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; | |
31b51510 JA |
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; |
1b4a51b6 | 189 | bool drain_next; |
2b188cc1 | 190 | |
75b28aff HV |
191 | /* |
192 | * Ring buffer of indices into array of io_uring_sqe, which is | |
193 | * mmapped by the application using the IORING_OFF_SQES offset. | |
194 | * | |
195 | * This indirection could e.g. be used to assign fixed | |
196 | * io_uring_sqe entries to operations and only submit them to | |
197 | * the queue when needed. | |
198 | * | |
199 | * The kernel modifies neither the indices array nor the entries | |
200 | * array. | |
201 | */ | |
202 | u32 *sq_array; | |
2b188cc1 JA |
203 | unsigned cached_sq_head; |
204 | unsigned sq_entries; | |
205 | unsigned sq_mask; | |
6c271ce2 | 206 | unsigned sq_thread_idle; |
498ccd9e | 207 | unsigned cached_sq_dropped; |
206aefde | 208 | atomic_t cached_cq_overflow; |
2b188cc1 | 209 | struct io_uring_sqe *sq_sqes; |
de0617e4 JA |
210 | |
211 | struct list_head defer_list; | |
5262f567 | 212 | struct list_head timeout_list; |
1d7bb1d5 | 213 | struct list_head cq_overflow_list; |
fcb323cc JA |
214 | |
215 | wait_queue_head_t inflight_wait; | |
2b188cc1 JA |
216 | } ____cacheline_aligned_in_smp; |
217 | ||
206aefde JA |
218 | struct io_rings *rings; |
219 | ||
2b188cc1 | 220 | /* IO offload */ |
561fb04a | 221 | struct io_wq *io_wq; |
6c271ce2 | 222 | struct task_struct *sqo_thread; /* if using sq thread polling */ |
2b188cc1 | 223 | struct mm_struct *sqo_mm; |
6c271ce2 | 224 | wait_queue_head_t sqo_wait; |
75b28aff | 225 | |
6b06314c JA |
226 | /* |
227 | * If used, fixed file set. Writers must ensure that ->refs is dead, | |
228 | * readers must ensure that ->refs is alive as long as the file* is | |
229 | * used. Only updated through io_uring_register(2). | |
230 | */ | |
65e19f54 | 231 | struct fixed_file_table *file_table; |
6b06314c JA |
232 | unsigned nr_user_files; |
233 | ||
edafccee JA |
234 | /* if used, fixed mapped user buffers */ |
235 | unsigned nr_user_bufs; | |
236 | struct io_mapped_ubuf *user_bufs; | |
237 | ||
2b188cc1 JA |
238 | struct user_struct *user; |
239 | ||
181e448d JA |
240 | struct cred *creds; |
241 | ||
206aefde JA |
242 | /* 0 is for ctx quiesce/reinit/free, 1 is for sqo_thread started */ |
243 | struct completion *completions; | |
244 | ||
0ddf92e8 JA |
245 | /* if all else fails... */ |
246 | struct io_kiocb *fallback_req; | |
247 | ||
206aefde JA |
248 | #if defined(CONFIG_UNIX) |
249 | struct socket *ring_sock; | |
250 | #endif | |
251 | ||
252 | struct { | |
253 | unsigned cached_cq_tail; | |
254 | unsigned cq_entries; | |
255 | unsigned cq_mask; | |
256 | atomic_t cq_timeouts; | |
257 | struct wait_queue_head cq_wait; | |
258 | struct fasync_struct *cq_fasync; | |
259 | struct eventfd_ctx *cq_ev_fd; | |
260 | } ____cacheline_aligned_in_smp; | |
2b188cc1 JA |
261 | |
262 | struct { | |
263 | struct mutex uring_lock; | |
264 | wait_queue_head_t wait; | |
265 | } ____cacheline_aligned_in_smp; | |
266 | ||
267 | struct { | |
268 | spinlock_t completion_lock; | |
def596e9 JA |
269 | bool poll_multi_file; |
270 | /* | |
271 | * ->poll_list is protected by the ctx->uring_lock for | |
272 | * io_uring instances that don't use IORING_SETUP_SQPOLL. | |
273 | * For SQPOLL, only the single threaded io_sq_thread() will | |
274 | * manipulate the list, hence no extra locking is needed there. | |
275 | */ | |
276 | struct list_head poll_list; | |
eac406c6 | 277 | struct rb_root cancel_tree; |
31b51510 | 278 | |
fcb323cc JA |
279 | spinlock_t inflight_lock; |
280 | struct list_head inflight_list; | |
2b188cc1 | 281 | } ____cacheline_aligned_in_smp; |
2b188cc1 JA |
282 | }; |
283 | ||
284 | struct sqe_submit { | |
285 | const struct io_uring_sqe *sqe; | |
fcb323cc JA |
286 | struct file *ring_file; |
287 | int ring_fd; | |
8776f3fa | 288 | u32 sequence; |
2b188cc1 | 289 | bool has_user; |
ba5290cc | 290 | bool in_async; |
6c271ce2 | 291 | bool needs_fixed_file; |
2b188cc1 JA |
292 | }; |
293 | ||
09bb8394 JA |
294 | /* |
295 | * First field must be the file pointer in all the | |
296 | * iocb unions! See also 'struct kiocb' in <linux/fs.h> | |
297 | */ | |
221c5eb2 JA |
298 | struct io_poll_iocb { |
299 | struct file *file; | |
300 | struct wait_queue_head *head; | |
301 | __poll_t events; | |
8c838788 | 302 | bool done; |
221c5eb2 JA |
303 | bool canceled; |
304 | struct wait_queue_entry wait; | |
305 | }; | |
306 | ||
ad8a48ac JA |
307 | struct io_timeout_data { |
308 | struct io_kiocb *req; | |
309 | struct hrtimer timer; | |
310 | struct timespec64 ts; | |
311 | enum hrtimer_mode mode; | |
cc42e0ac | 312 | u32 seq_offset; |
ad8a48ac JA |
313 | }; |
314 | ||
5262f567 JA |
315 | struct io_timeout { |
316 | struct file *file; | |
ad8a48ac | 317 | struct io_timeout_data *data; |
5262f567 JA |
318 | }; |
319 | ||
09bb8394 JA |
320 | /* |
321 | * NOTE! Each of the iocb union members has the file pointer | |
322 | * as the first entry in their struct definition. So you can | |
323 | * access the file pointer through any of the sub-structs, | |
324 | * or directly as just 'ki_filp' in this struct. | |
325 | */ | |
2b188cc1 | 326 | struct io_kiocb { |
221c5eb2 | 327 | union { |
09bb8394 | 328 | struct file *file; |
221c5eb2 JA |
329 | struct kiocb rw; |
330 | struct io_poll_iocb poll; | |
5262f567 | 331 | struct io_timeout timeout; |
221c5eb2 | 332 | }; |
2b188cc1 JA |
333 | |
334 | struct sqe_submit submit; | |
335 | ||
336 | struct io_ring_ctx *ctx; | |
eac406c6 JA |
337 | union { |
338 | struct list_head list; | |
339 | struct rb_node rb_node; | |
340 | }; | |
9e645e11 | 341 | struct list_head link_list; |
2b188cc1 | 342 | unsigned int flags; |
c16361c1 | 343 | refcount_t refs; |
8449eeda | 344 | #define REQ_F_NOWAIT 1 /* must not punt to workers */ |
def596e9 | 345 | #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */ |
6b06314c | 346 | #define REQ_F_FIXED_FILE 4 /* ctx owns file */ |
4d7dd462 | 347 | #define REQ_F_LINK_NEXT 8 /* already grabbed next link */ |
e2033e33 SB |
348 | #define REQ_F_IO_DRAIN 16 /* drain existing IO first */ |
349 | #define REQ_F_IO_DRAINED 32 /* drain done */ | |
9e645e11 | 350 | #define REQ_F_LINK 64 /* linked sqes */ |
2665abfd | 351 | #define REQ_F_LINK_TIMEOUT 128 /* has linked timeout */ |
f7b76ac9 | 352 | #define REQ_F_FAIL_LINK 256 /* fail rest of links */ |
1b4a51b6 | 353 | #define REQ_F_DRAIN_LINK 512 /* link should be fully drained */ |
5262f567 | 354 | #define REQ_F_TIMEOUT 1024 /* timeout request */ |
491381ce JA |
355 | #define REQ_F_ISREG 2048 /* regular file */ |
356 | #define REQ_F_MUST_PUNT 4096 /* must be punted even for NONBLOCK */ | |
93bd25bb | 357 | #define REQ_F_TIMEOUT_NOSEQ 8192 /* no timeout sequence */ |
fb4b3d3f LT |
358 | #define REQ_F_INFLIGHT 16384 /* on inflight list */ |
359 | #define REQ_F_COMP_LOCKED 32768 /* completion under lock */ | |
94ae5e77 | 360 | #define REQ_F_FREE_SQE 65536 /* free sqe if not async queued */ |
2b188cc1 | 361 | u64 user_data; |
9e645e11 | 362 | u32 result; |
de0617e4 | 363 | u32 sequence; |
2b188cc1 | 364 | |
fcb323cc JA |
365 | struct list_head inflight_entry; |
366 | ||
561fb04a | 367 | struct io_wq_work work; |
2b188cc1 JA |
368 | }; |
369 | ||
370 | #define IO_PLUG_THRESHOLD 2 | |
def596e9 | 371 | #define IO_IOPOLL_BATCH 8 |
2b188cc1 | 372 | |
9a56a232 JA |
373 | struct io_submit_state { |
374 | struct blk_plug plug; | |
375 | ||
2579f913 JA |
376 | /* |
377 | * io_kiocb alloc cache | |
378 | */ | |
379 | void *reqs[IO_IOPOLL_BATCH]; | |
380 | unsigned int free_reqs; | |
381 | unsigned int cur_req; | |
382 | ||
9a56a232 JA |
383 | /* |
384 | * File reference cache | |
385 | */ | |
386 | struct file *file; | |
387 | unsigned int fd; | |
388 | unsigned int has_refs; | |
389 | unsigned int used_refs; | |
390 | unsigned int ios_left; | |
391 | }; | |
392 | ||
561fb04a | 393 | static void io_wq_submit_work(struct io_wq_work **workptr); |
78e19bbe | 394 | static void io_cqring_fill_event(struct io_kiocb *req, long res); |
4fe2c963 | 395 | static void __io_free_req(struct io_kiocb *req); |
ec9c02ad | 396 | static void io_put_req(struct io_kiocb *req); |
78e19bbe | 397 | static void io_double_put_req(struct io_kiocb *req); |
978db57e | 398 | static void __io_double_put_req(struct io_kiocb *req); |
94ae5e77 JA |
399 | static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req); |
400 | static void io_queue_linked_timeout(struct io_kiocb *req); | |
de0617e4 | 401 | |
2b188cc1 JA |
402 | static struct kmem_cache *req_cachep; |
403 | ||
404 | static const struct file_operations io_uring_fops; | |
405 | ||
406 | struct sock *io_uring_get_socket(struct file *file) | |
407 | { | |
408 | #if defined(CONFIG_UNIX) | |
409 | if (file->f_op == &io_uring_fops) { | |
410 | struct io_ring_ctx *ctx = file->private_data; | |
411 | ||
412 | return ctx->ring_sock->sk; | |
413 | } | |
414 | #endif | |
415 | return NULL; | |
416 | } | |
417 | EXPORT_SYMBOL(io_uring_get_socket); | |
418 | ||
419 | static void io_ring_ctx_ref_free(struct percpu_ref *ref) | |
420 | { | |
421 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
422 | ||
206aefde | 423 | complete(&ctx->completions[0]); |
2b188cc1 JA |
424 | } |
425 | ||
426 | static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) | |
427 | { | |
428 | struct io_ring_ctx *ctx; | |
429 | ||
430 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
431 | if (!ctx) | |
432 | return NULL; | |
433 | ||
0ddf92e8 JA |
434 | ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL); |
435 | if (!ctx->fallback_req) | |
436 | goto err; | |
437 | ||
206aefde JA |
438 | ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL); |
439 | if (!ctx->completions) | |
440 | goto err; | |
441 | ||
21482896 | 442 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
206aefde JA |
443 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) |
444 | goto err; | |
2b188cc1 JA |
445 | |
446 | ctx->flags = p->flags; | |
447 | init_waitqueue_head(&ctx->cq_wait); | |
1d7bb1d5 | 448 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
206aefde JA |
449 | init_completion(&ctx->completions[0]); |
450 | init_completion(&ctx->completions[1]); | |
2b188cc1 JA |
451 | mutex_init(&ctx->uring_lock); |
452 | init_waitqueue_head(&ctx->wait); | |
453 | spin_lock_init(&ctx->completion_lock); | |
def596e9 | 454 | INIT_LIST_HEAD(&ctx->poll_list); |
eac406c6 | 455 | ctx->cancel_tree = RB_ROOT; |
de0617e4 | 456 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 457 | INIT_LIST_HEAD(&ctx->timeout_list); |
fcb323cc JA |
458 | init_waitqueue_head(&ctx->inflight_wait); |
459 | spin_lock_init(&ctx->inflight_lock); | |
460 | INIT_LIST_HEAD(&ctx->inflight_list); | |
2b188cc1 | 461 | return ctx; |
206aefde | 462 | err: |
0ddf92e8 JA |
463 | if (ctx->fallback_req) |
464 | kmem_cache_free(req_cachep, ctx->fallback_req); | |
206aefde JA |
465 | kfree(ctx->completions); |
466 | kfree(ctx); | |
467 | return NULL; | |
2b188cc1 JA |
468 | } |
469 | ||
9d858b21 | 470 | static inline bool __req_need_defer(struct io_kiocb *req) |
7adf4eaf | 471 | { |
a197f664 JL |
472 | struct io_ring_ctx *ctx = req->ctx; |
473 | ||
498ccd9e JA |
474 | return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped |
475 | + atomic_read(&ctx->cached_cq_overflow); | |
7adf4eaf JA |
476 | } |
477 | ||
9d858b21 | 478 | static inline bool req_need_defer(struct io_kiocb *req) |
de0617e4 | 479 | { |
9d858b21 BL |
480 | if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) == REQ_F_IO_DRAIN) |
481 | return __req_need_defer(req); | |
de0617e4 | 482 | |
9d858b21 | 483 | return false; |
de0617e4 JA |
484 | } |
485 | ||
7adf4eaf | 486 | static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx) |
de0617e4 JA |
487 | { |
488 | struct io_kiocb *req; | |
489 | ||
7adf4eaf | 490 | req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list); |
9d858b21 | 491 | if (req && !req_need_defer(req)) { |
de0617e4 JA |
492 | list_del_init(&req->list); |
493 | return req; | |
494 | } | |
495 | ||
496 | return NULL; | |
497 | } | |
498 | ||
5262f567 JA |
499 | static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx) |
500 | { | |
7adf4eaf JA |
501 | struct io_kiocb *req; |
502 | ||
503 | req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list); | |
93bd25bb JA |
504 | if (req) { |
505 | if (req->flags & REQ_F_TIMEOUT_NOSEQ) | |
506 | return NULL; | |
fb4b3d3f | 507 | if (!__req_need_defer(req)) { |
93bd25bb JA |
508 | list_del_init(&req->list); |
509 | return req; | |
510 | } | |
7adf4eaf JA |
511 | } |
512 | ||
513 | return NULL; | |
5262f567 JA |
514 | } |
515 | ||
de0617e4 | 516 | static void __io_commit_cqring(struct io_ring_ctx *ctx) |
2b188cc1 | 517 | { |
75b28aff | 518 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 519 | |
75b28aff | 520 | if (ctx->cached_cq_tail != READ_ONCE(rings->cq.tail)) { |
2b188cc1 | 521 | /* order cqe stores with ring update */ |
75b28aff | 522 | smp_store_release(&rings->cq.tail, ctx->cached_cq_tail); |
2b188cc1 | 523 | |
2b188cc1 JA |
524 | if (wq_has_sleeper(&ctx->cq_wait)) { |
525 | wake_up_interruptible(&ctx->cq_wait); | |
526 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); | |
527 | } | |
528 | } | |
529 | } | |
530 | ||
561fb04a | 531 | static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe) |
18d9be1a | 532 | { |
561fb04a JA |
533 | u8 opcode = READ_ONCE(sqe->opcode); |
534 | ||
535 | return !(opcode == IORING_OP_READ_FIXED || | |
536 | opcode == IORING_OP_WRITE_FIXED); | |
537 | } | |
538 | ||
94ae5e77 JA |
539 | static inline bool io_prep_async_work(struct io_kiocb *req, |
540 | struct io_kiocb **link) | |
18d9be1a | 541 | { |
561fb04a | 542 | bool do_hashed = false; |
54a91f3b | 543 | |
6cc47d1d JA |
544 | if (req->submit.sqe) { |
545 | switch (req->submit.sqe->opcode) { | |
546 | case IORING_OP_WRITEV: | |
547 | case IORING_OP_WRITE_FIXED: | |
561fb04a | 548 | do_hashed = true; |
5f8fd2d3 JA |
549 | /* fall-through */ |
550 | case IORING_OP_READV: | |
551 | case IORING_OP_READ_FIXED: | |
552 | case IORING_OP_SENDMSG: | |
553 | case IORING_OP_RECVMSG: | |
554 | case IORING_OP_ACCEPT: | |
555 | case IORING_OP_POLL_ADD: | |
f8e85cf2 | 556 | case IORING_OP_CONNECT: |
5f8fd2d3 JA |
557 | /* |
558 | * We know REQ_F_ISREG is not set on some of these | |
559 | * opcodes, but this enables us to keep the check in | |
560 | * just one place. | |
561 | */ | |
562 | if (!(req->flags & REQ_F_ISREG)) | |
563 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
6cc47d1d JA |
564 | break; |
565 | } | |
561fb04a JA |
566 | if (io_sqe_needs_user(req->submit.sqe)) |
567 | req->work.flags |= IO_WQ_WORK_NEEDS_USER; | |
54a91f3b JA |
568 | } |
569 | ||
94ae5e77 | 570 | *link = io_prep_linked_timeout(req); |
561fb04a JA |
571 | return do_hashed; |
572 | } | |
573 | ||
a197f664 | 574 | static inline void io_queue_async_work(struct io_kiocb *req) |
561fb04a | 575 | { |
a197f664 | 576 | struct io_ring_ctx *ctx = req->ctx; |
94ae5e77 JA |
577 | struct io_kiocb *link; |
578 | bool do_hashed; | |
579 | ||
580 | do_hashed = io_prep_async_work(req, &link); | |
561fb04a JA |
581 | |
582 | trace_io_uring_queue_async_work(ctx, do_hashed, req, &req->work, | |
583 | req->flags); | |
584 | if (!do_hashed) { | |
585 | io_wq_enqueue(ctx->io_wq, &req->work); | |
586 | } else { | |
587 | io_wq_enqueue_hashed(ctx->io_wq, &req->work, | |
588 | file_inode(req->file)); | |
589 | } | |
94ae5e77 JA |
590 | |
591 | if (link) | |
592 | io_queue_linked_timeout(link); | |
18d9be1a JA |
593 | } |
594 | ||
5262f567 JA |
595 | static void io_kill_timeout(struct io_kiocb *req) |
596 | { | |
597 | int ret; | |
598 | ||
ad8a48ac | 599 | ret = hrtimer_try_to_cancel(&req->timeout.data->timer); |
5262f567 JA |
600 | if (ret != -1) { |
601 | atomic_inc(&req->ctx->cq_timeouts); | |
842f9612 | 602 | list_del_init(&req->list); |
78e19bbe | 603 | io_cqring_fill_event(req, 0); |
ec9c02ad | 604 | io_put_req(req); |
5262f567 JA |
605 | } |
606 | } | |
607 | ||
608 | static void io_kill_timeouts(struct io_ring_ctx *ctx) | |
609 | { | |
610 | struct io_kiocb *req, *tmp; | |
611 | ||
612 | spin_lock_irq(&ctx->completion_lock); | |
613 | list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list) | |
614 | io_kill_timeout(req); | |
615 | spin_unlock_irq(&ctx->completion_lock); | |
616 | } | |
617 | ||
de0617e4 JA |
618 | static void io_commit_cqring(struct io_ring_ctx *ctx) |
619 | { | |
620 | struct io_kiocb *req; | |
621 | ||
5262f567 JA |
622 | while ((req = io_get_timeout_req(ctx)) != NULL) |
623 | io_kill_timeout(req); | |
624 | ||
de0617e4 JA |
625 | __io_commit_cqring(ctx); |
626 | ||
627 | while ((req = io_get_deferred_req(ctx)) != NULL) { | |
628 | req->flags |= REQ_F_IO_DRAINED; | |
a197f664 | 629 | io_queue_async_work(req); |
de0617e4 JA |
630 | } |
631 | } | |
632 | ||
2b188cc1 JA |
633 | static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx) |
634 | { | |
75b28aff | 635 | struct io_rings *rings = ctx->rings; |
2b188cc1 JA |
636 | unsigned tail; |
637 | ||
638 | tail = ctx->cached_cq_tail; | |
115e12e5 SB |
639 | /* |
640 | * writes to the cq entry need to come after reading head; the | |
641 | * control dependency is enough as we're using WRITE_ONCE to | |
642 | * fill the cq entry | |
643 | */ | |
75b28aff | 644 | if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries) |
2b188cc1 JA |
645 | return NULL; |
646 | ||
647 | ctx->cached_cq_tail++; | |
75b28aff | 648 | return &rings->cqes[tail & ctx->cq_mask]; |
2b188cc1 JA |
649 | } |
650 | ||
1d7bb1d5 JA |
651 | static void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
652 | { | |
653 | if (waitqueue_active(&ctx->wait)) | |
654 | wake_up(&ctx->wait); | |
655 | if (waitqueue_active(&ctx->sqo_wait)) | |
656 | wake_up(&ctx->sqo_wait); | |
657 | if (ctx->cq_ev_fd) | |
658 | eventfd_signal(ctx->cq_ev_fd, 1); | |
659 | } | |
660 | ||
c4a2ed72 JA |
661 | /* Returns true if there are no backlogged entries after the flush */ |
662 | static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force) | |
1d7bb1d5 JA |
663 | { |
664 | struct io_rings *rings = ctx->rings; | |
665 | struct io_uring_cqe *cqe; | |
666 | struct io_kiocb *req; | |
667 | unsigned long flags; | |
668 | LIST_HEAD(list); | |
669 | ||
670 | if (!force) { | |
671 | if (list_empty_careful(&ctx->cq_overflow_list)) | |
c4a2ed72 | 672 | return true; |
1d7bb1d5 JA |
673 | if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) == |
674 | rings->cq_ring_entries)) | |
c4a2ed72 | 675 | return false; |
1d7bb1d5 JA |
676 | } |
677 | ||
678 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
679 | ||
680 | /* if force is set, the ring is going away. always drop after that */ | |
681 | if (force) | |
682 | ctx->cq_overflow_flushed = true; | |
683 | ||
c4a2ed72 | 684 | cqe = NULL; |
1d7bb1d5 JA |
685 | while (!list_empty(&ctx->cq_overflow_list)) { |
686 | cqe = io_get_cqring(ctx); | |
687 | if (!cqe && !force) | |
688 | break; | |
689 | ||
690 | req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb, | |
691 | list); | |
692 | list_move(&req->list, &list); | |
693 | if (cqe) { | |
694 | WRITE_ONCE(cqe->user_data, req->user_data); | |
695 | WRITE_ONCE(cqe->res, req->result); | |
696 | WRITE_ONCE(cqe->flags, 0); | |
697 | } else { | |
698 | WRITE_ONCE(ctx->rings->cq_overflow, | |
699 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
700 | } | |
701 | } | |
702 | ||
703 | io_commit_cqring(ctx); | |
704 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
705 | io_cqring_ev_posted(ctx); | |
706 | ||
707 | while (!list_empty(&list)) { | |
708 | req = list_first_entry(&list, struct io_kiocb, list); | |
709 | list_del(&req->list); | |
ec9c02ad | 710 | io_put_req(req); |
1d7bb1d5 | 711 | } |
c4a2ed72 JA |
712 | |
713 | return cqe != NULL; | |
1d7bb1d5 JA |
714 | } |
715 | ||
78e19bbe | 716 | static void io_cqring_fill_event(struct io_kiocb *req, long res) |
2b188cc1 | 717 | { |
78e19bbe | 718 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
719 | struct io_uring_cqe *cqe; |
720 | ||
78e19bbe | 721 | trace_io_uring_complete(ctx, req->user_data, res); |
51c3ff62 | 722 | |
2b188cc1 JA |
723 | /* |
724 | * If we can't get a cq entry, userspace overflowed the | |
725 | * submission (by quite a lot). Increment the overflow count in | |
726 | * the ring. | |
727 | */ | |
728 | cqe = io_get_cqring(ctx); | |
1d7bb1d5 | 729 | if (likely(cqe)) { |
78e19bbe | 730 | WRITE_ONCE(cqe->user_data, req->user_data); |
2b188cc1 | 731 | WRITE_ONCE(cqe->res, res); |
c71ffb67 | 732 | WRITE_ONCE(cqe->flags, 0); |
1d7bb1d5 | 733 | } else if (ctx->cq_overflow_flushed) { |
498ccd9e JA |
734 | WRITE_ONCE(ctx->rings->cq_overflow, |
735 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
1d7bb1d5 JA |
736 | } else { |
737 | refcount_inc(&req->refs); | |
738 | req->result = res; | |
739 | list_add_tail(&req->list, &ctx->cq_overflow_list); | |
2b188cc1 JA |
740 | } |
741 | } | |
742 | ||
78e19bbe | 743 | static void io_cqring_add_event(struct io_kiocb *req, long res) |
2b188cc1 | 744 | { |
78e19bbe | 745 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
746 | unsigned long flags; |
747 | ||
748 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
78e19bbe | 749 | io_cqring_fill_event(req, res); |
2b188cc1 JA |
750 | io_commit_cqring(ctx); |
751 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
752 | ||
8c838788 | 753 | io_cqring_ev_posted(ctx); |
2b188cc1 JA |
754 | } |
755 | ||
0ddf92e8 JA |
756 | static inline bool io_is_fallback_req(struct io_kiocb *req) |
757 | { | |
758 | return req == (struct io_kiocb *) | |
