file, or at least that each file | |
1316 | * has more than one submission. | |
1317 | */ | |
1318 | static struct file *io_file_get(struct io_submit_state *state, int fd) | |
1319 | { | |
1320 | if (!state) | |
1321 | return fget(fd); | |
1322 | ||
1323 | if (state->file) { | |
1324 | if (state->fd == fd) { | |
1325 | state->used_refs++; | |
1326 | state->ios_left--; | |
1327 | return state->file; | |
1328 | } | |
3d6770fb | 1329 | io_file_put(state); |
9a56a232 JA |
1330 | } |
1331 | state->file = fget_many(fd, state->ios_left); | |
1332 | if (!state->file) | |
1333 | return NULL; | |
1334 | ||
1335 | state->fd = fd; | |
1336 | state->has_refs = state->ios_left; | |
1337 | state->used_refs = 1; | |
1338 | state->ios_left--; | |
1339 | return state->file; | |
1340 | } | |
1341 | ||
2b188cc1 JA |
1342 | /* |
1343 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1344 | * any file. For now, just ensure that anything potentially problematic is done | |
1345 | * inline. | |
1346 | */ | |
1347 | static bool io_file_supports_async(struct file *file) | |
1348 | { | |
1349 | umode_t mode = file_inode(file)->i_mode; | |
1350 | ||
1351 | if (S_ISBLK(mode) || S_ISCHR(mode)) | |
1352 | return true; | |
1353 | if (S_ISREG(mode) && file->f_op != &io_uring_fops) | |
1354 | return true; | |
1355 | ||
1356 | return false; | |
1357 | } | |
1358 | ||
267bc904 | 1359 | static int io_prep_rw(struct io_kiocb *req, bool force_nonblock) |
2b188cc1 | 1360 | { |
267bc904 | 1361 | const struct io_uring_sqe *sqe = req->submit.sqe; |
def596e9 | 1362 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 | 1363 | struct kiocb *kiocb = &req->rw; |
09bb8394 JA |
1364 | unsigned ioprio; |
1365 | int ret; | |
2b188cc1 | 1366 | |
09bb8394 JA |
1367 | if (!req->file) |
1368 | return -EBADF; | |
2b188cc1 | 1369 | |
491381ce JA |
1370 | if (S_ISREG(file_inode(req->file)->i_mode)) |
1371 | req->flags |= REQ_F_ISREG; | |
1372 | ||
1373 | /* | |
1374 | * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so | |
1375 | * we know to async punt it even if it was opened O_NONBLOCK | |
1376 | */ | |
1377 | if (force_nonblock && !io_file_supports_async(req->file)) { | |
1378 | req->flags |= REQ_F_MUST_PUNT; | |
1379 | return -EAGAIN; | |
1380 | } | |
6b06314c | 1381 | |
2b188cc1 JA |
1382 | kiocb->ki_pos = READ_ONCE(sqe->off); |
1383 | kiocb->ki_flags = iocb_flags(kiocb->ki_filp); | |
1384 | kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp)); | |
1385 | ||
1386 | ioprio = READ_ONCE(sqe->ioprio); | |
1387 | if (ioprio) { | |
1388 | ret = ioprio_check_cap(ioprio); | |
1389 | if (ret) | |
09bb8394 | 1390 | return ret; |
2b188cc1 JA |
1391 | |
1392 | kiocb->ki_ioprio = ioprio; | |
1393 | } else | |
1394 | kiocb->ki_ioprio = get_current_ioprio(); | |
1395 | ||
1396 | ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags)); | |
1397 | if (unlikely(ret)) | |
09bb8394 | 1398 | return ret; |
8449eeda SB |
1399 | |
1400 | /* don't allow async punt if RWF_NOWAIT was requested */ | |
491381ce JA |
1401 | if ((kiocb->ki_flags & IOCB_NOWAIT) || |
1402 | (req->file->f_flags & O_NONBLOCK)) | |
8449eeda SB |
1403 | req->flags |= REQ_F_NOWAIT; |
1404 | ||
1405 | if (force_nonblock) | |
2b188cc1 | 1406 | kiocb->ki_flags |= IOCB_NOWAIT; |
8449eeda | 1407 | |
def596e9 | 1408 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 JA |
1409 | if (!(kiocb->ki_flags & IOCB_DIRECT) || |
1410 | !kiocb->ki_filp->f_op->iopoll) | |
09bb8394 | 1411 | return -EOPNOTSUPP; |
2b188cc1 | 1412 | |
def596e9 JA |
1413 | kiocb->ki_flags |= IOCB_HIPRI; |
1414 | kiocb->ki_complete = io_complete_rw_iopoll; | |
1415 | } else { | |
09bb8394 JA |
1416 | if (kiocb->ki_flags & IOCB_HIPRI) |
1417 | return -EINVAL; | |
def596e9 JA |
1418 | kiocb->ki_complete = io_complete_rw; |
1419 | } | |
2b188cc1 | 1420 | return 0; |
2b188cc1 JA |
1421 | } |
1422 | ||
1423 | static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) | |
1424 | { | |
1425 | switch (ret) { | |
1426 | case -EIOCBQUEUED: | |
1427 | break; | |
1428 | case -ERESTARTSYS: | |
1429 | case -ERESTARTNOINTR: | |
1430 | case -ERESTARTNOHAND: | |
1431 | case -ERESTART_RESTARTBLOCK: | |
1432 | /* | |
1433 | * We can't just restart the syscall, since previously | |
1434 | * submitted sqes may already be in progress. Just fail this | |
1435 | * IO with EINTR. | |
1436 | */ | |
1437 | ret = -EINTR; | |
1438 | /* fall through */ | |
1439 | default: | |
1440 | kiocb->ki_complete(kiocb, ret, 0); | |
1441 | } | |
1442 | } | |
1443 | ||
ba816ad6 JA |
1444 | static void kiocb_done(struct kiocb *kiocb, ssize_t ret, struct io_kiocb **nxt, |
1445 | bool in_async) | |
1446 | { | |
1447 | if (in_async && ret >= 0 && nxt && kiocb->ki_complete == io_complete_rw) | |
1448 | *nxt = __io_complete_rw(kiocb, ret); | |
1449 | else | |
1450 | io_rw_done(kiocb, ret); | |
1451 | } | |
1452 | ||
edafccee JA |
1453 | static int io_import_fixed(struct io_ring_ctx *ctx, int rw, |
1454 | const struct io_uring_sqe *sqe, | |
1455 | struct iov_iter *iter) | |
1456 | { | |
1457 | size_t len = READ_ONCE(sqe->len); | |
1458 | struct io_mapped_ubuf *imu; | |
1459 | unsigned index, buf_index; | |
1460 | size_t offset; | |
1461 | u64 buf_addr; | |
1462 | ||
1463 | /* attempt to use fixed buffers without having provided iovecs */ | |
1464 | if (unlikely(!ctx->user_bufs)) | |
1465 | return -EFAULT; | |
1466 | ||
1467 | buf_index = READ_ONCE(sqe->buf_index); | |
1468 | if (unlikely(buf_index >= ctx->nr_user_bufs)) | |
1469 | return -EFAULT; | |
1470 | ||
1471 | index = array_index_nospec(buf_index, ctx->nr_user_bufs); | |
1472 | imu = &ctx->user_bufs[index]; | |
1473 | buf_addr = READ_ONCE(sqe->addr); | |
1474 | ||
1475 | /* overflow */ | |
1476 | if (buf_addr + len < buf_addr) | |
1477 | return -EFAULT; | |
1478 | /* not inside the mapped region */ | |
1479 | if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len) | |
1480 | return -EFAULT; | |
1481 | ||
1482 | /* | |
1483 | * May not be a start of buffer, set size appropriately | |
1484 | * and advance us to the beginning. | |
1485 | */ | |
1486 | offset = buf_addr - imu->ubuf; | |
1487 | iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len); | |
bd11b3a3 JA |
1488 | |
1489 | if (offset) { | |
1490 | /* | |
1491 | * Don't use iov_iter_advance() here, as it's really slow for | |
1492 | * using the latter parts of a big fixed buffer - it iterates | |
1493 | * over each segment manually. We can cheat a bit here, because | |
1494 | * we know that: | |
1495 | * | |
1496 | * 1) it's a BVEC iter, we set it up | |
1497 | * 2) all bvecs are PAGE_SIZE in size, except potentially the | |
1498 | * first and last bvec | |
1499 | * | |
1500 | * So just find our index, and adjust the iterator afterwards. | |
1501 | * If the offset is within the first bvec (or the whole first | |
1502 | * bvec, just use iov_iter_advance(). This makes it easier | |
1503 | * since we can just skip the first segment, which may not | |
1504 | * be PAGE_SIZE aligned. | |
1505 | */ | |
1506 | const struct bio_vec *bvec = imu->bvec; | |
1507 | ||
1508 | if (offset <= bvec->bv_len) { | |
1509 | iov_iter_advance(iter, offset); | |
1510 | } else { | |
1511 | unsigned long seg_skip; | |
1512 | ||
1513 | /* skip first vec */ | |
1514 | offset -= bvec->bv_len; | |
1515 | seg_skip = 1 + (offset >> PAGE_SHIFT); | |
1516 | ||
1517 | iter->bvec = bvec + seg_skip; | |
1518 | iter->nr_segs -= seg_skip; | |
99c79f66 | 1519 | iter->count -= bvec->bv_len + offset; |
bd11b3a3 | 1520 | iter->iov_offset = offset & ~PAGE_MASK; |
bd11b3a3 JA |
1521 | } |
1522 | } | |
1523 | ||
edafccee JA |
1524 | return 0; |
1525 | } | |
1526 | ||
87e5e6da JA |
1527 | static ssize_t io_import_iovec(struct io_ring_ctx *ctx, int rw, |
1528 | const struct sqe_submit *s, struct iovec **iovec, | |
1529 | struct iov_iter *iter) | |
2b188cc1 JA |
1530 | { |
1531 | const struct io_uring_sqe *sqe = s->sqe; | |
1532 | void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
1533 | size_t sqe_len = READ_ONCE(sqe->len); | |
edafccee JA |
1534 | u8 opcode; |
1535 | ||
1536 | /* | |
1537 | * We're reading ->opcode for the second time, but the first read | |
1538 | * doesn't care whether it's _FIXED or not, so it doesn't matter | |
1539 | * whether ->opcode changes concurrently. The first read does care | |
1540 | * about whether it is a READ or a WRITE, so we don't trust this read | |
1541 | * for that purpose and instead let the caller pass in the read/write | |
1542 | * flag. | |
1543 | */ | |
1544 | opcode = READ_ONCE(sqe->opcode); | |
1545 | if (opcode == IORING_OP_READ_FIXED || | |
1546 | opcode == IORING_OP_WRITE_FIXED) { | |
87e5e6da | 1547 | ssize_t ret = io_import_fixed(ctx, rw, sqe, iter); |
edafccee JA |
1548 | *iovec = NULL; |
1549 | return ret; | |
1550 | } | |
2b188cc1 JA |
1551 | |
1552 | if (!s->has_user) | |
1553 | return -EFAULT; | |
1554 | ||
1555 | #ifdef CONFIG_COMPAT | |
1556 | if (ctx->compat) | |
1557 | return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV, | |
1558 | iovec, iter); | |
1559 | #endif | |
1560 | ||
1561 | return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter); | |
1562 | } | |
1563 | ||
32960613 JA |
1564 | /* |
1565 | * For files that don't have ->read_iter() and ->write_iter(), handle them | |
1566 | * by looping over ->read() or ->write() manually. | |
1567 | */ | |
1568 | static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb, | |
1569 | struct iov_iter *iter) | |
1570 | { | |
1571 | ssize_t ret = 0; | |
1572 | ||
1573 | /* | |
1574 | * Don't support polled IO through this interface, and we can't | |
1575 | * support non-blocking either. For the latter, this just causes | |
1576 | * the kiocb to be handled from an async context. | |
1577 | */ | |
1578 | if (kiocb->ki_flags & IOCB_HIPRI) | |
1579 | return -EOPNOTSUPP; | |
1580 | if (kiocb->ki_flags & IOCB_NOWAIT) | |
1581 | return -EAGAIN; | |
1582 | ||
1583 | while (iov_iter_count(iter)) { | |
1584 | struct iovec iovec = iov_iter_iovec(iter); | |
1585 | ssize_t nr; | |
1586 | ||
1587 | if (rw == READ) { | |
1588 | nr = file->f_op->read(file, iovec.iov_base, | |
1589 | iovec.iov_len, &kiocb->ki_pos); | |
1590 | } else { | |
1591 | nr = file->f_op->write(file, iovec.iov_base, | |
1592 | iovec.iov_len, &kiocb->ki_pos); | |
1593 | } | |
1594 | ||
1595 | if (nr < 0) { | |
1596 | if (!ret) | |
1597 | ret = nr; | |
1598 | break; | |
1599 | } | |
1600 | ret += nr; | |
1601 | if (nr != iovec.iov_len) | |
1602 | break; | |
1603 | iov_iter_advance(iter, nr); | |
1604 | } | |
1605 | ||
1606 | return ret; | |
1607 | } | |
1608 | ||
267bc904 PB |
1609 | static int io_read(struct io_kiocb *req, struct io_kiocb **nxt, |
1610 | bool force_nonblock) | |
2b188cc1 JA |
1611 | { |
1612 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1613 | struct kiocb *kiocb = &req->rw; | |
1614 | struct iov_iter iter; | |
1615 | struct file *file; | |
31b51510 | 1616 | size_t iov_count; |
9d93a3f5 | 1617 | ssize_t read_size, ret; |
2b188cc1 | 1618 | |
267bc904 | 1619 | ret = io_prep_rw(req, force_nonblock); |
2b188cc1 JA |
1620 | if (ret) |
1621 | return ret; | |
1622 | file = kiocb->ki_filp; | |
1623 | ||
2b188cc1 | 1624 | if (unlikely(!(file->f_mode & FMODE_READ))) |
09bb8394 | 1625 | return -EBADF; |
2b188cc1 | 1626 | |
267bc904 | 1627 | ret = io_import_iovec(req->ctx, READ, &req->submit, &iovec, &iter); |
87e5e6da | 1628 | if (ret < 0) |
09bb8394 | 1629 | return ret; |
2b188cc1 | 1630 | |
9d93a3f5 | 1631 | read_size = ret; |
9e645e11 JA |
1632 | if (req->flags & REQ_F_LINK) |
1633 | req->result = read_size; | |
1634 | ||
31b51510 JA |
1635 | iov_count = iov_iter_count(&iter); |
1636 | ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count); | |
2b188cc1 JA |
1637 | if (!ret) { |
1638 | ssize_t ret2; | |
1639 | ||
32960613 JA |
1640 | if (file->f_op->read_iter) |
1641 | ret2 = call_read_iter(file, kiocb, &iter); | |
1642 | else | |
1643 | ret2 = loop_rw_iter(READ, file, kiocb, &iter); | |
1644 | ||
9d93a3f5 JA |
1645 | /* |
1646 | * In case of a short read, punt to async. This can happen | |
1647 | * if we have data partially cached. Alternatively we can | |
1648 | * return the short read, in which case the application will | |
1649 | * need to issue another SQE and wait for it. That SQE will | |
1650 | * need async punt anyway, so it's more efficient to do it | |
1651 | * here. | |
1652 | */ | |
491381ce JA |
1653 | if (force_nonblock && !(req->flags & REQ_F_NOWAIT) && |
1654 | (req->flags & REQ_F_ISREG) && | |
1655 | ret2 > 0 && ret2 < read_size) | |
9d93a3f5 JA |
1656 | ret2 = -EAGAIN; |
1657 | /* Catch -EAGAIN return for forced non-blocking submission */ | |
561fb04a | 1658 | if (!force_nonblock || ret2 != -EAGAIN) |
267bc904 | 1659 | kiocb_done(kiocb, ret2, nxt, req->submit.in_async); |
561fb04a | 1660 | else |
2b188cc1 JA |
1661 | ret = -EAGAIN; |
1662 | } | |
1663 | kfree(iovec); | |
2b188cc1 JA |
1664 | return ret; |
1665 | } | |
1666 | ||
267bc904 PB |
1667 | static int io_write(struct io_kiocb *req, struct io_kiocb **nxt, |
1668 | bool force_nonblock) | |
2b188cc1 JA |
1669 | { |
1670 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1671 | struct kiocb *kiocb = &req->rw; | |
1672 | struct iov_iter iter; | |
1673 | struct file *file; | |
31b51510 | 1674 | size_t iov_count; |
87e5e6da | 1675 | ssize_t ret; |
2b188cc1 | 1676 | |
267bc904 | 1677 | ret = io_prep_rw(req, force_nonblock); |
2b188cc1 JA |
1678 | if (ret) |
1679 | return ret; | |
2b188cc1 | 1680 | |
2b188cc1 JA |
1681 | file = kiocb->ki_filp; |
1682 | if (unlikely(!(file->f_mode & FMODE_WRITE))) | |
09bb8394 | 1683 | return -EBADF; |
2b188cc1 | 1684 | |
267bc904 | 1685 | ret = io_import_iovec(req->ctx, WRITE, &req->submit, &iovec, &iter); |
87e5e6da | 1686 | if (ret < 0) |
09bb8394 | 1687 | return ret; |
2b188cc1 | 1688 | |
9e645e11 JA |
1689 | if (req->flags & REQ_F_LINK) |
1690 | req->result = ret; | |
1691 | ||
31b51510 JA |
1692 | iov_count = iov_iter_count(&iter); |
1693 | ||
1694 | ret = -EAGAIN; | |
561fb04a | 1695 | if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT)) |
31b51510 | 1696 | goto out_free; |
31b51510 JA |
1697 | |
1698 | ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count); | |
2b188cc1 | 1699 | if (!ret) { |
9bf7933f RP |
1700 | ssize_t ret2; |
1701 | ||
2b188cc1 JA |
1702 | /* |
1703 | * Open-code file_start_write here to grab freeze protection, | |
1704 | * which will be released by another thread in | |
1705 | * io_complete_rw(). Fool lockdep by telling it the lock got | |
1706 | * released so that it doesn't complain about the held lock when | |
1707 | * we return to userspace. | |
1708 | */ | |
491381ce | 1709 | if (req->flags & REQ_F_ISREG) { |
2b188cc1 JA |
1710 | __sb_start_write(file_inode(file)->i_sb, |
1711 | SB_FREEZE_WRITE, true); | |
1712 | __sb_writers_release(file_inode(file)->i_sb, | |
1713 | SB_FREEZE_WRITE); | |
1714 | } | |
1715 | kiocb->ki_flags |= IOCB_WRITE; | |
9bf7933f | 1716 | |
32960613 JA |
1717 | if (file->f_op->write_iter) |
1718 | ret2 = call_write_iter(file, kiocb, &iter); | |
1719 | else | |
1720 | ret2 = loop_rw_iter(WRITE, file, kiocb, &iter); | |
561fb04a | 1721 | if (!force_nonblock || ret2 != -EAGAIN) |
267bc904 | 1722 | kiocb_done(kiocb, ret2, nxt, req->submit.in_async); |
561fb04a | 1723 | else |
9bf7933f | 1724 | ret = -EAGAIN; |
2b188cc1 | 1725 | } |
31b51510 | 1726 | out_free: |
2b188cc1 | 1727 | kfree(iovec); |
2b188cc1 JA |
1728 | return ret; |
1729 | } | |
1730 | ||
1731 | /* | |
1732 | * IORING_OP_NOP just posts a completion event, nothing else. | |
1733 | */ | |
