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> | |
59 | #include <linux/workqueue.h> | |
6c271ce2 | 60 | #include <linux/kthread.h> |
2b188cc1 | 61 | #include <linux/blkdev.h> |
edafccee | 62 | #include <linux/bvec.h> |
2b188cc1 JA |
63 | #include <linux/net.h> |
64 | #include <net/sock.h> | |
65 | #include <net/af_unix.h> | |
6b06314c | 66 | #include <net/scm.h> |
2b188cc1 JA |
67 | #include <linux/anon_inodes.h> |
68 | #include <linux/sched/mm.h> | |
69 | #include <linux/uaccess.h> | |
70 | #include <linux/nospec.h> | |
edafccee JA |
71 | #include <linux/sizes.h> |
72 | #include <linux/hugetlb.h> | |
2b188cc1 JA |
73 | |
74 | #include <uapi/linux/io_uring.h> | |
75 | ||
76 | #include "internal.h" | |
77 | ||
78 | #define IORING_MAX_ENTRIES 4096 | |
6b06314c | 79 | #define IORING_MAX_FIXED_FILES 1024 |
2b188cc1 JA |
80 | |
81 | struct io_uring { | |
82 | u32 head ____cacheline_aligned_in_smp; | |
83 | u32 tail ____cacheline_aligned_in_smp; | |
84 | }; | |
85 | ||
1e84b97b | 86 | /* |
75b28aff HV |
87 | * This data is shared with the application through the mmap at offsets |
88 | * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING. | |
1e84b97b SB |
89 | * |
90 | * The offsets to the member fields are published through struct | |
91 | * io_sqring_offsets when calling io_uring_setup. | |
92 | */ | |
75b28aff | 93 | struct io_rings { |
1e84b97b SB |
94 | /* |
95 | * Head and tail offsets into the ring; the offsets need to be | |
96 | * masked to get valid indices. | |
97 | * | |
75b28aff HV |
98 | * The kernel controls head of the sq ring and the tail of the cq ring, |
99 | * and the application controls tail of the sq ring and the head of the | |
100 | * cq ring. | |
1e84b97b | 101 | */ |
75b28aff | 102 | struct io_uring sq, cq; |
1e84b97b | 103 | /* |
75b28aff | 104 | * Bitmasks to apply to head and tail offsets (constant, equals |
1e84b97b SB |
105 | * ring_entries - 1) |
106 | */ | |
75b28aff HV |
107 | u32 sq_ring_mask, cq_ring_mask; |
108 | /* Ring sizes (constant, power of 2) */ | |
109 | u32 sq_ring_entries, cq_ring_entries; | |
1e84b97b SB |
110 | /* |
111 | * Number of invalid entries dropped by the kernel due to | |
112 | * invalid index stored in array | |
113 | * | |
114 | * Written by the kernel, shouldn't be modified by the | |
115 | * application (i.e. get number of "new events" by comparing to | |
116 | * cached value). | |
117 | * | |
118 | * After a new SQ head value was read by the application this | |
119 | * counter includes all submissions that were dropped reaching | |
120 | * the new SQ head (and possibly more). | |
121 | */ | |
75b28aff | 122 | u32 sq_dropped; |
1e84b97b SB |
123 | /* |
124 | * Runtime flags | |
125 | * | |
126 | * Written by the kernel, shouldn't be modified by the | |
127 | * application. | |
128 | * | |
129 | * The application needs a full memory barrier before checking | |
130 | * for IORING_SQ_NEED_WAKEUP after updating the sq tail. | |
131 | */ | |
75b28aff | 132 | u32 sq_flags; |
1e84b97b SB |
133 | /* |
134 | * Number of completion events lost because the queue was full; | |
135 | * this should be avoided by the application by making sure | |
136 | * there are not more requests pending thatn there is space in | |
137 | * the completion queue. | |
138 | * | |
139 | * Written by the kernel, shouldn't be modified by the | |
140 | * application (i.e. get number of "new events" by comparing to | |
141 | * cached value). | |
142 | * | |
143 | * As completion events come in out of order this counter is not | |
144 | * ordered with any other data. | |
145 | */ | |
75b28aff | 146 | u32 cq_overflow; |
1e84b97b SB |
147 | /* |
148 | * Ring buffer of completion events. | |
149 | * | |
150 | * The kernel writes completion events fresh every time they are | |
151 | * produced, so the application is allowed to modify pending | |
152 | * entries. | |
153 | */ | |
75b28aff | 154 | struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp; |
2b188cc1 JA |
155 | }; |
156 | ||
edafccee JA |
157 | struct io_mapped_ubuf { |
158 | u64 ubuf; | |
159 | size_t len; | |
160 | struct bio_vec *bvec; | |
161 | unsigned int nr_bvecs; | |
162 | }; | |
163 | ||
31b51510 JA |
164 | struct async_list { |
165 | spinlock_t lock; | |
166 | atomic_t cnt; | |
167 | struct list_head list; | |
168 | ||
169 | struct file *file; | |
170 | off_t io_end; | |
9310a7ba | 171 | size_t io_len; |
31b51510 JA |
172 | }; |
173 | ||
2b188cc1 JA |
174 | struct io_ring_ctx { |
175 | struct { | |
176 | struct percpu_ref refs; | |
177 | } ____cacheline_aligned_in_smp; | |
178 | ||
179 | struct { | |
180 | unsigned int flags; | |
181 | bool compat; | |
182 | bool account_mem; | |
183 | ||
75b28aff HV |
184 | /* |
185 | * Ring buffer of indices into array of io_uring_sqe, which is | |
186 | * mmapped by the application using the IORING_OFF_SQES offset. | |
187 | * | |
188 | * This indirection could e.g. be used to assign fixed | |
189 | * io_uring_sqe entries to operations and only submit them to | |
190 | * the queue when needed. | |
191 | * | |
192 | * The kernel modifies neither the indices array nor the entries | |
193 | * array. | |
194 | */ | |
195 | u32 *sq_array; | |
2b188cc1 JA |
196 | unsigned cached_sq_head; |
197 | unsigned sq_entries; | |
198 | unsigned sq_mask; | |
6c271ce2 | 199 | unsigned sq_thread_idle; |
2b188cc1 | 200 | struct io_uring_sqe *sq_sqes; |
de0617e4 JA |
201 | |
202 | struct list_head defer_list; | |
2b188cc1 JA |
203 | } ____cacheline_aligned_in_smp; |
204 | ||
205 | /* IO offload */ | |
206 | struct workqueue_struct *sqo_wq; | |
6c271ce2 | 207 | struct task_struct *sqo_thread; /* if using sq thread polling */ |
2b188cc1 | 208 | struct mm_struct *sqo_mm; |
6c271ce2 | 209 | wait_queue_head_t sqo_wait; |
a4c0b3de | 210 | struct completion sqo_thread_started; |
2b188cc1 JA |
211 | |
212 | struct { | |
2b188cc1 JA |
213 | unsigned cached_cq_tail; |
214 | unsigned cq_entries; | |
215 | unsigned cq_mask; | |
216 | struct wait_queue_head cq_wait; | |
217 | struct fasync_struct *cq_fasync; | |
9b402849 | 218 | struct eventfd_ctx *cq_ev_fd; |
2b188cc1 JA |
219 | } ____cacheline_aligned_in_smp; |
220 | ||
75b28aff HV |
221 | struct io_rings *rings; |
222 | ||
6b06314c JA |
223 | /* |
224 | * If used, fixed file set. Writers must ensure that ->refs is dead, | |
225 | * readers must ensure that ->refs is alive as long as the file* is | |
226 | * used. Only updated through io_uring_register(2). | |
227 | */ | |
228 | struct file **user_files; | |
229 | unsigned nr_user_files; | |
230 | ||
edafccee JA |
231 | /* if used, fixed mapped user buffers */ |
232 | unsigned nr_user_bufs; | |
233 | struct io_mapped_ubuf *user_bufs; | |
234 | ||
2b188cc1 JA |
235 | struct user_struct *user; |
236 | ||
237 | struct completion ctx_done; | |
238 | ||
239 | struct { | |
240 | struct mutex uring_lock; | |
241 | wait_queue_head_t wait; | |
242 | } ____cacheline_aligned_in_smp; | |
243 | ||
244 | struct { | |
245 | spinlock_t completion_lock; | |
def596e9 JA |
246 | bool poll_multi_file; |
247 | /* | |
248 | * ->poll_list is protected by the ctx->uring_lock for | |
249 | * io_uring instances that don't use IORING_SETUP_SQPOLL. | |
250 | * For SQPOLL, only the single threaded io_sq_thread() will | |
251 | * manipulate the list, hence no extra locking is needed there. | |
252 | */ | |
253 | struct list_head poll_list; | |
221c5eb2 | 254 | struct list_head cancel_list; |
2b188cc1 JA |
255 | } ____cacheline_aligned_in_smp; |
256 | ||
31b51510 JA |
257 | struct async_list pending_async[2]; |
258 | ||
2b188cc1 JA |
259 | #if defined(CONFIG_UNIX) |
260 | struct socket *ring_sock; | |
261 | #endif | |
262 | }; | |
263 | ||
264 | struct sqe_submit { | |
265 | const struct io_uring_sqe *sqe; | |
266 | unsigned short index; | |
8776f3fa | 267 | u32 sequence; |
2b188cc1 | 268 | bool has_user; |
def596e9 | 269 | bool needs_lock; |
6c271ce2 | 270 | bool needs_fixed_file; |
2b188cc1 JA |
271 | }; |
272 | ||
09bb8394 JA |
273 | /* |
274 | * First field must be the file pointer in all the | |
275 | * iocb unions! See also 'struct kiocb' in <linux/fs.h> | |
276 | */ | |
221c5eb2 JA |
277 | struct io_poll_iocb { |
278 | struct file *file; | |
279 | struct wait_queue_head *head; | |
280 | __poll_t events; | |
8c838788 | 281 | bool done; |
221c5eb2 JA |
282 | bool canceled; |
283 | struct wait_queue_entry wait; | |
284 | }; | |
285 | ||
09bb8394 JA |
286 | /* |
287 | * NOTE! Each of the iocb union members has the file pointer | |
288 | * as the first entry in their struct definition. So you can | |
289 | * access the file pointer through any of the sub-structs, | |
290 | * or directly as just 'ki_filp' in this struct. | |
291 | */ | |
2b188cc1 | 292 | struct io_kiocb { |
221c5eb2 | 293 | union { |
09bb8394 | 294 | struct file *file; |
221c5eb2 JA |
295 | struct kiocb rw; |
296 | struct io_poll_iocb poll; | |
297 | }; | |
2b188cc1 JA |
298 | |
299 | struct sqe_submit submit; | |
300 | ||
301 | struct io_ring_ctx *ctx; | |
302 | struct list_head list; | |
9e645e11 | 303 | struct list_head link_list; |
2b188cc1 | 304 | unsigned int flags; |
c16361c1 | 305 | refcount_t refs; |
8449eeda | 306 | #define REQ_F_NOWAIT 1 /* must not punt to workers */ |
def596e9 | 307 | #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */ |
6b06314c | 308 | #define REQ_F_FIXED_FILE 4 /* ctx owns file */ |
31b51510 | 309 | #define REQ_F_SEQ_PREV 8 /* sequential with previous */ |
e2033e33 SB |
310 | #define REQ_F_IO_DRAIN 16 /* drain existing IO first */ |
311 | #define REQ_F_IO_DRAINED 32 /* drain done */ | |
9e645e11 | 312 | #define REQ_F_LINK 64 /* linked sqes */ |
f7b76ac9 ZL |
313 | #define REQ_F_LINK_DONE 128 /* linked sqes done */ |
314 | #define REQ_F_FAIL_LINK 256 /* fail rest of links */ | |
2b188cc1 | 315 | u64 user_data; |
9e645e11 | 316 | u32 result; |
de0617e4 | 317 | u32 sequence; |
2b188cc1 JA |
318 | |
319 | struct work_struct work; | |
320 | }; | |
321 | ||
322 | #define IO_PLUG_THRESHOLD 2 | |
def596e9 | 323 | #define IO_IOPOLL_BATCH 8 |
2b188cc1 | 324 | |
9a56a232 JA |
325 | struct io_submit_state { |
326 | struct blk_plug plug; | |
327 | ||
2579f913 JA |
328 | /* |
329 | * io_kiocb alloc cache | |
330 | */ | |
331 | void *reqs[IO_IOPOLL_BATCH]; | |
332 | unsigned int free_reqs; | |
333 | unsigned int cur_req; | |
334 | ||
9a56a232 JA |
335 | /* |
336 | * File reference cache | |
337 | */ | |
338 | struct file *file; | |
339 | unsigned int fd; | |
340 | unsigned int has_refs; | |
341 | unsigned int used_refs; | |
342 | unsigned int ios_left; | |
343 | }; | |
344 | ||
de0617e4 JA |
345 | static void io_sq_wq_submit_work(struct work_struct *work); |
346 | ||
2b188cc1 JA |
347 | static struct kmem_cache *req_cachep; |
348 | ||
349 | static const struct file_operations io_uring_fops; | |
350 | ||
351 | struct sock *io_uring_get_socket(struct file *file) | |
352 | { | |
353 | #if defined(CONFIG_UNIX) | |
354 | if (file->f_op == &io_uring_fops) { | |
355 | struct io_ring_ctx *ctx = file->private_data; | |
356 | ||
357 | return ctx->ring_sock->sk; | |
358 | } | |
359 | #endif | |
360 | return NULL; | |
361 | } | |
362 | EXPORT_SYMBOL(io_uring_get_socket); | |
363 | ||
364 | static void io_ring_ctx_ref_free(struct percpu_ref *ref) | |
365 | { | |
366 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
367 | ||
368 | complete(&ctx->ctx_done); | |
369 | } | |
370 | ||
371 | static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) | |
372 | { | |
373 | struct io_ring_ctx *ctx; | |
31b51510 | 374 | int i; |
2b188cc1 JA |
375 | |
376 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
377 | if (!ctx) | |
378 | return NULL; | |
379 | ||
21482896 RG |
380 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
381 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) { | |
2b188cc1 JA |
382 | kfree(ctx); |
383 | return NULL; | |
384 | } | |
385 | ||
386 | ctx->flags = p->flags; | |
387 | init_waitqueue_head(&ctx->cq_wait); | |
388 | init_completion(&ctx->ctx_done); | |
a4c0b3de | 389 | init_completion(&ctx->sqo_thread_started); |
2b188cc1 JA |
390 | mutex_init(&ctx->uring_lock); |
391 | init_waitqueue_head(&ctx->wait); | |
31b51510 JA |
392 | for (i = 0; i < ARRAY_SIZE(ctx->pending_async); i++) { |
393 | spin_lock_init(&ctx->pending_async[i].lock); | |
394 | INIT_LIST_HEAD(&ctx->pending_async[i].list); | |
395 | atomic_set(&ctx->pending_async[i].cnt, 0); | |
396 | } | |
2b188cc1 | 397 | spin_lock_init(&ctx->completion_lock); |
def596e9 | 398 | INIT_LIST_HEAD(&ctx->poll_list); |
221c5eb2 | 399 | INIT_LIST_HEAD(&ctx->cancel_list); |
de0617e4 | 400 | INIT_LIST_HEAD(&ctx->defer_list); |
2b188cc1 JA |
401 | return ctx; |
402 | } | |
403 | ||
de0617e4 JA |
404 | static inline bool io_sequence_defer(struct io_ring_ctx *ctx, |
405 | struct io_kiocb *req) | |
406 | { | |
407 | if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) != REQ_F_IO_DRAIN) | |
408 | return false; | |
409 | ||
75b28aff | 410 | return req->sequence != ctx->cached_cq_tail + ctx->rings->sq_dropped; |
de0617e4 JA |
411 | } |
412 | ||
413 | static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx) | |
414 | { | |
415 | struct io_kiocb *req; | |
416 | ||
417 | if (list_empty(&ctx->defer_list)) | |
418 | return NULL; | |
419 | ||
420 | req = list_first_entry(&ctx->defer_list, struct io_kiocb, list); | |
421 | if (!io_sequence_defer(ctx, req)) { | |
422 | list_del_init(&req->list); | |
423 | return req; | |
424 | } | |
425 | ||
426 | return NULL; | |
427 | } | |
428 | ||
429 | static void __io_commit_cqring(struct io_ring_ctx *ctx) | |
2b188cc1 | 430 | { |
75b28aff | 431 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 432 | |
75b28aff | 433 | if (ctx->cached_cq_tail != READ_ONCE(rings->cq.tail)) { |
2b188cc1 | 434 | /* order cqe stores with ring update */ |
75b28aff | 435 | smp_store_release(&rings->cq.tail, ctx->cached_cq_tail); |
2b188cc1 | 436 | |
2b188cc1 JA |
437 | if (wq_has_sleeper(&ctx->cq_wait)) { |
438 | wake_up_interruptible(&ctx->cq_wait); | |
439 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); | |
440 | } | |
441 | } | |
442 | } | |
443 | ||
de0617e4 JA |
444 | static void io_commit_cqring(struct io_ring_ctx *ctx) |
445 | { | |
446 | struct io_kiocb *req; | |
447 | ||
448 | __io_commit_cqring(ctx); | |
449 | ||
450 | while ((req = io_get_deferred_req(ctx)) != NULL) { | |
451 | req->flags |= REQ_F_IO_DRAINED; | |
452 | queue_work(ctx->sqo_wq, &req->work); | |
453 | } | |
454 | } | |
455 | ||
2b188cc1 JA |
456 | static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx) |
457 | { | |
75b28aff | 458 | struct io_rings *rings = ctx->rings; |
2b188cc1 JA |
459 | unsigned tail; |
460 | ||
461 | tail = ctx->cached_cq_tail; | |
115e12e5 SB |
462 | /* |
463 | * writes to the cq entry need to come after reading head; the | |
464 | * control dependency is enough as we're using WRITE_ONCE to | |
465 | * fill the cq entry | |
466 | */ | |
75b28aff | 467 | if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries) |
2b188cc1 JA |
468 | return NULL; |
469 | ||
470 | ctx->cached_cq_tail++; | |
75b28aff | 471 | return &rings->cqes[tail & ctx->cq_mask]; |
2b188cc1 JA |
472 | } |
473 | ||
474 | static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data, | |
c71ffb67 | 475 | long res) |
2b188cc1 JA |