759 | ((unsigned long) req->ctx->fallback_req & ~1UL); | |
760 | } | |
761 | ||
762 | static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx) | |
763 | { | |
764 | struct io_kiocb *req; | |
765 | ||
766 | req = ctx->fallback_req; | |
767 | if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req)) | |
768 | return req; | |
769 | ||
770 | return NULL; | |
771 | } | |
772 | ||
2579f913 JA |
773 | static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx, |
774 | struct io_submit_state *state) | |
2b188cc1 | 775 | { |
fd6fab2c | 776 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
2b188cc1 JA |
777 | struct io_kiocb *req; |
778 | ||
779 | if (!percpu_ref_tryget(&ctx->refs)) | |
780 | return NULL; | |
781 | ||
2579f913 | 782 | if (!state) { |
fd6fab2c | 783 | req = kmem_cache_alloc(req_cachep, gfp); |
2579f913 | 784 | if (unlikely(!req)) |
0ddf92e8 | 785 | goto fallback; |
2579f913 JA |
786 | } else if (!state->free_reqs) { |
787 | size_t sz; | |
788 | int ret; | |
789 | ||
790 | sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs)); | |
fd6fab2c JA |
791 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs); |
792 | ||
793 | /* | |
794 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
795 | * retry single alloc to be on the safe side. | |
796 | */ | |
797 | if (unlikely(ret <= 0)) { | |
798 | state->reqs[0] = kmem_cache_alloc(req_cachep, gfp); | |
799 | if (!state->reqs[0]) | |
0ddf92e8 | 800 | goto fallback; |
fd6fab2c JA |
801 | ret = 1; |
802 | } | |
2579f913 JA |
803 | state->free_reqs = ret - 1; |
804 | state->cur_req = 1; | |
805 | req = state->reqs[0]; | |
806 | } else { | |
807 | req = state->reqs[state->cur_req]; | |
808 | state->free_reqs--; | |
809 | state->cur_req++; | |
2b188cc1 JA |
810 | } |
811 | ||
0ddf92e8 | 812 | got_it: |
60c112b0 | 813 | req->file = NULL; |
2579f913 JA |
814 | req->ctx = ctx; |
815 | req->flags = 0; | |
e65ef56d JA |
816 | /* one is dropped after submission, the other at completion */ |
817 | refcount_set(&req->refs, 2); | |
9e645e11 | 818 | req->result = 0; |
561fb04a | 819 | INIT_IO_WORK(&req->work, io_wq_submit_work); |
2579f913 | 820 | return req; |
0ddf92e8 JA |
821 | fallback: |
822 | req = io_get_fallback_req(ctx); | |
823 | if (req) | |
824 | goto got_it; | |
6805b32e | 825 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
826 | return NULL; |
827 | } | |
828 | ||
def596e9 JA |
829 | static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr) |
830 | { | |
831 | if (*nr) { | |
832 | kmem_cache_free_bulk(req_cachep, *nr, reqs); | |
6805b32e | 833 | percpu_ref_put_many(&ctx->refs, *nr); |
def596e9 JA |
834 | *nr = 0; |
835 | } | |
836 | } | |
837 | ||
9e645e11 | 838 | static void __io_free_req(struct io_kiocb *req) |
2b188cc1 | 839 | { |
fcb323cc JA |
840 | struct io_ring_ctx *ctx = req->ctx; |
841 | ||
bbad27b2 PB |
842 | if (req->flags & REQ_F_FREE_SQE) |
843 | kfree(req->submit.sqe); | |
09bb8394 JA |
844 | if (req->file && !(req->flags & REQ_F_FIXED_FILE)) |
845 | fput(req->file); | |
fcb323cc JA |
846 | if (req->flags & REQ_F_INFLIGHT) { |
847 | unsigned long flags; | |
848 | ||
849 | spin_lock_irqsave(&ctx->inflight_lock, flags); | |
850 | list_del(&req->inflight_entry); | |
851 | if (waitqueue_active(&ctx->inflight_wait)) | |
852 | wake_up(&ctx->inflight_wait); | |
853 | spin_unlock_irqrestore(&ctx->inflight_lock, flags); | |
854 | } | |
ad8a48ac JA |
855 | if (req->flags & REQ_F_TIMEOUT) |
856 | kfree(req->timeout.data); | |
fcb323cc | 857 | percpu_ref_put(&ctx->refs); |
0ddf92e8 JA |
858 | if (likely(!io_is_fallback_req(req))) |
859 | kmem_cache_free(req_cachep, req); | |
860 | else | |
861 | clear_bit_unlock(0, (unsigned long *) ctx->fallback_req); | |
e65ef56d JA |
862 | } |
863 | ||
a197f664 | 864 | static bool io_link_cancel_timeout(struct io_kiocb *req) |
2665abfd | 865 | { |
a197f664 | 866 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd JA |
867 | int ret; |
868 | ||
ad8a48ac | 869 | ret = hrtimer_try_to_cancel(&req->timeout.data->timer); |
2665abfd | 870 | if (ret != -1) { |
78e19bbe | 871 | io_cqring_fill_event(req, -ECANCELED); |
2665abfd JA |
872 | io_commit_cqring(ctx); |
873 | req->flags &= ~REQ_F_LINK; | |
ec9c02ad | 874 | io_put_req(req); |
2665abfd JA |
875 | return true; |
876 | } | |
877 | ||
878 | return false; | |
e65ef56d JA |
879 | } |
880 | ||
ba816ad6 | 881 | static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr) |
9e645e11 | 882 | { |
2665abfd | 883 | struct io_ring_ctx *ctx = req->ctx; |
9e645e11 | 884 | struct io_kiocb *nxt; |
2665abfd | 885 | bool wake_ev = false; |
9e645e11 | 886 | |
4d7dd462 JA |
887 | /* Already got next link */ |
888 | if (req->flags & REQ_F_LINK_NEXT) | |
889 | return; | |
890 | ||
9e645e11 JA |
891 | /* |
892 | * The list should never be empty when we are called here. But could | |
893 | * potentially happen if the chain is messed up, check to be on the | |
894 | * safe side. | |
895 | */ | |
896 | nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list); | |
2665abfd | 897 | while (nxt) { |
76a46e06 | 898 | list_del_init(&nxt->list); |
94ae5e77 JA |
899 | |
900 | if ((req->flags & REQ_F_LINK_TIMEOUT) && | |
901 | (nxt->flags & REQ_F_TIMEOUT)) { | |
902 | wake_ev |= io_link_cancel_timeout(nxt); | |
903 | nxt = list_first_entry_or_null(&req->link_list, | |
904 | struct io_kiocb, list); | |
905 | req->flags &= ~REQ_F_LINK_TIMEOUT; | |
906 | continue; | |
907 | } | |
9e645e11 JA |
908 | if (!list_empty(&req->link_list)) { |
909 | INIT_LIST_HEAD(&nxt->link_list); | |
910 | list_splice(&req->link_list, &nxt->link_list); | |
911 | nxt->flags |= REQ_F_LINK; | |
912 | } | |
913 | ||
b18fdf71 | 914 | *nxtptr = nxt; |
94ae5e77 | 915 | break; |
9e645e11 | 916 | } |
2665abfd | 917 | |
4d7dd462 | 918 | req->flags |= REQ_F_LINK_NEXT; |
2665abfd JA |
919 | if (wake_ev) |
920 | io_cqring_ev_posted(ctx); | |
9e645e11 JA |
921 | } |
922 | ||
923 | /* | |
924 | * Called if REQ_F_LINK is set, and we fail the head request | |
925 | */ | |
926 | static void io_fail_links(struct io_kiocb *req) | |
927 | { | |
2665abfd | 928 | struct io_ring_ctx *ctx = req->ctx; |
9e645e11 | 929 | struct io_kiocb *link; |
2665abfd JA |
930 | unsigned long flags; |
931 | ||
932 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
9e645e11 JA |
933 | |
934 | while (!list_empty(&req->link_list)) { | |
935 | link = list_first_entry(&req->link_list, struct io_kiocb, list); | |
2665abfd | 936 | list_del_init(&link->list); |
9e645e11 | 937 | |
c826bd7a | 938 | trace_io_uring_fail_link(req, link); |
2665abfd JA |
939 | |
940 | if ((req->flags & REQ_F_LINK_TIMEOUT) && | |
941 | link->submit.sqe->opcode == IORING_OP_LINK_TIMEOUT) { | |
a197f664 | 942 | io_link_cancel_timeout(link); |
2665abfd | 943 | } else { |
78e19bbe | 944 | io_cqring_fill_event(link, -ECANCELED); |
978db57e | 945 | __io_double_put_req(link); |
2665abfd | 946 | } |
5d960724 | 947 | req->flags &= ~REQ_F_LINK_TIMEOUT; |
9e645e11 | 948 | } |
2665abfd JA |
949 | |
950 | io_commit_cqring(ctx); | |
951 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
952 | io_cqring_ev_posted(ctx); | |
9e645e11 JA |
953 | } |
954 | ||
4d7dd462 | 955 | static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt) |
9e645e11 | 956 | { |
4d7dd462 | 957 | if (likely(!(req->flags & REQ_F_LINK))) |
2665abfd | 958 | return; |
2665abfd | 959 | |
9e645e11 JA |
960 | /* |
961 | * If LINK is set, we have dependent requests in this chain. If we | |
962 | * didn't fail this request, queue the first one up, moving any other | |
963 | * dependencies to the next request. In case of failure, fail the rest | |
964 | * of the chain. | |
965 | */ | |
2665abfd JA |
966 | if (req->flags & REQ_F_FAIL_LINK) { |
967 | io_fail_links(req); | |
7c9e7f0f JA |
968 | } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) == |
969 | REQ_F_LINK_TIMEOUT) { | |
2665abfd JA |
970 | struct io_ring_ctx *ctx = req->ctx; |
971 | unsigned long flags; | |
972 | ||
973 | /* | |
974 | * If this is a timeout link, we could be racing with the | |
975 | * timeout timer. Grab the completion lock for this case to | |
7c9e7f0f | 976 | * protect against that. |
2665abfd JA |
977 | */ |
978 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
979 | io_req_link_next(req, nxt); | |
980 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
981 | } else { | |
982 | io_req_link_next(req, nxt); | |
9e645e11 | 983 | } |
4d7dd462 | 984 | } |
9e645e11 | 985 | |
c69f8dbe JL |
986 | static void io_free_req(struct io_kiocb *req) |
987 | { | |
944e58bf PB |
988 | struct io_kiocb *nxt = NULL; |
989 | ||
990 | io_req_find_next(req, &nxt); | |
70cf9f32 | 991 | __io_free_req(req); |
944e58bf PB |
992 | |
993 | if (nxt) | |
994 | io_queue_async_work(nxt); | |
c69f8dbe JL |
995 | } |
996 | ||
ba816ad6 JA |
997 | /* |
998 | * Drop reference to request, return next in chain (if there is one) if this | |
999 | * was the last reference to this request. | |
1000 | */ | |
f9bd67f6 | 1001 | __attribute__((nonnull)) |
ec9c02ad | 1002 | static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr) |
e65ef56d | 1003 | { |
f9bd67f6 | 1004 | io_req_find_next(req, nxtptr); |
4d7dd462 | 1005 | |
e65ef56d | 1006 | if (refcount_dec_and_test(&req->refs)) |
4d7dd462 | 1007 | __io_free_req(req); |
2b188cc1 JA |
1008 | } |
1009 | ||
e65ef56d JA |
1010 | static void io_put_req(struct io_kiocb *req) |
1011 | { | |
1012 | if (refcount_dec_and_test(&req->refs)) | |
1013 | io_free_req(req); | |
2b188cc1 JA |
1014 | } |
1015 | ||
978db57e JA |
1016 | /* |
1017 | * Must only be used if we don't need to care about links, usually from | |
1018 | * within the completion handling itself. | |
1019 | */ | |
1020 | static void __io_double_put_req(struct io_kiocb *req) | |
78e19bbe JA |
1021 | { |
1022 | /* drop both submit and complete references */ | |
1023 | if (refcount_sub_and_test(2, &req->refs)) | |
1024 | __io_free_req(req); | |
1025 | } | |
1026 | ||
978db57e JA |
1027 | static void io_double_put_req(struct io_kiocb *req) |
1028 | { | |
1029 | /* drop both submit and complete references */ | |
1030 | if (refcount_sub_and_test(2, &req->refs)) | |
1031 | io_free_req(req); | |
1032 | } | |
1033 | ||
1d7bb1d5 | 1034 | static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush) |
a3a0e43f | 1035 | { |
84f97dc2 JA |
1036 | struct io_rings *rings = ctx->rings; |
1037 | ||
1d7bb1d5 JA |
1038 | /* |
1039 | * noflush == true is from the waitqueue handler, just ensure we wake | |
1040 | * up the task, and the next invocation will flush the entries. We | |
1041 | * cannot safely to it from here. | |
1042 | */ | |
1043 | if (noflush && !list_empty(&ctx->cq_overflow_list)) | |
1044 | return -1U; | |
1045 | ||
1046 | io_cqring_overflow_flush(ctx, false); | |
1047 | ||
a3a0e43f JA |
1048 | /* See comment at the top of this file */ |
1049 | smp_rmb(); | |
75b28aff | 1050 | return READ_ONCE(rings->cq.tail) - READ_ONCE(rings->cq.head); |
a3a0e43f JA |
1051 | } |
1052 | ||
fb5ccc98 PB |
1053 | static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx) |
1054 | { | |
1055 | struct io_rings *rings = ctx->rings; | |
1056 | ||
1057 | /* make sure SQ entry isn't read before tail */ | |
1058 | return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head; | |
1059 | } | |
1060 | ||
def596e9 JA |
1061 | /* |
1062 | * Find and free completed poll iocbs | |
1063 | */ | |
1064 | static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1065 | struct list_head *done) | |
1066 | { | |
1067 | void *reqs[IO_IOPOLL_BATCH]; | |
1068 | struct io_kiocb *req; | |
09bb8394 | 1069 | int to_free; |
def596e9 | 1070 | |
09bb8394 | 1071 | to_free = 0; |
def596e9 JA |
1072 | while (!list_empty(done)) { |
1073 | req = list_first_entry(done, struct io_kiocb, list); | |
1074 | list_del(&req->list); | |
1075 | ||
78e19bbe | 1076 | io_cqring_fill_event(req, req->result); |
def596e9 JA |
1077 | (*nr_events)++; |
1078 | ||
09bb8394 JA |
1079 | if (refcount_dec_and_test(&req->refs)) { |
1080 | /* If we're not using fixed files, we have to pair the | |
1081 | * completion part with the file put. Use regular | |
1082 | * completions for those, only batch free for fixed | |
9e645e11 | 1083 | * file and non-linked commands. |
09bb8394 | 1084 | */ |
bbad27b2 PB |
1085 | if (((req->flags & |
1086 | (REQ_F_FIXED_FILE|REQ_F_LINK|REQ_F_FREE_SQE)) == | |
0ddf92e8 | 1087 | REQ_F_FIXED_FILE) && !io_is_fallback_req(req)) { |
09bb8394 JA |
1088 | reqs[to_free++] = req; |
1089 | if (to_free == ARRAY_SIZE(reqs)) | |
1090 | io_free_req_many(ctx, reqs, &to_free); | |
6b06314c | 1091 | } else { |
09bb8394 | 1092 | io_free_req(req); |
6b06314c | 1093 | } |
9a56a232 | 1094 | } |
def596e9 | 1095 | } |
def596e9 | 1096 | |
09bb8394 | 1097 | io_commit_cqring(ctx); |
def596e9 JA |
1098 | io_free_req_many(ctx, reqs, &to_free); |
1099 | } | |
1100 | ||
1101 | static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1102 | long min) | |
1103 | { | |
1104 | struct io_kiocb *req, *tmp; | |
1105 | LIST_HEAD(done); | |
1106 | bool spin; | |
1107 | int ret; | |
1108 | ||
1109 | /* | |
1110 | * Only spin for completions if we don't have multiple devices hanging | |
1111 | * off our complete list, and we're under the requested amount. | |
1112 | */ | |
1113 | spin = !ctx->poll_multi_file && *nr_events < min; | |
1114 | ||
1115 | ret = 0; | |
1116 | list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) { | |
1117 | struct kiocb *kiocb = &req->rw; | |
1118 | ||
1119 | /* | |
1120 | * Move completed entries to our local list. If we find a | |
1121 | * request that requires polling, break out and complete | |
1122 | * the done list first, if we have entries there. | |
1123 | */ | |
1124 | if (req->flags & REQ_F_IOPOLL_COMPLETED) { | |
1125 | list_move_tail(&req->list, &done); | |
1126 | continue; | |
1127 | } | |
1128 | if (!list_empty(&done)) | |
1129 | break; | |
1130 | ||
1131 | ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin); | |
1132 | if (ret < 0) | |
1133 | break; | |
1134 | ||
1135 | if (ret && spin) | |
1136 | spin = false; | |
1137 | ret = 0; | |
1138 | } | |
1139 | ||
1140 | if (!list_empty(&done)) | |
1141 | io_iopoll_complete(ctx, nr_events, &done); | |
1142 | ||
1143 | return ret; | |
1144 | } | |
1145 | ||
1146 | /* | |
1147 | * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a | |
1148 | * non-spinning poll check - we'll still enter the driver poll loop, but only | |
1149 | * as a non-spinning completion check. | |
1150 | */ | |
1151 | static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1152 | long min) | |
1153 | { | |
08f5439f | 1154 | while (!list_empty(&ctx->poll_list) && !need_resched()) { |
def596e9 JA |
1155 | int ret; |
1156 | ||
1157 | ret = io_do_iopoll(ctx, nr_events, min); | |
1158 | if (ret < 0) | |
1159 | return ret; | |
1160 | if (!min || *nr_events >= min) | |
1161 | return 0; | |
1162 | } | |
1163 | ||
1164 | return 1; | |
1165 | } | |
1166 | ||
1167 | /* | |
1168 | * We can't just wait for polled events to come to us, we have to actively | |
1169 | * find and complete them. | |
1170 | */ | |
1171 | static void io_iopoll_reap_events(struct io_ring_ctx *ctx) | |
1172 | { | |
1173 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
1174 | return; | |
1175 | ||
1176 | mutex_lock(&ctx->uring_lock); | |
1177 | while (!list_empty(&ctx->poll_list)) { | |
1178 | unsigned int nr_events = 0; | |
1179 | ||
1180 | io_iopoll_getevents(ctx, &nr_events, 1); | |
08f5439f JA |
1181 | |
1182 | /* | |
1183 | * Ensure we allow local-to-the-cpu processing to take place, | |
1184 | * in this case we need to ensure that we reap all events. | |
1185 | */ | |
1186 | cond_resched(); | |
def596e9 JA |
1187 | } |
1188 | mutex_unlock(&ctx->uring_lock); | |
1189 | } | |
1190 | ||
2b2ed975 JA |
1191 | static int __io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, |
1192 | long min) | |
def596e9 | 1193 | { |
2b2ed975 | 1194 | int iters = 0, ret = 0; |
500f9fba | 1195 | |
def596e9 JA |
1196 | do { |
1197 | int tmin = 0; | |
1198 | ||
a3a0e43f JA |
1199 | /* |
1200 | * Don't enter poll loop if we already have events pending. | |
1201 | * If we do, we can potentially be spinning for commands that | |
1202 | * already triggered a CQE (eg in error). | |
1203 | */ | |
1d7bb1d5 | 1204 | if (io_cqring_events(ctx, false)) |
a3a0e43f JA |
1205 | break; |
1206 | ||
500f9fba JA |
1207 | /* |
1208 | * If a submit got punted to a workqueue, we can have the | |
1209 | * application entering polling for a command before it gets | |
1210 | * issued. That app will hold the uring_lock for the duration | |
1211 | * of the poll right here, so we need to take a breather every | |
1212 | * now and then to ensure that the issue has a chance to add | |
1213 | * the poll to the issued list. Otherwise we can spin here | |
1214 | * forever, while the workqueue is stuck trying to acquire the | |
1215 | * very same mutex. | |
1216 | */ | |
1217 | if (!(++iters & 7)) { | |
1218 | mutex_unlock(&ctx->uring_lock); | |
1219 | mutex_lock(&ctx->uring_lock); | |
1220 | } | |
1221 | ||
def596e9 JA |
1222 | if (*nr_events < min) |
1223 | tmin = min - *nr_events; | |
1224 | ||
1225 | ret = io_iopoll_getevents(ctx, nr_events, tmin); | |
1226 | if (ret <= 0) | |
1227 | break; | |
1228 | ret = 0; | |
1229 | } while (min && !*nr_events && !need_resched()); | |
1230 | ||
2b2ed975 JA |
1231 | return ret; |
1232 | } | |
1233 | ||
1234 | static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, | |
1235 | long min) | |
1236 | { | |
1237 | int ret; | |
1238 | ||
1239 | /* | |
1240 | * We disallow the app entering submit/complete with polling, but we | |
1241 | * still need to lock the ring to prevent racing with polled issue | |
1242 | * that got punted to a workqueue. | |
1243 | */ | |
1244 | mutex_lock(&ctx->uring_lock); | |
1245 | ret = __io_iopoll_check(ctx, nr_events, min); | |
500f9fba | 1246 | mutex_unlock(&ctx->uring_lock); |
def596e9 JA |
1247 | return ret; |
1248 | } | |
1249 | ||
491381ce | 1250 | static void kiocb_end_write(struct io_kiocb *req) |
2b188cc1 | 1251 | { |
491381ce JA |
1252 | /* |
1253 | * Tell lockdep we inherited freeze protection from submission | |
1254 | * thread. | |
1255 | */ | |
1256 | if (req->flags & REQ_F_ISREG) { | |
1257 | struct inode *inode = file_inode(req->file); | |
2b188cc1 | 1258 | |
491381ce | 1259 | __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE); |
2b188cc1 | 1260 | } |
491381ce | 1261 | file_end_write(req->file); |
2b188cc1 JA |
1262 | } |
1263 | ||
ba816ad6 | 1264 | static void io_complete_rw_common(struct kiocb *kiocb, long res) |
2b188cc1 JA |
1265 | { |
1266 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
1267 | ||
491381ce JA |
1268 | if (kiocb->ki_flags & IOCB_WRITE) |
1269 | kiocb_end_write(req); | |
2b188cc1 | 1270 | |
9e645e11 JA |
1271 | if ((req->flags & REQ_F_LINK) && res != req->result) |
1272 | req->flags |= REQ_F_FAIL_LINK; | |
78e19bbe | 1273 | io_cqring_add_event(req, res); |
ba816ad6 JA |
1274 | } |
1275 | ||
1276 | static void io_complete_rw(struct kiocb *kiocb, long res, long res2) | |
1277 | { | |
1278 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
1279 | ||
1280 | io_complete_rw_common(kiocb, res); | |
e65ef56d | 1281 | io_put_req(req); |
2b188cc1 JA |
1282 | } |
1283 | ||
ba816ad6 JA |
1284 | static struct io_kiocb *__io_complete_rw(struct kiocb *kiocb, long res) |
1285 | { | |
1286 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
ec9c02ad | 1287 | struct io_kiocb *nxt = NULL; |
ba816ad6 JA |
1288 | |
1289 | io_complete_rw_common(kiocb, res); | |
ec9c02ad JL |
1290 | io_put_req_find_next(req, &nxt); |
1291 | ||
1292 | return nxt; | |
2b188cc1 JA |
1293 | } |
1294 | ||
def596e9 JA |
1295 | static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2) |
1296 | { | |
1297 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
1298 | ||
491381ce JA |
1299 | if (kiocb->ki_flags & IOCB_WRITE) |
1300 | kiocb_end_write(req); | |
def596e9 | 1301 | |
9e645e11 JA |
1302 | if ((req->flags & REQ_F_LINK) && res != req->result) |
1303 | req->flags |= REQ_F_FAIL_LINK; | |
1304 | req->result = res; | |
def596e9 JA |
1305 | if (res != -EAGAIN) |
1306 | req->flags |= REQ_F_IOPOLL_COMPLETED; | |
1307 | } | |
1308 | ||
1309 | /* | |
1310 | * After the iocb has been issued, it's safe to be found on the poll list. | |
1311 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
1312 | * find it from a io_iopoll_getevents() thread before the issuer is done | |
1313 | * accessing the kiocb cookie. | |
1314 | */ | |
1315 | static void io_iopoll_req_issued(struct io_kiocb *req) | |
1316 | { | |
1317 | struct io_ring_ctx *ctx = req->ctx; | |
1318 | ||
1319 | /* | |
1320 | * Track whether we have multiple files in our lists. This will impact | |
1321 | * how we do polling eventually, not spinning if we're on potentially | |
1322 | * different devices. | |
1323 | */ | |
1324 | if (list_empty(&ctx->poll_list)) { | |
1325 | ctx->poll_multi_file = false; | |
1326 | } else if (!ctx->poll_multi_file) { | |
1327 | struct io_kiocb *list_req; | |
1328 | ||
1329 | list_req = list_first_entry(&ctx->poll_list, struct io_kiocb, | |
1330 | list); | |
1331 | if (list_req->rw.ki_filp != req->rw.ki_filp) | |
1332 | ctx->poll_multi_file = true; | |
1333 | } | |
1334 | ||
1335 | /* | |
1336 | * For fast devices, IO may have already completed. If it has, add | |
1337 | * it to the front so we find it first. | |
1338 | */ | |
1339 | if (req->flags & REQ_F_IOPOLL_COMPLETED) | |
1340 | list_add(&req->list, &ctx->poll_list); | |
1341 | else | |
1342 | list_add_tail(&req->list, &ctx->poll_list); | |
1343 | } | |