78e19bbe | 1734 | static int io_nop(struct io_kiocb *req) |
2b188cc1 JA |
1735 | { |
1736 | struct io_ring_ctx *ctx = req->ctx; | |
2b188cc1 | 1737 | |
def596e9 JA |
1738 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
1739 | return -EINVAL; | |
1740 | ||
78e19bbe | 1741 | io_cqring_add_event(req, 0); |
ec9c02ad | 1742 | io_put_req(req); |
2b188cc1 JA |
1743 | return 0; |
1744 | } | |
1745 | ||
c992fe29 CH |
1746 | static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
1747 | { | |
6b06314c | 1748 | struct io_ring_ctx *ctx = req->ctx; |
c992fe29 | 1749 | |
09bb8394 JA |
1750 | if (!req->file) |
1751 | return -EBADF; | |
c992fe29 | 1752 | |
6b06314c | 1753 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
def596e9 | 1754 | return -EINVAL; |
edafccee | 1755 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) |
c992fe29 CH |
1756 | return -EINVAL; |
1757 | ||
c992fe29 CH |
1758 | return 0; |
1759 | } | |
1760 | ||
1761 | static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
ba816ad6 | 1762 | struct io_kiocb **nxt, bool force_nonblock) |
c992fe29 CH |
1763 | { |
1764 | loff_t sqe_off = READ_ONCE(sqe->off); | |
1765 | loff_t sqe_len = READ_ONCE(sqe->len); | |
1766 | loff_t end = sqe_off + sqe_len; | |
1767 | unsigned fsync_flags; | |
1768 | int ret; | |
1769 | ||
1770 | fsync_flags = READ_ONCE(sqe->fsync_flags); | |
1771 | if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC)) | |
1772 | return -EINVAL; | |
1773 | ||
1774 | ret = io_prep_fsync(req, sqe); | |
1775 | if (ret) | |
1776 | return ret; | |
1777 | ||
1778 | /* fsync always requires a blocking context */ | |
1779 | if (force_nonblock) | |
1780 | return -EAGAIN; | |
1781 | ||
1782 | ret = vfs_fsync_range(req->rw.ki_filp, sqe_off, | |
1783 | end > 0 ? end : LLONG_MAX, | |
1784 | fsync_flags & IORING_FSYNC_DATASYNC); | |
1785 | ||
9e645e11 JA |
1786 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1787 | req->flags |= REQ_F_FAIL_LINK; | |
78e19bbe | 1788 | io_cqring_add_event(req, ret); |
ec9c02ad | 1789 | io_put_req_find_next(req, nxt); |
c992fe29 CH |
1790 | return 0; |
1791 | } | |
1792 | ||
5d17b4a4 JA |
1793 | static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
1794 | { | |
1795 | struct io_ring_ctx *ctx = req->ctx; | |
1796 | int ret = 0; | |
1797 | ||
1798 | if (!req->file) | |
1799 | return -EBADF; | |
5d17b4a4 JA |
1800 | |
1801 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
1802 | return -EINVAL; | |
1803 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) | |
1804 | return -EINVAL; | |
1805 | ||
5d17b4a4 JA |
1806 | return ret; |
1807 | } | |
1808 | ||
1809 | static int io_sync_file_range(struct io_kiocb *req, | |
1810 | const struct io_uring_sqe *sqe, | |
ba816ad6 | 1811 | struct io_kiocb **nxt, |
5d17b4a4 JA |
1812 | bool force_nonblock) |
1813 | { | |
1814 | loff_t sqe_off; | |
1815 | loff_t sqe_len; | |
1816 | unsigned flags; | |
1817 | int ret; | |
1818 | ||
1819 | ret = io_prep_sfr(req, sqe); | |
1820 | if (ret) | |
1821 | return ret; | |
1822 | ||
1823 | /* sync_file_range always requires a blocking context */ | |
1824 | if (force_nonblock) | |
1825 | return -EAGAIN; | |
1826 | ||
1827 | sqe_off = READ_ONCE(sqe->off); | |
1828 | sqe_len = READ_ONCE(sqe->len); | |
1829 | flags = READ_ONCE(sqe->sync_range_flags); | |
1830 | ||
1831 | ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags); | |
1832 | ||
9e645e11 JA |
1833 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1834 | req->flags |= REQ_F_FAIL_LINK; | |
78e19bbe | 1835 | io_cqring_add_event(req, ret); |
ec9c02ad | 1836 | io_put_req_find_next(req, nxt); |
5d17b4a4 JA |
1837 | return 0; |
1838 | } | |
1839 | ||
0fa03c62 | 1840 | #if defined(CONFIG_NET) |
aa1fa28f | 1841 | static int io_send_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
ba816ad6 | 1842 | struct io_kiocb **nxt, bool force_nonblock, |
aa1fa28f JA |
1843 | long (*fn)(struct socket *, struct user_msghdr __user *, |
1844 | unsigned int)) | |
1845 | { | |
0fa03c62 JA |
1846 | struct socket *sock; |
1847 | int ret; | |
1848 | ||
1849 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1850 | return -EINVAL; | |
1851 | ||
1852 | sock = sock_from_file(req->file, &ret); | |
1853 | if (sock) { | |
1854 | struct user_msghdr __user *msg; | |
1855 | unsigned flags; | |
1856 | ||
1857 | flags = READ_ONCE(sqe->msg_flags); | |
1858 | if (flags & MSG_DONTWAIT) | |
1859 | req->flags |= REQ_F_NOWAIT; | |
1860 | else if (force_nonblock) | |
1861 | flags |= MSG_DONTWAIT; | |
1862 | ||
1863 | msg = (struct user_msghdr __user *) (unsigned long) | |
1864 | READ_ONCE(sqe->addr); | |
1865 | ||
aa1fa28f | 1866 | ret = fn(sock, msg, flags); |
0fa03c62 JA |
1867 | if (force_nonblock && ret == -EAGAIN) |
1868 | return ret; | |
1869 | } | |
1870 | ||
78e19bbe | 1871 | io_cqring_add_event(req, ret); |
f1f40853 JA |
1872 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1873 | req->flags |= REQ_F_FAIL_LINK; | |
ec9c02ad | 1874 | io_put_req_find_next(req, nxt); |
5d17b4a4 JA |
1875 | return 0; |
1876 | } | |
aa1fa28f JA |
1877 | #endif |
1878 | ||
1879 | static int io_sendmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
ba816ad6 | 1880 | struct io_kiocb **nxt, bool force_nonblock) |
aa1fa28f JA |
1881 | { |
1882 | #if defined(CONFIG_NET) | |
ba816ad6 JA |
1883 | return io_send_recvmsg(req, sqe, nxt, force_nonblock, |
1884 | __sys_sendmsg_sock); | |
aa1fa28f JA |
1885 | #else |
1886 | return -EOPNOTSUPP; | |
1887 | #endif | |
1888 | } | |
1889 | ||
1890 | static int io_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
ba816ad6 | 1891 | struct io_kiocb **nxt, bool force_nonblock) |
aa1fa28f JA |
1892 | { |
1893 | #if defined(CONFIG_NET) | |
ba816ad6 JA |
1894 | return io_send_recvmsg(req, sqe, nxt, force_nonblock, |
1895 | __sys_recvmsg_sock); | |
0fa03c62 JA |
1896 | #else |
1897 | return -EOPNOTSUPP; | |
1898 | #endif | |
1899 | } | |
5d17b4a4 | 1900 | |
17f2fe35 JA |
1901 | static int io_accept(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
1902 | struct io_kiocb **nxt, bool force_nonblock) | |
1903 | { | |
1904 | #if defined(CONFIG_NET) | |
1905 | struct sockaddr __user *addr; | |
1906 | int __user *addr_len; | |
1907 | unsigned file_flags; | |
1908 | int flags, ret; | |
1909 | ||
1910 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL))) | |
1911 | return -EINVAL; | |
1912 | if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index) | |
1913 | return -EINVAL; | |
1914 | ||
1915 | addr = (struct sockaddr __user *) (unsigned long) READ_ONCE(sqe->addr); | |
1916 | addr_len = (int __user *) (unsigned long) READ_ONCE(sqe->addr2); | |
1917 | flags = READ_ONCE(sqe->accept_flags); | |
1918 | file_flags = force_nonblock ? O_NONBLOCK : 0; | |
1919 | ||
1920 | ret = __sys_accept4_file(req->file, file_flags, addr, addr_len, flags); | |
1921 | if (ret == -EAGAIN && force_nonblock) { | |
1922 | req->work.flags |= IO_WQ_WORK_NEEDS_FILES; | |
1923 | return -EAGAIN; | |
1924 | } | |
8e3cca12 JA |
1925 | if (ret == -ERESTARTSYS) |
1926 | ret = -EINTR; | |
17f2fe35 JA |
1927 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1928 | req->flags |= REQ_F_FAIL_LINK; | |
78e19bbe | 1929 | io_cqring_add_event(req, ret); |
ec9c02ad | 1930 | io_put_req_find_next(req, nxt); |
17f2fe35 JA |
1931 | return 0; |
1932 | #else | |
1933 | return -EOPNOTSUPP; | |
1934 | #endif | |
1935 | } | |
1936 | ||
221c5eb2 JA |
1937 | static void io_poll_remove_one(struct io_kiocb *req) |
1938 | { | |
1939 | struct io_poll_iocb *poll = &req->poll; | |
1940 | ||
1941 | spin_lock(&poll->head->lock); | |
1942 | WRITE_ONCE(poll->canceled, true); | |
1943 | if (!list_empty(&poll->wait.entry)) { | |
1944 | list_del_init(&poll->wait.entry); | |
a197f664 | 1945 | io_queue_async_work(req); |
221c5eb2 JA |
1946 | } |
1947 | spin_unlock(&poll->head->lock); | |
1948 | ||
1949 | list_del_init(&req->list); | |
1950 | } | |
1951 | ||
1952 | static void io_poll_remove_all(struct io_ring_ctx *ctx) | |
1953 | { | |
1954 | struct io_kiocb *req; | |
1955 | ||
1956 | spin_lock_irq(&ctx->completion_lock); | |
1957 | while (!list_empty(&ctx->cancel_list)) { | |
1958 | req = list_first_entry(&ctx->cancel_list, struct io_kiocb,list); | |
1959 | io_poll_remove_one(req); | |
1960 | } | |
1961 | spin_unlock_irq(&ctx->completion_lock); | |
1962 | } | |
1963 | ||
47f46768 JA |
1964 | static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr) |
1965 | { | |
1966 | struct io_kiocb *req; | |
1967 | ||
1968 | list_for_each_entry(req, &ctx->cancel_list, list) { | |
1969 | if (req->user_data != sqe_addr) | |
1970 | continue; | |
1971 | io_poll_remove_one(req); | |
1972 | return 0; | |
1973 | } | |
1974 | ||
1975 | return -ENOENT; | |
1976 | } | |
1977 | ||
221c5eb2 JA |
1978 | /* |
1979 | * Find a running poll command that matches one specified in sqe->addr, | |
1980 | * and remove it if found. | |
1981 | */ | |
1982 | static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
1983 | { | |
1984 | struct io_ring_ctx *ctx = req->ctx; | |
47f46768 | 1985 | int ret; |
221c5eb2 JA |
1986 | |
1987 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1988 | return -EINVAL; | |
1989 | if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index || | |
1990 | sqe->poll_events) | |
1991 | return -EINVAL; | |
1992 | ||
1993 | spin_lock_irq(&ctx->completion_lock); | |
47f46768 | 1994 | ret = io_poll_cancel(ctx, READ_ONCE(sqe->addr)); |
221c5eb2 JA |
1995 | spin_unlock_irq(&ctx->completion_lock); |
1996 | ||
78e19bbe | 1997 | io_cqring_add_event(req, ret); |
f1f40853 JA |
1998 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1999 | req->flags |= REQ_F_FAIL_LINK; | |
ec9c02ad | 2000 | io_put_req(req); |
221c5eb2 JA |
2001 | return 0; |
2002 | } | |
2003 | ||
a197f664 | 2004 | static void io_poll_complete(struct io_kiocb *req, __poll_t mask) |
221c5eb2 | 2005 | { |
a197f664 JL |
2006 | struct io_ring_ctx *ctx = req->ctx; |
2007 | ||
8c838788 | 2008 | req->poll.done = true; |
78e19bbe | 2009 | io_cqring_fill_event(req, mangle_poll(mask)); |
8c838788 | 2010 | io_commit_cqring(ctx); |
221c5eb2 JA |
2011 | } |
2012 | ||
561fb04a | 2013 | static void io_poll_complete_work(struct io_wq_work **workptr) |
221c5eb2 | 2014 | { |
561fb04a | 2015 | struct io_wq_work *work = *workptr; |
221c5eb2 JA |
2016 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
2017 | struct io_poll_iocb *poll = &req->poll; | |
2018 | struct poll_table_struct pt = { ._key = poll->events }; | |
2019 | struct io_ring_ctx *ctx = req->ctx; | |
89723d0b | 2020 | struct io_kiocb *nxt = NULL; |
221c5eb2 JA |
2021 | __poll_t mask = 0; |
2022 | ||
561fb04a JA |
2023 | if (work->flags & IO_WQ_WORK_CANCEL) |
2024 | WRITE_ONCE(poll->canceled, true); | |
2025 | ||
221c5eb2 JA |
2026 | if (!READ_ONCE(poll->canceled)) |
2027 | mask = vfs_poll(poll->file, &pt) & poll->events; | |
2028 | ||
2029 | /* | |
2030 | * Note that ->ki_cancel callers also delete iocb from active_reqs after | |
2031 | * calling ->ki_cancel. We need the ctx_lock roundtrip here to | |
2032 | * synchronize with them. In the cancellation case the list_del_init | |
2033 | * itself is not actually needed, but harmless so we keep it in to | |
2034 | * avoid further branches in the fast path. | |
2035 | */ | |
2036 | spin_lock_irq(&ctx->completion_lock); | |
2037 | if (!mask && !READ_ONCE(poll->canceled)) { | |
2038 | add_wait_queue(poll->head, &poll->wait); | |
2039 | spin_unlock_irq(&ctx->completion_lock); | |
2040 | return; | |
2041 | } | |
2042 | list_del_init(&req->list); | |
a197f664 | 2043 | io_poll_complete(req, mask); |
221c5eb2 JA |
2044 | spin_unlock_irq(&ctx->completion_lock); |
2045 | ||
8c838788 | 2046 | io_cqring_ev_posted(ctx); |
89723d0b | 2047 | |
ec9c02ad | 2048 | io_put_req_find_next(req, &nxt); |
89723d0b JA |
2049 | if (nxt) |
2050 | *workptr = &nxt->work; | |
221c5eb2 JA |
2051 | } |
2052 | ||
2053 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, | |
2054 | void *key) | |
2055 | { | |
2056 | struct io_poll_iocb *poll = container_of(wait, struct io_poll_iocb, | |
2057 | wait); | |
2058 | struct io_kiocb *req = container_of(poll, struct io_kiocb, poll); | |
2059 | struct io_ring_ctx *ctx = req->ctx; | |
2060 | __poll_t mask = key_to_poll(key); | |
8c838788 | 2061 | unsigned long flags; |
221c5eb2 JA |
2062 | |
2063 | /* for instances that support it check for an event match first: */ | |
8c838788 JA |
2064 | if (mask && !(mask & poll->events)) |
2065 | return 0; | |
221c5eb2 | 2066 | |
8c838788 | 2067 | list_del_init(&poll->wait.entry); |
221c5eb2 | 2068 | |
7c9e7f0f JA |
2069 | /* |
2070 | * Run completion inline if we can. We're using trylock here because | |
2071 | * we are violating the completion_lock -> poll wq lock ordering. | |
2072 | * If we have a link timeout we're going to need the completion_lock | |
2073 | * for finalizing the request, mark us as having grabbed that already. | |
2074 | */ | |
8c838788 JA |
2075 | if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) { |
2076 | list_del(&req->list); | |
a197f664 | 2077 | io_poll_complete(req, mask); |
7c9e7f0f JA |
2078 | req->flags |= REQ_F_COMP_LOCKED; |
2079 | io_put_req(req); | |
8c838788 | 2080 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
221c5eb2 | 2081 | |
8c838788 | 2082 | io_cqring_ev_posted(ctx); |
8c838788 | 2083 | } else { |
a197f664 | 2084 | io_queue_async_work(req); |
221c5eb2 JA |
2085 | } |
2086 | ||
221c5eb2 JA |
2087 | return 1; |
2088 | } | |
2089 | ||
2090 | struct io_poll_table { | |
2091 | struct poll_table_struct pt; | |
2092 | struct io_kiocb *req; | |
2093 | int error; | |
2094 | }; | |
2095 | ||
2096 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, | |
2097 | struct poll_table_struct *p) | |
2098 | { | |
2099 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
2100 | ||
2101 | if (unlikely(pt->req->poll.head)) { | |
2102 | pt->error = -EINVAL; | |
2103 | return; | |
2104 | } | |
2105 | ||
2106 | pt->error = 0; | |
2107 | pt->req->poll.head = head; | |
2108 | add_wait_queue(head, &pt->req->poll.wait); | |
2109 | } | |
2110 | ||
89723d0b JA |
2111 | static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2112 | struct io_kiocb **nxt) | |
221c5eb2 JA |
2113 | { |
2114 | struct io_poll_iocb *poll = &req->poll; | |
2115 | struct io_ring_ctx *ctx = req->ctx; | |
2116 | struct io_poll_table ipt; | |
8c838788 | 2117 | bool cancel = false; |
221c5eb2 JA |
2118 | __poll_t mask; |
2119 | u16 events; | |
221c5eb2 JA |
2120 | |
2121 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
2122 | return -EINVAL; | |
2123 | if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index) | |
2124 | return -EINVAL; | |
09bb8394 JA |
2125 | if (!poll->file) |
2126 | return -EBADF; | |
221c5eb2 | 2127 | |
6cc47d1d | 2128 | req->submit.sqe = NULL; |
561fb04a | 2129 | INIT_IO_WORK(&req->work, io_poll_complete_work); |
221c5eb2 JA |
2130 | events = READ_ONCE(sqe->poll_events); |
2131 | poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP; | |
2132 | ||
221c5eb2 | 2133 | poll->head = NULL; |
8c838788 | 2134 | poll->done = false; |
221c5eb2 JA |
2135 | poll->canceled = false; |
2136 | ||
2137 | ipt.pt._qproc = io_poll_queue_proc; | |
2138 | ipt.pt._key = poll->events; | |
2139 | ipt.req = req; | |
2140 | ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */ | |
2141 | ||
2142 | /* initialized the list so that we can do list_empty checks */ | |
2143 | INIT_LIST_HEAD(&poll->wait.entry); | |
2144 | init_waitqueue_func_entry(&poll->wait, io_poll_wake); | |
2145 | ||
36703247 JA |
2146 | INIT_LIST_HEAD(&req->list); |
2147 | ||
221c5eb2 | 2148 | mask = vfs_poll(poll->file, &ipt.pt) & poll->events; |