476 | { |
477 | struct io_uring_cqe *cqe; | |
478 | ||
479 | /* | |
480 | * If we can't get a cq entry, userspace overflowed the | |
481 | * submission (by quite a lot). Increment the overflow count in | |
482 | * the ring. | |
483 | */ | |
484 | cqe = io_get_cqring(ctx); | |
485 | if (cqe) { | |
486 | WRITE_ONCE(cqe->user_data, ki_user_data); | |
487 | WRITE_ONCE(cqe->res, res); | |
c71ffb67 | 488 | WRITE_ONCE(cqe->flags, 0); |
2b188cc1 | 489 | } else { |
75b28aff | 490 | unsigned overflow = READ_ONCE(ctx->rings->cq_overflow); |
2b188cc1 | 491 | |
75b28aff | 492 | WRITE_ONCE(ctx->rings->cq_overflow, overflow + 1); |
2b188cc1 JA |
493 | } |
494 | } | |
495 | ||
8c838788 JA |
496 | static void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
497 | { | |
498 | if (waitqueue_active(&ctx->wait)) | |
499 | wake_up(&ctx->wait); | |
500 | if (waitqueue_active(&ctx->sqo_wait)) | |
501 | wake_up(&ctx->sqo_wait); | |
9b402849 JA |
502 | if (ctx->cq_ev_fd) |
503 | eventfd_signal(ctx->cq_ev_fd, 1); | |
8c838788 JA |
504 | } |
505 | ||
506 | static void io_cqring_add_event(struct io_ring_ctx *ctx, u64 user_data, | |
c71ffb67 | 507 | long res) |
2b188cc1 JA |
508 | { |
509 | unsigned long flags; | |
510 | ||
511 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
c71ffb67 | 512 | io_cqring_fill_event(ctx, user_data, res); |
2b188cc1 JA |
513 | io_commit_cqring(ctx); |
514 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
515 | ||
8c838788 | 516 | io_cqring_ev_posted(ctx); |
2b188cc1 JA |
517 | } |
518 | ||
519 | static void io_ring_drop_ctx_refs(struct io_ring_ctx *ctx, unsigned refs) | |
520 | { | |
521 | percpu_ref_put_many(&ctx->refs, refs); | |
522 | ||
523 | if (waitqueue_active(&ctx->wait)) | |
524 | wake_up(&ctx->wait); | |
525 | } | |
526 | ||
2579f913 JA |
527 | static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx, |
528 | struct io_submit_state *state) | |
2b188cc1 | 529 | { |
fd6fab2c | 530 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
2b188cc1 JA |
531 | struct io_kiocb *req; |
532 | ||
533 | if (!percpu_ref_tryget(&ctx->refs)) | |
534 | return NULL; | |
535 | ||
2579f913 | 536 | if (!state) { |
fd6fab2c | 537 | req = kmem_cache_alloc(req_cachep, gfp); |
2579f913 JA |
538 | if (unlikely(!req)) |
539 | goto out; | |
540 | } else if (!state->free_reqs) { | |
541 | size_t sz; | |
542 | int ret; | |
543 | ||
544 | sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs)); | |
fd6fab2c JA |
545 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs); |
546 | ||
547 | /* | |
548 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
549 | * retry single alloc to be on the safe side. | |
550 | */ | |
551 | if (unlikely(ret <= 0)) { | |
552 | state->reqs[0] = kmem_cache_alloc(req_cachep, gfp); | |
553 | if (!state->reqs[0]) | |
554 | goto out; | |
555 | ret = 1; | |
556 | } | |
2579f913 JA |
557 | state->free_reqs = ret - 1; |
558 | state->cur_req = 1; | |
559 | req = state->reqs[0]; | |
560 | } else { | |
561 | req = state->reqs[state->cur_req]; | |
562 | state->free_reqs--; | |
563 | state->cur_req++; | |
2b188cc1 JA |
564 | } |
565 | ||
60c112b0 | 566 | req->file = NULL; |
2579f913 JA |
567 | req->ctx = ctx; |
568 | req->flags = 0; | |
e65ef56d JA |
569 | /* one is dropped after submission, the other at completion */ |
570 | refcount_set(&req->refs, 2); | |
9e645e11 | 571 | req->result = 0; |
2579f913 JA |
572 | return req; |
573 | out: | |
2b188cc1 JA |
574 | io_ring_drop_ctx_refs(ctx, 1); |
575 | return NULL; | |
576 | } | |
577 | ||
def596e9 JA |
578 | static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr) |
579 | { | |
580 | if (*nr) { | |
581 | kmem_cache_free_bulk(req_cachep, *nr, reqs); | |
582 | io_ring_drop_ctx_refs(ctx, *nr); | |
583 | *nr = 0; | |
584 | } | |
585 | } | |
586 | ||
9e645e11 | 587 | static void __io_free_req(struct io_kiocb *req) |
2b188cc1 | 588 | { |
09bb8394 JA |
589 | if (req->file && !(req->flags & REQ_F_FIXED_FILE)) |
590 | fput(req->file); | |
e65ef56d JA |
591 | io_ring_drop_ctx_refs(req->ctx, 1); |
592 | kmem_cache_free(req_cachep, req); | |
593 | } | |
594 | ||
9e645e11 JA |
595 | static void io_req_link_next(struct io_kiocb *req) |
596 | { | |
597 | struct io_kiocb *nxt; | |
598 | ||
599 | /* | |
600 | * The list should never be empty when we are called here. But could | |
601 | * potentially happen if the chain is messed up, check to be on the | |
602 | * safe side. | |
603 | */ | |
604 | nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list); | |
605 | if (nxt) { | |
606 | list_del(&nxt->list); | |
607 | if (!list_empty(&req->link_list)) { | |
608 | INIT_LIST_HEAD(&nxt->link_list); | |
609 | list_splice(&req->link_list, &nxt->link_list); | |
610 | nxt->flags |= REQ_F_LINK; | |
611 | } | |
612 | ||
f7b76ac9 | 613 | nxt->flags |= REQ_F_LINK_DONE; |
9e645e11 JA |
614 | INIT_WORK(&nxt->work, io_sq_wq_submit_work); |
615 | queue_work(req->ctx->sqo_wq, &nxt->work); | |
616 | } | |
617 | } | |
618 | ||
619 | /* | |
620 | * Called if REQ_F_LINK is set, and we fail the head request | |
621 | */ | |
622 | static void io_fail_links(struct io_kiocb *req) | |
623 | { | |
624 | struct io_kiocb *link; | |
625 | ||
626 | while (!list_empty(&req->link_list)) { | |
627 | link = list_first_entry(&req->link_list, struct io_kiocb, list); | |
628 | list_del(&link->list); | |
629 | ||
630 | io_cqring_add_event(req->ctx, link->user_data, -ECANCELED); | |
631 | __io_free_req(link); | |
632 | } | |
633 | } | |
634 | ||
635 | static void io_free_req(struct io_kiocb *req) | |
636 | { | |
637 | /* | |
638 | * If LINK is set, we have dependent requests in this chain. If we | |
639 | * didn't fail this request, queue the first one up, moving any other | |
640 | * dependencies to the next request. In case of failure, fail the rest | |
641 | * of the chain. | |
642 | */ | |
643 | if (req->flags & REQ_F_LINK) { | |
644 | if (req->flags & REQ_F_FAIL_LINK) | |
645 | io_fail_links(req); | |
646 | else | |
647 | io_req_link_next(req); | |
648 | } | |
649 | ||
650 | __io_free_req(req); | |
651 | } | |
652 | ||
e65ef56d JA |
653 | static void io_put_req(struct io_kiocb *req) |
654 | { | |
655 | if (refcount_dec_and_test(&req->refs)) | |
656 | io_free_req(req); | |
2b188cc1 JA |
657 | } |
658 | ||
75b28aff | 659 | static unsigned io_cqring_events(struct io_rings *rings) |
a3a0e43f JA |
660 | { |
661 | /* See comment at the top of this file */ | |
662 | smp_rmb(); | |
75b28aff | 663 | return READ_ONCE(rings->cq.tail) - READ_ONCE(rings->cq.head); |
a3a0e43f JA |
664 | } |
665 | ||
def596e9 JA |
666 | /* |
667 | * Find and free completed poll iocbs | |
668 | */ | |
669 | static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
670 | struct list_head *done) | |
671 | { | |
672 | void *reqs[IO_IOPOLL_BATCH]; | |
673 | struct io_kiocb *req; | |
09bb8394 | 674 | int to_free; |
def596e9 | 675 | |
09bb8394 | 676 | to_free = 0; |
def596e9 JA |
677 | while (!list_empty(done)) { |
678 | req = list_first_entry(done, struct io_kiocb, list); | |
679 | list_del(&req->list); | |
680 | ||
9e645e11 | 681 | io_cqring_fill_event(ctx, req->user_data, req->result); |
def596e9 JA |
682 | (*nr_events)++; |
683 | ||
09bb8394 JA |
684 | if (refcount_dec_and_test(&req->refs)) { |
685 | /* If we're not using fixed files, we have to pair the | |
686 | * completion part with the file put. Use regular | |
687 | * completions for those, only batch free for fixed | |
9e645e11 | 688 | * file and non-linked commands. |
09bb8394 | 689 | */ |
9e645e11 JA |
690 | if ((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) == |
691 | REQ_F_FIXED_FILE) { | |
09bb8394 JA |
692 | reqs[to_free++] = req; |
693 | if (to_free == ARRAY_SIZE(reqs)) | |
694 | io_free_req_many(ctx, reqs, &to_free); | |
6b06314c | 695 | } else { |
09bb8394 | 696 | io_free_req(req); |
6b06314c | 697 | } |
9a56a232 | 698 | } |
def596e9 | 699 | } |
def596e9 | 700 | |
09bb8394 | 701 | io_commit_cqring(ctx); |
def596e9 JA |
702 | io_free_req_many(ctx, reqs, &to_free); |
703 | } | |
704 | ||
705 | static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
706 | long min) | |
707 | { | |
708 | struct io_kiocb *req, *tmp; | |
709 | LIST_HEAD(done); | |
710 | bool spin; | |
711 | int ret; | |
712 | ||
713 | /* | |
714 | * Only spin for completions if we don't have multiple devices hanging | |
715 | * off our complete list, and we're under the requested amount. | |
716 | */ | |
717 | spin = !ctx->poll_multi_file && *nr_events < min; | |
718 | ||
719 | ret = 0; | |
720 | list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) { | |
721 | struct kiocb *kiocb = &req->rw; | |
722 | ||
723 | /* | |
724 | * Move completed entries to our local list. If we find a | |
725 | * request that requires polling, break out and complete | |
726 | * the done list first, if we have entries there. | |
727 | */ | |
728 | if (req->flags & REQ_F_IOPOLL_COMPLETED) { | |
729 | list_move_tail(&req->list, &done); | |
730 | continue; | |
731 | } | |
732 | if (!list_empty(&done)) | |
733 | break; | |
734 | ||
735 | ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin); | |
736 | if (ret < 0) | |
737 | break; | |
738 | ||
739 | if (ret && spin) | |
740 | spin = false; | |
741 | ret = 0; | |
742 | } | |
743 | ||
744 | if (!list_empty(&done)) | |
745 | io_iopoll_complete(ctx, nr_events, &done); | |
746 | ||
747 | return ret; | |
748 | } | |
749 | ||
750 | /* | |
751 | * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a | |
752 | * non-spinning poll check - we'll still enter the driver poll loop, but only | |
753 | * as a non-spinning completion check. | |
754 | */ | |
755 | static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
756 | long min) | |
757 | { | |
08f5439f | 758 | while (!list_empty(&ctx->poll_list) && !need_resched()) { |
def596e9 JA |
759 | int ret; |
760 | ||
761 | ret = io_do_iopoll(ctx, nr_events, min); | |
762 | if (ret < 0) | |
763 | return ret; | |
764 | if (!min || *nr_events >= min) | |
765 | return 0; | |
766 | } | |
767 | ||
768 | return 1; | |
769 | } | |
770 | ||
771 | /* | |
772 | * We can't just wait for polled events to come to us, we have to actively | |
773 | * find and complete them. | |
774 | */ | |
775 | static void io_iopoll_reap_events(struct io_ring_ctx *ctx) | |
776 | { | |
777 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
778 | return; | |
779 | ||
780 | mutex_lock(&ctx->uring_lock); | |
781 | while (!list_empty(&ctx->poll_list)) { | |
782 | unsigned int nr_events = 0; | |
783 | ||
784 | io_iopoll_getevents(ctx, &nr_events, 1); | |
08f5439f JA |
785 | |
786 | /* | |
787 | * Ensure we allow local-to-the-cpu processing to take place, | |
788 | * in this case we need to ensure that we reap all events. | |
789 | */ | |
790 | cond_resched(); | |
def596e9 JA |
791 | } |
792 | mutex_unlock(&ctx->uring_lock); | |
793 | } | |
794 | ||
795 | static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, | |
796 | long min) | |
797 | { | |
500f9fba JA |
798 | int iters, ret = 0; |
799 | ||
800 | /* | |
801 | * We disallow the app entering submit/complete with polling, but we | |
802 | * still need to lock the ring to prevent racing with polled issue | |
803 | * that got punted to a workqueue. | |
804 | */ | |
805 | mutex_lock(&ctx->uring_lock); | |
def596e9 | 806 | |
500f9fba | 807 | iters = 0; |
def596e9 JA |
808 | do { |
809 | int tmin = 0; | |
810 | ||
a3a0e43f JA |
811 | /* |
812 | * Don't enter poll loop if we already have events pending. | |
813 | * If we do, we can potentially be spinning for commands that | |
814 | * already triggered a CQE (eg in error). | |
815 | */ | |
75b28aff | 816 | if (io_cqring_events(ctx->rings)) |
a3a0e43f JA |
817 | break; |
818 | ||
500f9fba JA |
819 | /* |
820 | * If a submit got punted to a workqueue, we can have the | |
821 | * application entering polling for a command before it gets | |
822 | * issued. That app will hold the uring_lock for the duration | |
823 | * of the poll right here, so we need to take a breather every | |
824 | * now and then to ensure that the issue has a chance to add | |
825 | * the poll to the issued list. Otherwise we can spin here | |
826 | * forever, while the workqueue is stuck trying to acquire the | |
827 | * very same mutex. | |
828 | */ | |
829 | if (!(++iters & 7)) { | |
830 | mutex_unlock(&ctx->uring_lock); | |
831 | mutex_lock(&ctx->uring_lock); | |
832 | } | |
833 | ||
def596e9 JA |
834 | if (*nr_events < min) |
835 | tmin = min - *nr_events; | |
836 | ||
837 | ret = io_iopoll_getevents(ctx, nr_events, tmin); | |
838 | if (ret <= 0) | |
839 | break; | |
840 | ret = 0; | |
841 | } while (min && !*nr_events && !need_resched()); | |
842 | ||
500f9fba | 843 | mutex_unlock(&ctx->uring_lock); |
def596e9 JA |
844 | return ret; |
845 | } | |
846 | ||
2b188cc1 JA |
847 | static void kiocb_end_write(struct kiocb *kiocb) |
848 | { | |
849 | if (kiocb->ki_flags & IOCB_WRITE) { | |
850 | struct inode *inode = file_inode(kiocb->ki_filp); | |
851 | ||
852 | /* | |
853 | * Tell lockdep we inherited freeze protection from submission | |
854 | * thread. | |
855 | */ | |
856 | if (S_ISREG(inode->i_mode)) | |
857 | __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE); | |
858 | file_end_write(kiocb->ki_filp); | |
859 | } | |
860 | } | |
861 | ||
862 | static void io_complete_rw(struct kiocb *kiocb, long res, long res2) | |
863 | { | |
864 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
865 | ||
866 | kiocb_end_write(kiocb); | |
867 | ||
9e645e11 JA |
868 | if ((req->flags & REQ_F_LINK) && res != req->result) |
869 | req->flags |= REQ_F_FAIL_LINK; | |
c71ffb67 | 870 | io_cqring_add_event(req->ctx, req->user_data, res); |
e65ef56d | 871 | io_put_req(req); |
2b188cc1 JA |
872 | } |
873 | ||
def596e9 JA |
874 | static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2) |
875 | { | |
876 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
877 | ||
878 | kiocb_end_write(kiocb); | |
879 | ||
9e645e11 JA |
880 | if ((req->flags & REQ_F_LINK) && res != req->result) |
881 | req->flags |= REQ_F_FAIL_LINK; | |
882 | req->result = res; | |
def596e9 JA |
883 | if (res != -EAGAIN) |
884 | req->flags |= REQ_F_IOPOLL_COMPLETED; | |
885 | } | |
886 | ||
887 | /* | |
888 | * After the iocb has been issued, it's safe to be found on the poll list. | |
889 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
890 | * find it from a io_iopoll_getevents() thread before the issuer is done | |
891 | * accessing the kiocb cookie. | |
892 | */ | |
893 | static void io_iopoll_req_issued(struct io_kiocb *req) | |
894 | { | |
895 | struct io_ring_ctx *ctx = req->ctx; | |
896 | ||
897 | /* | |
898 | * Track whether we have multiple files in our lists. This will impact | |
899 | * how we do polling eventually, not spinning if we're on potentially | |
900 | * different devices. | |
901 | */ | |
902 | if (list_empty(&ctx->poll_list)) { | |
903 | ctx->poll_multi_file = false; | |
904 | } else if (!ctx->poll_multi_file) { | |
905 | struct io_kiocb *list_req; | |
906 | ||
907 | list_req = list_first_entry(&ctx->poll_list, struct io_kiocb, | |
908 | list); | |
909 | if (list_req->rw.ki_filp != req->rw.ki_filp) | |
910 | ctx->poll_multi_file = true; | |
911 | } | |
912 | ||
913 | /* | |
914 | * For fast devices, IO may have already completed. If it has, add | |
