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 | ||
5277deaa | 78 | #define IORING_MAX_ENTRIES 32768 |
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; | |
6d5d5ac5 | 170 | off_t io_start; |
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; | |
5262f567 | 203 | struct list_head timeout_list; |
2b188cc1 JA |
204 | } ____cacheline_aligned_in_smp; |
205 | ||
206 | /* IO offload */ | |
54a91f3b | 207 | struct workqueue_struct *sqo_wq[2]; |
6c271ce2 | 208 | struct task_struct *sqo_thread; /* if using sq thread polling */ |
2b188cc1 | 209 | struct mm_struct *sqo_mm; |
6c271ce2 | 210 | wait_queue_head_t sqo_wait; |
a4c0b3de | 211 | struct completion sqo_thread_started; |
2b188cc1 JA |
212 | |
213 | struct { | |
2b188cc1 JA |
214 | unsigned cached_cq_tail; |
215 | unsigned cq_entries; | |
216 | unsigned cq_mask; | |
217 | struct wait_queue_head cq_wait; | |
218 | struct fasync_struct *cq_fasync; | |
9b402849 | 219 | struct eventfd_ctx *cq_ev_fd; |
5262f567 | 220 | atomic_t cq_timeouts; |
2b188cc1 JA |
221 | } ____cacheline_aligned_in_smp; |
222 | ||
75b28aff HV |
223 | struct io_rings *rings; |
224 | ||
6b06314c JA |
225 | /* |
226 | * If used, fixed file set. Writers must ensure that ->refs is dead, | |
227 | * readers must ensure that ->refs is alive as long as the file* is | |
228 | * used. Only updated through io_uring_register(2). | |
229 | */ | |
230 | struct file **user_files; | |
231 | unsigned nr_user_files; | |
232 | ||
edafccee JA |
233 | /* if used, fixed mapped user buffers */ |
234 | unsigned nr_user_bufs; | |
235 | struct io_mapped_ubuf *user_bufs; | |
236 | ||
2b188cc1 JA |
237 | struct user_struct *user; |
238 | ||
239 | struct completion ctx_done; | |
240 | ||
241 | struct { | |
242 | struct mutex uring_lock; | |
243 | wait_queue_head_t wait; | |
244 | } ____cacheline_aligned_in_smp; | |
245 | ||
246 | struct { | |
247 | spinlock_t completion_lock; | |
def596e9 JA |
248 | bool poll_multi_file; |
249 | /* | |
250 | * ->poll_list is protected by the ctx->uring_lock for | |
251 | * io_uring instances that don't use IORING_SETUP_SQPOLL. | |
252 | * For SQPOLL, only the single threaded io_sq_thread() will | |
253 | * manipulate the list, hence no extra locking is needed there. | |
254 | */ | |
255 | struct list_head poll_list; | |
221c5eb2 | 256 | struct list_head cancel_list; |
2b188cc1 JA |
257 | } ____cacheline_aligned_in_smp; |
258 | ||
31b51510 JA |
259 | struct async_list pending_async[2]; |
260 | ||
2b188cc1 JA |
261 | #if defined(CONFIG_UNIX) |
262 | struct socket *ring_sock; | |
263 | #endif | |
264 | }; | |
265 | ||
266 | struct sqe_submit { | |
267 | const struct io_uring_sqe *sqe; | |
268 | unsigned short index; | |
8776f3fa | 269 | u32 sequence; |
2b188cc1 | 270 | bool has_user; |
def596e9 | 271 | bool needs_lock; |
6c271ce2 | 272 | bool needs_fixed_file; |
2b188cc1 JA |
273 | }; |
274 | ||
09bb8394 JA |
275 | /* |
276 | * First field must be the file pointer in all the | |
277 | * iocb unions! See also 'struct kiocb' in <linux/fs.h> | |
278 | */ | |
221c5eb2 JA |
279 | struct io_poll_iocb { |
280 | struct file *file; | |
281 | struct wait_queue_head *head; | |
282 | __poll_t events; | |
8c838788 | 283 | bool done; |
221c5eb2 JA |
284 | bool canceled; |
285 | struct wait_queue_entry wait; | |
286 | }; | |
287 | ||
5262f567 JA |
288 | struct io_timeout { |
289 | struct file *file; | |
290 | struct hrtimer timer; | |
291 | }; | |
292 | ||
09bb8394 JA |
293 | /* |
294 | * NOTE! Each of the iocb union members has the file pointer | |
295 | * as the first entry in their struct definition. So you can | |
296 | * access the file pointer through any of the sub-structs, | |
297 | * or directly as just 'ki_filp' in this struct. | |
298 | */ | |
2b188cc1 | 299 | struct io_kiocb { |
221c5eb2 | 300 | union { |
09bb8394 | 301 | struct file *file; |
221c5eb2 JA |
302 | struct kiocb rw; |
303 | struct io_poll_iocb poll; | |
5262f567 | 304 | struct io_timeout timeout; |
221c5eb2 | 305 | }; |
2b188cc1 JA |
306 | |
307 | struct sqe_submit submit; | |
308 | ||
309 | struct io_ring_ctx *ctx; | |
310 | struct list_head list; | |
9e645e11 | 311 | struct list_head link_list; |
2b188cc1 | 312 | unsigned int flags; |
c16361c1 | 313 | refcount_t refs; |
8449eeda | 314 | #define REQ_F_NOWAIT 1 /* must not punt to workers */ |
def596e9 | 315 | #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */ |
6b06314c | 316 | #define REQ_F_FIXED_FILE 4 /* ctx owns file */ |
31b51510 | 317 | #define REQ_F_SEQ_PREV 8 /* sequential with previous */ |
e2033e33 SB |
318 | #define REQ_F_IO_DRAIN 16 /* drain existing IO first */ |
319 | #define REQ_F_IO_DRAINED 32 /* drain done */ | |
9e645e11 | 320 | #define REQ_F_LINK 64 /* linked sqes */ |
f7b76ac9 ZL |
321 | #define REQ_F_LINK_DONE 128 /* linked sqes done */ |
322 | #define REQ_F_FAIL_LINK 256 /* fail rest of links */ | |
4fe2c963 | 323 | #define REQ_F_SHADOW_DRAIN 512 /* link-drain shadow req */ |
5262f567 | 324 | #define REQ_F_TIMEOUT 1024 /* timeout request */ |
2b188cc1 | 325 | u64 user_data; |
9e645e11 | 326 | u32 result; |
de0617e4 | 327 | u32 sequence; |
2b188cc1 JA |
328 | |
329 | struct work_struct work; | |
330 | }; | |
331 | ||
332 | #define IO_PLUG_THRESHOLD 2 | |
def596e9 | 333 | #define IO_IOPOLL_BATCH 8 |
2b188cc1 | 334 | |
9a56a232 JA |
335 | struct io_submit_state { |
336 | struct blk_plug plug; | |
337 | ||
2579f913 JA |
338 | /* |
339 | * io_kiocb alloc cache | |
340 | */ | |
341 | void *reqs[IO_IOPOLL_BATCH]; | |
342 | unsigned int free_reqs; | |
343 | unsigned int cur_req; | |
344 | ||
9a56a232 JA |
345 | /* |
346 | * File reference cache | |
347 | */ | |
348 | struct file *file; | |
349 | unsigned int fd; | |
350 | unsigned int has_refs; | |
351 | unsigned int used_refs; | |
352 | unsigned int ios_left; | |
353 | }; | |
354 | ||
de0617e4 | 355 | static void io_sq_wq_submit_work(struct work_struct *work); |
5262f567 JA |
356 | static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data, |
357 | long res); | |
4fe2c963 | 358 | static void __io_free_req(struct io_kiocb *req); |
de0617e4 | 359 | |
2b188cc1 JA |
360 | static struct kmem_cache *req_cachep; |
361 | ||
362 | static const struct file_operations io_uring_fops; | |
363 | ||
364 | struct sock *io_uring_get_socket(struct file *file) | |
365 | { | |
366 | #if defined(CONFIG_UNIX) | |
367 | if (file->f_op == &io_uring_fops) { | |
368 | struct io_ring_ctx *ctx = file->private_data; | |
369 | ||
370 | return ctx->ring_sock->sk; | |
371 | } | |
372 | #endif | |
373 | return NULL; | |
374 | } | |
375 | EXPORT_SYMBOL(io_uring_get_socket); | |
376 | ||
377 | static void io_ring_ctx_ref_free(struct percpu_ref *ref) | |
378 | { | |
379 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
380 | ||
381 | complete(&ctx->ctx_done); | |
382 | } | |
383 | ||
384 | static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) | |
385 | { | |
386 | struct io_ring_ctx *ctx; | |
31b51510 | 387 | int i; |
2b188cc1 JA |
388 | |
389 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
390 | if (!ctx) | |
391 | return NULL; | |
392 | ||
21482896 RG |
393 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
394 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) { | |
2b188cc1 JA |
395 | kfree(ctx); |
396 | return NULL; | |
397 | } | |
398 | ||
399 | ctx->flags = p->flags; | |
400 | init_waitqueue_head(&ctx->cq_wait); | |
401 | init_completion(&ctx->ctx_done); | |
a4c0b3de | 402 | init_completion(&ctx->sqo_thread_started); |
2b188cc1 JA |
403 | mutex_init(&ctx->uring_lock); |
404 | init_waitqueue_head(&ctx->wait); | |
31b51510 JA |
405 | for (i = 0; i < ARRAY_SIZE(ctx->pending_async); i++) { |
406 | spin_lock_init(&ctx->pending_async[i].lock); | |
407 | INIT_LIST_HEAD(&ctx->pending_async[i].list); | |
408 | atomic_set(&ctx->pending_async[i].cnt, 0); | |
409 | } | |
2b188cc1 | 410 | spin_lock_init(&ctx->completion_lock); |
def596e9 | 411 | INIT_LIST_HEAD(&ctx->poll_list); |
221c5eb2 | 412 | INIT_LIST_HEAD(&ctx->cancel_list); |
de0617e4 | 413 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 414 | INIT_LIST_HEAD(&ctx->timeout_list); |
2b188cc1 JA |
415 | return ctx; |
416 | } | |
417 | ||
de0617e4 JA |
418 | static inline bool io_sequence_defer(struct io_ring_ctx *ctx, |
419 | struct io_kiocb *req) | |
420 | { | |
5262f567 JA |
421 | /* timeout requests always honor sequence */ |
422 | if (!(req->flags & REQ_F_TIMEOUT) && | |
423 | (req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) != REQ_F_IO_DRAIN) | |
de0617e4 JA |
424 | return false; |
425 | ||
75b28aff | 426 | return req->sequence != ctx->cached_cq_tail + ctx->rings->sq_dropped; |
de0617e4 JA |
427 | } |
428 | ||
5262f567 JA |
429 | static struct io_kiocb *__io_get_deferred_req(struct io_ring_ctx *ctx, |
430 | struct list_head *list) | |
de0617e4 JA |
431 | { |
432 | struct io_kiocb *req; | |
433 | ||
5262f567 | 434 | if (list_empty(list)) |
de0617e4 JA |
435 | return NULL; |
436 | ||
5262f567 | 437 | req = list_first_entry(list, struct io_kiocb, list); |
de0617e4 JA |
438 | if (!io_sequence_defer(ctx, req)) { |
439 | list_del_init(&req->list); | |
440 | return req; | |
441 | } | |
442 | ||
443 | return NULL; | |
444 | } | |
445 | ||
5262f567 JA |
446 | static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx) |
447 | { | |
448 | return __io_get_deferred_req(ctx, &ctx->defer_list); | |
449 | } | |
450 | ||
451 | static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx) | |
452 | { | |
453 | return __io_get_deferred_req(ctx, &ctx->timeout_list); | |
454 | } | |
455 | ||
de0617e4 | 456 | static void __io_commit_cqring(struct io_ring_ctx *ctx) |
2b188cc1 | 457 | { |
75b28aff | 458 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 459 | |
75b28aff | 460 | if (ctx->cached_cq_tail != READ_ONCE(rings->cq.tail)) { |
2b188cc1 | 461 | /* order cqe stores with ring update */ |
75b28aff | 462 | smp_store_release(&rings->cq.tail, ctx->cached_cq_tail); |
2b188cc1 | 463 | |
2b188cc1 JA |
464 | if (wq_has_sleeper(&ctx->cq_wait)) { |
465 | wake_up_interruptible(&ctx->cq_wait); | |
466 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); | |
467 | } | |
468 | } | |
469 | } | |
470 | ||
18d9be1a JA |
471 | static inline void io_queue_async_work(struct io_ring_ctx *ctx, |
472 | struct io_kiocb *req) | |
473 | { | |
6cc47d1d | 474 | int rw = 0; |
54a91f3b | 475 | |
6cc47d1d JA |
476 | if (req->submit.sqe) { |
477 | switch (req->submit.sqe->opcode) { | |
478 | case IORING_OP_WRITEV: | |
479 | case IORING_OP_WRITE_FIXED: | |
480 | rw = !(req->rw.ki_flags & IOCB_DIRECT); | |
481 | break; | |
482 | } | |
54a91f3b JA |
483 | } |
484 | ||
485 | queue_work(ctx->sqo_wq[rw], &req->work); | |
18d9be1a JA |
486 | } |
487 | ||
5262f567 JA |
488 | static void io_kill_timeout(struct io_kiocb *req) |
489 | { | |
490 | int ret; | |
491 | ||
492 | ret = hrtimer_try_to_cancel(&req->timeout.timer); | |
493 | if (ret != -1) { | |
494 | atomic_inc(&req->ctx->cq_timeouts); | |
495 | list_del(&req->list); | |
496 | io_cqring_fill_event(req->ctx, req->user_data, 0); | |
497 | __io_free_req(req); | |
498 | } | |
499 | } | |
500 | ||
501 | static void io_kill_timeouts(struct io_ring_ctx *ctx) | |
502 | { | |
503 | struct io_kiocb *req, *tmp; | |
504 | ||
505 | spin_lock_irq(&ctx->completion_lock); | |
506 | list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list) | |
507 | io_kill_timeout(req); | |
508 | spin_unlock_irq(&ctx->completion_lock); | |
509 | } | |
510 | ||
de0617e4 JA |
511 | static void io_commit_cqring(struct io_ring_ctx *ctx) |
512 | { | |
513 | struct io_kiocb *req; | |
514 | ||
5262f567 JA |
515 | while ((req = io_get_timeout_req(ctx)) != NULL) |
516 | io_kill_timeout(req); | |
517 | ||
de0617e4 JA |
518 | __io_commit_cqring(ctx); |
519 | ||
520 | while ((req = io_get_deferred_req(ctx)) != NULL) { | |
4fe2c963 JL |
521 | if (req->flags & REQ_F_SHADOW_DRAIN) { |
522 | /* Just for drain, free it. */ | |
523 | __io_free_req(req); | |
524 | continue; | |
525 | } | |
de0617e4 | 526 | req->flags |= REQ_F_IO_DRAINED; |
18d9be1a | 527 | io_queue_async_work(ctx, req); |
de0617e4 JA |
528 | } |
529 | } | |
530 | ||
2b188cc1 JA |
531 | static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx) |
532 | { | |
75b28aff | 533 | struct io_rings *rings = ctx->rings; |
2b188cc1 JA |
534 | unsigned tail; |
535 | ||
536 | tail = ctx->cached_cq_tail; | |
115e12e5 SB |
537 | /* |
538 | * writes to the cq entry need to come after reading head; the | |
539 | * control dependency is enough as we're using WRITE_ONCE to | |
540 | * fill the cq entry | |
541 | */ | |
75b28aff | 542 | if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries) |
2b188cc1 JA |
543 | return NULL; |
544 | ||
545 | ctx->cached_cq_tail++; | |
75b28aff | 546 | return &rings->cqes[tail & ctx->cq_mask]; |
2b188cc1 JA |
547 | } |
548 | ||
549 | static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data, | |
c71ffb67 | 550 | long res) |
2b188cc1 JA |
551 | { |
552 | struct io_uring_cqe *cqe; | |
553 | ||
554 | /* | |
555 | * If we can't get a cq entry, userspace overflowed the | |
556 | * submission (by quite a lot). Increment the overflow count in | |
557 | * the ring. | |
558 | */ | |
559 | cqe = io_get_cqring(ctx); | |
560 | if (cqe) { | |
561 | WRITE_ONCE(cqe->user_data, ki_user_data); | |
562 | WRITE_ONCE(cqe->res, res); | |
c71ffb67 | 563 | WRITE_ONCE(cqe->flags, 0); |
2b188cc1 | 564 | } else { |
75b28aff | 565 | unsigned overflow = READ_ONCE(ctx->rings->cq_overflow); |
2b188cc1 | 566 | |
75b28aff | 567 | WRITE_ONCE(ctx->rings->cq_overflow, overflow + 1); |
2b188cc1 JA |
568 | } |
569 | } | |
570 | ||
8c838788 JA |
571 | static void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
572 | { | |
573 | if (waitqueue_active(&ctx->wait)) | |
574 | wake_up(&ctx->wait); | |
575 | if (waitqueue_active(&ctx->sqo_wait)) | |
576 | wake_up(&ctx->sqo_wait); | |
9b402849 JA |
577 | if (ctx->cq_ev_fd) |
578 | eventfd_signal(ctx->cq_ev_fd, 1); | |
8c838788 JA |
579 | } |
580 | ||
581 | static void io_cqring_add_event(struct io_ring_ctx *ctx, u64 user_data, | |
c71ffb67 | 582 | long res) |
2b188cc1 JA |
583 | { |
584 | unsigned long flags; | |
585 | ||
586 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
c71ffb67 | 587 | io_cqring_fill_event(ctx, user_data, res); |
2b188cc1 JA |
588 | io_commit_cqring(ctx); |
589 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
590 | ||
8c838788 | 591 | io_cqring_ev_posted(ctx); |
2b188cc1 JA |
592 | } |
593 | ||
2579f913 JA |
594 | static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx, |
595 | struct io_submit_state *state) | |
2b188cc1 | 596 | { |
fd6fab2c | 597 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
2b188cc1 JA |
598 | struct io_kiocb *req; |
599 | ||
600 | if (!percpu_ref_tryget(&ctx->refs)) | |
601 | return NULL; | |
602 | ||
2579f913 | 603 | if (!state) { |
fd6fab2c | 604 | req = kmem_cache_alloc(req_cachep, gfp); |
2579f913 JA |
605 | if (unlikely(!req)) |
606 | goto out; | |
607 | } else if (!state->free_reqs) { | |
608 | size_t sz; | |
609 | int ret; | |
610 | ||
611 | sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs)); | |
fd6fab2c JA |
612 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs); |
613 | ||
614 | /* | |
615 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
616 | * retry single alloc to be on the safe side. | |
617 | */ | |
618 | if (unlikely(ret <= 0)) { | |
619 | state->reqs[0] = kmem_cache_alloc(req_cachep, gfp); | |
620 | if (!state->reqs[0]) | |
621 | goto out; | |
622 | ret = 1; | |
623 | } | |
2579f913 JA |
624 | state->free_reqs = ret - 1; |
625 | state->cur_req = 1; | |
626 | req = state->reqs[0]; | |
627 | } else { | |
628 | req = state->reqs[state->cur_req]; | |
629 | state->free_reqs--; | |
630 | state->cur_req++; | |
2b188cc1 JA |
631 | } |
632 | ||
60c112b0 | 633 | req->file = NULL; |
2579f913 JA |
634 | req->ctx = ctx; |
635 | req->flags = 0; | |
e65ef56d JA |
636 | /* one is dropped after submission, the other at completion */ |
637 | refcount_set(&req->refs, 2); | |
9e645e11 | 638 | req->result = 0; |
2579f913 JA |
639 | return req; |
640 | out: | |
6805b32e | 641 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
642 | return NULL; |
643 | } | |
644 | ||
def596e9 JA |
645 | static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr) |
646 | { | |
647 | if (*nr) { | |