1344 | ||
3d6770fb | 1345 | static void io_file_put(struct io_submit_state *state) |
9a56a232 | 1346 | { |
3d6770fb | 1347 | if (state->file) { |
9a56a232 JA |
1348 | int diff = state->has_refs - state->used_refs; |
1349 | ||
1350 | if (diff) | |
1351 | fput_many(state->file, diff); | |
1352 | state->file = NULL; | |
1353 | } | |
1354 | } | |
1355 | ||
1356 | /* | |
1357 | * Get as many references to a file as we have IOs left in this submission, | |
1358 | * assuming most submissions are for one file, or at least that each file | |
1359 | * has more than one submission. | |
1360 | */ | |
1361 | static struct file *io_file_get(struct io_submit_state *state, int fd) | |
1362 | { | |
1363 | if (!state) | |
1364 | return fget(fd); | |
1365 | ||
1366 | if (state->file) { | |
1367 | if (state->fd == fd) { | |
1368 | state->used_refs++; | |
1369 | state->ios_left--; | |
1370 | return state->file; | |
1371 | } | |
3d6770fb | 1372 | io_file_put(state); |
9a56a232 JA |
1373 | } |
1374 | state->file = fget_many(fd, state->ios_left); | |
1375 | if (!state->file) | |
1376 | return NULL; | |
1377 | ||
1378 | state->fd = fd; | |
1379 | state->has_refs = state->ios_left; | |
1380 | state->used_refs = 1; | |
1381 | state->ios_left--; | |
1382 | return state->file; | |
1383 | } | |
1384 | ||
2b188cc1 JA |
1385 | /* |
1386 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1387 | * any file. For now, just ensure that anything potentially problematic is done | |
1388 | * inline. | |
1389 | */ | |
1390 | static bool io_file_supports_async(struct file *file) | |
1391 | { | |
1392 | umode_t mode = file_inode(file)->i_mode; | |
1393 | ||
1394 | if (S_ISBLK(mode) || S_ISCHR(mode)) | |
1395 | return true; | |
1396 | if (S_ISREG(mode) && file->f_op != &io_uring_fops) | |
1397 | return true; | |
1398 | ||
1399 | return false; | |
1400 | } | |
1401 | ||
267bc904 | 1402 | static int io_prep_rw(struct io_kiocb *req, bool force_nonblock) |
2b188cc1 | 1403 | { |
267bc904 | 1404 | const struct io_uring_sqe *sqe = req->submit.sqe; |
def596e9 | 1405 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 | 1406 | struct kiocb *kiocb = &req->rw; |
09bb8394 JA |
1407 | unsigned ioprio; |
1408 | int ret; | |
2b188cc1 | 1409 | |
09bb8394 JA |
1410 | if (!req->file) |
1411 | return -EBADF; | |
2b188cc1 | 1412 | |
491381ce JA |
1413 | if (S_ISREG(file_inode(req->file)->i_mode)) |
1414 | req->flags |= REQ_F_ISREG; | |
1415 | ||
1416 | /* | |
1417 | * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so | |
1418 | * we know to async punt it even if it was opened O_NONBLOCK | |
1419 | */ | |
1420 | if (force_nonblock && !io_file_supports_async(req->file)) { | |
1421 | req->flags |= REQ_F_MUST_PUNT; | |
1422 | return -EAGAIN; | |
1423 | } | |
6b06314c | 1424 | |
2b188cc1 JA |
1425 | kiocb->ki_pos = READ_ONCE(sqe->off); |
1426 | kiocb->ki_flags = iocb_flags(kiocb->ki_filp); | |
1427 | kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp)); | |
1428 | ||
1429 | ioprio = READ_ONCE(sqe->ioprio); | |
1430 | if (ioprio) { | |
1431 | ret = ioprio_check_cap(ioprio); | |
1432 | if (ret) | |
09bb8394 | 1433 | return ret; |
2b188cc1 JA |
1434 | |
1435 | kiocb->ki_ioprio = ioprio; | |
1436 | } else | |
1437 | kiocb->ki_ioprio = get_current_ioprio(); | |
1438 | ||
1439 | ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags)); | |
1440 | if (unlikely(ret)) | |
09bb8394 | 1441 | return ret; |
8449eeda SB |
1442 | |
1443 | /* don't allow async punt if RWF_NOWAIT was requested */ | |
491381ce JA |
1444 | if ((kiocb->ki_flags & IOCB_NOWAIT) || |
1445 | (req->file->f_flags & O_NONBLOCK)) | |
8449eeda SB |
1446 | req->flags |= REQ_F_NOWAIT; |
1447 | ||
1448 | if (force_nonblock) | |
2b188cc1 | 1449 | kiocb->ki_flags |= IOCB_NOWAIT; |
8449eeda | 1450 | |
def596e9 | 1451 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 JA |
1452 | if (!(kiocb->ki_flags & IOCB_DIRECT) || |
1453 | !kiocb->ki_filp->f_op->iopoll) | |
09bb8394 | 1454 | return -EOPNOTSUPP; |
2b188cc1 | 1455 | |
def596e9 JA |
1456 | kiocb->ki_flags |= IOCB_HIPRI; |
1457 | kiocb->ki_complete = io_complete_rw_iopoll; | |
6873e0bd | 1458 | req->result = 0; |
def596e9 | 1459 | } else { |
09bb8394 JA |
1460 | if (kiocb->ki_flags & IOCB_HIPRI) |
1461 | return -EINVAL; | |
def596e9 JA |
1462 | kiocb->ki_complete = io_complete_rw; |
1463 | } | |
2b188cc1 | 1464 | return 0; |
2b188cc1 JA |
1465 | } |
1466 | ||
1467 | static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) | |
1468 | { | |
1469 | switch (ret) { | |
1470 | case -EIOCBQUEUED: | |
1471 | break; | |
1472 | case -ERESTARTSYS: | |
1473 | case -ERESTARTNOINTR: | |
1474 | case -ERESTARTNOHAND: | |
1475 | case -ERESTART_RESTARTBLOCK: | |
1476 | /* | |
1477 | * We can't just restart the syscall, since previously | |
1478 | * submitted sqes may already be in progress. Just fail this | |
1479 | * IO with EINTR. | |
1480 | */ | |
1481 | ret = -EINTR; | |
1482 | /* fall through */ | |
1483 | default: | |
1484 | kiocb->ki_complete(kiocb, ret, 0); | |
1485 | } | |
1486 | } | |
1487 | ||
ba816ad6 JA |
1488 | static void kiocb_done(struct kiocb *kiocb, ssize_t ret, struct io_kiocb **nxt, |
1489 | bool in_async) | |
1490 | { | |
f9bd67f6 | 1491 | if (in_async && ret >= 0 && kiocb->ki_complete == io_complete_rw) |
ba816ad6 JA |
1492 | *nxt = __io_complete_rw(kiocb, ret); |
1493 | else | |
1494 | io_rw_done(kiocb, ret); | |
1495 | } | |
1496 | ||
edafccee JA |
1497 | static int io_import_fixed(struct io_ring_ctx *ctx, int rw, |
1498 | const struct io_uring_sqe *sqe, | |
1499 | struct iov_iter *iter) | |
1500 | { | |
1501 | size_t len = READ_ONCE(sqe->len); | |
1502 | struct io_mapped_ubuf *imu; | |
1503 | unsigned index, buf_index; | |
1504 | size_t offset; | |
1505 | u64 buf_addr; | |
1506 | ||
1507 | /* attempt to use fixed buffers without having provided iovecs */ | |
1508 | if (unlikely(!ctx->user_bufs)) | |
1509 | return -EFAULT; | |
1510 | ||
1511 | buf_index = READ_ONCE(sqe->buf_index); | |
1512 | if (unlikely(buf_index >= ctx->nr_user_bufs)) | |
1513 | return -EFAULT; | |
1514 | ||
1515 | index = array_index_nospec(buf_index, ctx->nr_user_bufs); | |
1516 | imu = &ctx->user_bufs[index]; | |
1517 | buf_addr = READ_ONCE(sqe->addr); | |
1518 | ||
1519 | /* overflow */ | |
1520 | if (buf_addr + len < buf_addr) | |
1521 | return -EFAULT; | |
1522 | /* not inside the mapped region */ | |
1523 | if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len) | |
1524 | return -EFAULT; | |
1525 | ||
1526 | /* | |
1527 | * May not be a start of buffer, set size appropriately | |
1528 | * and advance us to the beginning. | |
1529 | */ | |
1530 | offset = buf_addr - imu->ubuf; | |
1531 | iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len); | |
bd11b3a3 JA |
1532 | |
1533 | if (offset) { | |
1534 | /* | |
1535 | * Don't use iov_iter_advance() here, as it's really slow for | |
1536 | * using the latter parts of a big fixed buffer - it iterates | |
1537 | * over each segment manually. We can cheat a bit here, because | |
1538 | * we know that: | |
1539 | * | |
1540 | * 1) it's a BVEC iter, we set it up | |
1541 | * 2) all bvecs are PAGE_SIZE in size, except potentially the | |
1542 | * first and last bvec | |
1543 | * | |
1544 | * So just find our index, and adjust the iterator afterwards. | |
1545 | * If the offset is within the first bvec (or the whole first | |
1546 | * bvec, just use iov_iter_advance(). This makes it easier | |
1547 | * since we can just skip the first segment, which may not | |
1548 | * be PAGE_SIZE aligned. | |
1549 | */ | |
1550 | const struct bio_vec *bvec = imu->bvec; | |
1551 | ||
1552 | if (offset <= bvec->bv_len) { | |
1553 | iov_iter_advance(iter, offset); | |
1554 | } else { | |
1555 | unsigned long seg_skip; | |
1556 | ||
1557 | /* skip first vec */ | |
1558 | offset -= bvec->bv_len; | |
1559 | seg_skip = 1 + (offset >> PAGE_SHIFT); | |
1560 | ||
1561 | iter->bvec = bvec + seg_skip; | |
1562 | iter->nr_segs -= seg_skip; | |
99c79f66 | 1563 | iter->count -= bvec->bv_len + offset; |
bd11b3a3 | 1564 | iter->iov_offset = offset & ~PAGE_MASK; |
bd11b3a3 JA |
1565 | } |
1566 | } | |
1567 | ||
5e559561 | 1568 | return len; |
edafccee JA |
1569 | } |
1570 | ||
87e5e6da JA |
1571 | static ssize_t io_import_iovec(struct io_ring_ctx *ctx, int rw, |
1572 | const struct sqe_submit *s, struct iovec **iovec, | |
1573 | struct iov_iter *iter) | |
2b188cc1 JA |
1574 | { |
1575 | const struct io_uring_sqe *sqe = s->sqe; | |
1576 | void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
1577 | size_t sqe_len = READ_ONCE(sqe->len); | |
edafccee JA |
1578 | u8 opcode; |
1579 | ||
1580 | /* | |
1581 | * We're reading ->opcode for the second time, but the first read | |
1582 | * doesn't care whether it's _FIXED or not, so it doesn't matter | |
1583 | * whether ->opcode changes concurrently. The first read does care | |
1584 | * about whether it is a READ or a WRITE, so we don't trust this read | |
1585 | * for that purpose and instead let the caller pass in the read/write | |
1586 | * flag. | |
1587 | */ | |
1588 | opcode = READ_ONCE(sqe->opcode); | |
1589 | if (opcode == IORING_OP_READ_FIXED || | |
1590 | opcode == IORING_OP_WRITE_FIXED) { | |
87e5e6da | 1591 | ssize_t ret = io_import_fixed(ctx, rw, sqe, iter); |
edafccee JA |
1592 | *iovec = NULL; |
1593 | return ret; | |
1594 | } | |
2b188cc1 JA |
1595 | |
1596 | if (!s->has_user) | |
1597 | return -EFAULT; | |
1598 | ||
1599 | #ifdef CONFIG_COMPAT | |
1600 | if (ctx->compat) | |
1601 | return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV, | |
1602 | iovec, iter); | |
1603 | #endif | |
1604 | ||
1605 | return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter); | |
1606 | } | |
1607 | ||
31b51510 | 1608 | /* |
32960613 JA |
1609 | * For files that don't have ->read_iter() and ->write_iter(), handle them |
1610 | * by looping over ->read() or ->write() manually. | |
31b51510 | 1611 | */ |
32960613 JA |
1612 | static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb, |
1613 | struct iov_iter *iter) | |
1614 | { | |
1615 | ssize_t ret = 0; | |
1616 | ||
1617 | /* | |
1618 | * Don't support polled IO through this interface, and we can't | |
1619 | * support non-blocking either. For the latter, this just causes | |
1620 | * the kiocb to be handled from an async context. | |
1621 | */ | |
1622 | if (kiocb->ki_flags & IOCB_HIPRI) | |
1623 | return -EOPNOTSUPP; | |
1624 | if (kiocb->ki_flags & IOCB_NOWAIT) | |
1625 | return -EAGAIN; | |
1626 | ||
1627 | while (iov_iter_count(iter)) { | |
311ae9e1 | 1628 | struct iovec iovec; |
32960613 JA |
1629 | ssize_t nr; |
1630 | ||
311ae9e1 PB |
1631 | if (!iov_iter_is_bvec(iter)) { |
1632 | iovec = iov_iter_iovec(iter); | |
1633 | } else { | |
1634 | /* fixed buffers import bvec */ | |
1635 | iovec.iov_base = kmap(iter->bvec->bv_page) | |
1636 | + iter->iov_offset; | |
1637 | iovec.iov_len = min(iter->count, | |
1638 | iter->bvec->bv_len - iter->iov_offset); | |
1639 | } | |
1640 | ||
32960613 JA |
1641 | if (rw == READ) { |
1642 | nr = file->f_op->read(file, iovec.iov_base, | |
1643 | iovec.iov_len, &kiocb->ki_pos); | |
1644 | } else { | |
1645 | nr = file->f_op->write(file, iovec.iov_base, | |
1646 | iovec.iov_len, &kiocb->ki_pos); | |
1647 | } | |
1648 | ||
311ae9e1 PB |
1649 | if (iov_iter_is_bvec(iter)) |
1650 | kunmap(iter->bvec->bv_page); | |
1651 | ||
32960613 JA |
1652 | if (nr < 0) { |
1653 | if (!ret) | |
1654 | ret = nr; | |
1655 | break; | |
1656 | } | |
1657 | ret += nr; | |
1658 | if (nr != iovec.iov_len) | |
1659 | break; | |
1660 | iov_iter_advance(iter, nr); | |
1661 | } | |
1662 | ||
1663 | return ret; | |
1664 | } | |
1665 | ||
267bc904 | 1666 | static int io_read(struct io_kiocb *req, struct io_kiocb **nxt, |
8358e3a8 | 1667 | bool force_nonblock) |
2b188cc1 JA |
1668 | { |
1669 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1670 | struct kiocb *kiocb = &req->rw; | |
1671 | struct iov_iter iter; | |
1672 | struct file *file; | |
31b51510 | 1673 | size_t iov_count; |
9d93a3f5 | 1674 | ssize_t read_size, ret; |
2b188cc1 | 1675 | |
267bc904 | 1676 | ret = io_prep_rw(req, force_nonblock); |
2b188cc1 JA |
1677 | if (ret) |
1678 | return ret; | |
1679 | file = kiocb->ki_filp; | |
1680 | ||
2b188cc1 | 1681 | if (unlikely(!(file->f_mode & FMODE_READ))) |
09bb8394 | 1682 | return -EBADF; |
2b188cc1 | 1683 | |
267bc904 | 1684 | ret = io_import_iovec(req->ctx, READ, &req->submit, &iovec, &iter); |
87e5e6da | 1685 | if (ret < 0) |
09bb8394 | 1686 | return ret; |
2b188cc1 | 1687 | |
9d93a3f5 | 1688 | read_size = ret; |
9e645e11 JA |
1689 | if (req->flags & REQ_F_LINK) |
1690 | req->result = read_size; | |
1691 | ||
31b51510 JA |
1692 | iov_count = iov_iter_count(&iter); |
1693 | ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count); | |
2b188cc1 JA |
1694 | if (!ret) { |
1695 | ssize_t ret2; | |
1696 | ||
32960613 JA |
1697 | if (file->f_op->read_iter) |
1698 | ret2 = call_read_iter(file, kiocb, &iter); | |
1699 | else | |
1700 | ret2 = loop_rw_iter(READ, file, kiocb, &iter); | |
1701 | ||
9d93a3f5 JA |
1702 | /* |
1703 | * In case of a short read, punt to async. This can happen | |
1704 | * if we have data partially cached. Alternatively we can | |
1705 | * return the short read, in which case the application will | |
1706 | * need to issue another SQE and wait for it. That SQE will | |
1707 | * need async punt anyway, so it's more efficient to do it | |
1708 | * here. | |
1709 | */ | |
491381ce JA |
1710 | if (force_nonblock && !(req->flags & REQ_F_NOWAIT) && |
1711 | (req->flags & REQ_F_ISREG) && | |
1712 | ret2 > 0 && ret2 < read_size) | |
9d93a3f5 JA |
1713 | ret2 = -EAGAIN; |
1714 | /* Catch -EAGAIN return for forced non-blocking submission */ | |
561fb04a | 1715 | if (!force_nonblock || ret2 != -EAGAIN) |
267bc904 | 1716 | kiocb_done(kiocb, ret2, nxt, req->submit.in_async); |
561fb04a | 1717 | else |
2b188cc1 JA |
1718 | ret = -EAGAIN; |
1719 | } | |
1720 | kfree(iovec); | |
2b188cc1 JA |
1721 | return ret; |
1722 | } | |
1723 | ||
267bc904 | 1724 | static int io_write(struct io_kiocb *req, struct io_kiocb **nxt, |
8358e3a8 | 1725 | bool force_nonblock) |
2b188cc1 JA |
1726 | { |
1727 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1728 | struct kiocb *kiocb = &req->rw; | |
1729 | struct iov_iter iter; | |
1730 | struct file *file; | |
31b51510 | 1731 | size_t iov_count; |
87e5e6da | 1732 | ssize_t ret; |
2b188cc1 | 1733 | |
267bc904 | 1734 | ret = io_prep_rw(req, force_nonblock); |
2b188cc1 JA |
1735 | if (ret) |
1736 | return ret; | |
2b188cc1 | 1737 | |
2b188cc1 JA |
1738 | file = kiocb->ki_filp; |
1739 | if (unlikely(!(file->f_mode & FMODE_WRITE))) | |
09bb8394 | 1740 | return -EBADF; |
2b188cc1 | 1741 | |
267bc904 | 1742 | ret = io_import_iovec(req->ctx, WRITE, &req->submit, &iovec, &iter); |
87e5e6da | 1743 | if (ret < 0) |
09bb8394 | 1744 | return ret; |
2b188cc1 | 1745 | |
9e645e11 JA |
1746 | if (req->flags & REQ_F_LINK) |
1747 | req->result = ret; | |
1748 | ||
31b51510 JA |
1749 | iov_count = iov_iter_count(&iter); |
1750 | ||
1751 | ret = -EAGAIN; | |
561fb04a | 1752 | if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT)) |
31b51510 | 1753 | goto out_free; |
31b51510 JA |
1754 | |
1755 | ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count); | |
2b188cc1 | 1756 | if (!ret) { |
9bf7933f RP |
1757 | ssize_t ret2; |
1758 | ||
2b188cc1 JA |
1759 | /* |
1760 | * Open-code file_start_write here to grab freeze protection, | |
1761 | * which will be released by another thread in | |
1762 | * io_complete_rw(). Fool lockdep by telling it the lock got | |
1763 | * released so that it doesn't complain about the held lock when | |
1764 | * we return to userspace. | |
1765 | */ | |
491381ce | 1766 | if (req->flags & REQ_F_ISREG) { |
2b188cc1 JA |
1767 | __sb_start_write(file_inode(file)->i_sb, |
1768 | SB_FREEZE_WRITE, true); | |
1769 | __sb_writers_release(file_inode(file)->i_sb, | |
1770 | SB_FREEZE_WRITE); | |
1771 | } | |
1772 | kiocb->ki_flags |= IOCB_WRITE; | |
9bf7933f | 1773 | |
32960613 JA |
1774 | if (file->f_op->write_iter) |
1775 | ret2 = call_write_iter(file, kiocb, &iter); | |
1776 | else | |
1777 | ret2 = loop_rw_iter(WRITE, file, kiocb, &iter); | |
561fb04a | 1778 | if (!force_nonblock || ret2 != -EAGAIN) |
267bc904 | 1779 | kiocb_done(kiocb, ret2, nxt, req->submit.in_async); |
561fb04a | 1780 | else |
9bf7933f | 1781 | ret = -EAGAIN; |
2b188cc1 | 1782 | } |
31b51510 | 1783 | out_free: |
2b188cc1 | 1784 | kfree(iovec); |
2b188cc1 JA |
1785 | return ret; |
1786 | } | |
1787 | ||
1788 | /* | |
1789 | * IORING_OP_NOP just posts a completion event, nothing else. | |
1790 | */ | |
78e19bbe | 1791 | static int io_nop(struct io_kiocb *req) |
2b188cc1 JA |
1792 | { |
1793 | struct io_ring_ctx *ctx = req->ctx; | |
2b188cc1 | 1794 | |
def596e9 JA |
1795 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
1796 | return -EINVAL; | |
1797 | ||
78e19bbe | 1798 | io_cqring_add_event(req, 0); |
e65ef56d | 1799 | io_put_req(req); |
2b188cc1 JA |
1800 | return 0; |
1801 | } | |
1802 | ||
c992fe29 CH |
1803 | static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
1804 | { | |
6b06314c | 1805 | struct io_ring_ctx *ctx = req->ctx; |
c992fe29 | 1806 | |
09bb8394 JA |
1807 | if (!req->file) |
1808 | return -EBADF; | |
c992fe29 | 1809 | |
6b06314c | 1810 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
def596e9 | 1811 | return -EINVAL; |
edafccee | 1812 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) |
c992fe29 CH |
1813 | return -EINVAL; |
1814 | ||
c992fe29 CH |
1815 | return 0; |
1816 | } | |
1817 | ||
1818 | static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
ba816ad6 | 1819 | struct io_kiocb **nxt, bool force_nonblock) |
c992fe29 CH |
1820 | { |
1821 | loff_t sqe_off = READ_ONCE(sqe->off); | |
1822 | loff_t sqe_len = READ_ONCE(sqe->len); | |
1823 | loff_t end = sqe_off + sqe_len; | |
1824 | unsigned fsync_flags; | |
1825 | int ret; | |
1826 | ||
1827 | fsync_flags = READ_ONCE(sqe->fsync_flags); | |
1828 | if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC)) | |
1829 | return -EINVAL; | |
1830 | ||
1831 | ret = io_prep_fsync(req, sqe); | |
1832 | if (ret) | |
1833 | return ret; | |
1834 | ||
1835 | /* fsync always requires a blocking context */ | |
1836 | if (force_nonblock) | |
1837 | return -EAGAIN; | |
1838 | ||
1839 | ret = vfs_fsync_range(req->rw.ki_filp, sqe_off, | |
1840 | end > 0 ? end : LLONG_MAX, | |
1841 | fsync_flags & IORING_FSYNC_DATASYNC); | |
1842 | ||
9e645e11 JA |
1843 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1844 | req->flags |= REQ_F_FAIL_LINK; | |
78e19bbe | 1845 | io_cqring_add_event(req, ret); |
ec9c02ad | 1846 | io_put_req_find_next(req, nxt); |
c992fe29 CH |
1847 | return 0; |
1848 | } | |
1849 | ||
5d17b4a4 JA |
1850 | static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
1851 | { | |
1852 | struct io_ring_ctx *ctx = req->ctx; | |
1853 | int ret = 0; | |
1854 | ||
1855 | if (!req->file) | |
1856 | return -EBADF; | |
5d17b4a4 JA |
1857 | |
1858 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
1859 | return -EINVAL; | |
1860 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) | |
1861 | return -EINVAL; | |
1862 | ||
5d17b4a4 JA |
1863 | return ret; |
1864 | } | |
1865 | ||
1866 | static int io_sync_file_range(struct io_kiocb *req, | |
1867 | const struct io_uring_sqe *sqe, | |
ba816ad6 | 1868 | struct io_kiocb **nxt, |
5d17b4a4 JA |
1869 | bool force_nonblock) |
1870 | { | |
1871 | loff_t sqe_off; | |
1872 | loff_t sqe_len; | |
1873 | unsigned flags; | |
1874 | int ret; | |
1875 | ||
1876 | ret = io_prep_sfr(req, sqe); | |
1877 | if (ret) | |
1878 | return ret; | |
1879 | ||
1880 | /* sync_file_range always requires a blocking context */ | |
1881 | if (force_nonblock) | |
1882 | return -EAGAIN; | |
1883 | ||
1884 | sqe_off = READ_ONCE(sqe->off); | |
1885 | sqe_len = READ_ONCE(sqe->len); | |
1886 | flags = READ_ONCE(sqe->sync_range_flags); | |
1887 | ||
1888 | ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags); | |
1889 | ||
9e645e11 JA |
1890 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1891 | req->flags |= REQ_F_FAIL_LINK; | |
78e19bbe | 1892 | io_cqring_add_event(req, ret); |
ec9c02ad | 1893 | io_put_req_find_next(req, nxt); |
5d17b4a4 JA |
1894 | return 0; |
1895 | } | |
1896 | ||
0fa03c62 | 1897 | #if defined(CONFIG_NET) |
aa1fa28f | 1898 | static int io_send_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
ba816ad6 | 1899 | struct io_kiocb **nxt, bool force_nonblock, |
aa1fa28f JA |
1900 | long (*fn)(struct socket *, struct user_msghdr __user *, |
1901 | unsigned int)) | |
1902 | { | |
0fa03c62 JA |
1903 | struct socket *sock; |
1904 | int ret; | |
1905 | ||
1906 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1907 | return -EINVAL; | |
1908 | ||
1909 | sock = sock_from_file(req->file, &ret); | |
1910 | if (sock) { | |
1911 | struct user_msghdr __user *msg; | |
1912 | unsigned flags; | |
1913 | ||
1914 | flags = READ_ONCE(sqe->msg_flags); | |
1915 | if (flags & MSG_DONTWAIT) | |
1916 | req->flags |= REQ_F_NOWAIT; | |
1917 | else if (force_nonblock) | |
1918 | flags |= MSG_DONTWAIT; | |
1919 | ||
1920 | msg = (struct user_msghdr __user *) (unsigned long) | |
1921 | READ_ONCE(sqe->addr); | |
1922 | ||
aa1fa28f | 1923 | ret = fn(sock, msg, flags); |
0fa03c62 JA |
1924 | if (force_nonblock && ret == -EAGAIN) |
1925 | return ret; | |
1926 | } | |
1927 | ||
78e19bbe | 1928 | io_cqring_add_event(req, ret); |
f1f40853 JA |
1929 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1930 | req->flags |= REQ_F_FAIL_LINK; | |