915 | * it to the front so we find it first. | |
916 | */ | |
917 | if (req->flags & REQ_F_IOPOLL_COMPLETED) | |
918 | list_add(&req->list, &ctx->poll_list); | |
919 | else | |
920 | list_add_tail(&req->list, &ctx->poll_list); | |
921 | } | |
922 | ||
3d6770fb | 923 | static void io_file_put(struct io_submit_state *state) |
9a56a232 | 924 | { |
3d6770fb | 925 | if (state->file) { |
9a56a232 JA |
926 | int diff = state->has_refs - state->used_refs; |
927 | ||
928 | if (diff) | |
929 | fput_many(state->file, diff); | |
930 | state->file = NULL; | |
931 | } | |
932 | } | |
933 | ||
934 | /* | |
935 | * Get as many references to a file as we have IOs left in this submission, | |
936 | * assuming most submissions are for one file, or at least that each file | |
937 | * has more than one submission. | |
938 | */ | |
939 | static struct file *io_file_get(struct io_submit_state *state, int fd) | |
940 | { | |
941 | if (!state) | |
942 | return fget(fd); | |
943 | ||
944 | if (state->file) { | |
945 | if (state->fd == fd) { | |
946 | state->used_refs++; | |
947 | state->ios_left--; | |
948 | return state->file; | |
949 | } | |
3d6770fb | 950 | io_file_put(state); |
9a56a232 JA |
951 | } |
952 | state->file = fget_many(fd, state->ios_left); | |
953 | if (!state->file) | |
954 | return NULL; | |
955 | ||
956 | state->fd = fd; | |
957 | state->has_refs = state->ios_left; | |
958 | state->used_refs = 1; | |
959 | state->ios_left--; | |
960 | return state->file; | |
961 | } | |
962 | ||
2b188cc1 JA |
963 | /* |
964 | * If we tracked the file through the SCM inflight mechanism, we could support | |
965 | * any file. For now, just ensure that anything potentially problematic is done | |
966 | * inline. | |
967 | */ | |
968 | static bool io_file_supports_async(struct file *file) | |
969 | { | |
970 | umode_t mode = file_inode(file)->i_mode; | |
971 | ||
972 | if (S_ISBLK(mode) || S_ISCHR(mode)) | |
973 | return true; | |
974 | if (S_ISREG(mode) && file->f_op != &io_uring_fops) | |
975 | return true; | |
976 | ||
977 | return false; | |
978 | } | |
979 | ||
6c271ce2 | 980 | static int io_prep_rw(struct io_kiocb *req, const struct sqe_submit *s, |
8358e3a8 | 981 | bool force_nonblock) |
2b188cc1 | 982 | { |
6c271ce2 | 983 | const struct io_uring_sqe *sqe = s->sqe; |
def596e9 | 984 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 | 985 | struct kiocb *kiocb = &req->rw; |
09bb8394 JA |
986 | unsigned ioprio; |
987 | int ret; | |
2b188cc1 | 988 | |
09bb8394 JA |
989 | if (!req->file) |
990 | return -EBADF; | |
2b188cc1 | 991 | |
09bb8394 JA |
992 | if (force_nonblock && !io_file_supports_async(req->file)) |
993 | force_nonblock = false; | |
6b06314c | 994 | |
2b188cc1 JA |
995 | kiocb->ki_pos = READ_ONCE(sqe->off); |
996 | kiocb->ki_flags = iocb_flags(kiocb->ki_filp); | |
997 | kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp)); | |
998 | ||
999 | ioprio = READ_ONCE(sqe->ioprio); | |
1000 | if (ioprio) { | |
1001 | ret = ioprio_check_cap(ioprio); | |
1002 | if (ret) | |
09bb8394 | 1003 | return ret; |
2b188cc1 JA |
1004 | |
1005 | kiocb->ki_ioprio = ioprio; | |
1006 | } else | |
1007 | kiocb->ki_ioprio = get_current_ioprio(); | |
1008 | ||
1009 | ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags)); | |
1010 | if (unlikely(ret)) | |
09bb8394 | 1011 | return ret; |
8449eeda SB |
1012 | |
1013 | /* don't allow async punt if RWF_NOWAIT was requested */ | |
1014 | if (kiocb->ki_flags & IOCB_NOWAIT) | |
1015 | req->flags |= REQ_F_NOWAIT; | |
1016 | ||
1017 | if (force_nonblock) | |
2b188cc1 | 1018 | kiocb->ki_flags |= IOCB_NOWAIT; |
8449eeda | 1019 | |
def596e9 | 1020 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 JA |
1021 | if (!(kiocb->ki_flags & IOCB_DIRECT) || |
1022 | !kiocb->ki_filp->f_op->iopoll) | |
09bb8394 | 1023 | return -EOPNOTSUPP; |
2b188cc1 | 1024 | |
def596e9 JA |
1025 | kiocb->ki_flags |= IOCB_HIPRI; |
1026 | kiocb->ki_complete = io_complete_rw_iopoll; | |
1027 | } else { | |
09bb8394 JA |
1028 | if (kiocb->ki_flags & IOCB_HIPRI) |
1029 | return -EINVAL; | |
def596e9 JA |
1030 | kiocb->ki_complete = io_complete_rw; |
1031 | } | |
2b188cc1 | 1032 | return 0; |
2b188cc1 JA |
1033 | } |
1034 | ||
1035 | static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) | |
1036 | { | |
1037 | switch (ret) { | |
1038 | case -EIOCBQUEUED: | |
1039 | break; | |
1040 | case -ERESTARTSYS: | |
1041 | case -ERESTARTNOINTR: | |
1042 | case -ERESTARTNOHAND: | |
1043 | case -ERESTART_RESTARTBLOCK: | |
1044 | /* | |
1045 | * We can't just restart the syscall, since previously | |
1046 | * submitted sqes may already be in progress. Just fail this | |
1047 | * IO with EINTR. | |
1048 | */ | |
1049 | ret = -EINTR; | |
1050 | /* fall through */ | |
1051 | default: | |
1052 | kiocb->ki_complete(kiocb, ret, 0); | |
1053 | } | |
1054 | } | |
1055 | ||
edafccee JA |
1056 | static int io_import_fixed(struct io_ring_ctx *ctx, int rw, |
1057 | const struct io_uring_sqe *sqe, | |
1058 | struct iov_iter *iter) | |
1059 | { | |
1060 | size_t len = READ_ONCE(sqe->len); | |
1061 | struct io_mapped_ubuf *imu; | |
1062 | unsigned index, buf_index; | |
1063 | size_t offset; | |
1064 | u64 buf_addr; | |
1065 | ||
1066 | /* attempt to use fixed buffers without having provided iovecs */ | |
1067 | if (unlikely(!ctx->user_bufs)) | |
1068 | return -EFAULT; | |
1069 | ||
1070 | buf_index = READ_ONCE(sqe->buf_index); | |
1071 | if (unlikely(buf_index >= ctx->nr_user_bufs)) | |
1072 | return -EFAULT; | |
1073 | ||
1074 | index = array_index_nospec(buf_index, ctx->nr_user_bufs); | |
1075 | imu = &ctx->user_bufs[index]; | |
1076 | buf_addr = READ_ONCE(sqe->addr); | |
1077 | ||
1078 | /* overflow */ | |
1079 | if (buf_addr + len < buf_addr) | |
1080 | return -EFAULT; | |
1081 | /* not inside the mapped region */ | |
1082 | if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len) | |
1083 | return -EFAULT; | |
1084 | ||
1085 | /* | |
1086 | * May not be a start of buffer, set size appropriately | |
1087 | * and advance us to the beginning. | |
1088 | */ | |
1089 | offset = buf_addr - imu->ubuf; | |
1090 | iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len); | |
bd11b3a3 JA |
1091 | |
1092 | if (offset) { | |
1093 | /* | |
1094 | * Don't use iov_iter_advance() here, as it's really slow for | |
1095 | * using the latter parts of a big fixed buffer - it iterates | |
1096 | * over each segment manually. We can cheat a bit here, because | |
1097 | * we know that: | |
1098 | * | |
1099 | * 1) it's a BVEC iter, we set it up | |
1100 | * 2) all bvecs are PAGE_SIZE in size, except potentially the | |
1101 | * first and last bvec | |
1102 | * | |
1103 | * So just find our index, and adjust the iterator afterwards. | |
1104 | * If the offset is within the first bvec (or the whole first | |
1105 | * bvec, just use iov_iter_advance(). This makes it easier | |
1106 | * since we can just skip the first segment, which may not | |
1107 | * be PAGE_SIZE aligned. | |
1108 | */ | |
1109 | const struct bio_vec *bvec = imu->bvec; | |
1110 | ||
1111 | if (offset <= bvec->bv_len) { | |
1112 | iov_iter_advance(iter, offset); | |
1113 | } else { | |
1114 | unsigned long seg_skip; | |
1115 | ||
1116 | /* skip first vec */ | |
1117 | offset -= bvec->bv_len; | |
1118 | seg_skip = 1 + (offset >> PAGE_SHIFT); | |
1119 | ||
1120 | iter->bvec = bvec + seg_skip; | |
1121 | iter->nr_segs -= seg_skip; | |
99c79f66 | 1122 | iter->count -= bvec->bv_len + offset; |
bd11b3a3 | 1123 | iter->iov_offset = offset & ~PAGE_MASK; |
bd11b3a3 JA |
1124 | } |
1125 | } | |
1126 | ||
edafccee JA |
1127 | return 0; |
1128 | } | |
1129 | ||
87e5e6da JA |
1130 | static ssize_t io_import_iovec(struct io_ring_ctx *ctx, int rw, |
1131 | const struct sqe_submit *s, struct iovec **iovec, | |
1132 | struct iov_iter *iter) | |
2b188cc1 JA |
1133 | { |
1134 | const struct io_uring_sqe *sqe = s->sqe; | |
1135 | void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
1136 | size_t sqe_len = READ_ONCE(sqe->len); | |
edafccee JA |
1137 | u8 opcode; |
1138 | ||
1139 | /* | |
1140 | * We're reading ->opcode for the second time, but the first read | |
1141 | * doesn't care whether it's _FIXED or not, so it doesn't matter | |
1142 | * whether ->opcode changes concurrently. The first read does care | |
1143 | * about whether it is a READ or a WRITE, so we don't trust this read | |
1144 | * for that purpose and instead let the caller pass in the read/write | |
1145 | * flag. | |
1146 | */ | |
1147 | opcode = READ_ONCE(sqe->opcode); | |
1148 | if (opcode == IORING_OP_READ_FIXED || | |
1149 | opcode == IORING_OP_WRITE_FIXED) { | |
87e5e6da | 1150 | ssize_t ret = io_import_fixed(ctx, rw, sqe, iter); |
edafccee JA |
1151 | *iovec = NULL; |
1152 | return ret; | |
1153 | } | |
2b188cc1 JA |
1154 | |
1155 | if (!s->has_user) | |
1156 | return -EFAULT; | |
1157 | ||
1158 | #ifdef CONFIG_COMPAT | |
1159 | if (ctx->compat) | |
1160 | return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV, | |
1161 | iovec, iter); | |
1162 | #endif | |
1163 | ||
1164 | return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter); | |
1165 | } | |
1166 | ||
31b51510 JA |
1167 | /* |
1168 | * Make a note of the last file/offset/direction we punted to async | |
1169 | * context. We'll use this information to see if we can piggy back a | |
1170 | * sequential request onto the previous one, if it's still hasn't been | |
1171 | * completed by the async worker. | |
1172 | */ | |
1173 | static void io_async_list_note(int rw, struct io_kiocb *req, size_t len) | |
1174 | { | |
1175 | struct async_list *async_list = &req->ctx->pending_async[rw]; | |
1176 | struct kiocb *kiocb = &req->rw; | |
1177 | struct file *filp = kiocb->ki_filp; | |
1178 | off_t io_end = kiocb->ki_pos + len; | |
1179 | ||
1180 | if (filp == async_list->file && kiocb->ki_pos == async_list->io_end) { | |
9310a7ba | 1181 | unsigned long max_bytes; |
31b51510 JA |
1182 | |
1183 | /* Use 8x RA size as a decent limiter for both reads/writes */ | |
9310a7ba ZL |
1184 | max_bytes = filp->f_ra.ra_pages << (PAGE_SHIFT + 3); |
1185 | if (!max_bytes) | |
1186 | max_bytes = VM_READAHEAD_PAGES << (PAGE_SHIFT + 3); | |
1187 | ||
1188 | /* If max len are exceeded, reset the state */ | |
1189 | if (async_list->io_len + len <= max_bytes) { | |
31b51510 | 1190 | req->flags |= REQ_F_SEQ_PREV; |
9310a7ba | 1191 | async_list->io_len += len; |
31b51510 JA |
1192 | } else { |
1193 | io_end = 0; | |
9310a7ba | 1194 | async_list->io_len = 0; |
31b51510 JA |
1195 | } |
1196 | } | |
1197 | ||
1198 | /* New file? Reset state. */ | |
1199 | if (async_list->file != filp) { | |
9310a7ba | 1200 | async_list->io_len = 0; |
31b51510 JA |
1201 | async_list->file = filp; |
1202 | } | |
1203 | async_list->io_end = io_end; | |
1204 | } | |
1205 | ||
e0c5c576 | 1206 | static int io_read(struct io_kiocb *req, const struct sqe_submit *s, |
8358e3a8 | 1207 | bool force_nonblock) |
2b188cc1 JA |
1208 | { |
1209 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1210 | struct kiocb *kiocb = &req->rw; | |
1211 | struct iov_iter iter; | |
1212 | struct file *file; | |
31b51510 | 1213 | size_t iov_count; |
9d93a3f5 | 1214 | ssize_t read_size, ret; |
2b188cc1 | 1215 | |
8358e3a8 | 1216 | ret = io_prep_rw(req, s, force_nonblock); |
2b188cc1 JA |
1217 | if (ret) |
1218 | return ret; | |
1219 | file = kiocb->ki_filp; | |
1220 | ||
2b188cc1 | 1221 | if (unlikely(!(file->f_mode & FMODE_READ))) |
09bb8394 | 1222 | return -EBADF; |
2b188cc1 | 1223 | if (unlikely(!file->f_op->read_iter)) |
09bb8394 | 1224 | return -EINVAL; |
2b188cc1 JA |
1225 | |
1226 | ret = io_import_iovec(req->ctx, READ, s, &iovec, &iter); | |
87e5e6da | 1227 | if (ret < 0) |
09bb8394 | 1228 | return ret; |
2b188cc1 | 1229 | |
9d93a3f5 | 1230 | read_size = ret; |
9e645e11 JA |
1231 | if (req->flags & REQ_F_LINK) |
1232 | req->result = read_size; | |
1233 | ||
31b51510 JA |
1234 | iov_count = iov_iter_count(&iter); |
1235 | ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count); | |
2b188cc1 JA |
1236 | if (!ret) { |
1237 | ssize_t ret2; | |
1238 | ||
2b188cc1 | 1239 | ret2 = call_read_iter(file, kiocb, &iter); |
9d93a3f5 JA |
1240 | /* |
1241 | * In case of a short read, punt to async. This can happen | |
1242 | * if we have data partially cached. Alternatively we can | |
1243 | * return the short read, in which case the application will | |
1244 | * need to issue another SQE and wait for it. That SQE will | |
1245 | * need async punt anyway, so it's more efficient to do it | |
1246 | * here. | |
1247 | */ | |
1248 | if (force_nonblock && ret2 > 0 && ret2 < read_size) | |
1249 | ret2 = -EAGAIN; | |
1250 | /* Catch -EAGAIN return for forced non-blocking submission */ | |
31b51510 | 1251 | if (!force_nonblock || ret2 != -EAGAIN) { |
2b188cc1 | 1252 | io_rw_done(kiocb, ret2); |
31b51510 JA |
1253 | } else { |
1254 | /* | |
1255 | * If ->needs_lock is true, we're already in async | |
1256 | * context. | |
1257 | */ | |
1258 | if (!s->needs_lock) | |
1259 | io_async_list_note(READ, req, iov_count); | |
2b188cc1 | 1260 | ret = -EAGAIN; |
31b51510 | 1261 | } |
2b188cc1 JA |
1262 | } |
1263 | kfree(iovec); | |
2b188cc1 JA |
1264 | return ret; |
1265 | } | |
1266 | ||
e0c5c576 | 1267 | static int io_write(struct io_kiocb *req, const struct sqe_submit *s, |
8358e3a8 | 1268 | bool force_nonblock) |
2b188cc1 JA |
1269 | { |
1270 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1271 | struct kiocb *kiocb = &req->rw; | |
1272 | struct iov_iter iter; | |
1273 | struct file *file; | |
31b51510 | 1274 | size_t iov_count; |
87e5e6da | 1275 | ssize_t ret; |
2b188cc1 | 1276 | |
8358e3a8 | 1277 | ret = io_prep_rw(req, s, force_nonblock); |
2b188cc1 JA |
1278 | if (ret) |
1279 | return ret; | |
2b188cc1 | 1280 | |
2b188cc1 JA |
1281 | file = kiocb->ki_filp; |
1282 | if (unlikely(!(file->f_mode & FMODE_WRITE))) | |
09bb8394 | 1283 | return -EBADF; |
2b188cc1 | 1284 | if (unlikely(!file->f_op->write_iter)) |
09bb8394 | 1285 | return -EINVAL; |
2b188cc1 JA |
1286 | |
1287 | ret = io_import_iovec(req->ctx, WRITE, s, &iovec, &iter); | |
87e5e6da | 1288 | if (ret < 0) |
09bb8394 | 1289 | return ret; |
2b188cc1 | 1290 | |
9e645e11 JA |
1291 | if (req->flags & REQ_F_LINK) |
1292 | req->result = ret; | |
1293 | ||
31b51510 JA |
1294 | iov_count = iov_iter_count(&iter); |
1295 | ||
1296 | ret = -EAGAIN; | |
1297 | if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT)) { | |
1298 | /* If ->needs_lock is true, we're already in async context. */ | |
1299 | if (!s->needs_lock) | |
1300 | io_async_list_note(WRITE, req, iov_count); | |
1301 | goto out_free; | |
1302 | } | |
1303 | ||
1304 | ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count); | |
2b188cc1 | 1305 | if (!ret) { |
9bf7933f RP |
1306 | ssize_t ret2; |
1307 | ||
2b188cc1 JA |
1308 | /* |
1309 | * Open-code file_start_write here to grab freeze protection, | |
1310 | * which will be released by another thread in | |
1311 | * io_complete_rw(). Fool lockdep by telling it the lock got | |
1312 | * released so that it doesn't complain about the held lock when | |
1313 | * we return to userspace. | |
1314 | */ | |
1315 | if (S_ISREG(file_inode(file)->i_mode)) { | |
1316 | __sb_start_write(file_inode(file)->i_sb, | |