221c5eb2 JA |
2149 | |
2150 | spin_lock_irq(&ctx->completion_lock); | |
8c838788 JA |
2151 | if (likely(poll->head)) { |
2152 | spin_lock(&poll->head->lock); | |
2153 | if (unlikely(list_empty(&poll->wait.entry))) { | |
2154 | if (ipt.error) | |
2155 | cancel = true; | |
2156 | ipt.error = 0; | |
2157 | mask = 0; | |
2158 | } | |
2159 | if (mask || ipt.error) | |
2160 | list_del_init(&poll->wait.entry); | |
2161 | else if (cancel) | |
2162 | WRITE_ONCE(poll->canceled, true); | |
2163 | else if (!poll->done) /* actually waiting for an event */ | |
2164 | list_add_tail(&req->list, &ctx->cancel_list); | |
2165 | spin_unlock(&poll->head->lock); | |
2166 | } | |
2167 | if (mask) { /* no async, we'd stolen it */ | |
221c5eb2 | 2168 | ipt.error = 0; |
a197f664 | 2169 | io_poll_complete(req, mask); |
221c5eb2 | 2170 | } |
221c5eb2 JA |
2171 | spin_unlock_irq(&ctx->completion_lock); |
2172 | ||
8c838788 JA |
2173 | if (mask) { |
2174 | io_cqring_ev_posted(ctx); | |
ec9c02ad | 2175 | io_put_req_find_next(req, nxt); |
221c5eb2 | 2176 | } |
8c838788 | 2177 | return ipt.error; |
221c5eb2 JA |
2178 | } |
2179 | ||
5262f567 JA |
2180 | static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer) |
2181 | { | |
2182 | struct io_ring_ctx *ctx; | |
11365043 | 2183 | struct io_kiocb *req; |
5262f567 JA |
2184 | unsigned long flags; |
2185 | ||
2186 | req = container_of(timer, struct io_kiocb, timeout.timer); | |
2187 | ctx = req->ctx; | |
2188 | atomic_inc(&ctx->cq_timeouts); | |
2189 | ||
2190 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
ef03681a | 2191 | /* |
11365043 JA |
2192 | * We could be racing with timeout deletion. If the list is empty, |
2193 | * then timeout lookup already found it and will be handling it. | |
ef03681a | 2194 | */ |
842f9612 | 2195 | if (!list_empty(&req->list)) { |
11365043 | 2196 | struct io_kiocb *prev; |
5262f567 | 2197 | |
11365043 JA |
2198 | /* |
2199 | * Adjust the reqs sequence before the current one because it | |
2200 | * will consume a slot in the cq_ring and the the cq_tail | |
2201 | * pointer will be increased, otherwise other timeout reqs may | |
2202 | * return in advance without waiting for enough wait_nr. | |
2203 | */ | |
2204 | prev = req; | |
2205 | list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list) | |
2206 | prev->sequence++; | |
11365043 | 2207 | list_del_init(&req->list); |
11365043 | 2208 | } |
842f9612 | 2209 | |
78e19bbe | 2210 | io_cqring_fill_event(req, -ETIME); |
842f9612 | 2211 | io_commit_cqring(ctx); |
5262f567 JA |
2212 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
2213 | ||
842f9612 | 2214 | io_cqring_ev_posted(ctx); |
f1f40853 JA |
2215 | if (req->flags & REQ_F_LINK) |
2216 | req->flags |= REQ_F_FAIL_LINK; | |
ec9c02ad | 2217 | io_put_req(req); |
11365043 JA |
2218 | return HRTIMER_NORESTART; |
2219 | } | |
2220 | ||
47f46768 JA |
2221 | static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data) |
2222 | { | |
2223 | struct io_kiocb *req; | |
2224 | int ret = -ENOENT; | |
2225 | ||
2226 | list_for_each_entry(req, &ctx->timeout_list, list) { | |
2227 | if (user_data == req->user_data) { | |
2228 | list_del_init(&req->list); | |
2229 | ret = 0; | |
2230 | break; | |
2231 | } | |
2232 | } | |
2233 | ||
2234 | if (ret == -ENOENT) | |
2235 | return ret; | |
2236 | ||
2237 | ret = hrtimer_try_to_cancel(&req->timeout.timer); | |
2238 | if (ret == -1) | |
2239 | return -EALREADY; | |
2240 | ||
2241 | io_cqring_fill_event(req, -ECANCELED); | |
2242 | io_put_req(req); | |
2243 | return 0; | |
2244 | } | |
2245 | ||
11365043 JA |
2246 | /* |
2247 | * Remove or update an existing timeout command | |
2248 | */ | |
2249 | static int io_timeout_remove(struct io_kiocb *req, | |
2250 | const struct io_uring_sqe *sqe) | |
2251 | { | |
2252 | struct io_ring_ctx *ctx = req->ctx; | |
11365043 | 2253 | unsigned flags; |
47f46768 | 2254 | int ret; |
11365043 JA |
2255 | |
2256 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
2257 | return -EINVAL; | |
2258 | if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len) | |
2259 | return -EINVAL; | |
2260 | flags = READ_ONCE(sqe->timeout_flags); | |
2261 | if (flags) | |
2262 | return -EINVAL; | |
2263 | ||
11365043 | 2264 | spin_lock_irq(&ctx->completion_lock); |
47f46768 | 2265 | ret = io_timeout_cancel(ctx, READ_ONCE(sqe->addr)); |
11365043 | 2266 | |
47f46768 | 2267 | io_cqring_fill_event(req, ret); |
11365043 JA |
2268 | io_commit_cqring(ctx); |
2269 | spin_unlock_irq(&ctx->completion_lock); | |
5262f567 | 2270 | io_cqring_ev_posted(ctx); |
47f46768 JA |
2271 | if (ret < 0 && req->flags & REQ_F_LINK) |
2272 | req->flags |= REQ_F_FAIL_LINK; | |
ec9c02ad | 2273 | io_put_req(req); |
11365043 | 2274 | return 0; |
5262f567 JA |
2275 | } |
2276 | ||
2277 | static int io_timeout(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
2278 | { | |
5da0fb1a | 2279 | unsigned count; |
5262f567 JA |
2280 | struct io_ring_ctx *ctx = req->ctx; |
2281 | struct list_head *entry; | |
a41525ab | 2282 | enum hrtimer_mode mode; |
bdf20073 | 2283 | struct timespec64 ts; |
a1f58ba4 | 2284 | unsigned span = 0; |
a41525ab | 2285 | unsigned flags; |
5262f567 JA |
2286 | |
2287 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
2288 | return -EINVAL; | |
a41525ab JA |
2289 | if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len != 1) |
2290 | return -EINVAL; | |
2291 | flags = READ_ONCE(sqe->timeout_flags); | |
2292 | if (flags & ~IORING_TIMEOUT_ABS) | |
5262f567 | 2293 | return -EINVAL; |
bdf20073 AB |
2294 | |
2295 | if (get_timespec64(&ts, u64_to_user_ptr(sqe->addr))) | |
5262f567 JA |
2296 | return -EFAULT; |
2297 | ||
11365043 JA |
2298 | if (flags & IORING_TIMEOUT_ABS) |
2299 | mode = HRTIMER_MODE_ABS; | |
2300 | else | |
2301 | mode = HRTIMER_MODE_REL; | |
2302 | ||
2303 | hrtimer_init(&req->timeout.timer, CLOCK_MONOTONIC, mode); | |
2304 | ||
5262f567 JA |
2305 | /* |
2306 | * sqe->off holds how many events that need to occur for this | |
2307 | * timeout event to be satisfied. | |
2308 | */ | |
2309 | count = READ_ONCE(sqe->off); | |
2310 | if (!count) | |
2311 | count = 1; | |
2312 | ||
2313 | req->sequence = ctx->cached_sq_head + count - 1; | |
5da0fb1a | 2314 | /* reuse it to store the count */ |
2315 | req->submit.sequence = count; | |
5262f567 JA |
2316 | req->flags |= REQ_F_TIMEOUT; |
2317 | ||
2318 | /* | |
2319 | * Insertion sort, ensuring the first entry in the list is always | |
2320 | * the one we need first. | |
2321 | */ | |
5262f567 JA |
2322 | spin_lock_irq(&ctx->completion_lock); |
2323 | list_for_each_prev(entry, &ctx->timeout_list) { | |
2324 | struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list); | |
5da0fb1a | 2325 | unsigned nxt_sq_head; |
2326 | long long tmp, tmp_nxt; | |
5262f567 | 2327 | |
5da0fb1a | 2328 | /* |
2329 | * Since cached_sq_head + count - 1 can overflow, use type long | |
2330 | * long to store it. | |
2331 | */ | |
2332 | tmp = (long long)ctx->cached_sq_head + count - 1; | |
2333 | nxt_sq_head = nxt->sequence - nxt->submit.sequence + 1; | |
2334 | tmp_nxt = (long long)nxt_sq_head + nxt->submit.sequence - 1; | |
2335 | ||
2336 | /* | |
2337 | * cached_sq_head may overflow, and it will never overflow twice | |
2338 | * once there is some timeout req still be valid. | |
2339 | */ | |
2340 | if (ctx->cached_sq_head < nxt_sq_head) | |
8b07a65a | 2341 | tmp += UINT_MAX; |
5da0fb1a | 2342 | |
a1f58ba4 | 2343 | if (tmp > tmp_nxt) |
5262f567 | 2344 | break; |
a1f58ba4 | 2345 | |
2346 | /* | |
2347 | * Sequence of reqs after the insert one and itself should | |
2348 | * be adjusted because each timeout req consumes a slot. | |
2349 | */ | |
2350 | span++; | |
2351 | nxt->sequence++; | |
5262f567 | 2352 | } |
a1f58ba4 | 2353 | req->sequence -= span; |
5262f567 | 2354 | list_add(&req->list, entry); |
5262f567 | 2355 | req->timeout.timer.function = io_timeout_fn; |
a41525ab | 2356 | hrtimer_start(&req->timeout.timer, timespec64_to_ktime(ts), mode); |
842f9612 | 2357 | spin_unlock_irq(&ctx->completion_lock); |
5262f567 JA |
2358 | return 0; |
2359 | } | |
2360 | ||
62755e35 JA |
2361 | static bool io_cancel_cb(struct io_wq_work *work, void *data) |
2362 | { | |
2363 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
2364 | ||
2365 | return req->user_data == (unsigned long) data; | |
2366 | } | |
2367 | ||
e977d6d3 | 2368 | static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr) |
62755e35 | 2369 | { |
62755e35 | 2370 | enum io_wq_cancel cancel_ret; |
62755e35 JA |
2371 | int ret = 0; |
2372 | ||
62755e35 JA |
2373 | cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr); |
2374 | switch (cancel_ret) { | |
2375 | case IO_WQ_CANCEL_OK: | |
2376 | ret = 0; | |
2377 | break; | |
2378 | case IO_WQ_CANCEL_RUNNING: | |
2379 | ret = -EALREADY; | |
2380 | break; | |
2381 | case IO_WQ_CANCEL_NOTFOUND: | |
2382 | ret = -ENOENT; | |
2383 | break; | |
2384 | } | |
2385 | ||
e977d6d3 JA |
2386 | return ret; |
2387 | } | |
2388 | ||
47f46768 JA |
2389 | static void io_async_find_and_cancel(struct io_ring_ctx *ctx, |
2390 | struct io_kiocb *req, __u64 sqe_addr, | |
2391 | struct io_kiocb **nxt) | |
2392 | { | |
2393 | unsigned long flags; | |
2394 | int ret; | |
2395 | ||
2396 | ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr); | |
2397 | if (ret != -ENOENT) { | |
2398 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
2399 | goto done; | |
2400 | } | |
2401 | ||
2402 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
2403 | ret = io_timeout_cancel(ctx, sqe_addr); | |
2404 | if (ret != -ENOENT) | |
2405 | goto done; | |
2406 | ret = io_poll_cancel(ctx, sqe_addr); | |
2407 | done: | |
2408 | io_cqring_fill_event(req, ret); | |
2409 | io_commit_cqring(ctx); | |
2410 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
2411 | io_cqring_ev_posted(ctx); | |
2412 | ||
2413 | if (ret < 0 && (req->flags & REQ_F_LINK)) | |
2414 | req->flags |= REQ_F_FAIL_LINK; | |
2415 | io_put_req_find_next(req, nxt); | |
2416 | } | |
2417 | ||
e977d6d3 JA |
2418 | static int io_async_cancel(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2419 | struct io_kiocb **nxt) | |
2420 | { | |
2421 | struct io_ring_ctx *ctx = req->ctx; | |
e977d6d3 JA |
2422 | |
2423 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
2424 | return -EINVAL; | |
2425 | if (sqe->flags || sqe->ioprio || sqe->off || sqe->len || | |
2426 | sqe->cancel_flags) | |
2427 | return -EINVAL; | |
2428 | ||
47f46768 | 2429 | io_async_find_and_cancel(ctx, req, READ_ONCE(sqe->addr), NULL); |
62755e35 JA |
2430 | return 0; |
2431 | } | |
2432 | ||
a197f664 | 2433 | static int io_req_defer(struct io_kiocb *req) |
de0617e4 | 2434 | { |
267bc904 | 2435 | const struct io_uring_sqe *sqe = req->submit.sqe; |
de0617e4 | 2436 | struct io_uring_sqe *sqe_copy; |
a197f664 | 2437 | struct io_ring_ctx *ctx = req->ctx; |
de0617e4 | 2438 | |
9d858b21 BL |
2439 | /* Still need defer if there is pending req in defer list. */ |
2440 | if (!req_need_defer(req) && list_empty(&ctx->defer_list)) | |
de0617e4 JA |
2441 | return 0; |
2442 | ||
2443 | sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL); | |
2444 | if (!sqe_copy) | |
2445 | return -EAGAIN; | |
2446 | ||
2447 | spin_lock_irq(&ctx->completion_lock); | |
9d858b21 | 2448 | if (!req_need_defer(req) && list_empty(&ctx->defer_list)) { |
de0617e4 JA |
2449 | spin_unlock_irq(&ctx->completion_lock); |
2450 | kfree(sqe_copy); | |
2451 | return 0; | |
2452 | } | |
2453 | ||
2454 | memcpy(sqe_copy, sqe, sizeof(*sqe_copy)); | |
2455 | req->submit.sqe = sqe_copy; | |
2456 | ||
c826bd7a | 2457 | trace_io_uring_defer(ctx, req, false); |
de0617e4 JA |
2458 | list_add_tail(&req->list, &ctx->defer_list); |
2459 | spin_unlock_irq(&ctx->completion_lock); | |
2460 | return -EIOCBQUEUED; | |
2461 | } | |
2462 | ||
a197f664 JL |
2463 | static int __io_submit_sqe(struct io_kiocb *req, struct io_kiocb **nxt, |
2464 | bool force_nonblock) | |
2b188cc1 | 2465 | { |
e0c5c576 | 2466 | int ret, opcode; |
267bc904 | 2467 | struct sqe_submit *s = &req->submit; |
a197f664 | 2468 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 | 2469 | |
2b188cc1 JA |
2470 | opcode = READ_ONCE(s->sqe->opcode); |
2471 | switch (opcode) { | |
2472 | case IORING_OP_NOP: | |
78e19bbe | 2473 | ret = io_nop(req); |
2b188cc1 JA |
2474 | break; |
2475 | case IORING_OP_READV: | |
edafccee JA |
2476 | if (unlikely(s->sqe->buf_index)) |
2477 | return -EINVAL; | |
267bc904 | 2478 | ret = io_read(req, nxt, force_nonblock); |
2b188cc1 JA |
2479 | break; |
2480 | case IORING_OP_WRITEV: | |
edafccee JA |
2481 | if (unlikely(s->sqe->buf_index)) |
2482 | return -EINVAL; | |
267bc904 | 2483 | ret = io_write(req, nxt, force_nonblock); |
edafccee JA |
2484 | break; |
2485 | case IORING_OP_READ_FIXED: | |
267bc904 | 2486 | ret = io_read(req, nxt, force_nonblock); |
edafccee JA |
2487 | break; |
2488 | case IORING_OP_WRITE_FIXED: | |
267bc904 | 2489 | ret = io_write(req, nxt, force_nonblock); |
2b188cc1 | 2490 | break; |
c992fe29 | 2491 | case IORING_OP_FSYNC: |
ba816ad6 | 2492 | ret = io_fsync(req, s->sqe, nxt, force_nonblock); |
c992fe29 | 2493 | break; |
221c5eb2 | 2494 | case IORING_OP_POLL_ADD: |
89723d0b | 2495 | ret = io_poll_add(req, s->sqe, nxt); |
221c5eb2 JA |
2496 | break; |
2497 | case IORING_OP_POLL_REMOVE: | |
2498 | ret = io_poll_remove(req, s->sqe); | |
2499 | break; | |
5d17b4a4 | 2500 | case IORING_OP_SYNC_FILE_RANGE: |
ba816ad6 | 2501 | ret = io_sync_file_range(req, s->sqe, nxt, force_nonblock); |
5d17b4a4 | 2502 | break; |
0fa03c62 | 2503 | case IORING_OP_SENDMSG: |
ba816ad6 | 2504 | ret = io_sendmsg(req, s->sqe, nxt, force_nonblock); |
0fa03c62 | 2505 | break; |
aa1fa28f | 2506 | case IORING_OP_RECVMSG: |
ba816ad6 | 2507 | ret = io_recvmsg(req, s->sqe, nxt, force_nonblock); |
aa1fa28f | 2508 | break; |
5262f567 JA |
2509 | case IORING_OP_TIMEOUT: |
2510 | ret = io_timeout(req, s->sqe); | |
2511 | break; | |
11365043 JA |
2512 | case IORING_OP_TIMEOUT_REMOVE: |
2513 | ret = io_timeout_remove(req, s->sqe); | |
2514 | break; | |
17f2fe35 JA |
2515 | case IORING_OP_ACCEPT: |
2516 | ret = io_accept(req, s->sqe, nxt, force_nonblock); | |
2517 | break; | |
62755e35 JA |
2518 | case IORING_OP_ASYNC_CANCEL: |
2519 | ret = io_async_cancel(req, s->sqe, nxt); | |
2520 | break; | |
2b188cc1 JA |
2521 | default: |
2522 | ret = -EINVAL; | |
2523 | break; | |
2524 | } | |
2525 | ||
def596e9 JA |
2526 | if (ret) |
2527 | return ret; | |
2528 | ||
2529 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
9e645e11 | 2530 | if (req->result == -EAGAIN) |
def596e9 JA |
2531 | return -EAGAIN; |
2532 | ||
2533 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
ba5290cc | 2534 | if (s->in_async) |
def596e9 JA |
2535 | mutex_lock(&ctx->uring_lock); |
2536 | io_iopoll_req_issued(req); | |
ba5290cc | 2537 | if (s->in_async) |
def596e9 JA |
2538 | mutex_unlock(&ctx->uring_lock); |
2539 | } | |
2540 | ||
2541 | return 0; | |
2b188cc1 JA |
2542 | } |
2543 | ||
561fb04a | 2544 | static void io_wq_submit_work(struct io_wq_work **workptr) |
2b188cc1 | 2545 | { |
561fb04a | 2546 | struct io_wq_work *work = *workptr; |
2b188cc1 | 2547 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
561fb04a JA |
2548 | struct sqe_submit *s = &req->submit; |
2549 | const struct io_uring_sqe *sqe = s->sqe; | |
2550 | struct io_kiocb *nxt = NULL; | |
2551 | int ret = 0; | |
2b188cc1 | 2552 | |
561fb04a JA |
2553 | /* Ensure we clear previously set non-block flag */ |
2554 | req->rw.ki_flags &= ~IOCB_NOWAIT; | |
2b188cc1 | 2555 | |
561fb04a JA |