ec9c02ad | 1931 | io_put_req_find_next(req, nxt); |
5d17b4a4 JA |
1932 | return 0; |
1933 | } | |
aa1fa28f JA |
1934 | #endif |
1935 | ||
1936 | static int io_sendmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
ba816ad6 | 1937 | struct io_kiocb **nxt, bool force_nonblock) |
aa1fa28f JA |
1938 | { |
1939 | #if defined(CONFIG_NET) | |
ba816ad6 JA |
1940 | return io_send_recvmsg(req, sqe, nxt, force_nonblock, |
1941 | __sys_sendmsg_sock); | |
aa1fa28f JA |
1942 | #else |
1943 | return -EOPNOTSUPP; | |
1944 | #endif | |
1945 | } | |
1946 | ||
1947 | static int io_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
ba816ad6 | 1948 | struct io_kiocb **nxt, bool force_nonblock) |
aa1fa28f JA |
1949 | { |
1950 | #if defined(CONFIG_NET) | |
ba816ad6 JA |
1951 | return io_send_recvmsg(req, sqe, nxt, force_nonblock, |
1952 | __sys_recvmsg_sock); | |
0fa03c62 JA |
1953 | #else |
1954 | return -EOPNOTSUPP; | |
1955 | #endif | |
1956 | } | |
5d17b4a4 | 1957 | |
17f2fe35 JA |
1958 | static int io_accept(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
1959 | struct io_kiocb **nxt, bool force_nonblock) | |
1960 | { | |
1961 | #if defined(CONFIG_NET) | |
1962 | struct sockaddr __user *addr; | |
1963 | int __user *addr_len; | |
1964 | unsigned file_flags; | |
1965 | int flags, ret; | |
1966 | ||
1967 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL))) | |
1968 | return -EINVAL; | |
8042d6ce | 1969 | if (sqe->ioprio || sqe->len || sqe->buf_index) |
17f2fe35 JA |
1970 | return -EINVAL; |
1971 | ||
1972 | addr = (struct sockaddr __user *) (unsigned long) READ_ONCE(sqe->addr); | |
1973 | addr_len = (int __user *) (unsigned long) READ_ONCE(sqe->addr2); | |
1974 | flags = READ_ONCE(sqe->accept_flags); | |
1975 | file_flags = force_nonblock ? O_NONBLOCK : 0; | |
1976 | ||
1977 | ret = __sys_accept4_file(req->file, file_flags, addr, addr_len, flags); | |
1978 | if (ret == -EAGAIN && force_nonblock) { | |
1979 | req->work.flags |= IO_WQ_WORK_NEEDS_FILES; | |
1980 | return -EAGAIN; | |
1981 | } | |
8e3cca12 JA |
1982 | if (ret == -ERESTARTSYS) |
1983 | ret = -EINTR; | |
17f2fe35 JA |
1984 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1985 | req->flags |= REQ_F_FAIL_LINK; | |
78e19bbe | 1986 | io_cqring_add_event(req, ret); |
ec9c02ad | 1987 | io_put_req_find_next(req, nxt); |
17f2fe35 | 1988 | return 0; |
0fa03c62 JA |
1989 | #else |
1990 | return -EOPNOTSUPP; | |
1991 | #endif | |
1992 | } | |
5d17b4a4 | 1993 | |
f8e85cf2 JA |
1994 | static int io_connect(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
1995 | struct io_kiocb **nxt, bool force_nonblock) | |
1996 | { | |
1997 | #if defined(CONFIG_NET) | |
1998 | struct sockaddr __user *addr; | |
1999 | unsigned file_flags; | |
2000 | int addr_len, ret; | |
2001 | ||
2002 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL))) | |
2003 | return -EINVAL; | |
2004 | if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags) | |
2005 | return -EINVAL; | |
2006 | ||
2007 | addr = (struct sockaddr __user *) (unsigned long) READ_ONCE(sqe->addr); | |
2008 | addr_len = READ_ONCE(sqe->addr2); | |
2009 | file_flags = force_nonblock ? O_NONBLOCK : 0; | |
2010 | ||
2011 | ret = __sys_connect_file(req->file, addr, addr_len, file_flags); | |
2012 | if (ret == -EAGAIN && force_nonblock) | |
2013 | return -EAGAIN; | |
2014 | if (ret == -ERESTARTSYS) | |
2015 | ret = -EINTR; | |
2016 | if (ret < 0 && (req->flags & REQ_F_LINK)) | |
2017 | req->flags |= REQ_F_FAIL_LINK; | |
2018 | io_cqring_add_event(req, ret); | |
2019 | io_put_req_find_next(req, nxt); | |
2020 | return 0; | |
2021 | #else | |
2022 | return -EOPNOTSUPP; | |
2023 | #endif | |
2024 | } | |
2025 | ||
eac406c6 JA |
2026 | static inline void io_poll_remove_req(struct io_kiocb *req) |
2027 | { | |
2028 | if (!RB_EMPTY_NODE(&req->rb_node)) { | |
2029 | rb_erase(&req->rb_node, &req->ctx->cancel_tree); | |
2030 | RB_CLEAR_NODE(&req->rb_node); | |
2031 | } | |
2032 | } | |
2033 | ||
221c5eb2 JA |
2034 | static void io_poll_remove_one(struct io_kiocb *req) |
2035 | { | |
2036 | struct io_poll_iocb *poll = &req->poll; | |
2037 | ||
2038 | spin_lock(&poll->head->lock); | |
2039 | WRITE_ONCE(poll->canceled, true); | |
2040 | if (!list_empty(&poll->wait.entry)) { | |
2041 | list_del_init(&poll->wait.entry); | |
a197f664 | 2042 | io_queue_async_work(req); |
221c5eb2 JA |
2043 | } |
2044 | spin_unlock(&poll->head->lock); | |
eac406c6 | 2045 | io_poll_remove_req(req); |
221c5eb2 JA |
2046 | } |
2047 | ||
2048 | static void io_poll_remove_all(struct io_ring_ctx *ctx) | |
2049 | { | |
eac406c6 | 2050 | struct rb_node *node; |
221c5eb2 JA |
2051 | struct io_kiocb *req; |
2052 | ||
2053 | spin_lock_irq(&ctx->completion_lock); | |
eac406c6 JA |
2054 | while ((node = rb_first(&ctx->cancel_tree)) != NULL) { |
2055 | req = rb_entry(node, struct io_kiocb, rb_node); | |
221c5eb2 JA |
2056 | io_poll_remove_one(req); |
2057 | } | |
2058 | spin_unlock_irq(&ctx->completion_lock); | |
2059 | } | |
2060 | ||
47f46768 JA |
2061 | static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr) |
2062 | { | |
eac406c6 | 2063 | struct rb_node *p, *parent = NULL; |
47f46768 JA |
2064 | struct io_kiocb *req; |
2065 | ||
eac406c6 JA |
2066 | p = ctx->cancel_tree.rb_node; |
2067 | while (p) { | |
2068 | parent = p; | |
2069 | req = rb_entry(parent, struct io_kiocb, rb_node); | |
2070 | if (sqe_addr < req->user_data) { | |
2071 | p = p->rb_left; | |
2072 | } else if (sqe_addr > req->user_data) { | |
2073 | p = p->rb_right; | |
2074 | } else { | |
2075 | io_poll_remove_one(req); | |
2076 | return 0; | |
2077 | } | |
47f46768 JA |
2078 | } |
2079 | ||
2080 | return -ENOENT; | |
2081 | } | |
2082 | ||
221c5eb2 JA |
2083 | /* |
2084 | * Find a running poll command that matches one specified in sqe->addr, | |
2085 | * and remove it if found. | |
2086 | */ | |
2087 | static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
2088 | { | |
2089 | struct io_ring_ctx *ctx = req->ctx; | |
47f46768 | 2090 | int ret; |
221c5eb2 JA |
2091 | |
2092 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
2093 | return -EINVAL; | |
2094 | if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index || | |
2095 | sqe->poll_events) | |
2096 | return -EINVAL; | |
2097 | ||
2098 | spin_lock_irq(&ctx->completion_lock); | |
47f46768 | 2099 | ret = io_poll_cancel(ctx, READ_ONCE(sqe->addr)); |
221c5eb2 JA |
2100 | spin_unlock_irq(&ctx->completion_lock); |
2101 | ||
78e19bbe | 2102 | io_cqring_add_event(req, ret); |
f1f40853 JA |
2103 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
2104 | req->flags |= REQ_F_FAIL_LINK; | |
e65ef56d | 2105 | io_put_req(req); |
221c5eb2 JA |
2106 | return 0; |
2107 | } | |
2108 | ||
b0dd8a41 | 2109 | static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error) |
221c5eb2 | 2110 | { |
a197f664 JL |
2111 | struct io_ring_ctx *ctx = req->ctx; |
2112 | ||
8c838788 | 2113 | req->poll.done = true; |
b0dd8a41 JA |
2114 | if (error) |
2115 | io_cqring_fill_event(req, error); | |
2116 | else | |
2117 | io_cqring_fill_event(req, mangle_poll(mask)); | |
8c838788 | 2118 | io_commit_cqring(ctx); |
221c5eb2 JA |
2119 | } |
2120 | ||
561fb04a | 2121 | static void io_poll_complete_work(struct io_wq_work **workptr) |
221c5eb2 | 2122 | { |
561fb04a | 2123 | struct io_wq_work *work = *workptr; |
221c5eb2 JA |
2124 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
2125 | struct io_poll_iocb *poll = &req->poll; | |
2126 | struct poll_table_struct pt = { ._key = poll->events }; | |
2127 | struct io_ring_ctx *ctx = req->ctx; | |
89723d0b | 2128 | struct io_kiocb *nxt = NULL; |
221c5eb2 | 2129 | __poll_t mask = 0; |
b0dd8a41 | 2130 | int ret = 0; |
221c5eb2 | 2131 | |
b0dd8a41 | 2132 | if (work->flags & IO_WQ_WORK_CANCEL) { |
561fb04a | 2133 | WRITE_ONCE(poll->canceled, true); |
b0dd8a41 JA |
2134 | ret = -ECANCELED; |
2135 | } else if (READ_ONCE(poll->canceled)) { | |
2136 | ret = -ECANCELED; | |
2137 | } | |
561fb04a | 2138 | |
b0dd8a41 | 2139 | if (ret != -ECANCELED) |
221c5eb2 JA |
2140 | mask = vfs_poll(poll->file, &pt) & poll->events; |
2141 | ||
2142 | /* | |
2143 | * Note that ->ki_cancel callers also delete iocb from active_reqs after | |
2144 | * calling ->ki_cancel. We need the ctx_lock roundtrip here to | |
2145 | * synchronize with them. In the cancellation case the list_del_init | |
2146 | * itself is not actually needed, but harmless so we keep it in to | |
2147 | * avoid further branches in the fast path. | |
2148 | */ | |
2149 | spin_lock_irq(&ctx->completion_lock); | |
b0dd8a41 | 2150 | if (!mask && ret != -ECANCELED) { |
221c5eb2 JA |
2151 | add_wait_queue(poll->head, &poll->wait); |
2152 | spin_unlock_irq(&ctx->completion_lock); | |
2153 | return; | |
2154 | } | |
eac406c6 | 2155 | io_poll_remove_req(req); |
b0dd8a41 | 2156 | io_poll_complete(req, mask, ret); |
221c5eb2 JA |
2157 | spin_unlock_irq(&ctx->completion_lock); |
2158 | ||
8c838788 | 2159 | io_cqring_ev_posted(ctx); |
89723d0b | 2160 | |
fba38c27 JA |
2161 | if (ret < 0 && req->flags & REQ_F_LINK) |
2162 | req->flags |= REQ_F_FAIL_LINK; | |
ec9c02ad | 2163 | io_put_req_find_next(req, &nxt); |
89723d0b JA |
2164 | if (nxt) |
2165 | *workptr = &nxt->work; | |
221c5eb2 JA |
2166 | } |
2167 | ||
2168 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, | |
2169 | void *key) | |
2170 | { | |
2171 | struct io_poll_iocb *poll = container_of(wait, struct io_poll_iocb, | |
2172 | wait); | |
2173 | struct io_kiocb *req = container_of(poll, struct io_kiocb, poll); | |
2174 | struct io_ring_ctx *ctx = req->ctx; | |
2175 | __poll_t mask = key_to_poll(key); | |
8c838788 | 2176 | unsigned long flags; |
221c5eb2 JA |
2177 | |
2178 | /* for instances that support it check for an event match first: */ | |
8c838788 JA |
2179 | if (mask && !(mask & poll->events)) |
2180 | return 0; | |
221c5eb2 | 2181 | |
8c838788 | 2182 | list_del_init(&poll->wait.entry); |
221c5eb2 | 2183 | |
7c9e7f0f JA |
2184 | /* |
2185 | * Run completion inline if we can. We're using trylock here because | |
2186 | * we are violating the completion_lock -> poll wq lock ordering. | |
2187 | * If we have a link timeout we're going to need the completion_lock | |
2188 | * for finalizing the request, mark us as having grabbed that already. | |
2189 | */ | |
8c838788 | 2190 | if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) { |
eac406c6 | 2191 | io_poll_remove_req(req); |
b0dd8a41 | 2192 | io_poll_complete(req, mask, 0); |
7c9e7f0f JA |
2193 | req->flags |= REQ_F_COMP_LOCKED; |
2194 | io_put_req(req); | |
8c838788 | 2195 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
221c5eb2 | 2196 | |
8c838788 | 2197 | io_cqring_ev_posted(ctx); |
8c838788 | 2198 | } else { |
a197f664 | 2199 | io_queue_async_work(req); |
221c5eb2 JA |
2200 | } |
2201 | ||
221c5eb2 JA |
2202 | return 1; |
2203 | } | |
2204 | ||
2205 | struct io_poll_table { | |
2206 | struct poll_table_struct pt; | |
2207 | struct io_kiocb *req; | |
2208 | int error; | |
2209 | }; | |
2210 | ||
2211 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, | |
2212 | struct poll_table_struct *p) | |
2213 | { | |
2214 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
2215 | ||
2216 | if (unlikely(pt->req->poll.head)) { | |
2217 | pt->error = -EINVAL; | |
2218 | return; | |
2219 | } | |
2220 | ||
2221 | pt->error = 0; | |
2222 | pt->req->poll.head = head; | |
2223 | add_wait_queue(head, &pt->req->poll.wait); | |
2224 | } | |
2225 | ||
eac406c6 JA |
2226 | static void io_poll_req_insert(struct io_kiocb *req) |
2227 | { | |
2228 | struct io_ring_ctx *ctx = req->ctx; | |
2229 | struct rb_node **p = &ctx->cancel_tree.rb_node; | |
2230 | struct rb_node *parent = NULL; | |
2231 | struct io_kiocb *tmp; | |
2232 | ||
2233 | while (*p) { | |
2234 | parent = *p; | |
2235 | tmp = rb_entry(parent, struct io_kiocb, rb_node); | |
2236 | if (req->user_data < tmp->user_data) | |
2237 | p = &(*p)->rb_left; | |
2238 | else | |
2239 | p = &(*p)->rb_right; | |
2240 | } | |
2241 | rb_link_node(&req->rb_node, parent, p); | |
2242 | rb_insert_color(&req->rb_node, &ctx->cancel_tree); | |
2243 | } | |
2244 | ||
89723d0b JA |
2245 | static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2246 | struct io_kiocb **nxt) | |
221c5eb2 JA |
2247 | { |
2248 | struct io_poll_iocb *poll = &req->poll; | |
2249 | struct io_ring_ctx *ctx = req->ctx; | |
2250 | struct io_poll_table ipt; | |
8c838788 | 2251 | bool cancel = false; |
221c5eb2 JA |
2252 | __poll_t mask; |
2253 | u16 events; | |
221c5eb2 JA |
2254 | |
2255 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
2256 | return -EINVAL; | |
2257 | if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index) | |
2258 | return -EINVAL; | |
09bb8394 JA |
2259 | if (!poll->file) |
2260 | return -EBADF; | |
221c5eb2 | 2261 | |
6cc47d1d | 2262 | req->submit.sqe = NULL; |
561fb04a | 2263 | INIT_IO_WORK(&req->work, io_poll_complete_work); |
221c5eb2 JA |
2264 | events = READ_ONCE(sqe->poll_events); |
2265 | poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP; | |
eac406c6 | 2266 | RB_CLEAR_NODE(&req->rb_node); |
221c5eb2 | 2267 | |
221c5eb2 | 2268 | poll->head = NULL; |
8c838788 | 2269 | poll->done = false; |
221c5eb2 JA |
2270 | poll->canceled = false; |
2271 | ||
2272 | ipt.pt._qproc = io_poll_queue_proc; | |
2273 | ipt.pt._key = poll->events; | |
2274 | ipt.req = req; | |
2275 | ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */ | |
2276 | ||
2277 | /* initialized the list so that we can do list_empty checks */ | |
2278 | INIT_LIST_HEAD(&poll->wait.entry); | |
2279 | init_waitqueue_func_entry(&poll->wait, io_poll_wake); | |
2280 | ||
36703247 JA |
2281 | INIT_LIST_HEAD(&req->list); |
2282 | ||
221c5eb2 | 2283 | mask = vfs_poll(poll->file, &ipt.pt) & poll->events; |
221c5eb2 JA |
2284 | |
2285 | spin_lock_irq(&ctx->completion_lock); | |
8c838788 JA |
2286 | if (likely(poll->head)) { |
2287 | spin_lock(&poll->head->lock); | |
2288 | if (unlikely(list_empty(&poll->wait.entry))) { | |
2289 | if (ipt.error) | |
2290 | cancel = true; | |
2291 | ipt.error = 0; | |
2292 | mask = 0; | |
2293 | } | |
2294 | if (mask || ipt.error) | |
2295 | list_del_init(&poll->wait.entry); | |
2296 | else if (cancel) | |
2297 | WRITE_ONCE(poll->canceled, true); | |
2298 | else if (!poll->done) /* actually waiting for an event */ | |
eac406c6 | 2299 | io_poll_req_insert(req); |
8c838788 JA |
2300 | spin_unlock(&poll->head->lock); |
2301 | } | |
2302 | if (mask) { /* no async, we'd stolen it */ | |
221c5eb2 | 2303 | ipt.error = 0; |
b0dd8a41 | 2304 | io_poll_complete(req, mask, 0); |
221c5eb2 | 2305 | } |
221c5eb2 JA |
2306 | spin_unlock_irq(&ctx->completion_lock); |
2307 | ||
8c838788 JA |
2308 | if (mask) { |
2309 | io_cqring_ev_posted(ctx); | |
ec9c02ad | 2310 | io_put_req_find_next(req, nxt); |
221c5eb2 | 2311 | } |
8c838788 | 2312 | return ipt.error; |
221c5eb2 JA |
2313 | } |
2314 | ||
5262f567 JA |
2315 | static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer) |
2316 | { | |
ad8a48ac JA |
2317 | struct io_timeout_data *data = container_of(timer, |
2318 | struct io_timeout_data, timer); | |
2319 | struct io_kiocb *req = data->req; | |
2320 | struct io_ring_ctx *ctx = req->ctx; | |
5262f567 JA |
2321 | unsigned long flags; |
2322 | ||
5262f567 JA |
2323 | atomic_inc(&ctx->cq_timeouts); |
2324 | ||
2325 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
ef03681a | 2326 | /* |
11365043 JA |
2327 | * We could be racing with timeout deletion. If the list is empty, |
2328 | * then timeout lookup already found it and will be handling it. | |
ef03681a | 2329 | */ |
842f9612 | 2330 | if (!list_empty(&req->list)) { |
11365043 | 2331 | struct io_kiocb *prev; |
5262f567 | 2332 | |
11365043 JA |
2333 | /* |
2334 | * Adjust the reqs sequence before the current one because it | |
2335 | * will consume a slot in the cq_ring and the the cq_tail | |
2336 | * pointer will be increased, otherwise other timeout reqs may | |
2337 | * return in advance without waiting for enough wait_nr. | |
2338 | */ | |
2339 | prev = req; | |
2340 | list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list) | |
2341 | prev->sequence++; | |
11365043 | 2342 | list_del_init(&req->list); |
11365043 | 2343 | } |
5262f567 | 2344 | |
78e19bbe | 2345 | io_cqring_fill_event(req, -ETIME); |
5262f567 JA |
2346 | io_commit_cqring(ctx); |
2347 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
2348 | ||
2349 | io_cqring_ev_posted(ctx); | |
f1f40853 JA |
2350 | if (req->flags & REQ_F_LINK) |
2351 | req->flags |= REQ_F_FAIL_LINK; | |
5262f567 JA |
2352 | io_put_req(req); |
2353 | return HRTIMER_NORESTART; | |
2354 | } | |
2355 | ||
47f46768 JA |
2356 | static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data) |
2357 | { | |
2358 | struct io_kiocb *req; | |
2359 | int ret = -ENOENT; | |
2360 | ||
2361 | list_for_each_entry(req, &ctx->timeout_list, list) { | |
2362 | if (user_data == req->user_data) { | |
2363 | list_del_init(&req->list); | |
2364 | ret = 0; | |
2365 | break; | |
2366 | } | |
2367 | } | |
2368 | ||
2369 | if (ret == -ENOENT) | |
2370 | return ret; | |
2371 | ||
ad8a48ac | 2372 | ret = hrtimer_try_to_cancel(&req->timeout.data->timer); |
47f46768 JA |
2373 | if (ret == -1) |
2374 | return -EALREADY; | |
2375 | ||
fba38c27 JA |
2376 | if (req->flags & REQ_F_LINK) |
2377 | req->flags |= REQ_F_FAIL_LINK; | |
47f46768 JA |
2378 | io_cqring_fill_event(req, -ECANCELED); |
2379 | io_put_req(req); | |
2380 | return 0; | |
2381 | } | |
2382 | ||
11365043 JA |
2383 | /* |
2384 | * Remove or update an existing timeout command | |
2385 | */ | |
2386 | static int io_timeout_remove(struct io_kiocb *req, | |
2387 | const struct io_uring_sqe *sqe) | |
2388 | { | |
2389 | struct io_ring_ctx *ctx = req->ctx; | |
11365043 | 2390 | unsigned flags; |
47f46768 | 2391 | int ret; |
11365043 JA |
2392 | |
2393 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
2394 | return -EINVAL; | |
2395 | if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len) | |
2396 | return -EINVAL; | |
2397 | flags = READ_ONCE(sqe->timeout_flags); | |
2398 | if (flags) | |
2399 | return -EINVAL; | |
2400 | ||
11365043 | 2401 | spin_lock_irq(&ctx->completion_lock); |
47f46768 | 2402 | ret = io_timeout_cancel(ctx, READ_ONCE(sqe->addr)); |
11365043 | 2403 | |
47f46768 | 2404 | io_cqring_fill_event(req, ret); |
11365043 JA |
2405 | io_commit_cqring(ctx); |
2406 | spin_unlock_irq(&ctx->completion_lock); | |
5262f567 | 2407 | io_cqring_ev_posted(ctx); |
47f46768 JA |
2408 | if (ret < 0 && req->flags & REQ_F_LINK) |
2409 | req->flags |= REQ_F_FAIL_LINK; | |
ec9c02ad | 2410 | io_put_req(req); |
11365043 | 2411 | return 0; |
5262f567 JA |
2412 | } |
2413 | ||
ad8a48ac | 2414 | static int io_timeout_setup(struct io_kiocb *req) |
5262f567 | 2415 | { |
ad8a48ac JA |
2416 | const struct io_uring_sqe *sqe = req->submit.sqe; |
2417 | struct io_timeout_data *data; | |
a41525ab | 2418 | unsigned flags; |
5262f567 | 2419 | |
ad8a48ac | 2420 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5262f567 | 2421 | return -EINVAL; |
ad8a48ac | 2422 | if (sqe->ioprio || sqe->buf_index || sqe->len != 1) |
a41525ab JA |
2423 | return -EINVAL; |
2424 | flags = READ_ONCE(sqe->timeout_flags); | |
2425 | if (flags & ~IORING_TIMEOUT_ABS) | |
5262f567 | 2426 | return -EINVAL; |
bdf20073 | 2427 | |
ad8a48ac JA |
2428 | data = kzalloc(sizeof(struct io_timeout_data), GFP_KERNEL); |
2429 | if (!data) | |
2430 | return -ENOMEM; | |
2431 | data->req = req; | |
2432 | req->timeout.data = data; | |
2433 | req->flags |= REQ_F_TIMEOUT; | |
2434 | ||
2435 | if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr))) | |
5262f567 JA |
2436 | return -EFAULT; |
2437 | ||
11365043 | 2438 | if (flags & IORING_TIMEOUT_ABS) |
ad8a48ac | 2439 | data->mode = HRTIMER_MODE_ABS; |
11365043 | 2440 | else |
ad8a48ac | 2441 | data->mode = HRTIMER_MODE_REL; |
11365043 | 2442 | |
ad8a48ac JA |
2443 | hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode); |
2444 | return 0; | |
2445 | } | |
2446 | ||
2447 | static int io_timeout(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
2448 | { | |
2449 | unsigned count; | |
2450 | struct io_ring_ctx *ctx = req->ctx; | |
2451 | struct io_timeout_data *data; | |
2452 | struct list_head *entry; | |
2453 | unsigned span = 0; | |
2454 | int ret; | |
2455 | ||
2456 | ret = io_timeout_setup(req); | |
2457 | /* common setup allows flags (like links) set, we don't */ | |
2458 | if (!ret && sqe->flags) | |
2459 | ret = -EINVAL; | |
2460 | if (ret) | |
2461 | return ret; | |
93bd25bb | 2462 | |
5262f567 JA |
2463 | /* |
2464 | * sqe->off holds how many events that need to occur for this | |
93bd25bb JA |
2465 | * timeout event to be satisfied. If it isn't set, then this is |
2466 | * a pure timeout request, sequence isn't used. | |
5262f567 JA |
2467 | */ |
2468 | count = READ_ONCE(sqe->off); | |
93bd25bb JA |
2469 | if (!count) { |
2470 | req->flags |= REQ_F_TIMEOUT_NOSEQ; | |
2471 | spin_lock_irq(&ctx->completion_lock); | |
2472 | entry = ctx->timeout_list.prev; | |
2473 | goto add; | |
2474 | } | |
5262f567 JA |
2475 | |
2476 | req->sequence = ctx->cached_sq_head + count - 1; | |
cc42e0ac | 2477 | req->timeout.data->seq_offset = count; |
5262f567 JA |
2478 | |
2479 | /* | |
2480 | * Insertion sort, ensuring the first entry in the list is always | |
2481 | * the one we need first. | |
2482 | */ | |
5262f567 JA |
2483 | spin_lock_irq(&ctx->completion_lock); |
2484 | list_for_each_prev(entry, &ctx->timeout_list) { | |
2485 | struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list); | |
5da0fb1a | 2486 | unsigned nxt_sq_head; |