1317 | SB_FREEZE_WRITE, true); | |
1318 | __sb_writers_release(file_inode(file)->i_sb, | |
1319 | SB_FREEZE_WRITE); | |
1320 | } | |
1321 | kiocb->ki_flags |= IOCB_WRITE; | |
9bf7933f RP |
1322 | |
1323 | ret2 = call_write_iter(file, kiocb, &iter); | |
1324 | if (!force_nonblock || ret2 != -EAGAIN) { | |
1325 | io_rw_done(kiocb, ret2); | |
1326 | } else { | |
1327 | /* | |
1328 | * If ->needs_lock is true, we're already in async | |
1329 | * context. | |
1330 | */ | |
1331 | if (!s->needs_lock) | |
1332 | io_async_list_note(WRITE, req, iov_count); | |
1333 | ret = -EAGAIN; | |
1334 | } | |
2b188cc1 | 1335 | } |
31b51510 | 1336 | out_free: |
2b188cc1 | 1337 | kfree(iovec); |
2b188cc1 JA |
1338 | return ret; |
1339 | } | |
1340 | ||
1341 | /* | |
1342 | * IORING_OP_NOP just posts a completion event, nothing else. | |
1343 | */ | |
1344 | static int io_nop(struct io_kiocb *req, u64 user_data) | |
1345 | { | |
1346 | struct io_ring_ctx *ctx = req->ctx; | |
1347 | long err = 0; | |
1348 | ||
def596e9 JA |
1349 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
1350 | return -EINVAL; | |
1351 | ||
c71ffb67 | 1352 | io_cqring_add_event(ctx, user_data, err); |
e65ef56d | 1353 | io_put_req(req); |
2b188cc1 JA |
1354 | return 0; |
1355 | } | |
1356 | ||
c992fe29 CH |
1357 | static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
1358 | { | |
6b06314c | 1359 | struct io_ring_ctx *ctx = req->ctx; |
c992fe29 | 1360 | |
09bb8394 JA |
1361 | if (!req->file) |
1362 | return -EBADF; | |
c992fe29 | 1363 | |
6b06314c | 1364 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
def596e9 | 1365 | return -EINVAL; |
edafccee | 1366 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) |
c992fe29 CH |
1367 | return -EINVAL; |
1368 | ||
c992fe29 CH |
1369 | return 0; |
1370 | } | |
1371 | ||
1372 | static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
1373 | bool force_nonblock) | |
1374 | { | |
1375 | loff_t sqe_off = READ_ONCE(sqe->off); | |
1376 | loff_t sqe_len = READ_ONCE(sqe->len); | |
1377 | loff_t end = sqe_off + sqe_len; | |
1378 | unsigned fsync_flags; | |
1379 | int ret; | |
1380 | ||
1381 | fsync_flags = READ_ONCE(sqe->fsync_flags); | |
1382 | if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC)) | |
1383 | return -EINVAL; | |
1384 | ||
1385 | ret = io_prep_fsync(req, sqe); | |
1386 | if (ret) | |
1387 | return ret; | |
1388 | ||
1389 | /* fsync always requires a blocking context */ | |
1390 | if (force_nonblock) | |
1391 | return -EAGAIN; | |
1392 | ||
1393 | ret = vfs_fsync_range(req->rw.ki_filp, sqe_off, | |
1394 | end > 0 ? end : LLONG_MAX, | |
1395 | fsync_flags & IORING_FSYNC_DATASYNC); | |
1396 | ||
9e645e11 JA |
1397 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1398 | req->flags |= REQ_F_FAIL_LINK; | |
c71ffb67 | 1399 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
e65ef56d | 1400 | io_put_req(req); |
c992fe29 CH |
1401 | return 0; |
1402 | } | |
1403 | ||
5d17b4a4 JA |
1404 | static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
1405 | { | |
1406 | struct io_ring_ctx *ctx = req->ctx; | |
1407 | int ret = 0; | |
1408 | ||
1409 | if (!req->file) | |
1410 | return -EBADF; | |
5d17b4a4 JA |
1411 | |
1412 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
1413 | return -EINVAL; | |
1414 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) | |
1415 | return -EINVAL; | |
1416 | ||
5d17b4a4 JA |
1417 | return ret; |
1418 | } | |
1419 | ||
1420 | static int io_sync_file_range(struct io_kiocb *req, | |
1421 | const struct io_uring_sqe *sqe, | |
1422 | bool force_nonblock) | |
1423 | { | |
1424 | loff_t sqe_off; | |
1425 | loff_t sqe_len; | |
1426 | unsigned flags; | |
1427 | int ret; | |
1428 | ||
1429 | ret = io_prep_sfr(req, sqe); | |
1430 | if (ret) | |
1431 | return ret; | |
1432 | ||
1433 | /* sync_file_range always requires a blocking context */ | |
1434 | if (force_nonblock) | |
1435 | return -EAGAIN; | |
1436 | ||
1437 | sqe_off = READ_ONCE(sqe->off); | |
1438 | sqe_len = READ_ONCE(sqe->len); | |
1439 | flags = READ_ONCE(sqe->sync_range_flags); | |
1440 | ||
1441 | ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags); | |
1442 | ||
9e645e11 JA |
1443 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1444 | req->flags |= REQ_F_FAIL_LINK; | |
c71ffb67 | 1445 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
5d17b4a4 JA |
1446 | io_put_req(req); |
1447 | return 0; | |
1448 | } | |
1449 | ||
0fa03c62 | 1450 | #if defined(CONFIG_NET) |
aa1fa28f JA |
1451 | static int io_send_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
1452 | bool force_nonblock, | |
1453 | long (*fn)(struct socket *, struct user_msghdr __user *, | |
1454 | unsigned int)) | |
1455 | { | |
0fa03c62 JA |
1456 | struct socket *sock; |
1457 | int ret; | |
1458 | ||
1459 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1460 | return -EINVAL; | |
1461 | ||
1462 | sock = sock_from_file(req->file, &ret); | |
1463 | if (sock) { | |
1464 | struct user_msghdr __user *msg; | |
1465 | unsigned flags; | |
1466 | ||
1467 | flags = READ_ONCE(sqe->msg_flags); | |
1468 | if (flags & MSG_DONTWAIT) | |
1469 | req->flags |= REQ_F_NOWAIT; | |
1470 | else if (force_nonblock) | |
1471 | flags |= MSG_DONTWAIT; | |
1472 | ||
1473 | msg = (struct user_msghdr __user *) (unsigned long) | |
1474 | READ_ONCE(sqe->addr); | |
1475 | ||
aa1fa28f | 1476 | ret = fn(sock, msg, flags); |
0fa03c62 JA |
1477 | if (force_nonblock && ret == -EAGAIN) |
1478 | return ret; | |
1479 | } | |
1480 | ||
c71ffb67 | 1481 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
5d17b4a4 JA |
1482 | io_put_req(req); |
1483 | return 0; | |
1484 | } | |
aa1fa28f JA |
1485 | #endif |
1486 | ||
1487 | static int io_sendmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
1488 | bool force_nonblock) | |
1489 | { | |
1490 | #if defined(CONFIG_NET) | |
1491 | return io_send_recvmsg(req, sqe, force_nonblock, __sys_sendmsg_sock); | |
1492 | #else | |
1493 | return -EOPNOTSUPP; | |
1494 | #endif | |
1495 | } | |
1496 | ||
1497 | static int io_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
1498 | bool force_nonblock) | |
1499 | { | |
1500 | #if defined(CONFIG_NET) | |
1501 | return io_send_recvmsg(req, sqe, force_nonblock, __sys_recvmsg_sock); | |
0fa03c62 JA |
1502 | #else |
1503 | return -EOPNOTSUPP; | |
1504 | #endif | |
1505 | } | |
5d17b4a4 | 1506 | |
221c5eb2 JA |
1507 | static void io_poll_remove_one(struct io_kiocb *req) |
1508 | { | |
1509 | struct io_poll_iocb *poll = &req->poll; | |
1510 | ||
1511 | spin_lock(&poll->head->lock); | |
1512 | WRITE_ONCE(poll->canceled, true); | |
1513 | if (!list_empty(&poll->wait.entry)) { | |
1514 | list_del_init(&poll->wait.entry); | |
1515 | queue_work(req->ctx->sqo_wq, &req->work); | |
1516 | } | |
1517 | spin_unlock(&poll->head->lock); | |
1518 | ||
1519 | list_del_init(&req->list); | |
1520 | } | |
1521 | ||
1522 | static void io_poll_remove_all(struct io_ring_ctx *ctx) | |
1523 | { | |
1524 | struct io_kiocb *req; | |
1525 | ||
1526 | spin_lock_irq(&ctx->completion_lock); | |
1527 | while (!list_empty(&ctx->cancel_list)) { | |
1528 | req = list_first_entry(&ctx->cancel_list, struct io_kiocb,list); | |
1529 | io_poll_remove_one(req); | |
1530 | } | |
1531 | spin_unlock_irq(&ctx->completion_lock); | |
1532 | } | |
1533 | ||
1534 | /* | |
1535 | * Find a running poll command that matches one specified in sqe->addr, | |
1536 | * and remove it if found. | |
1537 | */ | |
1538 | static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
1539 | { | |
1540 | struct io_ring_ctx *ctx = req->ctx; | |
1541 | struct io_kiocb *poll_req, *next; | |
1542 | int ret = -ENOENT; | |
1543 | ||
1544 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1545 | return -EINVAL; | |
1546 | if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index || | |
1547 | sqe->poll_events) | |
1548 | return -EINVAL; | |
1549 | ||
1550 | spin_lock_irq(&ctx->completion_lock); | |
1551 | list_for_each_entry_safe(poll_req, next, &ctx->cancel_list, list) { | |
1552 | if (READ_ONCE(sqe->addr) == poll_req->user_data) { | |
1553 | io_poll_remove_one(poll_req); | |
1554 | ret = 0; | |
1555 | break; | |
1556 | } | |
1557 | } | |
1558 | spin_unlock_irq(&ctx->completion_lock); | |
1559 | ||
c71ffb67 | 1560 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
e65ef56d | 1561 | io_put_req(req); |
221c5eb2 JA |
1562 | return 0; |
1563 | } | |
1564 | ||
8c838788 JA |
1565 | static void io_poll_complete(struct io_ring_ctx *ctx, struct io_kiocb *req, |
1566 | __poll_t mask) | |
221c5eb2 | 1567 | { |
8c838788 | 1568 | req->poll.done = true; |
c71ffb67 | 1569 | io_cqring_fill_event(ctx, req->user_data, mangle_poll(mask)); |
8c838788 | 1570 | io_commit_cqring(ctx); |
221c5eb2 JA |
1571 | } |
1572 | ||
1573 | static void io_poll_complete_work(struct work_struct *work) | |
1574 | { | |
1575 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
1576 | struct io_poll_iocb *poll = &req->poll; | |
1577 | struct poll_table_struct pt = { ._key = poll->events }; | |
1578 | struct io_ring_ctx *ctx = req->ctx; | |
1579 | __poll_t mask = 0; | |
1580 | ||
1581 | if (!READ_ONCE(poll->canceled)) | |
1582 | mask = vfs_poll(poll->file, &pt) & poll->events; | |
1583 | ||
1584 | /* | |
1585 | * Note that ->ki_cancel callers also delete iocb from active_reqs after | |
1586 | * calling ->ki_cancel. We need the ctx_lock roundtrip here to | |
1587 | * synchronize with them. In the cancellation case the list_del_init | |
1588 | * itself is not actually needed, but harmless so we keep it in to | |
1589 | * avoid further branches in the fast path. | |
1590 | */ | |
1591 | spin_lock_irq(&ctx->completion_lock); | |
1592 | if (!mask && !READ_ONCE(poll->canceled)) { | |
1593 | add_wait_queue(poll->head, &poll->wait); | |
1594 | spin_unlock_irq(&ctx->completion_lock); | |
1595 | return; | |
1596 | } | |
1597 | list_del_init(&req->list); | |
8c838788 | 1598 | io_poll_complete(ctx, req, mask); |
221c5eb2 JA |
1599 | spin_unlock_irq(&ctx->completion_lock); |
1600 | ||
8c838788 JA |
1601 | io_cqring_ev_posted(ctx); |
1602 | io_put_req(req); | |
221c5eb2 JA |
1603 | } |
1604 | ||
1605 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, | |
1606 | void *key) | |
1607 | { | |
1608 | struct io_poll_iocb *poll = container_of(wait, struct io_poll_iocb, | |
1609 | wait); | |
1610 | struct io_kiocb *req = container_of(poll, struct io_kiocb, poll); | |
1611 | struct io_ring_ctx *ctx = req->ctx; | |
1612 | __poll_t mask = key_to_poll(key); | |
8c838788 | 1613 | unsigned long flags; |
221c5eb2 JA |
1614 | |
1615 | /* for instances that support it check for an event match first: */ | |
8c838788 JA |
1616 | if (mask && !(mask & poll->events)) |
1617 | return 0; | |
221c5eb2 | 1618 | |
8c838788 | 1619 | list_del_init(&poll->wait.entry); |
221c5eb2 | 1620 | |
8c838788 JA |
1621 | if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) { |
1622 | list_del(&req->list); | |
1623 | io_poll_complete(ctx, req, mask); | |
1624 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
221c5eb2 | 1625 | |
8c838788 JA |
1626 | io_cqring_ev_posted(ctx); |
1627 | io_put_req(req); | |
1628 | } else { | |
1629 | queue_work(ctx->sqo_wq, &req->work); | |
221c5eb2 JA |
1630 | } |
1631 | ||
221c5eb2 JA |
1632 | return 1; |
1633 | } | |
1634 | ||
1635 | struct io_poll_table { | |
1636 | struct poll_table_struct pt; | |
1637 | struct io_kiocb *req; | |
1638 | int error; | |
1639 | }; | |
1640 | ||
1641 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, | |
1642 | struct poll_table_struct *p) | |
1643 | { | |
1644 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
1645 | ||
1646 | if (unlikely(pt->req->poll.head)) { | |
1647 | pt->error = -EINVAL; | |
1648 | return; | |
1649 | } | |
1650 | ||
1651 | pt->error = 0; | |
1652 | pt->req->poll.head = head; | |
1653 | add_wait_queue(head, &pt->req->poll.wait); | |
1654 | } | |
1655 | ||
1656 | static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
1657 | { | |
1658 | struct io_poll_iocb *poll = &req->poll; | |
1659 | struct io_ring_ctx *ctx = req->ctx; | |
1660 | struct io_poll_table ipt; | |
8c838788 | 1661 | bool cancel = false; |
221c5eb2 JA |
1662 | __poll_t mask; |
1663 | u16 events; | |
221c5eb2 JA |
1664 | |
1665 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1666 | return -EINVAL; | |
1667 | if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index) | |
1668 | return -EINVAL; | |
09bb8394 JA |
1669 | if (!poll->file) |
1670 | return -EBADF; | |
221c5eb2 JA |
1671 | |
1672 | INIT_WORK(&req->work, io_poll_complete_work); | |
1673 | events = READ_ONCE(sqe->poll_events); | |
1674 | poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP; | |
1675 | ||
221c5eb2 | 1676 | poll->head = NULL; |
8c838788 | 1677 | poll->done = false; |
221c5eb2 JA |
1678 | poll->canceled = false; |
1679 | ||
1680 | ipt.pt._qproc = io_poll_queue_proc; | |
1681 | ipt.pt._key = poll->events; | |
1682 | ipt.req = req; | |
1683 | ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */ | |
1684 | ||
1685 | /* initialized the list so that we can do list_empty checks */ | |
1686 | INIT_LIST_HEAD(&poll->wait.entry); | |
1687 | init_waitqueue_func_entry(&poll->wait, io_poll_wake); | |
1688 | ||
36703247 JA |
1689 | INIT_LIST_HEAD(&req->list); |
1690 | ||
221c5eb2 | 1691 | mask = vfs_poll(poll->file, &ipt.pt) & poll->events; |
221c5eb2 JA |
1692 | |
1693 | spin_lock_irq(&ctx->completion_lock); | |
8c838788 JA |
1694 | if (likely(poll->head)) { |
1695 | spin_lock(&poll->head->lock); | |
1696 | if (unlikely(list_empty(&poll->wait.entry))) { | |
1697 | if (ipt.error) | |
1698 | cancel = true; | |
1699 | ipt.error = 0; | |
1700 | mask = 0; | |
1701 | } | |
1702 | if (mask || ipt.error) | |
1703 | list_del_init(&poll->wait.entry); | |
1704 | else if (cancel) | |
1705 | WRITE_ONCE(poll->canceled, true); | |
1706 | else if (!poll->done) /* actually waiting for an event */ | |
1707 | list_add_tail(&req->list, &ctx->cancel_list); | |
1708 | spin_unlock(&poll->head->lock); | |
1709 | } | |
1710 | if (mask) { /* no async, we'd stolen it */ | |
221c5eb2 | 1711 | ipt.error = 0; |
8c838788 | 1712 | io_poll_complete(ctx, req, mask); |
221c5eb2 | 1713 | } |
221c5eb2 JA |
1714 | spin_unlock_irq(&ctx->completion_lock); |
1715 | ||
8c838788 JA |
1716 | if (mask) { |
1717 | io_cqring_ev_posted(ctx); | |
e65ef56d | 1718 | io_put_req(req); |
221c5eb2 | 1719 | } |
8c838788 | 1720 | return ipt.error; |
221c5eb2 JA |
1721 | } |
1722 | ||
de0617e4 JA |
1723 | static int io_req_defer(struct io_ring_ctx *ctx, struct io_kiocb *req, |
1724 | const struct io_uring_sqe *sqe) | |
1725 | { | |
1726 | struct io_uring_sqe *sqe_copy; | |
1727 | ||
1728 | if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list)) | |
1729 | return 0; | |
1730 | ||
1731 | sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL); | |
1732 | if (!sqe_copy) | |
1733 | return -EAGAIN; | |
1734 | ||
1735 | spin_lock_irq(&ctx->completion_lock); | |
1736 | if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list)) { | |
1737 | spin_unlock_irq(&ctx->completion_lock); | |
1738 | kfree(sqe_copy); | |
1739 | return 0; | |
1740 | } | |
1741 | ||
1742 | memcpy(sqe_copy, sqe, sizeof(*sqe_copy)); | |
1743 | req->submit.sqe = sqe_copy; | |
1744 | ||
1745 | INIT_WORK(&req->work, io_sq_wq_submit_work); | |