2556 | if (work->flags & IO_WQ_WORK_CANCEL) |
2557 | ret = -ECANCELED; | |
31b51510 | 2558 | |
561fb04a JA |
2559 | if (!ret) { |
2560 | s->has_user = (work->flags & IO_WQ_WORK_HAS_MM) != 0; | |
2561 | s->in_async = true; | |
2562 | do { | |
a197f664 | 2563 | ret = __io_submit_sqe(req, &nxt, false); |
561fb04a JA |
2564 | /* |
2565 | * We can get EAGAIN for polled IO even though we're | |
2566 | * forcing a sync submission from here, since we can't | |
2567 | * wait for request slots on the block side. | |
2568 | */ | |
2569 | if (ret != -EAGAIN) | |
2570 | break; | |
2571 | cond_resched(); | |
2572 | } while (1); | |
2573 | } | |
31b51510 | 2574 | |
561fb04a | 2575 | /* drop submission reference */ |
ec9c02ad | 2576 | io_put_req(req); |
817869d2 | 2577 | |
561fb04a | 2578 | if (ret) { |
f1f40853 JA |
2579 | if (req->flags & REQ_F_LINK) |
2580 | req->flags |= REQ_F_FAIL_LINK; | |
78e19bbe | 2581 | io_cqring_add_event(req, ret); |
ec9c02ad | 2582 | io_put_req(req); |
2b188cc1 | 2583 | } |
31b51510 | 2584 | |
561fb04a JA |
2585 | /* async context always use a copy of the sqe */ |
2586 | kfree(sqe); | |
31b51510 | 2587 | |
561fb04a JA |
2588 | /* if a dependent link is ready, pass it back */ |
2589 | if (!ret && nxt) { | |
2590 | io_prep_async_work(nxt); | |
2591 | *workptr = &nxt->work; | |
31b51510 | 2592 | } |
2b188cc1 JA |
2593 | } |
2594 | ||
09bb8394 JA |
2595 | static bool io_op_needs_file(const struct io_uring_sqe *sqe) |
2596 | { | |
2597 | int op = READ_ONCE(sqe->opcode); | |
2598 | ||
2599 | switch (op) { | |
2600 | case IORING_OP_NOP: | |
2601 | case IORING_OP_POLL_REMOVE: | |
a320e9fa PB |
2602 | case IORING_OP_TIMEOUT: |
2603 | case IORING_OP_TIMEOUT_REMOVE: | |
2604 | case IORING_OP_ASYNC_CANCEL: | |
2605 | case IORING_OP_LINK_TIMEOUT: | |
09bb8394 JA |
2606 | return false; |
2607 | default: | |
2608 | return true; | |
2609 | } | |
2610 | } | |
2611 | ||
65e19f54 JA |
2612 | static inline struct file *io_file_from_index(struct io_ring_ctx *ctx, |
2613 | int index) | |
2614 | { | |
2615 | struct fixed_file_table *table; | |
2616 | ||
2617 | table = &ctx->file_table[index >> IORING_FILE_TABLE_SHIFT]; | |
2618 | return table->files[index & IORING_FILE_TABLE_MASK]; | |
2619 | } | |
2620 | ||
a197f664 | 2621 | static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req) |
09bb8394 | 2622 | { |
267bc904 | 2623 | struct sqe_submit *s = &req->submit; |
a197f664 | 2624 | struct io_ring_ctx *ctx = req->ctx; |
09bb8394 JA |
2625 | unsigned flags; |
2626 | int fd; | |
2627 | ||
2628 | flags = READ_ONCE(s->sqe->flags); | |
2629 | fd = READ_ONCE(s->sqe->fd); | |
2630 | ||
4fe2c963 | 2631 | if (flags & IOSQE_IO_DRAIN) |
de0617e4 | 2632 | req->flags |= REQ_F_IO_DRAIN; |
4fe2c963 JL |
2633 | /* |
2634 | * All io need record the previous position, if LINK vs DARIN, | |
2635 | * it can be used to mark the position of the first IO in the | |
2636 | * link list. | |
2637 | */ | |
2638 | req->sequence = s->sequence; | |
de0617e4 | 2639 | |
60c112b0 | 2640 | if (!io_op_needs_file(s->sqe)) |
09bb8394 | 2641 | return 0; |
09bb8394 JA |
2642 | |
2643 | if (flags & IOSQE_FIXED_FILE) { | |
65e19f54 | 2644 | if (unlikely(!ctx->file_table || |
09bb8394 JA |
2645 | (unsigned) fd >= ctx->nr_user_files)) |
2646 | return -EBADF; | |
b7620121 | 2647 | fd = array_index_nospec(fd, ctx->nr_user_files); |
65e19f54 JA |
2648 | req->file = io_file_from_index(ctx, fd); |
2649 | if (!req->file) | |
08a45173 | 2650 | return -EBADF; |
09bb8394 JA |
2651 | req->flags |= REQ_F_FIXED_FILE; |
2652 | } else { | |
2653 | if (s->needs_fixed_file) | |
2654 | return -EBADF; | |
c826bd7a | 2655 | trace_io_uring_file_get(ctx, fd); |
09bb8394 JA |
2656 | req->file = io_file_get(state, fd); |
2657 | if (unlikely(!req->file)) | |
2658 | return -EBADF; | |
2659 | } | |
2660 | ||
2661 | return 0; | |
2662 | } | |
2663 | ||
a197f664 | 2664 | static int io_grab_files(struct io_kiocb *req) |
fcb323cc JA |
2665 | { |
2666 | int ret = -EBADF; | |
a197f664 | 2667 | struct io_ring_ctx *ctx = req->ctx; |
fcb323cc JA |
2668 | |
2669 | rcu_read_lock(); | |
2670 | spin_lock_irq(&ctx->inflight_lock); | |
2671 | /* | |
2672 | * We use the f_ops->flush() handler to ensure that we can flush | |
2673 | * out work accessing these files if the fd is closed. Check if | |
2674 | * the fd has changed since we started down this path, and disallow | |
2675 | * this operation if it has. | |
2676 | */ | |
2677 | if (fcheck(req->submit.ring_fd) == req->submit.ring_file) { | |
2678 | list_add(&req->inflight_entry, &ctx->inflight_list); | |
2679 | req->flags |= REQ_F_INFLIGHT; | |
2680 | req->work.files = current->files; | |
2681 | ret = 0; | |
2682 | } | |
2683 | spin_unlock_irq(&ctx->inflight_lock); | |
2684 | rcu_read_unlock(); | |
2685 | ||
2686 | return ret; | |
2687 | } | |
2688 | ||
2665abfd JA |
2689 | static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer) |
2690 | { | |
2691 | struct io_kiocb *req = container_of(timer, struct io_kiocb, | |
2692 | timeout.timer); | |
2693 | struct io_ring_ctx *ctx = req->ctx; | |
2694 | struct io_kiocb *prev = NULL; | |
2695 | unsigned long flags; | |
2665abfd JA |
2696 | |
2697 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
2698 | ||
2699 | /* | |
2700 | * We don't expect the list to be empty, that will only happen if we | |
2701 | * race with the completion of the linked work. | |
2702 | */ | |
2703 | if (!list_empty(&req->list)) { | |
2704 | prev = list_entry(req->list.prev, struct io_kiocb, link_list); | |
76a46e06 JA |
2705 | if (refcount_inc_not_zero(&prev->refs)) |
2706 | list_del_init(&req->list); | |
2707 | else | |
2708 | prev = NULL; | |
2665abfd JA |
2709 | } |
2710 | ||
2711 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
2712 | ||
2713 | if (prev) { | |
47f46768 | 2714 | io_async_find_and_cancel(ctx, req, prev->user_data, NULL); |
76a46e06 | 2715 | io_put_req(prev); |
47f46768 JA |
2716 | } else { |
2717 | io_cqring_add_event(req, -ETIME); | |
2718 | io_put_req(req); | |
2665abfd | 2719 | } |
2665abfd JA |
2720 | return HRTIMER_NORESTART; |
2721 | } | |
2722 | ||
76a46e06 JA |
2723 | static void io_queue_linked_timeout(struct io_kiocb *req, struct timespec64 *ts, |
2724 | enum hrtimer_mode *mode) | |
2665abfd | 2725 | { |
76a46e06 | 2726 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd | 2727 | |
76a46e06 JA |
2728 | /* |
2729 | * If the list is now empty, then our linked request finished before | |
2730 | * we got a chance to setup the timer | |
2731 | */ | |
2732 | spin_lock_irq(&ctx->completion_lock); | |
2733 | if (!list_empty(&req->list)) { | |
2734 | req->timeout.timer.function = io_link_timeout_fn; | |
2735 | hrtimer_start(&req->timeout.timer, timespec64_to_ktime(*ts), | |
2736 | *mode); | |
2665abfd | 2737 | } |
76a46e06 | 2738 | spin_unlock_irq(&ctx->completion_lock); |
2665abfd | 2739 | |
2665abfd | 2740 | /* drop submission reference */ |
76a46e06 JA |
2741 | io_put_req(req); |
2742 | } | |
2665abfd | 2743 | |
76a46e06 JA |
2744 | static int io_validate_link_timeout(const struct io_uring_sqe *sqe, |
2745 | struct timespec64 *ts) | |
2746 | { | |
2747 | if (sqe->ioprio || sqe->buf_index || sqe->len != 1 || sqe->off) | |
2748 | return -EINVAL; | |
2749 | if (sqe->timeout_flags & ~IORING_TIMEOUT_ABS) | |
2750 | return -EINVAL; | |
2751 | if (get_timespec64(ts, u64_to_user_ptr(sqe->addr))) | |
2752 | return -EFAULT; | |
2665abfd | 2753 | |
76a46e06 | 2754 | return 0; |
2665abfd JA |
2755 | } |
2756 | ||
76a46e06 JA |
2757 | static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req, |
2758 | struct timespec64 *ts, | |
2759 | enum hrtimer_mode *mode) | |
2665abfd JA |
2760 | { |
2761 | struct io_kiocb *nxt; | |
76a46e06 | 2762 | int ret; |
2665abfd JA |
2763 | |
2764 | if (!(req->flags & REQ_F_LINK)) | |
2765 | return NULL; | |
2766 | ||
2767 | nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list); | |
76a46e06 JA |
2768 | if (!nxt || nxt->submit.sqe->opcode != IORING_OP_LINK_TIMEOUT) |
2769 | return NULL; | |
2665abfd | 2770 | |
76a46e06 JA |
2771 | ret = io_validate_link_timeout(nxt->submit.sqe, ts); |
2772 | if (ret) { | |
2773 | list_del_init(&nxt->list); | |
2774 | io_cqring_add_event(nxt, ret); | |
2775 | io_double_put_req(nxt); | |
2776 | return ERR_PTR(-ECANCELED); | |
2777 | } | |
2778 | ||
2779 | if (nxt->submit.sqe->timeout_flags & IORING_TIMEOUT_ABS) | |
2780 | *mode = HRTIMER_MODE_ABS; | |
2781 | else | |
2782 | *mode = HRTIMER_MODE_REL; | |
2783 | ||
2784 | req->flags |= REQ_F_LINK_TIMEOUT; | |
2785 | hrtimer_init(&nxt->timeout.timer, CLOCK_MONOTONIC, *mode); | |
2786 | return nxt; | |
2665abfd JA |
2787 | } |
2788 | ||
a197f664 | 2789 | static int __io_queue_sqe(struct io_kiocb *req) |
2b188cc1 | 2790 | { |
76a46e06 | 2791 | enum hrtimer_mode mode; |
2665abfd | 2792 | struct io_kiocb *nxt; |
76a46e06 | 2793 | struct timespec64 ts; |
e0c5c576 | 2794 | int ret; |
2b188cc1 | 2795 | |
76a46e06 JA |
2796 | nxt = io_prep_linked_timeout(req, &ts, &mode); |
2797 | if (IS_ERR(nxt)) { | |
2798 | ret = PTR_ERR(nxt); | |
2799 | nxt = NULL; | |
2800 | goto err; | |
2665abfd JA |
2801 | } |
2802 | ||
a197f664 | 2803 | ret = __io_submit_sqe(req, NULL, true); |
491381ce JA |
2804 | |
2805 | /* | |
2806 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
2807 | * doesn't support non-blocking read/write attempts | |
2808 | */ | |
2809 | if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) || | |
2810 | (req->flags & REQ_F_MUST_PUNT))) { | |
267bc904 | 2811 | struct sqe_submit *s = &req->submit; |
2b188cc1 JA |
2812 | struct io_uring_sqe *sqe_copy; |
2813 | ||
954dab19 | 2814 | sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL); |
2b188cc1 | 2815 | if (sqe_copy) { |
2b188cc1 | 2816 | s->sqe = sqe_copy; |
fcb323cc | 2817 | if (req->work.flags & IO_WQ_WORK_NEEDS_FILES) { |
a197f664 | 2818 | ret = io_grab_files(req); |
fcb323cc JA |
2819 | if (ret) { |
2820 | kfree(sqe_copy); | |
2821 | goto err; | |
2822 | } | |
2823 | } | |
e65ef56d JA |
2824 | |
2825 | /* | |
2826 | * Queued up for async execution, worker will release | |
9e645e11 | 2827 | * submit reference when the iocb is actually submitted. |
e65ef56d | 2828 | */ |
a197f664 | 2829 | io_queue_async_work(req); |
76a46e06 JA |
2830 | |
2831 | if (nxt) | |
2832 | io_queue_linked_timeout(nxt, &ts, &mode); | |
2833 | ||
e65ef56d | 2834 | return 0; |
2b188cc1 JA |
2835 | } |
2836 | } | |
e65ef56d | 2837 | |
fcb323cc | 2838 | err: |
76a46e06 | 2839 | /* drop submission reference */ |
ec9c02ad | 2840 | io_put_req(req); |
e65ef56d | 2841 | |
76a46e06 JA |
2842 | if (nxt) { |
2843 | if (!ret) | |
2844 | io_queue_linked_timeout(nxt, &ts, &mode); | |
2845 | else | |
2846 | io_put_req(nxt); | |
2847 | } | |
2848 | ||
e65ef56d | 2849 | /* and drop final reference, if we failed */ |
9e645e11 | 2850 | if (ret) { |
78e19bbe | 2851 | io_cqring_add_event(req, ret); |
9e645e11 JA |
2852 | if (req->flags & REQ_F_LINK) |
2853 | req->flags |= REQ_F_FAIL_LINK; | |
ec9c02ad | 2854 | io_put_req(req); |
9e645e11 | 2855 | } |
2b188cc1 JA |
2856 | |
2857 | return ret; | |
2858 | } | |
2859 | ||
a197f664 | 2860 | static int io_queue_sqe(struct io_kiocb *req) |
4fe2c963 JL |
2861 | { |
2862 | int ret; | |
2863 | ||
a197f664 | 2864 | ret = io_req_defer(req); |
4fe2c963 JL |
2865 | if (ret) { |
2866 | if (ret != -EIOCBQUEUED) { | |
78e19bbe JA |
2867 | io_cqring_add_event(req, ret); |
2868 | io_double_put_req(req); | |
4fe2c963 JL |
2869 | } |
2870 | return 0; | |
2871 | } | |
2872 | ||
a197f664 | 2873 | return __io_queue_sqe(req); |
4fe2c963 JL |
2874 | } |
2875 | ||
a197f664 | 2876 | static int io_queue_link_head(struct io_kiocb *req, struct io_kiocb *shadow) |
4fe2c963 JL |
2877 | { |
2878 | int ret; | |
2879 | int need_submit = false; | |
a197f664 | 2880 | struct io_ring_ctx *ctx = req->ctx; |
4fe2c963 JL |
2881 | |
2882 | if (!shadow) | |
a197f664 | 2883 | return io_queue_sqe(req); |
4fe2c963 JL |
2884 | |
2885 | /* | |
2886 | * Mark the first IO in link list as DRAIN, let all the following | |
2887 | * IOs enter the defer list. all IO needs to be completed before link | |
2888 | * list. | |
2889 | */ | |
2890 | req->flags |= REQ_F_IO_DRAIN; | |
a197f664 | 2891 | ret = io_req_defer(req); |
4fe2c963 JL |
2892 | if (ret) { |
2893 | if (ret != -EIOCBQUEUED) { | |
78e19bbe JA |
2894 | io_cqring_add_event(req, ret); |
2895 | io_double_put_req(req); | |
7b20238d | 2896 | __io_free_req(shadow); |
4fe2c963 JL |
2897 | return 0; |
2898 | } | |
2899 | } else { | |
2900 | /* | |
2901 | * If ret == 0 means that all IOs in front of link io are | |
2902 | * running done. let's queue link head. | |
2903 | */ | |
2904 | need_submit = true; | |
2905 | } | |
2906 | ||
2907 | /* Insert shadow req to defer_list, blocking next IOs */ | |
2908 | spin_lock_irq(&ctx->completion_lock); | |
c826bd7a | 2909 | trace_io_uring_defer(ctx, shadow, true); |
4fe2c963 JL |
2910 | list_add_tail(&shadow->list, &ctx->defer_list); |
2911 | spin_unlock_irq(&ctx->completion_lock); | |
2912 | ||
2913 | if (need_submit) | |
a197f664 | 2914 | return __io_queue_sqe(req); |
4fe2c963 JL |
2915 | |
2916 | return 0; | |
2917 | } | |
2918 | ||
9e645e11 JA |
2919 | #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK) |
2920 | ||
a197f664 JL |
2921 | static void io_submit_sqe(struct io_kiocb *req, struct io_submit_state *state, |
2922 | struct io_kiocb **link) | |
9e645e11 JA |
2923 | { |
2924 | struct io_uring_sqe *sqe_copy; | |
267bc904 | 2925 | struct sqe_submit *s = &req->submit; |
a197f664 | 2926 | struct io_ring_ctx *ctx = req->ctx; |
9e645e11 JA |
2927 | int ret; |
2928 | ||
78e19bbe JA |
2929 | req->user_data = s->sqe->user_data; |
2930 | ||
9e645e11 JA |
2931 | /* enforce forwards compatibility on users */ |
2932 | if (unlikely(s->sqe->flags & ~SQE_VALID_FLAGS)) { | |
2933 | ret = -EINVAL; | |
196be95c | 2934 | goto err_req; |
9e645e11 JA |
2935 | } |
2936 | ||
a197f664 | 2937 | ret = io_req_set_file(state, req); |
9e645e11 JA |
2938 | if (unlikely(ret)) { |
2939 | err_req: | |
78e19bbe JA |
2940 | io_cqring_add_event(req, ret); |
2941 | io_double_put_req(req); | |
9e645e11 JA |
2942 | return; |
2943 | } | |
2944 | ||
9e645e11 JA |
2945 | /* |
2946 | * If we already have a head request, queue this one for async | |
2947 | * submittal once the head completes. If we don't have a head but | |
2948 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
2949 | * submitted sync once the chain is complete. If none of those | |
2950 | * conditions are true (normal request), then just queue it. | |
2951 | */ | |
2952 | if (*link) { | |
2953 | struct io_kiocb *prev = *link; | |
2954 | ||
2955 | sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL); | |
2956 | if (!sqe_copy) { | |
2957 | ret = -EAGAIN; | |
2958 | goto err_req; | |
2959 | } | |
2960 | ||
2961 | s->sqe = sqe_copy; | |
c826bd7a | 2962 | trace_io_uring_link(ctx, req, prev); |
9e645e11 JA |
2963 | list_add_tail(&req->list, &prev->link_list); |
2964 | } else if (s->sqe->flags & IOSQE_IO_LINK) { | |
2965 | req->flags |= REQ_F_LINK; | |
2966 | ||
9e645e11 JA |
2967 | INIT_LIST_HEAD(&req->link_list); |
2968 | *link = req; | |
2665abfd JA |
2969 | } else if (READ_ONCE(s->sqe->opcode) == IORING_OP_LINK_TIMEOUT) { |
2970 | /* Only valid as a linked SQE */ | |
2971 | ret = -EINVAL; | |
2972 | goto err_req; | |
9e645e11 | 2973 | } else { |
a197f664 | 2974 | io_queue_sqe(req); |
9e645e11 JA |
2975 | } |
2976 | } | |
2977 | ||
9a56a232 JA |
2978 | /* |
2979 | * Batched submission is done, ensure local IO is flushed out. | |
2980 | */ | |
2981 | static void io_submit_state_end(struct io_submit_state *state) | |