2487 | long long tmp, tmp_nxt; | |
cc42e0ac | 2488 | u32 nxt_offset = nxt->timeout.data->seq_offset; |
5262f567 | 2489 | |
93bd25bb JA |
2490 | if (nxt->flags & REQ_F_TIMEOUT_NOSEQ) |
2491 | continue; | |
2492 | ||
5da0fb1a | 2493 | /* |
2494 | * Since cached_sq_head + count - 1 can overflow, use type long | |
2495 | * long to store it. | |
2496 | */ | |
2497 | tmp = (long long)ctx->cached_sq_head + count - 1; | |
cc42e0ac PB |
2498 | nxt_sq_head = nxt->sequence - nxt_offset + 1; |
2499 | tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1; | |
5da0fb1a | 2500 | |
2501 | /* | |
2502 | * cached_sq_head may overflow, and it will never overflow twice | |
2503 | * once there is some timeout req still be valid. | |
2504 | */ | |
2505 | if (ctx->cached_sq_head < nxt_sq_head) | |
8b07a65a | 2506 | tmp += UINT_MAX; |
5da0fb1a | 2507 | |
a1f58ba4 | 2508 | if (tmp > tmp_nxt) |
5262f567 | 2509 | break; |
a1f58ba4 | 2510 | |
2511 | /* | |
2512 | * Sequence of reqs after the insert one and itself should | |
2513 | * be adjusted because each timeout req consumes a slot. | |
2514 | */ | |
2515 | span++; | |
2516 | nxt->sequence++; | |
5262f567 | 2517 | } |
a1f58ba4 | 2518 | req->sequence -= span; |
93bd25bb | 2519 | add: |
5262f567 | 2520 | list_add(&req->list, entry); |
ad8a48ac JA |
2521 | data = req->timeout.data; |
2522 | data->timer.function = io_timeout_fn; | |
2523 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode); | |
5262f567 | 2524 | spin_unlock_irq(&ctx->completion_lock); |
5262f567 JA |
2525 | return 0; |
2526 | } | |
5262f567 | 2527 | |
62755e35 JA |
2528 | static bool io_cancel_cb(struct io_wq_work *work, void *data) |
2529 | { | |
2530 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
2531 | ||
2532 | return req->user_data == (unsigned long) data; | |
2533 | } | |
2534 | ||
e977d6d3 | 2535 | static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr) |
62755e35 | 2536 | { |
62755e35 | 2537 | enum io_wq_cancel cancel_ret; |
62755e35 JA |
2538 | int ret = 0; |
2539 | ||
62755e35 JA |
2540 | cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr); |
2541 | switch (cancel_ret) { | |
2542 | case IO_WQ_CANCEL_OK: | |
2543 | ret = 0; | |
2544 | break; | |
2545 | case IO_WQ_CANCEL_RUNNING: | |
2546 | ret = -EALREADY; | |
2547 | break; | |
2548 | case IO_WQ_CANCEL_NOTFOUND: | |
2549 | ret = -ENOENT; | |
2550 | break; | |
2551 | } | |
2552 | ||
e977d6d3 JA |
2553 | return ret; |
2554 | } | |
2555 | ||
47f46768 JA |
2556 | static void io_async_find_and_cancel(struct io_ring_ctx *ctx, |
2557 | struct io_kiocb *req, __u64 sqe_addr, | |
b0dd8a41 | 2558 | struct io_kiocb **nxt, int success_ret) |
47f46768 JA |
2559 | { |
2560 | unsigned long flags; | |
2561 | int ret; | |
2562 | ||
2563 | ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr); | |
2564 | if (ret != -ENOENT) { | |
2565 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
2566 | goto done; | |
2567 | } | |
2568 | ||
2569 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
2570 | ret = io_timeout_cancel(ctx, sqe_addr); | |
2571 | if (ret != -ENOENT) | |
2572 | goto done; | |
2573 | ret = io_poll_cancel(ctx, sqe_addr); | |
2574 | done: | |
b0dd8a41 JA |
2575 | if (!ret) |
2576 | ret = success_ret; | |
47f46768 JA |
2577 | io_cqring_fill_event(req, ret); |
2578 | io_commit_cqring(ctx); | |
2579 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
2580 | io_cqring_ev_posted(ctx); | |
2581 | ||
2582 | if (ret < 0 && (req->flags & REQ_F_LINK)) | |
2583 | req->flags |= REQ_F_FAIL_LINK; | |
2584 | io_put_req_find_next(req, nxt); | |
2585 | } | |
2586 | ||
e977d6d3 JA |
2587 | static int io_async_cancel(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2588 | struct io_kiocb **nxt) | |
2589 | { | |
2590 | struct io_ring_ctx *ctx = req->ctx; | |
e977d6d3 JA |
2591 | |
2592 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
2593 | return -EINVAL; | |
2594 | if (sqe->flags || sqe->ioprio || sqe->off || sqe->len || | |
2595 | sqe->cancel_flags) | |
2596 | return -EINVAL; | |
2597 | ||
b0dd8a41 | 2598 | io_async_find_and_cancel(ctx, req, READ_ONCE(sqe->addr), nxt, 0); |
5262f567 JA |
2599 | return 0; |
2600 | } | |
2601 | ||
a197f664 | 2602 | static int io_req_defer(struct io_kiocb *req) |
de0617e4 | 2603 | { |
267bc904 | 2604 | const struct io_uring_sqe *sqe = req->submit.sqe; |
de0617e4 | 2605 | struct io_uring_sqe *sqe_copy; |
a197f664 | 2606 | struct io_ring_ctx *ctx = req->ctx; |
de0617e4 | 2607 | |
9d858b21 BL |
2608 | /* Still need defer if there is pending req in defer list. */ |
2609 | if (!req_need_defer(req) && list_empty(&ctx->defer_list)) | |
de0617e4 JA |
2610 | return 0; |
2611 | ||
2612 | sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL); | |
2613 | if (!sqe_copy) | |
2614 | return -EAGAIN; | |
2615 | ||
2616 | spin_lock_irq(&ctx->completion_lock); | |
9d858b21 | 2617 | if (!req_need_defer(req) && list_empty(&ctx->defer_list)) { |
de0617e4 JA |
2618 | spin_unlock_irq(&ctx->completion_lock); |
2619 | kfree(sqe_copy); | |
2620 | return 0; | |
2621 | } | |
2622 | ||
2623 | memcpy(sqe_copy, sqe, sizeof(*sqe_copy)); | |
bbad27b2 | 2624 | req->flags |= REQ_F_FREE_SQE; |
de0617e4 JA |
2625 | req->submit.sqe = sqe_copy; |
2626 | ||
915967f6 | 2627 | trace_io_uring_defer(ctx, req, req->user_data); |
de0617e4 JA |
2628 | list_add_tail(&req->list, &ctx->defer_list); |
2629 | spin_unlock_irq(&ctx->completion_lock); | |
2630 | return -EIOCBQUEUED; | |
2631 | } | |
2632 | ||
f9bd67f6 | 2633 | __attribute__((nonnull)) |
d732447f PB |
2634 | static int io_issue_sqe(struct io_kiocb *req, struct io_kiocb **nxt, |
2635 | bool force_nonblock) | |
2b188cc1 | 2636 | { |
e0c5c576 | 2637 | int ret, opcode; |
267bc904 | 2638 | struct sqe_submit *s = &req->submit; |
a197f664 | 2639 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
2640 | |
2641 | opcode = READ_ONCE(s->sqe->opcode); | |
2642 | switch (opcode) { | |
2643 | case IORING_OP_NOP: | |
78e19bbe | 2644 | ret = io_nop(req); |
2b188cc1 JA |
2645 | break; |
2646 | case IORING_OP_READV: | |
edafccee JA |
2647 | if (unlikely(s->sqe->buf_index)) |
2648 | return -EINVAL; | |
267bc904 | 2649 | ret = io_read(req, nxt, force_nonblock); |
2b188cc1 JA |
2650 | break; |
2651 | case IORING_OP_WRITEV: | |
edafccee JA |
2652 | if (unlikely(s->sqe->buf_index)) |
2653 | return -EINVAL; | |
267bc904 | 2654 | ret = io_write(req, nxt, force_nonblock); |
edafccee JA |
2655 | break; |
2656 | case IORING_OP_READ_FIXED: | |
267bc904 | 2657 | ret = io_read(req, nxt, force_nonblock); |
edafccee JA |
2658 | break; |
2659 | case IORING_OP_WRITE_FIXED: | |
267bc904 | 2660 | ret = io_write(req, nxt, force_nonblock); |
2b188cc1 | 2661 | break; |
c992fe29 | 2662 | case IORING_OP_FSYNC: |
ba816ad6 | 2663 | ret = io_fsync(req, s->sqe, nxt, force_nonblock); |
c992fe29 | 2664 | break; |
221c5eb2 | 2665 | case IORING_OP_POLL_ADD: |
89723d0b | 2666 | ret = io_poll_add(req, s->sqe, nxt); |
221c5eb2 JA |
2667 | break; |
2668 | case IORING_OP_POLL_REMOVE: | |
2669 | ret = io_poll_remove(req, s->sqe); | |
2670 | break; | |
5d17b4a4 | 2671 | case IORING_OP_SYNC_FILE_RANGE: |
ba816ad6 | 2672 | ret = io_sync_file_range(req, s->sqe, nxt, force_nonblock); |
5d17b4a4 | 2673 | break; |
0fa03c62 | 2674 | case IORING_OP_SENDMSG: |
ba816ad6 | 2675 | ret = io_sendmsg(req, s->sqe, nxt, force_nonblock); |
0fa03c62 | 2676 | break; |
aa1fa28f | 2677 | case IORING_OP_RECVMSG: |
ba816ad6 | 2678 | ret = io_recvmsg(req, s->sqe, nxt, force_nonblock); |
aa1fa28f | 2679 | break; |
5262f567 JA |
2680 | case IORING_OP_TIMEOUT: |
2681 | ret = io_timeout(req, s->sqe); | |
2682 | break; | |
11365043 JA |
2683 | case IORING_OP_TIMEOUT_REMOVE: |
2684 | ret = io_timeout_remove(req, s->sqe); | |
2685 | break; | |
17f2fe35 JA |
2686 | case IORING_OP_ACCEPT: |
2687 | ret = io_accept(req, s->sqe, nxt, force_nonblock); | |
2688 | break; | |
f8e85cf2 JA |
2689 | case IORING_OP_CONNECT: |
2690 | ret = io_connect(req, s->sqe, nxt, force_nonblock); | |
2691 | break; | |
62755e35 JA |
2692 | case IORING_OP_ASYNC_CANCEL: |
2693 | ret = io_async_cancel(req, s->sqe, nxt); | |
2694 | break; | |
2b188cc1 JA |
2695 | default: |
2696 | ret = -EINVAL; | |
2697 | break; | |
2698 | } | |
2699 | ||
def596e9 JA |
2700 | if (ret) |
2701 | return ret; | |
2702 | ||
2703 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
9e645e11 | 2704 | if (req->result == -EAGAIN) |
def596e9 JA |
2705 | return -EAGAIN; |
2706 | ||
2707 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
ba5290cc | 2708 | if (s->in_async) |
def596e9 JA |
2709 | mutex_lock(&ctx->uring_lock); |
2710 | io_iopoll_req_issued(req); | |
ba5290cc | 2711 | if (s->in_async) |
def596e9 JA |
2712 | mutex_unlock(&ctx->uring_lock); |
2713 | } | |
2714 | ||
2715 | return 0; | |
2b188cc1 JA |
2716 | } |
2717 | ||
b76da70f JA |
2718 | static void io_link_work_cb(struct io_wq_work **workptr) |
2719 | { | |
2720 | struct io_wq_work *work = *workptr; | |
2721 | struct io_kiocb *link = work->data; | |
2722 | ||
2723 | io_queue_linked_timeout(link); | |
2724 | work->func = io_wq_submit_work; | |
2725 | } | |
2726 | ||
561fb04a | 2727 | static void io_wq_submit_work(struct io_wq_work **workptr) |
2b188cc1 | 2728 | { |
561fb04a | 2729 | struct io_wq_work *work = *workptr; |
2b188cc1 | 2730 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
561fb04a | 2731 | struct sqe_submit *s = &req->submit; |
561fb04a JA |
2732 | struct io_kiocb *nxt = NULL; |
2733 | int ret = 0; | |
2b188cc1 | 2734 | |
561fb04a JA |
2735 | /* Ensure we clear previously set non-block flag */ |
2736 | req->rw.ki_flags &= ~IOCB_NOWAIT; | |
2b188cc1 | 2737 | |
561fb04a JA |
2738 | if (work->flags & IO_WQ_WORK_CANCEL) |
2739 | ret = -ECANCELED; | |
31b51510 | 2740 | |
561fb04a JA |
2741 | if (!ret) { |
2742 | s->has_user = (work->flags & IO_WQ_WORK_HAS_MM) != 0; | |
2743 | s->in_async = true; | |
2744 | do { | |
d732447f | 2745 | ret = io_issue_sqe(req, &nxt, false); |
561fb04a JA |
2746 | /* |
2747 | * We can get EAGAIN for polled IO even though we're | |
2748 | * forcing a sync submission from here, since we can't | |
2749 | * wait for request slots on the block side. | |
2750 | */ | |
2751 | if (ret != -EAGAIN) | |
2752 | break; | |
2753 | cond_resched(); | |
2754 | } while (1); | |
2755 | } | |
31b51510 | 2756 | |
561fb04a | 2757 | /* drop submission reference */ |
ec9c02ad | 2758 | io_put_req(req); |
817869d2 | 2759 | |
561fb04a | 2760 | if (ret) { |
f1f40853 JA |
2761 | if (req->flags & REQ_F_LINK) |
2762 | req->flags |= REQ_F_FAIL_LINK; | |
78e19bbe | 2763 | io_cqring_add_event(req, ret); |
817869d2 | 2764 | io_put_req(req); |
edafccee | 2765 | } |
2b188cc1 | 2766 | |
561fb04a JA |
2767 | /* if a dependent link is ready, pass it back */ |
2768 | if (!ret && nxt) { | |
94ae5e77 JA |
2769 | struct io_kiocb *link; |
2770 | ||
2771 | io_prep_async_work(nxt, &link); | |
561fb04a | 2772 | *workptr = &nxt->work; |
b76da70f JA |
2773 | if (link) { |
2774 | nxt->work.flags |= IO_WQ_WORK_CB; | |
2775 | nxt->work.func = io_link_work_cb; | |
2776 | nxt->work.data = link; | |
2777 | } | |
31b51510 | 2778 | } |
2b188cc1 JA |
2779 | } |
2780 | ||
09bb8394 JA |
2781 | static bool io_op_needs_file(const struct io_uring_sqe *sqe) |
2782 | { | |
2783 | int op = READ_ONCE(sqe->opcode); | |
2784 | ||
2785 | switch (op) { | |
2786 | case IORING_OP_NOP: | |
2787 | case IORING_OP_POLL_REMOVE: | |
5683e540 | 2788 | case IORING_OP_TIMEOUT: |
a320e9fa PB |
2789 | case IORING_OP_TIMEOUT_REMOVE: |
2790 | case IORING_OP_ASYNC_CANCEL: | |
2791 | case IORING_OP_LINK_TIMEOUT: | |
09bb8394 JA |
2792 | return false; |
2793 | default: | |
2794 | return true; | |
2795 | } | |
2796 | } | |
2797 | ||
65e19f54 JA |
2798 | static inline struct file *io_file_from_index(struct io_ring_ctx *ctx, |
2799 | int index) | |
2800 | { | |
2801 | struct fixed_file_table *table; | |
2802 | ||
2803 | table = &ctx->file_table[index >> IORING_FILE_TABLE_SHIFT]; | |
2804 | return table->files[index & IORING_FILE_TABLE_MASK]; | |
2805 | } | |
2806 | ||
a197f664 | 2807 | static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req) |
09bb8394 | 2808 | { |
267bc904 | 2809 | struct sqe_submit *s = &req->submit; |
a197f664 | 2810 | struct io_ring_ctx *ctx = req->ctx; |
09bb8394 JA |
2811 | unsigned flags; |
2812 | int fd; | |
2813 | ||
2814 | flags = READ_ONCE(s->sqe->flags); | |
2815 | fd = READ_ONCE(s->sqe->fd); | |
2816 | ||
4fe2c963 | 2817 | if (flags & IOSQE_IO_DRAIN) |
de0617e4 | 2818 | req->flags |= REQ_F_IO_DRAIN; |
4fe2c963 JL |
2819 | /* |
2820 | * All io need record the previous position, if LINK vs DARIN, | |
2821 | * it can be used to mark the position of the first IO in the | |
2822 | * link list. | |
2823 | */ | |
2824 | req->sequence = s->sequence; | |
de0617e4 | 2825 | |
60c112b0 | 2826 | if (!io_op_needs_file(s->sqe)) |
09bb8394 | 2827 | return 0; |
09bb8394 JA |
2828 | |
2829 | if (flags & IOSQE_FIXED_FILE) { | |
65e19f54 | 2830 | if (unlikely(!ctx->file_table || |
09bb8394 JA |
2831 | (unsigned) fd >= ctx->nr_user_files)) |
2832 | return -EBADF; | |
b7620121 | 2833 | fd = array_index_nospec(fd, ctx->nr_user_files); |
65e19f54 JA |
2834 | req->file = io_file_from_index(ctx, fd); |
2835 | if (!req->file) | |
08a45173 | 2836 | return -EBADF; |
09bb8394 JA |
2837 | req->flags |= REQ_F_FIXED_FILE; |
2838 | } else { | |
2839 | if (s->needs_fixed_file) | |
2840 | return -EBADF; | |
c826bd7a | 2841 | trace_io_uring_file_get(ctx, fd); |
09bb8394 JA |
2842 | req->file = io_file_get(state, fd); |
2843 | if (unlikely(!req->file)) | |
2844 | return -EBADF; | |
2845 | } | |
2846 | ||
2847 | return 0; | |
2848 | } | |
2849 | ||
a197f664 | 2850 | static int io_grab_files(struct io_kiocb *req) |
fcb323cc JA |
2851 | { |
2852 | int ret = -EBADF; | |
a197f664 | 2853 | struct io_ring_ctx *ctx = req->ctx; |
fcb323cc JA |
2854 | |
2855 | rcu_read_lock(); | |
2856 | spin_lock_irq(&ctx->inflight_lock); | |
2857 | /* | |
2858 | * We use the f_ops->flush() handler to ensure that we can flush | |
2859 | * out work accessing these files if the fd is closed. Check if | |
2860 | * the fd has changed since we started down this path, and disallow | |
2861 | * this operation if it has. | |
2862 | */ | |
2863 | if (fcheck(req->submit.ring_fd) == req->submit.ring_file) { | |
2864 | list_add(&req->inflight_entry, &ctx->inflight_list); | |
2865 | req->flags |= REQ_F_INFLIGHT; | |
2866 | req->work.files = current->files; | |
2867 | ret = 0; | |
2868 | } | |
2869 | spin_unlock_irq(&ctx->inflight_lock); | |
2870 | rcu_read_unlock(); | |
2871 | ||
2872 | return ret; | |
2873 | } | |
2874 | ||
2665abfd | 2875 | static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer) |
2b188cc1 | 2876 | { |
ad8a48ac JA |
2877 | struct io_timeout_data *data = container_of(timer, |
2878 | struct io_timeout_data, timer); | |
2879 | struct io_kiocb *req = data->req; | |
2665abfd JA |
2880 | struct io_ring_ctx *ctx = req->ctx; |
2881 | struct io_kiocb *prev = NULL; | |
2882 | unsigned long flags; | |
2665abfd JA |
2883 | |
2884 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
2885 | ||
2886 | /* | |
2887 | * We don't expect the list to be empty, that will only happen if we | |
2888 | * race with the completion of the linked work. | |
2889 | */ | |
2890 | if (!list_empty(&req->list)) { | |
2891 | prev = list_entry(req->list.prev, struct io_kiocb, link_list); | |
5d960724 | 2892 | if (refcount_inc_not_zero(&prev->refs)) { |
76a46e06 | 2893 | list_del_init(&req->list); |
5d960724 JA |
2894 | prev->flags &= ~REQ_F_LINK_TIMEOUT; |
2895 | } else | |
76a46e06 | 2896 | prev = NULL; |
2665abfd JA |
2897 | } |
2898 | ||
2899 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
2900 | ||
2901 | if (prev) { | |
fba38c27 JA |
2902 | if (prev->flags & REQ_F_LINK) |
2903 | prev->flags |= REQ_F_FAIL_LINK; | |
b0dd8a41 JA |
2904 | io_async_find_and_cancel(ctx, req, prev->user_data, NULL, |
2905 | -ETIME); | |
76a46e06 | 2906 | io_put_req(prev); |
47f46768 JA |
2907 | } else { |
2908 | io_cqring_add_event(req, -ETIME); | |
2909 | io_put_req(req); | |
2665abfd | 2910 | } |
2665abfd JA |
2911 | return HRTIMER_NORESTART; |
2912 | } | |
2913 | ||
ad8a48ac | 2914 | static void io_queue_linked_timeout(struct io_kiocb *req) |
2665abfd | 2915 | { |
76a46e06 | 2916 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd | 2917 | |
76a46e06 JA |
2918 | /* |
2919 | * If the list is now empty, then our linked request finished before | |
2920 | * we got a chance to setup the timer | |
2921 | */ | |
2922 | spin_lock_irq(&ctx->completion_lock); | |
2923 | if (!list_empty(&req->list)) { | |
94ae5e77 JA |
2924 | struct io_timeout_data *data = req->timeout.data; |
2925 | ||
ad8a48ac JA |
2926 | data->timer.function = io_link_timeout_fn; |
2927 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), | |
2928 | data->mode); | |
2665abfd | 2929 | } |
76a46e06 | 2930 | spin_unlock_irq(&ctx->completion_lock); |
2665abfd | 2931 | |
2665abfd | 2932 | /* drop submission reference */ |
76a46e06 JA |
2933 | io_put_req(req); |
2934 | } | |
2665abfd | 2935 | |
ad8a48ac | 2936 | static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) |
2665abfd JA |
2937 | { |
2938 | struct io_kiocb *nxt; | |
2939 | ||
2940 | if (!(req->flags & REQ_F_LINK)) | |
2941 | return NULL; | |
2942 | ||
2943 | nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list); | |
76a46e06 JA |
2944 | if (!nxt || nxt->submit.sqe->opcode != IORING_OP_LINK_TIMEOUT) |
2945 | return NULL; | |
2665abfd | 2946 | |
76a46e06 | 2947 | req->flags |= REQ_F_LINK_TIMEOUT; |
76a46e06 | 2948 | return nxt; |
2665abfd JA |
2949 | } |
2950 | ||
0e0702da | 2951 | static void __io_queue_sqe(struct io_kiocb *req) |
2b188cc1 | 2952 | { |
f9bd67f6 PB |
2953 | struct io_kiocb *linked_timeout = io_prep_linked_timeout(req); |
2954 | struct io_kiocb *nxt = NULL; | |
e0c5c576 | 2955 | int ret; |
2b188cc1 | 2956 | |
f9bd67f6 PB |
2957 | ret = io_issue_sqe(req, &nxt, true); |
2958 | if (nxt) | |
2959 | io_queue_async_work(nxt); | |
491381ce JA |
2960 | |
2961 | /* | |
2962 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
2963 | * doesn't support non-blocking read/write attempts | |
2964 | */ | |
2965 | if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) || | |
2966 | (req->flags & REQ_F_MUST_PUNT))) { | |
267bc904 | 2967 | struct sqe_submit *s = &req->submit; |
2b188cc1 JA |
2968 | struct io_uring_sqe *sqe_copy; |
2969 | ||
954dab19 | 2970 | sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL); |
bbad27b2 PB |
2971 | if (!sqe_copy) |
2972 | goto err; | |
e65ef56d | 2973 | |
bbad27b2 PB |
2974 | s->sqe = sqe_copy; |
2975 | req->flags |= REQ_F_FREE_SQE; | |
2976 | ||
2977 | if (req->work.flags & IO_WQ_WORK_NEEDS_FILES) { | |
2978 | ret = io_grab_files(req); | |
2979 | if (ret) | |
2980 | goto err; | |
2b188cc1 | 2981 | } |
bbad27b2 PB |
2982 | |
2983 | /* | |
2984 | * Queued up for async execution, worker will release | |
2985 | * submit reference when the iocb is actually submitted. | |
2986 | */ | |
2987 | io_queue_async_work(req); | |
2988 | return; | |
2b188cc1 | 2989 | } |
e65ef56d | 2990 | |
fcb323cc | 2991 | err: |
76a46e06 | 2992 | /* drop submission reference */ |
ec9c02ad | 2993 | io_put_req(req); |
e65ef56d | 2994 | |
f9bd67f6 | 2995 | if (linked_timeout) { |
76a46e06 | 2996 | if (!ret) |
f9bd67f6 | 2997 | io_queue_linked_timeout(linked_timeout); |
76a46e06 | 2998 | else |
f9bd67f6 | 2999 | io_put_req(linked_timeout); |
76a46e06 JA |
3000 | } |
3001 | ||
e65ef56d | 3002 | /* and drop final reference, if we failed */ |
9e645e11 | 3003 | if (ret) { |
78e19bbe | 3004 | io_cqring_add_event(req, ret); |
9e645e11 JA |
3005 | if (req->flags & REQ_F_LINK) |
3006 | req->flags |= REQ_F_FAIL_LINK; | |
e65ef56d | 3007 | io_put_req(req); |
9e645e11 | 3008 | } |
2b188cc1 JA |
3009 | } |
3010 | ||
0e0702da | 3011 | static void io_queue_sqe(struct io_kiocb *req) |
4fe2c963 JL |
3012 | { |
3013 | int ret; | |
3014 | ||
1b4a51b6 PB |
3015 | if (unlikely(req->ctx->drain_next)) { |
3016 | req->flags |= REQ_F_IO_DRAIN; | |
3017 | req->ctx->drain_next = false; | |
3018 | } | |
3019 | req->ctx->drain_next = (req->flags & REQ_F_DRAIN_LINK); | |
3020 | ||
a197f664 | 3021 | ret = io_req_defer(req); |
4fe2c963 JL |
3022 | if (ret) { |
3023 | if (ret != -EIOCBQUEUED) { | |
78e19bbe | 3024 | io_cqring_add_event(req, ret); |
d3b35796 PB |
3025 | if (req->flags & REQ_F_LINK) |
3026 | req->flags |= REQ_F_FAIL_LINK; | |
78e19bbe | 3027 | io_double_put_req(req); |
4fe2c963 | 3028 | } |
0e0702da JA |
3029 | } else |
3030 | __io_queue_sqe(req); | |
4fe2c963 JL |
3031 | } |
3032 | ||
1b4a51b6 | 3033 | static inline void io_queue_link_head(struct io_kiocb *req) |
4fe2c963 | 3034 | { |
94ae5e77 | 3035 | if (unlikely(req->flags & REQ_F_FAIL_LINK)) { |
1b4a51b6 PB |
3036 | io_cqring_add_event(req, -ECANCELED); |
3037 | io_double_put_req(req); | |
3038 | } else | |
0e0702da | 3039 | io_queue_sqe(req); |
4fe2c963 JL |
3040 | } |
3041 | ||
1b4a51b6 | 3042 | |
9e645e11 JA |
3043 | #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK) |
3044 | ||
a197f664 JL |
3045 | static void io_submit_sqe(struct io_kiocb *req, struct io_submit_state *state, |
3046 | struct io_kiocb **link) | |
9e645e11 | 3047 | { |
267bc904 | 3048 | struct sqe_submit *s = &req->submit; |
a197f664 | 3049 | struct io_ring_ctx *ctx = req->ctx; |
9e645e11 JA |
3050 | int ret; |
3051 | ||
78e19bbe JA |
3052 | req->user_data = s->sqe->user_data; |
3053 | ||
9e645e11 JA |
3054 | /* enforce forwards compatibility on users */ |
3055 | if (unlikely(s->sqe->flags & ~SQE_VALID_FLAGS)) { | |
3056 | ret = -EINVAL; | |