1746 | list_add_tail(&req->list, &ctx->defer_list); | |
1747 | spin_unlock_irq(&ctx->completion_lock); | |
1748 | return -EIOCBQUEUED; | |
1749 | } | |
1750 | ||
2b188cc1 | 1751 | static int __io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, |
8358e3a8 | 1752 | const struct sqe_submit *s, bool force_nonblock) |
2b188cc1 | 1753 | { |
e0c5c576 | 1754 | int ret, opcode; |
2b188cc1 | 1755 | |
9e645e11 JA |
1756 | req->user_data = READ_ONCE(s->sqe->user_data); |
1757 | ||
2b188cc1 JA |
1758 | if (unlikely(s->index >= ctx->sq_entries)) |
1759 | return -EINVAL; | |
2b188cc1 JA |
1760 | |
1761 | opcode = READ_ONCE(s->sqe->opcode); | |
1762 | switch (opcode) { | |
1763 | case IORING_OP_NOP: | |
1764 | ret = io_nop(req, req->user_data); | |
1765 | break; | |
1766 | case IORING_OP_READV: | |
edafccee JA |
1767 | if (unlikely(s->sqe->buf_index)) |
1768 | return -EINVAL; | |
8358e3a8 | 1769 | ret = io_read(req, s, force_nonblock); |
2b188cc1 JA |
1770 | break; |
1771 | case IORING_OP_WRITEV: | |
edafccee JA |
1772 | if (unlikely(s->sqe->buf_index)) |
1773 | return -EINVAL; | |
8358e3a8 | 1774 | ret = io_write(req, s, force_nonblock); |
edafccee JA |
1775 | break; |
1776 | case IORING_OP_READ_FIXED: | |
8358e3a8 | 1777 | ret = io_read(req, s, force_nonblock); |
edafccee JA |
1778 | break; |
1779 | case IORING_OP_WRITE_FIXED: | |
8358e3a8 | 1780 | ret = io_write(req, s, force_nonblock); |
2b188cc1 | 1781 | break; |
c992fe29 CH |
1782 | case IORING_OP_FSYNC: |
1783 | ret = io_fsync(req, s->sqe, force_nonblock); | |
1784 | break; | |
221c5eb2 JA |
1785 | case IORING_OP_POLL_ADD: |
1786 | ret = io_poll_add(req, s->sqe); | |
1787 | break; | |
1788 | case IORING_OP_POLL_REMOVE: | |
1789 | ret = io_poll_remove(req, s->sqe); | |
1790 | break; | |
5d17b4a4 JA |
1791 | case IORING_OP_SYNC_FILE_RANGE: |
1792 | ret = io_sync_file_range(req, s->sqe, force_nonblock); | |
1793 | break; | |
0fa03c62 JA |
1794 | case IORING_OP_SENDMSG: |
1795 | ret = io_sendmsg(req, s->sqe, force_nonblock); | |
1796 | break; | |
aa1fa28f JA |
1797 | case IORING_OP_RECVMSG: |
1798 | ret = io_recvmsg(req, s->sqe, force_nonblock); | |
1799 | break; | |
2b188cc1 JA |
1800 | default: |
1801 | ret = -EINVAL; | |
1802 | break; | |
1803 | } | |
1804 | ||
def596e9 JA |
1805 | if (ret) |
1806 | return ret; | |
1807 | ||
1808 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
9e645e11 | 1809 | if (req->result == -EAGAIN) |
def596e9 JA |
1810 | return -EAGAIN; |
1811 | ||
1812 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
1813 | if (s->needs_lock) | |
1814 | mutex_lock(&ctx->uring_lock); | |
1815 | io_iopoll_req_issued(req); | |
1816 | if (s->needs_lock) | |
1817 | mutex_unlock(&ctx->uring_lock); | |
1818 | } | |
1819 | ||
1820 | return 0; | |
2b188cc1 JA |
1821 | } |
1822 | ||
31b51510 JA |
1823 | static struct async_list *io_async_list_from_sqe(struct io_ring_ctx *ctx, |
1824 | const struct io_uring_sqe *sqe) | |
1825 | { | |
1826 | switch (sqe->opcode) { | |
1827 | case IORING_OP_READV: | |
1828 | case IORING_OP_READ_FIXED: | |
1829 | return &ctx->pending_async[READ]; | |
1830 | case IORING_OP_WRITEV: | |
1831 | case IORING_OP_WRITE_FIXED: | |
1832 | return &ctx->pending_async[WRITE]; | |
1833 | default: | |
1834 | return NULL; | |
1835 | } | |
1836 | } | |
1837 | ||
edafccee JA |
1838 | static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe) |
1839 | { | |
1840 | u8 opcode = READ_ONCE(sqe->opcode); | |
1841 | ||
1842 | return !(opcode == IORING_OP_READ_FIXED || | |
1843 | opcode == IORING_OP_WRITE_FIXED); | |
1844 | } | |
1845 | ||
2b188cc1 JA |
1846 | static void io_sq_wq_submit_work(struct work_struct *work) |
1847 | { | |
1848 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
2b188cc1 | 1849 | struct io_ring_ctx *ctx = req->ctx; |
31b51510 JA |
1850 | struct mm_struct *cur_mm = NULL; |
1851 | struct async_list *async_list; | |
1852 | LIST_HEAD(req_list); | |
edafccee | 1853 | mm_segment_t old_fs; |
2b188cc1 JA |
1854 | int ret; |
1855 | ||
31b51510 JA |
1856 | async_list = io_async_list_from_sqe(ctx, req->submit.sqe); |
1857 | restart: | |
1858 | do { | |
1859 | struct sqe_submit *s = &req->submit; | |
1860 | const struct io_uring_sqe *sqe = s->sqe; | |
d0ee8791 | 1861 | unsigned int flags = req->flags; |
2b188cc1 | 1862 | |
8449eeda | 1863 | /* Ensure we clear previously set non-block flag */ |
31b51510 JA |
1864 | req->rw.ki_flags &= ~IOCB_NOWAIT; |
1865 | ||
1866 | ret = 0; | |
1867 | if (io_sqe_needs_user(sqe) && !cur_mm) { | |
1868 | if (!mmget_not_zero(ctx->sqo_mm)) { | |
1869 | ret = -EFAULT; | |
1870 | } else { | |
1871 | cur_mm = ctx->sqo_mm; | |
1872 | use_mm(cur_mm); | |
1873 | old_fs = get_fs(); | |
1874 | set_fs(USER_DS); | |
1875 | } | |
1876 | } | |
1877 | ||
1878 | if (!ret) { | |
1879 | s->has_user = cur_mm != NULL; | |
1880 | s->needs_lock = true; | |
1881 | do { | |
8358e3a8 | 1882 | ret = __io_submit_sqe(ctx, req, s, false); |
31b51510 JA |
1883 | /* |
1884 | * We can get EAGAIN for polled IO even though | |
1885 | * we're forcing a sync submission from here, | |
1886 | * since we can't wait for request slots on the | |
1887 | * block side. | |
1888 | */ | |
1889 | if (ret != -EAGAIN) | |
1890 | break; | |
1891 | cond_resched(); | |
1892 | } while (1); | |
1893 | } | |
817869d2 JA |
1894 | |
1895 | /* drop submission reference */ | |
1896 | io_put_req(req); | |
1897 | ||
31b51510 | 1898 | if (ret) { |
c71ffb67 | 1899 | io_cqring_add_event(ctx, sqe->user_data, ret); |
e65ef56d | 1900 | io_put_req(req); |
31b51510 JA |
1901 | } |
1902 | ||
1903 | /* async context always use a copy of the sqe */ | |
1904 | kfree(sqe); | |
1905 | ||
f7b76ac9 | 1906 | /* req from defer and link list needn't decrease async cnt */ |
d0ee8791 | 1907 | if (flags & (REQ_F_IO_DRAINED | REQ_F_LINK_DONE)) |
f7b76ac9 ZL |
1908 | goto out; |
1909 | ||
31b51510 JA |
1910 | if (!async_list) |
1911 | break; | |
1912 | if (!list_empty(&req_list)) { | |
1913 | req = list_first_entry(&req_list, struct io_kiocb, | |
1914 | list); | |
1915 | list_del(&req->list); | |
1916 | continue; | |
1917 | } | |
1918 | if (list_empty(&async_list->list)) | |
1919 | break; | |
1920 | ||
1921 | req = NULL; | |
1922 | spin_lock(&async_list->lock); | |
1923 | if (list_empty(&async_list->list)) { | |
1924 | spin_unlock(&async_list->lock); | |
1925 | break; | |
1926 | } | |
1927 | list_splice_init(&async_list->list, &req_list); | |
1928 | spin_unlock(&async_list->lock); | |
1929 | ||
1930 | req = list_first_entry(&req_list, struct io_kiocb, list); | |
1931 | list_del(&req->list); | |
1932 | } while (req); | |
edafccee JA |
1933 | |
1934 | /* | |
31b51510 JA |
1935 | * Rare case of racing with a submitter. If we find the count has |
1936 | * dropped to zero AND we have pending work items, then restart | |
1937 | * the processing. This is a tiny race window. | |
edafccee | 1938 | */ |
31b51510 JA |
1939 | if (async_list) { |
1940 | ret = atomic_dec_return(&async_list->cnt); | |
1941 | while (!ret && !list_empty(&async_list->list)) { | |
1942 | spin_lock(&async_list->lock); | |
1943 | atomic_inc(&async_list->cnt); | |
1944 | list_splice_init(&async_list->list, &req_list); | |
1945 | spin_unlock(&async_list->lock); | |
1946 | ||
1947 | if (!list_empty(&req_list)) { | |
1948 | req = list_first_entry(&req_list, | |
1949 | struct io_kiocb, list); | |
1950 | list_del(&req->list); | |
1951 | goto restart; | |
1952 | } | |
1953 | ret = atomic_dec_return(&async_list->cnt); | |
edafccee | 1954 | } |
edafccee | 1955 | } |
2b188cc1 | 1956 | |
f7b76ac9 | 1957 | out: |
31b51510 | 1958 | if (cur_mm) { |
edafccee | 1959 | set_fs(old_fs); |
31b51510 JA |
1960 | unuse_mm(cur_mm); |
1961 | mmput(cur_mm); | |
2b188cc1 | 1962 | } |
31b51510 | 1963 | } |
2b188cc1 | 1964 | |
31b51510 JA |
1965 | /* |
1966 | * See if we can piggy back onto previously submitted work, that is still | |
1967 | * running. We currently only allow this if the new request is sequential | |
1968 | * to the previous one we punted. | |
1969 | */ | |
1970 | static bool io_add_to_prev_work(struct async_list *list, struct io_kiocb *req) | |
1971 | { | |
1972 | bool ret = false; | |
1973 | ||
1974 | if (!list) | |
1975 | return false; | |
1976 | if (!(req->flags & REQ_F_SEQ_PREV)) | |
1977 | return false; | |
1978 | if (!atomic_read(&list->cnt)) | |
1979 | return false; | |
1980 | ||
1981 | ret = true; | |
1982 | spin_lock(&list->lock); | |
1983 | list_add_tail(&req->list, &list->list); | |
c0e48f9d ZL |
1984 | /* |
1985 | * Ensure we see a simultaneous modification from io_sq_wq_submit_work() | |
1986 | */ | |
1987 | smp_mb(); | |
31b51510 JA |
1988 | if (!atomic_read(&list->cnt)) { |
1989 | list_del_init(&req->list); | |
1990 | ret = false; | |
1991 | } | |
1992 | spin_unlock(&list->lock); | |
1993 | return ret; | |
2b188cc1 JA |
1994 | } |
1995 | ||
09bb8394 JA |
1996 | static bool io_op_needs_file(const struct io_uring_sqe *sqe) |
1997 | { | |
1998 | int op = READ_ONCE(sqe->opcode); | |
1999 | ||
2000 | switch (op) { | |
2001 | case IORING_OP_NOP: | |
2002 | case IORING_OP_POLL_REMOVE: | |
2003 | return false; | |
2004 | default: | |
2005 | return true; | |
2006 | } | |
2007 | } | |
2008 | ||
2009 | static int io_req_set_file(struct io_ring_ctx *ctx, const struct sqe_submit *s, | |
2010 | struct io_submit_state *state, struct io_kiocb *req) | |
2011 | { | |
2012 | unsigned flags; | |
2013 | int fd; | |
2014 | ||
2015 | flags = READ_ONCE(s->sqe->flags); | |
2016 | fd = READ_ONCE(s->sqe->fd); | |
2017 | ||
de0617e4 JA |
2018 | if (flags & IOSQE_IO_DRAIN) { |
2019 | req->flags |= REQ_F_IO_DRAIN; | |
8776f3fa | 2020 | req->sequence = s->sequence; |
de0617e4 JA |
2021 | } |
2022 | ||
60c112b0 | 2023 | if (!io_op_needs_file(s->sqe)) |
09bb8394 | 2024 | return 0; |
09bb8394 JA |
2025 | |
2026 | if (flags & IOSQE_FIXED_FILE) { | |
2027 | if (unlikely(!ctx->user_files || | |
2028 | (unsigned) fd >= ctx->nr_user_files)) | |
2029 | return -EBADF; | |
2030 | req->file = ctx->user_files[fd]; | |
2031 | req->flags |= REQ_F_FIXED_FILE; | |
2032 | } else { | |
2033 | if (s->needs_fixed_file) | |
2034 | return -EBADF; | |
2035 | req->file = io_file_get(state, fd); | |
2036 | if (unlikely(!req->file)) | |
2037 | return -EBADF; | |
2038 | } | |
2039 | ||
2040 | return 0; | |
2041 | } | |
2042 | ||
9e645e11 JA |
2043 | static int io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, |
2044 | struct sqe_submit *s) | |
2b188cc1 | 2045 | { |
e0c5c576 | 2046 | int ret; |
2b188cc1 | 2047 | |
a982eeb0 JL |
2048 | ret = io_req_defer(ctx, req, s->sqe); |
2049 | if (ret) { | |
2050 | if (ret != -EIOCBQUEUED) { | |
2051 | io_free_req(req); | |
2052 | io_cqring_add_event(ctx, s->sqe->user_data, ret); | |
2053 | } | |
2054 | return 0; | |
2055 | } | |
2056 | ||
8358e3a8 | 2057 | ret = __io_submit_sqe(ctx, req, s, true); |
8449eeda | 2058 | if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) { |
2b188cc1 JA |
2059 | struct io_uring_sqe *sqe_copy; |
2060 | ||
2061 | sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL); | |
2062 | if (sqe_copy) { | |
31b51510 JA |
2063 | struct async_list *list; |
2064 | ||
2b188cc1 JA |
2065 | memcpy(sqe_copy, s->sqe, sizeof(*sqe_copy)); |
2066 | s->sqe = sqe_copy; | |
2067 | ||
2068 | memcpy(&req->submit, s, sizeof(*s)); | |
31b51510 JA |
2069 | list = io_async_list_from_sqe(ctx, s->sqe); |
2070 | if (!io_add_to_prev_work(list, req)) { | |
2071 | if (list) | |
2072 | atomic_inc(&list->cnt); | |
2073 | INIT_WORK(&req->work, io_sq_wq_submit_work); | |
2074 | queue_work(ctx->sqo_wq, &req->work); | |
2075 | } | |
e65ef56d JA |
2076 | |
2077 | /* | |
2078 | * Queued up for async execution, worker will release | |
9e645e11 | 2079 | * submit reference when the iocb is actually submitted. |
e65ef56d JA |
2080 | */ |
2081 | return 0; | |
2b188cc1 JA |
2082 | } |
2083 | } | |
e65ef56d JA |
2084 | |
2085 | /* drop submission reference */ | |
2086 | io_put_req(req); | |
2087 | ||
2088 | /* and drop final reference, if we failed */ | |
9e645e11 JA |
2089 | if (ret) { |
2090 | io_cqring_add_event(ctx, req->user_data, ret); | |
2091 | if (req->flags & REQ_F_LINK) | |
2092 | req->flags |= REQ_F_FAIL_LINK; | |
e65ef56d | 2093 | io_put_req(req); |
9e645e11 | 2094 | } |
2b188cc1 JA |
2095 | |
2096 | return ret; | |
2097 | } | |
2098 | ||
9e645e11 JA |
2099 | #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK) |
2100 | ||
2101 | static void io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s, | |
2102 | struct io_submit_state *state, struct io_kiocb **link) | |
2103 | { | |
2104 | struct io_uring_sqe *sqe_copy; | |
2105 | struct io_kiocb *req; | |
2106 | int ret; | |
2107 | ||
2108 | /* enforce forwards compatibility on users */ | |
2109 | if (unlikely(s->sqe->flags & ~SQE_VALID_FLAGS)) { | |
2110 | ret = -EINVAL; | |
2111 | goto err; | |
2112 | } | |
2113 | ||
2114 | req = io_get_req(ctx, state); | |
2115 | if (unlikely(!req)) { | |
2116 | ret = -EAGAIN; | |
2117 | goto err; | |
2118 | } | |
2119 | ||
2120 | ret = io_req_set_file(ctx, s, state, req); | |
2121 | if (unlikely(ret)) { | |
2122 | err_req: | |
2123 | io_free_req(req); | |
2124 | err: | |
2125 | io_cqring_add_event(ctx, s->sqe->user_data, ret); | |
2126 | return; | |
2127 | } | |
2128 | ||
9e645e11 JA |
2129 | /* |
2130 | * If we already have a head request, queue this one for async | |
2131 | * submittal once the head completes. If we don't have a head but | |
2132 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
2133 | * submitted sync once the chain is complete. If none of those | |
2134 | * conditions are true (normal request), then just queue it. | |
2135 | */ | |
2136 | if (*link) { | |
2137 | struct io_kiocb *prev = *link; | |
2138 | ||
2139 | sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL); | |
2140 | if (!sqe_copy) { | |
2141 | ret = -EAGAIN; | |
2142 | goto err_req; | |
2143 | } | |
2144 | ||
2145 | s->sqe = sqe_copy; | |
2146 | memcpy(&req->submit, s, sizeof(*s)); | |
2147 | list_add_tail(&req->list, &prev->link_list); | |
2148 | } else if (s->sqe->flags & IOSQE_IO_LINK) { | |
2149 | req->flags |= REQ_F_LINK; | |
2150 | ||
2151 | memcpy(&req->submit, s, sizeof(*s)); | |
2152 | INIT_LIST_HEAD(&req->link_list); | |
2153 | *link = req; | |
2154 | } else { | |
2155 | io_queue_sqe(ctx, req, s); | |
2156 | } | |
2157 | } | |
2158 | ||
9a56a232 JA |
2159 | /* |
2160 | * Batched submission is done, ensure local IO is flushed out. | |
2161 | */ | |
2162 | static void io_submit_state_end(struct io_submit_state *state) | |
2163 | { | |
2164 | blk_finish_plug(&state->plug); | |
3d6770fb | 2165 | io_file_put(state); |
2579f913 JA |
2166 | if (state->free_reqs) |
2167 | kmem_cache_free_bulk(req_cachep, state->free_reqs, | |
2168 | &state->reqs[state->cur_req]); | |
9a56a232 JA |
2169 | } |
2170 | ||
2171 | /* | |
2172 | * Start submission side cache. | |
2173 | */ | |
2174 | static void io_submit_state_start(struct io_submit_state *state, | |
2175 | struct io_ring_ctx *ctx, unsigned max_ios) | |
2176 | { | |
2177 | blk_start_plug(&state->plug); | |
2579f913 | 2178 | state->free_reqs = 0; |
9a56a232 JA |
2179 | state->file = NULL; |
2180 | state->ios_left = max_ios; | |
2181 | } | |
2182 | ||
2b188cc1 JA |
2183 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
2184 | { | |