2982 | { | |
2983 | blk_finish_plug(&state->plug); | |
3d6770fb | 2984 | io_file_put(state); |
2579f913 JA |
2985 | if (state->free_reqs) |
2986 | kmem_cache_free_bulk(req_cachep, state->free_reqs, | |
2987 | &state->reqs[state->cur_req]); | |
9a56a232 JA |
2988 | } |
2989 | ||
2990 | /* | |
2991 | * Start submission side cache. | |
2992 | */ | |
2993 | static void io_submit_state_start(struct io_submit_state *state, | |
2994 | struct io_ring_ctx *ctx, unsigned max_ios) | |
2995 | { | |
2996 | blk_start_plug(&state->plug); | |
2579f913 | 2997 | state->free_reqs = 0; |
9a56a232 JA |
2998 | state->file = NULL; |
2999 | state->ios_left = max_ios; | |
3000 | } | |
3001 | ||
2b188cc1 JA |
3002 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
3003 | { | |
75b28aff | 3004 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 3005 | |
75b28aff | 3006 | if (ctx->cached_sq_head != READ_ONCE(rings->sq.head)) { |
2b188cc1 JA |
3007 | /* |
3008 | * Ensure any loads from the SQEs are done at this point, | |
3009 | * since once we write the new head, the application could | |
3010 | * write new data to them. | |
3011 | */ | |
75b28aff | 3012 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); |
2b188cc1 JA |
3013 | } |
3014 | } | |
3015 | ||
2b188cc1 JA |
3016 | /* |
3017 | * Fetch an sqe, if one is available. Note that s->sqe will point to memory | |
3018 | * that is mapped by userspace. This means that care needs to be taken to | |
3019 | * ensure that reads are stable, as we cannot rely on userspace always | |
3020 | * being a good citizen. If members of the sqe are validated and then later | |
3021 | * used, it's important that those reads are done through READ_ONCE() to | |
3022 | * prevent a re-load down the line. | |
3023 | */ | |
3024 | static bool io_get_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s) | |
3025 | { | |
75b28aff HV |
3026 | struct io_rings *rings = ctx->rings; |
3027 | u32 *sq_array = ctx->sq_array; | |
2b188cc1 JA |
3028 | unsigned head; |
3029 | ||
3030 | /* | |
3031 | * The cached sq head (or cq tail) serves two purposes: | |
3032 | * | |
3033 | * 1) allows us to batch the cost of updating the user visible | |
3034 | * head updates. | |
3035 | * 2) allows the kernel side to track the head on its own, even | |
3036 | * though the application is the one updating it. | |
3037 | */ | |
3038 | head = ctx->cached_sq_head; | |
e523a29c | 3039 | /* make sure SQ entry isn't read before tail */ |
75b28aff | 3040 | if (head == smp_load_acquire(&rings->sq.tail)) |
2b188cc1 JA |
3041 | return false; |
3042 | ||
75b28aff | 3043 | head = READ_ONCE(sq_array[head & ctx->sq_mask]); |
2b188cc1 | 3044 | if (head < ctx->sq_entries) { |
fcb323cc | 3045 | s->ring_file = NULL; |
2b188cc1 | 3046 | s->sqe = &ctx->sq_sqes[head]; |
8776f3fa | 3047 | s->sequence = ctx->cached_sq_head; |
2b188cc1 JA |
3048 | ctx->cached_sq_head++; |
3049 | return true; | |
3050 | } | |
3051 | ||
3052 | /* drop invalid entries */ | |
3053 | ctx->cached_sq_head++; | |
498ccd9e JA |
3054 | ctx->cached_sq_dropped++; |
3055 | WRITE_ONCE(rings->sq_dropped, ctx->cached_sq_dropped); | |
2b188cc1 JA |
3056 | return false; |
3057 | } | |
3058 | ||
fb5ccc98 | 3059 | static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr, |
ae9428ca PB |
3060 | struct file *ring_file, int ring_fd, |
3061 | struct mm_struct **mm, bool async) | |
6c271ce2 JA |
3062 | { |
3063 | struct io_submit_state state, *statep = NULL; | |
9e645e11 | 3064 | struct io_kiocb *link = NULL; |
4fe2c963 | 3065 | struct io_kiocb *shadow_req = NULL; |
9e645e11 | 3066 | int i, submitted = 0; |
95a1b3ff | 3067 | bool mm_fault = false; |
6c271ce2 | 3068 | |
1d7bb1d5 JA |
3069 | if (!list_empty(&ctx->cq_overflow_list)) { |
3070 | io_cqring_overflow_flush(ctx, false); | |
3071 | return -EBUSY; | |
3072 | } | |
3073 | ||
6c271ce2 JA |
3074 | if (nr > IO_PLUG_THRESHOLD) { |
3075 | io_submit_state_start(&state, ctx, nr); | |
3076 | statep = &state; | |
3077 | } | |
3078 | ||
3079 | for (i = 0; i < nr; i++) { | |
196be95c | 3080 | struct io_kiocb *req; |
50585b9a | 3081 | unsigned int sqe_flags; |
fb5ccc98 | 3082 | |
196be95c PB |
3083 | req = io_get_req(ctx, statep); |
3084 | if (unlikely(!req)) { | |
3085 | if (!submitted) | |
3086 | submitted = -EAGAIN; | |
fb5ccc98 | 3087 | break; |
196be95c | 3088 | } |
50585b9a | 3089 | if (!io_get_sqring(ctx, &req->submit)) { |
196be95c PB |
3090 | __io_free_req(req); |
3091 | break; | |
3092 | } | |
fb5ccc98 | 3093 | |
50585b9a | 3094 | if (io_sqe_needs_user(req->submit.sqe) && !*mm) { |
95a1b3ff PB |
3095 | mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm); |
3096 | if (!mm_fault) { | |
3097 | use_mm(ctx->sqo_mm); | |
3098 | *mm = ctx->sqo_mm; | |
3099 | } | |
3100 | } | |
3101 | ||
50585b9a PB |
3102 | sqe_flags = req->submit.sqe->flags; |
3103 | ||
3104 | if (link && (sqe_flags & IOSQE_IO_DRAIN)) { | |
4fe2c963 JL |
3105 | if (!shadow_req) { |
3106 | shadow_req = io_get_req(ctx, NULL); | |
a1041c27 JL |
3107 | if (unlikely(!shadow_req)) |
3108 | goto out; | |
4fe2c963 JL |
3109 | shadow_req->flags |= (REQ_F_IO_DRAIN | REQ_F_SHADOW_DRAIN); |
3110 | refcount_dec(&shadow_req->refs); | |
3111 | } | |
50585b9a | 3112 | shadow_req->sequence = req->submit.sequence; |
4fe2c963 JL |
3113 | } |
3114 | ||
a1041c27 | 3115 | out: |
50585b9a PB |
3116 | req->submit.ring_file = ring_file; |
3117 | req->submit.ring_fd = ring_fd; | |
3118 | req->submit.has_user = *mm != NULL; | |
3119 | req->submit.in_async = async; | |
3120 | req->submit.needs_fixed_file = async; | |
3121 | trace_io_uring_submit_sqe(ctx, req->submit.sqe->user_data, | |
3122 | true, async); | |
a197f664 | 3123 | io_submit_sqe(req, statep, &link); |
95a1b3ff | 3124 | submitted++; |
e5eb6366 PB |
3125 | |
3126 | /* | |
3127 | * If previous wasn't linked and we have a linked command, | |
3128 | * that's the end of the chain. Submit the previous link. | |
3129 | */ | |
50585b9a | 3130 | if (!(sqe_flags & IOSQE_IO_LINK) && link) { |
a197f664 | 3131 | io_queue_link_head(link, shadow_req); |
e5eb6366 PB |
3132 | link = NULL; |
3133 | shadow_req = NULL; | |
3134 | } | |
6c271ce2 JA |
3135 | } |
3136 | ||
9e645e11 | 3137 | if (link) |
a197f664 | 3138 | io_queue_link_head(link, shadow_req); |
6c271ce2 JA |
3139 | if (statep) |
3140 | io_submit_state_end(&state); | |
3141 | ||
ae9428ca PB |
3142 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
3143 | io_commit_sqring(ctx); | |
3144 | ||
6c271ce2 JA |
3145 | return submitted; |
3146 | } | |
3147 | ||
3148 | static int io_sq_thread(void *data) | |
3149 | { | |
6c271ce2 JA |
3150 | struct io_ring_ctx *ctx = data; |
3151 | struct mm_struct *cur_mm = NULL; | |
3152 | mm_segment_t old_fs; | |
3153 | DEFINE_WAIT(wait); | |
3154 | unsigned inflight; | |
3155 | unsigned long timeout; | |
c1edbf5f | 3156 | int ret; |
6c271ce2 | 3157 | |
206aefde | 3158 | complete(&ctx->completions[1]); |
a4c0b3de | 3159 | |
6c271ce2 JA |
3160 | old_fs = get_fs(); |
3161 | set_fs(USER_DS); | |
3162 | ||
c1edbf5f | 3163 | ret = timeout = inflight = 0; |
2bbcd6d3 | 3164 | while (!kthread_should_park()) { |
fb5ccc98 | 3165 | unsigned int to_submit; |
6c271ce2 JA |
3166 | |
3167 | if (inflight) { | |
3168 | unsigned nr_events = 0; | |
3169 | ||
3170 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
2b2ed975 JA |
3171 | /* |
3172 | * inflight is the count of the maximum possible | |
3173 | * entries we submitted, but it can be smaller | |
3174 | * if we dropped some of them. If we don't have | |
3175 | * poll entries available, then we know that we | |
3176 | * have nothing left to poll for. Reset the | |
3177 | * inflight count to zero in that case. | |
3178 | */ | |
3179 | mutex_lock(&ctx->uring_lock); | |
3180 | if (!list_empty(&ctx->poll_list)) | |
3181 | __io_iopoll_check(ctx, &nr_events, 0); | |
3182 | else | |
3183 | inflight = 0; | |
3184 | mutex_unlock(&ctx->uring_lock); | |
6c271ce2 JA |
3185 | } else { |
3186 | /* | |
3187 | * Normal IO, just pretend everything completed. | |
3188 | * We don't have to poll completions for that. | |
3189 | */ | |
3190 | nr_events = inflight; | |
3191 | } | |
3192 | ||
3193 | inflight -= nr_events; | |
3194 | if (!inflight) | |
3195 | timeout = jiffies + ctx->sq_thread_idle; | |
3196 | } | |
3197 | ||
fb5ccc98 | 3198 | to_submit = io_sqring_entries(ctx); |
c1edbf5f JA |
3199 | |
3200 | /* | |
3201 | * If submit got -EBUSY, flag us as needing the application | |
3202 | * to enter the kernel to reap and flush events. | |
3203 | */ | |
3204 | if (!to_submit || ret == -EBUSY) { | |
6c271ce2 JA |
3205 | /* |
3206 | * We're polling. If we're within the defined idle | |
3207 | * period, then let us spin without work before going | |
c1edbf5f JA |
3208 | * to sleep. The exception is if we got EBUSY doing |
3209 | * more IO, we should wait for the application to | |
3210 | * reap events and wake us up. | |
6c271ce2 | 3211 | */ |
c1edbf5f JA |
3212 | if (inflight || |
3213 | (!time_after(jiffies, timeout) && ret != -EBUSY)) { | |
9831a90c | 3214 | cond_resched(); |
6c271ce2 JA |
3215 | continue; |
3216 | } | |
3217 | ||
3218 | /* | |
3219 | * Drop cur_mm before scheduling, we can't hold it for | |
3220 | * long periods (or over schedule()). Do this before | |
3221 | * adding ourselves to the waitqueue, as the unuse/drop | |
3222 | * may sleep. | |
3223 | */ | |
3224 | if (cur_mm) { | |
3225 | unuse_mm(cur_mm); | |
3226 | mmput(cur_mm); | |
3227 | cur_mm = NULL; | |
3228 | } | |
3229 | ||
3230 | prepare_to_wait(&ctx->sqo_wait, &wait, | |
3231 | TASK_INTERRUPTIBLE); | |
3232 | ||
3233 | /* Tell userspace we may need a wakeup call */ | |
75b28aff | 3234 | ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP; |
0d7bae69 SB |
3235 | /* make sure to read SQ tail after writing flags */ |
3236 | smp_mb(); | |
6c271ce2 | 3237 | |
fb5ccc98 | 3238 | to_submit = io_sqring_entries(ctx); |
c1edbf5f | 3239 | if (!to_submit || ret == -EBUSY) { |
2bbcd6d3 | 3240 | if (kthread_should_park()) { |
6c271ce2 JA |
3241 | finish_wait(&ctx->sqo_wait, &wait); |
3242 | break; | |
3243 | } | |
3244 | if (signal_pending(current)) | |
3245 | flush_signals(current); | |
3246 | schedule(); | |
3247 | finish_wait(&ctx->sqo_wait, &wait); | |
3248 | ||
75b28aff | 3249 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
3250 | continue; |
3251 | } | |
3252 | finish_wait(&ctx->sqo_wait, &wait); | |
3253 | ||
75b28aff | 3254 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
3255 | } |
3256 | ||
fb5ccc98 | 3257 | to_submit = min(to_submit, ctx->sq_entries); |
1d7bb1d5 JA |
3258 | ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true); |
3259 | if (ret > 0) | |
3260 | inflight += ret; | |
6c271ce2 JA |
3261 | } |
3262 | ||
3263 | set_fs(old_fs); | |
3264 | if (cur_mm) { | |
3265 | unuse_mm(cur_mm); | |
3266 | mmput(cur_mm); | |
3267 | } | |
06058632 | 3268 | |
2bbcd6d3 | 3269 | kthread_parkme(); |
06058632 | 3270 | |
6c271ce2 JA |
3271 | return 0; |
3272 | } | |
3273 | ||
bda52162 JA |
3274 | struct io_wait_queue { |
3275 | struct wait_queue_entry wq; | |
3276 | struct io_ring_ctx *ctx; | |
3277 | unsigned to_wait; | |
3278 | unsigned nr_timeouts; | |
3279 | }; | |
3280 | ||
1d7bb1d5 | 3281 | static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush) |
bda52162 JA |
3282 | { |
3283 | struct io_ring_ctx *ctx = iowq->ctx; | |
3284 | ||
3285 | /* | |
3286 | * Wake up if we have enough events, or if a timeout occured since we | |
3287 | * started waiting. For timeouts, we always want to return to userspace, | |
3288 | * regardless of event count. | |
3289 | */ | |
1d7bb1d5 | 3290 | return io_cqring_events(ctx, noflush) >= iowq->to_wait || |
bda52162 JA |
3291 | atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
3292 | } | |
3293 | ||
3294 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
3295 | int wake_flags, void *key) | |
3296 | { | |
3297 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, | |
3298 | wq); | |
3299 | ||
1d7bb1d5 JA |
3300 | /* use noflush == true, as we can't safely rely on locking context */ |
3301 | if (!io_should_wake(iowq, true)) | |
bda52162 JA |
3302 | return -1; |
3303 | ||
3304 | return autoremove_wake_function(curr, mode, wake_flags, key); | |
3305 | } | |
3306 | ||
2b188cc1 JA |
3307 | /* |
3308 | * Wait until events become available, if we don't already have some. The | |
3309 | * application must reap them itself, as they reside on the shared cq ring. | |
3310 | */ | |
3311 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
3312 | const sigset_t __user *sig, size_t sigsz) | |
3313 | { | |
bda52162 JA |
3314 | struct io_wait_queue iowq = { |
3315 | .wq = { | |
3316 | .private = current, | |
3317 | .func = io_wake_function, | |
3318 | .entry = LIST_HEAD_INIT(iowq.wq.entry), | |
3319 | }, | |
3320 | .ctx = ctx, | |
3321 | .to_wait = min_events, | |
3322 | }; | |
75b28aff | 3323 | struct io_rings *rings = ctx->rings; |
e9ffa5c2 | 3324 | int ret = 0; |
2b188cc1 | 3325 | |
1d7bb1d5 | 3326 | if (io_cqring_events(ctx, false) >= min_events) |
2b188cc1 JA |
3327 | return 0; |
3328 | ||
3329 | if (sig) { | |
9e75ad5d AB |
3330 | #ifdef CONFIG_COMPAT |
3331 | if (in_compat_syscall()) | |
3332 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 3333 | sigsz); |
9e75ad5d AB |
3334 | else |
3335 | #endif | |
b772434b | 3336 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 3337 | |
2b188cc1 JA |
3338 | if (ret) |
3339 | return ret; | |
3340 | } | |
3341 | ||
bda52162 | 3342 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
c826bd7a | 3343 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 JA |
3344 | do { |
3345 | prepare_to_wait_exclusive(&ctx->wait, &iowq.wq, | |
3346 | TASK_INTERRUPTIBLE); | |
1d7bb1d5 | 3347 | if (io_should_wake(&iowq, false)) |
bda52162 JA |
3348 | break; |
3349 | schedule(); | |
3350 | if (signal_pending(current)) { | |
e9ffa5c2 | 3351 | ret = -EINTR; |
bda52162 JA |
3352 | break; |
3353 | } | |
3354 | } while (1); | |
3355 | finish_wait(&ctx->wait, &iowq.wq); | |
3356 | ||
e9ffa5c2 | 3357 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 3358 | |
75b28aff | 3359 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
3360 | } |
3361 | ||
6b06314c JA |
3362 | static void __io_sqe_files_unregister(struct io_ring_ctx *ctx) |
3363 | { | |
3364 | #if defined(CONFIG_UNIX) | |
3365 | if (ctx->ring_sock) { | |
3366 | struct sock *sock = ctx->ring_sock->sk; | |
3367 | struct sk_buff *skb; | |
3368 | ||
3369 | while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) | |
3370 | kfree_skb(skb); | |
3371 | } | |
3372 | #else | |
3373 | int i; | |
3374 | ||
65e19f54 JA |
3375 | for (i = 0; i < ctx->nr_user_files; i++) { |
3376 | struct file *file; | |
3377 | ||
3378 | file = io_file_from_index(ctx, i); | |
3379 | if (file) | |
3380 | fput(file); | |
3381 | } | |
6b06314c JA |
3382 | #endif |
3383 | } | |
3384 | ||
3385 | static int io_sqe_files_unregister(struct io_ring_ctx *ctx) | |
3386 | { | |
65e19f54 JA |
3387 | unsigned nr_tables, i; |
3388 | ||
3389 | if (!ctx->file_table) | |
6b06314c JA |
3390 | return -ENXIO; |
3391 | ||
3392 | __io_sqe_files_unregister(ctx); | |
65e19f54 JA |
3393 | nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE); |
3394 | for (i = 0; i < nr_tables; i++) | |
3395 | kfree(ctx->file_table[i].files); | |
3396 | kfree(ctx->file_table); | |
3397 | ctx->file_table = NULL; | |
6b06314c JA |
3398 | ctx->nr_user_files = 0; |
3399 | return 0; | |
3400 | } | |
3401 | ||
6c271ce2 JA |
3402 | static void io_sq_thread_stop(struct io_ring_ctx *ctx) |
3403 | { | |
3404 | if (ctx->sqo_thread) { | |