648 | kmem_cache_free_bulk(req_cachep, *nr, reqs); | |
6805b32e | 649 | percpu_ref_put_many(&ctx->refs, *nr); |
def596e9 JA |
650 | *nr = 0; |
651 | } | |
652 | } | |
653 | ||
9e645e11 | 654 | static void __io_free_req(struct io_kiocb *req) |
2b188cc1 | 655 | { |
09bb8394 JA |
656 | if (req->file && !(req->flags & REQ_F_FIXED_FILE)) |
657 | fput(req->file); | |
6805b32e | 658 | percpu_ref_put(&req->ctx->refs); |
e65ef56d JA |
659 | kmem_cache_free(req_cachep, req); |
660 | } | |
661 | ||
9e645e11 JA |
662 | static void io_req_link_next(struct io_kiocb *req) |
663 | { | |
664 | struct io_kiocb *nxt; | |
665 | ||
666 | /* | |
667 | * The list should never be empty when we are called here. But could | |
668 | * potentially happen if the chain is messed up, check to be on the | |
669 | * safe side. | |
670 | */ | |
671 | nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list); | |
672 | if (nxt) { | |
673 | list_del(&nxt->list); | |
674 | if (!list_empty(&req->link_list)) { | |
675 | INIT_LIST_HEAD(&nxt->link_list); | |
676 | list_splice(&req->link_list, &nxt->link_list); | |
677 | nxt->flags |= REQ_F_LINK; | |
678 | } | |
679 | ||
f7b76ac9 | 680 | nxt->flags |= REQ_F_LINK_DONE; |
9e645e11 | 681 | INIT_WORK(&nxt->work, io_sq_wq_submit_work); |
18d9be1a | 682 | io_queue_async_work(req->ctx, nxt); |
9e645e11 JA |
683 | } |
684 | } | |
685 | ||
686 | /* | |
687 | * Called if REQ_F_LINK is set, and we fail the head request | |
688 | */ | |
689 | static void io_fail_links(struct io_kiocb *req) | |
690 | { | |
691 | struct io_kiocb *link; | |
692 | ||
693 | while (!list_empty(&req->link_list)) { | |
694 | link = list_first_entry(&req->link_list, struct io_kiocb, list); | |
695 | list_del(&link->list); | |
696 | ||
697 | io_cqring_add_event(req->ctx, link->user_data, -ECANCELED); | |
698 | __io_free_req(link); | |
699 | } | |
700 | } | |
701 | ||
702 | static void io_free_req(struct io_kiocb *req) | |
703 | { | |
704 | /* | |
705 | * If LINK is set, we have dependent requests in this chain. If we | |
706 | * didn't fail this request, queue the first one up, moving any other | |
707 | * dependencies to the next request. In case of failure, fail the rest | |
708 | * of the chain. | |
709 | */ | |
710 | if (req->flags & REQ_F_LINK) { | |
711 | if (req->flags & REQ_F_FAIL_LINK) | |
712 | io_fail_links(req); | |
713 | else | |
714 | io_req_link_next(req); | |
715 | } | |
716 | ||
717 | __io_free_req(req); | |
718 | } | |
719 | ||
e65ef56d JA |
720 | static void io_put_req(struct io_kiocb *req) |
721 | { | |
722 | if (refcount_dec_and_test(&req->refs)) | |
723 | io_free_req(req); | |
2b188cc1 JA |
724 | } |
725 | ||
75b28aff | 726 | static unsigned io_cqring_events(struct io_rings *rings) |
a3a0e43f JA |
727 | { |
728 | /* See comment at the top of this file */ | |
729 | smp_rmb(); | |
75b28aff | 730 | return READ_ONCE(rings->cq.tail) - READ_ONCE(rings->cq.head); |
a3a0e43f JA |
731 | } |
732 | ||
def596e9 JA |
733 | /* |
734 | * Find and free completed poll iocbs | |
735 | */ | |
736 | static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
737 | struct list_head *done) | |
738 | { | |
739 | void *reqs[IO_IOPOLL_BATCH]; | |
740 | struct io_kiocb *req; | |
09bb8394 | 741 | int to_free; |
def596e9 | 742 | |
09bb8394 | 743 | to_free = 0; |
def596e9 JA |
744 | while (!list_empty(done)) { |
745 | req = list_first_entry(done, struct io_kiocb, list); | |
746 | list_del(&req->list); | |
747 | ||
9e645e11 | 748 | io_cqring_fill_event(ctx, req->user_data, req->result); |
def596e9 JA |
749 | (*nr_events)++; |
750 | ||
09bb8394 JA |
751 | if (refcount_dec_and_test(&req->refs)) { |
752 | /* If we're not using fixed files, we have to pair the | |
753 | * completion part with the file put. Use regular | |
754 | * completions for those, only batch free for fixed | |
9e645e11 | 755 | * file and non-linked commands. |
09bb8394 | 756 | */ |
9e645e11 JA |
757 | if ((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) == |
758 | REQ_F_FIXED_FILE) { | |
09bb8394 JA |
759 | reqs[to_free++] = req; |
760 | if (to_free == ARRAY_SIZE(reqs)) | |
761 | io_free_req_many(ctx, reqs, &to_free); | |
6b06314c | 762 | } else { |
09bb8394 | 763 | io_free_req(req); |
6b06314c | 764 | } |
9a56a232 | 765 | } |
def596e9 | 766 | } |
def596e9 | 767 | |
09bb8394 | 768 | io_commit_cqring(ctx); |
def596e9 JA |
769 | io_free_req_many(ctx, reqs, &to_free); |
770 | } | |
771 | ||
772 | static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
773 | long min) | |
774 | { | |
775 | struct io_kiocb *req, *tmp; | |
776 | LIST_HEAD(done); | |
777 | bool spin; | |
778 | int ret; | |
779 | ||
780 | /* | |
781 | * Only spin for completions if we don't have multiple devices hanging | |
782 | * off our complete list, and we're under the requested amount. | |
783 | */ | |
784 | spin = !ctx->poll_multi_file && *nr_events < min; | |
785 | ||
786 | ret = 0; | |
787 | list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) { | |
788 | struct kiocb *kiocb = &req->rw; | |
789 | ||
790 | /* | |
791 | * Move completed entries to our local list. If we find a | |
792 | * request that requires polling, break out and complete | |
793 | * the done list first, if we have entries there. | |
794 | */ | |
795 | if (req->flags & REQ_F_IOPOLL_COMPLETED) { | |
796 | list_move_tail(&req->list, &done); | |
797 | continue; | |
798 | } | |
799 | if (!list_empty(&done)) | |
800 | break; | |
801 | ||
802 | ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin); | |
803 | if (ret < 0) | |
804 | break; | |
805 | ||
806 | if (ret && spin) | |
807 | spin = false; | |
808 | ret = 0; | |
809 | } | |
810 | ||
811 | if (!list_empty(&done)) | |
812 | io_iopoll_complete(ctx, nr_events, &done); | |
813 | ||
814 | return ret; | |
815 | } | |
816 | ||
817 | /* | |
818 | * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a | |
819 | * non-spinning poll check - we'll still enter the driver poll loop, but only | |
820 | * as a non-spinning completion check. | |
821 | */ | |
822 | static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
823 | long min) | |
824 | { | |
08f5439f | 825 | while (!list_empty(&ctx->poll_list) && !need_resched()) { |
def596e9 JA |
826 | int ret; |
827 | ||
828 | ret = io_do_iopoll(ctx, nr_events, min); | |
829 | if (ret < 0) | |
830 | return ret; | |
831 | if (!min || *nr_events >= min) | |
832 | return 0; | |
833 | } | |
834 | ||
835 | return 1; | |
836 | } | |
837 | ||
838 | /* | |
839 | * We can't just wait for polled events to come to us, we have to actively | |
840 | * find and complete them. | |
841 | */ | |
842 | static void io_iopoll_reap_events(struct io_ring_ctx *ctx) | |
843 | { | |
844 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
845 | return; | |
846 | ||
847 | mutex_lock(&ctx->uring_lock); | |
848 | while (!list_empty(&ctx->poll_list)) { | |
849 | unsigned int nr_events = 0; | |
850 | ||
851 | io_iopoll_getevents(ctx, &nr_events, 1); | |
08f5439f JA |
852 | |
853 | /* | |
854 | * Ensure we allow local-to-the-cpu processing to take place, | |
855 | * in this case we need to ensure that we reap all events. | |
856 | */ | |
857 | cond_resched(); | |
def596e9 JA |
858 | } |
859 | mutex_unlock(&ctx->uring_lock); | |
860 | } | |
861 | ||
862 | static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, | |
863 | long min) | |
864 | { | |
500f9fba JA |
865 | int iters, ret = 0; |
866 | ||
867 | /* | |
868 | * We disallow the app entering submit/complete with polling, but we | |
869 | * still need to lock the ring to prevent racing with polled issue | |
870 | * that got punted to a workqueue. | |
871 | */ | |
872 | mutex_lock(&ctx->uring_lock); | |
def596e9 | 873 | |
500f9fba | 874 | iters = 0; |
def596e9 JA |
875 | do { |
876 | int tmin = 0; | |
877 | ||
a3a0e43f JA |
878 | /* |
879 | * Don't enter poll loop if we already have events pending. | |
880 | * If we do, we can potentially be spinning for commands that | |
881 | * already triggered a CQE (eg in error). | |
882 | */ | |
75b28aff | 883 | if (io_cqring_events(ctx->rings)) |
a3a0e43f JA |
884 | break; |
885 | ||
500f9fba JA |
886 | /* |
887 | * If a submit got punted to a workqueue, we can have the | |
888 | * application entering polling for a command before it gets | |
889 | * issued. That app will hold the uring_lock for the duration | |
890 | * of the poll right here, so we need to take a breather every | |
891 | * now and then to ensure that the issue has a chance to add | |
892 | * the poll to the issued list. Otherwise we can spin here | |
893 | * forever, while the workqueue is stuck trying to acquire the | |
894 | * very same mutex. | |
895 | */ | |
896 | if (!(++iters & 7)) { | |
897 | mutex_unlock(&ctx->uring_lock); | |
898 | mutex_lock(&ctx->uring_lock); | |
899 | } | |
900 | ||
def596e9 JA |
901 | if (*nr_events < min) |
902 | tmin = min - *nr_events; | |
903 | ||
904 | ret = io_iopoll_getevents(ctx, nr_events, tmin); | |
905 | if (ret <= 0) | |
906 | break; | |
907 | ret = 0; | |
908 | } while (min && !*nr_events && !need_resched()); | |
909 | ||
500f9fba | 910 | mutex_unlock(&ctx->uring_lock); |
def596e9 JA |
911 | return ret; |
912 | } | |
913 | ||
2b188cc1 JA |
914 | static void kiocb_end_write(struct kiocb *kiocb) |
915 | { | |
916 | if (kiocb->ki_flags & IOCB_WRITE) { | |
917 | struct inode *inode = file_inode(kiocb->ki_filp); | |
918 | ||
919 | /* | |
920 | * Tell lockdep we inherited freeze protection from submission | |
921 | * thread. | |
922 | */ | |
923 | if (S_ISREG(inode->i_mode)) | |
924 | __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE); | |
925 | file_end_write(kiocb->ki_filp); | |
926 | } | |
927 | } | |
928 | ||
929 | static void io_complete_rw(struct kiocb *kiocb, long res, long res2) | |
930 | { | |
931 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
932 | ||
933 | kiocb_end_write(kiocb); | |
934 | ||
9e645e11 JA |
935 | if ((req->flags & REQ_F_LINK) && res != req->result) |
936 | req->flags |= REQ_F_FAIL_LINK; | |
c71ffb67 | 937 | io_cqring_add_event(req->ctx, req->user_data, res); |
e65ef56d | 938 | io_put_req(req); |
2b188cc1 JA |
939 | } |
940 | ||
def596e9 JA |
941 | static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2) |
942 | { | |
943 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw); | |
944 | ||
945 | kiocb_end_write(kiocb); | |
946 | ||
9e645e11 JA |
947 | if ((req->flags & REQ_F_LINK) && res != req->result) |
948 | req->flags |= REQ_F_FAIL_LINK; | |
949 | req->result = res; | |
def596e9 JA |
950 | if (res != -EAGAIN) |
951 | req->flags |= REQ_F_IOPOLL_COMPLETED; | |
952 | } | |
953 | ||
954 | /* | |
955 | * After the iocb has been issued, it's safe to be found on the poll list. | |
956 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
957 | * find it from a io_iopoll_getevents() thread before the issuer is done | |
958 | * accessing the kiocb cookie. | |
959 | */ | |
960 | static void io_iopoll_req_issued(struct io_kiocb *req) | |
961 | { | |
962 | struct io_ring_ctx *ctx = req->ctx; | |
963 | ||
964 | /* | |
965 | * Track whether we have multiple files in our lists. This will impact | |
966 | * how we do polling eventually, not spinning if we're on potentially | |
967 | * different devices. | |
968 | */ | |
969 | if (list_empty(&ctx->poll_list)) { | |
970 | ctx->poll_multi_file = false; | |
971 | } else if (!ctx->poll_multi_file) { | |
972 | struct io_kiocb *list_req; | |
973 | ||
974 | list_req = list_first_entry(&ctx->poll_list, struct io_kiocb, | |
975 | list); | |
976 | if (list_req->rw.ki_filp != req->rw.ki_filp) | |
977 | ctx->poll_multi_file = true; | |
978 | } | |
979 | ||
980 | /* | |
981 | * For fast devices, IO may have already completed. If it has, add | |
982 | * it to the front so we find it first. | |
983 | */ | |
984 | if (req->flags & REQ_F_IOPOLL_COMPLETED) | |
985 | list_add(&req->list, &ctx->poll_list); | |
986 | else | |
987 | list_add_tail(&req->list, &ctx->poll_list); | |
988 | } | |
989 | ||
3d6770fb | 990 | static void io_file_put(struct io_submit_state *state) |
9a56a232 | 991 | { |
3d6770fb | 992 | if (state->file) { |
9a56a232 JA |
993 | int diff = state->has_refs - state->used_refs; |
994 | ||
995 | if (diff) | |
996 | fput_many(state->file, diff); | |
997 | state->file = NULL; | |
998 | } | |
999 | } | |
1000 | ||
1001 | /* | |
1002 | * Get as many references to a file as we have IOs left in this submission, | |
1003 | * assuming most submissions are for one file, or at least that each file | |
1004 | * has more than one submission. | |
1005 | */ | |
1006 | static struct file *io_file_get(struct io_submit_state *state, int fd) | |
1007 | { | |
1008 | if (!state) | |
1009 | return fget(fd); | |
1010 | ||
1011 | if (state->file) { | |
1012 | if (state->fd == fd) { | |
1013 | state->used_refs++; | |
1014 | state->ios_left--; | |
1015 | return state->file; | |
1016 | } | |
3d6770fb | 1017 | io_file_put(state); |
9a56a232 JA |
1018 | } |
1019 | state->file = fget_many(fd, state->ios_left); | |
1020 | if (!state->file) | |
1021 | return NULL; | |
1022 | ||
1023 | state->fd = fd; | |
1024 | state->has_refs = state->ios_left; | |
1025 | state->used_refs = 1; | |
1026 | state->ios_left--; | |
1027 | return state->file; | |
1028 | } | |
1029 | ||
2b188cc1 JA |
1030 | /* |
1031 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1032 | * any file. For now, just ensure that anything potentially problematic is done | |
1033 | * inline. | |
1034 | */ | |
1035 | static bool io_file_supports_async(struct file *file) | |
1036 | { | |
1037 | umode_t mode = file_inode(file)->i_mode; | |
1038 | ||
1039 | if (S_ISBLK(mode) || S_ISCHR(mode)) | |
1040 | return true; | |
1041 | if (S_ISREG(mode) && file->f_op != &io_uring_fops) | |
1042 | return true; | |
1043 | ||
1044 | return false; | |
1045 | } | |
1046 | ||
6c271ce2 | 1047 | static int io_prep_rw(struct io_kiocb *req, const struct sqe_submit *s, |
8358e3a8 | 1048 | bool force_nonblock) |
2b188cc1 | 1049 | { |
6c271ce2 | 1050 | const struct io_uring_sqe *sqe = s->sqe; |
def596e9 | 1051 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 | 1052 | struct kiocb *kiocb = &req->rw; |
09bb8394 JA |
1053 | unsigned ioprio; |
1054 | int ret; | |
2b188cc1 | 1055 | |
09bb8394 JA |
1056 | if (!req->file) |
1057 | return -EBADF; | |
2b188cc1 | 1058 | |
09bb8394 JA |
1059 | if (force_nonblock && !io_file_supports_async(req->file)) |
1060 | force_nonblock = false; | |
6b06314c | 1061 | |
2b188cc1 JA |
1062 | kiocb->ki_pos = READ_ONCE(sqe->off); |
1063 | kiocb->ki_flags = iocb_flags(kiocb->ki_filp); | |
1064 | kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp)); | |
1065 | ||
1066 | ioprio = READ_ONCE(sqe->ioprio); | |
1067 | if (ioprio) { | |
1068 | ret = ioprio_check_cap(ioprio); | |
1069 | if (ret) | |
09bb8394 | 1070 | return ret; |
2b188cc1 JA |
1071 | |
1072 | kiocb->ki_ioprio = ioprio; | |
1073 | } else | |
1074 | kiocb->ki_ioprio = get_current_ioprio(); | |
1075 | ||
1076 | ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags)); | |
1077 | if (unlikely(ret)) | |
09bb8394 | 1078 | return ret; |
8449eeda SB |
1079 | |
1080 | /* don't allow async punt if RWF_NOWAIT was requested */ | |
1081 | if (kiocb->ki_flags & IOCB_NOWAIT) | |
1082 | req->flags |= REQ_F_NOWAIT; | |
1083 | ||
1084 | if (force_nonblock) | |
2b188cc1 | 1085 | kiocb->ki_flags |= IOCB_NOWAIT; |
8449eeda | 1086 | |
def596e9 | 1087 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 JA |
1088 | if (!(kiocb->ki_flags & IOCB_DIRECT) || |
1089 | !kiocb->ki_filp->f_op->iopoll) | |
09bb8394 | 1090 | return -EOPNOTSUPP; |
2b188cc1 | 1091 | |
def596e9 JA |
1092 | kiocb->ki_flags |= IOCB_HIPRI; |
1093 | kiocb->ki_complete = io_complete_rw_iopoll; | |
1094 | } else { | |
09bb8394 JA |
1095 | if (kiocb->ki_flags & IOCB_HIPRI) |
1096 | return -EINVAL; | |
def596e9 JA |
1097 | kiocb->ki_complete = io_complete_rw; |
1098 | } | |
2b188cc1 | 1099 | return 0; |
2b188cc1 JA |
1100 | } |
1101 | ||
1102 | static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) | |
1103 | { | |
1104 | switch (ret) { | |
1105 | case -EIOCBQUEUED: | |
1106 | break; | |
1107 | case -ERESTARTSYS: | |
1108 | case -ERESTARTNOINTR: | |
1109 | case -ERESTARTNOHAND: | |
1110 | case -ERESTART_RESTARTBLOCK: | |
1111 | /* | |
1112 | * We can't just restart the syscall, since previously | |
1113 | * submitted sqes may already be in progress. Just fail this | |