196be95c | 3057 | goto err_req; |
9e645e11 JA |
3058 | } |
3059 | ||
a197f664 | 3060 | ret = io_req_set_file(state, req); |
9e645e11 JA |
3061 | if (unlikely(ret)) { |
3062 | err_req: | |
78e19bbe JA |
3063 | io_cqring_add_event(req, ret); |
3064 | io_double_put_req(req); | |
9e645e11 JA |
3065 | return; |
3066 | } | |
3067 | ||
9e645e11 JA |
3068 | /* |
3069 | * If we already have a head request, queue this one for async | |
3070 | * submittal once the head completes. If we don't have a head but | |
3071 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
3072 | * submitted sync once the chain is complete. If none of those | |
3073 | * conditions are true (normal request), then just queue it. | |
3074 | */ | |
3075 | if (*link) { | |
3076 | struct io_kiocb *prev = *link; | |
bbad27b2 | 3077 | struct io_uring_sqe *sqe_copy; |
9e645e11 | 3078 | |
1b4a51b6 PB |
3079 | if (s->sqe->flags & IOSQE_IO_DRAIN) |
3080 | (*link)->flags |= REQ_F_DRAIN_LINK | REQ_F_IO_DRAIN; | |
3081 | ||
94ae5e77 JA |
3082 | if (READ_ONCE(s->sqe->opcode) == IORING_OP_LINK_TIMEOUT) { |
3083 | ret = io_timeout_setup(req); | |
3084 | /* common setup allows offset being set, we don't */ | |
3085 | if (!ret && s->sqe->off) | |
3086 | ret = -EINVAL; | |
3087 | if (ret) { | |
3088 | prev->flags |= REQ_F_FAIL_LINK; | |
3089 | goto err_req; | |
3090 | } | |
3091 | } | |
3092 | ||
9e645e11 JA |
3093 | sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL); |
3094 | if (!sqe_copy) { | |
3095 | ret = -EAGAIN; | |
3096 | goto err_req; | |
3097 | } | |
3098 | ||
3099 | s->sqe = sqe_copy; | |
94ae5e77 | 3100 | req->flags |= REQ_F_FREE_SQE; |
c826bd7a | 3101 | trace_io_uring_link(ctx, req, prev); |
9e645e11 JA |
3102 | list_add_tail(&req->list, &prev->link_list); |
3103 | } else if (s->sqe->flags & IOSQE_IO_LINK) { | |
3104 | req->flags |= REQ_F_LINK; | |
3105 | ||
9e645e11 JA |
3106 | INIT_LIST_HEAD(&req->link_list); |
3107 | *link = req; | |
3108 | } else { | |
a197f664 | 3109 | io_queue_sqe(req); |
9e645e11 JA |
3110 | } |
3111 | } | |
3112 | ||
9a56a232 JA |
3113 | /* |
3114 | * Batched submission is done, ensure local IO is flushed out. | |
3115 | */ | |
3116 | static void io_submit_state_end(struct io_submit_state *state) | |
3117 | { | |
3118 | blk_finish_plug(&state->plug); | |
3d6770fb | 3119 | io_file_put(state); |
2579f913 JA |
3120 | if (state->free_reqs) |
3121 | kmem_cache_free_bulk(req_cachep, state->free_reqs, | |
3122 | &state->reqs[state->cur_req]); | |
9a56a232 JA |
3123 | } |
3124 | ||
3125 | /* | |
3126 | * Start submission side cache. | |
3127 | */ | |
3128 | static void io_submit_state_start(struct io_submit_state *state, | |
3129 | struct io_ring_ctx *ctx, unsigned max_ios) | |
3130 | { | |
3131 | blk_start_plug(&state->plug); | |
2579f913 | 3132 | state->free_reqs = 0; |
9a56a232 JA |
3133 | state->file = NULL; |
3134 | state->ios_left = max_ios; | |
3135 | } | |
3136 | ||
2b188cc1 JA |
3137 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
3138 | { | |
75b28aff | 3139 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 3140 | |
75b28aff | 3141 | if (ctx->cached_sq_head != READ_ONCE(rings->sq.head)) { |
2b188cc1 JA |
3142 | /* |
3143 | * Ensure any loads from the SQEs are done at this point, | |
3144 | * since once we write the new head, the application could | |
3145 | * write new data to them. | |
3146 | */ | |
75b28aff | 3147 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); |
2b188cc1 JA |
3148 | } |
3149 | } | |
3150 | ||
2b188cc1 JA |
3151 | /* |
3152 | * Fetch an sqe, if one is available. Note that s->sqe will point to memory | |
3153 | * that is mapped by userspace. This means that care needs to be taken to | |
3154 | * ensure that reads are stable, as we cannot rely on userspace always | |
3155 | * being a good citizen. If members of the sqe are validated and then later | |
3156 | * used, it's important that those reads are done through READ_ONCE() to | |
3157 | * prevent a re-load down the line. | |
3158 | */ | |
3159 | static bool io_get_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s) | |
3160 | { | |
75b28aff HV |
3161 | struct io_rings *rings = ctx->rings; |
3162 | u32 *sq_array = ctx->sq_array; | |
2b188cc1 JA |
3163 | unsigned head; |
3164 | ||
3165 | /* | |
3166 | * The cached sq head (or cq tail) serves two purposes: | |
3167 | * | |
3168 | * 1) allows us to batch the cost of updating the user visible | |
3169 | * head updates. | |
3170 | * 2) allows the kernel side to track the head on its own, even | |
3171 | * though the application is the one updating it. | |
3172 | */ | |
3173 | head = ctx->cached_sq_head; | |
e523a29c | 3174 | /* make sure SQ entry isn't read before tail */ |
9835d6fa | 3175 | if (unlikely(head == smp_load_acquire(&rings->sq.tail))) |
2b188cc1 JA |
3176 | return false; |
3177 | ||
75b28aff | 3178 | head = READ_ONCE(sq_array[head & ctx->sq_mask]); |
9835d6fa | 3179 | if (likely(head < ctx->sq_entries)) { |
fcb323cc | 3180 | s->ring_file = NULL; |
2b188cc1 | 3181 | s->sqe = &ctx->sq_sqes[head]; |
8776f3fa | 3182 | s->sequence = ctx->cached_sq_head; |
2b188cc1 JA |
3183 | ctx->cached_sq_head++; |
3184 | return true; | |
3185 | } | |
3186 | ||
3187 | /* drop invalid entries */ | |
3188 | ctx->cached_sq_head++; | |
498ccd9e JA |
3189 | ctx->cached_sq_dropped++; |
3190 | WRITE_ONCE(rings->sq_dropped, ctx->cached_sq_dropped); | |
2b188cc1 JA |
3191 | return false; |
3192 | } | |
3193 | ||
fb5ccc98 | 3194 | static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr, |
ae9428ca PB |
3195 | struct file *ring_file, int ring_fd, |
3196 | struct mm_struct **mm, bool async) | |
6c271ce2 JA |
3197 | { |
3198 | struct io_submit_state state, *statep = NULL; | |
9e645e11 | 3199 | struct io_kiocb *link = NULL; |
9e645e11 | 3200 | int i, submitted = 0; |
95a1b3ff | 3201 | bool mm_fault = false; |
6c271ce2 | 3202 | |
c4a2ed72 JA |
3203 | /* if we have a backlog and couldn't flush it all, return BUSY */ |
3204 | if (!list_empty(&ctx->cq_overflow_list) && | |
3205 | !io_cqring_overflow_flush(ctx, false)) | |
1d7bb1d5 | 3206 | return -EBUSY; |
6c271ce2 JA |
3207 | |
3208 | if (nr > IO_PLUG_THRESHOLD) { | |
3209 | io_submit_state_start(&state, ctx, nr); | |
3210 | statep = &state; | |
3211 | } | |
3212 | ||
3213 | for (i = 0; i < nr; i++) { | |
196be95c | 3214 | struct io_kiocb *req; |
50585b9a | 3215 | unsigned int sqe_flags; |
fb5ccc98 | 3216 | |
196be95c PB |
3217 | req = io_get_req(ctx, statep); |
3218 | if (unlikely(!req)) { | |
3219 | if (!submitted) | |
3220 | submitted = -EAGAIN; | |
fb5ccc98 | 3221 | break; |
196be95c | 3222 | } |
50585b9a | 3223 | if (!io_get_sqring(ctx, &req->submit)) { |
196be95c PB |
3224 | __io_free_req(req); |
3225 | break; | |
3226 | } | |
fb5ccc98 | 3227 | |
50585b9a | 3228 | if (io_sqe_needs_user(req->submit.sqe) && !*mm) { |
95a1b3ff PB |
3229 | mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm); |
3230 | if (!mm_fault) { | |
3231 | use_mm(ctx->sqo_mm); | |
3232 | *mm = ctx->sqo_mm; | |
3233 | } | |
9e645e11 | 3234 | } |
9e645e11 | 3235 | |
50585b9a PB |
3236 | sqe_flags = req->submit.sqe->flags; |
3237 | ||
50585b9a PB |
3238 | req->submit.ring_file = ring_file; |
3239 | req->submit.ring_fd = ring_fd; | |
3240 | req->submit.has_user = *mm != NULL; | |
3241 | req->submit.in_async = async; | |
3242 | req->submit.needs_fixed_file = async; | |
3243 | trace_io_uring_submit_sqe(ctx, req->submit.sqe->user_data, | |
3244 | true, async); | |
a197f664 | 3245 | io_submit_sqe(req, statep, &link); |
95a1b3ff | 3246 | submitted++; |
e5eb6366 PB |
3247 | |
3248 | /* | |
3249 | * If previous wasn't linked and we have a linked command, | |
3250 | * that's the end of the chain. Submit the previous link. | |
3251 | */ | |
50585b9a | 3252 | if (!(sqe_flags & IOSQE_IO_LINK) && link) { |
1b4a51b6 | 3253 | io_queue_link_head(link); |
e5eb6366 | 3254 | link = NULL; |
6c271ce2 | 3255 | } |
6c271ce2 JA |
3256 | } |
3257 | ||
9e645e11 | 3258 | if (link) |
1b4a51b6 | 3259 | io_queue_link_head(link); |
6c271ce2 JA |
3260 | if (statep) |
3261 | io_submit_state_end(&state); | |
3262 | ||
ae9428ca PB |
3263 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
3264 | io_commit_sqring(ctx); | |
3265 | ||
6c271ce2 JA |
3266 | return submitted; |
3267 | } | |
3268 | ||
3269 | static int io_sq_thread(void *data) | |
3270 | { | |
6c271ce2 JA |
3271 | struct io_ring_ctx *ctx = data; |
3272 | struct mm_struct *cur_mm = NULL; | |
181e448d | 3273 | const struct cred *old_cred; |
6c271ce2 JA |
3274 | mm_segment_t old_fs; |
3275 | DEFINE_WAIT(wait); | |
3276 | unsigned inflight; | |
3277 | unsigned long timeout; | |
c1edbf5f | 3278 | int ret; |
6c271ce2 | 3279 | |
206aefde | 3280 | complete(&ctx->completions[1]); |
a4c0b3de | 3281 | |
6c271ce2 JA |
3282 | old_fs = get_fs(); |
3283 | set_fs(USER_DS); | |
181e448d | 3284 | old_cred = override_creds(ctx->creds); |
6c271ce2 | 3285 | |
c1edbf5f | 3286 | ret = timeout = inflight = 0; |
2bbcd6d3 | 3287 | while (!kthread_should_park()) { |
fb5ccc98 | 3288 | unsigned int to_submit; |
6c271ce2 JA |
3289 | |
3290 | if (inflight) { | |
3291 | unsigned nr_events = 0; | |
3292 | ||
3293 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
2b2ed975 JA |
3294 | /* |
3295 | * inflight is the count of the maximum possible | |
3296 | * entries we submitted, but it can be smaller | |
3297 | * if we dropped some of them. If we don't have | |
3298 | * poll entries available, then we know that we | |
3299 | * have nothing left to poll for. Reset the | |
3300 | * inflight count to zero in that case. | |
3301 | */ | |
3302 | mutex_lock(&ctx->uring_lock); | |
3303 | if (!list_empty(&ctx->poll_list)) | |
3304 | __io_iopoll_check(ctx, &nr_events, 0); | |
3305 | else | |
3306 | inflight = 0; | |
3307 | mutex_unlock(&ctx->uring_lock); | |
6c271ce2 JA |
3308 | } else { |
3309 | /* | |
3310 | * Normal IO, just pretend everything completed. | |
3311 | * We don't have to poll completions for that. | |
3312 | */ | |
3313 | nr_events = inflight; | |
3314 | } | |
3315 | ||
3316 | inflight -= nr_events; | |
3317 | if (!inflight) | |
3318 | timeout = jiffies + ctx->sq_thread_idle; | |
3319 | } | |
3320 | ||
fb5ccc98 | 3321 | to_submit = io_sqring_entries(ctx); |
c1edbf5f JA |
3322 | |
3323 | /* | |
3324 | * If submit got -EBUSY, flag us as needing the application | |
3325 | * to enter the kernel to reap and flush events. | |
3326 | */ | |
3327 | if (!to_submit || ret == -EBUSY) { | |
6c271ce2 JA |
3328 | /* |
3329 | * We're polling. If we're within the defined idle | |
3330 | * period, then let us spin without work before going | |
c1edbf5f JA |
3331 | * to sleep. The exception is if we got EBUSY doing |
3332 | * more IO, we should wait for the application to | |
3333 | * reap events and wake us up. | |
6c271ce2 | 3334 | */ |
c1edbf5f JA |
3335 | if (inflight || |
3336 | (!time_after(jiffies, timeout) && ret != -EBUSY)) { | |
9831a90c | 3337 | cond_resched(); |
6c271ce2 JA |
3338 | continue; |
3339 | } | |
3340 | ||
3341 | /* | |
3342 | * Drop cur_mm before scheduling, we can't hold it for | |
3343 | * long periods (or over schedule()). Do this before | |
3344 | * adding ourselves to the waitqueue, as the unuse/drop | |
3345 | * may sleep. | |
3346 | */ | |
3347 | if (cur_mm) { | |
3348 | unuse_mm(cur_mm); | |
3349 | mmput(cur_mm); | |
3350 | cur_mm = NULL; | |
3351 | } | |
3352 | ||
3353 | prepare_to_wait(&ctx->sqo_wait, &wait, | |
3354 | TASK_INTERRUPTIBLE); | |
3355 | ||
3356 | /* Tell userspace we may need a wakeup call */ | |
75b28aff | 3357 | ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP; |
0d7bae69 SB |
3358 | /* make sure to read SQ tail after writing flags */ |
3359 | smp_mb(); | |
6c271ce2 | 3360 | |
fb5ccc98 | 3361 | to_submit = io_sqring_entries(ctx); |
c1edbf5f | 3362 | if (!to_submit || ret == -EBUSY) { |
2bbcd6d3 | 3363 | if (kthread_should_park()) { |
6c271ce2 JA |
3364 | finish_wait(&ctx->sqo_wait, &wait); |
3365 | break; | |
3366 | } | |
3367 | if (signal_pending(current)) | |
3368 | flush_signals(current); | |
3369 | schedule(); | |
3370 | finish_wait(&ctx->sqo_wait, &wait); | |
3371 | ||
75b28aff | 3372 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
3373 | continue; |
3374 | } | |
3375 | finish_wait(&ctx->sqo_wait, &wait); | |
3376 | ||
75b28aff | 3377 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
3378 | } |
3379 | ||
fb5ccc98 | 3380 | to_submit = min(to_submit, ctx->sq_entries); |
1d7bb1d5 JA |
3381 | ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true); |
3382 | if (ret > 0) | |
3383 | inflight += ret; | |
6c271ce2 JA |
3384 | } |
3385 | ||
3386 | set_fs(old_fs); | |
3387 | if (cur_mm) { | |
3388 | unuse_mm(cur_mm); | |
3389 | mmput(cur_mm); | |
3390 | } | |
181e448d | 3391 | revert_creds(old_cred); |
06058632 | 3392 | |
2bbcd6d3 | 3393 | kthread_parkme(); |
06058632 | 3394 | |
6c271ce2 JA |
3395 | return 0; |
3396 | } | |
3397 | ||
bda52162 JA |
3398 | struct io_wait_queue { |
3399 | struct wait_queue_entry wq; | |
3400 | struct io_ring_ctx *ctx; | |
3401 | unsigned to_wait; | |
3402 | unsigned nr_timeouts; | |
3403 | }; | |
3404 | ||
1d7bb1d5 | 3405 | static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush) |
bda52162 JA |
3406 | { |
3407 | struct io_ring_ctx *ctx = iowq->ctx; | |
3408 | ||
3409 | /* | |
3410 | * Wake up if we have enough events, or if a timeout occured since we | |
3411 | * started waiting. For timeouts, we always want to return to userspace, | |
3412 | * regardless of event count. | |
3413 | */ | |
1d7bb1d5 | 3414 | return io_cqring_events(ctx, noflush) >= iowq->to_wait || |
bda52162 JA |
3415 | atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
3416 | } | |
3417 | ||
3418 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
3419 | int wake_flags, void *key) | |
3420 | { | |
3421 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, | |
3422 | wq); | |
3423 | ||
1d7bb1d5 JA |
3424 | /* use noflush == true, as we can't safely rely on locking context */ |
3425 | if (!io_should_wake(iowq, true)) | |
bda52162 JA |
3426 | return -1; |
3427 | ||
3428 | return autoremove_wake_function(curr, mode, wake_flags, key); | |
3429 | } | |
3430 | ||
2b188cc1 JA |
3431 | /* |
3432 | * Wait until events become available, if we don't already have some. The | |
3433 | * application must reap them itself, as they reside on the shared cq ring. | |
3434 | */ | |
3435 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
3436 | const sigset_t __user *sig, size_t sigsz) | |
3437 | { | |
bda52162 JA |
3438 | struct io_wait_queue iowq = { |
3439 | .wq = { | |
3440 | .private = current, | |
3441 | .func = io_wake_function, | |
3442 | .entry = LIST_HEAD_INIT(iowq.wq.entry), | |
3443 | }, | |
3444 | .ctx = ctx, | |
3445 | .to_wait = min_events, | |
3446 | }; | |
75b28aff | 3447 | struct io_rings *rings = ctx->rings; |
e9ffa5c2 | 3448 | int ret = 0; |
2b188cc1 | 3449 | |
1d7bb1d5 | 3450 | if (io_cqring_events(ctx, false) >= min_events) |
2b188cc1 JA |
3451 | return 0; |
3452 | ||
3453 | if (sig) { | |
9e75ad5d AB |
3454 | #ifdef CONFIG_COMPAT |
3455 | if (in_compat_syscall()) | |
3456 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 3457 | sigsz); |
9e75ad5d AB |
3458 | else |
3459 | #endif | |
b772434b | 3460 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 3461 | |
2b188cc1 JA |
3462 | if (ret) |
3463 | return ret; | |
3464 | } | |
3465 | ||
bda52162 | 3466 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
c826bd7a | 3467 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 JA |
3468 | do { |
3469 | prepare_to_wait_exclusive(&ctx->wait, &iowq.wq, | |
3470 | TASK_INTERRUPTIBLE); | |
1d7bb1d5 | 3471 | if (io_should_wake(&iowq, false)) |
bda52162 JA |
3472 | break; |
3473 | schedule(); | |
3474 | if (signal_pending(current)) { | |
e9ffa5c2 | 3475 | ret = -EINTR; |
bda52162 JA |
3476 | break; |
3477 | } | |
3478 | } while (1); | |
3479 | finish_wait(&ctx->wait, &iowq.wq); | |
3480 | ||
e9ffa5c2 | 3481 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 3482 | |
75b28aff | 3483 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
3484 | } |
3485 | ||
6b06314c JA |
3486 | static void __io_sqe_files_unregister(struct io_ring_ctx *ctx) |
3487 | { | |
3488 | #if defined(CONFIG_UNIX) | |
3489 | if (ctx->ring_sock) { | |
3490 | struct sock *sock = ctx->ring_sock->sk; | |
3491 | struct sk_buff *skb; | |
3492 | ||
3493 | while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) | |
3494 | kfree_skb(skb); | |
3495 | } | |
3496 | #else | |
3497 | int i; | |
3498 | ||
65e19f54 JA |
3499 | for (i = 0; i < ctx->nr_user_files; i++) { |
3500 | struct file *file; | |
3501 | ||
3502 | file = io_file_from_index(ctx, i); | |
3503 | if (file) | |
3504 | fput(file); | |
3505 | } | |
6b06314c JA |
3506 | #endif |
3507 | } | |
3508 | ||
3509 | static int io_sqe_files_unregister(struct io_ring_ctx *ctx) | |
3510 | { | |
65e19f54 JA |
3511 | unsigned nr_tables, i; |
3512 | ||
3513 | if (!ctx->file_table) | |
6b06314c JA |
3514 | return -ENXIO; |
3515 | ||
3516 | __io_sqe_files_unregister(ctx); | |
65e19f54 JA |
3517 | nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE); |
3518 | for (i = 0; i < nr_tables; i++) | |
3519 | kfree(ctx->file_table[i].files); | |
3520 | kfree(ctx->file_table); | |
3521 | ctx->file_table = NULL; | |
6b06314c JA |
3522 | ctx->nr_user_files = 0; |
3523 | return 0; | |
3524 | } | |
3525 | ||
6c271ce2 JA |
3526 | static void io_sq_thread_stop(struct io_ring_ctx *ctx) |
3527 | { | |
3528 | if (ctx->sqo_thread) { | |
206aefde | 3529 | wait_for_completion(&ctx->completions[1]); |
2bbcd6d3 RP |
3530 | /* |
3531 | * The park is a bit of a work-around, without it we get | |
3532 | * warning spews on shutdown with SQPOLL set and affinity | |
3533 | * set to a single CPU. | |
3534 | */ | |
06058632 | 3535 | kthread_park(ctx->sqo_thread); |
6c271ce2 JA |
3536 | kthread_stop(ctx->sqo_thread); |
3537 | ctx->sqo_thread = NULL; | |
3538 | } | |
3539 | } | |
3540 | ||
6b06314c JA |
3541 | static void io_finish_async(struct io_ring_ctx *ctx) |
3542 | { | |
6c271ce2 JA |
3543 | io_sq_thread_stop(ctx); |
3544 | ||
561fb04a JA |
3545 | if (ctx->io_wq) { |
3546 | io_wq_destroy(ctx->io_wq); | |
3547 | ctx->io_wq = NULL; | |
6b06314c JA |
3548 | } |
3549 | } | |
3550 | ||
3551 | #if defined(CONFIG_UNIX) | |
3552 | static void io_destruct_skb(struct sk_buff *skb) | |
3553 | { | |
3554 | struct io_ring_ctx *ctx = skb->sk->sk_user_data; | |
8a997340 | 3555 | |
561fb04a JA |
3556 | if (ctx->io_wq) |
3557 | io_wq_flush(ctx->io_wq); | |
6b06314c | 3558 | |
6b06314c JA |
3559 | unix_destruct_scm(skb); |
3560 | } | |
3561 | ||
3562 | /* | |
3563 | * Ensure the UNIX gc is aware of our file set, so we are certain that | |
3564 | * the io_uring can be safely unregistered on process exit, even if we have | |
3565 | * loops in the file referencing. | |
3566 | */ | |
3567 | static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset) | |
3568 | { | |
3569 | struct sock *sk = ctx->ring_sock->sk; | |
3570 | struct scm_fp_list *fpl; | |
3571 | struct sk_buff *skb; | |
08a45173 | 3572 | int i, nr_files; |
6b06314c JA |
3573 | |
3574 | if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) { | |
3575 | unsigned long inflight = ctx->user->unix_inflight + nr; | |
3576 | ||
3577 | if (inflight > task_rlimit(current, RLIMIT_NOFILE)) | |
3578 | return -EMFILE; | |
3579 | } | |
3580 | ||
3581 | fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); | |
3582 | if (!fpl) | |
3583 | return -ENOMEM; | |
3584 | ||
3585 | skb = alloc_skb(0, GFP_KERNEL); | |
3586 | if (!skb) { | |
3587 | kfree(fpl); | |
3588 | return -ENOMEM; | |
3589 | } | |
3590 | ||
3591 | skb->sk = sk; | |
6b06314c | 3592 | |
08a45173 | 3593 | nr_files = 0; |
6b06314c JA |
3594 | fpl->user = get_uid(ctx->user); |
3595 | for (i = 0; i < nr; i++) { | |
65e19f54 JA |
3596 | struct file *file = io_file_from_index(ctx, i + offset); |
3597 | ||
3598 | if (!file) | |
08a45173 | 3599 | continue; |
65e19f54 | 3600 | fpl->fp[nr_files] = get_file(file); |
08a45173 JA |
3601 | unix_inflight(fpl->user, fpl->fp[nr_files]); |
3602 | nr_files++; | |
6b06314c JA |
3603 | } |
3604 | ||
08a45173 JA |
3605 | if (nr_files) { |
3606 | fpl->max = SCM_MAX_FD; | |
3607 | fpl->count = nr_files; | |
3608 | UNIXCB(skb).fp = fpl; | |
3609 | skb->destructor = io_destruct_skb; | |
3610 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); | |
3611 | skb_queue_head(&sk->sk_receive_queue, skb); | |
6b06314c | 3612 | |
08a45173 JA |
3613 | for (i = 0; i < nr_files; i++) |
3614 | fput(fpl->fp[i]); | |
3615 | } else { | |
3616 | kfree_skb(skb); | |
3617 | kfree(fpl); | |
3618 | } | |
6b06314c JA |
3619 | |
3620 | return 0; | |
3621 | } | |
3622 | ||
3623 | /* | |
3624 | * If UNIX sockets are enabled, fd passing can cause a reference cycle which | |
3625 | * causes regular reference counting to break down. We rely on the UNIX | |
3626 | * garbage collection to take care of this problem for us. | |
3627 | */ | |
3628 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
3629 | { | |
3630 | unsigned left, total; | |
3631 | int ret = 0; | |
3632 | ||
3633 | total = 0; | |
3634 | left = ctx->nr_user_files; | |
3635 | while (left) { | |
3636 | unsigned this_files = min_t(unsigned, left, SCM_MAX_FD); | |
6b06314c JA |
3637 | |
3638 | ret = __io_sqe_files_scm(ctx, this_files, total); | |
3639 | if (ret) | |
3640 | break; | |
3641 | left -= this_files; | |
3642 | total += this_files; | |
3643 | } | |
3644 | ||
3645 | if (!ret) | |
3646 | return 0; | |
3647 | ||
3648 | while (total < ctx->nr_user_files) { | |
65e19f54 JA |