206aefde | 3405 | wait_for_completion(&ctx->completions[1]); |
2bbcd6d3 RP |
3406 | /* |
3407 | * The park is a bit of a work-around, without it we get | |
3408 | * warning spews on shutdown with SQPOLL set and affinity | |
3409 | * set to a single CPU. | |
3410 | */ | |
06058632 | 3411 | kthread_park(ctx->sqo_thread); |
6c271ce2 JA |
3412 | kthread_stop(ctx->sqo_thread); |
3413 | ctx->sqo_thread = NULL; | |
3414 | } | |
3415 | } | |
3416 | ||
6b06314c JA |
3417 | static void io_finish_async(struct io_ring_ctx *ctx) |
3418 | { | |
6c271ce2 JA |
3419 | io_sq_thread_stop(ctx); |
3420 | ||
561fb04a JA |
3421 | if (ctx->io_wq) { |
3422 | io_wq_destroy(ctx->io_wq); | |
3423 | ctx->io_wq = NULL; | |
6b06314c JA |
3424 | } |
3425 | } | |
3426 | ||
3427 | #if defined(CONFIG_UNIX) | |
3428 | static void io_destruct_skb(struct sk_buff *skb) | |
3429 | { | |
3430 | struct io_ring_ctx *ctx = skb->sk->sk_user_data; | |
8a997340 | 3431 | |
561fb04a JA |
3432 | if (ctx->io_wq) |
3433 | io_wq_flush(ctx->io_wq); | |
6b06314c | 3434 | |
6b06314c JA |
3435 | unix_destruct_scm(skb); |
3436 | } | |
3437 | ||
3438 | /* | |
3439 | * Ensure the UNIX gc is aware of our file set, so we are certain that | |
3440 | * the io_uring can be safely unregistered on process exit, even if we have | |
3441 | * loops in the file referencing. | |
3442 | */ | |
3443 | static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset) | |
3444 | { | |
3445 | struct sock *sk = ctx->ring_sock->sk; | |
3446 | struct scm_fp_list *fpl; | |
3447 | struct sk_buff *skb; | |
08a45173 | 3448 | int i, nr_files; |
6b06314c JA |
3449 | |
3450 | if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) { | |
3451 | unsigned long inflight = ctx->user->unix_inflight + nr; | |
3452 | ||
3453 | if (inflight > task_rlimit(current, RLIMIT_NOFILE)) | |
3454 | return -EMFILE; | |
3455 | } | |
3456 | ||
3457 | fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); | |
3458 | if (!fpl) | |
3459 | return -ENOMEM; | |
3460 | ||
3461 | skb = alloc_skb(0, GFP_KERNEL); | |
3462 | if (!skb) { | |
3463 | kfree(fpl); | |
3464 | return -ENOMEM; | |
3465 | } | |
3466 | ||
3467 | skb->sk = sk; | |
6b06314c | 3468 | |
08a45173 | 3469 | nr_files = 0; |
6b06314c JA |
3470 | fpl->user = get_uid(ctx->user); |
3471 | for (i = 0; i < nr; i++) { | |
65e19f54 JA |
3472 | struct file *file = io_file_from_index(ctx, i + offset); |
3473 | ||
3474 | if (!file) | |
08a45173 | 3475 | continue; |
65e19f54 | 3476 | fpl->fp[nr_files] = get_file(file); |
08a45173 JA |
3477 | unix_inflight(fpl->user, fpl->fp[nr_files]); |
3478 | nr_files++; | |
6b06314c JA |
3479 | } |
3480 | ||
08a45173 JA |
3481 | if (nr_files) { |
3482 | fpl->max = SCM_MAX_FD; | |
3483 | fpl->count = nr_files; | |
3484 | UNIXCB(skb).fp = fpl; | |
3485 | skb->destructor = io_destruct_skb; | |
3486 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); | |
3487 | skb_queue_head(&sk->sk_receive_queue, skb); | |
6b06314c | 3488 | |
08a45173 JA |
3489 | for (i = 0; i < nr_files; i++) |
3490 | fput(fpl->fp[i]); | |
3491 | } else { | |
3492 | kfree_skb(skb); | |
3493 | kfree(fpl); | |
3494 | } | |
6b06314c JA |
3495 | |
3496 | return 0; | |
3497 | } | |
3498 | ||
3499 | /* | |
3500 | * If UNIX sockets are enabled, fd passing can cause a reference cycle which | |
3501 | * causes regular reference counting to break down. We rely on the UNIX | |
3502 | * garbage collection to take care of this problem for us. | |
3503 | */ | |
3504 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
3505 | { | |
3506 | unsigned left, total; | |
3507 | int ret = 0; | |
3508 | ||
3509 | total = 0; | |
3510 | left = ctx->nr_user_files; | |
3511 | while (left) { | |
3512 | unsigned this_files = min_t(unsigned, left, SCM_MAX_FD); | |
6b06314c JA |
3513 | |
3514 | ret = __io_sqe_files_scm(ctx, this_files, total); | |
3515 | if (ret) | |
3516 | break; | |
3517 | left -= this_files; | |
3518 | total += this_files; | |
3519 | } | |
3520 | ||
3521 | if (!ret) | |
3522 | return 0; | |
3523 | ||
3524 | while (total < ctx->nr_user_files) { | |
65e19f54 JA |
3525 | struct file *file = io_file_from_index(ctx, total); |
3526 | ||
3527 | if (file) | |
3528 | fput(file); | |
6b06314c JA |
3529 | total++; |
3530 | } | |
3531 | ||
3532 | return ret; | |
3533 | } | |
3534 | #else | |
3535 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
3536 | { | |
3537 | return 0; | |
3538 | } | |
3539 | #endif | |
3540 | ||
65e19f54 JA |
3541 | static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables, |
3542 | unsigned nr_files) | |
3543 | { | |
3544 | int i; | |
3545 | ||
3546 | for (i = 0; i < nr_tables; i++) { | |
3547 | struct fixed_file_table *table = &ctx->file_table[i]; | |
3548 | unsigned this_files; | |
3549 | ||
3550 | this_files = min(nr_files, IORING_MAX_FILES_TABLE); | |
3551 | table->files = kcalloc(this_files, sizeof(struct file *), | |
3552 | GFP_KERNEL); | |
3553 | if (!table->files) | |
3554 | break; | |
3555 | nr_files -= this_files; | |
3556 | } | |
3557 | ||
3558 | if (i == nr_tables) | |
3559 | return 0; | |
3560 | ||
3561 | for (i = 0; i < nr_tables; i++) { | |
3562 | struct fixed_file_table *table = &ctx->file_table[i]; | |
3563 | kfree(table->files); | |
3564 | } | |
3565 | return 1; | |
3566 | } | |
3567 | ||
6b06314c JA |
3568 | static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, |
3569 | unsigned nr_args) | |
3570 | { | |
3571 | __s32 __user *fds = (__s32 __user *) arg; | |
65e19f54 | 3572 | unsigned nr_tables; |
6b06314c JA |
3573 | int fd, ret = 0; |
3574 | unsigned i; | |
3575 | ||
65e19f54 | 3576 | if (ctx->file_table) |
6b06314c JA |
3577 | return -EBUSY; |
3578 | if (!nr_args) | |
3579 | return -EINVAL; | |
3580 | if (nr_args > IORING_MAX_FIXED_FILES) | |
3581 | return -EMFILE; | |
3582 | ||
65e19f54 JA |
3583 | nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE); |
3584 | ctx->file_table = kcalloc(nr_tables, sizeof(struct fixed_file_table), | |
3585 | GFP_KERNEL); | |
3586 | if (!ctx->file_table) | |
6b06314c JA |
3587 | return -ENOMEM; |
3588 | ||
65e19f54 JA |
3589 | if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) { |
3590 | kfree(ctx->file_table); | |
46568e9b | 3591 | ctx->file_table = NULL; |
65e19f54 JA |
3592 | return -ENOMEM; |
3593 | } | |
3594 | ||
08a45173 | 3595 | for (i = 0; i < nr_args; i++, ctx->nr_user_files++) { |
65e19f54 JA |
3596 | struct fixed_file_table *table; |
3597 | unsigned index; | |
3598 | ||
6b06314c JA |
3599 | ret = -EFAULT; |
3600 | if (copy_from_user(&fd, &fds[i], sizeof(fd))) | |
3601 | break; | |
08a45173 JA |
3602 | /* allow sparse sets */ |
3603 | if (fd == -1) { | |
3604 | ret = 0; | |
3605 | continue; | |
3606 | } | |
6b06314c | 3607 | |
65e19f54 JA |
3608 | table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT]; |
3609 | index = i & IORING_FILE_TABLE_MASK; | |
3610 | table->files[index] = fget(fd); | |
6b06314c JA |
3611 | |
3612 | ret = -EBADF; | |
65e19f54 | 3613 | if (!table->files[index]) |
6b06314c JA |
3614 | break; |
3615 | /* | |
3616 | * Don't allow io_uring instances to be registered. If UNIX | |
3617 | * isn't enabled, then this causes a reference cycle and this | |
3618 | * instance can never get freed. If UNIX is enabled we'll | |
3619 | * handle it just fine, but there's still no point in allowing | |
3620 | * a ring fd as it doesn't support regular read/write anyway. | |
3621 | */ | |
65e19f54 JA |
3622 | if (table->files[index]->f_op == &io_uring_fops) { |
3623 | fput(table->files[index]); | |
6b06314c JA |
3624 | break; |
3625 | } | |
6b06314c JA |
3626 | ret = 0; |
3627 | } | |
3628 | ||
3629 | if (ret) { | |
65e19f54 JA |
3630 | for (i = 0; i < ctx->nr_user_files; i++) { |
3631 | struct file *file; | |
6b06314c | 3632 | |
65e19f54 JA |
3633 | file = io_file_from_index(ctx, i); |
3634 | if (file) | |
3635 | fput(file); | |
3636 | } | |
3637 | for (i = 0; i < nr_tables; i++) | |
3638 | kfree(ctx->file_table[i].files); | |
3639 | ||
3640 | kfree(ctx->file_table); | |
3641 | ctx->file_table = NULL; | |
6b06314c JA |
3642 | ctx->nr_user_files = 0; |
3643 | return ret; | |
3644 | } | |
3645 | ||
3646 | ret = io_sqe_files_scm(ctx); | |
3647 | if (ret) | |
3648 | io_sqe_files_unregister(ctx); | |
3649 | ||
3650 | return ret; | |
3651 | } | |
3652 | ||
c3a31e60 JA |
3653 | static void io_sqe_file_unregister(struct io_ring_ctx *ctx, int index) |
3654 | { | |
3655 | #if defined(CONFIG_UNIX) | |
65e19f54 | 3656 | struct file *file = io_file_from_index(ctx, index); |
c3a31e60 JA |
3657 | struct sock *sock = ctx->ring_sock->sk; |
3658 | struct sk_buff_head list, *head = &sock->sk_receive_queue; | |
3659 | struct sk_buff *skb; | |
3660 | int i; | |
3661 | ||
3662 | __skb_queue_head_init(&list); | |
3663 | ||
3664 | /* | |
3665 | * Find the skb that holds this file in its SCM_RIGHTS. When found, | |
3666 | * remove this entry and rearrange the file array. | |
3667 | */ | |
3668 | skb = skb_dequeue(head); | |
3669 | while (skb) { | |
3670 | struct scm_fp_list *fp; | |
3671 | ||
3672 | fp = UNIXCB(skb).fp; | |
3673 | for (i = 0; i < fp->count; i++) { | |
3674 | int left; | |
3675 | ||
3676 | if (fp->fp[i] != file) | |
3677 | continue; | |
3678 | ||
3679 | unix_notinflight(fp->user, fp->fp[i]); | |
3680 | left = fp->count - 1 - i; | |
3681 | if (left) { | |
3682 | memmove(&fp->fp[i], &fp->fp[i + 1], | |
3683 | left * sizeof(struct file *)); | |
3684 | } | |
3685 | fp->count--; | |
3686 | if (!fp->count) { | |
3687 | kfree_skb(skb); | |
3688 | skb = NULL; | |
3689 | } else { | |
3690 | __skb_queue_tail(&list, skb); | |
3691 | } | |
3692 | fput(file); | |
3693 | file = NULL; | |
3694 | break; | |
3695 | } | |
3696 | ||
3697 | if (!file) | |
3698 | break; | |
3699 | ||
3700 | __skb_queue_tail(&list, skb); | |
3701 | ||
3702 | skb = skb_dequeue(head); | |
3703 | } | |
3704 | ||
3705 | if (skb_peek(&list)) { | |
3706 | spin_lock_irq(&head->lock); | |
3707 | while ((skb = __skb_dequeue(&list)) != NULL) | |
3708 | __skb_queue_tail(head, skb); | |
3709 | spin_unlock_irq(&head->lock); | |
3710 | } | |
3711 | #else | |
65e19f54 | 3712 | fput(io_file_from_index(ctx, index)); |
c3a31e60 JA |
3713 | #endif |
3714 | } | |
3715 | ||
3716 | static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file, | |
3717 | int index) | |
3718 | { | |
3719 | #if defined(CONFIG_UNIX) | |
3720 | struct sock *sock = ctx->ring_sock->sk; | |
3721 | struct sk_buff_head *head = &sock->sk_receive_queue; | |
3722 | struct sk_buff *skb; | |
3723 | ||
3724 | /* | |
3725 | * See if we can merge this file into an existing skb SCM_RIGHTS | |
3726 | * file set. If there's no room, fall back to allocating a new skb | |
3727 | * and filling it in. | |
3728 | */ | |
3729 | spin_lock_irq(&head->lock); | |
3730 | skb = skb_peek(head); | |
3731 | if (skb) { | |
3732 | struct scm_fp_list *fpl = UNIXCB(skb).fp; | |
3733 | ||
3734 | if (fpl->count < SCM_MAX_FD) { | |
3735 | __skb_unlink(skb, head); | |
3736 | spin_unlock_irq(&head->lock); | |
3737 | fpl->fp[fpl->count] = get_file(file); | |
3738 | unix_inflight(fpl->user, fpl->fp[fpl->count]); | |
3739 | fpl->count++; | |
3740 | spin_lock_irq(&head->lock); | |
3741 | __skb_queue_head(head, skb); | |
3742 | } else { | |
3743 | skb = NULL; | |
3744 | } | |
3745 | } | |
3746 | spin_unlock_irq(&head->lock); | |
3747 | ||
3748 | if (skb) { | |
3749 | fput(file); | |
3750 | return 0; | |
3751 | } | |
3752 | ||
3753 | return __io_sqe_files_scm(ctx, 1, index); | |
3754 | #else | |
3755 | return 0; | |
3756 | #endif | |
3757 | } | |
3758 | ||
3759 | static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg, | |
3760 | unsigned nr_args) | |
3761 | { | |
3762 | struct io_uring_files_update up; | |
3763 | __s32 __user *fds; | |
3764 | int fd, i, err; | |
3765 | __u32 done; | |
3766 | ||
65e19f54 | 3767 | if (!ctx->file_table) |
c3a31e60 JA |
3768 | return -ENXIO; |
3769 | if (!nr_args) | |
3770 | return -EINVAL; | |
3771 | if (copy_from_user(&up, arg, sizeof(up))) | |
3772 | return -EFAULT; | |
3773 | if (check_add_overflow(up.offset, nr_args, &done)) | |
3774 | return -EOVERFLOW; | |
3775 | if (done > ctx->nr_user_files) | |
3776 | return -EINVAL; | |
3777 | ||
3778 | done = 0; | |
3779 | fds = (__s32 __user *) up.fds; | |
3780 | while (nr_args) { | |
65e19f54 JA |
3781 | struct fixed_file_table *table; |
3782 | unsigned index; | |
3783 | ||
c3a31e60 JA |
3784 | err = 0; |
3785 | if (copy_from_user(&fd, &fds[done], sizeof(fd))) { | |
3786 | err = -EFAULT; | |
3787 | break; | |
3788 | } | |
3789 | i = array_index_nospec(up.offset, ctx->nr_user_files); | |
65e19f54 JA |
3790 | table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT]; |
3791 | index = i & IORING_FILE_TABLE_MASK; | |
3792 | if (table->files[index]) { | |
c3a31e60 | 3793 | io_sqe_file_unregister(ctx, i); |
65e19f54 | 3794 | table->files[index] = NULL; |
c3a31e60 JA |
3795 | } |
3796 | if (fd != -1) { | |
3797 | struct file *file; | |
3798 | ||
3799 | file = fget(fd); | |
3800 | if (!file) { | |
3801 | err = -EBADF; | |
3802 | break; | |
3803 | } | |
3804 | /* | |
3805 | * Don't allow io_uring instances to be registered. If | |
3806 | * UNIX isn't enabled, then this causes a reference | |
3807 | * cycle and this instance can never get freed. If UNIX | |
3808 | * is enabled we'll handle it just fine, but there's | |
3809 | * still no point in allowing a ring fd as it doesn't | |
3810 | * support regular read/write anyway. | |
3811 | */ | |
3812 | if (file->f_op == &io_uring_fops) { | |
3813 | fput(file); | |
3814 | err = -EBADF; | |
3815 | break; | |
3816 | } | |
65e19f54 | 3817 | table->files[index] = file; |
c3a31e60 JA |
3818 | err = io_sqe_file_register(ctx, file, i); |
3819 | if (err) | |
3820 | break; | |
3821 | } | |
3822 | nr_args--; | |
3823 | done++; | |
3824 | up.offset++; | |
3825 | } | |
3826 | ||
3827 | return done ? done : err; | |
3828 | } | |
3829 | ||
7d723065 JA |
3830 | static void io_put_work(struct io_wq_work *work) |
3831 | { | |
3832 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
3833 | ||
3834 | io_put_req(req); | |
3835 | } | |
3836 | ||
3837 | static void io_get_work(struct io_wq_work *work) | |
3838 | { | |
3839 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
3840 | ||
3841 | refcount_inc(&req->refs); | |
3842 | } | |
3843 | ||
6c271ce2 JA |
3844 | static int io_sq_offload_start(struct io_ring_ctx *ctx, |
3845 | struct io_uring_params *p) | |
2b188cc1 | 3846 | { |
561fb04a | 3847 | unsigned concurrency; |
2b188cc1 JA |
3848 | int ret; |
3849 | ||
6c271ce2 | 3850 | init_waitqueue_head(&ctx->sqo_wait); |
2b188cc1 JA |
3851 | mmgrab(current->mm); |
3852 | ctx->sqo_mm = current->mm; | |
3853 | ||
6c271ce2 | 3854 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
3ec482d1 JA |
3855 | ret = -EPERM; |
3856 | if (!capable(CAP_SYS_ADMIN)) | |
3857 | goto err; | |
3858 | ||
917257da JA |
3859 | ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle); |
3860 | if (!ctx->sq_thread_idle) | |
3861 | ctx->sq_thread_idle = HZ; | |
3862 | ||
6c271ce2 | 3863 | if (p->flags & IORING_SETUP_SQ_AFF) { |
44a9bd18 | 3864 | int cpu = p->sq_thread_cpu; |
6c271ce2 | 3865 | |
917257da | 3866 | ret = -EINVAL; |
44a9bd18 JA |
3867 | if (cpu >= nr_cpu_ids) |
3868 | goto err; | |
7889f44d | 3869 | if (!cpu_online(cpu)) |
917257da JA |
3870 | goto err; |
3871 | ||
6c271ce2 JA |
3872 | ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread, |
3873 | ctx, cpu, | |
3874 | "io_uring-sq"); | |
3875 | } else { | |
3876 | ctx->sqo_thread = kthread_create(io_sq_thread, ctx, | |
3877 | "io_uring-sq"); | |
3878 | } | |