1114 | * IO with EINTR. | |
1115 | */ | |
1116 | ret = -EINTR; | |
1117 | /* fall through */ | |
1118 | default: | |
1119 | kiocb->ki_complete(kiocb, ret, 0); | |
1120 | } | |
1121 | } | |
1122 | ||
edafccee JA |
1123 | static int io_import_fixed(struct io_ring_ctx *ctx, int rw, |
1124 | const struct io_uring_sqe *sqe, | |
1125 | struct iov_iter *iter) | |
1126 | { | |
1127 | size_t len = READ_ONCE(sqe->len); | |
1128 | struct io_mapped_ubuf *imu; | |
1129 | unsigned index, buf_index; | |
1130 | size_t offset; | |
1131 | u64 buf_addr; | |
1132 | ||
1133 | /* attempt to use fixed buffers without having provided iovecs */ | |
1134 | if (unlikely(!ctx->user_bufs)) | |
1135 | return -EFAULT; | |
1136 | ||
1137 | buf_index = READ_ONCE(sqe->buf_index); | |
1138 | if (unlikely(buf_index >= ctx->nr_user_bufs)) | |
1139 | return -EFAULT; | |
1140 | ||
1141 | index = array_index_nospec(buf_index, ctx->nr_user_bufs); | |
1142 | imu = &ctx->user_bufs[index]; | |
1143 | buf_addr = READ_ONCE(sqe->addr); | |
1144 | ||
1145 | /* overflow */ | |
1146 | if (buf_addr + len < buf_addr) | |
1147 | return -EFAULT; | |
1148 | /* not inside the mapped region */ | |
1149 | if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len) | |
1150 | return -EFAULT; | |
1151 | ||
1152 | /* | |
1153 | * May not be a start of buffer, set size appropriately | |
1154 | * and advance us to the beginning. | |
1155 | */ | |
1156 | offset = buf_addr - imu->ubuf; | |
1157 | iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len); | |
bd11b3a3 JA |
1158 | |
1159 | if (offset) { | |
1160 | /* | |
1161 | * Don't use iov_iter_advance() here, as it's really slow for | |
1162 | * using the latter parts of a big fixed buffer - it iterates | |
1163 | * over each segment manually. We can cheat a bit here, because | |
1164 | * we know that: | |
1165 | * | |
1166 | * 1) it's a BVEC iter, we set it up | |
1167 | * 2) all bvecs are PAGE_SIZE in size, except potentially the | |
1168 | * first and last bvec | |
1169 | * | |
1170 | * So just find our index, and adjust the iterator afterwards. | |
1171 | * If the offset is within the first bvec (or the whole first | |
1172 | * bvec, just use iov_iter_advance(). This makes it easier | |
1173 | * since we can just skip the first segment, which may not | |
1174 | * be PAGE_SIZE aligned. | |
1175 | */ | |
1176 | const struct bio_vec *bvec = imu->bvec; | |
1177 | ||
1178 | if (offset <= bvec->bv_len) { | |
1179 | iov_iter_advance(iter, offset); | |
1180 | } else { | |
1181 | unsigned long seg_skip; | |
1182 | ||
1183 | /* skip first vec */ | |
1184 | offset -= bvec->bv_len; | |
1185 | seg_skip = 1 + (offset >> PAGE_SHIFT); | |
1186 | ||
1187 | iter->bvec = bvec + seg_skip; | |
1188 | iter->nr_segs -= seg_skip; | |
99c79f66 | 1189 | iter->count -= bvec->bv_len + offset; |
bd11b3a3 | 1190 | iter->iov_offset = offset & ~PAGE_MASK; |
bd11b3a3 JA |
1191 | } |
1192 | } | |
1193 | ||
edafccee JA |
1194 | return 0; |
1195 | } | |
1196 | ||
87e5e6da JA |
1197 | static ssize_t io_import_iovec(struct io_ring_ctx *ctx, int rw, |
1198 | const struct sqe_submit *s, struct iovec **iovec, | |
1199 | struct iov_iter *iter) | |
2b188cc1 JA |
1200 | { |
1201 | const struct io_uring_sqe *sqe = s->sqe; | |
1202 | void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
1203 | size_t sqe_len = READ_ONCE(sqe->len); | |
edafccee JA |
1204 | u8 opcode; |
1205 | ||
1206 | /* | |
1207 | * We're reading ->opcode for the second time, but the first read | |
1208 | * doesn't care whether it's _FIXED or not, so it doesn't matter | |
1209 | * whether ->opcode changes concurrently. The first read does care | |
1210 | * about whether it is a READ or a WRITE, so we don't trust this read | |
1211 | * for that purpose and instead let the caller pass in the read/write | |
1212 | * flag. | |
1213 | */ | |
1214 | opcode = READ_ONCE(sqe->opcode); | |
1215 | if (opcode == IORING_OP_READ_FIXED || | |
1216 | opcode == IORING_OP_WRITE_FIXED) { | |
87e5e6da | 1217 | ssize_t ret = io_import_fixed(ctx, rw, sqe, iter); |
edafccee JA |
1218 | *iovec = NULL; |
1219 | return ret; | |
1220 | } | |
2b188cc1 JA |
1221 | |
1222 | if (!s->has_user) | |
1223 | return -EFAULT; | |
1224 | ||
1225 | #ifdef CONFIG_COMPAT | |
1226 | if (ctx->compat) | |
1227 | return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV, | |
1228 | iovec, iter); | |
1229 | #endif | |
1230 | ||
1231 | return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter); | |
1232 | } | |
1233 | ||
6d5d5ac5 JA |
1234 | static inline bool io_should_merge(struct async_list *al, struct kiocb *kiocb) |
1235 | { | |
1236 | if (al->file == kiocb->ki_filp) { | |
1237 | off_t start, end; | |
1238 | ||
1239 | /* | |
1240 | * Allow merging if we're anywhere in the range of the same | |
1241 | * page. Generally this happens for sub-page reads or writes, | |
1242 | * and it's beneficial to allow the first worker to bring the | |
1243 | * page in and the piggy backed work can then work on the | |
1244 | * cached page. | |
1245 | */ | |
1246 | start = al->io_start & PAGE_MASK; | |
1247 | end = (al->io_start + al->io_len + PAGE_SIZE - 1) & PAGE_MASK; | |
1248 | if (kiocb->ki_pos >= start && kiocb->ki_pos <= end) | |
1249 | return true; | |
1250 | } | |
1251 | ||
1252 | al->file = NULL; | |
1253 | return false; | |
1254 | } | |
1255 | ||
31b51510 JA |
1256 | /* |
1257 | * Make a note of the last file/offset/direction we punted to async | |
1258 | * context. We'll use this information to see if we can piggy back a | |
1259 | * sequential request onto the previous one, if it's still hasn't been | |
1260 | * completed by the async worker. | |
1261 | */ | |
1262 | static void io_async_list_note(int rw, struct io_kiocb *req, size_t len) | |
1263 | { | |
1264 | struct async_list *async_list = &req->ctx->pending_async[rw]; | |
1265 | struct kiocb *kiocb = &req->rw; | |
1266 | struct file *filp = kiocb->ki_filp; | |
31b51510 | 1267 | |
6d5d5ac5 | 1268 | if (io_should_merge(async_list, kiocb)) { |
9310a7ba | 1269 | unsigned long max_bytes; |
31b51510 JA |
1270 | |
1271 | /* Use 8x RA size as a decent limiter for both reads/writes */ | |
9310a7ba ZL |
1272 | max_bytes = filp->f_ra.ra_pages << (PAGE_SHIFT + 3); |
1273 | if (!max_bytes) | |
1274 | max_bytes = VM_READAHEAD_PAGES << (PAGE_SHIFT + 3); | |
1275 | ||
1276 | /* If max len are exceeded, reset the state */ | |
1277 | if (async_list->io_len + len <= max_bytes) { | |
31b51510 | 1278 | req->flags |= REQ_F_SEQ_PREV; |
9310a7ba | 1279 | async_list->io_len += len; |
31b51510 | 1280 | } else { |
6d5d5ac5 | 1281 | async_list->file = NULL; |
31b51510 JA |
1282 | } |
1283 | } | |
1284 | ||
1285 | /* New file? Reset state. */ | |
1286 | if (async_list->file != filp) { | |
6d5d5ac5 JA |
1287 | async_list->io_start = kiocb->ki_pos; |
1288 | async_list->io_len = len; | |
31b51510 JA |
1289 | async_list->file = filp; |
1290 | } | |
31b51510 JA |
1291 | } |
1292 | ||
32960613 JA |
1293 | /* |
1294 | * For files that don't have ->read_iter() and ->write_iter(), handle them | |
1295 | * by looping over ->read() or ->write() manually. | |
1296 | */ | |
1297 | static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb, | |
1298 | struct iov_iter *iter) | |
1299 | { | |
1300 | ssize_t ret = 0; | |
1301 | ||
1302 | /* | |
1303 | * Don't support polled IO through this interface, and we can't | |
1304 | * support non-blocking either. For the latter, this just causes | |
1305 | * the kiocb to be handled from an async context. | |
1306 | */ | |
1307 | if (kiocb->ki_flags & IOCB_HIPRI) | |
1308 | return -EOPNOTSUPP; | |
1309 | if (kiocb->ki_flags & IOCB_NOWAIT) | |
1310 | return -EAGAIN; | |
1311 | ||
1312 | while (iov_iter_count(iter)) { | |
1313 | struct iovec iovec = iov_iter_iovec(iter); | |
1314 | ssize_t nr; | |
1315 | ||
1316 | if (rw == READ) { | |
1317 | nr = file->f_op->read(file, iovec.iov_base, | |
1318 | iovec.iov_len, &kiocb->ki_pos); | |
1319 | } else { | |
1320 | nr = file->f_op->write(file, iovec.iov_base, | |
1321 | iovec.iov_len, &kiocb->ki_pos); | |
1322 | } | |
1323 | ||
1324 | if (nr < 0) { | |
1325 | if (!ret) | |
1326 | ret = nr; | |
1327 | break; | |
1328 | } | |
1329 | ret += nr; | |
1330 | if (nr != iovec.iov_len) | |
1331 | break; | |
1332 | iov_iter_advance(iter, nr); | |
1333 | } | |
1334 | ||
1335 | return ret; | |
1336 | } | |
1337 | ||
e0c5c576 | 1338 | static int io_read(struct io_kiocb *req, const struct sqe_submit *s, |
8358e3a8 | 1339 | bool force_nonblock) |
2b188cc1 JA |
1340 | { |
1341 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1342 | struct kiocb *kiocb = &req->rw; | |
1343 | struct iov_iter iter; | |
1344 | struct file *file; | |
31b51510 | 1345 | size_t iov_count; |
9d93a3f5 | 1346 | ssize_t read_size, ret; |
2b188cc1 | 1347 | |
8358e3a8 | 1348 | ret = io_prep_rw(req, s, force_nonblock); |
2b188cc1 JA |
1349 | if (ret) |
1350 | return ret; | |
1351 | file = kiocb->ki_filp; | |
1352 | ||
2b188cc1 | 1353 | if (unlikely(!(file->f_mode & FMODE_READ))) |
09bb8394 | 1354 | return -EBADF; |
2b188cc1 JA |
1355 | |
1356 | ret = io_import_iovec(req->ctx, READ, s, &iovec, &iter); | |
87e5e6da | 1357 | if (ret < 0) |
09bb8394 | 1358 | return ret; |
2b188cc1 | 1359 | |
9d93a3f5 | 1360 | read_size = ret; |
9e645e11 JA |
1361 | if (req->flags & REQ_F_LINK) |
1362 | req->result = read_size; | |
1363 | ||
31b51510 JA |
1364 | iov_count = iov_iter_count(&iter); |
1365 | ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count); | |
2b188cc1 JA |
1366 | if (!ret) { |
1367 | ssize_t ret2; | |
1368 | ||
32960613 JA |
1369 | if (file->f_op->read_iter) |
1370 | ret2 = call_read_iter(file, kiocb, &iter); | |
1371 | else | |
1372 | ret2 = loop_rw_iter(READ, file, kiocb, &iter); | |
1373 | ||
9d93a3f5 JA |
1374 | /* |
1375 | * In case of a short read, punt to async. This can happen | |
1376 | * if we have data partially cached. Alternatively we can | |
1377 | * return the short read, in which case the application will | |
1378 | * need to issue another SQE and wait for it. That SQE will | |
1379 | * need async punt anyway, so it's more efficient to do it | |
1380 | * here. | |
1381 | */ | |
1382 | if (force_nonblock && ret2 > 0 && ret2 < read_size) | |
1383 | ret2 = -EAGAIN; | |
1384 | /* Catch -EAGAIN return for forced non-blocking submission */ | |
31b51510 | 1385 | if (!force_nonblock || ret2 != -EAGAIN) { |
2b188cc1 | 1386 | io_rw_done(kiocb, ret2); |
31b51510 JA |
1387 | } else { |
1388 | /* | |
1389 | * If ->needs_lock is true, we're already in async | |
1390 | * context. | |
1391 | */ | |
1392 | if (!s->needs_lock) | |
1393 | io_async_list_note(READ, req, iov_count); | |
2b188cc1 | 1394 | ret = -EAGAIN; |
31b51510 | 1395 | } |
2b188cc1 JA |
1396 | } |
1397 | kfree(iovec); | |
2b188cc1 JA |
1398 | return ret; |
1399 | } | |
1400 | ||
e0c5c576 | 1401 | static int io_write(struct io_kiocb *req, const struct sqe_submit *s, |
8358e3a8 | 1402 | bool force_nonblock) |
2b188cc1 JA |
1403 | { |
1404 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
1405 | struct kiocb *kiocb = &req->rw; | |
1406 | struct iov_iter iter; | |
1407 | struct file *file; | |
31b51510 | 1408 | size_t iov_count; |
87e5e6da | 1409 | ssize_t ret; |
2b188cc1 | 1410 | |
8358e3a8 | 1411 | ret = io_prep_rw(req, s, force_nonblock); |
2b188cc1 JA |
1412 | if (ret) |
1413 | return ret; | |
2b188cc1 | 1414 | |
2b188cc1 JA |
1415 | file = kiocb->ki_filp; |
1416 | if (unlikely(!(file->f_mode & FMODE_WRITE))) | |
09bb8394 | 1417 | return -EBADF; |
2b188cc1 JA |
1418 | |
1419 | ret = io_import_iovec(req->ctx, WRITE, s, &iovec, &iter); | |
87e5e6da | 1420 | if (ret < 0) |
09bb8394 | 1421 | return ret; |
2b188cc1 | 1422 | |
9e645e11 JA |
1423 | if (req->flags & REQ_F_LINK) |
1424 | req->result = ret; | |
1425 | ||
31b51510 JA |
1426 | iov_count = iov_iter_count(&iter); |
1427 | ||
1428 | ret = -EAGAIN; | |
1429 | if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT)) { | |
1430 | /* If ->needs_lock is true, we're already in async context. */ | |
1431 | if (!s->needs_lock) | |
1432 | io_async_list_note(WRITE, req, iov_count); | |
1433 | goto out_free; | |
1434 | } | |
1435 | ||
1436 | ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count); | |
2b188cc1 | 1437 | if (!ret) { |
9bf7933f RP |
1438 | ssize_t ret2; |
1439 | ||
2b188cc1 JA |
1440 | /* |
1441 | * Open-code file_start_write here to grab freeze protection, | |
1442 | * which will be released by another thread in | |
1443 | * io_complete_rw(). Fool lockdep by telling it the lock got | |
1444 | * released so that it doesn't complain about the held lock when | |
1445 | * we return to userspace. | |
1446 | */ | |
1447 | if (S_ISREG(file_inode(file)->i_mode)) { | |
1448 | __sb_start_write(file_inode(file)->i_sb, | |
1449 | SB_FREEZE_WRITE, true); | |
1450 | __sb_writers_release(file_inode(file)->i_sb, | |
1451 | SB_FREEZE_WRITE); | |
1452 | } | |
1453 | kiocb->ki_flags |= IOCB_WRITE; | |
9bf7933f | 1454 | |
32960613 JA |
1455 | if (file->f_op->write_iter) |
1456 | ret2 = call_write_iter(file, kiocb, &iter); | |
1457 | else | |
1458 | ret2 = loop_rw_iter(WRITE, file, kiocb, &iter); | |
9bf7933f RP |
1459 | if (!force_nonblock || ret2 != -EAGAIN) { |
1460 | io_rw_done(kiocb, ret2); | |
1461 | } else { | |
1462 | /* | |
1463 | * If ->needs_lock is true, we're already in async | |
1464 | * context. | |
1465 | */ | |
1466 | if (!s->needs_lock) | |
1467 | io_async_list_note(WRITE, req, iov_count); | |
1468 | ret = -EAGAIN; | |
1469 | } | |
2b188cc1 | 1470 | } |
31b51510 | 1471 | out_free: |
2b188cc1 | 1472 | kfree(iovec); |
2b188cc1 JA |
1473 | return ret; |
1474 | } | |
1475 | ||
1476 | /* | |
1477 | * IORING_OP_NOP just posts a completion event, nothing else. | |
1478 | */ | |
1479 | static int io_nop(struct io_kiocb *req, u64 user_data) | |
1480 | { | |
1481 | struct io_ring_ctx *ctx = req->ctx; | |
1482 | long err = 0; | |
1483 | ||
def596e9 JA |
1484 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
1485 | return -EINVAL; | |
1486 | ||
c71ffb67 | 1487 | io_cqring_add_event(ctx, user_data, err); |
e65ef56d | 1488 | io_put_req(req); |
2b188cc1 JA |
1489 | return 0; |
1490 | } | |
1491 | ||
c992fe29 CH |
1492 | static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
1493 | { | |
6b06314c | 1494 | struct io_ring_ctx *ctx = req->ctx; |
c992fe29 | 1495 | |
09bb8394 JA |
1496 | if (!req->file) |
1497 | return -EBADF; | |
c992fe29 | 1498 | |
6b06314c | 1499 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
def596e9 | 1500 | return -EINVAL; |
edafccee | 1501 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) |
c992fe29 CH |
1502 | return -EINVAL; |
1503 | ||
c992fe29 CH |
1504 | return 0; |
1505 | } | |
1506 | ||
1507 | static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
1508 | bool force_nonblock) | |
1509 | { | |
1510 | loff_t sqe_off = READ_ONCE(sqe->off); | |
1511 | loff_t sqe_len = READ_ONCE(sqe->len); | |
1512 | loff_t end = sqe_off + sqe_len; | |
1513 | unsigned fsync_flags; | |
1514 | int ret; | |
1515 | ||
1516 | fsync_flags = READ_ONCE(sqe->fsync_flags); | |
1517 | if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC)) | |
1518 | return -EINVAL; | |
1519 | ||
1520 | ret = io_prep_fsync(req, sqe); | |
1521 | if (ret) | |
1522 | return ret; | |
1523 | ||
1524 | /* fsync always requires a blocking context */ | |
1525 | if (force_nonblock) | |
1526 | return -EAGAIN; | |
1527 | ||
1528 | ret = vfs_fsync_range(req->rw.ki_filp, sqe_off, | |
1529 | end > 0 ? end : LLONG_MAX, | |
1530 | fsync_flags & IORING_FSYNC_DATASYNC); | |
1531 | ||
9e645e11 JA |
1532 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1533 | req->flags |= REQ_F_FAIL_LINK; | |
c71ffb67 | 1534 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
e65ef56d | 1535 | io_put_req(req); |
c992fe29 CH |
1536 | return 0; |
1537 | } | |
1538 | ||
5d17b4a4 JA |
1539 | static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
1540 | { | |
1541 | struct io_ring_ctx *ctx = req->ctx; | |
1542 | int ret = 0; | |
1543 | ||
1544 | if (!req->file) | |
1545 | return -EBADF; | |
5d17b4a4 JA |
1546 | |
1547 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
1548 | return -EINVAL; | |
1549 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) | |
1550 | return -EINVAL; | |
1551 | ||
5d17b4a4 JA |
1552 | return ret; |
1553 | } | |
1554 | ||