75b28aff | 2185 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 2186 | |
75b28aff | 2187 | if (ctx->cached_sq_head != READ_ONCE(rings->sq.head)) { |
2b188cc1 JA |
2188 | /* |
2189 | * Ensure any loads from the SQEs are done at this point, | |
2190 | * since once we write the new head, the application could | |
2191 | * write new data to them. | |
2192 | */ | |
75b28aff | 2193 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); |
2b188cc1 JA |
2194 | } |
2195 | } | |
2196 | ||
2b188cc1 JA |
2197 | /* |
2198 | * Fetch an sqe, if one is available. Note that s->sqe will point to memory | |
2199 | * that is mapped by userspace. This means that care needs to be taken to | |
2200 | * ensure that reads are stable, as we cannot rely on userspace always | |
2201 | * being a good citizen. If members of the sqe are validated and then later | |
2202 | * used, it's important that those reads are done through READ_ONCE() to | |
2203 | * prevent a re-load down the line. | |
2204 | */ | |
2205 | static bool io_get_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s) | |
2206 | { | |
75b28aff HV |
2207 | struct io_rings *rings = ctx->rings; |
2208 | u32 *sq_array = ctx->sq_array; | |
2b188cc1 JA |
2209 | unsigned head; |
2210 | ||
2211 | /* | |
2212 | * The cached sq head (or cq tail) serves two purposes: | |
2213 | * | |
2214 | * 1) allows us to batch the cost of updating the user visible | |
2215 | * head updates. | |
2216 | * 2) allows the kernel side to track the head on its own, even | |
2217 | * though the application is the one updating it. | |
2218 | */ | |
2219 | head = ctx->cached_sq_head; | |
e523a29c | 2220 | /* make sure SQ entry isn't read before tail */ |
75b28aff | 2221 | if (head == smp_load_acquire(&rings->sq.tail)) |
2b188cc1 JA |
2222 | return false; |
2223 | ||
75b28aff | 2224 | head = READ_ONCE(sq_array[head & ctx->sq_mask]); |
2b188cc1 JA |
2225 | if (head < ctx->sq_entries) { |
2226 | s->index = head; | |
2227 | s->sqe = &ctx->sq_sqes[head]; | |
8776f3fa | 2228 | s->sequence = ctx->cached_sq_head; |
2b188cc1 JA |
2229 | ctx->cached_sq_head++; |
2230 | return true; | |
2231 | } | |
2232 | ||
2233 | /* drop invalid entries */ | |
2234 | ctx->cached_sq_head++; | |
75b28aff | 2235 | rings->sq_dropped++; |
2b188cc1 JA |
2236 | return false; |
2237 | } | |
2238 | ||
6c271ce2 JA |
2239 | static int io_submit_sqes(struct io_ring_ctx *ctx, struct sqe_submit *sqes, |
2240 | unsigned int nr, bool has_user, bool mm_fault) | |
2241 | { | |
2242 | struct io_submit_state state, *statep = NULL; | |
9e645e11 JA |
2243 | struct io_kiocb *link = NULL; |
2244 | bool prev_was_link = false; | |
2245 | int i, submitted = 0; | |
6c271ce2 JA |
2246 | |
2247 | if (nr > IO_PLUG_THRESHOLD) { | |
2248 | io_submit_state_start(&state, ctx, nr); | |
2249 | statep = &state; | |
2250 | } | |
2251 | ||
2252 | for (i = 0; i < nr; i++) { | |
9e645e11 JA |
2253 | /* |
2254 | * If previous wasn't linked and we have a linked command, | |
2255 | * that's the end of the chain. Submit the previous link. | |
2256 | */ | |
2257 | if (!prev_was_link && link) { | |
2258 | io_queue_sqe(ctx, link, &link->submit); | |
2259 | link = NULL; | |
2260 | } | |
2261 | prev_was_link = (sqes[i].sqe->flags & IOSQE_IO_LINK) != 0; | |
2262 | ||
6c271ce2 | 2263 | if (unlikely(mm_fault)) { |
9e645e11 JA |
2264 | io_cqring_add_event(ctx, sqes[i].sqe->user_data, |
2265 | -EFAULT); | |
6c271ce2 JA |
2266 | } else { |
2267 | sqes[i].has_user = has_user; | |
2268 | sqes[i].needs_lock = true; | |
2269 | sqes[i].needs_fixed_file = true; | |
9e645e11 | 2270 | io_submit_sqe(ctx, &sqes[i], statep, &link); |
6c271ce2 | 2271 | submitted++; |
6c271ce2 | 2272 | } |
6c271ce2 JA |
2273 | } |
2274 | ||
9e645e11 JA |
2275 | if (link) |
2276 | io_queue_sqe(ctx, link, &link->submit); | |
6c271ce2 JA |
2277 | if (statep) |
2278 | io_submit_state_end(&state); | |
2279 | ||
2280 | return submitted; | |
2281 | } | |
2282 | ||
2283 | static int io_sq_thread(void *data) | |
2284 | { | |
2285 | struct sqe_submit sqes[IO_IOPOLL_BATCH]; | |
2286 | struct io_ring_ctx *ctx = data; | |
2287 | struct mm_struct *cur_mm = NULL; | |
2288 | mm_segment_t old_fs; | |
2289 | DEFINE_WAIT(wait); | |
2290 | unsigned inflight; | |
2291 | unsigned long timeout; | |
2292 | ||
a4c0b3de JL |
2293 | complete(&ctx->sqo_thread_started); |
2294 | ||
6c271ce2 JA |
2295 | old_fs = get_fs(); |
2296 | set_fs(USER_DS); | |
2297 | ||
2298 | timeout = inflight = 0; | |
2bbcd6d3 | 2299 | while (!kthread_should_park()) { |
6c271ce2 JA |
2300 | bool all_fixed, mm_fault = false; |
2301 | int i; | |
2302 | ||
2303 | if (inflight) { | |
2304 | unsigned nr_events = 0; | |
2305 | ||
2306 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
6c271ce2 | 2307 | io_iopoll_check(ctx, &nr_events, 0); |
6c271ce2 JA |
2308 | } else { |
2309 | /* | |
2310 | * Normal IO, just pretend everything completed. | |
2311 | * We don't have to poll completions for that. | |
2312 | */ | |
2313 | nr_events = inflight; | |
2314 | } | |
2315 | ||
2316 | inflight -= nr_events; | |
2317 | if (!inflight) | |
2318 | timeout = jiffies + ctx->sq_thread_idle; | |
2319 | } | |
2320 | ||
2321 | if (!io_get_sqring(ctx, &sqes[0])) { | |
2322 | /* | |
2323 | * We're polling. If we're within the defined idle | |
2324 | * period, then let us spin without work before going | |
2325 | * to sleep. | |
2326 | */ | |
2327 | if (inflight || !time_after(jiffies, timeout)) { | |
2328 | cpu_relax(); | |
2329 | continue; | |
2330 | } | |
2331 | ||
2332 | /* | |
2333 | * Drop cur_mm before scheduling, we can't hold it for | |
2334 | * long periods (or over schedule()). Do this before | |
2335 | * adding ourselves to the waitqueue, as the unuse/drop | |
2336 | * may sleep. | |
2337 | */ | |
2338 | if (cur_mm) { | |
2339 | unuse_mm(cur_mm); | |
2340 | mmput(cur_mm); | |
2341 | cur_mm = NULL; | |
2342 | } | |
2343 | ||
2344 | prepare_to_wait(&ctx->sqo_wait, &wait, | |
2345 | TASK_INTERRUPTIBLE); | |
2346 | ||
2347 | /* Tell userspace we may need a wakeup call */ | |
75b28aff | 2348 | ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP; |
0d7bae69 SB |
2349 | /* make sure to read SQ tail after writing flags */ |
2350 | smp_mb(); | |
6c271ce2 JA |
2351 | |
2352 | if (!io_get_sqring(ctx, &sqes[0])) { | |
2bbcd6d3 | 2353 | if (kthread_should_park()) { |
6c271ce2 JA |
2354 | finish_wait(&ctx->sqo_wait, &wait); |
2355 | break; | |
2356 | } | |
2357 | if (signal_pending(current)) | |
2358 | flush_signals(current); | |
2359 | schedule(); | |
2360 | finish_wait(&ctx->sqo_wait, &wait); | |
2361 | ||
75b28aff | 2362 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
2363 | continue; |
2364 | } | |
2365 | finish_wait(&ctx->sqo_wait, &wait); | |
2366 | ||
75b28aff | 2367 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
2368 | } |
2369 | ||
2370 | i = 0; | |
2371 | all_fixed = true; | |
2372 | do { | |
2373 | if (all_fixed && io_sqe_needs_user(sqes[i].sqe)) | |
2374 | all_fixed = false; | |
2375 | ||
2376 | i++; | |
2377 | if (i == ARRAY_SIZE(sqes)) | |
2378 | break; | |
2379 | } while (io_get_sqring(ctx, &sqes[i])); | |
2380 | ||
2381 | /* Unless all new commands are FIXED regions, grab mm */ | |
2382 | if (!all_fixed && !cur_mm) { | |
2383 | mm_fault = !mmget_not_zero(ctx->sqo_mm); | |
2384 | if (!mm_fault) { | |
2385 | use_mm(ctx->sqo_mm); | |
2386 | cur_mm = ctx->sqo_mm; | |
2387 | } | |
2388 | } | |
2389 | ||
2390 | inflight += io_submit_sqes(ctx, sqes, i, cur_mm != NULL, | |
2391 | mm_fault); | |
2392 | ||
2393 | /* Commit SQ ring head once we've consumed all SQEs */ | |
2394 | io_commit_sqring(ctx); | |
2395 | } | |
2396 | ||
2397 | set_fs(old_fs); | |
2398 | if (cur_mm) { | |
2399 | unuse_mm(cur_mm); | |
2400 | mmput(cur_mm); | |
2401 | } | |
06058632 | 2402 | |
2bbcd6d3 | 2403 | kthread_parkme(); |
06058632 | 2404 | |
6c271ce2 JA |
2405 | return 0; |
2406 | } | |
2407 | ||
2b188cc1 JA |
2408 | static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit) |
2409 | { | |
9a56a232 | 2410 | struct io_submit_state state, *statep = NULL; |
9e645e11 JA |
2411 | struct io_kiocb *link = NULL; |
2412 | bool prev_was_link = false; | |
5c8b0b54 | 2413 | int i, submit = 0; |
2b188cc1 | 2414 | |
9a56a232 JA |
2415 | if (to_submit > IO_PLUG_THRESHOLD) { |
2416 | io_submit_state_start(&state, ctx, to_submit); | |
2417 | statep = &state; | |
2418 | } | |
2b188cc1 JA |
2419 | |
2420 | for (i = 0; i < to_submit; i++) { | |
2421 | struct sqe_submit s; | |
2422 | ||
2423 | if (!io_get_sqring(ctx, &s)) | |
2424 | break; | |
2425 | ||
9e645e11 JA |
2426 | /* |
2427 | * If previous wasn't linked and we have a linked command, | |
2428 | * that's the end of the chain. Submit the previous link. | |
2429 | */ | |
2430 | if (!prev_was_link && link) { | |
2431 | io_queue_sqe(ctx, link, &link->submit); | |
2432 | link = NULL; | |
2433 | } | |
2434 | prev_was_link = (s.sqe->flags & IOSQE_IO_LINK) != 0; | |
2435 | ||
2b188cc1 | 2436 | s.has_user = true; |
def596e9 | 2437 | s.needs_lock = false; |
6c271ce2 | 2438 | s.needs_fixed_file = false; |
5c8b0b54 | 2439 | submit++; |
9e645e11 | 2440 | io_submit_sqe(ctx, &s, statep, &link); |
2b188cc1 JA |
2441 | } |
2442 | io_commit_sqring(ctx); | |
2443 | ||
9e645e11 JA |
2444 | if (link) |
2445 | io_queue_sqe(ctx, link, &link->submit); | |
9a56a232 JA |
2446 | if (statep) |
2447 | io_submit_state_end(statep); | |
2b188cc1 | 2448 | |
5c8b0b54 | 2449 | return submit; |
2b188cc1 JA |
2450 | } |
2451 | ||
2b188cc1 JA |
2452 | /* |
2453 | * Wait until events become available, if we don't already have some. The | |
2454 | * application must reap them itself, as they reside on the shared cq ring. | |
2455 | */ | |
2456 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
2457 | const sigset_t __user *sig, size_t sigsz) | |
2458 | { | |
75b28aff | 2459 | struct io_rings *rings = ctx->rings; |
2b188cc1 JA |
2460 | int ret; |
2461 | ||
75b28aff | 2462 | if (io_cqring_events(rings) >= min_events) |
2b188cc1 JA |
2463 | return 0; |
2464 | ||
2465 | if (sig) { | |
9e75ad5d AB |
2466 | #ifdef CONFIG_COMPAT |
2467 | if (in_compat_syscall()) | |
2468 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 2469 | sigsz); |
9e75ad5d AB |
2470 | else |
2471 | #endif | |
b772434b | 2472 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 2473 | |
2b188cc1 JA |
2474 | if (ret) |
2475 | return ret; | |
2476 | } | |
2477 | ||
75b28aff | 2478 | ret = wait_event_interruptible(ctx->wait, io_cqring_events(rings) >= min_events); |
b772434b | 2479 | restore_saved_sigmask_unless(ret == -ERESTARTSYS); |
97abc889 ON |
2480 | if (ret == -ERESTARTSYS) |
2481 | ret = -EINTR; | |
2b188cc1 | 2482 | |
75b28aff | 2483 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
2484 | } |
2485 | ||
6b06314c JA |
2486 | static void __io_sqe_files_unregister(struct io_ring_ctx *ctx) |
2487 | { | |
2488 | #if defined(CONFIG_UNIX) | |
2489 | if (ctx->ring_sock) { | |
2490 | struct sock *sock = ctx->ring_sock->sk; | |
2491 | struct sk_buff *skb; | |
2492 | ||
2493 | while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) | |
2494 | kfree_skb(skb); | |
2495 | } | |
2496 | #else | |
2497 | int i; | |
2498 | ||
2499 | for (i = 0; i < ctx->nr_user_files; i++) | |
2500 | fput(ctx->user_files[i]); | |
2501 | #endif | |
2502 | } | |
2503 | ||
2504 | static int io_sqe_files_unregister(struct io_ring_ctx *ctx) | |
2505 | { | |
2506 | if (!ctx->user_files) | |
2507 | return -ENXIO; | |
2508 | ||
2509 | __io_sqe_files_unregister(ctx); | |
2510 | kfree(ctx->user_files); | |
2511 | ctx->user_files = NULL; | |
2512 | ctx->nr_user_files = 0; | |
2513 | return 0; | |
2514 | } | |
2515 | ||
6c271ce2 JA |
2516 | static void io_sq_thread_stop(struct io_ring_ctx *ctx) |
2517 | { | |
2518 | if (ctx->sqo_thread) { | |
a4c0b3de | 2519 | wait_for_completion(&ctx->sqo_thread_started); |
2bbcd6d3 RP |
2520 | /* |
2521 | * The park is a bit of a work-around, without it we get | |
2522 | * warning spews on shutdown with SQPOLL set and affinity | |
2523 | * set to a single CPU. | |
2524 | */ | |
06058632 | 2525 | kthread_park(ctx->sqo_thread); |
6c271ce2 JA |
2526 | kthread_stop(ctx->sqo_thread); |
2527 | ctx->sqo_thread = NULL; | |
2528 | } | |
2529 | } | |
2530 | ||
6b06314c JA |
2531 | static void io_finish_async(struct io_ring_ctx *ctx) |
2532 | { | |
6c271ce2 JA |
2533 | io_sq_thread_stop(ctx); |
2534 | ||
6b06314c JA |
2535 | if (ctx->sqo_wq) { |
2536 | destroy_workqueue(ctx->sqo_wq); | |
2537 | ctx->sqo_wq = NULL; | |
2538 | } | |
2539 | } | |
2540 | ||
2541 | #if defined(CONFIG_UNIX) | |
2542 | static void io_destruct_skb(struct sk_buff *skb) | |
2543 | { | |
2544 | struct io_ring_ctx *ctx = skb->sk->sk_user_data; | |
2545 | ||
2546 | io_finish_async(ctx); | |
2547 | unix_destruct_scm(skb); | |
2548 | } | |
2549 | ||
2550 | /* | |
2551 | * Ensure the UNIX gc is aware of our file set, so we are certain that | |
2552 | * the io_uring can be safely unregistered on process exit, even if we have | |
2553 | * loops in the file referencing. | |
2554 | */ | |
2555 | static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset) | |
2556 | { | |
2557 | struct sock *sk = ctx->ring_sock->sk; | |
2558 | struct scm_fp_list *fpl; | |
2559 | struct sk_buff *skb; | |
2560 | int i; | |
2561 | ||
2562 | if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) { | |
2563 | unsigned long inflight = ctx->user->unix_inflight + nr; | |
2564 | ||
2565 | if (inflight > task_rlimit(current, RLIMIT_NOFILE)) | |
2566 | return -EMFILE; | |
2567 | } | |
2568 | ||
2569 | fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); | |
2570 | if (!fpl) | |
2571 | return -ENOMEM; | |
2572 | ||
2573 | skb = alloc_skb(0, GFP_KERNEL); | |
2574 | if (!skb) { | |
2575 | kfree(fpl); | |
2576 | return -ENOMEM; | |
2577 | } | |
2578 | ||
2579 | skb->sk = sk; | |
2580 | skb->destructor = io_destruct_skb; | |
2581 | ||
2582 | fpl->user = get_uid(ctx->user); | |
2583 | for (i = 0; i < nr; i++) { | |
2584 | fpl->fp[i] = get_file(ctx->user_files[i + offset]); | |
2585 | unix_inflight(fpl->user, fpl->fp[i]); | |
2586 | } | |
2587 | ||
2588 | fpl->max = fpl->count = nr; | |
2589 | UNIXCB(skb).fp = fpl; | |
2590 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); | |
2591 | skb_queue_head(&sk->sk_receive_queue, skb); | |
2592 | ||
2593 | for (i = 0; i < nr; i++) | |
2594 | fput(fpl->fp[i]); | |
2595 | ||
2596 | return 0; | |
2597 | } | |
2598 | ||
2599 | /* | |
2600 | * If UNIX sockets are enabled, fd passing can cause a reference cycle which | |
2601 | * causes regular reference counting to break down. We rely on the UNIX | |
2602 | * garbage collection to take care of this problem for us. | |
2603 | */ | |
2604 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
2605 | { | |
2606 | unsigned left, total; | |
2607 | int ret = 0; | |
2608 | ||
2609 | total = 0; | |
2610 | left = ctx->nr_user_files; | |
2611 | while (left) { | |
2612 | unsigned this_files = min_t(unsigned, left, SCM_MAX_FD); | |
6b06314c JA |
2613 | |
2614 | ret = __io_sqe_files_scm(ctx, this_files, total); | |
2615 | if (ret) | |
2616 | break; | |
2617 | left -= this_files; | |
2618 | total += this_files; | |
2619 | } | |
2620 | ||
2621 | if (!ret) | |
2622 | return 0; | |
2623 | ||
2624 | while (total < ctx->nr_user_files) { | |