3879 | if (IS_ERR(ctx->sqo_thread)) { | |
3880 | ret = PTR_ERR(ctx->sqo_thread); | |
3881 | ctx->sqo_thread = NULL; | |
3882 | goto err; | |
3883 | } | |
3884 | wake_up_process(ctx->sqo_thread); | |
3885 | } else if (p->flags & IORING_SETUP_SQ_AFF) { | |
3886 | /* Can't have SQ_AFF without SQPOLL */ | |
3887 | ret = -EINVAL; | |
3888 | goto err; | |
3889 | } | |
3890 | ||
561fb04a JA |
3891 | /* Do QD, or 4 * CPUS, whatever is smallest */ |
3892 | concurrency = min(ctx->sq_entries, 4 * num_online_cpus()); | |
7d723065 JA |
3893 | ctx->io_wq = io_wq_create(concurrency, ctx->sqo_mm, ctx->user, |
3894 | io_get_work, io_put_work); | |
975c99a5 JA |
3895 | if (IS_ERR(ctx->io_wq)) { |
3896 | ret = PTR_ERR(ctx->io_wq); | |
3897 | ctx->io_wq = NULL; | |
2b188cc1 JA |
3898 | goto err; |
3899 | } | |
3900 | ||
3901 | return 0; | |
3902 | err: | |
54a91f3b | 3903 | io_finish_async(ctx); |
2b188cc1 JA |
3904 | mmdrop(ctx->sqo_mm); |
3905 | ctx->sqo_mm = NULL; | |
3906 | return ret; | |
3907 | } | |
3908 | ||
3909 | static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages) | |
3910 | { | |
3911 | atomic_long_sub(nr_pages, &user->locked_vm); | |
3912 | } | |
3913 | ||
3914 | static int io_account_mem(struct user_struct *user, unsigned long nr_pages) | |
3915 | { | |
3916 | unsigned long page_limit, cur_pages, new_pages; | |
3917 | ||
3918 | /* Don't allow more pages than we can safely lock */ | |
3919 | page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
3920 | ||
3921 | do { | |
3922 | cur_pages = atomic_long_read(&user->locked_vm); | |
3923 | new_pages = cur_pages + nr_pages; | |
3924 | if (new_pages > page_limit) | |
3925 | return -ENOMEM; | |
3926 | } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages, | |
3927 | new_pages) != cur_pages); | |
3928 | ||
3929 | return 0; | |
3930 | } | |
3931 | ||
3932 | static void io_mem_free(void *ptr) | |
3933 | { | |
52e04ef4 MR |
3934 | struct page *page; |
3935 | ||
3936 | if (!ptr) | |
3937 | return; | |
2b188cc1 | 3938 | |
52e04ef4 | 3939 | page = virt_to_head_page(ptr); |
2b188cc1 JA |
3940 | if (put_page_testzero(page)) |
3941 | free_compound_page(page); | |
3942 | } | |
3943 | ||
3944 | static void *io_mem_alloc(size_t size) | |
3945 | { | |
3946 | gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP | | |
3947 | __GFP_NORETRY; | |
3948 | ||
3949 | return (void *) __get_free_pages(gfp_flags, get_order(size)); | |
3950 | } | |
3951 | ||
75b28aff HV |
3952 | static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries, |
3953 | size_t *sq_offset) | |
3954 | { | |
3955 | struct io_rings *rings; | |
3956 | size_t off, sq_array_size; | |
3957 | ||
3958 | off = struct_size(rings, cqes, cq_entries); | |
3959 | if (off == SIZE_MAX) | |
3960 | return SIZE_MAX; | |
3961 | ||
3962 | #ifdef CONFIG_SMP | |
3963 | off = ALIGN(off, SMP_CACHE_BYTES); | |
3964 | if (off == 0) | |
3965 | return SIZE_MAX; | |
3966 | #endif | |
3967 | ||
3968 | sq_array_size = array_size(sizeof(u32), sq_entries); | |
3969 | if (sq_array_size == SIZE_MAX) | |
3970 | return SIZE_MAX; | |
3971 | ||
3972 | if (check_add_overflow(off, sq_array_size, &off)) | |
3973 | return SIZE_MAX; | |
3974 | ||
3975 | if (sq_offset) | |
3976 | *sq_offset = off; | |
3977 | ||
3978 | return off; | |
3979 | } | |
3980 | ||
2b188cc1 JA |
3981 | static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries) |
3982 | { | |
75b28aff | 3983 | size_t pages; |
2b188cc1 | 3984 | |
75b28aff HV |
3985 | pages = (size_t)1 << get_order( |
3986 | rings_size(sq_entries, cq_entries, NULL)); | |
3987 | pages += (size_t)1 << get_order( | |
3988 | array_size(sizeof(struct io_uring_sqe), sq_entries)); | |
2b188cc1 | 3989 | |
75b28aff | 3990 | return pages; |
2b188cc1 JA |
3991 | } |
3992 | ||
edafccee JA |
3993 | static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx) |
3994 | { | |
3995 | int i, j; | |
3996 | ||
3997 | if (!ctx->user_bufs) | |
3998 | return -ENXIO; | |
3999 | ||
4000 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
4001 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
4002 | ||
4003 | for (j = 0; j < imu->nr_bvecs; j++) | |
27c4d3a3 | 4004 | put_user_page(imu->bvec[j].bv_page); |
edafccee JA |
4005 | |
4006 | if (ctx->account_mem) | |
4007 | io_unaccount_mem(ctx->user, imu->nr_bvecs); | |
d4ef6475 | 4008 | kvfree(imu->bvec); |
edafccee JA |
4009 | imu->nr_bvecs = 0; |
4010 | } | |
4011 | ||
4012 | kfree(ctx->user_bufs); | |
4013 | ctx->user_bufs = NULL; | |
4014 | ctx->nr_user_bufs = 0; | |
4015 | return 0; | |
4016 | } | |
4017 | ||
4018 | static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, | |
4019 | void __user *arg, unsigned index) | |
4020 | { | |
4021 | struct iovec __user *src; | |
4022 | ||
4023 | #ifdef CONFIG_COMPAT | |
4024 | if (ctx->compat) { | |
4025 | struct compat_iovec __user *ciovs; | |
4026 | struct compat_iovec ciov; | |
4027 | ||
4028 | ciovs = (struct compat_iovec __user *) arg; | |
4029 | if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) | |
4030 | return -EFAULT; | |
4031 | ||
4032 | dst->iov_base = (void __user *) (unsigned long) ciov.iov_base; | |
4033 | dst->iov_len = ciov.iov_len; | |
4034 | return 0; | |
4035 | } | |
4036 | #endif | |
4037 | src = (struct iovec __user *) arg; | |
4038 | if (copy_from_user(dst, &src[index], sizeof(*dst))) | |
4039 | return -EFAULT; | |
4040 | return 0; | |
4041 | } | |
4042 | ||
4043 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg, | |
4044 | unsigned nr_args) | |
4045 | { | |
4046 | struct vm_area_struct **vmas = NULL; | |
4047 | struct page **pages = NULL; | |
4048 | int i, j, got_pages = 0; | |
4049 | int ret = -EINVAL; | |
4050 | ||
4051 | if (ctx->user_bufs) | |
4052 | return -EBUSY; | |
4053 | if (!nr_args || nr_args > UIO_MAXIOV) | |
4054 | return -EINVAL; | |
4055 | ||
4056 | ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf), | |
4057 | GFP_KERNEL); | |
4058 | if (!ctx->user_bufs) | |
4059 | return -ENOMEM; | |
4060 | ||
4061 | for (i = 0; i < nr_args; i++) { | |
4062 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
4063 | unsigned long off, start, end, ubuf; | |
4064 | int pret, nr_pages; | |
4065 | struct iovec iov; | |
4066 | size_t size; | |
4067 | ||
4068 | ret = io_copy_iov(ctx, &iov, arg, i); | |
4069 | if (ret) | |
a278682d | 4070 | goto err; |
edafccee JA |
4071 | |
4072 | /* | |
4073 | * Don't impose further limits on the size and buffer | |
4074 | * constraints here, we'll -EINVAL later when IO is | |
4075 | * submitted if they are wrong. | |
4076 | */ | |
4077 | ret = -EFAULT; | |
4078 | if (!iov.iov_base || !iov.iov_len) | |
4079 | goto err; | |
4080 | ||
4081 | /* arbitrary limit, but we need something */ | |
4082 | if (iov.iov_len > SZ_1G) | |
4083 | goto err; | |
4084 | ||
4085 | ubuf = (unsigned long) iov.iov_base; | |
4086 | end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
4087 | start = ubuf >> PAGE_SHIFT; | |
4088 | nr_pages = end - start; | |
4089 | ||
4090 | if (ctx->account_mem) { | |
4091 | ret = io_account_mem(ctx->user, nr_pages); | |
4092 | if (ret) | |
4093 | goto err; | |
4094 | } | |
4095 | ||
4096 | ret = 0; | |
4097 | if (!pages || nr_pages > got_pages) { | |
4098 | kfree(vmas); | |
4099 | kfree(pages); | |
d4ef6475 | 4100 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), |
edafccee | 4101 | GFP_KERNEL); |
d4ef6475 | 4102 | vmas = kvmalloc_array(nr_pages, |
edafccee JA |
4103 | sizeof(struct vm_area_struct *), |
4104 | GFP_KERNEL); | |
4105 | if (!pages || !vmas) { | |
4106 | ret = -ENOMEM; | |
4107 | if (ctx->account_mem) | |
4108 | io_unaccount_mem(ctx->user, nr_pages); | |
4109 | goto err; | |
4110 | } | |
4111 | got_pages = nr_pages; | |
4112 | } | |
4113 | ||
d4ef6475 | 4114 | imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec), |
edafccee JA |
4115 | GFP_KERNEL); |
4116 | ret = -ENOMEM; | |
4117 | if (!imu->bvec) { | |
4118 | if (ctx->account_mem) | |
4119 | io_unaccount_mem(ctx->user, nr_pages); | |
4120 | goto err; | |
4121 | } | |
4122 | ||
4123 | ret = 0; | |
4124 | down_read(¤t->mm->mmap_sem); | |
932f4a63 IW |
4125 | pret = get_user_pages(ubuf, nr_pages, |
4126 | FOLL_WRITE | FOLL_LONGTERM, | |
4127 | pages, vmas); | |
edafccee JA |
4128 | if (pret == nr_pages) { |
4129 | /* don't support file backed memory */ | |
4130 | for (j = 0; j < nr_pages; j++) { | |
4131 | struct vm_area_struct *vma = vmas[j]; | |
4132 | ||
4133 | if (vma->vm_file && | |
4134 | !is_file_hugepages(vma->vm_file)) { | |
4135 | ret = -EOPNOTSUPP; | |
4136 | break; | |
4137 | } | |
4138 | } | |
4139 | } else { | |
4140 | ret = pret < 0 ? pret : -EFAULT; | |
4141 | } | |
4142 | up_read(¤t->mm->mmap_sem); | |
4143 | if (ret) { | |
4144 | /* | |
4145 | * if we did partial map, or found file backed vmas, | |
4146 | * release any pages we did get | |
4147 | */ | |
27c4d3a3 JH |
4148 | if (pret > 0) |
4149 | put_user_pages(pages, pret); | |
edafccee JA |
4150 | if (ctx->account_mem) |
4151 | io_unaccount_mem(ctx->user, nr_pages); | |
d4ef6475 | 4152 | kvfree(imu->bvec); |
edafccee JA |
4153 | goto err; |
4154 | } | |
4155 | ||
4156 | off = ubuf & ~PAGE_MASK; | |
4157 | size = iov.iov_len; | |
4158 | for (j = 0; j < nr_pages; j++) { | |
4159 | size_t vec_len; | |
4160 | ||
4161 | vec_len = min_t(size_t, size, PAGE_SIZE - off); | |
4162 | imu->bvec[j].bv_page = pages[j]; | |
4163 | imu->bvec[j].bv_len = vec_len; | |
4164 | imu->bvec[j].bv_offset = off; | |
4165 | off = 0; | |
4166 | size -= vec_len; | |
4167 | } | |
4168 | /* store original address for later verification */ | |
4169 | imu->ubuf = ubuf; | |
4170 | imu->len = iov.iov_len; | |
4171 | imu->nr_bvecs = nr_pages; | |
4172 | ||
4173 | ctx->nr_user_bufs++; | |
4174 | } | |
d4ef6475 MR |
4175 | kvfree(pages); |
4176 | kvfree(vmas); | |
edafccee JA |
4177 | return 0; |
4178 | err: | |
d4ef6475 MR |
4179 | kvfree(pages); |
4180 | kvfree(vmas); | |
edafccee JA |
4181 | io_sqe_buffer_unregister(ctx); |
4182 | return ret; | |
4183 | } | |
4184 | ||
9b402849 JA |
4185 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg) |
4186 | { | |
4187 | __s32 __user *fds = arg; | |
4188 | int fd; | |
4189 | ||
4190 | if (ctx->cq_ev_fd) | |
4191 | return -EBUSY; | |
4192 | ||
4193 | if (copy_from_user(&fd, fds, sizeof(*fds))) | |
4194 | return -EFAULT; | |
4195 | ||
4196 | ctx->cq_ev_fd = eventfd_ctx_fdget(fd); | |
4197 | if (IS_ERR(ctx->cq_ev_fd)) { | |
4198 | int ret = PTR_ERR(ctx->cq_ev_fd); | |
4199 | ctx->cq_ev_fd = NULL; | |
4200 | return ret; | |
4201 | } | |
4202 | ||
4203 | return 0; | |
4204 | } | |
4205 | ||
4206 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) | |
4207 | { | |
4208 | if (ctx->cq_ev_fd) { | |
4209 | eventfd_ctx_put(ctx->cq_ev_fd); | |
4210 | ctx->cq_ev_fd = NULL; | |
4211 | return 0; | |
4212 | } | |
4213 | ||
4214 | return -ENXIO; | |
4215 | } | |
4216 | ||
2b188cc1 JA |
4217 | static void io_ring_ctx_free(struct io_ring_ctx *ctx) |
4218 | { | |
6b06314c | 4219 | io_finish_async(ctx); |
2b188cc1 JA |
4220 | if (ctx->sqo_mm) |
4221 | mmdrop(ctx->sqo_mm); | |
def596e9 JA |
4222 | |
4223 | io_iopoll_reap_events(ctx); | |
edafccee | 4224 | io_sqe_buffer_unregister(ctx); |
6b06314c | 4225 | io_sqe_files_unregister(ctx); |
9b402849 | 4226 | io_eventfd_unregister(ctx); |
def596e9 | 4227 | |
2b188cc1 | 4228 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
4229 | if (ctx->ring_sock) { |
4230 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 4231 | sock_release(ctx->ring_sock); |
355e8d26 | 4232 | } |
2b188cc1 JA |
4233 | #endif |
4234 | ||
75b28aff | 4235 | io_mem_free(ctx->rings); |
2b188cc1 | 4236 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
4237 | |
4238 | percpu_ref_exit(&ctx->refs); | |
4239 | if (ctx->account_mem) | |
4240 | io_unaccount_mem(ctx->user, | |
4241 | ring_pages(ctx->sq_entries, ctx->cq_entries)); | |
4242 | free_uid(ctx->user); | |
206aefde | 4243 | kfree(ctx->completions); |
0ddf92e8 | 4244 | kmem_cache_free(req_cachep, ctx->fallback_req); |
2b188cc1 JA |
4245 | kfree(ctx); |
4246 | } | |
4247 | ||
4248 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
4249 | { | |
4250 | struct io_ring_ctx *ctx = file->private_data; | |
4251 | __poll_t mask = 0; | |
4252 | ||
4253 | poll_wait(file, &ctx->cq_wait, wait); | |
4f7067c3 SB |
4254 | /* |
4255 | * synchronizes with barrier from wq_has_sleeper call in | |
4256 | * io_commit_cqring | |
4257 | */ | |
2b188cc1 | 4258 | smp_rmb(); |
75b28aff HV |
4259 | if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head != |
4260 | ctx->rings->sq_ring_entries) | |
2b188cc1 | 4261 | mask |= EPOLLOUT | EPOLLWRNORM; |
daa5de54 | 4262 | if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail) |
2b188cc1 JA |
4263 | mask |= EPOLLIN | EPOLLRDNORM; |
4264 | ||
4265 | return mask; | |
4266 | } | |
4267 | ||
4268 | static int io_uring_fasync(int fd, struct file *file, int on) | |
4269 | { | |
4270 | struct io_ring_ctx *ctx = file->private_data; | |
4271 | ||
4272 | return fasync_helper(fd, file, on, &ctx->cq_fasync); | |
4273 | } | |
4274 | ||
4275 | static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) | |
4276 | { | |
4277 | mutex_lock(&ctx->uring_lock); | |
4278 | percpu_ref_kill(&ctx->refs); | |
4279 | mutex_unlock(&ctx->uring_lock); | |
4280 | ||
5262f567 | 4281 | io_kill_timeouts(ctx); |
221c5eb2 | 4282 | io_poll_remove_all(ctx); |
561fb04a JA |
4283 | |
4284 | if (ctx->io_wq) | |
4285 | io_wq_cancel_all(ctx->io_wq); | |
4286 | ||
def596e9 | 4287 | io_iopoll_reap_events(ctx); |
15dff286 JA |
4288 | /* if we failed setting up the ctx, we might not have any rings */ |
4289 | if (ctx->rings) | |
4290 | io_cqring_overflow_flush(ctx, true); | |
206aefde | 4291 | wait_for_completion(&ctx->completions[0]); |
2b188cc1 JA |
4292 | io_ring_ctx_free(ctx); |
4293 | } | |
4294 | ||
4295 | static int io_uring_release(struct inode *inode, struct file *file) | |
4296 | { | |
4297 | struct io_ring_ctx *ctx = file->private_data; | |
4298 | ||
4299 | file->private_data = NULL; | |
4300 | io_ring_ctx_wait_and_kill(ctx); | |
4301 | return 0; | |
4302 | } | |
4303 | ||
fcb323cc JA |
4304 | static void io_uring_cancel_files(struct io_ring_ctx *ctx, |
4305 | struct files_struct *files) | |
4306 | { | |
4307 | struct io_kiocb *req; | |
4308 | DEFINE_WAIT(wait); | |
4309 | ||
4310 | while (!list_empty_careful(&ctx->inflight_list)) { | |
768134d4 | 4311 | struct io_kiocb *cancel_req = NULL; |
fcb323cc JA |
4312 | |
4313 | spin_lock_irq(&ctx->inflight_lock); | |
4314 | list_for_each_entry(req, &ctx->inflight_list, inflight_entry) { | |
768134d4 JA |
4315 | if (req->work.files != files) |
4316 | continue; | |
4317 | /* req is being completed, ignore */ | |
4318 | if (!refcount_inc_not_zero(&req->refs)) | |
4319 | continue; | |
4320 | cancel_req = req; | |
4321 | break; | |
fcb323cc | 4322 | } |
768134d4 | 4323 | if (cancel_req) |
fcb323cc | 4324 | prepare_to_wait(&ctx->inflight_wait, &wait, |
768134d4 | 4325 | TASK_UNINTERRUPTIBLE); |
fcb323cc JA |
4326 | spin_unlock_irq(&ctx->inflight_lock); |
4327 | ||
768134d4 JA |
4328 | /* We need to keep going until we don't find a matching req */ |
4329 | if (!cancel_req) | |
fcb323cc | 4330 | break; |
2f6d9b9d BL |
4331 | |
4332 | io_wq_cancel_work(ctx->io_wq, &cancel_req->work); | |
4333 | io_put_req(cancel_req); | |
fcb323cc JA |
4334 | schedule(); |
4335 | } | |
768134d4 | 4336 | finish_wait(&ctx->inflight_wait, &wait); |
fcb323cc JA |
4337 | } |
4338 | ||
4339 | static int io_uring_flush(struct file *file, void *data) | |
4340 | { | |
4341 | struct io_ring_ctx *ctx = file->private_data; | |
4342 | ||
4343 | io_uring_cancel_files(ctx, data); | |