1555 | static int io_sync_file_range(struct io_kiocb *req, | |
1556 | const struct io_uring_sqe *sqe, | |
1557 | bool force_nonblock) | |
1558 | { | |
1559 | loff_t sqe_off; | |
1560 | loff_t sqe_len; | |
1561 | unsigned flags; | |
1562 | int ret; | |
1563 | ||
1564 | ret = io_prep_sfr(req, sqe); | |
1565 | if (ret) | |
1566 | return ret; | |
1567 | ||
1568 | /* sync_file_range always requires a blocking context */ | |
1569 | if (force_nonblock) | |
1570 | return -EAGAIN; | |
1571 | ||
1572 | sqe_off = READ_ONCE(sqe->off); | |
1573 | sqe_len = READ_ONCE(sqe->len); | |
1574 | flags = READ_ONCE(sqe->sync_range_flags); | |
1575 | ||
1576 | ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags); | |
1577 | ||
9e645e11 JA |
1578 | if (ret < 0 && (req->flags & REQ_F_LINK)) |
1579 | req->flags |= REQ_F_FAIL_LINK; | |
c71ffb67 | 1580 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
5d17b4a4 JA |
1581 | io_put_req(req); |
1582 | return 0; | |
1583 | } | |
1584 | ||
0fa03c62 | 1585 | #if defined(CONFIG_NET) |
aa1fa28f JA |
1586 | static int io_send_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
1587 | bool force_nonblock, | |
1588 | long (*fn)(struct socket *, struct user_msghdr __user *, | |
1589 | unsigned int)) | |
1590 | { | |
0fa03c62 JA |
1591 | struct socket *sock; |
1592 | int ret; | |
1593 | ||
1594 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1595 | return -EINVAL; | |
1596 | ||
1597 | sock = sock_from_file(req->file, &ret); | |
1598 | if (sock) { | |
1599 | struct user_msghdr __user *msg; | |
1600 | unsigned flags; | |
1601 | ||
1602 | flags = READ_ONCE(sqe->msg_flags); | |
1603 | if (flags & MSG_DONTWAIT) | |
1604 | req->flags |= REQ_F_NOWAIT; | |
1605 | else if (force_nonblock) | |
1606 | flags |= MSG_DONTWAIT; | |
1607 | ||
1608 | msg = (struct user_msghdr __user *) (unsigned long) | |
1609 | READ_ONCE(sqe->addr); | |
1610 | ||
aa1fa28f | 1611 | ret = fn(sock, msg, flags); |
0fa03c62 JA |
1612 | if (force_nonblock && ret == -EAGAIN) |
1613 | return ret; | |
1614 | } | |
1615 | ||
c71ffb67 | 1616 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
5d17b4a4 JA |
1617 | io_put_req(req); |
1618 | return 0; | |
1619 | } | |
aa1fa28f JA |
1620 | #endif |
1621 | ||
1622 | static int io_sendmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
1623 | bool force_nonblock) | |
1624 | { | |
1625 | #if defined(CONFIG_NET) | |
1626 | return io_send_recvmsg(req, sqe, force_nonblock, __sys_sendmsg_sock); | |
1627 | #else | |
1628 | return -EOPNOTSUPP; | |
1629 | #endif | |
1630 | } | |
1631 | ||
1632 | static int io_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe, | |
1633 | bool force_nonblock) | |
1634 | { | |
1635 | #if defined(CONFIG_NET) | |
1636 | return io_send_recvmsg(req, sqe, force_nonblock, __sys_recvmsg_sock); | |
0fa03c62 JA |
1637 | #else |
1638 | return -EOPNOTSUPP; | |
1639 | #endif | |
1640 | } | |
5d17b4a4 | 1641 | |
221c5eb2 JA |
1642 | static void io_poll_remove_one(struct io_kiocb *req) |
1643 | { | |
1644 | struct io_poll_iocb *poll = &req->poll; | |
1645 | ||
1646 | spin_lock(&poll->head->lock); | |
1647 | WRITE_ONCE(poll->canceled, true); | |
1648 | if (!list_empty(&poll->wait.entry)) { | |
1649 | list_del_init(&poll->wait.entry); | |
18d9be1a | 1650 | io_queue_async_work(req->ctx, req); |
221c5eb2 JA |
1651 | } |
1652 | spin_unlock(&poll->head->lock); | |
1653 | ||
1654 | list_del_init(&req->list); | |
1655 | } | |
1656 | ||
1657 | static void io_poll_remove_all(struct io_ring_ctx *ctx) | |
1658 | { | |
1659 | struct io_kiocb *req; | |
1660 | ||
1661 | spin_lock_irq(&ctx->completion_lock); | |
1662 | while (!list_empty(&ctx->cancel_list)) { | |
1663 | req = list_first_entry(&ctx->cancel_list, struct io_kiocb,list); | |
1664 | io_poll_remove_one(req); | |
1665 | } | |
1666 | spin_unlock_irq(&ctx->completion_lock); | |
1667 | } | |
1668 | ||
1669 | /* | |
1670 | * Find a running poll command that matches one specified in sqe->addr, | |
1671 | * and remove it if found. | |
1672 | */ | |
1673 | static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
1674 | { | |
1675 | struct io_ring_ctx *ctx = req->ctx; | |
1676 | struct io_kiocb *poll_req, *next; | |
1677 | int ret = -ENOENT; | |
1678 | ||
1679 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1680 | return -EINVAL; | |
1681 | if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index || | |
1682 | sqe->poll_events) | |
1683 | return -EINVAL; | |
1684 | ||
1685 | spin_lock_irq(&ctx->completion_lock); | |
1686 | list_for_each_entry_safe(poll_req, next, &ctx->cancel_list, list) { | |
1687 | if (READ_ONCE(sqe->addr) == poll_req->user_data) { | |
1688 | io_poll_remove_one(poll_req); | |
1689 | ret = 0; | |
1690 | break; | |
1691 | } | |
1692 | } | |
1693 | spin_unlock_irq(&ctx->completion_lock); | |
1694 | ||
c71ffb67 | 1695 | io_cqring_add_event(req->ctx, sqe->user_data, ret); |
e65ef56d | 1696 | io_put_req(req); |
221c5eb2 JA |
1697 | return 0; |
1698 | } | |
1699 | ||
8c838788 JA |
1700 | static void io_poll_complete(struct io_ring_ctx *ctx, struct io_kiocb *req, |
1701 | __poll_t mask) | |
221c5eb2 | 1702 | { |
8c838788 | 1703 | req->poll.done = true; |
c71ffb67 | 1704 | io_cqring_fill_event(ctx, req->user_data, mangle_poll(mask)); |
8c838788 | 1705 | io_commit_cqring(ctx); |
221c5eb2 JA |
1706 | } |
1707 | ||
1708 | static void io_poll_complete_work(struct work_struct *work) | |
1709 | { | |
1710 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
1711 | struct io_poll_iocb *poll = &req->poll; | |
1712 | struct poll_table_struct pt = { ._key = poll->events }; | |
1713 | struct io_ring_ctx *ctx = req->ctx; | |
1714 | __poll_t mask = 0; | |
1715 | ||
1716 | if (!READ_ONCE(poll->canceled)) | |
1717 | mask = vfs_poll(poll->file, &pt) & poll->events; | |
1718 | ||
1719 | /* | |
1720 | * Note that ->ki_cancel callers also delete iocb from active_reqs after | |
1721 | * calling ->ki_cancel. We need the ctx_lock roundtrip here to | |
1722 | * synchronize with them. In the cancellation case the list_del_init | |
1723 | * itself is not actually needed, but harmless so we keep it in to | |
1724 | * avoid further branches in the fast path. | |
1725 | */ | |
1726 | spin_lock_irq(&ctx->completion_lock); | |
1727 | if (!mask && !READ_ONCE(poll->canceled)) { | |
1728 | add_wait_queue(poll->head, &poll->wait); | |
1729 | spin_unlock_irq(&ctx->completion_lock); | |
1730 | return; | |
1731 | } | |
1732 | list_del_init(&req->list); | |
8c838788 | 1733 | io_poll_complete(ctx, req, mask); |
221c5eb2 JA |
1734 | spin_unlock_irq(&ctx->completion_lock); |
1735 | ||
8c838788 JA |
1736 | io_cqring_ev_posted(ctx); |
1737 | io_put_req(req); | |
221c5eb2 JA |
1738 | } |
1739 | ||
1740 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, | |
1741 | void *key) | |
1742 | { | |
1743 | struct io_poll_iocb *poll = container_of(wait, struct io_poll_iocb, | |
1744 | wait); | |
1745 | struct io_kiocb *req = container_of(poll, struct io_kiocb, poll); | |
1746 | struct io_ring_ctx *ctx = req->ctx; | |
1747 | __poll_t mask = key_to_poll(key); | |
8c838788 | 1748 | unsigned long flags; |
221c5eb2 JA |
1749 | |
1750 | /* for instances that support it check for an event match first: */ | |
8c838788 JA |
1751 | if (mask && !(mask & poll->events)) |
1752 | return 0; | |
221c5eb2 | 1753 | |
8c838788 | 1754 | list_del_init(&poll->wait.entry); |
221c5eb2 | 1755 | |
8c838788 JA |
1756 | if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) { |
1757 | list_del(&req->list); | |
1758 | io_poll_complete(ctx, req, mask); | |
1759 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
221c5eb2 | 1760 | |
8c838788 JA |
1761 | io_cqring_ev_posted(ctx); |
1762 | io_put_req(req); | |
1763 | } else { | |
18d9be1a | 1764 | io_queue_async_work(ctx, req); |
221c5eb2 JA |
1765 | } |
1766 | ||
221c5eb2 JA |
1767 | return 1; |
1768 | } | |
1769 | ||
1770 | struct io_poll_table { | |
1771 | struct poll_table_struct pt; | |
1772 | struct io_kiocb *req; | |
1773 | int error; | |
1774 | }; | |
1775 | ||
1776 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, | |
1777 | struct poll_table_struct *p) | |
1778 | { | |
1779 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
1780 | ||
1781 | if (unlikely(pt->req->poll.head)) { | |
1782 | pt->error = -EINVAL; | |
1783 | return; | |
1784 | } | |
1785 | ||
1786 | pt->error = 0; | |
1787 | pt->req->poll.head = head; | |
1788 | add_wait_queue(head, &pt->req->poll.wait); | |
1789 | } | |
1790 | ||
1791 | static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
1792 | { | |
1793 | struct io_poll_iocb *poll = &req->poll; | |
1794 | struct io_ring_ctx *ctx = req->ctx; | |
1795 | struct io_poll_table ipt; | |
8c838788 | 1796 | bool cancel = false; |
221c5eb2 JA |
1797 | __poll_t mask; |
1798 | u16 events; | |
221c5eb2 JA |
1799 | |
1800 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1801 | return -EINVAL; | |
1802 | if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index) | |
1803 | return -EINVAL; | |
09bb8394 JA |
1804 | if (!poll->file) |
1805 | return -EBADF; | |
221c5eb2 | 1806 | |
6cc47d1d | 1807 | req->submit.sqe = NULL; |
221c5eb2 JA |
1808 | INIT_WORK(&req->work, io_poll_complete_work); |
1809 | events = READ_ONCE(sqe->poll_events); | |
1810 | poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP; | |
1811 | ||
221c5eb2 | 1812 | poll->head = NULL; |
8c838788 | 1813 | poll->done = false; |
221c5eb2 JA |
1814 | poll->canceled = false; |
1815 | ||
1816 | ipt.pt._qproc = io_poll_queue_proc; | |
1817 | ipt.pt._key = poll->events; | |
1818 | ipt.req = req; | |
1819 | ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */ | |
1820 | ||
1821 | /* initialized the list so that we can do list_empty checks */ | |
1822 | INIT_LIST_HEAD(&poll->wait.entry); | |
1823 | init_waitqueue_func_entry(&poll->wait, io_poll_wake); | |
1824 | ||
36703247 JA |
1825 | INIT_LIST_HEAD(&req->list); |
1826 | ||
221c5eb2 | 1827 | mask = vfs_poll(poll->file, &ipt.pt) & poll->events; |
221c5eb2 JA |
1828 | |
1829 | spin_lock_irq(&ctx->completion_lock); | |
8c838788 JA |
1830 | if (likely(poll->head)) { |
1831 | spin_lock(&poll->head->lock); | |
1832 | if (unlikely(list_empty(&poll->wait.entry))) { | |
1833 | if (ipt.error) | |
1834 | cancel = true; | |
1835 | ipt.error = 0; | |
1836 | mask = 0; | |
1837 | } | |
1838 | if (mask || ipt.error) | |
1839 | list_del_init(&poll->wait.entry); | |
1840 | else if (cancel) | |
1841 | WRITE_ONCE(poll->canceled, true); | |
1842 | else if (!poll->done) /* actually waiting for an event */ | |
1843 | list_add_tail(&req->list, &ctx->cancel_list); | |
1844 | spin_unlock(&poll->head->lock); | |
1845 | } | |
1846 | if (mask) { /* no async, we'd stolen it */ | |
221c5eb2 | 1847 | ipt.error = 0; |
8c838788 | 1848 | io_poll_complete(ctx, req, mask); |
221c5eb2 | 1849 | } |
221c5eb2 JA |
1850 | spin_unlock_irq(&ctx->completion_lock); |
1851 | ||
8c838788 JA |
1852 | if (mask) { |
1853 | io_cqring_ev_posted(ctx); | |
e65ef56d | 1854 | io_put_req(req); |
221c5eb2 | 1855 | } |
8c838788 | 1856 | return ipt.error; |
221c5eb2 JA |
1857 | } |
1858 | ||
5262f567 JA |
1859 | static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer) |
1860 | { | |
1861 | struct io_ring_ctx *ctx; | |
1862 | struct io_kiocb *req; | |
1863 | unsigned long flags; | |
1864 | ||
1865 | req = container_of(timer, struct io_kiocb, timeout.timer); | |
1866 | ctx = req->ctx; | |
1867 | atomic_inc(&ctx->cq_timeouts); | |
1868 | ||
1869 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
1870 | list_del(&req->list); | |
1871 | ||
1872 | io_cqring_fill_event(ctx, req->user_data, -ETIME); | |
1873 | io_commit_cqring(ctx); | |
1874 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
1875 | ||
1876 | io_cqring_ev_posted(ctx); | |
1877 | ||
1878 | io_put_req(req); | |
1879 | return HRTIMER_NORESTART; | |
1880 | } | |
1881 | ||
1882 | static int io_timeout(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
1883 | { | |
1884 | unsigned count, req_dist, tail_index; | |
1885 | struct io_ring_ctx *ctx = req->ctx; | |
1886 | struct list_head *entry; | |
bdf20073 | 1887 | struct timespec64 ts; |
5262f567 JA |
1888 | |
1889 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
1890 | return -EINVAL; | |
1891 | if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->timeout_flags || | |
1892 | sqe->len != 1) | |
1893 | return -EINVAL; | |
bdf20073 AB |
1894 | |
1895 | if (get_timespec64(&ts, u64_to_user_ptr(sqe->addr))) | |
5262f567 JA |
1896 | return -EFAULT; |
1897 | ||
1898 | /* | |
1899 | * sqe->off holds how many events that need to occur for this | |
1900 | * timeout event to be satisfied. | |
1901 | */ | |
1902 | count = READ_ONCE(sqe->off); | |
1903 | if (!count) | |
1904 | count = 1; | |
1905 | ||
1906 | req->sequence = ctx->cached_sq_head + count - 1; | |
1907 | req->flags |= REQ_F_TIMEOUT; | |
1908 | ||
1909 | /* | |
1910 | * Insertion sort, ensuring the first entry in the list is always | |
1911 | * the one we need first. | |
1912 | */ | |
1913 | tail_index = ctx->cached_cq_tail - ctx->rings->sq_dropped; | |
1914 | req_dist = req->sequence - tail_index; | |
1915 | spin_lock_irq(&ctx->completion_lock); | |
1916 | list_for_each_prev(entry, &ctx->timeout_list) { | |
1917 | struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list); | |
1918 | unsigned dist; | |
1919 | ||
1920 | dist = nxt->sequence - tail_index; | |
1921 | if (req_dist >= dist) | |
1922 | break; | |
1923 | } | |
1924 | list_add(&req->list, entry); | |
1925 | spin_unlock_irq(&ctx->completion_lock); | |
1926 | ||
1927 | hrtimer_init(&req->timeout.timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
1928 | req->timeout.timer.function = io_timeout_fn; | |
bdf20073 | 1929 | hrtimer_start(&req->timeout.timer, timespec64_to_ktime(ts), |
5262f567 JA |
1930 | HRTIMER_MODE_REL); |
1931 | return 0; | |
1932 | } | |
1933 | ||
de0617e4 JA |
1934 | static int io_req_defer(struct io_ring_ctx *ctx, struct io_kiocb *req, |
1935 | const struct io_uring_sqe *sqe) | |
1936 | { | |
1937 | struct io_uring_sqe *sqe_copy; | |
1938 | ||
1939 | if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list)) | |
1940 | return 0; | |
1941 | ||
1942 | sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL); | |
1943 | if (!sqe_copy) | |
1944 | return -EAGAIN; | |
1945 | ||
1946 | spin_lock_irq(&ctx->completion_lock); | |
1947 | if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list)) { | |
1948 | spin_unlock_irq(&ctx->completion_lock); | |
1949 | kfree(sqe_copy); | |
1950 | return 0; | |
1951 | } | |
1952 | ||
1953 | memcpy(sqe_copy, sqe, sizeof(*sqe_copy)); | |
1954 | req->submit.sqe = sqe_copy; | |
1955 | ||
1956 | INIT_WORK(&req->work, io_sq_wq_submit_work); | |
1957 | list_add_tail(&req->list, &ctx->defer_list); | |
1958 | spin_unlock_irq(&ctx->completion_lock); | |
1959 | return -EIOCBQUEUED; | |
1960 | } | |
1961 | ||
2b188cc1 | 1962 | static int __io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, |
8358e3a8 | 1963 | const struct sqe_submit *s, bool force_nonblock) |
2b188cc1 | 1964 | { |
e0c5c576 | 1965 | int ret, opcode; |
2b188cc1 | 1966 | |
9e645e11 JA |
1967 | req->user_data = READ_ONCE(s->sqe->user_data); |
1968 | ||
2b188cc1 JA |
1969 | if (unlikely(s->index >= ctx->sq_entries)) |
1970 | return -EINVAL; | |
2b188cc1 JA |
1971 | |
1972 | opcode = READ_ONCE(s->sqe->opcode); | |
1973 | switch (opcode) { | |
1974 | case IORING_OP_NOP: | |
1975 | ret = io_nop(req, req->user_data); | |
1976 | break; | |
1977 | case IORING_OP_READV: | |
edafccee JA |
1978 | if (unlikely(s->sqe->buf_index)) |
1979 | return -EINVAL; | |
8358e3a8 | 1980 | ret = io_read(req, s, force_nonblock); |
2b188cc1 JA |
1981 | break; |
1982 | case IORING_OP_WRITEV: | |
edafccee JA |
1983 | if (unlikely(s->sqe->buf_index)) |
1984 | return -EINVAL; | |
8358e3a8 | 1985 | ret = io_write(req, s, force_nonblock); |
edafccee JA |
1986 | break; |
1987 | case IORING_OP_READ_FIXED: | |
8358e3a8 | 1988 | ret = io_read(req, s, force_nonblock); |
edafccee JA |
1989 | break; |
1990 | case IORING_OP_WRITE_FIXED: | |
8358e3a8 | 1991 | ret = io_write(req, s, force_nonblock); |
2b188cc1 | 1992 | break; |
c992fe29 CH |
1993 | case IORING_OP_FSYNC: |
1994 | ret = io_fsync(req, s->sqe, force_nonblock); | |
1995 | break; | |
221c5eb2 JA |
1996 | case IORING_OP_POLL_ADD: |
1997 | ret = io_poll_add(req, s->sqe); | |
1998 | break; | |