2625 | fput(ctx->user_files[total]); | |
2626 | total++; | |
2627 | } | |
2628 | ||
2629 | return ret; | |
2630 | } | |
2631 | #else | |
2632 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
2633 | { | |
2634 | return 0; | |
2635 | } | |
2636 | #endif | |
2637 | ||
2638 | static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, | |
2639 | unsigned nr_args) | |
2640 | { | |
2641 | __s32 __user *fds = (__s32 __user *) arg; | |
2642 | int fd, ret = 0; | |
2643 | unsigned i; | |
2644 | ||
2645 | if (ctx->user_files) | |
2646 | return -EBUSY; | |
2647 | if (!nr_args) | |
2648 | return -EINVAL; | |
2649 | if (nr_args > IORING_MAX_FIXED_FILES) | |
2650 | return -EMFILE; | |
2651 | ||
2652 | ctx->user_files = kcalloc(nr_args, sizeof(struct file *), GFP_KERNEL); | |
2653 | if (!ctx->user_files) | |
2654 | return -ENOMEM; | |
2655 | ||
2656 | for (i = 0; i < nr_args; i++) { | |
2657 | ret = -EFAULT; | |
2658 | if (copy_from_user(&fd, &fds[i], sizeof(fd))) | |
2659 | break; | |
2660 | ||
2661 | ctx->user_files[i] = fget(fd); | |
2662 | ||
2663 | ret = -EBADF; | |
2664 | if (!ctx->user_files[i]) | |
2665 | break; | |
2666 | /* | |
2667 | * Don't allow io_uring instances to be registered. If UNIX | |
2668 | * isn't enabled, then this causes a reference cycle and this | |
2669 | * instance can never get freed. If UNIX is enabled we'll | |
2670 | * handle it just fine, but there's still no point in allowing | |
2671 | * a ring fd as it doesn't support regular read/write anyway. | |
2672 | */ | |
2673 | if (ctx->user_files[i]->f_op == &io_uring_fops) { | |
2674 | fput(ctx->user_files[i]); | |
2675 | break; | |
2676 | } | |
2677 | ctx->nr_user_files++; | |
2678 | ret = 0; | |
2679 | } | |
2680 | ||
2681 | if (ret) { | |
2682 | for (i = 0; i < ctx->nr_user_files; i++) | |
2683 | fput(ctx->user_files[i]); | |
2684 | ||
2685 | kfree(ctx->user_files); | |
25adf50f | 2686 | ctx->user_files = NULL; |
6b06314c JA |
2687 | ctx->nr_user_files = 0; |
2688 | return ret; | |
2689 | } | |
2690 | ||
2691 | ret = io_sqe_files_scm(ctx); | |
2692 | if (ret) | |
2693 | io_sqe_files_unregister(ctx); | |
2694 | ||
2695 | return ret; | |
2696 | } | |
2697 | ||
6c271ce2 JA |
2698 | static int io_sq_offload_start(struct io_ring_ctx *ctx, |
2699 | struct io_uring_params *p) | |
2b188cc1 JA |
2700 | { |
2701 | int ret; | |
2702 | ||
6c271ce2 | 2703 | init_waitqueue_head(&ctx->sqo_wait); |
2b188cc1 JA |
2704 | mmgrab(current->mm); |
2705 | ctx->sqo_mm = current->mm; | |
2706 | ||
6c271ce2 | 2707 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
3ec482d1 JA |
2708 | ret = -EPERM; |
2709 | if (!capable(CAP_SYS_ADMIN)) | |
2710 | goto err; | |
2711 | ||
917257da JA |
2712 | ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle); |
2713 | if (!ctx->sq_thread_idle) | |
2714 | ctx->sq_thread_idle = HZ; | |
2715 | ||
6c271ce2 | 2716 | if (p->flags & IORING_SETUP_SQ_AFF) { |
44a9bd18 | 2717 | int cpu = p->sq_thread_cpu; |
6c271ce2 | 2718 | |
917257da | 2719 | ret = -EINVAL; |
44a9bd18 JA |
2720 | if (cpu >= nr_cpu_ids) |
2721 | goto err; | |
7889f44d | 2722 | if (!cpu_online(cpu)) |
917257da JA |
2723 | goto err; |
2724 | ||
6c271ce2 JA |
2725 | ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread, |
2726 | ctx, cpu, | |
2727 | "io_uring-sq"); | |
2728 | } else { | |
2729 | ctx->sqo_thread = kthread_create(io_sq_thread, ctx, | |
2730 | "io_uring-sq"); | |
2731 | } | |
2732 | if (IS_ERR(ctx->sqo_thread)) { | |
2733 | ret = PTR_ERR(ctx->sqo_thread); | |
2734 | ctx->sqo_thread = NULL; | |
2735 | goto err; | |
2736 | } | |
2737 | wake_up_process(ctx->sqo_thread); | |
2738 | } else if (p->flags & IORING_SETUP_SQ_AFF) { | |
2739 | /* Can't have SQ_AFF without SQPOLL */ | |
2740 | ret = -EINVAL; | |
2741 | goto err; | |
2742 | } | |
2743 | ||
2b188cc1 JA |
2744 | /* Do QD, or 2 * CPUS, whatever is smallest */ |
2745 | ctx->sqo_wq = alloc_workqueue("io_ring-wq", WQ_UNBOUND | WQ_FREEZABLE, | |
2746 | min(ctx->sq_entries - 1, 2 * num_online_cpus())); | |
2747 | if (!ctx->sqo_wq) { | |
2748 | ret = -ENOMEM; | |
2749 | goto err; | |
2750 | } | |
2751 | ||
2752 | return 0; | |
2753 | err: | |
6c271ce2 | 2754 | io_sq_thread_stop(ctx); |
2b188cc1 JA |
2755 | mmdrop(ctx->sqo_mm); |
2756 | ctx->sqo_mm = NULL; | |
2757 | return ret; | |
2758 | } | |
2759 | ||
2760 | static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages) | |
2761 | { | |
2762 | atomic_long_sub(nr_pages, &user->locked_vm); | |
2763 | } | |
2764 | ||
2765 | static int io_account_mem(struct user_struct *user, unsigned long nr_pages) | |
2766 | { | |
2767 | unsigned long page_limit, cur_pages, new_pages; | |
2768 | ||
2769 | /* Don't allow more pages than we can safely lock */ | |
2770 | page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
2771 | ||
2772 | do { | |
2773 | cur_pages = atomic_long_read(&user->locked_vm); | |
2774 | new_pages = cur_pages + nr_pages; | |
2775 | if (new_pages > page_limit) | |
2776 | return -ENOMEM; | |
2777 | } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages, | |
2778 | new_pages) != cur_pages); | |
2779 | ||
2780 | return 0; | |
2781 | } | |
2782 | ||
2783 | static void io_mem_free(void *ptr) | |
2784 | { | |
52e04ef4 MR |
2785 | struct page *page; |
2786 | ||
2787 | if (!ptr) | |
2788 | return; | |
2b188cc1 | 2789 | |
52e04ef4 | 2790 | page = virt_to_head_page(ptr); |
2b188cc1 JA |
2791 | if (put_page_testzero(page)) |
2792 | free_compound_page(page); | |
2793 | } | |
2794 | ||
2795 | static void *io_mem_alloc(size_t size) | |
2796 | { | |
2797 | gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP | | |
2798 | __GFP_NORETRY; | |
2799 | ||
2800 | return (void *) __get_free_pages(gfp_flags, get_order(size)); | |
2801 | } | |
2802 | ||
75b28aff HV |
2803 | static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries, |
2804 | size_t *sq_offset) | |
2805 | { | |
2806 | struct io_rings *rings; | |
2807 | size_t off, sq_array_size; | |
2808 | ||
2809 | off = struct_size(rings, cqes, cq_entries); | |
2810 | if (off == SIZE_MAX) | |
2811 | return SIZE_MAX; | |
2812 | ||
2813 | #ifdef CONFIG_SMP | |
2814 | off = ALIGN(off, SMP_CACHE_BYTES); | |
2815 | if (off == 0) | |
2816 | return SIZE_MAX; | |
2817 | #endif | |
2818 | ||
2819 | sq_array_size = array_size(sizeof(u32), sq_entries); | |
2820 | if (sq_array_size == SIZE_MAX) | |
2821 | return SIZE_MAX; | |
2822 | ||
2823 | if (check_add_overflow(off, sq_array_size, &off)) | |
2824 | return SIZE_MAX; | |
2825 | ||
2826 | if (sq_offset) | |
2827 | *sq_offset = off; | |
2828 | ||
2829 | return off; | |
2830 | } | |
2831 | ||
2b188cc1 JA |
2832 | static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries) |
2833 | { | |
75b28aff | 2834 | size_t pages; |
2b188cc1 | 2835 | |
75b28aff HV |
2836 | pages = (size_t)1 << get_order( |
2837 | rings_size(sq_entries, cq_entries, NULL)); | |
2838 | pages += (size_t)1 << get_order( | |
2839 | array_size(sizeof(struct io_uring_sqe), sq_entries)); | |
2b188cc1 | 2840 | |
75b28aff | 2841 | return pages; |
2b188cc1 JA |
2842 | } |
2843 | ||
edafccee JA |
2844 | static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx) |
2845 | { | |
2846 | int i, j; | |
2847 | ||
2848 | if (!ctx->user_bufs) | |
2849 | return -ENXIO; | |
2850 | ||
2851 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
2852 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
2853 | ||
2854 | for (j = 0; j < imu->nr_bvecs; j++) | |
27c4d3a3 | 2855 | put_user_page(imu->bvec[j].bv_page); |
edafccee JA |
2856 | |
2857 | if (ctx->account_mem) | |
2858 | io_unaccount_mem(ctx->user, imu->nr_bvecs); | |
d4ef6475 | 2859 | kvfree(imu->bvec); |
edafccee JA |
2860 | imu->nr_bvecs = 0; |
2861 | } | |
2862 | ||
2863 | kfree(ctx->user_bufs); | |
2864 | ctx->user_bufs = NULL; | |
2865 | ctx->nr_user_bufs = 0; | |
2866 | return 0; | |
2867 | } | |
2868 | ||
2869 | static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, | |
2870 | void __user *arg, unsigned index) | |
2871 | { | |
2872 | struct iovec __user *src; | |
2873 | ||
2874 | #ifdef CONFIG_COMPAT | |
2875 | if (ctx->compat) { | |
2876 | struct compat_iovec __user *ciovs; | |
2877 | struct compat_iovec ciov; | |
2878 | ||
2879 | ciovs = (struct compat_iovec __user *) arg; | |
2880 | if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) | |
2881 | return -EFAULT; | |
2882 | ||
2883 | dst->iov_base = (void __user *) (unsigned long) ciov.iov_base; | |
2884 | dst->iov_len = ciov.iov_len; | |
2885 | return 0; | |
2886 | } | |
2887 | #endif | |
2888 | src = (struct iovec __user *) arg; | |
2889 | if (copy_from_user(dst, &src[index], sizeof(*dst))) | |
2890 | return -EFAULT; | |
2891 | return 0; | |
2892 | } | |
2893 | ||
2894 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg, | |
2895 | unsigned nr_args) | |
2896 | { | |
2897 | struct vm_area_struct **vmas = NULL; | |
2898 | struct page **pages = NULL; | |
2899 | int i, j, got_pages = 0; | |
2900 | int ret = -EINVAL; | |
2901 | ||
2902 | if (ctx->user_bufs) | |
2903 | return -EBUSY; | |
2904 | if (!nr_args || nr_args > UIO_MAXIOV) | |
2905 | return -EINVAL; | |
2906 | ||
2907 | ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf), | |
2908 | GFP_KERNEL); | |
2909 | if (!ctx->user_bufs) | |
2910 | return -ENOMEM; | |
2911 | ||
2912 | for (i = 0; i < nr_args; i++) { | |
2913 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
2914 | unsigned long off, start, end, ubuf; | |
2915 | int pret, nr_pages; | |
2916 | struct iovec iov; | |
2917 | size_t size; | |
2918 | ||
2919 | ret = io_copy_iov(ctx, &iov, arg, i); | |
2920 | if (ret) | |
a278682d | 2921 | goto err; |
edafccee JA |
2922 | |
2923 | /* | |
2924 | * Don't impose further limits on the size and buffer | |
2925 | * constraints here, we'll -EINVAL later when IO is | |
2926 | * submitted if they are wrong. | |
2927 | */ | |
2928 | ret = -EFAULT; | |
2929 | if (!iov.iov_base || !iov.iov_len) | |
2930 | goto err; | |
2931 | ||
2932 | /* arbitrary limit, but we need something */ | |
2933 | if (iov.iov_len > SZ_1G) | |
2934 | goto err; | |
2935 | ||
2936 | ubuf = (unsigned long) iov.iov_base; | |
2937 | end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
2938 | start = ubuf >> PAGE_SHIFT; | |
2939 | nr_pages = end - start; | |
2940 | ||
2941 | if (ctx->account_mem) { | |
2942 | ret = io_account_mem(ctx->user, nr_pages); | |
2943 | if (ret) | |
2944 | goto err; | |
2945 | } | |
2946 | ||
2947 | ret = 0; | |
2948 | if (!pages || nr_pages > got_pages) { | |
2949 | kfree(vmas); | |
2950 | kfree(pages); | |
d4ef6475 | 2951 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), |
edafccee | 2952 | GFP_KERNEL); |
d4ef6475 | 2953 | vmas = kvmalloc_array(nr_pages, |
edafccee JA |
2954 | sizeof(struct vm_area_struct *), |
2955 | GFP_KERNEL); | |
2956 | if (!pages || !vmas) { | |
2957 | ret = -ENOMEM; | |
2958 | if (ctx->account_mem) | |
2959 | io_unaccount_mem(ctx->user, nr_pages); | |
2960 | goto err; | |
2961 | } | |
2962 | got_pages = nr_pages; | |
2963 | } | |
2964 | ||
d4ef6475 | 2965 | imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec), |
edafccee JA |
2966 | GFP_KERNEL); |
2967 | ret = -ENOMEM; | |
2968 | if (!imu->bvec) { | |
2969 | if (ctx->account_mem) | |
2970 | io_unaccount_mem(ctx->user, nr_pages); | |
2971 | goto err; | |
2972 | } | |
2973 | ||
2974 | ret = 0; | |
2975 | down_read(¤t->mm->mmap_sem); | |
932f4a63 IW |
2976 | pret = get_user_pages(ubuf, nr_pages, |
2977 | FOLL_WRITE | FOLL_LONGTERM, | |
2978 | pages, vmas); | |
edafccee JA |
2979 | if (pret == nr_pages) { |
2980 | /* don't support file backed memory */ | |
2981 | for (j = 0; j < nr_pages; j++) { | |
2982 | struct vm_area_struct *vma = vmas[j]; | |
2983 | ||
2984 | if (vma->vm_file && | |
2985 | !is_file_hugepages(vma->vm_file)) { | |
2986 | ret = -EOPNOTSUPP; | |
2987 | break; | |
2988 | } | |
2989 | } | |
2990 | } else { | |
2991 | ret = pret < 0 ? pret : -EFAULT; | |
2992 | } | |
2993 | up_read(¤t->mm->mmap_sem); | |
2994 | if (ret) { | |
2995 | /* | |
2996 | * if we did partial map, or found file backed vmas, | |
2997 | * release any pages we did get | |
2998 | */ | |
27c4d3a3 JH |
2999 | if (pret > 0) |
3000 | put_user_pages(pages, pret); | |
edafccee JA |
3001 | if (ctx->account_mem) |
3002 | io_unaccount_mem(ctx->user, nr_pages); | |
d4ef6475 | 3003 | kvfree(imu->bvec); |
edafccee JA |
3004 | goto err; |
3005 | } | |
3006 | ||
3007 | off = ubuf & ~PAGE_MASK; | |
3008 | size = iov.iov_len; | |
3009 | for (j = 0; j < nr_pages; j++) { | |
3010 | size_t vec_len; | |
3011 | ||
3012 | vec_len = min_t(size_t, size, PAGE_SIZE - off); | |
3013 | imu->bvec[j].bv_page = pages[j]; | |
3014 | imu->bvec[j].bv_len = vec_len; | |
3015 | imu->bvec[j].bv_offset = off; | |
3016 | off = 0; | |
3017 | size -= vec_len; | |
3018 | } | |
3019 | /* store original address for later verification */ | |
3020 | imu->ubuf = ubuf; | |
3021 | imu->len = iov.iov_len; | |
3022 | imu->nr_bvecs = nr_pages; | |
3023 | ||
3024 | ctx->nr_user_bufs++; | |
3025 | } | |
d4ef6475 MR |
3026 | kvfree(pages); |
3027 | kvfree(vmas); | |
edafccee JA |
3028 | return 0; |
3029 | err: | |
d4ef6475 MR |
3030 | kvfree(pages); |
3031 | kvfree(vmas); | |
edafccee JA |
3032 | io_sqe_buffer_unregister(ctx); |
3033 | return ret; | |
3034 | } | |
3035 | ||
9b402849 JA |
3036 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg) |
3037 | { | |
3038 | __s32 __user *fds = arg; | |
3039 | int fd; | |
3040 | ||
3041 | if (ctx->cq_ev_fd) | |
3042 | return -EBUSY; | |
3043 | ||
3044 | if (copy_from_user(&fd, fds, sizeof(*fds))) | |
3045 | return -EFAULT; | |
3046 | ||
3047 | ctx->cq_ev_fd = eventfd_ctx_fdget(fd); | |
3048 | if (IS_ERR(ctx->cq_ev_fd)) { | |
3049 | int ret = PTR_ERR(ctx->cq_ev_fd); | |
3050 | ctx->cq_ev_fd = NULL; | |
3051 | return ret; | |
3052 | } | |
3053 | ||
3054 | return 0; | |
3055 | } | |
3056 | ||
3057 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) | |
3058 | { | |
3059 | if (ctx->cq_ev_fd) { | |
3060 | eventfd_ctx_put(ctx->cq_ev_fd); | |
3061 | ctx->cq_ev_fd = NULL; | |
3062 | return 0; | |
3063 | } | |
3064 | ||
3065 | return -ENXIO; | |
3066 | } | |
3067 | ||
2b188cc1 JA |
3068 | static void io_ring_ctx_free(struct io_ring_ctx *ctx) |
3069 | { | |
6b06314c | 3070 | io_finish_async(ctx); |
2b188cc1 JA |
3071 | if (ctx->sqo_mm) |
3072 | mmdrop(ctx->sqo_mm); | |
def596e9 JA |
3073 | |
3074 | io_iopoll_reap_events(ctx); | |
edafccee | 3075 | io_sqe_buffer_unregister(ctx); |
6b06314c | 3076 | io_sqe_files_unregister(ctx); |
9b402849 | 3077 | io_eventfd_unregister(ctx); |
def596e9 | 3078 | |
2b188cc1 | 3079 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
3080 | if (ctx->ring_sock) { |
3081 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 3082 | sock_release(ctx->ring_sock); |
355e8d26 | 3083 | } |
2b188cc1 JA |
3084 | #endif |
3085 | ||
75b28aff | 3086 | io_mem_free(ctx->rings); |
2b188cc1 | 3087 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
3088 | |
3089 | percpu_ref_exit(&ctx->refs); | |
3090 | if (ctx->account_mem) | |
3091 | io_unaccount_mem(ctx->user, | |
3092 | ring_pages(ctx->sq_entries, ctx->cq_entries)); | |
3093 | free_uid(ctx->user); | |
3094 | kfree(ctx); | |
3095 | } | |
3096 | ||
3097 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
3098 | { | |
3099 | struct io_ring_ctx *ctx = file->private_data; | |