1d7bb1d5 JA |
4344 | if (fatal_signal_pending(current) || (current->flags & PF_EXITING)) { |
4345 | io_cqring_overflow_flush(ctx, true); | |
fcb323cc | 4346 | io_wq_cancel_all(ctx->io_wq); |
1d7bb1d5 | 4347 | } |
fcb323cc JA |
4348 | return 0; |
4349 | } | |
4350 | ||
2b188cc1 JA |
4351 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) |
4352 | { | |
4353 | loff_t offset = (loff_t) vma->vm_pgoff << PAGE_SHIFT; | |
4354 | unsigned long sz = vma->vm_end - vma->vm_start; | |
4355 | struct io_ring_ctx *ctx = file->private_data; | |
4356 | unsigned long pfn; | |
4357 | struct page *page; | |
4358 | void *ptr; | |
4359 | ||
4360 | switch (offset) { | |
4361 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
4362 | case IORING_OFF_CQ_RING: |
4363 | ptr = ctx->rings; | |
2b188cc1 JA |
4364 | break; |
4365 | case IORING_OFF_SQES: | |
4366 | ptr = ctx->sq_sqes; | |
4367 | break; | |
2b188cc1 JA |
4368 | default: |
4369 | return -EINVAL; | |
4370 | } | |
4371 | ||
4372 | page = virt_to_head_page(ptr); | |
a50b854e | 4373 | if (sz > page_size(page)) |
2b188cc1 JA |
4374 | return -EINVAL; |
4375 | ||
4376 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
4377 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
4378 | } | |
4379 | ||
4380 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, | |
4381 | u32, min_complete, u32, flags, const sigset_t __user *, sig, | |
4382 | size_t, sigsz) | |
4383 | { | |
4384 | struct io_ring_ctx *ctx; | |
4385 | long ret = -EBADF; | |
4386 | int submitted = 0; | |
4387 | struct fd f; | |
4388 | ||
6c271ce2 | 4389 | if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP)) |
2b188cc1 JA |
4390 | return -EINVAL; |
4391 | ||
4392 | f = fdget(fd); | |
4393 | if (!f.file) | |
4394 | return -EBADF; | |
4395 | ||
4396 | ret = -EOPNOTSUPP; | |
4397 | if (f.file->f_op != &io_uring_fops) | |
4398 | goto out_fput; | |
4399 | ||
4400 | ret = -ENXIO; | |
4401 | ctx = f.file->private_data; | |
4402 | if (!percpu_ref_tryget(&ctx->refs)) | |
4403 | goto out_fput; | |
4404 | ||
6c271ce2 JA |
4405 | /* |
4406 | * For SQ polling, the thread will do all submissions and completions. | |
4407 | * Just return the requested submit count, and wake the thread if | |
4408 | * we were asked to. | |
4409 | */ | |
b2a9eada | 4410 | ret = 0; |
6c271ce2 | 4411 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
c1edbf5f JA |
4412 | if (!list_empty_careful(&ctx->cq_overflow_list)) |
4413 | io_cqring_overflow_flush(ctx, false); | |
6c271ce2 JA |
4414 | if (flags & IORING_ENTER_SQ_WAKEUP) |
4415 | wake_up(&ctx->sqo_wait); | |
4416 | submitted = to_submit; | |
b2a9eada | 4417 | } else if (to_submit) { |
ae9428ca | 4418 | struct mm_struct *cur_mm; |
2b188cc1 | 4419 | |
ae9428ca | 4420 | to_submit = min(to_submit, ctx->sq_entries); |
2b188cc1 | 4421 | mutex_lock(&ctx->uring_lock); |
ae9428ca PB |
4422 | /* already have mm, so io_submit_sqes() won't try to grab it */ |
4423 | cur_mm = ctx->sqo_mm; | |
4424 | submitted = io_submit_sqes(ctx, to_submit, f.file, fd, | |
4425 | &cur_mm, false); | |
2b188cc1 | 4426 | mutex_unlock(&ctx->uring_lock); |
2b188cc1 JA |
4427 | } |
4428 | if (flags & IORING_ENTER_GETEVENTS) { | |
def596e9 JA |
4429 | unsigned nr_events = 0; |
4430 | ||
2b188cc1 JA |
4431 | min_complete = min(min_complete, ctx->cq_entries); |
4432 | ||
def596e9 | 4433 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 | 4434 | ret = io_iopoll_check(ctx, &nr_events, min_complete); |
def596e9 JA |
4435 | } else { |
4436 | ret = io_cqring_wait(ctx, min_complete, sig, sigsz); | |
4437 | } | |
2b188cc1 JA |
4438 | } |
4439 | ||
6805b32e | 4440 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
4441 | out_fput: |
4442 | fdput(f); | |
4443 | return submitted ? submitted : ret; | |
4444 | } | |
4445 | ||
4446 | static const struct file_operations io_uring_fops = { | |
4447 | .release = io_uring_release, | |
fcb323cc | 4448 | .flush = io_uring_flush, |
2b188cc1 JA |
4449 | .mmap = io_uring_mmap, |
4450 | .poll = io_uring_poll, | |
4451 | .fasync = io_uring_fasync, | |
4452 | }; | |
4453 | ||
4454 | static int io_allocate_scq_urings(struct io_ring_ctx *ctx, | |
4455 | struct io_uring_params *p) | |
4456 | { | |
75b28aff HV |
4457 | struct io_rings *rings; |
4458 | size_t size, sq_array_offset; | |
2b188cc1 | 4459 | |
75b28aff HV |
4460 | size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset); |
4461 | if (size == SIZE_MAX) | |
4462 | return -EOVERFLOW; | |
4463 | ||
4464 | rings = io_mem_alloc(size); | |
4465 | if (!rings) | |
2b188cc1 JA |
4466 | return -ENOMEM; |
4467 | ||
75b28aff HV |
4468 | ctx->rings = rings; |
4469 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
4470 | rings->sq_ring_mask = p->sq_entries - 1; | |
4471 | rings->cq_ring_mask = p->cq_entries - 1; | |
4472 | rings->sq_ring_entries = p->sq_entries; | |
4473 | rings->cq_ring_entries = p->cq_entries; | |
4474 | ctx->sq_mask = rings->sq_ring_mask; | |
4475 | ctx->cq_mask = rings->cq_ring_mask; | |
4476 | ctx->sq_entries = rings->sq_ring_entries; | |
4477 | ctx->cq_entries = rings->cq_ring_entries; | |
2b188cc1 JA |
4478 | |
4479 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
4480 | if (size == SIZE_MAX) | |
4481 | return -EOVERFLOW; | |
4482 | ||
4483 | ctx->sq_sqes = io_mem_alloc(size); | |
52e04ef4 | 4484 | if (!ctx->sq_sqes) |
2b188cc1 | 4485 | return -ENOMEM; |
2b188cc1 | 4486 | |
2b188cc1 JA |
4487 | return 0; |
4488 | } | |
4489 | ||
4490 | /* | |
4491 | * Allocate an anonymous fd, this is what constitutes the application | |
4492 | * visible backing of an io_uring instance. The application mmaps this | |
4493 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
4494 | * we have to tie this fd to a socket for file garbage collection purposes. | |
4495 | */ | |
4496 | static int io_uring_get_fd(struct io_ring_ctx *ctx) | |
4497 | { | |
4498 | struct file *file; | |
4499 | int ret; | |
4500 | ||
4501 | #if defined(CONFIG_UNIX) | |
4502 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, | |
4503 | &ctx->ring_sock); | |
4504 | if (ret) | |
4505 | return ret; | |
4506 | #endif | |
4507 | ||
4508 | ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
4509 | if (ret < 0) | |
4510 | goto err; | |
4511 | ||
4512 | file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx, | |
4513 | O_RDWR | O_CLOEXEC); | |
4514 | if (IS_ERR(file)) { | |
4515 | put_unused_fd(ret); | |
4516 | ret = PTR_ERR(file); | |
4517 | goto err; | |
4518 | } | |
4519 | ||
4520 | #if defined(CONFIG_UNIX) | |
4521 | ctx->ring_sock->file = file; | |
6b06314c | 4522 | ctx->ring_sock->sk->sk_user_data = ctx; |
2b188cc1 JA |
4523 | #endif |
4524 | fd_install(ret, file); | |
4525 | return ret; | |
4526 | err: | |
4527 | #if defined(CONFIG_UNIX) | |
4528 | sock_release(ctx->ring_sock); | |
4529 | ctx->ring_sock = NULL; | |
4530 | #endif | |
4531 | return ret; | |
4532 | } | |
4533 | ||
4534 | static int io_uring_create(unsigned entries, struct io_uring_params *p) | |
4535 | { | |
4536 | struct user_struct *user = NULL; | |
4537 | struct io_ring_ctx *ctx; | |
4538 | bool account_mem; | |
4539 | int ret; | |
4540 | ||
4541 | if (!entries || entries > IORING_MAX_ENTRIES) | |
4542 | return -EINVAL; | |
4543 | ||
4544 | /* | |
4545 | * Use twice as many entries for the CQ ring. It's possible for the | |
4546 | * application to drive a higher depth than the size of the SQ ring, | |
4547 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
4548 | * some flexibility in overcommitting a bit. If the application has |
4549 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
4550 | * of CQ ring entries manually. | |
2b188cc1 JA |
4551 | */ |
4552 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
4553 | if (p->flags & IORING_SETUP_CQSIZE) { |
4554 | /* | |
4555 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
4556 | * to a power-of-two, if it isn't already. We do NOT impose | |
4557 | * any cq vs sq ring sizing. | |
4558 | */ | |
4559 | if (p->cq_entries < p->sq_entries || p->cq_entries > IORING_MAX_CQ_ENTRIES) | |
4560 | return -EINVAL; | |
4561 | p->cq_entries = roundup_pow_of_two(p->cq_entries); | |
4562 | } else { | |
4563 | p->cq_entries = 2 * p->sq_entries; | |
4564 | } | |
2b188cc1 JA |
4565 | |
4566 | user = get_uid(current_user()); | |
4567 | account_mem = !capable(CAP_IPC_LOCK); | |
4568 | ||
4569 | if (account_mem) { | |
4570 | ret = io_account_mem(user, | |
4571 | ring_pages(p->sq_entries, p->cq_entries)); | |
4572 | if (ret) { | |
4573 | free_uid(user); | |
4574 | return ret; | |
4575 | } | |
4576 | } | |
4577 | ||
4578 | ctx = io_ring_ctx_alloc(p); | |
4579 | if (!ctx) { | |
4580 | if (account_mem) | |
4581 | io_unaccount_mem(user, ring_pages(p->sq_entries, | |
4582 | p->cq_entries)); | |
4583 | free_uid(user); | |
4584 | return -ENOMEM; | |
4585 | } | |
4586 | ctx->compat = in_compat_syscall(); | |
4587 | ctx->account_mem = account_mem; | |
4588 | ctx->user = user; | |
4589 | ||
4590 | ret = io_allocate_scq_urings(ctx, p); | |
4591 | if (ret) | |
4592 | goto err; | |
4593 | ||
6c271ce2 | 4594 | ret = io_sq_offload_start(ctx, p); |
2b188cc1 JA |
4595 | if (ret) |
4596 | goto err; | |
4597 | ||
2b188cc1 | 4598 | memset(&p->sq_off, 0, sizeof(p->sq_off)); |
75b28aff HV |
4599 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
4600 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
4601 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
4602 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
4603 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
4604 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
4605 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
4606 | |
4607 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
4608 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
4609 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
4610 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
4611 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
4612 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
4613 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
ac90f249 | 4614 | |
044c1ab3 JA |
4615 | /* |
4616 | * Install ring fd as the very last thing, so we don't risk someone | |
4617 | * having closed it before we finish setup | |
4618 | */ | |
4619 | ret = io_uring_get_fd(ctx); | |
4620 | if (ret < 0) | |
4621 | goto err; | |
4622 | ||
1d7bb1d5 | 4623 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP; |
c826bd7a | 4624 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
4625 | return ret; |
4626 | err: | |
4627 | io_ring_ctx_wait_and_kill(ctx); | |
4628 | return ret; | |
4629 | } | |
4630 | ||
4631 | /* | |
4632 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
4633 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
4634 | * params structure passed in. | |
4635 | */ | |
4636 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
4637 | { | |
4638 | struct io_uring_params p; | |
4639 | long ret; | |
4640 | int i; | |
4641 | ||
4642 | if (copy_from_user(&p, params, sizeof(p))) | |
4643 | return -EFAULT; | |
4644 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
4645 | if (p.resv[i]) | |
4646 | return -EINVAL; | |
4647 | } | |
4648 | ||
6c271ce2 | 4649 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
33a107f0 | 4650 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE)) |
2b188cc1 JA |
4651 | return -EINVAL; |
4652 | ||
4653 | ret = io_uring_create(entries, &p); | |
4654 | if (ret < 0) | |
4655 | return ret; | |
4656 | ||
4657 | if (copy_to_user(params, &p, sizeof(p))) | |
4658 | return -EFAULT; | |
4659 | ||
4660 | return ret; | |
4661 | } | |
4662 | ||
4663 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
4664 | struct io_uring_params __user *, params) | |
4665 | { | |
4666 | return io_uring_setup(entries, params); | |
4667 | } | |
4668 | ||
edafccee JA |
4669 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
4670 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
4671 | __releases(ctx->uring_lock) |
4672 | __acquires(ctx->uring_lock) | |
edafccee JA |
4673 | { |
4674 | int ret; | |
4675 | ||
35fa71a0 JA |
4676 | /* |
4677 | * We're inside the ring mutex, if the ref is already dying, then | |
4678 | * someone else killed the ctx or is already going through | |
4679 | * io_uring_register(). | |
4680 | */ | |
4681 | if (percpu_ref_is_dying(&ctx->refs)) | |
4682 | return -ENXIO; | |
4683 | ||
edafccee | 4684 | percpu_ref_kill(&ctx->refs); |
b19062a5 JA |
4685 | |
4686 | /* | |
4687 | * Drop uring mutex before waiting for references to exit. If another | |
4688 | * thread is currently inside io_uring_enter() it might need to grab | |
4689 | * the uring_lock to make progress. If we hold it here across the drain | |
4690 | * wait, then we can deadlock. It's safe to drop the mutex here, since | |
4691 | * no new references will come in after we've killed the percpu ref. | |
4692 | */ | |
4693 | mutex_unlock(&ctx->uring_lock); | |
206aefde | 4694 | wait_for_completion(&ctx->completions[0]); |
b19062a5 | 4695 | mutex_lock(&ctx->uring_lock); |
edafccee JA |
4696 | |
4697 | switch (opcode) { | |
4698 | case IORING_REGISTER_BUFFERS: | |
4699 | ret = io_sqe_buffer_register(ctx, arg, nr_args); | |
4700 | break; | |
4701 | case IORING_UNREGISTER_BUFFERS: | |
4702 | ret = -EINVAL; | |
4703 | if (arg || nr_args) | |
4704 | break; | |
4705 | ret = io_sqe_buffer_unregister(ctx); | |
4706 | break; | |
6b06314c JA |
4707 | case IORING_REGISTER_FILES: |
4708 | ret = io_sqe_files_register(ctx, arg, nr_args); | |
4709 | break; | |
4710 | case IORING_UNREGISTER_FILES: | |
4711 | ret = -EINVAL; | |
4712 | if (arg || nr_args) | |
4713 | break; | |
4714 | ret = io_sqe_files_unregister(ctx); | |
4715 | break; | |
c3a31e60 JA |
4716 | case IORING_REGISTER_FILES_UPDATE: |
4717 | ret = io_sqe_files_update(ctx, arg, nr_args); | |
4718 | break; | |
9b402849 JA |
4719 | case IORING_REGISTER_EVENTFD: |
4720 | ret = -EINVAL; | |
4721 | if (nr_args != 1) | |
4722 | break; | |
4723 | ret = io_eventfd_register(ctx, arg); | |
4724 | break; | |
4725 | case IORING_UNREGISTER_EVENTFD: | |
4726 | ret = -EINVAL; | |
4727 | if (arg || nr_args) | |
4728 | break; | |
4729 | ret = io_eventfd_unregister(ctx); | |
4730 | break; | |
edafccee JA |
4731 | default: |
4732 | ret = -EINVAL; | |
4733 | break; | |
4734 | } | |
4735 | ||
4736 | /* bring the ctx back to life */ | |
206aefde | 4737 | reinit_completion(&ctx->completions[0]); |
edafccee JA |
4738 | percpu_ref_reinit(&ctx->refs); |
4739 | return ret; | |
4740 | } | |
4741 | ||
4742 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
4743 | void __user *, arg, unsigned int, nr_args) | |
4744 | { | |
4745 | struct io_ring_ctx *ctx; | |
4746 | long ret = -EBADF; | |
4747 | struct fd f; | |
4748 | ||
4749 | f = fdget(fd); | |
4750 | if (!f.file) | |
4751 | return -EBADF; | |
4752 | ||
4753 | ret = -EOPNOTSUPP; | |
4754 | if (f.file->f_op != &io_uring_fops) | |
4755 | goto out_fput; | |
4756 | ||
4757 | ctx = f.file->private_data; | |
4758 | ||
4759 | mutex_lock(&ctx->uring_lock); | |
4760 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
4761 | mutex_unlock(&ctx->uring_lock); | |
c826bd7a DD |
4762 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, |
4763 | ctx->cq_ev_fd != NULL, ret); | |
edafccee JA |
4764 | out_fput: |
4765 | fdput(f); | |
4766 | return ret; | |
4767 | } | |
4768 | ||
2b188cc1 JA |
4769 | static int __init io_uring_init(void) |
4770 | { | |
4771 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC); | |
4772 | return 0; | |
4773 | }; | |
4774 | __initcall(io_uring_init); |