1999 | case IORING_OP_POLL_REMOVE: | |
2000 | ret = io_poll_remove(req, s->sqe); | |
2001 | break; | |
5d17b4a4 JA |
2002 | case IORING_OP_SYNC_FILE_RANGE: |
2003 | ret = io_sync_file_range(req, s->sqe, force_nonblock); | |
2004 | break; | |
0fa03c62 JA |
2005 | case IORING_OP_SENDMSG: |
2006 | ret = io_sendmsg(req, s->sqe, force_nonblock); | |
2007 | break; | |
aa1fa28f JA |
2008 | case IORING_OP_RECVMSG: |
2009 | ret = io_recvmsg(req, s->sqe, force_nonblock); | |
2010 | break; | |
5262f567 JA |
2011 | case IORING_OP_TIMEOUT: |
2012 | ret = io_timeout(req, s->sqe); | |
2013 | break; | |
2b188cc1 JA |
2014 | default: |
2015 | ret = -EINVAL; | |
2016 | break; | |
2017 | } | |
2018 | ||
def596e9 JA |
2019 | if (ret) |
2020 | return ret; | |
2021 | ||
2022 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
9e645e11 | 2023 | if (req->result == -EAGAIN) |
def596e9 JA |
2024 | return -EAGAIN; |
2025 | ||
2026 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
2027 | if (s->needs_lock) | |
2028 | mutex_lock(&ctx->uring_lock); | |
2029 | io_iopoll_req_issued(req); | |
2030 | if (s->needs_lock) | |
2031 | mutex_unlock(&ctx->uring_lock); | |
2032 | } | |
2033 | ||
2034 | return 0; | |
2b188cc1 JA |
2035 | } |
2036 | ||
31b51510 JA |
2037 | static struct async_list *io_async_list_from_sqe(struct io_ring_ctx *ctx, |
2038 | const struct io_uring_sqe *sqe) | |
2039 | { | |
2040 | switch (sqe->opcode) { | |
2041 | case IORING_OP_READV: | |
2042 | case IORING_OP_READ_FIXED: | |
2043 | return &ctx->pending_async[READ]; | |
2044 | case IORING_OP_WRITEV: | |
2045 | case IORING_OP_WRITE_FIXED: | |
2046 | return &ctx->pending_async[WRITE]; | |
2047 | default: | |
2048 | return NULL; | |
2049 | } | |
2050 | } | |
2051 | ||
edafccee JA |
2052 | static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe) |
2053 | { | |
2054 | u8 opcode = READ_ONCE(sqe->opcode); | |
2055 | ||
2056 | return !(opcode == IORING_OP_READ_FIXED || | |
2057 | opcode == IORING_OP_WRITE_FIXED); | |
2058 | } | |
2059 | ||
2b188cc1 JA |
2060 | static void io_sq_wq_submit_work(struct work_struct *work) |
2061 | { | |
2062 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
2b188cc1 | 2063 | struct io_ring_ctx *ctx = req->ctx; |
31b51510 JA |
2064 | struct mm_struct *cur_mm = NULL; |
2065 | struct async_list *async_list; | |
2066 | LIST_HEAD(req_list); | |
edafccee | 2067 | mm_segment_t old_fs; |
2b188cc1 JA |
2068 | int ret; |
2069 | ||
31b51510 JA |
2070 | async_list = io_async_list_from_sqe(ctx, req->submit.sqe); |
2071 | restart: | |
2072 | do { | |
2073 | struct sqe_submit *s = &req->submit; | |
2074 | const struct io_uring_sqe *sqe = s->sqe; | |
d0ee8791 | 2075 | unsigned int flags = req->flags; |
2b188cc1 | 2076 | |
8449eeda | 2077 | /* Ensure we clear previously set non-block flag */ |
31b51510 JA |
2078 | req->rw.ki_flags &= ~IOCB_NOWAIT; |
2079 | ||
2080 | ret = 0; | |
2081 | if (io_sqe_needs_user(sqe) && !cur_mm) { | |
2082 | if (!mmget_not_zero(ctx->sqo_mm)) { | |
2083 | ret = -EFAULT; | |
2084 | } else { | |
2085 | cur_mm = ctx->sqo_mm; | |
2086 | use_mm(cur_mm); | |
2087 | old_fs = get_fs(); | |
2088 | set_fs(USER_DS); | |
2089 | } | |
2090 | } | |
2091 | ||
2092 | if (!ret) { | |
2093 | s->has_user = cur_mm != NULL; | |
2094 | s->needs_lock = true; | |
2095 | do { | |
8358e3a8 | 2096 | ret = __io_submit_sqe(ctx, req, s, false); |
31b51510 JA |
2097 | /* |
2098 | * We can get EAGAIN for polled IO even though | |
2099 | * we're forcing a sync submission from here, | |
2100 | * since we can't wait for request slots on the | |
2101 | * block side. | |
2102 | */ | |
2103 | if (ret != -EAGAIN) | |
2104 | break; | |
2105 | cond_resched(); | |
2106 | } while (1); | |
2107 | } | |
817869d2 JA |
2108 | |
2109 | /* drop submission reference */ | |
2110 | io_put_req(req); | |
2111 | ||
31b51510 | 2112 | if (ret) { |
c71ffb67 | 2113 | io_cqring_add_event(ctx, sqe->user_data, ret); |
e65ef56d | 2114 | io_put_req(req); |
31b51510 JA |
2115 | } |
2116 | ||
2117 | /* async context always use a copy of the sqe */ | |
2118 | kfree(sqe); | |
2119 | ||
f7b76ac9 | 2120 | /* req from defer and link list needn't decrease async cnt */ |
d0ee8791 | 2121 | if (flags & (REQ_F_IO_DRAINED | REQ_F_LINK_DONE)) |
f7b76ac9 ZL |
2122 | goto out; |
2123 | ||
31b51510 JA |
2124 | if (!async_list) |
2125 | break; | |
2126 | if (!list_empty(&req_list)) { | |
2127 | req = list_first_entry(&req_list, struct io_kiocb, | |
2128 | list); | |
2129 | list_del(&req->list); | |
2130 | continue; | |
2131 | } | |
2132 | if (list_empty(&async_list->list)) | |
2133 | break; | |
2134 | ||
2135 | req = NULL; | |
2136 | spin_lock(&async_list->lock); | |
2137 | if (list_empty(&async_list->list)) { | |
2138 | spin_unlock(&async_list->lock); | |
2139 | break; | |
2140 | } | |
2141 | list_splice_init(&async_list->list, &req_list); | |
2142 | spin_unlock(&async_list->lock); | |
2143 | ||
2144 | req = list_first_entry(&req_list, struct io_kiocb, list); | |
2145 | list_del(&req->list); | |
2146 | } while (req); | |
edafccee JA |
2147 | |
2148 | /* | |
31b51510 JA |
2149 | * Rare case of racing with a submitter. If we find the count has |
2150 | * dropped to zero AND we have pending work items, then restart | |
2151 | * the processing. This is a tiny race window. | |
edafccee | 2152 | */ |
31b51510 JA |
2153 | if (async_list) { |
2154 | ret = atomic_dec_return(&async_list->cnt); | |
2155 | while (!ret && !list_empty(&async_list->list)) { | |
2156 | spin_lock(&async_list->lock); | |
2157 | atomic_inc(&async_list->cnt); | |
2158 | list_splice_init(&async_list->list, &req_list); | |
2159 | spin_unlock(&async_list->lock); | |
2160 | ||
2161 | if (!list_empty(&req_list)) { | |
2162 | req = list_first_entry(&req_list, | |
2163 | struct io_kiocb, list); | |
2164 | list_del(&req->list); | |
2165 | goto restart; | |
2166 | } | |
2167 | ret = atomic_dec_return(&async_list->cnt); | |
edafccee | 2168 | } |
edafccee | 2169 | } |
2b188cc1 | 2170 | |
f7b76ac9 | 2171 | out: |
31b51510 | 2172 | if (cur_mm) { |
edafccee | 2173 | set_fs(old_fs); |
31b51510 JA |
2174 | unuse_mm(cur_mm); |
2175 | mmput(cur_mm); | |
2b188cc1 | 2176 | } |
31b51510 | 2177 | } |
2b188cc1 | 2178 | |
31b51510 JA |
2179 | /* |
2180 | * See if we can piggy back onto previously submitted work, that is still | |
2181 | * running. We currently only allow this if the new request is sequential | |
2182 | * to the previous one we punted. | |
2183 | */ | |
2184 | static bool io_add_to_prev_work(struct async_list *list, struct io_kiocb *req) | |
2185 | { | |
6d5d5ac5 | 2186 | bool ret; |
31b51510 JA |
2187 | |
2188 | if (!list) | |
2189 | return false; | |
2190 | if (!(req->flags & REQ_F_SEQ_PREV)) | |
2191 | return false; | |
2192 | if (!atomic_read(&list->cnt)) | |
2193 | return false; | |
2194 | ||
2195 | ret = true; | |
2196 | spin_lock(&list->lock); | |
2197 | list_add_tail(&req->list, &list->list); | |
c0e48f9d ZL |
2198 | /* |
2199 | * Ensure we see a simultaneous modification from io_sq_wq_submit_work() | |
2200 | */ | |
2201 | smp_mb(); | |
31b51510 JA |
2202 | if (!atomic_read(&list->cnt)) { |
2203 | list_del_init(&req->list); | |
2204 | ret = false; | |
2205 | } | |
2206 | spin_unlock(&list->lock); | |
2207 | return ret; | |
2b188cc1 JA |
2208 | } |
2209 | ||
09bb8394 JA |
2210 | static bool io_op_needs_file(const struct io_uring_sqe *sqe) |
2211 | { | |
2212 | int op = READ_ONCE(sqe->opcode); | |
2213 | ||
2214 | switch (op) { | |
2215 | case IORING_OP_NOP: | |
2216 | case IORING_OP_POLL_REMOVE: | |
2217 | return false; | |
2218 | default: | |
2219 | return true; | |
2220 | } | |
2221 | } | |
2222 | ||
2223 | static int io_req_set_file(struct io_ring_ctx *ctx, const struct sqe_submit *s, | |
2224 | struct io_submit_state *state, struct io_kiocb *req) | |
2225 | { | |
2226 | unsigned flags; | |
2227 | int fd; | |
2228 | ||
2229 | flags = READ_ONCE(s->sqe->flags); | |
2230 | fd = READ_ONCE(s->sqe->fd); | |
2231 | ||
4fe2c963 | 2232 | if (flags & IOSQE_IO_DRAIN) |
de0617e4 | 2233 | req->flags |= REQ_F_IO_DRAIN; |
4fe2c963 JL |
2234 | /* |
2235 | * All io need record the previous position, if LINK vs DARIN, | |
2236 | * it can be used to mark the position of the first IO in the | |
2237 | * link list. | |
2238 | */ | |
2239 | req->sequence = s->sequence; | |
de0617e4 | 2240 | |
60c112b0 | 2241 | if (!io_op_needs_file(s->sqe)) |
09bb8394 | 2242 | return 0; |
09bb8394 JA |
2243 | |
2244 | if (flags & IOSQE_FIXED_FILE) { | |
2245 | if (unlikely(!ctx->user_files || | |
2246 | (unsigned) fd >= ctx->nr_user_files)) | |
2247 | return -EBADF; | |
2248 | req->file = ctx->user_files[fd]; | |
2249 | req->flags |= REQ_F_FIXED_FILE; | |
2250 | } else { | |
2251 | if (s->needs_fixed_file) | |
2252 | return -EBADF; | |
2253 | req->file = io_file_get(state, fd); | |
2254 | if (unlikely(!req->file)) | |
2255 | return -EBADF; | |
2256 | } | |
2257 | ||
2258 | return 0; | |
2259 | } | |
2260 | ||
4fe2c963 | 2261 | static int __io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, |
c5766668 | 2262 | struct sqe_submit *s, bool force_nonblock) |
2b188cc1 | 2263 | { |
e0c5c576 | 2264 | int ret; |
2b188cc1 | 2265 | |
c5766668 | 2266 | ret = __io_submit_sqe(ctx, req, s, force_nonblock); |
8449eeda | 2267 | if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) { |
2b188cc1 JA |
2268 | struct io_uring_sqe *sqe_copy; |
2269 | ||
954dab19 | 2270 | sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL); |
2b188cc1 | 2271 | if (sqe_copy) { |
31b51510 JA |
2272 | struct async_list *list; |
2273 | ||
2b188cc1 | 2274 | s->sqe = sqe_copy; |
2b188cc1 | 2275 | memcpy(&req->submit, s, sizeof(*s)); |
31b51510 JA |
2276 | list = io_async_list_from_sqe(ctx, s->sqe); |
2277 | if (!io_add_to_prev_work(list, req)) { | |
2278 | if (list) | |
2279 | atomic_inc(&list->cnt); | |
2280 | INIT_WORK(&req->work, io_sq_wq_submit_work); | |
18d9be1a | 2281 | io_queue_async_work(ctx, req); |
31b51510 | 2282 | } |
e65ef56d JA |
2283 | |
2284 | /* | |
2285 | * Queued up for async execution, worker will release | |
9e645e11 | 2286 | * submit reference when the iocb is actually submitted. |
e65ef56d JA |
2287 | */ |
2288 | return 0; | |
2b188cc1 JA |
2289 | } |
2290 | } | |
e65ef56d JA |
2291 | |
2292 | /* drop submission reference */ | |
2293 | io_put_req(req); | |
2294 | ||
2295 | /* and drop final reference, if we failed */ | |
9e645e11 JA |
2296 | if (ret) { |
2297 | io_cqring_add_event(ctx, req->user_data, ret); | |
2298 | if (req->flags & REQ_F_LINK) | |
2299 | req->flags |= REQ_F_FAIL_LINK; | |
e65ef56d | 2300 | io_put_req(req); |
9e645e11 | 2301 | } |
2b188cc1 JA |
2302 | |
2303 | return ret; | |
2304 | } | |
2305 | ||
4fe2c963 | 2306 | static int io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, |
c5766668 | 2307 | struct sqe_submit *s, bool force_nonblock) |
4fe2c963 JL |
2308 | { |
2309 | int ret; | |
2310 | ||
2311 | ret = io_req_defer(ctx, req, s->sqe); | |
2312 | if (ret) { | |
2313 | if (ret != -EIOCBQUEUED) { | |
2314 | io_free_req(req); | |
2315 | io_cqring_add_event(ctx, s->sqe->user_data, ret); | |
2316 | } | |
2317 | return 0; | |
2318 | } | |
2319 | ||
c5766668 | 2320 | return __io_queue_sqe(ctx, req, s, force_nonblock); |
4fe2c963 JL |
2321 | } |
2322 | ||
2323 | static int io_queue_link_head(struct io_ring_ctx *ctx, struct io_kiocb *req, | |
c5766668 JA |
2324 | struct sqe_submit *s, struct io_kiocb *shadow, |
2325 | bool force_nonblock) | |
4fe2c963 JL |
2326 | { |
2327 | int ret; | |
2328 | int need_submit = false; | |
2329 | ||
2330 | if (!shadow) | |
c5766668 | 2331 | return io_queue_sqe(ctx, req, s, force_nonblock); |
4fe2c963 JL |
2332 | |
2333 | /* | |
2334 | * Mark the first IO in link list as DRAIN, let all the following | |
2335 | * IOs enter the defer list. all IO needs to be completed before link | |
2336 | * list. | |
2337 | */ | |
2338 | req->flags |= REQ_F_IO_DRAIN; | |
2339 | ret = io_req_defer(ctx, req, s->sqe); | |
2340 | if (ret) { | |
2341 | if (ret != -EIOCBQUEUED) { | |
2342 | io_free_req(req); | |
2343 | io_cqring_add_event(ctx, s->sqe->user_data, ret); | |
2344 | return 0; | |
2345 | } | |
2346 | } else { | |
2347 | /* | |
2348 | * If ret == 0 means that all IOs in front of link io are | |
2349 | * running done. let's queue link head. | |
2350 | */ | |
2351 | need_submit = true; | |
2352 | } | |
2353 | ||
2354 | /* Insert shadow req to defer_list, blocking next IOs */ | |
2355 | spin_lock_irq(&ctx->completion_lock); | |
2356 | list_add_tail(&shadow->list, &ctx->defer_list); | |
2357 | spin_unlock_irq(&ctx->completion_lock); | |
2358 | ||
2359 | if (need_submit) | |
c5766668 | 2360 | return __io_queue_sqe(ctx, req, s, force_nonblock); |
4fe2c963 JL |
2361 | |
2362 | return 0; | |
2363 | } | |
2364 | ||
9e645e11 JA |
2365 | #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK) |
2366 | ||
2367 | static void io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s, | |
c5766668 JA |
2368 | struct io_submit_state *state, struct io_kiocb **link, |
2369 | bool force_nonblock) | |
9e645e11 JA |
2370 | { |
2371 | struct io_uring_sqe *sqe_copy; | |
2372 | struct io_kiocb *req; | |
2373 | int ret; | |
2374 | ||
2375 | /* enforce forwards compatibility on users */ | |
2376 | if (unlikely(s->sqe->flags & ~SQE_VALID_FLAGS)) { | |
2377 | ret = -EINVAL; | |
2378 | goto err; | |
2379 | } | |
2380 | ||
2381 | req = io_get_req(ctx, state); | |
2382 | if (unlikely(!req)) { | |
2383 | ret = -EAGAIN; | |
2384 | goto err; | |
2385 | } | |
2386 | ||
2387 | ret = io_req_set_file(ctx, s, state, req); | |
2388 | if (unlikely(ret)) { | |
2389 | err_req: | |
2390 | io_free_req(req); | |
2391 | err: | |
2392 | io_cqring_add_event(ctx, s->sqe->user_data, ret); | |
2393 | return; | |
2394 | } | |
2395 | ||
9e645e11 JA |
2396 | /* |
2397 | * If we already have a head request, queue this one for async | |
2398 | * submittal once the head completes. If we don't have a head but | |
2399 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
2400 | * submitted sync once the chain is complete. If none of those | |
2401 | * conditions are true (normal request), then just queue it. | |
2402 | */ | |
2403 | if (*link) { | |
2404 | struct io_kiocb *prev = *link; | |
2405 | ||
2406 | sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL); | |
2407 | if (!sqe_copy) { | |
2408 | ret = -EAGAIN; | |
2409 | goto err_req; | |
2410 | } | |
2411 | ||
2412 | s->sqe = sqe_copy; | |
2413 | memcpy(&req->submit, s, sizeof(*s)); | |
2414 | list_add_tail(&req->list, &prev->link_list); | |
2415 | } else if (s->sqe->flags & IOSQE_IO_LINK) { | |
2416 | req->flags |= REQ_F_LINK; | |
2417 | ||
2418 | memcpy(&req->submit, s, sizeof(*s)); | |
2419 | INIT_LIST_HEAD(&req->link_list); | |
2420 | *link = req; | |
2421 | } else { | |
c5766668 | 2422 | io_queue_sqe(ctx, req, s, force_nonblock); |
9e645e11 JA |
2423 | } |
2424 | } | |
2425 | ||
9a56a232 JA |
2426 | /* |
2427 | * Batched submission is done, ensure local IO is flushed out. | |
2428 | */ | |
2429 | static void io_submit_state_end(struct io_submit_state *state) | |
2430 | { | |
2431 | blk_finish_plug(&state->plug); | |
3d6770fb | 2432 | io_file_put(state); |
2579f913 JA |
2433 | if (state->free_reqs) |
2434 | kmem_cache_free_bulk(req_cachep, state->free_reqs, | |
2435 | &state->reqs[state->cur_req]); | |
9a56a232 JA |
2436 | } |
2437 | ||
2438 | /* | |
2439 | * Start submission side cache. | |
2440 | */ | |
2441 | static void io_submit_state_start(struct io_submit_state *state, | |
2442 | struct io_ring_ctx *ctx, unsigned max_ios) | |
2443 | { | |
2444 | blk_start_plug(&state->plug); | |
2579f913 | 2445 | state->free_reqs = 0; |
9a56a232 JA |
2446 | state->file = NULL; |
2447 | state->ios_left = max_ios; | |
2448 | } | |
2449 | ||
2b188cc1 JA |
2450 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
2451 | { | |
75b28aff | 2452 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 2453 | |
75b28aff | 2454 | if (ctx->cached_sq_head != READ_ONCE(rings->sq.head)) { |
2b188cc1 JA |
2455 | /* |
2456 | * Ensure any loads from the SQEs are done at this point, | |
2457 | * since once we write the new head, the application could | |
2458 | * write new data to them. | |
2459 | */ | |
75b28aff | 2460 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); |
2b188cc1 JA |
2461 | } |
2462 | } | |
2463 | ||
2b188cc1 JA |
2464 | /* |