3100 | __poll_t mask = 0; | |
3101 | ||
3102 | poll_wait(file, &ctx->cq_wait, wait); | |
4f7067c3 SB |
3103 | /* |
3104 | * synchronizes with barrier from wq_has_sleeper call in | |
3105 | * io_commit_cqring | |
3106 | */ | |
2b188cc1 | 3107 | smp_rmb(); |
75b28aff HV |
3108 | if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head != |
3109 | ctx->rings->sq_ring_entries) | |
2b188cc1 | 3110 | mask |= EPOLLOUT | EPOLLWRNORM; |
75b28aff | 3111 | if (READ_ONCE(ctx->rings->sq.head) != ctx->cached_cq_tail) |
2b188cc1 JA |
3112 | mask |= EPOLLIN | EPOLLRDNORM; |
3113 | ||
3114 | return mask; | |
3115 | } | |
3116 | ||
3117 | static int io_uring_fasync(int fd, struct file *file, int on) | |
3118 | { | |
3119 | struct io_ring_ctx *ctx = file->private_data; | |
3120 | ||
3121 | return fasync_helper(fd, file, on, &ctx->cq_fasync); | |
3122 | } | |
3123 | ||
3124 | static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) | |
3125 | { | |
3126 | mutex_lock(&ctx->uring_lock); | |
3127 | percpu_ref_kill(&ctx->refs); | |
3128 | mutex_unlock(&ctx->uring_lock); | |
3129 | ||
221c5eb2 | 3130 | io_poll_remove_all(ctx); |
def596e9 | 3131 | io_iopoll_reap_events(ctx); |
2b188cc1 JA |
3132 | wait_for_completion(&ctx->ctx_done); |
3133 | io_ring_ctx_free(ctx); | |
3134 | } | |
3135 | ||
3136 | static int io_uring_release(struct inode *inode, struct file *file) | |
3137 | { | |
3138 | struct io_ring_ctx *ctx = file->private_data; | |
3139 | ||
3140 | file->private_data = NULL; | |
3141 | io_ring_ctx_wait_and_kill(ctx); | |
3142 | return 0; | |
3143 | } | |
3144 | ||
3145 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
3146 | { | |
3147 | loff_t offset = (loff_t) vma->vm_pgoff << PAGE_SHIFT; | |
3148 | unsigned long sz = vma->vm_end - vma->vm_start; | |
3149 | struct io_ring_ctx *ctx = file->private_data; | |
3150 | unsigned long pfn; | |
3151 | struct page *page; | |
3152 | void *ptr; | |
3153 | ||
3154 | switch (offset) { | |
3155 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
3156 | case IORING_OFF_CQ_RING: |
3157 | ptr = ctx->rings; | |
2b188cc1 JA |
3158 | break; |
3159 | case IORING_OFF_SQES: | |
3160 | ptr = ctx->sq_sqes; | |
3161 | break; | |
2b188cc1 JA |
3162 | default: |
3163 | return -EINVAL; | |
3164 | } | |
3165 | ||
3166 | page = virt_to_head_page(ptr); | |
3167 | if (sz > (PAGE_SIZE << compound_order(page))) | |
3168 | return -EINVAL; | |
3169 | ||
3170 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
3171 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
3172 | } | |
3173 | ||
3174 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, | |
3175 | u32, min_complete, u32, flags, const sigset_t __user *, sig, | |
3176 | size_t, sigsz) | |
3177 | { | |
3178 | struct io_ring_ctx *ctx; | |
3179 | long ret = -EBADF; | |
3180 | int submitted = 0; | |
3181 | struct fd f; | |
3182 | ||
6c271ce2 | 3183 | if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP)) |
2b188cc1 JA |
3184 | return -EINVAL; |
3185 | ||
3186 | f = fdget(fd); | |
3187 | if (!f.file) | |
3188 | return -EBADF; | |
3189 | ||
3190 | ret = -EOPNOTSUPP; | |
3191 | if (f.file->f_op != &io_uring_fops) | |
3192 | goto out_fput; | |
3193 | ||
3194 | ret = -ENXIO; | |
3195 | ctx = f.file->private_data; | |
3196 | if (!percpu_ref_tryget(&ctx->refs)) | |
3197 | goto out_fput; | |
3198 | ||
6c271ce2 JA |
3199 | /* |
3200 | * For SQ polling, the thread will do all submissions and completions. | |
3201 | * Just return the requested submit count, and wake the thread if | |
3202 | * we were asked to. | |
3203 | */ | |
3204 | if (ctx->flags & IORING_SETUP_SQPOLL) { | |
3205 | if (flags & IORING_ENTER_SQ_WAKEUP) | |
3206 | wake_up(&ctx->sqo_wait); | |
3207 | submitted = to_submit; | |
3208 | goto out_ctx; | |
3209 | } | |
3210 | ||
2b188cc1 JA |
3211 | ret = 0; |
3212 | if (to_submit) { | |
3213 | to_submit = min(to_submit, ctx->sq_entries); | |
3214 | ||
3215 | mutex_lock(&ctx->uring_lock); | |
3216 | submitted = io_ring_submit(ctx, to_submit); | |
3217 | mutex_unlock(&ctx->uring_lock); | |
2b188cc1 JA |
3218 | } |
3219 | if (flags & IORING_ENTER_GETEVENTS) { | |
def596e9 JA |
3220 | unsigned nr_events = 0; |
3221 | ||
2b188cc1 JA |
3222 | min_complete = min(min_complete, ctx->cq_entries); |
3223 | ||
def596e9 | 3224 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 | 3225 | ret = io_iopoll_check(ctx, &nr_events, min_complete); |
def596e9 JA |
3226 | } else { |
3227 | ret = io_cqring_wait(ctx, min_complete, sig, sigsz); | |
3228 | } | |
2b188cc1 JA |
3229 | } |
3230 | ||
3231 | out_ctx: | |
3232 | io_ring_drop_ctx_refs(ctx, 1); | |
3233 | out_fput: | |
3234 | fdput(f); | |
3235 | return submitted ? submitted : ret; | |
3236 | } | |
3237 | ||
3238 | static const struct file_operations io_uring_fops = { | |
3239 | .release = io_uring_release, | |
3240 | .mmap = io_uring_mmap, | |
3241 | .poll = io_uring_poll, | |
3242 | .fasync = io_uring_fasync, | |
3243 | }; | |
3244 | ||
3245 | static int io_allocate_scq_urings(struct io_ring_ctx *ctx, | |
3246 | struct io_uring_params *p) | |
3247 | { | |
75b28aff HV |
3248 | struct io_rings *rings; |
3249 | size_t size, sq_array_offset; | |
2b188cc1 | 3250 | |
75b28aff HV |
3251 | size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset); |
3252 | if (size == SIZE_MAX) | |
3253 | return -EOVERFLOW; | |
3254 | ||
3255 | rings = io_mem_alloc(size); | |
3256 | if (!rings) | |
2b188cc1 JA |
3257 | return -ENOMEM; |
3258 | ||
75b28aff HV |
3259 | ctx->rings = rings; |
3260 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
3261 | rings->sq_ring_mask = p->sq_entries - 1; | |
3262 | rings->cq_ring_mask = p->cq_entries - 1; | |
3263 | rings->sq_ring_entries = p->sq_entries; | |
3264 | rings->cq_ring_entries = p->cq_entries; | |
3265 | ctx->sq_mask = rings->sq_ring_mask; | |
3266 | ctx->cq_mask = rings->cq_ring_mask; | |
3267 | ctx->sq_entries = rings->sq_ring_entries; | |
3268 | ctx->cq_entries = rings->cq_ring_entries; | |
2b188cc1 JA |
3269 | |
3270 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
3271 | if (size == SIZE_MAX) | |
3272 | return -EOVERFLOW; | |
3273 | ||
3274 | ctx->sq_sqes = io_mem_alloc(size); | |
52e04ef4 | 3275 | if (!ctx->sq_sqes) |
2b188cc1 | 3276 | return -ENOMEM; |
2b188cc1 | 3277 | |
2b188cc1 JA |
3278 | return 0; |
3279 | } | |
3280 | ||
3281 | /* | |
3282 | * Allocate an anonymous fd, this is what constitutes the application | |
3283 | * visible backing of an io_uring instance. The application mmaps this | |
3284 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
3285 | * we have to tie this fd to a socket for file garbage collection purposes. | |
3286 | */ | |
3287 | static int io_uring_get_fd(struct io_ring_ctx *ctx) | |
3288 | { | |
3289 | struct file *file; | |
3290 | int ret; | |
3291 | ||
3292 | #if defined(CONFIG_UNIX) | |
3293 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, | |
3294 | &ctx->ring_sock); | |
3295 | if (ret) | |
3296 | return ret; | |
3297 | #endif | |
3298 | ||
3299 | ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
3300 | if (ret < 0) | |
3301 | goto err; | |
3302 | ||
3303 | file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx, | |
3304 | O_RDWR | O_CLOEXEC); | |
3305 | if (IS_ERR(file)) { | |
3306 | put_unused_fd(ret); | |
3307 | ret = PTR_ERR(file); | |
3308 | goto err; | |
3309 | } | |
3310 | ||
3311 | #if defined(CONFIG_UNIX) | |
3312 | ctx->ring_sock->file = file; | |
6b06314c | 3313 | ctx->ring_sock->sk->sk_user_data = ctx; |
2b188cc1 JA |
3314 | #endif |
3315 | fd_install(ret, file); | |
3316 | return ret; | |
3317 | err: | |
3318 | #if defined(CONFIG_UNIX) | |
3319 | sock_release(ctx->ring_sock); | |
3320 | ctx->ring_sock = NULL; | |
3321 | #endif | |
3322 | return ret; | |
3323 | } | |
3324 | ||
3325 | static int io_uring_create(unsigned entries, struct io_uring_params *p) | |
3326 | { | |
3327 | struct user_struct *user = NULL; | |
3328 | struct io_ring_ctx *ctx; | |
3329 | bool account_mem; | |
3330 | int ret; | |
3331 | ||
3332 | if (!entries || entries > IORING_MAX_ENTRIES) | |
3333 | return -EINVAL; | |
3334 | ||
3335 | /* | |
3336 | * Use twice as many entries for the CQ ring. It's possible for the | |
3337 | * application to drive a higher depth than the size of the SQ ring, | |
3338 | * since the sqes are only used at submission time. This allows for | |
3339 | * some flexibility in overcommitting a bit. | |
3340 | */ | |
3341 | p->sq_entries = roundup_pow_of_two(entries); | |
3342 | p->cq_entries = 2 * p->sq_entries; | |
3343 | ||
3344 | user = get_uid(current_user()); | |
3345 | account_mem = !capable(CAP_IPC_LOCK); | |
3346 | ||
3347 | if (account_mem) { | |
3348 | ret = io_account_mem(user, | |
3349 | ring_pages(p->sq_entries, p->cq_entries)); | |
3350 | if (ret) { | |
3351 | free_uid(user); | |
3352 | return ret; | |
3353 | } | |
3354 | } | |
3355 | ||
3356 | ctx = io_ring_ctx_alloc(p); | |
3357 | if (!ctx) { | |
3358 | if (account_mem) | |
3359 | io_unaccount_mem(user, ring_pages(p->sq_entries, | |
3360 | p->cq_entries)); | |
3361 | free_uid(user); | |
3362 | return -ENOMEM; | |
3363 | } | |
3364 | ctx->compat = in_compat_syscall(); | |
3365 | ctx->account_mem = account_mem; | |
3366 | ctx->user = user; | |
3367 | ||
3368 | ret = io_allocate_scq_urings(ctx, p); | |
3369 | if (ret) | |
3370 | goto err; | |
3371 | ||
6c271ce2 | 3372 | ret = io_sq_offload_start(ctx, p); |
2b188cc1 JA |
3373 | if (ret) |
3374 | goto err; | |
3375 | ||
3376 | ret = io_uring_get_fd(ctx); | |
3377 | if (ret < 0) | |
3378 | goto err; | |
3379 | ||
3380 | memset(&p->sq_off, 0, sizeof(p->sq_off)); | |
75b28aff HV |
3381 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
3382 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
3383 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
3384 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
3385 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
3386 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
3387 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
3388 | |
3389 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
3390 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
3391 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
3392 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
3393 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
3394 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
3395 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
ac90f249 JA |
3396 | |
3397 | p->features = IORING_FEAT_SINGLE_MMAP; | |
2b188cc1 JA |
3398 | return ret; |
3399 | err: | |
3400 | io_ring_ctx_wait_and_kill(ctx); | |
3401 | return ret; | |
3402 | } | |
3403 | ||
3404 | /* | |
3405 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
3406 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
3407 | * params structure passed in. | |
3408 | */ | |
3409 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
3410 | { | |
3411 | struct io_uring_params p; | |
3412 | long ret; | |
3413 | int i; | |
3414 | ||
3415 | if (copy_from_user(&p, params, sizeof(p))) | |
3416 | return -EFAULT; | |
3417 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
3418 | if (p.resv[i]) | |
3419 | return -EINVAL; | |
3420 | } | |
3421 | ||
6c271ce2 JA |
3422 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
3423 | IORING_SETUP_SQ_AFF)) | |
2b188cc1 JA |
3424 | return -EINVAL; |
3425 | ||
3426 | ret = io_uring_create(entries, &p); | |
3427 | if (ret < 0) | |
3428 | return ret; | |
3429 | ||
3430 | if (copy_to_user(params, &p, sizeof(p))) | |
3431 | return -EFAULT; | |
3432 | ||
3433 | return ret; | |
3434 | } | |
3435 | ||
3436 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
3437 | struct io_uring_params __user *, params) | |
3438 | { | |
3439 | return io_uring_setup(entries, params); | |
3440 | } | |
3441 | ||
edafccee JA |
3442 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
3443 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
3444 | __releases(ctx->uring_lock) |
3445 | __acquires(ctx->uring_lock) | |
edafccee JA |
3446 | { |
3447 | int ret; | |
3448 | ||
35fa71a0 JA |
3449 | /* |
3450 | * We're inside the ring mutex, if the ref is already dying, then | |
3451 | * someone else killed the ctx or is already going through | |
3452 | * io_uring_register(). | |
3453 | */ | |
3454 | if (percpu_ref_is_dying(&ctx->refs)) | |
3455 | return -ENXIO; | |
3456 | ||
edafccee | 3457 | percpu_ref_kill(&ctx->refs); |
b19062a5 JA |
3458 | |
3459 | /* | |
3460 | * Drop uring mutex before waiting for references to exit. If another | |
3461 | * thread is currently inside io_uring_enter() it might need to grab | |
3462 | * the uring_lock to make progress. If we hold it here across the drain | |
3463 | * wait, then we can deadlock. It's safe to drop the mutex here, since | |
3464 | * no new references will come in after we've killed the percpu ref. | |
3465 | */ | |
3466 | mutex_unlock(&ctx->uring_lock); | |
edafccee | 3467 | wait_for_completion(&ctx->ctx_done); |
b19062a5 | 3468 | mutex_lock(&ctx->uring_lock); |
edafccee JA |
3469 | |
3470 | switch (opcode) { | |
3471 | case IORING_REGISTER_BUFFERS: | |
3472 | ret = io_sqe_buffer_register(ctx, arg, nr_args); | |
3473 | break; | |
3474 | case IORING_UNREGISTER_BUFFERS: | |
3475 | ret = -EINVAL; | |
3476 | if (arg || nr_args) | |
3477 | break; | |
3478 | ret = io_sqe_buffer_unregister(ctx); | |
3479 | break; | |
6b06314c JA |
3480 | case IORING_REGISTER_FILES: |
3481 | ret = io_sqe_files_register(ctx, arg, nr_args); | |
3482 | break; | |
3483 | case IORING_UNREGISTER_FILES: | |
3484 | ret = -EINVAL; | |
3485 | if (arg || nr_args) | |
3486 | break; | |
3487 | ret = io_sqe_files_unregister(ctx); | |
3488 | break; | |
9b402849 JA |
3489 | case IORING_REGISTER_EVENTFD: |
3490 | ret = -EINVAL; | |
3491 | if (nr_args != 1) | |
3492 | break; | |
3493 | ret = io_eventfd_register(ctx, arg); | |
3494 | break; | |
3495 | case IORING_UNREGISTER_EVENTFD: | |
3496 | ret = -EINVAL; | |
3497 | if (arg || nr_args) | |
3498 | break; | |
3499 | ret = io_eventfd_unregister(ctx); | |
3500 | break; | |
edafccee JA |
3501 | default: |
3502 | ret = -EINVAL; | |
3503 | break; | |
3504 | } | |
3505 | ||
3506 | /* bring the ctx back to life */ | |
3507 | reinit_completion(&ctx->ctx_done); | |
3508 | percpu_ref_reinit(&ctx->refs); | |
3509 | return ret; | |
3510 | } | |
3511 | ||
3512 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
3513 | void __user *, arg, unsigned int, nr_args) | |
3514 | { | |
3515 | struct io_ring_ctx *ctx; | |
3516 | long ret = -EBADF; | |
3517 | struct fd f; | |
3518 | ||
3519 | f = fdget(fd); | |
3520 | if (!f.file) | |
3521 | return -EBADF; | |
3522 | ||
3523 | ret = -EOPNOTSUPP; | |
3524 | if (f.file->f_op != &io_uring_fops) | |
3525 | goto out_fput; | |
3526 | ||
3527 | ctx = f.file->private_data; | |
3528 | ||
3529 | mutex_lock(&ctx->uring_lock); | |
3530 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
3531 | mutex_unlock(&ctx->uring_lock); | |
3532 | out_fput: | |
3533 | fdput(f); | |
3534 | return ret; | |
3535 | } | |
3536 | ||
2b188cc1 JA |
3537 | static int __init io_uring_init(void) |
3538 | { | |
3539 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC); | |
3540 | return 0; | |
3541 | }; | |
3542 | __initcall(io_uring_init); |