3649 | struct file *file = io_file_from_index(ctx, total); |
3650 | ||
3651 | if (file) | |
3652 | fput(file); | |
6b06314c JA |
3653 | total++; |
3654 | } | |
3655 | ||
3656 | return ret; | |
3657 | } | |
3658 | #else | |
3659 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
3660 | { | |
3661 | return 0; | |
3662 | } | |
3663 | #endif | |
3664 | ||
65e19f54 JA |
3665 | static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables, |
3666 | unsigned nr_files) | |
3667 | { | |
3668 | int i; | |
3669 | ||
3670 | for (i = 0; i < nr_tables; i++) { | |
3671 | struct fixed_file_table *table = &ctx->file_table[i]; | |
3672 | unsigned this_files; | |
3673 | ||
3674 | this_files = min(nr_files, IORING_MAX_FILES_TABLE); | |
3675 | table->files = kcalloc(this_files, sizeof(struct file *), | |
3676 | GFP_KERNEL); | |
3677 | if (!table->files) | |
3678 | break; | |
3679 | nr_files -= this_files; | |
3680 | } | |
3681 | ||
3682 | if (i == nr_tables) | |
3683 | return 0; | |
3684 | ||
3685 | for (i = 0; i < nr_tables; i++) { | |
3686 | struct fixed_file_table *table = &ctx->file_table[i]; | |
3687 | kfree(table->files); | |
3688 | } | |
3689 | return 1; | |
3690 | } | |
3691 | ||
6b06314c JA |
3692 | static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, |
3693 | unsigned nr_args) | |
3694 | { | |
3695 | __s32 __user *fds = (__s32 __user *) arg; | |
65e19f54 | 3696 | unsigned nr_tables; |
6b06314c JA |
3697 | int fd, ret = 0; |
3698 | unsigned i; | |
3699 | ||
65e19f54 | 3700 | if (ctx->file_table) |
6b06314c JA |
3701 | return -EBUSY; |
3702 | if (!nr_args) | |
3703 | return -EINVAL; | |
3704 | if (nr_args > IORING_MAX_FIXED_FILES) | |
3705 | return -EMFILE; | |
3706 | ||
65e19f54 JA |
3707 | nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE); |
3708 | ctx->file_table = kcalloc(nr_tables, sizeof(struct fixed_file_table), | |
3709 | GFP_KERNEL); | |
3710 | if (!ctx->file_table) | |
6b06314c JA |
3711 | return -ENOMEM; |
3712 | ||
65e19f54 JA |
3713 | if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) { |
3714 | kfree(ctx->file_table); | |
46568e9b | 3715 | ctx->file_table = NULL; |
65e19f54 JA |
3716 | return -ENOMEM; |
3717 | } | |
3718 | ||
08a45173 | 3719 | for (i = 0; i < nr_args; i++, ctx->nr_user_files++) { |
65e19f54 JA |
3720 | struct fixed_file_table *table; |
3721 | unsigned index; | |
3722 | ||
6b06314c JA |
3723 | ret = -EFAULT; |
3724 | if (copy_from_user(&fd, &fds[i], sizeof(fd))) | |
3725 | break; | |
08a45173 JA |
3726 | /* allow sparse sets */ |
3727 | if (fd == -1) { | |
3728 | ret = 0; | |
3729 | continue; | |
3730 | } | |
6b06314c | 3731 | |
65e19f54 JA |
3732 | table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT]; |
3733 | index = i & IORING_FILE_TABLE_MASK; | |
3734 | table->files[index] = fget(fd); | |
6b06314c JA |
3735 | |
3736 | ret = -EBADF; | |
65e19f54 | 3737 | if (!table->files[index]) |
6b06314c JA |
3738 | break; |
3739 | /* | |
3740 | * Don't allow io_uring instances to be registered. If UNIX | |
3741 | * isn't enabled, then this causes a reference cycle and this | |
3742 | * instance can never get freed. If UNIX is enabled we'll | |
3743 | * handle it just fine, but there's still no point in allowing | |
3744 | * a ring fd as it doesn't support regular read/write anyway. | |
3745 | */ | |
65e19f54 JA |
3746 | if (table->files[index]->f_op == &io_uring_fops) { |
3747 | fput(table->files[index]); | |
6b06314c JA |
3748 | break; |
3749 | } | |
6b06314c JA |
3750 | ret = 0; |
3751 | } | |
3752 | ||
3753 | if (ret) { | |
65e19f54 JA |
3754 | for (i = 0; i < ctx->nr_user_files; i++) { |
3755 | struct file *file; | |
6b06314c | 3756 | |
65e19f54 JA |
3757 | file = io_file_from_index(ctx, i); |
3758 | if (file) | |
3759 | fput(file); | |
3760 | } | |
3761 | for (i = 0; i < nr_tables; i++) | |
3762 | kfree(ctx->file_table[i].files); | |
6b06314c | 3763 | |
65e19f54 JA |
3764 | kfree(ctx->file_table); |
3765 | ctx->file_table = NULL; | |
6b06314c JA |
3766 | ctx->nr_user_files = 0; |
3767 | return ret; | |
3768 | } | |
3769 | ||
3770 | ret = io_sqe_files_scm(ctx); | |
3771 | if (ret) | |
3772 | io_sqe_files_unregister(ctx); | |
3773 | ||
3774 | return ret; | |
3775 | } | |
3776 | ||
c3a31e60 JA |
3777 | static void io_sqe_file_unregister(struct io_ring_ctx *ctx, int index) |
3778 | { | |
3779 | #if defined(CONFIG_UNIX) | |
65e19f54 | 3780 | struct file *file = io_file_from_index(ctx, index); |
c3a31e60 JA |
3781 | struct sock *sock = ctx->ring_sock->sk; |
3782 | struct sk_buff_head list, *head = &sock->sk_receive_queue; | |
3783 | struct sk_buff *skb; | |
3784 | int i; | |
3785 | ||
3786 | __skb_queue_head_init(&list); | |
3787 | ||
3788 | /* | |
3789 | * Find the skb that holds this file in its SCM_RIGHTS. When found, | |
3790 | * remove this entry and rearrange the file array. | |
3791 | */ | |
3792 | skb = skb_dequeue(head); | |
3793 | while (skb) { | |
3794 | struct scm_fp_list *fp; | |
3795 | ||
3796 | fp = UNIXCB(skb).fp; | |
3797 | for (i = 0; i < fp->count; i++) { | |
3798 | int left; | |
3799 | ||
3800 | if (fp->fp[i] != file) | |
3801 | continue; | |
3802 | ||
3803 | unix_notinflight(fp->user, fp->fp[i]); | |
3804 | left = fp->count - 1 - i; | |
3805 | if (left) { | |
3806 | memmove(&fp->fp[i], &fp->fp[i + 1], | |
3807 | left * sizeof(struct file *)); | |
3808 | } | |
3809 | fp->count--; | |
3810 | if (!fp->count) { | |
3811 | kfree_skb(skb); | |
3812 | skb = NULL; | |
3813 | } else { | |
3814 | __skb_queue_tail(&list, skb); | |
3815 | } | |
3816 | fput(file); | |
3817 | file = NULL; | |
3818 | break; | |
3819 | } | |
3820 | ||
3821 | if (!file) | |
3822 | break; | |
3823 | ||
3824 | __skb_queue_tail(&list, skb); | |
3825 | ||
3826 | skb = skb_dequeue(head); | |
3827 | } | |
3828 | ||
3829 | if (skb_peek(&list)) { | |
3830 | spin_lock_irq(&head->lock); | |
3831 | while ((skb = __skb_dequeue(&list)) != NULL) | |
3832 | __skb_queue_tail(head, skb); | |
3833 | spin_unlock_irq(&head->lock); | |
3834 | } | |
3835 | #else | |
65e19f54 | 3836 | fput(io_file_from_index(ctx, index)); |
c3a31e60 JA |
3837 | #endif |
3838 | } | |
3839 | ||
3840 | static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file, | |
3841 | int index) | |
3842 | { | |
3843 | #if defined(CONFIG_UNIX) | |
3844 | struct sock *sock = ctx->ring_sock->sk; | |
3845 | struct sk_buff_head *head = &sock->sk_receive_queue; | |
3846 | struct sk_buff *skb; | |
3847 | ||
3848 | /* | |
3849 | * See if we can merge this file into an existing skb SCM_RIGHTS | |
3850 | * file set. If there's no room, fall back to allocating a new skb | |
3851 | * and filling it in. | |
3852 | */ | |
3853 | spin_lock_irq(&head->lock); | |
3854 | skb = skb_peek(head); | |
3855 | if (skb) { | |
3856 | struct scm_fp_list *fpl = UNIXCB(skb).fp; | |
3857 | ||
3858 | if (fpl->count < SCM_MAX_FD) { | |
3859 | __skb_unlink(skb, head); | |
3860 | spin_unlock_irq(&head->lock); | |
3861 | fpl->fp[fpl->count] = get_file(file); | |
3862 | unix_inflight(fpl->user, fpl->fp[fpl->count]); | |
3863 | fpl->count++; | |
3864 | spin_lock_irq(&head->lock); | |
3865 | __skb_queue_head(head, skb); | |
3866 | } else { | |
3867 | skb = NULL; | |
3868 | } | |
3869 | } | |
3870 | spin_unlock_irq(&head->lock); | |
3871 | ||
3872 | if (skb) { | |
3873 | fput(file); | |
3874 | return 0; | |
3875 | } | |
3876 | ||
3877 | return __io_sqe_files_scm(ctx, 1, index); | |
3878 | #else | |
3879 | return 0; | |
3880 | #endif | |
3881 | } | |
3882 | ||
3883 | static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg, | |
3884 | unsigned nr_args) | |
3885 | { | |
3886 | struct io_uring_files_update up; | |
3887 | __s32 __user *fds; | |
3888 | int fd, i, err; | |
3889 | __u32 done; | |
3890 | ||
65e19f54 | 3891 | if (!ctx->file_table) |
c3a31e60 JA |
3892 | return -ENXIO; |
3893 | if (!nr_args) | |
3894 | return -EINVAL; | |
3895 | if (copy_from_user(&up, arg, sizeof(up))) | |
3896 | return -EFAULT; | |
3897 | if (check_add_overflow(up.offset, nr_args, &done)) | |
3898 | return -EOVERFLOW; | |
3899 | if (done > ctx->nr_user_files) | |
3900 | return -EINVAL; | |
3901 | ||
3902 | done = 0; | |
3903 | fds = (__s32 __user *) up.fds; | |
3904 | while (nr_args) { | |
65e19f54 JA |
3905 | struct fixed_file_table *table; |
3906 | unsigned index; | |
3907 | ||
c3a31e60 JA |
3908 | err = 0; |
3909 | if (copy_from_user(&fd, &fds[done], sizeof(fd))) { | |
3910 | err = -EFAULT; | |
3911 | break; | |
3912 | } | |
3913 | i = array_index_nospec(up.offset, ctx->nr_user_files); | |
65e19f54 JA |
3914 | table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT]; |
3915 | index = i & IORING_FILE_TABLE_MASK; | |
3916 | if (table->files[index]) { | |
c3a31e60 | 3917 | io_sqe_file_unregister(ctx, i); |
65e19f54 | 3918 | table->files[index] = NULL; |
c3a31e60 JA |
3919 | } |
3920 | if (fd != -1) { | |
3921 | struct file *file; | |
3922 | ||
3923 | file = fget(fd); | |
3924 | if (!file) { | |
3925 | err = -EBADF; | |
3926 | break; | |
3927 | } | |
3928 | /* | |
3929 | * Don't allow io_uring instances to be registered. If | |
3930 | * UNIX isn't enabled, then this causes a reference | |
3931 | * cycle and this instance can never get freed. If UNIX | |
3932 | * is enabled we'll handle it just fine, but there's | |
3933 | * still no point in allowing a ring fd as it doesn't | |
3934 | * support regular read/write anyway. | |
3935 | */ | |
3936 | if (file->f_op == &io_uring_fops) { | |
3937 | fput(file); | |
3938 | err = -EBADF; | |
3939 | break; | |
3940 | } | |
65e19f54 | 3941 | table->files[index] = file; |
c3a31e60 JA |
3942 | err = io_sqe_file_register(ctx, file, i); |
3943 | if (err) | |
3944 | break; | |
3945 | } | |
3946 | nr_args--; | |
3947 | done++; | |
3948 | up.offset++; | |
3949 | } | |
3950 | ||
3951 | return done ? done : err; | |
3952 | } | |
3953 | ||
7d723065 JA |
3954 | static void io_put_work(struct io_wq_work *work) |
3955 | { | |
3956 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
3957 | ||
3958 | io_put_req(req); | |
3959 | } | |
3960 | ||
3961 | static void io_get_work(struct io_wq_work *work) | |
3962 | { | |
3963 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
3964 | ||
3965 | refcount_inc(&req->refs); | |
3966 | } | |
3967 | ||
6c271ce2 JA |
3968 | static int io_sq_offload_start(struct io_ring_ctx *ctx, |
3969 | struct io_uring_params *p) | |
2b188cc1 | 3970 | { |
576a347b | 3971 | struct io_wq_data data; |
561fb04a | 3972 | unsigned concurrency; |
2b188cc1 JA |
3973 | int ret; |
3974 | ||
6c271ce2 | 3975 | init_waitqueue_head(&ctx->sqo_wait); |
2b188cc1 JA |
3976 | mmgrab(current->mm); |
3977 | ctx->sqo_mm = current->mm; | |
3978 | ||
6c271ce2 | 3979 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
3ec482d1 JA |
3980 | ret = -EPERM; |
3981 | if (!capable(CAP_SYS_ADMIN)) | |
3982 | goto err; | |
3983 | ||
917257da JA |
3984 | ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle); |
3985 | if (!ctx->sq_thread_idle) | |
3986 | ctx->sq_thread_idle = HZ; | |
3987 | ||
6c271ce2 | 3988 | if (p->flags & IORING_SETUP_SQ_AFF) { |
44a9bd18 | 3989 | int cpu = p->sq_thread_cpu; |
6c271ce2 | 3990 | |
917257da | 3991 | ret = -EINVAL; |
44a9bd18 JA |
3992 | if (cpu >= nr_cpu_ids) |
3993 | goto err; | |
7889f44d | 3994 | if (!cpu_online(cpu)) |
917257da JA |
3995 | goto err; |
3996 | ||
6c271ce2 JA |
3997 | ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread, |
3998 | ctx, cpu, | |
3999 | "io_uring-sq"); | |
4000 | } else { | |
4001 | ctx->sqo_thread = kthread_create(io_sq_thread, ctx, | |
4002 | "io_uring-sq"); | |
4003 | } | |
4004 | if (IS_ERR(ctx->sqo_thread)) { | |
4005 | ret = PTR_ERR(ctx->sqo_thread); | |
4006 | ctx->sqo_thread = NULL; | |
4007 | goto err; | |
4008 | } | |
4009 | wake_up_process(ctx->sqo_thread); | |
4010 | } else if (p->flags & IORING_SETUP_SQ_AFF) { | |
4011 | /* Can't have SQ_AFF without SQPOLL */ | |
4012 | ret = -EINVAL; | |
4013 | goto err; | |
4014 | } | |
4015 | ||
576a347b JA |
4016 | data.mm = ctx->sqo_mm; |
4017 | data.user = ctx->user; | |
181e448d | 4018 | data.creds = ctx->creds; |
576a347b JA |
4019 | data.get_work = io_get_work; |
4020 | data.put_work = io_put_work; | |
4021 | ||
561fb04a JA |
4022 | /* Do QD, or 4 * CPUS, whatever is smallest */ |
4023 | concurrency = min(ctx->sq_entries, 4 * num_online_cpus()); | |
576a347b | 4024 | ctx->io_wq = io_wq_create(concurrency, &data); |
975c99a5 JA |
4025 | if (IS_ERR(ctx->io_wq)) { |
4026 | ret = PTR_ERR(ctx->io_wq); | |
4027 | ctx->io_wq = NULL; | |
2b188cc1 JA |
4028 | goto err; |
4029 | } | |
4030 | ||
4031 | return 0; | |
4032 | err: | |
54a91f3b | 4033 | io_finish_async(ctx); |
2b188cc1 JA |
4034 | mmdrop(ctx->sqo_mm); |
4035 | ctx->sqo_mm = NULL; | |
4036 | return ret; | |
4037 | } | |
4038 | ||
4039 | static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages) | |
4040 | { | |
4041 | atomic_long_sub(nr_pages, &user->locked_vm); | |
4042 | } | |
4043 | ||
4044 | static int io_account_mem(struct user_struct *user, unsigned long nr_pages) | |
4045 | { | |
4046 | unsigned long page_limit, cur_pages, new_pages; | |
4047 | ||
4048 | /* Don't allow more pages than we can safely lock */ | |
4049 | page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
4050 | ||
4051 | do { | |
4052 | cur_pages = atomic_long_read(&user->locked_vm); | |
4053 | new_pages = cur_pages + nr_pages; | |
4054 | if (new_pages > page_limit) | |
4055 | return -ENOMEM; | |
4056 | } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages, | |
4057 | new_pages) != cur_pages); | |
4058 | ||
4059 | return 0; | |
4060 | } | |
4061 | ||
4062 | static void io_mem_free(void *ptr) | |
4063 | { | |
52e04ef4 MR |
4064 | struct page *page; |
4065 | ||
4066 | if (!ptr) | |
4067 | return; | |
2b188cc1 | 4068 | |
52e04ef4 | 4069 | page = virt_to_head_page(ptr); |
2b188cc1 JA |
4070 | if (put_page_testzero(page)) |
4071 | free_compound_page(page); | |
4072 | } | |
4073 | ||
4074 | static void *io_mem_alloc(size_t size) | |
4075 | { | |
4076 | gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP | | |
4077 | __GFP_NORETRY; | |
4078 | ||
4079 | return (void *) __get_free_pages(gfp_flags, get_order(size)); | |
4080 | } | |
4081 | ||
75b28aff HV |
4082 | static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries, |
4083 | size_t *sq_offset) | |
4084 | { | |
4085 | struct io_rings *rings; | |
4086 | size_t off, sq_array_size; | |
4087 | ||
4088 | off = struct_size(rings, cqes, cq_entries); | |
4089 | if (off == SIZE_MAX) | |
4090 | return SIZE_MAX; | |
4091 | ||
4092 | #ifdef CONFIG_SMP | |
4093 | off = ALIGN(off, SMP_CACHE_BYTES); | |
4094 | if (off == 0) | |
4095 | return SIZE_MAX; | |
4096 | #endif | |
4097 | ||
4098 | sq_array_size = array_size(sizeof(u32), sq_entries); | |
4099 | if (sq_array_size == SIZE_MAX) | |
4100 | return SIZE_MAX; | |
4101 | ||
4102 | if (check_add_overflow(off, sq_array_size, &off)) | |
4103 | return SIZE_MAX; | |
4104 | ||
4105 | if (sq_offset) | |
4106 | *sq_offset = off; | |
4107 | ||
4108 | return off; | |
4109 | } | |
4110 | ||
2b188cc1 JA |
4111 | static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries) |
4112 | { | |
75b28aff | 4113 | size_t pages; |
2b188cc1 | 4114 | |
75b28aff HV |
4115 | pages = (size_t)1 << get_order( |
4116 | rings_size(sq_entries, cq_entries, NULL)); | |
4117 | pages += (size_t)1 << get_order( | |
4118 | array_size(sizeof(struct io_uring_sqe), sq_entries)); | |
2b188cc1 | 4119 | |
75b28aff | 4120 | return pages; |
2b188cc1 JA |
4121 | } |
4122 | ||
edafccee JA |
4123 | static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx) |
4124 | { | |
4125 | int i, j; | |
4126 | ||
4127 | if (!ctx->user_bufs) | |
4128 | return -ENXIO; | |
4129 | ||
4130 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
4131 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
4132 | ||
4133 | for (j = 0; j < imu->nr_bvecs; j++) | |
27c4d3a3 | 4134 | put_user_page(imu->bvec[j].bv_page); |
edafccee JA |
4135 | |
4136 | if (ctx->account_mem) | |
4137 | io_unaccount_mem(ctx->user, imu->nr_bvecs); | |
d4ef6475 | 4138 | kvfree(imu->bvec); |
edafccee JA |
4139 | imu->nr_bvecs = 0; |
4140 | } | |
4141 | ||
4142 | kfree(ctx->user_bufs); | |
4143 | ctx->user_bufs = NULL; | |
4144 | ctx->nr_user_bufs = 0; | |
4145 | return 0; | |
4146 | } | |
4147 | ||
4148 | static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, | |
4149 | void __user *arg, unsigned index) | |
4150 | { | |
4151 | struct iovec __user *src; | |
4152 | ||
4153 | #ifdef CONFIG_COMPAT | |
4154 | if (ctx->compat) { | |
4155 | struct compat_iovec __user *ciovs; | |
4156 | struct compat_iovec ciov; | |
4157 | ||
4158 | ciovs = (struct compat_iovec __user *) arg; | |
4159 | if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) | |
4160 | return -EFAULT; | |
4161 | ||
4162 | dst->iov_base = (void __user *) (unsigned long) ciov.iov_base; | |
4163 | dst->iov_len = ciov.iov_len; | |
4164 | return 0; | |
4165 | } | |
4166 | #endif | |
4167 | src = (struct iovec __user *) arg; | |
4168 | if (copy_from_user(dst, &src[index], sizeof(*dst))) | |
4169 | return -EFAULT; | |
4170 | return 0; | |
4171 | } | |
4172 | ||
4173 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg, | |
4174 | unsigned nr_args) | |
4175 | { | |
4176 | struct vm_area_struct **vmas = NULL; | |
4177 | struct page **pages = NULL; | |
4178 | int i, j, got_pages = 0; | |
4179 | int ret = -EINVAL; | |
4180 | ||
4181 | if (ctx->user_bufs) | |
4182 | return -EBUSY; | |
4183 | if (!nr_args || nr_args > UIO_MAXIOV) | |
4184 | return -EINVAL; | |
4185 | ||
4186 | ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf), | |
4187 | GFP_KERNEL); | |
4188 | if (!ctx->user_bufs) | |
4189 | return -ENOMEM; | |
4190 | ||
4191 | for (i = 0; i < nr_args; i++) { | |
4192 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
4193 | unsigned long off, start, end, ubuf; | |
4194 | int pret, nr_pages; | |
4195 | struct iovec iov; | |
4196 | size_t size; | |
4197 | ||
4198 | ret = io_copy_iov(ctx, &iov, arg, i); | |
4199 | if (ret) | |
a278682d | 4200 | goto err; |
edafccee JA |
4201 | |
4202 | /* | |
4203 | * Don't impose further limits on the size and buffer | |
4204 | * constraints here, we'll -EINVAL later when IO is | |
4205 | * submitted if they are wrong. | |
4206 | */ | |
4207 | ret = -EFAULT; | |
4208 | if (!iov.iov_base || !iov.iov_len) | |
4209 | goto err; | |
4210 | ||
4211 | /* arbitrary limit, but we need something */ | |
4212 | if (iov.iov_len > SZ_1G) | |
4213 | goto err; | |
4214 | ||
4215 | ubuf = (unsigned long) iov.iov_base; | |
4216 | end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
4217 | start = ubuf >> PAGE_SHIFT; | |
4218 | nr_pages = end - start; | |
4219 | ||
4220 | if (ctx->account_mem) { | |
4221 | ret = io_account_mem(ctx->user, nr_pages); | |
4222 | if (ret) | |
4223 | goto err; | |
4224 | } | |
4225 | ||
4226 | ret = 0; | |
4227 | if (!pages || nr_pages > got_pages) { | |
4228 | kfree(vmas); | |
4229 | kfree(pages); | |
d4ef6475 | 4230 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), |
edafccee | 4231 | GFP_KERNEL); |
d4ef6475 | 4232 | vmas = kvmalloc_array(nr_pages, |
edafccee JA |
4233 | sizeof(struct vm_area_struct *), |
4234 | GFP_KERNEL); | |
4235 | if (!pages || !vmas) { | |
4236 | ret = -ENOMEM; | |
4237 | if (ctx->account_mem) | |
4238 | io_unaccount_mem(ctx->user, nr_pages); | |
4239 | goto err; | |
4240 | } | |
4241 | got_pages = nr_pages; | |
4242 | } | |
4243 | ||
d4ef6475 | 4244 | imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec), |
edafccee JA |
4245 | GFP_KERNEL); |
4246 | ret = -ENOMEM; | |
4247 | if (!imu->bvec) { | |
4248 | if (ctx->account_mem) | |
4249 | io_unaccount_mem(ctx->user, nr_pages); | |
4250 | goto err; | |
4251 | } | |
4252 | ||
4253 | ret = 0; | |
4254 | down_read(¤t->mm->mmap_sem); | |
932f4a63 IW |
4255 | pret = get_user_pages(ubuf, nr_pages, |
4256 | FOLL_WRITE | FOLL_LONGTERM, | |
4257 | pages, vmas); | |
edafccee JA |
4258 | if (pret == nr_pages) { |
4259 | /* don't support file backed memory */ | |
4260 | for (j = 0; j < nr_pages; j++) { | |
4261 | struct vm_area_struct *vma = vmas[j]; | |
4262 | ||
4263 | if (vma->vm_file && | |
4264 | !is_file_hugepages(vma->vm_file)) { | |
4265 | ret = -EOPNOTSUPP; | |
4266 | break; | |
4267 | } | |
4268 | } | |
4269 | } else { | |
4270 | ret = pret < 0 ? pret : -EFAULT; | |
4271 | } | |
4272 | up_read(¤t->mm->mmap_sem); | |
4273 | if (ret) { | |
4274 | /* | |
4275 | * if we did partial map, or found file backed vmas, | |
4276 | * release any pages we did get | |
4277 | */ | |
27c4d3a3 JH |
4278 | if (pret > 0) |
4279 | put_user_pages(pages, pret); | |
edafccee JA |
4280 | if (ctx->account_mem) |
4281 | io_unaccount_mem(ctx->user, nr_pages); | |
d4ef6475 | 4282 | kvfree(imu->bvec); |
edafccee JA |
4283 | goto err; |
4284 | } | |
4285 | ||
4286 | off = ubuf & ~PAGE_MASK; | |
4287 | size = iov.iov_len; | |
4288 | for (j = 0; j < nr_pages; j++) { | |
4289 | size_t vec_len; | |
4290 | ||
4291 | vec_len = min_t(size_t, size, PAGE_SIZE - off); | |
4292 | imu->bvec[j].bv_page = pages[j]; | |
4293 | imu->bvec[j].bv_len = vec_len; | |
4294 | imu->bvec[j].bv_offset = off; | |
4295 | off = 0; | |
4296 | size -= vec_len; | |
4297 | } | |
4298 | /* store original address for later verification */ | |
4299 | imu->ubuf = ubuf; | |
4300 | imu->len = iov.iov_len; | |
4301 | imu->nr_bvecs = nr_pages; | |
4302 | ||
4303 | ctx->nr_user_bufs++; | |
4304 | } | |
d4ef6475 MR |
4305 | kvfree(pages); |
4306 | kvfree(vmas); | |
edafccee JA |
4307 | return 0; |
4308 | err: | |
d4ef6475 MR |
4309 | kvfree(pages); |
4310 | kvfree(vmas); | |
edafccee JA |