2465 | * Fetch an sqe, if one is available. Note that s->sqe will point to memory | |
2466 | * that is mapped by userspace. This means that care needs to be taken to | |
2467 | * ensure that reads are stable, as we cannot rely on userspace always | |
2468 | * being a good citizen. If members of the sqe are validated and then later | |
2469 | * used, it's important that those reads are done through READ_ONCE() to | |
2470 | * prevent a re-load down the line. | |
2471 | */ | |
2472 | static bool io_get_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s) | |
2473 | { | |
75b28aff HV |
2474 | struct io_rings *rings = ctx->rings; |
2475 | u32 *sq_array = ctx->sq_array; | |
2b188cc1 JA |
2476 | unsigned head; |
2477 | ||
2478 | /* | |
2479 | * The cached sq head (or cq tail) serves two purposes: | |
2480 | * | |
2481 | * 1) allows us to batch the cost of updating the user visible | |
2482 | * head updates. | |
2483 | * 2) allows the kernel side to track the head on its own, even | |
2484 | * though the application is the one updating it. | |
2485 | */ | |
2486 | head = ctx->cached_sq_head; | |
e523a29c | 2487 | /* make sure SQ entry isn't read before tail */ |
75b28aff | 2488 | if (head == smp_load_acquire(&rings->sq.tail)) |
2b188cc1 JA |
2489 | return false; |
2490 | ||
75b28aff | 2491 | head = READ_ONCE(sq_array[head & ctx->sq_mask]); |
2b188cc1 JA |
2492 | if (head < ctx->sq_entries) { |
2493 | s->index = head; | |
2494 | s->sqe = &ctx->sq_sqes[head]; | |
8776f3fa | 2495 | s->sequence = ctx->cached_sq_head; |
2b188cc1 JA |
2496 | ctx->cached_sq_head++; |
2497 | return true; | |
2498 | } | |
2499 | ||
2500 | /* drop invalid entries */ | |
2501 | ctx->cached_sq_head++; | |
75b28aff | 2502 | rings->sq_dropped++; |
2b188cc1 JA |
2503 | return false; |
2504 | } | |
2505 | ||
6c271ce2 JA |
2506 | static int io_submit_sqes(struct io_ring_ctx *ctx, struct sqe_submit *sqes, |
2507 | unsigned int nr, bool has_user, bool mm_fault) | |
2508 | { | |
2509 | struct io_submit_state state, *statep = NULL; | |
9e645e11 | 2510 | struct io_kiocb *link = NULL; |
4fe2c963 | 2511 | struct io_kiocb *shadow_req = NULL; |
9e645e11 JA |
2512 | bool prev_was_link = false; |
2513 | int i, submitted = 0; | |
6c271ce2 JA |
2514 | |
2515 | if (nr > IO_PLUG_THRESHOLD) { | |
2516 | io_submit_state_start(&state, ctx, nr); | |
2517 | statep = &state; | |
2518 | } | |
2519 | ||
2520 | for (i = 0; i < nr; i++) { | |
9e645e11 JA |
2521 | /* |
2522 | * If previous wasn't linked and we have a linked command, | |
2523 | * that's the end of the chain. Submit the previous link. | |
2524 | */ | |
2525 | if (!prev_was_link && link) { | |
c5766668 JA |
2526 | io_queue_link_head(ctx, link, &link->submit, shadow_req, |
2527 | true); | |
9e645e11 | 2528 | link = NULL; |
5f5ad9ce | 2529 | shadow_req = NULL; |
9e645e11 JA |
2530 | } |
2531 | prev_was_link = (sqes[i].sqe->flags & IOSQE_IO_LINK) != 0; | |
2532 | ||
4fe2c963 JL |
2533 | if (link && (sqes[i].sqe->flags & IOSQE_IO_DRAIN)) { |
2534 | if (!shadow_req) { | |
2535 | shadow_req = io_get_req(ctx, NULL); | |
a1041c27 JL |
2536 | if (unlikely(!shadow_req)) |
2537 | goto out; | |
4fe2c963 JL |
2538 | shadow_req->flags |= (REQ_F_IO_DRAIN | REQ_F_SHADOW_DRAIN); |
2539 | refcount_dec(&shadow_req->refs); | |
2540 | } | |
2541 | shadow_req->sequence = sqes[i].sequence; | |
2542 | } | |
2543 | ||
a1041c27 | 2544 | out: |
6c271ce2 | 2545 | if (unlikely(mm_fault)) { |
9e645e11 JA |
2546 | io_cqring_add_event(ctx, sqes[i].sqe->user_data, |
2547 | -EFAULT); | |
6c271ce2 JA |
2548 | } else { |
2549 | sqes[i].has_user = has_user; | |
2550 | sqes[i].needs_lock = true; | |
2551 | sqes[i].needs_fixed_file = true; | |
c5766668 | 2552 | io_submit_sqe(ctx, &sqes[i], statep, &link, true); |
6c271ce2 | 2553 | submitted++; |
6c271ce2 | 2554 | } |
6c271ce2 JA |
2555 | } |
2556 | ||
9e645e11 | 2557 | if (link) |
c5766668 | 2558 | io_queue_link_head(ctx, link, &link->submit, shadow_req, true); |
6c271ce2 JA |
2559 | if (statep) |
2560 | io_submit_state_end(&state); | |
2561 | ||
2562 | return submitted; | |
2563 | } | |
2564 | ||
2565 | static int io_sq_thread(void *data) | |
2566 | { | |
2567 | struct sqe_submit sqes[IO_IOPOLL_BATCH]; | |
2568 | struct io_ring_ctx *ctx = data; | |
2569 | struct mm_struct *cur_mm = NULL; | |
2570 | mm_segment_t old_fs; | |
2571 | DEFINE_WAIT(wait); | |
2572 | unsigned inflight; | |
2573 | unsigned long timeout; | |
2574 | ||
a4c0b3de JL |
2575 | complete(&ctx->sqo_thread_started); |
2576 | ||
6c271ce2 JA |
2577 | old_fs = get_fs(); |
2578 | set_fs(USER_DS); | |
2579 | ||
2580 | timeout = inflight = 0; | |
2bbcd6d3 | 2581 | while (!kthread_should_park()) { |
6c271ce2 JA |
2582 | bool all_fixed, mm_fault = false; |
2583 | int i; | |
2584 | ||
2585 | if (inflight) { | |
2586 | unsigned nr_events = 0; | |
2587 | ||
2588 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
6c271ce2 | 2589 | io_iopoll_check(ctx, &nr_events, 0); |
6c271ce2 JA |
2590 | } else { |
2591 | /* | |
2592 | * Normal IO, just pretend everything completed. | |
2593 | * We don't have to poll completions for that. | |
2594 | */ | |
2595 | nr_events = inflight; | |
2596 | } | |
2597 | ||
2598 | inflight -= nr_events; | |
2599 | if (!inflight) | |
2600 | timeout = jiffies + ctx->sq_thread_idle; | |
2601 | } | |
2602 | ||
2603 | if (!io_get_sqring(ctx, &sqes[0])) { | |
2604 | /* | |
2605 | * We're polling. If we're within the defined idle | |
2606 | * period, then let us spin without work before going | |
2607 | * to sleep. | |
2608 | */ | |
2609 | if (inflight || !time_after(jiffies, timeout)) { | |
9831a90c | 2610 | cond_resched(); |
6c271ce2 JA |
2611 | continue; |
2612 | } | |
2613 | ||
2614 | /* | |
2615 | * Drop cur_mm before scheduling, we can't hold it for | |
2616 | * long periods (or over schedule()). Do this before | |
2617 | * adding ourselves to the waitqueue, as the unuse/drop | |
2618 | * may sleep. | |
2619 | */ | |
2620 | if (cur_mm) { | |
2621 | unuse_mm(cur_mm); | |
2622 | mmput(cur_mm); | |
2623 | cur_mm = NULL; | |
2624 | } | |
2625 | ||
2626 | prepare_to_wait(&ctx->sqo_wait, &wait, | |
2627 | TASK_INTERRUPTIBLE); | |
2628 | ||
2629 | /* Tell userspace we may need a wakeup call */ | |
75b28aff | 2630 | ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP; |
0d7bae69 SB |
2631 | /* make sure to read SQ tail after writing flags */ |
2632 | smp_mb(); | |
6c271ce2 JA |
2633 | |
2634 | if (!io_get_sqring(ctx, &sqes[0])) { | |
2bbcd6d3 | 2635 | if (kthread_should_park()) { |
6c271ce2 JA |
2636 | finish_wait(&ctx->sqo_wait, &wait); |
2637 | break; | |
2638 | } | |
2639 | if (signal_pending(current)) | |
2640 | flush_signals(current); | |
2641 | schedule(); | |
2642 | finish_wait(&ctx->sqo_wait, &wait); | |
2643 | ||
75b28aff | 2644 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
2645 | continue; |
2646 | } | |
2647 | finish_wait(&ctx->sqo_wait, &wait); | |
2648 | ||
75b28aff | 2649 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
2650 | } |
2651 | ||
2652 | i = 0; | |
2653 | all_fixed = true; | |
2654 | do { | |
2655 | if (all_fixed && io_sqe_needs_user(sqes[i].sqe)) | |
2656 | all_fixed = false; | |
2657 | ||
2658 | i++; | |
2659 | if (i == ARRAY_SIZE(sqes)) | |
2660 | break; | |
2661 | } while (io_get_sqring(ctx, &sqes[i])); | |
2662 | ||
2663 | /* Unless all new commands are FIXED regions, grab mm */ | |
2664 | if (!all_fixed && !cur_mm) { | |
2665 | mm_fault = !mmget_not_zero(ctx->sqo_mm); | |
2666 | if (!mm_fault) { | |
2667 | use_mm(ctx->sqo_mm); | |
2668 | cur_mm = ctx->sqo_mm; | |
2669 | } | |
2670 | } | |
2671 | ||
2672 | inflight += io_submit_sqes(ctx, sqes, i, cur_mm != NULL, | |
2673 | mm_fault); | |
2674 | ||
2675 | /* Commit SQ ring head once we've consumed all SQEs */ | |
2676 | io_commit_sqring(ctx); | |
2677 | } | |
2678 | ||
2679 | set_fs(old_fs); | |
2680 | if (cur_mm) { | |
2681 | unuse_mm(cur_mm); | |
2682 | mmput(cur_mm); | |
2683 | } | |
06058632 | 2684 | |
2bbcd6d3 | 2685 | kthread_parkme(); |
06058632 | 2686 | |
6c271ce2 JA |
2687 | return 0; |
2688 | } | |
2689 | ||
c5766668 JA |
2690 | static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit, |
2691 | bool block_for_last) | |
2b188cc1 | 2692 | { |
9a56a232 | 2693 | struct io_submit_state state, *statep = NULL; |
9e645e11 | 2694 | struct io_kiocb *link = NULL; |
4fe2c963 | 2695 | struct io_kiocb *shadow_req = NULL; |
9e645e11 | 2696 | bool prev_was_link = false; |
5c8b0b54 | 2697 | int i, submit = 0; |
2b188cc1 | 2698 | |
9a56a232 JA |
2699 | if (to_submit > IO_PLUG_THRESHOLD) { |
2700 | io_submit_state_start(&state, ctx, to_submit); | |
2701 | statep = &state; | |
2702 | } | |
2b188cc1 JA |
2703 | |
2704 | for (i = 0; i < to_submit; i++) { | |
c5766668 | 2705 | bool force_nonblock = true; |
2b188cc1 JA |
2706 | struct sqe_submit s; |
2707 | ||
2708 | if (!io_get_sqring(ctx, &s)) | |
2709 | break; | |
2710 | ||
9e645e11 JA |
2711 | /* |
2712 | * If previous wasn't linked and we have a linked command, | |
2713 | * that's the end of the chain. Submit the previous link. | |
2714 | */ | |
2715 | if (!prev_was_link && link) { | |
c5766668 JA |
2716 | io_queue_link_head(ctx, link, &link->submit, shadow_req, |
2717 | force_nonblock); | |
9e645e11 | 2718 | link = NULL; |
5f5ad9ce | 2719 | shadow_req = NULL; |
9e645e11 JA |
2720 | } |
2721 | prev_was_link = (s.sqe->flags & IOSQE_IO_LINK) != 0; | |
2722 | ||
4fe2c963 JL |
2723 | if (link && (s.sqe->flags & IOSQE_IO_DRAIN)) { |
2724 | if (!shadow_req) { | |
2725 | shadow_req = io_get_req(ctx, NULL); | |
a1041c27 JL |
2726 | if (unlikely(!shadow_req)) |
2727 | goto out; | |
4fe2c963 JL |
2728 | shadow_req->flags |= (REQ_F_IO_DRAIN | REQ_F_SHADOW_DRAIN); |
2729 | refcount_dec(&shadow_req->refs); | |
2730 | } | |
2731 | shadow_req->sequence = s.sequence; | |
2732 | } | |
2733 | ||
a1041c27 | 2734 | out: |
2b188cc1 | 2735 | s.has_user = true; |
def596e9 | 2736 | s.needs_lock = false; |
6c271ce2 | 2737 | s.needs_fixed_file = false; |
5c8b0b54 | 2738 | submit++; |
c5766668 JA |
2739 | |
2740 | /* | |
2741 | * The caller will block for events after submit, submit the | |
2742 | * last IO non-blocking. This is either the only IO it's | |
2743 | * submitting, or it already submitted the previous ones. This | |
2744 | * improves performance by avoiding an async punt that we don't | |
2745 | * need to do. | |
2746 | */ | |
2747 | if (block_for_last && submit == to_submit) | |
2748 | force_nonblock = false; | |
2749 | ||
2750 | io_submit_sqe(ctx, &s, statep, &link, force_nonblock); | |
2b188cc1 JA |
2751 | } |
2752 | io_commit_sqring(ctx); | |
2753 | ||
9e645e11 | 2754 | if (link) |
c5766668 | 2755 | io_queue_link_head(ctx, link, &link->submit, shadow_req, |
bf7ec93c | 2756 | !block_for_last); |
9a56a232 JA |
2757 | if (statep) |
2758 | io_submit_state_end(statep); | |
2b188cc1 | 2759 | |
5c8b0b54 | 2760 | return submit; |
2b188cc1 JA |
2761 | } |
2762 | ||
2b188cc1 JA |
2763 | /* |
2764 | * Wait until events become available, if we don't already have some. The | |
2765 | * application must reap them itself, as they reside on the shared cq ring. | |
2766 | */ | |
2767 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
2768 | const sigset_t __user *sig, size_t sigsz) | |
2769 | { | |
75b28aff | 2770 | struct io_rings *rings = ctx->rings; |
5262f567 | 2771 | unsigned nr_timeouts; |
2b188cc1 JA |
2772 | int ret; |
2773 | ||
75b28aff | 2774 | if (io_cqring_events(rings) >= min_events) |
2b188cc1 JA |
2775 | return 0; |
2776 | ||
2777 | if (sig) { | |
9e75ad5d AB |
2778 | #ifdef CONFIG_COMPAT |
2779 | if (in_compat_syscall()) | |
2780 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 2781 | sigsz); |
9e75ad5d AB |
2782 | else |
2783 | #endif | |
b772434b | 2784 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 2785 | |
2b188cc1 JA |
2786 | if (ret) |
2787 | return ret; | |
2788 | } | |
2789 | ||
5262f567 JA |
2790 | nr_timeouts = atomic_read(&ctx->cq_timeouts); |
2791 | /* | |
2792 | * Return if we have enough events, or if a timeout occured since | |
2793 | * we started waiting. For timeouts, we always want to return to | |
2794 | * userspace. | |
2795 | */ | |
2796 | ret = wait_event_interruptible(ctx->wait, | |
2797 | io_cqring_events(rings) >= min_events || | |
2798 | atomic_read(&ctx->cq_timeouts) != nr_timeouts); | |
b772434b | 2799 | restore_saved_sigmask_unless(ret == -ERESTARTSYS); |
97abc889 ON |
2800 | if (ret == -ERESTARTSYS) |
2801 | ret = -EINTR; | |
2b188cc1 | 2802 | |
75b28aff | 2803 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
2804 | } |
2805 | ||
6b06314c JA |
2806 | static void __io_sqe_files_unregister(struct io_ring_ctx *ctx) |
2807 | { | |
2808 | #if defined(CONFIG_UNIX) | |
2809 | if (ctx->ring_sock) { | |
2810 | struct sock *sock = ctx->ring_sock->sk; | |
2811 | struct sk_buff *skb; | |
2812 | ||
2813 | while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) | |
2814 | kfree_skb(skb); | |
2815 | } | |
2816 | #else | |
2817 | int i; | |
2818 | ||
2819 | for (i = 0; i < ctx->nr_user_files; i++) | |
2820 | fput(ctx->user_files[i]); | |
2821 | #endif | |
2822 | } | |
2823 | ||
2824 | static int io_sqe_files_unregister(struct io_ring_ctx *ctx) | |
2825 | { | |
2826 | if (!ctx->user_files) | |
2827 | return -ENXIO; | |
2828 | ||
2829 | __io_sqe_files_unregister(ctx); | |
2830 | kfree(ctx->user_files); | |
2831 | ctx->user_files = NULL; | |
2832 | ctx->nr_user_files = 0; | |
2833 | return 0; | |
2834 | } | |
2835 | ||
6c271ce2 JA |
2836 | static void io_sq_thread_stop(struct io_ring_ctx *ctx) |
2837 | { | |
2838 | if (ctx->sqo_thread) { | |
a4c0b3de | 2839 | wait_for_completion(&ctx->sqo_thread_started); |
2bbcd6d3 RP |
2840 | /* |
2841 | * The park is a bit of a work-around, without it we get | |
2842 | * warning spews on shutdown with SQPOLL set and affinity | |
2843 | * set to a single CPU. | |
2844 | */ | |
06058632 | 2845 | kthread_park(ctx->sqo_thread); |
6c271ce2 JA |
2846 | kthread_stop(ctx->sqo_thread); |
2847 | ctx->sqo_thread = NULL; | |
2848 | } | |
2849 | } | |
2850 | ||
6b06314c JA |
2851 | static void io_finish_async(struct io_ring_ctx *ctx) |
2852 | { | |
54a91f3b JA |
2853 | int i; |
2854 | ||
6c271ce2 JA |
2855 | io_sq_thread_stop(ctx); |
2856 | ||
54a91f3b JA |
2857 | for (i = 0; i < ARRAY_SIZE(ctx->sqo_wq); i++) { |
2858 | if (ctx->sqo_wq[i]) { | |
2859 | destroy_workqueue(ctx->sqo_wq[i]); | |
2860 | ctx->sqo_wq[i] = NULL; | |
2861 | } | |
6b06314c JA |
2862 | } |
2863 | } | |
2864 | ||
2865 | #if defined(CONFIG_UNIX) | |
2866 | static void io_destruct_skb(struct sk_buff *skb) | |
2867 | { | |
2868 | struct io_ring_ctx *ctx = skb->sk->sk_user_data; | |
8a997340 JA |
2869 | int i; |
2870 | ||
2871 | for (i = 0; i < ARRAY_SIZE(ctx->sqo_wq); i++) | |
2872 | if (ctx->sqo_wq[i]) | |
2873 | flush_workqueue(ctx->sqo_wq[i]); | |
6b06314c | 2874 | |
6b06314c JA |
2875 | unix_destruct_scm(skb); |
2876 | } | |
2877 | ||
2878 | /* | |
2879 | * Ensure the UNIX gc is aware of our file set, so we are certain that | |
2880 | * the io_uring can be safely unregistered on process exit, even if we have | |
2881 | * loops in the file referencing. | |
2882 | */ | |
2883 | static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset) | |
2884 | { | |
2885 | struct sock *sk = ctx->ring_sock->sk; | |
2886 | struct scm_fp_list *fpl; | |
2887 | struct sk_buff *skb; | |
2888 | int i; | |
2889 | ||
2890 | if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) { | |
2891 | unsigned long inflight = ctx->user->unix_inflight + nr; | |
2892 | ||
2893 | if (inflight > task_rlimit(current, RLIMIT_NOFILE)) | |
2894 | return -EMFILE; | |
2895 | } | |
2896 | ||
2897 | fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); | |
2898 | if (!fpl) | |
2899 | return -ENOMEM; | |
2900 | ||
2901 | skb = alloc_skb(0, GFP_KERNEL); | |
2902 | if (!skb) { | |
2903 | kfree(fpl); | |
2904 | return -ENOMEM; | |
2905 | } | |
2906 | ||
2907 | skb->sk = sk; | |
2908 | skb->destructor = io_destruct_skb; | |
2909 | ||
2910 | fpl->user = get_uid(ctx->user); | |
2911 | for (i = 0; i < nr; i++) { | |
2912 | fpl->fp[i] = get_file(ctx->user_files[i + offset]); | |
2913 | unix_inflight(fpl->user, fpl->fp[i]); | |
2914 | } | |
2915 | ||
2916 | fpl->max = fpl->count = nr; | |
2917 | UNIXCB(skb).fp = fpl; | |
2918 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); | |
2919 | skb_queue_head(&sk->sk_receive_queue, skb); | |
2920 | ||