4311 | io_sqe_buffer_unregister(ctx); |
4312 | return ret; | |
4313 | } | |
4314 | ||
9b402849 JA |
4315 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg) |
4316 | { | |
4317 | __s32 __user *fds = arg; | |
4318 | int fd; | |
4319 | ||
4320 | if (ctx->cq_ev_fd) | |
4321 | return -EBUSY; | |
4322 | ||
4323 | if (copy_from_user(&fd, fds, sizeof(*fds))) | |
4324 | return -EFAULT; | |
4325 | ||
4326 | ctx->cq_ev_fd = eventfd_ctx_fdget(fd); | |
4327 | if (IS_ERR(ctx->cq_ev_fd)) { | |
4328 | int ret = PTR_ERR(ctx->cq_ev_fd); | |
4329 | ctx->cq_ev_fd = NULL; | |
4330 | return ret; | |
4331 | } | |
4332 | ||
4333 | return 0; | |
4334 | } | |
4335 | ||
4336 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) | |
4337 | { | |
4338 | if (ctx->cq_ev_fd) { | |
4339 | eventfd_ctx_put(ctx->cq_ev_fd); | |
4340 | ctx->cq_ev_fd = NULL; | |
4341 | return 0; | |
4342 | } | |
4343 | ||
4344 | return -ENXIO; | |
4345 | } | |
4346 | ||
2b188cc1 JA |
4347 | static void io_ring_ctx_free(struct io_ring_ctx *ctx) |
4348 | { | |
6b06314c | 4349 | io_finish_async(ctx); |
2b188cc1 JA |
4350 | if (ctx->sqo_mm) |
4351 | mmdrop(ctx->sqo_mm); | |
def596e9 JA |
4352 | |
4353 | io_iopoll_reap_events(ctx); | |
edafccee | 4354 | io_sqe_buffer_unregister(ctx); |
6b06314c | 4355 | io_sqe_files_unregister(ctx); |
9b402849 | 4356 | io_eventfd_unregister(ctx); |
def596e9 | 4357 | |
2b188cc1 | 4358 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
4359 | if (ctx->ring_sock) { |
4360 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 4361 | sock_release(ctx->ring_sock); |
355e8d26 | 4362 | } |
2b188cc1 JA |
4363 | #endif |
4364 | ||
75b28aff | 4365 | io_mem_free(ctx->rings); |
2b188cc1 | 4366 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
4367 | |
4368 | percpu_ref_exit(&ctx->refs); | |
4369 | if (ctx->account_mem) | |
4370 | io_unaccount_mem(ctx->user, | |
4371 | ring_pages(ctx->sq_entries, ctx->cq_entries)); | |
4372 | free_uid(ctx->user); | |
181e448d | 4373 | put_cred(ctx->creds); |
206aefde | 4374 | kfree(ctx->completions); |
0ddf92e8 | 4375 | kmem_cache_free(req_cachep, ctx->fallback_req); |
2b188cc1 JA |
4376 | kfree(ctx); |
4377 | } | |
4378 | ||
4379 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
4380 | { | |
4381 | struct io_ring_ctx *ctx = file->private_data; | |
4382 | __poll_t mask = 0; | |
4383 | ||
4384 | poll_wait(file, &ctx->cq_wait, wait); | |
4f7067c3 SB |
4385 | /* |
4386 | * synchronizes with barrier from wq_has_sleeper call in | |
4387 | * io_commit_cqring | |
4388 | */ | |
2b188cc1 | 4389 | smp_rmb(); |
75b28aff HV |
4390 | if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head != |
4391 | ctx->rings->sq_ring_entries) | |
2b188cc1 | 4392 | mask |= EPOLLOUT | EPOLLWRNORM; |
daa5de54 | 4393 | if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail) |
2b188cc1 JA |
4394 | mask |= EPOLLIN | EPOLLRDNORM; |
4395 | ||
4396 | return mask; | |
4397 | } | |
4398 | ||
4399 | static int io_uring_fasync(int fd, struct file *file, int on) | |
4400 | { | |
4401 | struct io_ring_ctx *ctx = file->private_data; | |
4402 | ||
4403 | return fasync_helper(fd, file, on, &ctx->cq_fasync); | |
4404 | } | |
4405 | ||
4406 | static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) | |
4407 | { | |
4408 | mutex_lock(&ctx->uring_lock); | |
4409 | percpu_ref_kill(&ctx->refs); | |
4410 | mutex_unlock(&ctx->uring_lock); | |
4411 | ||
5262f567 | 4412 | io_kill_timeouts(ctx); |
221c5eb2 | 4413 | io_poll_remove_all(ctx); |
561fb04a JA |
4414 | |
4415 | if (ctx->io_wq) | |
4416 | io_wq_cancel_all(ctx->io_wq); | |
4417 | ||
def596e9 | 4418 | io_iopoll_reap_events(ctx); |
15dff286 JA |
4419 | /* if we failed setting up the ctx, we might not have any rings */ |
4420 | if (ctx->rings) | |
4421 | io_cqring_overflow_flush(ctx, true); | |
206aefde | 4422 | wait_for_completion(&ctx->completions[0]); |
2b188cc1 JA |
4423 | io_ring_ctx_free(ctx); |
4424 | } | |
4425 | ||
4426 | static int io_uring_release(struct inode *inode, struct file *file) | |
4427 | { | |
4428 | struct io_ring_ctx *ctx = file->private_data; | |
4429 | ||
4430 | file->private_data = NULL; | |
4431 | io_ring_ctx_wait_and_kill(ctx); | |
4432 | return 0; | |
4433 | } | |
4434 | ||
fcb323cc JA |
4435 | static void io_uring_cancel_files(struct io_ring_ctx *ctx, |
4436 | struct files_struct *files) | |
4437 | { | |
4438 | struct io_kiocb *req; | |
4439 | DEFINE_WAIT(wait); | |
4440 | ||
4441 | while (!list_empty_careful(&ctx->inflight_list)) { | |
768134d4 | 4442 | struct io_kiocb *cancel_req = NULL; |
fcb323cc JA |
4443 | |
4444 | spin_lock_irq(&ctx->inflight_lock); | |
4445 | list_for_each_entry(req, &ctx->inflight_list, inflight_entry) { | |
768134d4 JA |
4446 | if (req->work.files != files) |
4447 | continue; | |
4448 | /* req is being completed, ignore */ | |
4449 | if (!refcount_inc_not_zero(&req->refs)) | |
4450 | continue; | |
4451 | cancel_req = req; | |
4452 | break; | |
fcb323cc | 4453 | } |
768134d4 | 4454 | if (cancel_req) |
fcb323cc | 4455 | prepare_to_wait(&ctx->inflight_wait, &wait, |
768134d4 | 4456 | TASK_UNINTERRUPTIBLE); |
fcb323cc JA |
4457 | spin_unlock_irq(&ctx->inflight_lock); |
4458 | ||
768134d4 JA |
4459 | /* We need to keep going until we don't find a matching req */ |
4460 | if (!cancel_req) | |
fcb323cc | 4461 | break; |
2f6d9b9d BL |
4462 | |
4463 | io_wq_cancel_work(ctx->io_wq, &cancel_req->work); | |
4464 | io_put_req(cancel_req); | |
fcb323cc JA |
4465 | schedule(); |
4466 | } | |
768134d4 | 4467 | finish_wait(&ctx->inflight_wait, &wait); |
fcb323cc JA |
4468 | } |
4469 | ||
4470 | static int io_uring_flush(struct file *file, void *data) | |
4471 | { | |
4472 | struct io_ring_ctx *ctx = file->private_data; | |
4473 | ||
4474 | io_uring_cancel_files(ctx, data); | |
1d7bb1d5 JA |
4475 | if (fatal_signal_pending(current) || (current->flags & PF_EXITING)) { |
4476 | io_cqring_overflow_flush(ctx, true); | |
fcb323cc | 4477 | io_wq_cancel_all(ctx->io_wq); |
1d7bb1d5 | 4478 | } |
fcb323cc JA |
4479 | return 0; |
4480 | } | |
4481 | ||
2b188cc1 JA |
4482 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) |
4483 | { | |
4484 | loff_t offset = (loff_t) vma->vm_pgoff << PAGE_SHIFT; | |
4485 | unsigned long sz = vma->vm_end - vma->vm_start; | |
4486 | struct io_ring_ctx *ctx = file->private_data; | |
4487 | unsigned long pfn; | |
4488 | struct page *page; | |
4489 | void *ptr; | |
4490 | ||
4491 | switch (offset) { | |
4492 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
4493 | case IORING_OFF_CQ_RING: |
4494 | ptr = ctx->rings; | |
2b188cc1 JA |
4495 | break; |
4496 | case IORING_OFF_SQES: | |
4497 | ptr = ctx->sq_sqes; | |
4498 | break; | |
2b188cc1 JA |
4499 | default: |
4500 | return -EINVAL; | |
4501 | } | |
4502 | ||
4503 | page = virt_to_head_page(ptr); | |
a50b854e | 4504 | if (sz > page_size(page)) |
2b188cc1 JA |
4505 | return -EINVAL; |
4506 | ||
4507 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
4508 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
4509 | } | |
4510 | ||
4511 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, | |
4512 | u32, min_complete, u32, flags, const sigset_t __user *, sig, | |
4513 | size_t, sigsz) | |
4514 | { | |
4515 | struct io_ring_ctx *ctx; | |
4516 | long ret = -EBADF; | |
4517 | int submitted = 0; | |
4518 | struct fd f; | |
4519 | ||
6c271ce2 | 4520 | if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP)) |
2b188cc1 JA |
4521 | return -EINVAL; |
4522 | ||
4523 | f = fdget(fd); | |
4524 | if (!f.file) | |
4525 | return -EBADF; | |
4526 | ||
4527 | ret = -EOPNOTSUPP; | |
4528 | if (f.file->f_op != &io_uring_fops) | |
4529 | goto out_fput; | |
4530 | ||
4531 | ret = -ENXIO; | |
4532 | ctx = f.file->private_data; | |
4533 | if (!percpu_ref_tryget(&ctx->refs)) | |
4534 | goto out_fput; | |
4535 | ||
6c271ce2 JA |
4536 | /* |
4537 | * For SQ polling, the thread will do all submissions and completions. | |
4538 | * Just return the requested submit count, and wake the thread if | |
4539 | * we were asked to. | |
4540 | */ | |
b2a9eada | 4541 | ret = 0; |
6c271ce2 | 4542 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
c1edbf5f JA |
4543 | if (!list_empty_careful(&ctx->cq_overflow_list)) |
4544 | io_cqring_overflow_flush(ctx, false); | |
6c271ce2 JA |
4545 | if (flags & IORING_ENTER_SQ_WAKEUP) |
4546 | wake_up(&ctx->sqo_wait); | |
4547 | submitted = to_submit; | |
b2a9eada | 4548 | } else if (to_submit) { |
ae9428ca | 4549 | struct mm_struct *cur_mm; |
2b188cc1 | 4550 | |
ae9428ca | 4551 | to_submit = min(to_submit, ctx->sq_entries); |
2b188cc1 | 4552 | mutex_lock(&ctx->uring_lock); |
ae9428ca PB |
4553 | /* already have mm, so io_submit_sqes() won't try to grab it */ |
4554 | cur_mm = ctx->sqo_mm; | |
4555 | submitted = io_submit_sqes(ctx, to_submit, f.file, fd, | |
4556 | &cur_mm, false); | |
2b188cc1 | 4557 | mutex_unlock(&ctx->uring_lock); |
2b188cc1 JA |
4558 | } |
4559 | if (flags & IORING_ENTER_GETEVENTS) { | |
def596e9 JA |
4560 | unsigned nr_events = 0; |
4561 | ||
2b188cc1 JA |
4562 | min_complete = min(min_complete, ctx->cq_entries); |
4563 | ||
def596e9 | 4564 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 | 4565 | ret = io_iopoll_check(ctx, &nr_events, min_complete); |
def596e9 JA |
4566 | } else { |
4567 | ret = io_cqring_wait(ctx, min_complete, sig, sigsz); | |
4568 | } | |
2b188cc1 JA |
4569 | } |
4570 | ||
6805b32e | 4571 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
4572 | out_fput: |
4573 | fdput(f); | |
4574 | return submitted ? submitted : ret; | |
4575 | } | |
4576 | ||
4577 | static const struct file_operations io_uring_fops = { | |
4578 | .release = io_uring_release, | |
fcb323cc | 4579 | .flush = io_uring_flush, |
2b188cc1 JA |
4580 | .mmap = io_uring_mmap, |
4581 | .poll = io_uring_poll, | |
4582 | .fasync = io_uring_fasync, | |
4583 | }; | |
4584 | ||
4585 | static int io_allocate_scq_urings(struct io_ring_ctx *ctx, | |
4586 | struct io_uring_params *p) | |
4587 | { | |
75b28aff HV |
4588 | struct io_rings *rings; |
4589 | size_t size, sq_array_offset; | |
2b188cc1 | 4590 | |
75b28aff HV |
4591 | size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset); |
4592 | if (size == SIZE_MAX) | |
4593 | return -EOVERFLOW; | |
4594 | ||
4595 | rings = io_mem_alloc(size); | |
4596 | if (!rings) | |
2b188cc1 JA |
4597 | return -ENOMEM; |
4598 | ||
75b28aff HV |
4599 | ctx->rings = rings; |
4600 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
4601 | rings->sq_ring_mask = p->sq_entries - 1; | |
4602 | rings->cq_ring_mask = p->cq_entries - 1; | |
4603 | rings->sq_ring_entries = p->sq_entries; | |
4604 | rings->cq_ring_entries = p->cq_entries; | |
4605 | ctx->sq_mask = rings->sq_ring_mask; | |
4606 | ctx->cq_mask = rings->cq_ring_mask; | |
4607 | ctx->sq_entries = rings->sq_ring_entries; | |
4608 | ctx->cq_entries = rings->cq_ring_entries; | |
2b188cc1 JA |
4609 | |
4610 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 JA |
4611 | if (size == SIZE_MAX) { |
4612 | io_mem_free(ctx->rings); | |
4613 | ctx->rings = NULL; | |
2b188cc1 | 4614 | return -EOVERFLOW; |
eb065d30 | 4615 | } |
2b188cc1 JA |
4616 | |
4617 | ctx->sq_sqes = io_mem_alloc(size); | |
eb065d30 JA |
4618 | if (!ctx->sq_sqes) { |
4619 | io_mem_free(ctx->rings); | |
4620 | ctx->rings = NULL; | |
2b188cc1 | 4621 | return -ENOMEM; |
eb065d30 | 4622 | } |
2b188cc1 | 4623 | |
2b188cc1 JA |
4624 | return 0; |
4625 | } | |
4626 | ||
4627 | /* | |
4628 | * Allocate an anonymous fd, this is what constitutes the application | |
4629 | * visible backing of an io_uring instance. The application mmaps this | |
4630 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
4631 | * we have to tie this fd to a socket for file garbage collection purposes. | |
4632 | */ | |
4633 | static int io_uring_get_fd(struct io_ring_ctx *ctx) | |
4634 | { | |
4635 | struct file *file; | |
4636 | int ret; | |
4637 | ||
4638 | #if defined(CONFIG_UNIX) | |
4639 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, | |
4640 | &ctx->ring_sock); | |
4641 | if (ret) | |
4642 | return ret; | |
4643 | #endif | |
4644 | ||
4645 | ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
4646 | if (ret < 0) | |
4647 | goto err; | |
4648 | ||
4649 | file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx, | |
4650 | O_RDWR | O_CLOEXEC); | |
4651 | if (IS_ERR(file)) { | |
4652 | put_unused_fd(ret); | |
4653 | ret = PTR_ERR(file); | |
4654 | goto err; | |
4655 | } | |
4656 | ||
4657 | #if defined(CONFIG_UNIX) | |
4658 | ctx->ring_sock->file = file; | |
6b06314c | 4659 | ctx->ring_sock->sk->sk_user_data = ctx; |
2b188cc1 JA |
4660 | #endif |
4661 | fd_install(ret, file); | |
4662 | return ret; | |
4663 | err: | |
4664 | #if defined(CONFIG_UNIX) | |
4665 | sock_release(ctx->ring_sock); | |
4666 | ctx->ring_sock = NULL; | |
4667 | #endif | |
4668 | return ret; | |
4669 | } | |
4670 | ||
4671 | static int io_uring_create(unsigned entries, struct io_uring_params *p) | |
4672 | { | |
4673 | struct user_struct *user = NULL; | |
4674 | struct io_ring_ctx *ctx; | |
4675 | bool account_mem; | |
4676 | int ret; | |
4677 | ||
4678 | if (!entries || entries > IORING_MAX_ENTRIES) | |
4679 | return -EINVAL; | |
4680 | ||
4681 | /* | |
4682 | * Use twice as many entries for the CQ ring. It's possible for the | |
4683 | * application to drive a higher depth than the size of the SQ ring, | |
4684 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
4685 | * some flexibility in overcommitting a bit. If the application has |
4686 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
4687 | * of CQ ring entries manually. | |
2b188cc1 JA |
4688 | */ |
4689 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
4690 | if (p->flags & IORING_SETUP_CQSIZE) { |
4691 | /* | |
4692 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
4693 | * to a power-of-two, if it isn't already. We do NOT impose | |
4694 | * any cq vs sq ring sizing. | |
4695 | */ | |
4696 | if (p->cq_entries < p->sq_entries || p->cq_entries > IORING_MAX_CQ_ENTRIES) | |
4697 | return -EINVAL; | |
4698 | p->cq_entries = roundup_pow_of_two(p->cq_entries); | |
4699 | } else { | |
4700 | p->cq_entries = 2 * p->sq_entries; | |
4701 | } | |
2b188cc1 JA |
4702 | |
4703 | user = get_uid(current_user()); | |
4704 | account_mem = !capable(CAP_IPC_LOCK); | |
4705 | ||
4706 | if (account_mem) { | |
4707 | ret = io_account_mem(user, | |
4708 | ring_pages(p->sq_entries, p->cq_entries)); | |
4709 | if (ret) { | |
4710 | free_uid(user); | |
4711 | return ret; | |
4712 | } | |
4713 | } | |
4714 | ||
4715 | ctx = io_ring_ctx_alloc(p); | |
4716 | if (!ctx) { | |
4717 | if (account_mem) | |
4718 | io_unaccount_mem(user, ring_pages(p->sq_entries, | |
4719 | p->cq_entries)); | |
4720 | free_uid(user); | |
4721 | return -ENOMEM; | |
4722 | } | |
4723 | ctx->compat = in_compat_syscall(); | |
4724 | ctx->account_mem = account_mem; | |
4725 | ctx->user = user; | |
181e448d | 4726 | ctx->creds = prepare_creds(); |
2b188cc1 JA |
4727 | |
4728 | ret = io_allocate_scq_urings(ctx, p); | |
4729 | if (ret) | |
4730 | goto err; | |
4731 | ||
6c271ce2 | 4732 | ret = io_sq_offload_start(ctx, p); |
2b188cc1 JA |
4733 | if (ret) |
4734 | goto err; | |
4735 | ||
2b188cc1 | 4736 | memset(&p->sq_off, 0, sizeof(p->sq_off)); |
75b28aff HV |
4737 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
4738 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
4739 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
4740 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
4741 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
4742 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
4743 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
4744 | |
4745 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
4746 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
4747 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
4748 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
4749 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
4750 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
4751 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
ac90f249 | 4752 | |
044c1ab3 JA |
4753 | /* |
4754 | * Install ring fd as the very last thing, so we don't risk someone | |
4755 | * having closed it before we finish setup | |
4756 | */ | |
4757 | ret = io_uring_get_fd(ctx); | |
4758 | if (ret < 0) | |
4759 | goto err; | |
4760 | ||
1d7bb1d5 | 4761 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP; |
c826bd7a | 4762 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
4763 | return ret; |
4764 | err: | |
4765 | io_ring_ctx_wait_and_kill(ctx); | |
4766 | return ret; | |
4767 | } | |
4768 | ||
4769 | /* | |
4770 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
4771 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
4772 | * params structure passed in. | |
4773 | */ | |
4774 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
4775 | { | |
4776 | struct io_uring_params p; | |
4777 | long ret; | |
4778 | int i; | |
4779 | ||
4780 | if (copy_from_user(&p, params, sizeof(p))) | |
4781 | return -EFAULT; | |
4782 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
4783 | if (p.resv[i]) | |
4784 | return -EINVAL; | |
4785 | } | |
4786 | ||
6c271ce2 | 4787 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
33a107f0 | 4788 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE)) |
2b188cc1 JA |
4789 | return -EINVAL; |
4790 | ||
4791 | ret = io_uring_create(entries, &p); | |
4792 | if (ret < 0) | |
4793 | return ret; | |
4794 | ||
4795 | if (copy_to_user(params, &p, sizeof(p))) | |
4796 | return -EFAULT; | |
4797 | ||
4798 | return ret; | |
4799 | } | |
4800 | ||
4801 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
4802 | struct io_uring_params __user *, params) | |
4803 | { | |
4804 | return io_uring_setup(entries, params); | |
4805 | } | |
4806 | ||
edafccee JA |
4807 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
4808 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
4809 | __releases(ctx->uring_lock) |
4810 | __acquires(ctx->uring_lock) | |
edafccee JA |
4811 | { |
4812 | int ret; | |
4813 | ||
35fa71a0 JA |
4814 | /* |
4815 | * We're inside the ring mutex, if the ref is already dying, then | |
4816 | * someone else killed the ctx or is already going through | |
4817 | * io_uring_register(). | |
4818 | */ | |
4819 | if (percpu_ref_is_dying(&ctx->refs)) | |
4820 | return -ENXIO; | |
4821 | ||
edafccee | 4822 | percpu_ref_kill(&ctx->refs); |
b19062a5 JA |
4823 | |
4824 | /* | |
4825 | * Drop uring mutex before waiting for references to exit. If another | |
4826 | * thread is currently inside io_uring_enter() it might need to grab | |
4827 | * the uring_lock to make progress. If we hold it here across the drain | |
4828 | * wait, then we can deadlock. It's safe to drop the mutex here, since | |
4829 | * no new references will come in after we've killed the percpu ref. | |
4830 | */ | |
4831 | mutex_unlock(&ctx->uring_lock); | |
206aefde | 4832 | wait_for_completion(&ctx->completions[0]); |
b19062a5 | 4833 | mutex_lock(&ctx->uring_lock); |
edafccee JA |
4834 | |
4835 | switch (opcode) { | |
4836 | case IORING_REGISTER_BUFFERS: | |
4837 | ret = io_sqe_buffer_register(ctx, arg, nr_args); | |
4838 | break; | |
4839 | case IORING_UNREGISTER_BUFFERS: | |
4840 | ret = -EINVAL; | |
4841 | if (arg || nr_args) | |
4842 | break; | |
4843 | ret = io_sqe_buffer_unregister(ctx); | |
4844 | break; | |
6b06314c JA |
4845 | case IORING_REGISTER_FILES: |
4846 | ret = io_sqe_files_register(ctx, arg, nr_args); | |
4847 | break; | |
4848 | case IORING_UNREGISTER_FILES: | |
4849 | ret = -EINVAL; | |
4850 | if (arg || nr_args) | |
4851 | break; | |
4852 | ret = io_sqe_files_unregister(ctx); | |
4853 | break; | |
c3a31e60 JA |
4854 | case IORING_REGISTER_FILES_UPDATE: |
4855 | ret = io_sqe_files_update(ctx, arg, nr_args); | |
4856 | break; | |
9b402849 JA |
4857 | case IORING_REGISTER_EVENTFD: |
4858 | ret = -EINVAL; | |
4859 | if (nr_args != 1) | |
4860 | break; | |
4861 | ret = io_eventfd_register(ctx, arg); | |
4862 | break; | |
4863 | case IORING_UNREGISTER_EVENTFD: | |
4864 | ret = -EINVAL; | |
4865 | if (arg || nr_args) | |
4866 | break; | |
4867 | ret = io_eventfd_unregister(ctx); | |
4868 | break; | |
edafccee JA |
4869 | default: |
4870 | ret = -EINVAL; | |
4871 | break; | |
4872 | } | |
4873 | ||
4874 | /* bring the ctx back to life */ | |
206aefde | 4875 | reinit_completion(&ctx->completions[0]); |
edafccee JA |
4876 | percpu_ref_reinit(&ctx->refs); |
4877 | return ret; | |
4878 | } | |
4879 | ||
4880 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
4881 | void __user *, arg, unsigned int, nr_args) | |
4882 | { | |
4883 | struct io_ring_ctx *ctx; | |
4884 | long ret = -EBADF; | |
4885 | struct fd f; | |
4886 | ||
4887 | f = fdget(fd); | |
4888 | if (!f.file) | |
4889 | return -EBADF; | |
4890 | ||
4891 | ret = -EOPNOTSUPP; | |
4892 | if (f.file->f_op != &io_uring_fops) | |
4893 | goto out_fput; | |
4894 | ||
4895 | ctx = f.file->private_data; | |
4896 | ||
4897 | mutex_lock(&ctx->uring_lock); | |
4898 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
4899 | mutex_unlock(&ctx->uring_lock); | |
c826bd7a DD |
4900 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, |
4901 | ctx->cq_ev_fd != NULL, ret); | |
edafccee JA |
4902 | out_fput: |
4903 | fdput(f); | |
4904 | return ret; | |
4905 | } | |
4906 | ||
2b188cc1 JA |
4907 | static int __init io_uring_init(void) |
4908 | { | |
4909 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC); | |
4910 | return 0; | |
4911 | }; | |
4912 | __initcall(io_uring_init); |