2921 | for (i = 0; i < nr; i++) | |
2922 | fput(fpl->fp[i]); | |
2923 | ||
2924 | return 0; | |
2925 | } | |
2926 | ||
2927 | /* | |
2928 | * If UNIX sockets are enabled, fd passing can cause a reference cycle which | |
2929 | * causes regular reference counting to break down. We rely on the UNIX | |
2930 | * garbage collection to take care of this problem for us. | |
2931 | */ | |
2932 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
2933 | { | |
2934 | unsigned left, total; | |
2935 | int ret = 0; | |
2936 | ||
2937 | total = 0; | |
2938 | left = ctx->nr_user_files; | |
2939 | while (left) { | |
2940 | unsigned this_files = min_t(unsigned, left, SCM_MAX_FD); | |
6b06314c JA |
2941 | |
2942 | ret = __io_sqe_files_scm(ctx, this_files, total); | |
2943 | if (ret) | |
2944 | break; | |
2945 | left -= this_files; | |
2946 | total += this_files; | |
2947 | } | |
2948 | ||
2949 | if (!ret) | |
2950 | return 0; | |
2951 | ||
2952 | while (total < ctx->nr_user_files) { | |
2953 | fput(ctx->user_files[total]); | |
2954 | total++; | |
2955 | } | |
2956 | ||
2957 | return ret; | |
2958 | } | |
2959 | #else | |
2960 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
2961 | { | |
2962 | return 0; | |
2963 | } | |
2964 | #endif | |
2965 | ||
2966 | static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, | |
2967 | unsigned nr_args) | |
2968 | { | |
2969 | __s32 __user *fds = (__s32 __user *) arg; | |
2970 | int fd, ret = 0; | |
2971 | unsigned i; | |
2972 | ||
2973 | if (ctx->user_files) | |
2974 | return -EBUSY; | |
2975 | if (!nr_args) | |
2976 | return -EINVAL; | |
2977 | if (nr_args > IORING_MAX_FIXED_FILES) | |
2978 | return -EMFILE; | |
2979 | ||
2980 | ctx->user_files = kcalloc(nr_args, sizeof(struct file *), GFP_KERNEL); | |
2981 | if (!ctx->user_files) | |
2982 | return -ENOMEM; | |
2983 | ||
2984 | for (i = 0; i < nr_args; i++) { | |
2985 | ret = -EFAULT; | |
2986 | if (copy_from_user(&fd, &fds[i], sizeof(fd))) | |
2987 | break; | |
2988 | ||
2989 | ctx->user_files[i] = fget(fd); | |
2990 | ||
2991 | ret = -EBADF; | |
2992 | if (!ctx->user_files[i]) | |
2993 | break; | |
2994 | /* | |
2995 | * Don't allow io_uring instances to be registered. If UNIX | |
2996 | * isn't enabled, then this causes a reference cycle and this | |
2997 | * instance can never get freed. If UNIX is enabled we'll | |
2998 | * handle it just fine, but there's still no point in allowing | |
2999 | * a ring fd as it doesn't support regular read/write anyway. | |
3000 | */ | |
3001 | if (ctx->user_files[i]->f_op == &io_uring_fops) { | |
3002 | fput(ctx->user_files[i]); | |
3003 | break; | |
3004 | } | |
3005 | ctx->nr_user_files++; | |
3006 | ret = 0; | |
3007 | } | |
3008 | ||
3009 | if (ret) { | |
3010 | for (i = 0; i < ctx->nr_user_files; i++) | |
3011 | fput(ctx->user_files[i]); | |
3012 | ||
3013 | kfree(ctx->user_files); | |
25adf50f | 3014 | ctx->user_files = NULL; |
6b06314c JA |
3015 | ctx->nr_user_files = 0; |
3016 | return ret; | |
3017 | } | |
3018 | ||
3019 | ret = io_sqe_files_scm(ctx); | |
3020 | if (ret) | |
3021 | io_sqe_files_unregister(ctx); | |
3022 | ||
3023 | return ret; | |
3024 | } | |
3025 | ||
6c271ce2 JA |
3026 | static int io_sq_offload_start(struct io_ring_ctx *ctx, |
3027 | struct io_uring_params *p) | |
2b188cc1 JA |
3028 | { |
3029 | int ret; | |
3030 | ||
6c271ce2 | 3031 | init_waitqueue_head(&ctx->sqo_wait); |
2b188cc1 JA |
3032 | mmgrab(current->mm); |
3033 | ctx->sqo_mm = current->mm; | |
3034 | ||
6c271ce2 | 3035 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
3ec482d1 JA |
3036 | ret = -EPERM; |
3037 | if (!capable(CAP_SYS_ADMIN)) | |
3038 | goto err; | |
3039 | ||
917257da JA |
3040 | ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle); |
3041 | if (!ctx->sq_thread_idle) | |
3042 | ctx->sq_thread_idle = HZ; | |
3043 | ||
6c271ce2 | 3044 | if (p->flags & IORING_SETUP_SQ_AFF) { |
44a9bd18 | 3045 | int cpu = p->sq_thread_cpu; |
6c271ce2 | 3046 | |
917257da | 3047 | ret = -EINVAL; |
44a9bd18 JA |
3048 | if (cpu >= nr_cpu_ids) |
3049 | goto err; | |
7889f44d | 3050 | if (!cpu_online(cpu)) |
917257da JA |
3051 | goto err; |
3052 | ||
6c271ce2 JA |
3053 | ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread, |
3054 | ctx, cpu, | |
3055 | "io_uring-sq"); | |
3056 | } else { | |
3057 | ctx->sqo_thread = kthread_create(io_sq_thread, ctx, | |
3058 | "io_uring-sq"); | |
3059 | } | |
3060 | if (IS_ERR(ctx->sqo_thread)) { | |
3061 | ret = PTR_ERR(ctx->sqo_thread); | |
3062 | ctx->sqo_thread = NULL; | |
3063 | goto err; | |
3064 | } | |
3065 | wake_up_process(ctx->sqo_thread); | |
3066 | } else if (p->flags & IORING_SETUP_SQ_AFF) { | |
3067 | /* Can't have SQ_AFF without SQPOLL */ | |
3068 | ret = -EINVAL; | |
3069 | goto err; | |
3070 | } | |
3071 | ||
2b188cc1 | 3072 | /* Do QD, or 2 * CPUS, whatever is smallest */ |
54a91f3b JA |
3073 | ctx->sqo_wq[0] = alloc_workqueue("io_ring-wq", |
3074 | WQ_UNBOUND | WQ_FREEZABLE, | |
2b188cc1 | 3075 | min(ctx->sq_entries - 1, 2 * num_online_cpus())); |
54a91f3b JA |
3076 | if (!ctx->sqo_wq[0]) { |
3077 | ret = -ENOMEM; | |
3078 | goto err; | |
3079 | } | |
3080 | ||
3081 | /* | |
3082 | * This is for buffered writes, where we want to limit the parallelism | |
3083 | * due to file locking in file systems. As "normal" buffered writes | |
3084 | * should parellelize on writeout quite nicely, limit us to having 2 | |
3085 | * pending. This avoids massive contention on the inode when doing | |
3086 | * buffered async writes. | |
3087 | */ | |
3088 | ctx->sqo_wq[1] = alloc_workqueue("io_ring-write-wq", | |
3089 | WQ_UNBOUND | WQ_FREEZABLE, 2); | |
3090 | if (!ctx->sqo_wq[1]) { | |
2b188cc1 JA |
3091 | ret = -ENOMEM; |
3092 | goto err; | |
3093 | } | |
3094 | ||
3095 | return 0; | |
3096 | err: | |
54a91f3b | 3097 | io_finish_async(ctx); |
2b188cc1 JA |
3098 | mmdrop(ctx->sqo_mm); |
3099 | ctx->sqo_mm = NULL; | |
3100 | return ret; | |
3101 | } | |
3102 | ||
3103 | static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages) | |
3104 | { | |
3105 | atomic_long_sub(nr_pages, &user->locked_vm); | |
3106 | } | |
3107 | ||
3108 | static int io_account_mem(struct user_struct *user, unsigned long nr_pages) | |
3109 | { | |
3110 | unsigned long page_limit, cur_pages, new_pages; | |
3111 | ||
3112 | /* Don't allow more pages than we can safely lock */ | |
3113 | page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
3114 | ||
3115 | do { | |
3116 | cur_pages = atomic_long_read(&user->locked_vm); | |
3117 | new_pages = cur_pages + nr_pages; | |
3118 | if (new_pages > page_limit) | |
3119 | return -ENOMEM; | |
3120 | } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages, | |
3121 | new_pages) != cur_pages); | |
3122 | ||
3123 | return 0; | |
3124 | } | |
3125 | ||
3126 | static void io_mem_free(void *ptr) | |
3127 | { | |
52e04ef4 MR |
3128 | struct page *page; |
3129 | ||
3130 | if (!ptr) | |
3131 | return; | |
2b188cc1 | 3132 | |
52e04ef4 | 3133 | page = virt_to_head_page(ptr); |
2b188cc1 JA |
3134 | if (put_page_testzero(page)) |
3135 | free_compound_page(page); | |
3136 | } | |
3137 | ||
3138 | static void *io_mem_alloc(size_t size) | |
3139 | { | |
3140 | gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP | | |
3141 | __GFP_NORETRY; | |
3142 | ||
3143 | return (void *) __get_free_pages(gfp_flags, get_order(size)); | |
3144 | } | |
3145 | ||
75b28aff HV |
3146 | static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries, |
3147 | size_t *sq_offset) | |
3148 | { | |
3149 | struct io_rings *rings; | |
3150 | size_t off, sq_array_size; | |
3151 | ||
3152 | off = struct_size(rings, cqes, cq_entries); | |
3153 | if (off == SIZE_MAX) | |
3154 | return SIZE_MAX; | |
3155 | ||
3156 | #ifdef CONFIG_SMP | |
3157 | off = ALIGN(off, SMP_CACHE_BYTES); | |
3158 | if (off == 0) | |
3159 | return SIZE_MAX; | |
3160 | #endif | |
3161 | ||
3162 | sq_array_size = array_size(sizeof(u32), sq_entries); | |
3163 | if (sq_array_size == SIZE_MAX) | |
3164 | return SIZE_MAX; | |
3165 | ||
3166 | if (check_add_overflow(off, sq_array_size, &off)) | |
3167 | return SIZE_MAX; | |
3168 | ||
3169 | if (sq_offset) | |
3170 | *sq_offset = off; | |
3171 | ||
3172 | return off; | |
3173 | } | |
3174 | ||
2b188cc1 JA |
3175 | static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries) |
3176 | { | |
75b28aff | 3177 | size_t pages; |
2b188cc1 | 3178 | |
75b28aff HV |
3179 | pages = (size_t)1 << get_order( |
3180 | rings_size(sq_entries, cq_entries, NULL)); | |
3181 | pages += (size_t)1 << get_order( | |
3182 | array_size(sizeof(struct io_uring_sqe), sq_entries)); | |
2b188cc1 | 3183 | |
75b28aff | 3184 | return pages; |
2b188cc1 JA |
3185 | } |
3186 | ||
edafccee JA |
3187 | static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx) |
3188 | { | |
3189 | int i, j; | |
3190 | ||
3191 | if (!ctx->user_bufs) | |
3192 | return -ENXIO; | |
3193 | ||
3194 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
3195 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
3196 | ||
3197 | for (j = 0; j < imu->nr_bvecs; j++) | |
27c4d3a3 | 3198 | put_user_page(imu->bvec[j].bv_page); |
edafccee JA |
3199 | |
3200 | if (ctx->account_mem) | |
3201 | io_unaccount_mem(ctx->user, imu->nr_bvecs); | |
d4ef6475 | 3202 | kvfree(imu->bvec); |
edafccee JA |
3203 | imu->nr_bvecs = 0; |
3204 | } | |
3205 | ||
3206 | kfree(ctx->user_bufs); | |
3207 | ctx->user_bufs = NULL; | |
3208 | ctx->nr_user_bufs = 0; | |
3209 | return 0; | |
3210 | } | |
3211 | ||
3212 | static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, | |
3213 | void __user *arg, unsigned index) | |
3214 | { | |
3215 | struct iovec __user *src; | |
3216 | ||
3217 | #ifdef CONFIG_COMPAT | |
3218 | if (ctx->compat) { | |
3219 | struct compat_iovec __user *ciovs; | |
3220 | struct compat_iovec ciov; | |
3221 | ||
3222 | ciovs = (struct compat_iovec __user *) arg; | |
3223 | if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) | |
3224 | return -EFAULT; | |
3225 | ||
3226 | dst->iov_base = (void __user *) (unsigned long) ciov.iov_base; | |
3227 | dst->iov_len = ciov.iov_len; | |
3228 | return 0; | |
3229 | } | |
3230 | #endif | |
3231 | src = (struct iovec __user *) arg; | |
3232 | if (copy_from_user(dst, &src[index], sizeof(*dst))) | |
3233 | return -EFAULT; | |
3234 | return 0; | |
3235 | } | |
3236 | ||
3237 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg, | |
3238 | unsigned nr_args) | |
3239 | { | |
3240 | struct vm_area_struct **vmas = NULL; | |
3241 | struct page **pages = NULL; | |
3242 | int i, j, got_pages = 0; | |
3243 | int ret = -EINVAL; | |
3244 | ||
3245 | if (ctx->user_bufs) | |
3246 | return -EBUSY; | |
3247 | if (!nr_args || nr_args > UIO_MAXIOV) | |
3248 | return -EINVAL; | |
3249 | ||
3250 | ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf), | |
3251 | GFP_KERNEL); | |
3252 | if (!ctx->user_bufs) | |
3253 | return -ENOMEM; | |
3254 | ||
3255 | for (i = 0; i < nr_args; i++) { | |
3256 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
3257 | unsigned long off, start, end, ubuf; | |
3258 | int pret, nr_pages; | |
3259 | struct iovec iov; | |
3260 | size_t size; | |
3261 | ||
3262 | ret = io_copy_iov(ctx, &iov, arg, i); | |
3263 | if (ret) | |
a278682d | 3264 | goto err; |
edafccee JA |
3265 | |
3266 | /* | |
3267 | * Don't impose further limits on the size and buffer | |
3268 | * constraints here, we'll -EINVAL later when IO is | |
3269 | * submitted if they are wrong. | |
3270 | */ | |
3271 | ret = -EFAULT; | |
3272 | if (!iov.iov_base || !iov.iov_len) | |
3273 | goto err; | |
3274 | ||
3275 | /* arbitrary limit, but we need something */ | |
3276 | if (iov.iov_len > SZ_1G) | |
3277 | goto err; | |
3278 | ||
3279 | ubuf = (unsigned long) iov.iov_base; | |
3280 | end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
3281 | start = ubuf >> PAGE_SHIFT; | |
3282 | nr_pages = end - start; | |
3283 | ||
3284 | if (ctx->account_mem) { | |
3285 | ret = io_account_mem(ctx->user, nr_pages); | |
3286 | if (ret) | |
3287 | goto err; | |
3288 | } | |
3289 | ||
3290 | ret = 0; | |
3291 | if (!pages || nr_pages > got_pages) { | |
3292 | kfree(vmas); | |
3293 | kfree(pages); | |
d4ef6475 | 3294 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), |
edafccee | 3295 | GFP_KERNEL); |
d4ef6475 | 3296 | vmas = kvmalloc_array(nr_pages, |
edafccee JA |
3297 | sizeof(struct vm_area_struct *), |
3298 | GFP_KERNEL); | |
3299 | if (!pages || !vmas) { | |
3300 | ret = -ENOMEM; | |
3301 | if (ctx->account_mem) | |
3302 | io_unaccount_mem(ctx->user, nr_pages); | |
3303 | goto err; | |
3304 | } | |
3305 | got_pages = nr_pages; | |
3306 | } | |
3307 | ||
d4ef6475 | 3308 | imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec), |
edafccee JA |
3309 | GFP_KERNEL); |
3310 | ret = -ENOMEM; | |
3311 | if (!imu->bvec) { | |
3312 | if (ctx->account_mem) | |
3313 | io_unaccount_mem(ctx->user, nr_pages); | |
3314 | goto err; | |
3315 | } | |
3316 | ||
3317 | ret = 0; | |
3318 | down_read(¤t->mm->mmap_sem); | |
932f4a63 IW |
3319 | pret = get_user_pages(ubuf, nr_pages, |
3320 | FOLL_WRITE | FOLL_LONGTERM, | |
3321 | pages, vmas); | |
edafccee JA |
3322 | if (pret == nr_pages) { |
3323 | /* don't support file backed memory */ | |
3324 | for (j = 0; j < nr_pages; j++) { | |
3325 | struct vm_area_struct *vma = vmas[j]; | |
3326 | ||
3327 | if (vma->vm_file && | |
3328 | !is_file_hugepages(vma->vm_file)) { | |
3329 | ret = -EOPNOTSUPP; | |
3330 | break; | |
3331 | } | |
3332 | } | |
3333 | } else { | |
3334 | ret = pret < 0 ? pret : -EFAULT; | |
3335 | } | |
3336 | up_read(¤t->mm->mmap_sem); | |
3337 | if (ret) { | |
3338 | /* | |
3339 | * if we did partial map, or found file backed vmas, | |
3340 | * release any pages we did get | |
3341 | */ | |
27c4d3a3 JH |
3342 | if (pret > 0) |
3343 | put_user_pages(pages, pret); | |
edafccee JA |
3344 | if (ctx->account_mem) |
3345 | io_unaccount_mem(ctx->user, nr_pages); | |
d4ef6475 | 3346 | kvfree(imu->bvec); |
edafccee JA |
3347 | goto err; |
3348 | } | |
3349 | ||
3350 | off = ubuf & ~PAGE_MASK; | |
3351 | size = iov.iov_len; | |
3352 | for (j = 0; j < nr_pages; j++) { | |
3353 | size_t vec_len; | |
3354 | ||
3355 | vec_len = min_t(size_t, size, PAGE_SIZE - off); | |
3356 | imu->bvec[j].bv_page = pages[j]; | |
3357 | imu->bvec[j].bv_len = vec_len; | |
3358 | imu->bvec[j].bv_offset = off; | |
3359 | off = 0; | |
3360 | size -= vec_len; | |
3361 | } | |
3362 | /* store original address for later verification */ | |
3363 | imu->ubuf = ubuf; | |
3364 | imu->len = iov.iov_len; | |
3365 | imu->nr_bvecs = nr_pages; | |
3366 | ||
3367 | ctx->nr_user_bufs++; | |
3368 | } | |
d4ef6475 MR |
3369 | kvfree(pages); |
3370 | kvfree(vmas); | |
edafccee JA |
3371 | return 0; |
3372 | err: | |
d4ef6475 MR |
3373 | kvfree(pages); |
3374 | kvfree(vmas); | |
edafccee JA |
3375 | io_sqe_buffer_unregister(ctx); |
3376 | return ret; | |
3377 | } | |
3378 | ||
9b402849 JA |
3379 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg) |
3380 | { | |
3381 | __s32 __user *fds = arg; | |
3382 | int fd; | |
3383 | ||
3384 | if (ctx->cq_ev_fd) | |
3385 | return -EBUSY; | |
3386 | ||
3387 | if (copy_from_user(&fd, fds, sizeof(*fds))) | |
3388 | return -EFAULT; | |
3389 | ||
3390 | ctx->cq_ev_fd = eventfd_ctx_fdget(fd); | |
3391 | if (IS_ERR(ctx->cq_ev_fd)) { | |
3392 | int ret = PTR_ERR(ctx->cq_ev_fd); | |
3393 | ctx->cq_ev_fd = NULL; | |
3394 | return ret; | |
3395 | } | |
3396 | ||
3397 | return 0; | |
3398 | } | |
3399 | ||
3400 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) | |
3401 | { | |
3402 | if (ctx->cq_ev_fd) { | |
3403 | eventfd_ctx_put(ctx->cq_ev_fd); | |
3404 | ctx->cq_ev_fd = NULL; | |
3405 | return 0; | |
3406 | } | |
3407 | ||
3408 | return -ENXIO; | |
3409 | } | |
3410 | ||
2b188cc1 JA |
3411 | static void io_ring_ctx_free(struct io_ring_ctx *ctx) |
3412 | { | |
6b06314c | 3413 | io_finish_async(ctx); |
2b188cc1 JA |
3414 | if (ctx->sqo_mm) |
3415 | mmdrop(ctx->sqo_mm); | |
def596e9 JA |
3416 | |
3417 | io_iopoll_reap_events(ctx); | |
edafccee | 3418 | io_sqe_buffer_unregister(ctx); |
6b06314c | 3419 | io_sqe_files_unregister(ctx); |
9b402849 | 3420 | io_eventfd_unregister(ctx); |
def596e9 | 3421 | |
2b188cc1 | 3422 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
3423 | if (ctx->ring_sock) { |
3424 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 3425 | sock_release(ctx->ring_sock); |
355e8d26 | 3426 | } |
2b188cc1 JA |
3427 | #endif |
3428 | ||
75b28aff | 3429 | io_mem_free(ctx->rings); |
2b188cc1 | 3430 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
3431 | |
3432 | percpu_ref_exit(&ctx->refs); | |
3433 | if (ctx->account_mem) | |
3434 | io_unaccount_mem(ctx->user, | |
3435 | ring_pages(ctx->sq_entries, ctx->cq_entries)); | |
3436 | free_uid(ctx->user); | |
3437 | kfree(ctx); | |
3438 | } | |
3439 | ||
3440 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
3441 | { | |
3442 | struct io_ring_ctx *ctx = file->private_data; | |
3443 | __poll_t mask = 0; | |
3444 | ||
3445 | poll_wait(file, &ctx->cq_wait, wait); | |
4f7067c3 SB |
3446 | /* |
3447 | * synchronizes with barrier from wq_has_sleeper call in | |
3448 | * io_commit_cqring | |
3449 | */ | |
2b188cc1 | 3450 | smp_rmb(); |
75b28aff HV |
3451 | if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head != |
3452 | ctx->rings->sq_ring_entries) | |
2b188cc1 | 3453 | mask |= EPOLLOUT | EPOLLWRNORM; |
75b28aff | 3454 | if (READ_ONCE(ctx->rings->sq.head) != ctx->cached_cq_tail) |
2b188cc1 JA |
3455 | mask |= EPOLLIN | EPOLLRDNORM; |
3456 | ||
3457 | return mask; | |
3458 | } | |
3459 | ||
3460 | static int io_uring_fasync(int fd, struct file *file, int on) | |
3461 | { | |
3462 | struct io_ring_ctx *ctx = file->private_data; | |
3463 | ||
3464 | return fasync_helper(fd, file, on, &ctx->cq_fasync); | |
3465 | } | |
3466 | ||
3467 | static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) | |
3468 | { | |
3469 | mutex_lock(&ctx->uring_lock); | |
3470 | percpu_ref_kill(&ctx->refs); | |
3471 | mutex_unlock(&ctx->uring_lock); | |
3472 | ||
5262f567 | 3473 | io_kill_timeouts(ctx); |
221c5eb2 | 3474 | io_poll_remove_all(ctx); |
def596e9 | 3475 | io_iopoll_reap_events(ctx); |
2b188cc1 JA |
3476 | wait_for_completion(&ctx->ctx_done); |
3477 | io_ring_ctx_free(ctx); | |
3478 | } | |
3479 | ||
3480 | static int io_uring_release(struct inode *inode, struct file *file) | |
3481 | { | |
3482 | struct io_ring_ctx *ctx = file->private_data; | |
3483 | ||
3484 | file->private_data = NULL; | |
3485 | io_ring_ctx_wait_and_kill(ctx); | |
3486 | return 0; | |
3487 | } | |
3488 | ||
3489 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
3490 | { | |
3491 | loff_t offset = (loff_t) vma->vm_pgoff << PAGE_SHIFT; | |
3492 | unsigned long sz = vma->vm_end - vma->vm_start; | |
3493 | struct io_ring_ctx *ctx = file->private_data; | |
3494 | unsigned long pfn; | |
3495 | struct page *page; | |
3496 | void *ptr; | |
3497 | ||
3498 | switch (offset) { | |
3499 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
3500 | case IORING_OFF_CQ_RING: |
3501 | ptr = ctx->rings; | |
2b188cc1 JA |
3502 | break; |
3503 | case IORING_OFF_SQES: | |
3504 | ptr = ctx->sq_sqes; | |
3505 | break; | |
2b188cc1 JA |
3506 | default: |
3507 | return -EINVAL; | |
3508 | } | |
3509 | ||
3510 | page = virt_to_head_page(ptr); | |
a50b854e | 3511 | if (sz > page_size(page)) |
2b188cc1 JA |
3512 | return -EINVAL; |
3513 | ||
3514 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
3515 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
3516 | } | |
3517 | ||
3518 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, | |
3519 | u32, min_complete, u32, flags, const sigset_t __user *, sig, | |
3520 | size_t, sigsz) | |
3521 | { | |
3522 | struct io_ring_ctx *ctx; | |
3523 | long ret = -EBADF; | |
3524 | int submitted = 0; | |
3525 | struct fd f; | |
3526 | ||
6c271ce2 | 3527 | if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP)) |
2b188cc1 JA |
3528 | return -EINVAL; |
3529 | ||
3530 | f = fdget(fd); | |
3531 | if (!f.file) | |
3532 | return -EBADF; | |
3533 | ||
3534 | ret = -EOPNOTSUPP; | |
3535 | if (f.file->f_op != &io_uring_fops) | |
3536 | goto out_fput; | |
3537 | ||
3538 | ret = -ENXIO; | |
3539 | ctx = f.file->private_data; | |
3540 | if (!percpu_ref_tryget(&ctx->refs)) | |
3541 | goto out_fput; | |
3542 | ||
6c271ce2 JA |
3543 | /* |
3544 | * For SQ polling, the thread will do all submissions and completions. | |
3545 | * Just return the requested submit count, and wake the thread if | |
3546 | * we were asked to. | |
3547 | */ | |
b2a9eada | 3548 | ret = 0; |
6c271ce2 JA |
3549 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
3550 | if (flags & IORING_ENTER_SQ_WAKEUP) | |
3551 | wake_up(&ctx->sqo_wait); | |
3552 | submitted = to_submit; | |
b2a9eada | 3553 | } else if (to_submit) { |
c5766668 JA |
3554 | bool block_for_last = false; |
3555 | ||
2b188cc1 JA |
3556 | to_submit = min(to_submit, ctx->sq_entries); |
3557 | ||
c5766668 JA |
3558 | /* |
3559 | * Allow last submission to block in a series, IFF the caller | |
3560 | * asked to wait for events and we don't currently have | |
3561 | * enough. This potentially avoids an async punt. | |
3562 | */ | |
3563 | if (to_submit == min_complete && | |
3564 | io_cqring_events(ctx->rings) < min_complete) | |
3565 | block_for_last = true; | |
3566 | ||
2b188cc1 | 3567 | mutex_lock(&ctx->uring_lock); |
c5766668 | 3568 | submitted = io_ring_submit(ctx, to_submit, block_for_last); |
2b188cc1 | 3569 | mutex_unlock(&ctx->uring_lock); |
2b188cc1 JA |
3570 | } |
3571 | if (flags & IORING_ENTER_GETEVENTS) { | |
def596e9 JA |
3572 | unsigned nr_events = 0; |
3573 | ||
2b188cc1 JA |
3574 | min_complete = min(min_complete, ctx->cq_entries); |
3575 | ||
def596e9 | 3576 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 | 3577 | ret = io_iopoll_check(ctx, &nr_events, min_complete); |
def596e9 JA |
3578 | } else { |
3579 | ret = io_cqring_wait(ctx, min_complete, sig, sigsz); | |
3580 | } | |
2b188cc1 JA |
3581 | } |
3582 | ||
6805b32e | 3583 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
3584 | out_fput: |
3585 | fdput(f); | |
3586 | return submitted ? submitted : ret; | |
3587 | } | |
3588 | ||
3589 | static const struct file_operations io_uring_fops = { | |
3590 | .release = io_uring_release, | |
3591 | .mmap = io_uring_mmap, | |
3592 | .poll = io_uring_poll, | |
3593 | .fasync = io_uring_fasync, | |
3594 | }; | |
3595 | ||
3596 | static int io_allocate_scq_urings(struct io_ring_ctx *ctx, | |
3597 | struct io_uring_params *p) | |
3598 | { | |
75b28aff HV |
3599 | struct io_rings *rings; |
3600 | size_t size, sq_array_offset; | |
2b188cc1 | 3601 | |
75b28aff HV |
3602 | size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset); |
3603 | if (size == SIZE_MAX) | |
3604 | return -EOVERFLOW; | |
3605 | ||
3606 | rings = io_mem_alloc(size); | |
3607 | if (!rings) | |
2b188cc1 JA |
3608 | return -ENOMEM; |
3609 | ||
75b28aff HV |
3610 | ctx->rings = rings; |
3611 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
3612 | rings->sq_ring_mask = p->sq_entries - 1; | |
3613 | rings->cq_ring_mask = p->cq_entries - 1; | |
3614 | rings->sq_ring_entries = p->sq_entries; | |
3615 | rings->cq_ring_entries = p->cq_entries; | |
3616 | ctx->sq_mask = rings->sq_ring_mask; | |
3617 | ctx->cq_mask = rings->cq_ring_mask; | |
3618 | ctx->sq_entries = rings->sq_ring_entries; | |
3619 | ctx->cq_entries = rings->cq_ring_entries; | |
2b188cc1 JA |
3620 | |
3621 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
3622 | if (size == SIZE_MAX) | |
3623 | return -EOVERFLOW; | |
3624 | ||
3625 | ctx->sq_sqes = io_mem_alloc(size); | |
52e04ef4 | 3626 | if (!ctx->sq_sqes) |
2b188cc1 | 3627 | return -ENOMEM; |
2b188cc1 | 3628 | |
2b188cc1 JA |
3629 | return 0; |
3630 | } | |
3631 | ||
3632 | /* | |
3633 | * Allocate an anonymous fd, this is what constitutes the application | |
3634 | * visible backing of an io_uring instance. The application mmaps this | |
3635 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
3636 | * we have to tie this fd to a socket for file garbage collection purposes. | |
3637 | */ | |
3638 | static int io_uring_get_fd(struct io_ring_ctx *ctx) | |
3639 | { | |
3640 | struct file *file; | |
3641 | int ret; | |
3642 | ||
3643 | #if defined(CONFIG_UNIX) | |
3644 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, | |
3645 | &ctx->ring_sock); | |
3646 | if (ret) | |
3647 | return ret; | |
3648 | #endif | |
3649 | ||
3650 | ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
3651 | if (ret < 0) | |
3652 | goto err; | |
3653 | ||
3654 | file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx, | |
3655 | O_RDWR | O_CLOEXEC); | |
3656 | if (IS_ERR(file)) { | |
3657 | put_unused_fd(ret); | |
3658 | ret = PTR_ERR(file); | |
3659 | goto err; | |
3660 | } | |
3661 | ||
3662 | #if defined(CONFIG_UNIX) | |
3663 | ctx->ring_sock->file = file; | |
6b06314c | 3664 | ctx->ring_sock->sk->sk_user_data = ctx; |
2b188cc1 JA |
3665 | #endif |
3666 | fd_install(ret, file); | |
3667 | return ret; | |
3668 | err: | |
3669 | #if defined(CONFIG_UNIX) | |
3670 | sock_release(ctx->ring_sock); | |
3671 | ctx->ring_sock = NULL; | |
3672 | #endif | |
3673 | return ret; | |
3674 | } | |
3675 | ||
3676 | static int io_uring_create(unsigned entries, struct io_uring_params *p) | |
3677 | { | |
3678 | struct user_struct *user = NULL; | |
3679 | struct io_ring_ctx *ctx; | |
3680 | bool account_mem; | |
3681 | int ret; | |
3682 | ||
3683 | if (!entries || entries > IORING_MAX_ENTRIES) | |
3684 | return -EINVAL; | |
3685 | ||
3686 | /* | |
3687 | * Use twice as many entries for the CQ ring. It's possible for the | |
3688 | * application to drive a higher depth than the size of the SQ ring, | |
3689 | * since the sqes are only used at submission time. This allows for | |
3690 | * some flexibility in overcommitting a bit. | |
3691 | */ | |
3692 | p->sq_entries = roundup_pow_of_two(entries); | |
3693 | p->cq_entries = 2 * p->sq_entries; | |
3694 | ||
3695 | user = get_uid(current_user()); | |
3696 | account_mem = !capable(CAP_IPC_LOCK); | |
3697 | ||
3698 | if (account_mem) { | |
3699 | ret = io_account_mem(user, | |
3700 | ring_pages(p->sq_entries, p->cq_entries)); | |
3701 | if (ret) { | |
3702 | free_uid(user); | |
3703 | return ret; | |
3704 | } | |
3705 | } | |
3706 | ||
3707 | ctx = io_ring_ctx_alloc(p); | |
3708 | if (!ctx) { | |
3709 | if (account_mem) | |
3710 | io_unaccount_mem(user, ring_pages(p->sq_entries, | |
3711 | p->cq_entries)); | |
3712 | free_uid(user); | |
3713 | return -ENOMEM; | |
3714 | } | |
3715 | ctx->compat = in_compat_syscall(); | |
3716 | ctx->account_mem = account_mem; | |
3717 | ctx->user = user; | |
3718 | ||
3719 | ret = io_allocate_scq_urings(ctx, p); | |
3720 | if (ret) | |
3721 | goto err; | |
3722 | ||
6c271ce2 | 3723 | ret = io_sq_offload_start(ctx, p); |
2b188cc1 JA |
3724 | if (ret) |
3725 | goto err; | |
3726 | ||
3727 | ret = io_uring_get_fd(ctx); | |
3728 | if (ret < 0) | |
3729 | goto err; | |
3730 | ||
3731 | memset(&p->sq_off, 0, sizeof(p->sq_off)); | |
75b28aff HV |
3732 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
3733 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
3734 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
3735 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
3736 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
3737 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
3738 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
3739 | |
3740 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
3741 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
3742 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
3743 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
3744 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
3745 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
3746 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
ac90f249 JA |
3747 | |
3748 | p->features = IORING_FEAT_SINGLE_MMAP; | |
2b188cc1 JA |
3749 | return ret; |
3750 | err: | |
3751 | io_ring_ctx_wait_and_kill(ctx); | |
3752 | return ret; | |
3753 | } | |
3754 | ||
3755 | /* | |
3756 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
3757 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
3758 | * params structure passed in. | |
3759 | */ | |
3760 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
3761 | { | |
3762 | struct io_uring_params p; | |
3763 | long ret; | |
3764 | int i; | |
3765 | ||
3766 | if (copy_from_user(&p, params, sizeof(p))) | |
3767 | return -EFAULT; | |
3768 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
3769 | if (p.resv[i]) | |
3770 | return -EINVAL; | |
3771 | } | |
3772 | ||
6c271ce2 JA |
3773 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
3774 | IORING_SETUP_SQ_AFF)) | |
2b188cc1 JA |
3775 | return -EINVAL; |
3776 | ||
3777 | ret = io_uring_create(entries, &p); | |
3778 | if (ret < 0) | |
3779 | return ret; | |
3780 | ||
3781 | if (copy_to_user(params, &p, sizeof(p))) | |
3782 | return -EFAULT; | |
3783 | ||
3784 | return ret; | |
3785 | } | |
3786 | ||
3787 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
3788 | struct io_uring_params __user *, params) | |
3789 | { | |
3790 | return io_uring_setup(entries, params); | |
3791 | } | |
3792 | ||
edafccee JA |
3793 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
3794 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
3795 | __releases(ctx->uring_lock) |
3796 | __acquires(ctx->uring_lock) | |
edafccee JA |
3797 | { |
3798 | int ret; | |
3799 | ||
35fa71a0 JA |
3800 | /* |
3801 | * We're inside the ring mutex, if the ref is already dying, then | |
3802 | * someone else killed the ctx or is already going through | |
3803 | * io_uring_register(). | |
3804 | */ | |
3805 | if (percpu_ref_is_dying(&ctx->refs)) | |
3806 | return -ENXIO; | |
3807 | ||
edafccee | 3808 | percpu_ref_kill(&ctx->refs); |
b19062a5 JA |
3809 | |
3810 | /* | |
3811 | * Drop uring mutex before waiting for references to exit. If another | |
3812 | * thread is currently inside io_uring_enter() it might need to grab | |
3813 | * the uring_lock to make progress. If we hold it here across the drain | |
3814 | * wait, then we can deadlock. It's safe to drop the mutex here, since | |
3815 | * no new references will come in after we've killed the percpu ref. | |
3816 | */ | |
3817 | mutex_unlock(&ctx->uring_lock); | |
edafccee | 3818 | wait_for_completion(&ctx->ctx_done); |
b19062a5 | 3819 | mutex_lock(&ctx->uring_lock); |
edafccee JA |
3820 | |
3821 | switch (opcode) { | |
3822 | case IORING_REGISTER_BUFFERS: | |
3823 | ret = io_sqe_buffer_register(ctx, arg, nr_args); | |
3824 | break; | |
3825 | case IORING_UNREGISTER_BUFFERS: | |
3826 | ret = -EINVAL; | |
3827 | if (arg || nr_args) | |
3828 | break; | |
3829 | ret = io_sqe_buffer_unregister(ctx); | |
3830 | break; | |
6b06314c JA |
3831 | case IORING_REGISTER_FILES: |
3832 | ret = io_sqe_files_register(ctx, arg, nr_args); | |
3833 | break; | |
3834 | case IORING_UNREGISTER_FILES: | |
3835 | ret = -EINVAL; | |
3836 | if (arg || nr_args) | |
3837 | break; | |
3838 | ret = io_sqe_files_unregister(ctx); | |
3839 | break; | |
9b402849 JA |
3840 | case IORING_REGISTER_EVENTFD: |
3841 | ret = -EINVAL; | |
3842 | if (nr_args != 1) | |
3843 | break; | |
3844 | ret = io_eventfd_register(ctx, arg); | |
3845 | break; | |
3846 | case IORING_UNREGISTER_EVENTFD: | |
3847 | ret = -EINVAL; | |
3848 | if (arg || nr_args) | |
3849 | break; | |
3850 | ret = io_eventfd_unregister(ctx); | |
3851 | break; | |
edafccee JA |
3852 | default: |
3853 | ret = -EINVAL; | |
3854 | break; | |
3855 | } | |
3856 | ||
3857 | /* bring the ctx back to life */ | |
3858 | reinit_completion(&ctx->ctx_done); | |
3859 | percpu_ref_reinit(&ctx->refs); | |
3860 | return ret; | |
3861 | } | |
3862 | ||
3863 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
3864 | void __user *, arg, unsigned int, nr_args) | |
3865 | { | |
3866 | struct io_ring_ctx *ctx; | |
3867 | long ret = -EBADF; | |
3868 | struct fd f; | |
3869 | ||
3870 | f = fdget(fd); | |
3871 | if (!f.file) | |
3872 | return -EBADF; | |
3873 | ||
3874 | ret = -EOPNOTSUPP; | |
3875 | if (f.file->f_op != &io_uring_fops) | |
3876 | goto out_fput; | |
3877 | ||
3878 | ctx = f.file->private_data; | |
3879 | ||
3880 | mutex_lock(&ctx->uring_lock); | |
3881 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
3882 | mutex_unlock(&ctx->uring_lock); | |
3883 | out_fput: | |
3884 | fdput(f); | |
3885 | return ret; | |
3886 | } | |
3887 | ||
2b188cc1 JA |
3888 | static int __init io_uring_init(void) |
3889 | { | |
3890 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC); | |
3891 | return 0; | |
3892 | }; | |
3893 | __initcall(io_uring_init); |